Set-theoretic absoluteness for analysts and separable -algebras without Choice

Let x_n, for n\in \mathbb{N}, be an enumeration of a dense subset of the unit ball of X.

(?) It suffices to prove this in the case when X is a real Banach space. The standard proof, using Zorn’s Lemma (e.g., [pedersen2012analysis, Lemma 2.3.2]), can be performed using a countable dense set instead, as follows. Then Y_n=\operatorname{span}(Y\cup \{x_j\mid j\leq n\}) defines an increasing sequence of linear subspaces of X with dense union. Let Y_0=Y and \psi_0=\varphi. The standard proof of the Hahn–Banach theorem shows that for every n and linear functional \psi_n on Y_n there is a closed interval [\alpha,\beta] in \mathbb R such that for any \gamma\in [\alpha,\beta], \psi_{n+1}(y+tx_n)=\psi_n(y)+t\gamma extends \psi_n to Y_{n+1} and satisfies \|\psi_n\|=\|\psi_{n+1}\|. Use \gamma=\alpha to define \psi_{n+1}. This defines \psi on \bigcup_nY_n, and \psi extends to X by continuity.

(?) For the Banach–Alaoglu theorem, note that the evaluation map from \mathbf B(X^*) into \mathbb{D}^n (where \mathbb{D} denotes the closed unit disk in \mathbb{C}) defined by \varphi\mapsto (\varphi(x_n)) is, by the standard proof, a homeomorphism onto its range. Since \mathbb{D} is homeomorphic to [0,1]^2, and since [0,1]^\mathbb{N} is compact by Tychonoff’s theorem (no choice needed, cf. [BlackadarReal, XII.11.6.9]), \mathbb{D}^\mathbb{N} is compact. The range of the evaluation map is closed, hence compact.

(?) The standard metric on \mathbf B(X^*),

is as required, by the standard proof. \mathbf B(X^*) is second countable, hence separable by [BlackadarReal, XII.15.9.8].

(?) and (?): While we don’t know ordinary proofs of these assertions that do not use the Axiom of Choice, we can prove that a proof exists. See ? and ?.

(?): The Baire Category Theorem for separable complete metric spaces requires no Choice. The proofs of the Open Mapping Theorem and Uniform Boundedness Theorem (second proof) as written in [BlackadarReal] require no additional choice (in the proof in XVIII.8.2.5 the x_n can be systematically chosen from a dense sequence), and the Closed Graph Theorem is a simple corollary of the Open Mapping Theorem.

If X is separable, then Y is separable (Proposition), and a countable dense set in X works for C. Conversely, let C be such a subset of X, and D a countable dense subset of Y. Then C+D is countable. Let \pi:X\to X/Y be the quotient map. Let x\in X and \epsilon>0. Then there is a c\in C with \|\pi(x)-\pi(c)\|_{X/Y}=\|\pi(x-c)\|_{X/Y}<\epsilon. From the definition of the quotient norm and density of D in Y, there is a d\in D with

and thus C+D is dense in X.

Fix a separable \mathrm{C}^*-algebra A. We may assume that A is unital. Fix x\in A. We will prove that there is a state \varphi on A such that \varphi(x^*x)=\|x\|^2. In particular, if A is a unital \mathrm{C}^*-algebra and x\in A, then x^*x\in A_+ (Proposition Proposition), and it follows from functional calculus that \|\lambda 1+x^*x\|=\lambda+\|x^*x\| for any \lambda\geq0. Thus, if S is the two-dimensional subspace of A spanned by 1 and x^*x, the linear functional \phi on S defined by \phi(1)=1 and \phi(x^*x)=\|x^*x\|=\|x\|^2 has norm 1.

The Hahn–Banach Theorem for separable Banach spaces does not require any Choice (Proposition Proposition); in fact \phi has an algorithmically constructed extension to A of norm 1 (once a countable dense sequence is fixed). Since a linear functional \psi on A is a state if and only if it has norm 1 and \psi(1)=1 [BlackadarOperator, II.6.2.5], we conclude that a separable \mathrm{C}^*-algebra has sufficiently many states.

