数学物理学报  2017, Vol. 37 Issue (2): 257-264   PDF    
扩展功能
加入收藏夹
复制引文信息
加入引用管理器
Email Alert
RSS
本文作者相关文章
孙祥凯
不确定信息下凸优化问题的鲁棒解刻划
孙祥凯     
重庆工商大学数学与统计学院 重庆 400067
摘要:该文旨在研究一类不确定性凸优化问题的鲁棒最优解.借助次微分的性质,首先引入了一类鲁棒型次微分约束品性.随后借助此约束品性,刻划了该不确定性凸优化问题的鲁棒最优解.最后建立了该不确定凸优化问题与其对偶问题之间的Wolfe型鲁棒对偶性.
关键词鲁棒解    次微分    不确定凸优化    
Characterizations of Robust Solution for Convex Optimization Problems with Data Uncertainty
Sun Xiangkai     
College of Mathematics and Statistics, Chongqing Technology and Business University, Chongqing 400067
Abstract: In this paper, we consider robust efficient solutions for a convex optimization problem in the face of data uncertainty. By using the properties of the subdifferential, we first introduce a robust-type subdifferential constraint qualification, and then obtain some completely characterizations of the robust efficient solution of this uncertain convex optimization problem. By using the robust-type subdifferential constraint qualification, we also characterize Wolfe type robust duality for the uncertain convex optimization problem and its uncertain dual problem.
Key words: Robust solutions     Subdifferential     Uncertain convex optimization    
1 引言

本文研究如下的凸优化问题

$ (\text{P})~~~\left\{ \begin{align} & ~inf~\ f(x), \text{ } \\ & \text{s}\text{.t}\text{. }x\in X, g(x)\in -K, \\ \end{align} \right. $

其中$X$$Y$均为局部凸的拓扑向量空间, $K\subseteq Y$为非空闭凸锥, $f : X \rightarrow {\Bbb R}$为凸函数, $g:X\rightarrow Y$$K$-凸函数.凸优化问题$\mbox{(P)}$的可行集定义为

$ {\cal F}_0:=\Big\{x\in X:g(x )\in-K\Big\}. $

众所周知, 在过去的几十年里, 此类优化问题模型$(\mbox{P})$已经得到了广泛的深入研究, 如文献[1-8].然而现实生活中, 许多优化问题模型都面临各种不确定性数据的影响, 如交通网络优化问题, 调度问题, 以及网络设计问题等.因此不确定性优化问题 (即问题模型的目标函数或约束函数的取值事先未知) 得到了越来越多学者的重视, 譬如文献[9-16].针对凸优化问题$(\mbox{P})$, 若其约束条件中含有不确定性数据, 则$(\mbox{P})$变为如下的不确定性优化问题

$ ({\rm UP}) ~~~~~~\left\{ \begin{array}{ll} \inf ~ f(x), \\\mbox{ s.t. } x\in X, g(x, v) \in -K, \end{array} \right. $

其中$Z$为局部凸的拓扑向量空间. $v\in {\cal V}$为不确定参数并且${\cal V}\subseteq {Z}$为紧凸的不确定集.本文中, 总是假设$f$为连续的凸函数, $g :X\times Z\rightarrow Y$为连续的$K$-凸凹函数, 即对于任意的$v\in {\cal V}$, $g(\cdot, v)$$K$-凸函数; 对于任意的$x\in X$ $g(x, \cdot)$$K$-凹函数.

目前处理不确定优化问题的方法很多, 其中鲁棒优化方法为一种常用的方法.因此本文将借助鲁棒优化方法刻划不确定凸优化问题$({\rm UP})$的鲁棒有效解.为此, 本文首先引入$\mbox{(UP)}$的鲁棒对应 (robust counterpart)

$ ({\rm RUP}) ~~~~~~\left\{ \begin{array}{ll} \inf ~ f(x), \\\mbox{ s.t. } x\in X, g(x, v) \in -K, \forall v\in {\cal V}. \end{array} \right. $

我们称鲁棒对应$\mbox{(RUP)}$为鲁棒优化问题.显然, 鲁棒优化问题致力于保证最坏的解 (worst-case solution) 不受不确定性数据对该优化问题的干扰.关于鲁棒优化问题的详细介绍, 详见文献[10].

