Simone Brugiapaglia, Rick Archibald and I have been investigating the robustness of constrained l1-minimization regularization to unknown measurement errors. The vast majority of existing theory for such problems requires an a priori bound on the noise level. Yet in many, if not most, real-world applications such a bound is unknown. Our work provides the first recovery guarantees for a large class of practical measurement matrices, thereby extending existing results for subgaussian random matrices. Besides this, our work sheds new light on sparse regularization in practice, addressing questions about the relationship between model fidelity, noise parameter estimation and reconstruction error.
In it we demonstrate how the structured sparsity of polynomial coefficients of high-dimensional functions can be exploited via weighted l1 minimization techniques. This yields approximation algorithms whose sample complexities are essentially independent of the dimension d. Hence the curse of dimensionality is mitigated to a significant extent.
I am part of the organizing committee for the 1st Biennial Meeting of the SIAM Pacific Northwest Section to be held from Oct 27 to 29, 2017 at Oregon State University. We look forward to seeing you in Corvallis in the Fall!
Frames of Hilbert spaces are ubiquitous in image and signal processing, coding theory and sampling theory. However, they are far less widely known in numerical analysis.
In a new paper with Daan Huybrechs, we take a look at frames from a numerical analyst’s perspective:
First, we point out that frames can be useful tools in numerical analysis where orthonormal bases may be difficult or impossible to construct. Second, we investigate issues concerning stability and accuracy in frame approximations. Our main result is that frame approximations are stable and accurate, provide the function being approximated has representations in the frame with small-norm coefficients.
One application of this work is meshfree approximation of functions on complex geometries using so-called Fourier extensions. Daan maintains a GitHub page with fast algorithms for computing such approximations.
In it, we generalize Rodrigo’s previous work (with Arno Kuijlaars and Nick Trefethen) from equispaced nodes to arbitrary nonequispaced nodes. In particular, our result quantifies the tradeoff between convergence rates and ill-conditioning for nodes distributed according to modified Jacobi weight functions. We also determine a necessary and sufficient sampling rate for stable approximation with polynomial least-squares fitting.
SFU Science wrote a news article about my latest project on MRI reconstruction:
This project is a joint collaboration between me, Rick Archibald, Anne Gelb, Jan Hesthaven, Rodrigo Platte, Guohui Song and Ed Walsh.
I gave a minitutorial at the 2016 SIAM Conference on Uncertainty Quantification on compressed sensing and its application to UQ. My slides are available here:
I have also written a preliminary bibliography of papers relevant to the tutorial: PolynomialCSBibliography
Please feel free to contact me with comments and questions!
My postdoc Il Yong Chun, visiting student Chen Li and I have developed some new and improved recovery guarantees for parallel acquisition systems. These are described in two submitted conference papers:
These new results build on a previous paper of ours which introduced a compressed sensing framework for multi-sensor systems – see this past news item. The new results explain the practical recovery capabilities of a much broader class of sensing scenarios.
Bernard Deconinck (President), Malgorzata Peszynska (Vice President), Leming Qu (Treasurer) and myself (Secretary) have started a new SIAM section for the Pacific Northwest region (SIAMPNWS)! The aim of the section is to bring together communities in the Pacific Northwest that are active in industrial and applied mathematics. If you are in the Pacific Northwest and would like to be involved, please drop us a line.
My postdoc Il Yong Chun and I recently completed a new paper titled Compressed sensing and parallel acquisition. In it we present a theoretical framework for the use of compressed sensing in multi-sensor systems. These systems are found numerous applications, ranging from parallel MRI, to multi-view imaging and generalized sampling theory. Our work provides a first mathematical analysis of the gains offered by parallel acquisition and compressed sensing, and presents new insight into questions of optimal sensor design.