Priors on beta matrix rows

[NicholasWE]

Thanks for the inputs, @Bob_Carpenter!

@Bob_Carpenter

I’m afraid not. But the loop isn’t the problem here, it’s the multi_normal_cholesky solves that will dominate the computation here (much better than dense matrices as it’s only quadratic instead of cubic).

So, what you meant here is that this way I’m doing is probably the best one we have so far and that, besides it consumes a lot of computational effort, the use of multi_normal_cholesky distribution is the best choice as well, right?

@Bob_Carpenter

I don’t think you can vectorize this in a useful way, but it’s not going to matter. Stan’s loops are very fast. The only thing that’s useful about vectorizing is that it can reduce the size of the autodiff expression graph for derivatives, but that requires customization on a per-function basis, and we don’t have anything like that built-in. The penalty for loops here will be negligible compared to the cost of the density evals and gradients.

Do you think I can reduce this computational time/cost by using reduce_sum functions?

I tried this approach here with a simpler example, but it seems the code didn’t get faster.
I’d appreciate your opinion in this topic if you can, as well.

Thanks a lot!