The quadratic image field and its contractions¶
Paper 1, Eqs. nearfield and exfs and supplement (L14); Paper 2, Eq. madelung_decomp and supplement (C23). Work in physical Cartesian bohr, in a fixed full-rank three-dimensional BvK cell of volume V with the mean-zero background. Use displacements on a common local lift. Write R(x) = Re G(x) − 1/|x| for x ≠ 0, R(0) = ξ from the Madelung limit. The real part is the electrostatic potential. This is the untwisted supercell kernel, not a separate point kernel at each orbital transfer.
Regularity: written argument and boundary of the Lean proof¶
Let d be a lower bound on nonzero direct-lattice lengths and η > 0. The Gaussian-layer formula for the real Ewald term extends its noncentral images to complex displacements z by integrating exp(−u²(R₀+z)·(R₀+z)) from η to infinity with prefactor 2/√π; here the dot product is bilinear, without conjugation. For |Re z|, |Im z| < d/4 and a nonzero image R₀, the real part of the squared displacement is at least |R₀|²/2. The resulting Gaussian majorant is summable after integration in u: one can split its exponential into two equal factors, evaluate the first at η and integrate the second to obtain a bound proportional to exp(−η²|R₀|²/4)/|R₀|. At the central image, subtracting the bare potential leaves the entire function −(2/√π)∫₀^η exp(−u²z·z) du. In the reciprocal sum the Gaussian likewise dominates exp(|k| |Im z|)/|k|² on compact sets. Local uniform convergence of holomorphic functions gives a real analytic extension. In particular this argument does not differentiate a conditionally convergent bare lattice series.
Away from the source the Laplacian of erfc(ηr)/r is (4η³/√π)exp(−η²r²); the Laplacian of 1/r vanishes there. The Fourier series of the periodized Gaussian, by the Gaussian transform and unfolding a fundamental cell, is (π^(3/2)/(Vη³)) times the sum of exp(−|k|²/(4η²))exp(ik·x) over all reciprocal k. Each term with k ≠ 0 cancels the Laplacian of the damped reciprocal kernel. The surviving zero is +4π/V. This three-dimensional argument uses the full dual lattice, not the implementation’s finite mode list.
The full kernel hypotheses used below are now theorems: see the analyticity and classical Poisson proofs. In VibeQCAICCM/NearField.lean, near_kernel_even and near_regular_even reindex the actual Ewald sums by inversion. An independent unconditional estimate is near_central_quadratic: for 0 < ηr ≤ 1 the central regular term differs from −2η/√π + 2η³r²/(3√π) by at most 2η⁵r⁴/(5√π). It follows by integrating the bound |exp(−t²)−1+t²| ≤ t⁴ (near_erfc_error).
Under local analyticity, near_even_taylor and near_ewald_expansion average the multivariate Taylor polynomial at x and −x. Odd degrees cancel, and the power-series tail is O(|x|⁴) uniformly over direction. The quadratic term is half the second Fréchet derivative on (x,x). near_trace requires only C² regularity and the punctured classical equation: the Hessian is continuous, so its trace at zero is 4π/V in any orthonormal Cartesian basis. near_lattice_symmetry transports both actual image sums by an isometry preserving the two lattices bijectively. near_hessian_symmetry transports the second derivative by a linear symmetry. near_cubic_tensor uses coordinate reflections to kill mixed entries and axis permutations to equate diagonal entries; only with these symmetries does Q = 4πI/(3V).
Direct and exchange quadratic energies¶
For finite weights a,b and Cartesian components u,v, expand (u_i−u_j)(v_i−v_j) and separate its four moments (near_pair_moments). In the direct energy, both the pair-energy half and the Taylor half are present. For one neutral charge measure with dipole p, the coefficient is −pᵀQp/2 in the periodic-minus-bare direction (near_neutral_dipole). Neutrality removes the monopole terms; a nonzero dipole need not cancel. The reverse correction for cubic symmetry is +2π|p|²/(3V). None of this defines the polarization of a bulk solid without choosing a lift.
For the spin-restricted exchange convention of Paper 1, take occupied spatial orbitals orthonormal in the stated measure and W_ij = conjugate(φ_i)φ_j. The second density is W_ji. The same four moments now include ∫W_ij = δ_ij. Summing over i,j leaves the raw second moment minus the product of occupied position matrix elements; the two cross terms agree after exchanging i and j. With the negative exchange sign this gives −Tr(Q Ω_I) for the quadratic component. Contraction against symmetric Q removes any antisymmetric part of the mixed moments. near_exchange_spread is the finite quadrature version. near_integral_moments and near_exchange_integral prove the result for actual iterated integrals with integrable W, Wu, Wv and Wuv (NearMoments); there is no implicit collocation or real-orbital assumption. This use of occupied-space spread is the localized Cartesian expression, not a periodic position operator.
A local Taylor law cannot replace the kernel across the entire orbital domain without a bound on the remaining pairs. near_weighted_remainder and near_integrated_remainder give the absolute fourth-moment budget on support where |e| ≤ C|x−y|⁴. The integral version also proves absolute integrability for measurable weighted errors. A noncompact tail outside that ball requires its own estimate. Neither neutral signed moments nor the spread alone supplies it. In particular no new self-consistent RHF or MP2 accuracy bound follows merely from the measured curvatures.
Native comparison¶
Run the standard-library replay from either extracted archive:
python3 formal/check_aiccm_nearfield.py nearfield-input.json replay.json
The native capture calls ewald_point_charge_potential and ewald_point_charge_energy in cpp/src/ewald.cpp. In a skew cell, its (2,1,1) supercell and a cube, it evaluates opposite points along seven directions at four radii and two positive splittings. Remove 1/r, average opposite potentials, difference successive radii to eliminate ξ and extrapolate the curvature in r². The first six directions determine six Hessian entries; the seventh is a separate mixed-direction check. The reference differentiates a Gaussian-smoothed reciprocal Poisson potential at the origin: H_ab = −(4π/V) Σ’ k_a k_b exp(−s|k|²)/|k|² + δ_ab/(6√π s^(3/2)). The added term removes the free Gaussian potential’s Hessian. Compare two heat times; the finite numerical errors are not certified enclosures.
For neutral point clusters scaled by r, subtract their known free pair energy from the native energy, divide by r², and extrapolate. Test both vanishing and nonvanishing dipoles. For exchange, form two specified orthonormal complex orbitals on nodes 0,r,2r along a mixed direction. Use native off-site potentials and the independent heat regular part at coincidence. Compare the full exchange pair sum after removal of its constant with the orbital spread; reusing W_ij instead of W_ji changes the coefficient. This last test uses a constructed orbital probe, not a native SCF exchange matrix or an MP2 calculation. Numerical results and tolerances are in nearfield-results.json.