The Orbit Processing Unit: a new class of processor for correlated problems

Plain-language explainer for doi:10.5281/zenodo.21360838 (#598)

The central image

A classical computer writes symbols onto a one-dimensional tape. An Orbit Processing Unit (OPU) writes onto a curved surface — the Grassmannian Gr(k,n), the space of all k-dimensional subspaces of an n-dimensional vector space.

The Grassmannian is not an exotic abstraction. It is the natural state space for any problem that involves choosing an active subspace: quantum chemistry’s active-space selection, scattering amplitude computation, topological phase classification, and kinetic proofreading all live naturally on Gr(k,n). The OPU reads and writes this space directly, rather than simulating it from a flat tape.

The problem with flat tapes

A molecule with 50 orbitals has a wavefunction that lives in a Hilbert space of dimension ~10¹⁴. Writing it onto a flat tape requires listing all 10¹⁴ coefficients. That is the exponential wall of classical quantum chemistry.

Current methods (CASSCF, CASPT2, MRCI) avoid this by choosing a small active space — a k-dimensional subspace of the n available orbitals. This reduces the problem to Gr(k,n), which has C(n,k) distinct “cells” (Schubert cells), each representing a qualitatively different correlation regime. For CASSCF(10,10): C(10,10) = 252 cells instead of 10¹⁰ basis functions.

What was missing: a formal name for the processor that operates on this space, a canonical alphabet (the cells), a canonical memory register, and a canonical halting condition. This paper provides all four.

The Schubert cell alphabet

The Grassmannian Gr(k,n) has a natural cell decomposition — the Schubert cells — indexed by Young diagrams with at most k rows and (n−k) columns. For Gr(2,4) (the two-orbital active space relevant to H₂), there are C(4,2) = 6 cells:

Cell NOON signature Chemistry ISA opcode
∅ (top cell) σ₀² ≈ 1 Hartree-Fock reference ORBIT (H⁰)
□ σ₀² ∈ (0.88, 1) Weak correlation TWIST (H¹)
□□ σ₀² ≈ 0.5 Fully correlated BIND (H²)

Moving between cells is the OPU head transition. The ORBIT opcode moves within a cell (orbital rotation without cell-crossing); TWIST crosses the H⁰/H¹ boundary; BIND crosses into H² territory where classical simulation fails.

The NOON spectrum as memory

The OPU’s memory register is the NOON spectrum — the natural orbital occupation numbers {σ₀², σ₁², …, σ_{k-1}²} obtained by diagonalising the one-particle density matrix. These are the singular values of the active-space coefficient matrix, squared.

Three key properties make the NOON spectrum the right memory:

  1. Basis-independent — the NOON values do not change when you rotate the orbital basis. They are a property of the physical state, not the representation.
  2. Directly observable — NOONs are measurable by photon counting, NMR, or STM. No full wavefunction reconstruction needed.
  3. Classifies the correlation regime — σ₀² ≈ 1: uncorrelated (H⁰); σ₀² ∈ (0.88, 1): weakly correlated (H¹); σ₀² < 0.88: strongly correlated (H²). This is the C/T classification of Papers 560/596.

The OPU reads the NOON spectrum at each step, uses it to determine which Schubert cell it is in, and halts when the spectrum stabilises.

The β* snap: a canonical halting condition

The OPU halts when σ₀² crosses the threshold 0.88 — the β* snap. Below this threshold, the system has entered the H² Schubert stratum, where:

  • DFT fails (the derivative discontinuity is a topological boundary, not a basis-set error — see Paper 596)
  • Classical simulation of the full active space becomes exponentially hard
  • BIND opcodes are required — quantum hardware is needed

Above the threshold, CASSCF (= an OPU emulator running on GPU via batched SVD + GEMM) is sufficient. The snap at σ₀² = 0.88 is not a numerical parameter — it is the Schubert variety boundary in Gr(k,n): the locus where two Schubert cells meet and the geodesic between them changes qualitative character.

For H₂ dissociation, this snap occurs at R* ≈ 1.7–1.8 Å — precisely where DFT errors exceed 5 kcal/mol (confirmed in Paper 596 experiment x596d).

