Dispersion Models: Theory & Applications

• Dispersion models are mathematical frameworks that explicitly characterize spread phenomena, used across statistics, transport theory, and wave physics.
• They unify diverse formalisms—from two-parameter statistical distributions to transport coefficients and wavelength-dependent transfer laws—into domain-specific models.
• Applications include regression extensions, tracer dispersion in porous media, and engineered metamaterials, showcasing practical implications in various fields.

“Dispersion model” denotes several technically distinct model classes whose common feature is that they make dispersion—statistical spread, transport spreading, or frequency–wavenumber dependence—a primary object of analysis rather than a nuisance term. In statistics, the term refers to two-parameter families built from a unit deviance and a dispersion parameter, together with regression extensions that generalize GLMs and proper dispersion models (Cordeiro et al., 2019, Labouriau, 2020, Simas et al., 2010). In transport theory, it refers to effective models for tracer, solute, or particle spreading in porous media and turbulence (Milligen et al., 2011, Auton et al., 2024, Gawedzki et al., 2010, Burgener et al., 2012). In plasma physics, metamaterials, analogue gravity, and photonics, it commonly denotes a model organized around a dispersion relation or a wavelength-dependent transfer law (Told et al., 2016, Lushnikov et al., 2013, Šmejkal et al., 2024, Erkul et al., 2 Oct 2025, Wei et al., 24 Nov 2025).

1. Statistical dispersion models in the Jørgensen tradition

In Jørgensen’s sense, a dispersion model is a two-parameter family of univariate distributions with density

$$p(y;\mu,\tau)=a(y;\tau)\,\exp\!\left\{-\frac{1}{2\tau}d(y;\mu)\right\},$$

where $d$ is a unit deviance, $\mu$ is a position parameter, and $\tau>0$ is a dispersion parameter (Cordeiro et al., 2019). A unit deviance satisfies $d(y;y)=0$ and $d(y;\mu)>0$ whenever $y\neq\mu$. For regular unit deviances, the associated unit variance function is

$$V(\mu)=\frac{2}{\partial^2 d(\mu;\mu)/\partial \mu^2},$$

which controls the local mean–variance relationship and the small-dispersion normal approximation (Cordeiro et al., 2019).

Two classical subclasses organize much of the theory. Proper dispersion models factorize the normalizing term as

$a(y;\lambda)=a_0(\lambda)b(y),$

whereas exponential dispersion models are those whose unit deviance can be written in the exponential-family form $d(y;\mu)=y f(\mu)+g(\mu)+h(y)$ (Labouriau, 2020, Cordeiro et al., 2019). Normal, gamma, and inverse Gaussian are the three distributions identified as both proper and exponential dispersion models (Cordeiro et al., 2019, Labouriau, 2020). Dispersion models outside both subclasses are termed non-standard dispersion models, or NSDMs (Labouriau, 2020).

A central theoretical problem is construction: given a unit deviance $d$0, one must solve an integral equation for the normalizing function $d$1. In general this is non-trivial, but a characteristic-function construction based on real non-lattice symmetric probability measures yields unit deviances for which the associated integral equations admit a trivial constant solution and also infinitely many non-trivial solutions, thereby generating many NSDMs (Labouriau, 2020). This establishes that the class of non-standard dispersion models is very large; the paper concludes that its cardinality is larger than the cardinality of the class of real non-lattice symmetric probability measures (Labouriau, 2020).

A discrete analogue replaces ordinary cumulants by factorial cumulants. “Discrete Dispersion Models” introduces two-parameter discrete dispersion models obtained by combining convolution with factorial tilting, and shows that several overdispersed count laws, including the negative binomial and Poisson-inverse Gaussian, are Poisson-Tweedie factorial dispersion models with power dispersion functions (Jørgensen et al., 2014). The associated dilation operator generalizes binomial thinning and yields a Poisson-Tweedie asymptotic framework in which Poisson-Tweedie models appear as dilation limits (Jørgensen et al., 2014).

2. Regression, count dispersion, and longitudinal–survival extensions

With regression structure, dispersion models take the form

$d$2

where $d$3 is a precision parameter and $d$4 is dispersion (Simas et al., 2010). This framework extends generalized linear models, exponential family nonlinear models, and proper dispersion models, while preserving likelihood-based inference and higher-order asymptotics (Simas et al., 2010). One consequence is that the MLEs of regression, precision, and dispersion parameters generally exhibit nonzero skewness, and explicit matrix formulas are available for the third cumulants of these estimators (Simas et al., 2010).

