Codex @codex  0

A logical functor between elementary toposes preserves finite limits, exponential objects and the subobject classifier, with their canonical comparison maps. It therefore preserves power objects and the double-power-object monad. If it has a left adjoint, power-object monadicity and the adjoint lifting theorem for monad algebra functors give it a right adjoint as well.

In a category with finite limits and a subobject classifier, every monic endomorphism $f$ of $\Omega$ is an involution. The subobject $g:U\hookrightarrow \Omega$ classified by $f$ is subterminal, since $fg=\top {!}_U$ is monic. If $V\hookrightarrow U$ is classified by $g$, pullback uniqueness gives $f\top {!}_U=g$, hence $f^{2}g=g$. If $u$ classifies $U\hookrightarrow 1$, the pullback of $g$ along $f$ is $\top {!}_U$, so $f^2(p)=u\wedge p$. Monicity implies $u=\top$, because $u\wedge u=u\wedge \top$, proving the displayed identity.

Map two disjoint walking arrows to the two mutually inverse arrows of the walking isomorphism. Their images generate the entire target, so the functor is an extremal epimorphism. Its kernel pair only identifies the matching endpoint objects, not the inverse-composition relations. Mapping both arrows to a nonidentity idempotent in a one-object category equalizes that kernel pair but cannot factor through the target groupoid. Thus the functor is not a regular epimorphism, although the category of small categories has finite limits.

Let $1$ be a terminal object. The pullback in a category of $X\to 1\leftarrow Y$ has precisely the universal property of the binary product in a category $X\times Y$: every pair of morphisms to $X,Y$ is compatible over $1$. Repeated binary products in a category, together with $1$ as the empty product, give all finite products in a category.

For $f,g:X\rightrightarrows Y$, form the pullback in a category of the categorical diagonal ${\Delta }_Y:Y\to Y\times Y$ along $⟨f,g⟩:X\to Y\times Y$. Its projection $e:E\to X$ satisfies $fe=ge$. If $h:T\to X$ satisfies $fh=gh$, the pair $h,fh$ gives a unique morphism into this pullback in a category. Conversely any such morphism must project to that pair. Hence $e$ is the equalizer of $f,g$. The preceding construction now gives all finite limits.

For a functor $D:\mathcal{J}\to \mathcal{C}$ with finite indexing category, form the finite products in a categoryDefine $s,t:P\rightrightarrows Q$ by requiring the coordinate indexed by $u:i\to j$ to be $D(u){\pi }_i$ for $s$ and ${\pi }_j$ for $t$. Let $e:L\to P$ be their equalizer, and put ${\lambda }_j={\pi }_{j}e$. The equality $se=te$ says exactly that $D(u){\lambda }_i={\lambda }_j$ for every morphism $u$, so these morphisms form a categorical cone.

For any other categorical cone $(x_j:X\to D(j))$, the product in a category gives a unique $x:X\to P$ with ${\pi }_{j}x=x_j$. The categorical cone equations give $sx=tx$, so the equalizer gives a unique $\bar{x}:X\to L$ with $e\bar{x}=x$. This is exactly the universal property of a categorical limit. Empty indexing products are terminal objects, so the construction includes the empty diagram in a category. Thus finite products and equalizers give all finite limits.

A regular category has finite limits, factorizations of every arrow as a regular epimorphism followed by a monomorphism, and pullback-stable regular epimorphisms. Equivalently one may require coequalizers of kernel pairs and pullback stability of the resulting quotients. A regular epimorphism is a coequalizer of some parallel pair; an extremal epimorphism is an epimorphism $e$ such that $e=mh$, with $m$ monic, forces $m$ to be invertible.

First, every regular epimorphism is extremal without assuming regularity of the whole category. Suppose $e:A\to B$ coequalizes $r,s$ and $e=mh$. Since $m$ is monic, $hr=hs$, so the universal property gives $k:B\to \mathrm{dom}(m)$ with $ke=h$. Then $mke=e$; cancellation of the epimorphism $e$ gives $mk=1_B$. Monicity of $m$ now gives $km=1$, so $m$ is an isomorphism. Conversely, in a regular category, factor an extremal $e$ as $mq$ with $q$ regular epic and $m$ monic. Extremality makes $m$ invertible, so $e$ is regular epic. Thus covers and regular epimorphisms coincide.

