The Singular Value Decomposition

Similar documents
The Singular Value Decomposition and Least Squares Problems

Notes on singular value decomposition for Math 54. Recall that if A is a symmetric n n matrix, then A has real eigenvalues A = P DP 1 A = P DP T.

Fall TMA4145 Linear Methods. Exercise set Given the matrix 1 2

(a) If A is a 3 by 4 matrix, what does this tell us about its nullspace? Solution: dim N(A) 1, since rank(a) 3. Ax =

Math 407: Linear Optimization

MA 265 FINAL EXAM Fall 2012

Linear Algebra Formulas. Ben Lee

ELE/MCE 503 Linear Algebra Facts Fall 2018

MAT Linear Algebra Collection of sample exams

SINGULAR VALUE DECOMPOSITION

Recall the convention that, for us, all vectors are column vectors.

Math Bootcamp An p-dimensional vector is p numbers put together. Written as. x 1 x =. x p

Chapter 7: Symmetric Matrices and Quadratic Forms

σ 11 σ 22 σ pp 0 with p = min(n, m) The σ ii s are the singular values. Notation change σ ii A 1 σ 2

235 Final exam review questions

Introduction to Numerical Linear Algebra II

The Singular Value Decomposition

be a Householder matrix. Then prove the followings H = I 2 uut Hu = (I 2 uu u T u )u = u 2 uut u

Summary of Week 9 B = then A A =

UNIT 6: The singular value decomposition.

Math 520 Exam 2 Topic Outline Sections 1 3 (Xiao/Dumas/Liaw) Spring 2008

Chapter 3 Transformations

MATH 304 Linear Algebra Lecture 34: Review for Test 2.

Computational math: Assignment 1

Final Review Written by Victoria Kala SH 6432u Office Hours R 12:30 1:30pm Last Updated 11/30/2015

Lecture 4 Orthonormal vectors and QR factorization

Problem # Max points possible Actual score Total 120

7. Symmetric Matrices and Quadratic Forms

Applied Mathematics 205. Unit II: Numerical Linear Algebra. Lecturer: Dr. David Knezevic

Algorithms to Compute Bases and the Rank of a Matrix

Parallel Singular Value Decomposition. Jiaxing Tan

Properties of Matrices and Operations on Matrices

ANSWERS. E k E 2 E 1 A = B

MATH 240 Spring, Chapter 1: Linear Equations and Matrices

Cheat Sheet for MATH461

1 Last time: least-squares problems

Solutions to Review Problems for Chapter 6 ( ), 7.1

Lecture 7: Positive Semidefinite Matrices

Computational Methods. Eigenvalues and Singular Values

2. Every linear system with the same number of equations as unknowns has a unique solution.

1. The Polar Decomposition

Problem Set (T) If A is an m n matrix, B is an n p matrix and D is a p s matrix, then show

Lecture notes on Quantum Computing. Chapter 1 Mathematical Background

Control Systems. Linear Algebra topics. L. Lanari

2. Linear algebra. matrices and vectors. linear equations. range and nullspace of matrices. function of vectors, gradient and Hessian

Glossary of Linear Algebra Terms. Prepared by Vince Zaccone For Campus Learning Assistance Services at UCSB

Math 308 Practice Test for Final Exam Winter 2015

AMS526: Numerical Analysis I (Numerical Linear Algebra)

MATH36001 Generalized Inverses and the SVD 2015

Orthonormal Transformations and Least Squares

Linear Algebra. Session 12

Warm-up. True or false? Baby proof. 2. The system of normal equations for A x = y has solutions iff A x = y has solutions

Part I True or False. (One point each. A wrong answer is subject to one point deduction.)

Linear Algebra Review. Vectors

Least Squares. Tom Lyche. October 26, Centre of Mathematics for Applications, Department of Informatics, University of Oslo

Singular Value Decomposition (SVD)

ELEMENTS OF MATRIX ALGEBRA

Orthogonal Transformations

MATH 304 Linear Algebra Lecture 23: Diagonalization. Review for Test 2.

Chapters 5 & 6: Theory Review: Solutions Math 308 F Spring 2015

Chapter 1. Matrix Algebra

Matrix decompositions

The Singular Value Decomposition

Topic 2 Quiz 2. choice C implies B and B implies C. correct-choice C implies B, but B does not imply C

Eigenspaces. (c) Find the algebraic multiplicity and the geometric multiplicity for the eigenvaules of A.

NOTES on LINEAR ALGEBRA 1

Review problems for MA 54, Fall 2004.

Mathematical Foundations of Applied Statistics: Matrix Algebra

Conceptual Questions for Review

Numerical Linear Algebra A Solution Manual. Georg Muntingh and Christian Schulz

Lecture notes: Applied linear algebra Part 1. Version 2

Orthonormal Bases; Gram-Schmidt Process; QR-Decomposition

Lecture 2: Linear Algebra

Linear Least Squares. Using SVD Decomposition.

