A Note on the Complexity of the PCG Algorithm for Solving Toeplitz Systems with a Fisher-Hartwig Singularity Seak-Weng VONG1∗ , Wei WANG2 , Xiao-Qing JIN3 1 Department of Mathematics, University of Macau, Macao, China. E-mail: [email protected] 2 Department of Mathematics, University of Macau, Macao, China. 3 Department of Mathematics, University of Macau, Macao, China, (The research of this author is supported by the Grant 050/2005/A from FDCT.) E-mail: [email protected] Abstract Recently, Y. Lu and C. Hurvich showed that the complexity of T. Chan’s preconditioned ¡ ¢ conjugate gradient algorithm for solving the Toeplitz system Tn (f )x = b is O n log3 n where the generating function f is given by f (ω) = |1 − e−iω |−2d h(ω) ¡ ¢ with d ∈ − 12 , 12 \{0} and h(ω) being positive continuous on [−π, π] and differentiable on [−π, π]\{0}. Although their results are interesting, there exist some improper expressions in their proofs needed to be corrected. In this paper, we try to improve those improper expressions and demonstrate these important results by some numerical tests. Key Words: PCG algorithm, Toeplitz matrix, circulant matrix, Fisher-Hartwig singularity, T. Chan’s circulant preconditioner. AMS subject classifications: 65F10, 65F15, 65L05, 65N22. 1 Introduction In this paper, we consider to use the preconditioned conjugate gradient (PCG) algorithm to solve some special Toeplitz systems. The PCG algorithm has been a popular and effective iterative method for solving Toeplitz systems Tn x = b since 1986 ([11]). It is well known that Toeplitz systems arise in a variety of applications in science and engineering [2, 6, 7, 8, 10]. The Toeplitz matrix is defined as follows: t0 t1 .. . t−1 t0 t1 Tn = tn−2 · · · tn−1 tn−2 · · · t2−n t1−n t−1 · · · t2−n .. .. . t0 . .. .. . . t−1 ··· t1 t0 (1) where the entries are given by Z π √ 1 tk = f (ω)e−ikω dω, i = −1, 2π −π for k = 0, ±1, ±2, · · ·. The function f is called the generating function of Tn . In the following, we assume that Tn is symmetric, i.e., tk = t−k . The complexity of the PCG algorithm with some suitable preconditioners, for instance, T. Chan’s circulant preconditioner and superoptimal circulant preconditioner, is known to be O (n log n) for Toeplitz systems where the generating function f (ω) is positive continuous on [−π, π], see [2, 6, 10]. Recently Y. Lu and C. Hurvich considered ([9]) the CG algorithm with T. Chan’s preconditioner for solving Tn x = b with the generating function f (ω) satisfying the following assumption: Assumption 1 ([4, 9]) The generating function f is given by f (ω) = |1 − e−iω |−2d h(ω), ω ∈ [−π, π], ¡ 1 1¢ where d ∈ − 2 , 2 and h(ω) is even positive continuous on [−π, π] and differentiable on [−π, π]\{0}. Also log h is Riemann integrable on [−π, π], i.e., ¯ Z π¯ 0 ¯ h (ω) ¯ ¯ ¯ ¯ h(ω) ¯ dω < ∞, −π and there exists c ∈ (0, ∞) such that |h0 (ω)| ≤ c|ω|−1 , ω ∈ [−π, π]\{0}. Moreover, h(ω) is assumed to be in B11 1 h(ω) . T L∞ and all conditions for h(ω) are also applied to We recall that the Besov space B11 includes all functions g such that g ∈ L1 and Z π Z 1 π |g(ω + t) − 2g(ω) + g(ω − t)|dωdt < ∞. 