Derived Cats
Derived Cats
Derived Cats
VIEW
J.P. MAY
Contents
1. Cell A-modules 2
2. Whiteheads theorem and the derived category 5
3. Derived tensor product and Hom functors: Tor and Ext 8
4. Some spectral sequences 9
5. Commutative DGAs and duality 11
6. Browns representability theorem 13
References 14
This is an edited version of Part III of [8]. It is an elementary introduction to
the theory of derived categories that is based on slavish immitation of the theory
of CW complexes, or rather cell complexes, in algebraic topology. The basic theory
is in Sections 1, 2, 3, and 5. The other sections, although important, are a little
more advanced.
Let k be a commutative ring. We consider Z-graded chain complexes of k-
modules, which we abbreviate to k-complexes. Such an X = {X
q
} has a dierential
d : X
q
X
q1
. The cohomologically minded reader can reindex by setting X
q
=
X
q
and d
q
= d
q
, so that the dierential raises degree. A k-chain map f : X Y
is a sequence of maps f : X
q
Y
q
that commute with the dierentials, df = f d;
f is a quasi-isomorphism if it induces an isomorphism on homology. The tensor
product (over k) X Y of k-complexes X and Y is given by
(X Y )
n
=
p+q=n
X
p
Y
q
,
with dierential
d(x y) = d(x) y + (1)
p
x d(y)
if deg(x) = p.
Let A be a dierential graded associative and unital k-algebra (= DGA). This
means that A is a k-complex with a unit element 1 A
0
and an associative and
unital product map AA A of chain complexes. Writing products by juxtapo-
sition, d(xy) = d(x)y+(1)
p
xd(y). The novice in homological algebra may think of
the simple case A = k, concentrated in degree zero and given zero dierential, but
the general case is no more dicult and is very important in modern mathematics.
By an A-module M, we mean a left A-module such that the action A M M
is given by a k-chain map.
As many topologists recognize, there is a close analogy between the derived cat-
egory D
A
of dierential graded A-modules and the homotopy category of spaces
1
2 J.P. MAY
in algebraic topology. The analogy becomes much closer if one considers the sta-
ble homotopy category of spectra, but the novice is not expected to know about
that. We here give a topologically motivated, but purely algebraic, exposition of
the classical derived categories associated to DGAs. These categories admit re-
markably simple and explicit descriptions in terms of cell A-complexes. These
are the precise algebraic analogs of cell complexes in topology, which are dened
in the same way as CW complexes except that cells need not be attached only to
cells of lower dimension.
Such familiar topological results as Whiteheads theorem and Browns repre-
sentability theorem transcribe directly into algebra. There is also a theory of CW
modules, but these are less useful, due to the limitations of the algebraic cellular
approximation theorem. Derived tensor products and Hom functors, together with
dierential Tor and Ext functors and Eilenberg-Moore (or hyperhomology) spectral
sequences for their computation, drop out quite easily. Since [8] assumed familiarity
with spectral sequences, which the novice will not have seen, we put the relevant
material into a small section of its own. In the lectures, I intend to interpolate a
brief discussion of the classical Tor and Ext functors from the present point of view
by specialization of Sections 13.
Our methods can be abstracted and applied more generally, and some of what
we do can be formalized in Quillens context of closed model categories. We prefer
to be more concrete and less formal. We repeat that many topologists have long
known some of this material. For the expert, we emphasize that k is an arbitrary
commutative ring and we nowhere impose boundedness or atness hypotheses. The
novice will wonder why such hypotheses were ever thought to be needed.
1. Cell A-modules
We begin with some trivial notions, expressed so as to show the analogy with
topology. Let I denote the unit interval k-complex. It is free on generators [0]
and [1] of degree 0 and [I] of degree 1, with d[I] = [0] [1]. A homotopy is a map
XI Y . The cone CX is the quotient module X(I/k[1]) and the suspension
X is X (I/I), where I has basis [0] and [1]. Additively, CX is the sum of
copies of X and X, but with dierential arranged so that H
. Explicitly, (Cf)
q
= Y
q
X
q1
, with dierential
d(y, x) = (dy + (1)
q
fx, dx).
The sequence
X Y Cf X
is called a cober sequence, or an exact triangle.
