================================================================================
WHY NOTHING HIGHER CAN GO WRONG: THE KADISON-RINGROSE COHOMOLOGY CONJECTURE,
PROVEN FOR UNDERGRADUATES (AND DETERMINED HIGH SCHOOLERS)
================================================================================
Author note for the reader
--------------------------
This file explains, from first principles, a single theorem that answers a
question operator algebraists asked for about fifty years. The theorem says
roughly this: if you take a complex von Neumann algebra M and a bounded
multilinear map f from k copies of M to M (k >= 2) that satisfies the rule
"df = 0" (the exact analogue of a cocycle in ordinary cohomology), then f is
automatically "exact": there is a bounded (k-1)-linear map g with f = dg.
If that sentence means nothing to you yet, do not worry. The next few
sections build every word of it. By the end you will understand:
- what a von Neumann algebra is, concretely;
- what a bounded multilinear map and its Hochschild differential are;
- what a cocycle and a coboundary are;
- what H^k_b(M,M) means;
- why H^1_b(M,M) = 0 was already known (Kadison, Sakai, 1966);
- why H^k_b(M,M) = 0 for k >= 2 resisted all attacks for so long;
- what the new idea is (a single random walk on the unitary group whose
long products look like free Haar unitaries), and
- how that idea turns into a bounded primitive for any cocycle.
Nothing in the proof is hidden behind "it is well known". Classical facts
that we use (Akemann's continuity theorem, Johnson-Kadison-Ringrose normal
reduction, Popa's ultraproduct freeness, Haagerup's length bound) are stated
precisely, and each is either proven here or cited with exact location.
The theorem itself is: for every complex von Neumann algebra M and every
integer k >= 2, H^k_b(M,M) = 0. Combined with the degree-one inner
derivation theorem, this proves the full Kadison-Ringrose cohomology
conjecture in all positive degrees, with ordinary operator-norm boundedness
only. No complete-boundedness condition is assumed.
Part I. Groundwork: what the objects really are.
================================================================================
1. Hilbert space, operators, norms
-----------------------------------
Throughout, H is a complex Hilbert space with inner product <x,y> (conjugate
in the first slot) and norm ||x|| = <x,x>^{1/2}. B(H) is the algebra of
bounded linear maps H -> H, with the operator norm
||T|| = sup { ||Tx|| : x in H, ||x|| <= 1 }.
The *-operation is the adjoint: <T*x, y> = <x, T y>. A C*-algebra is a
closed subalgebra of B(H) closed under *. For example, the n x n complex
matrices M_n(C) form the C*-algebra B(C^n).
Every operator T with ||T|| <= 1 and T*T = TT* = 1 is a unitary. Unitaries
preserve inner products and form a group under multiplication.
A von Neumann algebra M on H is a *-subalgebra of B(H) closed in the weak
operator topology (equivalently, equal to its own double commutant; this is
von Neumann's bicommutant theorem). Concrete examples:
- B(H) itself.
- The n x n matrices M_n(C), acting on C^n.
- The diagonal matrices on C^n; this is abelian and isomorphic to the
functions on an n-point set.
- L^\infty(X, mu) acting by pointwise multiplication on L^2(X, mu) for a
standard measure space (X, mu). This is abelian.
- The hyperfinite II_1 factor R: an infinite-dimensional von Neumann
algebra built as the norm closure of the union of the matrix algebras
M_2(C) in M_4(C) in M_8(C) in ..., each embedded block-diagonally.
A factor is a von Neumann algebra whose center (elements commuting with
everything) is just the scalars. R is the smallest interesting non-abelian
finite factor, and it is the playground for almost every tool in this paper.
2. Traces, the 2-norm, and finite algebras
-------------------------------------------
A trace on a von Neumann algebra M is a linear functional tau: M -> C with
tau(ab) = tau(ba) for all a,b in M,
tau(1) = 1.
It is faithful if tau(x*x) = 0 only when x = 0. It is normal if it is
continuous for the weak operator topology. A finite von Neumann algebra is
one that admits a faithful normal trace; every such algebra is a direct sum
of a finite-dimensional center algebra and a bunch of II_1 summands. A
II_1 factor admits a unique such trace, and every II_1 factor is either
R itself (hyperfinite) or infinite-dimensional and wild.
For a tracial algebra (M, tau), define the tracial 2-norm
||x||_2 = tau(x*x)^{1/2}.
This is a norm on M (faithfulness), and it makes M into a (not complete)
pre-Hilbert space. Its completion is the Hilbert space L^2(M, tau) on which
M acts by left multiplication. We have the inequality, used constantly:
||x||_2 <= ||x||, and |tau(x)| <= ||x||_2,
||xyz||_2 <= ||x|| ||y||_2 ||z||, ||x|| ||y||_2 = ||xy||_2 for
unitary x.
The 2-norm is the natural topology in which "almost equal up to trace" is
measured; it is much finer than the operator norm in arguments (it lets you
approximate an operator norm bound by trace information) and much finer in
the other direction (an operator-norm limit is automatically a 2-norm limit).
A family of unital subalgebras A_1, A_2, ... of a tracial algebra M is
called free if any alternating product
a_1 a_2 ... a_n, a_i in A_{k_i} centered (tau(a_i) = 0), k_i != k_{i+1}
has tau(a_1 a_2 ... a_n) = 0. Free independence is the noncommutative
analogue of ordinary independence; it is the source of all the probabilistic
rigidity in this paper.
A unitary u is Haar if tau(u^k) = 0 for every nonzero integer k. A family
of unitaries (u_i) is a free Haar family if the unital von Neumann algebras
they generate are free and each u_i is Haar. Equivalently, for every finite
set of indices and every nonidentity reduced word q in the free group on
those indices,
tau(q(u)) = 0.
The Haar condition makes distinct reduced words orthonormal objects in
L^2(M, tau). This orthonormality is the fundamental computational tool.
3. Free group length notation
------------------------------
Let F_r be the free group on r generators x_1, ..., x_r. A word is a string
x_{i_1}^{e_1} ... x_{i_q}^{e_q} with e_j = +-1. It is reduced if no
adjacent pair x_i^{+1} x_i^{-1} or x_i^{-1} x_i^{+1} appears after
cancellation. The length |w| is the number of letters in the reduced form;
the identity has length 0. Distinct reduced words are exactly the distinct
elements of F_r, so the whole group is an object of combinatorial geometry.
Notation for the paper, fixed once and for all:
* epsilon = +-1, sigma = +-1.
* "Signed word" means w = x_{i_1}^{e_1} ... x_{i_t}^{e_t} with pairwise
distinct unsigned indices i_1, ..., i_t (a "multiplicative basis" word).
* "Degree-q word" means |w| = q at the level of F_r.
================================================================================
4. Bounded multilinear maps and their norms
--------------------------------------------
Let M be a von Neumann algebra. For k >= 1, a bounded k-linear map
f : M x M x ... x M -> M (k copies of M)
is a function of k algebra elements, linear in each slot separately, with
finite norm
||f|| = sup { ||f(a_1, ..., a_k)|| :
a_j in M, ||a_j|| <= 1 for j = 1, ..., k }.
This is exactly the operator-theoretic analogue of a bounded k-linear form
on a Banach space. For k = 0 we decree C_b^0(M,M) = M with its operator
norm, so a "0-linear bounded map" is just an element of M.
Two structural facts we will use without comment:
(i) A bounded k-linear map is uniquely determined by its values on
k-tuples of unitaries. Indeed, every element of a unital C*-algebra
is a finite linear combination of unitaries (this is the classic
decomposition x = (Re x)^+ - (Re x)^- + i (Im x)^+ - i (Im x)^- with
self-adjoint parts split by positive/negative spectral pieces);
multilinearity extends the values to all k-tuples.
(ii) If M is finite tracial with faithful trace tau, the inequality
||f(a_1, ..., a_k)||_2 <= ||f|| ||a_1|| ... ||a_k|| holds, and
the same estimates pass through 2-norm limits whenever f extends
continuously.
A multilinear map is called separately normal if, for each j, fixing the
other slots leaves an ultraweakly continuous map of the j-th slot. Every
bounded multilinear map has a canonical separately normal "core", but
the core of a cocycle need not be a cocycle. The Johnson-Kadison-Ringrose
reduction (Lemma 5.4 and Theorem 5.6 of JKR 1972, recalled in Section 20
of this text below) sharpens this: for any bounded cocycle f of degree k >= 2,
there is a separately normal bounded cocycle f_0 and a bounded
(k-1)-cochain b satisfying f = f_0 + db. So when constructing a bounded
primitive for f, it suffices to construct one for f_0 and add b. Separate
normality is what lets Akemann's continuity theorem apply, giving uniform
2-norm continuity on operator-norm balls, a hypothesis used in every
averaging step.
5. The Hochschild complex and its cohomology
---------------------------------------------
The Hochschild differential d is the operator
d : C_b^k(M,M) -> C_b^{k+1}(M,M)
defined, at the cochain level, by the alternating formula
(df)(a_1, ..., a_{k+1})
= a_1 f(a_2, ..., a_{k+1})
+ sum_{j=1}^{k} (-1)^j f(a_1, ..., a_j a_{j+1}, ..., a_{k+1})
+ (-1)^{k+1} f(a_1, ..., a_k) a_{k+1}.
The first and last terms are the outer M-bimodule actions (multiply by a_1
on the left of the output, or by a_{k+1} on the right). The middle terms
merge adjacent slots; each such merge carries a sign that reflects how far
to the right the merge sits. In degree k = 0 the formula degenerates to
(db)(a) = a b - b a, b in M,
which is the commutator or inner derivation. In degree k = 1 the formula
is the ordinary two-term Leibniz rule for a linear map f: M -> M:
(df)(a,b) = a f(b) - f(ab) + f(a) b.
The key algebraic identity is d^2 = 0: each face of a (k+2)-cube appears
twice with opposite signs, so the sum of the faces cancels. This is just
the simplicial boundary calculation, and it holds verbatim in the operator
algebra setting because d is defined purely algebraically. Consequently
the image of d sits inside the kernel of d, and we can form quotients.
A cocycle of degree k is a bounded k-cochain f with df = 0. A coboundary
of degree k is a bounded k-cochain f of the form f = dg, where g is a
bounded (k-1)-cochain. The k-th bounded Hochschild cohomology group of M
with coefficients in M is
H^k_b(M,M) = ker(d: C_b^k -> C_b^{k+1}) / im(d: C_b^{k-1} -> C_b^k).
The denominator is the honest algebraic image, not its closure. So the
statement H^k_b(M,M) = 0 for k >= 2 says: any bounded cocycle is a bounded
coboundary, with the primitive explicitly in C_b^{k-1}(M,M), no "modulo
closure" tolerated.
The variant "continuous Hochschild cohomology" without a subscript b is
the same construction with the ordinary continuous k-linear maps replaced
by norm-continuous ones; for the bounded case these two coincide by
definition. So the paper's H^k_b = 0 is the strongest possible statement
in the category of bounded cochains.
6. What was already known, and what was open
---------------------------------------------
The question goes back to Kadison and Ringrose. In degree one, Kadison
(1966) and Sakai (1966) proved, independently, that every derivation of a
von Neumann algebra M into itself is inner, i.e. has the form a -> a b - b a
for some b in M. In the notation above, this is exactly H^1_b(M,M) = 0.
The higher degrees resisted.
