FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA

SAMUEL EILENBERG AND J. C. MOORE 1)

Introduction
The notion of a derived functor for a fuoctor T: (1'- 93 (with suitable condi-
tions imposed on the categories (j and :.B and on the (unctor T) is one of the key
notions of homological algebra. Since its original definition [4] this notion has
undergone a multitude of generalizations. Some of these generalizations went in
the direction of avoiding the use of projective resolutions, following an idea of
Yoneda [12]. This will not be the point of view adopted here. Granting that the
definition of derived functors is to use projective resolutions, the procedure breaks
up neatly into two quite well separated steps.
The first of these steps is the definition of projective resolutions, their exis-
tence and basic properties. The second step is the definition of the derived func-
tors and the study of their properties. In this paper we shall not be concerned with
the second step at all; this will be deferred to subsequent papers. The first step
will be our main concern here.
In each category, the notion of projective objects is inherent. However it has
been recognized for some time that more latitude in the choice of projective ob-
jects (or equivalently in the choice of exact sequences) should be permitted. Thus
Hochschild [8] in studying the category of modules over an algebra A considered
r-projective modules where r is a subalgebra of A. Heller [7] considered addi-
tive categories with a distinguished class of "proper" morphisms. Buchsbaum [2]
considered abelian categories in which a class of morphisms called an "h. f."
ciass was given subject to a number of natural conditions. Butler and Horrocks
[3] modify and extend Buchsbaum's approach by starting with an abelian category
and a distinguished class of short exact sequences.
The main notion of this paper is that of a "projective class of sequences"
in an arbitrary (pointed) category. Each such class carries with it its own projec-
tive objects. One can then talk about projective resolutions, and if the category
is additive, all the usual properties of the resolutions hold. In particular, this
will permit the development of homological algebra in some additive categories
Received by the editors April 23, 1964.
I)The first author was partially supported by the Office of Naval Research and the
National Science Foundation, while the second author was partially supported by the Air
Force Office of Scientific Research during the period while this research was in progress.
I
2 FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
which are not abelian, e.g., the category of comodules over a coalgebra over an
arbitrary commutative ring.
With the notion of a projective class established in Chapter I, the balance -of
the paper is devoted to methods of finding such projective classes and their pro-
jective objects. The main tool here is the "adjoint theorem" of Chapter 11.
Chapter I begins with a brief review of categories and continues with the def-
inition and some basic properties of projective classes. Chapter 11 begins with a
review of adjoint functors and o n ~ n u s with the adjoint theorem (Chapter 11, 2)
which tums out to be the main tool for constructing projective classes. Chapter
III gives a varied assortment of examples of projective classes in various cate-
gories, all arrived at by the adjoint theorem. Chapter IV studies the category cet
of complexes over an abelian category et and the various sub-categories of cet.
In particular, it is shown that the "double resolutions" of Cartan-Eilenberg [4,
Chapter XVII] are simply resolutions in cet relative to a suitable projective class.
CHAPTER I. RELATIVE HOMOLOGICAL ALGEBRA
1. Review of categories
Let et be a category. We shall use the abbreviated notation et(A, B) instead
of the usual notation Homet(A, B) for the set of morphisms f: A-+ B of objects
A and B in et. The dual category is denoted by et*.
All categories et considered here will be assumed pointed, i.e., containing
an object A
o
such that et(A, A
o
) and et(A
o
' A) consist of single elements for
every object A in [11, p. 504]. For any two such trivial objects A
O
and A~
there is a unique isomorphism A
O
~ A~ in et ,and we shall write 0 for any
object A
o
with the property above. -For any two objects A, A' in et we define
the trivial map A-+ A' as the composition A-+ 0 -+ A'. -The trivial map will be
written as 0A' A or again simply as O.
A typical example is the category S of sets with basepoints with morphisms
being maps of sets preserving the basepoints. -Thus for any category et and for
any objects A, A' in et, et (A, A') is any object in the category S. This yields
the "representation" functor et* x et -... S.
A family Po-: A-...Ao-' a- I of morphisms in a category et is called a prod-
uct if for any family of morphisms f0-: B-+ A0- in et there is a unique morphism
f: B-. A such that Po-f = fo- for all a I. Dually a family io-: A0--+ A, a I
is a coproduct if for each family f0-: A0-- B, a I there exists a unique f: A-+ B
such that fi =f for all a I.
0- 0-
Let et be a category and let Po-: A-+Ao-' a- I be a product. For each
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
a': there exists then a unique morphism A -. A such that p i = lA '
v 0- 0-0- 0-
Prio- = 0 for r f.a.If the family lio-l is a coproduct then we say that
io-
Ao-
I
' A, a:
Po-
3
is a biproduct.
A pre-additive category is a category a such that each set a(A, A') is
given the structure of an abelian group in such a way that composition is distribu-
tive: (gI + g2)f=gd + g21, g(I
I
+1
2
) = gll + gh for 1,11' 1
2
: A'-.A and g, gl'
g2: A-. A". It is easy to see that in a pre-additive category and for a finite set
of indices there is no need to distinguish between products, coproducts and biprod-
ucts. Further one can verify that
i
g
A0". ' A, a= 1, 2
PO"
is a biproduct if and only if
PIi
l
= 1
Ar
P2
i
2 = 1
A2
, ilPI + i
2
P2 = lA
A pre-additive category a in which biproducts exist for any two objects Al
and A
2
is called additive. It can be shown that a category in which biproducts
exist for any two objects Al and A
2
can be converted in at most one way into an
additive category [11, p. 511-512].,
Given morphisms
such that rc = lA" we say that r is a retraction, c is a coretraction and that A'
is a retract of A.
Amorphism [:A'-.A is called an epimorphismif a([, B):a(A,
is injective (as a mapping of sets) for every B in a. Monomorphisms are defined
dually.
Let
(1.1)
be morphisms in a. We say that (1.1) is a sequence if ji = O. We say that i is a
kernel of j if ji = 0 and for any I: B-. A such that j[ =0, [admits a unique
factorization 1= ig. It follows that i is a monomorphism. If i: A' -. A is another
kernel of j then there exists a unique cI>: A' ---. A' such that icl> =, i, and this cl>
cl> is an isomorphism. Cokernels are defined dually.
The sequence (1.1) is said to be exact if i admits a factorization

such that is an epimorphism and k is a kernel of j.If further l is a retraction
4 SAMUEL EILENBERG AND J. C. MOORE
then we say that (1.1) is split exact.
Coaxactness and split coexactness are defined dually.
The notions of "sequence", "exact sequence", etc. are carried over to
longer @agrams
:-+ A
2
- Al - A
O
- A_I - A_
2
-+
(terminating or non-terminating at either end) by applying them to each consecu-
tive pair. In particular a sequence non-terminating in both directions is called a
complex.
It is easy to see that in the category S exactness has the usual meaning and
all exact sequences split because all epimorphisms are retractions. Further in
(1.1) i is a kernel of j if and only if for every 8 (1 the sequence
0- (1(8, A') - (1(8, A) - (1(8, (1")
is an exact sequence in S.
Let T: (1-. $ be a functor. We shall always assume that T(O) = 0 (i.e., T
carries trivial objects into trivial objects). If (1 and $ are pre-additive categories
and T(f
I
+ h) = T(f
I
) + T(f
2
} then we say that T is additive. It can be shown
that if (1 and $ are additive categories and the functor T: (1-. $ preserves
finite biproducts then T is an additive functor.
A functor T: (1 - $ is said to be faithful if fl' f
2
: AI - A in (1 and TU
I
) =
T(f
2
} imply f
l
= f
2

Proposition 1.1. A faithful functor T: (1-. $ reflects epimorphisms and mono-

morphisms, i.e., if f: A-. AI and T(f) is a monomorphism (epimorphism) then f
itself is a monomorphism (epimorphism).
Proof. Let gl' g2: AII_ A be such that fgI = fg
2
Then T(f} T(gl} =
T(fg
2
) = T(f) T(g2)' If T(f) is a monomorphism then T(gI) = T(g2} and since T
is faithful, gI = gr Thus f is a monomorphism.
Proposition 1.2. Let T: (1- $ be a faithful and kemel preserving functor.
If the category (1 has kemels then T reflects exact sequences, i.e., given
A'!.... A.!. A" (1.2)
such that
(1.3)
is an exact sequence, then (1.2) is an exact sequence.
Proof. Since T (ji) = T (J) T (i) = 0 and T is faithful, it follows that ji =0.
There is then a factorization
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
5
of i such that k is a kernel of j. Then T(k) is a kernel of T(j) and since (1.3)
is exact, it follows that T(1) is an epimorphism. Since T is faithful, it follows
that 1 is an epimorphism and (1.2) is exact.
2. dosed and projective classes
Let (1 be a category, P an object of (1 and E a sequence in (1. Then
(1 (P, E) is a sequence in the category S. Given a class & of sequences of (1,
let P be the class of all objects P in (1 for which (1(P, E) is exact {or every
E &. We then write & P. Similarly, given any class P of objects in (1 let
& be the class of all sequences E in (1 such that (1(P, E) is exact for every
P P. We then write P==? &.
Clearly, for any class & we have
and for any class P we have
P=9&=9P=9 &=:;;.." ,
A class & of sequences is called closed if & = &, i.e., if P=9 &.
Similarly, a class P of objects is called closed if P= Pi.e., if P=:;;. & P.
If P=i> & then & is closed and if & P then P is closed.
The objects of P will usually be called &projective, while the sequences
in & are called P-exact.
The proof of the following two propositions is elementary and is left to the
reader.
Proposition 2.1. If P is a closed class of objects in (1, P P and P' is a
retract of P, then P' P.
Proposition 2.2. If P is a closed class of objects in (1 and i
er
: Per -. P,
a I is a coproduct in (1, then P P if Ol7,d only if Po- P for every a I.
Let & be a closed class of sequences in a category (1 and let & P. We
shall say that & is projective if the following condition holds.
For every morphism A -. A' in (1 there exists amorphism
P -.. A with P P and with P-.. A-..A' in &.
(2.1)
All these notions may be dualized by passing to the category (1*. Thus we
shall write & * =9 if &* in the category (1*; we shall then say that the
objects of are &-injective. We shall say that & is an injective class in (1 if
&'" is a projective class in (1"', etc.
H & is a projective class of sequences in (1 and &' is an injective class of
6 SAMUEL EILENBERG AND J. C. MOORE
sequences in athen we say that &, and &" are complementary if the complexes
in &, coincide with the complexes in &". If&, ~ P and &" * ~ ~ then we also
say that P and ~ are complementary.
