HOPF ALGEBRAS IN COMBINATORICS

DARIJ GRINBERG AND VICTOR REINER

Contents
Introduction 2
1. What is a Hopf algebra? 4
1.1. Algebras 4
1.2. Coalgebras 6
1.3. Morphisms, tensor products, and bialgebras 6
1.4. Antipodes and Hopf algebras 11
1.5. Commutativity, cocommutativity 20
1.6. Duals 22
1.7. Infinite sums and Leray’s theorem 26
2. Review of symmetric functions Λ as Hopf algebra 35
2.1. Definition of Λ 35
2.2. Other Bases 37
2.3. Comultiplications 44
2.4. The antipode, the involution ω, and algebra generators 46
2.5. Cauchy product, Hall inner product, self-duality 48
2.6. Bialternants, Littlewood-Richardson: Stembridge’s concise proof 58
2.7. The Pieri and Assaf-McNamara skew Pieri rule 62
2.8. Skewing and Lam’s proof of the skew Pieri rule 65
2.9. Assorted exercises on symmetric functions 69
3. Zelevinsky’s structure theory of positive self-dual Hopf algebras 82
3.1. Self-duality implies polynomiality 82
3.2. The decomposition theorem 85
3.3. Λ is the unique indecomposable PSH 88
4. Complex representations for Sn , wreath products, GLn (Fq ) 95
4.1. Review of complex character theory 95
4.2. Three towers of groups 103
4.3. Bialgebra and double cosets 105
4.4. Symmetric groups 112
4.5. Wreath products 115
4.6. General linear groups 117
4.7. Steinberg’s unipotent characters 118
4.8. Examples: GL2 (F2 ) and GL3 (F2 ) 119
4.9. The Hall algebra 121
5. Quasisymmetric functions and P -partitions 127
5.1. Definitions, and Hopf structure 127
5.2. The fundamental basis and P -partitions 132
5.3. The Hopf algebra NSym dual to QSym 141
6. Polynomial generators for QSym and Lyndon words 148
6.1. Lyndon words 148
6.2. Shuffles and Lyndon words 162
6.3. Radford’s theorem on the shuffle algebra 174

Date: August 20, 2017 (with minor corrections made on April 2, 2018).
Key words and phrases. Hopf algebra, combinatorics.
1
2 DARIJ GRINBERG AND VICTOR REINER

6.4. Polynomial freeness of QSym: statement and easy parts 177

6.5. Polynomial freeness of QSym: the general case 180
7. Aguiar-Bergeron-Sottile character theory Part I: QSym as a terminal object 189
7.1. Characters and the universal property 189
7.2. Example: Ehrenborg’s quasisymmetric function of a ranked poset 192
7.3. Example: Stanley’s chromatic symmetric function of a graph 196
7.4. Example: The quasisymmetric function of a matroid 202
8. The Malvenuto-Reutenauer Hopf algebra of permutations 211
8.1. Definition and Hopf structure 211
9. Further topics 218
10. Some open problems and conjectures 219
11. Appendix: Some basics 219
11.1. Linear expansions and triangularity 219
Acknowledgements 223
References 223

Introduction
The concept of a Hopf algebra crystallized out of algebraic topology and the study of algebraic groups
in the 1940s and 1950s (see [8] and [33] for its history). Being a fairly elementary algebraic notion itself, it
subsequently found applications in other mathematical disciplines, and is now particularly commonplace in
representation theory1.
These notes concern themselves (after a brief introduction into the algebraic foundations of Hopf algebra
theory in Chapter 1) with the Hopf algebras that appear in combinatorics. These Hopf algebras tend to have
bases naturally parametrized by combinatorial objects (partitions, compositions, permutations, tableaux,
graphs, trees, posets, polytopes, etc.), and their Hopf-algebraic operations often encode basic operations on
these objects2. Combinatorial results can then be seen as particular cases of general algebraic properties
of Hopf algebras (e.g., the multiplicativity of the Möbius function can be recovered from the fact that the
antipode of a Hopf algebra is an algebra anti-endomorphism), and many interesting invariants of combina-
torial objects turn out to be evaluations of Hopf morphisms. In some cases (particularly that of symmetric
functions), the rigidity in the structure of a Hopf algebra can lead to enlightening proofs.
One of the most elementary interesting examples of a combinatorial Hopf algebra is that of the symmetric
functions. We will devote all of Chapter 2 to studying it, deviating from the usual treatments (such as in
Stanley [183, Ch. 7], Sagan [165] and Macdonald [125]) by introducing the Hopf-algebraic structure early
on and using it to obtain combinatorial results. Chapter 3 will underpin the importance of this algebra
by proving Zelevinsky’s main theorem of PSH theory, which (roughly) claims that a Hopf algebra over
Z satisfying a certain set of axioms must be a tensor product of copies of the Hopf algebra of symmetric
functions. These axioms are fairly restrictive, so this result is far from curtailing the diversity of combinatorial
Hopf algebras; but they are natural enough that, as we will see in Chapter 4, they are satisfied for a
Hopf algebra of representations of symmetric groups. As a consequence, this Hopf algebra will be revealed
isomorphic to the symmetric functions – this is the famous Frobenius correspondence between symmetric
functions and characters of symmetric groups, usually obtained through other ways ([60, §7.3], [165, §4.7]).
We will further elaborate on the representation theories of wreath products and general linear groups over
finite fields; while Zelevinsky’s PSH theory does not fully explain the latter, it illuminates it significantly.
In the next chapters, we will study further examples of combinatorial Hopf algebras: the quasisymmetric
functions and the noncommutative symmetric functions in Chapter 5, various other algebras (of graphs,
posets, matroids, etc.) in Chapter 7, and the Malvenuto-Reutenauer Hopf algebra of permutations in Chap-
ter 8.

1where it provides explanations for similarities between group representations and Lie algebra representations
2such as concatenating two compositions, or taking the disjoint union of two graphs – but, more often, operations which
return a multiset of results, such as cutting a composition into two pieces at all possible places, or partitioning a poset into two
subposets in every way that satisfies a certain axiom
 HOPF ALGEBRAS IN COMBINATORICS 3

The main prerequisite for reading these notes is a good understanding of graduate algebra3, in partic-
ular multilinear algebra (tensor products, symmetric powers and exterior powers)4 and basic categorical
language5. In Chapter 4, familiarity with representation theory of finite groups (over C) is assumed, along
with the theory of finite fields and (at some places) the rational canonical form of a matrix. Only basic
knowledge of combinatorics is required (except for a few spots in Chapter 7), and familiarity with geometry
and topology is needed only to understand some tangential remarks. The concepts of Hopf algebras and
coalgebras and the basics of symmetric function theory will be introduced as needed. We will work over a
commutative base ring most of the time, but no commutative algebra (besides, occasionally, properties of
modules over a PID) will be used.
These notes began as an accompanying text for Fall 2012 Math 8680 Topics in Combinatorics, a graduate
class taught by the second author at the University of Minnesota. The first author has since added many
exercises (and solutions6) as well as Chapter 6 on Lyndon words and the polynomiality of QSym. The notes
might still grow, and any comments, corrections and complaints are welcome!
The course was an attempt to focus on examples that we find interesting, but which are hard to find fully
explained currently in books or in one paper. Much of the subject of combinatorial Hopf algebras is fairly
recent (1990s onwards) and still spread over research papers, although sets of lecture notes do exist, such
as Foissy’s [57]. A reference which we discovered late, having a great deal of overlap with these notes is
Hazewinkel, Gubareni, and Kirichenko [78]. References for the purely algebraic theory of Hopf algebras are
much more frequent (see the beginning of Chapter 1 for a list). Another recent text that has a significant
amount of material in common with ours (but focuses on representation theory and probability applications)
is Méliot’s [135].
Be warned that our notes are highly idiosyncratic in choice of topics, and they steal heavily from the
sources in the bibliography.

Warnings: Unless otherwise specified ...

• k here usually denotes a commutative ring7.
• all maps between k-modules are k-linear.
• every ring or k-algebra is associative and has a 1, and every ring morphism or k-algebra morphism
preserves the 1’s.
• all k-algebras A have the property that (λ1A ) a = a (λ1A ) = λa for all λ ∈ k and a ∈ A.
• all tensor products are over k (unless a subscript specifies a different base ring).
• 1 will denote the multiplicative identity in some ring like k or in some k-algebra (sometimes also the
identity of a group written multiplicatively).
• for any set S, we denote by idS (or by id) the identity map on S.
• The symbols ⊂ (for “subset”) and < (for “subgroup”) don’t imply properness (so Z ⊂ Z and Z < Z).
• the n-th symmetric group (i.e., the group of all permutations of {1, 2, . . . , n}) is denoted Sn .
• The product of permutations a ∈ Sn and b ∈ Sn is defined by (ab)(i) = a(b(i)) for all i.
• Words over (or in) an alphabet I simply mean finite tuples of elements of a set I. It is custom-
ary to write such a word (a1 , a2 , . . . , ak ) as a1 a2 . . . ak when this is not likely to be confused for
multiplication.
• N := {0, 1, 2, . . .}.
• if i and j are any two objects, then δi,j denotes the Kronecker delta of i and j; this is the integer 1
if i = j and 0 otherwise.
• several objects s1 , s2 , . . . , sk are said to be distinct if every i 6= j satisfy si 6= sj .
• similarly, several sets S1 , S2 , . . . , Sk are said to be disjoint if every i 6= j satisfy Si ∩ Sj = ∅.

3William Schmitt’s expositions [171] are tailored to a reader interested in combinatorial Hopf algebras; his notes on modules
and algebras cover a significant part of what we need from abstract algebra, whereas those on categories cover all category
theory we will use and much more.
4Keith Conrad’s expository notes [37] are useful, even if not comprehensive, sources for the latter.
5We also will use a few nonstandard notions from linear algebra that are explained in the Appendix (Chapter 11).
6The version of the notes you are reading does not contain said solutions. The version that does can be downloaded from
[Link] or compiled from the sourcecode.
7As explained below, “ring” means “associative ring with 1”. The most important cases are when k is a field or when k = Z.
4 DARIJ GRINBERG AND VICTOR REINER

F
• the symbol t (and the corresponding quantifier ) denotes a disjoint union of sets or posets. For
Fk
example, if S1 , S2 , . . . , Sk are k sets, then i=1 Si is their disjoint union. This disjoint union can
mean either of the following two things:
Sk
• It can mean the union i=1 Si in the case when the sets S1 , S2 , . . . , Sk are disjoint. This is
called an “internal disjoint union”, and is simply a way to refer to the union of sets while
simultaneously claiming that these sets are disjoint. Thus, of course, it is only well-defined if
the sets are disjoint.
Sk
• It can also mean the union i=1 {i} × Si . This is called an “external disjoint union”, and is
well-defined whether or not the sets S1 , S2 , . . . , Sk are disjoint; it is a way to assemble the sets
S1 , S2 , . . . , Sk into a larger set which contains a copy of each of their elements that “remembers”
which set this element comes from.
The two meanings are different, but in the case when S1 , S2 , . . . , Sk are disjoint, they are isomorphic.
We hope the reader will not have a hard time telling which of them we are trying to evoke.
Similarly, the notion of a direct sum of k-modules has two meanings (“internal direct sum” and
“external direct sum”).
• A sequence (w1 , w2 , . . . , wk ) of numbers (or, more generally, of elements of a poset) is said to be
strictly increasing (or, for short, increasing) if it satisfies w1 < w2 < · · · < wk . A sequence
(w1 , w2 , . . . , wk ) of numbers (or, more generally, of elements of a poset) is said to be weakly in-
creasing (or nondecreasing) if it satisfies w1 ≤ w2 ≤ · · · ≤ wk . Reverting the inequalities, we obtain
the definitions of a strictly decreasing (a.k.a. decreasing) and of a weakly decreasing (a.k.a. nonin-
creasing) sequence. All these definitions extend in an obvious way to infinite sequences. Note that
“nondecreasing” is not the same as “not decreasing”; for example, any sequence having at most one
entry is both decreasing and nondecreasing, whereas the sequence (1, 3, 1) is neither.
Hopefully context will resolve some of the ambiguities.

1. What is a Hopf algebra?

The standard references for Hopf algebras are Abe [1] and Sweedler [189], and some other good ones are
[32, 34, 44, 78, 92, 103, 139, 156, 174, 201]. See also Foissy [57] and Manchon [132] for introductions to
Hopf algebras tailored to combinatorial applications. Most texts only study Hopf algebras over fields (with
exceptions such as [34, 32, 201]). We will work over arbitrary commutative rings8, which requires some more
care at certain points (but we will not go deep enough into the algebraic theory to witness the situation over
commutative rings diverge seriously from that over fields).
Let’s build up the definition of Hopf algebra structure bit-by-bit, starting with the more familiar definition
of algebras.

1.1. Algebras. The following definition of k-algebras may look unfamiliar, but it is merely a restatement
of their classical definition using tensors9 and k-linear maps:
m
Definition 1.1.1. An associative k-algebra A is a k-module with a k-linear associative operation A⊗A → A,
u
and a k-linear unit k → A sending 1 in k to the two-sided multiplicative identity element 1 in A. One can
rephrase this by saying that these diagrams commute:

(1.1.1) A⊗A⊗A
m⊗id id ⊗m

x &
A⊗A A⊗A
m m

& x
A

8and we will profit from this generality in Chapters 3 and 4, where we will be applying the theory of Hopf algebras to k = Z
in a way that would not be possible over k = Q
9i.e., the multiplication is encoded by a k-linear map A ⊗ A → A instead of a k-bilinear map A × A → A
 HOPF ALGEBRAS IN COMBINATORICS 5

(1.1.2) A⊗k o A / k⊗A

id ⊗u id u⊗id
  
A⊗A
m /Ao m
A⊗A
where the maps A → A ⊗ k and A → k ⊗ A are the isomorphisms sending a 7→ a ⊗ 1 and a 7→ 1 ⊗ a.
We abbreviate “associative k-algebra” as “k-algebra” (associativity is assumed unless otherwise specified)
or as “algebra” (when k is clear from the context).
Well-known examples of k-algebras are tensor and symmetric algebras, which we can think of as algebras
of words and multisets, respectively.
Example 1.1.2. If V is a k-module and n ∈ N, then the n-fold tensor power V ⊗n of V is the k-module
V ⊗ V ⊗ · · · ⊗ V . (For n = 0, this is the k-module k, spanned by the “empty tensor” 1k .)
| {z }
n times
The tensor algebra T (V ) = n≥0 V ⊗n on a k-module V is an associative k-algebra spanned (as k-module)
L
by decomposable tensors v1 v2 · · · vk := v1 ⊗ v2 ⊗ · · · ⊗ vk with k ∈ N and v1 , v2 , . . . , vk ∈ V . Its multiplication
is defined k-linearly by
m (v1 v2 · · · vk ⊗ w1 w2 · · · w` ) := v1 v2 · · · vk w1 w2 · · · w`
10
for all k, ` ∈ N and v1 , v2 , . . . , vk , w1 , w2 , . . . , w` in V . The unit map u : k → T (V ) sends 1k to the empty
tensor 1T (V ) = 1k ∈ k = V ⊗0 .
If V is a free k-module, say with k-basis {xi }i∈I , then T (V ) has a k-basis of decomposable tensors
xi1 · · · xik := xi1 ⊗ · · · ⊗ xik indexed by words (i1 , . . . , ik ) in the alphabet I, and the multiplication on this
basis is given by concatenation of words:
m(xi1 · · · xik ⊗ xj1 · · · xj` ) = xi1 · · · xik xj1 · · · xj` .
Recall that in an algebra A, when one has a two-sided ideal J ⊂ A, meaning a k-submodule with m(J ⊗
A), m(A ⊗ J) ⊂ J, then one can form a quotient algebra A/J.
L n
Example 1.1.3. The symmetric algebra Sym(V ) = n≥0 Sym (V ) is the quotient of T (V ) by the two-
sided ideal generated by all elements xy − yx with x, y in V . When V is a free k-module with basis {xi }i∈I ,
this symmetric algebra S (V ) can be identified with a (commutative) polynomial algebra k[xi ]i∈I , having a
k-basis of (commutative) monomials xi1 · · · xik as (i1 , . . . , ik ) runs through all finite multisubsets11 of I, and
with multiplication defined k-linearly via multiset union.
Note that the k-module k itself canonically becomes a k-algebra. Its associative operation m : k ⊗ k → k
is the canonical isomorphism k ⊗ k → k, and its unit u : k → k is the identity map.
Topology and group theory give more examples.
Example 1.1.4. The cohomology algebra H ∗ (X; k) = i≥0 H i (X; k) with coefficients in k for a topological
L
u
space X has an associative cup product. Its unit k = H ∗ (pt; k) → H ∗ (X; k) is induced from the unique
(continuous) map X → pt, where pt is a one-point space.
Example 1.1.5. For a group G, the group algebra kG has k-basis {tg }g∈G and multiplication defined
k-linearly by tg th = tgh , and unit defined by u(1) = te , where e is the identity element of G.

10Some remarks about our notation (which we are using here and throughout these notes) are in order.
Since we are working with tensor products of k-modules like T (V ) – which themselves are made of tensors – here, we must
specify what the ⊗ sign means in expressions like a ⊗ b where a and b are elements of T (V ). Our convention is the following:
When a and b are elements of a tensor algebra T (V ), we always understand a ⊗ b to mean the pure tensor a ⊗ b ∈ T (V ) ⊗ T (V )
rather than the product of a and b inside the tensor algebra T (V ). The latter product will plainly be written ab.
The operator precedence between ⊗ and multiplication in T (V ) is such that multiplication in T (V ) binds more tightly than
the ⊗ sign; e.g., the term ab ⊗ cd means (ab) ⊗ (cd). The same convention applies to any algebra instead of T (V ).
11By a multisubset of a set S, we mean a multiset each of whose elements belongs to S (but can appear arbitrarily often).
6 DARIJ GRINBERG AND VICTOR REINER

1.2. Coalgebras. If we are to think of the multiplication A ⊗ A → A in an algebra as putting together two
basis elements of A to get a sum of basis elements of A, then coalgebra structure should be thought of as
taking basis elements apart.
Definition 1.2.1. A co-associative k-coalgebra C is a k-module C with a comultiplication, that is, a k-linear
∆ 
map C → C ⊗ C, and a k-linear counit C → k making commutative the diagrams as in (1.1.1), (1.1.2) but
with all arrows reversed:

(1.2.1) C⊗
8 C ⊗f C
∆⊗id id ⊗∆

C ⊗ Cf C8 ⊗ C
∆ ∆

(1.2.2) C ⊗O k /Co k ⊗O C
O
id ⊗ id ⊗id

C ⊗C o C / C ⊗C
∆ ∆

Here the maps C ⊗ k → C and k ⊗ C → C are the isomorphisms sending c ⊗ 1 7→ c and 1 ⊗ c 7→ c.

We abbreviate “co-associative k-coalgebra” as “k-coalgebra” (co-associativity, i.e., the commutativity of
the diagram (1.2.1), is assumed unless otherwise specified) or as “coalgebra” (when k is clear from the
context).
Sometimes, the word “coproduct” is used as a synonym for “comultiplication”12.
One often uses the Sweedler notation
X X
∆(c) = c1 ⊗ c2 = c1 ⊗ c2
(c)

to
P abbreviate formulas involving ∆. For example, commutativity of the left square in (1.2.2) asserts that
(c) c1 (c2 ) = c.
The k-module k itself canonically becomes a k-coalgebra, with its comultiplication ∆ : k → k ⊗ k being
the canonical isomorphism k → k ⊗ k, and its counit  : k → k being the identity map.
L
Example 1.2.2. Let k be a field. The homology H∗ (X; k) = i≥0 Hi (X; k) for a topological space X
is naturally a coalgebra: the (continuous) diagonal embedding X → X × X sending x 7→ (x, x) induces a
coassociative map
H∗ (X; k) → H∗ (X × X; k) ∼
= H∗ (X; k) ⊗ H∗ (X; k)
in which the last isomorphism comes from the Künneth theorem with field coefficients k. As before, the

unique (continuous) map X → pt induces the counit H∗ (X; k) → H∗ (pt; k) ∼ = k.
Exercise 1.2.3. Given a k-module C and a k-linear map ∆ : C → C ⊗ C. Prove that there exists at most
one k-linear map  : C → k such that the diagram (1.2.2) commutes.
1.3. Morphisms, tensor products, and bialgebras.
ϕ
Definition 1.3.1. A morphism of algebras A → B makes these diagrams commute:

(1.3.1) AO
ϕ
/B A_
ϕ
/B
O ?
mA mB
uA uB

A⊗A
ϕ⊗ϕ
/ B⊗B k

12although the word “coproduct” already has a different meaning in algebra

 HOPF ALGEBRAS IN COMBINATORICS 7

Here the subscripts on mA , mB , uA , uB indicate for which algebra they are part of the structure– we will
occasionally use such conventions from now on.
ϕ
Similarly a morphism of coalgebras is a k-linear map C → D making the reverse diagrams commute:

(1.3.2) C
ϕ
/D C
ϕ
/D

∆C ∆D
C D
   
C ⊗C
ϕ⊗ϕ
/ D⊗D k
f
Example 1.3.2. Let k be a field. Continuous maps X → Y of topological spaces induce algebra morphisms
H ∗ (Y ; k) → H ∗ (X; k), and coalgebra morphisms H∗ (X; k) → H∗ (Y ; k).
Coalgebra morphisms behave similarly to algebra morphisms in many regards: For example, the inverse
of an invertible coalgebra morphism is again a coalgebra morphism13. Thus, invertible coalgebra morphisms
are called coalgebra isomorphisms.
Definition 1.3.3. Given two k-algebras A, B, their tensor product A ⊗ B also becomes a k-algebra defining
the multiplication bilinearly via
m((a ⊗ b) ⊗ (a0 ⊗ b0 )) := aa0 ⊗ bb0
or in other words mA⊗B is the composite map

A⊗B⊗A⊗B
id ⊗T ⊗id
/ A⊗A⊗B⊗B mA ⊗mB
/ A⊗B
where T is the twist map B ⊗ A → A ⊗ B that sends b ⊗ a 7→ a ⊗ b. (See Exercise 1.3.4(a) below for a proof
that this k-algebra A ⊗ B is well-defined.)
Here we are omitting the topologist’s sign in the twist map which should be present for graded algebras
and coalgebras that come from cohomology and homology: for homogeneous elements a and b the topologist’s
twist map sends
(1.3.3) b ⊗ a 7−→ (−1)deg(a) deg(b) a ⊗ b.
This means that, if one is using the topologists’ conventions, most of our examples which we later call graded
should actually be considered to live in only even degrees (which can be achieved, e.g., by artificially doubling
their grading). We will, however, keep to our own definitions (so the twist map sends b ⊗ a 7→ a ⊗ b) unless
otherwise noted.
uA⊗B
The unit element of A ⊗ B is 1A ⊗ 1B , meaning that the unit map k → A ⊗ B is the composite

k / k⊗k uA ⊗uB
/ A⊗B .
Similarly, given two coalgebras C, D, one can make C ⊗ D a coalgebra in which the comultiplication and
counit maps are the composites of

C ⊗D
∆C ⊗∆D
/ C ⊗C ⊗D⊗D id ⊗T ⊗id
/ C ⊗D⊗C ⊗D
and
C ⊗D
C ⊗D
/ k⊗k /k.
(See Exercise 1.3.4(b) below for a proof that this k-coalgebra C ⊗ D is well-defined.)
Exercise 1.3.4. (a) Let A and B be two k-algebras. Show that the k-algebra A ⊗ B introduced in
Definition 1.3.3 is actually well-defined (i.e., its multiplication and unit satisfy the axioms of a
k-algebra).
(b) Let C and D be two k-coalgebras. Show that the k-coalgebra C ⊗ D introduced in Definition 1.3.3
is actually well-defined (i.e., its comultiplication and counit satisfy the axioms of a k-coalgebra).
It is straightforward to show that the concept of tensor products of algebras and of coalgebras satisfy the
properties one would expect:
13The easy proof of this fact is left to the reader.
8 DARIJ GRINBERG AND VICTOR REINER

• For any three k-coalgebras C, D and E, the k-linear map

(C ⊗ D) ⊗ E → C ⊗ (D ⊗ E) , (c ⊗ d) ⊗ e 7→ c ⊗ (d ⊗ e)
is a coalgebra isomorphism. This allows us to speak of the k-coalgebra C ⊗ D ⊗ E without worrying
about the parenthesization.
• For any two k-coalgebras C and D, the k-linear map
T : C ⊗ D → D ⊗ C, c ⊗ d 7→ d ⊗ c
is a coalgebra isomorphism.
• For any k-coalgebra C, the k-linear maps
C → k ⊗ C, c 7→ 1 ⊗ c and
C → C ⊗ k, c 7→ c ⊗ 1
are coalgebra isomorphisms.
• Similar properties hold for algebras instead of coalgebras.
One of the first signs that these definitions interact nicely is the following straightforward proposition.

Proposition 1.3.5. When A is both a k-algebra and a k-coalgebra, the following are equivalent:
• (∆, ) are morphisms for the algebra structure (m, u).
• (m, u) are morphisms for the coalgebra structure (∆, ).
• These four diagrams commute:
A⊗A
∆⊗∆

w
A⊗A⊗A⊗A m


id ⊗T ⊗id A


A⊗A⊗A⊗A ∆

(1.3.4) m⊗m
' 
A⊗A
A⊗A
⊗
/ k⊗k k
u /A

m m ∆ ∆
   
A /k k⊗k / A⊗A
 u⊗u

k
id /k
?
u 

A
Exercise 1.3.6. (a) If A, A0 , B and B 0 are four k-algebras, and f : A → A0 and g : B → B 0 are two
k-algebra homomorphisms, then show that f ⊗ g : A ⊗ B → A0 ⊗ B 0 is a k-algebra homomorphism.
(b) If C, C 0 , D and D0 are four k-coalgebras, and f : C → C 0 and g : D → D0 are two k-coalgebra
homomorphisms, then show that f ⊗ g : C ⊗ D → C 0 ⊗ D0 is a k-coalgebra homomorphism.
Definition 1.3.7. Call the k-module A a k-bialgebra if it is a k-algebra and k-coalgebra satisfying the three
equivalent conditions in Proposition 1.3.5.

Example 1.3.8. For a group G, one can make the group algebra kG a coalgebra with counit kG → k
∆
mapping tg 7→ 1 for all g in G, and with comultiplication kG → kG ⊗ kG given by ∆(tg ) := tg ⊗ tg .
 HOPF ALGEBRAS IN COMBINATORICS 9

Checking the various diagrams in (1.3.4) commute is easy. For example, one can check the pentagonal
diagram on each basis element tg ⊗ th :

tg ⊗ th
∆⊗∆

v
tg ⊗ tg ⊗ th ⊗ th m


id ⊗T ⊗id tgh


∆
tg ⊗ th ⊗ tg ⊗ th

m⊗m
( 
tgh ⊗ tgh

Remark 1.3.9. In fact, one can think of adding a bialgebra structure to a k-algebra A as a way of making
A-modules M, N have an A-module structure on their tensor product M ⊗ N : the algebra A ⊗ A already
acts naturally on M ⊗ N , so one can let a in A act via ∆(a) in A ⊗ A. In the theory of group representations
over k, that is, kG-modules M , this is how one defines the diagonal action of G on M ⊗ N , namely tg acts
as tg ⊗ tg .

Definition 1.3.10. An element x in a coalgebra for which ∆(x) = x ⊗ x and (x) = 1 is called group-like.
An element x in a bialgebra for which ∆(x) = 1 ⊗ x + x ⊗ 1 is called primitive. We shall also sometimes
abbreviate “primitive element” as “primitive”.

Example 1.3.11. The tensor algebra T (V ) = n≥0 V ⊗n is a coalgebra, with counit  equal to the identity
L

on V ⊗0 = k and the zero map on V ⊗n for n > 0, and with comultiplication defined to make the elements x
in V ⊗1 = V all primitive:
∆(x) := 1 ⊗ x + x ⊗ 1 for x ∈ V ⊗1 .
Since the elements of V generate T (V ) as a k-algebra, and since T (V )⊗T (V ) is also an associative k-algebra,
the universal property of T (V ) as the free associative k-algebra on the generators V allows one to define
∆
T (V ) → T (V ) ⊗ T (V ) arbitrarily on V , and extend it as an algebra morphism.
It may not be obvious that this ∆ is coassociative, but one can note that

((id ⊗∆) ◦ ∆) (x) = x ⊗ 1 ⊗ 1 + 1 ⊗ x ⊗ 1 + 1 ⊗ 1 ⊗ x = ((∆ ⊗ id) ◦ ∆) (x)

for every x in V . Hence the two maps (id ⊗∆) ◦ ∆ and (∆ ⊗ id) ◦ ∆, considered as algebra morphisms
T (V ) → T (V ) ⊗ T (V ) ⊗ T (V ), must coincide on every element of T (V ) since they coincide on V . We leave
it as an exercise to check the map  defined as above satisfies the counit axioms (1.2.2).
Here is a sample calculation in T (V ) when V has basis {x, y, z}:

∆(xyz) = ∆(x)∆(y)∆(z)
= (1 ⊗ x + x ⊗ 1)(1 ⊗ y + y ⊗ 1)(1 ⊗ z + z ⊗ 1)
= (1 ⊗ xy + x ⊗ y + y ⊗ x + xy ⊗ 1)(1 ⊗ z + z ⊗ 1)
= 1 ⊗ xyz + x ⊗ yz + y ⊗ xz + z ⊗ xy
+ xy ⊗ z + xz ⊗ y + yz ⊗ x + xyz ⊗ 1.

This illustrates the idea that comultiplication “takes basis elements apart”. Here for any v1 , v2 , . . . , vn in V
one has
X
∆ (v1 v2 · · · vn ) = vj1 · · · vjr ⊗ vk1 · · · vkn−r
10 DARIJ GRINBERG AND VICTOR REINER

where the sum is over ordered pairs (j1 , j2 , . . . , jr ) , (k1 , k2 , . . . , kn−r ) of complementary subwords of the word
(1, 2, . . . , n). 14 Equivalently (and in a more familiar language),
X
(1.3.5) ∆ (v1 v2 · · · vn ) = vI ⊗ v{1,2,...,n}\I ,
I⊂{1,2,...,n}

where vJ (for J a subset of {1, 2, . . . , n}) denotes the product of all vj with j ∈ J in the order of increasing
j.
Recall one can quotient a k-algebra A by a two-sided ideal J to obtain a quotient algebra A/J.
Definition 1.3.12. In a coalgebra C, a two-sided coideal is a k-submodule J ⊂ C for which
∆(J) ⊂ J ⊗ C + C ⊗ J
(J) = 0
The quotient k-module C/J then inherits a coalgebra structure15. Similarly, in a bialgebra A, a subset
J ⊂ A which is both a two-sided ideal and two-sided coideal gives rise to a quotient bialgebra A/J.
Exercise 1.3.13. Let A and C be two k-coalgebras, and f : A → C a surjective coalgebra homomorphism.
(a) If f is surjective, then show that ker f is a two-sided coideal of A.
(b) If k is a field, then show that ker f is a two-sided coideal of A.
Example 1.3.14. The symmetric algebra Sym(V ) was the quotient of T (V ) by the two-sided ideal J
generated by all commutators [x, y] = xy − yx for x, y in V . Note that x, y are primitive elements in T (V ),
and the following very reusable calculation shows that the commutator of two primitives is primitive:
∆[x, y] = ∆(xy − yx)
= (1 ⊗ x + x ⊗ 1)(1 ⊗ y + y ⊗ 1) − (1 ⊗ y + y ⊗ 1)(1 ⊗ x + x ⊗ 1)
= 1 ⊗ xy − 1 ⊗ yx + xy ⊗ 1 − yx ⊗ 1
(1.3.6)
+x⊗y+y⊗x−x⊗y−y⊗x
= 1 ⊗ (xy − yx) + (xy − yx) ⊗ 1
= 1 ⊗ [x, y] + [x, y] ⊗ 1.
In particular, the commutators [x, y] have ∆[x, y] in J ⊗T (V )+T (V )⊗J. They also satisfy ([x, y]) = 0. Since
they are generators for J as a two-sided ideal, it is not hard to see this implies ∆(J) ⊂ J ⊗ T (V ) + T (V ) ⊗ J,
and (J) = 0. Thus J is also a two-sided coideal, and Sym(V ) = T (V )/J inherits a bialgebra structure.
In fact we will see in Section 3.1 that symmetric algebras are the universal example of bialgebras which
are graded, connected, commutative, cocommutative. But first we should define some of these concepts.
L
Definition 1.3.15. A graded k-module V is one with a k-module direct sum decomposition V = n≥0 Vn .
Elements x in Vn are called homogeneous of degree n, or deg(x) = n.
One endowsLtensor products V ⊗ W of graded k-modules V, W with graded module structure in which
(V ⊗ W )n := i+j=n Vi ⊗ Wj .
ϕ
A k-linear map V → W between two graded k-modules is called graded if ϕ(Vn ) ⊂ Wn for all n. Say that
a k-algebra (coalgebra, bialgebra) is graded if it is a graded k-module and all of the relevant structure maps
(u, , m, ∆) are graded.
Say that a graded module V is connected if V0 ∼ = k.
Example 1.3.16. Let k be a field. A path-connected space X has its homology and cohomology
M
H∗ (X; k) = Hi (X; k)
i≥0
M
∗
H (X; k) = H i (X; k)
i≥0

14More formally speaking, the sum is over all permutations (j , j , . . . , j , k , k , . . . , k

1 2 r 1 2 n−r ) of (1, 2, . . . , n) satisfying j1 <
j2 < · · · < jr and k1 < k2 < · · · < kn−r .
15since J ⊗ C + C ⊗ J is contained in the kernel of the canonical map C ⊗ C → (C/J) ⊗ (C/J)
 HOPF ALGEBRAS IN COMBINATORICS 11

carrying the structure of connected graded coalgebras and algebras, respectively. If in addition, X is a
topological group, or even less strongly, a homotopy-associative H-space (e.g. the loop space ΩY on some
other space Y ), the continuous multiplication map X × X → X induces an algebra structure on H∗ (X; k)
and a coalgebra structure on H ∗ (X; k), so that each become bialgebras in the topologist’s sense (i.e., with
the twist as in (1.3.3)), and these bialgebras are dual to each other in a sense soon to be discussed. This
was Hopf’s motivation: the (co-)homology of a compact Lie group carries bialgebra structure that explains
why it takes a certain form; see Cartier [33, §2].
Example 1.3.17. Tensor algebras T (V ) and symmetric algebras Sym(V ) are graded, once one picks a
graded module structure for V ; then
deg(xi1 · · · xik ) = deg(xi1 ) + · · · + deg(xik )
if {xi }i∈I is a graded basis (and, more generally, v1 v2 · · · vk is homogeneous of degree i1 + i2 + · · · + ik if
each vj is a homogeneous element of V of degree ij ). Assuming that V0 = 0, the graded algebras T (V ) and
Sym(V ) are connected. For example, we will often say that all elements of V are homogeneous of degree 1,
but at other times, it will make sense to have V live in different (positive) degrees.
L
Exercise 1.3.18. Let A = n≥0 An be a graded k-bialgebra. We denote by p the set of all primitive
elements of A.
L
(a) Show that p is a graded k-submodule of A (that is, we have p = n≥0 (p ∩ An )).
(b) Show that p is a two-sided coideal of A.
Exercise 1.3.19. Let A be a connected graded k-bialgebra. Show that
(a) the k-submodule k = k · 1A of A lies in A0 ,
u
(b) u is an isomorphism k → A0 ,
(c) we have A0 = k · 1A , while L
(d) the two-sided ideal ker  is the k-module of positive degree elements I = n>0 An .

(e)  restricted to A0 is the inverse isomorphism A0 → k to u.
(f) for every x ∈ A, we have
∆(x) ∈ x ⊗ 1 + A ⊗ I;
(g) every x in I has comultiplication of the form
∆(x) = 1 ⊗ x + x ⊗ 1 + ∆+ (x)
where ∆+ (x) lies in I ⊗ I.
(h) every n > 0 and every x ∈ An satisfy ∆ (x) = 1 ⊗ x + x ⊗ 1 + ∆+ (x), where ∆+ (x) lies in
Pn−1
k=1 Ak ⊗ An−k .
(Use only the gradedness of the unit u and counit  maps, along with commutativity of diagrams (1.2.2),
and (1.3.4) and the connectedness of A.)
The tensor product of two bialgebras is canonically a bialgebra, as the following proposition shows:
Proposition 1.3.20. Let A and B be two k-bialgebras. Then, A ⊗ B is both a k-algebra and a k-coalgebra
(by Definition 1.3.3). These two structures, combined, turn A ⊗ B into a k-bialgebra.
Exercise 1.3.21. (a) Prove Proposition 1.3.20.
(b) Let G and H be two groups. Show that the k-bialgebra kG ⊗ kH (defined as in Proposition 1.3.20)
is isomorphic to the k-bialgebra k [G × H]. (The notation k [S] is a synonym for kS.)
1.4. Antipodes and Hopf algebras. There is one more piece of structure needed to make a bialgebra a
Hopf algebra, although it will come for free in the connected graded case.
Definition 1.4.1. For any coalgebra C and algebra A, one can endow the k-linear maps Hom(C, A) with
an associative algebra structure
P called the convolution algebra: Define the productP f ? g of two maps f, g in
Hom(C, A) by (f ? g)(c) = f (c1 )g(c2 ), using the Sweedler notation ∆(c) = c1 ⊗ c2 . Equivalently, f ? g
is the composite
f ⊗g
C
∆ / C ⊗C / A⊗A m /A.
The associativity of this multiplication ? is easy to check (see Exercise 1.4.2 below).
12 DARIJ GRINBERG AND VICTOR REINER

The map u ◦  is a two-sided identity element for ?, meaning that every f ∈ Hom(C, A) satisfies
X X
f (c1 )(c2 ) = f (c) = (c1 )f (c2 )

for all c ∈ C. One sees this by adding a top row to (1.2.2):

(1.4.1) A ⊗O k /Ao k ⊗O A
O
f ⊗id f id ⊗f

C ⊗O k /Co k ⊗O C
O
id ⊗ id ⊗id

C ⊗C o C / C ⊗C
∆ ∆

In particular, when one has a bialgebra A, the convolution product ? gives an associative algebra structure
on End(A) := Hom(A, A).
Exercise 1.4.2. Let C be a k-coalgebra and A be a k-algebra. Show that the binary operation ? on
Hom (C, A) is associative.
The following simple (but useful) property of convolution algebras says essentially that the k-algebra
(Hom (C, A) , ?) is a covariant functor in A and a contravariant functor in C, acting on morphisms by pre-
and post-composition:

Proposition 1.4.3. Let C and C 0 be two k-coalgebras, and let A and A0 be two k-algebras. Let γ : C → C 0
be a k-coalgebra morphism. Let α : A → A0 be a k-algebra morphism.
The map
Hom (C 0 , A) → Hom (C, A0 ) , f 7→ α ◦ f ◦ γ
is a k-algebra homomorphism from the convolution algebra (Hom (C 0 , A) , ?) to the convolution algebra
(Hom (C, A0 ) , ?).

Proof of Proposition 1.4.3. Denote this map by ϕ. We must show that ϕ is a k-algebra homomorphism.
Recall that α is an algebra morphism; thus, α ◦ mA = mA0 ◦ (α ⊗ α) and α ◦ uA = uA0 . Also, γ is a
coalgebra morphism; thus, ∆C 0 ◦ γ = (γ ⊗ γ) ◦ ∆C and C 0 ◦ γ = C .
Now, the definition of ϕ yields ϕ(uA ◦ C 0 ) = α ◦ uA ◦ C 0 ◦ γ = uA0 ◦ C ; in other words, ϕ sends the unity
| {z } | {z }
=uA0 =C
of the algebra (Hom (C 0 , A) , ?) to the unity of the algebra (Hom (C, A0 ) , ?).
Furthermore, every f ∈ Hom (C 0 , A) and g ∈ Hom (C 0 , A) satisfy
ϕ(f ? g) = α ◦ (f ? g) ◦γ
| {z }
=mA ◦(f ⊗g)◦∆C 0

= α ◦ mA ◦(f ⊗ g) ◦ ∆C 0 ◦ γ
| {z } | {z }
=mA0 ◦(α⊗α) =(γ⊗γ)◦∆C

= mA0 ◦ (α ⊗ α) ◦ (f ⊗ g) ◦ (γ ⊗ γ) ◦∆C
| {z }
=(α◦f ◦γ)⊗(α◦g◦γ)

= mA0 ◦ ((α ◦ f ◦ γ) ⊗ (α ◦ g ◦ γ)) ◦ ∆C

(1.4.2) = (α ◦ f ◦ γ) ? (α ◦ g ◦ γ) = ϕ(f ) ? ϕ(g).
| {z } | {z }
=ϕ(f ) =ϕ(g)

Thus, ϕ is a k-algebra homomorphism (since ϕ is a k-linear map and sends the unity of the algebra
(Hom (C 0 , A) , ?) to the unity of the algebra (Hom (C, A0 ) , ?)). 

Exercise 1.4.4. Let C and D be two k-coalgebras, and let A and B be two k-algebras. Prove that:
 HOPF ALGEBRAS IN COMBINATORICS 13

(a) If f : C → A, f 0 : C → A, g : D → B and g 0 : D → B are four k-linear maps, then

(f ⊗ g) ? (f 0 ⊗ g 0 ) = (f ? f 0 ) ⊗ (g ? g 0 )
in the convolution algebra Hom (C ⊗ D, A ⊗ B).
(b) Let R be the k-linear map (Hom (C, A) , ?) ⊗ (Hom (D, B) , ?) → (Hom (C ⊗ D, A ⊗ B) , ?) which
sends every tensor f ⊗ g ∈ (Hom (C, A) , ?) ⊗ (Hom (D, B) , ?) to the map f ⊗ g : C ⊗ D → A ⊗ B.
(Notice that the tensor f ⊗ g and the map f ⊗ g are different things which happen to be written in
the same way.) Then, R is a k-algebra homomorphism.
Exercise 1.4.5. Let C and D be two k-coalgebras. Let A be a k-algebra. Let Φ be the canonical k-
module isomorphism Hom (C ⊗ D, A) → Hom (C, Hom (D, A)) (defined by ((Φ (f )) (c)) (d) = f (c ⊗ d) for
all f ∈ Hom (C ⊗ D, A), c ∈ C and d ∈ D). Prove that Φ is a k-algebra isomorphism
(Hom (C ⊗ D, A) , ?) → (Hom (C, (Hom (D, A) , ?)) , ?) .
Definition 1.4.6. A bialgebra A is called a Hopf algebra if there is an element S (called an antipode for
A) in End(A) which is a 2-sided inverse under ? for the identity map idA . In other words, this diagram
commutes:

(1.4.3) A< ⊗ A
S⊗idA
/ A⊗A
∆ m

"
A
 /k u /A
<
m
∆
"
A⊗A / A⊗A
idA ⊗S
P
Or equivalently, if ∆(a) = a1 ⊗ a2 , then
X X
(1.4.4) S(a1 )a2 = u((a)) = a1 S(a2 ).
(a) (a)

Example 1.4.7. For a group algebra kG, one can define an antipode k-linearly via S(tg ) = tg−1 . The top
pentagon in the above diagram commutes because
(S ? id)(tg ) = m((S ⊗ id)(tg ⊗ tg )) = S(tg )tg = tg−1 tg = te = (u ◦ )(tg ).
Note that when it exists, the antipode S is unique, as with all 2-sided inverses in associative algebras: if
S, S 0 are both 2-sided ?-inverses to idA then
S 0 = (u ◦ ) ? S 0 = (S ? idA ) ? S 0 = S ? (idA ?S 0 ) = S ? (u ◦ ) = S.
On the other hand, the next property is not quite as obvious, but is useful when one wants to check that
a certain map is the antipode in a particular Hopf algebra, by checking it on an algebra generating set.

Proposition 1.4.8. The antipode S in a Hopf algebra A is an algebra anti-endomorphism: S(1) = 1, and
S(ab) = S(b)S(a) for all a, b in A.

Proof. (see [189, Chap. 4]) Since ∆ is an algebra map, one has ∆(1) = 1 ⊗ 1, and therefore 1 = u(1) =
S(1) · 1 = S(1).
To show S(ab) = S(b)S(a), consider A ⊗ A as a coalgebra and A as an algebra. Then Hom(A ⊗ A, A) is an
associative algebra with a convolution product ~ (to be distinguished from the convolution ? on End(A)),
having two-sided identity element uA A⊗A . We will show below that these three elements of Hom(A ⊗ A, A)
f (a ⊗ b) = ab
g(a ⊗ b) = S(b)S(a)
h(a ⊗ b) = S(ab)
have the property that
(1.4.5) h~f = uA A⊗A = f ~g
14 DARIJ GRINBERG AND VICTOR REINER

which would then show the desired equality h = g via associativity:

h = h~(uA A⊗A ) = h~(f ~g) = (h~f )~g = (uA A⊗A )~g = g.
P P
So we evaluate the three elements in (1.4.5) on a ⊗ b, assuming ∆(a) = (a) a1 ⊗ a2 and ∆(b) = (b) b1 ⊗ b2 ,
P
and hence ∆(ab) = (a),(b) a1 b1 ⊗ a2 b2 . One has
(uA A⊗A )(a ⊗ b) = uA (A (a)A (b)) = uA (A (ab)).

X
(h~f )(a ⊗ b) = h(a1 ⊗ b1 )f (a2 ⊗ b2 )
(a),(b)
X
= S(a1 b1 )a2 b2
(a),(b)

= (S ? idA )(ab) = uA (A (ab)).

X
(f ~g)(a ⊗ b) = f (a1 ⊗ b1 )g(a2 ⊗ b2 )
(a),(b)
X
= a1 b1 S(b2 )S(a2 )
(a),(b)
X
= a1 · (idA ?S)(b) · S(a2 )
(a)
X
= uA (A (b)) a1 S(a2 ) = uA (A (b))uA (A (a)) = uA (A (ab)).
(a)


Remark 1.4.9. Recall from Remark 1.3.9 that the comultiplication on a bialgebra A allows one to define an
A-module structure on the tensor product M ⊗N of two A-modules M, N . Similarly, the anti-endomorphism
S in a Hopf algebra allows one to turn left A-modules into right A-modules, or vice-versa.16 E.g., left A-
modules M naturally have a right A-module structure on the dual k-module M ∗ := Hom(M, k), defined
via (f a)(m) := f (am) for f in M ∗ and a in A. The antipode S can be used to turn this back into a left
A-module M ∗ , via (af )(m) = f (S(a)m).
For groups G and left kG-modules (group representations) M , this is how one defines the contragredient
action of G on M ∗ , namely tg acts as (tg f )(m) = f (tg−1 m).
More generally, if A is a Hopf algebra and M and N are two left A-modules, then Hom (M, N ) (the Hom
here means Homk , not HomA ) canonically becomes a left A-module by setting
X
(af ) (m) = a1 f (S (a2 ) m) for all a ∈ A, f ∈ Hom (M, N ) and m ∈ M.
(a)
17
When A is the group algebra kG of a group G, this leads to


(tg f ) (m) = tg f tg−1 m for all g ∈ G, f ∈ Hom (M, N ) and m ∈ M.

16Be warned that these two transformations are not mutually inverse! Turning a left A-module into a right one and then
again into a left one using the antipode might lead to a non-isomorphic A-module, unless the antipode S satisfies S 2 = id.
17In more abstract terms, this A-module structure is given by the composition

A
∆ / A⊗A idA ⊗S
/ A ⊗ Aop / End (Hom (M, N )) ,
where the last arrow is the morphism
A ⊗ Aop −→ End (Hom (M, N )) ,
a ⊗ b 7−→ (f 7→ (M → N, m 7→ af (bm))) .
Here, Aop denotes the opposite algebra of A, which is the k-algebra differing from A only in the multiplication being twisted
(the product of a and b in Aop is defined to be the product of b and a in A). As k-modules, Aop = A, but we prefer to use Aop
instead of A here to ensure that all morphisms in the above composition are algebra maps.
 HOPF ALGEBRAS IN COMBINATORICS 15

This is precisely how one commonly makes Hom (M, N ) a representation of G for two representations M
and N .

Along the same lines, whenever A is a k-bialgebra, we are supposed to think of the counit A → k as giving
a way to make k into a trivial A-module. This A-module k behaves as one would expect: the canonical
isomorphisms k ⊗ M → M , M ⊗ k → M and (if A is a Hopf algebra) Hom (M, k) → M ∗ are A-module
isomorphisms for any A-module M .

Corollary 1.4.10. Commutativity of A implies the antipode is an involution: S 2 = idA .

Proof. One checks that S 2 = S ◦ S is a right ?-inverse to S, as follows:

X
(S ? S 2 )(a) = S(a1 )S 2 (a2 )
(a)
 
X
=S S(a2 )a1  (by Proposition 1.4.8)
(a)
 
X
=S a1 S(a2 ) (by commutativity of A)
(a)

= S (u((a)))
= u((a)) (since S(1) = 1 by Proposition 1.4.8) .
Since S itself is the ?-inverse to idA , this shows that S 2 = idA . 
Remark 1.4.11. We won’t need it, but it is easy to adapt the above proof to show that S 2 = idA also holds
for cocommutative Hopf algebras; see [139, Corollary 1.5.12] or [189, Chapter 4]. For a general Hopf algebra
which is not finite-dimensional over a field k, the antipode S may not even have finite order, even in the
connected graded setting. E.g., Aguiar and Sottile [7] show that the Malvenuto-Reutenauer Hopf algebra of
permutations has antipode of infinite order. In general, antipodes need not even be invertible [190].

Proposition 1.4.12. Let A and B be two Hopf algebras. Then, the k-bialgebra A ⊗ B (defined as in
Proposition 1.3.20) is a Hopf algebra. The antipode of this Hopf algebra A ⊗ B is the map SA ⊗ SB :
A ⊗ B → A ⊗ B, where SA and SB are the antipodes of the Hopf algebras A and B.

Exercise 1.4.13. Prove Proposition 1.4.12.

In our frequent setting of connected graded bialgebras, antipodes come for free.

Proposition 1.4.14. A connected graded bialgebra A has a unique antipode S, which is a graded map
S
A −→ A, endowing it with a Hopf structure.

Proof. Let us try to define a (k-linear) left ?-inverse S to idA on each homogeneous component An , via
induction on n.
In the base case n = 0, Proposition 1.4.8 and its proof show that one must define S(1) = 1 so S is the
identity on A0 = k.
In the inductive
P step, recall from Exercise 1.3.19(h) that a homogeneous element a of degree n > 0 has
∆(a) = a ⊗ 1 + a1 ⊗ a2 , with each P deg(a1 ) < n. Hence in order to have S ? idA
P= u, one must define
S(a) in such a way that S(a) · 1 + S(a1 )a2 = u(a) = 0 and hence S(a) := − S(a1 )a2 , where S(a1 )
have already been uniquely defined by induction. This does indeed define such a left ?-inverse S to idA , by
induction. It is also a graded map by induction.
The same argument shows how to define a right ?-inverse S 0 to idA . Then S = S 0 is a two-sided ?-inverse
to idA by the associativity of ?. 
Here is another consequence of the fact that S(1) = 1.

Proposition 1.4.15. In bialgebras, primitive elements x have (x) = 0, and in Hopf algebras, they have
S(x) = −x.
16 DARIJ GRINBERG AND VICTOR REINER

Proof. In a bialgebra, (1) = 1. Hence ∆(x) = 1 ⊗ x + x ⊗ 1 implies via (1.2.2) that 1 · (x) + (1)x = x, so
(x) = 0. It also implies via (1.4.3) that S(x)1 + S(1)x = u(x) = u(0) = 0, so S(x) = −x. 
Thus whenever A is a Hopf algebra generated as an algebra by its primitive elements, S is the unique
anti-endomorphism that negates all primitive elements.
Example 1.4.16. The tensor and symmetric algebras T (V ) and Sym(V ) are each generated by V , which
contains only primitive elements in either case. Hence one has in T (V ) that
(1.4.6) S(xi1 xi2 · · · xik ) = (−xik ) · · · (−xi2 )(−xi1 ) = (−1)k xik · · · xi2 xi1
for each word (i1 , . . . , ik ) in the alphabet I if V is a free k-module with basis {xi }i∈I . The same holds in
Sym(V ) for each multiset (i1 , . . . , ik ), recalling that the monomials are now commutative. In other words,
for a commutative polynomial f (x) in Sym(V ), the antipode S sends f to f (−x), negating all the variables.
The antipode for a connected graded Hopf algebra has an interesting formula due to Takeuchi [190],
reminiscent of P. Hall’s formula for the Möbius function of a poset18. For the sake of stating this, consider
(for every k ∈ N) the k-fold tensor power A⊗k = A ⊗ · · · ⊗ A (defined in Example 1.1.2) and define iterated
multiplication and comultiplication maps

A⊗k
m(k−1) /A

A
∆(k−1) / A⊗k
by induction over k, setting m(−1) = u, ∆(−1) = , m(0) = ∆(0) = idA , and
m(k) = m ◦ (idA ⊗m(k−1) ) for every k ≥ 1;
∆(k) = (idA ⊗∆(k−1) ) ◦ ∆ for every k ≥ 1.
Using associativity and coassociativity, one can see that for k ≥ 1 these maps also satisfy
m(k) = m ◦ (m(k−1) ⊗ idA ) for every k ≥ 1;
∆(k) = (∆(k−1) ⊗ idA ) ◦ ∆ for every k ≥ 1
(so we could just as well have used idA ⊗m(k−1) instead of m(k−1) ⊗ idA in defining them) and further
symmetry properties (see Exercise 1.4.17 and Exercise 1.4.18). They are how one gives meaning to the right
sides of these equations:
m(k) (a(1) ⊗ · · · ⊗ a(k+1) ) = a(1) · · · a(k+1)
X
∆(k) (b) = b1 ⊗ · · · ⊗ bk+1 in Sweedler notation.

Exercise 1.4.17. Let A be a k-algebra. Let us define, for every k ∈ N, a k-linear map m(k) : A⊗(k+1) → A.
Namely, we define these maps by induction
over k, with the induction base m(0) = idA , and with the
(k) (k−1)
induction step m = m ◦ idA ⊗m for every k ≥ 1. (This generalizes our definition of m(k) for Hopf
(−1)
algebras A given above, except for m which we have omitted.)
(k) (i) (k−1−i)


(a) Show that m = m ◦ m ⊗ m for every k ≥ 0 and 0 ≤ i ≤ k − 1.


(b) Show that m(k) = m ◦ m(k−1) ⊗ idA for every k ≥ 1.
(c) Show that m(k) = m(k−1) ◦ (idA⊗i ⊗m ⊗ idA⊗(k−1−i) ) for every k ≥ 0 and 0 ≤ i ≤ k − 1.
(d) Show that m(k) = m(k−1) ◦ (idA⊗(k−1) ⊗m) = m(k−1) ◦ (m ⊗ idA⊗(k−1) ) for every k ≥ 1.
Exercise 1.4.18. Let C be a k-coalgebra. Let us define, for every k ∈ N, a k-linear map ∆(k) : C → C ⊗(k+1) .
(0)

by induction over k, with the induction base ∆ = idC ,(k)
Namely, we define these maps and with the induction
(k) (k−1)
step ∆ = idC ⊗∆ ◦ ∆ for every k ≥ 1. (This generalizes our definition of ∆ for Hopf algebras A
given above, except for ∆(−1) which we have omitted.)


(a) Show that ∆(k) = ∆(i) ⊗ ∆(k−1−i) ◦ ∆ for every k ≥ 0 and 0 ≤ i ≤ k − 1.


(b) Show that ∆(k) = ∆(k−1) ⊗ idC ◦ ∆ for every k ≥ 1.
(c) Show that ∆(k) = (idC ⊗i ⊗∆ ⊗ idC ⊗(k−1−i) ) ◦ ∆(k−1) for every k ≥ 0 and 0 ≤ i ≤ k − 1.
(d) Show that ∆(k) = (idC ⊗(k−1) ⊗∆) ◦ ∆(k−1) = (∆ ⊗ idC ⊗(k−1) ) ◦ ∆(k−1) for every k ≥ 1.
18In fact, for incidence Hopf algebras, Takeuchi’s formula generalizes Hall’s formula– see Corollary 7.2.3.
 HOPF ALGEBRAS IN COMBINATORICS 17

Remark 1.4.19. Exercise 1.4.17 holds more generally for nonunital associative algebras A (that is, k-modules
A equipped with a k-linear map m : A ⊗ A → A such that the diagram (1.1.1) is commutative, but not
necessarily admitting a unit map u). Similarly, Exercise 1.4.18 holds for non-counital coassociative coalgebras
C. The existence of a unit in A, respectively a counit in C, allows slightly extending these two exercises by
additionally introducing maps m(−1) = u : k → A and ∆(−1) =  : C → k; however, not much is gained from
this extension.19
(k)
Exercise 1.4.20. For every k ∈ N and every k-bialgebra H, consider the map ∆H : H → H ⊗(k+1) (this is
(k)
the map ∆(k) defined as in Exercise 1.4.18 for C = H), and the map mH : H ⊗(k+1) → H (this is the map
(k)
m defined as in Exercise 1.4.17 for A = H).
Let H be a k-bialgebra. Let k ∈ N. Show that:20
(k)
(a) The map mH : H ⊗(k+1) → H is a k-coalgebra homomorphism.
(k)
(b) The map ∆H : H → H ⊗(k+1) is a k-algebra homomorphism.
 ⊗(`+1)
(`) (k) (k) (`)
(c) We have mH ⊗(k+1) ◦ ∆H = ∆H ◦ mH for every ` ∈ N.
 ⊗(k+1)
(`) (k) (k) (`)
(d) We have mH ◦ ∆H ⊗(`+1) = ∆H ◦ mH for every ` ∈ N.

The iterated multiplication and comultiplication maps allow explicitly computing the convolution of mul-
tiple maps; the following formula will often be used without explicit mention:
Exercise 1.4.21. Let C be a k-coalgebra, and A be a k-algebra. Let k ∈ N. Let f1 , f2 , . . . , fk be k elements
of Hom (C, A). Show that
(k−1) (k−1)
f1 ? f 2 ? · · · ? f k = m A ◦ (f1 ⊗ f2 ⊗ · · · ⊗ fk ) ◦ ∆C .
We are now ready to state Takeuchi’s formula for the antipode:

Proposition 1.4.22. In a connected graded Hopf algebra A, the antipode has formula
X
S= (−1)k m(k−1) f ⊗k ∆(k−1)
(1.4.7) k≥0

= u − f + m ◦ f ⊗2 ◦ ∆ − m(2) ◦ f ⊗3 ◦ ∆(2) + · · ·
where f := idA −u in End(A).

Proof. We argue as in [190, proof of Lemma 14] or [7, §5]. For any f in End(A) one has this explicit formula
for its k-fold convolution power f ?k := f ? · · · ? f in terms of its tensor powers f ⊗k := f ⊗ · · · ⊗ f (according
to Exercise 1.4.21):
f ?k = m(k−1) ◦ f ⊗k ◦ ∆(k−1) .
Therefore any f annihilating A0 will be locally ?-nilpotent on A, meaning that for each n one has that An
is annihilated by f ?m for every m > n: homogeneity forces that for a in An , every summand of ∆(m−1) (a)
must contain among its m tensor factors at least one factor lying in A0 , so each summand is annihilated by
f ⊗m , and f ?m (a) = 0.
In particular such f have the property that u + f has as two-sided ?-inverse
(u + f )?(−1) = u − f + f ? f − f ? f ? f + · · ·
X X
= (−1)k f ?k = (−1)k m(k−1) ◦ f ⊗k ◦ ∆(k−1) .
k≥0 k≥0

The proposition follows upon taking f := idA −u, which annihilates A0 . 

Remark 1.4.23. In fact, one can see that Takeuchi’s formula applies more generally to define an antipode
S
A −→ A in any (not necessarily graded) bialgebra A where the map idA −u is locally ?-nilpotent.
It is also worth noting that the proof of Proposition 1.4.22 gives an alternate proof of Proposition 1.4.14.
19The identity m(k) = m ◦ id ⊗m(k−1)
for a k-algebra A still holds when k = 0 if it is interpreted in the right way (viz.,
A
if A is identified with A ⊗ k using the canonical homomorphism).
20The following statements are taken from [147]; specifically, part (c) is [147, Lem. 1.8].
18 DARIJ GRINBERG AND VICTOR REINER

To finish our discussion of antipodes, we mention some properties (taken from [189, Chap. 4]) relat-
ing antipodes to convolutional inverses. It also shows that a bialgebra morphism between Hopf algebras
automatically respects the antipodes.

Proposition 1.4.24. Let H be a Hopf algebra with antipode S.

α
(a) For any algebra A and algebra morphism H → A, one has α ◦ S = α?−1 , the convolutional inverse
to α in Hom(H, A).
γ
(b) For any coalgebra C and coalgebra morphism C → H, one has S ◦ γ = γ ?−1 , the convolutional
inverse to γ in Hom(C, H).
β
(c) If H1 , H2 are Hopf algebras with antipodes S1 , S2 , then any bialgebra morphism H1 → H2 is a Hopf
(a) (b)
morphism, that is, it commutes with the antipodes, since β ◦ S1 = β ?−1 = S2 ◦ β.

Proof. We prove (a); the proof of (b) is similar, and (c) follows immediately from (a),(b) as indicated in its
statement.
For assertion (a), note that Proposition 1.4.3 (applied to H, H, H, A, idH and α instead of C, C 0 , A, A0 ,
γ and α) shows that the map
Hom (H, H) → Hom (H, A) , f 7→ α ◦ f
is a k-algebra homomorphism from the convolution algebra (Hom (H, H) , ?) to
the convolution algebra
?−1
(Hom (H, A) , ?). Denoting this homomorphism by ϕ, we thus have ϕ (idH )?−1 = (ϕ(idH )) (since
k-algebra homomorphisms preserve inverses). Now,
?−1 ?−1
α ◦ S = ϕ(S) = ϕ (idH )?−1 = (ϕ(idH )) = α?−1 .


= (α ◦ idH )


Exercise 1.4.25. Prove that the antipode S of a Hopf algebra A is a coalgebra anti-endomorphism, i.e.,
that it satisfies  ◦ S =  and ∆ ◦ S = T ◦ (S ⊗ S) ◦ ∆, where T : A ⊗ A → A ⊗ A is the twist map sending
every a ⊗ b to b ⊗ a.
Exercise 1.4.26. If C is a k-coalgebra and if A is a k-algebra, then a k-linear map f : C → A is said to be
?-invertible if it is invertible as an element of the k-algebra (Hom (C, A) , ?). In this case, the multiplicative
inverse f ?(−1) of f in (Hom (C, A) , ?) is called the ?-inverse of f .
For any two k-modules U and V , let TU,V : U ⊗ V → V ⊗ U be the twist map (i.e., the k-linear map
U ⊗ V → V ⊗ U sending every u ⊗ v to v ⊗ u).
(a) If C is a k-bialgebra, if A is a k-algebra, and if r : C → A is a ?-invertible k-algebra homomorphism,
then prove that the ?-inverse r?(−1) of r is a k-algebra anti-homomorphism21.
(b) If C is a k-bialgebra, if A is a k-coalgebra, and if r : A → C is a ?-invertible k-coalgebra homomor-
phism, then prove that the ?-inverse r?(−1) of r is a k-coalgebra anti-homomorphism22.
(c) Derive Proposition 1.4.8 from Exercise 1.4.26(a), and derive Exercise 1.4.25 from Exercise 1.4.26(b).
(d) Prove Corollary 1.4.10 again using Proposition 1.4.24.
(e) Prove that the antipode S of a cocommutative Hopf algebra A satisfies S 2 = idA . (This was a
statement made in Remark 1.4.11.)
Exercise 1.4.27. (a) Let A be a Hopf algebra. If P : A → A is a k-linear map such that every a ∈ A
satisfies
X
P (a2 ) · a1 = u ( (a)) ,
(a)

then prove that the antipode S of A is invertible and its inverse is P .

21A k-algebra anti-homomorphism means a k-linear map ϕ : P → Q between two k-algebras P and Q which satisfies
ϕ ◦ mP = mQ ◦ (ϕ ⊗ ϕ) ◦ TP,P and ϕ ◦ uP = uQ .
22A k-coalgebra anti-homomorphism means a k-linear map ϕ : P → Q between two k-coalgebras P and Q which satisfies
∆Q ◦ ϕ = TQ,Q ◦ (ϕ ⊗ ϕ) ◦ ∆P and Q ◦ ϕ = P .
 HOPF ALGEBRAS IN COMBINATORICS 19

(b) Let A be a Hopf algebra. If P : A → A is a k-linear map such that every a ∈ A satisfies
X
a2 · P (a1 ) = u ( (a)) ,
(a)

then prove that the antipode S of A is invertible and its inverse is P .

(c) Show that the antipode of a connected graded Hopf algebra is invertible.
(Compare this exercise to [139, Lemma 1.5.11].)
Definition 1.4.28. Let C be a k-coalgebra. A subcoalgebra of C means a k-coalgebra D such that D ⊂ C
and such that the canonical inclusion map D → C is a k-coalgebra homomorphism23. When k is a field, we
can equivalently define a subcoalgebra of C as a k-submodule D of C such that ∆C (D) is a subset of the
k-submodule D ⊗ D of C ⊗ C; however, this might no longer be equivalent when k is not a field24.
Similarly, a subbialgebra of a bialgebra C is a k-bialgebra D such that D ⊂ C and such that the canonical
inclusion map D → C is a k-bialgebra homomorphism. Also, a Hopf subalgebra of a Hopf algebra C is a
k-Hopf algebra D such that D ⊂ C and such that the canonical inclusion map D → C is a k-Hopf algebra
homomorphism.25
Exercise 1.4.29. Let C be a k-coalgebra. Let D be a k-submodule of C such that D is a direct summand
of C as a k-module (i.e., there exists a k-submodule E of C such that C = D ⊕ E). (This is automatically
satisfied if k is a field.) Assume that ∆ (D) ⊂ C ⊗ D and ∆ (D) ⊂ D ⊗ C. (Here, we are abusing the notation
C ⊗ D to denote the k-submodule of C ⊗ C spanned by tensors of the form c ⊗ d with c ∈ C and d ∈ D;
similarly, D ⊗ C should be understood.) Show that there is a canonically defined k-coalgebra structure on
D which makes D a subcoalgebra of C.
The next exercise is implicit in [4, §5]:
Exercise 1.4.30. Let k be a field. Let C be a k-coalgebra, and let U be any k-module. Let f : C → U be
a k-linear map. Recall the map ∆(2) : C → C ⊗3 from Exercise 1.4.18. Let K = ker (idC ⊗f ⊗ idC ) ◦ ∆(2) .

(a) Show that K is a k-subcoalgebra of C.

(b) Show that every k-subcoalgebra of C which is a subset of ker f must be a subset of K.
L
Exercise 1.4.31. (a) Let C = n≥0 Cn be a graded k-coalgebra, and A be any k-algebra. Notice that
C0 itself is a k-subcoalgebra of C. Let h : C → A be a k-linear map such that the restriction h |C0
is a ?-invertible map in Hom (C0 , A). Prove that h is a ?-invertible map in Hom (C, A). (This is a
weaker version
L of Takeuchi’s [190, Lemma 14].)
(b) Let A = n≥0 An be a graded k-bialgebra. Notice that A0 is a subbialgebra of A. Assume that A0
is a Hopf algebra. Show that A is a Hopf algebra.
(c) Obtain yet another proof of Proposition 1.4.14.
L
Exercise 1.4.32. Let A = n≥0 An be a connected graded k-bialgebra. Let p be the k-submodule of A
consisting of the primitive elements of A.
L
(a) If I is a two-sided coideal of A such that I ∩ p = 0 and such that I = n≥0 (I ∩ An ), then prove
that I = 0.
(b) Let f : A → C be a graded surjective coalgebra homomorphism from A to a graded k-coalgebra C.
If f |p is injective, then prove that f is injective.
(c) Assume that k is a field. Show that the claim of Exercise 1.4.32(b) is valid even without requiring
f to be surjective.
Remark 1.4.33. Exercise 1.4.32 (b) and (c) are often used in order to prove that certain coalgebra homo-
morphisms are injective.
The word “bialgebra” can be replaced by “coalgebra” in Exercise 1.4.32, provided that the notion of a
connected graded coalgebra is defined correctly (namely, as a graded coalgebra such that the restriction of
23In this definition, we follow [143, p. 55] and [201, §6.7]; other authors may use other definitions.
24This is because the k-submodule D ⊗ D of C ⊗ C is generally not isomorphic to the k-module D ⊗ D. See [143, p. 56] for
specific counterexamples for the non-equivalence of the two notions of a subcoalgebra. Notice that the equivalence is salvaged
if D is a direct summand of C as a k-module (see Exercise 1.4.29 for this).
25By Proposition 1.4.24(c), we can also define it as a subbialgebra of C which happens to be a Hopf algebra.
20 DARIJ GRINBERG AND VICTOR REINER

 to the 0-th graded component is an isomorphism), and the notion of the element 1 of a connected graded
coalgebra is defined accordingly (namely, as the preimage of 1 ∈ k under the restriction of  to the 0-th
graded component).
1.5. Commutativity, cocommutativity.
Definition 1.5.1. Say that the k-algebra A is commutative if ab = ba, that is, this diagram commutes:

(1.5.1) A⊗A
T / A⊗A
m m

" |
A
This, of course, is a mere restatement of the classical definition of a commutative k-algebra using tensors
instead of pairs of elements.
Say that the k-coalgebra C is cocommutative if this diagram commutes:

(1.5.2) C ⊗C
T / C ⊗C
c ;

∆ ∆
C
Example 1.5.2. Group algebras kG are always cocommutative, but commutative if and only if G is abelian.
Tensor algebras T (V ) are always cocommutative, but not generally commutative26.
Symmetric algebras Sym(V ) are always cocommutative and commutative.
Homology and cohomology of H-spaces are always cocommutative and commutative in the topologist’s
T
sense where one reinterprets that twist map A ⊗ A → A ⊗ A to have the extra sign as in (1.3.3).
Note how the cocommutative Hopf algebras T (V ), Sym(V ) have much of their structure controlled by their
k-submodules V , which consist of primitive elements only (although, in general, not of all their primitive
elements). This is not far from the truth in general, and closely related to Lie algebras.
Exercise 1.5.3. Recall that a Lie algebra over k is a k-module g with a k-bilinear map [·, ·] : g × g → g
that satisfies [x, x] = 0 for x in g, and the Jacobi identity
[x, [y, z]] = [[x, y], z] + [y, [x, z]], or equivalently
[x, [y, z]] + [z, [x, y]] + [y, [z, x]] = 0.
This k-bilinear map [·, ·] is called the Lie bracket of g.
(a) Check that any associative algebra A gives rise to a Lie algebra by means of the commutator operation
[a, b] := ab − ba.
(b) If A is also a bialgebra, show that the k-submodule of primitive elements p ⊂ A is closed under the
Lie bracket, that is, [p, p] ⊂ p, and hence forms a Lie subalgebra.
Conversely, given a Lie algebra p, one constructs the universal enveloping algebra U(p) := T (p)/J as the
quotient of the tensor algebra T (p) by the two-sided ideal J generated by all elements xy − yx − [x, y] for
x, y in p.
(c) Show that J is also a two-sided coideal in T (p) for its usual coalgebra structure, and hence the
quotient U(p) inherits the structure of a cocommutative bialgebra.
(d) Show that the antipode S on T (p) preserves J, meaning that S(J) ⊂ J, and hence U(p) inherits the
structure of a (cocommutative) Hopf algebra.
There are theorems, discussed in [33, §3.8], [139, Chap. 5], [51, §3.2] giving various mild hypotheses in
addition to cocommutativity which imply that the inclusion of the k-module p of primitives in a Hopf algebra
A extends to a Hopf isomorphism U(p) ∼= A.
Exercise 1.5.4. Let C be a cocommutative k-coalgebra. Let A be a commutative k-algebra. Show that
the convolution algebra (Hom (C, A) , ?) is commutative (i.e., every f, g ∈ Hom (C, A) satisfy f ? g = g ? f ).
26If k is a field, then T (V ) is commutative if and only if dim V ≤ 1.
k
 HOPF ALGEBRAS IN COMBINATORICS 21

Exercise 1.5.5. (a) Let C be a k-coalgebra. Show that C is cocommutative if and only if its comulti-
plication ∆C : C → C ⊗ C is a k-coalgebra homomorphism.
(b) Let A be a k-algebra. Show that A is commutative if and only if its multiplication mA : A ⊗ A → A
is a k-algebra homomorphism.
Remark 1.5.6. If C is a k-coalgebra, then C : C → k is always a k-coalgebra homomorphism. Similarly,
uA : k → A is a k-algebra homomorphism whenever A is a k-algebra.
Exercise 1.5.7. Let A be a commutative k-algebra, and let k ∈ N. The symmetric group Sk acts on the k-
fold tensor power A⊗k by permuting the tensor factors: σ (v1 ⊗ v2 ⊗ · · · ⊗ vk ) = vσ−1 (1) ⊗vσ−1 (2) ⊗· · ·⊗vσ−1 (k)
for all v1 , v2 , . . . , vk ∈ A and σ ∈ Sk . For every π ∈ Sk , denote by ρ (π) the action of π on A⊗k (this is
an endomorphism of A⊗k ). Show that every π ∈ Sk satisfies m(k−1) ◦ (ρ (π)) = m(k−1) . (Recall that
m(k−1) : A⊗k → A is defined as in Exercise 1.4.17 for k ≥ 1, and by m(−1) = u : k → A for k = 0.)
Exercise 1.5.8. State and solve the analogue of Exercise 1.5.7 for cocommutative k-coalgebras.
Exercise 1.5.9. (a) If H is a k-bialgebra and A is a commutative k-algebra, and if f and g are two
k-algebra homomorphisms H → A, then prove that f ? g also is a k-algebra homomorphism H → A.
(b) If H is a k-bialgebra and A is a commutative k-algebra, and if f1 , f2 , . . . , fk are several k-algebra
homomorphisms H → A, then prove that f1 ? f2 ? · · · ? fk also is a k-algebra homomorphism H → A.
(c) If H is a Hopf algebra and A is a commutative k-algebra, and if f : H → A is a k-algebra ho-
momorphism, then prove that f ◦ S : H → A (where S is the antipode of H) is again a k-algebra
homomorphism, and is a ?-inverse to f .
(d) If A is a commutative k-algebra, then show that m(k) is a k-algebra homomorphism for every k ∈ N.
(The map m(k) : A⊗(k+1) → A is defined as in Exercise 1.4.17.)
(e) If C 0 and C are two k-coalgebras, if γ : C → C 0 is a k-coalgebra homomorphism, if A and A0 are
two k-algebras, if α : A → A0 is a k-algebra homomorphism, and if f1 , f2 , . . . , fk are several k-linear
maps C 0 → A, then prove that
α ◦ (f1 ? f2 ? · · · ? fk ) ◦ γ = (α ◦ f1 ◦ γ) ? (α ◦ f2 ◦ γ) ? · · · ? (α ◦ fk ◦ γ) .
(f) If H is a commutative k-bialgebra, and k and ` are two nonnegative integers, then prove that
?(k`)
id?k ?`
H ◦ idH = idH .
(g) If H is a commutative k-Hopf algebra, and k and ` are two integers, then prove that id?k ?`
H ◦ idH =
?(k`) ?(k`)
idH . (These powers id?k ?`
H , idH and idH are well-defined since idH is ?-invertible.)
(h) State and prove the duals of parts (a)–(g) of this exercise.

Remark 1.5.10. The maps id?k H for k ∈ N are known as the Adams operators of the bialgebra H; they are
studied, inter alia, in [5]. Particular cases (and variants) of Exercise 1.5.9(f) appear in [147, Corollaire II.9]
and [65, Theorem 1]. Exercise 1.5.9(f) and its dual are [118, Prop. 1.6].
Exercise 1.5.11. Let A be a cocommutative graded Hopf algebra with antipode S. Define a k-linear map
E : A → A by having E (a) = (deg a) · a for every homogeneous element a of A.
(a) Prove that for every a ∈ A, the elements (S ? E) (a) and (E ? S) (a) (where ? denotes convolution in
Hom (A, A)) are primitive.
(b) Prove that for every primitive p ∈ A, we have (S ? E) (p) = (E ? S) (p) = E (p).
(c) Prove that for every a ∈ A and every primitive p ∈ A, we have (S ? E) (ap) = [(S ? E) (a) , p] +
 (a) E (p), where [u, v] denotes the commutator uv − vu of u and v.
(d) If A is connected and Q is a subring of k, prove that the k-algebra A is generated by the k-submodule
p consisting of the primitive elements of A.
(e) Assume that A is the tensor algebra T (V ) of a k-module V , and that the k-submodule V =
V ⊗1 of T (V ) is the degree-1 homogeneous component of A. Show that (S ? E) (x1 x2 . . . xn ) =
[. . . [[x1 , x2 ] , x3 ] , . . . , xn ] for any n ≥ 1 and any x1 , x2 , . . . , xn ∈ V .
Remark 1.5.12. Exercise 1.5.11 gives rise to a certain idempotent map A → A when k is a commutative
Q-algebra and A is a cocommutative connected graded k-Hopf algebra. Namely, the k-linear map A → A
1
sending every homogeneous a ∈ A to deg a (S ? E) (a) (or 0 if deg a = 0) is idempotent and is a projection
22 DARIJ GRINBERG AND VICTOR REINER

on the k-module of primitive elements of A. It is called the Dynkin idempotent; see [148] for more of its
properties.27 Part (c) of the exercise is more or less Baker’s identity.
1.6. Duals. Recall that for k-modules V , taking the dual k-module V ∗ := Hom(V, k) reverses k-linear
ϕ ϕ∗
maps. That is, every k-linear map V → W induces an adjoint map W ∗ → V ∗ defined uniquely by
(f, ϕ(v)) = (ϕ∗ (f ), v)
in which (f, v) is the bilinear pairing V ∗ × V → k sending (f, v) 7→ f (v). If V and W are finite free k-
modules28, more can be said: When ϕ is expressed in terms of a basis {vi }i∈I for V and a basis {wj }j∈J for
W by some matrix, the map ϕ∗ is expressed by the transpose matrix in terms of the dual bases of these two
bases29.
ϕ ϕ∗
The correspondence ϕ 7→ ϕ∗ between k-linear maps V → W and k-linear maps W ∗ → V ∗ is one-to-one
when W is finite free. However, this is not the case in many combinatorial situations (in which W is usually
free but not finite free). Fortunately, many of the good properties of finite free modules carry over to a
certain class of graded modules as long as the dual V ∗ is replaced by a smaller module V o called the graded
dual. Let us first
L introduce the latter:
When V = n≥0 Vn is a graded k-module, note that the dual V ∗ = n≥0 (Vn )∗ can contain functionals
Q
∗
f supported on infinitely many Vn . However, we can consider the k-submodule V o :=
L
n≥0 (Vn ) ⊂
∗ ∗ 30
Q
n≥0 (Vn ) = V , sometimes called the graded dual , consisting of the functions f that vanish on all but
ϕ
finitely many Vn . Notice that V o is graded, whereas V ∗ (in general) is not. If V → W is a graded k-linear
∗
ϕ
map, then the adjoint map W ∗ → V ∗ restricts to a graded k-linear map W o → V o , which we (abusively)
still denote by ϕ∗ .
When the graded k-module V = n≥0 Vn is of finite type, meaning that each Vn is a finite free k-module31,
L

the graded k-module V o is again of finite type32 and satisfies (V o ) ∼

o
= V . Many other properties of finite
free modules are salvaged in this situation; most importantly: The correspondence ϕ 7→ ϕ∗ between graded
k-linear maps V → W and graded k-linear maps W o → V o is one-to-one when W is of finite type33.
Reversing the diagrams should then make it clear that, in the finite free or finite-type situation, duals
of algebras are coalgebras, and vice-versa, and duals of bialgebras or Hopf algebras are bialgebras or Hopf
algebras. For example, the product in a Hopf algebra A of finite type uniquely defines the coproduct of Ao
via adjointness:
(∆Ao (f ), a ⊗ b)A⊗A = (f, ab)A .
Thus if A has a basis {ai }i∈I with product structure constants {cij,k }, meaning
X
aj ak = cij,k ai ,
i∈I

then the dual basis {fi }i∈I has the same {cij,k }
as its coproduct structure constants:
X
∆Ao (fi ) = cij,k fj ⊗ fk .
(j,k)∈I×I

27We will see another such idempotent in Exercise 5.3.6.

28A k-module is said to be finite free if it has a finite basis. If k is a field, then a finite free k-module is the same as a
finite-dimensional k-vector space.
29If {v } ∗
i i∈I is a basis of a finite free k-module V , then the dual basis of this basis is defined as the basis {fi }i∈I of V that
satisfies (fi , vj ) = δi,j for all i and j. (Recall that δi,j is the Kronecker delta: δi,j = 1 if i = j and 0 else.)
30Do not mistake this for the coalgebraic restricted dual A◦ of [189, §6.0].
31This meaning of “finite type” can differ from the standard one.
32More precisely, let V = L
n≥0 Vn be of finite type, and let {vi }i∈I be a graded basis of V , that is, a basis of the k-module
V such that the indexing set I is partitioned into subsets I0 , I1 , I2 , . . . (which are allowed to be empty) with the property that,
for every n ∈ N, the subfamily {vi }i∈In is a basis of the k-module Vn . Then, we can define a family {fi }i∈I of elements of V o
by setting (fi , vj ) = δi,j for all i, j ∈ I. This family {fi }i∈I is a graded basis of the graded k-module V o . (Actually, for every
n ∈ N, the subfamily {fi }i∈In is a basis of the k-submodule (Vn )∗ of V o – indeed the dual basis to the basis {vi }i∈In of Vn .)
This basis {fi }i∈I is said to be the dual basis to the basis {vi }i∈I of V .
33Only W has to be of finite type here; V can be any graded k-module.
 HOPF ALGEBRAS IN COMBINATORICS 23

The assumption that A be of finite type was indispensable here; in general, the dual of a k-algebra does
not become a k-coalgebra. However, the dual of a k-coalgebra still becomes a k-algebra, as shown in the
following exercise:
∗
Exercise 1.6.1. For any two k-modules U and V , let ρU,V : U ∗ ⊗ V ∗ → (U ⊗ V ) be the k-linear map
f ⊗g m
which sends every tensor f ⊗ g ∈ U ∗ ⊗ V ∗ to the composition U ⊗ V −→ k ⊗ k −→ k
k of the map34 f ⊗ g
mk
with the canonical isomorphism k ⊗ k −→ k. When k is a field and U is finite-dimensional, this map ρU,V is
a k-vector space isomorphism (and usually regarded as the identity); more generally, it is injective whenever
k is a field35. Also, let s : k → k∗ be the canonical isomorphism. Prove that:
(a) If C is a k-coalgebra, then C ∗ becomes a k-algebra if we define its associative operation by mC ∗ =
∆∗C ◦ ρC,C : C ∗ ⊗ C ∗ → C ∗ and its unit map to be ∗C ◦ s : k → C ∗ . 36

(b) The k-algebra structure defined on C in part (a) is precisely the one defined on Hom (C, k) = C ∗
∗

in Definition 1.4.1 applied to A = k.

(c) If C is a graded k-coalgebra, then C o is a k-subalgebra of the k-algebra C ∗ defined in part (a).
(d) If f : C → D is a homomorphism of k-coalgebras, then f ∗ : D∗ → C ∗ is a homomorphism of
k-algebras.
(e) Let U be a graded k-module (not necessarily of finite type), and let V be a graded k-module of
finite type. Then, there is a 1-to-1 correspondence between graded k-linear maps U → V and graded
k-linear maps V o → U o given by f 7→ f ∗ .
(f) Let C be a graded k-coalgebra (not necessarily of finite type), and let D be a graded k-coalgebra
of finite type. Part (e) of this exercise shows that there is a 1-to-1 correspondence between graded
k-linear maps C → D and graded k-linear maps Do → C o given by f 7→ f ∗ . This correspondence
has the property that a given graded k-linear map f : C → D is a k-coalgebra map if and only if
f ∗ : Do → C o is a k-algebra map.
Another example of a Hopf algebra is provided by the so-called shuffle algebra. Before we introduce it,
let us define the shuffles of two words:
Definition 1.6.2. Given two words a = (a1 , a2 , . . . , an ) and b = (b1 , b2 , . . . , bm ), the multiset of shuffles of
a and b is defined as the multiset


cw(1) , cw(2) , . . . , cw(n+m) : w ∈ Shn,m multiset ,
where (c1 , c2 , . . . , cn+m ) is the concatenation a · b = (a1 , a2 , . . . , an , b1 , b2 , . . . , bm ), and where Shn,m is the
subset37
w ∈ Sn+m : w−1 (1) < w−1 (2) < · · · < w−1 (n) ; w−1 (n + 1) < w−1 (n + 2) < · · · < w−1 (n + m)


of the symmetric group Sn+m . Informally speaking, the shuffles of the two words a and b are the words
obtained by overlaying the words a and b, after first moving their letters apart so that no letters get
superimposed when the words are overlayed38. In particular,
any shuffle of a and b contains a and b as
subsequences. The multiset of shuffles of a and b has m+n elements (counted with multiplicity) and is
denoted by a
n
b. For instance, the shuffles of (1, 2, 1) and (3, 2) are
(1, 2, 1, 3, 2) , (1, 2, 3, 1, 2) , (1, 2, 3, 2, 1) , (1, 3, 2, 1, 2) , (1, 3, 2, 2, 1) ,
(1, 3, 2, 2, 1) , (3, 1, 2, 1, 2) , (3, 1, 2, 2, 1) , (3, 1, 2, 2, 1) , (3, 2, 1, 2, 1) ,

⊗g
34Keep in mind that the tensor f ⊗ g ∈ U ∗ ⊗ V ∗ is not the same as the map U ⊗ V f−→ k ⊗ k.
35Over arbitrary rings it does not have to be even that!
36If C is a finite free k-module, then this k-algebra structure is the same as the one defined above by adjointness. But the
advantage of the new definition is that it works even if C is not a finite free k-module.
37Warning: This definition of Sh
n,m is highly nonstandard, and many authors define Shn,m to be the set of the inverses
of the permutations belonging to what we call Shn,m .
38For instance, if a = (1, 3, 2, 1) and b = (2, 4), then the shuffle (1, 2, 3, 2, 4, 1) of a and b can be obtained by moving the
letters of a and b apart as follows:
a= 1 3 2 1
b= 2 4
and then overlaying them to obtain 1 2 3 2 4 1 . Other ways of moving letters apart lead to further shuffles (not
always distinct).
24 DARIJ GRINBERG AND VICTOR REINER

listed here as often as they appear in the multiset (1, 2, 1)

(3, 2). Here we have underlined the letters
taken from a – that is, the letters at positions w−1 (1), w−1 (2), . . ., w−1 (n).
Example 1.6.3. When A = T (V ) is the tensor algebrafor a finite free k-module V , having k-basis {xi }i∈I ,
its graded dual Ao is another Hopf algebra whose basis y(i1 ,...,i` ) (the dual basis of the basis {xi1 · · · xi` } of
A = T (V )) is indexed by words in the alphabet I. This Hopf algebra Ao could be called the shuffle algebra
of V ∗ . (To be more precise, it is isomorphic to the shuffle algebra of V ∗ introduced in Proposition 1.6.7
further below; we prefer not to call Ao itself the shuffle algebra of V ∗ , since Ao has several disadvantages39.)
Duality shows that the cut coproduct in Ao is defined by
`
X
(1.6.1) ∆y(i1 ,...,i` ) = y(i1 ,...,ij ) ⊗ y(ij+1 ,ij+2 ,...,i` ) .
j=0

For example,
∆yabcb = y∅ ⊗ yabcb + ya ⊗ ybcb + yab ⊗ ycb + yabc ⊗ yb + yabcb ⊗ y∅
Duality also shows that the shuffle product in Ao will be given by
X
(1.6.2) y(i1 ,...,i` ) y(j1 ,...,jm ) = y(k1 ,...,k`+m )
k=(k1 ,...,k`+m )∈i
j
where i

j (as in Definition 1.6.2) denotes the multiset of the `+m `


words obtained as shuffles of the two
words i = (i1 , . . . , i` ) and j = (j1 , . . . , jm ). For example,
yab ycb = yabcb + yacbb + ycabb + ycabb + yacbb + ycbab
= yabcb + 2yacbb + 2ycabb + ycbab
Equivalently, one has
X
(1.6.3) y(i1 ,i2 ,...,i` ) y(i`+1 ,i`+2 ,...,i`+m ) = y(i )
w−1 (1) ,iw−1 (2) ,...,iw−1 (`+m)
w∈S`+m :
w(1)<···<w(`),
w(`+1)<···<w(`+m)
X
(1.6.4) = y(iσ(1) ,iσ(2) ,...,iσ(`+m) )
σ∈Sh`,m

(using the notations of Definition 1.6.2 again). Lastly, the antipode S of Ao is the adjoint of the antipode of
A = T (V ) described in (1.4.6):
Sy(i1 ,i2 ,...,i` ) = (−1)` y(i` ,...,i2 ,i1 ) .
o
Since the coalgebra T (V ) is cocommutative, its graded dual T (V ) is commutative.
Exercise 1.6.4. Let V be a 1-dimensional free k-module with basis element x, so Sym(V ) ∼
= k[x], with
k-basis {1 = x0 , x1 , x2 , . . .}.
(a) Check that the powers xi satisfy
xi · xj = xi+j
X n
∆(xn ) = xi ⊗ xj
i+j=n
i
S(xn ) = (−1)n xn
(b) Check that the dual basis elements {f (0) , f (1) , f (2) , . . .} for Sym(V )o , defined by f (i) (xj ) = δi,j ,
satisfy  
i + j (i+j)
f (i) f (j) = f
i
X
∆(f (n) ) = f (i) ⊗ f (j)
i+j=n
(n)
S(f ) = (−1)n f (n)
39Specifically, Ao has the disadvantages of being defined only when V ∗ is the dual of a finite free k-module V , and depending
on a choice of basis, whereas Proposition 1.6.7 will define shuffle algebras in full generality and canonically.
 HOPF ALGEBRAS IN COMBINATORICS 25

n
(c) Show that if Q is a subring of k, then the k-linear map Sym(V )o → Sym(V ) sending f (n) 7→ xn! is
a graded Hopf isomorphism.
For this reason, the Hopf structure on Sym(V )o is called a divided power algebra.
(d) Show that when k is a field of characteristic p > 0, one has (f (1) )p = 0, and hence why there can be
no Hopf isomorphism Sym(V )o → Sym(V ).
Exercise 1.6.5. Let V have k-basis {x1 , . . . , xn }, and let V ⊕ V have k-basis {x1 , . . . , xn , y1 , . . . , yn }, so
that one has isomorphisms
Sym(V ⊕ V ) ∼
= k[x, y] ∼ = k[x] ⊗ k[y] ∼= Sym(V ) ⊗ Sym(V ).
(a) Show that our usual coproduct on Sym(V ) can be re-expressed as follows:
Sym(V ) Sym(V ) ⊗ Sym(V )
k k
∆
k[x] −→ k[x, y]
f (x1 , . . . , xn ) 7−→ f (x1 + y1 , . . . , xn + yn )
In other words, it is induced from the diagonal map
V −→ V ⊕V
(1.6.5)
xi 7−→ xi + yi
(b) One can similarly define a coproduct on the exterior algebra ∧V , which is the quotient T (V )/J where
J is the two-sided ideal generated by the elements {x2 (= x ⊗ x)}x∈V in T 2 (V ). This becomes a
graded commutative algebra
∞
n
!
M M
∧V = ∧d V = ∧d V ,
d=0 d=0

if one views the elements of V = ∧1 V as having odd degree, and uses the topologist’s sign conventions
(as in (1.3.3)). One again has ∧(V ⊕ V ) = ∧V ⊗ ∧V as graded algebras. Show that one can again
let the diagonal map (1.6.5) induce a map
∆
∧(V ) −→ ∧V ⊗ ∧V
(1.6.6) f (x1 , . . . , xn ) 7−→ f (x1 + y1 , . . . , xn + yn )
P k P k
ci1 ,...,id · xi1 ∧ · · · ∧ xid ci1 ,...,id · (xi1 + yi1 ) ∧ · · · ∧ (xid + yid )
which makes ∧V into a connected graded Hopf algebra.
(c) Show that in the tensor algebra T (V ), if one views the elements of V = V ⊗1 as having odd degree,
and uses the convention (1.3.3) in the twist map when defining T (V ), then for any x in V one has
∆(x2 ) = 1 ⊗ x2 + x2 ⊗ 1.
(d) Use part (c) to show that the two-sided ideal J ⊂ T (V ) generated by {x2 }x∈V is also a two-sided
coideal, and hence the quotient ∧V = T (V )/J inherits the structure of a bialgebra. Check that the
coproduct on ∧V inherited from T (V ) is the same as the one defined in part (b).
[Hint: The ideal J in part (b) is a graded k-submodule of T (V ), but this is not completely obvious (not
all elements of V have to be homogeneous!).]
Exercise 1.6.6. Let C be a k-coalgebra. As we know from Exercise 1.6.1(a), this makes C ∗ into a k-algebra.
Let A be a k-algebra which is finite free as k-module. This makes A∗ into a k-coalgebra.
∗
Let f : C → A and g : C → A be two k-linear maps. Show that f ∗ ? g ∗ = (f ? g) .
The above arguments might have created the impression that duals of bialgebras have good properties only
under certain restrictive conditions (e.g., the dual of a bialgebra H does not generally become a bialgebra
unless H is of finite type), and so they cannot be used in proofs and constructions unless one is willing to
sacrifice some generality (e.g., we had to require V to be finite free in Example 1.6.3). While the first part
of this impression is true, the second is not always; often there is a way to gain back the generality lost from
using duals. As an example of this, let us define the shuffle algebra of an arbitrary k-module (not just of a
dual of a finite free k-module as in Example 1.6.3):
26 DARIJ GRINBERG AND VICTOR REINER

Proposition 1.6.7. Let V be a k-module. Define a k-linear map ∆
: T (V ) → T (V ) ⊗ T (V ) by setting

n
∆
(v1 v2 · · · vn ) =
X
(v1 v2 · · · vk ) ⊗ (vk+1 vk+2 · · · vn ) for all n ∈ N and v1 , v2 , . . . , vn ∈ V.
k=0
40
Define a k-bilinear map
: T (V ) × T (V ) → T (V ), which will be written in infix notation (that is, we



will write a b instead of (a, b)), by setting41


X
(v1 v2 · · · v` ) (v`+1 v`+2 · · · v`+m ) = vσ(1) vσ(2) · · · vσ(`+m)
σ∈Sh`,m
for all `, m ∈ N and v1 , v2 , . . . , v`+m ∈ V.
42
Consider also the comultiplication  of the Hopf algebra T (V ).


Then, the k-module T (V ), endowed with the multiplication , the unit 1T (V ) ∈ V ⊗0 ⊂ T (V ), the
comultiplication ∆
and the counit , becomes a commutative Hopf algebra. This Hopf algebra is called
the shuffle algebra of V , and denoted by Sh (V ). The antipode of the Hopf algebra Sh (V ) is precisely the
antipode S of T (V ).

Exercise 1.6.8. Prove Proposition 1.6.7.

[Hint: When V is a finite free k-module, Proposition 1.6.7 follows from Example 1.6.3. The trick is to
derive the general case from this specific one. Every k-linear map f : W → V between two k-modules W


and V induces a map T (f ) : T (W ) → T (V ) which preserves ∆
, , 1T (W ) ,  and S (in the appropriate
meanings – e.g., preserving ∆
means ∆
◦ T (f ) = (T (f ) ⊗ T (f )) ◦ ∆
). Show that each of the equalities
that need to be proven in order to verify Proposition 1.6.7 can be “transported” along such a map T (f )
from a T (W ) for a suitably chosen finite free k-module W .]
It is also possible to prove Proposition 1.6.7 “by foot”, as long as one is ready to make combinatorial
arguments about cutting shuffles.
o
Remark 1.6.9. (a) Let V be a finite free k-module. The Hopf algebra T (V ) (studied in Example 1.6.3) is
naturally isomorphic to the shuffle algebra Sh (V ∗ ) (defined as in Proposition 1.6.7 but for V ∗ instead
of V ) as Hopf algebras, by the obvious isomorphism (namely, the direct sum of the isomorphisms
∗ ⊗n
(V ⊗n ) → (V ∗ ) over all n ∈ N). 43

(b) The same statement applies to the case when V is a graded k-module of finite type satisfying V0 = 0
∗ o
rather than a finite free k-module, provided that V ∗ and (V ⊗n ) are replaced by V o and (V ⊗n ) .
We shall return to shuffle algebras in Section 6.3, where we will show that under certain conditions (Q
being a subring of k, and V being a free k-module) the algebra structure on a shuffle algebra Sh(V ) is a
polynomial algebra in an appropriately chosen set of generators44.
1.7. Infinite sums and Leray’s theorem. In this section (which can be skipped, as it will not be used
except in a few exercises), we will see how a Hopf algebra structure on a k-algebra reveals knowledge about the
k-algebra itself. Specifically, we will show that if k is a commutative Q-algebra, and if A is any commutative
connected graded k-Hopf algebra, then A as a k-algebra must be (isomorphic to) a symmetric algebra of a
k-module45. This is a specimen of a class of facts which are commonly called Leray theorems; for different
specimens, see [138, Theorem 7.5] or [33, p. 17, “Hopf’s theorem”] or [33, §2.5, A, B, C] or [33, Theorem
3.8.3].46 In a sense, these facts foreshadow Zelevinsky’s theory of positive self-dual Hopf algebras, which we
shall encounter in Chapter 3; however, the latter theory works in a much less general setting (and makes
much stronger claims).
40This is well-defined, because the right hand side is n-multilinear in v , v , . . . , v , and because any n-multilinear map
1 2 n
V ×n → M into a k-module M gives rise to a unique k-linear map V ⊗n → M .
41Many authors use the symbol

instead of

here, but we prefer to reserve the former notation for the shuffle product of
words.
42Again, this is well-defined by the ` + m-multilinearity of the right hand side.

43This can be verified by comparing (1.6.1) with the definition of ∆ , and comparing (1.6.4) with the definition of .

44This says nothing about the coalgebra structure on Sh(V ) – which is much more complicated in these generators.
45If k is a field, then this simply means that A as a k-algebra must be a polynomial ring over k.
46
Notice that many of these sources assume k to be a field; some of their proofs rely on this assumption.
 HOPF ALGEBRAS IN COMBINATORICS 27

We shall first explore the possibilities of applying a formal power series v to a linear map f : C → A from a
coalgebra C to an algebra A. We have already seen an example of this in the proof of Proposition 1.4.7 above
k
(where the power series k≥0 (−1) T k ∈ k [[T ]] was applied to the locally ?-nilpotent map idA −uA A : A →
P

A); we shall now take a more systematic approach and establish general criteria for when such applications
are possible. First, we will have to make sense of infinite sums of maps from a coalgebra to an algebra. This
is somewhat technical, but the effort will pay off.
Definition 1.7.1. Let A be an abelian group (written additively).
We say that a family (aq )q∈Q ∈ AQ of elements of A is finitely supported if all but finitely many q ∈ Q
satisfy aq = 0. Clearly, if (aq )q∈Q ∈ AQ is a finitely supported family, then the sum q∈Q aq is well-defined
P

(since all but finitely many of its addends are 0). Sums like this satisfy the usual rules for sums, even
though their indexing set Q may be infinite. (For example, if (aq )q∈Q and (bq )q∈Q are two finitely supported
families in AQ , then the family (aq + bq )q∈Q is also finitely supported, and we have q∈Q aq + q∈Q bq =
P P
P
q∈Q (aq + bq ).)

Definition 1.7.2. Let C and A be two k-modules.

Q
We say that a family (fq )q∈Q ∈ (Hom (C, A)) of maps fq ∈ Hom (C, A) is pointwise finitely supported
if for each x ∈ C, the family (fq (x))q∈Q ∈ AQ of elements of A is finitely supported.47 If (fq )q∈Q ∈
Q P
(Hom (C, A)) is a pointwise finitely supported family, then the sum q∈Q fq is defined to be the map
C → A sending each x ∈ C to q∈Q fq (x). 48
P

Q
Note that the concept of a “pointwise finitely supported” family (fq )q∈Q ∈ (Hom (C, A)) is precisely the
concept of a “summable” family in [51, Definition 1].
Definition 1.7.3. For the rest of Section 1.7, we shall use the following conventions:
• Let C be a k-coalgebra. Let A be a k-algebra.
• We shall avoid our standard practice of denoting the unit map uA : k → A of a k-algebra A by u;
instead, we will use the letter u (without the subscript A) for other purposes.
Definition 1.7.2 allows us to work with infinite sums in Hom (C, A), provided that we are summing a
pointwise finitely supported family. We shall next state some properties of such sums:49
Q
Proposition 1.7.4. Let (fq )q∈Q ∈ (Hom (C, A)) be a pointwise finitely supported family. Then, the map
P
q∈Q fq belongs to Hom (C, A).

Q
Proposition 1.7.5. Let (fq )q∈Q and (gq )q∈Q be two pointwise finitely supported families in (Hom (C, A)) .
Q
Then, the family (fq + gq )q∈Q ∈ (Hom (C, A)) is also pointwise finitely supported, and satisfies
X X X
fq + gq = (fq + gq ) .
q∈Q q∈Q q∈Q

Q R
Proposition 1.7.6. Let (fq )q∈Q ∈ (Hom (C, A)) and (gr )r∈R ∈ (Hom (C, A)) be two pointwise finitely
Q×R
supported families. Then, the family (fq ? gr )(q,r)∈Q×R ∈ (Hom (C, A)) is pointwise finitely supported,

47Here are some examples of pointwise finitely supported families:

• If Q is a finite set, then any family (fq )q∈Q ∈ (Hom (C, A))Q is pointwise finitely supported.
• More generally, any finitely supported family (fq )q∈Q ∈ (Hom (C, A))Q is pointwise finitely supported.
• If C is a graded k-module, and if (fn )n∈N ∈ (Hom (C, A))N is a family of maps such that fn (Cm ) = 0 whenever
n 6= m, then the family (fn )n∈N is pointwise finitely supported.
• If C is a graded k-coalgebra and A is any k-algebra, and if f ∈ Hom (C, A) satisfies f (C0 ) = 0, then the family
(f ?n )n∈N ∈ (Hom (C, A))N is pointwise finitely supported. (This will be proven in Proposition 1.7.11(h).)

48This definition of P P
q∈Q fq generalizes the usual definition of q∈Q fq when Q is a finite set (because if Q is a finite set,
then any family (fq )q∈Q ∈ (Hom (C, A))Q is pointwise finitely supported).
49See Exercise 1.7.9 below for the proofs of these properties.
28 DARIJ GRINBERG AND VICTOR REINER

and satisfies   !
X X X
(fq ? gr ) =  fq  ? gr .
(q,r)∈Q×R q∈Q r∈R
P
Roughly speaking, the above three propositions say that sums of the form q∈Q fq (where (fq )q∈Q is
a pointwise finitely supported family) satisfy the usual rules for finite sums. Furthermore, the following
properties of pointwise finitely supported families hold:
Q
Proposition 1.7.7. Let (fq )q∈Q ∈ (Hom (C, A)) be a pointwise finitely supported family. Let (λq )q∈Q ∈
Q Q
k be any family of elements of k. Then, the family (λq fq )q∈Q ∈ (Hom (C, A)) is pointwise finitely
supported.
Q Q
Proposition 1.7.8. Let (fq )q∈Q ∈ (Hom (C, A)) and (gq )q∈Q ∈ (Hom (C, A)) be two families such that
Q
(fq )q∈Q is pointwise finitely supported. Then, the family (fq ? gq )q∈Q ∈ (Hom (C, A)) is also pointwise
finitely supported.
Exercise 1.7.9. Prove Propositions 1.7.4, 1.7.5, 1.7.6, 1.7.7 and 1.7.8.
We can now define the notion of a “pointwise ?-nilpotent” map. Roughly speaking, these will be the
elements of (Hom (C, A) , ?) that can be substituted into any power series because their powers (with respect
to the convolution ?) form a pointwise finitely supported family. Here is the definition:
Definition 1.7.10. (a) A map f ∈ Hom (C, A) is said to be pointwise ?-nilpotent if and only if the fam-
ily (f ?n )n∈N ∈ (Hom (C, A)) is pointwise finitely supported. Equivalently, a map f ∈ Hom (C, A)
N

is pointwise ?-nilpotent if and only if for each x ∈ C, the family (f ?n (x))n∈N of elements of A is
finitely supported.
(b) If f ∈ Hom (C, A) is a pointwise ?-nilpotent map, and if (λn )n∈N ∈ kN is any family of scalars, then
the family (λn f ?n )n∈N ∈ (Hom (C, A)) is pointwise finitely supported50, and thus the infinite sum
N

?n
= n∈N λn f ?n is well-defined and belongs to Hom (C, A) (by Proposition 1.7.4).51
P P
n≥0 λn f
(c) We let n (C, A) be the set of all pointwise ?-nilpotent maps f ∈ Hom (C, A). Note that this is not
necessarily a k-submodule of Hom (C, A).
(d) Consider the ring k [[T ]] of formal power series in an indeterminate T over k. For anyP power series
u ∈ k [[T ]] and any f ∈ n (C, A), we define a map u? (f ) ∈ Hom (C, A) by u? (f ) = n≥0 un f ?n ,
n ?n
P P
where u is written in the form u = n≥0 un T with (un )n≥0 ∈ kN . (This sum n≥0 un f is
well-defined in Hom (C, A), since f is pointwise ?-nilpotent.)
The following proposition gathers some properties of pointwise ?-nilpotent maps52:
Proposition 1.7.11. (a) For any f ∈ n (C, A) and k ∈ N, we have

?
(1.7.1) T k (f ) = f ?k .
(b) For any f ∈ n (C, A) and u, v ∈ k [[T ]], we have
?
(1.7.2) (u + v) (f ) = u? (f ) + v ? (f ) and
? ? ?
(1.7.3) (uv) (f ) = u (f ) ? v (f ) .
Also, for any f ∈ n (C, A) and u ∈ k [[T ]] and λ ∈ k, we have
?
(1.7.4) (λu) (f ) = λu? (f ) .
50This follows easily from Proposition 1.7.7 above. (In fact, the map f is pointwise ?-nilpotent, and thus the family
(f ?n )n∈N∈ (Hom (C, A))N is pointwise finitely supported (by the definition of “pointwise ?-nilpotent”). Hence, Proposition 1.7.7
(applied to Q = N and (fq )q∈Q = (f ?n )n∈N and (λq )q∈Q = (λn )n∈N ) shows that the family (λn f ?n )n∈N ∈ (Hom (C, A))N is
pointwise finitely supported.)
51Notice that the concept of “local ?-nilpotence” we used in the proof of Proposition 1.4.22 serves the same function (viz.,
ensuring that the sum n∈N λn f ?n is well-defined). But local ?-nilpotence is only defined when a grading is present, whereas
P

pointwise ?-nilpotence is defined in the general case. Also, local ?-nilpotence is more restrictive (i.e., a locally ?-nilpotent map
is always pointwise ?-nilpotent, but the converse does not always hold).
52See Exercise 1.7.13 below for the proofs of these properties.
 HOPF ALGEBRAS IN COMBINATORICS 29

Also, for any f ∈ n (C, A), we have

(1.7.5) 0? (f ) = 0 and
?
(1.7.6) 1 (f ) = uA C .
(c) If f, g ∈ n (C, A) satisfy f ? g = g ? f , then f + g ∈ n (C, A).
(d) For any λ ∈ k and f ∈ n (C, A), we have λf ∈ n (C, A).
(e) If f ∈ n (C, A) and g ∈ Hom (C, A) satisfy f ? g = g ? f , then f ? g ∈ n (C, A).
(f) If v ∈ k [[T ]] is a power series whose constant term is 0, then v ? (f ) ∈ n (C, A) for each f ∈ n (C, A).
(g) If u, v ∈ k [[T ]] are two power series such that the constant term of v is 0, and if f ∈ n (C, A) is
arbitrary, then
?
(1.7.7) (u [v]) (f ) = u? (v ? (f )) .
Here, u [v] denotes the composition of u with v; this is the power series obtained by substituting v
for T in u. (This power series is well-defined, since v has constant term 0.) Furthermore, notice that
the right hand side of (1.7.7) is well-defined, since Proposition 1.7.11(f) shows that v ? (f ) ∈ n (C, A).
(h) If C is a graded k-coalgebra, and if f ∈ Hom (C, A) satisfies f (C0 ) = 0, then f ∈ n (C, A).
(i) If B is any k-algebra, and if s : A → B is any k-algebra homomorphism, then every u ∈ k [[T ]] and
f ∈ n (C, A) satisfy
s ◦ f ∈ n (C, B) and u? (s ◦ f ) = s ◦ (u? (f )) .
(j) If C is a connected graded k-bialgebra, and if F : C → A is a k-algebra homomorphism, then
F − uA C ∈ n (C, A).

Example 1.7.12. Let C be a graded k-coalgebra. Let f ∈ Hom (C, A) be such that f (C0 ) = 0. Then, we
claim that the map uA C + f : C → A is ?-invertible. (This observation has already been made in the proof
of Proposition 1.4.22, at least in the particular case when C = A.)
Let us see how this claim follows from Proposition 1.7.11. First, Proposition 1.7.11(h) shows that f ∈
n (C, A). Now, define a power series u ∈ k [[T ]] by u = 1 + T . Then, the power series u has Pconstant nterm
1, and thus has a multiplicative inverse v = u−1 ∈ k [[T ]]. Consider this v. (Explicitly, v = n≥0 (−1) T n ,
?
but this does not matter for us.) Now, (1.7.3) yields (uv) (f ) = u? (f ) ? v ? (f ). Since uv = 1 (because
−1 ? ?
v = u ), we have (uv) (f ) = 1 (f ) = uA C (by (1.7.6)). Thus, u? (f ) ? v ? (f ) = (uv) (f ) = uA C . Hence,
?
?
the map u (f ) has a right ?-inverse.
Also, from u = 1 + T , we obtain
?
u? (f ) = (1 + T ) (f ) = 1? (f ) + T ? (f ) (by (1.7.2))
| {z } | {z }
=uA C =f?1

(by (1.7.1), applied to k=1)

= uA C + f ?1 = uA C + f.
|{z}
=f

Thus, the map uA C +f has a right ?-inverse (since the map u? (f ) has a right ?-inverse). A similar argument
shows that this map uA C + f has a left ?-inverse. Consequently, the map uA C + f is ?-invertible.
Exercise 1.7.13. Prove Proposition 1.7.11.
Definition 1.7.14. (a) For the rest of Section 1.7, we assume that k is a commutative Q-algebra. Thus,
n−1
1 (−1)
the two formal power series exp = n≥0 T n ∈ k [[T ]] and log (1 + T ) = n≥1 T n ∈ k [[T ]]
P P
n! n
are well-defined.
(b) Define two power series exp ∈ k [[T ]] and log ∈ k [[T ]] by exp = exp −1 and log = log (1 + T ).
(c) If u and v are two power series in k [[T ]] such that v has constant term 0, then u [v] denotes the
composition of u with v; this is the power series obtained by substituting v for T in u.
The following proposition is just a formal analogue of the well-known fact that the exponential function
and the logarithm are mutually inverse (on their domains of definition):53
53See Exercise 1.7.20 below for the proof of this proposition, as well as of the lemma and proposition that follow afterwards.
30 DARIJ GRINBERG AND VICTOR REINER

 
Proposition 1.7.15. Both power series exp and log have constant term 0 and satisfy exp log = T and
log [exp] = T .

