请输入您要查询的字词:

 

单词 RayleighRitzTheorem
释义

Rayleigh-Ritz theorem


Let A𝐂n×n be a Hermitian matrixMathworldPlanetmath. Then itseigenvectorsMathworldPlanetmathPlanetmathPlanetmath are the critical points (vectors) of the ”Rayleigh quotient”,which is the real function R:n\\{𝟎}

R(𝐱)=𝐱HA𝐱𝐱H𝐱,𝐱0

and its eigenvaluesMathworldPlanetmathPlanetmathPlanetmathPlanetmath are its values at such critical points.

As a consequence, we have:

λmax=max𝐱0𝐱HA𝐱𝐱H𝐱

and

λmin=min𝐱0𝐱HA𝐱𝐱H𝐱

Proof:

First of all, let’s observe that for a hermitian matrix, the number 𝐱HA𝐱 is a real one (actually, <𝐱,A𝐱 𝐱HA𝐱=(AH𝐱)H𝐱=(A𝐱)H𝐱=<A𝐱,𝐱<𝐱,A𝐱>,whence <𝐱,A𝐱𝐱HA𝐱 is real), so that theRayleigh quotient is real as well.

Let’s now compute the critical points 𝐱¯ of theRayleigh quotient, i.e. let’s solve the equations system dR(𝐱¯)d𝐱=𝟎T. Let’s write 𝐱=𝐱(R)+j𝐱(I), 𝐱(R) and 𝐱(I) beingrespectively the real and imaginary part of 𝐱. We have:

dR(𝐱)d𝐱=dR(𝐱)d𝐱(R)+jdR(𝐱)d𝐱(I)

so that we must have:

dR(𝐱¯)d𝐱(R)=dR(𝐱¯)d𝐱(I)=𝟎T

Using derivativesPlanetmathPlanetmath rules, we obtain:

dR(𝐱)d𝐱(R)=dd𝐱(R)(𝐱HA𝐱𝐱H𝐱)=d(𝐱HA𝐱)d𝐱(R)𝐱H𝐱-𝐱HA𝐱d(𝐱H𝐱)d𝐱(R)(𝐱H𝐱)2=d(𝐱HA𝐱)d𝐱(R)-R(𝐱)d(𝐱H𝐱)d𝐱(R)𝐱H𝐱.

Applying matrix calculus rules, we find:

d(𝐱HA𝐱)d𝐱(R)=𝐱HAd𝐱d𝐱(R)+𝐱TATd𝐱d𝐱(R)=𝐱HA+𝐱TAT=𝐱HA+(𝐱HAH)=

and since A=AH,

=𝐱HA+(𝐱HA)=2(𝐱HA).

In a similarPlanetmathPlanetmath way, we get:

d(𝐱H𝐱)d𝐱(R)=2(𝐱H).

Substituting, we obtain:

dR(𝐱)d𝐱(R)=2(𝐱HA)-R(𝐱)(𝐱H)𝐱H𝐱

and, after a transpositionMathworldPlanetmath, equating to the null column vectorMathworldPlanetmath,

𝟎=((𝐱¯HA)-R(𝐱¯)(𝐱¯H))T=
=(AT𝐱¯)-R(𝐱¯)(𝐱¯)=((AH𝐱¯))-R(𝐱¯)(𝐱¯)=
=((A𝐱¯))-R(𝐱¯)(𝐱¯)=(A𝐱¯)-R(𝐱¯)(𝐱¯)

and, since R(𝐱) is real,

(A𝐱¯-R(𝐱¯)𝐱¯)=𝟎

Let’s then evaluate dR(𝐱)d𝐱(I):

dR(𝐱)d𝐱(I)=dd𝐱(I)(𝐱HA𝐱𝐱H𝐱)=d(𝐱HA𝐱)d𝐱(I)𝐱H𝐱-𝐱HA𝐱d(𝐱H𝐱)d𝐱(I)(𝐱H𝐱)2=d(𝐱HA𝐱)d𝐱(I)-R(𝐱)d(𝐱H𝐱)d𝐱(I)𝐱H𝐱.