The state space is nonempty by Lemma Lemma. The standard proof that \mathsf S(A) is a weak*-closed subset of the unit ball of the Banach space dual of A requires no choice. By the Banach–Alaoglu theorem (Proposition Proposition (?) and (?)), the latter is a separable compact metric space.

Finally, by the Krein–Milman theorem for separable compact convex sets (a proof is provided in ?), \mathsf S(A) is the closed convex hull of the space of its extreme points, which is the space of pure states.

That \mathsf P(A) is a G_\delta is a standard result, see. e.g. [BlackadarReal, XVIII.9.5.10].

The proof in case when A is not necessarily separable but has a well-ordered dense subset is analogous and uses Proposition Proposition (although \mathsf P(A) will not be Borel in general).

Finally, if A is unital and commutative then the standard proof that the space of characters is closed in \mathsf S(A) and that it coincides with \mathsf P(A) applies.

We may assume A is unital. Then \mathsf S(A) is weak-* separable (Lemma). Let (\phi_n) be a dense sequence in \mathsf S(A), and set \phi=\sum_{n=1}^\infty 2^{-n}\phi_n. Then \phi is a state on A. We claim that \phi is faithful. Fix a positive a\in A of norm 1. Then there is an (algorithmically constructed) state \psi on A with \psi(a)=1. Approximating \psi by a \phi_n, there is an n with \phi_n(a)>1/2, and we have \phi(a)\geq2^{-n}\phi_n(a)>2^{-n-1}>0.

By the \mathrm{C}^*-axiom, \|x^{2^k}\|^{2^{-k}}=\|x\| for any k. So we have that \lim_{n\to\infty} \|x^n\|^{1/n} exists and equals \inf_n\|x^n\|^{1/n}, and submultiplicativity of the norm implies \|x^n\|^{1/n} \leq \|x\| for all n, so we have \|x^n\|^{1/n}=\|x\| for all n. Thus the sequence in the statement of Theorem Theorem is constant.

There is a canonical a=x^{1/2}\in C^*(x)\subseteq A with a^2=x (Proposition Proposition (vii)). It remains to prove uniqueness of the square root in A. Since a is a uniform limit of polynomials in x, it commutes with any element of A which commutes with x. If b is any positive element of A with b^2=x, then b commutes with x and hence with a. Thus, replacing A with C^*(a,b), we may assume A is commutative. Since

we have a-b\perp a+b, and thus a-b\perp(a+b)^{1/2^n} for any n\in\mathbb{N}. But a and b, hence a-b, are in the hereditary \mathrm{C}^*-subalgebra of A generated by a+b (which is positive), and \{(a+b)^{1/2^n}:n\in\mathbb{N}\} is an approximate unit for this hereditary \mathrm{C}^*-subalgebra, so we have

and hence a-b=0.

If A is representable, then vector states determine the norm. Conversely, if states determine the norm, a direct sum of GNS representations is faithful (isometric).

By Proposition ? it suffices to prove that A has sufficiently many states, and this was shown in Lemma Lemma.

We first prove the assertion with the additional assumption that A is unital. By Lemma Lemma, A has sufficiently many states. Since the state space of A is compact and metrizable, the Krein–Milman theorem (?) implies that it is equal to the closed convex hull of the space of pure states of A. Therefore the Gelfand transformation is isometric. The image of A is a *-subalgebra of C(\mathsf P(A)) that contains all scalars and (by definition) separates the points. Since the Stone–Weierstrass theorem can be proved without the Axiom of Choice (?), the map is an isomorphism as required.

If A is not unital, consider its unitization \tilde A. Then \mathsf P(A) is naturally identified with an open dense subspace of \mathsf P(\tilde A) that separates the points in the image of A under the Gelfand map, and A is identified with C_0(\mathsf P(A)). The result follows from the already proved result for the unitization of A.

(?) Let X=\bigsqcup_n X_n be a Russell set, where each X_n is a two-element set. We first define a universal \mathrm{C}^*-algebra A_0 given by generators and relations; the existence of such algebra does not require any Choice. We will then define a state \tau on A_0, and the required algebra will be the image of A_0 under the GNS representation associated with \tau.

