Several stabilized finite element methods for the Stokes eigenvalue problem based on the lowest equal-order finite element pair are numerically investigated. They are penalty, regular, multiscale enrichment, and local Gauss integration method. Comparisons between them are carried out, which show that the local Gauss integration method has good stability, efficiency, and accuracy properties, and it is a favorite method among these methods for the Stokes eigenvalue problem.
1. Introduction
It is well known that numerical approximation of eigenvalue problems plays an important role in the analysis of the stability of nonlinear PDEs. Meanwhile, they are wildly used in many application areas: structural mechanics and fluid mechanics (see [1]). Thus, development of an efficient and effective computational method for investigating these problems has practical significance and has drawn the attention of many peoples. At the time of writing, numerous works are devoted to these problems (see [2–9], and the references cited therein). Yin et al. [10] derived a general procedure to produce an asymptotic expansion for eigenvalues of the Stokes problem on rectangular mesh. Chen and Lin [11] proposed the stream function-vorticity-pressure method for the eigenvalue problem. Rate of convergence estimates were derived for the approximation of eigenvalues and eigenvectors by Mercier et al. [12]. Xu and Zhou [13] proposed a two-grid discretization scheme for solving eigenvalue problems.
Mixed finite element methods are a natural choice for solving fluid mechanics equations because these equations naturally appear in mixed form in terms of velocity and pressure [14, 15]. In the analysis and practice of employing mixed finite element methods in solving the Stokes equations, the inf-sup condition has played an important role because it ensures a stability and accuracy of the underlying numerical schemes. A pair of finite element spaces that are used to approximate the velocity and the pressure unknowns are said to be stable if they satisfy the inf-sup condition. Intuitively speaking, the inf-sup condition is something that enforces a certain correlation between two finite element spaces so that they both have the required properties when employed for the Stokes equations. However, due to computational convenience and efficiency in practice, some mixed finite element pairs which do not satisfy the inf-sup condition are also popular. Thus, much attention has been paid to the study of the stabilized method for the Stokes problem.
Recent studies have focused on stabilization techniques, which include penalty [16–18], regular [19], multiscale enrichment [20], and local Gauss integration method [21, 22]. There exist a lot of theoretical results for the stabilized mixed finite element methods for the Stokes equations, and the comparisons between them are also given (see [17, 19, 20, 22–24], and the references cited therein). In this paper, we mainly focus on the Stokes eigenvalue problem solved by these stabilized finite element methods based on the lowest equal-order finite element space pair. Moreover, we present the comparisons between these methods for the considered problem.
The remainder of this paper is organized as follows. In the Section 2, we introduce the studied Stokes eigenvalue problem, the notations, and some well-known results used throughout this paper. Then several stabilized mixed finite element methods are reviewed, and their key stabilization techniques are recalled in Section 3. Comparisons between these stabilized methods are performed numerically in Section 4. Finally, we end with some short conclusions in Section 5.
2. Preliminaries
Let Ω be a bounded, convex, and open subset of ℝ2 with a Lipschitz continuous boundary ∂Ω. We consider the Stokes eigenvalue problem as follows. Find (u,p;λ), such that-νΔu+∇p=λuinΩ,div u=0inΩ,u=0on∂Ω,
where u=(u1(x),u2(x)) represents the velocity vector, p=p(x) the pressure, ν>0 the viscosity, and λ∈ℝ the eigenvalue.
We will introduce the following Hilbert spaces:
X=H01(Ω)2,Y=L2(Ω)2,M=L02(Ω)={q∈L2(Ω):∫Ωqdx=0}.
The spaces L2(Ω)m, m=1,2, are equipped with the L2-scalar product (·,·) and L2-norm ∥·∥L2 or ∥·∥0. The space X is endowed with the usual scalar product (∇u,∇v) and the norm ∥∇u∥0. Standard definitions are used for the Sobolev spaces Wm,p(Ω), with the norm ∥·∥m,p, m,p≥0. We will write Hm(Ω) for Wm,2(Ω) and ∥·∥m for ∥·∥m,2.
