A relaxed splitting preconditioner based on matrix splitting is introduced in this paper for linear systems of saddle point problem arising from numerical solution of the incompressible Navier-Stokes equations. Spectral analysis of the preconditioned matrix is presented, and numerical experiments are carried out to illustrate the convergence behavior of the preconditioner for solving both steady and unsteady incompressible flow problems.
1. Introduction
We consider systems of linear equations arising from the finite-element discretization of the incompressible Navier-Stokes equations governing the flow of viscous Newtonian fluids. The primitive variables formulation of the Navier-Stokes equations is
(1.1)∂u∂t-υΔu+(u⋅∇)u+∇p=fonΩ×(0,T],(1.2)divu=0onΩ×[0,T],(1.3)u=gon∂Ω×[0,T],(1.4)u(x,0)=u0(x)onΩ,
where Ω⊂ℝ2 is an open bounded domain with sufficiently smooth boundary ∂Ω, [0,T] is an time interval of interest, u(x,t) and p(x,t) are unknown velocity and pressure fields, υ is the kinematic viscosity, Δ is the vector Laplacian, ∇ is the gradient, div is the divergence, and f, g, and u0 are given functions. The Stokes problem is obtained by dropping the nonlinearity (u·∇)u from the momentum equation (1.1). Refer to [1] for an introduction to the numerical solution of the Navier-Stokes equations. Implicit time discretization and linearization of the Navier-Stokes equations by Picard or Newton fixed iteration result in a sequence of (generalized) Oseen problems. The Oseen problems by spatial discretization with LBB-stable finite elements (see [1, 2]) are reduced to a series of large sparse systems of linear equations with a saddle point matrix structure as follows:
(1.5)A~x=b,
with
(1.6)A~=[ABT-B0],x=(up),b=(f-g),
where u and p represent the discrete velocity and pressure, respectively. In two-dimensional cases, A=diag(A1,A2) denotes the discretization of the reaction diffusion, and each diagonal submatrix Ai is a scalar discrete convection-diffusion operator represented as
(1.7)Ai=σV+υL+Ni(i=1,2),
where V denotes the velocity mass matrix, L the discrete (negative) Laplacian, and Ni the convective terms. The matrix A is positive definite in the sense that AT+A is symmetric positive definite. Matrix BT=(B1T,B2T) denotes the discrete gradient with B1T, B2T being discretizations of the partial derivatives ∂/∂x, ∂/∂y, respectively. f=(f1,f2)T and g contain the forcing and boundary terms.
In the past few years, a considerable amount of work has been spent in developing efficient solvers for systems of linear equations in the form of (1.5); see [3] for a comprehensive survey. Here we consider preconditioned Krylov subspace methods, in particular preconditioned GMRES [4] in this paper. The convergence performance of this method is mainly determined by the underlying preconditioner employed. An important class of preconditioners is based on the block LU factorization of the coefficient matrix, including a variety of block diagonal and triangular preconditioners. A crucial ingredient in all these preconditioners is an approximation to the Schur complement S=BA-1BT. This class of preconditioners includes the pressure convection diffusion (PCD) preconditioner, the least-squares commutator (LSC) preconditioner, and their variants [5–7]. Somewhat related to this class of preconditioners are those based on the augmented Lagrangian (AL) reformulation of the saddle point problem; see [8–11]. Other types of preconditioners for the saddle point problems include those based on the Hermitian and skew-Hermitian splitting (HSS) [12–15] and the dimensional splitting (DS) [16] of the coefficient matrix A~. In [17], a relaxed dimensional factorization preconditioner is introduced.
The remainder of the paper is organized as follows. In Section 2, we present a relaxed splitting preconditioner based on matrix splitting and prove that the preconditioned matrix has eigenvalue 1 of algebraic multiplicity at least n (recall that n is the number of velocity degrees of freedom). In Section 3, we show the results of a series of numerical experiments indicating the convergence behavior of the relaxed splitting preconditioner. In the final section, we draw our conclusions.
