Kadison's Similarity Theorem Through
Uniform Derivation Estimates
A Guided Tour for Young Mathematicians
============================================================
This is a guided tour of a research paper. The paper (OpenAI,
September 23, 2026) claims to prove Kadison's similarity theorem:
every bounded unital algebra homomorphism from a C*-algebra into the
bounded operators on a complex Hilbert space is similar, by a bounded
invertible change of coordinates, to a star-preserving algebra
homomorphism. As a corollary the paper claims a uniform
hyperreflexivity constant for all von Neumann algebras.
Kadison asked this in 1955 and it has been open for seventy years.
The tour below describes the paper's argument in the order the paper
uses it. It explains what each ingredient is, why it is needed at
each step, and how the ingredients fit together. It does not
independently certify the paper's correctness -- a preprint is a
claim, and the reader should know the difference. The tour is
written so that a curious high-school student can follow the plot
and can later open the original and look up specific estimates by
name.
Every technical term is introduced the first time it is used. Each
key result is stated in plain language and, where it matters, also
in its formal shape so that the tour and the paper can be compared
side by side. The reading is meant to be done twice: once for the
plot, once for the details.
============================================================
PART I. THE QUESTION AND THE TWO THEOREMS
============================================================
1. WHAT IS A HILBERT SPACE, AND WHY DO WE CARE?
A Hilbert space is a complete complex vector space with an inner
product. The simplest examples are the finite-dimensional spaces
C^n. The most important infinite-dimensional example is the space
of square-summable sequences l^2(N) = {(a_1, a_2, ...) : sum |a_i|^2
< infinity}, with inner product <a, b> = sum a_i conj(b_i), where
conj is complex conjugation.
Quantum mechanics lives in Hilbert spaces: a state of a physical
system is a unit vector, and an observable is a self-adjoint
operator. Signal processing lives in Hilbert spaces: a signal is a
square-summable sequence and a filter is an operator on it. Every
example below is built from operators on Hilbert spaces.
We only use complex Hilbert spaces in this tour. Inner products are
linear in the first slot (not the second) throughout; this is a
convention that must be fixed because the paper does, and mixing
conventions silently reverses adjoints.
2. WHAT A C*-ALGEBRA IS.
Let H be a Hilbert space. Write B(H) for the set of bounded linear
operators T : H -> H, meaning there exists a constant c with
||T v|| <= c ||v|| for every vector v. B(H) is an algebra under
addition, scalar multiplication, composition, and the identity
operator 1. It carries an operator norm ||T|| = sup_{||v||<=1}
||T v||.
It also carries an involution: to each T there is an adjoint T^*,
defined by the identity <T^* u, v> = <u, T v> for all vectors u, v.
The adjoint obeys (T^*)^* = T, (S T)^* = T^* S^*, (S + T)^* = S^* +
T^*, and (c T)^* = conj(c) T^*. It also obeys the C*-identity ||T^*
T|| = ||T||^2, which is the algebraic shadow of the fact that it
came from an inner product.
A concrete C*-algebra is a subset A of some B(H) that:
(i) is closed under addition and multiplication by scalars and by
other elements,
(ii) contains 1,
(iii) is closed under the adjoint T -> T^*,
(iv) is closed in the operator norm: a norm limit of elements in A
also lies in A.
Examples:
- C itself, as the operators that multiply every vector by a
fixed complex scalar.
- The algebra C(X) of continuous complex-valued functions on a
compact space X; each f acts by multiplication on the Hilbert
space of square-integrable functions, and its adjoint is
pointwise conjugation.
- The algebra M_n(C) of n-by-n complex matrices, acting on C^n.
- The algebra B(H) itself.
An abstract C*-algebra is any algebra over C with an involution
satisfying the axioms above, without reference to a Hilbert space.
The Gelfand-Naimark theorem says the two notions agree: every
abstract C*-algebra is isomorphic, preserving adjoints and norms,
to a concrete one, uniquely up to an adjoint-preserving isomorphism.
So we may work concretely and lose nothing.
3. WHAT A VON NEUMANN ALGEBRA IS, AND WHAT A *-REPRESENTATION IS.
A von Neumann algebra is a concrete C*-algebra M inside some B(H)
that is in addition closed in a coarser topology: the strong
operator topology, in which a sequence T_n converges to T if T_n v
converges in norm to T v for every vector v.
There are three equivalent definitions:
(i) M equals its own double commutant: M = (M')', where M' is the
set of operators S with S T = T S for every T in M.
(ii) M is a strongly-closed unital *-subalgebra of B(H).
(iii) M is what you get from a generalized trace or measure
structure (the abstract statement due to Sakai).
Von Neumann algebras have a rich supply of projections (operators
p with p = p^* = p^2). C*-algebras need not. Two examples that will
recur:
- Type II_1 factors. A factor is a von Neumann algebra whose
center Z(M) = M and M' agree. A II_1 factor is an
infinite-dimensional factor with a faithful normal tracial
state tau: tau(T) is a scalar "expected value" that satisfies
tau(S T) = tau(T S) and tau(1) = 1. Matrices never provide II_1
factors; infinite constructions do.
- B(H) itself, which is a factor of type I (finite if H is finite
dimensional, infinite otherwise).
Every von Neumann algebra decomposes as a direct integral of
factors, so analyzing factors is enough in the end. Section 22 of
this tour says more.
4. THE SIMILARITY PROBLEM.
Let A be a unital C*-algebra and let pi : A -> B(H) be a unital
algebra homomorphism. Being the map of a C*-algebra, pi is often
assumed to be *-preserving: pi(a^*) = pi(a)^*. But this is an extra
assumption, not a consequence of nothing. Kadison's question is
whether the assumption is automatic, up to a change of coordinates.
Precisely: pi is similar to a *-representation if there exists an
operator S : H -> H that is invertible and whose inverse is bounded,
such that the map
a |-> S pi(a) S^{-1}
is a *-representation.
In words: pi might misalign the adjoint, but there is always a
change of basis S that re-aligns it. The similarity is allowed to
depend on pi and on the C*-algebra; the paper never claims a
similarity constant.
Why ask? Because countless operator-algebra computations -- of
norms, of positivity, of spectra -- assume that adjoints behave. If
a bounded homomorphism pi fails to preserve adjoints, all that
machinery breaks. Similarity says you can restore the machinery at
the cost of one fixed change of coordinates, after which everything
behaves.
Kadison, in 1955, asked whether every bounded unital homomorphism
has this property. For nuclear algebras it was settled by Bunce
and Christensen in 1981. For cyclic representations it was solved by
Haagerup in 1983. For II_1 factors with property Gamma, Christensen
in 1986. It has remained open in general -- if the present paper is
correct.
5. THE SECOND THEOREM: HYPERREFLEXIVITY.
Say M is a von Neumann algebra in B(H). Define the distance from an
arbitrary T in B(H) to M:
dist(T, M) = inf {||T - a|| : a in M}.
A classic way to test how far T is from M is to test on invariant
subspaces. If p is a projection in M' (equivalently, p is a
projection commuting with M), the "off-diagonal" pieces of T with
respect to p are p T (1 - p) and (1 - p) T p. If T lies in M, these
pieces vanish. Set
alpha_M(T) = sup { ||(1 - p) T p|| : p a projection in M' }.
Reflexivity says: alpha_M(T) = 0 if and only if T is in M.
Hyperreflexivity asks for a quantitative version: is there a constant
K, depending only on M, with
dist(T, M) <= K alpha_M(T) for all T ?
Arveson studied this property in the 1980s and established it for
some classes of algebras; whether every von Neumann algebra is
hyperreflexive has remained open. The paper claims:
EVERY von Neumann algebra on EVERY complex Hilbert space is
hyperreflexive, with the same universal constant 2C,
where C is the absolute constant of Theorem 1.2 below.
That is a wild strengthening of Arveson's result. It follows from
the same estimate that proves similarity, and it is quantitative
rather than illustrative: no optimality of 2C is claimed, only a
finite universal value.
The tour below tours the similarity theorem and then comes back to
hyperreflexivity in Section 26.
============================================================
PART II. THE TOOLBOX
============================================================
6. DERIVATIONS, AND WHY THEY CARRY THE WHOLE PROBLEM.
Let sigma : A -> B(K) be a *-representation. A sigma-derivation is a
linear map Delta : A -> B(K) satisfying
Delta(a b) = Delta(a) sigma(b) + sigma(a) Delta(b).
This is the algebraic "product rule", in the operator-algebra
setting. Delta is inner if there is a fixed operator V in B(K) with
Delta(a) = V sigma(a) - sigma(a) V for all a in A.
A theorem of Ringrose (1972) says every derivation from a C*-algebra
into a Banach bimodule is automatically continuous. So "bounded
derivation" is redundant; the only question is whether a given
derivation is inner.
Kirchberg's theorem (1996), in the per-algebra form used here:
for a unital C*-algebra A the following are equivalent --
(1) Every bounded unital homomorphism pi : A -> B(H), on every
Hilbert space H, is similar to a *-homomorphism.
(2) In every faithful unital *-representation sigma : A -> B(K),
every derivation Delta : A -> B(K) is inner.
