bundles / scipy latest / scipy / integrate / _cubature / cubature

function

`scipy.integrate._cubature:cubature`

source: /scipy/integrate/_cubature.py :60

Signature

   def     cubature (    f    ,    a    ,    b    ,  * ,     rule    =  gk21   ,    rtol    =  1e-08   ,    atol    =  0   ,    max_subdivisions    =  10000   ,    args    =  ()   ,    workers    =  1   ,    points    =  None     )

Summary

Adaptive cubature of multidimensional array-valued function.

Extended Summary

Given an arbitrary integration rule, this function returns an estimate of the integral to the requested tolerance over the region defined by the arrays a and b specifying the corners of a hypercube.

Convergence is not guaranteed for all integrals.

Parameters

f : callable

Function to integrate. f must have the signature

f(x : ndarray, *args) -> ndarray

f should accept arrays x of shape

(npoints, ndim)

and output arrays of shape

(npoints, output_dim_1, ..., output_dim_n)

In this case, cubature will return arrays of shape

(output_dim_1, ..., output_dim_n)

a, b : array_like

Lower and upper limits of integration as 1D arrays specifying the left and right endpoints of the intervals being integrated over. Limits can be infinite.

rule : str, optional

Rule used to estimate the integral. If passing a string, the options are "gauss-kronrod" (21 node), or "genz-malik" (degree 7). If a rule like "gauss-kronrod" is specified for an n-dim integrand, the corresponding Cartesian product rule is used. "gk21", "gk15" are also supported for compatibility with quad_vec. See Notes.

rtol, atol : float, optional

Relative and absolute tolerances. Iterations are performed until the error is estimated to be less than atol + rtol * abs(est). Here rtol controls relative accuracy (number of correct digits), while atol controls absolute accuracy (number of correct decimal places). To achieve the desired rtol, set atol to be smaller than the smallest value that can be expected from rtol * abs(y) so that rtol dominates the allowable error. If atol is larger than rtol * abs(y) the number of correct digits is not guaranteed. Conversely, to achieve the desired atol, set rtol such that rtol * abs(y) is always smaller than atol. Default values are 1e-8 for rtol and 0 for atol.

max_subdivisions : int, optional

Upper bound on the number of subdivisions to perform. Default is 10,000.

args : tuple, optional

Additional positional args passed to f, if any.

workers : int or map-like callable, optional

If workers is an integer, part of the computation is done in parallel subdivided to this many tasks (using python:multiprocessing.pool.Pool). Supply -1 to use all cores available to the Process. Alternatively, supply a map-like callable, such as python:multiprocessing.pool.Pool.map for evaluating the population in parallel. This evaluation is carried out as workers(func, iterable).

points : list of array_like, optional

List of points to avoid evaluating f at, under the condition that the rule being used does not evaluate f on the boundary of a region (which is the case for all Genz-Malik and Gauss-Kronrod rules). This can be useful if f has a singularity at the specified point. This should be a list of array-likes where each element has length ndim. Default is empty. See Examples.

Returns

res : object

Object containing the results of the estimation. It has the following attributes:

estimate: estimate
error: error
status: status
subdivisions: subdivisions
atol,rtol: atol, rtol
regions: list of object: List of objects containing the estimates of the integral over smaller regions of the domain.

Each object in regions has the following attributes:

a,b: a, b
estimate: estimate
error: error

Notes

The algorithm uses a similar algorithm to quad_vec, which itself is based on the implementation of QUADPACK's DQAG* algorithms, implementing global error control and adaptive subdivision.

The source of the nodes and weights used for Gauss-Kronrod quadrature can be found in ^[1], and the algorithm for calculating the nodes and weights in Genz-Malik cubature can be found in ^[2].

The rules currently supported via the rule argument are:

"gauss-kronrod", 21-node Gauss-Kronrod
"genz-malik", n-node Genz-Malik

If using Gauss-Kronrod for an n-dim integrand where n > 2, then the corresponding Cartesian product rule will be found by taking the Cartesian product of the nodes in the 1D case. This means that the number of nodes scales exponentially as 21^n in the Gauss-Kronrod case, which may be problematic in a moderate number of dimensions.

Genz-Malik is typically less accurate than Gauss-Kronrod but has much fewer nodes, so in this situation using "genz-malik" might be preferable.

Infinite limits are handled with an appropriate variable transformation. Assuming a = [a_1, ..., a_n] and b = [b_1, ..., b_n]:

If $a_{i} = - \infty$ and $b_{i} = \infty$ , the i-th integration variable will use the transformation $x = \frac{1 - ∣ t ∣}{t}$ and $t \in (- 1, 1)$ .

