Hi,

I am aware this is a vague question and am only looking for ideas and suggestions.

I am using NUTS for sampling spherical harmonics (SH). I am currently playing around with SH-degree 20-30, resulting in 500-1000 model parameters.

I use default settings: 4 chains with 1000 warm-up and 1000 post warm-up iterations, 10 tree-depth (never been necessary to increase). Always get Rhat=1.

Here’s the question:

Sometimes a run can take 4 hours and sometimes 24 hours (Independent of whether I increase of decrease the amount of model paramters).

I am not using the computer for anything else, thus nothing should interfere with the resources available. The applied model and data does not change. The posterior is uni-modal.

Do any of you have any suggestions as to why it can vary so much?

Code included below, but data is ~400 MB, so I have not included it

Thanks in advance

Michael

```
data {
int<lower=0> L; // Length of data-array
int<lower=0> n; // SH-degree
int<lower=0> mMax; // Amount of model parameters
vector[L] d; // Declare size of data-array
matrix[L, mMax] G; // Declare size of design matrix
vector[mMax] limits;
}
parameters {
vector[mMax] m;
}
model {
vector[L] res;
// Prior distribution
for (i in 1:mMax) {
m[i] ~ normal(0, limits[i]/2);
}
// Likelihood
res <- d-G*m;
res ~ normal(0,1);
}
```