Kadison and Ringrose (1971, 1972) proved vanishing for type I von Neumann
algebras (all of B(H), or direct sums with central projections whose fibers
are B(H_s)), and for hyperfinite algebras (those obtained as limits of
finite-dimensional matrix algebras). Johnson, Kadison, and Ringrose (1972)
proved that any bounded cocycle differs from its separately normal core by
a bounded coboundary. Akemann (1967) gave the continuity theorem used in
every averaging step. These are the classical inputs.
The next generation of results attacked the II_1 summand. Christensen,
Effros, and Sinclair (1987) proved vanishing for any M that absorbs the
hyperfinite II_1 factor R tensorially; this covers every algebra without a
II_1 summand. Sinclair and Smith (1998) proved vanishing for separable
predual M with a Cartan maximal abelian subalgebra, i.e. an MASA A for which
there is a trace-preserving conditional expectation M -> A and a cyclic
vector in L^2(M) for A. Cameron (2009) removed separability from that
result for II_1 factors. Christensen, Pop, Sinclair, and Smith (2003) proved
vanishing for every II_1 factor with property Gamma (a weak asymptotic
abelianness traced back to Murray and von Neumann), separability not needed.
Qian and Shen (2015) extended this to generalised property Gamma for
separable predual, and to type II_1 algebras with nontrivial center under
mild spectral hypotheses. Pop and Smith (2010) proved H^2_b = 0 for
tensor products A x B of type II_1 algebras, separability not needed.
Every one of these uses a structural hypothesis: type I, or hyperfinite
absorption, or a Cartan subalgebra, or property Gamma, or a tensor product
factorisation. The conjecture asks for one uniform statement covering
every II_1 algebra, i.e. every finite factor at every possible depth of
non-hyperfiniteness. The point of the present work is that no such
hypothesis is needed.
7. The main theorem
--------------------
Theorem (Kadison-Ringrose cohomology conjecture, positive answer).
Let M be a complex von Neumann algebra and let k >= 2 be an integer.
Every bounded k-cocycle f in C_b^k(M,M) is a bounded coboundary. Precisely,
there is a bounded (k-1)-cochain g in C_b^{k-1}(M,M) with dg = f, and one
may choose g with ||g|| <= ||f||. Hence H^k_b(M,M) = 0.
Combined with the degree-one inner-derivation theorem (Kadison-Sakai) this
gives H^k_b(M,M) = 0 for every k >= 1 and every complex von Neumann algebra
M. Boundedness is the ordinary operator norm; complete boundedness is not
assumed, and the primitive lives in the actually bounded cochains.
The next part of the text builds the new object that drives the proof: a
single random walk on the unitary group U(M) whose long products are
asymptotically free.
Part II. The random walk: one measure whose long products are free.
================================================================================
8. The idea, and why we want this
----------------------------------
The proof rests on a single probabilistic object. We build a probability
measure mu on the unitary group U(M) of a tracial von Neumann algebra (M,
tau) with the following two features at once:
(a) mu is symmetric, i.e. the law of U^{-1} is the law of U, and the
positive-mass atoms of mu are 2-norm dense in U(M). The second half
says mu's support generates M as a von Neumann algebra; without
this, average values of a linear map T over mu would only be
sensible on a subgroup, not on M.
(b) There is an increasing sequence n_1 < n_2 < ... in the positive
integers such that the n_l-th convolution power mu^{*n_l} - i.e.,
the law of U_1 U_2 ... U_{n_l} for independent U_i sim mu - has
approximate free Haar joint moments in probability.
The second condition is the substance. Fix r independent random products
G_1^(l), ..., G_r^(l) of length n_l. For every finite set of words w in
F_r and every eta > 0, we want
Pr( max_w |tau(w(G_1^(l), ..., G_r^(l)))| >= eta ) -> 0 (l -> inf).
That is: the joint scalar trace moments of (G_1^(l), ..., G_r^(l)) become
those of a free Haar family as l grows. Scalar moments are exactly the
information needed by the Liouville theorem in Part III: to identify the
value T(1) x at every x in M, the proof needs to know how T behaves on
long, statistically free unitary products.
The measure is FIXED IN ADVANCE, before any T (harmonic map) or f
(cocycle) is chosen. Only the family of moment tests changes with l; the
underlying randomness is one single mu. That is what makes the resulting
Liouville theorem usable iteratively in Part V: we do not re-randomise
for each application.
Theorem (walk: statement of thm:walk).
Let (M, tau) be a finite type II_1 von Neumann algebra with separable
predual and faithful normal tracial state. There exist a symmetric
countably supported probability measure mu on U(M) and an increasing
n_l -> infinity with the properties above.
The rest of Part II proves this. The construction is explicit: mu is a
mixture of a fixed dense base measure mu_0 and a sequence of level-j
measures mu_j, with weights p_j, in such a way that each finite moment
test is satisfied at a fixed level with overwhelming probability.
9. Where free Haar unitaries come from: Popa's freeness
--------------------------------------------------------
The only source of true freeness in the proof is a theorem of Popa (1995,
2014) about the tracial ultrapower. Let (P, tau_P) be a tracial von
Neumann algebra and omega a free ultrafilter on the naturals. The tracial
ultrapower is
P^omega = l^inf(N, P) / I_omega, I_omega = {(x_m):
lim_{m -> omega} ||x_m||_2 = 0}.
Equivalently P^omega is the quotient of bounded sequences by those whose
2-norm vanishes along omega. It is a tracial algebra with trace
tau^omega((x_m)) = lim_{m -> omega} tau_P(x_m), and P embeds as the
subalgebra of constant sequences. Coordinates that differ on a set of
omega-measure zero become equal in P^omega; anything the ultrafilter
"sees" becomes exact.
Lemma (Popa's ultrapower freeness; Lemma 0.1(a) of Popa 2014).
Let P be a II_1 factor, and let B be a unital von Neumann subalgebra of
P^omega with separable predual. Then there is a diffuse abelian von
Neumann subalgebra of P^omega free from B. Moreover, for every d >= 1,
the ultrapower P^omega contains Haar unitaries u_1, ..., u_d such that
W*(B, W*(u_1), ..., W*(u_d))
is a free family of subalgebras of P^omega.
Here "diffuse abelian" means a commutative von Neumann algebra isomorphic to
L^inf(0,1), hence with no minimal projections, and such an algebra always
contains a Haar unitary (e.g. the generator exp(2 pi i t) on L^2(0,1)). The
idea of the proof is: apply Popa 2014, Theorem 0.1(a), with the target
"coordinate algebra" and "coordinate subalgebra" both equal to P. The
required non-intertwining condition holds because P is a factor, so no
piece of it embeds into a scalar corner. Then adjoint one Haar unitary
to B, repeat d times.
That freeness is genuine, not approximate, is the crucial point. Popa's
theorem trades the abstract statement "we can pick free unitaries" for the
concrete coordinate statement that "a typical n-tuple of unitaries chosen
from some m-approximant is nearly free, and the leftover error vanishes
along omega".
10. One fresh label per path
-----------------------------
The next building block is combinatorial. Suppose we are handed free
unital subalgebras B, W*(u_1), ..., W*(u_d) of some tracial algebra with
each u_j Haar. Let y_1, ..., y_r be path expressions in the alphabet
B-cup {u_1^{+-1}, ..., u_d^{+-1}}. A "path expression" simply means a
finite product of fixed coefficients (from U(B)) and formal letters
u_j^{+1} or u_j^{-1}. Occurrences and labels are counted BEFORE any
algebraic cancellation. We ask: when is
tau(w(y_1, ..., y_r)) = 0 for every nonidentity w in F_r?
The answer is that it is enough for a single fresh label to appear in each
path.
Lemma (distinguished letters).
Let B, W*(u_1), ..., W*(u_d) be free subalgebras of a tracial algebra, with
each u_j Haar. Let y_1, ..., y_r be path expressions with coefficients in
U(B). Suppose that for each i there is a label h_i that occurs exactly
once in the entire list of path expressions, and that its occurrence belongs
to y_i. Then tau(w(y_1, ..., y_r)) = 0 for every nonidentity w in F_r.
The proof puts all non-distinguished letters into a coefficient algebra,
B_0 = W*(B, u_j : j not in {h_1, ..., h_r}), C = W*(u_{h_1}, ..., u_{h_r}),
so that B_0 and C are free. Grouping expansions of each y_i via the
distinguished letter h_i makes each y_i the outer factor times a
distinguished letter times an inner factor. When we iterate w on these
groupings, the resulting string of surviving distinct letters is itself
reduced: an inverse pair would require adjacent subscripts to match, which
a reduced w excludes. Centering every coefficient breaks the surviving
alternating product into alternating centered pieces from B_0 and C, with
at least one C-piece, so freeness plus Haar-ness gives trace zero. This is
a covariance statement: "distinct fresh labels behave as independent
centered generators, even when the coefficient algebra is large".
The corollary we actually use is "finite simultaneous path requirements":
for a finite set F of fixed unitaries in M, a d-tuple of formal labels, and
a collection of tests, each consisting of a finite list y_1, ..., y_r of
path expressions and a choice of nonidentity w in F_r, we can find
u_1, ..., u_d in U(M) satisfying all tests to tolerance eps simultaneously,
as long as every test has a distinguished letter as above. This is where
Popa's theorem and the distinguished-letters lemma dovetail: for a fixed
finite collection of tests, the conclusion lives in M^omega; lifting a
self-adjoint logarithm coordinatewise brings it back into M with finite (but
admissible) error. For a genuinely non-factor M with a central direct
integral decomposition, one uses measurable selection + Kaplansky density to
assemble the coordinate unitaries into a measurable section, then integrates
the trace.
11. The measure, level by level
--------------------------------
The walk itself is built in stages. The idea is that "makes a fixed family
of tests succeed" only requires finitely many random-unitaries of bounded
size, so we can pay a small weight for each stage and let later stages
dominate.
Fix a 2-norm dense sequence (v_j) in U(M) and let
mu_0 = sum_{j >= 1} 2^{-j-1} (delta_{v_j} + delta_{v_j^{-1}}).
This mu_0 is symmetric and carries positive weight on each v_j, so its
support is 2-hull-dense in U(M). mu_0 is our "background": a fixed, mild
source of density.
For each level l >= 1, set p_l = 2^{-2^l} (double exponentially decaying
level weights), n_l = p_l^{-3/2} = 2^{3 * 2^{l-1}} (path length at level l),
K_l = l * n_l (total increments across l paths), and d_l = K_l^4 (number of
labels at level l). Then construct a finite symmetric measure
mu_l = (1/(2 d_l)) sum_{j=1}^{d_l} (delta_{u_{l,j}} + delta_{u_{l,j}^{-1}}).
The level-l unitaries u_{l,j} are chosen inductively at the (l)-th stage.
The full measure is
mu = p_0 mu_0 + sum_{l >= 1} p_l mu_l,
where p_0 = 1 - sum_{l >= 1} p_l > 0 because p_l decays fast.
The choice at level l goes as follows. Fix a finite symmetric set F_l of
base atoms with mu_0(F_l) > 1 - K_l^{-2}, and a finite coefficient alphabet
consisting of F_l and the unitaries u_{j,h}^{+1}, u_{j,h}^{-1} for earlier
stages j < l. Introduce d_l unknown unitary labels. Enumerate every
ordered list of l paths of length n_l in this alphabet and in these signed
labels satisfying two combinatorial conditions:
(p1) each path contains at least one unknown label;
(p2) no unknown label occurs twice anywhere in the whole list.
For every such list, and every nonidentity reduced word of length at most l
on l generators, impose the inequality
|tau(w(y_1(u), ..., y_r(u)))| < 1/l.
