vibe-qcDocumentation

Complete finite-fixture occupied ladder

The experimental vibeqc.dlpno.bipole_ccsd module completes the occupied family on the accepted four-band BIPOLE fixture. It retains the original Bloch occupied_ladder T1-free action and exposes local pair terms through oooo_occupied_terms and complete_occupied_ladder.

Equation and projections

The shared native cpp/src/dlpno_ccsd.cpp block forms `W = Boo_i.transpose()*Boo_j + Boo_i.transpose()*Wj

  • (Boo_j.transpose()Wi).transpose() + 0.5(Z+Z1)and contracts it withtau_proj = tproj + t1p(m)*t1p(n).transpose()`. For target pair \(p=(ij)\), set

\[B_{mn}^{(p)}=(C_p^TSC_{mn})T_{mn}(C_p^TSC_{mn})^T,\qquad d_m=C_{mm}t_m,\qquad s_m=C_p^TSd_m.\]

The two newly exposed contributions are

\[O_T^{(p)}=\sum_{mn}(mi|nj)B_{mn}^{(p)},\qquad O_s^{(p)}=\sum_{mn}(mi|nj)s_ms_n^T.\]

Both coefficients are +1. All 36 ordered occupied pairs appear once: six diagonals and thirty off-diagonal ordered pairs. There is no half, exchange, orbit or cell multiplier. Source reversal transposes the double; target reversal transposes each result. The complex Bloch integral \((m^*i|n^*j)\) conjugates the first and third orbital columns, while ket columns and amplitude products remain unconjugated. The real pair representation uses ordinary transposes. Every single starts in its complete diagonal frame before projection onto the target.

The disconnected term is independent of doubles, vanishes exactly at zero singles and scales by \(\beta^2\) under uniform singles scaling. The bare term is independent of singles and vanishes at zero doubles.

This reuses the primary derivation of Stanton et al., Sec. II, Eqs. (2), (6), (10), the pair metric of Riplinger and Neese, Sec. II C, and McClain et al., Sec. 4, Eqs. (23)-(24). Spatial coefficients follow the shared closed-shell spin integration. Signed finite panels replace auxiliary factors; no positive-Gram assumption is made.

Assembly and snapshot boundaries

The complete occupied increment is exactly

\[O_T+O_s+X+Y+C_X+C_Y+\tfrac12\sum_{mn} (B_{mn}^{(p)}+s_ms_n^T)(Z_{mn}+Z1_{mn}).\]

Here X/Y are the fixed-T2 ooov terms, C_X/C_Y the cubic ooov terms, and the last expression is the accepted full half term. The API returns the earlier term arrays without modifying them. It also returns the T1-free pair sum \(O_T+H_{TT}\). Adding the full half to that T1-free sum would count \(H_{TT}\) twice: use \(O_T\) directly, or replace \(H_{TT}\) with the full half expression.

The actual reference boundary is occupied_ladder minus bare. The bare snapshot contains the nonzero \((ia|jb)\) seed. The subsequent virtual increment is virtual_ladder minus occupied_ladder and is excluded from the occupied family. The regression checks these boundaries with independent negative controls, including the duplicated half TT term.

Qualification and resource contract

The complete regression first constructs the independent dense 12-AO finite-supercell operator, placed PAO/PNO frames and AO-metric projections. It compares every one of 174 components on all 36 ordered targets, including both ranks, against that oracle and the isolated shared-native snapshot. The native reference uses a signed spectral resolution only inside the harness. It separately checks the bare and disconnected oooo terms and isolates each from the native increment by subtracting the other independent terms.

The accepted two-H2 fixture retains full unshifted (3,1,1), frozen=0, tcut_mkn=0.01, MP2-norm tcut_pno=1e-6, no pair screening, unequal domains and ranks (2,2,2,2,2,2,3). Home singles are (0.143,-0.081) and (-0.097,0.169) with seed-586 doubles. The predeclared component gate is 1e-10 Ha, with 1e-12 Ha for forbidden momentum. Missing reverse pairs and doubled diagonals must fail the unchanged gate. Zero amplitudes, positive and negative quadratic singles scaling, reversal, stale state/generation/frame, foreign handles and exact/one-below memory and work admission are checked.

The oooo action uses the authenticated band provider and transforms its small blocks to occupied weights. It streams one source-pair projection at a time and allocates no production dense AO ERI. Its conservative numerical census includes band copies, transforms, projections, outputs and retained earlier arrays, inside the existing 1 MiB transform slot. Each action admits 42,991,616 bytes and 231,276,511,232 work units. The full assembly calls four actions sequentially, reserving 925,106,044,928 work units. This is neither a fused pass nor a process-RSS bound. Each lease visits 256 domains and 6,912 normalized elements. Signed transfers, omitted G=0 and finite erfc conventions are unchanged.

The occupied/virtual Fock halves and their singles dressings are qualified on the accepted finite fixture against independent dense and isolated native references. Their precisely covered assembly adds the bare seed, this complete occupied family, the complete virtual family and half_ij_fock + half_ji_fock.T once each. The exact next ring increment is half_ij_ring - half_ij_fock, followed by its ji counterpart, using mixed PNO blocks for W1/W2/WX, complete singles and oriented source doubles. Later singles/integral half updates, all R1 families, complete R1/R2 evaluation, convergence, triples and materials acceptance remain open under #586, #592 and #1049.

Finite-fixture verdict

The complete occupied family is qualified on this finite fixture. All 36 ordered targets and 174 components pass the predeclared 1e-10 Ha gate. The single complete-family receipt on #586/#592, linked to #1049, retains signed production/reference/error arrays, source/native identities, measured resources and every refusal/control. Earlier linear, cubic, half and partial-sum output arrays are bit-for-bit unchanged.

Comparison

Worst absolute error (Ha)

bare_oooo against dense operator

5.85642645e-15

oooo_ss against dense operator

8.70831185e-16

t1_free against dense operator

5.93969318e-15

occupied against dense operator

6.27276009e-15

Complete family against isolated native snapshot

6.66133815e-15

The missing-reverse-pair and doubled-diagonal controls separate by 3.10510512e-03 and 1.74226712e-02 Ha. The numerical array census is 245,024 bytes. This verdict does not qualify a complete R1/R2 evaluation or CCSD iteration.