ρ some λ THE INVERSE POWER METHOD (or INVERSE ITERATION) , for , or (more usually) to

Similar documents
Lecture 3. Ax x i a i. i i

Errors for Linear Systems

CME 302: NUMERICAL LINEAR ALGEBRA FALL 2005/06 LECTURE 13

U.C. Berkeley CS294: Beyond Worst-Case Analysis Luca Trevisan September 5, 2017

STAT 309: MATHEMATICAL COMPUTATIONS I FALL 2018 LECTURE 16

Norms, Condition Numbers, Eigenvalues and Eigenvectors

Singular Value Decomposition: Theory and Applications

P A = (P P + P )A = P (I P T (P P ))A = P (A P T (P P )A) Hence if we let E = P T (P P A), We have that

Lectures - Week 4 Matrix norms, Conditioning, Vector Spaces, Linear Independence, Spanning sets and Basis, Null space and Range of a Matrix

Inexact Newton Methods for Inverse Eigenvalue Problems

Chapter 5. Solution of System of Linear Equations. Module No. 6. Solution of Inconsistent and Ill Conditioned Systems

Some Comments on Accelerating Convergence of Iterative Sequences Using Direct Inversion of the Iterative Subspace (DIIS)

Stanford University CS359G: Graph Partitioning and Expanders Handout 4 Luca Trevisan January 13, 2011

Eigenvalues of Random Graphs

Structure and Drive Paul A. Jensen Copyright July 20, 2003

Composite Hypotheses testing

= = = (a) Use the MATLAB command rref to solve the system. (b) Let A be the coefficient matrix and B be the right-hand side of the system.

Salmon: Lectures on partial differential equations. Consider the general linear, second-order PDE in the form. ,x 2

MATH Sensitivity of Eigenvalue Problems

Finding The Rightmost Eigenvalues of Large Sparse Non-Symmetric Parameterized Eigenvalue Problem

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

1 GSW Iterative Techniques for y = Ax

princeton univ. F 17 cos 521: Advanced Algorithm Design Lecture 7: LP Duality Lecturer: Matt Weinberg

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

U.C. Berkeley CS294: Spectral Methods and Expanders Handout 8 Luca Trevisan February 17, 2016

Lecture Notes on Linear Regression

Hongyi Miao, College of Science, Nanjing Forestry University, Nanjing ,China. (Received 20 June 2013, accepted 11 March 2014) I)ϕ (k)

Point cloud to point cloud rigid transformations. Minimizing Rigid Registration Errors

Prof. Dr. I. Nasser Phys 630, T Aug-15 One_dimensional_Ising_Model

APPENDIX A Some Linear Algebra

Linear Approximation with Regularization and Moving Least Squares

PHYS 215C: Quantum Mechanics (Spring 2017) Problem Set 3 Solutions

Lecture 12: Discrete Laplacian

LINEAR REGRESSION ANALYSIS. MODULE IX Lecture Multicollinearity

A Hybrid Variational Iteration Method for Blasius Equation

2.3 Nilpotent endomorphisms

LECTURE 9 CANONICAL CORRELATION ANALYSIS

Georgia Tech PHYS 6124 Mathematical Methods of Physics I

Vector Norms. Chapter 7 Iterative Techniques in Matrix Algebra. Cauchy-Bunyakovsky-Schwarz Inequality for Sums. Distances. Convergence.

Developing an Improved Shift-and-Invert Arnoldi Method

A New Refinement of Jacobi Method for Solution of Linear System Equations AX=b

332600_08_1.qxp 4/17/08 11:29 AM Page 481

Matrix Approximation via Sampling, Subspace Embedding. 1 Solving Linear Systems Using SVD

Quantum Mechanics I - Session 4

Report on Image warping

The Exact Formulation of the Inverse of the Tridiagonal Matrix for Solving the 1D Poisson Equation with the Finite Difference Method

Deriving the X-Z Identity from Auxiliary Space Method

n α j x j = 0 j=1 has a nontrivial solution. Here A is the n k matrix whose jth column is the vector for all t j=0

Fall 2012 Analysis of Experimental Measurements B. Eisenstein/rev. S. Errede

A property of the elementary symmetric functions

However, since P is a symmetric idempotent matrix, of P are either 0 or 1 [Eigen-values

Lecture 21: Numerical methods for pricing American type derivatives

A Tuned Preconditioner for Inexact Inverse Iteration Applied to Hermitian Eigenvalue Problems

Some modelling aspects for the Matlab implementation of MMA

MMA and GCMMA two methods for nonlinear optimization

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

Hidden Markov Models & The Multivariate Gaussian (10/26/04)

The Prncpal Component Transform The Prncpal Component Transform s also called Karhunen-Loeve Transform (KLT, Hotellng Transform, oregenvector Transfor

The Gaussian classifier. Nuno Vasconcelos ECE Department, UCSD

C4B Machine Learning Answers II. = σ(z) (1 σ(z)) 1 1 e z. e z = σ(1 σ) (1 + e z )

10-801: Advanced Optimization and Randomized Methods Lecture 2: Convex functions (Jan 15, 2014)

