flamp 1.0.1

Faster linear algebra with multiple precision

flamp - Faster linear algebra with multiple precision

flamp contains ports of many real and complex linear algebra routines from the mpmath package, but using numpy object arrays containing gmpy2 multiprecision floating point numbers instead of the mpmath floating point numbers. The resulting linear algebra routines are typically by a factor of 10x-15x faster than those in mpmath.

flamp is based on mpmath by Fredrik Johansson and mpmath contributors; in particular, the eigenvalue algorithms therein are by Timo Hartmann. flamp is BSD-licensed.

Installation

The package is written in pure Python and can simply be installed by

pip install flamp

Its only dependencies are numpy and gmpy2, both of which have pre-built packages readily available.

Benchmarks

Here are some timings for computing the eigenvalues and right eigenvectors of an n x n real matrix with prec binary digits. Results using mpmath, flamp and arb via the python-flint bindings are reported. Percentages are relative to the arb baseline.

n = 50 prec = 169
mpmath:  12.13s  (2766.1%)
flamp:    1.13s  (257.4%)
arb:      0.44s  (100.0%)

n = 100 prec = 169
mpmath:  90.55s  (2929.1%)
flamp:    8.55s  (276.4%)
arb:      3.09s  (100.0%)

n = 50 prec = 336
mpmath:  14.20s  (2073.9%)
flamp:    1.57s  (229.7%)
arb:      0.68s  (100.0%)

n = 100 prec = 336
mpmath: 106.47s  (2177.7%)
flamp:   11.59s  (237.1%)
arb:      4.89s  (100.0%)

Conclusion: for raw speed, use arb which is written in pure C. If you want a pure Python package which provides reasonable performance (usually within a factor of ~2.5x of arb), use flamp.

List of functions

The following is a list of supported functions in the flamp module by category. All matrix and vector arguments should be supplied as numpy arrays of gmpy2 numbers, although standard floating point numpy arrays will be automatically converted in most cases.

Refer to the docstrings for further information.

Linear algebra

These behave essentially like the corresponding functions in mpmath, with some slight modifications. For instance, all functions for solving linear systems accept either a single vector or an array of vectors for the right-hand side.

• lu_solve(A, b) - solve a linear system using LU decomposition
• qr_solve(A, b) - solve a linear system using QR decomposition
• cholesky_solve(A, b) - solve a symmetric positive definite system using Cholesky decomposition
• L_solve(L, b, unit_diag=False) - solve a lower triangular system
• U_solve(U, y) - solve an upper triangular system
• inverse(A) - compute inverse of a square matrix
• det(A) - compute determinant of a square matrix
• lu(A) - compute LU decomposition of a square matrix
• qr(A, mode='full') - compute QR decomposition of a matrix; mode=full, reduced, raw
• cholesky(A) - compute Cholesky decomposition of a symmetric positive definite matrix
• eig(A, left=False, right=True) - compute eigenvalues and (optionally) left and right eigenvectors of a matrix
• eigh(A, eigvals_only=False) - compute eigenvalues and (optionally) the orthonormal eigenvectors of a real symmetric or complex Hermitian square matrix
• hessenberg(A) - compute Hessenberg decomposition (Q, R) of a square matrix
• schur(A) - compute Schur decomposition of a square matrix
• svd(A, full_matrices=False, compute_uv=True) - compute singular value decomposition (singular values and optionally the left and right singular vectors) of a matrix

Array functions

Most of these behave essentially like their numpy counterparts, but work with gmpy2 extended precision numbers.

• zeros(shape)
• ones(shape)
• empty(shape)
• eye(n)
• linspace(start, stop, num, endpoint=True)
• vector_norm(x) - computes Euclidean norm of a vector
• to_mp(A) - converts an arbitrary numpy array (or list/tuple) into an array of gmpy2 numbers, copying the input

Utility functions

These functions are used to manipulate the working precision of the gmpy2 library.

• prec_to_dps(n) - number of accurate decimals that can be represented with a precision of n bits
• dps_to_prec(n) - number of bits required to represent n decimals accurately
• get_precision() - get the current precision in binary digits
• set_precision(prec) - set the working precision in binary digits
• get_dps() - get the current precision in decimal digits (approximate)
• set_dps(dps) - set the working precision in decimal digits (approximate)
• extraprec(n) - returns a context manager (for use in a with statement) which temporarily increases the working precision by the given amount

Array-aware special functions

These functions work much like the corresponding functions in numpy in that they automatically distribute over numpy arrays while computing in extended precision.

• exp(x)
• sqrt(x)
• sin(x)
• cos(x)
• tan(x)
• sinh(x)
• cosh(x)
• tanh(x)
• square(x)
• log(x)
• log2(x)
• log10(x)