The Story of Two Entropies

Open AI addressed also another one of the big questions in free probability, namely the relation between the two versions – microstates and non-microrstates – of free entropy. Although the negative solution of Connes embedding problem already had the consequence that they cannot be equal in general, it was still possible that in the Connes-embeddable subclass they could agree. However, AI showed us a counter-example where they are different. This might not have been so unexpected as the isomorphism of the free group factors, but is still a major result, which we should understand better.

And again, this understanding is possible, by a very nice recent preprint by David Jekel.

The Story of the Free Group Factor

The question whether the free group factors are isomorphic for different numbers of generators was the problem which led Dan Voiculescu to the creation of free probability. Many tools were developed in the last 40 years for attacking this problem; there were waves of optimism, when a solution seemed within reach, followed by periods when it seemed as distant as ever. Now the breakthrough has come through AI. And surprisingly for many of us, this solution is not distinguishing the free group factors but tells us that there is only one.

The Open AI solution has the usual problems of an AI-produced and written proof, but luckily enough, Dima Shlyakhtenko took on the task of understanding and improving the AI manuscript; see here for his preprint On the isomorphism of semicircular algebras. And the improvement is not just on the level of exposition, but he observed that the arguments extend from the von Neumann to the C*-algebra level and thus he can show that actually also the C*-algebras generated by at least two free semicirculars are isomorphic. Of course, we know that for free Haar unitaries this is not true, since the reduced C*-algebras of the free groups are not isomorphic, by Pimsner-Voiculescu. All this is quite unexpected and opens many new directions for research.

Hopefully somebody will have to say a bit more about all this; for now I only want to point out that Dima will give a talk on these results at the POAS seminar today, on Monday November 16 at 9am Pacific time (6pm European time)

The Story of the Determinant (and the Brown measure): Convergence of Polynomials in GUE

I have already started a story on the Fuglede-Kadison determinant of operator-valued semicircular elements, which also touched upon the Brown measure of polynomials in free circular operators. There has been some recent (mostly human, as I understand) progress on the question whether the eigenvalue distribution of a polynomial in independent Ginibre matrices converges to the Brown measure of the polynomial in the limiting circular operators. This was posted as an explicit question in my paper with Serban Belinschi and Piotr Sniady. For polynomials which only involve the Ginibre matrices, but not their adjoints this was proved a few months ago (which now feels like years ago) by Yi Han in this preprint.

A few days ago a new preprint Brown Measures and Spectral Convergence for Arbitrary GUE Polynomials by Collins, Kitagawa, Miyagawa, Yuan, Zhong appeared on the arXiv. This solves the problem in the affirmative for general *-polynomials in Ginibre matrices. They use quite a bit of free probability machinery; in particular, a formula for the Fuglede-Kadison determinant of operator-valued semicirculars appears also in their Lemma 5.7.

Some More Stories Around Free Probability

I am still quite a bit paralyzed by the overwhelming AI-supported progress in mathematics in general and in our field in particular. I also feel a bit too old and too rusty on technical details for working out by myself the details of what is going on. But luckily enough, there are many young(er) and more energetic people in our subject which bravely take on this task and give us human versions of the progress.

What I want to contribute to the present situation is to keep track of a few stories on progress on some of the problems which I find interesting. This list of stories might increase over time, but it should not become a huge collection of everything, but focus on what we perceive as the core problems of free probability. Storytelling by guests is also welcome.

The Story of the Determinant: Why the Determinant Formula is True

post by Vladislav Kargin

In his post Roland asked whether the Mai–Speicher formula for the Fuglede–Kadison determinant of a matrix semicircle survives when a constant matrix is added. With Sa=a⊗1+∑jAj⊗sjS_a=a\otimes1+\sum_jA_j\otimes s_j and η(X)=∑jAjXAj\eta(X)=\sum_jA_jXA_j, the proposed formula is

log⁡Δ(Sa−z)=−12+infb⁡{trm⁡log⁡b+12⟨a−z,(𝒞b+η)−1(a−z)⟩},𝒞b(X)=bXb,\log\Delta(S_a-z)=-\frac12+\inf_{b}\Bigl\{\operatorname{tr}_m\log b +\frac12\langle a-z,(\mathcal C_b+\eta)^{-1}(a-z)\rangle \Bigr\}, \qquad \mathcal C_b(X)=bXb,

where the infimum is over positive matrices bb with 𝒞b−η⪰0\mathcal C_b-\eta\succeq0. It is true, for every real zz. The proof is in the manuscript. Here I only want to say how I think about it, because the formula looked strange to me at first and now it looks a bit less strange.

