QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA J. DANIEL CHRISTENSEN AND MARK HOVEY Abstract. An important example of a model category is the category of unbounded chain complexes of R-modules, which has as its homotopy category the derived cat- egory of the ring R. This example shows that traditional homological algebra is encompassed by Quillen's homotopical algebra. The goal of this paper is to show that more general forms of homological algebra also fit into Quillen's framework. Specifically, a projective class on a complete and cocomplete abelian category A is exactly the information needed to do homological algebra in A. The main result is that, under weak hypotheses, the category of chain complexes of objects of A has a model category structure that reflects the homological algebra of the projective class in the sense that it encodes the Ext groups and more general derived functors. Examples include the "pure derived category" of a ring R, and derived categories capturing relative situations, including the projective class for Hochschild homology and cohomology. We characterize the model structures that are cofibrantly gener- ated, and show that this fails for many interesting examples. Finally, we explain how the category of simplicial objects in a possibly non-abelian category can be equipped with a model category structure reflecting a given projective class, and give examples that include equivariant homotopy theory and bounded below derived categories. Contents Introduction 2 0. Notation and conventions 3 1. Projective classes 4 1.1. Definition and some examples 4 1.2. Homological algebra 5 1.3. Pullbacks 5 1.4. Strong projective classes 6 2. The relative model structure 7 2.1. Properties of the relative model structure 11 2.2. Why projective classes? 15 3. Case A: Projective classes coming from adjoint pairs 15 3.1. Examples 17 4. Case B: Projective classes with enough small projectives 19 5. Cofibrant generation 21 5.1. Background 21 5.2. Projective classes with sets of enough small projectives 22 5.3. The pure and categorical derived categories 24 5.4. Failure to be cofibrantly generated 24 6. Simplicial objects and the bounded below derived category 26 6.1. Projective classes in pointed categories 27 6.2. The model structure 27 References 28 _________________ Date: April 6, 2002. 1991 Mathematics Subject Classification. Primary 18E30; Secondary 18G35, 55U35, 18G25, 55U15. Key words and phrases. Derived category, chain complex, relative homological algebra, projective class, model category, non-cofibrant generation, pure homological algebra. The first author was supported in part by NSF grant DMS 97-29992. The second author was supported in part by NSF grant DMS 99-70978. 1 2 J. DANIEL CHRISTENSEN AND MARK HOVEY Introduction An important example of a model category is the category Ch (R) of unbounded chain complexes of R-modules, which has as its homotopy category the derived category D(R) of the associative ring R. The formation of a projective resolution is an example of cofibrant replacement, traditional derived functors are examples of derived functors in the model category sense, and Ext groups appear as groups of maps in the derived category. This example shows that traditional homological algebra is encompassed by Quillen's homotopical algebra, and indeed this unification was one of the main points of Quillen's influential work [Qui67 ]. The goal of this paper is to illustrate that more general forms of homological algebra also fit into Quillen's framework. In any abelian category A there is a natural notion of "projective object" and "epimorphism." However, it is sometimes useful to impose different definitions of these terms. If this is done in a way that satisfies some natural axioms, what is obtained is a "projective class," which is exactly the information needed to do homological algebra in A. Our main result shows that for a wide variety of projec- tive classes (including all those that arise in examples) the category of unbounded chain complexes of objects of A has a model category structure that reflects the homological algebra of the projective class in the same way that ordinary homological algebra is captured by the usual model structure on Ch (R). When A has enough projectives, the projective objects and epimorphisms form a projective class. Therefore the results of this paper apply to traditional homological algebra as well. Even in this special case, it is not a trivial fact that the category of unbounded chain complexes can be given a model category structure, and indeed Quillen restricted himself to the bounded below case. We know of three other written proofs that the category of unbounded chain complexes is a model category [Gro , Hin97 , Hov98 ], which do the case of R-modules, but this was probably known to others as well. An important corollary of the fact that a derived category D(A) is the homotopy category of a model category is that the group D(A)(X, Y ) of maps is a set (as opposed to a proper class) for any two chain complexes X and Y . This is not the case in general, and much work on derived categories ignores this possibility. The importance of this point is that if one uses the morphisms in the derived category to index constructions in other categories or to define cohomology groups, one needs to know that the indexing class is actually a set. Recently, D(A)(X, Y ) has been shown to be a set under various assumptions on A. (See Weibel [Wei94 ] Remark 10.4.5, which credits Gabber, and Exercise 10.4.5, which credits Lewis, May and Steinberger [LMS86 ]. See also Kriz and May [KM95 , Part III].) The assumptions that appear in the present paper are different from those that have appeared before and the proof is somewhat easier (because of our use of the theory of cofibrantly generated model categories), so this paper may be of some interest even in this special case. Another consequence of the fact that Ch (R) is a model category is the existence of resolutions coming from cofibrant and fibrant approximations, and the related derived functors. Some of these are discussed in [AFH97 ] and [Spa88 ]. We do not discuss these topics here, but just mention that these resolutions are immediate once you have the model structure, so our approach gives these results with very little work. While our results include new examples of traditional homological algebra, our focus is on more general projective classes. For example, let A be an algebra over a commutative ring k. We call a map of A-bimodules a relative epimorphism if it is split epic as a map of k-modules, and we call an A-bimodule a relative projective if maps from it lift over relative epimorphisms. These definitions give a projective class, and Theorem 2.2 tells us that there is a model category, and therefore a derived category, that captures the homological algebra of this situation. For example, Hochschild cohomology groups appear as Hom sets in this derived category (see Example 3.7). QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 3 We also discuss pure homological algebra and construct the "pure derived category" of a ring. Pure homological algebra has applications to phantom maps in the stable homo- topy category [CS98 ] and in the (usual) derived category of a ring [Chr98 ], connections to Kasparov KK-theory [Sch01 ], and is actively studied by algebraists and representation theorists. In the last section we describe a model category structure on the category of non- negatively graded chain complexes that works for an arbitrary projective class on an abelian category, without any hypotheses. More generally, we show that under appro- priate hypotheses a projective class on a possibly non-abelian category A determines a model category structure on the category of simplicial objects in A. As an example, we deduce that the category of equivariant simplicial sets has various model category structures. We now briefly outline the paper. In Section 1 we give the axioms for a projective class and mention many examples that will be discussed further in Subsections 3.1 and 5.3. In Section 2 we describe the desired model structure coming from a projective class and state our main theorem, which says that the model structure exists as long as cofibrant replacements exist. We also give two hypotheses that each imply the existence of cofibrant replacements. The first hypothesis handles situations coming from adjoint pairs, and is proved to be sufficient in Section 3, where we also give many examples involving relative situations. The second hypothesis deals with projective classes that have enough small projectives and is proved to be sufficient in Section 4. In Section 5 we prove that the model structure that one gets is cofibrantly generated if and only if there is a set of enough small projectives. We do this using the recognition lemma for cofibrantly generated categories, which is recalled in Subsection 5.1. This case is proved from scratch, independent of the main result in Section 2, since the proof is not long. In Subsection 5.3 we give two examples, the traditional derived category of R-modules and the pure derived category. We describe how the two relate and why the pure derived category is interesting. In the final section we discuss the bounded below case, which works for any projective class, and describe a result for simplicial objects in a possibly non-abelian category. We thank Haynes Miller for asking the question that led to this paper and Haynes Miller and John Palmieri for fruitful and enjoyable discussions. 