Naimark's problem: pure states and selective extensions

Source and verification

This article explains family 297 of the October 6, 2026 OpenAI collection. Research conclusions are attributed to its manuscripts. The review covers statements, introductions, and selected proof mechanisms; it is not an independent verification of the complete proofs.

🏷️ Preliminaries: representations and pure states

A representation of a complex $C^{\ast }$-algebra $A$ is a map $\pi :A\to B(H)$ preserving linear operations, multiplication and adjoints. It is irreducible if it is nonzero and the only closed subspaces invariant under every $\pi (a)$ are $\{0\}$ and $H$. Two representations are unitarily equivalent if a unitary between their Hilbert spaces intertwines their actions.

The compact operators $K(H)$ have exactly one nonzero irreducible representation up to unitary equivalence: the defining action on $H$. Naimark’s problem asks whether this property characterizes algebras of compact operators.

For a unital algebra, a state is a positive linear functional $h$ with $h(1)=1$. It is pure if it cannot be written as a nontrivial convex combination of distinct states. The Gelfand–Naimark–Segal construction represents a state as

$$h(a)=⟨{\pi }_h(a){\xi }_h,{\xi }_h⟩,$$

where ${\xi }_h$ is a cyclic unit vector: applying the represented algebra to it gives a dense subspace. Purity of $h$ is equivalent to irreducibility of ${\pi }_h$. Write $h\sim k$ when their GNS representations are unitarily equivalent.

Thus the representation problem can be expressed through pure states. Equivalent pure states need not be equal. They are different vector states in equivalent irreducible representations; in a unital algebra, Kadison’s transitivity theorem makes them conjugate by an algebra unitary.

A faithful tracial state $\tau$ satisfies $\tau (ab)=\tau (ba)$ and $\tau (a^{\ast }a)>0$ for every nonzero $a$. An algebra is simple if it has no nonzero proper closed two-sided ideal.

🏷️ The counterexample asserted by the manuscript

A tracial counterexample

The manuscript constructs a unital, infinite-dimensional, simple complex $C^{\ast }$-algebra $A$ with a faithful tracial state such that all nonzero irreducible representations of $A$ are unitarily equivalent. Its construction uses ZFC, the usual set-theoretic foundation with choice, without an additional prediction principle. (OpenAI, 2026)

The obstruction to being compact operators is elementary. If $K(H)$ is unital, then $H$ is finite-dimensional: the identity on an infinite-dimensional Hilbert space is not compact. Hence the asserted infinite-dimensional unital $A$ cannot be isomorphic to $K(H)$ for any $H$.

There is no separability assertion for $A$. The manuscript also recalls that uniqueness of an irreducible representation acting on a separable Hilbert space would force the compact-operator conclusion. That condition concerns the irreducible representation, rather than an arbitrary faithful representation.

The exposition below concentrates on the operator-algebra mechanism. The long indexing of the construction serves to ensure that every pure state is eventually handled; the decisive input is an analytic theorem making each such state uniquely determined on a separable subalgebra.

🏷️ Unique extension and peaking projections

For an inclusion $B\subset C$, the key preservation property is:

Unique extension

Every pure state of $B$ has exactly one extension to a state of $C$. Uniqueness is among all states, not only among pure extensions.

The unique extension is automatically pure. In a convex decomposition of it, purity forces both restricted states to equal the original state, and uniqueness then identifies both full states with the extension.

A decreasing sequence of projections $(p_n)$ peaks to a state $h$ if $h$ is the only state satisfying $h(p_n)=1$ for every $n$. The same convex-decomposition argument proves purity. If these projections lie in a smaller algebra, any state on that algebra satisfying the constraints can be extended to the larger one. Peaking therefore also gives purity of the restriction and uniqueness of its extension.

The analytic tool is excision. A pure state has positive contractions $x_{\lambda }$ with $h(x_{\lambda })=1$ such that

$$\begin{array}{c}\Vert x_{\lambda }(a-h(a)1)x_{\lambda }\Vert ⟶0\\ (a\in B).\end{array}$$

The index may be a directed set rather than the natural numbers. Compression near the state turns each operator into its scalar state value. For a peaking projection sequence, the projections themselves excise the state.

