Mean-Variance Mixture of Normals

• Mean-variance mixture of normals is a class of distributions that uses a positive latent variable to simultaneously mix both the mean and covariance, enabling skewness and heavy tails.
• The model extends conventional mean or variance mixtures by allowing a single latent variable to adjust both location and scale, preserving Gaussian tractability while capturing complex dependencies.
• Applications span finance and risk management, where the model supports portfolio optimization, tail risk assessment, and enhanced inferential techniques like EM-based estimation.

Mean-variance mixture of normals (MVMN), also called the location-scale mixture of normals, is a class of distributions obtained by letting a positive latent scalar simultaneously shift the mean and scale the covariance of a multivariate normal vector. In its standard form, if $Z\sim N_p(0,I_p)$, $W>0$ is independent of $Z$, $\mu\in\mathbb R^p$ is a location vector, $\delta\in\mathbb R^p$ is a shape or skewness vector, and $\Sigma$ is positive definite, then

$Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$

Equivalently,

$Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$

with marginal density

$$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$

This construction extends variance mixtures by allowing the same latent variable to govern both mean and covariance, thereby accommodating skewness, heavy tails, and multivariate dependence within a single hierarchical mechanism (Lee et al., 2020).

1. Formal construction and equivalent representations

The defining feature of MVMN is that the same positive scalar mixing variable controls both conditional location and conditional scale. The notation

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$

is used when $W>0$0 has the density above. In finance-oriented notation, the same class is often written as

$W>0$1

where $W>0$2, $W>0$3, and $W>0$4 is positive and independent of $W>0$5; conditionally on $W>0$6, one has

$W>0$7

This is the same latent-Gaussian structure with a different symbol set for the mixing variable and skewness vector (Sayit, 2022).

A broader continuous-mixture formulation writes

$W>0$8

or, more generally,

$W>0$9

with two independent univariate random variables $Z$0 and $Z$1. The classical mean-variance mixture corresponds to the special choice $Z$2 and $Z$3, that is,

$Z$4

This general framework places mean mixtures, variance mixtures, and mean-variance mixtures inside a common family of continuous mixtures of multivariate normals (Arellano-Valle et al., 2020).

The same representation underlies several application-specific formulations. In risk and insurance, an $Z$5-dimensional normal mean-variance mixture is written as

$Z$6

so that

$Z$7

Because linear combinations preserve the class, aggregate losses $Z$8 are again normal mean-variance mixtures (Calderín-Ojeda et al., 2 Jan 2026).

2. Distinction from mean mixtures and variance mixtures

MVMN sits between two simpler normal-mixture constructions. In a variance mixture of normals (VMN),

$Z$9

so the mean is constant and only the covariance is mixed. In a mean mixture of normals (MMN),

$\mu\in\mathbb R^p$0

so the covariance is fixed and only the mean is mixed. MVMN combines both mechanisms: $\mu\in\mathbb R^p$1 Accordingly, MVMN reduces to VMN when $\mu\in\mathbb R^p$2, but MMN is not a special case of MVMN. The skew-normal distribution belongs to the MMN framework rather than to MVMN (Lee et al., 2020).

The basic moment structure makes the dual role of the mixing variable explicit. If $\mu\in\mathbb R^p$3, then

$\mu\in\mathbb R^p$4

and

$\mu\in\mathbb R^p$5

The moment generating function is

$\mu\in\mathbb R^p$6

showing that the MVMN mgf is inherited from the mgf of the mixing variable. In the broader $\mu\in\mathbb R^p$7 formulation, the same decomposition appears as

$\mu\in\mathbb R^p$8

These formulas isolate mean mixing through $\mu\in\mathbb R^p$9 and variance mixing through $\delta\in\mathbb R^p$0 (Lee et al., 2020, Arellano-Valle et al., 2020).

The family retains several normal-like closure properties. Affine transformations preserve the class: $\delta\in\mathbb R^p$1 Linear combinations with independent normal vectors remain MVMN; marginals are again MVMN; and conditionals are again MVMN with the expected partitioned parameters. This is one reason the family is used when Gaussian tractability is desired but Gaussian symmetry and light tails are inadequate (Lee et al., 2020).

Because the same positive latent variable drives both shift and spread, asymmetry arises when $\delta\in\mathbb R^p$2, while heavy tails arise through random variance scaling. In this sense MVMN is designed to model skewness and heavy tails simultaneously (Lee et al., 2020).

