Finite BIPOLE R2 Fock halves¶
This bounded action covers half_ij_fock and half_ji_fock in the
shared native residual in cpp/src/dlpno_ccsd.cpp. It uses the accepted
four-band, full unshifted (3,1,1) fixture, all six occupied orbitals,
seven unequal pair spaces of ranks (2,2,2,2,2,2,3), and all 36 ordered
targets with 174 components. It does not evaluate a complete CCSD residual
or perform an iteration.
Equations and snapshot boundaries¶
Let V_p be a target pair frame, S the AO overlap, and
B_mn = (V_p^T S V_mn) T_mn (V_p^T S V_mn)^T.
Complete diagonal singles are d_m = V_mm t_m; their target coordinates
are s_m = V_p^T S d_m. In a complete orthonormal virtual frame, write
t_m^a for those complete singles and T_mn^ab for each source pair’s
lifted doubles. Define tau_mn^ab = T_mn^ab + t_m^a t_n^b.
The two dressed Fock blocks, after combining the explicit half corrections in the native code, are
Fh_ae = f_ae - sum_m t_m^a f_me
+ sum_mf t_m^f [2(mf|ae) - (me|af)]
- sum_mnf tau_mn^af [2(me|nf) - (mf|ne)]
Fh_mx = f_mx + sum_e f_me t_x^e
+ sum_ne t_n^e [2(mx|ne) - (nx|me)]
+ sum_nef tau_xn^ef [2(me|nf) - (mf|ne)]
This tau is the algebraic result of combining native
taut = T2 + 0.5*T1*T1 with Fbe - 0.5*t1*Fme and
Fmx + 0.5*Fme.dot(t1). Substituting tau into Fbe before applying
its explicit half correction would count singles twice. Every occupied
sum runs over all six orbitals. Every ordered occupied pair occurs once,
including its diagonal; there is no orbit, cell or extra half factor.
Project Fh_ae onto both target axes to obtain Fh_p. The raw halves are
H_ij = T_ij Fh_p^T - sum_m Fh_mj B_im
H_ji = T_ij^T Fh_p^T - sum_m Fh_mi B_jm
R_F = H_ij + H_ji^T
half_ij_fock and half_ji_fock record these local halves immediately
after their initialization and occupied contraction, before any ring.
Their preceding local value is zero. The preceding cumulative R2 snapshot
is virtual_ladder, but it is not part of either half and must not be
subtracted from a local half. For i=j, native evaluates H_ii once and
adds H_ii + H_ii^T; there is no separate half_ji_fock key.
Reversing an ordered target exchanges the raw halves without a transpose:
H_ij(ji)=H_ji(ij). Their assembled sum transposes.
The terminal f_oo, f_ov, and full projected f_vv all enter this
residual, including their diagonals. Orbital energies supplied solely to
the denominator do not replace them. The accepted orbital-energy-only probe
does not enter these terminal Fock matrix elements. With singles and integrals zero,
the diagonal part is -D_ij^ab T_ij^ab, where
D_ij^ab = f_ii + f_jj - f_aa - f_bb. The serial iteration uses
T_new = T + R/D. Removing that diagonal part from R, or adding it
again during assembly, would change the serial equation.
Stanton et al., Sec. II, Eqs. (2)-(5), (9)-(10), (12)-(13)
move diagonal Fock terms to the left-hand side. That convention is the
material difference from this complete-residual implementation. Spatial
coefficients follow the actual native closed-shell exchange convention.
Riplinger and Neese, Sec. II D, Eqs. (26)-(30)
define the pair metric projections; complete input singles and doubles are
retained until the specified free legs are projected. The finite signed
BIPOLE operator replaces auxiliary factors without assuming a positive
Gram matrix. Signed transfers obey k1-k2 = k4-k3; first and third
Bloch coefficients conjugate, while ket coefficients and amplitudes do not.
Implementation and precisely covered assembly¶
fock_terms authenticates the
result-owned terminal handle, pair build, occupied and pair frames, and
terminal Fock record. It reuses the ovov provider and streams ooov and
ovvv from the same native panel contract. The two leases are sequential.
It exposes both raw halves and their assembled sum. Intermediate orbital
blocks are small finite-fixture objects; no dense production AO ERI is built.
ladders_fock_assembly returns exactly
bare seed + complete occupied-ladder increment
+ complete virtual-ladder increment + H_ij + H_ji^T
The bare seed is the authenticated pair-build integral. Each qualified
ladder includes its own 0.5*sum_mn tau_mn*(Z_mn+Z1_mn) occurrence.
