Source-linked AI summary

On a generalization of the Jensen-Shannon divergence and the JS-symmetrization of distances relying on abstract means

Frank Nielsen

arXiv:1904.04017v5cs.ITcs.LG

TL;DR

The paper addresses the lack of closed-form Jensen-Shannon divergence for some parametric families, especially Gaussian distributions. It introduces generalized JS divergences and JS-symmetrizations from abstract means, obtaining closed forms with geometric means for exponential families and harmonic means for scale Cauchy distributions.

  • Problem

    Ordinary Jensen-Shannon divergence lacks a closed-form expression for generic exponential-family cases such as distinct Gaussian mixtures.

  • Method

    The paper defines generalized (M, N)-Jensen-Shannon divergences and JS-symmetrizations of arbitrary distances using statistical mixtures derived from abstract means.

  • Results

    The paper reports closed-form geometric Jensen-Shannon formulas for exponential families and a closed-form harmonic Jensen-Shannon formula for scale Cauchy distributions.

  • Takeaways & Limitations

    Choosing an abstract mean according to the parametric family provides a route to closed-form Jensen-Shannon-type divergences and supports clustering applications.

  • Takeaways & Limitations

    Matrix extensions face nonuniqueness because geometric matrix means are not uniquely defined, yielding different matrix Hellinger distances.

Abstract

from arXiv · show

The Jensen-Shannon divergence is a renown bounded symmetrization of the unbounded Kullback-Leibler divergence which measures the total Kullback-Leibler divergence to the average mixture distribution. However the Jensen-Shannon divergence between Gaussian distributions is not available in closed-form. To bypass this problem, we present a generalization of the Jensen-Shannon (JS) divergence using abstract means which yields closed-form expressions when the mean is chosen according to the parametric family of distributions. More generally, we define the JS-symmetrizations of any distance using generalized statistical mixtures derived from abstract means. In particular, we first show that the geometric mean is well-suited for exponential families, and report two closed-form formula for (i) the geometric Jensen-Shannon divergence between probability densities of the same exponential family, and (ii) the geometric JS-symmetrization of the reverse Kullback-Leibler divergence. As a second illustrating example, we show that the harmonic mean is well-suited for the scale Cauchy distributions, and report a closed-form formula for the harmonic Jensen-Shannon divergence between scale Cauchy distributions. We also define generalized Jensen-Shannon divergences between matrices (e.g., quantum Jensen-Shannon divergences) and consider clustering with respect to these novel Jensen-Shannon divergences.

1 Introduction and motivations

The paper motivates generalized Jensen-Shannon constructions by contrasting bounded Jensen-Shannon divergence with asymmetric, potentially unbounded Kullback-Leibler divergence and related symmetrizations. It then extends symmetrization to arbitrary distances using abstract means, targeting closed-form formulas for structured distribution families.

  • Kullback-Leibler divergence is asymmetric, unbounded, and may be infinite, motivating bounded or symmetric alternatives.
  • Jeffreys divergence symmetrizes Kullback-Leibler divergence but remains unbounded and can cause numerical sensitivity.
  • Jensen-Shannon divergence measures divergence to an average distribution and is always bounded; its square root is a metric.
  • The paper generalizes JS-symmetrization to arbitrary distances through abstract weighted means, recovering ordinary Jensen-Shannon divergence with arithmetic means.
  • It selects means suited to distribution families, using geometric means for exponential families and harmonic means for scale Cauchy distributions.

2 Jensen-Shannon divergence in mixture and exponential families

For mixture families, Jensen-Shannon divergence admits a Jensen-Bregman representation, whereas generic exponential-family cases lack a practical closed form because ordinary mixtures generally leave the family. This limitation motivates generalized divergences based on abstract means.

  • Mixture families use prescribed component distributions and include categorical distributions as a special case.
  • 69 Theorem 1 gives a closed-form Jensen-Shannon expression for distributions in the same mixture family.
  • For mixture families, the Jensen-Shannon divergence can be expressed as a Jensen-Bregman divergence for the negentropy generator.
  • The mixture of two distinct Gaussian components is generally not Gaussian, so its ordinary Jensen-Shannon divergence lacks an obvious closed-form expression.
  • The paper introduces generalized Jensen-Shannon divergences to obtain closed-form formulas by choosing abstract means matched to the relevant parametric family.

3 Generalized Jensen-Shannon divergences

This section generalizes Jensen-Shannon divergences by constructing statistical mixtures from abstract weighted means. It also establishes boundedness conditions and extends JS symmetrization to arbitrary distances.

  • Abstract means and statistical mixtures: Abstract means generate statistical M-mixtures, which serve as the basis for generalized Jensen-Shannon divergences.The construction begins with weighted means and their normalized density interpolations.
  • Abstract means and statistical mixtures: Quasi-arithmetic means unify common weighted means, including the geometric mean obtained with the logarithm generator.The arithmetic, geometric, and harmonic means are treated within this broader framework.
  • Generalized Jensen-Shannon divergence: The M-Jensen-Shannon divergence replaces the usual arithmetic mixture with an M-mixture, recovering the ordinary JSD when M is arithmetic.The framework also extends from two-component to k-component mixtures.
  • Generalized JS symmetrizations: The same construction defines M-JS symmetrizations for arbitrary distances, including generalized K-divergences and generalized Jeffreys divergences.These definitions combine distance evaluations through weighted means rather than restricting the construction to Kullback-Leibler divergence.
  • Boundedness: M-JS divergences are bounded by log ZM, while the arithmetic case recovers the α-skew bound −log(1 −α), including log 2 at α = 1/2.The normalization factor is controlled using total variation, with ZM α(p, q) ∈[0, 2].
  • Family-adapted means: Choosing the abstract mean according to the distribution family is proposed as a route to closed-form statistical distances.The paper motivates this choice because ordinary Gaussian mixtures are generally not Gaussian, obstructing closed-form JSD expressions.

