Line Detection Instead of Template Fitting for Reactor Antineutrino Mass Splittings

Reactor neutrino experiments measure how many antineutrinos survive the trip from a power plant to a detector, as a function of the particle's energy. The mass differences between neutrino types appear as ripples in that energy spectrum. The standard analysis bins the events into a histogram and fits a predicted spectrum template. That fit is only as good as the predicted unoscillated flux, which is the spectrum the reactor would produce with no oscillation, and the predicted flux is known to be wrong in places. The best-known failure is the "5 MeV bump," an unexplained excess relative to conversion models (Neutrino physics with JUNO). Near/far detector pairs such as Daya Bay cancel this problem by taking a ratio. A single medium-baseline detector like JUNO cannot, and it has to read sub-percent mass splittings and the neutrino mass ordering off fast ripples that sit on top of a poorly known flux shape.

The idea is to change variables so the problem becomes one signal processing already solves well. Written as a function of inverse energy $x = 1/E$, each oscillation term is a pure sinusoid whose frequency is proportional to a mass splitting times the baseline. The reactor flux, bump included, is a smooth curve whose content is broadband. Recovering a mass splitting then becomes the classic task of estimating a spectral line against an unknown smooth background. Multitaper spectral estimation was built for that task. It averages over a set of orthogonal "Slepian" weighting windows that concentrate their sensitivity inside a chosen frequency band. Here we apply it to the raw event list, treated as a random point process, and do no histogramming. The open question this proposal answers is how to choose the analysis bandwidth: how wide a frequency band each estimate pools, and how many windows to use. That choice decides whether the method is robust to flux errors or merely noisier than the standard fit.

Core Claim

We claim that a band-limited likelihood built from multitaper coefficients of the inverse-energy event distribution will estimate the atmospheric mass splitting with near-nominal frequentist coverage when the reactor flux model is wrong in ways the template fit does not anticipate. The template fit loses coverage in those same cases. The improvement depends on two ingredients. The first is a trimmed taper count, so that leakage from the enormous low-frequency flux continuum into the oscillation band is suppressed by orders of magnitude. The second is a bandwidth chosen to minimize predicted mean squared error. That error combines Poisson variance with the bias from treating the flux continuum as locally flat across the band.

We do not claim the method beats a correctly specified template fit. When the flux model is right, the full likelihood is optimal and our estimator must lose some efficiency, because it discards the low-frequency band. The proposal stands or falls on whether that efficiency cost is small and the robustness gain under realistic misspecification is large.

A back-of-envelope estimate, our own and not from the file, shows why this regime needs care. At a 53 km baseline, the fast oscillation sits near 27 cycles per MeV⁻¹ in $x$. The usable window spans roughly 0.4 MeV⁻¹. The two nearly equal atmospheric splittings, whose interference carries the mass-ordering signal, are therefore closer together than the window's natural frequency resolution. A one-MeV-wide bump also has a non-negligible tail at that frequency. The analysis cannot "resolve the doublet." It must model the doublet jointly inside one band while controlling leakage from the continuum and from the bump's own high-frequency tail.

Research Program

We model the detected events as a Poisson process in $x$ with intensity

$$\lambda(x) = \Phi(x)\,\big[1 - \textstyle\sum_j A_j \sin^2(\omega_j x)\big] \ast R,$$

where $\Phi$ is flux times cross section, $A_j$ and $\omega_j = 1.27\,\Delta m^2_j L$ are the amplitude and angular frequency of each oscillation term, and $\ast R$ denotes smearing by the detector energy response.

For each Slepian taper $h_k$ on the observation window, with half-bandwidth $W$, the eigencoefficient at frequency $f$ is $y_k(f) = \sum_i h_k(x_i)\, e^{-2\pi i f x_i}$, summed over events $i$. Its Poisson variance is $\int h_k^2 \lambda\,dx$, so no Gaussian or stationarity assumption is needed.

