Fisher Information Metric Research Articles

A note on the Poisson structures of the space of probability measures

AbstractThe space of probability measures on a Riemannian manifold is endowed with the Fisher information metric. In [4] T. Friedrich showed that this space admits also Poisson structures {, }. In this note, we give directly another proof for the structure {, } being Poisson. (© 2007 WILEY‐VCH Verlag GmbH & Co. KGaA, Weinheim)

Distances and Riemannian Metrics for Spectral Density Functions

We introduce a differential-geometric structure for spectral density functions of discrete-time random processes. This is quite analogous to the Riemannian structure of information geometry, which is used to study perturbations of probability density functions, and which is based on the Fisher information metric. Herein, we introduce an analogous Riemannian metric, which we motivate with a problem in prediction theory. It turns out that this problem also provides a prediction theoretic interpretation to the Itakura distortion measure, which relates to our metric. Geodesies and geodesic distances are characterized in closed form and, hence, the geodesic distance between two spectral density functions provides an explicit, intrinsic (pseudo)metric on the cone of density functions. Certain other distortion measures that involve generalized means of spectral density functions are shown to lead to the same Riemannian metric. Finally, an alternative Riemannian metric is introduced, which is motivated by an analogous problem involving smoothing instead of prediction.

Quantum and Fisher information from the Husimi and related distributions

The two principal/immediate influences—which we seek to interrelate here—upon the undertaking of this study are papers of Życzkowski and Słomczyński [J. Phys. A 34, 6689 (2001)] and of Petz and Sudár [J. Math. Phys. 37, 2262 (1996)]. In the former work, a metric (the Monge one, specifically) over generalized Husimi distributions was employed to define a distance between two arbitrary density matrices. In the Petz-Sudár work (completing a program of Chentsov), the quantum analog of the (classically unique) Fisher information (monotone) metric of a probability simplex was extended to define an uncountable infinitude of Riemannian (also monotone) metrics on the set of positive definite density matrices. We pose here the questions of what is the specific/unique Fisher information metric for the (classically defined) Husimi distributions and how does it relate to the infinitude of (quantum) metrics over the density matrices of Petz and Sudár? We find a highly proximate (small relative entropy) relationship between the probability distribution (the quantum Jeffreys’ prior) that yields quantum universal data compression, and that which (following Clarke and Barron) gives its classical counterpart. We also investigate the Fisher information metrics corresponding to the escort Husimi, positive-P and certain Gaussian probability distributions, as well as, in some sense, the discrete Wigner pseudoprobability. The comparative noninformativity of prior probability distributions—recently studied by Srednicki [Phys. Rev. A 71, 052107 (2005)]—formed by normalizing the volume elements of the various information metrics, is also discussed in our context.

Strong symplectic structures on spaces of probability measures with positive density function

Spaces of probability measures with positive density function on a compact Riemannian manifold are endowed with a closed 2-form associated with the Fisher information metric by using a divergence-free vector field. In this note we give a necessary and sufficient condition on the vector field that this 2-form is a strong symplectic structure.

Dynamics of the Fisher information metric

We present a method to generate probability distributions that correspond to metrics obeying partial differential equations generated by extremizing a functional J [g(mu nu) (theta(i)) ] , where g(mu nu) (theta(i)) is the Fisher metric. We postulate that this functional of the dynamical variable g(mu nu) (theta(i)) is stationary with respect to small variations of these variables. Our approach enables a dynamical approach to the Fisher information metric. It allows one to impose symmetries on a statistical system in a systematic way.

Discriminative Components of Data

A simple probabilistic model is introduced to generalize classical linear discriminant analysis (LDA) in finding components that are informative of or relevant for data classes. The components maximize the predictability of the class distribution which is asymptotically equivalent to 1) maximizing mutual information with the classes, and 2) finding principal components in the so-called learning or Fisher metrics. The Fisher metric measures only distances that are relevant to the classes, that is, distances that cause changes in the class distribution. The components have applications in data exploration, visualization, and dimensionality reduction. In empirical experiments, the method outperformed, in addition to more classical methods, a Renyi entropy-based alternative while having essentially equivalent computational cost.

