Veried eigenvalue evaluation for Laplace operator on arbitrary polygonal domain

/CREST,JST (Xuefeng Liu) (Shinichi Oishi) Faculty of Science and Engineering, Waseda University

Abstract For eigenvalue problem of Laplace operator over polygonal domain ( R2 ) of arbitrary shape, we proposed an algorithm based on nite element method to bound the leading eigenvalues with indices guaranteed. The algorithm is developed by well use of max-min and min-max principles and newly constructed a priori error estimation for FEM solution. The eciency of the algorithm is demonstrated by several computation examples.

Introduction

The eigenvalue problem of Laplacian has been well investigated in history from various viewpoints. Here, we pay attention to giving accurate bounds for eigenvalues with indices guaranteed. For such a purpose, Lehmann-Goerisch method is well known as a eective way to give sharp bounds for eigenvalues once a quantity satisfying k < k+1 is available, where k denotes the eigenvalues with increasing order on magnitude. To nd such a is not an easy work. In [8], M. Plum developed homotopy method based on operator comparison theorem to give a computable . As base problem with explicit eigenvalues is necessary when we apply the homotopy method, domain mapping is used to construct the base problem, which brings diculties to solving problems over general domain. In this paper, starting from an early work of Birkho, de Boor, Swartz and Wendro [2], we propose a new method to give guaranteed estimation for leading k-th eigenvalue on arbitrary polygonal domain, where the nite element method(FEM) is used to give approximate eigenvalues with computable error bounds. The estimate of k obtained by using sparse domain triangulation will be relatively rough, but it can work as a good candidate of mentioned above. Thus, if needed, the bounds can be sharpened by further applying Lehmann-Goerisch method. The method proposed here can deal with three types homogeneous boundary conditions associated with function space: Dirichlet, Neumann and mixed one. To explain the method in a concise way, we only show details for the Dirichlet case. Also, it is possible to extend the method for general elliptic problem. At the end of this paper, we show computation results on triangle and L-shaped domain.

Preliminaries

Let be polygonal domain with arbitrary shape, convex or non-convex. We in1 troduce function space V = H0 () = {v H 1 ()|v = 0 on }. The notation v L2 denotes the L2 norm of v L2 and |v|H k () (k = 1, 2) the semi-norms in H k (). Let

T h be one triangularization of , which has polygon boundary. The variational form of eigenvalue problem is dened as below: Find R and u V s.t. (u, v) = (u, v), v V. (1)

The classical continuous piecewise linear nite element(FE) space Vh V will be used as approximation space. The Ritz method is to solve the variational problem in Vh , Find h R and uh Vh s.t. (uh , vh ) = h (uh , vh ), vh Vh . (2) Supposing the bases of Vh to be {i }n , the problem of (2) is in fact a generalized i=1 matrix eigenvalue problem:
h Ah x = h B h , where Ah = (i , j )L2 , Bi,j = (i , j )L2 . i,j

(3)

The eigenvalues h can be evaluated accurately by applying veried computations, c.f., k [1, 7, 9]. Denote by {i , ui } (resp. {h , uh } ) the eigenpairs of (1) (resp. (2) ) with i i eigenfunction being orthogonally normalized under L2 -norm. These eigenpairs are just the stationary values and critical points of Rayleigh quotient on space V ( resp. Vh ): R(u) := (u, u)/(u, u) . (4)

Since an upper bound for i as i h is easy to obtain from min-max principle, we i will pay attention to nd satisfactory lower bounds for eigenvalues. The eigenfunction estimation will not be discussed here. Lets introduce two constants Ci,h (i = 0, 1) to be used later, which are related to function interpolations i (i = 0, 1) over triangle element K. For u L2 (K), 0 u is constant function s.t. 0 u u(x) dx/ 1 dx , (5)
K K

and for u H (K), 1 u is linear function s.t.

(1 u) (x) = u(x) on each vertex of K .

