Skip to content

Latest commit

 

History

History
 
 

m1_leakage

Full-interface leakage and general-noise source refinement

This package removes code preservation from the quantum-noise assumptions of the declared source. Its existing complete failure/reset services provide noisy leakage reduction. The proof also permits a fixed nonzero elementary error below the credited threshold: growing concatenation depth replaces the parent's decreasing per-location error, with the additional local drive, inventory and lifetime costs charged. It also composes with the noisy-record compiler: general Markov quantum noise, native leakage, noisy runtime control and a live noisy public archive coexist in one constructed process at fixed finite physical rates. The contract states the scope, evidence requirements and interpretation of these self-reading source patches and their retained records.

Finite evidence includes:

  • Both native proper codes on every full-M6 matrix unit, with public outcomes separated from private Kraus indices, and two complete two-carrier gates on all 1,296 input matrix units each.
  • 26,880 exact integer decoder identities spanning every single-site map from a code qubit to its full six-level interface. Twenty-one complete channels check erasure, dissipative leakage and coherent leakage on every code site.
  • Six continuous full-M6 noise channels, with noncommuting drives and return from leaked states; a common-bath memory example; an analytic witness that amplitude damping cannot contain a positive ideal identity component.
  • A corrected full-carrier accounting effect, a two-site failure witness, a nonvanishing fractional-control-error example, and exact level/resource inequalities for both refinement regimes.
  • A coherent fault-tail identity replacing the archive's probabilistic estimate, its exact rational bounds, and the combined six-level resource and convergence calculation; recursive replay of the computation code, adaptive instruments and 8,100 archive fault cases from that construction.
  • A coherent counterexample to an additive bad-region norm bound, with the required intersection terms and disjoint location ownership checked. The joint-history proof derives its product bound explicitly and counts every physical interval once, including export-to-refresh handoffs.
PYTHONPATH=code python -m m1_leakage.build
PYTHONPATH=code python -m m1_leakage.verify
python -m pytest -q code/m1_leakage

PowerShell: set $env:PYTHONPATH='code' first. The verifier imports no local producer. Full channel matrices are serialized sparsely and checked within 3e-10; source custody is byte-exact, and integer identities are exact. It does not interpret a Bell-state trace distance as a diamond norm: the diamond upper bound follows from the generator estimate in the proof.

The arbitrary-size leakage/general-noise accuracy results are credited Aliferis--Terhal and AGP theorems; the channel-to-dilation estimate is credited Kretschmann--Schlingemann--Werner. The finite controls do not execute an entire universal concatenated machine or measure its threshold. Five fixed levels suffice for the conservative full-Markov route, four for bounded coherent couplings; a fixed subthreshold elementary strength uses O(log log q) levels. Both yield joint decoded/named-history error O(q^-12 log^6 q) and additional accounting error O(q^-8 log^6 q), within the declared source family.

The combined fixed-rate construction uses the noisy-record parent's larger measurement-free computation code, six protection levels and its actual noisy classical archive. A norm fault-path argument handles damping without inventing an ideal-operation probability. Joint error is O(q^-20 log^16 q), accounting error O(q^-16 log^16 q), and complete diagnostic storage-time O(q^6 log^25 q). This is an analytic composition, not a finite simulation of the entire six-level universal machine.

Reliable runtime classical control is retained only in the separate fixed-per-operation-error and common-bath variants, not in that combined fixed-rate result. The static schedule, source/CCG capability and accounting reference remain declared. Extra diagnostic histories and raw counts need not match the ideal experiment. Unknown raw-input encoding, autonomous clocks and physical energy or source selection are not claimed.