replicates should >= 0

[walterxie]

Add the validation to check if replicates >= 1 when using IID. Otherwise throw IllegalArgumentException (Runtime) and catch at the top level.

Update to replicates >= 0

[bredelings]

Hmm... requiring replicates to be >=1 makes it impossible to have the number of replicates be random.

If that is intentional, perhaps you should also check that replicates is random and complain about that as well?

Also, perhaps you should introduce an iid distribution for when people want to allow replicates to be 0 or 1?

@alexeid

[walterxie]

You can sample it from a Poisson distribution, and set min=1. This is an example from Alexei:

// replicates = 1, 2, 3, ...
k ~ Poisson(lambda=1.0, min=1);
pi ~ Dirichlet(conc=[1, 1, 1, 1], replicates=k);
L = length(pi);

But I cannot figure out what replicates=0 means?

[bredelings]

Oops, somehow I misread the issue title as "replicates should be > 1". Sorry. I was thinking that you were going to forbid replicates=1 in order to avoid a choice of how to interpret that. But now I see my mistake.

So, if replicates=1 returns x, then I also do not know what replicates=0 should return.

But if replicates=1 returns [x], then replicates=0 would return [].

The second case makes more sense to me because the return type does not depend on a parameter value. However, this is not as big a deal as I originally thought, since replicates=1 is still allowed.

One question is if there are models where replicates=0 occurs. If we are constructing an n-component mixture model by drawing n replicates from some distribution, then we do need n >= 1.

However, suppose we are looking at the number of change points in some interval. At each change point the evolutionary rate is multiplied by some value. Then we might find the rate at the end of the interval as follows:

n ~ Poisson(lambda)
scaleFactors ~ LogNormal(mu,sigma,replicates=n)
finalRate = initialRate * product(scaleFactors)

If there are no change points (n==0) then we want to get scaleFactors = [] and finalRate = initialRate.

This assumes that the product([]) == 1. I think this is pretty standard.