Nedelec elements for computational electromagnetics

Per Jacobsson, June 5, 2007

Maxwells equations
E = jH H = jE + J (E) = (H) = 0 J = j (1) (2) (3) (4) (5)

Only three of the equations are needed, (1), (2) and (3) or (1), (2) and (5). Combining (1) and (2) gives the wave equation (1 E) 2 E = j J for E and similarly for H. E, J H (curl; ) corresponds to nite energy solutions, where H (curl; ) = {u [L2 ()]3 : u [L2 ()]3 } in three dimensions. The norm associated with H (curl; ) is dened as ||v||H (curl;) = (||v||[L2 ()]3 + || v||[L2 ()]3 )1/2 Some other vector spaces used: H 1 () = {v L2 () : v [L2 ()]3 }
1 H0 ()

(6)

(7)

(8)

(9) (10) (11) (12) (13) (14)

= {v H () : v = 0 on } H0 (curl; ) = {v H (curl; ) : v = 0} H (div; , ) = {v [L2 ()]3 : (v) L2 ()} H0 (div; , ) = {v [L ()] : (v) = 0} V = H0 (curl; ) H (div; , )
2 3

where is the outward normal to .

Problems when using nodal elements

When using nodal elements spurious solutions can occur and we can also get convergence to solutions of a problem dierent from our original problem. As an example we look at a simple cavity eigenvalue problem with perfectly conducting walls (1 E) = 2 E (E) = 0 E=0 1 in in on (15) (16) (17)

The weak form of this eigenvalue problem is to nd R, E H0 (curl; ) H0 (div; , ), E = 0 such that (1 E, v) = 2 (E, v) v H0 (curl; ) H0 (div; , ). (18)

We see that 2 = 0 is not an eigenvalue to (18) since that would imply E = 0. This means that we cannot have a physical mode with non-zero energy and = 0, the only physical modes are the ones with > 0. For this case, (15) implies (16) and we can write the problem on weak form as Find R, E H0 (curl; ), E = 0 s.t. (1 E, v) = 2 (E, v) v H0 (curl; ). (19)

This adds an innite-dimensional space of non-physical eigensolutions that coincides with 1 H0 () with zero eigenvalues. It is important to recognize these non-physical solutions so that they can be discarded. When using FE we seek solutions in a nite-dimensional subspace Vh of H0 (curl; ): Find h R, Eh Vh , Eh = 0 s.t. (1 Eh , v) = h (Eh , v) v Vh . (20)

For nodal elements the approximation space is Vh = [Pk ]3 H0 (curl; ). When using nodal based FE on an unstructured grid, we obtain many spurious eigenvalues approaching zero and it can be hard to recognize these so that they can be discarded. For certain structured grids, the zero eigenvalues can be well approximated but other spurious solutions appear in the spectrum. Since these spurious eigenvalues can be distributed all over the spectrum, they are hard to distinguish from the physical eigenvalues. For non-convex domains, it has been proved that the discrete eigenvalues computed using nodal elements converge to the eigenvalues of a problem dierent than the original problem. One possible explanation for the failure of nodal based FE in the case of non-convex domains in two dimensions is that the nodal elements are constrained in [H 1 ()]2 , which is a proper and closed subset of V . This means that there are eigenfunctions in V that cannot be approximated by nodal elements in [H 1 ()]2 . Another important drawback of nodal-based FE when solving Maxwells equations is the inability to model eld singularities at conducting corners and tips. The boundary condition for the electric eld at a conducting surface is E = 0, but enforcing this boundary condition at each node also enforces normal continuity, which is not desired.

Nedelecs edge elements

We start by introducing a subspace Sk of homogeneous vector polynomials of degree k in d dimensions by k ]d : x p = 0}, Sk = {p [P (21) k is the space of homogeneous polynomials of order k . We can now introduce the space where P Rk which is needed for Nedelecs vector elements, Rk = [Pk1 ]d Sk . (22)

The dimension of the space Rk in three dimensions is dim(Rk ) = 1 2 (k + 3)(k + 2)k , so for k = 1 we have dim(R1 ) = 6. In two dimensions, a polynomial p R1 has dimension 3 and can be represented as p = (c1 + c2 y, c3 c2 x). For k = 1, the local degrees of freedom in two dimensions are given by Ni (p) =
ei

(23)

p i ds, 2

(24)

where i is the tangential unit vector of edge ei . The local shape functions can be expressed in terms of barycentric coordinates. In two dimensions we can dene the local shape function 1 associated with edge 1 oriented from node 1 to node 2 as 1 = 1 2 2 1 , (25) where the barycentric coordinates for node 1 and 2 can be written as 1 = a1 + b1 x + c1 y and 2 = a2 + b2 x + c2 y respectively, where ai and bi depend on the nodal coordinates of the triangle. Now 1 = 1 2 2 1 = (a1 + b1 x + c1 y )(b2 , c2 ) (a2 + b2 x + c2 y )(b1 , c1 ) = (d1 + d2 y, d3 d2 x) R1 . The local shape functions for triangular elements in two dimensions are shown in Fig. 1. (26)

Figure 1: The local shape functions for k = 1 in two dimensions on a triangle.