2 −π t −π Now let us give the definition of T. Chan’s circulant preconditioner proposed in [3]: Definition 1 ([2, 3]) For Tn given as in (1), the diagonals of T. Chan’s circulant preconditioner cF (Tn ) are given by (n − k)tk + ktk−n , if n ck = cn+k , if 0 ≤ k < n, 0 < −k < n. Under Assumption 1, Tn has a Fisher-Hartwig singularity caused by the term |1−e−iω |−2d in the generating function f and the condition number of Tn approaches ∞ as n increases, see [4]. Y. Lu and C. Hurvich showed in ([9]) that the complexity of the CG algorithm for solving Tn x = b without any preconditioning grows asymptotically as n1+|d| log n. With T. Chan’s optimal circulant preconditioner cF (Tn ), the complexity of the PCG algorithm ¡ ¢ is O n log3 n . But there exist some improper expressions in their proofs needed to be corrected or revised. In this paper, we try to improve their proofs and demonstrate these important results by some numerical examples. 2 Spectral analysis We first introduce the following lemma and the proof can be found in [9]. Lemma 1 ([9]) If f satisfies Assumption 1, we then have h i −1/2 −1/2 λmax cF (Tn (f )) Tn (f )cF (Tn (f )) = O(log2 n), ¡ ¢¤ £ −1 λmin c−1 Tn (f −1 ) ≥ C1 , F (Tn (f )) cF h ¢i ¢ 1/2 ¡ 1/2 ¡ λmin cF Tn (f −1 ) Tn−1 (f −1 )cF Tn (f −1 ) ≥ C2 · log−2 n, where Ci , i = 1, 2, represent positive constants independent of n, λmin (M ) and λmax (M ) denote the smallest and the largest eigenvalues of M respectively. Lemma 2 ([9]) If f satisfies Assumption 1, then Tn (f ) − Tn−1 (f −1 ) is positive semidefinite. Proof: Let ³ ´T 1 un (ω) = √ f −1/2 (ω) 1, eiω , · · · , ei(n−1)ω , 2π and ³ ´T 1 vn (ω) = √ f 1/2 (ω) 1, eiω , · · · , ei(n−1)ω . 2π Then, Z π un (ω)u∗n (ω)dω = Tn (f −1 ), Z −π and Z π −π π −π vn (ω)vn∗ (ω)dω = Tn (f ) un (ω)vn∗ (ω)dω = In . Note that for any α(ω) = (α1 (ω), α2 (ω), · · · , αn (ω))T , the matrix Z B = α(ω)α∗ (ω)dω is positive semidefinite since Z ∗ x Bx = |x∗ α|2 dω ≥ 0, for any x ∈ Cn . Therefore, for any n × n matrix A, we have Z π B= [Aun (ω) + vn (ω)][Aun (ω) + vn (ω)]∗ dω = ATn (f −1 )A∗ + A + A∗ + T (f ). (2) −π In particular, if we take A = −T −1 (f −1 ) in (2), we can conclude that Tn (f ) − Tn−1 (f −1 ) is positive semidefinite. Theorem 1 ([9]) If f satisfies Assumption 1, then we have h i −1/2 −1/2 λmin cF (Tn (f )) Tn (f )cF (Tn (f )) ≥ C · log−2 n, where C is a positive constant independent of n. −1/2 Proof: We can factorize cF −1/2 cF −1/2 (Tn (f )) Tn (f )cF −1/2 (Tn (f )) = B1 B2 B3 + B4 , −1/2 (Tn (f −1 )), (Tn (f )) Tn (f )cF (Tn (f )) as where −1/2 B1 = cF (Tn (f ))cF 1/2 1/2 B2 = cF (Tn (f −1 ))Tn−1 (f −1 )cF (Tn (f −1 )), −1/2 (Tn (f −1 ))cF −1/2 (Tn (f ))[Tn (f ) − Tn−1 (f −1 )]cF B3 = cF B4 = cF −1/2 (Tn (f )), −1/2 (Tn (f )). (3) By using Lemmas 1 and 2, we notice that B1 , B2 , B3 are all Hermitian positive definite and B4 is Hermitian positive semidefinite. Therefore, we have n o −1/2 −1/2 λmin cF (Tn (f ))Tn (f )cF (Tn (f )) ≥ λmin (B1 B2 B3 ) + λmin (B4 ) ≥ λmin (B1 )λmin (B2 )λmin (B3 ) + λmin (B4 ) ≥ C · log−2 n + λmin (B4 ) ≥ C · log−2 n. Thus, the condition number of T. Chan’s preconditioned matrix satisfies ¢ ¡ κ c−1 F (Tn (f ))Tn (f ) h i −1/2 −1/2 = κ cF (Tn (f ))Tn (f )cF (Tn (f )) µ = O log2 n C · log−2 n ¶ = O(log4 n). Therefore, the number of iterations in the PCG algorithm is bounded by O(log2 n), see [5, 7]. Since each iteration needs O(n log n) operations using the FFT, the total complexity of the PCG algorithm is still O(n log3 n). 