Release of CmdStan 2.37.0

Announcements
1

We are very happy to announce that CmdStan 2.37.0 is out!

This release cycle brings improvements to Stan’s constrained types, quality-of-life improvements to CmdStan, and many stability improvements and bug fixes. Thank you to all the developers who contributed to this release!

For details, see the blog post: Release of CmdStan 2.37 – The Stan Blog

18 Likes
2

Thank you!

I am sorry, but I didn’t understand how to use, for example, this one and what it means: <transform>_unconstrain. Are there any examples?

3

Starting in this version, there are functions available for each of the built-in transforms. For example, you have long been able to declare a simplex[N] theta;. Now, you can call the functions simplex_jacobian (or simplex_constrain, which does the same transform but does not perform a change of variables adjustment) that takes an unconstrained vector of length N-1 and returns an N-simplex, or simplex_unconstrain which is the inverse operation.

These are useful in some situations where you want to have several constrained variables of different sizes – you can declare one parameter that has the size you need in total, and then constrain slices of that parameter. @mitzimorris used this trick in her case study on the sum to zero vector, but she had to write the sum_to_zero_constrain function herself to do it. Now, Stan gives it to you out of the box

1 Like
4

FWIW, I have an example here of the functions to produce multiple simplexes in a matrix. For each group m I need a simplex of size J[m]. So I first produce a vector of probabilities that’s size sum(J - 1).

parameters {
  vector[J_sum - M] u;  // unnormalised entry probabilities
}

Then in the transformed parameters I do the following, slicing out of u what I need for each group m:

transformed parameters {
  matrix[J_max, M] beta;  // matrix of entry probabilities (simplexes)
  for (m in 1:M) {
    beta[:J[m], m] = simplex_jacobian(u[sum(Jm1[:m - 1]) + 1:sum(Jm1[:m])]);
  }
}

Now in the model block I can put a Dirichlet prior on the simplex in the matrix beta. Hope this helps!

EDIT: In my original post I defined the unnormalised probabilities as vector<lower=0, upper=1>[J_sum - M] u, but the lower and upper constraints shouldn’t have been there.

1 Like
5

We’ve updated the tarballs associated with this release this morning after we noticed a build machine issue had lead to the wrong version being reported by stanc –version. No code has changed, and there’s no reason explicitly why someone would need to re-download if they already have, I just wanted to note it here in case someone was wondering why the files had changed

6

9 posts were split to a new topic: Differences in simplex constraining function vs normal simplex