efficiency of sample() with prob.

It seems the distribution used for your tests is maximally favourable to 
your proposal (not uncommon in papers, but not very honest).  

I did not have time to do a thorough test. I was simply reporting what
I had done so any implied dishonesty is unfair to me.

Changing how this is done will break the reproducibility of past programs,
and we don't really want to introduce yet more options.  So it seems only
worth doing when there are substantial speed gains. 

I agree that Walker's method may be slow at some cases but there is
nothing to brag about R's current method either. The bottom line is: a
bisection method will be uniformly better than the linear search
method that R uses right now. This does not qualify as 'substantial
speed gain' though.

Cheers,
Bo