(6)

Global interpolations 0,h and 1,h are just the extension of 0 and 1 . Dene h by the mesh size and C0,h and C1,h the constants over triangulation Th , Ci,h := max Ci (K)/h (i = 0, 1) ,
KT h

(7)

where C0 (K) := sup

|0 u u|L2 /|u|H 1 ,

C1 (K) :=

sup

vH 2 (K)\{0}

|1 u u|H 1 /|u|H 2 .

Lower bound of eigenvalues by adopting minmax and max-min principle

In this section, we will introduce two methods to give lower bound for eigenvalues, all of them adopting computable a priori estimate of Ritz-Galerkin solution of Poissons 1 problem. Let u H0 () be the solution of following variation problem,
1 (u, v) = (f, v) v H0 () .

(8)

1 The solution u H0 (), in meaning of distribution, satises the partial dierential equation u = f . Whether u belongs to H 2 () or not depends on the domain shape. Let Ph be the orthogonal projection of u V into Vh ,

(u Ph u, vh ) = 0 vh Vh . We will deduce a computable a apriori error estimate in the form as below, |u Ph u|H 1 M f
L2 ,

(9)

u Ph u

L2

M |u Ph u|H 1 M 2 f

L2

(10)

where M is quantity to be evaluated in Section 4. In the following, we will introduce two methods to bound eigenvalues based on this a priori error estimation.

3.1

Birkho s method: application of Min-Max principle

Birkho, de Boor, Swartz and Wendro [2] considered eigenvalue problem in form of Rayleigh quotient R(u) := N (u)/D(u) , where N (u) and D(u) are quadratic forms of u V and D(u) > 0 for u = 0. Suppose {k , uk } ( resp. {k , uh }) the stationary h k values and critical points of R(u) on space V (resp. Vh ), with increasing order on k (resp. k ). Birkho et al deduced an estimate for h by applying Min-Max principle: h k h h h h Theorem 1. Given any v1 , v2 , , vk Vh ( V ) satisfying k D(vi ui ) < 1, we i=1 have, for k 1, ( k ) / ( k )1/2 2 h h 1 k k k + N (vi ui ) D(vi ui ) (11) h
i=1 i=1

For model problem of Laplacian in (1), N (u) = (u, u) and D(u) = (u, u). It h is natural to select vi = Ph ui (i = 1, , k) for each eigenfunction ui , and apply the error estimate of (10): |ui Ph ui |H 1 M ui L2 = M i ui L2 = M i . ui Ph ui L2 M |ui Ph ui |H 1 M 2 i . Thus, we obtain an a priori estimate for h . k Theorem 2. Let k and h be the ones dened in Section 2. If 1M 2 ( k we have /( )2 k k h k + M 2 2 1 M 2( 2 )1/2 . k i i
i=1 i=1

k
i=1

2 )1/2 > 0, i (12)

Dene function 1 on variable tk with parameters {t1 , , tk1 } as below, /( )2 k k 2 2 2 2 1/2 1 (tk ; t1 , , tk1 ) := tk + M ti 1M ( ti )
i=1 i=1

(13)

Noticing that 1 is increasing as varible tk increases, 1 (tk ) has increasing inverse function. Therefore, 1 (h ; 1 , , k1 ) k . As i h for i 1, we can further 1 i k see 1 (h ; h , , h ) k . k1 1 k 1 3

 Remark 3.1. In practical computation, instead of verifying 1M 2 ( k 2 )1/2 > 0, we i=1 i k 2 h 2 1/2 h will check a stronger condition 1M ( i=1 i ) > 0 since i ( i )s are computable ones. Remark 3.2. Birkho, de Boor, Swartz and Wendro [2] obtained the estimation of form (11) with Vh the space constructed by spline functions. By applying the error estimate for spline interpolation, quantitative estimate for eigenvalue problem of 1D Sturm-Liouville system is successfully done. However, it is dicult to apply their method to solve problem on general 2D domain. As a comparison, our estimation in Theorem 2 and 3 can deal with eigenvalue problem with domain of general shape, which inherits advantages from the nite element method.

