vibe-qcDocumentation

Finite BIPOLE R1 doubles times dressed Fme

The experimental doubles_fme_terms action in bipole_r1_doubles_fme.py returns bare_fov, direct, exchange and their increment. Each entry contains six vectors in the placed diagonal PNO frames. This is the difference between two consecutive raw R1 snapshots, doubles_fme - occupied_fock, at supplied amplitudes on the four-band (3,1,1) fixture. No occupied-Fock contribution enters the difference.

From spin orbitals to pair frames

Stanton et al., Sec. II, Eqs. (1) and (5) contract the doubles with an occupied-virtual Fock intermediate. With restricted spatial orbitals the two occupied spin choices give two direct terms and only the like-spin choice contributes exchange. The native shared_residual_singles in dlpno_ccsd.cpp uses the following chemists’ integral convention. Write C for the columns of the source pair [im], O_m for occupied column m, and D_n = C_nn t_n for the complete singles in the native diagonal frame of n. For each column e of C, form the three rows in that source frame:

b(e) = O_m^T F C_e
j(e) = +2 sum_n (O_m C_e | O_n D_n)
k(e) =   -sum_n (O_m D_n | O_n C_e)

The bare row is the physical Fock matrix element, without an extra AO overlap. In the native auxiliary representation the direct row contracts B_m[im] with w_n(n), while exchange contracts B_n[im] with w_n(m). This explains why interchanging only the two virtual legs changes the answer. The bounded periodic provider reconstructs the same integrals from one signed ovov block lease with conjugation on Bloch legs one and three. It does not require a positive auxiliary factorization.

For each row x separately the update is

R_i += (C_ii^T S C_im) [(2 T_im - T_im^T) x].

T_im is oriented for the ordered pair (i,m) before spin adaptation. Compute the bracket in [im], then transport its free output leg. The metric between distinct PNO frames follows Riplinger and Neese, Sec. II D, Eqs. (26)-(30). In a complete orthonormal virtual basis V the action uses U_im = V^T S C_im, so U_ii^T U_im is exactly the mixed metric. Singles D_n are not truncated into [im] or [ii] before their integral contraction.

Both m and n range over the occupied neighborhood of the target diagonal pair [ii]. A missing [im] skips that entire m update; a missing [nn] skips only n’s dressing. No off-diagonal [mn] is needed for dressing. The native solver evaluates R_i only for an existing diagonal [ii]. The bounded action admits the accepted unscreened spaces only: both ranges have six members, all 36 ordered source placements exist, and the seven home matrices have ranks (2,2,2,2,2,2,3). There is no orbit, spin or cell factor beyond the displayed coefficients.

Qualification and limits

The executable comparison lifts independently constructed pair frames and the supplied doubles into AO dyads. It contracts three AO covectors and projects only their free leg. A separate signed spectral decomposition feeds the actual native evaluator and differences the two snapshots in each sector, including Fock only once. The gates are absolute: 1e-10 Ha on each of twelve components for each term and the sum; 1e-12 on imaginary orbital-integral reconstruction. ORCA and CRYSTAL do not expose this native finite-operator boundary.

The physical source and amplitudes remain unchanged. A separate Hermitian perturbation with known occupied-virtual block exercises the nearly zero physical bare term. Zero doubles must remove all pieces, zero singles must leave just the bare row, and scaling singles at fixed doubles must scale only direct and exchange. A reversed oriented doubles matrix and a missing exchange term must each separate above tolerance.

One application reserves 42,991,616 bytes and 231,293,288,448 work units, including one pass over 256 domains and 6,912 normalized elements. Its numerical arrays use at most 323,584 bytes beyond the reserved domain and panel slots. This cap does not bound process RSS. Exact budgets admit; one-below budgets, changed source or frames, and inconsistent block metadata or shape refuse. The next native difference after this action is singles_integrals - doubles_fme. Complete R1, CCSD iteration and triples are outside its scope.

Finite-fixture verdict

All twelve components at the six occupied placements passed. The sum agreed with the actual native snapshot difference to 7.515255778801e-15 Ha. The maximum imaginary component in the reconstructed ovov block was 4.209912112047e-17.

Contribution

Maximum absolute AO error (Ha)

bare_fov

1.959183362049e-15

direct

1.510011733707e-14

exchange

6.888803763538e-15

increment

6.862566070964e-15

The physical bare contribution had maximum magnitude 2.344802102541e-15 Ha. On the separate Hermitian control the bare contribution changed by 1.009471335173e-02 Ha. Transposing the oriented doubles incorrectly changed the physical increment by 3.842850692279e-02 Ha; omitting exchange changed it by 2.241782911214e-02 Ha. Zero doubles, zero singles, singles scaled by -1.7, invalid blocks, stale sources and frames, and exact versus one-below byte and work limits passed their checks.

The physical calculation used 481.432 CPU seconds, 488.666 elapsed seconds, and 296.336 MiB peak RSS. Its source, frames, amplitudes and native identity match the earlier virtual-Fock fixture. There is no matched occupied-Fock receipt, so the verdict is isolated.