It should be noted that if &, is an exact ptojective class in an abelian cate-
gory et then the morphisms f: A-> A' such that the sequence
Ker f- A-> A' ->0
is in &, form an h.t class in the sense of Buchsbaum [2]. This establishes a bi-
jection between all the exact projective classes and all the h.f. classes which
possess enough projectives. The details are left to the reader.
3. Resolutions and derived functors
Let &, be a projective class in a category et, and let A be an object of et.
A left complex X over A is a complex
d
n
--. X
n
-> X
n
-
l
-> ........... X
o
-.. 0 -..
together with amorphism f: X
o
-> A such that fd
l
=0. The morphism f is called
the augmentation and frequently is regarded as amorphism f: X-> A of complexes
where A is treated as a complex with A in degree zero and zero l ~ w h r The
sequence --. X -> ..-. X
o
-.. A-.. 0 -> .will be denoted by X.
n ~
The left complex X over A is said to be &'-acyclic if X is in &'. It is said
to be &'-projective if each X
n
is &'projective.If X is both &,projective and
&,acyclic then X (taken together with its augmentation f: X-.. A) is called an
&,-projective resolution of A.
Proposition 3.1. If &, is a projective class in a category et, then every ob
ject of ahas &'.projective resolutions.
Proof. Let A et. Applying (2.1) successively we find sequences in &,
X
o
~ A -+ 0,
dl
Xl -+ X
o
-.. A,
d2 dl
X
2
-> Xl -+ X
o
'
...... ..... ,. ... ,.
such that X
n
is &,projective for n = 0, 1, Combining these into a single
sequence yields the desired resolution.
Proposition 3.2. Let f: X-+ A be an &,-projective left complex over A and
let Tf: Y-+ B be an &'-acyclic left complex over B. Then for any morphism
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
7
f: A--... B there exists amorphism F: X--... Y of complexes such that "IF = fi. If
further the category (f is pre-additive then any two such morphisms F are homo-
topic.
The standard inductive proof [4, p. 76] applies without modification and. will
not be repeated.
Let T: (f- 93 be an additive functor defined on a pre-additive category (!
and with values in an abelian category 93. Let & be a projective class in (f and
let A be an object in (f with an &-projective resolution f: X_ A. Then T(X) is
a complex in 93 and its homology depends only on A. The nth derived functor of
T relative to & is defined as (L& 1') (A ) =H (T (X )). Since these matters will
n n
be discussed in greater generality somewhere else, we shall not enter into this
any further here. If the category (f is additive, then the usual properties of reso-
lutions of sequences 0 -+ A' - A--... A"--... 0 in & hold and we obtain in the usual
way the connecting morphism L ~ T(A
IJ
) - L ~ _ l T(A") with the usual properties.
Let now (f be a pre-additive category. We shall consider (f(A, B) as a func
tor (f x (f* --... M* where M is the category of abelian groups, and M* is its dual.
If & is a projective class in (f then Ext{i;(A, B) is defined as Rn fU, B))
where X is an &projective resolution of A.If &' is an injective class in (f
then Ext{i;,(A, B) is defined as Rn (f(A, Y)) where Y is an &'-injective resolu-
tion of B. If the classes & and &' are complementary then Ext& and Ext&, co-
incide. This fundamental property of Ext is established in the usual way using
the fact that gis in &' andY is in &. One then proves that the complexes
(f (X, B) and (f (A, Y) have the same homology as a suitably defined complex
(f(X, Y).
4. Categories with kernels
In this section (f will denote a category with kernels. Let & be a class of
sequences. Let mbe the class of all morphisms such that the sequence A'-
A_ 0 is in &.We shall write & ~ m.
Let m be a class of morphisms.Given a sequence
A'!..ALA"
we have a factorization
A i ~ A
(4.1)
where k is a kernel of j. Let & be the class of all those sequences (4.1) for
which 1 m.We shall write m==> &.
Proposition 4.1. If m==>& ==> P then P P if and only if (! (P, f) lS sur-
jective for every f m. If P==> &=9 mthen f mif and only if (! (P, f) is
8
SAMUEL EILENBERG AND J. C. MOORE
surjective for every P P. If &, is closed then &, =9 m=9 &,. The closed class
&, with &, =9 P, &, =9 mis projective if and only if for every A (( there exists
amorphism f: P--... A in mwith P P. The class & is exact (i.e., composed of
exact sequences only) if and only if mis composed of epimorphisms.
The proof follows from the observation that
0-. (1(8, A) --... (((8, A) --... (((8, A")
is exact for every object 8. Therefore
(((8, A') --... (((8, A) --... (((8, A")
is exact if and only if (1(8, l) is surjective.
Proposition 4.2. Let m
o
be the class of all retractions in ((, and let
m
o
=9 &0 =9 Po' Then &0 is the class of split exact sequences and is projec-
tive, while Po is the class of all objects of ((.
The proof follows trivially from the observation rhat f is a retraction if and
only if (((8, f) is surjective for every object 8 in (1.
An object P of (( is called projective if (f(P, f) is surjective for all epimor-
phisms f.
Proposition 4.3. Let m
1
be the class of all epimorphisms and let m
1
='9
&1 = :J\. Then &'1 is the class of all exact sequences while PI is the class
of all projective objects in ((. The class &'1 is projective if and only if for every
A (( there exists an epimorphism f: P --... A with P projective.
Proof. We clearly have m
1
&'1 =9 m
1
and &'1 is the class of all exact
sequences. Further &'1 =9 PI and PI is the class of all projective objects.
Suppose now that for every A (( there exists an epimorphism f: P --... A with
P PI' Suppose that (4.1) is in &1' Then (( (P, 1) is surjective for every P PI'
Choose an epimorphism f: P --... Awith P PI' Then (((P, 1): (((P, A') --...
(( (P, A) is surjective and thus f = 1m for some m: P --... A'. Therefore l is an
epimorphism and (4.1) is in &1' The remaining assertions are obvious.
A category (( in which the class &'1 of all exact sequences is projective is
called projectively perfect.
Clearly &1 is the largest possible exact projective class. In the category
S of pointed sets we have &0 = &1' since every epimorphism is a retraction.
Thus S has only one exact projective class. It is easy to see that the only other
projective class in S is the class &, consisting of all sequences in S. The only
&projective objects are the trivial objects consisting of a single point.
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
9
5. Subcategories
Let & be a projective class in a category 11 and let & =9 P. A full subcat-
egory (11 of (1 will be called an &-subcategory if &' = (t' ('\ & is a projective
class for (1'. If further &' =9 (1' ('\ P (in (1') then we say that 11' is a normal
&-subcategory of 11.
If (1' is a normal &-subcategoryof (1 and if (:X'--.A' is an &'_
pro
-
jective resolution of A' (1' in (1' then (: X' --. AI also is an &-projec-
tive resolution of A' in (1. Thus if (1 is a pre-additive category and (i"
is a pre-additive subcategory of (1 (i.e., the inclusion functor ]: (1' --. (1
is additive), then for any additive functor T: Cl' --. with values in an abe-
lian category the derived functors L &, (Tj) coincide with (L&T)].
Proposition 5.1. Let & be a projective class in a category (j, let (i"
be an &-subcategory of (f and let a" be a subcategory of a' which is a
normal &-subcategory of a. Then a" is a normal (Ct' ('\ &)-subcategory
of <1'.
Proof. We have (111 ('\ (a' ('\ &) = a" f\ &, and thus a" is a (a' ('\ &)-subcat-
egory of (f'. If p" in (1" is (a" f\ &)-projective then p" is &-projective and thus
also (a
'
('\ &)-projective.
Proposition 5.2. Let (f' be a full subcategory of (1, let & be a projective
class in (1 and let &=9 P. If Pc (1' then (1' is a normal &-subcategory of (t.
Proof. Let &'= (1' ('\ & and let &'===;;. p' (in (1'). Then PeP'. If E' is a
a sequence in &' such that (1'(P', E') is exact for every P'in pi, then it fol-
lows that E' & and therefore E' &'. Thus the class &' is closed (in (f').
If A'1 -. A'2 is a morphism in (1 then there exists a sequence P-. A'1 -. AI
Z
in
& with P P. Then P P' and condition (2.1) holds. Thus &' is a projective
class in (j". Now let P' P'. There exists then in & a sequence p-Lp '-.0
with P P. Since this sequence is also in &', it follows that a'(p' , f) is sur-
jectivp.. Thus we have fg = 1 for some g: p' -. P. Thus P' is a retract of P and
and consequently P' P. Thus P' = P as required.
6. Resolvents
Given a category (1, we denote by (12 the category in which the objects are
morphisms f: Al -. A
Z
in (1 and in which amorphism (CPl' CPZ) : f-. f' is given
by a commutative diagram
Al
f
A
Z
----+
CPl
1
1CPZ'
A' -....... A'
1
f'
z
10
SAMUEL EILENBERG AND J. C. MOORE
We define the "range" and "domain" functors
R, D: (12 --+ (1
by setting
R(f) = A
2
,
D(f} = AI'
R(c/>l' c/>2) = c/>2'
D(c/>l' c/>2) =c/>l'
Then J(f) = f defines amorphism J: D--+ R.
Let ff, be a projective class in (1 and let ff, ==9 P. A resolvent for ff, is a
functor
satisfying the following conditions:
(R.l) Re= D
(R.2) For any morphism f, De (f) P
(R.3) For any f: A---> A'
De(f) e(f). A LA'
is a sequence in ff,.
Given an object A in (1 let 0 A denote the trivial map 0 A : A--+ O.Define
eo(A) = e(OA)' en(A) =e(e
n
_
1
(A))
for n = 1, 2, "'. Then the sequence
e2(A) . el(A). e O A ~
, , ~ De
2
(A) --. Del (A) -------. De
o
(A) A - 0
is an ff,-projective resolution of A, which is functorial. It is called the canonical
resolution of A relative to the resolvent e.
Proposition 6.1. A functor e: ({2 --+ (12 is a resolvent of a projective class
ff, in (1 if and only if Re = D and for any morphisms f: A--+ AI and g in (1
(1(De(g), De(f}} (1(De(g),e(f. (1(De(g},A) (1(De(g),fl (1 (De (g), A') (6.0
is an exact sequence., The ff,-projective objects of (1 are then the retracts of ob-
jects De (g) where g ranges over all morphisms of (1, Thus ff, is unique.
