Identifiability of Complete Dictionary Learning

TL;DR

This paper establishes new deterministic bounds for the sample complexity needed for unique recovery in complete dictionary learning, especially when data has low-rank structure, improving over previous combinatorial bounds.

Contribution

It provides the first non-asymptotic sample bounds for identifiability in deterministic low-rank dictionary learning scenarios, reducing sample complexity from combinatorial to polynomial in r.

Findings

Sample complexity bound of O(r^3/(r-k)^2) for identifiability.

Necessary lower bounds that challenge previous assumptions.

Constant proportion of zeros in B requires only O(r) samples for recovery.

Abstract

Sparse component analysis (SCA), also known as complete dictionary learning, is the following problem: Given an input matrix $M$ and an integer $r$, find a dictionary $D$ with $r$ columns and a matrix $B$ with $k$-sparse columns (that is, each column of $B$ has at most $k$ non-zero entries) such that $M \approx DB$. A key issue in SCA is identifiability, that is, characterizing the conditions under which $D$ and $B$ are essentially unique (that is, they are unique up to permutation and scaling of the columns of $D$ and rows of $B$). Although SCA has been vastly investigated in the last two decades, only a few works have tackled this issue in the deterministic scenario, and no work provides reasonable bounds in the minimum number of samples (that is, columns of $M$) that leads to identifiability. In this work, we provide new results in the deterministic scenario when the data has a low-rank structure, that is, when $D$ is (under)complete. While previous bounds feature a combinatorial term $r \choose k$, we exhibit a sufficient condition involving $\mathcal{O}(r^3/(r-k)^2)$ samples that yields an essentially unique decomposition, as long as these data points are well spread among the subspaces spanned by $r-1$ columns of $D$. We also exhibit a necessary lower bound on the number of samples that contradicts previous results in the literature when $k$ equals $r-1$. Our bounds provide a drastic improvement compared to the state of the art, and imply for example that for a fixed proportion of zeros (constant and independent of $r$, e.g., 10\% of zero entries in $B$), one only requires $\mathcal{O}(r)$ data points to guarantee identifiability.