AN INFORMAL INTRODUCTION TO DG CATEGORIES

XIAOFA CHEN, XIAO-WU CHEN

Abstract. In this informal introduction to dg categories, the slogan is that

arXiv:1908.04599v2 [[Link]] 9 Aug 2021

dg categories are more rudimentary than triangulated categories. We recall

some details on the dg quotient category introduced by Bernhard Keller and
Vladimir Drinfeld.

Contents
1. Introduction 1
2. Preliminaries 2
2.1. The definitions and examples 2
2.2. Tensor products and functor categories 3
3. The category of dg modules 5
3.1. DG modules 5
3.2. The suspension and cones 6
3.3. DG bimodules 7
4. The derived category 7
4.1. The definitions and resolutions 7
4.2. Derived functors and recollements 8
5. The dg quotient category 9
5.1. The construction of Drinfeld 9
5.2. An example 10
6. Exact dg categories 11
6.1. The suspensions and cones in general 11
6.2. The definition 12
7. The dg quotient vs Verdier quotient 13
7.1. A triangle equivalence 13
7.2. A sketched proof 14
References 16

1. Introduction
Why dg categories? There are two levels of reasons!

Level One.

• a dg category = a dg algebra with several objects

• the dg endomorphism algebras of certain complexes arise naturally in the
derived Morita theory and Koszul duality

August 10, 2021.

1
2 XIAOFA CHEN, XIAO-WU CHEN

• each triangulated category appearing naturally in algebra is essentially a

triangulated subcategory of D(A), the derived category of dg modules over
a dg category A

Level Two.

• the non-functoriality of cones in a triangulated category causes many prob-

lems, e.g., tensor products and functor categories between triangulated cat-
egories are not triangulated
• any exact (= strongly pretriangulated) dg category A has functorial cones,
and gives rise to a triangulated category H 0 (A), by taking cohomologies
• tensor products and functor categories of dg categories are naturally dg
• the homotopy category Hodgcat of small dg categories provides a suitable
framework

For algebraists, the slogan is

dg categories are more rudimentary than triangulated categories!

The main references are [5, 4, 7]. This note might serve as a preparation to read
the survey [7]. The dg quotient category is introduced in [6, 4]. We work over a
commutative ring k.

2. Preliminaries
In this section, we recall basic facts on dg categories.
2.1. The definitions and examples.
Definition 2.1.1. A dg category A is a category subject to
• obj(A): usually denoted by x, y, z · · ·
• each Hom set A(x, y) is a complex, denoted by
M
A(x, y) = ( A(x, y)p , dA ).
p∈Z
p
The element f ∈ A(x, y) is homogeneous of degree p with the notation
|f | = p,
such that the composition
A(y, z) ⊗ A(x, y) −→ A(x, z)
is a chain map. In more details, for any homogeneous morphisms f : x → y and
g : y → z, we have
|g ◦ f | = |g| + |f | and dA (g ◦ f ) = dA (g) ◦ f + (−1)|g| g ◦ dA (f ).
Remark 2.1.2. (1) A nice exercise: each identity endomorphism 1x is homo-
geneous of degree zero satisfying dA (1x ) = 0.
(2) The complex A(x, y) is sometimes denoted by HomA (x, y), the Hom-complex.
In a dg category A, a homogeneous morphism f is closed provided that dA (f ) =
0; a dg isomorphism means a closed isomorphism of degree zero.
We will later omit the word “homogeneous” by the following convention.

In the dg setting, we only consider homogeneous morphisms and elements!

 AN INFORMAL INTRODUCTION TO DG CATEGORIES 3

Definition 2.1.3. For a dg category A, its ordinary category Z 0 (A) is defined as

