Compressive Sensing

Consider a one-dimensional, finite-length signal  $$extract_itex$$x \in \mathbb{C}^N$$/extract_itex$$. We will vectorize a two-dimensional image or higher-dimensional data into a long one-dimensional vectors. Many real-world signals can be well-approximated by sparse or compressible under a suitable basis. Let $$extract_itex$$\Phi=$$\phi_1 \vert \phi_2 \vert \cdots \phi_N$$$$/extract_itex$$ be an orthonormal basis. Then signal $$extract_itex$$x$$/extract_itex$$ can be expressed as $$extract_itex$$\displaystyle x=\sum_{n=1}^N {<x,\phi_N>\phi_N}$$/extract_itex$$. We say that $$extract_itex$$x$$/extract_itex$$ is $$extract_itex$$K$$/extract_itex$$-sparse under $$extract_itex$$\Phi$$/extract_itex$$ if $$extract_itex$${\lbrace s_n=<x,\phi_n>\phi_n \rbrace}_{n=1,\dots,N}$$/extract_itex$$ has only $$extract_itex$$K$$/extract_itex$$-nonzero coefficients, and that $$extract_itex$$x$$/extract_itex$$ is compressible under $$extract_itex$$\Phi$$/extract_itex$$  if $$extract_itex$${\lbrace <x,\phi_n>\phi_n \rbrace}_{n=1,\dots,N}$$/extract_itex$$  has a few large coefficients.

Compressive sensing is that a sparse signal can be recovered from what was previously believed to be incomplete information. Consider $$extract_itex$$\Psi==$$\psi_1 \vert \psi_2 \vert \cdots \psi_N$$ \in \mathbb{C}^{M \times N}$$/extract_itex$$ for some $$extract_itex$$M < M$$/extract_itex$$. Then, we can obtains $$extract_itex$$y=\Psi x = \Psi \Phi f = As$$/extract_itex$$ where $$extract_itex$$A=\Psi \Phi$$/extract_itex$$. The measurements $$extract_itex$$\Psi is fixed and does not depend on the signal$$/extract_itex$$x$$extract_itex$$and then$$/extract_itex$$A$$extract_itex$$is selected independent of$$/extract_itex$$s$$extract_itex$$.$$/extract_itex$$A$$extract_itex$$is referred to as the encoder and obviously encoder is linear. In encoder, we need to design a good sensing matrix$$/extract_itex$$A$$extract_itex$$. The decoder is the attempted recovery of$$/extract_itex$$s$$extract_itex$$from its sensing matrix$$/extract_itex$$A$$extract_itex$$and$$/extract_itex$$y$$extract_itex$$. We define$$/extract_itex$$\Vert s \Vert _0 := \vert$$extract_itex$$supp$$/extract_itex$$s \vert$$extract_itex$$for a signals. The quantity$$/extract_itex$$\Vert \cdot \Vert _0$$extract_itex$$is often called$$/extract_itex$$\ell _0$$extract_itex$$-norm although it is not a norm. With a sparsity prior, a natural decoder is to search for the sparsest vector$$/extract_itex$$s$$extract_itex$$that$$/extract_itex$$y=Asmin \Vert s \Vert _0$$extract_itex$$subject to$$/extract_itex$$y=As$$extract_itex$$. (1) It is known that if the problem (1) has a solution for small$$/extract_itex$$K$$extract_itex$$, the solution is unique. Since decoder is well-defined for small$$/extract_itex$$K$$extract_itex$$, we need an efficient reconstruction algorithm. Unfortunately, the problem (1) is combinatorial problem and NP-hard in general. Essentially two approaches have mainly been pursued : greedy algorithm and convex relaxation. A greedy algorithm computes the support of signal iteratively, at each step finding one or more new elements and subtracting their contribution from the measurement vector. Examples include Matching pursuit(MP), Orthogonal Matching pursuit(OMP), stagewise OMP, regularized OMP, weak OMP. The$$/extract_itex$$\ell_1$$extract_itex$$minimization approach considers the solution of$$/extract_itex$$min \Vert s \Vert_1$$extract_itex$$subject to$$/extract_itex$$y=As$$extract_itex$$.(2) This is a convex optimization problem and can be seen as a convex relaxation of (1). The problem (2) is called “basis pursuit”. In the real-valued case, (2) is casted by a linear program and in the complex-valued case, it is casted by a second order cone program. The theoretical results of compressive sensing is that the solution of basis pursuit (2) equals the sparsest solution under the condition about$$/extract_itex$$A\$ with overwhelm probability.