If $a_{i} \neq = \pm \infty$ and $b_{i} = \infty$ , the i-th integration variable will use the transformation $x = a_{i} + \frac{1 - t}{t}$ and $t \in (0, 1)$ .

If $a_{i} = - \infty$ and $b_{i} \neq = \pm \infty$ , the i-th integration variable will use the transformation $x = b_{i} - \frac{1 - t}{t}$ and $t \in (0, 1)$ .

Array API Standard Support

cubature has experimental support for Python Array API Standard compatible backends in addition to NumPy. Please consider testing these features by setting an environment variable SCIPY_ARRAY_API=1 and providing CuPy, PyTorch, JAX, or Dask arrays as array arguments. The following combinations of backend and device (or other capability) are supported.

====================  ====================  ====================
Library               CPU                   GPU
====================  ====================  ====================
NumPy                 ✅                     n/a                 
CuPy                  n/a                   ✅                   
PyTorch               ✅                     ✅                   
JAX                   ⚠️ no JIT             ⚠️ no JIT           
Dask                  ⚠️ computes graph     n/a                 
====================  ====================  ====================

See dev-arrayapi for more information.

Examples

**1D integral with vector output**: .. math:: \int^1_0 \mathbf f(x) \text dx Where ``f(x) = x^n`` and ``n = np.arange(10)`` is a vector. Since no rule is specified, the default "gk21" is used, which corresponds to Gauss-Kronrod integration with 21 nodes.

import numpy as np
from scipy.integrate import cubature
def f(x, n):
   # Make sure x and n are broadcastable
   return x[:, np.newaxis]**n[np.newaxis, :]
res = cubature(
    f,
    a=[0],
    b=[1],
    args=(np.arange(10),),
)

✓

res.estimate

✗

**7D integral with arbitrary-shaped array output**:: f(x) = cos(2*pi*r + alphas @ x) for some ``r`` and ``alphas``, and the integral is performed over the unit hybercube, :math:`[0, 1]^7`. Since the integral is in a moderate number of dimensions, "genz-malik" is used rather than the default "gauss-kronrod" to avoid constructing a product rule with :math:`21^7 \approx 2 \times 10^9` nodes.

import numpy as np
from scipy.integrate import cubature
def f(x, r, alphas):
    # f(x) = cos(2*pi*r + alphas @ x)
    # Need to allow r and alphas to be arbitrary shape
    npoints, ndim = x.shape[0], x.shape[-1]
    alphas = alphas[np.newaxis, ...]
    x = x.reshape(npoints, *([1]*(len(alphas.shape) - 1)), ndim)
    return np.cos(2*np.pi*r + np.sum(alphas * x, axis=-1))
rng = np.random.default_rng()
r, alphas = rng.random((2, 3)), rng.random((2, 3, 7))
res = cubature(
    f=f,
    a=np.array([0, 0, 0, 0, 0, 0, 0]),
    b=np.array([1, 1, 1, 1, 1, 1, 1]),
    rtol=1e-5,
    rule="genz-malik",
    args=(r, alphas),
)

✓

res.estimate

✗

**Parallel computation with** `workers`:

from concurrent.futures import ThreadPoolExecutor
with ThreadPoolExecutor() as executor:
    res = cubature(
        f=f,
        a=np.array([0, 0, 0, 0, 0, 0, 0]),
        b=np.array([1, 1, 1, 1, 1, 1, 1]),
        rtol=1e-5,
        rule="genz-malik",
        args=(r, alphas),
        workers=executor.map,
     )

✓

res.estimate

✗

**2D integral with infinite limits**: .. math:: \int^{ \infty }_{ -\infty } \int^{ \infty }_{ -\infty } e^{-x^2-y^2} \text dy \text dx

def gaussian(x):
    return np.exp(-np.sum(x**2, axis=-1))
res = cubature(gaussian, [-np.inf, -np.inf], [np.inf, np.inf])

✓

res.estimate

✗

**1D integral with singularities avoided using** `points`: .. math:: \int^{ 1 }_{ -1 } \frac{\sin(x)}{x} \text dx It is necessary to use the `points` parameter to avoid evaluating `f` at the origin.

def sinc(x):
    return np.sin(x)/x
res = cubature(sinc, [-1], [1], points=[[0]])

✓

res.estimate

✗

Aliases

scipy.integrate.cubature

Referenced by

This package

release:1.15.0-notes