Slide 3 - HKBU MATH
Transcription
Slide 3 - HKBU MATH
Tridiagonal matrices Gérard MEURANT October, 2008 1 Similarity 2 Cholesky factorizations 3 Eigenvalues 4 Inverse Similarity Let α1 β1 Tk = ω1 α2 .. . ω2 .. . .. . βk−2 αk−1 βk−1 ωk−1 αk and βi 6= ωi , i = 1, . . . , k − 1 Proposition Assume that the coefficients ωj , j = 1, . . . , k − 1 are different from zero and the products βj ωj are positive. Then, the matrix Tk is similar to a symmetric tridiagonal matrix. Therefore, its eigenvalues are real. Proof. Consider Dk−1 Tk Dk which is similar to Tk , Dk diagonal matrix with diag. el. δj Take βj−1 · · · β1 , j = 2, . . . k δ1 = 1, δj2 = ωj−1 · · · ω1 Let α1 β1 Jk = β1 α2 .. . β2 .. . .. . βk−2 αk−1 βk−1 βk−1 αk where the values βj , j = 1, . . . , k − 1 are assumed to be nonzero Proposition det(Jk+1 ) = αk+1 det(Jk ) − βk2 det(Jk−1 ) with initial conditions det(J1 ) = α1 , det(J2 ) = α1 α2 − β12 . The eigenvalues of Jk are the zeros of det(Jk − λI ) The zeros do not depend on the signs of the coefficients βj , j = 1, . . . , k − 1 We suppose βj > 0 and we have a Jacobi matrix Cholesky factorizations Let ∆k be a diagonal matrix with diagonal elements δj , j = 1, . . . , k and 1 l1 1 . . .. .. Lk = lk−2 1 lk−1 1 Jk = Lk ∆k LT k δ1 = α1 , δj = αj − 2 βj−1 δj−1 , j = 2, . . . , k, l1 = β1 /δ1 lj = βj /δj , j = 2, . . . , k − 1 The factorization can be completed if no δj is zero for j = 1, . . . , k − 1 This does not happen if the matrix Jk is positive definite, all the elements δj are positive and the genuine Cholesky factorization can be obtained from ∆k Jk = LCk (LCk )T 1/2 with LCk = Lk ∆k which is √ δ1 √β1 δ1 C Lk = √ δ2 .. . .. . √βk−2 δk−2 p δk−1 √βk−1 δk−1 √ δk The factorization can also be written as −1 D T Jk = LD k ∆k (Lk ) with δ1 β1 LD = k δ2 .. . .. . βk−2 δk−1 βk−1 δk Clearly, the only elements we have to compute and store are the diagonal elements δj , j = 1, . . . , k To solve a linear system Jk x = c, we successively solve LD k y = c, T (LD k ) x = ∆k y The previous factorizations proceed from top to bottom (LU) We can also proceed from bottom to top (UL) −1 Jk = L̄T k Dk L̄k with (k) d1 β1 L̄k = (k) d2 .. . .. . (k) βk−2 dk−1 (k) βk−1 dk (k) and Dk a diagonal matrix with elements dj (k) dk = αk , (k) dj = αj − βj2 (k) dj+1 , j = k − 1, . . . , 1 From LU and UL factorizations we can obtain all the so-called “twisted” factorizations of Jk Jk = Mk Ωk MkT Mk is lower bidiagonal at the top for rows with index smaller than l and upper bidiagonal at the bottom for rows with index larger than l 2 βj−1 ω1 = α1 , ωj = αj − , j = 2, . . . , l − 1 ωj−1 ωk = αk , ωj = αj − βj2 ωj+1 , j = k − 1, . . . , l + 1 2 βl−1 β2 ωl = αl − − l ωl−1 ωl+1 Eigenvalues The eigenvalues of Jk are the zeros of det(Jk − λI ) (k) (k) det(Jk − λI ) = δ1 (λ) · · · δk (λ) = d1 (λ) · · · dk (λ) This shows that δk (λ) = Theorem det(Jk − λI ) , det(Jk−1 − λI ) (k+1) (k) d1 (λ) = det(Jk − λI ) det(J2,k − λI ) The eigenvalues θi of Jk+1 strictly interlace the eigenvalues of Jk (k+1) (k) (k+1) (k) (k) (k+1) θ1 < θ1 < θ2 < θ2 < · · · < θk < θk+1 (Cauchy interlacing theorem) Proof. Eigenvector x = (y ζ)T of Jk+1 corresponding to θ Jk y + βk ζe k = θy βk yk + αk+1 ζ = θζ Eliminating y from these relations, we obtain (αk+1 − βk2 ((e k )T (Jk − θI )−1 e k ))ζ = θζ αk+1 − βk2 k X ξj2 (k) j=1 θj −θ =0 −θ where ξj is the last component of the jth eigenvector of Jk (k) The zeros of this function interlace the poles θj Inverse Theorem There exist two sequences of numbers {ui }, {vi }, i = 1, . . . , k such that u1 v1 u1 v2 u1 v3 . . . u1 vk u1 v2 u2 v2 u2 v3 . . . u2 vk −1 Jk = u1 v3 u2 v3 u3 v3 . . . u3 vk .. .. .. .. .. . . . . . u1 vk u2 vk u3 vk . . . uk vk Moreover, u1 can be chosen arbitrarily, for instance u1 = 1 see Baranger and Duc-Jacquet; Meurant Hence, we just have to compute Jk−1 e 1 and Jk−1 e k Jk v = e 1 Use UL factorization of Jk v1 = 1 (k) d1 , vj = (−1)j−1 β1 · · · βj−1 (k) (k) , j = 2, . . . , k d1 · · · dj For vk J k u = e k use LU factorization uk = 1 , δ k vk uk−j = (−1)j βk−j · · · βk−1 , j = 1, . . . , k − 1 δk−j · · · δk vk Theorem The inverse of the symmetric tridiagonal matrix Jk is characterized as (k) (k) dj+1 · · · dk −1 j−i (Jk )i,j = (−1) βi · · · βj−1 , ∀i, ∀j > i δi · · · δk (k) (Jk−1 )i,i (k) d · · · dk = i+1 , ∀i δi · · · δk Proof. (k) ui = (−1)−(i+1) (k) d1 · · · dk 1 β1 · · · βi−1 δi · · · δk The diagonal elements of Jk−1 can also be obtained using twisted factorizations Theorem Let l be a fixed index and ωj the diagonal elements of the corresponding twisted factorization Then 1 (Jk−1 )l,l = ωl In the sequel we will be interested in (Jk−1 )1,1 (Jk−1 )1,1 = 1 (k) d1 Can we compute (Jk−1 )1,1 incrementally? Theorem −1 (Jk+1 )1,1 = (Jk−1 )1,1 + Proof. Jk+1 = (β1 · · · βk )2 (δ1 · · · δk )2 δk+1 Jk βk (e k )T βk e k αk+1 −1 The upper left block of Jk+1 is the inverse of the Schur complement −1 βk2 k k T e (e ) Jk − αk+1 Inverse of a rank-1 modification of Jk Use the Sherman–Morrison formula (A + αxy T )−1 = A−1 − α A−1 xy T A−1 1 + αy T A−1 x This gives Jk − βk2 k k T e (e ) αk+1 −1 = Jk−1 + (Jk−1 e k )((e k )T Jk−1 ) αk+1 βk2 − (e k )T Jk−1 e k Let l k = Jk−1 e k −1 (Jk+1 )1,1 = (Jk−1 )1,1 + βk2 (l1k )2 αk+1 − βk2 lkk l1k = (−1)k−1 β1 · · · βk−1 , δ1 · · · δk lkk = 1 δk To simplify the formulas, we note that αk+1 − βk2 lkk = αk+1 − βk2 = δk+1 δk We start with (J1−1 )1,1 = π1 = 1/α1 and c1 = 1 t = βk2 πk , δk+1 = αk+1 − t, πk+1 = 1 , δk+1 This gives −1 (Jk+1 )1,1 = (Jk−1 )1,1 + ck+1 πk+1 ck+1 = t ck πk J. Baranger and M. Duc-Jacquet, Matrices tridiagonales symétriques et matrices factorisables, RIRO, v 3, (1971), pp 61–66 G.H. Golub and C. Van Loan, Matrix Computations, Third Edition, Johns Hopkins University Press, (1996) G. Meurant, A review of the inverse of tridiagonal and block tridiagonal matrices, SIAM J. Matrix Anal. Appl., v 13 n 3, (1992), pp 707–728