Two inequivalent pure states have a complementary cross-excision property. If $x_{\lambda },y_{\nu }$ are their decreasing excision families, then

$$\Vert x_{\lambda }a{y}_{\nu }\Vert ⟶0(a\in B)$$

on the product directed set. Operationally, after choosing sufficiently late indices in each family, the bound holds for every pair of later indices. This controls coefficients between inequivalent representations.

A faithful trace provides another useful vanishing statement. In the enveloping von Neumann algebra $B^{\ast \ast }$, the support of a pure state is the least projection $r_h$ on which its normal extension takes value one. It is minimal, meaning that it has no proper nonzero subprojection. If $B$ has no finite-dimensional irreducible representation, then

$$\tau (r_h)=0.$$

Here $\tau$ on $B^{\ast \ast }$ denotes its normal extension. That extension need not be faithful on $B^{\ast \ast }$, even though the original trace is faithful on $B$, so a nonzero support projection can have trace zero. To prove the identity, the GNS space contains arbitrarily many orthogonal unit vectors. Transitivity carries the cyclic vector to them by algebra unitaries, yielding arbitrarily many orthogonal conjugates of $r_h$, all with the same trace. Their total trace is at most one.

🏷️ The selective extension theorem

Suppose $B$ is unital, faithfully tracial, and has no finite-dimensional irreducible representations. Let decreasing projection sequences $(p_n^1),(p_n^2)$ peak to inequivalent pure states $f_1,f_2$, with

$$\begin{array}{c}{p}_0^1=p_0^2=1,\\ \tau (p_n^1)=\tau (p_n^2)=t_n,\\ 1=t_0>t_1>\cdots >0,\\ t_n⟶0.\end{array}$$

The manuscript adjoins a unitary $v$ implementing

$$v{p}_n^1{v}^{\ast }=p_n^2.$$

The resulting algebra $H(B)=C^{\ast }(B,v)$ has a faithful trace extending $\tau$ and a trace-preserving conditional expectation back to $B$. A conditional expectation $E:C\to B$ fixes $B$, preserves the unit, satisfies $E(b_{1}a{b}_2)=b_{1}E(a)b_2$, and is completely positive: applying $E$ entrywise preserves positive matrices of every size. It is faithful if $E(a^{\ast }a)=0$ forces $a=0$.

Every old pure state has a unique state extension, and the only old equivalence classes that merge are the designated two:

$$\begin{array}{c}\tilde{h}\sim \tilde{k}\\ ⟺\\ h\sim k\text{ or }\{[h],[k]\}=\{[f_1],[f_2]\}.\end{array}$$

If $B$ is simple, so is $H(B)$. This selective bridge is much stronger than merely making $f_1$ and $f_2$ equivalent. Its exclusion of unintended identifications is needed later when steps are reordered.

To describe the extension, put $D_i=C^{\ast }(p_n^i:n\ge 0)$. These commuting projections identify $D_i$ with the continuous functions on a convergent sequence and its limit point. For $d_n^i=p_n^i-p_{n+1}^i$, the expectation onto $D_i$ is

$$\begin{array}{c}{E}_i(a)(n)=\frac{\tau (d_n^{i}a{d}_n^i)}{\tau (d_n^i)},\\ E_i(a)(\infty )=f_i(a).\end{array}$$

Excision makes the finite values converge to the value at infinity, so the output is continuous. Equality of the projection traces makes $p_n^1\mapsto p_n^2$ a trace-preserving isomorphism $\theta :D_1\to D_2$.

The reduced HNN construction, used here in Ueda’s matrix-corner form, supplies the stable unitary $v$ and an expectation $E_B:H(B)\to B$. Its operational rule is that nonempty reduced words have zero expectation. A word is

$$\begin{array}{c}{a}_0{v}^{{\epsilon }_1}{a}_1\cdots v^{{\epsilon }_m}{a}_m,\\ {\epsilon }_j\in \{+1,-1\}.\end{array}$$

At a $+,-$ turn the intervening coefficient must have $E_1(a_j)=0$; at a $-,+$ turn it must have $E_2(a_j)=0$. Equal consecutive signs impose no condition. Splitting a coefficient into its expectation and centered part either reduces the word length or produces such a reduced word.