Inside a band around the expected oscillation frequency, the expected eigencoefficients are written as the exact Fourier image of the fast oscillation terms plus a locally flat complex continuum. The fast terms include the sidebands generated by the slow solar oscillation and by smearing. The flat continuum absorbs flux-shape error, and its amplitude is a nuisance parameter. This generalizes Thomson's harmonic line test to a parametric multi-line model. The likelihood uses only the $K$ coefficients in that band.

The statistical machinery comes from three sources:

  • Leakage control. Thomson's Multitaper Method Revisited shows that the method is a projection onto a band-limited subspace. It also shows that the textbook choice $K = 2NW - 1$ includes poorly concentrated tapers whose leakage let a $10^9$-strength source mask weaker ones, and that trimming to $K = 2NW - O(\log NW)$ fixes this. That paper treats only Gaussian stationary processes and tunes the log constant by hand. We derive the constant from a Poisson leakage bound: the trimmed tapers' out-of-band energy times the total event count must stay below a fixed fraction of the expected line amplitude.
  • Point-process precedent. Multitapering has already been applied to spatial point processes for structure-factor estimation (On estimating the structure factor of a point process, with applications to hyperuniformity). That work shows the eigencoefficient construction carries over to event data. It does not address parametric line inference.
  • Bandwidth selection. The bandwidth rule follows the local mean-squared-error-minimizing selector of Weak Convergence and Adaptive Peak Estimation for Spectral Densities. That work handled kernel-smoothed periodograms and warned that data-driven bandwidths can collapse to extremes. We replace its curvature term with a bound on the continuum's departure from flatness across $[f-W, f+W]$, estimated from coefficients in adjacent bands. We also cap $W$ by the requirement that the band contain both atmospheric lines and their solar sidebands.

The helioseismic finding that large taper counts broaden narrow modes until they cannot be fitted (Multitaper Spectral Analysis and Wavelet Denoising Applied to Helioseismic Data) constrains us. It is why we fit a parametric line model to the coefficients rather than reading the frequency off a smoothed peak, and why $W$ is bounded above by an explicit resolution requirement rather than by an empirical rule such as $K<10$.

Confidence intervals are calibrated by toy Monte Carlo in the Feldman-Cousins style. Because the band likelihood has few parameters, this is cheap. If it is not, the Gaussian-process active learning of Efficient Neutrino Oscillation Parameter Inference with Gaussian Process applies unchanged downstream. That work treats the binned spectrum as a fixed input, while ours changes the input statistic itself.

Baselines

Naive reference. The do-nothing reference is the raw periodogram of the inverse-energy event list, a single untapered transform with the peak frequency read off directly. It shows what changing variables alone buys.

Strongest competitor. This is the standard binned Poisson-likelihood template fit with full detector response and nominal flux, as used by Daya Bay (Full Data Release of the Daya Bay Reactor Neutrino Experiment) and planned for JUNO. We run it in two variants. The first uses a nominal flux model. The second adds a flexible flux nuisance: a free spline correction with a number of knots tuned on held-out toys. The flexible variant is the true competitor, because it is what an experiment would do about the bump. Both variants and our method use the same simulated event lists, the same response model and energy-nonlinearity nuisance, and the same toy budget for coverage calibration. The spline knot count gets the same tuning budget as our bandwidth.

The relevant reported numbers are design targets or results from a different geometry, not direct comparisons. JUNO's design reports 3–4σ mass-ordering sensitivity in six years and a sub-1% precision goal on $|\Delta m^2_{ee}|$ (Neutrino physics with JUNO). Daya Bay's near/far result is $\Delta m^2_{32} = (2.466 \pm 0.060)\times10^{-3}$ eV² (2.4%) under normal ordering (Full Data Release of the Daya Bay Reactor Neutrino Experiment). No reported result was found on Lacuna for multitaper or any frequency-domain estimator applied to reactor antineutrino spectra.

What counts as beating the competitor. Under injected flux distortions, our method must keep 68% and 95% interval coverage within a few percent of nominal where the spline fit does not. Its interval width under correct specification must also stay within roughly 20% of the template fit's width. If the spline fit retains coverage at equal or smaller width, our method does not beat it.

