How do I generate samples from multivariate kernel density estimated distribution?

Unlike the univariate counterpart, there is no documentation for how to draw random samples from a multivariate kernel density estimation, as obtained from mvksdensity.

One possibility would be to query the mvksdensity at uniform random points, and accept the samples with the right probability.

Presumably one could replicate the estimated density using gmdistribution, with the number of components equal to the number of samples used in the kernel density estimation. But what is the right variance to use, and how does this relate to the bandwidth parameter used in mvksdensity?

Answer 1

[Linus Schumacher]

Ok, I've found the answer. The right sigma to use for gmdistribution seems to be bandwidth.^2

[Thomas Alderson]

How to do this?

[Sterling Baird]

Do you have a reference for using bw.^2 ?

[Linus Schumacher]

I can't remember, I either looked this up in the Matlab documentation, or tried it out with different bandwidth to make sure gmdistribution gives me the same results as mvksdensity – probably the latter

Answer 2

[Thomas Alderson]

How to do this?

[Linus Schumacher]

To sample from the KDE I built my own using gmdistribution, with one Gaussian distribution for each sample, and the standard deviation = bandwidth.^2