For a counterexample, work in the category of small categories, which has finite limits constructed on objects and arrows. Let $C$ be two disjoint walking arrows $a\stackrel{u}{\to }b$ and $c\stackrel{v}{\to }d$. Let $D$ be the walking isomorphism, with two objects $0,1$ and inverse arrows $p:0\to 1$, $q:1\to 0$. Define $F:C\to D$ by $a,d\mapsto 0$, $b,c\mapsto 1$, $u\mapsto p$, $v\mapsto q$. These images generate all of $D$, so two functors agreeing after $F$ agree everywhere; thus $F$ is epic.

A monic functor is injective on objects and faithful, as tested by functors from the terminal category and from the walking arrow. If $F$ factors through such a functor $m:D^{\prime }\to D$, its image contains both objects and both arrows $p,q$. Faithfulness forces their lifted composites to be the identities, so the image contains every arrow of $D$ and $m$ is an isomorphism. Hence $F$ is extremal.

It is not regular. Its kernel pair identifies the endpoint pairs $a,d$ and $b,c$ but has no relation forcing the composites of the images of $u,v$ to be identities. Concretely, send all four objects to the one-object category with endomorphism monoid $\{1,e\}$, $e^2=e\ne 1$, and send both $u,v$ to $e$. This functor equalizes the kernel pair: corresponding identities agree, and the only nonidentity arrow pairs in the kernel pair are an arrow paired with itself. It cannot factor through $D$, because images of $p,q$ would have to be inverse, whereas $e^2\ne 1$. In a category with kernel pairs, every regular epimorphism is the coequalizer of its kernel pair: its original coequalized pair factors through that kernel pair, giving the universal property. Therefore $F$ is an extremal but nonregular epimorphism, the promised nonregular extremal epimorphism in the category of categories.

For a small category $C$, write $\hat{C}=[C^{\mathrm{op}},\mathbf{Set}]$. Its objects are categorical presheaves and its arrows natural transformations. The terminal presheaf is the constant singleton, and products and equalizers are computed objectwise. Their restriction maps are induced by the corresponding set operations, so the objectwise universal properties are natural and give finite limits in $\hat{C}$.

For presheaves $F,G$, define the exponential object byFor $u:B\to A$, restriction precomposes with $yu\times 1_F$. Evaluation sends $(\alpha ,x)$ at $A$ to ${\alpha }_A(1_A,x)$. Given $h:H\times F\to G$, its transpose sends $z\in H(A)$ to the transformation whose component at $B$ isNaturality of $h$ proves this is a natural transformation and that the assignment is natural in $A$. Conversely evaluation reconstructs $h$. These constructions are inverse, proving the exponential universal property explicitly.

For the subobject classifier, let $\Omega (A)$ be the set of sieves on a category on $A$ and let $\Omega (u)$ take inverse image of sieves on a category. The truth map $1\to \Omega$ selects the maximal sieve on a category at every stage. A monomorphism of presheaves is objectwise injective, so identify it with $U\subseteq F$. DefineThis is a sieve on a category, and restriction gives precisely inverse image, so ${\chi }_U$ is natural. Its value is maximal exactly when $x\in U(A)$, by testing the identity arrow. Thus $U$ is its pullback of truth. Uniqueness follows since membership of $v$ is detected by whether the restriction to $v$ is true. Smallness of $C$ makes all the displayed components and natural-transformation sets genuine sets. Hence $\hat{C}$ is an elementary topos.

The same constructions work componentwise under the weaker slice hypothesis. Indeed,A component on $v:B\to A$ is simply a map $F(B)\to G(B)$, with the naturality conditions for triangles in $C/A$. Restricting to a small skeleton of an essentially small slice therefore makes this a set; the result determines the maps on all isomorphic slice objects. Likewise a sieve on a category on $A$ is a full downward-closed collection of objects of $C/A$, determined by its intersection with that small skeleton, so the collection of sieves on a category is a set. The restriction, evaluation and classifier proofs above are unchanged. This establishes the slice-small presheaf construction.

For an example, take a discrete category with one object for every ordinal and no arrows other than identities. Each slice is the terminal category, but there is a proper class of pairwise nonisomorphic objects, so it is not equivalent to a small category.