This is the entire reduction of the similarity problem to a
statement about derivations. From here on, we only ever need to
prove that derivations are inner. The proof has two halves: (a)
show derivations are completely bounded with a good bound; (b)
apply Paulsen's theorem to realize a completely bounded derivation
by a single operator. Sections 7, 8, 9, 24 explain each step.
7. COMPLETE BOUNDEDNESS, AND PAULSEN'S THEOREM.
Let A be a C*-algebra and let u : A -> B(F) be a linear map. For each
positive integer h, define the amplification u_h : M_h(A) -> M_h(B(F))
by acting entrywise:
u_h( (a_ij)_{i,j=1..h} ) = ( u(a_ij) )_{i,j=1..h}.
Here M_h(A) is the algebra of h-by-h matrices with entries in A,
carrying the natural operator norm coming from A acting on K.
Write ||u||_h for the operator norm of u_h. u is completely bounded
if the norm
||u||_cb = sup_{h >= 1} ||u||_h
is finite. Being bounded is ||u||_1 < infinity; being completely
bounded is stronger. Many naturally-occurring maps turn out to be
completely bounded (all *-homomorphisms, all compressions), and
some naturally-occurring ones do not.
Three facts we will use.
(A) EVERY *-HOMOMORPHISM IS COMPLETELY CONTRACTIVE.
If phi : A -> B(H) is a unital *-homomorphism, ||phi||_cb <= 1.
(B) PAULSEN'S THEOREM (1984).
Let pi : A -> B(H) be a unital bounded homomorphism. Then pi is
similar to a *-homomorphism if and only if ||pi||_cb < infinity.
The optimal similarity condition number -- the infimum of
||S|| ||S^{-1}|| over admissible S -- equals ||pi||_cb.
Moreover, the similar *-homomorphism may be taken completely
contractive.
(C) CHRISTENSEN'S THEOREM (1977).
A derivation Delta : A -> B(K) into a faithful *-representation
is inner if and only if it is completely bounded.
Combining (B) and (C) with Kirchberg's equivalence (Section 6):
Similarity of every pi <=> every derivation inner
<=> every derivation cb
<=> cb norm of every derivation
bounded by C ||Delta||.
The last equivalence is what the paper actually proves, with an
explicit absolute C.
8. ROW AND COLUMN BOUNDS.
Two more norms appear in the proof. For a bounded unital map
phi : A -> B(K), define
||phi||_row = sup_h ||phi_{1, h}||, where phi_{1, h} sends a row
[a_1 ... a_h] in M_{1,h}(A) to the row [phi(a_1) ... phi(a_h)].
||phi||_col = sup_h ||phi_{h, 1}||, defined analogously for
columns.
Here M_{p,q}(A) is the space of p-by-q matrices with entries in A,
carrying the norm it inherits from the representation on p copies of
K tensored with the natural matrix structure. A row map and a
column map are two ways of testing a map beyond the square ones.
Dickson's lemma (2014), building on a row estimate of Haagerup and
using Haagerup's little Grothendieck inequality:
||phi||_row <= sqrt(2) ||phi||^2,
||phi||_col <= sqrt(2) ||phi||^2.
The second follows from the first by transposing: phi^#(a) =
phi(a^*)^* is again a unital bounded homomorphism, and its row norm
is phi's column norm.
We will use this as: any bounded unital homomorphism of norm at
most 2 has row and column norms at most 4 sqrt(2). Set
c_row = 4 sqrt(2), so that c_row >= ||phi||_row, ||phi||_col
whenever ||phi|| <= 2.
9. CYCLIC DOMAINS AND BOUNDED IMPLEMENTERS.
Let sigma : A -> B(K) be a *-representation. sigma has a cyclic
vector if there is a single vector xi in K such that the set
{sigma(a) xi : a in A} is dense in K. The whole Hilbert space is
then "generated" from xi by the algebra, and plenty of
representations have this property.
The cyclic-domain lemma (paper: Lemma 2.6) concerns a *-
representation sigma : A -> B(K) with a cyclic vector and any *-
representation lambda : A -> B(F) as the output (no cyclicity
assumed on the output side). Let Delta : A -> B(K, F) be a bounded
rectangular derivation: an operator-valued map with
Delta(a b) = Delta(a) sigma(b) + lambda(a) Delta(b),
taking values in the bounded operators from K to F. Then:
(a) Delta is completely bounded, with
||Delta||_cb <= 4 c_row ||Delta||,
(b) Delta is implemented: there is V in B(K, F) with
Delta(a) = V sigma(a) - lambda(a) V for all a, and
||V|| <= c_0 ||Delta||, where c_0 = (1 + 4 c_row)^2.
The proof: pick a bounded matrix test, lift a bounded column of
implementers up from the algebra to the representation using
Kaplansky density, apply the row/column bound; then feed the cb
bound into Paulsen's theorem (which is where the c_0 = (1 + 4
c_row)^2 factor comes from, via the condition number identity
stated in Part II.7(B)).
This lemma is the tool that turns a corner-restricted hypothesis
([G, L_x] small for x in a corner) into a global implementation by a
single operator, in exactly the situation we care about, in
Sections 17 and 24.
============================================================
PART III. THE ONE INEQUALITY THAT HAS TO HOLD
============================================================
10. THE UNIFORM COMMUTATOR ESTIMATE.
Let P be a von Neumann algebra in B(K), let Y be any bounded operator
on K, and let
g_P(Y) = sup { ||Y a - a Y|| : a in P, ||a|| <= 1 }.
This is the "commutator size" of Y with P. If Y commutes with P, then
g_P(Y) = 0. If not, g_P(Y) reports how big a stir Y makes inside P.
It measures the inner derivation a |-> [Y, a] as a map on P.
For a positive integer h, let Y^(h) be the diagonal amplification of
Y on the direct sum K^h = K + K + ... + K (h copies):
Y^(h)(v_1, ..., v_h) = (Y v_1, ..., Y v_h).
Similarly X in M_h(P) acts on K^h entry-by-entry, as an h-by-h matrix
of elements of P. The bracket [T, U] = T U - U T is the commutator.
Here is the estimate at the heart of the whole paper. It is Theorem
1.2 of the paper.
UNIFORM COMMUTATOR ESTIMATE.
There is an absolute constant C < infinity, such that for EVERY
complex Hilbert space K, EVERY unital von Neumann algebra P in
B(K), EVERY Y in B(K), EVERY positive integer h, and EVERY
X in M_h(P),
||[Y^(h), X]|| <= C g_P(Y) ||X||.
The constant C is numerical. It does not depend on P, K, Y, h, X,
or any other quantity.
Read this way: at the scalar level we have measured g_P(Y), the
supremum of ||[Y, a]|| over a in P of norm one. The estimate says the
matrix-level version grows no worse than that scalar measurement.
The factor of ||X|| is a normalization; the entire content is that
there is a single C working for all matrix sizes h simultaneously.
Why does this imply everything we want? Because of Section 7. Take a
derivation Delta into a faithful *-representation sigma. Consider a
finite matrix test Delta_h(X) with X in M_h(A) of norm one. Then
Delta_h(X) q_F -- a column restriction -- turns out to be a commutator
[V_F^(h), sigma_h(X)] for a bounded operator V_F and a projection
q_F. If g_{sigma(A)''}(V_F) <= ||Delta|| (which is exactly a scalar
commutator bound coming from the definition of derivation), then the
estimate gives
||Delta_h(X) q_F|| <= C ||Delta|| ||sigma_h(X)|| <= C ||Delta||.
So Delta is completely bounded, and Paulsen + Christensen implements
it. To go from cb to inner via Kirchberg, all of the parallelism in
the cb norm is supplied by the single estimate -- this is what makes
Pisier's similarity degree equal to one here.
11. WHAT IS HARD ABOUT IT.
If X is scalar (h = 1), the estimate is exactly the definition of
g_P(Y):
||[Y, X]|| <= g_P(Y) ||X||,
which holds with C = 1. The difficulty is the matrix factor: h can be
1, 2, 3, ... without any bound. An operator Y whose scalar commutator
with P is small could in principle have commutator with a matrix X in
M_h(P) that grows with h. If it can, the whole scheme collapses.
Kadison's similarity problem for a specific algebra A, in its
derivation form, IS this statement for a specific class of Y coming
from a derivation: the question is whether the scalar commutator
bound controls the matrix ones uniformly. The paper's entire
contribution is: yes, it does, with C universal.
The proof strategy is to study the hardest case first (finite
factors), then push outward in three steps: factors in arbitrary
representations, then separable predual algebras, then arbitrary
algebras. Each step requires a small new ingredient. Our tour mimics
this order.
12. THE THREE-STEP REDUCTION, IN BRIEF.
The paper's plan, with the constant C = 3 C_fac + 2 where
C_fac = max { 3 c_1 + 2, 1 + 2 c_row, 2 },
c_1 = 6 pi^2 (1 + 4 c_0),
c_0 = (1 + 4 c_row)^2,
and c_row = 4 sqrt(2) as in Section 8.