4 Some closed-form formula for the M-Jensen-Shannon divergences

The paper chooses abstract means that preserve parametric families under generalized mixing, yielding closed-form Jensen-Shannon divergences and related symmetrizations. It develops geometric formulas for exponential families, including Gaussians, and a harmonic formula for scale Cauchy distributions, alongside clustering and matrix-distance applications.

  • Abstract means are chosen to match parametric families, so generalized M-Jensen-Shannon divergences can admit closed-form formulas.The section focuses on geometric means for exponential families and harmonic means for the Cauchy scale family.
  • 4.1 The geometric Jensen-Shannon divergence: G-JSD: The weighted geometric mean Gα(x, y) = x1−αyα is used for exponential families because normalized weighted products remain in the same family.Convexity of the natural parameter space ensures the geometrically mixed distribution remains an exponential-family member.
  • 4.1.1 Case study: The multivariate Gaussian family: For multivariate Gaussians, the exponential-family representation converts KL into Bregman divergence and yields explicit G-JS formulas for KL and reverse KL.The construction uses composite vector-and-matrix parameters and the Gaussian log-normalizer.
  • 4.1 The geometric Jensen-Shannon divergence: G-JSD: The geometric JS-symmetrization of reverse KL equals a Jensen/Burbea-Rao divergence between the corresponding natural parameters.The reverse divergence swaps the calling arguments, producing a distinct JS construction from the ordinary one.
  • 4.1.2 Applications to k-means clustering: The generalized divergences are applied to generalized k-means clustering, while matrix extensions face non-unique geometric matrix means and multiple Hellinger-distance notions.The generalized k-means objective is NP-hard to optimize exactly.
  • 4.3 The skewed Bhattacharyya distance interpreted as a geometric Jensen-Shannon symmetrization (G-JS): The skewed Bhattacharyya distance is identified as a geometric JS divergence for reverse KL, with a unique maximizing exponent α∗ called the Chernoff information.At α∗, KL(pα∗: p1) = KL(pα∗: p0).
  • 4.4 The harmonic Jensen-Shannon divergence (H-JS): The harmonic mean is well-suited to the Cauchy scale family and gives a closed-form harmonic Jensen-Shannon divergence.KL divergence within the same scale family is scale-invariant, supporting the scale-family treatment.

5 Conclusion and perspectives

The paper generalizes Jensen-Shannon divergences and JS-symmetrizations through abstract means, deriving closed forms for exponential and Cauchy scale families and extending the framework to other distances and applications.

  • The (M, N)-Jensen-Shannon divergences use abstract means for statistical mixtures and for symmetrizing asymmetric Kullback-Leibler divergence.
  • Arithmetic and geometric weighted means yield closed-form M-Jensen-Shannon divergences for mixture and exponential families, respectively.
  • The paper reports closed-form geometric Jensen-Shannon formulas for exponential-family densities and a harmonic Jensen-Shannon formula for scale Cauchy distributions.
  • The framework extends JS-symmetrization to any base distance D, including skew N-Jeffreys and skew (M, N)-JS constructions.
  • The geometric Jensen-Shannon divergence has applications in machine learning, while the paper also discusses clustering and extensions involving matrix divergences.

A Summary of distances and their notations

This summary defines the notation for generalized means, statistical distances, divergences, and their Jensen-Shannon, Jeffreys, and Bregman symmetrizations.

  • The arithmetic and geometric means are Aα(x, y) = (1 − α)x + αy and Gα(x, y) = x^(1−α)y^α.
  • A quasi-arithmetic mean is Mα(x, y) = f^−1((1 − α)f(x) + αf(y)) for strictly monotone f.
  • The reverse distance is D*(p : q) = D(q : p), and the Jeffreys divergence combines forward and reverse Kullback-Leibler divergences.
  • The skew Bhattacharyya divergence uses the logarithm of an α-weighted product integral, while the Mahalanobis distance is listed separately among statistical distances.
  • The skew Jeffreys-Bregman divergence averages two directed Bregman divergences, and the skew Jensen divergence is defined alongside it.
  • Generalized JS notation includes skew JS-symmetrization, skew M-Jensen-Shannon divergence, N-Jeffreys divergence, and skew (M, N)-JS divergence.

B Symbolic calculations in Maxima

The appendix uses Maxima code to calculate the normalizer for harmonic mixtures of Cauchy distributions.

  • The Maxima program calculates the normalizer Z for harmonic H-mixtures of Cauchy distributions.
  • The code defines the Cauchy density, harmonic mean, positive scale parameters, and positive α before integration.
  • The normalizer is obtained by integrating the harmonic-mixture expression over x from −∞ to ∞.
Loading 1904.04017v5…