Optimal Parameter in Hermitian and Skew-Hermitian Splitting Method for Certain Two-by-Two Block Matrices

Transcript

1 Optimal Parameter in Hermitian Skew-Hermitian Splitting Method for Certain Two-by-Two Block Matrices Zhong-Zhi Bai 1 Department of Mathematics, Fudan University Shanghai , P.R. China State Key Laboratory of Scientific/Engineering Computing Institute of Computational Mathematics Scientific/Engineering Computing Academy of Mathematics Systems Science Chinese Academy of Sciences, P.O. Box 2719, Beijing , P.R. China 2 bzz@lsec.cc.ac.cn Gene H. Golub 3 Scientific Computing Computational Mathematics Program Department of Computer Science Stanford University, Stanford, CA , USA golub@sccm.stanford.edu Chi-Kwong Li 4 Department of Mathematics, The College of William & Mary P.O. Box 8795, Williamsburg, VA , USA ckli@math.wm.edu October 30, Supported by The National Basic Research Program (No. 2005CB321702) The National Natural Science Foundation (No ), P.R. China; The 2004 Ky Yu-Fen Fan Fund Travel Grant of American Mathematical Society 2 Correspondence permanent address 3 The work of this author was in part supported by the Department of Energy: DE-FC02-01ER The research of Li was partially supported by an NSF grant. He is an honorary professor of the Heilongjiang University, an honorary professor of the University of Hong Kong

2 Abstract The optimal parameter of the Hermitian/skew-Hermitian splitting (HSS) iteration method for a real 2-by-2 linear system is obtained. The result is used to determine the optimal parameters for linear systems associated with certain 2-by-2 block matrices, to estimate the optimal parameters of the HSS iteration method for linear systems with n-by-n real coefficient matrices. Numerical examples are given to illustrate the results. Keywords: Non-Hermitian matrix, Hermitian matrix, skew-hermitian matrix, splitting iteration method, optimal iteration parameter. AMS(MOS) Subject Classifications: 65F10, 65F50; CR: G1.3.

3 2 Z.-Z. Bai, G.H. Golub C.-K. Li 1 Introduction To solve the large sparse non-hermitian positive definite system of linear equations Ax = f, A C n n positive definite, A A, x, f C n, (1.1) Bai, Golub Ng[2] recently proposed the Hermitian/skew-Hermitian splitting (HSS) iteration method based on the fact that the coefficient matrix A naturally possesses the Hermitian/skew-Hermitian (HS) splitting[10, 19] where A = H + S, H = 1 2 (A + A ) S = 1 2 (A A ), with A being the conjugate transpose of the matrix A. They showed that this HSS iteration converges unconditionally to the exact solution of the system of linear equations (1.1), with the upper bound on convergence speed about the same as that of the conjugate gradient method when applied to Hermitian matrices. Moreover, the upper bound of the contraction factor is dependent on the spectrum of the Hermitian part H, but is independent of the spectrum of the skew-hermitian part S as well as the eigenvalues of the matrices H, S A. Numerical experiments have shown that the HSS iteration method is very efficient robust both as a solver as a preconditioner (to Krylov subspace methods such as GMRES BiCGSTAB; see [15, 18]) for solving non-hermitian positive definite linear systems. To further improve the efficiency of the method, it is desirable to determine or find a good estimate for the optimal parameter α. Unfortunately, there is no good method in doing that. In this paper, we analyze 2-by-2 real matrices in detail, obtain the optimal parameter α that minimizes the spectral radius of the iteration matrix of the corresponding HSS method. We then use the results to determine the optimal parameters for linear systems associated with certain 2-by-2 block matrices, to estimate the optimal parameter α of the HSS method for general n-by-n nonsymmetric positive definite system of linear equations (1.1). Numerical examples are given to show that our estimations improve previous results are close to the values of the optimal parameters. Unless specified otherwise, we assume throughout the paper that the non-hermitian matrix A C n n is positive definite, i.e., A A its Hermitian part H = 1 2 (A + A ) is Hermitian positive definite. 2 The HSS Iteration Let us first review the HSS iteration method presented in Bai, Golub Ng[2]. The HSS Iteration Method. Given an initial guess x (0) C n, compute x (k) for k = 0, 1, 2,... using the following iteration scheme until {x (k) } satisfies the stopping criterion: { (αi + H)x (k+ 1 2 ) = (αi S)x (k) + f, (αi + S)x (k+1) = (αi H)x (k+ 1 2 ) + f,

4 The Optimal Parameter in HSS Method 3 where α is a given positive constant. In matrix-vector form, the above HSS iteration method can be equivalently rewritten as x (k+1) = M(α)x (k) + G(α)f, k = 0, 1, 2,..., (2.1) where M(α) = (αi + S) 1 (αi H)(αI + H) 1 (αi S) (2.2) G(α) = 2α(αI + S) 1 (αi + H) 1. Here, M(α) is the iteration matrix of the HSS method. In fact, (2.1) may also result from the splitting A = B(α) C(α) of the coefficient matrix A, with { B(α) = 1 2α (αi + H)(αI + S), C(α) = 1 2α (αi H)(αI S). The following theorem established in [2] describes the convergence property of the HSS iteration. Theorem 2.1. Let A C n n be a positive definite matrix, H = 1 2 (A + A ) S = 1 2 (A A ) be its Hermitian skew-hermitian parts, respectively, α be a positive constant. Then the spectral radius ρ(m(α)) of the iteration matrix M(α) of the HSS iteration (see (2.2)) is bounded by σ(α) = α λ j max λ j λ(h) α + λ j, where λ( ) represents the spectrum of the corresponding matrix. Consequently, we have ρ(m(α)) σ(α) < 1, α > 0, i.e., the HSS iteration converges to the exact solution x C n of the system of linear equations (1.1). Moreover, if γ min γ max are the lower the upper bounds of the eigenvalues of the matrix H, respectively, then { } α arg min max α λ α γ min λ γ max α + λ = γ min γ max σ( α) = γmax γ min κ(h) 1 γmax + =, γ min κ(h) + 1 where κ(h) is the spectral condition number of H.

