数学杂志  2023, Vol. 43 Issue (4): 336-346   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
史小波
高英
李林廷
吴春
拟凸向量值映射次微分性质及优化问题的最优性条件
史小波, 高英, 李林廷, 吴春    
重庆师范大学数学科学学院, 重庆 401331
摘要:本文研究了拟凸向量值映射的次微分及其拟凸向量优化问题的最优性条件.首先,引进恰当K-拟凸的概念,并利用Δ函数对其进行标量化,得到恰当K-拟凸的等价刻画.然后,给出拟凸向量值映射的四种次微分的定义,并研究了它们的性质.最后,利用拟凸向量值映射的次微分研究拟凸向量优化问题弱有效解的最优性条件,并用例子说明其合理性.
关键词拟凸向量值映射    次微分    弱有效解    最优性条件    
SUBDIFFERENTIAL PROPERTIES OF QUASICONVEX VECTOR VALUED MAPS AND OPTIMALITY CONDITIONS FOR OPTIMIZATION PROBLEMS
SHI Xiao-bo, GAO Ying, LI Lin-ting, WU Chun    
School of Mathematical Sciences, Chonqing Normal University, Chongqing 401331, China
Abstract: In this paper we study the subdifferentiation of quasiconvex vector valued mapping and the optimality conditions of quasiconvex vector optimization problems. Firstly, we introduce the concept of proper K-quasiconvex, scalar it by using Δ function, and obtain the equivalent characterization of proper K-quasiconvex. Then, four subdifferential definitions of quasi convex vector valued mappings are given and their properties are studied. Finally, the optimality conditions of weak efficient solutions for quasiconvex vector optimization problems are studied by using the subdifferential of quasiconvex vector valued mappings, and an example is given to illustrate its rationality.
Keywords: quasiconvex vector valued mapping     subdifferential     weak efficient solution     optimality condition    
1 引言

向量值映射的研究一直是研究者们高度关注并深入研究的重要课题. 目前, 各种广义锥凸性的概念已得到了充分的认识和广泛的应用. 许多领域都与各种锥凸性概念联系紧密, 如: 1985年, Jahn [1]给出的分离定理; 1999年, Craven [2]给出的择一定理等. 在这其中, 锥拟凸性的概念是最重要的锥凸性概念之一, 1989年Luc [3]和Ferro [4]分别提出了向量值映射的锥拟凸性和恰当锥拟凸性的定义. 由于推广途径不一, 许多文献提出了不同的锥拟凸向量概念, 主要包括: 锥拟凸、恰当锥拟凸、自然锥拟凸、标量锥拟凸等. 2012年, Lin [5]给出了几种不同锥拟凸性之间的关系. 尽管给出了拟凸向量值的多种锥拟凸性的概念, 但是对于它的次微分及其他在拟凸向量优化中的应用基本是处于空白的, 因此研究拟凸向量值映射的次微分以及它在拟凸向量优化问题中的应用是非常有必要的.

此外, 虽然直接研究和求解锥拟凸优化问题存在巨大困难, 但是我们可以使用标量化方法将向量优化问题转化为标量优化问题. 到目前为止, 最为常用的两个非线性标量化函数是Gerstewitz和$ \Delta $函数. $ \Delta $函数又称为径向距离函数, 1979年由Hiriart-Urnity [6]提出, Zaffaroni [7]给出了其详细的分析及应用, 包括重要的分离性质、次线性性质、单调性及连续性. 因此, 我们可以利用$ \Delta $函数对拟凸向量优化问题进行标量化, 再利用已有数值优化问题的结果, 给出拟凸向量优化问题的研究.

2 预备知识

$ X $为实向量空间, $ Y $为赋范向量空间, $ K\subset Y $为内部非空的凸锥.

空间$ Y $上的偏序由$ K $确定: 对任何的$ y, z\in Y $

$ \begin{array}{lll} &&y<_{K}z\Leftrightarrow z-y\in intK;\; y\nless_{K}z\Leftrightarrow z-y\notin intK, \\ &&y\leqq_{K}z\Leftrightarrow z-y\in K;\; y\nleqq_{K}z\Leftrightarrow z-y\notin K, \\ &&y\leq_{K}z\Leftrightarrow z-y\in K\backslash\{0\};\; y\nleq_{K}z\Leftrightarrow z-y\notin K\backslash\{0\}.\\ \end{array} $

$ R^{n} $表示$ n $维欧几里德空间. $ R_{+}^{n} $$ R_{++}^{n} $分别表示$ R^{n} $的非负象限和正象限, 即

$ \begin{array}{ll}R_{+}^{n}=\{(x_{1}, \ldots, x_{n})^{\mathrm{T}}: x_{i}\geq 0, i=1, \ldots, n\}, \\R_{++}^{n}=\mathrm {int}R_{+}^{n}=\{(x_{1}, \ldots, x_{n})^{\mathrm{T}}: x_{i}>0, i=1, \ldots, n\}. \end{array} $

$ x=(x_{1}, \cdots, x_{n})^{T}, y=(y_{1}, \cdots, y_{n})^{T}\in{R^{n}} $, 定义向量$ x $, $ y $的序关系:

$ \begin{array}{llll} x< y\; \; \; \mathrm{\Longleftrightarrow}\; \; \; y-x\in{\mathrm {int}R_{+}^{n}};\\ x\leq y\; \; \; \mathrm{\Longleftrightarrow}\; \; \; y-x\in{R_{+}^{n}}\setminus\{0\};\\ x\leqq y\; \; \; \mathrm{\Longleftrightarrow}\; \; \; y - x \in{R^{n}_{+}};\\ x \nless y \; \; \; \mathrm{\Longleftrightarrow}\; \; \; y-x\notin{\mathrm {int}R_{+}^{n}};\\ x \nleq y \; \; \; \mathrm{\Longleftrightarrow}\; \; \; y-x\notin{R_{+}^{n}}\setminus\{0\}. \end{array} $

