Professional Documents
Culture Documents
Free Probability
Moritz Weber, Editors
and Operator
areas (quantum mechanics, free group algebras, random matrices
etc). Thirty years after its foundation, it is a well-established and very
active field of mathematics. Originating from Voiculescu’s attempt
to solve the free group factor problem in operator algebras, free
probability has important connections with random matrix theory,
combinatorics, harmonic analysis, representation theory of large
groups, and wireless communication.
Algebras
These lecture notes arose from a masterclass in Münster, Germany
and present the state of free probability from an operator algebraic
Dan-Virgil Voiculescu
Nicolai Stammeier
matrices and combinatorics of free probability (Speicher), free
monotone transport (Shlyakhtenko), free group factors (Dykema),
free convolution (Bercovici), easy quantum groups (Weber), and a
historical review with an outlook (Voiculescu). In order to make it
more accessible, the exposition features a chapter on basics in free
Moritz Weber
probability, and exercises for each part.
ISBN 978-3-03719-165-1
www.ems-ph.org
Voiculescu / Font: NewsGothic / Pantone: 287 / Pantone: 116 / Format: 170 x 240 / RB: 7.2 mm
Münster Lectures in Mathematics
Edited by
Christopher Deninger (c.deninger@uni-muenster.de) and
Linus Kramer (linus.kramer@uni-muenster.de), Universität Münster, Germany
Dan-Virgil Voiculescu
Nicolai Stammeier
Moritz Weber
Editors
Editors:
2010 Mathematics Subject Classification: Primary 46L54; secondary 60B20, 47C15, 20G42
Key words: Free probability, operator algebras, random matrices, free monotone transport, free group
factors, free convolution, compact quantum groups, easy quantum groups, noncrossing partitions, free
independence, entropy, max-stable laws, exchangeability
ISBN 978-3-03719-165-1
The Swiss National Library lists this publication in The Swiss Book, the Swiss national bibliography, and the
detailed bibliographic data are available on the Internet at http://www.helveticat.ch.
This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
reproduction on microfilms or in other ways, and storage in data banks. For any kind of use permission
of the copyright owner must be obtained.
Contact address:
European Mathematical Society Publishing House
Seminar for Applied Mathematics
ETH-Zentrum SEW A27
CH-8092 Zürich
Switzerland
Typeset using the authors’ TEX files: Nicolai Stammeier (Oslo), Moritz Weber (Saarbrücken)
Printing and binding: Beltz Bad Langensalza GmbH, Bad Langensalza, Germany
∞ Printed on acid free paper
987654321
“Small observations can lead to big discoveries.”
(Dan-V. Voiculescu)
Contents
Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
Contributors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Preface
The present lecture notes arose from the masterclass “Free Probability and
Operator Algebras” held September 2–6, 2013 in Münster, Germany. We would
like to express our deep gratitude to the lecturers
• Dan-V. Voiculescu for providing a comprehensive account on early days of
free probability and hints where this theory may lead in the future,
• Roland Speicher for explaining intriguing connections of free probability to
the theory of random matrices and combinatorics,
• Dima Shlyakhtenko for presenting the state of the art theory concerning free
monotone transport,
• Ken Dykema for telling us the fascinating, yet unfinished story of free group
factors,
• Hari Bercovici for discussing free convolution, the free way of dealing with
sums and products of independent random variables, and finally
• Moritz Weber for introducing us to easy quantum groups and explaining to
us why we might want to care about them.
The lectures were attended by roughly 50 participants from various coun-
tries whose seniority ranged from master student to full professor, but most
participants were PhD students and young postdoctoral researchers. Taken
into account the impressions we got during the week as well as the feedback
we collected afterwards, we feel that this event has been very successful in
stimulating sustainable interactions between distinguished experts in the field
and young emerging researchers.
As the lecturers conveyed their themes with great enthusiasm that struck
the audience, a lot of work was carried out behind the curtains, both during
and before this event: Since the idea of having such a masterclass in Münster
came to life on a pleasant evening spent in the leisure room of the MFO (Ober-
wolfach) in October 2012, the organizers received advice and support from their
mentors Joachim Cuntz and Roland Speicher.
Many small and not so small things were taken care of by our phenomenal
team of secretaries: Elke Ernsting, Gabriele Dierkes and Lisa Steggemann.
Without their well-structured, competent work and their patience, this event
would have been impossible.
In order to ease the preparation of these lecture notes for the lecturers, the
participants Cédric Schonard and Jonas Wahl took notes and typeset a useful
first draft for each lecturer. We would like to thank both of them for their
contribution as well as Siegfried Echterhoff for the financial support granted to
undertake this step. In addition, we would like to thank Linus Kramer (WWU)
for all his efforts that led to the creation of this new lecture notes series within
the framework of the EMS Publishing House, Karin Halupczok (WWU) and
Simon Winter (Dimler & Albroscheit) for their valuable editing, and Thomas
Hintermann (EMS Publishing House) for his efficient and competent handling.
x Preface
Finally and quite importantly, we wish to thank the SFB Groups, Geometry
and Actions at the Mathematics Department in Münster for hosting the event
and providing us with generous support. This enabled us to invite renowned
specialists to Münster as well as to offer support for young talented researchers
from distant places that otherwise would not have had access to sufficient
funding in order to attend.
We are convinced that this masterclass added to the outstanding reputation
of the Mathematics Department at the University of Münster. On the other
hand, it also served the mathematical community as a whole by stimulating
scientific interaction and spreading knowledge. With this perspective in mind,
the creation of a lecture notes for this masterclass is nothing but the canonical
next step. The result of our efforts is right in front of you, and we hope that
it will prove itself an enjoyable and valuable source.
It is supposed to serve as an introduction into free probability from an
operator algebraic point of view as well as a reference book for this approach.
This is why we also inserted a lecture on basics of free probability which was
not part of the original masterclass lectures.
Again, we thank all speakers not only for giving the lectures in Münster but
also for all their efforts to improve these lecture notes and all their detailed
proof-reading. Finally, we thank Dan-V. Voiculescu for co-editing these lecture
notes together with us.
June 2015
Background and outlook
Dan-Virgil Voiculescu
While working on this paper the author was supported in part by NSF Grant
DMS 1001881.
2 D.-V. Voiculescu
on R such that Z
ϕ(P (a)) = P (t)dµa (t)
for all polynomials P . Thus, in the case of one hermitian random variable, we
get a probability measure, like in classical probability theory. It is completely
determined by the collection of the moments.
What distinguishes free probability from other noncommutative probability
theories is the definition of independence, which is different from the one used
in quantum mechanics (and in classical probability).
Indeed, in quantum mechanics, independence is modeled on tensor products
and we shall refer to it as classical independence. Two subalgebras 1 ∈ B, 1 ∈ C
in (A, ϕ) are classically independent if they commute (i.e. [B, C] = 0) and if
ϕ(bc) = ϕ(b)ϕ(c)
holds for all b ∈ B and c ∈ C. Note that in classical probability, this amounts
to the fact that independent variables factorize under the expectation.
In free probability we have free independence. A family of subalgebras
1 ∈ Ai , i ∈ I in (A, ϕ) is freely independent (or free) if
ϕ(a1 a2 . . . ak ) = 0
whenever aj ∈ Ai(j) , 1 ≤ j ≤ k are such that i(j) 6= i(j + 1), 1 ≤ j < k and all
ϕ(aj ) = 0, 1 ≤ j ≤ k. Sets of variables in (A, ϕ) are free, by definition, if the
algebras they generate are free.
Recently, I developed a version of freeness with left and right variables,
which I called bi-freeness. Since this extension of free probability is at a very
early stage, I took a look, to get some inspiration, at my old first free probability
paper which I presented at a conference in Buşteni.1 It was a well-attended
international conference in a mountain resort in Romania, which took place
August 29–September 9, 1983, that is almost exactly 30 years ago, from the
days of our master class in Münster.
Just before starting in this new direction, I had worked with Mihai Pimsner,
computing the K-theory of the reduced C ∗ -algebras of free groups. From the
K-theory work I had acquired a taste for operator algebras associated with free
groups and I became interested in a famous problem about the von Neumann
algebras L(Fn ) generated by the left regular representations of free groups,
which appears in Kadison’s Baton–Rouge problem list. The problem, which
may have already been known to Murray and von Neumann, is:
Are L(Fm ) and L(Fn ) nonisomorphic if m 6= n?
This is still an open problem. Fortunately, after trying in vain to solve it, I
realized it was time to be more humble and to ask: is there anything I can do,
which may be useful in connection with this problem? Since I had come across
computations of norms and spectra of certain convolution operators on free
1D. Voiculescu, Symmetries of some reduced free product C ∗ -algebras, in Operator al-
gebras and their connections with topology and ergodic theory (Buşteni, 1983), 556–588,
Lecture Notes in Math., 1132, Springer, Berlin, 1985. MR0799593
Background and outlook 3
groups (i.e., elements of L(Fn )), I thought of finding ways to streamline some
of these computations and perhaps to be able to compute more complicated
examples. This, of course, meant computing expectations of powers of such
operators with respect to the von Neumann tracial state τ (T ) = hT ee , ee i,
eg being the canonical basis of the ℓ2 space ℓ2 (Fn ), and e ∈ Fn the neutral
element.
The key remark I made was that if T1 , T2 are convolution operators on
Fm and Fn , then the operator on Fm+n = Fm ∗ Fn which is T1 + T2 has
moments τ ((T1 + T2 )p ) which depend only on the moments τ (Tjk ), j = 1, 2,
but not on the actual T1 and T2 . This was like the addition of independent
random variables, only classical independence had to be replaced by a notion
of free independence, which led to a free central limit theorem, a free analog of
the Gaussian functor, free convolution, an abstract existence theorem for one
variable free cumulants, etc.
An important consequence of the central limit theorem was that I realized
that the analog of the Gauss law in free probability was the semicircle law. By
Wigner’s work this law is known to play a key role in random matrix theory
as a limit of eigenvalue distributions. After wondering for a few years about
this coincidence, I understood what the connection was: large independent
Gaussian random matrices give rise to freely independent random variables
asymptotically. On a suitable algebra F (X , µ, Mn ) of n × n matrix-valued
functions T : X → Mn , the probability measure µ on X gives rise to an
expectation functional
Z
ϕ(T ) = n−1 Tr(T (ω))dµ(ω).
X
∗
When T = T , its distribution is easily seen to be the average of the probability
measures giving mass 1/n to the eigenvalues of T (ω).
A consequence of the asymptotic freeness of independent large Gaussian
matrices with i.i.d. entries was that the von Neumann algebra L(Fn ) could
be viewed as being asymptotically generated by an n-tuple of such random
matrices. With this asymptotic random matrix model I could prove, for in-
stance, the isomorphism of L(F∞ ) and P L(F∞ )P where P 6= 0 is a projection
with τ (P ) ∈ Q. Florin Radulescu was then able to remove the restriction
that τ (P ) be rational, and Ken Dykema and Florin Radulescu then discovered
the interpolated free group factors L(Fr ) where r > 1 does not need to be
an integer. These enjoy many of the properties of the L(Fn ), for instance,
L(Fr ) ∗ L(Fs ) ∼ L(Fr+s ).
In the meantime, Roland Speicher developed a theory of free stochastic in-
tegration on the Cuntz algebra and began developing the combinatorial side
of free probability. He discovered that, at the combinatorial level, the pas-
sage from classical probability to free probability meant replacing the lattice
of all partitions of {1, . . . , n} by the lattice of noncrossing partitions. This
was precisely how the combinatorial formulae for classical cumulants turned
into combinatorial formulae for free cumulants based on noncrossing partitions.
4 D.-V. Voiculescu
The combinatorial development was joined by Alexandru Nica with his many
essential contributions to the subject. On the other hand, the first classi-
cal probabilist to join in the effort was Philippe Biane with a wide range of
contributions from advanced stochastics to free probability aspects of asymp-
totics of group representations and processes with free increments. At present
the combinatorial side has also reached beyond noncrossing partitions to more
complicated diagrammatics like in the developing connections with subfactor
theory and in the study of second-order freeness.
The early successes of free probability were marked by several other deep
results.
I should mention here the deep analytic theorem of Hari Bercovici and
Vittorino Pata about the correspondence between classical and free infinitely
divisible laws in which the domains of attractions of corresponding laws are
equal.
In another direction, rather recently, a sweeping generalization of the almost
sure results about largest eigenvalues from one to several Gaussian random
matrices was obtained by Uffe Haarerup and Steen Thorbjørnsen which gave
a demonstration that free probability could make important analytically hard
contributions to random matrix theory.
By the end of the 80s and beginning of the 90s a parallel between classical
and free probability had emerged and it is natural to ask:
How far does the parallel between classical and free probability extend?
At present, we know that the answer is: very far and it is one of the most
surprising things about free probability.
Let me give a few examples of items on the list of classical probability items
with free probability analogs:
• limit laws,
• stochastic processes with independent increments,
• addition and multiplication of independent random variables,
• stochastic integration,
• combinatorics of cumulants,
• continuous entropy,
• max-stable laws,
• exchangeability.
There is an amazing parallelism in all this. However, when one takes a closer
look there are often serious differences and open questions, which one would
like to understand. I would like to illustrate this with some comments about
the last three entries of the list.
Entropy. A few minutes are certainly not enough time for a lecture on free
entropy. Looking at the parallelism, this should be a quantity χ(X1 , . . . , Xn ),
where Xj = Xj∗ are in a von Neumann algebra (M, τ ) endowed with a faithful
normal tracial state and which behaves like H(f1 , . . . , fn ), the continuous or
Background and outlook 5
are called max-stable laws (if they exist). A free analog has been found in joint
work of Gerard Ben Arous and myself. The definition for hermitian random
variables in a von Neumann algebra with faithful tracial state (M, τ ) of the
max is to define the spectral projections of X ∨ Y by
E(X ∨ Y ; (−∞, a)) = E(X; (−∞, a)) ∧ E(Y ; (−∞, a)).
We were able to classify these laws and show that, like in the classical case,
there are deeper things like the connection to Poisson processes which have
free analogs. This may be viewed as the half-full glass of the parallel for max-
stable laws, but there is also a half-empty glass to be considered which concerns
the applications. The well-known question about how high to build a dam in
Amsterdam in order that the probability of a flooding within the next 100
years be less than 1% based on the data on flooding for a certain number of
years, is answered using max-stable laws. One may wonder whether there is a
free analog to this application. What is a free dam in free Amsterdam to avoid
a free flooding with free probability . . . ?
Exchangeability. Claus Köstler and Roland Speicher discovered a free
de Finetti type theorem where invariance under classical permutations (which,
in the classical setting, gives conditional classical independence with respect to
a tail algebra of events) is replaced by invariance under the free quantum per-
mutation group (which then yields free independence with amalgamation over
a tail algebra). Due to work of Roland Speicher and Teodor Banica noncross-
ing partitions also make their appearance in this setting. It remains an open
6 D.-V. Voiculescu
Moritz Weber
1. Introduction
This small chapter provides some basic definitions and notions in free prob-
ability theory. It may be used as a warm-up for the following lectures as well
as a reference chapter for the most frequently used terms appearing in the
sequel. Most of the presented material is composed from Nica and Speicher’s
book [5]. Other sources are the books by Voiculescu, Dykema, and Nica [7]
and the one by Hiai and Petz [1]. Several aspects mentioned in this chapter
will be explained in more detail throughout the lectures of the other authors.
The author thanks Guillaume Cébron and Roland Speicher for helping to improve this
chapter, in particular the section on examples of noncommutative distributions.
8 M. Weber
Note that this product naturally extends the group multiplication when
identifying g ∈ G with g ∈ CG. We also have an involution on CG given
by X ∗ X
αg g := αg g −1 .
g∈G g∈G
We put
X X
A := CG, ϕ αg g := τG αg g := αe .
One can check that τG is actually a faithful trace.
(d) Let H be a Hilbert space with inner product h·, ·i and B(H) the algebra
of bounded linear operators on H. Let ξ ∈ H be a vector of norm one. We
put
A := B(H), ϕ(a) := haξ, ξi.
Replacing the scalars C somehow by a subalgebra B, we obtain a notion of
operator-valued probability spaces.
Definition 2.3 (Operator-valued noncommutative probability spaces). An
operator-valued noncommutative probability space (A, B, E) is the triple of a
unital algebra A over C, a subalgebra 1 ∈ B ⊂ A containing the unit of A, and
a conditional expectation E : A → B, i.e. E is a linear map such that E(b) = b
for all b ∈ B and E(b1 ab2 ) = b1 E(a)b2 for all a ∈ A and all b1 , b2 ∈ B.
Example 2.4. Let (A0 , ϕ) be a noncommutative probability space. Let
M2 (A0 ) be the algebra of 2 × 2-matrices with entries in A0 . Then the fol-
lowing is an operator-valued noncommutative probability space:
a b ϕ(a) ϕ(b)
A := M2 (A0 ), B := M2 (C), E := .
c d ϕ(c) ϕ(d)
Basics in free probability 9
3. Freeness
Definition 3.1 (Freeness). Let (A, ϕ) be a noncommutative probability space
and let 1 ∈ Ai ⊂ A be subalgebras containing the unit of A, for i ∈ I and some
index set I.
(a) The algebras (Ai )i∈I are tensor independent or classically independent, if
• ab = ba for all Qa ∈ Ai , b ∈ Aj with i 6= j,
n
• ϕ(a1 . . . an ) = j=1 ϕ(aj ) whenever aj ∈ Aij and all ij are mutually
different; for all n ∈ N.
(b) The algebras (Ai )i∈I are free or freely independent, if
ϕ(a1 . . . an ) = 0
whenever ϕ(aj ) = 0 for all j and aj ∈ Aij with i1 6= i2 6= · · · 6= in ; for all
n ∈ N.
(c) Random variables xi ∈ A, i ∈ I are called free, if the algebras Ai :=
alg(xi , 1) ⊂ A they generate are free. The random variables xi are called
∗
-free, if the ∗ -algebras Ai := alg(xi , x∗i , 1) ⊂ A are so, provided (A, ϕ) is
a ∗ -probability space. Likewise we define (∗ -)freeness for sets of random
variables via the unital (∗ -)algebras they generate.
Remark 3.2. There is also a definition of Boolean independence, as well as of
monotone and anti-monotone independence, see [4]. Further related concepts
are Male’s traffic freeness [3] or Lenczewski’s matricial freeness [2]. The very
recent concept of bi-freeness was initiated in 2013 by Voiculescu [6].
Example 3.3. (a) Let (L∞ (Ω, P), E) be the noncommutative probability space
as in Example 2.2 (a), and a, b ∈ L∞ (Ω, P) two random variables which are in-
dependent in the usual sense in probability theory. Then E(ak bl ) = E(ak )E(bl )
for all k, l ∈ N0 . Hence a and b are tensor independent in the sense of Defini-
tion 3.1 (a).
(b) Let us now come to the key example for freeness, namely the one from
which the definition of freeness was derived. Consider the free groups Fn and
Fm on n and m generators respectively. How do we capture the fact that Fn+m
can be written as the free products of Fn and Fm ? In other words, how to
formalize the fact that Fn and Fm sit freely inside Fn+m ?
Consider G := Fn+m with generators x1 , . . . , xn+m . We view G1 := Fn
and G2 := Fm as subgroups of G generated by x1 , . . . , xn and xn+1 , . . . , xn+m
respectively. Denote by e the neutral element in G. Then G1 , G2 ⊂ G are free
in the following sense. Whenever we take elements gj ∈ Gij with i1 6= i2 6=
· · · 6= in such that gj 6= e, then
g1 . . . gn 6= e.
In other words, elements from G1 and G2 share no relations. We take this as
a definition of freeness for groups.
This can easily be generalized to the group algebras: The algebras A1 :=
CFn and A2 := CFm in A := CFn+m have the property that whenever we take
10 M. Weber
P
aj = αjg g ∈ Aij with i1 6= i2 6= · · · 6= in such that αje = 0, then
X
a1 . . . an = βg g with βe = 0.
Now, how to generalize it again, to an operator algebraic setting? How to
formalize that the free group factors LFn and LFm are supposed to sit freely
inside LFn+m ? This is different from the case of group algebras, we cannot
simply say: “If the neutral element does not appear in the ai , it shall not
appear in the product of them.” The way out is provided by the trace τG as
defined in Example 2.2 (c). Reformulating the above property of freeness of
the group algebras, we obtain: The algebras A1 := CFn and AP 2 := CFm in
A := CFn+m are free in the sense that whenever we take aj = αjg g ∈ Aij
with i1 6= i2 6= · · · 6= in such that τG (aj ) = 0, then
τG (a1 . . . an ) = 0.
But this is exactly Definition 3.1 of freeness! Using τG , we can thus extend the
idea of freeness of groups to freeness of the enveloping von Neumann algebras.