There are finitely many lists and finitely many words, and every test has
its distinguished label (choose any single occurrence of an unknown label
in each path). So by the finite-path-tests corollary, we can choose
u_{l,1}, ..., u_{l,d_l} in U(M) so that all of them hold simultaneously.
Now sample l independent paths of length n_l, so that K_l = l * n_l
increments are involved. An increment has four "bad" ways to fail:
(b1) a level greater than l is chosen;
(b2) a level-zero increment falls outside F_l;
(b3) some level-l label is repeated;
(b4) some path misses level l entirely.
Union-bounding each probability gives
Pr(bad at level l)
<= K_l sum_{j > l} p_j
+ K_l p_0 mu_0(U(M) \ F_l)
+ (K_l choose 2) / d_l
+ l (1 - p_l)^{n_l}.
The right-hand side tends to 0 as l -> infinity:
- sum_{j > l} p_j <= (16/15) p_l^2 because p_{l+1} = p_l^2 and p_l <= 1/4;
hence K_l * sum_{j > l} p_j <= (16/15) l sqrt(p_l).
- K_l p_0 mu_0(U(M) \ F_l) <= K_l / K_l^2 = 1/K_l.
- (K_l choose 2) / d_l <= K_l^2 / (2 d_l) = 1/(2 K_l^2).
- l (1 - p_l)^{n_l} <= l exp(-p_l^{1/2}) -> 0.
On the good event (complement of the four failures), the sampled paths are
exactly one of the lists tested at stage l, and their scalar traces meet
the 1/l tolerance for every word of length at most l. So for a fixed
finite family W of words on r generators and a tolerance eta > 0, if we
sample any r independent paths of length n_l and extend to l paths, we get
the desired bound as soon as l >= max{|w| : w in W} and 1/l < eta.
Finally: mu is symmetric and countably supported by construction. Its base
part mu_0 has positive weight, so each v_j is a positive-mass atom, and the
von Neumann algebra generated by them is M (the v_j are 2-norm dense, hence
strong-dense, so generate every unitary, hence M). The convergence above
concerns scalar moments only; it does not assert operator-norm convergence
of sampled polynomials. This is enough for our application, and it is the
only kind of convergence the Liouville argument uses.
================================================================================
Part III. The Liouville theorem: harmonic maps are left multipliers.
================================================================================
12. Continuity: bounded-ball 2-continuity and ultrapower transfer
------------------------------------------------------------------
Let (N, rho) be a finite tracial von Neumann algebra with faithful
normal trace rho; write ||x||_2 = rho(x*x)^{1/2}. A linear map R: N -> N
is bounded-ball 2-continuous if
sup_j ||x_j|| < infinity and ||x_j||_2 -> 0
implies ||R(x_j)||_2 -> 0.
This is 2-norm continuity restricted to bounded operator-norm balls. It
is strictly weaker than full 2-norm continuity: a map can be fine on
bounded sets and blow up on unbounded ones. It is exactly the hypothesis
that makes the ultrapower lift work.
Lemma (equivalence with a uniform modulus).
For a bounded linear R: N -> N, bounded-ball 2-continuity is equivalent to:
for every r < infinity and eps > 0 there is delta > 0 such that
||x|| <= r, ||x||_2 < delta => ||R(x)||_2 < eps.
Proof sketch. If the modulus fails for some (r, eps), pick x_j with
||x_j|| <= r, ||x_j||_2 < 1/j, ||R(x_j)||_2 >= eps; that contradicts
bounded-ball 2-continuity. Conversely the sequential formulation is the
particular case "r fixed, eps = 1/j" plus a diagonal argument.
With the uniform modulus in hand, use it twice. First, it shows that
the lift of R to the tracial ultrapower,
R^omega: N^omega -> N^omega, R^omega([(x_j)]) = [(R(x_j))],
is well-defined: if (x_j) has vanishing 2-norm along omega and is uniformly
bounded, the modulus drives ||R(x_j)||_2 to zero as well, so [(R(x_j))]
is zero in N^omega. Second, the map R^omega inherits the same modulus:
if [(x_j)] has quotient-norm bounded by r and 2-norm < delta/2, we may
assume sup_j ||x_j|| <= r + 1 and Omega-large norms < delta, so the pointwise
modulus gives ||R^omega([(x_j)])||_2 <= eps/2 < eps.
bounded-ball 2-continuous in N, so is R^omega in N^omega; applying the
lemma twice gives any ultrapower lift (R^omega)^nu on (N^omega)^nu.
13. From a harmonic map to a martingale
----------------------------------------
Now let (M, tau) be type II_1 with separable predual and faithful normal
tracial state, let mu be the walk from Part II, and let T: M -> M be
bounded, 2-continuous on bounded balls, and HARMONIC for mu:
T(x) = E_{U ~ mu}[T(x U) U^*] for all x in M. (eq:harm)
Note the shape: T is averaged over right-multiplication by a random
unitary, and the result is then right-multiplied by the inverse of the
same unitary. This is the operator-algebra version of "T is invariant
under the random walk".
Write C = ||T||. Define D: U(M) -> M by
D(g) = T(g) g^*.
The map D takes values in the operator-norm ball of radius C, in M, hence
in the Hilbert space H = L^2(M, tau). Consider the random walk W_n =
U_1 U_2 ... U_n on U(M), where U_i are independent with law mu, and the
sigma-fields F_n = sigma(U_1, ..., U_n). Let Z_n = D(W_n).
Proposition (D(W_n) is a martingale).
The sequence (Z_n) satisfies E[Z_{n+1} | F_n] = Z_n almost surely, and
the L^2 norms E||Z_n||_2^2 are bounded by C^2.
Proof. Conditioning on F_n fixes W_n and W_n^*. By the harmonicity
property applied to x = W_n (which is unitary, hence bounded),
E[T(W_n U_{n+1}) U_{n+1}^* | F_n] = T(W_n) = Z_n W_n.
Then
E[D(W_{n+1}) | F_n] = E[Z_n W_n U_{n+1} (W_n U_{n+1})^* | F_n]
= Z_n W_n U_{n+1} U_{n+1}^* W_n^*
= Z_n W_n W_n^*
= Z_n.
Bound follows from ||D(g)|| <= C. ()
Because (Z_n) is a bounded martingale in the Hilbert space L^2(Omega; H),
it converges almost surely and in L^2 to some Z_infty:
E ||Z_infty - Z_n||_2^2 -> 0. (eq:mtg)
This is the standard L^2 martingale convergence theorem (Doob). Note
that the target Z_infty is an H-valued random variable, not (yet) an
element of M. The plan is to prove Z_infty is deterministic, then
recover T.
14. The constant-endpoint case and the nonconstant case
-------------------------------------------------------
Case A: Z_infty = c almost surely for some fixed c in H. Then by Doob,
Z_n = E[Z_infty | F_n] = c almost surely for every n; in particular
Z_0 = c. Because Z_0 = D(W_0) = D(1) = T(1), we get c = T(1). Fix any
atom u_j of mu. Since u_j is a possible value of W_1 (with positive
probability), Z_1 = T(u_j) u_j^* must equal c in L^2, so tau-a.e. T(u_j) =
c u_j. Because each u_j has ||u_j|| = 1 and both sides are bounded
operators, equality in L^2 upgrades to operator equality (both sides live
in M; differences with zero 2-norm vanish by faithfulness). So T(u_j) =
c u_j for every atom.
Now extend to arbitrary unitaries u in U(M) by the density of atoms and
the bounded-ball 2-continuity of T. Pick atoms u_j with ||u_j - u||_2 -> 0;
also ||u_j^{-1} - u^{-1}||_2 = ||u_j^* - u^*||_2 = ||u_j - u||_2 -> 0 by the
unitary invariance of the 2-norm. Then
||T(u) - T(u_j)||_2 -> 0 (bounded-ball 2-continuity)
||c u - c u_j||_2 <= ||c|| ||u - u_j||_2 -> 0 (unitaries).
So ||T(u) - c u||_2 = 0, whence T(u) = c u by faithfulness. Every element
of a unital complex C*-algebra is a finite linear combination of unitaries,
so multilinearity (linearity) gives T(x) = c x for all x in M. Setting
c = T(1) recovers T(x) = T(1) x. This is the Liouville conclusion for
Case A.
Case B: Z_infty is NOT deterministic. Then the pushforward of Z_infty's
law is a Borel probability measure on H not concentrated on a single point,
so its topological support contains two distinct points a, b. We will
show that any pair of distinct support points leads to a contradiction using
the rigidity theorem of Part IV.
First: a and b live in M, and in the operator-norm ball of radius C. A
subsequence of Z_n converges almost surely in H to Z_infty (Doob). The
set K_C = {x in M : ||x|| <= C} is 2-norm closed in H: given a 2-convergent
sequence y_j -> y in K_C, ultraweak compactness of the operator-norm ball
gives a subnet y_j(alpha) -> x ultraweakly, with x in K_C. Testing against
tau(v^* .) for every v in M identifies the ultraweak limit with the 2-norm
limit y (since M is dense in H). With Z_infty in K_C almost surely, and
a, b in the support, a, b are in K_C.
Second: for any delta > 0, there are eta_delta > 0 and N_delta such that
whenever n >= N_delta and G has law mu^{*n},
Pr( ||D(G) - a||_2 < delta AND ||D(G^{-1}) - b||_2 < delta )
>= eta_delta. (eq:2ep)
The proof uses FORWARD and REVERSE endpoints of the walk. Write G =
U_1 ... U_n, h = floor(n/2), and consider
A_n = D(U_1 ... U_h), B_n = D(U_n^{-1} ... U_{n-h+1}^{-1}).
These depend on disjoint increments, hence are independent, and by
symmetry each has the law of Z_h. Applying the martingale Cauchy
estimate to the forward string of increments and separately to the
reversed inverse string:
E ||D(G) - A_n||_2^2 <= e_h,
E ||D(G^{-1}) - B_n||_2^2 <= e_h,
where e_h = E ||Z_infty - Z_h||_2^2 -> 0. Fix delta and use the support
property to get positive probabilities
p_a = Pr( ||Z_infty - a||_2 < delta/3) > 0,
p_b = Pr( ||Z_infty - b||_2 < delta/3) > 0.
For h large, mean-square convergence gives
Pr( ||Z_h - a||_2 < 2 delta/3 ) >= p_a/2,
Pr( ||Z_h - b||_2 < 2 delta/3 ) >= p_b/2,
and independence of A_n, B_n (from the walk, not from G itself) plus Markov
gives
Pr(both in delta-balls of a, b) >= p_a p_b / 4 - 18 e_h / delta^2.
For h large enough, that lower bound is at least p_a p_b / 8. ()
Third: the finite-requirements selection. Fix m >= 1, use tolerance 1/m
in (eq:2ep), and let E_{m,n} be the event that m independent length-n
products all get D(G_i) within 1/m of a and D(G_i^{-1}) within 1/m of b.
Independence gives Pr(E_{m,n}) >= c_m for large n, where c_m = (eta_{1/m})^m.
On an ultrapower sample, impose moreover:
- scalar trace moments: |tau(q(G_1, ..., G_m))| < 1/m for every nonidentity
reduced word q of length at most m;
- block identity estimates |tau(...)| < 1/m for every signed word with
distinct unsigned indices (block law of whole tuple vs first block);
- the block identity ||D(v(G)) - D(G_{i_1}^{sigma_1})||_2 < 1/m.
