HomerReid / libBeyn

C++ and Julia implementations of Beyn's contour-integral method for nonlinear eigenproblems

libBeyn: A C++ Implementation of Beyn's Contour-Integration Method for Nonlinear Eigenproblems

libBeyn is a C++ library that implements the contour-integration method for nonlinear eigenproblems described by W-J. Beyn in this paper:

• Wolf-Jürgen Beyn, ``An integral method for solving nonlinear eigenvalue problems.'' Linear Algebra and its Applications 436 3839 (May 2012).

• DOI: https://doi.org/10.1016/j.laa.2011.03.030

• ArXiV: https://arxiv.org/abs/1003.1580

More specifically, using a modified version of Beyn's notation, we consider nonlinear eigenproblems of the form

T(z) v = 0

where T is a D×D-matrix-valued function of a complex variable z and v is an D-dimensional complex vector. (Beyn's notation uses lower-case m for the dimension D and does not use boldface for matrices or vectors.)

libBeyn includes both a C++ library and a (totally separate) native Julia code that both implement Beyn's algorithm.

Dependencies, building, installation

• The libBeyn library and the C++ application codes that use it depend on SCUFF-EM, so you will need to download and install that code first if you want to use the C++ version of the code.

• The repository includes a simple handwritten Makefile, which you will want to modify to configure things like the location of your SCUFF-EM installation

• Then just make to build the library libBeyn.a and the executables tBeyn411 and scuff-spectrum.

• Alternatively, you can bypass the C++ code and use the separate, native julia implementation provided in the julia folder to solve nonlinear eigenproblems.

API

The libBeyn API is extremely simple; it basically just exports a single routine, plus some associated support routines.

Initialize data structure

 BeynSolver *CreateBeynSolver(int D, int L);

Allocate, initialize, and return a BeynSolver structure for a problem of dimension D in which we expect no more than L eigenvalues to reside within the integration we specify.

Solve nonlinear eigenproblem (circular contour)

 int BeynSolve(BeynSolver *Solver, BeynFunction UserFunction, void *UserData,
               cdouble z0, double R, int N);

Compute all eigenvalues inside a circular contour, centered at z0 with radius R, using N-point trapezoidal-rule quadrature for contour integrals. (Here cdouble is short for std::complex<double>).

The parameters are as follows:

• BeynSolver *Solver

  • Data structure created by a prior call to CreateBeynSolver().

• BeynFunction UserFunction

• void *UserData

  • User-supplied function that performs linear algebra involving the T matrix (see below).

• cdouble z0

• double R

• int N

  • Together these parameters specify the integration contour (a circle centered at z0 with radius R) and the quadrature strategy (N-point rectangular rule).

The return value of BeynSolve is the number of eigenvalues identified within the given contour. The actual eigenvalues and eigenvectors are stored within the Solver structure (see below).

Solve nonlinear eigenproblem (elliptical contour)

 int BeynSolve(BeynSolver *Solver, BeynFunction UserFunction, void *UserData,
               cdouble z0, double Rx, double Ry, int N);

A variant of the above routine in which the integration contour is an ellipse centered at z0 with horizontal and vertical radii Rx, Ry.

User function

The prototype of the user-supplied function passed to BeynSolve is

  void UserFunction(cdouble z, void *UserData, HMatrix *VHat, HMatrix *MVHat);

where VHat is an M×L complex-valued matrix.

If MVHat is non-null, the user's function should overwrite MVHat with T(z) * VHat, i.e. the function should left-multiply VHat by T(z) and store the result in MVHat.

If MVHat is null, the function should instead overwrite VHat with T(z) \ VHat, i.e. the function should left-multiply the matrix VHat by the inverse of T(z) and store the result in VHat.

Note: HMatrix is a simple LAPACK wrapper class provided by the libhmat support library for SCUFF-EM, but you don't really need to use libhmat if you don't want to; just work directly with the three main fields of HMatrix:

 int NR      = VHat->NR;   // number of rows
 int NC      = VHat->NC;   // number of columns
 cdouble *ZM = VHat->ZM;   // row-major array of length NR*NC

Eigenvalues and Eigenvectors

The eigenvalues and eigenvectors are stored in the Lambda and Eigenvectors fields of Solver. If the integer K returned by BeynSolve is nonzero,

• Data->Lambda->GetEntry(k) returns the kth eigenvalue (k=0,1,...,K-1)

• Data->Eigenvectors->GetEntry(m,k) returns the mth component of the kth eigenvector.

• Data->Residuals->GetEntryD(k) is the residual of the kth eigenvector, i.e. |T(z) vk| where vk is the kth eigenvector and || denotes L2 norm.

Environment variables

The following environment variables may be set to customize the behavior of the Beyn solver.

• SCUFF_BEYN_RANK_TOL=1.0e-4 The threshold magnitude for singular values to be considered relevant.

• SCUFF_BEYN_RES_TOL=1.0e-4 The threshold magnitude for the residual. Candidate eigenvectors for which |T(z) vk| exceeds this value are discarded. By default this is set to 0, in which case all eigenvectors are retained.

• SCUFF_BEYN_VERBOSE=1 Print more verbose messages to logfiles.

