Auditing the Rare Bad Cycle: Tail-Risk Estimation for Small Quantum Engines from Two-Point Measurement Data

Small quantum machines, such as a single driven spin, a resonator, or a two-qubit gate, do not deliver the same amount of energy every run. Their run-to-run energy statistics are lopsided and heavy-tailed, so the average and the spread hide how often a cycle fails badly. The standard way to measure those statistics on hardware is the two-point measurement protocol. It measures the system's energy before a driven operation and again after it, and records the difference as the work done in that run. In practice, though, the protocol is used to check textbook equalities on average, not to answer the engineer's question: how often will this device give me less than I need, and is running it faster worth that risk?

We propose a statistically calibrated way to estimate per-cycle failure risk from two-point measurement data. It corrects for the two ways the measurement pipeline itself distorts the rare-event tail, readout error and post-selection, and attaches confidence bounds to the risk estimates. On top of this estimator we define a "marginal risk per unit power" curve that compares a fast protocol against a slow baseline, borrowing the marginal-risk idea from AI policy. The central falsifiable question is whether tail-aware risk picks different operating points than mean–variance criteria. If it does not, the whole apparatus is unnecessary, and we say so up front.

Core Claim

In strongly driven small systems, the work distribution develops large skewness and kurtosis. Heat, work, and fluctuations in a driven quantum resonator computes the first four cumulants of photon exchange using full counting statistics, a technique that tracks the complete distribution of energy quanta exchanged with a bath. It shows that higher cumulants grow sharply outside linear response. The paper only diagnoses this, using continuous counting rather than two-point measurement. It never asks what the skew does to failure rates.

Work extraction from quantum systems with bounded fluctuations in work motivates our concern. It proves that ignoring fluctuations makes average free-energy bounds operationally misleading, and it defines "c-bounded work," a hard cap on deviations. That treatment is deterministic and worst-case, and the paper itself lists probabilistic failure as an open case. It also offers no measurement procedure.

Our claim is that the operationally relevant quantity is a lower-tail risk of extracted work, and that it can be estimated from realistic two-point measurement datasets only if two distortions are corrected. The first is readout error. The second is post-selection: in Experimental test of the Crooks fluctuation theorem in a single nuclear spin, only about 5.4% of full sequences survived, with 16,000 trials per data point. Both distortions act most strongly on exactly the rare outcomes that dominate risk.

Concretely, let $W_{\rm out}$ be the work extracted in one cycle. We use the conditional value-at-risk

$$\mathrm{CVaR}_\alpha = \mathbb{E}\left[W_{\rm out} \mid W_{\rm out} \le q_\alpha\right],$$

where $q_\alpha$ is the $\alpha$-quantile of $W_{\rm out}$, so the quantity is the average output over the worst $\alpha$ fraction of cycles. We also use the failure probability $P(W_{\rm out}<w^*)$ for a required output $w^*$. The marginal risk of a fast protocol $f$ relative to a slow baseline $s$ is the ratio

$$\rho = \frac{\mathrm{CVaR}_\alpha(s)-\mathrm{CVaR}_\alpha(f)}{\bar P(f)-\bar P(s)},$$

where $\bar P$ is mean power. This is tail loss bought per unit of extra average power. The ratio makes concrete the rubric of Position: On the Societal Impact of Open Foundation Models, which argues that risk should be measured relative to an existing baseline rather than in absolute terms. That paper offers no quantitative weighting formula. Here the physics supplies one.

Research Program

The estimator has three parts.

Readout correction. Projective energy readout is modeled as a confusion matrix between true and recorded energy levels, calibrated on separately prepared eigenstates. The recorded joint (initial, final) histogram is inverted subject to a positivity constraint, rather than treating readout error as a fixed calibration offset.

Post-selection correction. The acceptance probability is modeled as a function of the true energy pair. This dependence is plausible, because state-dependent spin flips during readout, which the NV experiment suppressed with a roughly 7500 G field, will not affect all levels equally. The function is estimated from calibration runs, and accepted samples are reweighted by its inverse. Rejected runs therefore become a quantified bias rather than silently disappearing.

Tail extrapolation. The empirical tail is too sparse to measure directly. An event with probability $10^{-3}$ appears roughly once per thousand accepted runs, and the NV acceptance rate means thousands of attempts are needed for each accepted run. We therefore fit the measured cumulant generating function $\ln\langle e^{-\lambda W}\rangle$ over the range of $\lambda$ where it is statistically reliable. From that fit we obtain Chernoff-type upper bounds on $P(W_{\rm out}<w^*)$ that remain valid where no events were observed, and we bootstrap confidence intervals over the whole pipeline.