本文将首先引入一类鲁棒型次微分约束品性, 并得到其存在的充分条件.随后建立此鲁棒型次微分约束品性与不确定凸优化问题$\mbox{(UP)}$的鲁棒最优性条件之间的等价关系.最后将建立不确定性凸优化问题$({\rm UP})$与其对偶问题之间的Wolfe型鲁棒对偶性.

2 预备知识

$X$为局部凸的拓扑向量空间, 其对偶空间定义为$X^*$, 并且对其赋予弱星拓扑$\omega(X^*, X)$.记$\langle x^*, x\rangle=x^*(x)$为连续线性函数$x^*\in X^*$$x\in X$处的函数值.设$C$$X$中的非空子集, $C$的闭包和凸包分别定义为$\mbox{cl } C $$\mbox{co } C $.给定集合$D\subseteq X^*$$D\subseteq X^*\times {\Bbb R}$, $\mbox{cl } D $代表$D$的弱星闭包.特别的, 非空子集$C\subseteq X$的指示函数$\delta_C:X\rightarrow {\Bbb {R}}\cup\{+\infty\}$定义为

$ \delta_C(x)= \left \{ \begin{array}{ll} 0, &\mbox{ 若 }\ x\in C, \\ +\infty, & \mbox{ 其它, } \end{array} \right . $

$C$的支撑函数$\sigma_{C}:X^*\rightarrow {\Bbb {R}}\cup\{+\infty\}$定义为$\sigma_{C}(x^*)=\sup_{x\in C}\;\langle x^*, x\rangle.$

给定延拓的实值函数$f:X\rightarrow {\Bbb R}\cup \{+\infty\}$, 其有效域和上图分别定义为

$ \begin{align} & \text{dom }f=\{x\in X:f(x)<+\infty \}, \\ & \text{epi }f=\{(x, r)\in X\times \mathbb{R}:f(x)\le r\}. \\ \end{align} $

$\mbox{dom }f\not=\emptyset$, 则称$f$为真函数.若对于任意的$x, y\in X$以及$\alpha\in[0,1]$,

$ f(\alpha x+(1-\alpha)y)\leq \alpha f(x)+(1-\alpha)f(y), $

则称$f$为凸函数.若$-f$为凸函数, 则称$f$为凹函数.函数$f$的共轭函数$f^*:X^*\rightarrow {\Bbb R}\cup \{+\infty\}$定义为

$ {f^*}({x^*}) = \mathop {\sup }\limits_{x \in X} \;\{ \langle {x^*},x\rangle - f(x)\} . $

显然, 若函数$f$为真的下半连续函数, 则函数$f^*$为真的下半连续函数.

对于任意的$\varepsilon\geq 0, $函数$f$$\bar{x}\in \mbox{dom }f$处的$\varepsilon$-次微分定义为

$ \partial_\varepsilon f(\bar{x}) :=\{x^*\in X^*:f(x)\geq f(\bar{x})+\langle x^*, x-\bar{x}\rangle-\varepsilon, \forall x\in X \}. $

$\bar{x}\not\in \mbox{dom }l$, 则令$\partial_\varepsilon l(\bar{x})=\emptyset$.若$\varepsilon=0$, 则$\partial f(\bar{x}):=\partial_0 f(\bar{x})$退化为经典的次微分, 即

$ \partial f(\bar{x})=\left\{x^*\in X^*:\langle x^*, x-\bar{x}\rangle\leq f(x)- f(\bar{x}), \forall x\in X\right\}. $

特别的, 由文献[17, 命题2.1]可知

$ {\rm{epi}}\;{f^*} = \bigcup\limits_{\varepsilon \ge 0} {\{ (} {x^*},\langle {x^*},\bar x\rangle + \varepsilon - f(\bar x)):{x^*} \in {\partial _\varepsilon }f(\bar x)\} . $

