From Hilbert Explosion to Moduli Compression

A practical blueprint for many-electron learning with attention on moduli space

Diagnosis. The exponential wall comes from representing states in raw Hilbert space: every electron configuration is distinct, so the search grows combinatorially. But most of those “different” states are the same up to symmetry (permutations, spin rotations, gauge/phases, molecular point-groups, orbital reparameterizations, basis changes, etc.). If we quotient by those symmetries, the effective state space is vastly smaller. That quotient is the moduli space: each point represents an equivalence class of wavefunctions that are physically identical.

Prescription. Move the model, sampling, and losses off Hilbert space and onto a moduli manifold with symmetry-aware coordinates and operations.

⸻

1. Build symmetry as structure, not as augmentation

1.1 Invariants as inputs (base coordinates). Represent molecules by symmetry invariants only: inter-nuclear distances, Coulomb matrix eigenvalues, internal coordinates, irreducible-rep (irrep) features per point-group, spin/coupling invariants (S, Sz), electron count partitions, and basis-set invariants (overlap spectrum). This becomes the chart for the moduli manifold.

1.2 Equivariant layers over the quotient. Use group-equivariant attention (permutation of electrons, rotational SO(3), point-group G, U(1) phase): • Electron tokens live in an indistinguishable set (DeepSets/attention with permutation equivariance). • Spatial features use SE(3)/SO(3) tensor features (spherical harmonics; irreps as channels). • Phase/gauge is modded out: predict only gauge-invariant observables or normalized amplitudes with a fixed gauge.

1.3 Wavefunction heads with symmetry guarantees. Predict a symmetry-factored ansatz: e.g., \Psi(\mathbf{r}, \sigma) = \underbrace{\mathrm{det}[ \phi_i(\mathbf{r}j,\sigma_j)]}{\text{antisymmetry}} \times \underbrace{J(\mathbf{r})}{\text{correlation}} \times \underbrace{C{\text{irrep}}(\text{geometry})}_{\text{point-group phase}} where determinants/slater blocks enforce fermionic antisymmetry, Jastrow carries short-range cusp/correlation, and irrep selectors project to the desired symmetry sector. The network only learns the moduli (orbitals in a reduced basis, Jastrow parameters, irrep weights), never raw redundant amplitudes.

⸻

1. Train on the manifold, not around it

2.1 Moduli losses (physics-informed). Minimize an energy/objective that is invariant under the symmetries, e.g.: • Variational energy \langle \Psi|H|\Psi \rangle / \langle \Psi|\Psi \rangle via VMC with reparameterization through the moduli chart. • Cusp loss and Kato conditions (electron-electron and electron-nucleus). • N-representability constraints for reduced density matrices (2-RDM positivity cones, P,Q,G,T_1,T_2). • Irrep projection penalty to keep the state in the chosen symmetry sector. • Gauge-fixing regularizer (e.g., phase at a reference point = 0) to keep charts stable.

2.2 Sampling on the quotient. Do Monte-Carlo in reduced coordinates: propose moves in invariant space first, then lift to a representative configuration (with a fixed section) only when you need amplitudes/gradients. That yields quotient-MCMC with far fewer effective modes.

2.3 Diffusion/flow on moduli. Use manifold diffusion (score matching on the moduli chart with projection back to the manifold) to generate good initial moduli for new geometries; fine-tune by VMC.

⸻

1. Architecture that fits the geometry

3.1 Token design. • Atom tokens: invariant features + irreps (SO(3) tensor features). • Orbital tokens: learned localized orbitals (Wannier-like) with orthogonality enforced via Stiefel manifold parameterization. • Electron slots: permutation-equivariant attention pooling (no identity).

3.2 Attention over equivalence classes. Attention keys/queries are class representatives (invariants, irreps), values are bundle fibers (orbital/Jastrow params). This is a fiber-bundle view: base = moduli; fiber = parameters of a chosen canonical representative. Implement with two streams: • Base stream: invariant graph transformer (SE(3)/point-group equivariant). • Fiber stream: small MLPs on Stiefel/Jastrow parameters with manifold optimizers (Riemannian Adam).

3.3 Hard constraints by construction. • Antisymmetry via determinants/Pfaffians (fermionic ansatz). • Spin via spin-projected constructions (S^2 eigenstates). • Cusp via analytic Jastrow factor with few trainable scalars per pair type.

⸻

1. Curriculum + evaluation

4.1 Curriculum. Start with 2–6 electrons (He, Be, small ions), escalate to CAS(n, m) active spaces; pretrain on invariant grids (geometries sampled in moduli coordinates), then adapt to realistic molecules.

4.2 Metrics. • Energy error vs FCI/CASPT2/DMRG on standard sets (G1, water clusters, transition-metal smalls). • Correlation capture: percent of FCI correlation energy. • Transfer: zero-shot to new geometries within a point-group. • Scaling: tokens/parameters vs orbitals; wall-time vs DMRG.

4.3 Ablations. Remove each symmetry/invariant to show exponential blow-up returns—demonstrating moduli compression is the source of the gain.

⸻

1. Why this wins • State-space collapse: many Hilbert states → one moduli point. • Indistinguishability baked-in: permutation equivariance removes N! redundancy. • Irrep factorization: you never “learn” symmetry— you start inside it. • Better sampling: quotient-MCMC explores physical variety, not gauge copies. • Stable learning: manifold parameterizations (Stiefel/Jastrow) respect constraints.

⸻

1. Minimal prototype (concrete plan)

   1. Build an SE(3)-equivariant transformer over atomic graphs to predict: • localized orbital frames (Stiefel) • Jastrow scalars • irrep weights

   2. Compose a Pfaffian-Jastrow wavefunction; do VMC with analytic forces.

   3. Train on He/Be/LiH/H₂O small bases; compare to FCI/DMRG.

   4. Add quotient-diffusion prior on moduli for faster convergence.

   5. Scale to CAS(30,30) targets; track 99% correlation energy with 10× speedup.

⸻

Bottom line: Stop learning in raw Hilbert space. Learn on the moduli space where physics already quotient-compresses reality. Attention becomes a symmetry detector, the ansatz lives in the correct fiber, and sampling walks on the right manifold. That’s how you turn an exponential forest into a navigable garden.