3.2

Bounding eigenvalues by adopting Max-Min principle

h h h h Theorem 3 (Liu). Let v1 , , vk1 be arbitrary functions of Vh and Vk1 := span{v1 , , vk1 }. Dene k by Rayleigh quotient on Vh Vk1 ( Vk1 : complement space of Vk1 in V )

k =

vh Vh Vk1

min

(vh , vh ) . (vh , vh )

Then, an a posteriori estimate for k is available, ( )2 / ( ) k k M k 1 + M 2 k , where M is the one in (10). Proof. From Max-Min principle, we have k = max W V,dim(W )k1
vW

(14)

min

(v, v) . (v, v)

h h Thus, for specied Vk1 := span{v1 , , vk1 }, a lower bound for k is given as

k min

vVk1

(v, v) . (v, v)

(15)

For any v Vk1 , Ph v Vh . Let wh be arbitrary one in Vk1 ( Vh ). Then (v, wh ) = 0. Thus (Ph v, wh ) = (v, wh ) = 0, which implies that Ph v Vh Vk1 . Considering (10) and the denition of k ,

v v Hence,
2 L2

L2

Ph v L2 + v Ph v L2 k Ph v L2 + M (v Ph v) L2 . ) ) ( ( 1 + M 2 ( Ph v 2 2 + (v Ph v) 2 2 ) = 1 + M 2 v 2 2 . L L L k k v
2 L2 /

1/2

2 L2

( ) k / 1 + M 2 k for any v Vk1 . ) ( k k / 1 + M 2 k . (16)

The equation (15) tells us,

Now, it is trivious to formulate the result in (14). 4

Remark 3.3. The subspace Vk1 in Theorem 3 can be taken as the one spanned by rst k 1 eigenfunction of (2). In this case, k = h and lower bound of k is given k as: ( ) h / 1 + M 2 h k h . (17) k k k It is obvious that the above estimate based on Max-Min principle gives better estimate than the one of (11).

A prioiri error estimate for Ritz-Galerkin solution of Poissons problem

The following section will be devoted to evaluating M appearing in a priori error estimate (10) for projection Ph . The discussion will be divided into two parts, the one with regular solution on convex domain and the one with singular solution on non-convex domain.

4.1

Convex domain

First, we quote a well known result on a priori estimation for Laplacian. Lemma 4. [4] Assume is bounded convex polygonal domain in R2 . For u H 2 () 1 H0 () or u H 2 () and u/n = 0 on , let f := u. Then, we have |u|H 2 u
L2

= f

L2

Theorem 5. Let be convex polygonal domain and u be the solution of (8). The error estimate for (u Ph u) is given as |u Ph u|H 1 C1,h h f
L2 ,

u Ph u

L2

2 C1,h h|u Ph u|H 1 C1,h h2 f

L2

Thus, we can take M := C1,h h under current assumptions. Proof. Under the given assumptions, the solution u belongs to H 2 (). By using interpolation error estimate for 1,h and the Lemma 4, we have, |u Ph u|H 1 |u 1,h u|H 1 C1,h h|u|H 2 C1,h h f
L2

(18)

where the constant C1,h is the one dened in (7). The L2 -norm error estimation can be easily done by adopting Aubin-Nitsues technique.

4.2

Non-convex domain

To deal with problem on non-convex domain, which has singular solution not belonging to H 2 (), we adopt the hypercircle equation to deduce a computable a priori error estimate. Let W h be the lowest order Raviart-Thomas FEM space over domain triangulation T h and M h the space of piecewise constant. Also, dene subspace of Wh for h fh in M h , Wfh := {ph W h | div ph = fh }. Recall the denition of 0,h : L2 () M h in Section 2, (u 0,h u, vh ) = 0, vh M h . 5

From the denition, we have u

2 L2

= 0,h u

2 L2

+ u 0,h u

2 L2

and

u 0,h u

L2

C0,h h|u|H 1

if u H 1 .

where C0,h is the constant dened in (7). Lets introduce a computable quantity over nite dimensional spaces: :=
fh M h \{0}

max

vh Vh ph W h f

min

min

ph vh

L2 /

fh

L2

(19)

Lemma 6. Given fh M h , let u H 1 and uh Vh ( V ) be the solutions of varia tional problems, (, v) = (fh , v), u v V . (h , vh ) = (fh , vh ), u vh Vh , (20)

respectively. Then we have a computable error estimate as below: | uh |H 1 fh u

(21)

h Proof. From Prager-Synges theorem, we have, for u in (20) and any vh Vh , ph Wfh , such a hypercircle equation holds,

vh u Thus, vh u
L2

2 L2

+ ph u
L2 ,

2 L2

= ph vh

2 L2

(22) (23)

vh ph

h vh Vh , ph Wfh .