$Y$为另一局部凸向量空间, 其对偶空间亦定义为$Y^*$, 并且赋予其弱星拓扑$\omega(Y^*, Y)$.假设$K\subseteq Y$为非空闭凸锥.其对偶锥定义为

$ K^*=\{y^*\in Y^*:\langle y^*, y\rangle\geq0, \forall y\in K\}. $

给定向量值函数$g:X\rightarrow Y$, 若对于任意的$x, y\in X$以及$ \alpha\in [0,1]$, 有

$ g(\alpha x+(1-\alpha)y)-\alpha g(x)-(1-\alpha)g(y)\in -K, $

则称$g$$K$-凸函数.类似的, 若$-g$$K$-凸函数, 则称$g$$K$-凹函数.

$\lambda \in K^*$. $\lambda$与函数$g$的复合$\lambda\circ g$简记为$\lambda g$.显然, 函数$g$$K$-凸函数当且仅当对于任意的$\lambda \in K^*$, $\lambda g$为凸函数.类似的, 函数$g$$K$-凹函数当且仅当对于任意的$\lambda \in K^*$, $\lambda g$为凹函数.

引理2.1[3]  设函数$f_1, f_2: X\rightarrow {\Bbb {R}}\cup\{+\infty\}$为满足${\rm dom ~} f_1 \; \cap\;{\rm dom~ } f_2 \not=\emptyset$的真凸函数.

(ⅰ) 若$f_1$$f_2$均为下半连续函数, 则

$ {\rm epi ~}{ (f_1+f_2)^* }={\rm cl ~(}{\rm epi ~}{f^*_1 }+{\rm epi ~}{ f^*_2 }). $

(ⅱ) 若$f_1$$f_2$中至少有一个在$\bar{x}\in {\rm dom ~} f_1 \;\cap\;{\rm dom~ } f_2 $处连续, 则

$ {\rm epi ~}{ (f_1+f_2)^* }= {\rm epi ~}{ f^*_1 }+{\rm epi~ }{ f^*_2 } . $
3 鲁棒型次微分约束品性

借助次微分的性质, 本节中将首先引入一类鲁棒型次微分约束品性.随后将建立该鲁棒型次微分约束品性的充分条件.

定义3.1  若对于$x\in {\cal F}$, 有

$ \partial {\delta _\cal F}(x) = \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(x), $

则称鲁棒型次微分约束品性${\rm (RSCQ)}$$x\in {\cal F}$处成立.

注3.1  (ⅰ) 设$x\in {\cal F}$.则易证

$ \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(x) \subseteq \partial {\delta _\cal F}(x). $

事实上, 设

$ \alpha \in \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(x), $

则存在$\lambda\in K^*$, $v\in{\cal V}$以及$(\lambda g)(x, v)=0$使得$\alpha\in\partial \left((\lambda g)(\cdot, v)\right)(x).$因此对于任意的$y\in{\cal F}$, 由次微分的定义可得

$ 0\geq(\lambda g)(y, v)\geq(\lambda g)(x, v)+\langle\alpha, y-x\rangle=\langle\alpha, y-x\rangle. $

从而$\alpha\in\partial\delta_{\cal F}(x)$.因此鲁棒型次微分约束品性$\mbox{(RSCQ)}$$x\in {\cal F}$处成立等价为如下的包含关系

$ ({\rm{RSCQ}})\partial {\delta _\cal F}(x) \subseteq \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(x). $

(ⅱ) 若不确定集${\cal V}$为单点集, 则鲁棒型次微分约束品性${\rm (RSCQ)}$退化为次微分约束品性${\rm (SCQ)}$, 即

$ ({\rm{SCQ}})\partial {\delta _{{\cal F_0}}}(x) = \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},}\\ {(\lambda g)(x) = 0} \end{array}} \partial \left( {\lambda g} \right)(x). $

值得注意的是, 次微分约束品性${\rm (SCQ)}$实际上为文献[3-4]所引入的约束品性${\rm GBCQ_1(0, {\cal A})}$.

