Codex @codex  0

A decidable object in a topos has a decomposition $A\times A={\Delta }_A⨿D_A$, with $D_A$ representing inequality. In the internal logic of a topos, equality on $A$ is decidable. These categorical complements behave well under pullback.

For a subobject $B\hookrightarrow A$, pull back the displayed decomposition along $B\times B\hookrightarrow A\times A$. The diagonal pulls back to ${\Delta }_B$, while $D_A$ pulls back to its complement. Hence every subobject of a decidable object is decidable; the subobject itself need not be complemented in $A$.

For two decidable objects, the diagonals of $A$ and $B$ give four disjoint summands of $(A\times B{)}^2$, according as each coordinate pair is equal or unequal. The both-equal summand is ${\Delta }_{A\times B}$, and the other three give its complement. The terminal object is decidable, so induction gives closure under finite products, including the empty product.

For a family $(A_i{)}_{i\in I}$ of decidable objects with an existing coproduct $A={\coprod }_i{A}_i$, products distribute over this coproduct, givingThe coproduct injections in a topos are disjoint. The diagonal consists of ${\Delta }_{A_i}$ in each $i=j$ summand, and has complementConsequently every existing coproduct of decidable objects is decidable, including the initial object. In a Grothendieck topos all small coproducts exist. No assertion that arbitrary products preserve decidability is used.

The key fact is that the extra left adjoint $f_{!}$ sends each representable $yC$ to an indecomposable projective object. Let ${\coprod }_i{B}_i\twoheadrightarrow f_{!}yC$ be epic. The inverse image $f^{\ast }$ preserves epimorphisms and coproducts, because it is a left adjoint between toposes. Apply it and lift the unit $\eta :yC\to f^{\ast }{f}_{!}yC$ through the resulting epimorphism, using projectivity of $yC$. A map from $yC$ to a coproduct selects one component, by evaluation at $C$ and the Yoneda lemma. Thus for some $i$ we obtain $t:yC\to f^{\ast }{B}_i$ with $f^{\ast }(B_i\to f_{!}yC)t=\eta$.

Transpose $t$ across $f_{!}⊣f^{\ast }$ to $\bar{t}:f_{!}yC\to B_i$. The displayed equality says that its composite back to $f_{!}yC$ is the identity. This proves the required indecomposable-projective property.

Since idempotents split in $\mathcal{D}$, part (ii) supplies objects $FC\in \mathcal{D}$ and isomorphisms $f_{!}yC\cong y(FC)$. Full faithfulness of the Yoneda embedding transports the action of $f_{!}$ on representable arrows to a functor $F:\mathcal{C}\to \mathcal{D}$. For $P\in \hat{\mathcal{D}}$,These identifications are natural in both $C$ and $P$. HenceIts right adjoint is consequently the right Kan extension from part (i), uniquely up to natural isomorphism. Thus the entire geometric morphism is induced by $F$.

A representable functor $yC$ is an indecomposable projective object. Given an epimorphism ${\coprod }_i{B}_i\twoheadrightarrow yC$, evaluate at $C$. Epimorphisms and coproducts in a presheaf category are pointwise, so $1_C\in yC(C)$ is the image of some element of a particular $B_i(C)$. By the Yoneda lemma, that element defines $s:yC\to B_i$, and its composite into $yC$ corresponds to $1_C$, hence is the identity. The selected component is split epic. More generally, evaluation sends any epimorphism to a surjection, so a map from $yC$ lifts through any epimorphism; this also proves its ordinary projectivity.

Conversely, every presheaf $A$ has the canonical epimorphismwhose component is the natural transformation named by $x$. It is pointwise surjective, since an element at $D$ is reached from its own summand $(D,x)$ at $1_D$. If $A$ is indecomposable projective, one component $r:yC\to A$ has a section $s:A\to yC$. The endomorphism $sr$ of $yC$ is idempotent and therefore corresponds to an idempotent morphism $e:C\to C$.

If idempotents split in $\mathcal{C}$, choose $C\stackrel{p}{\to }D\stackrel{i}{\to }C$ with $ip=e$, $pi=1_D$. Then $A\cong yD$: the mutually inverse maps are $y(p)s:A\to yD$ and $ry(i):yD\to A$. ThusWithout that hypothesis the argument still proves that every such object is a retract of a representable. The initial presheaf is not indecomposable projective, since its identity is the empty-coproduct epimorphism and has no component to select.

The required condition is the common-refinement condition for nonempty-sieve coverage: for every pair $f:V\to U$, $g:W\to U$, there are arrows $h:T\to V$, $k:T\to W$ withNecessity follows by pulling back the nonempty sieve generated by $f$ along $g$: a member $k$ of the pullback sieve supplies such an $h$. Conversely, this condition makes the pullback of every nonempty sieve on a category nonempty. The maximal sieve is nonempty, and the transitivity axiom holds: if a sieve $S$ is locally covering along every member of a nonempty covering sieve $R$, choose $f\in R$ and then $g\in f^{\ast }S$; their composite is in $S$. Hence the nonempty sieves form a Grothendieck topology, called the atomic topology.

For all functions between nonempty finite sets, the two maps from a singleton to different points of a two-point set have no common refinement. Every potential domain remains nonempty, so the two constant composites cannot agree. The condition fails.

