Distributions in SciPy

This page summarizes how to work with univariate probability distributions using Python’s SciPy library. See also notes on working with distributions in Mathematica, Excel, and R/S-PLUS.

Probability distribution classes are located in scipy.stats.

The methods on continuous distribution classes are as follows.

Method Meaning
pdf Probability density function
cdf Cumulative distribution function
sf Survival function = complementary CDF
ppf Percentile point function (i.e. CDF inverse)
isf Inverse survival function (Complementary CDF inverse)
stats Mean, variance, skew, or kurtosis
moment Non-central moments
rvs Random samples

Functions such as pdf and cdf are defined over the entire real line. For example, the beta distribution is commonly defined on the interval [0, 1]. If you ask for the pdf outside this interval, you simply get 0. If you ask for the cdf to the left of the interval you get 0, and to the right of the interval you get 1.

Distributions have a general form and a “frozen” form. The general form is stateless: you supply the distribution parameters as arguments to every call. The frozen form creates an object with the distribution parameters set. For example, you could evaluate the PDF of a normal(3, 4) distribution at the value 5 by

	stats.norm.pdf(5, 3, 4)

or by

	mydist = stats.norm(3, 4)
	mydist.pdf(5)

Note that the argument of the PDF, in this example 5, comes before the distribution parameters. Note also that for discrete distributions, one would call pmf (probability mass function) rather than the pdf (probability density function).

Distributions and parameterizations

SciPy makes every continuous distribution into a location-scale family, including some distributions that typically do not have location scale parameters. This unusual approach has its advantages. For example, the question of whether an exponential distribution is parameterized in terms of its mean or its rate goes away: there is no mean or rate parameter per se, only a scale parameter like every other continuous distribution.

The table below only lists parameters in addition to location and scale.

Distribution SciPy name    Parameters
beta beta shape1, shape2
binomial binom size, prob
Cauchy cauchy
chi-squared chi2 df
exponential expon
F f df1, df2
gamma gamma shape
geometric geom p
hypergeometric hypergeom M, n, N
inverse gamma invgamma shape
log-normal lognorm sdlog
logistic logistic
negative binomial nbinom size, prob
normal norm
Poisson poisson lambda
Student t t df
uniform uniform
Weibull exponweib exponent, shape

SciPy does not have a simple Weibull distribution but instead has a generalization of the Weibull called the exponentiated Weibull. Set the exponential parameter to 1 and you get the ordinary Weibull distribution.

The hypergeometric distribution gives the probability of various numbers of red balls when N balls are taken from an urn containing n red balls and Mn blue balls. Note that another popular convention uses the number of red and blue balls rather than the number of red balls and the total number of balls.

Note that the parameters for the log-normal are the mean and standard deviation of the log of the distribution, not the mean and standard deviation of the distribution itself.

The PDF or PMF of a distribution is contained in the extradoc string. For example:

	>>> stats.poisson.extradoc
	Poisson distribution
	poisson.pmf(k, mu) = exp(-mu) * mu**k / k!
	for k >= 0

The lognormal distribution as implemented in SciPy may not be the same as the lognormal distribution implemented elsewhere. When the location parameter is 0, the stats.lognorm with parameter s corresponds to a lognormal(0, s) distribution as defined here. But if the location parameter is not 0, stats.lognorm does not correspond to a log-normal distribution under the other distribution. The difference is whether the PDF contains log(x-μ) or log(x) – μ.

For more information, see scipy.stats online documentation.