5 4 Z.-Z. Bai, G.H. Golub C.-K. Li Of course, α is usually different from the optimal parameter it always holds that α = arg min ρ(m(α)), α ρ(m(α )) σ( α). Numerical experiments in [2] have confirmed that in most situation, ρ(m(α )) σ( α). See [3, 12, 6, 16, 1, 13, 8, 14, 11, 7] for further applications generalizations of the HSS iteration method. 3 The Real Two-By-Two Case In this section, we study linear systems associated with a real 2-by-2 matrix A with positive definite symmetric part. We first determine the eigenvalues of M(α) defined in (2.2). The following theorem is stated in general terms so that it can be used more conveniently for future discussion. Theorem 3.1. Let A = H +S R 2 2 be such that H is symmetric positive definite S is skew-symmetric. Suppose H has eigenvalues λ 1 λ 2 > 0 det(s) = q 2 with q R. Then the two eigenvalues of the iteration matrix M(α) defined in (2.2) are λ ± = (α2 λ 1 λ 2 )(α 2 q 2 ) ± (α 2 λ 1 λ 2 ) 2 (α 2 q 2 ) 2 (α 2 λ 2 1 )(α2 λ 2 2 )(α2 + q 2 ) 2 (α + λ 1 )(α + λ 2 )(α 2 + q 2. ) As a result, if then ρ(m(α)) equals to if (α 2 λ 1 λ 2 ) 2 (α 2 q 2 ) 2 (α 2 λ 2 1)(α 2 λ 2 2)(α 2 + q 2 ) 2, α 2 λ 1 λ 2 α 2 q 2 + (α 2 λ 1 λ 2 ) 2 (α 2 q 2 ) 2 (α 2 λ 2 1 )(α2 λ 2 2 )(α2 + q 2 ) 2 (α + λ 1 )(α + λ 2 )(α 2 + q 2 ; ) then ρ(m(α)) equals to (α 2 λ 1 λ 2 ) 2 (α 2 q 2 ) 2 < (α 2 λ 2 1)(α 2 λ 2 2)(α 2 + q 2 ) 2, (α λ 1 )(α λ 2 ) (α + λ 1 )(α + λ 2 ). Proof. Let A = H + S R 2 2, where H is symmetric positive definite, S is skewsymmetric. Then there is an orthogonal matrix Q R 2 2 such that Q t HQ is a diagonal matrix with diagonal entries λ 1 λ 2, where Q t denotes the transpose matrix of Q. We may replace A by Q t AQ without changing the assumptions conclusions of our theorem. So, assume that [ ] [ ] λ1 0 0 q H = S = with q R. 0 λ 2 q 0

6 The Optimal Parameter in HSS Method 5 Then (αi + H) 1 (αi H)(αI + S) 1 (αi S) equals 1 α 2 + q 2 [ (α 2 q 2 )(α λ 1 ) α+λ 1 2qα(α λ 1) α+λ 1 2qα(α λ 2 ) (α 2 q 2 )(α λ 2 ) α+λ 2 α+λ 2 ]. The formula for λ ± the assertion on ρ(m(α)) follow. One may want to use the formula of ρ(m(α)) in Theorem 3.1 to determine the optimal choice of α. It turns out that the analysis is very complicated not productive. The main difficulty is the expression (α 2 λ 1 λ 2 ) 2 (α 2 q 2 ) 2 (α 2 λ 2 1 )(α2 λ 2 2 )(α2 + q 2 ) 2 (3.1) in the formula of ρ(m(α)). For example, one may see [5] for the analysis of a similar simpler problem. Here, we use a different approach that allows us to avoid the complicated expression (3.1). For notational simplicity, we write Define Evidently, Moreover, Thus, { } trace (M(α)) 2 φ(α) = = 2 ρ(α) = ρ(m(α)). { (α 2 q 2 )(α 2 } 2 λ 1 λ 2 ) (α 2 + q 2, )(α + λ 1 )(α + λ 2 ) ψ(α) = det (M(α)) = (α λ 1)(α λ 2 ) (α + λ 1 )(α + λ 2 ), ω(α) = max{φ(α), ψ(α) }. ρ(α) 2 ω(α). 1 = φ(0) = lim φ(α) 1 = ψ(0) = lim ψ(α). α + α + lim ω(α) = ω(0) = 1 > ω(ξ) for all ξ > 0. α + Since ω(α) is continuous nonnegative, there exists α > 0 such that We will show that ω(α ) = min{ω(α) α > 0}. φ(α ) = ψ(α ). (3.2) As a result, the eigenvalues of M(α ) have the same modulus, thus ρ(α) 2 ω(α) ω(α ) = ρ(α ) 2, for all α > 0. By the above discussion, establishing (3.2) will lead to the following theorem.

7 6 Z.-Z. Bai, G.H. Golub C.-K. Li Theorem 3.2. Let the assumptions of Theorem 3.1 be satisfied define the functions φ ψ as above. Then the optimal α > 0 satisfying ρ(m(α )) = min{ρ(m(α)) α > 0} lies in the finite set S = {α > 0 φ(α) = ψ(α) }, which consists of numbers α > 0 satisfying (α 2 + q 2 ) 2 (α 2 λ 2 1)(α 2 λ 2 2) = (α 2 q 2 ) 2 (α 2 λ 1 λ 2 ) 2 (3.3) or (α 2 + q 2 ) 2 (λ 2 1 α 2 )(α 2 λ 2 2) = (α 2 q 2 ) 2 (α 2 λ 1 λ 2 ) 2. (3.4) Proof. We only need to demonstrate the validity of (3.2), i.e., φ(α ) = ψ(α ). Note that φ(α) is continuously differentiable for α > 0, ψ(α) is continuous for α > 0 differentiable except for α = λ 1 λ 2. Since ω(α) = max{φ(α), ψ(α) }, if α satisfies ω(α ) = min{ω(α) α > 0}, then one of the following holds. (i) φ(α ) = ψ(α ) ; (ii) φ(α ) > ψ(α ) φ(α ) is a local minimum of φ(α); (iii) ψ(α ) > φ(α ) ψ(α ) is a local minimum of ψ(α). First, we claim that (iii) cannot happen. To see this, note that ψ (α) = 2(λ 1 + λ 2 )(α 2 λ 1 λ 2 ) [(α + λ 1 )(α + λ 2 )] 2, which is positive or negative according to α > λ 1 λ 2 or α < λ 1 λ 2, respectively, ψ ( λ 1 λ 2 ) = 0. Because λ1 λ 1 λ 2 λ 2, we see that ψ(α) = ψ(α) is differentiable decreasing on (0, λ 2 ), ψ(α) = ψ(α) is differentiable increasing on (λ 1, + ). Thus, there cannot be α in (0, λ 2 ) (λ 1, + ) satisfying (iii). Furthermore, ψ(λ j ) = 0 for j = 1, 2; so, it is impossible to have α = λ j satisfying (iii). Finally, ψ(α) = ψ(α) is differentiable on (λ 2, λ 1 ) with a local maximum at λ 1 λ 2. Thus, there cannot be α in (λ 2, λ 1 ) satisfying (iii). Next, we show that (ii) cannot happen. The analysis is more involved. Instead of φ(α), we consider its square root F (α) = (α2 q 2 )(α 2 ) ( λ 1 λ 2 ) (1 (α 2 + q 2 )(α + λ 1 )(α + λ 2 ) = 2q2 α 2 + q 2 1 λ 1 λ ) 2. α + λ 1 α + λ 2