0. Notation and conventions We make a few blanket assumptions. With the exception of Section 6, A and B will denote abelian categories. We will assume that our abelian categories are bicomplete; this assumption is stronger than strictly necessary, but it simplifies the statements of our results. For any category C, we write C(A, B) for the set of maps from A to B in C. We write Ch (A) for the category of unbounded chain complexes of objects of A and degree zero chain maps. To fix notation, assume that the differentials lower degree. For an object X of Ch (A), define Zn X := ker (d : Xn -! Xn-1 ) and Bn X := im (d : Xn+1 -! Xn ), and write Hn X for the quotient. A map inducing an isomorphism in Hn for all n is a quasi-isomorphism. The suspension X of X has ( X)n = Xn-1 and d X = -dX . The functor is defined on morphisms by ( f )n = fn-1 . Given a map f : X -! Y of chain complexes, the cofibre of f is the chain complex C with Cn = Yn Xn-1 and with differential d(y, x) = (dy + f x, -dx). There are natural maps Y -! C -! X, and the sequence X -! Y -! C -! X is called a cofibre sequence. Two maps f, g : X -! Y are chain homotopic if there is a collection of maps sn : Xn -! Yn+1 such that f - g = ds + sd. We write [X, Y ] for the chain homo- topy classes of maps from X to Y and K(A) for the category of chain complexes and chain homotopy classes of maps. Two complexes are chain homotopy equivalent if they are isomorphic in K(A), and a complex is contractible if it is chain homotopy 4 J. DANIEL CHRISTENSEN AND MARK HOVEY equivalent to 0. K(A) is a triangulated category, with triangles the sequences chain homotopy equivalent to the cofibre sequences above. The functors Hn (-), [X, -] and [-, Y ] are defined on K(A) and send triangles to long exact sequences. For P in A, Dk P denotes the (contractible) complex such that (Dk P )n = P if n = k or n = k - 1 but (Dk P )n = 0 for other values of n, and whose differential is the identity in degree k. The functor Dk is left adjoint to the functor X 7! Xk , and right adjoint to the functor X 7! Xk-1 . The path complex P X of a complex X is the contractible complex such that (P X)n = Xn Xn+1 , where d(x, y) = (dx, x - dy). We assume knowledge of the basics of model categories, for which [DS95 ] is an excel- lent reference. We use the definition of model category that requires that the category be complete and cocomplete, and that the factorizations be functorial. Correspond- ingly, when we say that cofibrant replacements exist, we implicitly mean that they are functorial. 1. Projective classes 1.1. Definition and some examples. In this subsection we explain the notion of a projective class, which is the information necessary in order to do homological algebra. Intuitively, a projective class is a choice of which sort of "elements" we wish to think about. In this section we focus on the case of an abelian category, but this definition works for any pointed category with kernels. The elements of a set X correspond bijectively to the maps from a singleton to X, and the elements of an abelian group A correspond bijectively to the maps from Z to A. Motivated by this, we call a map P -! A in any category a P -element of A. If we don't wish to mention P , we call such a map a generalized element of A. A map A -! B in any category is determined by what it does on generalized elements. If P is a collection of objects, then a P-element means a P -element for some P in P. Let A be an abelian category. A map B -! C is said to be P -epic if it induces a surjection of P -elements, that is, if the induced map A(P, B) -! A(P, C) is a surjection of abelian groups. The map B -! C is P-epic if it is P -epic for all P in P. Definition 1.1. A projective class on A is a collection P of objects of A and a collection E of maps in A such that (i ) E is precisely the collection of all P-epic maps; (ii ) P is precisely the collection of all objects P such that each map in E is P -epic; (iii ) for each object B there is a map P -! B in E with P in P. When a collection P is part of a projective class (P, E ), the projective class is unique, and so we say that P determines a projective class or even that P is a projective class. An object of P is called a P-projective, or, if the context is clear, a relative projective. A sequence A -! B -! C is said to be P -exact if the composite A -! C is zero and A(P, A) -! A(P, B) -! A(P, C) is an exact sequence of abelian groups. The latter can be rephrased as the condition that A -! B -! C induces an exact sequence of P -elements. A P-exact sequence is one that is P -exact for all P in P. Example 1.2. For an associative ring R, let A be the category of left R-modules, let P be the collection of all summands of free R-modules and let E be the collection of all surjections of R-modules. Then E is precisely the collection of P-epimorphisms, and P is a projective class. The P-exact sequences are the usual exact sequences. QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 5 Example 1.2 is a categorical projective class in the sense that the P-epimorphisms are just the epimorphisms and the P-projectives are the categorical projectives, i.e., those objects P such that maps from P lift through epimorphisms. Here are two examples of non-categorical projective classes. Example 1.3. If A is any abelian category, P is the collection of all objects, and E is the collection of all split epimorphisms B -! C, then P is a projective class. It is called the trivial projective class. A sequence A -! B -! C is P-exact if and only if A -! ker(B -! C) is split epic. Example 1.4. Let A be the category of left R-modules, as in Example 1.2. Let P consist of all summands of sums of finitely presented modules and define E to consist of all P-epimorphisms. Then P is a projective class. A sequence is P-exact iff it is exact after tensoring with every right module. Examples 1.2 and 1.4 will be discussed further in Subsection 5.3. Example 1.3 is important because many interesting examples are "pullbacks" of this projective class (see Subsection 1.3). Let P be a projective class. If S is a subcollection of P (not necessarily a set), and if a map is S-epic iff it is P-epic, then we say that P is determined by S and that S is a collection of enough projectives. Some projective classes, such as Examples 1.2 and 1.4, are determined by a set, and the lemma below shows that any set of objects determines a projective class. The trivial projective class is sometimes not determined by a set (see Subsection 5.4). Lemma 1.5. Suppose F is any set of objects in an abelian category with coproducts. Let E be the collection of F -epimorphisms and let P be the collection of all objects P such that every map in E is P -epic. Then P is the collection of retracts of coproducts of objects of F and (P, E ) is a projective class. ` Proof. Given an object X, let P be the coproduct F indexed by all maps F -! X and all objects F in F . The natural map P -! X is clearly an F -epimorphism. Moreover, if X is in P, then this map is split epic, and so X is a retract of a coproduct of objects of F . These two facts show that (P, E ) is a projective class. 