![Arxiv:2012.00853V2 [Math.CT] 25 Apr 2021 Introduction Ha Hs Ups St Ahrtoelclojcs N Tdefin Eq It It, and Under Spaces”](http://data.docslib.org/img/f8cff51bbd3f934309372a1e81034cec-1.webp)
arXiv:2012.00853v2 [math.CT] 25 Apr 2021 Introduction ha hs ups st ahrtoelclojcs n tdefin eq it it, and under spaces”. objects, family “structured local canonical of those categories this gather between with to index sense. functor is rather some points purpose in up whose whose universal ends space sheaf is one a which bu is object, object: data”, object ambient “local ambient an as an “ambie under see algebraic under to objects of wan object category one local a them one with between starts maps one and objects following: the is uation ino -iesoa pcr.W ol u ptecnrlphilosop central the up o sum claim: “spectra could order following We of the and s study spectra. the lattices of 2-dimensional sense flavors distributive of some different of tion in the is spectra theory of about model study for categorical is the duality mathematics: as Stone of sense while regions some several in in resume spectrum role could of prominent notion a the played gener about have more approaches and current multi-adjoint, synthesise right to of notion the through spectra em utdfrtpscntutos nDesapoc,testa the of approach, notion Diers the in was constructions; tra constru topos for to could suited used step terms first is called the variously content h approach, was a theoretic geometric what topos with this the situations where b In algebraic However, step spectrum. identifying second construction. step a the first and to a processes “scenario”: it similar way the in [ point-set in [ synthesis of approaches a topos-theoretic provide the to Contrasting done. been hnamsiiiy n hs emti enn smr ute u mo But as subtle. refer more will is we meaning - geometric assumption whose and admissibility, than rm en on-e,cudse i hnmd” h leri s algebraic the “handmade”, t while bit and a data, seem semantical could on proce point-set, based to and being “univers (admissibility) concrete trum, to situation more more way is complicated also way a concrete but Diers from and abstract start natural more and very is data, a approach in theoretic leads topos situation this practice, hsppr oehrwt [ with together paper, This ni o,svrladrte needn rpsl oconstru to proposals independent rather and several now, Until ihtento fsal uco.W lorcl h ikwit finite link locally situation and the a multi-adjoints recall in right involved also accessible We aspects between factorization functor. a describe stable We and of purpose. notion geometric the multi for relatin with of on conditions Beck-Chevalley theory later proving o adjoint, the used local be case around assum will revolves jointly the that part units In results first local of This one. cone unique junction. canonical a a than deploys rather o object Situations solutions canonical spectra. of of construction point-set egorical, revisited. ir eeoe eea hoyo ih ut-don fun multi-adjoint right of theory general a developed Diers tbefntr n ih multi-adjoints right and functors Stable pcr rs hnfe osrcinfails construction free when arise spectra ih multi-adjoint right 20 nDesster fSetu I Spectrum of theory Diers’s On ,Desapoc smr ueyctgrcli t rmss n stric and premises, its in categorical purely more is approach Diers ], admissibility ir condition Diers 19 ,i h rtpr fatoodwr nDescntuto of construction Diers on work twofold a of part first the is ], n[ in xlOsmond Axel coe 2020 October hc sa rtsgtfrmr btatadalgebraic and abstract more far sight first at is which , 5 Abstract ], geometry ntescn at-wihi aiyecutrdin encountered easily is which - part second the in 1 oa ducin n h equivalence the and adjunctions local g n[ in mliesl rprisrtr sets return properties “multiversal” f ymlirsnal aeoisi also is categories multipresentable ly 16 5 fmliajnto.Terelation The multi-adjunction. of n h oeo h nto nad- an of unit the of role the ing ,[ ], eybodoeve ftesit- the of overview broad very A . ], esnhszda eovn around revolving as synthesized be ajitadrclso precises or recalls and -adjoint h repoutcompletion, free-product the h n o hyaerltd Spectra related. are they how and s td h ekrnto of notion weaker the study lso 6 hois,frasilt-en no- still-to-define a for theories”, f triples tn on o osrcigspec- constructing for point rting tr edn oaprl cat- purely a to leading ctors ly oasre fppr devoted papers of series a to ally, ,[ ], salf don oacomparison a to adjoint left a es ih ut-don,each multi-adjoint, right a f tsetai eea a have way general a in spectra ct t oino pcr olw a follows spectra of notion oth l si sbsdo syntactic on based is it as al” 3 al oascaecanonically associate to fails t ybhn hsnto though notion this behind hy tacrepnignto of notion corresponding a ct etao omttv rings, commutative of pectra ,o [ or ], h pcrmo nambient an of spectrum The ntne leri geometry algebraic instance, st aua construction; natural a to ss taini trswt sfar is with starts it ituation ecntuto ftespec- the of construction he oino pcrm The spectrum. of notion a aiyo aoia local canonical of family a dsrcue oeloosely, more structure; ed t bet,adacasof class a and objects, nt” n[ in ipdwt structural a with uipped de emti content, geometric idden 16 uoasih additional slight a dulo 6 ,wihwssae in stated was which ], ,o hc ewl also will we which of ], tly more natural. Moreover this method subsumes not only most of the usual examples of algebraic geometry or the structured versions of Stone dualities, but also a vast list of exotic examples which were unsuspected before Diers investigation, and are not necessarily suited for the general topos theoretic approach. Indeed, the later requires the categories it use as “local data” to construct the spectrum to be axiomatisable by geometric theories, as well as the factorization system need to be left generated - in some sense, also axiomatisable. In particular local objects must form a (non full) subcategory of ambient objects, and are models of a geometric extension of the theory behind the ambient objects. In Diers this condition is largely relaxed as local objects are related to ambient objects through a functor that is not required to be faithful nor injective on objects.
Those two papers will hence deals with Diers approach. While Diers work on multiversal con- structions in category theory has been rather well known amongst category theory community, his presentation of the spectral construction seems to have been poorly acknowledged, perhaps due to a restricted circulation of the main paper [11]: this contrasts with the quality of the paper itself and the highly practical and comprehensive construction he proposed. We hope those two new papers will help to make more people aware of Diers work on spectra.
This first paper will revolves essentially about the first step in Diers approach. More precisely, we will focus on its notion of interest and its variations, providing different presentations of this situation and a purely categorical analysis of it, postponing the actual construction of the spectrum and the geometric analysis in the second paper. Most of the content of this paper is expository, and aimed at synthesizing as much as possible of the “algebraic” aspects of the construction, gath- ering and relating different notions dispatched in several papers as [8], [7], [22], [2]... However we try to present them as explicitly and originally as possible, providing alternative presentation and proofs of some already known results, and also providing new observations at some points. The last section will also provide a totally new method, whose relevance will however appear later with a 2-categorical version.
Right multi-adjoint were introduced by Diers and extensively studied in [8], [11] or [7], amongst “multi” versions of universal properties and usual categorical constructions. Multiversal properties are analogous of universal properties where, rather than having a unique solution representing a construction, one has a canonical small set of solution jointly assuming the universal property. The prototypical situation is the notion of multirepresentable functor into Set, that is, a functor that decompose as a coproduct of representable functors: the other situations are constructed from this as well as universal construction are done by representing a functor into Set. For instance, as well as a (co)limit over a diagram is an object representing the functor assigning to any object the set of (co)cones over the diagram with this object as tip, a multi (co)limit exists when this functor is multi-representable, and the “local representing objects” form a small family of (co)cone such that any other (co)cone factorizes uniquely through exactly one of them. The other main example of multi-construction is the notion of right multi-adjoint, to which revolves the present paper. In an ordinary adjunction, any object in the category where the right adjoint lands admits one unit uniquely factorizing any map from this object toward an object in the range of the right adjoint: and the left adjoint is used to provide the codomain of this unit. In a multi-adjonction, there is no global left adjoint, hence no uniquely defined unit under a fixed object: one rather has a small cone of local units under a fixed object, which jointly play the role of the unit in the sense that any arrow from this object toward the righ adjoint uniquely factorizes through exactly one of those units, followed by a morphism in the range of the right adjoint. This is a special situation of the more general notion of local right adjoint, were again one lacks a global left adjoint, but is able to construct local left adjoints to restrictions at slices: however in this case, while any object still posses a cone of local units, one cannot in general enforce the smallness of this cone, and right multi-adjointess amounts in fact exactly to “local adjointness plus small solution set condition”.
The second part of this work concerns a characterization of right multi-adjoint through free coproduct completion, and we reprove as explicitly as possible a result, already known in [8] but proved in a different way, stating that a functor is right multi-adjoint if and only if its free co- product extension is right adjoint. We will see in the second part that this result is the “discrete version” of the spectral adjunction.
The third section of this work is about the orthogonality aspects of local adjunctions, and gives some characterizations and properties of “diagonally universal morphism” as defined in [11], which will play a central role in the second part when defining the topology of the spectrum.
2 The fourth section contains some new results and improvements; we consider the case of accessi- ble right multi-adjoint in the context of locally finitely multipresentable categories -which were also introduced by Diers - and revisit some results expressing how to construct locally finitely multi- presentable categories from accessible right multi-adjoint satisfying some relative full faithfulness property; we also provide a ”right accessible multi-adjoint theorem” in theorem 4.4, and finally improve a theorem of [8] on a method for constructing locally finitely multipresentable categories from some orthogonality conditions.
The last part contains new results. We examine a situation, in the context of a factorization system in the presence of a terminal object, producing a case of right multi-adjunction. However in this last section, to emphasize the geometric interpretation, we will work with a convention that produces actually left multi-adjonction. The interest of this construction will be revealed in a future work applying its bicategorical analog to the bicategory of Grothendieck toposes in order to construct notions of “2-geometries” and spectra for toposes.
1 Local right adjoints and stable functors
In this first section, we recall our three notions of interest, namely local right adjoints, right multi-adjoints and stable functors. We first give some technical points about the behavior of the local units of the local adjunctions. We also prove that for a local right adjoint, the local adjunctions enjoy automatically a Beck-Chevalley condition, which was seemingly unnoticed until now. Then we turn to different characterization of local adjointness in term of nerves and initial family, and introduce the stronger notion of right multi-adjoint and recall a variant of Freyd adjoint functor theorem for multi-adjoint. Finally we turn to the notion of stable functor, as studied by Taylor in [22], and also in [25], and we prove equivalence with the notion of local right adjoint. Definition 1.1. A functor U : A→B is said to be a local right adjoint if for each object A of A the restriction of U to the slice A/A has a left adjoint
LA
A/A ⊥ B/U(A)
U/A where we denote Af the domain of the arrow LA(f) in A/A. In the following we will also denote A U/A as UA for concision. The maps ηf for f : B → U(A) are called local units under B. The definition of a local right adjoint means that for any arrow f : B → U(A) in B ↓ U and u : A′ → A in A/A we have triangles in B and A respectively
f BU(A) AU(u) A ǫu LA(U(u)) A ηf UALA(f) U(A ) ′ f A u A satisfying the triangle identities
ηA A UA LA(η ) UA UALAUA LA LAUALA
A ǫA UA(ǫ ) LA
UA LA In other words we have the following retractions
UA(u) LA(f) ′ U(A ) Af A L (ηA) ηU(u) A f
U L (U(u)) LAUALA(f) U(AU(u)) A A U(A) AUALA(f) A A A UA(ǫ ) ǫ u LA(f) ′ U(A ) Af LA(f) UA(u)
3 defining an isomorphism A/A[LA(f),u] ≃B/U(A)[f,U(u)] sending an arrow v : LA(f) → u, resp. an arrow g : f → U(u), to the composite triangle on the left, resp. on the right
A η U (v) A f A ′ LA(g) ǫu ′ BU(Af ) U(A ) Af AU(u) A
UALA(f) L (U(u)) f U (u) A u A LA(f) U(A) A Remark. Beware that in general we cannot enforce the counits to be pointwise iso, that is, to require each UA to be full and faithful. Hence the factorization of a morphism in the range of U may not be trivial. Morally, the factorization through the unit only takes in account the object of A whose strict image is the codomain, while, even when the domain is in the image of U, the factorization may not remember from which precise object in A it comes from.
