Finite BIPOLE final R2 half updates¶
This experimental action evaluates half_ij - half_ij_ring and its ji
counterpart in shared_residual_pair in cpp/src/dlpno_ccsd.cpp.
half_update_terms exposes the four terms, each raw orientation and each
ij + ji.T sum. doubles_residual adds the final update exactly once to
the accepted bare, complete occupied/virtual ladder, Fock and ring assembly.
It evaluates a residual at supplied amplitudes; it does not iterate CCSD.
Equations and frames¶
For target pair p, write its orthonormal PNO columns as C_p, the AO metric as S, complete diagonal singles as d_m = C_mm t_m, and projected output singles as s_m = C_p^T S d_m. In chemists’ integral ordering, each raw ij half has the following four contributions:
Name |
Target-frame expression |
Degree in singles |
|---|---|---|
|
-sum_m s_m(a) (m d_i|j b) |
2 |
|
-sum_m s_m(a) (m i|b d_j) |
2 |
|
+(a d_i|j b) |
1 |
|
-sum_m s_m(a) (m i|j b) |
1 |
Each m sum covers all six occupied orbitals, including each diagonal once. Coefficients are -1, -1, +1, -1, with no additional half, cell or spin factor. Integral singles inputs d_i/d_j remain in their complete diagonal frames. Only free a/b output legs and the output singles s_m are projected into the target. The implementation uses a complete six-dimensional virtual coordinate system; this changes coordinates without truncating the inputs.
The ji half swaps i and j in these expressions without first transposing its output. Assembly is ij + ji.T. On pair reversal each raw orientation exchanges with the other without transposition, while the assembled sum transposes. A diagonal target uses h_ii + h_ii.T once.
The source is Stanton et al., Sec. II, Eq. (2), (8)-(13), with the shared native closed-shell spatial coefficients. PNO metric projections follow Riplinger and Neese, Sec. II C-D, Eq. (26)-(30). The spin-orbital antisymmetrizers of Stanton are resolved into the two spatial orientations here. Unlike Stanton’s right-hand side with diagonal Fock entries removed, the native raw residual contains full physical Fock diagonals. Its diagonal limit already contains -D*T. Denominator energies belong to the subsequent iteration, so no additional denominator or diagonal subtraction is made by this action.
Native inventory and reference¶
The native R2 function performs exactly these operations, in order:
bare: the target (ia|jb) seed.occupied_ladder - bare: the complete occupied family, including its half contraction.virtual_ladder - occupied_ladder: the complete virtual family, including its separate half contraction.Local halves initialized at zero, recorded at
half_ij_fockandhalf_ji_fock.Ring increments, recorded at
half_ij_ringandhalf_ji_ring.The four updates above, recorded at
half_ijandhalf_ji.Addition of
half_ij + half_ji.Tto the cumulative virtual-ladder accumulator, then return.
Thus the complete raw R2 is virtual_ladder + half_ij + half_ji.T.
The cumulative ladder snapshots are not subtracted from local half snapshots.
No further native R2 term occurs after the final half assembly.
The executable qualification example builds independent dense finite-supercell AO contractions using independently constructed pair frames and the AO metric. Signed spectral resolution is used only by the native reference, never by production. Native runs with Bov alone isolate the first term; Boo and Bvv isolate the second; Bov and Bvv minus Bov alone isolate the third; the remaining full half difference isolates the fourth without using oracle values. Returned native raw R2 is checked separately against the recorded-snapshot inventory and the independently expanded full expression.
The predeclared absolute gate is 1e-10 Ha for every component of all 36 ordered targets / 174 components. The real-reconstruction gate is 1e-12. Zero singles, uniform scales -1.7 and +0.6, pair reversal, wrong sign, early input projection and a missing ji transpose are checked. Exact budgets pass; one-below budgets, stale terminal state, foreign handles, changed frames and invalid block identity/issuer/transfer are refused. Earlier signed arrays may be reused only after exact source, physical provenance, pair-build, frame and amplitude equality. The qualification record distinguishes replayed earlier actions from new provider leases.
Provider and resource boundary¶
The accepted authenticated ovov provider and finite operator stream supply ovov, oovv, ovvv and ooov. Signed momentum transfers, conjugated first and third Bloch coefficient legs, omitted G=0 mode and finite erfc subtraction are unchanged. Production does not materialize dense AO ERIs or assume a positive auxiliary factorization.
