[Paper Review] On Normal Variance-Mean Mixtures

Yaming Yu|arXiv (Cornell University)|Jun 12, 2011

Bayesian Methods and Mixture Models12 references3 citations

TL;DR

This paper establishes sufficient conditions for unimodality and log-concavity in univariate and multivariate normal variance-mean mixtures, providing a unified framework that proves the unimodality of generalized hyperbolic densities via a short, direct argument. The key contribution is a general theorem showing that unimodality and log-concavity are preserved under normal variance-mean mixing, with applications to heavy-tailed distributions like Student's t and variance gamma.

ABSTRACT

Normal variance-mean mixtures encompass a large family of useful distributions such as the generalized hyperbolic distribution, which itself includes the Student t, Laplace, hyperbolic, normal inverse Gaussian, and variance gamma distributions as special cases. We study shape properties of normal variance-mean mixtures, in both the univariate and multivariate cases, and determine conditions for unimodality and log-concavity of the density functions. This leads to a short proof of the unimodality of all generalized hyperbolic densities. We also interpret such results in practical terms and discuss discrete analogues.

Motivation & Objective

• To establish sufficient conditions for unimodality and log-concavity in univariate and multivariate normal variance-mean mixtures.
• To provide a direct, general proof of unimodality for generalized hyperbolic (GH) densities, avoiding reliance on deep results from self-decomposability.
• To clarify the inheritance of shape properties—such as tail heaviness and mode location—from the mixing distribution to the resulting mixture.
• To extend results to multivariate settings and characterize convex contours and mode locations in multivariate normal variance-mean mixtures.
• To derive discrete analogues and discuss practical implications for statistical modeling and random variate generation.

Proposed method

• The paper uses a general representation of normal variance-mean mixtures as Y = μ + βX + σ√X Z, where X is a mixing random variable with a density on (0, ∞), and Z ~ N(0,1).
• It proves that if the mixing density g is unimodal or log-concave, then the resulting mixture density f inherits these properties.
• The proof relies on analyzing the mixture density via integral representations and applying known results on log-concave and unimodal transformations.
• For multivariate extensions, the method uses linear transformations and analyzes the density along lines in R^p, reducing the problem to univariate mixtures.
• It introduces a transformed mixing density g*(x) = x^{-(p-1)/2}g(x) to study mode locations and shape properties in higher dimensions.
• The analysis leverages properties of modified Bessel functions and generalized inverse Gaussian (GIG) distributions to derive explicit conditions for log-concavity and log-convexity.

Experimental results

Research questions

• RQ1 Under what conditions on the mixing distribution is a normal variance-mean mixture unimodal?
• RQ2 When is the density of a normal variance-mean mixture log-concave, and how does this relate to tail behavior?
• RQ3 Where is the mode of the mixture located, and when is it at the location parameter μ?
• RQ4 How do shape properties like unimodality and log-concavity behave under linear transformations in the multivariate case?
• RQ5 What are the necessary and sufficient conditions for log-convexity of the mixture density on each side of the mode?

Key findings

• The paper proves that if the mixing density g is unimodal, then the resulting normal variance-mean mixture f is also unimodal, providing a general and direct proof of unimodality for generalized hyperbolic distributions.
• If the mixing density g is decreasing on (0, ∞) or β = 0, then the unique mode of f is located at μ.
• The mixture density f is log-concave if and only if the transformed mixing density g*(x) = x^{-(p-1)/2}g(x) is log-concave, which for the generalized inverse Gaussian case corresponds to λ ≥ (p+1)/2 in the multivariate setting.
• For the generalized hyperbolic distribution, log-concavity holds if and only if λ ≥ 1 in the univariate case, and λ ≥ (p+1)/2 in the multivariate case.
• The mixture density is log-convex on (−∞, μ) and (μ, ∞) if and only if δ = 0 and 0 < λ ≤ 1 in the generalized hyperbolic case.
• In the multivariate case, the density f(y) has convex level sets (convex contours) if and only if the transformed mixing density g* is log-concave, which is equivalent to λ ≥ (p+1)/2.