Finite auxiliary-image and reciprocal domains¶
VibeQCAICCM/DomainTails.lean adds 19 kernel-checked claims. It completes
quantitative tails for the modern analytic SR/LR construction at a fixed,
finite relative-AO domain. It does not certify the floating evaluator or
identify that finite AO domain with the complete Bloch density. The preceding
ScreenedPotential.lean and ThreePeriodic.lean supply the complete potential
and the separate relative-image convergence statements.
Notation and hypotheses¶
Use atomic units, direct translations A n and angular reciprocal translations B n, n in Z^3. A is nonsingular, B = 2 pi A^(-T), V = |det A| > 0, eta > 0, and K = sqrt(2 Ecut) > 0. Direct lengths have units bohr; eta, K and q have units inverse bohr. Positive lower bounds sigma_A and sigma_B satisfy |L n|^2 >= sigma_L^2 |n|^2. The inverse Frobenius norm 1/||L^(-1)||_F is sufficient; component-wise nearest-image rounding is neither used nor justified. All complex phases have modulus at most one.
Let h, f and g be continuous localized functions with |h(x)| <= E exp(-a |x-O|^2/2), |f(x)| <= C exp(-a |x-P|^2), |g(x)| <= D exp(-a |x-Q|^2), with nonnegative E,C,D and a > 0. Choose 0 < c <= min(a/12, eta^2/3). A contracted polynomial Gaussian requires an actual envelope, as constructed in AOSupport, not its exponent alone. For each retained relative AO label r, set Q_r = Q+A r. The auxiliary selector must contain every m whose centre O+A m lies within distance Raux >= 0 of the closed segment [P,Q_r]. Extra retained images are allowed. The finite relative AO domain P_AO is held fixed.
Define m(a)=(pi/a)^(3/2), C_s=exp(1)+2/sqrt(pi), C_C(a,b)=m(a)[2 pi+m(b)], and
t(c,N) = exp(-c(N+1)^2)/(1-exp(-c(2N+3)))
Z(c) = (1+2t(c,0))^3
Q(c,N) = 6t(c,N)(1+2t(c,0))^2
T(M,c,R,H,sigma) = M exp(-c R^2/2+c H^2/2) Z(c sigma^2/4).
These are explicit majorants, not the exact theta sum. The finite-domain coefficients retain their original AO normalization and physical units.
Auxiliary-image derivation¶
tail_auxiliary_integral combines the Gaussian product estimate
|f(x)g(x)| <= CD exp(-a|P-Q_r|^2/4) exp(-a|x-P|^2/2)
with shell_pair_bound. The full-space erfc three-centre integral I_mr
is absolutely integrable and satisfies
|I_mr| <= M_r exp(-c|O+A m-P|^2),
M_r = E C D C_s C_C(a/4,a/4) exp(-a|P-Q_r|^2/4).
Conjugation does not change this norm estimate. tail_phase_bound retains
the production phases exp(-i q.A m) exp(i k_ket.A r) without enlarging it.
tail_capsule_endpoint proves that the capsule contains its endpoint ball,
including its boundary. tail_capsule_excluded consequently places every
omitted image strictly outside that ball. tail_auxiliary proves summability
and error <= T(M_r,c,Raux,|O-P|,sigma_A). tail_auxiliary_finite sums these
errors over P_AO. Summability is proved, not assumed as a fitted-model axiom.
The estimate is conservative because it replaces a capsule by one endpoint
ball. It is a bound on the auxiliary images at the specified AO domain.
tail_selected_majorant is the common Gaussian selector lemma: combine
|L n+d|^2 >= R^2/2-H^2/2+sigma^2|n|^2/4 on the omitted set with the
separable lattice Gaussian mass and support_selection. The only selector
hypothesis needed is that no omitted point has distance below R.
Reciprocal derivation¶
tail_fourier_mass bounds a Fourier coefficient by its L1 mass.
tail_pair_coefficient gives, for the actual finite relative-image coefficient,
|rho_hat(p)| <= B_rho = C D m(a/2)
sum_{r in P_AO} exp(-a|A r+Q-P|^2/4).
If |h_hat(p)| <= A_h and |rho_hat(p)| <= B_rho, the damped reciprocal term is
J_n=(4 pi/V) exp(-|p|^2/(4 eta^2)) conj(h_hat(p)) rho_hat(p)/|p|^2,
p=B n+q. Its exact zero value is defined as zero. Outside K>0,
tail_lr_majorant bounds |J_n| by
M_K exp(-c_K |B n+q|^2), where M_K=4 pi A_h B_rho/(V K^2) and
c_K=1/(4 eta^2). tail_reciprocal constructs the complete sum and bounds
all omitted modes by
epsilon_LR = 4 pi A_h B_rho/(V K^2)
exp(-K^2/(8 eta^2)+|q|^2/(8 eta^2)) Z(sigma_B^2/(16 eta^2)).
The shift q cannot be dropped from this argument. The same theorem applies to a metric entry with B_rho replaced by the second auxiliary Fourier mass; the metric SR integral uses the preceding two-centre shell bound. A matrix norm bound still requires aggregation over all entries and their units.
Finite annulus with an infinite remainder¶
The simple radius bounds can be loose. tail_box_refinement proves the
sharper bound, for a finite selector S contained in the integer cube C_N,
|sum_all F - sum_S F| <=
sum_{n in C_N\S} M exp(-c|L n+d|^2)
+ M exp(c|d|^2/2) Q(c sigma_L^2/4,N).
