When Unmixing Fakes Thermodynamics: Separation Artifacts in Two-Point Work Statistics

Small quantum and nanoscale systems do not exchange energy smoothly. The amount of work done on a single spin or a single resonator varies widely from run to run, and the histogram of these values is lopsided and heavy-tailed rather than bell-shaped. The headline laws of this regime are fluctuation theorems, which relate those histograms to equilibrium free energies. They are tested by measuring a system's energy before and after a drive and subtracting the two readings. That subtraction is not clean. The recorded energy change mixes work from the drive, heat leaked to or from a bath, and spurious jumps caused by misread measurements. Turning that one recorded number into a work distribution is therefore a source-separation problem: recovering individual signals from their sum. Audio machine learning has learned the hard way that separators leave artifacts, such as ghost remnants, bleed between channels, and plausible textures with no basis in the input, and that downstream models absorb these artifacts as if they were real.

This note proposes to study that failure mode where it matters most, in tests of fluctuation theorems. These tests are dominated by rare tail events: the Jarzynski average weights each run by $e^{-\beta W}$, so a few spurious negative-work runs can swing the answer. The core claim is that the error measures used to judge separators, which compare whole histograms, are nearly blind to the tail errors that decide whether a fluctuation theorem looks satisfied or violated. A second claim follows from the first. A separator equipped with a learned physical prior can "confirm" a fluctuation theorem on data that actually violates it. We will build a simulated benchmark with exact ground truth, measure how separation artifacts move into the higher statistics of the work distribution (its skewness, its tail weight, and the fluctuation-theorem residuals), and test a constrained estimator designed to keep its errors out of the tails.

Background and Gap

Three bodies of work set up the problem. The first establishes that energy statistics in this regime are genuinely non-Gaussian and that the deviations carry the physics. Heat, work, and fluctuations in a driven quantum resonator computes full counting statistics, meaning the complete distribution of exchanged quanta, for a frequency-modulated oscillator weakly coupled to a bath. It shows that strong driving produces large third and fourth cumulants (skewness and kurtosis) that the mean and variance miss. In the classical nanoscale setting, Nonlinear Constitutive Models for Nano-Scale Heat Conduction and The Mori-Zwanzig formalism for the derivation of a fluctuating heat conduction model from molecular dynamics show that local block energies follow shifted Gamma distributions. They also show that additive Gaussian noise models permit unphysical negative energies. Both results motivate the proposal: they say the tails and the positivity constraint are where the physics lives. Work extraction from quantum systems with bounded fluctuations in work and Mapping current fluctuations of stochastic pumps to nonequilibrium steady states add that systems identical on average can differ sharply in their rare fluctuations. Any estimator that is good "on average" is therefore suspect.

The second body is experimental two-point measurement (TPM), in which the energy is measured projectively at the start and the end and work is defined as the difference. Experimental test of the Crooks fluctuation theorem in a single nuclear spin and Experimental test of the Jarzynski equality in a single spin-1 system using high-fidelity single-shot readouts reach single-shot readout fidelities of about 0.98 on a diamond nitrogen-vacancy (NV) center. The second paper reports $\langle e^{-\beta W}\rangle = 0.97(1)$ against a theoretical 0.966(3). Experimental characterization of the energetics of quantum logic gates reconstructs five moments of a photonic gate's energy distribution. Energy fluctuation relations and repeated quantum measurements shows that the Jarzynski equality survives intermediate measurements only for unital dynamics, that is, dynamics that leave the maximally mixed state unchanged. That result supplies us with a principled way to construct ground truths that violate the naive equality. The measurement-engine papers (Measurement-driven single temperature engine, Quantum collapse and the second law of thermodynamics) establish that measurement itself injects energy, which counts as a third channel in the recorded difference. None of these papers treats the recorded energy difference as a mixture to be unmixed, or asks how the choice of unmixing method biases higher cumulants.

The third body documents artifact propagation in audio. SingFake: Singing Voice Deepfake Detection and SVDD 2024: The Inaugural Singing Voice Deepfake Detection Challenge show that separating vocals before detection can hide exactly the anomalies a detector needs. MSRBench: A Benchmarking Dataset for Music Source Restoration shows that perceptual and waveform metrics can disagree wildly, with a perceptual score (FAD CLAP) near 0.74 alongside an SI-SNR of −23.4 dB for Band-Split RNN, and that transients suffer most. On the method side, Multi-Source Diffusion Models for Simultaneous Music Generation and Separation (MSDM) replaces a Gaussian mixture likelihood with an exact sum constraint to avoid muffled outputs. SURF: Separation via Unsupervised Remixing Flow warns that generative separators can hallucinate structure with no correspondence to the input. These papers supply both the methods we will test and the failure taxonomy we will look for. They operate on waveforms and never on physical energy records, where a hallucinated tail is not a cosmetic flaw but a false physical claim.

