scipy 1.16.2__cp312-cp312-win_arm64.whl

scipy/integrate/_cubature.py

@@ -0,0 +1,729 @@
+import math
+import heapq
+import itertools
+
+from dataclasses import dataclass, field
+from types import ModuleType
+from typing import Any, TypeAlias
+
+from scipy._lib._array_api import (
+    array_namespace,
+    xp_size,
+    xp_copy,
+    xp_promote
+)
+from scipy._lib._util import MapWrapper
+
+from scipy.integrate._rules import (
+    ProductNestedFixed,
+    GaussKronrodQuadrature,
+    GenzMalikCubature,
+)
+from scipy.integrate._rules._base import _split_subregion
+
+__all__ = ['cubature']
+
+Array: TypeAlias = Any  # To be changed to an array-api-typing Protocol later
+
+
+@dataclass
+class CubatureRegion:
+    estimate: Array
+    error: Array
+    a: Array
+    b: Array
+    _xp: ModuleType = field(repr=False)
+
+    def __lt__(self, other):
+        # Consider regions with higher error estimates as being "less than" regions with
+        # lower order estimates, so that regions with high error estimates are placed at
+        # the top of the heap.
+
+        this_err = self._xp.max(self._xp.abs(self.error))
+        other_err = self._xp.max(self._xp.abs(other.error))
+
+        return this_err > other_err
+
+
+@dataclass
+class CubatureResult:
+    estimate: Array
+    error: Array
+    status: str
+    regions: list[CubatureRegion]
+    subdivisions: int
+    atol: float
+    rtol: float
+
+
+def cubature(f, a, b, *, rule="gk21", rtol=1e-8, atol=0, max_subdivisions=10000,
+             args=(), workers=1, points=None):
+    r"""
+    Adaptive cubature of multidimensional array-valued function.
+
+    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 :class:`python:multiprocessing.pool.Pool`).
+        Supply `-1` to use all cores available to the Process. Alternatively, supply a
+        map-like callable, such as :meth:`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 : ndarray
+            Estimate of the value of the integral over the overall region specified.
+        error : ndarray
+            Estimate of the error of the approximation over the overall region
+            specified.
+        status : str
+            Whether the estimation was successful. Can be either: "converged",
+            "not_converged".
+        subdivisions : int
+            Number of subdivisions performed.
+        atol, rtol : float
+            Requested tolerances for the approximation.
+        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 : ndarray
+            Points describing the corners of the region. If the original integral
+            contained infinite limits or was over a region described by `region`,
+            then `a` and `b` are in the transformed coordinates.
+        estimate : ndarray
+            Estimate of the value of the integral over this region.
+        error : ndarray
+            Estimate of the error of the approximation over this region.
+
+    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 :math:`a_i = -\infty` and :math:`b_i = \infty`, the i-th integration variable
+    will use the transformation :math:`x = \frac{1-|t|}{t}` and :math:`t \in (-1, 1)`.
+
+    If :math:`a_i \ne \pm\infty` and :math:`b_i = \infty`, the i-th integration variable
+    will use the transformation :math:`x = a_i + \frac{1-t}{t}` and
+    :math:`t \in (0, 1)`.
+
+    If :math:`a_i = -\infty` and :math:`b_i \ne \pm\infty`, the i-th integration
+    variable will use the transformation :math:`x = b_i - \frac{1-t}{t}` and
+    :math:`t \in (0, 1)`.
+
+    References
+    ----------
+    .. [1] R. Piessens, E. de Doncker, Quadpack: A Subroutine Package for Automatic
+        Integration, files: dqk21.f, dqk15.f (1983).
+
+    .. [2] A.C. Genz, A.A. Malik, Remarks on algorithm 006: An adaptive algorithm for
+        numerical integration over an N-dimensional rectangular region, Journal of
+        Computational and Applied Mathematics, Volume 6, Issue 4, 1980, Pages 295-302,
+        ISSN 0377-0427
+        :doi:`10.1016/0771-050X(80)90039-X`
+
+    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
+     array([1.        , 0.5       , 0.33333333, 0.25      , 0.2       ,
+            0.16666667, 0.14285714, 0.125     , 0.11111111, 0.1       ])
+
+    **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
+     array([[-0.79812452,  0.35246913, -0.52273628],
+            [ 0.88392779,  0.59139899,  0.41895111]])
+
+    **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
+     array([[-0.79812452,  0.35246913, -0.52273628],
+            [ 0.88392779,  0.59139899,  0.41895111]])
+
+    **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
+     3.1415926
+
+    **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
+     1.8921661
+    """
+
+    # It is also possible to use a custom rule, but this is not yet part of the public
+    # API. An example of this can be found in the class scipy.integrate._rules.Rule.