The first family is what thm:walk gives; the second and third use the
orthogonality of martingale increments applied to concatenated independent
blocks: if v = G_{i_1}^{sigma_1} ... G_{i_t}^{sigma_t} has pairwise
distinct unsigned indices, then concatenating the reversed/inverted
individual blocks gives a (t n)-long walk, and the pair
(D(v(G)), D(G_{i_1}^{sigma_1})) has the law of (Z_{t n}, Z_n). That gives
E ||D(v(G)) - D(G_{i_1}^{sigma_1})||_2^2 <= e_n,
and Markov plus a union bound over finitely many v in the test family
gives the desired event with probability at least c_m / 2. Choose a
DETERMINISTIC tuple (u_1^(m), ..., u_m^(m)) in that intersection.
Fourth: pass to the ultrapower. Fix a free ultrafilter omega_1 on m,
set u_i^(m) = 1 for m < i, and take
N = M^{omega_1}, g_i = [(u_i^(m))], R = T^{omega_1}.
Then g_i are unitaries in N; the moment conditions become EXACT:
tau^{omega_1}(q(g_1, ..., g_r)) = 0 for every nonidentity q;
i.e. (g_i) is a genuinely free Haar family in N. The selected multipliers
become the first-letter identity: for every nonempty signed word w =
g_{i_1}^{sigma_1} ... g_{i_t}^{sigma_t} with pairwise distinct unsigned
indices,
R(w) = a w if sigma_1 = +1,
R(w) = b w if sigma_1 = -1. (eq:1st)
The constants a, b are the trace-preserving constant images of a, b from M.
By the rigidity theorem (Part IV, Theorem 15.1 below), applied to (N,
tau^{omega_1}), R, (g_i), a, b, we conclude a = b in N. The constant
embedding is trace-preserving and faithful, so a = b in M, contradicting
distinctness. Hence Case B is impossible.
The Liouville conclusion follows.
Theorem (Liouville). Let (M, tau) be type II_1 with separable predual
and faithful normal trace, mu the walk from Part II, and T: M -> M
bounded, linear, and 2-continuous on every operator-norm ball. If T is
harmonic for mu, then T(x) = T(1) x for every x in M.
The rigidity theorem invoked above is stated and proven in the next part.
Note that the entire argument uses only scalar trace moments of (g_i) and
the first-letter identity (eq:1st). No freeness from a or b is used.
================================================================================
Part IV. Rigidity: the first-letter identities force a = b.
================================================================================
15. The rigidity theorem, and its plan
---------------------------------------
We now prove the abstract structural statement that pins a, b together
from the data delivered by the walk. Let (N, tau) be a finite von Neumann
algebra with faithful normal tracial state; the reader should think of N
as the ultrapower M^{omega_1} of Part III.
Theorem (Rigidity; Theorem 4.1 of the paper).
Let (N, tau) be finite with faithful normal trace. Let R: N -> N be
bounded and complex linear, satisfying bounded-ball 2-continuity: if
sup_n ||x_n|| < infinity and ||x_n||_2 -> 0, then ||R(x_n)||_2 -> 0.
Let (g_j) be a free Haar family in N, and let a, b in N. Suppose that for
every nonempty signed word w = g_{i_1}^{e_1} ... g_{i_t}^{e_t} with
e_j in {+1, -1} and pairwise distinct indices i_1, ..., i_t,
R(w) = a w if e_1 = +1,
R(w) = b w if e_1 = -1. (eq:first-letter)
Then a = b.
The plan has four steps.
Step 1 (averaging over many free generators). Set s_r = r^{-1/2}
sum_{j=1}^r g_j. Orthogonality of Haar words gives ||s_r||_2 = 1, and the
Haagerup free-group length estimate gives ||s_r|| <= 2 uniformly in r. Pass
to a free ultrafilter omega_2: put s = (s_r)_{omega_2} in N^omega_2, and
S = R^omega_2. Then s is a well-behaved element of the ultrapower whose
scalar moments match a "quarter-circle" spectral measure on [0, 4].
Step 2 (blocks). Expanding powers of s in terms of (s^* s)^d and its
variants introduces repeated unsigned indices in the surviving words. A
combinatorial argument ("block lemma") shows those repetitions contribute
an operator-norm error tending to zero, so the first-letter identity
extends to ALL polynomial expressions s p(s^* s) and their adjoints.
Step 3 (polar unitary). Compute the moments of s^* s, identify them with
the Marchenko-Pastur law dnu(x) = (1/(2pi)) sqrt((4-x)/x) dx on (0, 4),
observe no atom at 0, and conclude s = V |s| with V a unitary in N^omega_2.
Bounded-ball 2-continuity passes the block identities down to V^k and V^{-k}.
Step 4 (spectral). Show that a unitary V with S(V^k) = a V^k and
S(V^{-k}) = b V^{-k} for all k >= 1 forces a = b, using uniformly
bounded sine sums and a partition of the unit circle into short arcs.
Sections 16-18 supply the four steps.
16. The Haagerup free-group length estimate
--------------------------------------------
This is the key quantitative tool: it bounds the operator norm of a
length-q word polynomial in free Haar unitaries by the L^2 norm of the
coefficients times a linear factor in q.
Lemma (Haagerup; Lemma 1.4 of Haagerup 1979).
Let F_r = <x_1, ..., x_r> be the free group on r generators and lambda its
left regular representation on l^2(F_r). For a finitely supported family
(c_w) on the reduced words of length exactly q >= 0:
|| sum_{|w|=q} c_w lambda(w) || <= (q + 1) (sum_{|w|=q} |c_w|^2)^{1/2}.
(eq:haagerup)
Moreover, if g_1, ..., g_r have free Haar joint moments in a faithfully
tracial von Neumann algebra, every scalar-coefficient word polynomial in
them has the same norm as the corresponding polynomial in lambda(x_1), ...,
lambda(x_r), and the same bound holds.
Proof of the length estimate, in case the reader wants to see where q+1
comes from. Write L = sum_{|w|=q} c_w lambda(w) and act on the standard
basis e_x of l^2(F_r). For a product wx, classify the intersection of w
and x by the number t of cancellations at the joining boundary (t in
{0, 1, ..., q}). After cancellations, we can factor (uniquely):
wx = u h, w = u v, x = v^{-1} h,
with |u| = q - t and |v| = t, all three products uv, v^{-1} h, uh reduced.
Focus on fixed t. An advertised output word uh (with type index h) is
produced from inputs x = v^{-1} h with |v| = t. Grouping by the suffix h
of length >= t makes L_t a block-diagonal sum of matrices B_h, whose
entries are indexed by v (column, |v|=t) and u (row, |u|=q-t). Within a
matrix B_h, the map (u, v) -> uv is injective because |u| is fixed, so
each column has at most one nonzero entry, and ||B_h|| <= ||B_h||_HS. Sum
over h to get ||L_t|| <= (sum |c_w|^2)^{1/2}. There are q + 1 possible
values of t, so triangle inequality gives (eq:haagerup).
For the norm transfer to a tracial algebra: the operator norm on a finite
tracial algebra satisfies ||A|| = lim_{n -> infinity} tau(A^n)^{1/n} for
A >= 0 (proved by writing the spectral projection onto (c, infinity) for
c < ||A||, which has positive trace; both inequalities follow from that).
But the moments (p(g)^* p(g))^n for a word polynomial p in a free Haar
family match those of p(lambda) computed in the free group von Neumann
algebra, because the traces of all *-monomials agree and both traces are
faithful. So ||p(g)|| = ||p(lambda)|| in the free Haar case. The 2-norm
is even easier, from orthonormality of reduced words.
17. Blocks: from distinct indices to ordered products
-----------------------------------------------------
Under the hypothesis of the rigidity theorem, put
s_r = r^{-1/2} sum_{j=1}^r g_j. (||s_r||_2 = 1, ||s_r|| <= 2.)
s = (s_r)_{omega_2} in N^omega_2.
S = R^omega_2.
Our goal in this section is to extend the first-letter identity from words
with distinct unsigned indices to a larger class: products of "blocks"
B_d(z) = z (z^* z)^d, d >= 0, evaluated at s.
Lemma (block identity).
For every k >= 1 and every d_1, ..., d_k >= 0, consider
P(z) = B_{d_1}(z) B_{d_2}(z) ... B_{d_k}(z).
Then in N^omega_2,
S(P(s)) = a P(s),
S(P(s)^*) = b P(s)^*. (eq:block-ids)
The proof is a coefficient estimate plus a small-norm error. Fix the
product, with total degree L = sum_{j=1}^k (2 d_j + 1).
Step 1 (block structure). Each block B_d(s_r) expands as
B_d(s_r) = s_r (s_r^* s_r)^d = sum over words of length 2d+1.
Every term in this expansion has an alternating sign pattern +, -, +, -,
..., + across the letters of s_r, s_r^*, ..., s_r. Free reduction of a
nonempty word of the form +, -, +, ..., + can only delete adjacent
opposite-sign pairs; deleting such pairs preserves the alternating pattern,
and a word that begins and ends positively stays nonempty (the excess of
positive letters always exceeds the count of negatives by exactly one).
So every reduced word coming from a single block is nonempty and begins
and ends positively. At the boundary between consecutive blocks, two
positive-endpoint letters meet, so no cancellation occurs across the
boundary. Therefore every reduced word appearing in the expansion of
P(s_r) is nonempty and begins and ends positively. Adjoints reverse the
block order and all signs, so P(s_r)^* has negative endpoints.
Step 2 (coefficient bound). Write
P(s_r) = sum_{q=1}^L sum_{|w|=q} c_w(r) w(g).
For a fixed reduced word w of length q, the set of free reductions of
original terms in the expansion that produce w corresponds to a
noncrossing partial matching on the L original positions: (L-q)/2 pairs of
canceled positions and L - q unmatched positions that spell w. Noncrossing
partial matchings on L positions can be encoded injectively by labeling
each position as "opening endpoint", "closing endpoint", or "unmatched"
(a closing endpoint must pair with the last unpaired opening endpoint), so
there are at most 3^L of them. For a fixed matching pattern and fixed w,
each canceled pair has equal unsigned labels and opposite signs, so has
at most r choices; unmatched positions are pinned by w. Every original
term in the expansion of P(s_r) has coefficient r^{-L/2} (from the r^{-1/2}
in s_r times L factors). Hence for every q and every w:
|c_w(r)| <= 3^L r^{(L-q)/2} r^{-L/2} = 3^L r^{-q/2}.
(eq:coef)
The same bound holds for the adjoint polynomial.
Step 3 (discard repeated surviving indices). Let E_{r,q} be the sum of
all length-q terms in the expansion whose surviving reduced word has some
repeated unsigned index. The number of possible surviving words of this
kind is at most 2^q binom(q, 2) r^{q-1}: choose the pair of positions that
share an index, choose the common label, choose the labels of the remaining
q - 2 positions, choose signs. (This count is an overestimate; it just
counts strings, not reduced ones.) By orthonormality of Haar words and the
coefficient bound above,
||E_{r,q}||_2^2 <= 3^{2L} 2^q binom(q,2) r^{-1}, 2 <= q <= L.
Applying the Haagerup length inequality (Section 16) once more, aggregated
over q:
E_r = sum_{q=1}^L E_{r,q}, ||E_r|| <= sum_{q=2}^L (q+1) ||E_{r,q}||_2
= O_L(r^{-1/2}) -> 0. (eq:discard)
The bound is uniform in r because L is fixed before choosing r.