Remark. For any u : A1 → A2 in A, functoriality of U makes the following square commute up to equality
UA1 A/A1 B/U(A1)
A/u = B/U(u)
A/A2 B/U(A2) UA2 its corresponding mate
LA1 A/A1 B/U(A1)
A/u σu B/U(u)
A/A2 B/U(A2) LA2 defined as the composite
A2 A1 ǫA/uL LA2 B/U(u)(η ) A1 LA2 B/U(u) LA2 B/U(u)UA1 LA1 = LA2 UA2 A/uLA1 A/uLA1
This mates relates in a canonical way the unit of any f : B → U(A1) and the unit of the composite U(u)f : B → U(A2) as seen in the following diagram
LA1 (f) Af A1
A2 ǫA/uL (f) u A1
A LA2 UA2 A/uLA1 (f) A UA2 A/uLA1 (f) 2
LA B/U(u)UA LA (f) 2 1 1 LA2 (B/U(u)(f))
A A B/U(u)UA1 LA1 (f) A1 U(u)f LA2 B/U(u)(ηf ) But surprisingly, this mates is automatically an isomorphism because of the universal property of the units, as stated in the following proposition:
Theorem 1.2. Let be U : A→B a local right adjoint. Then for any u : A1 → A2 in A, we have the Beck-Chevalley condition at u, that is, the canonical transformation σu is a point-wise isomorphism. u Proof. Remark that for each u : A1 → A2 and f : B → U(A1), the morphism σf : LA2 (B/U(u)(f)) →
LA1 (f) is in A, and we have a factorization f BU(A1) A U(u) η 1 f UA1 LA1 (f)
ηA2 U(u)f U(Af ) U(A2)
u U(σf ) UA2 LA2 (U(u)f) U(AU(u)f )
4 u A2 Observe that σf is the unique arrow in A provided by the universal property of the unit ηU(u)f A1 at ηf seen as an arrow U(u)f → U(u)UA1 LA1 (f) in B/U(A2). But on the other hand, by the A1 A2 u universal property of ηf at ηU(u)f seen as an arrow f → UA1 LA1 (f)U(σf ) in B/U(A1), there exists a unique arrow w : Af → AU(u)f in A such that
f BU(A1) A1 η UA LA (f) f A2 1 1 ηU(u)f UA1 LA1 (f)
U(Af ) U(AU(u)f ) u U(Af ) U(w) U(σf )
u Now we prove that w and σf are mutual inverses in A. First, as
u A2 A1 U(σf )ηU(u)f = ηf and 1 is the unique map induced by ηA1 seen as an arrow f → U L (f), then necessarily LA1 (f) f A1 A1 we have a retraction in A Af Af
w u σf AU(u)f but again, as now ηA2 = U(w)ηA1 and 1 is the unique map induced by ηA2 as an U(u)f f LA2 (U(u)f) U(u)f arrow U(u)f → UA2 LA2 (U(u)f), we have a retraction in A
Af u σf w
AU(u)f AU(u)f
u and σf defines both an iso AU(u)f ≃ Af in A and LA2 A/u(f) ≃ B/U(u)LA1 (f) which can be shown to be natural. Remark. Beck-Chevalley condition says that factorization through local unit are not modified by postcomposing with an arrow in the range of U: that is, for f : B → U(A1) and u : A1 → A2, then the Beck chevalley transformation provides an isomorphism Af ≃ AU(u)f .
Af A Corollary 1.3. Let be f : B → U(A): then we have Af ≃ AηA in A and η A ≃ ηf in B ↓ U. f ηf Proof. Consider the following diagram
A ηf UALA(f) BU(Af ) U(A)
Af A η UA LA (η ) ηA f f f f U(A A ) ηf
UALA(f) Then by what precedes we have σ A is an iso as seen in the following diagram ηf
A ηf UALA(f) BU(Af ) U(A)
Af η A ηA U L (η ) f Af Af f
ηA U L (f)ηA U(A A ) A A f ηf
U L (f) σ A A ηA f A ≃ UALA(UALA(f)ηf )
U(A A ) UALA(f)ηf
5 A But UALA(f)ηf = f, exhibiting an isomorphism
B Af A η η ηA f f
U(Af ) ≃ U(AηA ) U L (f) f σ A A ηA f
A consequence is that local units that are related by an arrow in the range of U must actually be isomorphic as objects under their domain:
Corollary 1.4. Let be f1 : B → U(A1) and f2 : B → U(A2), such that there exists a morphism u in A and a triangle B ηA1 ηA2 f1 f2
U(Af ) U(Af ) 1 U(u) 2 then u is an isomorphism. Proof. By theorem 1.2, we have that
A A η f1 ≃ ηA1 η f2 ≃ ηA2 ηA1 f1 ηA2 f2 f1 f2 But by Beck-Chevalley condition at u, we also have
A A A η f1 ≃ η f2 = η f2 ηA1 U(u)ηA1 ηA2 f1 f1 f2
From their very universal property, locals units under a given object have to live isolated from each other, each one in its connected component:
Corollary 1.5. If U : A→B is local right adjoint, then for any B, any two f1 : B → U(A1), f2 : B → U(A2) in the same connected component of B ↓ U factorize through the same unit, that is ηA1 ≃ ηA2 , A ≃ A . f1 f2 f1 f2
A Corollary 1.6. For any f : B → U(A), ηf is initial in (B ↓ U) ↓ f.
Definition 1.7. Let be C a category; multi-initial family in C is a family of objects (Xi)i∈I such that for any C in C there is a unique i ∈ I and a unique arrow Xi → C. A Xi for i ∈ I is a local initial object.
Remark. Observe that in this definition, if one has an arrow f : C1 → C2, then C1 and C2 lie under the same local initial object. More generally, if two objects C1, C2 are in the same connected component and Xi → C1, Xj → C2 are the initial maps, then they lie under the same initial object: for there is a zigzag
B1 ... Bn
C1 B2 Bn−1 C2 and by uniqueness of the local initial object over a given object, necessarily the local object over B1 is the same as the local initial object over B3 because they both lies over B2 and so on. Conversely any two objects under a same local initial object are in the same connected component. Hence there is exactly one local initial object by connected component. Proposition 1.8. Let U : A→B be a local right adjoint and B in B: then the coma category B ↓ U has a multi-initial family.
6 Proof. We claim that the (large) class of local units under B is a multi-initial family in B ↓ U. First, for any f : B → U(A), we have by local adjunction in A an arrow LA(f) in B ↓ U; but now ′ A′ suppose there is an other g : B → U(A ) such that ηg has a map g → f in B ↓ U, that is there A1 is a map u in A such that f = U(u)ηg ; but then by the universal property of the unit there is a unique factorization f BU(A)
A ηf UALA(f)
U(Af ) ′ UA(u) ηA g U(w)
U(Ag)
A A′ But by corollary 1.4, this forces ηf ≃ ηg . Remark. Observe that without further assumption, the multi-initial family of local units at a given B may not be small. This is the point of the following notion. Definition 1.9. A functor U : A→B is said to be a right multi-adjoint if for any B in B there is a small multi-initial family in the comma B ↓ U. Remark. Observe that this definition is indexed by the domain, that is, by the object B in B, while the definition of local right adjoint was indexed by the objects A of A. However it is easy to see A that any right multi-adjoint is in particular a local right adjoint: for any f : B → U(A) define ηf A A to be the unique local initial object over f, and LA(f) to be the unique map ηf → f. Then ηf has the universal property of the unit as any triangle f → U(u) in B/U(A)
f BU(A)
g U(u) U(A′) can be seen as a triangle g → f in B ↓ U forcing g and f to be in the same connected component. A A Therefore ηf is also the initial object over g, inducing an unique arrow ηf → g in B ↓ U, that is a unique arrow LA(f) → u in A/A as desired. Existence of a multi-initial object in the comma is reminiscent of the so called solution set condition in Freyd Adjoint Functor Theorem. Let us precise this fact, in order to retrieve an analogous multi-adjoint theorem, in the vein of [22]. First we need a weakening of the notion of initial family
Definition 1.10. Let be C a category; a weakly initial family is a family (Xi)i∈I such that for any object C, there some i ∈ I and some arrow Xi → C. Remark. Observe that in this definition, there is no requirement of uniqueness, nor for the index, nor for the arrow, so there may be several weakly initial objects over an arbitrary one. In the following we are interested in small weakly initial families: Definition 1.11. A functor U : A→B is said to satisfy the Solution Set Condition if each of the coma B ↓ U admits a small weakly initial family: that is a family (ni : B → U(Ai))i∈IB such that any map from B to U factorizes through some (non necessarily unique) ni. Proposition 1.12. A functor U : A→B is a right multi-adjoint if and only if it is local right adjoint and satisfies the Solution Set Condition. Proof. It is obvious that a right multi-adjoint satisfies the Solution Set Condition as for any object B the small multi-initial family of B ↓ U is in particular a small weakly initial family, and any right multi-adjoint is trivially local right adjoint.
For the converse, observe that the small family of local units of a right multi-adjoint produces a small (weakly) initial family, so that it always satisfies the solution set condition. Preservation of wide pullback just come from the fact that they are ordinary products in the slices, hence preserved by the restriction as it is right adjoint.
7 Definition 1.13. A connected limit is a limit indexed by a connected category; a wide pullback is a limit of a diagram over a set of arrows with common codomain. In particular any wide-pullback is a connected limit. Proposition 1.14. A category has wide pullbacks if and only if each slice A/A has products. A functor preserves connected limits if and only if it preserves wide pullbacks. In Freyd adjoint functor theorem, assuming completeness and local smallness of the domain category A, one can prove that a functor preserving limits is right adjoint if it satisfies the solu- tion set condition : this is achieved by constructing an initial object in any B ↓ U as the limit lim U(Ai) ≃ U(lim Ai). Similarly we can give a corresponding right multi-adjoint theorem, an i∈I i∈I accessible version of will be given in theorem 4.4 : Proposition 1.15. Let A be a category with wide pullbacks; then a functor U : A→B is a right multi-adjoint if and only if it satisfies the solution set condition and preserves wide pullbacks. Proof. Indeed, assuming that A has connected limit and that U preserves them, the solution set condition will enable us to prove that its restriction on any slice UA : A/A → B/U(A) is right adjoint by constructing an initial object in any of the coma f ↓ UA whit f : B → U(A) ∈B/U(A) (this is nothing but the category of factorizations of f through U) as the following. If SB = (ni :
B → U(Ai))i∈IB is the small weakly initial family in B ↓ U given by the solution set condition, define Sf = SB ↓ f consisting of all the pairs (u,i) with i ∈ IB and f = F (u)ni a factorization of f through ni: then we can do the wide pullback of the Ai over A and it is preserved by U, so that we have
U(×A(Ai)i∈Sf ) ≃×U(A)(U(Ai))i∈Sf
U(pi) U(pj ) ∃! n U(A ) BUi (A ) i nj j f
U(A)
Now we claim that this unique arrow B → U(×A(Ai)i∈Sf ) is the desired initial object in f ↓ UA, as any factorization of f factorizes itself through some of the ni in Sf hence through the wide pullback. For the converse we will make use of stability in order to recover the Solution Set Condition, while the preservation of wide pullback just come from the fact that they are ordinary products in the slices, hence preserved by the restriction as it is right adjoint. Now we turn to another facet of the local adjunction.
Definition 1.16. Let U : A→B be a functor; then the co-nerve of U is the functor
U B N [A, Set]
U sending each B in B to the functor NB = B[B,U(−)] : A → Set. For any B in B, we have a discrete opfibration
π : N U → A B Z B whose objects are pairs (A, f) with A in A and f : B → U(A), and morphisms (A1,f1) → (A2,f2) are u : A1 → A2 with U(u)f1 = f2. There is a general result saying that a functor is a right adjoint if the projection from the category of elements of its co-nerve at each object has a limit; in fact this says that the co-nerve functor is representable, that is, there is an initial object in the category of elements, which is the unit. We give here the corresponding statement for a right multi-adjoint. Definition 1.17. Let U : A→B be a functor; then the co-nerve of U is the functor
U B N [A, Set]
U sending each B in B to the functor NB = B[B,U(−)] : A → Set.
8 For any B in B, we have a discrete opfibration
π : N U → A B Z B whose objects are pairs (A, f) with A in A and f : B → U(A), and morphisms (A1,f1) → (A2,f2) are u : A1 → A2 with U(u)f1 = f2. There is a general result saying that a functor is a right adjoint if the projection from the category of elements of its co-nerve at each object has a limit; in fact this says that the co-nerve functor is representable, that is, there is an initial object in the category of elements, which is the unit. We give here the corresponding statement for a right multi-adjoint. Definition 1.18. Let F : I → A be a functor: then a multilimit of F is a small family of cones
j (p : Lj → F (i)) i∈I i j∈J such that for any cone (fi : X → F (i))i∈I there is a unique j ∈ J and a unique factorization of j the cone (fi)i∈I through the cone (pi )i∈I .
A functor U : A→B preserves multilimits (or also, is multicontinuous) if for any multilimit j k (pi : Lj → F (i))i∈I, j∈J in A, there is a multilimit (qi : Mk → UF (i))i∈I,k∈K in B and for each k ∈ K we have Mk ≃ U(Lj)
ja∈Jk j where Jk is the set of j ∈ J such that the cone (U(pi ): U(Lj ) → UF (i))i∈I factorizes through Mk. Before going further, we think relevant to introduce the dual notion of multilimits, for we are going to use them also at the end later in this paper and also in a closure theorem for multireflective subcategories in a moment: Definition 1.19. Let F : I → A be a functor. Then a multicolimit of F is a small family of cocones j ((qi : F (i) → Xj )i∈I )j∈J that is multi-initial the the category of cocones over F : that is, any other cocone (fi : F (i) → A)i∈I j factorizes uniquely through one of the (qi : F (i) → Xj )i∈I for a unique j ∈ J.
A category is (finitely) multicocomplete if any (finite) diagram admits a multi-colimit. Dually, a multicomplete category is a category where any (finite) diagram has a multi-limits.
j A functor U : A→B is multicocontinuous if for any multicolimit ((qi : F (i)i → Xj )i∈I )j∈J in k A, then the composite UF : I → B has a multicolimit ((si : UF (i) → Yk)i∈I )k∈K such that for any k ∈ K we have a coproduct decomposition
Yk ≃ U(Xj )
YJk
j k where Jk = {j ∈ J | (U(qi ): UF (i) → U(Xj ))i∈I factorizes through (si )i∈I }. Remark. The universal property of the multicolimits can be encapsulated, that for any other X in B, in the following isomorphism
B[F (i),X] ≃ lim B[Xj,X] i∈I ja∈J Remark. Remark that any (co)limit is in particular a multi-(co)limit with a single cone. In particu- lar (co)completeness implies multi-(co)completeness, and (co)continuity implies multi-(co)continuity. Then in particular any corepresentable functors are multicontinuous. Let us get back to right multi-adjoints. The following observation just gives the obvious analog of the characterization of right adjoint in terms of the existence of the limit of the projection of the category of elements of the nerve: Proposition 1.20. A functor U : A→B is a local right adjoint if and only if for any B in B, U each connected component of NB has an initial object. Moreover, U is a right multi-adjoint if U NB has a set of connected components.R Equivalently, U is a right multi-adjoint if and only if U Rthe functor πB : NB → A has a multi-limit in A and U preserves it. R 9 The first half of this fact is tautological; for the second part, one can adapt [4][proposition 3.3.2].