8 The Optimal Parameter in HSS Method 7 Then where F (α) = = 4q 2 α (α 2 λ 1 λ 2 ) (α 2 + q 2 ) 2 (α + λ 1 )(α + λ 2 ) + (α2 q 2 ) (α 2 + q 2 ) λ 1 + λ 2 (α 2 + q 2 ) 2 (α + λ 1 ) 2 (α + λ 2 ) 2 P (α), P (α) = α 6 + 4(λ 1λ 2 + q 2 ) λ 1 + λ 2 α 5 + (λ 1 λ 2 + 4q 2 )α 4 ( ) λ 1 (α + λ 1 ) 2 + λ 2 (α + λ 2 ) 2 q 2 (q 2 + 4λ 1 λ 2 )α 2 4q2 λ 1 λ 2 (q 2 + λ 1 λ 2 ) λ 1 + λ 2 α λ 1 λ 2 q 4. Suppose q λ 1 λ 2. Then F (α) < 0 for α (0, q), F (α) > 0 for α ( λ 1 λ 2, + ). So, F (α) is decreasing on (0, q) increasing on [ λ 1 λ 2, + ). Hence, φ(α) = F (α) 2 cannot have a local minimum in these two intervals. Thus, there cannot be an α in these intervals satisfying (ii). Because φ(q) = φ ( ) λ 1 λ 2 = 0, we cannot have α = q or λ 1 λ 2 satisfying (ii). Next, we claim that F (α) has only one critical point in ( q, λ 1 λ 2 ), which is a local minimum. This point will be the unique critical point for φ(α) = F (α) 2 on ( q, λ 1 λ 2 ), which is a local maximum. Thus, there cannot be an α in this interval satisfying (ii). To prove our claim, note that α > 0 is a critical point of F (α) if only if α is a zero of P (α). So, it suffices to show that P (α) only has one positive zero. Now, P (α) = 30α (λ 1λ 2 + q 2 ) λ 1 + λ 2 α (λ 1 λ 2 + 4q 2 )α 2 2q 2 (q 2 + 4λ 1 λ 2 ). Since P (0) < 0 lim α + P (α) = +, we see that P (α) has a positive zero. However, P (α) cannot have two or more positive zeros (counting multiplicity). Otherwise, P (α) = 120α (λ 1λ 2 + q 2 ) λ 1 + λ 2 α (λ 1 λ 2 + 4q 2 )α has a positive zero, which is impossible. Next, we argue that P (α) = 6α (λ 1λ 2 + q 2 ) λ 1 + λ 2 α 4 + 4(λ 1 λ 2 + 4q 2 )α 3 2q 2 (q 2 + 4λ 1 λ 2 )α 4q2 λ 1 λ 2 (q 2 + λ 1 λ 2 ) λ 1 + λ 2 has exactly one positive zero. Since P (0) < 0 lim α + P (α) = +, we see that P (α) has a positive zero. Since P (0) < 0 P (α) only has one positive zero α 1, we see that P (α) is decreasing on (0, α 1 ) increasing on (α 1, + ). So, P (α) can only have one positive zero α 2 > α 1. We can apply the same argument to P (α). Since P (0) < 0 lim α + P (α) = +, we see that P (α) has a positive zero. Since P (0) < 0 P (α) only has one positive zero α 2, we see that P (α) is decreasing on (0, α 2 ) increasing on (α 2, + ). So, P (α) can only have one positive zero α 3 > α 2, our claim is proved. One can use a similar argument to get the desired conclusion if q > λ 1 λ 2. We omit the discussion.

9 8 Z.-Z. Bai, G.H. Golub C.-K. Li Remark 3.3. Using the absolute value function, we can combine (3.3) (3.4) to a single equation (α 2 + q 2 ) 2 (λ 2 1 α 2 )(α 2 λ 2 2) = [(α 2 q 2 )(α 2 λ 1 λ 2 )] 2. Nonetheless, the polynomial equations are easier to solve use. In fact, if λ 1 = λ 2 = λ, then (3.3) (3.4) only has one positive solution, namely, α = λ. Suppose λ 1 > λ 2. If we use the substitution β = α 2, then (3.3) reduces to the quadratic equation which has solutions [(λ 1 λ 2 ) 2 4q 2 ]β 2 + 2q 2 (λ 1 + λ 2 ) 2 β + q 2 [q 2 (λ 1 λ 2 ) 2 4λ 2 1λ 2 2] = 0, q(2λ 1 λ 2 qλ 1 + qλ 2 ) λ 1 λ 2 + 2q q(2λ 1 λ 2 + qλ 1 qλ 2 ) λ 1 λ 2 2q if λ 1 λ 2 2q. Otherwise, the equation is linear has a solution β = 2(q4 λ 2 1 λ2 2 ) (λ 1 + λ 2 ) 2 = 2(q2 + λ 1 λ 2 )(q 2 λ 1 λ 2 ) (λ 1 + λ 2 ) 2 = (λ 1 + λ 2 ) 2 (q 2 λ 1 λ 2 ) 2(λ 1 + λ 2 ) 2 = 1 2 (q2 λ 1 λ 2 ). Of course, these solutions will be useful only if they are positive lie outside the interval [λ 2 2, λ2 1 ]. Similarly, by the substitution β = α2, (3.4) reduces to the quartic equation 2β 4 (λ 1 + λ 2 ) 2 β 3 + 2[λ 2 1λ 2 2 q 2 (λ 1 λ 2 ) 2 + q 4 ]β 2 q 4 (λ 1 + λ 2 ) 2 β + 2q 4 λ 2 1λ 2 2 = 0, (3.5) which has exactly two solutions µ 1 µ 2 with µ 1 [λ 2 2, λ 1λ 2 ) µ 2 ((λ λ2 2 )/2, λ2 1 ]. Furthermore, if q 2 [λ 2 2, λ2 1 ] then µ 1 q 2 µ 2 q 2. In particular, if q = λ i for i {1, 2} then β = λ 2 i is a solution. The verification of the above statements will be given in the last section. As an illustration of Theorems as well as Remark 3.3, we consider a simple example for which λ 1 = 2, λ 2 = 1 q = 1. By straightforward computations, we know that the only positive roots of the equation (3.3) are α1 = 1 α 2 = 5, those of the equation (3.4) are α3 = 1 6 (7 α4 = ) We remark that now equation (3.4) is equivalent to (β 1)(2β 3 7β 2 + β 8) = 0. Based on Theorem 3.1, from direct calculations we have 3(7 3 5) ρ(m(α1)) = ρ(m(α3)) = 0, ρ(m(α2)) = ρ(m(α4)) Therefore, for the HSS iteration method, the optimal parameter is α = 1 the corresponding optimal convergence factor is ρ(m(α )) = 0. On the other h, from Theorem 2.1 we can easily obtain α = σ( α) = Obviously, it holds that ρ(m(α )) < σ( α).