Note that we have: If subgroups Gi ⊂ G for i ∈ I are free in the above
sense of freeness for groups, then also the group algebras (CGi )i∈I are free in
the noncommutative probability space (CG, τG ) in the sense of Definition 3.1.
This goes over to the C ∗ - and von Neumann algebras.
(c) Let H beL a Hilbert space. The full Fock space over H is by definition
∞
F (H) := CΩ ⊕ n=1 H ⊗n . Here, Ω is some vector of norm one, the so-called
vacuum vector. Following Example 2.2 (d), we remark that B(F (H)) can be
turned into a noncommutative probability space endowed with the vector state
ϕ(x) = hxΩ, Ωi. Given a vector ξ ∈ H, we define the (left) creation operator
l(ξ) ∈ F (H) by
l(ξ)Ω := ξ, l(ξ)η1 ⊗ · · · ⊗ ηn := ξ ⊗ η1 ⊗ · · · ⊗ ηn for n ≥ 1.
Its adjoint is called the (left) annihilation operator. Now, if ξ1 , . . . , ξn ∈ H
form an orthonormal system, then l(ξ1 ), . . . , l(ξn ) are ∗ -free in (B(F (H)), ϕ).
The example of creation operators on a Fock space is quite instructive in free
probability theory. Moreover, we can also define the right creation operator by
placing ξ to the right of the vectors. The interplay of left and right creation
operators is the starting point for Voiculescu’s definition of bi-freeness.
We now list a few basic properties of freeness. Freeness may be understood
as a rule for computing mixed moments, like classical independence (in the
sense of Definition 3.1). We begin with the classical situation. Let a, b ∈ (A, ϕ)
be tensor independent, and assume we know all moments {ϕ(am ) | m ∈ N}
of a and also all of b. Then we know all mixed moments in a and b, since
ϕ(an bm ) = ϕ(an )ϕ(bm ).
How about the free case? Let a, b ∈ (A, ϕ) be freely independent, and again
assume that we know all moments of a and b. Then we also know all mixed
moments in a and b, i.e. we can express ϕ(an1 bn2 . . . anm ) with ni ∈ N0 as a
polynomial in the moments of a and b. For instance, centering the variables
Basics in free probability 11
yields ϕ(a − ϕ(a)1) = 0 and likewise for b. Then, ϕ((a − ϕ(a)1)(b − ϕ(b)1)) = 0
by the definition of freeness, which yields
ϕ(ab) = ϕ(a)ϕ(b).
Likewise we infer
ϕ(aba) = ϕ(a2 )ϕ(b)
or
ϕ(abab) = ϕ(a2 )ϕ(b)2 + ϕ(a)2 ϕ(b2 ) − ϕ(a)2 ϕ(b)2 .
Proposition 3.4. Let (Ai )i∈I be free in (A, ϕ) and denote by B the algebra
generated by all Ai . Then the restriction ϕ|B is uniquely determined by ϕ|Ai ,
i ∈ I. Hence, knowing the moments of the variables in the Ai implies knowing
the moments of the elements in the algebra generated by all Ai .
Proposition 3.5. Let x, y, z ∈ (A, ϕ) be random variables.
(a) If x and y are freely independent and xy = yx, then we have
ϕ((x − ϕ(x)1)2 ) = 0 or ϕ((y − ϕ(y)1)2 ) = 0.
Thus, if x and y are in addition classically independent, selfadjoint, and
if ϕ is faithful, at least one of these variables is almost surely constant.
Hence, freeness is not a generalization of classical independence. It is
somehow the other extreme, suitable for the noncommutative situation.
(b) If x ∈ C1 is a constant, then x and y are free.
(c) The concept of freeness is commutative: If x and y are free, then also y
and x are free.
(d) The concept of freeness is associative: The variables x, y, z are free if and
only if x and {y, z} are free as well as y and z are free.
Definition 3.6 (Operator-valued freeness). Let (A, B, E) be an operator-
valued noncommutative probability space and let B ⊂ Ai ⊂ A be subalgebras
of A containing B. The algebras (Ai )i∈I ⊂ A are free with amalgamation over
B or free with respect to E, if
E(a1 . . . an ) = 0
whenever E(aj ) = 0 for all j and aj ∈ Aij with i1 6= i2 6= · · · 6= in ; for all
n ∈ N.
Random variables xi ∈ A, i ∈ I are called free, if the algebras alg(xi , B) are
free; and likewise ∗ -freeness is defined via the ∗ -algebras they generate.
In the case B = C the above definition simply boils down to freeness as in
Definition 3.1.
4. Noncommutative distributions
Denote by ChX1 , . . . , Xn i the polynomials in the noncommuting variables
X1 , . . . , Xn .
Definition 4.1 (Noncommutative distribution). Let (A, ϕ) be a noncommu-
tative probability space and let a1 , . . . , an ∈ A.
12 M. Weber
(h) An element a ∈ (A, ϕ) is a free Poisson with rate λ ≥ 0 and jump size
α ∈ R or the Marchenko–Pastur law, if its moments are given by
Xn
λk n n−1
ϕ(an ) = αn .
n−k+1 k k−1
k=1
The cumulants are
κn (a, . . . , a) = λαn .
The measure of the free Poisson law is given by
(
(1 − λ)δ0 + ν, if 0 ≤ λ ≤ 1,
ν, if λ > 1,
where ν has density
1 p √ √
t 7→ 4λα2 − (t − α(1 + λ))2 on [α(1 − λ)2 , α(1 + λ)2 ].
2παt
The square of a semicircular element of variance σ 2 is a free Poisson element
with rate λ = 1 and jump size α = σ 2 . For the definition of a compound
Poisson, see [5, Lecture 12].
(i) A selfadjoint variable b ∈ (A, ϕ) is called a symmetric Bernoulli variable,
if its moments are, with α > 0,
ϕ(b2m ) = α2m , ϕ(b2m+1 ) = 0.
The cumulants are
κ2m (b, . . . , b) = (−1)m−1 Cm−1 α2m , κ2m+1 (b, . . . , b) = 0
for m ∈ N0 . Its measure according to Proposition 4.2 is
1
(δ−α + δα ).
2
(j) Let p ∈ (A, ϕ) be a projection, i.e. p = p∗ = p2 , with ϕ(p) = t ∈ [0, 1].
Its moments are
ϕ(pn ) = t.
Its measure is
(1 − t)δ0 + tδ1 .
(k) The free Cauchy distribution is the distribution of an unbounded vari-
able. It is the same as the classical Cauchy distribution.
References
[1] F. Hiai and D. Petz, The semicircle law, free random variables and entropy, Math.
Surveys Monogr., 77, Amer. Math. Soc., Providence, RI, 2000. MR1746976
[2] R. Lenczewski, Matricially free random variables, J. Funct. Anal. 258 (2010), no. 12,
4075–4121. MR2609539
[3] C. Male, The distributions of traffics and their free product. arXiv:1111.4662 [math.PR]
(2011).
[4] N. Muraki, The five independences as natural products, Infin. Dimens. Anal. Quantum
Probab. Relat. Top. 6 (2003), no. 3, 337–371. MR2016316
16 M. Weber
[5] A. Nica and R. Speicher, Lectures on the combinatorics of free probability, London Math.
Soc. Lecture Note Ser., 335, Cambridge Univ. Press, Cambridge, 2006. MR2266879
[6] D.-V. Voiculescu, Free probability for pairs of faces I, Comm. Math. Phys. 332 (2014),
no. 3, 955–980. MR3262618
[7] D.-V. Voiculescu, K. J. Dykema, and A. Nica, Free random variables, CRM Monogr.
Ser., 1, Amer. Math. Soc., Providence, RI, 1992. MR1217253
Random matrices and combinatorics
Roland Speicher
1. Introduction
The notion of free independence was introduced by Voiculescu [7] in 1983
in the context of operator algebras, giving rise to free probability theory. In
1991, Voiculescu [10] discovered that this notion of freeness also appeared in
the context of random matrices. The latter had already been a subject of in-
vestigation in statistics (Wishart [13], 1928) and physics (Wigner [12], 1955)
for quite some time. One of the basic results in random matrix theory was
Wigner’s discovery that the eigenvalue distribution of a Gaussian unitary en-
semble is asymptotically given by the semicircular law. Since the semicircle
law is also the limit in the free version of a central limit theorem, this pointed
to a connection between free probability theory and random matrices. We
will first present a short introduction to random matrices and show Wigner’s
semicircle law, and then switch to the free probability side and show that the
semicircle shows also up as the limit in a free central limit theorem. This mo-
tivated Voiculescu to look for a deeper relation between random matrices and
asymptotic freeness. We will present a few examples of this connection in the
final Section 6. However, before coming to this, we will give a more thorough
treatment of the combinatorial structure of free probability theory, based on
the lattice of noncrossing partitions and the notion of free cumulants.
denotes the space of random variables for which all moments exist. Moreover,
tr denotes the normalized trace on MN (C) and E denotes the expectation on
∞−
(Ω, P). Hence we have A = (aij )Ni,j=1 ∈ A if and only if aij ∈ L (Ω, P) for
all i, j = 1, . . . , N and
N
1 X
ϕ(A) = tr ⊗ E(A) = E[tr(A)] = E[aii ].
N i=1
where P2 (m) denotes the set of pair-partitions of m elements (i.e., the de-
composition of the set {1, . . . , m} into disjoint pairs (r1 , s1 ), (r2 , s2 ), . . .). This
combinatorial formula, which expresses all higher moments of a Gaussian fam-
ily in terms of second moments, is usually called the Wick formula.
It might be interesting to note that, while the work of Wick (in physics) is
from 1950, the same formula was also shown by Isserlis in 1918 in probability
theory—thus the name “Isserlis formula” might also be appropriate.
Random matrices and combinatorics 19
second, since E[xi xj ] = δij σ 2 , the Wick formula only counts pairings of the
same xi ’s, hence it factorizes for different ones, corresponding to the indepen-
dence of different xi ’s.
Thus, the entries Re aij , Im aij (i, j = 1, . . . , N ) of a GUE A = (aij )N
i,j=1
form a Gaussian family, where the second moments are given by
1
E[aij akl ] = δil δjk for i, j, k, l = 1, . . . , N .
N
Now we can calculate the moments of the GUE A = (aij )N
i,j=1 for m even:
N
X
1
ϕ(Am ) = E[ai(1)i(2) ai(2)i(3) . . . ai(m)i(1) ]
N
i(1),...,i(m)=1
N
X X Y
1
= E[ai(r)i(r+1) ai(s)i(s+1) ]
N
i(1),...,i(m)=1 π∈P2 (m) (r,s)∈π
N
X X Y
1 1
= δi(r)i(s+1) δi(r+1)i(s)
N N
i(1),...,i(m)=1 π∈P2 (m) (r,s)∈π
X N
X m
Y
1
= 1+ m
δi(r)i(π(r)+1)
N 2
π∈P2 (m) i(1),...,i(m)=1 r=1
X N
X m
Y
1
= m δi(r)i(γπ(r)) ,
N 1+ 2
π∈P2 (m) i(1),...,i(m)=1 r=1
where #(γπ) denotes the number of cycles of γπ, we get the following theorem.
Theorem 2.7. For an N × N GUE random matrix A = (aij )N
i,j=1 we have
the genus expansion
X m
ϕ(Am ) = N #(γπ)−1− 2 .
π∈P2 (m)
Example 2.8. For m = 2, P2 (m) only consists of the element π = (1, 2) and
we have γ = (1, 2). Hence γπ = e and #γπ = 2 which yields the second
moment
ϕ(A2 ) = N 2−1−1 = 1.
In the case m = 4, P2 (m) contains the elements
π1 = (1, 2)(3, 4), π2 = (1, 3)(2, 4), π3 = (1, 4)(2, 3)
and we have γ = (1, 2, 3, 4). Hence
#(γπ1 ) − 3 = 0, #(γπ2 ) − 3 = −2, #(γπ3 ) − 3 = 0,
which yields
1 N →∞
ϕ(A4 ) = 2 + −−−−→ 2.
N2
In the same way we obtain
10 N →∞
ϕ(A6 ) = 5 + −−−−→ 5,
N2
70 21 N →∞
ϕ(A8 ) = 14 + 2 + 4 −−−−→ 14.
N N
More general, one has #(γπ) − 1 − m 2 ≤ 0 for all π ∈ P2 (m) and equality
holds exactly for the so-called noncrossing π. A pair-partition π ∈ P2 (m) is
crossing if we can find two blocks (r1 , s1 ) and (r2 , s2 ) of π which cross, i.e.,
with r1 < r2 < s1 < s2 .
Thus, for N → ∞, the moments of a GUE A = (aij )N
i,j=1 are given by
One can show that these are exactly the moments for the semicircular law, i.e.
Z 2 p
1
cm = tm 4 − t2 dt
2π −2
and hence we get the following theorem; see Figure 1.
Theorem 2.9 (Wigner’s semicircle law). The asymptotic eigenvalue distribu-
tion of a GUE A is given by the semicircle law, i.e.
lim µA = µS (weak convergence),
N →∞
where
1 p
dµS (t) = 4 − t2 dt on [−2, 2].
2π
In order to illuminate the parallels (and also the differences) between clas-
sical and free, we will give a uniform treatment of both the classical and the
free central limit theorem.
Consider a sequence (ai )∞ i=1 of elements of a noncommutative probability
space (A, ϕ) which are
• identically distributed,
• centered, i.e. ϕ(a1 ) = 0,
• normalized, i.e. ϕ(a21 ) = 1,
• either classically independent or freely independent.
Note that we require the existence of moments here.
What can we say about
a1 + · · · + aN
SN = √ ,
N
when N → ∞?
Definition 3.1. We say that SN ∈ (AN , ϕN ), N ∈ N converges in distribution
to x ∈ (A, ϕ), if
m
lim ϕN (SN ) = ϕ(xm )
N →∞
dist.
for all m ∈ N. This convergence is denoted by SN −−−→ x.
Let us see whether we can control the moments of SN when N goes to ∞.
We have
N
X
m 1 m 1
ϕ(SN )= m ϕ((a1 + · · · + aN ) )= m ϕ(ai(1) . . . ai(m) ).
N2 N2
i(1),...,i(m)=1
for large N (i.e. these sequences tend to the same limit), where P(m) denotes
the set of all partitions of {1, . . . , m} (for a formal definition see Definition 4.1),
Random matrices and combinatorics 23
we obtain
κπ = ϕ(a1 a2 a3 a3 a2 a1 )
= ϕ(a3 a3 )ϕ(a1 a2 a2 a1 ) = ϕ(a3 a3 )ϕ(a2 a2 )ϕ(a1 a1 ) = 1,
24 R. Speicher
and if
π = {(1, 5), (2, 3), (4, 6)} =
we have
κπ = ϕ(a1 a2 a2 a3 a1 a3 ) = ϕ(a2 a2 )ϕ(a1 a3 a1 a3 ) = 0,
by definition of freeness and the fact that all ϕ(ai ) = 0.
So we get in the limit of the free central limit theorem that the moments are
counted by the number of noncrossing pair-partitions; hence the same moments
as in the limit of Gaussian random matrices.
Theorem 3.2. Assume a1 , a2 , . . . ∈ (A, ϕ) are free and identically distributed
with ϕ(a1 ) = 0 and ϕ(a21 ) = 1. Then
a1 + · · · + aN dist.
√ −−−→ s,
N
where s is a semicircular element, i.e.
Z 2 p
1
ϕ(sm ) = tm 4 − t2 dt = #N C 2 (m).
2π −2
The free central limit theorem can be easily generalized to a multivariate
version:
(i) (i)
Theorem 3.3. Let {a1 | i ∈ I}, {a2 | i ∈ I}, . . . ⊂ (A, ϕ) be a sequence of
families of freely independent random variables with identical distribution and
(i)
ϕ(ar ) = 0 for all r ∈ N, i ∈ I. We denote the covariance by
cij = ϕ(a(i) (j)
r ar ), i, j ∈ I.
Then
a(i) + · · · + a(i)
1 N dist.
√ −−−→ (si )i∈I ,
N i∈I
where (si )i∈I is a semicircular family of covariance (cij )i,j∈I , i.e.
X Y
ϕ(si(1) . . . si(m) ) = ci(r)i(p)
π∈N C 2 (m) (r,p)∈π
for all m ∈ N.
This partial order induces a lattice structure on N C(n), i.e. for all π, σ ∈
N C(n) there is a minimal partition π ∨ σ that is larger than π and larger than
σ (called the join of π and σ) and a maximal partition π ∧ σ that is smaller
than π and smaller than σ (called the meet of π and σ).
Example 4.2. We have
∧ =
and
∨ = .
The lattice N C(n) has the maximal element 1n consisting of one block of size n
and the minimal element 0n consisting of n blocks of size one.
Definition 4.3. Let (A, ϕ) be a noncommutative probability space. The free
cumulants κn : An → C, n ≥ 1, are inductively defined by the moment-cumu-
lant formulas X
ϕ(a1 . . . an ) = κπ (a1 , . . . , an ),
π∈N C(n)
Hence κπ is defined by factorization into the blocks, where each κ|Vi | is applied
to those elements aj whose indices j lie in the block Vi .
Example 4.4. For n = 1, we get
κ1 (a1 ) = ϕ(a1 )
26 R. Speicher
where i, j belong to the same block of σ if and only if ai and aj are factors in
the same Ak .
Remark 4.6. We note that the condition π ∨ σ = 1n appearing in the sum in
the last theorem means that one has to consider all partitions π that couple
all blocks of σ.
Now we present the main result on free cumulants, which connects them to
freeness.
Theorem 4.7. Let (A, ϕ) be a noncommutative probability space. Unital sub-
algebras A1 , . . . , As ⊂ A are free if and only if all mixed cumulants vanish,
i.e.
κn (a1 , . . . , an ) = 0
whenever there are k, l ∈ {1, . . . , n} such that ak ∈ Ai(k) , al ∈ Ai(l) , and
i(k) 6= i(l).
Proof. It is easy to show by induction that A1 , . . . , As are free whenever
mixed cumulants vanish. On the other hand, if A1 , . . . , As are free, we note
that another easy inductive argument shows that κn (a1 , . . . , an ) = 0 when-
ever ϕ(ai ) = 0 for all i = 1, . . . , n, aj ∈ Ai(j) and i(j) 6= i(j + 1) for all
Random matrices and combinatorics 27
j = 1, . . . , n−1. The difficult part of the proof is to weaken this condition to the
one needed in the theorem. To do so, one shows first that κn (a1 , . . . , an ) = 0
if 1 ∈ {a1 , . . . , an } and n ≥ 2. But then we get that
κn (a1 , . . . , an ) = κn (a1 − ϕ(a1 )1, . . . , an − ϕ(an )1)
and hence we can get rid of the condition ϕ(ai ) = 0 for all i = 1, . . . , n.
Therefore κn (a1 , . . . , an ) = 0 whenever i(j) 6= i(j + 1) for all j = 1, . . . , n − 1.
Let there be k, l ∈ {1, . . . , n} such that i(k) 6= i(l). We multiply neigh-
bors from the same algebra together to get elements A1 , . . . , Am such that
A1 · · · Am = a1 · · · an and Aj , Aj+1 are from different algebras for all j =
1, . . . , m − 1. Hence
κm (A1 , . . . , Am ) = 0
but we also have
X
κm (A1 , . . . , Am ) = κπ (a1 , . . . , an )
π∈N C(n)
π∨σ=1n
X
= κn (a1 , . . . , an ) + κπ (a1 , . . . , an ).
π6=1n
π∨σ=1n
Now we assume that we know the statement for κl , l < n. Then we have that
κπ (a1 , . . . , an ) 6= 0 only for partitions π that couple elements ai from the same
algebra. But as σ also does so, we then have π ∨ σ 6= 1.
Applying the product formula once more, we also get the following result.
Theorem 4.8. Let (A, ϕ) be a noncommutative probability space. Elements
b1 , . . . , bs ∈ A are free if and only if, for all n,
κn (bi(1) , . . . , bi(n) ) = 0
whenever there are k, l ∈ {1, . . . , n} such that i(k) 6= i(l).
as mixed cumulants vanish by the results of the last section. Hence cumulants
are additive with respect to what we want to call free convolution. However,
at the moment the relation between moments and cumulants is just given on
a combinatorial level by summations over large sets of noncrossing partitions.
In order to be really useful, we need analytic tools for a better understanding
of this relation between moments and cumulants.
Theorem 5.2. We denote the n-th moment ϕ(an ) of a ∈ A by mn and we
consider the formal power series
X∞
M (z) = 1 + mn z n (moment series),
n=1
X∞
C(z) = 1 + κan z n (cumulant series).
n=1
Then the moment-cumulant relation
X
mn = κaπ
π∈N C(n)
is equivalent to
M (z) = C(zM (z)).
