In this paper, we consider a compressed sensing problem of reconstructing a sparse signal from an undersampled set of noisy linear measurements. The regularized least squares or least absolute shrinkage and selection operator (LASSO) formulation is used for signal estimation. The measurement matrix is assumed to be constructed by concatenating several randomly orthogonal bases, referred to as structurally orthogonal matrices. Such measurement matrix is highly relevant to large-scale compressive sensing applications because it facilitates fast computation and also supports parallel processing. Using the replica method from statistical physics, we derive the mean-squared-error (MSE) formula of reconstruction over the structurally orthogonal matrix in the large-system regime. Extensive numerical experiments are provided to verify the analytical result. We then use the analytical result to study the MSE behaviors of LASSO over the structurally orthogonal matrix, with a particular focus on performance comparisons to matrices with independent and identically distributed (i.i.d.) Gaussian entries. We demonstrate that the structurally orthogonal matrices are at least as well performed as their i.i.d. Gaussian counterparts, and therefore the use of structurally orthogonal matrices is highly motivated in practical applications.
Index TermsCompressed sensing, LASSO, orthogonal measurement matrix, the replica method. ‡ C. Wen is with the Institute (negligible) vector addition. The complexity for evaluating the soft-thresholding function η is negligible.This kind of iterative thresholding algorithm requires few computations per-iteration, and therefore enables the application of LASSO in large-scale problems.Much of the theoretical work on (2) has focused on studying how aggressively a sparse signal can be undersampled while still guaranteeing perfect signal recovery. The existing results include those based on the restricted isometry property (RIP) [1,8], polyhedral geometry [9, 10], message passing [11], and the replica method [12][13][14]. Although RIP provides sufficient conditions for sparse signal reconstruction, the results provided by RIP analysis are often conservative in practice. In contrast, using combinational geometry, message passing, or the replica method, it is possible to compute the exact necessary and sufficient condition for measuring the sparsity-undersampling tradeoff performance of (2) in the limit N → ∞. However, the theoretical work largely focused on the case of having a measurement matrix A with independent and identically distributed (i.i.d.) entries. A natural question would be "how does the choice of the measurement matrix affect the typical sparsity-undersampling tradeoff performance?".There are strong reasons to consider different types of measurement matrix. Although the proximal gradient method performs efficiently in systems of medium size, the implementation of (4) will become prohibitively complex if the signal size is very large. This is not only because performing (4) requires matrix-ve...