A GENERALIZATION OF INJECTIVE MODULES

SEBASTIANO TRONTO

Abstract. The underlying abelian group of the field of rational numbers Q has an
interesting property: it is divisible, which means that for every x ∈ Q and every positive
integer n there is a y ∈ Q such that ny = x. On the other hand, if we only care about
dividing by the powers of a certain prime, then also the underlying abelian group of the
ring Z[p−1 ] has a similar property: it is p-divisible, that is for every x ∈ Z[p−1 ] there is
y ∈ Z[p−1 ] such that py = x. If one tries to generalize these concepts to modules over a
general (associative, unitary) ring R, things may not work so well, among other things
due to the possible presence of zero-divisors in the base ring. There is however a natural
(or categorical) concept that works well over any ring, which is injectivity. Indeed an
abelian group is divisible if and only if it is injective as a Z-module. What is in this
setting a suitable generalization for p-divisibility? Is there a more general property that
includes divisibility and p-divisibility as special cases, and that also works well for R-
modules? In this talk I will propose a definition that provides a positive answer to the
two questions above. If time permits I will also show an analogue of Morita duality using
this more general definition.

1. Divisible abelian groups and injective modules

Consider the abelian group Q. If x ∈ Q and n ∈ Z \ {0}, then there is y ∈ Q such that
ny = x; namely, we can take y = nx . This holds also, for example, for the abelian group
Q/Z. In general, an abelian group satisfying this property is called divisible.
Definition 1.1. An abelian group A is called divisible if for every x ∈ A and every
n ∈ Z \ {0} there is y ∈ A such that ny = x.
For modules over a general ring R this definition might not scale so well. For example,
taking R = Z × Z, the R-module M = Q × Q (with action of R given by multiplication
component-wise) does not satisfy the property above for every x ∈ M : if x = (1, 1) and
r = (0, 1) then there is clearly no y ∈ M such that ry = x.
There is however a property that plays the same role in many circumstances.
Definition 1.2. An R module Q is called injective if for every injective R-module ho-
momorphism i : M ,→ N and every R-module homomorphism f : M → Q there is an
R-module homomorphism g such that g ◦ i = f .
f
M Q
i
g
N
For Z-modules being injective is equivalent to being divisible.
Proposition 1.3. A Z-module is injective if and only if it is divisible as an abelian group.
Proof. Let A be an abelian group and assume that it is injective as a Z-module. Let x ∈ A
and n ∈ Z \ {0}. Consider the inclusion i : nZ ,→ Z and the map f : nZ → A which sends
n to x. Then since A is injective f extends to a map g : Z → A which sends n to x, so
letting y = g(1) we have ny = x, as required.
1
2 SEBASTIANO TRONTO

Assume now that A is divisible and let J : M ,→ N be and injective homomorphism

of abelian groups and f : M → A any homomorphism. In order to extend f to a map
g : N → A we will use Zorn’s Lemma. Let S be the set of pairs (N 0 , ϕ) with N 0 a sugroup
of N containing M and ϕ a homomorphism N 0 → A that extends f . The set S admits a
partial order
(N 0 , ϕ) 6 (N 00 , ψ) ⇐⇒ N 0 ⊆ N 00 and ψ|N 0 = ϕ
Every chain in S has an upper bound. Namely, if C ⊆ S is a chain, i.e. a totally ordered
subset of S, then we can take N 0 to be the union of all N 0 for (N 0 , ϕ) ∈ C and we let
Φ : N0 → A
x 7→ ϕ(x), if there is any (N 0 , ϕ) ∈ C with x ∈ N 0
which is well-defined because C is totally ordered (which means that if x belongs to N 0 and
to N 00 for (N 0 , ϕ) ∈ C and (N 00 , ψ) ∈ C, then either (N 0 , ϕ) 6 (N 00 ψ) or (N 00 , ψ) 6 (N 0 , ψ),
and in any case ϕ and ψ are compatible on x).
By Zorn’s lemma there is then a maximal element (N 0 , ϕ) ∈ S, and we want to show
that N 0 = N , so that f extends to the whole N . Assume that N 0 6= N and let x ∈ N \ N 0 ;
if we manage to extend ϕ to ϕ+ : N 0 + hxi → A this will yield a contradiction with the
maximality of (N 0 , ϕ), and thus we would be able to conclude that indeed N 0 = N .
If hxi ∩ N 0 = 0, we may simply define ϕ+ (x) = 0. Otherwise hxi ∩ N 0 contains some
nx 6= 0 for some positive integer n which we may assume minimal with respect to this
property. Since A is divisible there is y ∈ A such that ny = ϕ(nx), and one easily checks
that defining ϕ+ as ϕ+ (x) = y is compatible with ϕ. As explained above, this concludes
the proof. 
For a prime number p, the abelian groups Z[p−1 ] and Z[p−1 ]/Z have a property similar
to the divisibility of Q and Q/Z, but only if we restrict to dividing by (powers of) p.
Definition 1.4. Let p be a prime number. An abelian group A is called p-divisible if for
every x ∈ A there is y ∈ A such that py = x.
Is there any property of R-modules that generalizes p-divisibility, in a way similar to
how injectivity generalizes divisibility?

