def _cofactor_solve(a, b):
"""Equivalent to det(a)*solve(a, b) for nonsingular mat.
Intermediate function used for jvp and vjp of det.
This function borrows heavily from jax.numpy.linalg.solve and
jax.numpy.linalg.slogdet to compute the gradient of the determinant
in a way that is well defined even for low rank matrices.
This function handles two different cases:
* rank(a) == n or n-1
* rank(a) < n-1
For rank n-1 matrices, the gradient of the determinant is a rank 1 matrix.
Rather than computing det(a)*solve(a, b), which would return NaN, we work
directly with the LU decomposition. If a = p @ l @ u, then
det(a)*solve(a, b) =
prod(diag(u)) * u^-1 @ l^-1 @ p^-1 b =
prod(diag(u)) * triangular_solve(u, solve(p @ l, b))
If a is rank n-1, then the lower right corner of u will be zero and the
triangular_solve will fail.
Let x = solve(p @ l, b) and y = det(a)*solve(a, b).
Then y_{n}
x_{n} / u_{nn} * prod_{i=1...n}(u_{ii}) =
x_{n} * prod_{i=1...n-1}(u_{ii})
So by replacing the lower-right corner of u with prod_{i=1...n-1}(u_{ii})^-1
we can avoid the triangular_solve failing.
To correctly compute the rest of y_{i} for i != n, we simply multiply
x_{i} by det(a) for all i != n, which will be zero if rank(a) = n-1.
For the second case, a check is done on the matrix to see if `solve`
returns NaN or Inf, and gives a matrix of zeros as a result, as the
gradient of the determinant of a matrix with rank less than n-1 is 0.
This will still return the correct value for rank n-1 matrices, as the check
is applied *after* the lower right corner of u has been updated.
Args:
a: A square matrix or batch of matrices, possibly singular.
b: A matrix, or batch of matrices of the same dimension as a.
Returns:
det(a) and cofactor(a)^T*b, aka adjugate(a)*b
"""
a = _promote_arg_dtypes(jnp.asarray(a))
b = _promote_arg_dtypes(jnp.asarray(b))
a_shape = jnp.shape(a)
b_shape = jnp.shape(b)
a_ndims = len(a_shape)
if not (a_ndims >= 2 and a_shape[-1] == a_shape[-2]
and b_shape[-2:] == a_shape[-2:]):
msg = ("The arguments to _cofactor_solve must have shapes "
"a=[..., m, m] and b=[..., m, m]; got a={} and b={}")
raise ValueError(msg.format(a_shape, b_shape))
if a_shape[-1] == 1:
return a[..., 0, 0], b
# lu contains u in the upper triangular matrix and l in the strict lower
# triangular matrix.
# The diagonal of l is set to ones without loss of generality.
lu, pivots, permutation = lax_linalg.lu(a)
dtype = lax.dtype(a)
batch_dims = lax.broadcast_shapes(lu.shape[:-2], b.shape[:-2])
x = jnp.broadcast_to(b, batch_dims + b.shape[-2:])
lu = jnp.broadcast_to(lu, batch_dims + lu.shape[-2:])
# Compute (partial) determinant, ignoring last diagonal of LU
diag = jnp.diagonal(lu, axis1=-2, axis2=-1)
parity = jnp.count_nonzero(pivots != jnp.arange(a_shape[-1]), axis=-1)
sign = jnp.asarray(-2 * (parity % 2) + 1, dtype=dtype)
