Complete screened quasiperiodic potential and the modern SR/LR domains¶
This continuation proves 26 claims in VibeQCAICCM/ScreenedPotential.lean.
The full audit has 650 claims, including 568 AICCM claims. The pinned Lean,
mathlib and transitive axiom allowlist are unchanged. These are exact-real
results; they do not verify the production engine.
Equations and hypotheses¶
Use atomic units, a full-rank three-dimensional lattice L, a tiling primitive cell Ω, positive η (inverse length), and angular Bloch wave vector q. Let ρ be measurable, bounded by B and satisfy ρ(x+n)=exp(i q·n)ρ(x) for n in L. For an AO pair, ρ=conj(B_k f) B_l g and q=l-k. The Bloch sums here are unnormalized; the previous finite-cell normalization results still apply. The localized auxiliary h must be integrable. Define
sη(z) = erfc(η |z|)/|z|,
Qη[ρ](x) = ∫R³ sη(z) ρ(x-z) dz.
The source singleton can be assigned zero without changing an integral. The kernel has inverse-length units; its integral is π/η², with area units. Thus Q has the density’s units times area. This is the complete convolution, not a home-cell or finite-image potential.
The Gaussian-layer representation of erfc and its existing Fubini proof construct integrability of sη and give both ∫sη and ∫|sη| equal to π/η². Domination by B|sη| then proves absolute integrability at every x and |Qηρ|≤Bπ/η². Joint measurability gives measurable Q. Substituting the translation character in ρ(x+n-z) proves Q(x+n)=exp(iq·n)Q(x). The localized product conj(h)Q is dominated by |h|Bπ/η². Consequently
∫Ω conj(B_q h) Qη[ρ] = ∫R³ conj(h) Qη[ρ].
The integrable product majorant |h(x)||sη(z)|B also proves Fubini and the six-dimensional ERI representation in variables (x,z), y=x-z. No sum or integral is exchanged solely because a totalized Lean integral exists.
For Gaussian-enveloped localized f, |f(y)|≤C exp(-a|y-P|²), a>0, and a bijective integer lattice enumeration with |L n|²≥σ²|n|², σ>0, bounded representatives |x-m-P|≤H give the explicit uniform Bloch bound
|B_k f(x)| ≤ C exp(a H²/2) supportMass(a σ²/4).
screened_basis_cover constructs H=Σi|b_i| from a real lattice basis by
fractional flooring. This is only a bounded representative construction,
not a triclinic nearest-image algorithm. Measurability follows from
continuity of the localized orbital. Bounds for two Bloch sums multiply;
conjugation gives the exact pair character. These facts supply the density
hypotheses above for finite contracted Gaussians using the earlier AO
Gaussian-envelope results.
For a Fourier mode exp(2π i w·x), a=4π²|w|²=|p|², the convolution multiplier
is ewaldShortWeight η a. The complete mode identity is
Qη[exp(i p·x)] + exp(i p·x) [ewaldLong η a - ewaldMean η a]
= exp(i p·x) poissonWeight(a).
At a=0, the SR multiplier π/η² is subtracted once and LR is zero. At a>0, the sum is 4π/a. A density Fourier integral coefficient requires the separate 1/volume Fourier reconstruction factor. No inverse volume is inserted into the AO-pair coefficient itself.
Claim correspondence (L24 / C33)¶
Lean claims (all prefixed |
Result |
|---|---|
|
Source null set, absolute integrability and exact SR mass |
|
Complete potential and legitimate primitive-cell unfolding |
|
Gaussian Bloch pair supplies the density hypotheses |
|
SR multiplier and neutral SR/LR identity, including zero |
|
Product-space ERI and conditional absolutely convergent image expansion |
|
Equal closed physical pair selectors, exact ties and auxiliary bounding radius |
|
A uniform density perturbation ε gives potential error πε/η² |
screened_images requires the actual a.e. density expansion and summable
integrals of its absolute terms. ThreePeriodic.lean supplies a Gaussian
majorant for the specified two-image family and the relative-label
reindexing. This generic bridge does not certify every caller’s expansion,
array layout or finite selected domain. The native tests below check one
concrete modern implementation.