Proof. To simplify notation, we will write κπ instead of κaπ . The crucial ob-
servation is now that we can encode a noncrossing partition by its first block
V = (j1 = 1, j2 , . . . , js ) and noncrossing partitions π1 , . . . , πs of the points be-
tween consecutive points of V ; i.e., πr is a noncrossing partition of ir :=
jr+1 − jr − 1 points (where we put js+1 := n). This leads to the following
rewriting of our moment-cumulant formula:
X Xn X X X
mn = κπ = ··· κs κπ1 · · · κπs
π∈N C(n) s=1 i1 ,...,is ≥0 π1 ∈N C(i1 ) πs ∈N C(is )
i1 +···+is +s=n
n
X X
= κs mi1 · · · mis .
s=1 i1 ,...,is ≥0
i1 +···+is +s=n
In terms of
X∞ X∞
mn n cn n
A(z) = 1 + z and B(z) = 1 + z ,
n=1
n! n=1
n!
this is equivalent to
B(z) = log A(z).
Thus classical cumulants are essentially the coefficients of the logarithm of
the Fourier transform (or characteristic function) of the considered random
variable.
To be able to use analytic methods, it is useful to rewrite M (z) and C(z)
in terms of the Cauchy transform
1 X ∞
ϕ(an ) M ( 1z )
G(z) = ϕ = =
z−a n=0
z n+1 z
and Voiculescu’s R-transform
∞
X C(z) − 1
R(z) = κn+1 z n = .
n=0
z
Then, as
G( 1z )
M (z) = and C(z) = zR(z) + 1,
z
the relation M (z) = C(zM (z)) can be rewritten as
G( 1z ) 1 1
= zM (z)R(zM (z)) + 1 = G R G + 1.
z z z
1
Replacing z by z leads to
zG(z) = G(z)R(G(z)) + 1
and hence
1
R(G(z)) + = z,
G(z)
1
i.e. R(z) + z and G(z) are inverses under composition. Thus,
1
G R(z) + =z
z
also holds.
The advantage of G(z) over M (z) is that
1 Z 1
G(z) = ϕ = dµa (t)
z−a z−t
30 R. Speicher
Exercise 5.5.
(1) Calculate ( 12 δ−1 + 21 δ1 )⊞n for n = 2, 3, . . . .
(2) Prove: If a, b are free, then Ga+b (z) = Ga [z − Rb (Ga+b (z))].
5.6. Products. Let (A, ϕ) be a noncommutative probability space and let
a, b ∈ A be free. As in the sum case, we want to find a way of recovering
µab = µa ⊠ µb from µa and µb . We note that this is not an operation on real
probability measures as ab is in general not selfadjoint, even if a and b are.
However, if b ≥ 0, then b1/2 ab1/2 is selfadjoint and has the same moments as
ab (since ϕ is tracial on the algebra generated by a and b) and thus, we also
have
µa ⊠ µb = µb1/2 ab1/2 .
Calculating the moments directly, again, turns out to be rather complicated,
so we need other methods to determine µa ⊠ µb .
Let σ denote the pair-partition {{1, 2}, {3, 4}, . . . , {2n − 1, 2n}} for n ∈ N.
Then, for {a1 , . . . , an }, {b1 , . . . , bn } free, we have, by Theorem 4.5,
X
κn (a1 b1 , a2 b2 , . . . , an bn ) = κπ (a1 , b1 , a2 , b2 , . . . , an , bn ),
π∈N C(2n)
π∨σ=12n
and
X
ϕ(a1 b1 a2 b2 · · · an bn ) = κπ (a1 , . . . , an ) · ϕK(π) (b1 , . . . , bn ).
π∈N C(n)
Translating this into power series gives Voiculescu’s description via the S-
transform.
Theorem 5.8. Let (A, ϕ) be a noncommutative probability space and a ∈ A.
We denote the moment series of a by Ma (z) and the S-transform of a by
1 + z h−1i
Sa (z) = Ma (z),
z
h−1i
where Ma denotes the inverse with respect to composition. Then, if b, c ∈ A
are free, we have
Sbc (z) = Sb (z)Sc (z).
(p)
where P2 (m) denotes the set of pair-partitions π that respect the “coloring”
of the matrices, i.e. those π ∈ P2 (m) for which (r, s) ∈ π implies p(r) = p(s).
Thus
(p)
lim ϕ(A(p(1)) . . . A(p(m)) ) = #N C 2 (m).
N →∞
Random matrices and combinatorics 33
These limiting mixed moments of A(1) , . . . , A(n) are exactly those of a semi-
circular family s1 , . . . , sn with diagonal covariance, i.e.
X Y
ϕ(si(1) . . . si(m) ) = ϕ(si(r) si(p) )
π∈N C 2 (m) (r,p)∈π
and ϕ(si sj ) = δij for i, j = 1, . . . , n. Note that we can also rewrite this in
terms of cumulants as κn (si(1) , . . . , si(n) ) = 0 for n 6= 2 and
(
1, if i = j,
κ2 (si , sj ) =
0, if i 6= j.
Now we want to go further and find asymptotic freeness not just for semi-
circular distributions. For this we consider a GUE of N × N -random matri-
ces (AN )N ∈N and a sequence of deterministic matrices (DN )N ∈N (i.e. DN ∈
MN (C)) such that
m
lim tr(DN )
N →∞
dist.
exists for all m ∈ N, i.e. DN −−−→ d, where
where
1
E[aij akl ] = δil δjk .
N
Now, we consider a deterministic N × N -matrix D = (dij )N
i,j=1 and write
(k)
Dq(k) = (dij )N
i,j=1
for q(k) ∈ N, k = 1, . . . , m.
34 R. Speicher
X N
X m
Y
1 (1) (m)
= m δi(r)j(π(r)) dj(1)i(2) · · · dj(m)i(1)
N 1+ 2
π∈P2 (m) i(1),...,i(m)=1 r=1
j(1),...,j(m)=1
X N
X
1 (1) (m)
= 1+ m
dj(1)j(πγ(1)) · · · dj(m)j(πγ(m)) ,
N 2
π∈P2 (m) j(1),...,j(m)=1
as (
1, if π ∈ N C 2 (m),
κπ (s, . . . , s) =
0, otherwise.
Hence the asymptotic value of ϕ(ADq(1) . . . ADq(m) ) is given by the moment
ϕ(sdq(1) . . . sdq(m) ), provided that K(π) and πγ coincide. This is indeed the
case for general π ∈ N C 2 (m). We will check this in the following example.
Random matrices and combinatorics 35
1 1̄ 2 2̄ 3 3̄ 4 4̄ 5 5̄ 6 6̄ 7 7̄ 8 8̄.
where u1 , . . . , up are Haar unitaries and {u1 , u∗1 }, . . . , {up , u∗p }, {d1 , . . . , dq } are
free.
Remark 6.7. Note that, if u is a Haar unitary and {u, u∗ } is free from {a, b},
then a and ubu∗ are free. Thus, if (AN )N ∈N and (BN )N ∈N are two sequences of
deterministic matrices with limit distributions and (UN )N ∈N is a sequence of
∗
Haar unitary random matrices, then (AN )N ∈N and (UN BN UN )N ∈N are asymp-
totically free.
7. Comments
Random matrices have been studied in statistics and in physics since the
influential papers of Wishart [13] and Wigner [12], respectively. Random ma-
trices appear nowadays in different fields of mathematics and physics (such as
combinatorics, probability theory, statistics, operator theory, number theory,
quantum chaos, quantum field theory, etc.) or applied fields (such as electrical
engineering). Some idea of the diversity of random matrix appearances can be
gotten by looking on the collection of surveys in [1].
The genus expansion for Gaussian random matrices is a folklore result in
physics; for a mathematical exposition see, for example, [14].
The notion of “asymptotic freeness” was introduced by Voiculescu in [10].
Our presentation of the asymptotic freeness results for Gaussian random ma-
trices follows essentially the ideas of Voiculescu’s original proofs in [10, 11];
however, our presentation is more streamlined by using the Wick formula and
the genus expansion to make contact with the combinatorial description of
freeness.
The combinatorial approach to free probability theory originated in my
work [5] on free limit theorems. (At this time I was not aware of the work of
Kreweras on noncrossing partitions and addressed the latter, not very imagina-
tively, as “admissible” partitions.) Inspired by the work of Rota [4] around the
combinatorial structure of classical probability theory, featuring in particular
multiplicative functions on the lattice of all partitions, I developed a few years
later in [6] the full combinatorial description of freeness, resting on the notion
of multiplicative functions on the lattice of noncrossing partitions and “free cu-
mulants”. Andu Nica showed a bit later in [2] how this combinatorial approach
connects in general to Voiculescu’s operator-theoretic approach in terms of cre-
ation and annihilation operators on the full Fock space. I teamed then up with
Nica, pushing the combinatorial approach much further. Whereas in the be-
ginning we were mainly driven by the desire to understand Voiculescu’s work
by giving new and “simpler” (at least for the combinatorially inclined) proofs
of existing results of Voiculescu (like his R- and S-transform descriptions in
[8, 9] for the free additive and multiplicative convolutions, respectively), later
we could also initiate new directions in free probability. Prominent examples
here are the determination of the distribution of the free commutator, the
introduction of R-diagonal elements or the proof of the existence of the free
Random matrices and combinatorics 37
convolution power semigroup (µ⊞t )t≥1 . A good source for these developments
and the combinatorial approach in general is our monograph [3].
References
[1] G. Akemann, J. Baik, and P. Di Francesco, The Oxford handbook of random matrix
theory, Oxford Univ. Press, Oxford, 2011. MR2920518
[2] A. Nica, R-transforms of free joint distributions and non-crossing partitions, J. Funct.
Anal. 135 (1996), no. 2, 271–296. MR1370605
[3] A. Nica and R. Speicher, Lectures on the combinatorics of free probability, London Math.
Soc. Lecture Note Ser., 335, Cambridge Univ. Press, Cambridge, 2006. MR2266879
[4] G.-C. Rota, Gian-Carlo Rota on combinatorics, Contemporary Mathematicians, Birk-
häuser Boston, Boston, MA, 1995. MR1392961
[5] R. Speicher, A new example of “independence” and “white noise”, Probab. Theory
Related Fields 84 (1990), no. 2, 141–159. MR1030725
[6] R. Speicher, Multiplicative functions on the lattice of noncrossing partitions and free
convolution, Math. Ann. 298 (1994), no. 4, 611–628. MR1268597
[7] D. Voiculescu, Symmetries of some reduced free product C ∗ -algebras, in Operator alge-
bras and their connections with topology and ergodic theory (Buşteni, 1983), 556–588,
Lecture Notes in Math., 1132, Springer, Berlin, 1985. MR0799593
[8] D. Voiculescu, Addition of certain noncommuting random variables, J. Funct. Anal. 66
(1986), no. 3, 323–346. MR0839105
[9] D. Voiculescu, Multiplication of certain noncommuting random variables, J. Operator
Theory 18 (1987), no. 2, 223–235. MR0915507
[10] D. Voiculescu, Limit laws for random matrices and free products, Invent. Math. 104
(1991), no. 1, 201–220. MR1094052
[11] D. Voiculescu, A strengthened asymptotic freeness result for random matrices with ap-
plications to free entropy, Internat. Math. Res. Notices 1998, no. 1, 41–63. MR1601878
[12] E. P. Wigner, Characteristic vectors of bordered matrices with infinite dimensions, Ann.
of Math. (2) 62 (1955), no. 3, 548–564. MR77805
[13] J. Wishart, The generalised product moment distribution in samples from a normal
multivariate population, Biometrika 20A (1928), no. 12, 32–52.
[14] A. Zvonkin, Matrix integrals and map enumeration; an accessible introduction, Combi-
natorics and physics (Marseilles, 1995), Math. Comput. Modelling 26 (1997), no. 8-10,
281–304. MR1492512
Free monotone transport
Dimitri Shlyakhtenko
1. Introduction
The subject of optimal transportation [20] is an extremely rich and well-
developed theory.
Since its introduction by Voiculescu in the early 1980s [21, 26], free proba-
bility has established itself as a fertile source of both noncommutative analogs
of classical probability results as well as applications to operator algebras. It
is thus an important question to understand if free probability analogs of clas-
sical transportation results are available. In these notes we describe our joint
work with A. Guionnet [14] in which we were able to describe the beginnings
of such a theory.
subject to the condition that we end up with the desired configuration of dirt,
i.e., that the push-forward is
f∗ µ = ν.
(There are various choices for whatR is meant by the cost of f . In fact, Monge
originally considered the L1 -cost, |x − f (x)|dµ(x).)
More generally, let µ and ν be probability measures on Rn . One of the
first theorems from measure theory states that under mild assumptions on the
measures, there is always a function f : Rn → Rn such that the push-forward
of µ via the function f is exactly the measure ν (i.e. f∗ µ = ν).
Theorem 2.1. If µ is a nonatomic probability measure on Rn , then for any
probability measure ν there exists a function f : Rn → Rn such that f∗ µ = ν.
Proof. The “high-tech proof” goes as follows. The algebra A = L∞ (Rn , µ) may
be embedded into B(L2 (Rn , µ)) as multiplication operators. It is a weak closed
C ∗ -subalgebra, hence it is an abelian W ∗ -algebra. Furthermore, it is diffuse
(all minimal projections are zero) because µ has no atoms. The measure µ
gives one a state (even a trace) on A. From the general theory of W ∗ -algebras
it follows that A is universal in the class of separable abelian von Neumann
algebras: Any other separable abelian von Neumann algebra B with a trace ν
admits a trace-preserving embedding into A. But such an embedding exactly
corresponds to a map f from Rn → Rn so that f∗ µ = ν.
The “low-tech proof” (which of course underlies the high-tech one) proceeds
by cutting Rn into subsets of measures 2n , thus giving rise to an isomorphism
with a Cantor set.
In fact, there are a lot of transport maps—given f with f∗ µ = ν we can
replace f by f ◦ g for any µ-preserving g, but we want to have an optimal
solution.
Theorem 2.2 (Brenier 1991 [4]). Assume that µ and ν are probability mea-
sures on Rn (with compact support) and Lebesgue absolutely continuous.1 Then
there exists a unique function f : Rn → Rn such that f∗ µ = ν and f is the
gradient of a convex function g. Moreover, f solves the Monge problem (i.e. f
is an optimal transport):
kf − id kL2 (µ) = inf kfˆ − id kL2 (µ) .
fˆ∗ µ=ν
An approach to the proof of Brenier’s theorem comes with the theory of Kan-
torevich duality (which is a kind of convex optimization problem, see [4, 20]).
It is convenient to introduce the following definition.
Definition 2.3. We call f : Rn → Rn monotone if the Jacobian satisfies
Jac f ≥ 0 (i.e., it is positive semi-definite as a matrix).
1A weaker condition is that they do not give weight to small sets, i.e., sets of dimension
less than or equal to n − 1.
Free monotone transport 41
One can check that τ is both positive and satisfies our growth condition.
42 D. Shlyakhtenko
are adjoints of one another and satisfy L†j Lj = 1 and are thus bounded. Let
τ = h·P0 , P0 i. One easily checks that τ (Q(X1 , . . . , Xn )) = τ (Q(S1 , . . . , Sn ))
for any noncommutative polynomial Q. Indeed, it is sufficient to check this for
Q = Pr (i1 , . . . , ir ). In this case, τ (Q(X1 , . . . , Xn )) = 0 if r 6= 0 by orthogo-
nality. On the other hand, one easily proves recursively that τ (Q(S)) = 0 as
well. Boundedness of Xj follows from Xj Q = (Lj + L†j )Q and boundedness of
Lj , L†j .
Recall that there is a measure µ such that for any measure ν there exists
a function f such that ν = f∗ µ. In particular, L∞ (Rn , µ), where µ is the
Gauss measure, contains all abelian W ∗ -algebras (see Theorem 2.1). Hence,
it is natural to ask whether there exists a faithful trace τ such that W ∗ (τ )
contains all tracial separable W ∗ -algebras. The next theorem states that there
is no noncommutative analog of this.
Free monotone transport 43
Proof. First note that we can view the W ∗ -algebra M1 generated by a free
semicircular system S1 , S3 , . . . , S2n−1 as a W ∗ -subalgebra of
M := W ∗ (S1 , S2 , S3 , . . . , S2n ).
The von Neumann algebra M in turn is the free product of the algebras M1
and M2 := W ∗ (S2 , S4 , . . . , S2n ). Furthermore, M1 is isomorphic to the free
44 D. Shlyakhtenko
M1 L
2
(M1 ∗ M2 )M2 ∼
= M1 (L2 (M1 ) ⊗ L2 (M2 ))⊕∞
M2
N1 L
2
(M1 ∗ M2 )N2 ∼
= N1 (L2 (N1 ) ⊗ L2 (N2 ))⊕∞
N2
∼ N HS(L2 (N1 ), L2 (N2 ))⊕∞
= 1 N2
∼
= N HS(L2 (N ), L2 (N ))⊕∞
N .
yields
X X
Xj Dj (Xi1 . . . Xin ) = Xj Xik+1 . . . Xin Xi1 . . . Xik−1
j k
= nS(Xi1 . . . Xin ).
Free monotone transport 51
Let NPbe the operator that multiplies each monomial by its degree. We infer
that Xj Dj g = N g (cp. [23]). This allows us to re-write our equation as
h 1X i
g = N −1 S (τ ⊗ 1 + 1 ⊗ τ )(Tr log(I + J Dg)) − (Dj g)2 − βW (X + Dg) ,
2
i.e., in the form
g = F (g).
It turns out that our F is contractive on a ball for k·kR , hence we have a
fixpoint and thus a solution by Picard iteration.
As β → 0, the solution F = X + f converges to X and at some point it
becomes invertible. So for small β there exists F such that Y = F (X1 , . . . , Xn )
has law τVβ and F h−1i such that F h−1i (Y1 , . . . , Yn ) has semicircular law τ and
J F is positive.
6. Applications
It is a ten years old question asked by Voiculescu [25] if for all β we have
W ∗ (τβ ) ∼
= W ∗ (τsc ) with τsc the semicircle law. Now we know the following
result.
Corollary 6.1. For small β we have that W ∗ (τβ ) is isomorphic to W ∗ (τsc ).
It turns out that the so-called q-deformed free group factors of Bozejko and
Speicher [3] are particular examples of factors of the form W ∗ (τβ ). Thus we
can formulate the following corollaries.
Corollary 6.2. The q-deformed free group factors are isomorphic to free group
factors for small q.
Corollary 6.3. C ∗ (τβ ) is isomorphic to C ∗ (τsc ) for small values of β.
As C ∗ (τsc ) is projectionless by a result by Pimsner and Voiculescu (1981), so
is C ∗ (τβ ). This has important consequences for the histogram of eigenvalues.
In the classical case we know that for all measures µ on Rn with den-
sity ρ(x1 , . . . , xn ) there exists a unique transport map f which transports the
Gaussian distribution to the measure µ. The Jacobian of this transport map
more or less is the density ρ. Hence, having a transportation map is even
better than having a density, since it exists in more general settings.
In free probability however, we do not have densities, but given a law τ there
may be a transportation map F such that τ = F∗ τsc . In this case, we have the
following relation between the free entropy χ and the transport map F :
χ(τ ) = (τ ⊗ τ ) Tr log(J F ) + χ(τsc ).
(N ) sa n
We define a measure µV on (MN ) via
(N ) 1 −N Tr(V (X1 ,...,Xn ))
µV (dA1 · · · dAn ) = e .
ZV,N
52 D. Shlyakhtenko
2 2
By Brenier’s theorem, we find a transportation map F̃ (N ) : RnN → RnN
(N )
transporting the Gaussian measure µG to µV . We use the following random
matrix model:
(N )
H−N Tr(V (A1 ,...,An )) (µV ) → χV (τV ) as N → ∞.
Here, χV (τV ) is the free entropy relative to V . Applying F̃ (N ) to the random
matrices via functional calculus yields a map F (N ) with
kF (N ) − F̃ (N ) kL2 ( N1 Tr ◦EµG ) → 0 as N → ∞.
Here, the infimum runs over all tuples X1 , . . . , Xn which are distributed by µ
and likewise Y1 , . . . , Yn according to ν.
In this context the Brenier map is optimal.
In the noncommutative case the metric dW2 is generalized by the Biane–
Voiculescu–Wasserstein distance [2]
X
d(τ, τ ′ )2 = x ,...,x
inf ∼τ kxi − yj k2L2 (M(τ )∗M(τ ′ )) .
1 n
y1 ,...,yn ∼τ ′
But it is still an open question if in this case our transport map is optimal.
8. Comments
Our result on the existence of free monotone transport is an example of the
use of analysis tools from free probability theory in the context of von Neumann
algebra theory, a topic that has a long history. Indeed, such applications were
the original motivation of Voiculescu in developing free probability.
