We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.IT

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Information Theory

Title: Data-Time Tradeoffs for Optimal k-Thresholding Algorithms in Compressed Sensing

Abstract: Optimal $k$-thresholding algorithms are a class of $k$-sparse signal recovery algorithms that overcome the shortcomings of traditional hard thresholding algorithms caused by the oscillation of the residual function. In this paper, a novel convergence analysis for optimal $k$-thresholding algorithms is established, which reveals the data-time tradeoffs of these algorithms. Both the analysis and numerical results demonstrate that when the number of measurements is small, the algorithms cannot converge; when the number of measurements is suitably large, the number of iterations required for successful recovery has a negative correlation with the number of measurements, and the algorithms can achieve linear convergence. Furthermore, the main theorems indicate that the number of measurements required for successful recovery is of the order of $k \log({n}/{k})$, where $n$ is the dimension of the target signal.
Comments: 13 pages, 2 figures, accepted by IEEE ISIT 2022
Subjects: Information Theory (cs.IT)
Cite as: arXiv:2110.06460 [cs.IT]
  (or arXiv:2110.06460v3 [cs.IT] for this version)

Submission history

From: Xu Zhang [view email]
[v1] Wed, 13 Oct 2021 02:53:18 GMT (46kb)
[v2] Fri, 28 Jan 2022 03:33:13 GMT (94kb)
[v3] Sun, 19 Jun 2022 08:16:17 GMT (108kb)

Link back to: arXiv, form interface, contact.