There had to be some formula

The logarithmic potential is an integral of the Stieltjes transform. The Stieltjes transform comes from the Dyson equation G−1+η(G)=z−aG^{-1}+\eta(G)=z-a. And the Dyson equation is the Euler–Lagrange equation of a functional, the Dyson action

𝒜z(M)=−trm⁡log⁡M−12trm⁡(Mη(M))+trm⁡((z−a)M)−1.\mathcal A_z(M)=-\operatorname{tr}_m\log M-\frac12\operatorname{tr}_m\bigl(M\eta(M)\bigr) +\operatorname{tr}_m\bigl((z-a)M\bigr)-1 .

If you differentiate the action at its critical point with respect to zz, only the explicit zz contributes, and you get trm⁡G\operatorname{tr}_mG. So the action at the Dyson solution is also an integral of the Stieltjes transform, and it is not a surprise that log⁡Δ(Sa−z)=Re⁡𝒜z(G(z))\log\Delta(S_a-z)=\operatorname{Re}\mathcal A_z(G(z)). The shift aa plays no role here. It rides along with zz.

So the determinant is a critical value. The trouble is that it is the wrong kind of critical value. The action is holomorphic in MM, so the Dyson solution is a saddle of its real part, with as many directions going down as going up. A formula with an “inf⁡\inf” in it cannot come from the action directly.

The idea: free the scale

In the centered case the answer is det⁡b\det b for the matrix bb that scales η\eta to a doubly stochastic map. This suggested that GG should come with a positive “scale” also in general. It does: there is a unique b≻0b\succ0 with

G=b−1/2Ub−1/2,U unitary.G=b^{-1/2}\,U\,b^{-1/2},\qquad U\ \text{unitary}.

For scalars this is just b=1/|G|b=1/|G| and UU is the phase.

The step that was unexpected for me is to forget that bb is determined by GG and treat it as a new independent variable. Write G=−u−iγG=-u-i\gamma and consider

𝒦(b,u,γ)=trm⁡log⁡b+⟨a−z,u⟩−12⟨u,(𝒞b+η)u⟩−12⟨γ,(𝒞b−η)γ⟩.\mathcal K(b,u,\gamma)=\operatorname{tr}_m\log b+\langle a-z,u\rangle -\frac12\langle u,(\mathcal C_b+\eta)u\rangle -\frac12\langle \gamma,(\mathcal C_b-\eta)\gamma\rangle .

When bb is the true scale of −u−iγ-u-i\gamma this is the real part of the Dyson action (plus 12\frac12), by a direct computation. And the Dyson solution, together with its scale, is still a critical point of 𝒦\mathcal K in the three variables. So nothing is lost. What is gained is that the saddle has been taken apart: 𝒦\mathcal K is concave in uu and in γ\gamma, so all the descending directions sit in these two variables, and one can remove them by maximizing.

Two things that come for free

The first is the constraint. I had thought of 𝒞b−η⪰0\mathcal C_b-\eta\succeq0 as a condition somebody cleverly imposed. But look at the last term of 𝒦\mathcal K. Its supremum over γ⪰0\gamma\succeq0 is 00 if 𝒞b−η⪰0\mathcal C_b-\eta\succeq0 and +∞+\infty otherwise. The feasible set is simply the set of bb for which the problem makes sense. It is the trace that the imaginary part of GG leaves behind.

Maximizing over uu then gives exactly the function in Roland’s formula.

The second is convexity, and this is the heart of the matter. The function of bb that remains is not convex in the usual sense. It is convex along the geodesics b1/2etXb1/2b^{1/2}e^{tX}b^{1/2} of the positive cone, and it is so precisely on the feasible set. The scalar semicircle shows what this means. There the function is

Fx(b)=log⁡b+x22(b2+1),b≥1.F_x(b)=\log b+\frac{x^2}{2(b^2+1)},\qquad b\ge1 .

For positive numbers the geodesics are simply t↦bett\mapsto be^{t}, so geodesic convexity is ordinary convexity in the variable s=log⁡bs=\log b. In this variable

Fx=s+x24(1−tanh⁡s).F_x=s+\frac{x^2}{4}\,\bigl(1-\tanh s\bigr).