Generators of A_0 are p_x, for x\in X. The relations require that each p_x is a projection (that is, p_x=p_x^*=p_x^2), that for every n we have \sum_{x\in X_n} p_x=1, and that [p_x,p_y]=0 for all x and y. Thus p_x p_y=0 if and only if x and y are distinct and belong to the same X_n. The space of linear combinations of products \prod_{x\in F} p_x, where F\subset X intersects each X_n in at most one point (every such transversal is finite because X is a Russell set), is a dense *-subalgebra of A_0. On such monomials define a functional by \tau(\prod_{x\in F} p_x)=2^{-|F|}. We claim that \|\tau\|=\tau(1)=1. To see this, note that \tau(1)=1 and therefore \|\tau\|\geq 1. To see that \|\tau\|\leq 1, take a finite linear combination a of monomials in projections p_x. For a large enough n, a is of the form a=\sum_{x\in \prod_{j\leq n} X_j} \lambda_x \prod_{j\leq n} p_{x(j)} with complex coefficients \lambda_{x}. Then

and \|\tau\|\leq 1 follows. It follows that \tau is a contraction on a dense subalgebra of A_0. Its continuous extension to A_0 (denoted \tau for convenience) still satisfies \|\tau\|=\tau(1)=1, and is therefore a state. Since A_0 is commutative, it is tracial.

Then A=\pi_\tau[A_0] is a concrete commutative \mathrm{C}^*-algebra. Fix x\in X. Since \tau(p_x)=1/2, \pi_\tau(p_x) is a nonzero projection. In the following, for the simplicity of notation we will write p_x for \pi_\tau(p_x).

We need to prove that the pure state space of A is empty. Assume \varphi is a pure state on A. Then for every n we would have \varphi(p_x)=1 for some x\in X_n, otherwise \varphi would be a convex combination of \psi_x(a)=\varphi (p_x a)/\varphi(p_x) and \psi_y(a)=\varphi (p_y a)/\varphi(p_y), where X_n=\{x,y\}. Thus from \varphi one could define a subset of X, \{x\in X\mid \varphi(p_x)=1\} that intersects each X_n in exactly one point; contradiction.

This implies that the Gelfand spectrum of A is empty, and therefore A is not isomorphic to C(Z) for a compact Hausdorff space Z.

For each n\geq 1, let Y_n=\bigsqcup_{1\leq k\leq n}X_k. Then Y_n is a finite set with 2n elements, and X is the increasing union of the Y_n. Let A_n be the \pi_\tau-image of set of linear combinations of projections of the form \prod_{x\in F} p_x, where F\subset Y_n intersects each X_k in at most one point. Then A_n is a unital \mathrm{C}^*-subalgebra of A isomorphic to {\mathbb C}^{2^n}, A_n\subseteq A_{n+1}, and \bigcup_nA_n is dense in A. In fact, if B_n is the \mathrm{C}^*-subalgebra generated by \{p_x:x\in X\setminus Y_n\}, then A is naturally isomorphic to A_n\otimes B_n.

(?) Let \mathsf S(A) be the state space of the \mathrm{C}^*-algebra A as in (?). \mathsf S(A) can be described as an inverse limit of finite-dimensional simplexes: there is an inverse system

where the connecting maps are given by restriction. \mathsf S(A_n) is a (2^n-1)-simplex, and the connecting maps send extreme points to extreme points. The space \mathsf S(A) can be naturally put at the index -\infty, and the restriction maps from \mathsf S(A) to \mathsf S(A_n) separate points. They are even surjective. This is because for every \psi\in \mathsf S(A_n), the tensor product of \psi and the restriction of \tau to B_n belongs to \mathsf S(A).