3. Canonical families, special cases, and extensions

The generalized hyperbolic (GH) distribution is the most prominent special case of MVMN. It is obtained when the mixing variable follows a generalized inverse Gaussian law,

$\delta\in\mathbb R^p$3

In the alternative parameterization used in related work,

$\delta\in\mathbb R^p$4

which yields the multivariate GH distribution. GH is the canonical multivariate mean-variance mixture family in this literature (Lee et al., 2020, Arellano-Valle et al., 2020).

Several important distributions arise as GH subfamilies or limits. The literature summarized here identifies the normal inverse Gaussian, variance gamma, and asymmetric Laplace as notable special cases of GH, and also states that GH encompasses Student $\delta\in\mathbb R^p$5, Laplace, and hyperbolic distributions [(Lee et al., 2020); (Yu, 2011)]. The multivariate skewed variance gamma (MSVG) model is a gamma-mixed instance with

$\delta\in\mathbb R^p$6

which gives

$\delta\in\mathbb R^p$7

This places MSVG squarely inside the normal mean-variance mixture framework (Nitithumbundit et al., 2015).

Less familiar examples noted in the survey literature include MVN Birnbaum–Saunders (MVNBS), where $\delta\in\mathbb R^p$8 is Birnbaum–Saunders, and MVN Lindley (MVNL), where $\delta\in\mathbb R^p$9 has a Lindley distribution (Lee et al., 2020). At the opposite extreme, if the mixing law degenerates at $\Sigma$0, the model collapses to the multivariate normal (Lee et al., 2020).

The MVMN architecture has also been generalized by replacing the Gaussian conditional kernel. The multivariate Mixed Tempered Stable model keeps the same mean-variance mixing form,

$\Sigma$1

but replaces the normal innovation with a standardized Classical Tempered Stable innovation. When $\Sigma$2, Mixed Tempered Stable reduces to a Normal Variance Mean Mixture; with Gamma mixing and $\Sigma$3, one recovers the Variance Gamma model (Hitaj et al., 2016). This suggests that classical MVMN occupies a central position inside a wider family of latent scale-and-shift mixture constructions.

4. Estimation, identifiability, and inferential issues

For parametric MVMN models, the EM algorithm is the standard estimation tool. The conditional Gaussian representation

$\Sigma$4

treats $\Sigma$5 as missing data and supports likelihood-based estimation by iterating conditional expectation and maximization steps. The survey literature identifies explicit EM implementations for GH distributions, MVNBS, and MVNL, even though it does not derive a full generic MVMN EM scheme in closed form (Lee et al., 2020).

Concrete ECM and ECME schemes have been developed for specific MVMN families. For the MSVG model, the complete-data likelihood splits into a normal part and a gamma part; the conditional law of $\Sigma$6 is generalized inverse Gaussian, and closed-form CM updates are available for $\Sigma$7, $\Sigma$8, and $\Sigma$9, while $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$0 is updated by Newton–Raphson or by direct maximization of the observed likelihood. The same paper extends the mean to include autoregressive terms, proposes a delta-region bounding device when the density is unbounded for $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$1, inserts an extra E-step before updating $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$2, and computes standard errors by Louis’s method (Nitithumbundit et al., 2015).

A different inferential direction is semiparametric estimation. In the univariate model

$Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$3

the coefficient $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$4 can be identified as the unique zero of a monotone functional $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$5, estimated by the empirical root

$Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$6

and then the unknown mixing density can be recovered nonparametrically by Mellin-transform inversion. The paper proves an $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$7 convergence rate for $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$8 under a moment condition and shows that the plug-in error from using $Y=\mu+W\delta+\sqrt{W}\,\Sigma^{1/2}Z.$9 in the second-stage density estimator is asymptotically negligible in the stated normalization (Belomestny et al., 2017).

Identifiability is more delicate than the latent-Gaussian hierarchy might suggest. In the model

$Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$0

with i.i.d. latent pairs $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$1, the mixing distribution $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$2 is generally not identifiable when the latent shift is unbounded. The same work proves identifiability under bounded-shift conditions in the model class it studies, by comparing characteristic functions and using uniqueness of the Laplace transform. It also shows that generalized maximum likelihood can be inconsistent even when the model is identifiable, and that independence of shift and scale does not remove this inconsistency. By contrast, if there is more than one observation per latent realization, the resulting model becomes a mixture of a bounded full-rank exponential family, and the MLE/GMLE exists, is unique, and is consistent (Ritov, 2024).