The first term is linear. The second is a reflected tanh⁡\tanh: it is convex for s>0s>0 and concave for s<0s<0, and its inflection point is at s=0s=0, that is, at b=1b=1, exactly the boundary of the feasible set. So the constraint that came out of the imaginary part is the same as the condition that makes the problem convex. I do not think this is an accident, but I cannot say in one sentence why it must be so.

Once we have convexity, a critical point is a global minimum, and the critical point is the scale of the Dyson solution. This is the whole proof in the upper half-plane. To get to the real axis one adds εtrm⁡γ\varepsilon\operatorname{tr}_m\gamma to 𝒦\mathcal K, which adds a positive term of order ε2\varepsilon^2 to the function of bb, and lets ε→0\varepsilon\to0. Everything is monotone, and one never needs to know what the minimizers do.

What the picture looks like

In the scalar example the equation Fx′(b)=0F_x'(b)=0 is b+1/b=|x|b+1/b=|x|, which is the Dyson equation for b=1/|G|b=1/|G|. For |x|>2|x|>2 it has a solution with b>1b>1. For |x|<2|x|<2 it has none, the minimum sits at the wall b=1b=1, and its value is x2/4x^2/4, as it should be.

The same holds for matrices, at least at points where the density is nondegenerate. Outside the spectrum the minimizer is in the interior of the feasible set. Inside the spectrum it is on the boundary, and the kernel of 𝒞b−η\mathcal C_b-\eta at the minimizer contains γ=−Im⁡G\gamma=-\operatorname{Im}G. In other words, the spectrum is where the constraint is active, and the density of states is its Lagrange multiplier. This is the form in which I remember the result.

What I do not understand

Is the infimum attained at spectral edges, and at points where the matrix density is nonzero but singular? I do not know. The scales bεb_\varepsilon might run away, and at points where the density blows up they do. It would be nice to read the type of the singularity from how fast they run.

The question I find most interesting is about the Brown measure. By hermitization the formula applies to any S=a0⊗1+∑ai⊗siS=a_0\otimes1+\sum a_i\otimes s_i with arbitrary matrices aia_i and any complex zz. For each fixed bb the function under the infimum is a quadratic polynomial in Re⁡z\operatorname{Re} z and Im⁡z\operatorname{Im}z. So the logarithmic potential of the Brown measure is an infimum of quadratic polynomials over a fixed convex set.

Now we can ask whether the minimizer bzb_z sits in the interior or on the boundary of the feasible set. By the picture above this should detect the spectrum of SS: if S−zS-z is invertible, the minimizer is in the interior. Is this enough to detect the support of the Brown measure? That is, is the Brown measure of SS supported exactly where the constraint is active? For general operators the spectrum can be strictly larger than the support of the Brown measure. Haagerup and Larsen found this for RR-diagonal elements, where it happens when the element has an inverse which is in L2L^2 but is not bounded. Miyagawa has recently shown that a rational function in free semicirculars which is in L2L^2 is bounded, so this mechanism is not available here, and he asks whether there is any gap at all in that setting. Hall and Ho proved that there is none for a self-adjoint element plus a free elliptic one. For matrix semicirculars I do not know the answer. If there is no gap, the variational problem should show it: one would have to prove that the potential cannot be harmonic at a point where the constraint binds.

While this post was waiting to appear, Sheng Yin sent me a different formula for the same determinant, also found with the help of AI. In his formula the infimum is over all positive matrices, with no constraint, and it is attained exactly when the operator is invertible; in Roland’s formula the minimizer sits on the wall inside the spectrum. For self-adjoint coefficients the two are related by an explicit change of variables, which turns the free variable of one into the constrained variable of the other. His manuscript is on Hexagon (hexagon:2610.00180). How the two pictures fit together in general is something we would like to understand.

I should add that I worked on this in long conversations with language models. It felt like working with a co-author, and I cannot separate who contributed what. The responsibility for the correctness is mine.

A Story of the Determinant of Operator-Valued Semicircular Elements

Operator-valued semicircular elements are the powerhouse of free probability and random matrix theory. In particular, the intrinsic freeness principle typically replaces random matrices by operator-valued semicircular elements and shows that, in many respects, the former are close to the latter. Thus it is important to understand as much as possible about operator-valued semicirculars.