For any map f ∈ n (C, A), the power series exp, exp and log give rise to three further maps exp? f , exp? f
?
and log f . We can also define a map log? g whenever g is a map in Hom (C, A) satisfying g − uA C ∈ n (C, A)
(but we cannot define log? f for f ∈ n (C, A), since log is not per se a power series); in order to do this, we
need a simple lemma:
?
Lemma 1.7.16. Let g ∈ Hom (C, A) be such that g − uA C ∈ n (C, A). Then, log (g − uA C ) is a well-
defined element of n (C, A).

Definition 1.7.17. If g ∈ Hom (C, A) is a map satisfying g − uA C ∈ n (C, A), then we define a map
?
log? g ∈ n (C, A) by log? g = log (g − uA C ). (This is well-defined, according to Lemma 1.7.16.)

Proposition 1.7.18. (a) Each f ∈ n (C, A) satisfies exp? f − uA C ∈ n (C, A) and

log? (exp? f ) = f.
(b) Each g ∈ Hom (C, A) satisfying g − uA C ∈ n (C, A) satisfies
exp? (log? g) = g.
(c) If f, g ∈ n (C, A) satisfy f ? g = g ? f , then f + g ∈ n (C, A) and exp? (f + g) = (exp? f ) ? (exp? g).
(d) The k-linear map 0 : C → A satisfies 0 ∈ n (C, A) and exp? 0 = uA C .
?n
(e) If f ∈ n (C, A) and n ∈ N, then nf ∈ n (C, A) and exp? (nf ) = (exp? f ) .
(f) If f ∈ n (C, A), then
X (−1)n−1
(1.7.8) log? (f + uA C ) = f ?n .
n
n≥1

Example 1.7.19. Consider again the Hopf algebra k [x] from Exercise 1.6.4. Let c1 : k [x] → k be the
k-linear map sending each polynomial p ∈ k [x] to the coefficient of x1 in p. (In other words, c1 sends each
polynomial p ∈ k [x] to its derivative at 0.)
Then, c1 ((k [x])0 ) = 0 (as can easily be seen). Hence, Proposition 1.7.11(h) shows that c1 ∈ n (k [x] , k).
Thus, a map exp? (c1 ) : k [x] → k is well-defined. It is not hard to see that this map is explicitly given by
(exp? (c1 )) (p) = p (1) for every p ∈ k [x] .
(In fact, this follows easily after showing that each n ∈ N satisfies
?n
(c1 ) (p) = n! · (the coefficient of xn in p) for every p ∈ k [x] ,
which in turn is easily seen by induction.)
Note that the equality (exp? (c1 )) (p) = p (1) shows that the map exp? (c1 ) is a k-algebra homomorphism.
This is a particular case of a fact that we will soon see (Proposition 1.7.23).
Exercise 1.7.20. Prove Proposition 1.7.15, Lemma 1.7.16 and Proposition 1.7.18.
Next, we state another sequence of facts (some of which have nothing to do with Hopf algebras), beginning
with a fact about convolutions which is similar to Proposition 1.4.3:54

Proposition 1.7.21. Let C and C 0 be two k-coalgebras, and let A and A0 be two k-algebras. Let γ : C → C 0
be a k-coalgebra morphism. Let α : A → A0 be a k-algebra morphism.
(a) If f ∈ Hom (C, A), g ∈ Hom (C, A), f 0 ∈ Hom (C 0 , A0 ) and g 0 ∈ Hom (C 0 , A0 ) satisfy f 0 ◦ γ = α ◦ f
and g 0 ◦ γ = α ◦ g, then (f 0 ? g 0 ) ◦ γ = α ◦ (f ? g).
?n
(b) If f ∈ Hom (C, A) and f 0 ∈ Hom (C 0 , A0 ) satisfy f 0 ◦ γ = α ◦ f , then each n ∈ N satisfies (f 0 ) ◦ γ =
?n
α◦f .

54See Exercise 1.7.28 below for their proofs.

 HOPF ALGEBRAS IN COMBINATORICS 31

Proposition 1.7.22.  Let C be a k-bialgebra. Let A be a commutative k-algebra. Let f ∈ Hom (C, A) be
2
such that f (ker ) = 0 and f (1) = 0. Then, any x, y ∈ C and n ∈ N satisfy
n  
?n
X n ?i
f (xy) = f (x) f ?(n−i) (y) .
i=0
i

Proposition
 1.7.23.
 Let C be a k-bialgebra. Let A be a commutative k-algebra. Let f ∈ n (C, A) be such
2
that f (ker ) = 0 and f (1) = 0. Then, exp? f : C → A is a k-algebra homomorphism.

Lemma 1.7.24. Let V be any torsionfree abelian group (written additively). Let N ∈ N. For every
k ∈ {0, 1, . . . , N }, let wk be an element of V . Assume that
N
X
(1.7.9) wk nk = 0 for all n ∈ N.
k=0
Then, wk = 0 for every k ∈ {0, 1, . . . , N }.
Lemma 1.7.25. Let V be a torsionfree abelian group (written additively). Let (wk )k∈N ∈ V N be a finitely
supported family of elements of V . Assume that
X
wk nk = 0 for all n ∈ N.
k∈N
Then, wk = 0 for every k ∈ N.
Proposition 1.7.26. Let C be a graded k-bialgebra. Let A be a commutative k-algebra. Let f ∈Hom (C,A)
2
be such that f (C0 ) = 0. Assume that55 exp? f : C → A is a k-algebra homomorphism. Then, f (ker ) =
0.
Proposition 1.7.27. Let C  be a connected
 graded k-bialgebra. Let A be a commutative k-algebra. Let
2
f ∈ n (C, A) be such that f (ker ) = 0 and f (1) = 0. Assume further that f (C) generates the k-algebra
A. Then, exp? f : C → A is a surjective k-algebra homomorphism.
Exercise 1.7.28. Prove Lemmas 1.7.24 and 1.7.25 and Propositions 1.7.21, 1.7.22, 1.7.23, 1.7.26 and 1.7.27.
?n
[Hint: For Proposition 1.7.26, show first that exp? (nf ) = (exp? f ) is a k-algebra homomorphism for
each n ∈ N. Turn this into an equality between polynomials in n, and use Lemma 1.7.25.]
With these preparations, we can state our version of Leray’s theorem:
Theorem 1.7.29. Let A be a commutative connected graded k-bialgebra.56
(a) We have idA −uA A ∈ n (A, A); thus, the map log? (idA ) ∈ n (A, A) is well-defined. We denote this
map log? (idA ) by e.
(b) We have ker e = k · 1A + (ker ) and e (A) ∼
2 2
= (ker ) / (ker ) (as k-modules).
(c) For each k-module V , let ιV be the canonical inclusion V → Sym V . Let q be the map
e ιe(A)
A −→ e (A) −→ Sym (e (A)) .
Then, q ∈ n (A, Sym (e (A))) 57.
(d) Let i be the canonical inclusion e (A) → A. Recall the universal property of the symmetric algebra:
If V is a k-module, if W is a commutative k-algebra, and if ϕ : V → W is any k-linear map, then
there exists a unique k-algebra homomorphism Φ : Sym V → W satisfying ϕ = Φ ◦ ιV . Applying this
to V = e (A), W = A and ϕ = i, we conclude that there exists a unique k-algebra homomorphism
Φ : Sym (e (A)) → A satisfying i = Φ ◦ ιe(A) . Denote this Φ by s. Then, the maps exp? q : A →
Sym (e (A)) and s : Sym (e (A)) → A are mutually inverse k-algebra isomorphisms.
55Notice that exp? f is well-defined, since Proposition 1.7.11(h) yields f ∈ n (C, A).
56Keep in mind that k is assumed to be a commutative Q-algebra.
57Do not mistake the map q for e. While every a ∈ A satisfies q (a) = e (a), the two maps q and e have different target sets,
and thus we do not have (exp? q) (a) 6= (exp? e) (a) for every a ∈ A.
32 DARIJ GRINBERG AND VICTOR REINER

 
(e) We have A ∼
2
= Sym (ker ) / (ker ) as k-algebras.
(f) The map e : A → A is a projection (i.e., it satisfies e ◦ e = e).

Remark 1.7.30. (a) The main upshot of Theorem 1.7.29 is that any commutative connected graded
k-bialgebra A (where k is a commutative Q-algebra) is isomorphic as a k-algebra to the sym-
metric algebra Sym W of some k-module W . (Specifically, Theorem 1.7.29(e) claims this for W =
2
(ker ) / (ker ) , whereas Theorem 1.7.29(d) claims this for W = e (A); these two modules W are
isomorphic by Theorem 1.7.29(b).) This is a useful statement even without any specific knowledge
about W , since symmetric algebras are a far tamer class of algebras than arbitrary commutative
algebras. For example, if k is a field, then symmetric algebras are just polynomial algebras (up
to isomorphism). This can be applied, for example, to the case of the shuffle algebra Sh (V ) of a
k-module V . The consequence is that the shuffle algebra Sh (V ) of any k-module V (where k is a
commutative Q-algebra) is isomorphic as a k-algebra to a symmetric algebra Sym W . When V is
a free k-module, one can actually show that Sh (V ) is isomorphic as a k-algebra to the symmetric
algebra of a free k-module W (that is, to a polynomial ring over k); however, this W is not easy
to characterize. Such a characterization is given by Radford’s theorem (Theorem 6.3.4 below) using
the concept of Lyndon words. Notice that if V has rank ≥ 2, then W is not finitely generated.
(b) The isomorphism in Theorem 1.7.29(e) is generally not an isomorphism of Hopf algebras. However,
with a little (rather straightforward) work, it reveals to be an isomorphism of graded k-algebras.
Actually, all maps mentioned in Theorem 1.7.29 are graded, provided that we use the appropriate
gradings for e (A) and Sym (e (A)). (To define the appropriate grading for e (A), we must show that e
is a graded map, whence e (A) is a homogeneous submodule of A; this provides e (A) with the grading
we seek. The grading on Sym (e (A)) then follows from the usual definition of the grading on the
symmetric algebra Sym V of a graded k-module V : Namely, if V is a graded k-module, then the n-th
graded component of Sym V is defined to be the span of all products of the form v1 v2 · · · vk ∈ Sym V ,
where v1 , v2 , . . . , vk ∈ V are homogeneous elements satisfying deg (v1 )+deg (v2 )+· · ·+deg (vk ) = n.)
(c) The map e : A → A from Theorem 1.7.29 is called the Eulerian idempotent of A.
(d) Theorem 1.7.29 is concerned with commutative bialgebras. Most of its claims have a “dual version”,
concerning cocommutative bialgebras. Again, the Eulerian idempotent plays a crucial role; but the
result characterizes not the k-algebra structure on A, but the k-coalgebra structure on A. This leads
to the Cartier-Milnor-Moore theorem; see [33, §3.8] and [51, §3.2]. We shall say a bit about the
Eulerian idempotent for a cocommutative bialgebra in Exercises 5.3.6 and 5.3.8.
Example 1.7.31. Consider the symmetric algebra Sym V of a k-module V . Then, Sym V is a commutative
connected graded k-bialgebra, and thus Theorem 1.7.29 can be applied to A = Sym V . What is the projection
e : A → A obtained in this case?
Theorem 1.7.29(b) shows that its kernel is
2
X X
(1.7.10) Ker e = k · 1A + (ker ) = Sym0 V + Symn V = Symn V.
| {z } | {z }
n≥2 n6=1
=Sym0 V Symn V
P
= n≥2

This does not yet characterize e completely, because we have yet to determine the action of e on Sym1 V .
Fortunately, the elements of Sym1 V are all primitive (recall that ∆Sym V (v) = 1 ⊗ v + v ⊗ 1 for each v ∈ V ),
and it can easily be shown that the map e fixes any primitive element of A 58P . Therefore, the map e fixes
all elements of Sym1 V . Since we also know that e annihilates all elements of n6=1 Symn V (by (1.7.10)),
we thus conclude that e is the canonical projection from the direct sum Sym V = n∈N Symn V onto its
L

addend Sym1 V .
Example 1.7.32. For this example, let A be the shuffle algebra Sh (V ) of a k-module V . (See Proposition


1.6.7 for its definition, and keep in mind that its product is being denoted by , whereas the notation uv is
still being used for the product of two elements u and v in the tensor algebra T (V ).)
Theorem 1.7.29 can be applied to A = Sh (V ). What is the projection e : A → A obtained in this case?
58See Exercise 5.3.6(f) further below for this proof. (While Exercise 5.3.6 requires A to be cocommutative, this requirement
is not used in the solution to Exercise 5.3.6(f). That said, this requirement is actually satisfied for A = Sym V , so we do not
even need to avoid it here.)
 HOPF ALGEBRAS IN COMBINATORICS 33

 v1 , v2 ∈ V . Indeed, define a map id : A → A by id = idA −uA A .

Let us compute e (v1 v2 ) fortwo elements e e

e ∈ n (A, A) and log? id

Then, id e + u   = log? (idA ) = e. Hence, (1.7.8) (applied to C = A and f = id)
| {zA A}
e
=idA
shows that
X (−1)n−1 ?n
(1.7.11) e= id
e .
n
n≥1
?n
Thus, we need to compute id e (v1 v2 ) for each n ≥ 1.
Notice that the map id e annihilates A0 , but fixes any element of Ak for k > 0. Thus,
(
w1 w2 · · · wk , if k > 0;
id (w1 w2 · · · wk ) =
e for any w1 , w2 , . . . , wk ∈ V.
0, if k = 0
e ?n : A → A annihilates Ak whenever n > k. In particular, for every
But it is easy to see that the map id
?n
e : A → A annihilates A2 , and therefore satisfies
n > 2, the map id
(1.7.12) e ?n (v1 v2 ) = 0
id (since v1 v2 ∈ A2 ) .
?n
e (v1 v2 ) for n ∈ {1, 2}.
It remains to find id
?1
We have id
e = id e and thus
e ?1 (v1 v2 ) = id
id e (v1 v2 ) = v1 v2
and
e ?2 (v1 v2 ) = id
id e (1)
| {z }

id|e (v{zv }) + |ide {z(v })
id|e {z(v }) + id|e (v{zv })
id|e{z(1)}
1 2 1 2 1 2

=0 v 1 v2 =v1 =v2 =v1 v2 =0

(since ∆Sh V (v1 v2 ) = 1 ⊗ v1 v2 + v1 ⊗ v2 + v1 v2 ⊗ 1)
=0
|

(v{z v }) +
1 2

v1 v2 + (v1 v2 )
| {z } | {z

0}
=0 =v1 v2 +v2 v1 =0
= v 1 v2 + v2 v1 .
Now, applying both sides of (1.7.11) to v1 v2 , we find
e (v1 v2 )
X (−1)n−1 ?n 1−1 2−1 X (−1)n−1 ?n
= e (v1 v2 ) = (−1)
id e ?1 (v1 v2 ) + (−1)
id e ?2 (v1 v2 ) +
id id
e (v1 v2 )
n 1 } | {z } | {z 2 } | {z } n | {z }
n≥1 | {z n≥3
=v1 v2 =v1 v2 +v2 v1 =0
=1 −1 (by (1.7.12))
=
2
n−1
−1 X (−1) 1
= v1 v2 + (v1 v2 + v2 v1 ) + 0= (v1 v2 − v2 v1 ) .
2 n 2
n≥3
| {z }
=0

This describes the action of e on the graded component A2 of A = Sh (V ).

Similarly, we can describe e acting on any other graded component:
e (1) = 0;
e (v1 ) = v1 for each v1 ∈ V ;
1
e (v1 v2 ) = (v1 v2 − v2 v1 ) for any v1 , v2 ∈ V ;
2
1
e (v1 v2 v3 ) = (2v1 v2 v3 − v1 v3 v2 − v2 v1 v3 − v2 v3 v1 − v3 v1 v2 + 2v3 v2 v1 ) for any v1 , v2 , v3 ∈ V,
6
....
34 DARIJ GRINBERG AND VICTOR REINER

With some more work, one can show the following formula for the action of e on any nontrivial pure tensor:
 
n k−1  −1


X X (−1) n − 1 − des σ
e (v1 v2 · · · vn ) =   vσ(1) vσ(2) · · · vσ(n)
−1
k k − 1 − des (σ −1 )
σ∈Sn k=1+des(σ )

X (−1) des(σ −1 )
 −1
n
= vσ(1) vσ(2) · · · vσ(n)
des (σ −1 ) + 1 des (σ −1 ) + 1
σ∈Sn
for any n ≥ 1 and v1 , v2 , . . . , vn ∈ V,
where we use the notation des π for the number of descents59 of any permutation π ∈ Sn . (A statement
essentially dual to this appears in [169, Theorem 9.5].)
2 2
Theorem 1.7.29(b) yields ker e = k · 1A + (ker ) . Notice, however, that (ker ) means the square of

2
the ideal ker  with respect to the shuffle multiplication ; thus, (ker ) is the k-linear span of all shuffle


products of the form a b with a ∈ ker  and b ∈ ker .
Exercise 1.7.33. Prove Theorem 1.7.29.
[Hint: (a) is easy. For (b), define an element id e = idA −uA A . Observe that e =
e of n (A, A) by id
n−1
P (−1) e ?n , and draw the conclusions that e (1A ) = 0 and that each x ∈ A satisfies id e (x) − e (x) ∈
n≥1 id
n
2 e ?n 2 2
(ker ) (because
 id (x)  ∈ (ker ) for every n ≥ 2). Use this to prove ker e ⊂ k · 1A + (ker ) . On the other
2 2
hand, prove e (ker ) = 0 by applying Proposition 1.7.26. Combine to obtain ker e = k · 1A + (ker ) .
 
Finish (b) by showing that A/ k · 1A + (ker ) ∼
2 2
= (ker ) / (ker ) as k-modules. Part (c) is easy again.
For (d), first apply Proposition 1.7.11(i) to show that exp? (s ◦ q) = s ◦ (exp? q). In light of s ◦ q = e and
exp? e = idA , this becomes idA = s ◦ (exp? q). To obtain part (d), it remains to show that exp? q is a
surjective k-algebra homomorphism; but this follows from Proposition 1.7.27. For (e), combine (d) and (b).
e (x) − e (x) ∈ (ker )2 .]
For (f), use once again the observation that each x ∈ A satisfies id

59A descent of a permutation π ∈ S means an i ∈ {1, 2, . . . , n − 1} satisfying π (i) > π (i + 1).

n
 HOPF ALGEBRAS IN COMBINATORICS 35

2. Review of symmetric functions Λ as Hopf algebra

Here we review the ring of symmetric functions, borrowing heavily from standard treatments, such as
Macdonald [125, Chap. I], Sagan [165, Chap. 4], Stanley [183, Chap. 7], and Mendes and Remmel [136],
but emphasizing the Hopf structure early on. Other recent references for this subject are [200], [167] and
[135, Chapters 2–3].
2.1. Definition of Λ. As before, k here is a commutative ring (hence could be a field or the integers Z;
these are the usual choices). Given an infinite variable set x = (x1 , x2 , . . .), a monomial xα := xα 1 α2
1 x2 · · · is
∞ 60
indexed by an element α = (α1 , α2 , . . .) in N having finite support ; such α are called weak compositions.
The nonzero ones among the integers α1 , α2 , . . . areP called the parts of the weak composition α. We will
consider the ring R(x) of formal power series f (x) = αP cα xα with cα in k of bounded degree, that is, where
α
there exists some bound d = d(f ) for which deg(x ) := i αi > d implies cα = 0. It is easy to see that the
product of two such power series is well-defined, and also has bounded degree.
The symmetric group Sn permuting
S the first n variables x1 , . . . , xn acts as a group of automorphisms on
R(x), as does the union S(∞) = n≥0 Sn of the infinite ascending chain S0 ⊂ S1 ⊂ S2 ⊂ · · · of symmetric
groups61. This group S(∞) can also be described as the group of all permutations of the set {1, 2, 3, . . .}
which leave all but finitely many elements invariant.
Definition 2.1.1. The ring of symmetric functions in x with coefficients in k, denoted Λ = Λk = Λ(x) =
Λk (x), is the S(∞) -invariant subalgebra R(x)S(∞) of R(x):
( )
X
α
Λ := f = cα x ∈ R(x) : cα = cβ if α, β lie in the same S(∞) -orbit .
α

We refer to the elements of Λ as symmetric functions (over k); however, despite this terminology, they
are not functions in the usual sense.62
Note that Λ is a graded k-algebra, since Λ = n≥0 Λn where Λn are the symmetric functions f = α cα xα
L P

which are homogeneous of degree n, meaning deg(xα ) = n for all cα 6= 0.

Exercise 2.1.2. Let f ∈ R (x). Let A be a commutative k-algebra, and a1 , a2 , . . . , ak be finitely many
elements of A. Show that substituting a1 , a2 , . . . , ak , 0, 0, 0, . . . for x1 , x2 , x3 , . . . in f yields an infinite sum in
which all but finitely many addends are zero. Hence, this sum has a value in A, which is commonly denoted
by f (a1 , a2 , . . . , ak ).
Definition 2.1.3. A partition λ = (λ1 , λ2 , . . . , λ` , 0, 0, . . .) is a weak composition whose entries weakly
decrease: λ1 ≥ · · · ≥ λ` > 0. The (uniquely defined) ` is said to be the length of the partition λ and
denoted by ` (λ). One sometimes omits trailing zeroes from a partition: e.g., one can write the partition
(3, 1, 0, 0, 0, . . .) as (3, 1). We will often (but not always) write λi for the i-th entry of the partition λ (for
instance, λ2 = 3 if λ = (5, 3, 1, 1)) and call it the i-th part of λ. The sum λ1 + λ2 + · · · + λ` = λ1 + λ2 + · · ·
(where ` = ` (λ)) of all parts of λ is called the size of λ and denoted by |λ|; for a given integer n, the
partitions of size n are referred to as the partitions of n. The empty partition () is denoted by ∅.
Every weak composition α lies in the S(∞) -orbit of a unique partition λ = (λ1 , λ2 , . . . , λ` , 0, 0, . . .) with
λ1 ≥ · · · ≥ λ` > 0. For any partition λ, define the monomial symmetric function
X
(2.1.1) mλ := xα .
α∈S(∞) λ

Letting λ run through the set Par of all partitions, this gives the monomial k-basis {mλ } of Λ. Letting λ
run only through the set Parn of partitions of n gives the monomial k-basis for Λn .
60The support of a sequence α = (α , α , α , . . .) ∈ N∞ is defined to be the set of all positive integers i for which α 6= 0.
1 2 3 i
61This ascending chain is constructed as follows: For every n ∈ N, there is an injective group homomorphism ι : S → S
n n n+1
which sends every permutation σ ∈ Sn to the permutation ιn (σ) = τ ∈ Sn defined by
(
σ (i) , if i ≤ n;
τ (i) = for all i ∈ {1, 2, . . . , n + 1} .
i, if i = n + 1
0 ι 1 ι 2 ι
These homomorphisms ιn for all n form a chain S0 −→ S1 −→ S2 −→ · · · , which is often regarded as a chain of inclusions.
62Being power series, they can be evaluated at appropriate families of variables. But this does not make them functions (no
more than polynomials are functions). The terminology “symmetric function” is thus not well-chosen; but it is standard.
36 DARIJ GRINBERG AND VICTOR REINER

Example 2.1.4. For n = 3, one has

m(3) = x31 + x32 + x33 + · · ·
m(2,1) = x21 x2 + x1 x22 + x21 x3 + x1 x23 + · · ·
m(1,1,1) = x1 x2 x3 + x1 x2 x4 + x1 x3 x4 + x2 x3 x4 + x1 x2 x5 + · · ·
Remark 2.1.5. We have defined the symmetric functions as the elements of R (x) invariant under the group
S(∞) . However, they also are the elements of R (x) invariant under the group S∞ of all permutations of
63
the set {1, 2, 3, . . .}.
Remark 2.1.6. It is sometimes convenient to work with finite variable sets x1 , . . . , xn , which one justifies as
follows. Note that the algebra homomorphism
R(x) → R(x1 , . . . , xn ) = k[x1 , . . . , xn ]
which sends xn+1 , xn+2 , . . . to 0 restricts to an algebra homomorphism
Λk (x) → Λk (x1 , . . . , xn ) = k[x1 , . . . , xn ]Sn .
Furthermore, this last homomorphism is a k-linear isomorphism when restricted to Λi for 0 ≤ i ≤ n, since
it sends the monomial basis elements mλ (x) to the monomial basis elements mλ (x1 , . . . , xn ). Thus when
one proves identities in Λk (x1 , . . . , xn ) for all n, they are valid in Λ, that is, Λ is the inverse limit of the
Λ(x1 , . . . , xn ) in the category of graded k-algebras.64
This characterization of Λ as an inverse limit of the graded k-algebras Λ(x1 , . . . , xn ) can be used as an
alternative definition of Λ. The definitions used by Macdonald [125] and Wildon [200] are closely related (see
[125, §1.2, p. 19, Remark 1], [75, §A.11] and [200, §1.7] for discussions of this definition). It also suggests
that much of the theory of symmetric functions can be rewritten in terms of the Λ(x1 , . . . , xn ) (at the cost
of extra complexity); and this indeed is possible65.
One can also define a comultiplication on Λ as follows. Note that when one decomposes the variables into
two sets (x, y) = (x1 , x2 , . . . , y1 , y2 , . . .), one has a ring homomorphism
R(x) ⊗ R(x) −→ R(x, y)
f (x) ⊗ g(x) 7−→ f (x)g(y).
This restricts to an isomorphism
(2.1.2) Λ ⊗ Λ = R(x)S(∞) ⊗ R(x)S(∞) −→ R(x, y)S(∞) ×S(∞)
where S(∞) × S(∞) denotes permutations of (finite subsets of) the x and separate permutations of (finite
subsets of) the y, because R(x, y)S(∞) ×S(∞) has k-basis {mλ (x)mµ (y)}λ,µ∈Par . As S(∞) × S(∞) is a
subgroup of the group S(∞,∞) (the group of all permutations of {x1 , x2 , . . . , y1 , y2 , . . .} leaving all but
finitely many variables invariant) acting on all of (x, y), one gets an inclusion of rings
Λ(x, y) = R(x, y)S(∞,∞) ,→ R(x, y)S(∞) ×S(∞) ∼ =Λ⊗Λ
63Proof. We need to show that Λ = R (x)S∞ . Since
( )
X
α
Λ= f = cα x ∈ R (x) : cα = cβ if α, β lie in the same S(∞) -orbit
α
and ( )
X
S∞ α
R (x) = f = cα x ∈ R (x) : cα = cβ if α, β lie in the same S∞ -orbit ,
α
this will follow immediately if we can show that two weak compositions α and β lie in the same S(∞) -orbit if and only if they
lie in the same S∞ -orbit. But this is straightforward to check (in fact, two weak compositions α and β lie in the same orbit
under either group if and only if they have the same multiset of nonzero entries).
64Warning: The word “graded” here is crucial. Indeed, Λ is not the inverse limit of the Λ(x , . . . , x ) in the category of
1 n
k-algebras. In fact, the latter limit is the k-algebra of all symmetric power series f in k [x] with the following property: For
each g ∈ N, there exists a d ∈ N such that every monomial in f that involves exactly g distinct indeterminates has degree at
most d. For example, the power series (1 + x1 ) (1 + x2 ) (1 + x3 ) · · · and m(1) + m(2,2) + m(3,3,3) + · · · satisfy this property,
although they do not lie in Λ (unless k is a trivial ring).
65See, for example, [104, Chapter SYM], [154] and [121, Chapters 10–11] for various results of this present chapter rewritten
in terms of symmetric polynomials in finitely many variables.
 HOPF ALGEBRAS IN COMBINATORICS 37

where the last isomorphism is the inverse of the one in (2.1.2). This gives a comultiplication
∆
Λ = Λ(x) −→ Λ(x, y) ,→ Λ ⊗ Λ
f (x) = f (x1 , x2 , . . .) 7−→ f (x, y) = f (x1 , x2 , . . . , y1 , y2 , . . .).
Here, f (x1 , x2 , . . . , y1 , y2 , . . .) means the result of choosing some bijection
φ : {x1 , x2 , x3 , . . .} → {x1 , x2 , . . . , y1 , y2 , . . .} and substituting φ (xi ) for every xi in f . (The choice of φ is
irrelevant since f is symmetric.66)
Example 2.1.7. One has
∆m(2,1) = m(2,1) (x1 , x2 , . . . , y1 , y2 , . . .)
= x21 x2 + x1 x22 + · · ·
+ x21 y1 + x21 y2 + · · ·
+ x1 y12 + x1 y22 + · · ·
+ y12 y2 + y1 y22 + · · ·
= m(2,1) (x) + m(2) (x)m(1) (y) + m(1) (x)m(2) (y) + m(2,1) (y)
= m(2,1) ⊗ 1 + m(2) ⊗ m(1) + m(1) ⊗ m(2) + 1 ⊗ m(2,1) .
This example generalizes easily to the following formula
X
(2.1.3) ∆mλ = mµ ⊗ mν ,
(µ,ν):
µtν=λ

in which µ t ν is the partition obtained by taking the multiset union of the parts of µ and ν, and then
reordering them to make them weakly decreasing.
Checking that ∆ is coassociative amounts to checking that
(∆ ⊗ id) ◦ ∆f = f (x, y, z) = (id ⊗∆) ◦ ∆f

inside Λ(x, y, z) as a subring of Λ ⊗ Λ ⊗ Λ. The
L counit Λ → k is defined in the usual fashion for connected
graded coalgebras, namely  annihilates I = n>0 Λn , and  is the identity on Λ0 = k; alternatively  sends
a symmetric function f (x) to its constant term f (0, 0, . . .).
Note that ∆ is an algebra morphism Λ → Λ ⊗ Λ because it is a composition of maps which are all algebra
morphisms. As the unit and counit axioms are easily checked, Λ becomes a connected graded k-bialgebra of
finite type, and hence also a Hopf algebra by Proposition 1.4.14. We will identify its antipode more explicitly
in Section 2.4 below.
2.2. Other Bases. We introduce the usual other bases of Λ, and explain their significance later.
Definition 2.2.1. Define the families of power sum symmetric functions pn , elementary symmetric functions
en , and complete homogeneous symmetric functions hn , for n = 1, 2, 3, . . . by
(2.2.1) pn := xn1 + xn2 + · · · = m(n)
X
(2.2.2) en := xi1 · · · xin = m(1n )
i1 <···<in
X X
(2.2.3) hn := x i1 · · · x in = mλ
i1 ≤···≤in λ∈Parn

Here, we are using the multiplicative notation for partitions: whenever (m1 , m2 , m3 , · · · ) is a weak compo-
sition, (1m1 2m2 3m3 · · · ) denotes the partition λ such that for every i, the multiplicity of the part i in λ
is
 mi . The  imi satisfying mi = 0 are often omitted from this notation, and so the (1n ) in (2.2.2) means



1, 1, . . . , 1. (For another example, 12 31 43 = 12 20 31 43 50 60 70 · · · means the partition (4, 4, 4, 3, 1, 1).)
| {z }
n ones

66To be more precise, the choice of φ is irrelevant because f is S -invariant, with the notations of Remark 2.1.5.
∞
38 DARIJ GRINBERG AND VICTOR REINER

By convention, also define h0 = e0 = 1, and hn = en = 0 if n < 0. Extend these multiplicatively to partitions

λ = (λ1 , λ2 , . . . , λ` ) with λ1 ≥ · · · ≥ λ` > 0:

pλ := pλ1 pλ2 · · · pλ`

eλ := eλ1 eλ2 · · · eλ`
hλ := hλ1 hλ2 · · · hλ`

Also define the Schur function

X
(2.2.4) sλ := xcont(T )
T

where T runs through all column-strict tableaux of shape λ, that is, T is an assignment of entries in {1, 2, 3, . . .}
to the cells of the Ferrers diagram67 for λ, weakly increasing left-to-right in rows, and strictly increasing
top-to-bottom in columns. Here cont(T ) denotes the weak composition |T −1 (1)|, |T −1 (2)|, |T −1 (3)|, . . . , so


Q |T −1 (i)|
that xcont(T ) = i xi . For example,68

1 1 1 4 7
2 3 3
T =
4 4 6
6 7

is a column-strict tableau of shape λ = (5, 3, 3, 2) with xcont(T ) = x31 x12 x23 x34 x05 x26 x27 .

Example 2.2.2. One has

m(1) = p(1) = e(1) = h(1) = s(1) = x1 + x2 + x3 + · · ·

s(n) = hn
s(1n ) = en

67The Ferrers diagram of a partition λ is defined as the set of all pairs (i, j) ∈ {1, 2, 3, . . .}2 satisfying j ≤ λ . This is a
i
set of cardinality |λ|. Usually, one visually represents a Ferrers diagram by drawing its elements (i, j) as points on the plane,
although (unlike the standard convention for drawing points on the plane) one lets the x-axis go top-to-bottom (i.e., the point
(i + 1, j) is one step below the point (i, j)), and the y-axis go left-to-right (i.e., the point (i, j + 1) is one step to the right of
the point (i, j)). (This is the so-called English notation, also known as the matrix notation because it is precisely the way one
labels the entries of a matrix. Other notations appear in literature, such as the French notation used, e.g., in Malvenuto’s [128],
and the Russian notation used, e.g., in parts of Kerov’s [93].) These points are drawn either as dots or as square boxes; in the
latter case, the boxes are centered at the points they represent, and they have sidelength 1 so that the boxes centered around
(i, j) and (i, j + 1) touch each other along a sideline. For example, the Ferrers diagram of the partition (3, 2, 2) is represented
as
• • •
• • (using dots) or as (using boxes).
• •

The Ferrers diagram of a partition λ uniquely determines λ. One refers to the elements of the Ferrers diagram of λ as the
cells (or boxes) of this diagram (which is particularly natural when one represents them by boxes) or, briefly, as the cells of λ.
Notation like “west”, “north”, “left”, “right”, “row” and “column” concerning cells of Ferrers diagrams normally refers to their
visual representation.
Ferrers diagrams are also known as Young diagrams.
One can characterize the Ferrers diagrams of partitions as follows: A finite subset S of {1, 2, 3, . . .}2 is the Ferrers diagram of
some partition if and only if for every (i, j) ∈ S and every (i0 , j 0 ) ∈ {1, 2, 3, . . .}2 satisfying i0 ≤ i and j 0 ≤ j, we have (i0 , j 0 ) ∈ S.
In other words, a finite subset S of {1, 2, 3, . . .}2 is the Ferrers diagram of some partition if and only if it is a lower set of the
poset {1, 2, 3, . . .}2 with respect to the componentwise order.
68To visually represent a column-strict tableau T of shape λ, we draw the same picture as when representing the Ferrers
diagram of λ, but with a little difference: a cell (i, j) is no longer represented by a dot or box, but instead is represented by the
entry of T assigned to this cell. Accordingly, the entry of T assigned to a given cell c is often referred to as the entry of T in c.
 HOPF ALGEBRAS IN COMBINATORICS 39

Example 2.2.3. One has for λ = (2, 1) that

p(2,1) = p2 p1 = (x21 + x22 + · · · )(x1 + x2 + · · · )
= m(2,1) + m(3)

e(2,1) = e2 e1 = (x1 x2 + x1 x3 + · · · )(x1 + x2 + · · · )

= m(2,1) + 3m(1,1,1)

h(2,1) = h2 h1 = (x21 + x22 + · · · + x1 x2 + x1 x3 + · · · )(x1 + x2 + · · · )

= m(3) + 2m(2,1) + 3m(1,1,1)
and
s(2,1) = x21 x2 +x21 x3 +x1 x22 +x1 x23 +x1 x2 x3 +x1 x2 x3 +x1 x2 x4 +···
11 11 12 13 12 13 12
2 3 2 3 3 2 4
= m(2,1) + 2m(1,1,1)
In fact, one has these transition matrices for n = 3 expressing elements in terms of the monomial basis mλ :
p(3) p(2,1) p(1,1,1) e(3) e(2,1) e(1,1,1)
   
m(3) 1 1 1 m(3) 0 0 1
m(2,1)  0 1 3  m(2,1)  0 1 3 
m(1,1,1) 0 0 6 m(1,1,1) 1 3 6

h h(2,1) h(1,1,1) s s(2,1) s(1,1,1)

 (3)   (3) 
m(3) 1 1 1 m(3) 1 0 0
m(2,1)  1 2 3  m(2,1)  1 1 0 
m(1,1,1) 1 3 6 m(1,1,1) 1 2 1
Our next goal is to show that eλ , sλ , hλ (and, under some conditions, the pλ as well) all give bases for Λ.
However at the moment it is not yet even clear that sλ are symmetric!

Proposition 2.2.4. Schur functions sλ are symmetric, that is, they lie in Λ.

Proof. It suffices to show sλ is symmetric under swapping the variables xi , xi+1 , by providing an involution
ι on the set of all column-strict tableaux T of shape λ which switches the cont(T ) for (i, i + 1) cont(T ).
Restrict attention to the entries i, i + 1 in T , which must look something like this:
i i i i i+1 i+1
i i i i i i+1 i+1 i+1 i+1 i+1
i+1 i+1 i+1
i
One finds several vertically aligned pairs . If one were to remove all such pairs, the remaining entries
i+1
would be a sequence of rows, each looking like this:
(2.2.5) i, i, . . . , i , i + 1, i + 1, . . . , i + 1
| {z } | {z }
r occurrences s occurrences

i
An involution due to Bender and Knuth tells us to leave fixed all the vertically aligned pairs , but
i+1
change each sequence as in (2.2.5) to this:
i, i, . . . , i , i + 1, i + 1, . . . , i + 1
| {z } | {z }
s occurrences r occurrences

For example, the above configuration in T would change to

i i i i i i+1
i i i i i+1 i+1 i+1 i+1 i+1 i+1
i i+1 i+1
40 DARIJ GRINBERG AND VICTOR REINER

It is easily checked that this map is an involution, and that it has the effect of swapping (i, i + 1) in
cont(T ). 
Remark 2.2.5. The symmetry of Schur functions allows one to reformulate them via column-strict tableaux
defined with respect to any total ordering L on the positive integers, rather than the usual 1 < 2 < 3 < · · · .
For example, one can use the reverse order69 · · · < 3 < 2 < 1, or even more exotic orders, such as
1 < 3 < 5 < 7 < ··· < 2 < 4 < 6 < 8 < ··· .
Say that an assignment T of entries in {1, 2, 3, . . .} to the cells of the Ferrers diagram of λ is an L-column-
strict tableau if it is weakly L-increasing left-to-right in rows, and strictly L-increasing top-to-bottom in
columns.

Proposition 2.2.6. For any total order L on the positive integers,

X
(2.2.6) sλ = xcont(T )
T

as T runs through all L-column-strict tableaux of shape λ.

Proof. Given a weak composition α = (α1 , α2 , . . .) with αn+1 = αn+2 = · · · = 0, assume that the integers
1, 2, . . . , n are totally ordered by L as w(1) <L · · · <L w(n) for some w in Sn . Then the coefficient of
w−1 (α)
x α = xα αn
1 · · · xn on the right side of (2.2.6) is the same as the coefficient of x
1
on the right side
of (2.2.4) defining sλ , which by symmetry of sλ is the same as the coefficient of xα on the right side of
(2.2.4). 
It is now not hard to show that pλ , eλ , sλ give bases by a triangularity argument70. For this purpose, let
us introduce a useful partial order on partitions.
Definition 2.2.7. The dominance or majorization order on Parn is the partial order on the set Parn whose
greater-or-equal relation . is defined as follows: For two partitions λ and µ of n, we set λ . µ (and say that
λ dominates, or majorizes, µ) if and only if
λ1 + λ2 + · · · + λk ≥ µ1 + µ2 + · · · + µk for k = 1, 2, . . . , n.
(The definition of dominance would not change if we would replace “for k = 1, 2, . . . , n” by “for every
positive integer k” or by “for every k ∈ N”.)
Definition 2.2.8. For a partition λ, its conjugate or transpose partition λt is the one whose Ferrers diagram
is obtained from that of λ by exchanging rows for columns71. Alternatively, one has this formula for its ith
part:
(2.2.7) (λt )i := |{j : λj ≥ i}|.
Exercise 2.2.9. Let λ, µ ∈ Parn . Show that λ . µ if and only if µt . λt .

Proposition 2.2.10. The sets {eλ }, {sλ } as λ runs through all partitions give k-bases for Λk for any
commutative ring k. The same holds for {pλ } when Q is a subring of k.

Our proof of this proposition will involve three separate arguments, one for each of the three alleged bases
{sλ }, {eλ } and {pλ }; however, all these three arguments fit the same mold: Each one shows that the alleged
basis expands invertibly triangularly72 in the basis {mλ } (possibly after reindexing), with an appropriately
chosen partial order on the indexing set. We will simplify our life by restricting ourselves to Parn for a given
n ∈ N, and by stating the common part of the three arguments in a greater generality (so that we won’t
have to repeat it thrice):

69This reverse order is what one uses when one defines a Schur function as a generating function for reverse semistandard
tableaux or column-strict plane partitions; see Stanley [183, Proposition 7.10.4].
70See Section 11.1 for some notions and notations that will be used in this argument.
71In more rigorous terms: The cells of the Ferrers diagram of λt are the pairs (j, i), where (i, j) ranges over all cells of λ. It
is easy to see that this indeed uniquely determines a partition λt .
72i.e., triangularly, with all diagonal coefficients being invertible
 HOPF ALGEBRAS IN COMBINATORICS 41

Lemma 2.2.11. Let S be a finite poset. We write ≤ for the smaller-or-equal relation of S.
Let M be a free k-module with a basis (bλ )λ∈S . Let (aλ )λ∈S be a further family of elements of M .
For each λ ∈ S, let (gλ,µ )µ∈S be the family of the coefficients in the expansion of aλ ∈ M in the basis
(bµ )µ∈S ; in other words, let (gλ,µ )µ∈S ∈ kS be such that aλ =
P
gλ,µ bµ . Assume that:
µ∈S
• Assumption A1: Any λ ∈ S and µ ∈ S satisfy gλ,µ = 0 unless µ ≤ λ.
• Assumption A2: For any λ ∈ S, the element gλ,λ of k is invertible.
Then, the family (aλ )λ∈S is a basis of the k-module M .
Proof of Lemma 2.2.11. Use the notations of Section 11.1. Assumptions A1 and A2 yield that the S × S-
matrix (gλ,µ )(λ,µ)∈S×S ∈ kS×S is invertibly triangular. But the definition of the gλ,µ yields that the family
(aλ )λ∈S expands in the family (bλ )λ∈S through this matrix (gλ,µ )(λ,µ)∈S×S . Since the latter matrix is
invertibly triangular, this shows that the family (aλ )λ∈S expands invertibly triangularly in the family (bλ )λ∈S .
Therefore, Corollary 11.1.19(e) (applied to (es )s∈S = (aλ )λ∈S and (fs )s∈S = (bλ )λ∈S ) shows that (aλ )λ∈S is
a basis of the k-module M (since (bλ )λ∈S is a basis of the k-module M ). 
Proof of Proposition 2.2.10. We can restrict our attention to each homogeneous component Λn and partitions
λ of n. Thus, we have to prove that, for each n ∈ N, the families (eλ )λ∈Parn and (sλ )λ∈Parn are bases of the
k-module Λn , and that the same holds for (pλ )λ∈Parn if Q is a subring of k.
Fix n ∈ N. We already know that (mλ )λ∈Parn is a basis of the k-module Λn .
1. We shall first show that the family (sλ )λ∈Par
P n is a basis of the k-module Λn .
For every partition λ, we have sλ = µ∈Par Kλ,µ mµ , where the coefficient Kλ,µ is the Kostka
number counting the column-strict tableaux T of shape λ having cont(T ) = µ; this follows because
both sides are symmetric functions, and Kλ,µ is the coefficient of xµ on both sides73. Thus, for every
λ ∈ Parn , one has
X
(2.2.8) sλ = Kλ,µ mµ
µ∈Parn
74
(since sλ is homogeneous of degree n). But if λ and µ are partitions satisfying Kλ,µ 6= 0,
then there exists a column-strict tableau T of shape λ having cont(T ) = µ (since Kλ,µ counts such
tableaux), and therefore we must have λ1 + λ2 + · · · + λk ≥ µ1 + µ2 + · · · + µk for each positive
integer k (since the entries 1, 2, . . . , k in T must all lie within the first k rows of λ); in other words,
75
λ . µ (if Kλ,µ 6= 0) . In other words,
(2.2.9) any λ ∈ Parn and µ ∈ Parn satisfy Kλ,µ = 0 unless λ . µ.
76
One can also check that Kλ,λ = 1 for any λ ∈ Parn . Hence,
(2.2.10) for any λ ∈ Parn , the element Kλ,λ of k is invertible.
Now, let us regard the set Parn as a poset, whose greater-or-equal relation is .. Lemma 2.2.11
(applied to S = Parn , M = Λn , aλ = sλ , bλ = mλ and gλ,µ = Kλ,µ ) shows that the family (sλ )λ∈Parn
is a basis of the k-module Λn (because the Assumptions A1 and A2 of Lemma 2.2.11 are satisfied77).
2. Before we show that (eλ )λ∈Parn is a basis, we define a few notations regarding integer matrices. A
{0, 1}-matrix means a matrix whose entries belong to the set {0, 1}. If A ∈ N`×m is a matrix, then
the row sums of A means the `-tuple (r1 , r2 , . . . , r` ), where each ri is the sum of all entries in the
i-th row of A; similarly, the column sums of A means the m-tuple (c1 , c2 , . . . , cm ), where each cj
 the sum of all entries in the j-th column of A. (For instance, the row sums of the {0, 1}-matrix
is
0 1 1 0 0
is (2, 3), whereas its column sums is (1, 2, 1, 1, 0).) We identify any k-tuple of
1 1 0 1 0
73In general, in order to prove that two symmetric functions f and g are equal, it suffices to show that, for every µ ∈ Par,
the coefficients of xµ in f and in g are equal. (Indeed, all other coefficients are determined by these coefficients because of the
symmetry.)
74See Exercise 2.2.13(c) below for a detailed proof of (2.2.8).
75See Exercise 2.2.13(d) below for a detailed proof of this fact.
76See Exercise 2.2.13(e) below for a proof of this.
77Indeed, they follow from (2.2.9) and (2.2.10), respectively.
42 DARIJ GRINBERG AND VICTOR REINER

nonnegative integers (a1 , a2 , . . . , ak ) with the weak composition (a1 , a2 , . . . , ak , 0, 0, 0, . . .); thus, the
`×m
row sums and the column sums of a matrix  in N canbe viewed as weak compositions. (For
0 1 1 0 0
example, the column sums of the matrix is the 5-tuple (1, 2, 1, 1, 0), and can be
1 1 0 1 0
viewed as the weak composition (1, 2, 1, 1, 0, 0, 0, . . .).)
For every λ ∈ Parn , one has
X
(2.2.11) eλ = aλ,µ mµ ,
µ∈Parn

where aλ,µ counts {0, 1}-matrices (of size ` (λ) × ` (µ)) having row sums λ and column sums µ:
indeed, when one expands eλ1 eλ2 · · · , choosing the monomial xj1 . . . xjλi in the eλi factor corresponds
to putting 1’s in the ith row and columns j1 , . . . , jλi of the {0, 1}-matrix 78. Applying (2.2.11) to λt
instead of λ, we see that
X
(2.2.12) eλt = aλt ,µ mµ
µ∈Parn

for every λ ∈ Parn .

It is not hard to check79 that aλ,µ vanishes unless λt . µ. Applying this to λt instead of λ, we
conclude that
(2.2.13) any λ ∈ Parn and µ ∈ Parn satisfy aλt ,µ = 0 unless λ . µ.
80
Moreover, one can show that aλt ,λ = 1 for each λ ∈ Parn . Hence,
(2.2.14) for any λ ∈ Parn , the element aλt ,λ of k is invertible.
Now, let us regard the set Parn as a poset, whose greater-or-equal relation is .. Lemma 2.2.11
(applied to S = Parn , M = Λn , aλ = eλt , bλ = mλ and gλ,µ = aλt ,µ ) shows that the family
(eλt )λ∈Parn is a basis of the k-module Λn (because the Assumptions A1 and A2 of Lemma 2.2.11 are
satisfied81). Hence, (eλ )λ∈Parn is a basis of Λn .
3. Assume now that Q is a subring of k. For every λ ∈ Parn , one has
X
(2.2.15) pλ = bλ,µ mµ ,
µ∈Parn

where bλ,µ counts the ways to partition the nonzero parts λ1 , . . . , λ` into blocks such that the sums
blocks give µ; more formally, bλ,µ is the number of maps ϕ : {1, 2, . . . , `} → {1, 2, 3, . . .} having
of theP
82
µj = i:ϕ(i)=j λi for j = 1, 2, . . . . Again it is not hard to check that
(2.2.16) any λ ∈ Parn and µ ∈ Parn satisfy bλ,µ = 0 unless µ . λ.
83
Furthermore, for any λ ∈ Parn , the element bλ,λ is a positive integer84, and thus invertible in k
(since Q is a subring of k). Thus,
(2.2.17) for any λ ∈ Parn , the element bλ,λ of k is invertible
(although we don’t always have bλ,λ = 1 this time).
Now, let us regard the set Parn as a poset, whose smaller-or-equal relation is .. Lemma 2.2.11
(applied to S = Parn , M = Λn , aλ = pλ , bλ = mλ and gλ,µ = bλ,µ ) shows that the family (pλ )λ∈Parn
is a basis of the k-module Λn (because the Assumptions A1 and A2 of Lemma 2.2.11 are satisfied85).
78See Exercise 2.2.13(g) below for a detailed proof of (2.2.11).
79See Exercise 2.2.13(h) below for a proof of this. This is the easy implication in the Gale-Ryser Theorem. (The hard
implication is the converse: It says that if λ, µ ∈ Parn satisfy λt . µ, then there exists a {0, 1}-matrix having row sums λ and
column sums µ, so that aλ,µ is a positive integer. This is proven, e.g., in [99], in [43, Theorem 2.4] and in [200, Section 5.2].)
80See Exercise 2.2.13(i) below for a proof of this.
81Indeed, they follow from (2.2.13) and (2.2.14), respectively.
82See Exercise 2.2.13(k) below for a detailed proof of (2.2.15) (and see Exercise 2.2.13(j) for a proof that the numbers b
λ,µ
are well-defined).
83See Exercise 2.2.13(l) below for a proof of this.
84This is proven in Exercise 2.2.13(m) below.
85Indeed, they follow from (2.2.16) and (2.2.17), respectively.
 HOPF ALGEBRAS IN COMBINATORICS 43

Remark 2.2.12. When
Q is not a subring of k, the family {pλ } is not (in general) a basis of Λk ; for instance,
e2 = 21 p(1,1) − p2 ∈ ΛQ is not in the Z-span of this family. However, if we define bλ,µ as in the above
proof, then the Z-linear span of all pλ equals the Z-linear span of all bλ,λ mλ . Indeed, if µ = (µ1 , µ2 , . . . , µk )
with k = `(µ), then bµ,µ is the size of the subgroup of Sk consisting of all permutations σ ∈ Sk having each
86
i satisfy µσ(i) = µi . As a consequence, bµ,µ divides bλ,µ for every partition µ of the same size as λ
87
(because
  this group acts freely on the set which is enumerated by bλ,µ ) 88. Hence, the Parn × Parn -matrix
bλ,µ
has integer entries. Furthermore, this matrix is unitriangular89 (indeed, (2.2.16)
bµ,µ (λ,µ)∈Parn × Parn
shows that it is triangular, but its diagonal entries are clearly 1) and thus invertibly triangular. But (2.2.15)
shows that the family (pλ )λ∈Parn expands in the family (bλ,λ mλ )λ∈Parn through this matrix. Hence, the
family (pλ )λ∈Parn expands invertibly triangularly in the family (bλ,λ mλ )λ∈Parn . Thus, Corollary 11.1.19(b)
(applied to Z, Λn , Parn , (pλ )λ∈Parn and (bλ,λ mλ )λ∈Parn instead of k, M , S, (es )s∈S and (fs )s∈S ) shows that
the Z-submodule of Λn spanned by (pλ )λ∈Parn is the Z-submodule of Λn spanned by (bλ,λ mλ )λ∈Parn .

The purpose of the following exercise is to fill in some details omitted from the proof of Proposition 2.2.10.

Exercise 2.2.13. Let n ∈ N.

(a) Show that every f ∈ Λn satisfies
X
f= ([xµ ] f ) mµ .
µ∈Parn

Here, [xµ ] f denotes the coefficient of the monomial xµ in the power series f .
Now, we introduce a notation (which generalizes the notation Kλ,µ from the proof of Proposition 2.2.10):
For any partition λ and any weak composition µ, we let Kλ,µ denote the number of all column-strict tableaux
T of shape λ having cont (T ) = µ.
(b) Prove that this number Kλ,µ is well-defined (i.e., there are only finitely many column-strict tableaux
T of shape λ having
P cont (T ) = µ).
(c) Show that sλ = µ∈Parn Kλ,µ mµ for every λ ∈ Parn .
(d) Show that Kλ,µ = 0 for any partitions λ ∈ Parn and µ ∈ Parn that don’t satisfy λ . µ.
(e) Show that Kλ,λ = 1 for any λ ∈ Parn .
Next, we recall a further notation: For any two partitions λ and µ, we let aλ,µ denote the number of all
{0, 1}-matrices of size ` (λ)×` (µ) having row sums λ and column sums µ. (See the proof of Proposition 2.2.10
for the concepts of {0, 1}-matrices and of row sums and column sums.)
(f) Prove that this number aλ,µ is well-defined (i.e., there are only finitely many {0, 1}-matrices of size
` (λ) × ` (µ) having
P row sums λ and column sums µ).
(g) Show that eλ = µ∈Parn aλ,µ mµ for every λ ∈ Parn .
(h) Show that aλ,µ = 0 for any partitions λ ∈ Parn and µ ∈ Parn that don’t satisfy λt . µ.
(i) Show that aλt ,λ = 1 for any λ ∈ Parn .
Next, we introduce a further notation (which generalizes the notation bλ,µ from the proof of Propo-
sition 2.2.10): For any partition λ and 
any weak composition µ, we letbλ,µ be the number of all maps
P
ϕ : {1, 2, . . . , `} → {1, 2, 3, . . .} satisfying µj = λi for all j ≥ 1, where ` = ` (λ).
 
i∈{1,2,...,`};
ϕ(i)=j

86See Exercise 2.2.13(n) below for a proof of this.

87Specifically, an element σ of the group takes ϕ : {1, 2, . . . , `} → {1, 2, 3, . . .} to σ ◦ ϕ.
88See Exercise 2.2.13(o) below for a detailed proof of this.
89Here, we are using the terminology defined in Section 11.1, and we are regarding Par as a poset whose smaller-or-equal
n
relation is ..
44 DARIJ GRINBERG AND VICTOR REINER

 only finitely many maps ϕ : {1, 2, . . . , `} →

(j) Prove that this numberbλ,µ is well-defined (i.e., there are
P
{1, 2, 3, . . .} satisfying µj = λi for all j ≥ 1).
 
i∈{1,2,...,`};
ϕ(i)=j
P
(k) Show that pλ = µ∈Parn bλ,µ mµ for every λ ∈ Parn .
(l)Show that bλ,µ = 0 for any partitions λ ∈ Parn and µ ∈ Parn that don’t satisfy µ . λ.
(m) Show that bλ,λ is a positive integer for any λ ∈ Parn .
(n) Show that for any partition µ = (µ1 , µ2 , . . . , µk ) ∈ Parn with k = ` (µ), the integer bµ,µ is the size
of the subgroup of Sk consisting of all permutations σ ∈ Sk having each i satisfy µσ(i) = µi . (In
particular, show that this subgroup is indeed a subgroup.)
(o) Show that bµ,µ | bλ,µ for every λ ∈ Parn and µ ∈ Parn .
∆
2.3. Comultiplications. Thinking about comultiplication Λ → Λ ⊗ Λ on Schur functions forces us to
immediately confront the following.
Definition 2.3.1. For partitions µ, λ say that µ ⊆ λ if µi ≤ λi for i = 1, 2, . . ., so the Ferrers diagram for µ
is a subset of the cells for the Ferrers diagram of λ. In this case, define the skew (Ferrers) diagram λ/µ to
be their set difference.90
Then define the skew Schur function sλ/µ (x) to be the sum sλ/µ := T xcont(T ) , where the sum ranges
P
over all column-strict tableaux T of shape λ/µ, that is, assignments of a value in {1, 2, 3, . . .} to each cell of
λ/µ, weakly increasing left-to-right in rows, and strictly increasing top-to-bottom in columns.
Example 2.3.2.
· · · 2 5
· 1 1
T =
2 2 4
4 5
is a column-strict tableau of shape λ/µ = (5, 3, 3, 2)/(3, 1, 0, 0) and it has xcont(T ) = x21 x32 x03 x24 x25 .
Remark 2.3.3. If µ and λ are partitions such that µ ⊆ λ, then sλ/µ ∈ Λ. (This is proven similarly as
Proposition 2.2.4.) Actually, if µ ⊆ λ, then sλ/µ ∈ Λ|λ/µ| , where |λ/µ| denotes the number of cells of the
skew shape λ/µ (so |λ/µ| = |λ| − |µ|).
It is customary to define sλ/µ to be 0 if we don’t have µ ⊆ λ. This can also be seen by a literal reading
of the definition sλ/µ := T xcont(T ) , as long as we understand that there are no column-strict tableaux of
P
shape λ/µ when λ/µ is not defined.
Clearly, every partition λ satisfies sλ = sλ/∅ .
Exercise 2.3.4. (a) State and prove an analogue of Proposition 2.2.6 for skew Schur functions.
(b) Let λ, µ, λ0 and µ0 be partitions such that µ ⊆ λ and µ0 ⊆ λ0 . Assume that the skew Ferrers
diagram λ0 /µ0 can be obtained from the skew Ferrers diagram λ/µ by a 180◦ rotation.91 Prove that
sλ/µ = sλ0 /µ0 .
Exercise 2.3.5. Let λ and µ be two partitions, and let k ∈ N be such that92 µk ≥ λk+1 . Let F be the skew
Ferrers diagram λ/µ. Let Frows≤k denote the subset of F consisting of all (i, j) ∈ F satisfying i ≤ k. Let
Frows>k denote the subset of F consisting of all (i, j) ∈ F satisfying i > k. Let α and β be two partitions
such that β ⊆ α and such that the skew Ferrers diagram α/β can be obtained from Frows≤k by parallel
90In other words, the skew Ferrers diagram λ/µ is the set of all (i, j) ∈ {1, 2, 3, . . .}2 satisfying µ < j ≤ λ .
i i
While the Ferrers diagram for a single partition λ uniquely determines λ, the skew Ferrers diagram λ/µ does not uniquely
determine µ and λ. (For instance, it is empty whenever λ = µ.) When one wants to keep µ and λ in memory, one speaks of the
skew shape λ/µ; this simply means the pair (µ, λ). Every notion defined for skew Ferrers diagrams also makes sense for skew
shapes, because to any skew shape λ/µ we can assign the skew Ferrers diagram λ/µ (even if not injectively). For instance, the
cells of the skew shape λ/µ are the cells of the skew Ferrers diagram λ/µ.
One can characterize the skew Ferrers diagrams as follows: A finite subset S of {1, 2, 3, . . .}2 is a skew Ferrers diagram (i.e.,
there exist two partitions λ and µ such that µ ⊆ λ and such that S is the skew Ferrers diagram λ/µ) if and only if for every
(i, j) ∈ S, every (i0 , j 0 ) ∈ {1, 2, 3, . . .}2 and every (i00 , j 00 ) ∈ S satisfying i00 ≤ i0 ≤ i and j 00 ≤ j 0 ≤ j, we have (i0 , j 0 ) ∈ S.
91For example, this happens when λ = (3, 2), µ = (1), λ0 = (5, 4) and µ0 = (3, 1).
92As usual, we write ν for the k-th entry of a partition ν.
k
 HOPF ALGEBRAS IN COMBINATORICS 45

translation. Let γ and δ be two partitions such that δ ⊆ γ and such that the skew Ferrers diagram γ/δ can
be obtained from Frows>k by parallel translation.93 Prove that sλ/µ = sα/β sγ/δ .
∆
Proposition 2.3.6. The comultiplication Λ → Λ ⊗ Λ has the following effect on the symmetric functions
discussed so far94:
(i) ∆pn = P1 ⊗ pn + pn ⊗ 1 for every n ≥ 1, that is, the power sums pn are primitive.
(ii) ∆en = i+j=n ei ⊗ ej for every n ∈ N.
P
(iii) ∆hn = i+j=n hi ⊗ hj for every n ∈ N.
P
(iv) ∆sλ = µ⊆λ sµ ⊗ sλ/µ for any partition λ.
P
(v) ∆sλ/ν = µ∈Par: sµ/ν ⊗ sλ/µ for any partitions λ and ν.
ν⊆µ⊆λ
Proof. Recall that ∆ sends f (x) 7→ f (x, y), and one can easily check that
(i) pn (x, y) = P i xni + i yin = pn (x) · 1 + 1 · pn (y)
P P
(ii) en (x, y) = i+j=n ei (x)ej (y)
P
(iii) hn (x, y) = i+j=n hi (x)hj (y)
For assertion (iv), note that by (2.2.6), one has
X
(2.3.1) sλ (x, y) = (x, y)cont(T ) ,
T
where the sum is over column-strict tableaux T of shape λ having entries in the linearly ordered alphabet
(2.3.2) x1 < x2 < · · · < y1 < y2 < · · · .
95
For example,
x1 x1 x1 y2 y5
x2 y1 y1
T =
y2 y2 y4
y4 y5
is such a tableau of shape λ = (5, 3, 3, 2). Note that the restriction of T to the alphabet x gives a column-
strict tableau Tx of some shape µ ⊆ λ, and the restriction of T to the alphabet y gives a column-strict
tableau Ty of shape λ/µ (e.g. for T in the example above, the tableau Ty appeared in Example 2.3.2).
Consequently, one has
X
sλ (x, y) = xcont(Tx ) · ycont(Ty )
T
! 
X X X X
(2.3.3) = xcont(Tx )  ycont(Ty )  = sµ (x)sλ/µ (y).
µ⊆λ Tx Ty µ⊆λ

P Assertion (v) is obvious in the case when we don’t have ν ⊆ λ (in fact, in this case, both sλ/ν and
µ∈Par: sµ/ν ⊗ sλ/µ are clearly zero). In the remaining case, the proof of assertion (v) is similar to that of
ν⊆µ⊆λ

93Here is an example of the situation: λ = (6, 5, 5, 2, 2), µ = (4, 4, 3, 1), k = 3 (satisfying µ = µ = 3 ≥ 2 = λ = λ

k 3 4 k+1 ),
α = (3, 2, 2), β = (1, 1), γ = (2, 2), and δ = (1).
94 The abbreviated summation indexing P
i+j=n ti,j used here is intended to mean
X
ti,j .
(i,j):
0≤i,j≤n,
i+j=n

95Here, (x, y)cont(T ) means the monomial Q |T −1 (a)| , where A denotes the totally ordered alphabet x1 < x2 < · · · <
a∈A a
y1 < y2 < · · · . In other words, (x, y)cont(T ) is the product of all entries of the tableau T (which is a monomial, since the entries
of T are not numbers but variables).
The following rather formal argument should allay any doubts as to why (2.3.1) holds: Let L denote the totally ordered
set which is given by the set {1, 2, 3, . . .} of positive integers, equipped with the total order 1 <L 3 <L 5 <L 7 <L · · · <L
2 <L 4 <L 6 <L 8 <L · · · . Then, (2.2.6) yields sλ = T xcont(T ) as T runs through all L-column-strict tableaux of shape λ.
P

Substituting the variables x1 , y1 , x2 , y2 , x3 , y3 , . . . for x1 , x2 , x3 , x4 , x5 , x6 , . . . (that is, substituting xi for x2i−1 and yi for x2i )
in this equality, we obtain (2.3.1).
46 DARIJ GRINBERG AND VICTOR REINER

(iv). (Of course, the tableaux T and Tx now have skew shapes λ/ν and µ/ν, and instead of (2.2.6), we need
to use the answer to Exercise 2.3.4(a).) 
Notice that parts (ii) and (iii) of Proposition 2.3.6 are particular cases of part (iv), since hn = s(n) and
en = s(1n ) .
Exercise 2.3.7. (a) Show
P that the Hopf algebra Λ is cocommutative.
(b) Show that ∆sλ/ν = µ∈Par: sλ/µ ⊗ sµ/ν for any partitions λ and ν.
ν⊆µ⊆λ

Exercise 2.3.8. Let n ∈ N. Consider the finite variable set (x1 , x2 , . . . , xn ) as a subset of x = (x1 , x2 , x3 , . . .).
Recall that f (x1 , x2 , . . . , xn ) is a well-defined element of k [x1 , x2 , . . . , xn ] for every f ∈ R (x) (and therefore
also for every f ∈ Λ, since Λ ⊂ R (x)), according to Exercise 2.1.2.
(a) Show that any two partitions λ and µ satisfy
X
sλ/µ (x1 , x2 , . . . , xn ) = xcont(T ) .
T is a column-strict
tableau of shape λ/µ;
all entries of T belong
to {1,2,...,n}

(b) If λ is a partition having more than n parts (where the word “parts” means “nonzero parts”), then
show that sλ (x1 , x2 , . . . , xn ) = 0.
Remark 2.3.9. An analogue of Proposition 2.2.10 holds for symmetric polynomials in finitely many variables:
Let N ∈ N. Then, we have
(a) The set {mλ (x1 , x2 , . . . , xN )}, as λ runs through all partitions having length ≤ N , is a basis of the
S
k-module Λ (x1 , x2 , . . . , xN ) = k [x1 , x2 , . . . , xN ] N .
(b) For any partition λ having length > N , we have mλ (x1 , x2 , . . . , xN ) = 0.
(c) The set {eλ (x1 , x2 , . . . , xN )}, as λ runs through all partitions whose parts are all ≤ N , is a basis of
the k-module Λ (x1 , x2 , . . . , xN ).
(d) The set {sλ (x1 , x2 , . . . , xN )}, as λ runs through all partitions having length ≤ N , is a basis of the
k-module Λ (x1 , x2 , . . . , xN ).
(e) If Q is a subring of k, then the set {pλ (x1 , x2 , . . . , xN )}, as λ runs through all partitions having
length ≤ N , is a basis of the k-module Λ (x1 , x2 , . . . , xN ).
(f) If Q is a subring of k, then the set {pλ (x1 , x2 , . . . , xN )}, as λ runs through all partitions whose parts
are all ≤ N , is a basis of the k-module Λ (x1 , x2 , . . . , xN ).
Indeed, the claims (a) and (b) are obvious, while the claims (c), (d) and (e) are proven similarly to our proof
of Proposition 2.2.10. We leave the proof of (f) to the reader; this proof can also be found in [121, Theorem
10.86]96.
Claim (c) can be rewritten as follows: The elementary symmetric polynomials ei (x1 , x2 , . . . , xN ), for
i ∈ {1, 2, . . . , N }, form an algebraically independent generating set of Λ (x1 , x2 , . . . , xN ). This is precisely
the well-known theorem (due to Gauss)97 that every symmetric polynomial in N variables x1 , x2 , . . . , xN can
be written uniquely as a polynomial in the N elementary symmetric polynomials.
2.4. The antipode, the involution ω, and algebra generators. Since Λ is a connected graded k-
S
bialgebra, it will have an antipode Λ → Λ making it a Hopf algebra by Proposition 1.4.14. However, we can
identify S more explicitly now.

Proposition 2.4.1. Each of {en }n=1,2,... , {hn }n=1,2,... are algebraically independent, and generate Λk as a
polynomial algebra for any commutative ring k. The same holds for {pn }n=1,2,... when Q is a subring of k.
Furthermore, the antipode S acts as follows:
(i) S(pn ) = −pn
(ii) S(en ) = (−1)n hn
(iii) S(hn ) = (−1)n en
96See [121, Remark 10.76] for why [121, Theorem 10.86] is equivalent to our claim (f).
97See, e.g., [37, Symmetric Polynomials, Theorem 5 and Remark 17] or [196, §5.3] or [25, Theorem 1]. In a slightly different
form, it also appears in [104, Theorem (5.10)].
 HOPF ALGEBRAS IN COMBINATORICS 47

Proof. The assertions that {en }, {pn } are algebraically independent and generate Λ are equivalent to Propo-
sition 2.2.10 asserting {eλ }, {pλ } give bases for Λ. The assertion S(pn ) = −pn follows from Proposition 1.4.15
since pn is primitive by Proposition 2.3.6(i).
For the remaining assertions, start with the easy generating function identities
∞
Y X
(2.4.1) H(t) := (1 − xi t)−1 = 1 + h1 (x)t + h2 (x)t2 + · · · = hn (x)tn
i=1 n≥0
Y∞ X
(2.4.2) E(t) := (1 + xi t) = 1 + e1 (x)t + e2 (x)t2 + · · · = en (x)tn
i=1 n≥0

which shows that

  
X X
(2.4.3) 1 = E(−t)H(t) =  en (x)(−t)n   hn (x)tn 
n≥0 n≥0

and hence, equating coefficients of powers of t, that for n = 0, 1, 2, . . . one has

X
(2.4.4) (−1)i ei hj = δ0,n .
i+j=n

This lets one recursively express the en in terms of hn and vice-versa:

e0 := 1 =: h0
(2.4.5) en = en−1 h1 − en−2 h2 + en−3 h3 − · · ·
hn = hn−1 e1 − hn−2 e2 + hn−3 e3 − · · ·
for n = 1, 2, . . . Thus if one uses the algebraic independence of the generators {en } for Λ to define an algebra
endomorphism as follows
ω
Λ → Λ
(2.4.6)
en 7−→ hn ,
then the identical form of the two recursions in (2.4.5) shows that ω also sends hn 7→ en . Therefore ω is an
involutive automorphism of Λ, and the {hn } are another algebraically independent generating set for Λ.
For the assertion about the antipode S applied to en or hn , note that the coproduct formulas for en , hn
in Proposition 2.3.6(ii),(iii) show that the defining relations for their antipodes (1.4.4) will in this case be
X X
S(ei )ej = δ0,n = ei S(ej )
i+j=n i+j=n
X X
S(hi )hj = δ0,n = hi S(hj )
i+j=n i+j=n

because u(en ) = u(hn ) = δ0,n . Comparing these to (2.4.4), one concludes via induction on n that S(en ) =
(−1)n hn and S(hn ) = (−1)n en . 
The k-algebra endomorphism ω of Λ defined in the proof of Proposition 2.4.1 is known as the fundamental
involution on Λ. We record for future use two results that were shown in the above proof: that ω is an
involution, and that it sends en to hn and vice versa for every positive integer n.
Proposition 2.4.1 shows that the antipode S on Λ is, up to sign, the same as the fundamental involution
ω: one has
(2.4.7) S(f ) = (−1)n ω(f ) for f ∈ Λn
since this formula holds for all elements of the generating set {en } (or {hn }).
Remark 2.4.2. Up to now we have not yet derived how the involution ω and the antipode S act on (skew)
Schur functions, which is quite beautiful: If λ and µ are partitions satisfying µ ⊆ λ, then
ω(sλ/µ ) = sλt /µt
(2.4.8)
S(sλ/µ ) = (−1)|λ/µ| sλt /µt
48 DARIJ GRINBERG AND VICTOR REINER

where recall that λt is the transpose or conjugate partition to λ, and |λ/µ| is the number of squares in the
skew diagram λ/µ, that is, |λ/µ| = n − k if λ, µ lie in Parn , Park respectively.
We will deduce this later in three ways (once as an exercise using the Pieri rules in Exercise 2.7.11, once
again using skewing operators in Exercise 2.8.5, and for the third time from the action of the antipode in
QSym on P -partition enumerators in Corollary 5.2.22). However, one could also deduce it immediately from
our knowledge of the action of ω and S on en , hn , if we were to prove the following famous Jacobi-Trudi and
dual Jacobi-Trudi formulas98:
Theorem 2.4.3. Skew Schur functions are the following polynomials in {hn }, {en }:
(2.4.9) sλ/µ = det(hλi −µj −i+j )i,j=1,2,...,`
(2.4.10) sλt /µt = det(eλi −µj −i+j )i,j=1,2,...,`
for any two partitions λ and µ and any ` ∈ N satisfying ` (λ) ≤ ` and ` (µ) ≤ `.
Since we appear not to need these formulas in the sequel, we will not prove them right away. However, a
proof is sketched in the solution to Exercise 2.7.13, and various proofs are well-explained in [110, (39) and
(41)], [125, §I.5], [163, Thm. 7.1], [165, §4.5], [183, §7.16], [195, Thms. 3.5 and 3.5∗ ]; also, a simultaneous
generalization of both formulas is shown in [70, Theorem 11], and three others in [160, 1.9], [73, Thm. 3.1]
and [90]. An elegant treatment of Schur polynomials taking the Jacobi-Trudi formula (2.4.9) as the definition
of sλ is given by Tamvakis [191].
2.5. Cauchy product, Hall inner product, self-duality. The Schur functions, although a bit unmo-
tivated right now, have special properties with regard to the Hopf structure. One property is intimately
connected with the following Cauchy identity.
Theorem 2.5.1. In the power series ring k [[x, y]] := k [[x1 , x2 , . . . , y1 , y2 , . . .]], one has the following expan-
sion:
Y∞ X
(2.5.1) (1 − xi yj )−1 = sλ (x)sλ (y).
i,j=1 λ∈Par

Remark 2.5.2. The left hand side of (2.5.1) is known as the Cauchy product, or Cauchy kernel.
An equivalent version of the equality (2.5.1) is obtained by replacing each xi by xi t, and writing the
resulting identity in the power series ring R(x, y)[[t]]:
Y∞ X
(2.5.2) (1 − txi yj )−1 = t|λ| sλ (x)sλ (y).
i,j=1 λ∈Par

(Recall that |λ| = λ1 + λ2 + · · · + λ` for any partition λ = (λ1 , λ2 , . . . , λ` ).)

Proof of Theorem 2.5.1. We follow the standard combinatorial proof (see [165, §4.8],[183, §7.11,7.12]), which
rewrites the left and right sides of (2.5.2), and then compares them with the Robinson-Schensted-Knuth
(RSK) bijection.99 On the left side, expanding out each geometric series
(1 − txi yj )−1 = 1 + txi yj + (txi yj )2 + (txi yj )3 + · · ·
and thinking of (xi yj )m as m occurrences of a biletter100 ji , we see that the left hand side can be rewritten


n

o
as the sum of t` (xi1 yj1 ) (xi2 yj2 ) · · · (xi` yj` ) over all multisets ji11 , . . . , ji`` of biletters. Order the biletters
in such a multiset in the lexicographic order ≤lex , which is the total order on the set of all biletters defined
by    
i1 i2
≤lex ⇐⇒ (we have i1 ≤ i2 , and if i1 = i2 , then j1 ≤ j2 ) .
j1 j2
98The second of the following identities is also known as the von Nägelsbach-Kostka identity.
99The RSK bijection has been introduced by Knuth [96], where what we call “biletters” is referred to as “two-line arrays”.
The most important ingredient of this algorithm – the RS-insertion operation – however goes back to Schensted. The special
case of the RSK algorithm where the biword has to be a permutation (written in two-line notation) and the two tableaux
have to be standard (i.e., each of them has content (1n ), where n is the size of their shape) is the famous Robinson-Schensted
correspondence [114]. More about these algorithms can be found in [165, Chapter 3], [136, Chapter 5], [183, §7.11-7.12], [121,
Sections 10.9–10.22], [60, Chapters 1 and A], [27, §3, §6] and various other places.
100A biletter here simply means a pair of letters, written as a column vector. A letter means a positive integer.
 HOPF ALGEBRAS IN COMBINATORICS 49

= ji11 ···i
i i1 i`





Defining a biword to be an array ···j` in which the biletters are ordered j1 ≤lex · · · ≤lex j` , then
`
j
over all biwords ji , where ` stands for the number
P ` cont(i) cont(j)

the left side of (2.5.2) is the sum tx y
of
P biletters in the biword. On the right side, expanding out the Schur functions as sums of tableaux gives
` cont(Q) cont(P )
(P,Q) t x y in which the sum is over all ordered pairs (P, Q) of column-strict tableaux having
the same shape101, with ` cells. (We shall refer to such pairs as tableau pairs from now on.)
The Robinson-Schensted-Knuth algorithm gives us a bijection between the biwords ji and the tableau

pairs (P, Q), which has the property that

cont(i) = cont(Q),
cont(j) = cont(P )
i


(and that the length ` of the biword j equals the size |λ| of the common shape of P and Q; but this follows
automatically from cont(i) = cont(Q)). Clearly, once such a bijection is constructed, the equality (2.5.2)
will follow.
Before we define this algorithm, we introduce a simpler operation known as RS-insertion (short for
Robinson-Schensted insertion). RS-insertion takes as input a column-strict tableau P and a letter j, and
returns a new column-strict tableau P 0 along with a corner cell102 c of P 0 , which is constructed as follows:
Start out by setting P 0 = P . The letter j tries to insert itself into the first row of P 0 by either bumping out
the leftmost letter in the first row strictly larger than j, or else placing itself at the right end of the row if
no such larger letter exists. If a letter was bumped from the first row, this letter follows the same rules to
insert itself into the second row, and so on103. This series of bumps must eventually come to an end104. At
the end of the bumping, the tableau P 0 created has an extra corner cell not present in P . If we call this
corner cell c, then P 0 (in its final form) and c are what the RS-insertion operation returns. One says that
P 0 is the result of inserting 105 j into the tableau P . It is straightforward to see that this resulting filling P 0
is a column-strict tableau106.
Example 2.5.3. To give an example of this operation, let us insert the letter j = 3 into the column-strict
1 1 3 3 4
2 2 4 6
tableau (we are showing all intermediate states of P 0 ; the underlined letter is always the one
3 4 7
5
that is going to be bumped out at the next step):
1 1 3 3 4 1 1 3 3 3 1 1 3 3 3
7−→ 7−→
2 2 4 6 insert 3; 2 2 4 6 insert 4; 2 2 4 4
3 4 7 bump out 4 3 4 7 bump out 6 3 4 7
5 5 5
1 1 3 3 3 1 1 3 3 3
7−→ 7−→
insert 6; 2 2 4 4 insert 7; 2 2 4 4
bump out 7 done .
3 4 6 3 4 6
5 5 7

101And this shape should be the Ferrers diagram of a partition (not just a skew partition).
102A corner cell of a tableau or a Ferrers diagram is defined to be a cell c which belongs to the tableau (resp. diagram) but
whose immediate neighbors to the east and to the south don’t.
103Here, rows are allowed to be empty – so it is possible that a letter is bumped from the last nonempty row of P 0 and
settles in the next, initially empty, row.
104
since we can only bump out entries from nonempty rows
105This terminology is reminiscent of insertion into binary search trees, a basic operation in theoretical computer science.
This is more than superficial similarity; there are, in fact, various analogies between Ferrers diagrams (and their fillings) and
unlabelled plane binary trees (resp. their labellings), and one of them is the analogy between RS-insertion and binary search
tree insertion. See [82, §4.1].
106Indeed, the reader can check that P 0 remains a column-strict tableau throughout the algorithm that defines RS-insertion.
(The only part of this that isn’t obvious is showing that when a letter t bumped out of some row k is inserted into row k + 1,
the property that the letters increase strictly down columns is preserved. Argue that the bumping-out of t from row k was
caused by the insertion of another letter u < t, and that the cell of row k + 1 into which t is then being inserted is in the same
column as this u, or in a column further left than it.)
50 DARIJ GRINBERG AND VICTOR REINER

The last tableau in this sequence is the column-strict tableau that is returned. The corner cell that is
returned is the second cell of the fourth row (the one containing 7).
RS-insertion will be used as a step in the RSK algorithm; the construction will rely on a simple fact known
as the row bumping lemma. Let us first define the notion of a bumping path (or bumping route): If P is a
column-strict tableau, and j is a letter, then some letters are inserted into some cells when RS-insertion is
applied to P and j. The sequence of these cells (in the order in which they see letters inserted into them) is
called the bumping path for P and j. This bumping path always ends with the corner cell c which is returned
by RS-insertion. As an example, when j = 1 is inserted into the tableau P shown below, the result P 0 is
shown with all entries on the bumping path underlined:
1 1 1 2 3
1 1 2 2 3
2 2 2 4
2 2 4 4 insert
P = 7−→ P0 = 3 4 4
3 4 5 j=1
4 5 6
4 6 6
6
A first simple observation about bumping paths is that bumping paths trend weakly left – that is, if the
bumping path of P and j is (c1 , c2 , . . . , ck ), then, for each 1 ≤ i < k, the cell ci+1 lies in the same column as
ci or in a column further left.107 A subtler property of bumping paths is the following row bumping lemma
([60, p. 9]):
Row bumping lemma: Let P be a column-strict tableau, and let j and j 0 be two letters.
Applying RS-insertion to the tableau P and the letter j yields a new column-strict tableau
P 0 and a corner cell c. Applying RS-insertion to the tableau P 0 and the letter j 0 yields a
new column-strict tableau P 00 and a corner cell c0 .
(a) Assume that j ≤ j 0 . Then, the bumping path for P 0 and j 0 stays strictly to the right,
within each row, of the bumping path for P and j. The cell c0 (in which the bumping
path for P 0 and j 0 ends) is in the same row as the cell c (in which the bumping path
for P and j ends) or in a row further up; it is also in a column further right than c.
(b) Assume instead that j > j 0 . Then, the bumping path for P 0 and j 0 stays weakly to
the left, within each row, of the bumping path for P and j. The cell c0 (in which the
bumping path for P 0 and j 0 ends) is in a row further down than the cell c (in which
the bumping path for P and j ends); it is also in the same column as c or in a column
further left.
108
This lemma can be easily proven by induction over the
row.
i
We can now define the actual RSK algorithm. Let j be a biword. Starting with the pair (P0 , Q0 ) = (∅, ∅)
and m = 0, the algorithm applies the following steps (see Example 2.5.4 below):
• If im+1 does not exist (that is, m is the length of i), stop.
• Apply RS-insertion to the column-strict tableau Pm and the letter jm+1 (the bottom letter of jim+1


m+1
).
Let Pm+1 be the resulting column-strict tableau, and let cm+1
be the resulting corner cell.
• Create Qm+1 from Qm by adding the top letter im+1 of jim+1 m+1
to Qm in the cell cm+1 (which, as we
recall, is the extra corner cell of Pm+1 not present in Pm ).
107This follows easily from the preservation of column-strictness during RS-insertion.
108We leave the details to the reader, only giving the main idea for (a) (the proof of (b) is similar). To prove the first claim
of (a), it is enough to show that for every i, if any letter is inserted into row i during RS-insertion for P 0 and j 0 , then some letter
is also inserted into row i during RS-insertion for P and j, and the former insertion happens in a cell strictly to the right of the
cell where the latter insertion happens. This follows by induction over i. In the induction step, we need to show that if, for a
positive integer i, we try to consecutively insert two letters k and k0 , in this order, into the i-th row of a column-strict tableau,
possibly bumping out existing letters in the process, and if we have k ≤ k0 , then the cell into which k is inserted is strictly to
the left of the cell into which k0 is inserted, and the letter bumped out by the insertion of k is ≤ to the letter bumped out by
the insertion of k0 (or else the insertion of k0 bumps out no letter at all – but it cannot happen that k0 bumps out a letter but
k does not). This statement is completely straightforward to check (by only studying the i-th row). This way, the first claim
of (a) is proven, and this entails that the cell c0 (being the last cell of the bumping path for P 0 and j 0 ) is in the same row as
the cell c or in a row further up. It only remains to show that c0 is in a column further right than c. This follows by noticing
that, if k is the row in which the cell c0 lies, then c0 is in a column further right than the entry of the bumping path for P and
j in row k (by the first claim of (a)), and this latter entry is further right than or in the same column as the ultimate entry c
of this bumping path (since bumping paths trend weakly left).
 HOPF ALGEBRAS IN COMBINATORICS 51

• Set m to m + 1.
After all of the biletters have been thus processed, the result of the RSK algorithm is (P` , Q` ) =: (P, Q).
Example 2.5.4. The term in the expansion of the left side of (2.5.1) corresponding to
(x1 y2 )1 (x1 y4 )1 (x2 y1 )1 (x4 y1 )1 (x4 y3 )2 (x5 y2 )1
i 1124445



is the biword j = 2411332 , whose RSK algorithm goes as follows:

P0 = ∅ Q0 = ∅

P1 = 2 Q1 = 1

P2 = 2 4 Q2 = 1 1

1 4 1 1
P3 = Q3 =
2 2

1 1 1 1
P4 = Q4 =
2 4 2 4

1 1 3 1 1 4
P5 = Q5 =
2 4 2 4

1 1 3 3 1 1 4 4
P6 = Q6 =
2 4 2 4

1 1 2 3 1 1 4 4
P := P7 = 2 3 Q := Q7 = 2 4
4 5
The bumping rule obviously maintains the property that Pm is a column-strict tableau of some Ferrers
shape throughout. It should be clear that (Pm , Qm ) have the same shape at each stage. Also, the construction
of Qm shows that it is at least weakly increasing in rows and weakly increasing in columns throughout. What
is perhaps least clear is that Qm remains strictly increasing down columns. That is, when one has a string of
equal letters on top im = im+1 = · · · = im+r , so that on bottom one bumps in jm ≤ jm+1 ≤ · · · ≤ jm+r , one
needs to know that the new cells form a horizontal strip, that is, no two of them lie in the same column109.
This follows from (the last claim of) part (a) of the row bumping lemma. Hence, the result (P, Q) of the
RSK algorithm is a tableau pair.
To see that the RSK map is a bijection, we show how to recover ji from (P, Q). This is done by reverse

bumping from (Pm+1 , Qm+1 ) to recover both the biletter jim+1

m+1
and the tableaux (Pm , Qm ), as follows.
Firstly, im+1 is the maximum entry of Qm+1 , and Qm is obtained by removing the rightmost occurrence of
this letter im+1 from Qm+1 . 110 To produce Pm and jm+1 , find the position of the rightmost occurrence
of im+1 in Qm+1 , and start reverse bumping in Pm+1 from the entry in this same position, where reverse
bumping an entry means inserting it into one row higher by having it bump out the rightmost entry which
is strictly smaller.111 The entry bumped out of the first row is jm+1 , and the resulting tableau is Pm .

109Actually, each of these new cells (except for the first one) is in a column further right than the previous one. We will use
this stronger fact further below.
110It necessarily has to be the rightmost occurrence, since (according to the previous footnote) the cell into which i
m+1 was
filled at the step from Qm to Qm+1 lies further right than any existing cell of Qm containing the letter im+1 .
111Let us give a few more details on this “reverse bumping” procedure. Reverse bumping (also known as RS-deletion or
reverse RS-insertion) is an operation which takes a column-strict tableau P 0 and a corner cell c of P 0 , and constructs a column-
strict tableau P and a letter j such that RS-insertion for P and j yields P 0 and c. It starts by setting P = P 0 , and removing
the entry in the cell c from P . This removed entry is then denoted by k, and is inserted into the row of P above c, bumping
out the rightmost entry which is smaller than k. The letter which is bumped out – say, ` –, in turn, is inserted into the row
above it, bumping out the rightmost entry which is smaller than `. This procedure continues in the same way until an entry
52 DARIJ GRINBERG AND VICTOR REINER

Finally, to see that the RSK map is surjective, one needs to show that the reverse bumping procedure can
be applied to any pair (P,
Q) of column-strict tableaux of the same shape, and will result in a (lexicograph-
ically ordered) biword ji . We leave this verification to the reader.112 

This is by far not the only known proof of Theorem 2.5.1. Two further proofs will be sketched in
Exercise 2.7.10 and Exercise 2.7.8.
Before we move on to extracting identities in Λ from Theorem 2.5.1, let us state (as an exercise) a simple
technical fact that will be useful:
Exercise 2.5.5. Let (qλ )λ∈Par be a basis of the k-module Λ. Assume that for each partition λ, the element
qλ ∈ Λ is homogeneous of degree |λ|.
(a) If two families (aλ )λ∈Par ∈ kPar and (bλ )λ∈Par ∈ kPar satisfy
X X
(2.5.3) aλ qλ (x) = bλ qλ (x)
λ∈Par λ∈Par

in k [[x]], then (aλ )λ∈Par = (bλ )λ∈Par . 113

(b) Consider a further infinite family y = (y1 , y2 , y3 , . . .) of indeterminates (disjoint from x). If two
2 2
families (aµ,ν )(µ,ν)∈Par2 ∈ kPar and (bµ,ν )(µ,ν)∈Par2 ∈ kPar satisfy
X X
(2.5.4) aµ,ν qµ (x) qν (y) = bµ,ν qµ (x) qν (y)
(µ,ν)∈Par2 (µ,ν)∈Par2

in k [[x, y]], then (aµ,ν )(µ,ν)∈Par2 = (bµ,ν )(µ,ν)∈Par2 .

(c) Consider a further infinite family z = (z1 , z2 , z3 , . . .) of indeterminates (disjoint from x and y). If
3 3
two families (aλ,µ,ν )(µ,ν,λ)∈Par3 ∈ kPar and (bλ,µ,ν )(µ,ν,λ)∈Par3 ∈ kPar satisfy
X X
(2.5.5) aλ,µ,ν qµ (x) qν (y) qλ (z) = bλ,µ,ν qµ (x) qν (y) qλ (z)
(µ,ν,λ)∈Par3 (µ,ν,λ)∈Par3

in k [[x, y, z]], then (aλ,µ,ν )(µ,ν,λ)∈Par3 = (bλ,µ,ν )(µ,ν,λ)∈Par3 .

is bumped out of the first row (which will eventually happen). The reverse bumping operation returns the resulting tableau P
and the entry which is bumped out of the first row.
It is straightforward to check that the reverse bumping operation is well-defined (i.e., P does stay a column-strict tableau
throughout the procedure) and is the inverse of the RS-insertion operation. (In fact, these two operations undo each other step
by step.)
112It is easy to see that repeatedly applying reverse bumping to (P, Q) will result in a sequence i`
, i`−1
, . . . , i1
of
j` j j1
···i`

`−1
biletters such that applying the RSK algorithm to ji1 ···j gives back (P, Q). The question is why we have ji1 ≤lex · · · ≤lex ji` .


1 ` 1 `
Since the chain of inequalities i1 ≤ i2 ≤ · · · ≤ i` is clear from the choice of entry to reverse-bump, it only remains to show
that for every string im = im+1 = · · · = im+r of equal top letters, the corresponding bottom letters weakly increase (that is,
jm ≤ jm+1 ≤ · · · ≤ jm+r ). One way to see this is the following:
Assume the contrary; i.e., assume that the bottom letters corresponding to some string im = im+1 = · · · = im+r of equal
top letters do not weakly increase. Thus, jm+p > jm+p+1 for some p ∈ {0, 1, . . . , r − 1}. Consider this p.
Let us consider the cells containing the equal letters im = im+1 = · · · = im+r in the tableau Qm+r . Label these cells as
cm , cm+1 , . . . , cm+r from left to right (noticing that no two of them lie in the same column, since Qm+r is column-strict). By
the definition of reverse bumping, the first entry to be reverse bumped from Pm+r is the entry in position cm+r (since this is
the rightmost occurrence of the letter im+r in Qm+r ); then, the next entry to be reverse bumped is the one in position cm+r−1 ,
etc., moving further and further left. Thus, for each q ∈ {0, 1, . . . , r}, the tableau Pm+q−1 is obtained from Pm+q by reverse
bumping the entry in position cm+q . Hence, conversely, the tableau Pm+q is obtained from Pm+q−1 by RS-inserting the entry
jm+q , which creates the corner cell cm+q .
But recall that jm+p > jm+p+1 . Hence, part (b) of the row bumping lemma (applied to Pm+p−1 , jm+p , jm+p+1 , Pm+p ,
cm+p , Pm+p+1 and cm+p+1 instead of P , j, j 0 , P 0 , c, P 00 and c0 ) shows that the cell cm+p+1 is in the same column as the cell
cm+p or in a column further left. But this contradicts the fact that the cell cm+p+1 is in a column further right than the cell
cm+p (since we have labeled our cells as cm , cm+1 , . . . , cm+r from left to right, and no two of them lied in the same column).
This contradiction completes our proof.
113Note that this does not immediately follow from the linear independence of the basis (q )
λ λ∈Par . Indeed, linear indepen-
dence would help if the sums in (2.5.3) were finite, but they are not. A subtler argument (involving the homogeneity of the qλ )
thus has to be used.
 HOPF ALGEBRAS IN COMBINATORICS 53

Remark 2.5.6. Clearly, for any n ∈ N, we can state an analogue of Exercise 2.5.5 for n infinite families
xi = (xi,1 , xi,2 , xi,3 , . . .) of indeterminates (with i ∈ {1, 2, . . . , n}). The three parts of Exercise 2.5.5 are the
particular cases of this analogue for n = 1, for n = 2 and for n = 3. We have shied away from stating this
analogue in full generality because these particular cases are the only ones we will need.

Corollary 2.5.7. In the Schur function basis {sλ } for Λ, the structure constants for multiplication and
comultiplication are the same, that is, if one defines scalars cλµ,ν , ĉλµ,ν via the unique expansions
X
(2.5.6) sµ sν = cλµ,ν sλ ,
λ
X
(2.5.7) ∆(sλ ) = ĉλµ,ν sµ ⊗ sν ,
µ,ν

then cλµ,ν = ĉλµ,ν .

Proof. Work in the ring k [[x, y, z]], where y = (y1 , y2 , y3 , . . .) and z = (z1 , z2 , z3 , . . .) are two new sets of
variables. The identity (2.5.1) lets one interpret both cλµ,ν , ĉλµ,ν as the coefficient114 of sµ (x)sν (y)sλ (z) in the
product
∞ ∞
! !
Y Y (2.5.1) X X
(1 − xi zj )−1 (1 − yi zj )−1 = sµ (x)sµ (z) sν (y)sν (z)
i,j=1 i,j=1 µ ν
X
= sµ (x)sν (y) · sµ (z)sν (z)
µ,ν
!
X X
= sµ (x)sν (y) cλµ,ν sλ (z)
µ,ν λ
since, regarding x1 , x2 , . . . , y1 , y2 , . . . as lying in a single variable set (x, y), separate from the variables z,
the Cauchy identity (2.5.1) expands the same product as
Y∞ ∞
Y X
(1 − xi zj )−1 (1 − yi zj )−1 = sλ (x, y)sλ (z)
i,j=1 i,j=1 λ
!
X X
= ĉλµ,ν sµ (x)sν (y) sλ (z).
λ µ,ν


Definition 2.5.8. The coefficients cλµ,ν = ĉλµ,ν appearing in the expansions (2.5.6) and (2.5.7) are called
Littlewood-Richardson coefficients.
Remark 2.5.9. We will interpret cλµ,ν combinatorially in Section 2.6. By now, however, we can already prove
some properties of these coefficients:
We have
(2.5.8) cλµ,ν = cλν,µ for all λ, µ, ν ∈ Par
(by comparing coefficients in λ cλµ,ν sλ = sµ sν = sν sµ = λ cλν,µ sλ ). Furthermore, let λ and µ be two
P P
partitions (not necessarily satisfying µ ⊆ λ). Comparing the expansion
!
X X X
λ λ
sλ (x, y) = ∆(sλ ) = cµ,ν sµ (x)sν (y) = cµ,ν sν (y) sµ (x)
µ,ν µ∈Par ν∈Par

with X X
sλ (x, y) = sµ (x)sλ/µ (y) = sµ (x)sλ/µ (y)
µ⊆λ µ∈Par

114Let us explain why speaking of coefficients makes sense here:

P
We want to use the fact that if a power series f ∈ k [[x, y, z]] is written in the form f = (µ,ν,λ)∈Par3 aλ,µ,ν sµ (x) sν (y) sλ (z)
for some coefficients aλ,µ,ν ∈ k, then these coefficients aλ,µ,ν are uniquely determined by f . But this fact is precisely the claim
of Exercise 2.5.5(c) above (applied to qλ = sλ ).
54 DARIJ GRINBERG AND VICTOR REINER

115
one concludes that
!
X X X X
cλµ,ν sν (y) sµ (x) = sµ (x)sλ/µ (y) = sλ/µ (y)sµ (x).
µ∈Par ν∈Par µ∈Par µ∈Par

Treating the indeterminates y as constants, and comparing coefficients before sµ (x) on both sides of this
equality116, we arrive at another standard interpretation for cλµ,ν :
X
sλ/µ = cλµ,ν sν .
ν

In particular, cλµ,ν
vanishes unless µ ⊆ λ. Consequently, cλµ,ν vanishes unless ν ⊆ λ as well (since cλµ,ν = cλν,µ )
and furthermore vanishes unless the equality |µ| + |ν| = |λ| holds117. Altogether, we conclude that cλµ,ν
vanishes unless µ, ν ⊆ λ and |µ| + |ν| = |λ|.
Exercise 2.5.10. Show that any four partitions κ, λ, ϕ and ψ satisfy
X ρ ρ X
cκ,λ cϕ,ψ = cλβ,δ cϕ ψ
α,β cγ,δ .
ρ∈Par (α,β,γ,δ)∈Par4

Exercise 2.5.11. (a) For any partition µ, prove that

∞
−1
X Y
sλ (x) sλ/µ (y) = sµ (x) · (1 − xi yj )
λ∈Par i,j=1

in the power series ring k [[x, y]] = k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . .]].

(b) Let α and β be two partitions. Show that
 
∞
−1
X X Y
sλ/α (x) sλ/β (y) =  sβ/ρ (x) sα/ρ (y) · (1 − xi yj )
λ∈Par ρ∈Par i,j=1

in the power series ring k [[x, y]] = k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . .]].

[Hint: For (b), expand the product
∞ ∞ ∞ ∞
−1 −1 −1 −1
Y Y Y Y
(1 − xi yj ) (1 − xi wj ) (1 − zi yj ) (1 − zi wj )
i,j=1 i,j=1 i,j=1 i,j=1

in the power series ring k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . . , z1 , z2 , z3 , . . . , w1 , w2 , w3 , . . .]] in two ways: once by

applying Theorem 2.5.1 to the two variable sets (z, x) and (w, y) and then using (2.3.3); once again by
applying (2.5.1) to the two variable sets z and w and then applying Exercise 2.5.11(a) twice.]
The statement of Exercise 2.5.11(b) is known as the skew Cauchy identity, and appears in Sagan-Stanley
[166, Cor. 6.12], Stanley [183, exercise 7.27(c)] and Macdonald [125, §I.5, example 26]; it seems to be
due to Zelevinsky. It generalizes the statement of Exercise 2.5.11(a), which in turn is a generalization of
Theorem 2.5.1.
Definition 2.5.12. Define the Hall inner product on Λ to be the k-bilinear form (·, ·) which makes {sλ } an
orthonormal basis, that is, (sλ , sν ) = δλ,ν .
Exercise 2.5.13. (a) If n and m are two distinct nonnegative integers, and if f ∈ Λn and g ∈ Λm , then
show that (f, g) = 0.
(b) If n ∈ N and f ∈ Λn , then prove that (hn , f ) = f (1) (where f (1) is defined as in Exercise 2.1.2).
The Hall inner product induces a k-module homomorphism Λ → Λo (sending every f ∈ Λ to the k-linear
map Λ → k, g 7→ (f, g)). This homomorphism is invertible (since the Hall inner product has an orthonormal
basis), so that Λo ∼
= Λ as k-modules. But in fact, more can be said:
115In the last equality, we removed the condition µ ⊆ λ on the addends of the sum; this does not change the value of the
sum (because we have sλ/µ = 0 whenever we don’t have µ ⊆ λ).
116“Comparing coefficients” means applying Exercise 2.5.5(a) to q = s in this case (although the base ring k is now
λ λ
replaced by k [[y]], and the index µ is used instead of λ, since λ is already taken).
117In fact, this is clear when we don’t have µ ⊆ λ. When we do have µ ⊆ λ, this follows from observing that s
λ/µ ∈ Λ|λ/µ|
has zero coefficient before sν whenever |µ| + |ν| 6= |λ|.
 HOPF ALGEBRAS IN COMBINATORICS 55

Corollary 2.5.14. The isomorphism Λo ∼

= Λ induced by the Hall inner product is an isomorphism of Hopf
algebras.
Proof. We have seen that the orthonormal basis {sλ } of Schur functions is self-dual, in the sense that its
multiplication and comultiplication structure constants are the same. Thus the isomorphism Λo ∼= Λ induced
by the Hall inner product is an isomorphism of bialgebras118, and hence also a Hopf algebra isomorphism by
Proposition 1.4.24(c). 
We next identify two other dual pairs of bases, by expanding the Cauchy product in two other ways.
Proposition 2.5.15. One can also expand
∞
Y X X
(2.5.11) (1 − xi yj )−1 = hλ (x)mλ (y) = zλ−1 pλ (x)pλ (y)
i,j=1 λ∈Par λ∈Par
m1 m2 m1 m2
where zλ := m1 ! · 1 · m2 ! · 2 · · · if λ = (1 , 2 , . . .) with multiplicity mi for the part i. (Here, we
assume that Q is a subring of k for the last equality.)
Remark 2.5.16. It is relevant later (and explains the notation) that zλ is the size of the Sn -centralizer
subgroup for a permutation having cycle type119 λ with |λ| = n. This is a classical (and fairly easy) result
(see, e.g., [165, Prop. 1.1.1] or [183, Prop. 7.7.3] for a proof).
118Here are some details on the proof:
Let γ : Λ → Λo be the k-module isomorphism Λ → Λo induced by the Hall inner product. We want to show that γ is an
isomorphism of bialgebras.
Let s∗λ be the basis of Λo dual to the basis {sλ } of Λ. Thus, for any partition λ, we have


(2.5.9) γ (sλ ) = s∗λ

(since any partition µ satisfies (γ (sλ )) (sµ ) = (sλ , sµ ) = δλ,µ = s∗λ (sµ ), and thus the two k-linear maps γ (sλ ) : Λ → k and
s∗λ : Λ → k are equal to each other on the basis {sµ } of Λ, which forces them to be identical).
The coproduct structure constants of the basis s∗λ of Λo equal the product structure constants of the basis {sλ } of Λ


(according to our discussion of duals in Section 1.6). Since the latter are the Littlewood-Richardson numbers cλ µ,ν (because of
(2.5.6)), we thus conclude that the former are cλ µ,ν as well. In other words, every λ ∈ Par satisfies
X
(2.5.10) ∆Λo s∗λ = cλ ∗ ∗
µ,ν sµ ⊗ sν
µ,ν

(where the sum is over all pairs (µ, ν) of partitions). On the other hand, applying the map γ ⊗ γ : Λ ⊗ Λ → Λo ⊗ Λo to the
equality (2.5.7) yields
 
X X X
λ ∗ ∗
(γ ⊗ γ) (∆(sλ )) = (γ ⊗ γ)  ĉµ,ν sµ ⊗ sν  = ĉλ
µ,ν γ (sµ ) ⊗ γ (sν ) = cλ
µ,ν sµ ⊗ sν
µ,ν µ,ν |{z} µ,ν
| {z } | {z }
=cλ =s∗µ =s∗ ν
µ,ν
(by (2.5.9)) (by (2.5.9))

= ∆Λo s∗λ (by (2.5.10))

|{z}
=γ(sλ )
(by (2.5.9))
= ∆Λo (γ (sλ ))
for each λ ∈ Par. In other words, the two k-linear maps (γ ⊗ γ) ◦ ∆ and ∆Λo ◦ γ are equal to each other on each sλ with
λ ∈ Par. Hence, these two maps must be identical (since the sλ form a basis of Λ). Hence, ∆Λo ◦ γ = (γ ⊗ γ) ◦ ∆.
Our next goal is to show that Λo ◦ γ = . Indeed, each λ ∈ Par satisfies
(Λo ◦ γ) (sλ ) = Λo (γ (sλ )) = (γ (sλ )) (1) (by the definition of Λo )
 

= sλ , 1  = (sλ , s∅ ) = δλ,∅ =  (sλ ) .

 
|{z}
=s∅

Hence, Λo ◦ γ = . Combined with ∆Λo ◦ γ = (γ ⊗ γ) ◦ ∆, this shows that γ is a k-coalgebra homomorphism. Similar reasoning
can be used to prove that γ is a k-algebra homomorphism. Altogether, we thus conclude that γ is a bialgebra homomorphism.
Since γ is a k-module isomorphism, this yields that γ is an isomorphism of bialgebras. Qed.
119If σ is a permutation of a finite set X, then the cycle type of σ is defined as the list of the lengths of all cycles of σ (that
is, of all orbits of σ acting on X) written in decreasing order. This is clearly a partition of |X|. (Some other authors write it
in increasing order instead, or treat it as a multiset.)
For instance, the permutation of the set {0, 3, 6, 9, 12} which sends 0 to 3, 3 to 9, 6 to 6, 9 to 0, and 12 to 12 is (3, 1, 1),
since the cycles of this permutation have lengths 3, 1 and 1.
56 DARIJ GRINBERG AND VICTOR REINER

Proof of Proposition 2.5.15. For the first expansion, note that (2.4.1) shows
∞
Y ∞ X
Y
(1 − xi yj )−1 = hn (x)yjn
i,j=1 j=1 n≥0
X
= (hn1 (x)hn2 (x) · · · )(y1n1 y2n2 · · · )
weak
compositions
(n1 ,n2 ,...)
X X
= (hn1 (x)hn2 (x) · · · ) (y1n1 y2n2 · · · )
| {z } | {z }
λ∈Par weak
compositions =hλ (x) =y(n1 ,n2 ,...)
(n1 ,n2 ,...) (since (n1 ,n2 ,...)∈S(∞) λ)
satisfying
(n1 ,n2 ,...)∈S(∞) λ
X X
= hλ (x) y(n1 ,n2 ,...)
λ∈Par weak
compositions
(n1 ,n2 ,...)
satisfying
(n1 ,n2 ,...)∈S(∞) λ
| {z }
=mλ (y)
X
= hλ (x)mλ (y).
λ∈Par

For the second expansion (and for later use in the proof of Theorem 4.9.5) note that
∞ ∞ ∞ X ∞ ∞
Y X X (xi t)m X 1
(2.5.12) log H(t) = log (1 − xi t)−1 = − log(1 − xi t) = = pm (x)tm
i=1 i=1 i=1 m=1
m m=1
m
d
so that taking dt then shows that
X H 0 (t)
(2.5.13) P (t) := pm+1 tm = = H 0 (t)E(−t).
H(t)
m≥0

A similar calculation shows that

∞ ∞
Y X 1
(2.5.14) log (1 − xi yj )−1 = pm (x)pm (y)
i,j=1 m=1
m

and hence
∞ ∞ ∞
!  
Y
−1
X 1 Y 1
(1 − xi yj ) = exp pm (x)pm (y) = exp pm (x)pm (y)
i,j=1 m=1
m m=1
m
∞ X ∞   k ∞  km !
Y 1 1 X Y 1 1
= pm (x)pm (y) = pm (x)pm (y)
m=1
k! m m=1
km ! m
k=0 weak compositions
(k1 ,k2 ,k3 ,...)

(by the product rule)

∞ k Q∞ km Q∞ km
X Y (pm (x)pm (y)) m X
m=1 (pm (x)) m=1 (pm (y))
= = Q∞
m=1
km !mkm m=1 (km !mkm )
weak compositions weak compositions
(k1 ,k2 ,k3 ,...) (k1 ,k2 ,k3 ,...)

X p(1k1 2k2 3k3 ··· ) (x) p(1k1 2k2 3k3 ··· ) (y) X pλ (x)pλ (y)
= =
z(1k1 2k2 3k3 ··· ) zλ
weak compositions λ∈Par
(k1 ,k2 ,k3 ,...)

It is known that two permutations in Sn have the same cycle type if and only if they are conjugate. Thus, for a given
partition λ with |λ| = n, any two permutations in Sn having cycle type λ are conjugate and therefore their Sn -centralizer
subgroups have the same size.
 HOPF ALGEBRAS IN COMBINATORICS 57

due to the fact that every partition can be uniquely written in the form 1k1 2k2 3k3 · · · with (k1 , k2 , k3 , . . .)
a weak composition. 

Corollary 2.5.17. (a) With respect to the Hall inner product on Λ, one also has dual bases {hλ } and
{mλ }.
(b) If Q is a subring of k, then {pλ } and zλ−1 pλ are also dual bases with respect to the Hall inner


product on Λ.  
pλ
(c) If R is a subring of k, then √ is an orthonormal basis of Λ with respect to the Hall inner
zλ
product.

