Kadison similarity: uniform commutators and free-product corners

Source and verification

This article explains family 288 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.

Overview

Kadison’s similarity problem asks whether a bounded representation of the multiplication in a $C^{\ast }$-algebra can be made compatible with adjoints by one bounded change of Hilbert-space coordinates. This manuscript claims an affirmative answer without separability assumptions. Its central estimate controls every matrix amplification of a commutator by its scalar norm. The proof passes through cyclic modules, free-product words, and a Fourier compression whose trace factors cancel.

🏷️ Similarity and complete boundedness

A unital complex $C^{\ast }$-algebra $A$ has an adjoint operation $a\mapsto a^{\ast }$ and norm satisfying $\Vert a^{\ast }a\Vert =\Vert a{\Vert }^2$. A bounded unital algebra homomorphism $\pi :A\to \mathcal{B}(H)$ is complex-linear, multiplicative, and sends $1$ to the identity, but is not initially assumed to preserve adjoints.

It is similar to a $\ast$-homomorphism if some bounded invertible $S$ makes $\rho (a)=S\pi (a)S^{-1}$ preserve adjoints for every $a$. Equivalently, one positive invertible metric operator $D=S^{\ast }S$ satisfies

$$D\pi (a)=\pi (a^{\ast }{)}^{\ast }D(a\in A).$$

Thus a single equivalent Hilbert-space inner product corrects the whole representation.

For a linear map $u$ on an operator algebra, its matrix amplification $u_h$ applies $u$ entrywise to $h$-by-$h$ matrices. The completely bounded norm is $\Vert u{\Vert }_{\mathrm{cb}}={\sup }_{h\ge 1}\Vert u_h\Vert$. Ordinary boundedness controls only $h=1$; complete boundedness requires a bound independent of matrix size.

Similarity conclusion claimed by the manuscript

Every bounded complex-linear unital algebra homomorphism from any unital complex $C^{\ast }$-algebra to $\mathcal{B}(H)$, for any complex Hilbert space $H$, is similar to a $\ast$-homomorphism by a bounded invertible operator. No faithfulness, normality, nuclearity, or separability hypothesis is imposed. The conclusion is existence of one similarity; no optimal condition-number estimate is asserted. (OpenAI, 2026)

The manuscript obtains this through a uniform derivation estimate. A derivation obeys a Leibniz rule; an inner derivation is a commutator $a\mapsto [Y,a]=Ya-aY$. For a unital represented von Neumann algebra $P\subset \mathcal{B}(K)$, set

$$g_P(Y)=\underset{a\in P,\Vert a\Vert \le 1}{\sup }\Vert [Y,a]\Vert .$$

The claimed quantitative theorem gives an absolute $C<\infty$ such that

$$\begin{array}{c}\Vert [Y^{(h)},X]\Vert \le C{g}_P(Y)\Vert X\Vert ,\\ X\in M_h(P),\end{array}$$

for all $P,K,Y,h$. Here $Y^{(h)}$ acts diagonally on $K^h$. Uniformity in the algebra, representation, and matrix size is the issue, rather than boundedness of any one finite matrix test.

🏷️ Cyclic domains and rectangular derivations

For unital $\ast$-representations $\sigma :A\to \mathcal{B}(K)$ and $\lambda :A\to \mathcal{B}(F)$, a bounded rectangular derivation takes values in $\mathcal{B}(K,F)$ and satisfies

$$\Delta (ab)=\Delta (a)\sigma (b)+\lambda (a)\Delta (b).$$

The domain representation is cyclic if some vector has dense orbit span under $\sigma (A)$. The output representation may have arbitrary multiplicity.

The source begins with an established uniform row estimate for bounded homomorphisms. Applying it also to $a\mapsto \phi (a^{\ast }{)}^{\ast }$ gives the corresponding column estimate. To transfer these bounds to $\Delta$, put $d=\Vert \Delta \Vert >0$ and use the triangular homomorphism

$$\Phi (a)=\left(\begin{array}{cc}\lambda (a)& d^{-1}\Delta (a)\\ 0& \sigma (a)\end{array}\right).$$

Its norm is at most two. Compressing its row and column amplifications gives row and column norms of $\Delta$ bounded by $c_{\mathrm{row}}d$, with $c_{\mathrm{row}}=4\sqrt{2}$. The zero derivation is treated separately.

Why does a cyclic domain allow all square matrices? Every unit vector $\xi \in K^h$ can be written $\xi =v\zeta$, where $\zeta \in K$ is a unit vector and $v$ is a contraction column over $\sigma (A{)}^{\prime \prime }$. The proof represents the sum of the coordinate vector functionals on the commutant by one vector, then constructs the corresponding intertwining column.