An operator-valued semicircular element – say over B=Mm(ℂ)B=M_m(\mathbb C) – is of the form

S=a⊗1+∑i=1nai⊗siwitha,a1,…,an∈Mmsa(𝐂),S=a\otimes 1+\sum_{i=1}^n a_i\otimes s_i\qquad\text{with} \qquad a,a_1,\dots,a_n\in M_m^{sa}(\mathbf C),

where s1,…,sn s_1,\dots,s_n are free semicirculars. We denote by trm{\rm tr}_m the normalized matrix trace on BB, and by τ\tau the trace on the algebra, where the sis_i live. All information about such an element SS is contained in its mean

(id⊗τ)[S]=a ({\rm id} \otimes \tau)[S]=a

and its covariance

η(b)=:(id⊗τ)[(S−a)b(S−a)]=∑i=1naibai.\eta(b)=:({\rm id}\otimes \tau)[(S-a)b(S-a)]=\sum_{i=1}^n a_i b a_i.

The covariance η\eta is a completely positive map from BB to BB.

We also know how, in principle, to calculate the distribution μS\mu_S of SS. Its scalar-valued Cauchy transform is

gS(z)=(tr⁡m⊗τ)[(z1−S)−1].g_S(z)=(\tr_m\otimes \tau)[(z1-S)^{-1}].

This can be extracted from the operator-valued Cauchy transform

GS(b)=(id⊗τ)[(b−S)−1]G_S(b)=({\rm id}\otimes\tau)[(b-S)^{-1}]

simply by

gS(z)=tr⁡m[GS(z1)].g_S(z)=\tr_m[G_S(z1)].

And GSG_S is determined by the nice quadratic matrix-valued equation, usually called the Dyson equation,

(b−a)⋅GS(b)=1+η(GS(b))⋅GS(b). (b-a)\cdot G_S(b)=1+\eta(G_S(b))\cdot G_S(b).

So everything is determined by mean and covariance. Of course, some quantities are easier, others harder to extract in a meaningful way. For moments of SS, for example, one has the operator-valued free version of the Wick/Isserlis formula. The operator norm of SS, on the other hand, is not so easily accessible. In principle, this is given by the asymptotics of high moments, but that is not very concrete. So the following celebrated formula of Franz Lehner

λmax⁡(a⊗1+∑i=1nai⊗si)=infb∈B,b>0⁡λmax⁡(b+a+η(b−1))\lambda_{\max} \left(a\otimes 1+\sum_{i=1}^n a_i\otimes s_i\right) = \inf_{b\in B, b \gt 0}\lambda_{\max} \left(b+a+\eta(b^{-1})\right)

comes as a nice surprise, as it reduces the upper spectral edge of the infinite-dimensional SS to a variational problem for the same quantity on the finite-dimensional matrix algebra BB. Here λmax⁡\lambda_{\max} denotes the upper edge of the spectrum; for positive aa, Lehner’s original formulation gives the corresponding operator norm.

A few years ago together with Tobias Mai I was trying to understand the determinant (more precisely, the Fuglede–Kadison determinant) of an operator-valued semicircular element, in order to get some info about the accumulation of mass of the distribution of such a semicircular element close to zero. And we succeeded in finding a nice formula in terms of the covariance; more precisely the determinant was essentially given by the capacity of the covariance map. Capacity here is the quantity introduced by Gurvits in connection with operator scaling; it should not be confused with the various notions of channel capacity in quantum information. Unwinding this notion of capacity one can write our result concretely as

Δ(∑i=1nai⊗si)=e−1/2⋅infb∈B,b>0⁡Δ(b⋅η(b−1))1/2.\Delta(\sum_{i=1}^n a_i\otimes s_i) = e^{-1/2}\cdot \inf_{b\in B, b>0} \Delta\left(b\cdot \eta(b^{-1})\right)^{1/2}.

Δ\Delta is here the Fuglede–Kadison determinant, which reduces in the finite-dimensional situation to

Δ(b)=det⁡(|b|)1/mforb∈Mm(ℂ).\Delta(b)=\det (\vert b\vert)^{1/m}\qquad\text{for}\qquad b\in M_m(\mathbb C).

