topoi

These notes are from chapter IV of Sheaves in Geometry and Logic by Mac Lane and Moerdijk.

background

subobject functor

In a category $\mathcal C$, there may exist a subobject functor $\operatorname{Sub}(X)\colon \mathcal C^{\mathrm{op}} \to \mathrm{Set}$ which takes $X$ into the set of subobjects of $X$ (equivalence classes of monomorphisms into $X$). It forms a functor in the sense that for all $f\colon X \to Y$ in $\mathcal C$, there exists a function $\operatorname{Sub}(f)\colon \operatorname{Sub}(Y) \to \operatorname{Sub}(X)$.

A mono $m\colon S \rightarrowtail Y$ (an element of $\operatorname{Sub}(Y)$) can be pulled back along $f$ into a mono $m'\colon X \times_Y S \rightarrowtail X$ (an element of $\operatorname{Sub}(X)$) like this:

I think of this like representing the subobject that describes the largest part of $X$ that maps through $f$ into the image of $m$.

This is the first of three ways to write a subobject: as an equivalence class of monics $m\colon S \rightarrowtail X$.

subobject classifier

In a category $\mathcal C$, there may also exist a subobject classifier, or an object $\Omega$ and a mono $\mathrm{true}\colon 1 \rightarrowtail \Omega$ such that

for any mono $m\colon S \rightarrowtail X$ there exists a unique characteristic map or classifying map $\phi\colon X \to \Omega$ (also written $\operatorname{char} S$ or $\operatorname{char} m$) such that this square forms a pullback:

The characteristic map sends those parts of $X$ in the image of $m$ to the image of $\mathrm{true}$ in the subobject classifier.

This is the second way to write a subobject: as a characteristic map $\phi\colon X \to \Omega$ (a "predicate" of $X$).

By uniqueness this induces a bijection $\operatorname{Sub}(X) \cong \operatorname{Hom}(X, \Omega)$.

power object

In a category $\mathcal C$ with products and a subobject classifier, there may be a functor $P\colon\mathcal C^{\mathrm{op}} \to \mathcal C$ called the power functor sending objects $X$ to their power objects $P(X)$. These power objects function the same way as an exponential $\Omega^X$ would.

To express how this works on morphisms, we must also talk about how power objects each induce morphisms $\in_X\colon X \times P(X) \to \Omega$ such that for every $f\colon X \times Y \to \Omega$ there exists a unique $g\colon Y \to P(X)$ (the $P$-transpose of $f$) such that this triangle commutes:

Informally: $f$ is basically an uncurried $Y$-indexing of subobjects of $X$. By the product-hom adjunction, we know that $f$ has a corresponding $h\colon Y \to \operatorname{Hom}(X, \Omega)$. $\operatorname{Hom}(X, \Omega)$ can be viewed as isomorphic to $P(X)$.

This induces a bijection $\operatorname{Hom}(X \times Y, \Omega) \cong \operatorname{Hom}(Y, P(X))$.

Setting $Y$ to $1$ so $g$ describes a single subobject $s$ and viewing $f$ as a characteristic map $\phi$, the correspondence becomes clear:

So $\in_X$ essentially describes whether the first projection is in the image of the second projection (the subobject).

This is the third way to write a subobject: as a global element of a power object $s\colon 1 \to P(X)$.

morphisms? (not super important)

$P$ functions like a functor in the sense that for any $f\colon X \to Y$ there exists a $P(f)\colon P(Y) \to P(X)$ such that this diagram commutes:

definition

A(n elementary) topos is a category $\mathcal T$ in which there exist

The bijections described in previous sections can be bundled into a single one through the fact that $\Omega \simeq P(1)$:

$$ \operatorname{Sub}(X \times Y) \cong \operatorname{Hom}(X, P(Y)) $$

From these three concepts you can construct all exponentials $A^B$ (relatively easy) and all finite colimits (ungodly difficulty).

the "is singular" predicate

We can derive the "singleton arrow" $\{\cdot\}_X\colon X \to \mathcal P(X)$ as follows:

First, we can derive the diagonal morphism $\Delta_X\colon X \rightarrowtail X \times X$ as the universal morphism of this pullback (or product just as well):

Now, we can derive the characteristic map of $\Delta_X$ since it is monic, resulting in this pullback:

Taking the $\mathcal P$-transpose (currying) of that characteristic map gives us the map $\{\cdot\}_X\colon X \to \mathcal P(X)$.

