Neutrino physicists estimate a few fundamental constants by fitting a physics model to histograms of detected events across energy bins. When part of the spectrum does not match the model, the standard fit has two options. It can keep trusting the model and absorb the mismatch through a fixed, hand-built uncertainty matrix, or an analyst can cut the offending energy range by hand. We propose a third option taken from two other fields. In speech processing, microphone-array methods decide frame by frame which frequency bins to trust. In robust matrix factorization, the corruption is modeled as a learned mixture of Gaussians, which can approximate almost any noise shape. Combined, these ideas let the fit learn, per energy bin and per detector, how much to trust the model. The fit is constrained so that this learned "distrust" cannot soak up the oscillation signal itself.
The key point is that the constraint does most of the work, not the mixture. Reactor experiments like Daya Bay see the same reactors from several distances. Oscillation depends on distance divided by energy, while flux-model errors such as the anomalous excess near 5 MeV depend on energy alone. We place the flexible mixture noise only where oscillation cannot live. The fit's posterior belief that a bin is "clean" then acts as a soft, adaptive bin weight. We test whether this gives unbiased, well-covered estimates of $\sin^2 2\theta_{13}$ and $\Delta m^2_{32}$ in the cases where standard fits are fragile: mismatches that are specific to one detector or that drift in time.
Core Claim
A structured mixture-noise likelihood gives oscillation estimates that are no worse than the standard fit when the model is correct, and substantially less biased, with correct interval coverage, when unmodeled distortions are injected that do not cancel between near and far detectors. In such a likelihood, each bin's residual carries a narrow "well-modeled" component plus broad components for unmodeled distortions, and the residuals are tied across detectors by baseline structure. It should also report automatically which bins it has down-weighted.
The claim is only interesting because of a known fact. When the likelihood is correct, it already weights bins optimally, so any adaptive reweighting can only lose information. The proposal is therefore about robustness under misspecification, and it has to be measured as a trade: how much precision is paid in the well-specified case, and how much bias is removed in the misspecified one.
Relation to Prior Work
Two acoustic papers supply the bin-trust idea. Potamitis et al. estimate talker direction by keeping only the highest-SNR frequency bins before independent component analysis. They find that this hard selection becomes unstable at 0 dB and below and with few frames. That is a constraining result: it argues against hard thresholds. Xue et al. instead compute continuous per-bin weights from a recursively smoothed a priori SNR, carrying information from previous frames, so that interference-dominated bins receive near-zero weight. We borrow the soft, recursive form. Our "frames" are Daya Bay's data-taking periods, and our weights are posterior responsibilities rather than heuristic SNR estimates. Xue et al. also report that recursion propagates errors from earlier frames. That warns us to let the time coupling be learned rather than fixed.
Two mixture-noise papers supply the noise model. MoG-RPCA models each residual entry of a low-rank matrix as a Gaussian mixture fitted by variational Bayes. NMoG-LRMF gives each hyperspectral band its own mixture, which in effect down-weights noisy bands, and couples bands through a shared low-rank structure. Its band-specific noise is the closest analogue to our detector-specific residuals. Both papers learn noise with no physical constraint. The research file flags exactly this as the danger in physics: a flexible noise model could absorb the signal. Our contribution is the constraint, not the mixture.
On the physics side, the full Daya Bay data release documents time-dependent "muon-x" low-energy backgrounds. The flux measurement documents a persistent mismatch with the Huber-Mueller flux model, $R = 0.952 \pm 0.014 \pm 0.023$. Both motivate the work.
The long-baseline analyses handle out-of-model effects differently. The T2K–NOvA joint fit and NOvA's Bayesian analysis use hand-designed "nightmare" inflated-uncertainty tests and fixed binning. We aim to replace that manual stress test with a learned one, but we test on reactor data first, because there the identifiability argument is clean.