Now from what was said, it appears that a right multi-adjoint is a functor such that the associated conerve in any object is “locally representable”. Indeed, any arrow from an object B toward U is determined first by the connected component it lies in, which corresponds to the local unit it factorizes through, and secondly by a choice of map in A. This amounts to the following: Proposition 1.21. Let U be a multi-right adjoint: then for each B one has
U NB ≃ A[Ai, −] ia∈IB with IB the set of connected components of B and ni : B → U(Ai) the initial object of the connected component i. We finally end this subsection on a result on multireflective subcategories. Recall that full reflectives subcategories inherits limits and colimits from their ambient category in the following way. Limits can be computed directly in the subcategory and are created by the inclusion, which is proven by proving that the reflection of a limit computed in the ambient category is a limit in the subcategory, and observing the later must be preserved by inclusion, so that this reflection was actually an iso. For colimits, one compute the colimit in the ambient category, then take its reflection. In the context of multireflection, the correct analog of those statement are in term of limits and multicolimits. :Theorem 1.22. Let be U : A ֒→B a full multireflective subcategory; then − if B has colimits, then A has multicolimits − if B has limits, then A has connected limits and they are created by U.
Proof. For multicolimits, take a functor F : I → A; then one can compute the colimit (qi : UF (i) → colimi∈I UF (i))i∈I in B. Then consider its small cone of local units nx : (colimi∈I UF (i) → U(Ax))x∈I . Then by fullness of U each composite nxqi : UF (i) → U(Ax) comes from a colimi∈I UF (i) x x unique map q : F (i) → Ax and (q : F (i) → Ax)i∈I,x∈I is a multicolimit of F . i i colimi∈I UF (i)
Now for connected limits, if we suppose F : I → A with I connected, then the category limi∈I UF (i) ↓ UF is connected; then by corollary 1.5, all the limit projections pi factorize through a same local unit nF p lim UF (i) i UF (i) i∈I
nF F U(ui )
U(AF )
F Then (U(ui ): AF → Ai)i∈I is a limiting cone in A: indeed, any cone (ui : A → F (i))i∈I in A is transported by U to a cone in B, were it induces a unique arrow (U(ui))i∈I U(A) → limi∈I UF (i) and the composite nF (U(ui))i∈I : U(A) → U(AF ) comes uniquely from an arrow A → AF as desired. Hence AF is a limit of the connected diagram F , but as U preserves connected limits, nF was actually an isomorphism limi∈I UF (i) ≃ U(AF ). Remark. In fact, concerning connected limit, the condition that U is full can be slightly relaxed into the condition of being relative full and faithful, which will be defined later in definition 3.3 Finally we come to an alternative notion encapsulating the property of being a local right adjoint, but in a way that is more related to factorization systems. This was studied in [22] and [25], and we prove there that this notion coincides with local right adjointness. It relies on an atlernative presentation of local unit in a more “orthogonality structure” spirit. Definition 1.23. A candidate (diagonally universal toward U in the terminology of Diers), is a morphism n : B → U(A) such that for any square of the following form there exist an unique morphism w : A → A1 such that U(w) diagonalizes uniquely the square and the left triangle already commutes in A
f BU(A1) A1 U(w) n U(v) ∃!w v u U(A) U(A2) A A2 U(u)
10 Definition 1.24. A functor U : A→B is stable when any morphism f : B → U(A) factorizes uniquely through the range of U as
f BU(A)
nf U(uf ) U(Af ) where nf : B → U(Af ) is a candidate. We refer to this factorization as the stable factorization of f and to nf as the candidate of f. Remark. Then the candidate for f is the initial object in the category of factorizations of f through the range of U. Indeed, for any other factorization through U
f BU(A)
g U(u) U(A′) one gets a square as below, where nf produces a unique w such that U(w) is a filler
g BU(A′)
nf U(v) U(u)
U(Af ) U(A) U(uf )
Proposition 1.25. For a functor U : A→B and B in B we have the following
− If a candidate n1 : B → U(A1) admits an arrow n2 → n1 from another candidate n2 : B → U(A2) in B ↓ U, then we have n1 ≃ n2 in B ↓ U and A1 ≃ A2 in A. − In particular, any two candidates in a same connected component of B ↓ U are isomorphic. − If f : B → U(A) admits a stable factorization, then it is unique up to unique isomorphism. − In particular, when U is stable, the stable factorization of any arrow is unique up to unique isomorphism.
Proof. The first item is easily shown to implies the other ones. Suppose we have n1, n2 and a triangle n1 BU(A1)
n2 U(u)
U(A2) Then we have a unique filler n2 BU(A2)
n1 U(v) U(u)
U(A1) U(A1) U(1A1 ) But now there is a unique filler of the square
n1 BU(A1)
n2 U(w) U(v)
U(A2) U(A2) U(1A2 ) so that u : A2 → A2 is both a retraction and a section in AA, hence an isomorphism, so that n1 ≃ n2 in B ↓ U.
11 Proposition 1.26. For any square as below
f1 B1 U(A1)
f U(u)
B2 U(A2) f2 the stable factorizations of f1 and f2 are related by a unique morphism in A such that
nf1 U(uf1 ) B1 U(Af1 ) U(A1)
f U(wg,u) U(u)
B U(A ) U(A ) 2 nf f2 2 2 U(uf2 )
Proof. The desired wg,u is the filler of the square
nf1 B1 U(Af1 )
n f f2 U(uuf1 )
U(Af2 ) U(A2) U(uf2 )
Theorem 1.27. Stable functors and local right adjoints coincide. Proof. Let U : A→B be a stable functor. For each A defines the functor
LA A/A ⊥ B/U(A)
U/A where La returns the left part of the initial factorization of an arrow by its associated candidate:
LA : B/U(A) → A/A f BU(A) uf 7→ Af → A nf U(uf ) U(Af )
We can easily prove this functor is left adjoint to U/A, but it is more direct to observe that the family of candidates under B is a multi-initial family in B ↓ U. Hence U is a local right adjoint.
For the converse, suppose U is a local right adjoint. We claim that candidates are arrows A n : B → U(A) such that LA(n) provides an iso Af ≃ A in A, hence ηn ≃ n in B ↓ U. Let be a square g BU(A1)
n U(u)
U(A) U(A2) U(v) Recall by theorem 1.2 we also have that composing with U(v) does not modify the unit, as we u A2 A have an isomorphism σn : ηU(v)n ≃ ηf . But the triangle
g BU(A1)
U(u) U(u)n
U(A2)
12 provides us with a unique arrow w : AU(v)n → A1 such that
g A1 BU(A1)
w A2 U(u) ηU(v)n U(w) U(u)
AU(v)n A2 U(AU(v)n) U(A2) LA(U(u)n) UA2 LA2 (U(u)n)
A u Then by inserting the local inverses of ηn and σn in the square above and using the universal An property of η A at the triangle ηn g B U(A1)
A A η n n η ηA n f U(w)
u −1 U(A) U(LA(n)) U(An) U((σn) ) U(AU(v)n) U(w)
−1 u U(LA(n) ) U(σn)
u −1 U(A) U(LA(n)) U(An) U((σn) ) U(AU(v)n) U(u)
−1 u U(LA(n) ) σn
U(An) UA2 LA2 (U(v)n)
U(LA(n))
U(A) U(A2) provides us with a triangle as below and a diagonalization in B
A1 u −1 −1 w(σf ) LA(n) u
Av A2
In particular, local units are candidates by corollary 1.3. Hence for any arrow f : B → U(A), A the factorization through the unit as f = UALA(f)ηf provides a stable factorization through a candidate. This achieve to prove that stable functors and local right adjoint can be used indifferently and are two ways of encapsulating the same property.
However in the following, and especially in the second paper, we will give more interest to right multi-adjoint for the smallness condition allows us to manipulate local units without size issue.
2 Right multi-adjoints through free product completion
In this section, we give the characterization of right multi-adjoints though the free product completion, following loosely [8] and [23]. In the second part of this work, we are going to show how the notion of spectrum is a way to turn a local right adjoint into a right adjoint, the spectrum functor being the desired left adjoint. But this construction, motivated by geometric and duality theoretic conceptions, process by extracting as much as possible topological and geometric infor- mation from a situation of local adjunction: in some way, it exploits the defect of universality on the algebraic side in order to produce richer structure on the geometric side. In this section, we recall another way to turn a situation of multi-adjunction into a honest adjunction, which is purely algebraic and purer in some sense, but also devoid of any geometric content for this very reason. The relation between those two approach will be studied in more detail in the second part of this work, where this approach through free product completion will appear as the “discrete version” of the spectral adjunction.
The main intuition of this part is that, for a right multi-adjoint U : A→B, the cone of local units under a given object B defines a family of objects in A given by the codomains of those local
13 units. Hence, at the level of families of objects, the multiversality of the construction can be fixed and U will induce an honest adjunction between categories of families of objects of A and B. The good notion of “category of families” here is the one provided by the free product completion, the beginning of this part is devoted to. Definition 2.1. For a category A, the free product completion of A is the category ΠA whose
− objects are functors A(−) : I → A (also denoted (Ai)i∈I ) with I a set,
− and arrows (Ai)i∈I → (Bj )j∈J consist of the data of an application α : J → I and a natural transformation I A(−)
α f A
B J (−)
that is, a J-indexed family (fj : Aα(j) → Bj )j∈J . Proposition 2.2. We have the following properties of the free product completion, for a given category A: − ΠA has small products
There is a codense full embedding ιA : A ֒→ ΠA whose essential image is the subcategory − (ΠA)coco of co-connected objects. − Moreover, the embedding A → ΠA has a right adjoint if and only if A already had products We have a full embedding ΠA ֒→ [A, Set]op whose essential image consists of all small − products of representable − For any category B with small products, we have an equivalence of categories
[A, B] ≃ [ΠA, B]Π
(where [ΠA, B]Π is the category of functors preserving small products) sending any F : A→B
on its right Kan extension ran ιA F and any G : ΠA→B on its restriction GιA. j Proof. For the first item: the product in ΠA of a family of families ((Ai )i∈Ij )j∈J has as indexing i set the disjoint union j∈J Ij and whose member of index (j, i) is the object A(j,i) = Aj ; the projections are given for` each j ∈ J as the transformation
Ij j (Ai )i∈Ij q j pj
Ij A (A ) j∈J (j,i) (j,i)∈ ` Ij ` j∈J
j where pi is the pointwise equality A(j,i) = Ai .
For the second item, the embedding sends an object A of A to the one element family A :1 → A and a morphism f : A1 → A2 to the natural transformation
A1
1 f A
A2
Now we prove objects in the image of this embedding are coconnected, which says that for a family j of families ((Ai )i∈Ij )j∈J , we have
j j ΠA[ (Ai )i∈Ij , ιA(A)] ≃ ΠA[(Ai )i∈I , ιA(A)] jY∈J ja∈J
14 j Indeed, any arrow Πj∈J (Ai )i∈Ij → ιA(A) defines an arrow 1 → j∈J Ij pointing at some pair (j, i) with j ∈ J and i ∈ Ij , and a natural transformation `
1
(j,i) A Ij
qj (Ai)i∈Ij
Ij A (A ) j∈J (j,i) (j,i)∈ ` Ij ` j∈J which is nothing but an arrow f : A(j,i) → A. But in such a case, as 1 is a connected object in Set, this arrow (j, i):1 → j∈J Ij factorizes through Ij for some j ∈ J, pointing the corre- sponding i ∈ Ij , and the natural` transformation f factorizes through the componentwise identity j j j p : A(i,j) = Ai so we have an arrow (Ai )i∈Ij → ιA(A) as desired. Conversely, any coconnected object is of the form ιA(A): indeed a family (Ai)i∈I : I → A is nothing but the product in ΠA of the family (ιA(Ai))i∈I as the set I decomposes as the coproduct I 1 in Set; and any family should be indexed by a connected set to be a coconnected object in ΠA`, but 1 is the only connected set. This also suffice to prove that any object is a product of objects in the range of ιA.
Now suppose that A has products. Then for any family (Ai)i∈I in ΠA one can compute the product in A, i∈I Ai. Now for an object A in A, we have Q ΠA[ιA(A), (Ai)i∈I ] ≃ A[A, Ai] Yi∈I sending a family of arrows (A → Ai)i∈I to the universal map A → i∈I . The unit of this adjunction is iso as ιA is full and faithful, while the counit is the transformationQ ∗ Q Ai i∈I ! ǫ (Ai)i∈I I A (Ai)i∈I
where ǫ(Ai)i∈I has the projection pi : i∈I Ai → Ai has component in i. For the converse, it is easy to see that any right adjoint of theQ embedding ιA sends a family on an object in A with the universal property of the product.
op ょ∗ The embedding ΠA ֒→ [A, Set] justs sends a family (Ai)i∈I to the coproduct i∈I Ai and an arrow (α, (fj )j∈J ) : (Ai)i∈I → (Bj )j∈J to the opposite of the induced map between` the corresponding coproducts in [A, Set] as depicted below
ょ∗ qα(j) ょ∗ Aα(j) Ai i`∈I ∗ ょ ∗ fj hq ょ ij∈J α(j) fj ょ∗ ょ∗ Bj q′ Bj j j`∈J Finally, for a functor U : A→B with B having products; we claim that the right Kan extensions of U is pointwise and can be computed as
ran ιA U(Ai)i∈I = U(Ai) Yi∈I
Indeed for any (Ai)i∈I the comma category (Ai)i∈I ↓ ιA has a small initial I-indexed subcategory consisting of the objects (i, 1Ai ) for i ∈ I, and this subcategory is discrete. Hence calculating the poinwise right Kan extension resumes to calculating the product above. Moreover, as ιA is full and faithful, restricting back this Kan extension along ιA gives back U, in fact up to equality in this context. Proposition 2.3. The embedding A ֒→ ΠA creates connected limits in A. Moreover, ΠA is complete if and only if A has connected limits.
