The real fold tensor: exchange order without symmetric factors¶
AICCM/CCM remains experimental. VibeQCAICCM/FoldFit.lean adds nine
kernel-checked claims about the fitted tensor in the order the real-Gamma
route contracts it. The first AICCM tranche states the Gram order under a
row-symmetry hypothesis. These claims need no such hypothesis. No reported
energy changes.
Why the fold order is stated¶
The first tranche (Fitting.lean, derivation) writes the
fitted integral as gramERI,
V(ij|kl) = sum_P B_P[i,j] B_P[k,l],
and assumes that every factor row is symmetric, B_P[i,j] = B_P[j,i]
(hypothesis sym). With the exchange convention K_ij = sum_kl V(ik|jl) D_kl
this gives K = sum_P B_P D B_P^T.
The production factors are the real fold stack of
periodic/ccm/neutral.py::ccm_neutral_cderi_fold. Its fold identity is
(ab|cd) = sum_q sum_P B_q[P,a,b] conj(B_q[P,d,c]),
with the second pair reversed, and the program evaluates
K = sum_P L_P D L_P (ccm_ri_k_neutral), which is the exchange map of
foldERI(ij|kl) = sum_P B_P[i,j] B_P[l,k].
The statements below are written for that order and need no symmetry of the factor rows. For symmetric rows the two tensors are the same.
Correction of 2026-11-14. This note first said that the rows of the fold
stack are not symmetric once the basis has functions beyond s shells, on the
strength of a comment in ccm_ri_k_neutral_factored. That was not measured
when it was written, and the measurement in the last section of this note
does not reproduce it: on every cell tested the rows are symmetric to
rounding and the hypothesis sym holds numerically. The first tranche
therefore describes the program on those cells. The fold statements remain
the ones that hold for any real factors.
Statements¶
Indices are real supercell AOs, Aux is the retained auxiliary set, and all
scalars are real. Units and conventions are those of the first tranche:
S and the spin-summed D dimensionless, energies in Hartree, xi the
signed self-potential, one positive cell divisor.
Claim |
Statement |
|---|---|
|
Under |
|
The exchange contraction of |
|
The exchange contraction of |
|
For symmetric |
|
For symmetric |
|
A pair kernel with |
|
The RHF kernel of |
|
Exact quadratic expansion of the electronic energy along |
|
For arbitrary real factors and symmetric |
The last line replaces fitted_per_cell_derivative for the production
tensor. It needs no symmetry of individual factor rows. The reversal
symmetry is automatic: reversing both pairs and exchanging them maps
B_P[i,j] B_P[l,k] to B_P[l,k] B_P[i,j]. On symmetric matrices this is
trace cyclicity, tr(X L D L) = tr(D L X L).
What is not covered¶
The Fock matrix built from a general stack need not be symmetric. Its symmetry requires the bra-ket and first-pair symmetries of the complete folded tensor, which come from the conjugate pairing of transfer channels, not from these claims. The program symmetrises
Fafter the build.VibeQCRHF.rhf_fock_symmetricapplies once those symmetries are supplied.Positivity is stated for the Coulomb form on symmetric densities only.
The fit-response statements (
gamma_fit_difference,gamma_fit_bound, supplement Eq. (L9)) are written for arbitrary four-index tensors and need no change.formal/check_aiccm_fit_response.pystill builds its tensors in the Gram order.formal/aiccm_reference.pycontracts in the fold order since the measurement below and no longer rejects non-symmetric factors.Whether a factor stack reproduces a continuum integral is a separate obligation, as before.
Measured row symmetry of the dense fold stack¶
Measured on 2026-11-13 with ccm_neutral_cderi_fold (dense storage) on
rocksalt LiH, a = 4.084 Angstrom, native core sha256 4dbf397e..., a source
check under the stale-core override. max|L - L^T| is the largest element
of L_P - L_P^T over the stack; the exchange column is
max|sum_P L D L - sum_P L D L^T| on a random symmetric density.
| Basis | Mesh | Fit | AOs | max|L - L^T| | Exchange difference |
|—|—|—|—:|—:|—:|
| STO-3G | (2,1,1) | classic, tail resolved | 12 | 2.0e-14 | 3.2e-15 |
| STO-3G | (2,1,1) | SR/LR | 12 | 4.6e-11 | 5.6e-13 |
| STO-3G | (2,2,1) | SR/LR | 24 | 2.0e-11 | 3.2e-13 |
| pob-DZVP-rev2 | (2,1,1) | SR/LR | 22 | 8.1e-12 | 4.7e-13 |
| pob-TZVP-rev2 | (2,1,1) | SR/LR | 26 | 6.4e-11 | 3.7e-12 |
| pob-TZVP-rev2 | (2,2,1) | classic, no tail | 52 | 8.2e-15 | 1.8e-15 |
| pob-TZVP-rev2 | (2,2,1) | classic, no tail, historical_index_ball | 52 | 2.1e-14 | 2.7e-15 |
| pob-TZVP-rev2 | (2,2,2) | classic, no tail | 104 | 5.3e-15 | 2.2e-15 |
On every cell tested, with p functions and with the polarized bases, the
rows of the dense real fold stack are symmetric to rounding on the classic
fit and to the integral screen of the builder on the SR/LR fit, and the
Gram-order and fold-order exchange agree to the same level. The statement
above that the rows are not symmetric beyond s shells is therefore not
reproduced on the current dense fold: for these factors the sym
hypothesis of the first tranche holds numerically and gramERI and
foldERI are one tensor. The 5.6e-6 quoted in
periodic/ccm/ri.py::ccm_ri_k_neutral_factored was measured on an earlier
build and was not reproduced on the cell it names. The fold statements
remain the ones that need no such hypothesis, and they are what the
program evaluates.
formal/check_aiccm.py now contains both cases. A synthetic stack with
non-symmetric rows checks the program’s J, K, factored K, energy formula
and directional derivative against an independent explicit fold-order
contraction, with the Gram order as the negative control. A complete
two-cell calculation on LiH/STO-3G beside the H2 cell, on both named fits,
compares the returned Fock and energy with the same reference and reports
max|L - L^T| (1.3e-14 classic, 4.0e-11 SR/LR). 54 residuals pass (largest
3.7e-9 against 2e-8), 19 controls separate.
Check¶
python formal/check_proofs.py with the pinned Lake: 685 claims pass with
the allowed axioms propext, Classical.choice and Quot.sound. No
quantum-chemistry calculation is involved; these are finite identities.