Overview

SLEPc for Python (slepc4py) is a Python package that provides convenient access to the functionality of SLEPc.

SLEPc [1], [2] implements algorithms and tools for the numerical solution of large, sparse eigenvalue problems on parallel computers. It can be used for linear eigenvalue problems in either standard or generalized form, with real or complex arithmetic. It can also be used for computing a partial SVD of a large, sparse, rectangular matrix, and to solve nonlinear eigenvalue problems (polynomial or general). Additionally, SLEPc provides solvers for the computation of the action of a matrix function on a vector.

SLEPc is intended for computing a subset of the spectrum of a matrix (or matrix pair). One can for instance approximate the largest magnitude eigenvalues, or the smallest ones, or even those eigenvalues located near a given region of the complex plane. Interior eigenvalues are harder to compute, so SLEPc provides different methodologies. One such method is to use a spectral transformation. Cheaper alternatives are also available.

[1]J. E. Roman, C. Campos, L. Dalcin, E. Romero, A. Tomas. SLEPc Users Manual. DSIC-II/24/02 - Revision 3.16 D. Sistemas Informaticos y Computacion, Universitat Politecnica de Valencia. 2021.
[2]Vicente Hernandez, Jose E. Roman and Vicente Vidal. SLEPc: A Scalable and Flexible Toolkit for the Solution of Eigenvalue Problems, ACM Trans. Math. Softw. 31(3), pp. 351-362, 2005.

Features

Currently, the following types of eigenproblems can be addressed:

  • Standard eigenvalue problem, Ax=kx, either for Hermitian or non-Hermitian matrices.
  • Generalized eigenvalue problem, Ax=kBx, either Hermitian positive-definite or not.
  • Partial singular value decomposition of a rectangular matrix, Au=sv.
  • Polynomial eigenvalue problem, P(k)x=0.
  • Nonlinear eigenvalue problem, T(k)x=0.
  • Computing the action of a matrix function on a vector, w=f(alpha A)v.

For the linear eigenvalue problem, the following methods are available:

  • Krylov eigensolvers, particularly Krylov-Schur, Arnoldi, and Lanczos.
  • Davidson-type eigensolvers, including Generalized Davidson and Jacobi-Davidson.
  • Subspace iteration and single vector iterations (inverse iteration, RQI).
  • Conjugate gradient for the minimization of the Rayleigh quotient.
  • A contour integral solver.

For singular value computations, the following alternatives can be used:

  • Use an eigensolver via the cross-product matrix A’A or the cyclic matrix [0 A; A’ 0].
  • Explicitly restarted Lanczos bidiagonalization.
  • Implicitly restarted Lanczos bidiagonalization (thick-restart Lanczos).

For polynomial eigenvalue problems, the following methods are available:

  • Use an eigensolver to solve the generalized eigenvalue problem obtained after linearization.
  • TOAR and Q-Arnoldi, memory efficient variants of Arnoldi for polynomial eigenproblems.

For general nonlinear eigenvalue problems, the following methods can be used:

  • Solve a polynomial eigenproblem obtained via polynomial interpolation.
  • Rational interpolation and linearization (NLEIGS).
  • Newton-type methods such as SLP or RII.

Computation of interior eigenvalues is supported by means of the following methodologies:

  • Spectral transformations, such as shift-and-invert. This technique implicitly uses the inverse of the shifted matrix (A-tI) in order to compute eigenvalues closest to a given target value, t.
  • Harmonic extraction, a cheap alternative to shift-and-invert that also tries to approximate eigenvalues closest to a target, t, but without requiring a matrix inversion.

Other remarkable features include:

  • High computational efficiency, by using NumPy and SLEPc under the hood.
  • Data-structure neutral implementation, by using efficient sparse matrix storage provided by PETSc. Implicit matrix representation is also available by providing basic operations such as matrix-vector products as user-defined Python functions.
  • Run-time flexibility, by specifying numerous setting at the command line.
  • Ability to do the computation in parallel.

Components

SLEPc provides the following components, which are mirrored by slepc4py for its use from Python. The first five components are solvers for different classes of problems, while the rest can be considered auxiliary object.

EPS:The Eigenvalue Problem Solver is the component that provides all the functionality necessary to define and solve an eigenproblem. It provides mechanisms for completely specifying the problem: the problem type (e.g. standard symmetric), number of eigenvalues to compute, part of the spectrum of interest. Once the problem has been defined, a collection of solvers can be used to compute the required solutions. The behaviour of the solvers can be tuned by means of a few parameters, such as the maximum dimension of the subspace to be used during the computation.
SVD:This component is the analog of EPS for the case of Singular Value Decompositions. The user provides a rectangular matrix and specifies how many singular values and vectors are to be computed, whether the largest or smallest ones, as well as some other parameters for fine tuning the computation. Different solvers are available, as in the case of EPS.
PEP:This component is the analog of EPS for the case of Polynomial Eigenvalue Problems. The user provides the coefficient matrices of the polynomial. Several parameters can be specified, as in the case of EPS. It is also possible to indicate whether the problem belongs to a special type, e.g., symmetric or gyroscopic.
NEP:This component covers the case of general nonlinear eigenproblems, T(lambda)x=0. The user provides the parameter-dependent matrix T via the split form or by means of callback functions.
MFN:This component provides the functionality for computing the action of a matrix function on a vector. Given a matrix A and a vector b, the call MFNSolve(mfn,b,x) computes x=f(A)b, where f is a function such as the exponential.
ST:The Spectral Transformation is a component that provides convenient implementations of common spectral transformations. These are simple transformations that map eigenvalues to different positions, in such a way that convergence to wanted eigenvalues is enhanced. The most common spectral transformation is shift-and-invert, that allows for the computation of eigenvalues closest to a given target value.
BV:This component encapsulates the concept of a set of Basis Vectors spanning a vector space. This component provides convenient access to common operations such as orthogonalization of vectors. The BV component is usually not required by end-users.
DS:The Dense System (or Direct Solver) component, used internally to solve dense eigenproblems of small size that appear in the course of iterative eigensolvers.
FN:A component used to define mathematical functions. This is required by the end-user for instance to define function T(.) when solving nonlinear eigenproblems with NEP in split form.