3.1

Ane maps

When transforming to and from the reference element we need to transform in a special way since ) with vectors in Rk . For the ane map FK x + bK and = BK x we are working in H (curl; K Rk , the transformation is given by u
T 1 u FK = (BK ) u

(27)

and the curl transforms as u=

1 u . BK det(BK )

(28)

Lemma 1. The space Rk is invariant under the transformation (27) Rk we can write u =p 1 + p 2 , where p 1 [Pk1 ]3 and p 2 Sk . Now, Proof. Since for u
1 T 1 T 1 2 ](FK 1 + (BK (x)) ) p u(x) = [(BK ) p 1 1 1 T 1 T 1 2 (BK 1 ](FK x BK bK ). (x)) + (BK ) p = [(BK ) p

(29)

1 1 1 k ]3 , we have that p 2 [P 2 (BK 2 (BK 3 (x), where p 3 [Pk1 ]3 . Since p x BK bK ) = p x) + p This means 1 1 T 1 T 1 1 (FK 2 (BK 3 (x)] + (BK u(x) = [(BK ) p (x)) + p ) p x). (30) 1 1 1 T 1 T 1 3 [Pk1 ]3 and also that (BK 2 (BK We have that (BK ) p1 FK +p ) p2 (BK x) x = p x) 1 2 Sk . Thus, u Rk . (BK x) = 0, since p

Lemma 2. If h is a regular mesh and s 0, then we have for v transformed by (27) to give v
2 |[H s (K |v )]3 ChK |v|[H s (K )]3

s 1

(31) (32)

and

2 v |[H s (K | )]3 ChK | v|[H s (K )]3 .

s+ 1

3.2

Degrees of freedom

, which is a tetrahedron dened by the vertices a 1 , . . . , a 4 given by For the reference element K T 1 = (0, 0, 0) , a 2 = (1, 0, 0)T , a 3 = (0, 1, 0)T and a 4 = (0, 0, 1)T , we dene the curl-conforming a P , N ) by nite element (K, K K is the reference element K PK = Rk of K , faces f and K itself. Three types of degrees of freedom associated with edges e of K . The unit vector along edge e is denoted Me u) = (
e

q u ds , 1 ) area(f
K f

q Pk1 ( e) , )]3 and q [Pk2 (f =0 , q

(33) (34) (35)

Mf u) = ( MK u) = (

q dA, u

, q dV u

)]3 . [Pk3 (K q

The total set of degrees of freedom is then NK u ) Mf u ) MK u). = Me ( ( ( By using the transformation (27) we can dene the nite element (K, PK , NK ) for a general tetrahedron K as K is a tetrahedron PK = Rk Three types of degrees of freedom associated with edges e of K , faces f of K and K itself. The unit vector along edge e is denoted . Me (u) =
e

u qds, 1 area(f )
K

q Pk1 (e) , , q = BK q )]3 and q [Pk2 (f =0 , q 1 , BK q det(BK )

(36) (37)

Mf (u) = MK (u) =

u qdA,
f

u qdV,

q mapped by q FK =

[Pk3 (K )]3 . q (38)

The total set of degrees of freedom for a general tetrahedron is then NK = Me (u) Mf (u) MK (u). and the transformation (27) is used, the degrees of . freedom (36)(38) on K are identical to those on K 4 Lemma 3. If det(BK ) 0, =
BK | |BK

3.3

Unisolvence

Theorem 1. If u Rk is such that all degrees of freedom (36)(38)vanish, then u = 0. To prove Theorem 1 we need two Lemmas. Lemma 4. If u Rk is such that the degrees of freedom (37) vanish on f and the degrees of freedom (36) vanish on all edges of f , then u = 0 on f . Lemma 5. If u Rk is such that u = 0, then u = p for p Pk . Proof of Theorem 1. Since the degrees of freedom on the general tetrahedron and the reference tetrahedron are identical, we prove the theorem on the reference element. Because of Lemma 4 we . By integration by parts and using that the degrees of freedom (38) = 0 on K know that u are zero, we have
K

u = q dV

q = 0, dV u

[Pk2 ]3 . q

(39)

of K and that the degrees of freedom (37) are zero, we have Using Stokes theorem on each face f = u T q dA f = 0, q T dA u f ), q Pk1 (f (40)

q = is dened as . The surface vector curl operator T = ( ) on K where u u f f and hence on K q . This means that u u . Because of this T = ( ) = 0 on f
f f

u [Pk1 ]3 , we can write and the fact that u = ( x1 1 , x 2 , 2 , x 3 3 )T , (41)

u . = in (39) shows that = 0 in K where = (1 , 2 , 3 )T [Pk2 ]3 . Picking q = p =0 Using Lemma 5 this means that we can write u for some p Pk . The fact that u implies that we can take p and write p on K = 0 on K as p = x 1 x 2 x 3 r , for some r Pk3 . (42)