2. A Relaxed Splitting Preconditioner 2.1. A Splitting of the Matrix
In this paper, we limit to 2D case. The system matrix A~ admits the following splitting:
(2.1)A~=[A10B1T0A2B2T-B1-B20]=[A100000-B100]+[00B1T0A2B2T0-B20]=H+S,
where A1∈ℝn1×n1, A2∈ℝn2×n2, B1∈ℝm×n1, and B2∈ℝm×n2. Thus, A~∈ℝ(n+m)×(n+m) is of dimension n=n1+n2. Let α>0 be a parameter and denote by I the identity matrix of order n1+n2+m. Then, H+αI and S+αI are both nonsingular, nonsymmetric, and positive definite. Consider the two splittings of A~:
(2.2)A~=(H+αI)-(αI-S),A~=(S+αI)-(αI-H).
Associated to these splittings is the alternating iteration, k=0,1,…,
(2.3)(H+αI)xk+1/2=(αI-S)xk+b,(S+αI)xk+1=(αI-H)xk+1/2+b.
Eliminating xk+1/2 from these, we can rewrite (2.3) as the stationary scheme:
(2.4)xk+1=Tαxk+c,k=0,1,…,
where
(2.5)Tα=(S+αI)-1(αI-H)(H+αI)-1(αI-S)
is the iteration matrix and c=2α(S+αI)-1(H+αI)-1. The iteration matrix Tα can be rewritten as follows:
(2.6)Tα=(S+αI)-1(H+αI)-1(αI-H)(αI-S)=(S+αI)-1(H+αI)-1[(αI+H)(αI+S)-2αA~]=I-[12α(H+αI)(S+αI)]-1A~=I-Pα-1A~,
where Pα=(1/2α)(H+αI)(S+αI).
Obviously, Pα is nonsingular and c=Pα-1b. As in [18], one can show there is a unique splitting A~=Pα-Qα such that the iteration Tα is the matrix induced by that splitting, that is, Tα=Pα-1Qα=I-Pα-1A~. Matrix Qα is given by Qα=(1/2α)(αI-H)(αI-S).
2.2. A Relaxed Splitting Preconditioner
The relaxed splitting preconditioner is defined as follows:
(2.7)M=[A101αA1B1T0A2B2T-B1-B2αI-1αB1B1T].
It is important to note that the preconditioner M can be written in a factorized form as
(2.8)M=1α[A1000αI0-B10αI][αI0B1T0A2B2T0-B2αI]=[A1000I0-B10I][I01αB1T0A2B2T0-B2αI]=[A1000I0-B10I][I01α2B1T0I1αB2T00I][I1α2B1TB200A^2000αI][I000I00-1αB2I],
where A^2=A2+(1/α)B2TB2. Note that both factors on the right-hand side are invertible provided that A1 have A^2 have positive definite symmetric parts. Hence, the new preconditioner is nonsingular. This condition is satisfied for both Stokes and Oseen problems. We can see from (2.1) and (2.7) that the difference between M and A~ is given by
(2.9)R=M-A~=[001αA1B1T-B1T00000αI-1αB1B1T].
This observation suggests that M could be a good preconditioner, since the appropriate values for the parameters involved in the new preconditioners are estimated. Furthermore, the structure of (2.9) somewhat facilitates the analysis of the eigenvalue distribution of the preconditioned matrix. In the following, we analyze the spectral properties of the preconditioned matrix T=A~M-1.
Theorem 2.1.
The preconditioned matrix T=A~M-1 has an eigenvalue 1 with multiplicity at least n, and the remaining eigenvalues are λi, where λi are the eigenvalues of an m×m matrix Zα:=(1/α)(S1+S2)-(1/α2)S2S1 with S1=B1A1-1B1T and S2=B2A^2-1B2T.
Proof.
First of all, from T^:=M-1(A~M-1)M=M-1A~ we see that the right-preconditioned matrix T is similar to the left-preconditioned one T^, then T and T^ have the same eigenvalues. Furthermore, we have
(2.10)T^=I-M-1R=I-[I000I001αB2I][I-1α2B1TB2A^2-100A^2-10001αI][I0-1α2B1T0I-1αB2T00I]×[A1-1000I0B1A1-10I][001αA1B1T-B1T00000αI-1αB1B1T]=[I0*0I*001α(S1+S2)-1α2S2S1].
Therefore, from (2.10) we can see that the eigenvalues of T are given by 1 (with multiplicity at least n=n1+n2) and by the λi’s.
Lemma 2.2.
Let Aα=(A1-(1/α)B1TB1-(1/α)B1TB20A2)∈ℝn×n, α>σmax(B1TB1)/σmin(A1), and A1, and A2 be positive definite. Then Aα is positive definite.
Lemma 2.3.