15 Proof. Let be D a connected category and F : D → A; we prove that the singleton ιA(lim F ) is the limit of ιAF in ΠA. Let be a cone (αd, (fd)d∈D : (Ai)i∈I ιAF in ΠA consisting for each d ∈ D of an arrow fd : Aαd → F (d) where αd : 1 → I points to some index; but as D is connected and I is a set, necessarily the αd are all equal to the same index α, so that we actually have a cone (fd : Aα → F (d))d∈D in A, inducing a unique arrow f : Aα → lim F in A. This defines a unique arrow (α, f) : (Ai)i∈I → ιA(lim F ). By what precedes, it is also clear that any connected cone that ιA sends to a limiting cone was already limiting.
Now, recall that a category is complete if and only if it has connected limits and products. But ΠA always has products, so we just have to show that ΠA has connected limits if and only if A does. Let be D a connected category, and F : D → ΠA a functor, with Fd : Id → A its component in d and with transition morphism
Id1 Fd1
Is Fs A
Fd2 Id2 for each s : d1 → d2 in D. Then F defines an oplax cocone (Fd : Id → A)d∈D in Cat, defining uniquely a functor hF i I −→d d∈D A Z where I is the category of elements of the functor I : Dop → Set returning the indexing set
Id : Id2R→ Id1 and the transition map Is for s : d1 → d2: it is indeed well known that the category of elements is the oplax colimit in Cat, and we see here the Id as discrete categories. Now, as D was small and each Id was a set, the category I has a small set π0( I) of connected components. In this context, one can describe the connectedR components as follows.R In set, the colimit of the diagram I is the quotient
colim Id ≃ Id/ ∼D d∈D da∈D ′ ′ ′ where (d, i) ∼D (d ,i ) if there is a zigzag in D relating i and i : this exactly amounts to say that (d, i) and (d′,i′) are in the same connected component of I, so we also have that the connected components of I are exactly equivalence classes [(d, i)]∼DR and R I / ∼ ≃ π ( I) d D 0 Z da∈D
Now, if we restrict the induced functor hFdid∈D along the inclusion of a connected component
I R hFdid∈D
i[(d,i)] ∼D A
F[(d,i)] ∼D [(d, i)]∼D we can compute the limit lim F in A, and this limit is preserved by the inclusion functor [(d,i)]∼D ιA. So the desired limit of F in ΠA is the family
π ( I) → A 0 Z sending the connected component [(d, i)]∼ to the connected limit lim F , and this actually D [(d,i)]∼D coincides with the product in ΠA of the family (lim F[(d,i)] :1 → A) . ∼D [(d,i)]∼D ∈π0(R I)
Proposition 2.4. Any functor U : A→B extends uniquely into a functor ΠU, called its free product extension, making the square below to commute up to equality
ABU
ιA ιB ΠA ΠB ΠU
16 Proof. The functor ΠU just is the right Kan extension ran ιA ιBU, and is defined by sending a family (Ai)i∈I to (U(Ai))i∈I . The following proposition is tautological: Proposition 2.5. A has a multi-initial family if and only if ΠA has an initial object. The following proposition is the core idea of [8][part 4], though we present here a quite different proof. Proposition 2.6. For a functor U : A→B, the following are equivalent: 1. U is a right multi-adjoint
2. U has a relative left adjoint along ιA 3. its free product extension ΠU : ΠA → ΠB is a right adjoint
Proof. Suppose that U is a right multi-adjoint, with IB the set of local units ηx : B → U(Ax) and πB : IB → A its projection sending x to Ax. Define a functor L : B → ΠA sending an object B to the family πB : IB → A, and an arrow f : B1 → B2 to the transformation
IB1 πB1
If Lf A
πb2 IB2 where I sends x to the index of the unit ηAx : B → U(A ) = n and L has component f nxf nxf If (x) f
LAx (nxf): Anxf → Ax as provided in each x ∈ IB2 by the factorization
f B1 B2
Ax η nx nxf
U(Anxf ) U(Ax) UAx LAx (nxf) Observe that the local units of B define in particular a morphism of families
1 B B
! n U I A B πB
corresponding to the family (nx : B → U(Ax))x∈IB . Now it is easy to see that this functor is a relative left adjoint to U along ιA, that is, that for any B in B and A in A we have
ΠA[L(B), ιA(A)] ≃B[B,U(A)]
Indeed, any arrow f : B → U(A) factorizes through a unique nx : B → U(Ax), where x is the A index of the unit ηf , while LA(f): Ax → A provides a morphism in ΠA
IB πB L (f) x A A
A 1 while any arrow (x, u): L(B) → ιA(A) can be pasted with the family of units
1 B B
! n U πB IB A LA(f) x A 1
17 to return an arrow B → U(A), and one just has to use the universal properties of the local units to see that those processes are mutual inverses.
This functor L extends to ΠB as follows: for a family (Bi)i∈I , define L(Bi)i∈I as the family
hπBi ii∈I IBi −→ A ai∈I
sending (i, x) with i ∈ I and x ∈ IBi to the associated Ax; for an arrow(α, f = (fi)i∈I ) : (Bi)i∈I1 → ′ ′ (B ) , that is a family (f : B → B ) , each pair (i, x) ∈ I ′ defines uniquely some i i∈I2 i α(i) i i∈I2 i∈I2 Bi
If (x) ∈ IBα(i) indexing the unit through which factorizes the composite` nxfi : Bα(i) → U(Ax), that is such that fi ′ Bα(i) Bi n I(α,f)(x) nx
U(AIf (x)) U(Ax) U(LAx (nxf)) and define I : I ′ → I as sending (i, x) to this I (x), and define the desired (α,f) i∈I2 Bi i∈I Bi (α,f) morphism L(α, f)` as `
IBi i∈I1 ` hπBi ii∈I1
L I(α,f) f A
hπB′ ii∈I i 2 I ′ Bi i∈`I2 where L denotes the family (L (n f): A → A ) . In particular we have a f Ax x I(α,f) x (i,x)∈` IB′ i∈I2 i morphism (B ) I i i∈I B
πI U n
IBi π A (Bi)i∈I i`∈I where πI : i∈I IBi → I is the projection sending (i, x) on i ∈ I, π(Bi)i∈I : i∈I IBi → A sends (i, x) on Ax`, and n has as components `
(n = n : B → U(A )) (i,x) x i x (i,x)∈ ` IBi i∈I
Then for any (Aj )j∈J in ΠA and (Bi)i∈I in ΠB we have an isomorphism
ΠA[L(Bi)i∈I , (Aj )j∈J ] ≃ ΠB[(Bi)i∈I , (U(Aj ))j∈J ]
Indeed a morphism of family (α, f) : (Bi)i∈I → (U(Aj ))j∈J , that is a family (fj : Bα(j) →
U(Aj ))j∈J , defines an application ξ : J → i∈I IBi sending i to the index of the local unit n = ηAi : B → A , and a morphism of` families ξ(i) fi α(i) ξ(i)
IBi i∈I π ` (Bi)i∈I
L (f ) ξ (Aj )j∈J j j∈J A
(A ) J j j∈J
where L(Aj )j∈J (fj )j∈J consists of all right part in A of their factorization
(LAj (fj ): Aξ(j) → Aj )j∈J
For the converse, use the same argument as for the proof of the first implication by pasting.
Now we prove that if U is such that ΠU is right adjoint to a functor L, then U is multi-adjoint. Observe that with this hypothesis, we have in particular for any B in B a unit η(∗,B) : (∗,B) →
18 ΠUL(∗,B). So if IB denote the indexing set of ΠUL(∗,B) and Ai is the object in A corresponding to the ith index of IB in this family, then we have a cone (ηi : B → U(Ai))i∈IB . Now we prove this is a cone of local units. For any A in A, the unit property of ΠUL(∗,B) provides a factorization
∗B B
n f IB U
if (Ai)i∈IB LA(f) ∗ A A
A for if : ∗ → IB pointing at the index of the local unit ηf and LA(f) returning the image of f along the local left adjoint at A.
3 Factorization aspects
As suggested by the definition of candidates in the notion of stability, orthogonality structures are hidden in the notion of local adjunction. In the same vein, one could ask whether the stable factorization of arrows toward U can be generalized to any arrow, that is, if the orthogonality structure provided by the candidates on the left and the morphisms in the range of U on the right can be completed into a factorization system. In the context studied in [11], this is possible through a small object argument in the context of locally finitely presentable category. This step is essential in general in the construction of spectra, and also takes place in the topos-theoretic approach of [6], though it is mostly left implicit. The reference for this is [3], we mostly follow there modulo some adaptations, and in combination with elements from [11].
Definition 3.1. Let U : A→B a local right adjoint. A morphism n : B → C is said to be diagonally universal if it is left orthogonal to morphisms in the range of U, that is, if for any morphism u : A1 → A2 in A and any square as below, there exists a unique filler d : C → U(A1) making both the upper and lower triangles to commute
f BU(A1)
n d U(u)
CUg (A2) As a left class in an orthogonality structure, diagonally universal morphisms enjoy the following properties, which are standards and then do not need proofs here. Proposition 3.2. We have the following: − diagonally universal morphisms are stable under composition and contain isomorphisms − if n : B → C is diagonally universal and m : C → D is such that mn is diagonally universal, then m is also diagonally universal. In particular, diagonally universal morphisms are stable under retracts. − diagonally universal morphisms are stable under pushout along arbitrary morphisms − diagonally universal morphisms are stable under colimits and retracts in the arrow category B2 Remark. Beware that without further assumption, a diagonally universal morphism with codomain in the range of U, that is of the form n : B → U(A), needs not to be a candidate, as the filler needs not to arise from a morphism in A. Definition 3.3. A functor U : A→B is said to be relatively full and faithful if for any triangle as below f U(A1) U(A2)
U(u1) U(u2) U(A) then f comes uniquely from some u : A1 → A2 such that U(u)= f.
19 Proposition 3.4. For any u : A1 → A2, U(u) is diagonally universal if and only if U(u) is an isomorphism. If moreover U is relatively full and faithful, then U(u) is diagonally universal if and only if u is an isomorphism. Proof. The unique filler of the square
U(A1) U(A1)
U(u) d U(u)
U(A2) U(A2) is both a right and left inverse to U(u); and if moreover U is relatively full and faithful, it comes from a unique morphism d = U(v) which is both a section of u from the lower triangle, but it is also a retraction because there is a unique morphism in A lifting U(u) in the upper triangle, and this must be u.
Proposition 3.5. A morphism n : B → U(A) is diagonally universal if and only if UALA(n) is an isomorphism. Proof. The unique lifter d in the following square
A ηn BU(An)
n d UALA(n) U(A) U(A) is a section of UALA(n). This provide also a filler of the square
A ηn BU(An)
A ηn dUALA(n) UALA(n)
U(An) U(A) UALA(n)
A But 1U(A) = U(1A) is the only filler of the square because ηn is a candidate. So d is also a retraction of UALA, which is hence an isomorphism. Conversely, if UALA(n) has an inverse, then A one can use the candidate property at ηf to get a filler in any square with a morphism in the range of U on the right.
Remark. Beware that U needs not be conservative, so that the inverse of UALA(n) needs not comes from a morphism in A making LA(f) an isomorphism itself, so that the remark above does not A says that n ≃ ηn in B ↓ U; in particular n : B → U(A) may be diagonally universal without being a candidate. However, in case where U is relatively full and faithful, the filler we had above must come from a unique morphism d = U(v) which must satisfies the same commutations, hence provides an inverse of LA(n): hence the following corollary. Corollary 3.6. Suppose that U is relatively full and faithful; then for a morphism n : B → U(A), the following are equivalent: − n is diagonally universal − n is a candidate
− LA(n) is an isomorphism We defined Diag as the left orthogonal ⊥U(A2). Hence we end with an orthogonality strucutre (Diag, Diag⊥) where Diag⊥ is the double-orthogonal (⊥U(A2))⊥. Arrows in Diag⊥ lies now out of the essential image of U and may have arbitrary domain and codomain. However, we have the following fullness property of the essential image of U relatively to Diag⊥:
⊥ Proposition 3.7. Let be u : U(A1) → U(A2) be an arrow in Diag . Then u ≃ U(LA2 (u)) in A2 B/U(A2) and ηu is an isomorphism. In particular u is an arrow in the essential image of U.
20 Proof. Indeed, u is right orthogonal to the local unit in its stable factorization, so there exists a unique w as below
U(A1) U(A1)
A2 ηu w u
U(Au) U(A2) U(LA2 (u)) which is both diagonally universal by left cancellation, and in Diag⊥ by right cancellation, and is A2 hence an isomorphism. In particular ηu is an iso, being section of an iso. Now, we explain how, in a suitable context, the stable factorization of morphisms towards U extends to a factorization system in B, where the diagonally universal morphisms form the left class. To do so, we are going to adapt [3] version of the small object argument in the context of locally presentable categories.