$ \langle x, y\rangle $$ x^{T}y $都表示$ R^{n} $中的向量$ x=(x_{1}, \cdots, x_{n})^T $$ y=(y_{1}, \cdots, y_{n})^{T}\in{R^{n}} $的内积, 即

$ \langle x, y\rangle=x^{T}y=\sum\limits_{i=1}^{n}x_{i}y_{i}. $

设非空集合$ C\subset R^{n} $, $ clC, coneC $分别表示$ C $的闭包和锥包. 对任意的$ \overline{x}\in cl C $, $ C $$ \overline{x} $的切锥和法锥分别定义为

$ \begin{array}{lll} &&T(C, \bar{x})=\{d\in R^{n}:\exists \{x_{i}\}\subset C, t_{i}\downarrow 0, s.t.\; \; \lim\limits_{i\rightarrow \infty}\frac{x_{i}-\overline{x}}{t_{i}}= d\}, \\ &&N(C, \overline{x})=\{d\in R^{n}:d^{T}\overline{x}\leq0, \forall \overline{x}\in T(C, \bar{x})\}. \end{array} $

特别地, 当$ C $是凸集时, 法锥退化为$ N(C, \overline{x})=\{x^{*}\in R^{n}:\langle x^{*}, x-\overline{x}\rangle\leq0, \forall x\in C\}. $此外, 文献[16] 中定义了如下的$ \varepsilon $-法锥

$ {N_\varepsilon }(C, {x_0}) = \left\{ {x_0^* \in {R^n}:\langle {x_0^*, x - {x_0}} \rangle \le \varepsilon , \forall x \in C} \right\}. $

定义2.1[8]  设$ \varphi:C\rightarrow R $. 若对任意的$ x_{1}, x_{2}\in \Omega, \; \lambda\in[0, 1] $

$ \varphi(\lambda x_{1}+(1-\lambda)x_{2})\leq \max\{\varphi(x_{1}), \varphi(x_{2})\}.\; \nonumber $

则称$ \varphi(x) $$ C $上的拟凸函数.

定义2.2[9]-[12]  函数$ \varphi:C\rightarrow \overline{R}=R\cup\{ \pm \infty\} $$ \overline{x}\in \text{dom}\; \varphi $处的Greenberg-Pierskalla次微分定义为

$ x^{*}\in\partial^{*}\varphi(\overline{x})\Longleftrightarrow\langle x^{*}, x-\overline{x}\rangle< 0, \forall x\in S_{\varphi}^{<}(\overline{x}), $

星型次微分为

$ x^{*}\in\partial^{\circledast}\varphi(\overline{x})\Longleftrightarrow\langle x^{*}, x-\overline{x}\rangle\leq 0, \forall x\in S_{\varphi}^{<}(\overline{x}), $

Gutiérrez次微分定义为

$ x^{*}\in\partial^{\leq}\varphi(\overline{x})\Longleftrightarrow\langle x^{*}, x-\overline{x}\rangle\leq \varphi(x)-\varphi(\overline{x}), \forall x\in S_{\varphi}^{\leq}(\overline{x}), $

Plastria下次微分为

$ x^{*}\in\partial^{<}\varphi(\overline{x})\Longleftrightarrow\langle x^{*}, x-\overline{x}\rangle\leq \varphi(x)-\varphi(\overline{x}), \forall x\in S_{\varphi}^{<}(\overline{x}). $

其中, $ dom\; \varphi=\{x\in R^{n}:\varphi(x)<+\infty\} $, $ S_{\varphi}^{<}(\overline{x})=\{x\in C:\varphi(x)< \varphi(\overline{x})\}, S_{\varphi}^{\leq}(\overline{x})=\{x\in C:\varphi(x)\leq \varphi(\overline{x})\} $.

2000年, Penot [13]给出了Greenberg-Pierskalla次微分, Gutiérrez次微分, 星型次微分, Plastria下次微分以及这四种次微分之间的关系.

$ \begin{array}{lll} &&\partial^{\leq}\varphi(\bar{x})\subset\partial^{<}\varphi(\bar{x})\subset\partial^{\ast}\varphi(\bar{x})\subset\partial^{\circledast}\varphi(\bar{x}). \end{array} $

定义2.3[14]  设$ f:X\rightarrow Y $为向量值映射, 若对任意的$ x_{1}, x_{2}\in X, \lambda\in[0, 1] $

$ \begin{eqnarray} f(\lambda x_{1}+(1-\lambda)x_{2})\leqq_{K}f(x_{1}) \text{或者} f(\lambda x_{1}+(1-\lambda)x_{2})\leqq_{K}f(x_{2}).\; \end{eqnarray} $

$ f(x) $$ X $上是恰当$ K $-拟凸的.

引理2.1[14]  设$ f:X\rightarrow Y $为向量值映射, 则$ f(x) $$ X $上是恰当$ K $-拟凸的当且仅当集合$ \{x\in X:f(x)\ngeqq_{K}y\} $是凸集, $ \forall y\in Y. $

定义2.4[14]  设$ g:Y\rightarrow R, $对任意的$ y, z\in Y $, 如果$ y\leqq_{K}z\Rightarrow g(y)\leq g(z), \; \nonumber $则称$ g $是单调的; 如果$ y<_{K}z\Rightarrow g(y)< g(z), \; \nonumber $则称$ g $是严格单调的.

2003年, Zaffaroni在文献[7] 中研究了如下一类非线性标量化函数.

定义2.5[7]  径向距离函数$ \Delta_{-K}:Y\rightarrow R $定义为$ \Delta_{-K}(y)=d_{-K}(y)-d_{Y\setminus-K}(y), y\in Y, \; $其中$ d_{A}(y)=\inf\limits_{a\in A}\|y-a\|, y\in Y.\; \nonumber $

引理2.2[7]  设$ y, z\in Y, \; K\neq Y $, 则下面的叙述成立.

