Finite BIPOLE first R1 Fock contributions¶
The experimental vibeqc.dlpno.bipole_r1_fock action evaluates the bare
fock seed and the isolated virtual_fock - fock update of
shared_residual_singles in cpp/src/dlpno_ccsd.cpp. Each result contains
six vectors, one per occupied placement in its diagonal [ii] PNO frame.
The accepted four-band, full unshifted (3,1,1) fixture has twelve singles
components. Inputs are the supplied nonconverged seed-586 doubles and the
complete home singles (0.143, -0.081) and (-0.097, 0.169).
Exact equations and source¶
Let O contain the six occupied columns, C_mn the oriented pair PNO columns, S the physical AO overlap and F the physical AO Fock matrix. These matrices come from the result-owned authenticated terminal generation and its bound pair build. The source adapter reconstructs real finite-supercell F/S from the terminal Bloch matrices. Orbital energies do not replace F.
The seed is r_i(a) = O_i^dagger F C_ii(a). There is no extra overlap
matrix between the AO Fock matrix and its orbital columns. The overlap
belongs in coordinate changes between frames. In a complete orthonormal
virtual coordinate system V, define
U_mn = V^T S C_mn
D_m = U_mm t_m
T_mn = U_mn t_mn U_mn^T
P_i = U_ii^T
Every reversed source double is oriented before lifting. Complete diagonal
singles and both source-double axes retain their original pair spaces.
The mixed output projection is equivalent to C_ii^T S C_mn; the contracted
source leg is not truncated to [ii]. With chemists’ integrals, the exact
increment is P_i Fv D_i, where
Fv_ae = f_ae - 1/2 sum_m D_m(a) f_me
+ sum_mf D_m(f) [2(mf|ae) - (me|af)]
- sum_mnf [T_mn(a,f) + 1/2 D_m(a)D_n(f)]
[2(me|nf) - (mf|ne)].
All six occupied indices and all 36 ordered pairs are summed once. There is no additional cell, orbit, spin or pair half factor. Full physical Fock diagonals enter this raw residual. The occupied diagonal contribution and denominator update belong to later boundaries.
This is Stanton et al., Sec. II, Eqs. (1), (3), (9), (12)
with the shared native closed-shell spatial direct/exchange coefficients.
Stanton removes diagonal Fock elements onto the equation’s left side; this
native raw residual retains them. The half disconnected amplitude in Fv
must remain a half. The qualified R2 Fock-half action combines additional
explicit corrections into a full disconnected product, so that R2
intermediate cannot replace this one. Mixed metric projections follow
Riplinger and Neese, Sec. II D, Eqs. (26)-(30).
The existing citation keys are stanton_gauss_watts_bartlett_ccsd_1991
and riplinger_dlpno_2013 in the citation database.
Bounded actions¶
fock_seed(result, prepared, max_bytes=..., max_work=...) uses authenticated
physical F/S and occupied/diagonal PNO columns without an integral lease.
virtual_fock_terms(result, prepared, doubles, singles, ...) exposes the
seed, five pieces, isolated increment and seed plus increment. That sum
corresponds exactly to the native virtual_fock snapshot.
The five pieces are linear_fvv, quadratic_fov, quadratic_ovvv,
doubles and cubic. Under T1 -> alpha T1 and T2 -> beta T2 their
multipliers are alpha, alpha squared, alpha squared, alpha beta and
alpha cubed. At zero singles the entire increment is exactly zero,
including for nonzero doubles, while the physical seed is unchanged.
Two sequential leases reuse authenticated ovov blocks and stream ovvv from the accepted terminal finite operator. Signed momentum transfers, conjugated first and third Bloch legs, omitted G=0 and finite erfc subtraction remain unchanged. Production builds orbital blocks without materializing AO ERIs or requiring a positive auxiliary Gram matrix. Source state, generation, occupied/PNO frames, overlap and Fock binding are checked. Invalid integral block identity, issuer or transfer refuses.
The seed work reservation is 1,048,576 units. The update uses the accepted
conservative two-lease Fock envelope, 463,006,007,296 work units and
42,991,616 bytes. Each lease visits 256 domains and 6,912 normalized
elements. resource_census() includes live complex accumulators and
real copies, outer-product scratch, authenticated blocks, all 36 rotations
and lifted source doubles, complete singles, five simultaneous virtual
intermediates, projections and retained outputs. The numerical bound is
separate from measured process RSS.
Qualification contract¶
The executable independent oracle contracts finite-supercell AO integrals in independently constructed PNO frames. A separate signed spectral resolution feeds the actual native snapshot evaluator; the physical Fock is included exactly once. The gates are absolute 1e-10 Ha per component and 1e-12 for real reconstruction. Each seed, isolated increment and combined snapshot covers all six occupied placements and all twelve components. Every production piece is also compared separately with its AO expression.