🏷️ Why the extension is selective

The source realizes $H(B)$ as a corner of the reduced amalgamated free product

$$\begin{array}{c}Z=F{\ast }_{D}G,\\ F=M_2(B),G=M_2(D_1),\\ D=D_1\oplus D_2.\end{array}$$

The two copies of $D$ are identified using $\theta$. The reduced representation is built from alternating tensor words of centered vectors from $F$ and $G$. Its expectation onto $F$ annihilates each nonempty alternating word containing a centered $G$ factor.

The distinguished pure states become the two corner states ${\varphi }_i(a)=f_i(a_{ii})$. Write $s=e_{21}^G$ and $s^{\ast }=e_{12}^G$. In this matrix argument, $h,k$ are pure states of $F=M_2(B)$, with the two distinguished classes represented by ${\varphi }_1,{\varphi }_2$. For an equal-state pair $h=k$, or an inequivalent pair other than the designated pair, the main estimate is

$$\Vert x_{\lambda }W{y}_{\nu }\Vert ⟶0$$

for every nonempty reduced matrix word $W$ built from $s,s^{\ast }$ and $F$ coefficients. Here $x_{\lambda },y_{\nu }$ excise the endpoint states, using peaking projections for the distinguished representatives.

There are two parts to the estimate. Let $P_G$ project onto tensor words beginning with a centered $G$ factor. For $d\in F$,

$$\Vert d{P}_G\Vert \le \Vert E_D^F(d^{\ast }d){\Vert }^{1/2}.$$

A centered alternating core $w$ satisfies

$$w=P_{G}w+(1-P_G)w{P}_G.$$

Consequently a compressed word is small if the corresponding expectation norms decay at both ends. At each finite point of the spectrum of $D$, this decay follows from the trace-null pure-state support. At the limit point, it follows from inequivalence to the relevant distinguished state. Decreasing continuous functions on the compact convergent sequence then tend to zero uniformly, by the compactness proof of Dini’s theorem.

The exceptional situation is a matched endpoint, where the endpoint state is the distinguished state attached to that matrix corner. Split its coefficient into its state value and a zero-value remainder. The remainder has the required decay. For the scalar term, move the peaking projection through the adjacent matrix letter:

$$P_n^{\ell (b)}b=b{P}_n^{r(b)}(b=s\text{ or }{s}^{\ast }).$$

This exposes a formerly internal coefficient, which is centered and hence has zero value under both distinguished states. The endpoint criterion now applies. If no core remains, the one-letter case uses cross-excision; the two-letter, two-transfer case uses ordinary excision of its centered middle coefficient. The one-letter case with both ends matched is exactly the designated pair excluded from the test.

For $h=k$, every extension fixes its excision vectors and therefore evaluates all these nonempty words as zero. Its values on $F$ are already specified, proving uniqueness on a dense span. For a prohibited inequivalent pair, the same decay would make every matrix coefficient between their cyclic vectors vanish if the lifted representations were equivalent, contradicting cyclicity.

Finally, the stable-unitary relation makes ${\tilde{f}}_2\circ \mathrm{Ad}(v)$ extend $f_1$, where $\mathrm{Ad}(v)(a)=va{v}^{\ast }$. Unique extension gives

$${\tilde{f}}_1={\tilde{f}}_2\circ \mathrm{Ad}(v).$$

Thus the desired pair does merge, and the others do not. Simplicity follows from unique extension and the faithful expectation: a proper quotient is faithful on the simple base, so each base pure state extends through it; uniqueness forces the expectation to annihilate its positive kernel, and faithfulness makes that kernel zero.

🏷️ Countable data inside a long construction

Start with the norm closure of the increasing matrix algebras

$$B_0=\underset{m=1}{\overset{\infty }{⨂}}{M}_2(\mathbb{C}).$$

Set $p_0=1$. The projections $p_n=e_{11}^{\otimes n}\otimes 1$ for $n\ge 1$ peak to a pure state and have trace $2^{-n}$. The algebra is separable, simple and infinite-dimensional. Any unital algebra containing it has no finite-dimensional irreducible representations, since such a representation would embed the simple infinite-dimensional $B_0$ into a finite-dimensional matrix algebra.

At an active step $j$, adjoin one stable unitary $v_j$ joining the fixed $p_n$-state to another state peaked by projections $q_n^j$. At a limit stage, take the norm closure of the union. Unique extension composes along this chain; simplicity and faithfulness of the compatible traces also persist.