follows: the same objects with A, and morphisms are given by closed morphisms
of degree zero.
Similarly, the homotopy category H 0 (A) has the same objects whose morphism
spaces are given by the zeroth cohomologies of the Hom-complexes in A.
Remark 2.1.4. (1) H 0 (A) is a factor category of Z 0 (A).
(2) The graded version: Z(A) and H(A).
Definition 2.1.5. For two dg categories A and B, a dg functor F : A → B is a
k-linear functor such that
Fx,y : A(x, y) −→ B(F x, F y)
is a chain map. In other words, |F (f )| = |f | and F (dA (f )) = dB (F f ).
We might define the notion of a dg equivalence, and prove the following theorem:
a dg functor is a dg equivalence if and only if it is fully faithful and dg dense (i.e.
every object in the target category is dg isomorphic to some object in the image).
However, the more useful notions are as follows.
Definition 2.1.6. A dg functor F : A → B is quasi-fully faithful, if each Fx,y is a
quasi-isomorphism. If in addition H 0 (F ) : H 0 (A) → H 0 (B) is dense, F is called a
quasi-equivalence.
Example 2.1.7. (1) A dg algebra = a dg category with a single object. Then
dg functors between these dg categories are just dg algebra homomorphisms.
Accordingly, for each object x ∈ A, A(x, x) is a dg algebra, the dg
endomorphism algebra of x.
(2) For each n ∈ Z, the n-th disc D(n) is defined as follows: the object set is
{x, y} such that D(n)(x, x) = k1x , D(n)(y, y) = k1y , D(n)(y, x) = 0 and
D(n)(x, y) = kδ ⊕ kǫ, with |δ| = −n and |ǫ| = −n + 1. The differential is
determined by d(δ) = ǫ.
The (n − 1)-th sphere dg category S(n − 1) is the non-full subcategory
of D(n) with the same object set and spanned by {1x , 1y , ǫ}.
(3) The opposite dg category Aop : a ◦op b = (−1)|a|·|b|b ◦ a.
(4) The
L dg category Cdg (k) of k-modules: the objects are complexes V =
( n∈Z V n , dV ) of k-modules; the p-th component of the Hom-complex
Hom(V, W ) is given by
Y
Hom(V, W )p = Homk (V n , W n+p )
n∈Z
= {f : V → W | f is homogeneous of degree p}.
The differential is given by
d(f ) = dW ◦ f − (−1)|f | f ◦ dV .
Remark 2.1.8. (1) In Example 2.1.7 (4), the dg algebra End(V ) = Hom(V, V )
is the well-known dg endomorphism algebra of the complex V !
(2) Fact: Z 0 (Cdg (k)) coincides with the abelian category C(k-Mod) of cochain
complexes, and H 0 (Cdg (k)) is the classical homotopy category K(k-Mod)!
2.2. Tensor products and functor categories. The tensor product A⊗B of two
dg categories: the object class is obj(A) × obj(B), whose element (x, y) is denoted
by x ⊗ y, and the Hom-complex is
A ⊗ B(x ⊗ y, x′ ⊗ y ′ ) = A(x, x′ ) ⊗ B(y, y ′ ).
4 XIAOFA CHEN, XIAO-WU CHEN

The Koszul sign rule:

′
(a ⊗ b) ◦ (a′ ⊗ b′ ) = (−1)|b|·|a | a ◦ a′ ⊗ b ◦ b′
and
d(a ⊗ b) = dA (a) ⊗ b + (−1)|a| a ⊗ dB (b).
Remark 2.2.1. The commutativity holds:
(a ⊗ 1y′ ) ◦ (1x ⊗ b) = a ⊗ b = (−1)|a|·|b| (1x′ ⊗ b) ◦ (a ⊗ 1y )
for any a : x → x′ in A and b : y → y ′ in B.
The functor dg category Hom(A, B): the objects are dg functors from A to B;
a morphism η : F → G of degree p between two such functors is given by a family
η = (ηx )x∈obj(A) of morphisms ηx : F (x) → G(x) of degree p in B satisfying

G(a) ◦ ηx = (−1)|η|·|a|ηx′ ◦ F (a)

for all a : x → x′ . These morphisms form the p-th component of the Hom-complex
Hom(F, G). The differential is given by
(dη)x = dB (ηx ).
Lemma 2.2.2. For any dg categories A, B and C, there is a canonical isomorphism
of dg categories
Hom(A ⊗ B, C) −→ Hom(A, Hom(B, C)), F 7→ (x 7→ F (x ⊗ −)).
Example 2.2.3. Let A and B be two dg algebras, viewed as dg categories. Then
A ⊗ B is just the tensor product of dg algebras. The objects in the dg categories
Hom(A, B) are just dg algebra homomorphisms from A to B. For two homomor-
phisms θ, θ′ : A → B and p ∈ Z, we have
Hom(θ, θ′ )p = {b ∈ B p | θ′ (a)b = (−1)|a|·|b| bθ(a) for all a ∈ A}.
In particular, Hom(θ, θ′ ) is a subcomplex of B.
The morphism dg category mor(A) of A is introduced in [4, Subsection 2.9] : the
objects are triples (x, y; a), with x, y ∈ obj(A) and a : x → y a closed morphism of
degree zero; the p-th component of the Hom-complex is given by
 
α 0
Hom((x, y; a), (x′ , y ′ ; a′ ))p = { | α ∈ A(x, x′ )p , β ∈ A(y, y ′ )p , h ∈ A(x, y ′ )p−1 },
h β
with the differential given by
   
α 0 −dA (α) 0
d = .
h β dA (h) + a′ ◦ α − (−1)p β ◦ a dA (β)
The composition is given by the matrix multiplication.
Denote by mor(H 0 (A)) the usual morphism category of the homotopy category
0
H (A) of A.
Lemma 2.2.4. The canonical functor
H 0 (mor(A)) −→ mor(H 0 (A)), (x, y; a) 7→ (x, y; ā)
is full and dense, and its kernel ideal is square zero.
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 5

Proof. For the fullness, just notice that a commutative square in H 0 (A) is given
by a square in Z 0 (A), which is commutative  up to homotopy. 
dA (u) 0
Take two morphisms in the kernel ideal: : (x, y; a) → (x′ , y ′ ; a′ )
h dA (v)
dA (u′ )
 
0
and : (x′ , y ′ ; a′ ) → (x′′ , y ′′ ; a′′ ). Using dA (h) = dA (v) ◦ a − a′ ◦
h′ dA (v ′ )
dA (u) and dA (h′ ) = dA (v ′ ) ◦ a′ − a′′ ◦ dA (u′ ), we have
dA (u′ ) −dA (u′ ) ◦ u
     
0 dA (u) 0 0
◦ = d ,
h′ dA (v ′ ) h dA (v) h̃ v ′ ◦ dA (v)

where h̃ = v ′ ◦ h − h′ ◦ u + v ′ ◦ a′ ◦ u. 