To see that \mathsf S(A) is not compact (in the weak-* topology), let K_n be the set of states of A whose restriction to A_n is a pure state. Then K_n is closed since the pure state space of A_n is closed, and nonempty since the restriction map from \mathsf S(A) to \mathsf S(A_n) is surjective. So (K_n) is a decreasing sequence of nonempty closed subsets of \mathsf S(A), and any element of \bigcap_n K_n would define a choice function for X. Therefore \bigcap_nK_n=\emptyset and \mathsf S(A) is not compact.

Let (P_n) be a sequence of mutually orthogonal nonzero projections in {\mathcal B}({\mathcal H}), and let {\mathcal H}_n be the range of P_n. Then ({\mathcal H}_n) is a sequence of mutually orthogonal closed subspaces of {\mathcal H}. For each S\subseteq\mathbb{N}, let {\mathcal H}_S be the closed span of \{{\mathcal H}_n:n\in S\}, and P_S the projection onto {\mathcal H}_S. Then \{P_S:S\subseteq\mathbb{N}\} is a 1-discrete subset of {\mathcal B}({\mathcal H}) of cardinality 2^{\aleph_0}. Since 2^{\aleph_0}>\aleph_0, {\mathcal B}({\mathcal H}) cannot be separable by ?.

Since the axioms of the theory T can be recursively enumerated, and since a proof in T is a finite sequence of sentences that satisfies strict and clear rules, one can enumerate all (formal) proofs in T. Knowing that \varphi is provable in T, one knows that a proof for \varphi will eventually appear in this list.

If \mathsf X is a countable standard Borel space, then its Borel \sigma-algebra is equal to {\mathcal P}(\mathsf X), thus every two countable standard Borel spaces of the same cardinality are Borel-isomorphic. The second part is in this case vacuous.

So it suffices to prove that every uncountable standard Borel space \mathsf X is Borel-isomorphic to the Cantor space K. By Cantor–Bendixson analysis, \mathsf X contains a perfect subset ([Ke:Classical, Theorem 6.4]), hence there is a continuous injection from K into \mathsf X. On the other hand, there is a Borel injection from \mathsf X into K (for an elegant proof of this fact due to Jenna Zomback see [farah2019combinatorial, Lemma B.2.3]). By the Borel version of the Schröder–Bernstein theorem ([Ke:Classical, Theorem 15.7]), \mathsf X is Borel-isomorphic to K.

It remains to prove that a Borel isomorphism between Polish spaces preserves the projective hierarchy. Clearly, a Borel isomorphism sends Borel sets to Borel sets. Suppose that f\colon \mathsf X\to \mathsf Y is a Borel-measurable map between Polish spaces and that \mathsf A\subseteq \mathsf X is a \boldsymbol\Sigma^1_1 set. Fix a standard Borel space \mathsf Z and a Borel \mathsf B\subseteq \mathsf X\times \mathsf Z such that \mathsf A is the projection of \mathsf B. The graph of f is a Borel subset of \mathsf X\times \mathsf Y. The following subset of \mathsf X\times \mathsf Z\times \mathsf Y:

is an intersection of two Borel sets and therefore Borel, and its projection to \mathsf Y is equal to f[\mathsf A]. This implies that \boldsymbol\Sigma^1_1 sets are preserved by Borel-isomorphisms, and the statement for \boldsymbol\Pi^1_1 sets follows by taking complements.

Suppose that n is such that every Borel isomorphism between standard Borel spaces sends \boldsymbol\Sigma^1_n sets to \boldsymbol\Sigma^1_n sets and \boldsymbol\Pi^1_n sets to \boldsymbol\Pi^1_n sets. Fix a Borel isomorphism between standard Borel spaces f\colon \mathsf X\to \mathsf Y. The proof that the image of a \boldsymbol\Pi^1_n set \mathsf A\subseteq \mathsf X under a Borel map f is \boldsymbol\Sigma^1_{n+1} is analogous to the proof in case when n=0: Fix a \boldsymbol\Pi^1_{n} set \mathbf B\subseteq \mathsf X\times \mathsf Z, note that the set in (?) is an intersection of a \boldsymbol\Pi^1_n set and a Borel set, and therefore \boldsymbol\Pi^1_n. Its projection to \mathsf Y is equal to f[\mathsf A]. By taking complements, this conclusion extends to \boldsymbol\Pi^1_{n+1} sets and completes the proof of the inductive step.

