Compressive Sensing and Algebraic Coding: Connections and Challenges

Abstract

Compressive sensing refers to the reconstruction of high dimensional but low-complexity objects from relatively few measurements. Examples of such objects include: high dimensional but sparse vectors, large images with very few sharp edges, and high-dimensional matrices of low rank. One of the most popular methods for reconstruction is to solve a suitably constrained \(\ell _1\)-norm minimization problem, otherwise known as basis pursuit (BP). In this approach, a key role is played by the measurement matrix, which converts the high dimensional but sparse vector (for example) into a low-dimensional real-valued measurement vector. The widely used sufficient conditions for guaranteeing that BP recovers the unknown vector are the restricted isometry property (RIP), and the robust null space property (RNSP). It has recently been shown that the RIP implies the RNSP. There are two approaches for generating matrices that satisfy the RIP, namely, probabilistic and deterministic. Probabilistic methods are older. In this approach, the measurement matrix consists of samples of a Gaussian or sub-Gaussian random variable. This approach leads to measurement matrices that are “order optimal,” in that the number of measurements required is within a constant factor of the optimum achievable. However, in practice, such matrices have no structure, which leads to enormous storage requirements and CPU time. Recently, the emphasis has shifted to the use of sparse binary matrices, which require less storage and are much faster than randomly generated matrices. A recent trend has been the use of methods from algebraic coding theory, in particular, expander graphs and low-density parity-check (LDPC) codes, to construct sparse binary measurement matrices. In this chapter, we will first briefly summarize the known results on compressed sensing using both probabilistic and deterministic approaches. In the first part of the chapter, we introduce some new constructions of sparse binary measurement matrices based on low-density parity-check (LDPC) codes. Then, we describe some of our recent results that lead to the fastest available algorithms for compressive sensing in specific situations. We suggest some interesting directions for future research.