On a criterion of Riemannian distance for singularity and absolute continuity of probability measures

Parametric statistical models, with suitable regularity conditions, have a natural Riemannian manifold structure given by the Fisher information metric. This paper is concerned with the Riemannian distance for geometric conditions on singularity or absolute continuity of probability measures which depend on the parameters.

Exact Bures probabilities that two quantum bits are classically correlated

In previous studies, we have explored the ansatz that the volume elements of the Bures metrics over quantum systems might serve as prior distributions, in analogy to the (classical) Bayesian role of the volume elements ("Jeffreys' priors") of Fisher information metrics. Continuing this work, we obtain exact Bures probabilities that the members of certain low-dimensional subsets of the fifteen-dimensional convex set of 4 x 4 density matrices are separable or classically correlated. The main analytical tools employed are symbolic integration and a formula of Dittmann (quant-ph/9908044) for Bures metric tensors. This study complements an earlier one (quant-ph/9810026) in which numerical (randomization) --- but not integration --- methods were used to estimate Bures separability probabilities for unrestricted 4 x 4 or 6 x 6 density matrices. The exact values adduced here for pairs of quantum bits (qubits), typically, well exceed the estimate (.1) there, but this disparity may be attributable to our focus on special low-dimensional subsets. Quite remarkably, for the q = 1 and q = 1/2 states inferred using the principle of maximum nonadditive (Tsallis) entropy, the separability probabilities are both equal to 2^{1/2} - 1. For the Werner qubit-qutrit and qutrit-qutrit states, the probabilities are vanishingly small, while in the qubit-qubit case it is 1/4.

On the curvature of monotone metrics and a conjecture concerning the Kubo-Mori metric

The manifold of trace one positive complex n×n-matrices represents the space of faithful mixed states of a finite dimensional quantum system. Riemannian monotone metrics on these manifolds are in one-to-one correspondence with operator monotone functions. These metrics generalize the classical Fisher metric of statistical distinguishability of probability distributions to the quantum case and are of interest in quantum statistics and information theory. Thus, it is natural to ask for curvature quantities of these Riemannian metrics. It is known that the classical Fisher metric is of positive constant curvature. The quantum case considered here is much more complicated. Most informations are available for the Bures metric, a special distinguished monotone metric. Partial curvature results are also known for the Kubo-Mori metric. It is the aim of this article to determine curvature quantities of an arbitrary Riemannian monotone metric. The resulting scalar curvature is explained in more detail for three examples. In particular, we consider an important conjecture of Petz asserting that the scalar curvature of the Kubo-Mori metric increases if one goes to more mixed states. This assertion is shown up to a formal proof of the concavity of a certain function on R + 3 . However, this concavity seems to be numerically evident and we give certain function plots supporting this concavity.

Multivariate Normal Distributions Parametrized as a Riemannian Symmetric Space

The construction of a distance function between probability distributions is of importance in mathematical statistics and its applications. The distance function based on the Fisher information metric has been studied by a number of statisticians, especially in the case of the multivariate normal distribution (Gaussian) on Rn. It turns out that, except in the case n=1, where the Fisher metric describes the hyperbolic plane, it is difficult to obtain an exact formula for the distance function (although this can be achieved for special families with fixed mean or fixed covariance). We propose to study a slightly different metric on the space of multivariate normal distributions on Rn. Our metric is based on the fundamental idea of parametrizing this space as the Riemannian symmetric space SL(n+1)/SO(n+1). Symmetric spaces are well understood in Riemannian geometry, allowing us to compute distance functions and other relevant geometric data.

High-temperature expansions of Bures and Fisher information priors

For certain infinite and finite-dimensional thermal systems, we obtain-incorporating quantum-theoretic considerations into Bayesian thermostatistical investigations of Lavenda-high-temperature expansions over inverse temperature beta induced by volume elements (quantum Jeffreys' priors) of Bures metrics. Similarly to Lavenda's results based on volume elements (Jeffreys' priors) of (classical) Fisher information metrics, we find that in the limit beta-->0 the quantum-theoretic priors either conform to Jeffreys' rule for variables over [0,infinity], by being proportional to 1/beta, or to the Bayes-Laplace principle of insufficient reason, by being constant. Whether a system adheres to one rule or to the other appears to depend upon its number of degrees of freedom.

