How many matrices can be spectrally balanced simultaneously?

Abstract: We prove that any

e l l

positive definite

d i m e s d

matrices,

M_{1}, l d o t s, M_{e} l l

, of full rank, can be simultaneously spectrally balanced in the following sense: for any

k < d

such that

e l l l e q l f l o o r f r a c d - 1 k - 1 f l o o r

, there exists a matrix

A

satisfying

f r a c l a m b d a_{1} (A^{T} M_{i} A) m a t h r m T r (A^{T} M_{i} A) < f r a c 1 k

for all

i

, where

l a m b d a_{1} (M)

denotes the largest eigenvalue of a matrix

M

. This answers a question posed by Peres, Popov and Sousi and completes the picture described in that paper regarding sufficient conditions for transience of self-interacting random walks. Furthermore, in some cases we give quantitative bounds on the transience of such walks.

Review text

In this paper the authors answer a question posed in [\textit{Y. Peres} et al., Bull. Braz. Math. Soc. (N.S.) 44, No. 4, 841--867 (2013; Zbl 1301.60060)]: given a set of real symmetric bilinear positive-definite forms acting on a real \(d\)-dimensional linear space, under which basis of this space are the corresponding matrices as spectrally balanced as possible, in the sense that the ratio between the trace and the operator norm is maximal? That is, a study of the quantity \[ \frac {\|A^T M_i A\| }{{\mathrm{Tr}}\, (A^T M_i A)}, \] where \(M_1, \dots, M_\ell\) are given real positive definite matrices, and \(\|\cdot\|\) is the operator norm, as \(A\) runs through the space of \(d\times d\) real matrices. Since \(A^T M_i A\) is real positive definite, \(\|A^T M_i A\|\) coincides with the largest eigenvalue \(\lambda_1(A^T M_i A)\) of \(A^T M_i A\). Thus, the quantity of interest is indeed \[ \frac {\lambda_1(A^T M_i A)}{{\mathrm{Tr}}\, (A^T M_i A)}. \] The authors establish the following result: Theorem. Let \(k\), \(d\), \(\ell\) be positive integers such that \(k < d\) and \(\ell\leq \lfloor \frac {d-1}{k-1} \rfloor\). For any real positive definite \(d\times d\) matrices \(M_1, \dots, M_\ell\), there exists a real matrix \(A\) satisfying \[ \frac {\lambda_1(A^T M_i A)}{{\mathrm{Tr}}\, (A^T M_i A)} < \frac 1k\quad\text{ for all } i. \] The upper bound \(\frac 1k\) is sharp in the sense that if \(k\), \(\ell\), \(d\) satisfy \((k-1)\ell\geq d\), then there exist \(d\times d\) real positive definite matrices \(M_1, \dots, M_\ell\) such that for every \(d\times d\) real matrix \(A\), there is some \(1\leq i\leq \ell\) that satisfies \[ \frac {\lambda_1(A^T M_i A)}{{\mathrm{Tr}}\, (A^T M_i A)} > \frac 1k . \] The authors then apply the result to complete the picture described in [\textit{Y. Peres} et al., Bull. Braz. Math. Soc. (N.S.) 44, No. 4, 841--867 (2013; Zbl 1301.60060)] regarding sufficient conditions for transience of self-interacting random walks. The results are nice and the proofs are nontrivial and analytic.

Recommendations

Cites work

This page was built for publication: How many matrices can be spectrally balanced simultaneously?

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q1650021)