Since the times of these lectures, there have been a number of important
developments, both around free transport and this more general area of ap-
plications of free probability theory. One such development is the work of
Free monotone transport 53
B. Nelson [16], who was able to extend our free monotone transportation re-
sults to nontracial context. The tracial requirement is absolutely essential in
our work, and the understanding of how to handle it in the nontracial context
is a big step forward. In another direction, Bekerman, Figalli and Guionnet
[1, 9] were able to substantially improve our results from [14] relating transport
between random matrix models of finite size and their free probability limit;
using these ideas they were able to obtain universality results for eigenvalues
of certain polynomials in arbitrary GUE matrices.
Although we did not mention it in our lecture, a lot of the work we have
described in the free probability setting has an extension to Jones subfactor
theory. The starting point [10] is the replacement of the ring of noncommu-
tative polynomials by a different ring, coming from a so-called Jones planar
algebra [15] of a subfactor inclusion. In a certain sense, this is akin to passage
from analysis on Rn to analysis on a certain symmetric space Rn /G. The
significant difference in the noncommutative case is the generality of the sym-
metry group G, which can be a “quantum symmetry” in the subfactor theory
sense. It turns out that both random matrix models [11], free Gibbs states
[5, 11] and transportation theory [17] are available in this greater generality.
Finally, it is worth mentioning that advanced tools from free probability
theory—such as stochastic differential equations—continue to play an impor-
tant role in von Neumann algebra theory. Let us just mention the work of
Dabrowski and Ioana [8], which builds up on applications of free probability
found in Dabrowski’s earlier work [6, 7].
An important future direction for research in free transportation theory is
the analysis of free transportation beyond the perturbative regime. This is a
subject of active research.
9. Exercises
Recall from Definitions 4.1 and 4.2 that for the algebra of noncommutative
polynomials A = C[X1 , . . . , Xn ], we define the derivations ∂j : A → A ⊗ A
given on monomials by
X
∂j P = A ⊗ B.
P =AXi B
Define also Dj P by
X
Dj P = BA.
P =AXi B
References
[1] F. Bekerman, A. Figalli, and A. Guionnet, Transport maps for β-matrix models and
universality, Comm. Math. Phys. 338 (2015), no. 2, 589–619. MR3351052
[2] P. Biane and D. Voiculescu, A free probability analogue of the Wasserstein metric on
the trace-state space, Geom. Funct. Anal. 11 (2001), no. 6, 1125–1138. MR1878316
[3] M. Bożejko and R. Speicher, An example of a generalized Brownian motion, Comm.
Math. Phys. 137 (1991), no. 3, 519–531. MR1105428
56 D. Shlyakhtenko
Ken Dykema
1. Introduction
Infinite-dimensional von Neumann algebras that have trivial center and tra-
cial states are called II1 -factors. In the beginning, there were two of them:
Murray and von Neumann proved [16] that the hyperfinite II1 -factor (obtained
as a limit of finite-dimensional matrix algebras) is not isomorphic to the group
von Neumann algebra of the free group of two generators, L(F2 ). At the 1967
Batan Rouge conference, R. V. Kadison gave a list of open questions, includ-
ing the question of whether L(F2 ) and L(F3 ) are isomorphic (see [12]). Several
of Kadison’s questions have been answered, and today much more is known
about II1 -factors; in particular, there are uncountably many nonisomorphic ex-
amples, and some of them have quite exotic properties. But the isomorphism
question for free group factors remains open.
Voiculescu’s free probability theory and his random matrix results (of the
1980s and early 1990s) opened up new ways to try to understand these free
group factors. These lectures aim to provide an introduction to these random
matrix techniques and applications, by way of summarizing results obtained
in the 1990s and providing proofs (or sketches of proofs) of some of them. We
also include a section listing some more recent results about free group factors.
The author would like to thank Nicolai Stammeier and Moritz Weber for organizing the
masterclass and for assisting mightily in the production of these notes. (Any mistakes are,
of course, the responsibility of the author.)
58 K. Dykema
only if a = b∗ b for some b ∈ A. C∗ -algebras are the best places to work with
positivity in a noncommutative context.
Definition 2.1. A C∗ -noncommutative probability space is a pair (A, φ), where
A is a unital C∗ -algebra and φ is a state on A, i.e., φ : A → C is a linear
functional such that φ(1) = 1 and φ(a) ≥ 0 for all positive elements a ∈ A.
Our first goal is the following. Given two C∗ -noncommutative probability
spaces (A1 , φ1 ) and (A2 , φ2 ), we want to find a C∗ -n.c.p.s. (A, φ) such that
• Ai ֒→ A for i = 1, 2,
• φ|Ai = φi for i = 1, 2, and
• A1 and A2 are free with respect to φ.
We can do it if the GNS-representations of φ1 and φ2 are faithful. Recall
that given a C∗ -n.c.p.s. (A, φ), we can define a sesquilinear form on A by
1
ha1 , a2 iφ := φ(a∗2 a1 ). This yields a semi-norm kak2 := φ(a∗ a) 2 . Now separa-
tion and completion leads to a Hilbert space L2 (A, φ) and a canonical map
A → L2 (A, φ), denoted a 7→ â. The GNS-representation is the map
Here, vj ∈ Hi0j , ij 6= ij+1 . But that is not quite correct. More precisely, for a
unit vector ηi , we set
M
K(i) := Cηi ⊕ Hi01 ⊗ Hi02 ⊗ · · · ⊗ Hi0n
n≥1
i1 ,...,in ∈I
ij 6=ij+1 ,i1 6=i
By iteration we obtain
φ(λi1 (a1 ) · · · λin (an )) = hλi1 (a1 ) · · · λin (an )ξ, ξi
= hâ1 ⊗ · · · ⊗ ân , ξi
= 0.
Uniqueness: Note that the linear span of the set
{1} ∪ {λi1 (a1 ) · · · λin (an ) | aj ∈ Aij ∩ ker φij , ij 6= ij+1 }
is dense in A and freeness determines φ and h·, ·i uniquely.
Proposition 3.2. Some facts about the above construction:
(1) If the φi are traces for all i ∈ I, then φ is a trace (i.e. the tracial property
φ(ab) = φ(ba) is fulfilled).
(2) If the φi are faithful for all i ∈ I, then φ is faithful.
We leave the proof of (1) as an exercise. The first proof of (2) was in [8],
but a better proof is by E. Ricard, including the case of amalgamated free
products. His argument is reproduced in the paper [14].
Example 3.3. For a discrete group G we define the reduced group C∗ -algebra
by
Cr∗ (G) = spank·k {λ(g) | g ∈ G} ⊆ B(ℓ2 (G)).
Here λ is the left regular representation of G, given by λ(g)δh = δgh . A state
on Cr∗ (G) is given by τG (·) = h· δe , δe i. Thus
(
1, if g = e,
τG (λ(g)) =
0, otherwise.
Let (Gi )i∈I be a family of discrete groups such that G = ∗i∈I Gi , then
∗ (Cr∗ (Gi ), τGi ) = (Cr∗ (G), τG ).
i∈I
ψ ◦ π = φ. This is not necessarily true if the φi are nonfaithful, while still hav-
ing faithful GNS representations, of course (see [10] for a counter example).
However, by [4], if (Bi , ψi )i∈I are C∗ -noncommutative probability spaces with
faithful GNS-constructions and if there exist ∗-homomorphisms πi : Ai → Bi
satisfying ψi ◦ πi = φi , then there exists a “free product” ∗-homomorphism
π : A → B, where (B, ψ) := ∗i∈I (Bi , ψi ), satisfying π|Ai = πi and ψ ◦ π = φ.
We will also use the notion of a W ∗ -noncommutative probability space (ab-
breviated W ∗ -n.c.p.s.), which is a pair (M, φ), where M is a von Neumann
algebra and φ is a normal state. Recall that a von Neumann algebra is a
unital ∗ -subalgebra of B(H) which is closed in the weak operator topology.
Since M contains plenty of projections, this is the right place to do noncom-
mutative measure theory. The following is a von Neumann algebra analog of
Theorem 3.1.
Theorem 3.4 (Voiculescu [26]). Let I be a nonempty set and for all i ∈ I, let
(Ai , φi ) be W ∗ -noncommutative probability spaces with faithful GNS-represen-
tations. Then there exists a unique W ∗ -n.c.p.s. (A, φ) equipped with injective,
normal ∗ -homomorphisms λi : Ai → A such that
(1) φ ◦ λi = φi ,
(2) the (λi (AS i ))i∈I are free (with respect to φ),
(3) A = W ∗ ( i∈I λi (Ai )),
(4) the GNS-representation of φ is faithful.
Proof. Let (A, φ̊) = ∗i∈I (Ai , φi ) be the C∗ -algebraic free product with A rep-
resented on the free product Hilbert space L2 (A, φ̊). By construction, φ̊ is the
s.o.t.
restriction to A of the vector state h·ξ, ξi, where ξ = 1̂. Let A = A and let
φ be the vector state h·ξ, ξi on A.
The analog of Proposition 3.2 was proved by Voiculescu.
Proposition 3.5 (Voiculescu [26]). About the W∗ -algebra free product:
(1) If the φi are traces for all i ∈ I, then φ is a trace.
(2) If the φi are faithful for all i ∈ I, then φ is faithful.
Example 3.6. For a discrete group G we define the group von Neumann
algebra by
s.o.t.
L(G) = Cr (G) ⊆ B(ℓ2 (G)).
Then the canonical trace τ is given by τG (·) = h· δe , δe i. If (Gi )i∈I is a family
of discrete groups such that G = ∗i∈I Gi , then
∗ (L(Gi ), τGi ) = (L(G), τG ).
i∈I
R
Let E : L → C be the expectation E(f ) = f dω. For f ∈ L, we write
f ∼ N (0, σ 2 ) to mean that f has a Gaussian distribution with first moment
zero and second moment σ 2 . The algebra MN (L) of N × N -matrices over L
endowed with the tracial expectation φN = E ◦ trN is a ∗-noncommutative
probability space.
Definition 4.1. We define the following sorts of random matrices.
(1) A selfadjoint Gaussian random matrix2 is X = (xij ) ∈ MN (L) such that
xii ∼ N (0, σ 2 ), Re xij , Im xij ∼ N (0, σ2 ) for i < j; xij = xji ; the random
variables (xii )i , (Im xij )i<j and (Re xij )i<j are independent. We denote
this by X ∈ SGRM(N, σ 2 ).
(2) A Gaussian random matrix is X = (xij ) ∈ MN (L) such that Re xij ,
2
Im xij ∼ N (0, σ2 ) and (Im xij )i,j and (Re xij )i,j are independent. We
denote this by X ∈ GRM(N, σ 2 ).
(3) A Haar unitary random matrix is a UN -valued random matrix U ∈ MN (L)
(i.e., U (ω) is a unitary for all ω ∈ Ω) with Haar measure distribution. We
denote this by U ∈ HURM(N ).
Definition 4.2. Let K be an index set and let a(k, N ) ∈ MN (L). We say
that the family (a(k, N ))k∈K converges in ∗ -moments (or in ∗ -distribution) if
lim φN (a(k1 , N )ε1 a(k2 , N )ε2 · · · a(kp , N )εp )
N →∞
exists for all p ∈ N, εj ∈ {1, ∗} and kj ∈ K. The family converges in ∗ -moments
to a family (ak )k∈K of random variables in a ∗-n.c.p.s. (A, φ) (which may be
called limit random variables) if
ε
lim φN (a(k1 , N )ε1 a(k2 , N )ε2 · · · a(kp , N )εp ) = φ(aεk11 aεk22 · · · akpp )
N →∞
for all εj ∈ {1, ∗} and kj ∈ K. We further say that the family (a(k, N ))k∈K
is asymptotically free (respectively, asymptotically ∗-free) if it converges in ∗ -
moments to a family of random variables that is free (respectively, ∗-free).
Analogously, a family of sets of random variables is asymptotically free or
∗-free if the corresponding family of sets of limit random variables is free or
∗-free.
Voiculescu’s discovery [28] that certain random matrices are asymptotically
free (we call this the matrix model ) was a key result for applications of free
probability theory to von Neumann algebras. A strengthening of Voiculescu’s
theorem from [28] is the following.
Theorem 4.3 (Voiculescu [33]). For s ∈ N let X(s, N ) ∈ SGRM(N, N1 ),
Z(s, N ) ∈ GRM(N, N1 ) and U (s, N ) ∈ HURM(N ). Assume that they are all
independent, i.e. for each N , the σ-algebras
(ΣXs,N )∞ ∞ ∞
s=1 , (ΣZs,N )s=1 , (ΣUs,N )s=1
generated by these random matrices are independent. Suppose that B(s, N ) ∈
MN (C) is such that the family (B(s, N ))s∈N converges in ∗ -moments and for
all s, kB(s, N )k remains bounded as N → ∞.
Free group factors 63
GRM( N2 , N1 ) and these are independent, applying Theorem 4.3, we obtain that
e11 Me11 is generated by four free semicircular elements e11 xe11 , Re(e11 xe21 ),
Im(e11 xe21 ) and e12 xe21 . Hence,
4
e11 Me11 = ∗ (L(Z), τZ ) ∼
= (L(F4 ), τF4 ).
i=1
Finally, we have M ∼
= L(F4 ) ⊗ M2 (C).
What happens if we take the first isomorphism of Theorem 4.4 and let k
S to infinity? Recall that the hyperfinite II1 -factor R is given by the closure
tend
of k∈N M2k (C) in the weak operator topology. Could we hope to deduce an
isomorphism L(Z) ∗ R ∼ = L(F∞ ) ⊗ R from the above theorem? Such hope
would be in vain, since L(F∞ ) ⊗ R has property Γ whereas L(Z) ∗ R has not.
It is a question of central sequences. (Note that Murray and von Neumann
showed already in [16] that L(F2 ) ∼6 R because L(F2 ) does not have central
=
sequences.) Instead, we have the following theorem.
Theorem 4.5 ([6]). The free product von Neumann algebra L(Z) ∗ R is iso-
morphic to the free group factor L(F2 ). Consequently, for n ≥ 2,
L(Fn ) ∗ R ∼
= L(Fn+1 ).
The proof is similar in spirit to the proof of the above theorem, dividing
the semicircular element x = x∗ ∈ L(Z) into finer and finer matrix blocks and
using asymptotic freeness results for random matrices.
Using his matrix model, Voiculescu proved another theorem on the free
group factors.
Theorem 4.6 (Voiculescu [27]). If k ∈ N and n ∈ {2, 3, . . . , ∞}, then
L(Fn ) ∼
= L(F1+k2 (n−1) ) ⊗ Mk (C).
Let us now turn to Murray and von Neumann’s [16] rescaling of a II1 -
factor M with unique tracial state τ . Recall that a II1 -factor is an infinite-
dimensional von Neumann algebra M with trivial center (i.e. Z(M ) = C1) and
a normal faithful tracial state τ : M → C. II1 -factors have nice properties,
in some sense they behave like matrix algebras but infinite ones. We have
already encountered the free group factors L(Fn ), which are of type II1 , and
the hyperfinite II1 -factor R.
For 0 < t < 1 and a projection p ∈ M with τ (p) = t we define Mt = pM p. If
0 < t < ∞, then Mt = p(M ⊗ Mn (C))p , where p ∈ M ⊗ Mn (C) is a projection
such that τ ⊗ Trn (p) = t. Here, Trn is the unnormalized trace on Mn (C). This
is the rescaling of the II1 -factor M by t, and the result Mt is a II1 -factor. Note
that if p, q ∈ M are projections such that τ (p) = τ (q) then there exists v ∈ M
such that v ∗ v = p and vv ∗ = q, which implies that the choice of p is irrelevant,
so long as τ ⊗ Trn (p) = t. We have (Ms )t ∼ = Mst .
Definition 4.7. The fundamental group of a II1 -factor M is the subgroup of
the multiplicative group R∗+ := {t ∈ R | t > 0} defined by
F (M ) = {t ∈ R∗ | Mt = ∼ M }.
+
Free group factors 65
Furthermore, the isomorphism problem for the free group factors boils down to
the following dichotomy.
• either L(Fr ) ∼
= L(Fs ) for all 1 < r, s ≤ ∞ and then F (L(Fr )) = R∗+ ,
• or L(Fr ) ≇ L(Fs ) for all 1 < r < s ≤ ∞ and then F (L(Fr )) = {1}.
Proof. First we construct the interpolated free group factors. We consider
R ∗ L(F∞ ). Recall that L(F∞ ) is generated by a free semicircular P family
(xn )n∈N . Let p1 , p2 , . . . ∈ R be projections such that r = 1 + j τ (pj )2 . We
define
L(Fr ) := W ∗ (R ∪ {pj xj pj | 1 ≤ j < ∞}) ⊂ R ∗ L(F∞ ).
Using the matrix model, we can show that this is well-defined (i.e., it is inde-
pendent of the choice of the pj ).
To prove (2), consider L(Fr ) for 1 < r ≤ ∞ and let t ∈ (0, 1). Choose
a realization of L(Fr ) as above, such that there is a projection q ∈ R with
τ (q) = t and pj ≤ q for all j. We obtain
L(Fr )t ∼
= qL(Fr )q = W ∗ (qRq ∪ {pj xj pj | 1 ≤ j < ∞}) ∼
= L(Fx )
where, since qL(Fr )q is endowed with the trace τ (q)−1 τ |qL(Fr )q , we have
X
x=1+ (t−1 τ (pj ))2 = 1 + t−2 (r − 1).
j
This proves (2) in the case t ≤ 1, and the case t > 1 follows from this.
Assertion (3) follows from the result R ∗ R ∼ = L(Z) ∗ R and the above
construction of the interpolated free group factors.
In order to show the dichotomy result, let us first assume that L(Fr ) is
isomorphic to L(F∞ ) for some 1 < r < ∞. For every number s ∈ (1, ∞) we
find a 0 < t < ∞ such that s = 1 + t−2 (r − 1) and hence we obtain
= L(Fr )t ∼
L(Fs ) ∼ = L(F∞ )t = L(F∞ ).
Second, if L(Fr ) is isomorphic to L(Fs ) for some 1 < r < s < ∞, then for all
0 < t < ∞,
= L(Fs )t ∼
= L(Fr )t ∼
L(F1+t−2 (r−1) ) ∼ = L(F1+t−2 (s−1) ).
Now, let 0 < ε < 1 and choose 0 < t < ∞ such that ε = t−2 (r − 1). Then
s − 1
1 + t−2 (s − 1) = 1 + ε
r−1
and
L(F3 ) ∼
= L(F1+ε ) ∗ L(F2−ε ) ∼
= L(F s−1 ) ∗ L(F2−ε )
1+ε( r−1 )
∼ ∼
s−r ) = L(F3 )α
= L(F3+ε( r−1 )
in [24]). We can choose projections pj ∈ R such that τ (pj ) > 0 for all j ≥ 1,
P ∞ 2
j=1 τ (pj ) = 2 and p2k−1 = p2k for all k ≥ 1. Thus,
M := W ∗ (R ∪ {pj xj pj | j ≥ 1}) ∼
= L(F3 )
and for each k, the von Neumann subalgebra having pk for identity element and
generated by {p2k−1 x2k−1 p2k−1 , p2k x2k p2k } is isomorphic to L(F2 ). However,
from what we have already shown, it follows that L(F2 ) ∼ = L(Fm ) for all integers
m ≥ 2. We may replace {p2k x2k−1 p2k , p2k x2k p2k } by any set that generates
the same von Neumann algebra, and a set of m free semicircular generators
in this von Neumann algebra can be realized as the compressions by p2k of
m free semicircular elements in M. Thus, choosing integers mk ≥ 2, if we
renumber the semicircular family (xj )∞j=1 to be indexed by N × N, we see that
M is isomorphic to the von Neumann algebra generated by
R ∪ {p2k xk,j p2k | k ∈ N, 1 ≤ j ≤ mk },
and this von Neumann algebra
P∞is isomorphic to L(F∞ ) provided that we choose
the values of mk to ensure k=1 mk τ (p2k )2 = ∞.
Recall we have
L(Z) ∗ R ∼
= L(F2 ),
∼
R ∗ R = L(F2 ),
M2 (C) ∗ M2 (C) ∼
= L(F3 ) ⊗ M2 (C) ∼
= L(F 23 ),
L(Z) ∗ M2 (C) ∼= L(F4 ) ⊗ M2 (C) ∼
= L(F 7 ). 4
Thus, taking the free product with the hyperfinite factor R behaves like taking
the free product with L(Z), whereas taking the free product with M2 (C) rather
looks like taking the free product with L(F 34 ), whatever that may be. This
pattern is part of a larger picture, which is summarized in the next result and
ensuing description.
Theorem 5.2 ([5]). Let A and B be von Neumann algebras with normal,
faithful tracial states τA and τB respectively, and assume dim A ≥ 2, dim B ≥ 3.
Suppose each is
(1) either finite dimensional,
(2) or hyperfinite,
(3) or an interpolated free group factor,
(4) or a (possibly infinite) direct sum of these.
