vibe-qcDocumentation

Replacing a finite fitted interaction

FitResponse.lean compares two specified fits. Neither one is assumed exact. Paper 1 (L9) and Paper 2 (C18) use interaction differences, not differences of arbitrary auxiliary eigenvectors. If two legs have a common auxiliary frame, write the product change as (u-a)v+a(v-b). If their differences have entrywise radii r and s, the summed product has radius at most sum[r (|b|+s)+|a|s]. This keeps the mixed term even when both old legs vanish. Otherwise compare the contracted tensors directly. In a convention with a conjugated second leg, apply the formula to that conjugated leg.

For real Gamma supercell AO pairs p=(i,j), q=(k,l), put J = delta(ij|kl) and X = delta(ik|jl). At the same spin-summed density, unchanged core, seam, metric and nuclear energy,

delta E(D) = sum Dp Dq (J-X/2)/(2 Nc).

Bound each term by its absolute product and the two corresponding ERI radii. Nothing in this identity licenses comparing energies at different densities as though they were the same. For a common admissible space, let x minimize the old energy E and y minimize the new F. Comparing each functional at both points gives F(y)-E(y) <= F(y)-E(x) <= F(x)-E(x). At relaxed densities the Fock change is the change of the interaction map applied to the old density plus the new map applied to the density change. A small fitted tensor difference alone does not bound orbital relaxation.

For chi, compare the same ordered excitations in matched orbital gauges on a common full mesh with an exchange-closed active set. Frames and energies may differ. For a raw projected direct integral z and the exchanged integral y, w=(|z|^2+|z-y|^2/2)/Nc^3. Given radii r,s on their changes, b=[A(z,r)+A(z-y,r+s)/2]/Nc^3 bounds the change in w, where A(z,r)=(2|z|+r)r. This follows by factoring the difference of squared magnitudes and applying the reverse triangle inequality.

For each old pair (w,d) and new pair (W,D) of weight and denominator, the exact change is (W-w)/D-w(D-d)/(d D). Sum the absolute bounds b/h+|w|e/(g h) when g and h are positive lower bounds on |d| and |D| and e bounds |D-d|. In particular h=|d|-e suffices if positive. The energy denominators need not all shift by one common amount.

One way to obtain e uses four band-specific radii at matched orbitals. For old row l=c_old* and new column c in a common AO metric S, define m=lSc, a=l(F_new-F_old)c, R=l(F_new c-e_new Sc), L=(l F_old-e_old l S)c. Direct subtraction gives (e_new-e_old)m=a-R+L. Thus (|a|+|R|+|L|)/q bounds the eigenvalue change if q is positive and q <= |m|. The two residuals matter for rounded data. An unresolved degenerate band or a zero overlap is not covered by this band matching rule; endpoint energies can instead furnish separate radii. The estimate uses the actual Fock change, including the density response.

Run formal/capture_aiccm_fit_response.py via vq on the identified qualified runtime to repeat the complete SCF and MP2 calculations. From either proof archive run:

python3 formal/check_aiccm_fit_response.py fit-response-input.json fit-response-results.json
python3 formal/check_proofs.py

The checker reconstructs both real AO Gram tensors using exact dyadics. For chi the captured raw MO integrals come from independent four-label loops on the native fit, not the production MP2 result. The replay uses rational complex arithmetic on these supplied rounded projections and the stored AO Fock matrices, metrics and rephased orbitals. Complex norms are bracketed on a decimal grid by integer square roots. Scalar output summaries are rounded for display. Comparisons to native energies are numerical tests, not a proof of native integral evaluation or a bound on the complete-basis or infinite-crystal energy.

For the unfitted finite-cell target, see the all-mode Coulomb error bound.