10 The Optimal Parameter in HSS Method 9 Remark 3.4. Note that in our proof of Theorem 3.2, we get much information for the function ω(α) = max{φ(α), ψ(α) } with α > 0. In particular, we show that all the local minima of ω(α) satisfy φ(α) = ψ(α), they are the roots of (3.3) (3.4). Moreover, the local minima of ω(α) are also local minima of ρ(α), that the global minimum of ρ(α) ω(α) occur at the same α. If one can prove independently that the local minima ρ(α) always occur when the eigenvalues of M(α) have the same magnitude, then one can conclude that the functions ω(α) ρ(α) have the same local global minima. Once again, this is difficult to do because of the expression (3.1) in ρ(α) = ρ(m(α)). Using the notations defined in Theorem 3.1, we immediately get the following conclusion from Theorem 3.2. Corollary 3.5. Let the assumptions of Theorem 3.1 be satisfied. Then the optimal α > 0 satisfying ρ(m(α )) = min{ρ(m(α)) α > 0} is a positive root of the equation or (α 2 + q 2 ) 2 (α 2 λ 2 1)(α 2 λ 2 2) = (α 2 q 2 ) 2 (α 2 λ 1 λ 2 ) 2 (α 2 + q 2 ) 2 (λ 2 1 α 2 )(α 2 λ 2 2) = (α 2 q 2 ) 2 (α 2 λ 1 λ 2 ) 2. Remark 3.6. The first equation in Corollary 3.5 can be reduced to a quadratic equation the second one can be reduced to a quartic equation with respect to β = α 2, respectively, analogously to those in Remark 3.3. Remark 3.7. Note that the results in this section are also valid if λ 1 > 0 = λ 2 q 0. Remark 3.8. We should point out again that Theorem 3.2 has been established only for real matrices. For complex matrices, how to determine the optimal iteration parameter of the HSS iteration method is still an open problem. 4 Two-By-Two Block Matrices In this section, we determine the optimal parameter α for a 2-by-2 block matrix of the form [ ] λ1 I A = r E E, (4.1) λ 2 I s where λ 1 > λ 2 > 0. Note that the case when λ 1 = λ 2 is trivial 1. Systems of linear equations with the 2-by-2 block matrices (4.1) arise in many applications; for details we 1 If λ 1 = λ 2 λ, then the iteration matrix of the HSS iteration method is given by M(α) = α λ Q(α), [ α+λ ] where Q(α) := (αi +S) 1 0 E (αi S) is a Cayley transform of S is, thus, unitary, S = E ; 0 see (2.2). Obviously, it holds that ρ(m(α)) = α λ. Therefore, for the HSS iteration method applied to α+λ this special linear system, the optimal parameter is α = λ the corresponding optimal convergence factor is ρ(m(α )) = 0.

11 10 Z.-Z. Bai, G.H. Golub C.-K. Li refer the readers to [17, Chapter 6], [9, Chapters 4, 5 10], [20, 10, 4, 5] references therein. According to Young[20] Varga[17], the matrix A in (4.1) is called a 2-cyclic matrix, is connected with property A. Theorem 4.1. Suppose that the matrix A C n n defined in (4.1) satisfies λ 1 > λ 2 > 0, the nonzero matrix E C r s has nonzero singular values q 1 q 2 q k. Let H = 1 [ ] 2 (A + λ1 I A ) = r 0 S = 1 [ ] 0 λ 2 I s 2 (A 0 E A ) = E 0 be the Hermitian the skew-hermitian parts of the matrix A, respectively. Then, for the correspondingly induced HSS iteration method, the spectral radius ρ(m(α)) of its iteration matrix M(α) (see (2.2)) attains the minimum at α, which is a root of one of the following equations: (α λ 1 λ 2 )(α q 1 q k ) = 0, or where j = 1, k. (α 2 + q 2 j ) 2 (α 2 λ 2 1)(α 2 λ 2 2) = (α 2 q 2 j ) 2 (α 2 λ 1 λ 2 ) 2 (α 2 + q 2 j ) 2 (λ 2 1 α 2 )(α 2 λ 2 2) = (α 2 q 2 j ) 2 (α 2 λ 1 λ 2 ) 2, Proof. Suppose A satisfies the hypotheses of the theorem. Then A is unitarily [ similar ] λ1 q to A 1 A k λ 1 I u λ 2 I v, where denotes the matrix direct sum, A j = j, q j λ 2 u = r k v = s k. For α > 0, define ρ(α) = ρ(m(α)) ρ j (α) = ρ(m j (α)), where M(α) is unitarily similar to M 1 (α) M k (α) α λ 1 α+λ 1 I u α λ 2 α+λ 2 I v, with M j (α) being the HSS iteration matrix associated with A j for j = 1, 2,..., k. Furthermore, define Consider four cases. f j (α) = α λ j / α + λ j for j = 1, 2. Case 1. If r > k s > k, then f 1 (α) f 2 (α) are eigenvalues of M(α) is the largest singular value of M(α). Thus, ρ(α) = max{ f 1 (α), f 2 (α) } ( ) ρ(α ) = ρ λ1 λ 2 = λ1 λ 2 λ1 + λ 2. Case 2. Suppose r > k s = k. Let ρ 0 (α) = f 1 (α). Then ρ(α) = max{ρ j (α) 0 j k}. For α ( 0, λ 1 λ 2 ], ρ(α) = ρ 0 (α) = f 1 (α)

12 The Optimal Parameter in HSS Method 11 is the largest singular value of M(α). If α (λ 3 1 /λ 2) 1/2, then f 1 (α) = α λ 1 α + λ 1 = 1 2λ 1 α + λ 1 2λ 1 ( ) 1 = ρ λ1 λ 2. λ1 λ 2 + λ 1 Thus, ρ(α) ρ ( ) [ λ 1 λ 2 if α 0, (λ1 λ 2 ) 1/2] [ (λ 3 1 /λ 2) 1/2, + ). So, we focus on the interval ( J = (λ 1 λ 2 ) 1/2, (λ 3 1/λ 2 ) 1/2). For α J, let { } { trace (Mj (α)) 2 (α 2 qj 2 φ j (α) = = )(α2 λ 1 λ 2 ) 2 (α 2 + qj 2)(α + λ 1)(α + λ 2 ) ψ(α) = det (M j (α)) = (α λ 1)(α λ 2 ), j = 1, 2,..., k. (α + λ 1 )(α + λ 2 ) Clearly, ψ(α) is independent of the index j. By Theorem 3.1, if φ j (α) ψ(α), then ρ j (α) = φ j (α) + φ j (α) ψ(α); } 2 otherwise, We claim that ρ j (α) = (α λ 1 )(α λ 2 ) (α + λ 1 )(α + λ 2 ). max{ρ j (α) : 1 j k} = max{ρ 1 (α), ρ k (α)}, α J. Suppose 1 < j < k. If φ j (α) < ψ(α) then (α λ 1 )(α λ 2 ) ρ j (α) = (α + λ 1 )(α + λ 2 ), min{φ 1 (α), φ k (α)} φ j (α) < ψ(α). It follows that ρ 1 (α) = ρ j (α) or ρ k (α) = ρ j (α). If φ j (α) ψ(α) α q 2 j, then hence ρ j (α) = 0 φ j (α) φ k (α), φ j (α) + φ j (α) ψ(α) φ k (α) + φ k (α) ψ(α) = ρ k (α); if φ j (α) ψ(α) α < q 2 j, then hence ρ j (α) = 0 < φ j (α) φ 1 (α) φ j (α) + φ j (α) ψ(α) φ 1 (α) + φ 1 (α) ψ(α) = ρ 1 (α).

