数学物理学报, 2020, 40(1): 31-43 doi:

论文

Riemannian流形中DE算法算子最优特征量的量子渐进估计

王凯光,1, 高岳林2

Quantum Asymptotic Estimation of the Optimal Eigenvalues of DE Operators in Riemannian Manifolds

Wang Kaiguang,1, Gao Yuelin2

通讯作者: 高岳林

收稿日期: 2019-04-1  

基金资助: 国家自然科学基金.  61561001
北方民族大学重大科研专项资助项目.  ZDZX201901
北方民族大学研究生创新项目.  YCX19120
宁夏高等教育一流学科建设资助项目.  NXYLXK2017B09

Received: 2019-04-1  

Fund supported: 国家自然科学基金.  61561001
北方民族大学重大科研专项资助项目.  ZDZX201901
北方民族大学研究生创新项目.  YCX19120
宁夏高等教育一流学科建设资助项目.  NXYLXK2017B09

作者简介 About authors

王凯光,E-mail:wkg13759842420@foxmail.com , E-mail:wkg13759842420@foxmail.com

摘要

该文主要分析和探讨了差分进化算法(Differential Eveolutionary Algorithm,DE)在Riemannian流形中的几何关系,对P-ε条件下Riemannian流形中的种群个体进行了收敛性分析,得到了迭代个体收敛精度与收敛速度的量子不确定渐进估计,如下式

其中,Δv2为种群个体的速度分辨率,Δxβε2为种群个体带有误差的位置分辨率,(λεii=1,2,…,n.从本质上说明了Riemannian流形中迭代个体的局部特征量是不能从收敛精度和收敛速度同时达到算法高效.

关键词: DE算法 ; Riemannian流形 ; 收敛精度 ; 收敛速度 ; 量子不确定渐进估计

Abstract

In this paper, the geometric relations of differential evolution algorithm in Riemannian manifolds are analyzed and discussed. The convergence of populations in Riemannian manifolds with P-ε is analyzed. A quantum uncertain asymptotic estimation of the convergence accuracy and convergence speed of the iterative individual is obtained as follows

where, Δv2 is speed resolution of individual populations, Δxβε2 is position resolution with error ε of individual populations, (λε)i, i=1, 2, …, n. The theorem expression essentially shows that the local eigenvalues of iterated individuals in Riemann manifolds can not achieve high convergent accuracy and convergent speed at the same time.

Keywords: DE algorithm ; Riemannian manifolds ; Convergent accuracy ; Convergent speed ; Quantum uncertain asymptotic estimation

PDF (485KB) 元数据 多维度评价 相关文章 导出 EndNote| Ris| Bibtex  收藏本文

本文引用格式

王凯光, 高岳林. Riemannian流形中DE算法算子最优特征量的量子渐进估计. 数学物理学报[J], 2020, 40(1): 31-43 doi:

Wang Kaiguang, Gao Yuelin. Quantum Asymptotic Estimation of the Optimal Eigenvalues of DE Operators in Riemannian Manifolds. Acta Mathematica Scientia[J], 2020, 40(1): 31-43 doi:

1 引言

我们在论文中讨论一个关于元启发式进化算法的重要问题,以差分进化算法研究为基础进行这项研究,重点讨论进化算法在完备Riemannian流形中的关于最优特征量收敛精度与收敛速度之间的渐进量子估计.

差分进化算法(Differential Eveolution,简称DE[1-3]),由Storn和Price于1995年提出的为解决切比雪夫不等式(Chebyshev inequality)而采用浮点矢量编码的在连续空间进行搜索的全局优化算法[4-6],是通过差分的方式进行迭代搜索的全局性进化算法,具有简单、易实现、收敛性好、鲁棒性强等全局性优化的优点[7-11].

为保证分析的严谨性,本文所分析的随机进化种群为闭生态种群,但这样的生态种群具有不完全性态,因此将闭生态种群扩充为Riemannian流形,这里重点要讨论的是关于差分进化算法结构参数和适应度函数在Riemannian流形上的不同进化阶段特征量的量子特性渐进估计[12].在本文中需要解决2个问题

(1) $ P_{-\varepsilon} $条件下Riemannian流形中个体的收敛性分析;

(2) DE算法在Riemannian流形中不同进化阶段的特征量的量子特性渐进估计.

注1.1   Riemannian流形强调的是在闭生态种群中允许存在可数个非收敛点存在的完备赋范空间的基础上,进一步保持闭生态种群的共性变换性质和等距变换性质的空间,也即保持闭生态种群的保角性和保距性.种群个体的随机性要求算子在不同进化阶段的最优特征量分析必须建立在$ P_{-\varepsilon} $条件下Riemannian流形中个体收敛的基础上,这是一个必要条件,在文献[6]中作者已经系统推导了关于$ DE $算法的收敛数学原理,这个结论就是本论文的研究基础;

注1.2  由于进化算子在推进寻优过程时一方面改变了某一空间的阶段性特征(收敛速度和收敛精度),另一方面改变了不同阶段的最优特征量(局部最优特征量),那么就提出了第一个猜想:不同阶段的最优特征量在某一阶段性特征的变化条件下是否具有极限意义上的绝对不变性,即个体收敛精度与收敛速度是否满足量子不确定原理,这将是我们重点探讨的结论.

2 预备知识

一般地, $ DE $算法极小化优化问题如下表示

$ \begin{eqnarray} &&\min \; f(X_{i}^{t}+P_{-\varepsilon}), X_{i}^{t} = \{x_{ij}^{t}|i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D\}, \end{eqnarray} $

$ \begin{eqnarray} &&s.t. \; a_{ij}\leq x_{ij}^{t}\leq b_{ij}, i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D, \end{eqnarray} $

其中, $ D $为决策变量的维数, $ NP $为种群规模, $ f(X_{i}+P_{\varepsilon}) $为适应度函数, $ P_{-\varepsilon} $为种群中的带有相对误差$ \varepsilon $的个体摄动变量.一般来说,摄动变量$ P_{-\varepsilon} $是一个无穷小量.为方便起见,假设每一个体在外界环境扰动下的摄动变量$ P_{-\varepsilon} $是相同的,它表示最优值在受条件影响时的可调节幅度,摄动变量$ P_{-\varepsilon} $越大,表示算法在逼近最优值时,个体离散程度较高.摄动变量$ P_{-\varepsilon} $越小,表示算法在逼近最优值时,个体离散程度较小.在本文研究中约定无穷小量的取值是一个固定值, $ X_{i}(i = 1, 2, \cdots, NP) $$ D $维参量矢量, $ x_{ij}(i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D) $为第$ i $个个体的第$ j $个分量, $ a_{ij}, b_{ij} $分别为寻优范围的下界和上界,差分进化算法(Differential Eveolution, DE)基本操作原理详见文献[1, 4, 7].

2.1 初始化种群

设置$ DE $算法的种群为$ X(t) $,则个体可表示为

$ \begin{eqnarray} &&X_{i}^{t} = (x_{i1}^{t}, x_{i2}^{t}, \cdots, x_{iD}^{t}), i = 1, 2, \cdots, NP, \end{eqnarray} $

其中, $ t $为进化代数, $ NP $为种群规模.

初始化种群:确定所求优化问题的维数$ D $,最大进代数$ T $,种群规模$ NP $,设置生成的初值寻优向量如下所示

$ \begin{eqnarray} &&X_{i}^{0} = (x_{i1}^{0}, x_{i2}^{0}, \cdots, x_{iD}^{0}), \end{eqnarray} $

$ \begin{eqnarray} &&x_{ij}^{0} = a_{ij}+rand(0, 1)\cdot (b_{ij}-a_{ij}), i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D, \end{eqnarray} $

其中, $ a_{ij}, b_{ij}\in {{\Bbb R}}^{+} $.