Every element of the resulting algebra depends on only countably many stable unitaries: approximate it by a sequence of finite polynomials. In particular, choose a countable set $I_j$ of earlier steps containing the data $q_n^j$.

A set of steps $S$ is closed under dependencies if $j\in S$ implies $I_j\subset S$. Write

$$B[S]=C^{\ast }(B_0,\{v_j:j\in S\}).$$

Closing a countable set under dependencies still gives a countable set. Its algebra $B[S]$ is therefore separable.

The important theorem is that every pure state on $B[S]$ has a unique state extension to the whole iteration. Unique extension along the original chain alone would not give this for an arbitrary selected set of steps.

The proof moves each omitted adjunction after the retained ones. Two adjunctions whose data lie in a common base can be interchanged: their reduced-word rules give the same joint trace moments, and faithfulness recovers polynomial norms through

$$\Vert x\Vert =\underset{r\to \infty }{\lim }\tau ((x^{\ast }x{)}^r{)}^{1/(2r)}.$$

The selective bridge theorem keeps the omitted step’s two states inequivalent while it is postponed. A retained active step joins the fixed class to a different class, and so cannot accidentally perform the omitted join. This is where the “no other classes merge” part of the analytic theorem is used.

Now let $h$ be any pure state on the full iteration. Repeatedly add countably many excision witnesses for a countable dense set in the current $B[S]$, and close the new supports under dependencies. In the resulting countable union, the restriction of $h$ is pure: any convex decomposition must fix every witness, and the excision bounds force both components to agree on a dense set. The preceding unique-extension theorem then says that this restriction uniquely determines $h$ on the whole algebra.

This conclusion concerns a suitably chosen separable subalgebra for each state. Restriction of a pure state to an arbitrary subalgebra need not be pure.

🏷️ Treating every final state

Let $\mathfrak{c}$ be the cardinality of the real numbers and let $\kappa ={\mathfrak{c}}^{+}$ be the next larger cardinal. Below stage $\kappa$, there are at most $\mathfrak{c}$ countable sets of earlier steps. Each corresponding separable algebra has at most $\mathfrak{c}$ states, determined by their values on a countable dense set. Countable determination therefore bounds the number of pure states at every earlier stage by $\mathfrak{c}$.

Fix in advance a schedule assigning every pair “earlier stage, state-list index” to a distinct stage no earlier than the one listed. When a state is reached, join it to the fixed class if it is not equivalent already.

The required second peaking sequence is obtained on its separable simple determining algebra. An imported homogeneity theorem of Kishimoto, Ozawa and Sakai gives an automorphism $\gamma$ carrying the appropriate pure state to the fixed one. Set $q_n=\gamma (p_n)$. Peaking follows from the automorphism and unique extension.

The trace matching requires care: an arbitrary automorphism need not preserve the specified trace on the entire algebra. Instead, $p_n$ is a minimal projection in a unital copy of $M_{2^n}$, and $\gamma (p_n)$ is minimal in its image matrix algebra. The ambient trace restricts to the unique normalized matrix trace on both, so

$$\tau (q_n)=2^{-n}=\tau (p_n).$$

At the final stage, every pure state is determined on a countable set of steps. Such a set is bounded below $\kappa$, so the state restricts purely at an earlier stage and appears on that stage’s list. Its scheduled join is therefore performed. Unique extension preserves the resulting equivalence to the final algebra.

Every nonzero irreducible representation of a unital algebra is the GNS representation of any one of its unit vector states. Hence equivalence of all final pure-state representations gives the asserted uniqueness of irreducible representation. The final algebra remains unital, simple, faithfully tracial and infinite-dimensional. The recursion length is ${\mathfrak{c}}^{+}$; the paper does not identify that length with the algebra’s exact density. (OpenAI, 2026)

