bvn, MVNXPB, TruncatedMultivariateNormal, and UnifiedSkewNormal (pyto…

…rch#1394)

Summary:
Pull Request resolved: pytorch#1394

Introduces `utils/probability` submodule with the following offers:

- `bvn`: Methods for computing bivariate normal probabilities and moments.
- `MVNXPB`: Approximate solver for Multivariate Normal CDF.
- `LinearEllipticalSliceSampler`: Class for sampling trMVN random variables.
- `TruncatedMultivariateNormal`: Truncated multivariate normal Distribution class
- `UnifiedSkewNormal`: Unified skew normal Distribution class

Reviewed By: Balandat

Differential Revision: D39326106

fbshipit-source-id: e605ff7e3b484f128fa0a86cb6088a85a96e5969

botorch/utils/constants.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from __future__ import annotations
+
+from functools import lru_cache
+from numbers import Number
+from typing import Iterator, Optional, Tuple, Union
+
+import torch
+from torch import Tensor
+
+
+@lru_cache(maxsize=None)
+def get_constants(
+    values: Union[Number, Iterator[Number]],
+    device: Optional[torch.device] = None,
+    dtype: Optional[torch.dtype] = None,
+) -> Union[Tensor, Tuple[Tensor, ...]]:
+    r"""Returns scalar-valued Tensors containing each of the given constants.
+    Used to expedite tensor operations involving scalar arithmetic. Note that
+    the returned Tensors should not be modified in-place."""
+    if isinstance(values, Number):
+        return torch.full((), values, dtype=dtype, device=device)
+
+    return tuple(torch.full((), val, dtype=dtype, device=device) for val in values)
+
+
+def get_constants_like(
+    values: Union[Number, Iterator[Number]],
+    ref: Tensor,
+) -> Union[Tensor, Iterator[Tensor]]:
+    return get_constants(values, device=ref.device, dtype=ref.dtype)

botorch/utils/probability/__init__.py

@@ -0,0 +1,25 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+from botorch.utils.probability.bvn import bvn, bvnmom
+from botorch.utils.probability.lin_ess import LinearEllipticalSliceSampler
+from botorch.utils.probability.mvnxpb import MVNXPB
+from botorch.utils.probability.truncated_multivariate_normal import (
+    TruncatedMultivariateNormal,
+)
+from botorch.utils.probability.unified_skew_normal import UnifiedSkewNormal
+from botorch.utils.probability.utils import ndtr
+
+
+__all__ = [
+    "bvn",
+    "bvnmom",
+    "LinearEllipticalSliceSampler",
+    "MVNXPB",
+    "ndtr",
+    "TruncatedMultivariateNormal",
+    "UnifiedSkewNormal",
+]