As you see, our result is only for mean a=0a=0. But in this case it has a striking similarity with Lehner’s formula:

‖∑i=1nai⊗si‖=infb∈B,b>0⁡‖b+η(b−1)‖,\| \sum_{i=1}^n a_i\otimes s_i\| = \inf_{b\in B, b>0} \left\|b+\eta(b^{-1})\right\|,

And of course it raises the question: Can we extend our formula also to the case of general mean. We thought a little bit about this way back then, but did not see how to do this.

Now, in the times of AI hype, it is tempting to come back to this question and ask ChatGPT what it has to say on this. And actually, it has something to say. It proposes the following generalization

Δ(a⊗1+∑i=1nai⊗si)=e−1/2infb>0, 𝒞b−η⪰0⁡Δ(b)exp⁡{12⟨a,(𝒞b+η)−1(a)⟩2},\Delta\left(a\otimes 1+\sum_{i=1}^n a_i\otimes s_i\right) = e^{-1/2} \inf_{b \gt 0,\ \mathcal C_b-\eta\succeq 0} \Delta(b)\, \exp\left\{ \frac12 \left\langle a, (\mathcal C_b+\eta)^{-1}(a) \right\rangle_2 \right\},

where

𝒞b(x)=bxb,and⟨x,y⟩2=trm⁡(x∗y)forx,y∈Mm(ℂ)\mathcal C_b(x)=bxb,\qquad\text{and}\qquad\langle x,y\rangle_2 = \operatorname{tr}_m(x^*y)\qquad \text{for} \qquad x,y\in M_m(\mathbb C)

is the Hilbert-Schmidt inner product.

The condition 𝒞b−η⪰0\mathcal C_b-\eta\succeq 0 means positivity as an operator on the Hilbert–Schmidt space Mm(ℂ)M_m(\mathbb C), i.e.

⟨x,𝒞b(x)⟩2≥⟨x,η(x)⟩2,or equivalentlytrm⁡(x∗bxb)≥trm⁡(x∗η(x)),for all x∈Mm(ℂ).\left\langle x,\mathcal C_b(x)\right\rangle_2 \geq \left\langle x,\eta(x)\right\rangle_2,\qquad\text{or equivalently}\qquad \operatorname{tr}_m(x^*bxb) \geq \operatorname{tr}_m\left(x^*\eta(x)\right), \qquad\text{for all }x\in M_m(\mathbb C).

ChatGPT also produces what it claims is a proof of all this — note that even the reduction of this to my formula with Tobias in the case of a=0a=0 is non-trivial — and makes some remarks on the relevance for the Brown measure of an operator-valued circular element. Recall that the Brown measure of an operator TT is encoded by the logarithmic potential

z↦log⁡Δ(T−z1).z\mapsto \log \Delta (T-z1).

So understanding determinants of shifted operators is exactly what one needs here.

All this looks somehow reasonable, but not very enlightening; and even if I had a Lean certificate for all this (which I don’t), I would still not be too satisfied.

So let’s talk about all this and hope that, together, we can get a better understanding of what is really going on.

Doing Mathematics with AI — but How?

AI is revolutionizing mathematics, and it will not go away. So somehow we have to arrange ourselves with it and find our way of doing mathematics under these changed conditions.

I am of course also fascinated by asking AI (ChatGPT in my case) to solve all my problems. And sometimes there are indeed answers which seem to be okay and which promise some real progress. But those answers are usually complicated and quite technical, and I don’t really feel like becoming a little helper and argument checker for AI.

I agree with many voices saying that asking the meaningful questions, and understanding and presenting the answers, will remain some of the main tasks for us. But this should happen in a way that we still feel good about going along with it. For me this means that I would like to take from AI mainly some ideas, but then think myself about whether they could be true, what they actually mean, how one could prove them, and what is the best way to present them and convince others.

Of course, like everybody else, I don’t have a final answer to how this should work in practice, or whether this is a route which will help mathematics to survive in anything resembling its present form.

Anyhow, I suppose we just have to try different routes. Here is one experiment I would like to make.

At the moment I am thinking about determinants of operator-valued semicircular elements. Together with Tobias Mai, we derived a few years ago a formula for the determinant of such an element S in terms of its covariance map. This was, however, for the case of vanishing mean of S, and I wondered whether there could be a nice and useful extension to the case of general mean.

Okay, so I asked ChatGPT. And it did what two years ago would have seemed totally out of range, but has now become almost ordinary: it gave me an answer, proved it, and also made some comments on possible implications for questions about the Brown measure of operator-valued circular elements.

It even wrote a manuscript about all this.