Two sequential leases evaluate the four updates together. A fresh full R2
evaluation uses fourteen sequential leases. resource_census() counts
integral accumulators and copies, outer-product scratch, authenticated
blocks, complete singles, rotations, live target projections, mixed
contractions, four accumulators and retained earlier/new outputs. Admission
uses the existing 42,991,616-byte cap and the explicit work bound. These
are numerical-array bounds, not a process-RSS guarantee or a general
scaling claim.
The accepted finite fixture is qualified below. R1, iteration, bulk convergence and
triples remain unqualified. The first subsequent R1 boundary is the bare
fock seed, f_ia in diagonal pair [ii]. It needs the authenticated physical
Fock/overlap, occupied columns and diagonal PNO frame, with source and
frame binding; orbital eigenvalues alone are insufficient. The following
virtual_fock - fock family additionally needs complete diagonal singles,
source doubles, ovov/ovvv blocks and the appropriate mixed pair metrics.
Finite-fixture verdict¶
The accepted two-H2, four-band, full unshifted (3,1,1) fixture passes every component of all 36 ordered targets / 174 components at the predeclared 1e-10 Ha absolute gate. Singles are (0.143, -0.081) and (-0.097, 0.169) in their complete home diagonal frames; seed-586 doubles and unequal PNO ranks are retained. The following maxima are in Ha.
Term |
Dense ij |
Dense ji |
Dense sum |
Native ij |
Native ji |
Native sum |
|---|---|---|---|---|---|---|
quadratic_ovov |
6.1040582311e-16 |
6.1083950398e-16 |
6.1083950398e-16 |
6.0281640790e-16 |
6.0281640790e-16 |
6.1842891919e-16 |
quadratic_oovv |
1.3678294608e-15 |
1.3643600139e-15 |
1.1778772402e-15 |
1.3760693973e-15 |
1.3760693973e-15 |
1.1770098785e-15 |
linear_ovvv |
1.1609636863e-15 |
1.1609636863e-15 |
2.3219273726e-15 |
1.0295583830e-15 |
1.0295583830e-15 |
1.0963452368e-15 |
linear_ooov |
1.2247147740e-15 |
1.2247147740e-15 |
1.8821749714e-15 |
1.1136924716e-15 |
1.1136924716e-15 |
8.4741241801e-16 |
half |
1.3660947373e-15 |
1.3630589712e-15 |
2.3921836734e-15 |
1.3500485452e-15 |
1.3500485452e-15 |
2.3427440543e-15 |
Complete raw R2 agrees with the independently assembled full AO expression to 1.0819123375e-13 Ha and with the final native raw R2 to 1.0780265569e-13 Ha. The independent full expression versus native error is 2.7200464103e-15 Ha. The native snapshot inventory closes to 1.3877787808e-16 Ha. Raw MP2 is -0.040746630709696315 Ha/cell.
All four terms vanish exactly at zero singles. Uniform scales -1.7 and +0.6 pass the quadratic and linear laws in the production contraction, independent AO oracle and separately isolated native differences. Raw pair reversal exchanges the orientations; the combined update transposes. Wrong-sign, early-projection and missing-ji-transpose controls separate by 2.3243161251e-02, 2.6825497952e-03 and 5.2801214830e-03 Ha respectively.
Exact and one-below byte/work budgets and all stale-state/frame/block refusals pass. The four-term action uses 463,458,992,128 work units in two leases. A fresh complete assembly uses 3,238,663,880,704 work units in fourteen leases. The conservative simultaneous numerical-array bound is 657,664 bytes inside the existing 42,991,616-byte reservation.
Qualification ran through vq on a pinned available runtime. The calculation used 826.997000 CPU seconds and 836.335477 monotonic elapsed seconds, recorded separately. The earlier action sources and primary derivations are reused; earlier numerical actions were rerun because this qualification uses a different native binary. The signed receipt records the loaded Python/native identities, complete component arrays, controls and resource measurements. The source-only asymmetric check also passed to 1.39e-17 Ha against the dense oracle and 3.97e-17 Ha against native isolation. No R1, iteration, bulk or triples qualification is inferred from this finite residual test.