2. Division in modules
Fix for this and the following sections a unitary ring R.
Definition 2.1. If M ⊆ N are left R-modules and I is an ideal of R, we call the R-
submodule of N
(M :N I) := {x ∈ N | Ix ⊆ M }
the I-division module of M (in N ).
Notice that (M :N 0) = N and (M :N R) = M . If I 0 ⊇ I we have (M :N I 0 ) ⊆ (M :N I).
In general we might want to work with (possibly infinite) unions of division modules.
For example if R = Z we are interested in working with
[ 
M :N (pk )
k>0
or with
[
(M :N (n))
n>1
So it makes sense to give the following definition.
 A GENERALIZATION OF INJECTIVE MODULES 3

Definition 2.2. An ideal filter of R is a non-empty set J of two-sided ideals of R such

that:
(1) If I, I 0 ∈ J then I ∩ I 0 ∈ J and
(2) If I ∈ J and I 0 / R contains I, then I 0 ∈ J.
If J is an ideal filter of R and M ⊆ N are R-modules, we let
[
(M :N J) := (M :N I)
I∈J
which we call the J-division module of M in N , and
M [J] := (0 :M J)
which we call the J-torsion submodule of M .
Notice that if the zero ideal belongs to an ideal filter J, then every ideal of R belongs
to J, that is J is the maximal ideal filter. We will denote this ideal filter by 0, and we will
denote the minimal ideal filter {R} by 1. We have (M :N 0) = N and (M :N R) = M ,
and if J 0 ⊆ J we have (M :N J 0 ) ⊆ (M :N J).
Given a set of ideals S of R, we may consider the ideal filter J generated by S, that is
the minimal (with respect to inclusion) ideal filter of R that contains S. If S = {I} we
have (M :N J) = (M :N I).
Example 2.3. For any unitary ring R, there are two interesting examples: the ideal filter
generated by the powers of a given prime number p
p∞ := {I / R | I ⊇ pn R for some n ∈ N}
and the one generated by all non-zero integers
b := {I / R | I ⊇ nR for some n ∈ N>0 } .
n
Notice that some power of p is equal to 0 in R (respectively n = 0 for some n ∈ Nn>0 )
then p∞ (resp. n
b) is simply the set of all two-sided ideals of R.
Thus ideal filters allow us to consider the possibly infinite unions of division modules
mentioned above. We would also like to have a way to distinguish those ideal filters that
describe a complete iteration of the division process, as p∞ and n b do and (n) or (pk ) do
not. We propose two definition that might capture this concept, and we show that, under
certain condition, one is stronger than the other.
Definition 2.4. We call an ideal filter J of R:
• Complete if for every left R-module N and every submodule M ⊆ N we have
((M :N J) :N J) = (M :N J) .
• Product-closed if for any I, I 0 ∈ J we have II 0 ∈ J.
Proposition 2.5. Let J be a product-closed ideal filter of R. If every ideal in J is finitely
generated, then J is complete.
Proof. Let J be a product-closed ideal filter of R and let M ⊆ N be left R-modules. The
inclusion (M :N J) ⊆ ((M :N J) :N J) is always true, so in order to show that equality
holds we need to prove the other inclusion. Let x ∈ N be such that there is I ∈ J with
Ix ⊆ (M :N J). Let {y1 , . . . yn } be a set of generators for I. Then for every i = 1, . . . n
there isT an ideal Ii ∈ J such that Ii yi x ⊆ M . By definition of ideal filter we have
I 0 := ni=1 Ii ∈ J and since J is product-closed we have I 0 I ∈ J. But we also have
I 0 Ix ⊆ M , which shows that J is complete. 
4 SEBASTIANO TRONTO

The ideal filters introduced in Example 2.3 are both product-closed. If, for example, R
is Noetherian, then they are also complete.

