InverseGamma logcdf method fails with invalid parameters when array is used

[ricardoV94]

The theano.tensor function gammaincc fails with a C code assertion that exits the process, when called with negative arguments. This function is now being called in the logcdf method of the InverseGamma distribution (and after #4331, in the logcdf of the Poisson distribution).

Using bound avoids the exception when a single value is evaluated. However, it fails when an array/tensor with one or multiple values is used, since bound is not lazy.

Should we be concerned about this? From the docstrings, it seems that all logcdf methods are intended to work with both scalars and arrays/tensors values.

pm.InverseGamma.dist(alpha=-1, beta=5).logcdf(1).eval()  # returns -np.inf
pm.InverseGamma.dist(alpha=1, beta=5).logcdf(-1).eval()  # returns -np.inf

pm.InverseGamma.dist(alpha=-1, beta=5).logcdf(np.array([1])).eval()  # fails assertion and exits process
pm.InverseGamma.dist(alpha=1, beta=5).logcdf(np.array([-1])).eval()  # fails assertion and exits process
pm.InverseGamma.dist(alpha=-1, beta=5).logcdf(np.array([1, 1])).eval()  # fails assertion and exits process

Error message:

python3: /home/ricardo/.theano/compiledir_Linux-5.4--generic-x86_64-with-glibc2.29-x86_64-3.8.5-64/tmpci20dxir/mod.cpp:240: double GammaQ(double, double): Assertion `(n > 0) && (x >= 0)' failed.

[ricardoV94]

Same issue occurs with the Gamma distribution logcdf method as it relies on the gammainc, which also asserts values at the C level code

[ricardoV94]

Fixed by #4393. Fix can be reverted once this Theano issue is fixed: aesara-devs/aesara#224

In the meantime feel free to close this issue.

[twiecki]

Thanks @ricardoV94 for the deep dive into this!