Jorge Quitério Search
Mathematics

The γ-order generalized normal distribution

One additional parameter allows a model's errors to have lighter or heavier tails than the normal distribution. This article presents the origin and properties of the γ-order generalized normal, with two interactive figures showing the behaviour of the tails.

The normal distribution is the most widely used distribution in statistics. It is also frequently an inadequate model.

Daily returns of stock indices and exchange rates have heavier tails than the normal allows: days that a Gaussian model calls "once in a thousand years" turn up every few years. Measurement noise from some physical sensors goes the other way, with tails lighter than the normal. In both cases the centre of the data is well described; the discrepancy is in the tails.

Two everyday cases mark the ends of the range. A kitchen scale that displays whole grams makes a rounding error that is never larger than half a gram and is equally likely anywhere within that range: the error has no tails at all. At the other end, a saver holding a stock-index fund sees mostly small daily movements and, every few years, a day that a normal model treats as practically impossible. On 19 October 1987 the Dow Jones Industrial Average fell by about 22.6% in a single day, which is more than twenty standard deviations of a typical day's movement at the time.

This article introduces a family of distributions designed for that situation, the γ-order generalized normal (γGN). It contains the normal as a special case and adds a single parameter for the shape of the tails. This article examines what that parameter does and how the difference appears in data.

The relevant literature is as follows.

Generalising Gauss. The idea of replacing the square in the normal's exponent with another power is a century old; the astronomer M. T. Subbotin proposed such a law of errors in 1923. A separate note in this series compares those relatives.

The γGN. Kitsos and Tavoularis introduced the γGN in 2009, originally under the name hyper normal. It was not obtained by adding a parameter for flexibility. It arises from a generalised logarithmic Sobolev inequality, a result linking entropy with Fisher's information. In the classical inequality the normal is the case of equality; in the generalised, γ-order version the γGN family is. Kitsos and Stamatiou (2025) review that origin and argue it is what distinguishes the family from other generalisations.

Tools built on it. Maximum likelihood estimation for the family was worked out by Kitsos, Vassiliadis and Toulias (2014); and the Hellinger distance between members by Kitsos and Toulias (2017).

The density

For location μ\mu, scale σ>0\sigma > 0 and shape γ\gamma:

f(x)=Cγσ exp⁡{−γ−1γ∣x−μσ∣γ/(γ−1)},Cγ=(γ−1γ)(γ−1)/γ2 Γ ⁣(γ−1γ+1). f(x) = \frac{C_\gamma}{\sigma}\,\exp\left\{-\frac{\gamma-1}{\gamma}\left|\frac{x-\mu}{\sigma}\right|^{\gamma/(\gamma-1)}\right\}, \qquad C_\gamma = \frac{\left(\frac{\gamma-1}{\gamma}\right)^{(\gamma-1)/\gamma}}{2\,\Gamma\!\left(\frac{\gamma-1}{\gamma}+1\right)} .

The behaviour is determined by the exponent γ/(γ−1)\gamma/(\gamma-1) applied to the standardised distance from the centre:

  • at γ=2\gamma = 2 the exponent is 2 and the density is the normal;
  • as γ→1+\gamma \to 1^{+} the exponent grows without bound and the density flattens into the uniform on [μ−σ, μ+σ][\mu-\sigma,\ \mu+\sigma];
  • as γ→∞\gamma \to \infty the exponent falls to 1 and the density becomes the Laplace, with heavier tails than the normal.

So 1<γ<21 < \gamma < 2 gives lighter tails than the normal, and γ>2\gamma > 2 heavier ones. (Negative values of γ\gamma are also allowed and give tails heavier still; They are discussed in a later note.)

Measures of shape

The excess kurtosis, a standard measure of tail weight, runs from −1.2 at the uniform end, through 0 at the normal, to +3 at the Laplace end. A more intuitive measure is the probability of landing more than three standard deviations from the centre: 0.27% for the normal, zero for the uniform, and about 1.4% for the Laplace, roughly five times the normal value.

Importance of the tails

Forecasts, risk measures and control rules usually assume normal shocks somewhere. When the shocks are heavier-tailed:

  • prediction intervals that are accurate in the centre are too narrow in the tails, where they are used to plan for adverse outcomes;
  • risk measures derived from the normal understate extreme losses;
  • estimators tuned to normal errors, least squares above all, give occasional extreme observations excessive weight.

Part of these heavy tails comes from volatility that changes over time, as described in volatility clustering and the GARCH model; part remains even after that is accounted for.

For the index-fund holder, the consequence is that a risk estimate built on the normal understates how often a large fall occurs, which is the figure that matters when deciding how much of one's savings to keep in shares. For the kitchen scale, a normal model overstates the chance of an error of a full gram, which cannot happen.

A family such as the γGN allows the data to determine the weight of the tails, and because the normal is a member rather than a limit, every result has an internal check: at γ=2\gamma = 2 it should reduce to the classical one.

The two figures below illustrate two properties. First, one parameter moves the shape continuously from uniform through normal to Laplace. Second, heavier tails produce more extreme observations in data, by the amount the formula predicts.

  1. Moving from "Uniform" to "Laplace" changes the excess kurtosis from −1.2 to +3.
  2. Switch on the log scale. In the middle the curves look alike; in the tails they differ by orders of magnitude.

Samples from the distribution:

  1. Extreme observations. At γ=4\gamma = 4, theory says about 0.9% of observations fall beyond 3 standard deviations, against 0.27% for the normal. In samples of 2,000, roughly eight years of daily market data, the count varies around 18 rather than 5.
  2. Lighter than normal. Set γ=1.3\gamma = 1.3. The sample hardly ever produces a 3-SD observation, and the histogram has "shoulders" instead of tails, as rounding errors do.
  3. Small samples. Reduce the sample size to 100 and press "New sample" repeatedly. Sometimes a heavy-tailed sample shows no extremes at all. This is why estimating tail shape takes a lot of data, a topic for a later note.

Exact sampling. If XX has the γGN density with location 0 and scale 1, then G=γ−1γ∣X∣γ/(γ−1)G = \frac{\gamma-1}{\gamma}|X|^{\gamma/(\gamma-1)} follows a gamma distribution with shape (γ−1)/γ(\gamma-1)/\gamma. The widget draws GG, inverts the transformation and attaches a random sign, so each value comes from the γGN exactly, not approximately.

The γGN adds one parameter to the normal distribution, and that parameter controls the tails, where departures from normality usually matter most. Because it comes from an information inequality, it arrives with its information measures already worked out, and its normal member gives every result a reference point. The practical lesson is that data which appear normal in the centre should be checked in the tails before a normal model is adopted.

References

  1. M. T. Subbotin, "On the law of frequency of error", Matematicheskii Sbornik, 31(2), 296-301, 1923.
  2. C. P. Kitsos and N. K. Tavoularis, "Logarithmic Sobolev inequalities for information measures", IEEE Transactions on Information Theory, 55(6), 2554-2561, 2009.
  3. C. P. Kitsos, V. G. Vassiliadis and T. L. Toulias, "MLE for the γ-order generalized normal distribution", Discussiones Mathematicae Probability and Statistics, 34(1-2), 143-158, 2014.
  4. C. P. Kitsos and T. L. Toulias, "Hellinger distance between generalized normal distributions", British Journal of Mathematics & Computer Science, 21(2), 1-16, 2017.
  5. C. P. Kitsos and I. S. Stamatiou, "On the emerged from LSI γ-order generalized normal distribution", Journal of the Indian Society for Probability and Statistics, 26(2), 737-762, 2025.