Can every bounded representation be made to respect adjoints?

← All digests

Family 288 · Operator algebras
Importance 5/5 · Provisional editorial rating
Sources consulted 2026-10-25

1. The question

Matrices carry two kinds of structure: multiplication and the adjoint operation, which transposes a matrix and conjugates its entries. A map can preserve multiplication without preserving adjoints. Sometimes the apparent mismatch comes only from using the wrong inner product.

Kadison’s similarity problem asks whether this is always the explanation for a bounded representation of a C*-algebra. Can one bounded, invertible change of coordinates make every represented adjoint correct simultaneously?

Family 288 claims a positive answer without restrictions on the algebra or the Hilbert space beyond the stated complex, unital setting. Its central estimate also gives a universal way to measure how close an operator is to an arbitrary von Neumann algebra by testing how badly it mixes invariant subspaces.

2. Background and essential definitions

A C-algebra* is a complete normed algebra with an adjoint operation satisfying $\|a^*a\|=\|a\|^2$. Bounded operators on a Hilbert space form a basic example. A unital algebra homomorphism $\pi:A\to B(H)$ preserves multiplication, complex linear combinations and the identity. Boundedness means that one constant controls $\|\pi(a)\|$ by $\|a\|$.

An adjoint-preserving homomorphism additionally satisfies $\pi(a^*)=\pi(a)^*$. Two representations are similar if one is obtained from the other by conjugating with a bounded invertible operator $S$:

$$\rho(a)=S\pi(a)S^{-1}.$$

Such a change corresponds to replacing the original Hilbert norm by the equivalent norm $\|\xi\|_S=\|S\xi\|$. “Equivalent” means that each norm bounds the other by fixed constants; it excludes a singular or unbounded change of geometry.

For a linear map $u$, its matrix amplification $u^{(n)}$ applies $u$ entry by entry to $n$-by-$n$ matrices. It is completely bounded if

$$\|u\|_{\mathrm{cb}}=\sup_{n\geq1}\|u^{(n)}\|<\infty.$$

Ordinary boundedness controls individual elements. Complete boundedness additionally controls all finite matrix arrangements of elements with one constant. The distinction is central to the problem.

A derivation behaves like differentiation of a product: for a represented algebra, $d(ab)=d(a)\pi(b)+\pi(a)d(b)$. It is inner if it has the form $d(a)=Y\pi(a)-\pi(a)Y$. The difference $[Y,X]=YX-XY$ is called a commutator and measures failure to commute.

3. Development history and previous partial results

Kadison formulated the similarity problem in 1955. The question concerns one simultaneous coordinate change for an entire representation, rather than separately adjusting each operator. Multiplicativity is a strong constraint, but ordinary boundedness does not directly provide the matrix-level estimates needed to construct that change.

Bunce and Christensen established the nuclear C*-algebra case. Nuclearity is a finite-dimensional approximation property, and it gives especially strong similarity estimates. Haagerup solved the cyclic representation case, where one vector generates a dense subspace under the representation, and linked similarity to complete boundedness. He also obtained the similarity property for C*-algebras without tracial states. Christensen covered type $\mathrm{II}_1$ factors with property Gamma, a condition providing suitable approximately central unitaries.

Paulsen’s similarity theorem gives the general completely bounded homomorphism criterion and controls the best similarity condition number. Christensen established the corresponding criterion for derivations. Kirchberg then connected the similarity property of a given C*-algebra with innerness of derivations in its faithful *-representations. These results turn the original coordinate question into a problem about uniform operator estimates.

Pisier organized quantitative versions using similarity degree and matrix factorization length. More recently, work of Eleftherakis and Paulsen connected the similarity problem to hyperreflexivity of von Neumann algebras: whether distance to an algebra can be controlled by invariant-subspace tests.

The remaining difficulty addressed by M1 is uniformity across finite factors and their representations. Good estimates for a fixed matrix size or a favorable cyclic representation do not automatically survive arbitrary multiplicity and unbounded matrix size.

4. What the new manuscripts claim

Similarity in the unrestricted unital setting

M1 states that for every unital complex C*-algebra $A$, every complex Hilbert space $H$, and every bounded complex-linear unital algebra homomorphism $\pi:A\to B(H)$, there is a bounded invertible $S$ such that

$$S\pi(a^*)S^{-1}=\bigl(S\pi(a)S^{-1}\bigr)^*$$

for every $a\in A$.

There is no separability, nuclearity, finite-generation, faithfulness or normality assumption. The operator $S$ may depend on the whole homomorphism. The headline theorem asserts existence; it does not state an optimal universal bound for $\|S\|\|S^{-1}\|$.

A uniform estimate at every matrix size

For a represented unital von Neumann algebra $P\subseteq B(K)$, define

$$g_P(Y)=\sup_{a\in P,\,\|a\|\leq1}\|Ya-aY\|.$$

M1 claims an absolute constant $C<\infty$ such that, for every positive integer $h$ and every matrix $X\in M_h(P)$,

$$\|[Y^{(h)},X]\|\leq C\,g_P(Y)\,\|X\|.$$

Here $Y^{(h)}$ is diagonal repetition of $Y$ on $K^h$. The constant is independent of the algebra, its representation, the Hilbert space and $h$. Scalar commutator control therefore controls all amplifications.