(ⅰ) $ \Delta_{-K}(y)<0\Leftrightarrow y\in -intK;\Delta_{-K}(y)=0\Leftrightarrow y\in -bdK;\Delta_{-K}(y)>0\Leftrightarrow y\notin -K $;

(ⅱ) 若$ y<_{K}z $, 则$ \Delta_{-K}(y)<\Delta_{-K}(z) $;

(ⅲ)若$ K $是一个闭集, $ \Delta_{-K}(\cdot) $是次线性函数;

(ⅳ) $ \Delta_{-K}(\cdot) $是正齐次函数.

3 拟凸向量值映射的性质

李飞在文献[14] 中利用非线性标量化函数$ \eta_{e} $对恰当$ K $-拟凸向量值映射进行标量化处理, 得到一系列结论. 由于径向距离函数也是一种非线性标量化函数, 因此我们在本节首先考虑利用径向距离函数$ \Delta_{-K} $对恰当$ K $-拟凸向量值映射进行标量化刻画.

定理3.1  设$ x\in X, \; f:X\rightarrow Y $为向量值映射, $ f(x) $$ X $上是恰当$ K $-拟凸的当且仅当$ (\Delta_{-K}\circ f)(x) $是拟凸的.

  先证若$ f(x) $$ X $上是恰当$ K $-拟凸的, 则$ (\Delta_{-K}\circ f)(x) $是拟凸的.

要证$ (\Delta_{-K}\circ f)(x) $是拟凸, 只需证对任意的$ \alpha\in R, \; x\in X $, 水平集$ L_{\alpha}=\{(\Delta_{-K}\circ f)(x)\leq\alpha\} $是凸集. 对任意的$ x_{1}, x_{2}\in L_{\alpha}, \lambda\in[0, 1] $, 设$ (\Delta_{-K}\circ f)(x_{1})\leq\alpha, (\Delta_{-K}\circ f)(x_{2})\leq\alpha $. 由$ f(x) $是恰当$ K $-拟凸的定义, 不妨假设$ f(\lambda x_{1}+(1-\lambda)x_{2})\leqq_{K}f(x_{1}), $所以$ f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})\in -K $, 则

$ \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})]\leq0.\; \nonumber $

现只需要证明

$ \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})] - \Delta_{-K}[f(x_{1})]\leq \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})], \; \nonumber $

$ \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})] \leq \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})]+ \Delta_{-K}[f(x_{1})].\; \nonumber $

$ \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})+f(x_{1})]\leq \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})-f(x_{1})]+ \Delta_{-K}[f(x_{1})], $

从而

$ \Delta_{-K}[f(\lambda x_{1}+(1-\lambda)x_{2})]\leq\Delta_{-K}[f(x_{1})]\leq\alpha.\; \nonumber $

因此$ \lambda x_{1}+(1-\lambda)x_{2}\in L_{\alpha} $, 即$ L_{\alpha} $是凸集.

下面证明若$ (\Delta_{-K}\circ f)(x) $是拟凸的, 则$ f(x) $$ X $上是恰当$ K $-拟凸的.

反证, 假设$ f(x) $不是恰当$ K $-拟凸的, 则存在$ x_{1}, x_{2}, \lambda\in[0, 1] $$ f(\lambda x_{1}+(1-\lambda)x_{2})\nleqq_{K}f(x_{1}), \text{且} f(\lambda x_{1}+(1-\lambda)x_{2})\nleqq_{K}f(x_{2}), $

$ \begin{align} f(x_{1})-f(\lambda x_{1}+(1-\lambda)x_{2})\notin K, f(x_{2})-f(\lambda x_{1}+(1-\lambda)x_{2})\notin K. \end{align} $ (3.1)

$ (\Delta_{-K}\circ f)(x) $是拟凸可知

$ (\Delta_{-K}\circ f)(\lambda x_{1}+(1-\lambda) x_{2})\leq max\{(\Delta_{-K}\circ f)(x_{1}), (\Delta_{-K}\circ f)(x_{2})\}. $

假定$ (\Delta_{-K}\circ f)(x_{1})>(\Delta_{-K}\circ f)(x_{2}) $, 则$ (\Delta_{-K}\circ f)(\lambda x_{1}+(1-\lambda) x_{2})\leq(\Delta_{-K}\circ f)(x_{1}), $则有

$ (\Delta_{-K}\circ f)(x_{1})-(\Delta_{-K}\circ f)(\lambda x_{1}+(1-\lambda) x_{2})\geq0, $

从而$ \Delta_{-K}(f(x_{1})-f(\lambda x_{1}+(1-\lambda) x_{2})\geq0. $因此

$ \begin{align} f(x_{1})-f(\lambda x_{1}+(1-\lambda) x_{2})\in K. \end{align} $ (3.2)

(3.2)式和(3.1)式矛盾, 因此结论成立.

定理3.2  设$ f:X\rightarrow Y $为向量值映射, $ \; g:Y\rightarrow R $为单调递增函数, 且对$ y, z\in Y $满足

$ y\nleqq_{K}z\Rightarrow g(y)>g(z), $

$ f(x) $$ X $上是恰当$ K $-拟凸的当且仅当$ (g\circ f)(x) $是拟凸的.

  先证若$ f(x) $$ X $上是恰当$ K $-拟凸的, 则$ (g\circ f)(x) $是拟凸的. 由$ f(x) $是恰当$ K $-拟凸定义知, 对任意$ \lambda\in[0, 1], x_{1}, x_{2}\in X $$ f(\lambda x_{1}+(1-\lambda)x_{2})\leqq_{K}f(x_{1}) \text{或者} f(\lambda x_{1}+(1-\lambda)x_{2})\leqq_{K}f(x_{2}). $$ g $函数为单调递增的有

$ (g\circ f)(\lambda x_{1}+(1-\lambda)x_{2})\leq (g\circ f)(x_{1}), $

或者

$ (g\circ f)(\lambda x_{1}+(1-\lambda)x_{2})\leq (g\circ f)(x_{2}). $

所以$ (g\circ f)(\lambda x_{1}+(1-\lambda)x_{2})\leq max\{(g\circ f)(x_{1}), (g\circ f)(x_{2})\} $. 因此$ (g\circ f)(x) $是拟凸的.