botorch/utils/probability/bvn.py

@@ -0,0 +1,300 @@
+#!/usr/bin/env python3
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+r"""
+Methods for computing bivariate normal probabilities and statistics.
+
+.. [Drezner1990computation]
+    Z. Drezner and G. O. Wesolowsky. On the computation of the bivariate normal
+    integral. Journal of Statistical Computation and Simulation, 1990.
+
+.. [Genz2004bvnt]
+    A. Genz. Numerical computation of rectangular bivariate and trivariate normal and
+    t probabilities. Statistics and Computing, 2004.
+
+.. [Rosenbaum1961moments]
+    S. Rosenbaum. Moments of a Truncated Bivariate Normal Distribution. Journal of the
+    Royal Statistical Society (Series B), 1961.
+
+.. [Muthen1990moments]
+    B. Muthen. Moments of the censored and truncated bivariate normal distribution.
+    British Journal of Mathematical and Statistical Psychology, 1990.
+"""
+
+from __future__ import annotations
+
+from math import pi as _pi
+from typing import Optional, Tuple
+
+import torch
+from botorch.exceptions import UnsupportedError
+from botorch.utils.probability.utils import (
+    case_dispatcher,
+    get_constants_like,
+    leggauss,
+    ndtr as Phi,
+    phi,
+    STANDARDIZED_RANGE,
+)
+from botorch.utils.safe_math import (
+    div as safe_div,
+    exp as safe_exp,
+    mul as safe_mul,
+    sub as safe_sub,
+)
+from torch import Tensor
+
+# Some useful constants
+_inf = float("inf")
+_2pi = 2 * _pi
+_sqrt_2pi = _2pi**0.5
+_inv_2pi = 1 / _2pi
+
+
+def bvn(r: Tensor, xl: Tensor, yl: Tensor, xu: Tensor, yu: Tensor) -> Tensor:
+    r"""A function for computing bivariate normal probabilities.
+
+    Calculates `P(xl < x < xu, yl < y < yu)` where `x` and `y` are bivariate normal with
+    unit variance and correlation coefficient `r`. See Section 2.4 of [Genz2004bvnt]_.
+
+    This method uses a sign flip trick to improving numerical performance. Many of
+    `bvnu`s internal branches rely on evaluations `Phi(-bound)`. For `a < b < 0`,
+    the term `Phi(-a) - Phi(-b)` goes to zero faster than `Phi(b) - Phi(a)` because
+    `finfo(dtype).epsneg` is typically much larger than `finfo(dtype).tiny`. In these
+    cases, flipping the sign can prevent situations where `bvnu(...) - bvnu(...)` would
+    otherwise be zero due to round-off error.
+
+    Args:
+        r: Tensor of correlation coefficients.
+        xl: Tensor of lower bounds for `x`, same shape as `r`.
+        yl: Tensor of lower bounds for `y`, same shape as `r`.
+        xu: Tensor of upper bounds for `x`, same shape as `r`.
+        yu: Tensor of upper bounds for `y`, same shape as `r`.
+
+    Returns:
+        Tensor of probabilities `P(xl < x < xu, yl < y < yu)`.
+
+    """
+    if not (r.shape == xl.shape == xu.shape == yl.shape == yu.shape):
+        raise UnsupportedError("Arguments to `bvn` must have the same shape.")
+
+    # Sign flip trick
+    _0, _1, _2 = get_constants_like(values=(0, 1, 2), ref=r)
+    flip_x = xl.abs() > xu  # is xl more negative than xu is positive?
+    flip_y = yl.abs() > yu
+    flip = (flip_x & (~flip_y | yu.isinf())) | (flip_y & (~flip_x | xu.isinf()))
+    if flip.any():  # symmetric calls to `bvnu` below makes swapping bounds unnecessary
+        sign = _1 - _2 * flip.to(dtype=r.dtype)
+        xl = sign * xl  # becomes `-xu` if flipped
+        xu = sign * xu  # becomes `-xl`
+        yl = sign * yl  # becomes `-yu`
+        yu = sign * yu  # becomes `-yl`
+
+    p = bvnu(r, xl, yl) - bvnu(r, xu, yl) - bvnu(r, xl, yu) + bvnu(r, xu, yu)
+    return p.clip(_0, _1)
+
+
+def bvnu(r: Tensor, h: Tensor, k: Tensor) -> Tensor:
+    r"""Solves for `P(x > h, y > k)` where `x` and `y` are standard bivariate normal
+    random variables with correlation coefficient `r`. In [Genz2004bvnt]_, this is (1)
+    ```
+    L(h, k, r) = P(x < -h, y < -k)
+               = 1/(a 2\pi) \int_{h}^{\infty} \int_{k}^{\infty} f(x, y, r) dy dx,
+    ```