Since M is transitive, we have \mathsf X\subseteq M. If \mathsf T has an infinite branch x in M, then x is still an infinite branch in N. Assume \mathsf T is well-founded in M. Recursively define rank function \rho_\mathsf T from \mathsf T into the ordinals by

The supremum of the range of \rho_\mathsf T is an ordinal of cardinality no greater than that of \mathsf T, and since M is a model of ZF this ordinal is included in M.

Since \mathsf T is well-founded, the domain of \rho_\mathsf T is equal to \mathsf T and s\sqsubset t implies \rho_\mathsf T(s)>\rho_\mathsf T(t). If x is an infinite branch of \mathsf T in N, then \rho_\mathsf T(x\restriction n), for n\in \mathbb{N}, is an infinite decreasing sequence of ordinals; contradiction.

It suffices to prove the statement for \boldsymbol\Sigma^1_1 sets. By transitivity (of logical equivalence, pun not intended), it also suffices to prove absoluteness between a fixed transitive model of {ZF} and the universe. We first prove absoluteness for \boldsymbol\Sigma^1_1 subsets of the Baire space {\mathbb{N}}^{\mathbb{N}}. For every analytic subset \mathsf A of {\mathbb{N}}^{\mathbb{N}} there is a closed \mathsf B\subseteq ({\mathbb{N}}^{\mathbb{N}})^2 such that \mathsf A is the projection of \mathsf B. Let

Then \mathsf T is a tree on \mathbb{N}\times \mathbb{N} and

In other words, writing (by |s| we denote the length of s)

we have that \mathsf A is the set of all x such that \mathsf T_x is ill-founded. By Lemma Lemma, the latter assertion is absolute for transitive models of ZF and therefore for x\in M we have that x\in \mathsf A^M if and only if x\in \mathsf A, as required.

For subsets of an arbitrary Polish space \mathsf X, Theorem Theorem implies that \mathsf X is Borel-isomorphic to {\mathbb{N}}^{\mathbb{N}} and that the isomorphism is definable from an enumeration of a countable generating subset of \Sigma, and absolute for transitive models of ZF. Since Borel-isomorphism sends \boldsymbol\Sigma^1_n sets to \boldsymbol\Sigma^1_n sets, the conclusion follows.

Fix \varphi that is \boldsymbol\Sigma^1_2 or \boldsymbol\Pi^1_2 and provable in ZF + V=L[x], where x\in {\mathcal P}(\mathbb{N}) codes the parameters of \varphi. Fix a model M of ZF that contains x. This code can be chosen in any recursively defined Polish space, and it will be convenient to choose x in {\mathcal P}(\mathbb{N}). By our assumption, \varphi holds is L[x]^M (i.e., L[x] as computed in M) and by Corollary Corollary, \varphi is absolute between M and L[X]^M. Since M was an arbitrary model that contains x, this proves that \varphi holds in every model of ZF that contains x. By the Completeness Theorem of first-order logic, it follows that \varphi is a theorem of ZF.

For finite, stably finite, and purely infinite and simple one only needs to observe that the defining property of the class is already first order; see the proof of [Muenster, Theorem 2.5.1].

(?) Let A be a code for a separable \mathrm{C}^*-algebra. Since the space \mathsf S(A) of states of A is a compact metrizable space (Lemma Lemma) and the evaluation function is continuous, the space of all tracial states of A is closed and the assertion that it is nonempty is \boldsymbol\Sigma^1_1.

(?) The class \mathrm{C}^*-algebras with CPAP is Borel by a result of Effros ([Kec:C*, Theorem 1.2]). This result is stated using the equivalence of nuclearity and CPAP, but from the proof it is clear that CPAP is used. An alternative proof of this, and Borelness of related approximation properties (such as for example small nuclear dimension) can also be extracted from [Muenster, Theorem 5.7.3].