Li et al. fit a Gaussian-process surrogate to the Feldman-Cousins critical-value surface and cut the toy Monte Carlo budget by 70% in a 2D toy. They model a test statistic, not spectral noise, so they are not a competitor. We may reuse their method to afford the coverage studies.
Research Program
Let $n_{d,t,i}$ be the event count in detector $d$, data period $t$, and prompt-energy bin $i$. Let $\mu_{d,t,i}(\theta,\eta)$ be the standard Daya Bay prediction, where $\theta$ are the oscillation parameters and $\eta$ are the usual nuisance parameters with their existing priors. We model
$$ n_{d,t,i} \sim \mathrm{Poisson}\big(\mu_{d,t,i}(\theta,\eta)\,(1 + s_i + r_{d,t,i})\big), $$
where $s_i$ is a shape distortion shared by all detectors and $r_{d,t,i}$ is a detector- and period-specific distortion. Each gets a mixture prior:
$$ r_{d,t,i} \sim \sum_{k=0}^{K} \pi_{d,t,k}\,\mathcal{N}(0,\sigma_k^2), $$
with $\sigma_0$ fixed at the size of the existing uncorrelated detector uncertainty and $\sigma_{k\ge1}$ learned. The prior on $s_i$ has the same form.
The posterior probability that bin $i$ of detector $d$ in period $t$ belongs to the narrow component $k=0$ is the bin-trust weight, our analogue of Xue et al.'s SNR weight. The mixing proportions $\pi_{d,t,k}$ follow a random walk over periods $t$, which gives a learned version of their recursive smoothing.
Identifiability is argued in two steps. The shared term $s_i$ is a function of energy alone, identical at all baselines, so it can mimic an overall flux-shape error but not the baseline-dependent oscillation pattern. This is the same cancellation the near-far method already relies on. The detector-specific term $r$ is where the risk lies. We deliberately scan the width of its broadest allowed component and trace a curve of oscillation-parameter bias against precision. That curve is itself a result: it states how much detector-specific freedom the data can tolerate before $\theta_{13}$ information disappears.
Inference uses Hamiltonian Monte Carlo in Stan, as NOvA already does. A variational Bayes version following MoG-RPCA serves as a fast approximation, checked against the Monte Carlo result. $K$ is chosen by held-out predictive likelihood across data periods rather than by the heuristic pruning the mixture literature relies on.
Baselines
The naive reference is the standard full-data fit: a Poisson or covariance-based likelihood over fixed bins with existing systematic pulls, as in the Daya Bay full release, which reports $\sin^2 2\theta_{13} = 0.0851 \pm 0.0024$ and $\Delta m^2_{32} = (2.466 \pm 0.060)\times10^{-3}\,\mathrm{eV}^2$ under normal ordering. We reimplement it on the released data products and must first reproduce these central values and uncertainties closely. Otherwise every later comparison is suspect.
The strongest existing competitor is the same fit plus a free per-bin shape nuisance with a Gaussian prior, both shared and per detector. This is the conventional way to absorb unknown distortions, and it is equivalent to our model with $K=0$ and an inflated $\sigma_0$. It gets the same data, the same sampler, and the same budget for tuning its prior width: an equal-sized grid of widths selected on the same held-out periods.
The manual-cut baseline excludes the 4–6 MeV window and, per period, the low-energy region affected by muon-x.
The ablations are designed to isolate the two key ingredients. The first removes the baseline constraint by giving the mixture full per-detector freedom, including over the baseline-dependent pattern. We expect this version to absorb oscillation and bias $\theta_{13}$, which would demonstrate the Gap-4 danger directly. The second replaces the mixture with a single Student-t residual.
We count "beating" a baseline as follows. In injection tests, bias in $\sin^2 2\theta_{13}$ must drop below 0.3 of the statistical error while 68% and 95% coverage stays within Monte Carlo error of nominal. In clean toys, the uncertainty may inflate by no more than 10% relative to the naive fit. I found no reported injection-test result on Lacuna for any of these baselines, so all baseline numbers beyond the published central fit will be produced by us.
