Message Board

Dear readers, authors and reviewers,you can add a message on this page. We will reply to you as soon as possible!

2020 Volume 45 Issue 9
Article Contents

Yong-kai TAO, Jian-wen PENG. A New Proof of Linear Convergence of the Douglas-Rachford Splitting Method[J]. Journal of Southwest China Normal University(Natural Science Edition), 2020, 45(9): 13-18. doi: 10.13718/j.cnki.xsxb.2020.09.003
Citation: Yong-kai TAO, Jian-wen PENG. A New Proof of Linear Convergence of the Douglas-Rachford Splitting Method[J]. Journal of Southwest China Normal University(Natural Science Edition), 2020, 45(9): 13-18. doi: 10.13718/j.cnki.xsxb.2020.09.003

A New Proof of Linear Convergence of the Douglas-Rachford Splitting Method

More Information
  • Corresponding author: Jian-wen PENG
  • Received Date: 01/03/2019
    Available Online: 20/09/2020
  • MSC: O221.2

  • In this paper, the relationship between the proximity operator and the maximal monotone Operator has been combined, and the convex optimization and compression operator theory used to prove that the Douglas-Rachford splitting method with unconstrained optimization problems has global linear convergence under strong convex and smooth functions, and give the corresponding convergence rate.
  • 加载中
  • [1] COMBETTES P L, WAJS V R. Signal Recovery by Proximal Forward-Backward Splitting [J]. Multiscale Modeling Simulation, 2005, 4(4): 1168-1200. doi: 10.1137/050626090

    CrossRef Google Scholar

    [2] PEACEMAN D W, RACHFORD H H. The Numerical Solution of Parabolic and Elliptic Differential Equations [J]. Journal of the Society for Industrial and Applied Mathematics, 1955, 3(1): 28-41.

    Google Scholar

    [3] COMBETTES P L, PESQUET J C. Proximal Splitting Methods in Signal Processing [M]. Berlin: Springer, 2011:185-212.

    Google Scholar

    [4] BOYD S. Distributed Optimization and Statistical Learning via the Alternating Direction Method of Multipliers [J]. Foundations and Trends in Machine Learning, 2010, 3(1): 1-122.

    Google Scholar

    [5] LIONS P L, MERCIER B. Splitting Algorithms for the Sum of Two Nonlinear Operators [J]. SIAM Journal on Numerical Analysis, 1979, 16(6): 964-979. doi: 10.1137/0716071

    CrossRef Google Scholar

    [6] ROCKAFELLAR R T. Monotone Operators and the Proximal Point Algorithm [J]. SIAM Journal on Control and Optimization, 1976, 14(5): 877-898. doi: 10.1137/0314056

    CrossRef Google Scholar

    [7] COMBETTES P L, PESQUET J C. A Douglas-Rachford Splitting Approach to Nonsmooth Convex Variational Signal Recovery [J]. IEEE Journal of Selected Topics in Signal Processing, 2007, 1(4): 564-574. doi: 10.1109/JSTSP.2007.910264

    CrossRef Google Scholar

    [8] LI S J, QI H R. A Douglas-Rachford Splitting Approach to Compressed Sensing Image Recovery Using Low-Rank Regularization [J]. IEEE Transactions on Image Processing, 2015, 24(11): 4240-4249. doi: 10.1109/TIP.2015.2459653

    CrossRef Google Scholar

    [9] STEIDL G, TEUBER T. Removing Multiplicative Noise by Douglas-Rachford Splitting Methods [J]. Journal of Mathematical Imaging and Vision, 2010, 36(2): 168-184. doi: 10.1007/s10851-009-0179-5

    CrossRef Google Scholar

    [10] BRICEÑO-ARIAS L M. A Douglas-Rachford Splitting Method for Solving Equilibrium Problems [J]. Nonlinear Analysis: Theory, Methods & Applications, 2012, 75(16): 6053-6059.

    Google Scholar

    [11] PARIKH N. Proximal Algorithms [J]. Foundations and Trends® in Optimization, 2014, 1(3): 127-239. doi: 10.1561/2400000003

    CrossRef Google Scholar

    [12] DAVIS D, YIN W T. Convergence Rate Analysis of Several Splitting Schemes[M]//Splitting Methods in Communication, Imaging, Science, and Engineering. Berlin: Springer, 2016: 115-163.