Transfer Functions. Convenient representation of a linear, dynamic model. A transfer function (TF) relates one input and one output: ( ) system

e - c o m p a n i o n

A GLOBAL ARNOLDI METHOD FOR LARGE NON-HERMITIAN EIGENPROBLEMS WITH SPECIAL APPLICATIONS TO MULTIPLE EIGENPROBLEMS. Congying Duan and Zhongxiao Jia*

Least Squares Fitting of Data

The equation of motion of a dynamical system is given by a set of differential equations. That is (1)

Perron Vectors of an Irreducible Nonnegative Interval Matrix

Math 217 Fall 2013 Homework 2 Solutions

Module 3 LOSSY IMAGE COMPRESSION SYSTEMS. Version 2 ECE IIT, Kharagpur

CHALMERS, GÖTEBORGS UNIVERSITET. SOLUTIONS to RE-EXAM for ARTIFICIAL NEURAL NETWORKS. COURSE CODES: FFR 135, FIM 720 GU, PhD

First day August 1, Problems and Solutions

Lecture 6/7 (February 10/12, 2014) DIRAC EQUATION. The non-relativistic Schrödinger equation was obtained by noting that the Hamiltonian 2

Lecture 4: Constant Time SVD Approximation

Solutions to exam in SF1811 Optimization, Jan 14, 2015

Least squares cubic splines without B-splines S.K. Lucas

Consistency & Convergence

Accurate eigenvalue decomposition of arrowhead matrices and applications

Chapter 12 Analysis of Covariance

A combinatorial problem associated with nonograms

Review of Taylor Series. Read Section 1.2

Time-Varying Systems and Computations Lecture 6

Chat eld, C. and A.J.Collins, Introduction to multivariate analysis. Chapman & Hall, 1980

The Geometry of Logit and Probit

Min Cut, Fast Cut, Polynomial Identities

CSci 6974 and ECSE 6966 Math. Tech. for Vision, Graphics and Robotics Lecture 21, April 17, 2006 Estimating A Plane Homography

In this section is given an overview of the common elasticity models.

The Finite Element Method: A Short Introduction

Dr. Shalabh Department of Mathematics and Statistics Indian Institute of Technology Kanpur

Unified Subspace Analysis for Face Recognition

Problem Set 9 Solutions

Grover s Algorithm + Quantum Zeno Effect + Vaidman

ALGORITHM FOR THE CALCULATION OF THE TWO VARIABLES CUBIC SPLINE FUNCTION

Feature Selection: Part 1

Non-negative Matrices and Distributed Control

Statistical Mechanics and Combinatorics : Lecture III

form, and they present results of tests comparng the new algorthms wth other methods. Recently, Olschowka & Neumaer [7] ntroduced another dea for choo

A FORMULA FOR COMPUTING INTEGER POWERS FOR ONE TYPE OF TRIDIAGONAL MATRIX

Complex Numbers. x = B B 2 4AC 2A. or x = x = 2 ± 4 4 (1) (5) 2 (1)

Linear Regression Analysis: Terminology and Notation

College of Computer & Information Science Fall 2009 Northeastern University 20 October 2009

Transcription:

THE INVERSE POWER METHOD (or INVERSE ITERATION) -- applcaton of the Power method to A some fxed constant ρ (whch s called a shft), x λ ρ If the egenpars of A are { ( λ, x ) } ( ), or (more usually) to, then the egenpars of ( ( A ρ I), for A ρ I) are that Let (0) be the startng vector and (assumng that A s nondefectve) suppose ( 0) () () ( n) α x + α x + + α n x = L Then () = ( A ρi ) α = x () (0) α + x () α n + L+ x n ( n) If ρ some λ, then >> for, whch mples that the teraton wll converge very fast NOTE Shfts are much more effectve for the nverse Power method than for the Power method They enable you to compute any egenvector, not ust those assocated wth the largest or the smallest egenvalue 96

97 EXAMPLE = 5 0 0 4 0 A has egenpars 4,,, 4,

Use of the INVERSE POWER METHOD You need a reasonably good approxmaton to some egenvalue -- ths s obtaned by other means You do not explctly compute lnear system ( A ρ I) Instead, for example, you solve the () (0) ( A ρ I) = () for You need to compute the LU factorzaton of A ρi only once, and then use t to solve a lnear system n each teraton If A s a full matrx, the cost s ( / ) n flops () for the LU factorzaton and n flops to compute each vector These costs are much cheaper f A s upper Hessenberg or trdagonal As ρ some egenvalue λ, A ρi becomes very close to a sngular matrx, whch suggests that ( + ) ( ) ( A ρ I) = may be dffcult to solve; A ρi wll lely be very ll-condtoned However, ths algorthm wors well n practce -- see the dscusson on page 4 Usng, for example, Gaussan elmnaton wth partal pvotng, the computed ( +) approxmaton to s the exact soluton of some perturbed lnear system δa ( + ) ( ) ( A + δ A ρi ) ˆ =, where s very small In ths case, ρ s very lely A ρi ( ) also a good approxmaton to an egenvalue of A + δa, and thus ˆ + should be very close to an egenvector of A + δa, whch should then also be very close to an egenvector of A Thus, nverse teraton s usually very effectve From above, how effectve t s depends on whether or not a small perturbaton δ A n A ρi results n only a small perturbaton n ts egenvalue and egenvector That s, ths depends on the condton number of ths egenvalue and egenvector of A ρi and not on the condton number of A ρi (wth respect to solvng the above lnear system) In other words, the effectveness of nverse teraton does not depend on the condton number κ ( A ρi) THE RAYLEIGH QUOTIENT (page 4) Gven a (complex) vector x wth n entres and a (complex) Raylegh uotent s n n matrx A, the 98

