周 彪,李素蘭
(浙江工業(yè)大學(xué) 理學(xué)院,浙江 杭州 310023)
解變分不等式的簡(jiǎn)單牛頓法的收斂性
周彪,李素蘭
(浙江工業(yè)大學(xué) 理學(xué)院,浙江 杭州 310023)
摘要:牛頓法作為求解變分不等式問(wèn)題的一個(gè)重要方法,它的收斂性一直是各位學(xué)者研究的一個(gè)核心問(wèn)題.當(dāng)變分不等式中的函數(shù)F在內(nèi)滿足γ-條件時(shí),證明了由牛頓法產(chǎn)生的迭代點(diǎn)列是適定的,而且解的序列可以被構(gòu)造出來(lái)的{tn}序列控制收斂到一個(gè)變分不等式的最優(yōu)解.考慮到F在內(nèi)滿足γ-條件這個(gè)區(qū)域性條件給進(jìn)一步的研究帶來(lái)了困難,因此引入了解析函數(shù),給出了F是解析函數(shù)條件下的牛頓法的收斂性結(jié)果.數(shù)值實(shí)驗(yàn)表明算法是有效的.
關(guān)鍵詞:牛頓法;變分不等式;γ-條件;半局部收斂;解析函數(shù)
變分不等式(Variational inequality)首先由Hartman和Stampacchia[1]在1966年提出,最初是用來(lái)研究偏微分方程的,它在數(shù)學(xué)、工程、物理和經(jīng)濟(jì)等領(lǐng)域有著廣泛的應(yīng)用背景.作為非線性互補(bǔ)問(wèn)題的推廣,它的提出統(tǒng)一了優(yōu)化問(wèn)題和均衡問(wèn)題的研究,并且在數(shù)學(xué)領(lǐng)域中作為大量數(shù)學(xué)問(wèn)題實(shí)際求解的統(tǒng)一框架.實(shí)際生活中的許多問(wèn)題都可以歸結(jié)為(或松弛成)一個(gè)凸優(yōu)化問(wèn)題,而凸優(yōu)化的一階必要性條件就是一個(gè)單調(diào)變分不等式.因此用變分不等式研究凸優(yōu)化的求解問(wèn)題,就像微積分中用導(dǎo)數(shù)求一元函數(shù)的極值,常常會(huì)帶來(lái)很大的方便,這使得它成為數(shù)學(xué)規(guī)劃中一個(gè)十分熱門的研究課題[2-6].
1理論基礎(chǔ)與符號(hào)說(shuō)明
1.1符號(hào)說(shuō)明
1)Rn表示n維歐式向量空間,令Ω→Rn,F(xiàn):Ω→Rn表示映射F的值域和定義域都是Rn的非空子集.
2)‖x‖:向量x的2-范數(shù).
3)‖A‖:矩陣A的2-范數(shù).
B(x0,ρ)∶={y∈Rn|y-x0 1.2理論基礎(chǔ) 下面給出變分不等式的一般形式: 給定一個(gè)Rn中的非空子集Ω和一個(gè)函數(shù)F:Ω→Rn,變分不等式問(wèn)題記為VIP(Ω,F),是指求x*∈Ω,使得 (y-x*)TF(x*)≥0?y∈Ω (1) 文獻(xiàn)[13]中提到了很多數(shù)值解法,其中牛頓法是最重要的方法之一.牛頓法解函數(shù)F(x)=0時(shí)的一般迭代格式為 F′(xn-1)T(xn-xn-1)=-F(xn-1) n=1,2,3,…,∞ (2) 而在求解變分不等式時(shí),序列{xn}是由以下不等式的解產(chǎn)生的:假設(shè)已知迭代起點(diǎn)x0∈Ω,則x1的解為 (y-x1)T(F(x1)+F′(x0)(x1-x0))≥0 ?y∈Ω 那么假設(shè)xn-1存在,則xn為 (y-xn)T(F(xn-1)+F′(xn-1)(xn-xn-1))≥0 ?y∈Ω (3) 的解. 定義1設(shè)Ω是Rn的一個(gè)非空的凸閉集,令ρ>0,γ>0,ργ<1.函數(shù)F:Ω→Rn在Ω上二階可微且存在一個(gè)x0∈Ω,F(xiàn)′(x0)-1存在,若不等式 (4) 定義2數(shù)學(xué)上如果一個(gè)函數(shù)在鄰域內(nèi)的每個(gè)點(diǎn)都能夠展開成冪級(jí)數(shù),且收斂半徑大于零,則我們稱函數(shù)在鄰域內(nèi)是解析的. 2由牛頓法產(chǎn)生的序列{xn}的收斂性 2.1γ-條件 給定一個(gè)Rn中的非空子集Ω和一個(gè)函數(shù)F:Ω→Rn,變分不等式問(wèn)題記為VIP(Ω,F)是指求x*∈Ω,使得 (y-x*)TF(x*)≥0?y∈Ω (5) 牛頓法作為求解變分不等式(5)最重要的方法之一,在求解xn時(shí)文獻(xiàn)[13]考慮的是F在xn-1處的一階導(dǎo)數(shù)信息,將算法簡(jiǎn)化為每次迭代都只考慮F在x0處的一階導(dǎo)數(shù)信息.故其迭代格式如下: 假設(shè)已知迭代起點(diǎn)x0∈Ω,則x1的解為 (y-x1)T(F(x1)+F′(x0)(x1-x0))≥0 ?y∈Ω (6) 那么假設(shè)xn-1存在,則xn為 (y-xn)T(F(xn-1)+F′(x0)(xn-xn-1))≥0 ?y∈Ω (7) 的解.即xn為VIP(Ω,Fn-1)的一個(gè)解.記Fn-1(xn)∶=F(xn-1)+F′(x0)(xn-xn-1). 王興華和韓丹夫教授[15]通過(guò)引入函數(shù) (8) 改進(jìn)了Smale點(diǎn)估計(jì)的結(jié)論.其中β>0,R滿足 此時(shí)若將式(8)中的函數(shù)L定義為 則函數(shù)h(t)為 (9) 并且定義了一個(gè)序列{tn}為 t0=0,n=0,1,… (10) 其中h(t)滿足式(9). 引理1假設(shè)A∈Rn×Rn是一個(gè)正定矩陣,x∈Rn,則有 證明回顧矩陣2-范數(shù)的定義 ‖xn+1-xn‖ (11) 且{xn}收斂到變分不等式的一個(gè)解x*. 證明首先證明由牛頓法產(chǎn)生的序列{xn}是適定的:因?yàn)镕′(x0)是正定的,那么顯然Fn-1(x)是嚴(yán)格遞增的.