下证若$ (g\circ f)(x) $是拟凸的, 则$ f(x) $$ X $上是恰当$ K $-拟凸的.

对任意的$ y\in Y, \; x\in X $, 令集合$ L_{y}=\{x|f(x)\ngeqq_{K}y\} $, 根据引理2.1知, 要证$ f(x) $是恰当$ K $-拟凸的, 只需证$ L_{y} $是凸集. 假设$ L_{y} $不是凸集, 则存在$ y\in Y $, 存在$ x_{1}, x_{2}\in L_{y}, \lambda\in[0, 1] $, 使得$ \lambda x_{1}+(1-\lambda)x_{2}\notin L_{y}. $从而$ \; f(\lambda x_{1}+(1-\lambda)x_{2})\geqq_{K} y. $根据$ g $的单调性, 有

$ \begin{align} (g\circ f)(\lambda x_{1}+(1-\lambda)x_{2})\geq g(y). \end{align} $ (3.3)

$ (g\circ f)(x) $是拟凸可知下水平集$ \{x\in X|(g\circ f)(x)<g(y)\} $是凸集. 又由$ f(x_{1})\ngeqq_{K} y, f(x_{2})\ngeqq_{K} y, $可以得出$ (g\circ f)(x_{1})<g(y), (g\circ f)(x_{2})<g(y). $从而

$ \begin{align} (g\circ f)(\lambda x_{1}+(1-\lambda)x_{2})<g(y). \end{align} $ (3.4)

显然(3.4)式和(3.3)式矛盾, 因此$ f(x) $是恰当$ K $-拟凸的.

4 拟凸向量值映射的次微分

本节内容主要定义了恰当$ K $-拟凸映射的次微分, 并简单讨论了该次微分的一些性质.

1984年, 陈光亚在文献[15] 中给出了向量值函数有效次微分的定义.

定义4.1[15]  给定集合$ A\subset R^{p} $, $ \bar{a}\in A $, 如果不存在$ a'\in A $, 使得$ a'<\bar{a} $, 则$ \bar{a} $称为$ A $的弱有效点. $ A $的全体弱有效点组成的集合记为$ \mathrm{eff}A $.

定义4.2[15]  称$ F:U\rightarrow R^{p} $$ u\in U $点是有效次可微的, 其中$ U\subset R^{n} $, 如果存在一个向量$ u^{*}\in R^{n} $使得$ F(u)-\langle u, u^{*}\rangle\in \mathrm{eff}\{F(v)-\langle v, u^{*}\rangle:v\in U\} $, 则向量$ u^{*} $称为$ F $$ u $点的有效次梯度, $ F $$ u $点的有效次梯度的全体, 记为$ \partial F(u) $, 称为$ F $$ u $点的有效次微分.

受该定义启发, 我们给出如下恰当$ K $-拟凸映射的次微分.

定义4.3  函数$ f:X\rightarrow Y $$ \bar{x}\in X $处的次微分定义为

$ v\in\partial^{<}_{v}f(\bar{x})\Longleftrightarrow v^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \forall x\in[f(x)\ngeq_{K} f(\bar{x})], \\ $
$ v\in\partial^{\leq}_{v}f(\bar{x})\Longleftrightarrow v^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \forall x\in[f(x)\ngtr_{K} f(\bar{x})], \\ $
$ v\in\partial^{*}_{v}f(\bar{x})\Longleftrightarrow v^{T}(x-\bar{x})\ngtr_{K} 0, \forall x\in[f(x)\ngeq_{K} f(\bar{x})], \\ $
$ v\in\partial^{\circledast}_{v}f(\bar{x})\Longleftrightarrow v^{T}(x-\bar{x})\ngeq_{K} 0, \forall x\in[f(x)\ngeq_{K} f(\bar{x})].\\ $

其中下水平集

$ [f(x)\ngeq_{K} f(\overline{x})]=\{x\in X:f(x)\ngeq_{K} f(\overline{x})\}, \\ $
$ [f(x)\ngtr_{K} f(\overline{x})]=\{x\in X:f(x)\ngtr_{K} f(\overline{x})\}. $

注4.1  (ⅰ) 陈光亚老师定义的次微分是包含在上述次微分里面的.

(ⅱ) 当$ f $为拟凸数值函数时, 定义4.3退化为定义2.2. 而且根据次微分的定义显然有$ \partial^{\leq}_{v}f(\overline{x})\subset\partial^{<}_{v}f(\overline{x})\subset\partial^{*}_{v}f(\overline{x})\subset\partial^{\circledast}_{v}f(\overline{x}) $, 这也是结论$ \partial^{\leq}\varphi(\overline{x})\subset\partial^{<}\varphi(\overline{x})\subset\partial^{*}\varphi(\overline{x})\subset\partial^{\circledast}\varphi(\overline{x}) $的推广.

定理4.1  设$ X=R^{n}, \; Y=R^{p}, \; K=R^{p}_{+} $. $ \bar{x}\in R^{n}, \; f(x)=(f_{1}(x), \ldots, f_{p}(x))^{T}:R^{n}\rightarrow R^{p}, $其中$ f_{i}, i=1, \ldots, p $为拟凸函数. 则$ \partial^{\leq}f(\bar{x})\subset \partial^{\leq}_{v}f(\bar{x}) $, 其中$ \partial^{\leq}f(\bar{x})=\{(v_{1}, \ldots, v_{p})\in R^{p\times n}|v_{i}\in \partial^{\leq}f_{i}(\overline{x}), i=1, \ldots, p\} $.

  任取$ V=(v_{1}, \ldots, v_{p})\in\partial^{\leq}f(\bar{x}) $. 由拟凸函数次微分定义有

$ \begin{align} \; f_{i}(x)-f_{i}(\bar{x})\geq v_{i}(x-\bar{x}), \; \; \forall x\in\{x\in R^{n}|f_{i}(x)\leq f_{i}(\bar{x})\}. \end{align} $ (4.1)

任取$ x\in\{x\in R^{n}|f(x)\ngtr f(\overline{x})\} $, 下面证明$ V^{T}(x-\bar{x})\ngtr f(x)-f(\bar{x}) $.