But that map is obviously monic! So we can analyze the characteristic map of that transpose

This $\sigma_X\colon \mathcal P(X) \to \Omega$ tests if the given subobject is singular. This will be crucial for the next section.

deriving exponentials from power objects

graph of a function

In set theory, the graph of a function $f\colon X \to Y$ is defined as

$$ G_f = \{(x, f(x)) \mid x \in X\} $$

What we want to do is come up with a subobject of $X \times Y$ that describes graphs of functions. To do that, we need to find a way to test whether a subset of $X \times Y$ is a graph. This sounds like a characteristic map $\phi\colon X \times Y \to \Omega$.

Such a characteristic map (using $\mathrm{Set}$ terminology for the example) would function like, "we can just enumerate every $x \in X$ and check if the image of $x$ ($f[{x}] \in \mathcal P(Y)$) is a singular value $\{f(x)\}$".

the procedure

So first, for any particular subobject $\mathcal P(X \times Y)$, we need a map $v_s\colon X \to \mathcal P(Y)$ that takes the domain to subobjects in the codomain. We can merge these two into a single universal map $v\colon X \times \mathcal P(X \times Y) \to \mathcal P(Y)$. We can derive this by taking the $\mathcal P$-transpose of $\in_{X \times Y}\colon X \times Y \times \mathcal P(X \times Y) \to \Omega$.

Good. Step two is to test whether the image of that $\mathcal P(Y)$ we got is a singular element. We can do this by using that $\sigma_Y$ we derived earlier. So composing $\sigma_Y$ after $v$ gives us a $w\colon X \times \mathcal P(X \times Y) \to \Omega$ sending whatever in $X$ to $\mathrm{true}$ that has a singular image (e.g. in $\mathrm{Set}$ there is only one $(x, -)$).

Step three is to obtain all the $X$ (a subobject $\mathcal P(X)$) such that $w$ is sent to $\mathrm{true}$. We can do this using yet another $\mathcal P$-transpose on $w$, yielding us a map $u\colon \mathcal P(X \times Y) \to \mathcal P(X)$.

Now we want to define the subobject $Y^X$ containing all graphs. It is the subobject $\mathcal P(X \times Y)$ such that $u$ maps to the "trivial" subobject of $X$, which is $X$ itself.

Now we need to derive a characteristic map $\phi\colon X \to \Omega$ that sends everything to $\mathrm{true}$. There are several ways, but the simplest is to take the composition of the unique terminal map $!\colon X \to 1$ and then take $\mathrm{true}\colon 1 \to \Omega$. Then we can take the $\mathcal P$-transpose of that to get a global element $\mathrm{trivial}\colon 1 \to \mathcal P(X)$

Now we can define $Y^X$ as the pullback

Yay, let's celebrate!

But we're not done yet.

the evaluation map

For every exponential $Y^X$ there must be a corresponding evaluation map $\mathrm{ev}\colon X \times Y^X \to Y$. Thinking of this in terms of graphs, it's simply the singular element that the given $x \in X$ maps to. This calls for, you guessed it, another pullback

In the text there is an additional proof you have to make that this evaluation map has an "adjoint" relationship with products but I am too lazy to write this down.

monads and monadicity (extra stuff)

A monad is an endofunctor $T\colon \mathcal C \to \mathcal C$ with natural transformations $\eta_X\colon X \to T(X)$ (unit) and $\mu_X\colon T(T(X)) \to T(X)$ (multiplication) such that these diagrams commute:

It's a monoid in the category of endofunctors.

adjunctions and monads

All adjunctions $F \vdash G$ induce a monad $G \circ F\colon \mathcal C \to \mathcal C$ with the unit $\eta_C: C \to G(F(C))$.

Multiplication is a bit harder to derive, but we're looking for a $\mu_C\colon G(F(G(F(C)))) \to G(F(C))$ given an $\epsilon_D: F(G(D)) \to D$.

Setting $D$ to $F(C)$ we get $\epsilon_{F(C)}: F(G(F(C))) \to F(C)$ and we can lift that morphism through $G$ to yield the multiplication $\mu_C = G(\epsilon_{F(C)})\colon G(F(G(F(C)))) \to G(F(C))$

another way to look at the power functor

The power functor $\mathcal P\colon \mathcal T^{\mathrm{op}} \to \mathcal T$ has a left adjoint, the opposite functor $\mathcal P^{\mathrm{op}}\colon \mathcal T \to \mathcal T^{\mathrm{op}}$, meaning there are natural transformations $\eta_X: X \to \mathcal P(\mathcal P^{\mathrm{op}}(X))$ (unit) and $\epsilon_X: \mathcal P^{\mathrm{op}}(\mathcal P(X)) \to X$ (counit) satisfying the triangle inequalities.

deriving the initial object