next up previous contents
Next: Bayesian interpretation of pseudocount Up: Regularizers for Estimating Distributions Previous: References

Partial derivatives for Dirichlet mixtures

 

The derivative of the Beta function is

displaymath2206

If we introduce the variable R to be the expression

displaymath2210

then the derivative of tex2html_wrap_inline1734 is

displaymath2214

where tex2html_wrap_inline2216 is 1 if i=n and 0 otherwise.

Using Stirling's approximation

displaymath2220

lets us approximate the derivative of tex2html_wrap_inline2222

displaymath2224

Since Stirling's approximation is not good for small x, we may have to use

displaymath2228

to move the value of the argument up into a region where the approximation is adequate:

displaymath2230

Second partial derivatives are easily computed from the first partials.



Rey Rivera
Thu Aug 1 17:59:45 PDT 1996