vibe-qcDocumentation

Guarded Ewald ERI representation

This continuation closes the product-integral matching obligation left by the support-wide bounds. It does not identify an arbitrary fitted tensor with an unfitted ERI, or verify native arithmetic. The physical Ewald split is Ewald (1921), DOI 10.1002/andp.19213690304, Section III.2, equations 20-24. The new interchange argument below is a Lebesgue-integrability derivation, checked in the pinned Lean project.

Contract and notation

Let f and g be continuous complex pair densities on three-dimensional Euclidean space, with |f(x)| <= C exp(-a|x|²) and |g(x)| <= D exp(-a|x|²), C,D >= 0 and a > 0. A common smaller positive exponent can be used when the two constructed envelopes differ. Finite contracted positive-exponent Gaussian AO pairs satisfy this contract; an infinite Bloch or lattice expansion is not supplied by this theorem. The stored normalization and finite contraction coefficients must be included in C,D. No unit pair mass is assumed.

The direct labels are all n in Z³, with L(n)=An for a physical nonsingular cell and |L(n)|² >= sigma² |n|², sigma > 0. The core integrability results also hold for any map with that lower bound. Let eta > 0, V=|det A|>0, 0 <= delta <= R, and let S be the complete finite set of nonzero angular reciprocal vectors k=2 pi A^-T m with |k| <= K. Integrability holds for an arbitrary finite S; completeness and exact duality are additional requirements for the previous omitted-image error bound.

Write q(z)=erfc(eta|z|)/|z| and H_t(x)=sum_n 1_{|x-L(n)|<t} q(x-L(n)). At exact contact the Lean function uses the totalized inverse (value zero). The countable contact set has product measure zero; no finite physical source value is asserted. The strict guard excludes |z|<delta and the outer sphere retains |z|=R.

Define J_t as sum_n integral dy conjugate(g(y)) integral_{|z|<t} dz f(z+L(n)+y) q(z), and F_f(k)=integral f(x) exp(i k.x) dx. Then the actual guarded spherical Ewald kernel has the absolutely integrable ERI

integral dx dy f(x) conjugate(g(y)) K_(R,S,delta)(x-y)
  = J_R - J_delta
    + sum_(k in S) [4 pi exp(-|k|²/(4 eta²))/(V |k|²)]
        F_f(k) conjugate(F_g(k))
    - pi/(eta² V) F_f(0) conjugate(F_g(0)).

This is an ERI, with no particle, supercell or pair-multiplicity divisor. Those factors belong in the existing RHF/MP2 energy ledgers. For normalized AO pairs, f,g have units bohr^-3, their absolute masses are dimensionless, and the ERI and kernel-error radii have units bohr^-1 in atomic units. General Gaussian auxiliary functions in the synthetic checks have their stored basis normalization instead of unit pair mass.

Why the interchanges are valid

For |z|<t, the earlier contact estimate and the squared triangle inequality imply

|conjugate(g(y)) f(z+l+y)|
  <= C D exp(a t²) exp(-a |l|²/4) exp(-a |y|²/2).

Multiplication by q(z) on the ball is dominated by 1/|z| there. Its integral is exactly 2 pi t². The y factor integrates to M(a/2)=(pi/(a/2))^(3/2). Thus the absolute six-dimensional integral for image n is at most

C D exp(a t²) M(a/2) 2 pi t² exp(-a sigma² |n|²/4).

The final Gaussian lattice series is summable. This proves absolute integrability of every image term and summability of their absolute integrals, not merely convergence of the signed image integrals. Tonelli supplies a finite nonnegative integral of the absolute series. Bochner sum/integral interchange and Fubini therefore apply. Translation x=z+L(n)+y preserves volume. Applying the same argument to the outer integrals identifies J_t with both the product-space integral against H_t and the summed-inner-then-outer expression of the preceding tranche.

Every translated cutoff sphere has measure zero; their countable union also has zero product measure. Consequently subtraction of the two open balls gives the closed-outer/open-inner annulus ERI, including exact outer cutoff ties. This is an integral statement. A point evaluator must still retain an exact outer tie. Different tolerance-based selected domains require a separate error budget.

The plane waves have modulus one, so multiplication preserves L1 integrability. The identity exp(i k.(x-y))=exp(i k.x) conjugate(exp(i k.y)) factors the reciprocal term, including the complex conjugation. The finite reciprocal sum commutes with integration, and the constant background factors into the two pair integrals with its negative sign. It is not removed for charged transition pairs. No additional Madelung self term or exchange-seam shift is silently added to this ERI.

