Skip to content

Normal-inverse Gaussian distribution

A four-parameter probability law formed by mixing a normal mean and variance with an inverse Gaussian variable.

Version
v1 · 2026-10-31 · History
Domain-specific #
14538
Domain group
Formal Sciences
Origin domain
Mathematics
Subdomains
Probability Theory, Parametric Distributions → Mathematics
Aliases
NIG distribution

Core Idea

A normal-inverse Gaussian distribution (NIG law) is a continuous probability law obtained by taking a normal distribution whose conditional mean and variance both depend on a positive inverse Gaussian mixing variable, then averaging over that variable.[1] In one standard representation, conditional on V, X is normal with mean μ+βV and variance V, with the parameters of the inverse Gaussian law chosen consistently with the NIG scale and tail parameters.[2] The result is a marginal law for X; this mathematical representation does not by itself identify a measured physical mixing variable.

The conventional four parameters are α for tail shape, β for asymmetry, δ for scale, and μ for location. With γ=√(α²−β²), α>|β|, and δ>0, its density is

f(x)= [αδ K₁(α√(δ²+(x−μ)²)) / (π√(δ²+(x−μ)²))] exp(δγ+β(x−μ)),

where K₁ is a modified Bessel function. This is the λ=−1/2 member of the larger generalized-hyperbolic family.[3][2] The density identifies a family of possible laws, whether or not any data set is fitted to it. A corresponding homogeneous NIG Lévy process can be constructed.[1] Other NIG-marginal processes can have different dependence structures; the one-time law does not choose among them.[4]

How would you explain it like I'm…

 

No faithful explanation at this level. All three agree a five-year-old picture must present a real hidden thing that stretches and shifts a bell curve, collapsing the mathematical mixture representation into a physical mixing process.

Blended Bell Curves

A probability distribution describes which results are likely and which are rare. The normal-inverse Gaussian distribution is built with a math recipe: first a positive random number is chosen from one kind of distribution, and then that number sets both the center and the spread of a bell curve. Averaging over all the possible first numbers gives the final shape. Its four settings control how heavy the tails are, how lopsided it is, how wide it is and where it sits. The recipe is just a way to build the shape — it doesn't mean there's a real hidden number being measured.

Normal Mixed by Inverse Gaussian

The normal-inverse Gaussian (NIG) distribution is a continuous probability distribution built by mixing. Start with a positive random variable V from an inverse Gaussian distribution. Given V, let X be normal with mean μ + βV and variance V. Averaging over all values of V gives the distribution of X alone, called its marginal law. It has four parameters: α (tail shape), β (asymmetry), δ (scale) and μ (location), with α > |β| and δ > 0. This is a mathematical construction; it doesn't by itself mean there is a measurable physical mixing variable. The NIG is one special member of a larger family called generalized hyperbolic distributions.

 

The normal-inverse Gaussian distribution is a continuous law obtained as a normal variance-mean mixture: conditional on a positive inverse Gaussian mixing variable V, X is normal with mean μ + βV and variance V, with the inverse Gaussian parameters chosen consistently with the NIG scale and tail parameters; integrating over V yields the marginal law of X. The representation is mathematical and does not by itself identify a measured physical mixing variable. The conventional parameters are α (tail shape), β (asymmetry), δ (scale) and μ (location), subject to α > |β| and δ > 0, with γ = √(α² − β²). The density is f(x) = [αδ K₁(α√(δ² + (x − μ)²)) / (π√(δ² + (x − μ)²))] exp(δγ + β(x − μ)), where K₁ is a modified Bessel function of the second kind. The NIG is the λ = −1/2 member of the generalized hyperbolic family, and the density defines a family of laws independently of any fitted data. A homogeneous NIG Lévy process can be constructed, but other processes with NIG marginals can have different dependence structures, which the one-time law does not determine.