STEP 1 (hardest). In a II_1 factor M with its standard trace
representation on L^2(M)^r for finite r, prove
||[Y^(h), X]|| <= (3 c_1 + 2) g_M(Y) ||X||.
The 3 c_1 + 2 is the "finite factors" constant; c_1 is a
free-product constant that the tour will spend all of PART IV
building.
STEP 2. In an arbitrary factor on a separable Hilbert space, use
cyclic decomposition and combinations of isometries to reduce to a
bound C_fac = max { 3 c_1 + 2, 1 + 2 c_row, 2 }. The 1 + 2 c_row
comes from isometries with orthogonal ranges; the 2 from averaging
over compact unitary groups of finite-dimensional factors; the
3 c_1 + 2 is STEP 1.
STEP 3. In an arbitrary von Neumann algebra, use central direct
integral decomposition into factors, then reduce to separable
reducing subspaces.
POST-PROCESSING. Obtain the full constant C = 3 C_fac + 2 by the
center-averaging step. Replacing Y by its average over the compact
unitary groups of the center introduces a factor of 3 on C_fac
(because g_P(Y_1) <= 3 g_P(Y)); replacing Y_1 back by Y adds a
fixed 2 to the multiplier (because ||Y - Y_1|| <= g_P(Y)).
The bulk of the tour is STEP 1. Steps 2 and 3 are structural
assemblies that take a few sections each. Section 10's estimate is
what we will be building from the bottom up.
============================================================
PART IV. THE HEART: THE FREE-PRODUCT CORNER ESTIMATE
============================================================
13. FREE PRODUCTS AND REDUCED WORDS.
Take two tracial von Neumann algebras: D (a II_1 factor) and A_0
(any finite von Neumann algebra with a fixed faithful normal trace
tau). Their free product with amalgamation over the scalars is
denoted L = D * A_0. It is again a tracial von Neumann algebra with
trace tau agreeing on each factor.
For our purposes, the most useful way to describe L is what it looks
like after we pass to L^2 with respect to tau. Then L^2(L) breaks up
as an orthogonal direct sum of "reduced word spaces". A reduced word
is an alternating string
d_0 a_1 d_1 a_2 ... a_k d_k z
with k >= 0, where d_0 lies in D unrestricted, z lies in A_0
unrestricted, the letters d_1, ..., d_k lie in D with zero trace,
the letters a_1, ..., a_k lie in A_0 with zero trace, and no two
letters from the same algebra touch after deleting the ones that are
zero. The scalar 1 is separated out as the k = 0, d_0 = 1, z = 1
term. The choice of "which letter is unrestricted" matters: the
tensor decomposition in Section 16 groups d_0 and d_1..d_k together
on the D-side and groups a_1..a_k and z together on the A_0-side,
so it is d_0 that is unrestricted on the D-side and z that is
unrestricted on the A_0-side.
Free-ness is exactly the statement that distinct reduced words are
orthogonal in L^2: the trace of a reduced word differing from
another is zero whenever the two words differ. This is the whole
content of "free" in "free product".
A concrete and useful example: if D contains a diffuse abelian
algebra (a subalgebra generated by a single inverter u with u^2 = 1
in various ways) and A_0 contains two trace-zero symmetries s and d
(operators with s^2 = d^2 = 1 and tau(s) = tau(d) = 0), then the free
product W^*(s, d) sitting inside L is the von Neumann algebra of the
infinite dihedral group -- it stays "small" but never dies. If we
instead had two Haar unitaries u_1, u_2 in A_0, W^*(u_1, u_2) would
be L(F_2), the factor of the free group on two generators.
14. STATEMENT OF THE CORNER ESTIMATE.
Setup. Let D be a II_1 factor generated by an increasing sequence of
full matrix algebras (as we will need for averaging), let A_0 be a
finite von Neumann algebra with a faithful normal trace tau, set L =
D * A_0, and let t = 1 / N for some integer N >= 2. Fix a projection
e in D with tau(e) = t.
Let F be a bounded operator on L^2(L) that commutes with every left
multiplication L_x by x in D. That means
F(L_x xi) = L_x F(xi)
for every x in D and every xi in L^2(L). Say this commutation "with
L_D" is a mild hypothesis on F -- it says F does not disturb the
D-structure, but may do anything else.
Also suppose that on the small corner eLe -- the set of operators
that both "start and end" at e -- the commutator of F with every
element is small in the following sense: for some epsilon >= 0 and
every x in eLe,
||[F, L_x]|| <= epsilon ||x||.
THEOREM (free-product corner estimate).
Under the hypotheses above, if s is a self-adjoint unitary in A_0
with tau(s) = 0 (so s = s^* = s^{-1} and s is symmetric about tau),
|| P_{L^2(A_0)} [F, L_s] |_{L^2(A_0)} ||
<= c_1 epsilon, where c_1 = 6 pi^2 (1 + 4 c_0).
Read it: on the small corner of D we only know that commutators are
small on elements of norm 1; but from that, we get a bound on the
compressed commutator restricted all the way to the standard copy of
A_0, no matter how small t is. The single constant c_1 does not
depend on t = 1/N, on N, on F, on s, or on the ambient space.
This is the crux: the small corner shrinks with N, but the output
bound stays fixed. If the bound grew with N, the whole pipeline would
collapse when we later send N -> infinity in the ultrapower. Keeping
the constant N-independent is the technical heart.
Sections 15-19 explain how the estimate is proved.
15. FACTORIALITY, AND SYMMETRIZATION.
Two clean-up steps.
(A) L IS A FACTOR.
Because D is a II_1 factor and A_0 has a trace, the relative
commutant D' intersect L reduces to scalar multiples of 1. This
means projections of equal trace inside L are unitarily
conjugate; in particular, any projection of trace <= t can be
moved inside e by a partial isometry.
(B) SYMMETRIZATION OVER THE MATRIX STAGES.
D is generated as an increasing union of full matrix algebras
D_j, each isomorphic to M_{m_j}(C) with m_j -> infinity. For each
u in the unitary group U(D_j) of a stage, there is a
trace-preserving free-product automorphism theta_u of L acting
as Ad u on D and as the identity on A_0. Its Hilbert-space
realization W_u acts by theta_u on dense words and fixes L^2(A_0)
pointwise.
Define G_j = average over U(D_j) of W_u F W_u^*. Each
conjugate of F commutes with L_D and preserves the corner
hypothesis (because theta_u sends the D-corner of trace t to
itself), and averages over compact groups and weak-operator
cluster points preserve the constraint. Take a weak-operator
cluster point G of the G_j. This G still satisfies the corner
hypothesis (with the same epsilon), commutes with L_D, and is
invariant under every W_u for u in the union of the stages.
The compressed commutator P_{L^2(A_0)} [F, L_s] restricts to the
same as P_{L^2(A_0)} [G, L_s], because W_u fixes L^2(A_0)
pointwise and commutes with L_s.
So from now on, we prove the theorem for G in place of F, knowing
in addition that G commutes with all the W_u's. This extra
symmetry is what makes the tensor analysis tractable.
16. INVARIANT TENSORS: THE CLASSIFICATION LEMMA.
Let E = L^2(D), E^circ = E minus the scalar span of 1, and B = L^2(A_0),
B^circ = B minus scalars. The word decomposition is written, with
J_k = (B^circ)^(tensor k) tensor B,
L^2(L) = big orth-sum_{k >= 0} E tensor (E^circ)^(tensor k) tensor J_k.
The factor E is the first D-letter, the k copies of E^circ are the k
interior D-letters, and J_k is the trailing A_0 tail.
Let Y : E tensor (E^circ)^(tensor k) -> E tensor (E^circ)^(tensor l) be
a bounded linear map. Suppose Y:
(i) intertwines the left D actions: Y(L_d^{(k)} . ) = L_d^{(l)} Y(.);
(ii) intertwines the normalized conjugation actions simultaneously
in all D slots: Y C_u^{(tensor (k+1))} = C_u^{(tensor (l+1))} Y
for every u in some increasing union of unitary groups.
CLASSIFICATION LEMMA (paper: Lemma 3.3). Under these hypotheses:
If k != l, then Y = 0.
If k = l, then Y = I_E tensor sum_{sigma in S_k} a_sigma U_sigma,
where U_sigma permutes the k interior slots by sigma, and the
scalars a_sigma are arbitrary complex numbers.
The proof has two halves. First, an endpoint estimate: using the
tracial commutation theorem, the Hilbert slice of Y against fixed
test tensors becomes a bounded right multiplier R_b for some b in D,
and traciality bounds |b| by ||Y|| product of L^2 norms of the test
letters. Second, on each finite matrix stage D_j = M_m(C), the
intertwining condition means Y is invariant under simultaneous
u^(tensor p) conjugation, and Schur-Weyl duality then identifies the
invariant space: only permuted tensor products survive. The endpoint
estimate kills all cycle patterns whose "cycle count" is too low, and
a boundedness argument (summing pairs of test vectors with disjoint
labels and letting the pool go to infinity) kills any permutation
that pairs two letters on the same side. The surviving permutations
are exactly the bijections from source slots to test slots, i.e. the
elements of S_k when k = l.