Proof. Since (2.5.1) and (2.5.11) showed

∞
Y X X X X pλ (x) pλ (y)
(1 − xi yj )−1 = sλ (x)sλ (y) = hλ (x)mλ (y) = pλ (x)zλ−1 pλ (y) = √ √
i,j=1
zλ zλ
λ∈Par λ∈Par λ∈Par λ∈Par
120
it suffices to show that any pair of graded bases {uλ }, {vλ } having
X X
sλ (x)sλ (y) = uλ (x)vλ (y)
λ∈Par λ∈Par

will be dual with respect to (·, ·). Write transition matrices A = (aν,λ )(ν,λ)∈Par × Par and B = (bν,λ )(ν,λ)∈Par × Par
uniquely expressing
X
(2.5.15) uλ = aν,λ sν ,
ν
X
(2.5.16) vλ = bν,λ sν .
ν
F
Recall that Par = r∈N Parr . Hence, we can view A as a block matrix, where the blocks are indexed by pairs
of nonnegative integers, and the (r, s)-th block is (aν,λ )(ν,λ)∈Parr × Pars . For reasons of homogeneity121, we
have aν,λ = 0 for any (ν, λ) ∈ Par2 satisfying |ν| =6 |λ|. Therefore, the (r, s)-th block of A is zero whenever
r 6= s. In other words, the block matrix A is block-diagonal. Similarly, B can be viewed as a block-diagonal
matrix. The diagonal blocks of A and B are finite square matrices (since Parr is a finite set for each r ∈ N);
therefore, products such as At B, B t A and AB t are well-defined (since all sums involved in their definition
have only finitely many nonzero addends) and subject to the law of associativity. Moreover, the matrix A is
invertible (being a transition matrix between two bases), and its inverse is again block-diagonal (because A
is block-diagonal). P
The equalities (2.5.15) and P(2.5.16) show that (uα , vβ ) = ν aν,α bν,β (by the orthonormality of the sλ ).
Hence, we want to prove that ν aν,α bν,β = δα,β . In other words, we want to prove that At B = I, that is,
B −1 = At . On the other hand, one has
X X XX X
sλ (x)sλ (y) = uλ (x)vλ (y) = aν,λ sν (x) bρ,λ sρ (y).
λ λ λ ν ρ
122
of sν (x)sρ (y) forces λ aν,λ bρ,λ = δν,ρ , or in other words, AB t = I. Since A is
P
Comparing coefficients
invertible, this yields B A = I, and hence At B = I, as desired.123
t

120A basis {w } of Λ indexed by the partitions λ is said to be graded if every n ∈ N and λ ∈ Par satisfy w ∈ Λ .
λ n λ n
121More precisely: The power series u is homogeneous of degree |λ|, and the power series s is homogeneous of degree |ν|.
λ ν
122Comparing coefficients is legitimate because if a power series f ∈ k [[x, y]] is written in the form f =
P
(ν,ρ)∈Par2 aρ,ν sν (x) sρ (y) for some coefficients aρ,ν ∈ k, then these coefficients aρ,ν are uniquely determined by f . This is
just a restatement of Exercise 2.5.5(b).
123In our argument above, we have obtained the invertibility of A from the fact that A is a transition matrix between two
bases. Here is an alternative way to prove that A is invertible:
Recall that A and B t are block-diagonal matrices. Hence, the equality AB t = I rewrites as Ar,r B t r,r = I for all r ∈ N,

where we are using the notation Cr,s for the (r, s)-th block of a block matrix C. But this shows that each diagonal block Ar,r
of A is right-invertible. Therefore, each diagonal block Ar,r of A is invertible (because Ar,r is a square matrix of finite size, and
such matrices are always invertible when they are right-invertible). Consequently, the block-diagonal matrix A is invertible,
and its inverse is again a block-diagonal matrix (whose diagonal blocks are the inverses of the Ar,r ).
58 DARIJ GRINBERG AND VICTOR REINER

Corollary 2.5.17 is a known and fundamental fact124. However, our definition of the Hall inner product
is unusual; most authors (e.g., Macdonald in [125, §I.4, (4.5)], Hazewinkel/Gubareni/Kirichenko in [78, Def.
4.1.21], and Stanley in [183, (7.30)]) define the Hall inner product as the bilinear form satisfying (hλ , mµ ) =
δλ,µ (or, alternatively, (mλ , hµ ) = δλ,µ ), and only later prove that the basis {sλ } is orthonormal with respect
to this scalar product. (Of course, the fact that this definition is equivalent to our Definition 2.5.12 follows
either from this orthonormality, or from our Corollary 2.5.17(a).)
The tactic applied in the proof of Corollary 2.5.17 can not only be used to show that certain bases of Λ
are dual, but also, with a little help from linear algebra over rings (Exercise 2.5.18), it can be strengthened
to show that certain families of symmetric functions are bases to begin with, as we will see in Exercise 2.5.19.

Exercise 2.5.18. (a) Prove that if an endomorphism of a finitely generated k-module is surjective, then
this endomorphism is a k-module isomorphism.
(b) Let A be a finite free k-module with finite basis (γi )i∈I . Let (βi )i∈I be a family of elements of A
which spans the k-module A. Prove that (βi )i∈I is a k-basis of A.

Exercise 2.5.19. (a) Assume that for every partition λ, two homogeneous elements uλ and vλ of Λ,
both having degree |λ|, are given. Assume further that
X X
sλ (x) sλ (y) = uλ (x) vλ (y)
λ∈Par λ∈Par

in k [[x, y]] = k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . .]]. Show that (uλ )λ∈Par and (vλ )λ∈Par are k-bases of Λ,
and actually are dual bases with respect to the Hall inner product on Λ.
(b) Use this to give a new proof of the fact that (hλ )λ∈Par is a k-basis of Λ.

H 0 (t)
pm+1 tm =
P
Exercise 2.5.20. Prove that m≥0 . (This was proven in (2.5.13) in the case when Q is
H (t)
a subring of k, but here we make no requirements on k.)

2.6. Bialternants, Littlewood-Richardson: Stembridge’s concise proof. There is a more natural

way in which Schur functions arise as a k-basis for Λ, coming from consideration of polynomials in a finite
variable set, and the relation between those which are symmetric and those which are alternating.
For the remainder of this section, fix a nonnegative
P cont(T )integer n, and let x = (x1 , . . . , xn ) be a finite variable
set. This means that sλ/µ = sλ/µ (x) = T x is a generating function for column-strict tableaux
125
T as in Definition 2.3.1, but with the extra condition that T have entries in {1, 2, . . . , n}. As a
consequence, sλ/µ is a polynomial in k [x1 , x2 , . . . , xn ] (not just a power series), since there are only finitely
many column-strict tableaux T of shape λ/µ having all their entries in {1, 2, . . . , n}. We will assume without
further mention that all partitions appearing in the section have at most n parts.

Definition 2.6.1. Let k be the ring Z or a field of characteristic not equal to 2. (We require this to avoid
certain annoyances in the discussion of alternating polynomials in characteristic 2.)
Say that a polynomial f (x) = f (x1 , . . . , xn ) is alternating if for every permutation w in Sn one has that
(wf )(x) = f (xw(1) , . . . , xw(n) ) = sgn(w)f (x).

Let Λsgn ⊂ k[x1 , . . . , xn ] denote the subset of alternating polynomials126.

As with Λ andPits monomial basis {mλ }, there is an obvious k-basis for Λsgn , coming from the fact that a
polynomial f = α cα xα is alternating if and only if cw(α) = sgn(w)cα for every w in Sn and every α ∈ Nn .
This means that every alternating f is a k-linear combination of the following elements.

124
For example, Corollary 2.5.17(a) appears in [110, Corollary 3.3] (though the definition of Schur functions in [110] is
different from ours; we will meet this alternative definition later on), and parts (b) and (c) of Corollary 2.5.17 are equivalent to
[125, §I.4, (4.7)] (though Macdonald defines the Hall inner product using Corollary 2.5.17(a)).
125See Exercise 2.3.8(a) for this.
126When k has characteristic 2 (or, more generally, is an arbitrary commutative ring), it is probably best to define the
alternating polynomials Λsgn
k as the k-submodule Λsgn ⊗Z k of Z[x1 , . . . , xn ] ⊗Z k ∼
= k[x1 , . . . , xn ].
 HOPF ALGEBRAS IN COMBINATORICS 59

Definition 2.6.2. For α = (α1 , . . . , αn ) in Nn , define the alternant

 α1
xα

x1 ··· 1
n

xα 1
··· xα n
X  2 2 
aα := sgn(w)w(xα ) = det  . .. ..  .
 .. . . 
w∈Sn
xα
n
1
··· xα
n
n

Example 2.6.3. One has

a(1,5,0) = x11 x52 x03 − x51 x12 x03 − x11 x02 x53 − x01 x52 x13 + x01 x12 x53 + x51 x02 x13 = −a(5,1,0) .
Similarly, aw(α) = sgn(w)aα for every w ∈ Sn and every α ∈ Nn .
Meanwhile, a(5,2,2) = 0 since the transposition t = 123


132 fixes (5, 2, 2) and hence
a(5,2,2) = t(a(5,2,2) ) = sgn(t)a(5,2,2) = −a(5,2,2) .
127
Alternatively, a(5,2,2) = 0 as it is a determinant of a matrix with two equal columns. Similarly, aα = 0
for every n-tuple α ∈ Nn having two equal entries.
This example illustrates that, for a k-basis for Λsgn , one can restrict attention to alternants aα in which α is
a strict partition, i.e., in which α satisfies α1 > α2 > · · · > αn . One can therefore uniquely express α = λ + ρ,
where λ is a (weak) partition λ1 ≥ · · · ≥ λn ≥ 0 and where ρ := (n − 1, n − 2, . . . , 2, 1, 0) is sometimes called
the staircase partition128. For example α = (5, 1, 0) = (3, 0, 0) + (2, 1, 0) = λ + ρ.

Proposition 2.6.4. Let k be the ring Z or a field of characteristic not equal to 2.

The alternants {aλ+ρ } as λ runs through the partitions with at most n parts form a k-basis for Λsgn .
a
In addition, the bialternants { λ+ρ
aρ } as λ runs through the same set form a k-basis for Λ(x1 , . . . , xn ) =
k[x1 , . . . , xn ]Sn .

Proof. The first assertion should be clear from our previous discussion: the alternants {aλ+ρ } span Λsgn by
definition, and they are k-linearly independent because they are supported on disjoint sets of monomials xα .
The second assertion follows from the first, after proving the following Claim: f (x) lies in Λsgn if and
only if f (x) = aρ · g(x) where g(x) lies in k[x]Sn and where
Y
aρ = det(xin−j )i,j=1,2,...,n = (xi − xj )
1≤i<j≤n

is the Vandermonde determinant/product. In other words

Λsgn = aρ · k[x]Sn
is a free k[x]Sn -module of rank one, with aρ as its k[x]Sn -basis element.
To see the Claim, first note the inclusion
Λsgn ⊃ aρ · k[x]Sn
since the product of a symmetric polynomial and an alternating polynomial is an alternating polynomial. For
the reverse inclusion, note that since an alternating polynomial f (x) changes sign whenever one exchanges
two distinct variables xi , xj , it must vanish upon setting xi = xj , and therefore be divisible by xi − xj , so
divisible by the entire product 1≤i<j≤n (xi − xj ) = aρ . But then the quotient g(x) = fa(x)
Q
ρ
is symmetric, as
it is a quotient of two alternating polynomials. 
127One subtlety should be addressed: We want to prove that a
(5,2,2) = 0 in k [x1 , . . . , xn ] for every commutative ring k. It
is clearly enough to prove that a(5,2,2) = 0 in Z [x1 , . . . , xn ]. Since 2 is not a zero-divisor in Z [x1 , . . . , xn ], we can achieve this
by showing that a(5,2,2) = −a(5,2,2) . We would not be able to make this argument directly over an arbitrary commutative ring
k.
128The name is owed to its Ferrers shape. For instance, if n = 5, then the Ferrers diagram of ρ (represented using dots) has
the form
• • • •
• • •
.
• •
•
60 DARIJ GRINBERG AND VICTOR REINER

Let us now return to the general setting, where k is an arbitrary commutative ring. We are not requiring
that the assumptions of Proposition 2.6.4 be valid; we can still study the aα of Definition 2.6.2, but we
a
cannot use Proposition 2.6.4 anymore. We will show that the fraction λ+ρ aρ is nevertheless a well-defined
129
polynomial in Λ (x1 , . . . , xn ) whenever λ is a partition , and in fact equals the Schur function sλ (x). As
a
a consequence, the mysterious bialternant basis { λ+ρ aρ } of Λ (x1 , . . . , xn ) defined in Proposition 2.6.4 still
exists in the general setting, and is plainly the Schur functions {sλ (x)}. Stembridge [186] noted that one
could give a remarkably concise proof of an even stronger assertion, which simultaneously gives one of the
standard combinatorial interpretations for the Littlewood-Richardson coefficients cλµ,ν . For the purposes of
stating it, we introduce for a tableau T the notation T |cols≥j (resp. T |cols≤j ) to indicate the subtableau
which is the restriction of T to the union of its columns j, j + 1, j + 2, . . . (resp. columns 1, 2, . . . , j).
Theorem 2.6.5. For partitions λ, µ, ν with µ ⊆ λ, one has130
X
aν+ρ sλ/µ = aν+cont(T )+ρ
T
where T runs through all column-strict tableaux with entries in {1, 2, . . . , n} of shape λ/µ with the property
that for all j = 1, 2, . . . one has ν + cont(T |cols≥j ) a partition.
Before proving Theorem 2.6.5, let us see some of its consequences.
Corollary 2.6.6. For any partition λ, we have131
aλ+ρ
sλ (x) = .
aρ
Proof. Take ν = µ = ∅ in Theorem 2.6.5. Note that for any λ, there is only one column-strict tableau T of
shape λ having each cont(T |cols≥j ) a partition, namely the one having every entry in row i equal to i:
1 1 1 1 1
2 2 2
3 3 3
4 4
Furthermore, this T has cont(T ) = λ, so the theorem says aρ sλ = aλ+ρ . 
Example 2.6.7. For n = 2, so that ρ = (1, 0), if we take λ = (4, 2), then one has
aλ+ρ a(4,2)+(1,0) a(5,2)
= =
aρ a(1,0) a(1,0)
x51 x22 − x21 x52
=
x1 − x2
= x41 x22 + x31 x32 + x21 x42
     
1111 1112 1122
cont  cont  cont 
22 22 22
=x +x +x
= s(4,2) = sλ .
Some authors use the equality in Corollary 2.6.6 to define the Schur polynomial sλ (x1 , x2 , . . . , xn ) in n
variables; this definition, however, has the drawback of not generalizing easily to infinitely many variables
or to skew Schur functions132.
Next divide through by aρ on both sides of Theorem 2.6.5 (and use Corollary 2.6.6) to give the following.
129This can also be deduced by base change from the k = Z case of Proposition 2.6.4.
130Again, we can drop the requirement that µ ⊆ λ, provided that we understand that there are no column-strict tableaux
of shape λ/µ unless µ ⊆ λ.
131Notice that division by a is unambiguous in the ring k [x , . . . , x ], since a is not a zero-divisor (in fact, a =
ρ 1 n ρ ρ
Q
1≤i<j≤n (xi − xj ) is the product of the binomials xi − xj , none of which is a zero-divisor).
132With some effort, it is possible to use Corollary 2.6.6 in order to define the Schur function s in infinitely many variables.
λ
a
Indeed, one can define this Schur function as the unique element of Λ whose evaluation at (x1 , x2 , . . . , xn ) equals λ+ρ
aρ
for every
n ∈ N. If one wants to use such a definition, however, one needs to check that such an element exists. This is the approach to
defining sλ taken in [110, Definition 1.4.2] and in [125, §I.3].
 HOPF ALGEBRAS IN COMBINATORICS 61

Corollary 2.6.8. For partitions λ, µ, ν having at most n parts, one has

X
(2.6.1) sν sλ/µ = sν+cont(T )
T

where T runs through the same set as in Theorem 2.6.5. In particular, taking ν = ∅, we obtain
X
(2.6.2) sλ/µ = scont(T )
T

where in the sum T runs through all column-strict tableaux of shape λ/µ for which each cont(T |cols≥j ) is a
partition.

Proof of Theorem 2.6.5. Start by rewriting the left side of the theorem, and using the fact that w(sλ/µ ) =
sλ/µ for any w in Sn :
X
aν+ρ sλ/µ = sgn(w)xw(ν+ρ) w(sλ/µ )
w∈Sn
X X
= sgn(w)xw(ν+ρ) xw(cont(T ))
w∈Sn column-strict T
of shape λ/µ
X X
= sgn(w)xw(ν+cont(T )+ρ)
column-strict T w∈Sn
of shape λ/µ
X
= aν+cont(T )+ρ .
column-strict T
of shape λ/µ

We wish to cancel out all the summands indexed by column-strict tableaux T which fail any of the conditions
that ν + cont(T |cols≥j ) be a partition. Given such a T , find the maximal j for which it fails this condition133,
and then find the minimal k for which
νk + contk (T |cols≥j ) < νk+1 + contk+1 (T |cols≥j ).
Maximality of j forces
νk + contk (T |cols≥j+1 ) ≥ νk+1 + contk+1 (T |cols≥j+1 ).
Since column-strictness implies that column j of T can contain at most one occurrence of k or of k + 1 (or
neither or both), the previous two inequalities imply that column j must contain an occurrence of k + 1 and
no occurrence of k, so that
νk + contk (T |cols≥j ) + 1 = νk+1 + contk+1 (T |cols≥j ).
This implies that the adjacent transposition tk,k+1 swapping k and k+1 fixes the vector ν +cont(T |cols≥j )+ρ.
Now create a new tableau T ∗ from T by applying the Bender-Knuth involution (from the proof of Propo-
sition 2.2.4) on letters k, k + 1, but only to columns 1, 2, . . . , j − 1 of T , leaving columns j, j + 1, j + 2, . . .
unchanged. One should check that T ∗ is still column-strict, but this holds because column j of T has no
occurrences of letter k. Note that
tk,k+1 cont(T |cols≤j−1 ) = cont(T ∗ |cols≤j−1 )
and hence
tk,k+1 (ν + cont(T ) + ρ) = ν + cont(T ∗ ) + ρ
so that aν+cont(T )+ρ = −aν+cont(T ∗ )+ρ .
Because T, T ∗ have exactly the same columns j, j + 1, j + 2, . . ., the tableau T ∗ is also a violator of at
least one of the conditions that ν + cont(T ∗ |cols≥j ) be a partition, and has the same choice of maximal j
and minimal k as did T . Hence the map T 7→ T ∗ is an involution on the violators that lets one cancel their
summands aν+cont(T )+ρ and aν+cont(T ∗ )+ρ in pairs. 

133Such a j exists because ν + cont(T |

cols≥j ) is a partition for all sufficiently high j (in fact, ν itself is a partition).
62 DARIJ GRINBERG AND VICTOR REINER

So far (in this section) we have worked with a finite set of variables x1 , x2 , . . . , xn (where n is a fixed nonneg-
ative integer) and with partitions having at most n parts. We now drop these conventions and restrictions;
thus, partitions again mean arbitrary partitions, and x again means the infinite family (x1 , x2 , x3 , . . .) of
variables. In this setting, we have the following analogue of Corollary 2.6.8:

Corollary 2.6.9. For partitions λ, µ, ν (of any lengths), one has

X
(2.6.3) sν sλ/µ = sν+cont(T )
T

where T runs through all column-strict tableaux of shape λ/µ with the property that for all j = 1, 2, . . . one
has ν + cont(T |cols≥j ) a partition. In particular, taking ν = ∅, we obtain
X
(2.6.4) sλ/µ = scont(T )
T

where in the sum T runs through all column-strict tableaux of shape λ/µ for which each cont(T |cols≥j ) is a
partition.

Proof of Corollary 2.6.9. Essentially, Corollary 2.6.9 is obtained from Corollary 2.6.8 by “letting n (that is,
the number of variables) tend to ∞”. This can be formalized in different ways: One way is to endow the
ring of power series k [[x]] = k [[x1 , x2 , x3 , . . .]] with the coefficientwise topology134, and to show that the left
hand side of (2.6.1) tends to the left hand side of (2.6.3) when n → ∞, and the same holds for the right
hand sides. A different approach proceeds by regarding Λ as the inverse limit of the Λ (x1 , x2 , . . . , xn ). 

Comparing coefficients of a given Schur function sν in (2.6.4), we obtain the following version of the
Littlewood-Richardson rule.

Corollary 2.6.10. For partitions λ, µ, ν (of any lengths), the Littlewood-Richardson coefficient cλµ,ν counts
column-strict tableaux T of shape λ/µ with cont(T ) = ν having the property that each cont(T |cols≥j ) is a
partition.

2.7. The Pieri and Assaf-McNamara skew Pieri rule. The classical Pieri rule refers to two special
cases of the Littlewood-Richardson rule. To state them, recall that a skew shape is called a horizontal (resp.
vertical) strip if no two of its cells lie in the same column (resp. row). A horizontal (resp. vertical) n-strip
(for n ∈ N) shall mean a horizontal (resp. vertical) strip of size n (that is, having exactly n cells).

Theorem 2.7.1. For every partition λ and any n ∈ N, we have

X
(2.7.1) sλ hn = sλ+
λ+ :λ+ /λ is a
horizontal n-strip
X
(2.7.2) sλ en = sλ+
+ +
λ :λ /λ is a
vertical n-strip

134This topology is defined as follows:

We endow the ring k with the discrete topology. Then, we can regard the k-module k [[x]] as a direct product of infinitely
many copies of k (by identifying every power series in k [[x]] with the family of its coefficients). Hence, the product topology
is a well-defined topology on k [[x]]; this topology is denoted as the coefficientwise topology. Its name is due to the fact that a
sequence (an )n∈N of power series converges to a power series a with respect to this topology if and only if for every monomial
m, all sufficiently high n ∈ N satisfy

(the coefficient of m in an ) = (the coefficient of m in a) .

 HOPF ALGEBRAS IN COMBINATORICS 63

Example 2.7.2.
s h
    
•
 
 

s s s
         
=   +    +  
     
   

s s
        
+ +
    
   
Proof of Theorem 2.7.1. For the first Pieri formula involving hn , as hn = s(n) one has
X +
sλ hn = cλλ,(n) sλ+ .
λ+
+
Corollary 2.6.10 says cλλ,(n) counts column-strict tableaux T of shape λ+ /λ having cont(T ) = (n) (i.e. all
entries of T are 1’s), with an extra condition. Since its entries are all equal, such a T must certainly have
shape being a horizontal strip, and more precisely a horizontal n-strip (since it has n cells). Conversely,
for any horizontal n-strip, there is a unique such filling, and it will trivially satisfy the extra condition that
+
cont(T |cols≥j ) is a partition for each j. Hence cλλ,(n) is 1 if λ+ /λ is a horizontal n-strip, and 0 else.
For the second Pieri formula involving en , using en = s(n) one has
X +
sλ en = cλλ,(1n ) sλ+ .
λ+
+
Corollary 2.6.10 says cλλ,(1n ) counts column-strict tableaux T of shape λ+ /λ having cont(T ) = (1n ), so its
entries are 1, 2, . . . , n each occurring once, with the extra condition that 1, 2, . . . , n appear from right to left.
Together with the tableau condition, this forces at most one entry in each row, that is λ+ /λ is a vertical
strip, and then there is a unique way to fill it (maintaining column-strictness and the extra condition that
+
1, 2, . . . , n appear from right to left). Thus cλλ,(1n ) is 1 if λ+ /λ is a vertical n-strip, and 0 else. 

Assaf and McNamara [9] recently proved an elegant generalization.

Theorem 2.7.3. For any partitions λ and µ and any n ∈ N, we have135

X −
sλ/µ hn = (−1)|µ/µ | sλ+ /µ−
λ+ ,µ− :
λ+ /λ a horizontal strip
µ/µ− a vertical strip
|λ+ /λ|+|µ/µ− |=n
X −
sλ/µ en = (−1)|µ/µ | sλ+ /µ−
λ+ ,µ− :
λ+ /λ a vertical strip
µ/µ− a horizontal strip
|λ+ /λ|+|µ/µ− |=n

135Note that µ ⊆ λ is not required. (The left hand sides are 0 otherwise, but this does not trivialize the equalities.)
64 DARIJ GRINBERG AND VICTOR REINER

Example 2.7.4.
s h
   
•

 

s s s
      
=  +   + 
     
   

s s
      
+ +
  
   

s
s s
 
    
−   − −
    
 
   


s
  
+
 
 
Theorem 2.7.3 is proven in the next section, using an important Hopf algebra tool.
Exercise 2.7.5. Let λ = (λ1 , λ2 , λ3 , . . .) and µ = (µ1 , µ2 , µ3 , . . .) be two partitions such that µ ⊆ λ.
136
(a) Show that λ/µ is a horizontal strip if and only if every i ∈ {1, 2, 3, . . .} satisfies µi ≥ λi+1 .
(b) Show that λ/µ is a vertical strip if and only if every i ∈ {1, 2, 3, . . .} satisfies λi ≤ µi + 1.


Exercise 2.7.6. (a) Let λ and µ be two partitions such that µ ⊆ λ. Let n ∈ N. Show that hn , sλ/µ
equals 1 if λ/µ is a horizontal n-strip, and equals 0 otherwise.
(b) Use part (a) to give a new proof of (2.7.1).
Exercise 2.7.7. Prove Theorem 2.7.1 again using the ideas of the proof of Theorem 2.5.1.
Exercise 2.7.8. Let A be a commutative ring, and n ∈ N.
(a) Let a1 , a2 , . . . , an be n elements of A. Let b1 , b2 , . . . , bn be n further elements of A. If ai − bj is an
invertible element of A for every i ∈ {1, 2, . . . , n} and j ∈ {1, 2, . . . , n}, then prove that
! Q
1≤j<i≤n ((ai − aj ) (bj − bi ))
 
1
det = Q .
ai − bj i,j=1,2,...,n (i,j)∈{1,2,...,n}2 (ai − bj )

(b) Let a1 , a2 , . . . , an be n elements of A. Let b1 , b2 , . . . , bn be n further elements of A. If 1 − ai bj is an

invertible element of A for every i ∈ {1, 2, . . . , n} and j ∈ {1, 2, . . . , n}, then prove that
! Q
1≤j<i≤n ((ai − aj ) (bi − bj ))
 
1
det = Q .
1 − ai bj i,j=1,2,...,n (i,j)∈{1,2,...,n}2 (1 − ai bj )

136In other words, λ/µ is a horizontal strip if and only if (λ , λ , λ , . . .) ⊆ µ. This simple observation has been used by
2 3 4
Pak and Postnikov [145, §10] for a new approach to RSK-type algorithms.
 HOPF ALGEBRAS IN COMBINATORICS 65

(c) Use the result of part (b) to give a new proof for Theorem [Link]
The determinant on the left hand side of Exercise 2.7.8(a) is known as the Cauchy determinant.
Exercise 2.7.9. Prove that s(a,b) = ha hb − ha+1 hb−1 for any two integers a ≥ b ≥ 0 (where we set h−1 = 0
as usual).
(Note that this is precisely the Jacobi-Trudi formula (2.4.9) in the case when λ = (a, b) is a partition with
at most two entries and µ = ∅.)
Exercise 2.7.10. If λ is a partition and µ is a weak composition, let Kλ,µ denote the number of column-strict
tableaux T of shape λ having cont (T ) = µ. (This Kλ,µ is called the (λ, µ)-Kostka number.)
P
(a) Use Theorem 2.7.1 to show that every partition µ satisfies hµ = λ Kλ,µ sλ , where the sum ranges
over all partitions λ.
(b) Use this to give a new proof for Theorem [Link]
(c) Give a new proof of the fact that (hλ )λ∈Par is a k-basis of Λ.
Exercise 2.7.11. (a) Define a k-linear map Z : Λ → Λ by having it send sλ to sλt for every partition
λ. (This is clearly well-defined, since (sλ )λ∈Par is a k-basis of Λ.) Show that
Z (f hn ) = Z (f ) · Z (hn ) for every f ∈ Λ and every n ∈ N.
(b) Show that Z = ω.
t
(c) Show that cλµ,ν = cλµt ,ν t for any three partitions λ, µ and ν.
(d) Use this to prove (2.4.8).139
Exercise 2.7.12. (a) Show that
∞
Y X X
(1 + xi yj ) = sλ (x) sλt (y) = eλ (x) mλ (y)
i,j=1 λ∈Par λ∈Par

in the power series ring k [[x, y]] = k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . .]].

(b) Assume that Q is a subring of k. Show that
∞
|λ|−`(λ) −1
Y X
(1 + xi yj ) = (−1) zλ pλ (x) pλ (y)
i,j=1 λ∈Par

in the power series ring k [[x, y]] = k [[x1 , x2 , x3 , . . . , y1 , y2 , y3 , . . .]], where zλ is defined as in Propo-
sition 2.5.15.
The first equality of Exercise 2.7.12(a) appears in [183, Thm. 7.14.3], [165, Thm. 4.8.6] and several other
references under the name of the dual Cauchy identity, and is commonly proven using a “dual” analogue of
the Robinson-Schensted-Knuth algorithm.
Exercise 2.7.13. Prove Theorem 2.4.3.
[Hint:140 Switch x and y in the formula of Exercise 2.5.11(a), and specialize the resulting equality by
replacing y by a finite set of variables (y1 , y2 , . . . , y` ); then, set n = ` and ρ = (n − 1, n − 2, . . . , 0), and
multiply with the alternant aρ (y1 , y2 , . . . , y` ), using Corollary 2.6.6 to simplify the result; finally, extract the
coefficient of yλ+ρ .]
2.8. Skewing and Lam’s proof of the skew Pieri rule. We codify here the operation s⊥ µ of skewing by
sµ , acting on Schur functions via
s⊥
µ (sλ ) = sλ/µ
(where, as before, one defines sλ/µ = 0 if µ 6⊆ λ). These operations play a crucial role
• in Lam’s proof of the skew Pieri rule,
137This approach to Theorem 2.5.1 is taken in [41, §4] (except that [41] only works with finitely many variables).
138Of course, this gives a new proof of Theorem 2.5.1 only when coupled with a proof of Theorem 2.7.1 which does not rely
on Theorem 2.5.1. The proof of Theorem 2.7.1 we gave in the text above did not rely on Theorem 2.5.1, whereas the proof of
(2.7.1) given in Exercise 2.7.6(b) did.
139The first author learned this approach to (2.4.8) from Alexander Postnikov.
140This is the proof given in Stanley [183, §7.16, Second Proof of Thm. 7.16.1] and Macdonald [125, proof of (5.4)].
66 DARIJ GRINBERG AND VICTOR REINER

• in Lam, Lauve, and Sottile’s proof [105] of a more general skew Littlewood-Richardson rule that had
been conjectured by Assaf and McNamara, and
• in Zelevinsky’s structure theory of PSH’s to be developed in the next chapter.
We are going to define them in the general setting of any graded Hopf algebra.
Definition 2.8.1. Given a graded Hopf algebra A, and its (graded) dual Ao , let (·, ·) = (·, ·)A : Ao × A → k
be the pairing defined by (f, a) := f (a) for f in Ao and a in A. Then define for each f in Ao an operator
f⊥
A → A as follows141: for a in A with ∆(a) = a1 ⊗ a2 , let
P
X
f ⊥ (a) = (f, a1 )a2 .

In other words, f ⊥ is the composition

f ⊗id ∼
A
∆ / A⊗A / k⊗A = / A,

where the rightmost arrow is the canonical isomorphism k ⊗ A → A.

Now, recall that the Hall inner product induces an isomorphism Λo =∼ Λ (by Corollary 2.5.14). Hence, we
can regard any element f ∈ Λ as an element of Λo ; this allows us to define an operator f ⊥ : Λ → Λ for
each f ∈ Λ (by regarding f as an element of Λo , and applying Definition 2.8.1 to A = Λ). Explicitly, this
operator is given by
X X
(2.8.1) f ⊥ (a) = (f, a1 )a2 whenever ∆(a) = a1 ⊗ a2 ,

where the inner product (f, a1 ) is now understood as a Hall inner product.
Recall that each partition λ satisfies
X X X
∆sλ = sµ ⊗ sλ/µ = sν ⊗ sλ/ν = sν ⊗ sλ/ν
µ⊆λ ν⊆λ ν

(since sλ/ν = 0 unless ν ⊆ λ). Hence, for any two partitions λ and µ, we have
X
s⊥
µ (sλ ) = (sµ , sν ) sλ/ν (by (2.8.1), applied to f = sµ and a = sλ )
ν
| {z }
=δµ,ν
X
(2.8.2) = δµ,ν sλ/ν = sλ/µ .
ν

Thus, skewing acts on the Schur functions exactly as desired.

Proposition 2.8.2. Let A be a graded Hopf algebra. The f ⊥ operators A → A have the following properties.
f·
(i) For every f ∈ Ao , the map f ⊥ is adjoint to left multiplication Ao → Ao in the sense that
(g, f ⊥ (a)) = (f g, a).
(ii) For every f, g ∈ Ao , we have (f g)⊥ (a) = g ⊥ (f ⊥ (a)), that is, A becomes a right Ao -module via the
f ⊥ action.142
(iii) Assume that A isPof finite type (so Ao becomes a Hopf algebra, not just an algebra). If an f ∈ Ao
satisfies ∆(f ) = f1 ⊗ f2 , then
X
f ⊥ (ab) = f1⊥ (a)f2⊥ (b).

In particular, if f is primitive in Ao , so that ∆(f ) = f ⊗ 1 + 1 ⊗ f , then f ⊥ is a derivation:

f ⊥ (ab) = f ⊥ (a) · b + a · f ⊥ (b).

141This f ⊥ (a) is called a ( f in Montgomery [139, Example 1.6.5].

142This makes sense, since Ao is a k-algebra (by Exercise 1.6.1(c), applied to C = A).
 HOPF ALGEBRAS IN COMBINATORICS 67

Proof. For (i), note that

X
(g, f ⊥ (a)) = (f, a1 )(g, a2 ) = (f ⊗ g, ∆A (a)) = (mAo (f ⊗ g), a) = (f g, a).
For (ii), using (i) and considering any h in Ao , one has that
(h, (f g)⊥ (a)) = (f gh, a) = (gh, f ⊥ (a)) = (h, g ⊥ (f ⊥ (a))).
For (iii), noting that
  
X X X
∆(ab) = ∆(a)∆(b) =  a1 ⊗ a2   b1 ⊗ b2  = a1 b1 ⊗ a2 b2 ,
(a) (b) (a),(b)

one has that X X

f ⊥ (ab) = (f, a1 b1 )A a2 b2 = (∆(f ), a1 ⊗ b1 )A⊗A a2 b2
(a),(b) (a),(b)
X
= (f1 , a1 )A (f2 , b1 )A a2 b2
(f ),(a),(b)
  
X X X X
=  (f1 , a1 )A a2   (f2 , b1 )A b2  = f1⊥ (a)f2⊥ (b).
(f ) (a) (b) (f )

⊥
The following interaction between multiplication and h is the key to deducing the skew Pieri formula from
the usual Pieri formulas.
Lemma 2.8.3. For any f, g in Λ and any n ∈ N, one has
n
X
f · h⊥
n (g) = (−1)k h⊥ ⊥
n−k (ek (f ) · g).
k=0

Proof. Starting with the right side, first apply Proposition 2.8.2(iii):
Xn
(−1)k h⊥ ⊥
n−k (ek (f ) · g)
k=0
| {z }
⊥ ⊥ ⊥
= n−k
P
j=0 hj (ek (f ))·hn−k−j (g)
(by Proposition 2.8.2(iii), applied
to hn−k , e⊥
k (f ) and g instead of f , a and b)

n
X n−k
X
= (−1)k h⊥ ⊥ ⊥
j (ek (f )) · hn−k−j (g)
k=0 j=0
 
n
X Xn−i
= (−1)n−i  (−1)j h⊥ ⊥
j (en−i−j (f ))
 · h⊥
i (g) (reindexing i := n − k − j )
i=0 j=0
 ⊥
n
X Xn−i
= (−1)n−i  (−1)j en−i−j hj  (f ) · h⊥
i (g) (by Proposition 2.8.2(ii) )
i=0 j=0

= 1⊥ (f ) · h⊥ ⊥
n (g) = f · hn (g)

where the second-to-last equality used (2.4.4). 

Proof of Theorem 2.7.3. We prove the first skew Pieri rule; the second is analogous, swapping hi ↔ ei and
swapping the words “vertical” ↔ “horizontal”. For any f ∈ Λ, we have
sλ/µ , f = s⊥



µ (sλ ), f (by (2.8.2))
= f, s⊥


µ (sλ ) (by symmetry of (·, ·)Λ )
= (sµ f, sλ ) (by Proposition 2.8.2(i))
(2.8.3) = (sλ , sµ f ) (by symmetry of (·, ·)Λ ) .
68 DARIJ GRINBERG AND VICTOR REINER

Hence for any g in Λ, one can compute that

P rop. (2.8.3)
(hn sλ/µ , g) = (sλ/µ , h⊥
n g) = (sλ , sµ · h⊥
n g)
2.8.2(i)
n
Lemma
X
(2.8.4) = (−1)k (sλ , h⊥ ⊥
n−k (ek (sµ ) · g))
2.8.3
k=0
n
P rop. X
= (−1)k (hn−k sλ , e⊥
k (sµ ) · g)
2.8.2(i)
k=0
P
The first Pieri rule in Theorem 2.7.1 lets one rewrite hn−k sλ = λ+ sλ+ , with the sum running through
λ+ for which λ+ /λ is a horizontal (n − k)-strip. The second Pieri rule in Theorem 2.7.1 lets one rewrite
⊥
ek sµ = µ− sµ− , with the sum running through µ− for which µ/µ− is a vertical k-strip, since (sµ− , e⊥
P
k sµ ) =
(ek sµ− , sµ ). Thus the last line of (2.8.4) becomes
   
n n
X X X (2.8.3) X X
(−1)k  sλ+ , sµ− · g  =  (−1)k sλ+ /µ− , g 
k=0 λ+ µ− k=0 (λ+ ,µ− )

where the sum is over the pairs (λ+ , µ− ) for which λ+ /λ is a horizontal (n − k)-strip and µ/µ− is a vertical
k-strip. 

Exercise 2.8.4. Let n ∈ N.

(a) For every k ∈ N, let p (n, k) denote the number of partitions of n of length k. Let c (n) denote the
number of self-conjugate partitions of n (that is, partitions λ of n satisfying λt = λ). Show that
n
X
n k
(−1) c (n) = (−1) p (n, k) .
k=0

(This application of Hopf algebras was found by Aguiar and Lauve, [5, §5.1]. See also [183, Chapter
1, Exercise 22(b)] for an elementary proof.)
(b) For every partition λ, let C (λ) denote the number of corner cells of the Ferrers diagram of λ (these
are the cells of the Ferrers diagram whose neighbors to the east and to the south both lie outside of
the Ferrers diagram). For every partition λ, let µ1 (λ) denote the number of parts of λ equal to 1.
Show that
X X
C (λ) = µ1 (λ) .
λ∈Parn λ∈Parn

(This is also due to Stanley.)

Exercise 2.8.5. The goal of this exercise is to prove (2.4.8) using the skewing operators that we have
developed. Recall the involution ω : Λ → Λ defined in (2.4.6).
|λ|−`(λ)
(a) Show that ω (pλ ) = (−1) pλ for any λ ∈ Par, where ` (λ) denotes the length of the partition
λ.
(b) Show that ω is an isometry.
(c) Show that this same map ω : Λ → Λ is a Hopf automorphism.
⊥
(d) Prove that ω a⊥ b = (ω (a)) (ω (b)) for every a ∈ Λ and b ∈ Λ.

(e) For any partition λ = (λ1 , . . . , λ` ) with length ` (λ) = `, prove that

e⊥
` sλ = s(λ1 −1,λ2 −1,...,λ` −1) .

(f) For any partition λ = (λ1 , λ2 , . . .), prove that

h⊥
λ1 sλ = s(λ2 ,λ3 ,λ4 ,...) .

(g) Prove (2.4.8).

 HOPF ALGEBRAS IN COMBINATORICS 69

2.9. Assorted exercises on symmetric functions. Over a hundred exercises on symmetric functions are
collected in Stanley’s [183, chapter 7], and even more (but without any hints or references) on his website143.
Further sources for results related to symmetric functions are Macdonald’s work, including his monograph
[125] and his expository [126]. In this section, we gather a few exercises that are not too difficult to handle
with the material given above.
i
(a) Let m ∈ Z. Prove that, for every f ∈ Λ, the infinite sum i∈N (−1) hm+i e⊥
P
Exercise 2.9.1. i f is
convergent in the discrete topology (i.e., all but finitely many addends of this sum are zero). Hence,
we can define a map Bm : Λ → Λ by setting
X i
Bm (f ) = (−1) hm+i e⊥i f for all f ∈ Λ.
i∈N
Show that this map Bm is k-linear.
(b) Let λ = (λ1 , λ2 , λ3 , . . .) be a partition, and let m ∈ Z be such that m ≥ λ1 . Show that
X i
(−1) hm+i e⊥i sλ = s(m,λ1 ,λ2 ,λ3 ,...) .
i∈N

(c) Let n ∈ N. For every n-tuple (α1 , α2 , . . . , αn ) ∈ Zn , we define an element s(α1 ,α2 ,...,αn ) ∈ Λ by
 
s(α1 ,α2 ,...,αn ) = det (hαi −i+j )i,j=1,2,...,n .
Show that
(2.9.1) sλ = s(λ1 ,λ2 ,...,λn )
for every partition λ = (λ1 , λ2 , λ3 , . . .) having at most n parts (where “part” means “nonzero part”).
Furthermore, show that for every n-tuple (α1 , α2 , . . . , αn ) ∈ Zn , the symmetric function s(α1 ,α2 ,...,αn )
either is 0 or equals ±sν for some partition ν having at most n parts.
Finally, show that for any n-tuples (α1 , α2 , . . . , αn ) ∈ Zn and (β1 , β2 , . . . , βn ) ∈ Nn , we have
 
s⊥


(2.9.2) (β1 ,β2 ,...,βn ) s(α1 ,α2 ,...,αn ) = det hαi −βj −i+j i,j=1,2,...,n .
(d) For every n ∈ N, every m ∈ Z and every n-tuple (α1 , α2 , . . . , αn ) ∈ Zn , prove that
X i
(2.9.3) (−1) hm+i e⊥
i s(α1 ,α2 ,...,αn ) = s(m,α1 ,α2 ,...,αn ) ,
i∈N

where we are using the notations of Exercise 2.9.1(c).

(e) For every n ∈ N and every n-tuple (α1 , α2 , . . . , αn ) ∈ Zn , prove that
s(α1 ,α2 ,...,αn ) = (Bα1 ◦ Bα2 ◦ · · · ◦ Bαn ) (1) ,
where we are using the notations of Exercise 2.9.1(c) and Exercise 2.9.1(a).
Remark 2.9.2. The map Bm defined in Exercise 2.9.1(a) is the so-called m-th Bernstein creation oper-
ator ; it appears in Zelevinsky [203, §4.20(a)] and has been introduced by J.N. Bernstein, who found
the result of Exercise 2.9.1(b). It is called a “Schur row adder” in [61]. Exercise 2.9.1(e) appears in
Berg/Bergeron/Saliola/Serrano/Zabrocki [17, Theorem 2.3], where it is used as a prototype for defining
noncommutative analogues of Schur functions, the so-called immaculate functions. The particular case of
Exercise 2.9.1(e) for (α1 , α2 , . . . , αn ) a partition of length n (a restatement of Exercise 2.9.1(b)) is proven in
[125, §I.5, example 29].
Exercise 2.9.3. (a) Prove that there exists a unique family (xn )n≥1 of elements of Λ such that
∞
−1
Y
H (t) = (1 − xn tn ) .
n=1
Denote this family (xn )n≥1 by (wn )n≥1 . For instance,
w1 = s(1) , w2 = −s(1,1) , w3 = −s(2,1) ,
w4 = −s(1,1,1,1) − s(2,1,1) − s(2,2) − s(3,1) , w5 = −s(2,1,1,1) − s(2,2,1) − s(3,1,1) − s(3,2) − s(4,1) .
143 [Link] rstan/ec/[Link]
~
70 DARIJ GRINBERG AND VICTOR REINER

(b) Show that wn is homogeneous of degree n for every positive integer n.

(c) For every partition λ, define wλ ∈ Λ by wλ = wλ1 wλ2 · · · wλ` (where Pλ = (λ1 , λ2 , . . . , λ` ) with
` = ` (λ)). Notice that wλ is homogeneous of degree |λ|. Prove that λ∈Parn wλ = hn for every
n ∈ N.
(d) Show that {wλ }λ∈Par is a k-basis of Λ. (This basis is called the Witt basis144; it is studied in [75,
§9-§10].145)
P n/d P
(e) Prove that pn = d|n dwd for every positive integer n. (Here, the summation sign d|n means a
sum over all positive divisors d of n.)
(f) We are going to show that −wn is a sum of Schur functions (possibly with repetitions, but without
signs!) for every n ≥ 2. (For n = 1, the opposite is true: w1 is a single Schur function.) This proof
goes back to Doran [48]146. P
For any positive integers n and k, define fn,k ∈ Λ by fn,k = λ∈Parn , wλ , where min λ denotes
min λ≥k
the smallest nonzero part of λ. Show that
k−1
X
−fn,k = s(n−1,1) + fi,i fn−i,i for every n ≥ k ≥ 2.
i=2
Conclude that −fn,k is a sum of Schur functions for every n ∈ N and k ≥ 2. Conclude that −wn is
a sum of Schur functions for every n ≥ 2.

(g) For every partition λ, define rλ ∈ Λ by rλ = i≥1 hvi xi1 , xi2 , xi3 , . . . , where vi is the number of
Q
P Q∞ −1
occurrences of i in λ. Show that λ∈Par wλ (x) rλ (y) = i,j=1 (1 − xi yj ) .
(h) Show that {rλ }λ∈Par and {wλ }λ∈Par are dual bases of Λ.
Exercise 2.9.4. For this exercise, set k = Z, and consider Λ = ΛZ as a subring of ΛQ . Also, consider Λ ⊗Z Λ
147
as a subring of ΛQ ⊗Q ΛQ . Recall that the family (pn )n≥1 generates the Q-algebra ΛQ , but does not
generate the Z-algebra Λ.
(a) Define a Q-linear map Z : ΛQ → ΛQ by setting
Z (pλ ) = zλ pλ for every partition λ,
148
where zλ is defined as in Proposition 2.5.15. Show that Z (Λ) ⊂ Λ.
(b) Define a Q-algebra homomorphism ∆× : ΛQ → ΛQ ⊗Q ΛQ by setting
∆× (pn ) = pn ⊗ pn for every positive integer n.
149
Show that ∆× (Λ) ⊂ Λ ⊗Z Λ.
(c) Let r ∈ Z. Define a Q-algebra homomorphism r : ΛQ → Q by setting
r (pn ) = r for every positive integer n.
150
Show that r (Λ) ⊂ Z.
(d) Let r ∈ Z. Define a Q-algebra homomorphism ir : ΛQ → ΛQ by setting
ir (pn ) = rpn for every positive integer n.

144This is due to its relation with Witt vectors in the appropriate sense. Most of the work on this basis has been done by
Reutenauer and Hazewinkel.
145It also implicitly appears in [12, §5]. Indeed, the q of [12] are our w (for k = R).
n n
146See also Stanley [183, Exercise 7.46].
147Here is how this works: We have Λ ∼ Q ⊗ Λ. But fundamental properties of tensor products yield
Q = Z

(2.9.4) Q ⊗Z (Λ ⊗Z Λ) ∼
= (Q ⊗Z Λ) ⊗Q (Q ⊗Z Λ) ∼
= ΛQ ⊗Q ΛQ
| {z } | {z }
∼
=ΛQ ∼
=ΛQ

as Q-algebras. But Λ ⊗Z Λ is a free Z-module (since Λ is a free Z-module), and so the canonical ring homomorphism Λ ⊗Z Λ →
Q ⊗Z (Λ ⊗Z Λ) sending every u to 1Q ⊗Z u is injective. Composing this ring homomorphism with the Q-algebra isomorphism of
(2.9.4) gives an injective ring homomorphism Λ ⊗Z Λ → ΛQ ⊗Q ΛQ . We use this latter homomorphism to identify Λ ⊗Z Λ with
a subring of ΛQ ⊗Q ΛQ .
148This is well-defined, since (p )
λ λ∈Par is a Q-module basis of ΛQ .
149This is well-defined, since the family (p )
n n≥1 generates the Q-algebra ΛQ and is algebraically independent.
150This is well-defined, since the family (p )
n n≥1 generates the Q-algebra ΛQ and is algebraically independent.
 HOPF ALGEBRAS IN COMBINATORICS 71

151
Show that ir (Λ) ⊂ Λ.
(e) Define a Q-linear map Sq : ΛQ → ΛQ by setting
Sq (pλ ) = p2λ for every partition λ.
152
Show that Sq (Λ) ⊂ Λ.
(f) Let r ∈ Z. Define a Q-algebra homomorphism ∆r : ΛQ → ΛQ ⊗Q ΛQ by setting
n−1
X n
∆r (pn ) = pi ⊗ pn−i + r ⊗ pn + pn ⊗ r for every positive integer n.
i=1
i
153
Show that ∆r (Λ) ⊂ Λ ⊗Z Λ.
(g) Consider the map ∆× introduced in Exercise 2.9.4(b) and the map 1 introduced in Exercise 2.9.4(c).
Show that the Q-algebra ΛQ , endowed with the comultiplication ∆× and the counit 1 , becomes a
cocommutative Q-bialgebra.154
(h) Define a Q-bilinear map ∗ : ΛQ × ΛQ → ΛQ , which will be written in infix notation (that is, we will
write a ∗ b instead of ∗ (a, b)), by setting
pλ ∗ pµ = δλ,µ zλ pλ for any partitions λ and µ
(where zλ is defined as in Proposition 2.5.15). 155 Show that f ∗ g ∈ Λ for any f ∈ Λ and g ∈ Λ.
(i) Show that 1 (f ) = f (1) for every f ∈ ΛQ (where we are using the notation r defined in Exer-
cise 2.9.4(c)).
[Hint:
 
• For (b), show that, for every f ∈ ΛQ , the tensor ∆× (f ) is the preimage of f (xi yj )(i,j)∈{1,2,3,...}2 =
f (x1 y1 , x1 y2 , x1 y3 , . . . , x2 y1 , x2 y2 , x2 y3 , . . . , . . .) ∈ Q [[x, y]] under the canonical injection ΛQ ⊗Q
ΛQ → Q [[x, y]] which maps  every f ⊗ g to f (x) g (y). (This requires making sure that the evaluation
f (xi yj )(i,j)∈{1,2,3,...}2 is well-defined to begin with, i.e., converges as a formal power series.)
For an alternative solution to (b), compute ∆× (hn ) or ∆× (en ).
• For (c), compute r (en ) or r (hn ).
• Reduce (d) to (b) and (c) using Exercise 1.3.6.
• Reduce (e) to (b).
• (f) is the hardest part. It is tempting to try and interpret the definition
 of ∆r as a convoluted way of
saying that ∆r (f ) is the preimage of f (xi + yj )(i,j)∈{1,2,3,...}2 under the canonical injection ΛQ ⊗Q
ΛQ → Q [[x, y]]
 which maps every f  ⊗ g to f (x) g (y). However, this does not make sense since the
evaluation f (xi + yj )(i,j)∈{1,2,3,...}2 is (in general) not well-defined156 (and even if it was, it would
fail to explain the r). So we need to get down to finitely many variables. For every N ∈ N, define a
Q-algebra homomorphism EN : ΛQ ⊗Q ΛQ → Q [x1 , x2 , . . . , xN , y1 , y2 , . . . , yN ] by sending each f ⊗ g
−1
to f (x1 , x2 , . . . , xN ) g (y1 , y2 , . . . , yN ). Show that ∆N (Λ) ⊂ EN (Z [x1 , x2 , . . . , xN , y1 , y2 , . . . , yN ]).
This shows that, at least, the coefficients of ∆r (f ) in front of the mλ ⊗ mµ with ` (λ) ≤ r and
` (µ) ≤ r (in the Q-basis (mλ ⊗ mµ )λ,µ∈Par of ΛQ ⊗Q ΛQ ) are integral for f ∈ Λ. Of course, we want


all coefficients. Show that ∆a = ∆b ? ∆ΛQ ◦ ia−b in Hom (ΛQ , ΛQ ⊗Q ΛQ ) for any integers a and
b. This allows “moving” the r. This approach to (f) was partly suggested to the first author by
Richard Stanley.
• For (h), notice that Definition 3.1.1(b) (below) allows us to construct a bilinear form (·, ·)ΛQ ⊗Q ΛQ :
(ΛQ ⊗Q ΛQ ) × (ΛQ ⊗Q ΛQ ) → Q from the Hall inner product (·, ·) : ΛQ × ΛQ → Q. Show that
(2.9.5) (a ∗ b, c) = (a ⊗ b, ∆× (c))ΛQ ⊗Q ΛQ for all a, b, c ∈ ΛQ ,
and then use (b).
151This is well-defined, since the family (p )
n n≥1 generates the Q-algebra ΛQ and is algebraically independent.
152This is well-defined, since (p )
λ λ∈Par is a Q-module basis of ΛQ .
153This is well-defined, since the family (p )
n n≥1 generates the Q-algebra ΛQ and is algebraically independent.
154But unlike Λ with the usual coalgebra structure, it is neither graded nor a Hopf algebra.
Q
155This is well-defined, since (p )
λ λ∈Par is a Q-module basis of ΛQ .
156e.g., it involves summing infinitely many x ’s if f = e
1 1
72 DARIJ GRINBERG AND VICTOR REINER

]
Remark 2.9.5. The map ∆× defined in Exercise 2.9.4(b) is known as the internal comultiplication (or Kro-
necker comultiplication) on ΛQ . Unlike the standard comultiplication ∆ΛQ , it is not a graded map, but rather
sends every homogeneous component (ΛQ )n into (ΛQ )n ⊗ (ΛQ )n . The bilinear map ∗ from Exercise 2.9.4(h)
is the so-called internal multiplication (or Kronecker multiplication), and is similarly not graded but rather
takes (ΛQ )n × (ΛQ )m to (ΛQ )n if n = m and to 0 otherwise.
The analogy between the two internal structures is not perfect: While we saw in Exercise 2.9.4(g) how
the internal comultiplication yields another bialgebra structure on ΛQ , it is not true that the internal mul-
tiplication (combined with the usual coalgebra structure of ΛQ ) forms a bialgebra structure as well. What
is missing is a multiplicative unity; if we would take the closure of ΛQ with respect to the grading, then
1 + h1 + h2 + h3 + · · · would be such a unity.
The structure constants of the internal comultiplication on the Schur basis (sλ )λ∈Par are equal to the
structure constants of the internal multiplication on the Schur basis157, and are commonly referred to as the
Kronecker coefficients. They are known to be nonnegative integers (this follows from Exercise 4.4.7(c)158), but
no combinatorial proof is known for their nonnegativity. Combinatorial interpretations for these coefficients
akin to the Littlewood-Richardson rule have been found only in special cases (cf., e.g., [162]).
The map ∆r of Exercise 2.9.4(f) also has some classical theory behind it, relating to Chern classes of
tensor products ([133], [125, §I.4, example 5]).
Parts (b), (c), (d), (e) and (f) of Exercise 2.9.4 are instances of a general phenomenon: Many Z-algebra
homomorphisms Λ → A (with A a commutative ring, usually torsionfree) are easiest to define by first defining
a Q-algebra homomorphism ΛQ → A ⊗ Q and then showing that this homomorphism restricts to a Z-algebra
homomorphism Λ → A. One might ask for general criteria when this is possible; specifically, for what choices
of (bn )n≥1 ∈ A{1,2,3,...} does there exist a Z-algebra homomorphism Λ → A sending the pn to bn ? Such
choices are called ghost-Witt vectors in Hazewinkel [75], and we can give various equivalent conditions for a
family (bn )n≥1 to be a ghost-Witt vector:
Exercise 2.9.6. Let A be a commutative ring.
For every n ∈ {1, 2, 3, . . .}, let ϕn : A → A be a ring endomorphism of A. Assume that the following
properties hold:
• We have ϕn ◦ ϕm = ϕnm for any two positive integers n and m.
• We have ϕ1 = id.
• We have ϕp (a) ≡ ap mod pA for every a ∈ A and every prime number p.
(For example, when A = Z, one can set ϕn = id for all n; this simplifies the exercise somewhat. More
generally, setting ϕn = id works whenever A is a binomial ring159. However, the results of this exercise are at
their most useful when A is a multivariate polynomial ring Z [x1 , x2 , x3 , . . .] over Z and the homomorphism
ϕn sends every P ∈ A to P (xn1 , xn2 , xn3 , . . .).)
Let µ denote the number-theoretic Möbius function; this is the function {1, 2, 3, . . .} → Z defined by

0, if m is not squarefree;
µ (m) = (number of prime factors of m) for every positive integer m.
(−1) , if m is squarefree
Let φ denote the Euler totient function; this is the function {1, 2, 3, . . .} → N which sends every positive
integer m to the number of elements of {1, 2, . . . , m} coprime to m.
157This can be obtained, e.g., from (2.9.5).
158Their integrality can also be easily deduced from Exercise 2.9.4(b).
159A binomial ring is defined to be a torsionfree (as an additive group) commutative ring A which has one of the following
equivalent properties:
a
• For every n ∈ N and a ∈ A, we have a (a − 1) · · · (a − n + 1) ∈ n! · A. (That is, binomial coefficients with a ∈ A
n
and n ∈ N are defined in A.)
• We have ap ≡ a mod pA for every a ∈ A and every prime number p.
See [202] and the references therein for studies of these rings. It is not hard to check that Z and every localization of Z are
binomial rings, and so is any commutative Q-algebra as well as the ring
{P ∈ Q [X] | P (n) ∈ Z for every n ∈ Z}
(but not the ring Z [X] itself).
 HOPF ALGEBRAS IN COMBINATORICS 73

Let (bn )n≥1 ∈ A{1,2,3,...} be a family of elements of A. Prove that the following seven assertions are
equivalent:
• Assertion C: For every positive integer n and every prime factor p of n, we have
ϕp bn/p ≡ bn mod pvp (n) A.

Here, vp (n) denotes the exponent of p in the prime factorization of n.

• Assertion D: There exists a family (αn )n≥1 ∈ A{1,2,3,...} of elements of A such that every positive
integer n satisfies X n/d
bn = dαd .
d|n
160

• Assertion E: There exists a family (βn )n≥1 ∈ A{1,2,3,...} of elements of A such that every positive
integer n satisfies X
bn = dϕn/d (βd ) .
d|n
• Assertion F: Every positive integer n satisfies
X

µ (d) ϕd bn/d ∈ nA.
d|n

• Assertion G: Every positive integer n satisfies

X

φ (d) ϕd bn/d ∈ nA.
d|n

• Assertion H: Every positive integer n satisfies

Xn


ϕn/ gcd(i,n) bgcd(i,n) ∈ nA.
i=1
• Assertion J : There exists a ring homomorphism ΛZ → A which, for every positive integer n, sends
pn to bn .
[Hint: The following identities hold for every positive integer n:
X
(2.9.6) φ (d) = n;
d|n
X
(2.9.7) µ (d) = δn,1 ;
d|n
X n
(2.9.8) µ (d)
= φ (n) ;
d
d|n
X n
(2.9.9) dµ (d) φ = µ (n) .
d
d|n

Furthermore, the following simple lemma is useful: If k is a positive integer, and if p ∈ N, a ∈ A and b ∈ A
` `
are such that a ≡ b mod pk A, then ap ≡ bp mod pk+` A for every ` ∈ N.]
Remark 2.9.7. Much of Exercise 2.9.6 is folklore, but it is hard to pinpoint concrete appearances in literature.
The equivalence C ⇐⇒ D appears in Hesselholt [80, Lemma 1] and [81, Lemma 1.1] (in slightly greater
generality), where it is referred to as Dwork’s lemma and used in the construction of the Witt vector
functor. This equivalence is also [75, Lemma 9.93]. The equivalence D ⇐⇒ F ⇐⇒ G ⇐⇒ H in the case
A = Z is [49, Corollary on p. 10], where it is put into the context of Burnside rings and necklace counting.
The equivalence C ⇐⇒ F for finite families (bn )n∈{1,2,...,m} in lieu of (bn )n≥1 is [183, Exercise 5.2 a]. One
of the likely oldest relevant sources is Schur’s [173], which proves the equivalence C ⇐⇒ D ⇐⇒ F for finite
families (bn )n∈{1,2,...,m} , as well as a “finite version” of C ⇐⇒ J (Schur did not have Λ, but was working
with actual power sums of roots of polynomials).
160Here and in the following, summations of the form P
d|n range over all positive divisors of n.
74 DARIJ GRINBERG AND VICTOR REINER

Exercise 2.9.8. Let A denote the ring Z. For every n ∈ {1, 2, 3, . . .}, let ϕn denote the identity endomor-
phism id of A. Prove that the seven equivalent assertions C, D, E, F, G, H and J of Exercise 2.9.6 are
satisfied for each of the following families (bn )n≥1 ∈ Z{1,2,3,...} :
• the family (bn )n≥1 = (q n )n≥1 , where q is a given integer.
• the family (bn )n≥1 = (q)n≥1 , where q is a given integer.
 
qn
• the family (bn )n≥1 = , where r ∈ Q and q ∈ Z are given. (Here, a binomial coefficient
rn n≥1
 
a
has to be interpreted as 0 when b ∈ / N.)
b  
qn − 1
• the family (bn )n≥1 = , where r ∈ Z and q ∈ Z are given.
rn − 1 n≥1
Exercise 2.9.9. For every n ∈ {1, 2, 3, . . .}, define a map fn : Λ → Λ by setting
fn (a) = a (xn1 , xn2 , xn3 , . . .) for every a ∈ Λ.
(So what fn does to a symmetric function is replacing all variables x1 , x2 , x3 , . . . by their n-th powers.)
(a) Show that fn : Λ → Λ is a k-algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
(b) Show that fn ◦ fm = fnm for any two positive integers n and m.
(c) Show that f1 = id.
(d) Prove that fn : Λ → Λ is a Hopf algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
P2m i
(e) Prove that f2 (hm ) = i=0 (−1) hi h2m−i for every m ∈ N.
(f) Assume that k = Z. Prove that fp (a) ≡ ap mod pΛ for every a ∈ Λ and every prime number p.
(g) Use Exercise 2.9.6 to obtain new solutions to parts (b), (c), (d), (e) and (f) of Exercise 2.9.4.
The maps fn constructed in Exercise 2.9.9 are known as the Frobenius endomorphisms of Λ. They are a
(deceptively) simple particular case of the notion of plethysm ([183, Chapter 7, Appendix 2]), and are often
used as intermediate steps in computing more complicated plethysms161.
Exercise 2.9.10. For every n ∈ {1, 2, 3, . . .}, define a k-algebra homomorphism vn : Λ → Λ by

hm/n , if n | m;
vn (hm ) = for every positive integer m
0, if n - m
162
.
(a) Show that any positive integers n and m satisfy

npm/n , if n | m;
vn (pm ) = .
0, if n - m
(b) Show that any positive integers n and m satisfy
m−m/n

(−1) nem/n , if n | m;
vn (em ) = .
0, if n - m
(c) Prove that vn ◦ vm = vnm for any two positive integers n and m.
(d) Prove that v1 = id.
(e) Prove that vn : Λ → Λ is a Hopf algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
Now, consider also the maps fn : Λ → Λ defined in Exercise 2.9.9. Fix a positive integer n.
(f) Prove that the maps fn : Λ → Λ and vn : Λ → Λ are adjoint with respect to the Hall inner product
on Λ.
(g) Show that vn ◦ fn = id?n
Λ .
(h) Prove that fn ◦ vm = vm ◦ fn whenever m is a positive integer coprime to n.
Finally, recall the wm ∈ Λ defined in Exercise 2.9.3.

161In the notations of [183, (A2.160)], the value f (a) for an a ∈ Λ can be written as a [p ] or (when k = Z) as p [a].
n n n
162This is well-defined, since the family (h )
m m≥1 generates the k-algebra Λ and is algebraically independent.
 HOPF ALGEBRAS IN COMBINATORICS 75

(i) Show that any positive integer m satisfies


wm/n , if n | m;
vn (wm ) = .
0, if n - m
The homomorphisms vn : Λ → Λ defined in Exercise 2.9.10 are called the Verschiebung endomorphisms
of Λ; this name comes from German, where “Verschiebung” means “shift”. This terminology, as well as
that of Frobenius endomorphisms, originates in the theory of Witt vectors, and the connection between the
Frobenius and Verschiebung endomorphisms of Λ and the identically named operators on Witt vectors is
elucidated in [75, Chapter 13]163.
Exercise 2.9.11. Fix n ∈ N. For any n-tuple w = (w1 , w2 , . . . , wn ) of integers, define the descent set
Des (w) of w to be the set {i ∈ {1, 2, . . . , n − 1} : wi > wi+1 }.
(a) We say that an n-tuple (w1 , w2 , . . . , wn ) is Smirnov if every i ∈ {1, 2, . . . , n − 1} satisfies wi 6= wi+1 .
Fix k ∈ N, and let Xn,k ∈ k [[x]] denote the sum of the monomials xw1 xw2 · · · xwn over all Smirnov
n
n-tuples w = (w1 , w2 , . . . , wn ) ∈ {1, 2, 3, . . .} satisfying |Des (w)| = k. Prove that Xn,k ∈ Λ.
(b) For any n-tuple w = (w1 , w2 , . . . , wn ), define the stagnation set Stag (w) of w to be the set
{i ∈ {1, 2, . . . , n − 1} : wi = wi+1 }. (Thus, an n-tuple is Smirnov if and only if its stagnation set is
empty.)
For any d ∈ N and s ∈ N, define a power series Xn,d,s ∈ k [[x]] as the sum of the monomials
n
xw1 xw2 · · · xwn over all n-tuples w = (w1 , w2 , . . . , wn ) ∈ {1, 2, 3, . . .} satisfying |Des (w)| = d and
|Stag (w)| = s. Prove that Xn,d,s ∈ Λ for any nonnegative integers d and s.
(c) Assume that n is positive. For any d ∈ N and s ∈ N, define three further power series Un,d,s , Vn,d,s
and Wn,d,s in k [[x]] by the following formulas:
X
(2.9.10) Un,d,s = xw1 xw2 · · · xwn ;
w=(w1 ,w2 ,...,wn )∈{1,2,3,...}n ;
|Des(w)|=d; |Stag(w)|=s;
w1 <wn
X
(2.9.11) Vn,d,s = xw1 xw2 · · · xwn ;
w=(w1 ,w2 ,...,wn )∈{1,2,3,...}n ;
|Des(w)|=d; |Stag(w)|=s;
w1 =wn
X
(2.9.12) Wn,d,s = xw1 xw2 · · · xwn .
w=(w1 ,w2 ,...,wn )∈{1,2,3,...}n ;
|Des(w)|=d; |Stag(w)|=s;
w1 >wn

Prove that these three power series Un,d,s , Vn,d,s and Wn,d,s belong to Λ.
Remark 2.9.12. The function Xn,k in Exercise 2.9.11(a) is a simple example ([177, Example 2.5, Theorem
C.3]) of a chromatic quasisymmetric function that happens to be symmetric. See Shareshian/Wachs [177]
for more general criteria for such functions to be symmetric, as well as deeper results. For example, [177,
Theorem 6.3] gives an expansion for a wide class of chromatic quasisymmetric functions in the Schur basis
of Λ, which, in particular, shows that our Xn,k satisfies
X
Xn,k = aλ,k sλ ,
λ∈Parn

where aλ,k is the number of all assignments T of entries in {1, 2, . . . , n} to the cells of the Ferrers diagram
of λ such that the following four conditions are satisfied:
• Every element of {1, 2, . . . , n} is used precisely once in the assignment (i.e., we have cont (T ) = (1n )).
• Whenever a cell y of the Ferrers diagram lies immediately to the right of a cell x, we have T (y) −
T (x) ≥ 2.
• Whenever a cell y of the Ferrers diagram lies immediately below a cell x, we have T (y) − T (x) ≥ −1.
• There exist precisely k elements i ∈ {1, 2, . . . , n − 1} such that the cell T −1 (i) lies in a row below
T −1 (i + 1).

163which is also where most of the statements of Exercises 2.9.9 and 2.9.10 come from
76 DARIJ GRINBERG AND VICTOR REINER

Are there any such rules for the Xn,d,s of part (b)?
Smirnov n-tuples are more usually called Smirnov words, or (occasionally) Carlitz words.
Exercise 2.9.13. (a) Let n ∈ N. Define a matrix An = (ai,j )i,j=1,2,...,n ∈ Λn×n by

pi−j+1 , if i ≥ j;

2
ai,j = i, if i = j − 1; for all (i, j) ∈ {1, 2, . . . , n} .

0, if i < j − 1


This matrix An looks as follows:

 
p1 1 0 ··· 0 0
 p2
 p1 2 ··· 0 
0
 p3 p2 p1 ··· 0 0
An =  . .
 
.. .. .. .. ..
 .. . . . .  .
 
 pn−1 pn−2 pn−3 ··· p1 n−1 
pn pn−1 pn−2 ··· p2 p1
Show that det (An ) = n!en .
(b) Let n be a positive integer. Define a matrix Bn = (bi,j )i,j=1,2,...,n ∈ Λn×n by
(
iei , if j = 1; 2
bi,j = for all (i, j) ∈ {1, 2, . . . , n} .
ei−j+1 , if j > 1
The matrix Bn looks as follows:
 
e1 e0 e−1 ··· e−n+3 e−n+2

 2e 2 e1 e0 ··· e−n+4 e−n+3 

 3e3 e2 e1 ··· e−n+5 e−n+4 
Bn = 
 
.. .. .. .. .. .. 

 . . . . . . 

 (n − 1) en−1 en−2 en−3 ··· e1 e0 
nen en−1 en−2 ··· e2 e1
 
e1 1 0 ··· 0 0

 2e 2 e1 1 ··· 0 0 

 3e 3 e2 e1 ··· 0 0 
= .
 
.. .. .. .. .. ..

 . . . . . . 

 (n − 1) en−1 en−2 en−3 ··· e1 1 
nen en−1 en−2 ··· e2 e1
Show that det (Bn ) = pn .
The formulas of Exercise 2.9.13, for finitely many variables, appear in Prasolov’s [151, §4.1]164. In [151,
§4.2], Prasolov gives four more formulas, which express en as a polynomial in the h1 , h2 , h3 , . . ., or hn as a
polynomial in the e1 , e2 , e3 , . . ., or pn as a polynomial in the h1 , h2 , h3 , . . ., or n!hn as a polynomial in the
p1 , p2 , p3 , . . .. These are not novel for us, since the first two of them are particular cases of Theorem 2.4.3,
whereas the latter two can be derived from Exercise 2.9.13 by applying ω. (Note that ω is only well-defined
on symmetric functions in infinitely many indeterminates, so we need to apply ω before evaluating at finitely
many indeterminates; this explains why Prasolov has to prove the latter two identities separately.)
Exercise 2.9.14. In the following, if k ∈ N, we shall use the notation 1k for 1, 1, . . . , 1 (in contexts such as
| {z }

k times
(n, 1m )). So, for example, 3, 14 is the partition (3, 1, 1, 1, 1).
(a) Show that en hm = s(m+1,1n−1 ) + s(m,1n ) for any two positive integers n and m.

164where our symmetric functions e , h , p , evaluated in finitely many indeterminates, are denoted σ , p , s , respectively
k k k k k k
 HOPF ALGEBRAS IN COMBINATORICS 77

(b) Show that

b
X i
(−1) ha+i+1 eb−i = s(a+1,1b )
i=0

for any a ∈ N and b ∈ N.

(c) Show that
b
X i b
(−1) ha+i+1 eb−i = (−1) δa+b,−1
i=0

for any negative integer a and every b ∈ N. (As usual, we set hj = 0 for j < 0 here.)
(d) Show that
∆s(a+1,1b ) = 1 ⊗ s(a+1,1b ) + s(a+1,1b ) ⊗ 1
X X
+ s(c+1,1d ) ⊗ s(e+1,1f ) + s(c+1,1d ) ⊗ s(e+1,1f )
(c,d,e,f )∈N4 ; (c,d,e,f )∈N4 ;
c+e=a−1; c+e=a;
d+f =b d+f =b−1

for any a ∈ N and b ∈ N.

Our next few exercises survey some results on Littlewood-Richardson coefficients.
Exercise 2.9.15. Let m ∈ N and k ∈ N. Let λ and µ be two partitions such that ` (λ) ≤ k and ` (µ) ≤ k.
Assume that all parts of λ and all parts of µ are ≤ m (that is, λi ≤ m and µi ≤ m for every positive integer i)
165
. Let λ∨ and µ∨ denote the k-tuples (m − λk , m − λk−1 , . . . , m − λ1 ) and (m − µk , m − µk−1 , . . . , m − µ1 ),
respectively.
(a) Show that λ∨ and µ∨ are partitions, and that sλ/µ = sµ∨ /λ∨ .
∨
(b) Show that cλµ,ν = cµλ∨ ,ν for any partition ν.
(c) Let ν be a partition such that ` (ν) ≤ k, and such that all parts of ν are ≤ m. Let ν ∨ denote the
k-tuple (m − νk , m − νk−1 , . . . , m − ν1 ). Show that ν ∨ is a partition, and satisfies
∨ ∨ ∨ ∨
cλµ,ν = cλν,µ = cµλ∨ ,ν = cµν,λ∨ = cνµ,λ∨ = cνλ∨ ,µ .

(d) Show that

m
sλ∨ (x1 , x2 , . . . , xk ) = (x1 x2 · · · xk ) · sλ x−1 −1 −1


1 , x2 , . . . , xk

in the Laurent polynomial ring k x1 , x2 , . . . , xk , x−1 −1 −1

 
1 , x2 , . . . , xk .
(e) Let r be a nonnegative integer. Show that (r + λ1 , r + λ2 , . . . , r + λk ) is a partition and satisfies
r
s(r+λ1 ,r+λ2 ,...,r+λk ) (x1 , x2 , . . . , xk ) = (x1 x2 · · · xk ) · sλ (x1 , x2 , . . . , xk )
in the polynomial ring k [x1 , x2 , . . . , xk ].
Exercise 2.9.16. Let m ∈ N, n ∈ N and k ∈ N. Let µ and ν be two partitions such that ` (µ) ≤ k and
166
` (ν) ≤ k. Assume that all parts of µ are ≤ m (that is, µi ≤ m for every positive integer i) , and
∨{m}
that all parts of ν are ≤ n (that is, νi ≤ n for every positive integer i). Let µ denote the k-tuple
(m − µk , m − µk−1 , . . . , m − µ1 ), and let ν ∨{n} denote the k-tuple (n − νk , n − νk−1 , . . . , n − ν1 ).
(a) Show that µ∨{m} and ν ∨{n} are partitions.
Now, let λ be a further partition such that ` (λ) ≤ k.
(b) If not all parts of λ are ≤ m + n, then show that cλµ,ν = 0.
∨{m+n}
(c) If all parts of λ are ≤ m + n, then show that cλµ,ν = cλµ∨{m} ,ν ∨{n} , where λ∨{m+n} denotes the k-tuple
(m + n − λk , m + n − λk−1 , . . . , m + n − λ1 ).

165As usual, we are denoting by ν the i-th part of a partition ν here.

i
166As usual, we are denoting by ν the i-th part of a partition ν here.
i
78 DARIJ GRINBERG AND VICTOR REINER

The results of Exercise 2.7.11(c) and Exercise 2.9.15(c) are two symmetries of Littlewood-Richardson
coefficients 167; combining them yields further such symmetries. While these symmetries were relatively
easy consequences of our algebraic definition of the Littlewood-Richardson coefficients, it is a much more
challenging task to derive them bijectively from a combinatorial definition of these coefficients (such as the one
given in Corollary 2.6.10). Some such derivations appear in [194], in [11], in [16, Example 3.6, Proposition 5.11
and references therein], [60, §5.1, §A.1, §A.4] and [94, (2.12)] (though a different combinatorial interpretation
of cλµ,ν is used in the latter three).

Exercise 2.9.17. Recall our usual notations: For every partition λ and every positive integer i, the i-th
entry of λ is denoted by λi . The sign . stands for dominance order. We let λt denote the conjugate partition
of a partition λ.
For any two partitions µ and ν, we define two new partitions µ + ν and µ t ν of |µ| + |ν| as follows:
• The partition µ + ν is defined as (µ1 + ν1 , µ2 + ν2 , µ3 + ν3 , . . .).

• The partition µ t ν is defined as the result of sorting the list µ1 , µ2 , . . . , µ`(µ) , ν1 , ν2 , . . . , ν`(ν) in
decreasing order.
t t
(a) Show that any two partitions µ and ν satisfy (µ + ν) = µt t ν t and (µ t ν) = µt + ν t .
(b) Show that any two partitions µ and ν satisfy cµ+ν µtν
µ,ν = 1 and cµ,ν = 1.
(c) If k ∈ N and n ∈ N satisfy k ≤ n, and if µ ∈ Park , ν ∈ Parn−k and λ ∈ Parn are such that cλµ,ν =
6 0,
then prove that µ + ν . λ . µ t ν.
(d) If n ∈ N and m ∈ N and α, β ∈ Parn and γ, δ ∈ Parm are such that α . β and γ . δ, then show that
α + γ . β + δ and α t γ . β t δ.  


(e) Let m ∈ N and k ∈ N, and let λ be the partition mk = m, m, . . . , m. Show that any two
| {z }
k times
partitions µ and ν satisfy cλµ,ν ∈ {0, 1}.


(f) Let a ∈ N and b ∈ N, and let λ be the partition a + 1, 1b (using the notation of Exercise 2.9.14).
Show that any two partitions µ and ν satisfy cλµ,ν ∈ {0, 1}.
(g) If λ is any partition, and if µ and ν are two rectangular partitions168, then show that cλµ,ν ∈ {0, 1}.

Exercise 2.9.17(g) is part of Stembridge’s [187, Thm. 2.1]; we refer to that article for further results of its
kind.
The Littlewood-Richardson rule comes in many different forms, whose equivalence is not always immediate.
Our version (Corollary 2.6.10) has the advantage of being the simplest to prove and one of the simplest to
state. Other versions can be found in [183, appendix 1 to Ch. 7], Fulton’s [60, Ch. 5] and van Leeuwen’s [113].
We restrict ourselves to proving some very basic equivalences that allow us to restate parts of Corollary 2.6.10:

Exercise 2.9.18. We shall use the following notations:

• If T is a column-strict tableau and j is a positive integer, then we use the notation T |cols≥j for the
restriction of T to the union of its columns j, j + 1, j + 2, . . .. (This notation has already been used
in Section 2.6.)
• If T is a column-strict tableau and S is a set of cells of T , then we write T |S for the restriction of T
to the set S of cells.169
• If T is a column-strict tableau, then an NE-set of T means a set S of cells of T such that whenever
s ∈ S, every cell of T which lies northeast170 of s must also belong to S.

167The result of Exercise 2.9.16(c) can also be regarded as a symmetry of Littlewood-Richardson coefficients; see [10, §3.3].
 
168A partition is called rectangular if it has the form mk
= m, m, . . . , m for some m ∈ N and k ∈ N.
 
| {z }
k times
169This restriction T | is not necessarily a tableau of skew shape; it is just a map from S to {1, 2, 3, . . .}. The content
S
cont (T |S ) is nevertheless well-defined (in the usual way: (cont (T |S ))i = (T |S )−1 (i) ).
170A cell (r, c) is said to lie northeast of a cell (r 0 , c0 ) if and only if we have r ≤ r 0 and c ≥ c0 .
 HOPF ALGEBRAS IN COMBINATORICS 79

• The Semitic reading word 171 of a column-strict tableau T is the concatenation172 r1 r2 r3 · · · , where
ri is the word obtained by reading the i-th row of T from right to left.173
• If w = (w1 , w2 , . . . , wn ) is a word, then a prefix of w means a word of the form (w1 , w2 , . . . , wi ) for
some i ∈ {0, 1, . . . , n}. (In particular, both w and the empty word are prefixes of w.)
A word w over the set of positive integers is said to be Yamanouchi if for any prefix v of w and
any positive integer i, there are at least as many i’s among the letters of v as there are (i + 1)’s
among them.174
Prove the following two statements:
(a) Let µ be a partition. Let bi,j be a nonnegative integer for every two positive integers i and j. Assume
that bi,j = 0 for all but finitely many pairs (i, j).
The following two assertions are equivalent:
– Assertion A: There exist a partition λ and a column-strict tableau T of shape λ/µ such that
2
all (i, j) ∈ {1, 2, 3, . . .} satisfy
(2.9.13) bi,j = (the number of all entries i in the j-th row of T ) .
– Assertion B: The inequality
(2.9.14) µj+1 + (b1,j+1 + b2,j+1 + · · · + bi+1,j+1 ) ≤ µj + (b1,j + b2,j + · · · + bi,j )
holds for all (i, j) ∈ N × {1, 2, 3, . . .}.
(b) Let λ and µ be two partitions, and let T be a column-strict tableau of shape λ/µ. Then, the following
five assertions are equivalent:
– Assertion C: For every positive integer j, the weak composition cont (T |cols≥j ) is a partition.
– Assertion D: For every positive integers j and i, the number of entries i+1 in the first j rows175
of T is ≤ to the number of entries i in the first j − 1 rows of T .
– Assertion E: For every NE-set S of T , the weak composition cont (T |S ) is a partition.
– Assertion F: The Semitic reading word of T is Yamanouchi.
– Assertion G: There exists a column-strict tableau S whose shape is a partition and which
satisfies the following property: For any positive integers i and j, the number of entries i in the
j-th row of T equals the number of entries j in the i-th row of S.
Remark 2.9.19. The equivalence of Assertions C and F in Exercise 2.9.18(b) is the “not-too-difficult exer-
cise” mentioned in [186]. It yields the equivalence between our version of the Littlewood-Richardson rule
(Corollary 2.6.10) and that in [183, A1.3.3].
In the next exercises, we shall restate Corollary 2.6.9 in a different form. While Corollary 2.6.9 provided a
decomposition of the product of a skew Schur function with a Schur function into a sum of Schur functions,
the different form that we will encounter in Exercise 2.9.21(b) will give a combinatorial interpretation for
the Hall inner product between two skew Schur functions. Let us first generalize Exercise 2.9.18(b):
Exercise 2.9.20. Let us use the notations of Exercise 2.9.18. Let κ, λ and µ be three partitions, and let T
be a column-strict tableau of shape λ/µ.
(a) Prove that the following five assertions are equivalent:
171The notation comes from [113] and is a reference to the Arabic and Hebrew way of writing.
172If s , s , s , . . . are several words (finitely or infinitely many), then the concatenation s s s · · · is defined as the word
1 2 3 1 2 3
which is obtained by starting with the empty word, then appending s1 to its end, then appending s2 to the end of the result,
then appending s3 to the end of the result, etc.
173For example, the Semitic reading word of the tableau

3 4 4 5
1 4 6
3 5
is 544364153.
The Semitic reading word of a tableau T is what is called the reverse reading word of T in [183, §A.1.3].
174For instance, the words 11213223132 and 1213 are Yamanouchi, while the words 132, 21 and 1121322332111 are not. The
Dyck words (written using 1’s and 2’s) are precisely the Yamanouchi words whose letters are 1’s and 2’s.
Yamanouchi words are often called lattice permutations.
175The “first j rows” mean the 1-st row, the 2-nd row, etc., the j-th row (even if some of these rows are empty).
80 DARIJ GRINBERG AND VICTOR REINER

– Assertion C (κ) : For every positive integer j, the weak composition κ + cont (T |cols≥j ) is a
partition.
– Assertion D(κ) : For every positive integers j and i, we have
κi+1 + (the number of entries i + 1 in the first j rows of T )
≤ κi + (the number of entries i in the first j − 1 rows of T ) .
– Assertion E (κ) : For every NE-set S of T , the weak composition κ + cont (T |S ) is a partition.
– Assertion F (κ) : For every prefix v of the Semitic reading word of T , and for every positive
integer i, we have
κi + (the number of i’s among the letters of v)
≥ κi+1 + (the number of (i + 1) ’s among the letters of v) .
– Assertion G (κ) : There exist a partition ζ and a column-strict tableau S of shape ζ/κ which
satisfies the following property: For any positive integers i and j, the number of entries i in the
j-th row of T equals the number of entries j in the i-th row of S.
(b) Let τ be a partition such that τ = κ + cont T . Consider the five assertions C (κ) , D(κ) , E (κ) , F (κ) and
G (κ) introduced in Exercise 2.9.20(a). Let us also consider the following assertion:
– Assertion H(κ) : There exists a column-strict tableau S of shape τ /κ which satisfies the following
property: For any positive integers i and j, the number of entries i in the j-th row of T equals
the number of entries j in the i-th row of S.
Prove that the six assertions C (κ) , D(κ) , E (κ) , F (κ) , G (κ) and H(κ) are equivalent.
Clearly, Exercise 2.9.18(b) is the particular case of Exercise 2.9.20 when κ = ∅.
Using Exercise 2.9.20, we can restate Corollary 2.6.9 in several ways:
Exercise 2.9.21. Let λ, µ and κ be three partitions.
(a) Show that X
sκ sλ/µ = sκ+cont T ,
T
where the sum ranges over all column-strict tableaux T of shape λ/µ satisfying the five equivalent
assertions C (κ) , D(κ) , E (κ) , F (κ) and G (κ) introduced

in Exercise 2.9.20(a).
(b) Let τ be a partition. Show that sλ/µ , sτ /κ Λ is the number of all column-strict tableaux T of shape
λ/µ satisfying τ = κ + cont T and also satisfying the six equivalent assertions C (κ) , D(κ) , E (κ) , F (κ) ,
G (κ) and H(κ) introduced in Exercise 2.9.20.
Exercise 2.9.21(a) is merely Corollary 2.6.9, rewritten in light of Exercise 2.9.20. Various parts of it
appear in the literature. For instance, [110, (53)] easily reveals to be a restatement of the fact that sκ sλ/µ =
(κ)
P
T sν+cont T , where the sum ranges over all column-strict tableaux T of shape λ/µ satisfying Assertion D .
Exercise 2.9.21(b) is one version of a “skew Littlewood-Richardson rule” that goes back to Zelevinsky [204]
(although Zelevinsky’s version uses both a different language and a combinatorial interpretation which is not
obviously equivalent
to ours). It appears in various sources; for instance, [110, Theorem 5.2, second formula]
says that sλ/µ , sτ /κ Λ is the number of all column-strict tableaux T of shape λ/µ satisfying τ = κ + cont T


and the assertion H(κ) , whereas [62, Theorem 1.2] says that sλ/µ , sτ /κ Λ is the number of all all column-
strict tableaux T of shape λ/µ satisfying τ = κ + cont T and the assertion F (κ) . (Notice that Gasharov’s
proof of [62, Theorem 1.2] uses the same involutions as Stembridge’s proof of Theorem 2.6.5; it can thus
be regarded as a close precursor to Stembridge’s proof. However, it uses the Jacobi-Trudi identities, while
Stembridge’s does not.)
Exercise 2.9.22. Let K be a field.176 If N ∈ Kn×n is a nilpotent matrix, then the Jordan type of N is
defined to be the list of the sizes of the Jordan blocks in the Jordan normal form of N , sorted in decreasing
order177. This Jordan type is a partition of n, and uniquely determines N up to similarity (i.e., two nilpotent
176This field has no relation to the ring k, over which our symmetric functions are defined.
177The Jordan normal form of N is well-defined even if K is not algebraically closed, because N is nilpotent (so the
characteristic polynomial of N is X n ).
 HOPF ALGEBRAS IN COMBINATORICS 81

n×n-matrices N and N 0 are similar if and only if the Jordan types of N and N 0 are equal). If f is a nilpotent
endomorphism of a finite-dimensional K-vector space V , then we define the Jordan type of f as the Jordan
type of any matrix representing f (the choice of the matrix does not matter, since the Jordan type of a
matrix remains unchanged under conjugation).
(a) Let n ∈ N. Let N ∈ Kn×n be a nilpotent matrix. Let λ ∈ Parn . Show that the matrix N has Jordan
type λ if and only if every k ∈ N satisfies
dim ker N k = λt 1 + λt 2 + . . . + λt k .

(Here, we are using the notation λt for the transpose of a partition λ, and the notation νi for the
i-th entry of a partition ν.)
(b) Let f be a nilpotent endomorphism of a finite-dimensional K-vector space V . Let U be an f -stable
K-vector subspace of V (that is, a K-vector subspace of V satisfying f (U ) ⊂ U ). Then, restricting f
to U gives a nilpotent endomorphism f | U of U , and the endomorphism f also induces a nilpotent
endomorphism f of the quotient space V /U . Let λ, µ and ν be the Jordan types of f , f | U and f ,
respectively. Show that cλµ,ν 6= 0 (if Z is a subring of k).
t
[Hint: For (b), Exercise 2.7.11(c) shows that it is enough to prove that cλµt ,ν t 6= 0. Due to Corol-
lary 2.6.10, this only requires constructing a column-strict tableau T of shape λt /µt with cont T = ν t
which
 has the property that each cont (T |cols≥j ) is a partition. Construct this tableau by defining ai,j =
i −1

j


dim f (U ) ∩ ker f for all (i, j) ∈ N2 , and requiring that the number of entries i in the j-th row
2
of T be ai,j − ai,j−1 − ai−1,j + ai−1,j−1 for all (i, j) ∈ {1, 2, 3, . . .} . Use Exercise 2.9.18(a) to prove that
this indeed defines a column-strict tableau, and Exercise 2.9.18(b) to verify that it satisfies the condition on
cont (T |cols≥j ).]
Remark 2.9.23. Exercise 2.9.22 is a taste of the connections between the combinatorics of partitions and the
Jordan normal form. Much more can, and has, been said. Marc van Leeuwen’s [111] is dedicated to some of
these connections; in particular, our Exercise 2.9.22(a) is [111, Proposition 1.1], and a far stronger version of
Exercise 2.9.22(b) appears in [111, Theorem 4.3 (2)], albeit only for the case of an infinite K. One can prove
a converse to Exercise 2.9.22(b) as well: If cλµ,ν 6= 0, then there exist V , f and U satisfying the premises of
Exercise 2.9.22(b). When K is a finite field, we can ask enumerative questions, such as how many U ’s are
there for given V , f , λ, µ and ν; we will see a few answers in Section 4.9 (specifically, Proposition 4.9.4),
and a more detailed treatment is given in [125, Ch. 2].
The relationship between partitions and Jordan normal forms can be exploited to provide linear-algebraic
proofs of purely combinatorial facts. See [27, Sections 6 and 9] for some examples. Note that [27, Lemma
9.10] is the statement that, under the conditions of Exercise 2.9.22(b), we have ν ⊆ λ. This is a direct
consequence of Exercise 2.9.22(b) (since cλµ,ν 6= 0 can happen only if ν ⊆ λ).
Exercise 2.9.24. Let a ∈ Λ. Prove the following:
n o
⊥
(a) The set g ∈ Λ | g ⊥ a = (ω (g)) a is a k-subalgebra of Λ.
⊥
(b) Assume that e⊥ ⊥ ⊥
k a = hk a for each positive integer k. Then, g a = (ω (g)) a for each g ∈ Λ.

Exercise 2.9.25. Let n ∈ N. Let ρ be the partition (n − 1, n − 2, . . . , 1). Prove that sρ/µ = sρ/µt for every
µ ∈ Par.
Remark 2.9.26. Exercise 2.9.25 appears in [159, Corollary 7.32], and is due to John Stembridge. Using
Remark 2.5.9, we can rewrite it as yet another equality between Littlewood-Richardson coefficients: Namely,
cρµ,ν = cρµt ,ν for any µ ∈ Par and ν ∈ Par.
82 DARIJ GRINBERG AND VICTOR REINER

3. Zelevinsky’s structure theory of positive self-dual Hopf algebras

Chapter 2 showed that, as a Z-basis for the Hopf algebra Λ = ΛZ , the Schur functions {sλ } have two special
properties: they have the same structure constants cλµ,ν for their multiplication as for their comultiplication
(Corollary 2.5.7), and these structure constants are all nonnegative integers (Corollary 2.6.10). Zelevinsky
[203, §2,3] isolated these two properties as crucial.
Definition 3.0.1. Say that a connected graded Hopf algebra A over k = Z with a distinguished Z-basis
{σλ } consisting of homogeneous elements178 is a positive self-dual Hopf algebra (or PSH) if it satisfies the
two further axioms
• (self-duality) The same structure constants aλµ,ν appear for the product σµ σν = λ aλµ,ν σλ and
P

the coproduct ∆σλ = µ,ν aλµ,ν σµ ⊗ σν .

P

• (positivity) The aλµ,ν are all nonnegative (integers).

Call {σλ } the PSH-basis of A.
He then developed a beautiful structure theory for PSH’s, explaining how they can be uniquely expressed
as tensor products of copies of PSH’s each isomorphic to Λ after rescaling their grading. The next few
sections explain this, following his exposition closely.
3.1. Self-duality implies polynomiality. We begin with a property that forces a Hopf algebra to have
algebra structure which is a polynomial algebra, specifically the symmetric algebra Sym(p), where p is the
k-submodule of primitive elements. L∞
Recall from Exercise 1.3.19(g)
L that for a connected graded Hopf algebra A = n=0 An , every x in the
two-sided ideal I := ker  = n>0 An has the property that its comultiplication takes the form
∆(x) = 1 ⊗ x + x ⊗ 1 + ∆+ (x)
where ∆+ (x) lies in I ⊗ I. Recall also that the elements x for which ∆+ (x) = 0 are called the primitives.
Denote by p the k-submodule of primitive elements inside A.
Given a PSH A (over k = Z) with a PSH-basis {σλ }, we consider the bilinear form (·, ·)A : A × A → Z
on A that makes this basis orthonormal. Similarly, the elements {σλ ⊗ σµ } give an orthonormal basis for a
form (·, ·)A⊗A on A ⊗ A. The bilinear form (·, ·)A on the PSH A gives rise to a Z-linear map A → Ao , which
is easily seen to be injective and a Z-algebra homomorphism. We thus identify A with a subalgebra of Ao .
When A is of finite type, this map is a Hopf algebra isomorphism, thus allowing us to identify A with Ao .
This is an instance of the following notion of self-duality.
Definition 3.1.1. L If (·, ·) : V × W →
(a) L k is a bilinear form on the product V × W of two graded
k-modules V = n≥0 V n and W = n≥0 Wn , then we say that this form (·, ·) is graded if every
two distinct nonnegative integers n and m satisfy (Vn , Wm ) = 0 (that is, if every two homogeneous
elements v ∈ V and w ∈ W having distinct degrees satisfy (v, w) = 0).
(b) If (·, ·)V : V × V → k and (·, ·)W : W × W → k are two symmetric bilinear forms on some k-modules
V and W , then we can canonically define a symmetric bilinear form (·, ·)V ⊗W on the k-module
V ⊗ W by letting
(v ⊗ w, v 0 ⊗ w0 )V ⊗W = (v, v 0 )V (w, w0 )W for all v, v 0 ∈ V and w, w0 ∈ W.
This new bilinear form is graded if the original two forms (·, ·)V and (·, ·)W were graded (presuming
that V and W are graded).
(c) Say that a bialgebra A is self-dual with respect to a given symmetric bilinear form (·, ·) : A × A → k
if one has (a, m(b ⊗ c))A = (∆(a), b ⊗ c)A⊗A and (1A , a) = (a) for a, b, c in A. If A is a graded Hopf
algebra of finite type, and this form (·, ·) is graded, then this is equivalent to the k-module map
A → Ao induced by (·, ·)A giving a Hopf algebra homomorphism.
Thus, any PSH A is self-dual with respect to the bilinear form (·, ·)A that makes its PSH-basis orthonormal.
Notice also that the injective Z-algebra homomorphism A → Ao obtained from the bilinear form (·, ·)A
on a PSH A allows us to regard each f ∈ A as an element of Ao . Thus, for any PSH A and any f ∈ A, an
operator f ⊥ : A → A is well-defined (indeed, regard f as an element of Ao , and apply Definition 2.8.1).
178not necessarily indexed by partitions
 HOPF ALGEBRAS IN COMBINATORICS 83

Proposition 3.1.2. Let A be a Hopf algebra over k = Z or k = Q which is graded, connected, and self-dual
with respect to a positive definite graded179 bilinear form. Then:
(a) Within the ideal I, the k-submodule of primitives p is the orthogonal complement to the k-submodule
I 2.
(b) In particular, p ∩ I 2 = 0.
(c) When k = Q, one has I = p ⊕ I 2 .
Proof. (a) Note that I 2 = m(I ⊗ I). Hence an element x in I lies in the perpendicular space to I 2 if and
only if one has for all y in I ⊗ I that
0 = (x, m(y))A = (∆(x), y)A⊗A = (∆+ (x), y)A⊗A
where the second equality uses self-duality, while the third equality uses the fact that y lies in I ⊗ I and the
form (·, ·)A⊗A makes distinct homogeneous components orthogonal. Since y was arbitrary, this means x is
perpendicular to I 2 if and only if ∆+ (x) = 0, that is, x lies in p.
(b) This follows from (a), since the form (·, ·)A is positive definite.
(c) This follows from (a) using some basic linear algebra180 when A is of finite type (which is the only
case we will ever encounter in practice). See Exercise 3.1.6 for the general proof. 
Remark 3.1.3. One might wonder why we didn’t just say I = p ⊕ I 2 even when k = Z in Proposition 3.1.2(c).
However, this is false even for A = ΛZ : the second homogeneous component (p⊕I 2 )2 is the index 2 sublattice
of Λ2 which is Z-spanned by {p2 , e21 }, containing 2e2 , but not containing e2 itself.
Already the fact that p ∩ I 2 = 0 has a strong implication.
Lemma 3.1.4. A connected graded Hopf algebra A over any ring k having p ∩ I 2 = 0 must necessarily be
commutative (as an algebra).
Proof. The component A0 = k commutes with all of A. This forms the base case for an induction on i + j
in which one shows that any elements x in Ai and y in Aj with i, j > 0 will have [x, y] := xy − yx = 0. Since
[x, y] lies in I 2 , it suffices to show that [x, y] also lies in p:
∆[x, y] = [∆(x), ∆(y)]
= [1 ⊗ x + x ⊗ 1 + ∆+ (x), 1 ⊗ y + y ⊗ 1 + ∆+ (y)]
= [1 ⊗ x + x ⊗ 1, 1 ⊗ y + y ⊗ 1]
+ [1 ⊗ x + x ⊗ 1, ∆+ (y)] + [∆+ (x), 1 ⊗ y + y ⊗ 1] + [∆+ (x), ∆+ (y)]
= [1 ⊗ x + x ⊗ 1, 1 ⊗ y + y ⊗ 1]
= 1 ⊗ [x, y] + [x, y] ⊗ 1
showing that [x, y] lies in p. Here the second-to-last equality used the inductive hypotheses: homogeneity
implies that ∆+ (x) is a sum of homogeneous tensors of the form z1 ⊗ z2 satisfying deg(z1 ), deg(z2 ) < i,
so that by induction they will commute with 1 ⊗ y, y ⊗ 1, thus proving that [∆+ (x), 1 ⊗ y + y ⊗ 1] = 0; a
symmetric argument shows [1 ⊗ x + x ⊗ 1, ∆+ (y)] = 0, and, a similar argument shows [∆+ (x), ∆+ (y)] = 0.
The last equality is an easy calculation, and was done already in (1.3.6). 
Remark 3.1.5. Zelevinsky actually shows [203, Proof of A.1.3, p. 150] that the assumption of p ∩ I 2 = 0
(along with hypotheses of unit, counit, graded, connected, and ∆ being a morphism for multiplication)
already implies the associativity of the multiplication in A ! One shows by induction on i + j + k that
any x, y, z in Ai , Aj , Ak with i, j, k > 0 have vanishing associator assoc(x, y, z) := x(yz) − (xy)z. In the
inductive step, one first notes that assoc(x, y, z) lies in I 2 , and then checks that assoc(x, y, z) also lies in p,
by a calculation very similar to the one above, repeatedly using the fact that assoc(x, y, z) is multilinear in
its three arguments.
Exercise 3.1.6. Prove Proposition 3.1.2(c) in the general case.
179That is, (A , A ) = 0 for i 6= j.
i j
180Specifically, either the existence of an orthogonal projection on a subspace of a finite-dimensional inner-product space
over Q, or the fact that dim W ⊥ = dim V − dim W for a subspace W of a finite-dimensional inner-product space V over Q

can be used.
84 DARIJ GRINBERG AND VICTOR REINER

This leads to a general structure theorem.

Theorem 3.1.7. If a connected graded Hopf algebra A over a field k of characteristic zero has I = p ⊕ I 2 ,
then the inclusion p ,→ A extends to a Hopf algebra isomorphism from the symmetric algebra Symk (p) → A.
In particular, A is both commutative and cocommutative.

Note that the hypotheses of Theorem 3.1.7 are valid, using Proposition 3.1.2(c), whenever A is obtained
from a PSH (over Z) by tensoring with Q.

Proof of Theorem 3.1.7. Since Lemma 3.1.4 implies that A is commutative, the universal property of Symk (p)
as a free commutative algebra on generators p shows that the inclusion p ,→ A at least extends to an algebra
ϕ
morphism Symk (p) → A. Since the Hopf structure on Symk (p) makes the elements of p primitive (see
Example 1.3.14), this ϕ is actually a coalgebra morphism (since ∆ ◦ ϕ = (ϕ ⊗ ϕ) ◦ ∆ and  ◦ ϕ =  need
only to be checked on algebra generators), hence a bialgebra morphism, hence a Hopf algebra morphism (by
Proposition 1.4.24(c)). It remains to show that ϕ is surjective, and injective.
For the surjectivity of ϕ, note that the hypothesis I = p ⊕ I 2 implies that the composite p ,→ I → I/I 2
gives a k-vector space isomorphism. What follows is a standard argument to deduce that p generates A as
a commutative graded k-algebra. One shows by induction on n that any homogeneous element a in An lies
in the k-subalgebra generated by p. The base case n = 0 is trivial as a P lies in A0 = k · 1A . In the inductive
step where a lies in I, write a ≡ p mod I 2 for some p in p. Thus a = p + i bi ci , where bi , ci lie in I but have
strictly smaller degree, so that by induction they lie in the subalgebra generated by p, and hence so does a.
Note that the surjectivity argument did not use the assumption that k has characteristic zero, but we will
now use it in the injectivity argument for ϕ, to establish the following
(3.1.1) Claim: Every primitive element of Sym(p) lies in p = Sym1 (p).
Note that this claim fails in positive characteristic, e.g. if k has characteristic 2 then x2 lies in Sym2 (p),
however
∆(x2 ) = 1 ⊗ x2 + 2x ⊗ x + x2 ⊗ 1 = 1 ⊗ x2 + x2 ⊗ 1.
To see the claim, assume not, so that by gradedness, there must exist some primitive element y 6= 0 lying in
some Symn (p) with n ≥ 2. This would mean that the composite map f that follows the coproduct with a
component projection
∆
M
Symn (p) −→ Symi (p) ⊗ Symj (p) −→ Sym1 (p) ⊗ Symn−1 (p)
i+j=n

m
has f (y) = 0. However, one can check on a basis that the multiplication backward Sym1 (p) ⊗ Symn−1 (p) →
Symn (p) has the property that m ◦ f = n · idSymn (p) :
 
Xn
(m ◦ f )(x1 · · · xn ) = m  xj ⊗ x1 · · · xbj · · · xn  = n · x1 · · · xn
j=1

for x1 , . . . , xn in p. Then n · y = m(f (y)) = m(0) = 0 leads to the contradiction that y = 0, since k has
characteristic zero.
Now one can argue the injectivity of the (graded) map181 ϕ by assuming that one has a nonzero homoge-
neous element u in ker(ϕ) of minimum degree. In particular, deg(u) ≥ 1. Also since p ,→ A, one has that u
is not in Sym1 (p) = p, and hence u is not primitive by the previous Claim. Consequently ∆+ (u) 6= 0, and
one can find a nonzero component u(i,j) of ∆+ (u) lying in Sym(p)i ⊗ Sym(p)j for some i, j > 0. Since this
forces i, j < deg(u), one has that ϕ maps both Sym(p)i , Sym(p)j injectively into Ai , Aj . Hence the tensor
product map
ϕ⊗ϕ
Sym(p)i ⊗ Sym(p)j −→ Ai ⊗ Aj

181The grading on Sym(p) is induced from the grading on p, a homogeneous subspace of I ⊂ A as it is the kernel of the
∆+
graded map I −→ A ⊗ A.
 HOPF ALGEBRAS IN COMBINATORICS 85

is also injective182. This implies (ϕ ⊗ ϕ)(u(i,j) ) 6= 0, giving the contradiction that

Sym(p)
0 = ∆A A
+ (0) = ∆+ (ϕ(u)) = (ϕ ⊗ ϕ)(∆+ (u))
contains the nonzero Ai ⊗ Aj -component (ϕ ⊗ ϕ)(u(i,j) ).
(An alternative proof of the injectivity of ϕ proceeds as follows: By (3.1.1), the subspace of primitive
elements of Sym(p) is p, and clearly ϕ |p is injective. Hence, Exercise 1.4.32(b) (applied to the homomorphism
ϕ) shows that ϕ is injective.) 

Before closing this section, we mention one nonobvious corollary of the Claim (3.1.1), when applied to the
ring of symmetric functions ΛQ with Q-coefficients, since Proposition 2.4.1 says that ΛQ = Q[p1 , p2 , . . .] =
Sym(V ) where V = Q{p1 , p2 , . . .}.

Corollary 3.1.8. The subspace p of primitives in ΛQ is one-dimensional in each degree n = 1, 2, . . ., and

spanned by {p1 , p2 , . . .}.

We note in passing that this corollary can also be obtained in a simpler fashion and a greater generality:
Exercise 3.1.9. Let k be any commutative ring. Show that the primitive elements of Λ are precisely the
elements of the k-linear span of p1 , p2 , p3 , . . ..
3.2. The decomposition theorem. Our goal here is Zelevinsky’s theorem [203, Theorem 2.2] giving a
canonical decomposition of any PSH as a tensor product into PSH’s that each have only one primitive
element in their PSH-basis. For the sake of stating it, we introduce some notation.
Definition 3.2.1. Given a PSH A with PSH-basis Σ, let C := Σ ∩ p be the primitive elements in Σ. For
each ρ in C, let A(ρ) ⊂ A be the Z-span of
Σ(ρ) := {σ ∈ Σ : there exists n ≥ 0 with (σ, ρn ) 6= 0}.
Definition 3.2.2. The tensor product of two PSHs A1 and A2 with PSH-bases Σ1 and Σ2 is defined as the
graded Hopf algebra A1 ⊗ A2 with PSH-basis {σ1 ⊗ σ2 }(σ1 ,σ2 )∈Σ1 ×Σ2 . It is easy to see that this is again a
PSH. The tensor product of any finite family of PSHs is defined similarly183.

Theorem 3.2.3. Any PSH A has a canonical tensor product decomposition

O
A= A(ρ)
ρ∈C

with A(ρ) a PSH, and ρ the only primitive element in its PSH-basis Σ(ρ).

Although in all the applications, C will be finite, when C is infinite one should interpret the tensor product
in the theorem asNthe inductive limit of tensor products over finite subsets of C, that is, linear combinations
of basic tensors ρ aρ in which there are only finitely many factors aρ 6= 1.
The first step toward the theorem uses a certain unique factorization property.

Lemma 3.2.4. Let P be a set of pairwise orthogonal primitives in a PSH A. Then,

(ρ1 · · · ρr , π1 · · · πs ) = 0
for ρi , πj in P unless r = s and one can reindex so that ρi = πi .

ϕ
182One needs to know that for two injective maps V →i W of k-vector spaces V , W with i = 1, 2, the tensor product
i i i i
ϕ1 ⊗ ϕ2 is also injective. Factoring it as ϕ1 ⊗ ϕ2 = (id ⊗ϕ2 ) ◦ (ϕ1 ⊗ id) , one sees that it suffices to show that for an injective
ϕ ϕ⊗id
map V ,→ W of free k-modules, and any free k-module U , the map V ⊗ U −→ W ⊗ U is also injective. Since tensor products
commute with direct sums, and U is (isomorphic to) a direct sum of copies of k, this reduces to the easy-to-check case where
U = k.
ϕ1 =(·×2)
Note that some kind of freeness or flatness hypothesis on U is needed here since, e.g. the injective Z-module maps Z −→ Z
ϕ2 =id
and Z/2Z −→ Z/2Z have ϕ1 ⊗ ϕ2 = 0 on Z ⊗Z Z/2Z ∼ = Z/2Z 6= 0.
183For the empty family, it is the connected graded Hopf algebra Z with PSH-basis {1}.
86 DARIJ GRINBERG AND VICTOR REINER

Proof. Induct on r. For r > 0, one has

(ρ1 · · · ρr , π1 · · · πs ) = (ρ2 · · · ρr , ρ⊥
1 (π1 · · · πs ))
s
X
= (ρ2 · · · ρr , (π1 · · · πj−1 · ρ⊥
1 (πj ) · πj+1 · · · πs ))
j=1
184
from Proposition 2.8.2(iii) because ρ1 is primitive . On the other hand, since each πj is primitive, one has
ρ⊥1 (πj ) = (ρ1 , 1) · πj + (ρ1 , πj ) · 1 = (ρ1 , πj ) which vanishes unless ρ1 = πj . Hence (ρ1 · · · ρr , π1 · · · πs ) = 0
unless ρ1 ∈ {π1 , . . . , πs }, in which case after reindexing so that π1 = ρ1 , it equals
n · (ρ1 , ρ1 ) · (ρ2 · · · ρr , π2 · · · πs )
if there are exactly n occurrences of ρ1 among π1 , . . . , πs . Now apply induction. 
So far the positivity hypothesis for a PSH has played little role. Now we use it to introduce a certain
partial order on the PSH A, and then a semigroup grading.
Definition 3.2.5. For a subset S of an abelian group, let ZS (resp. NS) denote the subgroup of Z-linear
combinations (resp. submonoid of N-linear combinations185) of the elements of S.
In a PSH A with PSH-basis Σ, the subset NΣ forms a submonoid, and lets one define a partial order on
A via a ≤ b if b − a lies in NΣ.
We note a few trivial properties of this partial order:
• The positivity hypothesis implies that NΣ · NΣ ⊂ NΣ.
• Hence multiplication by an element c ≥ 0 (meaning c lies in NΣ) preserves the order: a ≤ b implies
ac ≤ bc since (b − a)c lies in NΣ.
• Thus 0 ≤ c ≤ d and 0 ≤ a ≤ b implies ac ≤ bc ≤ bd.
This allows one to introduce a semigroup grading on A.
Definition 3.2.6. Let NCfin denote the additive submonoid of NC consisting of those α = (αρ )ρ∈C with finite
support.
Note that for any α in NCfin , one has that the product ρ∈C ραρ ≥ 0. Define
Q
Y
Σ(α) := {σ ∈ Σ : σ ≤ ραρ },
ρ∈C
αρ
Q
that is, the subset of Σ on which ρ∈C ρ has support. Also define
A(α) := ZΣ(α) ⊂ A.
Proposition 3.2.7. The PSH A has an NCfin -semigroup-grading: one has an orthogonal direct sum decom-
position M
A= A(α)
α∈NC
fin

for which
(3.2.1) A(α) A(β) ⊂ A(α+β)
M
(3.2.2) ∆A(α) ⊂ A(β) ⊗ A(γ) .
α=β+γ

Proof. We will make free use of the fact that a PSH A is commutative, since it embeds in A ⊗Z Q, which is
commutative by Theorem 3.1.7.

Note that the orthogonality A(α) , A(β) = 0 for α 6= β is equivalent to the assertion that
 
Y Y
 ραρ , ρβρ  = 0,
ρ∈C ρ∈C

184Strictly speaking, this argument needs further justification since A might not be of finite type (and if it is not, Proposi-
tion 2.8.2(iii) cannot be applied). It is more adequate to refer to the proof of Proposition 2.8.2(iii), which indeed goes through
with ρ1 taking the role of f .
185Recall that N := {0, 1, 2, . . .}.
 HOPF ALGEBRAS IN COMBINATORICS 87

which follows from Lemma 3.2.4.

