Torsion of α \alpha -connections on the density manifold
Nihat Ay, Lorenz J. Schwachhöfer · 2025
Read Torsion of α \alpha -connections on the density manifold by Nihat Ay, Lorenz J. Schwachhöfer (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-23588.txt.
Introduction
Nihat Ay and Lorenz Schwachhöfer study a family of geometric structures used to compare probability distributions, in this 2025 mathematics paper. The structures, called alpha-connections, live on the manifold of densities, and the question is whether they twist, a property named torsion. For the Fisher-Rao metric, every alpha-connection proves torsion free. For the Otto metric, the answer flips: except for alpha equal to minus one, the connections do carry torsion.
Catalog record
| CREATOR | Nihat Ay, Lorenz J. Schwachhöfer |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 23 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-23588
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Torsion of α \alpha -connections on the density manifold Nihat Ay, Lorenz J. Schwachhöfer Abstract We study the torsion of the α \alpha -connections defined on the density manifold in terms of a regular Riemannian metric. In the case of the Fisher-Rao metric our results confirm the fact that all α \alpha -connections are torsion free. For the α \alpha -connections obtained by the Otto metric, we show that, except for α = − 1 \alpha=-1 , they are not torsion free. 1 Introduction Classical information-geometric structures are agnostic to the geometry of the sample space M which is given, for instance, in terms of a Riemannian metric {\rm g} . To address this problem, many works within information geometry have tried to incorporate Wasserstein geometry [ 5 ] . In this context, the so-called Otto metric has been studied as an alternative to the Fisher-Rao metric [ 7 ] . This line of research has been recently extended by the definition of an exponential connection as the dual of the mixture connection with respect to the Otto metric [ 2 ] . It was conjectured that the new exponential connection is torsion free, which would directly imply counterparts of important results within the classical theory, such as the generalised Pythagorean theorem in terms of a canonical contrast function. In the present article, we disprove this conjecture by deriving an explicit formula for the torsion of the new exponential connection. In fact, we study the torsion of all \alpha -connections associated with some regular Riemannian metric on the density manifold. For \alpha=0 , we do not recover the Levi-Civita connection whenever the connection for \alpha=1 is not torsion free. 2 The affine structure of P^{\infty}_{+}(M) We consider a finite-dimensional compact connected C^{\infty} -manifold M without boundary. A smooth measure on M is a measure which is induced by some smooth Riemannian metric on M . By compactness of M , such a measure is finite, and we are particularly interested in the space P_{+}^{\infty}(M) of smooth probability measures. Picking some \mu\in P_{+}^{\infty}(M) , we may identify the vector space S^{\infty}(M) of smooth signed measures on M with the space C^{\infty}(M ) of arbitrarily often differentiable real valued