We define the continuous bilinear forms a(·,·) and d(·,·) on X×X and X×M, respectively, by
a(u,v)=ν(∇u,∇v),∀u,v∈X,d(v,q)=(q,divv),∀v∈X,∀q∈M,
and a generalized bilinear form B((·,·);(·,·)) on (X×M)×(X×M) by B((u,p);(v,q))=a(u,v)-d(v,p)-d(u,q),∀(u,p),(v,q)∈X×M.
With the above notations, the variational formulation of problem (2.1) reads as follows. Find (u,p;λ)∈(X×M)×ℝ with ∥u∥0=1, such that for all (v,q)∈X×M,B((u,p);(v,q))=λ(u,v).
Moreover, the bilinear form d(·,·) satisfies the inf-sup condition [25] for all q∈M:supv∈X|d(v,q)|‖∇v‖0≥β1‖q‖0,
where β1 is a positive constant depending only on Ω. Therefore, the generalized bilinear form B((·,·);(·,·)) satisfies the continuity property and inf-sup condition: |B((u,p);(v,q))|≤c(ν‖∇u‖0+‖p‖0)(‖∇v‖0+‖q‖0),∀(u,p),(v,q)∈X×M,sup(v,q)∈X×M|B((u,p);(v,q))|‖∇v‖0+‖q‖0≥β2(ν‖∇u‖0+‖p‖0),∀(u,p)∈X×M,
where c and β2 are the positive constants depending only on Ω.
3. Stabilized Mixed Finite Element Methods
For h>0, we introduce the finite-dimensional subspaces Xh×Mh⊂X×M, which are characterized by Kh, a partitioning of Ω into triangles K with the mesh size h, assumed to be uniformly regular in the usual sense.
Then we define
Xh={u∈C0(Ω̅)2∩X:u|K∈P1(K)2,∀K∈Kh},Mh={q∈C0(Ω¯)∩M:q|K∈P1(K),∀K∈Kh},
where P1(K) represents the space of linear functions on K.
Remark 3.1 (Nonconforming finite element space).
Denote the boundary edge by Γj=∂Ω∩∂Kj and the interior boundary by Γjk=Γkj=∂Kj∩∂Kk. Set the centers of Γj and Γjk by ζj and ζjk, respectively. The nonconforming finite element space NCh for the velocity will be taken to be
NCh={v:vj=v∣Kj∈P1(Kj)2,vj(ζjk)=vk(ζkj),v(ζj)=0,Kj∈Kh,∀j,k},
where P1(Kj) is the set of all polynomials on Kj of degree less than 1. Note that the nonconforming finite element space NCh is not a subspace of X. In this nonconforming case, the pair of finite element spaces is NCh×Mh [26]; that is, the conforming space is still used for pressure.
Moreover, the discrete mixed finite element formulation for the Stokes eigenvalue problem reads: find (uh,ph;λh)∈(Xh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈Xh×Mh,B((uh,ph);(v,q))=λh(uh,v).
Next, we denote by U the array of the velocity and by P the array of the pressure. It is easy to see that (3.3) can be written in matrix form:[ABBT0][UP]=λh[E000][UP],
where the matrices A, B, and E are deduced in the usual manner, using the bases for Xh and Mh, from the bilinear forms a(·,·), d(·,·), and (·,·), respectively, and BT is the transpose of matrix B.
Remark 3.2.
In the nonconforming case, we define the bilinear forms
a′(u,v)=ν∑K∈Kh(∇u,∇v)K,d′(v,p)=∑K∈Kh(divv,q)K,
and a generalized bilinear form B′((·,·);(·,·)) on (X×M)×(X×M) by
B′((u,p);(v,q))=a′(u,v)-d′(v,p)-d′(u,q),∀(u,p),(v,q)∈X×M.
Accordingly, the discrete nonconforming formulation for the Stokes eigenvalue problem is: find (uh,ph;λh)∈(NCh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈NCh×Mh, B′((uh,ph);(v,q))=λh(uh,v).
Note that the lowest equal-order pair does not satisfy the discrete inf-sup condition:
supvh∈Xhd(vh,qh)‖∇vh‖0≥β3‖qh‖0,orsupvh∈NChd′(vh,qh)‖∇vh‖0,h≥β3‖qh‖0,∀qh∈Mh,
where the constant β3>0 is independent of h, and ∥∇vh∥0,h is the discrete energy seminorm in the nonconforming case. In order to fulfill this condition, several ways have been used to stabilize the lowest equal-order finite element space pair.
