Stable and unstable finite elements for the Maxwell eigenvalue problem

The Maxwell eigenvalue problem seeks eigenvalues \(\lambda\) and the corresponding nonzero vector-valued eigenfunctions \(u\) satisfying the partial differential equation

\[\operatorname{curl}\operatorname{curl} u = \lambda u \text{ in $\Omega$}\]

(we have simplified slightly by setting the material parameters equal to 1). The PDE is to be supplemented with boundary conditions, which we take here to be the essential boundary condition

\[u \times n = 0 \text{ on $\partial\Omega$}.\]

The eigenvalues \(\lambda\) are all real and non-negative, but only the positive ones are of interest, since, if \(\lambda >0\), then it follows from the PDE that \(\operatorname{div} u = 0\), which is also a requirement of Maxwell’s equations. There exist, in addition, an infinite-dimensional family of eigenfunctions with eigenvalue \(\lambda=0\), since for any smooth function \(\phi\) vanishing to second order on the boundary, \(u=\operatorname{grad}\phi\) is such an eigenfunction. But these functions are not divergence-free and should not be considered Maxwell eigenfunctions.

Model problem

In this demo we shall consider the Maxwell eigenvalue problem in two dimensions with the domain \(\Omega\) taken to be the square \((0,\pi)\times(0,\pi)\), since in that case the exact eigenpairs have a simple analytic expression. They are

\[u(x,y) = (\sin m x, \sin n y), \quad \lambda = m^2 + n^2,\]

for any non-negative integers \(m\) and \(n,\) not both zero. Thus the eigenvalues are

\[\lambda = 1, 1, 2, 4, 4, 5, 5, 8, 9, 9, 10, 10, 13, 13, \dots\]

In the demo program we compute the 12 eigenvalues nearest 5.5, and so should obtain the first 12 numbers on this list, ranging from 1 to 10.

The weak formulation and the finite element method

A weak formulation of the eigenvalue problem seeks \(0\ne u\in H_0(\operatorname{curl})\) and \(\lambda>0\) such that

\[\int_{\Omega} \operatorname{curl} u\, \operatorname{curl}v\, {\rm d} x = \lambda \int_{\Omega} u v\, {\rm d} x \quad \forall \ v\in H_0(\operatorname{curl}),\]

where \(H_0(\operatorname{curl})\) is the space of square-integrable vector fields with square-integrable curl and satisfying the essential boundary condition. If we replace \(H_0(\operatorname{curl})\) in this formulation by a finite element subspace, we obtain a finite element method.

Stable and unstable finite elements

We consider here two possible choices of finite element spaces. The first, the Nédélec edge elements, which are obtained in FEniCS as FunctionSpace(mesh, 'H1curl', 1), are well suited to this problem and give an accurate discretization. The second choice is simply the vector-valued Lagrange piecewise linears: VectorFunctionSpace(mesh, 'Lagrange', 1). To the uninitiated it usually comes as a surprise that the Lagrange elements do not provide an accurate discretization of the Maxwell eigenvalue problem: the computed eigenvalues do not converge to the true ones as the mesh is refined! This is a subtle matter connected to the stability theory of mixed finite element methods. See this paper for details.

While the Nédélec elements behave stably for any mesh, the failure of the Lagrange elements differs on different sorts of meshes. Here we compute with two structured meshes, the first obtained from a \(40\times 40\) grid of squares by dividing each with its positively-sloped diagonal, and the second the crossed mesh obtained by dividing each subsquare into four using both diagonals. The output from the first case is:

diagonal mesh
Nédélec:   [ 1.00  1.00  2.00  4.00  4.00  5.00  5.00  8.01  8.98  8.99  9.99  9.99]
Lagrange:  [ 5.16  5.26  5.26  5.30  5.39  5.45  5.53  5.61  5.61  5.62  5.71  5.73]
Exact:     [ 1.00  1.00  2.00  4.00  4.00  5.00  5.00  8.00  9.00  9.00 10.00 10.00]

Note that the eigenvalues calculated using the Nédélec elements are all correct to within a fraction of a percent. But the 12 eigenvalues computed by the Lagrange elements are certainly all wrong, since they are far from being integers!

On the crossed mesh, we obtain a different mode of failure:

crossed mesh
Nédélec:   [ 1.00  1.00  2.00  4.00  4.00  5.00  5.00  7.99  9.00  9.00 10.00 10.00]
Lagrange:  [ 1.00  1.00  2.00  4.00  4.00  5.00  5.00  6.00  8.01  9.01  9.01 10.02]
Exact:     [ 1.00  1.00  2.00  4.00  4.00  5.00  5.00  8.00  9.00  9.00 10.00 10.00]

Again the Nédélec elements are accurate. The Lagrange elements also approximate most of the eigenvalues well, but they return a totally spurious value of 6.00 as well. If we were to compute more eigenvalues, more spurious ones would be returned. This mode of failure might be considered more dangerous, since it is harder to spot.