Next let us deal with the assertion (3.2.1). It suffices to check that when τ, ω in Σ lie in A(α) , A(β) ,
respectively, then τ ω lies in A(α+β) . But note that any σ in Σ having σ ≤ τ ω will then have
Y Y Y
σ ≤ τω ≤ ραρ · ρβρ = ραρ +βρ
ρ∈C ρ∈C ρ∈C

so that σ lies in A(α+β) . This

L means that τ ω lies in A(α+β) .
This lets us check that α∈NC A(α) exhaust A. It suffices to check that any σ in Σ lies in some A(α) .
fin
Proceed by induction on deg(σ), with the case σ = 1 being trivial; the element 1 always lies in Σ, and hence
lies in A(α) for α = 0. For σ lying in I, one either has (σ, a) 6= 0 for some a in I 2 , or else σ lies in (I 2 )⊥ = p
(by Proposition 3.1.2(a)), so that σ is in C and we are done. If (σ, a) 6= 0 with a in I 2 , then σ appears in
the support of some Z-linear combination of elements τ ω where τ, ω lie in Σ and have strictly smaller degree
than σ has. There exists at least one such pair τ, ω for which (σ, τ ω) 6= 0, and therefore σ ≤ τ ω. Then by
induction τ, ω lie in some A(α) , A(β) , respectively, so τ ω lies in A(α+β) , and hence σ lies in A(α+β) also.
Self-duality shows that (3.2.1) implies (3.2.2): if a, b, c lie in A(α) , A(β) , A(γ) , respectively, then (∆a, b ⊗
c)A⊗A = (a, bc)A = 0 unless α = β + γ. 

Proposition 3.2.8. For α, β in NCfin with disjoint support, one has a bijection
Σ(α) × Σ(β) −→ Σ(α + β)
(σ, τ ) 7−→ στ.
Thus, the multiplication map A(α) ⊗ A(β) → A(α+β) is an isomorphism.

Proof. We first check that for σ1 , σ2 in Σ(α) and τ1 , τ2 in Σ(β), one has
(3.2.3) (σ1 τ1 , σ2 τ2 ) = δ(σ1 ,τ1 ),(σ2 ,τ2 ) .
Note that this is equivalent to showing both
• that στ lie in Σ(α + β) so that the map is well-defined, since it shows (στ, στ ) = 1, and
• that the map is injective.
One calculates
(σ1 τ1 , σ2 τ2 )A = (σ1 τ1 , m(σ2 ⊗ τ2 ))A
= (∆(σ1 τ1 ), σ2 ⊗ τ2 )A⊗A
= (∆(σ1 )∆(τ1 ), σ2 ⊗ τ2 )A⊗A
P
Note that due to (3.2.2), ∆(σ1 )∆(τ1 ) lies in A(α0 +β 0 ) ⊗ A(α00 +β 00 ) where
α0 + α00 = α
β 0 + β 00 = β.
Since σ2 ⊗ τ2 lies in A(α) ⊗ A(β) , the only nonvanishing terms in the inner product come from those with
α0 + β 0 = α
α00 + β 00 = β.
As α, β have disjoint support, this can only happen if
α0 = α, α00 = 0, β 0 = 0, β 00 = β,
that is, the only nonvanishing term comes from (σ1 ⊗ 1)(1 ⊗ τ1 ) = σ1 ⊗ τ1 . Hence
(σ1 τ1 , σ2 τ2 )A = (σ1 ⊗ τ1 , σ2 ⊗ τ2 )A⊗A = δ(σ1 ,τ1 ),(σ2 ,τ2 ) .
To see that the map is surjective, express
Y X
ραρ = σi
ρ∈C i
Y X
ρβρ = τj
ρ∈C j
88 DARIJ GRINBERG AND VICTOR REINER

with σi ∈ Σ(α) and τj in Σ(β). Then each product σi τj is in Σ(α + β) by (3.2.3), and
Y X
ραρ +βρ = σi τj
ρ∈C i,j

shows that {σi τj } exhausts Σ(α + β). This gives surjectivity. 

Proof of Theorem 3.2.3. Recall from Definition 3.2.1 that for each ρ in C, one defines A(ρ) ⊂ A to be the
Z-span of
Σ(ρ) := {σ ∈ Σ : there exists n ≥ 0 with (σ, ρn ) 6= 0}.
In other words, A(ρ) := n≥0 A(n·eρ ) where eρ in NCfin is the standard basis element indexed by ρ. Proposi-
L

then shows that A(ρ) is a Hopf subalgebra of A. Since every α in NCfin can be expressed
tion 3.2.7 P Nas the (fi-
nite) sum ρ αρ eρ , and the eρ have disjoint support, iterating Proposition 3.2.8 shows that A = ρ∈C A(ρ).
Lastly, Σ(ρ) is clearly a PSH-basis for A(ρ), and if σ is any primitive element in Σ(ρ) then (σ, ρn ) 6= 0 lets
one conclude via Lemma 3.2.4 that σ = ρ (and n = 1). 

3.3. Λ is the unique indecomposable PSH. The goal here is to prove the rest of Zelevinsky’s structure
theory for PSH’s. Namely, if A has only one primitive element ρ in its PSH-basis Σ, then A must be
isomorphic as a PSH to the ring of symmetric functions Λ, after one rescales the grading of A. Note that
every σ in Σ has σ ≤ ρn for some n, and hence has degree divisible by the degree of ρ. Thus one can divide
all degrees by that of ρ and assume ρ has degree 1.
The idea is to find within A and Σ a set of elements that play the role of
{hn = s(n) }n=0,1,2,... , {en = s(1n ) }n=0,1,2,...
within A = Λ and its PSH-basis of Schur functions Σ = {sλ }. Zelevinsky’s argument does this by isolating
some properties that turn out to characterize these elements:
(a) h0 = e0 = 1, and h1 = e1 =: ρ has ρ2 a sum of two elements of Σ, namely
ρ2 = h2 + e2 .
(b) For all n = 0, 1, 2, . . ., there exist unique elements hn , en in An ∩ Σ that satisfy

h⊥
2 en = 0,

e⊥
2 hn = 0

with h2 , e2 being the two elements of Σ introduced in (a).

(c) For k = 0, 1, 2, . . . , n one has

h⊥ ⊥
k hn = hn−k and σ hn = 0 for σ ∈ Σ \ {h0 , h1 , . . . , hn }

e⊥ ⊥
k en = en−k and σ en = 0 for σ ∈ Σ \ {e0 , e1 , . . . , en }.

In particular, e⊥ ⊥
k hn = 0 = hk en for k ≥ 2.
(d) Their coproducts are
X
∆(hn ) = hi ⊗ hj ,
i+j=n
X
∆(en ) = ei ⊗ ej .
i+j=n

We will prove Zelevinsky’s result [203, Theorem 3.1] as a combination of the following two theorems.

Theorem 3.3.1. Let A be a PSH with PSH-basis Σ containing only one primitive ρ, and assume that the
grading has been rescaled so that ρ has degree 1. Then, after renaming ρ = e1 = h1 , one can find unique
sequences {hn }n=0,1,2,... , {en }n=0,1,2,... of elements of Σ having properties (a),(b),(c),(d) listed above.

The second theorem uses the following notion.

 HOPF ALGEBRAS IN COMBINATORICS 89

ϕ
Definition 3.3.2. A PSH-morphism A → A0 between two PSH’s A, A0 having PSH-bases Σ, Σ0 is a graded
Hopf algebra morphism for which ϕ(NΣ) ⊂ NΣ0 . If A = A0 and Σ = Σ0 it will be called a PSH-endomorphism.
If ϕ is an isomorphism and restricts to a bijection Σ → Σ0 , it will be called a PSH-isomorphism; if it is both
a PSH-isomorphism and an endomorphism, it is a PSH-automorphism.186

Theorem 3.3.3. The elements {hn }n=0,1,2,... , {en }n=0,1,2,... in Theorem 3.3.1 also satisfy the following.
(e) The elements hn , en in A satisfy the same relation (2.4.4)
X
(−1)i ei hj = δ0,n
i+j=n

as their counterparts in Λ, along with the property that

A = Z[h1 , h2 , . . .] = Z[e1 , e2 , . . .].

ω
(f) There is exactly one nontrivial automorphism A → A as a PSH, swapping hn ↔ en .
(g) There are exactly two PSH-isomorphisms A → Λ,
• one sending hn to the complete homogeneous symmetric functions hn (x), while sending en to
the elementary symmetric functions en (x),
• the second one (obtained by composing the first with ω) sending hn 7→ en (x) and en 7→ hn (x).

Before embarking on the proof, we mention one more bit of convenient terminology: say that an element
σ in Σ is a constituent
P of a in NΣ when σ ≤ a, that is, σ appears with nonzero coefficient cσ in the unique
expansion a = τ ∈Σ cτ τ .

Proof of Theorem 3.3.1. One fact that occurs frequently is this:

(3.3.1) Every σ in Σ ∩ An is a constituent of ρn .

This follows from Theorem 3.2.3, since ρ is the only primitive element of Σ: one has A = A(ρ) and Σ = Σ(ρ),
so that σ is a constituent of some ρm , and homogeneity considerations force m = n.
Notice that A is of finite type (due to (3.3.1)). Thus, Ao is a graded Hopf algebra isomorphic to A.
Assertion (a). Note that
(ρ2 , ρ2 ) = (ρ⊥ (ρ2 ), ρ) = (2ρ, ρ) = 2
⊥
using the fact thatP ρ is a derivation since ρ is primitive (Proposition 2.8.2(iii)). On the other hand,
2
expressing ρ = σ∈Σ cσ σ with cσ in N, one has (ρ2 , ρ2 ) = σ c2σ . Hence exactly two of the cσ = 1, so ρ2
P
has exactly two distinct constituents. Denote them by h2 and e2 . One concludes that Σ ∩ A2 = {h2 , e2 }
from (3.3.1).
Note also that the same argument shows Σ ∩ A1 = {ρ}, so that A1 = Zρ. Since ρ⊥ h2 lies in A1 = Zρ and
(ρ h2 , ρ) = (h2 , ρ2 ) = 1, we have ρ⊥ h2 = ρ. Similarly ρ⊥ e2 = ρ.
⊥

Assertion (b). We will show via induction on n the following three assertions for n ≥ 1:

• There exists an element hn in Σ ∩ An with e⊥

2 hn = 0.
(3.3.2) • This element hn is unique.
• Furthermore ρ⊥ hn = hn−1 .

In the base cases n = 1, 2, it is not hard to check that our previously labelled elements, h1 , h2 (namely
h1 := ρ, and h2 as named in part (a)) really are the unique elements satisfying these hypotheses.

186The reader should be warned that not every invertible PSH-endomorphism is necessarily a PSH-automorphism. For
P
instance, it is an easy exercise to check that Λ ⊗ Λ → Λ ⊗ Λ, f ⊗ g 7→ (f ) f1 ⊗ f2 g is a well-defined invertible PSH-
endomorphism of the PSH Λ ⊗ Λ with PSH-basis (sλ ⊗ sµ )(λ,µ)∈Par × Par , but not a PSH-automorphism.
90 DARIJ GRINBERG AND VICTOR REINER

In the inductive step, it turns out that we will find hn as a constituent of ρhn−1 . Thus we again use the
derivation property of ρ⊥ to compute that ρhn−1 has exactly two constituents:
(ρhn−1 , ρhn−1 ) = (ρ⊥ (ρhn−1 ), hn−1 )
= (hn−1 + ρ · ρ⊥ hn−1 , hn−1 )
= (hn−1 + ρhn−2 , hn−1 )
= 1 + (hn−2 , ρ⊥ hn−1 )
= 1 + (hn−2 , hn−2 ) = 1 + 1 = 2
⊥
where the inductive hypothesis ρ hn−1 = hn−2 was used twice. We next show that exactly one of the two
constituents of ρhn−1 is annihilated by e⊥
2 . Note that since e2 lies in A2 , and A1 has Z-basis element ρ,
there is a constant c in Z such that
(3.3.3) ∆(e2 ) = e2 ⊗ 1 + cρ ⊗ ρ + 1 ⊗ e2 .
On the other hand, (a) showed
1 = (e2 , ρ2 )A = (∆(e2 ), ρ ⊗ ρ)A⊗A
so one must have c = 1. Therefore by Proposition 2.8.2(iii) again,
e⊥
2 (ρhn−1 ) = e⊥
2 (ρ)hn−1 + ρ⊥ (ρ)ρ⊥ (hn−1 ) + ρe⊥
2 (hn−1 )
(3.3.4) = 0 + hn−2 + 0
= hn−2 ,
where the first term vanished due to degree considerations and the last term vanished by the inductive
hypothesis. Bearing in mind that ρhn−1 lies in NΣ, and in a PSH with PSH-basis Σ, any skewing operator
σ ⊥ for σ in Σ will preserve NΣ, one concludes from (3.3.4) that
• one of the two distinct constituents of the element ρhn−1 must be sent by e⊥ 2 to hn−2 , and
• the other constituent of ρhn−1 must be annihilated by e⊥ 2 ; call this second constituent hn .
Lastly, to see that this hn is unique, it suffices to show that any element σ of Σ ∩ An which is killed by
e⊥ n
2 must be a constituent of ρhn−1 . This holds for the following reason. We know σ ≤ ρ by (3.3.1), and
n−1 ⊥
n
hence 0 6= (ρ , σ) = (ρ , ρ σ), implying that ρ σ 6= 0. On the other hand, since 0 = ρ⊥ e⊥
⊥ ⊥ ⊥
2 σ = e2 ρ σ,
⊥ ⊥ ⊥
one has that ρ σ is annihilated by e2 , and hence ρ σ must be a (positive) multiple of hn−1 by part of our
inductive hypothesis. Therefore (σ, ρhn−1 ) = (ρ⊥ σ, hn−1 ) is positive, that is, σ is a constituent of ρhn−1 .
The preceding argument, applied to σ = hn , shows that ρ⊥ hn = chn−1 for some c in {1, 2, . . .}. Since
(ρ hn , hn−1 ) = (hn , ρhn−1 ) = 1, this c must be 1, so that ρ⊥ hn = hn−1 . This completes the induction step
⊥

in the proof of (3.3.2).

One can then argue, swapping the roles of en , hn in the above argument, the existence and uniqueness of
a sequence {en }∞ n=0 in Σ satisfying the properties analogous to (3.3.2), with e0 := 1, e1 := ρ.

Assertion (c). Iterating the property from (b) that ρ⊥ hn = hn−1 shows that (ρk )⊥ hn = hn−k for 0 ≤ k ≤ n.
However one also has an expansion X
ρk = chk + cσ σ
σ∈Σ∩Ak :
σ6=hk

for some integers c, cσ > 0, since every σ in Σ ∩ Ak is a constituent of ρk . Hence

1 = (hn−k , hn−k ) = ((ρk )⊥ hn , (ρk )⊥ hn ) ≥ c2 (h⊥ ⊥
k hn , hk hn )

using Proposition 2.8.2(ii). Hence if we knew that h⊥

k hn 6= 0 this would force

h⊥ k ⊥
k hn = (ρ ) hn = hn−k

as well as σ ⊥ hn = 0 for all σ 6∈ {h0 , h1 , . . . , hn }. But

(ρn−k )⊥ h⊥ ⊥ n−k ⊥
k hn = hk (ρ ) hn = h⊥
k hk = 1 6= 0

so h⊥ ⊥
k hn 6= 0, as desired. The argument for ek en = en−k is symmetric.
The last assertion in (c) follows if one checks that en 6= hn for each n ≥ 2, but this holds since e⊥
2 (hn ) = 0
but e⊥
2 (en ) = en−2 .
 HOPF ALGEBRAS IN COMBINATORICS 91

Assertion (d). Part (c) implies that

(∆hn , σ ⊗ τ )A⊗A = (hn , στ )A = (σ ⊥ hn , τ )A = 0
unless σ = hk for some k = 0, 1, 2, . . . , n and τ = hn−k . Also one can compute
(c)
(∆hn , hk ⊗ hn−k ) = (hn , hk hn−k ) = (h⊥
k hn , hn−k ) = (hn−k , hn−k ) = 1.

This is equivalent to the assertion for ∆hn in (d). The argument for ∆en is symmetric. 
Before proving Theorem 3.3.3, we note some consequences of Theorem 3.3.1. Define for each partition
λ = (λ1 ≥ λ2 ≥ · · · ≥ λ` ) the elements of A
hλ = hλ1 hλ2 · · · ,
eλ = eλ1 eλ2 · · · .
Also, define the lexicographic order on Parn by saying λ <lex µ if λ 6= µ and the smallest index i for which
λi 6= µi has λi < µi . Recall also that λt denotes the conjugate or transpose partition to λ, obtained by
swapping rows and columns in the Ferrers diagram.
The following unitriangularity lemma will play a role in the proof of Theorem 3.3.3(e).

Lemma 3.3.4. Under the hypotheses of Theorem 3.3.1, for λ, µ in Parn , one has
(
⊥ 1 if µ = λt
(3.3.5) eµ hλ =
0 if µ >lex λt .
Consequently
(3.3.6) det [(eµt , hλ )]λ,µ∈Parn = 1.

Proof. Notice that A is of finite type (as shown in the proof of Theorem 3.3.1). Thus, Ao is a graded Hopf
algebra isomorphic to A.
Also, notice that any m ∈ N and any a1 , a2 , . . . , a` ∈ A satisfy
X
(3.3.7) e⊥
m (a1 a2 · · · a` ) = e⊥ ⊥
i1 (a1 ) · · · ei` (a` ) .
i1 +···+i` =m

Indeed, this follows by induction over ` using Proposition 2.8.2(iii) (and the coproduct formula for ∆(en ) in
Theorem 3.3.1(d)).
In order to prove (3.3.5), induct on the length of µ. If λ has length `, so that λt1 = `, then
e⊥ ⊥ ⊥


µ hλ = e(µ2 ,µ3 ,...) eµ1 (hλ1 · · · hλ` )
X
= e⊥
(µ2 ,µ3 ,...) e⊥ ⊥
i1 (hλ1 ) · · · ei` (hλ` ) (by (3.3.7))
i1 +···+i` =µ1
X
e⊥ e⊥ ⊥
since e⊥


= (µ2 ,µ3 ,...) i1 (hλ1 ) · · · ei` (hλ` ) k hn = 0 for k ≥ 2
i1 +···+i` =µ1 ;
each of i1 ,...,i` is ≤1
(
0 if µ1 > ` = λt1
=
e⊥
(µ2 ,µ3 ,...) h(λ1 −1,...,λ` −1) if µ1 = ` = λt1
where the last equality used (
hn−1 if k = 1,
e⊥
k (hn ) =
0 if k ≥ 2.
Now apply the induction hypothesis, since (λ1 − 1, . . . , λ` − 1)t = (λt2 , λt3 , . . .).
To prove (3.3.6), note that any λ, µ in Parn satisfy (eµt , hλ ) = (e⊥ ⊥
µt (hλ ), 1) = eµt (hλ ) (since degree
⊥
considerations enforce eµt (hλ ) ∈ A0 = k · 1), and thus
(
⊥ 1 if µt = λt
(eµt , hλ ) = eµt (hλ ) =
0 if µt >lex λt
92 DARIJ GRINBERG AND VICTOR REINER

(by (3.3.5)). This means that the matrix [(eµt , hλ )]λ,µ∈Parn is unitriangular with respect to some total order
on Parn (namely, the lexicographic order on the conjugate partitions), and hence has determinant 1. 
The following proposition will be the crux of the proof of Theorem 3.3.3(f) and (g), and turns out to be
closely related to Kerov’s asymptotic theory of characters of the symmetric groups [93].

L Given a PSH A with PSH-basis Σ containing only one primitive ρ, the two maps A → Z
Proposition 3.3.5.
defined on A = n≥0 An via
M
δh = h⊥n,
n
M
δe = e⊥
n
n
δ
are characterized as the only two Z-linear maps A → Z with the three properties of being
• positive: δ(NΣ) ⊂ N,
• multiplicative: δ(a1 a2 ) = δ(a1 )δ(a2 ), and
• normalized: δ(ρ) = 1.
Proof. Notice that A is of finite type (as shown in the proof of Theorem 3.3.1). Thus, Ao is a graded Hopf
algebra isomorphic to A.
It should be clear from their definitions that δh , δe are Z-linear, positive and normalized. To see that δh
is multiplicative, by Z-linearity, it suffices to check that for a1 , a2 in An1 , An2 with n1 + n2 = n, one has
X
δh (a1 a2 ) = h⊥
n (a1 a2 ) = h⊥ ⊥ ⊥ ⊥
i1 (a1 )hi2 (a2 ) = hn1 (a1 )hn2 (a2 ) = δh (a1 )δh (a2 )
i1 +i2 =n

in which the second equality used Proposition 2.8.2(iii) and Theorem 3.3.1(d). The argument for δe is
symmetric.
δ
Conversely, given A → Z which is Z-linear, positive, multiplicative, and normalized, note that
δ(h2 ) + δ(e2 ) = δ(h2 + e2 ) = δ(ρ2 ) = δ(ρ)2 = 12 = 1
and hence positivity implies that either δ(h2 ) = 0 or δ(e2 ) = 0. Assume the latter holds, and we will show
that δ = δh .
Given any σ in Σ ∩ An \ {hn }, note that e⊥ ⊥
2 σ 6= 0 by Theorem 3.3.1(b), and hence 0 6= (e2 σ, ρ
n−2
)=
n−2 n−2
(σ, e2 ρ ). Thus σ is a constituent of e2 ρ , so positivity implies
0 ≤ δ(σ) ≤ δ(e2 ρn−2 ) = δ(e2 )δ(ρn−2 ) = 0.
Thus δ(σ) = 0 for σ in Σ ∩ An \ {hn }. Since δ(ρn ) = δ(ρ)n = 1n = 1, this forces δ(hn ) = 1, for each n ≥ 0
(including n = 0, as 1 = δ(ρ) = δ(ρ · 1) = δ(ρ)δ(1) = 1 · δ(1) = δ(1)). Thus δ = δh . The argument when
δ(h2 ) = 0 showing δ = δe is symmetric. 
Proof of Theorem 3.3.3. Many of the assertions of parts (e) and (f) will come from constructing the unique
nontrivial PSH-automorphism ω of A from the antipode S: for homogeneous a in An , define ω(a) :=
(−1)n S(a). We now study some of the properties of S and ω.
Notice that A is of finite type (as shown in the proof of Theorem 3.3.1). Thus, Ao is a graded Hopf algebra
isomorphic to A.
Since A is a PSH, it is commutative by Theorem 3.1.7 (applied to A ⊗Z Q). This implies both that S, ω
are actually algebra endomorphisms by Proposition 1.4.8, and that S 2 = idA = ω 2 by Corollary 1.4.10.
Since A is self-dual and the defining diagram (1.4.3) satisfied by the antipode S is sent to itself when
one replaces A by Ao and all maps by their adjoints, one concludes that S = S ∗ (where S ∗ means the
restricted adjoint S ∗ : Ao → Ao ), i.e., S is self-adjoint. Since S is an algebra endomorphism, and S = S ∗ ,
in fact S is also a coalgebra endomorphism, a bialgebra endomorphism, and a Hopf endomorphism (by
Proposition 1.4.24(c)). The same properties are shared by ω.
Since idA = S 2 = SS ∗ , one concludes that S is an isometry, and hence so is ω.
Since ρ is primitive, one has S(ρ) = −ρ and ω(ρ) = ρ. Therefore ω(ρn ) = ρn for n = 1, 2, . . .. Use this as
follows to check that ω is a PSH-automorphism, which amounts to checking that every σ in Σ has ω(σ) in
Σ:
(ω(σ), ω(σ)) = (σ, σ) = 1
 HOPF ALGEBRAS IN COMBINATORICS 93

so that ±ω(σ) lies in Σ, but also if σ lies in An , then

(ω(σ), ρn ) = (σ, ω(ρn )) = (σ, ρn ) > 0.
In summary, ω is a PSH-automorphism of A, an isometry, and an involution.
Let us try to determine the action of ω on the {hn }. By similar reasoning as in (3.3.3), one has
∆(h2 ) = h2 ⊗ 1 + ρ ⊗ ρ + 1 ⊗ h2 .
Thus 0 = S(h2 ) + S(ρ)ρ + h2 , and combining this with S(ρ) = −ρ, one has S(h2 ) = e2 . Thus also
ω(h2 ) = (−1)2 S(h2 ) = e2 .
We claim that this forces ω(hn ) = en , because h⊥
2 ω(hn ) = 0 via the following calculation: for any a in A
one has
(h⊥2 ω(hn ), a) = (ω(hn ), h2 a)
= (hn , ω(h2 a))
= (hn , e2 ω(a))
= (e⊥
2 hn , ω(a)) = (0, ω(a)) = 0.
Consequently the involution ω swaps hn and en , while the antipode S has S(hn ) = (−1)n en and S(en ) =
(−1)n hn . Thus the coproduct formulas in (d) and definition of the antipode S imply the relation (2.4.4)
between {hn } and {en }.
This relation (2.4.4) also lets one recursively express the hn as polynomials with integer coefficients in the
{en }, and vice-versa, so that {hn } and {en } each generate the same Z-subalgebra A0 of A. We wish to show
that A0 exhausts A.
We argue that Lemma 3.3.4 implies the Gram matrix [(hµ , hλ )]µ,λ∈Parn has determinant ±1 as follows.
Since {hn } and {en } both generate A0 , there exists a Z-matrix (aµ,λ ) expressing eµt = λ aµ,λ hλ , and one
P
has
[(eµt , hλ )] = [aµ,λ ] · [(hµ , hλ )] .
Taking determinants of these three Z-matrices, and using the fact that the determinant on the left is 1 (by
(3.3.6)), both determinants on the right must also be ±1.
Now we will show that every σ ∈ Σ ∩ An lies in A0n . Uniquely express σ = σ 0 + σ 00 in which σ 0 lies in
the R-span RA0n and σ 00 lies in the real perpendicular space (RA0n )⊥ inside R ⊗Z An . One can compute
0
P
R-coefficients (cµ )µ∈Parn that express σ = µ cµ hµ by solving the system
!
X
cµ hµ , hλ = (σ, hλ ) for λ ∈ Parn .
µ

This linear system is governed by the Gram matrix [(hµ , hλ )]µ,λ∈Parn with determinant ±1, and its right
side has Z-entries since σ, hλ lie in A. Hence the solution (cµ )µ∈Parn will have Z-entries, so σ 0 lies in A0 .
Furthermore, σ 00 = σ − σ 0 will lie in A, and hence by the orthogonality of σ 0 , σ 00 ,
1 = (σ, σ) = (σ 0 , σ 0 ) + (σ 00 , σ 00 ).
One concludes that either σ 00 = 0, or σ 0 = 0. The latter cannot occur since it would mean that σ = σ 00 is
perpendicular to all of A0 . But ρn = hn1 lies in A0 , and (σ, ρn ) 6= 0. Thus σ 00 = 0, meaning σ = σ 0 lies in A0 .
This completes the proof of assertion (e). Note that in the process, having shown det(hµ , hλ )λ,µ∈Parn = ±1,
one also knows that {hλ }λ∈Parn are Z-linearly independent, so that {h1 , h2 , . . .} are algebraically independent,
and A = Z[h1 , h2 , . . .] is the polynomial algebra generated by {h1 , h2 , . . .}.
For assertion (f), we have seen that ω gives such a PSH-automorphism A → A, swapping hn ↔ en .
ϕ
Conversely, given a PSH-automorphism A → A, consider the positive, multiplicative, normalized Z-linear
map δ := δh ◦ ϕ : A → Z. Proposition 3.3.5 shows that either
• δ = δh , which then forces ϕ(hn ) = hn for all n, so ϕ = idA , or
• δ = δe , which then forces ϕ(en ) = hn for all n, so ϕ = ω.
For assertion (g), given a PSH A with PSH-basis Σ having exactly one primitive ρ, since we have seen
A = Z[h1 , h2 , . . .], where hn in A is as defined in Theorem 3.3.1, one can uniquely define an algebra morphism
ϕ
A → Λ that sends the element hn to the complete homogeneous symmetric function hn (x). Assertions (d)
and (e) show that ϕ is a bialgebra isomorphism, and hence it is a Hopf isomorphism. To show that it is a
94 DARIJ GRINBERG AND VICTOR REINER

PSH-isomorphism, we first note that it is an isometry because one can iterate Proposition 2.8.2(iii) together
with assertions (c) and (d) to compute all inner products
(hµ , hλ )A = (1, h⊥ ⊥ ⊥
µ hλ )A = (1, hµ1 hµ2 · · · (hλ1 hλ2 · · · ))A
for µ, λ in Parn . Hence
(hµ , hλ )A = (hµ (x), hλ (x))Λ = (ϕ(hµ ), ϕ(hλ ))Λ
Once one knows ϕ is an isometry, then elements ω in Σ ∩ An are characterized in terms of the form (·, ·) by
(ω, ω) = 1 and (ω, ρn ) > 0. Hence ϕ sends each σ in Σ to a Schur function sλ , and is a PSH-isomorphism. 
 HOPF ALGEBRAS IN COMBINATORICS 95

4. Complex representations for Sn , wreath products, GLn (Fq )

After reviewing the basics that we will need from representation and character theory of finite groups,
we give Zelevinsky’s three main examples of PSH’s arising as spaces of virtual characters for three towers of
finite groups:
• symmetric groups,
• their wreath products with any finite group, and
• the finite general linear groups.
Much in this chapter traces its roots to Zelevinsky’s book [203]. The results concerning the symmetric
groups, however, are significantly older and spread across the literature: see, e.g., [183, §7.18], [60, §7.3],
[125, §I.7], [165, §4.7], [98], for proofs using different tools.
4.1. Review of complex character theory. We shall now briefly discuss some basics of representation
(and character) theory that will be used below. A good source for this material, including the crucial Mackey
formula, is Serre [175, Chaps. 1-7].187
ϕ
4.1.1. Basic definitions, Maschke, Schur. For a group G, a representation is a homomorphism G → GL(V )
for some vector space V over a field. We will take the field to be C from now on, and we will also assume
that V is finite-dimensional over C. Thus a representation of G is the same as a finite-dimensional (left-
)CG-module V . (We use the notations CG and C [G] synonymously for the group algebra of G over C. More
generally, if S is a set, then CS = C [S] denotes the free C-module with basis S.)
We also assume that G is finite, so that Maschke’s Theorem188 says that CG is semisimple, meaning that
every CG-module U ⊂ V has a CG-module complement U 0 with V = U ⊕ U 0 . Equivalently, indecomposable
CG-modules are the same thing as simple (=irreducible) CG-modules.
Schur’s Lemma implies that for two simple CG-modules V1 , V2 , one has
(
C if V1 ∼ = V2 ,
HomCG (V1 , V2 ) ∼
=
0 if V1 6∼
= V2 .
4.1.2. Characters and Hom spaces. A CG-module V is completely determined up to isomorphism by its
character
χV
G −→ C
g 7−→ χV (g) := trace(g : V → V ).
This character χV is a class function, meaning it is constant on G-conjugacy classes. The space RC (G) of
class functions G → C has a Hermitian, positive definite form
1 X
(f1 , f2 )G := f1 (g)f2 (g).
|G|
g∈G

For any two CG-modules V1 , V2 ,

(4.1.1) (χV1 , χV2 )G = dimC HomCG (V1 , V2 ).
The set of all irreducible characters
Irr(G) = {χV : V is a simple CG-module}
forms an orthonormal basis of RC (G) with respect to this form, and spans a Z-sublattice
R(G) := Z Irr(G) ⊂ RC (G)
sometimes called the virtual characters of G. For every CG-module V , the character χV belongs to R(G).
Instead of working with the Hermitian form (·, ·)G on G, we could also (and some authors do) define a
C-bilinear form h·, ·iG on RC (G) by
1 X
hf1 , f2 iG := f1 (g)f2 (g −1 ).
|G|
g∈G

187More advanced treatments of representation theory can be found in [198] and [56].
188... which has a beautiful generalization to finite-dimensional Hopf algebras due to Larson and Sweedler; see Montgomery
[139, §2.2].
96 DARIJ GRINBERG AND VICTOR REINER

This form is not identical with (·, ·)G (indeed, h·, ·iG is bilinear while (·, ·)G is Hermitian), but it still satisfies
(4.1.1), and thus is identical with (·, ·)G on R(G) × R(G). Hence, for all we are going to do until Section 4.9,
we could just as well use the form h·, ·iG instead of (·, ·)G .

4.1.3. Tensor products. Given two groups G1 , G2 and CGi -modules Vi for i = 1, 2, their tensor product
V1 ⊗C V2 becomes a C[G1 × G2 ]-module via (g1 , g2 )(v1 ⊗ v2 ) = g1 (v1 ) ⊗ g2 (v2 ). This module is called the
(outer) tensor product of V1 and V2 . When V1 , V2 are both simple, then so is V1 ⊗ V2 , and every simple
C[G1 × G2 ]-module arises this way (with V1 and V2 determined uniquely up to isomorphism).189 Thus one
has identifications and isomorphisms
Irr(G1 × G2 ) = Irr(G1 ) × Irr(G2 ),
R(G1 × G2 ) ∼= R(G1 ) ⊗Z R(G2 );
here, χV1 ⊗ χV2 ∈ R(G1 ) ⊗Z R(G2 ) is being identified with χV1 ⊗V2 ∈ R(G1 × G2 ) for all CG1 -modules V1 and
all CG2 -modules V2 . The latter isomorphism is actually a restriction of the isomorphism RC (G1 × G2 ) ∼ =
RC (G1 ) ⊗C RC (G2 ) under which every pure tensor φ1 ⊗ φ2 ∈ RC (G1 ) ⊗C RC (G2 ) corresponds to the class
function G1 × G2 → C, (g1 , g2 ) 7→ φ1 (g1 ) ⊗ φ2 (g2 ).
Given two CG1 -modules V1 and W1 and two CG2 -modules V2 and W2 , we have
(4.1.2) (χV1 ⊗V2 , χW1 ⊗W2 )G1 ×G2 = (χV1 , χW1 )G1 (χV2 , χW2 )G2 .

4.1.4. Induction and restriction. Given a subgroup H < G and CH-module U , one can use the fact that CG
is a (CG, CH)-bimodule to form the induced CG-module
IndG
H U := CG ⊗CH U.

The fact that CG is free as a (right-)CH-module190 on basis elements {tg }gH∈G/H makes this tensor product
easy to analyze. For example one can compute its character
1 X
(4.1.3) χIndG
H U
(g) = χU (kgk −1 ).
|H|
k∈G:
kgk−1 ∈H

191
One can also recognize when a CG-module V is isomorphic to IndG H U for some CH-module
L U : this
happens if and only if there is an H-stable subspace U ⊂ V having the property that V = gH∈G/H gU .
The above construction of a CG-module IndG
H U corresponding to any CH-module U is part of a functor
IndG
H from the category of CH-modules to the category of CG-modules
192
; this functor is called induction.
Besides induction on CH-modules, one can define induction on class functions of H:
Exercise 4.1.1. Let G be a finite group, and H a subgroup of G. Let f ∈ RC (H) be a class function. We
define the induction IndG
H f of f to be the function G → C given by
  1 X
IndG f kgk −1


(4.1.4) H f (g) = for all g ∈ G.
|H|
k∈G:
kgk−1 ∈H

(a) Prove that this induction IndG

H f is a class function on G, hence belongsF to RC (G).
(b) Let J be a system of right coset representatives for H\G, so that G = j∈J Hj. Prove that
  X
IndG f jgj −1


H f (g) = for all g ∈ G.
j∈J:
jgj −1 ∈H

189This is proven in [175, §3.2, Thm. 10]. The fact that C is algebraically closed is essential for this!
190... which also has a beautiful generalization to finite-dimensional Hopf algebras due to Nichols and Zoeller; see [139, §3.1].
191See [175, §7.2, Prop. 20(ii)] for the proof of this equality. (Another proof is given in [56, Remark 5.9.2 (the Remark after
Theorem 4.32 in the arXiv version)], but [56] uses a different definition of IndG
H U ; see Remark 4.1.5 for why it is equivalent to
ours. Yet another proof of (4.1.3) is given in Exercise 4.1.14(k).)
192On morphisms, it sends any f : U → U 0 to id 0
CG ⊗CH f : CG ⊗CH U → CG ⊗CH U .
 HOPF ALGEBRAS IN COMBINATORICS 97

The induction IndG

H defined in Exercise 4.1.1 is a C-linear map RC (H) → RC (G). Since every CH-module
U satisfies
(4.1.5) χIndG
H U
= IndG
H (χU )

, this C-linear map IndG

193 G
H restricts to a Z-linear map R (H) → R (G) (also denoted IndH ) which sends the
character χU of any CH-module U to the character χIndG H U
of the induced CG-module IndG H U.

Exercise 4.1.2. Let G, H and I be three finite groups such that I < H < G. Let U be a CI-module. Prove
that IndG H ∼ G
H IndI U = IndI U . (This fact is often referred to as the transitivity of induction.)

Exercise 4.1.3. Let G1 and G2 be two groups. Let H1 < G1 and H2 < G2 be two subgroups. Let U1 be a
CH1 -module, and U2 be a CH2 -module. Show that
   
IndG1 ×G2 ∼ G1 G2
(4.1.6) H1 ×H2 (U1 ⊗ U2 ) = IndH1 U1 ⊗ IndH2 U2

as C [G1 × G2 ]-modules.
The restriction operation V 7→ ResG
H V restricts a CG-module V to a CH-module. Frobenius reciprocity
asserts the adjointness between IndG
H and ResG
H

(4.1.7) HomCG (IndG ∼ G

H U, V ) = HomCH (U, ResH V ),

as a special case (S = A = CG, R = CH, B = U, C = V ) of the general adjoint associativity

(4.1.8) HomS (A ⊗R B, C) ∼
= HomR (B, HomS (A, C))
for S, R two rings, A an (S, R)-bimodule, B a left R-module, C a left S-module.
We can define not just the restriction of a CG-module, but also the restriction of a class function f ∈
RC (G). When H is a subgroup of G, the restriction ResG H f of an f ∈ RC (G) is defined as the result of
restricting the map f : G → C to H. This ResG H f is easily seen to belong to RC (H), and so ResGH is
a C-linear map RC (G) → RC (H). This map restricts to a Z-linear map R(G) → R(H), since we have
ResGH χV = χResG H V
for any CG-module V . Taking characters in (4.1.7) (and recalling ResGH χV = χResGH V
and (4.1.5)), we obtain
(4.1.9) (IndG G
H χU , χV )G = (χU , ResH χV )H .

By bilinearity, this yields the equality

 
IndG = α, ResG


H α, β Hβ H
G

for any class functions α ∈ RC (H) and β ∈ RC (G) (since R(G) spans RC (G) as a C-vector space).
Exercise 4.1.4. Let G be a finite group, and let H < G. Let U be a CH-module. If A and B are two
algebras, P is a (B, A)-bimodule and Q is a left B-module, then HomB (P, Q) is a left A-module (since CG
is a (CH, CG)-bimodule). As a consequence, HomCH (CG, U ) is a CG-module. Prove that this CG-module
is isomorphic to IndGH U.

Remark 4.1.5. Some texts define the induction IndG H U of a CH-module U to be HomCH (CG, U ) (rather
than to be CG ⊗CH U , as we did).194 As Exercise 4.1.4 shows, this definition is equivalent to ours as long as
G is finite (but not otherwise).
Exercise 4.1.4 yields the following “wrong-way” version of Frobenius reciprocity:
Exercise 4.1.6. Let
 G be a finite
 group; let H < G. Let U be a CG-module, and let V be a CH-module.
Prove that HomCG U, IndG ∼ G


H V = Hom CH ResH U, V .

193This follows by comparing the value of χ G

IndG
H
U (g) obtained from (4.1.3) with the value of IndH (χU ) (g) found using
(4.1.4).
194Or they define it as a set of morphisms of H-sets from G to U (this is how [56, Def. 5.8.1 (Def. 4.28 in the arXiv version)]
defines it); this is easily seen to be equivalent to HomCH (CG, U ).
98 DARIJ GRINBERG AND VICTOR REINER

4.1.5. Mackey’s formula. Mackey gave an alternate description of a module which has been induced and
then restricted. To state it, for a subgroup H < G and g in G, let H g := g −1 Hg and g H := gHg −1 . Given
ϕ
a CH-module U , say defined by a homomorphism H → GL(U ), let U g denote the C[gHg −1 ]-module on the
same C-vector space U defined by the composite homomorphism
g ϕ
H −→ H −→ GL(U ).
h 7−→ g −1 hg

Theorem 4.1.7. (Mackey’s formula) Consider subgroups H, K < G, and any CH-module U . If {g1 , . . . , gt }
are double coset representatives for K\G/H, then
t 
g i 
∼
M
G
ResG
K IndH U = IndK
gi H∩K ResH
H∩K gi U
i=1

Proof. In this proof, all tensor product symbols ⊗ should be interpreted as ⊗CH . Recall CG has C-basis
{tg }g∈G . For subsets S ⊂ G, let C[S] denote the C-span of {tg }g∈S in CG.
Note that each double coset KgH gives rise to a sub-(K, H)-bimodule C[KgH] within CG, and one has
a CK-module direct sum decomposition
t
M
IndG
H U = CG ⊗ U = C[Kgi H] ⊗ U.
i=1
Hence it suffices to check for any element g in G that

g 
C[KgH] ⊗ U ∼
= IndK
g H∩K ResH
H∩K g U .
Note that g H ∩ K is the subgroup of K consisting of the elements k in K for which kgH = gH. Hence by
picking {k1 , . . . , ks } to be coset representatives for K/(g H ∩ K), one disjointly decomposes the double coset
s
G
KgH = kj (g H ∩ K)gH,
j=1

giving a C-vector space direct sum decomposition

M s
C[KgH] ⊗ U = C [kj (g H ∩ K) gH] ⊗ U
j=1
∼
= IndK g
g H∩K (C[( H ∩ K) gH] ⊗ U ) .

So it remains to check that one has a C[g H ∩ K]-module isomorphism

g
C[(g H ∩ K) gH] ⊗ U ∼
= ResH
H∩K g U .
g −1
Bearing in mind that, for each k in H ∩ K and h in H, one has g kg in H and hence
−1
tkgh ⊗ u = tg · tg−1 kg·h ⊗ u = tg ⊗ g kgh · u
one sees that this isomorphism can be defined by mapping
tkgh ⊗ u 7−→ g −1 kgh · u.

4.1.6. Inflation and fixed points. There are two (adjoint) constructions on representations that apply when
ϕ
one has a normal subgroup K / G. Given a C[G/K]-module U , say defined by the homomorphism G/K →
G
GL(U ), the inflation of U to a CG-module InflG/K U has the same underlying space U , and is defined by
ϕ
the composite homomorphism G → G/K → GL(U ). We will later use the easily-checked fact that when
H < G is any other subgroup, one has
G H G/K
(4.1.10) ResG
H InflG/K U = InflH/H∩K ResH/H∩K U.

(We regard H/H ∩ K as a subgroup of G/K, since the canonical homomorphism H/H ∩ K → G/K is
injective.)
 HOPF ALGEBRAS IN COMBINATORICS 99

Inflation turns out to be adjoint to the K-fixed space construction sending a CG-module V to the C[G/K]-
module
V K := {v ∈ V : k(v) = v for k ∈ K}
Note that V K is indeed a G-stable subspace: for any v in V K and g in G, one has that g(v) lies in V K since
an element k in K satisfies kg(v) = g · g −1 kg(v) = g(v) as g −1 kg lies in K. One has this adjointness
(4.1.11) HomCG (InflG K
G/K U, V ) = HomC[G/K] (U, V )

because any CG-module homomorphism ϕ on the left must have the property that kϕ(u) = ϕ(k(u)) = ϕ(u)
for all k in K, so that ϕ actually lies on the right.
We will also need the following formula for the character χV K in terms of the character χV :
1 X
(4.1.12) χV K (gK) = χV (gk).
|K|
k∈K

To see this, note that when one has a C-linear endomorphism ϕ on a space V that preserves some C-subspace
π
W ⊂ V , if V → W is any idempotent projection onto W , then the tracePof the restriction ϕ|W equals the
1 195
trace of ϕ ◦ π on V . Applying this to W = V K and ϕ = g, with π = |K| k∈K k, gives (4.1.12).
Another way to restate (4.1.12) is:
1 X
(4.1.13) χV K (gK) = χV (h).
|K|
h∈gK

Inflation and K-fixed space construction can also be defined on class functions. For inflation, this is par-
ticularly easy: Inflation InflG
G/K f of an f ∈ RC (G/K) is defined as the composition G
/ / G/K f
/C.
This is a class function of G and thus lies in RC (G). Thus, inflation InflG
is a C-linear map RC (G/K) →
G/K
RC (G). It restricts to a Z-linear map R(G/K) → R(G), since it is clear that every C(G/K)-module U
satisfies InflG
G/K χU = χInflG U.
G/K
We can also use (4.1.12) (or (4.1.13)) as inspiration for defining a “K-fixed space construction” on class
functions. Explicitly, for every class function f ∈ RC (G), we define a class function f K ∈ RC (G/K) by
1 X 1 X
f K (gK) = f (gk) = f (h).
|K| |K|
k∈K h∈gK

K
The map (·) : RC (G) → RC (G/K), f 7→ f K is C-linear, and restricts to a Z-linear map R(G) → R(G/K).
K
Again, we have a compatibility with the K-fixed point construction on modules: We have χV K = (χV ) for
every CG-module V .
Taking characters in (4.1.11), we obtain
(4.1.14) (InflG K
G/K χU , χV )G = (χU , χV )G/K

K
for any C [G/K]-module U and any CG-module V (since χInflG U = InflG
G/K χU and χV K = (χV ) ). By
G/K
Z-linearity, this implies that
 
InflG = α, β K G/K


G/K α, β
G
for any class functions α ∈ RC (G/K) and β ∈ RC (G).
There is also an analogue of (4.1.6):

Lemma 4.1.8. Let G1 and G2 be two groups, and K1 < G1 and K2 < G2 be two respective subgroups.
Let Ui be a CGi -module for each i ∈ {1, 2}. Then,
K1 ×K2
(4.1.15) (U1 ⊗ U2 ) = U1K1 ⊗ U2K2
(as subspaces of U1 ⊗ U2 ).

195For another proof of (4.1.12), see Exercise 4.1.14(l).

100 DARIJ GRINBERG AND VICTOR REINER

K
Proof. The subgroup K1 = K1 ×1 of G1 ×G2 acts on U1 ⊗U2 , and its fixed points are (U1 ⊗ U2 ) 1 = U1K1 ⊗U2
(because for a CK1 -module, tensoring with U2 is the same as taking a direct power, which clearly commutes
K
with taking fixed points). Similarly, (U1 ⊗ U2 ) 2 = U1 ⊗ U2K2 . Now,
   
K ×K K K
(U1 ⊗ U2 ) 1 2 = (U1 ⊗ U2 ) 1 ∩ (U1 ⊗ U2 ) 2 = U1K1 ⊗ U2 ∩ U1 ⊗ U2K2 = U1K1 ⊗ U2K2
according to the known linear-algebraic fact stating that if P and Q are subspaces of two vector spaces U
and V , respectively, then (P ⊗ V ) ∩ (U ⊗ Q) = P ⊗ Q. 
Exercise 4.1.9. (a) Let G1 and G2 be two groups. Let Vi and Wi be finite-dimensional CGi -modules
for every i ∈ {1, 2}. Prove that the C-linear map
HomCG1 (V1 , W1 ) ⊗ HomCG2 (V2 , W2 ) → HomC[G1 ×G2 ] (V1 ⊗ V2 , W1 ⊗ W2 )
sending each tensor f ⊗g to the tensor product f ⊗g of homomorphisms is a vector space isomorphism.
(b) Use part (a) to give a new proof of (4.1.2).
As an aside, (4.1.10) has a “dual” analogue:
Exercise 4.1.10. Let G be a finite group, and let K/G and H < G. Let U be a CH-module. As usual, regard
 K
H/ (H ∩ K) as a subgroup of G/K. Show that IndG ∼ G/K

H U = IndH/(H∩K) U H∩K as C [G/K]-modules.

Inflation also “commutes” with induction:

Exercise 4.1.11. Let G be a finite group, and let K < H < G be such that K / G. Thus, automatically,
K / H, and we regard the quotient H/K as a subgroup of G/K. Let V be a C [H/K]-module. Show that
InflG
G/K ∼ G H
G/K IndH/K V = IndH InflH/K V as CG-modules.

Exercise 4.1.12. Let G be a finite group, and let K / G. Let V be a CG-module. Let IV,K denote the
C-vector subspace of V spanned by all elements of the form v − kv for k ∈ K and v ∈ V .
(a) Show that IV,K is a CG-submodule of V .
(b) Let VK denote the quotient CG-module V /IV,K . (This module is occasionally called the K-coinvariant
module of V , a name it sadly shares with at least two other non-equivalent constructions in algebra.)
∼ G K


Show that VK = InflG/K V as CG-modules. (Use char C = 0.)
In the remainder of this subsection, we shall briefly survey generalized notions of induction and restriction,
defined in terms of a group homomorphism ρ rather than in terms of a group G and a subgroup H. These
generalized notions (defined by van Leeuwen in [112, §2.2]) will not be used in the rest of these notes, but
they shed some new light on the facts about induction, restriction, inflation and fixed point construction
discussed above. (In particular, they reveal that some of said facts have common generalizations.)
The reader might have noticed that the definitions of inflation and of restriction (both for characters and
for modules) are similar. In fact, they both are particular cases of the following construction:
Remark 4.1.13. Let G and H be two finite groups, and let ρ : H → G be a group homomorphism.
• If f ∈ RC (G), then the ρ-restriction Resρ f of f is defined as the map f ◦ ρ : H → C. This map is
easily seen to belong to RC (H).
• If V is a CG-module, then the ρ-restriction Resρ V of V is the CH-module with ground space V
and action given by
h · v = ρ (h) · v for every h ∈ H and v ∈ V.
This construction generalizes both inflation and restriction: If H is a subgroup of G, and if ρ : H → G is
the inclusion map, then Resρ f = ResG G
H f (for any f ∈ RC (G)) and Resρ V = ResH V (for any CG-module
V ). If, instead, we have G = H/K for a normal subgroup K of H, and if ρ : H → G is the projection map,
then Resρ f = InflH H
H/K f (for any f ∈ RC (H/K)) and Resρ V = InflH/K V (for any C [H/K]-module V ).

A subtler observation is that induction and fixed point construction can be generalized by a common
notion. This is the subject of Exercise 4.1.14 below.
Exercise 4.1.14. Let G and H be two finite groups, and let ρ : H → G be a group homomorphism. We
introduce the following notations:
 HOPF ALGEBRAS IN COMBINATORICS 101

• If f ∈ RC (H), then the ρ-induction Indρ f of f is a map G → C which is defined as follows:

1 X
(Indρ f ) (g) = f (h) for every g ∈ G.
|H|
(h,k)∈H×G;
kρ(h)k−1 =g

• If U is a CH-module, then the ρ-induction Indρ U of U is defined as the CG-module CG ⊗CH U ,

where CG is regarded as a (CG, CH)-bimodule according to the following rule: The left CG-module
structure on CG is plain multiplication inside CG; the right CH-module structure on CG is induced
by the C-algebra homomorphism C [ρ] : CH → CG (thus, it is explicitly given by γη = γ · (C [ρ]) η
for all γ ∈ CG and η ∈ CH).
Prove the following properties of this construction:
(a) For every f ∈ RC (H), we have Indρ f ∈ RC (G).
(b) For any finite-dimensional CH-module U , we have χIndρ U = Indρ χU .
(c) If H is a subgroup of G, and if ρ : H → G is the inclusion map, then Indρ f = IndG H f for every
f ∈ RC (H).
(d) If H is a subgroup of G, and if ρ : H → G is the inclusion map, then Indρ U = IndG H U for every
CH-module U .
(e) If G = H/K for some normal subgroup K of H, and if ρ : H → G is the projection map, then
Indρ f = f K for every f ∈ RC (H).
(f) If G = H/K for some normal subgroup K of H, and if ρ : H → G is the projection map, then
Indρ U ∼= U K for every CH-module U .
(g) Any class functions α ∈ RC (H) and β ∈ RC (G) satisfy
(4.1.16) (Indρ α, β)G = (α, Resρ β)H
and
(4.1.17) hIndρ α, βiG = hα, Resρ βiH .
(See Remark 4.1.13 for the definition of Resρ β.)
(h) We have HomCG (Indρ U, V ) ∼ = HomCH (U, Resρ V ) for every CH-module U and every CG-module
V . (See Remark 4.1.13 for the definition of Resρ V .)
(i) Similarly to how we made CG into a (CG, CH)-bimodule, let us make CG into a (CH, CG)-bimodule
(so the right CG-module structure is plain multiplication inside CG, whereas the left CH-module
structure is induced by the C-algebra homomorphism C [ρ] : CH → CG). If U is any CH-module,
then the CG-module HomCH (CG, U ) (defined as in Exercise 4.1.4 using the (CH, CG)-bimodule
structure on CG) is isomorphic to Indρ U .
(j) We have HomCG (U, Indρ V ) ∼ = HomCH (Resρ U, V ) for every CG-module U and every CH-module
V . (See Remark 4.1.13 for the definition of Resρ V .)
Furthermore:
(k) Use the above to prove the formula (4.1.3).
(l) Use the above to prove the formula (4.1.12).
[Hint: Part (b) of this exercise is hard. To solve it, it is useful to have a way of computing the trace of a
linear operator without knowing a basis of the vector space it is acting on. There is a way to do this using
a “finite dual generating system”, which is a somewhat less restricted notion than that of a basis196. Try to
create a finite dual generating system for Indρ U from one for U (and from the group G), and then use it to
compute χIndρ U .
196More precisely: Let K be a field, and V be a K-vector space. A finite dual generating system for V means a triple


I, (ai )i∈I , (fi )i∈I , where
• I is a finite set;
• (ai )i∈I is a family of elements of V ;
• (fi )i∈I is a family of elements of V ∗ (where V ∗ means HomK (V, K))
P
such that every v ∈ V satisfies v = i∈I fi (v) ai . For example, if (ej )j∈J is a finite basis of the vector space V , and if
     
∗
ej is the basis of V dual to this basis (ej )j∈J , then J, (ej )j∈J , e∗j
∗ is a finite dual generating system for V ;
j∈J j∈J
however, most finite dual generating systems are not obtained this way.
102 DARIJ GRINBERG AND VICTOR REINER

The solution of part (i) is a modification of the solution of Exercise 4.1.4, but complicated by the fact
that H is no longer (necessarily) a subgroup of G. Part (f) can be solved by similar arguments, or using
part (i), or using Exercise 4.1.12(b).]
The result of Exercise 4.1.14(h) generalizes (4.1.7) (because of Exercise 4.1.14(d)), but also generalizes
(4.1.11) (due to Exercise 4.1.14(f)). Similarly, Exercise 4.1.14(g) generalizes both (4.1.9) and (4.1.14). Sim-
ilarly, Exercise 4.1.14(i) generalizes Exercise 4.1.4, and Exercise 4.1.14(j) generalizes Exercise 4.1.6.
Similarly, Exercise 4.1.3 is generalized by the following exercise:
Exercise 4.1.15. Let G1 , G2 , H1 and H2 be four finite groups. Let ρ1 : H1 → G1 and ρ2 : H2 → G2
be two group homomorphisms. These two homomorphisms clearly induce a group homomorphism ρ1 × ρ2 :
H1 × H2 → G1 × G2 . Let U1 be a CH1 -module, and U2 be a CH2 -module. Show that
Indρ1 ×ρ2 (U1 ⊗ U2 ) ∼
= (Indρ1 U1 ) ⊗ (Indρ2 U2 )
as C [G1 × G2 ]-modules.
The Indρ and Resρ operators behave “functorially” with respect to composition. Here is what this means:
Exercise 4.1.16. Let G, H and I be three finite groups. Let ρ : H → G and τ : I → H be two group
homomorphisms.
(a) We have Indρ Indτ U ∼
= Indρ◦τ U for every CI-module U .
(b) We have Indρ Indτ f = Indρ◦τ f for every f ∈ RC (I).
(c) We have Resτ Resρ V = Resρ◦τ V for every CG-module V .
(d) We have Resτ Resρ f = Resρ◦τ f for every f ∈ RC (G).
Exercise 4.1.16(a), of course, generalizes Exercise 4.1.2.

4.1.7. Semidirect products. Recall that a semidirect product is a group G n K having two subgroups G, K
with
• K / (G n K) is a normal subgroup,
• G n K = GK = KG, and
• G ∩ K = {e}.
In this setting one has two interesting adjoint constructions, applied in Section 4.5.

Proposition 4.1.17. Fix a C[G n K]-module V .

(i) For any CG-module U , one has C[G n K]-module structure
Φ(U ) := U ⊗ V,
determined via
k(u ⊗ v) = u ⊗ k(v),
g(u ⊗ v) = g(u) ⊗ g(v).
(ii) For any C[G n K]-module W , one has CG-module structure
Ψ(W ) := HomCK (ResGnK
K V, ResGnK
K W ),
determined via g(ϕ) = g ◦ ϕ ◦ g −1 .
(iii) The maps
Φ
CG − mods C[G n K] − mods
Ψ

The crucial observation is now that if I, (ai )i∈I , (fi )i∈I is a finite dual generating system for a vector space V , and if T
is an endomorphism of V , then
X
trace T = fi (T ai ) .
i∈I

Prove this!
 HOPF ALGEBRAS IN COMBINATORICS 103

are adjoint in the sense that one has an isomorphism

HomCG (U, Ψ(W )) −→ HomC[GnK] (Φ(U ), W )
k k
GnK GnK
HomCG (U, HomCK (ResK V, ResK W )) HomC[GnK] (U ⊗ V, W )

ϕ 7−→ ϕ(u ⊗ v) := ϕ(u)(v)

(iv) One has a CG-module isomorphism
(Ψ ◦ Φ)(U ) ∼
= U ⊗ EndCK (ResGnK
K V ).
In particular, if ResGnK
K V is a simple CK-module, then (Ψ ◦ Φ)(U ) ∼
= U.

Proof. These are mostly straightforward exercises in the definitions. To check assertion (iv), for example,
note that K acts only in the right tensor factor in ResGnK
K (U ⊗ V ), and hence as CG-modules one has
(Ψ ◦ Φ)(U ) = HomCK (ResGnK
K
GnK
V, ResK (U ⊗ V ))
= HomCK (ResGnK
K
GnK
V, U ⊗ ResK V)
= U ⊗ HomCK (ResGnK
K
GnK
V, ResK V)
= U ⊗ EndCK (ResGnK
K V)


4.2. Three towers of groups. Here we consider three towers of groups

G∗ = (G0 < G1 < G2 < G3 < · · · )
where either
• Gn = Sn , the symmetric group197, or
• Gn = Sn [Γ], the wreath product of the symmetric group with some arbitrary finite group Γ, or
• Gn = GLn (Fq ), the finite general linear group198.
Here the wreath product Sn [Γ] can be thought of informally as the group of monomial n × n matrices whose
nonzero entries lie in Γ, that is, n × n matrices having exactly one nonzero entry in each row and column,
and that entry is an element of Γ. E.g.
    
0 g2 0 0 0 g6 0 g2 g5 0
g1 0 0   0 g5 0  =  0 0 g1 g6  .
0 0 g3 g4 0 0 g3 g4 0 0
More formally, Sn [Γ] is the semidirect product Sn n Γn in which Sn acts on Γn via σ(γ1 , . . . , γn ) =
(γσ−1 (1) , . . . , γσ−1 (n) ).
For each of the three towers G∗ , there are embeddings Gi × Gj ,→ Gi+j and we introduce maps indi+j i,j
taking C[Gi ×Gj ]-modules to CGi+j -modules, as well as maps resi+j
i,j carrying modules in the reverse direction
which are adjoint:
i+j
(4.2.1) HomCGi+j (indi,j U, V ) = HomC[Gi ×Gj ] (U, resi+j
i,j V ).

Definition 4.2.1. For Gn = Sn , one embeds Si × Sj into Si+j as the permutations that permute
{1, 2, . . . , i} and {i + 1, i + 2, . . . , i + j} separately. Here one defines
S
indi+j i+j
i,j := IndSi ×Sj ,
i+j S
resi,j := ResSi+j
i ×Sj
.

197The symmetric group S is the group of all permutations of the empty set {1, 2, . . . , 0} = ∅. It is a trivial group. (Note
0
that S1 is also a trivial group.)
198The group GL (F ) is a trivial group, consisting of the empty 0 × 0 matrix.
0 q
104 DARIJ GRINBERG AND VICTOR REINER

For Gn = Sn [Γ], similarly embed Si [Γ] × Sj [Γ] into Si+j [Γ] as block monomial matrices whose two diagonal
blocks have sizes i, j respectively, and define
S [Γ]
indi+j i+j
i,j := IndSi [Γ]×Sj [Γ] ,
S [Γ]
resi+j i+j
i,j := ResSi [Γ]×Sj [Γ] .

For Gn = GLn (Fq ), which we will denote just GLn , similarly embed GLi × GLj into GLi+j as block
diagonal matrices whose two diagonal blocks have sizes i, j respectively. However, one also introduces as an
intermediate the parabolic subgroup Pi,j consisting of the block upper-triangular matrices of the form
 
gi `
0 gj
where gi , gj lie in GLi , GLj , respectively, and ` in Fi×j
q is arbitrary. One has a quotient map Pi,j → GLi ×GLj
whose kernel Ki,j is the set of matrices of the form
 
Ii `
0 Ij
with ` again arbitrary. Here one defines
GL P
indi+j
i,j := IndPi,j
i+j i,j
InflGL i ×GLj
,
 Ki,j
GLi+j
resi+j
i,j := ResPi,j (−) .

In the case Gn = GLn , the operation indi+j

i,j is sometimes called parabolic induction or Harish-Chandra
i+j
induction. The operation resi,j is essentially just the Ki,j -fixed point construction V 7→ V Ki,j . However
writing it as the above two-step composite makes it more obvious, (via (4.1.7) and (4.1.11)) that resi+j
i,j is
again adjoint to indi+j
i,j .
Definition 4.2.2. For each of the three towers G∗ , define a graded Z-module
M
A := A(G∗ ) = R(Gn )
n≥0

with aF bilinear form (·, ·)A whose restriction to An := R(Gn ) is the usual form (·, ·)Gn , and such that
Σ := n≥0 Irr(Gn ) gives an orthonormal Z-basis. Notice that A0 = Z has its basis element 1 equal to the
unique irreducible character of the trivial group G0 .
Bearing in mind that An = R(Gn ) and
Ai ⊗ Aj = R(Gi ) ⊗ R(Gj ) ∼
= R(Gi × Gj )
one then has candidates for product and coproduct defined by
m := indi+j
i,j : Ai ⊗ Aj −→ Ai+j
∆ := i+j=n resi+j
L L
i,j : An −→ i+j=n Ai ⊗ Aj .
The coassociativity of ∆ is an easy consequence of transitivity of the constructions of restriction and fixed
points199. We could derive the associativity of m from the transitivity of induction and inflation, but this
would be more complicated200; we will instead prove it differently.
We first show that the maps m and ∆ are adjoint with respect to the forms (·, ·)A and (·, ·)A⊗A . In fact,
if U , V , W are modules over CGi , CGj , CGi+j , respectively, then we can write the C[Gi × Gj ]-module
resi+j
L
i,j W as a direct sum k Xk ⊗ Yk with Xk being CGi -modules and Yk being CGj -modules; we then have
X
(4.2.2) resi+j
i,j χW = χ X k ⊗ χ Yk
k

199More precisely, using this transitivity, it is easily reduced to proving that K

i+j,k · (Ki,j × {Ik }) = Ki,j+k · {Ii } × Kj,k
(an equality between subgroups of GLi+j+k ). This equality can be proven by realizing that both of its sides equal the set of
 
Ii ` m
all block matrices of the form  0 Ij n  with `, m and n being matrices of sizes i × j, i × k and j × k, respectively.
0 0 Ik
200See Exercise 4.3.11(c) for such a derivation.
 HOPF ALGEBRAS IN COMBINATORICS 105

and
   
i+j
(m (χU ⊗ χV ) , χW )A = indi,j (χU ⊗V ) , χW = indi+j
i,j (χU ⊗V ) , χW
A Gi+j
!
  X
= χU ⊗V , resi+j
i,j χW = χU ⊗V , χXk ⊗ χYk
Gi ×Gj
k Gi ×Gj
X X
= (χU ⊗V , χXk ⊗Yk )Gi ×Gj = (χU , χXk )Gi (χV , χYk )Gj
k k

(the third equality sign follows by taking dimensions in (4.2.1) and recalling (4.1.1); the fourth equality sign
follows from (4.2.2); the sixth one follows from (4.1.2)) and
!
  X
i+j
(χU ⊗ χV , ∆ (χW ))A⊗A = χU ⊗ χV , resi,j χW = χU ⊗ χV , χXk ⊗ χYk
A⊗A
k A⊗A
X X
= (χU , χXk )A (χV , χYk )A = (χU , χXk )Gi (χV , χYk )Gj
k k

(the first equality sign follows by removing all terms in ∆ (χW ) whose scalar product with χU ⊗ χV van-
ishes for reasons of gradedness; the second equality sign follows from (4.2.2)), which in comparison yield
(m (χU ⊗ χV ) , χW )A = (χU ⊗ χV , ∆ (χW ))A⊗A , thus showing that m and ∆ are adjoint maps. Therefore,
m is associative (since
L ∆ is coassociative).
Endowing A = n≥0 R(Gn ) with the obvious unit and counit maps, it thus becomes a graded, finite-type
Z-algebra and Z-coalgebra.
The next section addresses the issue of why they form a bialgebra. However,
F assuming this for the mo-
ment, it should be clear that each of these algebras A is a PSH having Σ = n≥0 Irr(Gn ) as its PSH-basis. Σ
is self-dual because m, ∆ are defined by adjoint maps, and it is positive because m, ∆ take irreducible repre-
sentations to genuine representations not just virtual ones, and hence have characters which are nonnegative
sums of irreducible characters.
Exercise 4.2.3. Let i, j and k be three nonnegative integers. Let U be a CSi -module, let V be a CSj -
module, and let W be a CSk -module. Show that there are canonical C [Si × Sj × Sk ]-module isomorphisms
 
S Si+j ∼ S
IndSi+j+k
i+j ×Sk
IndSi ×Sj (U ⊗ V ) ⊗ W = IndSi+j+k
i ×Sj ×Sk
(U ⊗ V ⊗ W )
 
∼ S Sj+k
= IndSi+j+k
i ×Sj+k
U ⊗ IndSj ×Sk
(V ⊗ W ) .

(Similar statements hold for the other two towers of groups and their respective ind functors, although
the one for the GL∗ tower is harder to prove. See Exercise 4.3.11(a) for a more general result.)
4.3. Bialgebra and double cosets. To show that the algebra and coalgebras A = A(G∗ ) are bialgebras,
the central issue is checking the pentagonal diagram in (1.3.4), that is, as maps A ⊗ A → A ⊗ A, one has
(4.3.1) ∆ ◦ m = (m ⊗ m) ◦ (id ⊗T ⊗ id) ◦ (∆ ⊗ ∆).
In checking this, it is convenient to have a lighter notation for various subgroups of the groups Gn
corresponding to compositions α.
Definition 4.3.1. (a) An almost-composition is a (finite) tuple α = (α1 , α2 , . . . , α` ) of nonnegative
integers. Its length is defined to be ` and denoted by `(α); its size is defined to be α1 + α2 + · · · + α`
and denoted by |α|; its parts are its entries α1 , α2 , . . . , α` . The almost-compositions of size n are
called the almost-compositions of n.
(b) A composition is a finite tuple of positive integers. Of course, any composition is an almost-
composition, and so all notions defined for almost-compositions (like size and length) make sense for
compositions.
Note that any partition of n (written without trailing zeroes) is a composition of n. We write ∅
(and sometimes, sloppily, (0), when there is no danger of mistaking it for the almost-composition
(0)) for the empty composition ().
106 DARIJ GRINBERG AND VICTOR REINER

Definition 4.3.2. Given an almost-composition α = (α1 , . . . , α` ) of n, define a subgroup

Gα ∼
= Gα1 × · · · × Gα` < Gn
via the block-diagonal embedding with diagonal blocks of sizes (α1 , . . . , α` ). This Gα is called a Young
subgroup Sα when Gn = Sn , and a Levi subgroup when Gn = GLn . In the case when Gn = Sn [Γ], we also
denote Gα by Sα [Γ]. In the case where Gn = GLn , also define the parabolic subgroup Pα to be the subgroup
of Gn consisting of block-upper triangular matrices whose diagonal blocks have sizes (α1 , . . . , α` ), and let Kα
be the kernel of the obvious surjection Pα → Gα which sends a block upper-triangular matrix to the tuple
of its diagonal blocks whose sizes are α1 , α2 , . . . , α` . Notice that P(i,j) = Pi,j for any i and j with i + j = n;
similarly, K(i,j) = Ki,j for any i and j with i + j = n. We will also abbreviate G(i,j) = Gi × Gj by Gi,j .
When (α1 , α2 , . . . , α` ) is an almost-composition, we abbreviate G(α1 ,α2 ,...,α` ) by Gα1 ,α2 ,...,α` (and similarly
for the P ’s).
Definition 4.3.3. Let K and H be two groups, τ : K → H a group homomorphism, and U a CH-module.
Then, U τ is defined as the CK-module with ground space U and action given by k · u = τ (k) · u for all k ∈ K
201
and u ∈ U . This very simple construction generalizes the definition of U g for an element g ∈ G, where
G is a group containing H as a subgroup; in fact, in this situation we have U g = U τ , where K = g H and
τ : K → H is the map k 7→ g −1 kg.
Using homogeneity, checking the bialgebra condition (4.3.1) in the homogeneous component (A ⊗ A)n
amounts to the following: for each pair of representations U1 , U2 of Gr1 , Gr2 with r1 + r2 = n, and for each
(c1 , c2 ) with c1 + c2 = n, one must verify that
resnc1 ,c2 indnr1 ,r2 (U1 ⊗ U2 )


 
(4.3.2) ∼
M
c1 c2

r1 r2

τA−1
= ind a11 ,a21 ⊗ inda12 ,a22 res U
a11 ,a12 1 ⊗ res U
a21 ,a22 2
A
 
a11 a12
where the direct sum is over all matrices A = in N2×2 with row sums (r1 , r2 ) and column sums
a21 a22
(c1 , c2 ), and where τA is the obvious isomorphism between the subgroups
Ga11 ,a12 ,a21 ,a22 (< Gr1 ,r2 )
(4.3.3)
Ga11 ,a21 ,a12 ,a22 (< Gc1 ,c2 )
(we are using the inverse τA−1
of this isomorphism τA to identify modules for the first subgroup with modules
for the second subgroup, according to Definition 4.3.3).
As one might guess, (4.3.2) comes from the Mackey formula (Theorem 4.1.7), once one identifies the
appropriate double coset representatives. This is just as easy to do in a slightly more general setting.
Definition 4.3.4. Given almost-compositions α, β of n having lengths `, m and a matrix A in N`×m with
row sums α and column sums β, define a permutation wA in Sn as follows. Disjointly decompose [n] =
{1, 2, . . . , n} into consecutive intervals of numbers
[n] = I1 t · · · t I`
[n] = J1 t · · · t Jm
such that |Ii | = αi , |Jj | = βj . For every j ∈ [m], disjointly decompose Jj into consecutive intervals of
numbers Jj = Jj,1 t Jj,2 t · · · t Jj,` such that every i ∈ [`] satisfies |Jj,i | = aij . For every i ∈ [`], disjointly
decompose Ii into consecutive intervals of numbers Ii = Ii,1 t Ii,2 t · · · t Ii,m such that every j ∈ [m] satisfies
|Ii,j | = aij . Now, for every i ∈ [`] and j ∈ [m], let πi,j be the increasing bijection from Jj,i to Ii,j (this is
well-defined since these two sets both have cardinality aij ). The disjoint union of these bijections πi,j over
all i and j is a bijection [n] → [n] (since the disjoint union of the sets Jj,i over all i and j is [n], and so is
the disjoint union of the sets Ii,j ), that is, a permutation of [n]; this permutation is what we call wA .
Example 4.3.5. Taking n = 9 and α = (4, 5), β = (3, 4, 2), one has
I1 = {1, 2, 3, 4}, I2 = {5, 6, 7, 8, 9}
J1 = {1, 2, 3}, J2 = {4, 5, 6, 7}, J3 = {8, 9}.
201We have already met this CK-module U τ in Remark 4.1.13, where it was called Res U .
τ
 HOPF ALGEBRAS IN COMBINATORICS 107

 
2 2 0
Then one possible matrix A having row and column sums α, β is A = , and its associated permu-
1 2 2
tation wA written in two-line notation is
 
1 2 3 | 4 5 6 7 | 8 9
1 2 5 | 3 4 6 7 | 8 9
with vertical lines dividing the sets Jj on top, and with elements of Ii underlined i times on the bottom.
Remark 4.3.6. Given almost-compositions α and β of n having lengths ` and m, and a permutation w ∈ Sn .
It is easy to see that there exists a matrix A ∈ N`×m satisfying wA = w if and only if the restriction of w
to each Jj and the restriction of w−1 to each Ii are increasing. In this case, the matrix A is determined by
aij = |w(Jj ) ∩ Ii |.
Among our three towers G∗ of groups, the symmetric group tower (Gn = Sn ) is the simplest one. We
will now see that it also embeds into the two others, in the sense that Sn embeds into Sn [Γ] for every Γ
and into GLn (Fq ) for every q.
First, for every n ∈ N and any group Γ, we embed the group Sn into Sn [Γ] by means of the canonical
embedding Sn → Sn n Γn = Sn [Γ]. If we regard elements of Sn [Γ] as n × n monomial matrices with
nonzero entries in Γ, then this boils down to identifying every π ∈ Sn with the permutation matrix of π (in
which the 1’s are read as the neutral element of Γ). If α is an almost-composition of n, then this embedding
Sn → Sn [Γ] makes the subgroup Sα of Sn become a subgroup of Sn [Γ], more precisely a subgroup of
Sα [Γ] < Sn [Γ].
For every n ∈ N and every q, we embed the group Sn into GLn (Fq ) by identifying every permutation
π ∈ Sn with its permutation matrix in GLn (Fq ). If α is an almost-composition of n, then this embedding
makes the subgroup Sα of Sn become a subgroup of GLn (Fq ). If we let Gn = GLn (Fq ), then Sα < Gα <
Pα .
The embeddings we have just defined commute with the group embeddings Gn < Gn+1 on both sides.
Proposition 4.3.7. The permutations {wA } as A runs over all matrices in N`×m having row, column sums
α, β give a system of double coset representatives for
Sα \Sn /Sβ
Sα [Γ]\Sn [Γ]/Sβ [Γ]
Pα \GLn /Pβ

Proof. First note that double coset representatives for Sα \Sn /Sβ should also provide double coset repre-
sentatives for Sα [Γ]\Sn [Γ]/Sβ [Γ], since
Sα [Γ] = Sα Γn = Γn Sα .
We give an algorithm to show that every double coset Sα wSβ contains some wA . Start by altering w within
its coset wSβ , that is, by permuting the positions within each set Jj , to obtain a representative w0 for wSβ
in which each set w0 (Jj ) appears in increasing order in the second line of the two-line notation for w0 . Then
alter w0 within its coset Sα w0 , that is, by permuting the values within each set Ii , to obtain a representative
wA having the elements of each set Ii appearing in increasing order in the second line; because the values
within each set Ii are consecutive, this alteration will not ruin the property that one had each set w0 (Jj )
appearing in increasing order. For example, one might have
 
1 2 3 | 4 5 6 7 | 8 9
w=
4 8 2 | 5 3 9 1 | 7 6
 
0 1 2 3 | 4 5 6 7 | 8 9
w = ∈ wSβ
2 4 8 | 1 3 5 9 | 6 7
 
1 2 3 | 4 5 6 7 | 8 9
wA = ∈ Sα w0 ⊂ Sα w0 Sβ = Sα wSβ
1 2 5 | 3 4 6 7 | 8 9
Next note that Sα wA Sβ = Sα wB Sβ implies A = B, since the quantities
ai,j (w) := |w(Jj ) ∩ Ii |
are easily seen to be constant on double cosets Sα wSβ .
108 DARIJ GRINBERG AND VICTOR REINER

A similar argument shows that Pα wA Pβ = Pα wB Pβ implies A = B: for g in GLn , the rank rij (g) of the
matrix obtained by restricting g to rows Ii t Ii+1 t · · · t I` and columns J1 t J2 t · · · t Jj is constant on
double cosets Pα gPβ , and for a permutation matrix w one can recover ai,j (w) from the formula
ai,j (w) = ri,j (w) − ri,j−1 (w) − ri+1,j (w) + ri+1,j−1 (w).
Thus it only remains to show that every double coset Pα gPβ contains some wA . Since Sα < Pα , and we
have seen already that every double coset Sα wSβ contains some wA , it suffices to show that every double
coset Pα gPβ contains some permutation w. However, we claim that this is already true for the smaller
double cosets BgB where B = P1n is the Borel subgroup of upper triangular invertible matrices, that is, one
has the usual Bruhat decomposition G
GLn = BwB.
w∈Sn
To prove this decomposition, we show how to find a permutation w in each double coset BgB. The freedom
to alter g within its coset gB allows one to scale columns and add scalar multiples of earlier columns to later
columns. We claim that using such column operations, one can always find a representative g 0 for coset gB
in which
• the bottommost nonzero entry of each column is 1 (call this entry a pivot),
• the entries to right of each pivot within its row are all 0, and
• there is one pivot in each row and each column, so that their positions are the positions of the 1’s
in some permutation matrix w.
In fact, we will see below that BgB = BwB in this case. The algorithm which produces g 0 from g is simple:
starting with the leftmost column, find its bottommost nonzero entry, and scale the column to make this
entry a 1, creating the pivot in this column. Now use this pivot to clear out all entries in its row to its right,
using column operations that subtract multiples of this column from later columns. Having done this, move
on to the next column to the right, and repeat, scaling to create a pivot, and using it to eliminate entries to
its right.202
For example, the typical matrix g lying in the double coset BwB where
 
1 2 3 | 4 5 6 7 | 8 9
w=
4 8 2 | 5 3 9 1 | 7 6
from before is one that can be altered within its coset gB to look like this:
 
∗ ∗ ∗ ∗ ∗ ∗ 1 0 0
∗ ∗ 1 0 0 0 0 0 0
 
∗ ∗ 0 ∗ 1 0 0 0 0
 
1 0 0 0 0 0 0 0 0
0
 
g = 0 ∗ 0 1 0 0 0 0 0  ∈ gB.
0 ∗ 0 0 0 ∗ 0 ∗ 1
 
0 ∗ 0 0 0 ∗ 0 1 0
 
0 1 0 0 0 0 0 0 0
0 0 0 0 0 1 0 0 0
Having found this g 0 in gB, a similar algorithm using left multiplication by B shows that w lies in Bg 0 ⊂
Bg 0 B = BgB. This time no scalings are required to create the pivot entries: starting with the bottom row,
one uses its pivot to eliminate all the entries above it in the same column (shown by stars ∗ above) by adding
202To see that this works, we need to check three facts:

(a) We will find a nonzero entry in every column during our algorithm.
(b) Our column operations preserve the zeroes lying to the right of already existing pivots.
(c) Every row contains exactly one pivot at the end of the algorithm.
But fact (a) simply says that our matrix can never have an all-zero column during the algorithm; this is clear (since the rank of
the matrix remains constant during the algorithm and was n at its beginning). Fact (b) holds because all our operations either
scale columns (which clearly preserves zero entries) or subtract a multiple of the column c containing the current pivot from a
later column d (which will preserve every zero lying to the right of an already existing pivot, because any already existing pivot
must lie in a column b < c and therefore both columns c and d have zeroes in its row). Fact (c) follows from noticing that
there are n pivots altogether at the end of the algorithm, but no row can contain two of them (since the entries to the right of
a pivot in its row are 0).
 HOPF ALGEBRAS IN COMBINATORICS 109

multiples of the bottom row to higher rows. Then do the same using the pivot in the next-to-bottom row,
etc. The result is the permutation matrix for w. 
F
Remark 4.3.8. The Bruhat decomposition GLn = w∈Sn BwB is related to the so-called LPU factorization
– one of a myriad of matrix factorizations appearing in linear algebra.203 It is actually a fairly general
phenomenon, and requires neither the finiteness of F, nor the invertibility, nor even the squareness of the
matrices (see Exercise 4.3.9(b) for an analogue holding in a more general setup).
Exercise 4.3.9. Let F be any field.
(a) For any n ∈ N and any A ∈ GLn (F), prove that there exist a lower-triangular matrix L ∈ GLn (F),
an upper-triangular matrix U ∈ GLn (F) and a permutation matrix P ∈ Sn ⊂ GLn (F) (here, we
identify permutations with the corresponding permutation matrices) such that A = LP U .
n×m
(b) Let n ∈ N and m ∈ N. Let Fn,m denote the set of all n × m-matrices B ∈ {0, 1} such that each
row of B contains at most one 1 and each column of B contains at most one 1. We regard Fn,m as
a subset of Fn×m by means of regarding {0, 1} as a subset of F.
For every k ∈ N, we let Bk denote the subgroup of GLk (F) consisting of all upper-triangular
matrices.
Prove that G
Fn×m = Bn f Bm .
f ∈Fn,m

Corollary 4.3.10. For each of the three towers of groups G∗ , the product and coproduct structures on
A = A(G∗ ) endow it with a bialgebra structure, and hence they form PSH’s.
Proof. The first two towers Gn = Sn and Gn = Sn [Γ] have product, coproduct defined by induction,
restriction along embeddings Gi × Gj < Gi+j . Hence the desired bialgebra equality (4.3.2) follows from
Mackey’s Theorem 4.1.7, taking G = Gn , H = G(r1 ,r2 ) , K = G(c1 ,c2 ) , U = U1 ⊗ U2 with double coset
representatives204
{g1 , . . . , gt } = wAt : A ∈ N2×2 , A has row sums (r1 , r2 ) and column sums (c1 , c2 )


and checking for a given double coset

KgH = (Gc1 ,c2 )wAt (Gr1 ,r2 )
2×2
indexed by a matrix A in N with row sums (r1 , r2 ) and column sums (c1 , c2 ), that the two subgroups
appearing on the left in (4.3.3) are exactly
H ∩ K wAt = Gr1 ,r2 ∩ (Gc1 ,c2 )wAt ,
wAt
H ∩ K = wAt (Gr1 ,r2 ) ∩ Gc1 ,c2 ,
respectively. One should also apply (4.1.6) and check that the isomorphism τA between the two subgroups
−1 wAt
in (4.3.3) is the conjugation isomorphism by wAt (that is, τA (g) = wAt gwA t for every g ∈ H ∩ K ). We
205
leave all of these bookkeeping details to the reader to check.
203Specifically, an LPU factorization of a matrix A ∈ GL (F) (for an arbitrary field F) means a way to write A as a product
n
A = LP U with L ∈ GLn (F) being lower-triangular, U ∈ GLn (F) being upper-triangular, and P ∈ Sn ⊂ GLn (F) being a
permutation matrix. Such a factorization always exists (although it is generally not unique). This can be derived from the
Bruhat decomposition (see Exercise 4.3.9(a) for a proof). See also [188] for related discussion.
204Proposition 4.3.7 gives as a system of double coset representatives for G
(c1 ,c2 ) \Gn /G(r1 ,r2 ) the elements
2×2

wA : A ∈ N , A has row sums (c1 , c2 ) and column sums (r1 , r2 )
= wAt : A ∈ N2×2 , A has row sums (r1 , r2 ) and column sums (c1 , c2 )


where At denotes the transpose matrix of A.

205It helps to recognize w t as the permutation written in two-line notation as
A
a022 | a022 + 1 a022 + 2
 
1 2 . . . a11 | a11 + 1 a11 + 2 . . . r1 | r1 + 1 r1 + 2 ... ... n
,
1 2 . . . a11 | c1 + 1 c1 + 2 . . . a022 | a11 + 1 a11 + 2 . . . c1 | a022 + 1 a022 + 2 ... n
 
Ia11 0 0 0
 0 0 Ia21 0
a022

where = r1 + a21 = c1 + a12 = n − a22 . In matrix form, wAt is the block matrix 
 0
.
Ia12 0 0 
0 0 0 Ia22
110 DARIJ GRINBERG AND VICTOR REINER

For the tower with Gn = GLn , there is slightly more work to be done to check the equality (4.3.2). Via
Mackey’s Theorem 4.1.7 and Proposition 4.3.7, the left side is

resnc1 ,c2 indnr1 ,r2 (U1 ⊗ U2 )

 Kc1 ,c2
Gn Pr1 ,r2
= ResG Pc1 ,c2 IndPr1 ,r2 InflGr1 ,r2 (U1 ⊗ U2 )
n

 τA−1 !!Kc1 ,c2

Pc1 ,c2 Pr1 ,r2 Pr1 ,r2
M
(4.3.4) = IndwAt Pr ,r ∩Pc ,c
1 2
Res
1 2
w t Infl
A Gr ,r (U1 ⊗ U2 )
Pr1 ,r2 ∩Pc1 ,c2 1 2
A

where A runs over the usual 2 × 2 matrices. The right side is a direct sum over this same set of matrices A:
 
M
τ −1
indca111 ,a21 ⊗ indca212 ,a22 resra111 ,a12 U1 ⊗ resra221 ,a22 U2 A

A
M G G
 
P Pa12 ,a22

= IndPac1 ⊗ IndPac2 ◦ InflGaa11 ,a21
⊗ InflGa ,a
11 ,a21 12 ,a22 ,a 11 21 12 22
A
Ka21 ,a22 τA−1
 !
Ka11 ,a12 
G G
ResPar1 ,a U1 ⊗ ResPar2 ,a U2
11 12 21 22

G
M
= IndPac1 ,c,a2
11 21 ×Pa12 ,a22
A
Ka11 ,a12 ×Ka21 ,a22 τA−1
 !
P 21 ×Pa12 ,a22 G
(4.3.5) InflGaa11 ,a ResPar1 ,r,a2 ×Pa ,a (U1 ⊗ U2 )
11 ,a21 ,a12 ,a22 11 12 21 22

(by (4.1.6), (4.1.15) and their obvious analogues for restriction and inflation). Thus it suffices to check for
each 2 × 2 matrix A that any CGc1 ,c2 -module of the form V1 ⊗ V2 has the same inner product with the
A-summands of (4.3.4) and (4.3.5). Abbreviate w := wAt and τ := τA−1 .
Notice that w Pr1 ,r2 is the group of all matrices having the block form
 
g11 h i j
 0 g21 0 k 
(4.3.6) 
 d

e g12 ` 
0 f 0 g22
in which the diagonal blocks gij for i, j = 1, 2 are invertible of size aij × aij , while the blocks h, i, j, k, `, d, e, f
are all arbitrary matrices206 of the appropriate (rectangular) block sizes. Hence, w Pr1 ,r2 ∩ Pc1 ,c2 is the group
of all matrices having the block form
 
g11 h i j
 0 g21 0 k 
(4.3.7) 
 0

0 g12 ` 
0 0 0 g22
in which the diagonal blocks gij for i, j = 1, 2 are invertible of size aij × aij , while the blocks h, i, j, k, `
are all arbitrary matrices of the appropriate (rectangular) block sizes; then w Pr1 ,r2 ∩ Gc1 ,c2 is the subgroup
where the blocks i, j, k all vanish. The canonical projection w Pr1 ,r2 ∩ Pc1 ,c2 → w Pr1 ,r2 ∩ Gc1 ,c2 (obtained by
restricting the projection Pc1 ,c2 → Gc1 ,c2 ) has kernel w Pr1 ,r2 ∩ Pc1 ,c2 ∩ Kc1 ,c2 . Consequently,

(4.3.8) (w Pr1 ,r2 ∩ Pc1 ,c2 ) / (w Pr1 ,r2 ∩ Pc1 ,c2 ∩ Kc1 ,c2 ) = w Pr1 ,r2 ∩ Gc1 ,c2 .

Similarly,

Pr1 ,r2 ∩ Pcw1 ,c2 / Pr1 ,r2 ∩ Pcw1 ,c2 ∩ Kr1 ,r2 = Gr1 ,r2 ∩ Pcw1 ,c2 .



(4.3.9)

206The blocks i and j have nothing to do with the indices i, j in g .

ij
 HOPF ALGEBRAS IN COMBINATORICS 111

Computing first the inner product of V1 ⊗ V2 with the A-summand of (4.3.4), and using adjointness
properties, one gets
 τ
P Pr1 ,r2
ResPrr1 ,r
,r
2
∩P w
c1 ,c2
Infl Gr ,r
(U 1 ⊗ U 2 ) ,
1 2 1 2

P P
ReswcP1r,c2,r ∩Pc ,c InflGcc1 ,c,c2 (V1 ⊗ V2 )
wP
r1 ,r2 ∩Pc1 ,c2
1 2 1 2 1 2
 w τ
(4.1.10) Pr1 ,r2 ∩Pc1 ,c2 Gr1 ,r2
= InflGr ,r ∩Pcw ,c ResGr ,r ∩Pcw ,c (U1 ⊗ U2 ) ,
1 2 1 2 1 2 1 2
w 
Pr1 ,r2 ∩Pc1 ,c2 Gc1 ,c2
Inflw Pr ,r ∩Gc ,c Resw Pr ,r ∩Gc ,c (V1 ⊗ V2 ) w
1 2 1 2 1 2 1 2 Pr1 ,r2 ∩Pc1 ,c2

(by (4.3.9) and (4.3.8)). One can compute this inner product by first recalling that w Pr1 ,r2 ∩ Pc1 ,c2 is the
group of matrices having the block form (4.3.7) in which the diagonal blocks gij for i, j = 1, 2 are invertible
of size aij × aij , while the blocks h, i, j, k, ` are all arbitrary matrices of the appropriate (rectangular) block
sizes; then w Pr1 ,r2 ∩ Gc1 ,c2 is the subgroup where the blocks i, j, k all vanish. The inner product above then
becomes
   
1 X g11 i g21 k
χ U1 χU2
|w Pr1 ,r2 ∩ Pc1 ,c2 | 0 g12 0 g22
(gij )
(4.3.10) (h,i,j,k,`)
   
g h g `
χV1 11 χV2 12 .
0 g21 0 g22
If one instead computes the inner product of V1 ⊗ V2 with the A-summand of (4.3.5), using adjointness
properties and (4.1.13) one gets
 Ka11 ,a12 ×Ka21 ,a22 τ
G
ResPar1 ,r,a2 ×Pa ,a (U1 ⊗ U2 ) ,
11 12 21 22

 Ka11 ,a21 ×Ka12 ,a22 

G
ResPac1 ,c,a2 ×Pa ,a (V1 ⊗ V2 )
11 21 12 22
Ga ,a
11 21 12 22 ,a ,a
   
1 X 1 X g11 i g21 k
= χU1 χU2
|Ga11 ,a21 ,a12 ,a22 | |Ka11 ,a12 × Ka21 ,a22 | 0 g12 0 g22
(gij ) (i,k)
   
1 X g h g `
χV1 11 χV2 12 .
|Ka11 ,a21 × Ka12 ,a22 | 0 g21 0 g22
(h,`)

But this right hand side can be seen to equal (4.3.10), after one notes that
|w Pr1 ,r2 ∩ Pc1 ,c2 | = |Ga11 ,a21 ,a12 ,a22 | · |Ka11 ,a12 × Ka21 ,a22 | · |Ka11 ,a21 × Ka12 ,a22 | · #{j ∈ Faq 11 ×a22 }
and that the summands in (4.3.10) are independent of the matrix j in the summation. 
L
We can also define a C-vector space AC as the direct sum n≥0 RC (Gn ). In the same way as we have
made A = n≥0 R(Gn ) into a Z-bialgebra, we can turn AC = n≥0 RC (Gn ) into a C-bialgebra207. There
L L

is a C-bilinear form (·, ·)AC on AC which can be defined either as the C-bilinear extension of the Z-bilinear
form (·, ·)A : A × A → Z to AC , or (equivalently) as the C-bilinear form on AC which restricts to h·, ·iSn on
every homogeneous component RC (Gn ) and makes different homogeneous components mutually orthogonal.
The obvious embedding of A into the C-bialgebra AC (obtained from the embeddings R(Gn ) → RC (Gn ) for

207The definitions of m and ∆ for this C-bialgebra look the same as for A: For instance, m is still defined to be indi+j
i,j
on (AC )i ⊗ (AC )j , where indi+j
i,j is defined by the same formulas as in Definition 4.2.1. However, the operators of induction,
restriction, inflation and K-fixed space construction appearing in these formulas now act on class functions as opposed to
modules.
The fact that these maps m and ∆ satisfy the axioms of a C-bialgebra is easy to check: they are merely the C-linear
extensions of the maps m and ∆ of the Z-bialgebra A (this is because, for instance, induction of class functions and induction
of modules are related by the identity (4.1.5)), and thus satisfy the same axioms as the latter.
112 DARIJ GRINBERG AND VICTOR REINER

all n) respects the bialgebra operations208, and the C-bialgebra AC can be identified with A ⊗Z C (the result
of extending scalars to C in A), because every finite group G satisfies RC (G) ∼
= R(G) ⊗Z C. The embedding
of A into AC also respects the bilinear forms.
Exercise 4.3.11. Let G∗ be one of the three towers.
For every almost-composition α = (α1 , α2 , . . . , α` ) of n ∈ N, let us define a map indnα which takes CGα -
modules to CGn -modules as follows: If G∗ = S∗ or G∗ = S∗ [Γ], we set
indnα := IndG n
Gα .

If G∗ = GL∗ , then we set

indnα := IndG n Pα
Pα InflGα .
(Note that indnα
= indni,j
if α has the form (i, j).)
Similarly, for every almost-composition α = (α1 , α2 , . . . , α` ) of n ∈ N, let us define a map resnα which
takes CGn -modules to CGα -modules as follows: If G∗ = S∗ or G∗ = S∗ [Γ], we set
resnα := ResG
Gα .
n

If G∗ = GL∗ , then we set

 Kα
resnα := ResG
Pα (−)
n
.
(Note that resnα = resni,j if α has the form (i, j).)
(a) If α = (α1 , α2 , . . . , α` ) is an almost-composition of an integer n ∈ N satisfying ` ≥ 1, and if Vi is a
CGαi -module for every i ∈ {1, 2, . . . , `}, then show that
 
α +α +···+α
indnα1 +α2 +···+α`−1 ,α` ind(α11 ,α22,...,α`−1`−1
) (V 1 ⊗ V 2 ⊗ · · · ⊗ V `−1 ) ⊗ V`
∼
= indnα (V1 ⊗ V2 ⊗ · · · ⊗ V` )
 
∼ α +α +···+α
= indnα1 ,α2 +α3 +···+α` V1 ⊗ ind(α22 ,α33,...,α` ) ` (V2 ⊗ V3 ⊗ · · · ⊗ V` ) .
(b) Solve Exercise 4.2.3 again using Exercise 4.3.11(a).
(c) We proved above that the map m : A ⊗ A → A (where A = A (G∗ )) is associative, by using the
adjointness of m and ∆. Give a new proof of this fact, which makes no use of ∆.
(d) If α = (α1 , α2 , . . . , α` ) is an almost-composition of an n ∈ N, and if χi ∈ R (Gαi ) for every i ∈
{1, 2, . . . , `}, then show that
χ1 χ2 · · · χ` = indnα (χ1 ⊗ χ2 ⊗ · · · ⊗ χ` )
in A = A (G∗ ).
(e) If n ∈ N, ` ∈ N and χ ∈ R (Gn ), then show that
X
∆(`−1) χ = resnα χ

in A⊗` , where A = A (G∗ ). Here, the sum on the right hand side runs over all almost-compositions
α of n having length `.
4.4. Symmetric groups. Finally, some payoff. Consider the tower of symmetric groups Gn = Sn , and
A = A(G∗ ) =: A(S). Denote by 1Sn , sgnSn the trivial and sign characters on Sn . For a partition λ of n,
denote by 1Sλ , sgnSλ the trivial and sign characters restricted to the Young subgroup Sλ = Sλ1 ×Sλ2 ×· · · ,
and denote by 1λ the class function which is the characteristic function for the Sn -conjugacy class of
permutations of cycle type λ.

Theorem 4.4.1. (a) Irreducible complex characters {χλ } of Sn are indexed by partitions λ in Parn ,
and one has a PSH-isomorphism, the Frobenius characteristic map209,
ch
A = A(S) −→ Λ

208This is because, for example, induction of class functions harmonizes with induction of modules (i.e., the equality (4.1.5)
holds).
209It is unrelated to the Frobenius endomorphisms from Exercise 2.9.9.
 HOPF ALGEBRAS IN COMBINATORICS 113

that for n ≥ 0 and λ ∈ Parn sends

1Sn 7−→ hn
sgnSn 7−→ en
χλ 7−→ sλ
IndS n
Sλ 1Sλ 7−→ hλ
IndSn
Sλ sgnSλ 7−→ eλ
pλ
1λ 7−→ zλ

(where ch is extended to a C-linear map AC → ΛC ) and for n ≥ 1 sends

pn
1(n) 7−→ n

Here, zλ is defined as in Proposition 2.5.15.

(b) For each n ≥ 0, the involution on class functions f : Sn → C sending f 7−→ sgnSn ∗f where
(sgnSn ∗f )(g) := sgn(g)f (g)
preserves the Z-sublattice R(SL n ) of genuine characters. The direct sum of these involutions induces
an involution on A = A(S) = n≥0 R(Sn ) that corresponds under ch to the involution ω on Λ.
F
Proof. (a) Corollary 4.3.10 implies that the set Σ = n≥0 Irr(Sn ) gives a PSH-basis for A. Since a character
χ of Sn has
M
(4.4.1) ∆(χ) = ResS
Si ×Sj χ,
n

i+j=n

such an element χ ∈ Σ ∩ An is never primitive for n ≥ 2. Hence the unique irreducible character ρ = 1S1 of
S1 is the only element of C = Σ ∩ p.
Thus Theorem 3.3.3(g) tells us that there are two PSH-isomorphisms A → Λ, each of which sends Σ to the
PSH-basis of Schur functions {sλ } for Λ. It also tells us that we can pin down one of the two isomorphisms
to call ch, by insisting that it map the two characters 1S2 , sgnS2 in Irr(S2 ) to h2 , e2 (and not e2 , h2 ).
Bearing in mind the coproduct formula (4.4.1), and the fact that 1Sn , sgnSn restrict, respectively, to
trivial and sign characters of Si × Sj for i + j = n, one finds that for n ≥ 2 one has sgn⊥ S2 annihilating 1Sn ,
and 1⊥S2 annihilating sgnSn . Therefore Theorem 3.3.1(b) (applied to Λ) implies 1 Sn , sgn Sn are sent under
Sn Sn
ch to hn , en . Then the fact that IndSλ 1Sλ , IndSλ sgnSλ are sent to hλ , eλ follows via induction products.
L
Recall that the C-vector space AC = n≥0 RC (Sn ) is a C-bialgebra, and can be identified with A ⊗Z C.
The multiplication and the comultiplicationL of AC are C-linear extensions of those of A, and are still given
i+j
by the same formulas m = indi,j and ∆ = i+j=n resi+j i,j as those of A (but now, induction and restriction
are defined for class functions, not just for representations). The C-bilinear form (·, ·)AC on AC extends both
the Z-bilinear form (·, ·)A on A and the C-bilinear forms h·, ·iSn on all RC (Sn ).
For the assertion about 1(n) , note that it is primitive in AC for n ≥ 1, because as a class function, the
indicator function of n-cycles vanishes upon restriction to Si × Sj for i + j = n if both i, j ≥ 1; these
subgroups contain no n-cycles. Hence Corollary 3.1.8 implies that ch(1(n) ) is a scalar multiple of pn . To pin
down the scalar, note pn = m(n) so (hn , pn )Λ = (hn , mn )Λ = 1, while ch−1 (hn ) = 1Sn has
1 1
(1Sn , 1(n) ) = · (n − 1)! = .
n! n
210
Thus ch(1(n) ) = pnn . The fact that ch(1λ ) = pλ
zλ then follows via induction product calculations211. Part
(b) follows from Exercise 4.4.4 below. 
210The first equality sign in this computation uses the fact that the number of all n-cycles in S is (n − 1)!. This is because
n
any n-cycle in Sn can be uniquely written in the form (i1 , i2 , . . . , in−1 , n) (in cycle notation) with (i1 , i2 , . . . , in−1 ) being a
permutation in Sn−1 (written in one-line notation).
211For instance, one can use (4.1.3) to show that z 1 = λ λ · · · λ ·1
λ λ 1 2 ` (λ1 ) 1(λ2 ) · · · 1(λ` ) if λ = (λ1 , λ2 , . . . , λ` ) with ` = ` (λ).
See Exercise 4.4.3(d) for the details.
114 DARIJ GRINBERG AND VICTOR REINER

Remark 4.4.2. The paper of Liulevicius [116] gives a very elegant alternate approach to the Frobenius map
ch
as a Hopf isomorphism A(S) −→ Λ, inspired by equivariant K-theory and vector bundles over spaces which
are finite sets of points!
Exercise 4.4.3. If P is a subset of a group G, we denote by 1P the map G → C which sends every element
212
of P to 1 and all remaining elements of G to 0. For any finite group G and any h ∈ G, we introduce
the following notations:
• Let ZG (h) denote the centralizer of h in G.
• Let ConjG (h) denote the conjugacy class of h in G.
• Define a map αG,h : G → C by αG,h = |ZG (h)| 1ConjG (h) . This map αG,h is a class function213.
(a) Prove that αG,h (g) = k∈G khk −1 = g for every finite group G and any h ∈ G and g ∈ G. Here,
P  

we are using the Iverson bracket notation (that is, for any statement A, we define [A] to be the
integer 1 if A is true, and 0 otherwise).
(b) Prove that if H is a subgroup of a finite group G, and if h ∈ H, then IndG H αH,h = αG,h .
(c) Prove that if G1 and G2 are finite groups, and if h1 ∈ G1 and h2 ∈ G2 , then the canonical isomorphism
RC (G1 ) ⊗ RC (G2 ) → RC (G1 × G2 ) sends αG1 ,h1 ⊗ αG2 ,h2 to αG1 ×G2 ,(h1 ,h2 ) .
(d) Fill in the details of the proof of ch(1λ ) = pzλλ in the proof of Theorem 4.4.1.
(e) Obtain an alternative proof of Remark 2.5.16.
(f) If G and H are two finite groups, and if ρ : H → G is a group homomorphism, then prove that
Indρ αH,h = αG,ρ(h) for every h ∈ H, where Indρ αH,h is defined as in Exercise 4.1.14.
Exercise 4.4.4. If G is a group and U1 and U2 are two CG-modules, then the tensor product U1 ⊗ U2 is
a C [G × G]-module, which can be made into a CG-module by letting g ∈ G act as (g, g) ∈ G × G. This
CG-module U1 ⊗ U2 is called the inner tensor product214 of U1 and U2 , and is a restriction of the outer tensor
product U1 ⊗ U2 using the inclusion map G → G × G, g 7→ (g, g).
Let n ≥ 0, and let sgnSn be the 1-dimensional CSn -module C on which everyLg ∈ Sn acts as multi-
plication by sgn(g). If V is a CSn -module, show that the involution on A(S) = n≥0 R(Sn ) defined in
Theorem 4.4.1(b) sends χV 7→ χsgnSn ⊗V where sgnSn ⊗V is the inner tensor product of sgnSn and V . Use
this to show that this involution is a nontrivial PSH-automorphism of A(S), and deduce Theorem 4.4.1(b).
Exercise 4.4.5. (a) Show that for every n ≥ 0, every g ∈ Sn and every finite-dimensional CSn -module
V , we have χV (g) ∈ Z.
(b) Show that for every n ≥ 0 and every finite-dimensional CSn -module V , there exists a QSn -module
W such that V ∼= C⊗Q W . (In the representation theorists’ parlance, this says that all representations
of Sn are defined over Q. This part of the exercise requires some familiarity with representation
theory.)
Remark 4.4.6. Parts (a) and (b) of Exercise 4.4.5 both follow from an even stronger result: For every n ≥ 0
and every finite-dimensional CSn -module V , there exists a ZSn -module W which is finitely generated and
free as a Z-module and satisfies V ∼ = C⊗Z W as CSn -modules. This follows from the combinatorial approach
to the representation theory of Sn , in which the irreducible representations of CSn (the Specht modules) are
constructed using Young tableaux and tabloids. See the literature on the symmetric group, e.g., [165], [60,
§7], [199] or [100, Section 2.2] for this approach.
The connection between Λ and A (S) as established in Theorem 4.4.1 benefits both the study of Λ and
that of A (S). The following two exercises show some applications to Λ:
Exercise 4.4.7. If G is a group and U1 and U2 are two CG-modules, then let U1  U2 denote the inner
tensor product of U1 and U2 (as defined in Exercise 4.4.4). Consider also the binary operation ∗ on ΛQ
defined in Exercise 2.9.4(h).
(a) Show that ch (χU1 U2 ) = ch (χU1 ) ∗ ch (χU2 ) for any n ∈ N and any two CSn -modules U1 and U2 .
212This is not in conflict with the notation 1 for the trivial character of G, since 1 = 1 for P = G. Note that 1 is a
G P G P
class function when P is a union of conjugacy classes of G.
213In fact, 1
ConjG (h) is a class function (since ConjG (h) is a conjugacy class), and so αG,h (being the scalar multiple
|ZG (h)| 1ConjG (h) of 1ConjG (h) ) must also be a class function.
214Do not confuse this with the inner product of characters.
 HOPF ALGEBRAS IN COMBINATORICS 115

(b) Use this to obtain a P

new solution for Exercise 2.9.4(h).
(c) Show that sµ ∗ sν ∈ λ∈Par Nsλ for any two partitions µ and ν.
[Hint: For any group G, introduce a binary operation ∗ on RC (G) which satisfies χU1 U2 = χU1 ∗ χU2
for any two CG-modules U1 and U2 .]
Exercise 4.4.8. Define a Q-bilinear map : ΛQ × ΛQ → ΛQ , which will be written in infix notation (that
is, we will write a b instead of (a, b)), by setting
`(λ) `(µ)
gcd(λ ,µ )
YY
pλ pµ = plcm(λii ,µjj ) for any partitions λ and µ.
i=1 j=1
215

(a) Show that ΛQ , equipped with the binary operation , becomes a commutative Q-algebra with unity
p1 .
(b) For every r ∈ Z, define the Q-algebra homomorphism r : ΛQ → Q as in Exercise 2.9.4(c). Show
that 1 f = 1 (f ) 1 for P every f ∈ ΛQ (where 1 denotes the unity of Λ).
(c) Show that sµ sν ∈ λ∈Par Nsλ for any two partitions µ and ν.
(d) Show that f g ∈ Λ for any f ∈ Λ and g ∈ Λ.
[Hint: For every set X, let SX denote the group of all permutations of X. For two sets X and
Y , there is a canonical group homomorphism SX × SY → SX×Y , which is injective if X and Y are
nonempty. For positive integers n and m, this yields an embedding Sn × Sm → S{1,2,...,n}×{1,2,...,m} ,
which, once S{1,2,...,n}×{1,2,...,m} is identified with Snm (using an arbitrary but fixed bijection {1, 2, . . . , n} ×
{1, 2, . . . , m} → {1, 2, . . . , nm}), can be regarded as an embedding Sn ×Sm → Snm and thus allows defining
a CSnm -module IndS Sn ×Sm (U ⊗ V ) for any CSn -module U and any CSm -module V . This gives a binary
nm

operation on A (S). Show that this operation corresponds to under the PSH-isomorphism ch : A (S) → Λ.]
Remark 4.4.9. The statements (and the idea of the solution) of Exercise 4.4.8 are due to Manuel Maia and
Miguel Méndez (see [127] and, more explicitly, [137]), who call the operation the arithmetic product. Li
[115, Thm. 3.5] denotes it by  and relates it to the enumeration of unlabelled graphs.
4.5. Wreath products. Next consider the tower of groups Gn = Sn [Γ] for a finite group Γ, and the Hopf
algebra A = A(G∗ ) =: A(S[Γ]). Recall (from Theorem 4.4.1) that irreducible complex representations
χλ of Sn are indexed by partitions λ in Parn . Index the irreducible complex representations of Γ as
Irr(Γ) = {ρ1 , . . . , ρd }.
Definition 4.5.1. Define for a partition λ in Parn and ρ in Irr(Γ) a representation χλ,ρ of Sn [Γ] in which
σ in Sn and γ = (γ1 , . . . , γn ) in Γn act on the space χλ ⊗ (ρ⊗n ) as follows
σ(u ⊗ (v1 ⊗ · · · ⊗ vn )) = σ(u) ⊗ (vσ−1 (1) ⊗ · · · ⊗ vσ−1 (n) )
(4.5.1)
γ(u ⊗ (v1 ⊗ · · · ⊗ vn )) = u ⊗ (γ1 v1 ⊗ · · · ⊗ γn vn )

Theorem 4.5.2. The irreducible CSn [Γ]-modules are the induced characters
 (1) (d)

S [Γ]
χλ := IndSndegs(λ) [Γ] χλ ,ρ1 ⊗ · · · ⊗ χλ ,ρd

as λ runs through all functions

λ
Irr(Γ) −→ Par
ρi 7−→ λ(i)
Pd

with the property that i=1 |λ(i) | = n. Here, degs(λ) denotes the d-tuple λ(1) , λ(2) , . . . , λ(d) ∈ Nd ,
and Sdegs(λ) is defined as the subgroup S|λ(1) | × S|λ(2) | × · · · × S|λ(d) | of Sn .
Furthermore, one has a PSH-isomorphism
A(S[Γ]) −→ Λ⊗d
χλ 7−→ sλ(1) ⊗ · · · ⊗ sλ(d) .
215This is well-defined, since (p )
λ λ∈Par is a Q-module basis of ΛQ .
116 DARIJ GRINBERG AND VICTOR REINER

Proof. We know from Corollary 4.3.10 that A(S[Γ]) is a PSH, with PSH-basis Σ given byNthe union of all
irreducible characters of all groups Sn [Γ]. Therefore Theorem 3.2.3 tells us that A(S[Γ]) ∼= ρ∈C A(S[Γ])(ρ)
where C is the set of irreducible characters which are also primitive. Just as in the case of Sn , it is clear
from the definition of the coproduct that an irreducible character ρ of Sn [Γ] is primitive if and only if n = 1,
that in this case Sn [Γ] = Γ, and ρ lies in Irr(Γ) = {ρ1 , . . . , ρd }.
The remaining assertions of the theorem will then follow from the definition of the induction product
algebra structure on A(S[Γ]), once we have shown that, for every ρ ∈ Irr(Γ), there is a PSH-isomorphism
sending
A(S) −→ A(S[Γ])(ρ)
(4.5.2)
χλ 7−→ χλ,ρ .
Such an isomorphism comes from applying Proposition 4.1.17 to the semidirect product Sn [Γ] = Sn n Γn ,
so that K = Γn , G = Sn , and fixing V = ρ⊗n as CSn [Γ]-module with structure as defined in (4.5.1) (but
with λ set to (n), so that χλ is the trivial 1-dimensional CSn -module). One obtains for each n, maps
Φ
R(Sn ) R(Sn [Γ])
Ψ

where
Φ
χ 7−→ χ ⊗ (ρ⊗n )
Ψ
α 7−→ HomCΓn (ρ⊗n , α).
Φ
Taking the direct sum of these maps for all n gives maps A(S) A(S[Γ]).
Ψ
These maps are coalgebra morphisms because of their interaction with restriction to Si × Sj . Since
Proposition 4.1.17(iii) gives the adjointness property that
(χ, Ψ(α))A(S) = (Φ(χ), α)A(S[Γ]) ,
one concludes from the self-duality of A(S), A(S[Γ]) that Φ, Ψ are also algebra morphisms. Since they
take genuine characters to genuine characters, they are PSH-morphisms. Since ρ being a simple CΓ-module
implies that V = ρ⊗n is a simple CΓn -module, Proposition 4.1.17(iv) shows that
(4.5.3) (Ψ ◦ Φ)(χ) = χ
for all Sn -characters χ. Hence Φ is an injective PSH-morphism. Using adjointness, (4.5.3) also shows that
Φ sends CSn -simples χ to C[Sn [Γ]]-simples Φ(χ):
(Φ(χ), Φ(χ))A(S[Γ]) = ((Ψ ◦ Φ)(χ), χ)A(S) = (χ, χ)A(S) = 1.
Since Φ(χ) = χ ⊗ (ρ⊗n ) has V = ρ⊗n as a constituent upon restriction to Γn , Frobenius Reciprocity shows
S [Γ]
that the irreducible character Φ(χ) is a constituent of IndΓnn ρ⊗n = ρn . Hence the entire image of Φ lies
in A(S[Γ])(ρ) (due to how we defined A(ρ) in the proof of Theorem 3.2.3), and so Φ must restrict to an
isomorphism as desired in (4.5.2). 

