Traktor/myenv/Lib/site-packages/scipy/stats/_ksstats.py

601 lines
20 KiB
Python
Raw Normal View History

2024-05-26 05:12:46 +02:00
# Compute the two-sided one-sample Kolmogorov-Smirnov Prob(Dn <= d) where:
# D_n = sup_x{|F_n(x) - F(x)|},
# F_n(x) is the empirical CDF for a sample of size n {x_i: i=1,...,n},
# F(x) is the CDF of a probability distribution.
#
# Exact methods:
# Prob(D_n >= d) can be computed via a matrix algorithm of Durbin[1]
# or a recursion algorithm due to Pomeranz[2].
# Marsaglia, Tsang & Wang[3] gave a computation-efficient way to perform
# the Durbin algorithm.
# D_n >= d <==> D_n+ >= d or D_n- >= d (the one-sided K-S statistics), hence
# Prob(D_n >= d) = 2*Prob(D_n+ >= d) - Prob(D_n+ >= d and D_n- >= d).
# For d > 0.5, the latter intersection probability is 0.
#
# Approximate methods:
# For d close to 0.5, ignoring that intersection term may still give a
# reasonable approximation.
# Li-Chien[4] and Korolyuk[5] gave an asymptotic formula extending
# Kolmogorov's initial asymptotic, suitable for large d. (See
# scipy.special.kolmogorov for that asymptotic)
# Pelz-Good[6] used the functional equation for Jacobi theta functions to
# transform the Li-Chien/Korolyuk formula produce a computational formula
# suitable for small d.
#
# Simard and L'Ecuyer[7] provided an algorithm to decide when to use each of
# the above approaches and it is that which is used here.
#
# Other approaches:
# Carvalho[8] optimizes Durbin's matrix algorithm for large values of d.
# Moscovich and Nadler[9] use FFTs to compute the convolutions.
# References:
# [1] Durbin J (1968).
# "The Probability that the Sample Distribution Function Lies Between Two
# Parallel Straight Lines."
# Annals of Mathematical Statistics, 39, 398-411.
# [2] Pomeranz J (1974).
# "Exact Cumulative Distribution of the Kolmogorov-Smirnov Statistic for
# Small Samples (Algorithm 487)."
# Communications of the ACM, 17(12), 703-704.
# [3] Marsaglia G, Tsang WW, Wang J (2003).
# "Evaluating Kolmogorov's Distribution."
# Journal of Statistical Software, 8(18), 1-4.
# [4] LI-CHIEN, C. (1956).
# "On the exact distribution of the statistics of A. N. Kolmogorov and
# their asymptotic expansion."
# Acta Matematica Sinica, 6, 55-81.
# [5] KOROLYUK, V. S. (1960).
# "Asymptotic analysis of the distribution of the maximum deviation in
# the Bernoulli scheme."
# Theor. Probability Appl., 4, 339-366.
# [6] Pelz W, Good IJ (1976).
# "Approximating the Lower Tail-areas of the Kolmogorov-Smirnov One-sample
# Statistic."
# Journal of the Royal Statistical Society, Series B, 38(2), 152-156.
# [7] Simard, R., L'Ecuyer, P. (2011)
# "Computing the Two-Sided Kolmogorov-Smirnov Distribution",
# Journal of Statistical Software, Vol 39, 11, 1-18.
# [8] Carvalho, Luis (2015)
# "An Improved Evaluation of Kolmogorov's Distribution"
# Journal of Statistical Software, Code Snippets; Vol 65(3), 1-8.
# [9] Amit Moscovich, Boaz Nadler (2017)
# "Fast calculation of boundary crossing probabilities for Poisson
# processes",
# Statistics & Probability Letters, Vol 123, 177-182.
import numpy as np
import scipy.special
import scipy.special._ufuncs as scu
from scipy._lib._finite_differences import _derivative
_E128 = 128
_EP128 = np.ldexp(np.longdouble(1), _E128)
_EM128 = np.ldexp(np.longdouble(1), -_E128)
_SQRT2PI = np.sqrt(2 * np.pi)
_LOG_2PI = np.log(2 * np.pi)
_MIN_LOG = -708
_SQRT3 = np.sqrt(3)