# partial_det[:, -1] contains the full determinant and
# partial_det[:, -2] contains det(u) / u_{nn}.
partial_det = jnp.cumprod(diag, axis=-1) * sign[..., None]
lu = lu.at[..., -1, -1].set(1.0 / partial_det[..., -2])
permutation = jnp.broadcast_to(permutation, batch_dims + (a_shape[-1], ))
iotas = jnp.ix_(*(lax.iota(jnp.int32, b) for b in batch_dims + (1, )))
# filter out any matrices that are not full rank
d = jnp.ones(x.shape[:-1], x.dtype)
d = lax_linalg.triangular_solve(lu, d, left_side=True, lower=False)
d = jnp.any(jnp.logical_or(jnp.isnan(d), jnp.isinf(d)), axis=-1)
d = jnp.tile(d[..., None, None], d.ndim * (1, ) + x.shape[-2:])
x = jnp.where(d, jnp.zeros_like(x), x) # first filter
x = x[iotas[:-1] + (permutation, slice(None))]
x = lax_linalg.triangular_solve(lu,
x,
left_side=True,
lower=True,
unit_diagonal=True)
x = jnp.concatenate(
(x[..., :-1, :] * partial_det[..., -1, None, None], x[..., -1:, :]),
axis=-2)
x = lax_linalg.triangular_solve(lu, x, left_side=True, lower=False)
x = jnp.where(d, jnp.zeros_like(x), x) # second filter
return partial_det[..., -1], x
def norm(x,
ord=None,
axis: Union[None, Tuple[int, ...], int] = None,
keepdims=False):
x = _promote_arg_dtypes(jnp.asarray(x))
x_shape = jnp.shape(x)
ndim = len(x_shape)
if axis is None:
# NumPy has an undocumented behavior that admits arbitrary rank inputs if
# `ord` is None: https://github.com/numpy/numpy/issues/14215
if ord is None:
return jnp.sqrt(
jnp.sum(jnp.real(x * jnp.conj(x)), keepdims=keepdims))
axis = tuple(range(ndim))
elif isinstance(axis, tuple):
axis = tuple(canonicalize_axis(x, ndim) for x in axis)
else:
axis = (canonicalize_axis(axis, ndim), )
num_axes = len(axis)
if num_axes == 1:
if ord is None or ord == 2:
return jnp.sqrt(
jnp.sum(jnp.real(x * jnp.conj(x)),
axis=axis,
keepdims=keepdims))
elif ord == jnp.inf:
return jnp.amax(jnp.abs(x), axis=axis, keepdims=keepdims)
elif ord == -jnp.inf:
return jnp.amin(jnp.abs(x), axis=axis, keepdims=keepdims)
elif ord == 0:
return jnp.sum(x != 0,
dtype=jnp.finfo(lax.dtype(x)).dtype,
axis=axis,
keepdims=keepdims)
elif ord == 1:
# Numpy has a special case for ord == 1 as an optimization. We don't
# really need the optimization (XLA could do it for us), but the Numpy
# code has slightly different type promotion semantics, so we need a
# special case too.
return jnp.sum(jnp.abs(x), axis=axis, keepdims=keepdims)
else:
abs_x = jnp.abs(x)
ord = lax._const(abs_x, ord)
out = jnp.sum(abs_x**ord, axis=axis, keepdims=keepdims)
return jnp.power(out, 1. / ord)
elif num_axes == 2:
row_axis, col_axis = cast(Tuple[int, ...], axis)
if ord is None or ord in ('f', 'fro'):
return jnp.sqrt(
jnp.sum(jnp.real(x * jnp.conj(x)),
axis=axis,
keepdims=keepdims))
elif ord == 1:
if not keepdims and col_axis > row_axis:
col_axis -= 1
return jnp.amax(jnp.sum(jnp.abs(x),
axis=row_axis,
keepdims=keepdims),
axis=col_axis,
keepdims=keepdims)
elif ord == -1:
if not keepdims and col_axis > row_axis:
col_axis -= 1
return jnp.amin(jnp.sum(jnp.abs(x),
axis=row_axis,
keepdims=keepdims),
axis=col_axis,
keepdims=keepdims)
elif ord == jnp.inf:
if not keepdims and row_axis > col_axis:
row_axis -= 1
return jnp.amax(jnp.sum(jnp.abs(x),
axis=col_axis,
keepdims=keepdims),
axis=row_axis,
keepdims=keepdims)
elif ord == -jnp.inf:
if not keepdims and row_axis > col_axis:
row_axis -= 1
return jnp.amin(jnp.sum(jnp.abs(x),
axis=col_axis,
keepdims=keepdims),
axis=row_axis,
keepdims=keepdims)
elif ord in ('nuc', 2, -2):
x = jnp.moveaxis(x, axis, (-2, -1))
if ord == 2:
reducer = jnp.amax
elif ord == -2:
reducer = jnp.amin
else:
reducer = jnp.sum
y = reducer(svd(x, compute_uv=False), axis=-1)
if keepdims:
result_shape = list(x_shape)
result_shape[axis[0]] = 1
result_shape[axis[1]] = 1
y = jnp.reshape(y, result_shape)
return y
else:
raise ValueError("Invalid order '{}' for matrix norm.".format(ord))
else:
raise ValueError(
"Invalid axis values ({}) for jnp.linalg.norm.".format(axis))
def _ndtri(p):
"""Implements ndtri core logic."""
