Norms, Condition Numbers, Eigenvalues and Eigenvectors

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

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

APPENDIX A Some Linear Algebra

Errors for Linear Systems

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

Singular Value Decomposition: Theory and Applications

MATH Homework #2

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

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

MEM 255 Introduction to Control Systems Review: Basics of Linear Algebra

2.3 Nilpotent endomorphisms

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

Inner Product. Euclidean Space. Orthonormal Basis. Orthogonal

Report on Image warping

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

Homework Notes Week 7

Perron Vectors of an Irreducible Nonnegative Interval Matrix

Lecture 12: Discrete Laplacian

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

Lecture 3. Ax x i a i. i i

= = = (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.

Quantum Mechanics I - Session 4

CSCE 790S Background Results

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

Linear Approximation with Regularization and Moving Least Squares

LECTURE 9 CANONICAL CORRELATION ANALYSIS

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

Formulas for the Determinant

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

Lecture 10 Support Vector Machines II

MATH Sensitivity of Eigenvalue Problems

Example: (13320, 22140) =? Solution #1: The divisors of are 1, 2, 3, 4, 5, 6, 9, 10, 12, 15, 18, 20, 27, 30, 36, 41,

5 The Rational Canonical Form

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

Lecture 20: Lift and Project, SDP Duality. Today we will study the Lift and Project method. Then we will prove the SDP duality theorem.

Convexity preserving interpolation by splines of arbitrary degree

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

8.4 COMPLEX VECTOR SPACES AND INNER PRODUCTS

The Order Relation and Trace Inequalities for. Hermitian Operators

Composite Hypotheses testing

NUMERICAL DIFFERENTIATION

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

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

BOUNDEDNESS OF THE RIESZ TRANSFORM WITH MATRIX A 2 WEIGHTS

Pattern Classification

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

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

MATH 241B FUNCTIONAL ANALYSIS - NOTES EXAMPLES OF C ALGEBRAS

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

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

Notes on Frequency Estimation in Data Streams

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

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

STAT 309: MATHEMATICAL COMPUTATIONS I FALL 2018 LECTURE 16

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

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

1 Matrix representations of canonical matrices

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

2 More examples with details

Bézier curves. Michael S. Floater. September 10, These notes provide an introduction to Bézier curves. i=0

SELECTED SOLUTIONS, SECTION (Weak duality) Prove that the primal and dual values p and d defined by equations (4.3.2) and (4.3.3) satisfy p d.

763622S ADVANCED QUANTUM MECHANICS Solution Set 1 Spring c n a n. c n 2 = 1.

Inexact Newton Methods for Inverse Eigenvalue Problems

Randić Energy and Randić Estrada Index of a Graph

C/CS/Phy191 Problem Set 3 Solutions Out: Oct 1, 2008., where ( 00. ), so the overall state of the system is ) ( ( ( ( 00 ± 11 ), Φ ± = 1

CHAPTER 5 NUMERICAL EVALUATION OF DYNAMIC RESPONSE

w ). Then use the Cauchy-Schwartz inequality ( v w v w ).] = in R 4. Can you find a vector u 4 in R 4 such that the

Norm Bounds for a Transformed Activity Level. Vector in Sraffian Systems: A Dual Exercise

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

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

Solutions to Problem Set 6

Google PageRank with Stochastic Matrix

p 1 c 2 + p 2 c 2 + p 3 c p m c 2

Econ107 Applied Econometrics Topic 3: Classical Model (Studenmund, Chapter 4)

6.854J / J Advanced Algorithms Fall 2008

Workshop: Approximating energies and wave functions Quantum aspects of physical chemistry

Lecture 10: May 6, 2013

Dynamic Programming. Preview. Dynamic Programming. Dynamic Programming. Dynamic Programming (Example: Fibonacci Sequence)

Tensor Analysis. For orthogonal curvilinear coordinates, ˆ ˆ (98) Expanding the derivative, we have, ˆ. h q. . h q h q

Calculation of time complexity (3%)

DISCRIMINANTS AND RAMIFIED PRIMES. 1. Introduction A prime number p is said to be ramified in a number field K if the prime ideal factorization

Time-Varying Systems and Computations Lecture 6

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

ON A DETERMINATION OF THE INITIAL FUNCTIONS FROM THE OBSERVED VALUES OF THE BOUNDARY FUNCTIONS FOR THE SECOND-ORDER HYPERBOLIC EQUATION

