Improved high-dimensional estimation with Langevin dynamics and stochastic weight averaging

TL;DR

This paper demonstrates that Langevin dynamics with iterate averaging can recover hidden signals in high-dimensional models at optimal sample complexity, emulating landscape smoothing effects without explicit smoothing.

Contribution

It shows Langevin dynamics with average iterates achieves optimal sample complexity in high-dimensional estimation, matching smoothed landscape results without explicit smoothing.

Findings

Langevin dynamics succeed with n ~ d^{k*/2} samples using average iterates.

Iterate averaging combined with noise mimics landscape smoothing effects.

Potential extension that minibatch SGD may also reach similar rates without added noise.

Abstract

Significant recent work has studied the ability of gradient descent to recover a hidden planted direction $\theta^\star \in S^{d-1}$ in different high-dimensional settings, including tensor PCA and single-index models. The key quantity that governs the ability of gradient descent to traverse these landscapes is the information exponent $k^\star$ (Ben Arous et al., (2021)), which corresponds to the order of the saddle at initialization in the population landscape. Ben Arous et al., (2021) showed that $n \gtrsim d^{\max(1, k^\star-1)}$ samples were necessary and sufficient for online SGD to recover $\theta^\star$, and Ben Arous et al., (2020) proved a similar lower bound for Langevin dynamics. More recently, Damian et al., (2023) showed it was possible to circumvent these lower bounds by running gradient descent on a smoothed landscape, and that this algorithm succeeds with $n \gtrsim d^{\max(1, k^\star/2)}$ samples, which is optimal in the worst case. This raises the question of whether it is possible to achieve the same rate without explicit smoothing. In this paper, we show that Langevin dynamics can succeed with $n \gtrsim d^{ k^\star/2 }$ samples if one considers the average iterate, rather than the last iterate. The key idea is that the combination of noise-injection and iterate averaging is able to emulate the effect of landscape smoothing. We apply this result to both the tensor PCA and single-index model settings. Finally, we conjecture that minibatch SGD can also achieve the same rate without adding any additional noise.