Using Detailed Balance to Self-Align Neutron Spectra for Spin-Liquid Diagnosis

Telling a quantum spin liquid apart from its competitors often comes down to how a magnet absorbs energy at very low frequencies. A quantum spin liquid is a magnet whose spins never freeze, even at absolute zero, because they stay quantum-entangled. Its competitors include frozen spin glasses and states with a hidden energy gap. Theory predicts sharply different low-frequency laws for these cases. A Kitaev-type spin liquid with generic perturbations should show a response that grows linearly with frequency at one momentum and as the cube of frequency at others. The idealized Kitaev model instead shows a gap, and a frozen glass shows a pile-up of signal near zero frequency. These laws are read off neutron scattering spectra in exactly the few-meV window nearest zero energy transfer. In that window, a small error in where "zero" sits on the energy axis can bend a straight line into a curve.

This note proposes to treat that problem as a temporal registration problem, the kind that breaks audio and video watermark decoders when a few samples are cropped. It borrows the fix those fields use: find a synchronization signal inside the data and align to it before decoding. For neutron spectra, physics provides the synchronization signal for free. The principle of detailed balance says the energy-gain and energy-loss sides of a spectrum must differ by a known Boltzmann factor at the sample temperature. That relation pins down the true zero of energy and any time-of-flight timing error. Our claim is that power-law exponents and phase labels estimated jointly with this self-alignment stay correct at offsets where the standard pipeline misclassifies. We will measure how large those offsets must be. That measurement decides whether the problem matters in practice.

Why this is a registration problem

The audio and video literature has documented how much damage tiny misalignments do. Finkelstein et al. showed that pointwise error metrics treat a one-sample shift as a large distortion, even when listeners hear no difference. Wang et al. showed that cropping a few leading samples moves the analysis grid, and frame-based watermark decoders then lose synchronization completely. Sloan et al. found that silent padding added by codecs makes clean speech score as degraded on quality metrics. In each case a fixed offset is mistaken for a change in content. The repairs share one structure. SIGMark recovers video frame order by cross-correlating against a known pseudorandom key spread across the whole sequence. Token-based audio watermarks such as ALIGNED-IS and Hidden in Plain Tokens scan over offsets and cluster confusable tokens. These papers supply the method template: a known structural constraint is used as an alignment reference. They say nothing about physics.

The physics side supplies the stakes. Low-Energy Spin Dynamics of the Honeycomb Spin Liquid Beyond the Kitaev Limit shows analytically that the textbook spin gap of the Kitaev honeycomb model is an artifact of an accidental symmetry. Under generic perturbations it predicts $S(q,\omega)\sim|\omega|$ at the zone center and $S\sim|\omega|^3$ near the zone corners. That paper assumes a perfectly registered continuous frequency axis. Bulk probes cannot settle the question. Cao et al. on honeycomb iridates, and the Ba3NiSb2O9 high-pressure study, show that susceptibility and heat capacity cannot separate a spin liquid from defect-driven freezing. Those results are motivation: frequency-resolved spectra are the evidence that discriminates, so their alignment matters.

The research file found no paper connecting these two areas. The closest physics machine-learning work, Quantum Loop Topography, classifies topological order from static Monte Carlo configurations. It never touches a measured time or frequency axis, so it neither competes with nor constrains this proposal.

Core Claim

In a direct-geometry time-of-flight spectrometer, energy transfer is computed from neutron arrival time. A timing-zero error $\delta t_0$ therefore warps the energy axis nonlinearly. It is not a simple shift, and an error in the nominal elastic energy adds a further offset $\delta E$. Theory predicts the true spectrum must satisfy

$$ S(q,-\omega) = e^{-\hbar\omega/k_B T}\, S(q,\omega), $$

where $T$ is the independently measured sample temperature. Only the correct $(\delta t_0, \delta E)$ makes the observed gain and loss sides consistent with this ratio. We will build a forward model. A parameterized true spectrum (gapped, linear, cubic, or a quasi-elastic glass-like Lorentzian) is passed through detailed balance, convolved with the instrument resolution, mapped through the time-of-flight warp, and converted to Poisson counts. The model is fit by maximum likelihood with the alignment parameters free. Both sides of the spectrum must be fit together. This coupling breaks the degeneracy between an offset and a change in exponent, which a one-sided power-law fit cannot do.

The constraint has a built-in limit. The gain side is suppressed by the Boltzmann factor. The method therefore only has leverage when $k_BT$ is comparable to the energy window being analyzed. That is the same low-energy window in which the competing laws differ, so the limit and the target overlap. We will map where the leverage disappears rather than assume it holds everywhere.

Research Program

The first stage is a sensitivity map, which can end the project on its own. Synthetic spectra will be generated from the honeycomb paper's three predicted forms and a glass-like quasi-elastic form. We will vary resolution width, counting statistics, temperature, and injected $(\delta t_0,\delta E)$, and measure two things: how large an offset, as a fraction of resolution width, flips the naive classification, and how far it biases the fitted exponent. The file contains no figures for typical residual calibration errors on real instruments. If the flipping offsets turn out much larger than any plausible calibration residual, we will report that as a negative result and stop.

The second stage builds the joint self-aligning estimator. The third stage tests it under realistic contamination: incoherent elastic scattering, phonon backgrounds, and a quasi-elastic component that pulls the apparent elastic peak off center.

Baselines

The naive reference applies nominal calibration and fits a power law to the energy-loss side alone. It shows how large the problem is.

