TY - JOUR
T1 - Multichannel Sparse Blind Deconvolution on the Sphere
AU - Li, Yanjun
AU - Bresler, Yoram
N1 - Funding Information:
Manuscript received May 29, 2018; revised March 18, 2019; accepted June 15, 2019. Date of publication July 15, 2019; date of current version October 18, 2019. This work was supported by the National Science Foundation (NSF) under Grant IIS 14-47879. This paper was presented in part at 2018 NeurIPS [1] and in part at 2019 ICASSP [2].
Publisher Copyright:
© 1963-2012 IEEE.
PY - 2019/11
Y1 - 2019/11
N2 - Multichannel blind deconvolution is the problem of recovering an unknown signal $f$ and multiple unknown channels $x_{i}$ from their circular convolution $y_{i}=x_{i} \circledast f$ ( $i=1,2, {\dots },N$ ). We consider the case where the $x_{i}$ 's are sparse, and convolution with $f$ is invertible. Our nonconvex optimization formulation solves for a filter $h$ on the unit sphere that produces sparse output $y_{i}\circledast h$. Under some technical assumptions, we show that all local minima of the objective function correspond to the inverse filter of $f$ up to an inherent sign and shift ambiguity, and all saddle points have strictly negative curvatures. This geometric structure allows successful recovery of $f$ and $x_{i}$ using a simple manifold gradient descent (MGD) algorithm. The same approach is also applicable to blind gain and phase calibration with a Fourier sensing matrix. Our algorithm and analysis require fewer assumptions than previous algorithms for the same problem. Our theoretical findings are complemented by numerical experiments, which demonstrate superior performance of the proposed approach over the previous methods. Empirically, our algorithm has low computation cost (converging in a small number of iterations) and low memory footprint (solving only for the inverse filter of $f$ ).
AB - Multichannel blind deconvolution is the problem of recovering an unknown signal $f$ and multiple unknown channels $x_{i}$ from their circular convolution $y_{i}=x_{i} \circledast f$ ( $i=1,2, {\dots },N$ ). We consider the case where the $x_{i}$ 's are sparse, and convolution with $f$ is invertible. Our nonconvex optimization formulation solves for a filter $h$ on the unit sphere that produces sparse output $y_{i}\circledast h$. Under some technical assumptions, we show that all local minima of the objective function correspond to the inverse filter of $f$ up to an inherent sign and shift ambiguity, and all saddle points have strictly negative curvatures. This geometric structure allows successful recovery of $f$ and $x_{i}$ using a simple manifold gradient descent (MGD) algorithm. The same approach is also applicable to blind gain and phase calibration with a Fourier sensing matrix. Our algorithm and analysis require fewer assumptions than previous algorithms for the same problem. Our theoretical findings are complemented by numerical experiments, which demonstrate superior performance of the proposed approach over the previous methods. Empirically, our algorithm has low computation cost (converging in a small number of iterations) and low memory footprint (solving only for the inverse filter of $f$ ).
KW - Manifold gradient descent
KW - Riemannian Hessian
KW - Riemannian gradient
KW - nonconvex optimization
KW - strict saddle points
KW - super-resolution fluorescence microscopy
UR - http://www.scopus.com/inward/record.url?scp=85077498936&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85077498936&partnerID=8YFLogxK
U2 - 10.1109/TIT.2019.2928576
DO - 10.1109/TIT.2019.2928576
M3 - Article
AN - SCOPUS:85077498936
SN - 0018-9448
VL - 65
SP - 7415
EP - 7436
JO - IRE Professional Group on Information Theory
JF - IRE Professional Group on Information Theory
IS - 11
M1 - 8762219
ER -