How to read a neutrino measurement
Neutrino results arrive with error bars, confidence levels and sigmas, and their meaning is often lost on the way to the headline. This lesson explains how to read them, and why a 6σ signal can still be wrong.
Beginner
Numbers with a warning label
When physicists publish a result, they almost never give a single number. They write something like 7.5 ± 0.2: a central value and an uncertainty. The central value is the best estimate; the uncertainty says how far the truth could plausibly lie from it. Without an uncertainty a result is almost meaningless, because nobody can tell whether a difference from another measurement is real or just noise.
Two kinds of uncertainty
- Statistical uncertainty comes from chance. Neutrino detectors often record only a handful of events per day, and counts fluctuate, just as 100 coin tosses may give 46 heads rather than 50. More data shrink this uncertainty.
- Systematic uncertainty comes from imperfect knowledge of the apparatus or of the physics. A bathroom scale that is slightly off gives a wrong weight however often you step on it. More data do not help here; only a better understanding of the instrument does.
When the answer is “less than”
Sometimes an experiment sees no effect at all and quotes an upper limit. In 2025 KATRIN reported that the neutrino mass is less than 0.45 electronvolts1. This does not mean the mass is 0.45. It means the data are compatible with zero and that larger values are excluded. The phrase “at 90 % confidence level” describes how strict the exclusion is: the method would capture the true value in 90 of 100 repeated experiments. A confidence level of 95 % gives a somewhat looser limit than 90 %, so limits can only be compared at the same level.
How sure is sure? Counting sigmas
For a possible discovery, physicists ask how likely it is that pure chance produced the signal. That probability is called a p-value, and it is usually translated into units of sigma (σ), the typical size of random fluctuations. A 3σ deviation arises by chance about once in 740 tries and counts as “evidence”. A 5σ deviation arises by chance only about once in 3.5 million tries; that is the usual threshold for a “discovery”.
Lessons from neutrino history
- In 2011 the OPERA experiment found that neutrinos travelling from CERN to central Italy seemed to arrive about 60 billionths of a second earlier than light would, at about 6σ3. The cause was a loose fibre-optic cable in the timing system: a systematic error that no amount of extra data could have revealed.
- LSND in the 1990s and later MiniBooNE saw more electron neutrinos than expected; together the excess exceeds 6σ4. Yet the simplest explanation, a new sterile neutrino, is now largely ruled out. A real excess does not prove the proposed explanation.
- Gallium detectors caught too few neutrinos from strong radioactive sources. This gallium anomaly is still unexplained.
Searching in many places
If you look for a signal in a hundred different places, a 3σ fluctuation will turn up somewhere by chance, much as someone in a large crowd will roll three sixes in a row. This look-elsewhere effect means that the significance of a discovery must account for how widely the search looked.
Putting all data together
No single experiment determines all neutrino properties. Groups such as NuFIT combine hundreds of measurements in global fits5. When this site gives a precise value for a mixing angle or a mass splitting, it usually comes from such a fit, and the ± tells you how well all data together pin it down.
Quiz
Test your knowledge
4 questions on this level · Print as worksheet
Intermediate
Error bars and the Gaussian
Most uncertainties are quoted as one standard deviation, 1σ. If errors follow a Gaussian distribution, the interval “central value ± 1σ” contains the true value in about 68 % of repeated experiments, ±2σ in about 95 % and ±3σ in 99.7 %. Asymmetric errors are given separately, as in KATRIN's best fit mβ² = −0.14 (+0.13, −0.15) eV²1. Results often split the error into parts, as in Daya Bay's 2012 value sin²2θ₁₃ = 0.092 ± 0.016 (stat) ± 0.005 (syst). Independent uncertainties are added in quadrature, as the square root of the sum of their squares.
Where the uncertainties come from
Event counts follow Poisson statistics: if N events are expected, the fluctuation is about √N. With 100 events the relative statistical uncertainty is 10 %, with 10,000 events 1 %. Systematic uncertainties come from energy scales, background estimates, neutrino fluxes and, in accelerator experiments, especially cross sections. Near detectors that measure the beam before oscillation exist largely to cancel them.