3. J-injective modules
Fix for this section a unitary ring R and a complete ideal filter J of R.
Definition 3.1. An injective R-module homomorphism i : M ,→ N such that (i(M ) :N J) =
N is called a J-extension.
We can finally give our definition of J-injective module. In words, one can say that an
injective module is one that admits extensions of maps into it along any injective map. A
J-injective module is one that admits extensions of maps into it along J-extensions.
Definition 3.2. A left R-module Q is called J-injective if for every J-extension i : M ,→ N
and every R-module homomorphism f : M → Q there exists a homomorphism g : N → Q
such that g ◦ i = f .
f
M Q
(J-extension) i
g
N
Notice that in case J = 0 the definition of J-injective module coincides with that of
injective module, because any injective homomorphism is a 0-extension. Moreover, if J 0 is
an ideal filter of R such that J 0 ⊆ J, then a J-injective module is also J 0 -injective, because
every J 0 -extension is also a J-extension.
The following Proposition is an analogue of the well-known Baer’s criterion in the
classical case of injective modules, and the proof is almost identical to the classical case.
Proposition 3.3. A left R-module Q is J-injective if and only if for every I ∈ J and every
R-module homomorphism f : I → Q there is an R-module homomorphism g : R → Q that
extends f .
Proof. The “only if” part is trivial, because any two-sided ideal of R is also a left R-
module. For the other implication, let i : M ,→ N be a J-extensions and let f : M → Q
be any R-module homomorphism. By Zorn’s Lemma there is a submodule N 0 of N and an
extension g 0 : N 0 → Q of f to N 0 that is maximal in the sense that it cannot be extended
to any larger submodule of N . If N 0 = N we are done, so assume that N 0 6= N and let
x ∈ N \ N 0.
Let I be the two-sided ideal of R generated by {r ∈ R | rx ∈ N 0 }. Since i(M ) ⊆ N 0 and
(i(M ) :N J) = N there is I 0 ∈ J such that I 0 x ⊆ N 0 , which implies I 0 ⊆ I, so also I ∈ J.
By assumption the map I → Q that sends y ∈ I to g 0 (yx) extends to a map h : R → Q.
Since ker(R → Rx) is contained in ker(h), the map h gives rise to a map h0 : Rx → Q by
sending rx ∈ Rx to h(r). By definition the restrictions of g 0 and h0 to N 0 ∩ Rx coincide, so
we can define a map g 00 : N 0 +Rx → Q that extends both. This contradicts the maximality
of g 0 , so we conclude that N 0 = N . 
Remark 3.4. One can adapt the proof or Proposition 1.3 to show that an abelian group
is p-divisible if and only if it is p∞ -injective (see Example 2.3) as a Z-module.
Let J = 0 be the maximal ideal filter of R and assume that J 0 = J \ 0 is an ideal
filter; this amounts to say that no two non-zero ideals of R have trivial intersection. Using
Proposition 3.3 one can easily show that an R-module Q is J-injective if and only if it is
J 0 -injective. Indeed, one implication holds, as remarked above, because J ⊆ J 0 , and for
 A GENERALIZATION OF INJECTIVE MODULES 5

the other it is enough to notice that the only map 0 → Q can always be extended to the
zero map on R.
One advantage of using J 0 instead of J is that the J 0 -torsion submodule may be different
from M [0] = M .
Example 3.5. Let M be an abelian group, let p be a prime and let J = p∞ be the ideal
filter of Z introduced in Example 2.3. Then the localization M [p−1 ] is a J-injective Z-
module. Indeed if i : N ,→ P is a J-extension and f : N → M [p−1 ] is any homomorphism
k
then for every x ∈ P there is k ∈ N such that pk x ∈ N , and one can define g(x) := f (ppk x) .
It is easy to check that g is then a well-defined group homomorphism such that g ◦ i = f .

