Download presentation
Presentation is loading. Please wait.
Published bySabina Lynne Davidson Modified over 9 years ago
1
Extremum Properties of Orthogonal Quotients Matrices By Achiya Dax Hydrological Service, Jerusalem, Israel e-mail: dax20@water.gov.il
2
The Eckart – Young Theorem ( 1936 ) says that the “ truncated SVD” matrix A k = U k D k V k T solves the least norm problem minimize F ( B ) = || A - B || F 2 subject to rank ( B ) k ( Also called the Schmidt-Mirsky Theorem. )
3
Ky Fan’s Maximum Principle ( 1949 ) S a symmetric positive semi-definite n x n matrix Y k the set of orthogonal n x k matrices 1 + … + k = max { trace ( Y k T S Y k ) | Y k Y k } Solution is obtained for the Spectral matrix V k = [v 1, v 2, …, v k ]. giving the sum of the largest k eigenvalues.
4
Outline * The Eckart – Young Theorem and Orthogonal Quotients matrices. * The Orthogonal Quotients Equality. * The Symmetric Quotients Equality and Ky Fan’s extremum principles. * An Extended Extremum Principle.
5
The Singular Value Decomposition A = U V T = diag { 1, 2, …, p }, p = min { m, n } U = [u 1, u 2, …, u p ], U T U = I V = [v 1, v 2, …, v p ], V T V = I A V = U A T U = V A v j = j u j, A T u j = j v j j = 1, …, p.
6
Low - Rank Approximations A k = U k D k V k T D k = diag { 1, 2, …, k }, U k = [u 1, u 2, …, u k ], U k T U k = I V k = [v 1, v 2, …, v k ], V k T V k = I
7
The Eckart – Young Theorem ( 1936 ) says that the “truncated SVD” matrix A k = U k D k V k T solves the least norm problem minimize F ( B ) = || A - B || F 2 subject to rank ( B ) k ( Also called the Schmidt-Mirsky Theorem. )
8
Rank - k Matrices B = X k R k Y k T where R k is a k x k matrix X k = [x 1, x 2, …, x k ], X k T X k = I Y k = [y 1, y 2, …, y k ], Y k T Y k = I
9
The Eckart – Young Problem can be rewritten as minimize F ( B ) = || A - X k R k Y k T || F 2 subject to X k T X k = I and Y k T Y k = I.
10
Theorem 1 : Given a pair of orthogonal matrices, X k and Y k, the related “Orthogonal Quotients Matrix” X k T A Y k = ( x i T Ay j ) solves the problem minimize F ( R k ) = || A - X k R k Y k T || F 2
11
Notation: X k - denotes the set of all real m x k orthogonal matrices X k, X k = [x 1, x 2, …, x k ], X k T X k = I Y k - denotes the set of all real n x k orthogonal matrices Y k, Y k = [y 1, y 2, …, y k ], Y k T Y k = I
12
Corollary 1 : The Eckart – Young Problem can be rewritten as minimize F ( X k, Y k ) = || A - X k R k Y k T || F 2 subject to X k X k and Y k Y k, where R k is the “Orthogonal Quotients Matrix” R k = X k T A Y k.
13
The Orthogonal Quotients Equality For any pair of orthogonal matrices, X k X k and Y k Y k, || A - X k R k Y k T || F 2 = || A || F 2 - || R k || F 2 where R k is the orthogonal quotients matrix R k = X k T A Y k.
14
Corollary : The Eckart – Young Problem minimize F ( X k, Y k ) = || A - X k R k Y k T || F 2 subject to X k X k and Y k Y k. is equivalent to maximize || X k T A Y k || F 2 subject to X k X k and Y k Y k, and the SVD matrices U k and V k solves both problems, giving the optimal value of 1 2 + 2 2 + … + k 2.
15
Question : Is the related minimum problem minimize || (X k ) T A Y k || F 2 subject to X k X k and Y k Y k, solvable ?
16
Using the Orthogonal Quotients Equality the last problem takes the form maximize F ( X k, Y k ) = || A - X k R k Y k T || F 2 subject to X k X k and Y k Y k.
17
Note that the more general problem maximize F ( B ) = || A - B || F 2 subject to rank ( B ) k is not solvable.
18
Returning to symmetric matrices How we extend the Orthogonal Quotients Equality to symmetric matrices ?
19
The Spectral Decomposition S = ( S ij ) a symmetric positive semi-definite n x n matrix With eigenvalues n and eigenvectors v 1, v 2, …, v n S v j = j v j, j = 1, …, n. S V = V D V = [v 1, v 2, …, v n ], V T V = V V T = I D = diag { 1, 2, …, n } S = V D V T = j v j v j T
20
Recall that Rayleigh Quotient Matrices S k = Y k T S Y k play important role in Ky Fan’s Extremum Principles.
21
Ky Fan’s Maximum Principle S a symmetric positive semi-definite n x n matrix Y k the set of orthogonal n x k matrices 1 + … + k = max { trace ( Y k T S Y k ) | Y k Y k } Solution is obtained for the Spectral matrix V k = [v 1, v 2, …, v k ]. which is related to the largest k eigenvalues.
22
Ky Fan’s Minimum Principle S a symmetric n x n matrix. Y k the set of orthogonal n x k matrices. n - k+1 + … + n = min { trace ( Y k T S Y k ) | Y k Y k } Solution is obtained for the Spectral matrix V k = [v n-k+1, …, v n ], which is related to the smallest k eigenvalues.