Then, the free product (M, τ ) := (A, τA ) ∗ (B, τB ) is of the form M ∼= L(Ft )
or M ∼ = L(Ft ) ⊕ N for some t ∈ (1, ∞] and a finite-dimensional von Neumann
algebra N (which can be described exactly).
To find the parameter t in the preceding theorem, we use a “free dimension”
with the following properties:
(1) fdim(M, τ ) = fdim(A, τA ) + fdim(B, τB ).
(2) fdim(Mk (C)) = 1 − k12 .
68 K. Dykema
This theorem has been strengthened by Ozawa and Popa. A II1 -factor M
is strongly solid, if for all diffuse hyperfinite subalgebras A ⊂ M of M , the
normalizer subalgebra NM (A) is hyperfinite.
Theorem 6.4 (Ozawa, Popa [19]). L(Ft ) is strongly solid, t ∈ (1, ∞).
In particular, this implies that free products such as (L(Ft ) ⊗ R) ∗ L(Z) and
(L(Ft ) ⊗ L(Z)) ∗ L(Z) are not interpolated free group factors.
From Theorem 5.1, we have L(Fr ) ⊗ R ∼ = L(Fs ) ⊗ R for all 1 < r, s < ∞
but not L(F2 ) ⊗ R ∼= L(F ∞ ) ⊗ R. Though it is not a priori clear whether such
an isomorphism would imply isomorphism of free group factors, this does, in
fact, follow from the following result.
There is also the following Kurosh-type theorem for II1 -factors. A II1 -factor
is semi-exact if it contains a weakly dense exact C∗ -algebra. It is semi-solid if
the relative commutant of every subfactor of type II1 is hyperfinite.
A1 ∗ A2 ∗ . . . ∗ An ≇ B1 ∗ B2 ∗ . . . ∗ Bm ,
A0 ∗ A1 ∗ A2 ∗ . . . ∗ An ≇ B0 ∗ B1 ∗ B2 ∗ . . . ∗ Bm .
and
Proof. Starting with L(F∞ ) ⊗ L(Z) and splitting off either from the one side
of the tensor product or the other, we have the isomorphisms
L(F∞ ) ⊗ L(Z) ∼
= (L(F∞ ) ⊗ L(Z)) ⊗ M2 (C) ∼
= (L(F∞ ) ⊗ L(Z)) ⊗ L(Z2 ).
70 K. Dykema
7. Exercises
Exercise 7.1. Suppose (A, φ) = ∗i∈I (Ai , φi ) is a reduced free product of C∗ -
algebras (or a free product of von Neumann algebras) and that for every i ∈ I,
φi is a trace. Show that φ is a trace.
Exercise 7.2. A useful tool for proving isomorphisms. Let (D, φ) = (A, φA ) ∗
(B, φB ) be a reduced free product of C∗ -algebras and suppose there is a central
projection p ∈ A. (For sake of clarity, assume φA and φB are faithful or even
traces, if you like.) Take the subalgebra A1 = Cp + (1 − p)A of A and let D1 be
the C∗ -subalgebra of D generated by A1 ∪ B. (Thus, D1 is the corresponding
reduced free product of A1 and B.) Show that pDp is generated by pD1 p and
pA, and that these two algebras are free in the C∗ -n.c.p.s. (pDp, φ(p)−1 φ|pDp ).
The next exercise is based partly on classical knowledge about two projec-
tions (or two subspaces) in Hilbert space that has been thoroughly explored in
the literature; see [13] and [25], for example.
Exercise 7.3. Let A be a unital C∗ -algebra that is generated by projections
p and q.
(a) The element 1 − p − q + pq + qp lies in the center of A.
(b) Every irreducible representation of A (namely, a ∗ -representation whose
image has no nontrivial reducing subspaces) must be one- or two-dimen-
sional.
(c) Every irreducible unital representation is unitarily equivalent to one of
these:
(1) p 7→ 0, q 7→ 0 in C,
(2) p 7→ 0, q 7→ 1 in C,
(3) p 7→ 1, q 7→ 0 in C,
(4) p 7→ 1, q 7→ 1 in C, √
t t(1−t)
(5) p 7→ 10 00 , q 7→ √ in M2 (C), for some t ∈ (0, 1).
t(1−t) 1−t
Free group factors 71
(d) Let
A = {f : [0, 1] → M2 (C) | f continuous, f (0) and f (1) diagonal}
with P, Q ∈ A given by the functions
√
t t(1−t)
P ∼ 10 00 , Q ∼ √ .
t(1−t) 1−t
References
[1] D. Avitzour, Free products of C ∗ -algebras, Trans. Amer. Math. Soc. 271 (1982), no. 2,
423–435. MR0654842
[2] H. Bercovici and D. Voiculescu, Lévy-Hinčin type theorems for multiplicative and ad-
ditive free convolution, Pacific J. Math. 153 (1992), no. 2, 217–248. MR1151559
[3] P. Biane, M. Capitaine, and A. Guionnet, Large deviation bounds for matrix Brownian
motion, Invent. Math. 152 (2003), no. 2, 433–459. MR1975007
[4] E. F. Blanchard and K. J. Dykema, Embeddings of reduced free products of operator
algebras, Pacific J. Math. 199 (2001), no. 1, 1–19. MR1847144
[5] K. Dykema, Free products of hyperfinite von Neumann algebras and free dimension,
Duke Math. J. 69 (1993), no. 1, 97–119. MR1201693
[6] K. Dykema, On certain free product factors via an extended matrix model, J. Funct.
Anal. 112 (1993), no. 1, 31–60. MR1207936
[7] K. Dykema, Interpolated free group factors, Pacific J. Math. 163 (1994), no. 1, 123–135.
MR1256179
[8] K. J. Dykema, Faithfulness of free product states, J. Funct. Anal. 154 (1998), no. 2,
323–329. MR1612705
[9] K. J. Dykema, Simplicity and the stable rank of some free product C ∗ -algebras, Trans.
Amer. Math. Soc. 351 (1999), no. 1, 1–40. MR1473439
[10] K. J. Dykema and M. Rørdam, Projections in free product C ∗ -algebras, Geom. Funct.
Anal. 8 (1998), no. 1, 1–16. MR1601917
Erratum in Geom. Funct. Anal. 10 (2000), no. 4, 975. MR1791146
[11] L. Ge, Applications of free entropy to finite von Neumann algebras. II, Ann. of Math.
(2) 147 (1998), no. 1, 143–157. MR1609522
[12] L. M. Ge, On “Problems on von Neumann algebras by R. Kadison, 1967”, Acta Math.
Sin. (Engl. Ser.) 19 (2003), no. 3, 619–624. MR2014042
[13] P. R. Halmos, Two subspaces, Trans. Amer. Math. Soc. 144 (1969), 381–389.
MR0251519
[14] N. A. Ivanov, On the structure of some reduced amalgamated free product C ∗ -algebras,
Internat. J. Math. 22 (2011), no. 2, 281–306. MR2782689
[15] K. Jung, The free entropy dimension of hyperfinite von Neumann algebras, Trans. Amer.
Math. Soc. 355 (2003), no. 12, 5053–5089 (electronic). MR1997595
72 K. Dykema
[16] F. J. Murray and J. von Neumann, On rings of operators. IV, Ann. of Math. (2) 44
(1943), 716–808. MR0009096
[17] N. Ozawa, Solid von Neumann algebras, Acta Math. 192 (2004), no. 1, 111–117.
MR2079600
[18] N. Ozawa, A Kurosh-type theorem for type II1 factors, Int. Math. Res. Not. 2006, Art.
ID 97560, 21 pp. MR2211141
[19] N. Ozawa and S. Popa, On a class of II1 factors with at most one Cartan subalgebra,
Ann. of Math. (2) 172 (2010), no. 1, 713–749. MR2680430
[20] S. Popa, Deformation and rigidity for group actions and von Neumann algebras, in
International Congress of Mathematicians. Vol. I, 445–477, Eur. Math. Soc., Zürich,
2007. MR2334200
[21] R. T. Powers, Simplicity of the C ∗ -algebra associated with the free group on two gen-
erators, Duke Math. J. 42 (1975), 151–156. MR0374334
[22] F. Rădulescu, The fundamental group of the von Neumann algebra of a free group with
infinitely many generators is R+ \ 0, J. Amer. Math. Soc. 5 (1992), no. 3, 517–532.
MR1142260
[23] F. Rădulescu, Stable equivalence of the weak closures of free groups convolution alge-
bras, Comm. Math. Phys. 156 (1993), no. 1, 17–36. MR1234103
[24] F. Rădulescu, Random matrices, amalgamated free products and subfactors of the von
Neumann algebra of a free group, of noninteger index, Invent. Math. 115 (1994), no. 2,
347–389. MR1258909
[25] I. Raeburn and A. M. Sinclair, The C ∗ -algebra generated by two projections, Math.
Scand. 65 (1989), no. 2, 278–290. MR1050869
[26] D. Voiculescu, Symmetries of some reduced free product C ∗ -algebras, in Operator alge-
bras and their connections with topology and ergodic theory (Buşteni, 1983), 556–588,
Lecture Notes in Math., 1132, Springer, Berlin, 1985. MR0799593
[27] D. Voiculescu, Circular and semicircular systems and free product factors, in Opera-
tor algebras, unitary representations, enveloping algebras, and invariant theory (Paris,
1989), 45–60, Progr. Math., 92, Birkhäuser, Boston, MA, 1990. MR1103585
[28] D. Voiculescu, Limit laws for random matrices and free products, Invent. Math. 104
(1991), no. 1, 201–220. MR1094052
[29] D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free
probability theory. I, Comm. Math. Phys. 155 (1993), no. 1, 71–92. MR1228526
[30] D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free
probability theory. II, Invent. Math. 118 (1994), no. 3, 411–440. MR1296352
[31] D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free
probability theory. III. The absence of Cartan subalgebras, Geom. Funct. Anal. 6 (1996),
no. 1, 172–199. MR1371236
[32] D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free
probability theory. IV. Maximum entropy and freeness, in Free probability theory (Wa-
terloo, ON, 1995), 293–302, Fields Inst. Commun., 12, Amer. Math. Soc., Providence,
RI, 1997. MR1426847
[33] D. Voiculescu, A strengthened asymptotic freeness result for random matrices with ap-
plications to free entropy, Internat. Math. Res. Notices 1998, no. 1, 41–63. MR1601878
[34] D. Voiculescu, The analogues of entropy and of Fisher’s information measure in free
probability theory. V. Noncommutative Hilbert transforms, Invent. Math. 132 (1998),
no. 1, 189–227. MR1618636
[35] D. Voiculescu, Free entropy, Bull. London Math. Soc. 34 (2002), no. 3, 257–278.
MR1887698
Free convolution
Hari Bercovici
1. Introduction
It may have seemed at the inception of free probability theory that the
connection with actual probability theory is somewhat tenuous. Subsequent
developments have shown that such connections exist and are quite deep. For
instance, the asymptotics (as the size of the matrices tends to infinity) of
eigenvalue distributions for sums of large random matrices can often be studied
by considering sums of freely independent random variables [53]. Wigner’s
semicircle law can be viewed as a manifestation of the central limit theorem
in free probability. In these notes we discuss the basic apparatus for studying
sums of freely independent random variables and the free counterparts of the
classical limit theorems of probability theory. We focus especially on the case of
unbounded random variables. Products of freely independent random variables
are discussed briefly. We use some of the material presented in earlier lectures,
particularly the existence of free cumulants as described by Roland Speicher.
The author was supported in part by a grant from the National Science Foundation.
74 H. Bercovici
(1) If X1 has finite expected value (that is, X1 is integrable) then the averages
X1 + X2 + · · · + Xn
n
converge weakly to EX1 . This is the weak law of large numbers.
(2) If EX1 = 0 and EX12 = 1, the central limit theorem states that
X1 + X2 + · · · + Xn
√
n
converges weakly to the standard Gaussian distribution N (0, 1).
(3) Suppose that the random variables√{Xnm | m ≤ n} are independent, and
P[Xnm = 1] = 1 − P[Xnm = 0] = 1/ n. In this case the sum Xn1 + Xn2 +
· · · + Xnn converges weakly to a Poisson distribution.
The proofs of these results are done easily using Fourier analysis. Under suit-
able assumptions on the random variables, a proof may be obtained using
classical cumulants.
Voiculescu [51] observed that an analog of (2) holds for freely independent
random variables. The limit distribution is semicircular in this case. There-
fore, it is natural to ask whether the other classical limit theorems have such
counterparts in free probability theory.
2
(2) Assume now √ that ϕ(x1 ) = 0 and ϕ(x1 ) = 1. An analogous calculation
shows that kyk converges in moments to the centered semicircular law γ
of variance one. This is an absolutely continuous distribution on the real
line, with density
dγ 1 p
= 4 − t2 1[−2,2] .
dt 2π
The cumulants of this law are equal to zero except for κ2 (γ) = 1. The
calculation done above yields immediately
√ k
κn ( kyk ) = n/2 κn (x1 ),
k
and one sees again that these cumulants converge to κn (γ).
Now, we know that general distributions on the real line are not uniquely de-
termined by their moments and, indeed, they may fail to have any moments at
all. In addition, convergence in moments is not equivalent to weak convergence
of probability distributions, even when moments of all orders exist. It is true
however that convergence in moments can be promoted to weak convergence
for many limit distributions, including for instance point masses and semicircle
laws.
It can be shown [50, Chap. 9, §2] that Ae has an algebra structure such that
A is a subalgebra provided that, for instance, A is closed in the weak operator
topology and ϕ(x∗ x) > 0 for every x ∈ A\ {0}. In other words, we can think of
A as sitting inside Ae if A is a von Neumann algebra and ϕ is a faithful tracial
state. Observe that, given a selfadjoint x ∈ Ae and a bounded Borel function
u on R, we have
Z
u(t)µx (dt) = ϕ(u(x)) = hu(x)ξ, ξi.
R
Definition 4.2. Two real, unbounded variables x, y are said to be freely in-
dependent if the sets {Ex (σ) | σ Borel} and {Ey (σ) | σ Borel} are freely inde-
pendent in (A, ϕ).
As seen in Ken Dykema’s lectures, the construction of free products shows
that there exist noncommutative probability spaces (A, ϕ) which contain an
infinite sequence (An )∞
n=1 of freely independent algebras, each of which is iso-
morphic to L∞ (0, 1) (with the usual expectation). Unbounded random vari-
ables associated to (An , ϕ|An ) can also be viewed as random variables in A, e
and thus it is possible to construct unbounded, freely independent random
variables. In particular, given two probability distributions µ, ν on R, there
exist free random variables x and y such that µx = µ and µy = ν. Note that in
the context of classical probability, independent random variables can be con-
structed using products of probability spaces. This amounts to considering the
tensor products of the corresponding algebras of bounded random variables.
A free counterpart of classical convolution can be constructed because of
the following result (see [52, 41, 24] or the book [62]).
Proposition 4.3. If x, y are freely independent selfadjoint random variables,
then µx+y depends only on µx and µy .
Proof. Let us assume first that x and y are bounded. We use the notation
µ ⊞ ν for the distribution of x + y when x and y are free, bounded, µx = µ,
and µy = ν. The proposition, and the existence of µ ⊞ ν in that case, is
an immediate consequence of moment calculations. Indeed, we know that
κn (x + y) = κn (x) + κn (y) for all n. This uniquely determines the moments
of x + y with respect to ϕ, and hence the measure µx+y itself because it has
compact support.
The case of unbounded random variables can be treated using a continuity
property of the operation ⊞. Recall that for compactly supported probability
distributions µ and µ′ , the Lévy metric d is given by the infimum of all ε > 0
such that
µ((−∞, t − ε)) − ε ≤ µ′ ((−∞, t)) ≤ µ((−∞, t + ε)) + ε
for all t ∈ R. We will make use of the fact (see [24]) that for compactly
supported probability distributions µ, µ′ , ν and ν ′ , we have
d(µ ⊞ ν, µ′ ⊞ ν ′ ) ≤ d(µ, µ′ ) + d(ν, ν ′ ).
Free convolution 77
But we will stick to the Lévy metric because it defines the topology of weak
convergence on probability distributions.
The uniform continuity of µ ⊞ ν in both variables, along with the fact that
every probability distribution can be approximated weakly by compactly sup-
ported ones, implies immediately that the operation ⊞ of free additive convo-
lution extends by continuity to arbitrary probability distributions.
We argue now that the equality µx+y = µx ⊞ µy persists for unbounded
random variables x and y. This will conclude the proof of the proposition. To
do this, set xn = Ex ([−n, n])x and yn = Ey ([−n, n])y for n ∈ N, so that xn
and yn are bounded, freely independent, and µxn → µx , µyn → µy weakly. We
have µxn +yn = µxn ⊞ µyn → µx ⊞ µy weakly, and the desired conclusion follows
because µxn +yn → µx+y weakly. The last assertion follows from the fact that
xn +yn coincides with x+y on the common range of the projections Ex ([−n, n])
and Ey ([−n, n]), and the projection Pn = Ex ([−n, n]) ∧ Ey ([−n, n]) satisfies
ϕ(Pn ) → 1 as n → ∞.
for every ε > 0. The question is then: Under what conditions do the measures
νn = µn1 ⊞ µn2 ⊞ · · · ⊞ µnkn converge weakly to a probability measure ν? For
simplicity, we restrict ourselves to the special case where the measures in each
row are identical: µnm = µn does not depend on m. This certainly covers
the analogs of the theorems described in Section 2. In this case νn = µ⊞k n
n
is the kn -th convolution power of µn , and one discards the trivial case when
(kn )n is a bounded sequence, in which case νn → δ0 as n → ∞ (note that
δ0 ⊞ δ0 = δ0 ). The convergence of νn can then be studied using Lemma 4.9.
Thus we need to study the limiting behavior of the functions kn ϕµn , according
to Proposition 4.7. By the remarks at the end of the preceding section, we can
equivalently analyze the limiting behavior of the functions kn (λ − Fµn (λ)):
1
kn (λ − Fµn (λ)) = kn λ −
Gµn (λ)
kn λ 1
= Gµn (λ) −
Gµn (λ) λ
1
∼ kn λ2 Gµn (λ) −
Z h λ
2 1 1i
= kn λ − dµn (t)
λ−t λ
ZR
λt
= −kn dµn (t)
R t−λ
Z
1 + λt
= βn − dσn (t),
R t−λ
82 H. Bercovici
where
Z
t t2
βn = kn 2
dµn (t) and dσn (t) = kn dµn (t),
R 1+t 1 + t2
and the approximation in the third line is uniform in n as λ → ∞ nontangen-
tially, building on Proposition 4.6. These considerations lead to the following
result, see [22].
Theorem 5.1. The following assertions are equivalent.
(1) The sequence (µ⊞k
n
n
)n converges weakly to a probability measure.
(2) The numerical sequence (βn )n converges, and the measures (σn )n have a
weak limit as n → ∞.
Remarkably, Gnedenko and Kolmogorov proved earlier that the conditions
in (2) are also equivalent to the weak convergence of the sequence µ∗k n , where
n
µ∗k
n
n
denotes the k n -fold classical convolution of µ n with itself. The above
arguments can in fact be refined to yield an even more general result, see [31].
Theorem 5.2. Consider an infinitesimal array {Xnm | n ∈ N, 1 ≤ m ≤ kn } of
classically independent random variables and an array {Ynm | n ∈ n, 1 ≤ m ≤
kn } of freely independent variables such that µXnm = µYnm for all n, m. Let
(cn )∞
n=1 be a sequence of real numbers. The following assertions are equivalent.
(1) The sequence cn + Xn1 + Xn2 + · · · + Xnkn converges in distribution.
(2) The sequence cn + Yn1 + Yn2 + · · · + Ynkn converges in distribution.
Thus the entire body of classical limit theorems transfers to the free context
simply by replacing classical independence by free independence. The methods
of proof in the free case have a vague family resemblance with the classical ones,
but they are quite different. One can hope that a common proof for these two
families of limit theorems will be found, but our optimism should be tempered,
for instance by the fact that the multiplicative analog of Theorem 5.2 is not
quite correct (but almost correct)—in fact, there are more limit theorems in
the free world [23].
8. Subordination
The study of limit laws only required understanding the behavior of the
functions Gµ and Fµ at infinity. To study finer analytic properties of free
convolutions of probability measures we need to understand the behavior of
Gµ near the real line. This requires the finer tool of analytic subordination.
Consider two analytic functions g1 , g2 : C+ → C. We say that g2 is sub-
ordinate to g1 if there exists an analytic function ω : C+ → C+ such that
∡ limλ→∞ ω(λ)/λ = 1 and g2 (λ) = g1 (ω(λ)) for all λ ∈ C+ . The following
Free convolution 85
result was proved under various forms and degrees of generality by Voiculescu,
Maassen, and Biane.