+
+    xp = array_namespace(a, b)
+    max_subdivisions = float("inf") if max_subdivisions is None else max_subdivisions
+    points = [] if points is None else points
+
+    # Convert a and b to arrays and convert each point in points to an array, promoting
+    # each to a common floating dtype.
+    a, b, *points = xp_promote(a, b, *points, broadcast=True, force_floating=True,
+                               xp=xp)
+    result_dtype = a.dtype
+
+    if xp_size(a) == 0 or xp_size(b) == 0:
+        raise ValueError("`a` and `b` must be nonempty")
+
+    if a.ndim != 1 or b.ndim != 1:
+        raise ValueError("`a` and `b` must be 1D arrays")
+
+    # If the rule is a string, convert to a corresponding product rule
+    if isinstance(rule, str):
+        ndim = xp_size(a)
+
+        if rule == "genz-malik":
+            rule = GenzMalikCubature(ndim, xp=xp)
+        else:
+            quadratues = {
+                "gauss-kronrod": GaussKronrodQuadrature(21, xp=xp),
+
+                # Also allow names quad_vec uses:
+                "gk21": GaussKronrodQuadrature(21, xp=xp),
+                "gk15": GaussKronrodQuadrature(15, xp=xp),
+            }
+
+            base_rule = quadratues.get(rule)
+
+            if base_rule is None:
+                raise ValueError(f"unknown rule {rule}")
+
+            rule = ProductNestedFixed([base_rule] * ndim)
+
+    # If any of limits are the wrong way around (a > b), flip them and keep track of
+    # the sign.
+    sign = (-1) ** xp.sum(xp.astype(a > b, xp.int8), dtype=result_dtype)
+
+    a_flipped = xp.min(xp.stack([a, b]), axis=0)
+    b_flipped = xp.max(xp.stack([a, b]), axis=0)
+
+    a, b = a_flipped, b_flipped
+
+    # If any of the limits are infinite, apply a transformation
+    if xp.any(xp.isinf(a)) or xp.any(xp.isinf(b)):
+        f = _InfiniteLimitsTransform(f, a, b, xp=xp)
+        a, b = f.transformed_limits
+
+        # Map points from the original coordinates to the new transformed coordinates.
+        #
+        # `points` is a list of arrays of shape (ndim,), but transformations are applied
+        # to arrays of shape (npoints, ndim).
+        #
+        # It is not possible to combine all the points into one array and then apply
+        # f.inv to all of them at once since `points` needs to remain iterable.
+        # Instead, each point is reshaped to an array of shape (1, ndim), `f.inv` is
+        # applied, and then each is reshaped back to (ndim,).
+        points = [xp.reshape(point, (1, -1)) for point in points]
+        points = [f.inv(point) for point in points]
+        points = [xp.reshape(point, (-1,)) for point in points]
+
+        # Include any problematic points introduced by the transformation
+        points.extend(f.points)
+
+    # If any problematic points are specified, divide the initial region so that these
+    # points lie on the edge of a subregion.
+    #
+    # This means ``f`` won't be evaluated there if the rule being used has no evaluation
+    # points on the boundary.
+    if len(points) == 0:
+        initial_regions = [(a, b)]
+    else:
+        initial_regions = _split_region_at_points(a, b, points, xp)
+
+    regions = []
+    est = 0.0
+    err = 0.0
+
+    for a_k, b_k in initial_regions:
+        est_k = rule.estimate(f, a_k, b_k, args)
+        err_k = rule.estimate_error(f, a_k, b_k, args)
+        regions.append(CubatureRegion(est_k, err_k, a_k, b_k, xp))
+
+        est += est_k
+        err += err_k
+
+    subdivisions = 0
+    success = True
+
+    with MapWrapper(workers) as mapwrapper:
+        while xp.any(err > atol + rtol * xp.abs(est)):
+            # region_k is the region with highest estimated error
+            region_k = heapq.heappop(regions)
+
+            est_k = region_k.estimate
+            err_k = region_k.error
+
+            a_k, b_k = region_k.a, region_k.b
+
+            # Subtract the estimate of the integral and its error over this region from
+            # the current global estimates, since these will be refined in the loop over
+            # all subregions.
+            est -= est_k
+            err -= err_k
+
+            # Find all 2^ndim subregions formed by splitting region_k along each axis,
+            # e.g. for 1D integrals this splits an estimate over an interval into an
+            # estimate over two subintervals, for 3D integrals this splits an estimate
+            # over a cube into 8 subcubes.
+            #
+            # For each of the new subregions, calculate an estimate for the integral and
+            # the error there, and push these regions onto the heap for potential
+            # further subdividing.