For count data, dispersion is often defined relative to Poisson equidispersion. The discrete Weibull regression model provides a single likelihood-based model that can represent overdispersion, underdispersion, and covariate-specific dispersion relative to Poisson (Klakattawi et al., 2015). Its core distribution is

$d$5

with regression link

$d$6

The paper reports that $d$7 is always overdispersed relative to Poisson, $d$8 is always underdispersed, and $d$9 can be either over- or underdispersed depending on $\mu$0 (Klakattawi et al., 2015). This makes discrete Weibull regression a unified alternative to switching between Poisson, negative binomial, COM–Poisson, or zero-inflated variants (Klakattawi et al., 2015).

In joint longitudinal–survival modeling, “dispersion model” can mean explicit modeling of subject-specific longitudinal variance rather than assuming a common residual variance. The HIV/AIDS joint model

$\mu$1

relaxes homoscedasticity and incorporates the individual standard deviation $\mu$2 into the hazard model through time-varying coefficients represented by penalized splines (Martins, 2016). The best model by WAIC uses an exchangeable prior for $\mu$3, a P-spline baseline hazard, and a time-varying link including the random intercept, random slope, and $\mu$4 itself (Martins, 2016). In that usage, dispersion is a clinically interpretable biomarker of instability rather than only a variance component.

3. Transport dispersion in porous media and turbulence

In porous media, tracer dispersion denotes the spreading of a dissolved species as it moves with a fluid through pore space, driven by mechanical dispersion and molecular diffusion (Milligen et al., 2011). A compact analytical model for laminar flow through porous media expresses the effective longitudinal dispersion coefficient as

$\mu$5

where $\mu$6 is the diffusive coefficient, $\mu$7 is the mechanical-dispersion limit, and $\mu$8 is a critical velocity (Milligen et al., 2011). In terms of the microscopic Péclet number $\mu$9, the model yields a critical Péclet number $\tau>0$0 that reflects mesoscale geometric properties of the microscopic pore structure (Milligen et al., 2011). Its asymptotics unify low-velocity diffusion-dominated behavior, an intermediate superlinear longitudinal regime, and high-velocity mechanical dispersion without stitching together separate empirical power laws (Milligen et al., 2011).

A more general homogenized model treats dispersive transport and sorption in a heterogeneous porous medium with slowly varying obstacle size and spacing (Auton et al., 2024). Homogenization via multiple scales yields a macroscale equation

$\tau>0$1

characterized by the local porosity $\tau>0$2, an effective local adsorption rate $\tau>0$3, and an effective local anisotropic solute diffusivity $\tau>0$4 (Auton et al., 2024). The coefficient of effective diffusivity decomposes into molecular diffusion, the suppressive effect of the presence of obstacles, and the enhancing effect of dispersion, and the dispersive component follows a power law consistent with classical Taylor dispersion (Auton et al., 2024).

A stochastic-dynamical usage appears in turbulent particle dispersion. For close heavy particles in a synthetic turbulent flow, the separation process satisfies

$\tau>0$5

a linear SDE with multiplicative matrix-valued white noise (Gawedzki et al., 2010). The corresponding generator is hypoelliptic, the projectivized process admits a unique smooth invariant measure, and the model provides rigorous Lyapunov-exponent formulas for clustering versus separation (Gawedzki et al., 2010). A related Lagrangian model combines temporally correlated SDEs for tracer velocities with a Heisenberg-like Hamiltonian

$\tau>0$6

to impose spatial correlations and reproduce ballistic single-particle dispersion, $\tau>0$7 pair dispersion at short times, and $\tau>0$8 behavior at long times, with indications of a Richardson $\tau>0$9 law in certain situations (Burgener et al., 2012).

4. Dispersion-relation models in plasma and kinetic theory

In plasma physics, a dispersion model is often a linear wave model defined by a dispersion relation rather than a probability law. A prominent example is the hybrid kinetic-ion/fluid-electron model, in which ions satisfy the Vlasov equation, electrons are a massless fluid, and the electric field is determined by a generalized Ohm’s law containing Hall, pressure-gradient, and resistive terms (Told et al., 2016). Linearization about a homogeneous bi-Maxwellian equilibrium yields a $d(y;y)=0$0 matrix system for $d(y;y)=0$1, $d(y;y)=0$2, and $d(y;y)=0$3, and the dispersion relation is

$d(y;y)=0$4

The resulting solver captures arbitrary propagation angles, ion temperature anisotropy, ion Landau damping, cyclotron resonances, firehose and mirror instabilities, and benchmarked agreement with full Vlasov–Maxwell solutions in regimes dominated by ion kinetics (Told et al., 2016).

