vibe-qcDocumentation

The signed regular potential at a lattice site

Paper 1, Eq. xiN and supplement (L13); Paper 2, supplement (C22). Use a fixed, full-rank three-dimensional Born-von-Kármán cell, direct lattice L, angular reciprocal lattice K, positive volume Ω, splitting parameter η > 0 and the mean-zero uniform-background convention. Take the limit through nonzero physical displacements x at a single source; other independent point charges must not coincide with it modulo L. This concerns the periodic supercell kernel. A transfer label of an orbital pair does not define a new point-charge Madelung constant.

Limit of the actual Ewald sums

Write r = |x|, h(r) = erfc(ηr)/r. The half-line Gaussian integral gives erfc(0) = 1. Differentiating its tail at z gives −2 exp(−z²)/√π. Thus h(r)−1/r tends to −2η/√π in every direction. Simply omitting the central image, as in an exactly coincident evaluation of the native point potential, misses this finite term.

The direct lattice is discrete. Choose d > 0 below or equal to the length of every nonzero vector. For |x| < d/2 and R ≠ 0, |R+x| ≥ |R|/2. For z > 0, majorize exp(−t²) on t ≥ z by exp(−zt) and integrate. Since exp(−z²) ≤ 1/z², erfc(z) ≤ 2/(√π z³). Consequently

h(|R+x|) ≤ 32/(√π η³ |R|⁴).

The fourth inverse powers are summable on the rank-three lattice. The noncentral sum therefore tends to its value at zero by dominated convergence, after splitting the central term from the literal all-image sum. Each reciprocal character tends to 1 and has modulus 1. The damped reciprocal weights are absolutely summable by the preceding Ewald proof, so here too the limit passes under the sum. The background is constant. The signed regular potential is

ξ = Σ(R≠0) h(|R|) + (1/Ω) Σ(k≠0) (4π/|k|²) exp(−|k|²/(4η²))
    − 2η/√π − π/(η²Ω).

In VibeQCAICCM/Madelung.lean, self_erfc_zero, self_erfc_deriv and self_central_limit derive the first limit. self_erfc_bound, self_lattice_gap, self_image_bound, self_off_images and self_split_images justify the direct sum; self_recip_limit and self_madelung_limit give the result for the literal, complex-valued kernel. Physically, K is the angular dual of L, and Ω its covolume; the limit algebra itself also holds for independently supplied discrete lattices of rank less than four. self_ae_unique proves that any two almost-everywhere equal physical representatives admitting these limits give the same regular part: the nonzero displacements at which they agree are dense. No arbitrary value at the source enters. This uniqueness lemma requires equality almost everywhere; it does not convert a finite reciprocal truncation into an exact Green kernel.

Potential, charging energy and exchange

The regular field for a single charge t and its compensating background is tξ. Integrating tξ from zero to Q at fixed geometry and gauge gives Q²ξ/2 (self_charging_energy). This charging work is not the exchange update. The exchange helper returns −ξ, not |ξ|; the earlier seam proof takes that signed coefficient and the chosen closed-shell density. In self_two_modes, retaining only one opposite reciprocal pair gives a lower bound since all terms in both sums are nonnegative. A 4 by 4 by 32 bohr cell has Ω = 512 and an axial pair |k| = π/16. At η = 1/2, the two terms after division by Ω give (4/π) exp(−(π/16)²). After both subtractions the bound still exceeds 1/8 bohr⁻¹ (self_elongated_positive). This rules out a universal negative sign; uniformly scaling that shape does not change its aspect ratio.

Numerical bridge

The capture uses ewald_point_charge_potential and ewald_point_charge_energy in cpp/src/ewald.cpp, madelung_constant_for_cell in python/vibeqc/madelung.py, and the actual k-mesh helper in python/vibeqc/periodic_k_gdf.py. Lengths are in bohr. The fixture compares three cells, two off-axis directions, four positive radii, two splits, and Q = −1.375. It also records the exact-coincidence diagnostic. For a nonzero displacement subtract 1/r from the native potential divided by Q. Extrapolate the last two radii as [4v(r/2)−v(r)]/3 and check refinement. This extrapolation tests the analytic limit; the theorem does not establish its convergence rate.

An independent reference uses no erfc image sum and no Ewald options. Convolving 1/r with the normalized heat kernel at time s gives 1/√(πs) at the centre. For the primed periodic Green kernel the off-source Laplacian is 4π/Ω, so the heat evolution adds 4πs/Ω up to noncentral Gaussian images. Subtract both terms from the smoothed reciprocal potential at the origin. Two heat times test convergence; the finite spectral and heat errors are not interval-certified. The wrong cell, wrong sign, missing charging half and omitted central finite part are distinguishable on these fixtures.

From either extracted archive, using the Python standard library, run:

python3 formal/check_aiccm_self.py self-input.json replay.json

Compare with self-results.json. The capture driver and the recorded operands identify the native files by digest. This is not a certification of roundoff, cutoff selection, the Hessian, surface terms or a thermodynamic limit.

The near-field derivation treats the quadratic term, its trace, and its direct and exchange contractions.

The analytic off-source hypothesis means a nonzero displacement modulo the lattice, not a finite distance cutoff. The native point potential skips distances below 1e-14 bohr; this is a numerical singularity guard, not a proof that every nonzero displacement is evaluated. The qualified probes are 0.01 to 0.08 bohr, well outside that guard. See the reference-contract audit for the signed seam, metric admission and stationarity distinctions.