而Ω是Rn的一個(gè)非空的凸閉集,所以VIP(Ω,Fn-1)有一個(gè)唯一解xn.即如果F′(x0)是正定的,則由式(7)牛頓法產(chǎn)生的序列是有意義的[12]. (xn+1-xn)TFn-1(xn)=(xn+1-xn)T· (F(xn-1)+F′(x0)(xn-xn-1))≥0 (12) (xn-xn+1)TFn(xn+1)=(xn-xn+1)T· (F(xn)+F′(x0)(xn+1-xn))≥0 (13) 合并式(12,13),并移項(xiàng)得 (xn-xn+1)TF′(x0)(xn-xn+1)≤ (xn-xn+1)T[F(xn)-F(xn-1)- F′(x0)(xn-xn-1)] (14) (xn-xn+1)TF′(x0)(xn-xn+1)≥ (15) 結(jié)合式(14,15),有 (xn-xn+1)≤ (xn-xn+1)T[F(xn)- F(xn-1)-F′(x0)· (xn-xn-1)] 移項(xiàng)化簡(jiǎn)得 ‖xn+1-xn‖≤‖F(xiàn)′(x0)-1‖‖F(xiàn)(xn)-F(xn-1)- F′(x0)(xn-xn-1)‖= t(xn-xn-1))-F′(x0)‖· ‖xn-xn-1‖dt (16) ‖xn-xn + 1‖‖xn - 1+ t(xn-xn - 1)-x0‖dλdt = tn+1-tn (y-x*)T(F(x*)+F′(x0)(xn-xn-1))= (y-x*)T(F(x*))≥0成立,即{xn}收斂到變分不等式的一個(gè)解x*.得證. 2.2 在解析函數(shù)上的應(yīng)用 (17) 如果F′(x0)不可逆,則定義γ(F,x0)=∞.因此,若 F′(x0)可逆,則由解析性得γ(F,x0)有界.以下引理說(shuō)明當(dāng)函數(shù)F是解析的,則F一定滿足γ-條件. 證明因?yàn)楹瘮?shù)F是解析的,可以寫成冪級(jí)數(shù)的形式為 所以 又因?yàn)镕′(x0)可逆,結(jié)合(17)式得 (18) (19)再次對(duì)(19)等式兩邊同時(shí)乘以γ(x-x0),并相減,得 移項(xiàng)得 再結(jié)合(18)式得 接下來(lái),我們得到了一個(gè)關(guān)于函數(shù)F是解析函數(shù)的牛頓法序列收斂性的推論. ‖xn+1-xn‖ (20) 且{xn}收斂到變分不等式的一個(gè)解x*. 3數(shù)值結(jié)果 任意x=(ξ1,ξ2)∈R2 表1牛頓迭代的次數(shù)與數(shù)值結(jié)果 Table1ThenumberofiterationsandNumericalresultsofNewton 迭代數(shù)/次序列{xn}0(0.95,0.96)1(0.9463,0.9421)2(0.9397,0.9377)3(0.9333,0.9308)4(0.9298,0.9285)5(0.9255,0.9266)6(0.9229,0.9232)7(0.9214,0.9221)8(0.9210,0.9219)9(0.9209,0.9213)10(0.9206,0.9206)11(0.9202,0.9205)12(0.9201,0.9203)13(0.9201,0.9202)14(0.9201,0.9202) 算法在第14步趨于最優(yōu)解x*=(0.92,0.92).說(shuō)明牛頓算法是有效的. 4結(jié)論 參考文獻(xiàn): [1]HARTMAN P, STAMPACCHIA G. On some nonlinear slliptic differential functional equations[J]. Acta mathematica,1966,115:153-188. [2]FAROUQ N E. Pseudomonotone variational inequatlities: convergence of proximal methods[J]. Journal of optimization theory and applications,2001,109(2):311-326. [3]SALMON G, NGUYEN V H,STRODIOT J J. Coupling the auxiliary problem principle and epiconvergence theory to solve general variational inequalities[J]. Journal of optimization theory and applications,2000,104(3):629-657. [4]PENG J M, FUKUSHIMA M. A hybrid Newton method for solving the variational inequalities problems via the d-gap functiona[J]. Mathmatical programming,1999,86:367-386. [5]WU J H. Long-step primal path-following algorithm for monotone variational inequalities problems[J]. Journal of optimization theory and applications,1998,99(2):509-531. [6]FAROUQ N E. Pseudomonotone variational inequalities: convergence of the auxiliary problem method[J]. Journal of optimization theory and applications,2001,111(2):305-326. [7]KANTOROVICH L V,AKILOV G