13 12 Z.-Z. Bai, G.H. Golub C.-K. Li Thus, our claim is proved. For each α J, we have α λ1 / α + λ 1 α λ 2 / α + λ 2, hence ρ 0 (α) = α λ 1 α + λ 1 (α λ 1 )(α λ 2 ) (α + λ 1 )(α + λ 2 ) max{ρ 1(α), ρ k (α)}. Combining the above, we see that Let Then ρ(α) = max{ρ 1 (α), ρ k (α)}, α J. Ω(α) = max{φ 1 (α), φ k (α), ψ(α) }. Ω(α) max{ρ 1 (α) 2, ρ k (α) 2 }. If α J is such that Ω(α ) Ω(α) for all α J, then one of the following holds. (1) φ 1 (α ) = φ k (α ) = ψ(α ) ; (2.a) φ 1 (α ) = φ k (α ) > ψ(α ), α is a local minimum of the function max{φ 1 (α), φ k (α)}, (2.b) φ 1 (α ) = ψ(α ) > φ k (α ), α is a local minimum of the function max{φ 1 (α), ψ(α) }, (2.c) φ k (α ) = ψ(α ) > φ 1 (α ), α is a local minimum of the function max{φ k (α), ψ(α) }; (3.a) max{φ 1 (α ), φ k (α )} < ψ(α ) α is a local minimum of the function ψ(α ), (3.b) max{φ 1 (α ), ψ(α ) } < φ k (α ) α is a local minimum of the function φ k (α ), (3.c) max{φ k (α ), ψ(α ) } < φ 1 (α ) α is a local minimum of the function φ 1 (α ). By the proof of Theorem 3.2, we see that the function ψ(α) has a differentiable local maximum at α > 0, two non-differentiable local minima at λ 2 λ 1, where ψ(λ 1 ) = ψ(λ 2 ) = 0. Thus, condition (3.a) cannot hold. Similarly, for j = 1, k, the proof of Theorem 3.2 shows that the function φ j (α) has a local maximum at α > 0, two local minima at q j λ 1 λ 2, where φ j ( q j ) = φ j ( λ 1 λ 2 ) = 0. Thus, none of conditions (3.b) or (3.c) holds. Now suppose that (1) or (2.a) holds. Then φ 1 (α ) = φ k (α ) implies that α = q 1 q k. In both cases, we have ρ 1 (α ) = φ 1 (α ) + φ 1 (α ) ψ(α ) = φ k (α ) + φ k (α ) ψ(α ) = ρ k (α ), Ω(α ) = max{ρ 1 (α ) 2, ρ k (α ) 2 }. Suppose that (2.b) holds. If φ k (α ) ψ(α ), then ρ k (α ) = φ k (α ) + φ k (α ) ψ(α ) < φ 1 (α ) + φ 1 (α ) ψ(α ) = ρ 1 (α );

14 The Optimal Parameter in HSS Method 13 otherwise, ρ k (α ) = (α λ 1 )(α λ 2 ) (α + λ 1 )(α + λ 2 ) ρ 1(α ). Thus, Ω(α ) = ρ 1 (α ) 2 = max{ρ 1 (α ) 2, ρ k (α ) 2 }. Moreover, since α is a local minimum of the function max{φ 1 (α), ψ(φ) }, by Remark 3.4 we see that α is a root of the equation (α 2 + q 2 1) 2 (α 2 λ 2 1)(α 2 λ 2 2) = (α 2 q 2 1) 2 (α 2 λ 1 λ 2 ) 2 or (α 2 + q 2 1) 2 (λ 2 1 α 2 )(α 2 λ 2 2) = (α 2 q 2 1) 2 (α 2 λ 1 λ 2 ) 2. Suppose that (2.c) holds. We can use an argument similar to the case of (2.b) to conclude that Ω(α ) = ρ k (α ) 2 = max{ρ 1 (α ) 2, ρ k (α ) 2 } that α is a root of the equation (α 2 + qk 2 )2 (α 2 λ 2 1)(α 2 λ 2 2) = (α 2 qk 2 )2 (α 2 λ 1 λ 2 ) 2 or (α 2 + q 2 k )2 (λ 2 1 α 2 )(α 2 λ 2 2) = (α 2 q 2 k )2 (α 2 λ 1 λ 2 ) 2. Note that in all the cases (1), (2.a), (2.b), (2.c), we have Ω(α ) = max{ρ 1 (α ) 2, ρ k (α ) 2 }. Consequently, if α J yields the smallest Ω(α), then ρ(α) 2 = max{ρ 1 (α) 2, ρ k (α) 2 } Ω(α) Ω(α ) = max{ρ 1 (α ) 2, ρ k (α ) 2 } = ρ(α ) 2. So, we only need to consider those α satisfying the specified equations in the theorem to determine the optimal parameter α. Case 3. Suppose r = k s > k. Let ρ 0 (α) = f 2 (α). Then ρ(α) = max{ρ j (α) 0 j k}. Similarly to the proof of Case 2, we can show that ρ(α) ρ( λ 1 λ 2 ) if α is positive lies outside the interval ( J = (λ 3 2/λ 1 ) 1/2, (λ 1 λ 2 ) 1/2). So, we can focus on the interval J. For α J, we have ρ(α) = max{ρ 1 (α), ρ k (α)}. Note that in this case, for each α J, we have α λ2 / α + λ 2 α λ 1 / α + λ 1,

