ebook img

Categorical Algebra and its Applications: Proceedings of a Conference, held in Louvain-La-Neuve, Belgium, July 26 – August 1, 1987 PDF

364 Pages·1988·13.298 MB·English
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Categorical Algebra and its Applications: Proceedings of a Conference, held in Louvain-La-Neuve, Belgium, July 26 – August 1, 1987

ARE ALL LIMIT-CLOSED SUBCATEGORIES OF LOCALLY PRESENTABLE CATEGORIES REFLECTIVE? J. Ad~mek, J. Rosick9 and V. Trnkov~ Teeh. Univ. Purkyne Univ. Math. Inst. of the Suehb~tarova 2 Jan~ckovo n~m. 2a Charles University Praha 16627 ,6 66295 Brno, CSSR Sokolovsk~ 83 CSSR Praha 18600 ,8 CSSR Introduction. It is not surprising that the answer to the question in the title is negative, under "reasonable" set-theoretical hypothesis (e.g., assuming the non-existence of measurable cardinals). We even construct two reflective subcategories ~) of the locally present- able category Graph of graphs whose intersection is not reflective in Graph (although it is, of course, closed under limits). What might be surprising is that under other set-theoretical hypothesis, the answer is affirmative. We introduce a condition called Weak Vop~nka Principle, consistency of which follows by the existence of huge cardinals and, as some set-theorists believe, may be added to the usual axioms of set theory. We prove that assuming Weak Vop~nka Principle, then each locally presentable category K has the following properties: (i) Every subcategory of K closed under limits is reflective in K; (ii) All reflective subcategories of K form a large complete lattice (ordered by inclusion); (iii) The intersection of two reflective subcategories of K is reflective. Conversely, assuming the negation of Weak Vop~nka Principle, none of the above statements holds in K = Graph. Weak Vop~nka Principle is the following statement: Ord °p does not have a full embedding into Graph. (Here Ord °p is the dual of the well-ordered category of all ordinals.) We have chosen that name because the principle is weaker than the well- known Vop~nka Principle which, as we show below, can be formulated as follows: ~)Subcategories are understood to be full throughout our paper. drO does not have full embedding into Graph. The position of Vop~nka Principle in set theory is discussed e.g. in [j]; there are good reasons to believe that Vop~nka Principle does not contradict the usual axioms of set theory. If so, tb~n we can add Weak Vop~nka Principle to the usual axioms, and in the resulting set theory the answer to our title question is affirmative. On the other hand, by results to be found in [PT] , every concrete category has a full embedding into Graph provided that we assume the following (M) There does not exist a proper class of measurable cardinals. Thus, (M) implies the negation of both of the above principles. Since the non-existence of measurable cardinals is certainly not contradict- ory to the usual axioms of set theory, we conclude that we can add the negation of Weak Vop~nka Principle to those axioms.In the resulting set theory, the answer to the above question is negative. A closely related result has been proved in [AR] where, under (M), a proper class of reflective subcategories of Graph was presented whose intersection is not reflective in Graph. The present result is a refinement of the previous one: the set-theoretical hypothesis is weaker (viz., as weak as possible) and we have two subcategories in place of a proper class. The price of that refinement is a much deeper delving into the results of Prague School. This has made our main construction technically quite involved. For a reader interested only in the equivalence of our title question and Weak Vop~nka Principle, it is not necessary to study our construction (i.e., he can skip parts II and III of our paper) since J. R. Isbell has made the observation in[Ill that the affirmative answer easily implies Weak Vop~nka Principle and our Theorem 2 states the converse. This converse also follows by a result of E. R. Fisher announced in [FI]. He unfortunately has not published the proof yet but he kindly sent it to us IF2] . We finally mention a recent result of M. Makkai and A. Pitts [MP] which holds without set-theoretical restrictions: let K be a locally finitely presentable category, then each subcategory of K closed under limits and filtered colimits is reflective in K. In contrast to locally presentable categorie~ in the category poT of topological spaces we have found, without additional set-theoretical hypothesis, two reflective subcategories with non-reflective intersect- ion, see [TAR]. I. Assumin~ Weak Vop~nka Principle Convention. We work, throughout our paper, within the usual G~del-Bernays theory of sets with AC. We will heavily use the category Graph of graphs (directed), i.e., pairs (V,E) where V is a set (of vertices) and E