🏷️ Key steps and proof outlines

The analytic hinge is decay at both ends of a reduced word

The bridge theorem needs a uniform statement about all old pure states, including states absent from the two peaking sequences used to define the extension. The following estimate explains how the source’s “Joining two classes” section achieves that. Work in its matrix enlargement $F\subset Z$, and let $P_G$ project onto reduced-word vectors whose first letter comes from the other free-product component. If $d\in F$, the Hilbert-module inner product gives

$$\Vert d{P}_G{\Vert }^2\le \Vert E_D^F(d^{\ast }d)\Vert .$$

Indeed, such a vector has the tensor form $1\otimes \eta$, and multiplication replaces $1$ by $d$; its squared norm is controlled by the positive coefficient $E_D^F(d^{\ast }d)$. This converts a noncommutative operator estimate into a supremum of scalar functions on two convergent sequences. (OpenAI, 2026)

For a centered alternating word with core $w$, the reduced-word decomposition gives $(1-P_G)w(1-P_G)=0$. Thus

$$\Vert dw{d}^{\prime }\Vert \le \Vert d{P}_G\Vert \Vert w\Vert \Vert d^{\prime }\Vert +\Vert d\Vert \Vert w\Vert \Vert P_G{d}^{\prime }\Vert .$$

Both endpoint expectation norms must tend to zero for this direct argument. At a finite atom, this follows from the trace-null support of the excised pure state. At the limiting point, it follows from inequivalence to the distinguished endpoint state, or from a coefficient whose value in that state is zero. Monotonicity upgrades these pointwise limits to uniform convergence: for each tolerance, compactness supplies finitely many indices that cover the spectrum; a common later index works everywhere. This is the exact role of the Dini argument.

Matched endpoints expose the centered coefficient

A matched endpoint is the exceptional case where its pure state equals the distinguished state attached to the adjacent letter. Split its coefficient as $a=h(a)1+a^0$, where $h(a^0)=0$. The zero-value part now decays by the previous estimate. For the scalar part, move the peaking projection through the adjacent matrix letter using

$$P_n^{\ell (b)}b=b{P}_n^{r(b)}.$$

The newly exposed coefficient was internal to the original reduced word. Its conditional expectation is zero, hence both distinguished states evaluate it to zero. Therefore the transferred endpoint also satisfies the required decay estimate, even if its new state matches the next letter.

Two short words need separate checks. For one letter, matching both ends is precisely the designated pair of classes, which is excluded from the no-unintended-join test. A single transfer leaves a compression between inequivalent states, so cross-excision applies. For two letters with both ends transferred, one obtains $P_n^{j}a{P}_t^j$ with distinguished-state value zero. Nestedness bounds its norm by $\Vert P_q^{j}a{P}_q^j\Vert$, where $q=\min (n,t)$, and excision makes this vanish. This also explains why convergence must hold on the product of the two index sets.

One estimate has two separate outputs

If the endpoint states agree, reduced-word decay proves uniqueness of state extension. If they are inequivalent and not the designated pair, the same decay prevents their lifted representations from becoming equivalent.

For uniqueness, an extending state’s cyclic vector is fixed by every positive contraction of state value one. Hence its value on a tested word equals its value on the compressed word and must be zero. The old algebra and these words have dense linear span. For inequivalence, a supposed common irreducible representation would have all coefficients $⟨\pi (z){\xi }_k,{\xi }_h⟩$ equal to zero; cyclicity of ${\xi }_k$ makes that impossible. Finally, the stable unitary explicitly merges the designated pair, and uniqueness identifies the resulting conjugate states. These analytic facts are the interfaces carried through the manuscript’s subsequent iteration; they do not by themselves verify that iteration’s separate bookkeeping arguments.

🏷️ Connections

Generators for finite factors concerns generation after weak operator closure and a separable-predual hypothesis. The present construction uses norm-closed $C^{\ast }$-algebras and makes no separability claim. Its pure-state extension theorem is the mechanism that turns countable local data into control of all final irreducible representations.

🏷️ Manuscripts and verification scope

• A counterexample to Naimark’s problem in ZFC — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selection is a unital infinite-dimensional simple complex C*-algebra with a faithful tracial state and exactly one unitary-equivalence class of nonzero irreducible representations, not isomorphic to compact operators on any Hilbert space. No additional set-theoretic assumption is used. The scope page makes no separability assertion. No Lean build, axiom audit, or manuscript-to-formal-statement equivalence check was performed for these notes.

OpenAI collection provides prerequisites and the reading sequence; coverage register records completed and pending families.

🐻  OpenAI 2026. A counterexample to Naimark’s problem in ZFC.