Applying again matrix calculus rules, we find:

d(𝐱HA𝐱)d𝐱(I)=𝐱HAd𝐱d𝐱(I)+𝐱TATd𝐱d𝐱(I)=j𝐱HA-j𝐱TAT=j(𝐱HA-(𝐱HAH))=

and since A=AH,

=j(𝐱HA-(𝐱HA))=j(2j(𝐱HA))=-2(𝐱HA).

In a similar way, we get:

d(𝐱H𝐱)d𝐱(I)=j𝐱H-j𝐱T=j(𝐱H-(𝐱H))=j(2j(𝐱H))=-2(𝐱H).

Substituting, we obtain:

dR(𝐱)d𝐱(I)=-2(𝐱HA)-R(𝐱)(𝐱H)𝐱H𝐱

and, after a transposition, equating to the null column vector,

𝟎=((𝐱¯HA)-R(𝐱¯)(𝐱¯H))T=
=(AT𝐱¯)-R(𝐱¯)(𝐱¯)=((AH𝐱¯))-R(𝐱¯)(𝐱¯)=
=((A𝐱¯))-R(𝐱¯)(𝐱¯)=-(A𝐱¯)+R(𝐱¯)(𝐱¯)

and, since R(𝐱) is real,

(A𝐱¯-R(𝐱¯)𝐱¯)=𝟎

In conclusionMathworldPlanetmath, we have that a stationary vector 𝐱¯ forthe Rayleigh quotient satisfies the complex eigenvalue equation

A𝐱¯-R(𝐱¯)𝐱¯=𝟎

whence the thesis.

Remarks:

1) The two relationsMathworldPlanetmath λmax=max𝐱0𝐱HA𝐱𝐱H𝐱 and λmin=min𝐱𝟎𝐱HA𝐱𝐱H𝐱 can also be obtained in a simpler way.By Schur’s canonical form theorem, any normal (and hence any hermitian)matrix is unitarily diagonalizable, i.e. a unitary matrixMathworldPlanetmath U exists suchthat A=UΛUH with Λ=diag(λ1,λ2,,λn). So, since all eigenvalues of a hermitian matrix arereal, it’s possible to write:

𝐱HA𝐱=𝐱UHΛUH𝐱=(UH𝐱)HΛ(UH𝐱)=𝐲HΛ𝐲=i=1nλi|yi|2λmaxi=1n|yi|2=
=λmax𝐲H𝐲=λmax(UH𝐱)H(UH𝐱)=λmax(𝐱HUUH𝐱)=λmax(𝐱H𝐱)

whence

λmax𝐱HA𝐱𝐱H𝐱

But, having defined A𝐯M=λmax𝐯M, wehave:

𝐯MHA𝐯M𝐯MH𝐯M=𝐯MHλmax𝐯M𝐯MH𝐯M=λmax

so that

λmax=max𝐱𝟎𝐱HA𝐱𝐱H𝐱

In a much similar way, we obtain

λmin=min𝐱𝟎𝐱HA𝐱𝐱H𝐱

2) The above relations yield the following noteworthing bounds for thediagonal entries of a hermitian matrix:

λminaiiλmax

In fact, having defined

𝐞i=[0,0,,0,i-11,0,,0]T

and observing that aii=𝐞iHA𝐞i𝐞iH𝐞i=R(𝐞i), we have:

λmin=min𝐱𝟎R(𝐱)R(𝐞i)max𝐱𝟎R(𝐱)=λmax.
随便看

 

数学辞典收录了18232条数学词条,基本涵盖了常用数学知识及数学英语单词词组的翻译及用法,是数学学习的有利工具。

 

Copyright © 2000-2023 Newdu.com.com All Rights Reserved
更新时间:2025/5/4 21:47:06