[Paper Review] Recoverability Analysis for Modified Compressive Sensing with Partially Known Support

Jun Zhang, Yuanqing Li+2 more|arXiv (Cornell University)|Jul 8, 2012

Sparse and Compressive Sensing Techniques14 references3 citations

TL;DR

This paper presents a sufficient and necessary condition for exact recovery in modified compressive sensing with partially known support, even when the known support contains errors. It enables explicit computation of recovery probability for sparse signals with ℓ non-zero entries, offering a quantitative measure of reliability under stochastic support uncertainty.

ABSTRACT

The recently proposed modified-compressive sensing (modified-CS), which utilizes the partially known support as prior knowledge, significantly improves the performance of recovering sparse signals. However, modified-CS depends heavily on the reliability of the known support. An important problem, which must be studied further, is the recoverability of modified-CS when the known support contains a number of errors. In this letter, we analyze the recoverability of modified-CS in a stochastic framework. A sufficient and necessary condition is established for exact recovery of a sparse signal. Utilizing this condition, the recovery probability that reflects the recoverability of modified-CS can be computed explicitly for a sparse signal with \ell nonzero entries, even though the known support exists some errors. Simulation experiments have been carried out to validate our theoretical results.

Motivation & Objective

• To address the limitation of existing sufficient conditions for modified-CS, which fail to accurately reflect recoverability when known support contains errors.
• To develop a probabilistic framework for analyzing the recoverability of modified-CS in realistic scenarios with uncertain prior knowledge.
• To derive a sufficient and necessary condition (SNC) for exact recovery that accounts for errors in the known support.
• To enable explicit computation of recovery probability for a sparse signal with ℓ non-zero entries, given a measurement matrix A and a known support T with p₁ errors.
• To provide a quantitative reliability metric for modified-CS in real-world applications such as dynamic MRI, where support evolves slowly but may contain errors.

Proposed method

• Proposes a sufficient and necessary condition (SNC) for exact recovery of a sparse signal x* in modified-CS, based on the structure of the measurement matrix A and the support error set Δe.
• Defines the recovery condition via an optimization problem over all subsets I of the unknown support part Δ, requiring the optimal value of a specific ℓ1-norm difference to be positive.
• Uses the null space property and duality theory to derive the SNC, ensuring that no alternative solution can have a smaller ℓ1-norm on the unknown support part.
• Introduces a probabilistic model where the support T is randomly drawn with p₁ errors, and computes the recovery probability as the fraction of favorable configurations.
• Applies the law of large numbers to estimate the recovery probability from a finite sample of random support configurations, using the expected value of indicator variables over quads of support patterns.
• Derives a closed-form expression for the recovery probability based on combinatorial counts of valid support patterns, involving binomial coefficients and powers of two.

Experimental results

Research questions

• RQ1 What is a sufficient and necessary condition for exact recovery in modified-CS when the known support contains errors?
• RQ2 How can the recovery probability of modified-CS be explicitly computed for a given sparse signal with ℓ non-zero entries and a known support with p₁ errors?
• RQ3 What is the impact of support error rate (p₁/p) on the recovery performance of modified-CS in a stochastic setting?
• RQ4 How does the recovery probability depend on the measurement matrix A, the sparsity level ℓ, and the size of the known support T?
• RQ5 Can the theoretical recovery probability be reliably estimated using a finite sample of random support configurations?

Key findings

• A sufficient and necessary condition (SNC) for exact recovery in modified-CS is derived, which depends on the feasibility and objective value of a specific ℓ1-norm optimization problem over subsets of the unknown support.
• The recovery probability can be computed explicitly for a sparse signal with ℓ non-zero entries, even when the known support T contains p₁ errors, by evaluating the fraction of favorable support configurations.
• The recovery probability is estimated using the law of large numbers, where the expected value of indicator variables over random support quads converges to the true recovery probability for large sample sizes.
• The theoretical recovery probability is expressed as a ratio of the number of recoverable configurations (Sw) to the total number of possible configurations, involving combinatorial terms CₙˡCₗᵖ²Cₙ₋ₗᵖ¹2ˡ⁻ᵖ².
• Simulation results validate the theoretical recovery probability, confirming that the proposed method accurately predicts the performance of modified-CS under support uncertainty.
• The method provides a quantitative reliability index for modified-CS, enabling performance evaluation in real-world applications such as dynamic MRI where support evolves slowly but may be corrupted by errors.