vibe-qcDocumentation

The Ewald kernel represents the neutral periodic Poisson inverse

Paper 1, Eq. GNewald and supplement (L12); Paper 2, supplement (C21). The weak interaction in the Poisson derivation can be evaluated with the real-image and damped reciprocal Ewald sums. Its Fourier coefficients do not depend on the positive splitting parameter η. The gauge matters even when there is no singularity at an observation point.

From erfc to the short-range multiplier

Work in three Euclidean spatial dimensions, with lengths in bohr, η in inverse bohr, and potential per unit charge in inverse bohr. Write p=2πw, a=|p|². Lean takes w in cycles per unit length. Define the complementary error function by its real Gaussian tail. For r>0 a change of variable gives

erfc(η r)/r = (2/√π) ∫[η,∞) exp(−u² r²) du.

The integral of each Gaussian layer over space is 2π/u³. Its spatial Fourier transform is that mass times exp(−a/(4u²)), by the three-dimensional Gaussian transform. Since the layer masses are integrable on (η,∞), Fubini applies even after multiplication by a complex character of norm one. An antiderivative of the mass times the Fourier exponential is (4π/a) exp(−a/(4u²)) for a>0, and −π/u² at a=0. Thus the physical coefficient of the short kernel is (4π/a)(1−exp(−a/(4η²))) for a>0 and π/η² at a=0. This is an actual free-space transform, not an assumption that a split reproduces the Coulomb multiplier.

In formal/rhf/VibeQCAICCM/Ewald.lean, ewald_erfc, ewald_mass_layer, ewald_layer_deriv, ewald_mass_integral, ewald_phase_norm, ewald_layer_integrable, ewald_fubini, ewald_fourier_layer and ewald_fourier_short give the change of variable, constants, integrability and transform. Values at a lattice contact are unnecessary in a weak integral. The theorem is not a prescription for a point-charge self-energy.

Periodization, twist and mean

Let C be a fundamental cell of a countable translation lattice L, of positive volume Ω. Use the image form Σ_R exp(−2πi q·R) h(x+R), with q also in cycles per length. This equals the usual sum with x−R and positive image phase by relabelling R. For any sector frequency w, (w−q)·R is an integer for every R. Its character on the images is therefore the q character. Multiplying by exp(−2πi w·x), integrating over C, and unfolding the translated cells recovers the full-space Fourier transform above. Absolute integrability, not termwise manipulation of bare 1/r, justifies this interchange. This step is ewald_phase_add, ewald_sector_character, ewald_images_erfc, and ewald_image_coeff.

With physical Fourier coefficients, the damped reciprocal kernel adds (4π/a)exp(−a/(4η²)) at nonzero momentum. On the zero it contributes nothing. The mean of the short kernel at Γ is π/(η²Ω), so subtract that spatial constant once. A test or source with nonzero total charge detects it even at off-source points. There is no mean subtraction in a sector without physical zero. In particular, the integer-zero label survives if p=B t is nonzero. The primed symbol is now 4π/a for a>0 and zero at a=0. The two factors from physical Fourier coefficients are paired with 1/Ω, exactly as in (L11)/(C20). The resulting interaction is independent of η for arbitrary L² pair envelopes under the reciprocal gap condition. No exchange seam or Madelung self-limit is derived here.

The corresponding declarations are ewald_zero_mean, ewald_recip_coeff, ewald_long_summable, ewald_infinite_coeff, ewald_split, ewald_symbol_eq, ewald_poisson_pairing and ewald_pairing_solution. In three dimensions absolute summability of the damped weights follows by bounding them by 16πη²/|p|⁴ on the shifted reciprocal lattice. The full-rank cell hypothesis gives the positive nonzero momentum gap in PeriodicPoisson. The result is for 3D periodic boundary conditions, not slab, wire, or surface-term gauges.

What the checks exercise

Use the proof archive with NumPy 2.2.6. Run:

python formal/check_aiccm_ewald.py ewald-input.json poisson-input.json poisson-results.json replay.json

The capture uses the native ewald_point_charge_potential on a skew lattice. It records absolute potentials with two dissimilar charges at three off-source points, with the short term both enabled and disabled, at two η values. The nontrivial character is projected using real and imaginary charge patterns in a three-cell BvK supercell. An independent primitive-cell reference evaluates the Fourier Poisson inverse for Gaussian-smoothed charges at heat times t and 2t and extrapolates to the off-source value. Subtracting the two smoothings as 2V(t)−V(2t) cancels the linear constant from the primed zero mode; changing the times tests convergence. These are numerical comparisons, not certified pointwise tail bounds. The long-only native option retains the background; it is not the free erf potential. Removing the Γ constant, dropping the central shifted vector, or reversing the character phase changes the comparison.

For the earlier smooth p-shell pair, radial quadrature of the erfc kernel uses 4π ∫ r erfc(ηr) sinc(|p|r) dr with sinc(x)=sin(x)/x. Combine with the damped reciprocal part and mean, then compare its contraction with independent Gauss-Hermite integration of the probe potential from the Poisson fixture. The classic fit routine primes squared momenta at 1e−12; it returns zero for both split reciprocal weights there. This is compatible with an ideal physical-zero rule only when the finite nonzero spectrum clears the threshold and the caller treats any required short-range constant consistently. No change to production code, proof of finite cutoff accuracy, or validation of all fit or energy routes is implied.

The self-subtracted limit treats the signed regular part at a supercell lattice site and distinguishes it from the charging energy.