+    where `f(x, y, r) = e^{-1/(2a^2) (x^2 - 2rxy + y^2)}` and `a = (1 - r^2)^{1/2}`.
+
+    [Genz2004bvnt]_ report the following integation scheme incurs a maximum of 5e-16
+    error when run in double precision. For strongly correlated variables |r| >= 0.925,
+    use a 20-point quadrature rule on a 5th order Taylor expansion. Elsewhere,
+    numerically integrate in polar coordinates using no more than 20 quadrature points.
+
+    Args:
+        r: Tensor of correlation coefficients.
+        h: Tensor of negative upper bounds for `x`, same shape as `r`.
+        k: Tensor of negative upper bounds for `y`, same shape as `r`.
+
+    Returns:
+        A tensor of probabilities `P(x > h, y > k)`.
+    """
+    if not (r.shape == h.shape == k.shape):
+        raise UnsupportedError("Arguments to `bvnu` must have the same shape.")
+    _0, _1, lower, upper = get_constants_like((0, 1) + STANDARDIZED_RANGE, r)
+    x_free = h < lower
+    y_free = k < lower
+    return case_dispatcher(
+        out=torch.empty_like(r),
+        cases=(  # Special cases admitting closed-form solutions
+            (lambda: (h > upper) | (k > upper), lambda mask: _0),
+            (lambda: x_free & y_free, lambda mask: _1),
+            (lambda: x_free, lambda mask: Phi(-k[mask])),
+            (lambda: y_free, lambda mask: Phi(-h[mask])),
+            (lambda: r == _0, lambda mask: Phi(-h[mask]) * Phi(-k[mask])),
+            (  # For |r| >= 0.925, use a Taylor approximation
+                lambda: r.abs() >= get_constants_like(0.925, r),
+                lambda m: _bvnu_taylor(r[m], h[m], k[m]),
+            ),
+        ),  # For |r| < 0.925, integrate in polar coordinates.
+        default=lambda mask: _bvnu_polar(r[mask], h[mask], k[mask]),
+    )
+
+
+def _bvnu_polar(
+    r: Tensor, h: Tensor, k: Tensor, num_points: Optional[int] = None
+) -> Tensor:
+    r"""Solves for `P(x > h, y > k)` by integrating in polar coordinates as
+    ```
+        L(h, k, r) = \Phi(-h)\Phi(-k) + 1/(2\pi) \int_{0}^{sin^{-1}(r)} f(t) dt
+        f(t) = e^{-0.5 cos(t)^{-2} (h^2 + k^2 - 2hk sin(t))}
+    ```
+    For details, see Section 2.2 of [Genz2004bvnt]_.
+    """
+    if num_points is None:
+        mar = r.abs().max()
+        num_points = 6 if mar < 0.3 else 12 if mar < 0.75 else 20
+
+    _0, _1, _i2, _i2pi = get_constants_like(values=(0, 1, 0.5, _inv_2pi), ref=r)
+
+    x, w = leggauss(num_points, dtype=r.dtype, device=r.device)
+    x = x + _1
+    asin_r = _i2 * torch.asin(r)
+    sin_asrx = (asin_r.unsqueeze(-1) * x).sin()
+
+    _h = h.unsqueeze(-1)
+    _k = k.unsqueeze(-1)
+    vals = safe_exp(
+        safe_sub(safe_mul(sin_asrx, _h * _k), _i2 * (_h.square() + _k.square()))
+        / (_1 - sin_asrx.square())
+    )
+    probs = Phi(-h) * Phi(-k) + _i2pi * asin_r * (vals @ w)
+    return probs.clip(min=_0, max=_1)  # necessary due to "safe" handling of inf
+
+
+def _bvnu_taylor(r: Tensor, h: Tensor, k: Tensor, num_points: int = 20) -> Tensor:
+    r"""Solves for `P(x > h, y > k)` via Taylor expansion.
+
+    Following [Drezner1990computation], the standard BVN problem may be rewritten as
+    ```
+        L(h, k, r) = L(h, k, s) - s/(2\pi) \int_{0}^{a} f(x) dx
+        f(x) = (1 - x^2){-1/2} e^{-0.5 ((h - sk)/ x)^2} e^{-shk/(1 + (1 - x^2)^{1/2})},
+    ```
+    where `s = sign(r)` and `a = sqrt(1 - r^{2})`. The term `L(h, k, s)` is analytic.
+    The second integral is approximated via Taylor expansion. See Sections 2.3 and
+    2.4 of [Genz2004bvnt]_.
+    """
+    _0, _1, _ni2, _i2pi, _sq2pi = get_constants_like(
+        values=(0, 1, -0.5, _inv_2pi, _sqrt_2pi), ref=r
+    )
+
+    x, w = leggauss(num_points, dtype=r.dtype, device=r.device)
+    x = x + _1
+
+    s = get_constants_like(2, r) * (r > _0).to(r) - _1  # sign of `r` where sign(0) := 1
+    sk = s * k
+    skh = sk * h
+    comp_r2 = _1 - r.square()
+
+    a = comp_r2.clip(min=0).sqrt()
+    b = safe_sub(h, sk)