Method 1 (Penalty method).
Find (uh,ph;λh)∈(Xh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈Xh×Mh,
B((uh,ph);(v,q))+εν(ph,q)=λh(uh,v),
where ε>0 is a penalty parameter. The performance of this method obviously depends on the choice of the penalty parameter ε. Then the matrix form of (3.9) can be expressed as
[ABBTενD][UP]=λh[E000][UP],
where the matrix D is deduced in the usual manner, using the base for Mh, from (ph,q). Because the coefficient matrix of (3.10) is usually large and sparse, it is not easy to compute the numerical solution using a direct method. In general, one can use the Uzawa algorithm. By some simple calculation, we get
Ei+1P=(I-τ(BT(A-λhE)-1B-ενD))EiP,
where EiP=P-Pi and τ is a positive constant.
Method 2 (Regular method).
Find (uh,ph;λh)∈(Xh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈Xh×Mh,
B((uh,ph);(v,q))-δ∑K∈Kh(∇ph-λhuh,∇q)K=λh(uh,v),
where δ=h2/(αν) is a stabilization parameter and α>0. The regular method uses a simple way to stabilize the mixed finite element approximation without a loss of accuracy. In fact, it can be treated the regular method as a special Douglas-Wang's scheme [19]. Then the matrix form of this stabilized version can be expressed as
[ABBT-δD1][UP]=λh[E0F0][UP],
where additional blocks D1 and F correspond to the following respective terms:
∑K∈Kh(∇ph,∇q)K,-δ∑K∈Kh(uh,∇q)K.
As (3.11),we also have
Ei+1P=(I-τ((BT-λhF)(A-λhE)-1B+δD1))EiP.
Method 3 (Multiscale enrichment method).
Find (uh,ph;λh)∈(Xh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈Xh×Mh,
B((uh,ph);(v,q))-δ1∑K∈Kh(∇ph-λhuh,∇q)K+δ2∑e=∂Kj∩∂Kk([ν∂nuh],[ν∂nv])e=λh(uh,v),
where δ1=h2/(α1ν), δ2=h/(α2ν)are the stabilization parameters, α1,α2>0, n is the normal outward vector, ∂n is normal derivative operator, and [v] denotes the jump of v across e. This stabilized method includes the usual Galerkin least squares stabilized terms on each finite element and positive jump terms at interelement boundaries. Moreover, a direct algebraic manipulation leads to the matrix form
[ABBT+δ2D2-δ1D1][UP]=λh[E0F0][UP],
where the matrix D2 is deduced in the usual manner, using the bases for Xh, from the term ∑e=∂Kj∩∂Kk([ν∂nuh],[ν∂nv])e. As (3.11), we have
Ei+1P=(I-τ((BT+δ2D2-λhF)(A-λhE)-1B+δD1))EiP.
Method 4 (Local Gauss integration method).
Find (uh,ph;λh)∈(Xh×Mh)×ℝ with ∥uh∥0=1, such that for all (v,q)∈Xh×Mh,
B((uh,ph);(v,q))-G(ph,q)=λh(uh,v),
where G(ph,q) is defined by
G(ph,q)=∑K∈Kh{∫K,2phqdx-∫K,1phqdx},∀ph,q∈Mh,
where ∫K,ig(x)dx indicates a local Gauss integral over K that is exact for polynomials of degree i, i=1,2. In particular, the trial function ph∈Mh must be projected to piecewise constant space when i=1. This stabilization technique is free of stabilization parameters and does not require any calculation of high-order derivatives, a specification of any mesh-dependent parameter or edge-based data structures. Then the corresponding matrix form of this stabilized method is
[ABBT-G][UP]=λh[E000][UP],
where the matrix G is deduced in the usual manner, using the bases for Mh, from the term G(ph,q). As (3.11), we get
Ei+1P=(I-τ(BT(A-λhE)-1B+G))EiP.
Remark 3.3.
It is well known that if τ is well chosen, then Ui and Pi converge, respectively, to U and P with a rate of convergence based on (3.11), (3.15), (3.18), and (3.22). From these equations, we can find that Method 1 converges faster than Method 4. Methods 2 and 3, whose coefficient matrices are not symmetric, may cost more time to converge.
Remark 3.4.