We mention that the morphism dg category might be viewed as a certain gluing

[9, Section 4] of dg categories.

3. The category of dg modules

In this section, the dg category A is required to be small.

3.1. DG modules.
Definition 3.1.1. A left dg A-module is a dg functor M : A → Cdg (k). The
dg category of left dg A-modules is given by A-DGMod = Hom(A, Cdg (k)). The
Hom-complex between two dg modules M and M ′ is denoted by HomA (M, M ′ ).
By definition, a right dg A-module is a left dg Aop -module. Then we have the
dg category DGMod-A of right dg A-modules.
By default, modules are left modules. Later, we will justify the following view-
point.

dg modules over dg categories are better viewed as modules than functors!

L 3.1.2. A left dg A-module M : A → Cdg (k) is identified with the formal

Remark
sum x∈obj(A) M (x) of complexes. Here, by a formal sum, we really mean a
complex graded by obj(A) rather than a coproduct of complexes. The left A-action
on M is defined as follows: for a : x → y and m ∈ M (x),
a.m = M (a)(m) ∈ M (y).
Recall that M (dA (a)) = d(M (a)), where d denotes the differential in Cdg (k). It is
a nice exercise to prove that it is equivalent to the Leibniz rule
dM (a.m) = dA (a).m + (−1)|a| [Link] (m).
We use the notation A M for the left dg module.
Similarly,
L a right dg A-module N : Aop → Cdg (k) is identified with the formal
sum x∈obj(A) N (x). We define the right A-action on N by

n.a = (−1)|n|·|a|N (a)(n) ∈ N (x)

for n ∈ N (y). The Leibniz rule for N has the following form
dN (n.a) = dN (n).a + (−1)|n| [Link] (a).
We use the notation NA for the right dg module.
6 XIAOFA CHEN, XIAO-WU CHEN

Example 3.1.3. For each object y, the free dg A-module A(y, −) is identified with
L
x∈obj(A) A(y, x), whose left A-action is given by the composition of morphisms.
We observe that in general, A(y, −) is not a projective object in the abelian category
Z 0 (A-DGMod).
For each A M , we have a canonical isomorphism of complexes

HomA (A(y, −), M ) −→ M (y), f 7→ fy (1y ) ∈ M (y).

Consequently, the Yoneda dg functor

Aop −→ A-DGMod, y 7→ A(y, −).

is fully faithful. A nice exercise: the Yoneda functor is a dg functor!

3.2. The suspension and cones. As we will see, the dg category A-DGMod has
rich structures: the suspension dg functor and functorial cones.
For a complex X, we denote by Σ(X) the suspended complex: Σ(X)p = X p+1
and dΣ(X) = −dX . For each element x ∈ X p+1 , the corresponding element in
Σ(X)p will be denoted by Σ(x), meaning that |Σ(x)| = |x| − 1.
WARNING: Σ(x) is just a symbol to remember the change of degrees, and is
not induced by any action on X!
For A M , we have the suspended dg module Σ(M ) defined as follows: as com-
plexes, Σ(M )(x) = Σ(M (x)) for each object x, and the new A-action is given
by
a.Σ(m) = (−1)|a| Σ(a.m)
for m ∈ M (x) and a ∈ A(x, y). For each morphism f : M → M ′ between dg
modules, we set Σ(f ) : Σ(M ) → Σ(N ) by Σ(f )x = (−1)|f | fx . Then we have the dg
endofunctor
Σ : A-DGMod −→ A-DGMod,
called the suspension dg functor.
For a closed morphism f : M → M ′ of degree zero, its mapping cone Cone(f )
is the dg A-module defined as follows: as a graded A-module, we have Cone(f ) =
M ′ ⊕ Σ(M ); the differential is given such that Cone(f )(x) = Cone(fx ). In other
words,
 
dM ′ f
dCone(f ) = .
0 dΣ(M)
Here, we implicitly use the identification Σ(M ) with M , and obtain f : Σ(M ) =
M → M ′ . The identification Σ(M ) → M is actually a closed isomorphism of degree
one. This will be clearer in the general consideration later.
The following diagram

(10) (0,1)
(3.2.1) M′ h / Cone(f ) / Σ(M )
j
(1,0) (01)

is a biproduct in the category of graded A-modules. The solid arrows are closed
morphisms, and the dotted arrows in general are not closed.
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 7

