12
7

Identifiability of Complete Dictionary Learning

Abstract

Sparse component analysis (SCA), also known as complete dictionary learning, is the following problem: Given an input matrix MM and an integer rr, find a dictionary DD with rr columns and a matrix BB with kk-sparse columns (that is, each column of BB has at most kk non-zero entries) such that MDBM \approx DB. A key issue in SCA is identifiability, that is, characterizing the conditions under which DD and BB are essentially unique (that is, they are unique up to permutation and scaling of the columns of DD and rows of BB). 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 MM) 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 DD is (under)complete. While previous bounds feature a combinatorial term (rk)r \choose k, we exhibit a sufficient condition involving O(r3/(rk)2)\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 r1r-1 columns of DD. We also exhibit a necessary lower bound on the number of samples that contradicts previous results in the literature when kk equals r1r-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 rr, e.g., 10\% of zero entries in BB), one only requires O(r)\mathcal{O}(r) data points to guarantee identifiability.

View on arXiv
Comments on this paper