Sample problem: Beyn example 4.11

For testing purposes, I consider the nonlinear eigenproblem of Example 4.11 of the Beyn paper referenced above, which in turn borrowed the example from this reference:

• Kressner, D. "A block Newton method for nonlinear eigenvalue problems." Numerische Mathematik 114 355 (2009)

• DOI: https://doi.org/10.1007/s00211-009-0259-x

This reference quotes 10-digit values for 5 eigenvalues on the real line, namely:

 Lambda1 = 201.88234012
 Lambda2 = 122.91317036
 Lambda3 = 63.692138408
 Lambda4 = 24.219005847
 Lambda5 = 4.4820338110

Solving the problem with libBeyn

The libBeyn distribution includes a C++ code named tBeyn411 that uses libBeyn to solve the nonlinear eigenproblem. Run tBeyn411 --help to summarize command-line options:

 unix% tBeyn411 --help

usage: tBeyn411 [options]

 options: 

  --z0 xx             (center of elliptical contour ({150,0}))
  --Rx xx             (horizontal radius of elliptical contour (148))
  --Ry xx             (vertical radius of elliptical contour (148))
  --N xx              (number of quadrature points (50))
  --L xx              (number of EVs expected in contour (10))
  --Dim xx            (dimension of problem (400))

Here's a sample run in which we ask for all eigenvalues in a circular contour of radius R=148 centered at z0=150+2i with 25 quadrature points:

unix % tBeyn411 --z0 150+2i --Rx 148 --Ry 148 --N 25
Found 8 eigenvalues: 
0: {+3.006036828652e+02,-4.125966235335e-10} 
1: {+2.018823401134e+02,-1.492634460476e-09} 
2: {+4.190845874670e+02,+3.012321138414e-04} 
3: {+1.229131703422e+02,-1.611239586197e-09} 
4: {+6.369213835606e+01,-2.175300739538e-08} 
5: {+2.421900544828e+01,-1.314105806394e-07} 
6: {+4.482017101574e+00,-7.884401266534e-06} 
7: {+4.570891096865e-01,-1.236450296460e-04} 
2054 hikari /home/homer/work/libBeyn <> 

So, we've found all 5 of the correct eigenvalues, with precision ranging from 10 digits (Lambda1) to 5 digits (Lambda5), plus 3 spurious eigenvalues.

Let's try again with the same number of quadrature points but now an elliptical contour, centered at the same point but now much wider than it is tall:

2059 hikari /home/homer/work/libBeyn <> tBeyn411 --z0 150+2i --Rx 148 --Ry 5 --N 25
Found 7 eigenvalues: 
0: {+3.006036828638e+02,-2.343307770047e-10} 
1: {+2.018823401181e+02,-1.838840191226e-11} 
2: {+1.229131703566e+02,+2.110533969812e-12} 
3: {+6.369213840724e+01,-2.950528710244e-11} 
4: {+2.421900584303e+01,-3.840048279358e-09} 
5: {+4.482033592848e+00,+1.790884427422e-07} 
6: {+4.573368338358e-01,-2.154159593548e-05} 

We get one fewer spurious eigenvalue, plus better accuracy in our determination of Lambda5.

Julia implementation

The libBeyn repository also includes a native Julia implementation of the Beyn method, structured similarly to the C++ library described above. The main solver routine is BeynSolve() in julia/BeynSolve.jl and a test code that solves the Beyn Example 4.11 problem is in julia/tBeyn411.jl.

Usage example:

 unix % julia

julia> include("tBeyn411.jl")
SolveBeyn411 (generic function with 1 method)

julia> SolveBeyn411()
Found 7 nonzero singular values.
 1: {+3.006036828641e+02, -1.864047544792e-12} 
 2: {+2.018823401181e+02, -3.890499332734e-12} 
 3: {+1.229131703566e+02, +1.285568188887e-11} 
 4: {+6.369213840779e+01, +4.993707422805e-12} 
 5: {+2.421900584750e+01, -3.318073056626e-10} 
 6: {+4.482033815245e+00, -3.973736028485e-09} 
 7: {+4.573182886588e-01, +4.649270462093e-08} 

Mode solver for SCUFF-EM

Also packaged with libBeyn is a first stab at a mode solver for SCUFF-EM that uses Beyn's method to search for electromagnetic resonance frequencies. This code, called scuff-spectrum, inputs:

• a SCUFF-EM geometry file

• a file containing a succession of (omega0, R, N) tuples. (For geometries with 1D or 2D lattice periodicity, each tuple also includes a 1D or 2D Bloch vector).

For each (omega0, R, N) tuple in the file, scuff-spectrum executes Beyn's method---for a circular contour of radius R, centered at omega0, using N-point rectangular-rule quadrature---to compute all resonance frequencies of the SIE impedance matrix M(ω)---lying within the contor. (A resonance frequency ω is simply a frequency at which M(ω) has a nontrivial nullspace, in which case the usual SIE equation Ms=f can have nonzero surface-current solutions s even for vanishing incident fields f.)

Example: Modes of a spherical dielectric cavity

As a simple test, I'll use scuff-spectrum to look for modes of a spherical dielectric cavity---namely, a sphere of radius R=1 μm, filled with a homogeneous lossless dielectric with relative permittivity εr=4. For this problem, I can compute the exact mode frequencies numerically by looking for roots of the denominator of the Mie scattering coefficients relating interior to incident fields for M-type and N-type spherical polarizations. This calculation is described in Section 5.3 in this memo.

Here's a plot of the lowest 8 resonant frequencies as computed via Mie theory and the eigenvalues identified by SCUFF-SPECTRUM for a radius-1 sphere with ε=4, meshed into 501 panels, for N=20 and N=40 quadrature points.

The two calculations do not agree! What am I doing wrong?