One of Zelevinsky’s sample applications of the theorem is this branching rule.

Pd
Corollary 4.5.3. Given λ = (λ(1) , . . . , λ(d) ) with i=1 |λ(i) | = n, one has
d
(1) (i)
S [Γ] ,...,λ− ,...,λ(d) )

X X
ResSnn−1 [Γ]×Γ χλ = χ(λ ⊗ ρi .
i=1 (i)
λ− ⊆λ(i) :
(i) (i)
|λ /λ− |=1


(We are identifying functions λ : Irr(Γ) → Par with the corresponding d-tuples λ(1) , λ(2) , . . . , λ(d) here.)

Example 4.5.4. For Γ a two-element group, so Irr(Γ) = {ρ1 , ρ2 } and d = 2, then

 
S [Γ]
ResS65 [Γ]×Γ χ((3,1),(1,1)) = χ((3),(1,1)) ⊗ ρ1 + χ((2,1),(1,1)) ⊗ ρ1 + χ((3,1),(1)) ⊗ ρ2 .
 HOPF ALGEBRAS IN COMBINATORICS 117

Proof of Corollary 4.5.3. By Theorem 4.5.2, this is equivalent to computing in the Hopf algebra A := Λ⊗d
the component of the coproduct of sλ(1) ⊗ · · · ⊗ sλ(d) that lies in An−1 ⊗ A1 . Working within each tensor
factor Λ, the Pieri formula implies that the Λ|λ|−1 ⊗ Λ1 -component of ∆(sλ ) is
X
sλ− ⊗ ρ.
λ− ⊆λ:
|λ/λ− |=1

One must apply this in each of the d tensor factors of A = Λ⊗d , then sum on i. 
4.6. General linear groups. We now consider the tower of finite general linear groups Gn = GLn =
GLn (Fq ) and A = A(G∗ ) =: A(GL). Corollary 4.3.10 tells us that A(GL) is a PSH, with PSH-basis Σ given
by the union of all irreducible characters of all groups GLn . Therefore Theorem 3.2.3 tells us that
A(GL) ∼
O
(4.6.1) = A(GL)(ρ)
ρ∈C

where C = Σ ∩ p is the set of primitive irreducible characters.

Definition 4.6.1. Call an irreducible representation ρ of GLn cuspidal for n ≥ 1 if it lies in C, that is, its
restriction to proper parabolic subgroups Pi,j with i+j = n and i, j > 0 contain no nonzero vectors which are
Ki,j -invariant. Given an irreducible character σ of GLn , say that d(σ) = n, and let Cn := {ρ ∈ C : d(ρ) = n}
for n ≥ 1 denote the subset of cuspidal characters of GLn .
Just as was the case for S1 and S1 [Γ] = Γ, every irreducible character ρ of GL1 (Fq ) = F× q is cuspidal.
However, this does not exhaust the cuspidal characters. In fact, one can predict the number of cuspidal
characters in Cn , using knowledge of the number of conjugacy classes in GLn . Let F denote the set of all
nonconstant monic irreducible polynomials f (x) 6= x in Fq [x]. Let Fn := {f ∈ F : deg(f ) = n} for n ≥ 1.
Proposition 4.6.2. The number |Cn | of cuspidal characters of GLn (Fq ) is the number of |Fn | of irreducible
monic degree n polynomials f (x) 6= x in Fq [x] with nonzero constant term.
Proof. We show |Cn | = |Fn | for n ≥ 1 by strong induction on n. For the base case216 n = 1, just as with
the families Gn = Sn and Gn = Sn [Γ], when n = 1 any irreducible character χ of G1 = GL1 (Fq ) gives a
primitive element of A = A(GL), and hence is cuspidal. Since GL1 (Fq ) = F× ×
q is abelian, there are |Fq | = q−1
such cuspidal characters in C1 , which agrees with the fact that there are q − 1 monic (irreducible) linear
polynomials f (x) 6= x in Fq [x], namely F1 := {f (x) = x − c : c ∈ F×
q }.
In the inductive step, use the fact that the number |Σn | of irreducible complex characters χ of GLn (Fq )
equals its number of conjugacy classes. These conjugacy classes are uniquely represented
P by rational canonical
forms, which are parametrized by functions λ : F → Par with the property that f ∈F deg(f )|λ(f )| = n.
On the other hand, P(4.6.1) tells us that |Σn | is similarly parametrized by the functions λ : C → Par having
the property that ρ∈C d(ρ)|λ(ρ)| = n. Thus we have parallel disjoint decompositions
F
C = n≥1 Cn where Cn = {ρ ∈ C : d(ρ) = n}
F
F = n≥1 Fn where Fn = {f ∈ F : deg(f ) = n}
and hence an equality for all n ≥ 1
   
   
λ X λ X
C −→ Par : d(ρ)|λ(ρ)| = n = |Σn | = F −→ Par : deg(f )|λ(f )| = n .
   
ρ∈C f ∈F

Since there is only one partition λ having |λ| = 1 (namely, λ = (1)), this leads to parallel recursions
 
n−1 
G λ X
|Cn | = |Σn | − Ci −→ Par : d(ρ)|λ(ρ)| = n
 
i=1 ρ∈C
 
n−1 
G λ X
|Fn | = |Σn | − Fi −→ Par : deg(f )|λ(f )| = n
 
i=1 f ∈F

216Actually, we don’t need any base case for our strong induction. We nevertheless handle the case n = 1 as a warmup.
118 DARIJ GRINBERG AND VICTOR REINER

and induction implies that |Cn | = |Fn |. 

We shall use the notation 1H for the trivial character of a group H whenever H is a finite group. This
generalizes the notations 1Sn and 1Sλ introduced above.
Example 4.6.3. Taking q = 2, let us list the sets Fn of monic irreducible polynomials f (x) 6= x in F2 [x] of
degree n for n ≤ 3, so that we know how many cuspidal characters of GLn (Fq ) in Cn to expect:
F1 = {x + 1}
F2 = {x2 + x + 1}
F3 = {x3 + x + 1, x3 + x2 + 1}
Thus we expect
• one cuspidal character of GL1 (F2 ), namely ρ1 (= 1GL1 (F2 ) ),
• one cuspidal character ρ2 of GL2 (F2 ), and
• two cuspidal characters ρ3 , ρ03 of GL3 (F2 ).
We will say more about ρ2 , ρ3 , ρ03 in the next section.
Exercise 4.6.4. (a) Show that for n ≥ 2,
1 X n d
(4.6.2) |Cn |(= |Fn |) = µ q
n d
d|n

where µ(m) is the number-theoretic Möbius function of m, that is µ(m) = (−1)d if m P = p1 · · · pd for
d distinct primes, and µ(m) = 0 if m is not squarefree. (Here, the summation sign d|n means a
sum over all positive divisors d of n.)
(b) Show that (4.6.2) also counts the necklaces with n beads of q colors (=equivalence classes under the
Z/nZ-action of cyclic rotation on sequences (a1 , . . . , an ) in Fnq ) which are primitive in the sense that
no nontrivial rotation fixes any of the sequences within the equivalence class. For example, when
q = 2, here are representatives of these primitive necklaces for n = 2, 3, 4:
n = 2 :{(0, 1)}
n = 3 :{(0, 0, 1), (0, 1, 1)}
n = 4 :{(0, 0, 0, 1), (0, 0, 1, 1), (0, 1, 1, 1)}.
4.7. Steinberg’s unipotent characters. Not surprisingly, the (cuspidal) character ι := 1GL1 of GL1 (Fq )
plays a distinguished role. The parabolic subgroup P(1n ) of GLn (Fq ) is the Borel subgroup B of upper
triangular matrices, and we have ιn = IndGL B
n
1B = C[GLn /B] (identifying representations with their
217
characters as usual) . The subalgebra A(GL)(ι) of A(GL) is the Z-span of the irreducible characters σ
that appear as constituents of ιn = IndGL
B
n
1B = C[GLn /B] for some n.
Definition 4.7.1. An irreducible character σ of GLn appearing as a constituent of IndGL B
n
1B = C[GLn /B]
is called a unipotent character. Equivalently, by Frobenius reciprocity, σ is unipotent if it contains a nonzero
B-invariant vector.
In particular, 1GLn is a unipotent character of GLn for each n.

Proposition 4.7.2. One can choose Λ ∼

= A(GL)(ι) in Theorem 3.3.3(g) so that hn 7−→ 1GLn .
 
217Proof. Exercise 4.3.11(d) (applied to G = GL , ` = n, α = (1n ) = 1, 1, . . . , 1 and χ = ι) gives
∗ ∗   i
| {z }
n times
P(1n )
ιn = indn
(1n ) ι
⊗n
= IndGn
P(1n ) InflG(1n ) ι
⊗n
= IndGL
B
n
1B = C [GLn /B] ,
| {z } | {z }
=IndB n =1P(1n ) =1B
GL

where the last equality follows from the general fact that if G is a finite group and H is a subgroup of G, then IndG ∼
H 1H = C [G/H]
as CG-modules.
 HOPF ALGEBRAS IN COMBINATORICS 119

GL2
Proof. Theorem 3.3.1(a) tells us ι2 = IndB 1B must have exactly two irreducible constituents, one of which
is 1GL2 ; call the other one St2 . Choose the isomorphism so as to send h2 7−→ 1GL2 . Then hn 7→ 1GLn follows
from the claim that St⊥ 2 (1GLn ) = 0 for n ≥ 2: one has

X  Ki,j X
∆(1GLn ) = ResG
Pi,j 1GLn
n
= 1GLi ⊗ 1GLj
i+j=n i+j=n

so that St⊥
2 (1GLn ) = (St2 , 1GL2 )1GLn−2 = 0 since St2 6= 1GL2 .


This subalgebra A(GL)(ι), and the unipotent characters χλq corresponding under this isomorphism to the
Schur functions sλ , were introduced by Steinberg [185]. He wrote down χλq as a virtual sum of induced
characters IndGL
Pα 1Pα (= 1Gα · · · 1Gα ), modelled on the Jacobi-Trudi determinantal expression for sλ =
n
1 `

det(hλi −i+j ). Note that IndGL n

Pα 1Pα is the transitive permutation representation C[G/Pα ] for GLn permuting
the finite partial flag variety G/Pα , that is, the set of α-flags of subspaces

{0} ⊂ Vα1 ⊂ Vα1 +α2 ⊂ · · · ⊂ Vα1 +α2 +···+α`−1 ⊂ Fnq

where dimFq Vd = d in each case. This character has dimension equal to |G/Pα |, with formula given by the
q-multinomial coefficient (see e.g. Stanley [183, §1.7]):
 
n [n]!q
=
α q [α1 ]!q · · · [α` ]!q
n
−1
where [n]!q := [n]q [n − 1]q · · · [2]q [1]q and [n]q := 1 + q + · · · + q n−1 = qq−1 .
Our terminology St2 is motivated by the n = 2 special case of the Steinberg character Stn , which is the
unipotent character corresponding under the isomorphism in Proposition 4.7.2 to en = s(1n ) . It can be
defined by the virtual sum
n X
Stn := χq(1 ) = (−1)n−`(α) IndGL n
Pα 1Pα
α

in which the sum runs through all compositions α of n. This turns out to be the genuine character for
GLn (Fq ) acting on the top homology group of its Tits building: the simplicial complex whose vertices are
nonzero proper subspaces V of Fnq , and whose simplices correspond to flags of nested subspaces. One needs
to know that this Tits building has only top homology, so that one can deduce the above character formula
from the Hopf trace formula; see Björner [22].

4.8. Examples: GL2 (F2 ) and GL3 (F2 ). Let’s get our hands dirty.
(2)
Example 4.8.1. For n = 2, there are two unipotent characters, χq = 1GL2 and

(4.8.1) St2 := χ(1,1)

q = 12GL1 − 1GL2 = IndGL
B
2
1B − 1GL2
 
h h2
since the Jacobi-Trudi formula (2.4.9) gives s(1,1) = det 1 = h21 − h2 . The description (4.8.1) for this
1 h1
Steinberg character St2 shows that it has dimension

|GL2 /B| − 1 = (q + 1) − 1 = q

and that one can think of it as follows: consider the permutation action of GL2 on the q+1 lines {`0 , `1 , . . . , `q }
in the projective space P1Fq = GL2 (Fq )/B, and take the invariant subspace perpendicular to the sum of basis
elements e`0 + · · · + e`q .

Example 4.8.2. Continuing the previous example, but taking q = 2, we find that we have constructed two
(2) (1,1)
unipotent characters: 1GL2 = χq=2 of dimension 1, and St2 = χq=2 of dimension q = 2. This lets us identify
the unique cuspidal character ρ2 of GL2 (F2 ), using knowledge of the character table of GL2 (F2 ) ∼
= S3 :
120 DARIJ GRINBERG AND VICTOR REINER

           
1 0 1 1 1 0 0 1 1 1 0 1
, , ,
0 1 0 1 1 1 1 0 1 0 1 1

(2)
1GL2 = χq=2 unipotent 1 1 1
(1,1)
St2 = χq=2 unipotent 2 0 −1
ρ2 cuspidal 1 −1 1
In other words, the cuspidal character ρ2 of GL2 (F2 ) corresponds under the isomorphism GL2 (F2 ) ∼
= S3 to
the sign character sgnS3 .

Example 4.8.3. Continuing the previous example to q = 2 and n = 3 lets us analyze the irreducible
characters of GL3 (F2 ). Recalling our labelling ρ1 , ρ2 , ρ3 , ρ03 from Example 4.6.3 of the cuspidal characters of
GLn (F2 ) for n = 1, 2, 3, Zelevinsky’s Theorem 3.2.3 tells us that the GL3 (F2 )-irreducible characters should
λ
be labelled by functions {ρ1 , ρ2 , ρ3 , ρ03 } −→ Par for which

1 · |λ(ρ1 )| + 2 · |λ(ρ2 )| + 3 · |λ(ρ3 )| + 3 · |λ(ρ03 )| = 3

0
We will label such an irreducible character χλ = χ(λ(ρ1 ),λ(ρ2 ),λ(ρ3 ),λ(ρ3 )) .
Three of these irreducibles will be the unipotent characters, mapping under the isomorphism from Propo-
sition 4.7.2 as follows:
• s(3) = h3 7−→ χ((3),∅,∅,∅) = 1GL3 of dimension 1.
•
 
h2 h3
s(2,1) = det = h2 h1 − h3 7−→ χ((2,1),∅,∅,∅) = IndGL
P2,1 1P2,1 − 1GL3 ,
3
1 h1
   
3 3 q=2
of dimension − = [3]q − 1 = q 2 + q 6.
2, 1 q 3 q
• Lastly,
 
h1 h2 h3
s(1,1,1) = det  1 h1 h2  = h31 − h2 h1 − h1 h2 + h3
0 1 h1
7−→ St3 = χ((1,1,1),∅,∅,∅) = IndGL
B
3
1B − IndGL GL3
P2,1 1P2,1 − IndP1,2 1P1,2 + 1GL3
3

of dimension
       
3 3 3 3
− − +
1, 1, 1 q 2, 1 q 1, 2 q 3 q
q=2
= [3]!q − [3]q − [3]q + 1 = q 3 8.
There should also be one non-unipotent, non-cuspidal character, namely
P
χ((1),(1),∅,∅) = ρ1 ρ2 = IndGL 1,2


P1,2 InflGL1 ×GL2 1GL1 ⊗ ρ2
3

 
3 q=2
having dimension · 1 · 1 = [3]q 7.
1, 2 q
Finally, we expect cuspidal characters ρ3 = χ(∅,∅,(1),∅) , ρ03 = χ(∅,∅,∅,(1)) , whose dimensions d3 , d03 can be
deduced from the equation

12 + 62 + 82 + 72 + d23 + (d03 )2 = |GL3 (F2 )| = (q 3 − q 0 )(q 3 − q 1 )(q 3 − q 2 ) q=2 = 168.

 

This forces d23 + (d03 )2 = 18, whose only solution in positive integers is d3 = d03 = 3.
We can check our predictions of the dimensions for the various GL3 (F2 )-irreducible characters since
GL3 (F2 ) is the finite simple group of order 168 (also isomorphic to P SL2 (F7 )), with known character table
(see James and Liebeck [89, p. 318]):
 HOPF ALGEBRAS IN COMBINATORICS 121

centralizer order 168 8 4 3 7 7

unipotent?/cuspidal?
1GL3 = χ((3),∅,∅,∅) unipotent 1 1 1 1 1 1
χ((2,1),∅,∅,∅) unipotent 6 2 0 0 −1 −1
St3 = χ((1,1,1),∅,∅,∅) unipotent 8 0 0 −1 1 1
χ((1),(1),∅,∅) 7 −1 −1 1 0 0
ρ3 = χ(∅,∅,(1),∅) cuspidal 3 −1 1 0 α α
ρ03 = χ(∅,∅,∅,(1)) cuspidal 3 −1 1 0 α α
√
Here α := −1/2 + i 7/2.
Remark 4.8.4. It is known (see e.g. Bump [29, Cor. 7.4]) that, for n ≥ 2, the dimension of any cuspidal
irreducible character ρ of GLn (Fq ) is
(q n−1 − 1)(q n−2 − 1) · · · (q 2 − 1)(q − 1).
Note that when q = 2,
• for n = 2 this gives 21 − 1 = 1 for the dimension of ρ2 , and
• for n = 3 it gives (22 − 1)(2 − 1) = 3 for the dimensions of ρ3 , ρ03 ,
agreeing with our calculations above. Much more is known about the character table of GLn (Fq ); see
Remark 4.9.14 below, Zelevinsky [203, Chap. 11], and Macdonald [125, Chap. IV].
4.9. The Hall algebra. There is another interesting Hopf subalgebra (and quotient Hopf algebra) of A(GL),
related to unipotent conjugacy classes in GLn (Fq ).
Definition 4.9.1. Say that an element g in GLn (Fq ) is unipotent if its eigenvalues are all equal to 1.
Equivalently, g ∈ GLn (Fq ) is unipotent if and only if g − idFnq is nilpotent. A conjugacy class in GLn (Fq ) is
unipotent if its elements are unipotent.
Denote by Hn the C-subspace of RC (GLnL ) consisting of those class functions which
Lare supported only
on unipotent conjugacy classes, and let H = n≥0 Hn as a C-subspace of AC (GL) = n≥0 RC (GLn ).

Proposition 4.9.2. The subspace H is a Hopf subalgebra of AC (GL), which is graded, connected, and of
finite type, and self-dual with respect to the inner product on class functions inherited from AC (GL). It is
also a quotient Hopf algebra of AC (GL), as the C-linear surjection AC (GL)  H restricting class functions
to unipotent classes is a Hopf algebra homomorphism. This surjection has kernel H⊥ , which is both an ideal
and a two-sided coideal.

Proof. It is immediately clear that H⊥ is a graded C-vector subspace of AC (GL), whose n-th homogeneous
component consists of those class functions on GLn whose values on all unipotent classes are 0. (This holds
no matter whether the perpendicular space is taken with respect to the Hermitian form (·, ·)G or with respect
to the bilinear form h·, ·iG .) In other words, H⊥ is the kernel of the surjection AC (GL)  H defined in the
proposition.
Given two class functions χi , χj on GLi , GLj and g in GLi+j , one has
1 X
(4.9.1) (χi · χj ) (g) = χi (gi )χj (gj ).
|Pi,j |
h∈GL
 i+j : 

−1 g i ∗
h gh= ∈Pi,j
0 gj

Since g is unipotent if and only if h−1 gh is unipotent if and only if both gi , gj are unipotent, the formula
(4.9.1) shows both that H is a subalgebra218 and that H⊥ is a two-sided ideal219. It also shows that the
surjection AC (GL)  H restricting every class function to unipotent classes is an algebra homomorphism220.

218Indeed, if χ and χ are both supported only on unipotent classes, then the same holds for χ · χ .
i j i j
219In fact, if one of χ and χ annihilates all unipotent classes, then so does χ · χ .
i j i j
220because if g is unipotent, then the only values of χ and χ appearing on the right hand side of (4.9.1) are those on
i j
unipotent elements
122 DARIJ GRINBERG AND VICTOR REINER

Similarly, for class functions χ on GLn and (gi , gj ) in GLi,j = GLi × GLj , one has
 
1 X gi k
∆(χ)(gi , gj ) = ij χ
q i×j
0 gj
k∈Fq

using (4.1.13). This shows both that H is a sub-coalgebra of A = AC (GL) (that is, it satisfies ∆H ⊂ H ⊗ H)
and that H⊥ is a two-sided coideal (that is, we have ∆(H⊥ ) ⊂ H⊥ ⊗ A + A ⊗ H⊥ ), since it shows that if χ is
supported only on unipotent classes, then ∆(χ) vanishes on (g1 , g2 ) that have either g1 or g2 non-unipotent.
It also shows that the surjection AC (GL)  H restricting every class function to unipotent classes is a
coalgebra homomorphism. The rest follows. 
The subspace H is called the Hall algebra. It has an obvious orthogonal C-basis, with interesting structure
constants.
Definition 4.9.3. Given a partition λ of n, let Jλ denote the GLn -conjugacy class of unipotent matrices
whose Jordan type (that is, the list of the sizes of the Jordan blocks, in decreasing order) is given by λ.
Furthermore, let zλ (q) denote the size of the centralizer of any element of this conjugacy class Jλ .
The indicator class functions221 {1Jλ }λ∈Par form a C-basis for H whose multiplicative structure constants
λ
are called the Hall coefficients gµ,ν (q):
X
λ
1Jµ 1Jν = gµ,ν (q) 1Jλ .
λ

Because the dual basis to {1Jλ } is {zλ (q)1Jλ }, self-duality of H shows that the Hall coefficients are (essen-
tially) also structure constants for the comultiplication:
X
λ zµ (q)zν (q)
∆1Jλ = gµ,ν (q) · 1Jµ ⊗ 1Jν .
µ,ν
zλ (q)
λ
The Hall coefficient gµ,ν (q) has the following interpretation.

Proposition 4.9.4. Fix any g in GLn (Fq ) acting unipotently on Fnq with Jordan type λ. Then gµ,ν λ
(q)
n
counts the g-stable Fq -subspaces V ⊂ Fq for which the restriction g|V acts with Jordan type µ, and the
induced map ḡ on the quotient space Fnq /V has Jordan type ν.

Proof. Given µ, ν partitions of i, j with i + j = n, taking χi , χj equal to 1Jµ , 1Jν in (4.9.1) shows that for
 
any g in GLn , the value of 1Jµ · 1Jν (g) is given by
   
1 −1 gi ∗
(4.9.2) h ∈ GLn : h gh = with gi ∈ Jµ , gj ∈ Jν .
|Pi,j | 0 gj
Let S denote the set appearing in (4.9.2), and let Fiq denote the i-dimensional subspace of Fnq spanned by
the first i standard basis vectors. Note that the condition on an element h in S saying that h−1 gh is in
block upper-triangular form can be re-expressed by saying that the subspace V := h(Fiq ) is g-stable. One
ϕ
then sees that the map h 7−→ V = h(Fiq ) surjects S onto the set of i-dimensional g-stable subspaces V of
Fnq for which g|V and ḡ are unipotent of types µ, ν, respectively. Furthermore, for any particular such V ,
its fiber ϕ−1 (V ) in S is a coset of the stabilizer within GLn of V , which is conjugate to Pi,j , and hence has
cardinality |ϕ−1 (V )| = |Pi,j |. This proves the assertion of the proposition. 
The Hall algebra H will turn out to be isomorphic to the ring ΛC of symmetric functions with C coefficients,
via a composite ϕ of three maps
ΛC −→ A(GL)(ι)C −→ A(GL)C −→ H
in which the first map is the isomorphism from Proposition 4.7.2, the second is inclusion, and the third is
the quotient map from Proposition 4.9.2.
221Here we use the following notation: Whenever P is a subset of a group G, we denote by 1 the map G → C which
P
sends every element of P to 1 and all remaining elements of G to 0. This is not in conflict with the notation 1G for the trivial
character of G, since 1P = 1G for P = G. Note that 1P is a class function when P is a union of conjugacy classes of G.
 HOPF ALGEBRAS IN COMBINATORICS 123

Theorem 4.9.5. The above composite ϕ is a Hopf algebra isomorphism, sending

P
hn 7−→ λ∈Parn 1Jλ ,
n
en 7−→ q 2 ) 1J(1n ) ,
(
P
pn 7−→ λ∈Parn (q; q)`(λ)−1 1Jλ (for n > 0) ,
where
(x; q)m := (1 − x)(1 − qx)(1 − q 2 x) · · · (1 − q m−1 x) for all m ∈ N.

Proof. That ϕ is a graded Hopf morphism follows because it is a composite of three such morphisms. We
claim that once one shows the formula for the (nonzero) image ϕ(pn ) given above is correct, then this
will already show ϕ is an isomorphism, by the following argument. Note first that ΛC and H both have
dimension | Parn | for their nth homogeneous components, so it suffices to show that the graded map ϕ is
injective. On the other hand, both ΛC and H are (graded, connected, finite type) self-dual Hopf algebras
(although with respect to a sesquilinear form), so Theorem 3.1.7 says that each is the symmetric algebra on
its space of primitive elements. Thus it suffices to check that ϕ is injective when restricted to their subspaces
of primitives.222 For ΛC , by Corollary 3.1.8 the primitives are spanned by {p1 , p2 , . . .}, with only one basis
element in each degree n ≥ 1. Hence ϕ is injective on the subspace of primitives if and only if it does not
annihilate any pn .
Thus it only remains to show the above formulas for the images of hn , en , pn under ϕ. This is clear for
hn , since Proposition 4.7.2 shows that it maps Punder the first two composites to the indicator function 1GLn
which then restricts to the sum of indicators λ∈Parn 1Jλ in H. For en , pn , we resort to generating functions.
Let h̃n , ẽn , p̃n denote the three putative images in H of hn , en , pn , appearing on the right side in the theorem,
and define generating functions
X X X
H̃(t) := h̃n tn , Ẽ(t) := ẽn tn , P̃ (t) := p̃n+1 tn in H[[t]].
n≥0 n≥0 n≥0

We wish to show that the map ϕ[[t]] : ΛC [[t]] → H[[t]] (induced by ϕ) maps H(t), E(t), P (t) in Λ[[t]] to these
three generating functions223. Since we have already shown this is correct for H(t), by (2.4.3), (2.5.13), it
suffices to check that in H[[t]] one has
Pn k
H̃(t)Ẽ(−t) = 1, or equivalently, k=0 (−1) ẽk h̃n−k = δ0,n ;
n
H̃ 0 (t)Ẽ(−t) = P̃ (t), or equivalently, k
P
k=0 (−1) (n − k)ẽk h̃n−k = p̃n .

Thus it would be helpful to evaluate the class function ẽk h̃n−k . Note that a unipotent g in GLn having `
Jordan blocks has an `-dimensional 1-eigenspace,

so that the number of k-dimensional g-stable Fq -subspaces
of Fnq on which g has Jordan type 1k (that is, on which g acts as the identity) is the q-binomial coefficient
 
` (q; q)`
= ,
k q (q; q)k (q; q)`−k
counting k-dimensional Fq -subspaces V of an `-dimensional Fq -vector space; see, e.g., [183, §1.7]. Hence, for
a unipotent g in GLn having ` Jordan blocks, we have
 
(k
)
 
(k
)
X  
( k
) `
(ẽk h̃n−k )(g) = q · 1J k · h̃n−k (g) = q ·
2 2 1J k · 1Jν (g) = q 2
(1 ) (1 ) k q
ν∈Parn−k

(by Proposition 4.9.4). Thus one needs for ` ≥ 1 that

`  
k `
(−1)k q (2)
X
(4.9.3) = 0,
k q
k=0
`  
X
k (k
) `
(4.9.4) (−1) (n − k)q 2 = (q; q)`−1 .
k q
k=0

222An alternative way to see that it suffices to check this is by recalling Exercise 1.4.32(c).
223See (2.4.1), (2.4.2), (2.5.13) for the definitions of H(t), E(t), P (t).
124 DARIJ GRINBERG AND VICTOR REINER

Identity (4.9.3) comes from setting x = 1 in the q-binomial theorem [183, Exer. 3.119]:
`  
X
k (k ) `
(−1) q 2 x`−k = (x − 1)(x − q)(x − q 2 ) · · · (x − q `−1 ).
k q
k=0
d
Identity (4.9.4) comes from taking dx in the q-binomial theorem, then setting x = 1, and finally adding
(n − `) times (4.9.3). 
(1) (2) (k)
Exercise
 4.9.6. Fix a prime  power q. For any k ∈ N, and any k partitions λ , λ , . . . , λ , we define a
family gλλ(1) ,λ(2) ,...,λ(k) (q) of elements of C by the equation
λ∈Par
X
1J (1) 1J (2) · · · 1J (k) = gλλ(1) ,λ(2) ,...,λ(k) (q) 1Jλ
λ λ λ
λ∈Par
λ
in H. This notation generalizes the notation (q) we introduced in Definition 4.9.3. Note that gµλ (q) = δλ,µ
gµ,ν
for any two partitions λ and µ, and that g (q) = δλ,∅ for any partition λ (where g λ (q) is to be understood
λ

as gλλ(1) ,λ(2) ,...,λ(k) (q) for k = 0).

(a) Let λ ∈ Par, and let n = |λ|. Let V be an n-dimensional Fq -vector space, and let g be a unipotent
endomorphism of V
having Jordan type λ. Let k ∈ N, and let λ(1) , λ(2) , . . . , λ(k) be k partitions. A
λ(1) , λ(2) , . . . , λ(k) -compatible g-flag will mean a sequence 0 = V0 ⊂ V1 ⊂ V2 ⊂ · · · ⊂ Vk = V of
g-invariant Fq -vector subspaces Vi of V such that for every i ∈ {1, 2, . . . , k}, the endomorphism of
224
Vi /Vi−1 induced by g has Jordan type λ(i) .
Show that gλλ(1) ,λ(2) ,...,λ(k) (q) is the number of λ(1) , λ(2) , . . . , λ(k) -compatible g-flags.225

(b) Let λ ∈ Par. Let k ∈ N, and let λ(1) , λ(2) , . . . , λ(k) be k partitions. Show that gλλ(1) ,λ(2) ,...,λ(k) (q) = 0
unless λ(1) + λ(2) + · · · + λ(k) = |λ| and λ(1) + λ(2) + · · · + λ(k) . λ. (Here and in the following,
we are using the notations of Exercise 2.9.17).
(c) Let λ ∈ Par, and let us write the transpose partition λt as λt = ((λt )1 , (λt )2 , . . . , (λt )` ). Show that
gλ (λt )   (λt )   (λt )  (q) 6= 0.
1 1 , 1 2 ,..., 1 `

(d) Let n ∈ N and λ ∈ Parn . Show that

X
ϕ (eλ ) = αλ,µ 1Jµ
µ∈Parn ; λt .µ

for some coefficients αλ,µ ∈ C satisfying αλ,λt 6= 0.

(e) Give another proof of the fact that the map ϕ is injective.
[Hint: For (b), use Exercise 2.9.22(b).]
We next indicate, without proof, how H relates to the classical Hall algebra.
Definition 4.9.7. Let p be a prime. The usual Hall algebra, or what Schiffmann [168, §2.3] calls Steinitz’s
classical Hall algebra (see also Macdonald [125, Chap. II]), has Z-basis elements {uλ }λ∈Par , with the multi-
λ
plicative structure constants gµ,ν (p) in
X
λ
uµ uν = gµ,ν (p) uλ
λ
defined as follows: fix a finite abelian p-group L of type λ, meaning that
`(λ)
L∼
M
= Z/pλi Z,
i=1

224This is well-defined. In fact, both V and V

i i−1 are g-invariant, so that g restricts to an endomorphism of Vi , which further
restricts to an endomorphism of Vi−1 , and thus gives rise to an endomorphism of Vi /Vi−1 .
225This can be seen as a generalization of Proposition 4.9.4. In fact, if µ and ν are two partitions, then a (µ, ν)-compatible
g-flag is a sequence 0 = V0 ⊂ V1 ⊂ V2 = V of g-invariant Fq -vector subspaces Vi of V such that the endomorphism of V1 /V0 ∼ = V1
induced by g has Jordan type µ, and the endomorphism of V2 /V1 ∼ = V /V1 induced by g has Jordan type ν. Choosing such
a sequence amounts to choosing V1 (since there is only one choice for each of V0 and V2 ), and the conditions on this V1 are
precisely the conditions on V in Proposition 4.9.4.
 HOPF ALGEBRAS IN COMBINATORICS 125

λ
and let gµ,ν (p) be the number of subgroups M of L of type µ, for which the quotient N := L/M is of type ν.
λ
In other words, gµ,ν (p) counts, for a fixed abelian p-group L of type λ, the number of short exact sequences
0 → M → L → N → 0 in which M, N have types µ, ν, respectively (modulo isomorphism of short exact
sequences restricting to the identity on L).
We claim that when one takes the finite field Fq of order q = p a prime, the Z-linear map
(4.9.5) uλ 7−→ 1Jλ
gives an isomorphism from this classical Hall algebra to the Z-algebra HZ ⊂ H. The key point is Hall’s
Theorem, a non-obvious statement for which Macdonald includes two proofs in [125, Chap. II], one of them
due to Zelevinsky226. To state it, we first recall some notions about discrete valuation rings.
Definition 4.9.8. A discrete valuation ring (DVR) o is a principal ideal domain having only one maximal
ideal m 6= 0, with quotient k = o/m called its residue field.
The structure theorem for finitely generated modules over a PID implies that an o-module L with finite
composition series of composition length n must have L ∼
L`(λ)
= i=1 o/mλi for some partition λ of n; say L has
type λ in this situation.
Here are the two crucial examples for us.
Example 4.9.9. For any field F, the power series ring o = F[[t]] is a DVR with maximal ideal m = (t)
and residue field k = o/m = F[[t]]/(t) ∼ = F. An o-module L of type λ is an F-vector space together with an
F-linear transformation T ∈ End L that acts on L nilpotently (so that g := T + 1 acts unipotently, where
1 = idL ) with Jordan blocks of sizes given by λ: each summand o/mλi = F[[t]]/(tλi ) of L has an F-basis
{1, t, t2 , . . . , tλi −1 } on which the map T that multiplies by t acts as a nilpotent Jordan block of size λi . Note
also that, in this setting, o-submodules are the same as T -stable (or g-stable) F-subspaces.
Example 4.9.10. The ring of p-adic integers o = Zp is a DVR with maximal ideal m = (p) and residue field
k = o/m = Zp /pZp ∼= Z/pZ. An o-module L of type λ is an abelian p-group of type λ: for each summand,
o/mλi = Zp /pλi Zp ∼
= Z/pλi Z. Note also that, in this setting, o-submodules are the same as subgroups.
P
One last notation: n(λ) := i≥1 (i − 1)λi , for λ in Par. Hall’s Theorem is as follows.

Theorem 4.9.11. Assume o is a DVR with maximal ideal m, and that its residue field k = o/m is finite of
cardinality q. Fix an o-module L of type λ. Then the number of o-submodules M of type µ for which the
quotient N = L/M is of type ν can be written as the specialization
λ
[gµ,ν (t)]t=q
λ
of a polynomial gµ,ν (t) in Z[t], called the Hall polynomial.
λ
Furthermore, the Hall polynomial gµ,ν (t) has degree at most n(λ) − (n(µ) + n(ν)), and its coefficient of
tn(λ)−(n(µ)+n(ν)) is the Littlewood-Richardson coefficient cλµ,ν .

Comparing what Hall’s Theorem says in Examples 4.9.9 and 4.9.10, shows that the map (4.9.5) gives the
desired isomorphism from the classical Hall algebra to HZ .
We close this section with some remarks on the vast literature on Hall algebras that we will not discuss
here.
Remark 4.9.12. Macdonald’s version of Hall’s Theorem [125, (4.3)] is stronger than Theorem 4.9.11, and use-
λ
ful for certain applications: he shows that gµ,ν (t) is the zero polynomial whenever the Littlewood-Richardson
λ
coefficient cµ,ν is zero.
λ
Remark 4.9.13. In general, not all coefficients of the Hall polynomials gµ,ν (t) are nonnegative (see But-
λ λ
ler/Hales [31] for a study of when they are); it often happens that gµ,ν (1) = 0 despite gµ,ν (t) not being the

226See also [168, Thm. 2.6, Prop. 2.7] for quick proofs of part of it, similar to Zelevinsky’s. Another proof, based on a
recent category-theoretical paradigm, can be found in [52, Theorem 3.53].
126 DARIJ GRINBERG AND VICTOR REINER

zero polynomial227. However, in [95, Thm. 4.2], Klein showed that the polynomial values gµ,ν
λ
(p) for p prime
λ
are always positive when cµ,ν 6= 0. (This easily yields the same result for p a prime power.)
Remark 4.9.14. Zelevinsky in [203, Chaps 10, 11] uses the isomorphism ΛC → H to derive J. Green’s formula
for the value of any irreducible character χ of GLn on any unipotent class Jλ . The answer involves values
of irreducible characters of Sn along with Green’s polynomials Qλµ (q) (see Macdonald [125, §III.7]; they are
denoted Q(λ, µ) by Zelevinsky), which express the images under the isomorphism of Theorem 4.9.5 of the
symmetric function basis {pµ } in terms of the basis {1Jλ }.
λ
Remark 4.9.15. The Hall polynomials gµ,ν (t) also essentially give the multiplicative structure constants for
Λ(x)[t] with respect to its basis of Hall-Littlewood symmetric functions Pλ = Pλ (x; t):
X
Pµ Pν = tn(λ)−(n(µ)+n(ν)) gµ,ν
λ
(t−1 )Pλ .
λ
See Macdonald [125, §III.3].
Remark 4.9.16. Schiffmann [168] discusses self-dual Hopf algebras which vastly generalize the classical Hall
algebra called Ringel-Hall algebras, associated to abelian categories which are hereditary. Examples come
from categories of nilpotent representations of quivers; the quiver having exactly one node and one arc
recovers the classical Hall algebra HZ discussed above.
Remark 4.9.17. The general linear groups GLn (Fq ) are one of four families of so-called classical groups.
Progress has been made on extending Zelevinsky’s PSH theory to the other families:
(a) Work of Thiem and Vinroot [193] shows that Lthe tower {G∗ } of finite unitary groups Un (Fq2 ) give
rise to another positive self-dual Hopf algebra A = n≥0 R(Un (Fq2 )), in which the role of Harish-Chandra
induction is played by Deligne-Lusztig induction. In this theory, character and degree formulas for Un (Fq2 )
are related to those of GLn (Fq ) by substituting q 7→ −q, along with appropriate scalings by ±1, a phenomenon
sometimes called Ennola duality. See also L [184, §4]. L L


(b) van Leeuwen [112] has studied n≥0 R (Sp2n (Fq )), n≥0 R (O2n (Fq )) and n≥0 R Un Fq2 not
as Hopf algebras, but rather as so-called twisted PSH-modules over the PSH A(GL) (a “deformed” version of
the older notion of Hopf modules). He classified these PSH-modules axiomatically similarly to Zelevinsky’s
above classification of PSH’s.
(c) In a recent honors thesis [178], Shelley-Abrahamson defined
L yet another variation
Lof the concept of Hopf
modules, named 2-compatible Hopf modules, and identified n≥0 R (Sp2n (Fq )) and n≥0 R (O2n+1 (Fq )) as
such modules over A(GL).

227Actually, Butler/Hales show in [31, proof of Prop. 2.4] that the values g λ (1) are the structure constants of the ring Λ
µ,ν
with respect to its basis (mλ )λ∈Par : we have
X
λ
mµ mν = gµ,ν (1)mλ
λ∈Par
for all partitions µ and ν.
 HOPF ALGEBRAS IN COMBINATORICS 127

5. Quasisymmetric functions and P -partitions

We discuss here our next important example of a Hopf algebra arising in combinatorics: the quasisym-
metric functions of Gessel [66], with roots in work of Stanley [180] on P -partitions.

5.1. Definitions, and Hopf structure. The definitions of quasisymmetric functions require a totally
ordered variable set. Usually we will use a variable set denoted x = (x1 , x2 , . . .) with the usual ordering
x1 < x2 < . . .. However, it is good to have some flexibility in changing the ordering, which is why we make
the following definition.
Definition 5.1.1. Given any totally ordered set I, create a totally ordered variable set {xi }i∈I , and then
let R({xi }i∈I ) denote the power series of bounded degree in {xi }i∈I having coefficients in k.
The quasisymmetric functions QSym({xi }i∈I ) over the alphabet ({xi }i∈I ) will be the k-submodule con-
α`
sisting of the elements f in R({xi }i∈I ) that have the same coefficient on the monomials xα i1 · · · xi` and
1

α1 α`
xj1 · · · xj` whenever both i1 < · · · < i` and j1 < · · · < j` in the total order on I. We write QSymk ({xi }i∈I )
instead of QSym({xi }i∈I ) to stress the choice of base ring k.
It immediately follows from this definition that QSym({xi }i∈I ) is a free k-submodule of R({xi }i∈I ), having
as k-basis elements the monomial quasisymmetric functions
X α`
Mα ({xi }i∈I ) := xαi1 · · · xi`
1

i1 <···<i` in I

for all compositions α satisfying `(α) ≤ |I|. When I is infinite, this means that the Mα for all compositions
α form a basis of QSym({xi }i∈IL ).
Note that QSym({xi }i∈I ) = n≥0 QSymn ({xi }i∈I ) is a graded k-module of finite type, where QSymn ({xi }i∈I )
is the k-submodule of quasisymmetric functions which are homogeneous of degree n. Letting Comp denote
the set of all compositions α, and Compn the compositions α of n (that is, compositions whose parts sum
to n), the subset {Mα }α∈Compn ; `(α)≤|I| gives a k-basis for QSymn ({xi }i∈I ).
Example 5.1.2. Taking the variable set x = (x1 < x2 < · · · ) to define QSym(x), for n = 0, 1, 2, 3, one has
these basis elements in QSymn (x):
M() = M∅ =1

M(1) = x1 + x2 + x3 + · · · = m(1) = s(1) = e1 = h1 = p1

M(2) = x21 + x22 + x23 + · · · = m(2) = p2

M(1,1) = x1 x2 + x1 x3 + x2 x3 + · · · = m(1,1) = e2

M(3) = x31 + x32 + x33 + · · · = m(3) = p3

M(2,1) = x21 x2 + x21 x3 + x22 x3 + · · ·
M(1,2) = x1 x22 + x1 x23 + x2 x23 + · · ·
M(1,1,1) = x1 x2 x3 + x1 x2 x4 + x1 x3 x4 + · · · = m(1,1,1) = e3
It is not obvious that QSym(x) is a subalgebra of R(x), but we will show this momentarily. For example,
M(a) M(b,c) = (xa1 + xa2 + xa3 + · · · )(xb1 xc2 + xb1 xc3 + xb2 xc3 + · · · )
= xa+b
1 xc2 + · · · + xb1 xa+c
3 + · · · + xa1 xb2 xc3 + · · · + xb1 xa2 xc3 + · · · + xb1 xc2 xa3 + · · ·
= M(a+b,c) + M(b,a+c) + M(a,b,c) + M(b,a,c) + M(b,c,a)

Proposition 5.1.3. For any infinite totally ordered set I, one has that QSym({xi }i∈I ) is a k-subalgebra of
R({xi }i∈I ), with multiplication in the {Mα }-basis as follows: Fix three disjoint chain posets (i1 < · · · < i` ),
(j1 < · · · < jm ) and (k1 < k2 < · · · ). Now, if α = (α1 , . . . , α` ) and β = (β1 , . . . , βm ) are two compositions,
then
X
(5.1.1) Mα Mβ = Mwt f
f
128 DARIJ GRINBERG AND VICTOR REINER

in which the sum is over all p ∈ N and all maps f from the disjoint union of two chains to a chain
f
(5.1.2) (i1 < · · · < i` ) t (j1 < · · · < jm ) −→ (k1 < · · · < kp )
which are both surjective and strictly order-preserving (x < P y implies f (x) < P f (y)), and where the compo-
sition wt(f ) := (wt1 (f ), . . . , wtp (f )) is defined by wts (f ) := iu ∈f −1 (ks ) αu + jv ∈f −1 (ks ) βv .

Example 5.1.4. For this example, set α = (2, 1) and β = (3, 4, 2). Let us compute Mα Mβ using (5.1.1).
Indeed, the length of α is ` = 2, and the length of β is m = 3, so the sum on the right hand side of
(5.1.1) is a sum over all p ∈ N and all surjective strictly order-preserving maps f from the disjoint union
(i1 < i2 ) t (j1 < j2 < j3 ) of two chains to the chain (k1 < k2 < · · · < kp ). Such maps can exist only when
p ≤ 5 (due to having to be surjective) and only for p ≥ 3 (since, being strictly order-preserving, they have
to be injective when restricted to (j1 < j2 < j3 )). Hence, enumerating them is a finite problem. The reader
can check that the value obtained fo Mα Mβ is
M(2,1,3,4,2) + M(2,3,1,4,2) + M(2,3,4,1,2) + M(2,3,4,2,1) + M(3,2,1,4,2)
+ M(3,2,4,1,2) + M(3,2,4,2,1) + M(3,4,2,1,2) + M(3,4,2,2,1) + M(3,4,2,2,1)
+ M(2,3,4,3) + M(2,3,5,2) + M(2,4,4,2) + M(3,2,4,3) + M(3,2,5,2) + M(3,4,2,3)
+ M(3,4,4,1) + M(3,6,1,2) + M(3,6,2,1) + M(5,1,4,2) + M(5,4,1,2) + M(5,4,2,1)
+ M(5,4,3) + M(5,5,2) + M(3,6,3) .
Here, we have listed the addends corresponding to p = 5 on the first two rows, the addends corresponding
to p = 4 on the next two rows, and those corresponding to p = 3 on the third row. The reader might notice
that the first two rows (i.e., the addends with p = 5) are basically a list of shuffles of α and β: In general,
228
the maps (5.1.2) for p = ` + m are in bijection with the elements of Sh`,m , and the corresponding
compositions wt f are the shuffles of α and β. Therefore the name “overlapping shuffle product”.
Proof of Proposition 5.1.3. It clearly suffices to prove the formula (5.1.1). Let α = (α1 , . . . , α` ) and β =
(β1 , . . . , βm ) be two compositions. Fix three disjoint chain posets (i1 < · · · < i` ), (j1 < · · · < jm ) and
(k1 < k2 < · · · ).
Thus, multiplying Mα = u1 <···<u` xα α` β1 βm
P P
u1 · · · xu` with Mβ = v1 <···<vm xv1 · · · xvm , we obtain
1

X X
xα α`

β1
xv1 · · · xβvm


M α Mβ = u1 · · · xu`
1 m

u1 <···<u` v1 <···<vm
X X
(5.1.3) = Nwγ 1 ,...,wp xγw11 · · · xγwpp ,
γ=(γ1 ,...,γp )∈Comp w1 <···<wp in I

where Nwγ 1 ,...,wp is the number of all pairs

(5.1.4) ((u1 < · · · < u` ) , (v1 < · · · < vm )) ∈ I ` × I m

of two strictly increasing tuples satisfying
xα α`
xβv11 · · · xβvm = xγw11 · · · xγwpp .



(5.1.5) u1 · · · xu`
1 m

229
Thus, we need to show that Nwγ 1 ,...,wp (for a given γ = (γ1 , . . . , γp ) ∈ Comp and a given (w1 < · · · < wp ) ∈
I p ) is also the number of all surjective strictly order-preserving maps
f
(5.1.6) (i1 < · · · < i` ) t (j1 < · · · < jm ) −→ (k1 < · · · < kp ) satisfying wt(f ) = γ
(because then, (5.1.3) will simplify to (5.1.1)).
In order to show this, it suffices to construct a bijection from the set of all pairs (5.1.4) satisfying (5.1.5)
to the set of all surjective strictly order-preserving maps (5.1.6). This bijection is easy to construct: Given

228The bijection takes a map f to the inverse of the permutation σ ∈ S which sends every x ∈ {1, 2, . . . , `} to the index y
p
satisfying f (ix ) = ky , and sends every x ∈ {` + 1, ` + 2, . . . , ` + m} to the index y satisfying f (jx−` ) = ky .
229In the second equality in (5.1.3), we have used the fact that each monomial can be uniquely written in the form xγ1 · · · xγp
w1 wp
for some composition γ = (γ1 , . . . , γp ) ∈ Comp and some strictly increasing tuple (w1 < · · · < wp ) ∈ I p .
 HOPF ALGEBRAS IN COMBINATORICS 129

a pair (5.1.4) satisfying (5.1.5), the bijection sends it to the map (5.1.6) determined by:

f
ig 7→ kh , where h is chosen such that ug = wh ;
f
jg 7→ kh , where h is chosen such that vg = wh .

Proving that this bijection is well-defined and bijective is straightforward230. 

The multiplication rule (5.1.1) shows that the k-algebra QSym({xi }i∈I ) does not depend much on I, as
long as I is infinite. More precisely, all such k-algebras are mutually isomorphic. We can use this to define
a k-algebra of quasisymmetric functions without any reference to I:

Definition 5.1.5. Let QSym be the k-algebra defined as having k-basis {Mα }α∈Comp and with multiplica-
tion defined k-linearly by (5.1.1). This is called the k-algebra of quasisymmetric functions. We write QSymk
instead of QSym to stress the choice of base ring k.
The k-algebra QSym is graded, and its n-th graded component QSymn has k-basis {Mα }α∈Compn .
For every infinite totally ordered set I, the k-algebra QSym is isomorphic to the k-algebra QSym({xi }i∈I ).
The isomorphism sends Mα 7−→ Mα ({xi }i∈I ).
In particular, we obtain the isomorphism QSym ∼ = QSym (x) for x being the infinite chain (x1 < x2 < x3 < · · · ).
We will identify QSym with QSym (x) along this isomorphism. This allows us to regard quasisymmetric
functions either as power series in a specific set of variables (“alphabet”), or as formal linear combinations
of Mα ’s, whatever is more convenient.

For any infinite alphabet {xi }i∈I and any f
∈ QSym, we denote by f {xi }i∈I the image of f under the
algebra isomorphism QSym → QSym {xi }i∈I defined in Definition 5.1.5.

The comultiplication of QSym will extend the one that we defined for Λ, but we need to take care about
the order of the variables this time. We consider the linear order from (2.3.2) on two sets of variables
(x, y) = (x1 < x2 < . . . < y1 < y2 < . . .), and we embed the k-algebra QSym(x) ⊗ QSym(y) into the
k-algebra R(x, y) by identifying every f ⊗ g ∈ QSym(x) ⊗ QSym(y) with f g ∈ R(x, y) (this embedding is
indeed injective231). It can then be seen that

QSym(x, y) ⊂ QSym(x) ⊗ QSym(y)

(where the right hand side is viewed as k-subalgebra of R(x, y) via said embedding)232, so that one can
∆
define QSym −→ QSym ⊗ QSym as the composite of the maps in the bottom row here:

R(x, y) = R(x, y)
∪ ∪
(5.1.7)
QSym ∼ = QSym(x, y) ,→ QSym(x) ⊗ QSym(y) ∼
= QSym ⊗ QSym
f 7−→ f (x, y) = f (x1 , x2 , . . . , y1 , y2 , . . .)

(Recall that f (x, y) is formally defined as the image of f under the algebra isomorphism QSym → QSym(x, y)
defined in Definition 5.1.5.)

230 The inverse of this bijection sends each map (5.1.6) to the pair (5.1.4) determined by

ug = wh , where h is chosen such that f (ig ) = kh ;

vg = wh , where h is chosen such that f (jg ) = kh .

231This is because it sends the basis elements M (x) ⊗ M (y) of the former k-algebra to the linearly independent power
β γ
series Mβ (x)Mγ (y).
232This is not completely obvious, but can be easily checked by verifying that M (x, y) = P
α (β,γ): Mβ (x) ⊗ Mγ (y) for
β·γ=α
every composition α (see the proof of Proposition 5.1.7 for why this holds).
130 DARIJ GRINBERG AND VICTOR REINER

Example 5.1.6. For example,

∆M(a,b,c) = M(a,b,c) (x1 , x2 , . . . , y1 , y2 , . . .)
= xa1 xb2 xc3 + xa1 xb2 xc4 + · · ·
+ xa1 xb2 · y1c + xa1 xb2 · y2c + · · ·
+ xa1 · y1b y2c + xa1 · y1b y3c + · · ·
+ y1a y2b y3c + y1a y2b y4c + · · ·
= M(a,b,c) (x) + M(a,b) (x)M(c) (y) + M(a) (x)M(b,c) (y) + M(a,b,c) (y)
= M(a,b,c) ⊗ 1 + M(a,b) ⊗ M(c) + M(a) ⊗ M(b,c) + 1 ⊗ M(a,b,c)
Defining the concatenation β · γ of two compositions β = (β1 , . . . , βr ), γ = (γ1 , . . . , γs ) to be the composition
(β1 , . . . , βr , γ1 , . . . , γs ), one has the following description of the coproduct in the {Mα } basis.

Proposition 5.1.7. For a composition α = (α1 , . . . , α` ), one has

`
X X
∆Mα = M(α1 ,...,αk ) ⊗ M(αk+1 ,...,α` ) = Mβ ⊗ Mγ
k=0 (β,γ):
β·γ=α

αk
Proof. This comes from expressing a monomial in ∆Mα = Mα (x, y) uniquely in the form xα i1 · · · xik ·
1

αk+1
yj1 · · · yjα`−k
`
for some k ∈ {0, 1, . . . , n} and some subscripts i1 < · · · < ik and j1 < · · · < j`−k . 

Proposition 5.1.8. The quasisymmetric functions QSym form a connected graded Hopf algebra of finite
type, which is commutative, and contains the symmetric functions Λ as a Hopf subalgebra.

Proof. To prove coassociativity of ∆, we need to be slightly careful. It seems reasonable to argue by

(∆ ⊗ id) ◦ ∆f = f (x, y, z) = (id ⊗∆) ◦ ∆f as in the case of Λ, but this would now require further justification,
as terms like f (x, y) and f (x, y, z) are no longer directly defined as evaluations of f on some sequences (but
rather are defined as images of f under certain homomorphisms). However, it is very easy to see that ∆ is
coassociative by checking (∆ ⊗ id) ◦ ∆ = (id ⊗∆) ◦ ∆ on the {Mα } basis: Proposition 5.1.7 yields
`
X
((∆ ⊗ id) ◦ ∆) Mα = ∆(M(α1 ,...,αk ) ) ⊗ M(αk+1 ,...,α` )
k=0
` k
!
X X
= M(α1 ,...,αi ) ⊗ M(αi+1 ,...,αk ) ⊗ M(αk+1 ,...,α` )
k=0 i=0
` X
X k
= M(α1 ,...,αi ) ⊗ M(αi+1 ,...,αk ) ⊗ M(αk+1 ,...,α` )
k=0 i=0

and the same expression for ((id ⊗∆) ◦ ∆) Mα .

The coproduct ∆ of QSym is an algebra morphism because it is defined as a composite of algebra mor-
phisms in the bottom row of (5.1.7). To prove that the restriction of ∆ to the subring Λ of QSym is the
comultiplication
Pn of Λ, it thus is enough to check that it sends the elementary symmetric function en to
i=0 ei ⊗ en−i for every n ∈ N. This again follows from Proposition 5.1.7, since en = M(1,1,...,1) (with n
times 1).
The counit is as usual for a connected graded coalgebra, and just as in the case of Λ, sends a quasisymmetric
function f (x) to its constant term f (0, 0, . . .). This is an evaluation, and hence an algebra map. Hence QSym
forms a bialgebra, and as it is graded and connected, also a Hopf algebra by Proposition 1.4.14. It is clearly
of finite type and contains Λ as a Hopf subalgebra. 

We will identify the antipode in QSym shortly, but we first deal with another slightly subtle issue. In
addition to the counit evaluation (f ) = f (0, 0, . . .), starting in Section 7.1, we will want to specialize elements
in QSym(x) by making other variable substitutions, in which all but a finite list of variables are set to zero.
We justify this here.
 HOPF ALGEBRAS IN COMBINATORICS 131

Proposition 5.1.9. Fix a totally ordered set I, a commutative k-algebra A, a finite list of variables
xi1 , . . . , xim , say with i1 < . . . < im in I, and an ordered list of elements (a1 , . . . , am ) ∈ Am .
Then there is a well-defined evaluation homomorphism
QSym({xi }i∈I ) −→ A
f 7−→ [f ] xi1 =a1 ,...,xim =am .
xj =0 for j6∈{i1 ,...,im }

Furthermore, this homomorphism depends only upon the list (a1 , . . . , am ), as it coincides with the following:
QSym({xi }i∈I ) ∼
= QSym(x1 , x2 , . . .) −→ A
f (x1 , x2 , . . .) 7−→ f (a1 , . . . , am , 0, 0 . . .).
(This latter statement is stated for the case when I is infinite; otherwise, read “x1 , x2 , . . . , x|I| ” for “x1 , x2 , . . .”,
and interpret (a1 , . . . , am , 0, 0 . . .) as an |I|-tuple.)

Proof. One already can make sense of evaluating xi1 = a1 , . . . , xim = am and xj = 0 for j 6∈ {i1 , . . . , im }
in the ambient ring R({xi }i∈I ) containing QSym({xi }i∈I ), since a power series f of bounded degree will
have finitely many monomials that only involve the variables xi1 , . . . , xim . The last assertion follows from
quasisymmetry of f , and is perhaps checked most easily when f = Mα ({xi }i∈I ) for some α. 

The antipode in QSym has a reasonably simple expression in the {Mα } basis, but requiring a definition.
Definition 5.1.10. For α, β in Compn , say that α refines β or β coarsens α if, informally, one can obtain
β from α by combining some of its adjacent parts. Alternatively, one has a bijection Compn → 2[n−1] where
[n − 1] := {1, 2, . . . , n − 1} which sends α = (α1 , . . . , α` ) having length `(α) = ` to its subset of partial sums
D(α) := {α1 , α1 + α2 , . . . , α1 + · · · + α`−1 } ,
and this sends the refinement ordering to the inclusion ordering on the Boolean algebra 2[n−1] (to be more
precise: a composition α refines a composition β if and only if D(α) ⊃ D(β)). There is also a bijection
sending α to its ribbon diagram: the skew diagram λ/µ having rows of sizes (α1 , . . . , α` ) read from bottom to
top with exactly one column of overlap between adjacent rows. These bijections and the refinement partial
order are illustrated here for n = 4:

{1, 2, 3} (1, 1, 1, 1) 



{1, 2} {1, 3} {2, 3} (1, 1, 2) (1, 2, 1) (2, 1, 1)

  
  
  

{1} {2} {3} (1, 3) (2, 2) (3, 1)

  
  
∅ (4)



Given α = (α1 , . . . , α` ), its reverse composition is rev(α) = (α` , α`−1 , . . . , α2 , α1 ). Note that α 7→ rev(α) is
a poset automorphism for the refinement ordering.

Theorem 5.1.11. For any composition α in Comp,

X
S(Mα ) = (−1)`(α) Mγ
γ∈Comp:
γ coarsens rev(α)

For example,


S(M(a,b,c) ) = − M(c,b,a) + M(b+c,a) + M(c,a+b) + M(a+b+c)
132 DARIJ GRINBERG AND VICTOR REINER

Proof. We give Ehrenborg’s proof233 [54, Prop. 3.4] via induction on ` = `(α). One has easy base cases
when `(α) = 0, where S(M∅ ) = S(1) = 1 = (−1)0 Mrev(∅) , and when `(α) = 1, where M(n) is primitive by
Proposition 5.1.7, so Proposition 1.4.15 shows S(M(n) ) = −M(n) = (−1)1 Mrev((n)) .
For the inductive step, apply the inductive definition of S from the proof of Proposition 1.4.14:
`−1
X
S(M(α1 ,...,α` ) ) = − S(M(α1 ,...,αi ) )M(αi+1 ,...,α` )
i=0
`−1
X X
= (−1)i+1 Mβ M(αi+1 ,...,α` )
i=0 β coarsening
(αi ,αi−1 ,...,α1 )

The idea will be to cancel terms of opposite sign that appear in the expansions of the products Mβ M(αi+1 ,...,α` ) .
Note that each composition β appearing above has first part β1 of the form αi + αi−1 + · · · + αh for some
h ≤ i (unless β = ∅), and hence each term Mγ in the expansion of the product Mβ M(αi+1 ,...,α` ) has γ1 (that
is, the first entry of γ) a sum that can take one of these three forms:
• αi + αi−1 + · · · + αh ,
• αi+1 + (αi + αi−1 + · · · + αh ),
• αi+1 .
Say that the type of γ is i in the first case, and i + 1 in the second two cases234; in other words, the type is
the largest subscript k on a part αk which was combined in the sum γ1 . It is not hard to see that a given
γ for which the type k is strictly smaller than ` arises from exactly two pairs (β, γ), (β 0 , γ), having opposite
signs (−1)k and (−1)k+1 in the above sum235. For example, if α = (α1 , . . . , α8 ), then the composition
γ = (α6 + α5 + α4 , α3 , α7 , α8 + α2 + α1 ) of type 6 can arise from either of

β = (α6 + α5 + α4 , α3 , α2 + α1 ) with i = 6 and sign (−1)7

β 0 = (α5 + α4 , α3 , α2 + α1 ) with i = 5 and sign (−1)6 .

Similarly, γ = (α6 , α5 + α4 , α3 , α7 , α8 + α2 + α1 ) can arise from either of

β = (α6 , α5 + α4 , α3 , α2 + α1 ) with i = 6 and sign (−1)7

β 0 = (α5 + α4 , α3 , α2 + α1 ) with i = 5 and sign (−1)6 .

Thus one can cancel almost all the terms, excepting those with γ of type ` among the terms Mγ in the
expansion of the last (i = ` − 1) summand Mβ M(α` ) . A bit of thought shows that these are the γ coarsening
rev(α), and all have sign (−1)` . 

5.2. The fundamental basis and P -partitions. There is a second important basis for QSym which arose
originally in Stanley’s P -partition theory [180].236

Definition 5.2.1. A labelled poset will here mean a partially ordered set P whose underlying set is some
f
finite subset of the integers. A P -partition is a function P → {1, 2, . . .} with the property that
• i <P j and i <Z j implies f (i) ≤ f (j), and
• i <P j and i >Z j implies f (i) < f (j).
P Q
Denote by A(P ) the set of all P -partitions f , and let FP (x) := f ∈A(P ) xf where xf := i∈P xf (i) . This
FP (x) is an element of k [[x]] := k [[x1 , x2 , . . .]].

233A different proof was given by Malvenuto and Reutenauer [129, Cor. 2.3], and is sketched in Remark 5.3.4 below.
234We imagine that we label the terms obtained by expanding M M by distinct labels, so that each term knows
β (αi+1 ,...,α` )
how exactly it was created (i.e., which i, which β and which map f as in (5.1.2) gave rise to it). Strictly speaking, it is these
triples (i, β, f ) that we should be assigning types to, not terms.
235Strictly speaking, this means that we have an involution on the set of our (i, β, f ) triples having type smaller than `, and
this involution switches the sign of (−1)i Mwt f .
236See [67] for a history of P -partitions; our notations, however, strongly differ from those in [67].
 HOPF ALGEBRAS IN COMBINATORICS 133

Example 5.2.2. Depicted is a labelled poset P , along with the relations among the four values f =
(f (1), f (2), f (3), f (4)) that define its P -partitions f :
2 f (2)
≤

4 1 f (4) f (1)
<
≤
3
f (3)

Remark 5.2.3. Stanley’s treatment of P -partitions in [183, §3.15 and §7.19] uses a language different from
ours. First, Stanley works not with labelled posets P , but with pairs (P, ω) of a poset P and a bijective
labelling ω : P → [n]. Thus, the relation <Z is not given on P a priori, but has to be pulled back from
[n] using ω (and it depends on ω, whence Stanley speaks of “(P, ω)-partitions”). Furthermore, Stanley uses
the notations FP and FP,ω for something different from what we denote by FP , whereas what we call FP is
dubbed KP,ω in [183, §7.19].
The so-called fundamental quasisymmetric functions are an important special case of the FP (x). We shall
first define them directly and then see how they are obtained as P -partition enumerators FP (x) for some
special labelled posets P .
Definition 5.2.4. Let n ∈ N and α ∈ Compn . We define the fundamental quasisymmetric function Lα =
Lα (x) ∈ QSym by
X
(5.2.1) Lα := Mβ .
β∈Compn :
β refines α

Example 5.2.5. The extreme cases for α in Compn give quasisymmetric functions Lα which are symmetric:
L(1n ) = M(1n ) = en ,
X
L(n) = Mα = hn
α∈Compn

Before studying the Lα in earnest, we recall a basic fact about finite sets, which is sometimes known as
the “principle of inclusion and exclusion” (although it is more general than the formula for the size of a
union of sets that commonly goes by this name):

Lemma 5.2.6. Let G be a finite set. Let V be a k-module. For each subset A of G, we let fA and gA be
two elements of V .
(a) If X
every A ⊂ G satisfies gA = fB ,
B⊂A
then
|A\B|
X
every A ⊂ G satisfies fA = (−1) gB .
B⊂A
(b) If X
every A ⊂ G satisfies gA = fB ,
B⊂G; B⊃A
then
|B\A|
X
every A ⊂ G satisfies fA = (−1) gB .
B⊂G; B⊃A

Proof. This can be proven by elementary arguments (easy exercise). Alternatively, Lemma 5.2.6 can be
viewed as a particular case of the Möbius inversion principle (see, e.g., [183, Propositions 3.7.1 and 3.7.2])
applied to the Boolean lattice 2G (whose Möbius function is very simple: see [183, Example 3.8.3]). (This is
spelled out in [121, Example 4.52], for example.) 
134 DARIJ GRINBERG AND VICTOR REINER

Lemma 5.2.6 can be translated into the language of compositions:

Lemma 5.2.7. Let n ∈ N. Let V be a k-module. For each α ∈ Compn , we let fα and gα be two elements
of V .
(a) If X
every α ∈ Compn satisfies gα = fβ ,
β coarsens α
then X `(α)−`(β)
every α ∈ Compn satisfies fα = (−1) gβ .
β coarsens α
(b) If X
every α ∈ Compn satisfies gα = fβ ,
β refines α
then X `(β)−`(α)
every α ∈ Compn satisfies fα = (−1) gβ .
β refines α

Proof of Lemma 5.2.7. Set [n − 1] = {1, 2, . . . , n − 1}. Recall (from Definition 5.1.10) that there is a bijection
D : Compn → 2[n−1] that sends each α ∈ Compn to D (α) ⊂ [n − 1]. This bijection D has the properties
that:
• a composition β refines a composition α if and only if D (β) ⊃ D (α);
• a composition β coarsens a composition α if and only if D (β) ⊂ D (α);
• any composition α ∈ Compn satisfies |D (α)| = ` (α) − 1 (unless n = 0), and thus
• any compositions α and β in Compn satisfy |D (α)| − |D (β)| = ` (α) − ` (β).
This creates a dictionary between compositions in Compn and subsets of [n − 1]. Now, apply Lemma 5.2.6
to G = [n − 1], fA = fD−1 (A) and gA = gD−1 (A) , and translate using the dictionary. 
Now, we can see the following about the fundamental quasisymmetric functions:
Proposition 5.2.8. The family {Lα }α∈Comp is a k-basis for QSym, and each n ∈ N and α ∈ Compn satisfy
X
(5.2.2) Mα = (−1)`(β)−`(α) Lβ .
β∈Compn :
β refines α

Proof of Proposition 5.2.8. Fix n ∈ N. Recall the equality (5.2.1). Thus, Lemma 5.2.7(b) (applied to
V = QSym, fα = Mα and gα = Lα ) yields (5.2.2).
Recall that the family (Mα )α∈Compn is a basis of the k-module QSymn . The equality (5.2.1) shows that
the family (Lα )α∈Compn expands invertibly triangularly237 with respect to the family (Mα )α∈Compn (where
Compn is equipped with the refinement order).238 Thus, Corollary 11.1.19(e) (applied to QSymn , Compn ,
(Mα )α∈Compn and (Lα )α∈Compn instead of M , S, (es )s∈S and (fs )s∈S ) shows that the family (Lα )α∈Compn is
a basis of the k-module QSymn . Combining this fact for all n ∈ N, we conclude that the family (Lα )α∈Comp
is a basis of the k-module QSym. This completes the proof of Proposition 5.2.8. 
Proposition 5.2.9. Let n ∈ N. Let α be a composition of n. Let I be an infinite totally ordered set. Then,

X
Lα {xi }i∈I = x i1 x i2 · · · x in ,
i1 ≤i2 ≤···≤in in I;
ij <ij+1 if j∈D(α)



where Lα {xi }i∈I is defined as the image of Lα under the isomorphism QSym → QSym {xi }i∈I obtained
in Definition 5.1.5. In particular, for the standard (totally ordered) variable set x = (x1 < x2 < . . .), we
obtain
X
(5.2.3) Lα = Lα (x) = xi1 xi2 · · · xin .
(1≤)i1 ≤i2 ≤···≤in ;
ij <ij+1 if j∈D(α)

237See Section 11.1 for a definition of this concept.

238In fact, it expands unitriangularly with respect to the latter family.
 HOPF ALGEBRAS IN COMBINATORICS 135

Proof. Every composition β = (β1 , . . . , β` ) of n satisfies

X X
(5.2.4) Mβ ({xi }i∈I ) = xβk11 · · · xβk`` = xi1 xi2 · · · xin .
k1 <···<k` in I i1 ≤i2 ≤···≤in in I;
ij <ij+1 if and only if j∈D(β)


Applying the ring homomorphism QSym → QSym {xi }i∈I to (5.2.1), we obtain

X
(5.2.4) X X
Lα {xi }i∈I = Mβ {xi }i∈I = xi1 xi2 · · · xin
β∈Compn : β∈Compn : i1 ≤i2 ≤···≤in in I;
β refines α β refines α ij <ij+1 if and only if j∈D(β)
X X
= xi1 xi2 · · · xin
β∈Compn : i1 ≤i2 ≤···≤in in I;
D(α)⊂D(β) ij <ij+1 if and only if j∈D(β)
X X X
= xi1 xi2 · · · xin = x i1 x i2 · · · x in .
Z⊂[n−1]: i1 ≤i2 ≤···≤in in I; i1 ≤i2 ≤···≤in in I;
D(α)⊂Z ij <ij+1 if and only if j∈Z ij <ij+1 if j∈D(α)

Proposition 5.2.10. When the labelled poset P is a total or linear order w = (w1 < . . . < wn ), the
generating function Fw (x) depends only upon the descent set
Des(w) := {i : wi >Z wi+1 } ⊂ {1, 2, . . . , n − 1}
and its associated composition α in Compn having partial sums D(α) = Des(w): one has that Fw (x) equals
the fundamental quasisymmetric function Lα .

E.g., total order w = 35142 has Des(w) = {2, 4} and composition α = (2, 2, 1), so
X
F35142 (x) = xf (3) xf (5) xf (1) xf (4) xf (2)
f (3)≤f (5)<f (1)≤f (4)<f (2)
X
= x i1 x i2 x i3 x i4 x i5
i1 ≤i2 <i3 ≤i4 <i5
= L(2,2,1) = M(2,2,1) + M(2,1,1,1) + M(1,1,2,1) + M(1,1,1,1,1) .

Proof of Proposition 5.2.10. Write Fw (x) as a sum of monomials xf (w1 ) · · · xf (wn ) over all P -partitions f .
These P -partitions are exactly the sequences f (w1 ) ≤ · · · ≤ f (wn ) having strict inequalities f (wi ) < f (wi+1 )
whenever i is in Des(w). Hence, comparison with (5.2.3) reveals that this sum equals Lα . 

The next proposition ([183, Cor. 7.19.5], [123, Cor. 3.3.24]) is an algebraic shadow of Stanley’s main
lemma [183, Thm. 7.19.4] in P -partition theory. It expands any FP (x) in the {Lα } basis, as a sum over
239
the set L(P ) of all linear extensions w of P . E.g., the poset P from Example 5.2.2 has L(P ) =
{3124, 3142, 3412}.

Theorem 5.2.11. For any labelled poset P ,

X
FP (x) = Fw (x).
w∈L(P )

239Let us explain what we mean by linear extensions and how we represent them.
If P is a finite poset, then a linear extension of P denotes a total order w on the set P having the property that i <P j
(for two elements i and j of P) implies i <w j. (In other words, it is a linear order on the ground set P which extends P as
a poset; therefore the name.) We identify such a total order w with the list (p1 , p2 , . . . , pn ) containing all elements of P in
w-increasing order (that is, p1 <w p2 <w · · · <w pn ).
(Stanley, in [183, §3.5], defines linear extensions in a slightly different way: For him, a linear extension of a finite poset P
is an order-preserving bijection from P to the subposet {1, 2, . . . , |P|} of Z. But this is equivalent to our definition, since a
bijection like this can be used to transport the order relation of {1, 2, . . . , |P|} back to P, thus resulting in a total order on P
which is a linear extension of P in our sense.)
136 DARIJ GRINBERG AND VICTOR REINER

Proof. We give Gessel’s proof [66, Thm. 1], via induction on the number of pairs i, j which are incomparable
in P . When this quantity is 0, then P is itself a linear order w, so that L(P ) = {w} and there is nothing to
prove.
In the inductive step, let i, j be incomparable elements. Consider the two posets Pi<j and Pj<i which are
obtained from P by adding in an order relation between i and j, and then taking the transitive closure; it is
not hard to see that these transitive closures cannot contain a cycle, so that these really do define two posets.
The result then follows by induction applied to Pi<j , Pj<i , once one notices that L(P ) = L(Pi<j ) t L(Pj<i )
since every linear extension w of P either has i before j or vice-versa, and A(P ) = A(Pi<j ) t A(Pj<i ) since,
assuming that i <Z j without loss of generality, every f in A(P ) either satisfies f (i) ≤ f (j) or f (i) > f (j). 

Example 5.2.12. To illustrate the induction in the above proof, consider the poset P from Example 5.2.2,
having L(P ) = {3124, 3142, 3412}. Then choosing as incomparable pair (i, j) = (1, 4), one has

4 2 f (4) f (2)
≤
≤
Pi<j = 1 f (1) L(Pi<j ) = {3124, 3142}

<

3 f (3)

2 f (2)
≤

Pj<i = 1 f (1) L(Pj<i ) = {3412}

<

4 f (4)
≤
3
f (3)

Exercise 5.2.13. Give an alternative proof for Theorem 5.2.11.

[Hint: For every f : P → {1, 2, 3, . . .}, we can define a binary relation ≺f on the set P by letting i ≺f j
hold if and only if

(f (i) < f (j) or (f (i) = f (j) and i <Z j)) .

Show that this binary relation ≺f is (the smaller relation of) a total order. When f is a P -partition, then
endowing the set P with this total order yields a linear extension of P . Use this to show that the set A (P )
is the union of its disjoint subsets A (w) with w ∈ L (P ).]

Various other properties of the quasisymmetric functions FP (x) are studied, e.g., in [134].
We next wish to describe the structure maps for the Hopf algebra QSym in the basis {Lα } of fundamental
quasisymmetric functions. For this purpose, two more definitions are useful.

Definition 5.2.14. Given two nonempty compositions α = (α1 , . . . , α` ), β = (β1 , . . . , βm ), their near-
concatenation is

α β := (α1 , . . . , α`−1 , α` + β1 , β2 , . . . , βm )
 HOPF ALGEBRAS IN COMBINATORICS 137

For example, the figure below depicts for α = (1, 3, 3) (black squares) and β = (4, 2) (white squares) the
concatenation and near-concatenation as ribbons:
 
   
α·β =   
  

 
      
α β=
  

Lastly, given α in Compn , let ω(α) be the unique composition in Compn whose partial sums D(ω(α))
form the complementary set within [n − 1] to the partial sums D(rev(α)); alternatively, one can check this
means that the ribbon for ω(α) is obtained from that of α by conjugation or transposing, that is, if α = λ/µ
then ω(α) = λt /µt . E.g. if α = (4, 2, 2) so that n = 8, then rev(α) = (2, 2, 4) has D(rev(α)) = {2, 4} ⊂ [7],
complementary to the set {1, 3, 5, 6, 7} which are the partial sums for ω(α) = (1, 2, 2, 1, 1, 1), and the ribbon
diagrams of α, ω(α) are


 

α=   and ω(α) =
 
   
 


Proposition 5.2.15. The structure maps for the Hopf algebra QSym in the basis {Lα } of fundamental
quasisymmetric functions are as follows:
X
(5.2.5) ∆Lα = Lβ ⊗ Lγ
(β,γ):
β·γ=α or β γ=α
X
(5.2.6) Lα Lβ = Lγ(w)
w∈wα
w β

(5.2.7) S(Lα ) = (−1)|α| Lω(α) .

Here we are making use of the following notations in (5.2.6) (recall also Definition 1.6.2):
• A labelled linear order will mean a labelled poset P whose order <P is a total order. We will identify
any labelled linear order P with the word (over the alphabet Z) obtained by writing down the
elements of P in increasing order (with respect to the total order <P ). This way, every word (over
the alphabet Z) which has no two equal letters becomes identified with a labelled linear order.
• wα is any labelled linear order with underlying set {1, 2, . . . , |α|} such that Des (wα ) = D (α).
• wβ is any labelled linear order with underlying set {|α| + 1, |α| + 2, . . . , |α| + |β|} such that Des (wβ ) =
D (β).
• γ(w) is the unique composition of |α| + |β| with D(γ(w)) = Des(w).
(The right hand side of (5.2.6) is to be read as a sum over all w, for a fixed choice of wα and wβ .)

At first glance the formula (5.2.5) for ∆Lα might seem more complicated than the formula of Proposition 5.1.7
for ∆Mα . However, it is equally simple when viewed in terms of ribbon diagrams: it cuts the ribbon diagram
α into two smaller ribbons β and γ, in all |α| + 1 possible ways, via horizontal cuts (β · γ = α) or vertical
cuts (β γ = α). For example,
∆L(3,2)
= 1 ⊗ L(3,2) +L(1) ⊗ L(2,2) +L(2) ⊗ L(1,2) +L(3) ⊗ L(2) +L(3,1) ⊗ L(1) +L(3,2) ⊗ 1
         |  
    |    |         
138 DARIJ GRINBERG AND VICTOR REINER

Example 5.2.16. To multiply L(1,1) L(2) , one could pick wα = 21 and wβ = 34, and then
L(1,1) L(2) = w∈21
34 Lγ(w) = Lγ(2134) +Lγ(2314) +Lγ(3214) +Lγ(2341) +Lγ(3241)
P
+Lγ(3421)
= L(1,3) +L(2,2) +L(1,1,2) +L(3,1) +L(1,2,1) +L(2,1,1) .
Before we prove Proposition 5.2.15, we state a simple lemma:

Lemma 5.2.17. Let Q and R be two labelled posets whose underlying sets are disjoint. Let Q t R be the
disjoint union of these posets Q and R; this is again a labelled poset. Then,
FQ (x) FR (x) = FQtR (x) .

Proof of Lemma 5.2.17. We identify the underlying set of Q t R with Q ∪ R (since the sets Q and R are
already disjoint). If f : Q t R → {1, 2, 3, . . .} is a Q t R-partition, then its restrictions f |Q and f |R are a
Q-partition and an R-partition, respectively. Conversely, any pair of a Q-partition and an R-partition can
be combined to form a Q t R-partition. Thus, there is a bijective correspondence between the addends in
the expanded sum FQ (x) FR (x) and the addends in FQtR (x). 

Proof of Proposition 5.2.15. To prove formula (5.2.5) for α in Compn , note that
n
X X
(5.2.8) ∆Lα = Lα (x, y) = xi1 · · · xik · yik+1 · · · yin
k=0 1≤i1 ≤···≤ik ,
1≤ik+1 ≤···≤in :
ir <ir+1 for r∈D(α)\{k}

by Proposition 5.2.9 (where we identify QSym ⊗ QSym with a k-subalgebra of R (x, y) by means of the
∼
=
embedding QSym ⊗ QSym → QSym (x) ⊗ QSym (y) ,→ R (x, y) as in the definition of the comultiplication
on QSym). One then realizes that the inner sums corresponding to values of k that lie (resp. do not lie) in
D(α) ∪ {0, n} correspond to the terms Lβ (x)Lγ (y) for pairs (β, γ) in which β · γ = α (resp. β γ = α).
For formula (5.2.6), let P be the labelled poset which is the disjoint union of linear orders wα , wβ . Then
X X
Lα Lβ = Fwα (x)Fwβ (x) = FP (x) = Fw (x) = Lγ(w)
w∈L(P ) w∈wα
w β

where the first equality used Proposition 5.2.10, the second equality comes from Lemma 5.2.17, the third
equality from Theorem 5.2.11, and the fourth from the equality L(P ) = wα wβ .

To prove formula (5.2.7), compute using Theorem 5.1.11 that
X X X X
S(Lα ) = S(Mβ ) = (−1)`(β) Mγ = Mγ (−1)`(β)
β refining α (β,γ): γ β
β refines α,
γ coarsens rev(β)

in which the last inner sum is over β for which

D(β) ⊃ D(α) ∪ D(rev(γ)).
The alternating signs make such inner sums vanish unless they have only the single term where D(β) = [n−1]
(that is, β = (1n )). This happens exactly when D(rev(γ)) ∪ D(α) = [n − 1] or equivalently, when D(rev(γ))
contains the complement of D(α), that is, when D(γ) contains the complement of D(rev(α)), that is, when
γ refines ω(α). Thus
X
S(Lα ) = Mγ · (−1)n = (−1)|α| Lω(α) .
γ∈Compn :
γ refines ω(α)

The antipode formula (5.2.7) for Lα leads to a general interpretation for the antipode of QSym acting on
P -partition enumerators FP (x).
Definition 5.2.18. Given a labelled poset P on {1, 2, . . . , n}, let the opposite or dual labelled poset P opp
have i <P opp j if and only if j <P i.
 HOPF ALGEBRAS IN COMBINATORICS 139

For example,
P = 2 P opp = 3

4 1 4 1

3 2
The following observation is straightforward.

Proposition 5.2.19. When P is a linear order corresponding to some permutation w = (w1 , . . . , wn ) in Sn ,

then wopp = ww0 where w0 ∈ Sn is the permutation that swaps i ↔ n + 1 − i (this is the so-called longest
permutation, thus named due to it having the highest “Coxeter length” among all permutations in Sn ).
Furthermore, in this situation one has Fw (x) = Lα , that is, Des(w) = D(α) if and only if Des(wopp ) =
D(ω(α)), that is Fwopp (x) = Lω(α) . Thus,

S(Fw (x)) = (−1)n Fwopp (x).

For example, given the compositions considered earlier



 

α = (4, 2, 2) =   ω(α) = (1, 2, 2, 1, 1, 1) =
 
   
 


if one picks w = 1235 · 47 · 68 (with descent positions marked by dots) having Des(w) = {4, 6} = D(α), then
wopp = ww0 = 8 · 67 · 45 · 3 · 2 · 1 has Des(wopp ) = {1, 3, 5, 6, 7} = D(ω(α)).

Corollary 5.2.20. For any labelled poset P on {1, 2, . . . , n}, one has

S (FP (x)) = (−1)n FP opp (x).

Proof. Since S is linear, one can apply Theorem 5.2.11 and Proposition 5.2.19
X X
S (FP (x)) = S(Fw (x)) = (−1)n Fwopp (x) = (−1)n FP opp (x).
w∈L(P ) w∈L(P )

as L(P opp ) = {wopp : w ∈ L(P )}. 

Remark 5.2.21. Malvenuto and Reutenauer, in [130, Theorem 3.1], prove an even more general antipode
formula, which encompasses our Corollary 5.2.20, Proposition 5.2.19, Theorem 5.1.11 and (5.2.7). See [71,
Theorem 4.2] for a restatement and a self-contained proof of this theorem (and [71, Theorem 4.7] for an even
further generalization).

We remark on a special case of Corollary 5.2.20 to which we alluded earlier, related to skew Schur functions.

Corollary 5.2.22. In Λ, the action of ω and the antipode S on skew Schur functions sλ/µ are as follows:

(5.2.9) ω(sλ/µ ) = sλt /µt

(5.2.10) S(sλ/µ ) = (−1)|λ/µ| sλt /µt .

Proof. Given a skew shape λ/µ, one can always create a labelled poset P which is its skew Ferrers poset,
together with one of many column-strict labellings, in such a way that FP (x) = sλ/µ (x). An example is
140 DARIJ GRINBERG AND VICTOR REINER

shown here for λ/µ = (4, 4, 2)/(1, 1, 0):

  
λ/µ =    P = 5 f (5)
<
 
8 4 2 ≤
f (8) f (4) f (2)
< <
7 3 1 ≤ ≤ ≤
f (7) f (3) f (1)
<
6 ≤
f (6)
The general definition is as follows: Let P be the set of all boxes of the skew diagram λ/µ. Label these
boxes by the numbers 1, 2, . . . , n (where n = |λ/µ|) row by row from bottom to top (reading every row from
left to right), and then define an order relation <P on P by requiring that every box be smaller (in P ) than
its right neighbor and smaller (in P ) than its lower neighbor. It is not hard to see that in this situation,
P cont(T )
FP (x) = T x
opp as T ranges over all reverse semistandard tableaux or column-strict plane partitions
t t
of λ /µ :

t t  
λ /µ = P opp = 6 f (6)
  <
  7 3 1 ≤
f (7) f (3) f (1)
< < <
8 4 2 ≤ ≤
f (8) f (4) f (2)
<
5 ≤
f (5)
But this means that FP opp (x) = sλt /µt (x), since the fact that skew Schur functions lie in Λ implies that they
can be defined either as generating functions for column-strict tableaux or reverse semistandard tableaux;
see Remark 2.2.5 above, or [183, Prop. 7.10.4].
Thus we have
FP (x) = sλ/µ (x)
FP opp (x) = sλt /µt (x).
Proposition 1.4.24(c) tell us that the antipode for QSym must specialize to the antipode for Λ (see also
Remark 5.3.11 below), so (5.2.10) is a special case of Corollary 5.2.20. Then (5.2.9) follows from the relation
(2.4.7) that S(f ) = (−1)n ω(f ) for f in Λn . 
Remark 5.2.23. Before leaving P -partitions temporarily, we mention two open questions about them.
The first is a conjecture of Stanley from his thesis [180]. As mentioned in the proof of Corollary 5.2.22,
each skew Schur function sλ/µ (x) is a special instance of P -partition enumerator FP (x).

Conjecture 5.2.24. A labelled poset P has FP (x) symmetric, and not just quasisymmetric, if and only if
P is a column-strict labelling of some skew Ferrers poset λ/µ.

A somewhat weaker result in this direction was proven by Malvenuto in her thesis [128, Thm. 6.4], showing
that if a labelled poset P has the stronger property that its set of linear extensions L(P ) is a union of plactic
or Knuth equivalence classes, then P must be a column-strict labelling of a skew Ferrers poset.
The next question is due to P. McNamara, and is suggested by the obvious factorizations of P -partition
enumerators FP1 tP2 (x) = FP1 (x)FP2 (x) (Lemma 5.2.17).
Question 5.2.25. If k is a field, does a connected labelled poset P always have FP (x) irreducible within the
ring QSym?
The phrasing of this question requires further comment. It is assumed here that x = (x1 , x2 , . . .) is infinite;
for example when P is a 2-element chain labelled “against the grain” (i.e., the bigger element of the chain
has the smaller label), then FP (x) = e2 (x) is irreducible, but its specialization to two variables x = (x1 , x2 )
is e2 (x1 , x2 ) = x1 x2 , which is reducible. If one wishes to work in finitely many variables x = (x1 , . . . , xm )
one can perhaps assume that m is at least |P | + 1.
 HOPF ALGEBRAS IN COMBINATORICS 141

When working in QSym = QSym(x) in infinitely many variables, it is perhaps not so clear where fac-
torizations occur. For example, if f lies in QSym and factors f = g · h with g, h in R(x), does this imply
that g, h also lie in QSym? The answer is “Yes” (for k = Z), but this is not obvious, and was proven by P.
Pylyavskyy in [155, Chap. 11].
One also might wonder whether QSymZ is a unique factorization domain, but this follows from the result
of M. Hazewinkel ([74] and [78, Thm. 6.7.5], and Theorem 6.4.3 further below) who proved a conjecture of
Ditters that QSymZ is a polynomial algebra; earlier Malvenuto and Reutenauer [129, Cor. 2.2] had shown
that QSymQ is a polynomial algebra. In fact, one can find polynomial generators {Pα } for QSymQ as a subset
of the dual basis to the Q-basis {ξα } for NSymQ which comes from taking products ξα := ξα1 · · · ξα` of the
elements {ξn } defined in Remark 5.3.4 below. Specifically, one takes those Pα for which the composition α
is a Lyndon composition; see the First proof of Proposition 6.4.4 for a mild variation on this construction.
Hazewinkel’s proof [78, Thm. 6.7.5] of the polynomiality of QSymZ also shows that QSym is a polynomial
ring over Λ (see Corollary 6.5.33); in particular, this yields that QSym is a free Λ-module.240
An affirmative answer to Question 5.2.25 is known at least in the special case where P is a connected
column-strict labelling of a skew Ferrers diagram, that is, when FP (x) = sλ/µ (x) for some connected skew
diagram λ/µ; see [13].
5.3. The Hopf algebra NSym dual to QSym. We introduce here the (graded) dual Hopf algebra to QSym.
This is well-defined, as QSym is connected graded of finite type.
(·,·)
Definition 5.3.1. Let NSym := QSymo , with dual pairing NSym ⊗ QSym −→ k. Let {Hα } be the k-basis
of NSym dual to the k-basis {Mα } of QSym, so that
(Hα , Mβ ) = δα,β .
When the base ring k is not clear from the context, we write NSymk in lieu of NSym.
The Hopf algebra NSym is known as the Hopf algebra of noncommutative symmetric functions. Its study
goes back to [64].
Theorem 5.3.2. Letting Hn := H(n) for n = 0, 1, 2, . . ., with H0 = 1, one has that
(5.3.1) NSym = ∼ khH1 , H2 , . . .i,
the free associative (but not commutative) algebra on generators {H1 , H2 , . . .} with coproduct determined
by241
X
(5.3.2) ∆Hn = Hi ⊗ Hj .
i+j=n
P
Proof. Since Proposition 5.1.7 asserts that ∆Mα = (β,γ):β·γ=α Mβ ⊗Mγ , and since {Hα } are dual to {Mα },
one concludes that for any compositions β, γ, one has
Hβ Hγ = Hβ·γ
Iterating this gives
(5.3.3) Hα = H(α1 ,...,α` ) = Hα1 · · · Hα` .
Since the Hα are a k-basis for NSym, this shows NSym ∼ = khH1 , H2 , . . .i.
Note that Hn = H(n) is dual to M(n) , so to understand ∆Hn , one should understand how M(n) can appear
as a term in the product Mα Mβ . By (5.1.1) this occurs only if α = (i), β = (j) where i + j = n, where
M(i) M(j) = M(i+j) + M(i,j) + M(j,i)
240The latter statement has an analogue in finitely many indeterminates, proven by Lauve and Mason in [109, Corollary



13]: The quasisymmetric functions QSym {xi }i∈I are free as a Λ {xi }i∈I -module for any totally ordered set I, infinite or


not. In the case of finite I, this cannot be derived by Hazewinkel’s arguments, as the ring QSym {xi }i∈I is not in general a
polynomial ring (e.g., when k = Q and I = {1, 2}, this ring is not even a UFD, as witnessed by x21 x2 · x1 x22 = (x1 x2 )3 ).



241The abbreviated summation indexing P
i+j=n ti,j used here is intended to mean
X
ti,j .
(i,j):
0≤i,j≤n,
i+j=n
142 DARIJ GRINBERG AND VICTOR REINER

(where the M(i,j) and M(j,i) addends have to be disregarded if one of i and j is 0). By duality, this implies
the formula (5.3.2). 

Corollary 5.3.3. The algebra homomorphism defined by

π
NSym −→ Λ
Hn 7−→ hn
i
is a Hopf algebra surjection, and adjoint to the inclusion Λ ,→ QSym (with respect to the dual pairing
(·,·)
NSym ⊗ QSym −→ k).

Proof. As an algebra map π may be identified with the surjection T (V ) → Sym(V ) from the tensor algebra
on a graded free k-module V with basis {H1 , H2 , . . .} to the symmetric algebra on V , since
NSym ∼
= khH1 , H2 , . . .i
Λ∼= k[h1 , h2 , . . .]

As (5.3.2) and Proposition 2.3.6(iii) assert that

X
∆Hn = Hi ⊗ Hj
i+j=n
X
∆hn = hi ⊗ hj
i+j=n

this map π is also a bialgebra morphism, and hence a Hopf morphism by Proposition 1.4.24(c).
To check π is adjoint to i, let λ(α) denote the partition which is the weakly decreasing rearrangement of
the composition α, and note that the bases {Hα } of NSym and {mλ } of Λ satisfy
 
 
1 if λ(α) = λ X
(π(Hα ), mλ ) = (hλ(α) , mλ ) = = Hα , Mβ  = (Hα , i(mλ )).
0 otherwise
β:λ(β)=λ

Remark 5.3.4. For those who prefer generating functions to sign-reversing involutions, we sketch here Mal-
venuto and Reutenauer’s elegant proof [129, Cor. 2.3] of the antipode formula (Theorem 5.1.11). One needs
to know that when Q is a subring of k, and A is a k-algebra (possibly noncommutative), in the ring of power
series A[[t]] where t commutes with all of A, one still has familiar facts, such as

a(t) = log b(t) if and only if b(t) = exp a(t)

and whenever a(t), b(t) commute in A[[t]], one has

(5.3.4) exp (a(t) + b(t)) = exp a(t) exp b(t)

(5.3.5) log (a(t)b(t)) = log a(t) + log b(t)

Start by assuming WLOG that k = Z (as NSymk = NSymZ ⊗Z k in the general case). Now, define in
NSymQ = NSym ⊗Z Q the elements {ξ1 , ξ2 , . . .} via generating functions in NSymQ [[t]]:
X
H(t)
e := Hn tn ,
n≥0
(5.3.6) X
ξ(t) := ξn tn = log H(t)
e
n≥1

One first checks that this makes each ξn primitive, via a computation in the ring (NSymQ ⊗ NSymQ )[[t]] (into
which we “embed” the ring (NSymQ [[t]]) ⊗Q[[t]] (NSymQ [[t]]) via the canonical ring homomorphism from the
 HOPF ALGEBRAS IN COMBINATORICS 143

242
latter into the former ):
   
X X X X
∆ξ(t) = ∆ log Hn tn  = log ∆(Hn )tn = log  Hi ⊗ Hj  tn
n≥0 n≥0 n≥0 i+j=n
      
X X X X
= log  Hi ti  ⊗  Hj tj  = log  Hi ti ⊗ 1 1 ⊗ Hj tj 
i≥0 j≥0 i≥0 j≥0
(5.3.5)
= e ⊗ 1 + 1 ⊗ log H(t)
log H(t) e = ξ(t) ⊗ 1 + 1 ⊗ ξ(t).
Comparing coefficients in this equality yields ∆(ξn ) = ξn ⊗ 1 + 1 ⊗ ξn . Thus S(ξn ) = −ξn , by Proposi-
tion 1.4.15. This allows one to determine S(Hn ) and S(Hα ), after one first inverts the relation (5.3.6) to get
that H(t)
e = exp ξ(t), and hence
(5.3.4)
S(H(t))
e = S(exp ξ(t)) = exp S(ξ(t)) = exp (−ξ(t)) = (exp ξ(t))−1
e −1 = 1 + H1 t + H2 t2 + · · · −1 .


= H(t)
Upon expanding the right side, and comparing coefficients of tn , this gives
X
S(Hn ) = (−1)`(β) Hβ
β∈Compn

and hence X
S(Hα ) = S(Hα` ) · · · S(Hα2 )S(Hα1 ) = (−1)`(γ) Hγ .
γ:
γ refines rev(α)
As SNSym , SQSym are adjoint, and {Hα }, {Mα } are dual bases, this is equivalent to Theorem 5.1.11:
X
S(Mα ) = (−1)`(α) Mγ
γ:
γ coarsens rev(α)

(because if µ and ν are two compositions, then µ coarsens ν if and only if rev(µ) coarsens rev(ν)). Thus,
Theorem 5.1.11 is proven once again.
Let us say a bit more about the elements ξn defined in (5.3.6) above. The elements nξn are noncommu-
tative analogues of the power sum symmetric functions pn (and, indeed, are lifts of the latter to NSym, as
Exercise 5.3.5 below shows). They are called the noncommutative power sums of the second kind in [64]243,
and their products form a basis of NSym. They are furthermore useful in studying the so-called Eulerian
idempotent of a cocommutative Hopf algebra, as shown in Exercise 5.3.6 below.
Exercise 5.3.5. Assume that Q is a subring of k. Define a sequence of elements ξ1 , ξ2 , ξ3 , . . . of NSym =
NSymk by (5.3.6).
(a) For every n ≥ 1, show that ξn is a primitive homogeneous element of NSym of degree n.
(b) For every n ≥ 1, show that π (nξn ) is the n-th power sum symmetric function pn ∈ Λ.
(c) For every n ≥ 1, show that
X `(α)−1 1
(5.3.7) ξn = (−1) Hα .
` (α)
α∈Compn

(d) For every composition α, define an element ξα of NSym by ξα = ξα1 ξα2 · · · ξα` , where α is written
in the form α = (α1 , α2 , . . . , α` ) with ` = ` (α). Show that
X 1
(5.3.8) Hn = ξα
` (α)!
α∈Compn

242This ring homomorphism might fail to be injective, whence the “embed” stands in quotation marks. This does not need
to worry us, since we will not draw any conclusions in (NSymQ [[t]]) ⊗Q[[t]] (NSymQ [[t]]) from our computation.
We are also somewhat cavalier with the notation ∆: we use it both for the comultiplication ∆ : NSymQ → NSymQ ⊗ NSymQ


of the Hopf algebra NSymQ and for the continuous k-algebra homomorphism NSymQ [[t]] → NSymQ ⊗ NSymQ [[t]] it induces.
243See Exercise 5.3.12 for the ones of the first kind.
144 DARIJ GRINBERG AND VICTOR REINER

for every n ∈ N.
Use this to prove that (ξα )α∈Compn is a k-basis of NSymn for every n ∈ N.
ExerciseL5.3.6. Assume that Q is a subring of k. Let A be a cocommutative connected graded k-bialgebra.
Let A = n≥0 An be the decomposition of A into homogeneous components. If f is any k-linear map A → A
n−1 1 ?n
annihilating A0 , then f is locally ?-nilpotent244, and so the sum log? (f + u) := n≥1 (−1)
P
nf is a
245 ?
well-defined endomorphism of A . Let e denote the endomorphism log (idA ) of A (obtained by setting
f = idA −u : A → A). Show that e is a projection from A to the k-submodule p of all primitive elements
of A (and thus, in particular, is idempotent).
Hint: For every n ≥ 0, let πn : A → A be the projection onto the n-th homogeneous component An .
Since NSym is the free k-algebra with generators H1 , H2 , H3 , . . ., we can define a k-algebra homomorphism
W : NSym → (End A, ?) by sending Hn to πn . Show that:
(a) The map e : A → A is graded. For every n ≥ 0, we will denote the map πn ◦ e = e ◦ πn : A → A by
en .
(b) We have W (ξn ) = en for all n ≥ 1, where ξn is definedPas in Exercise 5.3.5.
(c) If w is an element of NSym, and if we write ∆ (w) = (w) w1 ⊗ w2 using Sweedler’s notation, then
P 
∆ ◦ (W (w)) = (w) W (w1 ) ⊗ W (w2 ) ◦ ∆.
(d) We have en (A) ⊂ p for every n ≥ 0.
(e) We have e (A) ⊂ p.
(f) The map e fixes any element of p.
Remark 5.3.7. The endomorphism e of Exercise 5.3.6 is known as the Eulerian idempotent of A, and can be
contrasted with the Dynkin idempotent of Remark 1.5.12. It has been studied in [146], [149], [30] and [51],
and relates to the Hochschild cohomology of commutative algebras [117, §4.5.2].
Exercise 5.3.8. Assume that Q is a subring of k. Let A, An and e be as in Exercise 5.3.6.
(a) Show that e?n ◦ e?m = n!δn,m e?n for all n ∈ N and m ∈ N.
(b) Show that e?n ◦ id?m ?m
A = idA ◦e
?n
= mn e?n for all n ∈ N and m ∈ N.
We next explore the basis for NSym dual to the {Lα } in QSym.
Definition 5.3.9. Define the noncommutative ribbon functions {Rα } to be the k-basis of NSym dual to the
fundamental basis {Lα } of QSym, so that (Rα , Lβ ) = δα,β .
Theorem 5.3.10. (a) One has that
X
(5.3.9) Hα = Rβ
β coarsens α
X
(5.3.10) Rα = (−1)`(β)−`(α) Hβ
β coarsens α
π
(b) The surjection NSym −→ Λ sends Rα 7−→ sα , the skew Schur function associated to the ribbon α.
(c) Furthermore,
(5.3.11) Rα Rβ = Rα·β + Rα β if α and β are nonempty
|α|
(5.3.12) S(Rα ) = (−1) Rω(α)
Finally, R∅ is the multiplicative identity of NSym.
244See the proof of Proposition 1.4.22 for what this means.
245This definition of log? (f + u) is actually a particular case of Definition 1.7.17. This can be seen as follows:
We have f (A0 ) = 0. Thus, Proposition 1.7.11(h) (applied to C = A) yields f ∈ n (A, A) (where n (A, A) is defined as in
Section 1.7), so that (f + u) − u = f ∈ n (A, A). Therefore, Definition 1.7.17 defines a map log? (f + u) ∈ n (A, A). This
map is identical to the map log? (f + u) := n≥1 (−1)n−1 n 1 ?n
P
f we have just defined, because Proposition 1.7.18(f) (applied
?
to C = A) shows that the map log (f + u) defined using Definition 1.7.17 satisfies
X (−1)n−1 X 1
log? (f + u) = f ?n = (−1)n−1 f ?n .
n≥1
n n≥1
n
 HOPF ALGEBRAS IN COMBINATORICS 145

Proof. (a) For the first assertion, note that

 
X X X X
Hα = (Hα , Lβ )Rβ = Hα , Mγ  R β = Rβ .

β β γ: β:
γ refines β β coarsens α

The second assertion follows from the first by Lemma 5.2.7(a).

(b) Write α as (α1 , . . . , α` ). To show that π(Rα ) = sα , we instead examine π(Hα ):
π(Hα ) = π(Hα1 · · · Hα` ) = hα1 · · · hα` = s(α1 ) · · · s(α` ) = s(α1 )⊕···⊕(α` )
where (α1 ) ⊕ · · · ⊕ (α` ) is some skew shape which is a horizontal strip having rows of lengths α1 , . . . , α` from
bottom to top. We claim
X
s(α1 )⊕···⊕(α` ) = sβ
β:
β coarsens α

because column-strict tableaux T of shape (α1 ) ⊕ · · · ⊕ (α` ) biject to column-strict tableaux T 0 of some
ribbon β coarsening α, as follows: let ai , bi denote the leftmost, rightmost entries of the ith row from the
bottom in T , of length αi , and
• if bi ≤ ai+1 , merge parts αi , αi+1 in β, and concatenate the rows of length αi , αi+1 in T 0 , or
• if bi > ai+1 , do not merge parts αi , αi+1 in β, and let these two rows overlap in one column in T 0 .
E.g., if α = (3, 3, 2, 3, 2), this T of shape (α1 ) ⊕ · · · ⊕ (α` ) maps to this T 0 of shape β = (3, 8, 2):
3 4
4 4 5 3 4
T = 4 4 7−→ 2 2 3 4 4 4 4 5
2 2 3 1 1 3
1 1 3
0
The reverse bijection
P breaks the rows of T into the rows of T of lengths dictated by the parts of α. Having
shown π(Hα ) = β:β coarsens α sβ , we can now apply Lemma 5.2.7(a) to obtain
X `(α)−`(β)
sα = (−1) π (Hβ ) = π (Rα ) (by (5.3.10)) ;
β:β coarsens α

thus, π(Rα ) = sα is proven.

(c) Finally, (5.3.11) and (5.3.12) follow from (5.2.5) and (5.2.7) by duality. 

Remark 5.3.11. Since the maps

NSym QSym
<
π
"" - i
Λ
are Hopf morphisms, they must respect the antipodes SΛ , SQSym , SNSym , but it is interesting to compare
them explicitly using the fundamental basis for QSym and the ribbon basis for NSym.
On one hand (5.2.7) shows that SQSym (Lα ) = (−1)|α| Lω(α) extends the map SΛ since L(1n ) = en and
L(n) = hn , as observed in Example 5.2.5, and ω((n)) = (1n ).
On the other hand, (5.3.12) shows that SNSym (Rα ) = (−1)|α| Rω(α) lifts the map SΛ to SNSym : The-
orem 5.3.10(b) showed that Rα lifts the skew Schur function sα , while (2.4.8) asserted that S(sλ/µ ) =
(−1)|λ/µ| sλt /µt , and a ribbon α = λ/µ has ω(α) = λt /µt .
Exercise 5.3.12. (a) Show that any integers n and i with 0 ≤ i < n satisfy
i
X i−j
R(1i ,n−i) = (−1) R(1j ) Hn−j .
j=0

(Here, as usual, 1i stands for the number 1 repeated i times.)

146 DARIJ GRINBERG AND VICTOR REINER

(b) Show that any integers n and i with 0 ≤ i < n satisfy

i
X
i
(−1) R(1i ,n−i) = S (Hj ) Hn−j .
j=0

(c) For every positive integer n, define an element Ψn of NSym by

n−1
X i
Ψn = (−1) R(1i ,n−i) .
i=0

Show that Ψn = (S ? E) (Hn ), where the map E : NSym → NSym is defined as in Exercise 1.5.11
(for A = NSym). Conclude that Ψn is primitive.
(d) Prove that
n−1
X
Hk Ψn−k = nHn
k=0
for every n ∈ N.
(e) Define two power series ψ (t) and H
e (t) in NSym [[t]] by
X
ψ (t) = Ψn tn−1 ;
n≥1
X
H
e (t) = Hn tn .
n≥0

d e
Show that246 e (t) · ψ (t).
H (t) = H
dt
(The functions Ψn are called noncommutative power sums of the first kind ; they are studied in
[64]. The power sums of the second kind are the nξn in Remark 5.3.4.)
(f) Show that π (Ψn ) equals the power sum symmetric function pn for every positive integer n.
(g) Show that every positive integer n satisfies
n−1
X i
pn = (−1) s(n−i,1i ) in Λ.
i=0

(h) For every nonempty composition α, define a positive integer lp (α) by lp (α) = α` , where α is written
in the form α = (α1 , α2 , . . . , α` ) with ` = ` (α). (Thus, lp (α) is the last part of α.) Show that every
positive integer n satisfies
X `(α)−1
(5.3.13) Ψn = (−1) lp (α) Hα .
α∈Compn

(i) Assume that Q is a subring of k. For every composition α, define an element Ψα of NSym by
Ψα = Ψα1 Ψα2 · · · Ψα` , where α is written in the form α = (α1 , α2 , . . . , α` ) with ` = ` (α). For every
composition α, define πu (α) to be the positive integer α1 (α1 + α2 ) · · · (α1 + α2 + · · · + α` ), where
α is written in the form α = (α1 , α2 , . . . , α` ) with ` = ` (α). Show that
X 1
(5.3.14) Hn = Ψα
πu (α)
α∈Compn

for every n ∈ N.
Use this to prove that (Ψα )α∈Compn is a k-basis of NSymn for every n ∈ N.
(j) Assume that Q is a subring of k. Let V be the free k-module with basis (bn )n∈{1,2,3,...} . Define a k-
module homomorphism f : V → NSym by requiring that f (bn ) = Ψn for every n ∈ {1, 2, 3, . . .}. Let
F be the k-algebra homomorphism T (V ) → NSym induced by this f (using the universal property
of the tensor algebra T (V )). Show that F is a Hopf algebra isomorphism (where the Hopf algebra
structure on T (V ) is as in Example 1.4.16).

246The derivative d Q (t) of a power series Q (t) ∈ R [[t]] over a noncommutative ring R is defined just as in the case of R
dt
d
iqi ti−1 , where Q (t) is written in the form Q (t) = i
P P
commutative: by setting dt
Q (t) = i≥1 i≥0 qi t .
 HOPF ALGEBRAS IN COMBINATORICS 147

(k) Assume that Q is a subring of k. Let V be as in Exercise 5.3.12(j). Show that QSym is isomorphic
to the shuffle algebra Sh (V ) (defined as in Proposition 1.6.7) as Hopf algebras.
(l) Solve parts (a) and (b) of Exercise 2.9.14 again using the ribbon basis functions Rα .
One might wonder whether the Frobenius endomorphisms of Λ (defined in Exercise 2.9.9) and the Ver-
schiebung endomorphisms of Λ (defined in Exercise 2.9.10) generalize to analogous operators on either
QSym or NSym. The next two exercises (whose claims mostly come from [75, §13]) answer this question:
The Frobenius endomorphisms extend to QSym, and the Verschiebung ones lift to NSym.
Exercise 5.3.13. For every n ∈ {1, 2, 3, . . .}, define a map Fn : QSym → QSym by setting
Fn (a) = a (xn1 , xn2 , xn3 , . . .) for every a ∈ QSym .
(So what Fn does to a quasi-symmetric function is replacing all variables x1 , x2 , x3 , . . . by their n-th powers.)
(a) Show that Fn : QSym → QSym is a k-algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
(b) Show that Fn ◦ Fm = Fnm for any two positive integers n and m.
(c) Show that F1 = id.