We recall first some general properties of the left and right classes of a factorization system: Proposition 3.8. For a factorization system (L, R):
− L contains all isomorphisms and is closed − R contains all isomorphisms and is closed under composition, under composition,
− L is right-cancellative: if one has − R is left-cancellative: if one has
l f C1 C2 C1 C2
f r l′ r′ C3 C3
with l, l′ in L then f is also in L with r, r′ in R then f is also in R −→ −→ − L is closed under colimits in C − R is closed under limits in C
We also have the following constrains on the mutual factorizations of left and right maps: Lemma 3.9. If (L, R) is an orthogonality structure in C and we have a factorization as below with r in R and l in L Ar B
l f C Then l is a split monomorphism, and f factorizes through r. Dually, for any factorization as below
f A B r l C then r is a split epimorphism and f factorizes through l. Proof. Those statement just comes from the diagonalization of the squares
f A A A B
l d r l d r C B C C f
We also have the following usefull property: Lemma 3.10. Let be (L, R) an orthogonality structure in a category C with coequalizers: then if a parallel pair a,a′ : C ⇒ D in C is equalized by a morphism l : B → C in L, then its coequalizer ′ qa,a′ : D → coeq(a,a ) is in L.
21 Proof. Let be a square as below with r ∈ R
f D A
q(a,a′ ) r
′ ′ coeq(a,a ) g A
′ ′ Then we have fan = fa n and gq(a,a′)a = gq(a,a′)a , so that we have a commutative square
fan=fa′n B A n r a ′ ′ CD q ′ coeq(a,a ) g A a′ (a,a ) and both fa, fa′ provides diagonalization of this square: but such a diagonalization must be unique, so that fa = fa′. Hence there exists a unique d factorizing f through the coequalizer as below f D A
q(a,a′ ) d coeq(a,a′)
Moreover, we have udq(a,a′) = uf = gq(a,a′), but as a coequalizer, q(a,a′) is an epimorphism: thence necessarily ud = g so that d is the desired lift of the square above. Uniqueness of such a lift process from the uniqueness of the solution in the universal property of the coequalizer. Remark. Of course we have the dual statement saying that the equalizer of a parallel pair coequal- ized in R must be in R. Corollary 3.11. Any two arrows which is simultaneously equalized by a morphism in L and coequalized by a morphism in R must be equal. Proof. Suppose a,a′ : C ⇒ D are equalized by some l : B → C in L and coequalized by some ′ r : D → A in R. Then the coequalizer q(a,a′) : D → coeq(a,a ) is both in L and factorizes r as below a BCl Dr A a′ q(a,a′ ) f coeq(a,a′)
Then from lemma 3.9 we know q(a,a′) to be a split monomorphism; but a coequalizer being always ′ an epimorphism, this forces that actually q(a,a′) is an isomorphism, so that a = a . In the following we suppose that B is a locally finitely presentable category and U : A→B is a local right adjoint. Then denote D the class of diagonally universal morphisms between finitely presented objects. This coincides with the intersection of the class of diagonally universal 2 morphisms and the class of finitely presented morphisms, that is, D = Diag ∩Bω. We are going to use D to left-generate a factorization system, which will enjoy some degree of accessibility. Proposition 3.12. The class D has the following properties: − D is closed under composition and contains isomorphisms between finitely presented objects − D is right-cancellative − D is closed under pushouts along arbitrary finitely presented morphisms − D is closed under finite colimit in the arrow category B2 − Any filtered colimit in B2 of morphisms in D is diagonally universal. Proof. The two first propositions are obvious. The third is an easy consequence of the universal 2 property of the pushout. The fourth comes from the fact that Bω is itself closed under finite colimits 2 in B because Bω is so in B and colimits in the arrow category are sent to colimit by domain and codomain functors, while Diag is also closed under colimit as a left class in an orthogonality structure. This last argument also proves the last item.
22 Now we invoke results of [3] to construct a factorization system (Ind(D), D⊥). In our context, we can use the small class D of finitely presented diagonally universal morphism to left-generatea factorization system. We recall here the process:
2 Definition 3.13. A saturated class is a set V ⊆ Bω of finitely -presented maps such that: − V contains isomorphisms and is stable by composition, − V is right-cancellative − V is closed under finite colimits in B2 − V is closed under pushouts along arbitrary maps between finitely presented objects
2 A saturated class is always small, as lying in the essentially small generator Bω. In our case, the class D of finitely presented diagonally universal morphisms is a saturated class.
2 Proposition 3.14. Any set of finitely presented maps V ⊆Bω can be completed into a saturated class V such that V ⊥ = V⊥. Proof. It is clear that V⊥ ⊆ V ⊥ for V ⊆V. Moreover it is also easy to see that taking the closure by composition and iso does not modify the right class. Stability of the right class after closing the right class under finite colimit is a special case of stability of left class under colimits, but let ⊥ ′ 2 us give the detailed proof: let be a map l in V . If(li : Ki → Ki)i∈I is a diagram in B , then in the following diagram we have for each i ∈ I a lifting
qi u Ki colim Ki B i∈I
ki colim ki l i∈I di ′ ′ ′ Ki ′ colim Ki v B qi i∈I
′ and then by universal property of the colimit, this induces a unique map hdiii∈I with hdiii∈I qi = di. We prove this map is a filler. First, universal properties of colimits give us the following sequence of equalities: ′ ′ u = huqiii∈I = hdiqiliii∈I = hdiii∈I hqiliii∈I = hdiii∈I colim li i∈I so the upper triangle commutes; similarly we have
′ v = hvqiii∈I = hldiii∈I = lhdiii∈I so that the lower triangle commutes also. Finally, the stability of the right class after closing V under pushouts along arbitrary maps is an easy consequence of the universal property of the pushout.
Definition 3.15. Let be B in B: define DiagB the category of diagonally universal morphisms with domain B, and DB ֒→ DiagB the full subcategory whose objects are arrows n : B → C such that there exists some finitely presented diagonally universal morphism k : K → K′ in D and a : K → B such that we can exhibit n as a pushout
KKk ′
a k∗a p Bn C
Remark. By left cancellation, DiagB is itself a full subcategory of the coslice B ↓B.
Proposition 3.16. The category DB is closed under finite colimits in B ↓B.
Proof. Let be (ni,ai)i∈I with li ′ Ki Ki
ai B
23 a finite diagram in DB. We can use the fact that D is closed under finite colimit to compute the finite colimit of F as li ′ Ki Ki ′ qi qi
′ ai colim Ki colim Ki colim li i∈I i∈I i∈I
haiii∈I B ′ where (qi, qi): ni → colimi∈I ni is the inclusion in the colimit computed in the category of arrows and colimi∈I li is still in D. Then by commutation of pushouts with colimits we have
haiii∈I ∗colim li = colim ai∗li i∈I i∈I and this map is still in DB.
As a consequence, for any f : B → C, the category DB ↓ f of diagonally universal morphisms under B above f is filtered. Moreover, recall that the codomain functor B ↓ B preserves filtered colimits. Now we can construct the factorization of any arrow f in B: Proposition 3.17. For any f : B → C in B we have a factorization
f B C
colim DB↓f rf colim C DB ↓f
⊥ with lf in Ind(D) and rf is in D . For a complete proof of the statement, see [3][section 2.3] and in particular [theorem 14]. Proposition 3.18. The factorization above is orthogonal, that is, Ind(D)=⊥ (D⊥). Proof. First recall that diagonally universal are closed under (filtered) colimits. For l ∈⊥ (D⊥) with factorization f = rf lf , f is left orthogonal to its own right part, therefore there is a unique filler in the diagram below lf B Cf lf
Cf f rf
rf C C But now left cancellation of left maps, together with cancellation of right maps, enforces that this filler is an isomorphism, which forces rf to be iso, so that f is in Ind(D). Remark. At this point we should also give some details about the right classes, in particular Diag⊥ under the conditions of proposition 4.5. We have Diag⊥ = (⊥U(A2))⊥, exhibiting it as the closure of U(A2) under the axioms of the right classes. But from the specific assumptions the closure operations simplifies: − by functoriality U(A2) is allready stable under composition − it is also left-cancellative for U is relatively full and faithful − and closed in B2 under connected limits for A2 has connected limits and U preserves them. Therefore one has to add: − all the isomorphisms of B, comprising those between objects out of the essential image of U; in fact it is sufficient to add the identities
− in particular one must add the terminal map 11, the identity of 1, in order to have all limits in B2
24 − and add all pullbacks along arbitrary morphisms of B as below
f ∗U(A) U(A) y f ∗U(u) U(u) BU(A′) f
− finally one must also add retract in B2 of maps in U(A2) as below
B B f g r U(A) r
U(u) B′ B′
f ′ g′ U(A′)
Remark. Hence we have for any local right adjoint U a factorization system (Ind(D), D⊥). Observe that in the general case, the local units under a given object may not be obtained as filtered colimit of finitely presented diagonaly universal morphisms under them. If one successively take the stable factorization and the (Ind(D), D⊥)-factorization
f BU(A)
colim n A D ↓ηA ηf U(L (f)) B f A
colim CU(Af ) A uηA DB ↓ηf f then uf = U(LA(f))u A ηf is in Ind(D)⊥ by uniqueness of the factorization because Diag⊥ ⊆ Ind(D)⊥, so that the right part A of ηf is also the right part of f, so that we know that the functor
A DB ↓ ηf → DB ↓ f
⊥ is cofinal since it induces the same colimit. Remark also that u A is in Diag ∩ Ind(D) , which ηf however does not forces it however to be an isomorphism.
Moreover, this situation cannot even be improved in the case of a right multi-adjoint, where the local units under a given object form a small set. This is why, as we shall see in the second part, we have to impose explicitly the condition that local units are filtered colimits of finitely presented diagonally universal morphisms above them amongst conditions isolated by Diers to enable the construction of a spectra from a right multi-adjoint. Definition 3.19. A local right adjoint functor U is said to be diagonally axiomatisable if we have Diag = Ind(D). It is said to satisfy Diers condition if for any B and any f : B → U(A), the local A unit ηf is in Ind(DiagB). 2 2 Recall that B also is locally finitely presentable, with Bω as generator of finitely presented objects. We have now a functor preserving finite colimits
ιD 2 D ֒→Bω which extends into pair of adjoint functors
∗ ιD B2 ⊥ Ind(D)
ιD∗ where Ind(D) itself is locally finitely presentable. This gives rise to an idempotent comonad, one ∗ could see as returning the left part of a factorization system. In fact, the adjunction ιD ⊣ ιD∗ ∗ defines a morphism of locally finitely presentable categories, for the left adjoint ιD restricts to finitely presented objects, as morphisms in D are in particular finitely presented in B2.
25 Remark. It is well known that a left adjoint between locally finitely presented categories sends finitely presented object to finitely presented objects if and only if its right adjoint preserves filtered colimits. In our case, this says that the functor ιD∗, which returns the left part of the 2 factorization, preserves filtered colimits. This means that for any filtered diagram f(−) : I → B ′ 2 of arrows fi : Bi → Bi in B , the left part of the filtered colimit is the filtered colimit of the left parts
lcolim fi ≃ colim lfi i∈I i∈I 2 In particular, if we apply this to the canonical diagram of an arrow f ≃ colim Bω ↓ f, we have
lf ≃ colim lk 2 Bω↓f
′ 2 where the colimit is indexed by all the (k,a,a ) in Bω ↓ f. Hence this proves that arrows of the form lk for k ∈ K form a generator in Ind(D). Moreover, objects in Ind(D) are then models of a finite limit theory, which motivates the following definition: Definition 3.20. A diagonally universal morphism is said to be axiomatisable if it lies in Ind(D). Remark. Observe however that in general one cannot force arbitrary diagonally universal mor- phisms to be decomposable as a filtered colimit of morphisms in D. That is, we only have Ind(D) ⊆ Diag (and hence Diag⊥ ⊆ Ind(D)⊥) in the general case. For this reason, with the inclusion of right class a morphism between orthogonality structures, the factorization system (Ind(D), Ind(D)⊥) is the free left generated factorization system associated to the orthogonality structure (Diag, Diag⊥). Remark. Diers condition says that local units coincide with the diagonally axiomatisable morphism in the induced factorization. This is a strictly weaker condition than being diagonally axiomatisable as it does not requires any diagonally universal morphism to be axiomatisable. However in some situation we can produce local right adjoints with the desired property if we start from a left generated factorization system and a class of objects enjoying an adequate “gliding” condition: Definition 3.21. Let be a functor U : A→B and R a class of maps in B. We say that U lifts R maps if for any A in A and any r : B → U(A), there exists u : A0 → A and an isomorphism α : U(A0) ≃ B such that rα = U(u). Let be V a saturated class in a locally finitely presentable category B and (L, R) the associated ⊥ left generated system, with L = Ind(V) and R = V . Now suppose that U0 : A0 →B is a functor lifting R-maps. Then define ι0 : A ֒→ A0 as the wide subcategory whose arrows are those whose image under U0 are in R and U as the restriction U = U0ι0.