Nature already builds OPUs

The OPU is not a proposed device. Several natural systems are already running OPU programmes:

FeMoco (nitrogenase active site): A physical Gr(3,7) OPU — the special case where the Grassmannian has G₂ symmetry (the Fano plane). N₂ fixation requires 3 BIND operations in the OPU programme. The protein scaffold enforces the geodesic from [Fe-N₂] to [Fe-2NH₃] on Gr(14, n_orb) via parallel transport. Runs at 300 K with no decoherence problem because the relevant holonomy is geometric, not quantum-phase-coherent in the fragile sense.

DNA polymerase III / ribosome: Both run the ISA’s IG Carnot cycle on the Grassmannian of substrate orbital geometries. The ribosome’s codon-anticodon check is a Gr(3,6) discrimination (three independent 2-dimensional orbital matchings). Error rate < 10⁻⁴ per residue = OPU operating above β*.

FMO light harvesting complex: A Gr(1,7) OPU with the 6-731 topology. Exciton transfer efficiency η ≈ 18% follows from the IG Carnot bound on Gr(1,7) with one broken coupling line.

The RPU family: where the OPU sits

The OPU is one member of the RPU (Resonance Processing Unit) family (Paper 205):

Processor Tape manifold Acting group β regime Distinguishing structure
QPU Full Hilbert space CP^{2^n−1} U(2^n) β = it Unitary, interference
OPU Gr(k,n) U(n) finite β Schubert cells, NOON spectrum
PPU Bruhat-Tits tree PGL(2,ℚ_p) p-adic β Ultrametric, NTT
Lie group RPU G G finite β Root vectors, Cartan algebra

All RPUs share the same ISA (ORBIT/TWIST/BIND) and the same IG Carnot thermodynamics. What differs is the manifold — which determines the alphabet, memory register, and halting condition. The Grassmannian’s distinguishing feature is the Schubert structure: it is the only natural state space that combines a canonical cell decomposition with a canonical singular-value memory.

GPU emulation: what works and what does not

Current CASSCF codes (PySCF, GAMESS, ORCA) are implicit OPU emulators. The mapping:

OPU opcode ISA meaning GPU primitive
SPLIT Schmidt decomposition batched SVD (cuSOLVER gesvd)
ORBIT Geodesic step on Gr(k,n) matrix exponential (cuBLAS Padé)
TWIST Berry phase accumulation det(U†V) phase (cuBLAS)
LABEL Halt check: σ₀² > 0.88? threshold on NOON
SPLAT CI coefficient update batched GEMM (cuBLAS)

H⁰ and H¹ operations are GPU-emulable efficiently. The hard boundary is BIND (H²): tracking a full loop in Gr(k,n) with phase coherence requires either exponential precision or quantum hardware. This is the same barrier as the T-gate in fault-tolerant quantum computing — not a coincidence, since T-gate injection is a Schubert variety crossing in the quantum circuit Grassmannian.

Turing completeness

The OPU is Turing-complete on sufficiently large Gr(k,n) by the same argument as Tao’s Navier-Stokes result (Paper 512):

  • Each Schubert cell = a tape symbol (C(n,k) symbols total)
  • Each Schubert calculus product (Littlewood-Richardson rule) = a head transition
  • The β* snap = the halt condition

Honest caveat: complexity is conserved, not eliminated. Exponential time on a classical tape becomes exponential precision in specifying the geodesic on Gr(k,n). The OPU is not a polynomial-time machine for NP-hard problems. Its advantage is specifically for problems already defined on Gr(k,n), where the classical simulation requires simulating the Grassmannian geometry from a flat tape — an unnecessary overhead.

The dictionary

Classical Turing OPU / Grassmannian tape
Tape alphabet {0,1} Schubert cells of Gr(k,n)
Head state k-plane V ∈ Gr(k,n)
Tape symbol → state transition Schubert calculus product (LR rule)
Halting state Schubert variety crossing (β* snap)
Memory register NOON spectrum {σ₀², …, σ_{k−1}²}
Computation Parallel transport on Gr(k,n) accumulating Berry phase

See also:

For the full technical treatment, see doi:10.5281/zenodo.21360838