The previous all-image guard bound is now a bound on this actual ERI: |J_delta| <= M_g B_guard, where M_g=integral |g| and B_guard=2 pi delta² C exp(a delta²+a H²/2) G(a sigma²/4). It additionally requires the additive full lattice and a representative ball of radius H. Together with the previous exact spherical remainder radius E, the matched continuum kernel error is bounded by E M_f M_g + B_guard M_g. Quadrature, native roundoff and fit error are separate additions before applying the Gamma and Chi energy ledgers.

Audited statements

All statements are in VibeQCAICCM/GuardedERI.lean.

Claims (prefix eri_)

Role and derivation

guard_measurable, guard_bound

Measurable open guard and screened-Coulomb domination

contact_integrable, contact_mass

Constructed contact integrability and exact 2 pi t² mass

pair_decay

Gaussian product bound displayed above

shift_measurable, image_measurable

Joint measurability before and after translation

shift_majorant, shift_integrable

Integrable tensor-product majorant and absolute image budget

image_fubini

Volume-preserving translation, absolute norms and iterated integral identity

absolute_interchange

Tonelli domination and Bochner interchange from summable absolute integrals

images_absolute

Constructs that summability from the Gaussian lattice bound

guard_representation, guard_outer

Product-space, image-summed and outer-integrated expressions agree

guard_radius

Transfers the earlier uniform all-image contact radius to the actual ERI

cutoff_spheres_null, annulus_ball, annulus_representation

Closed outer cutoff, strict guard and exact ties

plane_integrable, plane_difference, reciprocal_factor

L1 plane waves, difference phase and conjugated Fourier product

smooth_representation, ewald_representation

Finite reciprocal/background assembly and guarded Ewald ERI

The 23 new claims give 554 project claims, including 472 AICCM claims. The same transitive-axiom audit and unchanged Lean/mathlib pins apply. Definitions and physical selector hypotheses are not advertised as derived physical laws. No unfinished proof or extra axiom is introduced.

Implementation bridge and checks

cpp/src/ewald.cpp::ewald_point_charge_potential implements the strict r<1e-14 guard, closed real/reciprocal spheres and -pi/(eta² V) background. Its direct displacement uses the opposite lattice sign; integer negation permutes the full additive lattice. Centered fractional reduction is not asserted to be a triclinic nearest-image algorithm. Earlier qualified cutoff/tie and far-unwrapped tests are reused.

cpp/src/aux_eri.cpp::compute_2c_eri_lattice_sr instead uses analytic libint erfc Coulomb integrals over all electron coordinates, with a shell-centre image cutoff (pair_complete_cells, pair_in_range). The 3-centre route likewise has its own AO-image domain. These are NOT the point-distance annulus and do not apply the point-potential guard to their analytic integrals. A fitted production tensor still needs its own domain, whitening, truncation and representation contracts. No engine defect was found in the tested conventions; treating these distinct cutoffs as the same mathematical object would be a contract error.

A new vq capture used qualified 0.18.11, source 5c89075026d59eb6b8235554594d9ae4f1831f2a, generation 45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49, native SHA256 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5, on 2026-11-11 UTC. The asymmetric skew-cell fixture uses two different signed contracted s functions, three eta values and two shell cutoffs. Twenty-four matrix entries agree with independent Gaussian convolution formulas within 1.07e-14 (tolerance 2e-11). The checker uses the stored coefficients, an independently bounded integer box and the closed shell selection, not production enumeration.

For primitive Gaussians, p=ab/(a+b), q=p eta²/(p+eta²), and distance d, the independent full-space integral is (pi/a)^(3/2)(pi/b)^(3/2) [erf(sqrt(p)d)-erf(sqrt(q)d)]/d, with the continuous d=0 limit. Independent radial Simpson integration of the erfc convolution agrees within 9.33e-12 in three fixtures, including contact; refinement residuals are retained. Six positive-radius guard integrals, at 1e-14 and 0.2 bohr, lie below their analytic contact budgets. The enlarged radius is a diagnostic mathematical fixture, not a native parameter. These numerical quadratures are not certified enclosures.

Independent Ewald sums agree with four complex source potentials within 4.58e-16 and with their conjugated contraction within 4.01e-16 (tolerance 2e-12). Omitting nonzero SR images, the conjugation or the background changes the results by 3.138, 1.196 and 0.02567, respectively. There is no new SCF/MP2 energy or matched ORCA/CRYSTAL claim: analytic Gaussian and Ewald references test these kernel operands directly.

Both paper archives retain the capture, checker, input and result JSON. The replay needs only the Python standard library:

python3 formal/check_aiccm_guarded_eri.py guarded-eri-input.json replay.json
python3 formal/check_proofs.py

Remaining obligations are certified quadrature and floating-point selection/arithmetic, representation matching for the actual fitted production tensors, infinite AO/Bloch image-density construction, and relaxed response/stability. The new result proves the guarded finite-cutoff continuum representation and all of its required image/integral interchanges; previous support-wide bounds control the omitted Ewald sums.