15 14 Z.-Z. Bai, G.H. Golub C.-K. Li hence ρ 0 (α) = α λ 2 α + λ 2 (α λ 1 )(α λ 2 ) (α + λ 1 )(α + λ 2 ) max{ρ 1(α), ρ k (α)}. Finally, if α J is such that ρ(α ) < ρ( λ 1 λ 2 ), we can show that α must satisfy one of the three specified equations using an argument similar to the one in Case 2. Case 4. Suppose r = s = k. Then ρ(α) = max{ρ j (α) 1 j k}. Note that if α is positive lies outside the interval ( J = (λ 3 2/λ 1 ) 1/2, (λ 3 1/λ 2 ) 1/2), then each of the singular values of M(α), which is equal to α λ 1 / α + λ 1 or α λ 2 / α + λ 2, has magnitude larger than ρ( λ 1 λ 2 ). Thus ρ(α) det(m(α)) 1/n ρ( λ 1 λ 2 ). So, we can focus on those α J. For each α J, we can show that ρ(α) = max{ρ 1 (α), ρ k (α)}. Moreover, if α J is such that ρ(α ) < ρ( λ 1 λ 2 ), we can show that α must satisfy one of the three specified equations. Remark 4.2. The second group of equations in Theorem 4.1 can be reduced to a group of quadratic equations the last one can be reduced to a group of quartic equations with respect to β = α 2, respectively, analogously to those in Remark 3.3. As an illustration of Theorem 4.1 Remark 4.2, we consider a simple example for which λ 1 = 2, λ 2 = 1, q 1 = 1 q k = 2. Obviously, α0 = is the root of the first equation in Theorem 4.1. In addition, by straightforward computations, we know that the only positive roots of the group of the equations with respect to q 1 = 1 in Theorem 4.1 are α1 = 1, α2 = (7 α3 = ) , those with respect to q k = 2 in Theorem 4.1 are α 6 = 20 α4 = 2, α5 = ( )

16 The Optimal Parameter in HSS Method 15 We remark that now the second one in the group of equations with respect to q 1 = 1 is equivalent to (β 1)(2β 3 7β 2 + β 8) = 0, the second one in the group of equations with respect to q k = 2 is equivalent to (β 4)(2β 3 β β 32) = 0. Based on Theorem 3.1, from direct calculations we have ρ(m(α 0)) = , ρ(m(α 1)) = 0, ρ(m(α 2)) = 3(7 3 5) ρ(m(α 3)) 0.201, ρ(m(α 4)) = 0, ρ(m(α 5)) 0.146, ρ(m(α 6)) Therefore, for the HSS iteration method, the optimal parameter is α = 1 or α = 2 the corresponding optimal convergence factor is ρ(m(α )) = 0. On the other h, from Theorem 2.1 we can easily obtain Obviously, it holds that ρ(m(α )) < σ( α). α = 2 σ( α) = Remark 4.3. Our proof techniques can be used to hle the case when λ 2 = 0, which also occur in applications; see [4] its references. In such case, we may normalize λ 1 = 1, we always assume that s = k to ensure that A is invertible. In such case, we can use the analysis of Case 2 Case 4 in our proof of Theorem 4.1. Note that in this case, we have { ψ(α) = α 1 α + 1, φ (α 2 qj 2 j(α) = )α } 2 (α 2 + qj 2)(α + λ for j = 1, k. 1) If q 1 = q k, then A is unitarily similar to with A j = A 1 A k I u [ 1 q1 q 1 0 ], j = 1, 2,..., k. So, the analysis reduces to the 2-by-2 case, Theorem 3.2 applies. Suppose q 1 > q k > 0. Then the optimal value α can be easily determined by checking whether ρ 1 (α) ρ k (α) intersect at a point α such that φ 1 (α ) ψ(α ) = φ k (α ) ψ(α ) 0. This happens if only if q 1 q k 1 2 (q 1 + q k ). In this case, α = q 1 q k is the optimal parameter, ρ(m(α )) = q 1 q k q1 q k q 1 + q k q1 q k (q 1 + q k ) 2 4q1 2q2 k (. q 1 q k + 1)(q 1 q k )

17 16 Z.-Z. Bai, G.H. Golub C.-K. Li Otherwise, α = q 1 2q1 1 is the optimal parameter such that φ 1(α ) = ψ(α ) > φ k (α ) ρ(m(α )) = q 1 1 q 1 + 2q 1 1. Note that the second case rarely appears in applications its discussion was not included in [4]. Also, note that there were some typos in the formula of ρ(m(α )) for the first case in [4]. 5 Estimation of Optimal Parameters for n-by-n Matrices Numerical Examples In general, for a nonsymmetric positive definite system of linear equations (1.1), the eigenvalues of its coefficient matrix A is evidently contained in the complex domain D A := [λ min, λ max ] ı[ q, q], where ı is the imaginary unit, λ min λ max are, respectively, the smallest the largest eigenvalues of the Hermitian part H, q is the largest module of the eigenvalues of the skew-hermitian part S, of the coefficient matrix A. If a reduced (simpler lower-dimensional) matrix A R whose eigenvalues possess the same contour as the domain D A is used to approximate the matrix A, then we may expect that the main mathematical numerical properties of the HSS iteration method for the original linear system with the coefficient matrix A can be roughly preserved by the HSS iteration method applied to the linear system with the reduced coefficient matrix A R. A simple choice of the reduced matrix is given by [ ] λmax q A R = q with q = S or q = ρ(h 1 S) λ min λ max. We can then use Theorem 3.2 Corollary 3.5 to estimate the optimal parameter α of the HSS iteration method as follows. Estimation 5.1. Let A R n n be a positive definite matrix, H, S R n n be its symmetric skew-symmetric parts, respectively. Let λ min = min{λ λ λ(h)} λ max = max{λ λ λ(h)}. Suppose λ min q = S or q = ρ(h 1 S) λ min λ max. Then one can use the positive roots of the equation or (α 2 + q 2 ) 2 (α 2 λ 2 max)(α 2 λ 2 min) = (α 2 q 2 ) 2 (α 2 λ min λ max ) 2 (α 2 + q 2 ) 2 (λ 2 max α 2 )(α 2 λ 2 min) = (α 2 q 2 ) 2 (α 2 λ min λ max ) 2 to estimate the optimal parameter α > 0 satisfying ρ(m(α )) = min{ρ(m(α)) α > 0}. Here, M(α) is the iteration matrix of the HSS iteration method, see (2.2).