_PI_SQUARED = np.pi ** 2
_PI_FOUR = np.pi ** 4
_PI_SIX = np.pi ** 6
# [Lifted from _loggamma.pxd.] If B_m are the Bernoulli numbers,
# then Stirling coeffs are B_{2j}/(2j)/(2j-1) for j=8,...1.
_STIRLING_COEFFS = [-2.955065359477124183e-2, 6.4102564102564102564e-3,
-1.9175269175269175269e-3, 8.4175084175084175084e-4,
-5.952380952380952381e-4, 7.9365079365079365079e-4,
-2.7777777777777777778e-3, 8.3333333333333333333e-2]
def _log_nfactorial_div_n_pow_n(n):
# Computes n! / n**n
# = (n-1)! / n**(n-1)
# Uses Stirling's approximation, but removes n*log(n) up-front to
# avoid subtractive cancellation.
# = log(n)/2 - n + log(sqrt(2pi)) + sum B_{2j}/(2j)/(2j-1)/n**(2j-1)
rn = 1.0/n
return np.log(n)/2 - n + _LOG_2PI/2 + rn * np.polyval(_STIRLING_COEFFS, rn/n)
def _clip_prob(p):
"""clips a probability to range 0<=p<=1."""
return np.clip(p, 0.0, 1.0)
def _select_and_clip_prob(cdfprob, sfprob, cdf=True):
"""Selects either the CDF or SF, and then clips to range 0<=p<=1."""
p = np.where(cdf, cdfprob, sfprob)
return _clip_prob(p)
def _kolmogn_DMTW(n, d, cdf=True):
r"""Computes the Kolmogorov CDF: Pr(D_n <= d) using the MTW approach to
the Durbin matrix algorithm.
Durbin (1968); Marsaglia, Tsang, Wang (2003). [1], [3].
"""
# Write d = (k-h)/n, where k is positive integer and 0 <= h < 1
# Generate initial matrix H of size m*m where m=(2k-1)
# Compute k-th row of (n!/n^n) * H^n, scaling intermediate results.
# Requires memory O(m^2) and computation O(m^2 log(n)).
# Most suitable for small m.
if d >= 1.0:
return _select_and_clip_prob(1.0, 0.0, cdf)
nd = n * d
if nd <= 0.5:
return _select_and_clip_prob(0.0, 1.0, cdf)
k = int(np.ceil(nd))
h = k - nd
m = 2 * k - 1
H = np.zeros([m, m])
# Initialize: v is first column (and last row) of H
# v[j] = (1-h^(j+1)/(j+1)! (except for v[-1])
# w[j] = 1/(j)!
# q = k-th row of H (actually i!/n^i*H^i)
intm = np.arange(1, m + 1)
v = 1.0 - h ** intm
w = np.empty(m)
fac = 1.0
for j in intm:
w[j - 1] = fac
fac /= j # This might underflow. Isn't a problem.
v[j - 1] *= fac
tt = max(2 * h - 1.0, 0)**m - 2*h**m
v[-1] = (1.0 + tt) * fac
for i in range(1, m):
H[i - 1:, i] = w[:m - i + 1]
H[:, 0] = v
H[-1, :] = np.flip(v, axis=0)
Hpwr = np.eye(np.shape(H)[0]) # Holds intermediate powers of H
nn = n
expnt = 0 # Scaling of Hpwr
Hexpnt = 0 # Scaling of H
while nn > 0:
if nn % 2:
Hpwr = np.matmul(Hpwr, H)
expnt += Hexpnt
H = np.matmul(H, H)
Hexpnt *= 2
# Scale as needed.
if np.abs(H[k - 1, k - 1]) > _EP128:
H /= _EP128
Hexpnt += _E128
nn = nn // 2
p = Hpwr[k - 1, k - 1]
# Multiply by n!/n^n
for i in range(1, n + 1):
p = i * p / n
if np.abs(p) < _EM128:
p *= _EP128
expnt -= _E128
# unscale
if expnt != 0:
p = np.ldexp(p, expnt)
return _select_and_clip_prob(p, 1.0-p, cdf)
def _pomeranz_compute_j1j2(i, n, ll, ceilf, roundf):
"""Compute the endpoints of the interval for row i."""
