Introduction

NEP-PACK

NEP-PACK is a package with implementations of methods to solve and to manipulate nonlinear eigenvalue problems of the type: Find $(λ,v)\in\mathbb{C}\times\mathbb{C}^n$ such that

\[M(λ)v=0\]

and $v\neq 0$.

Getting started

Install it as a registered package in Julia's REPL package mode by typing ] add Nonline...:

julia> ]
(v1.0) pkg> add NonlinearEigenproblems

Then we can start to load the NEP-PACK package

julia> using NonlinearEigenproblems

As a first example we will solve the NEP associated with the matrix polynomial

\[M(λ)=\begin{bmatrix}1&3\newline5&6\end{bmatrix}+ λ\begin{bmatrix}3&4\newline6&6\end{bmatrix}+ λ^2\begin{bmatrix}1&0\newline0&1\end{bmatrix}\]

The following code creates this NEP, by constructing an object called PEP, an abbreviation for polynomial eigenvalue problem. It subsequently solves it using the NEP solution method implemented in the NEP-solver polyeig:

julia> A0=[1.0 3; 5 6]; A1=[3.0 4; 6 6]; A2=[1.0 0; 0 1.0];
julia> nep=PEP([A0,A1,A2])
PEP(2, Array{Float64,2}[[1.0 3.0; 5.0 6.0], [3.0 4.0; 6.0 6.0], [1.0 0.0; 0.0 1.0]])
julia> λ,v=polyeig(nep)
(Complex{Float64}[1.36267+0.0im, -0.824084+0.280682im, -0.824084-0.280682im, -8.7145+0.0im], Complex{Float64}[-1.0+0.0im 0.739183-0.196401im 0.739183+0.196401im 0.627138+0.0im; 0.821812+0.0im -0.501408-0.375337im -0.501408+0.375337im 1.0+0.0im])

You have now solved your first nonlinear eigenvalue problem with NEP-PACK.

In order to verify that we have a solution, we can check that $M(λ)$ is singular, with a singular vector $v$ such that $M(λ)v=0$:

julia> λ1=λ[1]; v1=v[:,1];
julia> using LinearAlgebra # the norm-function is in this Julia package
julia> norm(A0*v1+λ1*A1*v1+λ1^2*v1)/norm(v1)
1.1502634749464687e-14
Tip

MATLAB users: Do you have a NEP defined in MATLAB? You can solve MATLAB-defined NEPs with this package. See the MATLAB tutorial. We also have some MATLAB implementations of the solvers in NEP-PACK in a separate repository.

Accessing more complicated applications

We have made benchmark examples available through the function nep_gallery:

julia> nep=nep_gallery("dep0",100);
julia> size(nep)
(100, 100)
julia> λ,v=mslp(nep,tol=1e-10);
julia> λ
0.05046248970129549 - 7.60684247532422e-16im
julia> size(v)
(100,)
julia> resnorm=norm(compute_Mlincomb(nep,λ,v))
5.178780131881974e-13

Information about the gallery can be found by typing ?nep_gallery. The second argument in the call to nep_gallery is a problem parameter, in this case specifying that the size of the problem should be 100. The example solves the problem with the NEP-algorithm MSLP. The parameter tol specifies the tolerance for iteration termination.

Note

All the NEP-solvers have considerble documentation easily available. Every NEP-solver has documentation accompanied with at least one example, and references to corresponding research papers, which we strongly recommend you to cite if you use the method. This is available to you in Julia's repl-prompt. Type ?mslp and you will see an example how to use mslp and that citation credit should go to A. Ruhe, Algorithms for the nonlinear eigenvalue problem, SIAM J. Numer. Anal. 10 (1973) 674-689. This documentation is the same as the online documentation under the tab NEP-solvers.

A model of a neuron

The following (delay) differential equation models the interaction of two neurons

\[\dot{x}_1(t)=-\kappa x_1(t)+\beta\tanh(x_1(t-\tau_3))+a_1\tanh(x_2(t-\tau_2))\]
\[\dot{x}_2(t)=-\kappa x_2(t)+\beta\tanh(x_2(t-\tau_3))+a_2\tanh(x_1(t-\tau_1))\]

See L. P. Shayer and S. A. Campbell. Stability, bifurcation and multistability in a system of two coupled neurons with multiple time delays. SIAM J. Applied Mathematics , 61(2):673–700, 2000. It is also available as a first demo in DDE-BIFTOOL. The linear stability analysis of this problem requires the solution of a nonlinear eigenvalue problem

\[M(λ)=-λI+A_0+A_1e^{-\tau_1λ}+A_2e^{-\tau_2λ}+A_3e^{-\tau_3λ}\]

where the matrices are the Jacobian at the stationary solution. For the zero stationary solution, the matrices are

kappa=0.5; a2=2.34; a1=1; beta=-1;
A0=-kappa*[1 0; 0 1];
A1=a2*[0 0; 1 0];
A2=a1*[0 1; 0 0];
A3=beta*[1 0; 0 1];

