# 共轭梯度法

dTiAe(j)=0(i<j) d i T A e ( j ) = 0 ( i < j )

span{d1,...,dn}=span{ui,...,un}=span{r(1),...,r(n)} s p a n { d 1 , . . . , d n } = s p a n { u i , . . . , u n } = s p a n { r ( 1 ) , . . . , r ( n ) }

r(j) r ( j ) $r_{(j)}$ di(i<j) d i ( i < j ) $d_i(i张成的子空间正交，因此也和所有的 r(i)(i<j) r ( i ) ( i < j ) $r_{(i)}(i正交。

# 共轭梯度法的优势

${r}_{\left(i+1\right)}={r}_{\left(i\right)}-{\alpha }_{\left(i\right)}A{d}_{i}$

${D}_{\left(i\right)}=span\left\{{d}_{1},...,{d}_{n}\right\}=span\left\{{d}_{1},A{d}_{1},{A}^{2}{d}_{1},...,{A}^{i-1}{d}_{1}\right\}=span\left\{{r}_{\left(1\right)},A{r}_{\left(1\right)},{A}^{2}{r}_{\left(1\right)},...,{A}^{i-1}{r}_{\left(1\right)}\right\}$

βij=r(i)AdjdTjAdj β i j = r ( i ) A d j d j T A d j

rT(i)r(j+1)=rT(i)(r(j)α(j)Adj)=rT(i)r(j)α(j)rT(i)AdjrT(i)Adj=1α(j)[rT(i)r(j)rT(i)r(j+1)] r ( i ) T r ( j + 1 ) = r ( i ) T ( r ( j ) − α ( j ) A d j ) = r ( i ) T r ( j ) − α ( j ) r ( i ) T A d j r ( i ) T A d j = 1 α ( j ) [ r ( i ) T r ( j ) − r ( i ) T r ( j + 1 ) ]

(i=j+1):βij=1α(j)rT(i)r(i)dTjAdj(i>j+1):βij=0 ( i = j + 1 ) : β i j = − 1 α ( j ) r ( i ) T r ( i ) d j T A d j ( i > j + 1 ) : β i j = 0

α(j)=dTjr(j)dTjAdj α ( j ) = d j T r ( j ) d j T A d j

βi=βi,i1=rT(i)r(i)dTi1r(i1)=rT(i)r(i)rT(i1)r(i1) β i = β i , i − 1 = − r ( i ) T r ( i ) d i − 1 T r ( i − 1 ) = − r ( i ) T r ( i ) r ( i − 1 ) T r ( i − 1 )

• 2
点赞
• 0
评论
• 4
收藏
• 一键三连
• 扫一扫，分享海报

11-08
09-02 3079

04-09 1248
12-18 1万+
12-21 1058
08-31 2803
08-31 5137
07-14 1万+
07-18 210
06-27 1457
06-11 108
08-07 6376
11-16 5万+
04-18 319