An important problem in matrix analysis, completely solved in the early 2000's by A. Knutson & T. Tao (The honeycomb model of GLn(C) tensor products. I. Proof of the saturation conjecture. J. Amer. Math. Soc., 12 (1999), pp 1055–1090), is

Given two $n\times n$ Hermitian matrices $A$ and $B$, describe all the possible spectra of $A+U^* BU$, as $U$ runs over the unitary group.

The answer was conjectured by A. Horn (who is distinct from R. Horn, the co-author of the book with Johnson) in 1962. The possible spectra form in ${\mathbb R}^n$ a polytope defined by a rather intricate list of linear inequalities. The first ones were derived by H. Weyl (1912). Then came Ky Fan's inequalities (1949) and Wielandt's ones (1955).

My question is of a different nature:

What is the distribution of the spectrum of $A+U^* BU$ when $U^* BU$ is taken randomly among the matrices with fixed prescribed spectrum ?

By the way, what is a natural probabilistic framework ? One could take $U$ randomly, according to the Haar measure, but I do not see a clear argument for that.

Also, because the answers to questions about random matrices are often easier when the size $n$ tends to infinity, let me ask the following.

For $n>\!\!>1$, let us prescribe two real spectra $(a_1,\ldots,a_n)$ and $(b_1,\ldots,b_n)$ (multiplicities are allowed). Suppose that the empirical measures $$\mu_n=\frac1n\sum_1^n\delta_{a_j},\qquad\nu_n=\frac1n\sum_1^n\delta_{b_j}$$ converge to probabilities $\mu_\infty$ and $\nu_\infty$. How does the spectrum of $A+B$ behave asymptotically, when the Hermitian matrices $A$ and $B$ are chosen randomly with prescribed spectra $\vec a$ and $\vec b$ ?

  • $\begingroup$ I'm not expert in this field by no means and have a question rather than an insightful comment. Can you provide a link to Knutson & Tao's paper. I wonder if their answer changes if one replaces matrices with two self-adjoint compact operators on a separable Hilbert space. $\endgroup$
    – BigM
    Nov 25 '15 at 15:32
  • 1
    $\begingroup$ @Zoalroshd. For a reference, see my edits. The polytope his an exponentialcomplexity as $n\rightarrow+\infty$. Hence I doubt that it adapts easily to the case of a Hilbert space. $\endgroup$ Nov 25 '15 at 15:50
  • $\begingroup$ I'd like to mention that the initial solution to Horn's conjecture also relies heavily on Klyachko's paper "Stable bundles, representation theory and Hermitian operators", Selecta Math. (N.S.) 4 (1998), no. 3, 419–445, in addition to my paper with Knutson. (There have since been subsequent proofs of Horn's conjecture that avoid the difficult GIT machinery of Klyachko, though.) $\endgroup$
    – Terry Tao
    Nov 25 '15 at 18:25
  • $\begingroup$ Actually, the inequalities in the compact operator case are simply the union of the inequalities in the finite dimensional cases: see for instance the paper of Bercovici, Li, and Timotin in ams.org/mathscinet-getitem?mr=2567501 , together with the references therein. $\endgroup$
    – Terry Tao
    Nov 25 '15 at 18:34

The unitary group $U(n)$ acts on the space $X$ of all $n\times n$ matrices with fixed eigenvalues (and multiplicities, to be precise) by conjugation and this action is transitive. As a result there exists a unique probability measure on $X$ which is invariant under unitary conjugation. Sampling that measure is equivalent to sampling $U A U^*$ with $U$ distributed according to the Haar measure and $A\in X$ a fixed Hermitian matrix.

The asymptotic question is well-understood through the connection with Voiculescu's free probability theory. It is known that (under the assumptions that you make and also under the additional assumption that your measures $\mu_n$ and $\nu_n$ are supported on some compact set $K$ for all $n$) that the spectral measure of $A+B$ almost surely converges to the free additive convolution $\eta = \mu_\infty \boxplus \nu_\infty$. Furthermore, using Voiculescu's $R$-transform, it is possible to compute the Stieltjes transform of $\eta$ in terms of Stieltjes transforms for $\mu_\infty$ and $\nu_\infty$ (what is involved is inverting some analytic functions). In fact, since the matrices $A$ and $B$ become asymptotically freely independent, there is (at least in principle) a way to answer similar question for more complicated expressions (e.g. products, commutators).

The behavior for finite $n$ is also of significant interest. A very recent result of Marcus, Speilman and Srivastava computes precisely the expected value of the characteristic polynomial of $A+B$ in terms of those of $A$ and $B$. This does not quite give what you might want, but gives amazing control on spectral properties of the sum.

  • $\begingroup$ Ronald Speicher has a number of nice surveys on free probability which in particular address the asymptotic version of the original question: mast.queensu.ca/~speicher/survey.html $\endgroup$
    – Terry Tao
    Nov 25 '15 at 18:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.