Four-virtual action on the disconnected singles product¶
This bounded contribution uses the accepted finite four-band BIPOLE
RHF fixture, full unshifted (3,1,1) mesh, two inequivalent hydrogen
molecules, no frozen occupieds, tcut_mkn=0.01, MP2-norm PNOs at
tcut_pno=1e-6, and no pair screening. The seven representative ranks
are (2,2,2,2,2,2,3). It is an experimental residual component.
Let \(C_{ii}\) and \(C_{jj}\) be the placed diagonal input PNO frames and \(C_p\) the target frame. In the finite supercell, define
The coefficient is +1. There is no half factor, exchange multiplier, orbit weight or division by the number of cells. Reversal exchanges \(d_i\) and \(d_j\), hence \(D_{ji}=D_{ij}^T\) and \(R_{ji}=R_{ij}^T\) in the translated target frame. In complex Bloch coordinates the first and third ERI coefficients are conjugated; the two input singles coefficients are not. Transfer obeys \(k_a-k_e=k_f-k_b\) modulo the mesh. Signed transfers, the omitted zero mode and finite erfc subtraction belong to the existing terminal source and are consumed unchanged.
The shared native virtual_ladder block uses
u_i.transpose() * u_j, where u_i(P,a) contracts a diagonal-space
single with the four-virtual factor before projecting the output index.
This is the disconnected part of \(\tau_{ij}=T_{ij}+t_i t_j^T\).
The input spaces of its two summands differ. Replacing the singles by
\(C_p^T S C_{ii}t_i\) before applying the target-only four-virtual tensor
would insert two unwanted target projectors. The action therefore does
not pass a target-space outer product to bare_virtual_ladder.
The executable physical regression
builds independent diagonal bases \(V_{ii}\) and target bases \(V_p\). It
rotates input coordinates with \(V_{ii}^T S C_{ii}\), constructs the full
AO product, contracts explicit supercell ERIs, then rotates the output
with \(V_p^T S C_p\). It compares every component in all 36 ordered
placements (174 components), with a predeclared absolute gate of
1e-10 Ha. Supplying only Bvv and zero doubles to the shared native
evaluator independently isolates this term. The signed spectral
resolution used for that check is test-only and assumes no positive Gram
matrix. An external molecular total energy is not a reference for this
finite-source tensor.
The production entry point is disconnected_virtual_ladder in
vibeqc.dlpno.bipole_virtual_disconnected. Its two real input vectors
belong, in order, to the home diagonal frames of occupied bands 0 and 1.
It returns seven immutable representative matrices. Use
spaces.orient_tensor(i, j, result) for ordered placements. The source
generation and all pair coefficients must match the authenticated build.
The source reserves 8 MiB for the owner/domain, 32 MiB for one native
panel including backend margin, and 1 MiB for transforms, snapshots,
receipts and output: 42,991,616 bytes. The new consumer’s live numerical
arrays fit within a conservative 64 KiB subreservation of that last MiB.
This includes immutable singles, six placed AO inputs, their Bloch
projections, seven target projections, seven complex accumulators,
per-element products, and output copies. No four-index virtual tensor or
full AO product is allocated by production. One exclusive source lease
visits 256 domains and 6,912 normalized AO elements. Work is bounded by
WORK_BOUND, including setup and every contracted element; exact budgets
are accepted and one byte or work unit less is refused before the action.
This reservation is not a bound on total Python-process RSS.
The four-term result qualified by this regression covers exactly the bare
vvvv x T2, the two negative fixed-T2 ovvv updates, and the positive
vvvv x (t1_i t1_j).
The subsequently qualified
cubic dressings and six-term assembly
add the two contributions
\(-\sum_m (w_j^T u_i)_{ma}s_m^b\) and
\(-\sum_m s_m^a(w_i^T u_j)_{mb}\).
The subsequent complete virtual-ladder qualification
adds the entire half-weighted
\(\tfrac12\sum_{mn}\tau_{mn}^{ab}(Z_{mn}+Z1_{mn})\) contribution, with
all four doubles/singles products separately exposed and tested.
Here \(s_m=C_p^T S C_{mm}t_m\),
\(Z_{mn}=\sum_{ef}(me|nf)T_{ij}^{ef}\) and
\(Z1_{mn}=\sum_{ef}(me|nf)t_i^e t_j^f\) with the original diagonal
input frames. The historical four-term result above covers only its stated
components; the linked complete-family qualification adds the remaining
terms and isolates the native snapshot boundary. Occupied ladder, CCSD
iteration, triples and materials acceptance remain separate.
DePrince and Sherrill, Sec. II, Eqs. (11), (27)-(30) give the four-virtual contraction and singles transformation; McClain et al., Sec. 4, Eqs. (23)-(24) give the complex periodic integral and momentum convention. The explicit tau expansion here follows the shared native residual, while finite BIPOLE panels replace auxiliary factors. The papers’ molecular or bulk energies are not used as tensor references.