parser should flag use of _lxxf suffix on built-in functions in sampling statements

[mitzimorris]

Summary:

currently parser allows: foo ~ normal_lpdf(0, 1); due to suffix stripping logic; should flag as incorrect.

Description:

parser has suffix stripping logic for user-defined probability functions, but here this applies to built-in probability distributions as well.

Reproducible Steps:

this shouldn't compile but it does:

data {
  int<lower=0> N;
}
parameters {
  vector[N] foo;
}
model {
  foo ~ normal_lpdf(0, 1);
}

Current Output:

currently OK; should get flagged as error.

Expected Output:

Describe what you expect the output to be. Knowing the correct behavior is also very useful.

Additional Information:

Provide any additional information here.

Current Version:

v2.18.0

[VMatthijs]

Stanc3 currently says

Semantic error at file "test.stan", line 8, characters 2-26:
   -------------------------------------------------
     6:  }
     7:  model {
     8:    foo ~ normal_lpdf(0, 1);
           ^
     9:  }
   -------------------------------------------------

Ill-typed arguments to '~' statement. No distribution normal_lpdf was found with the correct signature.

[bob-carpenter]

This one may require a bit more care as it's a common user error. Assuming foo is real or a 1D container type, this is the right signature for normal_lpdf(foo | 0, 1). So the function with the correct signature exists, but it shouldn't be called with the "_lpdf" suffix. Is there a way to catch this case when users include _lpmf or _lpdf by mistake and signal that they should remove the suffix? Error messages are much more useful if they suggest a fix.

…

On Dec 13, 2018, at 7:41 AM, Matthijs Vákár ***@***.***> wrote: Stanc3 currently says Semantic error at file "test.stan", line 8, characters 2-26: ------------------------------------------------- 6: } 7: model { 8: foo ~ normal_lpdf(0, 1); ^ 9: } ------------------------------------------------- Ill-typed arguments to '~' statement. No distribution normal_lpdf was found with the correct signature. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub, or mute the thread.

[VMatthijs]

OK. Just added a custom error message for these cases:

Semantic error at file "test.stan", line 8, columns 8-19:
   -------------------------------------------------
     6:  }
     7:  model {
     8:    foo ~ normal_lpdf(0, 1);
                 ^
     9:  }
   -------------------------------------------------

~-statement expects a distribution name without '_lpdf' or '_lcdf' suffix.

[bob-carpenter]

It's "lpdf" and "lpmf" (log probability { density / mass } functions). I don't think users will be trying to use "lcdf" or "lccdf" (log {complementary / } cumulative distribution functions).

…

On Dec 13, 2018, at 9:40 AM, Matthijs Vákár ***@***.***> wrote: OK. Just added a custom error message for these cases: Semantic error at file "test.stan", line 8, columns 8-19: ------------------------------------------------- 6: } 7: model { 8: foo ~ normal_lpdf(0, 1); ^ 9: } ------------------------------------------------- ~-statement expects a distribution name without '_lpdf' or '_lcdf' suffix. — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[VMatthijs]

Thanks! That was definitely a typo. Fixed now.

[mcol]

Fixed in stanc3. The error message is now the following:

Semantic error in 'test.stan', line 8, column 8 to column 19:
   -------------------------------------------------
     6:  }
     7:  model {
     8:    foo ~ normal_lpdf(0, 1);
                 ^
     9:  }
   -------------------------------------------------


~ statement should refer to a distribution without its "_lpdf" or "_lpmf" suffix.
For example, "target += normal_lpdf(y, 0, 1)" should become "y ~ normal(0, 1)."