Primitive-cell unfolding and Bloch-pair normalization¶
This tranche closes the primitive-cell correspondence for the reciprocal AO-pair coefficient. It builds on the absolute Gaussian image estimates in the complete three-centre domain. It also supplies a generic dual Bloch unfolding identity for a matching quasiperiodic potential. Constructing that potential from the complete screened interaction and matching the actual SR/LR production routes remain separate obligations. No production engine is claimed formally verified.
Domain and conventions¶
Let Lambda be a countable additive lattice in R^3, and Omega a fundamental
domain for its translation action: translates cover full space and are
pairwise disjoint almost everywhere. cell_basis_tiles constructs such a
domain from any real basis, including a triclinic basis, using fractional
coordinates in [0,1). cell_boundary_null identifies the closed
parallelepiped with this cell almost everywhere. This null-boundary result
concerns continuum integration; it does not remove a discrete cutoff tie
or justify componentwise rounding as a nearest-image algorithm.
In Cartesian coordinates cell_volume gives |det A|. Lattice vectors are
columns of A and angular reciprocal vectors are columns of 2 pi A^(-T).
Coordinates, Gaussian exponents and all stored coefficients use the prior
atomic-unit convention. Cell volume has units bohr^3; it is not an image
count or an AO normalization factor.
Write e_k(n)=exp(i k.n) and
B_k f(x) = sum_(n in Lambda) exp(-i k.n) f(x+n)
= sum_(n in Lambda) exp(+i k.n) f(x-n).
The plus-translate convention is convenient in Lean. The native sign identity explicitly reindexes n to -n to recover the positive ket phase. Gaussian hypotheses are continuity and
|f(x)| <= C exp(-a|x-P|^2), |g(x)| <= D exp(-a|x-Q|^2),
a>0, C,D>=0.
An enumeration e:Z^3 -> Lambda is bijective and satisfies |e(n)|>=sigma |n| with sigma>0. The real-basis tiling and this quantitative lower bound have different roles: the former prevents spatial double counting; the latter controls image summability. Finite positive-exponent Gaussian contractions admit the envelopes, with their actual normalization and signed coefficients absorbed in C and D. No unit charge or individual unit orbital norm is assumed.
Absolute unfolding, without an assumed interchange¶
For every full-space integrable complex F, translation invariance and the fundamental-domain partition give
sum_n integral_Omega |F(x+n)| dx = integral_R3 |F(x)| dx < infinity.
cell_absolute constructs summability of these norm integrals.
cell_unfold then uses the preceding absolutely convergent interchange to
prove integrability on Omega and
integral_Omega sum_n F(x+n) dx = integral_R3 F(x) dx.
Bloch translation is B_k f(x+n)=e_k(n) B_k f(x). If Q(x+n)=e_k(n) Q(x) and conjugate(f) Q is integrable, this yields
integral_Omega conjugate(B_k f(x)) Q(x) dx
= integral_R3 conjugate(f(x)) Q(x) dx.
cell_dual_unfold states exactly these hypotheses. In particular, applying
it to a screened potential requires proving that potential’s covariance
and the localized product’s integrability. Calling an arbitrary potential
periodic does not establish those conditions.
Reciprocal pair coefficient¶
For Q(x)=B_l g(x) exp(-i K.x), the matching character condition is
exp(i l.n) exp(-i K.n) = exp(i k.n), for every n in Lambda.
It holds when K=l-k+b and b is angular reciprocal. cell_integer_dual
proves that b.n in 2 pi Z gives the trivial character. Without this
condition the tested pair need not be primitive-periodic, and the claimed
cell-to-localized equality is not justified.
The earlier Gaussian pair-mass bound is summable over relative images.
cell_pair_seed uses it to construct integrability of conjugate(f) Q,
not to assume the desired sum/integral interchange. cell_bloch_absolute
also proves pointwise absolute convergence of the individual Bloch sums.
The resulting theorem cell_pair_formula, followed by cell_native_sign, is
integral_Omega conjugate(B_k f(x)) B_l g(x) exp(-i K.x) dx
= sum_n exp(+i l.n)
integral_R3 conjugate(f(x)) g(x-n) exp(-i K.x) dx.
This is an integral Fourier coefficient, with no extra factor 1/|Omega|. A Fourier-series coefficient defined as a cell average would include that factor by definition. Coulomb reciprocal weights have their own 1/|Omega| and must not acquire a second copy from the AO-pair coefficient. At K=0 and k=l this identity gives the cell overlap, including all relative images. It does not identify that overlap with the molecular home-cell matrix. Complete-series equality does not identify two differently screened finite domains; the earlier tail and selector-error bounds still apply.
Cell and supercell normalization¶
At allowed transfer the Fourier-tested product is primitive-periodic.