Matrix Kaplansky density approximates $v$ strongly by columns from $\sigma (A)$. Lift them to columns $w$ over $A$ of norm at most two. The Leibniz rule gives

$$\begin{array}{rl}{\Delta }_h(X){\sigma }_{h,1}(w)& ={\Delta }_{h,1}(Xw)\\ & -{\lambda }_h(X){\Delta }_{h,1}(w).\end{array}$$

Column bounds now control its action on $\zeta$ uniformly in $h$. Passing to the strong limit proves $\Vert \Delta {\Vert }_{\mathrm{cb}}\le 4c_{\mathrm{row}}\Vert \Delta \Vert$. The lift argument does not require $\Delta$ to vanish on $\ker \sigma$.

The established completely bounded similarity theorem, applied to $\Phi$, supplies a metric $D$. Its off-diagonal block yields an implementer $V$:

$$\begin{array}{c}\Delta (a)=V\sigma (a)-\lambda (a)V,\\ \Vert V\Vert \le c_0\Vert \Delta \Vert ,\\ c_0=(1+4c_{\mathrm{row}}{)}^2.\end{array}$$

This cyclic result is proved before the unrestricted similarity theorem; using the established completely bounded theorem here does not assume ordinary bounded homomorphisms are already similar. (OpenAI, 2026)

🏷️ A free-product corner estimate

For a finite tracial algebra, $\Vert x{\Vert }_2=\tau (x^{\ast }x{)}^{1/2}$ gives its trace Hilbert space. Left and right multiplication are $L_a\xi =a\xi$ and $R_b\xi =\xi b$. They commute; the commutant of all left multiplications consists of the bounded right multipliers.

Let $D$ be a ${\mathrm{II}}_1$ factor generated by increasing full matrix algebras of unbounded sizes, and let $A_0$ be a finite algebra with faithful normal tracial state. Their tracial free product $\mathcal{L}=D\ast A_0$ has alternating centered reduced words as its orthogonal Hilbert-space model. Freeness means those alternating centered products have trace zero.

Suppose $G\in \mathcal{B}(L^2(\mathcal{L}))$ commutes with $L_D$. Fix a projection $e\in D$ of trace $t=1/N$, $N\ge 2$, and assume

$$\Vert [G,L_x]\Vert \le \epsilon \Vert x\Vert (x\in e\mathcal{L}e).$$

Because equal-trace projections of $D$ are unitarily conjugate and $G$ commutes with $D$, this hypothesis transfers to every such corner.

The central free-product theorem bounds the compression of $[G,L_s]$ to $L^2(A_0)$ by $c_1\epsilon$ whenever $s\in A_0$ is a trace-zero self-adjoint unitary. Its constant

$$c_1=6{\pi }^2(1+4c_0)$$

is independent of the small corner trace. The next three mechanisms explain this independence. (OpenAI, 2026)

Symmetrization and repeated-letter words

Average $G$ over the compact unitary groups of the matrix stages, acting by conjugation on $D$ and trivially on $A_0$. A weak operator limit gains those symmetries while preserving the desired compressed commutator exactly.

Group words by their number $k$ of internal centered $D$ letters. An invariant-tensor lemma shows that the symmetrized operator preserves these levels. On a word with a repeated centered $D$ letter, it acts only on the remaining $A_0$ factors, through an operator $B_k$ independent of that repeated letter.

There are two norm restrictions behind the lemma. On a finite matrix stage, conjugation-invariant coefficients are combinations of products of traces around permutation cycles. The distinguished endpoint has an $L^1$ bound, while the other letters have $L^2$ bounds. Testing on matrix units and increasing the matrix size eliminates long cycles. The remaining pairings within the source side or within the test side would grow too quickly on sums of orthonormal tensors. Boundedness eliminates those as well. Only source-to-test permutations survive, requiring equal word levels.

These are full-output identities, also for $G^{\ast }$, rather than identities after a favorable compression. That distinction permits uniform control on sums of levels later.

Propagation without a word-length loss

For any right-support projection $r\in \mathcal{L}$ with $\tau (r)\le t$, the cyclic-domain lemma produces a bounded right multiplier approximating $G$ on the entire column:

$$\begin{array}{c}\Vert (G-R_b){|}_{L^2(\mathcal{L})r}\Vert \le c_0\epsilon ,\\ b=rb.\end{array}$$

To see why the trace does not cost a factor $N$, first work on $e{L}^2(\mathcal{L})r$, which is a cyclic corner module by projection comparison. Matrix units in $D$ identify the full column with $N$ copies of this corner column. Since $G$ commutes with them, extending the intertwiner is diagonal amplification and preserves its norm.

