Rare-event experiments face a problem that per-event classifiers handle badly. A dataset of millions of detector pulses may contain a small subpopulation of events that differ subtly from the bulk: a new class of electronic glitch, a slowly drifting multi-site fraction, or an unanticipated background. No label exists for that subpopulation, and its signature is spread across many weak correlations rather than one clean feature. We propose a label-free, population-level test for "is there structure in this batch beyond known noise?"
The test builds on a recent physics-inspired detector. It treats the eigenvalue spectrum of the waveform covariance as the "momenta" of a statistical field theory. It then runs a renormalization group flow, which progressively averages out fine-scale structure. The new step is to push that flow onto the real frequency axis, where it yields a continuous, calibrated detection statistic instead of a yes/no phase label.
The starting point is the work of Lahoche and collaborators, Field Theoretical Approach for Signal Detection in Nearly Continuous Positive Spectra II: Tensorial Data. They consider noisy data whose covariance eigenvalues form a continuum with no clean outlier. In that setting a weak signal shows up as spontaneous breaking of a $Z_2$ symmetry in the effective potential $V_k(\phi)$, the scale-dependent energy landscape of a uniform field. They compute $V_k$ in the local potential approximation (LPA), which keeps the full shape of the potential but drops all gradient and momentum structure. Their criterion is binary and qualitative, and they report no numerical detection threshold ("no reported result found on Lacuna" for a spiked-tensor threshold). It has also never touched real detector data. We ask whether analytically continuing the flow of the two-point function, the dressed propagator describing how fluctuations respond at a given scale, into the complex spectral plane turns that criterion into a usable statistic. We also ask whether that statistic beats simpler random-matrix and pulse-shape tests on real germanium waveforms.
Core Claim
In the Lahoche construction, the loop integrals of the flow run over the empirical eigenvalue distribution of the data. That distribution plays the role that the momentum-space density of states plays in a field theory. The retarded two-point function is obtained by continuing the flow equation itself, following Tripolt, Kamikado, Strodthoff and collaborators. Its imaginary part is therefore a renormalization-group-resummed, nonlinear transform of the data spectrum, with a threshold, possibly an isolated pole, and a width.
Our claim is that a signal buried inside the noise continuum moves two things before it produces any empirical eigenvalue outlier: the pole residue and the threshold position of this dressed spectral function. Both are continuous quantities, so they give a statistic with a tunable false-alarm rate. The binary phase label of the static LPA analysis cannot provide that.
Continuation matters for a specific reason. The same quantity could in principle be read off a Euclidean flow followed by numerical continuation, for example with Padé approximants. However, Spectral functions from the functional renormalization group and Real-time correlation functions in the O(N) model from the functional renormalization group show that continuing the flow before integrating it turns an ill-posed inversion into a well-posed initial-value problem. The alternative is a Fredholm inversion of the first kind, which blurs sharp peaks. Sharp near-threshold structure is exactly what we need to resolve.
This claim could easily be false. The dressed spectral function might carry no more information than the empirical Stieltjes transform, the standard random-matrix object that already encodes the spectrum analytically. The research program is built so that this outcome is detected, not hidden.
Research Program
The first component is the method. We reimplement the Lahoche LPA flow for the matrix (order-2) case, where the elementary interaction is unambiguous. In this model the kinetic term is set by the empirical covariance eigenvalues of a waveform ensemble, and the regulator removes eigenvalues below a running scale.
We then derive the flow of the retarded two-point function by making the substitution $p_0 \to -i(\omega + i\eta)$ inside the one-loop integrand. This follows Real-time correlation functions in the O(N) model from the functional renormalization group, whose LPA-background construction we reuse directly. It feeds vertices taken as derivatives of $V_k$ at its running minimum into the real-frequency flow.
Poles of the dressed propagator migrate as the scale runs. When one crosses the integration contour, the flow breaks down. We therefore solve on a grid in the complex plane with back-coupling of pole positions, as developed by Strodthoff in paper. This is the most expensive step. It is tractable here because the model has a single scalar channel rather than a gauge theory.
The readout is the infrared spectral function $\rho(\omega) = -\frac{1}{\pi}\,\mathrm{Im}\,G^R(\omega)$, with $G^R$ the continued propagator at $k \to 0$. From it we define the test statistic $T$ as the log-ratio of spectral weight in a window just below the noise-only threshold to that predicted under a noise-only null. We calibrate the null by bootstrap resampling of known-clean ensembles.
The second component is the truncation question raised by both the file and the physics literature. Tripolt et al. report that static LPA can create artifacts near decay thresholds, and threshold structure is precisely what $T$ reads. We therefore run the flow at two levels. One is plain LPA. The other is LPA′, which adds a single scale-dependent field-rescaling factor $Z_k$. Towards quantitative precision in the chiral crossover: Masses and fluctuation scales found this level within 3% of full momentum-dependent results in QCD-like models. If $T$ changes sign or its threshold moves substantially between the two levels, the statistic is a truncation artifact and we report it as such.
The third component is the application. Germanium pulse-shape data are the target because good per-event labels exist to validate against. Deep learning based pulse shape discrimination for germanium detectors supplies a $^{228}$Th protocol in which double-escape-peak events are single-site (signal-like) and full-energy-peak events are multi-site (background-like). From such data we build batches of single-site waveforms contaminated with a fraction $\epsilon$ of a second population. We then measure the smallest $\epsilon$ each method detects at fixed false-alarm rate and batch size $N$.