Now assume given a DGA A over k. If X is a k-complex and M is an A-module,
then M X is an A-module, and the notion of a homotopy between maps of A-
modules is dened by taking X = I. Since we dened cober sequences in terms of
DERIVED CATEGORIES FROM A TOPOLOGICAL POINT OF VIEW 3
tensoring with k-complexes, the cober sequence generated by a map of A-modules
is clearly a sequence of A-modules. Let M
A
denote the category of A-modules
and hM
A
be its homotopy category. Its objects are the A-modules, and its maps
are the homotopy classes of maps of A-modules. Then the derived category D
A
is obtained from hM
A
by adjoining formal inverses to the quasi-isomorphisms of
A-modules. In Construction 2.7 below, we shall give an explicit description that
makes it clear that there are no set theoretic diculties. (This point is typically
ignored in algebraic geometry and obviated by concrete construction in algebraic
topology.)
The sequences isomorphic to cober sequences in the respective categories give
hM
A
and D
A
classes of exact triangles with respect to which they become tri-
angulated categories in the sense of Verdier [14] (see also [13]), but we do not go
seriously into that here.
It is also convenient to think of the suspension functors in a slightly dierent
way. Let S
q
be the free k-complex generated by a cycle i
q
of degree q, where q Z.
Then our suspension functors are just
q
M = M S
q
.
We think of the S
q
as sphere k-complexes. We let S
q
A
= AS
q
and think of the S
q
A
as sphere A-modules; they are free on the generating cycles i
q
. The reader puzzled
by the analogy with spaces might prefer to take all k-complexes X to satisfy X
q
= 0
for q < 0, as holds for the chains of a space.
Digressively, as noted before, the closer analogy is with stable homotopy theory,
since that is a place in which negative dimensional spheres live topologically. In
fact, one description of the stable homotopy category ([9]) translates directly into
our new description of the derived category. (The preamble of [9] explains the re-
lationship with earlier treatments of the stable homotopy category, which did not
have the same avor.) In brief, one sets up a category of spectra. In that cate-
gory, one denes a theory of cell and CW spectra that allows negative dimensional
spheres. One shows that a weak homotopy equivalence between cell spectra is a
homotopy equivalence and that every spectrum is weakly homotopy equivalent to
a cell spectrum. The stable homotopy category is obtained from the homotopy
category of spectra by formally inverting the weak homotopy equivalences, and it is
described more concretely as the homotopy category of cell spectra. With spectra
and weak homotopy equivalences replaced by A-modules and quasi-isomorphisms,
precisely the same pattern works algebraically, but far more simply.
Denitions 1.1. (i) A cell A-module M is the union of an expanding sequence of
sub A-modules M
n
such that M
0
= 0 and M
n+1
is the cober of a map
n
: F
n
M
n
, where F
n
is a direct sum of sphere modules S
q
A
(of varying degrees q). The
restriction of
n
to a summand S
q
A
is called an attaching map and is determined
by the attaching cycle
n
(i
q
). An attaching map S
q
A
M
n
induces a map
CS
q
A
= A CS
q
M
n+1
M,
and such a map is called a (q+1)-cell. Thus M
n+1
is obtained from M
n
by adding a
copy of S
q+1
A
for each attaching map with domain S
q
A
, but giving the new generators
j
q+1
= i
q
[I] the dierentials
d(j
q+1
) = (1)
q
n
(i
q
).
4 J.P. MAY
We call such a copy of S
q+1
A
in M an open cell; if we ignore the dierential, then
M is the direct sum of its open cells.
(ii) A map f : M N between cell A-modules is cellular if f(M
n
) N
n
for all n.
(iii) A submodule L of a cell A-module M is a cell submodule if L is a cell A-module
such that L
n
M
n
and the composite of each attaching map S
q
A
L
n
of L with
the inclusion L
n
M
n
is an attaching map of M. Every cell of L is a cell of M.
We call {M
n
} the sequential ltration of M. It is essential for inductive ar-
guments, but it should be regarded as exible and subject to change whenever
convenient. It merely records the order in which cells are attached and, as long
as the cycles to which attachment are made are already present, it doesnt matter
when we attach cells.
Lemma 1.2. Let f : M N be an A-map between cell A-modules. Then M
admits a new sequential ltration with respect to which f is cellular.
Proof. Assume inductively that M
n
has been ltered as a cell A-module M
n
=
M
r
such that f(M
r
) N
r
for all r. Let x M
n
be an attaching cycle for the
construction of M
n+1
from M
n
and let : CS
q
A
M
n+1
be the corresponding cell.
