Experimental Gamma MP2 spin ledger¶
The closed-shell neutral fitted path in periodic/ccm/mp2.py calls
fitted_mp2.fitted_pair_energy. It uses real Gamma orbitals and returns
extensive torus energies. It does not call a chi Hamiltonian. A neutral
Bloch SCF result first passes through ccm_reference_from_bloch, which
requires the full unshifted mesh, a supported retained fit, a real folded
operator, canonical metric-orthonormal orbitals and integer occupations.
Time-reversal-closed complex eigenspaces are realified, not projected onto
their real parts. The fit, exchange seam and physical energies are retained.
Contraction and normalization¶
For real factors, the occupied pair block is
G[a,b] = sum_P B[P,i,a] B[P,j,b]. Exchange is G.T, not an occupied-only
permutation. The denominator is eps[i]+eps[j]-eps[a]-eps[b].
The unordered occupied loop counts a diagonal pair once and an off-diagonal
pair twice. OS is sum(G*G/Delta); SS is sum(G*(G-G.T)/Delta).
These are the real limits of the complex numerators, without changing signs
or the occupied/virtual partition. This is the canonical RI pair sum of
Weigend and Haeser, Section 2, equations 1 to 4,
DOI:10.1007/s002140050269.
The shared native canonical_mp2_energy_from_lov in
cpp/src/aiccm2026dev_b_kernels_bindings.cpp has a historical method name.
Its algebra can also be checked using factors from the neutral Bloch
producer. Its use by a different method is not evidence of equal Hamiltonians.
For occupied momenta ki,kj, it sums every virtual momentum ka and sets
kb=ki+kj-ka modulo the regular mesh. The exchange involution is
(ka,a,b) -> (kb,b,a): both virtual bands and both momentum labels swap.
Its four factor keys are (ki,ka), (kj,kb), (ki,kb), (kj,ka).
The complete virtual sum is closed under this involution and Delta is
unchanged by it. Real Delta must be nonzero. Production requires the
stronger Delta < -1e-12 Ha; the fitted real entry now enforces this too.
The native kernel forms both integrals from raw primitive factors and
multiplies each by 1/Nk. Its amplitude is conj(g/Delta). OS receives
abs(g)^2/Delta; SS receives Re(conj(g)*(g-g_exchange))/Delta, equivalent
to the conjugate convention in the manuscript. A final 1/Nk makes its
return values per primitive cell: raw numerator normalization is Nk^-3.
The real Gamma factors already contain their finite Fourier normalization;
fitted_pair_energy introduces no cell divisor. CCMMP2Result holds
torus-total OS, SS and total correlation. Divide those by ccm.n_cells
once for a per-cell comparison. Per-atom fields instead divide by the total
torus atom count. bloch_reference.py multiplies the per-cell SCF energy by
n_cells before constructing the real reference; that is an HF conversion,
not a second correlation divisor.
Evidence and exclusions¶
tests/test_ccm_mp2.py, tests prefixed test_gamma_spin_ledger, separates
the following checks:
Check |
Contract and limit |
|---|---|
Independent ordered scalar and four-spin sum |
Sum all ordered spin orbitals with factor 1/4 and classify by occupied spins; nonzero OS and SS; absolute tolerance 2e-12 Ha/cell |
Complex native contraction |
Nontrivial phases, asymmetric exchange, conjugation, independent band relabeling, orbital gauges and the momentum automorphism k to -k |
Deliberate wrong formulas |
Omitting conjugation or swapping only virtual bands must change the measured components |
Same finite synthetic fit |
Three translated copies of real AO factors, checked Bloch realification, occupied tiles 1, 4 and 6, short final tile; omitting or duplicating the cell divisor fails |
Same physical neutral fit |
Three-cell H2/STO-3G, 40 Ha base, zero tail, retained classic fit; native Bloch, realified Gamma and direct Gamma OS, SS and total each within 1e-9 Ha/cell |
Admission |
Reject nonreal explicit orbitals, dense factors or energies before transformation; accept an exactly real complex dtype; reject zero, positive, near-zero or nonfinite denominators/energies |
The synthetic fit tests arithmetic and the finite representation map; it is not a neutral Coulomb construction. The physical fixture consumes the same SCF cache, then supplies the folded factor to the independent direct SCF. No external program with a different periodic kernel supplies its oracle. Earlier bounded block/dense and allocation evidence belongs to issue 1063; the new guard does not change accepted real contraction arithmetic.
The demonstrated admission defect was the unconditional .real on fitted
orbitals and real casts of complex energies/dense factors. Even multiplying
one occupied orbital by 1j could discard it. The bounded fix rejects
these inputs; it does not implement arbitrary complex Gamma correlation.
Complex internal retained channels are supported through their real fold.
The legacy explicit four-centre and unrestricted routes are outside this
fix, including their separate input-admission behavior. The low-level native
binding requires callers to supply finite real energy arrays and compatible
factor frames; a passing arithmetic test does not certify arbitrary raw input.
ChiMP2.lean proves the exchange-closed identity and mesh scaling for its
stated finite labels, real nonzero denominators and supplied raw factors.
The 445-claim audit at source 4858ac6e is not a production-engine
certificate. Matching the runtime factor frames, supports, pruning,
occupations, energy convention and complete mesh to those hypotheses
remains a theorem-to-code obligation. The tests provide bounded numerical
evidence, not a proof for all meshes or fits, nor a radius-convergence,
performance or experimental-status claim.