王鵬,胡潔儀,陳劍波
(五邑大學(xué) 數(shù)學(xué)與計(jì)算科學(xué)學(xué)院,廣東 江門 529020)
低秩矩陣恢復(fù)的近似迭代再加權(quán)算法
王鵬,胡潔儀,陳劍波
(五邑大學(xué) 數(shù)學(xué)與計(jì)算科學(xué)學(xué)院,廣東 江門 529020)
低秩矩陣恢復(fù)問題常常正則化為非凸非光滑的最優(yōu)化問題,并用迭代再加權(quán)算法求解.本文借助跡算子的次梯度提出了一種近似算法,避免了求解線性方程組而直接求解.
低秩矩陣恢復(fù);迭代再加權(quán);SVD分解
低秩矩陣恢復(fù)是近年來計(jì)算數(shù)學(xué)和信息科學(xué)中的一個(gè)熱門研究領(lǐng)域,相關(guān)成果廣泛應(yīng)用于圖像處理、子空間分割、協(xié)同濾波等領(lǐng)域.低秩矩陣恢復(fù)問題一般表述為
本文將上述問題正則化為以下非凸非光滑的優(yōu)化問題:
為了求解式(1),首先引入它的一種光滑近似形式[1]929:
引理1[4]設(shè),其中函數(shù)是正交不變函數(shù),是一個(gè)絕對對稱函數(shù),則在是次可微的,并且
由式(3)并結(jié)合跡算子的凹凸性[5]可知為凹函數(shù),則為凸函數(shù).由次梯度的定義得:
關(guān)于光滑函數(shù)的李普希茲連續(xù)有如下引理:
引理2[6]設(shè)是李普希茲連續(xù)可微函數(shù),李普希茲常數(shù)為L(f),則對于任意的可得:
本部分將給出求解最優(yōu)化問題(2)的近似迭代再加權(quán)算法,并結(jié)合式(6)和(7)給出Xk+1的迭代格式.
我們得到的低秩矩陣恢復(fù)的近似迭代再加權(quán)算法如下:
4)更新權(quán)重Wk+1,,其中,
5)輸出低秩矩陣kX.
對于低秩矩陣恢復(fù)問題,現(xiàn)有算法多將原非凸問題正則化為凸優(yōu)化問題,然而,非凸正則化要比凸正則化效果更好.本文將迭代再加權(quán)算法用于低秩恢復(fù)問題求解,借助跡算子的次梯度避免了求解線性方程組而直接求解.
[1] LAI Mingjun, XU Yangyang, YIN Wotao.Improved iteratively reweighted least squares for unconstrained smoothed lqminimization [J].SIAM Journal on Numerical Analysis, 2013, 51(2): 927-957.
[2] KONISHI K, URUMA K, TAKAHASHI T, et al.Iterative partial matrix shrinkage algorithm for matrix rank minimization [J].Signal Processing, 2014, 100: 124–131.
[3] MOHAN K, FAZEL M.Iterative reweighted algorithms for matrix rank minimization [J].Journal of Machine Learning Research, 2012, 13(1): 3441-3473.
[4] LI Yufan, ZHANG Yanjiao, HUANG Zhenghai.A reweighted nuclear norm minimization algorithm for low rank matrix recovery [J].Journal of Computational & Applied Mathematics, 2014, 263: 338-350.
[5] BEKJAN T N.On joint convexity of trace functions [J].Linear Algebra & Its Applications, 2004, 390: 321-327.
[6] TOH K C, YUN S W.An accelerated proximal gradient algorithm for nuclear norm regularized least squares problems [J].Pacific Journal of Optimization, 2010, 6(3): 615-640.
[責(zé)任編輯:熊玉濤]
Proximal Iteratively Reweighted Algorithm for Low Rank Matrix Recovery
WANG Peng, HU Jie-yi, CHEN Jian-bo
(School of Mathematics and Computational Science, Wuyi University, Jiangmen 529020, China)
Low rank matrix recovery is often regularized into a non-convex and non-smooth optimization problem, which can be solved with iteratively reweighted algorithms.This paper proposes a proximal algorithm with the subgradient of trace operators, and consequently the problem can be solved directly and the system of linear equations can be disposed of.
low rank matrix recovery; iteratively reweighted algorithms; singular value decomposition
O189.1
A
1006-7302(2017)02-0006-03
2017-02-26
五邑大學(xué)青年基金資助項(xiàng)目(2015ZK09);廣東省大學(xué)生創(chuàng)新創(chuàng)業(yè)項(xiàng)目(1134912031).廣東省高等教育教學(xué)改革項(xiàng)目(30527003,JG2014010);廣東省普通高校特色創(chuàng)新類項(xiàng)目(2016GXJK161).
王鵬(1980—),男,廣東江門人,講師,碩士,研究方向矩陣優(yōu)化計(jì)算、數(shù)字圖像處理與分析.