1.2. Homological algebra. A projective class is precisely the information needed to form projective resolutions and define derived functors. All of the usual definitions and theorems go through. A P-resolution of an object M is a P-exact sequence . . .-! P2 -! P1 -! P0 -! M -! 0 such that each Pi is in P. If B is an abelian category and T : A -! B is an additive functor, then the nth left derived functor of T with respect to P is defined by LPnT (M ) = Hn (T (P* )) where P* is a P-resolution of M . One has the usual uniqueness of resolutions up to chain homotopy and so this is well-defined. From a P-exact sequence 0 -! L -! M -! N -! 0 one gets a long exact sequence involving the derived functors. The abelian groups Ext nP(M, N ) can be defined in the usual two ways, as equivalence classes of P-exact sequences 0 -! N -! L1 -! . . .-! Ln -! M -! 0, or as LPnT (M ) where T (-) = A(-, N ). For further details and useful results we refer the reader to the classic reference [EM65 ]. 1.3. Pullbacks. A common setup in relative homological algebra is the following. We assume we have a functor U : A -! B of abelian categories, together with a left adjoint F : B -! A. Then U and F are additive, U is left exact and F is right exact. If (P0, E 0) is a projective class on B, we define P := {retracts of F P for P in P0 } and E := {B -! C such that U B -! U C is in E 0 }. Then one can easily show that (P, E ) is a projective class on A and that a sequence is P-exact if and only if it is sent to a P0-exact sequence by U . (P, E ) is called the pullback of (P0, E 0) along the right adjoint U . 6 J. DANIEL CHRISTENSEN AND MARK HOVEY The most common case is when (P0, E 0) is the trivial projective class (see Example 1.3). Then for any M in A the counit F U M -! M is a P-epimorphism from a P-projective. Example 1.6. Let R -! S be a map of associative rings. Write R-mod and S-mod for the categories of left R- and S-modules. Consider the forgetful functor U : S-mod -! R-mod and its left adjoint F that sends an R-module M to S R M . The pullback along U of the trivial projective class on R-mod gives a projective class P. The P-projectives are the S-modules P such that the natural map S R P -! P is split epic as a map of S-modules. The P-epimorphisms are the S-module maps that are split epic as maps of R-modules. Example 1.7. As above, let R -! S be a map of rings. The forgetful functor U : S-mod -! R-mod has a right adjoint G that sends an R-module M to the S-module R-mod (S, M ). We can pullback the trivial injective class along U to get an injective class on S-mod . (An injective class is just a projective class on the opposite category.) The relative injectives are the S-modules I such that the natural map I -! R-mod (S, I) is split monic as a map of S-modules, and the relative monomorphisms are the S-module maps that are split monic as maps of R-modules. We investigate these examples in detail in Section 3. 1.4. Strong projective classes. In Section 3 we will focus on projective classes that are the pullback of a trivial projective class along a right adjoint. In this subsection we describe a special property that these projective classes have. Definition 1.8. A projective class P is strong if for each P-projective P and each P-epimorphism M -! N , the surjection A(P, M ) -! A(P, N ) of abelian groups is split epic. It is clear that a trivial projective class is strong, and that the pullback of a strong projective class is strong. The importance of strong projective classes comes from the following lemma. Lemma 1.9. The following are equivalent for a projective class P on an abelian category A: (i ) P is strong. (ii ) For each complex C in Ch (A), if the complex A(P, C) in Ch (Ab ) has trivial ho- mology for each P in P, then it is contractible for each P in P. (iii ) For each map f in Ch (A), if the map A(P, f ) in Ch (Ab ) is a quasi-isomorphism for each P in P, then it is a chain homotopy equivalence for each P in P. Here A(P, C) denotes the chain complex with A(P, Ck ) in degree k. We also use the notation from Section 0. Proof. (i) =) (ii): Assume P is a strong projective class and let C be a complex in Ch (A) such that A(P, C) has trivial homology for each P-projective P . Then for each k and each P we have a short exact sequence 0 -! A(P, Zk C) -! A(P, Ck ) -! A(P, Zk-1 C) -! 0 of abelian groups. Because the projective class is strong, the sequence is split. This implies that the complex A(P, C) is isomorphic to k Dk+1 A(P, Zk C), and in particular that it is contractible. (ii) =) (iii): Let f : X -! Y be a map in Ch (A) such that A(P, X) -! A(P, Y ) is a quasi-isomorphism for each P in P, and let C be the cofibre of f . Then by the long exact sequence, A(P, C) has trivial homology for each P in P. By (ii) this complex is contractible. This implies that A(P, X) -! A(P, Y ) is a chain homotopy equivalence. (iii) =) (i): Let M -! N be a P-epimorphism with kernel L. Then the complex L -! M -! N has trivial homology after applying A(P, -), for each P-projective P . By QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 7 (iii), it is contractible after applying A(P, -), for each P . In particular, A(P, M ) -! A(P, N ) is split epic. 2. The relative model structure The object of this section is to construct a Quillen model structure on the category Ch (A) of chain complexes over A that reflects a given projective class P on A. If X is a chain complex, we write A(P, X) for the chain complex that has the abelian group A(P, Xn ) in degree n. This is the chain complex of P -elements of X. Definition 2.1. A map f : X -! Y in Ch (A) is a P-equivalence if A(P, f ) is a quasi- isomorphism in Ch (Z) for each P in P. The map f is a P-fibration if A(P, f ) is a surjection for each P in P. The map f is a P-cofibration if f has the left lifting property with respect to all maps that are both P-fibrations and P-equivalences (the P-trivial fibrations). The motivation for this definition is that it implies that a complex . . .-! P2 -! P1 -! P0 -! 0 -! . . . equipped with an augmentation P0 -! M to an object M is a cofibrant replacement if and only if it is a P-resolution in the sense of Subsection 1.2. This implies that if M and N are objects of A thought of as complexes concentrated in degree zero, then Ext nP(M, N ) can be identified with maps from n M to N in the homotopy category of the model category Ch (A). This will be described in more detail in Subsection 2.1. The main goal of this section is then to prove the following theorem. Theorem 2.2. Suppose P is a projective class on the abelian category A. Then the cat- egory Ch (A), together with the P-equivalences, the P-fibrations, and the P-cofibrations, forms a Quillen model category if and only if cofibrant replacements exist. When the model structure exists, it is proper. Cofibrant replacements exist in each of the following cases: A: P is the pullback of the trivial projective class along a right adjoint that preserves countable sums. B: There are enough ~-small P-projectives for some cardinal ~, and P-resolutions can be chosen functorially. The words "enough" and "~-small" will be explained in Section 4. We call this structure the relative model structure. We point out that we are using the modern definition of model category [DHK97 , Hov98 ], so our factorizations will be functorial. Correspondingly, we require our cofibrant replacements to be functo- rial as well. This theorem requires our blanket assumption that abelian categories are bicomplete. In Subsection 2.1 we will describe further properties of these model structures, in- cluding conditions under which they are monoidal. In Section 5 we show that if there is a set of enough small projectives, then the model structure is cofibrantly generated. On the other hand, we show in Subsection 5.4 that model categories coming from Case A are generally not cofibrantly generated. In Subsection 2.2 we explain why it is no loss of generality to start with a projective class (P, E ), rather than just an arbitrary class P of test objects. Proof. Some of the properties necessary for a model category are evident from the def- initions. It is clear that Ch (A) is bicomplete, since A is so. Also, P-equivalences have the two out of three property, and P-equivalences, P-fibrations, and