if i == 0:
j1, j2 = -ll - ceilf - 1, ll + ceilf - 1
else:
# i + 1 = 2*ip1div2 + ip1mod2
ip1div2, ip1mod2 = divmod(i + 1, 2)
if ip1mod2 == 0: # i is odd
if ip1div2 == n + 1:
j1, j2 = n - ll - ceilf - 1, n + ll + ceilf - 1
else:
j1, j2 = ip1div2 - 1 - ll - roundf - 1, ip1div2 + ll - 1 + ceilf - 1
else:
j1, j2 = ip1div2 - 1 - ll - 1, ip1div2 + ll + roundf - 1
return max(j1 + 2, 0), min(j2, n)
def _kolmogn_Pomeranz(n, x, cdf=True):
r"""Computes Pr(D_n <= d) using the Pomeranz recursion algorithm.
Pomeranz (1974) [2]
"""
# V is n*(2n+2) matrix.
# Each row is convolution of the previous row and probabilities from a
# Poisson distribution.
# Desired CDF probability is n! V[n-1, 2n+1] (final entry in final row).
# Only two rows are needed at any given stage:
# - Call them V0 and V1.
# - Swap each iteration
# Only a few (contiguous) entries in each row can be non-zero.
# - Keep track of start and end (j1 and j2 below)
# - V0s and V1s track the start in the two rows
# Scale intermediate results as needed.
# Only a few different Poisson distributions can occur
t = n * x
ll = int(np.floor(t))
f = 1.0 * (t - ll) # fractional part of t
g = min(f, 1.0 - f)
ceilf = (1 if f > 0 else 0)
roundf = (1 if f > 0.5 else 0)
npwrs = 2 * (ll + 1) # Maximum number of powers needed in convolutions
gpower = np.empty(npwrs) # gpower = (g/n)^m/m!
twogpower = np.empty(npwrs) # twogpower = (2g/n)^m/m!
onem2gpower = np.empty(npwrs) # onem2gpower = ((1-2g)/n)^m/m!
# gpower etc are *almost* Poisson probs, just missing normalizing factor.
gpower[0] = 1.0
twogpower[0] = 1.0
onem2gpower[0] = 1.0
expnt = 0
g_over_n, two_g_over_n, one_minus_two_g_over_n = g/n, 2*g/n, (1 - 2*g)/n
for m in range(1, npwrs):
gpower[m] = gpower[m - 1] * g_over_n / m
twogpower[m] = twogpower[m - 1] * two_g_over_n / m
onem2gpower[m] = onem2gpower[m - 1] * one_minus_two_g_over_n / m
V0 = np.zeros([npwrs])
V1 = np.zeros([npwrs])
V1[0] = 1 # first row
V0s, V1s = 0, 0 # start indices of the two rows
j1, j2 = _pomeranz_compute_j1j2(0, n, ll, ceilf, roundf)
for i in range(1, 2 * n + 2):
# Preserve j1, V1, V1s, V0s from last iteration
k1 = j1
V0, V1 = V1, V0
V0s, V1s = V1s, V0s
V1.fill(0.0)
j1, j2 = _pomeranz_compute_j1j2(i, n, ll, ceilf, roundf)
if i == 1 or i == 2 * n + 1:
pwrs = gpower
else:
pwrs = (twogpower if i % 2 else onem2gpower)
ln2 = j2 - k1 + 1
if ln2 > 0:
conv = np.convolve(V0[k1 - V0s:k1 - V0s + ln2], pwrs[:ln2])
conv_start = j1 - k1 # First index to use from conv
conv_len = j2 - j1 + 1 # Number of entries to use from conv
V1[:conv_len] = conv[conv_start:conv_start + conv_len]
