427 lines
15 KiB
Python
427 lines
15 KiB
Python
"""
|
|
Created on Fri Apr 2 09:06:05 2021
|
|
|
|
@author: matth
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
import math
|
|
import numpy as np
|
|
from scipy import special
|
|
from ._axis_nan_policy import _axis_nan_policy_factory, _broadcast_arrays
|
|
from scipy._lib._array_api import array_namespace
|
|
|
|
__all__ = ['entropy', 'differential_entropy']
|
|
|
|
|
|
@_axis_nan_policy_factory(
|
|
lambda x: x,
|
|
n_samples=lambda kwgs: (
|
|
2 if ("qk" in kwgs and kwgs["qk"] is not None)
|
|
else 1
|
|
),
|
|
n_outputs=1, result_to_tuple=lambda x: (x,), paired=True,
|
|
too_small=-1 # entropy doesn't have too small inputs
|
|
)
|
|
def entropy(pk: np.typing.ArrayLike,
|
|
qk: np.typing.ArrayLike | None = None,
|
|
base: float | None = None,
|
|
axis: int = 0
|
|
) -> np.number | np.ndarray:
|
|
"""
|
|
Calculate the Shannon entropy/relative entropy of given distribution(s).
|
|
|
|
If only probabilities `pk` are given, the Shannon entropy is calculated as
|
|
``H = -sum(pk * log(pk))``.
|
|
|
|
If `qk` is not None, then compute the relative entropy
|
|
``D = sum(pk * log(pk / qk))``. This quantity is also known
|
|
as the Kullback-Leibler divergence.
|
|
|
|
This routine will normalize `pk` and `qk` if they don't sum to 1.
|
|
|
|
Parameters
|
|
----------
|
|
pk : array_like
|
|
Defines the (discrete) distribution. Along each axis-slice of ``pk``,
|
|
element ``i`` is the (possibly unnormalized) probability of event
|
|
``i``.
|
|
qk : array_like, optional
|
|
Sequence against which the relative entropy is computed. Should be in
|
|
the same format as `pk`.
|
|
base : float, optional
|
|
The logarithmic base to use, defaults to ``e`` (natural logarithm).
|
|
axis : int, optional
|
|
The axis along which the entropy is calculated. Default is 0.
|
|
|
|
Returns
|
|
-------
|
|
S : {float, array_like}
|
|
The calculated entropy.
|
|
|
|
Notes
|
|
-----
|
|
Informally, the Shannon entropy quantifies the expected uncertainty
|
|
inherent in the possible outcomes of a discrete random variable.
|
|
For example,
|
|
if messages consisting of sequences of symbols from a set are to be
|
|
encoded and transmitted over a noiseless channel, then the Shannon entropy
|
|
``H(pk)`` gives a tight lower bound for the average number of units of
|
|
information needed per symbol if the symbols occur with frequencies
|
|
governed by the discrete distribution `pk` [1]_. The choice of base
|
|
determines the choice of units; e.g., ``e`` for nats, ``2`` for bits, etc.
|
|
|
|
The relative entropy, ``D(pk|qk)``, quantifies the increase in the average
|
|
number of units of information needed per symbol if the encoding is
|
|
optimized for the probability distribution `qk` instead of the true
|
|
distribution `pk`. Informally, the relative entropy quantifies the expected
|
|
excess in surprise experienced if one believes the true distribution is
|
|
`qk` when it is actually `pk`.
|
|
|
|
A related quantity, the cross entropy ``CE(pk, qk)``, satisfies the
|
|
equation ``CE(pk, qk) = H(pk) + D(pk|qk)`` and can also be calculated with
|
|
the formula ``CE = -sum(pk * log(qk))``. It gives the average
|
|
number of units of information needed per symbol if an encoding is
|
|
optimized for the probability distribution `qk` when the true distribution
|
|
is `pk`. It is not computed directly by `entropy`, but it can be computed
|
|
using two calls to the function (see Examples).
|
|
|
|
See [2]_ for more information.
|
|
|
|
References
|
|
----------
|
|
.. [1] Shannon, C.E. (1948), A Mathematical Theory of Communication.
|
|
Bell System Technical Journal, 27: 379-423.
|
|
https://doi.org/10.1002/j.1538-7305.1948.tb01338.x
|
|
.. [2] Thomas M. Cover and Joy A. Thomas. 2006. Elements of Information
|
|
Theory (Wiley Series in Telecommunications and Signal Processing).
|
|
Wiley-Interscience, USA.