3.3. DG bimodules.
Definition 3.3.1. A dg A-B-bimodule is a dg functor X : A ⊗ B op → Cdg (k). The
complex X(u ⊗ y) is usually denoted by X(y, u) for u ∈ obj(A) and y ∈ obj(B).
Consequently, X(y, −) is a left dg A-module and X(−, u) is a right dg B-module.
Remark 3.3.2. The dg bimodule A XB might be viewed as a bunch of left dg A-
modules indexed by obj(B), and in the same time a bunch of right dg B-modules
indexed by obj(A). Strictly speaking, the notations A X and XB are not well-
defined.
Example 3.3.3. For a dg functor F : A → B, we have the dg A-B-bimodule F B1
given by: F B1 (y, u) = B(y, F (u)). Similarly, we have the dg B-A-bimodule 1 BF .
For a dg A-B-bimodule X, we have the Hom dg functor
HomA (X, −) : A-DGMod −→ B-DGMod.
The left dg B-module HomA (X, M ) is described as follows: for y ∈ obj(B), we have
Hom
L A (X, M )(y) = HomA (X(y, −), M ), so, it is identified with the formal sum
y∈obj(B) HomA (X(y, −), M ); the left B-action on HomA (X, M ) is induced by the
right B-action on X, that is, for a morphism b : y → y ′ in B, f : X(y, −) → M and
x ∈ X(y ′ , −), we have
(b.f )(x) = (−1)|b|·(|f |+|x|)f (x.b).
The tensor dg functor is given by
X ⊗B − : B-DGMod −→ A-DGMod.
L
Here, the tensor product X ⊗B N is identified with u∈obj(A) X(−, u) ⊗B N . The
complex X(−, u) ⊗B N is defined by
(⊕y∈obj(B) X(y, u) ⊗ N (y))/hx.b ⊗ n − x ⊗ [Link].
We observe a dg isomorphism
X ⊗B B(y, −) ≃ X(y, −)
of left A-modules.
Lemma 3.3.4. There is a canonical isomorphism of complexes
HomA (X ⊗B N, M ) −→ HomB (N, HomA (X, M )).

4. The derived category

The standard reference for the derived category of a dg category is [5].
4.1. The definitions and resolutions. The ordinary category Z 0 (A-DGMod)
is abelian. Here, we observe that a closed morphism of degree zero f : M → M ′
necessarily respects the grading, differentials and A-actions. An exact sequence
f g
0 −→ X −→ Y −→ Z −→ 0
in Z 0 (A-DGMod) is graded-split, provided that there exists t ∈ HomA (Z, Y )0 satis-
fying g ◦t = 1Z , or equivalently, there exists s ∈ HomA (Y, X)0 satisfying s◦f = 1X .
Proposition 4.1.1. The category Z 0 (A-DGMod) is a Frobenius exact category,
whose conflations are given by graded-split short exact sequences. The resulting
stable category coincides with H 0 (A-DGMod).
Proof. The ingredient is the canonical adjoint triple between Z 0 (A-DGMod) and
the abelian category of graded A-modules; see [5, Lemma 2.2]. 
8 XIAOFA CHEN, XIAO-WU CHEN

We will write K(A) = H 0 (A-DGMod), the homotopy category of dg A-modules.

A dg A-module M is acyclic if each complex M (x) is acyclic. These dg modules
form a triangulated subcategory Kac (A).
Definition 4.1.2. The derived category of left dg A-modules
D(A) = K(A)/Kac (A).
It turns out that D(A) is compactly generated by the modules A(x, −). The
perfect derived category of dg A-modules is by definition the subcategory perf (A) =
D(A)c of compact objects.
The following is a fundamental theorem due to Keller; see [5, Theorem 4.3].
Theorem 4.1.3. Each algebraic compactly generated triangulated category T is of
the form D(A).
Proof. Lift the subcategory T c of compact objects to a dg category! 
To actually work with D(A), we need dg-projective resolutions and dg-injective
resolutions.
A dg A-module P is dg-projective, if for each surjective quasi-isomorphism π : X →
Y and every morphism f : P → Y both in Z 0 (A-DGMod), there is a lifting
f˜: P → X, also in Z 0 (A-DGMod).
Lemma 4.1.4. A dg A-module P is dg-projective if and only if P is projective as
a graded A-module and HomK(A) (P, N ) = 0 for any acyclic module N .
The dg-projective resolution is as follows.
Lemma 4.1.5. Each dg module X admits a surjective quasi-isomorphism
p(X) −→ X
with p(X) dg-projective, which is unique up to a unique isomorphism in K(A).
Denote by Kp (A) the homotopy category of dg-projective modules. Then the
dg-projective resolution yields a triangle equivalence
p : D(A) −→ Kp (A).
Dually, we have the dg-injective resolution X → i(X) and the equivalence
i : D(A) −→ Ki (A).
To summarize, we have a recollement
t p
u
(4.1.1) Kac (A) inc / K(A)
i
can / D(A).
j
i

4.2. Derived functors and recollements. We will discuss derived functors.

resolutions make life easier, but not really easier!

For a dg A-B-bimodule X, we have

p X⊗ − can
X ⊗LB − : D(B) −→ K(B) −−−B−→ K(A) −→ D(A)
and
i HomA (X,−) can
RHomA (X, −) : D(A) −→ K(A) −−−−−−−−→ K(B) −→ D(B).
They form an adjoint pair (X ⊗LB −, RHomA (X, −)).
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 9

Example 4.2.1. Let F : B → A be a dg functor. Then we have an adjoint triple

u F∗
D(A) i / D(B),
F!

where F∗ = 1 AF ⊗LB −, F! = RHomB (F A1 , −) and the unnamed arrow sends M to

the composition M F . We observe that it is isomorphic to both RHomA (1 AF , −)
and F A1 ⊗LA −.
Proposition 4.2.2. Assume that j : B → A is a quasi-fully faithful dg functor.
Then there is a recollement
u j∗
w
(4.2.1) Kg inc / D(A)
i
res / D(B).
j!

