Fix invalid sparse Hessian dimensions while dropping fixed variables

[chrhansk]

Describe the change(s)

The internal sparse Hessian function assumes wrong dimensions when creating the Hessian matrix. A difference occurs when n != n_full, i.e., if variables are dropped (e.g. instance BOX2). This results in a ValueError('row index exceeds matrix dimesions') being thrown by scipy.

Have you updated the documentation?

Yes

Have you included relevant unit tests?

No, I cannot get the test framework to work. But loading BOX2 and executing sphess should be sufficient.

[jfowkes]

Many thanks for the proposed fix @chrhansk!

@lindonroberts would you be able to look into this? I'm not that familiar with the n_free/n_fixed variable handling, and I suspect a similar issue may be present in the other sparse matrix functions (e.g. gradsphess etc).

[chrhansk]

That is a valid point, maybe we should add the BOX2 example and call pretty much all functions (Jacobian, Hessian, and the combinations) to ensure everything works correctly.

[jfowkes]

That is a valid point, maybe we should add the BOX2 example and call pretty much all functions (Jacobian, Hessian, and the combinations) to ensure everything works correctly.

Yes indeed, might be good to add BOX2 to the sparse problem functions unit test here: pycutest/tests/test_sparse_functionality.py

[lindonroberts]

Thanks @chrhansk - yes, I think all the underscore sparse functions (scons, slagjac, ...) would have the same issue. And I agree with @jfowkes that it would worthwhile adding BOX2 to the sparse unit tests in pycutest/tests/test_sparse_functionality.

[jfowkes]

@lindonroberts many thanks, I've made the suggested changes and all seems to work, please do review when you have some time.

[lindonroberts]

Looks good to me, thanks @jfowkes!