The unmodified converged RHF fixture remains the scientific reference.
A controlled Hermitian perturbation
deltaF = S (O K V^T + V K^T O^T) S, with a known nonzero occupied-virtual
block K, checks the seed kernel and actual native snapshots independently
of the small physical occupied-virtual Fock. This diagnostic does not
replace or mutate the authenticated physical source. Complex orbital gauge
controls expose missing bra conjugations; sign, frame, extra-metric,
early-projection, source-double reversal and incorrect full-weight R2
substitution must separate above the gate. Stale source/frame and exact
versus one-below byte/work budgets are exercised for both public actions.
Independent analytic AO and actual native isolated tensors are the matched references. ORCA/CRYSTAL do not expose these bespoke finite-operator snapshots. Reuse of complete raw R2 evidence requires matching unchanged shared source and native binary identities.
Residual boundary¶
The first two R1 contributions are qualified on the accepted finite fixture.
The next unqualified R1 increment is
occupied_fock - virtual_fock, the negative occupied-Fock contraction
with projected output singles. It requires physical f_oo/f_ov, complete
singles, oriented source doubles, ooov/ovov and mixed pair metrics. Later
boundaries are doubles_fme, singles_integrals, three_external and
complete. Iterative convergence, broader materials acceptance and triples
remain open under #586/#592/#1049.
Finite-fixture verdict¶
Every singles component at all six occupied placements passes the predeclared 1e-10 Ha absolute gate. The following errors are in Ha.
Contribution |
Independent AO error |
Native snapshot error |
|---|---|---|
seed |
8.903384284624e-15 |
9.386343669982e-15 |
linear_fvv |
6.994405055138e-15 |
included in increment |
quadratic_fov |
6.099676519521e-17 |
included in increment |
quadratic_ovvv |
7.743914016978e-17 |
included in increment |
doubles |
9.450990337556e-16 |
included in increment |
cubic |
2.766070792554e-17 |
included in increment |
increment |
6.189493362285e-15 |
5.870304242706e-15 |
total |
1.262878690511e-14 |
1.468269950067e-14 |
The unmodified physical seed has maximum magnitude 4.080949860034e-15 Ha in the independent AO oracle; the production seed maximum is 1.130345816946e-14 Ha. The known occupied-virtual Hermitian Fock perturbation produces a seed of 3.490469310957e-02 Ha, with known-block error 1.318389841742e-16 Ha. Complex frame covariance passes to 8.879328308950e-15 Ha. These controls exercise the same seed/contraction kernels and actual native snapshots; the physical provider remains unmodified.
Deliberately incorrect action |
Separation (Ha) |
|---|---|
R2_full_instead_of_R1_half |
5.038772990094e-04 |
early_input_projection |
2.573096141559e-03 |
integral_missing_bra_conjugations |
5.562871982821e-02 |
reversed_source_doubles |
3.716401746868e-03 |
seed_extra_metric |
3.982205006729e-02 |
seed_missing_bra_conjugation |
6.383773940892e-02 |
seed_wrong_frame |
4.515246008775e-02 |
seed_wrong_sign |
6.980938621914e-02 |
wrong_update_sign |
3.980721163630e-01 |
Zero singles gives exactly zero increment with unchanged physical seed. The combinations (alpha, beta) = (0, 1), (-1.7, 0.6), (0.6, -1.25) and (1, 0) pass production, independent AO, actual native and the term-specific polynomial laws. Each comparison includes signed component arrays in the qualification receipt. Exact caps pass; one-below caps, stale source/frame/Fock, foreign handle/generation and invalid integral block identity/issuer/transfer all refuse.
The simultaneous numerical-array bound is 482,560 bytes within the 42,991,616-byte reservation. The seed needs no integral lease; each update uses two sequential leases. The qualification job used 669.004128 calculation CPU seconds and 672.648285 monotonic elapsed seconds, with 282.449219 MiB peak process RSS. The receipt records vq accounting separately and retains loaded source/native hashes.
Complete raw R2 evidence is reused without rerunning it: the native binary, shared residual and finite-panel source blobs, all previously loaded product modules, physical source record, pair-build identity and supplied amplitudes match the accepted final-R2 qualification. No shared product code or native runtime behavior changed within this pinned qualification runtime. Raw MP2 remains -0.040746630709696315 Ha/cell.
Only the seed, isolated virtual-Fock increment and their sum at supplied amplitudes are qualified here. Remaining R1 terms, iterative convergence, bulk/material acceptance and triples stay open.