Step 4 (apply the first-letter identity). Set Q_r = P(s_r) - E_r.
Every surviving word in Q_r has distinct unsigned indices and begins
positively. By (eq:first-letter), R(Q_r) = a Q_r. Moreover,
||P(s_r)|| <= 2^L (bounded because each B_d is a contraction times
appropriate power, and ||s_r|| <= 2), so (Q_r) is bounded. Then
||R(P(s_r)) - a P(s_r)|| <= (||R|| + ||a||) ||E_r|| -> 0.
Pass to the ultrapower to conclude S(P(s)) = a P(s). The adjoint case is
identical with b: every surviving word of Q_r^* has distinct unsigned
indices and begins negatively, so R(Q_r^*) = b Q_r^*, and the same bound
with b gives S(P(s)^*) = b P(s)^*.
================================================================================
18. A polar unitary, and the transfer of the block identities
-------------------------------------------------------------
We now push S(P(s)) = a P(s) from polynomial expressions in s down to
powers of the polar part of s. The key facts are: (i) the operator s^* s
has no kernel, so polar decomposition of s yields a genuine unitary; (ii)
polynomial approximations to (s^* s)^{1/2} converge in operator norm.
Write A = s^* s, so 0 <= A <= 4 in N^omega_2. First we compute the
moments of A. For fixed d >= 1, tau^omega_2(A^d) is the limit, along
omega_2, of tau((s_r^* s_r)^d). Expand (s_r^* s_r)^d as a sum of 2d-letter
words in g_j and g_j^*; each contributes if and only if its reduced form
is the identity. Count the labelings of the 2d positions - for which
the word reduces to identity.
Claim: the number of such labelings is C_d (r)_d + O_d(r^{d-1}), where
(r)_d = r (r-1) ... (r-d+1), and C_d is the Catalan number binom(2d,d)/(d+1).
Reason: an identity reduction pairs all 2d positions noncrossingly, so uses
at most d distinct unsigned labels. The count with AT MOST d - 1 labels is
O_d(r^{d-1}). The count with EXACTLY d labels is the count of noncrossing
perfect matchings on 2d positions, times the number (r)_d of ways to assign
distinct labels to the d pairs. Every such matching has the property that
the two endpoints of any pair have opposite signs in the alternating
pattern, because the number of positions strictly inside a pair is even.
So tau((s_r^* s_r)^d) -> C_d, and
tau^omega_2(A^d) = C_d. (eq:catalan)
Now consider the probability measure on (0, 4) defined by
dnu(x) = (1 / (2 pi)) sqrt((4 - x) / x) dx.
Its d-th moment is computed by substituting x = 4t and using the Beta
integral:
integral_0^4 x^d dnu(x) = 4^{d+1} / (2 pi) Beta(d + 1/2, 3/2)
= binom(2d, d) / (d + 1) = C_d.
So dnu and the spectral distribution of A have matching moments. Both
are compactly supported (A in [0, 4], dnu supported in [0, 4]), so equality
of moments extends by uniform polynomial approximation on [0, 4] to
equality of the two measures. In particular:
tau^omega_2(1_{0}(A)) = nu({0}) = 0, hence 1_{0}(A) = 0.
Via polar decomposition, s = V A^{1/2} with V a partial isometry and
V^* V = 1_A(0)^perp = 1. So V is an isometry. In a finite tracial
algebra, tau^omega_2(1 - VV^*) = 1 - tau^omega_2(V^*V) = 0, so by
faithfulness VV^* = 1. So V is a UNITARY in N^omega_2.
Now transfer the block identities to powers of V. For eps > 0 define
x_eps = s (A + eps 1)^{-1/2} = V ( A / (A + eps 1) )^{1/2}.
Since A >= 0 and eps > 0, the right-hand factor is a contraction, so
||x_eps|| <= 1. For any polynomial p(t) = sum_{d=0}^D c_d t^d and any
k >= 1,
[s p(A)]^k = sum_{d_1, ..., d_k = 0}^D c_{d_1} ... c_{d_k}
[s A^{d_1}] [s A^{d_2}] ... [s A^{d_k}].
Every summand is a product covered by the block lemma of Section 17. By
that lemma and complex linearity,
S([s p(A)]^k) = a [s p(A)]^k,
S(([s p(A)]^k)^*) = b ([s p(A)]^k)^*.
Choose polynomials p_n converging uniformly on [0, 4] to (t + eps)^{-1/2}.
Functional calculus for A plus operator-norm boundedness of S pass these
identities to the norm limit:
S(x_eps^k) = a x_eps^k, S((x_eps^k)^*) = b (x_eps^k)^*. (eq:reg)
Finally pass eps -> 0. The x_eps are all contractions. Their 2-norm
distance to V is
||x_eps - V||_2^2 = integral_0^4 ( sqrt(x/(x+eps)) - 1 )^2 dnu(x),
and this goes to zero by dominated convergence and nu({0}) = 0.
Multiplying by contractions is 2-norm continuous:
||x_eps^k - V^k||_2 <= k ||x_eps - V||_2 -> 0,
||(x_eps^k)^* - (V^k)^*||_2 -> 0.
All the elements x_eps^k, V^k, x_eps^{-k}, V^{-k} are contractions
(bound 1). Bounded-ball 2-continuity of S, applied to the identity
(eq:reg) along eps = 1/n, gives
S(V^k) = a V^k,
S(V^{-k}) = b V^{-k}. (eq:pow)
Note: V^{-k} is V^{*k}, so this is the second identity applied to the
adjoint powers.
19. The spectral conclusion, and why the finite hypothesis is essential
------------------------------------------------------------------------
We now prove the last step of the rigidity theorem.
Lemma (spectral). Let (A, rho) be a finite von Neumann algebra with
faithful normal trace, V in A a unitary, S: A -> A bounded and complex
linear. If a, b in A and
S(V^k) = a V^k, S(V^{-k}) = b V^{-k} for every k >= 1,
then a = b.
The proof uses harmonic partial sums.
Define for m >= 1 and z on the unit circle T = {|z| = 1}:
H_m(z) = sum_{k=1}^m z^k / k, h_m = sum_{k=1}^m 1/k.
First, a uniform sine bound:
sup_{m >= 1} sup_{theta in R} | sum_{k=1}^m sin(k theta) / k | <= B := 1 + pi.
(eq:sine)
Proof. By periodicity and oddness it suffices to use t in [0, pi], the
angular distance from 2 pi Z. At t = 0 the sum is zero. For t > 0 choose
K = floor(1/t). The first K terms of the sum are bounded by
K t <= 1 because |sin(k t)| <= k t. For the tail from K + 1 to m, use
geometric summation: any block sum sum_{k=p}^q sin(k t) has absolute value
at most 1/sin(t/2) <= pi/t (because sin(t/2) >= t/pi for t in [0, pi]).
Abel summation with that bound compares sum_{k=K+1}^m sin(k t)/k to pi/t
times the sum of the weights 1/(k(k+1)), which telescopes to at most 1.
Total: at most 1 + pi.
Fix m and any half-open arc I subset T of angular length at most 1/m.
Let e = 1_I(V). Pick lambda in T to rotate the midpoint of I to 1. On
the rotated arc, the argument theta satisfies |theta| <= 1/m, so
Re H_m(e^{i theta}) = sum_{k=1}^m cos(k theta) / k
>= cos(1) h_m.
Write H_m(lambda V) = X_{m,lambda} + i Y_{m,lambda}, with both parts
self-adjoint. Continuous functional calculus and the sine bound give
||Y_{m,lambda}|| <= B.
By the power identities,
S( H_m(lambda V) - H_m(lambda V)^* )
= a H_m(lambda V) - b H_m(lambda V)^*
= (a - b) X_{m,lambda} + i (a + b) Y_{m,lambda}.
The argument of S (the left-hand element) has norm at most 2B. Putting
c = a - b, and using ||Y_{m,lambda}|| <= B:
||c X_{m,lambda}|| <= C := B (2 ||S|| + ||a|| + ||b||). (eq:bound)
Now use the compression. On the arc I, we have Re H_m(e^{i theta}) >=
cos(1) h_m, so X_{m,lambda} is invertible in the corner e A e. Define
F = 1_I(V) X_{m,lambda}^{-1} in W^*(V) subset e A e.
Then X_{m,lambda} F = e and ||F|| <= 1/(cos(1) h_m). Right-multiply
(eq:bound) by F:
||c e|| = ||c X_{m,lambda} F|| <= C / (cos(1) h_m) =: delta_m.
Squaring and reading this in the trace form, in the corner e A e:
e c^* c e <= delta_m^2 e.
Partition T into finitely many disjoint half-open arcs I_1, ..., I_J of
angular length at most 1/m (J depends on m). Their projections
e_j = 1_{I_j}(V) are orthogonal and sum to 1. Traciality gives
||c||_2^2 = sum_j rho(c^* c e_j)
= sum_j rho(e_j c^* c e_j)
<= delta_m^2 sum_j rho(e_j)
= delta_m^2.
As m -> infinity, h_m >= log(m + 1) -> infinity, so delta_m -> 0. So
||c||_2 = 0, and by faithfulness c = 0. This gives a = b.
The right multiplication by F and the tracial compression are legitimate
throughout: X_{m,lambda} and F commute with V (both are functions of V)
and lie in A; boundedness of S is used only through (eq:bound); the
argument works for arbitrary a, b (no commutation of a, b with V is
required).
Remark (the finite-trace hypothesis is essential). On H = l^2(Z), let V
e_n = e_{n+1} be the bilateral shift, and let p, q be the projections
p = projection onto C e_0, q = projection onto closure of span{e_n : n < 0}.
Consider the contraction S: B(H) -> B(H) defined by S(X) = p X q. Then
for every k >= 1,
S(V^k) = p V^k,
S(V^{-k}) = p V^{-k} q = 0.
(The first identity is p V^k q, which equals p V^k because p V^k = 0 on
q^perp: for n >= 0, p V^k e_n involves e_{-k} with -k < 0, outside the
range of p. The second is p V^{-k} q = 0 because V^{-k} q H lies in the
span of e_n with n < -k < 0, always outside p.) So the power identities
hold with
a = p != 0 = b,
in B(H) with the trace tr(X) = <e_0, X e_0> (not faithful on B(H)! it
dies on everything outside the range of p). The failure is exactly that
tr is not faithful on B(H): B(H) is a properly infinite algebra, and the
infinity is what lets a and b pull apart. In a II_1 factor, the trace is
faithful and finite, and the spectral-argument trick closes the gap. This
remark justifies keeping the finite hypothesis in the rigidity theorem
without further comment.
================================================================================
Part V. Bounded primitives: assembling the cohomology theorem.
================================================================================
20. Two classical inputs, and the continuity lemma
---------------------------------------------------
Two cohomological facts are imported from the literature. We state them
precisely, since they are the only external cohomological inputs used.
Proposition (normal reduction; JKR 1972, Lemma 5.4 and Theorem 5.6;
restated as (i) of the classical-reductions proposition of the paper).
Let M be a von Neumann algebra and k >= 2. For every bounded cocycle
f in C_b^k(M, M), there are a bounded separately normal cocycle f_0 and
a bounded (k - 1)-cochain b such that
f = f_0 + db.
The proof exchanges a multilinear map for its separately normal core via
averaging against the predual, which changes the map by a coboundary in
degree k >= 2. In degree one the analogous statement is the definition of
the derivation equation, so the conclusion is sharper for k >= 2.