Let r be minimal such that both x M
r
and f has image in N
r+1
. Extend
the ltration of M
n
to M
n+1
by taking x to be a typical attaching cycle of a cell
CS
q
A
M
r+1
.
Denition 1.3. The dimension of a cell CS
q
A
M
n+1
is q+1. A cell A-module M
is said to be a CW A-module if each cell is attached only to cells of lower dimension,
in the sense that the dening cycles
n
(i
q
) are elements in the sum of the images
of cells of dimension at most q. The n-skeleton M
n
of a CW A-module is the sum
of the images of its cells of dimension at most n, so that M
n
M
n+1
. We require
of cellular maps f : M N between CW A-modules that they be bicellular, in
the sense that both f(M
n
) N
n
and f(M
n
) N
n
for all n. By Lemma 1.2, the
latter condition can be arranged by changing the order in which the cells of M are
attached.
Denition 1.4. A cell A-module is nite dimensional if it has cells in nitely many
dimensions. It is nite if it has nitely many cells.
Just as nite cell spectra are central to the topological theory, so nite cell
A-modules are central here, especially when we restrict to commutative DGAs
and discuss duality. The collection of cell A-modules enjoys the following closure
properties, which imply many others.
Proposition 1.5. (i) A direct sum of cell A-modules is a cell A-module.
(ii) If L is a cell submodule of a cell A-module M, N is a cell A-module, and
f : L N is a cellular map, then the pushout N
f
M is a cell A-module with
sequential ltration {N
n
f
M
n
}. It contains N as a cell submodule and has one
cell for each cell of M not in L.
(iii) If L is a cell submodule of a cell A-module M and X is a cell submodule
of a cell k-complex Y , then M Y is a cell A-module with sequential ltration
{
p
(M
p
Y
np
)}. It contains LY +MX as a cell submodule and has a (q+r)-
cell for each pair consisting of a q-cell of M
p
and an r-cell of Y
np
, 0 p n.
(iv) The mapping cylinder Mf = N
f
(LI) of f : L N is the pushout dened
by taking L = Lk[0] LI. If f is a cellular map between cell A-modules, then
DERIVED CATEGORIES FROM A TOPOLOGICAL POINT OF VIEW 5
Mf is a cell A-module, L = L k[1] is a cell submodule, the inclusion N Mf
is a homotopy equivalence, and Cf = Mf/L.
Proof. Parts (i) and (ii) are easy and (iv) follows from (ii) and (iii). For (iii),
observe that there are evident canonical isomorphisms
S
q
S
r
= S
q+r
and S
q
A
S
r
= S
q+r
A
.
MY has an open cell S
q+r
A
for each open cell S
q
A
of M and S
r
of Y ; the dierential
on its canonical basis element is the cycle
d(j
q
) j
r
+ (1)
q
j
q
d(j
r
).
2. Whiteheads theorem and the derived category
A quick space level version of some of the results of this section may be found
in [11], and the spectrum level model is given in [9, I 5]. We construct the derived
category explicitly in terms of cell modules. As in topology, the homotopy ex-
tension and lifting property is pivotal. It is a direct consequence of the following
trivial observation. Let i
0
and i
1
be the evident inclusions of M in M I.
Lemma 2.1. Let e : N P be a map such that e
: H
(N) H
(P) is a
monomorphism in degree q and an epimorphism in degree q +1. Then, given maps
f, g, and h such that f|S
q
A
= hi
0
and eg = hi
1
in the following diagram, there are
maps g and
h that make the entire diagram commute.
S
q
A
i0
/
S
q
A
I
h
.{w
w
w
w
w
w
w
w
w
S
q
A
i1
o
g
}{
{
{
{
{
{
{
{
P N
e
o
CS
q
A
i0
/
f
=
{
{
{
{
{
{
{
{
CS
q
A
I
h
cH
H
H
H
H
H
H
H
H
CS
q
A
i0
o
g
aC
C
C
C
C
C
C
C
Proof. Let i = i
q
[0] and j = i
q
[I] be the basis elements of CS
q
A
, so that
d(j) = (1)
q
i. Then eg(i) = h(i [1]) and f(i) = h(i [0]), hence
d(h(i [I]) f(j)) = (1)
q+1
eg(i).
Since eg(i) bounds in P, g(i) must bound in N, say d(n
) = g(i). Then
p e(n
) + (1)
q
(h(i [I]) f(j))
is a cycle. There must be a cycle n N and a chain c P such that
d(c) = p e(n).