The second term is a proved infinite remainder. The reference cube is never
assumed complete. tail_envelope_transfer allows any summable nonnegative
majorant of the omitted terms. For the auxiliary capsule, use the omitted
endpoint-ball Gaussian as this majorant, evaluate its finite annulus, and
add its cube remainder. This avoids enumerating the omitted capsules
separately for every relative AO image.
Matching the production path¶
The inspected modern path is ThreeCenterImageDomain and
inside_auxiliary_capsule in cpp/src/periodic_gdf_short_range.cpp, together
with _rsgdf_bounded_reciprocal_sphere in python/vibeqc/aux_basis.py.
The physical AO-pair radius is tested on Q+A r-P. The auxiliary search box
uses inverse-row endpoint extrema and row norms before the capsule test.
The reciprocal enumerator uses K=sqrt(2 Ecut), inverse-row bounds, and a
nonnegative outward boundary tolerance. Candidate/output limits raise an
error rather than silently truncate. tail_padded_sphere proves that extra
boundary points preserve a tail certificate based on the nominal radius;
tail_kinetic_radius proves the cutoff conversion. Exact ties are included.
Floating inclusion is a separately checked condition, not a Lean result.
Additional SR integral screening is disabled in the capture. A positive
screening budget or an optional LR shell-pair cutoff needs its own bound.
The program’s tolerance-based zero test and |p|^2 <= 1e-24 omission must
agree with the exact zero convention, or their difference must be bounded.
tail_split_error adds SR and LR tails when the same zero subtraction is
used. tail_split_zero_error adds an explicit zero discrepancy otherwise.
tail_evaluated_error adds any separately established finite arithmetic or
screening bound. tail_refinement bounds the difference between two finite
approximations by their two mathematical tail budgets.
The source convention was compared with Ye and Berkelbach (2021), DOI 10.1063/5.0046617, Eqs. 11-13 and SI S4, read in the local primary library. The bounds above are independent conservative Gaussian estimates; the paper’s screening estimates are not substituted for proved hypotheses.
Reproduction and scope¶
Run python3 formal/check_proofs.py with the pinned Lean/mathlib project.
Capture formal/capture_aiccm_domain_tails.py through vq with an identified
qualified runtime. Replay with Python 3.13 and optional python-flint 0.8.0:
python formal/check_aiccm_domain_tails.py domain-tails-input.json \
screened-potential-input.json domain-tails-results.json
The checker imports the archived cutoff and screened-potential checkers;
all three source hashes, both input hashes and the preceding result hash are
recorded. Keep screened-potential-results.json beside its input; the
checker verifies its input and independent-checker hashes before reuse. Arb at 112 bits
encloses the finite positive majorants and infinite remainders, interpreting
captured operands as exact dyadics. Domain comparisons reject unresolved
interval ties. Independent inverse-row boxes prove complete inner spheres.
Captured reciprocal vectors are compared with the ideal angular dual;
that floating discrepancy is reported, not identified with exact equality.
The preceding independently checked SR fixture is reused only after source,
cell, basis and native identity checks and an entrywise comparison. The new
finite LR contraction and auxiliary metric are evaluated independently using
Gaussian Fourier and two-centre Boys formulas. Those native residuals are floating observations, not enclosures
of the compiled evaluator. Negative controls remove an inner reciprocal
point, confuse energy and radius, and assert an unjustifiably enlarged
auxiliary radius.
No new RHF/MP2 energy calculation or external ORCA/CRYSTAL comparison is claimed. The next obligation is to aggregate these entry bounds with the relative-AO tail and finite evaluation/screening errors, certify the metric spectral gap and cutoff stability, and propagate through whitening to the existing RHF/MP2 energy bounds. Complete production certification remains open.
Executed check¶
Eight cases completed through vq on 2026-11-12 UTC using qualified version
0.18.11, native source 5c89075026d59eb6b8235554594d9ae4f1831f2a,
SHA-256 499575ab1782406d9ab85a664b74a368a7f300feb8a40c746fadcbb3266bd0d5,
artifact 45fcbab08e7f24d5f4ddbdfceb7da8cbc7dd00d9d7e3f66660983e4bbe828c49.
The skew-cell unnormalized primitive fixture uses eta 0.65, pair radius 12,
auxiliary radii 8/18 and reciprocal radii 2/4. The production enumerator
retained 7/77 vectors at Gamma and 9/73 at nonzero transfer. Independent
finite LR contractions differ by at most 7.325e-15, and metric SR/LR
formulas by at most 1.191e-13, versus tolerance 3e-11. The repeated SR
three-centre entries agree exactly with the previously checked capture.
At radii 18/4, outward-rounded maximum auxiliary and reciprocal
three-centre bounds are 0.006505 and 0.1445 in these primitive units.
They are conservative mathematical truncation bounds, not energy targets.
All finite refinements passed comparison with their summed budgets;
the comparison allowance 6e-11 is not a certified native roundoff bound.
The new control removes a nonzero interior mode. Exact ties are tested
separately on a rational sphere. A false radius 64 for an actual radius-8
capsule is rejected by an explicit omitted image inside its endpoint ball;
a finite refinement difference is not used as a lower bound on the full
error. Captured operands, per-entry bounds, infinite remainders, hashes
and controls are retained as domain-tails-input.json and
domain-tails-results.json in both paper source archives. The native and
proof source revisions are separate identities.