2.2 变异操作

$ DE $算法的个体变异成分是父代个体的差分矢量,每次差分变异个体均来源于第$ t $代父代个体种群中的两个个体分量$ (x_{i_1}^{t}, x_{i_2}^{t}) $,其中$ i_1, i_2\in NP $,则差分矢量定义为$ D_{i_{1, 2}} = (x_{i_{1}}^{t}-x_{i_{2}}^{t}) $,那么对任意矢量个体$ X_{i}^{t} $,定义变异操作为

$ \begin{eqnarray} V_{i}^{t+1} = X_{i_{3}}^{t}+F\cdot (X_{i_{1}}^{t}-X_{i_{2}}^{t}), \end{eqnarray} $

其中, $ {i_{1}}, {i_{2}}, {i_{3}}\in \{ 1, 2, \cdots, NP\} $$ {i_{1}}, {i_{2}}, {i_{3}} $互不相同,种群规模$ NP\geq 4 $, $ F $为收缩因子.在种群中随机选取不为零的相异矢量,通过差分操作得到变异个体,变异个体将实现调整种群多样性的可能.

2.3 交叉操作

对于种群目标矢量个体$ X_{i}^{t} $,与变异个体$ V_{i}^{t+1} $进行交叉操作,产生试验个体$ U_{i}^{t+1} $,为保持种群多样性,引入交叉概率$ CR $和随机函数$ rand(0, 1) $对变异个体$ V_{i}^{t+1} $和目标矢量个体$ X_{i}^{t} $进行交叉选择,保证试验个体$ U_{i}^{t+1} $至少有一位由变异个体$ V_{i}^{t+1} $贡献,对于其他位点,通过交叉概率$ CR $决定变异个体$ V_{i}^{t+1} $和目标矢量个体分量$ (x_{i}^{t}) $对试验个体$ U_{i}^{t+1} $中某些位点的贡献,交叉操作方程如下

$ \begin{equation} (u_{ij}^{t+1}) = \left\{ \begin{array}{ll} (v_{ij}^{t+1}), &{\rm if} \; rand_{j}[0, 1]\leq CR \; or \; j = j_{rand}, \\ (x_{ij}^{t}), &{\rm otherwise, } \end{array} \right. i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D. \end{equation} $

(2.7)式中$ rand_{j}[0, 1], CR\in (0, 1) $为交叉概率, $ CR $取值越大,表明矢量个体在不同位点发生交叉而产生新矢量个体的概率就越大, $ CR = 0 $时, $ U_{i}^{t+1} = X_{i}^{t} $,表明没有发生任何交叉操作,有利于保持种群的多样性和全局群搜索能力, $ CR = 1 $时, $ U_{i}^{t+1} = V_{i}^{t+1} $,表明一定在某些位点发生交叉操作,有利于保持全局搜索和加快收敛速度, $ CR = 0 $或1是发生交叉操作的两种极端情况, $ j = j_{rand} $为随机选择的位点,这是为了试验个体$ U_{i}^{t} $要从变异个体$ V_{i}^{t} $至少获得一个发生基因变异的位点,确保变异个体$ V_{i}^{t+1} $、目标矢量个体$ X_{i}^{t} $、试验个体$ U_{i}^{t+1} $三者矢量个体之间互不相同,表明交叉操作引起种群间的交叉是有效交叉.

2.4 选择操作

$ DE $算法的选择操作是一种基于贪婪算法的选择机制,经过变异和选择操作之后生成的试验个体$ U_{i}^{t+1} $与目标矢量个体$ X_{i}^{t} $进行竞争和选择,若当$ U_{i}^{t+1} $的适应度值好于$ X_{i}^{t} $的适应度值,那么$ U_{i}^{t+1} $作为最优个体会被种群继承到下一代,否则保留到下一代的个体将是$ X_{i}^{t} $,选择算子的作用由下述方程进行描述

$ \begin{eqnarray} X_{i}^{t+1} = \left\{ \begin{array}{ll}U_{i}^{t+1}, &{\rm if} \; f(U_{i}^{t+1})\leq f(X_{i}^{t}), \\ X_{i}^{t} , &{\rm otherwise, } \end{array} \right. i = 1, 2, \cdots, NP . \end{eqnarray} $

2.5 Riemannian流形

定义2.1[13]  设$ {\cal A} = \{(U_{1}, f_{U_{1}}), (U_{2}, f_{U_{2}}), \cdots\} $是定义在$ n $维流形$ M $上一个笛卡尔坐标集,称$ {\cal A} $$ M $的一个$ C^{r} $微分结构,若满足

(ⅰ) $ \{U_{1}, U_{2}, \cdots\} $$ M $上的一个开覆盖;

(ⅱ) $ {\cal A} $任一坐标卡$ (U_{1}, f_{U_{1}}) $, $ (U_{2}, f_{U_{2}}) $两两$ C^{r_{-}} $相容;

(ⅲ) $ {\cal A} $是最大坐标卡集,即$ M $上任一坐标卡$ (\hat{U}, f_{\hat{U}}) $如果与属于$ {\cal A} $的全部坐标卡$ C^{r_{-}} $相容,那么$ (\hat{U}, f_{\hat{U}})\in {\cal A} $,则$ {\cal A} $$ M $的一个$ C^{r} $微分结构.其中, $ f $是两个集合之间的双射.$ C^{r_{-}} $表示微分到$ r $次都连续.

定义2.2[13]   $ M $上有一个$ C^{r} $微分结构,那么$ M $是一个$ C^{r_{-}} $的微分流形.

定义2.3   (Riemannian流形)[13]  在Riemannian几何中, $ n $维流形上邻近两点之间的距离平方为

其中, $ g_{ij}(x) $是Riemannian度量,且处处可微,并满足$ \det|g_{ij}|\neq 0 $.我们称具有Riemannian度量的流形称为Riemannian流形.

性质2.1   Riemannian流形具有保距性、对称性,同时具有共形变换、等距变换的优良性质,且由共性变换、等距变换产生的子群同样具有保距性和对称性[13].

3 $ P_{-\varepsilon} $条件下高阶微分方程特征量的摄动

下面建立带有$ P_{-\varepsilon} $条件高阶微分方程,从生物角度来看是具有个体上的离散特征,从物理角度来看却具有时间上的连续特征,这为寻找最优存在个体提供多样性的搜索空间.由于差分进化算法具有搜索的连续性特征,因此任意种群中具有适应性最优个体必定是种群中所有个体的极限逼近,则存在种群摄动$ P_{-\varepsilon} $在理论上是合理的,用极限语言来描述就如下式所示

$ \begin{eqnarray} &&\min \; f(X_{i}^{t}+P_{-\varepsilon}), X_{i}^{t} = \{x_{ij}^{t}|i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D\}, \end{eqnarray} $

$ \begin{eqnarray} &&s.t. \; a_{ij}\leq x_{ij}^{t}\leq b_{ij}, i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D , \end{eqnarray} $

等价于

$ \begin{eqnarray} &&\lim\limits_{t\rightarrow +\infty} f((X_{i}^{t})_\varepsilon) = \lim\limits_{t\rightarrow +\infty} f_{\varepsilon}(X_{i}^{t}) = f_{\varepsilon}(X_{i})\in(f(X_{i})-\delta_\varepsilon, f(X_{i})+\delta_\varepsilon) , \end{eqnarray} $

$ \begin{eqnarray} &&s.t. \; a_{ij}\leq x_{ij}^{t}\leq b_{ij}, i = 1, 2, \cdots, NP;j = 1, 2, \cdots, D , \end{eqnarray} $

其中, $ f(X_{i}) $是差分进化算法在$ t\rightarrow+\infty $时的最优值,那么$ f_{\varepsilon}(X_{i}) $必然介于$ (f_{\varepsilon}(X_{i})-\delta_{\varepsilon}, $$ f_{\varepsilon}(X_{i}) +\delta_{\varepsilon}) $, $ \delta_{\varepsilon} $指的是最优值上下浮动的最大幅度.在同一种群,最优值只有一个,它继承了这个种群所有个体的所有适应性特征,与之相适应的的适应度函数$ f_{\varepsilon}(X_{i}^{t}) $度量了它在种群中的适应性,本文分别称之为种群的特征值与特征函数,之后根据需要可以统称为闭种群特征量.为了建立种群特征值与特征函数之间的连续性特征关系,需要通过$ P_{-\varepsilon} $条件下建立高阶常微分方程的特征值与特征函数的摄动方程,以此构建种群最优值与适应度函数之间的关系.