Upper limits and physical boundaries
KATRIN fits mβ², the quantity that shapes the spectrum. If the true mass is zero, fluctuations push the best fit below zero half of the time, and for a small mass negative values remain common; −0.14 eV² is such a case. A naive interval would then be empty or absurdly narrow. The Feldman–Cousins method, published by Gary Feldman and Robert Cousins in 1998, builds intervals that respect the physical boundary and switch automatically from an upper limit to a two-sided interval as a signal grows2. Boundary-aware frequentist constructions of this kind underlie limits such as mβ < 0.45 eV at 90 % CL. Cosmological limits such as DESI's Σmν < 0.0642 eV (95 %) are Bayesian instead and depend on a prior, here Σmν > 0.
Significance and p-values
For a claimed signal the question is: without any effect, how often would fluctuations produce data at least this extreme? That probability is the p-value, translated into a significance Z via the one-sided Gaussian tail: 3σ means p ≈ 1.3 × 10⁻³, 5σ means p ≈ 2.9 × 10⁻⁷. A p-value is not the probability that the hypothesis is wrong. Daya Bay's 5.2σ observation of non-zero θ₁₃ crossed the discovery threshold; T2K's 2020 result, excluding CP conservation at about 95 % confidence, did not.
When significance misleads
- OPERA, 2011: neutrinos crossing about 730 km arrived roughly 60 ns early, at about 6σ. In 2012 a loose fibre connector and a clock fault were found; corrected measurements by OPERA and ICARUS agreed with the speed of light3. The σ only covered known uncertainties.
- LSND reported a 3.8σ ν̄e excess in 2001; MiniBooNE reached 4.8σ in 2018 and 6.0σ combined with LSND4. MicroBooNE found no matching νe excess. The excess may be real while its interpretation stays open.
- SAGE, GALLEX and BEST measured about 20 % fewer captures than expected from intense sources. Interpreted as a sterile neutrino, this gallium anomaly is in 4–5σ tension with reactor and solar data6.
Look-elsewhere effect and blind analyses
When a search scans many energies, sky positions or candidate sources, a large fluctuation somewhere becomes likely. The local p-value at the best spot must be converted into a global one with a trials factor; IceCube's 2022 evidence for neutrinos from the galaxy NGC 1068 is quoted at 4.2σ after this correction. In blind analyses the signal region stays hidden until all selection criteria are fixed, which prevents unconscious tuning.
Global fits
Global fits such as NuFIT, whose authors include Concha González-García and Thomas Schwetz, combine solar, atmospheric, reactor and accelerator data in one likelihood and publish allowed ranges for all oscillation parameters5.
Quiz
Test your knowledge
4 questions on this level · Print as worksheet
Expert
Likelihoods and profiled systematics
Modern neutrino analyses are built on a likelihood L(μ, θ) with parameters of interest μ, such as mβ², δCP or sin²θ₂₃, and nuisance parameters θ that encode systematic effects: flux normalisations, energy scales, cross-section model parameters, constrained by near-detector or external data. The nuisance parameters are profiled out:
By Wilks' theorem Δχ² follows a χ² distribution with k degrees of freedom if the model is regular, the hypotheses are nested and the true value lies away from boundaries. For k = 1 the thresholds are Δχ² = 1, 4 and 9 for 1, 2 and 3σ, and 2.71 for 90 % CL; for k = 2 they are 2.30, 4.61 and 5.99 for 68.3, 90 and 95 %. Global fits quote their allowed ranges this way5.
When Wilks fails
Neutrino physics violates these conditions routinely: physical boundaries (mβ² ≥ 0, Σmν ≥ 0), a periodic parameter (δCP), strongly non-linear dependence near maximal θ₂₃ (the octant) and discrete hypotheses (the mass ordering). The remedy is a full Neyman construction with pseudo-experiments: for each true μ, simulate many experiments, compute the test statistic and find the critical value containing the desired fraction. Feldman and Cousins order outcomes by a likelihood ratio2:
T2K and NOvA apply such corrections to δCP intervals. For the ordering, the Δχ² between the best fits in NO and IO does not follow a χ² distribution; it is approximately Gaussian with mean ±T₀ and width 2√T₀, so the naive √Δχ² is only a rough guide. NuFIT's Δχ² = 6.1 against the inverted ordering, with Super-Kamiokande atmospheric data5, should be read in that light.