# Constants used in piece-wise rational approximations. Taken from the cephes
# library:
# https://root.cern.ch/doc/v608/SpecFuncCephesInv_8cxx_source.html
p0 = list(
reversed([
-5.99633501014107895267E1, 9.80010754185999661536E1,
-5.66762857469070293439E1, 1.39312609387279679503E1,
-1.23916583867381258016E0
]))
q0 = list(
reversed([
1.0, 1.95448858338141759834E0, 4.67627912898881538453E0,
8.63602421390890590575E1, -2.25462687854119370527E2,
2.00260212380060660359E2, -8.20372256168333339912E1,
1.59056225126211695515E1, -1.18331621121330003142E0
]))
p1 = list(
reversed([
4.05544892305962419923E0, 3.15251094599893866154E1,
5.71628192246421288162E1, 4.40805073893200834700E1,
1.46849561928858024014E1, 2.18663306850790267539E0,
-1.40256079171354495875E-1, -3.50424626827848203418E-2,
-8.57456785154685413611E-4
]))
q1 = list(
reversed([
1.0, 1.57799883256466749731E1, 4.53907635128879210584E1,
4.13172038254672030440E1, 1.50425385692907503408E1,
2.50464946208309415979E0, -1.42182922854787788574E-1,
-3.80806407691578277194E-2, -9.33259480895457427372E-4
]))
p2 = list(
reversed([
3.23774891776946035970E0, 6.91522889068984211695E0,
3.93881025292474443415E0, 1.33303460815807542389E0,
2.01485389549179081538E-1, 1.23716634817820021358E-2,
3.01581553508235416007E-4, 2.65806974686737550832E-6,
6.23974539184983293730E-9
]))
q2 = list(
reversed([
1.0, 6.02427039364742014255E0, 3.67983563856160859403E0,
1.37702099489081330271E0, 2.16236993594496635890E-1,
1.34204006088543189037E-2, 3.28014464682127739104E-4,
2.89247864745380683936E-6, 6.79019408009981274425E-9
]))
dtype = lax.dtype(p).type
shape = jnp.shape(p)
def _create_polynomial(var, coeffs):
"""Compute n_th order polynomial via Horner's method."""
coeffs = np.array(coeffs, dtype)
if not coeffs.size:
return jnp.zeros_like(var)
return coeffs[0] + _create_polynomial(var, coeffs[1:]) * var
maybe_complement_p = jnp.where(p > dtype(-np.expm1(-2.)), dtype(1.) - p, p)
# Write in an arbitrary value in place of 0 for p since 0 will cause NaNs
# later on. The result from the computation when p == 0 is not used so any
# number that doesn't result in NaNs is fine.
sanitized_mcp = jnp.where(maybe_complement_p <= dtype(0.),
jnp.full(shape, dtype(0.5)), maybe_complement_p)
# Compute x for p > exp(-2): x/sqrt(2pi) = w + w**3 P0(w**2)/Q0(w**2).
w = sanitized_mcp - dtype(0.5)
ww = lax.square(w)
x_for_big_p = w + w * ww * (_create_polynomial(ww, p0) /
_create_polynomial(ww, q0))
x_for_big_p *= -dtype(np.sqrt(2. * np.pi))
# Compute x for p <= exp(-2): x = z - log(z)/z - (1/z) P(1/z) / Q(1/z),
# where z = sqrt(-2. * log(p)), and P/Q are chosen between two different
# arrays based on whether p < exp(-32).
z = lax.sqrt(dtype(-2.) * lax.log(sanitized_mcp))
first_term = z - lax.log(z) / z
second_term_small_p = (_create_polynomial(dtype(1.) / z, p2) /
_create_polynomial(dtype(1.) / z, q2) / z)
second_term_otherwise = (_create_polynomial(dtype(1.) / z, p1) /
_create_polynomial(dtype(1.) / z, q1) / z)
x_for_small_p = first_term - second_term_small_p
x_otherwise = first_term - second_term_otherwise
x = jnp.where(sanitized_mcp > dtype(np.exp(-2.)), x_for_big_p,
jnp.where(z >= dtype(8.0), x_for_small_p, x_otherwise))
x = jnp.where(p > dtype(1. - np.exp(-2.)), x, -x)
infinity = jnp.full(shape, dtype(np.inf))
x_nan_replaced = jnp.where(p <= dtype(0.0), -infinity,
jnp.where(p >= dtype(1.0), infinity, x))
return x_nan_replaced
def log_ndtr(x, series_order=3):
r"""Log Normal distribution function.
For details of the Normal distribution function see `ndtr`.