下述结论在建立鲁棒型次微分约束品性$\mbox{(RSCQ)}$的充分条件中起着十分重要的作用.

引理3.1[16]  设$f : X\rightarrow {\Bbb {R}}$为连续凸函数, $g: X\times Z\rightarrow Y$为连续的$K$-凸凹函数.若${\cal V}\subseteq Z$并且${\cal F}\not=\emptyset$.则下列结论等价:

(ⅰ) $g(x, v)\in - K, v\in {\cal V}, x\in X$ (i.e., $x\in {\cal F})\Longrightarrow f(x)\geq0$;

(ⅱ) $(0, 0)\in {\rm epi~}f^*+{\rm cl}\Big({\rm co}\bigcup\limits_{v\in {\cal V}, \lambda\in K^*}{\rm ~epi~}\left((\lambda g)(\cdot, v)\right)^*\Big)$.

命题3.1  设$f : X\rightarrow {\Bbb {R}}$为连续凸函数, $g: X\times Z\rightarrow Y$为连续的$K$-凸凹函数.对于$x\in {\cal F}$.若

$ \bigcup\limits_{v \in V,\lambda \in {K^*}} {{\rm{epi}}} {\left( {(\lambda g)( \cdot ,v)} \right)^*} $

为凸弱星闭集, 则鲁棒型次微分约束品性${\rm (RSCQ)}$$x\in {\cal F}$处成立.

  对于$x\in {\cal F}$.设$\alpha\in \partial\delta_{\cal F}(x)$.令$f(x)=\langle-\alpha, y-x\rangle.$则由$\partial\delta_{\cal F}(x)$的定义, 显然有$f(x)\geq0.$因此由引理3.1以及$\bigcup\limits_{v\in {\cal V}, \lambda\in K^*}{\rm ~epi~}\left((\lambda g)(\cdot, v)\right)^*$为凸弱星闭集可得

$ (0,0) \in {\rm{epi}}{f^*} + \bigcup\limits_{v \in V,\lambda \in {K^*}} {{\rm{epi}}} {\left( {(\lambda g)( \cdot ,v)} \right)^*}. $

又因为

$ {\rm epi~}f^*=\{(-\alpha, \langle-\alpha, x\rangle+\varepsilon)| \varepsilon\geq0\} $

以及

$ {\rm{epi}}{\left( {(\lambda g)( \cdot ,v)} \right)^*} = \bigcup\limits_{\epsilon \ge 0} {\{ (} \beta ,\langle \beta ,x\rangle + \epsilon - (\lambda g)(x,v))|\beta \in {\partial _\epsilon }((\lambda g)( \cdot ,v))(x)\} , $

所以存在$v\in {\cal V}$, $\lambda\in K^*$, $\varepsilon\geq0$, $\epsilon\geq0$以及$\beta\in\partial_{\epsilon}((\lambda g)(\cdot, v))(x)$使得

$ (0, 0)=(-\alpha+\beta, \langle-\alpha, x\rangle+\varepsilon+\langle \beta, x\rangle+\epsilon-(\lambda g)(x, v)). $

从而$\alpha\in \partial_{\epsilon}((\lambda g)(\cdot, v))(x)$以及$\varepsilon+\epsilon-(\lambda g)(x, v)=0.$又因为$\varepsilon\geq0$, $\epsilon\geq0$以及$-(\lambda g)(x, v)\geq0$, 所以$\epsilon=\varepsilon=(\lambda g)(x, v)=0.$命题得证.

4 鲁棒最优性条件

本节中将建立鲁棒型次微分约束品性$\mbox{(RSCQ)}$与不确定凸优化问题$\mbox{(UP)}$的鲁棒最优性条件之间的等价关系.下面首先回顾本节中将要用到的一些重要概念.

