原创力文档- 1、本文档共6页，可阅读全部内容。
- 2、原创力文档（book118）网站文档一经付费（服务费），不意味着购买了该文档的版权，仅供个人/单位学习、研究之用，不得用于商业用途，未经授权，严禁复制、发行、汇编、翻译或者网络传播等，侵权必究。
- 3、本站所有内容均由合作方或网友上传，本站不对文档的完整性、权威性及其观点立场正确性做任何保证或承诺！文档内容仅供研究参考，付费前请自行鉴别。如您付费，意味着您自己接受本站规则且自行承担风险，本站不退款、不进行额外附加服务；查看《如何避免下载的几个坑》。如果您已付费下载过本站文档，您可以点击 这里二次下载。
- 4、如文档侵犯商业秘密、侵犯著作权、侵犯人身权等，请点击“版权申诉”（推荐），也可以打举报电话：400-050-0827(电话支持时间：9:00-18:30)。
查看更多
共轭梯度法的全局收敛性_英文_
JOURNAL OF EN GINEERIN G MATHEMATICS
Article ID :100523085 (2003) 0120082205
Convergence Properties of Conjugate
Ξ
Gra dient Methods
L IAN Shu2jun1 ,2 ,
WANG Chang2yu1 ,2 , L I J i2bao2
(12Department of Applied Mathematics , Dalian University of Technology , Liaoning 116024 ;
22Academy of Mathematics Sciences , Qufu Normal University , Shandong 273165)
Abstract : The global convergence is considered for any conjugate gradient method of the form d1 = - g1 , dk = - gk +
βk dk - 1 ( k Ε 2) with anyβk connected withβk , and with the strong Wolfe line search conditions. The suffi2
PR
cient descent property and the global convergence are proved for this method , without assuming the convexity
of the objective function.
Key words : conjugate gradient method ; strong Wolfe line search ; global convergence
Classif ication : AMS(2000) 90C25
CLC number : O221. 2
Document code : A
1
Introduction
Conjugate gradient methods are very useful for minimizing a smooth function of n variable , minf ( x)
x ∈Rn
Especially if n is large . They are of the following form
= xk + λk dk
(1)
xk + 1
-
-
gk ,
for
for
k = 1
k Ε 2
dk =
(2)
+ βk dk - 1 ,
gk
gk = A f ( xk ) , βk is a scalar and λk is a step size obtained by means of a one2dimensional
Where
search.
Well2known formulae for βk are called the Fletcher2Reeves ( FR) , Polak2Ribiere2Polyak ( PRP) ,
and Hestenes2Stiefel ( HS) formulas , which are gieven respectively by
βFR
2
2
= ‖gk ‖ /‖gk - 1 ‖
(3)
(4)
k
βPR
T
2
= [ g k ( gk
- gk - 1 ) ] /‖gk - 1 ‖
k
βHS
= [ gT ( g - g ) ] /[ dT ( g - g ) ]
(5)
k
k k k - 1
k - 1 k k - 1
Gilbert and Nocedal1 investigated the convergent properties of the Polak2Ribiére method. They con2
sidered the case whereβPR 0 for all sufficiently large k , and showed that the global convergence hol2
k
ds provided that the line search has the following properties :
( I) it satisfies the strong Wolfe conditions :
T
f ( xk ) - f ( xk +
文档评论（0）