3.1 Riemannian流形中特征量在摄动变量$ P_\varepsilon $条件下的收敛性分析

种群特征和种群特征函数在数学空间描述时相当于是微分算子的特征值与特征函数,差分进化寻优在Riemannian流形中可归结为特征值与特征函数展开成为积分(或是级数)的问题,为此,本文考虑这样的初值问题$ A_{\varepsilon} $,首先设置初始向量为$ (X_{i}^{0})_{\varepsilon} = (x_{i1}^{0}, x_{i2}^{0}, \cdots , x_{iD}^{0})_{\varepsilon} $$ = (0, 0, \cdots, 0)^{\top} $, $ f(0) = 0 $,为了算出特征值和特征函数的具体表示形式及其收敛性,同时在微分算子中增加另一个初始条件$ (X_{i}^{0})_{\varepsilon} = (x_{i1}^{0}, x_{i2}^{0}, \cdots , x_{iD}^{0})_{\varepsilon} = (1+\varepsilon, 1+\varepsilon, \cdots, 1+\varepsilon)^{\top} $, $ f(1+\varepsilon) = 0 $这并不影响相关证明.

设一个$ n $阶线性常微分方程

$ \begin{eqnarray} &&L_{\varepsilon}f_{\varepsilon}-\lambda_{\varepsilon}f_{\varepsilon}\equiv f_{\varepsilon}^{(n)}+p_1(x)f_{\varepsilon}^{(n-1)}+\cdots + p_{n-1}(x)f_{\varepsilon}'+p_n(x)f_{\varepsilon}-\lambda_{\varepsilon}f_{\varepsilon} = 0 , \end{eqnarray} $

其中, $ \lambda_{\varepsilon} $指的是种群个体寻优的带有误差变量$ \varepsilon $的局部最优值(局部最优个体), $ f_{\varepsilon} $指的是带有误差变量$ \varepsilon $的具有微分性质的连续适应度函数,用来描述种群个体在寻优过程中逼近局部最优个体的接近程度,即$ f_{\varepsilon} $越小,种群寻优越接近局部最优个体.$ f_{\varepsilon}^{n} $指的是每迭代一次逼近局部最优个体的数值比率,用来描述种群个体在寻优过程中逼近局部最优个体的相似程度,即$ f_{\varepsilon}^{n} $越小,种群寻优与局部最优个体相似程度越高.

本文设置的关于摄动变量$ \varepsilon $的初始条件如下所示

$ \begin{eqnarray} \begin{array}{l} f_{\varepsilon}(0) = f_{\varepsilon}'(0) = \cdots = f_{\varepsilon}^{(n_1-1)}(0) = 0, \\ f_{\varepsilon}(1+\varepsilon) = f_{\varepsilon}'(1+\varepsilon) = \cdots = f_{\varepsilon}^{(n_2-1)} (1+\varepsilon) = 0, \\ (n_1+n_2 = n), \end{array} \end{eqnarray} $

其中, $ p_1(x), p_2(x), \cdots, p_{n-1}(x), p_n(x) $$ [0, 1+\delta] $上的解析函数.

首先,将$ f_0, f_1, \cdots, f_k $$ p_1, p_2, \cdots, p_n $$ x = 1 $处展开为幂级数

$ \begin{equation} f_{s}(1+\varepsilon) = \sum\limits_{m = 0}^{\infty}\frac{1}{m!}f_{s}^{(m)}(1)\varepsilon^{m}, s = 0, 1, 2, \cdots, k, \end{equation} $

$ \begin{equation} p_{i}(1+\varepsilon) = \sum\limits_{m = 0}^{\infty}\frac{1}{m!}p_{i}^{(m)}(1)\varepsilon^{m}, i = 1, 2, \cdots, n , \end{equation} $

$ \begin{equation} f_{s}^{(i)}(1+\varepsilon) = \sum\limits_{m = i}^{\infty}\frac{1}{(m-i)!}f_{s}^{(m)}(1)\varepsilon^{m-i} = \sum\limits_{l = 0}^{\infty}\frac{1}{l!}f_{s}^{l+1}(1)\varepsilon^{l}, s = 0, 1, 2, \cdots, k. \end{equation} $

$ \begin{eqnarray} p_{n-i}(1+\varepsilon)f_{k}^{i}(1+\varepsilon)& = &\bigg(\sum\limits_{m = 0}^{\infty}\frac{1}{m!}p_{n-i}^{(m)} (1)\varepsilon^{m}\bigg)\bigg(\sum\limits_{m = i}^{\infty}\frac{1}{(m-i)!}f_{s}^{(m)}(1)\varepsilon^{m-i} \bigg)\\ & = &\sum\limits_{l = 0}^{\infty}\bigg(\sum\limits_{j = 0}^{l}\frac{p_{n-i}^{j}}{l!}\cdot\frac{f_{k}^{i+l-j}}{(l-j)!} \bigg)\varepsilon^l\sum\limits_{i = 1}^{k}\lambda_{i}f_{k-i}(1+\varepsilon)\\ & = &\sum\limits_{i = 1}^{k}\lambda_{i}\bigg(\sum\limits_{l = 0}^{\infty}\frac{1}{l!} f_{k-l}^{l}(1)\varepsilon^l\bigg){}\\ & = &\sum\limits_{l = 0}^{\infty}\bigg(\sum\limits_{i = 1}^{k}\frac{\lambda_{i}}{l!} f_{k-l}^{l}(1)\bigg)\varepsilon^l, {\quad} i = 1, 2, \cdots, n. \end{eqnarray} $

把(3.7)–(3.10)式代入(3.5)式得

$ \begin{equation} \sum\limits_{l = 0}^{\infty}\bigg[\frac{1}{l!}f_{k}^{(n+l)}(1)+\sum\limits_{i}^{n-1}\sum\limits_{j = 0}^{l}\frac{1}{j!}p_{n-i}^{j}(1)\cdot \frac{1}{(l-j)!}f_{k}^{(i+l-j)}(1)\bigg]\varepsilon^l = \sum\limits_{l = 0}^{\infty}\bigg(\frac{1}{l!}\sum\limits_{i = 1}^{k}\lambda_{i} f_{k-i}^{(l)}(1)\bigg)\varepsilon^l . \end{equation} $

比较$ \varepsilon $同次幂系数得

$ \begin{eqnarray} &&\frac{1}{l!}f_{k}^{(n+l)}(1)+\sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{1}{j!}p_{n-i}^{j}(1)\cdot \frac{1}{(l-j)!}f_{k}^{(i+l-j)}(1) = \frac{1}{l!}\sum\limits_{i = 1}^{k}\lambda_{i} f_{k-i}^{(l)}(1) \end{eqnarray} $