Proposition (classical vanishing; JKR 1972 Theorem 2.4, CES 1987,
CPSS 2003, restated as (ii) of the classical-reductions proposition of the paper).
Let M be a von Neumann algebra with no type II_1 central summand. For
every k >= 2, H^k_b(M, M) = 0. In the notation above, for every bounded
cocycle f_0 of degree k >= 2, there is a bounded (k - 1)-cochain c with
dc = f_0.
The complementary case is what really matters: M has a type II_1 central
summand. On that summand we now build bounded primitives explicitly. The
assembly then glues the summand results together using central projections.
Both reductions turn continuous Hochschild cohomology into a problem
about separately normal cocycles, and separate normality is what allows
Akemann's continuity theorem to be applied input-by-input. We state and
prove that next.
Lemma (Akemann continuity; Corollary III.10 of Akemann 1967).
Let (M, tau) be a finite von Neumann algebra with a faithful normal
tracial state. Let B: M -> M be bounded and normal. For every R <
infinity and eps > 0 there is delta > 0 such that
||x|| <= R and ||x||_2 < delta => ||B(x)||_2 < eps.
In particular, B is 2-norm continuous on every operator-norm ball.
Proof. Argue by contradiction. A failure of the modulus gives R <
infinity, eta > 0, and a sequence (x_j) with
||x_j|| <= R, ||x_j||_2 -> 0, ||B(x_j)||_2 >= eta.
Choose alpha > 0 with ||B|| alpha < eta / 4. Taking the polar decomposition
x_j = v_j |x_j|, truncate at alpha:
r_j = 1_{(alpha, infinity)}(|x_j|),
y_j = v_j |x_j| r_j,
l_j = v_j r_j v_j^*,
e_j = r_j vee l_j.
Then
y_j = e_j y_j e_j, (y_j lives in the e_j-corner)
||y_j|| <= R,
||x_j - y_j|| <= alpha, (truncated the small eigenvalue part)
||B(y_j)||_2 >= 3 eta / 4. (tail dominated by ||B|| alpha)
The trace of e_j satisfies
tau(e_j) <= tau(r_j) + tau(l_j) = 2 tau(r_j) <= 2 alpha^{-2} ||x_j||_2^2.
So sum_j tau(e_j) < infinity along a subsequence. Put
p_j = vee_{h >= j} e_h. Then p_j decreases to 0: tau(p_j) <= sum_{h>=j}
tau(e_h) -> 0, and a countable decreasing family of projections with
trace tending to 0 decreases strongly to 0 in a tracial algebra.
For fixed j and h > j, the compressions
(p_j - p_h) y_j (p_j - p_h)
converge ultraweakly to y_j as h -> infinity. Applying B preserves
ultraweak convergence. The 2-norm is ultraweakly lower semicontinuous,
because
||y||_2 = sup { |tau(b^* y)| : b in M, ||b||_2 <= 1 }
and tau(b^* .) is normal. So we can choose j_1 = 1 and then inductively
j_{s+1} > j_s such that, on the blocks t_s = (p_{j_s} - p_{j_{s+1}})
y_{j_s} (p_{j_s} - p_{j_{s+1}}),
||B(t_s)||_2 >= eta / 2.
These t_s have pairwise orthogonal left and right supports, so their sum
with arbitrary signs has operator norm at most R. Take independent
uniform signs eps_s in {+1, -1}. Hilbert-space orthogonality of these
scalar random variables, applied inside L^2(M, tau), gives
N eta^2 / 4 <= sum_s ||B(t_s)||_2^2
= E ||B(sum_s eps_s t_s)||_2^2
<= ||B||^2 R^2,
which is a contradiction for N large enough. So the modulus must exist.
Applying the modulus to differences (which have operator norm at most 2R)
gives 2-norm continuity on the ball of radius R.
This concludes the classical inputs. Now we average.
21. Averaging the last input: the separable-predial primitive
--------------------------------------------------------------
Throughout this section, (M, tau) is a type II_1 von Neumann algebra with
separable predual and faithful normal tracial state; k >= 2; and f in
C_b^k(M, M) is a separately normal bounded cocycle (so df = 0). Our
goal: a bounded (k - 1)-cochain g with dg = f and ||g|| <= ||f||.
Define the cochain averaging operator for the walk. Write
mu = sum_{s >= 1} p_s delta_{u_s} as before. For a bounded k-cochain Phi,
define P Phi in C_b^k(M, M) by
(P Phi)(a_1, ..., a_{k-1}, x)
= sum_{s >= 1} p_s Phi(a_1, ..., a_{k-1}, x u_s) u_s^*. (eq:P)
Because ||Phi|| < infinity and the sum has total mass 1 with unit-modulus
right multiplier u_s^*, the sum converges in operator norm, and
||P Phi|| <= ||Phi|| (P is a contraction).
Iterating, with W_j = U_1 U_2 ... U_j and U_i independent with law mu:
(P^j Phi)(a_1, ..., a_{k-1}, x)
= E_w[Phi(a_1, ..., a_{k-1}, x W_j) W_j^*] (induction in j).
In particular P^0 Phi = Phi (since W_0 = 1). The Cesaro averages
F_n = (1/n) sum_{j=0}^{n-1} P^j f (eq:F_n)
are bounded k-cochains still satisfying ||F_n|| <= ||f||, and
P F_n - F_n = (P^n f - f)/n,
||P F_n - F_n|| <= 2 ||f|| / n. (eq:PFn)
Step 1: compactness limit. For each fixed k-tuple (a, x) in M^k, the set
{F_n(a, x)} lies in the operator-norm ball of radius ||f|| prod ||a_i||
prod ||x||, which is ultraweakly compact (Banach-Alaoglu in the predual).
The product over all tuples is compact by Tychonoff. A multilinearity
closure argument (addition and scalar multiplication are ultraweakly
continuous) shows that the set of bounded multilinear maps is closed for
pointwise ultraweak convergence. So some subnet of (F_n) converges at
every tuple to a bounded k-cochain F with
||F|| <= ||f||.
Step 2: F is P-invariant. On any fixed operator-norm ball in the space
of k-cochains, P is continuous for pointwise ultraweak convergence: given
Phi_lambda -> Phi pointwise ultraweakly with uniform bound C, fix (a, x)
and a normal functional rho. Truncating (eq:P) after L terms, the finite
sum passes to the limit; the tail is at most
2 C ||rho|| ||x|| prod_i ||a_i|| sum_{s > L} p_s,
which we can make arbitrarily small. Applying this to the subnet
F_n -> F and using (eq:PFn), which sends ||P F_n - F_n|| to 0, we get
PF = F.
Step 3: F(a, .) is 2-continuous on balls. Fix a = (a_1, ..., a_{k-1})
and put B(x) = f(a, x). Since f is separately normal, B is bounded and
normal. By the Akemann lemma of Section 20, for every R and eps there
is delta such that ||x|| <= R and ||x||_2 < delta imply ||B(x)||_2 < eps.
For any unitary v in U(M), the map x -> B(x v) v^* = f(a, x v) v^* has the
same modulus, because right multiplication by v preserves both ||.|| and
||.||_2 on M, and left multiplication by v^* is an isometry of the 2-norm
on M. Therefore the j-th iterate (P^j f)(a, x) = E_w[f(a, x W_j) W_j^*]
inherits the same modulus, and so does the Cesaro average
F_n(a, x).
The subnet F_n -> F converges pointwise ultraweakly, and the 2-norm is
ultraweakly lower semicontinuous (Section 20). So if ||x|| <= R and
||x||_2 < delta (with eps/2 in place of eps), we get
||F(a, x)||_2 <= liminf_n ||F_n(a, x)||_2 <= eps/2 < eps.
This is exactly the uniform modulus for the map T(x) = F(a, x). So T is
bounded, linear, and 2-continuous on every operator-norm ball.
Step 4: apply Liouville. The identity PF = F, restricted to the last slot
with a fixed, is exactly the harmonicity relation (eq:harm) of Part III:
T(x) = F(a, x) = (PF)(a, x) = E_U[F(a, x U) U^*]
= E_U[T(x U) U^*].
Since mu is the walk from Part II, the Liouville theorem applies to T,
giving
T(x) = T(1) x = h(a) x,
where h(a) = F(a, 1) is a bounded (k - 1)-cochain (it inherits multi-
linearity and the norm bound from F). We have thus produced a bounded
(k - 1)-cochain h of norm <= ||f|| with F(a, x) = h(a) x for all (a, x).
Step 5: the cocycle equation for h. We now average the original equation
df = 0 in the last slot to extract dh in terms of f. Fix any k-tuple
(a_1, ..., a_k) in M^k and any unitary v in U(M). The equation df = 0,
evaluated at (a_1, ..., a_k, v), reads
0 = a_1 f(a_2, ..., a_k, v)
+ sum_{j=1}^{k-1} (-1)^j f(a_1, ..., a_j a_{j+1}, ..., a_k, v)
+ (-1)^k f(a_1, ..., a_k v)
+ (-1)^{k+1} f(a_1, ..., a_k) v.
Multiply on the right by v^* (unitary, so vv^* = 1):
0 = a_1 f(a_2, ..., a_k, v) v^*
+ sum_{j=1}^{k-1} (-1)^j f(a_1, ..., a_j a_{j+1}, ..., a_k, v) v^*
+ (-1)^k f(a_1, ..., a_k v) v^*
+ (-1)^{k+1} f(a_1, ..., a_k).
Now average over v with the Cesaro measure nu_n = (1/n) sum_{j=0}^{n-1} mu^{*j}:
- The first term average is a_1 F_n(a_2, ..., a_k), because
E_{v ~ mu^{*j}}[f(a_2, ..., a_k, v) v^*] = (P^j f)(a_2, ..., a_k, 1),
and the Cesaro average yields a_1 F_n(a_2, ..., a_k, 1) = a_1 h_n(a_2,...,a_k)
where h_n = F_n(., 1) is the (k-1)-cochain underlying F_n.
- Each middle-term average is the corresponding h_n evaluated at
(a_1, ..., a_j a_{j+1}, ..., a_k), with sign (-1)^j.
- The third-term average is F_n(a_1, ..., a_k). Indeed the map
v -> f(a_1, ..., a_{k-1}, a_k v) v^* is exactly (P f)(a_1, ..., a_k)
when v distributes as mu; averaging over nu_n gives (1/n) sum_j (P^j f) = F_n.
- The last term is v-independent.
So for every n:
0 = a_1 h_n(a_2, ..., a_k)
+ sum_{j=1}^{k-1} (-1)^j h_n(a_1, ..., a_j a_{j+1}, ..., a_k)
+ (-1)^k F_n(a_1, ..., a_k)
+ (-1)^{k+1} f(a_1, ..., a_k). (eq:cesaro)
Pass to the chosen subnet (of Step 1), which converges pointwise
ultraweakly to F. Multiplying a fixed element on the left preserves
ultraweak convergence, so each term converges. In the limit, using
F(a, x) = h(a) x with the last argument x = a_k:
0 = a_1 h(a_2, ..., a_k)
+ sum_{j=1}^{k-1} (-1)^j h(a_1, ..., a_j a_{j+1}, ..., a_k)
+ (-1)^k h(a_1, ..., a_{k-1}) a_k
+ (-1)^{k+1} f(a_1, ..., a_k).
The first four displayed terms together are exactly the value of the
Hochschild differential dh on (a_1, ..., a_k). Thus
(dh)(a_1, ..., a_k) = (-1)^k f(a_1, ..., a_k).