Theorem 8.1. Let µ, ν be two probability measures on R. Then Gµ⊞ν is
subordinate to both Gµ and Gν .
Proof. We need to prove that there exist analytic functions ω1 and ω2 such
that Fµ⊞ν = Fµ ◦ ω1 = Fν ◦ ω2 . We can define
ω1 = Fµh−1i ◦ Fµ⊞ν and ω2 = Fνh−1i ◦ Fµ⊞ν
in some domain of the form Λα,β . This yields functions with the desired
behavior at infinity, so the issue is whether the functions thus defined continue
analytically to selfmaps of C+ . Observe first that the equation ϕµ⊞ν (λ) =
h−1i
ϕµ (λ) + ϕν (λ) is equivalent to Fµ⊞ν (λ) = Fµh−1i (λ) + Fνh−1i (λ) − λ. Replacing
λ by Fµ⊞ν (λ), we obtain
λ = Fµh−1i (Fµ⊞ν (λ)) + Fνh−1i (Fµ⊞ν (λ)) − Fµ⊞ν (λ)
= ω1 (λ) + ω2 (λ) − Fµ⊞ν (λ).
Of course, this identity only holds in some domain Λα,β where all the functions
are already defined. Note however that once the functions ωj are continued
analytically to C+ , this identity will extend to the entire upper half-plane. This
suggests the following approach. Fix λ ∈ C+ and search for points ωλ,1 , ωλ,2
such that
Fµ⊞ν (λ) = Fµ (ωλ,1 ) = Fν (ωλ,2 )
and
ωλ,1 + ωλ,2 = λ + Fµ⊞ν (λ).
We already know that such points exist if λ belongs to some Λα,β . Next, we
eliminate ωλ,2 :
ωλ,1 = λ + Fµ⊞ν (λ) − ωλ,2
= λ + Fν (ωλ,2 ) − ωλ,2
= λ + Fν (λ + Fµ⊞ν (λ) − ωλ,1 ) − (λ + Fµ⊞ν (λ) − ωλ,1 ).
Setting
Hλ (ω) = λ + Fν (λ + Fµ⊞ν (λ) − ω) − (λ + Fµ⊞ν (λ) − ω),
we see that Hλ (ωλ,1 ) = ωλ,1 . Now, Hλ is easily seen to be an analytic selfmap
of C+ . Moreover, Im Hλ (w) ≥ Im λ > 0 holds. Indeed, using the application
of Nevanlinna’s result presented right after Proposition 4.10, we may rewrite
Fν as Z
1 + zt
Fν (z) = z + β + dρ(t), z ∈ C+ ,
R t−z
where β ∈ R and ρ is a finite positive Borel measure on R. Then
1 + zt 1 + t2
Im = Im z > 0
t−z |t − z|2
86 H. Bercovici
proves the claim. Now Im Hλ (ω) ≥ Im λ > 0 implies that Hλ is not a conformal
automorphism. Thus, Hλ has a fixed point for λ in some open set of the form
Λα,β . Therefore, the iterates of Hλ applied to the imaginary unit i (or any
other point of the upper half-plane) converge to that fixed point. Denote by
Hλ◦n the n-th iterate of Hλ , and set
Fn (λ) = Hλ◦n (i).
Then (Fn )∞ +
n=1 is a normal family of analytic selfmaps of C , and limn→∞ Fn (λ)
exists for λ in some open set of the form Λα,β . Montel’s theorem shows that
this limit exists for all λ ∈ C+ . Moreover, it is analytic and it extends ω1 .
The extension of ω2 is proved analogously or by using the relation between ω1
and ω2 .
Interestingly, in case µ = ν, we obtain a result in complex analysis which
had not been observed before its appearance in this context, see [13, 29].
Theorem 8.2. Let u : C+ → C be an analytic function such that
(1) limt↑+∞ u(it)
it = 1 and
(2) Im u(z) ≤ Im z for all z ∈ C+ .
Then there exists a map ω : C+ → C+ such that u(ω(λ)) = λ for all λ ∈ C+ .
Proof. Consider a probability measure µ on R, and consider the subordination
function ω satisfying Fµ⊞µ = Fµ ◦ ω. We have then 2ω(λ) = λ + Fµ⊞µ (λ), so
that λ = 2ω(λ) − Fµ (ω(λ)) = u(ω(λ)), where u(λ) = 2λ − Fµ (λ). To conclude
the proof, we only need to observe that conditions (1) and (2) imply that the
function u is of the form u(λ) = 2λ−Fµ (λ) for some probability measure µ.
We list a few consequences of the subordination result (Theorem 8.1).
(1) If µ is absolutely continuous relative to the Lebesgue measure, and its
density belongs to Lp for some p, then the same is true for µ ⊞ ν. This is
similar to results in classical probability. Note, however, that it is possible
that µ has an infinitely differentiable density on R while the density of
µ ⊞ ν is not everywhere differentiable.
(2) We have
(µ ⊞ ν)({s}) = max 0, max(µ({t}) + ν({s − t}) − 1) .
t∈R
Exercise 9.2. Prove the converse of the above statement. Use the statement
and its converse, along with a free product construction, to prove that ⊞ is
Lipschitz in both variables with respect to the Lévy metric d.
Consider an element x in a noncommutative probability space (A, ϕ), denote
by αn = ϕ(xn ) its moments, and let κn = κn (x) be the free cumulants. The
formula relating Rx and the Cauchy transform Gx can be written using the
generating series
X∞ ∞
X
M (λ) = 1 + αn λn , C(λ) = 1 + κn λn
n=1 n=1
associated to these sequences.
Exercise 9.3. Show that the relation Gx (Rx (λ) + 1/λ) = λ is equivalent to
C(λM (λ)) = M (λ). Deduce this last equality from the combinatorial formulas
relating αn and κn .
It is generally difficult to find explicit formulas for a measure for which
Rµ (λ) is known. Here are a couple of cases where this is possible.
Exercise 9.4. Use Stieltjes inversion in order to find the measures γ and ρ
for which Rγ (λ) = λ and Rρ (λ) = −i.
The relation Rµ⊞ν = Rµ + Rν gives a method for calculating µ ⊞ ν: First
calculate Rµ and Rν by inverting the Cauchy transforms, then calculate Gµ⊞ν
by inverting the function (1/λ) + Rµ (λ) + Rν (λ), and finally find µ ⊞ ν via
Stieltjes inversion.
Exercise 9.5. Let µ be defined by µ({0}) = ε ∈ (0, 1) and µ({1}) = 1 − ε.
Calculate explicitly the measure µ ⊞ µ. (You should obtain an absolutely
continuous measure except for one atom in case ε 6= 1/2.)
Herglotz proved that an analytic function f with positive real part in the
unit disk can be represented as
Z 2π it
e +λ
f (λ) = i Im f (0) + dµ(t), |λ| < 1,
0 eit − λ
for some finite, positive Borel measure on [0, 2π]. The argument consists in
writing the Poisson formula for Re f (rλ), where r ∈ (0, 1), and using the weak
compactness of the set of probability measures on a compact interval.
Exercise 9.6. Use a conformal map to deduce the Nevanlinna representation
from this result of Herglotz, compare Proposition 4.10.
Theorem 5.1 was proved in [22]. Special cases were considered earlier; see
for instance [45] for the general free central limit theorem. The more general
Theorem 5.2 is from [31]. See also [23, 27, 32] for counterparts of these results
in the context of free multiplicative convolution.
The proof of Theorem 5.1 yields some information about the limiting prob-
ability measure: Its ϕ transform is defined in the entire upper half-plane, and
Free convolution 89
was proved in full generality in [29]. The most conceptual approach is based on
coalgebras, see [58]. A direct analytic approach, inspired by the earlier work
in [41], is given in [33]. The argument we give in Section 8 is taken from [13].
The following exercise justifies the statement concerning fixed points in the
proof of Theorem 8.1 since C+ is conformally equivalent to D.
Exercise 9.12. Assume that f : D → D is analytic, f (0) = 0, but f is not a
bijection. Show that for every r ∈ (0, 1) there exists a constant c ∈ (0, 1) such
that |f (z)| ≤ c|z| for |z| ≤ r. Deduce that f ◦n (z) → 0 as n → ∞ for every
z ∈ D.
Subordination also holds for free multiplicative convolution. For measures
on T one can use the following result.
Exercise 9.13. Let f : D → C be an analytic function such that f (0) = 0
and |f (λ)| ≥ |λ| for all λ ∈ D. Show that there exists an analytic function
g : D → D such that f (g(λ)) = λ for all λ ∈ D.
The first regularity results using subordination, for instance item (1) in Sec-
tion 8, appeared in [54]. Details on the absence of a singular unitary part for
µ ⊞ ν can be found in [10]. The atoms of µ ⊞ ν were identified in [26]. The ex-
istence of µ⊞t for large t was proved in [25] when µ has compact support. The
extension to t ≥ 1 is taken from [42] for compact supports and from [12] for
arbitrary measures. The fact that connected components in the support of con-
volution powers tend to coalesce was shown in [36, 37, 66] in various contexts.
The fact that free convolutions have no “consecutive” atoms is explained in [28];
see also [33] for related results on freely indecomposable measures. We have not
mentioned many other interesting results, some of which are in the references;
see [3, 4, 5, 6, 11, 14, 17, 18, 19, 20, 21, 34, 38, 39, 40, 44, 47, 48, 59, 61, 63].
Exercise 9.14. Assume that µ({1}) = µ({−1}) = 1/2. Calculate the mea-
sures µ⊞t for all t ≥ 1. (Observe that Rµ⊞t = tRµ . The measure µ⊞t has two
atoms when t < 2.)
References
[1] N. I. Akhiezer, The classical moment problem and some related questions in analysis,
Translated by N. Kemmer, Hafner Publishing Co., New York, 1965. MR0184042
[2] M. Anshelevich, S. T. Belinschi, M. Février, and A. Nica, Convolution powers in the
operator-valued framework, Trans. Amer. Math. Soc. 365 (2013), no. 4, 2063–2097.
MR3009653
[3] M. Anshelevich, J.-C. Wang, and P. Zhong, Local limit theorems for multiplicative free
convolutions, J. Funct. Anal. 267 (2014), no. 9, 3469–3499. MR3261117
[4] O. Arizmendi and S. T. Belinschi, Free infinite divisibility for ultrasphericals, Infin. Di-
mens. Anal. Quantum Probab. Relat. Top. 16 (2013), no. 1, 1350001, 11 pp. MR3071453
[5] O. Arizmendi and T. Hasebe, Semigroups related to additive and multiplicative, free
and Boolean convolutions, Studia Math. 215 (2013), no. 2, 157–185. MR3071490
[6] O. Arizmendi and V. Pérez-Abreu, The S-transform of symmetric probability mea-
sures with unbounded supports, Proc. Amer. Math. Soc. 137 (2009), no. 9, 3057–3066.
MR2506464
Free convolution 91
[7] O. E. Barndorff-Nielsen and S. Thorbjørnsen, Lévy laws in free probability, Proc. Natl.
Acad. Sci. USA 99 (2002), no. 26, 16568–16575. MR1947756
[8] O. E. Barndorff-Nielsen and S. Thorbjørnsen, Self-decomposability and Lévy processes
in free probability, Bernoulli 8 (2002), no. 3, 323–366. MR1913111
[9] O. E. Barndorff-Nielsen and S. Thorbjørnsen, A connection between free and classical
infinite divisibility, Infin. Dimens. Anal. Quantum Probab. Relat. Top. 7 (2004), no. 4,
573–590. MR2105912
[10] S. T. Belinschi, The Lebesgue decomposition of the free additive convolution of two
probability distributions, Probab. Theory Related Fields 142 (2008), no. 1-2, 125–150.
MR2413268
[11] S. T. Belinschi, F. Benaych-Georges, and A. Guionnet, Regularization by free additive
convolution, square and rectangular cases, Complex Anal. Oper. Theory 3 (2009), no. 3,
611–660. MR2551632
[12] S. T. Belinschi and H. Bercovici, Partially defined semigroups relative to multiplicative
free convolution, Int. Math. Res. Not. 2005, no. 2, 65–101. MR2128863
[13] S. T. Belinschi and H. Bercovici, A new approach to subordinationresults in free prob-
ability, J. Anal. Math. 101 (2007), 357–365. MR2346550
[14] S. T. Belinschi, M. Bożejko, F. Lehner, and R. Speicher, The normal distribution is
⊞-infinitely divisible, Adv. Math. 226 (2011), no. 4, 3677–3698. MR2764902
[15] S. T. Belinschi, M. Popa, and V. Vinnikov, Infinite divisibility and a non-commutative
Boolean-to-free Bercovici-Pata bijection, J. Funct. Anal. 262 (2012), no. 1, 94–123.
MR2852257
[16] S. T. Belinschi, M. Popa, and V. Vinnikov, On the operator-valued analogues of the
semicircle, arcsine and Bernoulli laws, J. Operator Theory 70 (2013), no. 1, 239–258.
MR3085826
[17] F. Benaych-Georges, Failure of the Raikov theorem for free random variables, in
Séminaire de Probabilités XXXVIII, 313–319, Lecture Notes in Math., 1857, Springer,
Berlin, 2005. MR2126982
[18] F. Benaych-Georges, Infinitely divisible distributions for rectangular free convolution:
classification and matricial interpretation, Probab. Theory Related Fields 139 (2007),
no. 1-2, 143–189. MR2322694
[19] F. Benaych-Georges, Rectangular random matrices, related convolution, Probab. The-
ory Related Fields 144 (2009), no. 3-4, 471–515. MR2496440
[20] F. Benaych-Georges, Rectangular R-transform as the limit of rectangular spherical in-
tegrals, J. Theoret. Probab. 24 (2011), no. 4, 969–987. MR2851240
[21] F. Benaych-Georges and T. Cabanal-Duvillard, Marčenko-Pastur theorem and
Bercovici-Pata bijections for heavy-tailed or localized vectors, ALEA Lat. Am. J.
Probab. Math. Stat. 9 (2012), no. 2, 685–715. MR3069381
[22] H. Bercovici and V. Pata, Stable laws and domains of attraction in free probability
theory, Ann. of Math. (2) 149 (1999), no. 3, 1023–1060. MR1709310
[23] H. Bercovici and V. Pata, Limit laws for products of free and independent random
variables, Studia Math. 141 (2000), no. 1, 43–52. MR1782911
[24] H. Bercovici and D. Voiculescu, Free convolution of measures with unbounded support,
Indiana Univ. Math. J. 42 (1993), no. 3, 733–773. MR1254116
[25] H. Bercovici and D. Voiculescu, Superconvergence to the central limit and failure of the
Cramér theorem for free random variables, Probab. Theory Related Fields 103 (1995),
no. 2, 215–222. MR1355057
[26] H. Bercovici and D. Voiculescu, Regularity questions for free convolution. Nonselfadjoint
operator algebras, operator theory, and related topics, 37–47, Oper. Theory Adv. Appl.,
104, Birkhäuser, Basel, 1998. MR1639647
[27] H. Bercovici and J.-C. Wang, Limit theorems for free multiplicative convolutions, Trans.
Amer. Math. Soc. 360 (2008), no. 11, 6089–6102. MR2425704
92 H. Bercovici
[28] H. Bercovici and J.-C. Wang, On freely indecomposable measures, Indiana Univ. Math.
J. 57 (2008), no. 6, 2601–2610. MR2482992
[29] P. Biane, Processes with free increments, Math. Z. 227 (1998), no. 1, 143–174.
MR1605393
[30] P. Biane, Free probability for probabilists, in Quantum probability communications,
Vol. XI (Grenoble, 1998), 55–71, QP-PQ, XI, World Sci. Publ., River Edge, NJ, 2003.
MR2032363
[31] G. P. Chistyakov and F. Götze, Limit theorems in free probability theory. I, Ann.
Probab. 36 (2008), no. 1, 54–90. MR2370598
[32] G. P. Chistyakov and F. Götze, Limit theorems in free probability theory. II, Cent. Eur.
J. Math. 6 (2008), no. 1, 87–117. MR2379953
[33] G. P. Chistyakov and F. Götze, The arithmetic of distributions in free probability theory,
Cent. Eur. J. Math. 9 (2011), no. 5, 997–1050. MR2824443
[34] G. P. Chistyakov and F. Götze, Asymptotic expansions in the CLT in free probability,
Probab. Theory Related Fields 157 (2013), no. 1-2, 107–156. MR3101842
[35] B. V. Gnedenko and A. N. Kolmogorov, Limit distributions for sums of independent
random variables. Translated from the Russian, annotated, and revised by K. L. Chung.
With appendices by J. L. Doob and P. L. Hsu. Revised edition. Addison-Wesley Pub-
lishing Co., Reading, MA, 1968. MR0233400
[36] H.-W. Huang, Supports, regularity, and ⊞-infinitedivisibility for measures of the form
(µ⊞p )⊎q , arXiv:1209.5787 [math.CV] (2012).
[37] H.-W. Huang and P. Zhong, On the supports of measures in free multiplicative convo-
lution semigroups, Math. Z. 278 (2014), no. 1-2, 321–345. MR3267581
[38] V. Kargin, Berry-Esseen for free random variables, J. Theoret. Probab. 20 (2007), no. 2,
381–395. MR2324538
[39] V. Kargin, On superconvergence of sums of free random variables, Ann. Probab. 35
(2007), no. 5, 1931–1949. MR2349579
[40] V. Kargin, An inequality for the distance between densities of free convolutions, Ann.
Probab. 41 (2013), no. 5, 3241–3260. MR3127881
[41] H. Maassen, Addition of freely independent random variables, J. Funct. Anal. 106
(1992), no. 2, 409–438. MR1165862
[42] A. Nica and R. Speicher, On the multiplication of free N -tuples of noncommutative
random variables, Amer. J. Math. 118 (1996), no. 4, 799–837. MR1400060
[43] A. Nica and R. Speicher, Lectures on the combinatorics of free probability, London Math.
Soc. Lecture Note Ser., 335, Cambridge Univ. Press, Cambridge, 2006. MR2266879
[44] V. Pata, Lévy type characterization of stable laws for free random variables, Trans.
Amer. Math. Soc. 347 (1995), no. 7, 2457–2472. MR1311913
[45] V. Pata, The central limit theorem for free additive convolution, J. Funct. Anal. 140
(1996), no. 2, 359–380. MR1409042
[46] M. Popa and V. Vinnikov, Non-commutative functions and the non-commutative free
Lévy-Hinčin formula, Adv. Math. 236 (2013), 131–157. MR3019719
[47] M. Popa and J.-C. Wang, On multiplicative conditionally free convolution, Trans. Amer.
Math. Soc. 363 (2011), no. 12, 6309–6335. MR2833556
[48] R. Speicher, Free probability theory and random matrices, in Asymptotic combinatorics
with applications to mathematical physics (St. Petersburg, 2001), 53–73, Lecture Notes
in Math., 1815, Springer, Berlin, 2003. MR2009835
[49] M. Takesaki, Theory of operator algebras. I, Reprint of the first (1979) edition. Ency-
clopaedia Math. Sci., 124, Springer, Berlin, 2002. MR1873025
[50] M. Takesaki, Theory of operator algebras. II, Encyclopaedia Math. Sci., 125, Springer,
Berlin, 2003. MR1943006
[51] D. Voiculescu, Symmetries of some reduced free product C ∗ -algebras, in Operator alge-
bras and their connections with topology and ergodic theory (Buşteni, 1983), 556–588,
Lecture Notes in Math., 1132, Springer, Berlin, 1985. MR0799593
Free convolution 93
Moritz Weber
1. Introduction
In the field of noncommutative operator algebras, quantum groups are a
good notion of what “symmetries” should be. In 1987, Woronowicz [72] gave
a definition of compact matrix quantum groups based on the theory of C ∗ -
algebras. They generalize compact Lie groups G ⊂ Mn (C). About twenty
years later, Banica and Speicher [21] isolated a class of compact matrix quan-
tum groups with an intrinsic combinatorial structure. These so-called easy
quantum groups are determined by categories of partitions (via some Tannaka–
Krein type result).
They have been proven useful in order to understand various aspects of
quantum groups, in particular those linked to free probability theory. Further-
more, they open a way to find examples of compact quantum groups apart
from q-deformations and quantum isometry groups. The class of easy quantum
groups contains the symmetric group Sn and the orthogonal group On as well
as Wang’s quantum permutation group Sn+ and his free orthogonal quantum
group On+ .
We begin with an introduction to compact matrix quantum groups, before
we turn to easy quantum groups and their relation to free probability.
for all t. The next theorem is due to Woronowicz [72, 76] with an improvement
by Van Daele [61].
Theorem 2.2. Let G be a compact quantum group. Then there exists a unique
state h on C(G) (the Haar state) such that (h ⊗ id)∆ = (id ⊗h)∆ = h1.
The Haar state is a very useful tool. For instance, the representation theory
of quantum groups very much relies on the existence (and use) of a Haar state.