3 Numerical tests In this section, we verify Theorem 1 by the following problem. Problem 1 ([1]) The generating function is given by ³ ω ´¯−2d σ 2 ¯¯ ¯ f (ω) = s ¯2 sin −π ≤ ω ≤ π. ¯ , 2π 2 Then we obtain tk = σs2 Γ(1 − 2d)Γ(k + d) , Γ(d)Γ(1 − d)Γ(k − d + 1) k = 0, 1, 2, · · · . Also, we choose b = (t1 , t2 , · · · , tn )T as the right-hand side of Tn x = b. 250 30 d=0.37 d=0.1 25 Number of iterations: y=f(x) Number of iterations: y=f(x) 200 T. Chan Preconditioner 150 No Preconditioner 100 50 0 T. Chan Preconditioner 20 No Preconditioner 15 10 5 0 5 10 0 15 0 x 5 10 15 The number of matrix size (n=2x) The number of matrix size (n=2 ) Figure 1: Two examples of iteration numbers. All the experiments were performed in MATLAB. We used the MATLAB-provided M-file “pcg” to solve the system. In our tests, the zero vector is the initial guess and the stopping criterion is krj k2 /kr0 k2 < 10−10 where rj is the residual after the j-th iteration. We consider the following two cases in Problem 1: Case 1: Choose d = 0.37 and σs2 = 0.27. Case 2: Choose d = 0.1 and σs2 = 0.27. In Table 1, we give the number of iterations for convergence. In the table, n is the matrix size. We compare the preconditioned system by using T. Chan’s preconditoner with the system with no preconditioner. We note that when n increases, the number of iterations of the preconditioned system is much less than that of the system with no preconditioner. n Case1 No preconditioner Case1 T. Chan’s preconditioner Case2 No preconditioner Case2 T. Chan’s preconditioner 25 26 27 28 29 210 211 212 213 214 215 18 25 32 40 52 65 81 103 130 168 214 7 7 8 8 8 9 9 10 10 10 12 11 13 14 16 17 19 20 22 23 25 28 6 6 7 7 7 7 7 8 8 8 9 Table 1: Number of iterations for PCG algorithm. Acknowledgments. The authors would like to thank Prof. Clifford M. Hurvich for his useful suggestions on this topic. References [1] P. Brockwell and R. Davis, Time Series: Theory and Methods, Springer-Verlag, New York, 1991. [2] R. Chan and M. Ng, Conjugate Gradient Methods for Toeplitz Systems, SIAM Review, Vol. 38 (1996), pp. 427–482. [3] T. Chan, An Optimal Circulant Preconditioner for Toeplitz Systems, SIAM J. Sci. Statist. Comput., Vol. 9 (1988), pp. 766–771. [4] M. Fisher and R. Hartwig, Toeplitz Determinants–Some Applications, Theorems, and Conjectures, Adv. Chem. Phys., Vol. 32 (1968), pp. 190–225. [5] G. Golub and C. Loan, Matrix Computations, 3rd edition, Johns Hopkins University Press, Baltimore, 1996. [6] X. Jin, Developments and Applications of Block Toeplitz Iterative Solvers, Kluwer Academic Publishers, Dordrecht; and Science Press, Beijing, 2002. [7] X. Jin and Y. Wei, Numerical Linear Algebra and Its Applications, Science Press, Beijing, 2004. [8] R. King, M. Ahmadi, R. Gorgui-Naguib, A. Kwabwe and M. Azimi-Sadjadi, Digital Filtering in One and Two Dimensions: Design and Applications, Plenum, New York, 1989. [9] Y. Lu and C. Hurvich, On the complexity of the Preconditioned Conjugate Gradient Algorithm for Solving Toeplitz Systems with a Fisher–Hartwig Singularity, SIAM J. Matrix Anal. Appl., Vol. 27 (2005), pp. 638–653. [10] M. Ng, Iterative Methods for Toeplitz Systems, Oxford University Press, Oxford, 2004. [11] G. Strang, A Proposal for Toeplitz Matrix Calculations, Stud. Appl. Math., Vol. 74 (1986), pp. 171–176.
© Copyright 2026 Paperzz