Krzakala et al.'s TAP prediction for the free energy of Bayes linear regression

About 4 years old · traced to

Let β=(β1,…,βp)\beta=(\beta_1,\ldots,\beta_p) have product prior π0=π⊗p\pi_0=\pi^{\otimes p}, where π\pi is a probability distribution on [−1,1][-1,1]. Let Zp\mathcal{Z}_p denote the linear-regression partition function, with data matrix XX, observations yy, noise variance Δ\Delta, and aspect ratio α\alpha; write cˉ=p−1∑i=1pci\bar c=p^{-1}\sum_{i=1}^p c_i. For a∈Rpa\in\mathbb{R}^p, define probability measures QiQ_i by

dQidπ(β)∝exp⁡(−(β−Ri)22Σ∗2),\frac{\mathrm{d}Q_i}{\mathrm{d}\pi}(\beta)\propto\exp\left(-\frac{(\beta-R_i)^2}{2\Sigma_*^2}\right),

where RiR_i is chosen so that EQi[β]=ai\mathbb{E}_{Q_i}[\beta]=a_i, and set ci=Var⁡Qi(β)c_i=\operatorname{Var}_{Q_i}(\beta). The parameter Σ∗\Sigma_* depends on Δ\Delta and α\alpha and is the smallest solution of a fixed-point equation. Krzakala et al.'s TAP conjecture. The log-partition function satisfies

log⁡Zp=max⁡∥a∥∞≤1[−∥y−Xa∥22Δ−n2log⁡(1+cˉΔ)+∑i=1plog⁡Eπ[exp⁡(−(βi−Ri)22Σ∗2)]]+o(p).\log \mathcal{Z}_p=\max_{\|a\|_\infty\leq 1}\left[-\frac{\lVert y-Xa\rVert^2}{2\Delta}-\frac{n}{2}\log\left(1+\frac{\bar c}{\Delta}\right)+\sum_{i=1}^p\log\mathbb{E}_{\pi}\left[\exp\left(-\frac{(\beta_i-R_i)^2}{2\Sigma_*^2}\right)\right]\right]+o(p).

The additional logarithmic term is the Onsager correction absent from the naive mean-field approximation. The conjecture predicts that NMF fails when the numbers of samples and features grow proportionally; the stated bounded-support formulation extends directly to priors supported on any bounded set.

References

Primary source

Jiaze Qiu and Subhabrata Sen, “The TAP free energy for high-dimensional linear regression”, arXiv:2203.07539 (2022).

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.