98 M. Weber
uij ukl = ukl uij yields Wang’s [70] free symmetric quantum group (also called
the quantum permutation group) Sn+ given by
X X
C(Sn+ ) := C ∗ uij , 1 ≤ i, j ≤ n | u∗ij = uij = u2ij , uik = ukj = 1 ∀i, j
k k
P
together with the comultiplication ∆(uij ) = k uik ⊗ ukj .
Example 2.5. For the sake of historical completeness, we should also mention
Woronowicz’s seminal example SUq (2), the first example of a compact quantum
group in his setting. For this, we observe that the group SU (2) ⊂ M2 (C)
consists of unitary matrices of the form ac −c̄
ā . For q ∈ [−1, 1], Woronowicz
[73] defined the quantum group SUq (2) by the universal ∗unital
C ∗ -algebra
a −qc
generated by elements a and c such that the matrix c a∗ is unitary. The
comultiplication is given by
∆(a) = a ⊗ a − qc∗ ⊗ c and ∆(c) = c ⊗ a + a∗ ⊗ c.
This is an example of obtaining a quantum group via deformation, in con-
trast to Wang’s liberations. There exist generalizations to SUq (n), n ∈ N. See
also [58] for details.
The quantum groups On+ , Sn+ and SUq (2) are compact quantum groups
of a special type. The following definition (under the name compact matrix
pseudogroups) was introduced by Woronowicz [72] (see also [75] for an improve-
ment). In fact, the definition of compact matrix quantum groups preceded the
one of compact quantum groups (Definition 2.1) and one can check that the
latter one generalizes the first one [58, Prop. 6.1.4]. Once this is done, it is clear
that the above examples give rise to compact matrix quantum groups—and
hence to compact quantum groups.
Definition 2.6. Given n ∈ N, a compact matrix quantum group consists of a
unital C ∗ -algebra A and a ∗ -homomorphism ∆ : A → A ⊗min A such that
• A is generated by n2 elements uij , 1 ≤ i, j ≤ n in the sense that the
∗
-algebra generated by the uij is dense in A,
• the matrices u = (uij ) and ut = (uji ) are invertible, P
• the ∗ -homomorphism ∆ : A → A ⊗min A sends uij to k uik ⊗ ukj .
Again, we write A = C(G) even if A is noncommutative, and speak of G
as the quantum group. Since the matrix u = (uij ) contains the essential
data of a compact matrix quantum group (in particular, it determines the
comultiplication ∆), it is also common to denote a compact matrix quantum
group by (A, u) or (C(G), u). In the sequel, we will only deal with compact
matrix quantum groups and C(G) will typically be a universal C ∗ -algebra
generated by selfadjoint elements uij such that u is orthogonal (and hence
also ut ) and some further relations are fulfilled. Sometimes, it is convenient to
work with the matrix ū := (u∗ij ) instead of ut = ū∗ .
Compact matrix quantum groups carry a very nice feature: They obey a
Tannaka–Krein type result, proved by Woronowicz [74]. In order to understand
100 M. Weber
it, we again begin with the classical case. Let G be a compact group and let
U : G → B(H) be a unitary representation of G on some finite-dimensional
Hilbert space H. Hence B(H) = Mm (C) for some m ∈ N. Thus, we can view
U as an element in C(G, Mm (C)) = Mm (C) ⊗ C(G), i.e. as a matrix with
entries from
P C(G). Let eij denote the matrix units in Mm (C). We can write
U as U = eij ⊗ Uij .
Passing to quantum groups, a unitary finite-dimensional corepresentation
of a compact quantum group G is by definition P a unitary matrix v ∈ Mm (C) ⊗
C(G) for some m ∈ N such that ∆(vij ) = k vik ⊗ vkj . This reflects exactly
U (gh) = U (g)U (h) in the Pcase of groups. WePcan form the tensor product of
two representations u = euij ⊗ uij and v = evij ⊗ vij by
X
u⊗v = euij ⊗ evkl ⊗ uij vkl ∈ Mmu (C) ⊗ Mmv (C) ⊗ C(G)
∼
= Mm m (C) ⊗ C(G).
u v
3. Categories of partitions
We now want to calculate some intertwiner spaces in an explicit way. For
instance, what are the intertwiner spaces of Sn+ and of On+ ? What are concrete
intertwiners of these quantum groups? Following the work of [21], we shall see
that the theory of categories of partitions provides a useful framework for the
construction of such intertwiners. Moreover, this will lead to the definition of
easy quantum groups.
A partition p ∈ P (k, l) is given by k ∈ N0 upper and l ∈ N0 lower points
which are connected by some strings. This gives rise to a partition of the
ordered set on k + l points, with the additional information which points are
upper and which are lower. A block of p is a maximal subset of connected
points. Here are two examples p1 consisting of four blocks and p2 having five
blocks:
Easy quantum groups 101
• • • • • • • • •
p1 = ✄ p2 =
• • • • • • • • •
A partition is called noncrossing (p ∈ N C(k, l)), if the strings connecting the
points may be drawn in such a way that they do not cross. In the above
example, p1 ∈ P (3, 6) but p1 ∈ / N C(3, 6), whereas p2 ∈ N C(6, 3). The set of
all partitions is denoted by P , the set of all noncrossing partitions is N C.
Now, let e1 , . . . , en be the canonical basis of Cn and let p ∈ P (k, l) be a
partition. We define a linear map Tp : (Cn )⊗k → (Cn )⊗l by
n
X
Tp (ei1 ⊗ · · · ⊗ eik ) = δp (i, j)ej1 ⊗ · · · ⊗ ejl .
j1 ,...,jl =1
2 2 3 2 2 3
✄ δ=1 ✄ δ=0
2 5 3 5 5 2 3 5 3 5 5 2
As an example of such a linear map Tp , consider the partition
p= ✁❆ ∈ P (2, 2).
Here, δp ((i1 , i2 ), (j1 , j2 )) = 1 if and only if i1 = j2 and i2 = j1 . We thus have
the flip map Tp (ei1 ⊗ ei2 ) = ei2 ⊗ ei1 . As a second example, consider the
partition p ∈ P (2, 2) consisting of four points which are all connected. Then
Tp (ei1 ⊗ ei2 ) = δi1 i2 ei1 ⊗ ei1 .
We can define the tensor product p ⊗ q of two partitions p ∈ P (k, l) and
q ∈ P (k ′ , l′ ) by vertical concatenation, i.e. p ⊗ q ∈ P (k + k ′ , l + l′ ) is the
partition obtained by writing p and q side by side. We have Tp ⊗ Tq = Tp⊗q .
This is quite nice since intertwiner spaces are always closed under tensor
products, i.e. whenever we have two intertwiner maps S and T such that
′ ′ ′ ′
Su⊗k = u⊗l S and T u⊗k = u⊗l T , then also (S ⊗ T )u⊗k+k = u⊗l+l (S ⊗ T ).
Thus, we are able to model this operation on intertwiner maps already on the
partition level.
We define the composition of two partitions p ∈ P (k, l) and q ∈ P (l, m)
as the partition qp ∈ P (k, m) obtained from horizontal concatenation, i.e., we
first connect k upper points by p to l middle points, and then by q to m lower
points. The l middle points are removed which yields a partition connecting k
102 M. Weber
upper and m lower points. During this procedure, certain loops may arise—
there may be middle points which are neither connected to upper points nor
to lower points. Connected components of such middle points are called loops
and we denote their number by l(q, p). We have Tq Tp = nl(q,p) Tqp . As an
example, consider the composition of the above partitions p1 and p2 . One loop
appears (points marked in white):
• • •
✄ • • •
p2 p1 = • ◦ • ◦ ◦ • =
• • •
• • •
Note that in general, the intertwiner spaces of compact matrix quantum
groups are closed under composition of (composable) maps.
Finally, we also have an involution p∗ ∈ P (l, k) of a partition p ∈ P (k, l)
obtained by turning p upside down (reflection at the horizontal axis). This
yields (Tp )∗ = Tp∗ . We conclude that the operations on the partitions behave
nicely with the map p 7→ Tp :
Definition 3.1. Let C(k, l) ⊂ P (k, l) be subsets for all k, l ∈ N0 and let C ⊂ P
be the collection of these subsets. We say that C is a category of partitions,
if it is closed under taking tensor products, composition and involution (the
category operations), and if the pair partition ⊓ and the identity partition |
are in C.
We compute
X X
Tp u⊗2 (ei ⊗ ej ) = Tp (ek ⊗ el ) ⊗ uki ulj = el ⊗ ek ⊗ uki ulj ,
k,l k,l
X
u⊗2 Tp (ei ⊗ ej ) = u⊗2 (ej ⊗ ei ) = el ⊗ ek ⊗ ulj uki .
l,k
We infer that Tp ∈ HomG (2, 2) (i.e. Tp u⊗2 = u⊗2 Tp ) if and only if the
generators uij commute. Hence, if Tp is in the intertwiner space of a compact
matrix quantum group G with selfadjoint generators uij , it is actually a group.
One can convince oneself that if G is a quantum subgroup of H (i.e. there is
a surjection C(H) → C(G) mapping generators to generators), then HomG ⊃
HomH . Thus, for any compact matrix quantum group G with Sn ⊂ G ⊂ On+ ,
we have
span{Tp | p ∈ P (k, l)} ⊃ HomG (k, l) ⊃ span{Tp | p ∈ N C2 (k, l)}.
Amongst such quantum groups G there are some with a particularly nice
description of the intertwiner space. These so-called easy quantum groups were
defined by Banica and Speicher in 2009 [21].
Definition 3.3. A compact matrix quantum group Sn ⊂ G ⊂ On+ is called
easy if there is a category of partitions C ⊂ P such that
HomG (k, l) = span{Tp | p ∈ C(k, l)}, k, l ∈ N0 .
Hence, easy quantum groups are determined by the combinatorics of parti-
tions. This is why one could also think of them as “partition quantum groups”.
The structure of the intertwiner space is exactly reflected in the category of
partitions. One of the major philosophies of easy quantum groups is that “all
of the quantum group structure should be visible in the associated category of
partitions”. See Sections 7 and 8 for more on this.
104 M. Weber
Proposition 4.1 ([21, Lem. 2.7]). Categories of partitions are closed under
rotation of partitions.
Natural further examples of easy quantum groups are obtained from the
categories h i and h↑i. The first one gives rise to the hyperoctahedral quan-
tum group Hn+ of Banica, Bichon and Collins [9], whereas the second one is
the bistochastic quantum group Bn+ (see [21]). In [21] and [71], all free easy
quantum groups were classified, i.e. all easy quantum groups whose category
of partitions satisfies C ⊆ N C, see also Exercise 11.3. Besides the above four
examples, there are h↑ ⊗ ↑i, h i and h , ↑ ⊗ ↑i = h , i (for the latter
equality, see [71, Lem. 2.6 (c)]).
Theorem 4.2 ([21, 71]). For n ∈ N, there are exactly seven free easy quantum
groups. They contain Sn+ , On+ , Hn+ and Bn+ .
One should note that for small n ∈ N, certain quantum groups may coincide.
Hence, Theorem 4.2 is rather a statement on sequences (Gn )n∈N of quantum
groups.
A second natural subclass was completely classified in [21]—easy quantum
groups whose categories contain the partition ✁❆ . These categories give rise to
(classical) groups (see Example 3.2).
Theorem 4.3 ([21]). There are exactly six easy quantum groups which are
actually groups. They contain Sn , On , the hyperoctahedral group Hn = Z2 ≀ Sn
and the bistochastic group Bn .
These two theorems are yet another example of the fact that the noncommu-
tative world can be richer than the commutative one: The expected one-to-one
correspondence between classical easy groups and free easy quantum groups
breaks down; one of the groups splits into two quantum groups on the free
side, in the sense that there are two quantum groups which yield the same
classical group, if we add the relations that the generators commute. On the
combinatorial side, this is reflected by the fact that the categories h , ✁❆ i and
h↑ ⊗ ↑, ❆✁ i coincide, whereas their noncrossing counterparts h i and h↑ ⊗ ↑i
do not.
Another subclass of easy quantum groups is obtained by half-liberation, see
[21] for details. A category of partitions is called half-liberated, if it contains
the partition ❅ ❅ but not ❆✁ . The partition ❅❅ ∈ P (3, 3) gives rise to the relations
uij ukl ust = ust ukl uij in the sense of Example 3.2. The corresponding quantum
group is in general neither a classical group nor a free easy quantum group—in
case that the commutativity of the generators cannot be deduced from the
other relations. An example of such a half-liberated quantum group is On∗ ,
given by adding the above relations to those of On+ . The corresponding cate-
∗ +
gory is h ❅❅ i. For n ≥ 4, we have On ( On ( On .
Theorem 4.4 ([71]). The half-liberated easy quantum groups are exactly On∗ ,
(s)
Hn∗ , Bn#∗ and the hyperoctahedral series Hn , s ≥ 3 of [16].
In joint work with Raum [54, 55, 56], the author could completely classify
all easy quantum groups, see also [56] for an overview on the history of the
classification program. We briefly review the main results here.
106 M. Weber
as an equality in C(Sn+ ),
for all m, n ∈ N and all 1 ≤ i1 , . . . , im ≤ n. Note
that it would be more precise to speak of invariance under the coaction of the
sequence (Sn+ )n∈N .
Comparing this equation to the classical case, let u′ij be the coordinate
functions of C(Sn ) and σ ∈ Sn . Then u′ij (σ) = δiσ(j) . Hence, evaluating
Easy quantum groups 109
P
u′j1 i1 . . . u′jm im φ(xj1 . . . xjm ) at σ yields exactly φ(xσ(i1 ) . . . xσ(im ) ). There-
fore, quantum exchangeability implies classical exchangeability.
We are ready for the quantum version of de Finetti’s theorem. Köstler and
Speicher revealed a nice parallelism: While invariance under Sn is equivalent to
classical independence (with amalgamation), invariance under Sn+ is equivalent
to freeness (with amalgamation).
Theorem 5.3 ([47]). A sequence (xi )i∈N of selfadjoint random variables in a
noncommutative W ∗ -probability space (A, φ) is quantum exchangeable if and
only if it is conditionally free and identically distributed over the tail von Neu-
mannT algebra, i.e. with respect to the conditional expectation E : A → T , where
T := n∈N W ∗ (xk | k ≥ n) and φ ◦ E = φ.
Proof. Assume that (xi ) is free and identically distributed with respect to the
tail algebra T . The moment-cumulant formula (see Definition 4.3 of Speicher’s
lecture) can be extended to the operator-valued setting. Using it, together with
φ ◦ E = φ, we infer
Xn
uj1 i1 . . . ujm im φ(xj1 . . . xjm )
j1 ,...,jm =1
X X
= uj1 i1 . . . ujm im φ κE
π (xj1 , . . . , xjm ) .
π∈N C(0,m) j1 ,...,jm
Finally, the intertwiner map Tπ : (Cn )⊗m → C for π ∈ N C(m, 0) may simply
⊗m
P as Tπ (ei1 ⊗ · · · ⊗ eim ) = δker i≥π . Thus, the equality Tπ u
be written = Tπ
yields ker j≥π uj1 i1 . . . ujm im = δker i≥π in C(Sn+ ) and thus
X X
uj1 i1 . . . ujm im φ(xj1 . . . xjm ) = φ δker i≥π κE
π .
j1 ,...,jm π∈N C
Using again the identical distribution, freeness, the moment cumulant formula
and φ ◦ E = φ (in this order), we infer that (xi ) is quantum exchangeable.
The converse direction is a bit more elaborate. One first has to show that
such a conditional expectation E exists once the sequence is quantum ex-
changeable, and that quantum exchangeability with respect to φ is equiva-
lent to quantum exchangeability with respect to E (defined analogously as for
110 M. Weber
6. Laws of characters
An easy quantum group G comes with an algebra C(G) and a state, the
Haar state h. Hence, we can do free probability on the quantum group itself. A
first step is the computation of the laws of characters, which is, by the way, also
relevant for understanding the object from a quantum group perspective. P Let G
be a compact matrix quantum group of dimension n. Denote by χ := ni=1 uii
its character. If G ⊂ On+ , the uij are selfadjoint, hence also χ = χ∗ .
What are the moments of χ? Due to Woronowicz we know that h(χk ) is
exactly the dimension of the fixpoint algebra Fix(u⊗k ) = Hom(0, k). In the
case of easy quantum groups, this is given by the number of partitions in
C(0, k), see [21]. The computation of the law of characters of easy quantum
groups thus reduces to the problem of counting partitions. By [16, 21], we have
the following laws:
• On : real Gaussian
• On+ : semicircle
• Sn : Poisson
• Sn+ : free Poisson
Furthermore, the law on On∗ is given by the squeezed complex Gaussian, i.e.
the moments are given by φ((aa∗ )k ) for k even, and zero for k odd. Here, a is
a complex Gaussian.
There are several refinements of the study of laws of characters. For in-
stance, one can investigate Tr(um ) instead of Tr(u) = χ. One can also Ppass to
[nt]
tuples of elements, or one can study the truncated characters χt := i=1 uii .
See [17, 21] for research in this direction and further investigation of stochastic
aspects of easy quantum groups.
We want to point out that the investigation of the nature of these laws
of characters and of further possible links to free probability is still somehow
underdeveloped. We quote from Voiculescu’s lecture, contained in this volume:
It remains an open question whether the noncommutative dis-
tributions of the variables generating the free quantum groups
can be well integrated into the free probability framework. So,
do the distributions which arise in the free quantum group
setting fit in the free probability context or do these laws go
beyond?
112 M. Weber
Here, Wkn is the Weingarten matrix given as the inverse of the Gram matrix
Gkn (p, q) := hTp , Tq i.
projects onto the fixpoint algebra Fix(u⊗k ), which in turn has a basis
{Tp | p ∈ N C2 (0, k)}. By linear algebra, the entry Pij is of the form
X
Pij = G−1 (p, q),
p,q∈N C2 (0,k)
p≤ker i; q≤ker j
where G(p, q) is obtained from evaluating the inner product hTp , Tq i, for p, q ∈
N C2 (0, k).
The result can be extended to other free easy quantum groups, see [14,
17, 21], but the definition of the Gram matrix has to be adapted a bit. It is
then defined as Gkn (p, q) := n#{blocks in p∨q} where p ∨ q denotes the minimal
partition r with r ≥ p and r ≥ q. In principal, this allows us to compute the
Haar state on easy quantum groups by combinatorial means. In practice, it
is quite hard to invert the Gram matrix, see for instance [14] for some low-
dimensional computations in the case of Sn+ . √
It follows from the Weingarten formula, that the elements ( nuij )1≤i,j≤m
(with m ∈ N fixed) of C(On+ ) become asymptotically free semicircular elements,
as n → ∞ (see [13]).
Easy quantum groups 113
And here are two more projective partitions which are in N C(4, 4) but not
in N C2 (4, 4):
Maybe the reason for having three summands vs. five summands comes from
the fact that the above two partitions are in the category corresponding to Sn+
but not in the one of On+ ?
Even more enlightning is the following observation. Due to Banica’s inves-
tigation, we know that the fundamental representation u of Sn+ decomposes as
u = u0 ⊕ u1 . Now
u⊗2 = (u0 ⊕ u1 ) ⊗ (u0 ⊕ u1 ) = u0 ⊕ u1 ⊕ u1 ⊕ u0 ⊕ u1 ⊕ u2 .
The projective partitions in N C(2, 2) are exactly the following six:
Let us quickly sketch how the fusion rules for Sn+ may be deduced from the
above theorem. We may label the irreducible representations of Sn+ by projec-
tive partitions in N C. Now, two such representations are unitarily equivalent
if and only if they have the same number of through-blocks. Thus, the irre-
ducible representations may be labeled by N0 and we may take pk = |⊗k as a
representative for k 6= 0. Then, the elements of the set XC (pk , pl ) are mainly of
the same form as the above five examples of projective partitions in N C(4, 4).
This translates to the known fusion rules. The case of On+ is similar.
The above approach yields a unified proof for the fusion rules of all free easy
quantum groups. Furthermore, it can be extended to unitary easy quantum
groups. By the work of Freslon [38], operator algebraic properties like the
Haagerup property may be deduced from the fusion rules. Also, he studied
the possible fusion rules in more detail, isolating a “free part” and a “group
part”.
Note that this example of determining quantum algebraic properties by
combinatorial means is really a substantial one: The fusion rules of a quantum
group are an essential information not only of the full C ∗ -algebraic version of
the quantum group, but also of the reduced one, the von Neumann algebraic
one and the purely algebraic one. The fusion rules are intrinsic for the quantum
group as such.
10. Comments
For an introduction to algebraic quantum groups, Hopf algebras and the
Drinfeld–Jimbo approach, see [43, 44, 50]. For Kac algebras, we refer to [36].