Section 6.4. The Gram Schmidt Process

Feb 10/ True or false: if A is a n n matrix, Ax = 0 has non-zero solution if and only if there are some b such that Ax = b has no solution.

TBP MATH33A Review Sheet. November 24, 2018

EIGENVALUES AND SINGULAR VALUE DECOMPOSITION

Linear Algebra (Review) Volker Tresp 2018

MATH 31 - ADDITIONAL PRACTICE PROBLEMS FOR FINAL

Symmetric Matrices and Eigendecomposition

Math 102 Final Exam - Dec 14 - PCYNH pm Fall Name Student No. Section A0

YORK UNIVERSITY. Faculty of Science Department of Mathematics and Statistics MATH M Test #2 Solutions

linearly indepedent eigenvectors as the multiplicity of the root, but in general there may be no more than one. For further discussion, assume matrice

EE731 Lecture Notes: Matrix Computations for Signal Processing

Singular Value Decomposition and Polar Form

Lecture Summaries for Linear Algebra M51A

Math 265 Linear Algebra Sample Spring 2002., rref (A) =

LINEAR ALGEBRA SUMMARY SHEET.

Lecture 1: Review of linear algebra

Spring 2014 Math 272 Final Exam Review Sheet

SUMMARY OF MATH 1600

Math 102, Winter Final Exam Review. Chapter 1. Matrices and Gaussian Elimination

The QR Factorization

Let A an n n real nonsymmetric matrix. The eigenvalue problem: λ 1 = 1 with eigenvector u 1 = ( ) λ 2 = 2 with eigenvector u 2 = ( 1

This can be accomplished by left matrix multiplication as follows: I

Jim Lambers MAT 610 Summer Session Lecture 1 Notes

[3] (b) Find a reduced row-echelon matrix row-equivalent to ,1 2 2

EIGENVALE PROBLEMS AND THE SVD. [5.1 TO 5.3 & 7.4]

Preliminary/Qualifying Exam in Numerical Analysis (Math 502a) Spring 2012

Transcription:

CHAPTER 6 The Singular Value Decomposition Exercise 67: SVD examples (a) For A =[, 4] T we find a matrixa T A =5,whichhastheeigenvalue =5 Thisprovidesuswiththesingularvalue =+ p =5forA Hence the matrix A has rank and a SVD of the form A = U U 5 0 V, with U, U R,, V = V R The eigenvector of A T A that corresponds to the eigenvalue =5isgivenbyv =, providing us with V = Using part of Theorem 65, one finds u = [, 5 4]T Extending u to an orthonormal basis for R gives u = [ 4, 5 ]T A SVD of A is therefore A = 4 5 5 4 0 (b) One has A = 4 5, A T = 9 9, AT A = 9 9 The eigenvalues of A T A are the zeros of det(a T A I) =(9 ) 8, yielding =8and =0,andtherefore = p 8 and = 0 Note that since there is only one nonzero singular value, the rank of A is one Following the dimensions of A, one finds p 8 0 = 4 0 05 0 0 The normalized eigenvectors v, v of A T A corresponding to the eigenvalues the columns of the matrix V =[v v ]= p, are Using part of Theorem 65 one finds u,whichcanbeextendedtoanorthonormal basis {u, u, u } using Gram-Schmidt Orthogonalization (see Theorem 49) The vectors u, u, u constitute a matrix U =[u u u ]= 4 5 4

A SVD of A is therefore given by A = 4 p 8 0 5 4 0 05 p 0 0 Exercise 68: More SVD examples (a) We have A = e and A T A = e T e = This gives the eigenpair (, v )= (, ) of A T A Hence =and = e = A As = A and V = I we must have U = I m yielding a singular value decomposition A = I m e I (b) For A = e T n,thematrix 0 0 0 A T A = 6 7 4 0 0 0 5 0 0 has eigenpairs (0, e j )forj =,,n and(, e n ) Then = e T R,n and V = e n, e n,,e R n,n Using part of Theorem 65 we get u =,yielding U = A SVD for A is therefore given by A = e T n = e T en, e n,,e (c) In this exercise A = 0 0, AT = A, A T A = 0 0 9 The eigenpairs of A T A are given by (, v )=(9, e )and(, v )=(, e ), from which we find 0 0 = 0, V = 0 Using part of Theorem 65 one finds u = e and u = matrix 0 U = 0 A SVD of A is therefore given by 0 0 0 A = 0 0 0 e, which constitute the Exercise 66: Counting dimensions of fundamental subspaces Let A have singular value decomposition U V By parts and of Theorem 65, span(a) andspan(a )arevectorspacesof the same dimension r, implyingthatrank(a) =rank(a ) This statement is known as the rank-nullity theorem, anditfollowsimmediately from combining parts and 4 in Theorem 65 As rank(a )=rank(a) by,thisfollowsbyreplacinga by A in 5