|
|
|
|
|
|
Examples
|
|
--------
|
|
The outcome of a fair coin is the most uncertain:
|
|
|
|
>>> import numpy as np
|
|
>>> from scipy.stats import entropy
|
|
>>> base = 2 # work in units of bits
|
|
>>> pk = np.array([1/2, 1/2]) # fair coin
|
|
>>> H = entropy(pk, base=base)
|
|
>>> H
|
|
1.0
|
|
>>> H == -np.sum(pk * np.log(pk)) / np.log(base)
|
|
True
|
|
|
|
The outcome of a biased coin is less uncertain:
|
|
|
|
>>> qk = np.array([9/10, 1/10]) # biased coin
|
|
>>> entropy(qk, base=base)
|
|
0.46899559358928117
|
|
|
|
The relative entropy between the fair coin and biased coin is calculated
|
|
as:
|
|
|
|
>>> D = entropy(pk, qk, base=base)
|
|
>>> D
|
|
0.7369655941662062
|
|
>>> D == np.sum(pk * np.log(pk/qk)) / np.log(base)
|
|
True
|
|
|
|
The cross entropy can be calculated as the sum of the entropy and
|
|
relative entropy`:
|
|
|
|
>>> CE = entropy(pk, base=base) + entropy(pk, qk, base=base)
|
|
>>> CE
|
|
1.736965594166206
|
|
>>> CE == -np.sum(pk * np.log(qk)) / np.log(base)
|
|
True
|
|
|
|
"""
|
|
if base is not None and base <= 0:
|
|
raise ValueError("`base` must be a positive number or `None`.")
|
|
|
|
xp = array_namespace(pk) if qk is None else array_namespace(pk, qk)
|
|
|
|
pk = xp.asarray(pk)
|
|
with np.errstate(invalid='ignore'):
|
|
pk = 1.0*pk / xp.sum(pk, axis=axis, keepdims=True) # type: ignore[operator]
|
|
if qk is None:
|
|
vec = special.entr(pk)
|
|
else:
|
|
qk = xp.asarray(qk)
|
|
pk, qk = _broadcast_arrays((pk, qk), axis=None, xp=xp) # don't ignore any axes
|
|
sum_kwargs = dict(axis=axis, keepdims=True)
|
|
qk = 1.0*qk / xp.sum(qk, **sum_kwargs) # type: ignore[operator, call-overload]
|
|
vec = special.rel_entr(pk, qk)
|
|
S = xp.sum(vec, axis=axis)
|
|
if base is not None:
|
|
S /= math.log(base)
|
|
return S
|
|
|
|
|
|
def _differential_entropy_is_too_small(samples, kwargs, axis=-1):
|
|
values = samples[0]
|
|
n = values.shape[axis]
|
|
window_length = kwargs.get("window_length",
|
|
math.floor(math.sqrt(n) + 0.5))
|
|
if not 2 <= 2 * window_length < n:
|
|
return True
|
|
return False
|
|
|
|
|
|
@_axis_nan_policy_factory(
|
|
lambda x: x, n_outputs=1, result_to_tuple=lambda x: (x,),
|
|
too_small=_differential_entropy_is_too_small
|
|
)
|
|
def differential_entropy(
|
|
values: np.typing.ArrayLike,
|
|
*,
|
|
window_length: int | None = None,
|
|
base: float | None = None,
|
|
axis: int = 0,
|
|
method: str = "auto",
|
|
) -> np.number | np.ndarray:
|
|
r"""Given a sample of a distribution, estimate the differential entropy.
|
|
|
|
Several estimation methods are available using the `method` parameter. By
|
|
default, a method is selected based the size of the sample.
|
|
|
|
Parameters
|
|
----------
|
|
values : sequence
|
|
Sample from a continuous distribution.
|
|
window_length : int, optional
|
|
Window length for computing Vasicek estimate. Must be an integer
|
|
between 1 and half of the sample size. If ``None`` (the default), it
|
|
uses the heuristic value
|
|
|
|
.. math::
|
|
\left \lfloor \sqrt{n} + 0.5 \right \rfloor
|
|
|
|
where :math:`n` is the sample size. This heuristic was originally
|
|
proposed in [2]_ and has become common in the literature.
|
|
base : float, optional
|
|
The logarithmic base to use, defaults to ``e`` (natural logarithm).
|
|
axis : int, optional
|
|
The axis along which the differential entropy is calculated.
|
|
Default is 0.
|
|
method : {'vasicek', 'van es', 'ebrahimi', 'correa', 'auto'}, optional
|
|
The method used to estimate the differential entropy from the sample.
|
|
Default is ``'auto'``. See Notes for more information.
|
|
|
|
Returns
|
|
-------
|
|
entropy : float
|
|
The calculated differential entropy.
|
|
|
|
Notes
|
|
-----
|
|
This function will converge to the true differential entropy in the limit
|
|
|
|
.. math::
|
|
n \to \infty, \quad m \to \infty, \quad \frac{m}{n} \to 0
|
|
|
|
The optimal choice of ``window_length`` for a given sample size depends on
|
|
the (unknown) distribution. Typically, the smoother the density of the
|
|
distribution, the larger the optimal value of ``window_length`` [1]_.
