Solving underdetermined systems with error-correcting codes

TL;DR

This paper presents decoding algorithms for underdetermined systems using error-correcting codes, enabling the recovery of sparse solutions with specific matrix classes like Fourier and Vandermonde matrices, relevant for signal processing.

Contribution

It introduces new decoding algorithms for underdetermined systems based on error-correcting codes, applicable to matrices with certain properties, with efficient complexity.

Findings

Decoding algorithms exist for systems with evenly spaced known measurements.

Algorithms have complexity $O(nt)$ or $O(t^2)$ for specific matrix classes.

Applications include signal processing and compressed sensing.

Abstract

In an underdetermined system of equations $Ax=y$, where $A$ is an $m\times n$ matrix, only $u$ of the entries of $y$ with $u < m$ are known. Thus $E_jw$, called `measurements', are known for certain $j\in J \subset \{0,1,\ldots,m-1\}$ where $\{E_i, i=0,1,\ldots, m-1\}$ are the rows of $A$ and $|J|=u$. It is required, if possible, to solve the system uniquely when $x$ has at most $t$ non-zero entries with $u\geq 2t$. Here such systems are considered from an error-correcting coding point of view. The unknown $x$ can be shown to be the error vector of a code subject to certain conditions on the rows of the matrix $A$. This reduces the problem to finding a suitable decoding algorithm which then finds $x$. Decoding workable algorithms are shown to exist, from which the unknown $x$ may be determined, in cases where the known $2t$ values are evenly spaced (that is, when the elements of $J$ are in arithmetic progression) for classes of matrices satisfying certain row properties. These cases include Fourier $n\times n $ matrices where the arithmetic difference $k$ satisfies $\gcd(n,k)=1$, and classes of Vandermonde matrices $V(x_1,x_2,\ldots,x_n)$ (with $x_i\neq 0$) with arithmetic difference $k$ where the ratios $x_i/x_j$ for $i\neq j$ are not $k^{th}$ roots of unity. The decoding algorithm has complexity $O(nt)$ and in some cases, including the Fourier matrix cases, the complexity is $O(t^2)$. Matrices which have the property that the determinant of any square submatrix is non-zero are of particular interest. Randomly choosing rows of such matrices can then give $t$ error-correcting pairs to generate a `measuring' code $C^\perp=\{E_j | j\in J\}$ with a decoding algorithm which finds $x$. This has applications to signal processing and compressed sensing.