C V x V i~ a set (of edges). Morphisms f:(V~E) ~ (V',E'), called homomorphisms, are maps f:V - V' with (fxf)(E) ~ E'. An important property of that category is that every graph is a coproduct of its connected components (=maximal indecomposable subobjects). The following result explains why Graph can be used as a representative locally presentable category: Theorem .i Every locally presentable category can be fully embedd- ed into Graph. Proof. For each small category ,A the category S~t A has a full embedding into Graph, see [PT] (II.5.3 and Ex. 1.7.1). Consequently, every locally presentable category is equivalent to a subcategory of Graph, see [GU]. It is obvious that a category equivalent to a sub- category of Graph is also isomorphic to (another) subcategory of Graph. A large discrete category is defined by having a proper class of objects and no other morphisms than identities. Lemma .i (Formulations of Vop~nka Principle.) The following statements are equivalent: (i) Vop~nka Principle: for each first-order language, every class of models such that none of them has an elementary embedding into another one is a set} (ii) No locally presentable category has a large discrete subcategory) (iii) Graph does not have a large discrete subcategory; (iv) Ord cannot be fully embedded into Graph. Proof. i ~ iv: Suppose that, on the contrary, Ord has a full embedding into Graph (denoted by (Vi,Ei) on objects and by aij:(Vi,E i) -- (Vj,Ej) on morphisms, i ~ j). There clearly exists a proper class K --COrd such that for each j E K the maps ~ij for i 6 K, i < j, are not collectively onto~ let us choose a vertex xj E Vj - - ieK,i<jU (v~ ij i). Now consider the language of one binary relation and one nullary operation (and no axioms). Then each (vi,Ei,xi),i 6 K, is a model, and there are no elementary extensions in the resulting class. This contradicts (i). iv ~ iii: The category K of connected graphs has the property that Graph can be fully embedded into K, see [PT] (I.l.ll). Thus, if (iii) would be false, there would exist a large discrete category of connected graphs G , x E X. For any ordinal i, X. = {x 6 X \ the rank x l of x is smaller than i} is a set (see [J] .) (Using the rank, which is based on the axiom of regularity, we avoid the axiom of choice for classes.) There is an isotone map F:Ord ~ Ord such that XF(i) # XF(j) for i ~ j. The functor A:Ord ~ Graph defined on objects by i A = ~ G and on morphisms by the obvious coproduct injections x XeXF(i) is full. (In fact, given a homomorphism f:A ~ A then, since f i 3 preserves connected components, for each x 6 XF(i) there is y 6 XF(j) such that f maps Gx inside Gy. It follows that x = y and f(v) = v for each vertex v of G . Thus, i ~ j and F(v) = v for each vertex v of A..) x 1 This contradicts (iv). iii ~ ii by Theorem i. ii - i : Suppose that, on the contrary, there is a first-order language L and a proper class Mi, i @ X of L-models such that none of them has an elementary embedding into another. Add a new binary relation symbol to L and denote by ~ the resulting first order language. By [PT] II.3.12., the underlying set IMil of i M always carries a rigid binary relation r.1. Hence (Mi,r i), i E X form a large discrete subcategory of the category Mod(~) of ~-models and elementary embeddings. By the downward Lowenheim-Skolem theorem (cf. [CK] , 3.1.6.), Mod(~) has a small dense subcategory. Hence Mod(~) can be fully embedded into a category of algebras (cf. [I2] , 4.2.), i.e. into a locally presentable category. It contradicts (ii). L~mna 2. Vop~nka Principle implies Weak Vop~nka Principle (in other words, if Ord °p has a full embedding into Graph then also Ord has such an embedding). Proof. Let A:Ord °p ~ Graph be a full embedding denoted by A = l = (Vi,E i) on objects and eij:Ai - Aj (j N i) on morphisms. There clearly exists a proper class K COrd such that ~. is not one-to-one 13 for any i,j E K, i # j. Consider - the theory of two binary relations (and no axiom). For each i e K we have a model (Vi,Ei,Ei) where E[ = 1 = {(x,y) I x,y 6 Vi, x # y} and there exist no homomorphisms from one of those models into another one - thus, Vop~nka Principle does not hold. Open problem: Are the two principles equivalent? Theorem 2. Assuming that Weak Vop~nka Principle is true, each subcategory of a locally presentable category K which is closed under limits is reflective in K. Proof. We first observe that for each set H of objects of K, the least subcategory H* of K closed under limits and containing H is reflective in K. This follows from the Special Adjoint Functor Theorem [M]: H is a cogenerating set, and since K is wellpowered and complete (being locally presentable), so is H*. Let [ be a subcategory of K closed under limits. For j @ Ord, let [j be the set of all objects of i having the rank smaller than j (compare the proof of Lemma I; we are again avoiding the axiom of choice for classes). The categories i~ are reflective. Thus,for each 3 object K of K we have reflections r.:K ~ K. of K in [~ for all j 6 ... 