raleigh 1.1.0

RALEIGH: RAL EIGensolver for real symmetric and Hermitian problems

RALEIGH is a Python implementation of the block Jacobi-conjugated gradients algorithm for computing several eigenpairs (eigenvalues and corresponding eigenvectors) of large scale real symmetric and Hermitian problems.

Key features

• Can be applied to both standard eigenvalue problem for a real symmetric or Hermitian matrix A and generalized eigenvalue problems for matrix pencils A - λ B or A B - λ I with positive definite real symmetric or Hermitian B.
• Can employ either of the two known convergence improvement techniques for large sparse problems: shift-invert and preconditioning.
• Can also compute singular values and vectors, and is actually an especially efficient tool for Principal Component Analysis (PCA) of dense data, owing to the high efficiency of matrix multiplications on modern multicore and GPU architectures.
• The user can specify the number of wanted eigenvalues
  • on either margin of the spectrum (e.g. 5 on the left one, 10 on the right one)
  • of largest magnitude
  • on either side of a given real value
  • nearest to a given real value
• If the number of eigenvalues needed is not known in advance (as is normally the case with PCA), the computation will continue until user-specified stopping criteria are satisfied (e.g. PCA approximation to the data is satisfactory).
• PCA capabilities include quick update of principal components after arrival of new data and incremental computation of principal components, one chunk of data at a time.
• The core solver is written in terms of abstract vectors, owing to which it will work on any architecture verbatim, provided that basic linear algebra operations on vectors are implemented (currently MKL and CUDA implementations are provided with the package, in the absence of these libraries Numpy algebra being used).

Dependencies

For best performance, install MKL 10.3 or later (or, on Windows, numpy+mkl). On Linux, the folder containing libmkl_rt.so must be in LD_LIBRARY_PATH. On Windows, the one containing mkl_rt.dll must be in PATH. Large sparse problems can only be solved if MKL is available, PCA and other dense problems can be solved without it.

To use GPU (which must be CUDA-enabled), NVIDIA GPU Computing Toolkit needs to be installed. On Linux, the folder containing libcudart.so must be in LD_LIBRARY_PATH.

Basic usage

Subfolder interfaces contains user-friendly SciPy-like interfaces to core solver working in terms of NumPy and SciPy data objects. Subfolder examples contains scripts illustrating their use, as well as a script illustrating basic capabilities of the core solver.

Documentation

Basic usage of the package is briefly described in docstrings of modules in folder interfaces (the best starting point to learn about RALEIGH usage) and example scripts. Advanced users will find the description of basic principles of RALEIGH's design in core/solver.py.

The mathematical and numerical aspects of the algorithm implemented by RALEIGH are described in the papers by E. E. Ovtchinnikov in J. Comput. Phys. 227:9477-9497 and SIAM Numer. Anal. 46:2567-2619.

Issues

Please use GitHub issue tracker to report bugs and request features.

License

RALEIGH is released under 3-clause BSD licence.