(d) Prove that Fn M(β1 ,β2 ,...,βs ) = M(nβ1 ,nβ2 ,...,nβs ) for every n ∈ {1, 2, 3, . . .} and (β1 , β2 , . . . , βs ) ∈
Comp.
(e) Prove that Fn : QSym → QSym is a Hopf algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
(f) Consider the maps fn : Λ → Λ defined in Exercise 2.9.9. Show that Fn |Λ = fn for every n ∈
{1, 2, 3, . . .}.
(g) Assume that k = Z. Prove that fp (a) ≡ ap mod p QSym for every a ∈ QSym and every prime
number p.
(h) Give a new solution to Exercise 2.9.9(d).
Exercise 5.3.14. For every n ∈ {1, 2, 3, . . .}, define a k-algebra homomorphism Vn : NSym → NSym by

Hm/n , if n | m;
Vn (Hm ) = for every positive integer m
0, if n - m
247
.
(a) Show that any positive integers n and m satisfy

nΨm/n , if n | m;
Vn (Ψm ) = ,
0, if n - m
where the elements Ψm and Ψm/n of NSym are as defined in Exercise 5.3.12(c).
(b) Show that if Q is a subring of k, then any positive integers n and m satisfy

ξm/n , if n | m;
Vn (ξm ) = ,
0, if n - m
where the elements ξm and ξm/n of NSym are as defined in Exercise 5.3.5.
(c) Prove that Vn ◦ Vm = Vnm for any two positive integers n and m.
(d) Prove that V1 = id.
(e) Prove that Vn : NSym → NSym is a Hopf algebra homomorphism for every n ∈ {1, 2, 3, . . .}.
Now, consider also the maps Fn : QSym → QSym defined in Exercise 2.9.9. Fix a positive integer n.
(f) Prove that the maps Fn : QSym → QSym and Vn : NSym → NSym are adjoint with respect to the
(·,·)
dual pairing NSym ⊗ QSym −→ k.
(g) Consider the maps vn : Λ → Λ defined in Exercise 2.9.10. Show that the surjection π : NSym → Λ
satisfies vn ◦ π = π ◦ Vn for every n ∈ {1, 2, 3, . . .}.
(h) Give a new solution to Exercise 2.9.10(f).

247This is well-defined, since NSym is (isomorphic to) the free associative algebra with generators H , H , H , . . . (according
1 2 3
to (5.3.1)).
148 DARIJ GRINBERG AND VICTOR REINER

6. Polynomial generators for QSym and Lyndon words

In this chapter, we shall construct an algebraically independent generating set for QSym as a k-algebra,
thus showing that QSym is a polynomial ring over k. This has been done by Malvenuto [128, Cor. 4.19] when
k is a field of characteristic 0, and by Hazewinkel [74] in the general case. We will begin by introducing the
notion of Lyndon words (Section 6.1), on which both of these constructions rely; we will then (Section 6.2)
elucidate the connection of Lyndon words with shuffles, and afterwards (Section 6.3) apply it to prove
Radford’s theorem stating that the shuffle algebra of a free k-module over a commutative Q-algebra is a
polynomial algebra (Theorem 6.3.4). The shuffle algebra is not yet QSym, but Radford’s theorem on the
shuffle algebra serves as a natural stepping stone for the study of the more complicated algebra QSym. We
will prove – in two ways – that QSym is a polynomial algebra when Q is a subring of k in Section 6.4, and
then we will finally prove the general case in Section 6.5.
Strictly speaking, this whole Chapter 6 is a digression, as it involves almost no coalgebraic or Hopf-
algebraic structures, and its results will not be used in further chapters (which means it can be skipped if so
desired). However, it sheds additional light on QSym and serves as an excuse to study Lyndon words, which
are a combinatorial object of independent interest (and are involved in the study of free algebras and Hopf
algebras, apart from QSym – see [157] and [161]248).
We will take a scenic route to the proof of Hazewinkel’s theorem. A reader only interested in the proof
proper can restrict themselves to reading only the following:
• from Section 6.1, everything up to Corollary 6.1.6, then from Definition 6.1.13 up to Proposi-
tion 6.1.18, then from Definition 6.1.25 up to Lemma 6.1.28, and finally Theorem 6.1.30. (Proposi-
tion 6.1.19 and Theorem 6.1.20 are also relevant if one wants to use a different definition of Lyndon
words, as they prove the equivalence of most such definitions.)
• from Section 6.2, everything except for Exercise 6.2.25.
• from Section 6.3, Definition 6.3.1, Lemma 6.3.7, and Lemma 6.3.10.
• from Section 6.4, Definition 6.4.1, Theorem 6.4.3, then from Proposition 6.4.5 up to Definition 6.4.9,
and Lemma 6.4.11.
• all of Section 6.5.
6.1. Lyndon words. Lyndon words have been independently defined by Shirshov [179], Lyndon [124],
Radford [157, §2] and de Bruijn/Klarner [28] (though using different and sometimes incompatible notations).
They have since been surfacing in various places in noncommutative algebra (particularly the study of free
Lie algebras); expositions of their theory can be found in [122, §5], [161, §5.1] and [108, §1] (in German).
We will follow our own approach to the properties of Lyndon words that we need.
Definition 6.1.1. We fix a totally ordered set A, which we call the alphabet. Throughout Section 6.1 and
Section 6.2, we will understand “word” to mean a word over A.
We
F recall that a word is just a (finite) tuple of elements of A. In other words, a word is an element of the
set n≥0 An . We denote this set by A∗ .
The empty word is the unique tuple with 0 elements. It is denoted by ∅. If w ∈ An is a word and
i ∈ {1, 2, . . . , n}, then the i-th letter of w means the i-th entry of the n-tuple w. This i-th letter will be
denoted by wi .
The length ` (w) of a word w ∈ n≥0 An is defined to be the n ∈ N satisfying w ∈ An . Thus, w =
F


w1 , w2 , . . . , w`(w) for every word w.
Given two words u and v, we say that u is longer than v (or, equivalently, v is shorter than u) if and only
if ` (u) > ` (v).

The concatenation of two words u and v is defined to be the word u1 , u2 , . . . , u`(u) , v1 , v2 , . . . , v`(v) . This
concatenation is denoted by uv or u · v. The set A∗ of all words is a monoid with respect to concatenation,
with neutral element ∅. It is precisely the free monoid on generators A. If u is a word and i ∈ N, we will
understand ui to mean the i-th power of u in this monoid (that is, the word uu | {z · · · u}).
i times
The elements of A are called letters, and will be identified with elements of A1 ⊂ n≥0 An = A∗ .
F

This identification equates every letter u ∈ A with the one-letter word (u) ∈ A1 . Thus, every word
248They also are involved in indexing basis elements of combinatorial Hopf algebras other than QSym. See Bergeron/Zabrocki
[18].
 HOPF ALGEBRAS IN COMBINATORICS 149

(u1 , u2 , . . . , un ) ∈ A∗ equals the concatenation u1 u2 · · · un of letters, hence allowing us to use u1 u2 · · · un

as a brief notation for the word (u1 , u2 , . . . , un ).
If w is a word, then:
• a prefix of w means a word of the form (w1 , w2 , . . . , wi ) for some
i ∈ {0, 1, . . . , ` (w)};
• a suffix of w means a word of the form wi+1 , wi+2 , . . . , w`(w) for some
i ∈ {0, 1, . . . , ` (w)};
• a proper suffix of w means a word of the form wi+1 , wi+2 , . . . , w`(w) for some i ∈ {1, 2, . . . , ` (w)}.
In other words,
• a prefix of w ∈ A∗ is a word u ∈ A∗ such that there exists a v ∈ A∗ satisfying w = uv;
• a suffix of w ∈ A∗ is a word v ∈ A∗ such that there exists a u ∈ A∗ satisfying w = uv;
• a proper suffix of w ∈ A∗ is a word v ∈ A∗ such that there exists a nonempty u ∈ A∗ satisfying
w = uv.
Clearly, any proper suffix of w ∈ A∗ is a suffix of w. Moreover, if w ∈ A∗ is any word, then a proper suffix
of w is the same thing as a suffix of w distinct from w.
We define a relation ≤ on the set A∗ as follows: For two words u ∈ A∗ and v ∈ A∗ , we set u ≤ v to hold
if and only if
either there exists an i ∈ {1, 2, . . . , min {` (u) , ` (v)}}
such that (ui < vi , and every j ∈ {1, 2, . . . , i − 1} satisfies uj = vj ) ,
or the word u is a prefix of v.
This order relation (taken as the smaller-or-equal relation) makes A∗ into a poset (by Proposition 6.1.2(a)
below), and we will always be regarding A∗ as endowed with this poset structure (thus, notations such as <,
≤, > and ≥ will be referring to this poset structure). This poset is actually totally ordered (see Proposition
6.1.2(a)).
Here are some examples of words compared by the relation ≤:
113 ≤ 114, 113 ≤ 132, 19 ≤ 195, 41 ≤ 412,
41 ≤ 421, 539 ≤ 54, ∅ ≤ 21, ∅≤∅
(where A is the alphabet {1 < 2 < 3 < · · · }).
Notice that if u and v are two words of the same length (i.e., we have u, v ∈ An for one and the same
n), then u ≤ v holds if and only if u is lexicographically smaller-or-equal to v. In other words, the relation
≤ is an extension of the lexicographic order on every An to A∗ . This is the reason why this relation ≤
is usually called the lexicographic order on A∗ . In particular, we will be using this name.249 However,
unlike the lexicographic order on An , it does not always respect concatenation from the right: It can happen
that u, v, w ∈ A∗ satisfy u ≤ v but not uw ≤ vw. (For example, u = 1, v = 13 and w = 4, again with
A = {1 < 2 < 3 < · · · }.) We will see in Proposition 6.1.2 that this is rather an exception than the rule and
the relation ≤ still behaves mostly predictably with respect to concatenation.
Some basic properties of the order relation ≤ just defined are collected in the following proposition:

Proposition 6.1.2. (a) The order relation ≤ is (the smaller-or-equal relation of) a total order on the
set A∗ .
(b) If a, c, d ∈ A∗ satisfy c ≤ d, then ac ≤ ad.
(c) If a, c, d ∈ A∗ satisfy ac ≤ ad, then c ≤ d.
(d) If a, b, c, d ∈ A∗ satisfy a ≤ c, then either we have ab ≤ cd or the word a is a prefix of c.
(e) If a, b, c, d ∈ A∗ satisfy ab ≤ cd, then either we have a ≤ c or the word c is a prefix of a.
(f) If a, b, c, d ∈ A∗ satisfy ab ≤ cd and ` (a) ≤ ` (c), then a ≤ c.
(g) If a, b, c ∈ A∗ satisfy a ≤ b ≤ ac, then a is a prefix of b.
(h) If a ∈ A∗ is a prefix of b ∈ A∗ , then a ≤ b.
(i) If a and b are two prefixes of c ∈ A∗ , then either a is a prefix of b, or b is a prefix of a.
(j) If a, b, c ∈ A∗ are such that a ≤ b and ` (a) ≥ ` (b), then ac ≤ bc.
(k) If a ∈ A∗ and b ∈ A∗ are such that b is nonempty, then a < ab.
249The relation ≤ is also known as the dictionary order, due to the fact that it is the order in which words appear in a
dictionary.
150 DARIJ GRINBERG AND VICTOR REINER

Exercise 6.1.3. Prove Proposition 6.1.2.

[Hint: No part of Proposition 6.1.2 requires more than straightforward case analysis. However, the proof
of (a) can be simplified by identifying the order relation ≤ on A∗ as a restriction of the lexicographic order
on the set B∞ , where B is a suitable extension of the alphabet A. What is this extension, and how to embed
A∗ into B∞ ?]
Proposition 6.1.2 provides a set of tools for working with the lexicographic order without having to refer
to its definition; we shall use it extensively. Proposition 6.1.2(h) (and its equivalent form stating that a ≤ ac
for every a ∈ A∗ and c ∈ A∗ ) and Proposition 6.1.2(k) will often be used without explicit mention.
Before we define Lyndon words, let us show two more facts about words which will be used later. First,
when do words commute?
Proposition 6.1.4. Let u, v ∈ A∗ satisfy uv = vu. Then, there exist a t ∈ A∗ and two nonnegative integers
n and m such that u = tn and v = tm .
Proof. We prove this by strong induction on ` (u) + ` (v). We assume WLOG that ` (u) and ` (v) are positive
(because otherwise, one of u and v is the empty word, and everything is trivial). It is easy to see that either
u is a prefix of v, or v is a prefix of u 250. We assume WLOG that u is a prefix of v (since our situation
is symmetric). Thus,
 we can write v in the form v = uw for some w ∈ A∗ . Consider this w. Clearly,


` (u) + ` (w) = ` |{z}

uw  = ` (v) < ` (u) + ` (v) (since ` (v) is positive). Since v = uw, the equality uv = vu
=v
becomes uuw = uwu. Cancelling u from this equality, we obtain uw = wu. Now, we can apply Proposition
6.1.4 to w instead of v (by the induction assumption, since ` (u) + ` (w) < ` (u) + ` (v)), and obtain that
there exist a t ∈ A∗ and two nonnegative integers n and m such that u = tn and w = tm . Consider this t
and these n and m. Of course, u = tn and v = |{z} w = tn tm = tn+m . So the induction step is complete,
u |{z}
=tn =tm
and Proposition 6.1.4 is proven. 
∗
Proposition 6.1.5. Let u, v, w ∈ A be nonempty words satisfying uv ≥ vu, vw ≥ wv and wu ≥ uw. Then,
there exist a t ∈ A∗ and three nonnegative integers n, m and p such that u = tn , v = tm and w = tp .
Proof. We prove this by strong induction on ` (u) + ` (v) + ` (w). Clearly, ` (u), ` (v) and ` (w) are positive
(since u, v and w are nonempty). We assume WLOG that ` (u) = min {` (u) , ` (v) , ` (w)} (because there is
a cyclic symmetry in our situation). Thus, ` (u) ≤ ` (v) and ` (u) ≤ ` (w). But vu ≤ uv. Hence, Proposition
6.1.2(e) (applied to a = v, b = u, c = u and d = v) yields that either we have v ≤ u or the word u is a prefix
of v. But Proposition 6.1.2(f) (applied to a = u, b = w, c = w and d = u) yields u ≤ w (since uw ≤ wu
and ` (u) ≤ ` (w)). Furthermore, wv ≤ vw. Hence, Proposition 6.1.2(e) (applied to a = w, b = v, c = v and
d = w) yields that either we have w ≤ v or the word v is a prefix of w.
From what we have found so far, it is easy to see that u is a prefix of v 251. In other words, there exists
a v 0 ∈ A∗ such that v = uv 0 . Consider this v 0 .
If the word v 0 is empty, then the statement of Proposition 6.1.5 can be easily deduced from Proposition
6.1.4252. Thus, we assume WLOG that this is not the case. Hence, v 0 is nonempty.
Using v = uv 0 , we can rewrite uv ≥ vu as uuv 0 ≥ uv 0 u. That is, uv 0 u ≤ uuv 0 , so that v 0 u ≤ uv 0 (by
Proposition 6.1.2(c), applied to a = u, c = v 0 u and d = uv 0 ). That is, uv 0 ≥ v 0 u. But ` (uw) = ` (u) + ` (w) =
` (w) + ` (u) = ` (wu) ≥ ` (wu). Hence, Proposition 6.1.2(i) (applied to a = uw, b = wu and c = v 0 ) yields
250Proof. The word u is a prefix of uv. But the word v is also a prefix of uv (since uv = vu). Hence, Proposition 6.1.2(i)
(applied to a = u, b = v and c = uv) yields that either u is a prefix of v, or v is a prefix of u, qed.
251Proof. Assume the contrary. Then, u is not a prefix of v. Hence, we must have v ≤ u (since either we have v ≤ u or the
word u is a prefix of v), and in fact v < u (because v = u would contradict to u not being a prefix of v). Thus, v < u ≤ w. But
recall that either we have w ≤ v or the word v is a prefix of w. Thus, v must be a prefix of w (because v < w rules out w ≤ v).
In other words, there exists a q ∈ A∗ such that w = vq. Consider this q. We have v < u ≤ w = vq. Thus, Proposition 6.1.2(g)
(applied to a = v, b = u and c = q) yields that v is a prefix of u. In light of ` (u) ≤ ` (v), this is only possible if v = u, but this
contradicts v < u. This contradiction completes this proof.
252Proof. Assume that the word v 0 is empty. Then, v = uv 0 becomes v = u. Therefore, vw ≥ wv becomes uw ≥ wu.
Combined with wu ≥ uw, this yields uw = wu. Hence, Proposition 6.1.4 (applied to w instead of v) yields that there exist a
t ∈ A∗ and two nonnegative integers n and m such that u = tn and w = tm . Clearly, v = u = tn as well, and so the statement
of Proposition 6.1.5 is true.
 HOPF ALGEBRAS IN COMBINATORICS 151

uv 0 w = vw ≥ w |{z}
uwv 0 ≤ wuv 0 (since uw ≤ wu). Now, |{z} v = wuv 0 ≥ uwv 0 (since uwv 0 ≤ wuv 0 ), so that
=v =uv 0
uwv 0 ≤ uv 0 w. Hence, wv 0 ≤ v 0 w (by Proposition 6.1.2(c), applied to a = u, c = wv 0 and d = v 0 w), so that
v 0 w ≥ wv 0 . Now, we can apply Proposition 6.1.5 to v 0 instead of v (by the induction hypothesis, because
` (u) + ` (v 0 ) +` (w) = ` (v) + ` (w) < ` (u) + ` (v) + ` (w)). As a result, we see that there exist a t ∈ A∗ and
| {z }
=`(uv 0 )=`(v)
(since uv 0 =v)
three nonnegative integers n, m and p such that u = tn , v 0 = tm and w = tp . Clearly, this t and these n, m, p
satisfy v = |{z} v 0 = tn tm = tn+m , and so the statement of Proposition 6.1.5 is satisfied. The induction
u |{z}
=tn =tm
step is thus complete. 

Corollary 6.1.6. Let u, v, w ∈ A∗ be words satisfying uv ≥ vu and vw ≥ wv. Assume that v is nonempty.
Then, uw ≥ wu.

Proof. Assume the contrary. Thus, uw < wu, so that wu ≥ uw.

If u or w is empty, then everything is obvious. We thus WLOG assume that u and w are nonempty.
Thus, Proposition 6.1.5 shows that there exist a t ∈ A∗ and three nonnegative integers n, m and p such
tn |{z}
that u = tn , v = tm and w = tp . But this yields wu = tp tn = tp+n = tn+p = |{z} tp = uw, contradicting
=u =w
uw < wu. This contradiction finishes the proof. 

Exercise 6.1.7. Find an alternative proof of Corollary 6.1.6 which does not use Proposition 6.1.5.
The above results have a curious consequence, which we are not going to use:

Corollary 6.1.8. We can define a preorder on the set A∗ \{∅} of all nonempty words by defining a nonempty
word u to be greater-or-equal to a nonempty word v (with respect to this preorder) if and only if uv ≥ vu.
Two nonempty words u, v are equivalent with respect to the equivalence relation induced by this preorder if
and only if there exist a t ∈ A∗ and two nonnegative integers n and m such that u = tn and v = tm .

Proof. The alleged preorder is transitive (by Corollary 6.1.6) and reflexive (obviously), and hence is really a
preorder. The claim in the second sentence follows from Proposition 6.1.4. 

As another consequence of Proposition 6.1.5, we obtain a classical property of words [122, Proposition
1.3.1]:
Exercise 6.1.9. Let u and v be words and n and m be positive integers such that un = v m . Prove that
there exists a word t and positive integers i and j such that u = ti and v = tj .
Here is another application of Corollary 6.1.6:
Exercise 6.1.10. Let n and m be positive integers. Let u ∈ A∗ and v ∈ A∗ be two words. Prove that
uv ≥ vu holds if and only if un v m ≥ v m un holds.
Exercise 6.1.11. Let n and m be positive integers. Let u ∈ A∗ and v ∈ A∗ be two words satisfying
n` (u) = m` (v). Prove that uv ≥ vu holds if and only if un ≥ v m holds.
We can also generalize Propositions 6.1.4 and 6.1.5:
Exercise 6.1.12. Let u1 , u2 , . . . , uk be nonempty words such that every i ∈ {1, 2, . . . , k} satisfies ui ui+1 ≥
ui+1 ui , where uk+1 means u1 . Show that there exist a word t and nonnegative integers n1 , n2 , . . . , nk such
that u1 = tn1 , u2 = tn2 , . . ., uk = tnk .
Now, we define the notion of a Lyndon word. There are several definitions in literature, some of which
will be proven equivalent in Theorem 6.1.20.
Definition 6.1.13. A word w ∈ A∗ is said to be Lyndon if it is nonempty and satisfies the following
property: Every nonempty proper suffix v of w satisfies v > w.
152 DARIJ GRINBERG AND VICTOR REINER

For example, the word 113 is Lyndon (because its nonempty proper suffixes are 13 and 3, and these are
both > 113), and the word 242427 is Lyndon (its nonempty proper suffixes are 42427, 2427, 427, 27 and
7, and again these are each > 242427). The words 2424 and 35346 are not Lyndon (the word 2424 has a
nonempty proper suffix 24 ≤ 2424, and the word 35346 has a nonempty proper suffix 346 ≤ 35346). Every
word of length 1 is Lyndon (since it has no nonempty proper suffixes). A word w = (w1 , w2 ) with two letters
is Lyndon if and only if w1 < w2 . A word w = (w1 , w2 , w3 ) of length 3 is Lyndon if and only if w1 < w3 and
w1 ≤ w2 . A four-letter word w = (w1 , w2 , w3 , w4 ) is Lyndon if and only if w1 < w4 , w1 ≤ w3 , w1 ≤ w2 and
(if w1 = w3 then w2 < w4 ). (These rules only get more complicated as the words grow longer.)
We will show several properties of Lyndon words now. We begin with trivialities which will make some
arguments a bit shorter:
Proposition 6.1.14. Let w be a Lyndon word. Let u and v be words such that w = uv.
(a) If v is nonempty, then v ≥ w.
(b) If v is nonempty, then v > u.
(c) If u and v are nonempty, then vu > uv.
(d) We have vu ≥ uv.
Proof. (a) Assume that v is nonempty. Clearly, v is a suffix of w (since w = uv). If v is a proper suffix of w,
then the definition of a Lyndon word yields that v > w (since w is a Lyndon word); otherwise, v must be w
itself. In either case, we have v ≥ w. Hence, Proposition 6.1.14(a) is proven.
(b) Assume that v is nonempty. From Proposition 6.1.14(a), we obtain v ≥ w = uv > u (since v is
nonempty). This proves Proposition 6.1.14(b).
(c) Assume that u and v are nonempty. Since u is nonempty, we have vu > v ≥ w (by Proposition
6.1.14(a)). Since w = uv, this becomes vu > uv. This proves Proposition 6.1.14(c).
(d) We need to prove that vu ≥ uv. If either u or v is empty, vu and uv are obviously equal, and thus
vu ≥ uv is true in this case. Hence, we can WLOG assume that u and v are nonempty. Assume this. Then,
vu ≥ uv follows from Proposition 6.1.14(c). This proves Proposition 6.1.14(d). 
Corollary 6.1.15. Let w be a Lyndon word. Let v be a nonempty suffix of w. Then, v ≥ w.
Proof. Since v is a nonempty suffix of w, there exists u ∈ A∗ such that w = uv. Thus, v ≥ w follows from
Proposition 6.1.14(a). 
Our next proposition is [78, Lemma 6.5.4]; its part (a) is also [161, (5.1.2)]:
Proposition 6.1.16. Let u and v be two Lyndon words such that u < v. Then:
(a) The word uv is Lyndon.
(b) We have uv < v.
Proof. (b) The word u is Lyndon and thus nonempty. Hence, uv 6= v 253. If uv ≤ v∅, then Proposition
6.1.16(b) easily follows254. Hence, for the rest of this proof, we can WLOG assume that we don’t have
uv ≤ v∅. Assume this.
We have u < v. Hence, Proposition 6.1.2(d) (applied to a = u, b = v, c = v and d = ∅) yields that either
we have uv ≤ v∅ or the word u is a prefix of v. Since we don’t have uv ≤ v∅, we thus see that the word u
is a prefix of v. In other words, there exists a t ∈ A∗ satisfying v = ut. Consider this t. Then, t is nonempty
(else we would have v = u |{z}t = u in contradiction to u < v).
=∅
Now, v = ut. Hence, t is a proper suffix of v (proper because u is nonempty). Thus, t is a nonempty
proper suffix of v. Since every nonempty proper suffix of v is > v (because v is Lyndon), this shows that
t > v. Hence, v ≤ t. Thus, Proposition 6.1.2(b) (applied to a = u, c = v and d = t) yields uv ≤ ut = v.
Combined with uv 6= v, this yields uv < v. Hence, Proposition 6.1.16(b) is proven.
(a) The word v is nonempty (since it is Lyndon). Hence, uv is nonempty. It thus remains to check that
every nonempty proper suffix p of uv satisfies p > uv.
253Proof. Assume the contrary. Then, uv = v. Thus, uv = v = ∅v. Cancelling v from this equation, we obtain u = ∅. That
is, u is empty. This contradicts the fact that u is nonempty. This contradiction proves that our assumption was wrong, qed.
254Proof. Assume that uv ≤ v∅. Thus, uv ≤ v∅ = v. Since uv 6= v, this becomes uv < v, so that Proposition 6.1.16(b) is
proven.
 HOPF ALGEBRAS IN COMBINATORICS 153

So let p be a nonempty proper suffix of uv. We must show that p > uv. Since p is a nonempty proper
suffix of uv, we must be in one of the following two cases (depending on whether this suffix begins before
the suffix v of uv begins or afterwards):
Case 1: The word p is a nonempty suffix of v. (Note that p = v is allowed.)
Case 2: The word p has the form qv where q is a nonempty proper suffix of u.
Let us first handle Case 1. In this case, p is a nonempty suffix of v. Since v is Lyndon, this yields that
p ≥ v (by Corollary 6.1.15, applied to v and p instead of w and v). But Proposition 6.1.16(b) yields uv < v,
thus v > uv. Hence, p ≥ v > uv. We thus have proven p > uv in Case 1.
Let us now consider Case 2. In this case, p has the form qv where q is a nonempty proper suffix of u.
Consider this q. Clearly, q > u (since u is Lyndon and since q is a nonempty proper suffix of u), so that
u ≤ q. Thus, Proposition 6.1.2(d) (applied to a = u, b = v, c = q and d = v) yields that either we have
uv ≤ qv or the word u is a prefix of q. Since u being a prefix of q is impossible (in fact, q is a proper suffix
of u, thus shorter than u), we thus must have uv ≤ qv. Since uv 6= qv (because otherwise we would have
uv = qv, thus u = q (because we can cancel v from the equality uv = qv), contradicting q > u), this can be
strengthened to uv < qv = p. Thus, p > uv is proven in Case 2 as well.
Now that p > uv is shown to hold in both cases, we conclude that p > uv always holds.
Now, let us forget that we fixed p. We have thus shown that every nonempty proper suffix p of uv satisfies
p > uv. Since uv is nonempty, this yields that uv is Lyndon (by the definition of a Lyndon word). Thus,
the proof of Proposition 6.1.16(a) is complete. 
Proposition 6.1.16(b), combined with Corollary 6.1.6, leads to a technical result which we will find good
use for later:
Corollary 6.1.17. Let u and v be two Lyndon words such that u < v. Let z be a word such that zv ≥ vz
and uz ≥ zu. Then, z is the empty word.
Proof. Assume the contrary. Then, z is nonempty. Thus, Corollary 6.1.6 (applied to z and v instead of v and
w) yields uv ≥ vu. But Proposition 6.1.16(b) yields uv < v ≤ vu, contradicting uv ≥ vu. This contradiction
completes our proof. 
We notice that the preorder of Corollary 6.1.8 becomes particularly simple on Lyndon words:
Proposition 6.1.18. Let u and v be two Lyndon words. Then, u ≥ v if and only if uv ≥ vu.
Proof. We distinguish between three cases:
Case 1: We have u < v.
Case 2: We have u = v.
Case 3: We have u > v.
Let us consider Case 1. In this case, we have u < v. Thus,
uv < v (by Proposition 6.1.16(b))
≤ vu.
Hence, we have neither u ≥ v nor uv ≥ vu (because we have u < v and uv < vu). Thus, Proposition 6.1.18
is proven in Case 1.
In Case 2, we have u = v. Therefore, in Case 2, both inequalities u ≥ v and uv ≥ vu hold (and actually
are equalities). Thus, Proposition 6.1.18 is proven in Case 2 as well.
Let us finally consider Case 3. In this case, we have u > v. In other words, v < u. Thus,
vu < u (by Proposition 6.1.16(b), applied to v and u instead of u and v)
≤ uv.
Hence, we have both u ≥ v and uv ≥ vu (because we have v < u and vu < uv). Thus, Proposition 6.1.18 is
proven in Case 3.
Proposition 6.1.18 is now proven in all three possible cases. 
Proposition 6.1.19. Let w be a nonempty word. Let v be the (lexicographically) smallest nonempty suffix
of w. Then:
(a) The word v is a Lyndon word.
154 DARIJ GRINBERG AND VICTOR REINER

(b) Assume that w is not a Lyndon word. Then there exists a nonempty u ∈ A∗ such that w = uv,
u ≥ v and uv ≥ vu.
Proof. (a) Every nonempty proper suffix of v is ≥ v (since every nonempty proper suffix of v is a nonempty
suffix of w, but v is the smallest such suffix) and therefore > v (since a proper suffix of v cannot be = v).
Combined with the fact that v is nonempty, this yields that v is Lyndon. Proposition 6.1.19(a) is proven.
(b) Assume that w is not a Lyndon word. Then, w 6= v (since v is Lyndon (by Proposition 6.1.19(a))
while w is not). Now, v is a suffix of w. Thus, there exists an u ∈ A∗ such that w = uv. Consider this u.
Clearly, u is nonempty (since uv = w 6= v). Assume (for the sake of contradiction) that u < v. Let v 0 be
the (lexicographically) smallest nonempty suffix of u. Then, v 0 is a Lyndon word (by Proposition 6.1.19(a),
applied to u and v 0 instead of w and v) and satisfies v 0 ≤ u (since u is a nonempty suffix of u, whereas v 0
is the smallest such suffix). Thus, v 0 and v are Lyndon words such that v 0 ≤ u < v. Proposition 6.1.16(a)
(applied to v 0 instead of u) now yields that the word v 0 v is Lyndon. Hence, every nonempty proper suffix of
v 0 v is > v 0 v. Since v is a nonempty proper suffix of v 0 v, this yields that v > v 0 v.
But v 0 is a nonempty suffix of u, so that v 0 v is a nonempty suffix of uv = w. Since v is the smallest such
suffix, this yields that v 0 v ≥ v. This contradicts v > v 0 v. Our assumption (that u < v) therefore falls. We
conclude that u ≥ v.
It remains to prove that uv ≥ vu. Assume the contrary. Then, uv < vu. Thus, there exists at least one
suffix t of u such that tv < vt (namely, t = u). Let p be the minimum-length such suffix. Then, pv < vp.
Thus, p is nonempty.
Since p is a suffix of u, it is clear that pv is a suffix of uv = w. So we know that pv is a nonempty suffix
of w. Since v is the smallest such suffix, this yields that v ≤ pv < vp. Thus, Proposition 6.1.2(g) (applied
to a = v, b = pv and c = p) yields that v is a prefix of pv. In other words, there exists a q ∈ A∗ such
that pv = vq. Consider this q. This q is nonempty (because otherwise we would have pv = v q = v,
|{z}
=∅
contradicting the fact that p is nonempty). From vq = pv < vp, we obtain q ≤ p (by Proposition 6.1.2(c),
applied to a = v, c = q and d = p).
We know that q is a suffix of pv (since vq = pv), whereas pv is a suffix of w. Thus, q is a suffix of w.
So q is a nonempty suffix of w. Since v is the smallest such suffix, this yields that v ≤ q. We now have
v ≤ q ≤ p ≤ pv < vp. Hence, v is a prefix of p (by Proposition 6.1.2(g), applied to a = v, b = p and c = p).
In other words, there exists an r ∈ A∗ such that p = vr. Consider this r. Clearly, r is a suffix of p, while
p is a suffix of u; therefore, r is a suffix of u. Also, pv < vp rewrites as vrv < vvr (because p = vr). Thus,
Proposition 6.1.2(c) (applied to a = v, c = rv and d = vr) yields rv ≤ vr. Since rv 6= vr (because otherwise,
we would have rv = vr, thus v |{z} rv = vvr, contradicting vrv < vvr), this becomes rv < vr.
=vr
Now, r is a suffix of u such that rv  Since p is the minimum-length such suffix, this yields ` (r) ≥ ` (p).
< vr. 

But this contradicts the fact that `  p  = ` (vr) = ` (v) +` (r) > ` (r). This contradiction proves our
|{z} |{z}
=vr >0
assumption wrong; thus, we have shown that uv ≥ vu. Proposition 6.1.19(b) is proven. 
Theorem 6.1.20. Let w be a nonempty word. The following four assertions are equivalent:
• Assertion A: The word w is Lyndon.
• Assertion B: Any nonempty words u and v satisfying w = uv satisfy v > w.
• Assertion C: Any nonempty words u and v satisfying w = uv satisfy v > u.
• Assertion D: Any nonempty words u and v satisfying w = uv satisfy vu > uv.
Proof. Proof of the implication A =⇒ B: If Assertion A holds, then Assertion B clearly holds (in fact,
whenever u and v are nonempty words satisfying w = uv, then v is a nonempty proper suffix of w, and
therefore > w by the definition of a Lyndon word).
Proof of the implication A =⇒ C: This implication follows from Proposition 6.1.14(b).
Proof of the implication A =⇒ D: This implication follows from Proposition 6.1.14(c).
Proof of the implication B =⇒ A: Assume that Assertion B holds. If v is a nonempty proper suffix of w,
then there exists an u ∈ A∗ satisfying w = uv. This u is nonempty because v is a proper suffix, and thus
Assertion B yields v > w. Hence, every nonempty proper suffix v of w satisfies v > w. By the definition of
a Lyndon word, this yields that w is Lyndon, so that Assertion A holds.
 HOPF ALGEBRAS IN COMBINATORICS 155

Proof of the implication C =⇒ A: Assume that Assertion C holds. If w was not Lyndon, then Proposition
6.1.19(b) would yield nonempty words u and v such that w = uv and u ≥ v; this would contradict Assertion
C. Thus, w is Lyndon, and Assertion A holds.
Proof of the implication D =⇒ A: Assume that Assertion D holds. If w was not Lyndon, then Proposition
6.1.19(b) would yield nonempty words u and v such that w = uv and uv ≥ vu; this would contradict Assertion
D. Thus, w is Lyndon, and Assertion A holds.
Now we have proven enough implications to conclude the equivalence of all four assertions. 
Theorem 6.1.20 connects our definition of Lyndon words with some of the definitions appearing in litera-
ture. For example, Lothaire [122, §5.1], Shirshov [179] and de Bruijn/Klarner [28, §4] define Lyndon words
using Assertion D (note, however, that Shirshov takes < instead of > and calls Lyndon words “regular words”;
also, de Bruijn/Klarner call Lyndon words “normal words”). Chen-Fox-Lyndon [35, §1], Reutenauer [161]
and Radford [157] use our definition (but Chen-Fox-Lyndon call the Lyndon words “standard sequences”,
and Radford calls them “primes” and uses < instead of >).
Theorem 6.1.20 appears (with different notations) in Zhou-Lu [205, Proposition 1.4]. The equivalence
D ⇐⇒ A of our Theorem 6.1.20 is equivalent to [122, Proposition 5.12] and to [35, A00 = A000 ].
The following exercise provides a different (laborious) approach to Theorem 6.1.20:
Exercise 6.1.21. (a) Prove that if u ∈ A∗ and v ∈ A∗ are two words satisfying uv < vu, then there
exists a nonempty suffix s of u satisfying sv < v.
(b) Give a new proof of Theorem 6.1.20 (avoiding the use of Proposition 6.1.19).
[Hint: For (a), perform strong induction on ` (u) + ` (v), assume the contrary, and distinguish between
the case when u ≤ v and the case when v is a prefix of u. For (b), use part (a) in proving the implication
D =⇒ B, and factor v as v = um v 0 with m maximal in the proof of the implication C =⇒ B.]
The following two exercises are taken from [76]255.
Exercise 6.1.22. Let w be a nonempty word. Prove that w is Lyndon if and only if every nonempty word
t and every positive integer n satisfy (if w ≤ tn , then w ≤ t).
Exercise 6.1.23. Let w1 , w2 , . . ., wn be n Lyndon words, where n is a positive integer. Assume that
w1 ≤ w2 ≤ · · · ≤ wn and w1 < wn . Show that w1 w2 · · · wn is a Lyndon word.
The following exercise is a generalization (albeit not in an obvious way) of Exercise 6.1.23:
Exercise 6.1.24. Let w1 , w2 , . . ., wn be n Lyndon words, where n is a positive integer. Assume that
wi wi+1 · · · wn ≥ w1 w2 · · · wn for every i ∈ {1, 2, . . . , n}. Show that w1 w2 · · · wn is a Lyndon word.
We are now ready to meet the one of the most important features of Lyndon words: a bijection between all
words and multisets of Lyndon words256; it is clear that such a bijection is vital for constructing polynomial
generating sets of commutative algebras with bases indexed by words, such as QSym or shuffle algebras.
This is given by the Chen-Fox-Lyndon factorization:
Definition 6.1.25. Let w be a word. A Chen-Fox-Lyndon factorization (in short, CFL factorization) of w
means a tuple (a1 , a2 , . . . , ak ) of Lyndon words satisfying w = a1 a2 · · · ak and a1 ≥ a2 ≥ · · · ≥ ak .
Example 6.1.26. The tuple (23, 2, 14, 13323, 13, 12, 12, 1) is a CFL factorization of the word 23214133231312121
over the alphabet {1, 2, 3, . . .} (ordered by 1 < 2 < 3 < · · · ), since 23, 2, 14, 13323, 13, 12, 12 and 1 are
Lyndon words satisfying 23214133231312121 = 23 · 2 · 14 · 13323 · 13 · 12 · 12 · 1 and 23 ≥ 2 ≥ 14 ≥ 13323 ≥
13 ≥ 12 ≥ 12 ≥ 1.
The bijection is given by the following Chen-Fox-Lyndon theorem ([78, Theorem 6.5.5], [122, Thm. 5.1.5],
[157, part of Thm. 2.1.4]):

Theorem 6.1.27. Let w be a word. Then, there exists a unique CFL factorization of w.

Before we prove this, we need to state a lemma (which is [122, Proposition 5.1.6]):

255Exercise 6.1.22 is more or less [76, Lemma 4.3] with a converse added; Exercise 6.1.23 is [76, Lemma 4.2].
256And it is not even the only one: see [69, §3] for another.
156 DARIJ GRINBERG AND VICTOR REINER

Lemma 6.1.28. Let (a1 , a2 , . . . , ak ) be a CFL factorization of a nonempty word w. Let p be a nonempty
suffix of w. Then, p ≥ ak .
Proof of Lemma 6.1.28. We will prove Lemma 6.1.28 by induction over the (obviously) positive integer k.
Induction base: Assume that k = 1. Thus, (a1 , a2 , . . . , ak ) = (a1 ) is a tuple of Lyndon words satisfying
w = a1 a2 · · · ak . We have w = a1 a2 · · · ak = a1 (since k = 1), so that w is a Lyndon word (since a1 is a
Lyndon word). Thus, Corollary 6.1.15 (applied to v = p) yields p ≥ w = a1 = ak (since 1 = k). Thus,
Lemma 6.1.28 is proven in the case k = 1. The induction base is complete.
Induction step: Let K be a positive integer. Assume (as the induction hypothesis) that Lemma 6.1.28 is
proven for k = K. We now need to show that Lemma 6.1.28 holds for k = K + 1.
So let (a1 , a2 , . . . , aK+1 ) be a CFL factorization of a nonempty word w. Let p be a nonempty suffix of w.
We need to prove that p ≥ aK+1 .
By the definition of a CFL factorization, (a1 , a2 , . . . , aK+1 ) is a tuple of Lyndon words satisfying w =
a1 a2 · · · aK+1 and a1 ≥ a2 ≥ · · · ≥ aK+1 . Let w0 = a2 a3 · · · aK+1 ; then, w = a1 a2 · · · aK+1 = a1 (a2 a3 · · · aK+1 ) =
| {z }
=w0
a1 w0 . Hence, every nonempty suffix of w is either a nonempty suffix of w0 , or has the form qw0 for a nonempty
suffix q of a1 . Since p is a nonempty suffix of w, we thus must be in one of the following two cases:
Case 1: The word p is a nonempty suffix of w0 .
Case 2: The word p has the form qw0 for a nonempty suffix q of a1 .
Let us first consider Case 1. In this case, p is a nonempty suffix of w0 . The K-tuple (a2 , a3 , . . . , aK+1 ) of
Lyndon words satisfies w0 = a2 a3 · · · aK+1 and a2 ≥ a3 ≥ · · · ≥ aK+1 ; therefore, (a2 , a3 , . . . , aK+1 ) is a CFL
factorization of w0 . We can thus apply Lemma 6.1.28 to K, w0 and (a2 , a3 , . . . , aK+1 ) instead of k, w and
(a1 , a2 , . . . , ak ) (because we assumed that Lemma 6.1.28 is proven for k = K). As a result, we obtain that
p ≥ aK+1 . Thus, p ≥ aK+1 is proven in Case 1.
Let us now consider Case 2. In this case, p has the form qw0 for a nonempty suffix q of a1 . Consider this
q. Since a1 is a Lyndon word, we have q ≥ a1 (by Corollary 6.1.15, applied to a1 and q instead of w and v).
Thus, q ≥ a1 ≥ a2 ≥ · · · ≥ aK+1 , so that p = qw0 ≥ q ≥ aK+1 . Thus, p ≥ aK+1 is proven in Case 2.
We have now proven p ≥ aK+1 in all cases. This proves that Lemma 6.1.28 holds for k = K + 1. The
induction step is thus finished, and with it the proof of Lemma 6.1.28. 
Proof of Theorem 6.1.27. Let us first prove that there exists a CFL factorization of w.
Indeed, there clearly exists a tuple (a1 , a2 , . . . , ak ) of Lyndon words satisfying w = a1 a2 · · · ak 257. Fix
such a tuple with minimum k. We claim that a1 ≥ a2 ≥ · · · ≥ ak .
Indeed, if some i ∈ {1, 2, . . . , k − 1} would satisfy ai < ai+1 , then the word ai ai+1 would be Lyndon (by
Proposition 6.1.16(a), applied to u = ai and v = ai+1 ), whence (a1 , a2 , . . . , ai−1 , ai ai+1 , ai+2 , ai+3 , . . . , ak )
would also be a tuple of Lyndon words satisfying w = a1 a2 · · · ai−1 (ai ai+1 ) ai+2 ai+3 · · · ak but having length
k −1 < k, contradicting the fact that k is the minimum length of such a tuple. Hence, no i ∈ {1, 2, . . . , k − 1}
can satisfy ai < ai+1 . In other words, every i ∈ {1, 2, . . . , k − 1} satisfies ai ≥ ai+1 . In other words,
a1 ≥ a2 ≥ · · · ≥ ak . Thus, (a1 , a2 , . . . , ak ) is a CFL factorization of w, so we have shown that such a CFL
factorization exists.
It remains to show that there exists at most one CFL factorization of w. We shall prove this by induction
over ` (w). Thus, we fix a word w and assume that
(6.1.1) for every word v with ` (v) < ` (w) , there exists at most one CFL factorization of v.
We now have to prove that there exists at most one CFL factorization of w.
Indeed, let (a1 , a2 , . . . , ak ) and (b1 , b2 , . . . , bm ) be two CFL factorizations of w. We need to prove that
(a1 , a2 , . . . , ak ) = (b1 , b2 , . . . , bm ). If w is empty, then this is obvious, so we WLOG assume that it is not;
thus, k > 0 and m > 0.
Since (b1 , b2 , . . . , bm ) is a CFL factorization of w, we have w = b1 b2 · · · bm , and thus bm is a nonempty
suffix of w. Thus, Lemma 6.1.28 (applied to p = bm ) yields bm ≥ ak . The same argument (but with the
roles of (a1 , a2 , . . . , ak ) and (b1 , b2 , . . . , bm ) switched) shows that ak ≥ bm . Combined with bm ≥ ak , this
yields ak = bm . Now let v = a1 a2 · · · ak−1 . Then, (a1 , a2 , . . . , ak−1 ) is a CFL factorization of v (since
a1 ≥ a2 ≥ · · · ≥ ak−1 ).
257For instance, the tuple w , w , . . . , w

1 2 `(w) of one-letter words is a valid example (recall that one-letter words are always
Lyndon).
 HOPF ALGEBRAS IN COMBINATORICS 157

Since (a1 , a2 , . . . , ak ) is a CFL factorization of w, we have w = a1 a2 · · · ak = a1 a2 · · · ak−1 ak = vbm , so

| {z } |{z}
=v =bm
that
vbm = w = b1 b2 · · · bm = b1 b2 · · · bm−1 bm .
Cancelling bm yields v = b1 b2 · · · bm−1 . Thus, (b1 , b2 , . . . , bm−1 ) is a CFL factorization of v (since b1 ≥ b2 ≥
· · · ≥ bm−1 ). Since ` (v) < ` (w) (because v = a1 a2 · · · ak−1 is shorter than w = a1 a2 · · · ak ), we can apply
(6.1.1) to obtain that there exists at most one CFL factorization of v. But we already know two such CFL
factorizations: (a1 , a2 , . . . , ak−1 ) and (b1 , b2 , . . . , bm−1 ). Thus, (a1 , a2 , . . . , ak−1 ) = (b1 , b2 , . . . , bm−1 ), which,
combined with ak = bm , leads to (a1 , a2 , . . . , ak ) = (b1 , b2 , . . . , bm ). This is exactly what we needed to prove.
So we have shown (by induction) that there exists at most one CFL factorization of w. This completes the
proof of Theorem 6.1.27. 
The CFL factorization allows us to count all Lyndon words of a given length if A is finite:
Exercise 6.1.29. Assume that the alphabet A is finite. Let q = |A|. Show that the number of Lyndon words
1P
µ (d) q n/d for every positive integer n (where “ ” means a sum over all positive
P
of length n equals
n d|n d|n
divisors of n, and where µ is the number-theoretic Möbius function).258
Exercise 6.1.29 is a well-known result and appears, e.g., in [35, Theorem 1.5].
We will now study another kind of factorization: not of an arbitrary word into Lyndon words, but of a
Lyndon word into two smaller Lyndon words. This factorization is called standard factorization ([122, §5.1])
or canonical factorization ([78, Lemma 6.5.33]); we only introduce it from the viewpoint we are interested
in, namely its providing a way to do induction over Lyndon words259. Here is what we need to know:

Theorem 6.1.30. Let w be a Lyndon word of length > 1. Let v be the (lexicographically) smallest
nonempty proper suffix of w. Since v is a proper suffix of w, there exists a nonempty u ∈ A∗ such that
w = uv. Consider this u. Then:
(a) The words u and v are Lyndon.
(b) We have u < w < v.

Proof. Every nonempty proper suffix of v is ≥ v (since every nonempty proper suffix of v is a nonempty
proper suffix of w, but v is the smallest such suffix) and therefore > v (since a proper suffix of v cannot be
= v). Combined with the fact that v is nonempty, this yields that v is Lyndon.
Since w is Lyndon, we know that every nonempty proper suffix of w is > w. Applied to the nonempty
proper suffix v of w, this yields that v > w. Hence, w < v. Since v is nonempty, we have u < uv = w < v.
This proves Theorem 6.1.30(b).
Let p be a nonempty proper suffix of u. Then, pv is a nonempty proper suffix of uv = w. Thus, pv > w
(since every nonempty proper suffix of w is > w). Thus, pv > w = uv, so that uv < pv. Thus, Proposition
6.1.2(e) (applied to a = u, b = v, c = p and d = v) yields that either we have u ≤ p or the word p is a prefix
of u.
Let us assume (for the sake of contradiction) that p ≤ u. Then, p < u (because p is a proper suffix of u,
and therefore p 6= u). Hence, we cannot have u ≤ p. Thus, the word p is a prefix of u (since either we have
u ≤ p or the word p is a prefix of u). In other words, there exists a q ∈ A∗ such that u = pq. Consider this
q. We have w = |{z}u v = pqv = p (qv), and thus qv is a proper suffix of w (proper because p is nonempty).
=pq
Moreover, qv is nonempty (since v is nonempty). Hence, qv is a nonempty proper suffix of w. Since v is
the smallest such suffix, this entails that v ≤ qv. Proposition 6.1.2(b) (applied to a = p, c = v and d = qv)
thus yields pv ≤ pqv. Hence, pv ≤ pqv = w, which contradicts pv > w. This contradiction shows that our
assumption (that p ≤ u) was false. We thus have p > u.
We now have shown that p > u whenever p is a nonempty proper suffix of u. Combined with the fact that
u is nonempty, this shows that u is a Lyndon word. This completes the proof of Theorem 6.1.30(a). 

258In particular, 1 P µ (d) q n/d is an integer.

n d|n
259
e.g., allowing to solve Exercise 6.1.24 in a simpler way
158 DARIJ GRINBERG AND VICTOR REINER

Another approach to the standard factorization is given in the following exercise:

Exercise 6.1.31. Let w be a Lyndon word of length > 1. Let v be the longest proper suffix of w such that
v is Lyndon260. Since v is a proper suffix of w, there exists a nonempty u ∈ A∗ such that w = uv. Consider
this u. Prove that:
(a) The words u and v are Lyndon.
(b) We have u < w < v.
(c) The words u and v are precisely the words u and v constructed in Theorem 6.1.30.
Notice that a well-known recursive characterization of Lyndon words [35, A0 = A00 ] can be easily derived
from Theorem 6.1.30 and Proposition 6.1.16(a). We will not dwell on it.
The following exercise surveys some variations on the characterizations of Lyndon words261:
Exercise 6.1.32. Let w be a nonempty word. Consider the following nine assertions:
• Assertion A0 : The word w is a power of a Lyndon word.
• Assertion B 0 : If u and v are nonempty words satisfying w = uv, then either we have v ≥ w or the
word v is a prefix of w.
• Assertion C 0 : If u and v are nonempty words satisfying w = uv, then either we have v ≥ u or the
word v is a prefix of u.
• Assertion D0 : If u and v are nonempty words satisfying w = uv, then we have vu ≥ uv.
• Assertion E 0 : If u and v are nonempty words satisfying w = uv, then either we have v ≥ u or the
word v is a prefix of w.
• Assertion F 0 : The word w is a prefix of a Lyndon word in A∗ .
• Assertion F 00 : Let m be an object not in the alphabet A. Let us equip the set A ∪ {m} with a total
order which extends the total order on the alphabet A and which satisfies (a < m for every a ∈ A).
∗
Then, the word wm ∈ (A ∪ {m}) (the concatenation of the word w with the one-letter word m) is
a Lyndon word.
• Assertion G 0 : There exists a Lyndon word t ∈ A∗ , a positive integer ` and a prefix p of t (possibly
empty) such that w = t` p.
• Assertion H0 : There exists a Lyndon word t ∈ A∗ , a nonnegative integer ` and a prefix p of t
(possibly empty) such that w = t` p.
(a) Prove the equivalence A0 ⇐⇒ D0 .
(b) Prove the equivalence B 0 ⇐⇒ C 0 ⇐⇒ E 0 ⇐⇒ F 00 ⇐⇒ G 0 ⇐⇒ H0 .
(c) Prove the implication F 0 =⇒ B 0 .
(d) Prove the implication D0 =⇒ B 0 . (The implication B 0 =⇒ D0 is false, as witnessed by the word
11211.)
(e) Prove that if there exists a letter µ ∈ A such that (µ > a for every letter a of w), then the equivalence
F 0 ⇐⇒ F 00 holds.
(f) Prove that if there exists a letter µ ∈ A such that (µ > a for some letter a of w), then the equivalence
F 0 ⇐⇒ F 00 holds.
The next exercise (based on work of Hazewinkel [77]) extends some of the above properties of Lyndon
words (and words in general) to a more general setting, in which the alphabet A is no longer required to be
totally ordered, but only needs to be a poset:
Exercise 6.1.33. In this exercise, we shall loosen the requirement that the alphabet A be a totally ordered
set: Instead, we will only require A to be a poset. The resulting more general setting will be called the
partial-order setting, to distinguish it from the total-order setting in which A is required to be a totally
ordered set. All results in Chapter 6 so far address the total-order setting. In this exercise, we will generalize
some of them to the partial-order setting.
All notions that we have defined in the total-order setting (the notion of a word, the relation ≤, the notion
of a Lyndon word, etc.) are defined in precisely the same way in the partial-order setting. However, the
poset A∗ is no longer totally ordered in the partial-order setting.
260This is well-defined, because there exists at least one proper suffix v of w such that v is Lyndon. (Indeed, the last letter
of w forms such a suffix, because it is a proper suffix of w (since w has length > 1) and is Lyndon (since it is a one-letter word,
and since every one-letter word is Lyndon).)
261Compare this with [97, §7.2.11, Theorem Q].
 HOPF ALGEBRAS IN COMBINATORICS 159

(a) Prove that Proposition 6.1.2 holds in the partial-order setting, as long as one replaces “a total order”
by “a partial order” in part (a) of this Proposition.
(b) Prove (in the partial-order setting) that if a, b, c, d ∈ A∗ are four words such that the words ab and
cd are comparable (with respect to the partial order ≤), then the words a and c are comparable.
(c) Prove that Proposition 6.1.4, Proposition 6.1.5, Corollary 6.1.6, Corollary 6.1.8, Exercise 6.1.9,
Exercise 6.1.10, Exercise 6.1.11, Exercise 6.1.12, Proposition 6.1.14, Corollary 6.1.15, Proposition
6.1.16, Corollary 6.1.17, Proposition 6.1.18, Theorem 6.1.20, Exercise 6.1.21(a), Exercise 6.1.23,
Exercise 6.1.24, Exercise 6.1.31(a) and Exercise 6.1.31(b) still hold in the partial-order setting.
(d) Find a counterexample to Exercise 6.1.22 in the partial-order setting.
(e) Salvage Exercise 6.1.22 in the partial-order setting (i.e., find a statement which is easily equivalent
to this exercise in the total-order setting, yet true in the partial-order setting).
(f) In the partial-order setting, a Hazewinkel-CFL factorization of a word w will mean a tuple (a1 , a2 , . . . , ak )
of Lyndon words such that w = a1 a2 · · · ak and such that no i ∈ {1, 2, . . . , k − 1} satisfies ai < ai+1 .
Prove that every word w has a unique Hazewinkel-CFL factorization (in the partial-order setting).262
(g) Prove that Exercise 6.1.32 still holds in the partial-order setting.
The reader is invited to try extending other results to the partial-order setting (it seems that no research
has been done on this except for Hazewinkel’s [77]). We shall now, however, return to the total-order setting
(which has the most known applications).
Lyndon words are related to various other objects in mathematics, such as free Lie algebras (Subsection
6.1.1 below), shuffles and shuffle algebras (Sections 6.2 and 6.3 below), QSym (Sections 6.4 and 6.5), Markov
chains on combinatorial Hopf algebras ([47]), de Bruijn sequences ([59], [140], [141], [97, §7.2.11, Algorithm
F]), symmetric functions (specifically, the transition matrices between the bases (hλ )λ∈Par , (eλ )λ∈Par and
(mλ )λ∈Par ; see [102] for this), and the Burrows-Wheeler algorithm for data compression ([42], [68], [101]).
They are also connected to necklaces (in the combinatorial sense) – a combinatorial object that also happens
to be related to a lot of algebra ([164, Chapter 5], [45]). Let us survey the basics of this latter classical
connection in an exercise:
Exercise 6.1.34. Let A be any set (not necessarily totally ordered). Let C denote the infinite cyclic group,
263
written multiplicatively. Fix a generator c of C. Fix a positive integer n. The group C acts on An
from the left according to the rule
c · (a1 , a2 , . . . , an ) = (a2 , a3 , . . . , an , a1 ) for all (a1 , a2 , . . . , an ) ∈ An .
264
The orbits of this C-action will be called n-necklaces 265; they form a set partition of the set An .
The n-necklace containing a given n-tuple w ∈ An will be denoted by [w].
(a) Prove that every n-necklace N is a finite nonempty set and satisfies |N | | n. (Recall that N is an
orbit, thus a set; as usual, |N | denotes the cardinality of this set.)
The period of an n-necklace N is defined as the positive integer |N |. (This |N | is indeed a positive integer,
since N is a finite nonempty set.)
An n-necklace is said to be aperiodic if its period is n.
(b) Given any n-tuple w = (w1 , w2 , . . . , wn ) ∈ An , prove that the n-necklace [w] is aperiodic if and only
if every k ∈ {1, 2, . . . , n − 1} satisfies (wk+1 , wk+2 , . . . , wn , w1 , w2 , . . . , wk ) 6= w.
From now on, we assume that the set A is totally ordered. We use A as our alphabet to define the notions
of words, the lexicographic order, and Lyndon words. All notations that we introduced for words will thus
be used for elements of An .
(c) Prove that every aperiodic n-necklace contains exactly one Lyndon word.
262This result, as well as the validity of Proposition 6.1.16 in the partial-order setting, are due to Hazewinkel [77].
263So C is a group isomorphic to (Z, +), and the isomorphism (Z, +) → C sends every n ∈ Z to cn . (Recall that we write
the binary operation of C as · instead of +.)
264In other words, c rotates any n-tuple of elements of A cyclically to the left. Thus, cn ∈ C acts trivially on An , and so
this action of C on An factors through C/ hcn i (a cyclic group of order n).
265Classically, one visualizes them as necklaces of n beads of |A| colors. (The colors are the elements of A.) The intuition
behind this is that a necklace is an object that doesn’t really change when we rotate it in its plane. However, to make this
intuition match the definition, we need to think of a necklace as being stuck in its (fixed) plane, so that we cannot lift it up
and turn it around, dropping it back to its plane in a reflected state.
160 DARIJ GRINBERG AND VICTOR REINER

(d) If N is an n-necklace which is not aperiodic, then prove that N contains no Lyndon word.
(e) Show that the aperiodic n-necklaces are in bijection with Lyndon words of length n.
From now on, we assume that the set A is finite. Define the number-theoretic Möbius function µ and the
Euler totient function φ as in Exercise 2.9.6.
(f) Prove that the number of all aperiodic n-necklaces is
1X n/d
µ (d) |A| .
n
d|n

(g) Prove that the number of all n-necklaces is

1X n/d
φ (d) |A| .
n
d|n

(h) Solve Exercise 6.1.29 again.

(i) Forget that we fixed A. Show that every q ∈ Z satisfies n | d|n µ (d) q n/d and n | d|n φ (d) q n/d .
P P

[Hint: For (c), use Theorem 6.1.20. For (i), either use parts (f) and (g) and a trick to extend to q
negative; or recall Exercise 2.9.8.]
6.1.1. Free Lie algebras. In this brief subsection, we shall review the connection between Lyndon words and
free Lie algebras (following [108, Kap. 4], but avoiding the generality of Hall sets in favor of just using
Lyndon words). None of this material shall be used in the rest of these notes. We will only prove some basic
results; for more thorough and comprehensive treatments of free Lie algebras, see [161], [26, Chapter 2] and
[108, Kap. 4].
We begin with some properties of Lyndon words.
Exercise 6.1.35. Let w ∈ A∗ be a nonempty word. Let v be the longest Lyndon suffix of w 266. Let t be
a Lyndon word. Then, t is the longest Lyndon suffix of wt if and only if we do not have v < t.
(We have written “we do not have v < t” instead of “v ≥ t” in Exercise 6.1.35 for reasons of generalizability:
This way, Exercise 6.1.35 generalizes to the partial-order setting introduced in Exercise 6.1.33, whereas the
version with “v ≥ t” does not.)
Exercise 6.1.36. Let w ∈ A∗ be a word of length > 1. Let v be the longest Lyndon proper suffix of w
267
. Let t be a Lyndon word. Then, t is the longest Lyndon proper suffix of wt if and only if we do not have
v < t.
(Exercise 6.1.36, while being a trivial consequence of Exercise 6.1.35, is rather useful in the study of free
Lie algebras. It generalizes both [35, Lemma (1.6)] (which is obtained by taking w = c, v = b and t = d)
and [122, Proposition 5.1.4] (which is obtained by taking v = m and t = n).)
Definition 6.1.37. For the rest of Subsection 6.1.1, we let L be the set of all Lyndon words (over the
alphabet A).
Definition 6.1.38. Let w be a Lyndon word of length > 1. Let v be the longest proper suffix of w such
that v is Lyndon. (This is well-defined, as we know from Exercise 6.1.31.) Since v is a proper suffix of w,
there exists a nonempty u ∈ A∗ such that w = uv. Consider this u. (Clearly, this u is unique.) Theorem
6.1.30(a) shows that the words u and v are Lyndon. In other words, u ∈ L and v ∈ L. Hence, (u, v) ∈ L × L.
The pair (u, v) ∈ L × L is called the standard factorization of w, and is denoted by stf w.
For the sake of easier reference, we gather a few basic properties of the standard factorization:
Exercise 6.1.39. Let w be a Lyndon word of length > 1. Let (g, h) = stf w. Prove the following:
(a) The word h is the longest Lyndon proper suffix of w.
(b) We have w = gh.
(c) We have g < gh < h.
(d) The word g is Lyndon.
266Of course, a Lyndon suffix of w just means a suffix p of w such that p is Lyndon.
267Of course, a Lyndon proper suffix of w just means a proper suffix p of w such that p is Lyndon.
 HOPF ALGEBRAS IN COMBINATORICS 161

(e) We have g ∈ L, h ∈ L, ` (g) < ` (w) and ` (h) < ` (w).

(f) Let t be a Lyndon word. Then, t is the longest Lyndon proper suffix of wt if and only if we do not
have h < t.
Exercise 6.1.40. Let g be a Lie algebra. For every Lyndon word w, let bw be an element of g. Assume
that for every Lyndon word w of length > 1, we have
(6.1.2) bw = [bu , bv ] , where (u, v) = stf w.
Let B be the k-submodule of g spanned by the family (bw )w∈L .
(a) Prove that B is a Lie subalgebra of g.
(b) Let h be a k-Lie algebra. Let f : B → h be a k-module homomorphism. Assume that whenever w
is a Lyndon word of length > 1, we have
(6.1.3) f ([bu , bv ]) = [f (bu ) , f (bv )] , where (u, v) = stf w.
Prove that f is a Lie algebra homomorphism.
[Hint: Given two words w and w0 , write w ∼ w0 if and only if w0 is a permutation of w. Part (a) follows
from the fact that for any (p, q) ∈ L × L satisfying p < q, we have [bp , bq ] ∈ Bpq,q , where Bh,s denotes the
k-linear span of {bw | w ∈ L, w ∼ h and w < s} for any two words h and s. Prove this fact by a double
induction, first inducting over ` (pq), and then (for fixed ` (pq)) inducting over the rank of q in lexicographic
order (i.e., assume that the fact is already proven for every q 0 < q instead of q). In the induction step, assume
that (p, q) 6= stf (pq) (since
 otherwisethe claim is rather obvious) and conclude that p has length > 1; thus,

set (u, v) = stf p, so that  bp , bq  = [[bu , bv ] , bq ] = [[bu , bq ] , bv ] − [[bv , bq ] , bu ], and use Exercise 6.1.36 to
 
|{z}
=[bu ,bv ]
obtain v < q.
The proof of (b) proceeds by a similar induction, piggybacking on the [bp , bq ] ∈ Bpq,q claim.]
Exercise 6.1.41. Let V be the free k-module with basis (xa )a∈A . For every word w ∈ A∗ , let xw be the
tensor xw1 ⊗ xw2 ⊗ · · · ⊗ xw`(w) . As we know from Example 1.1.2, the tensor algebra T (V ) is a free k-module
with basis (xw )w∈A∗ . We regard V as a k-submodule of T (V ).
The tensor algebra T (V ) becomes a Lie algebra via the commutator (i.e., its Lie bracket is defined by
[α, β] = αβ − βα for all α ∈ T (V ) and β ∈ T (V )).
We define a sequence (g1 , g2 , g3 , . . .) of k-submodules of T (V ) as follows: Recursively, we set g1 = V , and
for every i ∈ {2, 3, 4, . . .}, we set gi = [V, gi−1 ]. Let g be the k-submodule g1 + g2 + g3 + · · · of T (V ).
Prove the following:
(a) The k-submodule g is a Lie subalgebra of T (V ).
(b) If k is any Lie subalgebra of T (V ) satisfying V ⊂ k, then g ⊂ k.
Now, for every w ∈ L, we define an element bw of T (V ) as follows: We define bw by recursion on the
length of w. If the length of w is 1 268, then we have w = (a) for some letter a ∈ A, and we set bw = xa
for this letter a. If the length of w is > 1, then we set bw = [bu , bv ], where (u, v) = stf w 269.
Prove the following:
(c) For every w ∈ L, we have X
bw ∈ x w + kxv .
v∈A`(w) ;
v>w
(d) The family (bw )w∈L is a basis of the k-module g.
(e) Let h be any k-Lie algebra. Let ξ : A → h be any map. Then, there exists a unique Lie algebra
homomorphism Ξ : g → h such that every a ∈ A satisfies Ξ (xa ) = ξ (a).
268The length of any w ∈ L must be at least 1. (Indeed, if w ∈ L, then the word w is Lyndon and thus nonempty, and hence
its length must be at least 1.)
269This is well-defined, because b and b have already been defined. [Proof. Let (u, v) = stf w. Then, Exercise 6.1.39(e)
u v
(applied to (g, h) = (u, v)) shows that u ∈ L, v ∈ L, ` (u) < ` (w) and ` (v) < ` (w). Recall that we are defining bw by recursion
on the length of w. Hence, bp is already defined for every p ∈ L satisfying ` (p) < ` (w). Applying this to p = u, we see that bu
is already defined (since u ∈ L and ` (u) < ` (w)). The same argument (but applied to v instead of u) shows that bv is already
defined. Hence, bu and bv have already been defined. Thus, bw is well-defined by bw = [bu , bv ], qed.]
162 DARIJ GRINBERG AND VICTOR REINER

Remark 6.1.42. Let V and g be as in Exercise 6.1.41. In the language of universal algebra, the statement
of Exercise 6.1.41(e) says that g (or, to be more precise, the pair (g, f ), where f : A → g is the map sending
each a ∈ A to xa ∈ g) satisfies the universal property of the free Lie algebra on the set A. Thus, this exercise
allows us to call g the free Lie algebra on A. Most authors define the free Lie algebra differently, but all
reasonable definitions of a free Lie algebra270 lead to isomorphic Lie algebras (because the universal property
determines the free Lie algebra uniquely up to canonical isomorphism).
Notice that the Lie algebra g does not depend on the total order on the alphabet A, but the basis (bw )w∈L
constructed in Exercise 6.1.41(d) does. There is no known basis of g defined without ordering A.
It is worth noticing that our construction of g proves not only that the free Lie algebra on A exists, but also
that this free Lie algebra can be realized as a Lie subalgebra of the (associative) algebra T (V ). Therefore, if
we want to prove that a certain identity holds in every Lie algebra, we only need to check that this identity
holds in every associative algebra (if all Lie brackets are replaced by commutators); the universal property
of the free Lie algebra (i.e., Exercise 6.1.41(e)) will then ensure that this identity also holds in every Lie
algebra h.
There is much more to say about free Lie algebras than what we have said here; in particular, there are
connections to symmetric functions, necklaces, representations of symmetric groups and NSym. See [122,
§5.3], [161], [26, Chapter 2], [108, §4] and [23] for further developments271.
6.2. Shuffles and Lyndon words. We will now connect the theory of Lyndon words with the notion of
shuffle products. We have already introduced the latter notion in Definition 1.6.2, but we will now study it
more closely and introduce some more convenient notations (e.g., we will need a notation for single shuffles,
not just the whole multiset).272
Definition 6.2.1. (a) Let n ∈ N and m ∈ N. Then, Shn,m denotes the subset
σ ∈ Sn+m : σ (1) < σ −1 (2) < · · · < σ −1 (n) ; σ −1 (n + 1) < σ −1 (n + 2) < · · · < σ −1 (n + m)
−1


of the symmetric group Sn+m .

(b) Let u = (u1 , u2 , . . . , un ) and v = (v1 , v2 , . . . , vm ) be two words. If σ ∈ Shn,m , then, u v will

σ


denote the word wσ(1) , wσ(2) , . . . , wσ(n+m) , where (w1 , w2 , . . . , wn+m ) is the concatenation u · v =
(u1 , u2 , . . . , un , v1 , v2 , . . . , vm ). We notice that the multiset of all letters of u
σ


v is the disjoint
union ofthe multiset of all letters of u with the multiset of all letters of v. As a consequence,



` u v = ` (u) + ` (v).
σ
(c) Let u = (u1 , u2 , . . . , un )and v = (v1 , v2 , . . . , vm ) be
two words. The multiset of shuffles of u and v is
defined as the multiset wσ(1) , wσ(2) , . . . , wσ(n+m) : σ ∈ Shn,m multiset , where (w1 , w2 , . . . , wn+m )
is the concatenation u · v = (u1 , u2 , . . . , un , v1 , v2 , . . . , vm ). In other words, the multiset of shuffles of
u and v is the multiset


n o
u v : σ ∈ Shn,m .
σ

v.
multiset
It is denoted by u
The next fact provides the main connection between Lyndon words and shuffles:

Theorem 6.2.2. Let u and v be two words.

Let (a1 , a2 , . . . , ap ) be the CFL factorization of u. Let (b1 , b2 , . . . , bq ) be the CFL factorization of v.
(a) Let (c1 , c2 , . . . , cp+q ) be the result of sorting the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ) in decreasing or-
der273. Then, the lexicographically highest element of the multiset u
v is c1 c2 · · · cp+q (and
(c1 , c2 , . . . , cp+q ) is the CFL factorization of this element).

270Here, we call a definition “reasonable” if the “free Lie algebra” it defines satisfies the universal property.
271The claim made in [23, page 2] that “{x , . . . , x } generates freely a Lie subalgebra of A ” is essentially our Exer-
1 n R
cise 6.1.41(e).
272
Parts (a) and (c) of the below Definition 6.2.1 define notions which have already been introduced in Definition 1.6.2. Of
course, the definitions of these notions are equivalent; however, the variables are differently labelled in the two definitions (for
example, the variables u, v, w and σ of Definition 6.2.1(c) correspond to the variables a, b, c and w of Definition 1.6.2). The
labels in Definition 6.2.1 have been chosen to match with the rest of Section 6.2.
273with respect to the total order on A∗ whose greater-or-equal relation is ≥
 HOPF ALGEBRAS IN COMBINATORICS 163

(b) Let L denote the set of all Lyndon words. If w is a Lyndon word and z is any word, let multw z
denote the number of terms in the CFL factorization of z which are equal to w. The multiplicity
with which the lexicographically

 highest element of the multiset u v appears in the multiset u v

Q multw u + multw v
is w∈L . (This product is well-defined because almost all of its factors are 1.)
multw u
(c) If ai ≥ bj for every i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}, then the lexicographically highest element


of the multiset u v is uv.
(d) If ai > bj for every i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}, then the multiplicity with which the word
uv appears in the multiset u v is 1.

(e) Assume that u is a Lyndon word. Also, assume that u ≥ bj for every j ∈ {1, 2, . . . , q}. Then, the


lexicographically highest element of the multiset u v is uv, and the multiplicity with which this


word uv appears in the multiset u v is multu v + 1.

Example 6.2.3. For this example, let u and v be the words u = 23232 and v = 323221 over the alpha-
bet A = {1, 2, 3, . . .} with total order given by 1 < 2 < 3 < · · · . The CFL factorizations of u and v
are (23, 23, 2) and (3, 23, 2, 2, 1), respectively. Thus, using the notations of Theorem 6.2.2, we have p = 3,
(a1 , a2 , . . . , ap ) = (23, 23, 2), q = 5 and (b1 , b2 , . . . , bq ) = (3, 23, 2, 2, 1). Thus, Theorem 6.2.2(a) predicts that


the lexicographically highest element of the multiset u v is c1 c2 c3 c4 c5 c6 c7 c8 , where c1 , c2 , c3 , c4 , c5 , c6 , c7 , c8
are the words 23, 23, 2, 3, 23, 2, 2, 1 listed in decreasing order (in other words, (c1 , c2 , c3 , c4 , c5 , c6 , c7 , c8 ) =
(3, 23, 23, 23, 2, 2, 2, 1)). In other words, Theorem 6.2.2(a) predicts that the lexicographically highest element


of the multiset u v is 32323232221.  We could verify this by brute force, but this would be laborious



5+6
since the multiset u v has = 462 elements (with multiplicities). Theorem 6.2.2(b) predicts that
5
this lexicographically
 highest
 element 32323232221  appears in the multiset
u v with a multiplicity of
Q multw u + multw v Q multw u + multw v
w∈L . This product w∈L is infinite, but all but finitely many
multw u multw u
of its factors are 1 and therefore can be omitted; the only factors which are not 1 are those corresponding to
Lyndon words w which appear both in the CFL factorization of u and in the CFL factorization of v (since for
any other
 factor, at least one  of the numbers multw u or multw v equals 0, and therefore  the binomial coeffi-

multw u + multw v Q multw u + multw v
cient equals 1). Thus, in order to compute the product w∈L ,
multw u multw u
we only need to multiply these factors. In our example, these are the factors for w = 23 and for w = 2
(these are the only Lyndon words which appear both in the CFL factorization (23, 23, 2) of u and in the
CFL factorization (3, 23, 2, 2, 1) of v). So we have
Y multw u + multw v  mult23 u + mult23 v  mult2 u + mult2 v 
= = 3 · 3 = 9.
multw u mult23 u mult2 u
w∈L | {z }| {z }
   
2+1 1+2
= =3 = =3
2 1
The word 32323232221 must thus appear in the multiset u
v with a multiplicity of 9. This, too, could be
checked by brute force.
Theorem 6.2.2 (and Theorem 6.2.22 further below, which describes more precisely how the lexicographi-


cally highest element of u v emerges by shuffling u and v) is fairly close to [157, Theorem 2.2.2] (and will
be used for the same purposes), the main difference being that we are talking about the shuffle product of
two (not necessarily Lyndon) words, while Radford (and most other authors) study the shuffle product of
many Lyndon words.
In order to prove Theorem 6.2.2, we will need to make some stronger statements, for which we first have
to introduce some more notation:
+
Definition 6.2.4. (a) If p and q are two integers, then [p : q] denotes the interval {p + 1, p + 2, . . . , q}
+
of Z. Note that [p : q] = q − p if q ≥ p.
(b) If I and J are two nonempty intervals of Z, then we say that I < J if and only if every i ∈ I and
j ∈ J satisfy i < j. This defines a partial order on the set of nonempty intervals of Z. (Roughly
speaking, I < J if the interval I ends before J begins.)
164 DARIJ GRINBERG AND VICTOR REINER

+
(c) If w is a word with n letters (for some n ∈ N), and I is an interval of Z such that I ⊂ [0 : n] ,
+
then w [I] will denote the word (wp+1 , wp+2 , . . . , wq ), where I is written in the form I = [p : q]
+
with q ≥ p. Obviously, ` (w [I]) = |I| = q − p. A word of the form w [I] for an interval I ⊂ [0 : n]
(equivalently, a word which is a prefix of a suffix of w) is called a factor of w.
(d) Let α be a composition. Then, we define a tuple intsys α of intervals of Z as follows: Write α in the
form (α1 , α2 , . . . , α` ) (so that ` = ` (α)). Then, set intsys α = (I1 , I2 , . . . , I` ), where
" i−1 i
#+
X X
Ii = αk : αk for every i ∈ {1, 2, . . . , `} .
k=1 k=1

This `-tuple intsys α is a tuple of nonempty intervals of Z. This tuple intsys α is called the interval
system corresponding to α. (This is precisely the `-tuple (I1 , I2 , . . . , I` ) constructed in Definition
4.3.4.) The length of the tuple intsys α is ` (α).
+ +
Example 6.2.5. (a) We have [2 : 4] = {3, 4} and [3 : 3] = ∅.
+ + + + + + +
(b) We have [2 : 4] < [4 : 5] < [6 :h8] , buti we have neither [2 : 4] < [3h: 5] nori [3 : 5] < [2 : 4] .
+ +
(c) If w is the word 915352, then w [0 : 3] = (w1 , w2 , w3 ) = 915 and w [2 : 4] = (w3 , w4 ) = 53.
(d) If α is the composition (4, 1, 4, 2, 3), then the interval system corresponding to α is
 
+ + + + +
intsys α = [0 : 4] , [4 : 5] , [5 : 9] , [9 : 11] , [11 : 14]
= ({1, 2, 3, 4} , {5} , {6, 7, 8, 9} , {10, 11} , {12, 13, 14}) .
The following properties of the notions introduced in the preceding definition are easy to check:
Remark 6.2.6. (a) If I and J are two nonempty intervals of Z satisfying I < J, then I and J are disjoint.
(b) If I and J are two disjoint nonempty intervals of Z, then either I < J or J < I.
(c) Let α be a composition. Write α in the form (α1 , α2 , . . . , α` ) (so that ` = ` (α)). The interval system
intsys α can be described as the unique `-tuple (I1 , I2 , . . . , I` ) of nonempty intervals of Z satisfying
the following three properties:
+
– The intervals I1 , I2 , . . ., I` form a set partition of the set [0 : n] , where n = |α|.
– We have I1 < I2 < · · · < I` .
– We have |Ii | = αi for every i ∈ {1, 2, . . . , `}.
Exercise 6.2.7. Prove Remark 6.2.6.
The following two lemmas are collections of more or less trivial consequences of what it means to be an
element of Shn,m and what it means to be a shuffle:

Lemma 6.2.8. Let n ∈ N and m ∈ N. Let σ ∈ Shn,m .

+ + +
(a) If I is an interval of Z such that I ⊂ [0 : n + m] , then σ (I) ∩ [0 : n] and σ (I) ∩ [n : n + m] are
intervals.
+ +
(b) Let K and L be nonempty intervals of Z such that K ⊂ [0 : n] and L ⊂ [0 : n] and K < L and such
that K ∪ L is an interval. Assume that σ −1 (K) and σ −1 (L) are intervals, but σ −1 (K) ∪ σ −1 (L)
+
is not an interval. Then, there exists a nonempty interval P ⊂ [n : n + m] such that σ −1 (P ),
σ (K) ∪ σ (P ) and σ (P ) ∪ σ (L) are intervals and such that σ (K) < σ −1 (P ) < σ −1 (L).
−1 −1 −1 −1 −1
+ + +
(c) Lemma 6.2.8(b) remains valid if “K ⊂ [0 : n] and L ⊂ [0 : n] ” and “P ⊂ [n : n + m] ” are
+ + +
replaced by “K ⊂ [n : n + m] and L ⊂ [n : n + m] ” and “P ⊂ [0 : n] ”, respectively.

Exercise 6.2.9. Prove Lemma 6.2.8.

Lemma 6.2.10. Let u and v be two words. Let n = ` (u) and m = ` (v). Let σ ∈ Shn,m .
+ +
(a) If I is an interval of Z satisfying either I ⊂ [0 : n] or I ⊂ [n : n + m] , and if σ −1 (I) is an interval,
then


 
u v σ −1 (I) = (uv) [I] .

(6.2.1)
σ
 HOPF ALGEBRAS IN COMBINATORICS 165

(b) Assume that u
v is the lexicographically highest element of the multiset u
v. Let I ⊂ [0 : n]
σ
+
and
+ −1 −1
J ⊂ [n : n + m] be two nonempty intervals. Assume that σ (I) and σ (J) are also intervals,
that σ −1 (I) < σ −1 (J), and that σ −1 (I) ∪ σ −1 (J) is an interval as well. Then, (uv) [I] · (uv) [J] ≥
(uv) [J] · (uv) [I].
+ + +
(c) Lemma 6.2.10(b) remains valid if “I ⊂ [0 : n] and J ⊂ [n : n + m] ” is replaced by “I ⊂ [n : n + m]
+
and J ⊂ [0 : n] ”.
Exercise 6.2.11. Prove Lemma 6.2.10.
[Hint: For (b), show that there exists a τ ∈ Shn,m such that u
v differs from u
v only in the order
τ σ
of the subwords (uv) [I] and (uv) [J].]
We are still a few steps away from stating our results in a way that allows comfortably proving Theorem
6.2.2. For the latter aim, we introduce the notion of α-clumping permutations, and characterize them in two
ways:
Definition 6.2.12. Let n ∈ N. Let α be a composition of n. Let ` = ` (α).
→
−
(a) For every set S of positive integers, let S denote the list of all elements of S in increasing order
→
−
(with each element appearing exactly once). Notice that this list S is a word over the set of positive
integers.
(b) For every τ ∈ S` , we define a permutation iper (α, τ ) ∈ Sn as follows:
The interval system corresponding to α is an `-tuple of intervals (since ` (α) = `); denote this
`-tuple by (I1 , I2 , . . . , I` ). Now, define iper (α, τ ) to be the permutation in Sn which (in one-line
−−→−−→ −−→
notation) is the word Iτ (1) Iτ (2) · · · Iτ (`) (a concatenation of ` words). This is well-defined274; hence,
iper (α, τ ) ∈ Sn is defined.
(c) The interval system corresponding to α is an `-tuple of intervals (since ` (α) = `); denote this `-tuple
by (I1 , I2 , . . . , I` ).
A permutation σ ∈ Sn is said to be α-clumping if every i ∈ {1, 2, . . . , `} has the two properties
that:
– the set σ −1 (Ii ) is an interval;
– the restriction of the map σ −1 to the interval Ii is increasing.
Example 6.2.13. For this example, let n = 7 and α = (2, 1, 3, 1). Then, ` = ` (α) = 4 and (I1 , I2 , I3 , I4 ) =
→
− →
−
({1, 2} , {3} , {4, 5, 6} , {7}) (where we are using the notations of Definition 6.2.12). Hence, I1 = 12, I2 = 3,
→
− →
−
I3 = 456 and I4 = 7.
(a) If τ ∈ S` = S4 is the permutation (2, 3, 1, 4), then iper (α, τ ) is the permutation in S7 which (in
−−→−−→−−→−−→ → −→ −→ −→ −
one-line notation) is the word Iτ (1) Iτ (2) Iτ (3) Iτ (4) = I2 I3 I1 I4 = 3456127.
If τ ∈ S` = S4 is the permutation (3, 1, 4, 2), then iper (α, τ ) is the permutation in S7 which (in
−−→−−→−−→−−→ → −→ −→ −→ −
one-line notation) is the word Iτ (1) Iτ (2) Iτ (3) Iτ (4) = I3 I1 I4 I2 = 4561273.
(b) The permutation σ = (3, 7, 4, 5, 6, 1, 2) ∈ S7 (given here in one-line notation) is α-clumping, because:
– every i ∈ {1, 2, . . . , `} = {1, 2, 3, 4} has the property that σ −1 (Ii ) is an interval (namely,
σ −1 (I1 ) = σ −1 ({1, 2}) = {6, 7}, σ −1 (I2 ) = σ −1 ({3}) = {1}, σ −1 (I3 ) = σ −1 ({4, 5, 6}) =
{3, 4, 5} and σ −1 (I4 ) = σ −1 ({7}) = {2}), and
– the restrictions of the map σ −1 to the intervals Ii are increasing (this means that σ −1 (1) <
σ −1 (2) and σ −1 (4) < σ −1 (5) < σ −1 (6), since the one-element intervals I2 and I4 do not
contribute anything to this condition).
Here is a more or less trivial observation:
Proposition 6.2.14. Let n ∈ N. Let α be a composition of n. Let ` = ` (α). Write α in the form
(α1 , α2 , . . . , α` ). The interval system corresponding to α is an `-tuple of intervals (since ` (α) = `); denote
this `-tuple by (I1 , I2 , . . . , I` ). Let τ ∈ S` . Set σ = iper (α, τ ).
274In fact, from the properties of interval systems, we know that the intervals I , I , . . ., I form a set partition of the
1 2 `
set [0 : n]+ . Hence, the intervals Iτ (1) , Iτ (2) , . . ., Iτ (`) form a set partition of the set [0 : n]+ . As a consequence, the word
−−→−−→ −−→
Iτ (1) Iτ (2) · · · Iτ (`) is a permutation of the word 12 . . . n, and so there exists a permutation in Sn which (in one-line notation)
is this word, qed.
166 DARIJ GRINBERG AND VICTOR REINER

hPj−1 Pj i+
(a) We have σ −1 Iτ (j) = k=1 ατ (k) : k=1 ατ (k) for every j ∈ {1, 2, . . . , `}.
(b) For every j ∈ {1, 2, . . . , `}, the restriction of the map σ −1 to the interval Iτ (j) is increasing.
(c) The permutation iper (α, τ ) is α-clumping.
σ −1 Iτ (i) , σ
−1 Iτ (i+1) and σ −1 Iτ (i) ∪ σ −1 Iτ (i+1) are





(d) Let i ∈ {1, 2, . . . , ` − 1}. Then, the sets
nonempty intervals. Also, σ −1 Iτ (i) < σ −1 Iτ (i+1) .

Exercise 6.2.15. Prove Proposition 6.2.14.

Proposition 6.2.16. Let n ∈ N. Let α be a composition of n. Let ` = ` (α).

(a) Define a map
iperα : S` −→ {ω ∈ Sn | ω is α-clumping} ,
τ 7−→ iper (α, τ )
275
. This map iperα is bijective.
(b) Let σ ∈ Sn be an α-clumping permutation. Then, there exists a unique τ ∈ S` satisfying σ =
iper (α, τ ).

Exercise 6.2.17. Prove Proposition 6.2.16.

Next, we recall that the concatenation α · β of two compositions α and β is defined in the same way as
the concatenation of two words; if we regard compositions as words over the alphabet {1, 2, 3, . . .}, then the
concatenation α · β of two compositions α and β is the concatenation αβ of the words α and β. Thus, we
are going to write αβ for the concatenation α · β of two compositions α and β from now on.

Proposition 6.2.18. Let n ∈ N and m ∈ N. Let α be a composition of n, and β be a composition of m.

Let p = ` (α) and q = ` (β). Let τ ∈ Sp+q . Notice that iper (αβ, τ ) ∈ Sn+m (since αβ is a composition of
n + m having length ` (αβ) = ` (α) + ` (β) = p + q). Then, τ ∈ Shp,q if and only if iper (αβ, τ ) ∈ Shn,m .

Exercise 6.2.19. Prove Proposition 6.2.18.

Here is one more simple fact:

Lemma 6.2.20. Let u and v be two words. Let n = ` (u) and m = ` (v). Let α be a composition of n, and
let β be a composition of m. Let p = ` (α) and q = ` (β). The concatenation αβ is a composition of n + m
having length ` (αβ) = ` (α) + ` (β) = p + q. Thus, the interval system corresponding to αβ is a (p + q)-tuple
+
of intervals which covers [0 : n + m] . Denote this (p + q)-tuple by (I1 , I2 , . . . , Ip+q ).
Let τ ∈ Shp,q . Set σ = iper (αβ, τ ). Then,

    
u v = (uv) Iτ (1) · (uv) Iτ (2) · · · · · (uv) Iτ (p+q) .
σ


Exercise 6.2.21. Prove Lemma 6.2.20.

Having these notations and trivialities in place, we can say a bit more about the lexicographically highest
element of a shuffle product than what was said in Theorem 6.2.2:

Theorem 6.2.22. Let u and v be two words. Let n = ` (u) and m = ` (v).
Let (a1 , a2 , . . . , ap ) be the CFL factorization of u. Let (b1 , b2 , . . . , bq ) be the CFL factorization Ppof v.
276
Let α be the p-tuple (` (a1 ) , ` (a2 ) , . . . , ` (ap )). Then, α is a composition of length p and size k=1 ` (ak ) =

` a1 a2 · · · ap  = ` (u) = n.
| {z }
=u Pq
Let β be the q-tuple (` (b1 ) , ` (b2 ) , . . . , ` (bq )). Then, β is a composition of length q and size k=1 ` (bk ) =
277
m.
275This map is well-defined because for every τ ∈ S , the permutation iper (α, τ ) is α-clumping (according to Proposition
`
6.2.14(c)).
276since Lyndon words are nonempty, and thus ` (a ) > 0 for every i
i
277The proof of this is the same as the proof of the fact that α is a composition of length p and size Pp
k=1 ` (αk ) = n.
 HOPF ALGEBRAS IN COMBINATORICS 167

Now, α is a composition of length p and size n, and β is a composition of length q and size m. Thus, the
concatenation αβ of these two tuples is a composition of length p + q and size n + m. The interval system
corresponding to this composition αβ is a (p + q)-tuple (since said composition has length p + q); denote
this (p + q)-tuple by (I1 , I2 , . . . , Ip+q ).
     
(a) If τ ∈ Shp,q satisfies (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) , and if we set σ = iper (αβ, τ ),


then σ ∈ Shn,m , and the word u v is the lexicographically highest element of the multiset u v.



σ
(b) Let σ ∈ Shn,m be a permutation such that u v is the lexicographically highest element of the multiset


σ    
u v. Then,  there
 exists a unique permutation τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥
· · · ≥ (uv) Iτ (p+q) and σ = iper (αβ, τ ).

Proof. Before we step to the actual proof, we need to make some preparation. First of all, (I1 , I2 , . . . , Ip+q )
is the interval system corresponding to the composition αβ. In other words,
(6.2.2) (I1 , I2 , . . . , Ip+q ) = intsys (αβ) .
But since α = (` (a1 ) , ` (a2 ) , . . . , ` (ap )) and β = (` (b1 ) , ` (b2 ) , . . . , ` (bq )), we have
αβ = (` (a1 ) , ` (a2 ) , . . . , ` (ap ) , ` (b1 ) , ` (b2 ) , . . . , ` (bq )) .
Thus, (6.2.2) rewrites as
(I1 , I2 , . . . , Ip+q ) = intsys (` (a1 ) , ` (a2 ) , . . . , ` (ap ) , ` (b1 ) , ` (b2 ) , . . . , ` (bq )) .
By the definition of intsys (` (a1 ) , ` (a2 ) , . . . , ` (ap ) , ` (b1 ) , ` (b2 ) , . . . , ` (bq )), we thus have
" i−1 i
#+
X X
Ii = ` (ak ) : ` (ak ) for every i ∈ {1, 2, . . . , p} ,
k=1 k=1
and besides
" j−1 j
#+
X X
Ip+j = n + ` (bk ) : n + ` (bk ) for every j ∈ {1, 2, . . . , q}
k=1 k=1
Pp
(since k=1 ` (ak ) = n). Moreover, Remark 6.2.6(c) shows that (I1 , I2 , . . . , Ip+q ) is a (p + q)-tuple of
nonempty intervals of Z and satisfies the following three properties:
+
• The intervals I1 , I2 , . . ., Ip+q form a set partition of the set [0 : n + m] .
• We have I1 < I2 < · · · < Ip+q .
• We have |Ii | = ` (ai ) for every i ∈ {1, 2, . . . , p} and |Ip+j | = ` (bj ) for every j ∈ {1, 2, . . . , q}.
Of course, every i ∈ {1, 2, . . . , p} satisfies
+
(6.2.3) Ii ⊂ [0 : n] and (uv) [Ii ] = u [Ii ] = ai .
Meanwhile, every i ∈ {p + 1, p + 2, . . . , p + q} satisfies
+
(6.2.4) Ii ⊂ [n : n + m] and (uv) [Ii ] = v [Ii − n] = bi−p
(where Ii − n denotes the interval {k − n | k ∈ Ii }). We thus see that
(6.2.5) (uv) [Ii ] is a Lyndon word for every i ∈ {1, 2, . . . , p + q}
278
.
By the definition of a CFL factorization, we have a1 ≥ a2 ≥ · · · ≥ ap and b1 ≥ b2 ≥ · · · ≥ bq .
We have σ ∈ Shn,m , so that σ −1 (1) < σ −1 (2) < · · · < σ −1 (n) and σ −1 (n + 1) < σ −1 (n + 2) < · · · <
+
σ −1 (n + m). In other words, the restriction of the map σ −1 to the interval [0 : n] is strictly increasing,
+
and so is the restriction of the map σ −1 to the interval [n : n + m] .
(b) We will first show that
+
(6.2.6) if J ⊂ [0 : n] is an interval such that the word (uv) [J] is Lyndon, then σ −1 (J) is an interval.
Proof of (6.2.6): We will prove (6.2.6) by strong induction over |J|.

278Indeed, when i ≤ p, this follows from (6.2.3) and the fact that a is Lyndon; whereas in the other case, this follows from
i
(6.2.4) and the fact that bi−p is Lyndon.
168 DARIJ GRINBERG AND VICTOR REINER

So, fix some N ∈ N. Assume (as the induction hypothesis) that (6.2.6) has been proven whenever |J| < N .
We now need to prove (6.2.6) when |J| = N .
+
Let J ⊂ [0 : n] be an interval such that the word (uv) [J] is Lyndon and such that |J| = N . We have to
prove that σ (J) is an interval. This is obvious if |J| = 1 (because in this case, σ −1 (J) is a one-element
−1

set, thus trivially an interval). Hence, we WLOG assume that we don’t have |J| = 1. We also don’t have
|J| = 0, because (uv) [J] has to be Lyndon (and the empty word is not). So we have |J| > 1. Now,
` ((uv) [J]) = |J| > 1, and thus (uv) [J] is a Lyndon word of length > 1. Let v 0 be the (lexicographically)
smallest nonempty proper suffix of (uv) [J]. Since v 0 is a proper suffix of w, there exists a nonempty u0 ∈ A∗
such that (uv) [J] = u0 v 0 . Consider this u0 .
Now, Theorem 6.1.30(a) (applied to (uv) [J], u0 and v 0 instead of w, u and v) yields that the words u0
and v 0 are Lyndon. Also, Theorem 6.1.30(b) (applied to (uv) [J], u0 and v 0 instead of w, u and v) yields that
u0 < (uv) [J] < v 0 .
But from the fact that (uv) [J] = u0 v 0 with u0 and v 0 both being nonempty, it becomes immediately clear
that we can write J as a union of two disjoint nonempty intervals K and L such that K < L, u0 = (uv) [K]
and v 0 = (uv) [L]. Consider these K and L. The intervals K and L are nonempty and have their sizes add up
to |J| (since they are disjoint and their union is J), and hence both must have size smaller than |J| = N . So
+
K ⊂ [0 : n] is an interval of size |K| < N having the property that (uv) [K] is Lyndon (since (uv) [K] = u0
is Lyndon). Thus, we can apply (6.2.6) to K instead of J (because of the induction hypothesis). As a
result, we conclude that σ −1 (K) is an interval. Similarly, we can apply (6.2.6) to L instead of J (we know
that (uv) [L] is Lyndon since (uv) [L] = v 0 ), and learn that σ −1 (L) is an interval. The intervals σ −1 (K)
and σ −1 (L) are both nonempty (since K and L are nonempty), and their union is σ −1 (J) (because the
+
union of K and L is J). The nonempty intervals K and L both are subsets of [0 : n] (since their union is
+
J ⊂ [0 : n] ), and their union K ∪ L is an interval (since their union K ∪ L is J, and we know that J is an
interval).
Now, assume (for the sake of contradiction) that σ −1 (J) is not an interval. Since J is the union of K
and L, we have J = K ∪ L and thus σ −1 (J) = σ −1 (K ∪ L) = σ −1 (K) ∪ σ −1 (L) (since σ is a bijection).
Therefore, σ −1 (K) ∪ σ −1 (L) is not an interval (since σ −1 (J) is not an interval). Thus, Lemma 6.2.8(b)
+
yields that there exists a nonempty interval P ⊂ [n : n + m] such that σ −1 (P ), σ −1 (K) ∪ σ −1 (P ) and
σ (P ) ∪ σ (L) are intervals and such that σ (K) < σ (P ) < σ −1 (L). Consider this P . Since P is
−1 −1 −1 −1

nonempty, we have |P | = 6 0.
Lemma 6.2.10(b) (applied to K and P instead of I and J) yields
(6.2.7) (uv) [K] · (uv) [P ] ≥ (uv) [P ] · (uv) [K] .
0
Since (uv) [K] = u , this rewrites as
(6.2.8) u0 · (uv) [P ] ≥ (uv) [P ] · u0 .
But Lemma 6.2.10(c) (applied to P and L instead of I and J) yields
(6.2.9) (uv) [P ] · (uv) [L] ≥ (uv) [L] · (uv) [P ] .
0
Since (uv) [L] = v , this rewrites as
(6.2.10) (uv) [P ] · v 0 ≥ v 0 · (uv) [P ] .
Recall also that u0 < v 0 , and that both words u0 and v 0 are Lyndon. Now, Corollary 6.1.17 (applied to
u , v 0 and (uv) [P ] instead of u, v and z) yields that (uv) [P ] is the empty word (because of (6.2.8) and
0

(6.2.10)), so that ` ((uv) [P ]) = 0. This contradicts ` ((uv) [P ]) = |P | =

6 0. This contradiction shows that our
assumption (that σ −1 (J) is not an interval) was wrong. Hence, σ −1 (J) is an interval. This completes the
induction step, and thus (6.2.6) is proven.
Similarly to (6.2.6), we can show that
(6.2.11)
+
if J ⊂ [n : n + m] is an interval such that the word (uv) [J] is Lyndon, then σ −1 (J) is an interval.
Now, let i ∈ {1, 2, . . . , p + q} be arbitrary. We are going to prove that
(6.2.12) σ −1 (Ii ) is an interval.
Proof of (6.2.12): We must be in one of the following two cases:
Case 1: We have i ∈ {1, 2, . . . , p}.
 HOPF ALGEBRAS IN COMBINATORICS 169

Case 2: We have i ∈ {p + 1, p + 2, . . . , p + q}.

+
Let us first consider Case 1. In this case, we have i ∈ {1, 2, . . . , p}. Thus, Ii ⊂ [0 : n] (by (6.2.3)). Also,
(6.2.3) yields that (uv) [Ii ] = ai is a Lyndon word. Hence, (6.2.6) (applied to J = Ii ) yields that σ −1 (Ii ) is
an interval. Thus, (6.2.12) is proven in Case 1.
Similarly, we can prove (6.2.12) in Case 2, using (6.2.4) and (6.2.11) instead of (6.2.3) and (6.2.6), respec-
tively. Hence, (6.2.12) is proven.
+ +
So we know that σ −1 (Ii ) is an interval. But we also know that either Ii ⊂ [0 : n] or Ii ⊂ [n : n + m]
−1
(depending on whether i ≤ p or i > p). As a consequence, the restriction of the map σ to the interval Ii
+
is increasing (because the restriction of the map σ −1 to the interval [0 : n] is strictly increasing, and so is
+
the restriction of the map σ −1 to the interval [n : n + m] ).
Now, let us forget that we fixed i. We thus have shown that every i ∈ {1, 2, . . . , p + q} has the two
properties that:
• the set σ −1 (Ii ) is an interval;
• the restriction of the map σ −1 to the interval Ii is increasing.
In other words, the permutation σ is (αβ)-clumping (since (I1 , I2 , . . . , Ip+q ) is the interval system corre-
sponding to the composition αβ). Hence, Proposition 6.2.16(b) (applied to n + m, αβ and p + q instead of
n, α and `) shows that there exists a unique τ ∈ Sp+q satisfying σ = iper (αβ, τ ). Thus, the uniqueness part
of Theorem 6.2.22(b) (i.e., the claim that the τ in Theorem 6.2.22(b) is unique if it exists) is proven.
It now remains to prove the existence part of Theorem
 6.2.22(b),
 i.e., toprove that there exists at least one
permutation τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) and σ = iper (αβ, τ ). We
already know that there
 existsa unique
 τ ∈ Sp+q satisfying
 σ = iper (αβ, τ ). Consider this τ . We will now
prove that (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) and τ ∈ Shp,q . Once this is done, the existence
part of Theorem 6.2.22(b) will be proven, and thus the proof of Theorem 6.2.22(b) will be complete.
Proposition 6.2.18 yields that τ ∈ Shp,q if and only if iper (αβ, τ ) ∈ Shn,m . Since we know that
iper (αβ, τ ) = σ ∈ Shn,m , we thus conclude that τ ∈ Shp,q . The only thing that remains to be proven
now is that
     
(6.2.13) (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) .

Proof of (6.2.13): We have τ ∈ Shp,q . In other words, τ −1 (1) < τ −1 (2) < · · · < τ −1 (p) and τ −1 (p + 1) <
+
τ (p + 2) < · · · < τ −1 (p + q). In other words, the restriction of the map τ −1 to the interval [0 : p] is
−1
−1 +
strictly increasing, and so is the restriction of the map τ to the interval [p : p + q] .
Let i ∈ {1, 2, . . . , p + q − 1}. We will show that
   
(6.2.14) (uv) Iτ (i) ≥ (uv) Iτ (i+1) .
Clearly, both τ (i) and τ (i + 1) belong to {1, 2, . . . , p + q} = {1, 2, . . . , p}∪{p + 1, p + 2, . . . , p + q}. Thus,
we must be in one of the following four cases:
Case 1: We have τ (i) ∈ {1, 2, . . . , p} and τ (i + 1) ∈ {1, 2, . . . , p}.
Case 2: We have τ (i) ∈ {1, 2, . . . , p} and τ (i + 1) ∈ {p + 1, p + 2, . . . , p + q}.
Case 3: We have τ (i) ∈ {p + 1, p + 2, . . . , p + q} and τ (i + 1) ∈ {1, 2, . . . , p}.
Case 4: We have τ (i) ∈ {p + 1, p + 2, . . . , p + q} and τ (i + 1) ∈ {p + 1, p + 2, . . . , p + q}.
Let us consider Case 1 first. In this case, we have τ (i) ∈ {1, 2, . . . , p} and τ (i + 1) ∈ {1, 2, . . . , p}. From
+
the fact that the restriction of the map τ −1 to the interval [0 : p] is strictly increasing, we can easily deduce
279
τ (i) < τ (i + 1)  . Therefore, aτ (i) ≥ aτ (i+1) (since a1 ≥ a2 ≥ · · · ≥ ap ). 
But (uv) Iτ (i) = aτ (i) (by (6.2.3), applied to τ (i) instead of i) and (uv) Iτ (i+1) = aτ (i+1)  (similarly). In
view of these equalities, the inequality aτ (i) ≥ aτ (i+1) rewrites as (uv) Iτ (i) ≥ (uv) Iτ (i+1) . Thus, (6.2.14)
is proven in Case 1.    
Similarly, we can show (6.2.14) in Case 4 (observing that (uv) Iτ (i) = bτ (i)−p and (uv) Iτ (i+1) =
bτ (i+1)−p in this case).

279Proof. Assume the contrary. Then, τ (i) ≥ τ (i + 1). Since both τ (i) and τ (i + 1) belong to {1, 2, . . . , p} = [0 : p]+ , this
yields τ −1 (τ (i)) ≥ τ −1 (τ (i + 1)) (since the restriction of the map τ −1 to the interval [0 : p]+ is strictly increasing), which
contradicts τ −1 (τ (i)) = i < i + 1 = τ −1 (τ (i + 1)). This contradiction proves the assumption wrong, qed.
170 DARIJ GRINBERG AND VICTOR REINER

Let us now consider Case 2. In this case, we have τ (i) ∈ {1, 2, . . . , p} and τ (i + 1) ∈ {p + 1, p + 2, . . . , p + q}.
+
From τ (i) ∈ {1, 2, . . . , p}, we conclude that Iτ (i) ⊂ [0 : n] . From τ (i + 1) ∈ {p + 1, p + 2, . . . , p + q}, we
+
conclude that Iτ (i+1) ⊂ [n : n + m] . The intervals Iτ (i) and Iτ (i+1) are clearly nonempty.
Proposition 6.2.14(d) (applied
to n+m, αβ,
p+q and (I1 ,
I2 , . . . , Ip+q ) instead of n, α, ` and (I1 , I2 , . . . , I` ))
yields that the sets
σ −1 Iτ (i) , σ −1
Iτ (i+1) and σ −1 Iτ (i) ∪ σ −1 Iτ (i+1) are nonempty intervals, and that

we have σ −1 Iτ (i) < σ −1 Iτ (i+1) . Hence, Lemma 6.2.10(b) (applied to I = Iτ (i) and J = Iτ (i+1) ) yields
       
(uv) Iτ (i) · (uv) Iτ (i+1) ≥ (uv) Iτ (i+1) · (uv) Iτ (i) .
   
But (uv) Iτ (i) and  (uv)
 Iτ (i+1)  are Lyndon
 words (as a consequence of (6.2.5)).  Thus,
 Proposition
  6.1.18
(applied to (uv) Iτ (i) and (uv)  Iτ (i+1)
 instead of u
 and v) shows that (uv) Iτ (i) ≥
 (uv)  Iτ (i+1)
 if and

only if (uv) Iτ (i) ·(uv)
 I
 τ (i+1) ≥ (uv) I τ (i+1) ·(uv) I τ (i) . Since we know
 that (uv)
 Iτ (i) ·(uv) I τ (i+1) ≥
(uv) Iτ (i+1) · (uv) Iτ (i) holds, we thus conclude that (uv) Iτ (i) ≥ (uv) Iτ (i+1) . Thus, (6.2.14) is proven
in Case 2.
The proof of (6.2.14) in Case 3 is analogous to that in Case 2 (the main difference being that Lemma
6.2.10(c) is used in lieu of Lemma 6.2.10(b)).  
Thus,
 (6.2.14)
 is proven in all possible cases. So we always have (6.2.14). In other words, (uv) Iτ (i) ≥
(uv) Iτ (i+1) .    
Now, forget that we fixed i. We hence have shown that (uv) Iτ (i) ≥ (uv) Iτ (i+1) for all i ∈ {1, 2, . . . , p + q − 1}.
This proves (6.2.13), and thus completes our proof of Theorem 6.2.22(b).
(a) Let τ ∈ Shp,q be such that
     
(6.2.15) (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) .
Set σ = iper (αβ, τ ). Then, Proposition 6.2.18 yields that τ ∈ Shp,q if and only if iper (αβ, τ ) ∈ Shn,m . Since
we know that τ ∈ Shp,q , we can deduce from this that iper (αβ, τ ) ∈ Shn,m , so that σ = iper (αβ, τ ) ∈ Shn,m .


It remains to prove that the word u v is the lexicographically highest element of the multiset u v.



σ
It is clear that the multiset u v has some lexicographically highest element. This element has the form


u v for some σ
σ
e
e ∈ Shn,m (because any element of this multiset has such a form). Consider this σ e. Theorem
6.2.22(b)
 (applied
 to σ e instead of σ) yields that  there exists a unique permutation τe ∈ Shp,q satisfying
(uv) Iτe(1) ≥ (uv) Iτe(2) ≥ · · · ≥ (uv) Iτe(p+q) and σ e = iper (αβ, τe). (What we call τe here is what has been
called τ in Theorem 6.2.22(b).)      
Now, the chain of inequalities
 (1) ≥ (uv) Iτe(2) ≥ · · · ≥ (uv) Iτe(p+q) shows that the list
(uv) Iτe

(uv) Iτe(1) , (uv) Iτe(2) , . . . , (uv) Iτe(p+q) is the result of sorting the list ((uv) [I1 ] , (uv) [I2 ] , . . . , (uv) [Ip+q ])
in decreasing
  order. But the chain of inequalities
(6.2.15) shows that the list
(uv) Iτ (1) , (uv) Iτ (2) , . . . , (uv) Iτ (p+q) is theresult of sorting  the  same list ((uv) [I1
] , (uv) [I2 ] , . . . , (uv) [Ip+q ])
in decreasing
  order.
 So each
 of thetwo lists

(uv) I τe(1) , (uv) I τe(2) , . . . , (uv) I τe(p+q) and
(uv) Iτ (1) , (uv) Iτ (2) , . . . , (uv) Iτ (p+q) is the result of sorting one and the same list
((uv) [I1 ] , (uv) [I2 ] , . . . , (uv) [Ip+q ]) in decreasing order. Since the result of sorting a given list in decreasing
order is unique, this yields
     
     

(uv) Iτe(1) , (uv) Iτe(2) , . . . , (uv) Iτe(p+q) = (uv) Iτ (1) , (uv) Iτ (2) , . . . , (uv) Iτ (p+q) .
Hence,
           
(6.2.16) (uv) Iτe(1) · (uv) Iτe(2) · · · · · (uv) Iτe(p+q) = (uv) Iτ (1) · (uv) Iτ (2) · · · · · (uv) Iτ (p+q) .
But Lemma 6.2.20 yields
(6.2.17) u
v = (uv) I
σ
τ (1)
    
· (uv) Iτ (2) · · · · · (uv) Iτ (p+q) .

Meanwhile, Lemma 6.2.20 (applied to τe and σ

e instead of τ and σ) yields

    
u v = (uv) Iτe(1) · (uv) Iτe(2) · · · · · (uv) Iτe(p+q)
σe


     
= (uv) Iτ (1) · (uv) Iτ (2) · · · · · (uv) Iτ (p+q) (by (6.2.16))
=u
v
σ
(by (6.2.17)) .
 HOPF ALGEBRAS IN COMBINATORICS 171

Thus, u
v is the lexicographically highest element of the multiset u
v (since we know that u
v is the
lexicographically highest element of the multiset u
v). This proves Theorem 6.2.22(a).
σ σ
e


Now, in order to prove Theorem 6.2.2, we record a very simple fact about counting shuffles:

Proposition 6.2.23. Let p ∈ N and q ∈ N. Let W be a totally ordered set, and let h : {1, 2, . . . , p + q} → W
be a map. Assume that h (1) ≥ h (2) ≥ · · · ≥ h (p) and h (p + 1) ≥ h (p + 2) ≥ · · · ≥ h (p + q).
For every w ∈ W, let a (w) denote the number of all i ∈ {1, 2, . . . , p} satisfying h (i) = w, and let b (w)
denote the number of all i ∈ {p + 1, p + 2, . . . , p + q} satisfying h (i) = w.  
Q a (w) + b (w)
Then, the number of τ ∈ Shp,q satisfying h (τ (1)) ≥ h (τ (2)) ≥ · · · ≥ h (τ (p + q)) is w∈W .
a (w)
(Of course, all but finitely many factors of this product are 1.)

Exercise 6.2.24. Prove Proposition 6.2.23.

Proof of Theorem 6.2.2. Let n = ` (u) and m = ` (v). Define α, β and (I1 , I2 , . . . , Ip+q ) as in Theorem
6.2.22.
Since (a1 , a2 , . . . , ap ) is the CFL factorization of u, we have a1 ≥ a2 ≥ · · · ≥ ap and a1 a2 · · · ap = u.
Similarly, b1 ≥ b2 ≥ · · · ≥ bq and b1 b2 · · · bq = v.
From (6.2.3), we see that (uv) [Ii ] = ai for every i ∈ {1, 2, . . . , p}. From (6.2.4), we see that (uv) [Ii ] = bi−p
for every i ∈ {p + 1, p + 2, . . . , p + q}. Combining these two equalities, we obtain

ai , if i ≤ p;
(6.2.18) (uv) [Ii ] = for every i ∈ {1, 2, . . . , p + q} .
bi−p , if i > p

In other words,

(6.2.19) ((uv) [I1 ] , (uv) [I2 ] , . . . , (uv) [Ip+q ]) = (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ) .