I could now just put my name on it, say that the ideas were developed in conversations with AI, and declare that I take full responsibility for the results. But this does not feel right to me. And, more importantly, it does not feel very satisfying.

So here is what I would rather like to do.

I will give some background, explain what ChatGPT proposes as the answer, and then open the problem for public discussion: Is the statement correct? Is it perhaps obvious? How is it related to things that somebody already knows? What is the right proof? And what other interesting questions or conjectures might come out of it?

My dream would be that the answers — and also the questions — are actually self-thought, and not just copies of something another AI conversation produced. Of course everybody could now ask AI about the problem and perhaps produce a paper on it. But this should not be about publishing or priority. It should be about understanding.

Maybe one could think of it as telling a mathematical story together, in a somewhat grassroots way: somebody starts with a question, somebody else recognizes a connection, another person finds an argument, somebody points out that the whole thing is wrong, or suggests the right formulation, and gradually we understand what is really going on. Of course, we can use AI in the background for information, references, or inspiration. But AI should not tell the story.

I have no idea whether this will work.

But let us try.

In my next post I will give the concrete mathematical problem and the beginning of the story.

The intermediate quantum permutation group problem

Guest post by Amaury Freslon

At a recent conference celebrating Roland Speicher’s birthday, I gave a talk on a problem from the theory of compact quantum groups that has fascinated me for years and on which Roland and I have been working, albeit unsuccessfully. I wrote detailed notes introducing the problem, which are available here and Roland suggested I write a blog post to publicize them. So here I am !

Let’s begin by unpacking the terminology in the title. The quantum permutation group SN+, introduced by Sh. Wang, is defined via a universal ∗-algebra 𝒪(SN+), which is generated by entries of a matrix (pij). These entries are subject to relations that resemble those of classical permutation matrices, but with the key difference that they do not commute. More precisely, the defining properties are as follows:

  • Each generator is idempotent and self-adjoint;
  • They are pairwise orthogonal on rows and columns;
  • They add up to 1 on rows and columns.

In contrast, let 𝒪(SN) be the algebra of all complex-valued function on the classical permutation group SN. The connection between SN+ and SN is given by a canonical ∗-homomorphism πab: 𝒪(SN+) → 𝒪(SN), which is defined by making all generators commute. In that sense, SN+ can be thought of as the free or quantum version of SN. We can now state our main problem:

Does there exist a compact quantum group 𝔾 such that SN<𝔾<SN+? In other words, is there a Hopf ∗-algebra 𝒪(𝔾) through which the map πab factors?

In algebraic terms, the problem asks whether there exists a non-trivial Hopf ideal inside the commutator ideal of 𝒪(SN+). However, understanding Hopf ideals in general is notoriously difficult, so we must look for alternative approaches. One promising direction involves exploiting the fact that the quantum permutation group is an easy quantum groups in the sense that all the relations in the algebra 𝒪(SN+) are encoded by partitions of finite sets. While we will not delve into the details here (see the notes and references therein), we can sketch a potential strategy.

  1. The relations in 𝒪(SN+) are exactly those given by linear combinations of non-crossing partitions;
  2. The relations in 𝒪(SN) are exactly those given by linear combinations of arbitrary partitions;
  3. The problem therefore boils down to this: if we are given a relation corresponding to a linear combination of arbitrary partitions, can we combine it with relations coming from non-crossing partitions to obtain all possible relations from partitions?


While this approach seems reasonable, it has not proven efficient thus far. The only known results on the problem, which rely on different techniques, are as follows:

  • For N<4, there is no intermediate quantum permutation group, because SN=SN+ in that case;
  • For N=4, there is again no intermediate quantum permutation group, as can be observed on the complete list of quantum subgroups of S4+;
  • For N=5, the absence of intermediate quantum permutation group was obtained by T. Banica relying on deep results from the classification of subfactors.

The notes outline an alternative approach developed by Roland Speicher and myself which is inspired by noncommutative probability theory. More precisely, any quantum subgroup 𝔾 of SN+ comes with a distinguished linear form, known as its Haar state. Pulling it back to 𝒪(SN+), we obtain a map with lots of symmetry properties coming from the fact that 𝔾 is assumed to contain SN. One might therefore try to prove that such a highly symmetric state must coincide with the Haar state of SN+ or with that of SN, thereby answering negatively the question.