Exercise 67: Rank and nullity relations Let A = U V be a singular value decomposition of a matrix A C m n By part 5 of Theorem 64, rank(a) isthenumberofpositiveeigenvaluesof AA = U V V U = UDU, where D := is a diagonal matrix with real nonnegative elements Since UDU is an orthogonal diagonalization of AA,thenumberofpositiveeigenvaluesofAA is the number of nonzero diagonal elements in D Moreover, rank(aa )isthenumber of positive eigenvalues of AA (AA ) = AA AA = U V = UD U, which is the number of nonzero diagonal elements in D,sothatrank(A) =rank(aa ) From a similar argument for rank(a A), we conclude that rank(a) =rank(aa )=rank(a A) Let r := rank(a) =rank(a )=rank(aa )=rank(a A) Applying Theorem 64, parts and 4, to the singular value decompositions A = U V, A = V U, AA = U U, A A = V V, one finds that {v r+,,v n } is a basis for both ker(a)andker(a A), while {u r+,u m } is a basis for both ker(a )andker(aa ) In particular it follows that null(a) =null(a A), null(a )=null(aa ), which is what needed to be shown Given is the matrix A = 4 4 6 5 Exercise 68: Orthonormal bases example With B as defined its is straightforward to find these spaces Let us therefore set B = A T instead We have that A = U V T and B = V T U T,with V = 4 5 0 0, =, U = 4 0 0 5 4 From Theorem 65 we know that V forms an orthonormal basis for span(a )= span(b), V an orthonormal basis for ker(a)andu an orthonormal basis for ker(a T )= ker(b) Hence span(b) = v + v, ker(a) = v and ker(b) =0 Exercise 69: Some spanning sets The matrices A C m n and A A have the same rank r since they have the same number of singular values, so that the vector spaces span(a A)andspan(A )have the same dimension It is immediate from the definition that span(a A) span(a ), and therefore span(a A) = span(a ) From Theorem 65 we know that the columns of V form an orthonormal basis for span(a ), and the result follows 6

Exercise 60: Singular values and eigenpair of composite matrix Given is a singular value decomposition A = U V Let r = rank(a), so that r > 0and r+ = = n =0 LetU =[U, U ]andv =[V, V ] be partitioned accordingly and =diag(,, r) asinequation(67),sothat A = U V forms a singular value factorization of A By Theorem 65, 0 A ui Avi Cp i = A = 0 v i A = ip i for i =,,r u i 0 p i for i = r +,,n 0 A Cq i = A 0 ui vi Avi = A = u i iq i for i =,,r 0 q i for i = r +,,n 0 A Cr j = A 0 uj 0 0 0 = A = u j 0 =0 r j, for j = n +,,m This gives a total of n + n +(m n) =m + n eigen pairs Exercise 66: Rank example We are given the singular value decomposition 6 0 0 A = U V T = 6 7 60 6 0 7 4 4 5 40 0 05 0 0 0 Write U =[u, u, u, u 4 ]andv =[v, v, v ] Clearly r =rank(a) = (a) A direct application of Theorem 65 with r =gives {u, u } is an orthonormal basis for span(a), {u, u 4 } is an orthonormal basis for ker(a T ), {v, v } is an orthonormal basis for span(a T ), {v } is an orthonormal basis for ker(a) Since U is orthogonal, {u, u, u, u 4 } is an orthonormal basis for R 4 In particular u, u 4 are orthogonal to u, u,sothattheyspantheorthogonalcomplementspan(a)? to span(a) =span{u, u } (b) Applying Theorem 65 with r =yields q ka Bk F + = p 6 +0 =6 (c) Following Section 6, with D 0 := diag(, 0,,0) R n,n,take D A = A 0 0 := U 0 VT = 6 7 4 5 5 7

Exercise 67: Another rank example (a) The matrix B =(b ij ) ij R n,n is defined by 8 if i = j; >< if i<j; b ij = >: n if (i, j) =(n, ); 0 otherwise while the column vector x =(x j ) j R n is given by if j = n; x j = n j otherwise For the final entry in the matrix product Bx one finds that (Bx) n = b nj x j = b n x + b nn x n = n n + =0 j= For any of the remaining indices i 6= n, thei-th entry of the matrix product Bx can be expressed as Xn (Bx) i = b ij x j = b in + n j b ij j= = + n i b ii + j= Xn = + n i Xn j=i+ j=i+ n j b ij n j i = + n i n i j 0 =0 = + n i n i j 0 n i (n i) = + n i n i =0 As B has a nonzero kernel, it must be singular The matrix A, ontheotherhand, is nonsingular, as its determinant is ( ) n 6=0 ThematricesA and B di er only in their (n, )-th entry, so one has ka Bk F = p a n b n = n Inotherwords,the tiniest perturbation can make a matrix with large determinant singular (b) Let n 0bethesingularvaluesofA Applying Theorem 65 for r =rank(b) <n, we obtain n q r+ + + n = min ka Ck F ka Bk F = n CR n,n rank(c)=r We conclude that the smallest singular value n can be at most n 8