|WikiProject Statistics||(Rated Stub-class)|
Definition of scale parameter for Dirichlet distributions...
There seems to be some confusion in the literature about how to define a concentration parameter for a multi-variable Dirichlet distribution.
In several sources I've seen it defined such that a scale parameter of 1 leads to a uniform distribution. I've only really seen this in reference to symmetric Dirichlet distributions. In this case, the Dirichlet parameter set has scale parameter 3.
In other sources (notably in the topic modelling literature), the scale parameter is defined as the sum of the Dirichlet parameters for each dimension. In this case, the parameter has scale parameter 9 and "base measure" (1/3 , 1/3 , 1/3) I've not spent too much time researching this and havn't much in the way of references for the first definition (just wikipedia and similar web pages). I immagine this fisrt definition can be found in text books?? I have a few topic modelling papers that use the second definition. I have edited the main page and added a reference that uses this second definition. It's probably not the best reference, as it doesn't specifically describe how it defines "concentration parameter", but it is using the term in this second sense. Feel free to find a better reference (a text book, for example).
Dirichet process v Dirichlet distribution
I think there could be a better distinction made in the text between Dirichlet distributions and Dirichlet processes. I think the current text is liable to confuse. Unfortunately, I'm not expert enough to rewrite. — Preceding unsigned comment added by 126.96.36.199 (talk) 08:57, 10 November 2013 (UTC)