From minimization principle and the denition of , we obtain h u u

min min

vh Vh ph W h f

ph vh

L2

fh

L2

(24)

Theorem 7. For any f L2 (), let u V and uh Vh be solution of variational problems v V, (uh , vh ) = (f, vh ), vh Vh . (25) 2 respectively. Introduce quantity M := C0,h h2 + 2 , where C0,h is the constant dened in (7). Then, we have, |u uh |H 1 M f
L2 ,

(u, v) = (f, v),

u uh

L2

M2 f

L2

(26)

Remark 4.1. The quantity M , independent of f , will decrease when mesh is rened. By theoretical analysis, we can show that M tends to 0 in the same order as the error of linear conforming FEM solution solution. Proof. We follow analogous framework with Kikuchi and Saito [6] to nish the proof. Let u and uh be the ones dened in Theorem 6 with fh = 0,h f . The minimization principle leads to |u uh |H 1 |u uh |H 1 . Decomposing u uh by (u u) + ( uh ), u we have |u uh |H 1 |u uh |H 1 |u u|H 1 + | uh |H 1 . u (27) 6

From the denitions of u and u, we have, for any v V ((u u), v) = (f 0,h f, v) = ((I 0,h )f, (I 0,h )v) . Taking v to be u u and applying the error estimate of interpolation (I 0,h )v, |u u|H 1 C0,h h (I 0,h )f Substitute (21) and (28) into (27), |u uh |H 1 C0,h h (I 0,h )f The estimate for u uh
L2 L2 L2

(28)

+ 0,h f

L2

2 C0,h h2 + 2 f

L2

can be easily done by applying Aubin-Nitsches method.

Computation

The computation of quantity turns to solving eigenvalue problem of matrix, for which we omit the details but point out that evaluation of consumes most of the total computation time. To obtain accurate numerical result, we adopt interval computation arithmetic to do the oating-point computation. The total framework is as below 1) Triangulate the domain and construct nite element space Vh . 2) Solve eigenvalues problem Ah x = h B h x under the bases of Vh . 3) Evaluate quantity M for the mesh and domain. 4) Calculate the lower and upper bounds of k by using (12) or (17) . In the following we display computation examples on several domains.

5.1

Triangle domain

In case of unit isosoceles right triangle domain, due to the symmetry of specied triangle domain, we can apply reecting techniques, e.g.,[5], to obtain the explicit eigenpairs as below: { = m2 + n2 , u = sin mx sin ny sin nx sin my}m>n1 . To compare the eciency of the methods basing on Max-min principle and the Min-max principle, we display the estimates of (12) and (17) in Table 1.

5.2

Computation Results on L-shaped domain

The domain is taken as = [0, 2][0, 2]\[1, 2][1, 2]. As a model problem, it has been well explored by many people, e.g., L. Fox, P. Henrici and C. Moler [3]. However, to the authors knowledge, most of the results are given only in the sense of approximate computation. Although our method gives a relatively rough evaluation, it can easily deal with more general domain and the result works as mathematically correct with indices guaranteed. 7

200 + + 100 + +

i 1 2 3 4 5

approx. 49.348 98.696 128.305 167.783 197.392

lower (12) 48.955 96.532 122.196 154.763 174.176

lower (17) 48.976 97.331 125.853 163.694 192.372

upper 49.553 99.633 129.729 170.312 201.577

150

Birkho(Minmax) Liu-Oishi(Maxmin) Exact value Upper Bound

50+

Table 1: Estimates by Min-max(12) and Max-min (17) principle. (h = 1/32)

i 1 2 3 4 5

h = 1/32 h = 1/64 lower upper lower upper 2 5 49.348 48.976 49.553 49.254 49.400 10 2 98.696 97.331 99.633 98.352 98.930 13 2 128.305 125.853 129.729 127.687 128.662 17 2 167.783 163.694 170.312 166.746 168.413 20 2 197.392 192.372 201.577 196.129 198.439 exact value

Table 2: Eigenvalue estimates for Laplacian on triangle domain (Dirichlet b.d.c.)