4. Injective hulls and J-hulls

Definition 4.1. A map of R-modules i : M ,→ N is called an essential extension if for
every nonzero submodule P of N we have P ∩ i(M ) 6= 0.
It is a well-known fact of commutative algebra that every R-module M admits an
injective hull ι : M ,→ Γ, which is an essential extension such that Γ is injective. Such an
extension, which is unique up to a not-necessarily-unique isomorphism that is the identity
on M , may be as well characterized by either of the following two properties:
(1) It is the largest essential extension of M , that is to say if i : M ,→ N is an essential
extension then there is an (injective) R-module homomorphism j : N ,→ Γ such
that ι ◦ i = j (the injectivity of j follows from the injectivity of ι and the fact that
i : M ,→ N is an essential extension).
(2) It is the smallest injective extension of M , that is to say if i : M ,→ N is an
injective R-module homomorphism and N is injective, then there is an injective
R-module homomorphism j : Γ ,→ N such that j ◦ ι = i (the existence of a map
Γ → N that commutes with i follows from the injectivity of N , but the fact that
this map is injective does not).
As an example, the standard map Zn ,→ Qn is an injective hull of the Z-module Zn .
There is an analogue construction for J-injectivity.
Definition 4.2. Let J be a complete ideal filter of R and let M be a left R-module. A
J-extension ι : M ,→ Ω is called a J-hull of M if it is an essential extension and Ω is
J-injective.
The following theorem is not a replacement for the classical one, since it relies on it.
Theorem 4.3. Every left R-module M admits a J-hull, which is unique up to a not-
necessarily-unique isomorphism that is the identity on M .
Sketch of proof. Let ι : M ,→ Γ be an injective hull of M and let Ω := (ι(M ) :Γ J). One
can show that ι maps M into Ω and ι : M ,→ Ω is indeed a J-hull of M , and that for any
∼
other J-hull ι0 : M ,→ Ω0 there is an isomorphism j : Ω → Ω0 such that j ◦ ι = ι0 . 
Example 4.4. Let M be an abelian group, let p be a prime and let J = p∞ be the ideal
filter of Z introduced in Example 2.3. Write M as
k
M
r
M =Z ⊕ Z/pei Z ⊕ M [n]
i=1
where n is a positive integer coprime to p and the ei ’s are suitable exponents. Let
Γ = (Z[p−1 ])r ⊕ (Z[p−1 ]/Z)k ⊕ M [n]
6 SEBASTIANO TRONTO

and
ι: Γ →   M  
z s
(z, (si mod pei )i , t) 7→ 1 , e
p i mod Z ,t
i

Then ι : Γ → M is a J-hull. To see this it is enough to show that f : Zr ,→ (Z[p−1 ])r and
gi : Z/pei Z ,→ Z[p−1 ]/Z for every i = 1, . . . , k are J-hulls, and that M [n] is J-injective,
being trivially an essential extension of itself. The assertions about f and M [n] follow
from Example 3.5, noticing that multiplication by p is an automorphism of M [n] and that
Zr ,→ (Z[p−1 ])r is an essential J-extension.
So we are left to show that for every positive integer e the map g : Z/pe Z ,→ Z[p−1 ]/Z de-
fined by (s mod pe ) 7→ ( pse mod Z) is a J-hull. It is a J-extension, because the Prüfer group
Z[p−1 ]/Z itself is J-torsion, and it is also essential because every subgroup of Z[p−1 ]/Z is
of the form p1d Z, so it intersects the image of g in pmin(e,d)
1
Z.
Finally, Z[p−1 ]/Z is divisible as an abelian group, so in particular it is J-injective, since
in this case it is equivalent to being p-divisible.
We can draw an interesting parallel between the J-hull of an R-module M and the
algebraic closure k of a field k. Indeed k is at the same time the smallest algebraically
closed extension and the largest algebraic extension of k. Similarly to J-hulls, an algebraic
closure is unique up to a not-necessarily-unique isomorphism that fixes the base field.