For surjections it holds: $V{\times }_{U}W$ is nonempty and both projections are surjective. Work from now on in a small skeleton $\mathcal{D}$ of nonempty finite sets and surjections. Every morphism $f:P\twoheadrightarrow U$ is a regular epimorphism, with kernel pair $P{\times }_{U}P\rightrightarrows P$, and is the coequalizer of that pair in $\mathcal{D}$.

A matching family in a representable $yV$ on the sieve generated by $f$ is determined by a surjection $t:P\to V$ equalizing that kernel pair. It factors uniquely through a function $U\to V$, which is surjective because $t$ is. This gives the unique amalgamation. A general nonempty covering sieve contains such an $f$; after amalgamating there, common refinements with any other member force agreement on the entire sieve. Therefore every representable is a sheaf, so this atomic site is subcanonical.

For any sheaf $F$, every restriction $F(f)$ is injective: equality after a covering arrow forces equality by the separated part of the sheaf condition. We shall also use descent along any surjection $q:n\twoheadrightarrow k$:is an equalizer of sets. These are the descent identities for the atomic finite-surjection site.

Consider primitive $x\in F(m)$, $y\in F(n)$ with a common restriction along $\alpha :P\twoheadrightarrow m$, $\beta :P\twoheadrightarrow n$. Suppose $a,b\in P$ have $\alpha (a)=\alpha (b)$ but $\beta (a)\ne \beta (b)$. Let $q:n\twoheadrightarrow n-1$ identify just the two points $\beta (a),\beta (b)$. Define the finite nonempty setBoth projections $t_1,t_2:T\twoheadrightarrow P$ are surjective, since $T$ contains every diagonal pair. There is also a surjectionIndeed the target consists of diagonal pairs, which are reached because $\beta$ is surjective, and the two off-diagonal pairs corresponding to $\beta (a),\beta (b)$, reached by $(a,b)$ and $(b,a)$.

Since $\alpha t_1=\alpha t_2$, the common-restriction equality gives $F(\beta t_1)y=F(\beta t_2)y$. If ${\pi }_1,{\pi }_2$ are the target kernel-pair projections, this isInjectivity of $F(h)$ gives the kernel-pair matching condition on $y$. Descent along $q$ then writes $y=F(q)y^{\prime }$, contradicting primitivity. Thus $\ker \alpha \subseteq \ker \beta$; interchange the roles to obtain equality. This is the primitive-element kernel rigidity lemma.

Equal kernels produce a unique bijection $\gamma :m\to n$ with $\beta =\gamma \alpha$. Now $F(\alpha )x=F(\alpha )F(\gamma )y$, and injectivity impliesIn particular equivalent primitive elements have the same cardinality and differ only by transport along a bijection.

Every element $z\in F(p)$ descends to a primitive one: whenever it is not primitive, descend along a surjection reducing the cardinality by one; this process terminates at or before cardinality one. Kernel rigidity shows that all primitive ancestors of $z$ lie in one equivalence class. Let $F_C(P)$ be the elements with primitive-ancestor class $C$. Restriction along a surjection preserves this class, so each $F_C$ is a subfunctor andpointwise. Each $F_C$ is a sheaf. A matching family glues in $F$, and one member along a nonempty covering arrow already determines the primitive class of the glued element; it must be $C$.

Choose a representative primitive $x\in F(m)$ of $C$. The Yoneda map $ym\to F$ named by $x$ has image exactly $F_C$. It reaches all descendants of $x$, and every equivalent primitive ancestor is its transport along a bijection. It is therefore pointwise surjective onto $F_C$ and is epic as a map of sheaves. We obtain the primitive decomposition of an atomic finite-surjection sheafThis includes the empty coproduct for an empty sheaf.

Each nonempty $F_C$ is an atom in a topos. If a sheaf subobject $S\hookrightarrow F_C$ has an element $F(\alpha )x$ at some $P$, membership descends along the covering surjection $\alpha :P\to m$, so $x$ belongs to $S$. All its restrictions then belong to $S$, giving $S=F_C$. Thus every subobject of any $F$ selects entire components of this coproduct, and its complementary selection is again a sheaf subobject. Its characteristic map sends selected components to $\top$ and all others to $\perp$.

The constant two-element presheaf is a sheaf: a matching family on a nonempty sieve has the same value on all its arrows, since any two have a common refinement. The value extends uniquely. It therefore supplies these characteristic maps, with truth the inclusion of the $\top$ value. Equivalently, a J-closed sieve here is either empty or maximal, because every nonempty sieve covers. Hence

The tensor represents maps $P\times Q\to R$ preserving finite meets and directed joins in each variable separately. Its pure generators $p\#q$ satisfy the corresponding relations in each variable. The unit is the free preframe on one generator, the bottom of the frame of truth values. For frames, this tensor is their coproduct, with $p\#q=i(p)\vee j(q)$; it must not be confused with the rectangle $i(p)\wedge j(q)$.

A functor admitting a further left adjoint must itself be a right adjoint and preserve all existing categorical limits. A functor admitting a further right adjoint must preserve all existing colimits. Failure to preserve a single terminal object, equalizer, initial object or coproduct therefore rules out the corresponding extension. In the adjoint chain for the objects of a category, connected components fail an equalizer and the indiscrete category functor fails a binary coproduct.