We can now create the nonlinear eigenvalue problem and determine the stability by first creating the problem

julia> tauv=[0;0.2;0.2;1.5];
julia> dep=DEP([A0, A1,   A2, A3],tauv);

The constructor DEP is an abbreviation for a delay eigenvalue problem, which is a NEP with exponential terms stemming from the stability analysis of a delay-differential equation. See Types and data-structures for other NEP-types. You can now solve this NEP, for instance, with the infinite Arnoldi method:

julia> λ,V=iar_chebyshev(dep,maxit=100); # This takes some time the first time is run due to JIT-compiler

The figure in a demo of DDE-BIFTOOL http://ddebiftool.sourceforge.net/demos/neuron/html/demo1_stst.html#3 can be directly generated by

using PyPlot
plot(real(λ),imag(λ),"*");
xlabel("real(λ)"); ylabel("imag(λ)");
┌ Warning: No working GUI backend found for matplotlib
└ @ PyPlot ~/.julia/packages/PyPlot/4wzW1/src/init.jl:165

Tip

This problem is also available in the Gallery by calling dep=nep_gallery("neuron0"). Most of the NEPs constructed in the tutorials are also available in corresponding gallery problems. See all gallery problems under NEP Gallery. In particular, note that the problems in the Berlin-Manchester collection of problems NLEVP are also directly available.

The "gun" benchmark problem

One of the most common benchmark problems for NEPs is the so-called "gun"-problem. It models an electromagnetic cavity, and it is directly available in the NEP-PACK gallery. (See gallery references or type ?nep_gallery at the repl-prompt.) This is how you can set it up and solve it with the block Newton method:

julia> nep=nep_gallery("nlevp_native_gun");
julia> n=size(nep,1)
julia> S=150^2*[1.0 0; 0 1]; V=[[1 0; 0 1]; zeros(n-2,2)];
julia> (Z,X)=blocknewton(nep,S=S,X=V,logger=1,armijo_factor=0.5,maxit=20)
Iteration 1: Error: 6.081316e+03
Iteration 2: Error: 1.701970e-02 Armijo scaling=0.031250
Iteration 3: Error: 1.814887e-02 Armijo scaling=0.250000
...
Iteration 13: Error: 6.257442e-09
Iteration 14: Error: 2.525942e-15

This algorithm returns a partial Schur factorization of the NEP, and therefore the eigenvalues of the small matrix Z are eigenvalues of the problem. An eigenpair of the NEP can be extracted by diagonalizing:

julia> using LinearAlgebra
julia> (Λ,P)=eigen(Z);
julia> VV=X*P;  # Construct the eigenvector matrix
julia> v=VV[:,1]; λ=Λ[1]
61330.208714730004 + 63185.15983933589im
julia> norm(compute_Mlincomb(nep,λ,v)) # Very small residual
1.8270553408452648e-16

If you use the NEP-algorithms for research, please give the author of the algorithm credit by citiation. The recommended citation can be found in the function documentation, e.g., ?blocknewton.

Your own NEP nonlinearity

As an application researcher, we recommend that you first try to express your problem in the following form since it gives access to several efficient routines associated with the NEP, in turn making it possible to use many NEP-solvers. A problem that can be expressed as a (short) S um of P roducts of M atrices and F unctions can be represented with the objects of type SPMF_NEP in NEP-PACK. For instance, a problem with three terms

\[M(λ) = A+λB+e^{\sin(λ/2)}C\]

can be created by

julia> A=(1:4)*(1:4)'+I; B=diagm(1 => [1,2,3]); C=ones(4,4);
julia> f1= λ-> one(λ);
julia> f2= λ-> λ;
julia> f3= λ-> exp(sin(λ/2));
julia> nep=SPMF_NEP([A,B,C],[f1,f2,f3]);

The NEP is solved by using the NEP-object as a parameter in a call to an algorithm, e.g.,

julia> v0 = 0.1*[1,-1,1,-1];
julia> λ,v=quasinewton(nep,λ=4,v=v0)
(3.1760990071435193 + 0.0im, Complex{Float64}[2.892363187499394 + 0.0im, -1.6573097795628646 + 0.0im, 0.00729776922332883 + 0.0im, -0.09002519738673213 + 0.0im])

As usual, you can check that we computed a sensible solution:

julia> (A+B*λ+C*exp(sin(λ/2)))*v
4-element Array{Complex{Float64},1}:
  -3.489601657766542e-12 + 0.0im
 -1.0118303586944344e-12 + 0.0im
  -9.480334553029193e-13 + 0.0im
  -5.912084880273861e-13 + 0.0im
Note

The functions f1,f2 and f3 in the example above have to be defined for scalar values and for matrices (in the matrix function sense, not elementwise sense). This is the reason f1 needs to be defined as one(λ), instead of just 1. Fortunately, many elementary functions in Julia already have matrix function implementations, e.g., exp([1 2 ; 3 4]) will return the matrix exponential of the given matrix.

