The authors released CohAlign, a design-time toolkit that predicts how quickly measured parameter gradients decay under noise in brickwork architectures that conserve a global charge. The computation needs only a classical description of the circuit, the chosen noise model, the prepared state, the readout operator and the initialization prior. No finite-noise circuit sweep is required. The predicted quantity is deliberately distinct from training convergence or final accuracy; the toolkit addresses the first-order decay of the ensemble second moment of a measured gradient.
The central observation driving the toolkit is concrete. Only the part of the channel’s action that the measurement can access affects the measured gradient. A constructed control channel in the study damages the operator carrying the gradient at nearly the worst possible rate while leaving the measured gradient exactly unchanged because the contracted directions are invisible to the readout. The software turns this intuition into computable diagnostics.
The paper studies a brickwork quantum neural network on a ring of n qubits, with an open-chain variant available for validation. Each layer applies trainable phase rotations on every site followed by fixed hopping gates on alternating edges. Every gate commutes with the total charge, so dynamics remain within a fixed excitation sector. The readout is the sector-projected observable with projection onto the r-excitation subspace; inputs in the single-excitation case are the basis state with the excitation at site zero. Trainable phases are sampled from a small-angle box to avoid noiseless barren plateaus; hopping angles are drawn once with a fixed seed and frozen for experiments.
Noise is injected after every layer via a per-layer channel at probe strength epsilon, with explicit conventions linking epsilon to elementary error probabilities. The toolkit’s reference set includes four single-qubit Kraus-pair channels that act identically on sites: amplitude damping, dephasing, depolarising and X-error (bit flip), collectively termed the restricted isotropic family because their restricted generators contract every direction of the sector coherence block at a common rate. Variants introduce structure: inhomogeneous dephasing and site-dependent amplitude damping scale site probabilities by a linear profile; biased Pauli noise breaks phase covariance by uneven Pauli weights. Two strict controls are provided: a coherent-dissipative channel (uniform Z rotation followed by amplitude damping) that acts as a null control within a sector, and a correlated dephasing control that preserves coherence within paired edges while rapidly removing coherence across pairs.
The toolkit restricts the Markovian generator to the off-diagonal block inside the chosen sector—the occupation-basis coherences—because the readout and the channels studied are diagonal-friendly in that basis. Three progressively refined quantities are computed from this restricted generator. First, the worst-case sector coherence rate is the variational supremum of the decay Rayleigh quotient over coherence directions and reports the fastest contraction any coherence direction could experience. Second, the operator mode contraction rate evaluates how rapidly the channel contracts the particular operator mode that carries the derivative of the readout with respect to a chosen parameter; it averages the decay quotient over the ensemble of noise slots encountered after the parameter insertion. Third, the response susceptibility (described below) links these operator-level effects to the measurable gradient decay.
The worst-case rate is finer than scalar whole-channel summaries because it resolves the channel’s action on the specific coherences visible to the measurement. For the reference isotropic channels the finite-probe values converge linearly to analytic limits as the probe is reduced; the shipped working probe is 10^-3 and Table 2 in the source documents the probe sensitivity. The toolkit therefore reports relative tolerances at the working probe and certifies isotropy by spectral spread of the Hermitian part of the restricted generator.
Rates describe operator contraction, but training depends on the measured gradient’s decay. CohAlign computes the first-order response of the derivative of the measured cost to an infinitesimal generator insertion at each noise slot and combines these responses with the noiseless derivative to form a cross-moment estimator. Concretely, expanding the noisy derivative to first order yields a sum over slots of a response term; the degradation of the ensemble second moment of the gradient is governed by the cross moment between the noiseless derivative and its first-order response.
The toolkit evaluates these closed-form response terms within the finite-probe approximation. For slots at or after the insertion layer the response involves the derivative operator propagated forward and the readout evolved backward; for earlier slots the adjoint response functional with the noiseless state appears. The estimator accumulates numerator and denominator of the susceptibility across draws rather than computing per-draw ratios, which keeps the estimator defined when individual draws have vanishing derivatives. The study used an estimator ensemble of forty draws and a separate measurement ensemble of thirty draws, each with recorded seeds and a numerical floor on squared derivatives to stabilise the measurement side.
Validation proceeds as a ladder of increasing difficulty. The sequence begins with exact analytic recovery and the engineered null control whose gradient is provably immune to its noise, continues through direct finite-difference validation of the susceptibility across parameter locations and a second circuit topology, then extends to larger system sizes and varied settings. The ladder concludes with a comparison across nine noise models and a study in the two-excitation sector, where scalar descriptions are least likely to suffice. The toolkit was validated in exact simulations at eight and ten qubits and on two circuit topologies; the released pipeline includes the notebook, module sources, per-phase outputs, figures and full run metadata with every seed and package version on GitHub.
The audit additionally demonstrates behavior under a pure coherent perturbation (a site-dependent Z over-rotation that is not proportional to the conserved charge). For such generators the worst-case rate can be numerically zero and the response becomes signed; the software reports signed values with explicit flags rather than normalised ratios.
The authors note practical constraints that shape the audit. CohAlign is a classical design-time calculation and requires a classical noiseless simulation of the circuit at the system sizes studied; this requirement limits immediate scalability to much larger systems unless classical simulation remains tractable. The finite-probe approximation introduces probe-dependent biases, which the paper quantifies and reports as relative tolerances at the working probe (the shipped probe is 10^-3). Symmetry-breaking channels can produce sector leakage that is discarded by the projected construction, and that leakage contributes to the finite-difference bias discussed in the methods.
All code, data and reproduction materials are publicly available in the project repository at https://github.com/ugail/CohAlign under the MIT licence, with deterministic generation of every dataset and recorded seeds and package versions. The paper documents these limitations explicitly and frames them as part of the reproducible design-time audit pipeline rather than as opaque, unquantified uncertainty.