OXFORD UNIVERSITY COMPUTING LABORATORY

Examples of pseudospectra

See also Examples from engineering, and the Pseudospectra Gateway

Three examples are shown here:


A mildly non-normal matrix

Here is an example of a matrix (the reaction-diffusion Brusselator model of dimension 800, taken from Matrix Market) where the computation indicates that non-normality is not significant. The smallest epsilon-pseudospectrum shown is 10-2 and the scale of the entire picture is O(1). The diameter of the 10-2-pseudospectrum is not much bigger than 2*10-2, which shows that the matrix is reasonably normal and that the Ritz values returned should be both accurate and physically meaningful.

Pseudospectra of a mildly non-normal matrix.


An extremely non-normal matrix

This example is very different to the above. The matrix is the `Grcar matrix' of dimension 400, and the smallest epsilon-pseudospectrum shown is 10-11 even though the axis scale is O(1). This indicates that this matrix is strongly non-normal. The eigenvalues returned by the Arnoldi iteration are shown as black dots. The actual eigenvalues of the matrix are shown as magenta stars. They lie nowhere near the Arnoldi values. This is a typical example of what can happen when computing eigenvalues of a very non-normal matrix. The Arnoldi iteration has converged to eigenvalues of a slightly perturbed matrix, but in this case the non-normality is so extreme that these are far from the true eigenvalues.

Pseudospectra of a highly non-normal matrix.


A moderately non-normal matrix

Perhaps the most important situation is where machine precision is sufficient to converge to eigenvalue estimates, but pronounced non-normality may nevertheless diminish the physical significance of some of them. An example of a case in which this is important is the matrix created by linearisation about the laminar solution of the Navier-Stokes equations for fluid flow in an infinite circular pipe. (Our matrix is obtained by a Petrov-Galerkin spectral discretisation of the Navier-Stokes problem due to Meseguer and Trefethen. The axial and azimuthal wave numbers are 0 and 1, respectively.) The pseudospectra are shown below, and although the eigenvalues all have negative real part, implying stability of the flow, the pseudospectra protrude far into the right half-plane. This implies pronounced transient growth of some perturbations of the velocity field in the pipe, which in the presence of nonlinearities in practice may lead to transition to turbulence.

Pseudospectra of 
a moderately non-normal matrix.


Pseudospectra GUI home page.


[Oxford Spires]



Oxford University Computing Laboratory Courses Research People About us News