Proof. If e is a resolvent for ff, then De(g) is &-projective by (R.2) and
thus by (R.3), (6.1) is exact. Conversely assume that (6.1) is exact. Taking
g = f it follows that fe (f) = O. Let P be the class of all retracts of objects
De (g) where g ranges over all morphisms in (1, and let P=9 K Then ff, is a
closed class of sequences and
De(f) e (f). A L ~
is in ff,. Since De (f) IS &-projective, it follows that the class ff, is projective
FOUNDAnONS OF RELAnVE HOMOLOGICAL ALGEBRA
11
F(A) A.
and that e is a resolvent for &.
Now let P be &'projective.Tben taking f: P-O we have the sequence
d(p, De (f}) -> d(p, P) - 0
so that P is a retract of De (f).
Let d be a category with kernels. Choosing a kernel
Dk (f) k (n, A
for any morphism f: A-. A' yields a functor
k: d
2
-. (1'2
such that Rk = D. Moreover, for any object B in (f the sequence
0-. I1(B, Dk(f}) _11(B, A) -. (j(B, A')
is exact. Thus by 6.1, k is a resolvent for a projective class &. As f ranges
overall morphisms of 11, the objects Dk(f} and their isomorphs range through
the class Po of all objects of I1.Tbus by 6.1,& =9 Po and thus from 4.2 it fol-
lows that & = &0 is the class of all split exact sequences in et. We have thus
Proposition 6.2. If 11 is a category with kernels then the kernel functor
k: cF _ (f2 is a resolvent for the class &0 of split exact sequences.
Let d be a category with kernels, and with kernel functor k. Let &be a
projective class in (1 with resolvent e.Then for any f: A- A' in (1 we have
fe (f) = 0 and therefore we have a commutative diagram
De(f) A L A'
a(f)\ I k(f}
Dk(f)
for a unique morphism a(f}.If a(f) = e(g)wbere g: Dk.(f)-O then we say that
the resolvent e factors through the kernel. For any A 11 let 0A : A - 0 and
write e (0A) as
Then F: 11- 11 is a functor and f: F-1(1 a this notation, if
e factors through the kernel then De(f} =F(A) where A = Dk(f) and e(f} is
the composition
F(ii) 1 !.Y.2..... A.
We say that the resolvent e is defined by the pair (F, f), Conditions (R.2) and
(R.3) yield conditions
(R. 2') F (A) P for any A (j
12
SAMUEL EILENBERG AND J. C. MOORE
(R.3
/
) F(A) f(A ), A --. 0 is in & for any A (f.
Conversely if (f has kernels and (F, i) is a pair consisting of a functor
F: (f --.. (f and amorphism f: F--.. 1(:j', satisfying (R.2 ') and (R.3') then the formu
lae above define a resolvent e for &.
Frequently the pair (F, i) will itself be referred as a resolvent for &.
Proposition 6.2. If & is an exact projective class in a category (j' and if
(F, i) is a resolvent, then F is faithful.
Proof. Let fl' h : A--.. A' be such that F([I) = F(f
1
). Since fidA) =
f (A ') F(f
i
) for i = 1, 2, it follows that f
1
i (A) = f
2
i (A). Since the sequence
F(A) A -----. 0 is in & it is exact and thus f (A) is an epimorphism. Thus
f
1
= fz
CHAPTER 11. THE ADJOINT THEOREM
1. Adjoint functors
We shall use the symbol
(a, 13) :.s -1 T: ((!, 53)
to designate the following situation: (f and 53 are categories, T: (f--.. 53,
5: 53 --.. (! are functors,
are morphisms of functors satisfying
(1.1)
We shall say that the functor 5 is the coadjoint of T or that T is the ad
joing of 5. We shall sometimes use the abbreviated notation (a, 13) :5 -1 T or
even just 5 -1 T.
An alternative approach may be obtained as follows. Given functors
T: (f--.. 93 and 5: 93--.. (1 and given amorphism 13: 153 --.. T5 define for A in (f
and B in 53
b:(f(5(B), A)--..53(B, T(A))
by setting
We have
b(l/J) = T(l/J) 13 (B) for l/J: 5(B) -. A.
b(at/f5(r)) = T(a)b(l/J)r for a:A-. A', r:B'-.B
t3(B) = b(I
S
(B)'
(1.2)
(1.3)
(1.4)
Conversely, given b as above satisfying (1.3), formula (1.4) defines amorphism
of functors 13: 153-' T5 and (1.2) holds. Thus the study of 13 may be reduced to
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
that of b and vice-versa.
13
Proposition 1.1. Given functors T: and given amorphism
f3: -- TS there exists a morphism a: ST-+ 1(1 such that (a, (3) : S--1 T: (1. 93)
if and only if for every A in (1 and B in the mapping
b: (1(S(8). A) -. T(A
defined by (1.2) is a bijection. If this is the case then a is dnique and is given
by a(A) = b-
1
(IT(A' The mapping
T(A -. (1(S(8), A)
defined by
a(ep) = a (A) Seep) (1.2*)
is the inverse of b.
Proof. Given a: ST-.1(1 we define a by (1.2*) and find that conditions
(1.1) imply ab = 1 and ba = 1. Conversely if b is a bijection we set a =b-1 and
a(A) = a(lT(A' Then a is a morphism of functors and (1.1) is proved by compu-
tation.
The condition that b is bijective when stated explicitly reads:
Given ep: B -. T(A) there is a unique t/J: S(B) -. A
such that commutativity holds in the triangle
B
TS(8)
1T(t/J)
T(A)
Stated in this way the condition asserts that the morphism f3(B): B-. TS(B)
is "universal" for all morphisms B-. T(A).
Proposition 1.2. If (a, f3): T then passing to the dual catego-
ries we have (f3*. a*): T* S*: (1*).
It suffices to record the duals of (1.1).
Proposition 1.3. Let (a, f3): S T: Then T preserves monomorph-
isms, products, and kernels while S preserves epimorphisms, coproducts and co-
kernels.
Proof. We shall use the notation (1.2). Let f: A-. A' be a monomorphism
and let gl' g2 : B-. T(A) be such that T(f) gl == T(n g2' Since
T(f)gj = T(f}ba(gj) == b(fa(gj
for i = 1, 2, it follows that f
a
(gl) = f a(g2)' Thus a(gl) = a(g2) and gl = g2'
14 SAMUEL EILENBERG AND J. C. MOORE
Let Per: A-. A
er
, a I, be a product. Thus for each B in 93 the natural
mapping
(1(S(B), A) -. x (1(S(B), A )
er er
is a bijection. Applying a to both sides yields that the natural mapping
93(B, T(A)) -. x 93(B, T(A ))
er er
is a bijection and thus T(p er): T(A)-. T(A
er
), a I,is a product.
Now let f: A' -. A be a kernel of g: A-. A". Sincegf =0, we have
T(g) T(n = T(O) = O. Suppose h: B-. T(A) is such that T(g)h = O. Then
b(ga(h)) = T(g) ba(h) = T(g)h = O. Consequently ga(h) = 0 and a(h) admits a
factorization a(h) = fh' with h': S(E) = A'. Then h =ba(h) = b(fh') = T(f}b(h').
Since T(f} is a monomorphism this factorization of h is unique and thus T(n
is a kernel of T(g).
The second half of 1.3 follows by duality.
Proposition 1.4. Let S -+ T: (1, 93) where er and 93 cire additive categories.
Then the functors T and S are additive.
Proof. By 1. 1 and 1. 2, T is a functor which preserves products. It is known
that such functors are additive. Similarly for S.
Proposition 1.5. Let (a, 13):S-4 T:(1, 93). Then the following conditions
are equivalent:
(i) T is faithful.
(H) T reflects epimorphisms.
(iii) a (A) : ST (A) -. A is an epimorphism for every A in (1.
Proof. (i) (ii) follows from 1. 1.
(ii) (Hi). Since (Ta) (f3T) = IT it follows that Ta(A) is an epimorphism.
Thus by (H) a(A) is an epimorphism.
(Hi) =9 (i). Let f
l
, h : A -. A and let T (f
1
) = T (f
2
). Then f
l
a (A) =
a(A')ST(f
l
) = a(A')5T(f
2
) = f2a(A), so that f
l
=1
2
since a(A) is an epimor-
phism.
2. The adjoint theorem
Let
(a, f3): 5 -4 T: (1, 93).
We shall utilize the natural isomorphisms
a: 93(B, T(A) (1 (5(B), A)
with b = a-I, as given in 1. Given any class ff, of sequences in 93 we denote
by T-lff, the class of all sequences E in (1 such that T (E) ff,.
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
15
(2.1)
Theorem 2.1. If & is a projective class in :B then T-
1
& is a projective
class in (f. The T-l&-projective objects of (f are the retracts of objects S(B)
where B ranges over aU the &-projective objects of :B.
We begin by proving
Lemma 2.2. Under the conditions above if f: A -. A' in (f and if
B -!-.. T(A) TCA')
is in & then
is in T-l&.
Indeed we have

fa (g) =a(TC[)g) =a(O) = 0
(2.2)
and thus (2.1) is a sequence. We must show that the sequence
TS(B) T(A) '0!.2. T(A') (2.3)
is in &. Let then h: P-. TCA), where P is &-projective, be such that TCn h = O.
Since (2.1) is in & we have h=gk for some k:P-.B. Since g= Ta(g){3(B),it
follows that h == Ta(g)k' for some 1/: P-. TS(B).Thus
:B(P, TS(B-. :B(P, T(A -. :B(P, T(A"
is exact and consequendy (2.3) is in &.Consequently (2.2) is in T-l& as re-
quired.
Proof of 2.1. Let & ==9 P, &' = T-
1
&==9 P'.If P P and E' &' then
T(E') & and :B(P, T(E' is exact. Therefore applying the isomorphism a we
find that (f(S(P), E') is exact and thus S(P) P'.Thus every retract of S(P)
also is in P'. Next, let E' be a sequence in (f such that (f(P', E') is exact
for every P' P'.Then (f(S(P), E') is exact for every P P. Thus applying
the isomorphism b we find that :B(P, TCE' is exact for every P P. Conse-
quendy T(E') & and thus E' &'. This shows that P' ==9 &' and thus &'
is closed.
For an arbitrary f: A-.A', a sequence (2.1) in & with B P exists since
the class & is projective. Then (2.2) is in &' and S(P) P'. Thus condition I,
(2.1) is fulfilled and the class &' is proj ective.