Dene g(j) = (1)
q
(n
n) and
h(j [I]) = c.
Theorem 2.2 (HELP). Let L be a cell submodule of a cell A-module M and let
e : N P be a quasi-isomorphism of A-modules. Then, given maps f : M P,
6 J.P. MAY
g : L N, and h : L I P such that f|L = hi
0
and eg = hi
1
in the following
diagram, there are maps g and
h that make the entire diagram commute.
L
i0
/
L I
h
.{x
x
x
x
x
x
x
x
x
L
i1
o
g
.~}
}
}
}
}
}
}
}
P N
e
o
M
i0
/
f
>
}
}
}
}
}
}
}
}
M I
h
cF
F
F
F
F
F
F
F
F
M
g
`A
A
A
A
A
A
A
A
i1
o
Proof. By induction up the ltration {M
n
} and pullback along cells not in L, this
quickly reduces to the case (M, L) = (CS
q
A
, S
q
A
) of the lemma.
For objects M and N of any category Cat, let Cat(M, N) denote the set of
morphisms in Cat from M to N.
Theorem 2.3 (Whitehead). If M is a cell A-module and e : N P is a quasi-
isomorphism of A-modules, then e
: hM
A
(M, N) hM
A
(M, P) is an isomor-
phism. Therefore a quasi-isomorphism between cell A-modules is a homotopy equiv-
alence.
Proof. Take L = 0 in HELP to see the surjectivity. Replace (M, L) by the pair
(M I, M (I)) to see the injectivity. When N and P are cell A-modules, we
may take M = P and obtain a homotopy inverse f : P N.
Theorem 2.4 (Cellular approximation). Assume that H
q
(N/N
q
) = 0 for all q and
all CW A-modules N. Let L be a cell submodule of a CW A-module M, let N be
a CW A-module , and let f : M N be a map whose restriction to L is cellular.
Then f is homotopic relative to L to a cellular map. Therefore any map M N
is homotopic to a cellular map, and any two homotopic cellular maps are cellularly
homotopic.
Proof. By Lemma 1.2, we may change the sequential ltration of M to one for
which f is sequentially cellular. Proceeding by induction up the ltration {M
n
}, we
construct compatible cellular maps g
n
: M
n
N
n
and a homotopy h
n
: M
n
I
N
n
from f|M
n
to g
n
. The result quickly reduces to the case of a single cell of M
that is not in L and thus to the case when (M, L) = (CS
q
A
, S
q
A
). The conclusion
follows by application of Lemma 2.1 to the inclusions e : (N
n
)
q+1
N
n
.
Remark 2.5. If H
q
(A) = 0 for all q < 0, then the homological hypothesis holds
and we can work throughout with CW A-modules and cellular maps rather than
with cell A-modules. This matches the intuition: CW theory works topologically
because the homotopy groups of a sphere S
q
are zero in degrees less than q. In
many of the motivating algebraic examples, the natural grading is cohomological,
with the cohomology groups of the spheres S
q
A
zero in degrees less than q. In
homological grading, this reverses the inequality, and the homological hypothesis
of the cellular approximation theorem then fails in general. That is why we focus
on cellular rather than CW theory here.
Theorem 2.6 (Approximation by cell modules). For any A-module M, there is a
cell A-module N and a quasi-isomorphism e : N M.
DERIVED CATEGORIES FROM A TOPOLOGICAL POINT OF VIEW 7
Proof. We construct an expanding sequence N
n
and compatible maps e
n
: N
n
M
inductively. Choose a cycle in each homology class of M, let N
1
be the direct
sum of A-modules S
q
A
, one for each of degree q, and let e
1
: N
1
M send the th
canonical basis element to the cycle . Inductively, suppose that e
n
: N
n
M has
been constructed. Choose a pair of cycles (,
(M). Let N
n+1
be
the homotopy coequalizer obtained by adjoining a copy of S
q
A
I to N
n
along
the evident map S
q
A
I N
n
determined by each such pair (,
) of degree
q. Proposition 1.5 implies that N
n+1
is a cell A-module such that N
n
is a cell
submodule. Any choice of chains M such that d() =
determines an
extension of e
n
: N
n
M to e
n+1
: N
n+1
M. Let N be the colimit of the
N
n
and e : N M be the resulting map. Clearly, N is a cell module, e induces
an epimorphism on homology since e
1
does, and e induces a monomorphism on
homology by construction.
