Jorge Quitério Search
Mathematics

The exponential power family and the γ-order generalized normal

The γ-order generalized normal has the same univariate shape as the exponential power family known since 1923. It differs in the origin of that shape and, in several dimensions, in the type of dependence between components. A translator between the two parametrisations is included.

When I first studied the γ-order generalized normal distribution (γGN), introduced in the γ-order generalized normal distribution, an obvious question arose: whether it is simply the exponential power distribution under a different name.

In one dimension the curves coincide exactly. The two families, however, have different origins, and the extension to several variables involves a modelling choice that matters far more than the name. This article sets out the correspondence between the two parametrisations and explains where the real difference lies.

The family has been rediscovered and renamed several times over a century.

  • Subbotin (1923), an astronomer, generalised Gauss's law of errors to densities proportional to exp⁡{−∣x−μ∣β/sβ}\exp\{-|x-\mu|^\beta / s^\beta\}.
  • Box and Tiao (1973) called it the exponential power distribution and used it to test how robust Bayesian conclusions are to the normal assumption.
  • Nelson (1991) used it, as the generalized error distribution, for the shocks in his EGARCH volatility model. Signal processing calls it the generalized Gaussian, and uses it to model the coefficients of image and audio transforms, which are sharply peaked at zero with heavy tails.
  • Gómez, Gómez-Villegas and Marín (1998) studied its elliptical multivariate version.
  • Nadarajah (2005) published a "generalized normal distribution" of the same shape, and Mineo and Ruggieri (2005) the R package normalp. Python's SciPy implements it as gennorm.
  • Kitsos and Tavoularis (2009) obtained the γGN from a generalized logarithmic Sobolev inequality, and Kitsos and Stamatiou (2025) reviewed the family and made the correspondence with the exponential power distribution explicit.

The γGN in the same coordinates

The univariate γGN with location μ\mu, scale σ\sigma and shape γ\gamma has density

f(x)=Cγσ exp⁡{−γ−1γ∣x−μσ∣γ/(γ−1)}. f(x) = \frac{C_\gamma}{\sigma}\,\exp\left\{-\frac{\gamma-1}{\gamma}\left|\frac{x-\mu}{\sigma}\right|^{\gamma/(\gamma-1)}\right\} .

Set β=γ/(γ−1)\beta = \gamma/(\gamma-1) and absorb the factor (γ−1)/γ(\gamma-1)/\gamma into the scale, and this is Subbotin's density. The dictionary:

γGN shape γ\gamma Exponent β\beta Member
γ→1+\gamma \to 1^{+} β→∞\beta \to \infty uniform
1<γ<21 < \gamma < 2 β>2\beta > 2 lighter tails than normal
γ=2\gamma = 2 β=2\beta = 2 normal
γ>2\gamma > 2 1<β<21 < \beta < 2 heavier tails than normal
γ→±∞\gamma \to \pm\infty β→1\beta \to 1 Laplace
γ<0\gamma < 0 0<β<10 < \beta < 1 heavier tails than Laplace

The γGN parameter space is the real line without [0,1][0, 1], and it covers every β>0\beta > 0. The negative values of γ\gamma are simply the very heavy-tailed members. Kitsos and Stamatiou show how extreme they get: for the standardised member with γ=−0.1\gamma = -0.1, the probability below −3 is about 0.17, against 0.0013 for the normal.

This is useful for software. Anything written for the exponential power distribution works for the γGN after converting β=γ/(γ−1)\beta = \gamma/(\gamma-1) and the scale s=σ ((γ−1)/γ)−(γ−1)/γs = \sigma\,\big((\gamma-1)/\gamma\big)^{-(\gamma-1)/\gamma}.

Difference in origin

Subbotin's family is a natural generalisation: replace the square in Gauss's exponent by any power. The γGN has a different origin. Kitsos and Tavoularis obtained it as the family of extremal functions of a generalized, γ-order logarithmic Sobolev inequality, the inequality that in its classical form singles out the normal. Read that way, the exponent γ/(γ−1)\gamma/(\gamma-1) is not a free choice but the conjugate exponent the inequality produces, and the constant (γ−1)/γ(\gamma-1)/\gamma follows from the same calculation. A later note in this series covers the inequality itself.

Kitsos and Stamatiou argue that this grounding is what separates the γGN from generalisations "obtained by trial and error". The significance of this is partly a matter of judgement, but it has a practical consequence: entropy, Fisher information and the Kullback-Leibler divergence take simple forms on the γGN, because the family was built from them.

Difference in the multivariate case

There are two natural ways to make a vector out of a univariate shape:

  • Product: multiply univariate densities, so the components are independent by construction.
  • Elliptical: let the density depend on the data only through the Mahalanobis distance.

