vibe-qcDocumentation

Unfitted finite-cell Coulomb error

Paper 1 (L10) and Paper 2 (C19) refer to the neutral periodic interaction on the finite torus. Its reciprocal basis is infinite even for one cell. Use the same lattice, AO functions, active labels, zero-mode convention and seam at both endpoints. These bounds do not estimate basis incompleteness or the thermodynamic limit.

For transfer q let p=G+q, excluding only p=0, and w=4π/(Ω|p|²). With primitive-cell normalization the direct integral of columns A and B is the sum of w ρ_A(p) conj(ρ_B(p)); the second column uses reversed orbital labels. Let H be the finite cube sum with approximate pair columns, I the change to exact columns inside that cube, and T the sum over its complement. For the actual fitted integral V the identity U−V=H−V+I+T gives the radius |H−V|+b+t, where |I|≤b and |T|≤t. The represented mismatch includes auxiliary fitting, whitening, support and any different cutoff. No alignment of auxiliary rows is required. For a true orthogonal projection in the same full Coulomb norm, expansion cancels both mixed products and Cauchy-Schwarz gives the sharper product of the two residual norms. Orthogonality on a truncated metric does not imply these cancellations in the full metric.

The kernel statements in CoulombError.lean are as follows:

  • coulomb_band_images: write the product error as (u−a)conj(v)+a conj(v−b). If the two column errors are bounded by r,s, the absolute weighted sum is bounded by Σ w[r(|b|+s)+|a|s].

  • coulomb_cube_weight bounds the shell count divided by (m−Q)² by 104; coulomb_gaussian_tail derives the geometric majorant from Gaussian decay.

  • coulomb_shell_tail, coulomb_series_radius: a complex shell ledger with |T_j|≤C r^j, 0≤r<1, is absolutely summable and its norm is at most C/(1−r). Combine this with the retained mismatch and the image error (coulomb_radius). coulomb_orthogonal_error proves the separate projection result under both mixed orthogonality hypotheses.

  • gamma_coulomb_bound: insert these entrywise radii in (L9), including the distinct exchanged indices and the factor 1/(2N). Unchanged one-body, nuclear and seam terms cancel at a common density. For global minima on the same admissible set use the two minimizing densities or a uniform bound, not a single SCF iterate.

  • chi_coulomb_bound: use the corresponding two MO radii for the direct integral and its exchange in (C18), then sum over occupied pairs. For changing frames H is evaluated in the target frame; its difference from V in the fitted frame then includes the orbital change. Denominator errors and positive lower bounds at both endpoints are required. Without a gap this comparison need not be finite.

  • coulomb_parity_pencil: matrices with rows (f,g),(g,f) and (s,t),(t,s) have generalized eigenvectors (1,p), p=±1, with eigenvalues (f+pg)/(s+pt) if the metric eigenvalue is positive.

Independent all-mode enclosure for periodic hydrogen

Run formal/capture_aiccm_coulomb.py through the qualified native runtime, or use the captured input from either proof supplement. Install python-flint 0.8.0 in a separate environment for the checker. Run python formal/check_aiccm_coulomb.py coulomb-input.json replay.json. The JSON bounds are rounded outwards. The checker does not import the engine. It uses 112-bit Arb complex balls, interpreting the captured binary64 shell parameters as exact dyadics. This is a numerical enclosure under the Gaussian product and periodic Fourier formulas, not a kernel proof of the library, shell normalization or the native code.

For a primitive pair (a,b) and ket image R the Gaussian product has exponent z=a+b, reduced exponent μ=ab/z and centre P=(a A+b(B+R))/z. Completing the square gives its Fourier coefficient c_a c_b (π/z)^(3/2) exp(−μ|A−B−R|²−|p|²/(4z)) exp(i k_ket R−i p P). Sum the coefficients over primitive pairs and all images before forming the Gram. Translating the common origin multiplies all columns by a common phase and leaves the Gram unchanged. The cell is cubic, side 6 bohr, with two contracted s sites separated by 1.5 bohr along z. The x mesh has one point for Γ, three for χ. All image and reciprocal cubes use integer indices including their boundaries; reflected y terms have multiplicity two except at zero.

For each coordinate and |m|≥1 the centre separation from image m is at least 4.5|m| bohr. Set c=μ(4.5)², U=1+2 exp(−c)/(1−exp(−c)). The full absolute image envelope B is the sum of |c_a c_b|(π/z)^(3/2) U³. For image cube R=4 the omission ε has the same prefactor times 3 U²·2 exp(−c(R+1)²)/(1−exp(−c(2R+3))). This is a union bound over three coordinate tails; the box and the true column are each bounded by B. Within the reciprocal cube, Σw(2Bε+ε²) suffices. Both B and ε can retain an additional exp(−|p|²/(4 z_max)) factor.

Outside reciprocal cube M=16, a shell of max index m contains 24m²+2 points. Write d=2π/6, Q=1/3, α=d²/(2 z_max), u=M+1−Q. Since |p|≥d(m−Q) and m/(m−Q)≤2, summing the geometric majorant yields the AO integral tail (4π/216) B² (104/d²) exp(−αu²)/(1−exp(−α(2u+1))). For each MO pair multiply the AO column radii by the sum of the absolute AO projection coefficients. The same estimates apply to every transfer, without removing nonzero p in the central reciprocal cell. These Gaussian bounds are specific to the fixture; angular momentum, nonorthogonal lattices, other gauges and screened interactions need their own envelopes.

Bridge to the implementation

The capture calls run_ccm_rhf_direct with tail completion disabled; its real factors are contracted separately for J and the permuted K, including the stored Ewald exchange scalar. For χ it calls prepare_chi_hf_reference with classic RI, physical-pair support and a 3.1 Ha reciprocal cutoff, then run_chi_mp2. The checker reconstructs the primitive-normalized direct and exchange Gram products from the returned sector factors and sums the 27 ordered excitations with weight 1/27. It also reconstructs spin components and checks the independent native total. The three-point orbital phases are changed in the replay to make a dropped conjugation observable.

For the unfitted spectrum, z-reflection interchanges the two sites, leaves each k along x fixed and commutes with the all-image Coulomb operator. Use the Hermitian reflection averages of the supplied overlap and core matrices. Each positive-metric parity block is one-dimensional. With the even band occupied at every k, the Hartree and exchange operators respect reflection. The normalized even and odd expectations of h+2J−K with the occupied Ewald shift therefore give the target canonical eigenvalues directly. Check ordering across the mesh and the excitation gaps before dividing. The check reports the one-body symmetrization, fitted eigenvalue replay and the small differences between the parity vectors and the native orbitals. This does not prove that an unrestricted SCF iteration reaches this branch. The native classic-GDF route emits its existing absolute-energy qualification warning; the controlled interaction replacement does not clear it. No molecular reference defines the infinite reciprocal sum on this particular torus.

The periodic Poisson derivation identifies this reciprocal target with the neutral real-space pairing in a fixed cell.