|
|
|
|
The following options are available for the `method` parameter.
|
|
|
|
* ``'vasicek'`` uses the estimator presented in [1]_. This is
|
|
one of the first and most influential estimators of differential entropy.
|
|
* ``'van es'`` uses the bias-corrected estimator presented in [3]_, which
|
|
is not only consistent but, under some conditions, asymptotically normal.
|
|
* ``'ebrahimi'`` uses an estimator presented in [4]_, which was shown
|
|
in simulation to have smaller bias and mean squared error than
|
|
the Vasicek estimator.
|
|
* ``'correa'`` uses the estimator presented in [5]_ based on local linear
|
|
regression. In a simulation study, it had consistently smaller mean
|
|
square error than the Vasiceck estimator, but it is more expensive to
|
|
compute.
|
|
* ``'auto'`` selects the method automatically (default). Currently,
|
|
this selects ``'van es'`` for very small samples (<10), ``'ebrahimi'``
|
|
for moderate sample sizes (11-1000), and ``'vasicek'`` for larger
|
|
samples, but this behavior is subject to change in future versions.
|
|
|
|
All estimators are implemented as described in [6]_.
|
|
|
|
References
|
|
----------
|
|
.. [1] Vasicek, O. (1976). A test for normality based on sample entropy.
|
|
Journal of the Royal Statistical Society:
|
|
Series B (Methodological), 38(1), 54-59.
|
|
.. [2] Crzcgorzewski, P., & Wirczorkowski, R. (1999). Entropy-based
|
|
goodness-of-fit test for exponentiality. Communications in
|
|
Statistics-Theory and Methods, 28(5), 1183-1202.
|
|
.. [3] Van Es, B. (1992). Estimating functionals related to a density by a
|
|
class of statistics based on spacings. Scandinavian Journal of
|
|
Statistics, 61-72.
|
|
.. [4] Ebrahimi, N., Pflughoeft, K., & Soofi, E. S. (1994). Two measures
|
|
of sample entropy. Statistics & Probability Letters, 20(3), 225-234.
|
|
.. [5] Correa, J. C. (1995). A new estimator of entropy. Communications
|
|
in Statistics-Theory and Methods, 24(10), 2439-2449.
|
|
.. [6] Noughabi, H. A. (2015). Entropy Estimation Using Numerical Methods.
|
|
Annals of Data Science, 2(2), 231-241.
|
|
https://link.springer.com/article/10.1007/s40745-015-0045-9
|
|
|
|
Examples
|
|
--------
|
|
>>> import numpy as np
|
|
>>> from scipy.stats import differential_entropy, norm
|
|
|
|
Entropy of a standard normal distribution:
|
|
|
|
>>> rng = np.random.default_rng()
|
|
>>> values = rng.standard_normal(100)
|
|
>>> differential_entropy(values)
|
|
1.3407817436640392
|
|
|
|
Compare with the true entropy:
|
|
|
|
>>> float(norm.entropy())
|
|
1.4189385332046727
|
|
|
|
For several sample sizes between 5 and 1000, compare the accuracy of
|
|
the ``'vasicek'``, ``'van es'``, and ``'ebrahimi'`` methods. Specifically,
|
|
compare the root mean squared error (over 1000 trials) between the estimate
|
|
and the true differential entropy of the distribution.
|
|
|
|
>>> from scipy import stats
|
|
>>> import matplotlib.pyplot as plt
|
|
>>>
|
|
>>>
|
|
>>> def rmse(res, expected):
|
|
... '''Root mean squared error'''
|
|
... return np.sqrt(np.mean((res - expected)**2))
|
|
>>>
|
|
>>>
|
|
>>> a, b = np.log10(5), np.log10(1000)
|
|
>>> ns = np.round(np.logspace(a, b, 10)).astype(int)
|
|
>>> reps = 1000 # number of repetitions for each sample size
|
|
>>> expected = stats.expon.entropy()
|
|
>>>
|
|
>>> method_errors = {'vasicek': [], 'van es': [], 'ebrahimi': []}
|
|
>>> for method in method_errors:
|
|
... for n in ns:
|
|
... rvs = stats.expon.rvs(size=(reps, n), random_state=rng)
|
|
... res = stats.differential_entropy(rvs, method=method, axis=-1)
|
|
... error = rmse(res, expected)
|
|
... method_errors[method].append(error)
|
|
>>>
|
|
>>> for method, errors in method_errors.items():
|
|
... plt.loglog(ns, errors, label=method)
|
|
>>>
|
|
>>> plt.legend()
|
|
>>> plt.xlabel('sample size')
|
|
>>> plt.ylabel('RMSE (1000 trials)')
|
|
>>> plt.title('Entropy Estimator Error (Exponential Distribution)')
|
|
|
|
"""
|
|
values = np.asarray(values)
|
|
values = np.moveaxis(values, axis, -1)
|
|
n = values.shape[-1] # number of observations
|
|
|
|
if window_length is None:
|
|
window_length = math.floor(math.sqrt(n) + 0.5)
|
|
|
|
if not 2 <= 2 * window_length < n:
|
|
raise ValueError(
|
|
f"Window length ({window_length}) must be positive and less "
|
|
f"than half the sample size ({n}).",
|
|
)
|
|
|
|
if base is not None and base <= 0:
|
|
raise ValueError("`base` must be a positive number or `None`.")