P-cofibrations are closed under retracts. Furthermore, P-cofibrations have the left lifting property with re- spect to P-trivial fibrations, by definition. It remains to show that P-trivial cofibrations 8 J. DANIEL CHRISTENSEN AND MARK HOVEY have the left lifting property with respect to P-fibrations, and that the two factoriza- tion axioms hold. The remaining lifting property will be proved in Proposition 2.8, and the two factorization axioms will be proved in Propositions 2.9 and 2.10, assuming that cofibrant replacements exist. Properness will be proved in Proposition 2.18, which also defines the term. That cofibrant replacements exist in cases A and B will be proved in Sections 3 and 4, respectively. Note that our Theorem 2.2 will also apply when we have an injective class, that is, a projective class on Aop , by dualizing the definition of the model structure. We continue the proof of Theorem 2.2 with a lemma that gives us a simple test of the lifting property. We use the notation from Section 0. Lemma 2.3. Suppose p : X -! Y is a P-fibration with kernel K, and i : A -! B is a degreewise split inclusion whose cokernel C is a complex of relative projectives. If every map C -! K is chain homotopic to 0, then i has the left lifting property with respect to p. Proof. We can write Bn ~= An Cn , where the differential is defined by d(a, c) = (da+o c, dc) (we use the element notation for convenience, but it is not strictly necessary), and o : Cn -! An-1 can be any family of maps such that do + o d = 0. Suppose we have a commutative square as below. A -- -f- ! X ? ? i?y ?yp B -- -g- ! Y In terms of the splitting Bn ~= An Cn , we have g(a, c) = pf (a) + ff(c), where the family ffn : Cn -! Yn satisfies dff = pf o + ffd. We are looking for a map h : B -! X making the diagram above commute. In terms of the splitting, this means we are looking for a family of maps fin : Cn -! Xn such that pfi = ff and dfi = f o + fid. Since Cn is relatively projective, and p is P-epic, there is a map fl : Cn -! Xn such that pfl = ff. The difference ffi = dfl - f o - fld may not be zero, but at least pffi = 0. Let j : K -! X denote the kernel of p. Then there is a map F : Cn -! Kn-1 such that jF = ffi. Furthermore, one can check that F d = -dF , so that F : C -! K is a chain map. By hypothesis, F is chain homotopic to 0 by a map D : Cn -! Kn , so that Dd - dD = F . Define fi = fl + jD. Then fi defines the desired lift, so i has the left lifting property with respect to p. Now we study the P-cofibrations. A complex C is called P-cofibrant if the map 0 -! C is a P-cofibration. A complex K is called weakly P-contractible if the map K -! 0 is a P-equivalence, or, equivalently, if all maps from a complex k P consisting of a relative projective concentrated in one degree to K are chain homotopic to 0. Lemma 2.4. A complex C is P-cofibrant if and only if each Cn is relatively projective and every map from C to a weakly P-contractible complex K is chain homotopic to 0. Proof. Suppose first that C is P-cofibrant. If M -! N is a P-epimorphism, then the map Dn+1 M -! Dn+1 N is a P-fibration. It is also a P-equivalence, since it is in fact a chain homotopy equivalence. Since C is P-cofibrant, the map Ch (A)(C, Dn+1 M ) -! Ch (A)(C, Dn+1 N ) is surjective. But this map is isomorphic to the map A(Cn , M ) -! A(Cn , N ), so Cn is relatively projective. If K is weakly P-contractible, then the natural map P K -! K is a P-trivial fibration. Since C is P-cofibrant, any map C -! K factors through P K, which means that it is chain homotopic to 0. QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 9 The converse follows immediately from Lemma 2.3, since the kernel of a P-trivial fibration is weakly P-contractible. Proposition 2.5. A map i : A -! B is a P-cofibration if and only if i is a degreewise split monomorphism with P-cofibrant cokernel. Proof. Suppose first that i is a P-cofibration with cokernel C. Since P-cofibrations are closed under pushouts, it is clear that C is P-cofibrant. The map Dn+1 An -! 0 is a P-fibration and a P-equivalence. Since i is a P-cofibration, the map A -! Dn+1 An that is the identity in degree n extends to a map B -! Dn+1 An . In degree n, this map defines a splitting of in . Conversely, suppose that i is a degreewise split monomorphism and the cokernel C of i is P-cofibrant. We need to show that i has the left lifting property with respect to any P-trivial fibration p : X -! Y . But this follows by combining Lemmas 2.3 and 2.4. The next lemma provides a source of P-cofibrant objects, including the P-cellular complexes. Definition 2.6. Call a complex C purely P-cellular if it is a colimit of a colimit- preserving diagram 0 = C0 -! C1 -! C2 -! . . . indexed by an ordinal fl, such that for each ff < fl the map Cff -! Cff+1 is degreewise split monic with cokernel a complex of relative projectives with zero differential. By "colimit- preserving" we mean that for each limit ordinal ~ < fl, the map colim ff<~ Cff -! C~ is an isomorphism. We say C is P-cellular if it is a retract of a purely P-cellular complex. Lemma 2.7. (a) If D in Ch (A) is a complex of relative projectives with zero dif- ferential, then D is P-cofibrant. (b) If D in Ch (A) is a bounded below complex of relative projectives, then D is P-cofibrant. (c) If D in Ch (A) is P-cellular, then D is P-cofibrant. Proof. (a) follows immediately from Lemma 2.4. (b) Let D be a bounded below complex of relative projectives and write D n for the truncation of D that agrees with D in degrees n and is 0 elsewhere. Then the map D n -! D n+1 is degreewise split monic and has a P-cofibrant cokernel (by (a)), so is a P-cofibration by Proposition 2.5. Since D n = 0 for n << 0 and 0 is P-cofibrant, each D n is P-cofibrant. Therefore, so is their colimit D. (c) The proof is just a transfinite version of the proof of (b), combined with the fact that a retract of a cofibrant object is cofibrant. We can now prove that the other lifting axiom holds. Proposition 2.8. A map i : A -! B has the left lifting property with respect to P- fibrations if and only if i is a P-trivial cofibration. Proof. Suppose first that i has the left lifting property with respect to P-fibrations. Then i is a P-cofibration, by definition, and the cokernel 0 -! C also has the left lifting property with respect to P-fibrations. In particular, since the map P C -! C is a P- fibration, C is contractible. Hence i is a chain homotopy equivalence, and in particular a P-trivial cofibration. Conversely, suppose that i is a P-trivial cofibration with cokernel C. By Lemma 2.3, in order to show that i has the left lifting property with respect to P-fibrations, it suffices to show that every map from C to any complex K is chain homotopic to 0. This is equivalent to showing that C is contractible. Since i is degreewise split monic, for each relative projective P there is a long exact sequence . . .-! [ k P, A] -! [ k P, B] -! [ k P, C] -! . . .. Since i is a P-equivalence, [ k P, C] = 0 for each relative projective P 10 J. DANIEL CHRISTENSEN AND MARK HOVEY and each k, and so P C -! C is a P-trivial fibration. Since C is P-cofibrant, the identity map of C factors through P C, and so C is contractible. Note that the proof shows that a P-trivial cofibration is in fact a chain homotopy equivalence. Now we proceed to prove the factorization axioms, under the assumption that we have cofibrant replacements. Proposition 2.9. If every object A has a cofibrant replacement qA : QA -! A, then every map in Ch (A) can be factored into a P-cofibration followed by a P-trivial fibration. Proof. Suppose f : A -! B is a map in Ch (A). Let C be the cofibre of f , so C = B A with d(b, a) = (db+f a, -da), and let E be the fibre of the composite g : QC -! C -! A, so E = A QC with d(a, q) = (da - gq, dq) (the desuspension of the cofibre). Consider the diagram A ---i- ! E -- - - ! QC ---g- ! A flfl ? fl fl qC?y flfl A ---f- ! B -- - - ! C --- - ! A whose rows are triangles in K(A). There is a natural fill-in map p : E -! B defined by p(a, q) = f (a) + ssB qC q, where ssB : C -! B is the projection. The map p makes the left- hand square commute in Ch (A) and the middle square commute in K(A) (with the chain homotopy s(a, q) = (0, a)). The map i : A -! E is a P-cofibration since it is degreewise split and its cokernel QC is P-cofibrant. Furthermore, since