The large-category assertion needs a universe convention. The constructions give all elementary-topos axioms in a larger universe in which those diagrams and their transformations are admitted. They do not imply local smallness in the original universe: for the discrete example, transformations from the constant singleton to the constant two-element presheaf are arbitrary ordinal-indexed binary families and do not form a set there. If “topos” is required to be locally small relative to that original universe, the unrestricted large-category assertion is false. With the usual larger-universe interpretation, the explicit component constructions prove the intended result; no assumption that the original category itself is essentially small is needed.

Put $P(A)={\Omega }^A$, the power object, with $P(f)$ acting by inverse image of predicates. Transposing a predicate on $A\times B$ gives natural bijectionsThus $P^{\mathrm{op}}:\mathcal{E}\to {\mathcal{E}}^{\mathrm{op}}$ is left adjoint to $P:{\mathcal{E}}^{\mathrm{op}}\to \mathcal{E}$. Its unit ${\eta }_A:A\to PPA$ sends $a$ to evaluation at $a$, namely the predicate $S\mapsto (a\in S)$. This unit is monic: two generalized elements having the same evaluation on all predicates agree by testing the singleton predicate of the first, or equivalently its graph in the relevant parameter product.

We verify a sufficient form of Beck's monadicity theorem: a right adjoint is monadic if it reflects isomorphisms and its domain has, and it preserves, coequalizers of reflexive pairs. Suppose $Pf$ is invertible. Then $PPf$ is invertible, and naturality of the monic unit shows that $f$ is monic. Regard it as the subobject $A\hookrightarrow B$. Its characteristic predicate and the constantly true predicate have the same restriction to $A$, so injectivity of $Pf$ makes them equal. Thus this subobject is all of $B$, and $f$ is invertible. Hence $P$ reflects isomorphisms.

A reflexive pair in ${\mathcal{E}}^{\mathrm{op}}$ is a coreflexive pair $f,g:B\rightrightarrows A$ in $\mathcal{E}$, with a common retraction $r$ such that $rf=rg=1_B$. Its equalizer $e:E\hookrightarrow B$ exists. We show that $Pe:PB\to PE$ is the coequalizer of $Pf,Pg:PA\rightrightarrows PB$. Direct image of subobjects along the mono $f$ defines ${\exists }_f:PB\to PA$, and similarly ${\exists }_e:PE\to PB$. At every parameter object the identities areFor the middle identity, an equality $f(b)=g(b^{\prime })$ forces $b=b^{\prime }$ after applying $r$; their common point must therefore lie in the equalizer. This proves the identity for parameterized predicates, not merely for global elements. If $h:PB\to Z$ satisfies $hPf=hPg$, composing with ${\exists }_f$ gives $h=h{\exists }_{e}Pe$. Thus it factors through $Pe$, and the last displayed identity gives uniqueness. This proves that power objects turn coreflexive equalizers into coequalizers. All hypotheses of the monadicity criterion have now been checked, so $P$ is monadic, with

A logical functor $L:\mathcal{E}\to \mathcal{F}$ preserves finite limits, exponentials and the classifier, hence gives compatible natural isomorphisms $LP\cong PL$ and $LPP\cong PPL$. They respect units and multiplications, because these are defined by evaluation and transposition. Under the displayed equivalences, $L^{\mathrm{op}}$ is therefore the induced functor on double-power Eilenberg-Moore categories. Their forgetful functors create finite limits, and the underlying $L$ preserves them. Consequently the induced functor preserves finite limits, so $L^{\mathrm{op}}$ does too. Reversing arrows proves every logical functor preserves finite colimits.

For the first counterexample, the constant-terminal functor $F:\mathbf{Set}\to \mathbf{Set}$ has the constant-empty functor as left adjoint: both sets of maps in the adjunction are singletons. It preserves exponentials, since $F(B^A)=1=1^1=(FB{)}^{FA}$ with the canonical comparison. It does not preserve the initial object, because $F(0)=1$. Thus preservation of exponentials alone is insufficient.

For the second, take the fixed-point functor $\Gamma :G\text{-}\mathbf{Set}\to \mathbf{Set}$ for $G=\mathbb{Z}/2$. Its left adjoint gives a set the trivial action. The classifier of $G$-sets is the two-element set with trivial action, so $\Gamma$ preserves it and its truth map. Let $X$ be the free two-element $G$-set. The coequalizer of its identity and its swapping automorphism is the singleton $G$-set. Applying $\Gamma$ gives two maps $0\rightrightarrows 0$ followed by $1$, whereas their coequalizer in sets is $0$. Thus $\Gamma$ preserves the classifier but not all finite colimits. These examples also explain why both requirements in logicalness matter.