Structural Signature

  • Real-valued marginal: X is the outcome whose probability density is specified.[1][3] An observed return or physical increment can fill this role, but no empirical data set is needed to define the law.
  • Conditional normal kernel: at a fixed positive V, the conditional law is normal.[2] Substituting another kernel leaves this defining normal/inverse Gaussian mixture representation; whether it yields the same marginal would need a separate proof.
  • Positive inverse Gaussian mixer: V supplies a random conditional variance. Replacing its law with an arbitrary positive distribution leaves a normal mixture, but not generally an NIG law.[1][3]
  • Mean–variance coupling: the conditional mean shifts by βV while conditional variance changes with V; β=0 selects a symmetric member, and nonzero coupling permits asymmetry.[2][3] A nonzero skew is possible, not mandatory.
  • Four-parameter output: integrating out V yields the NIG density whose admissible α,β,δ,μ settings control tail shape, asymmetry, scale, and location.[3] A fitted parameter vector is one member, not the identity of the whole family.

What It Is Not

NIG is narrower than a generic heavy-tailed law or normal mixture. Its defining mixture representation uses a conditional normal kernel and specifically inverse Gaussian positive mixing; the displayed density is the final law-level test.[1][2] A different mixer can retain a two-stage representation while generally producing a different marginal. An inverse Gaussian distribution by itself is the law of the positive mixer, not the resulting real-valued normal mixture.

An NIG process is also a different object: it adds a time index and a rule for joint behavior. Barndorff-Nielsen constructs a homogeneous Lévy process from NIG laws.[1] Other NIG-marginal processes can have different dependence structures.[4] A fit to one-increment density alone establishes neither independent increments nor a forecast model. The empirical studies fit return and turbulence densities.[5][3] Their fits do not reveal an actual inverse Gaussian physical cause inside a market or turbulent flow.

Scope of Application

In probability theory, the law is specified by its density or the equivalent normal/inverse Gaussian mixture, independent of an application. It is a generalized-hyperbolic subclass with shape index λ=−1/2; the four free parameters distinguish members within that class.[3] Statistical modeling can then estimate a member for a real-valued variable and compare its fit with alternatives. This application does not change the definition.

One financial use fits high-frequency IBEX35 index returns. In that study, NIG gave the best overall fit among the generalized-hyperbolic subclasses the authors tested; the result concerns those returns and those comparisons.[5] In a physically different use, WIND solar-wind and simulated magnetohydrodynamic Elsasser increments were modeled with NIG densities whose parameters vary with measurement scale.[3] Both are literal distributional uses, not claims that financial transactions and plasma flows share an observed hidden inverse Gaussian variable.

Clarity

The two words in the name have distinct roles: normal names the conditional kernel; inverse Gaussian names the positive mixing law.[1][2] Neither means “take the inverse of a Gaussian observation.” The class is defined by the resulting marginal law, not by a visual impression of thick tails.

The distinction between a family, a member, a fit, and a process prevents a common slide in reasoning. The family allows many admissible parameter settings; a member fixes them; a fit estimates a member for observed data; a process further specifies how observations relate over time. The IBEX35 and solar-wind papers report empirical fits, while Barndorff-Nielsen's process construction asks an additional temporal question.[5][3][1]

Manages Complexity

The mixture construction gives a two-stage account of a density that can be skewed and has heavier-than-Gaussian tails: sample a positive variance-mean mixer, then sample a conditional normal outcome. Marginalizing those stages yields one four-parameter density.[1][2] This is a calculational and interpretive compression; it is not evidence that the mixer is a physically observed variable.

The four parameters provide a common family for comparing distributions with different tail and asymmetry patterns. The solar-wind study uses the same NIG functional form at multiple scales while allowing fitted parameters to change.[3] Thus a common equation need not imply one universal parameter vector. A compact marginal description still leaves temporal dependence and model adequacy to be tested separately.

Abstract Reasoning

To test a proposed NIG model, first verify the law's defining form: an admissible four-parameter density, or an equivalent normal conditional law with inverse Gaussian mixing.[2][3] A merely non-Gaussian histogram is insufficient. Setting β=0 keeps the model within the family and removes the density's asymmetry about its location.[2] Replacing the inverse Gaussian mixer generally leaves this defining representation.

