The Eigenvalue Equation

4-17

The Eigenvalue Equation

The basic eigenvalue problem handled by the PDE Toolbox is -∇ ⋅ (c∇u) + au = λdu

where λ is an unknown complex number. In solid mechanics, this is a problem associated with wave phenomena describing, e.g., the natural modes of a vibrating membrane. In quantum mechanics λ is the energy level of a bound state in the potential well a(x).

The numerical solution is found by discretizing the equation and solving the resulting algebraic eigenvalue problem. Let us first consider the discretization.

Expand u in the FEM basis, multiply with a basis element, and integrate on the domain Ω. This yields the generalized eigenvalue equation

KU = λMU,

where the mass matrix corresponds to the right-hand side, i.e.,

The matrices K and M are produced by calling assema for the equations -∇ ⋅ (c∇u) + au = 0 and -∇ ⋅ (0∇u) + du = 0.

Note that in the most common case, when the function d(x) is positive, the mass matrix M is positive definite symmetric. Likewise, when c(x) is positive and we have Dirichlet boundary conditions, the stiffness matrix K is also positive definite.

The generalized eigenvalue problem, KU = λMU, is now solved by the Arnoldi algorithm applied to a shifted and inverted matrix with restarts until all eigenvalues in the user-specified interval have been found.

Let us describe how this is done in more detail. You may want to look at the example provided in the section “Examples of Eigenvalue Problems” on page 2-27, where an actual run is reported.

First a shift µ is determined close to where we want to find the eigenvalues.

When both K and M are positive definite, it is natural to take µ = 0, and get the smallest eigenvalues; in other cases take any point in the interval [lb,ub] where eigenvalues are sought. Subtract µM from the eigenvalue equation and get (K - µM)U = (λ − µ)MU. Then multiply with the inverse of this shifted matrix

M_{i j,} d( )φx _j( )φx _i( )x dx.

∫

=

4

4-18

and get . This is a standard eigenvalue problem AU

= θU, with the matrix A = (K - µM)^-1M and eigenvalues , i = 1, . . ., n. The largest eigenvalues θ_i of the transformed matrix A now correspond to the eigenvalues λ_i = µ + 1/θ_i of the original pencil (K,M) closest to the shift µ.

The Arnoldi algorithm computes an orthonormal basis V where the shifted and inverted operator A is represented by a Hessenberg matrix H,

AV_j=V_jH_j,j + E_j.

(The subscripts mean that V_j and E_j have j columns and H_j,j has j rows and columns. When no subscripts are used we deal with vectors and matrices of size n.)

Some of the eigenvalues of this Hessenberg matrix H_j,j eventually give good approximations to the eigenvalues of the original pencil (K,M) when the basis grows in dimension j, and less and less of the eigenvector is hidden in the residual matrix E_j.

The basis V is built one column v_j at a time. The first vector v₁ is chosen at random, as n normally distributed random numbers. In step j, the first j vectors are already computed and form the n × j matrix V_j. The next vector v_{j + 1} is computed by first letting A operate on the newest vector v_j, and then making the result orthogonal to all the previous vectors.

This is formulated as, h_{j + 1,j}v_{j + 1} =Av_j - V_jh_j, where the column vector h_j consists of the Gram-Schmidt coefficients, and h_{j + 1,j} is the normalization factor that gives v_{j + 1} unit length. Put the corresponding relations from previous steps in front of this and get,

where H_j,j is a j × j Hessenberg matrix with the vectors h_j as columns. The second term on the right-hand side has nonzeros only in the last column; the earlier normalization factors show up in the subdiagonal of H_j,j.

The eigensolution of the small Hessenberg matrix H_j,j gives approximations to some of the eigenvalues and eigenvectors of the large matrix operator A in the following way. Compute eigenvalues θi and eigenvectors s_i of H_j,j,

1 λ µ–

---U = (K–µM)^–1MU

θ_i 1

λ_i–µ

---=

AV_j = V_jH_{j j,} +v_j+1h_{j+1 j,} e_j^T,

The Eigenvalue Equation

4-19 H_j,js_i = s_iθ_i , i = 1, . . . ,j.

Then y_i = V_js_iis an approximate eigenvector of A, and its residual is, r_i = Ay_i - y_iθ_i = AV_js_i - V_js_iθ_i = (AV_j - V_jH_j,j)s_i = v_j+1h_j+1,js_i,j.

This residual has to be small in norm for θ_i to be a good eigenvalue

approximation. The norm of the residual is ||r_i|| = |h_j+1,js_j,i|, the product of the last subdiagonal element of the Hessenberg matrix and the last element of its eigenvector. It seldom happens that h_j+1,j gets particularly small, but after sufficiently many steps j there are always some eigenvectors s_i with small last elements. Note that the long vector V_j+1 is of unit norm.

It is not necessary to actually compute the eigenvector approximation y_i to get the norm of the residual; we only need to examine the short vectors s_i, and flag those with tiny last components as converged. In a typical case n may be 2000, while j seldom exceeds 50, so all computations that involve only matrices and vectors of size j are much cheaper than those involving vectors of length n.

This eigenvalue computation and test for convergence is done every few steps j, until all approximations to eigenvalues inside the interval [lb,ub] are flagged as converged. When n is much larger than j, this is done very often, for smaller n more seldom. When all eigenvalues inside the interval have converged, or when j has reached a prescribed maximum, the converged eigenvectors, or more appropriately Schur vectors, are computed and put in the front of the basis V.

After this, the Arnoldi algorithm is restarted with a random vector, if all approximations inside the interval are flagged as converged, or else with the best unconverged approximate eigenvector y_i. In each step j of this second Arnoldi run, the vector is made orthogonal to all vectors in V including the converged Schur vectors from the previous runs. This way, the algorithm is applied to a projected matrix, and picks up a second copy of any double eigenvalue there may be in the interval. If anything in the interval converges during this second run, a third is attempted and so on, until no more

approximate eigenvalues θ_i show up inside. Then the algorithm signals convergence. If there are still unconverged approximate eigenvalues after a prescribed maximum number of steps, the algorithm signals nonconvergence and reports all solutions it has found.

4

4-20

This is a heuristic strategy that has worked well on both symmetric, nonsymmetric, and even defective eigenvalue problems. There is a tiny theoretical chance of missing an eigenvalue, if all the random starting vectors happen to be orthogonal to its eigenvector. Normally, the algorithm restarts p times, if the maximum multiplicity of an eigenvalue is p. At each restart a new random starting direction is introduced.

Note that the shifted and inverted matrix A = (K - µM)^-1M is needed only to operate on a vector v_j in the Arnoldi algorithm. This is done by computing an LU factorization,

P(K - µM)Q = LU,

using the sparse MATLAB command lu ( P and Q are permutations that make the triangular factors L and U sparse and the factorization numerically stable).

This factorization needs to be done only once, in the beginning, then x = Av_j is computed as,

x = QU^-1L^-1PMv_j,

with one sparse matrix vector multiplication, a permutation, sparse forwards- and backsubstitutions, and a final renumbering.