scipy.stats.multinomial = <scipy.stats._multivariate.multinomial_gen object>[source]

A multinomial random variable.


x : array_like

Quantiles, with the last axis of x denoting the components.

n : int

Number of trials

p : array_like

Probability of a trial falling into each category; should sum to 1

random_state : None or int or np.random.RandomState instance, optional

If int or RandomState, use it for drawing the random variates. If None (or np.random), the global np.random state is used. Default is None.

See also

The binomial distribution.
Sampling from the multinomial distribution.


n should be a positive integer. Each element of p should be in the interval \([0,1]\) and the elements should sum to 1. If they do not sum to 1, the last element of the p array is not used and is replaced with the remaining probability left over from the earlier elements.

Alternatively, the object may be called (as a function) to fix the n and p parameters, returning a “frozen” multinomial random variable:

The probability mass function for multinomial is

\[f(x) = \frac{n!}{x_1! \cdots x_k!} p_1^{x_1} \cdots p_k^{x_k},\]

supported on \(x=(x_1, \ldots, x_k)\) where each \(x_i\) is a nonnegative integer and their sum is \(n\).

New in version 0.19.0.


>>> from scipy.stats import multinomial
>>> rv = multinomial(8, [0.3, 0.2, 0.5])
>>> rv.pmf([1, 3, 4])

The multinomial distribution for \(k=2\) is identical to the corresponding binomial distribution (tiny numerical differences notwithstanding):

>>> from scipy.stats import binom
>>> multinomial.pmf([3, 4], n=7, p=[0.4, 0.6])
>>> binom.pmf(3, 7, 0.4)

The functions pmf, logpmf, entropy, and cov support broadcasting, under the convention that the vector parameters (x and p) are interpreted as if each row along the last axis is a single object. For instance:

>>> multinomial.pmf([[3, 4], [3, 5]], n=[7, 8], p=[.3, .7])
array([0.2268945,  0.25412184])

Here, x.shape == (2, 2), n.shape == (2,), and p.shape == (2,), but following the rules mentioned above they behave as if the rows [3, 4] and [3, 5] in x and [.3, .7] in p were a single object, and as if we had x.shape = (2,), n.shape = (2,), and p.shape = (). To obtain the individual elements without broadcasting, we would do this:

>>> multinomial.pmf([3, 4], n=7, p=[.3, .7])
>>> multinomial.pmf([3, 5], 8, p=[.3, .7])

This broadcasting also works for cov, where the output objects are square matrices of size p.shape[-1]. For example:

>>> multinomial.cov([4, 5], [[.3, .7], [.4, .6]])
array([[[ 0.84, -0.84],
        [-0.84,  0.84]],
       [[ 1.2 , -1.2 ],
        [-1.2 ,  1.2 ]]])

In this example, n.shape == (2,) and p.shape == (2, 2), and following the rules above, these broadcast as if p.shape == (2,). Thus the result should also be of shape (2,), but since each output is a \(2 \times 2\) matrix, the result in fact has shape (2, 2, 2), where result[0] is equal to multinomial.cov(n=4, p=[.3, .7]) and result[1] is equal to multinomial.cov(n=5, p=[.4, .6]).


pmf(x, n, p) Probability mass function.
logpmf(x, n, p) Log of the probability mass function.
rvs(n, p, size=1, random_state=None) Draw random samples from a multinomial distribution.
entropy(n, p) Compute the entropy of the multinomial distribution.
cov(n, p) Compute the covariance matrix of the multinomial distribution.