[Paper Review] A null space property approach to compressed sensing with frames

Xuemei Chen, Rongrong Wang+1 more|arXiv (Cornell University)|Feb 28, 2013

Sparse and Compressive Sensing Techniques15 references3 citations

TL;DR

This paper introduces the dictionary-based null space property (D-NSP), the first sufficient and necessary condition for exact recovery of sparse signals in overcomplete dictionaries using ℓ¹-synthesis. It proves that when the dictionary D is full spark, D-NSP is equivalent to AD satisfying standard NSP, implying that highly coherent dictionaries cannot enable stable recovery under this framework, and admissibility is not robust to perturbations.

ABSTRACT

An interesting topic in compressive sensing concerns problems of sensing and recovering signals with sparse representations in a dictionary. In this note, we study conditions of sensing matrices A for the L1-synthesis method to accurately recover sparse, or nearly sparse signals in a given dictionary D. In particular, we propose a dictionary based null space property (D-NSP) which, to the best of our knowledge, is the first sufficient and necessary condition for the success of the L1 recovery. This new property is then utilized to detect some of those dictionaries whose sparse families cannot be compressed universally. Moreover, when the dictionary is full spark, we show that AD being NSP, which is well-known to be only sufficient for stable recovery via L1-synthesis method, is indeed necessary as well.

Motivation & Objective

• To establish a characterization of when ℓ¹-synthesis can exactly recover sparse signals represented in a dictionary D.
• To identify conditions under which the sensing matrix A ensures stable and accurate recovery of signals x₀ = Dz₀ via ℓ¹-minimization.
• To investigate the role of dictionary coherence and spark in determining the admissibility of sensing matrices for ℓ¹-synthesis.
• To clarify the relationship between D-NSP and standard NSP, particularly under full spark assumptions.
• To demonstrate that admissibility of dictionaries is not stable under small perturbations, especially when coherence increases.

Proposed method

• Proposes a new null space property tailored to dictionaries, called D-NSP, defined as: for all v ∈ ker(A) ∖ {0} and all T with |T| ≤ k, ||v_T||₁ < ||v_{T^c}||₁.
• Establishes that D-NSP is both necessary and sufficient for exact recovery of k-sparse signals via ℓ¹-synthesis under noiseless conditions.
• Proves that when D is full spark (every d columns are linearly independent), D-NSP is equivalent to AD satisfying standard NSP.
• Uses a contradiction argument based on ℓ¹-optimality and vector decomposition to show that failure of D-NSP leads to non-unique or incorrect recovery.
• Constructs explicit counterexamples of inadmissible dictionaries of size d×(d+1) to demonstrate non-admissibility under coherence.
• Applies Lemma IV.3 on vector decomposition and norm inequalities to prove that D-NSP fails for certain perturbed dictionaries.

Experimental results

Research questions

• RQ1 What is the necessary and sufficient condition for ℓ¹-synthesis to exactly recover a sparse signal represented in an overcomplete dictionary D?
• RQ2 How does the null space property of the sensing matrix A relate to the structure of the dictionary D?
• RQ3 Under what conditions is the ℓ¹-synthesis method stable and accurate when D is highly coherent or has dependent columns?
• RQ4 Is D-NSP equivalent to standard NSP of AD when D is full spark, and what are the implications for dictionary design?
• RQ5 Is the admissibility of a dictionary robust to small perturbations, especially when coherence increases?

Key findings

• D-NSP is the first known sufficient and necessary condition for exact recovery of sparse signals in dictionaries via ℓ¹-synthesis.
• When D is full spark, A satisfies D-NSP if and only if AD satisfies standard NSP, establishing a tight equivalence.
• For full spark dictionaries, the ℓ¹-synthesis method cannot succeed without also recovering the sparse representation z₀ accurately, contradicting the common belief that signal recovery is easier than coefficient recovery.
• Highly coherent full spark dictionaries are inadmissible, as they violate D-NSP and thus prevent stable recovery.
• Admissibility is not stable under perturbations: even arbitrarily small changes to a perfectly coherent dictionary can result in a full spark, highly coherent, and inadmissible dictionary.
• The paper constructs explicit examples of inadmissible dictionaries of size d×(d+1), and shows that such dictionaries can be extended to arbitrary dimensions via column addition.