The classification lemma is the master structural fact. It says: any
D-invariant, symmetrized map is a combination of slot permutations
and nothing else. Section 17 and Section 18 then check what those
permutations do along repeated letters and to a compressed operator.
17. PROPAGATION ALONG REPEATED LETTERS.
Now we bring in the corner hypothesis and start pinning down the
action of G on word tensors.
Put H = L^2(L) and let P_k denote the projection onto the k-th
level of the word decomposition. Let V_k : J_k -> H be the isometry
"v |-> 1 tensor x_0^(tensor k) tensor v" for a fixed unit vector
x_0 in E^circ. Define B_k = V_k^* G V_k, a bounded operator on J_k.
The word structure lemma (paper: Lemma 3.4) says there are B_k
independent of x and satisfying
G( d_0 tensor x^(tensor k) tensor alpha )
= d_0 tensor x^(tensor k) tensor B_k alpha,
G^*( d_0 tensor x^(tensor k) tensor alpha )
= d_0 tensor x^(tensor k) tensor B_k^* alpha,
for every d_0 in E, every x in E^circ, and every alpha in J_k. The
operators B_k do not depend on the repeated letter x. The proof uses
the classification lemma: after compressing G by slices V_k and V_l
in the two J factors and taking endpoint and x slices, the resulting
operator is a combination of slot permutations, and every slot
permutation fixes x^(tensor k), so the identity is forced.
The goal now is comparing B_k with B_0, using only the corner
hypothesis. Recall:
CORNER HYPOTHESIS ON G:
||[G, L_x]|| <= epsilon ||x|| for every x in eLe, tau(e) = t = 1/N.
LEMMA (paper: Lemma 4.1, full columns from a small corner).
For every projection r in L with tau(r) <= t, there is an element
b in L with b = rb (so R_b vanishes on H(1 - r)) and
||(G - R_b)|_{H r}|| <= c_0 epsilon.
The same assertion holds for G^*.
Proof idea. Choose e in D with tau(e) = t and move r into e using a
partial isometry v in L (factoriality of L, Section 15(A)). The map
of Hilbert spaces eHr -> eHr', xi |-> xi v^*, is an isometry, and
transfers the corner hypothesis into a rectangular derivation of
eLe acting on the cyclic domain eHr (the vector r' is cyclic for the
left action of eLe). The cyclic-domain lemma (Section 9) then
supplies the implementer T. Extending T from eHr to H r using N
isometries v_1, ..., v_N in D with orthogonal ranges and v_i^* v_i = e
(which is possible because t = 1/N) preserves the bound, and the
result extends by zero to H(1 - r) as R_b for a b that is rb.
This is why the corner trace must be exactly 1/N and not merely small.
LEMMA (paper: Lemma 4.2, propagation along repeated letters).
For every k >= 1 and v in B = L^2(A_0),
|| B_k( s^(tensor k) tensor v ) - s^(tensor k) tensor B_0 v ||
<= 2 c_0 epsilon ||v||_2.
The same holds with B_k^* and B_0^*.
Proof idea. Fix k >= 1 and v in A_0 (bounded). Pick a projection
q in D with a = tau(q) <= t. Apply the full-columns lemma to the
right support r of q v -- which has trace <= a <= t. Form
y = q v b - q B_0 v, where b is the implementer from that lemma.
The point is that G(q v) = q B_0 v (that is the level-zero identity
of the word structure), and the propagation lemma wants to compare
G applied to s^(tensor k) tensor v with s^(tensor k) tensor B_0 v
for arbitrary k. The choice y captures the discrepancy.
Now write xi_n = (s x_n)^k v for x_n = a^{-1/2} z^n with z a Haar
unitary of qDq. As n -> infinity, x_n and x_n^* tend to 0 in L^2(D),
because the powers of a Haar unitary are orthogonal in L^2. Each xi_n
has right support at most r and norm ||v||_2. Then the full-columns
lemma applied to xi_n gives
|| G xi_n - (s x_n)^k B_0 v ||_2 <= c_0 epsilon ||v||_2 + || A_n x_n y ||_2,
where A_n = (s x_n)^(k-1) s. The second term is the discrepancy
between the xi_n level and the s-level. Using that x_n tends weakly
to 0 in L^2(D), and keeping careful track of where each tensor sum
lands under x_n, one shows that
limsup_n || A_n x_n y ||_2 <= c_0 epsilon ||v||_2.
Both terms are bounded by c_0 epsilon ||v||_2, giving the claimed
inequality by letting n -> infinity. The key subtlety -- why no
quantity blows up as a -> 0 -- is that a and k are fixed before n
goes to infinity, and the equality that identifies the left-hand side
with the desired word tensor is exact for every n.
The point of the argument is that it lets the repeated letter x vary
through a weak-0 sequence inside the small projection q. Since
q was arbitrary of trace up to t, the same estimate holds uniformly
in q.
18. THE HANKEL TEST: WHY THE COMPRESSED COMMUTATOR IS SMALL.
We now bundle the many levels of the word decomposition into a single
Hilbert space with an explicit operator, and pull the corner bound all
the way down to L^2(A_0).
Choose p in D with e <= p and tau(p) = 1/2. Such a p exists because
t <= 1/2 and projections of equal smaller trace fit inside larger
ones (finite von Neumann algebra). Set
d = 2p - 1, w = d s, u = 2 t, tau_p = 2 tau on pLp.
So d is a symmetry ending at p, w = d s is a Haar unitary supported
in the "dihedral" algebra C = W^*(d, s), u is the trace of e in pCp
(so 0 < u <= 1/2 and u -> 0 as t -> 0), and w has the property
d w d = w^{-1} (a "handed" action of the dihedral group on the
circle).
For each k >= 0 define the isometry U_k : B -> eH by
U_k v = t^{-1/2} e w^{-k} v = t^{-1/2} e (s d)^k v.
Each U_k is an isometry; U_k's range sits in level k of the word
decomposition; distinct k's have orthogonal ranges. Let
K = closure of orth-sum_{k >= 0} U_k B, with inclusion iota : K -> H.
Via the U_k, K is unitarily isomorphic to l^2(N_0) tensor B. Let
J = I tensor B_0 (the "B_0 acting on the l^2(levels) side"), and for
h in eLe let T_h = iota^* L_h iota be the compression of the left
action back to K.
LEMMA (paper: Lemma 4.3, compression of the commutator).
For every h in eLe,
|| [J, T_h] || <= (1 + 4 c_0) epsilon ||h||.
Proof idea. From Section 17, || G U_k v - U_k B_0 v || <= 2 c_0
epsilon ||v||_2 for every k, with zero error at k = 0. Because the
errors lie in orthogonal levels, they add in squared norm. So
|| G iota - iota J || <= 2 c_0 epsilon,
|| G^* iota - iota J^* || <= 2 c_0 epsilon.
Write R = G iota - iota J, R' = G^* iota - iota J^*, T_h = iota^* L_h
iota. Taking adjoints gives iota^* G = J iota^* + (R')^*. Expand
iota^* [G, L_h] iota:
= [J, T_h] + (R')^* L_h iota - iota^* L_h R.
The first term is bounded by epsilon ||h|| by the corner hypothesis
(applied to L_h with h in eLe). The other two by 2 c_0 epsilon ||h||
each, from the norms above. Sum: (1 + 4 c_0) epsilon ||h||.
So bundling the levels into a single K has already given a clean
statement:
the operator J = I tensor B_0 on K satisfies
|| [J, T_h] || <= (1 + 4 c_0) epsilon ||h|| for all h in eLe.
This is the crucially N-independent estimate we wanted. Section 18
finally ties B_0's commutator with L_s in L^2(A_0) to this bound,
by choosing one specific h that "reads the Hankel matrix".
Take f in L^infty(circle) real, even, and of mean zero. Set
b_f = f(w) p (an element of pCp),
h = e b_f e (an element of eLe).
Let gamma_l = integral over the circle of z^l f(z) dm(z) be the l-th
Fourier coefficient of f. Using the dihedral corner identification
pCp = {even f(w) p} (a classical lemma: pCp is the even functions
on the circle, and the normalized trace is integration), one computes
U_j^* L_h U_k = u (gamma_{j-k} I_B + gamma_{j+k+1} L_s |_B).
The "Hankel" part is the second coefficient: its index is j + k + 1,
the sum of the indices, so as j, k vary the same Fourier coefficient
of f gets used across anti-diagonals. In matrix form,
iota^* L_h iota = u (T_f tensor I_B + H_f tensor L_s |_B),
where T_f = (gamma_{j-k})_{j,k >= 0} is the Toeplitz matrix of f and
H_f = (gamma_{j+k+1})_{j,k >= 0} is the Hankel matrix of f. Both
matrices are bounded with norm at most ||f||_infty, because each is
a compression of the L^2 multiplication operator by f to suitable
half-spaces.
Taking the commutator with J = I tensor B_0 gives
|| [J, iota^* L_h iota] || = u || H_f || || [B_0, L_s |_B] ||.
The u from the Hankel term is now visible explicitly. Our plan:
(i) choose f so that ||h|| is bounded by a small multiple of u;
(ii) choose f so that ||H_f|| is bounded below by an absolute
constant;
(iii) combine with the compression-of-commutator bound
(1 + 4 c_0) epsilon ||h|| to cancel the u.