T ρ x Ax x Ax = or n the real case T Theorem 54 (page 5) Gven A and x, let r ( µ ) = Ax µ x Then r (µ) s mnmzed when x Ax µ =, the Raylegh uotent Proof mn µ Ax µ x s a least-suares problem n varable, namely µ Its assocated lnear system s the n over-determned lnear system x µ = Ax The soluton can be obtaned from the normal euatons (a lnear system), whch s ( x x) µ = x Ax, from whch t follows that x Ax µ = Alternatve proof: Gven A and x, { Ax µx} s a lne (a -dmensonal subspace) n the vector space of complex vectors wth n entres Ax {Ax - µx} x 99

mn µ Therefore, Ax µ x s obtaned when µ s such that Ax µ x s orthogonal to x ( Ax µ x, x) = 0 x Ax µ = 0 x Ax µ = NOTE If x s an egenvector of A (suppose that sdes by x, Ax = λx ), then on multplyng both x Ax = λ λ = x Ax That s, the Raylegh uotent wth respect to an egenvector s eual to an egenvalue Thus, Ax µ x = 0 f and only f ( µ, x) s an egenpar of A, and f x s not an egenvector of A, then Ax µ x 0 but x Ax mn Ax µ x occurs when µ = µ The pont of the above dscusson: gven A and x (and thnng of x as an approxmaton to an egenvector), the Raylegh uotent s the best approxmaton (n the above sense) to an egenvalue Gven a vector, the followng result ndcates how good an approxmaton to an egenvalue the Raylegh uotent s, n terms of how close s to an egenvector Theorem 55 (page 6) Suppose Av = λv, where, and let ρ = A = A Then = A v = 00

Proof v Av Av = λ v λ = = v Av v v Therefore, = v Av A = v Av v A + v A A = v A( v ) + ( v ) A whch mples that v A( v ) + ( v ) A Recall the Cauchy-Schwarz neualty: ( x, y) x y, where ( x, y) = y x Ths gves Smlarly, v A( v ) v A( v ) = A( v ) A v ( v ) A v A Thus, from above, A v EXAMPLE If v < ε, then A ε That s, f v s O (ε ), then s also O (ε ) 0

THE RAYLEIGH QUOTIENT ITERATION -- a varaton of the nverse Power method -- a dfferent shft s used at each step, namely the Raylegh uotent of the current approxmate egenvector Algorthm Choose a startng vector x such that = (that s, x ) Repeat untl convergence: ρ x Ax apply the nverse Power method: solve ( A ρ I) xˆ = x for xˆ xˆ x (that s, normalze xˆ ) xˆ Because a dfferent shft s used at every step, ths algorthm s not guaranteed to converge to an egenvector However, n practce t very seldom fals, and usually converges very rapdly to an egenvector of ( A ρ I) correspondng to the domnant egenvalue of (A ρi), for some ρ Snce { ρ } computed n Step are Raylegh uotents, {} x an egenvector of A correspondng to an egenvalue λ ρ of A The Raylegh uotent of ths computed egenvector then gves the approxmaton to λ = 0

CONVERGENCE OF THE RAYLEIGH QUOTIENT ALGORITHM (pages 7-8) If convergent, then the Raylegh Quotent algorthm converges uadratcally (order ) The proof asumes that the egenvalue λ (to whch t converges) s smple (multplcty s ) A setch of the argument: Suppose that the egenpars of A are ( λ t, vt ) The Raylegh Quotent algorthm computes { ρt } λ (assumed to be a smple egenvalue) { t } egenvector v Suppose that λ s the closest egenvalue to λ Analyss of the Inverse Power method gves v + r v, where r s the rato of the largest egenvalues of error of the r ) ( A ρ I) st ( + approxmaton to + error of the th approxmaton to v s to v That s, For suffcently large, snce s the largest egenvalue of ( A ρ I), r = = λ λ A v λ λ by Thm 55 Therefore, v A + λ λ v v, whch s uadratc convergence NOTE If A s a real symmetrc (or Hermtan) matrx, the Raylegh Quotent algorthm s cubcally convergent: 0

v constant + v Reason: for Hermtan matrces A, the result of Theorem 55 s stronger -- f v = O( ) and ρ = A, then = O( ε ) For a dscusson of ths, see the ε mddle of page 8 COST OF THE RAYLEIGH QUOTIENT ALGORITHM (page 9) O ( n ) flops per teraton for a full matrx, snce the LU factorzaton must be recomputed for each teraton However, the cost s only O ( n ) flops per teraton f A s upper Hessenberg See Exercse 5 04