When can a few neutrino events become evidence?

Begin with a fully specified toy experiment

Imagine a counting experiment with a fixed observing period and a selection decided in advance. Events are independent and rare, and the background mean is known exactly. Under these teaching assumptions a Poisson model describes the count. None of the numbers below are IceCube data.

Suppose five events are observed when the background-only expectation is one. The probability of five or more under that model is about 0.00366. If the expectation is instead three, the same upper-tail probability is about 0.185. The observed count is unchanged; the evidential comparison is very different.

The tail includes outcomes at least as extreme as the observation under the chosen statistic. It is not the probability of exactly five, and it is not the probability that the background hypothesis is true. Real analyses must also justify the assumptions behind that background expectation.

P(N\geq n\mid b)=1-e^{-b}\sum_{k=0}^{n-1}\frac{b^k}{k!}

N is the count, n the observed nonnegative integer and b the exactly known dimensionless background mean; when the observed count is zero, the sum is empty.

Sources: [1]

Measure what can imitate the signal

A useful background estimate comes from a defined control measurement or validated physical model. For coincidence searches, time shifts can estimate accidental pairs. For optical telescopes, entering-muon controls and atmospheric-neutrino predictions constrain different components.

A control region needs a justified transfer into the signal region. An easier-to-measure low-energy sample does not automatically predict a rare high-energy tail. Differences in acceptance, detector period or physical origin can invalidate a simple scaling.

Record both the estimate and how it was obtained. If a later calibration changes the control selection, revisit the transfer and its correlations. A background number without that measurement chain is insufficient to evaluate whether a small excess is persuasive.

Sources: [2] · [3] · [4]

Decide the test before reading the answer

If cuts are adjusted until an excess looks strongest, the final procedure has searched more possibilities than its last plot reveals. Blinding reduces that pressure by withholding a signal region, scrambling information or hiding a fitted parameter while the method is developed.

Specify what is hidden, what control information remains accessible and what permits unblinding. Test reconstruction and selections using simulations or appropriate control samples. Keeping a development record makes it possible to identify choices made before and after seeing the signal.

Blinding cannot fix an incorrect response model or an overlooked systematic effect. It is one safeguard against biased choices, alongside validation and review. Different experiments use different mechanisms; a paper should describe its actual procedure rather than relying on the word “blind”.

Sources: [5] · [6]

Unknown backgrounds need another layer of inference

The simple example treated the background mean as exact. A real control sample has finite statistics, and detector efficiency or atmospheric production may be uncertain. These quantities enter as nuisance parameters constrained by additional measurements.

A frequentist analysis can profile a likelihood over those parameters; a Bayesian analysis can integrate over a specified prior distribution. The resulting tail probability or credible statement is conditional on that complete model. Assigning one arbitrary percentage is not an adequate substitute.

Also preserve correlations. The same optical calibration can alter signal acceptance, energy reconstruction and background prediction together. Checking alternate models and control distributions helps expose an apparent excess driven by a common response mismatch.

Sources: [1] · [4] · [7]

The hottest spot is the winner of a search

Searching one preselected source differs from inspecting an entire sky and reporting its most unusual patch. Even a background-only sky produces a winner. The global test must repeat the same search on background realizations and compare those winners with the observed one.

For 100 independent tests, each with a one-percent false-positive probability, the chance of at least one false positive is about 63 percent. This is another teaching example. Real neighboring sky locations are correlated, so simply counting pixels is not a valid trials correction.

Ask for the full package: selection, background controls, uncertainty model, search scope and replication. A small sample can be compelling when alternative explanations are strongly constrained. A large fitted excess can remain inconclusive when its background or search procedure is poorly specified.

Sources: [1] · [7] · [8]

Try it in the Lab

Supernova Neutrino Burst Observatory

Primary sources & revision

First published 2026-10-11; last revised 2026-10-11. Original explanatory text and diagrams by Sandbox Physics. Illustrations are schematic; no experimental event records are reproduced here. This is an independent educational publication, not an official Nobel or experiment collaboration publication.

Continue the story

Read IceCube’s discovery paper, one figure at a time How does a faint signal become a neutrino discovery? NGC 1068: what does a neutrino excess establish? Open a real neutrino event file and read it carefully