The strongest existing competitor is standard elastic-line calibration. It locates the center of the incoherent elastic peak, or a vanadium reference measurement, and redefines zero there before fitting. This is what practitioners actually do, and it is the real threat. It will receive the same simulated data, including a simulated vanadium run with matched statistics. Our estimator beats it if it keeps correct linear-versus-cubic-versus-gap labels, and exponent errors below 0.2, at offsets where elastic calibration fails. We expect those failures where a quasi-elastic signal drags the elastic peak, or where the time-of-flight warp is not captured by a single shift. This is our own target. The file reports no result for any of these methods on this task: no reported result found on Lacuna for neutron-spectrum exponent classification under registration error.

The key ablation fits the same forward model with alignment parameters free but no detailed-balance coupling. This mirrors the multi-offset scanning used by ALIGNED-IS and isolates whether the physics constraint, rather than mere freedom to shift, does the work.

A learned competitor completes the set: a small neural classifier trained with random-offset augmentation on identical synthetic data, with the same hyperparameter search budget as our fit. On a different task, the file shows how much explicit alignment can matter. SIGMark reports about 81–87% bit accuracy under 30-frame drops, against about 51% for VideoMark. That figure motivates the comparison but cannot serve as a target here.

Why not just…?

Why not just calibrate on the elastic line, as everyone already does? Because the elastic peak is itself the region where a glass-like phase puts quasi-elastic weight, and a time-zero error is a warp rather than a shift. This is still the alternative most likely to make the proposal unnecessary. The decisive experiment is a direct head-to-head on contaminated spectra. If elastic calibration matches our estimator across the realistic range, we will say so and narrow the contribution to the sensitivity map.

Why not just scan over offsets and keep the best-fitting one? A free offset and a free exponent trade off near zero once the spectrum is blurred by resolution, so the best-fitting offset can be wrong while still fitting well. This mirrors the finding in Hidden in Plain Tokens that re-shifting alone does not repair corrupted content. The no-detailed-balance ablation tests this claim directly.

Why not just train a shift-augmented classifier? It may work on in-distribution simulations. However, it learns invariance only to the offsets it was shown, and it gives no calibrated offset estimate to check against. Skoglund et al. found that alignment tuned to static assumptions breaks under real time-warping. We will test the classifier on warp forms held out of training.

Why not just raise counting statistics? More neutrons shrink noise but not bias. A misregistered axis converges confidently to the wrong exponent, and the sensitivity map will show this as statistics increase.

Decisive Checks

The project succeeds if three things hold. First, the sensitivity map shows label flips at offsets that are a modest fraction of resolution width. Second, the detailed-balance estimator recovers the injected offsets and correct labels where both elastic calibration and the unconstrained ablation fail. Third, its advantage disappears, as predicted, when $k_BT$ falls well below the analysis window. This last check is the falsifiable signature that the gain mechanism is detailed balance itself. If real spectra from honeycomb iridates of the kind studied by Cao et al. become available, the final test is whether self-alignment changes the reported low-energy exponent. The file contains no such public data, so this test depends on collaboration.

References

  1. A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences
  2. Formant enhancement based speech watermarking for tampering detection
  3. Objective Assessment of Perceptual Audio Quality Using ViSQOLAudio
  4. SIGMark: Scalable In-Generation Watermark with Blind Extraction for Video Diffusion
  5. Robust Distortion-Free Watermark for Autoregressive Audio Generation Models
  6. Hidden in Plain Tokens: Simply Robust, Gradient-Free Watermark for Synthetic Audio
  7. Low-Energy Spin Dynamics of the Honeycomb Spin Liquid Beyond the Kitaev Limit
  8. Evolution of magnetism in the single-crystal honeycomb iridates<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:mo>(</mml:mo><mml:msub><mml:mtext>Na</mml:mtext><mml:mrow><mml:mn>1</mml:mn><mml:mo>−</mml:mo><mml:mi>x</mml:mi></mml:mrow></mml:msub></mml:math>Li<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mrow/><mml:mi>x</mml:mi></mml:msub><mml:msub><mml:mo>)</mml:mo><mml:mn>2</mml:mn></mml:msub><mml:mtext>Ir</mml:mtext><mml:msub><mml:mtext>O</mml:mtext><mml:mn>3</mml:mn></mml:msub></mml:math>
  9. High-Pressure Sequence of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msub><mml:mi>Ba</mml:mi><mml:mn>3</mml:mn></mml:msub><mml:msub><mml:mi>NiSb</mml:mi><mml:mn>2</mml:mn></mml:msub><mml:msub><mml:mi mathvariant="normal">O</mml:mi><mml:mn>9</mml:mn></mml:msub></mml:math>Structural Phases: New<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>S</mml:mi><mml:mo>=</mml:mo><mml:mn>1</mml:mn></mml:math>Quantum Spin Liquids Based on<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:msup><mml:mi>Ni</mml:mi><mml:mrow><mml:mn>2</mml:mn><mml:mo>+</mml:mo></mml:mrow></mml:msup></mml:math>
  10. Machine learning <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"><mml:msub><mml:mi mathvariant="double-struck">Z</mml:mi><mml:mn>2</mml:mn></mml:msub></mml:math> quantum spin liquids with quasiparticle statistics
  11. Robustness of speech quality metrics to background noise and network degradations: Comparing ViSQOL, PESQ and POLQA