Measurement as mitigation risk. Towards Concrete and Connected AI Risk Assessment (C2AIRA) finds that across 16 frameworks the risk introduced by the measuring or mitigating mechanism itself is the most neglected. The physical analogue is that the first projective measurement erases initial coherence and can inject energy. Experimental characterization of the energetics of quantum logic gates links entropy production to coherence, and Measurement-driven single temperature engine shows that measurement back-action can destabilize engine cycles.

We make this auditable at the level of the mean. The unmeasured mean energy change $\langle H(\tau)\rangle-\langle H(0)\rangle$ can be obtained from separate expectation-value runs with no first projection. The gap between it and the two-point mean is reported as a "measurement-induced risk" term alongside the tail estimate. For states that are diagonal in the energy basis this term is zero, which gives a built-in sanity check.

Baselines

The do-nothing reference ranks operating points by mean power or mean efficiency alone. This is how devices are currently benchmarked, for example the efficiency-focused evaluation in Capturing dynamics and thermodynamics of a three-level quantum heat engine via programmable quantum circuits, which reports 65.39% efficiency for its optimized cycle against a theoretical 65.34%.

The Gaussian risk baseline computes value-at-risk and CVaR from the measured mean and variance only. It uses exactly the same corrected samples, so any difference comes from tail shape alone.

The worst-case baseline applies the c-bounded criterion of Richens and Masanes, implemented as a cap on the observed range of work at matched sample size.

The strongest direct competitor is the empirical CVaR computed straight from raw post-selected two-point samples. This is what a careful experimentalist would do today. It is given the same number of total trials and the same calibration budget as our estimator.

Two ablations isolate the key ingredients. One removes the post-selection reweighting. The other removes the cumulant-generating-function extrapolation, leaving only the empirical tail.

No tail-risk result for any of these platforms was found on Lacuna. The only reported hardware figures we can anchor to are readout fidelity 0.98(1), sequence success of about 5.4%, and 16,000 trials per point, all from the NV Crooks experiment.

Our estimator beats the strongest competitor if two conditions hold together. Its confidence intervals must cover the ground-truth failure probability at nominal rate. They must also be at least several-fold narrower than empirical CVaR's at $\alpha\le 10^{-2}$ under the NV-matched acceptance rate.

Why not just…?

Why not just use mean and variance? Gaussian metrics are cheap, and if the work distribution is close to Gaussian they suffice. That is exactly what happens in the denser carbon nanotube case in Nonlinear Constitutive Models for Nano-Scale Heat Conduction, where per-block statistics become "more Gaussian." This is a real threat in weak-driving regimes. We test it directly by sweeping drive amplitude and recording where the Gaussian baseline's operating-point choice diverges from the tail criterion. If it never diverges on our benchmarks, we report that Gaussian risk is adequate.

Why not just compute the distribution from the model with full counting statistics? When the model is exact, full counting statistics gives the true distribution, and we use it as ground truth for exactly that reason. But risk audits matter on hardware, where readout error, post-selection, and unmodeled noise are the problem. The three-level engine circuits produced unphysical cold-to-hot heat flow from hardware noise, which no clean model predicts. The audit must work from measured data.

Why not just collect more trials? Brute force works in principle. At roughly 5.4% acceptance, though, resolving a $10^{-3}$ tail with useful precision costs orders of magnitude more runs than the 16,000 per point reported. More data also does nothing about acceptance that depends on the energy outcome, which is a bias rather than variance. The no-reweighting ablation shows whether that bias is material.

Proposed Experiments

Experiment 1: driven resonator. We start with the driven resonator under square-wave, sawtooth, and harmonic drives, simulated with a tilted Lindblad equation that supplies exact distributions. We generate synthetic two-point datasets with NV-matched readout fidelity, energy-dependent acceptance, and trial counts. We then compare every baseline on interval coverage, interval width, and selected operating point, across drive amplitudes spanning linear response to strongly non-Gaussian.

Experiment 2: qubit Crooks protocol and three-level engine. We repeat the comparison on a qubit Crooks-style protocol with switching times of 25–300 μs and on the three-level engine circuits in Qiskit. We report $\rho$ curves alongside the paper's efficiency figures.

Experiment 3: coherent initial states. On the two-qubit controlled-unitary model, we vary initial coherence and check that the measurement-induced risk term tracks coherence and vanishes for diagonal states.

Decisive checks. Three outcomes would settle the question. Does the tail criterion choose a different drive speed than mean–variance in the non-Gaussian regime? Do our intervals keep nominal coverage where raw empirical CVaR fails? Does removing the post-selection correction measurably bias the failure probability under energy-dependent acceptance? A negative answer on the first check confines the method to a statistical tool for rare-event estimation and removes its claim to change engineering decisions.