    Google Scholar

    [13] GISELSSON P, BOYD S. Linear Convergence and Metric Selection for Douglas-Rachford Splitting and ADMM [J]. IEEE Transactions on Automatic Control, 2017, 62(2): 532-544. doi: 10.1109/TAC.2016.2564160

    CrossRef Google Scholar

    [14] GISELSSON P. Tight Global Linear Convergence Rate Bounds for Douglas-Rachford Splitting [J]. Journal of Fixed Point Theory and Applications, 2017, 19(4): 2241-2270. doi: 10.1007/s11784-017-0417-1

    CrossRef Google Scholar

    [15] GISELSSON P, BOYD S. Diagonal Scaling in Douglas-Rachford Splitting and ADMM[C]//53rd IEEE Conference on Decision and Control. Los Angeles: IEEE Press, 2014: 5033-5039.

    Google Scholar

    [16] BAUSCHKE H H, COMBETTES P L. Convex Analysis and Monotone Operator Theory in Hilbert Spaces[M]. Berlin: Springer, 2011.

    Google Scholar

    [17] ECKSTEIN J, BERTSEKAS D. On the Douglas—Rachford Splitting Method and the Proximal Point Algorithm for Maximal Monotone Operators[J]. Mathematical Programming, 1992, 55(1-3): 293-318. doi: 10.1007/BF01581204

    CrossRef Google Scholar

  • 加载中
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Article Metrics

Article views(1052) PDF downloads(90) Cited by(0)

Access History

Other Articles By Authors

A New Proof of Linear Convergence of the Douglas-Rachford Splitting Method

    Corresponding author: Jian-wen PENG

Abstract: In this paper, the relationship between the proximity operator and the maximal monotone Operator has been combined, and the convex optimization and compression operator theory used to prove that the Douglas-Rachford splitting method with unconstrained optimization problems has global linear convergence under strong convex and smooth functions, and give the corresponding convergence rate.

  • 考虑下列无约束优化问题

    其中:H是实的无穷维希尔伯特空间; $f, g:H\to \mathbb{R}\cup \{+\infty \} $为闭、真凸函数.对于优化问题(1),现在已有多种求解方法[1-5].文献[5]第一次提出寻求单调算子和为零的分裂方法(Douglas-Rachford分裂法)求解如下形式的变分包含问题

    这里$A, B: H \longrightarrow{ }^2{H} $是两个极大单调算子.因为闭的凸函数的次微分是极大单调的[6],根据一阶最优性条件,求解优化问题(1)等价于寻找$\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} \in H $使得$ 0 \in \partial f(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} ) + \partial g(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} )$.近年来,求解优化问题(1)的Douglas-Rachford分裂算法被广泛运用于诸多领域,例如信号处理[7]、压缩感知[8]、图像去噪[9]和平衡问题[10]等方面.

    对于问题(2)的求解,常常是通过算子AB相对应的预解式JAJB进行分离的,算子A预解式JA$ H \to H$定义为

    其中Id为恒等算子,若$ f:H\to \mathbb{R}\cup \{+\infty \}$是凸函数,当$A = \partial f $时,有JA=proxf,这里proxf为临近算子[11],其定义如下

    求解优化问题(1)的Douglas-Rachford分裂算法的迭代公式为

    其中α∈(0,2)通常称为松弛因子,Rf称为反射临近算子,其定义为

    对于凸优化问题(1),可以等价转化成变分包含问题(2).文献[5]证明了若A,B都为极大单调算子,则Douglas-Rachford分裂算法产生的序列弱收敛于最优点; 文献[12]证明了若A,B中有一个算子既是强单调的又是Lipschitz连续的,则Douglas-Rachford分裂算法具有收敛性,并给出了其收敛速度为$ O\left( {\sqrt k } \right)$; 文献[13]利用共轭函数和辅助函数的关系证明了当问题(1)中的一个函数同时具有强凸性和光滑性时,Duglas-Rachford分裂算法产生的迭代序列具有线性收敛性; 紧接着,文献[14]给出了3种情形下Douglas-Rachford分裂法具有全局线性收敛率.

    本文针对问题(1),若$f:H\to \mathbb{R}\cup \{+\infty \} $为光滑强凸函数时,给出了Douglas-Rachford分裂方法的线性收敛性的新证明,并得到不同于文献[13]的线性收敛结果.

