[Refactoring] Avoid calling block to tensor on diagonal Hamiltonians

[jpmoutinho]

When calling HamEvo on diagonal Hamiltonians the dense matrix is built but then only the diagonal is taken.

This is currently done in the PyQHamiltonianEvolution class in qadence/backends/pyqtorch/convert_ops.py, where the __init__ method first builds the full matrix hmat, and then the _unitary method defines the evolution function for diagonal operators which runs diagonal() on the full hamiltonian matrix.

We should instead build only the diagonal. For non-parametric block this is directly possible with the function block_to_diagonal, as done in the PyQObservable class in the same file.

Handling parametric blocks is out of the scope of this issue, and can be done separately. It requires further development in the block_to_tensor side as detailed in #394.

NOTE: Due to ongoing changes in the codebase, we might need to implement this directly in the PyQTorch backend (pasqal-io/pyqtorch#175).

[neogyk]

Dear all,
I am trying to solve this issue.
Is it will be enough to add the following lines to the PyQHamiltonianEvolution class constructor:

elif not block.generator.is_parametric and block.generator.diagonal:
           m = block.generator.to(dtype=cdouble)
           hmat = block_to_diagonal(
               MatrixBlock(m, qubit_support=block.qubit_support),
               qubit_support=self.qubit_support,
               use_full_support=False,
           )
           hmat = hmat.permute(1, 2, 0)
           self.register_buffer("hmat", hmat)
           self._hamiltonian = lambda self, values: self.hmat

[jpmoutinho]

Hey @neogyk, thanks for your interest in completing this issue!

The solution should be along those lines, but there is no need to use MatrixBlock. You should be able to run block_to_diagonal on the block.generator.

Feel free to open an MR and add me as a reviewer, then we can discuss it there.

[neogyk]

Dear @jpmoutinho, could you please review the changes. I made the PR - #441

[Roland-djee]

Dear @jpmoutinho, could you please review the changes. I made the PR - #441

Dear @neogyk, just for your information, @jpmoutinho is off today. He will catch-up as soon as he's back, ie. tomorrow. Thanks.

[jpmoutinho]

Closing this issue as completed by @neogyk. The solution will remain unmerged in #478 as explained there.