Complete finite-fixture virtual ladder¶
The experimental half_virtual_ladder action completes the virtual-ladder
family on the accepted four-band finite BIPOLE (3,1,1) fixture. It uses
the same authenticated terminal generation, seven pair representatives,
unequal AO domains, rank-two and rank-three PNO frames, deterministic
doubles, and two distinct nonzero diagonal singles vectors as the
six-term assembly.
This is a residual-family qualification, not a CCSD iteration or energy.
Equations and frames¶
Write the finite AO operator as \(E_{uvxy}=(uv|xy)\), its overlap as \(S\), the six occupied columns as \(O_m\), and the target frame as \(C_p=C_{ij}\). Each ordered source pair has its own frame \(C_{mn}\) and amplitude \(T_{mn}\). The diagonal input vectors retain their complete AO components:
Pair reversal transposes the ordered amplitude before projection. This
tau uses coefficient one for the singles product. The half-singles
taut used elsewhere in the residual is a different intermediate.
The two occupied-index intermediates are
The contracted singles in \(Z1\) are in their original diagonal frames. Replacing them by target projections inserts unwanted projectors. The new action is exactly
It exposes tt, t_z1, ss_z, ss_z1, and total separately. Every
monomial includes the factor one half. Six diagonal occupied pairs occur
once; all thirty off-diagonal ordered pairs occur once. Grouping an
off-diagonal pair requires both oriented contractions, not an assumed
scalar orbit multiplier. There is no extra exchange or cell factor.
Under \(T_2\mapsto\alpha T_2\) and \(T_1\mapsto\beta T_1\),
Zero singles leave \(H_{TT}\); zero doubles leave the quartic singles term. All four monomials reverse by transpose. In Bloch coordinates the first and third ERI columns are conjugated and the two contracted ket columns are not. Signed transfers, momentum conservation, the omitted \(G=0\) mode and the finite erfc convention are inherited unchanged from the source.
In cpp/src/dlpno_ccsd.cpp, tproj, t1p, and tau_proj implement these
frame projections. Q_m=Bov_p[m]*T_ij produces Z, while
Z1=Wi.transpose()*Wj uses the full diagonal singles built by the shared
evaluator. The final occupied-index loop inside the virtual block adds
the displayed half-weighted term. The preceding occupied block has its
own half-weighted term; that other copy is outside this action.
Stanton et al., Sec. II, Eqs. (2), (7), (9), (10) give the parent spin-orbital residual, virtual intermediate and effective two-particle amplitudes. Coefficients here follow their closed-shell spatial spin integration in the shared native equations, rather than copying an antisymmetrized spin-orbital prefactor. Riplinger and Neese, Sec. II C give the pair-space overlap construction. DePrince and Sherrill, Sec. II, Eqs. (11), (27)-(30) describe the singles-transformed ladder, and McClain et al., Sec. 4, Eqs. (23)-(24) give the periodic integral convention. Here signed finite BIPOLE panels replace auxiliary factors without a positive-Gram assumption.
Bounded action and qualification¶
The production entry point in vibeqc.dlpno.bipole_virtual_half accepts
the result, authenticated prepared pair build, seven real doubles
matrices, and two real home-diagonal singles vectors. The returned five
tuples contain immutable representative matrices; placed(spaces,i,j)
returns the ordered placement of every monomial and their sum.
One exclusive source lease visits 256 domains and 6,912 normalized
elements. It builds only two seven-by-six-by-six occupied intermediates.
After the panel pass, one source-pair amplitude is projected at a time.
No inventory of all projected tau matrices or dense AO ERIs is retained.
resource_census() accounts for caller and copied amplitudes, AO/Bloch
inputs, projected singles, Z/Z1, temporary products, accumulators, immutable
outputs and retained assembly arrays. These numerical arrays fit a 64 KiB
subreservation of the existing 1 MiB transform slot. The unchanged total
reservation is 42,991,616 bytes: 8 MiB owner/domain, 32 MiB native panel,
and 1 MiB transforms/output. It is not a process-RSS bound.
WORK_BOUND includes setup, every domain, every panel element, and 65,536
consumer arithmetic units per element. Exact budgets are admitted;
one byte or work unit less is rejected before contraction.
The executable qualification
constructs the dense AO oracle before invoking production. Independent
PAO/PNO bases carry each amplitude from its own frame using AO-metric
overlaps. It checks each monomial, the isolated half contribution, and
the complete assembly over 36 ordered targets and 174 components, with
an absolute tolerance of 1e-10 Ha declared before production. Tests
cover both ranks, diagonal/off-diagonal multiplicities, zero amplitudes,
the full polynomial scaling, reversal, stale frames and generations,
and exact/one-below budgets. Omitted singles products, wrong half factors,
double-counted diagonals and omitted reversed occupied pairs must each
produce an error above the gate.
The complete assembly adds this action to the unchanged six terms.
Calls remain sequential, with each action’s own work budget; their sum
is recorded as the assembly work allowance. There is no fused-pass claim.
The reference also evaluates the actual shared native residual using a
signed spectral resolution of the independent AO operator. Its complete
virtual increment is virtual_ladder - occupied_ladder. The bare and
occupied snapshots are nonzero, so subtracting bare or using a cumulative
snapshot is an effective negative control. Later Fock and ring families
are excluded by the adjacent snapshot boundary. Dense ERIs and their
spectral factors exist only in the harness.
The scope is the complete virtual-ladder snapshot on this finite fixture. The complete occupied family is also qualified on the same finite fixture, including both oooo terms, both fixed-T2 ooov terms, both cubic ooov terms and the full half expression. Its boundary is occupied_ladder minus bare; the complete virtual family and all earlier arrays remain unchanged. R2 Fock dressings, rings, R1, iteration, triples and materials acceptance remain open. An external molecular total energy cannot reference this isolated finite-source tensor; the independent AO oracle and shared native snapshot provide the numerical reference within the stated scope.