In Table 4, we list the rst 5 eigenvalues given by [3] and the veried bounds by h our proposed method. The values of , C0 and M , which are only depending on the mesh, are displayed in Table 3. We can see M tends to zero in order less than 1. Once lower bound for 5 is available, we can further apply the Lehmann-Goerisch method to obtain more precise bound for 1 , , 4 . Such a computation, although not veried, has been reported by Yuan and He [10] with very sharp bounds, while the lower bound for 5 is obtained in a dierent way.

Conclusion

For the classical eigenvalue problems of Laplace operator over 2-dimensional domain, we have proposed a novel and robust method to give accurate lower and upper bounds for eigenvalues. The method can deal with both convex and non-convex domains with general shape. To apply the Lehmann-Goerisch method for purpose of high precision, we still need pay eorts on constructing base functions over general domain. Acknowledgement The authors would like show great appreciation to Prof. M. Plum and Prof. K. Nagatou for fruitful discussion.

(0, 2)

(1, 2)

h=

1 4

(0, 1)

(2, 1)

h 1/4 1/8 1/16 1/32

0.1466 0.0882 0.0538 0.0332

h C0 0.080 0.040 0.020 0.010

M order of M 0.1668 0.0968 0.786 0.0574 0.754 0.0348 0.722

(0, 0)

(1, 0)

(2, 0)

Table 3: Uniform mesh of L-shaped domain and values i 1 2 3 4 5 lower bound approximate upper bound 9.5585 9.6397 9.6698 14.950 15.361 15.225 19.326 19.739 19.787 28.605 29.521 29.626 30.866 31.913 32.058 relative error 0.012 0.018 0.024 0.035 0.038

Table 4: Eigenvalue evaluation for L-shaped domain (h = 1/32)

References
[1] H. Behnke. The calculation of guaranteed bounds for eigenvalues using complementary variational principles. Springer-Verlag, 47(1):1127, 1991. [2] G. Birkho, C. De Boor, B. Swartz, and B. Wendro. Rayleigh-ritz approximation by piecewise cubic polynomials. SIAM Journal on Numerical Analysis, pages 188203, 1966. [3] L. Fox, P. Henrici, and C. Moler. Approximations and bounds for eigenvalues of elliptic operators. SIAM Journal on Numerical Analysis, 4(1):89102, 1967. [4] P. Grisvard. Elliptic problems in nonsmooth domains. Pitman, 1985. [5] F. Kikuchi and X. Liu. Determination of the Babuka-Aziz constant for the linear triangular s nite element. Japan Journal of Industrial and Applied Mathematics, 23:7582, 2006. [6] F. Kikuchi and H. Saito. Remarks on a posteriori error estimation for nite element solutions. Journal of Computational and Applied Mathematics, 199:329336, 2007. [7] K. Maruyama, T. Ogita, Y. Nakaya, and S. Oishi. Numerical inclusion method for all eigenvalues of real symmetric denite generalized eigenvalue problem. IEICE Trans, J97-A(8):11111119, 2004. [8] M. Plum. Bounds for eigenvalues of second-order elliptic dierential operators. The Journal of Applied Mathematics and Physics(ZAMP), 42(6):848863, 1991. [9] N. Yamamoto. A simple method for error bounds of eigenvalues of symmetric matrices. Linear Algebra and its Applications, 324(1-3):227234, 2001. [10] Q. Yuan and Z. He. Bounds to eigenvalues of the laplacian on L-shaped domain by variational methods. Journal of Computational and Applied Mathematics, 233(4):10831090, 2009.