A universal distance estimate

Let $M'$ denote the commutant of $M$, consisting of operators commuting with every element of $M$. For $T\in B(H)$, set

$$\alpha_M(T)=\sup_{e=e^*=e^2\in M'}\|(1-e)Te\|.$$

The projection $e$ selects a reducing subspace for $M$; the displayed block measures how much $T$ sends that subspace outside itself. M1 states

$$\operatorname{dist}(T,M)\leq2C\,\alpha_M(T).$$

This is universal hyperreflexivity. No optimality of $2C$ is claimed.

5. Why the advance matters

Importance: 5/5. If correct, the theorem resolves a long-standing structural question about all bounded representations of C*-algebras. It says that a representation preserving algebraic multiplication cannot possess an irreparable mismatch with the adjoint, provided it is bounded.

The result also unifies different languages. Similarity is a coordinate-change problem; complete boundedness concerns matrices of operators; derivations express infinitesimal algebraic behavior; hyperreflexivity concerns distance detected by subspaces. M1’s universal estimate would connect all of them without special assumptions on the represented algebra.

The distance conclusion has an especially concrete interpretation. If an operator almost respects every reducing subspace of $M$, it lies close to an element of $M$, with a constant independent of the particular algebra. This is stronger than merely knowing which operators respect all those subspaces exactly.

6. The main idea

First control cyclic pieces

The argument begins with derivations between two representation spaces when the domain representation is cyclic. Combining row and column estimates produces a bounded implementing operator with controlled norm. This is a local building block: arbitrary representations need not be cyclic, but can be analyzed through collections of cyclic pieces.

The obstacle is preventing constants from growing when those pieces are combined or when the representation is amplified to large matrices. M1 devotes its central argument to a dimension-independent estimate in a finite tracial setting.

Enlarge the algebra with free variables

A tracial free product joins two operator algebras so that alternating centered products have trace zero. M1 uses one algebra assembled from increasing full matrix algebras and another arbitrary finite tracial algebra. Their free-product word structure supplies a Hilbert-space decomposition on which commutator information can be tracked.

The key intermediate statement starts with control in a small corner, cut out by a projection of trace $1/N$, and concludes an estimate for compressed commutators with trace-zero self-adjoint unitaries. Crucially, its constant is independent of $N$.

Averaging over matrix unitary groups constrains the possible actions on words. Estimates on long words then propagate local control. The final comparison uses a Hankel matrix, whose entries depend on the sum of their indices; its lower bound cancels the small trace factors in the upper estimate. In the manuscript’s account, this cancellation prevents the bound from deteriorating as the corner becomes tiny.

Transfer the estimate and return to similarity

Popa’s relative independence theorem places the needed free-product configuration inside a tracial ultrapower after matrix amplification. An ultrapower packages a bounded sequence of approximations into a single limiting algebra. M1 passes the relevant operator through a Hilbert-space ultrapower and compresses it while retaining the required commutator data.

The argument next moves from separable finite factors to arbitrary normal factor representations, then uses central decomposition and separable reducing subspaces to remove the remaining restrictions. This is why the theorem’s arbitrary-Hilbert-space scope needs more than its initial tracial calculation.

Finally, cyclic implementations of a derivation are assembled using the uniform commutator bound. The derivation becomes completely bounded, established implementation results make it inner, and Kirchberg’s equivalence yields similarity. Applying the same estimate to a commutant gives the hyperreflexivity consequence through the standard commutant-distance formula.

7. Scope, limitations, and open questions

The statement requires a bounded complex-linear unital algebra homomorphism from a unital complex C*-algebra. It does not assert that every algebraic representation is automatically bounded, nor that the original representation already preserves adjoints. The necessary similarity can be nonunitary.

A uniformly bounded representation of a discrete group is another matter. Such a representation automatically extends to the group’s $\ell^1$-algebra, but need not extend boundedly to its full or reduced group C*-algebra. Consequently M1’s theorem does not make every uniformly bounded group representation unitarizable. The manuscript explicitly separates that issue from Kadison similarity.

The universal constant concerns commutators and hyperreflexivity; the paper does not claim its optimal value or an optimal condition-number formula for the resulting similarity. Quantitative improvements and concrete construction of the best equivalent inner product remain distinct questions.

The exposition reports the uniform estimates and their stated dependencies. The passage through free products and ultrapowers is part of the manuscript’s argument, not an independent verification of the claimed solution.

8. Sources and further reading

• Ozawa, An Invitation to the Similarity Problems, for the homomorphism, derivation and group-representation distinctions.
• Eleftherakis and Paulsen, Hyperreflexivity of von Neumann algebras and similarity of finitely generated C*-algebras, for the hyperreflexivity connection discussed by M1.
• M1’s introduction attributes the nuclear, cyclic, property-Gamma and complete-boundedness results and states the unrestricted hypotheses of its new estimate.

Source manuscripts

• M1. OpenAI, Kadison’s similarity theorem through uniform derivation estimates (September 23, 2026).

Sources consulted on 25 October 2026 at repository revision 1581d4c. Download the manuscript citations.