Its integral is independent of the choice of fundamental domain.
For any finite set S of distinct lattice translates, cell_supercell
constructs integrability on their union and proves
integral_(union_(n in S)(Omega+n)) F = |S| integral_Omega F.
Thus the finite-transform convention Phi_k=B_k/sqrt(Nc), Nc>0, contributes
1/Nc to an orbital pair and cancels the Nc distinct primitive cells in a
supercell integral. cell_normalization proves the scalar cancellation,
including complex conjugation. It does not assert that each Bloch orbital
has norm one in a nonorthogonal AO basis. No additional particle or pair
multiplicity is inserted here; the earlier RHF/MP2 accounting remains in
force. The reciprocal weight’s geometric volume and the finite transform’s
cell count are separate normalizations.
Lean mapping and provenance¶
VibeQCAICCM/CellUnfold.lean adds 22 audited declarations; the total is
624 claims, including 542 AICCM claims. The Lean/mathlib pins and transitive
axiom allowlist are unchanged.
Statements |
Contract |
|---|---|
|
Basis-generated cell, null boundary and Cartesian volume |
|
Constructed absolute cell decomposition and interchange |
|
Plane-wave signs and Bloch covariance |
|
Matching quasiperiodic dual and pair identities |
|
Exact allowed transfer, angular duality |
|
Gaussian convergence through the native relative-image coefficient |
|
Periodicity, domain independence and finite-cell normalization |
The primary comparison is Ye and Berkelbach (2021), DOI 10.1063/5.0046617, main Eqs. 1-2 and 11-13, and SI Section S4, Eqs. S6-S17, read from the local paper and SI. Their primitive-cell versus full-space integration domains must remain explicit. These results do not infer all three-centre conventions or approximations of that reference from the proved AO-pair identity.
Implementation bridge¶
python/vibeqc/_aopair_ft.py::ao_pair_fourier_transform_bloch and the native
ao_pair_fourier_transform_bloch_ss sum supplied relative images with the
positive ket phase and negative Fourier sign. They return an integral
coefficient without a cell-volume divisor. In contrast, standalone
python/vibeqc/aux_basis.py::rsgdf_lr_3c_tensor still uses the home-cell pair
transform, and its zero-mode calibration uses the molecular overlap.
Its explicit reciprocal weight already contains 4 pi/|Omega|. The present
proof makes the required full relative-image and overlap contracts precise;
it does not repair this producer or accept its model mismatch as zero.
Other fitting callers must be assessed individually.
The finite-transform normalization agrees with the earlier amplitude
tranche (amp_metric_block, amp_three_block, amp_factor_scatter) and
its retained asymmetric native checks. That earlier evidence is reused;
there is no new SCF/MP2 energy or ORCA/CRYSTAL comparison in this tranche.
The new capture and independent checker test native reciprocal pair values
against both analytic full-space Gaussian coefficients and direct
primitive-cell Gauss-Legendre quadrature of Bloch sums. They include Gamma,
nonzero transfer, nonzero reciprocal labels and signed s contractions on a
skew cell. Controls insert an extra volume divisor or omit bra conjugation.
Quadrature and finite image refinements are measured numerical checks,
not certified quadrature, infinite-tail or roundoff enclosures.
Replay with the Python standard library:
python3 formal/check_aiccm_cell_unfold.py cell-unfold-input.json replay.json
Remaining work: construct and unfold the full screened quasiperiodic potential for the three-centre pairing, match production SR/LR domains and zero-mode/background conventions, and certify finite selection, quadrature, roundoff and stable whitening before setting their error budgets to zero.
Executed checks¶
The qualified 0.18.11 runtime, source
5c89075026d59eb6b8235554594d9ae4f1831f2a, native SHA-256
499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5,
completed the vq capture on 2026-11-12 UTC. Sixteen complex coefficients
agree with independent analytic Gaussian formulas within 2.89e-15
(tolerance 3e-12) and primitive-cell quadrature within 7.16e-15
(tolerance 2e-9). Orders 18 and 24 differ by 5.76e-11; image cubes of
radii two and three differ by 3.13e-11. These finite refinements do not
bound an infinite remainder or floating-point error. The exact allowed
characters have a measured residual below 6.44e-16.
An extra inverse-volume factor produces an error of 1.0937, missing bra
conjugation 0.4979, and a nonreciprocal transfer 0.3363. The last control
compares the cell and localized formulas outside the theorem’s momentum
hypothesis. The cell volume is 102.816 bohr^3. Inputs, measured residuals,
tolerances and capture/checker/helper identities are retained in both
paper source archives as cell-unfold-input.json and
cell-unfold-results.json. Native and source identities are distinct.