Repository navigation
NumPy: numpy.linalg.LinAlgError not thrown #4859
Description
🐛 Bug
The following code:
import numpy as np
A = np.array([[1,-2j],[2j,5]])
print(np.linalg.cholesky(A))
B = np.array([[1,3],[2,4]])
np.linalg.cholesky(B)Produces an error on normal CPython, just like the docs state it should.
[[1.+0.j 0.+0.j]
[0.+2.j 1.+0.j]]
Traceback (most recent call last):
File "/Users/henryschreiner/git/webpages/iscinumpy.gitlab.io/np.py", line 7, in <module>
np.linalg.cholesky(B)
File "/usr/local/lib/python3.12/site-packages/numpy/linalg/linalg.py", line 779, in cholesky
r = gufunc(a, signature=signature, extobj=extobj)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.12/site-packages/numpy/linalg/linalg.py", line 115, in _raise_linalgerror_nonposdef
raise LinAlgError("Matrix is not positive definite")
numpy.linalg.LinAlgError: Matrix is not positive definitePyodide does not get the exception, though, and instead just sets NaN's.
To Reproduce
>>> import numpy as np
>>> A = np.array([[1,-2j],[2j,5]])
>>> print(np.linalg.cholesky(A))
[[1.+0.j 0.+0.j]
[0.+2.j 1.+0.j]]
>>> B = np.array([[1,3],[2,4]])
>>> np.linalg.cholesky(B)
array([[nan, nan],
[nan, nan]])The same code produces an error outside a pyodide.
Expected behavior
It should throw numpy.linalg.LinAlgError.
Additional context
See scikit-hep/iminuit#995, this is breaking iminuit's test suite.
Activity
hoodmane commentedon Jun 10, 2024
Unfortunately, I think this is very hard for us to fix without webassembly getting support for floating point exception modes. This would require a whole process with the wasm committees. Probably someone should inquire whether it's a change they'd consider...
henryiii commentedon Jun 11, 2024
Maybe NumPy should check for NaN in situations like this if running on WASM?
lesteve commentedon Jun 23, 2024
Just for explicitness:
np.linalg.choleskyusesnp.errstate. Inside Pyodidenp.errstate(ornp.seterr) will never raise an exception because of lack of floating point exception support in webassembly. This was noted in Issues with exception raising after setting numpy.errstate #156 for example. I have seen cases wherenp.errstate+ try/catch is used for control flow and you sometimes end-up in a very confusing code branch.lesteve commentedon Jun 23, 2024
As a potential work-around (likely not applicable in iminuit but maybe useful in other cases)
scipy.linalg.choleskydoes raise an exception on the same data inside Pyodide:agriyakhetarpal commentedon Jun 23, 2024
I have noticed the same problem when running the test suite for
statsmodelsand xfailed for now: statsmodels/statsmodels#9270Shorten note about pyodide/pyodide#4859
agriyakhetarpal commentedon Aug 6, 2026
This will be fixed in NumPy 2.6, via numpy/numpy#32036.
Skip wasm-unsupported SciPy tests and avoid destructor abort