The process of quantifying the confidence or error associated with a measurement or model prediction.Uncertainty estimation is the process of quantifying how confident we are in a measurement or a model's prediction. In physics, every measurement has some degree of error—whether from instrument limitations, environmental noise, or fundamental quantum limits—and uncertainty estimation provides a formal way to express that error. For models, it answers the question: 'How much can I trust this prediction?' This is essential in fields ranging from climate science to particle physics, where decisions are made based on predictions with inherent uncertainty.
The formal treatment of measurement uncertainty began in the 18th century with the work of Carl Friedrich Gauss on the method of least squares. In the 20th century, the development of Bayesian statistics by Thomas Bayes (published posthumously in 1763) and later by Pierre-Simon Laplace provided a rigorous framework for combining prior knowledge with data. The modern standard for expressing measurement uncertainty was established in 1993 with the ISO Guide to the Expression of Uncertainty in Measurement.
Uncertainty estimation typically involves statistical methods. For measurements, repeated trials yield a mean and a standard deviation, which quantifies random error; systematic errors are estimated through calibration. For models, techniques like Bayesian inference assign probability distributions to model parameters, and Monte Carlo simulations propagate these uncertainties through the model to produce a confidence interval on the output. In machine learning, methods like dropout or ensemble averaging provide uncertainty estimates for neural network predictions.
Used in medical imaging (e.g., diagnosing breast cancer from X-ray shadows), in climate modeling to predict solar flares with confidence intervals, and in particle physics to report the significance of new discoveries.
The 2012 discovery of the Higgs boson was only announced after the uncertainty in the measurement was reduced to a 5-sigma level, meaning there was less than a 1-in-3.5-million chance the signal was a statistical fluke.