Proposition 3.22. Suppose that U0 is relatively full and faithful; then the induced functor U : A→B is stable and diagonally axiomatisable. Proof. For any f : B → U(A), consider the axiomatisable factorization
f BU(A)
lf rf
Cf where lf is obtained as the filtered colimit lf = colimDB ↓ f in B ↓B. For U0 lifts along R-maps, there exists uf : Af → A in A sent by U to rf , and moreover, this morphism is essentially unique in A as U0 is relatively full and faithful. But then lf is the local unit of U, or equivalently, is a candidate for U for it is diagonally universal with its image in the range of U by corollary 3.6. There is also a converse property: Proposition 3.23. Let U : A→B a diagonally axiomatizable right multi-adjoint: then U lifts its local maps.
26 Proof. For any local map of the form r : C → U(A), any precomposition with a diagonally universal morphism with codomain C as below
BCUl r (A) must also coincides up to iso with its own factorization, for in the square below
A ηrl BU(Arl)
l ULA(rl)
CUr (A) we have both a diagonal C → U(Arl) induced from the fact that l is diagonally universal, and a A diagonal U(Arl) → C from the fact that ηrl is diagonally universal and r is a local map, hence is in Diag⊥ by diagonal axiomatizability. It is easy to see that those maps are mutual inverse. A Applying the result before with l = 1C, we get an isomorphism ηr : C ≃ U(Ar), and we have A r = ULA(r)ηr . To complete this section, we should give a word on the way conversely, a factorization system together with a convenient class of objects induces a stable functor. A typical example of non full multireflection is the following as pointed out in [22] - where we recall that a class of maps closed under composition can be seen as a subcategory with is bijective on objects: Proposition 3.24. For any factorization system (L, R) (resp. left generated factorization sys- .(tem), the inclusion R ֒→B is a relatively full and faithful stable functor (resp. right multi-adjoint
Conversely any stable functor (resp. right multi-adjoint) which is surjective on objects and faithful is the inclusion of a right class (resp. of a right class in a left generated factorisation system).
Proof. If we have a factorization system, any morphism f : B → ιR(A) factorizes uniquely since ιR(r)l and l is orthogonal to any morphism in R so we can see it as the desired candidate.
If U : A→B is stable, faithful and surjective on objects, any object is some U(A) and any f : B → U(A) factorizes through a candidate which is left orthogonal to the morphisms in the range of U, hence the class of candidates constitutes the left part and the morphisms in U(A2) the right part: this is a factorization on the whole category as U is surjective on objects. The following terminology was suggested by Anel, and was also identified in [11][part 4] amongst condition to produce a spectral construction: Definition 3.25. Let be R a class of maps in a category and A a class of objects. We say that A has the gliding property relatively to R if for any arrow l : B → A in R with A a n object of A, then B must also be in A. Theorem 3.26. Let B be a category equiped with a factorization system (L, R) and A be a class of objects of B with the gliding condition relative to R. Then the inclusion AR ֒→B of A objects equiped with arrows of R between them defines a relatively full and faithful stable functor.
Proof. This follows from the previous proposition: for any B in B and any arrow f : B → A with nf uf A an object in A, as there exists a unique factorization B → Af → A with nf in E and uf in R, then Af is also an object in A; this factorization is initial amongst those through a morphism in R on the right. Moreover nf is a morphism in L with an objects in A as codomain, and such arrows are exactly the candidates for the inclusion as for any square with A0, A1, A2 objects in A, we have the diagonalization
B A1
n d u
A0 A2 u′ But recall that R is left-cancellative as any right class, so that u must itself be in R. Hence in the factorization above nf is a candidate and the inclusion is stable; left-cancellativity of R also enforces that this inclusion is relatively full and faithful.
27 4 Locally multi-presentable categories and accessible right multi-adjoint
Locally multipresentables categories were introduced by Diers in [15] (under the name “cat´egories localisantes”) as an a generalization of locally presentable categories encompassing a wide class of non-locally presentable categories, as local rings, fields, integral domains, local lattices... They are defined in the same way as locally presentable categories, but in the language of multi-colimits. For their tight relation with multiadjoitness and their reccuring role as the categories of local objects in spectral situations, we recall Diers theory of locally multipresentable categories, also present in [7].
First of all, recall the dual notion of multi-limit as introduced in section 1. Definition 4.1. A category A is said to be locally finitely multipresentable if − it has filtered colimits − it has a small generator of finitely presented object such that any object A decomposes as the filtered colimit A = colim Aω ↓ A − it is multicocomplete As well as locally finitely presentable can be characterized as the finitely accessible categories that are moreover either complete or cocomplete, locally finitely multipresentable categories are also characterized Proposition 4.2. Locally finitely multipresentable categories are exactly the finitely accessible categories with connected limits, where finite connected limits commutes with filtered colimits. Remark. Recall that a category is complete if and only if it has connected limits and a terminal object. Then a locally finitely multipresentable category is a locally finitely presentable category if and only if it has a terminal object. Remark. If A is locally finitely multipresentable, then: − the arrow category A2 also is locally finitely multipresentable − for any object A in A the coslice A ↓ A also is finitely multipresentable and moreover, the codomain functor cod : A ↓A→A is finitary and right multiadjoint.
2 Proof. For the first item, observe that the category Aω is a generator of finitely presented objects. 2 ′ Now multicolimits in A are computed as follows: for I a finite category and (fi : Ai → Ai)i∈I 2 ′ ′ ′ a I-indexed diagram in A , the multicolimit (gij : Ai → Bj )i∈I,j∈J of the codomains induces a ′ multicocone (gij fi : Ai → Bj )i∈I,j∈J , which defines in each j ∈ J a cocone over the (Ai)i∈I , and if we chose a multicolimit (gik : Ai → Bk)i∈I,k∈K , then for each i ∈ I and j ∈ J there is a unique ′ k ∈ K such that the cocone (gij fi)i∈I factorizes uniquely through gik
fi ′ Ai Ai
′ gik gij ′ Bk Bj fj
′ 2 and the family (f j : Bk → Bj )j∈J is a multicolimit in A . Then in particular for Aω is closed 2 under finite multicolimits, Aω is so.
For the second item, we refer to [8][Proposition 8.4] concerning existence of multicolimits. However we emphasize the following: for any object A′ in A, the canonical cone of local unit ′ relatively to the codomain functor A ↓A→A is made of the inclusions (qi : A → Bi)i∈I into a multicoproduct of A′ with A. In particular, as we shall see in proposition 4.5, this allows us to exhibits the generator of finitely presented objects (A ↓ A)ω as consisting of members of the multicoproducts of A with finitely presented objects of A. For the sake of completeness, we recall here elements of (finite) Diers duality, which is the multi-version of (finite) Gabriel-Ulmer duality:
28 Proposition 4.3 (Diers duality). Let A be a locally finitely multipresentable category. Then op op Aω is finitely multicomplete and we have an equivalence of category A≃Mlex[Aω , Set] where Mlex denotes the 2-category category of small finitely multicomplete categories, small finitely mul- ticomplete functors and natural transformation. Conversely for any small multicomplete category op C, Mlex[C, Set] is locally finitely multipresentable and C is equivalent to its small generator of finitely presented objects. Another analog to locally presentable categories is the relation with adjointness. First, there is a well known variant of the adjoint functor theorem saying that a functor between locally presentable categories have a left adjoint if and only if it is accessible and preserves limits. Here we provide an analogous statement for right multiadjoint with a locally multipresented domain: Theorem 4.4. Let be U : A→B a functor with A and B locally finitely multipresentable categories. Then U is a right multiadjoint if and only if it is accessible and preserves connected limits. Remark. Beware that we cannot control the rank of accessibility of U, which will end up λ-accessible for some λ ≥ ℵ0 we do not know, even though A and B are finitely accessible. In the following proof, we make use of the following general fact: for κ ≤ λ two cardinals, λ-filtered categories are in particular κ-filtered; hence if a category has κ-filtered colimits, it has in particular λ-filtered colimits, and similarly, a functor preserving κ-filtered colimits preserves in particular λ-filtered colimits. In particular, κ-presented objects are also λ-presented. Proof. For the indirect sense, it is known (see [18] and [21]) that accessible functors between acces- sible categories satisfy the solution set condition. As A and B are in particular (finitely) accessible and U is accessible, U satisfies the solution set condition, and if moreover it preserves connected limits, then by proposition 1.15 it is right multi-adjoint.
For the direct sense, we propose this adaptation of [1][Theorem 1.66] in the context of multi- adjointness and multipresentability. Suppose that U is right multiadjoint (and hence preserves connected limits by proposition 1.15). As Bω is small and A is accessible, we can chose a cardinal λ ≥ℵ0 such that for any K in Bω and any local unit nx : K → U(Ax) with x ∈ IK , the object Ax is λ-presented in A. Then we prove that U is λ-accessible as follows.
Let be F : I → A a λ-filtered diagram: in particular, I is finitely filtered, and the colimit (qi : F (i) → colim F )i∈I exists in A. We prove that (U(qi): F (i) → U(colim F ))i∈I is a colimit in B - and in particular, for it is λ-filtered, this colimit is also finitely filtered. By [1][Exercise 1.o(1)] we know that it is sufficient to check that − for any finitely presented object K in B and f : K → U(colim F ), there is some lift a : K → UF (i) of a along qi for some i ∈ I − and that for any two such lifts b : K → UF (i) and b′ : K → F (i′) there is a common refinement d : i → i′′ and d′ : i′ → i′′ such that UF (d)b = UF (d′)b′. For a : K → U(colim F ), consider the local factorization
KUa (colim F )
nx(a) U(ua) U(Ax(a))
As K is finitely presented, Ax(a) is λ-presented in A, and as I is λ-filtered, there is a lift
F (i) b qi A colim F x(a) ua whose image along U provide the desired lift by precomposing with the local unit nx(a)
KUa (colim F )
U(u ) U(q ) nx(a) a i U(A ) U(F (i)) x(a) U(b)
29 ′ ′ Now for two such lifts b : K → UF (i), b : K → F (i ) of a, as the image of the inclusions U(qi), ′ U(qi′ ) are in the range of U, we know by theorem 1.2 that up to canonical isomorphism, b and b factorize through the same local unit as a, that is x(b)= x(a)= x(b′) and we have the commutation below
b UF (i) U(qi) U(ub)
n KUx(a) (Ax(a)) U(ua) U(colim F )
U(ub′ ) U(qi′ ) ′ b UF (i′)
This returns in A the following commutation
F (i)
ub qi
Ax(a) ua colim F
q ′ ub′ i F (i′)
But now, for Ax(a) is λ-presented and I is λ-directed, the lifts ub and ub′ factorize through a ′′ ′ ′ ′′ ′ common refinement d : i → i and d : i → i such that F (d)ub = F (d )ub′ , which provides in B the desired common refinement for b and b′.
Then, we know that the cocone (U(qi): UF (i) → U(colim F ))i∈I is a colimit in B; this proves that U preserves λ-filtered colimits, that is, is λ-accessible. Remark. In proposition 4.5, we are going to prove that if we suppose that U is finitely accessible, then the localizations Ax for x ∈ IK and K finitely presented in B are finitely presented. In fact, in the proof above, λ was dependent of the rank of presentability of the localizations of finitely presented objects: if all the Ax for x ∈ IK and K in Bω happen to be finitely presented, then the functor U can be certified as finitely accessible. Recall that any locally finitely presentable functor is a right adjoint preserving filtered colimits; the last condition amounts to saying that its left adjoint sends finitely presented objects to finitely presented objects. Now recall from [1][Theorem 1.39] that a full, reflective subcategory A of a locally finitely presentable category B which is moreover closed under filtered colimits is locally finitely presentable - and the inclusion functor is locally finitely presentable; moreover, one could use as the generator of finitely presented objects in A the reflections of the finitely presented objects of B. In [8][Theorem 8.3.1] is given the following analog - we propose here a slightly different adaptation of.
Proposition 4.5. Let B be a locally finitely presentable category and U : A ֒→B such that − U is a right multiadjoint − U is moreover faithful, and relatively full and faithful, − A closed under filtered colimits and U preserves them.