So set g = (-1)^k h (the sign adjustment depends on k mod 2, but is a
global sign that does not affect the bound). Then dg = f, and
||g|| = ||h|| <= ||F|| <= ||f||. ()
This is the paper's separable-predial primitive (Proposition 5.2 of the paper).
Two remarks on the argument.
(1) The subnet limit F is NOT claimed to be the original f; only its
combination with dh recovers f. The passage from "F(a, x) = h(a) x" to
"dh = +-f" uses the Cesaro averages to convert the cocycle equation into
the coboundary equation; no commutativity of P and d is required.
(2) All of the argument is uniform in k. The bound ||g|| <= ||f|| is what
makes the next two sections work: separability removal and central
decomposition are both driven by uniformity of the bound.
22. Removing the separability hypothesis
-----------------------------------------
Section 21 assumed separable predual. That was only used to invoke
the walk/Ultrapower construction at the two ultrapower steps of Part III.
We now show that none of the essential conclusions depend on separability:
the uniform bound ||g|| <= ||f|| supplies the compactness needed to lift
the primitive to the general case.
Throughout this section, (M, tau) is any type II_1 von Neumann algebra
with a faithful normal tracial state; separability is not assumed.
Fix k >= 2 and a bounded separately normal cocycle f in C_b^k(M, M).
Order the finite subsets E subset M by inclusion, so that (E)_E is a
directed poset. For each E, choose a countably generated unital von
Neumann subalgebra N_E subset M with the properties:
(c1) N_E contains E;
(c2) N_E contains f(E^k) (pointwise image of E^k under f);
(c3) N_E contains unital copies of M_m(C) for every m >= 2
(i.e. matrix systems of every size).
Such N_E exists because a type II_1 algebra admits unital copies of
M_m(C) for every m: split the identity into m mutually equivalent
orthogonal projections, and let the partial isometries between them be the
matrix units. Adjoin one system of every size, plus countably many
generators, to see that N_E is countably generated.
Claim: N_E itself is a type II_1 von Neumann algebra with separable
predual. Countably many generators give separable L^2(N_E) by
approximating *-polynomials with rational complex coefficients in norm and
then extending to the 2-norm. Moreover N_E has no type I central
summand: a type I_n summand would inherit unital copies of M_m(C) only for
m <= n, contradicting (c3). Every finite type I algebra is a direct sum
of homogeneous type I_n fibers, so no type I central summand means N_E is
type II_1.
Let E_E: M -> N_E be the (unique) normal trace-preserving conditional
expectation. It exists because N_E is a subalgebra of M with a faithful
normal trace, and E_E is M-bimodular: E_E(a x b) = a E_E(x) b for all
a, b in N_E and x in M. (This is Dixmier 1953, Thm 8, and Umegaki 1954,
Thm 1.)
Define f_E: N_E^k -> N_E by
f_E(x_1, ..., x_k) = E_E( f(x_1, ..., x_k) ) for x_j in N_E.
Then f_E is a bounded separately normal k-cochain on N_E (separate
normality inherited from f and E_E; the bound by ||f_E|| <= ||f||).
Moreover f_E is a cocycle on N_E: because E_E is N_E-bimodular and
f(x_1, ..., x_k) in N_E by (c2), the outer bimodule factors pass through
E_E, and the inner terms are in N_E, so
d f_E = E_E (df restricted to N_E^{k+1}) = 0.
Apply the result of Section 21 (which requires only separable predual, which N_E
has). We get a bounded (k-1)-cochain g_E in C_b^{k-1}(N_E, N_E) with
d g_E = f_E, ||g_E|| <= ||f_E|| <= ||f||.
Extend each g_E to a cochain G_E on all of M by precomposing with E_E:
G_E(x_1, ..., x_{k-1}) = g_E( E_E x_1, ..., E_E x_{k-1} ).
The norm bound is inherited: ||G_E|| = ||g_E|| <= ||f||. If
(a_1, ..., a_k) is a k-tuple in E^k, then each a_j and each adjacent
product a_j a_{j+1} lies in N_E, and f(a_1, ..., a_k) lies in N_E by
(c2). The Hochschild differential of G_E at (a_1, ..., a_k) therefore
equals dg_E at (E_E a_1, ..., E_E a_k) = dg_E(a_1, ..., a_k), which is
f_E(a_1, ..., a_k) = E_E f(a_1, ..., a_k) = f(a_1, ..., a_k). So
(d G_E)(a_1, ..., a_k) = f(a_1, ..., a_k) for every
(a_1, ..., a_k) in E^k. (eq:G_E_exact)
The common norm bound ||G_E|| <= ||f|| and the ultraweak compactness of
operator-norm balls (as in Section 21 Step 1) give a pointwise ultraweakly
convergent subnet of (G_E), with limit
g in C_b^{k-1}(M, M), ||g|| <= ||f||.
For any fixed k-tuple (a_1, ..., a_k), the finite set E_0 = {a_1, ..., a_k}
is contained in every E above it in the directed poset. So along the
subnet, the coboundary equation (eq:G_E_exact) holds eventually. Passing
to the limit in its finitely many terms gives
(dg)(a_1, ..., a_k) = f(a_1, ..., a_k).
This holds at every tuple, so dg = f. This is the paper's tracial
primitive (Proposition 5.3 of the paper).
23. Central decomposition and the assembly
-------------------------------------------
We now remove the trace assumption of Section 22. The idea is to split
M into central pieces indexed by a partition of the identity, each of
which DOES admit a faithful normal trace, apply the earlier primitive
piece by piece, and use a correction homotopy to fix mixed inputs across
pieces.
Setup. Let M be a complex von Neumann algebra (any), and z_{II_1} the
central projection of the type II_1 summand. Write z_0 = 1 - z_{II_1}.
Step A: split the type II_1 summand into tracial pieces. In the abelian
algebra Z(z_{II_1} M) of central projections, choose a maximal orthogonal
family (z_i)_{i in I} of nonzero projections supporting normal states.
The join of the family equals z_{II_1}: an orthogonal complement would
support another normal state, contradicting maximality. For each i, a
state on z_i Z(M) is faithful on z_i Z(M); composing it with the faithful
normal center-valued trace of z_i M yields a faithful normal tracial state
tau_i on z_i M. (Blackadar 2006, Theorem III.2.5.7 and Corollary III.2.5.8.)
The family I may be uncountable; this is harmless.
Together with the complementary summand z_0 M (which has no type II_1
central summand), we have a central partition of unity:
1 = z_0 + sum_{i in I} z_i.
Step B: piecewise primitives. For each i in I, restrict f to z_i M as
f_i(x_1, ..., x_k) = z_i f(x_1, ..., x_k) for x_j in z_i M.
Since multiplication by a central projection is a bimodule map, f_i is a
bounded separately normal k-cocycle on z_i M, and ||f_i|| <= ||f||.
Because z_i M has a faithful normal trace tau_i, the result of Section 22
gives b_i in C_b^{k-1}(z_i M, z_i M) with
d b_i = f_i, ||b_i|| <= ||f||.
For the complementary summand z_0 M, the classical vanishing theorem of
Section 20 gives b_0 in C_b^{k-1}(z_0 M, z_0 M) with d b_0 = f_0
(the restriction of f to z_0 M).
Step C: pull back to M. Define the pulled-back cochains
\tilde b_i (for i in I union {0}) by
\tilde b_i(x_1, ..., x_{k-1}) = b_i( z_i x_1, ..., z_i x_{k-1} ).
Then \tilde b_i has the same norm bound as b_i. Applying d,
d \tilde b_i (a_1, ..., a_k) = z_i f(z_i a_1, ..., z_i a_k)
=: cut_{z_i}(z_i f)(a_1, ..., a_k).
The right-hand side is the "central cut" of z_i f: only the z_i-parts of
the inputs survive. It is generally NOT z_i f itself, because the inputs
a_j may have nontrivial q_i := 1 - z_i parts. This is the source of
"mixed input" corrections.
Step D: correct mixed inputs by a homotopy. For a central projection z in
M with q = 1 - z, define the homotopy J_z on n-cochains psi in
C_b^n(M, zM) (for n >= 2) by
(J_z psi)(a_1, ..., a_{n-1})
= sum_{i=1}^{n-1} (-1)^i
psi( z a_1, ..., z a_{i-1}, q, q a_i, a_{i+1}, ..., a_{n-1} ).
Set J_z = 0 in degrees 0 and 1.
Lemma (central homotopy).
For every n and every psi in C_b^n(M, zM):
(d J_z + J_z d) psi = psi - cut_z psi, (eq:homotopy)
||J_z psi|| <= (n - 1) ||psi||. (eq:homotopy-norm)
Here cut_z psi(a_1, ..., a_n) = psi(z a_1, ..., z a_n).
Proof. The norm estimate is direct from the alternating sum and
||q|| = 1. For the identity, expand each a_j = z a_j + q a_j by
multilinearity: it suffices to check on tuples whose entries each lie in
zM or in qM.
Given such a tuple (a_1, ..., a_n), let r be the index of the first qM
entry (if any). Then cut_z psi vanishes on this tuple (some input is not
in zM, so the z-cut annihilates it). On the other side, (J_z psi) applied
to the n-tuple inserts one q between consecutive entries:
- d(psi) merges adjacent positions. The merged positions include the
inserted q and the next entry. The inserted q is in qM at position r
in the (n-th slot) psi call (indices: the psi sees the tuple with
q inserted between slots); a positive-indexed psi cancellation between
slots r-1 and r would set q * (zM entry) = 0; similarly for the
zM-q product. The only surviving merger is q * q a_r = q^2 a_r = q a_r,
which re-creates q at position r, matching psi(a_1, ..., a_n) after
the differential signs alternate to a total (+ 1).
- The remaining terms (J_z applied to d psi) insert q at one of the
n-1 possible positions of an (n+1)-slot differential; by direct
accounting, every such term either duplicates a matching term in
d J_z psi (with opposite sign, cancelling) or is zero because the
merged product z M * q M = 0.
- In the all-zM case, J_z psi = 0 by definition (all inserted q's are
multiplied by zM entries), and cut_z psi = psi. Both sides vanish.
A careful coordinatewise check (the alternating-sum bookkeeping is
standard but tedious) confirms that this covers all differential terms.
The result (eq:homotopy) is what remains. ()
Remarks on (eq:homotopy). (1) If psi is a cocycle (d psi = 0) and
psi(a) takes values in zM, then
d(J_z psi) = psi - cut_z psi, ||J_z psi|| <= (n-1)||psi||.
So J_z psi is a bounded primitive for the "correction" psi - cut_z psi,
with a norm bounded by (n-1)||psi||, independent of the projection z and
of the algebra M. This is exactly the "piece-level uniform bound" we
need. (2) The construction is symmetric in the sense that if psi is
already a "z-cut" cocycle, J_z psi = 0: if psi(a) = psi(z a), the
inserted q annihilates, and cut_z psi = psi. So no new error is
introduced where none is needed. (3) The sign convention is chosen so
that (eq:homotopy) reads with a + on the right; another convention would
flip the sign of the insertion terms and the same identity would still
hold.
Step E: assemble. For each piece i in I union {0}, form
H_i = \tilde b_i + J_{z_i}(z_i f). (eq:Hi)
By the previous step applied to the cocycle z_i f (which has values in
z_i M and satisfies d(z_i f) = z_i d f = 0),
d H_i = cut_{z_i}(z_i f) + (z_i f - cut_{z_i}(z_i f))
= z_i f.