$ \begin{eqnarray} &&\Rightarrow \frac{|f_{k}^{(n+l)}(1)|}{l!}\leq \sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{1}{j!}|p_{n-i}^{j}(1)|\cdot \frac{1}{(l-j)!}|f_{k}^{(i+l-j)}(1)|{}\\ &&{\qquad} {\qquad}{\qquad}{\quad}\ +\frac{1}{l!}\sum\limits_{i = 1}^{k}|\lambda_{i}||f_{k-i}^{(l)}(1)|, {\quad} (k\geq0, l\geq0)\\ &&\Rightarrow \frac{|f_{k}^{(n+l)}(1)|}{(n+l)!}\leq \sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{1}{j!}|p_{n-i}^{j}(1)|\cdot \frac{1}{(i+l-j)!}|f_{k}^{(i+l-j)}(1)|+\frac{1}{l!}\sum\limits_{i = 1}^{k}|\lambda_{i}||f_{k-i}^{(l)}(1)| . {\quad} \end{eqnarray} $

由此,有下面式子成立

$ \begin{eqnarray} &&\sum\limits_{l = 0}^{\infty}\frac{|f_{k}^{(n+l)}(1)|}{(n+l)!}\varepsilon^{n+l}{}\\ &< & \sum\limits_{l = 0}^{\infty}\sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{1}{j!}|p_{n-i}^{j}(1)|\cdot \frac{1}{(i+l-j)!}|f_{k}^{(i+l-j)}(1)|\varepsilon^{n+l}+ \sum\limits_{l = 0}^{\infty}\sum\limits_{i = 1}^{k}|\lambda_{i}|\frac{|f_{k-i}^{(l)}(1)|}{l!}\varepsilon^{n+l}{}\\ & = &\sum\limits_{i = 0}^{n-1}\varepsilon^{n-i}\bigg(\sum\limits_{j = 0}^{\infty}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j}\bigg) \bigg(\sum\limits_{m = 1}^{\infty}\frac{|f_{k}^{(m)}(1)|}{m!}\varepsilon^{m}\bigg)+ \varepsilon^{n}\sum\limits_{i = 1}^{k}|\lambda_{i}|\bigg(\sum\limits_{l = 0}^{\infty}\frac{|f_{k-i}^{(l)}(1)|}{l!}\varepsilon^{l}\bigg) . \end{eqnarray} $

进而,可以得到下面的式子

$ \begin{eqnarray} \sum\limits_{l = 1}^{\infty}\frac{|f_{k}^{(l)}(1)|}{l!}\varepsilon^{l}&<&\varepsilon^{l}\sum\limits_{l = 1}^{n-1}\frac{|f_{k}^{(l)}(1)|}{l!}+ \bigg[\sum\limits_{i = 0}^{n-1}\varepsilon^{n-i} \bigg(\sum\limits_{j = 0}^{\infty}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j}\bigg)\bigg]\cdot\sum\limits_{m = 1}^{\infty}\frac{|f_{k}^{(m)}(1)|}{m!}\varepsilon^{m}{}\\ &&+\varepsilon^{n}\sum\limits_{j = 0}^{\infty}\frac{|p_{n}^{(j)}(1)|}{j!}\varepsilon^{j}\cdot|f_{k}(1)| +\varepsilon^{n}\sum\limits_{i = 1}^{k}|\lambda_{i}|\bigg(\sum\limits_{l = 0}^{\infty}\frac{|f_{k-i}^{(l)}(1)|}{l!}\varepsilon^{l}\bigg). \end{eqnarray} $

(3.12)式中的第一个不等式两端同乘$ \frac{1}{(l+1)(l+2)\cdots(l+n-m)}, (1\leq m\leq n-1) $,得

$ \begin{eqnarray} \frac{|f_{k}^{(n+l)}(1)|}{(n+l-m)!}&\leq & \sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\cdot\frac{|f_{k}^{(i+l-j)}(1)|}{(l-j)!} +\sum\limits_{i = m}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\cdot\frac{|f_{k}^{(i+l-j)}(1)|}{(l-j+i-m)!}\\ &&\cdot \frac{(l-j+1)\cdots(l-j+i-m)}{(l+1)\cdots(l+n-m)}+\frac{1}{l!}\sum\limits_{i = 1}^{k}|\lambda_{i}||f_{k-i}^{(l)}(1)|{}\\ &\leq & \sum\limits_{i = 0}^{m-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\cdot \frac{|f_{k}^{(i+l-j)}(1)|}{(l-j)!}+\sum\limits_{i = m}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\cdot \frac{|f_{k}^{(i+l-j)}(1)|}{(l-j+i-m)!}\\ &&+\sum\limits_{i = 1}^{k}|\lambda_{i}||f_{k-i}^{(l)}(1)|. \end{eqnarray} $

(3.16)式两端同乘$ \varepsilon^{n+l-m} $,得

$ \begin{eqnarray} \frac{|f_{k}^{(n+l)}(1)|}{(n+l-m)!}\varepsilon^{n+l-m} &<&\bigg[\sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j}\cdot\frac{|f_{k}^{(i+l-j)}(1)|}{(l-j)!} \varepsilon^{l-j}\bigg]\varepsilon^{n-m}{}\\ &&+\sum\limits_{i = m}^{n-1}\varepsilon^{n-i}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j}\cdot \frac{|f_{k}^{(i+l-j)}(1)|}{(l-j+i-m)!}\varepsilon^{l-j+i-m}{}\\ &&+\varepsilon^{n-m}\bigg[ \sum\limits_{i = 1}^{k}|\lambda_{i}|\bigg(\frac{|f_{k-i}^{(n)}(1)|}{l!}\varepsilon^{l}\bigg)\bigg], \end{eqnarray} $

从而有

$ \begin{eqnarray} \sum\limits_{l = 0}^{\infty}\frac{|f_{k}^{(n+l)}(1)|}{(n+l-m)!}\varepsilon^{n+l-m} &<&\sum\limits_{l = 0}^{\infty}\sum\limits_{i = 0}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j} \cdot\frac{|f_{k}^{(i+l-j)}(1)|}{(l-j)!}\varepsilon^{l-j}\cdot\varepsilon^{n-m}\\ &&+\varepsilon^{n-i}\sum\limits_{l = 0}^{\infty}\sum\limits_{i = m}^{n-1}\sum\limits_{j = 0}^{l}\frac{|p_{n-i}^{(j)}(1)|}{j!}\varepsilon^{j} \cdot\frac{|f_{k}^{(i+l-j)}(1)|}{(l-j+i-m)!}\varepsilon^{l-j+i-m}{}\\ &&+\varepsilon^{n-m} \sum\limits_{i = 1}^{k}|\lambda_{i}|\bigg(\frac{|f_{k-i}^{(n)}(1)|}{l!}\varepsilon^{l}\bigg). \end{eqnarray} $

这里建立关于$ |\lambda_{k}|, $$ \|f_{k}\|_{p}, |f_{k, 0}(\varepsilon)| = \sum\limits_{l = 1}^{\infty} \frac{1}{l!}|f_{k}^{(l)}(1)|\varepsilon^{l}, $$ |f_{k, m}(\varepsilon) = \frac{1}{l!}|f_{k}^{(l+m)}(1)|\varepsilon^{l} $$ (m = $$ 1, 2, \cdots, n-1) $的优参数$ \bar{\lambda}_{k}, \bar{f}_{k}, \bar{f}_{k, 0}(\varepsilon) = \sum\limits_{l = 1}^{\infty}C_{k}^{l}\varepsilon^{l}, \bar{f}_{k, m}(\varepsilon) = \sum\limits_{l = 1}^{\infty}C_{k}^{m}\varepsilon^{l} $,由数学归纳法我们得到$ \bar{\lambda}_{k}>|\lambda_{k}|, $$ \bar{f}_{k}>\|f_{k}(k)\|_{p}, \bar{f}_{k, m}(\varepsilon)>\sum\limits_{j = 1}^{\infty} \frac{|f_{k}^{(m+j)}|}{j!}\varepsilon^{j} $$ (k\geq1, 0< m<-1) $.为了计算方便,给出一个形式化级数为

