Flatness of a statistical manifold with Fisher information metric

444 Views Asked by At

Let $\mathcal{M} = \{p_\theta := p(\cdot | \theta), \theta \in \Theta\}$ be a statistical manifold with Fisher information metric: $$g_{{jk}}(\theta )=\operatorname {E} \left[\left({\frac {\partial }{\partial \theta _{i}}}\log p(X;\theta )\right)\left({\frac {\partial }{\partial \theta _{j}}}\log p(X;\theta )\right) \right].$$

The Wikipedia article on the topic derives the metric form Euclidean metric by changing variables. I can understand the procedure but I have questions related to the flatness of $\mathcal{M}$. In Amari's book; "Information Geometry and its application", it is said that such manifold is flat (dually flat actually) so

1- Is the above derivation enough to conclude that the manifold is flat (I mean the fact that the metric is derived from the Euclidean metric)?

2- Is there a straightforward way to show that the curvature is 0 everywhere?

2

There are 2 best solutions below

0
On BEST ANSWER

This is a common source of confusion.

The flatness in information geometry refers to the dual (with respect to the Fisher-Rao metric) connections $\nabla$ and $\nabla^*$, called exponential and mixture connections, and not to the metric or Levi-Civita connection. For the latter, please see Bruveris, M., & Michor, P. W. (2019). Geometry of the Fisher-Rao metric on the space of smooth densities on a compact manifold. Mathematische Nachrichten, 292(3), 511–523. https://doi.org/10.1002/mana.201600523

Finite-dimensional statistical manifolds can have negative curvature, for example the information manifold of 1-dimensional Gaussians is isometric to the Poincaré half-plane.

1
On

You just have to read Jean-Louis Koszul Work who has developed elementary structure of Information Geometry. You will find main Koszul references in the paper "Jean-Louis Koszul and the elementary structures of Information Geometry": https://www.academia.edu/attachments/56029291/download_file?st=MTUyNTYzMDQ1NiwxOTIuNTQuMTQ0LjIyOSw0OTY2MjA3Mw%3D%3D&s=swp-toolbar

Fisher metric could be extended on homogeneous manifolds or on Lie group by Fisher-Souriau metric developed in the framework of Souriau Lie Group Thermodynamics: [A] Marle, C.-M. From Tools in Symplectic and Poisson Geometry to J.-M. Souriau’s Theories of Statistical Mechanics and Thermodynamics. Entropy 2016, 18, 370. http://www.mdpi.com/1099-4300/18/10/370/pdf [B] Barbaresco, F. Geometric Theory of Heat from Souriau Lie Groups Thermodynamics and Koszul Hessian Geometry: Applications in Information Geometry for Exponential Families. Entropy 2016, 18, 386. http://www.mdpi.com/1099-4300/18/11/386/pdf

Other references: GSI’17 Geometric Science of Information Third International Conference, GSI 2017, Paris, France, November 7-9, 2017, Proceedings Editors: Nielsen, Frank, Barbaresco, Frédéric (Eds.) http://www.springer.com/us/book/9783319684444 Videos: https://www.youtube.com/channel/UCnE9-LbfFRqtaes49cN2DVg/videos

CIRM seminar TGSI'17 on Topological & Geometrical Structures of Information http://forum.cs-dc.org/category/94/tgsi2017

Differential Geometrical Theory of Statistics Frédéric Barbaresco and Frank Nielsen (Eds.) Pages: XIV, 458, Published: 6 June 2017 (This book is a printed edition of the Special Issue Differential Geometrical Theory of Statistics that was published in Entropy) Free Download: http://www.mdpi.com/books/pdfdownload/book/313/1

Information, Entropy and Their Geometric Structures Frédéric Barbaresco and Ali Mohammad-Djafari (Eds.) Pages: XXIV, 528, Published: 1 September 2015 (This book is a printed edition of the Special Issue Information, Entropy and Their Geometric Structures that was published in Entropy) Free Download: http://www.mdpi.com/books/pdfdownload/book/127/1