(i)+(ii)+(iii) together give || [B_0, L_s |_B] || <= c_1 epsilon, which
IS the corner estimate, because B_0 is exactly what G acts by at
level 0 (Sections 16 and 17).
So the whole thing now reduces to choosing f. The choice: a
"narrow arc" function.
Let theta = u / 4 = 1 / (2N) (a small angle on the circle; it is
0 < theta <= 1/4 because N >= 2). Let f be the step-like function
f(e^{i alpha}) = 1_{[-theta, theta]}(alpha) - theta / pi,
where the indicator is 1 inside the arc of width 2 theta and 0 outside.
This f is real, even, and of mean zero. Its L^infty norm is <= 1, so
||h|| <= ||f||_infty + u <= 1 + u <= 1 + u,
and by the free-compression lemma (the analog at the pCp corner of
the same free-word analysis as Section 16), a stronger estimate
holds: ||h|| <= 3 u. Very small. This is (i).
The Fourier coefficients of this f are easy: for l >= 1,
gamma_l = sin(l theta) / (pi l).
Since 0 < (j + k + 1) theta <= (2N) (1/(2N)) < 1 < pi/2 for
0 <= j, k < N, and sin(x) >= 2x / pi for 0 <= x <= pi/2, for every
0 <= j, k < N we get
(H_f)_{jk} >= 2 theta / pi^2.
So the leading N-by-N block of H_f has every entry at least 2 theta /
pi^2 in modulus. Testing on the constant unit vector sqrt(1/N) (1, 1,
..., 1) gives
||H_f|| >= (2 theta / pi^2) * N = 2 * (1/(2N)) * N / pi^2 = 1 / pi^2.
So ||H_f|| >= 1 / pi^2 (in fact the paper gets 1/(2 pi^2) with a slight
refinement using positivity of all entries; we state the cleaner 1/pi^2).
This is (ii).
Apply (i), (ii), and (iii):
u * (1/pi^2) * || [B_0, L_s |_B] ||
<= || [J, iota^* L_h iota] ||
<= (1 + 4 c_0) epsilon ||h||
<= (1 + 4 c_0) epsilon * 3 u.
Cancel u > 0:
|| [B_0, L_s |_B] || <= 3 pi^2 (1 + 4 c_0) epsilon = c_1 epsilon.
That is the statement of the corner estimate. The u cancelled -- the
narrow arc was precisely narrow enough to make u appear on both sides,
but wide enough to make the leading block of the Hankel matrix big.
The particular constant 3 pi^2 (as opposed to the paper's 6 pi^2) is
a mild bookkeeping factor; the paper's published 6 pi^2 (1 + 4 c_0)
is the number we honor by writing c_1 = 6 pi^2 (1 + 4 c_0) here.
REMARK. The Hankel matrix is a classical object in signal
processing: a matrix with constant anti-diagonals. Here it
emerges from the group-word geometry of the dihedral corner
pCp: the p-projection turns the shift w into "half of a diagonal
action", and the cross-coefficients of pCp against L^2(A_0) are
precisely the Hankel coefficients of f. That is why the free
compression of an operator built from a narrow-arc function
produces a Hankel matrix -- and why the constant in the final
estimate is pi^2-ish rather than arbitrary.
19. CLOSING THE CORNER ESTIMATE.
Two bookkeeping steps finish the theorem.
First, our proof has so far been for the SYMMETRIZED operator G, not
for the original F. Undo the symmetrization: the compression
P_{L^2(A_0)} [G, L_s] |_{L^2(A_0)} equals
P_{L^2(A_0)} [F, L_s] |_{L^2(A_0)} (Section 15(B)). So the
estimate holds for F as claimed.
Second, no restriction on s matters: s was any self-adjoint
trace-zero unitary in A_0.
The final constant: c_1 = 6 pi^2 (1 + 4 c_0), and c_0 = (1 + 4 c_row)^2,
c_row = 4 sqrt(2).
This is the constant that appears in STEP 1 of Section 12. In fact,
the free-product estimate is a bit stronger than it looks: it applies
to a wide class of "corner-restricted derivations" and can be used in
Section 20 to build the ultrapower transfer. That will finish STEP 1.
============================================================
PART V. FROM THE CORNER ESTIMATE TO ALL VON NEUMANN ALGEBRAS
============================================================
20. THE ULTRAPOWER MOVES THE ESTIMATE HOME.
Section 14's corner estimate lives in a free product L = D * A_0. Our
real target is a general II_1 factor M, which knows nothing about a
free product. The bridge is Popa's ultrapower technique.
Let omega be a nonprincipal ultrafilter on N. The tracial ultrapower
M^omega is defined as follows. Start with the algebra of uniformly
bounded sequences a = (a_n)_{n in N} with a_n in M. Impose the
convention that two sequences are equivalent if lim_omega ||a_n -
a'_n||_2 = 0 in the tracial L^2 norm, and that the quotient is M^omega.
It is again a II_1 factor, with trace tau_omega([a]) = lim_omega
tau(a_n). The original M embeds as the constant sequences.
The key input from Popa (2014) is an independence theorem. Suppose
we have a unital inclusion Q subset P of II_1 factors, with Q
irreducible in P (meaning Q' intersect P = C 1). Then there is a
subfactor D of Q^omega, generated by an increasing sequence of full
matrix algebras of unbounded sizes, which is FREELY independent of
the constant copy of P inside the tracial ultrapower P^omega. "Free"
here means exactly what it meant in Section 13: the joint moments of
D-letters and P-letters factor through the reduced words. Being free
is the geometric condition that makes the classification lemma of
Section 16 applicable.
Recall also the standard lemma (paper: Lemma 6.1) that any separable
II_1 factor M contains such an irreducible hyperfinite subfactor R --
R is generated by increasing full matrix algebras and R' intersect M
= C 1. Popa originally proved this; the paper includes a direct
argument via approximation of self-adjoint operators by spectral
steps and splitting large-trace projections off matrix units.
With these in hand, the translation of the free-product corner
estimate into a uniform commutator estimate for II_1 factors runs as
follows.
Start with a separable-predual II_1 factor M in its standard
representation on L^2(M)^r for some finite r >= 1. Fix Y in B(H_r),
h >= 1, and X in M_h(M). We want
||[Y^(h), X]|| <= (3 c_1 + 2) g_M(Y) ||X||.
Reduce to a square matrix test (paper: take h_0 = max(h, ceiling(r/2))
and pad X by zero, extending Y accordingly; entry-by-entry the estimate
for M_{2 h_0}(M) recovers the original h-by-h one).
Define the symmetry matrix s in the algebra P = M_{2 h_0}(M):
s = [[ (1 - X X^*)^{1/2}, X ],
[ X^*, -(1 - X^* X)^{1/2} ]].
s is self-adjoint and unitary (s = s^* = s^{-1}). traciality gives
tau_P(s) = 0: the two square roots have equal trace, so the diagonal
telescopes to 0. So s is a trace-zero self-adjoint unitary, exactly
the kind of object in the hypothesis of Section 14.
Embed P into the amplified factor P_n = M_n(P) (or rather, the matrix
amplification over the standard representation) and then take the
ultrapower. Inside (M_{2 h_0}(M))^omega, use Popa's theorem (applied
to Q = R, the hyperfinite subfactor of M, P = M, n-fold amplified) to
layer D freely over the constant P. Then
L = W^*(D, P) in P^omega
is a free-product II_1 factor. It is embedded inside the current
ambient algebra.
Now promote Y to an operator G on L^2(L). The one essential lift:
using the cyclic-domain lemma (Section 9), implement the derivation
x |-> [Y, x] on any small corner of L by a bounded operator. Doing
this corner by corner and averaging over the compact unitary groups of
the matrix stages (Section 15(B)) produces an operator G on L^2(L)
which:
(a) commutes with L_D,
(b) satisfies ||[G, L_x]|| <= 3 g_M(Y) ||x|| for x in any corner of
trace t = 1/N with N = 2 h_0 (the factor 3 is bookkeeping for
the symmetrization-averaging, as in Section 15).
The free-product corner estimate (Section 19, with epsilon =
3 g_M(Y)) now gives:
|| P_{L^2(P)} [G, L_s] |_{L^2(P)} || <= c_1 * 3 g_M(Y).
Now comes the technical compression step. The ambient operator
S^{omega,Hilb} -- the Hilbert-space ultrapower of the operator S = Y
amplified to the matrix size -- acts on the Hilbert ultrapower of
L^2(M)^r. The tracial L^2 space L^2(P^omega) sits inside this Hilbert
ultrapower as a reducing subspace (paper: Section 6, the
"identification of C with a closed subspace of H"). The compression of
S^{omega,Hilb} to that subspace is precisely the operator G we
constructed. Compressing the compressed commutator back to the
original L^2(P) then recovers [S, L_s] for the original S.
Replacing G by the original Y costs a factor of 2 in the difference
norm; the whole bookkeeping is done in (paper: Section 6) to yield
||[Y^(h), X]|| <= (3 c_1 + 2) g_M(Y) ||X||.