Moreover, there is a dg category C such that K is triangle equivalent to D(C).

Consequently, we have a triangle equivalence up to direct summands
perf (A)/perf (B) −→ perf (C).
Proof. We observe that j∗ is fully faithful, since it extends H 0 (j). The last state-
ment follows from Theorem 4.1.3. 

The following problem will be central for us:

what is the above dg category C?

5. The dg quotient category

In this section, we recall the explicit construction of Drinfeld on the dg quotient
category; see [4, Section 3]. It solves the above problem in an elegant way.

5.1. The construction of Drinfeld. Let B ⊆ A be a full dg subcategory. Denote

by j : B → A the inclusion.
Definition 5.1.1. The dg quotient category A/B is defined as follows:
• obj(A/B) = obj(A);
• freely add new morphisms εU : U → U of degree −1 for each U ∈ obj(B),
and set d(εU ) = 1U .
Denote by π : A → A/B the canonical functor.
Remark 5.1.2. For any objects x and y, we have
M M
A/B(x, y) = A(Un , y)⊗kεUn ⊗· · ·⊗kεU2 ⊗A(U1 , U2 )⊗kεU1 ⊗A(x, U1 ).
n≥0 Ui ∈obj(B)

This is only a decomposition of graded k-modules, not of complexes. Indeed, it

gives rise to an ascending filtration
0 ≤1 ≤2
A(x, y) = A/B (x, y) ⊆ A/B (x, y) ⊆ A/B (x, y) ⊆ · · ·
n
of subcomplexes, whose factors are denoted by A/B (x, y) or HomnA/B (x, y). We
observe that each object U in B becomes contractible in A/B.
Under mild assumptions, the above dg quotient category describes the dg cate-
gory C in Proposition 4.2.2; see [4, Proposition 4.6(ii)].
10 XIAOFA CHEN, XIAO-WU CHEN

Theorem 5.1.3. Assume that A(x, U ) is a dg-projective k-module for each x ∈

obj(A) and U ∈ obj(B). Then there is a recollement
t π∗ j∗
u
(5.1.1) D(A/B) can / D(A)
i
res / D(B),
j
π! j!

where “can” sends M to the composition M π. Consequently, we have a triangle

equivalence up to direct summands
perf (A)/perf (B) −→ perf (A/B).
Proof. It suffices to prove that Im can = Ker res and that “can” is fully faithful.
We observe Ker res = {A X | the restriction X|B is acyclic}. Then Im can ⊆
Ker res. Conversely, let A X ∈ Ker res be dg-projective. The assumption implies
that the complex X(U ) of k-modules is dg-projective, and thus contractible. Take
a contracting homotopy sU of degree −1 for each U ∈ obj(B). Using the universal
property, we lift X to a dg module over A/B.
For the fully-faithfulness of “can”, we observe that can(f ) is an isomorphism if
and only if so is f . We apply Lemma 5.1.4 to each X ∈ Ker res. It follows that
Cone(θX ) ∈ Im j∗ ∩ Ker res = 0. So, θX is an isomorphism. This yields another
proof of the above equality.
Now, for each dg A/B-module Y , the counit δY : π∗ can(Y ) → Y has to be an
isomorphism, since can(δY ) is an isomorphism. 

The key observation is as follows.

Lemma 5.1.4. For each A X, the cone of the unit θX : X → canπ∗ (X) lies in
Im j∗ .
Proof. The general case follows from the case where X = A(x, −). We observe
canπ∗ (X) = A/B(x, −). Then the statement follows from the following exact se-
quence in Z 0 (A-DGMod)
inc ≥1
0 −→ A(x, −) −→ A/B(x, −) −→ A/B (x, −) −→ 0,
where the rightmost term is generated by A(U, −)’s, using the assumption. 

Remark 5.1.5. In view of [7, Theorem 4.11 (iii)], there is an exact sequence
j π
B −→ A −→ A/B
in Hmo, the homotopy category of small dg categories with respect to the Morita-
model structures. We mention that Hmo is equivalent to a full subcategory of
Hodgcat.
5.2. An example. We assume that k is a field. Let A be an ordinary k-algebra.
Fix a set {Pi | i ∈ Λ} of two-term complexes of finitely generated projective right
A-modules. Denote by π : A → B the universal localization.
b
The complexes {Pi | i ∈ Λ} form a full dg subcategory B of Cdg (Aop -proj).
b op
Denote by Γ the dg endomorphism algebra of AA in Cdg (A -proj)/B.
Proposition 5.2.1. There is a recollement
u t
D(Γ) i / D(A-Mod) / D(B),
j
Consequently, Γ is non-positively graded with H 0 (Γ) ≃ B. Furthermore, π is a
homological epimorphism if and only if Γ is quasi-isomorphic to B.
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 11

Proof. In view of Theorem 5.1.3, it suffices to identify D(A) = D(A-Mod) with

b
D(Cdg (Aop -proj)), D(Γ) with D(Cdg
b
(Aop -proj)/B); see Lemma 6.2.8. 
Let us consider a special case, where B is given by a single stalk complex eA for
some idempotent e in A. Then the dg algebra Γ is isomorphic to the tensor algebra
TA (Ae ⊗ kε ⊗ eA),
with |ε| = −1 and d(e ⊗ ε ⊗ e) = e. We observe that H 0 (Γ) = A/AeA, H −1 (Γ) =
Ker(Ae ⊗eAe eA → A), and H −p (Γ) = ToreAe p−1 (Ae, eA) for p ≥ 2. Then we recover
the classical result: the ideal AeA is stratifying if and only if the natural projection
A → A/AeA is a homological epimorphism.

