vibe-qcDocumentation

The regular Ewald potential in the Γ and χ cells

Paper 1, nearfield and (L15); Paper 2, madelung_decomp and (C24). The potential is that of one unit point charge with a compensating uniform background, in physical Cartesian bohr. Work with the untwisted BvK cell, positive volume V and splitting parameter η > 0. Write R(x) = Re G(x) − 1/|x| off contact and R(0) = ξ. The subtraction uses one local lift; other lattice points are still singular. This result supplies both hypotheses used in the quadratic image-field derivation.

Analyticity. In Regularity.lean, reg_transform_jet constructs an actual Fréchet power series for ∫ w(t) exp(a(t)y) dμ, for real-linear real or complex a(t). Its n-th diagonal derivative at zero is ∫ w(t)(a(t)v)^n dμ. An integrable |w| exp(r‖a‖), r > 0, bounds the multilinear coefficients: multiply the n-th coefficient norm by rⁿ and use sⁿ/n! ≤ exp(s). Dominated summation then identifies the series with the original integral. reg_laplace_analytic records its analyticity alone.

For noncentral images c, lift y to (x,|x|²). In a Gaussian layer the weight is exp(−u²|c|²) and a(t)y = −u²(2 c·x + |x|²). If |c| ≥ d > 0, choose r = min(d/8,d²/4) in this lifted space. The weighted exponential is bounded by exp(−u²|c|²/2) ≤ 8|c|⁻⁴u⁻⁴. The fourth powers are summable on a discrete lattice of rank at most three and u ≥ η. For a finite family the Gaussian itself is integrable even from u = 0. reg_images_field and reg_images_analytic return to x by analytic composition; reg_off_analytic identifies the noncentral erfc sum. reg_central_eq subtracts the bare term before taking the limit, leaving −(2/√π)∫₀^η exp(−u²|x|²)du, analytic also at zero (reg_central_analytic). For the reciprocal terms the Gaussian dominates the factor exp(|k|r) even on the unit ball (reg_recip_field and reg_recip_analytic). reg_kernel_analytic identifies their sum with R on a punctured ball, and uses the Madelung limit to identify its value at zero. Thus reg_quartic gives the even Taylor expansion with a uniform O(|x|⁴) remainder. No absolute convergence of a bare Coulomb image sum is used.

The classical equation. At an off-lattice probe z, set c = l + z. Trace the second derivative of the lifted exponential over three orthonormal directions. The first derivative in the extra coordinate contributes −6u², and the second derivatives in the Cartesian directions contribute 4u⁴|c|². Multiplying by exp(−u²|c|²) gives exactly the u derivative of −2u³ exp(−u²|c|²). The upper endpoint vanishes. With the factor 2/√π, the direct trace is (4η³/√π)Σ exp(−η²|l+z|²). Taking one image and lower endpoint zero proves Δ(1/|z|)=0 off contact (reg_bare_harmonic). The reciprocal wave has Laplacian −|k|² times itself; this cancels the Coulomb denominator only for nonzero k.

RegCell states the geometric identifications explicitly. For cell columns A they are the quotient x ↦ A⁻¹x mod one, its fibers x+L, a tiling fundamental domain, the bijective dual labeling k = 2π A⁻ᵀn with its character, and the measure change dx = V dt for unit Haar mass. These are inputs about geometry; the Gaussian identity and differential equation are conclusions. reg_theta_dual descends the continuous periodized Gaussian to the quotient, unfolds each Fourier coefficient over a fundamental domain, applies the Gaussian transform, then inverts an absolutely summable Fourier series. It includes k = 0. The reciprocal Laplacian omits k = 0, so reg_kernel_poisson gives ΔR(z) = +4π/V at every z outside L. reg_source_trace gives the same trace for the Hessian of the extended R at contact by continuity. This does not assign a classical value to the Laplacian of the bare point singularity there. Adding a constant changes neither Laplacian; it does change ξ, whose convention is fixed separately.

Implementation. The three-dimensional native point-potential evaluator in cpp/src/ewald.cpp has an include_short_range switch. The driver formal/capture_aiccm_regularity.py samples the full and long-only potentials on Cartesian stencils for a unit source at the origin. Cell matrices, probe coordinates, positive splittings, cutoff radii, values and the binary digest are in regularity-input.json in the supplements. The checker independently sums the direct Gaussian charge density in lattice coordinates (increasing the index box from five to six in each direction), subtracts 1/r from the full stencil, and uses a five-point second difference in each Cartesian direction. Richardson extrapolation at 0.016 and 0.008 bohr removes the leading fourth-order error. Both short and long parts, the total, and wrong-sign, missing-sector and wrong-volume controls are checked. The native additive background constant is invisible to this derivative test; the Madelung and zero-mean checks test that convention separately.

Across two nonsymmetric probes, a skew primitive cell and its doubled supercell, and η = 0.39 and 0.67 bohr⁻¹, the maximum total-trace residual was 6.11e-11 bohr⁻³; all three comparisons were below 6.23e-11 bohr⁻³ (tolerance 3e-9). This is evidence for the finite-cutoff point potential, not a bound on cutoff or floating-point error in other cells, a proof about the compiled engine, or a new SCF or MP2 calculation. To replay:

python3 formal/check_aiccm_regularity.py regularity-input.json replay.json

Compare with regularity-results.json. The input and result identify the native module and scripts by SHA-256. A next step is to turn the analytic majorants into explicit, cell-dependent bounds for omitted images and differentiated reciprocal shells, including the floating-point stencil error.