18 The Optimal Parameter in HSS Method 17 We first illustrate our estimates with the following example of a general nonsymmetric positive definite system of linear equations, see [2]. Consider the two-dimensional convection-diffusion equation (u xx + u yy ) + δ(u x + u y ) = g(x, y) on the unit square [0, 1] [0, 1], with constant coefficient δ subject to Dirichlettype boundary conditions. When the five-point centered finite difference discretization is applied to it, we get the system of linear equations (1.1) with the coefficient matrix A = T I + I T, (5.1) where the equidistant step-size h = 1 m+1 is used in the discretization on both directions the natural lexicographic ordering is employed to the unknowns. In addition, denotes the Kronecker product, T is a tridiagonal matrix given by T = tridiag( 1 R e, 2, 1 + R e ), R e = δh 2 is the mesh Reynolds number. We remark that here the first-order derivatives are also approximated by the centered difference scheme. From Lemma A.1 in [2, Appendix], we know that γ min λ min = min j,k(h) = 4(1 cos(πh)), 1 j,k m γ max λ max = max j,k(h) = 4(1 + cos(πh)) 1 j,k m min λ j,k(s) = 0, 1 j,k m max λ j,k(s) = 4R e cos(πh). 1 j,k m Therefore, the optimal parameter α that minimizes the upper bound σ(α) of the convergence factor ρ(m(α)) of the HSS iteration method is given by see Theorem 2.1. α γ min γ max = λ min λ max = 4 sin(πh), In Table 1 we list the experimental optimal parameter α (denoted by α exp ), the estimated optimal parameter α (denoted by α est ) determined by Estimation 5.1, the upperbound minimizer α, the corresponding spectral radii ρ(m(α)) of the HSS iteration matrix M(α) for α = α exp, α est α. From this table we see that α always overestimates ρ(m(α)) when compared with α est, that α est yields quite a good approximation to α exp. In Table 2 we list the number of iterations (denoted by IT ) the elapsed CPU time in seconds (denoted by CPU ) of the HSS iteration method when it is applied to the nonsymmetric positive definite system of linear equations (1.1) with coefficient matrix

19 18 Z.-Z. Bai, G.H. Golub C.-K. Li Table 1: α versus ρ(m(α)) when m = 32 δ α exp ρ(m(α exp )) α est ρ(m(α est )) α ρ(m( α)) Table 2: IT CPU when m = 32 ε = 10 6 δ α exp IT CPU α est IT CPU (5.1). From this table we see that the numerical results produced with α est coincide with those using α exp, the match is quite pertinent. Here in the next example, we choose the right-h-side vector f such that the exact solution of the system of linear equations is (1, 1,..., 1) T R n. In addition, all runs are initiated from the initial vector x (0) = 0, terminated if the current iteration satisfy f Ax (k) 2 f Ax (0) 2 ε. The experiments are run in MATLAB (version 6.1) with a machine precision The machine used is a Pentium-III 500 personal computer with 256M memory. Then, we use the following example of 2-by-2 block system of linear equations to further confirm the above observations. Consider the system of linear equations (1.1) with the coefficient matrix [ ] B E A = E T, (5.2) 0.5I where B = [ I TH + T H I 0 0 I T H + T H I E = [ I F F I ] R 2m2 2m 2, (5.3) ] R 2m2 m 2, (5.4)

20 The Optimal Parameter in HSS Method 19 T H = tridiag( 1, 2, 1) R m m, F = δh tridiag( 1, 1, 0) R m m, (5.5) with h = 1 m+1 the discretization meshsize, see [4]. For this example, we have r = 2m 2 s = m 2. Hence, the total number of variables is r + s = 3m 2. In Table 3 we list the experimental optimal parameter α exp, the estimated optimal parameter α est determined by Estimation 5.1, the corresponding spectral radii ρ(m(α)) of the HSS iteration matrix M(α) for α = α exp α est. Here, considering Theorem 4.1 the two-by-two block structure of the matrix A in (5.2), we may apply the equations in Estimation 5.1 to the q which is either the largest or the smallest singular value (denoted, respectively, by q 1 or q k ) of the matrix E, also consider the two points α = λ min λ max α = q 1 q k, to obtain a more accurate estimate to the optimal iteration parameter for the HSS iteration method. From this table we can see that α est yields quite a good approximation to α exp. Table 3: α versus ρ(m(α)) δ m α exp ρ(m(α exp )) α est ρ(m(α est )) In Table 4 we list the number of iterations the elapsed CPU time of the HSS iteration method when it is applied to the 2 2 block system of linear equations (1.1) with the coefficient matrix (5.2)-(5.5). From this table we can see that the numerical results produced with α est coincide with those using α exp, the match is quite pertinent. Table 4: IT CPU when ε = 10 6 δ m IT α exp CPU IT α est CPU

21 20 Z.-Z. Bai, G.H. Golub C.-K. Li 6 Proof of Remark 3.3 Here we give the proof of Remark 3.3. The assertion concerning λ 1 = λ 2 the assertion concerning equation (3.3) can be readily verified. We consider the equation (3.4) the reduced equation (3.5). Note that for α > 0, the right side of (3.4) is nonnegative, the left side is nonnegative on the interval [λ 2, λ 1 ] only. So, all the positive solutions β = α 2 of the equation (3.5) lie in [λ 2 2, λ2 1 ]. Define f(β) = (λ 2 1 β)(β λ 2 2), g 1 (β) = (β λ 1 λ 2 ) 2, g 2 (β) = (β q2 ) 2 (β + q 2 ) 2, g(β) = g 1 (β)g 2 (β) h(β) = f(β) g(β), for β > 0. Then β is a positive solution of (3.5) if only if β is a positive solution of h(β) = 0. Obviously, the graph of f(β) is a concave parabola which intersects the β-axis at β = λ 2 2 λ 2 1, while the graph of g 1(β) is a convex parabola touching the β-axis at β = λ 1 λ 2. Note that f(β) = g 1 (β) has two solutions in the intervals [λ 2 2, λ 1λ 2 ) ((λ λ2 2 )/2, λ2 1 ]. The graph of g(β) can be obtained from that of g 1 (β) by reducing the vertical height of each point by the factor g 2 (β) (0, 1). Thus, the solution of f(β) = g 1 (β) nearer to λ 2 2 will move further left to a solution of f(β) = g(β); the solution of f(β) = g 1 (β) nearer to λ 2 1 will move further right to a solution of f(β) = g(β). We claim that h(β) cannot have more than 2 positive solutions in [λ 2 2, λ2 1 ], equivalently, (3.5) cannot have more than 2 positive solutions in [λ 2 2, λ2 1 ]. First, we prove the claim if q 2 < λ 2 2 or q2 > λ 2 1. Let p(β) be the polynomial on the left side of (3.5) divided by 2. Then the product of the roots of (3.5) is the constant term of p(β) equals the positive number q 4 λ 2 1 λ2 2. So, if (3.5) have more than 2 positive roots, then it must have 4 positive roots, say, β 1 β 2 β 3 β 4. Suppose p (β) has zeros β 1 β 2 β 3, p (β) has zeros β 1 β 2, p (β) has zero β 1. We have β 1 β 1 β 2 β 2 β 3 β 1 β 1 β 2. Thus, β 1 β 1 β 3. Now, examining the constant terms of the polynomials p (β)/4 p (β)/24, we see that β 1β 2β 3 = q 4 (λ 1 + λ 2 ) 2 /8 (6.1) β 1 = (λ 1 + λ 2 ) 2 /8. If q 2 < λ 2 2, then q2 < β 1 β 1 β 2 hence β 1β 2β 3 > (q 2 )(q 2 )(β 1 ) = q 4 (λ 1 + λ 2 ) 2 /8, which contradicts (6.1). Similarly, if q 2 > λ 2 1, then q2 > β 4 β 3 β 2 hence β 1β 2β 3 < β 1 (q 2 )(q 2 ) = q 4 (λ 1 + λ 2 ) 2 /8, which again contradicts (6.1). Thus, our claim is proved in these cases.

