New technical note AGE-PN-001: reproducible benchmark for WDM stability See evidence
TM-03 · Lab simulator

Calibrating the minimum-norm receiver

How much must you measure for C2⋆ to be real? This bench calibrates the manuscript's maximum-capture receiver with a population-only protocol over K templates, with shot and phase noise, and pinpoints why it fails when it fails. Every figure cites its origin.

Worst-mode capture vs K

Noiseless, with noise (median and 16–84% band), and the exact optimum of Theorem 1. K = 3 reproduces C2.

Fraction of each mode outside the basis

Leakage ε_K = 1 − ‖P_Kφ‖². It is the capture lost by truncating the basis; it causes no crosstalk.

Selectivity vs shots

Median and 16–84% band over 40 realisations. Unit slope is the law Σ ∝ M.

Realisability of the most demanding filter

Normalised profile of the calibrated filter against the absorption limit ρ_B F. Red marks where the control does not exist.

Diagnosis: what fails, where and why

    Why it works

    Six results, each with its level of evidence

    DEM = proved in the manuscript; VER = verified numerically on this bench; CONJ = reasonable conjecture awaiting proof.

    1

    Capture of a normalised dualDEM

    ⟨h̃m|φn⟩ = δmn  ⟹  ηm = 1 / ‖h̃m‖²

    Any filter that exactly nulls the other modes pays its excess norm as lost capture. The problem is therefore choosing the minimum-norm dual.

    2

    Theorem 1: the optimum lives in span{φ}DEM

    ηm* = 1 / (G−1)mm,   Gkn = ⟨φk|φn⟩
    Proof idea

    A component orthogonal to span{φ} changes no overlap and spends norm. Inside the span, nulling the other modes forces h ∝ Qφm, with Q the projector orthogonal to them; ‖Qφm‖² = 1/(G−1)mm by the Schur complement.

    3

    Calibrating with K templates is projectingVER

    φ̂n = PK φn = Σj<K Sjn fj,   Ĝ = S†S  ⟹  ηm(K) ≤ ηm*

    The calibrated receiver is the C2⋆ of the projected modes. With K = 3 and span{φ̂} = span{f}, the dual is unique and equals C2. As K grows, the leakage εK = 1 − ‖PKφ‖² falls and η(K) climbs to η*.

    Why crosstalk is zero without noise

    The filter lies in span{f0…fK−1}, so ⟨h|φn⟩ = ⟨h|PKφn⟩: the part of φ outside the basis is invisible. Truncation therefore costs capture, not crosstalk.

    4

    Gauge invarianceDEM

    φn → eiαnφn  ⟹  G → D†GD,   (G−1)mm unchanged

    With D diagonal and unitary, the diagonal of G−1 is unchanged and each filter only picks up a global phase. The bench checks this with random phases: the difference is of order 10−16.

    5

    Noise law: Σ ∝ MVER

    δĜ ≈ 2 Re(S†δS),   ‖δ(Ĝ−1)‖ ≲ ‖G−1‖² ‖δĜ‖  ⟹  Pm≠n ∝ σq² ∝ 1/M

    Estimation error breaks duality at first order in δS, and crosstalk is quadratic in that error. With a well-conditioned G, Σ ≈ 2M in this model. An ill-conditioned G amplifies the error by cond(G).

    6

    The calibration templates must be absorbable tooCONJ

    ρB ≥ maxj,k,χ ρcrit[(fj + eiχfk)/√2] ≈ 1,061

    The population protocol absorbs template superpositions. The most demanding in the family is (f₀ + f₁)/√2, because of the same time asymmetry studied in bench TM-01. That this protocol suffices to recover G without quadrature access is the conjecture this bench supports but does not prove.

    Tests of the core and of the page's claims

    Auditor's findings

    What this bench does not model

    • The link is chiral with no delay; ETH uses a 30 m bidirectional link where the rejected mode returns after ≈ 292 ns.
    • The envelope distortion is a causal filter (proxy), not the manuscript's a = 0.150 miscalibration.
    • No emitter control noise, qubit decoherence, or drift between calibration and operation.
    • Calibration noise is a first-order model; the real population protocol has its own estimator and biases.
    • Templates f₃ … f₇ extend the ETH family by Gram–Schmidt of τk sech(τ/2); ETH uses only f₀ … f₂.
    • Calibration time uses an assumed repetition rate, not an ETH figure.
    • The manuscript’s Σ ≈ 3.6×10³ reference is a favourable noise realisation: over 40 seeds the median is ≈ 2.1×10³ (bench TM-05).

    Reproducible, sealed record

    Parameters, seed and results for the current configuration, with its SHA-256 fingerprint to pre-register a prediction.