3 J J • Ord. Since i ~ j implies L~ _ C 3 _ i "" C i we have the unique K-morphism eji: Kj - i K with i = r eji.r j. We are going to prove that the chain e.. is stationary, i.e., that there exists an ordinal o i such that 31 all eji with io ~ i ~ j are isomorphisms. It then follows that rio:K ~ KioiS the reflection of K in each L~, i ~ io, and hence, the reflection in L; this will conclude the proof. Suppose that, on the contrary, there is an object K such that eji is not stationary. From thewellpoweredness of K it follows that there is a transfinite sequence to < I < t .-. i < < t ... of ordinals such that for i < j the morphism e is not a monomorphism~ tj i t Now consider the comma-category KSK (of all K-arrows with domain K, see [M] :) we have a functor E:Ord °p ~ KiK rt. defined on objects by i E = (K i Kt ) and on morphisms by (i ~ j)~ et3 ti" This functor is a full embedding. In fact, let f:Ej ~ i E be a morphism, i.e., let f:Kt. ~ Kt ' fulfil f.rt. = rt . If i S j, 3 l 3 l then f.rt t = e .r t and this implies f t = e (by the uniqueness 3 J ti J jti requirement on reflections). And the case i > j cannot occur since we would have (f'etitj)'rtl f.rtj = rti, and hence, f.etitj = id, z although e is not a monomorphism. It is well-known that since tit j K is locally presentable, so is KSK (see, e.g., [MPa]6.1.1.), and hence, there is a full embedding of KiK in Graph by Theorem i. This embedding composed with E above, shows that Weak Vop~nka Principle is false - a contradiction. (There is,a more straightforward way to finish the proof because E induces a full embedding of KSK in G~Graph for G = EK and GiGraph has a full embedding into a locally presentable category of graphs with constants c indexed by vertices of G (sending f:G x X to (X,f(V))veV(G))). Corollary. Assuming Weak Vop~nka Principle, each intersection of reflective subcategories of a locally presentable category K is reflective in K. Thus, the lattice of all reflective subcategories of K is large-complete. II. Assuming Negation of Weak Vop~nk a Principle Theorem .3 Assuming that Weak Vop~nka Principle is false, the category Graph has two reflective subcategories whose intersection is not reflective. The aim of the following two sections is to prove the theorem. We shall construct acollection D (i E Ord) of graphs with the following 1 properties: (a) The subcategory i of Grap~ consisting of those graphs G such that for each ordinal i there is precisely one homomorphism from D to G is not reflective in Graph; 1 (b) Let i and L denote the subcategories of Gaaph consisting of e o those graphs G satisfying the following conditions: (bl) Given an even ordinal i (odd ordinal i, respectively) such that for each t < i there is a homomorphism from t D to G, then there is also a homomorphism from D to G, i (b2) For each ordinal i there is at most one homomorphism fror~ D to G, 1 (b3) There is a homomorphism from O D to G. Then e L and ° i are both reflective subcategories of Graph. Since, obviously, L e = N L Lo, this will prove Theorem .3 We will now present a condition guaranteeing that (b) is true. In the last section we will construct, assuming the negation of Weak Vop~nka Principle, a collection i D satisfying both that condition and (a) above. Recall that a collection of objects is rigid if the corresponding subcategory is discrete. Proposition i. Let i D (i 60rd) be a rigid collection of connected graphs such that for each graph G the following holds: if there is a homomorphism from Di+ 1 to G and a surjective homomorphism from k<~l Dk to G, then there is also a homomorphism from D. to G. Then the above subcategories L and L are reflective in Graph. e o Proof. We present a proof that Lo is reflective, the case of e i is analogous. Let i denote the category of all graphs satisfying (b2) above. Observe that % is obviously closed under products and subobjects in Graph, and therefore, ~ is epireflective in Graph. It is thus sufficient to prove that o L is reflective in .L We denote, for short, by [p,Q] the set of all homomorphisms from P to Q. We first show that for each graph G 6 ~ with [Di,G] # ~ for all even ordinals i there exists an ordinal io with [Di,G] # ~ for all i ~ io. Without loss of generality, we can assume that each vertex of G lies in the image of some homomorphism from i, D i E Ord. (For else we work with the subgraph G' of G consisting of all vertices lying in those images: proving the statement for G' would clearly prove it for G too.) We can, then, choose an ordinal io such that each vertex of G lies in the image of some homomorphism from Di, i < io. Let us show that for each ordinal j ~ io we have [Dj,G] # 9. Suppose, on the contrary, that [Dj,G] = 9. It follows