Finally let P' P'.Then in & we have a sequence
P.!. TCP') -. 0
with P P. Using this for (2.1) we obtain the sequence

in &'. Since P' P', (1.(P', S(P -. (1.(P', P') is surjective. Thus we have
16
SAMUEL EILENBERG AND J. C. MOORE
a(g) 11. = lp' for some 11.: P' -+ S(P). Thus P' is a retract of S(P) and the proof
is complete.
Suppose now in' 2.1, that e: -+ is a resolvent for liJ. Then for every
f: A--> A' in ({ the sequence
DeT(f):!J!.2. T(A)!:!l2. T(A')
is in liJ and DeT(f) is &-projective. It follows from 2.2 that
SDeT(f) a(eT(f. A L A'
is in T-1liJ, while 2.1 implies that SDe T(f) is T-l&-projective.
a(e T (fn yields a resolvent for T-1liJ. This yields
Corollary 2.3. If e is a resolvent for & in then e'(f) = a(eT(f)) for
f: A--> A' in ({ is a resolvent for T-1liJ in ({. Alternatively, e' (f) is the compo-
sition
SDeT(f) SeT(f). ST(A) a(A\ A
If (F, d is a resolvent pair for liJ in then (SFT, (') where / is the com-
position
. l . ( T-l c.
a reso vent palr Jor (!)
Corollary 2.4. If the category ({ has kernels, T is faithful and the class liJ
is exact, then the class T-1liJ is exact. If further T preserves epimorphisms and
liJ is the class of all exact sequences in then T-1liJ is the class of all exact
sequences in ({.
Indeed, by 1.3, T preserves kernels and therefore by I, 1.2, T reflects ex
act sequences. Thus T- lliJ is exact. If further T preserves epimorphisms, then
since it preserves kernels it also preserves exact sequences. This implies the
second statement.
Proposition 2.5. Let liJ be a projective class in a category ({ and let ({' be
a subcategory of ({. If the inclusion ftmctor I: ({' -+ ({ has a coadjoint P--1I,
then ({' is an liJ-subcategory of ({. Further ({' isa normal &-subcategory of ({
if and only if P(A) is liJ-projective if A is liJ-projective.
This follows from 2.1 since ({'(') liJ =r1liJ.
3. The multiple adjoint theorem
A family of functors S0-: 0--> ({, Q E I, is called cointegrable if for any
family of objects Ro-E Q I, the coproductEBSo-(Ro-) exists.
0-
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 17
Theorem 3.1. Let
(ao-' f30-) : So- -+ To-: (f, a' I
where the family ISo-I, u' I, is cointegrable. For each a I, let ff,o- be a pro-
jective class in and let ff,o- =9 Po-. Then the family ff,1t = Q is a pro-
jective class in (f. The ff,1t.projective objects are the retracts of coproducts
E!1
S
o-(P0-)' Po- Po-' a I.
0-
Proof. Let be the product of the categories with projections
p 0-: -. 0-' .and let T: (f -. be the functor defined by the conditions po-T = T 0-'
a I. Let S: (f be the coproduct EB S Po-: (f with injections
0- 0-
io-:So-Po--'S. Then we have
(a, f3) : S -1 T: 1,
where
a(i T) = a
0- 0-
Indeed the relations (1.1) follow by an easy computation.
Let ff, be the class of sequences E in such that Po-(E) &0- for all
a::, and let P be the class of all objects P such that p o-(P) Po- for
all a I. Then ff, is a projective class in and ff, =:> P.Since ff,1t = T-Iff"
the conclusion follows from 2.1-
We shall omit the statement of the analogue of 2.3.
Corollary. If the category (f has kernels, if the family IT0-
1
, a I is col-
lectively faithful (i.e., if f
I
, f
2
: A_ A' in (f and T0-(f
1
) = T0-(f
2
) for all a I
implies f
I
= f
2
) and if each class ff,o- is exact, then the class ff,,1t is exact.
This follows from 2.4 since T: (f-. is faithful.
Taking T0-= So- = la in 3.1 we obtain
Proposition 3.3. Let ff,C7' a':: be a family of projective classes in (f and
let ff, =:> Po-. Suppose that for each family lA I, Ao- <1, a I the coproduct EB Bo-
o- 0- 0-
exists. Then the class ff, = 0 ff, is projective and the ff,-projective objects in
0- 0-
(f are the retracts of coproducts E!1 Po-' Po- Po-' a ::.
0-
4. A refinement
In some rather special situations it is possible to refine the conclusions of
the adjoint theorem 2.1 as concerns the structure of projective objects.
Theorem 4.1. Under the conditions of Theorem 2.1 assume that a functor
L: (f-. is given such that
(I) LS is isomorphic with
(II) If y: A -. AI is a morphism in (f such that L (y) is an isomorphism and
18
SAMUEL EILENBERG AND J. C. MOORE
if A and AI are T-
1
fJ-pro jective then y is an isomorphism.
Then for any object A in (1' the following properties are equivalent:
(i) A is T-1fJ-projective
(ii) L (A) is fJ-projective and A ~ SL (A)
(Hi) A ~ S(B) for some fJ-projective object B in 93.
Proof. (ii) =9 (iii) is obvious and (iii) =9 (i) follows from 2.1. Thus only
(i) =9 (ii) needs to be established.
Let ell: LS-+ 193 be an isomorphism of functors.From the commutative diagram
LSeIl
LSLS LS
~ L S j
j~
LS
> 193
ell
we deduce that eIlLS = LScIl since ell is an isomorphism.
Let A be T-1fJ-projective.By 2.1, there exist morphisms
A !-.. 5(B) J:.... A
such that B is &projective and l/J8 = lA' Thus the composition
L A ~ LS(B) L(t/J) , L(A)
is IL(A)'Since ell (93) : LS(B) ~ B it follows that LS(B) is &projective and
thus L (A) is also fJ-projective as a retract of LS(B).
Let y denote the composition
SL(A) SL( 8). SLS(B) ~ SW) ~ A
Then
L (y) = L (l/J)LScIl (B)LSL (8) = L (l/J)eIl (LS(BLSL (8)
= L(l/J)L (o)eIl(L (A = eIl(L(A
and thus L (y) is an isomorphism. Since both A and SL (A) are T-l&projective
it follows from (11) that y: SL (A) -+ A is an isomorphism as required.
As a refinement of 3.1 we have
Proposition 4.2. Under the conditions of 3.1 let La: (1' -+ 93a' O' ~ be a
family of functors such that
(la) La preserves the coproducts E:B S/B
r
), T It
T
(Ib) L a S a ~ 1:B and LaST = 0 for T';'0'.
a
(11) If y: A -+ AI is a morphism in (1' such that L CT(y) is an isomorphism
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
19
for each and if A and A' are ti/"-projective then y is an isomorphism.
Then for each object A in (1 the following properties are equivalent:
(i) A is &,It.projective,
(ii) L(J(A) is &'(J-projective for all and
(J
(iii) A EB S(J(B(J) for some family IB(JI, where B(J is &'(J-projective for
all (j
Proof. Using the notation of the proof of 3.1 we consider the functor L: (1-.53
satisfying p L == L for all u
er (J
From (la) and (Ib) we deduce
PaLS == SrPr) == EBLaSrPr ==L(JSaPa P(J
r r
and therefore LS 153. Condition (11) obviously implies condition (11) of 4.1.
Since SL == EBS(JPaL =EBSaL
a
, the conclusions follow from 4.1.
a (J
5. Generators
Proposition 5.1. Let (1 be a category with coproducts and, P an object of
(1. Let (P) =9 &'p=?> Pp. Then &'p is a projective class in (1 and Pp is the
class of retracts of coproducts of copies of P.
Proof. Consider tbe functor T = (1(P, ): (1-. S. We construct a coadjoint
S--+T as follows. For S let = = EBA(J' u where ACT=O
CT
if u is the base point of and Aa = P if u is not the base point of Let
i(J: A(J-> be the natural injections. For every morphism cjJ: in S,de-
note by S(cjJ) = P4> the unique morphism satisfying P4>i(J= i4>J) if
cjJ(u) is not the base point of and P4>i
CT
=O if cjJ(u) is the base point of
This yields a functor S: S-. (1. Now define
a(A): ST(A) = P(1(P,A) -. A, a(A)i4> = cjJ: P -. A
a(A)i4> = 0:0-. A
-. (1(P, = P -4
0: P -.
Relations 1.1 are easily verified so that we have
if cjJ ,f 0,
if cjJ = O.
if u-,f 0,
if u-=O.
(a, 13):S--1 T: (1, S).
Let &'0 be the class of all (split) exact sequences in S and let &'0 ==9 PO.Then
by_I, 4.2 Po consists of all the objects of S. By the adjoint theorem 2.1, &,' =
T-l&,O is a projective class in (1, and if &" ==9 pi then P' is the class of re-
tracts of objects i.e., p' is the class of retracts of coproducts of P.
Since P Pp and Pp is closed under coproducts and retractions, it follows
20 SAMUEL EILENBERG AND J. C. MOORE
that PI C Pp.On the other hand (P) C pI implies Pp =(P) C P'. Thus Pp =pI
and therefore &p = &'. This completes the proof.
Let k be the kernel functor in the category S. By I, 6.2,k is a resolvent
for the class &0' Therefore 2.3 yields a resolvent e for the projective class &p.
Explicidy for any f: A-+ A', e (f} is the composition
SDkT(f} SkT(f). S T A ~ A.
A resolvent pair for &p is even easier to describe. It is simply (ST, a).
We shall say that the object P of (1 is a generator if the functor T = (1(P, )
is faithful.
Corollary 5.2. Let (1 be a category with kemels. If P is a generator, then
&p is an exact projective class. If further P is projective then &p =&1 is the
class of all exact sequences in the category (1, which is thus proje-ctively perfect.
This follows from 2.4 and the remark that P is projective if and only if
(1(P, f) is surjective for any epimorphism f.Thus P is projective if and only if
the functor T preserves epimorphisms.
OIAPTER III. EXAMPLES
1. Groups
Let be the category of groups and group morphisms with the usual compo-
sition.The groups consisting of the Unit element alone are ttivial. -The category
has products (namely, the usual cartesian products) and coproducts (namely,
the free products).
Every morphism f: G-+ GI has a kemel i: H-+ G, where H is the subgroup
of G composed of elements g G with f(g) = e', and i is the inclusion mor-
phism. Also f has a cokernel which is the natural morphism GI -+ G'/N where N
is the least invariant subgroup of G' containing the subgroup f(G). It should be
noted that an exact sequence in is also coexact, but a coexact sequence need
not be exact.
