.png){kind=icon}
.png){kind=icon}
Return
SELECTING REGULARIZATION PARAMETERS FOR NUCLEAR NORM-TYPE MINIMIZATION PROBLEMS
.png){kind=icon}
OA
AI
DOI：10.1137/21M143786X.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The reconstruction of a low-rank matrix from its noisy observation finds its usage in many applications. It can be reformulated into a constrained nuclear norm minimization problem, where the bound eta of the constraint is explicitly given or can be estimated by the probability distribution of the noise. When the Lagrangian method is applied to find the minimizer, the solution can be obtained by the singular value thresholding operator, where the thresholding parameter lambda is related to the Lagrangian multiplier. In this paper, we first show that the Frobenius norm of the discrepancy between the minimizer and the observed matrix is a strictly monotonically increasing function of lambda. From that we derive a closed form solution for lambda in terms of eta. Since lambda is the same as the regularization parameter for the unconstrained regularized problem, our results can be applied to automatically choosing a suitable regularization parameter for the nuclear norm-type regularized minimization problems using the discrepancy principle. The regularization parameters obtained are comparable to (and sometimes better than) those obtained by Stein's unbiased risk estimator approach, while the cost of solving the minimization problem can be reduced by 11-18 times. Numerical experiments with both synthetic data and real MRI data are performed to validate the proposed approach.
Keywords:
low-rank matrix
singular value thresholding
discrepancy principle
nuclear norm
regularization parameter
Journal
IF:
2.6
Papers:
5.1K
Citations:
1.8W
Organization
View more organizations