ENH: add quantile function with weights support

src/array_api_extra/__init__.py

@@ -11,6 +11,7 @@
     one_hot,
     pad,
     partition,
+    quantile,
     sinc,
 )
 from ._lib._at import at
@@ -48,6 +49,7 @@
     "one_hot",
     "pad",
     "partition",
+    "quantile",
     "setdiff1d",
     "sinc",
 ]

src/array_api_extra/_delegation.py

@@ -4,7 +4,8 @@
 from types import ModuleType
 from typing import Literal
 
-from ._lib import _funcs
+from ._lib import _funcs, _quantile
+from ._lib._backends import NUMPY_VERSION
 from ._lib._utils._compat import (
     array_namespace,
     is_cupy_namespace,
@@ -768,7 +769,7 @@ def argpartition(
         Axis along which to partition. The default is ``-1`` (the last axis).
         If ``None``, the flattened array is used.
     xp : array_namespace, optional
-        The standard-compatible namespace for `x`. Default: infer.
+        The standard-compatible namespace for `a`. Default: infer.
 
     Returns
     -------
@@ -895,3 +896,273 @@ def isin(
         return xp.isin(a, b, assume_unique=assume_unique, invert=invert)
 
     return _funcs.isin(a, b, assume_unique=assume_unique, invert=invert, xp=xp)
+
+
+def quantile(
+    a: Array,
+    q: float | Array,
+    /,
+    axis: int | None = None,
+    method: str = "linear",
+    keepdims: bool = False,
+    nan_policy: str = "propagate",
+    *,
+    weights: Array | None = None,
+    xp: ModuleType | None = None,
+) -> Array:
+    """
+    Compute the q-th quantile of the data along the specified axis.
+
+    Parameters
+    ----------
+    a : array_like of real numbers
+        Input array or object that can be converted to an array.
+    q : array_like of float
+        Probability or sequence of probabilities of the quantiles to compute.
+        Values must be between 0 and 1 inclusive.
+    axis : {int, tuple of int, None}, optional
+        Axis or axes along which the quantiles are computed. The default is
+        to compute the quantile(s) along a flattened version of the array.
+    method : str, optional
+        This parameter specifies the method to use for estimating the
+        quantile. There are many different methods.
+        The recommended options, numbered as they appear in [1]_, are:
+
+        1. 'inverted_cdf'
+        2. 'averaged_inverted_cdf'
+        3. 'closest_observation'
+        4. 'interpolated_inverted_cdf'
+        5. 'hazen'
+        6. 'weibull'
+        7. 'linear'  (default)
+        8. 'median_unbiased'
+        9. 'normal_unbiased'
+
+        The first three methods are discontinuous.
+        Only 'linear', 'inverted_cdf' and 'averaged_inverted_cdf' are implemented.
+
+    keepdims : bool, optional
+        If this is set to True, the axes which are reduced are left in
+        the result as dimensions with size one. With this option, the
+        result will broadcast correctly against the original array `a`.
+
+    nan_policy : str, optional
+        'propagate' (default) or 'omit'.
+        'omit' is support only when `weights` are provided.
+
+    weights : array_like, optional
+        An array of weights associated with the values in `a`. Each value in
+        `a` contributes to the quantile according to its associated weight.
+        The weights array can either be 1-D (in which case its length must be
+        the size of `a` along the given axis) or of the same shape as `a`.
+        If `weights=None`, then all data in `a` are assumed to have a
+        weight equal to one.
+        Only `method="inverted_cdf"` or `method="averaged_inverted_cdf"`
+        support weights. See the notes for more details.
+
+    xp : array_namespace, optional
+        The standard-compatible namespace for `a` and `q`. Default: infer.
+
+    Returns
+    -------
+    scalar or ndarray
+        If `q` is a single probability and `axis=None`, then the result
+        is a scalar. If multiple probability levels are given, first axis
+        of the result corresponds to the quantiles. The other axes are
+        the axes that remain after the reduction of `a`. If the input
+        contains integers or floats smaller than ``float64``, the output
+        data-type is ``float64``. Otherwise, the output data-type is the
+        same as that of the input. If `out` is specified, that array is