23
Question : Can we formulate the Symmetric Quotients Equality in terms of trace ( S k ) = trace ( Y k T S Y k ) = y j T S y j
24
The Symmetric Quotients Equality S a symmetric n x n matrix Y k Y k an orthogonal n x k matrix S k = Y k T S Y k the related “Rayleigh quotient matrix” trace ( S - Y k S k Y k T ) = trace ( S ) - trace( S k )
25
Corollary 1 : Ky Fan’s maximum problem maximize trace (Y k S Y k T ) subject to Y k Y k, is equivalent to minimize trace ( S - Y k S k Y k T ) subject to Y k Y k. The Spectral matrix V k = [v 1, v 2, …, v k ] solves both problems, giving the optimal value of 1 + … + k.
26
Recall that : The Eckart – Young Problem minimize F ( X k, Y k ) = || A - X k R k Y k T || F 2 subject to X k X k and Y k Y k. is equivalent to maximize || X k T A Y k || F 2 subject to X k X k and Y k Y k, and the SVD matrices U k and V k solves both problems, giving the optimal value of 1 2 + 2 2 + … + k 2.
27
Corollary 2 : Ky Fan’s minimum problem minimize trace (Y k S Y k T ) subject to Y k Y k, is equivalent to maximize trace ( S - Y k T S k Y k ) subject to Y k Y k. The matrix V k = [v n-k+1, …, v n ] solves both problems, giving the optimal value of n-k+1 + … + n.
28
Extended Exremum Principles Can we extend these extremums from eigenvalues of symmetric matrices to singular values of rectangular matrices ?
29
Notation: 1 m* m, 1 n* n, X m* - denotes the set of all real m x m* orthogonal matrices X m*, X m* = [x 1, x 2, …, x m* ], X m* T X m* = I Y n* - denotes the set of all real n x n* orthogonal matrices Y n*, Y n* = [y 1, y 2, …, y n* ], Y n* T Y * = I
30
Notations : Given X m* X m* and Y n* Y m*, the m* x n* matrix ( X m* ) T A Y n* = ( x i T Ay j ) is called “ Orthogonal Quotients Matrix”.
31
Notations : The singular values of the Orthogonal Quotients Matrix ( X m* ) T A Y n* = ( x i T Ay j ) are denoted as 1 2 … k 0, where k = min { m*, n* }.
32
Questions : Which choice of orthogonal matrices X m* X m* and Y n* Y m*, maximizes (or minimizes ) the sum ( 1 ) p + ( 2 ) p + … + ( k ) p where p > 0 is a given positive constant.
33
An Extended Maximum Principle : The SVD matrices U m* = [ u 1, u 2, …, u m* ] and V n* = [v 1, v 2, …, v n* ] solve the problem maximize F ( X m*, Y n* ) = ( 1 ) p + ( 2 ) p + … + ( k ) p subject to X m* X m* and Y n* Y n*, for any positive power p > 0, giving the optimal value of ( 1 ) p + ( 2 ) p + … + ( k ) p.
34
An Extended Maximum Principle That is, for any positive power p > 0, ( 1 ) p + ( 2 ) p + … + ( k ) p = max{ ( 1 ) p + ( 2 ) p +…+ ( k ) p | X m* X m* and Y n* Y n* }, and the maximal value is attained for the matrices U m* = [u 1, u 2, …, u m* ] and V n* = [v 1, v 2, …, v n * ].
35
The proof is based on “rectangular” versions of Cauchy Interlace Theorem and Poincare Separation Theorem.
36
Corollary 1 : When p = 1 the SVD matrices U m* = [ u 1, u 2, …, u m* ] and V n* = [v 1, v 2, …, v n* ] solve the “Rectangular Ky Fan problem” maximize F ( X m*, Y n* ) = 1 + 2 + … + k subject to X m* X m* and Y n* Y n*, giving the optimal value of 1 + 2 + … + k.
37
Corollary 2 : When p = 1 and m* = n* = k the SVD matrices U k = [ u 1, u 2, …, u k ] and V k = [v 1, v 2, …, v k ] solve the maximum trace problem maximize F ( X k, Y k ) = trace ( (X k ) T A Y k ) subject to X k X k and Y k Y k, giving the optimal value of 1 + 2 + … + k. * See also Horn & Johnson, “Topics in Matrix Analysis”, p. 195.
38
Corollary 3 : When p = 2 the SVD matrices U m* = [ u 1, u 2, …, u m* ] and V n* = [v 1, v 2, …, v n* ] solve the “rectangular Eckart – Young problem” maximize F ( X m*, Y n* ) = || (X m* ) T A Y n* || F 2 subject to X m* X m* and Y n* Y n*, giving the optimal value of ( 1 ) 2 + ( 2 ) 2 + … + ( k ) 2.
39
An Extended Minimum Principle Question : Can we prove a similar minimum principle ? Answer : Yes, but the solution matrices are more complicated.
40
An Extended Minimum Principle : Here we consider the problem minimize F( X m*, Y n* ) = ( 1 ) p + ( 2 ) p + … + ( k ) p subject to X m* X m* and Y n* Y n*, for any positive power p > 0. The solution matrices are obtained by deleting some columns from the SVD matrices U = [u 1, u 2, …, u m ] and V = [v 1, v 2, …, v n ].
41
Summary * The Eckart – Young Theorem and Orthogonal Quotients matrices. * The Orthogonal Quotients Equality. * The Symmetric Quotients Equality and Ky Fan’s extremum principles. * An Extended Extremum Principle.
42
The END Thank You
Similar presentations
© 2025 SlidePlayer.com. Inc.
All rights reserved.