# Scale to avoid underflow.
if 0 < np.max(V1) < _EM128:
V1 *= _EP128
expnt -= _E128
V1s = V0s + j1 - k1
# multiply by n!
ans = V1[n - V1s]
for m in range(1, n + 1):
if np.abs(ans) > _EP128:
ans *= _EM128
expnt += _E128
ans *= m
# Undo any intermediate scaling
if expnt != 0:
ans = np.ldexp(ans, expnt)
ans = _select_and_clip_prob(ans, 1.0 - ans, cdf)
return ans
def _kolmogn_PelzGood(n, x, cdf=True):
"""Computes the Pelz-Good approximation to Prob(Dn <= x) with 0<=x<=1.
Start with Li-Chien, Korolyuk approximation:
Prob(Dn <= x) ~ K0(z) + K1(z)/sqrt(n) + K2(z)/n + K3(z)/n**1.5
where z = x*sqrt(n).
Transform each K_(z) using Jacobi theta functions into a form suitable
for small z.
Pelz-Good (1976). [6]
"""
if x <= 0.0:
return _select_and_clip_prob(0.0, 1.0, cdf=cdf)
if x >= 1.0:
return _select_and_clip_prob(1.0, 0.0, cdf=cdf)
z = np.sqrt(n) * x
zsquared, zthree, zfour, zsix = z**2, z**3, z**4, z**6
qlog = -_PI_SQUARED / 8 / zsquared
if qlog < _MIN_LOG: # z ~ 0.041743441416853426
return _select_and_clip_prob(0.0, 1.0, cdf=cdf)
q = np.exp(qlog)
# Coefficients of terms in the sums for K1, K2 and K3
k1a = -zsquared
k1b = _PI_SQUARED / 4
k2a = 6 * zsix + 2 * zfour
k2b = (2 * zfour - 5 * zsquared) * _PI_SQUARED / 4
k2c = _PI_FOUR * (1 - 2 * zsquared) / 16
k3d = _PI_SIX * (5 - 30 * zsquared) / 64
k3c = _PI_FOUR * (-60 * zsquared + 212 * zfour) / 16
k3b = _PI_SQUARED * (135 * zfour - 96 * zsix) / 4
k3a = -30 * zsix - 90 * z**8
K0to3 = np.zeros(4)
# Use a Horner scheme to evaluate sum c_i q^(i^2)
# Reduces to a sum over odd integers.
maxk = int(np.ceil(16 * z / np.pi))
for k in range(maxk, 0, -1):
m = 2 * k - 1
msquared, mfour, msix = m**2, m**4, m**6
qpower = np.power(q, 8 * k)
coeffs = np.array([1.0,
k1a + k1b*msquared,
k2a + k2b*msquared + k2c*mfour,
k3a + k3b*msquared + k3c*mfour + k3d*msix])
K0to3 *= qpower
K0to3 += coeffs
K0to3 *= q
K0to3 *= _SQRT2PI
# z**10 > 0 as z > 0.04
K0to3 /= np.array([z, 6 * zfour, 72 * z**7, 6480 * z**10])
# Now do the other sum over the other terms, all integers k
# K_2: (pi^2 k^2) q^(k^2),
# K_3: (3pi^2 k^2 z^2 - pi^4 k^4)*q^(k^2)
# Don't expect much subtractive cancellation so use direct calculation
q = np.exp(-_PI_SQUARED / 2 / zsquared)
ks = np.arange(maxk, 0, -1)
ksquared = ks ** 2
sqrt3z = _SQRT3 * z
kspi = np.pi * ks
qpwers = q ** ksquared
k2extra = np.sum(ksquared * qpwers)
k2extra *= _PI_SQUARED * _SQRT2PI/(-36 * zthree)
K0to3[2] += k2extra
k3extra = np.sum((sqrt3z + kspi) * (sqrt3z - kspi) * ksquared * qpwers)
k3extra *= _PI_SQUARED * _SQRT2PI/(216 * zsix)
K0to3[3] += k3extra
powers_of_n = np.power(n * 1.0, np.arange(len(K0to3)) / 2.0)
K0to3 /= powers_of_n
if not cdf:
K0to3 *= -1
K0to3[0] += 1
Ksum = sum(K0to3)
return Ksum
def _kolmogn(n, x, cdf=True):
"""Computes the CDF(or SF) for the two-sided Kolmogorov-Smirnov statistic.
x must be of type float, n of type integer.
