baryrat 2.1.0

A Python package for barycentric rational approximation

Barycentric rational approximation

This is a pure Python package which provides routines for rational and polynomial approximation for real and complex functions through the so-called barycentric representation. The advantage of this representation is (often significantly) improved stability over classical approaches.

See the API documentation for an overview of the available functions.

Features

Best rational approximation using BRASIL

The package implements the novel BRASIL algorithm for best rational approximation; see the paper or the preprint to learn more.

The following example computes the best uniform rational approximation of degree 5 to a given function in the interval [0, pi]:

import numpy as np
import baryrat

def f(x): return np.sin(x) * np.exp(x)
r = baryrat.brasil(f, [0,np.pi], 5)

The rational function r can then be evaluated at arbitrary nodes, its poles computed, and more. See the documentation for details.

The AAA algorithm

The package includes a Python implementation of the AAA algorithm for rational approximation described in the paper "The AAA Algorithm for Rational Approximation" by Yuji Nakatsukasa, Olivier Sète, and Lloyd N. Trefethen, SIAM Journal on Scientific Computing 2018 40:3, A1494-A1522. (doi)

A MATLAB implementation of this algorithm is contained in Chebfun. The present Python version is a more or less direct port of the MATLAB version.

The "cleanup" feature for spurious poles and zeros is not currently implemented.

Further algorithms

The package includes functions for polynomial interpolation, rational interpolation with either fixed poles or fixed interpolation nodes, Floater-Hormann interpolation, and more.

Extended precision arithmetic

From baryrat 2.1 forward, most functions in the package support computing in extended precision using the gmpy2 package; linear algebra routines are provided through the flamp package.

To enable this, first install the flamp package:

pip install flamp

This will automatically install gmpy2 as well if it is not yet installed.

In your code, first set the desired number of decimal digits to compute with by

import flamp
flamp.set_dps(100)  # compute with 100 decimal digits precision

Arrays of numbers should be represented as numpy arrays with the object datatype containing gmpy2 floating point numbers. Some convenience functions to create such arrays are provided in flamp. For instance, use flamp.linspace(0, 1, 100) to create equispaced points in extended precision.

Most functions will autodetect if you pass such extended precision arrays and use the corresponding extended precision arithmetic in that case. There is also a use_mp flag for many functions, but it is only required to force the use of extended precision even when the inputs are in double precision.

Also the BarycentricRational class supports having its nodes, values, and weights stored in extended precision and will operate accordingly, for instance when computing the poles.

Installation

The package is implemented in pure Python and depends only on numpy and scipy, with gmpy2 and flamp as optional dependencies as discussed above. Install it using pip:

pip install baryrat

Usage

Here's an example of how to approximate a function in the interval [0,1] using the AAA algorithm:

import numpy as np
from baryrat import aaa

Z = np.linspace(0.0, 1.0, 1000)
F = np.exp(Z) * np.sin(2*np.pi*Z)

r = aaa(Z, F, mmax=10)

Instead of the maximum number of terms mmax, it's also possible to specify the error tolerance tol. Both arguments work exactly as in the MATLAB version.

The returned object r is an instance of the class baryrat.BarycentricRational and can be called like a function. For instance, you can compute the error on Z like this:

err = F - r(Z)
print(np.linalg.norm(err, np.inf))

If you are interested in the poles and residues of the computed rational function, you can query them like

pol, res = r.polres()

and the zeroes using

zer = r.zeros()

Finally, the nodes, values and weights used for interpolation (called zj, fj and wj in the original implementation) can be accessed as properties:

r.nodes
r.values
r.weights

Example: approximating the complex exponential

# create 9 interpolation nodes in a circle
n = 9
nodes = exp(arange(n) / n * 2j * pi)

# interpolate the complex exp function as a degree (4,4) rational function
r = baryrat.interpolate_rat(nodes, exp(nodes))
# compute poles and zeros
poles, zer = r.poles(), r.zeros()

# plot the approximation error and the nodes, poles and zeros
figsize(13.5, 5)

subplot(1, 2, 1)
Y, X = ogrid[-2:2:100j, -2:2:100j]
Z = X + 1j * Y
pcolormesh(X.flat, Y.flat, abs(r(Z) - exp(Z)), norm=mpl.colors.LogNorm())
colorbar();
axis('equal');

subplot(1, 2, 2)
scatter(real(nodes), imag(nodes))
scatter(real(poles), imag(poles), marker='x', c='r')
scatter(real(zer), imag(zer), marker='.', c='g')
axis('equal');

Citing baryrat

If you use this package in any published research, please cite the following publication where the package was first introduced:

• C. Hofreither. An algorithm for best rational approximation based on barycentric rational interpolation. Numerical Algorithms, 88(1):365--388, 2021. (doi)