Erasure Flags as Free Information for Noisy Quantum Gradients

Training a parameterized quantum circuit means estimating gradients from a finite number of measurements ("shots"). Two kinds of error get mixed together. The first is statistical: shot noise shrinks as you take more shots. The second is systematic: hardware decay bends the energy landscape, and no number of shots removes that. Current methods for reducing gradient variance handle only the first kind. The best of them, a Bayesian version of the standard parameter-shift gradient rule, explicitly assumes noiseless hardware.

This note proposes encoding each logical qubit in a decoherence-free or "dual-rail" form. In a dual-rail encoding, one excitation is shared between two physical modes, and the dominant decay process knocks the state into a detectable "erased" outcome. Physical decay then stops being a hidden bias and becomes a per-shot flag that the gradient estimator can read. The central question is when spending shots on flagged-and-discarded outcomes beats spending them on biased unencoded measurements. We aim to give a sharp, testable answer, and to show how a variance-aware estimator should use the flags.

Core Claim

Take a dual-rail code with $|0_L\rangle = |01\rangle$ and $|1_L\rangle = |10\rangle$. A single decay maps either code state to $|00\rangle$, which lies outside the code. Multi-Error-Correcting Amplitude Damping Codes analyzes this mapping, and Logical multi-qubit entanglement with dual-rail superconducting qubits demonstrates it on transmons.

If both rails decay at the same rate, the evolution conditioned on "no erasure" acts on the code space as a constant times the identity. Discarding flagged shots therefore yields estimates whose mean is the noiseless value. The cost is that fewer shots survive, so the variance grows. Unencoded qubits behave the opposite way: they keep every shot but carry a bias $b$ that does not shrink with shots.

For a single gradient component, write $\sigma^2$ for the single-shot variance, $N$ for total shots, $p$ for the erasure probability, and $b_r$ for the residual bias the code does not remove (for example, independent dephasing or unequal rail decay rates). The mean-squared errors are

$$\mathrm{MSE}_{\text{unenc}} = b^2 + \frac{\sigma^2}{N}, \qquad \mathrm{MSE}_{\text{enc}} = b_r^2 + \frac{\sigma^2}{(1-p)N}.$$

The encoded estimator wins once $N$ exceeds roughly $N^* = \sigma^2 p / \big((1-p)(b^2 - b_r^2)\big)$. This crossover formula follows directly from the two expressions above. It turns "is the encoding worth it for training?" into a quantity that can be measured.

The second, less obvious claim concerns the structure that Bayesian gradient estimators rely on. Bayesian Parameter Shift Rules in Variational Quantum Eigensolvers models the energy as a Gaussian process with a trigonometric kernel. This model remains valid under angle-independent noise; noise only shrinks the amplitude and moves the minimum. It also remains valid for postselected dual-rail estimates when the rails decay symmetrically.

When decay rates are asymmetric, the survival probability depends on the circuit parameters. The postselected energy then becomes a ratio of trigonometric functions, and the kernel is misspecified. In this way the known weakness of decoherence-free codes, imperfect symmetry, becomes a concrete failure of the variance-reduction machinery. We propose to fix it by modeling the erasure counts themselves.

Research Program

The work has three parts.

Analysis. We will derive the bias and variance of parameter-shift gradients for three cases: unencoded qubits under amplitude damping, phase damping and depolarizing channels; dual-rail encodings under the same physical channels plus residual independent dephasing; and a four-qubit collective-dephasing code of the kind built in Development of Quantum Private Queries Protocol on Collective-Dephasing Noise Channel under correlated dephasing. This gives $b$, $b_r$, $p$ and hence $N^*$ as functions of the noise rate and circuit depth. The analysis directly addresses the open interaction flagged in the file between Evaluating the noise resilience of variational quantum algorithms, which studied decoherence with exact density matrices and zero shot noise, and the Bayesian parameter-shift work, which studied shot noise with zero decoherence.

Estimator. We will build an erasure-aware extension of the GradCoRe shot-allocation strategy from the Bayesian parameter-shift paper. Each shifted circuit gets its own noise variance, set from its observed survival fraction. Shot allocation then favors parameters whose shifted circuits leak less per unit of information gained.

For asymmetric rails, we add a second, jointly modeled output: the per-circuit erasure rate. It is itself a trigonometric function of the parameters, and the energy model becomes a ratio of the two. Erasure counts are free, since every shot reports them. They therefore also act as a control variate for the first-order bias caused by asymmetry. This plays a role analogous to how Sticking the Landing removes a gradient term whose structure is known in advance. The analogy is only a motivation: here the known term is the flagged decay, not a zero-mean score term.

Simulation study. We will run a controlled study on the benchmarks used by both lines of work, with a stretch goal of a dual-rail hardware run.

Baselines

Naive reference. Unencoded fixed-shot parameter-shift stochastic gradient descent on physical qubits, under the same noise model. It shows the bias floor in its rawest form.