QC -! C is degreewise P-epic and ssB : C -! B is degreewise split epic, it follows that p is degreewise P-epic. Applying the functor [ k P, -] gives two long exact sequences, and from the five-lemma one sees that [ k P, p] is an isomorphism when P is a relative projective. Thus f = pi is the required factorization. Proposition 2.10. If every object A has a cofibrant replacement qA : QA -! A, then every map in Ch (A) can be factored into a P-trivial cofibration followed by a P-fibration. Proof. It is well-known that we can factor any map f : A -! B in Ch (A) into a degreewise split monomorphism that is also a chain homotopy equivalence, followed by a degreewise split epimorphism. Since every degreewise split epimorphism is a P-fibration, we may as well assume f is a degreewise split monomorphism and a chain homotopy equivalence. In this case, we apply Proposition 2.9 to factor f = pi, where p is a P-trivial fibration and i is a P-cofibration. Since f is a chain homotopy equivalence, i must be a P-trivial cofibration, and so the proof is complete. Note that the factorizations constructed in Proposition 2.9 and Proposition 2.10 are both functorial in the map f , since we are implicitly assuming that cofibrant replacement is functorial. The homotopy category of Ch (A), formed by inverting the P-equivalences, is called the derived category of A (with respect to P). It is denoted D(A), and a fundamental result in model category theory asserts that D(A)(X, Y ) is a set for each X and Y . In Exercise 10.4.5 of [Wei94 ], Weibel outlines an argument that proves that D(A)(X, Y ) is a set when there are enough (categorical) projectives, P is the categorical projective class, and A satisfies AB5. A connection between Weibel's hypotheses and Case B is that if A has enough projectives that are small with respect to all filtered diagrams in A, then AB5 holds. The smallness condition needed for our theorem is weaker than this. (See Section 4 for the precise hypothesis.) QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 11 2.1. Properties of the relative model structure. In this subsection, we investigate some of the properties of the relative model structure. We begin by showing that the model category notions of homotopy, derived functor, suspension and cofibre sequence agree with the usual notions. Then we study properness and monoidal structure. We discuss cofibrant generation in Section 5. We assume throughout that P is a projective class on an abelian category A such that the relative model structure on Ch (A) exists. We first show that the notion of homotopy determined from the model category structure corresponds to the usual notion of chain homotopy. Definition 2.11. ([DS95 ] or [Qui67 ].) If M is an object in a model category C, a good ` i p cylinder object for M is an object M x I and a factorization M M -! M x I -! M of the codiagonal map, with i a cofibration and p a weak equivalence. (Despite the notation, M x I is not in general a product of M with an object I.) A left homotopy between maps`f, g : M` -! N is a map H : M x I -! N such that the composite Hi is equal to f g : M M -! N , for some good cylinder object M x I. The notion of good path object N I for N is dual to that of good cylinder object and leads to the notion of right homotopy. The following standard result can be found in [DS95 , Section 4], for example. Lemma 2.12. For M cofibrant and N fibrant, two maps f, g : M -! N are left homo- topic if and only if they are right homotopic, and both of these relations are equivalence relations and respect composition. Moreover, if M x I is a fixed good cylinder object for M , then f and g are left homotopic if and only if they are left homotopic using M x I; similarly for a fixed good path object. Because of the lemma, for M cofibrant and N fibrant we have a well-defined relation of homotopy on maps M -! N . Quillen showed that the homotopy category of C, which is by definition the category of fractions formed by inverting the weak equivalences, is equivalent to the category consisting of objects that are both fibrant and cofibrant with morphisms being homotopy classes of morphisms. Now we return to the study of the model category Ch (A). Lemma 2.13. Let M and N be objects of Ch (A) with M P-cofibrant. Two maps M -! N are homotopic if and only if they are chain homotopic. Proof. We construct a factorization M M -! M x I -! M of the codiagonal map M M -! M in the following way. Let M xI be the chain complex that has Mn Mn-1 Mn in degree n. We describe the differential by saying that it sends a generalized element (m, ~m, m0) in (M x I)n to (dm + ~m, -dm~ , dm0 - ~m). Let i : M M -! M x I be the map that sends (m, m0) to (m, 0, m0) and let p : M x I -! M be the map that sends (m, ~m, m0) to m + m0. One can check easily that M x I is a chain complex and that i and p are chain maps whose composite is the codiagonal. The map i is degreewise split monic with cokernel M , so it is a P-cofibration, since we have assumed that M is cofibrant. The map p is a chain homotopy equivalence with chain homotopy inverse sending m to (m, 0, 0); this implies that it induces a chain homotopy equivalence of generalized elements and is thus a P-equivalence. Therefore M x I is a good cylinder object for M . It is easy to see that a chain homotopy between two maps M -! N is the same as a left homotopy using the good cylinder object M x I. By Lemma 2.12, two maps are homotopic if and only if they are left homotopic using M x I. Thus the model category notion of homotopy is the same as the notion of chain homotopy when the source is P-cofibrant. There is a dual proof that proceeds by constructing a specific good path object N I for N such that a right homotopy using N I is the same as a chain homotopy. 12 J. DANIEL CHRISTENSEN AND MARK HOVEY Corollary 2.14. Let A and B be objects of A considered as chain complexes concentrated in degree 0. Then D(A)(A, n B) ~= Ext nP(A, B). See Subsection 1.2 for the definition of the Ext groups. Proof. The group D(A)(A, n B) may be calculated by choosing a P-cofibrant replace- ment A0 for A and computing the homotopy classes of maps from A0 to n B. (Re- call that all objects are P-fibrant, so there is no need to take a fibrant replacement for n B.) A P-resolution P of A serves as a P-cofibrant replacement for A, and by Lemma 2.13 the homotopy relation on Ch (A)(P, n B) is chain homotopy, so it follows that D(A)(A, n B) is isomorphic to Ext nP(A, B). More generally, a similar argument shows that the derived functors of a functor F can be expressed as the cohomology of the derived functor of F in the model category sense. To make the story complete, we next show that the shift functor corresponds to the notion of suspension that the category D(A) obtains as the homotopy category of a pointed model category. Definition 2.15. Let C be a pointed model category.` If M is cofibrant, we define the suspension M of M to be the cofibre of the map M M -! M x I for some`good cylinder object M x I. (The cofibre of a map X -! Y is the pushout * X Y , where * is the zero object.) M is cofibrant and well-defined up to homotopy equivalence. The loop object N of a fibrant object N is defined dually. These operations induce adjoint functors on the homotopy category. A straightforward argument based on the cylinder object described above (and a dual path object) proves the following lemma. Lemma 2.16. In the model category Ch (A), the functor defined in Definition 2.15 can be taken to be the usual suspension, so that ( X)n = Xn-1 and d X = -dX . Similarly, X can be taken to be the complex -1 X. That is, ( X)n = Xn+1 and d X = -dX . In particular, and are inverse functors. The second author [Hov98 ] has shown that this implies that cofibre sequences and fibre sequences agree (up to the usual sign) and that and the cofibre sequences give rise to a triangulation of the homotopy category. (See [Qui67 , Section I.3] for the definition of cofibre and fibre sequences in any pointed model category.) Using the explicit cylinder object from the proof of Lemma 2.13, we can be more explicit. Corollary 2.17. The category D(A) is triangulated with the usual suspension. A se- quence L -! M -! N -! L is a triangle if and only if it is isomorphic in D(A) to the usual cofibre sequence on the map L -! M (see Section 0). Now we show