+        returned instead.
+
+    Notes
+    -----
+    Given a sample `a` from an underlying distribution, `quantile` provides a
+    nonparametric estimate of the inverse cumulative distribution function.
+
+    By default, this is done by interpolating between adjacent elements in
+    ``y``, a sorted copy of `a`::
+
+        (1-g)*y[j] + g*y[j+1]
+
+    where the index ``j`` and coefficient ``g`` are the integral and
+    fractional components of ``q * (n-1)``, and ``n`` is the number of
+    elements in the sample.
+
+    This is a special case of Equation 1 of H&F [1]_. More generally,
+
+    - ``j = (q*n + m - 1) // 1``, and
+    - ``g = (q*n + m - 1) % 1``,
+
+    where ``m`` may be defined according to several different conventions.
+    The preferred convention may be selected using the ``method`` parameter:
+
+    =============================== =============== ===============
+    ``method``                      number in H&F   ``m``
+    =============================== =============== ===============
+    ``interpolated_inverted_cdf``   4               ``0``
+    ``hazen``                       5               ``1/2``
+    ``weibull``                     6               ``q``
+    ``linear`` (default)            7               ``1 - q``
+    ``median_unbiased``             8               ``q/3 + 1/3``
+    ``normal_unbiased``             9               ``q/4 + 3/8``
+    =============================== =============== ===============
+
+    Note that indices ``j`` and ``j + 1`` are clipped to the range ``0`` to
+    ``n - 1`` when the results of the formula would be outside the allowed
+    range of non-negative indices. The ``- 1`` in the formulas for ``j`` and
+    ``g`` accounts for Python's 0-based indexing.
+
+    The table above includes only the estimators from H&F that are continuous
+    functions of probability `q` (estimators 4-9). NumPy also provides the
+    three discontinuous estimators from H&F (estimators 1-3), where ``j`` is
+    defined as above, ``m`` is defined as follows, and ``g`` is a function
+    of the real-valued ``index = q*n + m - 1`` and ``j``.
+
+    1. ``inverted_cdf``: ``m = 0`` and ``g = int(index - j > 0)``
+    2. ``averaged_inverted_cdf``: ``m = 0`` and
+       ``g = (1 + int(index - j > 0)) / 2``
+    3. ``closest_observation``: ``m = -1/2`` and
+       ``g = 1 - int((index == j) & (j%2 == 1))``
+
+    **Weighted quantiles:**
+    More formally, the quantile at probability level :math:`q` of a cumulative
+    distribution function :math:`F(y)=P(Y \\leq y)` with probability measure
+    :math:`P` is defined as any number :math:`x` that fulfills the
+    *coverage conditions*
+
+    .. math:: P(Y < x) \\leq q \\quad\\text{and}\\quad P(Y \\leq x) \\geq q
+
+    with random variable :math:`Y\\sim P`.
+    Sample quantiles, the result of `quantile`, provide nonparametric
+    estimation of the underlying population counterparts, represented by the
+    unknown :math:`F`, given a data vector `a` of length ``n``.
+
+    Some of the estimators above arise when one considers :math:`F` as the
+    empirical distribution function of the data, i.e.
+    :math:`F(y) = \\frac{1}{n} \\sum_i 1_{a_i \\leq y}`.
+    Then, different methods correspond to different choices of :math:`x` that
+    fulfill the above coverage conditions. Methods that follow this approach
+    are ``inverted_cdf`` and ``averaged_inverted_cdf``.
+
+    For weighted quantiles, the coverage conditions still hold. The
+    empirical cumulative distribution is simply replaced by its weighted
+    version, i.e.
+    :math:`P(Y \\leq t) = \\frac{1}{\\sum_i w_i} \\sum_i w_i 1_{x_i \\leq t}`.
+
+    References
+    ----------
+    .. [1] R. J. Hyndman and Y. Fan,
+       "Sample quantiles in statistical packages,"
+       The American Statistician, 50(4), pp. 361-365, 1996
+    """
+    if xp is None:
+        xp = array_namespace(a)
+    if is_pydata_sparse_namespace(xp):
+        msg = "Sparse backend not supported"
+        raise ValueError(msg)
+
+    methods = {"linear", "inverted_cdf", "averaged_inverted_cdf"}
+    if method not in methods:
+        msg = f"`method` must be one of {methods}"
+        raise ValueError(msg)
+    nan_policies = {"propagate", "omit"}