$ \begin{eqnarray} \lambda(\varepsilon) = \sum\limits_{k = 1}^{\infty}\bar{\lambda}_{k}\varepsilon^{k}, \bar{f}(\varepsilon) = \sum\limits_{k = 0}^{\infty}\bar{f}_{k}\varepsilon^{k}, u_{m}(\varepsilon) = \sum\limits_{k = 0}^{\infty}\bar{f}_{k, m} (\varepsilon)\varepsilon^{k}, (0\leq m \leq n-1). \end{eqnarray} $

(3.19)式满足下列方程组和初始条件

$ \begin{equation} \nu_{m}(0) = 0, \lambda_{\varepsilon}(0) = 0, \bar{f}_{\varepsilon}(0) = \bar{f}_{0}, (0\leq m \leq n-1), \end{equation} $

$ \begin{equation} T_{1}(\lambda_{\varepsilon}, \bar{f}_{\varepsilon}, \nu_{m})\equiv \bar{f}_{\varepsilon}-\bar{f}_{0}-M_{0}\lambda_{\varepsilon}\bar{f}_{\varepsilon}-M_{1}\sum\limits_{j = 0}^{n_2-1}\nu_{j} = 0 , \end{equation} $

$ \begin{equation} T_{2}(\lambda_{\varepsilon}, \bar{f}_{\varepsilon}, \nu_{m})\equiv \lambda_{\varepsilon}-N_{0}\sum\limits_{j = 0}^{n_2-1}\nu_{j} = 0 , \end{equation} $

$ \begin{eqnarray} \psi_{0}(\lambda_{\varepsilon}, \bar{f}_{\varepsilon}, \nu_{m})&\equiv& \rho_{0}(\varepsilon)(\bar{f}_{\varepsilon}-\bar{f}_{0})+\tau_{0}(\varepsilon) [\nu_{0}-\bar{f}_{0, 0}(\varepsilon)]+\varepsilon^{n}\bar{p}_{n}(\varepsilon)(\bar{f}_{\varepsilon}-\bar{f}_{0}){}\\ &&+\varepsilon^{n}\lambda_{\varepsilon}(\nu_{0}+\bar{f}_{\varepsilon})-\nu_{0}+\bar{f}_{0, 0}(\varepsilon) = 0 , \end{eqnarray} $

$ \begin{eqnarray} \psi{m}(\lambda_{\varepsilon}, \bar{f}_{\varepsilon}, \nu_{m}) &\equiv & \rho_{m}(\varepsilon)(\bar{f}_{\varepsilon}-\bar{f}_{0})+\varepsilon^{n-m}\sum\limits_{i = 0}^{m-1}\bar{p}_{n-i}(\varepsilon) [\nu_{i}-\bar{f}_{0, i}(\varepsilon)+(\bar{f}_{\varepsilon}-\bar{f}_{0})]{}\\ &&+\tau_{m}(\varepsilon)[\nu_{m}-\bar{f}_{0, m}(\varepsilon)] +\varepsilon^{n-m}\bar{p}_{n-m}(\varepsilon)(\bar{f}_{\varepsilon}-\bar{f}_{0}){}\\ &&+\varepsilon^{n-m}\lambda (\nu_{0}+\bar{f}_{\varepsilon})-\nu_{m}+\bar{f}_{0, m}(\varepsilon) = 0 . \end{eqnarray} $

由此我们知道$ \rho_{m}(0) = \tau_{m}(0) = 0, (0\leq m \leq n-1) $,因此,当$ \varepsilon = 0, \bar{f}_{\varepsilon} = \bar {f}_{0} $, $ \lambda_{\varepsilon} = \nu_{0} = u_{1} = \cdots = u_{n-1} = 0 $时有

$ \begin{equation} \frac{D(T_{1}, T_{2}, \psi_{0}, \cdots, \psi_{n-1})}{D(\bar{f}_{\varepsilon}, \lambda_{\varepsilon}, \nu_{0}, \cdots, \nu_{n-1})} = \left|\begin{array}{ccccccc} 1 &{\quad} -M\bar{T}_0-M{\quad} & \cdots &{\quad} \cdots {\quad}& \cdots & M_1\varepsilon\\ & 1 & -N_0\varepsilon & \cdots & \cdots & N_0\varepsilon\\ & & -1 &\varepsilon & \cdots & \varepsilon\\ & \bf{O} & & -1 & \cdots & \varepsilon\\ & & & & \cdots & \varepsilon\\ & & & & & -1 \end{array}\right| = (-1)^{n}\neq 0. \end{equation} $

因此, (3.21)–(3.24)式在初始条件(3.20)下有唯一的一组解析解$ \bar{f}(\varepsilon) $$ \lambda(\varepsilon) $$ \nu_{m}(\varepsilon) $$ (0\leq m \leq n-1) $以及(3.19)式在某个圆$ |\varepsilon|<\Gamma $内是收敛的.设$ 0<\frac{1}{r}<\Gamma $,那么有

其中$ a_{1}, a_{2}, \cdots, a_{n-1}\in \mathbb{R^+} $,我们由此得到$ |\lambda_{k}|<a_1\rho^k, \|f_{k}(x)\|_{p}<a_{2}\rho^k $,进而得到如下定理.

定理3.1  对(3.1)式所构建的种群个体,其特征值$ \lambda_{\varepsilon} $与特征函数$ f_{\varepsilon}(x) $是微分方程(3.5)给出的,并且当$ |\varepsilon|<\frac{1}{r}, r\in\mathbb{R^+} $时是收敛的,且为依摄动变量$ P_\varepsilon $随迭代数量递增而逐步收敛.

4 Riemannian流形中的Heisenberg测不准量子渐进估计

本节将重点探讨$ DE $在Riemannian流形中个体收敛精度与收敛速度的不确定量子性质,所有个体的量子特征都建立在上节的个体收敛性条件上, Heisenberg测不准原理是量子力学的一个基本原理[14],它从根本上说明了在一个量子系统中粒子的位置和它的动量不可能同时被测量,其基本形式为$ \Delta x\Delta p\geq \frac{h}{2} $, $ h $为约化$ Planck $常数.$ DE $在Riemannian流形中推动闭种群寻优时在空间上以基本算子变异、交叉、选择为拓扑映射对Riemannian流形中的个体进行测量,筛选出最优特征量$ x^{*} $,如果把整个Riemannian流形看做带有最佳信号源$ \beta $的完备空间,其中的每个个体都具有优质信号的特征,而$ DE $算法筛选的是所有带有优质信号源的个体中最优的,也就是最佳信号源,那么每一个体自身所携带信息量的多少与最佳信号源的频率和个体在寻优时间内依概率$ F $收敛保留最佳信号源信息量有关,随着寻优时间的延长,收敛迭代序列中的每一个体所携带的优质信息量不断累加,逐渐接近最佳信号源,这种具有极限特征的迭代过程可以逐步接近局部最优特征量.此时,会有两种情形出现,一种是当迭代收敛序列收敛速度变慢时,个体携带优质信息量累加速度变慢,但是与最佳信号源的位置精度$ \varepsilon $却在不断缩小;另一种情形时迭代收敛序列收敛速度变快时,个体携带优质信息量由于空间概率分布不均匀的原因会丢失一部分个体的优质信息量,导致与最佳信号源的位置精度$ \varepsilon $有增大趋势.下面给出Riemannian流形中$ DE $算法的Heisenberg测不准量子特性的具体表述.