This function calculates :math:`\log(\mathrm{ndtr}(x))` by either calling
:math:`\log(\mathrm{ndtr}(x))` or using an asymptotic series. Specifically:
- For `x > upper_segment`, use the approximation `-ndtr(-x)` based on
:math:`\log(1-x) \approx -x, x \ll 1`.
- For `lower_segment < x <= upper_segment`, use the existing `ndtr` technique
and take a log.
- For `x <= lower_segment`, we use the series approximation of `erf` to compute
the log CDF directly.
The `lower_segment` is set based on the precision of the input:
.. math::
\begin{align}
\mathit{lower\_segment} =&
\ \begin{cases}
-20 & x.\mathrm{dtype}=\mathit{float64} \\
-10 & x.\mathrm{dtype}=\mathit{float32} \\
\end{cases} \\
\mathit{upper\_segment} =&
\ \begin{cases}
8& x.\mathrm{dtype}=\mathit{float64} \\
5& x.\mathrm{dtype}=\mathit{float32} \\
\end{cases}
\end{align}
When `x < lower_segment`, the `ndtr` asymptotic series approximation is:
.. math::
\begin{align}
\mathrm{ndtr}(x) =&\ \mathit{scale} * (1 + \mathit{sum}) + R_N \\
\mathit{scale} =&\ \frac{e^{-0.5 x^2}}{-x \sqrt{2 \pi}} \\
\mathit{sum} =&\ \sum_{n=1}^N {-1}^n (2n-1)!! / (x^2)^n \\
R_N =&\ O(e^{-0.5 x^2} (2N+1)!! / |x|^{2N+3})
\end{align}
where :math:`(2n-1)!! = (2n-1) (2n-3) (2n-5) ... (3) (1)` is a
`double-factorial
<https://en.wikipedia.org/wiki/Double_factorial>`_ operator.
Args:
x: an array of type `float32`, `float64`.
series_order: Positive Python integer. Maximum depth to
evaluate the asymptotic expansion. This is the `N` above.
Returns:
an array with `dtype=x.dtype`.
Raises:
TypeError: if `x.dtype` is not handled.
TypeError: if `series_order` is a not Python `integer.`
ValueError: if `series_order` is not in `[0, 30]`.
"""
if not isinstance(series_order, int):
raise TypeError("series_order must be a Python integer.")
if series_order < 0:
raise ValueError("series_order must be non-negative.")
if series_order > 30:
raise ValueError("series_order must be <= 30.")
x = jnp.asarray(x)
dtype = lax.dtype(x)
if dtype == jnp.float64:
lower_segment = _LOGNDTR_FLOAT64_LOWER
upper_segment = _LOGNDTR_FLOAT64_UPPER
elif dtype == jnp.float32:
lower_segment = _LOGNDTR_FLOAT32_LOWER
upper_segment = _LOGNDTR_FLOAT32_UPPER
else:
raise TypeError("x.dtype={} is not supported.".format(np.dtype(dtype)))
# The basic idea here was ported from:
# https://root.cern.ch/doc/v608/SpecFuncCephesInv_8cxx_source.html
# We copy the main idea, with a few changes
# * For x >> 1, and X ~ Normal(0, 1),
# Log[P[X < x]] = Log[1 - P[X < -x]] approx -P[X < -x],
# which extends the range of validity of this function.
# * We use one fixed series_order for all of 'x', rather than adaptive.
# * Our docstring properly reflects that this is an asymptotic series, not a
# Taylor series. We also provided a correct bound on the remainder.
# * We need to use the max/min in the _log_ndtr_lower arg to avoid nan when
# x=0. This happens even though the branch is unchosen because when x=0
# the gradient of a select involves the calculation 1*dy+0*(-inf)=nan
# regardless of whether dy is finite. Note that the minimum is a NOP if
# the branch is chosen.
return jnp.where(
lax.gt(x, upper_segment),
-_ndtr(-x), # log(1-x) ~= -x, x << 1
jnp.where(lax.gt(x, lower_segment),
lax.log(_ndtr(lax.max(x, lower_segment))),
_log_ndtr_lower(lax.min(x, lower_segment), series_order)))
def polygamma(n, x):
assert jnp.issubdtype(lax.dtype(n), jnp.integer)
n, x = _promote_args_inexact("polygamma", n, x)
shape = lax.broadcast_shapes(n.shape, x.shape)
return _polygamma(jnp.broadcast_to(n, shape), jnp.broadcast_to(x, shape))
def _polygamma(n, x):
dtype = lax.dtype(n).type
n_plus = n + dtype(1)
sign = dtype(1) - (n_plus % dtype(2)) * dtype(2)
return jnp.where(n == 0, digamma(x),
sign * jnp.exp(gammaln(n_plus)) * zeta(n_plus, x))
def __init__(self, fn=None, mask=True):
if lax.dtype(mask) != 'bool':
raise ValueError("`mask` should be a bool array.")