Production correspondence and exclusions¶
The relevant route is aux_basis._build_lpq_range_separated_shared_q to
compute_gdf_range_separated_integrals, not the older standalone
rsgdf_lr_3c_tensor. Inspection of periodic_gdf_short_range.cpp identifies:
With
integral_screen_error=0and no additional shell-pair pruning, both SRThreeCenterImageDomainand LRao_pair_gaussian_fourier_panelselect the physical relative AO images |A-(B+R)|≤Rpair usingpair_in_range. Equal complete candidate sets after this shared predicate follow fromscreened_pair_domains; exact boundary ties are included. Completeness and correct floating evaluation are separate from equality of predicates.SR additionally selects auxiliary centres within Raux of the AO segment. Its candidate ball about the segment midpoint has radius Raux+Rpair/2.
screened_aux_candidateproves the triangle-inequality containment, andscreened_segment_midpointsupplies the segment witness bound.screened_capsule_coordinatealso proves the inverse-row projection interval used byeach_capsule_index: endpoint extrema expanded by Raux times the row norm. These results do not prove the C++ inverse-lattice enumeration or rounded segment test.The SR phase is exp(-i q·T) exp(i kket·R); LR uses the same relative AO domain and exp(i kket·R), then the auxiliary transform is conjugated. The Γ subtraction uses π/(Vη²) conj(hhat(0)) pairhat(0), with the same image-summed overlap as LR. A molecular home-pair overlap is insufficient.
Zero transfer is tested by distance to integral fractional reciprocal coordinates <1e-12; the primed sum omits |p|²≤1e-24. Exact agreement with the theorem requires these tests to classify the exact sectors correctly and no nonzero retained momentum to enter the singular band. Near-zero sectors require an explicit discrepancy budget, not a claim of exactness.
Positive screening budgets, optional shell cutoffs, finite auxiliary images, finite reciprocal spheres, source arithmetic and whitening each introduce separate obligations. Finite auxiliary truncation does not equal the complete SR potential. The legacy standalone LR routine’s home-pair transform and molecular calibration are not certified here.
The primary source read locally is Ye and Berkelbach (2021), DOI 10.1063/5.0046617, Eqs. 11-13 and SI S4 (S6-S17). Its double-image and zero-mode structure motivates this comparison. The proofs state their own integration domains explicitly; no silent identification of a localized full-space integral with the reference’s printed cell integral is used.
Reproduction and remaining work¶
Run python3 formal/check_proofs.py with the existing pinned toolchain.
Run capture_aiccm_screened_potential.py through an identified qualified
runtime via vq, then replay check_aiccm_screened_potential.py input.json results.json. The latter uses only the standard library and independent
Gaussian-product Fourier and Boys-function integrals. It checks skew-cell,
asymmetric off-diagonal pairs, Γ and nonzero transfer, two splitting
parameters and two auxiliary radii. Deliberate controls omit auxiliary
images, reverse their phase, replace LR by home pairs, omit the zero-mode
subtraction and add an inverse-volume factor. Retain captured inputs,
identities, tolerances and residuals in each paper’s source archive.
The subsequent domain-tail results bound the actual auxiliary segment selector and reciprocal truncation at a fixed relative-AO domain. Rounding, screening and the relative-AO tail still need their own budgets. The generic majorants and mode identities are not yet a production-wide fitted-tensor or RHF/MP2 error certificate. No new SCF/MP2 calculation or matched ORCA/CRYSTAL comparison is claimed by this integral-level check; earlier qualified energy evidence remains separate.
Source inspection also covered the newer main-branch scheduling changes to the same builder. The native evidence identifies its executed source revision explicitly; it does not claim execution of those later changes.
Executed integral evidence¶
The qualified 0.18.11 runtime executed source
5c89075026d59eb6b8235554594d9ae4f1831f2a on 2026-11-12 UTC. Native
SHA-256: 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5;
runtime artifact: 45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49.
Eight cases contain 64 complex three-centre entries, with zero numerical
screening, pair radius 12 bohr, auxiliary radii 8 and 18 bohr, eta 0.45
and 0.65 inverse bohr, and Gamma/nonzero transfer. The independent finite
SR/assembled residual is 5.417889e-14 (tolerance 3e-11). At radius 18,
comparison of metric and tensor with the larger bare-Coulomb reciprocal
reference gives 2.984280e-13 (tolerance 3e-10). Deliberate errors produce
0.959276 (missing auxiliary images), 0.574041 (reversed auxiliary phase),
2.106417 (home-pair LR), 9.471091 (missing zero subtraction), and 58.757841
(extra inverse volume). The full per-case residuals, captured operands,
input/checker hashes and replay driver are retained in both source archives.
These are numerical observations, not directed-rounding error certificates.