 Solution

Work inside a model $V$ of ZF. Let $\pi :V\to V$ be a definable class bijection, with definable inverse, so that its image and inverse image of every set are sets by Axiom schema of replacement. Define new membership byThis is the Rieger-Bernays permutation model convention used here. Its basic advantage is that every prescribed old set $b$ is the new extension of the object ${\pi }^{-1}(b)$. We verify the axioms rather than treating a membership permutation as automatically harmless.

The new axiom of extensionality follows because equality of the new extensions means $\pi (x)=\pi (y)$, hence $x=y$. The new empty set and unordered pair are ${\pi }^{-1}(\varnothing )$ and ${\pi }^{-1}(\{x,y\})$. Union and power set are represented byIndeed, the elements of the first object are exactly the new elements of new elements of $a$; the elements of the second are exactly the objects whose new extensions are subsets of $\pi (a)$.

Translate an arbitrary formula recursively by replacing each atomic $xEy$ by $x\in \pi (y)$. This is an ordinary definable formula in the old model. Old axiom schema of separation forms any specified subclass $b$ of $\pi (a)$, and ${\pi }^{-1}(b)$ supplies the new separated set. Old Axiom schema of replacement similarly collects the uniquely specified images of all elements of $\pi (a)$, and applying ${\pi }^{-1}$ supplies its new representative. This verifies the full schemas, including formulas with unbounded quantifiers.

For axiom of infinity, define $e_0={\pi }^{-1}(\varnothing )$ andThen $\pi (e_n)=\{e_0,\dots ,e_{n-1}\}$. These objects are distinct by induction: equality of $e_n$ with an earlier $e_m$ would equate old sets of respectively $n$ and $m$ distinct members. Old replacement collects $S=\{e_n:n\in \omega \}$, and ${\pi }^{-1}(S)$ is a new inductive set. Thus the construction satisfies every axiom of ZF except possibly Axiom of foundation.

Now interchange the old objects $\varnothing$ and $\{\varnothing \}$ and fix everything else. For $q=\varnothing$ we have $\pi (q)=\{q\}$, so its new sole member is itself: $q=\{q{\}}_E$. This Quine atom contradicts Axiom of foundation. The identity permutation preserves the original model and its foundation. Hence, relative to consistency of ZF, foundation is independent of the other ZF axioms.

For the further independence result, a full permutation of a model of choice is insufficient. If the old model satisfies axiom of choice, then for any new family $a$ of nonempty new sets, old choice selects $g(x)\in \pi (x)$ for every $x\in \pi (a)$. Encode its graph using the new ordered pairs, and represent that graph by ${\pi }^{-1}$. This is a new choice function. We therefore add a hereditary finite-support restriction to a universe with many Quine atoms.

Start now with a model of ZFC. Put $a_n=\omega +n$, and let $\pi$ interchange $a_n$ and $\{a_n\}$ for every $n\in \omega$, fixing other objects. All these transpositions are disjoint: the infinite ordinals $a_n$ are distinct and none is a singleton. In the resulting membership structure, $A=\{a_n:n\in \omega \}$ is an infinite family of Quine atoms.

Build its universe $W$ by the internal transfinite hierarchyWrite $W$ for the union over the old ordinals. Apart from the atomic self-loops, an object's new members have lower construction ranks. Every permutation $\sigma$ of $A$ therefore extends recursively to $W$: on atoms use the prescribed permutation, and on other objects putThis is well defined and preserves $E$. A nonatom cannot be sent to an atom, since an extension consisting solely of one atom already represents that atom. Recursing with ${\sigma }^{-1}$ gives the inverse, so these maps are genuine automorphisms.

Say that $x$ has finite support $S\subseteq A$ if every such permutation fixing $S$ pointwise fixes $x$. Let $H$ consist of the atoms and the objects with finite support all of whose new members, recursively away from atomic self-loops, belong to $H$. This is the hereditarily finite-supported Quine-atom model. We check why it still has the required set-theoretic axioms.

It is closed under taking new members, so extensionality remains valid. The empty object and the pure finite ordinals have empty support. Their internally represented infinite collection also has empty support, giving infinity. Pairs and unions use the finite unions of their constituents' supports; all their members remain hereditary members of $H$.

For a set $a\in H$, take all objects $b\in H$ whose new extensions are subsets of $\pi (a)$. They form an old set, since the unrestricted new power set is an old set and membership in $H$ is definable. Its representative has support contained in a support of $a$, because an automorphism fixing $a$ permutes precisely these hereditary subsets. All its members belong to $H$. Hence this is the power set required inside $H$, not a claim that unrestricted external subsets are retained.

For separation, relativize the formula to $H$. A support for $a$ together with supports of the finitely many parameters fixes the resulting subclass; its members are hereditary, so its representative belongs to $H$. For replacement, old replacement applied to this relativized definable functional relation collects the range. Its members have a common bound on their construction ranks, by taking the supremum of the set of ranks, so the range's representative lies in $W$. The same finite union of parameter and domain supports fixes the range by uniqueness of each image. It is therefore hereditary and belongs to $H$. These arguments verify the full schemas. Thus $H\models \mathrm{ZF}-\mathrm{Foundation}$.

Finally $A$ and the family of its two-element subsets belong to $H$ with empty support. Suppose a choice function $c$ for this family belongs to $H$, and let $S$ be a finite support for $c$. Choose distinct atoms $u,v\notin S$ and transpose them. This permutation fixes $c$ and fixes the unordered pair $\{u,v{\}}_E$, but interchanges its two possible selected members. Equivariance would requirewhich is impossible. Hence choice fails in $H$. Together with a model of ZFC, or with its full permuted model satisfying choice and failing foundation, this proves independence of axiom of choice from ZF minus foundation, relative to consistency of ZFC. All hierarchy and support arguments are internal to the starting model; no transitive-model assumption is needed.