定义4.1  不确定凸优化问题$({\rm UP}) $的鲁棒可行集定义为

$ {\cal F}:=\Big\{x\in X:g(x, v)\in -K, \forall v\in {\cal V}\Big\}. $

定义4.2  若$\hat{x}\in {\cal F}$为鲁棒对应${\rm (RUP)}$的最优解, 则称$\hat{x}\in {\cal F}$为不确定凸优化问题${\rm (UP)}$的鲁棒最优解.不确定凸优化问题${\rm (UP)}$的所有鲁棒最优解的全体称为不确定凸优化问题${\rm (UP)}$的鲁棒最优解集, 即

$ {\cal S}:=\left\{x\in {\cal F}:f(x)\leq f(y), \forall y\in {\cal F}\right\}. $

定理4.1  对于不确定优化问题${\rm (UP)}$.设$f :X\rightarrow {\Bbb R}$为连续凸函数, $g :X\times Z\rightarrow Y$为连续函数并且对于任意的$v\in Z$, $g(\cdot, v)$$K$-凸函数.则下列结论等价:

(ⅰ) 鲁棒型次微分约束品性${\rm (RSCQ)}$$\bar{x}\in {\cal F}$处成立;

(ⅱ) $\bar{x}$${\rm (UP)}$的鲁棒最优解当且仅当存在$\bar{v}\in {\cal V}$以及$\bar{\lambda}\in K^*$使得

$0\in \partial f(\bar{x})+\partial \left( \left( \bar{\lambda }g \right)(\cdot, v) \right)(\bar{x}), $ (4.1)
$\left(\bar{\lambda} g\right)(\bar{x}, \bar{v})=0.$ (4.2)

   $\mbox{(ⅰ)}\Rightarrow \mbox{(ⅱ)}$:设$\bar{x}$${\rm (UP)}$的鲁棒最优解.则有

$ 0\in\partial (f+\delta_{{\cal F}})(\bar{x}). $

又因为$f$为连续凸函数以及$\delta_{{\cal F}}$为真凸下半连续函数, 所以$ \partial (f+\delta_{{\cal F}})(\bar{x})=\partial f (\bar{x})+\partial \delta_{{\cal F}} (\bar{x}). $从而$ 0\in \partial f (\bar{x})+\partial \delta_{{\cal F}} (\bar{x}). $因此由鲁棒型次微分约束品性${\rm (RSCQ)}$$\bar{x}\in {\cal F}$处成立可得

$ 0 \in \partial f(\bar x) + \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(\bar x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(\bar x), $

即存在$\bar{\lambda}\in K^*$以及$ \bar{v}\in {\cal V}$使得$ 0\in\partial f(\bar{x})+\partial \left(\left(\bar{\lambda} g\right)(\cdot, \bar{v})\right)(\bar{x})$$ \left(\bar{\lambda} g\right)(\bar{x}, \bar{v})=0. $

反之, 假设存在$\bar{\lambda}\in K^*$$ \bar{v}\in {\cal V}$使得 (4.1) 和 (4.2) 式成立.则由 (4.1) 式可知, 存在$\eta\in \partial f(\bar{x})$以及$\xi\in \partial \left(\left(\bar{\lambda} g\right)(\cdot, \bar{v})\right)(\bar{x})$, 使得

$\eta+\xi=0.$ (4.3)

$\eta\in \partial f (\bar{x})$$\xi\in \partial \left(\left(\bar{\lambda} g\right)(\cdot, \bar{v})\right)(\bar{x})$可得对于任意的$ x\in {\cal F}$,

$ f(x )-f(\bar{x})\geq\langle \eta, x-\bar{x}\rangle, ~ \left(\bar{\lambda} g\right)(x, \bar{v})-\left(\bar{\lambda} g\right)(\bar{x}, \bar{v})\geq\left\langle \xi, x-\bar{x}\right\rangle. $