The group Z of additive integers plays a special role. The functor
T = (Z, ): -+ S is the functor which to each group assigns the underlying set.
From 11, .5 we know that Thas a coadjoint S -1 T, where the functor S: S-.
assigns to each pointed set I the free group generated by the elements of I
different from the base point.
The following properties of the functor T should be noted: T is faithful and
preserves and reflects monomorphisms, epimorphisms, kemels and exact sequences.
All these facts are obvious except the preservation of epimorphisms. This re-
quires proving that an epimorphism J: H-. G in is surjective, i.e., maps H
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 21
onto G. This fact has been established in [9]. We give another proof here which
has the advantage of being valid also for the full subcategory of g determined by
the finite groups.
Let then j: H-> G be an epimorphism in g.Replacing H by itsmage in G
we may assume that H is a subgroup of G and that j is the inclusion. 'We must
prove that H = G. First assume that H has index 2 in G; then H is an invariant
subgroup of G. Then 7Tj = OJ where 7T, 0: G-+ G/H are the factorization morphism
and the trivial morphism. Thus 7T =0, a contradiction.
Thus we may assume that the set G/H of right cosets of H in G has at
least 3 elements. Let ep be a permutation of the set G/H which leaves the coset
H and only the coset H fixed. Let 7T: G-+ G/ H be the natural map rrg = Hg and
let.,,: G/H -> G be such that 7T'" = lC/ H' ." (If) = e (where e is the identity ele-
ment of G). Every element g of G can then be written uniquely as a(g) .,,(7T(g
with a(g) H. Define ,\ (g) = a(g)." (ep (7T(g))).We verify that ,\ is a permutation
of the set G. Let P be the group of all the permutations of G and consider the
morphisms k, l: G-+ P defined by
k(g)x =gx for g, x G, l(g) =,\-lk(g)'\.
The condition k(g) = l(g) is equivalent with the condition ,\k(g) = k(g)'\, i.e.,
with the condition '\(gx) =g'\(x) for all x G.1f g H then 7T(gX) = rr(x) and
ga (x) = a (gx) so that ,\ (gx) = g,\ (x) and k (g) = l (g). Thus kj = lj and since j
is an epimorphism we have k = l. Thus ,\ (gx) = g'\ (x) holds for all g, x G.
Taking x =e we find '\(g) =g. Thus .,,(ep7T(g =.,,(7T(g and ep7T(g) =7T(g). Con-
sequendy g H and H = G.
Since the functor T = g (Z, ) is faithful and preserves epimorphisms, it fol-
lows that Z is a projective generator for the category . From 11, 5.2 and 11, 5.1
we deduce that the category g is projectively perfect and the projective objects
in g are the retracts of coproducts of copies of Z.Since the coproducts of copies
of Z are free groups and since a subgroup of a free group is free it follows that
the projective objects in are precisely the free groups. A resolvent pair (F, ()
for is obtained by defining F (G) to be the free group generated by the non-
trivial elements of G and taking d G) : F (G) -+ G to be the morphism which maps
each generator of F (G) into itself.
The full subcategory f determined by the finitely generated groups is.a
projectively normal subcategory of . Since the epimorphisms in f are also
epimorphisms in it follows that f is projectively perfect. However , it will
follow from a result in the next section that f does not have a resolvent pair.
Not only is the category not injectively perfect, but we shall show that
every injective object in is trivial. Indeed suppose G is injective and let
22 SAMUEL EILENBERG AND J. C. MOORE
g G, g t 1. Let H be a simple group and H' be an infinite cyclic subgroup of
H with generator h. Let f': H' -. G be the morphism defined by [' (h) =g. Since
G is injective, f' admits an extension [: H-. G. Since [(h) =g t 1, the kernel
of [ is a proper invariant subgroup of H and since H is simple it follows that [
is a monomorphism. Therefore card (H) ::; card (G). Thus to conclude the proof it
suffices to show that simple groups containing infinite cyclic subgroups can be
constructed with arbitrarily high cardinality. Let X be an infinite set and let P
be the group of permutations of X.For Tr P we denote the set of points of
Tr by FTr and its complement in X by M
Tr
By a theorem of R. Baer [I] the subset
N of P defined by the condition card (M
Tr
) < card (X) is a maximal invariant proper
subgroup of P, i.e., H= PIN is simple. It is clear that card(H) and
that H contains infinite cyclic subgroups.
2. Abelian groups
Let G' = ZM be the category of abelian groups with the usual morphisms and
composition. This is an abelian category. This category has arbitrary products
and coproducts. Indeed the products are the usual direct products (sometimes
called the "unrestricted" direct sums or products) and the coproducts are the
usual direct sum (sometimes called the "restricted" direct sums or products).
The group 2 is a projective generator for G' and therefore n, 5.2 shows that
G' is projectively perfect and the free abelian groups are the projective objects of
G'. Also from 11, 5 we obtain a resolving pair (F, f) for G'.The functor F is not
additive and indeed we shall show that G' does not have a resolvent (F, d with
F additive. Indeed suppose that F were additive. Let Q be the additive group
of rational numbers. Then the morphism n: Q-. Q is an isomorphism for any inte-
ger n t O.Thus F(n): F(Q)-.F(Q) is an isomorphism. Since F is additive we
have F(n) =n so that F(Q) is divisible. Since F(Q) is also free it follows that
F(Q) = 0, a contradiction.
It is easy to verify that the group Q
1
= QI2 is a cogenerator for G' i.e., a
generator for the dual category G'*.Since Q
1
itself is easily shown to be injec
tive it follows from 11, 5.2 that the category G' is injectively perfect. The injec-
tive objects of (f are the retracts of products of copies of QI.On the other hand
it is well known that the injectiveabelian groups are precisely the divisible abel-
ian groups. Again 11, 5 yields a coresolvent pair (f, G) in which the functor G
is however not additive. In fact a coresolvent (f, G) for (:j' with G additive does
not exist. Indeed, suppose that G is additive. Let 2 = 21n2 for some n> 1-
n
Then the morphism n: 2 -.2 is zero and thus G(n) : G(2 ) -. G(2 ) is zero.
n n n n
Since G is additive we have G(n) = n and since G(2 ) is divisible,
n
n:G(2
n
)-.G(2
n
) is an epimorphism. Thus G(2
n
) =0, a contradiction.
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 23
Let (11 be the full subcategory of er determined by the finitely generated
abelian groups. It is easy to see that err is a projectively normal subcategory of
er, and that err itself is projectively perfect. We shall show that err has no re-
solvent pair. Indeed suppose that (F, f) is a resolvent pair for (1r For any inte-
ger n ~ 1 let An denote the image of the morphism F (n) : F(Z) -. F(Z).Given
another integer m> 1 we consider the natural morphism TT: Z-. Zmn.Since
TTmn = 0 we have F (TT) F (mn) =0 and therefore F (TT) (A ) = O.On the other
mn
hand TTn;iO and since F is faithful (by 1,6.2) we have F(TT)F(n) = F(TTn);i 0
and thus F (TT) (An) ;io. Since A
mn
C An it follows that F(TT) defines a non-zero
morphism AniA
mn
-. F(Zmn)' Since F(Zmn) is a free abelian group it follows
that rank (A
t
/ A
mn
) > 0 and therefore
rank (An) > rank (A
mn
), if m> 1.
Thus Al = F(Z) has infinite rank, a contradiction.
Let j" denote the full subcategory of er determined by the torsion groups.
Given any abelian group A let r(A) denote the torsion subgroup :>f A. Then if A
is injective (i.e., divisible), it follows that r(A) also is injective. From this we
deduce easily that j" is an injectively normal subcategory of (1 and is injectively
perfect.
Not only does j" fail to be projectively perfect, but we shall show that the
only projective objects in j" are the trivial ones. Indeed suppose that A is pro-
jective in j" and A ;i o. Then there exists a monomorphism i: Zp -. A for some
integer p > 1. Consider the exact sequence
j p
0-+Zp-+Q1-+Q
1

Since Q
1
is injective there exists amorphism f: A-> Q
1
such that fi = j. Since
A is projective and p: Q
1
-+ Q
1
is an epimorphism there exists amorphism
g: A-+ Q
1
such that gp = f. Then gp = f and j = {i = gpi. However pi = 0 and
thus j = 0, a contradiction.
Given an abelian group A and an integer n > 1 we have the exact sequences
o-> nA -+ A -+ nA -+0,
o-+ nA -> A -> A -. 0
n
as well as the isomorphisms
Consider the groups
and let
~ = Z EBEBZ
n
, IT = Q
1
EB x Zn' n = 2, 3, .
n n
~ ) ~ ~ ~ P ~ , (IT) * ~ &fI * _ ~ j"II.
24
SAMUEL EILENBERG AND J. C. MOORE
Then by IT, 5.1 5, ~ is a projective class in a and 5,11 is an injective class in a.
A sequence E is in 5, ~ if and only if a ~ E) is exact. Since
6 ~ E) ~ (f(Z, E) EEl x a(Z , E) ~ EEEl x E, n", 2, 3, ...
n n n n
it follows that E is in 5 ~ if and only if E and each of the sequences nE (n> 1)
are exact. Similarly E is in 5,II if and only if (f(E, ll) is exact. Since
aCE, ll) ~ (1(E, Ql) EEl x (!(E, Zn) ~ (f(E, Ql) EEl x (f(E
n
, Ql)
n n
and since Q
1
is an injective cogenerator for (f, it follows that E is in 5,II if
and only if the sequences E and En (n> 1) are all exact.
Now suppose that the sequence E is a complex, i.e., is non-terminating in
both directions. From the exactness of the sequences
o-+ E -+ E -+ nE -+ 0 0 -+ nE -+ E -+ E -+ 0
n n
we deduce that E is in 5 ~ if and only if E is in 5,Il and if and only if nE is
exact for every integer n'" 1, 2, '. Thus the classes 5 ~ and 5,11 are comple-
mentary and the complexes in 5 ~ (or 5,11) are the pure exact sequences.
The objects of Po!. are retracts of coproducts of copies of ~ . Since ~ is a
coproduct of cyclic groups every coproduct of copies of !. is again a coproduct
of cyclic groups. Since a subgroup of such a group is again such a group [5, p. 46],
it follows that the objects of P ~ are precisely the coproducts of cyclic groups.
The objects of j'II are retracts of products of copies of n. A product of
copies of n has the form A EEl B where A is a product of copies of Q
1
while B
is a product of cyclic groups. Since every morphism A-> B is trivial, it follows
(by a general argument valid in any abelian category) that a retract of AEEl B
has the form A' EB B/ where A' is a retract of A and B' is a retract of B. Thus
any object in &11 has the form AEB B where A is an injective (i.e., a divisible)
group while B is a retract of a product of cyclic groups.