To tackle this, we focus on the moments of the Haar state, that is, its values on elements of the form x = pi1j1pi2j2...pikjk. These moments satisfy algebraic relations that arise from the ones satisfied by the generators, along with the symmetries induced by SN. Using these, we were able to show that the moments of order up to 5 of such a state must either match the corresponding moments of SN+, or those of SN. Although this is encouraging, our method breaks down at order 6, were we encounter complicated systems of quadratic equations. Nonetheless, there is a glimmer of hope because it may be possible to incorporate an additional ingredient: the fact that the Haar state sends elements of the form x*x to positive real numbers.

I hope that this motivates the reader to take a closer look at these notes and perhaps try their luck at the problem.

Understanding Non-Linearity in Random Matrices

Update: I will give an online talk on my work with Alex in Voiculescu’s Probabilistic Operator Algebra Seminar on Monday, Feb 3, at 9am Pacific time (which is 6pm German time). In order to get the online link for the talk you should write to jgarzav at caltech dot edu.

In the post “Structured random matrices and cyclic cumulants: A free probability approach” by Bernard and Hruza I mentioned the problem about the effect of non-linear functions on random matrices and showed some plots from my ongoing joint work with Alexander Wendel. Finally, a few days ago, we uploaded the preprint to the arXiv.

What is this about? Consider a random matrix ensemble XN which has a limiting eigenvalue distribution for N going to infinity. In the machine learning context applying non-linear functions to the entries of such matrices plays a prominent role and the question arises: what is the asymptotic effect of this operation. There are a couple of results in this direction; see, in particular, the paper by Pennington and Worah and the one of Peche. However, it seems that they deal with quite special choices for XN, like a product of two independent Wishart matrices. In those cases the asymptotic effect of applying the non-linearity is the same as taking a linear combination of the XN with an independent Gaussian matrix. The coefficients in this linear combination depend only on a few quantities calculated from the non-linear function. Such statements are often known as Gaussian equivalence principle.

Our main contribution is that this kind of result is also true much more generally, namely for the class of rotationally invariant matrices. The rotational invariance is kind of the underlying reason for the specific form of the result. Roughly, the effect of the non-linearity is a small deformation of the rotational invariance, so that the resulting random matrix ensemble still exhibits the main features of rotational invariance. This provides precise enough information to control the limiting eigenvalue distribution.

We consider the real symmetric case, that is, symmetric orthogonally invariant random matrices. Similar results hold for selfadjoint unitarily invariant random matrices, or also for corresponding random matrices without selfadjointness conditions.

In addition to what I already showed in the above mentioned previous post, we extended the results now also to a multi-variate setting. This means we can also take, again entrywise, a non-linear function of several jointly orthogonally invariant random matrices. The asymptotic eigenvalue distribution after applying such a function is the same as for a linear combination of the involved random matrices and an additional independent GOE. As an example, take the two random matrices XN=AN2– BN and YN=AN4+CNAN+ANCN, where AN, BN, CN are independent GOEs. Note that, by the appearance of AN in both of them, they are not independent, but have a correlation. Nevertheless, they are jointly orthogonally invariant — that is, the joint distribution of all their entries does not change if we conjugate both of them by the same orthogonal matrix. We consider now the non-linear random matrix max(XN,YN), where we take entrywise the maximum of the corresponding entries in XN and YN. Our results say that asymptotically this should have the same eigenvalue distribution as the linear model aXN+ bYN + cZN, where ZN is an additional GOE, which is independent from AN, BN, CN, and where a, b, c are explicitly known numbers. The next plot superimposes the two eigenvalue distributions.

If you want to know more, in particular, why this is true and how the coefficients in the linear model are determined in terms of the non-linear function, see our preprint Entrywise application of non-linear functions on orthogonally invariant random matrices.

The noncommutative Edmonds’ problem revised: how to calculate the inner rank of matrices in noncommuting variables

Update: The paper has now been published in Foundations of Computational Mathematics as an open access article. Here is a link.

My paper with Johannes and Tobias on the noncommutative Edmonds’ problem has got a total makeover; the new revised version appeared just on the arXiv. It relies now heavily on the recent results of Tobias and mine on the Fuglede-Kadison determinant of matrix-valued semicircular elements. But also, the organization of the paper is now much more reader-friendly – at least we hope so. Everything you need to know can be gotten from the first three sections; if you need proofs, read the rest.