+    if nan_policy not in nan_policies:
+        msg = f"`nan_policy` must be one of {nan_policies}"
+        raise ValueError(msg)
+
+    a = xp.asarray(a)
+    if not xp.isdtype(a.dtype, ("integral", "real floating")):
+        msg = "`a` must have real dtype."
+        raise ValueError(msg)
+    if not xp.isdtype(xp.asarray(q).dtype, "real floating"):
+        msg = "`q` must have real floating dtype."
+        raise ValueError(msg)
+    weights = None if weights is None else xp.asarray(weights)
+
+    ndim = a.ndim
+    if ndim < 1:
+        msg = "`a` must be at least 1-dimensional."
+        raise TypeError(msg)
+    if axis is not None and ((axis >= ndim) or (axis < -ndim)):
+        msg = "`axis` is not compatible with the dimension of `a`."
+        raise ValueError(msg)
+    if weights is None:
+        if nan_policy != "propagate":
+            msg = "When `weights` aren't provided, `nan_policy` must be 'propagate'."
+            raise ValueError(msg)
+    else:
+        if method not in {"inverted_cdf", "averaged_inverted_cdf"}:
+            msg = f"`method` '{method}' not supported with weights."
+            raise ValueError(msg)
+        if not xp.isdtype(weights.dtype, ("integral", "real floating")):
+            msg = "`weights` must have real dtype."
+            raise ValueError(msg)
+        if ndim > 2:
+            msg = "When weights are provided, dimension of `a` must be 1 or 2."
+            raise ValueError(msg)
+        if a.shape != weights.shape:
+            if axis is None:
+                msg = "Axis must be specified when shapes of `a` and ̀ weights` differ."
+                raise TypeError(msg)
+            if weights.shape != eager_shape(a, axis):
+                msg = (
+                    "Shape of weights must be consistent with shape"
+                    " of a along specified axis."
+                )
+                raise ValueError(msg)
+        if axis is None and ndim == 2:
+            msg = "Axis must be specified when `a` and ̀ weights` are 2d."
+            raise ValueError(msg)
+
+    # Align result dtype with what numpy does:
+    dtype = xp.result_type(
+        xp.float64 if xp.isdtype(a.dtype, "integral") else a,
+        xp.asarray(q),
+        xp.float64,  # at least float64
+    )
+    device = get_device(a)
+    a = xp.asarray(a, dtype=dtype, device=device)
+    q_arr = xp.asarray(q, dtype=dtype, device=device)
+    # TODO: cast weights here? Assert weights are on the same device as `a`?
+
+    if xp.any((q_arr > 1) | (q_arr < 0) | xp.isnan(q_arr)):
+        msg = "`q` values must be in the range [0, 1]"
+        raise ValueError(msg)
+
+    # Delegate when possible.
+    # Note: No delegation for dask: I couldn't make it work.
+    basic_case = method == "linear" and weights is None
+
+    np_2 = NUMPY_VERSION >= (2, 0)
+    np_handles_weights = np_2 and nan_policy == "propagate" and method == "inverted_cdf"
+    if weights is None:
+        if is_numpy_namespace(xp) and (basic_case or np_2):
+            quantile = xp.quantile if nan_policy == "propagate" else xp.nanquantile
+            return quantile(a, q_arr, axis=axis, method=method, keepdims=keepdims)
+    elif is_numpy_namespace(xp) and np_handles_weights:
+        # TODO: call nanquantile for nan_policy == "omit" once
+        # https://github.com/numpy/numpy/issues/29709 is fixed
+        return xp.quantile(
+            a, q_arr, axis=axis, method=method, keepdims=keepdims, weights=weights
+        )
+
+    jax_or_cupy = is_jax_namespace(xp) or is_cupy_namespace(xp)
+    if jax_or_cupy and basic_case and nan_policy == "propagate":
+        return xp.quantile(a, q_arr, axis=axis, method=method, keepdims=keepdims)
+    if is_torch_namespace(xp) and basic_case:
+        quantile = xp.quantile if nan_policy == "propagate" else xp.nanquantile
+        return quantile(a, q_arr, dim=axis, interpolation=method, keepdim=keepdims)
+
+    # Otherwise call our implementation (will sort data)
+    return _quantile.quantile(
+        # XXX: I'm not sure we want to support dask, it seems uterly slow...
+        a,
+        q_arr,
+        axis=axis,
+        method=method,
+        keepdims=keepdims,
+        nan_policy=nan_policy,
+        weights=weights,
+        xp=xp,
+    )