
Choosing the parameter values that make the observed data most probable under the model; the workhorse fit of experimental physics.MLE is the workhorse of data analysis in physics because it is both intuitive and statistically powerful. It provides a principled way to extract the best estimate of a physical quantity from noisy measurements, and it naturally yields uncertainties on those estimates through the curvature of the likelihood function. It is also the foundation for many advanced techniques, such as Bayesian inference, which extends MLE by incorporating prior knowledge. In particle physics, MLE is used to discover new particles by comparing the likelihood of a background-only model with one that includes a signal.
The method was introduced by Ronald Fisher in the 1920s, building on earlier work by Carl Friedrich Gauss and others. Fisher formalized the concept of likelihood and proved many of its desirable statistical properties, making it a cornerstone of modern statistics.
The method begins with a probability distribution that describes how the data are generated given the model parameters. For example, if you are measuring the decay times of a radioactive sample, you might assume they follow an exponential distribution with a rate parameter λ. The likelihood function L(λ) is the product of the probabilities of each observed data point, given λ. To find the maximum, you typically take the logarithm of the likelihood (since it is easier to work with sums than products) and then find the parameter value that makes the derivative zero. This can be done analytically for simple models or numerically for complex ones. The resulting estimate is called the maximum likelihood estimate, and its uncertainty is often derived from the second derivative of the log-likelihood, which gives the Fisher information.
MLE is used in virtually every experimental physics analysis, from fitting particle tracks in detectors to estimating the Hubble constant from supernova data, and it is also widely applied in fields like economics, biology, and machine learning.
The maximum likelihood estimate is not always unbiased—for example, the MLE of the variance of a normal distribution is slightly too small on average—but it is still the most efficient estimator in many cases, meaning it has the smallest possible variance among unbiased estimators.






