From Ewald kernel radii to electronic energy radii¶
VibeQCAICCM/KernelContraction.lean adds twelve audited claims. Paper 1
Eq. (L17) and Paper 2 Eq. (C26) are conditional continuum estimates, following
the point-potential cutoff bounds (L16)/(C25). They do not infer a uniform
bound from finitely many probes. The framework now audits 457 claims,
including 375 AICCM claims, with the existing pinned Lean/mathlib and
transitive-axiom audit. No production engine verification is claimed.
Integral contract and derivation¶
Let mu be the physical integration measure, w a complex weight, and
e = K_new - K_old. If w e is AE strongly measurable, |w| b integrable,
and |e| <= b almost everywhere, then
|integral w e| <= integral |w| b =: r.
This is kernel_weighted_error; domination also establishes integrability.
The derivation is the norm inequality for the Bochner integral followed by
|w e| = |w| |e| <= |w| b. To equate this integral to the difference of
two endpoint integrals, both endpoints must be integrable
(kernel_integral_difference). A formally defined integral is not itself
an integrability certificate.
For a measurable region U, kernel_region_tail gives
|integral w e| <= epsilon integral_U |w| + T,
provided the error integral is globally integrable, |e| <= epsilon AE on U,
|w| is integrable on U, and |integral_(complement U) w e| <= T.
The complement bound is a hypothesis, not a theorem about Gaussian AO tails.
Compact grids do not make Gaussian orbitals compactly supported.
On a product of S-finite measures take
w(x,y) = rho(x) conjugate(sigma(y)). Integrable pair densities and an AE
uniform kernel radius imply
|delta I| <= epsilon M_rho M_sigma, M_rho = integral |rho|.
kernel_pair_mass factors these absolute masses; kernel_pair_uniform
combines that equality with the first bound. Complex conjugation is kept
explicit. To match a chemical ERI convention, choose rho/sigma so that the
integral is exactly the entry used by that convention; the energy theorems
require that identity as a hypothesis. Net transition charge is not M_rho.
For rho = conjugate(phi) psi, the pointwise Young inequality gives
2 |rho| <= |phi|^2 + |psi|^2. Integrating proves
kernel_orbital_pair_mass; two unit-L2 orbitals therefore have M_rho <= 1.
kernel_pair_unitmass then gives an entrywise radius epsilon for two such
pair densities and epsilon >= 0. AO functions need their actual masses;
normalization is not inferred from an AO coefficient vector’s Euclidean norm.
kernel_common_singularity proves cancellation of an identical singular
summand before integration. Using it for Coulomb kernels additionally needs
a common regularization and an AE envelope on the resulting difference.
Off-diagonal equality suffices only if the singular set is null for the
chosen measure; discrete measures can put mass there. No contact limit,
self exclusion, zero-mode change or seam correction is supplied by this
algebraic lemma. For both endpoints keep the same physical zero convention,
cell, potential gauge and signed seam, or budget their changes separately.
kernel_quadrature_budget adds Q when a reported quadrature differs from
the approximate integral by at most Q. kernel_finite_budget gives the
finite analogue |sum w_j e_j| <= sum |w_j| b_j. It is a theorem about
supplied exact numbers, not a bound on arbitrary native floating-point sums.
Energy composition and normalization¶
kernel_gamma_energy composes the integral identity for every real AO ERI
with gamma_fit_bound. At fixed spin-summed density d, h, S, signed seam xi
and nuclear energy, the per-cell RHF radius is
sum_(ij,kl) |d_ij d_kl| (r_ij,kl + r_ik,jl/2) / (2 N).
This retains both the exchange permutation and closed-shell factor. The
existing gamma_fit_minima bracket applies if the bound holds at the two
minimizing admissible densities; no claim about arbitrary SCF convergence
or equality of relaxed orbitals follows.
kernel_chi_energy feeds raw primitive integral radii into
chi_fit_mp2_response (C18). With exchange involution t -> t’, raw z_t,
|delta z_t| <= r_t, denominator change |D_t-d_t| <= e_t, and positive
absolute denominator margins g_t <= |d_t|, h_t <= |D_t|, put
P(z,r) = (2 |z| + r) r,
B_t = [P(z_t,r_t) + P(z_t-z_t',r_t+r_t')/2] / N^3.
The per-cell canonical MP2 radius is
sum_t [B_t/h_t + |seamWeight(z)_t| e_t/(g_t h_t)].
The integral identity uses raw primitive normalization, before cell divisors.
Orbital response and denominator margins are external supplied bounds.
The local-MP2 stability/preparation theorems need their own projected block
radii; this canonical composition does not bypass those hypotheses.
Physical volume is in the integration measure/kernel conventions. Replacing
physical volume measure by probability Haar measure requires the matching
volume factors, not another unexplained division by the cell count.
Implementation and retained evidence¶
Source inspection at product revision 3e0b980f follows
cpp/src/ewald.cpp::ewald_point_charge_potential: linear signed-source sums,
full angular reciprocal dual, and background -pi sum(q)/(eta^2 V).
The method excludes source coincidences and uses explicit finite cutoffs.
The check uses off-source displacements. It does not assert triclinic nearest
images by fractional rounding. Electronic contractions remain those mapped
in formal_aiccm_fit_response.md; no native production code changed here.
capture_aiccm_contraction.py was executed through vq on 2026-11-10 UTC:
version 0.18.11, source 5c89075026d59eb6b8235554594d9ae4f1831f2a,
qualified generation
45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49,
native SHA-256
499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5.
This is qualified earlier-runtime evidence, not a test of a rebuilt current
main binary. The source identity and capture/checker digests are retained.
Twelve native multi-source contractions use two skew cells, two splittings,
three cutoff pairs and weights (5i, 5-5i). Separate real and imaginary calls
test the real native API against a complex linear extension. The independent
reference is the retained 112-bit interval Ewald evaluation in
cutoff-results.json, not another call to the native implementation.
Exact dyadic interval propagation gives a largest finite-native residual
below 4.73e-15 bohr^-1, tolerance 5e-10 bohr^-1 for these dimensionless weights.
The infinite-kernel comparison adds the weighted cutoff radii and retained
interval widths. Wrong conjugation is rejected in all twelve cases; all four
coarse-cutoff neutral controls have nonzero error. Analytic signed-density
and constant complex-density controls expose cancellation and mass factors.
ORCA/CRYSTAL energies are not matched references for these synthetic point
sources; no new SCF or MP2 energy calculation is reported.
Reproduce with Python 3.10 or later, from either supporting source archive:
python3 formal/check_proofs.py
python3 formal/check_aiccm_contraction.py contraction-input.json cutoff-results.json cutoff-input.json contraction-results.json
The second command reuses the retained interval certificate and verifies its
input/checker identities. Regenerate that certificate separately with
check_aiccm_cutoff.py and its pinned python-flint version as described in
the cutoff derivation. Repeat native capture through the qualified vq runtime
using the same submission pattern as earlier formal checks.
Remaining obligations: complete AE kernel envelopes for continuous pair supports (including contact treatment), explicit AO-tail radii, certified quadrature and native arithmetic, operator/denominator response under orbital relaxation, differentiated cutoff bounds and the bulk limit. This tranche proves how such bounds compose; finite samples do not supply them.