The source applies this to a vector with a small left support $q\in D$. Normalized high powers of a Haar unitary in $qDq$ serve as repeated centered letters. Their weak convergence kills the scalar component that could prevent left multiplication from being reduced-word concatenation. For each fixed word length, the remaining prefix acts isometrically. The result is

$$\begin{array}{rl}& \Vert B_k(s^{\otimes k}\otimes v)-s^{\otimes k}\otimes B_{0}v\Vert \\ & \le 2c_0\epsilon \Vert v{\Vert }_2.\end{array}$$

The bound is uniform in $k$. The auxiliary weak limit is taken after fixing the length and support; no convergence rate uniform in those choices is claimed. Orthogonality of the levels then makes their error norms add in squares, giving an operator bound on their full direct sum.

A Hankel test with cancelling trace factors

Choose $p\in D$ with $e\le p$ and $\tau (p)=1/2$. Put $d=2p-1$, $w=ds$, and $u={\tau }_p(e)=2t$, where ${\tau }_p$ is normalized corner trace. The free symmetries $d,s$ generate the infinite-dihedral word model; $w$ is Haar.

The corner generated by them identifies with even functions $f$ on the circle through $f(w)p$. For real mean-zero $f$, set $h=ef(w)pe$. Free compression gives

$$\Vert h\Vert \le 2\sqrt{u}\Vert f{\Vert }_{L^2}+u\Vert f{\Vert }_{\infty }.$$

An orthogonal sequence of normalized word embeddings identifies a subspace with ${\ell }^2({\mathbb{N}}_0)\otimes L^2(A_0)$. On it, the compressed $L_h$ has Toeplitz-plus-Hankel form

$$u(T_f\otimes I+H_f\otimes L_s),$$

where $(T_f{)}_{jk}={\gamma }_{j-k}$, $(H_f{)}_{jk}={\gamma }_{j+k+1}$, and ${\gamma }_l=\int z^{l}f(z)dm(z)$. The normalized expectation calculation supplies precisely the factor $u$, not $u^2$.

The propagated operator is $J=I\otimes B_0$. The Toeplitz term commutes with $J$, leaving a commutator norm equal to $u\Vert H_f\Vert \Vert [B_0,L_s]\Vert$. Propagation and the original corner hypothesis bound the same norm by $(1+4c_0)\epsilon \Vert h\Vert$.

Now take $f(e^{i\theta })$ to be the indicator of an arc of half-width $u/4$, minus its mean. Its $L^2$ norm is at most $\sqrt{u}$, so $\Vert h\Vert \le 3u$. Its Fourier coefficients with positive index are $\sin (lu/4)/(\pi l)$. On an $N$-by-$N$ Hankel block, testing the constant unit vector gives $\Vert H_f\Vert \ge 1/(2{\pi }^2)$, uniformly in $N$.

Both sides contain the same positive factor $u$. Cancelling it proves the corner estimate with constant $c_1$. This is the point at which a shrinking trace becomes compatible with a fixed commutator bound. (OpenAI, 2026)

🏷️ Transfer to arbitrary represented algebras

For a separable-predual ${\mathrm{II}}_1$ factor $M$, first consider finitely many standard copies $L^2(M{)}^r$. A matrix contraction can be placed in the off-diagonal block of a trace-zero self-adjoint unitary in a larger matrix algebra. Controlling commutators with that symmetry controls the original matrix test.

Average the tested operator over an irreducible increasing-matrix subfactor of $M$. This changes it by at most its scalar commutator norm. An imported relative-independence theorem then places another increasing-matrix subfactor, free from the constant ambient algebra, in a tracial ultrapower. This realizes the preceding free-product configuration.

Two ultrapowers must be distinguished. The tested bounded operator acts coordinatewise on a Hilbert-space ultrapower. The space $L^2(P^{\omega })$ of the tracial algebra ultrapower is a closed subspace of it, and the argument compresses to that subspace. The algebra action factors through its trace-null quotient there; this is not asserted on the entire Hilbert ultrapower. A second compression to the free-product space preserves the commutator coefficients on constant vectors exactly.