P. Functional analtsis[M].Qxford:Pergamon,1982. [8]KANTOROVICH L V. Functional analysis and applied mathematics[J]. Uspekhi matematicheskikh nauk, 1948(3):89-185. [9]SMALE S. Newton’s method estimates from data at one point[M]. New York: Spring,1986:185-196. [10]SMALE S. Compexity theory and numerical analysis[J] Acta number,1997(6):523-551. [11]WANG X H, HAN D F. Criterion α and Newton’s method[J]. Applied mathematics-a journal of chinese universities,1997,19:96-105. [12]CHANG D C, WANG J H, YAO J C. Newton’s method for variational inequality problems: Smale’s point estimate theory under the -condition[J]. Applicable analysis,2013(1):44-55. [13]FACCHINEI F, PANG J S. Finite-dimensioanl variation inequalities and complementarity problems[M]. New York: Springer-Verlag,2003. [14]王興華,韓丹夫.弱條件下的α判據(jù)和Newton法[J].計(jì)算數(shù)學(xué),1997(2):103-112. [15]WANG X H. Convergence of Newton’s method and inverse function theorem in banach space[J]. Mathematics of computation,1999,68(1):169-186. [16]WANG X H. Convergence of Newton's method and uniqueness of the solution of equations in Banach space[J]. IMA journal of numberical analysis,2000(20):123-134. [17]鮑吉鋒.平衡問(wèn)題和優(yōu)化問(wèn)題若干算法的收斂性分析[D].杭州:浙江大學(xué),2013. (責(zé)任編輯:陳石平) Convergence of Newton method for solving variational inequalities ZHOU Biao, LI Sulan (College of Science, Zhejiang University of Technology, Hangzhou 310023, China) Abstract:Newton method is an important method for solving the problem of variational inequalities, the convergence of the algorithm has been a core issue for the scholars. When the function F in variational inequalities satisfies γ-conditions in , it is proved that the iterative dot sequence produced by the Newton method is valid and can be controlled to converge to an optimal solution of the variational inequalities by a sequence {tn}, which is constructed out. Taking into account that the regional conditions of the function F satisfies γ-conditions in makes more difficult for further study, analytic functions is introduced. The convergence of Newton method under the condition that the F is analytic function is given. Numerical experiments show that the algorithm is effective. Keywords:Newton method; variational inequalities; γ-conditions; semilocal convergence; analytic functions 收稿日期:2015-12-17 基金項(xiàng)目:國(guó)家自然科學(xué)基金資助項(xiàng)目(11371325 ) 作者簡(jiǎn)介:周彪(1991—),男,浙江臺(tái)州人,碩士研究生,研究方向?yàn)樽顑?yōu)化算法與理論,E-mail:zhoubiaozjut@163.com.通信作者:李素蘭副教授,E-mail:sulanli@zjut.edu.cn. 中圖分類號(hào):O224 文獻(xiàn)標(biāo)志碼:A 文章編號(hào):1006-4303(2016)04-0461-05