Whitening with a threshold per transfer¶
AICCM/CCM remains experimental. VibeQCAICCM/WhiteningTransfer.lean adds 17
claims. They extend the whitening statements of Whitening.lean
(derivation) from one metric threshold shared by all
momentum transfers to a threshold per transfer.
Why¶
The range-separated fit is the three-dimensional default of the program and
the fit the papers report (author decision of 2026-11-14). It decides its
metric rule per transfer: every mode is kept when the Cholesky factorisation
succeeds and the smallest eigenvalue exceeds the rounding floor, and
otherwise the strict cut lambda > tau applies
(aux_basis.py::_rsgdf_select_metric_decomposition, _rsgdf_metric_keep).
The earlier claims assume one tau for all transfers and certify the
classic fit only.
Statements¶
p is a transfer and m its conjugate partner. “pair” is the unchanged
partner hypothesis of cut_folded_tensor.
Claim |
Statement |
|---|---|
|
The folded-tensor identity with |
|
The same under the weakest condition: partners retain the same modes. |
|
With a constant threshold this is |
|
A threshold below every eigenvalue keeps every mode, and the Gram is the full inverse. This is the Cholesky branch. |
|
The program’s rule: per transfer either the full-rank rows or the strict cut. The folded-tensor identity holds when partners take the same branch. |
|
A square isometric intertwiner makes partner spectra coincide, so partners need not share an artificial threshold. |
|
Per-frame versions of the earlier transport statements. |
What is not covered¶
That the program takes the same branch at a transfer and at its conjugate partner. That is the hypothesis
share. The branch is a floating-point decision.The short-range and long-range integral evaluation of the range-separated fit, and its cutoff planner.
Stability of the retained set near a threshold.
Check¶
python formal/check_proofs.py with the pinned Lake.