that the model structures we construct are proper. A good reference for proper model categories is [Hir00 , Chapter 11]. Proposition 2.18. Let P be any projective class on an abelian category A. Consider the commutative square in Ch (A) below. A -- -f- ! X ? ? q ?y ?yp B -- -g- ! Y (a) If the square is a pullback square, p is a P-fibration, and g is a P-equivalence, then f is a P-equivalence. That is, the relative model structure is right proper. (b) If the square is a pushout square, q is a degreewise split monomorphism, and f is a P-equivalence, then g is a P-equivalence. In particular, the pushout of a P-equivalence along a P-cofibration is a P-equivalence. That is, the relative model structure is left proper. QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 13 A model category that is both left and right proper is said to be proper. Note that we don't actually need to know that our P-cofibrations, P-fibrations and P-equivalences give a model structure to ask whether the structure is proper. Proof. Part (a) is an immediate consequence of [Hir00 , Corollary 11.1.3], since every object is P-fibrant. For part (b), let C be the cokernel of q. Since pushouts are computed degreewise, it follows that p is a degreewise split monomorphism with cokernel C. Thus we have a map of triangles A -- -f- ! X ? ? q ?y ?yp B -- -g- ! Y ?? ? y ?y C -- -i-d! C in the homotopy category K(A). The top and bottom maps are P-equivalences, so the middle map must be as well, by using the five-lemma and the long exact sequences obtained by applying the functors [ k P, -]. We now consider monoidal structure. Monoidal model categories are studied in [Hov98 , Chapter 4]. We will assume that A is a closed monoidal category. Thus it is equipped with a functor : A x A -! A such that both A - and - A have right adjoints for each A in A. In particular, what we need is that these functors preserve colimits. There is, of course, an induced closed monoidal structure on Ch (A), for which we also use the notation . Proposition 2.19. Let A be a closed monoidal abelian category with a projective class P such that cofibrant replacements exist and the unit is P-projective. Then the relative model structure on Ch (A) is monoidal if and only if the tensor product of two P-cofibrant complexes is always P-cofibrant. Proof. There are two conditions that must hold for a model category to be monoidal. One of them is automatically satisfied when the unit is cofibrant. The unit of the monoidal structure on Ch (A) is 0 S, where S is the unit of A, which has been assumed to be P-projective. Thus the unit is P-cofibrant, by Lemma 2.7 (a), Therefore, the relative model structure is monoidal if and only if whenever f : A -! B and g : X -! Y are P-cofibrations, then the map f g : (A Y ) qA X (B X) -! B Y is a P-cofibration, and is a P-trivial cofibration if either f or g is a P-trivial cofibration. It is easy to see that f g is a degreewise split monomorphism, with cokernel C Z. By taking f to be the map 0 -! C and g to be the map 0 -! Z, we see that if the relative model structure is monoidal, then the tensor product of two P-cofibrant complexes is P-cofibrant. Conversely, if C Z is P-cofibrant whenever C and Z are P-cofibrant, the preceding paragraph implies f g is a P-cofibration whenever f and g are P-cofibrations. If either f or g is a P-trivial cofibration, then one of C or Z is P-trivially cofibrant, and hence contractible. It follows that C Z is contractible, and so f g is a P-trivial cofibration. In particular, suppose the relative model structure is monoidal, and M, N are P- projective. Then 0 M 0 N ~= 0 (M N ) is P-cofibrant, and therefore M N is P- projective. With this in mind, we say that a projective class P on a closed monoidal cat- egory A is monoidal if the unit is P-projective and the tensor product of P-projectives is P-projective. 14 J. DANIEL CHRISTENSEN AND MARK HOVEY We do not know of any example of a monoidal projective class P where cofibrant replacements exist, but the relative model structure is not monoidal. Certainly this does not happen in either Case A or Case B of Theorem 2.2. In Case A, we have the following corollary. Corollary 2.20. Suppose F : B -! A is a monoidal functor between closed monoidal abelian categories, with right adjoint U that preserves countable coproducts. Then the relative model structure on Ch (A) is monoidal. Proof. Proposition 3.3 says that QX QY is P-cofibrant for any X and Y , where Q denotes the cofibrant replacement functor constructed in Section 3. If X and Y are already cofibrant, then lifting implies that X is a retract of QX and Y is a retract of QY . Thus X Y is a retract of QX QY , so X Y is P-cofibrant. Proposition 2.19 completes the proof. In Case B, we will show in Corollary 4.4 that every P-cofibrant object is P-cellular. Thus the following corollary applies. Corollary 2.21. Let A be a closed monoidal abelian category with a monoidal projective class P such that cofibrant replacements exist and every P-cofibrant object is P-cellular (Definition 2.6). Then the relative model structure on Ch (A) is monoidal. Proof. Let A and B be P-cofibrant. By Proposition 2.19, it suffices to show that A B is P-cofibrant. By assumption, A is a retract of a transfinite colimit of a colimit-preserving diagram 0 = A0 -! A1 -! . . . such that for each ff, the map Aff -! Aff+1 is degreewise split monic with cokernel a complex of relative projectives with zero differential. Since - B preserves retracts, degreewise split monomorphisms and colimits, it is enough to prove that iP B is P-cofibrant for each P-projective P and each i. Applying the same filtration argument to B, we find that it suffices to show that iP jQ = i+j (P Q) is P-cofibrant for all P-projectives P and Q and integers i and j. But this follows immediately from the fact that P is monoidal and Lemma 2.7 (a). We can also prove a dual statement in Case A. Proposition 2.22. Suppose U : A -! B is a monoidal functor of closed monoidal abelian categories, with right adjoint F . Assume that U preserves countable products. Then the injective relative model structure on Ch (A) is monoidal. By the "injective relative model structure", we mean the model structure obtained by dualizing Theorem 2.2. The (trivial) cofibrations in this model structure are called the B-injective (trivial) cofibrations. Proof. Suppose f : A -! B and g : X -! Y are B-injective cofibrations with cokernels C and Z, respectively. This means that U f and U g are degreewise split monomorphisms. Recall the definition of f g used in the proof of Proposition 2.19. Since U is monoidal and preserves pushouts, U (f g) ~= U f U g. One can easily check that U f U g is a degreewise split monomorphism, so f g is a B-injective cofibration. If f is a B-injective trivial cofibration, then the cokernel C of f has U C contractible. Let Z denote the cokernel of g. Since the cokernel of U f U g is U C U Z, which is contractible, f g is a B-injective trivial cofibration. Since everything is cofibrant in the injective relative model structure, the other condition in the definition of a monoidal model category is automatically satisfied. Another important property of a model category is whether it is cofibrantly generated. This is the topic of Section 5. QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 15 2.2. Why projective classes? The astute reader will notice that we haven't used the assumption that there is always a P-epimorphism from a P-projective. Indeed, all we have used is that we have a collection P of objects such that, with the definitions at the beginning of this section, cofibrant replacements exist. The following proposition explains why we start with a projective class. Proposition 2.23. Let P be a collection of objects in an abelian category A such that cofibrant replacements exist in Ch (A). Then there is a unique projective class (P0, E ) that gives rise to the same definitions of weak equivalence, fibration and cofibration. Proof. Let E be the collection of P-epimorphisms and let P0 be the collection of all objects P such that every map in E is P -epic. Then P0 contains P, and a map is P0- epic if and only if it is P-epic. It follows that the P0-exact sequences are the same as the P-exact sequences. A map f is a P-equivalence if and only if the cofibre of f is P-exact. The same