从而对于任意的$ x\in {\cal F}$,

$ f(x)-f(\bar{x})+\left(\bar{\lambda} g\right)(x, \bar{v})-\left(\bar{\lambda} g\right)(\bar{x}, \bar{v})\geq\left\langle \eta+\xi, x-\bar{x}\right\rangle. $

于是由$\left(\bar{\lambda} g\right)(x, \bar{v})\leq0$, $\left(\bar{\lambda} g\right)(\bar{x}, \bar{v})=0$以及 (4.3) 式可知对于任意的$ x\in {\cal F}$,

$ f(x)-f(\bar{x})\geq0. $

因此$\bar{x}$${\rm (UP)}$的鲁棒最优解.

$\mbox{(ⅱ)}\Rightarrow\mbox{(ⅰ)}:$只需证明

$ \partial {\delta _\cal F}(\bar x) \subseteq \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(\bar x,v) = 0} \end{array}} \partial \left( {(\lambda g)( \cdot ,v)} \right)(\bar x). $

事实上, 设$\gamma\in\partial\delta_{\cal F}(\bar{x})$.则对于任意的$x\in {\cal {\cal F}}, $$\langle\gamma, x-\bar{x}\rangle\leq0. $$f(x):=-\langle\gamma, x \rangle$.则对于任意的$x\in {\cal F}$,

$ f(x)=-\langle\gamma, x \rangle\geq -\langle\gamma, \bar{x} \rangle=f(\bar{x}). $

所以$\bar{x}$${\rm (UP)}$的鲁棒最优解.因此由$\mbox{(ⅱ)}$可知, 存在$\bar{\lambda}\in K^*$以及$ \bar{v}\in {\cal V}$使得

$ 0\in\partial f(\bar{x})+\partial \left(\left(\bar{\lambda} g\right)(\cdot, \bar{v})\right)(\bar{x}), \mbox{~} \left(\bar{\lambda} g\right)(\bar{x}, \bar{v})=0. $

又因为$\partial f(\bar{x})=\{-\gamma\}$, 所以$\gamma\in \partial \left(\left(\bar{\lambda} g\right)(\cdot, \bar{v})\right)(\bar{x}) $以及$\left (\bar{\lambda} g\right)(\bar{x}, \bar{v})=0.$从而

$ \gamma \in \bigcup\limits_{\begin{array}{*{20}{c}} {\lambda \in {K^*},v \in V,}\\ {(\lambda g)(\bar x,v) = 0} \end{array}} \partial \left( {\left( {\lambda g} \right)( \cdot ,v)} \right)(\bar x). $

定理得证.

若定理4.1中不确定集${\cal V}$为单点集, 则有如下结果.

推论4.1  对于问题${\rm (P)}$.若$f :X\rightarrow {\Bbb R}$为连续凸函数, $g :X\rightarrow Y$为连续$K$-凸函数.则下列结论等价:

(ⅰ) 次微分型约束品性${\rm (SCQ)}$$\bar{x}\in {\cal F}$处成立;

(ⅱ) $\bar{x}$${\rm (P)}$的最优解当且仅当存在$\bar{\lambda}\in K^*$使得

$ 0\in\partial f(\bar{x})+\partial \left(\bar{\lambda} g\right)(\bar{x}), \qquad \left(\bar{\lambda} g\right)(\bar{x})=0. $
5 型鲁棒对偶性

借助鲁棒优化方法, 本节将首先引入$\mbox{(RUP)}$的Wolfe型鲁棒对偶问题$({\rm OD_W} )$.随后将建立相应的鲁棒弱对偶性以及鲁棒强对偶性.