由于$ x\in\{x\in R^{n}|f(x)\ngtr f(\overline{x})\} $, 则存在$ i\in\{1, \ldots, p\} $使得$ f_{i}(x)\leq f_{i}(\bar{x}) $. 从而由(4.1)式可知

$ f_{i}(x)-f_{i}(\bar{x})\geq v_{i}(x-\bar{x}). $

这表明

$ V^{T}(x-\bar{x})\ngtr f(x)-f(\bar{x}). $

因此$ V\in \partial^{\leq}_{v}f(\overline{x}) $, 结论成立.

注4.2  定理4.1反包含关系不一定成立, 具体见例4.1.

例4.1  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+}, \bar{x}=0 $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $, 其中$ f_{1}(x)=(x-1)^3+1 $, $ f_{2}(x)=x $. 由定义可知

$ \begin{array}{lll} &&\partial^{\leq}_{v}f(\bar{x})=\{(x, y)\in R^{2}:x\in R, y\in(1, +\infty)\}, \\ \end{array} $
$ \begin{array}{lll} &&\partial^{\leq}f_{1}(\bar{x})=\emptyset, \; \partial^{\leq}f_{1}(\bar{x})=[1, +\infty).\\ \end{array} $

显然$ \partial^{\leq}_{v}f(\overline{x})\not\subset\partial^{\leq}f(\overline{x}) $.

下面我们主要讨论$ \partial^{\leq}_{v}f(\overline{x}) $的性质. 首先, 凸意义下的次微分是一个闭凸集, 但是我们这里给出的$ \partial^{\leq}_{v}f(\overline{x}) $并不一定是闭凸集, 下面例子可以说明这个问题.

例4.2  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+}, \bar{x}=0 $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $, 其中$ f_{1}(x)=x $, $ f_{2}(x)=x-1 $. 由定义可知$ \partial^{\leq}_{v}f(\bar{x})=A\cup B\cup C. $其中

$ \begin{array}{lll} &&A=\{(x, y)\in R^{2}:x\in R, y\in(1, +\infty)\}, \\ &&B=\{(x, y)\in R^{2}:x\in (1, +\infty), y\in R\}, \\ &&C=\{(x, y)\in R^{2}:x=1, y=1\}.\\ \end{array} $

图 4.1容易得出, $ f $$ \bar{x} $处的$ \partial^{\leq}_{v}f(\overline{x}) $次微分是非凸开集.

图 4.1  

定理4.2  设$ x, \bar{x}\in X $, $ f:X\rightarrow Y $是恰当$ K $-拟凸映射, 则对$ \lambda>0 $$ \partial^{\leq}_{v}(\lambda f)(\bar{x})=\lambda\partial^{\leq}_{v}f(\bar{x}). $

  先证$ \partial^{\leq}_{v}(\lambda f)(\bar{x})\subset\lambda\partial^{\leq}_{v}f(\bar{x}) $. 任取$ v\in\partial^{\leq}_{v}(\lambda f)(\bar{x}) $, 由定义可知

$ v^{T}(x-\bar{x})\ngtr_{K}\lambda f(x)-\lambda f(\bar{x}), \; \; \forall x\in\{x\in X|\lambda f(x)\ngtr_{K}\lambda f(\bar{x})\}, $

从而有

$ (v/\lambda)^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \; \; \forall x\in\{x\in X|\lambda f(x)\ngtr_{K}\lambda f(\bar{x})\}, $

$ (v/\lambda)^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \; \; \forall x\in\{x\in X| f(x)\ngtr_{K} f(\bar{x})\}. $

这表明$ v/\lambda\in\partial^{\leq}_{v}f(\bar{x}) $, 故$ v\in\lambda\partial^{\leq}_{v}f(\bar{x}). $

下证$ \lambda\partial^{\leq}_{v}f(\bar{x})\subset\partial^{\leq}_{v}(\lambda f)(\bar{x}) $. 设$ \lambda>0 $, 任取$ v\in\lambda \partial^{\leq}_{v}f(x) $, 由定义可知

$ (v/\lambda)^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \; \; \forall x\in\{x\in X|f(x)\ngtr_{K} f(\bar{x})\}. $

从而有

$ (v/\lambda)^{T}(x-\bar{x})\ngtr_{K} f(x)-f(\bar{x}), \; \; \forall x\in\{x\in X|\lambda f(x)\ngtr_{K} \lambda f(\bar{x})\}. $

$ v^{T}(x-\bar{x})\ngtr_{K} \lambda f(x)-\lambda f(\bar{x}), \; \; \forall x\in\{x\in X|\lambda f(x)\ngtr_{K}\lambda f(\bar{x})\}. $

这表明$ v\in\partial^{\leq}_{v}(\lambda f)(\bar{x}) $, 从而结论成立.

注4.3  下面举例说明, 当$ \lambda<0 $时, 定理4.2中的结论不一定成立.

例4.3  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+} $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $. 当$ x<0 $时, $ f_{1}(x)=x $, $ f_{2}(x)=-1 $. 当$ x\geq0 $时, $ f_{1}(x)=x $, $ f_{2}(x)=-2 $. 取$ \bar{x}=0, \lambda=-1 $时, 根据定义可知

$ \begin{array}{lll} &&\partial^{\leq}_{v}(f(\bar{x}))=A\cup B\cup C, \; \partial^{\leq}_{v}(\lambda f(\bar{x}))=D\cup E\cup F, \; \lambda\partial^{\leq}_{v}(f(\bar{x}))=G\cup H.\\ \end{array} $

其中

$ \begin{array}{lll} &&A=\{(x, y)\in R^{2}:x\in (1, +\infty), y\in R\}, \\ &&B=\{(x, y)\in R^{2}:x\in R, y\in [0, +\infty)\}, \\ &&C=\{(x, y)\in R^{2}:x=1, y=0\}, \\ &&D=\{(x, y)\in R^{2}:x\in(-1, +\infty), y\in(-\infty, 0)\}, \\ &&E=\{(x, y)\in R^{2}:x\in(-\infty, -1), y\in(0, +\infty)\}, \\ &&F=\{(x, y)\in R^{2}:x=-1, y=0\}, \\ &&G=\{(x, y)\in R^{2}:x\in (-\infty , -1), y\in R\}, \\ &&H=\{(x, y)\in R^{2}:x\in R, y\in (-\infty, 0]\}.\\ \end{array} $

显然$ \partial^{\leq}_{v}(\lambda f)(\bar{x})\neq\lambda\partial^{\leq}_{v}f(\bar{x}) $.