holds for the P0-equivalences, and since the two notions of exactness also agree, the two notions of equivalence agree. Finally, since cofibrations are defined in terms of the fibrations and weak equivalences, the two notions of cofibration agree. That the pair (P0, E ) is a projective class follows immediately from the existence of cofibrant replacements. Our work earlier in this section shows that the objects of P0 are precisely those objects P that are cofibrant when viewed as complexes concentrated in degree 0. Thus the projective class is the unique projective class giving rise to the same weak equivalences, fibrations and cofibrations. Thus by requiring our collection of objects to be part of a projective class, we in effect choose a canonical collection of objects determining each model structure we produce. This means that the relevant question is: which projective classes give rise to model structures? Theorem 2.2 gives a necessary and sufficient condition, namely the existence of cofibrant replacements, and we know of no projective classes that do not satisfy this condition. An additional advantage of having a projective class is that it provides us with the lan- guage to state results such as: Ext nP(M, N ) ~= Ho Ch (A)( n M, N ) (see Corollary 2.14). 3. Case A: Projective classes coming from adjoint pairs In this section we prove Case A of Theorem 2.2. Let U : A -! B be a functor of abelian categories, with a left adjoint F : B -! A. Then U and F are additive, U is left exact and F is right exact. Let P be the projective class on A that is the pullback of the trivial projective class on B (see Example 1.3 and Subsection 1.3). In this subsection we construct a cofibrant replacement functor for the projective class P under the assumption that U preserves countable coproducts. Because this projective class is strong, Lemma 1.9 tells us that the P-fibrations and P-equivalences have alternate characterizations. A map f in Ch (A) is a P-fibration if and only if A(P, f ) is degreewise split epic for each P in P, and is a P-equivalence if and only if A(P, f ) is a chain homotopy equivalence for each P in P. First we prove a lemma characterizing the fibrations and weak equivalences in this model structure and giving us a way to generate cofibrant objects. Lemma 3.1. (a) A map p : X -! Y is a P-fibration if and only if U p is a degreewise split epimorphism in Ch (B). (b) A map p : X -! Y is a P-equivalence if and only if U p is a chain homotopy equivalence in Ch (B). (c) If i is a map in Ch (B) that is degreewise split monic, then F i is a P-cofibration. (d) For any C in Ch (B), F C is P-cofibrant. 16 J. DANIEL CHRISTENSEN AND MARK HOVEY Proof. For part (a), note that the P-projectives are all retracts of F M for M 2 B. Hence p is a P-fibration if and only if A(F M, p) = B(M, U p) is a surjection for all M 2 B. This is true if and only if U p is a degreewise split epimorphism. For part (b), a similar argument shows that p is a P-equivalence if and only if B(M, U p) is a quasi- isomorphism for all M 2 B. We claim that this forces U p to be a chain homotopy equivalence. Indeed, let C denote the cofiber of U p. Then B(M, C) is exact for all M 2 B. By taking M = Zn C, we find that C is exact and that Cn+1 -! Zn C is a split epimorphism. It follows that C is contractible, as in the proof of Lemma 1.9. Thus U p is a chain homotopy equivalence. For part (c), let C be the cokernel of i. Then F i is degreewise split monic with cokernel F C. Suppose that p : X -! Y is a P-trivial fibration with kernel K. Then K is P-trivially fibrant, so part (b) implies that U K is contractible. By Lemma 2.3, to show that F i is a P-cofibration, it suffices to show that every chain map F C -! K is chain homotopic to 0. By adjointness, it suffices to show that every chain map C -! U K is chain homotopic to 0. Since U K is contractible, this is clear. Part (d) follows from part (c), We now provide a construction, given a complex X, of a P-cofibrant complex QX and a P-trivial fibration QX -! X. This comes out of the bar construction, which we now recall from [ML95 , Section IX.6]. Given an object N of A, define the complex BN by (BN )m = (F U )m+1 N when m -1 and 0 otherwise. We will define maps s = sm : U (BN )m-1 -! U (BN )m and ffi = ffim : (BN )m -! (BN )m-1 . For m < 0 we declare both to be zero. For m 0, sm : U (BN )m-1 = (U F )m U N -! U (BN )m = U F (U F )m U N is defined to be adjoint to the identity map of F (U F )m U N . For m 0, we can then inductively define ffim : (BN )m = F U (F U )m N -! (BN )m-1 = (F U )m N to be adjoint to the self-map 1 - sm-1 (U ffim-1 ) of U (F U )m N . Properties of adjoint functors then guarantee that (U ffim )sm + sm-1 (U ffim-1 ) = 1. Using this and the prop- erties of adjoint functors, one deduces a sequence of implications ffim-1 ffim = 0 =) (U ffim )sm (U ffim ) = U ffim =) (U ffim )(U ffim+1 )sm+1 = 0 =) ffim ffim+1 = 0. Therefore ffim-1 ffim = 0 for each m and so ffi makes BN into a chain complex. The construction of BN , ffi, and s is obviously natural in N . Thus, given a complex X in Ch (A), we get a bicomplex (BX)m,n = (BXn )m , where ffi : (BX)m,n -! (BX)m-1,n and d : (BX)m,n -! (BX)m,n-1 commute. Furthermore, s : U (BX)m,n -!_ U (BX)m+1,n_L and U d also commute. We can then_define a total complex Q X by (Q X)k = m+n=k (BX)m,n . The differen- tial @ of Q_X_takes the summand (BX)m,n into (BX)m-1,n_ (BX)m,n-1 by (ffi, (-1)m d). Note that Q X is a filtered complex,_ with F iQ X the subcomplex_consisting_of terms with m i. In particular, F -1 Q X = -1 X. We define QX to be Q X= -1 X, so that (QX)k = (F U )Xk (F U )2 Xk-1 . . .. The following proposition proves Case A of Theorem 2.2. Proposition 3.2. There is a natural P-trivial fibration qX : QX -! X, and QX is P-cofibrant. Proof. We first show that QX is P-cofibrant. There is an increasing filtration {F iQX}i 0 on QX, where F 0QX = F U X and F iQX=F i-1QX = (F U )i+1 iX. Fur- thermore, each inclusion F iQX -! F i+1QX is a degreewise split monomorphism. By Lemma 3.1 (d), each quotient (F U )i+1 iX is P-cofibrant, so each map F i-1QX -! F iQX is a P-cofibration. Hence F U X -! QX is a P-cofibration. Thus QX is P- cofibrant. The map qX is induced by ffi0 : (F U )Xn -! Xn , adjoint to the identity. The map qX sends the other summands of (QX)n to 0. We leave to the reader the check that this is a chain map. Since U ffi0 is a split epimorphism, U qX is a degreewise split epimorphism, QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 17 so is a P-fibration_ by Lemma 3.1 (a). To show qX is a P-equivalence,_ it suffices to show that the fiber Q X is P-contractible, or, equivalently, that U Q X is contractible (Lemma 3.1 (b)). The contracting homotopy is given by s. Indeed, on the summand U (BX)m,n , the U (BX)m,n component of s(U @) + (U @)s is s(U ffi) + (U ffi)s = 1, and the U (BX)m+1,n-1 component is s(-1)m (U d) + (-1)m+1 (U d)s = 0. (This is where we use that U commutes with coproducts.) The following proposition implies that the relative model structure is monoidal in this case, as explained in Corollary 2.20. Proposition 3.3. Suppose F : B -! A is a monoidal functor of closed monoidal abelian categories, with right adjoint U that preserves countable coproducts. Then, for any X and Y in Ch (A), QX QY is P-cofibrant. Proof. Recall the filtration F iQX on QX used in the proof of Proposition 3.2. Using this filtration, we find that QX QY is the colimit of F iQX QY , and each map F i-1QX QY -! F iQX QY is a degreewise split monomorphism with cokernel (F U )i+1 iX QY . It therefore suffcies to show that this cokernel is P-cofibrant for all i. A similar argument using the filtration on QY shows that it suffices to show that (F U )i+1 ( iX) (F U )j+1 ( jY ) ~= F (U (F U )i iX U (F U )j jY ) is P-cofibrant for all i, j 0. But this follow immediately from Lemma 3.1 (d). 