Continuous Gaussian AO support, contact and tail bounds¶
This continuation supplies 28 kernel-checked claims in
formal/rhf/VibeQCAICCM/AOSupport.lean, bringing the pinned project audit to
505 claims, 423 in AICCM. It supports Paper 1 (L18) and Paper 2 (C27),
following the conditional kernel contraction (L17)/(C26). It certifies
finite Gaussian expansions, including finite translated and phase-weighted
lists. It does not certify the omitted infinite Bloch image sum or the
production engine. The definitions are a basis representation, not a proof
of physical adequacy of the chosen basis or Hamiltonian.
Conventions and constructed envelope¶
Use three-dimensional Euclidean volume, complex coefficients, nonnegative integer Cartesian powers, positive real exponents and finite index sets. Coordinates below are numerical atomic-unit coordinates: x means x/ell, with ell = 1 bohr, and a means alpha ell^2. Restore ell^3 in volume integrals and ell^2 in the contact integral when using other length units. This choice is necessary for the polynomial comparison r^L <= 1 + r^(2L).
For a primitive p_j(x) = c_j product_k (x_k-A_jk)^(l_jk) exp(-a_j |x-A_j|^2), put L_j = sum_k l_jk and
C_j = |c_j| [1 + L_j!/(a_j/2)^L_j] exp((a_j/2)|A_j|^2),
0 < b <= min_j a_j/4, C = sum_j C_j.
Then |sum_j p_j(x)| <= C exp(-b |x|^2) everywhere. The coefficient c_j includes the stored primitive normalization, contraction coefficient, any Cartesian-to-pure coefficient, phase and image weight. No separate unit-norm assumption is made for each Cartesian component. Empty finite expansions are allowed with any positive b. A single nonzero Gaussian has support all of R^3; a selected finite translation list does not give compact spatial support. A periodic infinite sum need not be L1 on R^3: its appropriate cell measure and convergence require a separate theorem.
For r >= 0 and a > 0, the exponential series gives r^(2L) exp(-(a/2)r^2) <= L!/(a/2)^L. Combine this with r^L <= 1+r^(2L) and exp(-(a/2)r^2) <= 1. Each coordinate has absolute value at most r. Finally |x|^2 <= 2|x-A|^2 + 2|A|^2 moves the envelope to a common origin. The triangle inequality handles the finite contraction. Continuity is proved for each primitive and the finite sum, so its measurability is constructed. For two such orbitals, the pair density rho = conjugate(phi) psi has C_rho = C_phi C_psi and a_rho = b_phi+b_psi. This retains complex conjugation and uses absolute, not signed, coefficients.
The basis convention is consistent with McMurchie and Davidson, J. Comput. Phys. 26, 218-231 (1978), doi:10.1016/0021-9991(78)90092-X, Sec. 2, Eqs. 2.1-2.4 (Cartesian primitives, normalization and contraction). The primary paper was read locally; no SI is listed in the library catalog. Its component normalization must not be applied a second time to the stored coefficients used by the implementation. The bounds here are derived below; the citation does not assert these specific error certificates.
Mass, contact and tail derivations¶
For |rho(x)| <= C exp(-a|x|^2), C >= 0, a > 0, define
M = C (pi/a)^(3/2),
T(R) = C exp(-a R^2/2) (2 pi/a)^(3/2),
U(R) = C exp(-a R^2), R >= 0.
The Gaussian volume integral gives integrability and integral |rho| <= M. On |x| >= R, split the exponent into two halves and integrate the remaining Gaussian on all space to obtain the absolute tail mass <= T(R). The tail also has supremum <= U(R). These conservative constants can be improved by changing origin or retaining primitive-specific envelopes; they are not claims of chemical accuracy.
For delta >= 0, spherical integration in dimension three gives exactly integral_{|z|<delta} 1/|z| dz = 2 pi delta^2. For a measurable bounded complex density |w| <= C, the absolute contact contraction is at most 2 pi C delta^2. The assigned value at z=0 is immaterial for volume integration. The diagonal, and any countable set of translated contact diagonals x-y=L_i, have zero product volume. This nullity alone proves neither summability of infinitely many image contributions nor anything about a point grid with an atom at contact.