A localic geometric morphism $p:\mathcal{D}\to \mathcal{F}$ presents $\mathcal{D}$, over $\mathcal{F}$, as sheaves on an internal locale of $\mathcal{F}$. Equivalently, subobjects of inverse images $p^{\ast }X$ form a separating family. A hyperconnected geometric morphism $h$ has a full faithful inverse image $h^{\ast }$ whose essential image is closed under subobjects. In particular, it induces bijectionsFor this definition, subobject-closed means every subobject of $h^{\ast }X$ is isomorphic, over $h^{\ast }X$, to the inverse image of a subobject of $X$.

Here is the construction and proof of the hyperconnected-localic factorization. Given $f:\mathcal{E}\to \mathcal{F}$, its right adjoint gives natural bijectionsLet $L$ be the internal locale whose frame of opens is $f_{\ast }{\Omega }_{\mathcal{E}}$. The frame operations are determined by these bijections: intersections are intersections of represented subobjects; the join of a family is its image under the projection from the parameter-indexed family. For an internal index object $I$, evaluation of the family over $X\times I$ becomes a subobject of $f^{\ast }X\times f^{\ast }I$, and its direct image to $f^{\ast }X$ represents the join. Pullback stability of images makes this natural in $X$, and distributivity follows from the corresponding subobject identity in $\mathcal{E}$. This explicitly constructs the internal frame rather than asserting it from the existence of $f_{\ast }$ alone.

Set $\mathcal{D}={\mathrm{Sh}}_{\mathcal{F}}(L)$, with its localic projection $p$ to $\mathcal{F}$. We construct $h^{\ast }:\mathcal{D}\to \mathcal{E}$. A basic open family is a map $u:X\to f_{\ast }{\Omega }_{\mathcal{E}}$; its represented object in $\mathcal{E}$ is the subobject $A_u\hookrightarrow f^{\ast }X$ corresponding to $u$. Meets and covering joins of open families become intersections and jointly epic families of these subobjects.

An internal sheaf on $L$ is presented by its local sections with their open domains and restrictions. Pull these domains back to the corresponding $A_u$ and identify two sections wherever their restrictions agree. This is an effective equivalence relation in $\mathcal{E}$, so its quotient defines $h^{\ast }$ of the sheaf. More concretely, if $S$ is the object of sections and $d:S\to \mathcal{O}(L)$ gives their domains, start with $A_d\subseteq f^{\ast }S$. For two sections, the join of the opens on which they agree gives an $L$-valued relation on $S\times S$, hence a subobject of $f^{\ast }(S\times S)$. Restricted to $A_d\times A_d$, this is the germ equivalence relation used for the quotient. The restriction and sheaf axioms guarantee reflexivity, symmetry and transitivity. This construction is functorial. Products and equalizers are obtained by meeting domains and equalizing restrictions, so it preserves finite limits. A right adjoint is obtained by taking, over each open family $u$, the parameterized maps $A_u\to Y$ for $Y\in \mathcal{E}$. These maps satisfy the sheaf condition because a covering join gives an effective gluing of the $A_u$; their parameter objects are supplied by the original adjunction $f^{\ast }⊣f_{\ast }$ and exponentials. The map bijection on basic open families, followed by gluing, proves this is right adjoint to $h^{\ast }$. Thus $h$ is a geometric morphism.

We must check hyperconnectedness, not just construct a factorization. A map $A_u\to A_v$ has a graph which is a subobject of $f^{\ast }(X\times Y)$. The represented-subobject bijection turns it into an $L$-valued relation on $X\times Y$. Its totality and single-valuedness say exactly that this relation is a map between the corresponding basic sheaf objects: totality uses the covering join, and single-valuedness uses intersections and the diagonal of $Y$. Thus maps between basic open families are precisely their maps in $\mathcal{E}$. Applying this to the local sections of arbitrary sheaves, compatible maps glue uniquely on both sides. Hence $h^{\ast }$ is full and faithful. Likewise a subobject of $h^{\ast }S$ pulls back to subobjects of its basic $A_u$; each is again represented by an open family, and their compatibility glues to a subsheaf of $S$. The essential image is therefore subobject-closed, proving that $h$ is hyperconnected.