6. Exact dg categories
In this section, we justify the following fact: dg categories have functorial cones.
Indeed, an exact dg category has intrinsic suspensions and cones.
6.1. The suspensions and cones in general. The following notions are taken
from [2, Section 3].
Definition 6.1.1. For x ∈ obj(A), its suspension means an object x′ together with
a closed isomorphism ξx : x′ → x of degree one. Notation: x′ = Σ(x), as it is unique
up to a unique dg isomorphism.
Lemma 6.1.2. An object y is dg isomorphic to Σ(x) if and only if A(−, y) is dg
isomorphic to ΣA(−, x), as a right dg A-module.
Proof. A(−, y) ≃ ΣA(−, x) if and only if there is a closed isomorphism A(−, y) ≃
A(−, x) of degree one. Now, use the Yoneda embedding A ֒→ DGMod-A. 
Assume that each object has a suspension. Then we have a fully faithful dg
functor
Σ : A −→ A, x 7→ Σ(x).
For a morphism a : x → y, we have Σ(a) = (−1)|a| (ξy )−1 ◦ a ◦ ξx . We call Σ the
suspension dg functor.
Let f : x → y be a morphism in Z 0 (A). Assume that ξx : Σ(x) → x is the chosen
isomorphism.
Definition 6.1.3. The cone of f is defined to be an object Cone(f ) together with
a diagram
j p
yg / Cone(f ) / Σ(x)
j
t s

such that
(C1) the diagram is a biproduct in A0 , the underlying category with morphisms
of degree zero;
(C2) d(j) = 0 = d(p);
(C3) f = t ◦ d(s) ◦ ξx−1 = −d(t) ◦ s ◦ ξx−1 .
Remark 6.1.4. (1) Assume (C1) and (C2). A nice exercise: prove that (C3)
is equivalent to any of the following identities: d(s) = j ◦ f ◦ ξx , d(t) =
−f ◦ ξx ◦ p.
(2) The cone is unique up to a unique dg isomorphism. We refer to [8, Sub-
section 2.4, the second paragraph] for a characterization of Cone(f ) using
a pair of maps with a universal property.
12 XIAOFA CHEN, XIAO-WU CHEN

Lemma 6.1.5. An object z is dg isomorphic to Cone(f ) if and only if A(−, z) is

dg isomorphic to Cone(A(−, f ) : A(−, x) → A(−, y)).
Proof. We use (3.2.1) and the Yoneda embedding A ֒→ DGMod-A. 
As promised, we prove that cones are functorial in a dg category.
Proposition 6.1.6. Assume that each object in A has a suspension and each closed
morphism of degree zero has a cone. Then we have the cone dg functor
Cone : mor(A) −→ A, (x, y; f ) 7→ Cone(f ).
 
α 0
It sends a morphism : (x, y; f ) → (x′ , y ′ ; f ′ ) to
h β
 
β h ◦ ξx
: Cone(f ) −→ Cone(f ′ ),
0 (−1)|α| Σ(α)
where we identify Cone(f ) with y ⊕ Σ(x), Cone(f ′ ) with y ′ ⊕ Σ(x′ ).
 
β h ◦ ξx
Proof. Just remember that really means (−1)|α| s′ ◦ Σ(α) ◦ p +
0 (−1)|α| Σ(α)
j ′ ◦ β ◦ t + j ′ ◦ h ◦ ξx ◦ p. 
 
α 0
Remark 6.1.7. For a closed morphism of degree zero, we obtain a com-
h β
mutative diagram in Z 0 (A) as follows:
j p
y / Cone(f ) / Σ(x)
 
β h ◦ ξx
β 0 Σ(α)
Σ(α)
j′ p′
  
y′ / Cone(f ) / Σ(x).

6.2. The definition. The following notion, due to [6, Section 2], should be central.
Definition 6.2.1. A dg category A is exact provided that
(1) each object has a suspension, and Σ : A → A is dg dense;
(2) each closed morphism of degree zero has a cone.
Exact dg categories are also called strongly pretriangulated dg categories; see
Remark 7.1.4 and compare Definition 7.1.3.
Lemma 6.2.2. The dg category A is exact if and only if the essential image of the
Yoneda embedding A ֒→ DGMod-A is closed under Σ±1 and cones.
Proof. We use Lemmas 6.1.2 and 6.1.5. 
The following fact justifies the importance of exact dg categories; see [6, Lemma 2.3].
Proposition 6.2.3. Let A be an exact dg category. Then Z 0 (A) has a canonical
Frobenius exact structure, whose stable category coincides with H 0 (A). Therefore,
H 0 (A) is canonically triangulated.
Furthermore, let F : A → B be a dg functor between two exact dg categories.
Then Z 0 (F ) : Z 0 (A) → Z 0 (B) is an exact functor preserving projective objects.
Consequently, H 0 (F ) : H 0 (A) → H 0 (B) is naturally a triangle functor.
Proof. We observe Z 0 (A) is a full subcategory of Z 0 (DGMod-A). Each conflation
in Z 0 (A) is given by a sequence x → e → y, which is a part of a biproduct in A0 .
We observe that H 0 (A) ⊆ perf (Aop ) is a triangulated subcategory.
The second statement follows immediately, since everything is intrinsic. 
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 13