self.mask = mask
super().__init__(fn)
def get_dtype(x):
return canonicalize_dtype(lax.dtype(x))
def _rotate_left(x, d):
if lax.dtype(d) != dtype:
d = lax.convert_element_type(d, dtype)
if lax.dtype(x) != dtype:
x = lax.convert_element_type(x, dtype)
return lax.shift_left(x, d) | lax.shift_right_logical(x, nbits - d)
def triangular_solve(a, b, left_side=False, lower=False, transpose_a=False,
conjugate_a=False, unit_diagonal=False):
conjugate_a = conjugate_a and np.issubdtype(lax.dtype(a), np.complexfloating)
return triangular_solve_p.bind(
a, b, left_side=left_side, lower=lower, transpose_a=transpose_a,
conjugate_a=conjugate_a, unit_diagonal=unit_diagonal)
def gen_normalized_legendre(l_max, x):
r"""Computes the normalized associated Legendre functions (ALFs).
The ALFs of the first kind are used in spherical harmonics. The spherical
harmonic of degree `l` and order `m` can be written as
`Y_l^m(θ, φ) = N_l^m * P_l^m(cos(θ)) * exp(i m φ)`, where `N_l^m` is the
normalization factor and θ and φ are the colatitude and longitude,
repectively. `N_l^m` is chosen in the way that the spherical harmonics form
a set of orthonormal basis function of L^2(S^2). For the computational
efficiency of spherical harmonics transform, the normalization factor is
embedded into the computation of the ALFs. In addition, normalizing `P_l^m`
avoids overflow/underflow and achieves better numerical stability. Three
recurrence relations are used in the computation. Note that the factor of
\sqrt(1 / (4 𝛑)) is used in the formulation.
Args:
l_max: The maximum degree of the associated Legendre function. Both the
degrees and orders are `[0, 1, 2, ..., l_max]`.
x: A vector of type `float32`, `float64` containing the sampled points in
spherical coordinates, at which the ALFs are computed; `x` is essentially
`cos(θ)`.
Returns:
The 3D array of shape `(l_max + 1, l_max + 1, len(x))` containing the
normalized values of the ALFs at `x`.
"""
dtype = lax.dtype(x)
if dtype not in (jnp.float32, jnp.float64):
raise TypeError(
'x.dtype={} is not supported, see docstring for supported types.'.
format(dtype))
if x.ndim != 1:
raise ValueError('x must be a 1D array.')
p = np.zeros((l_max + 1, l_max + 1, x.shape[0]))
# The initial value p(0,0).
initial_value = 0.5 / np.sqrt(math.pi)
p[0, 0] = initial_value
# Compute the diagonal entries p(l,l) with recurrence.
y = np.sqrt(1.0 - x * x)
for l in range(1, l_max + 1):
a = -1.0 * np.sqrt(1.0 + 0.5 / l)
p[l, l] = a * y * p[l - 1, l - 1]
# Compute the off-diagonal entries with recurrence.
for l in range(l_max):
b = np.sqrt(2.0 * l + 3.0)
p[l + 1, l] = b * x * p[l, l]
# Compute the remaining entries with recurrence.
for m in range(l_max + 1):
for l in range(m + 2, l_max + 1):
c0 = l * l
c1 = m * m
c2 = 2.0 * l
c3 = (l - 1.0) * (l - 1.0)
d0 = np.sqrt((4.0 * c0 - 1.0) / (c0 - c1))
d1 = np.sqrt(((c2 + 1.0) * (c3 - c1)) / ((c2 - 3.0) * (c0 - c1)))
p[l, m] = d0 * x * p[l - 1, m] - d1 * p[l - 2, m]
return jnp.asarray(p)
def get_dtypes(*args):
return [canonicalize_dtype(lax.dtype(arg)) for arg in args]
def inv(a):
if jnp.ndim(a) < 2 or a.shape[-1] != a.shape[-2]:
raise ValueError(
f"Argument to inv must have shape [..., n, n], got {a.shape}.")
return solve(
a, lax.broadcast(jnp.eye(a.shape[-1], dtype=lax.dtype(a)), a.shape[:-2]))
