Two recent programs learn electronic structure by learning an operator, not a number. KS-FNO learns the Kohn–Sham map from potential to density, whereas the fragment 2-RDM sheaf learns the correlated two-electron marginal. They look unrelated, one mean-field and one correlated, one global spectral and one local sheaf, yet they are the same construction at two rungs, and the engine that makes both scalable is a single principle, Kohn nearsightedness.
A narrative version of this synthesis is two-rungs-of-a-learned-electronic-structure.
One operator, two rungs
Electronic-structure learning has climbed a ladder of targets: energy, density, Hamiltonian, one- then two-particle reduced density matrix. The lesson both programs draw is to learn the map, not the end-state. Fix the operator that would otherwise be computed by an iterative solve, then evaluate it.
- Density rung, the Kohn–Sham map. KS-FNO learns
, the map from a Kohn–Sham potential to its density and non-interacting kinetic energy. This is exactly the output of an orbital diagonalization, learned instead. A single SE(3)-equivariant Fourier neural operator on real-space grids replaces the diagonalization inside the SCF, giving a quasi-linear self-consistent solve with no explicit orbitals. - Correlation rung, the 2-RDM sheaf. The two-electron reduced density matrix is the minimally sufficient object. It fixes the energy and every one- and two-body observable at any correlation strength. The sheaf construction learns it fragment-local, one shared network evaluated on overlapping neighbourhoods.
KS-FNO stops at the mean-field density, since correlation is deferred to the exchange–correlation functional, whereas the 2-RDM carries genuine two-body correlation. The move is the same, learning the solution operator, one rung higher.
Nearsightedness is the shared engine
Both are scalable for the same physical reason, Kohn nearsightedness. The density and local observables at a point depend only on the potential nearby (Kohn 1996; Prodan–Kohn 2005). Locality is what turns a global solve into something learnable and near-linear.
- In KS-FNO nearsightedness is the assumption that buys scaling. The learned map is effectively local, so cost grows quasi-linearly and one model transfers across system sizes.
- In the 2-RDM sheaf nearsightedness is the quantity being measured. The consistency residual on fragment overlaps is a per-region estimator of exactly that locality, certifying where it holds and where it fails.
The two are therefore complementary halves of one story. One relies on nearsightedness to be fast, whereas the other certifies it from the model’s own outputs. The residual is, in principle, a diagnostic for when a KS-FNO-style local operator is trustworthy.
Two implementations of locality
The constructions differ in how locality is encoded, and that is the interesting axis.
| Kohn–Sham map (KS-FNO) | 2-RDM sheaf | |
|---|---|---|
| rung | density (mean-field) | 2-RDM (correlated) |
| carrier | function on a real-space grid | operator-valued stalks on fragments |
| locality | global Fourier kernel, SE(3)-equivariant, local in effect | explicit cellular-sheaf, local by construction |
| structure operator | the solution map as a fixed point | the coboundary |
A global spectral operator and an explicit local sheaf reach the same nearsightedness from opposite sides. The bridge is spectral. The sheaf Laplacian’s low modes are the discrete analogue of the FNO’s retained Fourier band, and SE(3)-equivariance is the ambient-symmetry counterpart of the sheaf’s restriction and transport.
The operator-learning view
Read both as learning a solution operator for a fixed-point problem. The SCF is the iteration of the Kohn–Sham map, and KS-FNO learns the map so the fixed point is reached without diagonalizing. This is the operator/observable stance of the koopman-operator and state-predicate-duality. A nonlinear problem is handled by a learned linear-ish operator on the right space of fields. The 2-RDM adds a second axis. Across geometry and conformers, the marginal is transported, a parameter-time evolution on top of the spatial sheaf, in the same spirit as pairing a spatial Laplacian with a temporal operator.
The learned 2-RDM and the measured 2-RDM
There is a third corner, and it is quantum. The same 2-RDM that the sheaf learns is the object a quantum computer measures. Because the Hamiltonian is two-body,
Two things make this more than an analogy.
- Representability is paid on opposite sides. A learned or variational classical 2-RDM must fight the N-representability cone, whose membership is QMA-complete, whereas a quantum device prepares a physical state, so its measured 2-RDM is representable for free. The learned model and the quantum measurement are complementary ways to obtain a correlated marginal without a full wavefunction. One approximates the cone, whereas the other lives inside it.
- Nearsightedness bounds both. The same locality the sheaf certifies through its consistency residual is what caps the quantum measurement budget (which off-diagonal blocks can be skipped) and what makes fragment embedding on hardware tractable. So the residual is, in principle, a guide to where a fragment-based quantum computation can be trusted, with the sheaf diagnosing the quantum algorithm and not just the classical one.
What the synthesis opens
- Can the 2-RDM be given a KS-FNO-style neural operator so the correlated rung inherits quasi-linear scaling, with the sheaf supplying the locality the FNO assumes?
- Is the consistency residual a usable trust score for a learned Kohn–Sham map, flagging per region where the local operator has left its nearsighted regime?
- Does an SE(3)-equivariant Fourier operator and a cellular sheaf define the same local operator in a suitable limit, so the two programs are one construction with two discretizations?
References
- D. Khan, M. D. Hanisch, N. Argatoff, E. Xie, S. Sharma, A. Anandkumar, Learning the Kohn–Sham map with neural operators for quasi-linear scaling density functional theory (2026).
- W. Kohn, Density functional and density matrix method scaling linearly with the number of atoms, Phys. Rev. Lett. 76 (1996) 3168; E. Prodan, W. Kohn, Nearsightedness of electronic matter, PNAS 102 (2005) 11635.
- Z. Li et al., Fourier Neural Operator for Parametric PDEs, ICLR 2021.