Chebyshev interpolation

In applications, NEP-nonlinearities may be complicated to implement. Directly using the SPMF-functionality where every function needs to be defined in a matrix function sense may require too much work. In this case you may want to use an approximation method to create a new different NEP object for which the matrix functions are easy to implement (or directly available in the package). We illustrate this property with NEP-PACKs Chebyshev interpolation feature.

Suppose you have the following NEP, which requires a Bessel function. The Bessel function is analytic, but its matrix function is not easily available.

julia> using SpecialFunctions; # for the besselj
julia> fv=Vector{Function}(undef,m);
julia> Av=Vector{Matrix{Float64}}(undef,3)
julia> fv[1]=s->one(s);
julia> Av[1]=[ -2.0  -1.0   8.0; -1.0  0  -1.0;   -2.0   -1.0  -2.0];
julia> fv[2]=s->s;
julia> Av[2]=[4.0 -7.0  14.0; 8.0  9.0 -13.0; -1.0 -1.0    10.0];
julia> fv[3]=s->besselj(0, s);
julia> Av[3]=[-7.0 -0.0 -9.0; 8.0  3.0 -3.0;  0.0 13.0  2.0]

We use SPMF_NEP again, but in order to suppress a warning message indicating that evaluation with a matrix function is not available we use the keyword check_consistency=false.

julia> nep=SPMF_NEP(Av,fv,check_consistency=false);

Note that we cannot directly use the nep object with most NEP-solvers, since we did not provide a matrix function implementation for besselj. Any method requiring a derivative will just throw an error message that a matrix function is not defined. Let us now construct an interpolating Chebyshev polynomial, which we can use instead (since its matrix functions are trivial). The command ChebPEP, by default interpolates a NEP in the interval [-1,1] using Chebyshev points and represent the approximation in a Chebyshev basis:

julia> cheb=ChebPEP(nep,9,cosine_formula_cutoff=9);

We can now use an arbitrary method to try to solve this problem, e.g., the newtonqr method.

julia> (λ,v)=newtonqr(cheb,λ=0.0,logger=1)
iter 1 err:0.20552458291903797 λ=0.0 + 0.0im
iter 2 err:0.10317368136012978 λ=-3.1180031985377803 + 0.0im
iter 3 err:0.03898166871714645 λ=-0.5814386400379581 + 0.0im
iter 4 err:0.001421286693333467 λ=-0.4572312118506711 + 0.0im
iter 5 err:1.599526685190599e-6 λ=-0.46101438033594805 + 0.0im
iter 6 err:1.9383172515233692e-12 λ=-0.4610101105535983 + 0.0im
iter 7 err:2.1034235144362163e-17 λ=-0.4610101105484241 + 0.0im
(-0.4610101105484241 + 0.0im, Complex{Float64}[-0.597958+0.0im, 0.322148+0.0im, 1.0+0.0im], Complex{Float64}[-0.257712+0.0im, -0.964465+0.0im, -0.0582387+0.0im])

This solved the interpolated problem quite accurately, which turns out to be a reasonable approximation of the original problem:

julia> norm(compute_Mlincomb(nep,λ,v))
1.148749763351579e-9

The function compute_Mlincomb returns the evaluation of M(λ)*v; see the manual section for compute functions.

What now?

Now you are ready to try out one of our tutorials on artificial boundary conditions, boundary element method, contour integration, or deflation. See also the other tutorials (in the side-bar), or have a look at the examples in NEP-solvers and NEP Gallery.

How do I cite it?

We have a preprint for this work. If you find this software useful please cite this preprint by using this citation data:

@Misc{,
  author = 	 {E. Jarlebring and M. Bennedich and G. Mele and E. Ringh and P. Upadhyaya},
  title = 	 {{NEP-PACK}: A {Julia} package for nonlinear eigenproblems},
  year = 	 {2018},
  note = 	 {https://github.com/nep-pack},
  eprint = 	 {arXiv:1811.09592},
}

If you use a specific NEP-solver, please also give credit to the algorithm researcher. Reference to a corresponding algorithm paper can be found by in, e.g., by writing ?resinv.