Let Aα∈ℝn×n and B∈ℝm×n (m≤n). Let α∈ℝ, and assume that matrices Aα, Aα+(1/α)BTB, BAα-1BTand B(Aα+(1/α)BTB)-1BT are all invertible. Then
(2.11)[B(Aα+1αBTB)-1BT]-1=(BAα-1BT)-1+1αI.
Theorem 2.4.
Let α>σmax(B1TB1)/σmin(A1). The remaining eigenvalues λi of Zα are of the form:(2.12)λi=μiα+μi,
where the μi’s satisfy the eigenvalue problem: BAα-1BTϕi=μiϕi.
Proof.
We note
(2.13)Zα=1α(S1+S2)-1α2S2S1=1α(B1B2)(A1-10-1αA^2-1B2TB1A1-1A^2-1)(B1TB2T)=1α(B1B2)(A101αB2TB1A^2)-1(B1TB2T)=1α(B1B2)((A1-1αB1TB1-1αB1TB20A2)+(1αB1TB11αB1TB21αB2TB11αB2TB2))-1(B1TB2T)=1αB(Aα+1αBTB)-1BT.
Thus, the remaining eigenvalues are the solutions of the eigenproblem:
(2.14)1αB(Aα+1αBTB)-1BTϕi=λiϕi.
By Lemma 2.3, we obtain
(2.15)1αϕi=λi(B(Aα+1αBTB)-1BT)-1ϕi=λi(BAα-1BT)-1ϕi+λiαϕi.
Hence, λi=μi/(α+μi), where μi’s satisfy the eigenvalue problem BAα-1BTϕi=μiϕi.
In addition, we obtain easily that the remaining eigenvalues λi→0 as α→∞. Figures 1 and 2 show this behavior, that is, the nonunity eigenvalues of the preconditioned matrix are increasingly clustered at the origin as the parameters become larger.
Spectrum of preconditioned steady Oseen matrix, 32 × 32 grid with υ=0.1.
Spectrum of preconditioned generalized steady Oseen matrix, 32 × 32 grid with υ=0.001.
2.3. Practical Implementation of the Relaxed Splitting Preconditioner
In this subsection, we outline the practical implementation of the relaxed splitting preconditioner in a subspace iterative method. The main step is applying the preconditioner, that is, solving linear systems with the coefficient matrix M. From (2.8), we can see that the relaxed splitting preconditioner can be factorized as follows:
(2.16)M=[A1000I0-B10I][I01α2B1T0I1αB2T00I][I1α2B1TB200A^2000αI][I000I00-1αB2I],
showing that the preconditioner requires solving two linear systems at each step, with coefficient matrices A1 and A^2=A2+(1/α)B2TB2. Several different approaches are available for solving linear systems involving A1 and A^2. We defer the discussion of these to Section 3.
We conclude this section with a discussion of diagonal scaling. We found that scaling can be beneficial for the relaxed splitting preconditioner. Unless otherwise specified, we perform a preliminary symmetric scaling of the linear systems A~x=b in the form D-1/2A~D-1/2y=D-1/2b with y=D1/2x, and D=diag(D1,D2,I), where diag(D1,D2) is the main diagonal of the velocity submatrix A. Incidentally, it is noted that diagonal scaling is very beneficial for the HSS preconditioner (see [13]) and the DS preconditioner (see [16, 17]).
3. Numerical Experiments
In this section, numerical experiments are carried out for solving the linear system coming from the finite-element discretization of the two-dimensional linearized Stokes and Oseen models of incompressible flow in order to verify the performance of our preconditioner. The test problem is the leaky lid-driven cavity problem generated by the IFISS software package [19]. We used a zero initial guess and stopped the iteration when ||rk||2/||b||2≤10-6, where rk is the residual vector. The relaxed splitting preconditioner is combined with restarted GMRES(m). We set m=30.
We consider the 2D leaky lid-driven cavity problem discretized by the finite-element method on uniform grids [1]. The subproblems arising from the application of the relaxed splitting preconditioner are solved by direct methods. We use AMD reordering technique [20, 21] for the degrees of freedom that makes the application of the Cholesky (for Stokes) or LU (for Oseen) factorization of A1 and A^2 relatively fast. For simplicity, we use α=100 for all numerical experiments.
