A lot of DTs when using offset/multiplier vs manual NCP

Modeling
1

Hi,

I’m using cmdstan v2.36.0 via {cmdstanr}. I’ve been seeing an unexpected difference between using the off/multiplier notation for NPC compared to doing it manually. Below is simple model that doesn’t use any data (just doing prior prediction) and it results in 75% of iterations with divergent transitions. Below you see me doing NCP manually, which works fine, but if I comment out the raw_pr lines and uncomment the lines needed for offset/multiplier I get the DTs.

This is how I run the model

mm$sample(list(), parallel_chains = 4, seed = 453678)

transformed data {
  int n = 200;
}

parameters {
  real m;            
  real<lower=0> s;
  // vector<offset = m, multiplier = s>[n] pr;
  vector[n] raw_pr;
}

transformed parameters {
  vector[n] pr = m + s * raw_pr;
}

model {
  m ~ normal(-0.55, 1.0);
  s ~ normal(0, 0.5);
  // pr ~ normal(m, s);
  raw_pr ~ std_normal();
 }
2

This is a real problem. It’s not a bug in the sense of an implementation error, it’s just a failure of our current initialization and warmup procedure. See the following current thread for an explanation of what’s going on: Offset multiplier initialization - #44 by spinkney

1 Like
3

Ok, thanks! Sorry for creating a new thread for the same problem. I thought that one was mostly about initialization. I’ll keep an eye on that thread to see how things develop.

4

It is—the initialization turns out to be the problem. If you sample from the stationary distribution to start, everything should be fine.