The resulting estimate is independent of both the original multiplicity $r$ and the test size. Countable sums and reducing-subspace embeddings cover arbitrary normal factor representations on separable Hilbert spaces. Properly infinite factors use orthogonal isometries to encode matrix tests, while finite type-I factors use compact-unitary averaging.

Central averaging then makes the operator decomposable over factor fibers. Countably many dense contraction tests establish one common conull set on which the scalar bound holds, permitting the factor estimate to be applied. Finally each finite matrix test and vector lies in a separable subspace reducing the finitely generated operator algebra needed for that test. Quotient norms and Kaplansky density retain the bound there. Taking the supremum over vectors removes separability entirely. (OpenAI, 2026)

🏷️ From commutators to similarity

The uniform theorem controls inner derivations. To handle a general bounded derivation $\Delta$ for a $\ast$-representation $\sigma$, decompose its Hilbert space into cyclic reducing subspaces $K_i$. On each input column, the cyclic result gives an implementer $V_i:K_i\to K$.

For a finite set $F$ of columns, let $q_F$ project onto their sum and extend the sum of implementers to an operator $V_F$ on $K$. The exact identity is

$$[V_F,\sigma (a)]=\Delta (a)q_F.$$

The norms $\Vert V_F\Vert$ need not stay bounded. Their scalar commutator norms do: $g_{\sigma (A{)}^{\prime \prime }}(V_F)\le \Vert \Delta \Vert$. Applying the universal estimate and taking the strong limit $q_F\to I$ yields $\Vert \Delta {\Vert }_{\mathrm{cb}}\le C\Vert \Delta \Vert$. The completely bounded triangular-metric argument now implements $\Delta$ globally.

The final imported criterion is Kirchberg’s per-algebra equivalence: innerness of all bounded derivations in every faithful unital $\ast$-representation is equivalent to the similarity property for all bounded unital homomorphisms of that algebra. Applying it completes the claimed conclusion. The faithful representation testing derivations is not the original possibly nonfaithful homomorphism.

A uniformly bounded group representation is a different input: it need not extend boundedly to the full or reduced group $C^{\ast }$-algebra. Thus the theorem does not automatically unitarize every uniformly bounded group representation. (OpenAI, 2026)

🏷️ Universal hyperreflexivity

For a unital von Neumann algebra $M\subset \mathcal{B}(H)$, define the off-diagonal seminorm

$${\alpha }_M(T)=\underset{\begin{array}{c}e\in M^{\prime }\\ e=e^{\ast }=e^2\end{array}}{\sup }\Vert (1-e)Te\Vert .$$

The projections in the commutant $M^{\prime }$ are exactly the projections onto invariant, hence reducing, subspaces of $M$. Hyperreflexivity asks whether this detectable failure to preserve subspaces bounds the distance from $T$ to $M$.

Both $e$ and $1-e$ occur in the supremum, so $\Vert [T,e]\Vert \le {\alpha }_M(T)$. Spectral approximation gives a factor two for self-adjoint contractions in $M^{\prime }$, and real/imaginary decomposition gives $g_{M^{\prime }}(T)\le 4{\alpha }_M(T)$. The established commutant-distance formula and the uniform theorem yield

$$\begin{array}{rl}2\mathrm{dist}(T,M)& =\Vert {\mathrm{ad}}_T{|}_{M^{\prime }}{\Vert }_{\mathrm{cb}}\\ & \le 4C{\alpha }_M(T).\end{array}$$

Thus one hyperreflexivity constant $2C$ works for every unital von Neumann algebra on every Hilbert space. No optimality is asserted. (OpenAI, 2026)

🏷️ Review scope

The README, main file, seven principal sections, and family scope page were read. The exposition follows the cyclic implementation, invariant-tensor restrictions, propagation, Hankel normalization, ultrapower transfer, and arbitrary-derivation arguments. The source review does not independently certify the complete theorem or the cited row, similarity, independence, standard-form, and distance theorems.

🏷️ Key steps and proof outlines

Encode a whole matrix test into one operator

The finite-factor estimate is the difficult analytic input. The source then extends it to all factor types with elementary operator identities. In a properly infinite factor $P$, choose isometries $v_1,\dots ,v_h$ with orthogonal ranges and write $V=[v_1\cdots v_h]$. Then $V^{\ast }V=I$, so $X\mapsto VX{V}^{\ast }$ preserves the norm of every $X\in M_h(P)$.

