Complex numbers and matrix inversion in Stan

[bgoodri]

Yeah, calling a C++ function from a Stan function is sort of unpredictable. I encountered the same thing here

Problem implementing Zipf (power law) distribution -- external C++

OK. I have the problem surrounded. This works, i.e. just declare real zeta(real alpha); but don’t define a power_law function that calls it: functions{ real zeta(real s); /*TODO: implement rng for prior and posterior predictive checks*/ } data{ int<lower=0> K; // number of unique values int values[K]; int<lower=0> frequencies[K]; } parameters{ real <lower=1> alpha; } model{ real constant = log(zeta(alpha)); for (k in 1:K) { target += frequencies[k] * (-alpha * log(values…

Anyway, it should be possible to wrap up everything into a C++ function that calls other C++ functions to do what you want. But at the moment, the complex number support is only in CmdStan 2.23.

[Isaac_Fradkin]

Ok. I see. I already have everything in external C++ functions that returns a logLL for a set of free parameters. Can I use this as a custom logLL function in the model block?

Also, I’m figuring that simply doing everything by using std::complex won’t do the job, since I need a stan var type complex for it to work, right? Is that because of the autodiff? What exactly is the var type? And how will a complex var type variable look like?

Thanks again!

Isaac