Simard & L'Ecuyer (2011) [7].
"""
if np.isnan(n):
return n # Keep the same type of nan
if int(n) != n or n <= 0:
return np.nan
if x >= 1.0:
return _select_and_clip_prob(1.0, 0.0, cdf=cdf)
if x <= 0.0:
return _select_and_clip_prob(0.0, 1.0, cdf=cdf)
t = n * x
if t <= 1.0: # Ruben-Gambino: 1/2n <= x <= 1/n
if t <= 0.5:
return _select_and_clip_prob(0.0, 1.0, cdf=cdf)
if n <= 140:
prob = np.prod(np.arange(1, n+1) * (1.0/n) * (2*t - 1))
else:
prob = np.exp(_log_nfactorial_div_n_pow_n(n) + n * np.log(2*t-1))
return _select_and_clip_prob(prob, 1.0 - prob, cdf=cdf)
if t >= n - 1: # Ruben-Gambino
prob = 2 * (1.0 - x)**n
return _select_and_clip_prob(1 - prob, prob, cdf=cdf)
if x >= 0.5: # Exact: 2 * smirnov
prob = 2 * scipy.special.smirnov(n, x)
return _select_and_clip_prob(1.0 - prob, prob, cdf=cdf)
nxsquared = t * x
if n <= 140:
if nxsquared <= 0.754693:
prob = _kolmogn_DMTW(n, x, cdf=True)
return _select_and_clip_prob(prob, 1.0 - prob, cdf=cdf)
if nxsquared <= 4:
prob = _kolmogn_Pomeranz(n, x, cdf=True)
return _select_and_clip_prob(prob, 1.0 - prob, cdf=cdf)
# Now use Miller approximation of 2*smirnov
prob = 2 * scipy.special.smirnov(n, x)
return _select_and_clip_prob(1.0 - prob, prob, cdf=cdf)
# Split CDF and SF as they have different cutoffs on nxsquared.
if not cdf:
if nxsquared >= 370.0:
return 0.0
if nxsquared >= 2.2:
prob = 2 * scipy.special.smirnov(n, x)
return _clip_prob(prob)
# Fall through and compute the SF as 1.0-CDF
if nxsquared >= 18.0:
cdfprob = 1.0
elif n <= 100000 and n * x**1.5 <= 1.4:
cdfprob = _kolmogn_DMTW(n, x, cdf=True)
else:
cdfprob = _kolmogn_PelzGood(n, x, cdf=True)
return _select_and_clip_prob(cdfprob, 1.0 - cdfprob, cdf=cdf)
def _kolmogn_p(n, x):
"""Computes the PDF for the two-sided Kolmogorov-Smirnov statistic.
x must be of type float, n of type integer.
"""
if np.isnan(n):
return n # Keep the same type of nan
if int(n) != n or n <= 0:
return np.nan
if x >= 1.0 or x <= 0:
return 0
t = n * x
if t <= 1.0:
# Ruben-Gambino: n!/n^n * (2t-1)^n -> 2 n!/n^n * n^2 * (2t-1)^(n-1)
if t <= 0.5:
return 0.0
if n <= 140:
prd = np.prod(np.arange(1, n) * (1.0 / n) * (2 * t - 1))
else:
prd = np.exp(_log_nfactorial_div_n_pow_n(n) + (n-1) * np.log(2 * t - 1))
return prd * 2 * n**2
if t >= n - 1:
# Ruben-Gambino : 1-2(1-x)**n -> 2n*(1-x)**(n-1)
return 2 * (1.0 - x) ** (n-1) * n
if x >= 0.5:
return 2 * scipy.stats.ksone.pdf(x, n)
# Just take a small delta.
# Ideally x +/- delta would stay within [i/n, (i+1)/n] for some integer a.
# as the CDF is a piecewise degree n polynomial.
# It has knots at 1/n, 2/n, ... (n-1)/n
# and is not a C-infinity function at the knots
delta = x / 2.0**16
delta = min(delta, x - 1.0/n)
delta = min(delta, 0.5 - x)
def _kk(_x):
return kolmogn(n, _x)
return _derivative(_kk, x, dx=delta, order=5)
def _kolmogni(n, p, q):
"""Computes the PPF/ISF of kolmogn.
