Super-Resolution Compressed Sensing: An Iterative Reweighted Algorithm for Joint Parameter Learning and Sparse Signal Recovery

Jun Fang, Jing Li, Yanning Shen, Hongbin Li, Shaoqian LiIEEE Signal Processing Letters 2014 · DOI · OpenReview

Abstract

In many practical applications such as direction-of-arrival (DOA) estimation and line spectral estimation, the sparsifying dictionary is usually characterized by a set of unknown parameters in a continuous domain. To apply the conventional compressed sensing to such applications, the continuous parameter space has to be discretized to a finite set of grid points. Discretization, however, incurs errors and leads to deteriorated recovery performance. To address this issue, we propose an iterative reweighted method which jointly estimates the unknown parameters and the sparse signals. Specifically, the proposed algorithm is developed by iteratively decreasing a surrogate function majorizing a given objective function, which results in a gradual and interweaved iterative process to refine the unknown parameters and the sparse signal. Numerical results show that the algorithm provides superior performance in resolving closely-spaced frequency components.

1. Introduction

Traditional compressed sensing (CS) relies on the assumption that a signal is sparse within a predefined dictionary. In domains such as radar (Direction-of-Arrival estimation) or audio processing (line spectral estimation), dictionary atoms are defined by continuous parameters like frequency or angle. Standard CS approaches typically discretize these continuous parameters into a fixed grid. However, "grid mismatch"—where the true signal parameter falls between grid points—often leads to significant recovery errors. While finer grids or Taylor expansions can mitigate this, they frequently introduce high computational costs or numerical instability. This paper proposes a method to work directly in the continuous parameter space to achieve super-resolution accuracy.

2. Proposed Iterative Reweighted Algorithm

The authors introduce an Iterative Reweighted Algorithm that treats dictionary parameters as unknown variables to be learned alongside the signal.

2.1 Log-Sum Penalty and Majorization-Minimization

Instead of the standard $\ell_1$-norm, the algorithm utilizes a log-sum sparsity-encouraging functional: $L(z) = \sum \log(|z_i|^2 + \epsilon)$. This functional is more effective at promoting sparsity than the $\ell_1$-norm. Since the log-sum objective is non-convex, a Majorization-Minimization (MM) approach is employed. A quadratic surrogate function—a simpler upper bound—is constructed and iteratively minimized.

2.2 Joint Optimization

The algorithm follows an interweaved process:

  • Weight Update: Weights for signal components are calculated based on the current estimate.
  • Parameter Learning: A gradient-based search refines the continuous parameters (e.g., frequencies) of the dictionary.
  • Signal Recovery: The sparse signal estimate is updated using the refined dictionary.

To avoid local minima, an $\epsilon$-decreasing strategy is used, starting with a larger $\epsilon$ to smooth the objective and gradually reducing it to a very small value.

3. Experimental Evaluation

The algorithm was evaluated using synthetic simulations of complex sinusoid mixtures. Frequencies were randomly generated in the continuous range $[0, 2\pi)$. Performance was measured using Reconstruction Signal-to-Noise Ratio (RSNR) and "Success Rate." Comparisons were made against Spectral Iterative Hard Thresholding (SIHT), Bayesian dictionary refinement (DicRefCS), and Atomic norm minimization (SDP).

3.1 Performance vs. Measurements

The proposed algorithm demonstrates significant superiority in low-measurement regimes. As shown in Figure 1, when the number of measurements ($M$) is small, the proposed method provides an RSNR gain of more than 15dB compared to DicRefCS and SDP methods.

Figure 1: RSNR (a) and Success Rate (b) vs. Number of measurements. The proposed algorithm outperforms baselines, especially with fewer measurements.

3.2 Resolving Closely-Spaced Frequencies

A key strength of the algorithm is its ability to break the Rayleigh limit. In scenarios with extremely close frequency spacing (spacing coefficient $\mu < 1$), the proposed method maintains a high success rate and RSNR, whereas competing methods experience sharp performance drops.

Figure 2: RSNR (a) and Success Rate (b) vs. Frequency spacing coefficient. The proposed algorithm maintains high performance even for very closely spaced components.

4. Conclusion

The proposed iterative reweighted algorithm effectively addresses the grid mismatch problem by jointly learning dictionary parameters and recovering sparse signals. By interweaving these processes, the algorithm achieves super-resolution accuracy and robustness, particularly in challenging scenarios with few measurements or closely spaced signal components. Future work may explore the application of this framework to more complex, non-linear parametric models in other sensing domains.

Summary updated

Abstract

In many practical applications such as direction-of-arrival (DOA) estimation and line spectral estimation, the sparsifying dictionary is usually characterized by a set of unknown parameters in a continuous domain. To apply the conventional compressed sensing to such applications, the continuous parameter space has to be discretized to a finite set of grid points. Discretization, however, incurs errors and leads to deteriorated recovery performance. To address this issue, we propose an iterative reweighted method which jointly estimates the unknown parameters and the sparse signals. Specifically, the proposed algorithm is developed by iteratively decreasing a surrogate function majorizing a given objective function, which results in a gradual and interweaved iterative process to refine the unknown parameters and the sparse signal. Numerical results show that the algorithm provides superior performance in resolving closely-spaced frequency components.
BibTeX
@article{fang2014super,
  author = {Jun Fang and Jing Li and Yanning Shen and Hongbin Li and Shaoqian Li},
  title = {{Super-Resolution Compressed Sensing: An Iterative Reweighted Algorithm for Joint Parameter Learning and Sparse Signal Recovery}},
  journal = {IEEE Signal Processing Letters},
  year = {2014},
  doi = {10.1109/lsp.2014.2316004},
}
Figure