Construction 2.7. For each A-module M, choose a cell A-module M and a quasi-
isomorphism : M M. By the Whitehead theorem, for a map f : M N,
there is a map f : M N, unique up to homotopy, such that the following
diagram is homotopy commutative.
M
f
/
M
f
/
N
Thus is a functor hM
A
hM
A
, and is natural. The derived category D
A
can
be described as the category whose objects are the A-modules and whose morphisms
are specied by
D
A
(M, N) = hM
A
(M, N),
with the evident composition. When M is a cell A-module,
D
A
(M, N)
= hM
A
(M, N).
Using the identity function on objects and on morphisms, we obtain a functor
i : hM
A
D
A
that sends quasi-isomorphisms to isomorphisms and is universal
with this property. Let C
A
be the full subcategory of M
A
whose objects are the cell
A-modules. Then the functor induces an equivalence of categories D
A
hC
A
with inverse the composite of i and the inclusion of hC
A
in hM
A
.
Therefore the derived category and the homotopy category of cell modules can
be used interchangeably. Homotopy-preserving functors on A-modules that do not
preserve quasi-isomorphisms are transported to the derived category by rst ap-
plying , then the given functor. Much more emphasis is placed on this simple
procedure in the algebraic than in the topological literature, as is reected in the
respective notational conventions.
Digressively, we note that topologists routinely transport constructions to the
stable homotopy category by passing to CW spectra, without change of notation.
In fact, while a great deal of modern work in stable homotopy theory depends
heavily on having a good underlying category of spectra, earlier constructions of
the stable homotopy category did not even allow spectra that were more general
than CW spectra. For this and other reasons, topologists are accustomed to work
8 J.P. MAY
with CW spectra and their cells in a concrete calculational way, not as something
esoteric but rather as something much more basic and down to earth than general
spectra. An analogous view of dierential graded A-modules is rather intriguing.
3. Derived tensor product and Hom functors: Tor and Ext
We rst record some elementary facts about tensor products with cell A-modules.
Lemma 3.1. Let N be a cell A-module. Then the functor M
A
N preserves exact
sequences and quasi-isomorphisms in the variable M.
Proof. With dierential ignored, N is a free A-module, and preservation of exact
sequences follows. The sequential ltration of N gives short exact sequences of free
A-modules
0 N
n
N
n+1
N
n+1
/N
n
0,
where the subquotients N
n+1
/N
n
are direct sums of sphere A-modules. The preser-
vation of quasi-isomorphisms holds trivially if N is a sphere A-module, and the gen-
eral case follows by passage to direct sums, induction up the ltration, and passage
to colimits.
It is usual to dene the derived tensor product, denoted M
L
A
N, by replacing the
left A-module N (or the right A-module M) by a suitable resolution P and taking
the ordinary tensor product M
A
P, in line with the standard rubric of derived
functors (see e.g. Verdier [14], who restricts to bounded below modules). Our
procedure is the same, except that we take approximation by quasi-isomorphic cell
A-modules as our version of a resolution and, following the pedantically imprecise
tradition in algebraic topology, we prefer not to change notation. That is, in D
k
,
M
A
N means M
A
N (or M
A
N or M
A
N: the three are canonically
isomorphic in D
k
). The lemma shows that the denition makes sense. We can
also use the lemma to show that the derived category D
A
depends only on the
quasi-isomorphism type of A.
Proposition 3.2. Let : A A
: D
A
D
A
is an equivalence of categories with inverse given
by the extension of scalars functor A
A
(?).
Proof. For M M
A
and M
M
A
, we have
M
A
(A
A
M, M
= M
A
(M,
).
The functor A
A
(?) preserves sphere modules and therefore cell modules. This
implies formally that the adjuction passes to derived categories, giving
D
A
(A
A
M, M
= D
A
(M,
).
If M is a cell A-module, then
id : M
= A
A
M
(A
A
M)
is a quasi-isomorphism of A-modules. These maps give the unit of the adjunction.
Its counit is given by the maps of A
-modules
id
: A
A
M
A
M
= M
,
where M
is a quasi-isomorphism of A-
modules. Since the composite of this map with the quasi-isomorphism id for
the A-module M
= M
k
(L, Hom
A
(M, N)),
where A acts on L M through its action on M (with the usual sign convention:
a( m) = (1)
degadeg
am). This isomorphism clearly passes to homotopy
categories. Letting L run through the sphere k-complexes and using (3.1) and
the Whitehead theorem, we see that if M is a cell A-module then the functor
Hom
A
(M, N) preserves quasi-isomorphisms in N.