In Table 1, we show iteration counts (referred to as “its”) for the relaxed splitting preconditioned GMRES(30) when solving the steady Stokes problem on a sequence of uniform grids. We see that the iteration count is independent of mesh size involved in the Q2-Q1 and the Q2-P1 finite-element scheme. The Q2-P1 finite-element scheme has much better profile than the Q2-Q1 finite-element scheme.
Iterations of preconditioned GMRES(30) for steady Stokes problem.
Grid
Q2-Q1
Q2-P1
16 × 16
25
15
32 × 32
26
12
64 × 64
23
10
128 × 128
19
11
256 × 256
16
10
In Tables 2 and 3, we show iteration counts for the steady Oseen problem on a sequence of uniform grids and for different values of υ, using Picard and Newton linearization of generalized Oseen problems, respectively. We found that the relaxed splitting preconditioner has difficulties dealing with low-viscosity, that is, the number of iterations increases with the decrease in the kinematic viscosity. In this case, it appears that the Q2-P1 finite-element scheme gives faster convergence results than the Q2-Q1 finite-element scheme.
Iterations of preconditioned GMRES(30) for steady Oseen problems (Picard).
Grid
υ=1
υ=0.1
υ=0.02
Q2-Q1
Q2-P1
Q2-Q1
Q2-P1
Q2-Q1
Q2-P1
16 × 16
27
14
29
16
48
27
32 × 32
27
12
29
14
54
26
64 × 64
23
10
26
12
49
24
128 × 128
19
11
21
10
37
20
256 × 256
16
10
10
8
17
14
Iterations of preconditioned GMRES(30) for steady Oseen problems (Newton).
Grid
υ=1
υ=0.1
υ=0.02
Q2-Q1
Q2-P1
Q2-Q1
Q2-P1
Q2-Q1
Q2-P1
16 × 16
27
15
29
16
46
26
32 × 32
27
12
29
14
53
24
64 × 64
23
10
26
12
46
23
128 × 128
19
11
21
10
34
18
256 × 256
16
10
10
8
15
13
Next, we report on analogous experiments involving the generalized Stokes problem and the generalized Oseen problem. As we can see from Table 4, for the generalized Stokes problem, the results are virtually the same as those obtained in the steady case. Indeed, we can see from the results in Table 1 that the rate of convergence for the relaxed splitting preconditioned GMRES (30) is essentially independent of mesh size involved in the Q2-Q1 and the Q2-P1 finite-element schemes.
Iterations of preconditioned GMRES(30) for generalized Stokes problems.
Grid
Q2-Q1
Q2-P1
16 × 16
25
13
32 × 32
26
12
64 × 64
23
10
128 × 128
19
11
256 × 256
16
10
In Tables 5 and 6, for generalized Oseen problems, we compare our preconditioner with the RDF preconditioner in [17]. The RDF preconditioner can be factorized as follows:
(3.1)P=[I0B1Tα0I000I][A^1000I0-B10I][I000A^2B2T00αI][I000I00-B2αI],
where A^1=A1+(1/α)B1TB1 and A^2=A2+(1/α)B2TB2. It shows that RDF preconditioner requires solving two linear systems at each step. The new preconditioner requires solving linear systems with A1 and A^2 at each step. We can see that the linear system with A1 is easier to solve than that with A^1. From Tables 5 and 6, we can see for 128 × 128 grid with different viscosities that the RDF preconditioner leads to slightly less iteration counts than the new preconditioner, but the new preconditioner is slightly faster in terms of elapsed CPU time.
Iterations of preconditioned GMRES(30) for generalized Oseen problem (Picard, Q2-Q1, 128 × 128 uniform grids).
Viscosity
RDF
New preconditioner
its
CPU
its
CPU
0.1
12
22.75374
13
15.90265
0.01
7
20.93871
9
14.30008
0.001
5
20.31654
6
13.36891
Iterations of preconditioned GMRES(30) for generalized Oseen problem (Newton, Q2-Q1, 128 × 128 uniform grids).
Viscosity
RDF
New preconditioner
its
CPU
its
CPU
0.1
12
22.86654
13
16.12583
0.01
7
21.02928
8
14.34972
0.001
12
22.81674
16
17.49554
From Figures 3 and 4, we found that for the relaxed splitting preconditioner the intervals containing values of parameter α are very wide. Those imply that the relaxed splitting preconditioner is not sensitive to the value of parameter. Noting that the optimal parameters of the relaxed splitting preconditioner are always larger than 50, we can always take α=100 to obtain essentially optimal results.
