ival for Python

Interval arithmetic over NumPy arrays: every result is the tightest interval of doubles around the exact one (IEEE 1788.1). It calls crnumerics' C library libival, whose functions are CORE-MATH's, correctly rounded.

import ival, numpy as np

x = ival.Interval(1, 2)
ival.exp(x) * x + 1                # Interval(3.718281828459045, 15.7781121978613)
ival.Interval(0.1) + 0.2           # Interval(0.3, 0.30000000000000004): the exact sum inside
ival.Interval.from_text("[0.1]")   # Interval(0.09999999999999999, 0.1): the decimal 0.1 inside
X = ival.Interval(np.linspace(0, 1, 10**6), np.linspace(0, 1, 10**6) + 1e-3)
ival.sin(X)                        # a million intervals in one call

Installing

On Linux x86-64 and aarch64 (glibc 2.26 or later), the wheels on PyPI carry the library, so nothing else needs installing. The package is crival on PyPI; the module is ival:

pip install crival
python3 -c 'import ival; print(ival.exp(ival.Interval(1)))'

Elsewhere, or to use a libival you installed, install libival first, from the root of crnumerics, then the package:

make install-ival PREFIX=$HOME/.local     # or /usr/local; installs libival.so.0 and ival.pc
pip install ./ival/python                 # numpy is the one dependency

ival/python/build-wheel.sh builds a wheel that carries the library, for the glibc it was built on and newer. The wheels on PyPI are built in the manylinux_2_28 containers by .github/workflows/wheels.yml.

The binding finds the library through IVAL_LIBRARY (a path), then a copy inside the package (a wheel's), then the system's linker paths, then pkg-config --variable=libdir ival (for a prefix like $HOME/.local, PKG_CONFIG_PATH must include its lib/pkgconfig). ival.library_path says which it loaded, and ival.version() the library's version. It works with libival 0.1.0 and later; 0.2.0 fixes sinRev, cosRev and tanRev, which in 0.1.0 could miss a set. Python 3.9 or later; tested on Linux x86-64 and aarch64.

What there is

• Interval(lo, hi), or Interval(a) for the smallest interval holding a. One interval holds two Python floats; an array of them two float64 arrays, broadcast as NumPy does. The empty interval is [nan, nan]; Interval(lo, hi) is empty unless lo <= hi (1788's numsToInterval). Interval.from_text("[0.1, 0.2]") reads 1788's text forms, a string or an array of strings, rounding each decimal outward.
• Numbers that are not doubles round outward. A Python int beyond 2^53, a Fraction, a Decimal, a long double, an int64 array: each becomes the two doubles around it, never the nearest one, as an end, as an operand (x + Fraction(1, 3)) or as a member to test (2**53 + 1 in x). A float is taken as the double it is, so 0.1 is already rounded: give decimals as text or Decimal.
• Arithmetic: + - * /, unary -, abs, ** (an integer power is pown, tight; any other pow), & (intersection), | (hull).
• 40 functions: exp exp2 exp10 expm1 log log2 log10 log1p sqrt cbrt rsqrt sin cos tan sinpi cospi tanpi asin acos atan asinpi acospi atanpi sinh cosh tanh asinh acosh atanh erf erfc tgamma sqr recip sign ceil floor trunc round roundeven, and atan2 hypot pow min max cancelminus cancelplus pown rootn fma.
• The accurate mode: acc_exp, acc_sin, ... (the first 32 above, and acc_atan2, acc_hypot, acc_pow) are at most one ulp wider than the tight result, and faster on arrays, when the environment variable IVAL_CRMVEC names crmvec's libmvec.so.1 (x86-64 with AVX2 and FMA). Without it they give the tight result.
• Reverse operations (the tightest interval around {x in X : f(x) in C}, X the whole line if omitted): sqrrev absrev coshrev sinrev cosrev tanrev pownrev mulrev powrev1 powrev2.
• Measures and tests: .inf .sup .mid .wid .rad .mag .mig, is_empty() is_entire() is_singleton(), subset interior disjoint equal, contains(X, v) and v in X.

The rounding mode is the library's business: it sets what it needs and restores the caller's, so nothing here changes how the rest of a program rounds. A call releases the GIL, and the library is thread-safe.

Speed, and pyinterval

pyinterval (on CRlibm) was Python's correctly rounded interval library. Its last release was in March 2017, and pip install pyinterval fails on Python 3.12: its crlibm imports distutils.command.upload, which Python 3.12 lacks and setuptools 75 dropped. Built with setuptools 74 and no build isolation, it and this binding were run on the same inputs by bench/pyinterval_compare.py, on an AMD Ryzen 5 PRO 5650U, 2026-10-09:

• Bounds: on 2,000 points each of exp, log, sin, cos, tan, atan, sqrt, sinh, cosh, expm1 and log1p, both tight and the same, bound for bound. pyinterval's tanh is a quotient of sinh and cosh, every bound rounded outward: valid, but not tight on any of the 2,000 points, up to 4 doubles too wide.

• Time per interval, for intervals of width 0.5 (and 1, for b):

  	pyinterval	ival, one at a time	ival, an array of 20,000
  exp	11.0 µs	1.8 µs	25 ns
  sin	11.7 µs	1.8 µs	141 ns
  atan	10.8 µs	1.7 µs	31 ns
  a*b + 2	22.2 µs	4.5 µs	20 ns

  One interval at a time costs about a microsecond of ctypes per call; arrays pay it once per call.

Tests

make ival-python-check at the root of crnumerics runs tests/ against the library just built (it needs pytest and mpmath): bounds against mpmath at 300 bits and exact fractions, arrays against one interval at a time, outward rounding of numbers that are not doubles, the reverse operations and the accurate mode.