n of type integer, n>= 1
p is the CDF, q the SF, p+q=1
"""
if np.isnan(n):
return n # Keep the same type of nan
if int(n) != n or n <= 0:
return np.nan
if p <= 0:
return 1.0/n
if q <= 0:
return 1.0
delta = np.exp((np.log(p) - scipy.special.loggamma(n+1))/n)
if delta <= 1.0/n:
return (delta + 1.0 / n) / 2
x = -np.expm1(np.log(q/2.0)/n)
if x >= 1 - 1.0/n:
return x
x1 = scu._kolmogci(p)/np.sqrt(n)
x1 = min(x1, 1.0 - 1.0/n)
def _f(x):
return _kolmogn(n, x) - p
return scipy.optimize.brentq(_f, 1.0/n, x1, xtol=1e-14)
def kolmogn(n, x, cdf=True):
"""Computes the CDF for the two-sided Kolmogorov-Smirnov distribution.
The two-sided Kolmogorov-Smirnov distribution has as its CDF Pr(D_n <= x),
for a sample of size n drawn from a distribution with CDF F(t), where
:math:`D_n &= sup_t |F_n(t) - F(t)|`, and
:math:`F_n(t)` is the Empirical Cumulative Distribution Function of the sample.
Parameters
----------
n : integer, array_like
the number of samples
x : float, array_like
The K-S statistic, float between 0 and 1
cdf : bool, optional
whether to compute the CDF(default=true) or the SF.
Returns
-------
cdf : ndarray
CDF (or SF it cdf is False) at the specified locations.
The return value has shape the result of numpy broadcasting n and x.
"""
it = np.nditer([n, x, cdf, None],
op_dtypes=[None, np.float64, np.bool_, np.float64])
for _n, _x, _cdf, z in it:
if np.isnan(_n):
z[...] = _n
continue
if int(_n) != _n:
raise ValueError(f'n is not integral: {_n}')
z[...] = _kolmogn(int(_n), _x, cdf=_cdf)
result = it.operands[-1]
return result
def kolmognp(n, x):
"""Computes the PDF for the two-sided Kolmogorov-Smirnov distribution.
Parameters
----------
n : integer, array_like
the number of samples
x : float, array_like
The K-S statistic, float between 0 and 1
Returns
-------
pdf : ndarray
The PDF at the specified locations
The return value has shape the result of numpy broadcasting n and x.
"""
it = np.nditer([n, x, None])
for _n, _x, z in it:
if np.isnan(_n):
z[...] = _n
continue
if int(_n) != _n:
raise ValueError(f'n is not integral: {_n}')
z[...] = _kolmogn_p(int(_n), _x)
result = it.operands[-1]
return result
def kolmogni(n, q, cdf=True):
"""Computes the PPF(or ISF) for the two-sided Kolmogorov-Smirnov distribution.
Parameters
----------
n : integer, array_like
the number of samples
q : float, array_like
Probabilities, float between 0 and 1
cdf : bool, optional
whether to compute the PPF(default=true) or the ISF.
Returns
-------
ppf : ndarray
PPF (or ISF if cdf is False) at the specified locations
The return value has shape the result of numpy broadcasting n and x.
"""
it = np.nditer([n, q, cdf, None])
for _n, _q, _cdf, z in it:
if np.isnan(_n):
z[...] = _n
continue
if int(_n) != _n:
raise ValueError(f'n is not integral: {_n}')
_pcdf, _psf = (_q, 1-_q) if _cdf else (1-_q, _q)
z[...] = _kolmogni(int(_n), _pcdf, _psf)
result = it.operands[-1]
return result