定义4.1[15]  对于一个$ 2n\times 2n $的扩充辛群$ Sp(2n+P_{t}, R) $中的矩阵$ Q = \left(\begin{array}{cc} A{\quad} & B \\ C{\quad} & D \end{array}\right) $, $ f(q')\in L^{2}({{\Bbb R}} ^{n}) $的线性正则变换定义为

其中, $ C(M)(q, q') = \frac{{\rm e}^{(-\frac{{\rm i}n\pi}{4})}}{(\sqrt{2\pi})^{n}\sqrt{\det(B)}}\cdot {\rm e}^{{\rm i}(\frac{q^{\top}DB^{-1}q}{2}-q^{\top}(B^{\top})^{-1}q'+\frac{q'^{\top}B^{-1}Aq'}{2})} $.其逆变换为

注4.1  这里对原有文献[15]中的$ 2n\times 2n $的辛群$ Sp(2n+P_{t}, R) $做了依概率改进,成为$ 2n\times 2n $的扩充辛群$ Sp(2n+P_{t}, R) $,这样做的目的是为了缓解算子在不同搜索阶段的在局部最优特征量的误差,其中$ t $是操作算子的迭代次数.

定义4.2   (一维不确定原理)[16]  若$ f $是一个完备空间中的连续函数,那么它在完备空间中的速度分辨率$ \Delta_{v}^{2} $和位置分别率$ \Delta_{x}^{2} $分别定义为

其中, $ v_{0} = \int_{{{\Bbb R}} _{n}}v|f(v)|^{2}{\rm d}v, x_{0} = \int_{{{\Bbb R}} _{n}}x|\hat{f}(x)|^{2}{\rm d}x $.此时一维完备空间的Heisenberg测不准原理为$ \Delta_{v}^{2}\cdot \Delta_{x}^{2}\geq \frac{b^{2}}{4} $.

定义4.3  设$ f_{\varepsilon} $是定义在Riemannian流形中的连续可微函数,那么它在Riemannian流形中的速度分辨率$ \Delta_{v}^{2} $和位置分别率$ \Delta_{x_{\beta}^{\varepsilon}}^{2} $分别定义为

$ \begin{eqnarray} \Delta_{v}^{2} = \int_{{{\Bbb R}} _{n}}(v-v_{0})^{\top}(v-v_{0})|f_{\varepsilon}(v)|^{2}{\rm d}v, \Delta_{x_{\beta}^{\varepsilon}}^{2} = \int_{{{\Bbb R}} _{n}}(x-x_{0})^{\top}|\hat{f}_{\varepsilon}(x)|^{2}{\rm d}x , \end{eqnarray} $

其中, $ v^{\top} = (v_{1}, v_{2}, \cdots, v_{n})^{\top}, $$ v_{0}^{\top} = (\int_{{{\Bbb R}} _{n}}v_{1}|f(v)|^{2}{\rm d}v, \cdots, \int_{{{\Bbb R}} _{n}}v_{n}|f(v)|^{2}{\rm d}v)^{\top}, $$ x^{\top} = (x_{1}, x_{2}, \cdots, $$ x_{n})^{\top}, $$ x_{0}^{\top} = (\int_{{{\Bbb R}} _{n}}x_{1}|f(x)|^{2}{\rm d}x, \cdots, \int_{{{\Bbb R}} _{n}}x_{n}|f(x)|^{2}{\rm d}x)^{\top} $.

注4.2   $ f_{\varepsilon} $指的是带有误差变量$ \varepsilon $的具有微分性质的连续适应度函数,用来描述种群个体在寻优过程中逼近局部最优个体的接近程度,即$ f_{\varepsilon} $越小,种群寻优越接近局部最优个体.$ f $指的是关于种群中全部个体的适应度函数,用来描述种群个体在寻优过程中逼近全局最优个体的接近程度.

定理4.1  设$ f_{\varepsilon} $是定义在Riemannian流形空间中的连续可微函数且$ f_{\varepsilon}(v_{1}, \cdots, v_{n})\in L_{2}({{\Bbb R}} _{n}), $$ M\in Sp(2n+P_{t}, {{\Bbb R}} ) $. $ M^{n} $$ n $维Riemannian流形空间, $ \{S_{i}^{n}|i = 1, 2, \cdots, n\} $$ n $维Riemannian流形子空间且$ M^{n} = S_{1}^{n}\cup S_{2}^{n}\cup\cdots \cup S_{n}^{n} $形成了一个关于$ M^{n} $的开覆盖,那么$ { } \forall \lambda_{i}^{t}\in \{\lambda_{i}^{t}|i = 1, 2, \cdots, n\}, \exists S_{i}^{n}, \lim_{t\rightarrow \infty}\lambda_{i}^{t}\sim\lambda_{i} $$ { } \lim_{t\rightarrow \infty}\lambda_{i}^{t} = (\lambda_{\varepsilon})_{i}, \lambda_{i}\in S_{i}^{n} $,当$ \det(B)\neq 0 $时,则有下式成立

$ \begin{eqnarray} \Delta_{v}^{2}\cdot \Delta_{x_{\beta}^{\varepsilon}}^{2}\geq \bigg(\frac{\sqrt{\lim\limits_{t\rightarrow \infty}\lambda_{1}^{t}}+\cdots +\sqrt{\lim\limits_{t\rightarrow \infty}\lambda_{n}^{t}}}{2}\bigg)^{2} = \bigg(\frac{\sqrt{(\lambda_{\varepsilon})_{1}}+\cdots +\sqrt{(\lambda_{\varepsilon})_{n}}}{2}\bigg)^{2} , \end{eqnarray} $

其中, $ t $为迭代次数, $ v_{0}^{\top} = (\int_{{{\Bbb R}} _{n}}v_{1}|f_{\varepsilon}(v)|^{2}{\rm d}v, \cdots, \int_{{{\Bbb R}} _{n}}v_{n}|f_{\varepsilon}(v)|^{2}{\rm d}v)^{\top}, $$ x_{0}^{\top} = (\int_{{{\Bbb R}} _{n}}x_{1}|f_{\varepsilon}(x)|^{2}{\rm d}x, $$ \cdots, \int_{{{\Bbb R}} _{n}}x_{n}|f_{\varepsilon}(x)|^{2}{\rm d}x)^{\top} $, $ (\lambda_{\varepsilon})_{i} $$ B^{\top}B $的特征量值.

注4.3  定理$ 4.1 $$ { }\lim_{t\rightarrow \infty}\lambda_{i}^{t}\sim\lambda_{i} $指的是种群迭代个体在带有误差$ \varepsilon $的最优特征量$ \lambda_{i} $周围摆动并无限接近最后特征量$ \lambda_{i} $.

定理4.1的证明  首先利用开覆盖定理[17]证明在$ n $个开区域$ S_{i}^{n} $中至少存在一个最优特征量$ \lambda_{i} $.以其中一个开区域为例进行证明,设$ \delta_{1} $为开区域$ S_{i}^{n} $的一个最大半径,那么开区域$ B_{1}(S_{i}^{n}, \delta_{1})\subseteq B_{0}(S_{i}^{n}, \delta_{0}) $,且满足$ \delta_{0}\geq \delta_{1} $,此时记$ \delta_{1} = \frac{1}{2}\delta_{0} $;再对开区域$ B_{1} $进行分割,令$ \delta_{2}\geq \delta_{1} $,那么开区域$ B_{2} = (S_{i}^{n}, \delta_{2})\subseteq B_{1}(S_{i}^{n}, \delta_{1})\subseteq B_{0}(S_{i}^{n}, \delta_{0}) $,此时记$ \delta_{2} = \frac{1}{2^{2}}\delta_{0} $;同理,对开区域$ B_{2} $进行分割,令$ \delta_{3}\geq \delta_{2} $,那么开区域$ B_{3} = (S_{i}^{n}, \delta_{3})\subseteq B_{2}(S_{i}^{n}, \delta_{2})\subseteq B_{1}(S_{i}^{n}, \delta_{1})\subseteq B_{0}(S_{i}^{n}, \delta_{0}) $,此时记$ \delta_{3} = \frac{1}{2^{3}}\delta_{0} $,如此循环,则得到一个关于开区域$ S_{i}^{n} $的闭球套$ \{B_{k} = (S_{i}^{n}, \frac{1}{2^{k}}\delta_{0})\} $,它满足如下性质:

