Issue with simplex_free in 2.28.0

[rok-cesnovar]

This issue was reported to me by @spinkney and it seems to have made its way to the 2.28.0 release.

The model:

data {
  int<lower=0> K;
  vector[10] y;
}
parameters {
  simplex[K] s;
  real<lower=0> lambda;
}
model {
  y ~ exponential(lambda);
}

If you sample without specifying any initial values, it samples fine, but if you specify the initial value only for lambda it outputs:

Rejecting initial value:
  Error evaluating the log probability at the initial value.
Exception: stan::math::simplex_free: Simplex variable is not a valid simplex. sum(Simplex variable) = 2.384263259, but should be 1 (found before start of program)
Rejecting initial value:
  Error evaluating the log probability at the initial value.
Exception: stan::math::simplex_free: Simplex variable is not a valid simplex. sum(Simplex variable) = 2.274321496, but should be 1 (found before start of program)

This however works fine with 2.27.0 (sampling finishes with 0 warnings or errors).

It also fails with cmdstan 2.27.0 and the 2.28.0 stanc3 binary which leads me to believe this is a stanc3 issue.

Here is a folder with all the required files to reproduce:
reprex.zip

Reproduce simply by running:

./model sample data file=data.json init=init.json

You can also run with cmdstanr:

library(cmdstanr)

code <- "
data {
  int<lower=0> K;
  vector[10] y;
}
parameters {
  simplex[K] s;
  real<lower=0> lambda;
}
model {
  y ~ exponential(lambda);
}
"
stan_file <- write_stan_file(code)
mod_test <- cmdstan_model(stan_file)

fit_opt <- mod_test$sample(
  data = list(K = 3, y = rexp(n = 10, 4)),
  init = list(list(lambda = 2)),
  chains = 1
)

[rok-cesnovar]

It works fine prior to #872, with that PR it seems to segfault and after #978 it fails the way its described above.

cc: @rybern @SteveBronder

[SteveBronder]

Yes I'm looking right now, I have can see what's happening and am working on a fix

[rok-cesnovar]

Thank you good sir!

[SteveBronder]

np! Figured it out and should have a fix this afternoon. I think this is worth putting out a 2.28.1 if we put out 2.28 already. Essentially the init returns one size too big for simplex and unit constrained types (which should be size K for input from the user when going to the constrained space and K - 1 when read to the read of stan after going through transform inits

[rok-cesnovar]

Yes, definitely worthy of 2.28.1.

[spinkney]

Thanks @SteveBronder!