This allows us to dene Hom
A
(M, N) in D
A
for arbitrary modules M and N by
rst replacing M by a cell approximation M and then taking Hom
A
(M, N) on
the level of modules. Thus, in D
k
, Hom
A
(M, N) means Hom
A
(M, N). This gives
a well-dened functor such that
(3.4) D
A
(L M, N)
= D
k
(L, Hom
A
(M, N)).
Remark 3.5. The argument we have just run through is a special case of a general
one. If S and T are left and right adjoint functors between two categories of the
sort that we are considering, then S preserves objects of the homotopy type of cell
modules if and only T preserves quasi-isomorphisms, and in that case the resulting
induced functors on derived categories are still adjoint. See [9, I.5.13] for a precise
categorical statement.
We can now dene dierential Tor and Ext (or hyperhomology and hypercoho-
mology) groups as follows. We ignore questions of justication in terms of standard
homological terminology, some of which we believe to be antiquated.
Denition 3.6. Working in derived categories, dene
Tor
A
(M, N) = H
(M
A
N) and Ext
A
(M, N) = H
(Hom
A
(M, N)).
It is usual to regrade Ext cohomologically, along with Hom. If we specialize by
setting k = Z and letting A be a ring, thought of as a DGA concentrated in degree
zero and with zero dierential, then these groups Tor and Ext are the Tor and Ext
groups of classical homological algebra. We can check this by comparing denitions,
but it is more satisfactory to make this treatment self-contained by checking that
our functors satisfy the axioms that characterize the classical functors.
4. Some spectral sequences
This section is addressed to those who have seen spectral sequences. It will
not be used in later sections. No matter how Tor and Ext are dened in the
generality of modules over DGAs, the essential point is to have Eilenberg-Moore,
or hyperhomology, spectral sequences for their calculation.
Theorem 4.1. There are natural spectral sequences of the form
(4.2) E
2
p,q
= Tor
HA
p,q
(H
M, H
N) =Tor
A
p+q
(M, N)
and
(4.3) E
2
p,q
= Ext
HA
p,q
(H
M, H
N) =Ext
A
p+q
(M, N).
10 J.P. MAY
These are both spectral sequences of homological type, with
(4.4) d
r
: E
r
p,q
E
r
pr,q+r1
.
In (4.2), p is the usual homological degree, the spectral sequence is non-zero only
in the right half-plane, and it converges strongly. In (4.3), p is the negative of
the usual cohomological degree, the spectral sequence is non-zero only in the left
half plane, and it converges strongly if, for each xed (p, q), only nitely many of
the dierentials (4.4) are non-zero. (The best study of the convergence of spectral
sequences, is given in [1].)
Our construction of the spectral sequences follows [6], which is a precursor of
the present approach to derived categories. Let : P N be a quasi-isomorphism
of left A-modules, where P is a cell A-module. Rewrite the cellular ltration of P
by setting F
n
P = P
n+1
. Thus
0 = F
1
P F
0
P F
1
P F
n
P .
The ltration gives rise to a spectral sequence that starts from
E
0
p,q
P = (F
p
P/F
p1
P)
p+q
= A (
P
p,
)
p+q
,
where
P
p,
is k-free on the canonical basis elements of the open cells of P
p
. The
denition of a cell module implies that d
0
= d 1. Therefore
E
1
p,
P
= H
(A)
P
p,
.
Thinking of N as ltered with F
1
N = 0 and F
p
N = N for p 0, we see that
E
1
,
P gives a complex of left H
(A)-modules
(4.5) E
1
p+1,
P E
1
p,
P E
1
0,
P H
(N) 0.
Denition 4.6. Let P be a cell A-module. A quasi-isomorphism : P N is said
to be a distinguished resolution of N if the sequence (4.5) is exact, so that {E
1
p,
P}
is a free H
(A)-resolution of H
(N).
Observe that : P N is necessarily a homotopy equivalence if N is a cell A-
module, by Whiteheads theorem. The following result, which is due to Gugenheim
and myself [6, 2.1] and will not be reproven here, should be viewed as a greatly
sharpened version of Theorem 2.6: it gives cell approximations with precisely pre-
scribed algebraic properties.