By using the regularity assumptions and well-established techniques for eigenvalue approximation [1, 8, 10, 12], the theoretical convergence rates should be of order O(h2) and O(h) for the velocity in the L2- and H1-norms, respectively, of order O(h) for the pressure in the L2-norm, and of order O(h2) for the eigenvalue by using all these stabilized methods.
4. Numerical Experiments
In this section we numerically compare the performance of the various stabilized mixed finite element methods discussed in the previous section. In all experiments, the algorithms are implemented using public domain finite element software [27] with some of our additional codes.
Let the computation be carried out in the region Ω={(x,y)∣0<x,y<1}. We consider the Stokes eigenvalue problem in the case of the viscidity ν=1, and it will be numerically solved by the stabilized mixed methods on uniform mesh (see Figure 1). Here, we just consider the first eigenvalue of the Stokes eigenvalue problem for the sake of simplicity. The exact solution of this problem is unknown. Thus, we take the numerical solution by the standard Galerkin method (P2-P1 element) computed on a very fine mesh (6742 grid points) as the “exact” solution for the purpose of comparison. Here, we take λ=52.3447 as the first exact eigenvalue.
Uniform finite element partitioning of the unit square.
As we know, the stabilized term of the regular and multiscale enrichment methods must be controlled by carefully designed stabilization parameters (i.e., δ,δ1,δ2), whose optimal values are often unknown. Hence, in Figures 2 and 3 we show the effect on the error of varying δ, δ1, and δ2 on a fixed mesh h=1/64 for the regular and multiscale enrichment approximations, respectively. An interesting thing can be observed that these two methods have an analogous convergence pattern with respect to the parameter δ and δ1. Because they involve a similar stabilization term with respect to these two parameters, we note that the errors can become large with some values of the stabilization parameters.
Effect of varying δ for the regular method.
Effect of varying δ1 (a) and δ2 (b) for the multiscale enrichment method.
Results gotten from the penalty, regular, multiscale enrichment, local Gauss integration, and nonconforming local Gauss integration methods are presented in Tables 1–5, respectively. Here, we choose ε=1.0e-5, α=8, α1=8, and α2=12. Because they can deal with the considered problem well. From Tables 1, 2, 3, 4, and 5, we can see that these methods work well and keep the convergence rates just like the theoretical analysis except the multiscale enrichment method. Meanwhile, it can be seen that the penalty method requires the least CPU-time, which validates the analysis in Remark 3.3. As expected, we have an interesting observation that the error of the nonconforming local Gauss integration method is better than the conforming version, which is not surprising since the degrees of freedom of the nonconforming method are nearly three times than that of conforming one on uniform mesh (see Figure 1). Hence, it is natural that the nonconforming local Gauss integration method is more accurate and costs more CPU-time.
Results get from the penalty method with ε=1.0e-5.
1/h
CPU-time
λh
|λh-λ|/λ
Rate
8
0.016
60.2628
0.151269
—
16
0.094
54.1688
0.0348472
2.1180
24
0.234
53.1426
0.015243
2.0393
32
0.437
52.7909
0.00852367
2.0205
40
0.751
52.6295
0.00544061
2.0120
48
1.156
52.5433
0.00377371
2.0065
56
1.734
52.4898
0.00277154
2.0023
64
2.437
52.4558
0.00212237
1.9986
Results get from the regular method with α=8.
1/h
CPU-time
λh
|λh-λ|/λ
Rate
8
0.031
56.7283
0.0837442
—
16
0.109
53.4803
0.0216951
1.9486
24
0.266
52.8531
0.00971199
1.9822
32
0.515
52.6314
0.00547673
1.9913
40
0.861
52.5284
0.00350903
1.9950
48
1.313
52.4723
0.00243822
1.9969
56
1.953
52.4385
0.00179192
1.9979
64
2.719
52.4165
0.00137219
1.9986
Results get from the multiscale enrichment method with α1=8 and α2=12.
1/h
CPU-time
λh
|λh-λ|/λ
Rate
8
0.063
67.8237
0.295713
—
16
0.281
57.1062
0.0909634
1.7008
24
0.656
54.7378
0.0457175
1.6968
32
1.265
53.8304
0.0283839
1.6569
40
2.156
53.3813
0.0198027
1.6133
48
3.359
53.1229
0.0148665
1.5725
56
5.125
52.9588
0.0117327
1.5357
64
7.187
52.8472
0.00959935
1.5029
Results get from the local Gauss integration method.