For data, one may estimate a parameter vector and compare fitted densities or tail behavior within a stated study. Suárez-García and Gómez-Ullate compare candidate laws for IBEX35 returns; Palacios and colleagues fit NIG forms at distinct Elsasser-increment scales.[5][3] The valid conclusion is about those marginal fits. A claim about serial correlation, independent increments, or a literal latent physical variance needs separate evidence and a process model.[1][4]

Knowledge Transfer

The law transfers literally from financial-return analysis to turbulent-increment analysis: in both, a real-valued observed variable is assigned an NIG marginal family, and the same normal/inverse Gaussian mixture defines candidate densities. What changes is the measured object, sampling scheme, and estimated parameter vector.[5][3]

Within probability and statistics, the mixture representation can also guide calculation or simulation without any empirical fit.[2] Calling a social pattern “NIG-like” because it has asymmetry or extremes would only be metaphor. The name continues to require a real-valued probability law with the specific conditional-normal and inverse Gaussian mixing construction.

Examples

Canonical — a symmetric member of the family

Choose an admissible parameter vector with β=0, α>0, δ>0, and any real μ. The density above is symmetric around μ because its Bessel and square-root terms depend on (x−μ)², and its exponential asymmetry factor is exp(0). It remains NIG: zero asymmetry is a permitted parameter value, not an exit from the family.[3][2]

Mapped back: real outcome X → conditional normal kernel → positive inverse Gaussian V → the βV mean coupling set to zero → four-parameter NIG marginal with symmetric shape. The mixer and kernel are mathematical random variables, whether or not data are fitted.

Applied — unlike return and plasma-increment fits

For high-frequency IBEX35 returns, the study estimates and compares candidate marginal laws and reports NIG as its best overall generalized-hyperbolic subclass fit.[5] For WIND solar-wind Elsasser increments, another study fits NIG probability densities at multiple scales, with scale-dependent parameters.[3] These are unlike observed variables and unlike measurement systems; the same distributional rule is used in each.

Mapped back: IBEX35 returns or Elsasser increments fill the real-outcome role; a conditional normal kernel and positive inverse Gaussian mixer define each candidate NIG law; the β slot controls possible asymmetry, including zero; and a study-specific admissible α,β,δ,μ vector defines each fitted marginal. Neither study thereby observes a literal inverse Gaussian physical cause or establishes a full NIG process.

Structural Tensions

T1 — Shared family vs. fixed numerical parameters. A common NIG equation makes different scales comparable, while a single fixed parameter vector would be a much stronger claim. Palacios and colleagues use the same family but estimate scale-dependent parameters.[3] As an inference, forcing one vector could miss a changing empirical density; allowing each scale its own fit gives up a universal vector and requires separate estimation. Diagnostic: Is the proposed comparison about one functional family or one fitted parameter vector?

T2 — Marginal economy vs. temporal explanation. One NIG density can compactly describe an observed increment, while a process model must also specify how increments relate over time.[1][4] Stopping at the marginal avoids extra assumptions but cannot support a dependence claim; adding a process supports temporal questions only after its additional assumptions are checked. Diagnostic: Does the task ask for the law of one observation or for joint behavior across time?

Structural–Framed Character

The NIG family is specified by a conditional normal kernel, a positive inverse Gaussian mixing law, and the resulting marginal density.[1] It is strongly structural inside probability without a preferred data set. Its evaluative weight is low; naming the law does not endorse a financial strategy or a physical theory. Human statistical practice selects models and estimates parameters, but no institution or practice defines the family itself. Its name records a mathematical construction rather than an institution or a fixed experiment.

Its vocabulary travels literally between probability calculations, return data, and turbulence PDFs because each has real-valued random variables and fitted laws. It does not travel unchanged into a nonprobabilistic setting whose “mixture” is only a metaphor. In a new data set, the law is recognized by checking the density or generative representation and admissible parameters, rather than imported by analogy from skewness alone. Its character: a structurally precise, weakly framed probability object with cross-field statistical applications, still domain-bound by normal conditioning and inverse Gaussian mixing.

