Time series implementation questions

[maedoc]

@ricardoV94

The Scan in the rv_op is only concerned with generating random draws.

thanks, that makes a lot more sense now!

@ricardoV94

Scan implementation (or it’s gradient) has a different speed or numerical precision than the vectorized trick.

also makes sense, I get the use of scan as a more general way than vectorization to keep the computational graph compact (it’s strange that Stan never got a scan).

@ricardoV94

how the logp is implemented when I said centered. So I probably misunderstood your point. I meant that the logp function expects absolute “concrete” values, not innovations as inputs

I think we agree on this: for the current implementation, the sampler (or optimizer or whatever) is moving in the space of the values of the time series rv, not the space of the innovations.

That means there could be an alternative implementation which distinguishes the innovations as the parameter and does a scan to compute the concrete values of the time series rv, which would be non-centered, and in principle better for effective sample size / second (with HMC at least). We also observed that the stochastic Heun method accelerates convergence despite require 2x function calls, so I was aiming for that. I think I see how to do it now with scan.