Iteration number versus parameter, steady Oseen problem, with υ=0.1. (a) 16 × 16 grid, (b) 32 × 32 grid.
Iteration number versus parameter, generalized Oseen problem, with υ=0.001. (a) 16 × 16 grid, (b) 32 × 32 grid.
4. Conclusions
In this paper, we have described a relaxed splitting preconditioner for the linear systems arising from discretizations of the Navier-Stokes equations and analyzed the spectral properties of the preconditioned matrix. The numerical experiments show good performance on a wide range of cases. We use direct methods for the solution of inner linear systems, but it is not a good idea to solve larger 2D or 3D problems at the constraint of memory and time requirement. In this case, exact solve can be replaced with inexact solve, which requires further research in the future.
Acknowledgments
This research is supported by NSFC (60973015 and 61170311), Chinese Universities Specialized Research Fund for the Doctoral Program (20110185110020), and Sichuan Province Sci. & Tech. Research Project (12ZC1802).
ElmanH. C.SilvesterD. J.WathenA. J.2005New York, NY, USAOxford University PressNumerical Mathematics and Scientific Computation2155549ZBL1203.76098GlowinskiR.CiarletP. G.LionsJ. L.Finite element methods for incompressible viscous flow20039Amsterdam, The NetherlandsNorth-HollandNumerical Methods for Fluids (part3)BenziM.GolubG. H.LiesenJ.Numerical solution of saddle point problems200514113710.1017/S09624929040002122168342ZBL1115.65034SaadY.20032ndPhiladelphia, Pa, USASociety for Industrial and Applied Mathematics1990645ElmanH. C.HowleV. E.ShadidJ.SilvesterD.TuminaroR.Least squares preconditioners for stabilized discretizations of the Navier-Stokes equations200730129031110.1137/0606557422377443ElmanH. C.TuminaroR. S.Boundary conditions in approximate commutator preconditioners for the Navier-Stokes equations2009352572802582816OlshanskiiM. A.VassilevskiY. V.Pressure schur complement preconditioners for the discrete Oseen problem20072962686270410.1137/0706797762357633BenziM.OlshanskiiM. A.An augmented Lagrangian-based approach to the oseen problem20062862095211310.1137/0506464212272253ZBL1126.76028BenziM.OlshanskiiM. A.WangZ.Modified augmented Lagrangian preconditioners for the incompressible Navier-Stokes equations201166448650810.1002/fld.22672827781HamiltonS. P.BenziM.HaberE.New multigrid smoothers for the Oseen problem2010172-355757610.1002/nla.7072650226BenziM.WangZ.Analysis of augmented Lagrangian-based preconditioners for the steady incompressible Navier-Stokes equations20113352761278410.1137/1007979892861647BaiZ. Z.GolubG. H.NgM. K.Hermitian and skew-Hermitian splitting methods for non-Hermitian positive definite linear systems200324360362610.1137/S08954798013954581972670ZBL1036.65032BenziM.GolubG. H.A preconditioner for generalized saddle point problems2004261204110.1137/S08954798024171062112850ZBL1082.65034BenziM.LiuJ.An efficient solver for the incompressible Navier-Stokes equations in rotation form20072951959198110.1137/0606588252350015ZBL1154.76041ChanL.NgM. K.TsingN. K.Spectral analysis for HSS preconditioners20081157772401667ZBL1174.65352BenziM.GuoX.-P.A dimensional split preconditioner for stokes and linearized Navier-Stokes equations2011611667610.1016/j.apnum.2010.08.0052735256BenziM.NgM.NiuQ.WangZ.A relaxed dimensional factorization preconditioner for the incompressible Navier-Stokes equations2011230166185620210.1016/j.jcp.2011.04.0012806221BenziM.SzyldD. B.Existence and uniqueness of splittings for stationary iterative methods with applications to alternating methods199776330932110.1007/s0021100502651452511ZBL0905.65048ElmanH. C.RamageA.SilvesterD. J.Algorithm 886: IFISS, a Matlab toolbox for modelling incompressible flow2007332, article 1410.1145/1236463.12364692326956AmestoyP. R.DavisT. A.DuffI. S.An approximate minimum degree ordering algorithm199617488690510.1137/S08954798942789521410707ZBL0861.65021DavisT. A.2006Philadelphia, Pa, USASociety for Industrial and Applied Mathematics2270673