+
+            executor_args = zip(
+                itertools.repeat(f),
+                itertools.repeat(rule),
+                itertools.repeat(args),
+                _split_subregion(a_k, b_k, xp),
+            )
+
+            for subdivision_result in mapwrapper(_process_subregion, executor_args):
+                a_k_sub, b_k_sub, est_sub, err_sub = subdivision_result
+
+                est += est_sub
+                err += err_sub
+
+                new_region = CubatureRegion(est_sub, err_sub, a_k_sub, b_k_sub, xp)
+
+                heapq.heappush(regions, new_region)
+
+            subdivisions += 1
+
+            if subdivisions >= max_subdivisions:
+                success = False
+                break
+
+        status = "converged" if success else "not_converged"
+
+        # Apply sign change to handle any limits which were initially flipped.
+        est = sign * est
+
+        return CubatureResult(
+            estimate=est,
+            error=err,
+            status=status,
+            subdivisions=subdivisions,
+            regions=regions,
+            atol=atol,
+            rtol=rtol,
+        )
+
+
+def _process_subregion(data):
+    f, rule, args, coord = data
+    a_k_sub, b_k_sub = coord
+
+    est_sub = rule.estimate(f, a_k_sub, b_k_sub, args)
+    err_sub = rule.estimate_error(f, a_k_sub, b_k_sub, args)
+
+    return a_k_sub, b_k_sub, est_sub, err_sub
+
+
+def _is_strictly_in_region(a, b, point, xp):
+    if xp.all(point == a) or xp.all(point == b):
+        return False
+
+    return xp.all(a <= point) and xp.all(point <= b)
+
+
+def _split_region_at_points(a, b, points, xp):
+    """
+    Given the integration limits `a` and `b` describing a rectangular region and a list
+    of `points`, find the list of ``[(a_1, b_1), ..., (a_l, b_l)]`` which breaks up the
+    initial region into smaller subregion such that no `points` lie strictly inside
+    any of the subregions.
+    """
+
+    regions = [(a, b)]
+
+    for point in points:
+        if xp.any(xp.isinf(point)):
+            # If a point is specified at infinity, ignore.
+            #
+            # This case occurs when points are given by the user to avoid, but after
+            # applying a transformation, they are removed.
+            continue
+
+        new_subregions = []
+
+        for a_k, b_k in regions:
+            if _is_strictly_in_region(a_k, b_k, point, xp):
+                subregions = _split_subregion(a_k, b_k, xp, point)
+
+                for left, right in subregions:
+                    # Skip any zero-width regions.
+                    if xp.any(left == right):
+                        continue
+                    else:
+                        new_subregions.append((left, right))
+
+                new_subregions.extend(subregions)
+
+            else:
+                new_subregions.append((a_k, b_k))
+
+        regions = new_subregions
+
+    return regions
+
+
+class _VariableTransform:
+    """
+    A transformation that can be applied to an integral.
+    """
+
+    @property
+    def transformed_limits(self):
+        """
+        New limits of integration after applying the transformation.
+        """
+
+        raise NotImplementedError
+
+    @property
+    def points(self):
+        """
+        Any problematic points introduced by the transformation.
+
+        These should be specified as points where ``_VariableTransform(f)(self, point)``
+        would be problematic.
+
+        For example, if the transformation ``x = 1/((1-t)(1+t))`` is applied to a
+        univariate integral, then points should return ``[ [1], [-1] ]``.
+        """
+
+        return []
+
+    def inv(self, x):
+        """
+        Map points ``x`` to ``t`` such that if ``f`` is the original function and ``g``
+        is the function after the transformation is applied, then::
+
+            f(x) = g(self.inv(x))
+        """
+
+        raise NotImplementedError
+
+    def __call__(self, t, *args, **kwargs):
+        """
+        Apply the transformation to ``f`` and multiply by the Jacobian determinant.
+        This should be the new integrand after the transformation has been applied so
+        that the following is satisfied::
+
+            f_transformed = _VariableTransform(f)
+
+            cubature(f, a, b) == cubature(
+                f_transformed,
+                *f_transformed.transformed_limits(a, b),
+            )
+        """
+
+        raise NotImplementedError
+
+
+class _InfiniteLimitsTransform(_VariableTransform):
+    r"""
+    Transformation for handling infinite limits.
+
+    Assuming ``a = [a_1, ..., a_n]`` and ``b = [b_1, ..., b_n]``:
+
+    If :math:`a_i = -\infty` and :math:`b_i = \infty`, the i-th integration variable
+    will use the transformation :math:`x = \frac{1-|t|}{t}` and :math:`t \in (-1, 1)`.
+
+    If :math:`a_i \ne \pm\infty` and :math:`b_i = \infty`, the i-th integration variable
+    will use the transformation :math:`x = a_i + \frac{1-t}{t}` and
+    :math:`t \in (0, 1)`.
