roughly

 

 

About

A majority of algorithms in randomized numerical linear algebra are sequential and additive in nature. However, many implementations do not effectively exploit this structure. Often, once an insufficiently accurate result is observed, the computation is restarted from the beginning with a modified parameter set. This results in highly inefficient workflows.

The goal of roughly is to collect the most widespread algorithms of randomized numerical linear algebra and wrap them into an easy to use package where previous computations are stored in memory and available for the user to be reused.

This project is based on the doctoral course Advanced Scientific Programming in Python by Jochen Hinz.

Example

Suppose you need to compute a basis of the Krylov subspace

$$ \mathcal{K}^{k}(\boldsymbol{A}, \boldsymbol{x}) = \mathrm{span} \lbrace \boldsymbol{x}, \boldsymbol{A}\boldsymbol{x}, \boldsymbol{A}^2\boldsymbol{x}, \dots, \boldsymbol{A}^{k-1}\boldsymbol{x} \rbrace. $$

We do this by running $k$ iterations of the Arnoldi method:

import numpy as np
from roughly.approximate.krylov import ArnoldiDecomposition

arnoldi = ArnoldiDecomposition()

A = np.random.randn(100, 100)  # Example matrix
x = np.random.randn(100)  # Example starting vector
basis, _ = arnoldi.compute(A, x, k=10)

After $k$ iterations of the Arnoldi method you proceed with your computations, but realize your basis is not sufficient for these purposes. In these cases, roughly makes it easy to "refine" the approximation with additional iterations:

refined_basis, _ = arnoldi.refine(k=10)

The refine() attribute also makes convergence studies of the methods easier to compute.

Usage

To install this package, simply use

pip install git+https://github.com/FMatti/roughly.git

and then import it with

import roughly as rly

You can also test the package with pytest by running the command

pytest

in the root directory of the repository.

Features

Most implementations in roughly also work for linear operator only available as function handles instead of matrices. Currently, roughly implements the Arnoldi, Lanczos, and blocked versions of them; the randomized SVD and Nyström approximation; the randomized range sketch; and the Girard-Hutchinson, subspace projection, and Hutch++ algorithms.