New technical note AGE-PN-001: reproducible benchmark for WDM stability See evidence
TM-04 · Lab simulator

Dynamical realisability and control budget

Can the receiver follow the filter the algebra asks for? This bench emits the three ETH modes through the manuscript’s distorted channel, synthesises the control that would absorb each filter and measures the capture and crosstalk that actually remain: minimum resonator width, peak coupling and generator bandwidth.

Rising edge: |w|² versus ρ_B F_w

Log scale. Red: Theorem 2 violation; grey: the F < ε zone where the control is tapered. The line marks τ*, where ρ_crit is set.

Control pulse |g_B(τ)|

In units of Γ. Dashed: ideal control (tapered at ε). Solid: applied control after the peak limit and the generator.

ρ_crit versus floor ε: design or tail?

A flat curve means a physical limit. If it climbs towards 1 as ε drops, it is set by the sech tail, not by the design.

Realised transfer matrix P_mn

By integrating the dynamics with the applied control, not from algebra.

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 and not yet published.

    1

    Theorem 2: the realisable filterDEM

    |w(τ)|² ≤ ρ Fw(τ)  ⟺  (d/dτ) ln Fw ≤ ρ

    A unit-norm filter is realisable by the node if and only if its rising edge grows no faster than the resonator width. It holds for freely programmable complex coupling, no loss and infinite time.

    2

    The control and its radicandDEM

    gB = −i D* eiθ/r,   D = (−ẇ + ρw/2)/√ρ,   r² = Fw − |w|²/ρ

    To absorb w without reflection the resonator must hold b = −w/√ρ, and the qubit stores r² = Fw − |w|²/ρ. If r² turns negative no g achieves it; near zero, g diverges. It is the same radicand as ETH’s Eq. 4, reversed in time.

    3

    Absorbing is emitting backwardsDEM

    ρabs[w] = ρemit[w(−τ)]

    That is why a parity mixture like (f₀ + f₁)/√2 is only absorbed with ρ ≥ 1.062, a limit set at the pulse centre that does not depend on the floor, whereas it is emitted with ρ ≈ 0.93 (at ε = 10⁻⁹). The emission requirement is set in the trailing tail and climbs towards 1 as the floor drops. It is the same f(t) versus f(−t) asymmetry studied in TM-01. Try it with the custom family.

    4

    The tail rules the higher modesVER

    w ∝ P(τ) eτ/2  (τ → −∞)  ⟹  (d/dτ) ln Fw → 1

    The whole sech family rises as eτ/2, so in the deep tail any filter needs ρ → 1. The manuscript’s ρ_crit of 0.96 and 0.88 for C2⋆ modes 1 and 2 are reached at τ ≈ −25 and −31, exactly where F = ε = 10⁻⁹: lowering ε makes them climb towards 1 (0.95 for h₂ at ε = 3×10⁻¹¹). The design-relevant value is the one that does not depend on the floor: 1.033 for C2⋆ mode 0 (τ ≈ −4.8, F ≈ 6×10⁻³) and 1.88 for C3 (τ ≈ −10.9, F ≈ 6×10⁻⁷, stable for every ε ≤ 10⁻⁷).

    5

    The practical cost of breaking the theoremVER

    C3: peak 56 Γ with tapering at ε; crosstalk 6×10⁻⁸ with |g| ≤ Γ

    Tapering the radicand at the floor removes 5.3×10⁻⁴ of probability, as the manuscript says, but blows up the peak. Clipping the peak instead recovers C3’s algebraic metrics. C2⋆’s advantage is not that C3 is impossible. It is more capture (mean 0.9954 versus 0.9925; worst mode 0.9932 versus 0.9905), a smooth control and no compensation step.

    6

    Control hardware budgetVER

    C2⋆: max|g| ≈ 0,53 Γ,   B ≳ 10–15 Γ for Σ ≥ 2×10³ – 4×10³

    In the deterministic model, clipping the peak at 0.45 Γ leaves Σ ≈ 2×10³ and at 0.40 Γ, Σ ≈ 4×10². A first-order generator with B = 5, 10, 15 and 20 Γ gives Σ ≈ 450, 1.9×10³, 4.6×10³ and 8.6×10³. With Γ/2π = 24 MHz, B = 10 Γ is a 240 MHz cut-off and B = 15 Γ is 360 MHz. For reference, the median with control noise in the manuscript scenario is Σ ≈ 2.1×10³ (TM-05); the quoted 3.6×10³ is a favourable realisation. The manuscript lists these limits as not modelled.

    Tests of the core and of the page's claims

    Auditor's findings

    What this bench does not model

    • Chiral, Markovian, zero-delay link, as in the manuscript; ETH uses a bidirectional link.
    • No control noise (σ = 0): selectivities are deterministic and higher than the noisy medians of the manuscript and TM-05.
    • No qubit decay or dephasing, and no thermal occupation.
    • The generator is an ideal first-order filter; a real AWG has sampling, quantisation and nonlinearities.
    • C2⋆ and C3 filters use the exact G; real calibration is studied in TM-03.
    • Below ε ≈ 10⁻¹¹ the finite window τ ∈ [−40, 40] introduces artefacts in ρ_crit; hence the minimum floor of 3×10⁻¹¹.
    • A ρ_crit is called “robust” if it is set where the absorbed probability is at least 100 times the floor; otherwise it is flagged as “set by the floor”.
    • Results 4 to 6 come from this bench, not the manuscript: they must be checked before being cited.

    Reproducible, sealed record

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