Analytic fitting shell images and Ewald ERIs¶
This continuation concerns compute_2c_eri_lattice_sr and
compute_3c_eri_lattice_sr, used by the analytic rsgdf fitting path in
build_lpq_native. It does not certify every Gamma or chi fitting route.
The reference is Ye and Berkelbach (2021), DOI 10.1063/5.0046617,
Eqs. 11-13 and SI Section S4, Eqs. S6-S13, read together: the periodic
three-centre expression has two AO-image indices, while its reciprocal
pair density includes a relative-image sum. The statements below are
mathematical consequences of explicitly supplied Gaussian envelopes.
They do not derive that physical model from a finite native tensor.
Domain and derivation¶
Work in atomic units with bohr coordinates. Let q(z)=erfc(eta|z|)/|z|, eta>0, and let continuous complex functions f,g obey centred envelopes |f(x)|<=C exp(-a|x-P|²), |g(x)|<=D exp(-a|x-Q|²), a>0, C,D>=0. Stored primitive normalization and signed contraction coefficients belong in C,D; no unit auxiliary charge or unit pair mass is assumed. Earlier AO-envelope results construct these hypotheses for finite contracted Gaussian functions, including polynomial angular factors after spending some exponent. Contact uses the totalized zero value at z=0; this does not affect the Lebesgue integral.
Put m(a)=(pi/a)^(3/2), U(a,b)=m(a)[2 pi+m(b)] and K=e+2/sqrt(pi). The constant 2 pi is the Coulomb contact integral over the unit-bohr ball; the outside-ball estimate uses 1/r<=1 in these units. Thus U is a conservative numerical atomic-unit bound, not a dimensionally homogeneous formula for arbitrary unscaled coordinate units. Splitting at t=1 in the unit and Mills bounds gives erfc(t)<=K exp(-t²). The triangle inequality implies
|P-Q|² <= 3 (|x-P|²+|y-Q|²+|x-y|²).
For 0<c<=min(a/6,eta²/3), spend these exponents before integrating the Coulomb singularity. The result is absolute six-dimensional integrability and
integral |f(x) conjugate(g(y)) q(x-y)| dx dy
<= C D K U(a/2,a/2) exp(-c|P-Q|²).
A local contact bound plus Gaussian mass constructs the integrable majorant; integrability is not inferred from cancellation of signed terms. For a fixed auxiliary h at O, |h|<=E exp(-b|x-O|²), b>0, the AO overlap inequality gives
|f(y) g(y)| <= C D exp(-a|P-Q|²/4) exp(-(a/2)|y-P|²),
integral |h(x) conjugate(f(y)g(y)) q(x-y)| dx dy
<= E C D U(b,a/2) exp(-a|P-Q|²/4).
This second bound concerns the implemented single relative-AO-image family. It is uniform in the fixed auxiliary centre, but does not sum independent auxiliary images. Conjugations must match the intended complex density; the real native fixture is not a test of general complex AO evaluation.
Let L(n)=An, n in Z³, |L(n)|>=sigma|n| with sigma>0; take |d|<=H, H>=0 and R>=0. Suppose the image integrands have absolute integral at most M exp(-c|L(n)+d|²), M>=0, c>0. The complete closed selector is S_R={n:|L(n)+d|<=R}. Write
G(t) = [1+2 exp(-t)/(1-exp(-3t))]^3,
T_R = M exp(-c R²/2+c H²/2) G(c sigma²/4).
The displaced-square estimate bounds each omitted image by a summable lattice Gaussian, so the full series exists and
|sum_all I_n - sum_(n in S_R) I_n| <= T_R.
For two centres d=Q-P and M=C D K U(a/2,a/2); for the native three-centre family d=Q-P, M=E C D U(b,a/2), c=a/4. This is a shell-centre cutoff, not a cutoff on the electronic separation |x-y-L(n)|. The bound tends to zero with R but is deliberately conservative. It is not a practical microhartree cutoff certificate for the supplied small-radius fixtures.
Connection to the Ewald interaction¶
For I_n=integral f(x) conjugate(g(y-L(n))) q(x-y) dx dy, the substitution y=z+L(n) gives an integral of f(x) conjugate(g(z)) q(x-z-L(n)). Translation preserves both the integral and its L1 norm. The constructed Gaussian majorant sums the absolute integrals over every n. Tonelli and the existing absolute-interchange theorem therefore prove
sum_n I_n = integral f(x) conjugate(g(y)) sum_n q(x-y-L(n)) dx dy.
This is the full unguarded screened component of the periodic Ewald ERI.