$ \begin{equation} \begin{array}{l} { } B_{k} = (S_{i}^{n}, \frac{1}{2^{k}}\delta_{0})\supseteq B_{k+1} = (S_{i}^{n}, \frac{1}{2^{k+1}}\delta_{0}), k = 1, 2, \cdots, n, \cdots, \\ [3mm] { } \lim\limits_{k\rightarrow \infty}\frac{1}{2^{k}} = 0, \end{array} \end{equation} $

那么,存在开区域$ (S_{i}^{n}, \frac{1}{2^{k}}\delta_{0}) $中的一个点$ (\lambda_{\varepsilon})_{i} $,使得$ (\lambda_{\varepsilon})_{i}\in B_{k} $且有$ { }\lim_{t\rightarrow \infty}\lambda_{i}^{t}\sim\lambda_{i}, \lambda_{i}\in S_{i}^{n} $$ { } \lim_{t\rightarrow \infty}\lambda_{i}^{t} = (\lambda_{\varepsilon})_{i}, \lambda_{i}\in S_{i}^{n} $成立.

这说明这样的点是存在的,下面来证明关于最优特征量的量子渐进不等式.在条件$ \det(B)\neq 0 $下,假设$ v_{0} = 0, x_{0} = 0 $.利用线性正则变换[18]我们得到

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2}& = &\int_{{{\Bbb R}} _{n}}x^{\top}x|\hat{f}_{\varepsilon}(u)|^{2}{\rm d}x{}\\ & = &\int_{{{\Bbb R}} _{n}}x^{\top}x\bigg|\int_{{{\Bbb R}} _{n}}f_{\varepsilon}(u)\frac{{\rm e}^{-\frac{{\rm i}n\pi}{4}}}{(\sqrt{2\pi})^{n}\sqrt{\det(B)}}{\rm e}^{-{\rm i}x^{\top}(B^{\top})^{-1}u+i\frac{u^{\top}B^{-1}Au}{2}}{\rm d}u\bigg|^{2}{\rm d}x. \end{eqnarray} $

$ t = B^{-1}x $,由积分变换得

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2} = \int_{{{\Bbb R}} _{n}}t^{\top}B^{\top}Bt\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}f_{\varepsilon}(u){\rm e}^{-{\rm i}t^{\top}u+{\rm i}\frac{u^{\top}B^{-1}Au}{2}}{\rm d}u\bigg|^{2}{\rm d}t . \end{eqnarray} $

我们记$ \widetilde{f_{\varepsilon}(u)} = f_{\varepsilon}(u){\rm e}^{{\rm i}\frac{u^{\top}B^{-1}Au}{2}} $,那么有

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2} = \int_{{{\Bbb R}} _{n}}t^{\top}B^{\top}Bt\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}\widetilde{f_{\varepsilon}(u)}{\rm e}^{-{\rm i}t^{\top}u}{\rm d}u\bigg|^{2}{\rm d}t \end{eqnarray} $

$ \begin{eqnarray} \Delta_{v}^{2} = \int_{{{\Bbb R}} _{n}}u^{\top}u|f_{\varepsilon}(u)|^{2}{\rm d}u = \int_{{{\Bbb R}} _{n}}u^{\top}u|\widetilde{f_{\varepsilon}(u)}|^{2}{\rm d}u . \end{eqnarray} $

因为$ \det(B)\neq 0 $$ B^{\top}B $是一个对称性正定矩阵,利用矩阵谱分解知存在正交矩阵$ P $,使得

$ \begin{eqnarray} B^{\top}B = P^{\top}\Lambda P , \end{eqnarray} $

其中$ \Lambda $是一个对角矩阵,且对角线上元素是$ B^{\top}B $的特征量值,从而有

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2} = \int_{{{\Bbb R}} _{n}}t^{\top}P^{\top}\Lambda Pt\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}\widetilde{f_{\varepsilon}(u)}{\rm e}^{-{\rm i}t^{\top}u}{\rm d}u\bigg|^{2}{\rm d}t . \end{eqnarray} $

$ \omega = Pt $,对(4.9)式做一次积分变换,同时令$ u = P^{\top}y $

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2}& = &\int_{{{\Bbb R}} _{n}}\omega^{\top}\Lambda \omega\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}\widetilde{f_{\varepsilon}(u)}{\rm e}^{-{\rm i}\omega^{\top}Pu}{\rm d}u\bigg|^{2}{\rm d}t{}\\ & = &\int_{{{\Bbb R}} _{n}}\omega^{\top}\Lambda \omega\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}\widetilde{f_{\varepsilon}(P^{\top}y)}{\rm e}^{-{\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}t , \end{eqnarray} $

同时有

$ \begin{eqnarray} \Delta_{v}^{2} = \int_{{{\Bbb R}} _{n}}u^{\top}u|\widetilde{f_{\varepsilon}(u)}|^{2}{\rm d}u = \int_{{{\Bbb R}} _{n}}y^{\top}y|\widetilde{f_{\varepsilon}(P^{\top}y)}|^{2}{\rm d}y \end{eqnarray} $

$ \widetilde{f_{\varepsilon}(P^{\top}y)} = h(y) $,则有

$ \begin{eqnarray} \Delta_{x_{\beta}^{\varepsilon}}^{2} = \int_{{{\Bbb R}} _{n}}\omega^{\top}\Lambda \omega\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}h(y){\rm e}^{-{\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}t, {\quad} \Delta_{v}^{2} = \int_{{{\Bbb R}} _{n}}y^{\top}y|h(y)|^{2}{\rm d}y. \end{eqnarray} $

从而有

$ \begin{eqnarray} \Delta_{v}^{2}\cdot\Delta_{x_{\beta}^{\varepsilon}}^{2}& = &\int_{{{\Bbb R}} _{n}}y^{\top}y|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}}\omega^{\top}\Lambda\omega\frac{1}{(2\pi)^{n}}\bigg|\int_{{{\Bbb R}} _{n}}h(y){\rm e}^{-{\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}t{}\\ & = &\int_{{{\Bbb R}} _{n}}(y_{1}^{2}+\cdots+y_{n}^{2})|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}} \bigg(\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}\omega_{1}^{2}+\cdots+\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}\omega_{n}^{2}\bigg) \frac{1}{(2\pi)^{n}}{}\\ &&\times\bigg |\int_{{{\Bbb R}} _{n}}h(y){\rm e}^{-{\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}\omega{}\\ & = &\int_{{{\Bbb R}} _{n}}(y_{1}^{2}+\cdots+y_{n}^{2})|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}}\bigg(\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}\bigg|\omega_{1}\int_{{{\Bbb R}} _{n}}h(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}+\cdots{}\\ &&+\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}\bigg|\omega_{n}\int_{{{\Bbb R}} _{n}}h(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}\bigg){\rm d}\omega , \end{eqnarray} $

其中$ \lambda_{i} $$ B^{\top}B $的第$ i $个特征量值,记$ h_{i} = \frac{\partial h(y)}{\partial y_{i}} $,由Fourier变换性质[18]

$ \begin{eqnarray} \Delta_{v}^{2}\cdot\Delta_{x_{\beta}^{\varepsilon}}^{2}& = &\int_{{{\Bbb R}} _{n}}(y_{1}^{2}+\cdots+y_{n}^{2})|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}}\bigg(\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}\bigg|\int_{{{\Bbb R}} _{n}}h_{1}(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}+\cdots{}\\ &&+\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}\bigg|\int_{{{\Bbb R}} _{n}}h_{n}(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}\bigg){\rm d}\omega. \end{eqnarray} $