Geometries of quantum states

The quantum analog of the Fisher information metric of a probability simplex is searched and several Riemannian metrics on the set of positive definite density matrices are studied. Some of them appeared in the literature in connection with Cramér–Rao-type inequalities or the generalization of the Berry phase to mixed states. They are shown to be stochastically monotone here. All stochastically monotone Riemannian metrics are characterized by means of operator monotone functions and it is proven that there exist a maximal and a minimal among them. A class of metrics can be extended to pure states and a constant multiple of the Fubini–Study metric appears in the extension.

Quantum Fisher - Bures information of two-level systems and a three-level extension

Braunstein and Caves have recently demonstrated that the Bures metric on the mixed quantum states is equivalent - up to a proportionality factor of four - to the statistical distinguishability or quantum Fisher information metric. The volume element of these metrics can then - adapting a fundamental Bayesian principle of Jeffreys to the quantum context - serve as a reparametrization-invariant prior measure over the quantum states. The implications of this line of reasoning for the two-level systems, in general, and an embedding of them into a certain set of three-level systems are investigated.

An Infinite-Dimensional Geometric Structure on the Space of all the Probability Measures Equivalent to a Given One

Let $\mathscr{M}_\mu$ be the set of all probability densities equivalent to a given reference probability measure $\mu$. This set is thought of as the maximal regular (i.e., with strictly positive densities) $\mu$-dominated statistical model. For each $f \in \mathscr{M}_\mu$ we define (1) a Banach space $L_f$ with unit ball $\mathscr{V}_f$ and (2) a mapping $s_f$ from a subset $\mathscr{U}_f$ of $\mathscr{M}_\mu$ onto $\mathscr{V}_f$, in such a way that the system $(s_f, \mathscr{U}_f, f \in \mathscr{M}_\mu)$ is an affine atlas on $\mathscr{M}_\mu$. Moreover each parametric exponential model dominated by $\mu$ is a finite-dimensional affine submanifold and each parametric statistical model dominated by $\mu$ with a suitable regularity is a submanifold. The global geometric framework given by the manifold structure adds some insight to the so-called geometric theory of statistical models. In particular, the present paper gives some of the developments connected with the Fisher information metrics (Rao) and the Hilbert bundle introduced by Amari.

Generation of accurate Einstein-equation solutions by five-dimensional rotations

Spherically symmetric solutions of the combined system of four-dimensional Einstein, Maxwell, and specific scalar-field equations are obtained by generalization of the Schwarzschild metric and its subsequent linear quasi-rotation in the (x0, x5) plane. The accurate Kramer—Legkii solution of the five-dimensional Einstein equations is used. New accurate solutions generalizing the well-known Reissner—Nordstrom and Fisher metrics are obtained.

Geometry of canonical correlation on the state space of a quantum system

A Riemannian metric is defined on the state space of a finite quantum system by the canonical correlation (or Kubo–Mori/Bogoliubov scalar product). This metric is infinitesimally induced by the (nonsymmetric) relative entropy functional or the von Neumann entropy of density matrices. Hence its geometry expresses maximal uncertainty. It is proven that the metric is monotone under stochastic mappings, however, an example shows that it is not the only such Riemannian metric. This fact is remarkable because in the probabilistic case, the Markovian monotonicity property characterizes the Fisher information metric. The essential difference appears in the curvatures of a classical state space and a quantum one. A conjecture is made that the scalar curvature is monotone with respect to the ‘‘more mixed’’ (statistical) partial order of density matrices. Furthermore, an information inequality resembling the Cramér–Rao inequality of classical statistics is established. The inequality provides a lower bound for the canonical correlation matrix (of an unbiased observable) and it is saturated when a (partial) observation level and the corresponding family of coarse-grained states are considered.