Adding the finite reciprocal/background term from eri_smooth_representation
preserves the shell-truncation difference exactly. A physical zero-mean
Ewald target additionally requires V=|det A|>0, angular dual vectors
2 pi A^-T m, exclusion of m=0, the background -pi/(eta² V), and either
the complete reciprocal series or its separately controlled remainder.
A guard, when present in the point-potential routine, requires the previous
contact budget. No factor of one-half, particle number or cell number is
introduced at this raw integral stage. Existing energy ledgers supply them.
Whitening and fitted-to-unfitted integral/energy propagation still require
the metric, domain and stability hypotheses of their own results.
A floating selector T need not equal S_R, even at an exact shell tie. The extra error is at most sum_(S_R symmetric_difference T) |I_n|. Unlike a sphere in electron-coordinate integration, a shell-centre tie selects a whole nonzero integral. Symmetrization averages the two directional budgets and targets. It cannot establish that the unsymmetrized target already has the desired reciprocity. The total error ledger is
|Ewald_target - native| <= model_discrepancy + shell_tail
+ selector_discrepancy + evaluation_error.
The first term is zero only after a separate representation identity.
Audited theorem mapping¶
All 21 claims are in VibeQCAICCM/ShellImages.lean, listed in claims.json
and Audit.lean. The cumulative audit is 575 claims, 493 AICCM, with the
unchanged transitive axiom allowlist and pinned Lean/mathlib dependencies.
Claims |
Property |
|---|---|
|
Screened Coulomb envelopes |
|
Full-space integrable singular majorants |
|
Two-centre absolute integral and shell decay |
|
Actual single-index three-centre family |
|
All omitted labels and absolute summability |
|
Change of variables, ERI identity and cutoff error |
|
Floating selection, directional averaging and target/evaluation separation |
hi and hb in the ERI theorem are supplied by shell_pair_bound for
translated g, whose centre is Q+L(n); |P-Q-L(n)|=|L(n)+(Q-P)|.
The abstract statement can also be reused with a sharper proved envelope.
No custom axiom, admitted claim or compiler-trust shortcut is used.
Implementation checks and unresolved defects¶
formal/capture_aiccm_shell_images.py captures native operands.
formal/check_aiccm_shell_images.py uses independent Gaussian product and
screened convolution formulas, exact rational centre selection for the
stored binary inputs, and a separately reconstructed literal floating
selector. It does not import the engine. The inverse-cell Frobenius norm
bounds every omitted label outside its enumerating box. It checks signed
asymmetric contractions, two eta values, three cutoffs, auxiliary lattice
relabelling and exact Pythagorean ties with adjacent radii. Retained article
inputs/results and archive hashes provide replay provenance. The qualified
native runtime is 0.18.11, source 5c89075026d59eb6b8235554594d9ae4f1831f2a,
native SHA-256 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5.
The extended capture ran through vq on 2026-11-12 UTC. The source is inspected separately
from this binary identity; a passing fixture is not a binary equivalence proof.
The native two-centre loop translates its second auxiliary shell. The
three-centre loop keeps the auxiliary and first orbital shell fixed,
translates only the second orbital shell, filters on their centre distance,
and averages the two orbital directions. rsgdf_lr_3c_tensor uses the
home-cell ao_pair_fourier_transform; it does not construct the relative
AO-image sum in the cited periodic pair-density formula. Consequently,
increasing this SR shell cutoff alone does not establish matched full
periodic SR/LR three-centre targets or eta independence. An explicit
restricted model could be defined, but it must not silently inherit the
full periodic Ewald correspondence.
The checked three-centre tensor changes under auxiliary-centre lattice relabelling, and its raw two orbital directions differ. The exact selector also disagrees with translated-origin floating arithmetic near a radius-6 boundary and at the exact radius-5 ties. The literal-domain residual is 2.14e-14 (tolerance 2e-11); the near-boundary three-centre discrepancy is 4.224e-4. At the exact radius-5 tie, nine exact images become seven floating-selected images, changing the metric by 3.09344. Auxiliary relabelling changes the tensor by up to 2.41054; raw directional asymmetry reaches 0.119918. Detailed operands/results are retained with the article. These are detected contract failures, not suppressed failing evidence. The native numerical routines are unchanged in this tranche. Next work is to implement or prove the required two-index periodic three-centre domain, match its reciprocal pair transform, and provide a certified selector and roundoff budget. Independent reference values here are ordinary floating calculations, not interval certificates. No relaxed SCF/MP2 calculation, ORCA/CRYSTAL energy agreement or production-wide verification is claimed.