5. Morita duality
Consider the following well-know fact about vector spaces and linear maps.
Proposition 5.1. Let V be a finite dimensional vector space over a field k and let
f1 , . . . , fnT: V → k be linear functions. If g : V → k is a linear function such that
ker(g) ⊇ ni=1 ker(fi ), then g is a linear combination of the fi .
Proof. Consider the map
F := (f1 , . . . , fn ) : V → k n
x 7→ (f1 (x), . . . , fn (x))
Tn
and notice that K := ker(F ) = i=1 ker(fi ). Then both g and F factor through V / ker(F )
as g : V / ker(F ) → k and F : V / ker(F ) → k n respectively, and F is injective. By
extending a basis of Im(F ) ⊆ k n to a basis of k n one can find a linear map
λ : kn → k
(x1 , . . . , xn ) 7→ e1 x1 + · · · + en xn
such that λ ◦ F = g, which implies λ ◦ F = g. Then for every v ∈ V we have
g(v) = λ(F (v)) = λ(f1 (v), . . . , fn (v)) = e1 f1 (v) + · · · + en fn (v)
which shows that g is a linear combination of the fi . 
We can give a much more general version of this result. Fix a ring R, a complete
ideal filter J of R and R-modules M and T , with T being J-injective and J-torsion (i.e.
T [J] = T ). Let E = EndR (T ), and notice that HomR (M, T ) is an E-module.
For every submodule M 0 of M we will identify HomR (M/M 0 , T ) with the
T E-submodule
{f ∈ HomR (M, T ) | ker(f ) ⊇ M } of HomR (M, T ), and we will denote f ∈V ker(f ) by
ker(V ) for any subset V ⊆ HomR (M, T ).
 A GENERALIZATION OF INJECTIVE MODULES 7

Proposition 5.2. If V is a finitely generated E-submodule of HomR (M, T ) we have V =

HomR (M/ ker(V ), T ).
Proof. Notice that the inclusion V ⊆ HomR (M/ ker(V ), T ) is obvious. For the other
inclusion we want to show that every homomorphism g : M → T with ker(g) ⊇ ker(V )
belongs to V . Let then g be such a map and let g : M/ ker(V ) → T be its factorization
through the quotient M/ ker(V ). Let {f1 , . . . , fn } be a set of generators for V as an
E-module and let
ε : M → Tn
x 7→ (f1 (x), . . . , fn (x))
We have ker(ε) = ker(V ), so that ε factors as an injective map ε : M/ ker(V ) → T n . Since
T is J-torsion, so is T n , hence ε is a J-extension. Since T is J-injective there is an R-linear
map λ : T n → T such that λ ◦ ε = g, or equivalently λ ◦ ε = g.
g
T
g

M M/ ker(V ) λ

ε
Tn
ε

Since HomR (T n , T ) ∼
Ln
= i=1 EndR (T ), there are elements e1 , . . . , en ∈ EndR (T ) such that
λ(t1 , . . . , tn ) = e1 (t1 ) + · · · + en (tn ) for every (t1 , . . . , tn ) ∈ T n . Then for x ∈ M we get
λ(ε(x)) = λ(f1 (x), . . . , fn (x))
= e1 (f1 (x)) + · · · + en (fn (x))
which means that g = e1 ◦ f1 + · · · + en ◦ fn ∈ V because V is an E-module. 

From a different point of view, we have two maps

k : {E-submodules of HomR (M, T )} → {R-submodules of M }

V 7 → ker(V )

and

h : {R-submodules of M } → {E-submodules of HomR (M, T )}

M0 7→ HomR (M/M 0 , T )
and the previous proposition shows that the restriction k 0 of k to the subset of finitely
generated E-submodules satisfies h ◦ k 0 = id. It is natural to ask whether the two maps
are inverse of each other, possibly after restricting h to a suitable subset.
Definition 5.3. We say that T is a cogenerator for an R-module N if
\
ker(f ) = 0 .
f ∈HomR (N,T )

Using this definition, we may formulate the following duality statement.

8 SEBASTIANO TRONTO

Theorem 5.4. Let R be a unitary ring and let J be a complete ideal filter on R. Let T
be a J-injective and J-torsion left R-module and let M be a left R-module. Assume that
T is a cogenerator for every quotient of M and that HomR (M, T ) is Noetherian as an
EndR (T )-module. The maps
{R-submodules of M } → {EndR (T )-sumbodules of HomR (M, T )}
M0 7→ HomR (M/M 0 , T )
ker(V ) ←[ V
define an inclusion-reversing bijection between the the set of R-submodules of M and that
of EndR (T )-submodules of HomR (M, T ).
Proof. The maps are clearly inclusion-reversing and the fact that they are inverse of each
other follows from Proposition 5.2 combined with the Noetherianity of M and from the
assumption that T is a cogenerator for every quotient of M . 

Department of Mathematics, University of Luxembourg, 6 av. de la Fonte, 4364 Esch-sur-

Alzette, Luxembourg
Email address: [Link]@[Link]