Theorem 4.7 (Gugenheim-May). For any A-module N, every free H
(A)-resolu-
tion of H
= D
A
(L, Hom
A
(M, N)).
We also discuss duality, characterizing the (strongly) dualizable objects or, in an-
other language, identifying the largest rigid tensored subcategory of D
A
. Again,
in D
A
, M
A
N means M
A
N. Since A is commutative, this is an A-module.
From our present point of view, it makes good sense to resolve both variables since
we now have canonical isomorphisms of A-modules
S
q
A
A
S
r
A
= S
q+r
A
As in Proposition 1.5(iii), this directly implies that tensor products of cell A-
modules are cell A-modules.
Proposition 5.2. If M and M
is a cell A-
module with sequential ltration {
p
(M
p
A
N
np
)}. It has a (q +r)-cell for each
pair consisting of a q-cell of M
p
and an r-cell of M
np
, 0 p n.
For A-modules M and N, Hom
A
(M, N) is an A-module such that
(5.3) M
A
(L
A
M, N)
= M
A
(L, Hom
A
(M, N)).
In D
A
, Hom
A
(M, N) means Hom
A
(M, N), and we have the isomorphism (5.1).
There are general accounts of duality theory in the context of symmetric monoidal
categories in the literature of both algebraic geometry [4, 1], [3] and algebraic
topology [5], [9, III1]. I have recently given what I hope is an easily readable
exposition [12]. I will recall some of the ideas. Observe rst that, by an easy direct
inspection of denitions, the functor Hom
A
(M, N) preserves cober sequences in
both variables. (Actually, in the variable M, the functor Hom
A
(M, N) converts an
exact triangle into the negative of an exact triangle.)
The dual of an A-module M, denoted M
M
A
DM
Hom
A
(M, M) DM
A
M
o
12 J.P. MAY
The denition has many purely formal implications. The map of (5.4) is an
isomorphism (in D
A
) if either L or N is dualizable. The map of (5.5) is an
isomorphism if and only if M is dualizable, and the coevaluation map is then the
composite
1
in (5.6). The natural map
: M DDM
is an isomorphism if M is dualizable. The natural map
: Hom
A
(M, N)
A
Hom
A
(M
, N
) Hom
A
(M
A
M
, N
A
N
)
is an isomorphism if M and M
.
Theorem 5.7. A cell A-module is dualizable if and only if it is a direct summand
up to homotopy of a nite cell A-module.
Proof. Observe rst that S
q
A
is dualizable with dual S
q
A
, hence any nite direct
sum of A-modules S
q
A
is dualizable. Observe next that the cober of a map between
dualizable A-modules is dualizable. In fact, the evaluation map induces a natural
map
#
: D
A
(L, N
A
DM) D
A
(L
A
M, N),
and M is dualizable if and only if
#
is an isomorphism for all L and N [9, III.3.6].
Since both sides turn cober sequences in the variable M into long exact sequences,
the ve lemma gives the observation. We conclude by induction on the number of
cells that a nite cell A-module is dualizable. It is formal that a direct summand
in D
A
of a dualizable A-module is nite. For the converse, let M be a dualizable
cell A-modulewith coevaluation map : A M
A
DM. Clearly factors through
N
A
DM for some nite cell subcomplex N of M. By a diagram chase ([9, III.1.2]),
the bottom composite in the following commutative diagram is the identity (in D
A
):
N
A
DM
A
M
1
/
N
A
A
=
/
N
M
= A
A
M
5
k
k
k
k
k
k
k
k
k
k
k
k
k
k
1
/
M
A
DM
A
M
1
/
M
A
A
=
/
M
Therefore M is a retract up to homotopy and thus, by a comparison of exact
triangles, a direct summand up to homotopy of N. (Retractions split in triangulated
categories.)
Let F
A
be the full subcategory of M
A
whose objects are the direct summands
up to homotopy of nite cell A-modules. In the language of [4, 1.7], the theorem
states that the homotopy category hF
A
is the largest rigid tensored subcategory
of the derived category D
A
. Note that the sequential ltration of a nite cell A-
module can be arranged so that a single cell is attached at each stage. That is,
such a module is just a nite sequence of extensions by free modules on a single
generator, and each quotient module M
n
/M
n1
has the form S
q
A
for some q. A
direct summand up to homotopy of a nite cell A-module, which is the appropriate
analog in D
A
of a nitely generated projective A-module, need not be an actual
direct summand and need not be isomorphic in D
A
to a nite cell A-module. The
DERIVED CATEGORIES FROM A TOPOLOGICAL POINT OF VIEW 13
situation demands the introduction and study of the K-theory group K
0
(F
A
), but
we shall desist.