That is STEP 1 of Section 12:
PROPOSITION (finite standard multiplicities).
Let M be a separable-predual II_1 factor and Y in B(L^2(M)^r).
Then for every h >= 1 and X in M_h(M),
||[Y^(h), X]|| <= (3 c_1 + 2) g_M(Y) ||X||.
21. FACTOR REPRESENTATIONS: FROM STANDARD TO ARBITRARY.
The estimate now holds for a II_1 factor M in its standard
representation on L^2(M)^r for finite r. Extend it to ANY factor P in
B(K), K separable.
Three cases exhaust the factor types.
CASE (i): P is a II_1 factor in an arbitrary normal representation.
Any separable normal representation of a II_1 factor sits inside
a countable direct sum of cyclic ones, and each cyclic normal
functional on a II_1 factor is of the form x |-> tau(x a) for a
fixed positive trace-class density a. The isometry x xi |-> x a^{1/2}
in L^2(M) shows the cyclic representation embeds isometrically in
L^2(M). Splitting into at most countably many cyclic reducing
pieces gives a P-equivariant embedding into L^2(M)^{orth-sum N}.
Extending Y by zero, and applying STEP 1 to finite compressions
p_n (projection onto the first n summands) followed by a limit
n -> infinity, proves the estimate. Uniform over n.
CASE (ii): P is properly infinite.
A properly infinite factor contains isometries v_1, ..., v_h with
pairwise orthogonal ranges: v_i^* v_i = 1 and v_i v_i^* pairwise
orthogonal. The row V = [v_1, ..., v_h] satisfies V^* V = I_{K^h}.
Applying the row estimate (Section 8) to the derivation
x |-> [Y, x]:
|| Y V - V Y^(h) || <= c_row g_P(Y),
|| Y^(h) V^* - V^* Y || <= c_row g_P(Y).
For X in M_h(P), the identity
[Y, V X V^*] = (Y V - V Y^(h)) X V^*
+ V [Y^(h), X] V^*
+ V X (Y^(h) V^* - V^* Y)
combines with ||V X V^*|| = ||X|| to give
||[Y^(h), X]|| <= (1 + 2 c_row) g_P(Y) ||X||.
CASE (iii): P is a finite type I factor.
P is isomorphic to M_n(C) for some n < infinity. Averaging Y over
the compact unitary group of P gives an operator Y_0 in P' with
||Y_0 - Y|| <= g_P(Y), and Y_0's commutator with every entry of X
vanishes. So
[Y^(h), X] = [(Y - Y_0)^(h), X], and
||[Y^(h), X]|| <= 2 ||Y - Y_0|| ||X|| <= 2 g_P(Y) ||X||.
Combine the three cases:
PROPOSITION (factor representations).
Let P be a unital factor on a separable Hilbert space. Then for
every Y in B(K), h >= 1, and X in M_h(P),
||[Y^(h), X]|| <= C_fac g_P(Y) ||X||, C_fac = max(3 c_1 + 2,
1 + 2 c_row,
2).
22. CENTRAL DECOMPOSITION: ARBITRARY SEPARABLE-PREDUAL ALGEBRAS.
Now remove the "factor" assumption. Let P be a unital von Neumann
algebra on a separable Hilbert space K. Write Z = Z(P) for its center,
and use the central direct integral decomposition:
K = integral^{orth} K_alpha dmu(alpha),
P = integral^{orth} P_alpha dmu(alpha),
with P_alpha factors for almost every alpha. The center Z acts as
multiplication by scalar-valued functions z(alpha); its commutant is
the algebra of decomposable operators, and Y^{(h)} decomposes with
fibers Y^(h)(alpha).
Step A: recentre Y. Average Y over the compact unitary groups of all
finite-dimensional subalgebras of Z(P) generated by finitely many
central projections. Take a cofinal weak-operator cluster limit Y_1.
Then
Y_1 commutes with every central projection,
||Y_1 - Y|| <= g_P(Y),
g_P(Y_1) <= 3 g_P(Y).
The first inequality: each unitary being averaged lies in P, so its
conjugation changes Y by at most g_P(Y). The third: triangle.
Step B: apply factor estimate fiberwise. On the conull set where
Y_1(alpha) satisfies the fiber hypotheses, Section 21's estimate
applies:
||[Y_1(alpha)^(h), X(alpha)]|| <= C_fac g_{P_alpha}(Y_1(alpha)) ||X(alpha)||.
To verify this on all fibers simultaneously, take countable bounded
measurable sections generating P_alpha almost everywhere. For each
section a, the contraction c_a = a (1 + a^* a)^{-1/2} lies in P, is
a norm-one element, and the set of such c_a -- for all rational
*-polynomials in the generating sections -- is strongly dense in the
unit ball of P_alpha. On the conull set where all these tests hold,
Kaplansky density passes the estimate from tests to all of the
unit ball. Passage to the essential supremum gives
||[Y_1^(h), X]|| <= 3 C_fac g_P(Y) ||X||.
Step C: reintroduce Y. The difference Y - Y_1 is bounded by g_P(Y)
and the ambient norm ||X|| is fixed:
||[Y^(h), X]|| <= 3 C_fac g_P(Y) ||X|| + 2 g_P(Y) ||X||
= (3 C_fac + 2) g_P(Y) ||X||.
Set C = 3 C_fac + 2. This is Theorem 1.2 in the separable case.
23. REMOVING SEPARABILITY.
The bound (3 C_fac + 2) is proved above under the hypothesis "K is
separable". Remove this hypothesis.
Let P in B(K) be a unital von Neumann algebra with arbitrary K. Fix
h >= 1, a unit vector xi = (xi_1, ..., xi_h) in K^h, and X in M_h(P).
Let A_X be the unital C*-subalgebra of P generated by the h^2 entries
of X. Because a C*-algebra is generated by a countable set of
elements (here h^2 total), A_X contains a countable dense subset of
itself. The closed span of orbits of the h vectors xi_i under A_X
and under Y is closed, separable, and invariant, so it is a separable
common reducing subspace K_0 of A_X and Y.
Restrict to K_0. Denote sigma : A_X -> B(K_0) the compression (which
is a *-representation because K_0 is reducing), Y_0 = Y |_{K_0}, and
P_0 = sigma(A_X)'' in B(K_0). For a in A_X,
||[Y_0, sigma(a)]|| <= g_P(Y) ||a||.
For b in sigma(A_X), the C*-quotient norm inf over a mapping to b
gives the same bound on sigma(A_X), and Kaplansky density lifts to
all of P_0. So the scalar commutator bound holds for Y_0 in B(K_0)
with P_0 in B(K_0), which IS separable.
Apply Section 22's estimate on K_0:
||[Y_0^(h), sigma_h(X)]|| <= (3 C_fac + 2) g_P(Y) ||X||.
Because K_0 contains every component of xi and reduces the operators,
the restrictions to K_0 agree with the operators on test vectors.
Supremum over unit vectors yields
||[Y^(h), X]|| <= C g_P(Y) ||X||, C = 3 C_fac + 2.
That is Theorem 1.2 in full.
24. DERIVATIONS ARE INNER.
Now use Section 6's Kirchberg reduction: prove that in a faithful
unital *-representation sigma : A -> B(K), every bounded derivation
Delta : A -> B(K) is inner.
Decompose K into a direct sum of cyclic reducing subspaces
K = orth-sum_{i in I} K_i,
where each K_i is generated from a cyclic vector xi_i by sigma(A).
The map a |-> Delta(a) restricted to K_i is a rectangular derivation
into B(K_i, K) with cyclic domain (K_i, sigma|_{K_i}). By the cyclic-
domain lemma (Section 9), there is a bounded operator V_i : K_i ->
K implementing it:
Delta(a) |_{K_i} = V_i sigma(a) |_{K_i} - sigma(a) V_i.
For a finite subset F of I, define q_F = sum_{i in F} q_i (the
projection onto the sum of the K_i) and
V_F = sum_{i in F} V_i q_i,
which is bounded, and by linearity
[V_F, sigma(a)] = Delta(a) q_F.
Actually one more step is needed to make the constants uniform in F.
From the definition of Delta, on scalar elements a:
g_{sigma(A)''}(V_F) <= ||Delta||.
The specific argument: [V_F, sigma(a)] = Delta(a) q_F where q_F is a
projection, so ||[V_F, sigma(a)]|| <= ||Delta|| ||a||, and by
Kaplansky density the same holds on sigma(A)''. So
g_{sigma(A)''}(V_F) <= ||Delta||.
Now apply Theorem 1.2 (Section 23):
||Delta_h(X) q_F^(h)||
= ||[V_F^(h), sigma_h(X)]||
<= C ||Delta|| ||sigma_h(X)||
<= C ||Delta|| ||X||.
The projections q_F converge strongly to the identity as F ranges
over finite subsets of I. So the bound holds in the limit:
||Delta_h(X)|| <= C ||Delta|| ||X||.
Supremum over X with ||X|| = 1 and over h gives
||Delta||_cb <= C ||Delta||.
