Both appear to be due to the tests depending on the sign of eigenvectors computed by numpy.linalg.eigh.
The second failure is simple: the first eigenvector, as computed by two separate routines, are checked to be equal. The sign of this as computed by numpy.linalg.eigh has now flipped, raising an error. This can easily be fixed by multiplying both by the sign of the first element (which has to be real) before comparing.
The first one is ultimately an error coming from the same place, but it's hidden a bit deeper in the function calls in the test case, and is arising due to new sign inconsistencies between the output of numpy.linalg.eigh and jax.numpy.linalg.eigh. (Both of course output mathematically correct values, but again, some of the eigenvectors output by numpy now have a flipped sign.) I think this can be fixed by more explicitly controlling the RotatingFrame objects being constructed in this test, so that the explicit diagonalizing basis details don't impact the checks in the test.
The tests:
and
now fail on
main.Both appear to be due to the tests depending on the sign of eigenvectors computed by
numpy.linalg.eigh.numpy.linalg.eighhas now flipped, raising an error. This can easily be fixed by multiplying both by the sign of the first element (which has to be real) before comparing.numpy.linalg.eighandjax.numpy.linalg.eigh. (Both of course output mathematically correct values, but again, some of the eigenvectors output bynumpynow have a flipped sign.) I think this can be fixed by more explicitly controlling theRotatingFrameobjects being constructed in this test, so that the explicit diagonalizing basis details don't impact the checks in the test.