关于拟凸数值函数次微分的运算法则, 显然有$ \partial^{\leq}f(\bar{x})+\partial^{\leq}g(\bar{x})\subset\partial^{\leq}(f+g)(\bar{x}) $, 但我们发现$ \partial^{\leq}(f(x)+g(x))\subset \partial^{\leq}(f(x))+\partial^{\leq}(g(x)) $是不一定成立的, 见如下例子.

例4.4  设$ X=R, \; Y=R, \; \bar{x}=0 $, $ f(x)=x^{3}+x^{2}+1 $, $ g(x)=-x^{3} $. 由定义可知

$ \begin{array}{lll} \partial^{\leq}(f(\bar{x}))=\emptyset, \; \partial^{\leq}(g(\bar{x}))=\emptyset, \partial^{\leq}(f(\bar{x})+g(\bar{x}))=[1, +\infty). \end{array} $

因此$ \partial^{\leq}(f(\bar{x})+g(\bar{x}))\not\subset \partial^{\leq}(f(\bar{x}))+\partial^{\leq}(g(\bar{x})) $.

下面我们举例说明对于拟凸向量值映射的次微分, $ \partial^{\leq}_{v}(f(x))+\partial^{\leq}_{v}(g(x))\subset \partial^{\leq}_{v}(f(x)+g(x)) $是不一定成立的, 见如下例子.

例4.5  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+} $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $. 当$ x<0 $时, $ f_{1}(x)=0 $, $ f_{2}(x)=0 $. 当$ x\geq0 $时, $ f_{1}(x)=(x-1)^3+1 $, $ f_{2}(x)=x $.

$ g(x)=(g_{1}(x), g_{2}(x))^{T} $. 当$ x<0 $时, $ g_{1}(x)=0 $, $ g_{2}(x)=0 $. 当$ x\geq0 $时, $ g_{1}(x)=-[(x-1)^3+1] $, $ g_{2}(x)=-x $.

$ h(x)=f(x)+g(x)=(h_{1}(x), h_{2}(x))^{T} $, 则$ h_{1}(x)=0 $, $ h_{2}(x)=0 $.

$ \bar{x}=0 $时, 根据定义可知$ \partial^{\leq}_{v}f(\bar{x})=A\cup B\cup C, \; \partial^{\leq}_{v}(g(\bar{x}))=D, \; \partial^{\leq}_{v}(h(x))=E\cup F\cup C. $其中

$ \begin{array}{lll} &&A=\{(x, y)\in R^{2}:x\in R, y\in(0, +\infty)\}, \\ &&B=\{(x, y)\in R^{2}:x\in (0, +\infty), y\in R\}, \\ &&C=\{(x, y)\in R^{2}:x=0, y=0\}, \\ &&D=\{(x, y)\in R^{2}:x\in(0, +\infty), y\in(-\infty, -1)\}, \\ &&E=\{(x, y)\in R^{2}:x\in(0, +\infty), y\in(-\infty, 0)\}, \\ &&F=\{(x, y)\in R^{2}:x\in(-\infty, 0), y\in(0, +\infty)\}.\\ \end{array} $

$ \partial^{\leq}_{v}(f(\bar{x}))+\partial^{\leq}_{v}(g(\bar{x}))\not\subset \partial^{\leq}_{v}(f(\bar{x})+g(\bar{x})) $.

5 拟凸向量优化的最优性条件

本节内容我们主要利用$ \partial^{\leq}_{v}f(\overline{x}) $次微分研究拟凸向量优化问题弱有效解的最优性条件.

考虑下面的向量值优化问题:

(VOP)

$ \begin{array}{lll} &&\min f(x) \quad s.t.\quad x\in C. \end{array} $

其中$ C\subset X $, $ f:X\rightarrow Y $是恰当$ K $-拟凸的.

定义5.1  对于(VOP)问题有以下解的概念.设$ \bar{x}\in C $

(ⅰ)若不存在$ x\in C $, 使得$ f(x)\leq_{K} f(\bar{x}) $, 则称$ \bar{x}\in C $是(VOP)问题的有效解.

(ⅱ)若不存在$ x\in C $, 使得$ f(x)<_{K}f(\bar{x}) $, 则称$ \bar{x}\in C $是(VOP)问题的弱有效解.

定义5.2  设$ \varepsilon = {({\varepsilon _1}, ..., {\varepsilon _m})^T} \geq 0, \; \overline{x}\in C, $

(ⅰ) 若不存在$ x\in C $使得$ {f_i}(x) <_{K} {f_i}({\overline{x}}) - \varepsilon_{i} , \forall i \in \{1, ..., m\}, $则称$ \overline{x} $是(VOP) 问题的$ \varepsilon $-弱有效解.

(ⅱ) 若不存在$ x\in C $使得

$ {f_i}(x) \le_{K} {f_i}({\overline{x}}) - \varepsilon_{i} , \forall i \in \{1, ..., m\}, $
$ {f_j}(x) <_{K} {f_j}({\overline{x}}) - \varepsilon_{j}, \exists j \in \{1, ..., m\}, $

则称$ \overline{x} $是(VOP)问题的$ \varepsilon $- 有效解.

首先, 我们给出无约束向量优化问题弱有效解的最优性条件.

