Neutrino Atlas
EN
  1. Terms
  2. Analysis & measurement
  3. Statistics & inference
  4. Maximum-likelihood fit

Maximum-likelihood fit

ConceptLevel 3 of 4Also called: χ² fit, profile likelihood

Finding the parameter values that make the observed data most probable under a model.

Binned or unbinned likelihoods combine Poisson event counts with Gaussian penalty terms for systematic nuisance parameters. The test statistic −2Δln L is used to build confidence regions, for example Δχ² = 2.71 for a one-parameter 90% CL. This is standard in all modern neutrino analyses.

In plain words

Every set of parameter values, such as a mixing angle and a mass splitting, predicts how many events should appear in each energy bin. The likelihood measures how probable the observed data are under that prediction. A maximum-likelihood fit tries all reasonable parameter values and keeps the ones that make the data least surprising. The way the likelihood falls off around that best point tells you the uncertainty.

In detail

−2 ln L = 2 Σᵢ [ μᵢ(θ, η) − nᵢ + nᵢ ln(nᵢ / μᵢ) ] + Σₖ (ηₖ − η̂ₖ)² / σₖ²nᵢ: observed events in bin i; μᵢ: predicted events; θ: physics parameters; ηₖ: nuisance parameters with external estimates η̂ₖ and uncertainties σₖ.

The first sum is the Poisson part, which reduces to a familiar χ² when counts are large. The second sum is the penalty for moving the systematic nuisance parameters away from their measured values. To quote a result for θ alone, the fit is minimised over all η at each value of θ, the profile likelihood. Unbinned fits use each event's measured properties directly and are preferred when events are few.

From likelihood to intervals

By Wilks' theorem, Δχ² = −2Δln L follows a χ² distribution, giving standard thresholds:

  • One parameter: 1.00 (68.3%), 2.71 (90%), 3.84 (95%), 9.00 (3σ).
  • Two parameters: 2.30 (68.3%), 4.61 (90%), 5.99 (95%), 11.83 (3σ).

The theorem fails near physical boundaries or for periodic parameters such as δCP, and for small samples. Experiments such as T2K and NOvA then calibrate the thresholds with pseudo-experiments in the spirit of Feldman–Cousins.

Why it matters

Almost every modern neutrino result, from atmospheric zenith-angle distributions at Super-Kamiokande to the beta spectrum at KATRIN, comes out of a likelihood fit. Combining the likelihoods of several experiments gives a global fit.

Frequently asked questions

What is the difference between a χ² fit and a likelihood fit?

For large event counts with Gaussian errors they are equivalent. For small counts the Poisson likelihood is correct, while a simple χ² can be biased.

Why are nuisance parameters needed?

They let the fit account for imperfect knowledge of flux, cross sections and detector response, so that the final uncertainty on the physics parameters includes these effects.

Original papers

  1. TheoryG. J. Feldman and R. D. Cousins, Unified approach to the classical statistical analysis of small signals, Phys. Rev. D 57, 3873 (1998)DOI 10.1103/PhysRevD.57.3873arXiv:physics/9711021

Go deeper

  • Global fitA combined analysis of results from many experiments to determine the oscillation parameters jointly.

Related terms

Mentioned in 10 pages

Last updated: · Editorial standard

Cite this page

Text

Neutrino Atlas (2026). “Maximum-likelihood fit”. Retrieved 22 November 2026, from https://neutrino-atlas.com/terms/likelihood-fit/

BibTeX

@misc{neutrinoatlas2026termslikelihoodfit,
  title        = {Maximum-likelihood fit},
  author       = {{Neutrino Atlas}},
  year         = {2026},
  url          = {https://neutrino-atlas.com/terms/likelihood-fit/},
  note         = {Accessed 2026-11-22},
  language     = {english}
}