(a) Let z be the lexicographically highest element of the multiset u v. We must prove that z =

c1 c2 · · · cp+q .



Since z ∈ u v, we can write z in the form u v for some σ ∈ Shn,m (since we can write any element

v in this form). Consider this σ. Then, u
v = z is the lexicographically highest element of the
σ
of u
 u
v. Hence,
σ
multiset  Theorem
 6.2.22(b)  that there exists a unique permutation τ ∈ Sh satisfying
 yields p,q
(uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) and σ = iper (αβ, τ). Consider  this τ.  

Now, τ ∈ Shp,q ⊂ Sp+q is a permutation, and thus the list (uv) Iτ (1) , (uv) Iτ (2) , . . . , (uv) Iτ (p+q) is
a rearrangement
  of the list ((uv)  [I1 ] , (uv)
 [I2 ] , 

. . . , (uv) [Ip+q ]). Due to (6.2.19), this rewrites as follows: The
list (uv) Iτ (1) , (uv)
 Iτ (2)
 , . .
 . , (uv) I τ (p+q)
 is
a rearrangement of the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ).
Hence, (uv) Iτ (1) , (uv) Iτ (2)  , . . . ,(uv) Iτ (p+q)   is the result ofsorting the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq )
in decreasing order (since (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) ). But since the result of sorting
the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ) in decreasing order is (c1 , c2 , . . . , cp+q ), this becomes
     

(uv) Iτ (1) , (uv) Iτ (2) , . . . , (uv) Iτ (p+q) = (c1 , c2 , . . . , cp+q ) .

Hence,
     
(uv) Iτ (1) · (uv) Iτ (2) · · · · · (uv) Iτ (p+q) = c1 · c2 · · · · · cp+q .
But Lemma 6.2.20 yields

u
v = (uv) I
σ
τ (1)
    
· (uv) Iτ (2) · · · · · (uv) Iτ (p+q) .

Altogether, we have

z=u
v = (uv) I
σ
τ (1)
    
· (uv) Iτ (2) · · · · · (uv) Iτ (p+q) = c1 · c2 · · · · · cp+q = c1 c2 · · · cp+q .

This proves Theorem 6.2.2(a).

172 DARIJ GRINBERG AND VICTOR REINER

v = u
v
n o
(b) Recall that u : σ ∈ Shn,m . Hence,
σ multiset

(the multiplicity with which the lexicographically highest element of the multiset
u
v appears in the multiset u
v)

v is the

= the number of all σ ∈ Shn,m such that u
σ
lexicographically highest element of the multiset u
v) .
But for a given σ ∈ Shn,m , we know that u
v is the lexicographically highest element of the multiset


σ 
u v if and only if σ can be written in the form σ = iper (αβ, τ ) for some τ ∈ Shp,q satisfying (uv) Iτ (1) ≥
(uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) . 280 Hence,



the number of all σ ∈ Shn,m such that u v is the
σ
lexicographically highest element of the multiset u
v)
= (the number of all σ ∈ Shn,m which can be written in the form σ = iper (αβ, τ )
     

for some τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q)
     

= the number of all τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q)
 
(because ifa σ ∈ Shn,m can
 be written in the form σ = iper (αβ, τ ) for some τ ∈ Shp,q satisfying (uv) Iτ (1) ≥
(uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) , then σ can be written uniquely in this form281). Thus,
 

(the multiplicity with which the lexicographically highest element of the multiset
u
v appears in the multiset u
v)

v is the

= the number of all σ ∈ Shn,m such that u
σ
lexicographically highest element of the multiset u v)
   

 

(6.2.20) = the number of all τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) .
Now, define a map h : {1, 2, . . . , p + q} → L by

ai , if i ≤ p;
h (i) = for every i ∈ {1, 2, . . . , p + q} .
bi−p , if i > p
Then, h (1) ≥ h (2) ≥ · · · ≥ h (p) (because this is just a rewriting of a1 ≥ a2 ≥ · · · ≥ ap ) and h (p + 1) ≥
h (p + 2) ≥ · · · ≥ h (p + q) (since this is just a rewriting of b1 ≥ b2 ≥ · · · ≥ bq ). For every w ∈ L, the number
of all i ∈ {1, 2, . . . , p} satisfying h (i) = w is
 

 

i ∈ {1, 2, . . . , p} | h (i) = w

 |{z} 

=ai

= |{i ∈ {1, 2, . . . , p} | ai = w}|

= (the number of terms in the list (a1 , a2 , . . . , ap ) which are equal to w)
= (the number of terms in the CFL factorization of u which are equal to w)
(since the list (a1 , a2 , . . . , ap ) is the CFL factorization of u)
= multw u

280In fact, the “if” part of this assertion follows from Theorem 6.2.22(a), whereas its “only if” part follows from Theorem
6.2.22(b).
281Proof. Let σ ∈ Sh
n,m be such that  σ can be written in the form σ = iper (αβ, τ ) for some τ ∈ Shp,q satisfying
  
(uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) . Then, the word u

v is the lexicographically highest element of the


σ  
multiset u v (according to Theorem 6.2.22(a)). Hence, there exists a unique permutation τ ∈ Shp,q satisfying (uv) Iτ (1) ≥
   
(uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) and σ = iper (αβ, τ ) (according to Theorem 6.2.22(b)). In other words, σ can be written
     
uniquely in the form σ = iper (αβ, τ ) for some τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) , qed.
 HOPF ALGEBRAS IN COMBINATORICS 173

(because multw u is defined as the number of terms in the CFL factorization of u which are equal to w).
Similarly, for every w ∈ L, the number of all i ∈ {p + 1, p + 2, . . . , p + q} satisfying h (i) = w equals
multw v. Thus, we can apply Proposition 6.2.23 to W = L, a (w) = multw u and b (w) = multw v. As
 we see that the number of τ ∈ Shp,q satisfying h (τ (1)) ≥ h (τ (2)) ≥ · · · ≥ h (τ (p + q)) is
a result,
Q multw u + multw v
w∈L . In other words,
multw u

(the number of all τ ∈ Shp,q satisfying h (τ (1)) ≥ h (τ (2)) ≥ · · · ≥ h (τ (p + q)))

Y multw u + multw v 
(6.2.21) = .
multw u
w∈L

But for every i ∈ {1, 2, . . . , p + q}, we have


ai , if i ≤ p;
h (i) = = (uv) [Ii ] (by (6.2.18)) .
bi−p , if i > p
 
Hence,
 for any
 τ ∈ Shp,q , the
 condition
 h (τ (1)) ≥ h (τ (2)) ≥ · · · ≥ h (τ (p + q)) is equivalent to (uv) Iτ (1) ≥
(uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q) . Thus,
 
 
 
the number of all τ ∈ Shp,q satisfying h (τ (1)) ≥ h (τ (2)) ≥ · · · ≥ h (τ (p + q))
 | {z }
 
this is equivalent to
(uv)[Iτ (1) ]≥(uv)[Iτ (2) ]≥···≥(uv)[Iτ (p+q) ]
     

= the number of all τ ∈ Shp,q satisfying (uv) Iτ (1) ≥ (uv) Iτ (2) ≥ · · · ≥ (uv) Iτ (p+q)
= (the multiplicity with which the lexicographically highest element of the multiset
u
v appears in the multiset u
v)
(by (6.2.20)). Compared with (6.2.21), this yields

(the multiplicity with which the lexicographically highest element of the multiset


u v appears in the multiset u
v)
Y multw u + multw v 

= .
multw u
w∈L

This proves Theorem 6.2.2(b).

(c) We shall use the notations of Theorem 6.2.2(a) and Theorem 6.2.2(b).
Assume that ai ≥ bj for every i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}. This, combined with a1 ≥ a2 ≥
· · · ≥ ap and b1 ≥ b2 ≥ · · · ≥ bq , yields that a1 ≥ a2 ≥ · · · ≥ ap ≥ b1 ≥ b2 ≥ · · · ≥ bq . Thus, the list
(a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ) is weakly decreasing. Thus, the result of sorting the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq )
in decreasing order is the list (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ) itself. But since this result is (c1 , c2 , . . . , cp+q ), this
shows that (c1 , c2 , . . . , cp+q ) = (a1 , a2 , . . . , ap , b1 , b2 , . . . , bq ). Hence, c1 c2 · · · cp+q = a1 a2 · · · ap b1 b2 · · · bq =
| {z } | {z }


=u =v
uv. Now, Theorem 6.2.2(a) yields that the lexicographically highest element of the multiset u v is
c1 c2 · · · cp+q = uv. This proves Theorem 6.2.2(c).
(d) We shall use the notations of Theorem 6.2.2(a) and Theorem 6.2.2(b).
Assume that ai > bj for every i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}. Thus, ai ≥ bj for every i ∈
{1, 2, . . . , p} and j ∈ {1, 2, . . . , q}. Hence, Theorem 6.2.2(c) yields that the lexicographically highest element


of the multiset u v is uv. Therefore, Theorem 6.2.2(b) shows
 that the multiplicity with which this word



Q multw u + multw v
uv appears in the multiset u v is w∈L .
multw u
174 DARIJ GRINBERG AND VICTOR REINER

 
multw u + multw v 282
Now, every w ∈ L satisfies = 1 . Thus, as we know, the multiplicity with
multw u


 
Q multw u + multw v Q
which this word uv appears in the multiset u v is w∈L = w∈L 1 = 1. This proves
multw u
| {z }
=1
Theorem 6.2.2(d).
(e) We shall use the notations of Theorem 6.2.2(a) and Theorem 6.2.2(b).
Since u is a Lyndon word, the 1-tuple (u) is the CFL factorization of u. Hence, we can apply Theorem
6.2.2(c) to 1 and (u) instead of p and (a1 , a2 , . . . , ap ). As a result, we conclude that the lexicographically


highest element of the multiset u v is uv. It remains to prove that the multiplicity with which this word


uv appears in the multiset u v is multu v + 1.
For every w ∈ L satisfying w 6= u, we have
(6.2.22) multw u = 0
283
. Also, multu u = 1 (for a similar reason). But uv is the lexicographically highest element of the multiset


u v. Hence, the multiplicity with which the word uv appears in the multiset u v is the multiplicity with



which the lexicographically highest element of the multiset u v appears in the multiset u v. According

to Theorem 6.2.2(b), the latter multiplicity is
Y multw u + multw v 
multw u
w∈L
  Y  
multu u + multu v multw u + multw v
= · (since u ∈ L)
multu u multw u
| {z } w∈L; | {z }
  w6=u  
1 + multu v 0 + multw v
= =
1 0
(since multu u=1) (since multw u=0 (by (6.2.22)))
  Y 0 + multw v 
1 + multu v Y
= · = (multu v + 1) · 1 = multu v + 1.
1 0
| {z } w∈L; | {z } w∈L;
w6=u w6=u
=1+multu v=multu v+1 =1 | {z }
=1
This proves Theorem 6.2.2(e). 
As an application of our preceding results, we can prove a further necessary and sufficient criterion for a
word to be Lyndon; this criterion is due to Chen/Fox/Lyndon [35, A00 = A0000 ]:
Exercise 6.2.25. Let w ∈ A∗ be a nonempty word. Prove that w is Lyndon if and only if for any two
nonempty words u ∈ A∗ and v ∈ A∗ satisfying w = uv, there exists at least one s ∈ u v satisfying s > w.

6.3. Radford’s theorem on the shuffle algebra. We recall that our goal in Chapter 6 is to exhibit an
algebraically independent generating set of the k-algebra QSym. Having the notion of Lyndon words –
which will, to some extent, but not literally, parametrize this generating set – in place, we could start the
construction of this generating set immediately. However, it might come off as rather unmotivated this way,
and so we begin with some warmups. First, we shall prove Radford’s theorem on the shuffle algebra.
Definition 6.3.1. A polynomial algebra will mean a k-algebra which is isomorphic to the polynomial ring
k [xi | i ∈ I] as a k-algebra (for some indexing set I). Note that I need not be finite.
mult u + mult v 
282Proof. Assume the contrary. Then, there exists at least one w ∈ L such that w w
6= 1. Consider this
multw u
mult u + mult v 
w w
w. Both multw u and multw v must be positive (since 6= 1). Since multw u is positive, there must be at
multw u
least one term in the CFL factorization of u which is equal to w. In other words, there is at least one i ∈ {1, 2, . . . , p} satisfying
ai = w (since (a1 , a2 , . . . , ap ) is the CFL factorization of u). Similarly, there is at least one j ∈ {1, 2, . . . , q} satisfying bj = w.
These i and j satisfy ai = w = bj , which contradicts ai > bj . This contradiction shows that our assumption was false, qed.
283Proof of (6.2.22): Let w ∈ L be such that w 6= u. Then, the number of terms in the list (u) which are equal to w is 0.
Since (u) is the CFL factorization of u, this rewrites as follows: The number of terms in the CFL factorization of u which are
equal to w is 0. In other words, multw u = 0. This proves (6.2.22).
 HOPF ALGEBRAS IN COMBINATORICS 175

Equivalently, a polynomial algebra can be defined as a k-algebra which has an algebraically independent
(over k) generating set. Yet equivalently, a polynomial algebra can be defined as a k-algebra which is
isomorphic to the symmetric algebra of a free k-module.
Keep in mind that when we say that a certain bialgebra A is a polynomial algebra, we are making no
statement about the coalgebra structure on A. The isomorphism from A to the symmetric algebra of a free
k-module need not be a coalgebra isomorphism, and the algebraically independent generating set of A need
not consist of primitives. Thus, showing that a bialgebra A is a polynomial algebra does not trivialize the
study of its bialgebraic structure.
Remark 6.3.2. Let V be a k-module, and let A be a totally ordered set. Let ba be an element of V for every
a ∈ A. Consider the shuffle algebra Sh (V ) (defined in Definition 1.6.7).
For every word w ∈ A∗ over the alphabet A, let us define an element bw of Sh (V ) by bw = bw1 bw2 · · · bw` ,
where ` is the length of w. (The multiplication used here is that of T (V ), not that of Sh (V ); the latter is


denoted by .)
Let u ∈ A∗ and v ∈ A∗ be two words over the alphabet A. Let n = ` (u) and m = ` (v). Then,

bu
v .
X
bu bv =
σ
σ∈Shn,m

Exercise 6.3.3. Prove Remark 6.3.2.

[Hint: This follows from the definition of
.]
We can now state Radford’s theorem [157, Theorem 3.1.1(e)]:
Theorem 6.3.4. Assume that Q is a subring of k. Let V be a free k-module with a basis (ba )a∈A , where
A is a totally ordered set. Then, the shuffle algebra Sh (V ) (defined in Definition 1.6.7) is a polynomial
k-algebra. An algebraically independent generating set of Sh (V ) can be constructed as follows:
For every word w ∈ A∗ over the alphabet A, let us define an element bw of Sh (V ) by bw = bw1 bw2 · · · bw` ,
where ` is the length of w. (The multiplication used here is that of T (V ), not that of Sh (V ); the latter


is denoted by .) Let L denote the set of all Lyndon words over the alphabet A. Then, (bw )w∈L is an
algebraically independent generating set of the k-algebra Sh (V ).
Example 6.3.5. For this example, let A be the alphabet {1, 2, 3, . . .} with total order given by 1 < 2 <
3 < · · · , and assume that Q is a subring of k. Let V be the free k-module with basis (ba )a∈A . We use
the notations of Theorem 6.3.4. Then, Theorem 6.3.4 yields that (bw )w∈L is an algebraically independent
generating set of the k-algebra Sh (V ). Here are some examples of elements of Sh (V ) written as polynomials
in this generating set:
b12 = b12 (the word 12 itself is Lyndon) ;


b21 = b1 b2 − b12 ;
1
b11 = b1 b1 ;
2


b123 = b123 (the word 123 itself is Lyndon) ;
b132 = b132 (the word 132 itself is Lyndon) ;


b213 = b2 b13 − b123 − b132 ;
b231 = b
b − b
b + b ;
23 1 2 13 132
b312 = b
b − b − b ;
3 12 123 132
b321 = b
b
b − b
b − b
b
1 2 3 23 1 3 12 + b123 ;
b112 = b112 (the word 112 itself is Lyndon) ;


b121 = b12 b1 − 2b112 ;
1


b1212 = b12 b12 − 2b1122 ;
2






b4321 = b1 b2 b3 b4 − b1 b2 b34 − b1 b23 b4 − b12 b3 b4




+ b
b 1


234 + b12 b34 + b123 b4 − b1234 .
176 DARIJ GRINBERG AND VICTOR REINER

284

Note that Theorem 6.3.4 cannot survive without the condition that Q be a subring of k. For instance,


for any v ∈ V , we have v v = 2vv in Sh (V ), which vanishes if 2 = 0 in k; this stands in contrast to the fact
that polynomial k-algebras are integral domains when k itself is one. We will see that QSym is less sensitive
towards the base ring in this regard (although proving that QSym is a polynomial algebra is much easier
when Q is a subring of k).
Remark 6.3.6. Theorem 6.3.4 can be contrasted with the following fact: If Q is a subring of k, then the shuffle
algebra Sh (V) of any k-module
 V (not necessarily free!) is isomorphic (as a k-algebra) to the symmetric
2
algebra Sym (ker ) / (ker ) (by Theorem 1.7.29(e), applied to A = Sh (V )). This fact is closely related
to Theorem 6.3.4, but neither follows from it (since Theorem 6.3.4 only considers the case of free k-modules
2
V ) nor yields it (since this fact does not provide explicit generators for the k-module (ker ) / (ker ) and
thus for the k-algebra Sh (V )).
In our proof of Theorem 6.3.4 (but not only there), we will use part (a) of the following lemma285, which
makes proving that certain families indexed by Lyndon words generate certain k-algebras more comfortable:

Lemma 6.3.7. Let A be a commutative k-algebra. Let A be a totally ordered set. Let L be the set of all
Lyndon words over the alphabet A. Let bw be an element of A for every w ∈ L. For every word u ∈ A∗ ,
define an element bu of A by bu = ba1 ba2 · · · bap , where (a1 , a2 , . . . , ap ) is the CFL factorization of u.
(a) The family (bw )w∈L is an algebraically independent generating set of the k-algebra A if and only if
the family (bu )u∈A∗ is a basis of the k-module A.
(b) The family (bw )w∈L generates the k-algebra A if and only if the family (bu )u∈A∗ spans the k-module
A.
(c) Assume that the k-algebra A is graded. Let wt : A → {1, 2, 3, . . .} be any map such that for every
N ∈ {1, 2, 3, . . .}, the set wt−1 (N ) is finite.
For every word w ∈ A∗ , define an element Wt (w) ∈ N by Wt (w) = wt (w1 ) + wt (w2 ) + · · · +
wt (wk ), where k is the length of w.
Assume that for every w ∈ L, the element bw of A is homogeneous of degree Wt (w).
Assume further that the k-module A has a basis (gu )u∈A∗ having the property that for every
u ∈ A∗ , the element gu of A is homogeneous of degree Wt (u).
Assume also that the family (bw )w∈L generates the k-algebra A.
Then, this family (bw )w∈L is an algebraically independent generating set of the k-algebra A.

Exercise 6.3.8. Prove Lemma 6.3.7.

[Hint: For (a) and (b), notice that the bu are the “monomials” in the bw . For (c), use Exercise 2.5.18(b)
in every homogeneous component of A.]
The main workhorse of our proof of Theorem 6.3.4 will be the following consequence of Theorem 6.2.2(c):

Proposition 6.3.9. Let V be a free k-module with a basis (ba )a∈A , where A is a totally ordered set.

284A pattern emerges in the formulas for b , b

21 321 and b4321 : for every n ∈ N, we have

· · ·
bd
X
b(n,n−1,...,1) = (−1)n−`(α) bd1 (α) bd2 (α) `(α) (α)
,
α∈Compn


where (d1 (α)) · (d2 (α)) · · · · · d`(α) (α) is the factorization of the word (1, 2, . . . , n) into factors of length α1 , α2 , . . ., α` (where
α = (α1 , α2 , . . . , α` )). This can be proved by an application of Lemma 5.2.7(a) (as it is easy to see that for any composition α
of n, we have


bd1 (α) bd2 (α)
· · ·
bd `(α) (α)

= the sum of bπ for all words π ∈ Sn satisfying Des π −1 ⊂ D (α)

X X
= bπ ,
β∈Compn ; π∈Sn ;
β coarsens α γ (π −1 )=β

where γ π −1 denotes a composition τ satisfying D (τ ) = Des π −1 ).

285And in a later proof, we will also use its part (c) (which is tailored for application to QSym).
 HOPF ALGEBRAS IN COMBINATORICS 177

For every word w ∈ A∗ over the alphabet A, let us define an element bw of Sh (V ) by bw = bw1 bw2 · · · bw` ,
where ` is the length of w. (The multiplication used here is that of T (V ), not that of Sh (V ); the latter is
denoted by .)





For every word u ∈ A∗ , define an element bu by bu = ba1 ba2 · · · bap , where (a1 , a2 , . . . , ap ) is the
CFL factorization of u.
`
If ` ∈ N and if x ∈ A` is a word, then there is a family (ηx,y )y∈A` ∈ NA of elements of N satisfying
X
bx = ηx,y by
y∈A` ;
y≤x

and ηx,x 6= 0 (in N).

Before we prove this, let us show a very simple lemma:
Lemma 6.3.10. Let A be a totally ordered set. Let n ∈ N and m ∈ N. Let σ ∈ Shn,m .
(a) If u, v and v 0 are three words satisfying ` (u) = n, ` (v) = m, ` (v 0 ) = m and v 0 < v, then u
v < u
v.
0

< u, then u
v < u
v.

σ σ
(b) If u, u0 and v are three words satisfying ` (u) = n, ` (u0 ) = n, ` (v) = m and u0 0

≤ v, then u
v ≤ u
v.
σ σ
(c) If u, v and v 0 are three words satisfying ` (u) = n, ` (v) = m, ` (v 0 ) = m and v 0 0
σ σ

Exercise 6.3.11. Prove Lemma 6.3.10.

Exercise 6.3.12. Prove Proposition 6.3.9.
[Hint: Proceed by induction over `. In the induction step, apply Theorem 6.2.2(c)286 to u = a1 and
v = a2 a3 · · · ap , where (a1 , a2 , . . . , ap ) is the CFL factorization of x. Use Lemma 6.3.10 to get rid of smaller
terms.]
Exercise 6.3.13. Prove Theorem 6.3.4.
[Hint: According to Lemma 6.3.7(a), it suffices to show that the family (bu )u∈A∗ defined in Proposition
6.3.9 is a basis of the k-module Sh (V ). When A is finite, the latter can be proven by triangularity using
Proposition 6.3.9. Reduce the general case to that of finite A.]
6.4. Polynomial freeness of QSym: statement and easy parts.
Definition 6.4.1. For the rest of Chapter 6, we introduce the following notations: We let A be the totally
ordered set {1, 2, 3, . . .} with its natural order (that is, 1 < 2 < 3 < · · · .) Thus, the words over A are
precisely the compositions. That is, A∗ = Comp. We let L denote the set of all Lyndon words over A. These
Lyndon words are also called Lyndon compositions.
A natural question is how many Lyndon compositions of a given size exist. While we will not use the
answer, we nevertheless record it:
Exercise 6.4.2. Show that the number of Lyndon compositions of size n equals
1X   1X
µ (d) 2n/d − 1 = µ (d) 2n/d − δn,1
n n
d|n d|n
P
for every positive integer n (where “ ” means a sum over all positive divisors of n, and where µ is the
d|n
number-theoretic Möbius function).
[Hint: One solution is similar to the solution of Exercise 6.1.29 using CFL factorization. Another proceeds
by defining a bijection between Lyndon compositions and Lyndon words over a two-letter alphabet {0, 1}
(with 0 < 1) which are 6= 1. 287]
286Or Theorem 6.2.2(e), if you prefer.
287This bijection is obtained by restricting the bijection

Comp → {w ∈ {0, 1}∗ | w does not start with 1} ,

(α1 , α2 , . . . , α` ) 7→ 01α1 −1 01α2 −1 · · · 01α` −1
(where 01k is to be read as 0 1k , not as (01)k ) to the set of Lyndon compositions. The idea behind this bijection is well-known

in the Grothendieck-Teichmüller community: see, e.g., [79, §3.1] (and see [64, Note 5.16] for a different appearance of this idea).
178 DARIJ GRINBERG AND VICTOR REINER

Let us now state Hazewinkel’s result ([74, Theorem 8.1], [78, §6.7]) which is the main goal of Chapter 6:

Theorem 6.4.3. The k-algebra QSym is a polynomial algebra. It is isomorphic, as a graded k-algebra, to
P`(w)
the k-algebra k [xw | w ∈ L]. Here, the grading on k [xw | w ∈ L] is defined by setting deg (xw ) = i=1 wi
for every w ∈ L.

We shall prove Theorem 6.4.3 in the next section (Section 6.5). But the particular case of Theorem 6.4.3
when Q is a subring of k can be proven more easily; we state it as a proposition:

Proposition 6.4.4. Assume that Q is a subring of k. Then, Theorem 6.4.3 holds.

We will give two proofs of Proposition 6.4.4 in this Section 6.4; a third proof of Proposition 6.4.4 will
immediately result from the proof of Theorem 6.4.3 in Section 6.5. (There is virtue in giving three different
proofs, as they all construct different isomorphisms k [xw | w ∈ L] → QSym.)
Our first proof – originating in Malvenuto’s [128, Corollaire 4.20] – can be given right away; it relies on
Exercise 5.3.12:
First proof of Proposition 6.4.4. Let V be the free k-module with basis (bn )n∈{1,2,3,...} . Endow the k-module
V with a grading by assigning to each basis vector bn the degree n. Exercise 5.3.12(k) shows that QSym
is isomorphic to the shuffle algebra Sh (V ) (defined as in Proposition 1.6.7) as Hopf algebras. By being a
bit more careful, we can obtain the slightly stronger result that QSym is isomorphic to the shuffle algebra
Sh (V ) as graded Hopf algebras288. In particular, QSym ∼ = Sh (V ) as graded k-algebras.
Theorem 6.3.4 (applied to ba = ba ) yields that the shuffle algebra Sh (V ) is a polynomial k-algebra, and
that an algebraically independent generating set of Sh (V ) can be constructed as follows:
For every word w ∈ A∗ over the alphabet A, let us define an element bw of Sh (V ) by bw = bw1 bw2 · · · bw` ,
where ` is the length of w. (The multiplication used here is that of T (V ), not that of Sh (V ); the latter is


denoted by .) Then, (bw )w∈L is an algebraically independent generating set of the k-algebra Sh (V ).
For every w ∈ A∗ , we have bw = bw1 bw2 · · · bw`(w) (by the definition of bw ). For every w ∈ A∗ , the element
P`(w) P`(w)
bw = bw1 bw2 · · · bw`(w) of Sh (V ) is homogeneous of degree i=1 deg (bwi ) = i=1 wi .
| {z }
=wi
P`(w)
Now, define a grading on the k-algebra k [xw | w ∈ L] by setting deg (xw ) = i=1 wi for every w ∈ L. By
the universal property of the polynomial algebra k [xw | w ∈ L], we can define a k-algebra homomorphism
Φ : k [xw | w ∈ L] → Sh (V ) by setting
Φ (xw ) = bw for every w ∈ L.
This homomorphism Φ is a k-algebra isomorphism (since (bw )w∈L is an algebraically independent gener-
ating set of the k-algebra Sh (V )) and is graded (because for every w ∈ L, the element bw of Sh (V ) is
P`(w)
homogeneous of degree i=1 wi = deg (xw )). Thus, Φ is an isomorphism of graded k-algebras. Hence,
Sh (V ) ∼
= k [xw | w ∈ L] as graded k-algebras. Altogether, QSym ∼ = Sh (V ) ∼
= k [xw | w ∈ L] as graded
k-algebras. Thus, QSym is a polynomial algebra. This proves Theorem 6.4.3 under the assumption that Q
be a subring of k. In other words, this proves Proposition 6.4.4. 
Our second proof of Proposition 6.4.4 comes from Hazewinkel/Gubareni/Kirichenko [78] (where Proposi-
tion 6.4.4 appears as [78, Theorem 6.5.13]). This proof will construct an explicit algebraically independent
289
family generating the k-algebra QSym. The generating set will be very unsophisticated: it will be
(Mα )α∈L , where A and L are as in Theorem 6.4.3. Here, we are using the fact that words over the alphabet
{1, 2, 3, . . .} are the same thing as compositions, so, in particular, a monomial quasisymmetric function Mα
is defined for every such word α.
288Proof. In the solution of Exercise 5.3.12(k), we have shown that QSym ∼ T (V )o as graded Hopf algebras. But Remark
=
1.6.9(b) shows that the Hopf algebra T (V )o is naturally isomorphic to the shuffle algebra Sh (V o ) as Hopf algebras; it is
easy to

o see that the natural isomorphism T (V )o → Sh (V o ) is graded (because it is the direct sum of the isomorphisms
V ⊗n → (V o )⊗n over all n ∈ N, and each of these isomorphisms is graded). Hence, T (V )o ∼ = Sh (V o ) as graded Hopf
algebras. But V o ∼ = V as graded k-modules (since V is of finite type), and thus Sh (V o ) ∼
= Sh (V ) as graded Hopf algebras.
Altogether, we obtain QSym ∼ = T (V ) ∼
o
= Sh (V o ) ∼
= Sh (V ) as graded Hopf algebras, qed.
289We could, of course, obtain such a family from our above proof as well (this is done by Malvenuto in [128, Corollaire
4.20]), but it won’t be a very simple one.
 HOPF ALGEBRAS IN COMBINATORICS 179

It takes a bit of work to show that this family indeed fits the bill. We begin with a corollary of Proposition
5.1.3 that is essentially obtained by throwing away all non-bijective maps f :
Proposition 6.4.5. Let α ∈ A∗ and β ∈ A∗ . Then,
Mα M β
X
= Mγ + (a sum of terms of the form Mδ with δ ∈ A∗ satisfying ` (δ) < ` (α) + ` (β)) .
γ∈α
β
290

Exercise 6.4.6. Prove Proposition 6.4.5.

[Hint: Recall what was said about the p = ` + m case in Example 5.1.4.]
Corollary 6.4.7. Let α ∈ A∗ and β ∈ A∗ . Then, Mα Mβ is a sum of terms of the form Mδ with δ ∈ A∗
satisfying ` (δ) ≤ ` (α) + ` (β).
Exercise 6.4.8. Prove Corollary 6.4.7.
We now define a partial order on the compositions of a given nonnegative integer:
Definition 6.4.9. Let n ∈ N. We define a binary relation ≤ on the set Compn as follows: For two
wll
compositions α and β in Compn , we set α ≤ β if and only if
wll
either ` (α) < ` (β) or (` (α) = ` (β) and α ≤ β in lexicographic order) .
This binary relation ≤ is the smaller-or-equal relation of a total order on Compn ; we refer to said total order
wll
as the wll-order on Compn , and we denote by < the smaller relation of this total order.
wll

Notice that if α and β are two compositions satisfying ` (α) = ` (β), then α ≤ β in lexicographic order if
and only if α ≤ β with respect to the relation ≤ defined in Definition 6.1.1.
A remark about the name “wll-order” is in order. We have taken this notation from [74, Definition 6.7.14],
where it is used for an extension of this order to the whole set Comp. We will never use this extension, as
we will only ever compare two compositions of the same integer.291
We now state a fact which is similar (and plays a similar role) to Proposition 6.3.9:
Proposition 6.4.10. For every composition u ∈ Comp = A∗ , define an element Mu ∈ QSym by Mu =
Ma1 Ma2 · · · Map , where (a1 , a2 , . . . , ap ) is the CFL factorization of the word u.
If n ∈ N and if x ∈ Compn , then there is a family (ηx,y )y∈Comp ∈ NCompn of elements of N satisfying
n
X
Mx = ηx,y My
y∈Compn ;
y≤x
wll

and ηx,x 6= 0 (in N).

Before we prove it, let us show the following lemma:
Lemma 6.4.11. Let n ∈ N and m ∈ N. Let u ∈ Compn and v ∈ Compm . Let z be the lexicographically
highest element of the multiset u v.

(a) We have z ∈ Compn+m .
(b) There exists a positive integer h such that
 
Mu Mv = hMz + a sum of terms of the form Mw with w ∈ Compn+m satisfying w < z .
wll

290The sum P


γ∈α
β Mγ ranges over the multiset α β; if an element appears several times in α β, then it has accordingly
many addends corresponding to it.
291In [74, Definition 6.7.14], the name “wll-order” is introduced as an abbreviation for “weight first, then length, then
lexicographic” (in the sense that two compositions are first compared by their weights, then, if the weights are equal, by their
lengths, and finally, if the lengths are also equal, by the lexicographic order). For us, the alternative explanation “word length,
then lexicographic” serves just as well.
180 DARIJ GRINBERG AND VICTOR REINER

(c) Let v 0 ∈ Compm be such that v 0 < v. Then,

wll
 
Mu Mv0 = a sum of terms of the form Mw with w ∈ Compn+m satisfying w < z .
wll

Exercise 6.4.12. Prove Lemma 6.4.11.

[Hint: For (b), set h to be the multiplicity with which the word z appears in the multiset u v, then use

Proposition 6.4.5 and notice that Mu Mv is homogeneous of degree n + m. For (c), use (b) for v 0 instead of
v and notice that Lemma 6.3.10(a) shows that the lexicographically highest element of the multiset u v 0

is < z.]
wll

Exercise 6.4.13. Prove Proposition 6.4.10.

[Hint: Proceed by strong induction over n. In the induction step, let (a1 , a2 , . . . , ap ) be the CFL factor-
Pv = a2 a3 · · · ap ; then apply Proposition 6.4.10 to v instead of x, and multiply
ization of x, and set u = a1 and
the resulting equality Mv = y∈Comp|v| ; ηv,y My with Mu to obtain an expression for Mu Mv = Mx . Use
y≤v
wll P
Lemma 6.4.11 to show that this expression has the form y∈Compn ; ηx,y My with ηx,x 6= 0; here it helps to
y≤x

v is uv = x (by Theorem 6.2.2(c)).]

wll
remember that the lexicographically highest element of the multiset u
We are almost ready to give our second proof of Proposition 6.4.4; our last step is the following proposition:

Proposition 6.4.14. Assume that Q is a subring of k. Then, (Mw )w∈L is an algebraically independent
generating set of the k-algebra QSym.

Exercise 6.4.15. Prove Proposition 6.4.14.

[Hint: Define Mu for every u ∈ Comp as in Proposition 6.4.10. Conclude from Proposition 6.4.10 that,
for every n ∈ N, the family (Mu )u∈Compn expands invertibly triangularly292 (with respect to the total order
≤ on Compn ) with respect to the basis (Mu )u∈Compn of QSymn . Conclude that this family (Mu )u∈Compn
wll
is a basis of QSymn itself, and so the whole family (Mu )u∈Comp is a basis of QSym. Conclude using Lemma
6.3.7(a).]
Second proof of Proposition 6.4.4. Proposition 6.4.14 yields that (Mw )w∈L is an algebraically independent
generating set of the k-algebra QSym.
P`(w)
Define a grading on the k-algebra k [xw | w ∈ L] by setting deg (xw ) = i=1 wi for every w ∈ L. By
the universal property of the polynomial algebra k [xw | w ∈ L], we can define a k-algebra homomorphism
Φ : k [xw | w ∈ L] → QSym by setting
Φ (xw ) = Mw for every w ∈ L.
This homomorphism Φ is a k-algebra isomorphism (since (Mw )w∈L is an algebraically independent generating
set of the k-algebra QSym) and is graded (because for every w ∈ L, the element Mw of QSym is homogeneous
of degree |w| = i=1 wi = deg (xw )). Thus, Φ is an isomorphism of graded k-algebras. Hence, QSym ∼
P`(w)
=
k [xw | w ∈ L] as graded k-algebras. In particular, this shows that QSym is a polynomial algebra. This
proves Theorem 6.4.3 under the assumption that Q be a subring of k. Proposition 6.4.4 is thus proven
again. 
6.5. Polynomial freeness of QSym: the general case. We now will prepare for proving Theorem 6.4.3
without any assumptions on k. In our proof, we follow [74] and [78, §6.7], but without using the language
of plethysm and Frobenius maps. We start with the following definition:
Definition 6.5.1. Let α be a composition. Write α in the form α = (α1 , α2 , . . . , α` ) with ` = ` (α).
(a) Let SIS (`) denote the set of all strictly increasing `-tuples (i1 , i2 , . . . , i` ) of positive integers.293 For
α`
every `-tuple i = (i1 , i2 , . . . , i` ) ∈ SIS (`), we denote the monomial xα 1 α2 α
i1 xi2 · · · xi` by xi . This xi is a
α

292See Definition 11.1.16(b) for the meaning of this.

293“Strictly increasing” means that i < i < · · · < i here. Of course, the elements of SIS (`) are in 1-to-1 correspondence
1 2 `
with `-element subsets of {1, 2, 3, . . .}.
 HOPF ALGEBRAS IN COMBINATORICS 181

monomial of degree α1 + α2 + · · · + α` = |α|. Then,

X
(6.5.1) Mα = xα
i .
i∈SIS(`)
294

(b) Consider the ring k [[x]] endowed with the coefficientwise topology295. The family (xα
i )i∈SIS(`) of
296
elements
 [[x]] is power-summable . Hence, for every f ∈ Λ, there is a well-defined power series
of k 
297
f (xα
i )i∈SIS(`) ∈ k [[x]] obtained by “evaluating” f at (xα
i )i∈SIS(`) . In particular, for every s ∈

294Proof of (6.5.1): By the definition of M , we have

α
α α α
xα 1 α2
xα 1 α2
xα 1 α2
X X X
` ` `
Mα = i 1 xi 2 · · · xi =` i1 xi2 · · · xi =
` i1 xi2 · · · xi`
i1 <i2 <···<i` in {1,2,3,...} (i1 ,i2 ,...,i` )∈SIS(`) i=(i1 ,i2 ,...,i` )∈SIS(`) | {z }
| {z } =xαi
(by the definition of xα
i )
P
= (i ,i ,...,i )∈SIS(`)
1 2 `
X X
= xα
i = xα
i ,
i=(i1 ,i2 ,...,i` )∈SIS(`) i∈SIS(`)

qed.
295This topology is defined as follows:
We endow the ring k with the discrete topology. Then, we can regard the k-module k [[x]] as a direct product of infinitely
many copies of k (by identifying every power series in k [[x]] with the family of its coefficients). Hence, the product topology is
a well-defined topology on k [[x]]; this topology is denoted as the coefficientwise topology. A sequence (an )n∈N of power series
converges to a power series a with respect to this topology if and only if for every monomial m, all sufficiently high n ∈ N satisfy
(the coefficient of m in an ) = (the coefficient of m in a) .
Note that this is not the topology obtained by taking the completion of k [x1 , x2 , x3 , . . .] with respect to the standard grading
(in which all xi have degree 1). (The latter completion is actually a smaller ring than k [[x]].)
296Let us define what “power-summable” means for us:
A family (ni )i∈I ∈ NI (where I is some set) is said to be finitely supported if all but finitely many i ∈ I satisfy ni = 0.
If (ni )i∈I ∈ NI is a finitely supported family, then i∈I ni is a well-defined element of N. If N ∈ N, then a family (ni )i∈I ∈ NI
P
P
will be called (≤ N )-supported if it is finitely supported and satisfies i∈I ni ≤ N .
I
We say that a family (si )i∈I ∈ R of elements of a topological commutative k-algebra R is power-summable if it satisfies the
following property: For every N ∈ N, the sum
X Y n
α(ni )i∈I si i
(ni )i∈I ∈NI ; i∈I
(ni )i∈I is (≤N )-supported

converges in the topology on R for every choice of scalars α(ni )i∈I ∈ k corresponding to all (≤ N )-supported (ni )i∈I ∈ NI . In
our specific case, we consider
k [[x]] as a topological commutative k-algebra, where the topology is the coefficientwise topology.
The fact that the family xα i i∈SIS(`) is power-summable then can be proven as follows:
• If α 6= ∅, then this fact follows from the (easily-verified) observation that every given monomial in the variables
x1 , x2 , x3 , . . . can be written as a product of monomials of the form xα i (with i ∈ SIS (`)) in only finitely many ways.
• If α = ∅, then this fact follows by noticing that xα


i i∈SIS(`) is a finite family (indeed, SIS (`) = SIS (0) = {()}), and
every finite family is power-summable.
 
297Here is how this power series f xα


is formally defined:
i i∈SIS(`)
Let R be any topological commutative k-algebra, and let (si )i∈I ∈ RI be any power-summable family of elements of R.
Assume that the indexing set I is countably infinite, and fix a bijection j : {1, 2, 3, . . .} → I. Let g ∈ R (x) be arbitrary. Then,
we can substitute sj(1) , sj(2) , sj(3) , . . . for the variables x1 , x2 , x3 , . . . in g, thus obtaining an infinite sum which converges in R
(in fact, its convergence follows from the fact that the family (si )i∈I ∈ RI is power-summable). The value of this sum will be



denoted by g (si )i∈I . In general, this value depends on the choice of the bijection j, so the notation g (si )i∈I is unambiguous
only if this
bijection j is chosen once and for all. However, when g ∈ Λ, one can easily see that the choice of j has no effect on
g (si )i∈I .


We can still define g (si )i∈I when the set I is finite instead of being countably infinite. In this case, we only need to modify
our above definition as follows: Instead of fixing a bijection j : {1, 2, 3, . . .} → I, we now fix a bijection j : {1, 2, . . . , |I|} → I, and
instead of substituting sj(1) , sj(2) , sj(3) , . . . for the variables x1 , x2 , x3 , . . . in g, we now substitute sj(1) , sj(2) , . . ., sj(|I|) , 0, 0,


0, . . . for the variables x1 , x2 , x3 , . . . in g. Again, the same observations hold as before: g (si )i∈I is independent on j if g ∈ Λ.


Hence, g (si )i∈I is well-defined for every g ∈ R (x), every countable (i.e., finite or countably infinite) set I, every topological
commutative k-algebra R and every power-summable family (si )i∈I ∈ RI of elements of R, as long as a bijection j is chosen. In
particular, we can apply this to g = f , I = SIS (`), R = k [[x]] and (si )i∈I = xα


i i∈SIS(`) , choosing j to be the bijection which
182 DARIJ GRINBERG AND VICTOR REINER

298
Z, we can evaluate the symmetric function es ∈ Λ at (xα
i )i∈SIS(`) . The resulting power series
 
α hsi
es (xi )i∈SIS(`) ∈ k [[x]] will be denoted Mα . Thus,
 
Mαhsi = es (xα
i i∈SIS(`) .
)
hsi
The power series Mα are the power series es (α) in [78]. We will shortly (in Corollary 6.5.8(a)) see that
hsi hsi
Mα ∈ QSym (although this is also easy to prove by inspection). Here are some examples of Mα :
Example 6.5.2. If α is a composition and ` denotes its length ` (α), then
   
Mαh0i = e0 (xα )
i i∈SIS(`) = 1 (x α
i i∈SIS(`) = 1
)
|{z}
=1
and   X
Mαh1i = e1 (xα
i i∈SIS(`) =
) xα
i = Mα (by (6.5.1))
i∈SIS(`)

and299   X
Mαh2i = e2 (xα
i )i∈SIS(`) = xα α
i xj
i∈SIS(`), j∈SIS(`);
i<j
(where the notation “i < j” should be interpreted with respect to an arbitrary but fixed total order on the
set SIS (`) – for example, the lexicographic order). Applying the last of these three equalities to α = (2, 1),
we obtain
h2i (2,1) (2,1) (2,1) (2,1)
X X
M(2,1) = xi xj = x(i1 ,i2 ) x(j1 ,j2 )
i∈SIS(2), j∈SIS(2), (i1 ,i2 )∈SIS(2), (j1 ,j2 )∈SIS(2); | {z } | {z }
i<j (i1 ,i2 )<(j1 ,j2 ) =x2i x1i =x2j x1j
1 2 1 2
X
= x2i1 x1i2 x2j1 x1j2
(i1 ,i2 )∈SIS(2), (j1 ,j2 )∈SIS(2);
(i1 ,i2 )<(j1 ,j2 )
X X
= x2i1 x1i2 x2j1 x1j2 + x2i1 x1i2 x2j1 x1j2
i1 <i2 ; j1 <j2 ; i1 <i2 ; j1 <j2 ;
i1 <j1 i1 =j1 , i2 <j2
| {z } | {z }
=M(2,1,2,1) +M(2,3,1) +2M(2,2,1,1) +M(2,2,2) =M(4,1,1)

(here, we have WLOG assumed that the order on SIS (2) is lexicographic)
= M(2,1,2,1) + M(2,3,1) + 2M(2,2,1,1) + M(2,2,2) + M(4,1,1) .
 
hsi
Of course, every negative integer s satisfies Mα = es (xα i )i∈SIS(`) = 0.
|{z}
=0
hsi hsi
There is a determinantal formula for the s!Mα (and thus also for Mα when s! is invertible in k), but
in order to state it, we need to introduce one more notation:
Definition 6.5.3. Let α = (α1 , α2 , . . . , α` ) be a composition, and let k be a positive integer. Then, α {k}
will denote the composition (kα1 , kα2 , . . . , kα` ). Clearly, ` (α {k}) = ` (α) and |α {k}| = k |α|.
Exercise 6.5.4. Let α be a composition. Write the composition α in the form α = (α1 , α2 , . . . , α` ) with
` = ` (α).
(a) Show that the s-th power-sum symmetric function ps ∈ Λ satisfies
 
ps (xα
i i∈SIS(`) = Mα{s}
)
for every positive integer s.

sends every positive integer k to the k-th smallest element of SIS (`) in the lexicographic order. (Of course, since f ∈ Λ, the
choice of j is irrelevant.)
298Recall that e = 1, and that e = 0 for s < 0.
0 s
299This is not completely obvious, but easy to check (see Exercise 6.5.4(b)).
 HOPF ALGEBRAS IN COMBINATORICS 183

(b) Let us fix a total order on the set SIS (`) (for example, the lexicographic order). Show that the s-th
elementary symmetric function es ∈ Λ satisfies
  X
Mαhsi = es (xαi )i∈SIS(`) = xα α α
i1 xi2 · · · xis
(i1 ,i2 ,...,is )∈(SIS(`))s ;
i1 <i2 <···<is

for every s ∈ N.
hni
(c) Let s ∈ N, and let n be a positive integer. Let es be the symmetric function i1 <i2 <···<is xni1 xni2 · · · xnis ∈
P
Λ. Then, show that
 
hsi
Mα{n} = ehnis (xαi )i∈SIS(`) .

(d) Let s ∈ N, and let n be

 a positive integer. Prove
 that there exists a polynomial P ∈ k [z1 , z2 , z3 , . . .]
hsi h1i h2i h3i
such that Mα{n} = P Mα , Mα , Mα , . . . .
 
300
[Hint: For (a), (b) and (c), apply the definition of f (xαi )i∈SIS(`) with f a symmetric function . For
(d), recall that Λ is generated by e1 , e2 , e3 , . . ..]
hsi
Exercise 6.5.5. Let s ∈ N. Show that the composition (1) satisfies M(1) = es .

Proposition 6.5.6. Let α = (α1 , α2 , . . . , 

α` ) be a composition.
hαi hαi
(a) Let n ∈ N. Define a matrix An = ai,j by
i,j=1,2,...,n

Mα{i−j+1} , if i ≥ j;

hαi 2
ai,j = i, if i = j − 1; for all (i, j) ∈ {1, 2, . . . , n} .

0, if i < j − 1


hαi
This matrix An looks as follows:
 
Mα{1} 1 0 ··· 0 0
 Mα{2}
 Mα{1} 2 ··· 0 0 

 Mα{3} Mα{2} Mα{1} ··· 0 0 
Ahαi
n = .
 
.. .. .. .. .. ..

 . . . . . .

 Mα{n−1} Mα{n−2} Mα{n−3} ··· Mα{1} n−1 
Mα{n} Mα{n−1} Mα{n−2} ··· Mα{2} Mα{1}
 
hαi hni
Then, det An = n!Mα .
 
hαi hαi
(b) Let n be a positive integer. Define a matrix Bn = bi,j by
i,j=1,2,...,n
(
hii
hαi iMα , if j = 1; 2
bi,j = hi−j+1i for all (i, j) ∈ {1, 2, . . . , n} .
Mα , if j > 1

300There are two subtleties that need to be addressed:

 
• the fact that the definition of f xα


i i∈SIS(`) distinguishes between two cases depending on whether or not SIS (`)
is finite;
• the fact
P that the total order on the set {1, 2, 3, . . .} (which appears in the summation subscript in the equality
es = (i1 ,i2 ,...,is )∈{1,2,3,...}s ; xi1 xi2 · · · xis ) has nothing to do with the total order on the set SIS (`) (which appears
i1 <i2 <···<is
α α α
P
in the summation subscript in (i1 ,i2 ,...,is )∈(SIS(`))s ; xi1 xi2 · · · xis ). For instance, the former total order is well-
i1 <i2 <···<is
founded, whereas the latter may and may not be. So there is (generally) no bijection between {1, 2, 3, . . .} and SIS (`)
preserving these orders (even if SIS (`) is infinite). Fortunately, this does not matter much, because the total order is
only being used to ensure that every product of s distinct elements appears exactly once in the sum.
184 DARIJ GRINBERG AND VICTOR REINER

hαi
The matrix Bn looks as follows:
 h1i h0i h−1i h−n+3i h−n+2i 
Mα Mα Mα ··· Mα Mα
h2i h1i h0i h−n+4i h−n+3i

 2Mα Mα Mα ··· Mα Mα 

h3i h2i h1i h−n+5i h−n+4i
3Mα Mα Mα ··· Mα Mα
 
Bnhαi = 
 
.. .. .. .. .. .. 
.
 
 . . . . . 
 (n − 1) Mαhn−1i hn−2i hn−3i h1i h0i
 
Mα Mα ··· Mα Mα 
hni hn−1i hn−2i h2i h1i
nMα Mα Mα ··· Mα Mα
 h1i 
Mα 1 0 ··· 0 0
h2i h1i

 2Mα Mα 1 ··· 0 0 

h3i h2i h1i
3Mα Mα Mα ··· 0 0
 
 
= .. .. .. .. .. .. .
.
 
 . . . . . 
 (n − 1) Mαhn−1i hn−2i hn−3i h1i
 
Mα Mα ··· Mα 1 
hni hn−1i hn−2i h2i h1i
nMα Mα Mα ··· Mα Mα
 
hαi
Then, det Bn = Mα{n} .

Exercise 6.5.7. Prove Proposition 6.5.6.

[Hint: Substitute (xα
i )i∈SIS(`) for the variable set in Exercise 2.9.13, and recall Exercise 6.5.4(a).]

Corollary 6.5.8. Let α be a composition. Let s ∈ Z.

hsi
(a) We have Mα ∈ QSym.
hsi
(b) We have Mα ∈ QSyms|α| .

Exercise 6.5.9. Prove Corollary 6.5.8.

We make one further definition:
Definition 6.5.10. Let α be a nonempty composition. Then, we denote by gcd α the greatest common
divisor of the parts of α. (For instance, gcd (8, 6, 4) = 2.) We also define red α to be the composition

α1 α2 α`
, ,..., , where α is written in the form (α1 , α2 , . . . , α` ).
gcd α gcd α gcd α
We say that a nonempty composition α is reduced if gcd α = 1.
We define RL to be the set of all reduced Lyndon compositions. In other words, RL = {w ∈ L | w is reduced}
(since L is the set of all Lyndon compositions).
Hazewinkel, in [78, proof of Thm. 6.7.5], denotes RL by eLY N , calling reduced Lyndon compositions
“elementary Lyndon words”.
Remark 6.5.11. Let α be a nonempty composition.
(a) We have α = (red α) {gcd α}.
(b) The composition α is Lyndon if and only if the composition red α is Lyndon.
(c) The composition red α is reduced.
(d) If α is reduced, then red α = α.
(e) If s ∈ {1, 2, 3, . . .}, then the composition α {s} is nonempty and satisfies red (α {s}) = red α and
gcd (α {s}) = s gcd α.
(f) We have (gcd α) |red α| = |α|.
Exercise 6.5.12. Prove Remark 6.5.11.
Our goal in this section is now to prove the following result of Hazewinkel:
 
hsi
Theorem 6.5.13. The family Mw is an algebraically independent generating set of
(w,s)∈RL×{1,2,3,...}
the k-algebra QSym.
 HOPF ALGEBRAS IN COMBINATORICS 185

This will (almost) immediately yield Theorem 6.4.3.

Our first step towards proving Theorem 6.5.13 is the following observation:
   
hsi hgcd αi
Lemma 6.5.14. The family Mw is a reindexing of the family Mred α .
(w,s)∈RL×{1,2,3,...} α∈L

Exercise 6.5.15. Prove Lemma 6.5.14.

Next, we show a lemma:

Lemma 6.5.16. Let α be a nonempty composition. Let s ∈ N. Then,

X
(6.5.2) s!Mαhsi − Mαs ∈ kMβ .
β∈Comps|α| ;
`(β)≤(s−1)`(α)

hsi
(That is, s!Mα −Mαs is a k-linear combination of terms of the form Mβ with β ranging over the compositions
of s |α| satisfying ` (β) ≤ (s − 1) ` (α).)

Exercise 6.5.17. Prove Lemma 6.5.16.

[Hint: There are two approaches: One is to apply Proposition 6.5.6(a) and expand the determinant; the
hsi
other is to argue which monomials can appear in s!Mα − Mαs .]
We now return to studying products of monomial quasisymmetric functions:

Lemma 6.5.18. Let n ∈ N and m ∈ N. Let u ∈ Compn and v ∈ Compm . Let z be the lexicographically


highest element of the multiset u v. Let h be the multiplicity with which the word z appears in the multiset


u v. Then,301
 
Mu Mv = hMz + a sum of terms of the form Mw with w ∈ Compn+m satisfying w < z .
wll

Proof of Lemma 6.5.18. Lemma 6.5.18 was shown during the proof of Lemma 6.4.11(b). 

Corollary 6.5.19. Let n ∈ N and m ∈ N. Let u ∈ Compn and v ∈ Compm . Regard u and v as words in
A∗ . Assume that u is a Lyndon word. Let (b1 , b2 , . . . , bq ) be the CFL factorization of the word v.
Assume that u ≥ bj for every j ∈ {1, 2, . . . , q}. Let
h = 1 + |{j ∈ {1, 2, . . . , q} | bj = u}| .
Then,
 
Mu Mv = hMuv + a sum of terms of the form Mw with w ∈ Compn+m satisfying w < uv .
wll

Exercise 6.5.20. Prove Corollary 6.5.19.

[Hint: Apply Lemma 6.5.18, and notice that uv is the lexicographically highest element of the multiset


u v (by Theorem 6.2.2(e)), and that h is the multiplicity with which this word uv appears in the multiset


u v (this is a rewriting of Theorem 6.2.2(e)).]

Corollary 6.5.21. Let k ∈ N and s ∈ N. Let x ∈ Compk be such that x is a Lyndon word. Then:


(a) The lexicographically highest element of the multiset x xs is xs+1 .
(b) We have
 
s+1
Mx Mxs = (s + 1) Mxs+1 + a sum of terms of the form Mw with w ∈ Comp(s+1)k satisfying w < x .
wll
s
(c) Let t ∈ Compsk be such that t < x . Then,
wll
 
s+1
Mx Mt = a sum of terms of the form Mw with w ∈ Comp(s+1)k satisfying w < x .
wll

301The following equality makes sense because we have z ∈ Comp

n+m (by Lemma 6.4.11(a)).
186 DARIJ GRINBERG AND VICTOR REINER

Exercise 6.5.22. Prove

 Corollary 
6.5.21.

[Hint: Notice that x, x, . . . , x is the CFL factorization of the word xs . Now, part (a) of Corollary
| {z }
s times
6.5.21 follows from Theorem 6.2.2(c), part (b) follows from Corollary 6.5.19, and part (c) from Lemma
6.4.11(c) (using part (a)).]

Corollary 6.5.23. Let n ∈ N and m ∈ N. Let u ∈ Compn and v ∈ Compm . Regard u and v as words in
A∗ . Let (a1 , a2 , . . . , ap ) be the CFL factorization of u. Let (b1 , b2 , . . . , bq ) be the CFL factorization of the
word v. Assume that ai > bj for every i ∈ {1, 2, . . . , p} and j ∈ {1, 2, . . . , q}. Then,
 
Mu Mv = Muv + a sum of terms of the form Mw with w ∈ Compn+m satisfying w < uv .
wll

Exercise 6.5.24. Prove Corollary 6.5.23.

[Hint: Combine Lemma 6.5.18 with the parts (c) and (d) of Theorem 6.2.2.]

Corollary 6.5.25. Let n ∈ N. Let u ∈ Compn be a nonempty composition. Regard u as a word in A∗ . Let
(a1 , a2 , . . . , ap ) be the CFL factorization of u. Let k ∈ {1, 2, . . . , p − 1} be such that ak > ak+1 . Let x be
the word a1 a2 · · · ak , and let y be the word ak+1 ak+2 · · · ap . Then,
 
Mu = Mx My − a sum of terms of the form Mw with w ∈ Compn satisfying w < u .
wll

Exercise 6.5.26. Prove Corollary 6.5.25.

[Hint: Apply Corollary 6.5.23 to x, y, |x|, |y|, k, p − k, (a1 , a2 , . . . , ak ) and (ak+1 , ak+2 , . . . , ap ) instead
of u, v, n, m, p, q, (a1 , a2 , . . . , ap ) and (b1 , b2 , . . . , bq ); then, notice that xy = u and |x| + |y| = n.]

Corollary 6.5.27. Let k ∈ N. Let x ∈ Compk be a composition. Assume that x is a Lyndon word. Let
s ∈ N. Then,
X
Mxs − s!Mxs ∈ kMw .
w∈Compsk ;
w < xs
wll

s
· · · x}.)
(Recall that x is defined to be the word |xx{z
s times

Exercise 6.5.28. Prove Corollary 6.5.27.

[Hint: Rewrite the claim of Corollary 6.5.27 in the form Mxs ∈ s!Mxs + w∈Compsk ; kMw . This can be
P
w < xs
wll
proven by induction over s, where in the induction step we need the following two observations:
P
(1) We have Mx Mxs ∈ (s + 1) Mxs+1 + w∈Comp(s+1)k ; kMw .
w < xs+1
wll
(2) For every t ∈ Compsk satisfying t < xs , we have Mx Mt ∈
P
w∈Comp(s+1)k ; kMw .
wll
w < xs+1
wll

These two observations follow from parts (b) and (c) of Corollary 6.5.21.]

Corollary 6.5.29. Let k ∈ N. Let x ∈ Compk be a composition. Assume that x is a Lyndon word. Let
s ∈ N. Then,
X
Mxhsi − Mxs ∈ kMw .
w∈Compsk ;
w < xs
wll

s
· · · x}.)
(Recall that x is defined to be the word |xx{z
s times

Exercise 6.5.30. Prove Corollary 6.5.29.

 HOPF ALGEBRAS IN COMBINATORICS 187

[Hint: Lemma 6.5.16 (applied to α = x) yields

X X X
s!Mxhsi − Mxs ∈ Mβ = kMw ⊂ kMw
β∈Compsk ; w∈Compsk ; w∈Compsk ;
`(β)≤(s−1)`(x) `(w)≤(s−1)`(x) w < xs
wll

302 hsi P
. Adding this to the claim of Corollary 6.5.27, obtain s!Mx − s!Mxs ∈ w∈Compsk ; kMw , that is,
w < xs
  wll
hsi P
s! Mx − Mxs ∈ w∈Compsk ; kM w . It remains to get rid of the s! on the left hand side. Assume
w < xs
wll P
WLOG that k = Z, and argue that every f ∈ QSym satisfying s! · f ∈ w∈Compsk ; kMw must itself lie
w < xs
P wll
in w∈Compsk ; kMw .]
w < xs
wll

We are now ready to prove Theorem 6.5.13:

Exercise 6.5.31. Prove Theorem 6.5.13.  

hsi
[Hint: Lemma 6.5.14 yields that the family Mw is a reindexing of the family
  (w,s)∈RL×{1,2,3,...}
 
hgcd wi hgcd wi
Mred w . Hence, it is enough to prove that the family Mred w is an algebraically independent
w∈L w∈L
303
generating set of the k-algebra QSym. The latter claim, in turn, will follow from Lemma 6.3.7(c) once
hgcd wi
it is proven that the family Mred w generates the k-algebra QSym. So it remains to show that the
  w∈L
hgcd wi
family Mred w generates the k-algebra QSym.
w∈L  
hgcd wi
Let U denote the k-subalgebra of QSym generated by Mred w . It then suffices to prove that
w∈L
U = QSym. To this purpose, it is enough to prove that

(6.5.3) Mβ ∈ U for every composition β.

hji
For every reduced Lyndon composition
 α and every j ∈ {1, 2, 3, . . .}, the quasisymmetric function Mα
hgcd wi
is an element of the family Mred w and thus belongs to U . Combine this with Exercise 6.5.4(d) to
w∈L
see that
hsi
(6.5.4) Mβ ∈U for every Lyndon composition β and every s ∈ {1, 2, 3, . . .}

(because every Lyndon composition β can be written as α {n} for a reduced Lyndon composition α and an
n ∈ {1, 2, 3, . . .}). Now, prove (6.5.3) by strong induction: first, induce over |β|, and then, for fixed |β|,
induce over β in the wll-order. The induction step looks as follows: Fix some composition α, and assume
(as induction hypothesis) that:
• (6.5.3) holds for every composition β satisfying |β| < |α|;
• (6.5.3) holds for every composition β satisfying |β| = |α| and β < α.
wll

It remains to prove that (6.5.3) holds for β = α. In other words, it remains to prove that Mα ∈ U . Let
(a1 , a2 , . . . , ap ) be the CFL factorization of the word α. Assume WLOG that p 6= 0 (else, all is trivial). We
are in one of the following two cases:
Case 1: All of the words a1 , a2 , . . ., ap are equal.
Case 2: Not all of the words a1 , a2 , . . ., ap are equal.

302since every w ∈ Comp s

sk with the property that ` (w) ≤ (s − 1) ` (x) must satisfy w < x
wll
303applied to A = QSym, b = M hgcd wi , wt (N ) = N and g = M
w red w u u
188 DARIJ GRINBERG AND VICTOR REINER

In Case 2, there exists a k ∈ {1, 2, . . . , p − 1} satisfying ak > ak+1 (since a1 ≥ a2 ≥ · · · ≥ ap ), and thus
Corollary 6.5.25 (applied to u = α, n = |α|, x = a1 a2 · · · ak and y = ak+1 ak+2 · · · ap ) shows that
Mα = Ma1 a2 ···ak Mak+1 ak+2 ···ap
| {z } | {z }
∈U ∈U
(by the induction (by the induction
hypothesis) hypothesis)
 
 
 
− a sum of terms of the form Mw with w ∈ Comp|α| satisfying w < α
 
 |{z} wll 
 ∈U 
(by the induction
hypothesis)

∈ U U − (a sum of terms in U ) ⊂ U.
Hence, it only remains to deal with Case 1. In this case, set x = a1 = a2 = · · · = ap . Thus, α =
a1 a2 · · · ap = xp , whence |α| = p |x|. But Corollary 6.5.29 (applied to s = p and k = |x|) yields
X X
Mxhpi − Mxp ∈ kMw = k Mw (since p |x| = |α| and xp = α)
|{z}
w∈Compp|x| ; w∈Comp|α| ; ∈U
w < xp w<α (by the induction
wll wll
hypothesis)
X
⊂ kU ⊂ U,
w∈CompN ;
w<α
wll

so that Mxp ∈ Mxhpi −U ⊂ U − U ⊂ U . This rewrites as Mα ∈ U (since α = xp ). So Mα ∈ U is proven

| {z }
∈U
(by (6.5.4))
in both Cases 1 and 2, and thus the induction proof of (6.5.3) is finished.]
Exercise 6.5.32. Prove Theorem 6.4.3.
Of course, this proof of Theorem 6.4.3 yields a new (third) proof for Proposition 6.4.4.
We notice the following corollary of our approach to Theorem 6.4.3:
Corollary 6.5.33. The Λ-algebra QSym is a polynomial algebra (over Λ).
Exercise 6.5.34. Prove Corollary 6.5.33.  
hsi
[Hint: The algebraically independent generating set Mw of QSym contains the
(w,s)∈RL×{1,2,3,...}
hsi
elements M(1) = es ∈ Λ for all s ∈ {1, 2, 3, . . .}.]
 HOPF ALGEBRAS IN COMBINATORICS 189

7. Aguiar-Bergeron-Sottile character theory Part I: QSym as a terminal object

It turns out that the universal mapping property of NSym as a free associative algebra leads via duality
to a universal property for its dual QSym, elegantly explaining several combinatorial invariants that take
the form of quasisymmetric or symmetric functions:
• Ehrenborg’s quasisymmetric function of a ranked poset [54],
• Stanley’s chromatic symmetric function of a graph [182],
• the quasisymmetric function of a matroid considered in [21].

7.1. Characters and the universal property.

ζ
Definition 7.1.1. Given a Hopf algebra A over k, a character is an algebra morphism A −→ k, that is,
• ζ(1A ) = 1k ,
• ζ is k-linear, and
• ζ(ab) = ζ(a)ζ(b) for a, b in A.
Example 7.1.2. A particularly important character for A = QSym is defined as follows:304
ζQ
QSym −→ k
f (x) 7−→ f (1, 0, 0, . . .) = [f (x)]x1 =1,x2 =x3 =···=0 .
Hence,
(
1 if α = (n) for some n,
ζQ (Mα ) = ζQ (Lα ) =
0 otherwise.
In other words, the restriction ζQ |QSymn coincides with the functional Hn in NSymn = Homk (QSymn , k):
one has for f in QSymn that
(7.1.1) ζQ (f ) = (Hn , f ).
It is worth remarking that there is nothing special about setting x1 = 1 and x2 = x3 = · · · = 0: for
quasisymmetric f , we could have defined the same character ζQ by picking any variable, say xn , and sending
f (x) 7−→ [f (x)] xn =1, and .
xm =0 for m6=n

ζQ
This character QSym −→ k has a certain universal property.
ζ
Theorem 7.1.3. A connected graded Hopf algebra A together with a character A −→ k induces a unique
Ψ
graded Hopf morphism A −→ QSym making this diagram commute:

(7.1.2) A
Ψ / QSym

ζ ζQ
 |
k
Furthermore, Ψ has this formula on elements of An :
X
(7.1.3) Ψ(a) = ζα (a)Mα
α∈Compn

where for α = (α1 , . . . , α` ), the map ζα is the composite

∆(`−1) π ζ ⊗`
An −→ A⊗` −→
α
Aα1 ⊗ · · · ⊗ Aα` −→ k
π
in which A⊗` −→
α
Aα1 ⊗ · · · ⊗ Aα` is the canonical projection.

304We are using the notation of Proposition 5.1.9 here, and we are still identifying QSym with QSym (x), where x denotes
the infinite chain (x1 < x2 < · · · ).
190 DARIJ GRINBERG AND VICTOR REINER

Proof. One argues that Ψ is unique, and has formula (7.1.3), using only that ζ is k-linear and sends 1 to 1
and that Ψ is a graded k-coalgebra map making (7.1.2) commute. Equivalently, consider the adjoint k-algebra
map305
Ψ∗
NSym = QSymo −→ Ao .
Commutativity of (7.1.2) implies that for a in An ,
(7.1.1)
(Ψ∗ (Hn ), a) = (Hn , Ψ(a)) = ζQ (Ψ(a)) = ζ(a),
whereas gradedness of Ψ∗ yields that (Ψ∗ (Hm ), a) = 0 whenever a ∈ An and m 6= n. In other words, Ψ∗ (Hn )
is the element of Ao defined as the following functional on A:
(
∗ ζ(a) if a ∈ An ,
(7.1.4) Ψ (Hn )(a) =
0 if a ∈ Am for some m 6= n.
By the universal property for NSym ∼
= khH1 , H2 , . . .i as free associative k-algebra, we see that any choice of
ζ
a k-linear map A → k uniquely produces a k-algebra map Ψ∗ : QSymo → Ao which satisfies (7.1.4) for all
n ≥ 1. It is easy to see that this Ψ∗ then automatically satisfies (7.1.4) for n = 0 as well if ζ sends 1 to 1
ζ
(it is here that we use ζ(1) = 1 and the connectedness of A). Hence, any given k-linear map A → k sending
1 to 1 uniquely produces a k-algebra map Ψ∗ : QSymo → Ao which satisfies (7.1.4) for all n ≥ 0. Formula
(7.1.3) follows as
X
Ψ(a) = (Hα , Ψ(a)) Mα
α∈Comp

and for a composition α = (α1 , . . . , α` ), one has

(Hα , Ψ(a)) = (Ψ∗ (Hα ), a) = (Ψ∗ (Hα1 ) · · · Ψ∗ (Hα` ), a)
 
= Ψ∗ (Hα1 ) ⊗ · · · ⊗ Ψ∗ (Hα` ), ∆(`−1) (a)
(7.1.4)
 
= ζ ⊗` ◦ πα ∆(`−1) (a) = ζα (a),
where the definition of ζα was used in the last equality.
ζ
We wish to show that if, in addition, A is a Hopf algebra and A −→ k is a character (algebra map), then
Ψ
A −→ QSym will be an algebra map, that is, the two maps A ⊗ A −→ QSym given by Ψ ◦ m and m ◦ (Ψ ⊗ Ψ)
coincide. To see this, consider these two diagrams having the two maps in question as the composites of
their top rows:

(7.1.5) A⊗A
m /A Ψ / QSym A⊗A
Ψ⊗Ψ
/ QSym⊗2 m / QSym

ζ ζQ ⊗ζQ
"  |
ζ⊗ζ ζQ
(  v
ζ⊗ζ ζQ
k k
The fact that ζ, ζQ are algebra maps makes the above diagrams commute, so that applying the uniqueness in
ζ⊗ζ
the first part of the proof to the character A ⊗ A −→ k proves the desired equality Ψ ◦ m = m ◦ (Ψ ⊗ Ψ). 
Remark 7.1.4. When one assumes in addition that A is cocommutative, it follows that the image of Ψ will lie
in the subalgebra Λ ⊂ QSym, e.g. from the explicit formula (7.1.3) and the fact that one will have ζα = ζβ
ζΛ
whenever β is a rearrangement of α. In other words, the character Λ −→ k defined by restricting ζQ to Λ,
or by (
1 if λ = (n) for some n,
ζΛ (mλ ) =
0 otherwise,
has a universal property as terminal object with respect to characters on cocommutative co- or Hopf algebras.

305Here we are using the fact that there is a 1-to-1 correspondence between graded k-linear maps A → QSym and graded
k-linear maps QSymo → Ao given by f 7→ f ∗ , and this correspondence has the property that a given graded map f : A → QSym
is a k-coalgebra map if and only if f ∗ is a k-algebra map. This is a particular case of Exercise 1.6.1(f).
 HOPF ALGEBRAS IN COMBINATORICS 191

We close this section by discussing a well-known polynomiality and reciprocity phenomenon; see, e.g.,
Humpert and Martin [88, Prop. 2.2], Stanley [182, §4].
Definition 7.1.5. The binomial Hopf algebra (over the commutative ring k) is the polynomial algebra
k [m]
in a single variable m, with a Hopf algebra structure transported from the symmetric algebra Sym k
1 (which
is a Hopf algebra by virtue of Example 1.3.14, applied to V = k1 ) along the isomorphism Sym k1 → k [m]
which sends the standard basis element of k1 to m. Thus the element m is primitive; that is, ∆m =
1 ⊗ m + m ⊗ 1 and S(m) = −m. As S is an algebra anti-endomorphism by Proposition 1.4.8 and k[m] is
commutative, one has S(g)(m) = g(−m) for all polynomials g(m) in k[m].
Definition 7.1.6. For an element f (x) in QSym and a nonnegative integer m, let ps1 (f )(m) denote the
element of k obtained by principal specialization at q = 1
ps1 (f )(m) = [f (x)] x1 =x2 =···=xm =1,
xm+1 =xm+2 =···=0
= f (1, 1, . . . , 1, 0, 0, . . .).
| {z }
m ones

Proposition 7.1.7. Assume that Q is a subring of k. The map ps1 has the following properties.
(i) Let f ∈ QSym. There is a unique polynomial in k[m] which agrees for each nonnegative integer
m with ps1 (f )(m), and which, by abuse of notation, we will also denote ps1 (f )(m). If f lies in
QSymn , then ps1 (f )(m) is a polynomial of degree at most n, taking these values on Mα , Lα for
α = (α1 , . . . , α` ) in Compn :
 
1 m
ps (Mα )(m) = ,
`
 
1 m−`+n
ps (Lα )(m) = .
n
ps1
(ii) The map QSym −→ k[m] is a Hopf morphism into the binomial Hopf algebra.
(iii) For all m in Z and f in QSym one has
?m
ζQ (f ) = ps1 (f )(m).
In particular, one also has
?(−m)
ζQ (f ) = ps1 (S(f ))(m) = ps1 (f )(−m).
ζ
(iv) For a graded Hopf algebra A with a character A −→ k, and any element a in An , the polynomial
ps1 (Ψ(a))(m) in k[m] has degree at most n, and when specialized to m in Z satisfies
ζ ?m (a) = ps1 (Ψ(a))(m).

Proof. To prove assertion (i), note that one has

 
X m
x i1 · · · x α
1
 α1 
ps (Mα )(m) = Mα (1, 1, . . . , 1, 0, 0, . . .) = `
i` x =
j =1 `
1≤i1 <···<i` ≤m
X
ps1 (Lα )(m) = Lα (1, 1, . . . , 1, 0, 0, . . .) = [xi1 · · · xin ]xj =1
1≤i1 ≤···≤in ≤m:
ik <ik+1 if k∈D(α)
 
m−`+n
= |{1 ≤ j1 ≤ j2 ≤ · · · ≤ jn ≤ m − ` + 1}| = .
n
As {Mα }α∈Compn form a basis for QSymn , and m


` is a polynomial function in m of degree `(≤ n), one
1
concludes that for f in QSymn one has that ps (f )(m) is a polynomial function in m of degree at most n.
The polynomial giving rise to this function is unique, since infinitely many of its values are fixed.
To prove assertion (ii), note that ps1 is an algebra morphism because it is an evaluation homomorphism.
To check that it is a coalgebra morphism, it suffices to check ∆ ◦ ps1 = (ps1 ⊗ ps1 ) ◦ ∆ on each Mα for
192 DARIJ GRINBERG AND VICTOR REINER

A+B A B

P


α = (α1 , . . . , α` ) in Compn . Using the Vandermonde summation ` = k k `−k , one has
    X`    X`    
m m⊗1+1⊗m m⊗1 1⊗m m m
(∆ ◦ ps1 )(Mα ) = ∆ = = = ⊗
` ` k `−k k `−k
k=0 k=0

while at the same time

` `   

1 1

X
1 1
X m m
(ps ⊗ ps ) ◦ ∆ (Mα ) = ps (M(α1 ,...,αk ) ) ⊗ ps (M(αk+1 ,...,α` ) ) = ⊗ .
k `−k
k=0 k=0

Thus ps1 is a bialgebra map, and hence also a Hopf map, by Proposition 1.4.24(c).
For assertion (iii), first assume m lies in {0, 1, 2, . . .}. Since ζQ (f ) = f (1, 0, 0, . . .), one has
 
?m ⊗m ⊗m
ζQ (f ) = ζQ ◦ ∆(m−1) f (x) = ζQ f (x(1) , x(2) , . . . , x(m) )
h i
= f (x(1) , x(2) , . . . , x(m) ) x(1) =x(2) =···=x(m) =1,
1 1 1
(j) (j)
x2 =x3 =···=0 for all j

= f (1, 0, 0, . . . , 1, 0, 0, . . . , · · · , 1, 0, 0, . . .) = f (1, 1, . . . , 1, 0, 0, . . .) = ps1 (f )(m).

| {z }
m ones
306
But then Proposition 1.4.24(a) also implies
 ?m
?(−m) ?(−1) ?m ?m
ζQ (f ) = ζQ (f ) = (ζQ ◦ S) (f ) = ζQ (S(f ))
= ps1 (S(f ))(m) = S(ps1 (f ))(m) = ps1 (f )(−m).
For assertion (iv), note that
ζ ?m (a) = (ζQ ◦ Ψ)?m (a) = (ζQ
?m
)(Ψ(a)) = ps1 (Ψ(a))(m).
where the three equalities come from (7.1.2), Proposition 1.4.24(a), and assertion (iii) above, respectively. 

Remark 7.1.8. Aguiar, Bergeron and Sottile give a very cute (third) proof of the QSym antipode formula
Theorem 5.1.11, via Theorem 7.1.3, in [4, Example 4.8]. They apply Theorem 7.1.3 to the coopposite
?(−1) Ψ
coalgebra QSymcop and its character ζQ . One can show that the map QSymcop → QSym induced by
?(−1)
ζQ is Ψ = S, the antipode of QSym, because S : QSym → QSym is a coalgebra anti-endomorphism (by
?(−1)
Exercise 1.4.25) satisfying ζQ = ζQ ◦ S. They then use the formula (7.1.3) for Ψ = S (together with the
polynomiality Proposition 7.1.7) to derive Theorem 5.1.11.
?m
Exercise 7.1.9. Show that ζQ (f ) = ps1 (f )(m) for all f ∈ QSym and m ∈ {0, 1, 2, . . .}. (This was already
proven in Proposition 7.1.7(iii); give an alternative proof using Proposition 5.1.7.)
7.2. Example: Ehrenborg’s quasisymmetric function of a ranked poset. Here we consider incidence
algebras, coalgebras and Hopf algebras generally, and then particularize to the case of graded posets, to
recover Ehrenborg’s interesting quasisymmetric function invariant via Theorem 7.1.3.

7.2.1. Incidence algebras, coalgebras, Hopf algebras.

Definition 7.2.1. Given a family P of finite partially ordered sets P , let k[P] denote the free k-module
whose basis consists of symbols [P ] corresponding to isomorphism classes of posets P in P.
We will assume throughout that each P in P is bounded, that is, it has a unique minimal element 0̂ := 0̂P
and a unique maximal element 1̂ := 1̂P . In particular, P 6= ∅, although it is allowed that |P | = 1, so that
0̂ = 1̂; denote this isomorphism class of posets with one element by [o].
If P is closed under taking intervals
[x, y] := [x, y]P := {z ∈ P : x ≤P z ≤P y}

306See Exercise 7.1.9 for an alternative way to prove this, requiring less thought to verify its soundness.
 HOPF ALGEBRAS IN COMBINATORICS 193

then one can easily that the following coproduct and counit endow k[P] with the structure of a coalgebra,
called the (reduced) incidence coalgebra:
X
∆[P ] := [0̂, x] ⊗ [x, 1̂],
x∈P
(
1 if |P | = 1
[P ] :=
0 otherwise.

The dual algebra k[P]∗ is generally called the reduced incidence algebra (modulo isomorphism) for the family
ζ
P (see, e.g., [170]). It contains the important element k[P] −→ k, called the ζ-function that takes the value
ζ[P ] = 1 for all P .
If P (is not empty and) satisfies the further property of being hereditary in the sense that for every P1 , P2
in P, the Cartesian product poset P1 × P2 with componentwise partial order is also in P, then one can check
that the following product and unit endow k[P] with the structure of a (commutative) algebra:
[P1 ] · [P2 ] := m([P1 ] ⊗ [P2 ]) := [P1 × P2 ],
1k[P] := [o].

Proposition 7.2.2. For any hereditary family P of finite posets, k[P] is a bialgebra, and even a Hopf
algebra with antipode S given as in (1.4.7) (Takeuchi’s formula):
X X
S[P ] = (−1)k [x0 , x1 ] · · · [xk−1 , xk ].
k≥0 0̂=x0 <···<xk =1̂

Proof. Checking the commutativity of the pentagonal diagram in (1.3.4) amounts to the fact that, for any
(x1 , x2 ) <P1 ×P2 (y1 , y2 ), one has a poset isomorphism

[(x1 , x2 ) , (y1 , y2 )]P1 ×P2 ∼

= [x1 , y1 ]P1 × [x2 , y2 ]P2 .

Commutativity of the remaining diagrams in (1.3.4) is straightforward, and so k[P] is a bialgebra. But then
Remark 1.4.23 implies that it is a Hopf algebra, with antipode S as in (1.4.7), because the map f := idk[P] −u
(sending the class [o] to 0, and fixing all other [P ]) is locally ?-nilpotent:
X
f ?k [P ] = [x0 , x1 ] · · · [xk−1 , xk ]
0̂=x0 <···<xk =1̂

will vanish due to an empty sum whenever k exceeds the maximum length of a chain in the finite poset
P. 

It is perhaps worth remarking how this generalizes the Möbius function formula of P. Hall. Note that the
ζ
zeta function k[P] −→ k is a character, that is, an algebra morphism. Proposition 1.4.24(a) then tells us
µ=ζ ?−1
that ζ should have a convolutional inverse k[P] −→ k, traditionally called the Möbius function, with the
formula µ = ζ ?−1 = ζ ◦ S. Rewriting this via the antipode formula for S given in Proposition 7.2.2 yields P.
Hall’s formula.

Corollary 7.2.3. For a finite bounded poset P , one has

X
µ[P ] = (−1)k |{chains 0̂ = x0 < · · · < xk = 1̂ in P }|.
k≥0

We can also notice that S is an algebra anti-endomorphism (by Proposition 1.4.8), thus an algebra
endomorphism (since k[P] is commutative). Hence, µ = ζ◦S is a composition of two algebra homomorphisms,
thus an algebra homomorphism itself. We therefore obtain the following classical fact:

Corollary 7.2.4. For two finite bounded posets P and Q, we have µ[P × Q] = µ[P ] · µ[Q].
194 DARIJ GRINBERG AND VICTOR REINER

7.2.2. The incidence Hopf algebras for ranked posets and Ehrenborg’s function.
Definition 7.2.5. Take P to be the class of bounded ranked finite posets P , that is, those for which all
maximal chains from 0̂ to 1̂ have the same length r(P ). This is a hereditary class, as it implies that any
interval is [x, y]P is also ranked, and the product of two bounded ranked posets is also bounded and ranked.
r
It also uniquely defines a rank function P −→ N in which r(0̂) = 0 and r(x) is the length of any maximal
chain from 0̂ to x.
Example 7.2.6. Consider a pyramid with apex vertex a over a square base with vertices b, c, d, e:
a

b e

c d
Ordering its faces by inclusion gives a bounded ranked poset P , where the rank of an element is one more
than the dimension of the face it represents:
rank:

abcd 4

abc acd ade abe bcde 3

ab ac ad ae be bc cd de 2

a b c d e 1

∅ 0
Definition 7.2.7. Ehrenborg’s quasisymmetric function Ψ[P ] for a bounded ranked poset P is the image of
Ψ ζ
[P ] under the map k[P] −→ QSym induced by the zeta function k[P] −→ k as a character, via Theorem 7.1.3.
The quasisymmetric function Ψ[P ] captures several interesting combinatorial invariants of P ; see Stanley
[183, Chap. 3] for more background on these notions.
Definition 7.2.8. Let P be a bounded ranked poset P of rank r(P ) := r(1̂). Define its rank-generating
function X
RGF (P, q) := q r(p) ,
p∈P
its characteristic polynomial X
χ(P, q) := µ(0̂, p)q r(p)
p∈P
(where µ(p, q) is shorthand for µ([p, q])), its zeta polynomial
(7.2.1) Z(P, m) = |{multichains 0̂ ≤P p1 ≤P · · · ≤P pm−1 ≤P 1̂}|
r(P )−1  
X m
(7.2.2) = |{chains 0̂ < p1 < · · · < ps < 1̂}|
s=0
s+1
 HOPF ALGEBRAS IN COMBINATORICS 195

307
and for a subset S ⊂ {1, 2, . . . , r(P )−1}, its flag number fS , as a component of its flag f -vector (fS )S⊂[r−1]
defined by
fS = |{chains 0̂ <P p1 <P · · · <P ps <P 1̂ with {r(p1 ), . . . , r(ps )} = S}|,
as well as the flag h-vector entry hT given by fS = T ⊂S hT , or by inclusion-exclusion308, hS = T ⊂S (−1)|S\T | fT .
P P

Example 7.2.9. For the poset P in Example 7.2.6, one has RGF (P, q) = 1 + 5q + 8q 2 + 5q 3 + q 4 . Since
P is the poset of faces of a polytope, the Möbius function values for its intervals are easily predicted:
µ(x, y) = (−1)r[x,y] , that is, P is an Eulerian ranked poset; see Stanley [183, §3.16]. Hence its characteristic
polynomial is trivially related to the rank generating function, sending q 7→ −q, that is,
χ(P, q) = RGF (P, −q) = 1 − 5q + 8q 2 − 5q 3 + q 4 .
Its flag f -vector and h-vector entries are given in the following table.
S fS hS
∅ 1 1
{1} 5 5−1= 4
{2} 8 8−1= 7
{3} 5 5−1= 4
{1, 2} 16 16 − (5 + 8) + 1 = 4
{1, 3} 16 16 − (5 + 5) + 1 = 7
{2, 3} 16 16 − (5 + 8) + 1 = 4
{1, 2, 3} 32 32 − (16 + 16 + 16) + (5 + 8 + 5) − 1 = 1
and using (7.2.2), its zeta polynomial is
m2 (2m − 1)(2m + 1)
       
m m m m
Z(P, m) = 1 + (5 + 8 + 5) + (16 + 16 + 16) + 32 = .
1 2 3 4 3
Theorem 7.2.10. Assume that Q is a subring of k. Ehrenborg’s quasisymmetric function Ψ[P ] for a
bounded ranked poset P encodes
(i) the flag f -vector entries fS and flag h-vector entries hS as its Mα and Lα expansion coefficients309 :
X X
Ψ[P ] = fD(α) (P ) Mα = hD(α) (P ) Lα ,
α α
(ii) the zeta polynomial as the specialization from Definition 7.1.6
Z(P, m) = ps1 (Ψ[P ])(m) = [Ψ[P ]] x1 =x2 =···=xm =1, , and
xm+1 =xm+2 =···=0

(iii) the rank-generating function as the specialization

RGF (P, q) = [Ψ[P ]] x1 =q,x2 =1, ,
x3 =x4 =···=0
(iv) the characteristic polynomial as the convolution
χ(P, q) = ((ψq ◦ S) ? ζQ ) ◦ Ψ[P ]
ψq
where QSym −→ k[q] maps f (x) 7−→ f (q, 0, 0, . . .).
P
Proof. In assertion (i), the expansionP
Ψ[P ] = α fD(α) (P ) Mα is (7.1.3),
P since ζα [P ] = fD(α) (P ). The Lα
expansion follows from this, as Lα = β:D(β)⊃D(α) Mβ and fS (P ) = T ⊂S hT .
Assertion (ii) is immediate from Proposition 7.1.7(iv), since Z(P, m) = ζ ?m [P ].
Assertion (iii) can be deduced from assertion (i), but it is perhaps more fun and in the spirit of things
to proceed as follows. Note that ψq (Mα ) = q n for α = (n), and ψq (Mα ) vanishes for all other α 6= (n) in
Compn . Hence for a bounded ranked poset P one has
(7.2.3) (ψq ◦ Ψ)[P ] = q r(P ) .
307Actually, (7.2.2) is false if |P | = 1 (but only then). We use (7.2.1) to define Z(P, m) in this case.
308Specifically, we are using the converse of Lemma 5.2.6(a) here.
309In fact, Ehrenborg defined Ψ[P ] in [54, Defn. 4.1] via this M expansion, and then showed that it gave a Hopf morphism.
α
196 DARIJ GRINBERG AND VICTOR REINER

But if we treat ζQ : QSym → k as a map QSym → k [q], then (1.4.2) (applied to k [P], QSym, k [q], k [q], Ψ,
idk[q] , ψq and ζQ instead of C, C 0 , A, A0 , γ, α, f and g) shows that
(7.2.4) (ψq ? ζQ ) ◦ Ψ = (ψq ◦ Ψ) ? (ζQ ◦ Ψ) ,
since Ψ : k [P] → QSym is a k-coalgebra homomorphism. Consequently, one can compute
(7.2.3),
X X (7.1.2) X
RGF (P, q) = q r(p) · 1 = q r([0̂,p]) · ζ[p, 1̂] = (ψq ◦ Ψ)[0̂, p] · (ζQ ◦ Ψ)[p, 1̂]
p∈P p∈P p∈P
(7.2.4)
= ((ψq ◦ Ψ) ? (ζQ ◦ Ψ)) [P ] = (ψq ? ζQ )(Ψ[P ]) = (ψq ⊗ ζQ ) (∆Ψ[P ])
= [Ψ[P ](x, y)]x1 =q,x2 =x3 =···=0 = [Ψ[P ](x)] x1 =q,x2 =1,
y1 =1,y2 =y3 =···=0 x3 =x4 =···=0

Similarly, for assertion (iv) first note that

(7.2.5) ((ψq ◦ S) ? ζQ ) ◦ Ψ = (ψq ◦ S ◦ Ψ) ? (ζQ ◦ Ψ) ,
(this is proven similarly to (7.2.4), but now using the map ψq ◦ S instead of ψq ). Now, Proposition 7.2.2 and
Corollary 7.2.3 let one calculate that
X X
(ψq ◦ Ψ ◦ S)[P ] = (−1)k (ψq ◦ Ψ)([x0 , x1 ]) · · · (ψq ◦ Ψ)([xk−1 , xk ])
k 0̂=x0 <···<xk =1̂
(7.2.3) X X
= (−1)k q r(P ) = µ(0̂, 1̂)q r(P ) .
k 0̂=x0 <···<xk =1̂

This is used in the penultimate equality here:

(7.2.5)
((ψq ◦ S) ? ζQ ) ◦ Ψ[P ] = ((ψq ◦ S ◦ Ψ) ? (ζQ ◦ Ψ))[P ] = ((ψq ◦ Ψ ◦ S) ? ζ)[P ]
X X
= (ψq ◦ Ψ ◦ S)[0̂, p] · ζ[p, 1̂] = µ[0̂, p]q r(p) = χ(P, q).
p∈P p∈P


7.3. Example: Stanley’s chromatic symmetric function of a graph. We introduce the chromatic
Hopf algebra of graphs and an associated character ζ so that the map Ψ from Theorem 7.1.3 sends a graph G
to Stanley’s chromatic symmetric function of G. Then principal specialization ps1 sends this to the chromatic
polynomial of the graph.
7.3.1. The chromatic Hopf algebra of graphs.
Definition 7.3.1. The chromatic Hopf algebra (see Schmitt [172, §3.2]) G is a free k-module whose k-
basis elements [G] are indexed by isomorphism classes of (finite) simple graphs G = (V, E). Define for
G1 = (V1 , E1 ), G2 = (V2 , E2 ) the multiplication
[G1 ] · [G2 ] := [G1 t G2 ]
where [G1 t G2 ] denote the isomorphism class of the disjoint union, on vertex set V = V1 t V2 which is a
disjoint union of copies of their vertex sets V1 , V2 , with edge set E = E1 t E2 . For example,
     
• • · • = • • •
• • • •
Thus the class [∅] of the empty graph ∅ having V = ∅, E = ∅ is a unit element.
Given a subset V 0 ⊂ V , the subgraph induced on vertex set V 0 is defined as the graph G|V 0 := (V 0 , E 0 )
with edge set E 0 = {e ∈ E : e = {v1 , v2 } ⊂ V 0 }. This lets one define a comultiplication
X
∆[G] := [G|V1 ] ⊗ [G|V2 ].
(V1 ,V2 ):V1 tV2 =V

Define a counit (
1 if G = ∅
[G] :=
0 otherwise.
 HOPF ALGEBRAS IN COMBINATORICS 197

Proposition 7.3.2. The above maps endow G with the structure of a connected graded finite type Hopf
algebra over k, which is both commutative and cocommutative.

Example 7.3.3. Here are some examples of these structure maps:

     
• • · • = • • •
• • • •
       

∆ • • =1⊗ • •  + 2[ • ] ⊗  •  + 2 •  ⊗ [ • ] + [ • • ]⊗[ • ]
• • • •
 

+[ • ]⊗[ • • ]+ • • ⊗1

•
Proof of Proposition 7.3.2. The associativity of the multiplication and comultiplication should be clear as
m(2) ([G1 ] ⊗ [G2 ] ⊗ [G3 ]) = [G1 t G2 t G3 ]
X
∆(2) [G] = [G|V1 ] ⊗ [G|V2 ] ⊗ [G|V3 ].
(V1 ,V2 ,V3 ):
V =V1 tV2 tV3

Checking the unit and counit conditions are straightforward. Commutativity of the pentagonal bialgebra
diagram in (1.3.4) comes down to check that, given graphs G1 , G2 on disjoint vertex sets V1 , V2 , when one
applies to [G1 ] ⊗ [G2 ] either the composite ∆ ◦ m or the composite (m ⊗ m) ◦ (id ⊗T ⊗ id) ◦ (∆ ⊗ ∆), the
result is the same: X
[G1 |V11 t G2 |V21 ] ⊗ [G1 |V12 t G2 |V22 ].
(V11 ,V12 ,V21 ,V22 ):
V1 =V11 tV12
V2 =V21 tV22
Letting Gn be the k-span of [G] having n vertices makes G a bialgebra which is graded and connected,
and hence also a Hopf algebra by Proposition 1.4.14. Cocommutativity should be clear, and commutativity
follows from the graph isomorphism G1 t G2 ∼ = G2 t G1 . Finally, G is of finite type since there are only
finitely many isomorphism classes of simple graphs on n vertices for every given n. 
Remark 7.3.4. Humpert and Martin [88, Theorem 3.1] gave the following expansion for the antipode in
the chromatic Hopf algebra, containing fewer terms than Takeuchi’s general formula (1.4.7): given a graph
G = (V, E), one has
X
(7.3.1) S[G] = (−1)|V |−rank(F ) acyc(G/F )[GV,F ].
F

Here F runs over all subsets of edges that form flats in the graphic matroid for G, meaning that if e = {v, v 0 }
is an edge in E for which one has a path of edges in F connecting v to v 0 , then e also lies in F . Here G/F
denotes the quotient graph in which all of the edges of F have been contracted, while acyc(G/F ) denotes its
number of acyclic orientations, and GV,F := (V, F ) as a simple graph.310
Remark 7.3.5. In [14], Benedetti, Hallam and Machacek define a Hopf algebra of simplicial complexes, which
contains G as a Hopf subalgebra (and also has G as a quotient Hopf algebra). They compute a formula for
its antipode similar to (and generalizing) (7.3.1).
Remark 7.3.6. The chromatic Hopf algebra G is used in [107] and [36, §14.4] to study Vassiliev invariants
of knots. In fact, a certain quotient of G (named F in [107] and L in [36, §14.4]) is shown to naturally host
invariants of chord diagrams and therefore Vassiliev invariants of knots.
310The notation rank(F ) denotes the rank of F in the graphic matroid of G. We can define it without reference to matroid
theory as the maximum cardinality of a subset F 0 of F such that the graph GV,F 0 is acyclic. Equivalently, rank(F ) is |V | − c(F ),
where c(F ) denotes the number of connected components of the graph GV,F . Thus, the equality (7.3.1) can be rewritten as
S[G] = F (−1)c(F ) acyc(G/F )[GV,F ]. In this form, this equality is also proven in [15, Thm. 7.1].
P
198 DARIJ GRINBERG AND VICTOR REINER

Remark 7.3.7. The k-algebra G is isomorphic to a polynomial algebra (in infinitely many indeterminates)
over k. Indeed, every finite graph can be uniquely written as a disjoint union of finitely many connected finite
graphs (up to order). Therefore, the basis elements [G] of G corresponding to connected finite graphs G are
algebraically independent in G and generate the whole k-algebra G (indeed, the disjoint unions of connected
finite graphs are precisely the monomials in these elements). Thus, G is isomorphic to a polynomial k-
algebra with countably many generators (one for each isomorphism class of connected finite graphs). As a
consequence, for example, we see that G is an integral domain if k is an integral domain.
7.3.2. A “ribbon basis” for G and selfduality. In this subsection, we shall explore a second basis of G and a
bilinear form on G. This material will not be used in the rest of these notes (except in Exercise 7.3.25), but
it is of some interest and provides an example of how a commutative cocommutative Hopf algebra can be
studied.
First, let us define a second basis of G, which is obtained by Möbius inversion (in an appropriate sense)
from the standard basis ([G])[G] is an isomorphism class of finite graphs :

Definition 7.3.8. For every finite graph G = (V, E), set

E 0 \E c |
(−1)|
]
X
[G] = [H] ∈ G,
H=(V,E 0 );
E 0 ⊃E c
]
where E c denotes the complement of the subset E in the set of all two-element subsets of V . Clearly, [G]
depends only on the isomorphism class [G] of G, not on G itself.

Proposition 7.3.9. (a) Every finite graph G = (V, E) satisfies

X ]
[G] = [H] .
H=(V,E 0 );
E 0 ∩E=∅
]
(b) The elements [G] , where [G] ranges over all isomorphism classes of finite graphs, form a basis of the
k-module G.
(c) For any graph H = (V, E), we have
]
X ] ]
(7.3.2) ∆ [H] = [H|V1 ] ⊗ [H|V2 ] .
(V1 ,V2 );
V =V1 tV2 ;
H=H|V1 tH|V2

(d) For any two graphs H1 = (V1 , E1 ) and H2 = (V2 , E2 ), we have

] ]
X ]
(7.3.3) [H1 ] [H2 ] = [H] .
H=(V1 tV2 ,E);
H|V1 =H1 ;
H|V2 =H2

For example,
 ]        
• • =• • − • • − • • + • •
• • • • •
     

=• •  − 2 • • + • • .
• • •
Proving Proposition 7.3.9 is part of Exercise 7.3.14 further below.
]
The equalities that express the elements [G] in terms of the elements [H] (as in Definition 7.3.8), and
vice versa (Proposition 7.3.9(a)), are reminiscent of the relations (5.3.10) and (5.3.9) between the bases (Rα )
]
and (Hα ) of NSym. In this sense, we can call the basis of G formed by the [G] a “ribbon basis” of G.
We now define a k-bilinear form on G:
 HOPF ALGEBRAS IN COMBINATORICS 199

Definition 7.3.10. For any two graphs G and H, let Iso (G, H) denote the set of all isomorphisms from G
to H 311. Let us now define a k-bilinear form (·, ·) : G × G → k on G by setting
 
]
[G] , [H] = |Iso (G, H)| .
312

Proposition 7.3.11. The form (·, ·) : G × G → k is symmetric.

Again, we refer to Exercise 7.3.14 for a proof of Proposition 7.3.11.

The basis of G constructed in Proposition 7.3.9(b) and the bilinear form (·, ·) defined in Definition 7.3.10
can be used to construct a Hopf algebra homomorphism from G to its graded dual G o :
Definition 7.3.12. For any finite graph G, let aut (G) denote the number |Iso (G, G)|. Notice that this is
a positive integer, since the set Iso (G, G) is nonempty (it contains idG ).
Now, recall that the Hopf algebra G is a connected graded Hopf algebra of finite type. The n-th homoge-
neous component is spanned by the [G] where
G ranges over the graphs with n vertices. Since G is of finite
∗
type, its graded dual G o is defined. Let [G] [G] is an isomorphism class of finite graphs be the basis of G o dual to
the basis ([G])[G] is an isomorphism class of finite graphs of G. Define a k-linear map ψ : G → G o by
 
] ∗
ψ [G] = aut (G) · [G] for every finite graph G.
313

Proposition 7.3.13. Consider the map ψ : G → G o defined in Definition 7.3.12.

(a) This map ψ satisfies (ψ (a)) (b) = (a, b) for all a ∈ G and b ∈ G.
(b) The map ψ : G → G o is a Hopf algebra homomorphism.
(c) If Q is a subring of k, then the map ψ is a Hopf algebra isomorphism G → G o .

Exercise 7.3.14. Prove Proposition 7.3.9, Proposition 7.3.11 and Proposition 7.3.13.
Remark 7.3.15. Proposition 7.3.13(c) shows that the Hopf algebra G is self-dual when Q is a subring of k.
On the other hand, if k is a field of positive characteristic, then G is never self-dual. Here is a quick way to
∗
see this: The elements [G] of G o defined in Definition 7.3.12 have the property that
∗
n ∗
X
[◦] = n! · [G]
[G] is an isomorphism
class of finite graphs on
n vertices

for every n ∈ N, where ◦ denotes the graph with one vertex.314 Thus, if p is a prime and k is a field of
∗
p
characteristic p, then [◦] = 0. Hence, the k-algebra G o has nilpotents in this situation. However, the
k-algebra G does not (indeed, Remark 7.3.7 shows that it is an integral domain whenever k is an integral
domain). Thus, when k is a field of characteristic p, then G and G o are not isomorphic as k-algebras (let
alone as Hopf algebras).
311We recall that if G = (V, E) and H = (W, F ) are two graphs, then an isomorphism from G to H means a bijection
ϕ : V → W such that ϕ∗ (E) = F . Here, ϕ∗ denotes the map from the powerset of V to the powerset of W which sends every
T ⊂ V to ϕ (T ) ⊂ W .
312This is well-defined, because:

• the number |Iso (G, H)| depends only on the isomorphism classes [G] and [H] of G and H, but not on G and H
themselves;
• the elements [G]] , where [G] ranges over all isomorphism classes of finite graphs, form a basis of the k-module G
(because of Proposition 7.3.9(b));
• the elements [G], where [G] ranges over all isomorphism classes of finite graphs, form a basis of the k-module G.
 
313This is well-defined, since [G]] is a basis of the k-module G (because of Propo-
[G] is an isomorphism class of finite graphs
sition 7.3.9(b)).
314To see this, observe that the tensor [◦]⊗n appears in the iterated coproduct ∆(n−1) ([G]) exactly n! times whenever G is
a graph on n vertices.
200 DARIJ GRINBERG AND VICTOR REINER

7.3.3. Stanley’s chromatic symmetric function of a graph.

Definition 7.3.16. Stanley’s chromatic symmetric function Ψ[G] for a simple graph G = (V, E) is the image
Ψ ζ
of [G] under the map G −→ QSym induced via Theorem 7.1.3 from the edge-free character G −→ k defined
by
(
1 if G has no edges, that is, G is an independent/stable set of vertices,
(7.3.4) ζ[G] =
0 otherwise.
Note that, because G is cocommutative, Ψ[G] is symmetric and not just quasisymmetric; see Remark 7.1.4.
Recall that for a graph G = (V, E), a (vertex-)coloring f : V → {1, 2, . . .} is called proper if no edge
e = {v, v 0 } in E has f (v) = f (v 0 ).
315
Proposition 7.3.17. For a graph G = (V, E), the symmetric function Ψ[G] has the expansion
X
Ψ[G] = xf
proper colorings
f :V →{1,2,...}
Q
where xf := v∈V xf (v) . In particular, its specialization from Proposition 7.1.6 gives the chromatic polyno-
mial of G:
ps1 Ψ[G](m) = χG (m) = |{proper colorings f : V → {1, 2, . . . , m}}|
∆(`−1)
Proof. The iterated coproduct G −→ G ⊗` sends
X
[G] 7−→ [G|V1 ] ⊗ · · · ⊗ [G|V` ]
(V1 ,...,V` ):
V =V1 t···tV`

and the map ζ ⊗` sends each addend on the right to 1 or 0, depending upon whether each Vi ⊂ V is a stable
set or not, that is, whether the assignment of color i to the vertices in Vi gives a proper coloring of G. Thus
α`
formula (7.1.3) shows that the coefficient ζα of xα1 · · · x` in Ψ[G] counts the proper colorings f in which
1

−1
|f (i)| = αi for each i. 
Example 7.3.18. For the complete graph Kn on n vertices, one has
Ψ[Kn ] = n!en
 
m
ps1 (Ψ[Kn ])(m) = n!en (1, 1, . . . , 1) = n!
| {z } n
m ones
= m(m − 1) · · · (m − (n − 1)) = χ(Kn , m).
In particular, the single vertex graph K1 has Ψ[K1 ] = e1 , and since the Hopf morphism Ψ is in particular
an algebra morphism, a graph K1tn having n isolated vertices and no edges will have Ψ[K1tn ] = en1 .
As a slightly more interesting example, the graph P3 which is a path having three vertices and two edges
will have
Ψ[P3 ] = m(2,1) + 6m(1,1,1) = e2 e1 + 3e3
One might wonder, based on the previous examples, when Ψ[G] is e-positive, that is, when does its unique
expansion in the {eλ } basis for Λ have nonnegative coefficients? This is an even stronger assertion than
s-positivity, that is, having nonnegative coefficients for the expansion in terms of Schur functions {sλ }, since
each eλ is s-positive. This weaker property fails, starting with the claw graph K3,1 , which has
Ψ[K3,1 ] = s(3,1) − s(2,2) + 5s(2,1,1) + 8s(1,1,1,1) .
On the other hand, a result of Gasharov [62, Theorem 2] shows that one at least has s-positivity for Ψ[inc(P )]
where inc(P ) is the incomparability graph of a poset which is (3 + 1)-free; we refer the reader to Stanley
[182, §5] for a discussion of the following conjecture, which remains open316:
315In fact, Stanley defined Ψ[G] in [182, Defn. 2.1] via this expansion.
316A recent refinement for incomparability graphs of posets which are both (3 + 1)- and (2 + 2)-free, also known as unit
interval orders is discussed by Shareshian and Wachs [176].
 HOPF ALGEBRAS IN COMBINATORICS 201

Conjecture 7.3.19. For any (3 + 1)-free poset P , the incomparability graph inc(P ) has Ψ[inc(P )] an
e-positive symmetric function.

Here is another question about Ψ[G]: how well does it distinguish nonisomorphic graphs? Stanley gave
this example of two graphs G1 , G2 having Ψ[G1 ] = Ψ[G2 ]:
G1 = • • G2 = • •
•
• • • • •
At least Ψ[G] appears to do better at distinguishing trees, much better than its specialization, the chromatic
polynomial χ(G, m), which takes the same value m(m − 1)n−1 on all trees with n vertices.
Question 7.3.20. Does the chromatic symmetric function (for k = Z) distinguish trees?
It has been checked that the answer is affirmative for trees on 23 vertices or less. There are also interesting
partial results on this question by Martin, Morin and Wagner [142].
We close this section with a few other properties of Ψ[G] proven by Stanley which follow easily from the
theory we have developed. For example, his work makes no explicit mention of the chromatic Hopf algebra G,
and the fact that Ψ is a Hopf morphism (although he certainly notes the trivial algebra morphism property
Ψ[G1 t G2 ] = Ψ[G1 ]Ψ[G2 ]). One property he proves is implicitly related to Ψ as a coalgebra morphism: he
considers (in the case when Q is a subring of k) the effect on Ψ of the operator ∂p∂ 1 : ΛQ −→ ΛQ which acts
by first expressing a symmetric function f ∈ ΛQ as a polynomial in the power sums {pn }, and then applies
the partial derivative operator ∂p∂ 1 of the polynomial ring Q [p1 , p2 , p3 , . . .]. It is not hard to see that ∂p∂ 1
is the same as the skewing operator s⊥ ⊥
(1) = p1 : both act as derivations on ΛQ = Q[p1 , p2 , . . .], and agree in
their effect on each pn , in that both send p1 7→ 1, and both annihilate p2 , p3 , . . ..

Proposition 7.3.21. (Stanley [182, Cor. 2.12(a)]) For any graph G = (V, E), one has
∂ X
Ψ[G] = Ψ[G|V \v ].
∂p1
v∈V

Proof. Since Ψ is a coalgebra homomorphism, we have

X
∆Ψ[G] = (Ψ ⊗ Ψ)∆[G] = Ψ[G|V1 ] ⊗ Ψ[G|V2 ].
(V1 ,V2 ):
V =V1 tV2

Using this expansion (and the equality ∂

∂p1 = s⊥
(1) ), we now compute

∂ X X
Ψ[G] = s⊥
(1) Ψ[G] = (s(1) , Ψ[G|V1 ]) · Ψ[G|V2 ] = Ψ[G|V \v ]
∂p1
(V1 ,V2 ): v∈V
V =V1 tV2

(since degree considerations force (s(1) , Ψ[G|V1 ]) = 0 unless |V1 | = 1, in which case Ψ[G|V1 ] = s(1) ). 

Definition 7.3.22. Given a graph G = (V, E), an acyclic orientation Ω of the edges E (that is, an orientation
of each edge such that the resulting directed graph has no cycles), and a vertex-coloring f : V → {1, 2, . . .},
say that the pair (Ω, f ) are weakly compatible if whenever Ω orients an edge {v, v 0 } in E as v → v 0 , one has
f (v) ≤ f (v 0 ). Note that a proper vertex-coloring f of a graph G = (V, E) is weakly compatible with a unique
acyclic orientation Ω.

