[Bug] Invalid boolean negation of PyTorch Tensor in bfgsHessianInverse.py

[dodowawa]

I encountered an issue related to boolean operations on PyTorch tensors while running the tests with PyGRANSO 2.0.0.

My environment is approximately as follows:

• PyGRANSO: 2.0.0
• Python: 3.13.7
• PyTorch: 2.8.0
• OS: Windows 11

When running the tests, the following error occurred at line 188 of pygranso/private/bfgsHessianInverse.py:

RuntimeError: Boolean value of Tensor with more than one value is ambiguous

The corresponding code is:

notInf_flag = torch.all(not torch.isinf(H_vec))
notNan_flag = torch.all(not torch.isnan(H_vec))

Here, torch.isinf(H_vec) and torch.isnan(H_vec) return boolean tensors.

If H_vec contains multiple elements, Python's not operator attempts to convert the entire tensor into a single boolean value, which causes the error above.

I changed lines 188 and 189 to:

notInf_flag = torch.all(~torch.isinf(H_vec))
notNan_flag = torch.all(~torch.isnan(H_vec))

After this change, the relevant tests passed successfully.

For reference, the corresponding code in PyGRANSO 1.2.0 used element-wise comparisons:

notInf_flag = torch.all(torch.isinf(H_vec) == False)
notNan_flag = torch.all(torch.isnan(H_vec) == False)

Therefore, the change from PyGRANSO 1.2.0 to 2.0.0:

torch.isinf(H_vec) == False

to:

not torch.isinf(H_vec)

changes the actual semantics of the code.

For a regular Python bool, these two forms are essentially equivalent. However, for a PyTorch tensor, the former performs an element-wise comparison, while the latter attempts to evaluate the entire tensor as a Python boolean value.

I suggest changing the current implementation to something like:

notInf_flag = torch.all(~torch.isinf(H_vec))
notNan_flag = torch.all(~torch.isnan(H_vec))

or:

notInf_flag = torch.all(torch.logical_not(torch.isinf(H_vec)))
notNan_flag = torch.all(torch.logical_not(torch.isnan(H_vec)))

It may also be worth checking PyGRANSO 2.0.0 for similar changes where:

tensor == False

was mechanically replaced with:

not tensor

In particular, expressions such as the following may be worth reviewing:

not torch.isinf(...)
not torch.isnan(...)
not torch.isfinite(...)
not torch.any(...)
not torch.all(...)

If the expression returns a multi-element tensor, the same issue may occur.

Thank you for maintaining PyGRANSO. I hope this report and the proposed fix are helpful.