Limits from negative best fits
KATRIN's likelihood in mβ² is close to Gaussian and extends to negative values. From 259 days the best fit is mβ² = −0.14 (+0.13, −0.15) eV²1; a frequentist construction with the boundary at zero yields mβ < 0.45 eV (90 % CL). Because the fluctuation is downward, the limit is somewhat stronger than the median expected limit, the sensitivity, which is why collaborations quote both. A Bayesian analysis instead integrates the posterior over mβ² ≥ 0 under a chosen prior. Cosmological limits such as DESI's Σmν < 0.0642 eV are of this type, and changing the prior, for example imposing the oscillation floor, moves them.
p-values and the trials factor
For N effectively independent trials, p_global = 1 − (1 − p_local)ᴺ ≈ N p_local. For continuous scans, the Gross–Vitells method estimates the trials factor from the number of upcrossings of the test statistic. IceCube's search in a catalogue of 110 candidate sources applies such a correction, and NGC 1068 is quoted at 4.2σ post-trial.
Case studies
OPERA's 2011 time-of-flight result, about 60 ns early at about 6σ, carried a total uncertainty near 10 ns; the unrecognised fibre-connector offset was several times larger3. The LSND–MiniBooNE excess, 6.0σ combined4, shows that a significance measures incompatibility with a background model, not the truth of a particular signal model: a mis-modelled photon background yields the same σ as a sterile neutrino. For the gallium anomaly the parameter goodness of fit, which compares the combined minimum with the sum of individual minima, puts the 3+1 interpretation in 4–5σ tension with reactor, β-decay and solar data6, pointing to cross-section or source systematics or to physics beyond one sterile state.
Quiz
Test your knowledge
4 questions on this level · Print as worksheet
Open questions
Statistics in neutrino physics is not fully settled. How to quote the significance of the mass ordering, a choice between two discrete hypotheses, and how to compare frequentist laboratory limits with Bayesian cosmological ones remain debated, and different conventions can shift a quoted σ noticeably.
Several long-standing anomalies, the MiniBooNE low-energy excess and the gallium deficit among them, are statistically significant but lack an accepted explanation. Whether they stem from unrecognised systematics, nuclear-physics inputs or new physics is still open.
Terms to explore
- Statistical uncertainty
- Systematic uncertainty
- Confidence level (CL)
- Upper limit / exclusion
- Feldman–Cousins method
- Statistical significance (σ)
- p-value
- Look-elsewhere effect
- Global oscillation fits
- Blind analysis
People
- Gary J. Feldman
- M. Concepción (Concha) González-García
- Thomas Schwetz
- Eligio Lisi
- William C. Louis
- Janet M. Conrad
On the timeline
- 1995LSND reports an unexpected oscillation signal
- 2011OPERA neutrino-speed anomaly and its resolution
- 2012Daya Bay discovers a non-zero θ₁₃
- 2018MiniBooNE's excess grows to 4.8σ
- 2022BEST strengthens the gallium anomaly
- 2025KATRIN: neutrinos weigh less than 0.45 eV
References
- M. Aker et al. (KATRIN Collaboration), Direct neutrino-mass measurement based on 259 days of KATRIN data, Science 388, 180 (2025), arXiv:2406.13516
- G. J. Feldman, R. D. Cousins, Unified approach to the classical statistical analysis of small signals, Phys. Rev. D 57, 3873 (1998), arXiv:physics/9711021
- 2011 OPERA faster-than-light neutrino anomaly, Wikipedia, en.wikipedia.org/wiki/2011_OPERA_faster-than-light_neutrino_anomaly
- A. A. Aguilar-Arevalo et al. (MiniBooNE Collaboration), Significant excess of electronlike events in the MiniBooNE short-baseline neutrino experiment, Phys. Rev. Lett. 121, 221801 (2018), arXiv:1805.12028
- I. Esteban, M. C. Gonzalez-Garcia, M. Maltoni, I. Martinez-Soler, J. P. Pinheiro, T. Schwetz, NuFit-6.0: updated global analysis of three-flavor neutrino oscillations, JHEP 12, 216 (2024), arXiv:2410.05380
- C. Giunti et al., Gallium anomaly: critical view from the global picture of ν_e and ν̄_e disappearance, JHEP 10 (2022) 164, doi:10.1007/JHEP10(2022)164
Last updated: · Editorial standard
Cite this page
Text
Neutrino Atlas (2026). “How to read a neutrino measurement”. Retrieved 22 November 2026, from https://neutrino-atlas.com/learn/reading-a-measurement/BibTeX
@misc{neutrinoatlas2026learnreadingameasurement,
title = {How to read a neutrino measurement},
author = {{Neutrino Atlas}},
year = {2026},
url = {https://neutrino-atlas.com/learn/reading-a-measurement/},
note = {Accessed 2026-11-22},
language = {english}
}