The γGN takes the elliptical route by construction, because its inequality lives directly in several dimensions:

f(x)∝∣Σ∣−1/2exp⁡{−γ−1γ [(x−μ)⊤Σ−1(x−μ)]γ/(2(γ−1))}. f(\mathbf{x}) \propto |\Sigma|^{-1/2} \exp\left\{-\frac{\gamma-1}{\gamma}\,\big[(\mathbf{x}-\boldsymbol\mu)^\top \Sigma^{-1} (\mathbf{x}-\boldsymbol\mu)\big]^{\gamma/(2(\gamma-1))}\right\} .

Up to parametrisation this is the elliptical power exponential of Gómez and colleagues; both belong to the Kotz family. The two constructions agree only for the normal. Otherwise they describe different dependence structures: in one, a large shock in one coordinate says nothing about the others; in the other, large shocks tend to occur together.

Sensors give a concrete case of each. Two thermometers in different rooms have independent electronic noise, and the product form describes them. Two accelerometers on the same vehicle both register the same pothole, so a large reading on one makes a large reading on the other more likely, even if their readings are uncorrelated on average; the elliptical form describes them. The name of the family does not settle which applies: the physical situation does.

The translator below illustrates three properties: every exponent β>0\beta > 0 corresponds to exactly one γ outside [0,1][0, 1]; the map runs off to infinity at the Laplace member; and tails get heavier as β falls.

The top panel draws γ=β/(β−1)\gamma = \beta/(\beta - 1); the bottom panel draws the density with unit variance so that only the shape changes.

Suggested settings:

  1. Presets. Uniform-ish, normal, γ = 4, Laplace, γ = −1. Note the probability of landing beyond three standard deviations: about 0.27% for the normal, 0.90% for γ = 4, 1.44% for the Laplace and 2.18% for γ = −1.
  2. Crossing β = 1. Move β slowly from 1.2 down to 0.8. The dot on the top panel rises to +∞+\infty and comes back from −∞-\infty, while the density below changes smoothly. The curve changes little while the label changes greatly; this is why estimating γ directly near the Laplace member is unstable.
  3. Correspondence. Find the β that gives γ = 3. The slider lands at β ≈ 1.5, and 1/γ=1−1/β1/\gamma = 1 - 1/\beta confirms it: 1/3=1−2/31/3 = 1 - 2/3.

Properties of the map. From β=γ/(γ−1)\beta = \gamma/(\gamma - 1) it follows that 1/β=1−1/γ1/\beta = 1 - 1/\gamma, so 1/γ=1−1/β1/\gamma = 1 - 1/\beta: γ and β are Hölder conjugates, 1/γ+1/β=11/\gamma + 1/\beta = 1. The map β↦1−1/β\beta \mapsto 1 - 1/\beta is strictly increasing from (0,∞)(0, \infty) onto (−∞,1)(-\infty, 1); taking reciprocals sends (0,1)(0, 1) to (1,∞)(1, \infty) and (−∞,0)(-\infty, 0) to (−∞,0)(-\infty, 0), and the value 1/γ=01/\gamma = 0 (γ infinite) corresponds to β = 1. So every β > 0 has exactly one γ in R∖[0,1]\mathbb{R} \setminus [0, 1], plus the Laplace limit. The tail readout uses the closed form P(∣X∣>c)=Q(1/β,(c/s)β)P(|X| > c) = Q(1/\beta, (c/s)^\beta), with QQ the regularised upper incomplete gamma function.

Practical implications:

  • Use the existing tools. In one dimension, decades of work on the exponential power family, its estimation and its software transfer to the γGN.
  • Mind the parametrisation. The same curve can be indexed by γ or by β, and near the Laplace member only β is well behaved. That matters for estimation, as a later note shows.
  • The multivariate choice is a modelling decision. Independent components or elliptical dependence is a statement about the data, whatever name the family carries.

References

  1. M. T. Subbotin, "On the law of frequency of error", Matematicheskii Sbornik, 31(2), 296-301, 1923.
  2. G. E. P. Box and G. C. Tiao, Bayesian Inference in Statistical Analysis, Addison-Wesley, 1973.
  3. D. B. Nelson, "Conditional heteroskedasticity in asset returns: a new approach", Econometrica, 59(2), 347-370, 1991.
  4. E. Gómez, M. A. Gómez-Villegas and J. M. Marín, "A multivariate generalization of the power exponential family of distributions", Communications in Statistics to Theory and Methods, 27(3), 589-600, 1998.
  5. S. Nadarajah, "A generalized normal distribution", Journal of Applied Statistics, 32(7), 685-694, 2005.
  6. A. M. Mineo and M. Ruggieri, "A software tool for the exponential power distribution: the normalp package", Journal of Statistical Software, 12(4), 1-24, 2005.
  7. C. P. Kitsos and N. K. Tavoularis, "Logarithmic Sobolev inequalities for information measures", IEEE Transactions on Information Theory, 55(6), 2554-2561, 2009.
  8. 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.