由Cauchy不等式可知

$ \begin{eqnarray} \Delta_{v}^{2}\cdot\Delta_{x_{\beta}^{\varepsilon}}^{2}&\geq & \bigg(\bigg(\int_{{{\Bbb R}} _{n}}y_{1}^{2}|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}}\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}\bigg|\int_{{{\Bbb R}} _{n}}h_{1}(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}y\bigg)^{\frac{1}{2}}{}\\ &&+\cdots+\bigg(\int_{{{\Bbb R}} _{n}}y_{n}^{2}|h(y)|^{2}{\rm d}y\cdot\int_{{{\Bbb R}} _{n}}\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}\bigg|\int_{{{\Bbb R}} _{n}}h_{n}(y){\rm e}^{2\pi {\rm i}\omega^{\top}y}{\rm d}y\bigg|^{2}{\rm d}y\bigg)^{\frac{1}{2}}\bigg)^{2} .{\quad} \end{eqnarray} $

利用积分形式的Cauchy不等式可知

$ \begin{eqnarray} \Delta_{v}^{2}\cdot\Delta_{x_{\beta}^{\varepsilon}}^{2}&\geq & \bigg(\int_{{{\Bbb R}} _{n}}(|y_{1}h(y)\sqrt{\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}}h_{1}^{*}(y)|+ \cdots+|y_{n}h(y)\sqrt{\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}}h_{n}^{*}(y)|){\rm d}y\bigg)^{2} {}\\ & = &\bigg(\sqrt{\lim\limits_{t\rightarrow\infty}\lambda_{1}^{t}}\int_{{{\Bbb R}} _{n}}(|y_{1}h(y)h_{1}^{*}(y)|{\rm d}y+\cdots+\sqrt{\lim\limits_{t\rightarrow\infty}\lambda_{n}^{t}}\int_{{{\Bbb R}} _{n}}|y_{n}h(y)h_{n}^{*}(y)|{\rm d}y\bigg)^{2} . {\qquad} \end{eqnarray} $

由一维Heisenberg不确定性原理可知

$ \begin{eqnarray} \int_{{{\Bbb R}} _{n}}|y_{1}h(y)\sqrt{\lim\limits_{t\rightarrow \infty}\lambda_{1}^{t}}h_{1}^{*}(y)|{\rm d}y\geq\frac{1}{2} . \end{eqnarray} $

综上可知

$ \begin{eqnarray} \Delta_{v}^{2}\cdot \Delta_{x_{\beta}^{\varepsilon}}^{2}\geq \bigg(\frac{\sqrt{\lim\limits_{t\rightarrow \infty}\lambda_{1}^{t}}+\cdots +\sqrt{\lim\limits_{t\rightarrow \infty}\lambda_{n}^{t}}}{2}\bigg)^{2} = \bigg(\frac{\sqrt{(\lambda_{\varepsilon})_{1}}+\cdots +\sqrt{(\lambda_{\varepsilon})_{n}}}{2}\bigg)^{2} . \end{eqnarray} $

证毕.

注4.4  这个基于Riemannian流形考虑的关于种群个体收敛速度和收敛精度与局部最优特征量的渐进估计,从本质上说明了个体迭代过程中收敛精度与速度呈现类似于不确定量子原理.文献[19]则从数值计算方面进一步肯定了这样的量子关系,同时在种群搜索空间方面分析了收敛的三种拓扑结构,对于探究种群的收敛性分析起到了推动作用.

5 结论

本文主要探讨了关于差分进化算法结构参数和适应度函数在Riemannian流形上的不同进化阶段特征量的量子特性渐进估计,得到了基于Riemannian流形考虑的关于种群个体收敛速度和收敛精度与局部最优特征量的量子渐进估计,从而本质上阐述了DE算法在量子层面的几何关系.

参考文献

Storn R , Price K .

Differential evolution -a simple and efficient heuristic for global optimization over continuous spaces

Journal of Global Optimization, 1997, 11 (4): 341- 359

URL     [本文引用: 2]

Yuan X , Cao B , Yang B , et al.

A modified differential evolution for constrained optimization

Information Computing and Automation, 2008, 3, 19- 22

URL    

Qing A. Advances in Differential Evolution. Germany: Springer Berlin Heidelberg, 2008

[本文引用: 1]

Rönkkönen J , Kukkonen S , Price K V .

Real-parameter optimization with differential evolution

IEEE Congress on Evolutionary Computation, 2005, 1 (1): 506- 513

URL     [本文引用: 2]

Price K, Storn R M, Lampinen J A. Differential Evolution: A Practical Approach to Global Optimization. Germany: Springer Science and Business Media, 2006

王凯光, 高岳林.

十进制整数编码的DE算法模式集定理研究

应用数学, 2019, 32 (02): 443- 451

URL     [本文引用: 2]

Wang K G , Gao Y L .

The schema sets theorem of DE algorithm for decimal integer coding

Mathematica Applicata, 2019, 32 (02): 443- 451

URL     [本文引用: 2]

Qin A K , Huang V L , Suganthan P N .

Differential evolution algorithm with strategy adaptation for global numerical optimization

IEEE transactions on Evolutionary Computation, 2008, 13 (2): 398- 417

URL     [本文引用: 2]

Das S , Suganthan P N .

Differential evolution:a survey of the state-of-the-art

IEEE Transactions on Evolutionary Computation, 2011, 15 (1): 4- 31

Storn R .

System design by constraint adaptation and differential evolution

IEEE Transactions on Evolutionary Computation, 1999, 3 (1): 22- 34

URL    

Thosen R. Flexible Ligand Docking Using Differential Evolution[C]//Proceedings of the 2003 IEEE Congress on Evolutionary Computation. Canberra, Australia, 2003, 4: 2354-2361

Yang M , Li C H , Cai Z H , et al.

Differential evolution with auto-enhanced population diversity

IEEE Transactions on Cybernetics, 2015, 45 (2): 302- 315

URL     [本文引用: 1]

Lin T , Zha H .

Riemannian manifold learning

IEEE Transactions on Pattern Analysis and Machine Intelligence, 2008, 30 (5): 796- 809

URL     [本文引用: 1]

汪容. 数学物理中的微分几何与拓扑学. 杭州: 浙江大学出版社, 2010

[本文引用: 4]

Wang R . Differential Geometry and Topology in Mathematical Physic. Hangzhou: Zhejiang University Press, 2010

[本文引用: 4]

Busch P , Heinonen T , Lahti P .

Heisenberg's uncertainty principle

Physics Reports, 2007, 452 (6): 155- 176

URL     [本文引用: 1]

Moshinsky M , Quesne C .

Linear canonical transformations and their unitary representations

Journal of Mathematical Physics, 1971, 12 (8): 1772- 1780

URL     [本文引用: 2]

Zhao J , Tao R , Wang Y .

On signal moments and uncertainty relations associated with linear canonical transform

Signal Processing, 2010, 90 (9): 2686- 2689

URL     [本文引用: 1]

庞学诚, . 数学分析(第四版·上). 北京: 高等教育出版社, 2001

[本文引用: 1]

Pang X C , et al. Mathematical Analysis (Fourth Edition Part I). Beijing: Higher Education Press, 2001

[本文引用: 1]

Bracewell R N , Bracewell R N . The Fourier Transform and Its Applications. New York: McGraw-Hill, 1986

[本文引用: 2]

Wang K G , Gao Y L .

Topology structure implied in β-Hilbert space, Heisenberg uncertainty quantum characteristics and numerical simulation of the DE algorithm

Mathematics, 2019, 7 (4): 330- 349

URL     [本文引用: 1]

/