Support for partitioned active space CI methods (RAS, ORMAS, GAS)

[brycewestheimer]

Is your feature request related to a problem? Please describe.

For systems with multiple distinct correlation domains (transition metal + ligand, multiple open-shell sites, etc.), the full CASCI expansion becomes intractable even though most configurations contribute negligibly. ASCI helps with this, but it doesn't produce a determinant space with structure that maps to quantum circuit ansatz or allows factorization into independent subspace calculations.

Describe the solution you'd like

Support for RAS/ORMAS/GAS variants of CASCI, which partition the active orbitals into subspaces with occupation constraints and restrict the CI expansion accordingly. These methods use fixed orbitals (consistent with the existing CASCI approach) and offer two benefits for quantum computing: shallower ansatz (excitation operators violating constraints are excluded by construction) and, for ORMAS-CI specifically, the potential to evaluate subspaces as separate smaller quantum calculations with inter-subspace coupling handled classically.

Describe alternatives you've considered

ASCI (already available via MACIS) handles larger active spaces through determinant selection, but the resulting determinant space is irregular and can't be mapped to structured ansatz or factored into subspace calculations. Manually reducing the active space works but sacrifices chemically important correlation.

Additional context

The existing PySCF plugin provides a natural integration point, since PySCF's CASCI supports swappable FCI solvers. A custom fcisolver implementing restricted determinant enumeration could plug into the existing pipeline without changes to orbital preparation or Hamiltonian construction.

If Psi4 integration is upcoming, support for restricted active space methods could also be added via the Forte plugin.

[wavefunction91]

Hi @brycewestheimer,

Thanks for the suggestion! This would definitely be a good feature, and indeed, we've considered it internally. I can see a few potential paths for adding this kind of infrastructure to the code, some of which you've already enumerated. On the classical side, we can absolutely add some functionality to enumerate partitioned active space configurations, and as you said, there are turn-key solutions we could hook up with relatively low lift (as for whether we're planning a Psi4/Forte connection, I would be happy to start a conversation with @loriab or @fevangelista to see what that would look like in another thread).

To properly scope this, it would be good to get a bit more concrete information on the quantum angle you propose here - are you primarily envisioning this in the context of variational algorithms? Perhaps as state preparation for fault tolerant algorithms such as QPE?

Happy to discuss more, thanks again!

[brycewestheimer]

Hi @wavefunction91

Thanks for the quick response! Glad to hear that this has already been on the radar internally.

This feature request actually came out of a recent conversation with @billti about QDK/Chemistry's active-space support, and he suggested that I open the issue here just in case it wasn't already on your radar.

On the quantum side, the main use case I have in mind is variational. For RAS/ORMAS, the occupation constraints translate directly into rules for which excitation operators belong in a UCC-style ansatz or operator pool. That gives you a more structured, chemically motivated circuit with fewer operators and parameters, and typically a shallower circuit than working with the full CASCI space.

I also think there is a fault-tolerant angle, since a partitioned-CI wavefunction should provide a better multireference starting state for QPE than a single-determinant reference, so there is a natural state-preparation benefit there as well.

The subspace-decomposition point that I mentioned in the original issue is probably less important with existing hardware right now compared to the circuit-structure side.

As far as implementation goes, I've also written a PySCF fcisolver plugin for ORMAS/RAS, which made the integration path through the existing CASCI/PySCF flow look fairly natural, and seems to be working with qdk-chemistry (though I'm still debugging some open shell issues). It still needs some heavy performance optimization, but it's useful for resource estimation and comparisons at this point. I'd be happy to discuss it further if you think it would be useful!