numpy-flint-arb 1.4.1

Arbitrary precision ball arithmetic (interval arithmetic) dtype in NumPy

numpy-flint-arb

---

Documentation: https://numpy-flint-arb.readthedocs.io

Source Code: https://github.com/34j/numpy-flint-arb

---

Arbitrary precision ball arithmetic (interval arithmetic) dtype in NumPy

Installation

Install this via pip (or your favourite package manager):

pip install numpy-flint-arb

Usage

Import numpy_flint_arb.np instead of numpy:

from numpy_flint_arb import np

A = np.random.normal(size=(2, 2))
b = np.random.normal(size=(2,))
x = np.linalg.solve(A, b)
b_approx = A @ x
assert np.all(np.contains(b_approx, b))

asarray() and Input Check

To avoid mixing ordinary floats like float or np.float, flarray for arb, acb only accepts integers, arb or acb and flarray for arf only accepts integers and arf, arb.

To relax this, allow_input() may be used:

import pytest
from numpy_flint_arb import allow_input
from flint import arb, arf

# arb array
with pytest.raises(Exception):
    np.asarray(0.5, dtype=arb)
with pytest.raises(Exception):
    with allow_input(float=True):
        np.asarray(0.5, dtype=arb)
with allow_input(interval=True, float=True):
    np.asarray(0.5, dtype=arb)

# arf array
with pytest.raises(Exception):
    np.asarray(0.5, dtype=arf)
with allow_input(float=True):
    np.asarray(0.5, dtype=arf)
with allow_input(interval=True, float=True):
    np.asarray(0.5, dtype=arf)

Note that allow_input() does not affect for arf(), arb(), acb() constructors, but only for np.asarray() and flarray().

str input

One can input str to asarray(). If dtype is not specified, it will be automatically detected as arb. However, specifying dtype explicitly is recommended.

asarray(dtype=acb)

asarray() does not support separated input for real and imaginary parts.

Do the following instead, as acb(1j) is exact.

from flint import arb, acb
from numpy_flint_arb import np

with pytest.raises(Exception):
    # python-flint does not support single argument str input for acb
    np.asarray("[0.5 +/- 0.001] + [0.5 +/- 0.001]j", dtype=acb)
with pytest.raises(Exception):
    # This is inexact and raises an error without allow_input()
    np.asarray(0.5 + 0.5j, dtype=acb)
with pytest.raises(Exception):
    # Mixing complex and arb is not supported by python-flint
    np.asarray("0.5 +/- 0.001", dtype=arb) + 1j * np.asarray("0.5 +/- 0.001", dtype=arb)

>>> # This is possible but not recommended
>>> np.asarray("0.5 +/- 0.001", dtype=arb) + 1j * np.asarray("0.5 +/- 0.001", dtype=acb)
flarray([0.50 +/- 1.01e-3] + [0.50 +/- 1.01e-3]j,
        dtype=<class 'flint.types.acb.acb'>)
>>> # Recommended
>>> np.asarray("0.5 +/- 0.001", dtype=arb) + acb(1j) * np.asarray("0.5 +/- 0.001", dtype=arb)
flarray([0.50 +/- 1.01e-3] + [0.50 +/- 1.01e-3]j,
        dtype=<class 'flint.types.acb.acb'>)

Randomness

Since python-flint does not support random number generation, the random module just uses np.random. Therefore, the return values may not be random up to the precision of arb, acb.

What it does

• This package adds a flarray which subclasses ndarray in order to

  • Override __array_namespace__ to numpy_flint_arb.np
  • Override dtype to return newly added _fl_dtype private attribute, since the actual internal dtype object cannot be overridden.
  • Override __array_finalize__ as recommended by the NumPy docs to return flarray with proper _fl_dtype instead of ndarray after Numpy operations.

• Partially supports linalg and (scipy.)special functions.

• Adds tomat() and frommat() to treat flarray as array of arb_mat / acb_mat, so that we can perform matrix operations like np.linalg.solve on flarray.

• Does not perform any parallelization to avoid complexity and to fully utilize the great python-flint library

  • Using arb_series and acb_series may be faster for additions but this is too hacky.
  • Defining custom dtype is way too complicated
  • Writing C extension would be theoretically also possible but is still too complicated.

• Does not support in operator since it tries to convert the return value to bool. Use newly added np.contains(x, y) and np.overlaps(x, y) instead.

• Currently dtype of resulting flarray is inferred from type(output.flat[0]). If the output array is empty or its element type is inconsistent, the result will be inaccurate.

  >>> (np.asarray([0], dtype=arb) * acb(1j)).dtype
  <class 'flint.types.acb.acb'>
  >>> (np.asarray([], dtype=arb) * acb(1j)).dtype
  <class 'flint.types.arb.arb'>
  

Contributors ✨

Thanks goes to these wonderful people (emoji key):

34j 💻 🤔 📖

This project follows the all-contributors specification. Contributions of any kind welcome!

Credits

This package was created with Copier and the browniebroke/pypackage-template project template.