functions on M . This is done by the linear isomorphism \imath_{\mu}:C^{\infty}(M)\longrightarrow S^{\infty}(M),\qquad f\longmapsto f\;\mu, (1) whose inverse is given by the Radon-Nikodym derivative w.r.t. \mu . We also consider the codimension one subspaces \displaystyle{S}^{\infty}_{0}(M)\; \displaystyle:= \displaystyle\;\left\{\nu\in S^{\infty}(M)\;:\;\int_{M}\nu=0\right\}, (2) \displaystyle C^{\infty}_{0}(M,\mu)\; \displaystyle:= \displaystyle\;\left\{f\in C^{\infty}(M)\;:\;\int_{M}f\;d\mu=0\right\}, so that obviously, \imath_{\mu} from ( 1 ) restricts to an isomorphism of C^{\infty}_{0}(M,\mu) and S^{\infty}_{0}(M) . Morover, P^{\infty}_{+}(M) is the image of the intersection of the affine hyperplane 1+C^{\infty}_{0}(M,\mu) with the convex open set of positive functions. As C^{\infty}(M) has a canonical Fréchet space structure [ 6 ] , it follows that \imath_{\mu} induces a Fréchet space structure on S^{\infty}(M) and S^{\infty}_{0}(M) , and P^{\infty}_{+}(M) has the structure of an open convex set in an affine hyperplane. It is straightforward to verify that these Fréchet structures are independent of the choice of \mu . Therefore, there is a canonical identification T_{\mu}P^{\infty}_{+}(M)\cong S^{\infty}_{0}(M),\qquad TP^{\infty}_{+}(M)\cong P^{\infty}_{+}(M)\times S^{\infty}_{0}(M), (3) and the canonical flat connection on the vector space S^{\infty}(M) restricts to a canonical flat affine connection \nabla^{({\rm m})} on P^{\infty}_{+}(M) , called the mixture connection . It is characterized by the fact that the constant vector fields (A_{\mu}):=(\mu,\nu_{0}) are \nabla^{({\rm m})} -parallel for each \nu_{0}\in S^{\infty}_{0}(M) . In particular, the \nabla^{({\rm m})} -geodesics are the straight line segments, and due to the convexity, any two elements in P^{\infty}_{+}(M) can be joined by a unique \nabla^{({\rm m})} -geodesic. 3 Regular metrics on P^{\infty}_{+}(M) A Riemannian metric on P^{\infty}_{+}(M) is a collection of inner products {\mathcal{G}}=({\mathcal{G}}_{\mu})_{\mu\in P^{\infty}_{+}(M)} on each T_{\mu}P^{\infty}_{+}(M)\cong S^{\infty}_{0}(M) which depends smoothly on \mu , in the sense that smooth vector fields A,B on P^{\infty}_{+}(M) (i.e., smooth families A=(A_{\mu})_{\mu\in P^{\infty}_{+}(M)} , B=(B_{\mu})_{\mu\in P^{\infty}_{+}(M)}\in S^{\infty}_{0}(M) ) yield a smooth function on P^{\infty}_{+}(M) \mu\longmapsto{\mathcal{G}}_{\mu}(A_{\mu},B_{\mu}). Example: The Fisher-Rao metric Given \mu\in P^{\infty}_{+}(M) , we use ( 1 ) to identify T_{\mu}P^{\infty}_{+}(M)\cong S^{\infty}_{0}(M)\cong C^{\infty}_{0}(M,\mu)\subset L^{2}(M,\mu) , where the latter is the Hilbert space of square integrable functions. The Fisher-Rao metric on P^{\infty}_{+}(M) is then the restricton of the inner product on L^{2}(M,\mu) to C^{\infty}_{0}(M,\mu) under this identification. That is, \llangle A,B\rrangle^{\rm FR}_{\mu}:=\int_{M}\frac{dA_{\mu}}{d\mu}\;\frac{dB_{\mu}}{d\mu}\;d\mu. (4) The Fisher-Rao metric has very natural properties; e.g. it is invariant under diffeomorphisms of M and is in fact (up to constant rescaling) the only Riemannian metric on P^{\infty}_{+}(M) with this property [ 4 ] . However, there are other important choices of such metrics, for