SL n (F ) Equals its Own Derived Group

Feb 14: Spatial analysis of data fields

November 5, 2002 SE 180: Earthquake Engineering SE 180. Final Project

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

Kernel Methods and SVMs Extension

First day August 1, Problems and Solutions

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

COMPLEX NUMBERS AND QUADRATIC EQUATIONS

The Second Anti-Mathima on Game Theory

Developing an Improved Shift-and-Invert Arnoldi Method

Deriving the X-Z Identity from Auxiliary Space Method

Complex Numbers Alpha, Round 1 Test #123

Lecture 5 Decoding Binary BCH Codes

An Inequality for the trace of matrix products, using absolute values

Math 217 Fall 2013 Homework 2 Solutions

Bezier curves. Michael S. Floater. August 25, These notes provide an introduction to Bezier curves. i=0

Transcription:

Norms, Condton Numbers, Egenvalues and Egenvectors 1 Norms A norm s a measure of the sze of a matrx or a vector For vectors the common norms are: N a 2 = ( x 2 1/2 the Eucldean Norm (1a b 1 = =1 N x (1b =1 c = max x (1c and these norms satsfy > 0 ff x 0 (1d 0 = 0 (1e cx = c where c = any scalar (1f x + y y (1g For matrces we would lke the norm to satsfy Ax A and the addtonal requrements Ax A for some partcular x (2a (2b A 0 A = 0 ff a j = 0 ca = c A where c s a real number A + B A + B (trangle nequalty AB A B (Schwartz nequalty (2c (2d (2e (2f (2g 1

The most popular norms are L m (A = ( a m j 1/m j (3a A A = max a,j (the absolute norm (3b,j A E = L 2 (A = a 2 j = tr(aa T (Eucldan, Frobenus (3c A S = max λ(aa T (Spectral (3d ρ(a = max (λ (A, = A S f A = A T (spectral radus (3e Snce tr(aa T =sum of egenvalues, clearly A S < A E Unless otherwse specfed, A means the Eucldean norm 11 Theorems Some useful relatons are: a A S = max x Ax (4a b A S ρ(a (4b c If A S < 1, I + A s non sngular and (4c (I + A 1 1 S 1 A S (4d d (I + A 1 I S 1 n A E A S A E A S 1 A S (4e (4f e If A s non sngular and A 1 B S < 1 then (A + B 1 A 1 S k A 1 S where (4g k = A 1 B S 1 A 1 B S 2 Condton Number The condton number s a measure of the effect on an approxmate nverse of a matrx A (or on an approxmate soluton of Ax = b, when the elements are changed slghtly Several dfferent condton numbers have been suggested 1 Turng s M-condton number M(A = n max,j a j max j a 1 j The average value for M(A s n log(n 2

2 Turnng s N-condton number, N(A = n 1 A A 1 where A = (,j a2,j 1/2 The average value for N(A s n 3 Von Neumann P-condton number P(A = λ max /λ mn The average value for P(A s n When a matrx egenvalues are close to each other, the matrx s well condtoned for solvng lnear equatons However t s ll condtoned for determnng the egenvalues 4 Determnant of the normalzed matrx The th row of A s dvded by ( n j=1 a2,j 1/2 If the determnant s small compared to ±1, the system s ll-condtoned The best condtoned matrces are orthogonal matrces, ther N-condton numbers are 1 and ther M condton numbers are about log(n For matrces wth elements chosen randomly from a normal dstrbuton, the N condton numbers are of the order of n and ther M condton numbers are of the order of n log(n 21 Relatons between the condton numbers 22 Insght nto Condton Numbers M(A n 2 N(A M(A (5 M(A P(A nm(a n (6 Condton numbers are somewhat mpractcal snce they requre ether nvertng the matrx, solvng for egenvalues, or calculatng the determnant all expensve operatons However, consder a small change n the rght hand sde vector, b, that causes a change n the soluton, x It can be shown that f the condton number s defned as A A 1 that x cond(a b b (7 Lkewse f a small change A causes a change x, then x A cond(a A 1 cond(a A A (8 provded that A A 1 < 1 3