Their identical arrays can be reused, but the native residual contains
one occurrence in each ladder. The occupied t1_free convenience sum is
not added: it already contains half.tt. Earlier contribution arrays are
returned separately for comparison with the accepted receipts.
Qualification contract¶
The executable dense finite-supercell oracle
is independent of production blocks and projection helpers. It contracts
analytic AO integrals in independently built PAO/PNO frames, with nonzero
home singles (0.143,-0.081) and (-0.097,0.169) and seed-586 nonsymmetric
doubles. The predeclared absolute component gate is 1e-10 Ha, and real
reconstruction uses 1e-12. Each half, their sum, zero-singles controls,
and the covered assembly must pass every ordered target.
The shared native oracle uses a signed spectral resolution only in the reference path. It evaluates Fock-only once separately; including the same Fock matrix in both signed ERI passes would cancel it. A deterministic symmetric Fock sentinel with asymmetric entries exposes bare virtual couplings hidden by semicanonical target frames. The unmodified physical calculation remains the scientific gate.
Controls include off-diagonal omission and sign reversal, wrong half
orientation, diagonal residual accounting, changed denominator-only
energies, stale state, foreign generation and handle, changed pair frame,
and exact versus one-below memory and work caps. All numerical arrays,
36 lifted source amplitudes, one live target-projected amplitude,
intermediate blocks and retained ladder outputs are
included in resource_census; this is not a process-RSS or fused-pass claim.
Residual-family boundary¶
The exact next ring increment is half_ij_ring - half_ij_fock, followed
by its ji counterpart. Its prerequisites are the mixed source/target
PNO blocks forming W1, W2, and WX, complete diagonal singles, source
pair doubles in both orientations, and all six occupied sums. Its three
contractions are S_im (T_im-T_mi) W1, S_im T_im W2, and
S_mj T_mj WX. The later half_ij - half_ij_ring singles/integral updates
are a separate boundary. None is included here. R1, complete R2,
iterations, triples and materials qualification remain open under
#586/#592/#1049.
Finite-fixture verdict¶
The two Fock halves and their specified partial assembly are qualified
on this finite fixture. Every comparison covers all 36 ordered targets
and 174 components at the predeclared 1e-10 Ha absolute gate. The single
signed numerical receipt on #586/#592, linked to #1049, retains the actual,
reference and signed-error arrays, native snapshots, runtime identities,
resource measurements and negative controls.
Comparison |
Worst absolute error (Ha) |
|---|---|
|
2.81441537e-14 |
|
2.77000645e-14 |
Fock sum against dense oracle |
4.68514116e-14 |
|
2.83661983e-14 |
|
2.83661983e-14 |
Fock sum against native snapshots |
4.72955008e-14 |
Covered assembly against dense oracle |
5.49560397e-14 |
Covered assembly against native boundaries |
5.57331958e-14 |
Zero-singles production errors are at most 4.76285678e-14 Ha.
The raw-half exchange and summed-transpose reversal errors are exactly zero.
All 14 occupied and 13 virtual ladder array groups are bit-for-bit unchanged
from the accepted receipts. The bare seed is authenticated and added once.
The omitted occupied off-diagonal, wrong occupied off-diagonal sign,
and omitted dressed virtual off-diagonal controls separate by
5.33869714e-02, 1.06773943e-01,
and 9.06649441e-03 Ha, respectively.
The wrong raw-half transpose separates by 3.61755170e-01 Ha.
The doubled diagonal residual separates by 6.02148904e-01 Ha;
changing denominator-only energies leaves the isolated native Fock residual
unchanged within the gate. The asymmetric Fock sentinel exposes a bare
virtual off-diagonal signal of 1.79152179e-02 Ha.
Exact caps pass and one-below caps refuse: 42,991,616 bytes,
463,006,007,296 work units for the two Fock leases, and
2,311,745,896,448 work units for the ten-lease assembly. The
conservative numerical array census is 503,296 bytes.
The zero-singles action adds two further leases to the qualification run.
The production job measured 222.234 MiB peak process RSS,
reported separately from the numerical reservation. Stale state, foreign
generation/handle, changed pair frame and altered bare seed all refuse.
The reference is the independent analytic finite AO operator plus the actual shared native residual. ORCA/CRYSTAL do not supply a matched isolated tensor for this finite operator; no external total-energy, complete CCSD, iteration, bulk convergence or triples verdict follows.