vibe-qcDocumentation

The fixed-doubles, linear-singles virtual ladder

On the finite four-band BIPOLE fixture, the virtual ladder has two terms linear in the CCSD singles at fixed ordered doubles. In target pair \((ij)\), write \(s_m = V_{ij}^T S V_{mm} t_m\) for a placed diagonal single projected into the target. The contribution to the unweighted residual is

\[-\sum_{mef} T_{ij}^{ef} \left[(a^* e|m^* f)s_m^b + s_m^a(m^* e|b^* f)\right].\]

The occupied sum runs over both bands in all three cells. The starred Bloch columns are conjugated, and the first bra minus ket momentum equals the second ket minus bra momentum modulo the mesh. Both signs are negative; there is no exchange multiplier or division by the cell count. Reversing the ordered occupied pair transposes its doubles.

linear_virtual_ladder_terms returns the two signed matrices (x and y) and their total for each representative in one bounded pass. placed(spaces, i, j) exchanges the two matrices and transposes them when the occupied pair is reversed. Even for a diagonal pair, only the sum must be symmetric. The per-leg regression compares both updates and their sum in each ordered pair’s independent supercell PNO frame. It also tests zero singles and the effect of omitting the required doubles transpose.

The finite-fixture example uses the terminal RHF source, the seven MP2-norm PNO frames and the doubles from the bare ladder. It holds these fixed as singles change. It reconstructs the supercell integrals independently and compares the change in the native cumulative snapshot difference, virtual_ladder - occupied_ladder. Besides the linear terms above, that change contains the four-virtual action on the disconnected singles product, its two mixed dressings, and the change in the half-weighted occupied-virtual contraction. The disconnected product uses the original diagonal PNO frames on its input legs. Projecting it into the target on those legs would change the equation.

DePrince and Sherrill, Sec. II, Eqs. (11) and (27)-(30) describe the singles transformation of the four-virtual ladder. Here ordered BIPOLE panels replace auxiliary factors. The signed spectral resolution of the reference ERIs is only a way to exercise the shared native factor evaluator. Momentum conservation follows McClain et al., Sec. 4. Only the two displayed terms are provided by this increment. The other powers of singles, the full residual and iterative CCSD remain outside it.

Cubic dressings and six-term assembly

The experimental cubic_virtual_ladder action in vibeqc.dlpno.bipole_virtual_cubic returns x, y and total, each with one matrix per representative. Its placed(spaces, i, j) accessor exchanges and transposes x and y on pair reversal. The accepted quadratic action in vibeqc.dlpno.bipole_virtual_disconnected remains separate and unchanged.

Let \(d_i=V_{ii}t_i\) and \(d_j=V_{jj}t_j\) retain their full AO components. For target \(V=V_{ij}\), all six occupied columns \(O\), and projected output singles \(s[:,m]=V^T S d_m\), define

\[K_{ux}=\sum_{vy}(uv|xy)d_i[v]d_j[y].\]

The two cubic updates are \(C_X=-V^T K O s^T\) and \(C_Y=-s O^T K V\). In cpp/src/dlpno_ccsd.cpp, T1full synthesizes the full diagonal single before the u/w contractions. u_i(P,a) projects only its free index \(a\) into the target, while w_j(P,m) contracts the full \(d_j\) with occupied \(m\). Thus the two subtractions after u_i.transpose()*u_j are precisely \(C_X\) and \(C_Y\), with coefficients minus one and no half, exchange or cell factor. The output t1p is target-projected; prematurely projecting either contracted diagonal input changes the equation. The literature and finite operator conventions above apply without a refit.

The cubic regression uses explicit 12-AO supercell integrals and independent placed PAO/PNO frames. It checks each cubic term, their sum and the six-term assembly on all 36 ordered placements, including rank-two and rank-three targets. Controls include zero singles, common cubic scaling, exchanged-leg reversal, stale and foreign state, one-below budgets, wrong contracted legs and premature input projection. Native cumulative snapshots are compared after separately accounting for the quadratic, linear and half-weighted changes.

The assembly calls the accepted bare, linear and quadratic actions and then adds the two cubic terms. Each action keeps its existing work and memory cap; calls run sequentially. There is no claim of a fused single pass for the assembly. The retained arrays fit the existing transform slot.

The six fixture contributions are bare vvvv*T2, both negative fixed-T2 linear-singles dressings, the disconnected quadratic action, and these two negative cubic dressings. The subsequent complete virtual-ladder action and qualification adds \(\frac12\sum_{mn}\tau_{mn}^{(target)}(Z_{mn}+Z1_{mn})\), including all four doubles/singles products, as a separately tested contribution. That full family uses the adjacent native snapshot difference virtual_ladder - occupied_ladder. Occupied-ladder, iterative CCSD, triples and materials acceptance remain outside this six-term increment.