定理5.1  当$ C=X $时, $ \bar{x}\in X $为(VOP)问题的弱有效解当且仅当$ 0\in \partial^{\leq}_{v}f(\bar{x}) $.

  充分性: 反证, 假设$ \bar{x}\in X $不是$ f(x) $的弱有效解, 则存在$ x\in X $, 使得$ f(x)<_{K}f(\bar{x}) $, 即$ f(x)-f(\bar{x})<_{K}0 $. 由$ 0\in \partial^{\leq}_{v}f(\bar{x}) $可知

$ \begin{array}{lll} &&f(x)-f(\bar{x})\nless_{K}0^{T}(x-\bar{x}).\\ \end{array} $

从而

$ \begin{array}{lll} &&f(x)-f(\bar{x})\nless_{K}0.\\ \end{array} $

这与$ f(x)-f(\bar{x})<_{K}0 $矛盾, 因此$ \bar{x} $为(VOP)问题的弱有效解.

必要性: 反证, 假设$ 0\notin \partial^{\leq}_{v}f(\bar{x}) $, 则存在$ x\in[f(x)\ngtr f(\bar{x})] $, 有

$ \begin{array}{lll} &&f(x)-f(\bar{x})<_{K}0^{T}(x-\bar{x})=0.\\ \end{array} $

这与弱有效解的定义矛盾, 因此$ 0\in \partial^{\leq}_{v}f(\bar{x}) $.

以下例子可以说明定理5.1的合理性.

例5.1  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+} $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $. 当$ x>0 $时, $ f_{1}(x)=x $, $ f_{2}(x)=x-1 $, 当$ x\leq0 $时, $ f_{1}(x)=-1 $, $ f_{2}(x)=-1 $.

$ \bar{x}=0 $时, 由定义可知$ \partial^{\leq}_{v}f(\bar{x})=A\cup B\cup C. $其中

$ \begin{array}{lll} &&A=\{(x, y)\in R^{2}:x\in R, y\in(0, +\infty)\}, \\ &&B=\{(x, y)\in R^{2}:x\in (0, +\infty), y\in R\}, \\ &&C=\{(x, y)\in R^{2}:x=0, y=0\}.\\ \end{array} $

$ 0\in \partial^{\leq}_{v}f(\bar{x}) $, 且容易验证$ \bar{x}=0 $为该问题的弱有效解.

下面我们考虑带有约束的拟凸向量优化问题的最优性条件.

定理5.2  设$ X=R^{n}, \; Y=R^{p}, \; K=R^{p}_{+} $, $ \bar{x}\in C $. 若存在$ \lambda\in intR_{+}^{p} $, 使得$ \partial^{\leq}_{v}f(\bar{x})\cap-\lambda N(C, \bar{x})\neq\varnothing $, 则$ \bar{x} $为(VOP)问题的弱有效解. 其中$ \lambda N(C, \bar{x})=\{\lambda q^{T}\in R^{p\times n}:q\in N(C, \bar{x})\} $.

证明  设存在$ \lambda\in intR_{+}^{p} $, 使得$ \partial^{\leq}_{v}f(\bar{x})\cap-\lambda N(C, \bar{x})\neq\varnothing $, 则存在$ v\in\partial^{\leq}_{v}f(\bar{x}) $$ q\in N(C, \bar{x}) $使得$ v=-\lambda q $. 由$ q\in N(C, \bar{x}) $可知, 对任意$ x\in C $$ \langle q, x-\bar{x}\rangle\leq0 $. 从而有

$ (-\lambda q)^{T}(x-\bar{x})=\begin{bmatrix}\lambda_{1}\langle q, x-\bar{x}\rangle\\ \lambda_{2}\langle q, x-\bar{x}\rangle \\ \vdots \\ \lambda_{p}\langle q, x-\bar{x}\rangle \end{bmatrix}\geqq0. $

这表明$ v^{T}(x-\bar{x})\geqq0 $.

假设$ \bar{x} $不是(VOP)问题的弱有效解, 则存在$ x\in C $, 有$ f(x)-f(\bar{x})<0 $. 由$ v\in\partial^{\leq}_{v}f(\bar{x}) $$ v^{T}(x-\bar{x})\ngtr f(x)-f(\bar{x}). $再由$ f(x)-f(\bar{x})<0 $可知, $ v^{T}(x-\bar{x})\rangle\ngeqq 0. $这与$ v^{T}(x-\bar{x})\geqq0 $矛盾, 因此$ \bar{x} $为(VOP)问题的弱有效解.

以下例子可以说明定理5.2的合理性.

例5.2  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+}, \; C=[0, +\infty]\subset X $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $. 其中$ f_{1}(x)=(x-1)^3+1 $, $ f_{2}(x)=x $.

特别取$ \bar{x}=0 $时容易得出

$ \begin{array}{lll} &&\partial^{\leq}_{v}f(\bar{x})=R^{2}, N(C, \bar{x})=(-\infty, 0].\\ \end{array} $

这表明存在$ \lambda=(1, 1)^{T}>0 $, 使得$ \partial^{\leq}_{v}f(\bar{x})\cap-\lambda N(C, \bar{x})\neq\varnothing $, 并且$ \bar{x} $为(VOP)问题的弱有效解.

注5.1  定理5.2的逆命题不一定成立, 见例5.3.

例5.3  设$ X=R, \; Y=R^{2}, \; K=R^{2}_{+} $, $ f(x)=(f_{1}(x), f_{2}(x))^{T} $. 当$ x<0 $时, $ f_{1}(x)=-x $, $ f_{2}(x)=1 $. 当$ x\geq0 $时, $ f_{1}(x)=-x $, $ f_{2}(x)=2 $.