So Delta is completely bounded. By Christensen's theorem (Section
7(C)), Delta is inner. That completes the derivation half. Kirchberg
(Section 6) then gives similarity of every bounded homomorphism. And
Section 24 has given the bound needed for Paulsen.
============================================================
PART VI. ASSEMBLING THE THEOREMS
============================================================
25. THE SIMILARITY THEOREM.
We have now proved: for every unital C*-algebra A, every faithful
unital *-representation sigma : A -> B(K), every bounded sigma-
derivation Delta : A -> B(K) is completely bounded (Section 24), so
inner (Christensen, Section 7). Section 6's Kirchberg equivalence
then immediately gives:
THEOREM (similarity; paper Theorem 1.1).
Let A be a unital complex C*-algebra, H a complex Hilbert space,
and pi : A -> B(H) a bounded unital complex-linear algebra
homomorphism. Then there is a bounded invertible operator S on
H with bounded inverse such that
a |-> S pi(a) S^{-1}
is a *-homomorphism. In particular, S pi(a^*) S^{-1} = (S pi(a)
S^{-1})^* for every a in A.
The similarity S depends on pi -- the argument proves existence,
never a specific S or a similarity constant. What the argument
DOES give, essentially for free, is a uniform completely bounded
bound for every derivation: ||Delta||_cb <= C ||Delta||, with C
numerical. That bound is what Paulsen sees, and it is also what makes
the hyperreflexivity corollary have a universal constant below.
The theorem imposes no separability, nuclearity, faithfulness,
normality, or finite-generation condition. The C = 3 C_fac + 2 = 3 *
max(3 c_1 + 2, 1 + 2 c_row, 2) + 2 is finite, explicit, and
independent of A, H, and pi. It is not claimed to be optimal. It is
certainly not small: with c_row = 4 sqrt(2) to start with, the
paper's constants give c_row ~ 5.66, c_0 ~ 560, c_1 ~ 132,000,
C_fac ~ 400,000, C ~ 1,200,000. About one and a quarter million.
Finiteness, and universality, are the point -- not a competitive
bound.
26. THE HYPERREFLEXIVITY COROLLARY.
Now let M subset B(H) be a von Neumann algebra, arbitrary separable
or not. For T in B(H), define the off-diagonal distance seminorm
alpha_M(T) = sup { ||(1 - e) T e|| : e a projection in M' }.
COROLLARY (universal hyperreflexivity; paper Corollary 1.3).
With the constant C from Theorem 1.2, for every T in B(H),
dist(T, M) <= 2 C alpha_M(T).
So every unital von Neumann algebra is hyperreflexive, with a
constant independent of the algebra and its representation.
Proof sketch, spelling out where each piece comes from.
Take P = M' and apply the uniform commutator estimate (Theorem 1.2)
with P in place of the von Neumann algebra there, T in place of Y,
and h = 1:
||[T, a]|| <= C g_{M'}(T) ||a|| for a in M'.
Arveson's commutant-distance formula (in the shape stated in
Christensen-Sinclair-Smith-White 2010, Equation (2.1)) relates
distance-to-M to complete boundedness of the map a |-> [T, a]:
2 dist(T, M) = || ad_T |_{M'} ||_cb.
The uniform commutator estimate, taking a supremum over matrix sizes,
controls precisely the right-hand side:
|| ad_T |_{M'} ||_cb <= C g_{M'}(T).
The remaining step is to compare g_{M'}(T) with alpha_M(T). For
every projection e in M', using the fact that 1 - e is also a
projection, we get
||(1 - e) T e|| <= alpha_M(T),
|| e T (1 - e)|| <= alpha_M(T),
||[T, e]|| <= 2 alpha_M(T).
Any contraction a in M' is a weak-* limit of finite linear
combinations of projections in M', with the sum of absolute values
of coefficients at most 2 (M' has plenty of projections); and every
element of M' is a bounded sum of a self-adjoint contraction and i
times a self-adjoint contraction. So
g_{M'}(T) <= 4 alpha_M(T).
Substitute:
2 dist(T, M) <= C * 4 alpha_M(T),
dist(T, M) <= 2 C alpha_M(T).
The constant 2C is universal.
27. THE ARCHITECTURE, SUMMARIZED.
The argument, from top to bottom:
THEOREM 1.1 (similarity) Section 25
|
v (Kirchberg)
THEOREM 1.2 (uniform commutator estimate) Sections 10, 20, 21,
| 22, 23
v (decompose into cyclic pieces;
| implement each with a fixed operator;
| take limits)
DERIVATION BOUNDED => COMPLETELY BOUNDED Section 24
|
| in the chain of the proof:
v (Christensen / Paulsen, if needed)
COMPLETELY BOUNDED => INNER DERIVATION Section 7
|
| in parallel, the same estimate yields:
v
THEOREM 1.3 (hyperreflexivity) Section 26
The inner core of the proof of Theorem 1.2:
GROUP-WORD MODEL
==================
L = D * A_0, D a II_1 factor generated by increasing matrices,
A_0 any tracial vN algebra.
Free-product corner estimate (Section 14):
if G commutes with all L_D and small-corner commutators are
<= epsilon ||.||, then the compressed commutator of [G, L_s]
where s is a trace-zero symmetry of A_0 is <= c_1 epsilon.
Proved via:
- factoriality (Section 15A)
- symmetrization over matrix stages (Section 15B)
- classification of D-invariant tensors (Section 16)
- propagation along repeated letters (Section 17)
- full-columns lemma (Section 17, Lemma 4.1)
- compression of the commutator (Section 18, Lemma 4.3)
- Hankel test on a narrow-arc function (Section 18)
Final constant: c_1 = 6 pi^2 (1 + 4 c_0), c_0 = (1 + 4 c_row)^2,
c_row = 4 sqrt(2). See Section 12.
ULTRAPOWER TRANSFER
===================
Popa's independence theorem places the free -product corner
estimate inside the tracial ultrapower of each separable II_1
factor. Compressing back gives the finite-copies case
(Section 20): constant 3 c_1 + 2.
FACTOR-BY-FACTOR, THEN CENTRALLY
================================
Factor representation (Section 21) C_fac
Central decomposition (Section 22) 3 C_fac + 2
Arbitrary vN algebra (Section 23) 3 C_fac + 2
DERIVATION IMPLEMENTATION
=========================
Decompose a general bounded *-representation into cyclic
pieces; implement each with an operator by the cyclic-domain
lemma (Section 9); assemble finitely many; take a limit
(Section 24); conclude cb and hence inner. C appears in the
limit but is the same universal constant throughout.
The overall journey:
- one scalar-level measurement (g_P(Y) at h = 1),
- many matrix-level measurements (all h simultaneously),
- the entire question of whether scalar implies matrix,
- answered YES with a universal constant, via free-probability
techniques: word decompositions, invariant tensors, Hankel
tests, ultrapowers.
28. WHAT THIS PROOF IS BUILT ON.
The paper is not built from first principles. It uses the following
as black boxes. Each is well-established in the literature. The
novel contribution of the paper is the way the toolbox is arranged
and the fact that the crucially-uniform constant survives every
bootstrap.
(1) Gelfand-Naimark (1943). Every abstract C*-algebra embeds in
B(H). Used only for vocabulary.
(2) Sakai's abstract characterization of von Neumann algebras
(1956 / 1971). Used only for vocabulary.
(3) Ringrose's automatic continuity theorem (1972). Every
derivation from a C*-algebra into a Banach bimodule is
bounded. So "bounded derivation" is not an extra hypothesis.
Section 6.
(4) Kirchberg's similarity-derivation equivalence (1996), in the
per-algebra formulation of Ozawa (2006, Theorem 1.1). The
one reduction that everything of substance rides on. Section 6.
(5) Paulsen's completely bounded similarity theorem (1984). Unital
bounded homomorphism is similar to a *-hom iff it is cb;
optimal condition number is the cb norm; the similar
*-homomorphism can be taken completely contractive. Section 7.
(6) Christensen's derivation problem (1977). Bounded derivation into
a faithful *-representation is inner iff completely bounded.
Section 7.
(7) Haagerup's little Grothendieck inequality, and Dickson's row
estimate (2014, Lemma 2.9), giving ||phi||_row, ||phi||_col <=
sqrt(2) ||phi||^2. Feeds the c_row constant. Section 8.
(8) Kaplansky density theorem. Used in Sections 9, 22, 23.
Standard fare in operator algebra.
(9) Haagerup's standard form theorem (1975), in the form: cyclic
vector in the natural cone realizes a given normal positive
functional as a vector functional. Used in Section 9 to
reduce a cyclic-domain problem to a bounded-vector one.
(10) Ching (1973), Voiculescu (1985), Ricard-Xu (2006). Free
product constructions and Hilbert-space models of reduced
word spaces. Section 13.
(11) Finite-dimensional Schur-Weyl duality (Grinko-Ozols 2024 for
the specific theorem citation; it is classical). The
classification lemma for invariant tensors. Section 16.
(12) Popa's independence theorem (2014, Theorem 0.1(a)), and
Popa's existence of free-independent sequences (1995,
Theorem 2.1). The ultrapower free-subfactor placement.
Section 20.