5. Shape theory, geometry, and relation to adjacent models

MVMN densities inherit important shape properties from the mixing distribution. In the univariate model

$Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$3

if the mixing density $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$4 is unimodal, then the mixture density $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$5 is unimodal; if $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$6 is log-concave, then $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$7 is log-concave; and if $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$8 is log-convex on $Y\mid W=w\sim N_p(\mu+w\delta,\;w\Sigma),$9, then $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$0 is log-convex on each of $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$1 and $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$2. If $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$3 is decreasing, or if $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$4, then the only mode is at $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$5. In the multivariate case, the corresponding results use the corrected mixing density

$$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$6

If $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$7 is unimodal, the density has only one local maximum, lying on the line $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$8; if $$f(y;\mu,\Sigma,\delta;h(w;\theta))=\int_0^\infty \phi_p(y;\mu+w\delta,w\Sigma)\,h(w;\theta)\,dw.$$9 is log-concave, then the multivariate density is log-concave (Yu, 2011).

The multivariate geometry is correspondingly rigid. For each $Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$0, the density has ellipsoidal contours on the hyperplane

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$1

so the density is spindle-like: it is constant on ellipsoids orthogonal to the direction $Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$2. In the GH case, these general shape results yield an especially clean classification: all GH densities are unimodal; in the univariate case GH is log-concave iff $Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$3; and in the multivariate case GH is log-concave iff

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$4

The same paper presents these results as a short proof of unimodality for all generalized hyperbolic densities (Yu, 2011).

Several neighboring models are related but not identical to MVMN. Mean-mixtures of multivariate normals randomize only the mean,

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$5

with fixed conditional covariance, and are explicitly described as not obtained from the MVMN class because they lack the $Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$6 variance mixing term (Abdi et al., 2020). Conversely, finite discrete mixtures of multivariate normals, such as two-regime portfolio models with

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$7

remain mixtures of Gaussians but are regime mixtures rather than continuous mean-variance mixtures driven by a single positive scalar (Kocuk et al., 2017).

6. Portfolio theory, tail risk, and capital allocation

In portfolio applications, MVMN is usually written as

$Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$8

with $Y\sim MVMN_p(\mu,\Sigma,\delta;h(w;\theta))$9 independent of $W>0$00. For any portfolio $W>0$01,

$W>0$02

A central stochastic-dominance result states that for

$W>0$03

the conditions

$W>0$04

are sufficient for second-order stochastic dominance under the stated integrability assumptions. This leads to a closed-form frontier theorem: if returns follow the NMVM class and $W>0$05 is any finite-valued, law-invariant, convex risk measure on $W>0$06, then the frontier portfolio for a target return $W>0$07 is obtained by solving the classical Markowitz mean-variance problem with adjusted mean

$W>0$08

and covariance

$W>0$09

The optimizer therefore depends on the shifted mean and on the Gaussian covariance component, not on the full covariance

$W>0$10

This applies in particular to $W>0$11 and to law-invariant coherent or spectral risk measures of Kusuoka type (Sayit, 2022).

A related portfolio-optimization result concerns mean-risk-skewness criteria. After transforming portfolio weights by $W>0$12, the latent representation yields

$W>0$13

and, under the condition

$W>0$14

the mean-risk-skewness problem for any law-invariant coherent risk measure reduces to the quadratic program

$W>0$15

The same work derives exact integral formulas and approximate closed-form formulas for portfolio VaR and CVaR under general NMVM returns. The approximations use only the one-dimensional risks of

$W>0$16

and are reported to be accurate and computationally efficient in the numerical study (Abudurexiti et al., 2021).

Recent work on capital allocation extends MVMN methodology beyond CTE. For aggregate loss $W>0$17, the $W>0$18-th order tail central moment is defined by

$W>0$19

and the corresponding capital allocation rule is

$W>0$20

Under the NMVM model, the paper derives explicit formulas for tail moments, tail central moments, and the allocation itself: $W>0$21 with coefficients determined by $W>0$22, $W>0$23, and $W>0$24. In the four-asset GH illustration, BA and CVX have fairly stable allocation proportions across CTE, TV, and $W>0$25, whereas AXP receives larger TV and $W>0$26 allocations than its CTE allocation, and XOM can have a negative $W>0$27 contribution (Calderín-Ojeda et al., 2 Jan 2026).

Across these applications, MVMN is used not merely as a heavier-tailed replacement for the multivariate normal, but as a latent-variable class in which skewness, tail thickness, and dependence are jointly parameterized while much of the Gaussian conditional structure remains available for optimization, conditioning, and likelihood-based inference (Lee et al., 2020).