vibe-qcDocumentation

Finite BIPOLE R2 ring increments

This experimental action evaluates exactly half_ij_ring - half_ij_fock and half_ji_ring - half_ji_fock in the shared native residual in cpp/src/dlpno_ccsd.cpp. The accepted four-band (3,1,1) finite fixture has six occupied orbitals, seven unequal pair spaces of ranks (2,2,2,2,2,2,3), and 36 ordered targets containing 174 components. The later half_ij - half_ij_ring updates are excluded.

Equations and mixed projections

Write g_mn(e,f) = (me|nf), h_mxn(e) = (mx|ne), k_mx(b,e) = (mx|be) and l_m(e,b,f) = (me|bf). The complete diagonal singles are d_n = V_nn t_n in AO coordinates. Each ordered double is P_nx = V_nx T_nx V_nx^T, so P_xn = P_nx^T but the two contractions are distinct for nonsymmetric amplitudes. In a complete orthonormal virtual basis, define

A1 = g_mn T_nx       A2 = g_mn T_xn
A3 = g_nm T_nx       A4 = g_nm T_xn
D  = (g_mn d_x) d_n^T
E  = (g_nm d_x) d_n^T

W1(m,x) = (me|xb) + sum_f d_x(f) (me|bf)
          - sum_n (nx|me) d_n(b)
          + sum_n [A1 - A2/2 - A3/2 - D]

W2(m,x) = (me|xb) - (mx|be)
          + sum_f d_x(f) [(me|bf) - (mf|be)]
          - sum_n [(nx|me) - (mx|ne)] d_n(b)
          + sum_n [A1 - A2/2 + A4/2 - A3/2 - D + E]

WX(m,x) = -(mx|be) - sum_f d_x(f) (mf|be)
          + sum_n (mx|ne) d_n(b) + sum_n [A4/2 + E]

Every occupied sum contains all six orbitals, including its diagonal. There is no cell, orbit or additional half weight. The 1/2 coefficients are internal to the spin-adapted intermediates. These are the spatial forms of Stanton et al., Sec. II, Eqs. (2), (8)-(10). The two source orientations must survive until their distinct contractions. The source uses real time-reversal frames. In the underlying Bloch integrals, the first and third coefficient legs conjugate; ket legs and amplitudes do not. Signed momentum transfers, the finite-source supports, omitted G=0, and finite erfc overlap subtraction are inherited unchanged from the authenticated terminal provider.

For target p=(i,j), let S_pq = V_p^T S V_q. A W matrix has its first free leg in source q and its second in target p. It is projected as W_qp = U_q^T W U_p, where U_q maps that pair into the complete virtual frame. The raw isolated half is

G_ij = sum_m S_p,im (T_im - T_mi) W1_[im],p(m,j)
     + sum_m S_p,im T_im W2_[im],p(m,j)
     + sum_m S_p,mj T_mj WX_[mj],p(m,i)
R_ring = G_ij + G_ji^T

The contracted e leg stays in its source pair frame. Only the free amplitude leg is projected first. Complete diagonal singles and source doubles enter the W dressings before either free leg is projected. The metric overlaps follow Riplinger and Neese, Sec. II C-D, Eqs. (26)-(30). The finite signed operator replaces auxiliary factors; this action makes no positive-Gram or general linear-scaling assumption.

Snapshot boundaries and reversal

Native records each Fock half, adds all three ring contractions, and then records that cumulative ring half. Subtract the matching Fock half before comparison. virtual_ladder is a different, cumulative R2 accumulator. It must not be subtracted from either local half.

Reversing a pair exchanges the raw halves without a transpose: G_ij(ji) = G_ji(ij). The assembled ring transposes. For i=j, native records only the ij half and adds that half plus its transpose.

ring_terms returns both isolated halves and their sum. ladders_fock_ring_assembly reuses the qualified Fock assembly unchanged and adds R_ring once. Its independent native reference is exactly

virtual_ladder + half_ij_ring + half_ji_ring.T

Equivalently it is the bare seed, each complete ladder increment, both Fock halves, and both isolated ring halves. Each ladder retains its own qualified half contraction. No earlier contribution is replaced or added twice.

