Wishart-prior matrix generalized linear model conjecture

From papers

Let Φd(G)\Phi_d^{(G)} be the free entropy of the Gaussian-equivalent matrix generalized linear model, with \bS\mcWm,d\bS^\star\sim\mcW_{m,d}, and let Q0Q_0, α\alpha, JqJ_q, and Ψ\Psi have the meanings in the model. Define

Ψ(\hq)14+limd1d2E\bYlogE\bS\mcWm,dexp(d4Tr[(\bY\hq\bS)2]),\Psi(\hq)\coloneqq\frac14+\lim_{d\to\infty}\frac1{d^2}\mathbb E_{\bY}\log\mathbb E_{\bS\sim\mcW_{m,d}}\exp\left(-\frac d4\operatorname{Tr}[(\bY-\sqrt{\hq}\bS)^2]\right),

where \bY=\hq\bS+\bxi\bY=\sqrt{\hq}\bS^\star+\bxi, \bxiGOE(d)\bxi\sim\operatorname{GOE}(d), and \bS\mcWm,d\bS^\star\sim\mcW_{m,d}. The free entropy of a matrix generalized linear model. Assuming the limit defining Ψ\Psi is well-defined,

limdΦd(G)=supq[1,Q0]inf\hq0[(Q0q)\hq4+Ψ(\hq)+αR×Rdy\mcDξJq(y,ξ)logJq(y,ξ)].\lim_{d\to\infty}\Phi_d^{(G)}=\sup_{q\in[1,Q_0]}\inf_{\hq\geq0}\left[\frac{(Q_0-q)\hq}{4}+\Psi(\hq)+\alpha\int_{\mathbb R\times\mathbb R}\mathrm dy\,\mcD\xi\,J_q(y,\xi)\log J_q(y,\xi)\right].

This extends the generalized-linear-model free-entropy formula to the dependent Wishart prior; the source identifies technical difficulties in establishing it rigorously, so it remains conjectural.

Progress summary

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

Sources & referencesView supporting material

Primary source

Antoine Maillard, Emanuele Troiani, Simon Martin, Florent Krzakala and Lenka Zdeborová, “Bayes-optimal learning of an extensive-width neural network from quadratically many samples”, arXiv:2408.03733 (2024).

Solutions 0

No solutions have been posted yet.