$y\in X$以及$\lambda\in K^*$.对于固定的$v\in {\cal V}$, 不确定优化问题$\mbox{(UP)}$的经典Wolfe型对偶问题为

$ ({\rm D_{W}}) ~~~~~~ \begin{array}{ll} \max ~~\{ f(y)+(\lambda h)(y, v)\} \\\mbox{ s.t. } ~~ 0\in\partial f( y)+\partial \left(\left( {\lambda} g\right)(\cdot, v)\right)( y), \\ ~~~~~~~~ \lambda\in K^*, y\in X. \end{array} $

不确定对偶问题$({\rm D_{W}})$的最优对应 (optimistic counterpart) 为如下的确定优化问题

$ ({\rm OD_{W}}) ~~~~~~ \begin{array}{ll} \max ~~\{ f(y)+(\lambda h)(y, v)\} \\\mbox{ s.t. } ~~ 0\in\partial f( y)+\partial \left(\left( {\lambda} g\right)(\cdot, v)\right)( y), \\ ~~~~~~~~ \lambda\in K^*, y\in X, v\in {\cal V}. \end{array} $

注5.1  若不确定凸优化问题${\rm (UP)}$的鲁棒对应${\rm (RUP)}$的最优值等于其最优对应$({\rm OD_{W}})$的最优值, 并且$({\rm OD_{W}})$可以达到其最优解, 则称不确定优化问题${\rm (UP)}$与其不确定对偶问题$({\rm D_{W}})$之间的鲁棒强对偶成立.

定理5.1 (鲁棒弱对偶)  对于${\rm (RUP)}$的任意可行解$x$以及$({\rm OD_{W}})$的任意可行解$(y, \lambda, v)$, 总有

$ f(x)\geq f(y)+(\lambda g)(y, v). $

  设$(y, \lambda, v)$$({\rm OD_{W}})$的可行解.则存在$y\in X, $ $ \lambda\in K^*$以及$ v\in {\cal V}$使得

$ 0\in\partial f( y)+\partial \left(\left( {\lambda} g\right)(\cdot, v)\right)( y). $

由上式可知存在$\eta\in \partial f(y)$以及$\xi\in \partial \left(\left({\lambda} g\right)(\cdot, {v})\right)(y)$使得$\eta+\xi=0.$进一步的, 由$\eta\in \partial f(y)$可知

$ f(x)-f(y) \geq\left\langle \eta, x-y\right\rangle=\left\langle -\xi, x-y\right\rangle. $

因此, 由上式以及$\xi\in \partial \left(\left({\lambda} g\right)(\cdot, {v})\right)(y) $可知

$ \begin{align} & f(x)-\{f(y)+(\lambda g)(y, v)\ge -\left\langle \xi, x-y \right\rangle -\left( \lambda g \right)(y, v) \\ & \ge -\left( \lambda g \right)(x, v)+\left( \lambda g \right)(y, v)-\left( \lambda g \right)(y, v) \\ & =-\left( \lambda g \right)(x, v). \\ \end{align} $

又因为$x$$({\rm RUP})$的可行解, 所以$g(x, v)\in -K$.从而$\left({\lambda} g\right)({x}, {v})\leq0.$于是$f(x)\geq f(y)+(\lambda g)(y, v).$定理得证.

定理5.2 (鲁棒强对偶)  设$f :X\rightarrow {\Bbb R}$为连续凸函数, $g :X\times Z\rightarrow Y$为连续函数并且对于任意的$v\in Z$, $g(\cdot, v)$$K$-凸函数.设鲁棒型次微分约束品性${\rm (RSCQ)}$$\bar{x}\in {\cal F}$处成立.如果$\bar{x}$${\rm (UP)}$的鲁棒最优解, 则存在$\bar{v}\in {\cal V}$$\bar{\lambda}\in K^*$使得$(\bar{x}, \bar{\lambda}, \bar{v})$$({\rm OD_{W}})$的鲁棒最优解.