Definition 6.2.4. The pretriangulated hull Apretr is the smallest dg subcategory of

DGMod-A containing A, closed under Σ± and cones. As Apretr is exact, we define
the triangulated hull of A to be Atr = H 0 (Apretr ).
Remark 6.2.5. (1) The dg category A is exact if and only if the canonical
embedding canA : A → Apretr is a dg equivalence.
(2) The inclusion Atr ⊆ perf (Aop ) is a triangle equivalence up to direct sum-
mands.
(3) There is a concrete description of the pretriangulated hull; see [1] and [4,
Subsection 2.4].
Lemma 6.2.6. The canonical embedding canA : A → Apretr has the following uni-
versal property: for any dg functor F : A → B into an exact dg category B, there
is a dg functor F̃ : Apretr → B with a natural dg isomorphism θ : F ≃ F̃ canA ;
moreover, the pair (F̃ , θ) is unique in the obvious sense.
The dg functor F̃ is a dg equivalence if and only if F is fully faithful and Im F
generates B as an exact dg category.
Proof. Set F̃ = (canB )−1 F pretr . 

Remark 6.2.7. The “correct” statement for the universal property is as follows;
see [7, Subsection 4.5]: for any dg category B, there is a dg equivalence
∼
Hom(Apretr , B) −→ Hom(A, B), G 7→ GcanA .
The following easy fact is useful.
Lemma 6.2.8. Assume that j : B → A is a fully-faithful dg functor such that j pretr
is a dg equivalence. Then the restriction functor
res: D(A) −→ D(B)
is a triangle equivalence.

7. The dg quotient vs Verdier quotient

In this section, we recall a fundamental theorem [4, Theorem 3.4], which shows
that the dg quotient category enhances the classical Verdier quotient category.

7.1. A triangle equivalence. The following triangle equivalence shows that:

dg quotient enhances Verdier quotient!

Theorem 7.1.1. Let B ⊆ A be a full dg subcategory. Assume that HomA (x, U ) is

homotopically flat for each x ∈ obj(A) and U ∈ obj(B). Then the canonical functor
∼
Φ : Atr /B tr −→ (A/B)tr
is a triangle equivalence.
Remark 7.1.2. (1) By duality, the equivalence also holds if HomA (U, x) is
homotopically flat for each x ∈ obj(A) and U ∈ obj(B).
(2) Assume that HomA (x, U ) is dg-projective for each x ∈ obj(A) and U ∈
obj(B), or dually, HomA (U, x) is dg-projective for each x ∈ obj(A) and
U ∈ obj(B). Thanks to Remark 6.2.5(2), the above equivalence follows
from the equivalence in Theorem 5.1.3.
14 XIAOFA CHEN, XIAO-WU CHEN

The following notions, due to [1], are very convenient for the study of the homo-
topy category Hodgcat. By definition, Hodgcat is the localization of dgcat with
respect to quasi-equivalences.
Definition 7.1.3. (1) A dg category A is pretriangulated, if the canonical em-
bedding A → Apretr is a quasi-equivalence, or equivalently, H 0 (A) is a
triangulated subcategory of H 0 (DGMod-A) via the Yoneda embedding.
(2) A dg enhancement of a triangulated category T is a pretriangulated dg
category A together with a triangle equivalence E : T → H 0 (A).
Remark 7.1.4. Let us try to clarify the confusion of terminologies in the lit-
erature. Exact dg categories are called +-pretriangulated in [1, p.105, Remark],
strongly pretriangulated in [4, p.650] and [2, p.1475], and pretriangulated in [7,
Subsection 4.5]. The common terminology in Definition 7.1.3(1) is taken from [4,
p.650] and [2, p.1475].
Remark 7.1.5. (1) Let F : A → B be a quasi-equivalence. Then A is pretri-
angulated if and only if so is B.
(2) A triangulated category has a dg enhancement if and only if it is algebraic.
The uniqueness of dg enhancements is an active topic; see [3].
In practice, the following immediate consequence will be useful.
Corollary 7.1.6. Let A be a pretriangulated dg category and B ⊆ A an pretrian-
gulated dg full subcategory. Assume that HomA (x, U ) is homotopically flat for each
x ∈ obj(A) and U ∈ obj(B). Then A/B is pretriangulated. Moreover, the canonical
functor
∼
H 0 (A)/H 0 (B) −→ H 0 (A/B)
is an isomorphism between triangulated categories.
Proof. We identify H 0 (A) with Atr , H 0 (B) with B tr . For the isomorphism, we
mention that the canonical functor acts on objects by the identity. 
As the bounded derived category of a module category is the central object in
homological algebra, the following concept should be central in dg categories: it
provides a canonical dg enhancement for the bounded derived category.
b
Example 7.1.7. Let E be an exact category. Denote by Cdg (E) the dg category
b,ac
of bounded complexes in E, and by Cdg (E) the dg subcategory formed by acyclic
complexes. Then the dg quotient
b,ac
Dbdg (E) = Cdg
b
(E)/Cdg (E)
is called the bounded dg derived category of E. By Corollary 7.1.6, under suitable
conditions, the category Dbdg (E) is pretriangulated, and we have an isomorphism of
triangulated categories
∼
Db (E) −→ H 0 (Dbdg (E)).
Therefore, the dg derived category canonically enhances the derived category.
7.2. A sketched proof. We now sketch the proof. For details, we refer to [4,
Section 8].
Proof of Theorem 7.1.1. It suffices to show that Φ is fully faithful. Furthermore,
it suffices to show that the natural map
ExtiAtr /Btr (x, y) −→ Exti(A/B)tr (x, y)
is an isomorphism for i ∈ Z, and x, y ∈ obj(A).
 AN INFORMAL INTRODUCTION TO DG CATEGORIES 15

