Modulus function for real / vector

Developers
1

Hello,

before I write some code to do this manually, I thought I would double check:

does stan currently support modulus arithmetic for real valued variables?

For example, in R I can write:

5.2 %% 2

And I get an answer of 1.2

The stan operator appears to be %, and only works on ints. Is this correct?

Thanks

2

This is correct. To support real numbers (in particular, parameters of the model) we would need to be able to take the derivative with respect to a % b which is undefined when b divides a. We have a few functions with misbehaved derivatives, but not any quite that bad I think

3

Thanks for the reply.

In my case, I am only using this function to create my predictor features when I come to compute generate quantities.

On the off chance that can point to a more sophisticated solution:

my model uses the angle between two vectors as a feature vector. When model fitting, I can calculate during pre-processing in R.

I am now adding an extensive generated quantities block and will be having to calculate these angular differences “on the fly” in Stan.

my code currently looks like this: (note that the last line is still R,… I am still in the process of working out how to convert this into Stan!)

// angle from item Q[jj-1] to each other item
psi_j = atan2((item_y[t] - item_y[t][Q[jj-1]]), (item_x[t] - item_x[t][Q[jj-1]])) * 180 / pi();

// now subtract the angle from Q[jj-2] and Q[jj-1]
psi_j = psi_j - atan2((item_y[t][Q[jj-1]] - item_y[t][Q[ jj-2]]), (item_x[t][Q[jj-1]] - item_x[t][Q[jj-2]])) * 180 / pi();
                
// now I want to scale psi_j to go from 0 to pi.
// 0 indicates that we are going forward... both vectors are going in the same direction!
// pi/2 means we have turned a right angle, either left or right.
// pi means that we are doubling back on ourself.
for (tt in 1:n_targets) {
    psi_j[tt] = pmin(abs((psi_j[tt] % 360)), abs((-psi_j[tt] % 360)))/180;
}