1.   预备知识
  • 全文结论建立在实的希尔伯特空间(记作H)上.对H中任意两点xy,其内积用〈xy〉表示,且由内积诱导出的范数定义为$ \parallel \mathit{\boldsymbol{x}}\parallel = \sqrt {\langle \mathit{\boldsymbol{x}}, \mathit{\boldsymbol{x}}\rangle } $.设XH的一个子集,集合X的子集的全体为2XFX上的集值映射$ F: X \longrightarrow 2^{X}$.函数$f:H\to \mathbb{R}\cup \{+\infty \} $xH处的次微分用$\partial f(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} ) $表示,定义$\partial f(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} ) $= $ \{\boldsymbol{u} \mid f(\boldsymbol{y}) \geqslant f(\boldsymbol{x})+\langle\boldsymbol{u}, \boldsymbol{y}-\boldsymbol{x}\rangle, \forall \boldsymbol{y} \in H\}$.

    定义1   (强单调算子)称算子$A: H \longrightarrow 2^{H} $σ -强单调的,如果$\exists \sigma > 0, \forall \mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}} \in H, \quad \forall \mathit{\boldsymbol{u}} \in A\mathit{\boldsymbol{x}}, \mathit{\boldsymbol{v}} \in A\mathit{\boldsymbol{y}} $都有

    注1   当σ=0时,称算子A是单调的.

    定义2    (强制算子)如果$ \exists \beta \ge 0, \forall \mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}} \in H$都有

    则称算子$A: H \longrightarrow 2^{H} $β -强制的.

    定义3    (Lipschitz连续)如果存在常数L≥0,使得

    则称算子A$ H \to H$为L-Lipschitz连续的.称L为Lipschitz常数.当L=1时,称算子A为非扩张的;当L∈[0,1)时,称算子AL-压缩的.

    定义4   如果$\exists \sigma > 0 $,对于$\exists \sigma>0, \text { 对于 } \forall u \in \partial f $都有

    且当f为连续可微函数时有$\partial f=\{\nabla f(\boldsymbol{x})\} $,其中$ \nabla f(\mathit{\boldsymbol{x}})$fx处的梯度,则称函数$f:H\to \mathbb{R}\cup \{+\infty \} $σ-强凸函数.

    显然,f是强凸函数等价于

    其中$\forall \mathit{\boldsymbol{u}} \in \partial f(\mathit{\boldsymbol{x}}), \forall \mathit{\boldsymbol{v}} \in \partial f(\mathit{\boldsymbol{y}}) $.由定义1可知,fσ-强凸函数的充要条件为$ \partial f$σ-强单调算子.

    定义5  如果对于$\forall \mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}} \in H, \beta > 0 $都有

    其中$ \nabla f(\mathit{\boldsymbol{x}})$fx处的梯度,则称函数fβ -光滑凸函数.

    对于不等式(12),交换xy的位置,将两个不等式相加得到fβ -光滑函数的另一种表达形式

    引理1 [15]   若$f, g:H\to \mathbb{R}\cup \{+\infty \} $是连续函数,fσ -强凸函数当且仅当$f-\frac{\sigma}{2}\|\cdot\|^{2} $是凸函数; gβ -光滑函数当且仅当$\frac{\beta}{2}\|\cdot\|^{2}-g $是凸函数.

    引理2 [16]   设fH中的实、闭、真凸函数,下列命题等价:

    1) fβ -光滑函数;

    2) $\nabla f $β -Lipschitz连续的;

    3) $\nabla f$$\frac{1}{\beta } $ -强制的.

    引理3 [16]   若算子$A: H \longrightarrow 2^{H} $是极大单调的,那么算子$ J_{A}, I d-J_{A}: H \longrightarrow H$是严格非扩张极大单调的; 反射预解式${{R}_{A}}:\mathit{\boldsymbol{x}}\to 2{{J}_{A}}\mathit{\boldsymbol{x}}-\mathit{\boldsymbol{x}} $是非扩张的.

    引理4 [16]   若算子$A: H \longrightarrow 2^{H} $σ -强单调的,则预解式JA是(1+σ)-强制的.