The contaminants come in two kinds. Known-type contamination uses multi-site events. Novel-type contamination uses waveform classes that no standard pulse-shape parameter targets: slow-trailing surface-like pulses, which delayed charge recovery (DCR) is built to catch per α-event characterization and rejection in point-contact HPGe detectors, and non-physical electronic pulses of the kind catalogued for LEGEND in Machine learning-powered data cleaning for LEGEND. Where real samples are not available, we use simulated injections.
Baselines
The naive reference is the classical outlier test. It flags a batch when the largest covariance eigenvalue exceeds the noise-only bulk edge, estimated from clean ensembles. This is the test Lahoche et al. argue fails for nearly continuous spectra. It fixes the floor any method must clear.
The strongest spectrum-only competitor is a likelihood-style test on linear spectral statistics. It compares the empirical Stieltjes transform of the batch spectrum with its noise-only bootstrap distribution, evaluated on the same complex-plane contour we use. It sees exactly the same input as our method, so beating it is the only evidence that the renormalization group resummation adds information. We give it the same bootstrap calibration budget and let it choose its evaluation points by cross-validation on held-out clean data.
The static-LPA baseline is Lahoche's own criterion, broken versus symmetric phase at $k \to 0$. We run it with an identical regulator and grid. Converted to a statistic by the size of the order parameter, it shows what the continuation adds over the Euclidean analysis.
The ablation that isolates the key ingredient runs the Euclidean flow of the two-point function and continues the result post hoc with Padé approximants. This design follows the ill-posedness diagnosis in paper. If flow-level continuation does not beat this, the second research direction contributes nothing here.
The strongest per-event competitor is a counting test. It applies the A/E cut (peak current over energy) or the autoencoder classifier of Holl et al. to each waveform, then tests whether the fraction flagged exceeds the calibrated false-positive rate. Holl et al. report that both reach 90% background rejection at 90% signal acceptance on $^{228}$Th double-escape versus full-energy peaks. For anomaly cleaning, León et al. report physics-event sacrifice of 0.024±0.004% and anomalous-event leakage of 15.034±0.224% on high-rate data.
All methods see the same batches and the same clean calibration set. Neural baselines get a matched hyperparameter search. Beating a baseline means a smaller detectable $\epsilon$ at 1% false-alarm rate across at least three batch sizes, with bootstrap confidence intervals that do not overlap.
Why not just…?
Why not just count events failing A/E or the autoencoder? For multi-site contamination this will very likely win. The discriminator was designed for that population, and counting with 90/90 performance is statistically efficient. We expect to lose on known-type contamination and will report it. The method's claim concerns novel contaminants, which a supervised or engineered discriminator was not built to flag. The split between the known-type and novel-type experiments settles whether that niche is real.
Why not just use the empirical Stieltjes transform or random-matrix spectral tests? This is the most serious threat. The dressed spectral function is ultimately a functional of the same empirical spectrum. If a well-tuned linear-spectral-statistic test matches $T$, the renormalization group machinery is an expensive reparameterization and we will say so. The head-to-head comparison at fixed calibration budget is the decisive experiment. A positive result would need to show that the nonlinear, scale-by-scale resummation concentrates weak in-bulk perturbations into a threshold shift that linear functionals average away.
Why not just continue a Euclidean result with Padé or maximum entropy? Both washing out sharp structure and the instability of noisy inversions are documented. Even learned inverters such as Noise enhanced neural networks for analytic continuation must inject noise to remain stable. Rather than assume this carries over, the Padé ablation measures it directly on our statistic.
Why not just train an unsupervised anomaly detector on waveforms? Autoencoder reconstruction error is a legitimate competitor for novel contaminants. We include it as an additional arm, pooled per batch, under the same data. Its weakness is shift from calibration data to physics data, which Holl et al. and León et al. both flag. That weakness is tested by training on one detector's calibration runs and evaluating on another's.
Decisive Checks
On synthetic spiked-matrix ensembles, we map the detectable signal strength versus aspect ratio for all spectrum-only methods. The outlier test's threshold is known, so this checks calibration. The method passes only if $T$ detects below the outlier threshold and is stable between LPA and LPA′.
On germanium batches, the central result is the pair of curves of detectable $\epsilon$ versus $N$ for known and novel contaminants. The proposal is supported if $T$ beats the Stieltjes-transform test and the Padé ablation on novel contaminants while remaining competitive elsewhere. It is refuted if the Stieltjes-transform test matches it, or if its gains vanish under the LPA′ correction.
Finally, we report wall-clock cost of the complex-plane flow per batch. A statistic that wins but cannot run inside a data-quality loop has limited practical value.
References
- Field Theoretical Approach for Signal Detection in Nearly Continuous Positive Spectra II: Tensorial Data
- Spectral functions from the functional renormalization group
- Real-time correlation functions in the $$O(N)$$ O ( N ) model from the functional renormalization group
- Self-consistent spectral functions in the <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi>O</mml:mi><mml:mo stretchy="false">(</mml:mo><mml:mi>N</mml:mi><mml:mo stretchy="false">)</mml:mo></mml:math> model from the functional renormalization group
- Towards quantitative precision in the chiral crossover: Masses and fluctuation scales
- Deep learning based pulse shape discrimination for germanium detectors
- $$\alpha $$-event characterization and rejection in point-contact HPGe detectors
- Machine learning-powered data cleaning for LEGEND: a semi-supervised approach using affinity propagation and support vector machines
- Superconductivity and non-Fermi liquid behavior near a nematic quantum critical point
- Noise enhanced neural networks for analytic continuation