|
|
|
|
sorted_data = np.sort(values, axis=-1)
|
|
|
|
methods = {"vasicek": _vasicek_entropy,
|
|
"van es": _van_es_entropy,
|
|
"correa": _correa_entropy,
|
|
"ebrahimi": _ebrahimi_entropy,
|
|
"auto": _vasicek_entropy}
|
|
method = method.lower()
|
|
if method not in methods:
|
|
message = f"`method` must be one of {set(methods)}"
|
|
raise ValueError(message)
|
|
|
|
if method == "auto":
|
|
if n <= 10:
|
|
method = 'van es'
|
|
elif n <= 1000:
|
|
method = 'ebrahimi'
|
|
else:
|
|
method = 'vasicek'
|
|
|
|
res = methods[method](sorted_data, window_length)
|
|
|
|
if base is not None:
|
|
res /= np.log(base)
|
|
|
|
return res
|
|
|
|
|
|
def _pad_along_last_axis(X, m):
|
|
"""Pad the data for computing the rolling window difference."""
|
|
# scales a bit better than method in _vasicek_like_entropy
|
|
shape = np.array(X.shape)
|
|
shape[-1] = m
|
|
Xl = np.broadcast_to(X[..., [0]], shape) # [0] vs 0 to maintain shape
|
|
Xr = np.broadcast_to(X[..., [-1]], shape)
|
|
return np.concatenate((Xl, X, Xr), axis=-1)
|
|
|
|
|
|
def _vasicek_entropy(X, m):
|
|
"""Compute the Vasicek estimator as described in [6] Eq. 1.3."""
|
|
n = X.shape[-1]
|
|
X = _pad_along_last_axis(X, m)
|
|
differences = X[..., 2 * m:] - X[..., : -2 * m:]
|
|
logs = np.log(n/(2*m) * differences)
|
|
return np.mean(logs, axis=-1)
|
|
|
|
|
|
def _van_es_entropy(X, m):
|
|
"""Compute the van Es estimator as described in [6]."""
|
|
# No equation number, but referred to as HVE_mn.
|
|
# Typo: there should be a log within the summation.
|
|
n = X.shape[-1]
|
|
difference = X[..., m:] - X[..., :-m]
|
|
term1 = 1/(n-m) * np.sum(np.log((n+1)/m * difference), axis=-1)
|
|
k = np.arange(m, n+1)
|
|
return term1 + np.sum(1/k) + np.log(m) - np.log(n+1)
|
|
|
|
|
|
def _ebrahimi_entropy(X, m):
|
|
"""Compute the Ebrahimi estimator as described in [6]."""
|
|
# No equation number, but referred to as HE_mn
|
|
n = X.shape[-1]
|
|
X = _pad_along_last_axis(X, m)
|
|
|
|
differences = X[..., 2 * m:] - X[..., : -2 * m:]
|
|
|
|
i = np.arange(1, n+1).astype(float)
|
|
ci = np.ones_like(i)*2
|
|
ci[i <= m] = 1 + (i[i <= m] - 1)/m
|
|
ci[i >= n - m + 1] = 1 + (n - i[i >= n-m+1])/m
|
|
|
|
logs = np.log(n * differences / (ci * m))
|
|
return np.mean(logs, axis=-1)
|
|
|
|
|
|
def _correa_entropy(X, m):
|
|
"""Compute the Correa estimator as described in [6]."""
|
|
# No equation number, but referred to as HC_mn
|
|
n = X.shape[-1]
|
|
X = _pad_along_last_axis(X, m)
|
|
|
|
i = np.arange(1, n+1)
|
|
dj = np.arange(-m, m+1)[:, None]
|
|
j = i + dj
|
|
j0 = j + m - 1 # 0-indexed version of j
|
|
|
|
Xibar = np.mean(X[..., j0], axis=-2, keepdims=True)
|
|
difference = X[..., j0] - Xibar
|
|
num = np.sum(difference*dj, axis=-2) # dj is d-i
|
|
den = n*np.sum(difference**2, axis=-2)
|
|
return -np.mean(np.log(num/den), axis=-1)
|