Why not just…?
Why not just trust the near-far ratio, which already cancels flux-shape errors like the 5 MeV bump? For shared distortions this is largely correct, and we expect our shared term $s_i$ to change little there. We will report that honestly as a null result. The method earns its place only on distortions that do not cancel, such as detector-specific, time-varying backgrounds like muon-x or relative energy-scale drifts. The injection suite is built around exactly those cases.
Why not just add a Gaussian shape nuisance per bin? This is the real threat. A Gaussian prior wide enough to absorb a localized bump also dilutes the information in every clean bin. A narrow one leaves the bump to bias the fit. A mixture avoids this by being narrow for most bins and wide for a few. The matched-tuning comparison above settles the question. If the tuned Gaussian nuisance reaches the same bias–precision curve, the mixture is unnecessary and we will say so.
Why not just use a heavy-tailed Student-t residual, or cut bad bins by hand? The Student-t is a one-parameter version of the same idea and is included as an ablation. If it matches the mixture, we adopt it. Hand cuts need the distortion's location to be known in advance. The instability Potamitis et al. report for hard SNR thresholds suggests binary exclusion also behaves badly near threshold. We test this directly by injecting distortions at locations the cut does not anticipate.
Proposed Experiments
The first check is reproduction: the naive fit must match the published Daya Bay full-data values.
The second is a toy study built from the Daya Bay prediction with realistic statistics for all eight detectors and all data periods. We inject four kinds of distortion: a shared bump, a detector-specific bump, a slowly growing low-energy background in the far hall, and a relative energy-scale drift in one detector. Each kind is injected at several amplitudes. For every method we measure bias, interval width, and coverage over at least several hundred toys. The GP surrogate of Li et al. is used if full Feldman-Cousins coverage is too expensive.
The decisive result is the bias-versus-precision curve as the width of the detector-specific component grows. It separates three outcomes: the structured mixture dominates the tuned Gaussian nuisance, it merely ties with it, or it absorbs oscillation once the width passes some threshold.
The third experiment is on real data. We fit the released spectra and report three things: the shift in $\sin^2 2\theta_{13}$ and $\Delta m^2_{32}$ relative to the published values, the uncertainty change, and the learned bin-trust map over energy and time. A map that singles out the known muon-x periods and the 5 MeV region without being told about them would be strong qualitative confirmation. A map that flags bins near the oscillation minimum would be a warning sign that the constraint is leaking.
The fourth experiment is a stress case. We refit using only the far hall plus a single near hall and repeat with the n-H capture sample (Daya Bay n-H analysis), to check whether weaker baseline leverage makes the method inflate its uncertainty, as it should, rather than bias its estimates. Only if these reactor tests pass would we carry the approach to the single-site regime of SNO+ or to long-baseline cross-section mismodeling.
References
- DOA estimation of speech signals using semi-blind source separation techniques
- Interference robust DOA estimation of human speech by exploiting historical information and temporal correlation
- Robust Principal Component Analysis with Complex Noise
- Denoising Hyperspectral Image With Non-i.i.d. Noise Structure
- Full Data Release of the Daya Bay Reactor Neutrino Experiment
- Improved measurement of the reactor antineutrino flux at Daya Bay
- Joint neutrino oscillation analysis from the T2K and NOvA experiments
- Expanding neutrino oscillation parameter measurements in NOvA using a Bayesian approach
- Efficient Neutrino Oscillation Parameter Inference with Gaussian Process
- Measurement of Electron Antineutrino Oscillation Amplitude and Frequency via Neutron Capture on Hydrogen at Daya Bay
- Measurement of Reactor Antineutrino Oscillation at <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mrow><mml:mi>SNO</mml:mi><mml:mo>+</mml:mo></mml:mrow></mml:math>