6. Browns representability theorem
We revert to a general DGA A, not necessarily commutative. Functors of coho-
mological type on D
A
are of considerable interest, and we here recall a categorical
result that characterizes when they can be represented in the form D
A
(?, N). The
topological analogue has long played an important role.
We have said that we think of the S
q
A
as analogs of sphere spectra. Just as
maps out of spheres calculate homotopy groups and therefore detect weak equiv-
alences, so maps out of the S
q
A
calculate homology groups and therefore detect
quasi-isomorphisms. We display several versions of this fact for later use: for all
A-modules N,
(6.1) H
q
(N)
= hM
k
(k, NS
q
)
= hM
k
(S
q
, N)
= hM
A
(S
q
A
, N)
= D
A
(S
q
A
, N).
The category D
A
has homotopy limits and colimits. These are weak limits
and colimits in the sense that they satisfy the existence but not the uniqueness
property of categorical limits and colimits. For example, the homotopy pushout of
maps f : L M and g : L N is obtained from M (L I) N by identifying
l[0] with f(l) and l[1] with g(l). More precisely, we rst apply cell approximation
and then apply the cited construction. We used a similar homotopy coequalizer in
the proof of Theorem 2.6. The homotopy colimit, or telescope TelM
i
, of a sequence
of maps f
i
: M
i
M
i+1
is the homotopy coequalizer of Id: M
i
M
i
and
f
i
: M
i
M
i
; equivalently, it is the cober of g : M
i
M
i
, where
g(m) = m f
i
(m) for m M
i
. We now have enough information to quote the
categorical form of Browns representability theorem given in [2], but we prefer to
run through a quick concrete version of the proof.
Theorem 6.2 (Brown). A contravariant functor J : D
A
Sets is representable in
the form J(M)
= D
A
(M, N) for some A-module N if and only if J converts direct
sums to direct products and converts homotopy pushouts to weak pullbacks.
Proof. Necessity is obvious. Thus assume given a functor J that satises the spec-
ied direct sum and Mayer-Vietoris axioms. Since homotopy coequalizers and tele-
scopes can be constructed from sums and homotopy pushouts, J converts homotopy
coequalizers to weak equalizers and telescopes to weak limits. Write f
= J(f) for
a map f. Consider pairs (M, ) where M is an A-module and J(M).
Starting with an arbitrary pair (N
0
,
0
), we construct a sequence of pairs (N
i
,
i
)
and maps f
i
: N
i
N
i+1
such that f
i
(
i+1
) =
i
. Let N
1
= N
0
(S
q
A
), where
there is a copy of S
q
A
for each element of each set J(S
q
A
). Let
1
have coordinates
and the elements , and let f
0
: N
0
N
1
be the inclusion. Inductively, given
(N
i
,
i
), let L
i
be the sum of a copy of S
q
A
for each q and each unequal pair (x, y) of
elements of H
q
(N
i
) such that, when thought of as maps S
q
A
N
i
in D
A
, x
(
i
) =
y
(
i
). Let f
i
: N
i
N
i+1
be the coequalizer of the pair of maps L
i
N
i
given by the xs and the ys. By the weak equalizer property, there is an element
i+1
J(N
i+1
) such that f
i
(
i+1
) =
i
.
Let N = Tel N
i
. By the weak limit property, there is an element J(N) that
pulls back to
i
for each i. For an A-module M, dene
: D
A
(M, N) J(M) by
14 J.P. MAY
(f) = f
(x) =
(y). Replacing M
by a cell approximation if necessary, we can assume that x and y are given by maps
M N. Let c : N N
0
be the homotopy coequalizer of x and y and choose an
element
0
J(N
0
) such that c
0
) = . Construct a pair (N
) by repeating
the construction above, but starting with the pair (N
0
,
0
). Let j : N
0
N
be
the evident map such that j
) =
0
. Then, since (jc)
) = and both
and
is an isomorphism in D
A
. Since
cx = cy by construction, it follows that x = y. Therefore
such that i
) = and a map j : N N
such that j
) = .
Again, j is an isomorphism in D
A
since both
and
() and