Re: functions

On Friday 26 March 2004 01:31, you wrote:
> Each weight is a binomially distributed random variate with
> probability 1/N and N trials.  For selecting the first weight, we
> simply select a random variate from that distribution.  For
> selecting each successive weight, we select a random variate from
> the conditional distribution given what we already know.  I
> suspect that this conditional distribution is itself a binomially
> distributed random variate with probability 1/(N - n) and N - r
> trials, where r is the number of weights assigned so far and n is
> the sum of those weights, but I have not tried to prove it and I
> could be completely wrong.
>
> Does this sound right?

It is right except you've got the definitions of r and n reversed.

You didn't respond to the next bit though:

> > Anyway, we can figure that out later. How about the general idea of
> > setting up mechanism whereby procedures can be piped together?  What's
> > the cost?  If it's too hard, I guess you should concentrate on setting up
> > procedures for multiple passes (which we need sooner).  I don't want to
> > be nuisance.  Do it if it's relatively easy, but as I mentioned, the
> > first goal should probably be compatability.

Do I take that to mean that you want to avoid this or were you more concerned 
about storing the weight vector in core memory?

- M.