(?) For nuclearity, fix a (code for) a separable \mathrm{C}^*-algebra A. Then A is not nuclear if and only if there is a \mathrm{C}^*-algebra B with \|\cdot\|_{\max}\neq\|\cdot\|_{\min} on A\odot B. For such a B, there is a z=\sum_{k=1}^n x_k\otimes y_k\in A\odot B with \|z\|_{\max}>\|z\|_{\min}. Then \|\cdot\|_{\max}\neq\|\cdot\|_{\min} on A\odot C^*(y_1,\dots,y_n). Therefore, we may assume B is separable. Since both A and B are separable, \|\cdot\|_{\min} is the norm in a GNS-representation by states of the form \psi\otimes \eta, where \psi and \eta are faithful states of A and B.

Finally, since \pi_\psi and \pi_\eta are faithful, the right-hand side in the displayed formula of the Claim is equal to \|x\|_{\min}. Therefore, the set of codes for non-nuclear \mathrm{C}^*-algebras is, as the projection of this set, \boldsymbol\Sigma^1_1. This implies that the set of codes for nuclear \mathrm{C}^*-algebras is \boldsymbol\Pi^1_1 as claimed.

(?) By Lemma Lemma, the set of codes in \hat\Xi corresponding to separable \mathrm{C}^*-algebras that are stably finite and have no tracial state is \boldsymbol\Pi^1_1. Thus the assertion that this set is empty is \boldsymbol\Pi^1_2, hence absolute between transitive models of ZF that contain all countable ordinals by Theorem Theorem.

(?) and (?) The set of codes for separable \mathrm{C}^*-algebras with UCT is \boldsymbol\Sigma^1_1 by [pi2026game, Theorem 3.27]. By Lemma Lemma, (?) asks whether a concrete Borel set is included in a concrete \boldsymbol\Sigma^1_1 set and (?) asks whether a concrete \boldsymbol\Pi^1_1 set is included in a concrete \boldsymbol\Sigma^1_1 set. Thus each of these two statements is equivalent to asserting that every code for a separable \mathrm{C}^*-algebra belongs to a certain \boldsymbol\Sigma^1_1 set. Such statement is equivalent to a \boldsymbol\Pi^1_2 statement, and therefore absolute for transitive models of ZF that contain all countable ordinals by Shoenfield’s Absoluteness Theorem, Theorem Theorem.

The property of having a tracial state is axiomatizable in continuous logic by [Muenster, Theorem 2.5.1]. This proof uses an abstract characterization of axiomatizability via ultraproducts.

We will prove that in some model of ZF there are elementarily equivalent \mathrm{C}^*-algebras A and B such that B has a tracial state and A does not. Once proved, this will imply that having a tracial state is not axiomatizable. As pointed out in ?, (see also ?) the commutative \mathrm{C}^*-algebra A=\ell^\infty(\mathbb{N})/c_0(\mathbb{N}) has no states (tracial or otherwise) if all sets of reals have the Property of Baire. A model with this property in which DC holds was constructed starting from a model of ZFC in [shelah1984can, Corollary 7.17]. Since DC implies _\omega, in this model \operatorname{Sep}(A) is \sigma-complete and therefore the downwards Löwenheim–Skolem Theorem(We would like to thank Asaf Karagila for pointing out that DC is equivalent to the downwards Löwenheim–Skolem theorem (Boolos, see [aspero2021dependent, Theorem 3.4]).) ([farah2019combinatorial, Theorem 7.1.9]) implies that A is elementarily equivalent to some of its separable \mathrm{C}^*-subalgebras. Every commutative, separable \mathrm{C}^*-algebra B is isomorphic to C(X) for a compact metrizable space X (Lemma Lemma) and in particular has tracial states.

It suffices to prove (using AC) that if there is a stably finite \mathrm{C}^*-algebra without a tracial state then there is one that is in addition separable. If A is a unital \mathrm{C}^*-algebra, and every separable unital \mathrm{C}^*-subalgebra B has a tracial state \tau_B, extend \tau_B to a state \phi_B on A; then any cluster point of the net (\phi_B) is a tracial state on A. Since being tracial is a closed condition, this cluster point is a tracial state.