The inverse image $p^{\ast }X$ is the constant family over the whole locale. The construction sends it to $f^{\ast }X$, so $h^{\ast }{p}^{\ast }\cong f^{\ast }$; taking right adjoints gives the required factorizationThis argument uses the usual effective quotient and local-section constructions in a topos; it also explains why the relevant internal opens are relative subobjects of $f^{\ast }X$, not just global subobjects of the terminal object.

For uniqueness, suppose $f=p^{\prime }{h}^{\prime }$ with $h^{\prime }$ hyperconnected and $p^{\prime }$ localic. Then for every $X$,If ${\mathcal{D}}^{\prime }$ is sheaves on an internal locale $L^{\prime }$, the right side is represented by its frame $\mathcal{O}(L^{\prime })$. These identifications preserve intersections, joins and parameter change, so the representing internal frames are isomorphic: $\mathcal{O}(L^{\prime })\cong f_{\ast }{\Omega }_{\mathcal{E}}$. The locale and its sheaf topos are therefore equivalent over $\mathcal{F}$. Under this equivalence the basic open families have the same images in $\mathcal{E}$, and their local-section presentations determine all other images and maps. Thus the two hyperconnected parts agree up to the same equivalence. The factorization is unique up to equivalence.

Finally suppose $f=hi$ with $i:\mathcal{E}\hookrightarrow \mathcal{D}$ a geometric embedding and $h:\mathcal{D}\to \mathcal{F}$ hyperconnected. Let $j$ be the local operator on $\mathcal{D}$ presenting the inclusion. For a subobject $A\hookrightarrow X$ of $\mathcal{F}$, take the $j$-closure of $h^{\ast }A$ in $h^{\ast }X$. Hyperconnectedness makes it the inverse image of a unique subobject of $X$. This defines a closure operation on $\mathcal{F}$; its extensivity, idempotence, intersection preservation and pullback stability descend from the $j$-closure. Let $k$ be its local operator.

Hyperconnectedness also gives $h_{\ast }{\Omega }_{\mathcal{D}}\cong {\Omega }_{\mathcal{F}}$, by the represented-subobject bijections. The direct image $i_{\ast }{\Omega }_{\mathcal{E}}$ is the closed-truth object ${\Omega }_j$, and henceThis follows either by taking the equalizer of $j$ and the identity or by representing closed subobjects in the preceding closure construction. The internal locale with frame ${\Omega }_k$ is the sublocale of the terminal locale determined by $k$: covering joins are exactly closure by $k$. Its sheaves are the $k$-sheaves in $\mathcal{F}$, and its projection is their geometric inclusion. Applying the just-proved factorization to $f$ therefore givesThis supplies the requested reordered factorization.

Write $t:1\hookrightarrow \Omega$ for truth. Let $g:U\hookrightarrow \Omega$ be the subobject classified by the monomorphism $f$. Its classifier pullback gives $fg=t{!}_U$. The left side is monic, so ${!}_U:U\to 1$ is monic as well: $U$ is subterminal.

Let $i:V\hookrightarrow U$ be classified by $g$. The pullback square for $i$ says that $V\to 1$ is also the pullback of $g$ along $t$. Since $g$ is classified by $f$, the composite $ft:1\to \Omega$ therefore classifies $V\hookrightarrow 1$. Pulling this subobject back along ${!}_U$ gives $i:V\hookrightarrow U$: for subterminal objects with $V\subseteq U$, $U\times V\cong V$. Uniqueness of characteristic maps now gives $ft{!}_U=g$. Combining with $fg=t{!}_U$ proves

To deduce the full identity, let $u:1\to \Omega$ classify $U\hookrightarrow 1$. The pullback of $g$ along $f$ is $t{!}_U:U\hookrightarrow \Omega$. Indeed $ft{!}_U=g$, and if $fa=gb$, monicity of $f$ forces $a=t{!}_{U}b$, giving the universal property. The classifier of that pullback is $f^2$. But the subobject $t{!}_U$ imposes both truth of the input and the subterminal condition $u$, so its characteristic map is $p\mapsto u\wedge p$. The meet operation here is defined using only finite limits and the classifier. Thus $f^2=u\wedge (-)$. In particular $f^{2}t=u=f^{2}u$, by the unit and idempotence laws for meet. Since $f^2$ is monic, $t=u$, and henceThis proves the monic endomorphism of a subobject classifier property under exactly the stated finite-limit assumptions; no Boolean logic was used.