Proposition 7.3.23. (Stanley [182, Prop. 4.1, Thm. 4.2]) The involution ω of Λ sends Ψ[G] to ω (Ψ[G]) =
P
(Ω,f ) xf in which the sum runs over weakly compatible pairs (Ω, f ) of an acyclic orientation Ω and vertex-
coloring f .
Furthermore, the chromatic polynomial χG (m) has the property that (−1)|V | χ(G, −m) counts all such
weakly compatible pairs (Ω, f ) in which f : V → {1, 2, . . . , m} is a vertex-m-coloring.
202 DARIJ GRINBERG AND VICTOR REINER

Proof. As observed above, a proper coloring f is weakly compatible with a unique acyclic orientation Ω of G.
Denote by PΩ the poset on V which is the transitive closure of Ω, endowed with a strict labelling by integers,
that is, i <P j implies i >Z j. Then proper colorings f that induce Ω are the same as PΩ -partitions, so that
X
(7.3.5) Ψ[G] = FPΩ (x).
Ω
Applying the antipode S and using Corollary 5.2.20 gives
X X
ω (Ψ[G]) = (−1)|V | S (Ψ[G]) = FPΩopp (x) = xf
Ω (Ω,f )

where in the last line one sums over weakly compatible pairs as in the proposition. The last equality comes
from the fact that since each PΩ has been given a strict labelling, PΩopp acquires a weak (or natural) labelling,
that is i <PΩopp j implies i <Z j.
The last assertion follows from Proposition 7.1.7(iii). 
Remark 7.3.24. The interpretation of χ(G, −m) in Proposition 7.3.23 is a much older result of Stanley [181].
The special case interpreting χ(G, −1) as (−1)|V | times the number of acyclic orientations of G has sometimes
been called Stanley’s (-1)-color theorem. It also follows (via Proposition 7.1.7) from Humpert and Martin’s
antipode formula for G discussed in Remark 7.3.4: taking ζ to be the character of G given in (7.3.4),
X
χ(G, −1) = ζ ?(−1) [G] = ζ(S[G]) = (−1)|V |−rank(F ) acyc(G/F )ζ[GV,F ] = (−1)|V | acyc(G)
F
where the last equality uses the vanishing of ζ on graphs that have edges, so only the F = ∅ term survives.
Exercise 7.3.25. If V and X are two sets, and if f : V → X is any map, then eqs f will denote the set
{{u, u0 } | u ∈ V, u0 ∈ V, u 6= u0 and f (u) = f (u0 )} .
This is a subset of the set of all two-element subsets of V .
If G = (V, E) is a finite graph, then show that the map Ψ introduced in Definition 7.3.16 satisfies
  X
]
Ψ [G] = xf ,
f :V →{1,2,3,...};
eqs f =E
Q ]
where xf := v∈V xf (v) . Here, [G] is defined as in Definition 7.3.8.
7.4. Example: The quasisymmetric function of a matroid. We introduce the matroid-minor Hopf
algebra of Schmitt [169], and studied extensively by Crapo and Schmitt [38, 39, 40]. A very simple character
ζ on this Hopf algebra will then give rise, via the map Ψ from Theorem 7.1.3, to the quasisymmetric function
invariant of matroids from the work of Billera, Jia and the second author [21].
7.4.1. The matroid-minor Hopf algebra. We begin by reviewing some notions from matroid theory; see Oxley
[144] for background, undefined terms and unproven facts.
Definition 7.4.1. A matroid M of rank r on a (finite) ground set E is specified by a nonempty collection
B(M ) of r-element subsets of E with the following exchange property:
For any B, B 0 in B(M ) and b in B, there exists b0 in B 0 with (B \ {b}) ∪ {b0 } in B(M ).
The elements of B (M ) are called the bases of the matroid M .
Example 7.4.2. A matroid M with ground set E is represented by a family of vectors S = (ve )e∈E in a
vector space if B (M ) is the collection of subsets B ⊂ E having the property that the subfamily (ve )e∈B is a
basis for the span of all of the vectors in S.
For example, if M is the matroid with B(M ) = {{a, b}, {a, c}, {a, d}, {b, c}, {b, d}} on the ground set
E = {a, b, c, d}, then M is represented by the family S = (va , vb , vc , vd ) of the four vectors va = (1, 0), vb =
(1, 1), vc = (0, 1) = vd in R2 depicted here
vc ,O vd ; vb .
O

/ va
 HOPF ALGEBRAS IN COMBINATORICS 203

Conversely, whenever E is a finite set and S = (ve )e∈E is a family of vectors in a vector space, then the
set 
B ⊂ E : the subfamily (ve )e∈B is a basis for the span of all of the vectors in S
is a matroid on the ground set E.
A matroid is said to be linear if there exists a family of vectors in a vector space representing it. Not all
matroids are linear, but many important ones are.
Example 7.4.3. A special case of matroids M represented by vectors are graphic matroids, coming from a
graph G = (V, E), with parallel edges and self-loops allowed. One represents these by vectors in RV with
standard basis {v }v∈V by associating the vector v − v0 to any edge connecting a vertex v with a vertex v 0 .
One can check (or see [144, §1.2]) that the bases B in B(M ) correspond to the edge sets of spanning forests
for G, that is, edge sets which are acyclic and contain one spanning tree for each connected component of
G. For example, the matroid B(M ) corresponding to the graph G = (V, E) shown below:
•
a b
c

• •

is exactly the matroid represented by the vectors in Example 7.4.2; indeed, the spanning trees of this graph
G are the edge sets {a, b}, {a, c}, {a, d}, {b, c}, {b, d}.
To define the matroid-minor Hopf algebra one needs the basic matroid operations of deletion and con-
traction. These model the operations of deleting or contracting an edge in a graph. For configurations of
vectors they model the deletion of a vector, or the passage to images in the quotient space modulo the span
of a vector.
Definition 7.4.4. Given a matroid M of rank r and an element e of its ground set E, say that e is loop
(resp. coloop) of M if e lies in no basis (resp. every basis) B in B(M ). If e is not a coloop, the deletion M \ e
is a matroid of rank r on ground set E \ {e} having bases
(7.4.1) B(M \ e) := {B ∈ B(M ) : e 6∈ B}.
If e is not a loop, the contraction M/e is a matroid of rank r − 1 on ground set E \ {e} having bases
(7.4.2) B(M/e) := {B \ {e} : e ∈ B ∈ B(M )}.
When e is a loop of M , then M/e has rank r instead of r − 1 and one defines its bases as in (7.4.1) rather
than (7.4.2); similarly, if e is a coloop of M then M \ e has rank r − 1 instead of r and one defines its bases
as in (7.4.2) rather than (7.4.1).
Example 7.4.5. Starting with the graph G and its graphic matroid M from Example 7.4.3, the deletion
M \ a and contraction M/c correspond to the graphs G \ a and G/c shown here:
G\a= • G/c = •
b
c a b

• • •
d
d

One has
• B(M \ a) = {{b, c}, {b, d}}, so that b has become a coloop in M \ a, and
• B(M/c) = {{a}, {b}}, so that d has become a loop in M/c.
Definition 7.4.6. Deletions and contractions commute with each other. Thus, given a matroid M with
ground set E, and a subset A ⊂ E, two well-defined matroids can be constructed:
• the restriction M |A , which is a matroid on ground set A, obtained from M by deleting all e ∈ E \ A
in any order, and
204 DARIJ GRINBERG AND VICTOR REINER

• the quotient/contraction M/A, which is a matroid on ground set E \ A, obtained from M by con-
tracting all e ∈ A in any order.
We will also need the direct sum M1 ⊕ M2 of two matroids M1 and M2 . This is the matroid whose ground
set E = E1 t E2 is the disjoint union of a copy of the ground sets E1 , E2 for M1 , M2 , and whose bases are
B(M1 ⊕ M2 ) := {B1 t B2 : Bi ∈ B(Mi ) for i = 1, 2}.
ϕ
Lastly, say that two matroids M1 , M2 are isomorphic if there is a bijection of their ground sets E1 −→ E2
having the property that ϕB(M1 ) = B(M2 ).
Now one can define the matroid-minor Hopf algebra, originally introduced by Schmitt [169, §15], and
studied further by Crapo and Schmitt [38, 39, 40].
Definition 7.4.7. Let M have k-basis elements [M ] indexed by isomorphism classes of matroids. Define
the multiplication via
[M1 ] · [M2 ] := [M1 ⊕ M2 ]
so that the class [∅] of the empty matroid ∅ having empty ground set gives a unit. Define the comultiplication
for M a matroid on ground set E via
X
∆[M ] := [M |A ] ⊗ [M/A],
A⊂E
and a counit (
1 if M = ∅
[M ] :=
0 otherwise.

Proposition 7.4.8. The above maps endow M with the structure of a connected graded finite type Hopf
algebra over k, which is commutative.

Proof. Checking the unit and counit conditions are straightforward. Associativity and commutativity of the
multiplication follow because the direct sum operation ⊕ for matroids is associative and commutative up
to isomorphism. Coassociativity follows because for a matroid M on ground set E, one has this equality
between the two candidates for ∆(2) [M ]
X
[M |A1 ] ⊗ [(M |A2 )/A1 ] ⊗ [M/A2 ]
∅⊆A1 ⊆A2 ⊆E
X
= [M |A1 ] ⊗ [(M/A1 )|A2 \A1 ] ⊗ [M/A2 ]
∅⊆A1 ⊆A2 ⊆E

due to the matroid isomorphism (M |A2 )/A1 ∼ = (M/A1 )|A2 \A1 . Commutativity of the bialgebra diagram in
(1.3.4) amounts to the fact that for a pair of matroids M1 , M2 and subsets A1 , A2 of their (disjoint) ground
sets E1 , E2 , one has isomorphisms
M1 |A1 ⊕ M2 |A2 ∼
= (M1 ⊕ M2 ) |A1 tA2 ,
M1 /A1 ⊕ M2 /A2 ∼= (M1 ⊕ M2 ) /(A1 t A2 ).
Letting Mn be the k-span of [M ] for matroids whose ground set E has cardinality |E| = n, one can then
easily check that M becomes a bialgebra which is graded, connected, and of finite type, hence also a Hopf
algebra by Proposition 1.4.14. 
7.4.2. A quasisymmetric function for matroids.
ζ
Definition 7.4.9. Define a character M −→ k by
(
1 if M has only one basis,
ζ[M ] =
0 otherwise.
ζ
It is easily checked that this is a character, that is, an algebra map M −→ k. Note that if M has only one
basis, say B(M ) = {B}, then B :=L coloops(M ) is the set of coloops of M , and E \ B = loops(M ) is the set
of loops of M . Equivalently, M = e∈E M |{e} is the direct sum of matroids each having one element, each
a coloop or loop.
 HOPF ALGEBRAS IN COMBINATORICS 205

Ψ
Define Ψ[M ] for a matroid M to be the image of [M ] under the map M −→ QSym induced via Theo-
rem 7.1.3 from the above character ζ.
It turns out that Ψ[M ] is intimately related with greedy algorithms and finding minimum cost bases. A
fundamental property of matroids (and one that characterizes them, in fact; see [144, §1.8]) is that no matter
how one assigns costs f : E → R to the elements of E, the following greedy algorithm (generalizing Kruskal’s
algorithm for finding minimum costP spanning trees) always succeeds in finding one basis B in B(M ) achieving
the minimum total cost f (B) := b∈B f (b):
Algorithm 7.4.10. Start with the empty subset I0 = ∅ of E. For j = 1, 2, . . . , r, having already defined
the set Ij−1 , let e be the element of E \ Ij−1 having the lowest cost f (e) among all those for which Ij−1 ∪ {e}
is independent, that is, still a subset of at least one basis B in B(M ). Then define Ij := Ij−1 ∪ {e}. Repeat
this until j = r, and B = Ir will be among the bases that achieve the minimum cost.
Definition 7.4.11. Say that a cost function f : E → {1, 2, . . .} is M -generic if there is a unique basis B in
B(M ) achieving the minimum cost f (B).
Example 7.4.12. For the graphic matroid M of Example 7.4.3, this cost function f1 : E → {1, 2, . . .}
•
f1 (a)=1 f1 (b)=3
f1 (c)=3
• •

f1 (d)=2

is M -generic, as it minimizes uniquely on the basis {a, d}, whereas this cost function f2 : E → {1, 2, . . .}
•
f2 (a)=1 f2 (b)=3
f2 (c)=2
• •

f2 (d)=2

is not M -generic, as it achieves its minimum value on the two bases {a, c}, {a, d}.

Proposition 7.4.13. For a matroid M on ground set E, one has this expansion317
X
Ψ[M ] = xf
M -generic
f :E→{1,2,...}

where xf := e∈E xf (e) . In particular, for m ≥ 0, its specialization ps1 from Definition 7.1.6 has this
Q
interpretation:
ps1 Ψ[M ](m) = |{M -generic f : E → {1, 2, . . . , m}}|.
∆(`−1)
Proof. The iterated coproduct M −→ M⊗` sends
X
[M ] 7−→ [M |A1 ] ⊗ [(M |A2 )/A1 ] ⊗ · · · ⊗ [(M |A` )/A`−1 ]
where the sum is over flags of nested subsets
(7.4.3) ∅ = A0 ⊆ A1 ⊆ · · · ⊆ A`−1 ⊆ A` = E.
⊗`
The map ζ sends each summand to 1 or 0, depending upon whether each (M |Aj )/Aj−1 has a unique basis
α`
or not. Thus formula (7.1.3) shows that the coefficient ζα of xα i1 · · · xi` in Ψ[M ] counts the flags of subsets
1

in (7.4.3) for which |Aj \ Aj−1 | = αj and (M |Aj )/Aj−1 has a unique basis, for each j.
Given a flag as in (7.4.3), associate the cost function f : E → {1, 2, . . .} whose value on each element
of Aj \ Aj−1 is ij ; conversely, given any cost function f , say whose distinct values are i1 < . . . < i` , one
associates the flag having Aj \ Aj−1 = f −1 (ij ) for each j.
317In fact, this expansion was the original definition of Ψ[M ] in [21, Defn. 1.1].
206 DARIJ GRINBERG AND VICTOR REINER

Now, apply the greedy algorithm (Algorithm 7.4.10) to find a minimum-cost basis of M for such a
cost function f . At each step of the greedy algorithm, one new element is added to the independent set;
these elements weakly increase in cost as the algorithm progresses318. Thus, the algorithm first adds some
elements of cost i1 , then adds some elements of cost i2 , then adds some elements of cost i3 , and so on. We
can therefore subdivide the execution of the algorithm into phases 1, 2, . . . , `, where each phase consists of
some finite number of steps, such that all elements added in phase k have cost ik . (A phase may be empty.)
For each k ∈ {1, 2, . . . , `}, we let βk be the number of steps in phase k; in other words, βk is the number of
elements of elements of cost ik added during the algorithm.
We will prove below, using induction on s = 0, 1, 2, . . . , ` the following claim: After having completed
phases 1, 2, . . . , s in the greedy algorithm (Algorithm 7.4.10), there is a unique choice for the independent
set produced thus far, namely
s
G
(7.4.4) Iβ1 +β2 +···+βs = coloops((M |Aj )/Aj−1 ),
j=1

if and only if each of the matroids (M |Aj )/Aj−1 for j = 1, 2, . . . , s has a unique basis.
The case s = ` in this claim would show what we want, namely that f is M -generic, minimizing uniquely
on the basis shown in (7.4.4) with s = `, if and only if each (M |Aj )/Aj−1 has a unique basis.
The assertion of the claim is trivially true for s = 0. In the inductive step, one may assume that
• the independent set Iβ1 +β2 +···+βs−1 takes the form in (7.4.4), replacing s by s − 1,
• it is the unique f -minimizing basis for M |As−1 , and
• (M |Aj )/Aj−1 has a unique basis for j = 1, 2, . . . , s − 1.
Since As−1 exactly consists of all of the elements e of E whose costs f (e) lie in the range {i1 , i2 , . . . , is−1 },
in phase s the algorithm will work in the quotient matroid M/As−1 and attempt to augment Iβ1 +β2 +···+βs−1
using the next-cheapest elements, namely the elements of As \ As−1 , which all have cost f equal to is . Thus
the algorithm will have no choices about how to do this augmentation if and only if (M |As )/As−1 has a
unique basis, namely its set of coloops, in which case the algorithm will choose to add all of these coloops,
giving Iβ1 +β2 +···+βs as described in (7.4.4). This completes the induction.
The last assertion follows from Proposition 7.1.7. 

Example 7.4.14. If M has one basis then every function f : E → {1, 2, . . .} is M -generic, and
|E|
X
Ψ[M ] = xf = (x1 + x2 + · · · )|E| = M(1) .
f :E→{1,2,...}

Example 7.4.15. Let Ur,n denote the uniform matroid of rank r on n elements E, having B(Ur,n ) equal to
all of the r-element subsets of E.
As U1,2 has E = {1, 2} and B = {{1}, {2}}, genericity means f (1) 6= f (2), so
X
Ψ[U1,2 ] = xf (1) xf (2) = x1 x2 + x2 x1 + x1 x3 + x3 x1 + · · · = 2M(1,1) .
(f (1),f (2)):
f (1)6=f (2)

Similarly U1,3 has E = {1, 2, 3} with B = {{1}, {2}, {3}}, and genericity means either that f (1), f (2), f (3)
are all distinct, or that two of them are the same and the third is smaller. This shows
X X
Ψ[U1,3 ] = 3 xi x2j + 6 xi xj xk
i<j i<j<k
= 3M(1,2) + 6M(1,1,1)
   
m m m(m − 1)(2m − 1)
ps1 Ψ[U1,3 ](m) = 3 +6 =
2 3 2

318Proof. Let e be the element added at step i, and let e0 be the element added at step i + 1. We want to show that
f (e) ≤ f (e0 ). But the element e0 could already have been added at step i. Since it wasn’t, we thus conclude that the element
e that was added instead must have been cheaper or equally expensive. In other words, f (e) ≤ f (e0 ), qed.
 HOPF ALGEBRAS IN COMBINATORICS 207

One can similarly analyze U2,3 and check that

Ψ[U2,3 ] = 3M(2,1) + 6M(1,1,1)
   
m m m(m − 1)(2m − 1)
ps1 Ψ[U2,3 ](m) = 3 +6 =
2 3 2
These last examples illustrate the behavior of Ψ under the duality operation on matroids.
Definition 7.4.16. Given a matroid M of rank r on ground set E, its dual or orthogonal matroid M ⊥ is a
matroid of rank |E| − r on the same ground set E, having
B(M ⊥ ) := {E \ B}B∈B(M ) .
Here are a few examples of dual matroids.
Example 7.4.17. The dual of a uniform matroid is another uniform matroid:
⊥
Ur,n = Un−r,n .
Example 7.4.18. If M is matroid of rank r represented by family of vectors {e1 , . . . , en } in a vector space
over some field k, one can find a family of vectors {e⊥ ⊥
1 , . . . , en } that represent M
⊥
in the following way. Pick
n r×n
a basis for the span of the vectors {ei }i=1 , and create a matrix A in k whose columns express the ei in
terms of this basis. Then pick any matrix A⊥ whose row space is the null space of A, and one finds that the
columns {e⊥ n ⊥ ⊥
i }i=1 of A represent M . See Oxley [144, §2.2].

Example 7.4.19. Let G = (V, E) be a graph embedded in the plane with edge set E, giving rise to a
graphic matroid M on ground set E. Let G⊥ be a planar dual of G, so that, in particular, for each edge e
in E, the graph G⊥ has one edge e⊥ , crossing e transversely. Then the graphic matroid of G⊥ is M ⊥ . See
Oxley [144, §2.3].

Proposition 7.4.20. If Ψ[M ] = α cα Mα then Ψ[M ⊥ ] = α cα Mrev(α) .

P P

Consequently, ps1 Ψ[M ](m) = ps1 Ψ[M ⊥ ](m).

Proof. First, let us prove that if Ψ[M ] = α cα Mα then Ψ[M ⊥ ] = α cα Mrev(α) . In other words, let us
P P
show that for any given composition α, the coefficient of Mα in Ψ[M ] (when Ψ[M ] is expanded in the basis
(Mβ )β∈Comp of QSym) equals the coefficient of rev(α) in Ψ[M ⊥ ]. This amounts to showing that for any
composition α = (α1 , . . . , α` ), the cardinality of the set of M -generic f having xf = xα is the same as the
cardinality of the set of M ⊥ -generic f ⊥ having xf ⊥ = xrev(α) . We claim that the map f 7−→ f ⊥ in which
f ⊥ (e) = ` + 1 − f (e) gives a bijection between these sets. To see this, note that any basis B of M satisfies
X
(7.4.5) f (B) + f (E \ B) = f (e)
e∈E
⊥
(7.4.6) f (E \ B) + f (E \ B) = (` + 1)(|E| − r),
where r denotes the rank of M . Thus B is f -minimizing if and only if E \ B is f -maximizing (by (7.4.5)) if
and only if E \B is f ⊥ -minimizing (by (7.4.6)). Consequently f is M -generic if and only if f ⊥ is M ⊥ -generic.
1
The

last assertion follows, for example, from the calculation in Proposition 7.1.7(i) that ps (Mα )(m) =
m
`(α) together with the fact that `(rev(α)) = `(α). 

Just as (7.3.5) showed that Stanley’s chromatic symmetric function of a graph has an expansion as a sum
of P -partition enumerators for certain strictly labelled posets P , the same holds for Ψ[M ].
Definition 7.4.21. Given a matroid M on ground set E, and a basis B in B(M ), define the base-cobase
poset PB to have b < b0 whenever b lies in B and b0 lies in E \ B and (B \ {b}) ∪ {b0 } is in B(M ).
P
Proposition 7.4.22. For any matroid M , one has Ψ[M ] = B∈B(M ) F(PB ,strict) (x) where F(P, strict) (x) for
a poset P means the P -partition enumerator for any strict labelling of P , i.e. a labelling such that the
P -partitions satisfy f (i) < f (j) whenever i <P j.
In particular, Ψ[M ] expands nonnegatively in the {Lα } basis.
208 DARIJ GRINBERG AND VICTOR REINER

Proof. A basic result about matroids, due P to Edmonds [53], describes the edges in the matroid base polytope
which is the convex hull of all vectors { b∈B b }B∈B(M ) inside RE with standard basis {e }e∈E . He shows
that all such edges connect two bases B, B 0 that differ by a single basis exchange, that is, B 0 = (B \ {b})∪{b0 }
for some b in B and b0 in E \ B.
Polyhedral theory then says that a cost function f on E will minimize uniquely at B if and only if one
has a strict increase f (B) < f (B 0 ) along each such edge B → B 0 emanating from B, that is, if and only if
f (b) < f (b0 ) whenever b <PB b0 in the base-cobase poset PB , that is, f lies in A(PB , strict). 
Example 7.4.23. The graphic matroid from Example 7.4.3 has this matroid base polytope, with the bases
B in B(M ) labelling the vertices:
cd

ac ad

bc bd
The base-cobase posets PB for its five vertices B are as follows:
a b
c d
b d a d a c b c
a c b c b d a d
One can label the first of these five strictly as
1 2
3 4
and compute its strict P -partition enumerator from the linear extensions {3412, 3421, 4312, 4321} as
L(2,2) + L(2,1,1) + L(1,1,2) + L(1,1,1,1)
while any of the last four can be labelled strictly as
1 2
3 4
and they each have an extra linear extension 3142 giving their strict P -partition enumerators as
L(2,2) + L(2,1,1) + L(1,1,2) + L(1,1,1,1) + L(1,2,1) .
Hence one has
Ψ[M ] = 5L(2,2) + 5L(1,1,2) + 4L(1,2,1) + 5L(2,1,1) + 5L(1,1,1,1) .
As M is a graphic matroid for a self-dual planar graph, one has a matroid isomorphism M ∼ = M ⊥ (see
Example 7.4.19), reflected in the fact that Ψ[M ] is invariant under the symmetry swapping Mα ↔ Mrev(α)
(and simultaneously swapping Lα ↔ Lrev(α) ).
This P -partition expansion for Ψ[M ] also allows us to identify its image under the antipode of QSym.

Proposition 7.4.24. For a matroid M on ground set E, one has

X
S(Ψ[M ]) = (−1)|E| |{f -maximizing bases B}| · xf
f :E→{1,2,...}

and X
ps1 Ψ[M ](−m) = (−1)|E| |{f -maximizing bases B}|.
f :E→{1,2,...,m}
In particular, the expected number of f -maximizing bases among all cost functions f : E → {1, 2, . . . , m} is
(−m)−|E| ps1 Ψ[M ](−m).
 HOPF ALGEBRAS IN COMBINATORICS 209

Proof. Corollary 5.2.20 implies

X X
S(Ψ[M ]) = S(F(PB ,strict) (x)) = (−1)|E| F(PBopp ,natural) (x)
B∈B(M ) B∈B(M )

where F(P,natural) (x) is the enumerator for P -partitions in which P has been naturally labelled, so that
they satisfy f (i) ≤ f (j) whenever i <P j. When P = PBopp , this is exactly the condition for f to achieve
its maximum value at f (B) (possibly not uniquely), that is, for f to lie in the closed normal cone to
the vertex indexed by B in the matroid base polytope; compare this with the discussion in the proof of
Proposition 7.4.22. Thus one has
X
S(Ψ[M ]) = (−1)|E| xf
(B,f ):
B∈B(M )
f maximizing at B

which agrees with the statement of the proposition, after reversing the order of the summation.
The rest follows from Proposition 7.1.7. 

Example 7.4.25. We saw in Example 7.4.23 that the matroid M from Example 7.4.3 has
Ψ[M ] = 5L(2,2) + 5L(1,1,2) + 4L(1,2,1) + 5L(2,1,1) + 5L(1,1,1,1) .
and therefore will have
m(m − 1)(2m2 − 2m + 1)
     
1 m−2+4 m−3+4 m−4+4
ps Ψ[M ](m) = 5 + (5 + 4 + 5) +5 = .
4 4 4 2

using ps1 (Lα )(m) = m−`+|α|

|α|
from Proposition 7.1.7 (i). Let us first do a reality-check on a few of its values
with m ≥ 0 using Proposition 7.4.13, and for negative m using Proposition 7.4.24:
m −1 0 1 2
ps1 Ψ[M ](m) 5 0 0 5
When m = 0, interpreting the set of cost functions f : E → {1, 2, . . . , m} as being empty explains why
the value shown is 0. When m = 1, there is only one function f : E → {1}, and it is not M -generic; any of
the 5 bases in B(M ) will minimize f (B), explaining both why the value for m = 1 is 0, but also explaining
the value of 5 for m = −1. The value of 5 for m = 2 counts these M -generic cost functions f : E → {1, 2}:
• • • • •
1 1 1 2 1 2 2 1 2 1
2 1 2 1 2

• • • • • • • • • •

2 2 1 2 1

Lastly, Proposition 7.4.24 predicts the expected number of f -minimizing bases for f : E → {1, 2, . . . , m} as
m(m + 1)(2m2 + 2m + 1) (m + 1)(2m2 + 2m + 1)
(−m)−|E| ps1 Ψ[M ](−m) = (−m)−4 = ,
2 2m3
whose limit as m → ∞ is 1, consistent with the notion that “most” cost functions should be generic with
respect to the bases of M , and maximize/minimize on a unique basis.
Remark 7.4.26. It is not coincidental that there is a similarity of results for Stanley’s chromatic symmet-
ric function of a graph Ψ[G] and for the matroid quasisymmetric function Ψ[M ], such as the P -partition
expansions (7.3.5) versus Proposition 7.4.22, and the reciprocity results Proposition 7.3.23 versus Propo-
sition 7.4.24. It was noted in [21, §9] that one can associate a similar quasisymmetric function invariant
to any generalized permutohedra in the sense of Postnikov [153]. Furthermore, recent work of Ardila and
Aguiar [3] has shown that there is a Hopf algebra of such generalized permutohedra, arising from a Hopf
monoid in the sense of Aguiar and Mahajan [6]. This Hopf algebra generalizes the chromatic Hopf algebra
210 DARIJ GRINBERG AND VICTOR REINER

of graphs319 and the matroid-minor Hopf algebra, and its quasisymmetric function invariant derives as usual
from Theorem 7.1.3. Their work [3] also provides a generalization of the chromatic Hopf algebra antipode
formula of Humpert and Martin [88] discussed in Remark 7.3.4 above.

319Aguiar and Ardila actually work with a larger Hopf algebra of graphs. Namely, their concept of graphs allows parallel
edges, and it also allows “half-edges”, which have only one endpoint. If G = (V, E) is such a graph (where E is the set of its
edges and its half-edges), and if V 0 is a subset of V , then they define G/V 0 to be the graph on vertex set V 0 obtained from G
by
• removing all vertices that are not in V 0 ,
• removing all edges that have no endpoint in V 0 , and all half-edges that have no endpoint in V 0 , and
• replacing all edges that have only one endpoint in V 0 by half-edges.
(This is to be contrasted with the induced subgraph G |V 0 , which is constructed in the same way but with the edges that have
only one endpoint in V 0 getting removed as well.) The comultiplication
P  they
 defineon the Hopf algebra of such graphs sends
the isomorphism class [G] of a graph G = (V, E) to G |V1 ⊗ G/V2 . This is no longer a cocommutative Hopf
(V1 ,V2 ):V1 tV2 =V
algebra; our Hopf algebra G is a quotient of it. In [3, Corollary 13.10], Ardila and Aguiar compute the antipode of the Hopf
monoid of such graphs; this immediately leads to a formula for the antipode of the corresponding Hopf algebra, because what
they call the Fock functor K preserves antipodes [3, Theorem 2.18].
 HOPF ALGEBRAS IN COMBINATORICS 211

8. The Malvenuto-Reutenauer Hopf algebra of permutations

Like so many Hopf algebras we have seen, the Malvenuto-Reutenauer Hopf algebra FQSym can be thought
of fruitfully in more than one way. One is that it gives a natural noncommutative lift of the quasisymmetric
P -partition enumerators and the fundamental basis {Lα } of QSym, rendering their product and coproduct
formulas even more natural.

8.1. Definition and Hopf structure.

L
Definition 8.1.1. Define FQSym = n≥0 FQSymn to be a graded k-module in which FQSymn has k-basis
{Fw }w∈Sn indexed by the permutations w = (w1 , . . . , wn ) in Sn .

We first attempt to lift the product and coproduct formulas (5.2.6), (5.2.5) in the {Lα } basis of QSym.
We attempt to define a product for u ∈ Sk , v ∈ S` as follows:

X
(8.1.1) Fu Fv := Fw
w∈u
v[k]

where we regard permutations as words (namely, every π ∈ Sn is identified with the word (π1 , π2 , . . . , πn )),


and where for v = (v1 , . . . , v` ) one sets v[k] := (k + v1 , . . . , k + v` ). Note that the multiset u v[k] is an
actual set in this situation (i.e., has each element appear only once) and is a subset of Sk+` .
The coproduct will be defined using the notation of standardization std(i) of a word i = (i1 , . . . , in ) in
some linearly ordered alphabet. This standardization is the permutation in Sn obtained by replacing all
the occurrences of the smallest letter in i by the numbers 1, 2, . . . , m1 from left to right (where m1 is the
number of times this smallest letter occurs in i), then replacing all occurrences of the next smallest letter by
the numbers m1 + 1, m1 + 2, . . . , m1 + m2 , from left to right (where m2 is how often this letter occurs in i),
etc.320

Example 8.1.2. Considering words in the Roman alphabet a < b < c < · · ·

std(b a c c b a a b a c b)
= (5 1 9 10 6 2 3 7 4 11 8).

Using this, define for w = (w1 , . . . , wn ) in Sn

n
X
(8.1.2) ∆Fw := Fstd(w1 ,w2 ,...,wk ) ⊗ Fstd(wk+1 ,wk+2 ,...,wn ) .
k=0

It is possible to check directly that the maps defined in (8.1.1) and (8.1.2) endow FQSym with the
structure of a connected graded finite type Hopf algebra; see Hazewinkel, Gubareni, Kirichenko [78, Thm.
7.1.8]. However in justifying this here, we will follow the approach of Duchamp, Hivert and Thibon [50, §3],
which exhibits FQSym as a subalgebra of a larger ring of (noncommutative) power series of bounded degree
in a totally ordered alphabet.

320Alternatively, one can define std(i) as the unique permutation σ ∈ S such that for every 1 ≤ a < b ≤ n, we have
n
σ (a) < σ (b) if and only if ia ≤ ib .
212 DARIJ GRINBERG AND VICTOR REINER

Definition 8.1.3. Given a totally ordered set I, create a totally ordered variable set {Xi }i∈I , and the ring
Rh{Xi }i∈I i of noncommutative power series of bounded degree in this alphabet321. Many times, we will use
a variable set X := (X1 < X2 < · · · ), and call the ring RhXi.
We first identify the algebra structure for FQSym as the subalgebra of finite type within Rh{Xi }i∈I i
spanned by the elements X
Fw = Fw ({Xi }i∈I ) := Xi
i=(i1 ,...,in ):
std(i)=w−1
S
where Xi := Xi1 · · · Xin , as w ranges over n≥0 Sn .
Example 8.1.4. For the alphabet X = (X1 < X2 < · · · ), in RhXi one has
X
F1 = Xi = X1 + X2 + · · ·
1≤i
X
F12 = Xi Xj = X12 + X22 + · · · + X1 X2 + X1 X3 + X2 X3 + X1 X4 + · · ·
1≤i≤j
X
F21 = Xj Xi = X2 X1 + X3 X1 + X3 X2 + X4 X1 + · · ·
1≤i<j
X X
F312 = Xi = Xj Xk Xi
i:std(i)=231 1≤i<j≤k

= X22 X1 + X32 X1 + X32 X2 + · · · + X2 X3 X1 + X2 X4 X1 + · · ·

S
Proposition 8.1.5. For any totally ordered infinite set I, the elements {Fw } as w ranges over n≥0 Sn
form the k-basis for a subalgebra FQSym({Xi }i∈I ) of RhXi, which is connected graded and of finite type,
having multiplication defined k-linearly by (8.1.1).
Consequently all such algebras are isomorphic to a single algebra FQSym, having basis {Fw } and multi-
plication given by the rule (8.1.1), with the isomorphism mapping Fw 7−→ Fw ({Xi }i∈I ).
For example,
F1 F21 = (X1 + X2 + X3 + · · · )(X2 X1 + X3 X1 + X3 X2 + X4 X1 + · · · )
= X1 · X3 X2 + X1 · X4 X2 + · · · + X1 · X2 X1 + X2 · X3 X2 + X2 · X4 X2 + · · ·
+ X2 · X3 X1 + X2 · X4 X1 + · · · + X2 · X2 X1 + X3 · X3 X1 + X3 · X3 X2 + · · ·
+ X3 · X2 X1 + X4 · X2 X1 + · · ·
X X X X
= Xi + Xi + Xi = F132 + F312 + F321 = Fw
i:std(i)=132 i:std(i)=231 i:std(i)=321 w∈1
32
Proof of Proposition 8.1.5. The elements {Fw ({Xi }i∈I )} are linearly independent as they are supported on
disjoint monomials, and so form a k-basis for their span. The fact that they multiply via rule (8.1.1) is the
equivalence of conditions (i) and (iii) in the following Lemma 8.1.6, from which all the remaining assertions
follow. 
321Let us recall the definition of Rh{X }
i i∈I i.
Let N denote the free monoid on the alphabet {Xi }i∈I ; it consists of words Xi1 Xi2 · · · Xik . We define a topological k-module
k hh{Xi }i∈I ii to be the Cartesian product kN (equipped with the product topology), but we identify its element (δw,u )u∈N
with the word w for every w ∈ N . Thus, every element (λw )w∈N ∈ kN = k hh{Xi }i∈I ii can be rewritten as the convergent
P
sum w∈N λw w. We call λw the coefficient of w in this element (or the coefficient of this element before w). The elements of
k hh{Xi }i∈I ii will be referred to as noncommutative power series. We define a multiplication on k hh{Xi }i∈I ii by the formula
    
X X X X
 λw w   µw w  =  λu µv  w.
w∈N w∈N w∈N (u,v)∈N 2 ; w=uv

(This is well-defined thanks to the fact that, for each w ∈ N , there are only finitely many (u, v) ∈ N 2 satisfying w = uv.) Thus,
k hh{Xi }i∈I ii becomes a k-algebra with unity 1 (the empty word). (It is similar to the monoid algebra kN of N over k, with
the only difference that infinite sums are allowed.)
P Now, we define Rh{Xi }i∈I i to be the k-subalgebra of k hh{Xi }i∈I ii consisting of all noncommutative power series
w∈N λw w ∈ k hh{Xi }i∈I ii of bounded degree (i.e., such that all words w ∈ N of sufficiently high length satisfy λw = 0).
 HOPF ALGEBRAS IN COMBINATORICS 213

Lemma 8.1.6. For a triple of permutations

u = (u1 , . . . , uk ) in Sk ,
v = (v1 , . . . , vn−k ) in Sn−k ,
w = (w1 , . . . , wn ) in Sn ,
the following conditions are equivalent:
(i) w−1 lies in the set u−1
v −1 [k].
(ii) u = std(w1 , . . . , wk ) and v = std(wk+1 , . . . , wn ),
(iii) for some word i = (i1 , . . . , in ) with std(i) = w one has u = std(i1 , . . . , ik ) and v = std(ik+1 , . . . , in ).

Proof of Lemma 8.1.6. The implication (ii) ⇒ (iii) is clear since std(w) = w. The reverse implication (iii)
⇒ (ii) is best illustrated by example, e.g. considering Example 8.1.2 as concatenated, with n = 11, k =
6, n − k = 5:
w = std (b a c c b a | a b a c b)
= (5 1 9 10 6 2 | 3 7 4 11 8)

u = std (5 1 9 10 6 2) v = std (3 7 4 11 8)
= (3 1 5 6 4 2) = (1 3 2 5 4)
= std (b a c c b a) = std (a b a c b)
The equivalence of (i) and (ii) is a fairly standard consequence of unique parabolic factorization W =
W J WJ where W = Sn and WJ = Sk × Sn−k , so that W J are the minimum-length coset representatives
for cosets xWJ (that is, the permutations x ∈ Sn satisfying x1 < · · · < xk and xk+1 < · · · < xn ). One can
uniquely express any w in W as w = xy with x in W J and y in WJ , which here means that y = u·v[k] = v[k]·u
for some u in Sk and v in Sn−k . Therefore w = xuv[k], if and only if w−1 = u−1 v −1 [k]x−1 , which means that
w−1 is the shuffle of the sequences u−1 in positions {x1 , . . . , xk } and v −1 [k] in positions {xk+1 , . . . , xn }. 
Example 8.1.7. To illustrate the equivalence of (i) and (ii) and the parabolic factorization in the preceding
proof, let n = 9 and k = 5 with
 
1 2 3 4 5 | 6 7 8 9
w=
4 9 6 1 5 | 8 2 3 7
   
1 2 3 4 5 | 6 7 8 9 1 2 3 4 5 6 7 8 9
=
1 4 5 6 9 | 2 3 7 8 2 5 4 1 3 9 6 7 8
= x · u · v[k]

 
1 2 3 4 5 6 7 8 9
w−1 =
4 9 6 1 5 8 2 3 7
   
1 2 3 4 5 6 7 8 9 1 2 3 4 5 6 7 8 9
=
4 1 5 3 2 7 8 9 6 1 6 7 2 3 4 8 9 5
= u−1 · v −1 [k] · x−1
Proposition 8.1.5 yields that FQSym is isomorphic to the k-subalgebra FQSym (X) of the k-algebra R hXi
when X is the variable set (X1 < X2 < · · · ). We identify FQSym with FQSym (X)
along this isomorphism.
For any infinite alphabet {Xi }i∈I and any f ∈ FQSym,

we denote by f {xi } i∈I the image of f under the
algebra isomorphism FQSym → FQSym {Xi }i∈I defined in Proposition 8.1.5.
One can now use this to define a coalgebra structure on FQSym. Roughly speaking, one wants to first
evaluate an element f in FQSym ∼ = FQSym (X) ∼ = FQSym (X, Y) as f (X, Y), using the linearly ordered
variable set (X, Y) := (X1 < X2 < · · · < Y1 < Y2 < · · · ). Then one should take the image of f (X, Y) after
imposing the partial commutativity relations
(8.1.3) Xi Yj = Yj Xi for every pair (Xi , Yj ) ∈ X × Y,
and hope that this image lies in a subalgebra isomorphic to
FQSym (X) ⊗ FQSym (Y) ∼ = FQSym ⊗ FQSym .
214 DARIJ GRINBERG AND VICTOR REINER

We argue this somewhat carefully. Start by considering the canonical monoid epimorphism
ρ
(8.1.4) F hX, Yi  M,
where F hX, Yi denotes the free monoid on the alphabet (X, Y) and M denotes its quotient monoid imposing
the partial commutativity relations (8.1.3). Let kM denote the k-module of all functions f : M → k, with
pointwise addition and scalar multiplication; similarly define kF hX,Yi . As both monoids F hX, Yi and M
enjoy the property that an element m has only finitely many factorizations as m = m1 m2 , one can define a
convolution algebra structure on both kF hX,Yi and kM via
X
(f1 ? f2 )(m) = f1 (m1 )f2 (m2 ),
(m1 ,m2 )∈N ×N :
m=m1 m2

where N is respectively F hX, Yi or M . As fibers of the map ρ in (8.1.4) are finite, it induces a map of
convolution algebras, which we also call ρ:
ρ
(8.1.5) kF hX,Yi  kM .
Now recall that RhXi denotes the algebra of noncommutative formal power series in the variable set X, of
bounded degree, with coefficients in k. One similarly has the ring RhX, Yi, which can be identified with
the subalgebra of kF hX,Yi consisting of the functions f : F hX, Yi → k having a bound on the length of the
words in their support (the value of f on a word in (X, Y) gives its power series coefficient corresponding
to said word). We let RhM i denote the analogous subalgebra of kM ; this can be thought of as the algebra
of bounded degree “partially commutative power series” in the variable sets X and Y. Note that ρ restricts
to a map
ρ
(8.1.6) RhX, Yi → RhM i.
Finally, we claim (and see Proposition 8.1.9 below for a proof) that this further restricts to a map
ρ
(8.1.7) FQSym (X, Y) → FQSym (X) ⊗ FQSym (Y)
in which the target is identified with its image under the (injective322) multiplication map
FQSym (X) ⊗ FQSym (Y) ,→ RhM i
f (X) ⊗ g(Y) 7→ f (X)g(Y).
Using the identification of FQSym with all three of FQSym (X) , FQSym (Y) , FQSym (X, Y), the map ρ in
(8.1.7) will then define a coproduct structure on FQSym. Abusing notation, for f in FQSym, we will simply
write ∆(f ) = f (X, Y) instead of ρ(f (X, Y)).
Example 8.1.8. Recall from Example 8.1.4 that one has
X X
F312 = Xi = Xj Xk Xi
i:std(i)=231 1≤i<j≤k

and therefore its coproduct is

∆F312 = F312 (X1 , X2 , . . . , Y1 , Y2 , . . .) (by our abuse of notation)
X X X X
= Xj Xk Xi + Xj Yk Xi + Yj Yk Xi + Yj Yk Yi
i<j≤k i<j, i, i<j≤k
k j≤k
X X X X
= Xj Xk Xi · 1 + Xj Xi · Yk + Xi · Yj Yk + 1 · Yj Yk Yi
i<j≤k i<j, i, i<j≤k
k j≤k
= F312 (X) · 1 + F21 (X) · F1 (Y) + F1 (X) · F12 (Y) + 1 · F312 (Y)
= F312 ⊗ 1 + F21 ⊗ F1 + F1 ⊗ F12 + 1 ⊗ F312 .
Proposition 8.1.9. The map ρ in (8.1.6) does restrict as claimed to a map as in (8.1.7), and hence defines a
coproduct on FQSym, acting on the {Fw } basis by the rule (8.1.2). This endows FQSym with the structure
of a connected graded finite type Hopf algebra.
322as images of the basis F (X) ⊗ F (Y) of FQSym(X) ⊗ FQSym(Y) are supported on disjoint monomials in RhM i, so
u v
linearly independent.
 HOPF ALGEBRAS IN COMBINATORICS 215

Proof.
n Let I be the totally ordered oset {1 < 2 < 3 < · · · }. Let J be the totally ordered set
1 < 2 < 3 < ··· < e
1<e 2<e 3 < · · · . We set Xei = Yi for every positive integer i. Then, the alphabet
(X, Y) can be written as {Xi }i∈J .
If i is a word over the alphabet I = {1 < 2 < 3 < · · · }, then we denote by ei the word over J obtained
from i by replacing every letter i by ei.
For the first assertion of Proposition 8.1.9, it suffices to check that Fw indeed has the image under ∆
claimed in (8.1.2). Let n ∈ N and w ∈ Sn . Then,
∆Fw = Fw (X, Y) (by our abuse of notation)
X X
= (X, Y)i = (X, Y)t
i∈J n :std(i)=w−1 t∈J n :std(t)=w−1
n
X X X
(8.1.8) = (X, Y)t
k=0 (i,j)∈I k ×I n−k t∈J n :
std(t)=w−1 ;


t∈i e j

(since for every t ∈ J , there exists exactly one choice of k ∈ {0, 1, . . . , n} and (i, j) ∈ I k × I n−k satisfying
n

t ∈ i ej; namely, i is the restriction of t to the subalphabet I of J, whereas j is the restriction of t to J \ I,
and k is the length of i). P
We now fix k and (i, j), and try to simplify the inner sum t∈J n : (X, Y)t on the right hand side of
std(t)=w−1 ;


t∈i ej


(8.1.8). First we notice that this sum is nonempty if and only if there exists some t ∈ i ej satisfying std(t) =
w−1 . This existence is easily seen to be equivalent to w−1 ∈ std(i)

std(j)[k] (since the standardization
of any shuffle in i
ej is the corresponding shuffle in std(i)

std(j)[k]). This, in turn, is equivalent to
std(i) = (std(w1 , . . . , wk ))−1 and std(j) = (std(wk+1 , . . . , wn ))−1 (according to the equivalence (i) ⇐⇒ (ii)
in Lemma 8.1.6). Hence, the inner sum on the right hand side of (8.1.8) is nonempty if and only if std(i) =
(std(w1 , . . . , wk ))−1 and std(j) = (std(wk+1 , . . . , wn ))−1 . When it is nonempty, it has only one addend323,
and this addend is (X, Y)t = Xi Yj (since t ∈ i

ej). Summarizing, we see that the inner sum on the right
hand side of (8.1.8) equals Xi Yj when std(i) = (std(w1 , . . . , wk ))−1 and std(j) = (std(wk+1 , . . . , wn ))−1 , and
is empty otherwise. Thus, (8.1.8) simplifies to
Xn X
∆Fw = Xi Yj
k=0 (i,j)∈I k ×I n−k :
std(i)=(std(w1 ,...,wk ))−1
std(j)=(std(wk+1 ,...,wn ))−1
n
X
= Fstd(w1 ,...,wk ) (X)Fstd(wk+1 ,...,wn ) (Y)
k=0
Xn
= Fstd(w1 ,...,wk ) ⊗ Fstd(wk+1 ,...,wn ) ∈ FQSym ⊗ FQSym .
k=0
This proves (8.1.2), and thus the first assertion of Proposition 8.1.9.
From this, it is easy to derive that ∆ satisfies coassociativity (i.e., the diagram (1.2.1) holds for C =
FQSym). (Alternatively, one can obtain this from the associativity of multiplication using Corollary 8.1.11.)
We have already verified the rule (8.1.2). The connected graded structure on FQSym gives a counit and an
antipode for free. 
Exercise 8.1.10. We say that a permutation w ∈ Sn is connected if n is a positive integer and if there
exists no i ∈ {1, 2, . . . , n − 1} satisfying f ({1, 2, . . . , i}) = {1, 2, . . . , i}. Let CS denote the set of all con-
nected permutations of all n ∈ N. Show that FQSym is a free (noncommutative) k-algebra with generators
(Fw )w∈CS . (This statement means that (Fw1 Fw2 · · · Fwk )k∈N; (w1 ,w2 ,...,wk )∈CSk is a basis of the k-module
FQSym.)
323In fact, the elements std (t) for t ∈ i
ej are distinct, and thus only one of them can equal w−1 .
216 DARIJ GRINBERG AND VICTOR REINER

[Hint: This is a result of Poirier and Reutenauer [152, Theorem 2.1]; it is much easier than the similar
Theorem 6.4.3.]

Corollary 8.1.11. The Hopf algebra FQSym is self-dual: Let {Gw } be the dual k-basis to the k-basis {Fw }
for FQSym. Then, the k-linear map sending Gw 7−→ Fw−1 is a Hopf algebra isomorphism FQSymo −→
FQSym.

Proof. For any 0 ≤ k ≤ n, any u ∈ Sk and any v ∈ Sn−k , one has

X X
Fu−1 Fv−1 = Fw−1 = Fw−1
w−1 ∈u−1
v −1 [k] w:
std(w1 ,...,wk )=u
std(wk+1 ,...,wn )=v

via the equivalence of (i) and (ii) in Lemma 8.1.6. On the other hand, in FQSymo , the dual k-basis {Gw }
to the k-basis {Fw } for FQSym should have product formula
X
Gu Gv = Gw
w:
std(w1 ,...,wk )=u
std(wk+1 ,...,wn )=v

coming from the coproduct formula (8.1.2) for FQSym in the {Fw }-basis. Comparing these equalities, we see
that the k-linear map τ sending Gw 7−→ Fw−1 is an isomorphism FQSymo −→ FQSym of k-algebras. Hence,
o
the adjoint τ ∗ : FQSymo → (FQSymo ) of this map is an isomorphism of k-coalgebras. But identifying
o o
(FQSym ) with FQSym in the natural way (since FQSym is of finite type), we easily see that τ ∗ = τ ,
whence τ itself is an isomorphism of both k-algebras and k-coalgebras, hence of k-bialgebras, hence of Hopf
algebras. 
We can now be a bit more precise about the relations between the various algebras
Λ, QSym, NSym, FQSym, RhXi, R(x).
Not only does FQSym allow one to lift the Hopf structure of QSym, it dually allows one to extend the Hopf
structure of NSym. To set up this duality, note that Corollary 8.1.11 motivates the choice of an inner product
on FQSym in which
(Fu , Fv ) := δu−1 ,v .
We wish to identify the images of the ribbon basis {Rα } of NSym when included in FQSym.
Definition 8.1.12. For any composition α, define an element Rα of FQSym by
X X X
Rα := Fw−1 = Xi = Xi
w:Des(w)=D(α) (w,i): i:Des(i)=D(α)
Des(w)=D(α)
std(i)=w

where the w in the sums are supposed to belong to S|α| , and where the descent set of a sequence i =
(i1 , . . . , in ) is defined by
Des(i) := {j ∈ {1, 2, . . . , n − 1} : ij > ij+1 } = Des(std(i)).
P
Alternatively, Rα = T XT in which the sum is over column-strict tableaux of the ribbon skew shape α, and
XT = Xi in which i is the sequence of entries of T read in order from the southwest toward the northeast.
Example 8.1.13. Taking α = (1, 3, 2), with ribbon shape and column-strict fillings T as shown
i5 ≤ i6
  ∧
   T = i2 ≤ i3 ≤ i4
 ∧
i1
one has that
X X X
R(1,3,2) = Xi = Xi1 Xi2 Xi3 Xi4 Xi5 Xi6 = XT
i=(i1 ,i2 ,i3 ,i4 ,i5 ,i6 ): i1 >i2 ≤i3 ≤i4 >i5 ≤i6 T
Des(i)=D(α)={1,4}
 HOPF ALGEBRAS IN COMBINATORICS 217

Corollary 8.1.14. For every n ∈ N and w ∈ Sn , we let γ(w) denote the unique composition α of n satisfying
D (α) = Des (w). The following Z-linear surjection and injection
π
FQSym  QSym
Fw 7−→ Lγ(w)
ι
NSym ,→ FQSym
Rα −
7 → Rα
are Hopf morphisms, and are also adjoint maps with respect to the above choice of inner product on FQSym
and the usual dual pairing between NSym and QSym. Furthermore, the second map ι lets one factor the
surjection NSym  Λ as follows
NSym → FQSym ,→ RhXi → R(x)
Rα 7−→ Rα 7−→ sα (x)
through the abelianization RhXi  R(x) sending the noncommutative variable Xi to the commutative xi .
π
Proof. The fact that FQSym  QSym is a Hopf map comes from checking that it respects the product
(compare (5.2.6) and (8.1.1)) and the coproducts (compare (5.2.5) and (8.1.2)) then applying Proposi-
ι
tion 1.4.24(c). It follows that NSym ,→ FQSym is a Hopf map once we check the adjointness assertion, via
the following calculation:
 
X 1 if Des(w) = D(α)
(ι(Rα ), Fw ) = (Rα , Fw ) = (Fu−1 , Fw ) =
0 otherwise
u:Des(u)=D(α)

= (Rα , Lγ(w) ) = (Rα , π(Fw )).

The last assertion is clear: the abelianization map sends the noncommutative tableau monomial XT to the
commutative tableau monomial xT . 

We summarize some of this picture as follows:

dual
FQSym
O FQSym
ι π
? 
dual
NSym QSym O
π
 ?
dual
Λ Λ
Furthermore, if we denote by ι the canonical inclusion Λ → QSym as well, then the diagram

FQSym
9
π

,
ι
%%
NSym QSym
9
π
ι
%% +
Λ
is commutative (according to the last sentence of Corollary 8.1.14).

Remark 8.1.15. Different notations for FQSym appear in the literature. In the book [23] (which presents an
unusual approach to the character theory of the symmetric group using FQSym), the Hopf algebra FQSym
is called P, and its basis that we call {Gw }w∈Sn is denoted {w}w∈Sn . In [78, Chapter 7], the Hopf algebra
FQSym and its basis {Fw }w∈Sn are denoted M P R and {w}w∈Sn , respectively.
218 DARIJ GRINBERG AND VICTOR REINER

9. Further topics
The following is a list of topics that were, at one point, planned to be touched in class, but did not make
the cut. They might get elaborated upon in a future version of these notes.
9.0.1. 0-Hecke algebras.
• Review of representation theory of finite-dimensional algebras.
Review the notions of indecomposables, simples, projectives, along with the theorems of Krull-
Remak-Schmidt, of Jordan-Hölder, and the two kinds of Grothendieck groups dual to each other.
• 0-Hecke algebra representation theory.
Describe the simples and projectives, following Denton, Hivert, Schilling, Thiery [46] on J -trivial
monoids.
• Nsym and Qsym as Grothendieck groups.
Give Krob and Thibon’s interpretation (see [192, §5] for a brief summary) of
• QSym and the Grothendieck group of composition series, and
• NSym and the Grothendieck group of projectives.
Remark 9.0.1. Mention P. McNamara’s interpretation, in the case of supersolvable lattices, of the
Ehrenborg quasisymmetric function as the composition series enumerator for an Hn (0)-action on the
maximal chains
9.0.2. Aguiar-Bergeron-Sottile character theory Part II: Odd and even characters, subalgebras.
9.0.3. Face enumeration, Eulerian posets, and cd-indices. Borrowing from Billera’s ICM notes.
• f-vectors, h-vectors
• flag f-vectors, flag h-vectors
• ab-indices and cd-indices
9.0.4. Other topics.
• Loday-Ronco Hopf algebra of planar binary trees
• Poirier-Reutenauer Hopf algebra of tableaux
• Reading Hopf algebra of Baxter permutations
• Hopf monoids, e.g. of Hopf algebra of generalized permutohedra, of matroids, of graphs, Stanley
chromatic symmetric functions and Tutte polynomials
• Lam-Pylyavskyy Hopf algebra of set-valued tableaux
• Connes-Kreimer Hopf algebra and renormalization
• Noncommutative symmetric functions and ΩΣCP ∞
• Maschke’s theorem and “integrals” for Hopf algebras
• Nichols-Zoeller structure theorem and group-like elements
• Cartier-Milnor-Moore structure theorem and primitive elements
• Quasi-triangular Hopf algebras and quantum groups
• The Steenrod algebra, its dual, and tree Hopf algebras
• Ringel-Hall algebras of quivers
• Ellis-Khovanov odd symmetric function Hopf algebras (see also Lauda-Russell)
Student talks given in class were:
(1) Al Garver, on Maschke’s theorem for finite-dimensional Hopf algebras
(2) Jonathan Hahn, on the paper by Humpert and Martin.
(3) Emily Gunawan, on the paper by Lam, Lauve and Sottile.
(4) Jonas Karlsson, on the paper by Connes and Kreimer
(5) Thomas McConville, on Butcher’s group and generalized Runge-Kutta methods.
(6) Cihan Bahran, on universal enveloping algebras and the Poincaré-Birkhoff-Witt theorem.
(7) Theodosios Douvropolos, on the Cartier-Milnor-Moore theorem.
(8) Alex Csar, on the Loday-Ronco Hopf algebra of binary trees
(9) Kevin Dilks, on Reading’s Hopf algebra of (twisted) Baxter permutations
(10) Becky Patrias, on the paper by Lam and Pylyavskyy
(11) Meng Wu, on multiple zeta values and Hoffman’s homomorphism from QSym
 HOPF ALGEBRAS IN COMBINATORICS 219

10. Some open problems and conjectures

• Is there a proof of the Assaf-McNamara skew Pieri rule that gives a resolution of Specht or Schur/Weyl
modules whose character corresponds to sλ/µ hn , whose terms model their alternating sum?
• Explicit antipodes in the Lam-Pylyavskyy Hopf algebras? (Answered by Patrias in [150].)
• P. McNamara’s question [134, Question 7.1]: are P -partition enumerators irreducible for connected
posets P ?
• Stanley’s question: are the only P -partition enumerators which are symmetric (not just quasisym-
metric) those for which P is a skew-shape with a column-strict labelling?
• Does Stanley’s chromatic symmetric function distinguish trees?
• Hoffman’s stuffle conjecture
• Billera-Brenti’s nonnegativity conjecture for the total cd-index of Bruhat intervals ([20, Conjecture
6.1])

11. Appendix: Some basics

In this appendix, we briefly discuss some basic notions from linear algebra and elementary combinatorics
that are used in these notes.
11.1. Linear expansions and triangularity. In this Section, we shall recall some fundamental results from
linear algebra (most importantly, the notions of a change-of-basis matrix and of a unitriangular matrix), but
in greater generality than how it is usually done in textbooks. We shall use these results later when studying
bases of combinatorial Hopf algebras; but per se, this section has nothing to do with Hopf algebras.
11.1.1. Matrices. Let us first define the notion of a matrix whose rows and columns are indexed by arbitrary
objects (as opposed to numbers):324
Definition 11.1.1. Let S and T be two sets. An S × T -matrix over k shall mean a family (as,t )(s,t)∈S×T ∈
kS×T of elements of k indexed by elements of S × T . Thus, the set of all S × T -matrices over k is kS×T .
We shall abbreviate “S × T -matrix over k” by “S × T -matrix” when the value of k is clear from the
context.
This definition of S × T -matrices generalizes the usual notion of matrices (i.e., the notion of n × m-
matrices): Namely, if n ∈ N and m ∈ N, then the {1, 2, . . . , n} × {1, 2, . . . , m}-matrices are precisely the
n × m-matrices (in the usual meaning of this word). We shall often use the word “matrix ” for both the usual
notion of matrices and for the more general notion of S × T -matrices.
Various concepts defined for n×m-matrices (such as addition and multiplication of matrices, or the notion
of a row) can be generalized to S × T -matrices in a straightforward way. The following four definitions are
examples of such generalizations:
Definition 11.1.2. Let S and T be two sets.
(a) The sum of two S × T -matrices is defined by (as,t )(s,t)∈S×T + (bs,t )(s,t)∈S×T = (as,t + bs,t )(s,t)∈S×T .
(b) If u ∈ k and if (as,t )(s,t)∈S×T ∈ kS×T , then we define u (as,t )(s,t)∈S×T to be the S × T -matrix
(uas,t )(s,t)∈S×T .
(c) Let A = (as,t )(s,t)∈S×T be an S × T -matrix. For every s ∈ S, we define the s-th row of A to be the
{1} × T -matrix (as,t )(i,t)∈{1}×T . (Notice that {1} × T -matrices are a generalization of row vectors.)
Similarly, for every t ∈ T , we define the t-th column of A to be the S × {1}-matrix (as,t )(s,i)∈S×{1} .
Definition 11.1.3. Let S be a set.
(a) The S × S identity matrix is defined to be the S × S-matrix (δs,t )(s,t)∈S×S . This S × S-matrix is
denoted by IS . (Notice that the n × n identity matrix In is I{1,2,...,n} for each n ∈ N.)
(b) An S × S-matrix (as,t )(s,t)∈S×S is said to be diagonal if every (s, t) ∈ S × T satisfying s 6= t satisfies
as,t = 0.
(c) Let A = (as,t )(s,t)∈S×S be an S × S-matrix. The diagonal of A means the family (as,s )s∈S . The
diagonal entries of A are the entries of this diagonal (as,s )s∈S .
324As before, k denotes a commutative ring.
220 DARIJ GRINBERG AND VICTOR REINER

Definition 11.1.4. Let S, T and U be three sets. Let A = (as,t )(s,t)∈S×T be an S × T -matrix, and
P
let B = (bt,u )(t,u)∈T ×U be a T × U -matrix. Assume that the sum t∈T as,t bt,u is well-defined for every
(s, u) ∈ S × U . (For example, this is guaranteed to hold if the set T is finite. For infinite T , it may and may
not hold.) Then, the S × U -matrix AB is defined by
!
X
AB = as,t bt,u .
t∈T (s,u)∈S×U

Definition 11.1.5. Let S and T be two finite sets. We say that an S × T -matrix A is invertible if and only
if there exists a T × S-matrix B satisfying AB = IS and BA = IT . In this case, this matrix B is unique; it
is denoted by A−1 and is called the inverse of A.
The definitions that we have just given are straightforward generalizations of the analogous definitions
for n × m-matrices; thus, unsurprisingly, many properties of n × m-matrices still hold for S × T -matrices.
For example:

Proposition 11.1.6. (a) Let S and T be two sets. Let A be an S × T -matrix. Then, IS A = A and
AIT = A.
(b) Let S, T and U be three sets such that T is finite. Let A and B be two S × T -matrices. Let C be a
T × U -matrix. Then, (A + B) C = AC + BC.
(c) Let S, T , U and V be four sets such that T and U are finite. Let A be an S × T -matrix. Let B be
a T × U -matrix. Let C be a U × V -matrix. Then, (AB) C = A (BC).

The proof of Proposition 11.1.6 (and of similar properties that will be left unstated) is analogous to the
proofs of the corresponding properties of n × m-matrices.325 As a consequence of these properties, it is easy
to see that if S is any finite set, then kS×S is a k-algebra.
In general, S × T -matrices (unlike n × m-matrices) do not have a predefined order on their rows and
their columns. Thus, the classical notion of a triangular n × n-matrix cannot be generalized to a notion of
a “triangular S × S-matrix” when S is just a set with no additional structure. However, when S is a poset,
such a generalization can be made:
Definition 11.1.7. Let S be a poset. Let A = (as,t )(s,t)∈S×S be an S × S-matrix.
(a) The matrix A is said to be triangular if and only if every (s, t) ∈ S × S which does not satisfy t ≤ s
must satisfy as,t = 0. (Here, ≤ denotes the smaller-or-equal relation of the poset S.)
(b) The matrix A is said to be unitriangular if and only if A is triangular and has the further property
that, for every s ∈ S, we have as,s = 1.
(c) The matrix A is said to be invertibly triangular if and only if A is triangular and has the further
property that, for every s ∈ S, the element as,s of k is invertible.
Of course, all three notions of “triangular”, “unitriangular” and “invertibly triangular” depend on the
partial order on S.
Clearly, every invertibly triangular S × S-matrix is triangular. Also, every unitriangular S × S-matrix is
invertibly triangular (because the element 1 of k is invertible).
We can restate the definition of “invertibly triangular” as follows: The matrix A is said to be invertibly
triangular if and only if it is triangular and its diagonal entries are invertible. Similarly, we can restate the
definition of “unitriangular” as follows: The matrix A is said to be unitriangular if and only if it is triangular
and all its diagonal entries equal 1.
325A little warning: In Proposition 11.1.6(c), the condition that T and U be finite can be loosened (we leave this to
the interested reader), but cannot be completely disposed of. It can happen that both (AB) C and A (BC) are defined, but
(AB) C(= A (BC) does ! not hold (if we remove this condition). For example,
( this happens
! if S = Z, T = Z, U = Z, V = Z,
1, if i ≥ j; 0, if i ≥ j;
A= , B = (δi,j − δi,j+1 )(i,j)∈Z×Z and C = . (Indeed, in this example, it
0, if i < j 1, if i < j
(i,j)∈Z×Z (i,j)∈Z×Z
is easy to check that AB = IZ and BC = −IZ and thus (AB) C = IZ C = C 6= −A = A (−IZ ) = A (BC).)
| {z } | {z }
=IZ =BC
This seeming paradox is due to the subtleties of rearranging infinite sums (similarly to how a conditionally convergent series
of real numbers can change its value when its entries are rearranged).
 HOPF ALGEBRAS IN COMBINATORICS 221

Definition 11.1.7(a) generalizes both the notion of upper-triangular matrices and the notion of lower-
triangular matrices. To wit:
Example 11.1.8. Let n ∈ N. Let N1 be the poset whose ground set is {1, 2, . . . , n} and whose smaller-or-
equal relation ≤1 is given by
s ≤1 t ⇐⇒ s ≤ t (as integers).
(This is the usual order relation on this set.) Let N2 be the poset whose ground set is {1, 2, . . . , n} and
whose order relation ≤2 is given by
s ≤2 t ⇐⇒ s ≥ t (as integers).
n×n
Let A ∈ k .
(a) The matrix A is upper-triangular if and only if A is triangular when regarded as an N1 × N1 -matrix.
(b) The matrix A is lower-triangular if and only if A is triangular when regarded as an N2 × N2 -matrix.
More interesting examples of triangular matrices are obtained when the order on S is not a total order:
Example 11.1.9. Let S be the poset whose ground set is {1, 2, 3} and whose smaller relation <S is given
by 1 <S 2 and 3 <S2. Then, the triangular S × S-matrices are precisely the 3 × 3-matrices of the form

a1,1 0 0
 a2,1 a2,2 a2,3  with a1,1 , a2,1 , a2,2 , a2,3 , a3,3 ∈ k.
0 0 a3,3
We shall now state some basic properties of triangular matrices:
Proposition 11.1.10. Let S be a finite poset.
(a) The triangular S × S-matrices form a subalgebra of the k-algebra kS×S .
(b) The invertibly triangular S × S-matrices form a group with respect to multiplication.
(c) The unitriangular S × S-matrices form a group with respect to multiplication.
(d) Any invertibly triangular S × S-matrix is invertible, and its inverse is again invertibly triangular.
(e) Any unitriangular S × S-matrix is invertible, and its inverse is again unitriangular.
Exercise 11.1.11. Prove Proposition 11.1.10.
11.1.2. Expansion of a family in another. We will often study situations where two families (es )s∈S and
(ft )t∈T of vectors in a k-module M are given, and the vectors es can be written as linear combinations
of the vectors ft . In such situations, we can form an S × T -matrix out of the coefficients of these linear
combinations; this is one of the ways how matrices arise in the theory of modules. Let us define the notations
we are going to use in such situations:
Definition 11.1.12. Let M be a k-module. Let (es )s∈S and (ft )t∈T be two families of elements of M . (The
sets S and T may and may not be finite.)
Let A = (as,t )(s,t)∈S×T be an S × T -matrix. Assume that, for every s ∈ S, all but finitely many t ∈ T
satisfy as,t = 0. (This assumption is automatically satisfied if T is finite.)
We say that the family (es )s∈S expands in the family (ft )t∈T through the matrix A if
X
(11.1.1) every s ∈ S satisfies es = as,t ft .
t∈T

In this case, we furthermore say that the matrix A is a change-of-basis matrix (or transition matrix ) from
the family (es )s∈S to the family (ft )t∈T .
Remark 11.1.13. The notation in Definition 11.1.12 is not really standard; even we ourselves will occasionally
deviate in its use. In the formulation “the family (es )s∈S expands in the family (ft )t∈T through the matrix
A”, the word “in” can be replaced by “with respect to”, and the word “through” can be replaced by “using”.
The notion of a “change-of-basis matrix” is slightly misleading, because neither of the families (es )s∈S
and (ft )t∈T has to be a basis. Our use of the words “transition matrix” should not be confused with the
different meaning that these words have in the theory of Markov chains. The indefinite article in “a change-
of-basis matrix” is due to the fact that, for given families (es )s∈S and (ft )t∈T , there might be more than
one change-of-basis matrix from (es )s∈S to (ft )t∈T . (There also might be no such matrix.) When (es )s∈S
and (ft )t∈T are bases of the k-module M , there exists precisely one change-of-basis matrix from (es )s∈S to
(ft )t∈T .
222 DARIJ GRINBERG AND VICTOR REINER

So a change-of-basis matrix A = (as,t )(s,t)∈S×T from one family (es )s∈S to another family (ft )t∈T allows
us to write the elements of the former family as linear combinations of the elements of the latter (using
(11.1.1)). When such a matrix A is invertible (and the sets S and T are finite326), it also (indirectly) allows
us to do the opposite: i.e., to write the elements of the latter family as linear combinations of the elements
of the former. This is because if A is an invertible change-of-basis matrix from (es )s∈S to (ft )t∈T , then A−1
is a change-of-basis matrix from (ft )t∈T to (es )s∈S . This is part (a) of the following theorem:

Theorem 11.1.14. Let M be a k-module. Let S and T be two finite sets. Let (es )s∈S and (ft )t∈T be two
families of elements of M .
Let A be an invertible S × T -matrix. Thus, A−1 is a T × S-matrix.
Assume that the family (es )s∈S expands in the family (ft )t∈T through the matrix A. Then:
(a) The family (ft )t∈T expands in the family (es )s∈S through the matrix A−1 .
(b) The k-submodule of M spanned by the family (es )s∈S is the k-submodule of M spanned by the
family (ft )t∈T .
(c) The family (es )s∈S spans the k-module M if and only if the family (ft )t∈T spans the k-module M .
(d) The family (es )s∈S is k-linearly independent if and only if the family (ft )t∈T is k-linearly independent.
(e) The family (es )s∈S is a basis of the k-module M if and only if the family (ft )t∈T is a basis of the
k-module M .

Exercise 11.1.15. Prove Theorem 11.1.14.

Definition 11.1.16. Let M be a k-module. Let S be a finite poset. Let (es )s∈S and (fs )s∈S be two families
of elements of M .
(a) We say that the family (es )s∈S expands triangularly in the family (fs )s∈S if and only if there exists
a triangular S × S-matrix A such that the family (es )s∈S expands in the family (fs )s∈S through the
matrix A.
(b) We say that the family (es )s∈S expands invertibly triangularly in the family (fs )s∈S if and only if
there exists an invertibly triangular S × S-matrix A such that the family (es )s∈S expands in the
family (fs )s∈S through the matrix A.
(c) We say that the family (es )s∈S expands unitriangularly in the family (fs )s∈S if and only if there
exists a unitriangular S × S-matrix A such that the family (es )s∈S expands in the family (fs )s∈S
through the matrix A.
Clearly, if the family (es )s∈S expands unitriangularly in the family (fs )s∈S , then it also expands invertibly
triangularly in the family (fs )s∈S (because any unitriangular matrix is an invertibly triangular matrix).
We notice that in Definition 11.1.16, the two families (es )s∈S and (fs )s∈S must be indexed by one and
the same set S.
The concepts of “expanding triangularly”, “expanding invertibly triangularly” and “expanding unitrian-
gularly” can also be characterized without referring to matrices, as follows:
Remark 11.1.17. Let M be a k-module. Let S be a finite poset. Let (es )s∈S and (fs )s∈S be two families
of elements of M . Let < denote the smaller relation of the poset S, and let ≤ denote the smaller-or-equal
relation of the poset S. Then:
(a) The family (es )s∈S expands triangularly in the family (fs )s∈S if and only if every s ∈ S satisfies
es = (a k-linear combination of the elements ft for t ∈ S satisfying t ≤ s) .
(b) The family (es )s∈S expands invertibly triangularly in the family (fs )s∈S if and only if every s ∈ S
satisfies
es = αs fs + (a k-linear combination of the elements ft for t ∈ S satisfying t < s)
for some invertible αs ∈ k.
(c) The family (es )s∈S expands unitriangularly in the family (fs )s∈S if and only if every s ∈ S satisfies
es = fs + (a k-linear combination of the elements ft for t ∈ S satisfying t < s) .
326We are requiring the finiteness of S and T mainly for the sake of simplicity. We could allow S and T to be infinite, but
then we would have to make some finiteness requirements on A and A−1 .
 HOPF ALGEBRAS IN COMBINATORICS 223

All three parts of Remark 11.1.17 follow easily from the definitions.
Example 11.1.18. Let n ∈ N. For this example, let S be the poset {1, 2, . . . , n} (with its usual order).
Let M be a k-module, and let (es )s∈S and (fs )s∈S be two families of elements of M . We shall identify
these families (es )s∈S and (fs )s∈S with the n-tuples (e1 , e2 , . . . , en ) and (f1 , f2 , . . . , fn ). Then, the family
(es )s∈S = (e1 , e2 , . . . , en ) expands triangularly in the family (fs )s∈S = (f1 , f2 , . . . , fn ) if and only if, for
every s ∈ {1, 2, . . . , n}, the vector es is a k-linear combination of f1 , f2 , . . . , fs . Moreover, the family
(es )s∈S = (e1 , e2 , . . . , en ) expands unitriangularly in the family (fs )s∈S = (f1 , f2 , . . . , fn ) if and only if, for
every s ∈ {1, 2, . . . , n}, the vector es is a sum of fs with a k-linear combination of f1 , f2 , . . . , fs−1 .
Corollary 11.1.19. Let M be a k-module. Let S be a finite poset. Let (es )s∈S and (fs )s∈S be two families
of elements of M . Assume that the family (es )s∈S expands invertibly triangularly in the family (fs )s∈S .
Then:
(a) The family (fs )s∈S expands invertibly triangularly in the family (es )s∈S .
(b) The k-submodule of M spanned by the family (es )s∈S is the k-submodule of M spanned by the
family (fs )s∈S .
(c) The family (es )s∈S spans the k-module M if and only if the family (fs )s∈S spans the k-module M .
(d) The family (es )s∈S is k-linearly independent if and only if the family (fs )s∈S is k-linearly indepen-
dent.
(e) The family (es )s∈S is a basis of the k-module M if and only if the family (fs )s∈S is a basis of the
k-module M .
Exercise 11.1.20. Prove Remark 11.1.17 and Corollary 11.1.19.
An analogue of Corollary 11.1.19 can be stated for unitriangular expansions, but we leave this to the
reader.

Acknowledgements
The authors thank the following for helpful comments and/or teaching them about Hopf algebras: Marcelo
Aguiar, Federico Ardila, Lou Billera, Richard Ehrenborg, Mark Haiman, Florent Hivert, Christophe Hohlweg,
Jia Huang, Jang Soo Kim, Aaron Lauve, Dominique Manchon, John Palmieri, Alexander Postnikov, Margie
Readdy, Nathan Reading, Hans-Jürgen Schneider, Richard Stanley, Muge Taskin, Jean-Yves Thibon.

References
[1] E. Abe. Hopf algebras. Cambridge Tracts in Mathematics 74. Cambridge University Press, Cambridge-New York, 1980.
[2] M. Aguiar, et. al. (28 authors). Supercharacters, symmetric functions in noncommuting variables, and related Hopf algebras.
Adv. Math. 229 (2012), 2310–2337. [Link] . Also available as arXiv:1009.4134v2.
[3] M. Aguiar and F. Ardila. Hopf monoids and generalized permutahedra, arXiv:1709.07504v1.
[4] M. Aguiar, N. Bergeron, and F. Sottile. Combinatorial Hopf algebras and generalized Dehn-Sommerville relations. Compos.
Math. 142 (2006), 1 – 30. A newer version of this paper appears at [Link]
pdf.
[5] M. Aguiar, A. Lauve. The characteristic polynomial of the Adams operators on graded connected Hopf algebras. Algebra
& Number Theory 9-3 (2015), 547–583. Also available at [Link] and as
arXiv:1403.7584v2.
[6] M. Aguiar and S. Mahajan. Monoidal functors, species and Hopf algebras. CRM Monograph Series 29. American Mathe-
matical Society, Providence, RI, 2010. Available at [Link]
[7] M. Aguiar and F. Sottile, Structure of the Malvenuto-Reutenauer Hopf algebra of permutations. Adv. Math. 191 (2005),
225–275. Available at [Link]
[8] N. Andruskiewitsch, W.F. Santos, The beginnings of the theory of Hopf algebras, Acta Appl Math 108 (2009), 3–17. See
also a corrected postprint published on arXiv as arXiv:0901.2460v3.
[9] S. Assaf and P. McNamara. A Pieri rule for skew shapes. J. Combin. Theory, Ser. A 118 (2011) 277–290. [Link]
org/10.1016/[Link].2010.03.010
[10] O. Azenhas, Littlewood-Richardson fillings and their symmetries, Matrices and group representations (Coimbra, 1998),
81–92, Textos Mat. Ser. B, 19, Univ. Coimbra, Coimbra, 1999. [Link]
[11] O. Azenhas, R. C. King, I. Terada, The involutive nature of the Littlewood-Richardson commutativity bijection.
arXiv:1603.05037v1
[12] A. Baker and B. Richter. Quasisymmetric functions from a topological point of view. Math. Scand. 103 (2008), 208–242.
[Link]
[13] F. Barekat, V. Reiner, S. van Willigenburg, Corrigendum to “Coincidences among skew Schur functions” [Adv. Math. 216
(2007), 118–152], Adv. Math. 220 (2009), 1655–1656. See also a corrected version of this paper on arXiv:math/0602634v4.
224 DARIJ GRINBERG AND VICTOR REINER

[14] C. Benedetti, J. Hallam, J. Machacek, Combinatorial Hopf Algebras of Simplicial Complexes, arXiv:1505.04458v2. (Pub-
lished in: SIAM J. Discrete Math. 30 (3), 1737–1757.)
[15] C. Benedetti, B. Sagan, Antipodes and involutions, arXiv:1410.5023v4. (Published in: Journal of Combinatorial Theory,
Series A 148 (2017), 275–315.)
[16] G. Benkart, F. Sottile, J. Stroomer, Tableau Switching: Algorithms and Applications. Journal of Combinatorial The-
ory, Series A 76, 1, October 1996, 11–43. Preprint available at [Link]
[Link]
[17] C. Berg, N. Bergeron, F. Saliola, L. Serrano, M. Zabrocki, A lift of the Schur and Hall-Littlewood bases to non-commutative
symmetric functions. Canad. J. Math. 66 (2014), 525–565. A preprint is arXiv:1208.5191v3.
[18] N. Bergeron, M. Zabrocki, The Hopf algebras of symmetric functions and quasi-symmetric functions in non-commutative
variables are free and co-free. Journal of Algebra and Its Applications 08, Issue 04, August 2009, 581–600. A preprint also
appears at arXiv:math/0509265v3.
[19] L.J. Billera. Flag enumeration in polytopes, Eulerian partially ordered sets and Coxeter groups. Proceedings of the Interna-
tional Congress of Mathematicians IV, 2389–2415, Hindustan Book Agency, New Delhi, 2010. [Link]
edu/~billera/papers/[Link]
[20] L.J. Billera, F. Brenti. Quasisymmetric functions and Kazhdan-Lusztig polynomials, arXiv:0710.3965v2.
[21] L.J. Billera, N. Jia, and V. Reiner, A quasisymmetric function for matroids. European J. Combin. 30 (2009), 1727 – 1757.
[Link]
[22] A. Björner, Some combinatorial and algebraic properties of Coxeter complexes and Tits buildings. Adv. in Math. 52 (1984),
173–212. [Link]
[23] D. Blessenohl, H. Laue. Algebraic combinatorics related to the free Lie algebra. Seminaire Lotharingien de Combinatoire
29 (1992), B29e. [Link]
[24] D. Blessenohl, M. Schocker. Noncommutative character theory of the symmetric group. Imperial College Press 2005.
[25] Ben Blum-Smith, Samuel Coskey. The Fundamental Theorem on Symmetric Polynomials: History’s First Whiff of Galois
Theory. arXiv:1301.7116v4.
[26] N. Bourbaki, Éléments de Mathématique: Groupes et algèbres de Lie, Chapitres 2 et 3. Springer, Heidelberg 2006.
[27] T. Britz, S. Fomin, Finite posets and Ferrers shapes. Adv. in Math. 158, Issue 1, 1 March 2001, 86–127. Better version to
be found on arXiv as arXiv:math/9912126v1.
[28] N.G. de Bruijn, D.A. Klarner, Multisets of aperiodic cycles. SIAM J. Alg. Disc. Math. 3 (1982), no. 3, 359–368. http:
//[Link]/597568
[29] D. Bump, Notes on representations of GL(r) over a finite field, available at [Link]
[30] E. Burgunder, Eulerian idempotent and Kashiwara-Vergne conjecture. Annales de l’institut Fourier 58 (2008), Issue 4,
1153–1184. [Link]
[31] L.M. Butler, A.W. Hales. Nonnegative Hall polynomials. Journal of Algebraic Combinatorics 2 (1993), Issue 2, 125–135.
[Link]
[32] S. Caenepeel, J. Vercruysse. Hopf algebras. Lecture notes, Vrije Universiteit Brussel 2013. [Link]
~scaenepe/[Link]
[33] P. F. Cartier, A primer of Hopf algebras. Frontiers in number theory, physics, and geometry. II, 537–615, Springer, Berlin,
2007.
[34] V. Chari and A. Pressley, A guide to quantum groups. Cambridge University Press, Cambridge, 1994.
[35] K.T. Chen, R.H. Fox, R.C. Lyndon, Free Differential Calculus, IV: The Quotient Groups of the Lower Central Series.
Annals of Mathematics 68 (1), 81–95.
[36] S. Chmutov, S. Duzhin, J. Mostovoy, Introduction to Vassiliev Knot Invariants. CUP 2012.
A preliminary version can also be found at [Link] and at
arXiv:1103.5628v3.
[37] K. Conrad, Expository papers (“Blurbs”), specifically Tensor Products I, Tensor Products II, Exterior Powers, http:
//[Link]/~kconrad/blurbs
[38] H. Crapo and W. Schmitt, Primitive elements in the matroid-minor Hopf algebra. J. Algebraic Combin. 28 (2008), 43–64.
A preprint is arXiv:math/0511033v1.
[39] , A unique factorization theorem for matroids. J. Combin. Theory Ser. A 112 (2005), 222–249. [Link]
10.1016/[Link].2005.02.004
[40] , A free subalgebra of the algebra of matroids. European J. Combin. 26 (2005), 1066–1085. [Link]
[Link]/science/article/pii/S019566980400112X
[41] W. Crawley-Boevey, Lectures on representation theory and invariant theory, Bielefeld 1989/90. Available from http:
//[Link]/~pmtwc/.
[42] Maxime Crochemore, Jacques Désarménien, Dominique Perrin, A note on the Burrows–Wheeler transformation, The-
oretical Computer Science 332 (2005), pp. 567–572. [Link] (A preprint is
arXiv:cs/0502073.)
[43] G. Dahl, Network flows and combinatorial matrix theory. Lecture notes, 4 September 2013. [Link]
emner/matnat/math/MAT-INF4110/h13/lecturenotes/[Link]
[44] S. Dascalescu, C. Nastasescu, S. Raianu, Hopf algebras. An introduction. Monographs and Textbooks in Pure and Applied
Mathematics 235. Marcel Dekker, Inc., New York, 2001.
[45] B. Dayton, Witt vectors, the Grothendieck Burnside ring, and Necklaces. [Link]
[46] T. Denton, F. Hivert, A. Schilling, and N. Thiery. On the representation theory of finite J-trivial monoids. Sém. Lothar.
Combin. 64 (2010/11), Art. B64d, 44 pp. [Link]
 HOPF ALGEBRAS IN COMBINATORICS 225

[47] P. Diaconis, A. Pang and A. Ram. Hopf algebras and Markov chains: Two examples and a theory. J. Algebraic Combin.
39, Issue 3, May 2014, 527–585. A newer version is available at [Link]
[48] W.F. Doran IV. A Proof of Reutenauer’s −q(n) Conjecture. J. Combin. Theory, Ser. A 74 (1996), 342–344. https:
//[Link]/10.1006/jcta.1996.0056
[49] A.W.M. Dress, C. Siebeneicher, The Burnside Ring of the Infinite Cyclic Group and Its Relations to the Necklace Algebra,
λ-Rings, and the Universal Ring of Witt Vectors. Advances in Mathematics 78 (1989), 1–41. [Link]
0001-8708(89)90027-3
[50] G. Duchamp, F. Hivert, and J.-Y. Thibon, Noncommutative symmetric functions VI. Free quasi-symmetric functions
and related algebras. Internat. J. Algebra Comput. 12 (2002), 671–717. A preprint is available at [Link]
[Link]/~hivert/PAPER/[Link].
[51] G. Duchamp, V. Minh, C. Tollu, B. Chien, N. Nghia, Combinatorics of ϕ-deformed stuffle Hopf algebras.
arXiv:1302.5391v7.
[52] T. Dyckerhoff, Hall Algebras - Bonn, Wintersemester 14/15, lecture notes, version February 5, 2015. [Link]
[Link]/people/dyckerho/[Link]
[53] J. Edmonds, Submodular functions, matroids, and certain polyhedra, in: Combinatorial Structures and their Applications
(Proc. Calgary Internat. Conf., Calgary, Alta., 1969), Gordon and Breach, New York, 1970, pp. 66–87; reprinted in
Combinatorial optimization: Eureka, you shrink!, pp. 11–26, Lecture Notes in Comput. Sci. 2570, Springer, Berlin, 2003.
[54] R. Ehrenborg. On posets and Hopf algebras. Adv. Math. 119 (1996), 1–25. [Link]
[55] A.E. Ellis and M. Khovanov. The Hopf algebra of odd symmetric functions. Adv. Math. 231 (2012), 965–999. A newer
version is available as arXiv:1107.5610v2.
[56] P. Etingof, O. Golberg, S. Hensel, T. Liu, A. Schwendner, D. Vaintrob, E. Yudovina. Introduction to representation theory.
Student Mathematical Library 59, Amer. Math. Soc., Providence, RI, 2011. [Link]
. (Parts of this book appear in arXiv:0901.0827v5.)
[57] L. Foissy. Algèbres de Hopf combinatoires. [Link]
[58] L. Foissy, Free and cofree Hopf algebras. Journal of Pure and Applied Algebra 216, Issue 2, February 2012, 480–494.
[Link] . A preprint is arXiv:1010.5402v3.
[59] H. Fredricksen, J. Maiorana, Necklaces of beads in k colors and k-ary de Bruijn sequences. Discrete Mathematics 23 (1978),
207–210. [Link]
[60] W. Fulton. Young Tableaux. London Mathematical Society Student Texts 35, Cambridge University Press, Cambridge-New
York, 1997.
[61] A.M. Garsia, Permutation q-enumeration with the Schur row adder. PU. M. A. (Pure Mathematics and Applications) 21
(2010), No. 2, 233–248. [Link]
[62] V. Gasharov, Incomparability graphs of (3+1)-free posets are s-positive. Proceedings of the 6th Conference on Formal
Power Series and Algebraic Combinatorics (New Brunswick, NJ, 1994). Discrete Math. 157 (1996), 193–197. https:
//[Link]/10.1016/S0012-365X(96)83014-7
[63] V. Gasharov, A Short Proof of the Littlewood-Richardson Rule. European Journal of Combinatorics, Volume 19, Issue 4,
May 1998, Pages 451–453. [Link]
[64] I. Gelfand, D. Krob, A. Lascoux, B. Leclerc, V. S. Retakh, J.-Y. Thibon. Noncommutative symmetric functions. Adv.
Math. 112 (1995), 218–348. [Link]
A preprint is available as arXiv:hep-th/9407124v1.
[65] M. Gerstenhaber, S.D. Schack, The shuffle bialgebra and the cohomology of commutative algebras. Journal of Pure and
Applied Algebra 70 (1991), 263–272. [Link]
[66] I.M. Gessel, Multipartite P-partitions and inner products of skew Schur functions. Combinatorics and algebra (Boul-
der, Colo., 1983), 289–317, Contemp. Math. 34, Amer. Math. Soc., Providence, RI, 1984. [Link]
~gessel/homepage/papers/[Link]
[67] I.M. Gessel, A Historical Survey of P-Partitions, 2014, to appear in a festschrift honoring Richard Stanley.
arXiv:1506.03508v1.
[68] I. Gessel, A. Restivo, C. Reutenauer, A Bijection between Words and Multisets of Necklaces, European Journal of Com-
binatorics 33 (2012), pp. 1537–1546. [Link]
[69] I.M. Gessel, C. Reutenauer. Counting Permutations with Given Cycle Structure and Descent Set. Journal of Combinatorial
Theory, Series A 64 (1993), 189–215. [Link]
[70] I.M. Gessel, X.G. Viennot, Determinants, Paths, and Plane Partitions. preprint, 1989, [Link]
[Link]/xavier/articles_files/determinant_89.pdf
[71] D. Grinberg, Double posets and the antipode of QSym. arXiv:1509.08355v3.
[72] D. Grinberg, A constructive proof of Orzech’s theorem. Preprint, 20 November 2016.
[Link]
[73] A.M. Hamel, I.P. Goulden. Planar Decompositions of Tableaux and Schur Function Determinants. Europ. J. Combinatorics
16 (1995), 461–477. [Link]
[74] M. Hazewinkel, The algebra of quasi-symmetric functions is free over the integers. Adv. Math. 164 (2001), 283 – 300.
[Link]
[75] M. Hazewinkel, Witt vectors. Part 1. In: M. Hazewinkel (ed.), Handbook of Algebra 6, Elsevier 2009. Also available at
arXiv:0804.3888v1.
[76] M. Hazewinkel, The Leibniz-Hopf Algebra and Lyndon Words. Preprint AM CWI 9612 (1996). [Link]
asset/4828/[Link]
226 DARIJ GRINBERG AND VICTOR REINER

[77] M. Hazewinkel, Chen-Fox-Lyndon Factorization for Words over Partially Ordered Sets. Journal of Mathematical Sciences
131 (12-2005), Issue 6, 6027–6031. [Link]
[78] M. Hazewinkel, N. Gubareni, and V.V. Kirichenko, Algebras, rings and modules. Lie algebras and Hopf algebras. Mathe-
matical Surveys and Monographs 168. American Mathematical Society, Providence, RI, 2010.
[79] R. Henderson, The Algebra Of Multiple Zeta Values. [Link]
5432
[80] L. Hesselholt, Lecture notes on Witt vectors. [Link]
[81] L. Hesselholt, The big de Rham–Witt complex. Acta Math. 214 (2015), 135–207. [Link]
s11511-015-0124-y
[82] F. Hivert, An introduction to combinatorial Hopf algebras: examples and realizations. Nato Advanced Study Institute School
on Physics and Computer Science, 2005, october, 17–29, Cargese, France. [Link]
[Link]
[83] F. Hivert, J.-C. Novelli and J.-Y. Thibon, Commutative combinatorial Hopf algebras. J. Algebraic Combin. 28 (2008), no.
1, 65–95. [Link]
Also available as arXiv:math/0605262v1.
[84] , The algebra of binary search trees. Theoret. Comput. Sci. 339 (2005), no. 1, 129–165. [Link]
[Link].2005.01.012
A preprint appears as arXiv:math/0401089v2.
[85] , Trees, functional equations, and combinatorial Hopf algebras. European J. Combin. 29 (2008), no. 7, 1682–1695.
[Link]
A preprint appears as arXiv:math/0701539v1.
[86] M.E. Hoffman, Combinatorics of rooted trees and Hopf algebras. Trans. AMS 355 (2003), 3795–3811. [Link]
10.1090/S0002-9947-03-03317-8
[87] , A character on the quasi-symmetric functions coming from multiple zeta values. The Electronic Journal of Com-
binatorics 15 (2008), R97. [Link]
[88] B. Humpert and J. Martin, The incidence Hopf algebra of graphs. SIAM Journal on Discrete Mathematics 26, no. 2
(2012), 555-570. Also available as arXiv:1012.4786v3.
[89] G. James and M. Liebeck, Representations and characters of groups. 2nd edition, Cambridge University Press, Cambridge-
New York, 2001.
[90] Emma Yu Jin, Outside nested decompositions of skew diagrams and Schur function determinants. European Journal of
Combinatorics 67 (2018), 239–267. [Link]
[91] S.A. Joni and G.-C. Rota, Coalgebras and bialgebras in combinatorics. Studies in Applied Mathematics 61 (1979), 93–139.
[Link]
[92] C. Kassel. Quantum groups. Graduate Texts in Mathematics 155. Springer, Berlin, 1995.
[93] S.V. Kerov. Asymptotic representation theory of the symmetric group and its applications in analysis. Translations of
Mathematical Monographs 219. American Mathematical Society, Providence, RI, 2003.
[94] A.N. Kirillov, A.D. Berenstein, Groups generated by involutions, Gelfand-Tsetlin patterns and the combinatorics of Young
tableaux. Algebra i Analiz 7 (1995), issue 1, 92–152. A preprint is available at [Link]
[95] T. Klein. The multiplication of Schur-functions and extensions of p-modules. J. London Math. Soc. 43 (1968), 280–284.
[Link]
[96] D.E. Knuth. Permutations, matrices, and generalized Young tableaux. Pacific J. Math. 34, Number 3 (1970), 709–727.
[Link]
[97] D.E. Knuth. The Art of Computer Programming, Volume 4A: Combinatorial Algorithms, Part 1. Pearson 2011.
[98] D. Knutson. λ-Rings and the Representation Theory of the Symmetric Group. Lecture Notes in Mathematics 308, Springer,
Berlin-Heidelberg-New York 1973. [Link]
[99] M. Krause, A Simple Proof of the Gale-Ryser Theorem. The American Mathematical Monthly 103 (1996), 335–337.
[Link]
[100] D. Krob. Eléments de combinatoire. Magistère 1-ère année, Ecole Normale Supérieure, version 1.0, Novembre 1995.
[Link]
[101] M. Kufleitner, On Bijective Variants of the Burrows-Wheeler Transform. Presented at the Prague Stringology Conference
2009 (PSC 2009). arXiv:0908.0239v1.
[102] A. Kulikauskas, J. Remmel, Lyndon words and transition matrices between elementary, homogeneous and monomial
symmetric functions. Electronic Journal of Combinatorics 13 (2006), Research Paper ]R18. [Link]
org/ojs/[Link]/eljc/article/view/v13i1r18
[103] K. Kytölä. Introduction to Hopf algebras and representations. Lecture notes, Spring 2011. [Link]
~kkytola/files_KK/lectures_files_KK/Hopf-lecture_notes.pdf
[104] D. Laksov, A. Lascoux, P. Pragacz, and A. Thorup. The LLPT Notes. Edited by A. Thorup, 1995. [Link]
dk/~thorup/notes/[Link]
[105] T. Lam, A. Lauve, and F. Sottile. Skew Littlewood-Richardson rules from Hopf Algebras. DMTCS Proceedings, 22nd
International Conference on Formal Power Series and Algebraic Combinatorics (FPSAC 2010) 2010, 355–366. A preprint
can also be found at arXiv:0908.3714v3.
[106] T. Lam and P. Pylyavskyy. Combinatorial Hopf algebras and K-homology of Grassmanians. International Mathematics
Research Notices, 2007 (2007), rnm 125, 48 pages. A preprint is arXiv:0705.2189v1.
[107] S.K. Lando, On a Hopf Algebra in Graph Theory. Journal of Combinatorial Theory, Series B 80 (2000), 104–121.
[Link]
 HOPF ALGEBRAS IN COMBINATORICS 227

[108] H. Laue, Freie algebraische Strukturen. Lecture notes, Mathematisches Seminar der Universität Kiel 2013, version 16 Sep
2013. [Link]
[109] A. Lauve and S.K. Mason. QSym over Sym has a stable basis. FPSAC 2010, San Francisco, USA. DMTCS proc. AN
2010, 367–378. Also available as arXiv:1003.2124v1.
[110] M. van Leeuwen. Schur functions and alternating sums. Electronic Journal of Combinatorics 11(2) A5 (2006). Also
available at [Link]
[111] M. van Leeuwen. Flag varieties, and interpretations of Young tableau algorithms. Journal of Algebra 224 (2000). Also
available at [Link]
[112] M. van Leeuwen. An application of Hopf-Algebra techniques to representations of finite Classical Groups. Journal of
Algebra 140, Issue 1, 15 June 1991, pp. 210-246. Also available at [Link]
[Link]
[113] M. van Leeuwen. The Littlewood-Richardson rule, and related combinatorics. Math. Soc. of Japan Memoirs 11, Interaction
of Combinatorics and Representation Theory. Also available at [Link]
pdf
[114] M. van Leeuwen. The Robinson-Schensted and Schützenberger algorithms, an elementary approach Electronic Journal of
Combinatorics, Foata Festschrift, 3 (no. 2), R15 (1996). Also available at [Link]
pdf/[Link]
[115] J. Li. Prime Graphs and Exponential Composition of Species. Journal of Combinatorial Theory, Series A 115, Issue 8,
November 2008, 1374–1401. See arXiv:0705.0038v4 for a preprint.
[116] A. Liulevicius. Arrows, symmetries and representation rings. Journal of Pure and Applied Algebra 19 (1980), 259–273.
[Link]
[117] J.-L. Loday. Cyclic Homology. Grundlehren der mathematischen Wissenschaften 301, 2nd edition, Springer, Berlin-
Heidelberg 1998.
[118] J.-L. Loday. Série de Hausdorff, idempotents Eulériens et algèbres de Hopf. Expo. Math. 12 (1994), 165–178. http:
//[Link]/~loday/PAPERS/94Loday%28Eulerien%[Link]
[119] J.-L. Loday and M.O. Ronco. Combinatorial Hopf algebras. Quanta of maths, Clay Math. Proc. 11, 347–383, Amer.
Math. Soc., Providence, RI, 2010. [Link]
[120] , Hopf algebra of the planar binary trees. Adv. Math. 139 (1998), no. 2, 293–309. [Link]
1998.1759
[121] N.A. Loehr, Bijective Combinatorics. CRC Press, 2011. See [Link] for
errata.
[122] M. Lothaire. Combinatorics on words. Corrected printing, Cambridge University Press, 1997.
[123] K. Luoto, S. Mykytiuk, S. van Willigenburg, An introduction to quasisymmetric Schur functions – Hopf algebras,
quasisymmetric functions, and Young composition tableaux. Springer, May 23, 2013. [Link]
papers/[Link]
[124] R.C. Lyndon, On Burnside’s Problem. Transactions of the AMS 77, 202–215. [Link]
S0002-9947-1954-0064049-X
[125] I.G. Macdonald. Symmetric functions and Hall polynomials. 2nd edition, Oxford University Press, Oxford-New York,
1995.
[126] I.G. Macdonald, Schur functions : theme and variations. Publ. I.R.M.A. Strasbourg, 1992, 498/S-27, Actes 28 e Seminaire
Lotharingien, 5–39. [Link]
[127] M. Maia, M. Mendez. On the arithmetic product of combinatorial species. Discrete Mathematics 308, Issue 23, 6 December
2008, 5407–5427. [Link] .
See arXiv:math/0503436v2 for a preprint.
[128] C. Malvenuto, Produits et coproduits des fonctions quasi-symétriques et de l’algèbre des descents, PhD dissertation, Univ.
du Québéc à Montreal, 1993. [Link]
[129] C. Malvenuto and C. Reutenauer, Duality between quasi-symmetric functions and the Solomon descent algebra. J. Algebra
177 (1995), 967–982. [Link]
[130] C. Malvenuto and C. Reutenauer, Plethysm and conjugation of quasi-symmetric functions. Discrete Mathematics 193,
Issues 1–3, 28 November 1998, 225–233. [Link]
[131] C. Malvenuto and C. Reutenauer, A self paired Hopf algebra on double posets and a Littlewood-Richardson rule. Journal
of Combinatorial Theory, Series A 118 (2011), 1322–1333. [Link]
[132] D. Manchon, Hopf algebras, from basics to applications to renormalization. Comptes Rendus des Rencontres Mathema-
tiques de Glanon 2001 (published in 2003). arXiv:math/0408405v2.
[133] L. Manivel, Chern classes of tensor products. arXiv:1012.0014v1.
[134] P.R.W. McNamara, R.E. Ward. Equality of P -partition generating functions. arXiv:1210.2412v2.
[135] Pierre-Loı̈c Méliot, Representation Theory of Symmetric Groups, Discrete Mathematics and its Applications, CRC Press
2017.
[136] A. Mendes, J. Remmel, Counting with Symmetric Functions. Developments in Mathematics 43, Springer 2015.
[137] M. Mendez, MathOverflow answer #139482. [Link]
[138] J. Milnor and J. Moore, On the structure of Hopf algebras. The Annals of Mathematics, Second Series 81, No. 2 (Mar.,
1965), 211–264.
[139] S. Montgomery, Hopf algebras and their actions on rings. Regional Conference Series in Mathematics 82, Amer. Math.
Soc., Providence, RI, 2010.
228 DARIJ GRINBERG AND VICTOR REINER

[140] E. Moreno, On the theorem of Fredricksen and Maiorana about de Bruijn sequences. Advances in Applied Mathematics
33, Issue 2, August 2004, 413–415. [Link]
[141] E. Moreno, D. Perrin, Corrigendum to “On the theorem of Fredricksen and Maiorana about de Bruijn sequences” [Adv.
in Appl. Math. 33 (2) (2004) 413–415]. Advances in Applied Mathematics 62, January 2015, Pages 184–187. [Link]
[Link]/science/article/pii/S0196885814000918
[142] J.L. Martin, M. Morin, J.D. Wagner, On distinguishing trees by their chromatic symmetric functions. Journal of Combi-
natorial Theory, Series A 115, Issue 2, February 2008, 237–253. [Link]
[143] R. Morris, Umbral Calculus and Hopf Algebras. Contemporary Mathematics 6, AMS, Providence 1982.
[144] J. Oxley, Matroid theory. Oxford University Press, Oxford-New York, 1992.
[145] I. Pak, A. Postnikov, Oscillating Tableaux, Sp × Sq -modules, and Robinson-Schensted-Knuth Correspondence, updated
(yet unfinished) version of FPSAC 1996 abstract, January 15, 1994. [Link]
[146] F. Patras. La décomposition en poids des algèbres de Hopf. Annales de l’institut Fourier 43, no 4 (1993), 1067–1087.
[Link]
[147] F. Patras. L’algèbre des descentes d’une bigèbre graduée. Journal of Algebra 170 (1994), 547–566. [Link]
1006/jabr.1994.1352
[148] F. Patras, C. Reutenauer. On Dynkin and Klyachko idempotents in graded bialgebras. Advanced in Applied Mathematics
28, Issues 3–4, April 2002, 560–579. [Link]
[149] F. Patras, C. Reutenauer. Higher Lie idempotents. J. Algebra 222 (1999), no. 1, 51–64. [Link]
1999.7887
[150] R. Patrias. Antipode formulas for combinatorial Hopf algebras. arXiv:1501.00710v2.
[151] V. Prasolov. Problems and theorems in linear algebra. Translations of mathematical monographs 134, 1st edition 1994,
AMS. [Link]
[152] S. Poirier, C. Reutenauer. Algèbres de Hopf de tableaux. Ann. Sci. Math. Québec 19 (1995), no. 1, 79–90. [Link]
[Link]/~christo/Publi%C3%A9s/1995/Alg%C3%A8bres%20de%20Hopf%20de%[Link]
[153] A. Postnikov, Permutohedra, associahedra, and beyond. Int. Math. Res. Notices 2009, No. 6, pp. 1026–1106. A preprint
appears at [Link] and as arXiv:math/0507163v1.
[154] Amritanshu Prasad, An Introduction to Schur Polynomials. arXiv:1802.06073v1.
[155] P. Pylyavskyy, Comparing products of Schur functions and quasisymmetric functions, PhD dissertation, MIT, 2007.
[Link]
[156] D.E. Radford, Hopf algebras. Series on Knots and Everything 49. World Scientific, 2012.
[157] D.E. Radford, A Natural Ring Basis for the Shuffle Algebra and an Application to Group Schemes. Journal of Algebra
58 (1979), 432–454. [Link]
[158] N. Reading. Lattice congruences, fans and Hopf algebras. Journal of Combinatorial Theory, Series A 110, Issue 2, May
2005, pp. 237–273. [Link] . A preprint is arXiv:math/0402063v1.
[159] Victor Reiner, Kristin M. Shaw, and Stephanie van Willigenburg. Coincidences among skew Schur functions.
arXiv:math/0602634v4. (Update of a paper published in Advances in Mathematics, 216(1):118–152, 2007.)
[160] J.B. Remmel. The combinatorics of (k, `)-hook Schur functions. In: C. Greene (ed.), Combinatorics and algebra, Proceed-
ings of the AMS-IMS-SIAM joint summer research conference in the mathematical sciences on combinatorics and algebra,
Colorado, Boulder, 1983, Contemporary Mathematics 34, 1984, 253–287.
[161] C. Reutenauer. Free Lie Algebras. London Mathematical Society Monographs, New Series 7. Clarendon Press, Oxford
1993.
[162] M.H. Rosas, The Kronecker Product of Schur Functions Indexed by Two-Row Shapes or Hook Shapes. Journal of Algebraic
Combinatorics 14 (2001), 153–173. A preprint is arXiv:math/0001084v1.
[163] M.H. Rosas, B.E. Sagan, Symmetric functions in noncommuting variables. Transactions of the American Mathematical
Society 358, 183-214. [Link]
[164] J.S. Kung, G.-C. Rota, Gian-Carlo Rota on Combinatorics: Introductory Papers and Commentaries. Birkhäuser 1995.
[165] B.E. Sagan, The symmetric group: representations, combinatorial algorithms, and symmetric functions. 2nd edition,
Springer, New York-Berlin-Heidelberg 2001.
[166] B.E. Sagan, R.P. Stanley. Robinson-Schensted Algorithms for Skew Tableaux, Journal of Combinatorial Theory, Series
A 55 (1990), 161–193. [Link]
[167] Steven V. Sam, Notes for Math 740 (Symmetric Functions), 27 April 2017. [Link]
pdf
[168] O. Schiffmann. Lectures on Hall algebras. arXiv:math/0611617v2.
[169] W. Schmitt. Incidence Hopf algebras. Journal of Pure and Applied Algebra 96 (1994), 299–330. [Link]
1016/0022-4049(94)90105-8 . A preprint appears at [Link]
[170] W. Schmitt. Antipodes and Incidence Coalgebras. Journal of Combinatorial Theory, Series A 46 (1987), 264–290. https:
//[Link]/10.1016/0097-3165(87)90006-9
[171] W. Schmitt. Expository notes, specifically “A concrete introduction to category theory” and “Notes on modules and
algebras”. [Link]
[172] , Hopf algebras of combinatorial structures. Canadian Journal of Mathematics 45 (1993), 412–428. [Link]
org/10.4153/CJM-1993-021-5 . A preprint appears at [Link]
[173] I. Schur. Arithmetische Eigenschaften der Potenzsummen einer algebraischen Gleichung. Compositio Mathematica 4
(1937), 432–444. [Link]
[174] C. Schweigert. Hopf algebras, quantum groups and topological field theory. Lecture notes, Winter term 2014/15, Hamburg.
Version of 16 May 2015. [Link]
 HOPF ALGEBRAS IN COMBINATORICS 229

[175] J.-P. Serre. Linear representations of finite groups. Springer, Berlin-Heidelberg-New York, 1977.
[176] J. Shareshian and M.L. Wachs, Chromatic quasisymmetric functions and Hessenberg varieties. In: A. Björner, F. Cohen,
C. De Concini, C. Procesi, M. Salvetti (Eds.), Configuration Spaces, Publications of the Scuola Normale Superiore 14,
Springer, Berlin-Heidelberg-New York 2013. A preprint is arXiv:1106.4287v3.
[177] J. Shareshian and M.L. Wachs, Chromatic quasisymmetric functions, Advances in Mathematics 295, pp. 497–551. A
preprint is arXiv:1405.4629v2.
[178] S. Shelley-Abrahamson, Hopf Modules and Representations of Finite Groups of Lie Type. Honors thesis, Stanford, May
2013. [Link]
[179] A.I. Shirshov, On Free Lie Rings, Mat. Sbornik N.S. 45 (87), (1958), no. 2, 113–122. Original at: [Link]
ru/msb4963. Translation in: L.A. Bokut, V. Latyshev, I. Shestakov, E. Zelmanov (eds.), Selected works of A.I. Shirshov,
Birkhäuser 2009.
[180] R.P. Stanley, Ordered structures and partitions. Memoirs of the Amer. Math. Soc. 119, American Mathematical Society,
Providence, R.I., 1972. [Link]
[181] , Acyclic orientations of graphs. Discrete Math. 5 (1973), 171–178. Reprinted in: Discrete Math. 306 (2006),
905–909. [Link]
[182] , A symmetric function generalization of the chromatic polynomial of a graph, Adv. Math. 111 (1995), 166 – 194.
[Link]
[183] , Enumerative Combinatorics, Volumes 1 and 2. Cambridge Studies in Advanced Mathematics, 49 and 62. Cam-
bridge University Press, Cambridge, 2nd edition 2011 (volume 1) and 1st edition 1999 (volume 2).
[184] Shishuo Fu, V. Reiner, Dennis Stanton, Nathaniel Thiem. The negative q-binomial. The Electronic Journal of Combina-
torics 19, Issue 1 (2012), P36. [Link]
[185] R. Steinberg, A geometric approach to the representations of the full linear group over a Galois field. Trans. Amer. Math.
Soc. 71, (1951), 274–282. [Link]
[186] J. Stembridge. A concise proof of the Littlewood-Richardson rule. The Electronic Journal of Combinatorics 9, 2002, N5.
[Link]
[187] J. Stembridge. Multiplicity-Free Products of Schur Functions. Annals of Combinatorics 5 (2001), 113–121. [Link]
[Link]/~jrs/papers/[Link]
[188] Gilbert Strang, The algebra of Elimination. [Link]
[189] M.E. Sweedler, Hopf algebras. W.A. Benjamin, New York, 1969.
[190] M. Takeuchi, Free Hopf algebras generated by coalgebras. J. Math. Soc. Japan 23 (1971), 561 – 582. http://
[Link]/[Link]/1259849779
[191] H. Tamvakis, The theory of Schur polynomials revisited. Enseign. Math. 58 (2012), 147–163. A preprint appears at
[Link]
[192] J.-Y. Thibon. An Introduction to Noncommutative Symmetric Functions. Cargese lecture, October 2005. J.-P. Gazeau,
J. Nesetril, B. Rovan (eds.): From Numbers and Languages to (Quantum) Cryptography, NATO Security through Sci-
ence Series: Information and Communication Security 7, IOS Press, 2007. Available at [Link]
ARTICLES/cargese_thibon.ps.
[193] N. Thiem and C.R. Vinroot, On the characteristic map of finite unitary groups. Advances in Mathematics 210, Issue
2, 1 April 2007, pp. 707–732. [Link] . A preprint is [Link]
~vinroot/[Link]
[194] H. Thomas, A. Yong, An S3 -symmetric Littlewood-Richardson rule. Math. Res. Lett. 15 (2008), no. 5, 1027–1037.
arXiv:0704.0817v1.
[195] M.L. Wachs. Flagged Schur Functions, Schubert Polynomials, and Symmetrizing Operators. Journal of Combinatorial
Theory, Series A 40 (1985), 276–289. [Link]
[196] B.L. van der Waerden. Algebra, Volume I. Translation of the 7th (German) edition. Springer 2003.
[197] Stijn Vermeeren, Sequences and nets in topology. Version of 11 September 2013. [Link]
mathematics/[Link]
[198] P. Webb. A Course in Finite Group Representation Theory. 23 February 2016. [Link]
[Link]
[199] M. Wildon. Representation theory of the symmetric group. 3 May 2014. [Link]
html
[200] M. Wildon. An involutive introduction to symmetric functions. 1 July 2017. [Link]
[Link]
[201] R. Wisbauer. Coalgebras and Bialgebras. The Egyptian Mathematical Society, The Mathematical Sciences Research
Centre (MSRC) Technical Reports No. 1, 2004. [Link]
[202] Q.R. Xantcha, Binomial Rings: Axiomatisation, Transfer, and Classification, arXiv:1104.1931v3.
[203] A.V. Zelevinsky. Representations of finite classical groups: a Hopf algebra approach. Lecture Notes in Mathematics 869.
Springer-Verlag, Berlin-New York, 1981.
[204] A.V. Zelevinsky, A Generalization of the Littlewood-Richardson Rule and the Robinson-Schensted-Knuth Correspondence,
Journal of Algebra 69 (1981), 82–94. [Link]
[205] G.-S. Zhou, D.-M. Lu. Lyndon words for Artin-Schelter regular algebras. arXiv:1403.0385v1.

E-mail address: darijgrinberg@[Link]

230 DARIJ GRINBERG AND VICTOR REINER

School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA

E-mail address: reiner@[Link]

School of Mathematics, University of Minnesota, Minneapolis, MN 55455, USA