Finite BIPOLE R1 singles times integrals¶
The experimental singles_integral_terms action
returns the signed direct term, the signed exchange term and their sum at
fixed amplitudes on the accepted four-band mesh (3,1,1). Each result has
six vectors in the placed diagonal PNO frames. The sum corresponds to the
native snapshot difference singles_integrals - doubles_fme.
For source occupied orbital n let D_n = C_nn t_n, with C_nn the full placed diagonal frame. For target orbital i let A be column a of C_ii. In chemists’ integral order the two components are
direct(i,a) = +2 sum_n (O_n D_n | O_i A)
exchange(i,a) = -sum_n (O_n O_i | A D_n)
The antisymmetrized singles-integral term of Stanton et al., Sec. II, Eq. (1) has two direct spin choices in a restricted closed shell and one like-spin exchange. In the native evaluator these are the contractions of B_ii with w_n(n), and of B_ni with u(n,[ii]). The latter projects the free virtual index only after contracting T1. The pair metric of Riplinger and Neese, Sec. II D, Eq. (30) does not justify inserting a projector from [nn] into [ii] on the input. The source and target use different physical columns even when their ranks coincide. Neither doubles nor a Fock row enters this increment.
The occupied sum uses L_ii. A missing [nn] gives no term; a missing [ii] has no output. A missing [in] alone removes nothing. The public action admits the unscreened fixture only. The executable sparse control drops [22] and [01] and removes occupied 4 from L_00.
One authenticated stream supplies both integral orderings. Transform each placed occupied and each diagonal frame with the inverse unitary cell Fourier matrix. For every ordered, conserving quartet, conjugate the coefficients on Bloch legs one and three. The provider includes the finite-torus normalization. Accumulate two mixed tiles indexed by (n,i,a in [ii],f in [nn]) and contract f with the corresponding home single. This keeps the complete source vector and needs no dense AO ERI.
Qualification¶
The fixture comparison constructs the signed terms independently from finite-supercell AO integrals and independently reconstructed PNOs. It also differences both actual native snapshots in each signed spectral sector of the AO ERI, including physical Fock once. ORCA and CRYSTAL do not expose this isolated native boundary. The gates are 1e-10 Ha per singles component and 1e-12 Ha for imaginary reconstruction. The comparison includes zero and rescaled T1, changed T2, wrong exchange ordering and omitted exchange.
One action reserves 42,991,616 bytes: 8 MiB for the domain, 32 MiB for the panel and 1 MiB for numerical workspace. The numerical bound is 120,960 bytes, including frame validation, Fourier projections, both complex accumulators, products, real reconstruction and all three output vectors per target. This reservation is separate from process RSS. The work bound is 231,276,511,232 units for setup and 256 domains with 27 elements each.
The next isolated boundary is three_external - singles_integrals. It
requires the ordered three-external integral (mf|ae): f and e in [im],
a in [ii]. For each retained [im] with m in L_ii the provider must supply
the target-specific X3 role, row f*n_im+e, and the oriented doubles
2 T_im(e,f) - T_im(f,e). Missing pairs or empty X3 blocks skip the update.
No three-external action is included here.
Finite-fixture verdict¶
All twelve components at six occupied placements passed. The table gives the largest absolute error for each physical comparison, with its sign (actual minus reference). Both indices start at zero.
Comparison |
Placement i |
Component a |
Signed error (Ha) |
|---|---|---|---|
direct versus dense AO |
1 |
0 |
-3.469446951954e-16 |
exchange versus dense AO |
2 |
1 |
+1.117161918529e-15 |
increment versus dense AO |
2 |
1 |
+1.297573160031e-15 |
increment versus native |
2 |
0 |
-1.229485263599e-15 |
native versus dense AO |
5 |
1 |
+3.157196726278e-16 |
The largest imaginary component of either mixed tile was 7.719519468097e-17 Ha. Omitting exchange differed from the reference by +8.362509867526e-02 Ha and replacing (ni|af) with (na|if) differed by +4.730499771076e-02 Ha at their respective worst components. The sparse target neighborhood, missing diagonal and missing off-diagonal controls were independently distinguishable. Zero T1 gave exact zero; scaling T1 by -1.7 scaled each piece, and scaling T2 by -2 with fixed T1 left the increment unchanged.
The physical qualification used 539.766 CPU seconds, 544.407 elapsed seconds and 295.016 MiB peak RSS, including repeated applications for the controls. Exact budgets admitted; one-below limits and stale or foreign input were refused. No matching occupied-Fock receipt was available, so no cumulative R1 result is claimed.