Conjectured log-partition function for the generalized rectangular spherical integral

Let Yϵ=\bY+ϵW\bm{Y}_{\epsilon}=\bY+\sqrt{\epsilon}\,\bm{W} be the generalized observation matrix, with limiting noise singular-value density μϵ\mu_\epsilon, limiting largest singular value γˉϵ\bar{\gamma}_\epsilon, limiting largest singular value of Yϵ\bm{Y}_\epsilon equal to νˉϵ\bar{\nu}_\epsilon, limiting eigenvalue density of \bYϵT\bYϵ\bY_\epsilon^\mathsf{T}\bY_\epsilon equal to ρ^ϵ\hat\rho_\epsilon, and hˉϵ:=limxγˉϵ2Dμϵ(x)\bar{h}_\epsilon:=\lim_{x\downarrow\bar{\gamma}_\epsilon^2}D_{\mu_\epsilon}(x). Define the sticking transition by

λˉϵ:=αlimzνˉϵDμϵ(z).\bar{\lambda}_\epsilon:=\alpha\lim_{z\downarrow\bar{\nu}_\epsilon}D_{\mu_\epsilon}(z).

Log-partition function conjecture. Almost surely,

limn1nlnZn(Yϵ)=limn1nlnIn(λα,Yϵ)=fϵ(α)(λ,λ),\lim_{n\to\infty}\frac{1}{n}\ln Z_n(\bm{Y}_\epsilon)=\lim_{n\to\infty}\frac{1}{n}\ln I_n\left(\sqrt{\frac{\lambda}{\alpha}},\bm{Y}_\epsilon\right)=f_\epsilon^{(\alpha)}(\lambda,\lambda_*),

where fϵ(α)f_\epsilon^{(\alpha)} is the piecewise function specified in the statement, with gλ,ϵ(α)g_{\lambda,\epsilon}^{(\alpha)} as defined there; moreover, n1ElnZn(Yϵ)fϵ(α)(λ,λ)n^{-1}\mathbb{E}\ln Z_n(\bm{Y}_\epsilon)\to f_\epsilon^{(\alpha)}(\lambda,\lambda_*). This conjecture gives the limiting free energy in the high- and low-temperature regimes and is used in the paper to derive the asymptotic MSE of the mismatched Bayes estimator; a rigorous proof is not supplied.

Sources & referencesView supporting material

Primary source

Teng Fu, YuHao Liu, Jean Barbier, Marco Mondelli, ShanSuo Liang and TianQi Hou, “Mismatched estimation of non-symmetric rank-one matrices corrupted by structured noise”, arXiv:2302.03306 (2023).

Progress summary

Never refreshed

Nothing recorded yet. Refresh searches the literature and the public web for attempts on this problem, and writes the first summary here.

Solutions 0

No solutions have been posted yet.