15
38

Tight Query Complexity Lower Bounds for PCA via Finite Sample Deformed Wigner Law

Max Simchowitz
A. Alaoui
Benjamin Recht
Abstract

We prove a \emph{query complexity} lower bound for approximating the top rr dimensional eigenspace of a matrix. We consider an oracle model where, given a symmetric matrix MRd×d\mathbf{M} \in \mathbb{R}^{d \times d}, an algorithm Alg\mathsf{Alg} is allowed to make T\mathsf{T} exact queries of the form w(i)=Mv(i)\mathsf{w}^{(i)} = \mathbf{M} \mathsf{v}^{(i)} for ii in {1,...,T}\{1,...,\mathsf{T}\}, where v(i)\mathsf{v}^{(i)} is drawn from a distribution which depends arbitrarily on the past queries and measurements {v(j),w(i)}1ji1\{\mathsf{v}^{(j)},\mathsf{w}^{(i)}\}_{1 \le j \le i-1}. We show that for every gap(0,1/2]\mathtt{gap} \in (0,1/2], there exists a distribution over matrices M\mathbf{M} for which 1) gapr(M)=Ω(gap)\mathrm{gap}_r(\mathbf{M}) = \Omega(\mathtt{gap}) (where gapr(M)\mathrm{gap}_r(\mathbf{M}) is the normalized gap between the rr and r+1r+1-st largest-magnitude eigenvector of M\mathbf{M}), and 2) any algorithm Alg\mathsf{Alg} which takes fewer than const×rlogdgap\mathrm{const} \times \frac{r \log d}{\sqrt{\mathtt{gap}}} queries fails (with overwhelming probability) to identity a matrix V^Rd×r\widehat{\mathsf{V}} \in \mathbb{R}^{d \times r} with orthonormal columns for which V^,MV^(1const×gap)i=1rλi(M)\langle \widehat{\mathsf{V}}, \mathbf{M} \widehat{\mathsf{V}}\rangle \ge (1 - \mathrm{const} \times \mathtt{gap})\sum_{i=1}^r \lambda_i(\mathbf{M}). Our bound requires only that dd is a small polynomial in 1/gap1/\mathtt{gap} and rr, and matches the upper bounds of Musco and Musco '15. Moreover, it establishes a strict separation between convex optimization and \emph{randomized}, "strict-saddle" non-convex optimization of which PCA is a canonical example: in the former, first-order methods can have dimension-free iteration complexity, whereas in PCA, the iteration complexity of gradient-based methods must necessarily grow with the dimension.

View on arXiv
Comments on this paper