vibe-qcDocumentation

Support-wide Ewald remainder and omitted-image bounds

VibeQCAICCM/SupportWide.lean adds 26 claims to the pinned project. The combined audit has 531 claims, including 449 AICCM claims. Paper 1 (L19) and Paper 2 (C28) use the same periodic kernel certificate; their RHF and MP2 energy normalizations remain different. This extends the pointwise cutoff and finite Gaussian support results. It certifies mathematical series and iterated integrals, not a production floating-point evaluator.

Target, conventions and hypotheses

Use a nonsingular three-dimensional cell A, V = |det A| > 0, angular dual B = 2 pi A^(-T), positive splitting eta and positive spherical cutoffs R,K. Direct labels n and reciprocal labels m range over Z^3, each exactly once. A n is in bohr; B m and the unit-charge potential are in inverse bohr. The same zero-mean kernel and background -pi/(eta^2 V) occur on both sides. Physical reciprocal zero is excluded (equivalently its summand is set to zero). The signed seam, cell, fit conventions and normalization are held fixed when these radii are fed into the energy ledger.

The primary definition is Ewald, Annalen der Physik 369, 253-287 (1921), doi:10.1002/andp.19213690304, Sec. III.2, Eqs. (20)-(24): direct and reciprocal Gaussian representations and their splitting. The primary paper was read in the local library; no separate SI is listed there. The present conservative error constants are derived below, not attributed to Ewald as published error estimates. The earlier Ewald/Poisson tranches identify the zero-mean physical target; these bounds do not prove the physical adequacy of that model.

Let sigma, tau > 0 satisfy |A n|^2 >= sigma^2 |n|^2 and |B m|^2 >= tau^2 |m|^2. The earlier inverse-row theorem supplies safe choices 1/||A^(-1)||_F and 1/||B^(-1)||_F. Use a bounded representative y of x, x=y+A j, with |y|<=H. A constructive choice is

t_i' = t_i - floor(t_i+1/2),  x = sum_i t_i a_i,
y = sum_i t_i' a_i,          H = (sum_i |a_i|)/2.

The new proofs show |t_i’|<=1/2, the integer displacement and the radius H. This is a parallelepiped representative, not a triclinic nearest image. The global theorem receives the exact representative identity and reciprocal integer-pairing witness; a nonsingular direct/dual cell supplies those data. It does not infer them from a finite probe test.

Absolute omitted-image bounds

For c>0 define

T(c) = exp(-c)/(1-exp(-3c)),    G(c) = [1+2 T(c)]^3.

The earlier one-dimensional Gaussian mass bound, multiplied in three coordinates, proves summability and sum_n exp(-c |n|^2) <= G(c). For a direct image omitted by the complete displaced sphere, r=|A n+y|>R. Combining r^2 >= R^2 and r^2 >= sigma^2 |n|^2/2-H^2 gives

r^2 >= R^2/2 + sigma^2 |n|^2/4 - H^2/2.

The earlier Mills bound erfc(eta r)/r <= 2 exp(-eta^2 r^2)/ [sqrt(pi) eta R^2] then proves

E_d = 2 exp[-eta^2 R^2/2 + eta^2 H^2/2]
      /[sqrt(pi) eta R^2] G(eta^2 sigma^2/4).

This sums every omitted direct image, including those outside any finite reference box. No supplied unknown tail constant remains. For an omitted reciprocal vector |B m|>K, split its squared norm between K^2 and tau^2 |m|^2. The unit-modulus phase gives

E_k = 4 pi exp[-K^2/(8 eta^2)]/[V K^2] G(tau^2/(8 eta^2)),
E = E_d + E_k.

Both series are absolutely controlled; no neutrality or cancellation is used. These deliberately loose bounds decay with the cutoffs at fixed cell and eta. R=0, K=0, a singular cell or a missing positive lattice bound are excluded. Different cutoffs or representations require different constants. The result is not a differentiated-kernel certificate.