Another characterization of the objects in j'II as the retracrs of abelian
groups capable of carrying a compact topology was given by Los [10].
3. Banach spaces
Let K denote the field of real numbers or the field of complex numbers. We
shall consider the category j3 whose objects are Banach spaces over K and
whose morphisms are continuous linear transformations I: B-+ B/. Each such
transformation I has a norm If 1 =' SUPlxlq II(x) I and with this norm :B(B, B')
is again a Banach space. If g: BI -+ 8" then jgll S Igl Ill. This implies that
j3([, C):j3(B', C)-+j3(B, C) and j3(C, f):j3(C, B)-.j3(C, B') are again mor-
phisms and 1j3(f, C) I Sill, 1j3(C, f} I sI
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 25
For each Banach space B we denote by B' the unit ball of B, i.e., the set of
all points x B with Ixl 1. We denote by the subcategory of obtained by
replacing B
'
) by (B, B
'
) = B' )'. Thus in morphisms have norm
1. Isomorphisms in have norm 1 and are isometries.
Given a family IB (J"J, a of Banach spaces we consider the Banach space
x B(J" whose elements are families IX(J"J, x(J" B(J"' a with IX(J"I bounded and
with Ilx(J"JI defined as sUPlx(J"l. The natural projections Pr:x B(J"-4B
r
are then
morphisms in and are a product. Indeed, given any family [(J": C-4 B(J" of mor-
phisms in and given y C we have I[(J"(Y) I :s I[(J"I Iyl :s Iyl and thus fey) =
I{(J"(Y) J is an element of x B(J". This defines f: C-4x B(J" in and pof = [(J" for
a Clearly f is unique.
We also define the Banach space EB B(J" whose elements are families IX(J"J,
x(J" B(J"' a with IX(J"I < 00. We define Ilx(J"ll = IX(J"I. The natural injec-
tions i ; B -4 ffi Bare morphisms in and are a coproduct. Indeed let
r r Q7 (J"
[(J": B(J"-4 C, a be a family of morphisms in . Then for each x = Ix(J"l
Et1 B(J" we have
If(J"(x(J") I :s If(J"I IX(J"I :s IX(J"I <00.
Thus [(x) = f(J"(x (J") is a well defined element of C and I[(x) I :S Ixl. Thus
f: Et1 B(J" -4 C is a morphism in and fi (J" = [(J" for all a Clearly [ is unique.
Each element of Et1 B is also an element of x B and there is a canonical
(J" (J"
morphism
(): EBB(J"-4 x B(J"
in .. If the family is finite then () is bijective and Et1 B(J" and X B(J" are iden-
tical as topological vector spaces but have different norms. Thus Et1 B and x B
(J" (J"
are isomorphic in the category but not in the category
Let [:B
1
-+B
2
be any morphism in .. Let B,=/l(O) and B"=[(B
1
).
the inclusion BI -+ B1 is a kernel of [ and the natural projection B
2
-> B
2
/ B"
is a cokernel of [. Further [ admits a factorization in
B
1

where TT is the natural projection, a is an inclusion and l is a linear isomorphism,
but not necessarily an isometry. If 1 is an isometry, then the morphism [ is
called normal. It is easy to see that in the notions of exactness and coexact-
ness coincide.
Consider
(K) =9 &'K =9 PK' (K) * =9 &,K * =9
By 11, 5.1, &,K is a projective class in &,K is an injective class in PK
consists of retracts of coproducts of copies of K while consists of retracts
26 SAMUEL EILENBERG AND J. C. MOORE
of products of copies of K.
A sequence
is in [hK if and only if
B ~ B1... B"
B'-+ B -+ B"
(3.1)
is exact in S. It is easy to verify that this holds if and only if i (B') = r I (0)
and i is normal.
The sequence (3.1) is in [hK if and only if
~ ( B , K). -+ ~ ( B , K)" -+ ~ ( B , K).
is exact. It is an elementary exercise involving the Hahn-Banach theorem to see
that this holds if and oply if HB') = r 1 (0) and j is normal. Thus as far as un-
limited sequences are concerned both [hK and [hK consist of exact sequences in
which all morphisms are normal. Thus {f,K and [hK are complementary.
4. Rings and modules
Let A be a ring and let AM be the category of left A-modules. This is an
abelian category if AM(A, A') is regarded as an abelian group in the usual fashion.
Given a ring morphism <1>: I -+ A we consider the functor
T: AM -+ IM
which assigns to each A-module A the I-module obtained from A be setting
a a = <I> (a )a. Clearly T is faithful and exact. We also consider the functors
5:
I
M-+ AM, 5' :IM-+ AM
defined for B and fin IM by
5(B) = AIB, S(f) = AIf,
5' (B) =I(A, B), 5' (n = I(A, f)
where I( , B) stands for IM( , B). In defining 5 we regard A as a bimodule
AAI while in defining 5' we regard A as a bimodule IAk The usual isomor-
phisms
A(A I B, A) ~ I(B, A A A) = I(B, A),
A(A, I(A, B)) ~ I(A A A, B) = I(A, B)
yield adjointness relations 5 ---i T and T ---i 5'. Specifically we have
(a, m: 5 ---i T: (AM, IM), (4.1)
(fJ': a'): T ---i 5': (IM, AM) (4.2)
where
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 27
a: ST -+ 1, a(A): A0 I A -+ A, A 0 a -+ Aa,
{3: 1-+ TS, {3(B):B-+A0IB, b-+ 10b,
(3': TS' -+ 1, (3'(B) :I(A, B) -+ B, cp -+ 4>(1) for cp I(A, B),
a': 1 -+ S'T, a'(A): A -+ I(A, A), a -+ cp, cp(A) = Aa.
Applying il, 2.1 we find in AM a projective class &1}> = T-
1
&0 where &0 is
the class of split-exact sequences in IM. Similarly if &0 is the class of co-
splitting exact sequences in IM, then &11> = T-l&O is an injective class. For
unlimited sequences, splitting and co-splitting exact sequences coincide and
thus &1}> and &1}> are complementary classes.
The &I}>-projective objects are the retracts of A 0! B while the &I}>-injec-
tive objects are the retracts of I(A, B) for any B IM.
5. Coalgebras, comodules and contramodules
Let K be a commutative ring. A K-coalgebra is a K-module A together with
morphisms
(A : A -+ K, cpA : A -+ A 0 A
such that the diagrams
~
4>1 l
A0cp
A 0A----. A 0A 0 A
d>t<!)A
are commutative. Here and in the sequel all tensor products are taken over 1\.
A left comodule A over A is a K-module A together with a K-morphism
'P: A--. A 0 A such that the diagrams
A ,A0A
'Pj lA0'P
~
4>A
are commutative.
Morphisms of A-comodules are defined in the usual way. There results a
category AM of left comodules over A. This is an additive category with coker-
nels. If A is K-flat, then AM is an abelian category.
We have the obvious functor T: AM -+ KM which to each comodule A assigns
A regarded as a K-module. We also have the functor 5': KM -+ AM which to each
28
SAMUEL EILENBERG AND J. C. MDDRE
K-module B assigns the comodule A 0 B with ep 0 B: A 0 B -+ A 0 A 0 B.
There results an adjointness relation
T -i 5': (AM, KM).
Indeed, for every A AM, the morphism t/J: A-+ A 0 A defines f3 (A) : A-+ 5' T(A).
For every B KM the morphism f 0 B : A 0 B_ B defines a(B): T5' (B) -+ B.
If in KM we consider the class lP of all exact sequences (resp. the class
&0 of all split exact sequences) then in AM we obtain the injective class
&== T-I&I of all sequences which are exact as sequences of K-modules (resp.
the injective class &== T-I&O of all sequences which are split exact as sequences
of K-modules). The &-injective comodules are the retracts of comodules A 0 B
with B injective K-module (resp. an arbitrary K-module). It should be noted that
since T is faithful &== T- 1&1 contains only coexact sequences. However, since
T need not preserve monomorphisms &== T-I &1 need not contain all the coexact
sequences of AM.
A contramodule A over A is a K-module A together with a K-morphism
t/J: CA,A)-A (where (A, A) stands for the K-module of K-morphisms A_A),
such that the diagrams
A CA, CA, A == (A 0 A, A) ~ CA, A)
(" Alj ~ A (A.ofrl I I ofr
CA, A) - A CA, A) ) A
f ~
are commutative. There results an additive category AM# with kernels. The func-
tor T: AM#_KM which to each contramodule assigns the underlying K-module
has a coadjoint 5 ---i T with 5 == CA, ), where for every K-module B, CA, B) is
regarded as a A-contramodule using
~ , B
(A, CA, B == CA 0 A, B) ---+ (A, B).
Thus the adjoint theorem yields projective classes in AM#.
If A is K-projective then the category AM# is abelian. If A is K-projective
and fini;ely generated then the categories AM, AM# and AM are isomorphic
where A == CA, K) has the structure of K-algebra induced by the coalgebra struc-
rure of A. The details are left to the reader.
CHAPTER IV. COMPLEXES IN AN ABELIAN CATEGORY
1. The category ccf
Let cf be an abelian category. A complex A in cf is a family I An I, n Z,
of obj ects of cf and a family of morphisms d
n
: An - An -I such that dnd
n
+1 == 0
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 29
for all n Z. Amorphism f: A--> A' of complexes is a family of morphisms
f : A --> A' in (f' such that d' f = f 1d for all n Z. With composition de-
n n n n n n- n
fined in the usual way we obtain an abelian category cl] of complexes over Cf.
Given a complex A, the morphisms d
n
: An --> An -1 lead to exact sequences
W
n
Un
o~ Z (A) ~ A --> B 1 (A) ~ 0,
n n n-
Vn X
n
,
o~ B (A) ~ A ~ Z (A) ~ 0
n n n
such that d
n
= v
n
-1 un' The relation dd = 0 then yields an exact sequence
o~ B A ) ~ Z (A) ~ Z' A ) ~ B leA) ---+ 0
n n n n-
such that wni
n
= v
n
' inxn = un' There result exact sequences
in k
n
0---. B
n
(A)---. Zn(A) ---.H
n
(A) ---. 0
in Z' jn (
0---. H
n
(A) ~ n (A) ---. B
n
-
l
A) ---+ 0
(1.1)
(1.2)
'(1.3)
(1.4)
n Z,
Note that a complex A is an exact sequence if and only if the homology ob-
jects Hn (A) are zero for all n.
