von Mises distribution fails for kappa=0

[mbrubake]

Summary:

The von Mises distribution returns a non-finite value when kappa=0. This should result in a uniform distribution and likely needs to be special cased.

Reproducible Steps:

The following model should produce a uniform distribution of test_var but instead, fails at initialization:

parameters {
    real test_var;
}
model {
    test_var ~ von_mises(0,0);
}

Current Output:

Initialization between (-2, 2) failed after 100 attempts. 
 Try specifying initial values, reducing ranges of constrained values, or reparameterizing the model.

Expected Output:

Should sample easily and produce a uniform distribution on test_var.

Current Version:

v2.15.0

[bgoodri]

I have this reworked on a branch

…

On Wed, Jun 21, 2017 at 1:09 PM, Marcus Brubaker ***@***.***> wrote: Summary: The von Mises distribution returns a non-finite value when kappa=0. This should result in a uniform distribution and likely needs to be special cased. Reproducible Steps: The following model should produce a uniform distribution of test_var but instead, fails at initialization: parameters { real test_var; } model { test_var ~ von_mises(0,0); } Current Output: Initialization between (-2, 2) failed after 100 attempts. Try specifying initial values, reducing ranges of constrained values, or reparameterizing the model. Expected Output: Should sample easily and produce a uniform distribution on test_var. Current Version: v2.15.0 — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#564>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADOrqly4c5ocbA_dWMMMmvvtSoV_YU0Zks5sGU45gaJpZM4OBQg4> .

[bob-carpenter]

@bgoodri Is it a branch that you could turn into a pull request for this issue? If not, is it a public branch?

[bgoodri]

Yeah, I can make push it as a public branch, but I still haven't implemented your suggestions.

…

On Wed, Jun 21, 2017 at 4:23 PM, Bob Carpenter ***@***.***> wrote: @bgoodri <https://github.com/bgoodri> Is it a branch that you could turn into a pull request for this issue? If not, is it a public branch? — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#564 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADOrqnhJWyG7zvg-KhdKAJqUnFh_eFOvks5sGXvfgaJpZM4OBQg4> .

[mcol]

The example model now produces the following error (repeated 100 times for each chain):

Chain 1 Exception: von_mises_lpdf: Scale parameter is 0, but must be > 0!

If I remove the check_positive_finite line for kappa, then CmdStan reports:

Exception initializing step size.
Posterior is improper. Please check your model.

which is correct, I think, as test_var would be distributed uniformly over all reals. Changing the model to the following, it samples correctly:

parameters {
  real<lower=-5, upper=5> test_var;
}
model {
  test_var ~ von_mises(0,0);
}

So, am I correct in thinking that @bgoodri's changes to the Bessel functions has effectively fixed this (modulo the overly stringent check on kappa)?