For a quick start to Woronowicz’s approach to C ∗ -algebraic compact quantum
groups, the author recommends the book by Timmermann [58]. The best is,
to jump directly to Chapters 4 and 5; the first chapters of that book are more
algebraic. Another excellent book on compact quantum groups is the recent
one by Neshveyev and Tuset [52]. There are also good surveys [48, 51] from
the 90s. The latter one contains also locally compact quantum groups (see
also [58]). Besides, it is certainly worthwhile to take a look at Woronowicz’s
original papers on compact matrix quantum groups [72, 75], compact quantum
groups [76], Tannaka–Krein [74] and SUq (2) [73]. See also the introduction of
[69] for an overview on the general history of quantum groups.
The story of liberation of Lie groups begins with the work of Wang [69, 70].
There he defined Sn+ , On+ and the free unitary quantum group Un+ . By the
way, see [69, §4.1] for a comment on why both matrices u and ut need to
be unitary in Un+ . Wang and van Daele [62] also defined deformations of
On+ and Un+ by matrices Q ∈ GLn (C), which are related to Woronowicz’s
SUq (2), see [1]. Please consider also Banica’s slight adaption of the definition
(which is the one used nowadays) [1]. Banica worked a lot on refinements and
further studies of Wang’s quantum groups, see [1, 2, 4, 5, 7, 8, 12, 19] and the
nice survey (joint with Bichon and Collins) [10]. Bichon was also one of the
pioneers in free quantum groups, see for instance [24, 25] besides the many
articles in joint work with Banica. The hyperoctahedral quantum group Hn+
was introduced in [9]. The Weingarten calculus on quantum groups—i.e. the
Haar state computations—was developed by Banica and Collins [13, 14], the
latter one also being an expert on the Weingarten calculus on other structures.
Building on all the above work on free quantum groups, easy quantum
groups were first defined in 2009 by Banica and Speicher [21]. Articles directly
related to the classification are [16, 54, 55, 56, 71]. Stochastic aspects of easy
quantum groups may be found in [11, 17, 53]. Further articles somehow related
to easy quantum groups are [6, 15, 20, 23, 26]. For first steps to extend the easy
quantum group setting from quantum subgroups of On+ to quantum subgroups
Easy quantum groups 117
of Un+ , we refer to the ongoing research by Tarrago and the author [57]. In
order to get a better understanding of the operations on partitions, consult
[56] or [57].
The de Finetti result for Sn+ was stated by Köstler and Speicher [47]. It
was extended to other easy quantum groups in [18]. Further exchangeability
studies have been performed by Köstler partially in joint work with Gohm (for
actions of Sn and the braid group) [40, 41, 45, 46] and Curran partially in joint
work with Speicher (quantum setting) [30, 31, 32, 33, 34]. For the fusion rules
of easy quantum groups and others, see [3, 22, 38, 39, 49]. For literature on
von Neumann algebraic or other operator algebraic aspects, see Section 9.
This list is not guaranteed to be complete, but it might help to get an
overview of the subject. Here are some open questions and calls for possible
further developments.
• Find more proofs of the meta conjecture: “All quantum algebraic proper-
ties of easy quantum groups should be visible in terms of partitions.”
• Formulate analogs of classical results for Sn , On or Un for their quantum
versions Sn+ , On+ or Un+ .
• Extend any result obtained so far for Sn+ , On+ or Un+ to the other of these
three quantum groups. Even better: Extend it to any easy quantum group,
ideally with a uniform proof.
• Extend it to all compact matrix quantum groups Sn ⊂ G ⊂ On+ (see [55]
for a first, partial result in this direction—Theorem 4.6 can be extended
to noneasy quantum groups).
• We need more actions! Find (co-)actions of easy quantum groups on some
C ∗ -algebras. Can it be done in some systematic way?
• Do more free probability on easy quantum groups; for instance, answer
Voiculescu’s question about the laws of characters.
11. Exercises
Exercise 11.1. Consider the universal C ∗ -algebra A generated by commut-
ing selfadjoint elements uij , 1 ≤ i, j ≤ n such that the matrix u = (uij ) is
orthogonal (Example 2.3). Let φ : A → C be a character (i.e. a nonzero
∗
-homomorphism). Check that the matrix (φ(uij )) ∈ Mn (C) is orthogonal
and that it determines φ completely. Conversely, every orthogonal matrix in
Mn (C) gives rise to a character. Conclude that the space of characters of A is
homeomorphic to On and that C(On ) ∼ = A. Check that the generators uij are
mapped to the coordinate functions u′ij in C(On ).
Exercise 11.2.
(a) Check that N C2 and P2 are categories of partitions in the sense of Def-
inition 3.1. Check that every noncrossing pair partition p ∈ N C(0, l)
having no upper points may be obtained inductively from composing p′ ∈
N C(0, l − 2) with |⊗α ⊗ ⊓ ⊗ |⊗β for suitable α and β. Using rotation
(Proposition 4.1), infer that N C2 = h⊓, |i. Verify that we can permute the
points of a partition using ✁❆ and composition, and deduce that P2 = h ✁❆ i.
118 M. Weber
(We omit ⊓ and | as generating partitions since they are always contained
in a category by definition.)
(b) Check that N C and P are categories of partitions. Follow the lines of the
argumentation following Proposition 4.1 to show that N C = h , ↑i and
P =h , ↑, ❆✁ i.
(c) Prove that h i ⊂ N C is the category of all noncrossing partitions with
blocks of even size, whereas h↑i is the category of all noncrossing partitions
with blocks of size one or two.
Details may be found in [71].
Exercise 11.3. Let C ⊆ N C be a category of noncrossing partitions. Show
the following implications:
(a) If ∈ C and ↑ ∈ C, then C = N C. (Use Exercise 11.2.)
(b) If ∈
/ C and ↑ ∈ C, then C = h↑i. (Hint: Assume that C contains a
partition with a block of size at least three. Deduce that ∈ C. Now,
use Exercise 11.2.)
(c) If ∈ C and ↑ ⊗ ↑ ∈ / C, then C = h i. (Hint: Assume that C con-
tains a partition with a block of odd size. Deduce that ↑ ⊗ ↑ ∈ C. Use
Exercise 11.2.)
(d) If ∈
/ C and ↑ ⊗ ↑ ∈
/ C, then C = N C2 .
Extending these observations, one can show that there are exactly seven cate-
gories of noncrossing partitions, see [71].
Exercise 11.4. Let Xn = {x1 , . . . , xn } be a finite set. We can write the
∗
algebra C(Xn ) of continuous functions over XP n as the universal C -algebra
n
generated by projections e1 , . . . , en such that i=1 ei = 1. Let A be a C ∗ -
algebra generated by elements uij for 1 ≤ i, j ≤ n. Assume that there is a
unital ∗-homomorphism α : C(Xn ) → A ⊗ C(Xn ) such that
n
X
α(ei ) = uji ⊗ ej .
j=1
Show that the elements uij in A have to fulfill the following relations:
n
X
uij = u∗ij = u2ij , uij = 1.
j=1
References
[1] T. Banica, Théorie des représentations du groupe quantique compact libre O(n), C. R.
Acad. Sci. Paris Sér. I Math. 322 (1996), no. 3, 241–244. MR1378260
[2] T. Banica, Le groupe quantique compact libre U(n), Comm. Math. Phys. 190 (1997),
no. 1, 143–172. MR1484551
Easy quantum groups 119
[3] T. Banica, Fusion rules for representations of compact quantum groups, Exposition.
Math. 17 (1999), no. 4, 313–337. MR1734250
[4] T. Banica, Symmetries of a generic coaction, Math. Ann. 314 (1999), no. 4, 763–780.
MR1709109
[5] T. Banica, A note on free quantum groups, Ann. Math. Blaise Pascal 15 (2008), no. 2,
135–146. MR2468039
[6] T. Banica, S. Belinschi, M. Capitaine, and B. Collins, Free Bessel laws, Canad. J. Math.
63 (2011), no. 1, 3–37. MR2779129
[7] T. Banica and J. Bichon, Quantum automorphism groups of vertex-transitive graphs of
order ≤ 11, J. Algebraic Combin. 26 (2007), no. 1, 83–105. MR2335703
[8] T. Banica and J. Bichon, Quantum groups acting on 4 points, J. Reine Angew. Math.
626 (2009), 75–114. MR2492990
[9] T. Banica, J. Bichon, and B. Collins, The hyperoctahedral quantum group, J. Ramanu-
jan Math. Soc. 22 (2007), no. 4, 345–384. MR2376808
[10] T. Banica, J. Bichon, and B. Collins, Quantum permutation groups: a survey, in Non-
commutative harmonic analysis with applications to probability, 13–34, Banach Center
Publ., 78, Polish Acad. Sci. Inst. Math., Warsaw, 2008. MR2402345
[11] T. Banica, J. Bichon, B. Collins, and S. Curran, A maximality result for orthogonal
quantum groups, Comm. Algebra 41 (2013), no. 2, 656–665. MR3011789
[12] T. Banica, J. Bichon, and J.-M. Schlenker, Representations of quantum permutation
algebras, J. Funct. Anal. 257 (2009), no. 9, 2864–2910. MR2559720
[13] T. Banica and B. Collins, Integration over compact quantum groups, Publ. Res. Inst.
Math. Sci. 43 (2007), no. 2, 277–302. MR2341011
[14] T. Banica and B. Collins, Integration over quantum permutation groups, J. Funct. Anal.
242 (2007), no. 2, 641–657. MR2274824
[15] T. Banica, B. Collins and P. Zinn-Justin, Spectral analysis of the free orthogonal matrix,
Int. Math. Res. Not. IMRN 2009, no. 17, 3286–3309. MR2534999
[16] T. Banica, S. Curran, and R. Speicher, Classification results for easy quantum groups,
Pacific J. Math. 247 (2010), no. 1, 1–26. MR2718205
[17] T. Banica, S. Curran, and R. Speicher, Stochastic aspects of easy quantum groups,
Probab. Theory Related Fields 149 (2011), no. 3-4, 435–462. MR2776622
[18] T. Banica, S. Curran, and R. Speicher, De Finetti theorems for easy quantum groups,
Ann. Probab. 40 (2012), no. 1, 401–435. MR2917777
[19] T. Banica and S. Moroianu, On the structure of quantum permutation groups, Proc.
Amer. Math. Soc. 135 (2007), no. 1, 21–29. MR2280170
[20] T. Banica and A. Skalski, Two-parameter families of quantum symmetry groups, J.
Funct. Anal. 260 (2011), no. 11, 3252–3282. MR2776569
[21] T. Banica and R. Speicher, Liberation of orthogonal Lie groups, Adv. Math. 222 (2009),
no. 4, 1461–1501. MR2554941
[22] T. Banica and R. Vergnioux, Fusion rules for quantum reflection groups, J. Noncommut.
Geom. 3 (2009), no. 3, 327–359. MR2511633
[23] T. Banica and R. Vergnioux, Invariants of the half-liberated orthogonal group, Ann.
Inst. Fourier (Grenoble) 60 (2010), no. 6, 2137–2164. MR2791653
[24] J. Bichon, Quantum automorphism groups of finite graphs, Proc. Amer. Math. Soc.
131 (2003), no. 3, 665–673 (electronic). MR1937403
[25] J. Bichon, Free wreath product by the quantum permutation group, Algebr. Represent.
Theory 7 (2004), no. 4, 343–362. MR2096666
[26] J. Bichon and M. Dubois-Violette, Half-commutative orthogonal Hopf algebras, Pacific
J. Math. 263 (2013), no. 1, 13–28. MR3069074
[27] M. Brannan, Approximation properties for free orthogonal and free unitary quantum
groups, J. Reine Angew. Math. 672 (2012), 223–251. MR2995437
[28] M. Brannan, Quantum symmetries and strong Haagerup inequalities, Comm. Math.
Phys. 311 (2012), no. 1, 21–53. MR2892462
120 M. Weber
[53] S. Raum, Isomorphisms and fusion rules of orthogonal free quantum groups and
their free complexifications, Proc. Amer. Math. Soc. 140 (2012), no. 9, 3207–3218.
MR2917093
[54] S. Raum and M. Weber, The combinatorics of an algebraic class of easy quantum groups,
Infin. Dimens. Anal. Quantum Probab. Relat. Top. 17 (2014), no. 3. MR3241025
[55] S. Raum and M. Weber, Easy quantum groups and quantum subgroups of a semi-direct
product quantum group, J. Noncommut. Geom. 9 (2015), no. 4, 1261–1293. MR3448336
[56] S. Raum and M. Weber, The full classification of orthogonal easy quantum groups,
Comm. Math. Phys., 341 (2016), no. 3, 751–779.
[57] P. Tarrago and M. Weber, Unitary easy quantum groups: the free case and the group
case. arXiv:1512.00195 [math.QA] (2015).
[58] T. Timmermann, An invitation to quantum groups and duality, EMS Textbk. Math.,
Eur. Math. Soc., Zürich, 2008. MR2397671
[59] S. Vaes and N. Vander Vennet, Poisson boundary of the discrete quantum group A\ u (F ),
Compos. Math. 146 (2010), no. 4, 1073–1095. MR2660685
[60] S. Vaes and R. Vergnioux, The boundary of universal discrete quantum groups, exact-
ness, and factoriality, Duke Math. J. 140 (2007), no. 1, 35–84. MR2355067
[61] A. Van Daele, The Haar measure on a compact quantum group, Proc. Amer. Math.
Soc. 123 (1995), no. 10, 3125–3128. MR1277138
[62] A. Van Daele and S. Wang, Universal quantum groups, Internat. J. Math. 7 (1996),
no. 2, 255–263. MR1382726
[63] R. Vergnioux, Orientation of quantum Cayley trees and applications, J. Reine Angew.
Math. 580 (2005), 101–138. MR2130588
[64] R. Vergnioux, The property of rapid decay for discrete quantum groups, J. Operator
Theory 57 (2007), no. 2, 303–324. MR2329000
[65] R. Vergnioux, Paths in quantum Cayley trees and L2 -cohomology, Adv. Math. 229
(2012), no. 5, 2686–2711. MR2889142
[66] R. Vergnioux and C. Voigt, The K-theory of free quantum groups, Math. Ann. 357
(2013), no. 1, 355–400. MR3084350
[67] C. Voigt, The Baum-Connes conjecture for free orthogonal quantum groups, Adv. Math.
227 (2011), no. 5, 1873–1913. MR2803790
[68] C. Voigt, On the structure of quantum automorphism groups, J. Reine Angew. Math.
(2015), DOI 10.1515/crelle-2014-0141.
[69] S. Wang, Free products of compact quantum groups, Comm. Math. Phys. 167 (1995),
no. 3, 671–692. MR1316765
[70] S. Wang, Quantum symmetry groups of finite spaces, Comm. Math. Phys. 195 (1998),
no. 1, 195–211. MR1637425
[71] M. Weber, On the classification of easy quantum groups, Adv. Math. 245 (2013), 500–
533. MR3084436
[72] S. L. Woronowicz, Compact matrix pseudogroups, Comm. Math. Phys. 111 (1987),
no. 4, 613–665. MR0901157
[73] S. L. Woronowicz, Twisted SU(2) group. An example of a noncommutative differential
calculus, Publ. Res. Inst. Math. Sci. 23 (1987), no. 1, 117–181. MR0890482
[74] S. L. Woronowicz, Tannaka-Kreı̆n duality for compact matrix pseudogroups. Twisted
SU(N ) groups, Invent. Math. 93 (1988), no. 1, 35–76. MR0943923
[75] S. L. Woronowicz, A remark on compact matrix quantum groups, Lett. Math. Phys. 21
(1991), no. 1, 35–39. MR1088408
[76] S. L. Woronowicz, Compact quantum groups, in Symétries quantiques (Les Houches,
1995), 845–884, North-Holland, Amsterdam, 1998. MR1616348
Participants of the Masterclass on
Free Probability and Operator Algebras
Hari Bercovici
Mathematics Department, Indiana University
Bloomington, IN 47405, USA
E-mail: bercovic@indiana.edu
Ken Dykema
Department of Mathematics, Texas A&M University
College Station, TX 77843-3368, USA
E-mail: kdykema@math.tamu.edu
Dimitri Shlyakhtenko
Department of Mathematics, UCLA,
Los Angeles, CA 90095, USA
E-mail: shlyakht@math.ucla.edu
Roland Speicher
Fachrichtung Mathematik, Saarland University,
Postfach 151150, 66041 Saarbrücken, Germany
E-mail: speicher@math.uni-sb.de
Dan-Virgil Voiculescu
Department of Mathematics, University of California, Berkeley,
970 Evans Hall #3840, Berkeley, CA 94720-3840, USA
E-mail: dvv@math.berkeley.edu
Moritz Weber
Faculty of Mathematics, Saarland University,
66123 Saarbrücken, Germany
E-mail: weber@math.uni-sb.de
Index
amenable, 116 free max-stable laws, 5
arcsine law, 14 free Monge-Ampère equation, 50
free orthogonal quantum group, 98
Baum-Connes conjecture, 115 free Poisson/Marchenko–Pastur law, 15,
Bercovici–Pata bijection, 82, 89 89
Bernoulli variable, 15, 31 compound, 15
free Poisson/Marchenko-Pastur law, 111
Cartan subalgebra, 68, 115
free product
Catalan number, 13, 20, 30, 54
C∗ -algebras, 59
category of partitions, 102, 117
Hilbert spaces, 58
Cauchy distribution, 15
von Neumann algebras, 61, 76
Cauchy transform, 29, 78
free symmetric quantum group, 99, 118
central limit theorem
free unitary quantum group, 107
classical, 23, 74
fundamental group, 64
free, 3, 17, 24, 36, 73, 87
fusion rules, 113
Chebychev polynomials, 42
circular element, 14, 63 Gaussian distribution, 23, 62, 74
compact matrix quantum group, 99 Gaussian family, 18, 32
compact quantum group, 97 Gaussian random matrix, 18, 21, 36, 62
conditional expectation, 8, 84 Gaussian unitary ensemble, 17, 18, 32,
convergence in distribution, 12, 22, 62, 35, 53
74 genus expansion, 20, 36
convergence in moments, see also group algebra, 8, 9
convergence in distribution GUE, see also Gaussian unitary
creation operator, 1, 10, 36 ensemble
cumulant series, 28
cumulants, 3, 13, 25, 29, 36, 74, 87 Haagerup approximation property, 43,
vanishing of, 26, 109 115
cyclic derivative, 46 Haar state, 97, 112
Haar unitary, 14, 35, 63
de Finetti theorem k-Haar unitary, 14
classical, 108 random matrix, 35, 62
free, 5, 109 hyperfinite factor, 64
easy quantum group, 103 ICC group, 43, 63, 68
eigenvalue distribution, 18 independence
eigenvalue distribution, 13 asymptotically free, 12, 32, 36, 62
entropy, see also free entropy bi-free, 2, 9, 10
exchangeable, 5, 108 Boolean, 9
classical/tensor, 2, 9
Fock space, 1, 10, 14, 36
free, 2, 9
free convolution
matricial free, 9
additive, 3, 28, 30, 36, 60, 77, 87
monotone, 9
multiplicative, 31, 36, 83
operator-valued free/amalgamated
free cumulants, see also cumulants
free, 11
free difference quotient, 46, 53
traffic free, 9
free entropy, 4, 48
infinitely divisible, 4, 89
dimension, 68
intertwiner space, 100
free Gibbs law, 47, 53
free group, 1 join, 25
free group factors, 2, 10, 57, 61, 63
q-deformed, 51 Kolmogorov metric, 77, 87
interpolated, 3, 65 Kreweras complement, 31
132
noncommutative distribution, 1, 12
operator-valued, 13
noncommutative probability space, 1, 7,
57, 61
operator-valued, 8, 84, 89
noncommutative random variable, 1, 9
unbounded, 75, 87
Free Probability
Moritz Weber, Editors
and Operator
areas (quantum mechanics, free group algebras, random matrices
etc). Thirty years after its foundation, it is a well-established and very
active field of mathematics. Originating from Voiculescu’s attempt
to solve the free group factor problem in operator algebras, free
probability has important connections with random matrix theory,
combinatorics, harmonic analysis, representation theory of large
groups, and wireless communication.
Algebras
These lecture notes arose from a masterclass in Münster, Germany
and present the state of free probability from an operator algebraic
Dan-Virgil Voiculescu
Nicolai Stammeier
matrices and combinatorics of free probability (Speicher), free
monotone transport (Shlyakhtenko), free group factors (Dykema),
free convolution (Bercovici), easy quantum groups (Weber), and a
historical review with an outlook (Voiculescu). In order to make it
more accessible, the exposition features a chapter on basics in free
Moritz Weber
probability, and exercises for each part.
ISBN 978-3-03719-165-1
www.ems-ph.org
Voiculescu / Font: NewsGothic / Pantone: 287 / Pantone: 116 / Format: 170 x 240 / RB: 7.2 mm