22 The Optimal Parameter in HSS Method 21 Next, we assume that q 2 [λ 2 2, λ2 1 ]. Then g(β) = (β q2 ) 2 (β λ 1 λ 2 ) 2 (λ 2 1 β)(β λ2 2 ) has two zeros in [λ 2 2, λ2 1 ], namely, β = q2 β = λ 1 λ 2. If λ 2 2 q2 λ 1 λ 2, then f is increasing on [λ 2 2, (λ2 1 +λ2 2 )/2] q2 λ 1 λ 2 (λ 2 1 +λ2 2 )/2. So, f is increasing on [λ 2 2, q2 ]. Since g(β) is decreasing on [λ 2 2, q2 ], the function h(β) have different signs at λ 2 2 q2, it follows that h(β) = 0 has a root in [λ 2 λ 1, q 2 ]. On (q 2, (λ λ2 2 )/2], we have f(β) > g 1(β) g 1 (β)g 2 (β). Thus, f(β) = g 1 (β)g 2 (β) has no root in this interval. Finally, on the interval [(λ λ2 2 )/2, λ2 1 ], f is decreasing g(β) is increasing, the function h(β) assume nonzero values with different signs at the end points. Thus, f(β) = g 1 (β)g 2 (β) has a root in ((λ λ2 2 )/2, λ2 1 ). Suppose λ 1 λ 2 < q 2 λ 2 1. On [λ2 2, λ 1λ 2 ], f is increasing g(β) is decreasing. Moreover, the function h(β) assume nonzero values with different signs at λ 2 2 λ 1λ 2, it follows that h(β) = 0 has a root in (λ 2 2, λ 1λ 2 ). Suppose q 2 < (λ λ2 2 )/2. On (λ 1λ 2, (λ λ2 2 )/2], we have f(β) > g 1(β) g 1 (β)g 2 (β). Thus, f(β) = g 1 (β)g 2 (β) has no root in this interval. On the interval [(λ λ2 2 )/2, λ2 1 ], f is decreasing g(β) is increasing, the function h(β) assume nonzero values with different signs at the end points. Thus, f(β) = g(β) has a root in ((λ λ2 2 )/2, λ2 1 ). Finally, if q 2 [(λ λ2 2 )/2, λ2 1 ], then for β (λ 1λ 2, q 2 ) we have hence (β λ 1 λ 2 ) 2 /(β + q) 2 (0, 1) f(β) > (β q) 2 > g(β). Thus, f(β) = g(β) has no solution in [(λ λ2 2 )/2, q2 ). Now, on the interval [q 2, λ 2 1 ], f is decreasing g(β) is increasing, the function h(β) assume values with different signs at the end points. Thus, f(β) = g(β) has a root in [q 2, λ 2 1 ]. Acknowledgement: Part of this work was done during the first author s visits to the Department of Mathematics of The College of William Mary during May June 2004, to Shanghai Key Laboratory of Contemporary Applied Mathematics of Fudan University. The authors are very much indebted to Jun-Feng Yin for his kind help on running the numerical results.

23 22 Z.-Z. Bai, G.H. Golub C.-K. Li References [1] Z.-Z. Bai, G.H. Golub, L.-Z. Lu J.-F. Yin, Block triangular skew-hermitian splitting methods for positive definite linear systems, SIAM J. Sci. Comput., 26:3(2005), [2] Z.-Z. Bai, G.H. Golub M.K. Ng, Hermitian skew-hermitian splitting methods for non-hermitian positive definite linear systems, SIAM J. Matrix Anal. Appl., 24(2003), [3] Z.-Z. Bai, G.H. Golub M.K. Ng, On successive-overrelaxation acceleration of the Hermitian skew-hermitian splitting iterations, Tech. Report SCCM-02-06, Scientific Computing Computational Mathematics Program, Department of Computer Science, Stanford University, Available on line at [4] Z.-Z. Bai, G.H. Golub J.-Y. Pan, Preconditioned Hermitian skew-hermitian splitting methods for non-hermitian positive semidefinite linear systems, Numer. Math., 98(2004), [5] M. Benzi, M.J. Ger G.H. Golub, Optimization of the Hermitian skew- Hermitian splitting iteration for saddle-point problems, BIT, 43(2003), [6] M. Benzi G.H. Golub, A preconditioner for generalized saddle point problems, SIAM J. Matrix Anal. Appl., 26(2004), [7] M. Benzi M.K. Ng, Preconditioned iterative methods for weighted Toeplitz least squares problems, SIAM J. Matrix Anal. Appl., in press. [8] D. Bertaccini, G.H. Golub, S. Serra-Capizzano C.T. Possio, Preconditioned HSS methods for the solution of non-hermitian positive definite linear systems applications to the discrete convection-diffusion equation, Numer. Math., 99(2005), [9] G.H. Golub C.F. Van Loan, Matrix Computations, 3rd Edition, The Johns Hopkins University Press, Baltimore, [10] G.H. Golub A.J. Wathen, An iteration for indefinite systems its application to the Navier-Stokes equations, SIAM J. Sci. Comput., 19(1998), [11] M.-K. Ho M. K. Ng, Splitting iterations for circulant-plus-diagonal systems, Numer. Linear Algebra Appl., 12:8(2005), [12] M.K. Ng, Circulant skew-circulant splitting methods for Toeplitz systems, J. Comput. Appl. Math., 159(2003), [13] J.-Y. Pan, Z.-Z. Bai M.K. Ng, Two-step waveform relaxation methods for implicit linear initial value problems, Numer. Linear Algebra Appl., 12(2005), [14] J.-Y. Pan, M.K. Ng Z.-Z. Bai, New preconditioners for saddle point problems, Appl. Math. Comput., in Press.

24 The Optimal Parameter in HSS Method 23 [15] Y. Saad, Iterative Methods for Sparse Linear Systems, PWS Publishing Company, Boston, [16] V. Simoncini M. Benzi, Spectral properties of the Hermitian skew-hermitian splitting preconditioner for saddle point problems, SIAM J. Matrix Anal. Appl., 26(2004), [17] R.S. Varga, Matrix Iterative Analysis, 2nd Revised Exped Edition, Springer- Verlag, Berlin, [18] H.A. van der Vorst, Iterative Krylov Methods for Large Linear Systems, Cambridge University Press, Cambridge, [19] C.-L. Wang Z.-Z. Bai, Sufficient conditions for the convergent splittings of non- Hermitian positive definite matrices, Linear Algebra Appl., 330(2001), [20] D.M. Young, Iterative Solution of Large Linear Systems, Academic Press, New York London, 1971.