Proposition 1.1. If f: A-- A' is a morphism in c(l and if Bn(f) and H
n
(f)
are isomorphisms for all n Z then f is an isomorphism.
Proof. From the conmutative diagram
0 __ B (A) ~ Z (A) ---. H (A) ~ 0
n n n
1Bn (f) 1Zn (f) 1Hn (n
o __ B A ) ~ Z A ) ~ H (A')_ 0
n n n
with exact rows, we deduce that Zn(f) is an isomorphism for all n Z. From
the commutative diagram
0_ Zn (A) -- An ---. B
n
-
l
(A) ----+ 0
1Zn(f) 1f
n
IBn-l(f)
0----+ Zn(A')-- A'n----+ Bn_l(A') ----+ 0
with exact rows, it then follows that f
n
is an isomorphism for all n Z. Thus f
is an isomorphism.
In addition to the functors
B
n
, Zn' Z ~ , H
n
: ccf - cf,
we consider also the functors
30 SAMUEL EILENBERG AND J. C. MOORE
C
n
:CCf--+(1, nEZ,
given by C
n
(A) = An' C
n
(f) = f
n
.
Next we introduce the functors
Rn : (1 - cl1, n E Z,
by defining Rn (D) for D 11 to be the complex with D in position n and 0 in
all other positions. We note that Z R = 1'1' We assert the adjointness relation-
n n U
ship
(1.5)
where w
n
: RnZ
n
--+ 1
e
l1 is given by the morphism w
n
: Zn (A) --+ An and
in: 111--+ ZnRn = 111 is the identity morphism. Relations 11, (1.1) are trivially
satisfied. Similarly we have
(in' x
n
): Z'n ---i Rn: (Cf, cl1)
Finally we introduce the functors
(1.6)
Q
n
: 11--+ cl1
by defining Qn (D) for DE (1 to be the complex
ID
--+O----+D--+D--+O-
with D in positions n and n - 1 and 0 elsewhere. We note that CnQ
n
= 111' We
assert the adjoinrness relationship
(1. 7)
where in: 1(1- CnQ
n
= 1(1 is the identity morphism and an (A): QnC
n
(A) --+ A
is given by the diagram
The relations 11, (1.1) are easily verified. Similarly we have the adjoint relation-
ship
(1.8)
Since each of the functors C
n
have both a left and a right adjoint it follows
from 11, 2.2 that C
n
preserves kernel and cokernels, i.e., C
n
preserves exact se-
quences.
A family A(T' a of objects in cll is called locally finite if for each
nE Z, the objects C (A ) are trivial for all but a finite number of indices a
n (T
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
31
(1.9)
i
a
Aa;:::::::! A.
Pa
denotes any of the {unctors C , Z , Z' , B , H ,
/' n n n n n
Further, if T
n
Proposition 1.2. For any locally finite family Aa' a I of objects in C(1
there exists a biproduct
Tn<i
a
}
Tn (A a) , 'Tn (A)
Tn(p a)
is a finite biproduct (i.e., Tn(A
a
) t- 0 only for a finite number of indices a I).
Proof. For each n Z, the family A ,a I has only a finite number of
a,n
nontrivial objects and therefore we have a biproduct
i
a
n
A ~ An'
n ~
Pa,n
( 1.10)
We define d
n
: An -> A
n
-
1
by setting
d -li d p
n - a a,n-l a,n a,n'
aI
where the summation is only superficially infinite since only a finite number of
the summands is nontrivial. Then did = 0 so that (A , d ) defines a complex
n- n n n
A. The morphisms in (1.10) define the morphisms in (1.9) and the verification that
(1.9) is indeed a biproduct is straightforward.
To prove the second part of 1.2, replace A and A by the complexes A'
a a
and A' which coincide with Aa and A for the indices n + 1, n, n - 1 but are
trivial for other indices. Then (1.9) is still valid for A'a and A', and (1.9) is
then essentially finite. Since T (A' ) = T (A ) and T (A') = T (A) the con-
n a n ann
clusion follows from the fact that Tn is additive.
In the sections that follow we shall discuss various projectively perfect
classes of epimorphisms in C(1, and in various subcategories of c(1. Results
for injectively perfect classes of monomorphisms can then be obtained by duality.
In carrying out this dualization it is important to a priori distinguish between the
categories C(1*) and (Ca)*. An object in C(1*) is a complex in (1 but with
a differentiation of degree + 1. If we agree to write such a "co-complex" as
an n +1
__ An-I __ A n ~ An+l __
and adopt the sign change rule A_n = An, d_n = d
n
+1 then C(6 *) and (C(1) *
become identified and we may write C(1 * without fear of ambiguity.
2. A general theorem
Theorem 2.1. Let (f be an abelian category and ~ a full subcategory of C(1.
Let M and N be subsets of Z such that
32 SAMUEL EILENBERG AND J. C. MOORE
(I) If A et, m M and N N, then Qm(A) and Rn(A) are in :D.
(11) If IACTl is a locally finite family of objects in :D then their biproduct in
Cet is in :D.
Then for any projective class &, in et the class :D n &,M.N where
&,M,N = [ nC-
1
&,] n[ nZ- I&,]
m m n n
is a projective class in :D.
Further, for any object A in :D the following properties are equivalent:
(i) A is :D n &,M.N-projective.
(H) B
m
-
1
(A) and Hn(A) are &'.projective for all m M, n Nand
A [E!:) QmBm-l (A)] E!:)[E!:)RnHn(A]].
m n
(Hi) There exist &'.projective objects E ,F (j, m M, n N such that
m n
Proof. Since the functors Q ,m M and R , n N have values in :D and
m n
since :D is a full subcategory of C(f, it follows that the adjointness relations
(1.7) and (1.5) imply
Q
m
--j Cm :(:D, et),
Rn --j Zn :(:D, (1),
m M,
n N.
Conditions (I) and (II) show that the functors Q , R , m M, n N are coin-
m n
tegrable. Thus we may apply 11, 3.1. This proves that &11: = :D n &,M.N is a pro-
jective class in :D, aDd that the &,Itprojective objects of :D are the retracts of
obj ects satisfying (Hi).
Since for A satisfying (iii) we have
B
j
-
1
(A) = 0 if i e M, Hj(A) = 0 if j eN (2.1)
it follows that (2.1) holds for any &,Itprojective object A in :D.
To prove the equivalence of (i), (ii) and (Hi) we consider the functors
Bm_1::D--..d', H
n
:5)-et, mM, nN
and show that the conditions (la), (Ib) and (11) of 11, 4.2 are satisfied. Condition
(la) is satisfied because of 1.2. Condition (Ib) is satisfied because
B
m
-
1
Q
m=let, Bm-1Qj=O for i,f,m,Bm_1Rn=O,
H
n
R
n
=I(f, HnRj=O for j,f,n,HnQj=O.
To prove condition (11) let y: A-> AI be a morphism in et such that A and AI
are &,Itprojective and B
m
-
1
(y) and H
n
(y) are isomorphisms for m-I M and
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 33
The class
for n N. Since A and A' satisfy (2.1) it follows that B/y) and H/y) are iso-
morphisms for all i Z. Thus by 1.1, Y is an isomorphism. The equivalence of
(i), (ii) and (iii) now follows from 11, 4.2 and the proof is complete.
Proposition 2.2. Let :0, M, Nand :0', M', N' be two systems satisfying the
conditions of 2.1 and assume that :0 C :0', McM', Ne N'. If for some projective
class {i; in (1 we have
:0 () {i;M,N C tb
M
', N'
then :0 is a normal :0' n {i;M',N' -subcategory of :0'.
Proof. Since {i;M',N' C &M,N it follows that
From 2.1 (iii) we deduce that every :0' n {i;M',N' -projective object also is
:0 ('\ {i;M,N projective. This yields the desired conclusion.
3. Projective classes h;l c(1
Let {i; be a projective class in an abelian category a. Taking :0 = cll in
2.1, we find two natural choices for the sets M and N. The first one is M= Z
and N = O. This yields the projective class
c{i; = {i;2,0 = nC ~ I {i;, n Z,
n
in c(1. The second choice is M= N = Z. This yields the projective class
C& = &2,2 = n (C-
I
& n Z-I&), n Z
n n n
in c(1. Comparing condition (ii) of 2.1 for these two classes we find
Proposition 3.1. An object A of c(1 is C&-projective if and only if it is
C&-projective and Hn (A) = 0 for all n Z.
Now assume that the category ais projectively perfect and let &1 be the
class of all exact sequences in (1. Since the functors C are collectively faith-
n
ful, it follows from 11, 3.2 that C&1 is exact. Since further the functors C
n
are
exact, it follows that C&1 is the class of all the exact sequences in Ca. Thus
we find
Theorem 3.2. If the abelian category (1 is projectively perfect, then so is the
category C(:f. The projective objects of crr are biproducts Efl Qm(Em) for pro-
jective objects Em of (1.
&s = C&1 = n(C2&1 n Z ~ I I
. n
is of special importance. The sequences of &s will be called strongly exact.
34
SAMUEL EILENBERG AND J. C. MOORE
A sequence E is strongly exact if and only if the sequences C
n
(E) and Zn (E)
are exact for every n Z.
Corollary 3.3. If A in c(f is projective tlien H(A) = O.
Proposition 3.4. If E is a non-terminating sequence in Ca (i.e., if E CC(f)
then E is strongly exact if and only if
Cn (E), Zn (E), Z (E), Bn (E), Hn (E)
are exact for all n A.
Proof. The exact sequence
o-> Z (E) -> C (E) -> B (E) -> 0
n n n
implies that B
n
(E) is exact for every n Z. The exact sequences
o---. B (E) -> Z (E) -> H (E) -> 0,
n n n
0-+ Bn(E) -> Cn(E) -> -> 0
now yield the same conclusion for Hand Z' .
n 11
Theorem 3.5. If the abelian category ais projectively perfect then the strong
exact sequences form a projective class fi,s in Ca. Further, for any object A in
C(f the following properties are equivalent.
(i) A is fi, s-projective,
(ii) C (A), Z (A), Z' (A), B (A) and H (A) are proJ'ective for every n Z.
n n n n n
(iii) Bn (A) and Hn (A) are projective for every n Z.
Proof. Only the last part requires a proof.
(i) ==> (ii). If A is fi, s-projective, then by (iii) of 2.1 we have
A EB[EBRn(F
n
)]
with projective objects Em' F
n
(f. This implies (ii).