Bounded execution and qualification

One authenticated ovov lease precedes one stream accumulating ooov, oovv, and ovvv blocks. W1/W2/WX and their mixed projections are formed one occupied neighborhood at a time. Source identity, terminal state, pair build, metric and pair frames are checked before and after execution. All 36 oriented lifted doubles, complete singles, retained earlier outputs, complex integral accumulators and copies, W intermediates, projected amplitudes, mixed blocks and target accumulators enter the simultaneous numerical census. The domain and native panel reservations are additional. This fixed finite-fixture bound is not a process-RSS guarantee.

The executable independent oracle constructs covariant AO kernels from analytic finite-supercell integrals and independently reconstructed PAO/PNO frames. It derives each source projection with the AO metric, without production block or projection helpers. Nonzero deterministic singles and seed-586 nonsymmetric doubles are unchanged from the accepted fixture. The absolute component tolerance, declared before runs, is 1e-10 Ha; real reconstruction uses 1e-12.

Each orientation and its sum is checked against both the dense oracle and actual native snapshot differences. Controls cover pair reversal, zero singles, doubles/singles scaling, wrong source orientation, omitted ij/ji, wrong ji transpose, stale source/frame, invalid block identity/issuer/transfer, and exact versus one-below memory/work caps. A separate asymmetric algebra fixture checks all source orientations without physical symmetry masking. CPU time and elapsed wall time are recorded separately.

Finite-fixture verdict

Both isolated ring orientations, their sum, and their specified partial assembly are qualified on this finite fixture. Each comparison covers all 36 ordered targets and 174 components at the predeclared 1e-10 Ha absolute gate. The single signed numerical receipt on #586/#592, linked to #1049, retains actual/reference/signed-error arrays, native cumulative snapshots, loaded-module identities, resource measurements and controls.

Comparison

Dense oracle max error (Ha)

Native snapshot max error (Ha)

Isolated ij

7.52176099e-15

7.47318873e-15

Isolated ji

7.56339436e-15

7.47318873e-15

ij + ji.T

8.82627305e-15

8.74300632e-15

Covered assembly

5.65103520e-14

5.70654635e-14

The physical ij/ji distinction is nonzero: component [3, 2, 1, 0] differs by 5.10063955e-02 Ha. Zero-singles production errors are at most 9.02056208e-15 Ha. The raw-half exchange and summed-transpose reversal errors are exactly zero. The production contraction at controlled doubles/singles scales (0,1), (-1.7,0.6) and (0.6,-1.7) agrees with the dense oracle and native snapshots within 2.27422248e-14 Ha. These scaling controls use independently transformed dense blocks; the unscaled and zero-singles actions exercise the streamed provider. Zero doubles give exactly zero ring contributions. All 14 occupied and 13 virtual earlier array groups, all three Fock groups and the earlier assembly are bit-for-bit unchanged. Raw MP2 remains -0.04074663070969609 Ha/cell.

The physical negative controls exceed the declared tolerance:

Deliberate defect

Separation (Ha)

omitted ij

1.12488235e-01

omitted ji

1.12488235e-01

wrong ji transpose

6.01550235e-02

wrong source orientation

2.84394660e-02

Exact caps pass and one-below caps refuse: 42,991,616 bytes, 463,458,992,128 work units for the two ring leases, and 2,775,204,888,576 work units for the twelve-lease assembly. The conservative simultaneous numerical-array census is 684,544 bytes. The zero-singles action adds two leases to qualification. Stale state, foreign generation/handle, changed pair frame and invalid block identity/issuer/transfer all refuse. Peak process RSS was 223.203 MiB, separate from the numerical reservation.

The complete job consumed 748.802 CPU seconds. The scheduler resource receipt reports 18586.691 elapsed seconds; calendar start-to-finish time was 39044.330 seconds. Stage timings use a monotonic clock and are retained separately. Effective scheduling remained at Utility QoS under contention; these timings establish no speed claim.

The matched references are the independent analytic finite AO operator and actual shared-native snapshots. ORCA/CRYSTAL do not supply a matched isolated tensor for this finite operator. No complete R2, R1, iterative CCSD, bulk convergence or triples claim follows from this increment.

Remaining half updates

The exact next half updates are

- sum_me t_i(e) t_m(a) (me|jb)
- sum_me t_j(e) t_m(a) (mi|be)
+ sum_e  t_i(e) (ae|jb)
- sum_m  t_m(a) (mi|jb)

and their ji counterpart. They belong to half_ij - half_ij_ring and remain separate under #586/#592 toward #1049.