Unbinned Maximum Likelihood Estimation for Signal Extraction

304 concepts291 papers25 proposals

When searching for rare subatomic particles or faint astrophysical phenomena, data is often scarce and precious. Every event recorded by a detector carries a wealth of continuous information: an exact energy, a specific arrival time, a precise spatial coordinate, or a unique angular trajectory.

The traditional way to analyze this data is to group it into a histogram—a process known as binning. While histograms are intuitive and excellent for visualization, they are fundamentally destructive to information. Binning forces continuous values into discrete boxes, erasing the exact position of an event within its bin and introducing arbitrary boundary effects. If a rare signal manifests as a narrow peak sitting right on the boundary between two bins, its statistical significance is artificially diluted. Furthermore, when analyzing multiple variables simultaneously, the number of required bins grows exponentially, rapidly leading to empty bins and a severe loss of statistical power.

To maximize the discovery potential of low-statistics datasets, physicists rely on unbinned maximum likelihood estimation. This approach eliminates bins entirely, evaluating the exact continuous properties of every single event against a theoretical model.

Evaluating Every Event Individually

Instead of asking "how many events fell into this box?", an unbinned fit asks "what is the exact probability density of observing an event at this precise coordinate?"

To do this, researchers define a Probability Density Function (PDF) that models the expected distribution of both the signal and the background. The total likelihood of the dataset is the mathematical product of the PDF values evaluated at the exact coordinates of every single observed event. As noted by Y. X. Zhao, this approach preserves vital information that would otherwise be lost in binned data, making it exceptionally powerful for analyzing experimental samples with limited event counts.

By avoiding discretization, the fit retains maximum sensitivity to the fine details of the data distribution. For example, the authors of a study on meson decays emphasize that traditional binning can easily obscure narrow resonances or complex quantum interference patterns, whereas applying a model directly to individual data points maximizes statistical sensitivity.

The Extended Likelihood and Multi-Variable Power

In many physical experiments, the total number of observed events is not fixed beforehand; it is a fluctuating random variable governed by Poisson statistics. An extended maximum likelihood fit incorporates this total event count directly into the likelihood function as an additional Poisson term. This allows the statistical model to simultaneously determine both the parameters of the distribution (such as a particle's mass or lifetime) and the absolute normalization (the exact number of signal and background events).

J. Lee demonstrates that implementing these unbinned extended maximum likelihood fits significantly enhances the sensitivity of rare event searches compared to traditional event-counting techniques by fully exploiting the specific distributions and correlations of kinematic variables.

Signals are rarely clean enough to isolate using a single variable alone. To robustly separate a faint signal from overwhelming noise, physicists look at multiple dimensions at once. Bergfeld et al. leverage this multi-variable approach to distinguish rare, charmless decays from background by calculating probabilities based on the joint distribution of several event characteristics rather than relying on simple, rigid cuts.

Similarly, D’Amico et al. show that incorporating specific event-level variables—such as spatial morphology or energy—into the likelihood function extracts far more information from a dataset than aggregate counting statistics, drastically improving precision in low signal-to-noise regimes.

Accounting for Detector Realities and Joint Datasets

To make an unbinned fit work in practice, the theoretical PDF cannot just represent ideal physics; it must be mapped to what the detector actually sees. This requires integrating instrument response functions, energy resolutions, and detection efficiencies directly into the likelihood model.

Michele Doro outlines how a comprehensive statistical framework can incorporate expected signal shapes, instrument responses, and spatial uncertainties to move beyond simple event counting and maximize detection sensitivity. In a similar vein, Yan et al. demonstrate how an unbinned likelihood approach can fit physical simulation parameters to experimental data by directly accounting for detector-specific variables like photon detection and electron extraction efficiencies.

Once this framework is established, it can be extended to perform simultaneous fits across entirely different decay channels or datasets. Instead of analyzing different experimental runs in isolation, a single global likelihood function can be constructed.

M. N. Achasov et al. utilize this capability to simultaneously analyze spectral data from multiple decay channels and energy points, enabling the precise extraction of resonance parameters. By fitting multiple datasets concurrently without the information loss associated with binning, as Jun Ge demonstrates when comparing particle and antiparticle samples, researchers can directly expose subtle physical asymmetries and extract the absolute maximum scientific value from their data.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 287 papers
  2. Version 1 · 2026-10-04 · 288 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_unbinned_extended_maximum_likelihood_fit_04c524b8,
  title = {unbinned extended maximum likelihood fit},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unbinned-extended-maximum-likelihood-fit/rdr_04c524b809e342d1b505a8af066eb1ba},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_unbinned_extended_maximum_likelihood_fit_04c524b8_v2,
  title = {unbinned extended maximum likelihood fit},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/unbinned-extended-maximum-likelihood-fit/rdr_04c524b809e342d1b505a8af066eb1ba/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake