The goal of this page is to prove that a non-extensional choice operator is conservative over a theory without AC. I’ll take the theory without AC to be SEAR, for definiteness and since that’s where I'm most comfortable.

For the theory with a non-extensional choice operator, consider the following theory, which is a variant of the version of SEAR without fundamental equality described at SEAR. There are four basic sorts: *pre-sets*, *pre-relations*, *elements*, and *operations* (or *pre-functions*). Each element belongs to a pre-set, and each pre-relation and operation has a source and target which are pre-sets. For a prerelation $\phi:A\looparrowright B$ and $x\in A$, $y\in B$, we have the assertion $\phi(x,y)$, and for an operation $f:A\rightsquigarrow B$ and $x\in A$ we have an element $f(x)\in B$. There is no predefined notion of equality of anything. Axioms 0 and 1 of SEAR are unmodified, except that the uniqueness clause of the latter is interpreted as a *definition* of when two parallel pre-relations are called “equal”. We also impose

If $\varphi:A\looparrowright B$ is a pre-relation such that for every $x\in A$ there exists a $y\in B$ with $\varphi(x,y)$, then there exists an operation $\varepsilon_\varphi :A\rightsquigarrow B$ such that $\varphi(x,\varepsilon_\varphi(x))$ for all $x\in A$.

A *set* is defined to be a pre-set $A$ equipped with an equivalence pre-relation $=_A$ which satisfies the principle of substitution and the identity of indiscernibles. A *relation* between sets is a pre-relation which is *extensional*, i.e. if $\varphi(x,y)$, $x'=_A x$, and $y'=_B y$, then $\varphi(x',y')$. Likewise, a *function* $f:A\to B$ between sets is an operation $f:A\rightsquigarrow B$ which is extensional, i.e. if $x' =_A x$ then $f(x')=_B f(x)$. We define two functions $f,g:A\to B$ to be *equal* if $f(x)=_B g(x)$ for all $x\in A$. Note that we have no notion of equality for arbitrary operations between pre-sets.

For any operation $f:A\rightsquigarrow B$ between *sets*, we have a pre-relation (its *graph*) defined by $\Gamma_f(x,y) \Leftrightarrow (f(x)=_B y)$, which is *entire* in the sense that for any $x\in A$, there is a $y\in B$ with $\Gamma_f(x,y)$. This pre-relation is extensional in $y$, but not in $x$ unless $f$ is actually a function (in which case $\Gamma_f$ is a *functional relation* in the usual sense). Conversely, Axiom $1+\varepsilon$ says that any entire pre-relation induces an operation, and for an entire and functional relation between sets, this induced operation is a function (and is unique, in the sense of equality of functions defined above).

Now Axiom 2 can be taken to read: For any sets $A,B$ and any relation $\varphi:A\looparrowright B$, there exists a pre-set $|\varphi|$ and operations $p:{|\varphi|}\rightsquigarrow A$ and $q:{|\varphi|}\rightsquigarrow B$ such that $\varphi(x,y)$ iff there exists $z\in {|\varphi|}$ with $p(z)=_A x$ and $q(z)=_B y$. We can then define $z=_{|\varphi|} z'$ iff $p(z)=_A p(z')$ and $q(z)=_B q(z')$ to make $|\varphi|$ into a set and $p,q$ into functions that are jointly injective. Axioms 3, 4, and 5 are easy to translate.

We call this theory $SEAR+\varepsilon$. Clearly the sets, elements (with defined equality), and relations in any model of $SEAR+\varepsilon$ satisfy the axioms of SEAR. Conversely, from any model of SEAR-C we can construct a model of $SEAR+\varepsilon$ by taking the pre-sets, pre-relations, and operations to be the SEAR-C sets, relations, and functions respectively. With this interpretation axiom $1+\varepsilon$ is precisely the SEAR-C axiom of choice. The question is whether we can get $SEAR+\varepsilon$ without assuming or implying choice.

Our goal is to prove that $SEAR+\varepsilon$ is conservative over SEAR. But since I haven’t quite figured out how to do that (or, to be fair, even whether it’s true), as a warm-up let’s prove that the same thing is true when we add COSHEP, a choice-like axiom notably weaker than full AC, to both sides.

Suppose we have a model of SEAR satisfying COSHEP, call it $V$. Recall that COSHEP means that in $V$, every set admits a surjection from a projective one. We define a model of $SEAR+\varepsilon$ as follows.

- The pre-sets are the projective $V$-sets.
- The elements are the $V$-elements.
- The pre-relations are the $V$-relations.
- The operations are the $V$-functions.

We will continue to qualify the notions in the old model $V$, using unqualified words such as “set” for the new notions defined above.

Axiom 0 of $SEAR+\varepsilon$ is obviously true. For Axiom 1, we need to translate a first-order formula in the language of $SEAR+\varepsilon$ into a formula in the language of SEAR in order to apply Axiom 1 of $V$, but the above dictionary tells us exactly how to do that. Axiom $1+\varepsilon$ is true because we have chosen the pre-sets to be the *projective* sets in $V$, so any entire $V$-relation between them contains a $V$-function.

Now a *set* $A$ in our putative model of $SEAR+\varepsilon$ is a $V$-set equipped with an equivalence relation $=_A$, so in particular it has a quotient $A/{=_A}$. Now any *relation* $A\looparrowright B$, meaning extensional relation, induces a $V$-relation $(A/{=_A}) \looparrowright (B/{=_B})$, and conversely any such $V$-relation induces a relation $A\looparrowright B$, setting up a meta-bijection. The same is true of functions $A\to B$ and $V$-functions $(A/{=_A}) \to (B/{=_B})$. It follows that we have meta-functors $Set\to Set_V$ and $Rel\to Rel_V$ given by “take quotients” that are fully faithful.

I claim that in fact these functors are essentially surjective as well. For given any $V$-set $A$, by COSHEP there is a projective set $P$ and a surjection $e:P\to A$. Since $Set_V$ is a regular category, it follows that $A$ is the quotient set of the kernel pair of $e$. But that means that if we call this kernel pair $=_P$, then $P$ becomes a new-style set which maps onto $A$ under the quotient functor; hence this functor is essentially surjective.

It follows that the meta-categories $Set$ and $Rel$ (in the new sense) are equivalent to the old meta-categories $Set_V$ and $Rel_V$. Since the remaining axioms of SEAR are essentially just statements about these categories, their truth in the “new world” follows immediately from their truth in $V$. (There’s no need to worry about a meta-axiom-of-choice, since these axioms are just statements about finite numbers of objects and morphisms, so “fully faithful and essentially surjective” is a perfectly sufficient notion of “equivalence.”)

Thus, we have constructed a model of $SEAR+\varepsilon$. Furthermore, its underlying SEAR-model is equivalent to $V$. (The notion of “equivalence” here means “(weak) equivalence of categories of sets,” as above. But since SEAR intrinsically aheres to the principle of equivalence, this suffices to show that exactly the same first-order statements are true in both.) It follows that the statements *in the language of SEAR* which are true in this new model of $SEAR+\varepsilon$ are precisely those which are true in $V$.

Therefore, any statement in the language of SEAR which is true in all models of SEAR that underlie models of $SEAR+\varepsilon$ is, in fact, true in *all* models of SEAR+COSHEP. This is the conservativity result we were looking for. Its practical upshot is that if you want to work in SEAR+COSHEP, you might as well work in $SEAR+\varepsilon$ if it’s more convenient to have a choice operator, since the theorems you can prove in either case will be exactly the same.

Note that the model of $SEAR+\varepsilon$ we’ve just constructed also satisfies COSHEP (as it must, given its equivalence to $V$). In fact, if a model of $SEAR+\varepsilon$ admits an “identity” equivalence pre-relation $\equiv_A$ on every pre-set, and relative to which all pre-relations and operations are extensional, then it must satisfy COSHEP—for then every set $(A,{=_A})$ is the surjective image of the set $(A,{\equiv_A})$, which is projective by Axiom $1+\varepsilon$.

Now I step in to say: $\mathbf{SEAR} + \varepsilon \vDash COSHEP$.

The reason is that every preset *does* admit an identity prerelation as in the last paragraph above; let $x \equiv_A y$ if $x \sim_R y$ for every reflexive prerelation $R: A \looparrowright A$ (or even for every equivalence prerelation). This will work also in $\mathbf{ETCS} - AC + \varepsilon$ by quantifying over prefunctions $f: A \to \mathcal{P}1$ and using the kernel of $f$ (relative to the standard equality on truth values) as the equivalence relation. (Of course, $\mathbf{SEAR}$ is capable of using this method too.) It will still work in versions with intuitionistic logic, but not (as far as I can see) in $\mathbf{CETCS} - COSHEP + \varepsilon$ (following Palmgren), where one cannot take power sets or quantify over subsets.

Mike Shulman: Ah, you’re right. I actually thought of that briefly, but then I didn’t immediately see how to prove the following.

Define $x \equiv_A y$ if $R(x,y)$ for every equivalence pre-relation $R: A \looparrowright A$. Then every operation and every pre-relation is extensional with respect to these identity relations.

Suppose $f:A\rightsquigarrow B$ is an operation. Define $K_f:A\looparrowright A$ by $K_f(x,x')$ iff $f(x)\equiv_B f(x')$. Then $R$ is an equivalence pre-relation, so if $x\equiv_A x'$, then $K_f(x,x')$, and hence $f(x)\equiv_B f(x')$.

Similarly, suppose $\varphi:A\looparrowright B$ is a pre-relation. Define $R_\varphi:A\looparrowright A$ by $R_\varphi(x,x')$ iff $\varphi(x,y)\Leftrightarrow\varphi(x',y)$ for all $y\in B$. Then $R_\varphi$ is an equivalence pre-relation, so if $x\equiv_A x'$, then $R_\varphi(x,x')$ and so $\varphi(x,y)\Leftrightarrow\varphi(x',y)$ for all $y\in B$.

Of course, as you point out, this is very impredicative. Does that mean that your original suggestion would only acceptable be for someone who either (1) accepts COSHEP or (2) doesn’t accept (quantification over) powersets?

*Toby*: That may be so; I hadn't thought through all of the implications, so I was hoping otherwise. However, I'm not sure that $\varepsilon$ has to work as in Axiom $1 + \varepsilon$; how about this?

Add a symbol (actually several symbols) to the language for an operation that assigns to each set $A$:

- a set $A'$,
- an injective function $i_A: A \to A'$,
- and an element $\epsilon_A$ of $A'$;

then add the axiom that $\epsilon_A$ belongs to the image of $i_A$ if (hence iff) $A$ is inhabited.

With excluded middle, this is a theorem with $A + 0^A$ for $A'$ as in the discussion at choice operator. But even this may be too strong; I really want to say that $\epsilon_A$ takes values in $A_\bot$ (the set of subsingleton subsets of $A$) with $\epsilon_A$ an inhabited subsingleton if (hence iff) $A$ is inhabited, but the existence of $A_\bot$ is itself impredicative (in the presence of function sets) since $1_\bot$ is the set of truth values.

Okay, second try. Define **constructive SEAR** to use intuitionistic logic and consist of Axioms 0, 1B, 2, and 4 of SEAR (no power sets) together with

- The existence of quotient sets (i.e. $Set$ is an exact category)
- The existence of disjoint unions (i.e. now $Set$ is a pretopos)
- The existence of dependent products (i.e. now $Set$ is a $\Pi$-pretopos).

Now modify it as above, removing basic equality and adding a basic notion of “operation,” to obtain **CSEAR+$\varepsilon$**. Note that the axiom of quotient sets becomes redundant, given the definition of “set” as a preset with an equivalence prerelation.

The same argument as above should show that CSEAR+$\varepsilon$ is conservative over CSEAR+COSHEP. (Just as Bounded SEAR is equivalent to ETCS, CSEAR+COSHEP is equivalent to Palmgren’s CETCS.) However, the converse argument above fails since we cannot define $\equiv_A$ by quantifying over relations (note that avoiding this requires not only throwing out powersets, but restricting axiom 1 to bounded quantification).

Now: is CSEAR+$\varepsilon$ conservative over CSEAR?

Last revised on July 16, 2023 at 14:06:17. See the history of this page for a list of all contributions to it.