vibe-qcDocumentation

Cubic occupied-ladder singles products

The experimental cubic_occupied_ladder_terms action selects the two ooov*ss dressings of the finite BIPOLE occupied ladder. It reuses the authenticated terminal provider from the fixed-T2 singles action.

Derivation from the shared native expression

In cpp/src/dlpno_ccsd.cpp, the occupied_ladder block constructs

\[W_{mnij}=(mi|nj)+(mi|nd_j)+(md_i|nj)+\tfrac12(Z_{mn}+Z1_{mn})\]

and adds \(\sum_{mn}W_{mnij}\tau_{mn}^{(p)}\), where \(p=(ij)\) and

\[d_m=C_{mm}t_m,\quad s_m=C_p^T S d_m,\quad B_{mn}^{(p)}=(C_p^T S C_{mn})T_{mn}(C_p^T S C_{mn})^T,\quad \tau_{mn}^{(p)}=B_{mn}^{(p)}+s_ms_n^T.\]

Here \(S\) is the AO overlap. The native t1p constructs \(s_m\) from the diagonal-pair overlap, while tau_proj adds its outer product to tproj. Replacing \(B_{mn}^{(p)}\) by the disconnected product \(s_ms_n^T\) in Boo_i.transpose()*Wj gives

\[C_X^{ab}=\sum_{mn}s_m^a s_n^b(mi|nd_j).\]

Making the same replacement in (Boo_j.transpose()*Wi).transpose() gives

\[C_Y^{ab}=\sum_{mn}s_m^a s_n^b(md_i|nj).\]

The second expression uses electron-pair exchange of \((nj|md_i)\); the transpose places its occupied axes in the same ordered \((m,n)\) convention. Both coefficients are +1. All six occupied placements participate in each sum: six diagonal ordered pairs once and thirty off-diagonal ordered pairs once. There is no half, orbit, exchange or cell multiplier. This is the disconnected part of tau, not the half-weighted singles product in taut.

The integral inputs \(d_i,d_j\) retain their complete diagonal frames. Only the output vectors \(s_m,s_n\) are projected into the target pair. Prematurely replacing an integral input by \(C_pC_p^TSd_i\) changes the operator. In complex Bloch coordinates \((p^*q|r^*s)\) conjugates the first and third orbital columns, not the ket columns. The accepted real, time-reversal pair representation uses ordinary transposes. Reversing the target exchanges the two contributions and transposes them: \(C_X(ji)=C_Y(ij)^T\), \(C_Y(ji)=C_X(ij)^T\). Uniform singles rescaling by \(\beta\) multiplies each term by \(\beta^3\); zero singles gives exact zero. Neither term depends on doubles.

This reuses the primary derivation of Stanton et al., Sec. II, Eqs. (2), (6), (10), with the shared native closed-shell spin integration fixing the spatial coefficients. Pair projections follow Riplinger and Neese, Sec. II C; the periodic convention follows McClain et al., Sec. 4, Eqs. (23)-(24). The finite signed BIPOLE panels replace auxiliary factors without a positive-Gram assumption. The previously reviewed primary reading and provider evidence are reused.

Streamed action and resources

One exclusive terminal lease visits 256 domains and 6,912 normalized elements. Two seven-by-six-by-six occupied intermediates hold the X/Y weights. The action then constructs six projected singles for one target and streams one outer product at a time. Its immutable x, y, total outputs use the accepted exchanged-leg placed convention. It does not allocate a dense production ERI or a target-by-source product inventory.

resource_census() counts inputs and copies, the AO metric and frames, AO/Bloch singles and occupied columns, occupied intermediates, projection scratch, singles products, accumulators, immutable outputs and retained earlier contributions. The conservative numerical bound is 30,496 bytes, within a 64 KiB subreservation of the existing 1 MiB transform slot. The total action cap remains 42,991,616 bytes and 231,276,511,232 work units. Exact limits pass; one byte or work unit below refuses before contraction. This is an array reservation, not a process-RSS bound. Terminal generation and pair frames are authenticated before and after the action. Signed transfers, omitted G=0 and finite erfc subtraction are unchanged.

Oracle and assembly boundary

The executable regression first constructs independent dense 12-AO supercell contractions with independently placed PAO/PNO frames. It evaluates each term and their sum on all 36 ordered targets, 174 components per comparison, at a predeclared absolute 1e-10 Ha tolerance. The accepted four-band unshifted (3,1,1) fixture retains unequal ranks, domains, and nonzero deterministic diagonal singles. It includes targets with distinct X/Y values. Premature input projection, transposed output legs, and a spurious half coefficient must each fail above tolerance.

For example, at zero-based target/component (5,3,1,0), the independent X and Y values are -0.0003141764942972162 and +0.00017278125150538157 Ha. Their difference is well above tolerance. The oracle-only run isolates the cubic sum from the native snapshot to 8.7105e-16 Ha. Premature input projection fails by 1.3564e-4 Ha and transposing the output product fails by 2.5992e-4 Ha.

The actual native reference boundary is occupied_ladder - bare. To isolate these cubic terms, explicitly subtract independent oooo*T2, oooo*ss, both fixed-T2 ooov dressings, and the complete half-weighted expression. The reference-only dense and signed-spectral calculations do not qualify the other production terms. Molecular total energies are not matching references for this isolated finite operator.

The precisely defined partial assembly is

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

Here X/Y are the accepted fixed-T2 terms. The half-weighted expression is its occurrence in the occupied block; the separately qualified full virtual ladder must not be added inside this snapshot. The subsequent complete occupied-family qualification adds oooo*T2 and oooo*ss, qualifies the T1-free sum and compares the whole occupied-minus-bare increment. The cubic and earlier arrays remain exactly unchanged. No occupied term remains on this finite fixture; R2 Fock dressings, rings, R1, iteration, triples and materials remain open.

The partial assembly runs three sequential actions at their individual caps, totaling 693,829,533,696 work units. It does not claim a fused pass. The qualification compares the earlier X, Y, linear sum, half term and earlier partial sum bit-for-bit with their accepted receipt; full virtual qualification and normalized ooov momentum evidence are reused.

The physical qualification passes all 36 targets and 174 components per comparison. Maximum production errors are 1.0460e-16 Ha for X, 1.0482e-16 Ha for Y, 1.1714e-16 Ha for their sum and 9.1681e-16 Ha for the partial occupied assembly. Earlier outputs are exactly unchanged. Zero singles, negative and positive cubic rescaling, exchanged-leg reversal, stale state/generation/frame checks and exact/one-below resource limits pass. The single operational receipt records every signed component, source/native identities and measured resources outside the product tree.