The implementation

Preamble. First we import dolfin and numpy and make sure that dolfin has been configured with PETSc and SLEPc (since we depend on the SLEPc eigenvalue solver).

from dolfin import *
import numpy as np
if not has_linear_algebra_backend("PETSc"):
    print("DOLFIN has not been configured with PETSc. Exiting.")
    exit()
if not has_slepc():
    print("DOLFIN has not been configured with SLEPc. Exiting.")
    exit()

Function eigenvalues. The function eigenvalues takes the finite element space V and the essential boundary conditions bcs for it, and returns a requested set of Maxwell eigenvalues (specified in the code below) as a sorted numpy array:

def eigenvalues(V, bcs):

We start by defining the bilinear forms on the right- and left-hand sides of the weak formulation:

#
    # Define the bilinear forms on the right- and left-hand sides
    u = TrialFunction(V)
    v = TestFunction(V)
    a = inner(curl(u), curl(v))*dx
    b = inner(u, v)*dx

Next we assemble the bilinear forms a and b into PETSc matrices A and B, so the eigenvalue problem is converted into a generalized matrix eigenvalue problem \(Ax=\lambda B x\). During the assembly step the essential boundary conditions are incorporated by modifying the rows and columns of the matrices corresponding to constrained boundary degrees of freedom. We use assemble_system rather than assemble to do the assembly, since it maintains the symmetry of the matrices. assemble_system is designed for source problems, rather than eigenvalue problems, and requires a right-hand side linear form, so we define a dummy form to feed it.

#
    # Assemble into PETSc matrices
    dummy = v[0]*dx
    A = PETScMatrix()
    assemble_system(a, dummy, bcs, A_tensor=A)
    B = PETScMatrix()
    assemble_system(b, dummy, bcs, A_tensor=B)

We zero out the rows of \(B\) corresponding to constrained boundary degrees of freedom, so as not to introduce spurious eigenpairs with nonzero boundary DOFs.

#
    [bc.zero(B) for bc in bcs]

Now we solve the generalized matrix eigenvalue problem using the SLEPc package. The behavior of the SLEPcEigenSolver is controlled by a parameter set (use info(solver, True) to see all possible parameters). We use parameters to set the eigensolution method to Krylov-Schur, which is good for computing a subset of the eigenvalues of a sparse matrix, and to tell SLEPc that the matrices A and B in the generalized eigenvalue problem are symmetric (Hermitian).

#
    solver = SLEPcEigenSolver(A, B)
    solver.parameters["solver"] = "krylov-schur"
    solver.parameters["problem_type"] = "gen_hermitian"

We specify that we want 12 eigenvalues nearest in magnitude to a target value of 5.5. Note that when the spectrum parameter is set to target magnitude, the spectral_transform parameter should be set to shift-and-invert and the spectral_shift parameter should be set equal to the target.

#
    solver.parameters["spectrum"] = "target magnitude"
    solver.parameters["spectral_transform"] = "shift-and-invert"
    solver.parameters["spectral_shift"] = 5.5
    neigs = 12
    solver.solve(neigs)

Finally we collect the computed eigenvalues in list which we convert to a numpy array and sort before returning. Note that we are not guaranteed to get the number of eigenvalues requested. The function solver.get_number_converged() reports the actual number of eigenvalues computed, which may be more or less than the number requested.

#
    # Return the computed eigenvalues in a sorted array
    computed_eigenvalues = []
    for i in range(min(neigs, solver.get_number_converged())):
        r, _ = solver.get_eigenvalue(i) # ignore the imaginary part
        computed_eigenvalues.append(r)
    return np.sort(np.array(computed_eigenvalues))

Function print_eigenvalues. Given just a mesh, the function print_eigenvalues calls the preceding function eigenvalues to solve the Maxwell eigenvalue problem for each of the two finite element spaces, Nédélec and Lagrange, and prints the results, together with the known exact eigenvalues:

def print_eigenvalues(mesh):

First we define the Nédélec edge element space and the essential boundary conditions for it, and call eigenvalues to compute the eigenvalues. Since the degrees of freedom for the Nédélec space are tangential components on element edges, we simply need to constrain all the DOFs associated to boundary points to zero.

#
    nedelec_V   = FunctionSpace(mesh, "N1curl", 1)
    nedelec_bcs = [DirichletBC(nedelec_V, Constant((0.0, 0.0)), DomainBoundary())]
    nedelec_eig = eigenvalues(nedelec_V, nedelec_bcs)

Then we do the same for the vector Lagrange elements. Since the Lagrange DOFs are both components of the vector, we must specify which component must vanish on which edges (the x-component on horizontal edges and the y-component on vertical edges).

#
    lagrange_V   = VectorFunctionSpace(mesh, "Lagrange", 1)
    lagrange_bcs = [DirichletBC(lagrange_V.sub(1), 0, "near(x[0], 0) || near(x[0], pi)"),
                    DirichletBC(lagrange_V.sub(0), 0, "near(x[1], 0) || near(x[1], pi)")]
    lagrange_eig = eigenvalues(lagrange_V, lagrange_bcs)

The true eigenvalues are just the 12 smallest numbers of the form \(m^2 + n^2\), \(m,n\ge0\), not counting 0.

#
    true_eig = np.sort(np.array([float(m**2 + n**2) for m in range(6) for n in range(6)]))[1:13]

Finally we print the results:

#
    np.set_printoptions(formatter={'float': '{:5.2f}'.format})
    print("Nedelec:  {}".format(nedelec_eig))
    print("Lagrange: {}".format(lagrange_eig))
    print("Exact:    {}".format(true_eig))

Calling the functions. To complete the program, we call print_eigenvalues for each of two different meshes

mesh = RectangleMesh(Point(0, 0), Point(pi, pi), 40, 40)
print("\ndiagonal mesh")
print_eigenvalues(mesh)

mesh = RectangleMesh(Point(0, 0), Point(pi, pi), 40, 40, "crossed")
print("\ncrossed mesh")
print_eigenvalues(mesh)