Let $\delta =g_P(Y)$ and let $c_{\mathrm{row}}$ be the earlier row/column derivation constant. Those estimates give

$$\Vert YV-V{Y}^{(h)}\Vert \le c_{\mathrm{row}}\delta ,\Vert Y^{(h)}{V}^{\ast }-V^{\ast }Y\Vert \le c_{\mathrm{row}}\delta .$$

Expanding one scalar commutator gives

$$\begin{array}{rl}[Y,VX{V}^{\ast }]=& (YV-V{Y}^{(h)})X{V}^{\ast }\\ & +V[Y^{(h)},X]V^{\ast }\\ & +VX(Y^{(h)}{V}^{\ast }-V^{\ast }Y).\end{array}$$

The first and third terms each cost $c_{\mathrm{row}}\delta \Vert X\Vert$, while the left side costs $\delta \Vert X\Vert$. Isometry of the middle term yields the matrix bound with constant $1+2c_{\mathrm{row}}$, independent of $h$. Surjectivity of $V$ is unnecessary. (OpenAI, 2026)

For a finite type-I factor, compact-unitary averaging instead produces $Y_0\in P^{\prime }$ with $\Vert Y-Y_0\Vert \le \delta$. Then $[Y^{(h)},X]=[(Y-Y_0{)}^{(h)},X]$ costs at most $2\delta \Vert X\Vert$. These two reductions isolate precisely where the finite type-II argument is needed.

Pass a scalar bound to almost every factor simultaneously

For a general algebra on a separable Hilbert space, average $Y$ over finite families of central projections and take a weak operator cluster limit $Y_1$. Since each averaging unitary belongs to $P$,

$$\Vert Y_1-Y\Vert \le \delta ,g_P(Y_1)\le 3\delta .$$

The operator $Y_1$ commutes with the center and is therefore decomposable over factor fibers. A separate null set for each of uncountably many contractions would not prove a fiberwise scalar bound. The source avoids this by taking countably many rational $\ast$-polynomials $b$ in measurable generating sections and normalizing them to contractions

$$c_b=b(1+b^{\ast }b{)}^{-1/2}.$$

This transformation parametrizes the open unit ball: its inverse there is $c\mapsto c(1-c^{\ast }c{)}^{-1/2}$. Hence these tests are norm dense in the generated algebra’s contraction ball, and Kaplansky density makes them strongly dense in each factor’s contraction ball. Remove their countably many exceptional null sets once. Strong approximation then passes the bound $3\delta$ to every fiber contraction on the same conull set.

Applying the factor estimate and restoring $Y$ gives the global constant $3C_{\mathrm{fac}}+2$. The final two comes from $\Vert [(Y-Y_1{)}^{(h)},X]\Vert \le 2\delta \Vert X\Vert$. This shows exactly how uniformity survives central decomposition.

Unbounded implementer norms do not obstruct the final limit

For an arbitrary bounded derivation $\Delta$, decompose the representation into cyclic reducing input spaces $K_i$. On each, the cyclic theorem supplies an implementer $V_i:K_i\to K$. For a finite set $F$ of columns, let $q_F$ be the corresponding reducing projection and define $V_F$ by those columns, extending it by zero on the others.

The Leibniz implementation identity gives

$$[V_F,\sigma (a)]=\Delta (a)q_F.$$

Even though $\Vert V_F\Vert$ may grow with $F$, its scalar commutator norm stays at most $\Vert \Delta \Vert$. Apply the established universal commutator estimate to this bounded operator $V_F$ for each finite $F$. It bounds all amplified maps ${\Delta }_h(X)q_F^{(h)}$ by the same $C\Vert \Delta \Vert \Vert X\Vert$. Since $q_F\to I$ strongly, the bound passes to ${\Delta }_h(X)$, uniformly in $h$.

Complete boundedness, rather than convergence of the implementers themselves, is the output needed for the triangular-metric theorem. That theorem implements the full derivation, and the cited derivation criterion then supplies similarity of bounded homomorphisms. The global section proves these reductions; the earlier free-product propagation and Hankel estimate are what make their common constant available.

See Also

• Free group factors and generating limits develops free-product norm decompositions and the distinction between operator norm and tracial $L^2$ norm.
• Complete Crouzeix bounds provides another setting where an estimate must remain uniform under matrix amplification.

🏷️ Manuscripts and verification scope

• Kadison’s similarity theorem through uniform derivation estimates — pinned manuscript and companion source.

The family scope page describes selected formalization coverage. The documented selection covers Kadison similarity on arbitrary Hilbert spaces, universal hyperreflexivity for unital von Neumann algebras, and a commutator bound uniform over all finite matrix amplifications. 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. Kadison’s similarity theorem through uniform derivation estimates.