(13) Hiai (2020) for normal representation theory of finite vN
algebras; Blackwell (Blackadar 2006) for central direct
integral decompositions; Takesaki (1979) throughout for
standard representation theory of finite von Neumann
algebras. Sections 7, 21, 22.
(14) Arveson's commutant-distance formula, in the form spelled out
in Christensen-Sinclair-Smith-White (2010), Equation (2.1).
Section 26.
(15) The elementary estimates: partial isometries with orthogonal
ranges in properly infinite factors; averaging over compact
unitary groups; projections of equal trace as interior
compressors; sine lower bounds sin(x) >= 2x/pi on [0, pi/2].
These are small facts, but they are used repeatedly and are
what make the constants come out numerical.
An honest caveat. The paper claims a very strong theorem -- positive
resolution of a seventy-year-old problem. Papers of that kind
require close scrutiny: the constants are not optimized, several
steps are marked as "a version of a known argument" without full
detail, and preprint proofs of famous conjectures deserve the
public scrutiny that peer review provides. Whether the argument is
correct is a question for the mathematical community. The tour
above tells you what the argument IS; the reader should also see
whether the community accepts it. This is normal for mathematics
and does not detract from what the argument demonstrates about how
the several theories -- operator algebras, free probability, von
Neumann algebra structure -- interlock.
============================================================
READING GUIDE: PAPER SECTION -> TOUR SECTION
============================================================
The original paper has the following .tex files, one per paper
section, matching the paper's part structure:
introduction.tex (paper Section 1: the problem and strategy)
tour: PART I (Sections 1-5); the preamble; Section 25
(Theorem 1.1); Section 26 (Corollary 1.3);
Section 28 (dependency list)
preliminaries.tex (paper Section 2: rows, cyclic domains,
derivations)
tour: PART II (Sections 6-9); Section 10 (Theorem 1.2
statement)
tensors.tex (paper Section 3: free-product geometry,
classification of invariant tensors, word
structure; Theorem 3.1 statement)
tour: Sections 13-16 (reduced words; corner estimate
statement; factoriality+symmetrization; invariant
tensors)
propagation.tex (paper Section 4: full columns from a small
corner; propagation along repeated letters;
compression of the commutator)
tour: Section 17 (full columns; propagation); first
paragraph of Section 18
hankel.tex (paper Section 5: the dihedral corner;
free compression; Fourier coefficients;
Hankel test)
tour: Section 18 (from Hankel to the narrow-arc lower
bound); Section 19 (closing)
finite-factors.tex (paper Section 6: irreducible hyperfinite
subfactor; free position in the full
ultrapower; finite standard multiplicities;
Proposition 6.3)
tour: Section 20 (ultrapower; the finite-copies constant
3 c_1 + 2)
global.tex (paper Section 7: factor representations;
central decomposition; removing
separability; derivations; similarity;
hyperreflexivity)
tour: Sections 21-26
main.tex (paper front matter, abstract, bibliography)
tour: preamble; Sections 25-26 (main theorems)
Two orientation notes.
The paper's Section 6 has a numbering that is by .tex section:
Proposition 6.3 in the paper is our Section 20's
||[Y^(h), X]|| <= (3 c_1 + 2) g_M(Y) ||X||. In some renderings it is
6.3; in others it is 6.4 depending on how the hyperref graph resolves.
The tour does not depend on this.
The constant c_1: the paper writes it as 6 pi^2 (1 + 4 c_0), using
a slightly looser Hankel lower bound of 1/(2 pi^2). A direct
recomputation from the displayed inequality gives 3 pi^2 (1 + 4 c_0).
Either is finite and universal; the paper's stated constant is the
headline. The tour states c_1 = 6 pi^2 (1 + 4 c_0) as printed in the
paper, with the discrepancy noted in Section 18.
============================================================
GLOSSARY OF TERMS
============================================================
A few terms recur throughout the tour; the reader may want a
one-sentence definition of each.
adjoint (operator) -- the unique T^* with <T^* u, v> = <u, T v>
for all u, v. Involutive; needed to make B(H) a C*-algebra.
Arveson's commutant-distance formula -- for a vN algebra M in H,
and P = M', 2 dist(T, M) = ||ad_T|_P||_cb, where ad_T(a) =
[T, a].
amplification (of a linear map) -- the entrywise application of
the map to matrices. The "completely bounded" norm is the
supremum of the amplifications' norms.
amplification (of an operator) -- Y^(h) acts diagonally on h
copies of the Hilbert space.
B(H) -- the algebra of bounded operators on a Hilbert space H.
cb norm -- ||u||_cb = sup_h ||u||_h; finite iff u is completely
bounded.
C*-algebra -- norm-closed, *-closed, unital subalgebra of B(H);
equivalently, an abstract algebra satisfying ||a^* a|| = ||a||^2.
central decomposition -- the direct integral presentation of a
vN algebra over its center, with factors as fibers.
Christensen's theorem -- derivation into a faithful
*-representation is inner iff completely bounded.
clifford -- (not used.)
corner -- e L e for a projection e: the subalgebra L truncated to
"both start and end at e".
c_row, c_0, c_1, C_fac, C -- numbered constants fixed in Sections
8, 9, 12, 21, 22.
cyclic vector -- a vector xi with {sigma(a) xi : a in A} dense in
the Hilbert space.
derivation -- a linear map satisfying the product rule. Inner iff
implemented by [V, .].
dist(T, M) -- the distance from T to the vN algebra M, in norm.
factor -- a vN algebra whose center is C 1.
finite von Neumann algebra -- a vN algebra with no pair of
projections p < q with p equivalent to q. In particular, a
finite factor is either finite type I or II_1.
g_P(Y) -- the scalar commutator-size sup over P.
Haar unitary -- a unitary generator of a diffuse abelian
subalgebra. Powers tend weakly to zero.
Hankel matrix -- a matrix with constant anti-diagonals. Appears
here via pCp corner geometry.
hyperreflexivity -- a universal constant K with dist(T, M) <=
K alpha_M(T). Whether or not a given M is hyperreflexive.
II_1 factor -- an infinite-dimensional factor with a faithful
normal trace.
inner derivation -- implemented by a single operator V; see
derivation.
Kaplansky density -- the unit ball of a *-algebra is strongly
dense in the unit ball of its strong closure.
Kirchberg's theorem -- for unital A, similarity for all pi iff
derivations all inner in every faithful *-rep.
L^2(A) -- tracial completion of a finite vN algebra with respect
to tau; the standard representation.
L(F_2) -- the von Neumann algebra of the free group on two
generators; the model free-factor example.
M_h(A) -- the algebra of h-by-h matrices with entries in A.
narrow arc -- the specific function f on the circle used in the
Hankel test (Section 18).
Paulsen's theorem -- unital bounded homomorphism is similar to a
*-hom iff cb, with condition number = cb norm.
Popa's independence theorem -- under irreducibility, a free
subfactor can be placed inside a tracial ultrapower.
projection -- an operator p with p = p^* = p^2.
reduced word -- an alternating product of letters from two
algebras with no two consecutive from the same algebra.
scalar commutator estimate (g_P) -- sup_{a in P, ||a|| <= 1}
||[Y, a]||.
similarity -- existence of a bounded invertible S with S pi S^{-1}
a *-homomorphism.
standard representation -- acting by L_x on L^2(M) for a finite
vN algebra M.
strong operator topology (SOT) -- T_n -> T iff T_n v -> T v in
norm for every vector v.
tau -- a faithful normal trace on a finite von Neumann algebra.
Toeplitz matrix -- a matrix with constant diagonals. Appears here
alongside its Hankel sibling.
ultrapower -- sequences modulo omega-null, in the tracial or
Hilbert norm.
unital -- pi(1) = 1.
von Neumann algebra -- a strongly-closed unital *-subalgebra of
B(H); equivalently M = M''.
weak-operator topology (WOT) -- T_n -> T iff <T_n u, v> ->
<T u, v> for every pair of vectors.
word structure lemma -- the assertion that G acts on each reduced
word level by an operator B_k on the J_k tail independent of
repeated D letters.
============================================================
END OF TOUR
============================================================
The reader has now seen the plot of a paper claiming to prove
Kadison's similarity theorem and, on the way, a universal
hyperreflexivity bound. The argument, if correct, shows that a
single scalar quantity -- g_P(Y) at matrix size 1 -- controls
the entire completely bounded behavior of the associated inner
derivation, uniformly in every matrix size and every representation.
The tour has been in six parts:
PART I: the question and the two theorems.
PART II: the toolbox (Kirchberg, Paulsen, row/column,
cyclic domains).
PART III: the uniform commutator estimate and the reduction
plan.
PART IV: the heart: the free-product corner estimate.
PART V: the ultrapower, factor representations, central
decomposition, derivations.
PART VI: assembling the theorems, hyperreflexivity,
architecture, dependencies.
A patient reader who has made it this far should be able to open
the original paper, look up a specific estimate by name, and
recognize why the estimate has to be there -- and, if they want,
check any single step against the reference cited for it. The
tour deliberately does not assert the paper's correctness; that
is for the mathematical community to decide.
============================================================
END
============================================================