Structural Core vs. Domain Accent

A more portable skeleton would be “condition a simple response on a varying input, then aggregate over the input to obtain a richer observable pattern.” Whether that skeleton merits a broader abstraction is a separate question. The examples here demonstrate one probability-law construction across unlike data sources; they do not demonstrate a nonprobabilistic identity with the same test.

The NIG identity depends on exact probability machinery: a real-valued conditional normal law, a positive inverse Gaussian distribution, a particular mean–variance coupling, and admissible parameters giving a named density.[1][2] Finance and plasma studies change the empirical accent, not this mathematical core.[5][3] The named family fails a cross-domain Prime test because removing probability densities or replacing the mixing law eliminates the very diagnostic that separates NIG from other mixtures.

The live Probability gives a broad vocabulary for uncertainty; the NIG law is one domain-bound way to specify a probability measure. Mixture Distribution describes the broader latent-selection and marginalization construction, and Probability Distribution names the wider kind of law. The generalized inverse Gaussian distribution describes a possible mixing-law family, while the hyperbolic distribution is another generalized-hyperbolic member; neither is identical to this marginal.[3]

Neighborhood in Abstraction Space

Normal-inverse Gaussian distribution sits in a sparse region of the domain-specific corpus (94th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2835 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-31

Not to Be Confused With

  • Inverse Gaussian distribution: a positive law used for the mixer; by itself it is not the NIG marginal.[1]
  • Generic normal variance–mean mixture: keeps the conditional-normal form, but a different positive mixing law can yield another family.[2][3]
  • Generalized-hyperbolic or hyperbolic distribution: NIG fixes λ=−1/2 in the larger generalized-hyperbolic family; a sibling subclass is not the same law merely because it also has skew or heavy tails.[3][5]
  • NIG stochastic process: supplies a time-indexed joint model; a marginal NIG fit alone does not choose one process or establish its increments' dependence.[1][4]
  • Gaussian limit or Gaussian-looking center: a Gaussian approximation at the center of an empirical curve is not proof that the complete law is Gaussian or NIG.[3]

References

[1] Ole E. Barndorff-Nielsen, “Normal Inverse Gaussian Distributions and Stochastic Volatility Modelling”, Scandinavian Journal of Statistics 24 (1997), 1–13. Publisher abstract and metadata inspected; the full original article was not available in this pass. The publisher lists a 2007 online date for the 1997 volume. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o

[2] SciPy, scipy.stats.norminvgauss documentation, v1.11.3, Notes. Official implementation documentation used as an independent check of the conditional representation and standardized/four-parameter conventions, not as the original discovery source. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m

[3] Juan C. Palacios, Jean C. Perez, and Sofiane Bourouaine, “Parametric description of intermittent probability distribution functions in solar wind and magnetohydrodynamic turbulence”, Monthly Notices of the Royal Astronomical Society 531 (2024), 24–34. Publisher full text §§3–5, especially §4.2 equations (26)–(29), inspected. The source uses Λ where this entry uses δ for scale. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h ↩i ↩j ↩k ↩l ↩m ↩n ↩o ↩p ↩q ↩r ↩s ↩t ↩u

[4] Ole E. Barndorff-Nielsen, “Processes of normal inverse Gaussian type”, Finance and Stochastics 2 (1997), 41–68. Publisher abstract and metadata inspected; it discusses multiple NIG-marginal process types with different dependence structures, not a particular empirical financial fit. registry ↩a ↩b ↩c ↩d ↩e

[5] Pablo Suárez-García and David Gómez-Ullate, “Scaling, stability and distribution of the high-frequency returns of the IBEX35 index”, author manuscript, published in Physica A 392 (2013), 1409–1417. Abstract, §2.2, and §3 inspected (one-based PDF pp. 1, 3–6); the comparative conclusion is limited to their data and tested families. registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g ↩h