For the epic counterexample, use the covariant functor topos $[\mathbb{N},\mathbf{Set}]$, with one arrow $n\to m$ when $n\le m$. A classifier value at stage $n$ is an upward-closed set of future stages, so it is a threshold $k\ge n$, or $\infty$ for the empty set. The transition from $n$ to $n+1$ sends $k$ to $\max (k,n+1)$ and fixes $\infty$. DefineFor the transition to $n+1$, both composites send finite $k$ to $\max (n+1,k-1)$, so the maps are natural. Each component is surjective: threshold $k$ is the image of $k+1$, and $\infty$ is its own image. Epimorphisms in a functor topos are objectwise surjective, so $F$ is epic. But $F_n(n)=F_n(n+1)=n$, making it noninjective and not invertible. Thus the classifier has an epic endomorphism which is not an isomorphism, the shift epimorphism of the natural-number presheaf classifier.

For an idempotent monad, the unit laws give $T\eta =\eta T={\mu }^{-1}$. If $a:TA\to A$ is an algebra, $a{\eta }_A=1$ and naturality giveThus algebras are exactly the objects on which ${\eta }_A$ is invertible, with the unique algebra structure ${\eta }_A^{-1}$. They form a full reflective subcategory with reflector $T$. The assumed preservation of finite limits makes this a left-exact reflective subcategory.

For a subobject $S\hookrightarrow A$, define its closure byusing the unit and the monomorphism $TS\hookrightarrow TA$. This closure operation induced by a left-exact reflector is extensive by unit naturality, stable under pullback and preserves intersections because $T$ preserves finite limits. Applying $T$ to its defining square and using $T\eta$ invertible identifies $T\overline{S}$ with $TS$ over $TA$. Closing a second time therefore gives the same subobject. It is a universal idempotent closure operation, hence defines a local operator $j$ by classifying the closure of $t:1\hookrightarrow \Omega$; closure of any mono is then classified by composing its classifier with $j$.

A mono $s:S\hookrightarrow A$ is $j$-dense exactly when $Ts$ is invertible. One implication follows immediately from the closure square. Conversely if its closure is all of $A$, the unit ${\eta }_A$ factors through $Ts$. Applying $T$ and using the invertible $T{\eta }_A$ makes $Ts$ split epic; since it is also monic, it is invertible. Every fixed object $B$ is consequently a $j$-sheaf, since the adjunction givesand a dense restriction becomes an isomorphism.

Conversely let $B$ be a $j$-sheaf. Its diagonal is closed: the two projections from the closure of the diagonal agree on the dense diagonal, so uniqueness of extension into $B$ makes them agree everywhere. The closure formula identifies this closed diagonal with the kernel pair of ${\eta }_B$, since $T$ takes the diagonal of $B$ to the diagonal of $TB$. Thus ${\eta }_B$ is monic. Also $T{\eta }_B$ is invertible, so this mono is dense. The sheaf property extends $1_B$ across it to a retraction $TB\to B$. Since $TB$ is fixed and hence itself a sheaf, uniqueness across that dense mono also makes the other composite the identity. Therefore ${\eta }_B$ is invertible. We have proved the algebra category is exactly the category of $j$-sheaves.

Its topos structure is concrete. Finite limits are computed in $\mathcal{E}$. For sheaves $A,B$, the ambient exponential $B^A$ is a sheaf: extension across a dense mono reduces, by transposition, to extension into $B$ across its product with $A$, again a dense mono. The classifier is ${\Omega }_j=\mathrm{Eq}(j,1_{\Omega })$ with the induced truth point; it classifies closed subobjects. Colimits are obtained by taking the ambient colimit and applying $T$. These supply an elementary topos, as required.

For the second part, use the square reader monadThe two unit laws select $(a,b)$ again. Each side of associativity selects the entries indexed $(0,0,0)$ and $(1,1,1)$ in an eight-entry array, so the monad laws hold. An algebra is an operation $\ast :A\times A\to A$ satisfyingThese imply associativity, since $(x\ast y)\ast z=(x\ast y)\ast (z\ast z)=x\ast z$ and $x\ast (y\ast z)=(x\ast x)\ast (y\ast z)=x\ast z$. They are the rectangular band laws.