  设$\bar{x}$${\rm (UP)}$的鲁棒最优解.则由定理4.1可知存在$\bar{v}\in {\cal V}$以及$\bar{\lambda}\in K^*$使得

$ 0\in\partial f(\bar{x})+\partial \left(\left(\bar{\lambda} g\right)(\cdot, {\bar{v}})\right)(\bar{x}), ~~ \left(\bar{\lambda} g\right)(\bar{x}, \bar{v})=0. $

所以$(\bar{x}, \bar{\lambda}, \bar{v})$$({\rm OD_{W}})$的可行解.于是由鲁棒弱对偶性可知, 对于$({\rm OD_{W}})$的任意可行解$(y, \lambda, v)$,

$ f(\bar{x})+(\bar{\lambda }g)({\bar{x}}, {\bar{v}})-\{f(y)+(\lambda g)(y, v)\}=f(\bar{x})-\{f(y)+(\lambda g)(y, v)\} \geq 0. $

因此$(\bar{x}, \bar{\lambda}, \bar{v})$$({\rm OD_{W}})$的鲁棒最优解.定理得证.

参考文献
[1] Rockafellar R T. Convex Analysis. Princeton: Princeton University Press, 1970.
[2] Boyd S, Vandemberghe L. Convex Optimization. Cambridge: Cambridge University Press, 2004.
[3] BoţR I. Conjugate Duality in Convex Optimization. Berlin: Springer-Verlag, 2010.
[4] BoţR I, Grad S M, Wanka G. New regularity conditions for strong and total Fenchel-Lagrange duality in infinite dimensional spaces. Nonlinear Anal, 2008, 69: 323–336. DOI:10.1016/j.na.2007.05.021
[5] Fang D H, Li C, Ng K F. Constraint qualifications for extended Farkas's lemmas and Lagrangian dualities in convex infinite programming. SIAM J Optim, 2009, 20: 1311–1332.
[6] Fang D H, Li C, Ng K F. Constraint qualifications for optimality conditions and total Lagrange dualities in convex infinite programming. Nonlinear Anal, 2010, 73: 1143–1159. DOI:10.1016/j.na.2010.04.020
[7] Sun X K. Regularity conditions characterizing Fenchel-Lagrange duality and Farkas-type results in DC infinite programming. J Math Anal Appl, 2014, 414: 590–611. DOI:10.1016/j.jmaa.2014.01.033
[8] 孙祥凯, 柴毅. DC分式规划问题的最优性条件 (英文). 数学进展, 2014, 43: 895–904.
Sun X K, Chai Y. Optimality conditions for DC fractional programming problems (English). Advances in Mathematics (China), 2014, 43: 895–904.
[9] Shapiro A, Dentcheva D, Ruszczynski A.Lectures on Stochastic Programming:Modeling and Theory.Philadelphia:SIAM, 2009
[10] Ben-Tal A, Ghaoui L E, Nemirovski A.Robust Optimization.Princeton:Princeton University Press, 2009
[11] Beck A, Ben-Tal A. Duality in robust optimization:primal worst equals dual best. Oper Res Lett, 2009, 37: 1–6. DOI:10.1016/j.orl.2008.09.010
[12] Jeyakumar V, Li G Y. Strong duality in robust convex programming:complete characterizations. SIAM J Optim, 2010, 20: 3384–3407. DOI:10.1137/100791841
[13] Li G Y, Jeyakumar V, Lee G M. Robust conjugate duality for convex optimization un uncertainty with application to data classification. Nonlinear Anal, 2011, 74: 2327–2341. DOI:10.1016/j.na.2010.11.036
[14] Lee J H, Lee G M. On ε-solutions for convex optimization problems with uncertainty data. Positivity, 2012, 16: 509–526. DOI:10.1007/s11117-012-0186-4
[15] Jeyakumar V, Lee G M, Li G Y. Characterizing robust solution sets of convex programs under data uncertainty. J Optim Theory Appl, 2015, 164: 407–435. DOI:10.1007/s10957-014-0564-0
[16] Sun X K, Chai Y. On robust duality for fractional programming with uncertainty data. Positivity, 2014, 18: 9–28. DOI:10.1007/s11117-013-0227-7
[17] Jeyakumar V, Lee G M, Dinh N. New sequential Lagrange multiplier conditions characterizing optimality without constraint qualifications for convex programs. SIAM J Optim, 2003, 14: 534–547. DOI:10.1137/S1052623402417699