23 Condton Number n terms of norms Let Ax M = max x, m = mn x and defne the condton number as Ax (9a Cond(A M m (9b If A(x + x = b + b, then b M and b m x and thus x < Cond(A b b and t can be shown that f we defne (10 that Cond(A = A S A 1 S (11 Cond(A < n 2 λ max λ mn = n 2 ρ(aρ(a 1 (12 If a 1 j s large, A 1 E >> 0, e, ll condtoned But snce A 1 S < A 1 E, A 1 S s not a good measure of ll condtonng However A 2 S = max λ (AA T 1 n A 2 E we have A E n A S A E (13 So A S s an acceptable measure of ll condtonng 24 Does makng A symmetrc mprove the soluton It s often suggested that the set of lnear equatons, Ax = b be solved by multplyng by the transpose of A, A T, gvng but A T Ax = A T b (14a P(A T A P(A and N(A T A N(A (14b so makng the orgnal problem nto one wth a symmetrc matrx actually worsens the method 4

25 Example Let A = ( 41 28 97 66 b = ( 41 97 ( 1 x = 0 (15 Now lettng ( 411 b = 970 ( 001 b = 000 ( 034 then x = 097 ( 066 and x = 097 (16 (17 The one norms are: b = 138, = 1, b = 001, x = 163 and Eq (7 gves or 163 A A 1 72464e 04 (18 A A 1 = 2249 (19 Now the value 2249 s only an approxmaton because dfferent choces of b wll gve dfferent results The P condton number s P(A = λ max = 107093 = 1515 (20 λ mn 00093 26 Matlab and SVD condton numbers Another defnton s that the condton number s the rato of the largest and smallest sngular values obtaned from SVD In ths case, t equals 1623 The condton number obtaned usng condest(a n Matlab s 2249 Also cond(a = cond(a, 2 = cond(a, fro = 1623 and condest(a = cond(a, 1 = cond(a, nf = 2249 Matlab also computes a condton number based upon the 1-norm, =2249, based upon the 2-norm for nvertng A, = 1623, and for computng the egenvalues, = 11893 5

3 Egenvalues and Egenvectors Gven a matrx A, f the vector x s such that Ax = λ x (21 where λ s a scalar, x s sad to be an egenvector and λ s the egenvalue That s for Ae = λe = λie (22a (A λie = 0 (22b e > 0 (22c A λi = 0 (22d If A s n n, there wll be n egenvalues and egenvectors, some of whch may be duplcates The egenvalues satsfy a λ(ab = λ(ba (23a b λ(a 1 = 1/λ(A (23b c λ(a + B = λ(a + λ(b (23c d λ(a m = λ m (A (23d If A = A T, and A s real, λ s real, the egenvectors are lnearly ndependent and span the space In addton λ max a j, aj 2, max j a j, max j a j (24 31 Condton Number related to Egenvalues Snce b must belong to the space the A spans, we can expand t n terms of the egenvectors, e Now b = β 1 e 1 + + β n e n (25 Ae 1 =λ 1 e 1 A 1 Ae 1 =λ 1 A 1 e 1 gvng e 1 =λ 1 A 1 e 1 6

A 1 e 1 = 1 λ 1 e 1 Thus x = A 1 b = β 1 λ 1 e + + β n λ n e n (27 and f a small change n b alters β j and f λ j = λ mn 0, A 1 b stretches the soluton x by a huge amount n the drecton of e j and ntroduces a large error n the soluton 4 Hlbert and Vandermonde Matrces The Vandermonde determnant s defnded as V,j = a j 1 or 1 a 1 a1 n 1 1 a 2 a n 1 1 a n an n 1 2 (28 and arses from nterpolatng the polynomal a 0 + a 1 x + + an 1 on equally spaced data ponts In Matlab the call A = vander(1 : 05 : 3 wll produce the matrx wth a largest value of 81 and a condton number of cond(a = 21217 10 4, condest(a = 39795 10 4 and an nverse wth a largest value of 10 3 Scalng the matrx has no effect on the condton number The Hlbert matrx 1 1/2 1/3 1/(m + 1 1/2 1/3 1/4 1/(m + 2 1/(m + 1 1/(m + 2 1/(m + 3 1/(2m + 1 (29 arses from fttng the polynomal f(x = m j=0 a jφ j (x where φ j (x = x j and s a classc example of an llcondtoned matr 7