1/h
CPU-time
λh
|λh-λ|/λ
Rate
8
0.031
57.3951
0.096482
—
16
0.109
53.6201
0.024366
1.9854
24
0.251
52.9119
0.0108368
1.9983
32
0.469
52.6638
0.00609553
2.0001
40
0.797
52.5489
0.00390065
2.0006
48
1.234
52.4865
0.00270843
2.0007
56
1.844
52.4488
0.00198963
2.0008
64
2.547
52.4244
0.00152315
2.0008
Results get from the local Gauss integration method with the nonconforming element.
1/h
CPU-time
λh
|λh-λ|/λ
Rate
8
0.031
50.2121
0.0407434
—
16
0.125
51.7355
0.0116391
1.8076
24
0.344
52.0619
0.00540181
1.8932
32
0.672
52.1825
0.00309932
1.9311
40
1.156
52.2397
0.00200547
1.9508
48
1.859
52.2713
0.00140228
1.9624
56
2.797
52.2905
0.00103505
1.9698
64
4.001
52.3031
0.000795121
1.9749
Besides, to show the stability and efficiency of these methods for the considered problem, we present the velocity streamlines and the pressure contours with h=1/64 in Figures 5, 6, 7, 8, and 9. Meanwhile, we present the results by the standard Galerkin method (P2-P1 element) computed on a very fine mesh (6742 grid points) for the purpose of comparison in Figure 4. From Figures 5(a)–9(a), five resolved vortices are captured, which is consistent with that in Figure 4(a). For the pressure, the penalty method is divergent from Figure 5(b), although it costs the least time. From Figures 6(b)-9(b), the nonconforming local Gauss integration method shows the best numerical stability.
Velocity streamlines (a) and pressure level lines (b) for the standard Galerkin method (P2-P1 element).
Velocity streamlines (a) and pressure level lines (b) for the penalty method with ε=1.0e-5.
Velocity streamlines (a) and pressure level lines (b) for the regular method with α=8.
Velocity streamlines (a) and pressure level lines (b) for the multiscale enrichment method with α1=8 and α2=12.
Velocity streamlines (a) and pressure level lines (b) for the local Gauss integration method.
Velocity streamlines (a) and pressure level lines (b) for the local Gauss integration method with the nonconforming element.
5. Conclusions
In this paper we have presented several stabilized mixed finite element methods in solving the Stokes eigenvalue problem based on the lowest equal-order finite element space pair. By being compared numerically, we get some conclusions as follows.
The stability and efficiency of the penalty method depend on the penalty parameter. The smaller this parameter, the more stable the method. However, if this parameter is too small, the condition number of the system matrix arising from this method will become too large to solve.
The performance of the regular and multiscale enrichment method heavily depends on the choice of the stabilization parameters, which is a difficult task in reality. Meanwhile, a poor choice of these stabilization parameters can also lead to serious deterioration in the convergence rates.
The local integration method is free of stabilization parameters and shows numerically the best performance among the methods considered for the given problem.
From Tables 4 and 5, we have an interesting observation that the value of λh by conforming method becomes small to converge to the exact solution and the one by nonconforming method becomes large to converge to the exact solution. We hold that it can obtain more accurate numerical solution by Lagrange interpolation between conforming and nonconforming results based on the same degrees of freedom of these two methods. It may get superconvergence result on this triangular mesh.
Acknowledgments
The authors would like to thank the editor and reviewers for their valuable comments and suggestions which helped us improve the results of this paper. This work is in parts supported by the NSF of China (no. 10901131, no. 10971166), the National High Technology Research and Development Program of China (863 Program, no. 2009AA01A135), the China Postdoctoral Science Foundation (no. 200801448, no. 20070421155), and the Natural Science Foundation of Xinjiang Province (no. 2010211B04).