Strongest existing method. GradCoRe from Bayesian Parameter Shift Rules in Variational Quantum Eigensolvers, run on unencoded qubits with its Gaussian-process noise term tuned to absorb hardware noise. It is the strongest competitor because it already reuses past measurements and allocates shots adaptively. Any gain we report must come from the flags, not from adaptive allocation itself. That paper reports reaching lower energies with "an order of magnitude fewer cumulative shots than standard fixed-shot SGD" on 1D Ising and Heisenberg chains up to 7 qubits, verified with Wilcoxon tests over 100 trials. The file reports no absolute energy-error values, and no reported result was found on Lacuna for GradCoRe under hardware noise.

We will rerun GradCoRe on the same Hamiltonians, the same noise rates and the same total shot budget. Erased shots count against the encoded arms. Each arm gets the same hyperparameter search of 20 configurations. Beating GradCoRe means a statistically significant lower final energy error at equal cumulative shots for budgets above the predicted $N^*$, across 100 seeds. It also means a plateau level that keeps falling with budget while GradCoRe's plateau stays flat.

Classical optimizer. Adam on unencoded qubits. Benchmarking VQE Configurations reports Adam reaching 4.06% relative error with the UCCSD ansatz on silicon, where plain gradient descent and SPSA stalled.

Ablations. We isolate the key ingredient with three ablations, all on the same dual-rail circuits:

  • Flags ignored. Dual-rail with erased shots mapped to a fixed outcome rather than discarded.
  • Plain postselection. Postselected dual-rail with standard homoscedastic GradCoRe.
  • No flags at all. A particle-conserving ansatz on physical qubits, analogous to the PCU2 ansatz in the silicon benchmark (about 7.45% relative error there). It keeps the symmetry restriction but has no erasure flag.

Why not just…?

Why not just use Adam or another adaptive optimizer on unencoded hardware? Adaptive step sizes rescale noise but cannot remove a bias that does not average out. Evaluating the noise resilience of variational quantum algorithms shows that above a noise threshold, reoptimization becomes chaotic and converges to physically wrong states. QMLP shows that QuantumFlow collapses from 69% to about 10% accuracy under device-calibrated noise. The Adam baseline tests this directly. If Adam's plateau falls with budget like the encoded arms, the bias argument is wrong.

Why not just take more shots? Extra shots fix only the variance term. The analysis predicts that unencoded error flattens at $b^2$. The experiment plots error against budget over two orders of magnitude to check whether that floor exists at realistic noise rates. This alternative is a genuine threat. If $N^*$ exceeds practical budgets, the honest conclusion is that encoding does not pay for training. We will report that conclusion if it holds.

Why not just give GradCoRe a larger noise term on unencoded qubits? A larger noise variance widens uncertainty but leaves the posterior mean centered on the biased landscape. Only an estimator that can tell decayed shots from good ones removes the bias. The comparison between the plain-postselection arm and tuned unencoded GradCoRe settles this.

Why not just use a symmetry-preserving ansatz without encoding? Restricting to a symmetric subspace changes which directions the optimizer can explore. It may also shrink the circuit's Lie algebra, a property that Hadamard Test is Sufficient for Efficient Quantum Gradient Estimation exploits for shot savings. But decay inside a physical particle-conserving circuit remains silent. The no-flags ablation separates the symmetry effect from the flag effect.

Proposed Experiments

Benchmarks. The main benchmarks are 1D transverse-field Ising and Heisenberg chains with 4 to 7 logical qubits, matching the Bayesian parameter-shift study. They are simulated with quantum trajectories. Because erasures are flagged, each trajectory either stays in the code space or is labeled and dropped.

Noise sweeps. Amplitude-damping rates span the range from the linear regime to the chaotic regime identified in the noise-resilience paper. Residual independent dephasing is swept from zero up to the damping rate. The rail asymmetry ratio $\gamma_a/\gamma_b$ is swept from 1 to 3. Correlated dephasing is run on the collective-dephasing code. Haar-random state preparation provides a second task that does not depend on a particular Hamiltonian.

Metrics. We measure:

  • final energy error and fidelity against cumulative shots, with all shots counted;
  • gradient mean-squared error against the exact noiseless gradient, split into bias and variance;
  • measured $N^*$ compared with the analytic prediction;
  • physical-qubit and gate overhead, reported separately so the doubled qubit count is never hidden.

Decisive checks. The project's claims fail if any of the following hold:

  • Unencoded GradCoRe shows no bias floor at realistic noise.
  • The measured crossover misses the predicted $N^*$ by more than a factor of two.
  • Erasure-aware GradCoRe does not beat plain postselection once rail asymmetry exceeds about 1.5. At that point the trigonometric kernel should visibly misfit, so a missing advantage there would refute the misspecification claim.

Hardware stretch goal. If a dual-rail transmon device like that in Logical multi-qubit entanglement with dual-rail superconducting qubits is accessible, a two-logical-qubit Ising run would test whether measured erasure rates behave as the model assumes. The file reports no hardware fidelity numbers to compare against.