Let me first recall the noncommutative Edmonds’ problem. Consider a matrix

A=a_1\otimes x_1+\ldots+a_n\otimes x_n,

where the x_i are noncommuting variables and the a_i are arbitrary matrices of the same size N; N is fixed, but arbitrary.

The noncommutative Edmonds’ problem asks whether A is invertible over the noncommutative rational functions in the x_i; this is equivalent (by deep results of Cohn) to asking whether A is full, i.e., its noncommutative rank is equal to N. The noncommutative rank is here the inner rank rank(A), i.e., the smallest integer r such that we can write A as a product of an Nxr and an rxN-matrix. This fullness of the inner rank can also be equivalently decided by a more analytic object: to the matrix A we associate a completely positive map

\eta:M_N(\mathbb{C})\to M_N(\mathbb{C}),\qquad \eta(b):=\sum_{i=1}^n a_iba_i^*.

In terms of \eta, the fullness condition for A is equivalent to the fact that \eta is rank non-decreasing (here we have of course the ordinary commutative rank on the complex NxN-matrices).

This noncommutative Edmonds’ problem has become quite prominent in recent years and there are now a couple of deterministic algorithms; see, in particular, the work of Garg, Gurvits, Oliveira and Widgerson.

Let me now recall our noncommutative probabilistic approach to the noncommutative Edmonds’ problem, by replacing the formal variables by concrete operators on infinite-dimensional Hilbert spaces. And a particular nice, and our favorite, choice for the analytic operators are freely independent semicircular variables s_1,\dots,s_n, which are the noncommutative analogues of independent Gaussian random variables. In the case where all the a_i are selfadjoint, the matrix

S=a_1\otimes s_1+\ldots+a_n\otimes s_n

is a matrix-valued semicircular element.

Since one can reduce the general case to the selfadjoint one, it suffices to consider in the following the selfadjoint situation. Then the corresponding S is also a selfadjoint operator, hence its distribution is a probability measure on the real line. By our work of the last ten years or so we know that the invertibility of the formal matrix A over the noncommutative rational functions is equivalent to the invertibility of S as an unbounded operator. But this is equivalent to the question whether S has a trivial kernel, which is equivalent to the question whether its distribution has no atom at zero.

And we know how to calculate the distribution \mu of our matrix-valued semicircular element S. Namely its Cauchy transform g(z), that is, the analytic function

g(z)=\int_{\mathbb{R}} \frac 1{z-t}d\mu(t)

on the complex upper half plane is of the form

g(z)=\frac 1N \text{Tr}(G(z))

where G(z) is given as the unique solution of the matrix-valued equation

zG(z)=1+\eta(G(z)) G(z),

in the lower half-plane of the NxN-matrices. One should note that for each z in the upper complex half plane there is (by some old results of mine with Reza Rashidi Far and Bill Helton) exactly one solution G(z) in the complex lower half-plane of the matrices. And, finally, we know that \mu can have an atom only at zero, and the inner rank of A is related to the mass of this atom; more precisely,

\text{rank}(A)=N(1-\mu(\{0\})).

So it seems, problem solved: Calculate the Cauchy transform and use the Stieltjes inversion formula to get the mass of the distribution at zero. We can even make this very concrete; what we need to consider is

\theta(y):= - y \Im g(iy))

in the limit where the positive real number y tends to zero. This limit is the mass of the atom at zero.

This is nice, but of course we cannot calculate anything analytically here, but have to rely on numerical approximations. So the question is, for which y should we calculate \theta(y) with which accuracy to be able to make a justified statement about the mass of the atom. And this is were things are getting tricky, but also nice.

Consider, for example, the matrix-valued semicircular element

\begin{pmatrix} s_1 + 2s_4& s_1 + s_3 + s_4& s_1 - s_2 + s_4& s_1\\ s_1 + s_3 + s_4& s_1 + s_3& s_1 - s_2 + s_3& s_1 + s_3 - s_4\\ s_1 - s_2 + s_4& s_1 - s_2 + s_3& s_1 - 2s_2& s_1 - s_2 - s_4\\ s_1& s_1 + s_3 - s_4& s_1 - s_2 - s_4& s_1 - 2s_4 \end{pmatrix}

If you want to know how we can get out of our machinery that the atom at zero of its distribution has mass 0.5 (that is, its noncommutative rank is 2), have a look at the revised version of our paper!