What Is Kurtosis?

Kurtosis is a summary statistic of a probability distribution that describes the weight of its tails relative to the tails of a normal distribution. Formally it is the fourth standardized moment, the expected value of the fourth power of a variable after that variable has been centered on its mean and divided by its standard deviation. Because deviations are raised to the fourth power, observations far from the center dominate the calculation, which makes kurtosis a tail statistic rather than a measure of how sharply peaked a density appears.

Kurtosis sits alongside the mean, the variance, and the skewness as one of the first four moments used to characterize a distribution. The mean fixes location, the variance fixes spread, skewness measures asymmetry, and kurtosis measures the propensity to produce values far from the center. The NIST/SEMATECH e-Handbook of Statistical Methods treats skewness and kurtosis together as the standard numerical tools for judging whether a data set departs from normality, complementing graphical checks such as the normal probability plot.

Definitions and Conventions

Two conventions are in circulation and confusing them is a common source of error. Under the original definition the normal distribution has a kurtosis of exactly 3. Under the excess kurtosis convention, 3 is subtracted so that the normal distribution sits at zero and the sign of the statistic immediately indicates the direction of departure. Distributions with positive excess kurtosis are called leptokurtic and produce outliers more often than a Gaussian of the same variance: the Laplace and Student t distributions are standard examples. Distributions with negative excess kurtosis are platykurtic, with the uniform distribution as the familiar case. Sample estimators add a further complication, since the naive plug-in estimator is biased and several bias-corrected forms are implemented in different statistical packages.

Estimation and Robustness

The fourth moment estimator is extremely sensitive to individual extreme observations, and a single mis-recorded value can shift a sample kurtosis substantially. For heavy-tailed populations the fourth moment may not exist at all, in which case the sample statistic diverges as more data are collected rather than converging on a fixed value. Practitioners therefore pair kurtosis with quantile-based alternatives, or with formal normality tests such as Anderson-Darling and Shapiro-Wilk, before drawing conclusions. Work on independent component analysis under weakened moment assumptions shows how algorithms that traditionally depend on finite fourth moments can be reformulated to survive genuinely heavy-tailed data.

Kurtosis as a Non-Gaussianity Contrast

In signal processing, kurtosis is used less as a descriptive statistic and more as an objective function. Blind source separation methods exploit the fact that a linear mixture of independent sources is closer to Gaussian than the sources themselves, so maximizing the absolute value of excess kurtosis along a projection direction recovers an underlying component. This projection-pursuit idea underpins the kurtosis-based contrast functions used in independent component analysis, and analyses of fourth-moment methods establish the conditions under which such estimators are consistent. Related uses appear in machine condition monitoring, where a rising kurtosis in a vibration signal flags the impulsive shocks produced by an incipient bearing fault, and in image quality assessment, where kurtosis of wavelet coefficients tracks texture.

Applications

Kurtosis has applications in a range of fields, including:

• Quality control and process capability analysis, where non-normality invalidates standard control limits
• Financial risk modeling, where fat tails drive value-at-risk estimates
• Blind source separation and independent component analysis
• Vibration-based fault detection in rotating machinery
• Radar and sonar clutter characterization
• Model selection and goodness-of-fit testing in applied statistics