+    b2 = b.square()
+    c = get_constants_like(1 / 8, r) * (get_constants_like(4, r) - skh)
+    d = get_constants_like(1 / 80, r) * (get_constants_like(12, r) - skh)
+
+    # ---- Solve for `L(h, k, s)`
+    int_from_0_to_s = case_dispatcher(
+        out=torch.empty_like(r),
+        cases=[(lambda: r > _0, lambda mask: Phi(-torch.maximum(h[mask], k[mask])))],
+        default=lambda mask: (Phi(sk[mask]) - Phi(h[mask])).clip(min=_0),
+    )
+
+    # ---- Solve for `s/(2\pi) \int_{0}^{a} f(x) dx`
+    # Analytic part
+    _a0 = _ni2 * (safe_div(b2, comp_r2) + skh)
+    _a1 = c * get_constants_like(1 / 3, r) * (_1 - d * b2)
+    _a2 = _1 - b2 * _a1
+    abs_b = b.abs()
+    analytic_part = torch.subtract(  # analytic part of solution
+        a * (_a2 + comp_r2 * _a1 + c * d * comp_r2.square()) * safe_exp(_a0),
+        _sq2pi * Phi(safe_div(-abs_b, a)) * abs_b * _a2 * safe_exp(_ni2 * skh),
+    )
+
+    # Quadrature part
+    _b2 = b2.unsqueeze(-1)
+    _skh = skh.unsqueeze(-1)
+    _q0 = get_constants_like(0.25, r) * comp_r2.unsqueeze(-1) * x.square()
+    _q1 = (_1 - _q0).sqrt()
+    _q2 = _ni2 * (_b2 / _q0 + _skh)
+
+    _b2 = b2.unsqueeze(-1)
+    _c = c.unsqueeze(-1)
+    _d = d.unsqueeze(-1)
+    vals = (_ni2 * (_b2 / _q0 + _skh)).exp() * torch.subtract(
+        _1 + _c * _q0 * (_1 + get_constants_like(5, r) * _d * _q0),
+        safe_exp(_ni2 * _q0 / (_1 + _q1).square() * _skh) / _q1,
+    )
+    mask = _q2 > get_constants_like(-100, r)
+    if not mask.all():
+        vals[~mask] = _0
+    quadrature_part = _ni2 * a * (vals @ w)
+
+    # Return `P(x > h, y > k)`
+    int_from_0_to_a = _i2pi * s * (analytic_part + quadrature_part)
+    return (int_from_0_to_s - int_from_0_to_a).clip(min=_0, max=_1)
+
+
+def bvnmom(
+    r: Tensor,
+    xl: Tensor,
+    yl: Tensor,
+    xu: Tensor,
+    yu: Tensor,
+    p: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor]:
+    r"""Computes the expected values of truncated, bivariate normal random variables.
+
+    Let `x` and `y` be a pair of standard bivariate normal random variables having
+    correlation `r`. This function computes `E([x,y] | [xl,yl] < [x,y] < [xu,yu])`.
+
+    Following [Muthen1990moments]_ equations (4) and (5), we have
+    ```
+    E(x | [xl, yl] < [x, y] < [xu, yu])
+        = Z^{-1} \phi(xl) P(yl < y < yu | x=xl) - \phi(xu) P(yl < y < yu | x=xu)
+    ```
+    where `Z = P([xl, yl] < [x, y] < [xu, yu])` and `\phi` is the standard normal PDF.
+
+    Args:
+        r: Tensor of correlation coefficients.
+        xl: Tensor of lower bounds for `x`, same shape as `r`.
+        xu: Tensor of upper bounds for `x`, same shape as `r`.
+        yl: Tensor of lower bounds for `y`, same shape as `r`.
+        yu: Tensor of upper bounds for `y`, same shape as `r`.
+        p: Tensor of probabilities `P(xl < x < xu, yl < y < yu)`, same shape as `r`.
+
+    Returns:
+        `E(x | [xl, yl] < [x, y] < [xu, yu])` and `E(y | [xl, yl] < [x, y] < [xu, yu])`.
+    """
+    if not (r.shape == xl.shape == xu.shape == yl.shape == yu.shape):
+        raise UnsupportedError("Arguments to `bvn` must have the same shape.")
+
+    if p is None:
+        p = bvn(r=r, xl=xl, xu=xu, yl=yl, yu=yu)
+
+    corr = r[..., None, None]
+    istd = (1 - corr.square()).rsqrt()
+    lower = torch.stack([xl, yl], -1)
+    upper = torch.stack([xu, yu], -1)
+    bounds = torch.stack([lower, upper], -1)
+    deltas = safe_mul(corr, bounds)
+
+    # Compute densities and conditional probabilities
+    density_at_bounds = phi(bounds)
+    prob_given_bounds = Phi(
+        safe_mul(istd, safe_sub(upper.flip(-1).unsqueeze(-1), deltas))
+    ) - Phi(safe_mul(istd, safe_sub(lower.flip(-1).unsqueeze(-1), deltas)))
+
+    # Evaluate Muthen's formula
+    p_diffs = -(density_at_bounds * prob_given_bounds).diff().squeeze(-1)
+    moments = (1 / p).unsqueeze(-1) * (p_diffs + r.unsqueeze(-1) * p_diffs.flip(-1))
+    return moments.unbind(-1)