$ \bar{x}=0 $, 则$ \bar{x}=0 $$ f(x) $的弱有效解. 由定义可知

$ \begin{array}{lll} &&\partial^{\leq}_{v}f(\bar{x})=A\cup B\cup C, N(C, \bar{x})=(-\infty, 0].\\ \end{array} $

其中

$ \begin{array}{lll} &&A=\{(x, y)\in R^{2}:x\in (-1, +\infty), y\in(-\infty, 0)\}, \\ &&B=\{(x, y)\in R^{2}:x\in (-\infty, -1), y\in (0, +\infty)\}, \\ &&C=\{(x, y)\in R^{2}:x=-1, y=0\}.\\ \end{array} $

但是, 取$ \lambda=(1, 1)^{T}>0 $时, $ \partial^{\leq}_{v}f(\bar{x})\cap-\lambda N(C, \bar{x})=\varnothing $.

推论5.1  设$ X=R^{n}, \; Y=R^{p}, \; K=R^{p}_{+} $, $ \bar{x}\in C $. 若存在$ \lambda\in intR_{+}^{p} $, 使得$ \partial^{<}_{v}f(\bar{x})\cap-\lambda N(C, \bar{x})\neq\varnothing $, 则$ \bar{x} $为(VOP)问题的弱有效解. 其中$ \lambda N(C, \bar{x})=\{\lambda q^{T}\in R^{p\times n}:q\in N(C, \bar{x})\} $.

注5.2  当法锥$ N(C, \overline{x}) $改为$ N_{\varepsilon}(C, \overline{x}) $时, 利用上述次微分可得出如下近似解的最优性条件.

定理5.3  设$ X=R^{n}, \; Y=R^{p}, \; K=R^{p}_{+} $, $ \bar{x}\in C $. 若有$ \partial^{\leq}_{v}f(\bar{x})\cap- N_{\varepsilon}(C, \bar{x})\neq\varnothing $, 则$ \bar{x} $为(VOP)问题的$ \varepsilon $-有效解.

  由$ \partial^{\leq}_{v}f(\bar{x})\cap-N_{\varepsilon}(C, \bar{x})\neq\varnothing $可知, 存在$ \xi $, 有$ \xi \in \partial^{\leq}_{v}f(\bar{x})\cap-N_{\varepsilon}(C, \bar{x}), $$ \xi \in \partial^{\leq}_{v}f(\bar{x}) $$ \xi \in- N_{\varepsilon}(C, \bar{x}), $$ -\xi \in N_{\varepsilon}(C, \bar{x}). $于是有$ (-\xi)^{T} (x-\overline{x}) \leq \varepsilon , $

$ \xi^{T} (x-\overline{x}) \geq -\varepsilon. $

假设$ \overline{x} $不是(VOP)问题的$ \varepsilon $-有效解, 则存在$ x \in C $$ f(x)-f(\overline{x})\leq -\varepsilon. $$ \xi \in \partial^{\leq}_{v}f(\overline{x}) $$ \xi^{T} (x-\overline{x}) \ngtr f(x)-f(\overline{x}). $再由$ f(x)-f(\overline{x})\leq -\varepsilon<0 $可知

$ \xi^{T} (x-\overline{x})\ngeqq -\varepsilon. $

这与$ \xi^{T} (x-\overline{x}) \geq -\varepsilon $矛盾, 因此$ \overline{x} $为(VOP)问题的$ \varepsilon $-有效解.

参考文献
[1] Jahn. J. Vector Optimization-theory, applications and extensions(2nd Edition)[M]. Heidelberg Berlin: Spring-Verlag, 2011.
[2] Craven. B. D, Jeyakumar. V. Equivalence of a ky fan type minimax theorem and a gordan type alternative theorem[J]. Operations Research Letters, 1986, 5(2): 99–102. DOI:10.1016/0167-6377(86)90111-2
[3] Luc. D. T. Theory of vertor optimization[M]. Lecture Notes in Economic and Mathematical Systems, Berlin: Spring-Verlag, 1989.
[4] Ferro. F. A minimax theorem for vector-valued function[J]. Journal of Optimization Theory and Application, 1989, 60(1): 19–31. DOI:10.1007/BF00938796
[5] Lin. Y. C, Ansari. Q. H, Lai. H. C. Minimax theorems for set-valued mappings under cone-convexities[J]. Abstract and Applied Analysis, 2012, 2012(1): 137–138.
[6] Hiriart-Urruty. J. B. Tangent cones, generalized gradients and mathematical programming in banach spaces[J]. Mathematics of Operations Research, 1979, 4(1): 79–97. DOI:10.1287/moor.4.1.79
[7] Zaffaroni. A. Degrees of efficiency and degrees of minimality[J]. SIAM Journal on Control and Optimization, 2003, 42(3): 1071–1086. DOI:10.1137/S0363012902411532
[8] 林锉云, 董家礼. 多目标最优化的方法和理论[M]. 吉林: 吉林教育出版社, 1992.
[9] Greenberg. H. P, Pierskalla. W. P. Quasi-conjugate functions and surrogate duality[J]. cahiers centre études recherche opér, 1973, 15: 437–448.
[10] Gutiérrez Díez. J. M. Infragradients and directions of decrease[J]. Revista de la Real Academia de Ciencias Exactas, Fisicas y Naturales de Madrid, 1984, 78: 523–532.
[11] Plastria. F. Lower subdifferentiable functions and their minimization by cutting planes[J]. Journal of Optimization Theory Applications, 1985, 46(1): 37–53. DOI:10.1007/BF00938758
[12] Penot. J. P, Zalinescu. C. Elements of quasiconvex subdifferential calculus[J]. Journal of Convex Analysis, 2000, 7(7): 243–269.
[13] Penot. J. P. What is quasiconvex analysis[J]. Optimization, 2000, 47(1/2): 35–110.
[14] 李飞. 向量优化理论中的非线性标量化函数相关研究及应用[D]. 内蒙古: 内蒙古大学, 2016.
[15] 陈光亚. 向量值函数有效次微分映象的闭性[J]. 科学通报, 1984, 29(12): 714.
[16] 陈瑞婷, 徐智会, 高英. 拟凸多目标优化问题近似解的最优性条件[J]. 运筹学学报, 2019, 23(1): 35–44.