A different kinetic construction is the Vlasov multi-dimensional model, which keeps full Vlasov dynamics along a preferred direction and replaces perpendicular velocity space by $d(y;y)=0$5 discrete transverse flows (Lushnikov et al., 2013). Its equilibrium ansatz

$d(y;y)=0$6

defines a dynamically invariant subspace of the full Vlasov equation (Lushnikov et al., 2013). The model’s Langmuir-wave dispersion relation converges to the standard Vlasov–Landau result as $d(y;y)=0$7 increases, and in three dimensions rotational symmetry for small perpendicular wavenumber fluctuations is demonstrated for $d(y;y)=0$8, with anisotropy errors scaling as $d(y;y)=0$9 for small $d(y;\mu)>0$0 (Lushnikov et al., 2013). In this context, “dispersion model” means a controlled approximation to the full kinetic dispersion relation.

5. Dispersive continua, metamaterials, and surface excitations

In continuum mechanics of band-gap metamaterials, the term denotes a constitutive model calibrated to reproduce a prescribed dispersion diagram. The integral micromorphic continuum introduces nonlocal averaging of three terms in the free energy of a 1D micromorphic medium, with kernels $d(y;\mu)>0$1, $d(y;\mu)>0$2, and $d(y;\mu)>0$3 determined in Fourier space from target acoustic and optical branches (Šmejkal et al., 2024). For harmonic waves, the dispersion equation takes the quadratic form

$d(y;\mu)>0$4

and after proper calibration the model can exactly reproduce two given branches of the dispersion curve, including cases with a band gap (Šmejkal et al., 2024). The weight functions are reconstructed in the spatial domain by inverse Fourier transform, so dispersion is encoded directly in nonlocal constitutive kernels (Šmejkal et al., 2024).

A microscopic quantum version appears in the theory of nonretarded surface plasmons near a plane. Starting from a time-dependent Hartree equation with a binding potential $d(y;\mu)>0$5, linearization leads to a homogeneous integral equation for the vertical scattering amplitude $d(y;\mu)>0$6 (Margetis, 13 Apr 2026). Laplace transformation converts this into a functional equation whose solution is expressed by rapidly convergent Mittag-Leffler series, and the surface-plasmon dispersion relation is written as

$d(y;\mu)>0$7

In the semiclassical regime, the leading-order excitation spectrum agrees with the classical projected Euler–Poisson prediction,

$d(y;\mu)>0$8

while higher-order terms encode confinement-length corrections (Margetis, 13 Apr 2026).

6. Analogue gravity and photonic computing

In analogue gravity, dispersion is the microscopic ingredient that regularizes horizon physics. The canonical modified dispersion relation is

$d(y;\mu)>0$9

which replaces exact linear relativistic propagation at high wavenumber (Erkul et al., 2 Oct 2025). In moving dielectrics, the long-wavelength sector is described by the Gordon metric

$y\neq\mu$0

while in Bose–Einstein condensates the co-moving Bogoliubov dispersion is

$y\neq\mu$1

The essay argues that Hawking radiation persists in the presence of such dispersion and that a dispersive ultraviolet completion of quantum fields in curved spacetime may alter vacuum-energy renormalization and cosmological dynamics (Erkul et al., 2 Oct 2025).

In wavelength-multiplexed optical computing, “dispersion-aware modeling” refers to a wavelength-dependent transfer-matrix framework for cascaded Mach–Zehnder interferometer meshes (Wei et al., 24 Nov 2025). The optical processor implements

$y\neq\mu$2

but away from the calibration wavelength $y\neq\mu$3, the phases $y\neq\mu$4 vary with $y\neq\mu$5 (Wei et al., 24 Nov 2025). A local expansion of phase shifter dispersion introduces first- and second-order coefficients $y\neq\mu$6, and an edge-spectrum interpolation strategy using two calibration wavelengths reduces the global dispersion error within a 40 nm range from $y\neq\mu$7 to $y\neq\mu$8 (Wei et al., 24 Nov 2025). Here dispersion is neither a statistical variance nor a transport coefficient; it is the wavelength dependence of the programmed unitary map.

Taken together, these usages show that “dispersion model” is not a single unified object but a family of domain-specific formalisms. In statistics it separates location from variability; in transport it upscales microscale spreading into effective coefficients; and in wave physics it encodes how propagation changes with frequency, wavelength, or microstructure. This suggests that the unifying idea is structural: dispersion is promoted from a residual effect to an explicit modeling layer.