(ii) ==> (iii) is obvious.
(iii) ==> (i). Since B (A) and H (A) are pro)' ective, the exact sequences
11 11
(1.1) and (1.3) split, so that we have biproducts
There results a biproduct
W
n
Z (A)=
n w'
n
in
B (A) ==
n .,
111
k'
n
Z (A) -- H (A).
n k n
n
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA
35
P
n
Q B (A) ----+ A
n n-l ----+'
P'
n
qn
R H (A) ==: A
n n I
qn
where the morphisms Pn' ~ qn' q ~ are defined using the diagrams
B
n
-
1
(A)
1
B
n
-
1
(A)
H (A)
n
1 ~
1Wn -l
i
n-l 1 k'
W
n
n
d
n
-----to
A .A 1 ---+
---+ A ~ ...
n n- n
1Un 1' ,
1k w' n-1
W
n-l
n n
B
n -1 (A) B n-l (A)
H
n
(A)
Note that the first of these diagrams is commutative because d = W l i lu.
n n- n- n
Condition (ii) of 2.1 is thus fulfilled and A is (;;s-projective.
Proposition 3.6. Let Cf be a projectively perfect abelian category and let
E: X - A be a left complex over the object A of cCf. Then E: X - A is a strong
projective resolution of A in cCf if and only if for everyone of the functors
" "
Proof. Consider the complex X. By 3.4 X is strongly exact if and only if
Tn(X) is exact for all the functors T
n
By 3.5. Xi is (;;sprojective if and only
if Tn (Xi) is projective for every n Z. This yields the conclusion.
We thus find that the socalled "double resolutions" of a complex considered
in [4. Chapter XVII] are precisely the strongly projective resolutions.
4. Subcategories of cCf
Let Cf be an abelian category. Given - 00 $ p and q $ 00 we consider the
full subcategory Cq(j of C(f determined by the complexes A with A = 0 for
p n
n < p and for n > q.
We usually omit the symbol p if P = - 00 and the symbol q if q = 00
Complexes A in which the differentiation d
n
: An - A
n
-
1
is zero for all
n Z, determine a full subcategory cCf of cCf. An object of cCf is thus a family
{AnI, n Z, of nbjects in Cf. We also define
CZCf = cCf n cZa.
Starting with a projective class (;; in (f and applying 2.1 to the various sub-
categories listed above, with suitable choices of the sets M and N, we shall
36
SAMUEL EILENBERG AND J. C. MOORE
obtain projective classes in these subcategories. We shall then be able to apply
2.3 to recognize which of these are normal subcategories of Ca.
cqa for p = - 00. If we set
p
M= {n In Z, n:S gl, N = 0,
then conditions (I) and (11) of 2.1 are fulfilled and we obtain the projective class
cq&; in cq(j. Since C
q
&; = cq(f nC&;, it follows from 2.3 that Cq(j is a normal
e.- p p p p
C(9-subcategory of Ca. If we take
M= N = {nln Z, n:S gI
then again 2.1 applies and we obtain a projective class cq&; in cqa. Since
A A PAP
C
q
&; = c
q
(1 n C&; it follows from 2.3 that cq(f is a normal C&subcategory of Ca.
P P P
cZa for p finite. If we set
M= {nln Z, p < n:S g\, N = {pI,
then 2.1 yields a projective class CZ&;. We still have CZ&; = cZa n C&; (because
C
p
= Zp on CzcO so that Cq(j is a C&;-subcategoryof C(j. However 2.3 no
longer applies since for Cq(f we have N = {pI while for Ca (in defining C&;) we
p .
took N = O. In fact it follows from 2.1 (iii) that the cq&proj ective objects of
p
cza are the objects
A = Rp (E
p
) EB Qp+l CE
p
+
1
) EB QP+2 CE
p
+
2
) EB
where En' p n g, are &projective objects of (1'. Since Hp{A) = E
p
it follows
that A is not C&projective unless E
p
= O. Thus unless the category (j is trivial,
CZ&; is not a normal C&;-subcategoryof C&;. However, CZ&; is a normal Cp&sub-
category of cpa.
If we set
M= {nln Z, p < n g\, N = {nln Z, p n gl
then we find again that the conditions of 2.1 are again satisfied and 2.1 yields
the projective class CZ&;. Again we verify that C
q
&; = cq(j n C&; and 2.2 implies
A A P P
that CZ&; is a normal C&;-subcategory of cCf.
cZ(j for all values of p and g. We set
M= 0, N = {nln Z, p n gl.
Applying 2.1, we obtain a projective class CZ&;. A sequence E in CZC1 is in
CZ&; if and only if CnCE) = ZnCE) is in &; for all n Z. Thus CZ&; = n C&;
and cza is a C&subcategory of Ca, however (unless a is it is not a
normal C&subcategoryof Ca. We also find that cq&; = Cqa n C&; and 2.2 im-
A p P
plies that cqa is a normal C&;-subcategory of Ca.
p
We summarize the above results in the following
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 37
Proposition 4.1. Let &, be a projective class in an abelian category a. The
following are normal e&'-subcategories of ca;
ez for p = - 00
The following are normal C&'-subcategories of ca
eqa cq(j
p' p
for all values of p and q.
5. An application
Let abe an abelian category, &, a projective class in aand let be the
full subcategory of e_1(f determined by the complexes in &'. We set
M = Inln Z, n;:: 01, N = O.
The conditions of 2.1 are then fulfilled and we obtain a proj ective class &,1t in
A sequence E is in &,1t if and only if en (E) is in &, for every n;:: O. The fj,tt.
projective objects in are of the form EBQ (E ), n;:: 0 where En are fj,projec-
n n n
tive objects of a.
Let
f
X: .. --.X.--.X. l--.--.Xo--.A--.O--. ...
I r
be an object in and let
1
.. -Y..
L,I
J
... -Y
1

t- ,I
1
be a
.......... ...... 7] A
W: ",--,Yi--,Yi-1--",,--,Y
o
--' X--.O--., ..
&,1tprojective resolution of X. There results a double complex
1 1 f. 1
- Yi,j-l -'" --. Yi,o --!... B
i
--- 0
y J 1 f i - 1 J
i-l,l-l --. - Yi-1,0- B
i
-
1
- 0
1 1 1
1 1
f
O
1
.. -Y.
- YO,j-l
-----+
Yo,o
B
o
-0
0,1
---
7) j 1 7) j-1l
7) 0 1
1
1
"'-X' -X
j
-
1
--+ ---.
X
o
-
A -0
I
1
1
I
1 1
0 0
0 0
We denote the entire double complex by W, the double complex with the bottom
row replaced by zeros by V, the double complex with the extreme right column
replaced by zeros by U, and the double complex with the bottom row and the
right column replaced by zeros by Y. We use the same letters for the single
38
SAMUEL EILENBERG AND J. C. MDDRE
complexes associated in the usual fashion with these double complexes.
We note the following facts:
Each column of U is in &.
Each row of V is {j;i'tprojective and therefore is split exact
and Y . . and B; are &projective.

We shall show that
Bis in &
(5.1)
(5.2)
A
where B is the extreme right column of W.
Indeed, let P be any &projective object in (f. Consider the functor
T=:DCP,). We must show that HT(B) =0. Since it suffices to show
that HT(W) = 0 and HTCU) = O. Since each column of U is in & we have HT = 0
on each column of U. Thus by a standard filtration argument HTCU) = O. Since X
is in & and each row of V is split exact, it follows that HT is Zero on each row
of W. Thus again HTCW} = O. This proves (5.3).
Now let T: et ----+:B be an additive functor with values in any abelian category
:B. We consider the double complex TCY), the usual two filtrations of this double
complex and the associated spectral sequences.
First consider the horizontal filtration. We have EO = T(Y } and El =
A p,q p,q p,q
HqTCY
p
}' Since the row Y
p
is split exact, we have HqTCY
p
} = 0 for q> 0 and
HoTCY
p
} TCB
p
}. Therefore the spectral sequence collapses and
Since B is in & and Bi is &projective, r: B-+ A is an &projective resolution
of A. Thus HT(B} =L{i;Y(A}. Consequently
HT(V) L& T(A}. (5.4)
Now consider the vertical filtration. The terms EO are given by the columns
of TCV). Since the pth column of V is an &projective resolution of X
p
we have
= TCX
p
}' Thus
(5.5)
We thus obtain a spectral sequence
H CL&n CX) L&TCA}. (5.6)
p q
We consider two special cases. If L& TCX.} = 0 for q> 0 and L &0 TCX.} =
q {i;
TCX), then the spectral sequence collapses and yields HTCX) L T(A} just as
if X were an &projective resolution of A.
For the second special case consider a projective class in et such that
C & and let X be an resolution of A. Since Xis in it also is
in & so that (5.6) applies and yields
FOUNDATIONS OF RELATIVE HOMOLOGICAL ALGEBRA 39
(L n (A) L&T{A). (5.7)
This spectral sequence is the one given by Burler and Horrocks [3, p. 171].
BIBLIOGRAPHY
[1] R. Baer, Die Kompositionsreihe der Gruppe aller eineindeutigen Abbildungen
einer unendlichen Menge auf sich, Studia Math. 5(1934), 15-17.
[2] D. A. Buchsbaum, A note on homology in categories, Ann. of Math. 69(1959),
66-74.
[3] M. R. C. Burler and G. Horrocks, Classes of extensions and resolutions,
Phi!. Trans. Royal Soc. London, Ser. A, 254(1961), 155-222.
[4] H. Cartan and S. Eilenberg, Homological Algebra, Princeton 1956.
[ 5] L. Fuchs, Abelian Groups, Budapest 1958.
[6] D. K. Harrison, Infinite abelian groups and homological methods, Ann. of
Math. 69 (1959), 366-391-
[7] A. HelIer, Homological algebra in abelian categories, Ann. of Math. 68 (1958),
484-525.
[8] G. Hochschild, Relative homological algebra, Trans. A.M.S. 82 (1956), 246-
269
[9] A. G. Kurosh, A. Kh. Livsliz and E. G. Schulheifer, The foundation of the
theory of categories, Uspehi Mat. Naut. XV (6), 3-58.
[10] J. Los, Abelian groups that are direct summands of every abelian group
which contains them as pure subgroups, Fund. Math. 44 (1957), 84-90.
[ll] S. MacLane, Duality for groups, Bull. A.M.S. 56(1950), 485-516.
[12] N. Yoneda, On the homology theory of modules, Jour. Fac. Sci. Univ. Tokyo,
sec. 1(1954), 193-227.