Counts vs. Probabilities as Input

[Padraig20]

(I am new to this, so it may be a stupid question!)

Hi there,

thanks for such an amazing toolbox! I have a question regarding the input data on which we try to estimate the Dirichlet priors.

In the paper "Estimating a Dirichlet distribution", it mentions the input being a "vector whose elements sum to 1", i.e. probabilities.

Before applying the tool on a real dataset, I tried sampling counts from a Dirichlet distribution with fixed parameters. I calculated frequencies from these counts and got quite bad estimates for the parameters. When feeding the counts themselves into the algorithm, though, the estimated parameters are quite close to the real ones. I kept trying with different configurations and constantly saw a similar trend.

So, to clarify, does the algorithm take in counts or probabilities? (Or should it even matter?)

Thanks in advance :)