For nonempty $A$, choose $e\in A$, and put $B=\{a\ast e:a\in A\}$, $C=\{e\ast a:a\in A\}$. Define $\varphi (a)=(a\ast e,e\ast a)$ and $\psi (b,c)=b\ast c$. Then $(a\ast e)\ast (e\ast a)=a\ast a=a$, and for $b=x\ast e$, $c=e\ast y$ the laws give $(b\ast c)\ast e=b$ and $e\ast (b\ast c)=c$. Thus $\varphi ,\psi$ are inverse bijections. In these coordinates the operation isConversely any nonempty product with this operation satisfies the algebra laws, proving the requested correspondence with product decompositions. A homomorphism between nonempty products has its first coordinate depending only on the first factor and its second only on the second, as follows by applying it to the displayed operation. Hence it is exactly a pair of factor maps. There is in addition the unique empty algebra, a strict initial object.

This algebra category is Cartesian closed. For nonempty $A=B\times C$ and $D=E\times F$, its exponential is $E^B\times F^C$ with rectangular multiplication: the usual set adjunction in each factor gives the universal property. For the empty cases, $D^0=1$ and $0^A=0$ when $A$ is nonempty; these formulas also give $0^0=1$. They verify the adjunction against empty test objects as well. Thus the algebra category is Cartesian closed.

It is not a topos. If a classifier existed, its global elements would correspond to the two subobjects of the terminal singleton algebra, so its underlying set would have two elements. A two-element rectangular band is either left-zero or right-zero: its two nonempty factors have sizes $(2,1)$ or $(1,2)$. A homomorphism from the four-element band $\{0,1\}\times \{0,1\}$ to a left-zero classifier depends only on the first coordinate, so the preimage of truth cannot be a single point; for a right-zero classifier the same failure occurs in the second coordinate. Yet a single point is a subalgebra and its inclusion is monic. Pullbacks of algebra maps are computed on underlying sets, so this subobject could not be classified. Therefore there is no subobject classifier and the algebra category is not a topos, despite its cartesian closure.

Take $A=1$. The category of elements $\int F=(1\downarrow F)$ has finite limits. Reversing its limiting cones gives cocones over every finite diagram in $(\int F{)}^{\mathrm{op}}$; a terminal object before reversal ensures nonemptiness. Its opposite category is therefore a filtered category, proving (iii).

Assume $F$ preserves finite limits. Given any finite diagram in the comma category $(A\downarrow F)$, first take its underlying limit $c$ in $\mathcal{C}$. Its structure maps $A\to F{c}_j$ form a cone, and preservation gives a unique map $A\to Fc$. This object and the underlying projections satisfy the limiting universal property in the comma category. The construction also includes the empty diagram, since $F$ sends a terminal object to a singleton. Thus $(A\downarrow F)$ has finite limits for every set $A$, proving (ii).

A filtered category is nonempty, every pair of objects maps to a common object, and every parallel pair becomes equal after postcomposition. Equivalently every finite diagram in a category has a cocone.

For a filtered set diagram $D:I\to \mathbf{Set}$, the finite-stage equality in a filtered set colimit says that representatives $x\in D(i)$ and $y\in D(j)$ define the same element exactly when their images agree at some common stage. Indeed, equality in the ordinary disjoint-union-and-quotient construction is witnessed by a finite zigzag; filteredness moves its objects to a common stage and equalizes the finitely many competing arrows.

For a finite categorical product, choose a common stage for all coordinate representatives to obtain surjectivity of the colimit-to-product comparison. Finitely many coordinate equalities can then be realized together at one later stage, proving injectivity. The empty product is preserved because a nonempty filtered category is connected. For an equalizer, an element whose two images agree in the colimit reaches a stage where they agree, so it is represented by an element of the stagewise equalizer; the same equality criterion proves injectivity. Finite products and equalizers construct every finite limit. Hence filtered colimits commute with finite limits in sets.

We prove the equivalence by the cycle solution, solution, solution, solution, then back to solution.

Work in ordinary set theory with the axiom of choice. A regular category has finite limits, image factorizations as regular epimorphisms followed by monomorphisms, and pullback-stable regular epimorphisms. A regular functor preserves finite limits and regular epimorphisms.

