Multivariate random variables with scipy.stats rvs() function

Question

The scipy.stats suite of statistical distributions ( scipy.stats.norm , scipy.stats.uniform , scipy.stats.t etc) all produce univariate data series using their own .rvs() function, and only one has a multivariate rendition: multivariate_normal , which corresponds to numpy 's numpy.random.randn((N,K)) . In fact, virtually all of the statistical distributions found in numpy.random can produce multivariate data.

How can I extend the univariate distribution functions found in scipy.stats to multivariate number generation, given that it possesses some distributions not found in numpy.random like johnsonsu ? Must I manually make a loop function myself that concatenates multiple univariates together? what should that look like

Answer 1

I think you just want to pass a size parameter to rvs . For example:

from scipy import stats

stats.norm.rvs(size=10)

will give you a vector filled with 10 standard normal variates.

note that multivariate means something specific in statistics, not just IID copies of the same (which is what size does). eg the cov parameter to multivariate_normal specifies the covariance matrix of all variates within one draw.

as another example, multinomial is similar, but the parameters of the distribution are obviously different.

stats.multinomial.rvs(n=5, p=[0.5, 0.5], size=10)

tells you how many heads & tails you get from throwing a coin 5 times, repeating this 10 times. ie the rows are separate draws, the columns are the values within a draw