For the norm:
||H_i|| <= ||b_i|| + (k-1) ||z_i f|| <= ||f|| + (k-1) ||f||
<= k ||f|| for i in I,
and for the piece i = 0, with finitely many ||b_0|| = M_0,
||H_0|| <= M_0 + (k-1) ||f||.
So there is a single constant C < infinity (the larger of the two
displayed bounds, or just M_0 + k||f||) with
||H_i|| <= C for every piece i.
Step F: assemble into M. For any central partition of unity, the map
M -> prod_i l^infty z_i M, x |-> (z_i x)_i
is an isometric *-isomorphism (in particular, onto the l^\infty-direct
product). Surjectivity: given a bounded family (y_i) in prod_i z_i M
with sup_i ||y_i|| < infinity, the finite partial sums sum_{i in F} y_i
over finite F subset I form a Cauchy net directed by inclusion. Because
the z_i are orthogonal central projections, the net converges strongly
to an element y in M with pi_i(y) = y_i.
For any (k-1)-tuple (x_1, ..., x_{k-1}) in M, the family of values
H_i(x_1, ..., x_{k-1}) is bounded by C prod_j ||x_j||. It therefore
defines an element G(x_1, ..., x_{k-1}) of M via the isomorphism. By
coordinatewise multilinearity and bound, G is a bounded multilinear
(k-1)-cochain with ||G|| <= C. Because central projection commutes with
every term of the differential,
z_i (d G) = d H_i = z_i f for every piece i.
As the z_i have join 1, this forces d G = f.
Step G: restore the original cocycle. Recall from JKR reduction
(Section 20) that the original bounded cocycle f_orig of degree k >= 2
is f_orig = f + db for some separately normal bounded cocycle f and
bounded (k-1)-cochain b. We have just produced G with dG = f. Then
d(G + b) = f + db = f_orig.
Both G and b are bounded (||G|| <= C depends on ||f|| only; ||b|| is
finite by the JKR reduction), so G + b is a bounded (k-1)-cochain with
coboundary f_orig. Hence f_orig is a bounded coboundary.
This completes the proof of the main theorem.
Part VI. Wrap-up: what the theorem says, what it does not say, and what to read next.
================================================================================
24. The theorem, restated once more
------------------------------------
We have proved:
Theorem (Kadison-Ringrose cohomology conjecture, positive answer).
Let M be a complex von Neumann algebra and let k >= 2 be an integer.
Every bounded k-cocycle f in C_b^k(M, M) is a bounded coboundary. That
is, there is a bounded (k-1)-cochain g in C_b^{k-1}(M, M) with dg = f;
and the primitive can be taken to satisfy ||g|| <= ||f||.
Together with the classical Kadison-Sakai degree-one inner-derivation
theorem, H^k_b(M, M) = 0 for every k >= 1 and every complex von Neumann
algebra M.
The word "bounded" throughout refers to the ordinary operator norm on
multilinear maps C_b^k(M, M). No complete-boundedness condition is
imposed; no structural hypothesis (type I, hyperfinite absorption, Cartan
subalgebra, property Gamma, tensor product factorisation, separability,
or any other) is needed. The primitive is an actual bounded cochain, not
a class modulo closure or normal reduction.
25. What each tool did, in one sentence each
--------------------------------------------
- Popa's ultrapower freeness (Part II) built a fixed countably supported
symmetric measure mu on U(M) whose long products have free Haar scalar
moments along a fixed subsequence. This delivered the only freeness
that the whole proof uses.
- The distinguished-letters combinatorial lemma (Part II, Section 10)
turned that abstract freeness into a concrete finite-path test: one
distinct fresh label per path suffices to kill the trace of any word.
- The martingale argument (Part III, Section 13) turned any bounded
linear 2-continuous map T harmonic for mu into a martingale in
L^2(Omega; L^2(M, tau)), whose limit Z_infty is what must be shown to be
deterministic.
- The forward/reverse endpoint inequality (Part III, Section 14) took
two distinct support points a, b of Z_infty and, by using DISJOINT
increments for the two halves, produced a joint probability of hitting
both simultaneously. Without this, the rigidity theorem would have no
inputs a != b to contradict.
- Haagerup's free-group length estimate (Part IV, Section 16) bounded
||s_r|| by 2, uniformly in r, for the averaged free generator s_r; this
bound is what makes the polar argument stable.
- The block coefficient estimate (Part IV, Section 17) turned
"polynomials in s_r" into a sum of words with distinct unsigned indices
plus a norm-small error. The size 3^L on non-crossing matchings, and
the r^{-q/2} coefficient decay, are the arithmetic heart.
- The Catalan moment computation and free Poisson law (Part IV, Section
18) showed that s^* s has no kernel, so the polar part V of s is a
genuine unitary rather than a proper partial isometry.
- The harmonic series H_m and the sine-sum bound (Part IV, Section 19)
proved the spectral lemma: a bounded S with S(V^k) = a V^k and
S(V^{-k}) = b V^{-k} forces a = b, using - crucially - the faithfulness
of the finite trace.
- Akemann's continuity theorem (Part V, Section 20) supplied the
bounded-ball 2-continuity of B(x) = f(a, x) that Part III requires, and
the random-signs argument that proves it is the same technique used in
many earlier averaging arguments.
- JKR normal reduction (Part V, Section 20) let us work with separately
normal cocycles without loss of generality, which is what Akemann
requires as input.
- The averaging operator P and its Cesaro subnet limit (Part V, Section
21) turned a bounded cocycle f into a harmonic map T = F(a, .) whose
Liouville value h(a) = T(1) is a candidate primitive.
- The Cesaro averaging of the cocycle equation (Part V, Section 21,
Step 5) converted the harmonicity of T into the coboundary relation
dh = +-f, giving an explicit primitive with the operator-norm bound
||g|| <= ||f||.
- Conditional expectations onto countably generated subalgebras (Part V,
Section 22) removed the separability hypothesis from the existence of
a bounded primitive, at the cost of adding one step of compactness.
- The central homotopy J_z (Part V, Section 23) fixed the "mixed inputs"
correction when the cocycle is pulled back from central pieces of M;
without it, a partition into tracial pieces would produce primitives
only for cut cocycles, not for the cocycle itself.
- Uniform compactness of operator-norm balls, applied in Sections 21, 22,
23, is what makes all limits "actual bounded cochains" rather than
"cochains that exist in the quotient".
26. What this text does NOT claim
----------------------------------
To be honest about the boundaries of the result:
(1) Only self-coefficients are treated. We have H^k_b(M, M) = 0. The
analogous vanishing question for H^k_b(M, N) with a general M-bimodule N
is not addressed by this argument. The averaging operator P uses the
bimodule structure on the last slot; for general coefficients, that
structure may not be compatible with the walk construction.
(2) Only boundedness, not complete boundedness, is proven. The theorem
asserts vanishing in the bounded-category. A completely bounded
refinement would require extra estimates on the primitive (in
particular, on completely bounded norms of g) that the argument does
not control.
(3) No quantitative estimate on primitives other than the bound
||g|| <= ||f|| is proven. In principle the constructed primitive is
unique modulo bounded coboundaries of degree k-1; no canonical
choice is claimed.
(4) No claim is made about infinite multiplicities in the central
decomposition. The result holds for uncountable families of central
pieces, but the "uniform bound C" for the piecewise primitives H_i
is only infinite-algebraically finite; no quantitative smallness for
C in terms of k is asserted.
(5) The theorem is a cohomological vanishing statement, not a solution
to operator-theoretic stability. Whether every bounded cocycle can
be "uniformly approximated" by nice (say normal or even finite-rank)
cocycles is a separate question that the proof does not address.
27. One-sentence outline of the argument (for a first pass)
-----------------------------------------------------------
If you only want to remember the spine: build a random walk on unitaries
whose long products are free Haar; use its Furstenberg martingale to show
that any harmonic T is a left multiplier; average the cocycle along the
walk in the last input to reach a harmonic map; by Liouville that map is a
right module map; the averaged cocycle equation then produces a bounded
primitive; separability is removed by conditional expectations on
countably generated subalgebras; and central decomposition partitions M
into tracial pieces whose primitives glue via a small homotopy.
28. Suggested reading order, if you want to see the paper itself
-----------------------------------------------------------------
The paper this text accompanies is:
Vanishing of higher bounded Hochschild cohomology, OpenAI, 2026.
Its structure mirrors this exposition:
Section 1 = this text Sections 1-7 (background and main theorem).
Section 2 = this text Sections 8-11 (the walk).
Section 3 = this text Sections 12-14 (Liouville theorem).
Section 4 = this text Sections 15-19 (rigidity).
Section 5 = this text Sections 20-23 (the cohomological assembly).
If you can only read one theorem, read rigidity (Theorem 4.1 of the paper,
and the rigidity theorem in Section 15 of this text). It is the only
place where the finite trace
of a von Neumann algebra is used in the structural way; everything else
is combinatorial or probabilistic or linear-algebraic. The spectral
argument in Section 19 is the piece of mathematics you are most likely to
remember a year later.
29. Small glossary for the reader
----------------------------------
2-norm: ||x||_2 = tau(x*x)^{1/2}, the Hilbert-space norm of the
tracial representation.
Akemann's theorem: bounded normal linear maps M -> M are 2-norm
continuous on operator-norm balls.
bounded-ball 2-continuity: R: N -> N is such that ||x_j|| <= R and
||x_j||_2 -> 0 implies ||R(x_j)||_2 -> 0.
Catalan number: C_d = binom(2d, d)/(d+1); counts noncrossing perfect
matchings on 2d points and indexes the free Poisson moments.
cocycle: bounded k-cochain f with df = 0.
coboundary: bounded k-cochain f of the form f = dg with g a bounded
(k-1)-cochain.
free Haar family: unitaries with respect to which reduced words are
orthonormal in the tracial 2-norm; equivalently tau(w) = 0 for every
nonidentity reduced word w.
Haagerup length estimate: ||sum_{|w|=q} c_w lambda(w)||
<= (q+1) (sum |c_w|^2)^{1/2} in l^2(F_r).
JKR normal reduction: any bounded cocycle differs from a separately
normal bounded cocycle by a bounded coboundary.
Marchenko-Pastur (quarter-circle-ish): dnu(x) = (1/(2pi))
sqrt((4-x)/x) dx on (0, 4); limit spectral distribution of s^* s for
s = r^{-1/2} sum_{j=1}^r g_j with free Haar g_j.
n-th convolution power mu^{*n}: the law of the sum (or in this operator
setting, the product) of n independent increments of law mu.
Popa's ultraproduct freeness: a unital separable-predial subalgebra of
a II_1 ultrapower is free from a diffuse abelian subalgebra, hence
from a Haar unitary; iterate to get d free Haar unitaries.
rigidity (part IV): two left multipliers attached to positive and
negative powers of a unitary that arise from free Haar sums must be
equal, provided the containing algebra has a finite faithful trace.
symmetric measure: the law of U^{-1} equals the law of U.
30. Acknowledgments
--------------------
This exposition follows the paper "Vanishing of higher bounded Hochschild
cohomology" by OpenAI (September 2026). The classical inputs are due to
Kadison and Ringrose, to Johnson, Kadison, and Ringrose, to Akemann, and
to Haagerup, and the freeness input is due to Popa. The random-walk
construction and the rigidity theorem are new in the cited paper; the
present text is a self-contained re-derivation for readers at an
early-undergraduate level, with some lemmas proved in more detail than
in the paper and others cited with exact location.
End of text.
================================================================================