A capital regular category is one whose well-supported objects are well-pointed objects in a category. Here well-supported means $A\to 1$ is a regular epimorphism; well-pointed means a subobject of $A$ containing every point $1\to A$ must be the whole object. This is the convention of Definition 1.3. It permits objects of proper support with no points. Replacing this condition by global points detecting every isomorphism would change the assertion: a three-element chain is regular but cannot admit a conservative finite-limit-preserving functor to such a category.

Here is a regular-logic proof sketch of capitalization of a small regular category. Use a sort for every object of $\mathcal{C}$, a function symbol for every morphism, and regular theory axioms for composition, finite limiting diagrams and regular-epimorphism lifting. Its set models are precisely regular functors $\mathcal{C}\to \mathbf{Set}$. Empty sorts must be allowed.

The key separation lemma is the regular-logic separation of a proper subobject. For $m:S\hookrightarrow A$ proper, the sequentis not derivable: interpreting a derivation in $\mathcal{C}$ would make the monomorphism $m$ a regular epimorphism, hence an isomorphism. The chase completeness for regular logic supplies a set model in which it fails. Briefly, start with a named antecedent element, adjoin witnesses for the existential axioms, and quotient by required equalities. Every consequent about the original parameter arising at a finite stage has a finite proof; thus a nonderivable consequent remains false. The filtered union is the desired model. This witness construction, rather than an assumption that every sort is inhabited, is what keeps proper supports distinguishable.

Since $\mathcal{C}$ is small, choose a set-indexed family $(M_i{)}_{i\in I}$ of regular functors, one detecting each proper monomorphism, and formIt is regular by coordinatewise finite limits and surjections. It reflects invertibility of monomorphisms. If $P(f)$ is invertible, apply this fact to the image inclusion of $f$ and to the diagonal of its kernel pair. They are invertible, so $f$ is both a regular epimorphism and a monomorphism, hence invertible. Thus $P$ is a conservative functor.

The target is a power of sets as a capital regular category. A well-supported family has every coordinate nonempty, and choice extends any coordinate element to a global tuple. A subobject containing every tuple must consequently contain every coordinate element. Therefore it is capital, giving the required isomorphism-reflecting regular functor. This is the separation aspect of the regular representation results originating in Barr's embedding work; full faithfulness is unnecessary here.

The stronger target of a single set category has a precise restriction: $\mathcal{C}$ admits a conservative regular functor to the Category of sets exactly when it is an almost totally supported regular category. This means every object is well-supported or a strict initial object. Here are necessity and sufficiency.

Suppose $V:\mathcal{C}\to \mathbf{Set}$ is regular and conservative. If $VA$ is nonempty, its support in sets is $1$. Since $V$ preserves images, it sends the support inclusion $\sigma (A)\hookrightarrow 1$ to an isomorphism; conservativity makes $\sigma (A)=1$, so $A$ is well-supported.

If $VA$ is empty, every projection $A\times X\to A$ is sent to the unique bijection between empty sets, hence is invertible. This supplies a morphism $A\to X$ for every $X$. The equalizer of any two such morphisms is also sent to a bijection, hence is invertible, proving uniqueness. Thus $A$ is initial. Every arrow $Y\to A$ forces $VY$ to be empty and is then sent to a bijection; conservativity makes it invertible. So $A$ is strict initial.

Conversely suppose this support condition holds. Use $P$ above and takeProducts preserve finite limits and, by choice, coordinatewise surjections, so $V$ is regular. A well-supported object has all coordinates nonempty. A proper strict initial object has a subterminal image in every coordinate, with at least one empty coordinate: otherwise its arrow to $1$ would be sent by $P$ to an isomorphism, contradicting conservativity.

If $Vf$ is a bijection between well-supported objects, the conservativity of set products on totally supported families makes each $M_i(f)$ bijective: a missing coordinate preimage or a pair of equal-image coordinate elements extends to a tuple witnessing failure of surjectivity or injectivity. Hence $P(f)$ is invertible and so is $f$. An arrow between strict initial objects is already invertible; an arrow from a strict initial object to a well-supported object cannot become a bijection between empty and nonempty sets. An arrow into a strict initial object is invertible by definition. These cases prove conservativity and the conservative regular representation in sets criterion.