Ablations. The ablation that isolates the key ingredient keeps everything fixed but uses untrimmed $K = 2NW - 1$. A second ablation fixes $W$ at a hand-picked value instead of the adaptive rule, which isolates bandwidth selection.

Why not just…?

Why not just fit the template with a flexible flux nuisance? This is the real threat. With enough spline freedom, the fit can absorb the bump. The open question is what happens on the high-frequency side: a spline flexible enough to follow a narrow distortion near 27 cycles per MeV⁻¹ can also partly absorb the oscillation itself. The fit must then either lose coverage (too stiff) or lose precision (too loose). Our method makes this separation explicitly in frequency, rather than implicitly through knot spacing. The misspecification sweep in the experiments directly measures which approach sits on the better coverage-versus-width frontier. If the spline fit dominates, we will report that and drop the claim.

Why not just Fourier-transform the inverse-energy spectrum without tapers? The low-frequency flux continuum carries essentially all events, while the fast line carries a few percent. A single untapered window leaks a fraction of the continuum into the oscillation band that is comparable to the line itself. This is the same mechanism by which the strongest source masked weaker ones in the dynamic-range test of Thomson's Multitaper Method Revisited. The naive-periodogram baseline and the untrimmed ablation quantify this directly.

Why not just use a near detector ratio? It works when available, which is why Daya Bay's 0.13% uncorrelated systematic (Full Data Release of the Daya Bay Reactor Neutrino Experiment) is so strong. The single-detector medium-baseline geometry that drives mass-ordering sensitivity has no such ratio within the scope of this file. A method that is robust to flux errors without a ratio therefore complements the ratio approach rather than competing with it.

Why not just fit a high-order parametric spectral model? Autoregressive fits produce spurious peaks with tiny changes in model order: order 27 created a pseudo-peak that order 26 did not (Nonparametric and parametric methods of spectral analysis). An AIC-selected autoregression had peak-location MSE of $8.923\times10^{-2}$, against $7.136\times10^{-3}$ for adaptive smoothing (Weak Convergence and Adaptive Peak Estimation for Spectral Densities). We keep the physics model only for the oscillation lines and leave the flux nonparametric.

Proposed Experiments

Main benchmark. The main test is a JUNO-like simulation following the published specification: 53 km baseline, 3% at 1 MeV resolution, and six-year statistics (Neutrino physics with JUNO). We inject families of flux distortions on top of a nominal flux: bumps of varying width and amplitude near 5 MeV, plus fine-structure ripples of varying frequency. For each method we measure four quantities across thousands of toys: bias in $\Delta m^2_{31}$, interval coverage, interval width, and median mass-ordering $\Delta\chi^2$.

Decisive check. The decisive result is a coverage-versus-distortion curve. Our method should stay near nominal across distortion widths down to the point where the bump's tail genuinely overlaps the line. That cutoff is predictable from our bias bound, and it is part of the test. The template fits should drift away from nominal earlier.

Real-data null test. This test uses Daya Bay. At kilometer baselines the atmospheric oscillation completes less than one cycle in inverse energy, so the high-frequency band should contain no line at all. If the full data release (Full Data Release of the Daya Bay Reactor Neutrino Experiment) provides event lists or sufficiently fine binning, near-hall spectra with millions of events and the real bump test whether real reactor fine structure produces false lines at the target frequency. That is a direct empirical check of the flat-continuum and leakage assumptions on data no simulation designed. If the released binning puts the band above the binning's frequency limit, we will state that and rely on simulation alone.

Applicability boundary. The same Daya Bay data also checks the method's honesty: with fewer than a Rayleigh cycle in the window, it should return wide, uninformative intervals rather than spurious precision.

Low-statistics stress test. SNO+'s 59-pair sample (Measurement of Reactor Antineutrino Oscillation at SNO+) serves only to test whether the Poisson coefficient variance remains well calibrated at very low counts. It is not a precision benchmark.