For delta > 0 and f in L1 with |f| <= C, split the Coulomb integral into the ball and its complement, where 1/|z| <= 1/delta:
|integral f(z)/|z| dz| <= 2 pi delta^2 C + (integral |f|)/delta.
Let rho_R = 1_{|x|>=R} rho. Translation preserves its absolute mass and supremum. Uniformly in the contact centre y, its Coulomb tail therefore obeys
|integral rho_R(z+y)/|z| dz| <= B(R,delta)
B(R,delta) = 2 pi delta^2 U(R) + T(R)/delta.
For continuous rho satisfying the envelope and integrable sigma, the iterated contraction is measurable, integrable in its outer variable, and
|integral conjugate(sigma(y)) [integral rho_R(z+y)/|z| dz] dy|
<= (integral |sigma|) B(R,delta).
This is the precise iterated-integral statement proved in
ao_coulomb_pair_tail. Identification with an independently defined
six-dimensional ERI, including any required Fubini change of variables,
remains an explicit matching obligation. The theorem does not silently
assume a quadrature grid integrates this expression exactly. For a finite
list of contacts or charges, apply the triangle inequality with the
absolute image/charge weights; infinite sums need summable majorants.
For eta >= 0, the Ewald erfc tail satisfies 0 <= erfc(eta r) <= 1 for r >= 0, by its defining Gaussian integral and the half-Gaussian mass. Thus the same contact bound holds for erfc(eta r)/r, given measurability of the composite integrand. For a bounded regular remainder H, its ball contribution is separately bounded by the density bound times a bound on H times the ball volume. Contact control does not supply the remainder’s bound or the far-field Ewald error envelope.
Statement correspondence¶
All names below are in VibeQCAICCM; each has its own manifest entry and
transitive-axiom audit. The displayed hypotheses above are part of the
contracts, not conclusions inferred from a few native samples.
Lean statement(s) |
Result and derivation |
|---|---|
|
Physical three-dimensional Gaussian mass and integrability for a > 0. |
|
Exponential-series domination of t^n exp(-ct), t >= 0, c > 0. |
|
Coordinate norm bounds and finite products. |
|
Polynomial absorption and common-origin shift above. |
|
Explicit primitive constant and finite absolute-coefficient sum. |
|
Actual primitive continuity, finite-sum measurability and integrable contraction. |
|
Gaussian support is all space, independent of a cutoff grid. |
|
Product of orbital envelopes, with complex conjugation. |
|
Absolute mass M, tail mass T and tail supremum U. |
|
Local Coulomb integrability, exact 2 pi delta^2 and weighted contact bound. |
|
Single and countably many translated contact diagonals are null. |
|
Near/far split, tail and arbitrary contact centre. |
|
Outer L1 mass times an inner uniform radius; constructed measurability for the continuous Gaussian tail. |
|
Product-volume error <= epsilon M_rho M_sigma for an AE uniform kernel bound and measurable weighted error. |
|
Gaussian-tail erfc in [0,1] and screened contact bound. |
Energy bridge and remaining assumptions¶
ao_pair_kernel_error supplies the actual Gaussian mass factors in
kernel_pair_uniform. For a common region with an applicable AE kernel
bound, add certified complement/contact radii and a separately certified
quadrature radius as in (L17)/(C26). Do not replace the all-support AE bound
by the earlier point-probe cutoff bound. The endpoint integrals must exist
and their difference must match the weighted error integral.
At fixed density, core, overlap, nuclear contribution and signed seam,
kernel_gamma_energy inserts ERI radii into the existing RHF direct and
exchange ledger with its half factors and per-cell divisor. Bounds at both
minimizers support the earlier minimum bracket. For chi, kernel_chi_energy
uses raw primitive integral radii, their exchange partners and the original
cell normalization. Denominator changes and positive absolute denominator
margins remain supplied. Local MP2 also needs projected preparation radii
and stability; arbitrary relaxed orbitals are not covered. These results
do not certify the whole RHF/MP2 implementation or the thermodynamic limit.