The research file found no prior work joining TPM with source separation. The only search hits were a classical smart-meter disaggregation paper (Non-Intrusive Energy Disaggregation Using Non-negative Matrix Factorization with Sum-to-k Constraint) and an unrelated spatial-correlation paper. The obvious move, "apply a deep separator to quantum energy data," has not been done. It is also not what we propose as the contribution. The contribution is the audit: showing when and why such separators corrupt the statistics that physicists actually test.

Core Claim

Per TPM run, the recorded energy difference is $\Delta E = W + Q + R$. Here $W$ is the work done by the drive, $Q$ is the heat exchanged with the bath, and $R$ is the spurious difference created when the initial or final eigenvalue is misassigned at readout. An experimenter who wants the work distribution $P(W)$ must remove $Q$ and $R$ using some combination of known physics, a calibrated readout confusion matrix, and learned priors. We claim three things. First, across a family of separators, the histogram-level error of the recovered $P(W)$ correlates weakly with the error in the Jarzynski estimate $\widehat{J} = \frac{1}{N}\sum_{n} e^{-\beta W_n}$ (where $N$ is the number of runs and $\beta$ the inverse bath temperature) and with the error in the third and fourth cumulants. The reason is that $e^{-\beta W}$ magnifies errors in exactly the negative-work tail that histogram metrics weight least. Second, separators that rely on priors learned from simulations obeying the Jarzynski equality will partially erase genuine violations, producing false confirmations. Third, an estimator that enforces exact distribution-level mixture consistency and nonnegativity, and is fitted with a tilt-weighted loss, will substantially reduce tail-induced error relative to histogram-fitted separators.

Research Program

The first component is a ground-truth benchmark that comes in three tiers. In the quantum spin tier, we simulate qubit and qutrit TPM protocols with parameters matched to the NV experiments: switching times from 25 to 300 µs, readout fidelity 0.98, and sample sizes of 16,000 runs per condition (the Crooks experiment) and 80,000 runs (the Jarzynski experiment). We add weak thermal relaxation during the drive, so that $Q \neq 0$. In the resonator tier, we follow the tilted-Lindblad model of Heat, work, and fluctuations in a driven quantum resonator under sinusoidal, square-wave, and sawtooth driving. We use a confusion matrix whose misassignments are concentrated between neighboring levels, and we take $Q$ from the bath-counting statistics. Because this tier has many levels and strongly non-Gaussian work, it is where simple fixes should break. In the classical tier, we use the 500-atom Fermi-Pasta-Ulam lattice from The Mori-Zwanzig formalism for the derivation of a fluctuating heat conduction model from molecular dynamics with Gamma-distributed block energies. Here a "two-point" energy record of one block mixes driven work with heat from neighboring blocks. Every tier also includes deliberately violating instances: non-unital channels with intermediate measurements, which by Energy fluctuation relations and repeated quantum measurements break the naive Jarzynski equality, and feedback protocols. In all cases the true $W$, $Q$, and $R$ are logged for each run.

The second component is a set of artifact diagnostics. For each separator and condition we report four kinds of error. The first is the histogram error of $P(W)$, measured by total variation and KL divergence, which plays the role that SI-SDR plays in audio. The second is the error in each of the first four cumulants. The third is the error in $\log \widehat{J}$ and in the Crooks residual $P^F(W)/P^R(-W) - e^{\beta(W-\Delta F)}$, where $P^F$ and $P^R$ are the forward and reverse work distributions and $\Delta F$ is the free-energy difference. The fourth is an artifact decomposition modeled on audio practice. "Ghosting" is probability mass in $\widehat{P}(W)$ at work values that the true distribution assigns no mass. "Bleed" is the correlation between the recovered $W$ and the true $Q$. "Holes" are negative or zeroed probabilities.

The third component is the proposed estimator. It treats the observed joint histogram of initial and final readouts as an exact linear image of the latent joint distribution of $(W, Q, R)$ under the known confusion matrix and a parametric or simulator-based heat channel. It solves for the latent distribution subject to exact consistency with the observations. This consistency constraint is the distribution-level analogue of the MSDM sum constraint. The solution is also required to be nonnegative, which is the Chu–Li lesson against additive Gaussian corrections. The fit minimizes a tilt-weighted divergence:

$$ \mathcal{L} = \sum_{w} \left(1 + \lambda e^{-\beta w}\right) \, \ell\big(\widehat{P}(w), P_{\text{model}}(w)\big), $$

where $w$ ranges over work bins, $\ell$ is a per-bin likelihood term, and $\lambda \ge 0$ controls how strongly the negative-work tail is weighted. Setting $\lambda=0$ recovers ordinary histogram fitting. Crucially, the fluctuation theorem itself is never used as a constraint, so a test run on the output is not circular.