3.1. Examples. Example 3.4. Let B be a bicomplete abelian category. Since the identity functor is ad- joint to itself and preserves coproducts, we can apply Theorem 2.2 to the trivial projective class P to conclude that Ch (B) is a model category. We call this the absolute model structure. The P-equivalences are the chain homotopy equivalences (Lemma 3.1), the P-fibrations are the degreewise split epimorphisms and the P-cofibrations are the de- greewise split monomorphisms. Every object is both P-cofibrant and P-fibrant, and the homotopy category is the usual homotopy category K(B) in which chain homotopic maps have been identified. That this model structure exists was also shown in [Col ]. Note that if B is closed monoidal, the absolute model structure is also monoidal, by Corol- lary 2.20. In particular, since every object is P-cofibrant, given a differential graded algebra R 2 Ch (B), we get a model structure on the category of differential graded R- modules, where weak equivalences are chain homotopy equivalences (of the underlying chain complexes) and fibrations are degreewise split epimorphisms. Now let A and B be bicomplete abelian categories and let U : A -! B be a coproduct preserving functor with left adjoint F . By Theorem 2.2, Ch (A) has a relative model structure whose weak equivalences, cofibrations and fibrations are called B-equivalences, B-cofibrations and B-fibrations. This structure is a lifting of the absolute model structure on Ch (B) in the sense that a map f in Ch (A) is a weak equivalence or fibration if and only if U f is so in Ch (B). It is often the case that one wants to lift a model structure along a right adjoint. When the model structure is cofibrantly generated, necessary and sufficient conditions for a lifting are known [DHK97 , 9.1], [Hir00 , 13.4.2]. Our main theorem says that it is also possible to lift the absolute model structure on Ch (B), even though it is not usually cofibrantly generated (see Subsection 5.4). It follows from the above that F preserves cofibrations and trivial cofibrations. The adjoint functors F and U form a Quillen pair. The category Ch (A) also has an absolute model structure. The identity functor sends absolute fibrations and weak equivalences to B-fibrations and B-equivalences. Thus the identity functor is a right Quillen functor from the absolute model structure to the relative model structure. 18 J. DANIEL CHRISTENSEN AND MARK HOVEY Example 3.5. Let R -! S be a map of rings. Write R-mod and S-mod for the categories of left R- and S-modules. Consider the forgetful functor U : S-mod -! R-mod and its left adjoint F that sends an R-module M to S R M . We saw in Example 1.6 that this gives a projective class whose relative projectives are the S-modules P such that the natural map S R M -! M is split epic as a map of S-modules. The functor U preserves coproducts, so Theorem 2.2 and Lemma 1.9 tell us that Ch (S) = Ch (S-mod ) has a model structure in which the weak equivalences are the maps that become chain homotopy equivalences after forgetting the S-module structure. The fibrations are the maps that in each degree are split epic as maps of R-modules. The cofibrations are defined by the left lifting property with respect to the trivial fibrations. Equivalently, by Proposition 2.5, they are the degreewise split monomorphisms whose cokernels are cofibrant. And by Lemma 2.4, a complex C is cofibrant if and only if each Cn is a relative projective and every map from C to a complex K such that U K is contractible is chain homotopic to 0. Lemma 2.7 gives us a ready supply of cofibrant objects. In particular, a relative resolution of an S-module M is a cofibrant replacement, so the group Ho Ch (S)(M, iN ) of maps in the homotopy category is isomorphic to the relative Ext i group [ML95 ]. When R and S are commutative, the functor F is monoidal, so this relative model structure is monoidal by Corollary 2.20. There is then a derived tensor product X L Y in Ho Ch (S), and if M and N are S-modules, Hi(M L N ) is isomorphic to the relative Tor i group [ML95 ]. Example 3.6. Again let R -! S be a map of rings and let U be the forgetful functor. We saw in Example 1.7 that U has a right adjoint G that sends an R-module M to R-mod (S, M ) and so we get an injective class on S-mod by pulling back the trivial injective class along U . U preserves products, so we can apply the duals of Theorem 2.2 and Lemma 1.9 to conclude that Ch (S) has a model structure with the same weak equivalences as in the previous example. The cofibrations are the maps that in each degree are split monic as maps of R-modules, and the fibrations are the maps with the right lifting property with respect to with respect to the trivial cofibrations. Fibrations and fibrant objects can also be characterized by the duals of Lemmas 2.4 and 2.7, and the homotopy category again encodes the relative Ext groups. However, this relative model structure is not monoidal, even when R and S are commutative, since U is not monoidal. Example 3.7. Suppose A is an algebra over a commutative ring k, and let C denote the category of chain complexes of A-bimodules. Then there are forgetful functors from A- bimodules to left A-modules, right A-modules, and k-modules. Each of these preserves coproducts, so we get three different relative model structures on C. When we forget to right or left A-modules, the cofibrant replacement functor Q applied to A gives us the usual un-normalized bar construction, with (QA)n = A n+2 . The complex QA is then also cofibrant in the relative model structure obtained by forgetting to k-modules, since it is a bounded below complex of relative projectives. The Hochschild cohomology HHn (A; N ) of A with coefficients in a bimodule N (or a chain complex of bimodules N ) is then equal to Ho C( n A, N ), where we can use any of the three relative model structures above. We can extend this definition by replacing A with an arbitrary complex of bimodules, but then the answer may depend on which relative model structure we use. It is most natural to use the relative model structure obtained from the forgetful functor to k-modules, as then we can define Hochschild homology as well. Indeed, the Hochschild homology groups are defined using the tensor product M N of bimodules, where we identify ma n with m an but also am n with m na. We emphasize that M N is only a k-module, not a bimodule. In particular, this tensor product can't be associative, but it is commutative and also unital in a weak sense, since M (A A) ~= M as a k-module. Furthermore, the functor - N has a right QUILLEN MODEL STRUCTURES FOR RELATIVE HOMOLOGICAL ALGEBRA 19 adjoint that takes a k-module L to the bimodule Hom k(N, L), where, if g 2 Hom (N, L), (ga)(n) = g(an) and (ag)(n) = g(na). Both of these functors then extend to functors on chain complexes. The proof of Proposition 2.19 applies to this case as well, and shows that if f and g are cofibrations in the relative model structure on C obtained by forgetting to k-modules, then f g is a cofibration in the absolute model structure on chain complexes of k-modules. If either f or g is a trivial cofibration, so is f g. This means the tensor product has a total left derived functor that is commutative and unital (in the above weak sense). We can therefore extend the usual definition of Hochschild homology to complexes X and Y of bimodules, by defining HHn (X, Y ) = Hn (QX QY ). In fact, it is possible to prove, using the technique of Proposition 2.19, that, if X is cofibrant, the functor X - takes weak equivalences in the relative model structure to chain homotopy equivalences. This implies that HHn (X, Y ) = Hn (QX Y ). This is a direct generalization of the way Hochschild homology is defined in [ML95 , Section X.4]. 4. Case B: Projective classes with enough small projectives In this section we prove that cofibrant replacements exist in Case B. We begin by introducing the terminology necessary for the precise statement of Case B. We think of an ordinal as the set of all previous ordinals, and of a cardinal as the first ordinal with that cardinality. Definition 4.1. Given a limit ordinal fl, the cofinality of fl, cofin fl, is the smallest cardinal ~ such that there exists subset T of fl of cardinality ~ with sup T = fl. The cofinality of a successor ordinal is defined to be 1. A colimit-preserving sequence from an ordinal fl to a category A is a diagram X0 -! X1 -! . . .-! Xff -! Xff+1 -! . . . of objects of A indexed by the ordinals less than fl, such that for each limit ordinal ~ less than fl the natural map colim ff<~ Xff -! X~ is an isomorphism. For ~ a cardinal, an object P is said to be ~-small relative to a subcategory M if for each ordinal fl with cofin fl > ~ and each colimit-preserving sequence X : fl -! A that factors through M, the natural map colim ff