Production comparison and reproducibility¶
Production source inspected at ce6d5692ee317990b95a707cf2f2d2375f1f5665:
cpp/src/ao_eval.cpp: shell-wide stored normalization, bare Cartesian monomials and the pure transformation; a mixed Cartesian d function is not independently normalized to one.cpp/include/vibeqc/aopair_ft.hppand the periodic AO-pair Fourier bindings: pure spherical shells and Cartesian s are supported; the Cartesian d Fourier route rejects the input.python/vibeqc/aux_basis.py::_ao_scales_for_rsgdf: scales raw pair Fourier outputs to the AO convention. Missing this scale is detected by the check.python/vibeqc/periodic_gapw_j.py::_ao_values_on_grid: finite image evaluation. Its selected image cube does not certify omitted images.cpp/src/ewald.cpp::ewald_point_charge_potential: the direct term is skipped for r < 1e-14 bohr. This is a positive-radius exclusion, not only a null point. Its omitted screened contact has the 2 pi C delta_0^2 integrated bound for each bounded density/contact. This does not change the mathematical Madelung limit or certify a sampled grid at contact.
The native capture ran through vq on 2026-11-11 UTC using qualified
vibeqc 0.18.11, source 5c89075026d59eb6b8235554594d9ae4f1831f2a,
runtime generation 45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49,
native SHA-256 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5.
This is an identified existing runtime, not a build of the newly inspected
source. No production numerical code changed in this tranche.
The fixture uses a shifted centre (0.31,-0.27,0.18) bohr, a signed contracted
d shell with exponents (0.44,1.1) and a diffuse s shell with exponent 0.36,
asymmetric near and far probes, and three reciprocal vectors. The captured
stored coefficients are the reference operands. An independent 112-bit
python-flint==0.8.0 reference evaluates Cartesian polynomials, separable
Gaussian overlap moments and analytic complex Fourier moments. It does
not call native integration routines for its reference values.
Native comparison |
Maximum absolute residual |
Tolerance |
|---|---|---|
Cartesian AO values |
1.463e-17 |
2e-11 |
Pure AO values |
5.563e-17 |
2e-11 |
Overlap matrix |
9.231e-16 |
2e-11 |
Scaled complex pair Fourier coefficients |
2.166e-15 |
2e-11 |
Values use the captured atomic-unit basis convention (AO values in bohr^(-3/2); overlap and pair Fourier integrals dimensionless). Wrong Cartesian normalization changes the fixture by 0.00818; omitting the Fourier scale changes it by 0.439. Controls also reject zero signed pair mass in place of absolute mass, compact support for a Gaussian, and a contact delta^3 law. The unsupported Cartesian d Fourier error is retained and checked; it is not silently recast as a successful native route.
For the selected Cartesian xy/s pair the conservative Gaussian constants are C <= 15.820053738689019, a = 0.2, M <= 984.891. At R=12 bohr, T <= 0.001552717 and B(R,0.25) <= 0.006210867. An independent primitive-moment/cube-tail upper bound is below 3.357e-16, illustrating how loose the universal envelope can be. The per-contact guard budget is below 9.941e-27 in atomic-unit contraction units; it is not a total RHF or MP2 energy certificate. Probes below/above the native guard confirm the branch, with relative residuals below 1.025e-16 above it. Finite native agreement is tested, not a proof of native rounding, parsing or compilation. No new SCF/MP2 energy or matched ORCA/CRYSTAL comparison is claimed; analytic Gaussian moments provide the independent reference for these AO operands.
Both paper archives retain ao-support-input.json, ao-support-results.json,
the drivers and this derivation. Reproduce the proofs with the unchanged
pinned Lake/cache via python3 formal/check_proofs.py. To replay the
numerical reference (Python 3.10 or later with the optional flint version):
python3 formal/check_aiccm_ao_support.py ao-support-input.json replay.json
To recapture, submit formal/capture_aiccm_ao_support.py through vq with
the identified runtime and execute it using that runtime’s Python. Record
the new runtime, source and binary identities; retain failures as evidence.
The next obligations are certified continuous quadrature/native arithmetic,
a support-wide Ewald remainder envelope with summable image contributions,
and the remaining ERI representation and response/stability matching.