Real Gamma PAO to PNO projector contract¶
AICCM/CCM remains experimental. This bounded bridge covers the supplied
finite real matrices used by semicanonical_pao_basis and
_build_ccm_pno_pairs through _finish_pno_pair. It adds 20 exact Lean
claims and seven production-path fixture cases. It changes no numerical
algorithm, threshold, domain, solver, resource policy or runtime dependency.
The existing Gamma bare-density and cutoff tests remain part of the gate.
Inputs and exact statement¶
Let the AO metric S be symmetric positive definite, the occupied coefficients C satisfy C^T S C = I, and the AO Fock matrix F be symmetric. Define
Q = I - C C^T S
P = Q[:, domain]
G = P^T S P
G E = E Lambda, E^T E = I
X = P E Lambda^(-1/2)
V = X U, U^T U = U U^T = I
W = V d R, R^T R = R R^T = I
Pi(Z) = Z Z^T S
E contains only supplied metric eigenvectors with eigenvalues strictly
greater than the nonnegative lindep threshold. U diagonalizes the Fock
restriction to X; d contains selected columns of the supplied pair-density
eigenframe; R diagonalizes the Fock restriction to V d. These are conditional
statements about supplied decompositions. The proofs do not establish an
eigensolver’s existence, implementation or floating-point accuracy. Individual
algebraic lemmas use only the hypotheses they need; positive retained values,
for example, suffice for inverse-square-root whitening.
Q is idempotent and self-adjoint in S, and C^T S Q = 0. Whitening gives X^T S X = I and preserves exactly the span of P E: multiplying X by Lambda^(1/2) reconstructs P E. The later column selections and orthogonal rotations preserve occupied orthogonality and metric normalization. The resulting Pi(W) is an S-self-adjoint idempotent whose range is the selected PNO space. In particular,
Pi(W) = V (d d^T) V^T S
K_final = R^T (d^T K_PAO d) R
The final Fock rotation changes coordinates within the selected span. At
zero PNO cutoff d is a complete eigenframe, so Pi(W) = Pi(V). This equality
refers to the supplied PAOs after domain selection and metric filtering.
It cannot restore an excluded AO-domain direction or one removed by
lindep; it alone is not a full-virtual canonical-limit claim.
Claims and production correspondence¶
All theorem names below have prefix VibeQCAICCM. and live in
GammaPAO.lean.
They are registered in the existing claims inventory and transitive audit.
Claims |
Exact obligation |
Production boundary |
|---|---|---|
|
Occupied complement and S-adjoint projector |
|
|
Any supplied domain remains occupied-orthogonal |
Full AO rows of Q, selected domain columns |
|
Positive retained metric normalization and reconstruction |
|
|
Isometric selections/rotations preserve metric and occupied orthogonality; supplied Fock eigenpairs yield a diagonal restriction |
Both PAO and final PNO rotations |
|
Distinct eigenvector columns are isometric; their product equals the diagonal selection mask in the eigenframe |
|
|
The normalized frame defines its S-projector |
AO-space interpretation of the returned pair V |
|
Selected AO projector, invariance under final rotation, zero-cutoff identity on supplied PAOs |
|
|
Pair-integral congruence under composed transformations |
|
The local PAO/PNO construction follows the finite-domain setting discussed
by Riplinger and Neese, JCP 138, 034106 (2013),
doi:10.1063/1.4773581, Sec. II B.
The Gamma implementation uses its explicitly documented bare density
(T T^T + T^T T)/(1 + delta_ij), not the paper’s spin-adapted density norm.
This tranche assumes the selected column set. Existing gamma_pno tests
pin positive strict cutoff, leading-direction fallback and zero-cutoff
retention, including null directions. No density threshold is transferred
between conventions and no threshold-to-energy bound is asserted.
Focused execution and scope¶
Using a qualified native module and the checkout’s Python sources, run:
python -m pytest tests/test_ccm_dlpno_mp2.py -k 'gamma_pno or gamma_pao' -q
python formal/check_proofs.py
The Lean command requires the pinned toolchain and existing dependency cache as described in the formal README. Keep runtime identities, logs and machine paths in private operational evidence.
The new tests observe the actual production eigensolves, PAO return and
PNO finish boundary. Only synthetic atom mapping/domain inputs and
observation wrappers are substituted; NumPy eigensolvers, projections,
whitening, contractions, density selection and rotations execute normally.
Fixtures use six AOs, two occupieds, an omitted fourth global virtual,
redundant projected columns, a nonorthogonal AO metric, a degenerate Fock
restriction, and zero, positive and all-discard PNO cutoffs. A diagonal
metric fixture places lindep just below, exactly at and just above 0.25;
an all-discard PAO case checks the existing empty-pair behavior.
Comparisons use absolute tolerance 1e-12 and zero relative tolerance. Projector comparisons permit eigenvector signs and rotations in degenerate spaces. Deliberately removing S from Q or the inverse square root from whitening violates the corresponding contract by more than 0.1. The gate checks actual final V and K, both metric/projector identities, eigenpair residuals, selected spectral masks and the excluded global direction.
The focused run passed all 24 cases (17 existing density cases and seven new projector cases); the largest new residual was 1.67e-15. The positive 0.02 and fallback fixtures retained one PNO per pair; zero cutoff retained three or two according to PAO filtering. An all-discard PAO domain emitted no pair. These are small-fixture measured checks, not a proof of NumPy, native integrals, solver convergence, radius convergence or production energy accuracy. The 5-cubed calculation program, cache and amplitudes are outside this task. No new chemistry calculation or scientific acceptance is claimed from these proof/test additions.