2.   Douglas-Rachford分裂法
  • 求解优化问题(1)等价寻找最优点$\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} $使得$ 0 \in \partial f(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} ) + \partial g(\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} )$,因此求解优化问题(1)转化为求解变分包含问题(2).利用Douglas-Rachford分裂法能解决问题(2),任意给定一个初始点z0,Douglas-Rachford分裂算法如下

    其中proxf为临近算子,其定义见(4)式.将(14)式的3式相加,利用Rf的定义得到如下的迭代结果

    其中α∈(0,2),当α=1时,即有${\mathit{\boldsymbol{z}}^{k + 1}} = {R_g}{R_f}{\mathit{\boldsymbol{z}}^k} $,此时产的算法称之为Peaceman-Rachford分裂法[2].由引理3知,RgRf是非扩张的,因此它们的复合RgRf也是非扩张的,故在一般情形下Peaceman-Rachford分裂算法产生的点列不具有收敛性.当$ \alpha = \frac{1}{2}$时,式(15)退化为${\mathit{\boldsymbol{z}}^{k + 1}} = \frac{1}{2}\left( {Id + {R_g}{R_f}} \right){\mathit{\boldsymbol{z}}^k} $,这就是文献[17]中提及的Douglas-Rachford分裂算法.在本文中引入松弛因子α,相比较于Peaceman-Rachford分裂法和特定的Douglas-Rachford分裂法,具有良好的收敛性.

3.   主要结果
  • 定理1   假设fσ -强凸函数、β -光滑函数,且满足$\beta+\frac{1}{\sigma} \geqslant 2 $,则算法(15)迭代产生的Rf$ \sqrt{\frac{\beta \sigma-2 \sigma+1}{\beta+2 \sigma+1}}$ -压缩的.

      因fσ -强凸函数,由式(3)和引理4知proxf是(1+σ)-强制的,令T=proxf,由定义2可知,对$ \forall \mathit{\boldsymbol{x}}, \mathit{\boldsymbol{y}} \in H$

    对式(16)两边同乘以2并减去‖xy2

    由式(6)知$ R_{f}=2 { prox }_{f}-I d=2 T-I d$,整理式(17)可得与之等价的表达

    对式(18)两边同乘以$\frac{2}{\sigma+1} $,并将不等式右侧第一项展开,整理得

    又因fβ -光滑函数,由引理2可知$\nabla f $$ \frac{1}{\beta }$ -强制的.则对uvH

    对式(20)两边同时加‖uv2,整理得

    对式(21),令$\mathit{\boldsymbol{x}} = (\nabla f + Id)\mathit{\boldsymbol{u}}, \mathit{\boldsymbol{y}} = (\nabla f + Id)\mathit{\boldsymbol{v}} $,由式(3),式(4)可知:$\mathit{\boldsymbol{x}} = (\nabla f + Id)\mathit{\boldsymbol{u}} $当且仅当$\mathit{\boldsymbol{u}} = {(\nabla f + Id)^{ - 1}}\mathit{\boldsymbol{x}} = pro{x_f}\mathit{\boldsymbol{x}};\quad \mathit{\boldsymbol{y}} = (\nabla f + Id)\mathit{\boldsymbol{v}} $当且仅当$\mathit{\boldsymbol{v}} = {(\nabla f + Id)^{ - 1}}\mathit{\boldsymbol{y}} = pro{x_f}\mathit{\boldsymbol{y}} $.则式(21)化为

    对式(22)两侧同乘以2再同时减去‖xy2,整理得

    将不等式右侧第一项展开,整理得

    将式(19)加上式(23)再乘以$\frac{\sigma}{\sigma+1} $

    $ \beta \sigma-2 \sigma+1 \geqslant 0$,即$\beta+\frac{1}{\sigma} \geqslant 2 $时,有$ 0 \leqslant \frac{\beta \sigma-2 \sigma+1}{\beta+2 \sigma+1} <1$,则有

    Rf是压缩的.

    定理2  假设fσ -强凸函数、β -光滑函数,$\alpha \in\left(0, 1+\frac{\beta+1}{2 \sigma}-\sqrt{\left(\frac{\beta+1}{2 \sigma}\right)^{2}-1}\right) $,则Douglas-Rachford分裂算法(15)线性收敛到不动点$ \overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} $,且收敛率至少为$|1 - \alpha | + \alpha \sqrt {\frac{{\beta \sigma - 2\sigma + 1}}{{\beta \sigma + 2\sigma + 1}}} $.即有