that j is odd, and hence j + 1 is even and therefore, [Dj+I,G] # 9- Put A = {i e Ordli < j, and [Di,S] = 9}. There is a surjective homomorphism from ~ k D to G + ~Di: for each k <j i eA k < j we either have a unique homomorphism from k D to G (if k g A) or k D is a summand of ~ i D (if k 6 A) and the resulting homomorphism 3<]~k:f Dk - G + ~ i D is surjective by the choice of io, since j ~ io. i We now use the above property of the graphs i D to conclude that there is a homomorphism from Dj to G + ~ i. D But this is a contradiction ieA since j ~ A and [Dj,G] = 9: each homomorphism g:Dj ~ G + i~ Di would, by the connectedness of Dj, map Dj either to G, or to some i D (but then i = j). We are ready to describe the reflection of any graph G 6 .i We denote by io either the least even ordinal with [Dio,G] = @, or (if no such even ordinal exists) the least ordinal with [Di,G] # ~ for all i ~ io. We will prove that tne refl~ction of G in io is the following coproduct injection r : G ~ G + je~B Dj where B j{ E Ordlj < io and [Dj,G] @}. = = (A) G + ~ Dj lies in L . Firsts this graph obviously lies in .L jeB o Now let i be an odd ordinal with [Dt, G + ~_[ Dj] # @ for all t < i. We are to show that [D i, G + ~ Dj] # ~. This is obvious whenever either [Di,G] ~ @, or i E B. Thus, we can suppose that i ~ io. (AI) If io is even with [Dio,G] = ~ then, since i is odd, we have i > io and then io is one of the t's above - thus, [Dio, G + + ~ Dj] ~ ~. This is a contradiction since [Di0,G] = ~ and io ~ B. J (A2) If io is such that [Dk,G] # ~ for all k Z io, we are ready since i ~ io. (B) Any ~omomorphism h:G ~ H with H O E i factors uniquely through r. Since H E ,L it is sufficient to show that [Dj,H] # ~ for all j E B: then we have a (necessarily unique~ extension of h to G + J~ D.. j~B 3 We will prove by induction on i < io that [Di,H] # @. The case i = O follows from the definition of Lo" Suppose i < io is such that [Dt,H]# # ~ for all t < i. Since H ~ io, if i is odd we conclude [Di,H] # @. If i is even then i < io implies [Di,G] # ~, and hence, [Di,H] # ~. III. Main Construction The construction of the graphs i D which would prove Theorem 3 has several levels, and we start with some auxiliary constructions. We denote by Graph o the subcategory of Graph formed by all non-trivial, acyclic, strongly connected graphs, i.e., graphs of at least two vertices and such that for arbitrary two distinct vertices x and y there exists either a directed path from x to y, or a directed path from y to x, but not both. The sets of vertices and edges of a graph G are denoted by V(G) and E(G), respectively. When speaking about subgraphs and quotient graphs, we understand the strong sense, i.e., regular monos and regular epis, respectively. A pair (G,@) consisting of a graph G and a vertex a of G is called a pointed graph; homomorphisms of pointed graphs are those graph homomorphisms which preserve the chosen vertices. Construction " . Let G and H be graphs, and let a be a vertex of H. We denote by a H ~ G the graph obtained from G by gluing onto each vertex a copy of the graph H using the gluing vertex a: a~ C V V V ~ G H a H~ G Formally, H ~, G is the quotient of the sum'G + I I H × {x}, where x~ ]GfT~ H x {x} is a copy of the graph H, under the equivalence ~ with x ~ (a, x) for each x 6 V(G) and with all other equivalence classes singleton sets. Construction #. For each pointed graph (H,b) denote by (H,b) # the qraph obtained from H by iteratively gluing a copy of H onto vertices (using b as the gluing point) in such a way that, in the end, each ~ertex X has a separate copy of (H,b) # glued to x. Thus, (H,b) # is the direct limit of the following c~-sequence Ho ~ 1 2 ~ ~ H H ... of subgraphs: Ho is the graph with single vertex b and no edge, I H = H, 2 H is obtained by gluing a copy of H on each vertex in V(H) - {b}, i.e., 1 H is the quotient of H + ~ H × {x} (where H×{x} is xeV(H),x~b a copy of H disjoint with H) under the least equivalence ~ with x ~ (b,x) for all x E V(H) - b: and, in general, Hn+ 1 is obtained by gluing a copy of H on each vertex in V(H n) - V(Hn_I), i.e., Hn+ 1 is the quotient of n + H ~ H x {x} x~V(Hn]-V(Hn_ )I (where H × {x} is a copy of H disjoint with H n) under the least equivalence ~ with x ~ (b,x) for all x E V(Hn)-V(Hn_I). 10 H H H H 2 H 3 H Construction ~. For each graph G we denote by G the following graph. I I I 4 98 I I I I G G without edges Formally, V(G) is a disjoint union of V(G), V(G) = {~Ix e V(G)} and {1,2,...,9} (we suppose, for simplicity, that the sets are disjoint), and E(G) = E(G) U E(K) U {x,~)}xeV(G) O {C~,I)}xeV(G) , where K denotes the above graph on the vertices 1,2,...,9. Construction 4. Given graphs G and H, we denote by G ~'~ H the graph obtained from H by first gluing a copy of the following 5 graph C3,5 6// ~ \ / 3

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.