Is a nonzero-κ Gaussian an ordinary normal variable?

No. It is a distinct distribution with different tail behavior.

Does symmetry guarantee finite variance?

No. For this form, κ must be below 2/3.

Is β always inverse twice the variance?

No. That familiar relationship applies to the ordinary κ = 0 Gaussian limit, not generally to the deformed family.

AI Glossary / Statistics and probability / Advanced

Kaniadakis Gaussian distribution

Definition

The Kaniadakis Gaussian distribution, or κ-Gaussian, is a symmetric probability family obtained by normalizing a κ-exponential of a negative quadratic expression. A common centered form is f(x) = C expκ(−βx²), where β is positive and C ensures total probability one. It approaches an ordinary Gaussian as κ tends to zero, while nonzero deformation produces power-law tails. Despite its name, a κ-Gaussian with nonzero κ is not an ordinary Gaussian distribution and does not automatically inherit Gaussian moment, independence or closure properties.

What it means

The familiar normal density decays exponentially in squared distance from its center. A κ-Gaussian changes that decay mechanism while retaining symmetry around the center. Its central appearance can resemble a bell curve, yet its remote tails can be much heavier. This makes it a possible model for symmetric residuals with unusually large deviations. A location parameter can shift the distribution, but a location or symmetry center should not be confused with an existing mean unless the relevant integral is finite. The parameter β controls width; its connection to variance depends on κ.

How it works

Define expκ(u) = exp[asinh(κu)/κ] for κ ≠ 0 and use exp(u) at κ = 0. For a centered model, normalize expκ(−βx²) over the entire real line. The usual κ-statistical domain uses 0 < κ < 1; specific applications may impose narrower bounds. Its density tail behaves proportionally to |x|^(−2/κ). Therefore an absolute moment of order r exists only when r < 2/κ − 1. In particular, finite variance requires κ < 2/3, and a finite fourth moment requires κ < 2/5. These requirements must be checked before standardizing to unit variance or discussing kurtosis.

Three practical examples

Residual modeling: compare a κ-Gaussian with a normal model when residuals are roughly symmetric but show more extreme deviations.
Inverse-problem research: derive a likelihood-based objective from a specified κ-Gaussian error model and evaluate its sensitivity to large residuals.
Simulation analysis: test how a procedure behaves under symmetric distributions whose variance or fourth moment may cease to exist. These applications need their own validation and do not establish universal robustness or superiority.

Think of it this way

Think of a bell-shaped profile whose outer shoulders thin more slowly than those of an ordinary normal curve. Looking only near the peak may hide the important difference. The distribution's name points to the quadratic structure and its limiting case, while the tails determine whether familiar quantities such as variance and kurtosis are actually defined. A similar-looking center does not imply interchangeable probabilities far away.

Worked example

Choose κ = 0.5 and β = 1. The ratio of the density at x = 2 to its density at the center is f(2)/f(0) = exp₀.₅(−4), because the normalizing constant cancels. This equals (√5 − 2)², approximately 0.0557281. For the ordinary Gaussian-shaped density with the same β, the corresponding ratio is exp(−4), approximately 0.0183156. These are density ratios, not tail probabilities, and the comparison holds β fixed rather than variance fixed. At κ = 0.5, the κ-Gaussian density has a |x|^−4 tail. Its second moment is finite because 2 < 2/0.5 − 1 = 3, while its fourth moment diverges because 4 is not less than 3. Consequently, ordinary finite kurtosis is unavailable even though a variance can be calculated.

Illustrative case study

Fictional case study: Elm Signal Lab investigates residuals from a measurement reconstruction system. A normal model understates several large errors, so analysts evaluate a κ-Gaussian candidate alongside other symmetric heavy-tailed models. Before comparing standardized residuals, they restrict attention to parameters with finite variance and calculate each model's actual scale relationship. They then evaluate held-out reconstruction behavior and inspect whether the large errors reflect contamination or an omitted physical effect. A better residual fit alone does not prove that the reconstruction model captures the underlying mechanism.

Common misconceptions and mistakes

Common mistakes include copying the normal relationship variance = 1/(2β) when κ is nonzero, assuming a Gaussian name implies Gaussian independence properties, and reading density ratios as event probabilities. The normalizing constant changes with κ and β. Unit-variance formulations require κ-dependent scaling and valid moment conditions. Symmetry does not remove heavy tails or guarantee that every moment exists. Products or sums of independent κ-Gaussian variables need not preserve the same family under ordinary operations. A heavy-tailed likelihood also does not repair omitted structure, dependence or unrepresentative observations by itself.

Business use

Businesses may encounter this distribution in specialist signal, measurement or inverse-problem work. Ask what residual quantity is modeled, why symmetry is plausible and how its tail behavior was checked. Compare relevant alternatives using held-out data, including performance on ordinary observations as well as extreme ones. Document whether scales are matched by β, variance or another measure. Where moments are infinite, use appropriate summaries and clearly explain their interpretation. An application-specific research result should not be presented as a general guarantee for all datasets or operational systems.

Check your understanding

Why can the example have finite variance but no finite kurtosis?

Answer: Variance requires a finite second moment, while standard kurtosis requires a finite fourth moment as well. The |x|^−4 tail satisfies the first requirement but fails the second.

Related terms

Heavy-tailed distribution
Residual
Likelihood
Variance
Kurtosis
Gaussian distribution
κ-exponential

Learn next

calculate the normalization and actual variance for an admissible parameter choice, then compare models at matched variance rather than merely matched β. Derive a likelihood from the chosen density and check whether assumptions about independent residuals fit the application.

Sources

Sources: S. L. E. F. da Silva and G. Kaniadakis, κ-statistics approach to optimal transport waveform inversion, https://iris.polito.it/retrieve/ac200616-bdb7-4fc8-a62f-a0f3605808d5/PRE.2022.daSilva.Kaniadakis.pdf. G. Kaniadakis, New power-law tailed distributions emerging in κ-statistics, https://arxiv.org/abs/2203.01743 (related normalized forms and moments). Example and fictional case are original.

Previous: Kaniadakis Gamma distribution Next: Kaniadakis Weibull distribution