Baselines

The naive reference point performs no separation. It takes the recorded difference $\Delta E$ as the work, which is effectively how a high-fidelity, weak-relaxation experiment is analyzed. It is the right floor because the NV experiments report agreement with theory under exactly this kind of analysis, for example $\langle e^{-\beta W}\rangle = 0.97(1)$ versus 0.966(3) in Experimental test of the Jarzynski equality in a single spin-1 system using high-fidelity single-shot readouts. Our proposal is only worth anything in regimes where this floor fails measurably. The benchmark is built to locate the boundary of those regimes.

The strongest physics baseline inverts the calibrated confusion matrix and then deconvolves a model-computed heat distribution in the Fourier domain. It is standard, it uses no learned prior, and it is the method a careful experimentalist would reach for first. It receives the same confusion matrix and the same heat model as our estimator, and its regularization strength is tuned on the same held-out simulated conditions.

The strongest machine-learning baseline is an MSDM-style diffusion separator: a score network trained on simulated per-run $(W, Q)$ pairs and sampled under an exact sum constraint, following Multi-Source Diffusion Models for Simultaneous Music Generation and Separation. We also run its Gaussian-likelihood variant, so that the muffling effect reported there can be checked in the tails. A nonnegative-factorization separator with a sum constraint, following Non-Intrusive Energy Disaggregation Using Non-negative Matrix Factorization with Sum-to-k Constraint, serves as a classical disaggregation reference. All learned methods get the same simulated training budget, the same compute, and the same hyperparameter search size.

No reported result exists on Lacuna for separating work from heat or readout error in TPM data. The closest reported separator number is from audio: MSDM reports 16.48 dB SI-SDRi on Slakh2100. It cannot be transferred and serves only to show that the method is competitive in its home domain. Beating a baseline means a lower error in $\log\widehat{J}$ and in $\kappa_3$ and $\kappa_4$ on held-out conditions with a confidence interval excluding zero, with no loss in histogram error, and with violating instances correctly flagged as violating. Two ablations isolate the key ingredients. One sets $\lambda=0$, removing tail weighting. The other replaces exact consistency with a Gaussian likelihood.

Why not just…?

Why not just invert the confusion matrix and stop? At fidelity 0.98 on a qubit, this may work as well as anything, and we expect it to. Its weakness is that the inverse produces negative probabilities whenever counts are finite. The factor $e^{-\beta W}$ then amplifies these negative entries in the negative-work tail. The resonator tier, where many neighboring levels are confusable, tests whether this weakness matters in practice. If inversion matches our estimator there too, we will report that the proposal reduces to "invert, then audit the tails." In that case the contribution is the diagnostic and the false-confirmation result, not the estimator.

Why not just measure the heat directly, by counting bath quanta? When the bath can be monitored, this is the right answer, and in simulation it is our ground truth. The TPM experiments cited here measure only the system, however. Bath counting would add another measurement with its own back-action, which Measurement-driven single temperature engine shows can itself act as an energy source. The proposal targets the common case in which only system readouts exist.

Why not just widen the error bars with bootstrapping? Bootstrapping measures sampling variance, not the systematic bias a separator introduces. The decisive comparison is the coverage of bootstrap intervals for $\log\widehat{J}$ under each separator. If the intervals cover the truth at nominal rates, this objection wins and our estimator is unnecessary.

Why not just use a Bayesian posterior with a simulator prior? That approach belongs to the same family as the diffusion baseline, and it is exactly where we expect false confirmations. We test it directly by training the prior only on unital, equality-obeying dynamics and then applying it to non-unital data.

Proposed Experiments and Decisive Checks

The first decisive check is the metric-blindness result. Across all separators and conditions, we compute the rank correlation between histogram error and $|\log\widehat{J} - \log J|$. A low correlation, concentrated in strongly driven resonator conditions with large kurtosis, confirms the first claim. A high correlation refutes it.

The second check is false confirmation. On the non-unital and feedback instances, we measure the fraction of runs in which each separator's output passes a Jarzynski test, judged at the precision reported for the NV experiment. If learned-prior separators pass violating data substantially more often than the naive or physics baselines, the hallucination concern documented for generative audio separators in SURF: Separation via Unsupervised Remixing Flow carries over into physics as false confirmation of fluctuation theorems.

The third check covers the estimator and its ablations: tail-cumulant and Jarzynski error against the confusion-inversion baseline, at the 16,000- and 80,000-run sample sizes.

The fourth check probes the scaling limit raised by the Chu–Li findings, namely that coarser blocks in the classical lattice make the statistics more Gaussian. As block size grows, all separators should converge, and the advantage of tail-aware fitting should vanish. Seeing that convergence would confirm that the effect really comes from non-Gaussianity rather than from estimator quirks.