BabuškaI.OsbornJ. E.CiarletP. G.LionsJ. L.Eigenvalue problems1991IINorth-Holland, AmsterdamFinite Element Method641787BabuškaI.OsbornJ. E.Finite element-Galerkin approximation of the eigenvalues and eigenvectors of selfadjoint problems19895218627529710.2307/2008468962210ZBL0675.65108LinQ.XieH.Asymptotic error expansion and Richardson extrapolation of eigenvalue approximations for second order elliptic problems by the mixed finite element method20095981884189310.1016/j.apnum.2009.01.0112536089ZBL1201.65202LinQ.Fourth order eigenvalue approximation by extrapolation on domains with reentrant corners199158663164010.1007/BF013856451083525ZBL0695.65064JiaS.XieH.YinX.GaoS.Approximation and eigenvalue extrapolation of Stokes eigenvalue problem by nonconforming finite element methods2009541115247601810.1007/s10492-009-0001-0ZBL1212.65434ChenH.JiaS.XieH.Postprocessing and higher order convergence for the mixed finite element approximations of the eigenvalue problem201161461562910.1016/j.apnum.2010.12.0072754580ZBL1209.65126ChenH.JiaS.XieH.Postprocessing and higher order convergence for the mixed finite element approximations of the Stokes eigenvalue problems200954323725010.1007/s10492-009-0015-7253054110.1007/s10492-009-0015-7ZBL1212.65431LovadinaC.LylyM.StenbergR.A posteriori estimates for the Stokes eigenvalue problem200925124425710.1002/num.20342247368810.1002/num.20342ZBL1169.65109CliffeK. A.HallE.HoustonP.Adaptive discontinuous Galerkin methods for eigenvalue problems arising in incompressible fluid flows20103164607463210.1137/0807319182594995YinX.XieH.JiaS.GaoS.Asymptotic expansions and extrapolations of eigenvalues for the Stokes problem by mixed finite element methods20082151127141240062310.1016/j.cam.2007.03.028ZBL1149.65090ChenW.LinQ.Approximation of an eigenvalue problem associated with the Stokes problem by the stream function-vorticity-pressure method20065117388219732410.1007/s10492-006-0006-xZBL1164.65489MercierB.OsbornJ.RappazJ.RaviartP. A.Eigenvalue approximation by mixed and hybrid methods19813615442745310.2307/200765160650510.1090/S0025-5718-1981-0606505-9ZBL0472.65080XuJ. C.ZhouA. H.A two-grid discretization scheme for eigenvalue problems2009702331725ChenZ. X.2005Heidelberg, GermanySpringer2158541GiraultV.RaviartP. A.1987Berlin, GermanySpringerBrefortB.GhidagliaJ. M.TemamR.Attractors for the penalized Navier-Stokes equations200119112192454110.1137/0519001ZBL0696.35131HeY. N.Optimal error estimate of the penalty finite element method for the time-dependent Navier-Stokes equations2005742511201121610.1090/S0025-5718-05-01751-52136999ZBL1065.35025LiJ.MeiL. Q.HeY. N.A pressure-Poisson stabilized finite element method for the non-stationary Stokes equations to circumvent the inf-sup condition200618212435229201510.1016/j.amc.2006.03.030ZBL1119.65092DouglasJ.Jr.WangJ.An absolutely stabilized finite element method for the Stokes problem19895218649550810.2307/2008478 ArayaR.BarrenecheaG. R.ValentinF.Stabilized finite element methods based on multiscale enrichment for the stokes problem200644132234810.1137/050623176LiJ.HeY. N.ChenZ. X.A new stabilized finite element method for the transient Navier-Stokes equations20071971–4223510.1016/j.cma.2007.06.029LiJ.HeY. N.A stabilized finite element method based on two local Gauss integrations for the Stokes equations20082141586510.1016/j.cam.2007.02.015FengX. L.KimI.NamH.SheenD.Locally stabilized P1-nonconforming quadrilateral and hexahedral finite element methods for the Stokes equationsJournal of Computational and Applied Mathematics. In press10.1016/j.cam.2011.06.009LiJ.HeY. N.ChenZ. X.Performance of several stabilized finite element methods for the Stokes equations based on the lowest equal-order pairs2009861375110.1007/s00607-009-0064-5BrezziF.FortinM.199115New York, NY, USASpringerSpringer Series in Computational MathematicsLiJ.ChenZ. X.A new local stabilized nonconforming finite element method for the Stokes equations2008822-315717010.1007/s00607-008-0001-zFreeFem++version 2.23, http://www.freefem.org/