Then A is a locally finitely multipresentable category. Moreover, the full, dense generator Aω of finitely presented object of A has as objects local reflections of finitely presented objects, that is
|Aω | = {Ax | nx : K → U(Ax) ∈ IK }
Ka∈Bω where IK is the set of local units of K for U. Proof. First, let us prove that A is finitely accessible. For it is supposed to have filtered colimits, we have to check that Aω is a small dense generator of finitely presented objects. Let be K in Bω and nx : K → U(Ax) a local unit under K. We first check that Ax is finitely presented in A:
30 let be a filtered diagram F : I → A, and an arrow a : Ax → colimF . Then for U is supposed to preserves filtered colimits and K is finitely presented, there is some i in I such that we have a lift
Kb UF (i)
n x U(qi)
U(Ax) colim UF U(a)
Then the local reflection of b provides us with a filler and morover we have a commutation
F (i)
LF (i)(b) qi
Ax a colim F
Hence Ax is finitely presented. Now we prove that for each A in A, we have a filtered colimit A ≃ Aω ↓ A. First, observe that there exist a canonical arrow in A as below
Ax a qa
A colim Aω ↓ A haia∈Aω ↓A induced from the fact that the diagram dom : Aω ↓ A → A is equipped with a canonical cocone of tip A. Moreover, for U preserves filtered colimit, this triangle is sent by U on a triangle
U(Ax) U(a) U(qa)
U(A) colim U(Aω ↓ A) hU(a)ia∈Aω ↓A
But in B, we have a filtered colimit U(A) ≃ colimBω ↓ U(A), where each b : K → U(A) is the inclusion at the index it defines in this canonical colimit. Moreover the local adjoint of U over A restricts to finitely presented objects as a functor
Bω ↓ U(A) → Aω ↓ A sending any b : K → U(A) to its local part LA(b): Ax(b) → A for the unique x(b) in IK corresponding to the local factorization of b. For each b the corresponding LA(b) is equipped with a canonical inclusion qLA(b) : Ax(b) → colimAω ↓ A and one has LA(b) = haiAω ↓AqLA(b). Then one has a factorization of the inclusion b
nx(b) KU(Ax(b)) U(q ) b LA(b) ULA(b)
U(A) colim U(Aω ↓ A) U(haiAω ↓A) which entails by the universal property of the colimit that we have a retraction
U(A) hU(q )n i LA(b) x(b) b∈Bω ↓U(A)
U(colim Aω ↓ A)
U(hai ) U(A) a∈Aω ↓A
Moreover, for U is relatively full and faithful, the map hU(qLA(b))nx(b)ib∈Bω↓U(A) must actually comes from a unique section uA in A
A ua
colim Aω ↓ A
hai A a∈Aω ↓A
31 On the other hand, the composites U(qLA(b))nx : K → colim U(Aω ↓ A) define altogether a universal map nx(b) KU(Ax(b)) U(q ) b LA(b)
U(A) colim U(Aω ↓ A) hU(q )n i LA(b) x(b) b∈Bω↓U(A)
But as we have hU(qLA(b))nx(b)ib∈Bω↓U(A) = U(uA) and for each b : K → U(A), nx(b) is a local unit, and the local factorization provides a diagonalization of this square
Ax(b) qL (b) LA(b) A
A colim A ↓ A uA ω
Moreover, for U is relatively full and faithful, we knwo that each UA is full and faithful, so that the counits of the local adjunctions are isomorphisms, and for any a : Ax → A in Aω ↓ A, we have a ≃ LA(UA(a)), and hence a = LA(UA(a)nx), and qa = qLA(UA(a)). Hence for any a ∈ Aω ↓ A we have a factorization of the inclusion qa given as
Ax
LA(U(a)nx) qa
A colim A ↓ A uA ω which entails by the universal property of the colimit that we have a retraction
colim Aω ↓ A haia∈Aω ↓A A
uA colim Aω ↓ A
Hence we have an isomorphism A ≃ colim Aω ↓ A as desired.
Now we prove that A has multicolimits. Let be F : I → A. Then, for B is locally finitely presentable, we can compute in B the colimit colim UF ; now consider the local units under colim UF , and observe that for each of those local unit x ∈ Icolim UF we have a cocone over UF
UF (d) UF (i) UF (j)
qi qj colim UF
nx
U(Ax) and we claim that each of those cocones comes uniquely from a cocone in A, which altogether form the multicolimit. For a given cocone (fi : F (i) → A)i∈I in A, take the universal map hU(fi)ii∈I : colim UF → U(A); for U is right multi-adjoint, this arrows factorizes uniquely through one of the local unit of index x(hU(fi)ii∈I ) ∈ Icolim UF . Then for each i ∈ I we have a factorization of U(fi) UF (i)
U(fi) qi
colim UFUhU(fi)ii∈I (A) n x(hU(fi)ii∈I ) UA(LA(hU(fi)ii∈I ))
U(Ax(hU(fi)ii∈I ))
x which forces the composite nx(hU(fi)ii∈I )qi to come from a unique arrow γi : F (i) → Ax(hU(fi)ii∈I ) in A for U is relatively full and faithful. Hence in particular each composite nxqi for i ∈ I and
32 x x ∈ Icolim UF comes from a unique γi : F (i) → Ax in A, and the family of cocones
x ((γi : F (i) → Ax)i∈I )x∈Icolim UF is exhibited as a multicolimit for F in A. Hence A is multicocomplete, and being finitely accessible, it is locally finitely multipresentable.
In particular observe that A has connected limits - and observe that U preserves them as a right multi-adjoint, so that they are computed as limits in B. In this context, we can in particular compute explicitly the factorization through local units associated to U by using the canonical cone of its domain: Corollary 4.6. With the hypothesis of proposition 4.5, for any f : B → U(A), we have
A A Af ≃ colim Afb ηf = colim ηfb LA(f)= hLA(fb)ib∈Bω↓B b∈Bω↓B b∈Bω ↓B
Proof. We have B ≃ colim Bω ↓ B, so that f = hfbib∈Bω↓B. Now for a given b : K → B, consider the factorization f KBUb (A)
A ηf Af η ULA(f) ηAb f U(A A ) U(Af ) ηf b A ULAf (ηf b) but by theorem 1.2 we know that actually
Af A A A Aη b ≃ Afb η A = ηfb LA(f)LAf (ηf b)= LA(fb) f ηf b
In [10], a specific process to construct locally finitely multipresentable categories is provided, which encompasses most of the interesting examples and enjoys Diers condition. Though this pro- cess is originally done inside an ambient locally multipresentable category, we consider here the special case of an ambient locally presentable category, where the construction simplifies slightly and meets the version of the small object argument presented above.
In the following, let B be a locally finitely presentable category. We consider a small class Γ of small cones in Bω, with VΓ = {gi | (gi : K → Ki)i∈I ∈ Γ,i ∈ I} the set of all arrows involved in cones of Γ.
Definition 4.7. An object A in B is said to be Γ-local if for any cone (gi : K → Ki)i∈I in Γ, we have a surjection
hB[gi,A]ii∈I B[Ki, A] B[K, A] i`∈I Remark. Observe that this notion is slightly weaker than the one considered in [8][Section 8.6], which we treat later under the name of strongly Γ-local. Our notion, beside being more general, is more suited to encompass the situations corresponding to geometric extensions of the finite limit theory behind B as we shall see below. We have also a notion of local morphism, which is actually the notion of right map relatively to the class of all maps involved in the cones of Γ; together with local objects they form a certain subcategory of B we are going to focus on:
⊥ Definition 4.8. A morphism is said to be Γ-local if it is in VΓ . We denote as BΓ the category of Γ-local objects and Γ-local morphisms between them, equiped with a faitfull, injective on object inclusion UΓ BΓ B
33 Remark. Observe that the category hence obtained is not a full subcategory: this allows in par- ticular to select a distinguished class of morphisms between models of an extension of the theory behind B, while considering only a finite limit extension of the theory of B in the same signature does not restrict morphisms between the models of the extension, producing always a full subcat- egory: this is because morphisms are determined by the signature and not by the axioms of the theory. We know then from proposition 3.17 that Γ-local morphisms are the right class of a left gener- ⊥ ated factorization system (Ind(VΓ), VΓ ) where VΓ is the saturated class generated from VΓ.
Then we have the following property, which is slightly more general than [10][Theorem 3.2] and whose proof simplifies also a bit thanks to the prior treatment of the small object argument:
Theorem 4.9. Let Γ be any small class of small cones in Bω. Then the inclusion UΓ is − accessible, − relatively full and faithful, − right multi-adjoint, − and satisfies Diers condition
In particular it exhibits BΓ as a locally finitely multipresentable category.
Proof. First observe that UΓ is faithful; moreover, it is easy to see it is relatively full and faithful as a consequence of left cancellability local maps enjoy as a right class.
We now prove that BΓ has filtered colimits and that they are preserved by UΓ.
Let be F : I → BΓ a small diagram with I filtered. Consider its filtered colimit colimUΓF in B. Now take a cone (gj : K → Kj )j∈J in Γ. Now for the joint surjectivity, consider an arrow a : K → colim UΓF ; as K is finitely presented, we have a lift b : K → F (i) for some i ∈ I, and by localness of F (i) there is some j in J and a factorization
Kb F (i)
gj q a i ¯b Kj colim UΓF and we have a = qib = qi¯bgj. Hence the localness of colim UΓF .
We must now prove the inclusions qi are Γ-local. Consider a square as below, for (gj : K → Kj)j∈I and j ∈ J: Ka F (i)
gj qi
Kj colim F (i) b i∈I
′ Then for Kj is finitely presented and I filtered, there is some i such that b factorizes through qi′ ′ ′ ′′ ′ ′ ′′ as b = qi′ b . But then qia = bgj = qi′ b gj, so that there exists some span s : i → i , s : i → i in I such that we have a factorization as follows
a K F (i) qi
F (s)
′′ gj F (i )qi′′ colim F (i) i∈I F (s′)
q ′ K F (i′) i j b
34 But then for gj is in VΓ and F (s) is Γ-local we have a unique diagonalization as below
a K F (i) qi
gj d F (s) ′ ′′ Kj F (i ) F (i )qi′′ colim F (i) b F (s′) i∈I where d provides also a diagonalization of the original square above.
Moreover such a diagonalization must always be unique: suppose indeed one has two parallel diagonalization Ka F (i) d gj qi d′ Kj colim F (i) b i∈I
′ ′ Then gj equalize d, d ; but we also have that qi coequalizes d, d , and from the fact that Kj is finitely presented and I is filtered, this implies there exists some arrow s : i → i′ in I such that ′ ⊥ ′ F (s) coequalizes d, d : but F (s) is Γ-local, that is, in VΓ , so by corollary 3.11, this forces d, d to be equal. This finish to prove that BΓ has filtered colimit and that UΓ preserves them, hence is accessible.
Now we have to prove UΓ is a right multi-adjoint. We consider as already constructed the fac- ⊥ torization system (Ind(VΓ), VΓ ). We prove that Γ-local objects have the gliding property relatively to local maps (see definition 3.25). Suppose that A is Γ-local and C is an object equipped with a local arrow u : C → A. Now for any cone (gi : K → Ki)i∈I in Γ and any f : K → C we have a factorization for some gi: f K C
gi u
Ki a A
Then for gi ⊥ u this induces a unique factorization Ki → C of f as desired. Then C is Γ-local.
Hence from theorem 3.26 we know that UΓ is stable, as BΓ consists precisely of a class of objects that have the gliding property to a right class. Hence for any arrow f : B → A with A a Γ-local ⊥ object, the (Ind(VΓ), VΓ )-factorization
f B A
nf uf
Af returns a local object Af . Moreover, as UΓ is relatively full and faithful, arrow n : B → A in Ind(VΓ) with A Γ-local are exactly the candidate for U : hence UΓ automatically satisfies Diers condition.
By theorem 1.27, UΓ is local right adjoint. But now, for UΓ is accessible, it satisfies the solution set condition: then by proposition 1.12, UΓ is a right multi-adjoint. As a consequence, BΓ is locally finitely multipresentable, from proposition 4.5.
Remark. Beware however that without additional assumption, UΓ may not be diagonally axiom- atizable; it may happens actually that the class of Γ-local object is empty, or at least does not ⊥ 2 contain enough objects to ensure that UΓ(BΓ) = Ind(VΓ).
Moreover, we also must restrict to the local maps between local objects in order to get a multireflective subcategory, for there is no reason for local objects to have the gliding property along arbitrary maps. Definition 4.10. An object A is said to be strongly Γ-local if it is Γ-local and satisfies moreover the condition that for any cone (gi : K → Ki)i∈I in Γ and any i ∈ I, one has an injection we have
35 in each i ∈ I an injection
B[gi,A] B[Ki, A] B[K, A]
strong strong Again one can define BΓ ֒→ B the subcategory of strongly UΓ : Γ-local objects and local maps between them. Then we have the analog of theorem 4.9, which is the original content of [10][Theorem 3.2]: strong Proposition 4.11. The inclusion UΓ is accessible, relatively full and faithful, and axiomatiz- strong able right multi-adjoint, and it exhibits BΓ as a locally finitely multipresentable category. Proof. The proof would be essentially the same as for theorem 4.9: we just have to control the additional injectivity condition.
Take a filtered diagram F : I →BΓ. We know that colimI F is local, and that the inclusion are local. We must prove that the colimit is strongly local. Suppose that, for some j ∈ J one has two ′ ′ ′ arrows a,a such that agj = a gj ; as Kj is finitely presented, there are respectively i,i in I and arrows b, b′ factorizing respectively a and a′ as below
F (i) F (i′) b ′ qi′ b qi a Ki colim UΓF a′
′ ′ But then observe that qibgj = agj = a gj = qi′ b gj, and as K is finitely presented and F is filtered, there is some i′′ in I and a cospan (d : i → i′′, d′ : i′ → i′′) in I such that we have a common factorization as below K ′ agj a gj
F (i) F (d) F (i′′) F (d′) F (i′)
qi qi′′ qi′
colimUΓF Now by localness of F (i′′), we know that F (d)b = F (d′)b′ for those arrows are equalized by precomposition with gj . But then we have
′ ′ ′ ′ a = qib = qi′′ F (d)b = qi′′ F (d )b = qi′ b = a which proves the injectivity of the applications B[gj, colim UΓF ].
strong Now it remains to prove that UΓ is a right multi-adjoint. Again we just have to prove that strongly Γ-local objects have the gliding property relatively to Γ-local maps. If one has u : C → A with u local and A a strongly Γ-local object, then we know from the last item of theorem 4.9 that C is Γ-local; now we must strengthen this condition. For any cone (gi : K → Ki)i∈I in Γ, suppose we have two parallel extension of a same f as below
f K C a gi a′ Ki
′ ′ Then one has lf = lagi = la gi, which enforces that la = la as A is strongly Γ-local: but then ′ ⊥ a,a are simultaneously equalized by gj which is in VΓ, and coequalized by u which is in VΓ , so that they are equal by corollary 3.11.