For finite selections S(x)={n:|A n+x|<=R} and T={m:|B m|<=K}, the selection lemma constructs full-series summability from the summable complement and the finite retained part, then identifies full minus retained with the omitted series. Including reciprocal zero in T is harmless because its summand is exactly zero; it does not add a physical zero-mode coefficient. The common background cancels algebraically. The direct singular summands are retained and cancel before taking the error; source values assigned by totalized division are not assertions that the physical source potential is finite. The physical comparison is off contact or almost everywhere.

Reindexing n by n+j proves direct-image periodicity. Exact duality gives exp(i B m dot A j)=1. These prove periodicity of the omitted error and extend |K_infinity(x)-K_RK(x)|<=E from |y|<=H to the entire spatial support. support_global_error applies to the explicitly defined omitted-series error, not an arbitrary supplied bounded error function. Thus a finite Gaussian AO expansion need not be truncated spatially to use this bound. Infinite AO/Bloch image expansions remain distinct: their cell-density construction and normalization are not proved by an Ewald-image bound.

Complete enumeration and exact cutoff ties

For an inverse row u_i and a displaced candidate A n+x,

|n_i + u_i dot x| <= |u_i| R,
ceil(-u_i dot x-|u_i|R) <= n_i <= floor(-u_i dot x+|u_i|R).

Cauchy-Schwarz and the norm cutoff prove this complete candidate interval. The selected sphere uses <=, including exact ties. These proofs neither certify floating-point inversion/nearbyint nor prove that a padding of a fixed number of machine epsilons covers every ill-conditioned input. A lattice-vector sphere |A n|<=R is generally incomplete for the displaced sphere; the implementation check includes a discriminating counterexample.

All-image contact-guard budget

The production point-potential route skips r<delta_0=1e-14 bohr. This leaves a positive-radius singular difference, so E alone is NOT a uniform bound on its guarded kernel. Its retained-contact contribution must be added. Assume 0<=delta<=R and eta>=0. For a continuous pair density |rho(x)|<=C exp(-a|x|^2), a>0, C>=0, and |z|<delta,

|rho(z+w)| <= C exp(a delta^2) exp(-a |w|^2/2).

Use erfc(eta |z|)<=1 and the exact integral of 1/|z| on the ball, 2 pi delta^2. Applying the Gaussian lattice mass a second time gives

J_n(y) = integral_(|z|<delta) rho(z+A n+y) erfc(eta |z|)/|z| dz,
|sum_n J_n(y)| <= B_guard,
B_guard = 2 pi delta^2 C exp(a delta^2+a H^2/2) G(a sigma^2/4).

The series of contact integrals is proved summable, and the bound extends to every unwrapped y by integer reindexing. With integrable sigma, the constructed measurable iterated contraction obeys

|integral conjugate(sigma(y)) sum_n J_n(y) dy|
  <= (integral |sigma|) B_guard.

This supplies an all-image integrated budget, not just nullity of countably many contacts. It is the summed-then-iterated integral statement in Lean. Identification with an independently defined product-space ERI, including any required integral/sum interchange, remains an explicit matching obligation. Contact balls may overlap: the image sum counts each image term, so no disjointness assumption is used. For multiple sources use absolute charge weights. Restoring analytic short-range integrals can avoid this point-potential guard, but their errors need their own certificates.

Energy composition

For the exact spherical approximation without the positive-radius guard, support_pair_error gives an integral radius E M_rho M_sigma on physical product volume, using actual absolute pair masses from (L18)/(C27). Measurability of the weighted error and endpoint/representation identities remain explicit hypotheses. Add M_sigma B_guard for the matched iterated guarded contraction and add the independent quadrature/roundoff radius.

These entrywise radii feed (L17)/(C26), keeping direct/exchange index permutations and the existing RHF/MP2 cell factors. RHF holds density, core, overlap, nuclear energy and signed seam fixed; minimum brackets need bounds at both minima. Canonical MP2 additionally requires denominator-change bounds and positive absolute margins. Local MP2 needs projected operand radii and stability. The result does not control arbitrary orbital response, prove all fitted integrals match the Ewald target, or certify the engine.

Per-statement mapping

Every listed theorem has a manifest entry and transitive-axiom audit.

