Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Cubic NK-SVD: An Algorithm for Designing Parametric Dictionary in Frequency Estimation

Xiaozhi Liu, Yong Xia

TL;DR

The paper tackles line spectral estimation by reframing it as parametric dictionary learning and introduces Cubic NK-SVD, which embeds cubic-regularized Newton refinements into the K-SVD framework to mitigate grid mismatch in frequency estimation for both SMV and MMV settings. It provides a rigorous convergence analysis within a block coordinate descent paradigm, leveraging higher-order regularization to guarantee descent and convergence to critical points. Empirical results across SMV/MMV DOA estimation and real-world NYUSIM data show that Cubic NK-SVD achieves superior accuracy, particularly for closely spaced frequencies, while maintaining competitive runtimes. The authors also supply code and delineate the potential to extend the framework to broader alternating minimization problems with higher-order regularization.

Abstract

We propose a novel parametric dictionary learning algorithm for line spectral estimation, applicable in both single measurement vector (SMV) and multiple measurement vectors (MMV) scenarios. This algorithm, termed cubic Newtonized K-SVD (NK-SVD), extends the traditional K-SVD method by incorporating cubic regularization into Newton refinements. The proposed Gauss-Seidel scheme not only enhances the accuracy of frequency estimation over the continuum but also achieves better convergence by incorporating higher-order derivative information. A key contribution of this work is the rigorous convergence analysis of the proposed algorithm within the Block Coordinate Descent (BCD) framework. To the best of our knowledge, this is the first convergence analysis of BCD with a higher-order regularization scheme. Moreover, the convergence framework we develop is generalizable, providing a foundation for designing alternating minimization algorithms with higher-order regularization techniques. Extensive simulations demonstrate that cubic NK-SVD outperforms state-of-the-art methods in both SMV and MMV settings, particularly excelling in the challenging task of recovering closely-spaced frequencies. The code for our method is available at https://github.com/xzliu-opt/Cubic-NK-SVD.

Cubic NK-SVD: An Algorithm for Designing Parametric Dictionary in Frequency Estimation

TL;DR

The paper tackles line spectral estimation by reframing it as parametric dictionary learning and introduces Cubic NK-SVD, which embeds cubic-regularized Newton refinements into the K-SVD framework to mitigate grid mismatch in frequency estimation for both SMV and MMV settings. It provides a rigorous convergence analysis within a block coordinate descent paradigm, leveraging higher-order regularization to guarantee descent and convergence to critical points. Empirical results across SMV/MMV DOA estimation and real-world NYUSIM data show that Cubic NK-SVD achieves superior accuracy, particularly for closely spaced frequencies, while maintaining competitive runtimes. The authors also supply code and delineate the potential to extend the framework to broader alternating minimization problems with higher-order regularization.

Abstract

We propose a novel parametric dictionary learning algorithm for line spectral estimation, applicable in both single measurement vector (SMV) and multiple measurement vectors (MMV) scenarios. This algorithm, termed cubic Newtonized K-SVD (NK-SVD), extends the traditional K-SVD method by incorporating cubic regularization into Newton refinements. The proposed Gauss-Seidel scheme not only enhances the accuracy of frequency estimation over the continuum but also achieves better convergence by incorporating higher-order derivative information. A key contribution of this work is the rigorous convergence analysis of the proposed algorithm within the Block Coordinate Descent (BCD) framework. To the best of our knowledge, this is the first convergence analysis of BCD with a higher-order regularization scheme. Moreover, the convergence framework we develop is generalizable, providing a foundation for designing alternating minimization algorithms with higher-order regularization techniques. Extensive simulations demonstrate that cubic NK-SVD outperforms state-of-the-art methods in both SMV and MMV settings, particularly excelling in the challenging task of recovering closely-spaced frequencies. The code for our method is available at https://github.com/xzliu-opt/Cubic-NK-SVD.
Paper Structure (14 sections, 7 theorems, 65 equations, 9 figures, 1 algorithm)

This paper contains 14 sections, 7 theorems, 65 equations, 9 figures, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Cubic NK-SVD Algorithm
  4. Convergence
  5. Simulation Results
  6. Frequency Estimation for SMV Model
  7. Frequency Estimation for MMV Model
  8. Real-World Dataset
  9. Conclusion
  10. Convergence Analysis
  11. Basic Properties
  12. Proof of Lemma \ref{['lem:convergence']}
  13. Proof of Lemma \ref{['lem:gradient_gap']}
  14. Properties of the Function in \ref{['eq:H_doa']}

Key Result

Lemma 1

Suppose that Assumption assum:1 hold. Let $\{\boldsymbol{z}^k\}_{k\in\mathbb{N}}$ be a sequence iteratively generated by eq:x_subproblem and eq:y_subproblem. The following assertions hold. (i) The sequence $\{H(\boldsymbol{z}^k)\}_{k\in\mathbb{N}}$ is nonincreasing, and in particular, for any $k\geq (ii) We have and hence $\lim_{k \rightarrow \infty}\left\|\boldsymbol{z}^{k+1}-\boldsymbol{z}^k\ri

Figures (9)

  • Figure 1: Norm of residuals versus the number of iterations in 10 Monte Carlo trials.
  • Figure 2: RSNRs of respective algorithms. (a) RSNRs vs. $M$, $K=7$ and PSNR=10 dB. (b) RSNRs vs. $K$, $M=32$ and PSNR=10 dB.
  • Figure 3: Success rates of respective algorithms vs. $K$, $M=32$ and PSNR=10 dB.
  • Figure 4: Frequency estimation using different algorithms when $M=24$, $K=3$ and PSNR=0 dB. (a) Ground truth. (b) Cubic NK-SVD. (c) NOMP. (d) ANM. (e) EMaC. (f) OMP.
  • Figure 5: Average running times (sec) of respective algorithms when $M=24$ and PSNR=10 dB.
  • ...and 4 more figures

Theorems & Definitions (15)

  • Remark 1
  • Remark 2
  • Lemma 1
  • proof
  • Lemma 2
  • proof
  • Theorem 3
  • proof
  • Remark 3
  • Lemma 4: cubic_nesterov_06
  • ...and 5 more