= 0. The vanishing of the degrees of freedom (38) implies that p = 0 and thus u

3.4

The interpolant
Vh = {u H (curl; ) : u|K Rk K h }. (43)

We can dene the global nite element space on a mesh h as

The local interpolant rK u Rk for K h is characterized by the vanishing of the degrees of freedom on u rK u, Me (u rK u) = Mf (u rK u) = MK (u rK u) = 0. The global interpolant is then dened by rh u|K = rK u, K h . (45) (44)

Lemma 6. If u [Lp (K )]3 , u [Lp (K )]3 and u [Lp (K )]3 for p > 2, then rh u is well-dened and bounded. Theorem 2. Let h be a regular mesh on . If u [H s ()]3 and u [H s ()]3 for s k , then ||u rh u||[L2 ()]3 + || (u rh u)||[L2 ()]3 Chs (||u||[H s ()]3 + || u||[H s ()]3 ). 5 (46)

The error estimates for the interpolant holds for (u rh u) and ( (u rh u) separately: ||u rh u||[L2 ()]3 Chs (||u||[H s ()]3 + || u||[H s ()]3 ) ||u rh u||[L2 ()]3 Chs ||u||[H s ()]3 (for s > 1) || (u rh u)||[L2 ()]3 Chs || u||[H s ()]3 . Theorem 3. Let h be a regular mesh on . If u [H 2 + (K )]3 , 0 < k1 x, then [Pk1 ]3 P
1 1

(47) (48) (49)

1 2

and u|K Dk =

2 ||u rh u||[L2 (K )]3 C (hK ||u||[H 1/2+ (K )]3 + hK || u||[L2 (K )]3 ).

(50)

The error estimates show that the convergence is ||u rh u||[L2 ()]3 = O(hk ), according to the Bramble-Hilbert lemma since [Pk1 ]3 Rk , but [Pk ]3 Rk . For triangular elements in two dimensions it has been shown that superconvergence is obtained, so that the order of convergence is similar to nodal based elements, which have O(hk+1 ) convergence for degree k . To obtain higher order convergence, a second family of vector elements was introduced by Nedelec: K is a tetrahedron PK = [Pk ]3 Three types of degrees of freedom associated with edges e of K , faces f of K and K itself. The unit vector along edge e is denoted . Me (u) =
e

u qds, 1 area(f )
K

q Pk1 (e) , q Dk1 (f ) ,

(51) (52) (53)

Mf (u) = MK (u) =

uT qdA,
f

u qdV,

q Dk1 (K ) .

The total set of degrees of freedom is then NK = Me (u) Mf (u) MK (u). The dimension of the second family of elements is 1 2 (k + 1)(k + 2)(k + 3).

References
[1] C. Amrouce, C. Bernardi, M. Dauge, and V. Girault. Vector potentials in three-dimensional non-smooth domains. Mathematical Methods in the Applied Sciences, 21:823864, 1998. [2] D. Bo, P. Fernandes, L. Gastaldi, and I. Perugia. Computational models of electromagnetic resonators: Analysis of edge element approximation. SIAM J. Numer. Anal., 36:12641290, 1999. [3] Daniele Bo, Marco Farina, and Lucia Gastaldi. On the approximation of Maxwells eigenproblem in general 2D domains. Computers and Structures, 79:10891096, 2001. [4] Daniele Bo and Lucia Gastaldi. Interpolation estimates for edge nite elements and application to band gap computation. Applied Numerical Mathematics, 56:12831292, 2006. [5] Martin Costabel and Monique Dauge. Maxwell and lame eigenvalues on polyhedra. Mathematical Methods in the Applied Sciences, 22:243258, 1999. [6] R. Hiptmair. Finite elements in computational electromagnetics. Acta Numerica, 11:237339, 2002. 6

[7] Jianming Jin. The Finite Element Method in Electromagnetics. John Wiley & Sons, 2 edition, 2002. [8] Daniel R. Lynch and Keith D. Paulsen. Origin of vector parasites in numerical Maxwell solutions. IEEE Transactions on Microwave Theory and Techniques, 39, 1991. [9] Peter Monk. Finite Element Methods for Maxwells Equations. Oxford Science Publications, 2003. [10] J. C. Nedelec. Mixed nite elements in R3 . Numerische Mathematik, 35(3):315341, September 1980. [11] J.C. Nedelec. A new family of mixed nite elements in R3 . Numerische Mathematik, 50:5781, 1986.