+
+    If :math:`a_i = -\infty` and :math:`b_i \ne \pm\infty`, the i-th integration
+    variable will use the transformation :math:`x = b_i - \frac{1-t}{t}` and
+    :math:`t \in (0, 1)`.
+    """
+
+    def __init__(self, f, a, b, xp):
+        self._xp = xp
+
+        self._f = f
+        self._orig_a = a
+        self._orig_b = b
+
+        # (-oo, oo) will be mapped to (-1, 1).
+        self._double_inf_pos = (a == -math.inf) & (b == math.inf)
+
+        # (start, oo) will be mapped to (0, 1).
+        start_inf_mask = (a != -math.inf) & (b == math.inf)
+
+        # (-oo, end) will be mapped to (0, 1).
+        inf_end_mask = (a == -math.inf) & (b != math.inf)
+
+        # This is handled by making the transformation t = -x and reducing it to
+        # the other semi-infinite case.
+        self._semi_inf_pos = start_inf_mask | inf_end_mask
+
+        # Since we flip the limits, we don't need to separately multiply the
+        # integrand by -1.
+        self._orig_a[inf_end_mask] = -b[inf_end_mask]
+        self._orig_b[inf_end_mask] = -a[inf_end_mask]
+
+        self._num_inf = self._xp.sum(
+            self._xp.astype(self._double_inf_pos | self._semi_inf_pos, self._xp.int64),
+        ).__int__()
+
+    @property
+    def transformed_limits(self):
+        a = xp_copy(self._orig_a)
+        b = xp_copy(self._orig_b)
+
+        a[self._double_inf_pos] = -1
+        b[self._double_inf_pos] = 1
+
+        a[self._semi_inf_pos] = 0
+        b[self._semi_inf_pos] = 1
+
+        return a, b
+
+    @property
+    def points(self):
+        # If there are infinite limits, then the origin becomes a problematic point
+        # due to a division by zero there.
+
+        # If the function using this class only wraps f when a and b contain infinite
+        # limits, this condition will always be met (as is the case with cubature).
+        #
+        # If a and b do not contain infinite limits but f is still wrapped with this
+        # class, then without this condition the initial region of integration will
+        # be split around the origin unnecessarily.
+        if self._num_inf != 0:
+            return [self._xp.zeros(self._orig_a.shape)]
+        else:
+            return []
+
+    def inv(self, x):
+        t = xp_copy(x)
+        npoints = x.shape[0]
+
+        double_inf_mask = self._xp.tile(
+            self._double_inf_pos[self._xp.newaxis, :],
+            (npoints, 1),
+        )
+
+        semi_inf_mask = self._xp.tile(
+            self._semi_inf_pos[self._xp.newaxis, :],
+            (npoints, 1),
+        )
+
+        # If any components of x are 0, then this component will be mapped to infinity
+        # under the transformation used for doubly-infinite limits.
+        #
+        # Handle the zero values and non-zero values separately to avoid division by
+        # zero.
+        zero_mask = x[double_inf_mask] == 0
+        non_zero_mask = double_inf_mask & ~zero_mask
+        t[zero_mask] = math.inf
+        t[non_zero_mask] = 1/(x[non_zero_mask] + self._xp.sign(x[non_zero_mask]))
+
+        start = self._xp.tile(self._orig_a[self._semi_inf_pos], (npoints,))
+        t[semi_inf_mask] = 1/(x[semi_inf_mask] - start + 1)
+
+        return t
+
+    def __call__(self, t, *args, **kwargs):
+        x = xp_copy(t)
+        npoints = t.shape[0]
+
+        double_inf_mask = self._xp.tile(
+            self._double_inf_pos[self._xp.newaxis, :],
+            (npoints, 1),
+        )
+
+        semi_inf_mask = self._xp.tile(
+            self._semi_inf_pos[self._xp.newaxis, :],
+            (npoints, 1),
+        )
+
+        # For (-oo, oo) -> (-1, 1), use the transformation x = (1-|t|)/t.
+        x[double_inf_mask] = (
+            (1 - self._xp.abs(t[double_inf_mask])) / t[double_inf_mask]
+        )
+
+        start = self._xp.tile(self._orig_a[self._semi_inf_pos], (npoints,))
+
+        # For (start, oo) -> (0, 1), use the transformation x = start + (1-t)/t.
+        x[semi_inf_mask] = start + (1 - t[semi_inf_mask]) / t[semi_inf_mask]
+
+        jacobian_det = 1/self._xp.prod(
+            self._xp.reshape(
+                t[semi_inf_mask | double_inf_mask]**2,
+                (-1, self._num_inf),
+            ),
+            axis=-1,
+        )
+
+        f_x = self._f(x, *args, **kwargs)
+        jacobian_det = self._xp.reshape(jacobian_det, (-1, *([1]*(len(f_x.shape) - 1))))
+
+        return f_x * jacobian_det