5 Co-stable functors from factorization systems
We saw in the previous section how stable functors towards a locally presentable category in- duced a factorization system through a small object argument. In this part we also make use of a factorization system, but for a different purpose: in the context of a category equipped with a factorization system and terminal object, we show how to construct a costable inclusion from a
36 class of “left objects’ (aka, objects with a left map as their terminal map). The motivation for this construction will be made clearer in a future work, where the bicategorical version of this process will be developed and applied in the bicategory of Grothendieck topoi relatively to several factorization systems, and provide a notion of “2-geometry” for Grothendieck topoi. However, as we chose to remain purely 1-categorical in this paper, we do not go into the 2-dimensional version of this construction and give it as an autonomous construction to prepare for its future involvement.
Throughout this section, we want to emphasize the “geometric” intuition leading our work. Hence, while the remaining of the paper followed “algebraic convention”, where we studied sta- bles functor in to categories of objects to be seen as algebraic (for instance, living in a locally presentable category), this section will produce co-stable functors rather than stable ones, and the object of the ambient category should be seen as spaces.
Before anything, let us precise that we call a functor F : A→C co-stable if the corresponding functor between the opposite categories F op : Aop →Bop is stable.
We fix a category C with a terminal object. This object should be seen as the generic point, as is the point ∗ in the category of topological spaces, or the topos of Sets in the bicategory of Grothendieck toposes, or the 2-elements lattice 2 seen as the trivial locale. Now we equip C with a factorization system (L, R).
Now the core idea of this section is that one can classify objects of C as left or right depending whether their terminal map is left or right; though arbitrary objects may not lie in either of those two classes, it is well known that right classes always define reflective subcategories by this process, which means that any object admits a “right replacement”.
Definition 5.1. Let C be a category equiped with a factorization system (L, R) and a terminal object. Let define
− the class of L-objects as those L such that L →!L 1 is in L; together with the L maps they form a subcategory LObj.
− the class of R-objects as those R such that R →!R 1 is in R; together with the R maps they form a subcategory RObj. Remark. Now observe the following: − 1 is, up to iso, the only object up to iso to be both left and right. − By right cancellation, any point 1 → R of a right object is a right map. − Any arrow toward a right object factorizes through a right object by right cancellability of the right class. − The category RObj of right objects and right maps is reflexive in C because of the factoriza- tion of terminal maps ! A A 1
r lA A R(A)
l Indeed the reflector A →A R(A) (which is a left map) is initial amongst those arrows toward f a right object: for any A → R, the statute of (lA, rA) as the terminal factorization of !A with a left map on the left induces the dashed arrow in the following:
f A R lA
lf rf R(A) R ∃∈L f !R !Rf rA 1
37 − In particular the right reflection of a left object is necessarily 1 by applying cancellation property of both classes to its terminal map. − Any object in C admits exactly one L-map into an R-object: its own reflection map, because post composing it with the terminal map of this R-object returns the factorization of the terminal map of C which is unique. − Both right and left objects possess a gliding condition along their associated map by stability by composition: in the following
f A B
!A !B 1
Then !A is either in R or L as soon as both f and !B are. However, as we are in the geometric side, this is not the condition from which we want to deduce stability. − Left objects have moreover the co-gliding condition relatively to the L-maps: in the following
Ll∈L A
!L !A 1
if L is a left object sending a left map toward an object A, then A has to be a left object by left cancellation. As a consequence, in the opposite side we have a stable inclusion:
LObjop ֒→Cop
However notify we cannot infer any smallness condition, so this has to be token as a local reflexivity. But this just forces us to admits the use of large spaces with a proper (at least a moderate) class of points. Most of those properties have a relativised version, that is, a corresponding statement where 1 has been replaced by an arbitrary base object B of C. Definition 5.2. For B in C, we define LObj[B] (resp. RObj[B]) as the categories whose objects are respectively left maps l : C → B ∈ L (resp. r : C → B ∈ R) and L-maps (resp R-maps) between them. Proposition 5.3. We have the same properties as over 1: − RObj[B] is a reflective subcategory of C/B where the free right object of a map f : C → B is rf : Cf → B and its unit is given by the left map lf : C → Cf ; − iso morphism B′ ≃ B are the unique object to be both right and left; − the category LObj[B] is co-stable in C/B because of the glidding condition induced by the left cancellation property of the left maps. Let (L, R) be a factorization system on C as above. Now we are going to see that we also can chose as left objects a subclass L′ ⊆ L behaving at least as a left class in an orthogonality structure, while not constituting necessarily the whole left class of a factorization system, plus some additional condition of closure by pullback and cancellability along L-maps. Definition 5.4. Let be C and (L, R) as above with a distinguished class L′ ⊆ L. We say that L′ has right L-cancellation property when for each triangle as below
Al∈L B
l′∈L′ f C then l being in L and l′ in L′ forces f to be in L′. In the following we fix a class of map L′ ⊆ L satisfying the right L-cancellation property which moreover we suppose stable by pullback and composition and containing all iso. In particular L′-maps are left orthogonal to R-maps.
38 Definition 5.5. We define the category L′ObjL as having L′-objects as objects and L-maps between them. Similarily, for any B in C, we define L′ObjL[B] as having L′-maps l : C → B for objects and triangles with L-maps between them as morphisms. Proposition 5.6. In a context above with L′ having right L-cancellation, we have a stable inclusion
L′ObjL)op ֒→Cop)
Indeed it is easy to see that we then have the co-gliding condition above any object by absorption of the terminal map of the intermediate object of any (L, R)-factorization: indeed in the situation below f L′ A r lf f Lf ′ ′ l ∈L lf C ′ ′ lf is in L as soon as l is. Definition 5.7. In the spirit of [3], we define as L′-form of an object C all R-maps r : L → C with a L′-object as domain. In particular any arrow between L′-forms is in R and should be seen as an inclusion of fondamental neighborhood, or equivalently, as generalized specialization order.
This means that any left form of A induces a point of R(A): to any rx : Lx→A we can associate px = R(rx):1 → R(A). Moreover, we could think this assignment as “continuous” in the sense ′ r that any R-neighborhood R → R of px induces a R-neighborhood of rx by pullback
Lx x
∗ ∗ lAr !Lx r A A y
1 lA px
R r R(A) However in general, it is clear by functoriality of R(−) that the reflection of an object A does not correctly distinguish local forms of A as it collapses the specialization order when this one is non trivial: for a triangle of R-maps
L1 r1 r r2 L2 A then R(r1) = R(r2r) = R(r2)!1 so that both define the same point of R(A). Intuitively, R col- lapses the local components and produces a disconnected set of points with trivial specialization order.
In this context, R-objects still play the role of “discrete” objects and L-objects the role of con- nected object, while L′-objects as local objects are to be seen as the “fundamental neighborhood” of the geometry, while the R-maps codes its etale domain.
The L′-maps from an arbitrary object toward an object will then be examples of “bundles of L′-objects” over a space, that is, bundles that are locally L′-objects in the sense that their fibers are L′-objects and L′-form of the domain of the bundle, from our key assumption that L-maps are closed by pullback: ∗ lC p∈R Lp C y ∗ p lC ∈L lc p 1 R
39 Definition 5.8. Define CSpaces as the category of “C-modelled spaces” whose − objects are arbitrary maps f : C → B,
♯ − and morphisms f1 → f2 are triples (g,a ,a♯) of the form
C1 a♯ a♯
∗ ∗ f2 g g C2 C2 y f1 ∗ g f2 f2 g B1 B2
♯ Remark. Actually, in the diagram above, a and a♯ are obviously mutually determined, but this emphasizes the analogy with modelled toposes as they correspond to the dual maps of the inverse op and direct comorphisms. Indeed, one would expect f1,f2 to code for sheaves of C objects over R-spaces with g a continuous map between them, along which one can either pull or push those ∗ ∗ sheaves. In this point of view, g f2 : g C2 → B1 is the inverse image of f2 over B1 and g ◦ f1 really is the direct image r∗f1 of f1 over B2. Now we turn to the objects we want to see as locally modelled objects: L Definition 5.9. For each object B, we define LocL′ [B] as the category whose: − objects are C → B that are stalkwise in L′ − maps are triangles that are stalkwise in L L L Then we define the category LocL′ as the wide subcategory of LocL′ [−] whose arrows are re- stricted to the (g,a♯) with a♯ stalkwise in L. Observe that in particularR this is a non full subcategory of C− spaces. Remark. Observe we have to require the comorphisms to be locally in L as requiring it globally would not be sufficient from the non stability along pullback; actually, they do not even need to be so globally. This mimics the idea of local morphism, which returns local maps at stalks, in the usual spectral construction. In the following diagram
l∈L C1 C2
f1 f2 C
Then, if f1 is in L-cancellation. If f1 locally is in L, then by right cancellation of pullback square we know that for any point p, p∗a♯ is in L so that one has over 1
∗ ∗ p l∈L ∗ p C1 p C2
∗ ∗ p f1∈L p f2 1 ∗ which forces each p f2 to be in L.
Proposition 5.10. One has a relatively full and faithful costable inclusion in each C
L LocL′ [C] ֒→C/C Proof. This just comes from the L-absorption property of L′ applied over 1 in each point of R. The L inclusion functor LocL′ [C] always is relatively full and faithful as indeed in the following situation
A2 l f
′ l′∈L l2∈L A1 A3
′ l1∈L l3 C
40 f is forced to be in L′ by the absorption property of L′ in L (this actually does not depends of ′ l1,l2,l3 being in L ).
References
[1] Jiˇri Adamek, J Adamek, J Rosicky, et al. Locally presentable and accessible categories. Vol. 189. Cambridge University Press, 1994. [2] Jiˇr´ıAd´amek and Jiˇr´ıRosick´y. How nice are free completions of categories? 2019. arXiv: 1806.02524 [math.CT]. [3] Mathieu Anel. “Grothendieck topologies from unique factorisation systems”. In: (2009). url: https://arxiv.org/abs/0902.1130. [4] Francis Borceux. Handbook of categorical algebra: volume 1, Basic category theory. Vol. 1. Cambridge University Press, 1994. [5] Julian Cole. “The bicategory of topoi and spectra”. In: Reprints in Theory and Applications of Categories (2016), pp. 1–16. url: http://www.tac.mta.ca/tac/reprints/articles/25/tr25abs.html. [6] Michel Coste. “Localisation, spectra and sheaf representation”. In: Applications of Sheaves. Ed. by Spinger. 1977, pp. 212–238. [7] Yves Diers. “Cat´egories localement multipr´esentables”. In: Archiv der Mathematik 34.1 (1980), pp. 344–356. [8] Yves Diers. “Cat´egories localisables”. PhD thesis. Paris 6 et Centre universitaire de Valen- ciennes et du Hainaut Cambr´esis, 1977. [9] Yves Diers. “Multimonads and multimonadic categories”. In: Journal of Pure and Applied Algebra 17.2 (1980), pp. 153–170. [10] Yves Diers. “Quelques constructions de cat´egories localement multipr´esentables”. In: Ann. SC. math. Qu´ebec 4.2 (1980), p. 79401. [11] Yves Diers. “Une construction universelle des spectres, topologies spectrales et faisceaux structuraux”. In: Communication in Algebra (1984). doi: 10.1080/00927878408823101. [12] Richard Garner. “Ionads”. In: Journal of Pure and Applied Algebra 216.8 (2012). Special Issue devoted to the International Conference in Category Theory ‘CT2010’, pp. 1734–1747. issn: 0022-4049. doi: https://doi.org/10.1016/j.jpaa.2012.02.013. url: http://www.sciencedirect.com/science/article/pii/S0022404912000527. [13] M Hebert, J Adamek, and Jiˇri Rosicky. “More on orthogonality in locally presentable cate- gories”. In: Cahiers de topologie et g´eom´etrie diff´erentielle cat´egoriques 42.1 (2001), pp. 51– 80. [14] Hongde Hu and Walter Tholen. “Limits in free coproduct completions”. In: Journal of pure and applied algebra 105.3 (1995), pp. 277–291. [15] C. Lair. “Sesqui-monades et monades locales”. fr. In: Diagrammes 9 (1983). talk:1. url: http://www.numdam.org/item/DIA_1983__9__A1_0. [16] Jacob Lurie. “Derived algebraic geometry V: Structured spaces”. In: arXiv preprint arXiv:0905.0459 (2009). [17] Kenji Maillard and Paul-Andr´eMellies. “A fibrational account of local states”. In: 2015 30th Annual ACM/IEEE Symposium on Logic in Computer Science. IEEE. 2015, pp. 402–413. [18] Michael Makkai and Robert Par´e. Accessible Categories: The Foundations of Categorical Model Theory: The Foundations of Categorical Model Theory. Vol. 104. American Mathe- matical Soc., 1989. [19] Axel Osmond. “On Diers theory of Spectra II: Geometries and Dualities”. In: (2020). [20] Axel Osmond. “The general construction of spectra”. In: (2020). [21] Jiˇri Rosicky and Walter Tholen. “Accessibility and the solution set condition”. In: Journal of Pure and Applied Algebra 98.2 (1995), pp. 189–208. [22] Paul Taylor. “The Trace Factorisation of Stable Functors”. In: (1998). url: http://www.paultaylor.eu/stable/trafsf.pdf. [23] Walter Tholen. “Pro-Categories and Multiadjoint Functors”. In: Canadian Journal of Math- ematics 36.1 (1984), pp. 144–155. doi: 10.4153/CJM-1984-010-2.
41 [24] Mark Weber. “Familial 2-functors and parametric right adjoints”. In: Theory Appl. Categ 18.22 (2007), pp. 665–732. [25] Mark Weber. “Generic morphisms, parametric representations and weakly Cartesian mon- ads”. In: Theory Appl. Categ 13.14 (2004), pp. 191–234.
42