Lean statement(s), namespace VibeQCAICCM

Derived property

support_gaussian_mass, support_gaussian_series

Full Gaussian lattice mass and dominated complex series.

support_displaced_cutoff

Split squared-distance bound using both cutoff and bounded representative.

support_direct_majorant, support_recip_majorant

Explicit omitted-summand Gaussian envelopes.

support_direct_tail, support_recip_tail

Summability and E_d, E_k for all omitted labels.

support_selection, support_kernel_bound

Complete closed selectors, full-minus-finite identity and E on the representative ball.

support_centered_coordinate, support_cell_radius, support_centered_decomposition

Constructive integer reduction and conservative cell radius.

support_image_translation, support_recip_translation, support_error_translation

Integer reindexing and exact dual phase periodicity.

support_recip_zero

Zero reciprocal summand is explicitly zero.

support_lift_bound, support_global_error

General representative transport and the concrete omitted Ewald series on all space.

support_contact_density, support_guard_contact

Gaussian contact-centre decay and screened ball contribution.

support_guard_images, support_guard_global, support_guard_pair

All-image summability, arbitrary centres and measurable outer contraction.

support_pair_error

Product-volume absolute-mass radius from the whole-support bound.

support_candidate_interval, support_integer_window

Complete displaced candidate window, including exact ties.

Implementation and retained checks

The inspected product source is a6c7f8209401af1a2f4a8b3043acce7404bbe1dc. cpp/src/ewald.cpp uses centered fractional representatives and for_each_pair_candidate with inverse-plane spans, then the displaced radius filter. enumerate_cells supplies the reciprocal sphere without zero. The background remains -pi/(eta^2 V). The default tolerances are heuristic selectors, not these support-wide error certificates. The point-potential comments direct AO grid users to include_short_range=false and analytic erfc nuclear integrals; this tranche does not certify that separate analytic/grid integration pipeline.

The qualified runtime is vibeqc 0.18.11, source 5c89075026d59eb6b8235554594d9ae4f1831f2a, generation 45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49, native SHA-256 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5. A new native capture ran through vq on 2026-11-11 UTC. Its 144 point comparisons use skew and doubled cells, eta=0.39/0.67, cutoff pairs (4,1), (12,4), (20,8), four asymmetric fractional probes and three lattice representatives, with the source separately unwrapped. Six further native values test an exact direct cutoff tie and its translated representative. No production code or dependency pin is changed.

The independent 112-bit python-flint==0.8.0 replay interprets captured binary64 operands as exact dyadics. It reduces them independently, enumerates larger boxes, applies the earlier certified omitted-box bounds, and compares direct and reciprocal finite sums separately. The maximum native residual is 3.392e-11 inverse bohr, tolerance 5e-10. The maximum native difference between translated representatives is 3.896e-11; real periodicity does not imply binary64 equality near contact. The cutoff-tie jump is 0.4795001221869535, and using an origin-centred lattice-vector selection misses a contribution greater than 3.521 inverse bohr.

For R=20 bohr, K=8 inverse bohr, all-support mathematical radii range from 1.466e-12 to 1.358e-8 inverse bohr across the four cell/splitting settings. Some are smaller than the observed native rounding discrepancy: they are exact-arithmetic truncation radii, NOT total native error bars. Coarse-cutoff bounds can be extremely loose (above 4e7 in one doubled-cell case). For the illustrative density envelope C=2, a=0.3 and delta=1e-14, the all-image integrated guard radius is below 2.254e-20 in atomic units. Neither illustrative bound is a new SCF/MP2 energy result. Analytic Ewald references are the appropriate independent comparison for these synthetic kernel operands; no matched ORCA/CRYSTAL electronic energy is claimed.

Both paper archives retain support-wide-input.json, support-wide-results.json, the capture, checker, reused cutoff reference and this derivation. Replay:

python3 formal/check_aiccm_support_wide.py support-wide-input.json replay.json
python3 formal/check_proofs.py

Use the unchanged pinned Lean/Lake and optional flint version for the two respective checks. Remaining obligations include numerical conditioning, certified quadrature/native roundoff, the guarded ERI sum/integral matching, AO/Bloch image-density construction and relaxed response/stability. The mathematical omitted Ewald image sums and their whole-support bound are now explicit, rather than supplied unknown tail budgets.