      由引理3和定理1可知Rg是非扩张的,Rf$\sqrt{\frac{\beta \sigma-2 \sigma+1}{\beta \sigma+2 \sigma+1}} $压缩的,则RgRf也是$\sqrt{\frac{\beta \sigma-2 \sigma+1}{\beta \sigma+2 \sigma+1}} $压缩的.因为对$\forall {\mathit{\boldsymbol{z}}_1}, {\mathit{\boldsymbol{z}}_2} \in H, \left\| {{R_g}{R_f}{\mathit{\boldsymbol{z}}_1} - {R_g}{R_f}{\mathit{\boldsymbol{z}}_2}} \right\| \le \left\| {{R_f}{\mathit{\boldsymbol{z}}_1} - {R_f}{\mathit{\boldsymbol{z}}_2}} \right\| \le \sqrt {\frac{{\beta - 2\sigma + 1}}{{\beta \sigma + 2\sigma + 1}}} \left\| {{\mathit{\boldsymbol{z}}_1} - {\mathit{\boldsymbol{z}}_2}} \right\| $.令$T=(1-\alpha) I d+\alpha R_{g} R_{f} $,则式(15)可化为zk+1=Tzk,因为$ \overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} $是Douglas-Rachford分裂算法(15)的不动点,故有$\overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} = T\overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} $

    其中第一个小于等于关系式是利用三角不等式,第二个小于等于关系式是利用RgRf的压缩性.令$ {\rm{ }}|1 - \alpha | + \alpha \sqrt {\frac{{\beta \sigma - 2\sigma + 1}}{{\beta + 2\sigma + 1}}} < 1$,当α≤1时,显然$|1 - \alpha | + \alpha \sqrt {\frac{{\beta \sigma - 2\sigma + 1}}{{\beta \sigma + 2\sigma + 1}}} < 1 $恒成立; 当α>1时,解得$ {\rm{ }}\alpha < 1 + \frac{{\beta + 1}}{{2\sigma }} - \sqrt {{{\left( {\frac{{\beta + 1}}{{2\sigma }}} \right)}^2} - 1} $.综上,当$ \alpha \in \left( {0, 1 + \frac{{\beta + 1}}{{2\sigma }} - \sqrt {{{\left( {\frac{{\beta + 1}}{{2\sigma }}} \right)}^2} - 1} } \right)$时,Douglas-Rachford分裂算法式(17)线性收敛率至少为$|1-\alpha|+\alpha \sqrt {\frac{{\beta \sigma - 2\sigma + 1}}{{\beta \sigma + 2\sigma + 1}}} $.

    注2   文献[13]的定理1说明当fσ -强凸函数、β -光滑函数,γ∈(0,+∞)时,算法(15)迭代产生的Rf$\max \left(\frac{\gamma \beta-1}{\gamma \beta+1}, \frac{1-\gamma_{\sigma}}{\gamma_{\sigma}+1}\right) $压缩的,而本文的定理1说明了当fσ -强凸函数、β -光滑函数,且$ \beta + \frac{1}{\sigma } \ge 2$时,算法(15)迭代产生的Rf$\sqrt{\frac{\beta \sigma-2 \sigma+1}{\beta \sigma+2 \sigma+1}} $ -压缩的.

    注3   文献[13]的定理2与本文的定理2都说明了在适当条件下,Douglas-Rachford分裂算法(15)产生的序列{zk}是线性收敛的.

    推论1   若fσ -强凸函数、β -光滑函数,设$\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} $是优化问题(1)的解,则利用Douglas-Rachford分裂算法式(16)产生的迭代序列{xk}满足

       fσ -强凸函数,则$ \partial f$σ -强单调算子,可知proxf是(1+σ)-强制的,由定义2及Cauchy-Schwarz不等式可得proxf$\frac{1}{{\sigma + 1}} $ -Lipschitz连续.因$\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} $是优化问题(1)的解,那么有$\overset{\wedge }{\mathop{\mathit{\boldsymbol{x}}}} $=proxf($ \overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} $),其中$ \overset{\wedge }{\mathop{\mathit{\boldsymbol{z}}}} $是Douglas-Rachford分裂算法(15)的不动点.因此,

    第一个不等式是利用proxf$ \frac{1}{{1 + \sigma }}$ -Lipschitz连续的,第二、三个不等式利用定理2的结论,则推论得证.

Reference (17)

Catalog

    /

    DownLoad:  Full-Size Img  PowerPoint
    Return
    Return