Set Qy to be the filtered category of Atr -morphisms α : y → z with Cone(α) ∈

tr
B . Then

LHS = colimα∈Qy ExtiAtr (x, z)

= colimα∈Qy H i HomApretr (x, z).

WARNING: Here, H i and colim can not commute, since Qy is a not filtered
category of morphisms in Z 0 (Apretr )!

RHS = H i HomA/B (x, y)

= H i HomApretr /B (x, y).

For every α : y → z ∈ Qy , we assume α = ᾱ′ , α′ ∈ Z 0 (Apretr ). So, Cone(α′ ) lies

in B pretr up to homotopy. We have a quasi-isomorphism

HomApretr /B (x, y) −→ HomApretr /B (x, z),

as HomApretr /B (x, Cone(α′ )) is acyclic (up to homotopy, Cone(α′ ) is an iterated

extension of objects in B; each U ∈ obj(B) is contractible in Apretr /B!).
So, we have
RHS = colimα∈Qy H i HomApretr /B (x, z).
Recall the following exact sequence

0 −→ HomApretr (x, z) −→ HomApretr /B (x, z) −→ Hom≥1

Apretr /B (x, z) −→ 0.

Taking H i and then colimα∈Qy , we will be done by proving

colimα∈Qy H i Hom≥1
Apretr /B (x, z) = 0.

Using the filtration, the above equality follows from the following claim

colimα∈Qy H i HomnApretr /B (x, z) = 0

for each n ≥ 1. We observe that

M
HomnApretr /B (x, z) = HomApretr (U, z) ⊗ Fx,U ,
U∈obj(B)

where the complex Fx,U is homotopically flat. But, we do have

colimα∈Qy H i HomApretr (U, z) = ExtiAtr /Btr (U, y) = 0.

Then the claim follows from Lemma 7.2.1, a general fact. 

Lemma 7.2.1. Let {Cα }α∈Λ be a filtered system of objects in K(k-Mod). As-
sume that colimα∈Λ H i (Cα ) = 0 for each i ∈ Z, and that F is a homotopically flat
complex. Then colimα∈Λ H i (Cα ⊗ F ) = 0 for each i ∈ Z.

Remark 7.2.2. The assumptions on the Hom-complexes in Theorems 5.1.3 and

7.1.1 might be “removed”. Indeed, it suffices to take a cofibrant replacement of
A. Similarly, to define the dg quotient category in the most general case, we
have to replace A by its cofibrant replacement and then freely add the contracting
homotopies as in Definition 5.1.1.
16 XIAOFA CHEN, XIAO-WU CHEN

References
[1] A.I. Bondal, and M.M. Kapranov, Enhanced triangulated categories, Mat. Sb. 181 (5)
(1990), 669–683.
[2] A.I. Bondal, M. Larsen, and V.A. Lunts, Grothendieck ring of pretriangulated categories,
IMRN 29 (2004), 1461–1495.
[3] A. Canonaco, and P. Stellari, A tour about existence and uniqueness of dg enhancements
and lifts, J. Geom. Phys. 122 (2017), 28–52.
[4] V. Drinfeld, DG quotients of DG categories, J. Algebra 272 (2) (2004), 643–691.
[5] B. Keller, Derived DG categories, Ann. Sci. École Norm. Sup. (4) 27(1) (1994), 63–102.
[6] B. Keller, On the cyclic homology of exact categorues, J. Pure Appl. Algebra 136 (1) (1999),
1–56.
[7] B. Keller, On differential graded categories, International Congress of Mathematicians. Vol.
II, 151–190, Eur. Math. Soc., Zürich, 2006.
[8] B. Keller, Deformed Calabi-Yau completions, with an appendix by Michel Van den Bergh,
J. reine angew. Math. 654 (2011), 125–180.
[9] A. Kuznetsov, and V.A. Lunts, Categorical resolutions of irrational singularities, Inter.
Math. Res. Not. IMRN 13 (2015), 4536–4625.

Key Laboratory of Wu Wen-Tsun Mathematics, Chinese Academy of Sciences, School

of Mathematical Sciences, University of Science and Technology of China, Hefei 230026,
Anhui, PR China
Email address: cxf2011@[Link], xwchen@[Link]