instance the Otto metric which we shall discuss in more detail in section 5 . For the moment, we shall stay fairly general in our discussion, but always keep the Otto metric in mind as our main application. Definition 1 A regular Riemannian metric on P^{\infty}_{+}(M) is a metric of the form {\mathcal{G}}_{\mu}(A,B):=\llangle\Phi^{\mathcal{G}}_{\mu}(A_{\mu}),B_{\mu}\rrangle^{\rm FR}_{\mu}, (5) where \Phi^{\mathcal{G}}_{\mu}:S^{\infty}_{0}(M)\to S^{\infty}_{0}(M) is an automorphism of Fréchet spaces. Evidently, this is only well defined if for each \nu\in S^{\infty}_{0}(M) the expression (\Phi^{\mathcal{G}}_{\mu})(\nu) depends smoothly on \mu , and if \llangle\Phi^{\mathcal{G}}_{\mu}(\nu^{1}),\nu^{2}\rrangle^{\rm FR}_{\mu}=\llangle\Phi^{\mathcal{G}}_{\mu}(\nu^{2}),\nu^{1}\rrangle^{\rm FR}_{\mu},\qquad\llangle\Phi^{\mathcal{G}}_{\mu}(\nu),\nu\rrangle^{\rm FR}_{\mu}>0,\;\nu\neq 0. (6) We also shall use the notation for the Gâteaux derivative \partial_{\nu^{1}}\Phi^{\mathcal{G}}_{\mu}(\nu^{2}):=\left.\frac{d}{dt}\right|_{t=0}\Phi^{\mathcal{G}}_{\mu+t\nu^{1}}(\nu^{2})\in S^{\infty}_{0}(M). (7) 4 The \alpha -connections of a regular metric on P^{\infty}_{+}(M) For a Riemannian metric {\mathcal{G}} on P^{\infty}_{+}(M) its covariant mixture derivative \nabla^{({\rm m})}{\mathcal{G}} is defined as the tensor \displaystyle{\mathcal{A}}^{\mathcal{G}}(A;B,C)\; \displaystyle:= \displaystyle\;-\left(\nabla^{({\rm m})}_{A}{\mathcal{G}}\right)(B,C) (8) \displaystyle= \displaystyle\;-A\left({\mathcal{G}}(B,C)\right)+{\mathcal{G}}\left(\nabla^{({\rm m})}_{A}B,C\right)+{\mathcal{G}}\left(B,\nabla^{({\rm m})}_{A}C\right). It is straightforward to verify that this expression is tensorial in all entries, so that {\mathcal{A}}^{\mathcal{G}}(A;B,C)_{\mu} only depends on A_{\mu},B_{\mu} and C_{\mu} . Furthermore, it is obviously symmetric in the last two entries. Lemma 1 We have the identity {\mathcal{A}}^{\mathcal{G}}(A;B,C)={\mathcal{G}}\left({\mathcal{K}}^{\mathcal{G}}(A,B),C\right), (9) where {\mathcal{K}}^{\mathcal{G}}(A,B)_{\mu}=(\Phi^{\mathcal{G}}_{\mu})^{-1}\left(\frac{d{A_{\mu}}}{d\mu}\;\Phi^{\mathcal{G}}_{\mu}(B_{\mu})-\partial_{A_{\mu}}\Phi^{\mathcal{G}}_{\mu}(B_{\mu})\right), (10) using the notation ( 7 ). Proof Let \mu\in P^{\infty}_{+}(M) and a,b,c\in C^{\infty}_{0}(M,\mu) , so that the constant vector fields A:\equiv a\mu , B:\equiv b\mu and C:\equiv c\mu on P^{\infty}_{+}(M) are \nabla^{({\rm m})} -parallel, i.e., \nabla^{({\rm m})}_{A}B=\nabla^{({\rm m})}_{A}C=0 . Then ( 8 ) implies \displaystyle{\mathcal{A}}^{\mathcal{G}}(A;B,C)\; \displaystyle= \displaystyle\;-A\llangle B,C\rrangle^{\mathcal{G}}_{\mu}=-\left.\frac{d}{dt}\right|_{t=0}\llangle\Phi^{\mathcal{G}}_{(1+ta)\mu}(B),C\rrangle^{\rm FR}_{(1+ta)\mu} \displaystyle= \displaystyle\;-\left.\frac{d}{dt}\right|_{t=0}\int_{M}\frac{d\Phi^{\mathcal{G}}_{(1+ta)\mu}(B)}{d((1+ta)\mu)}\frac{dC}{d((1+ta)\mu)}\;d((1+ta)\mu) \displaystyle= \displaystyle\;\int_{M}\left(a\;\frac{d\Phi^{\mathcal{G}}_{\mu}(B)}{d\mu}-\frac{d(\partial_{a\mu}\Phi^{\mathcal{G}}_{\mu}(B))}{d\mu}\right)\frac{dC}{d\mu}\;d\mu \displaystyle= \displaystyle\;\left\llangle\left(\frac{dA_{\mu}}{d\mu}\;\Phi^{\mathcal{G}}_{\mu}(B_{\mu})-\partial_{A_{\mu}}\Phi^{\mathcal{G}}_{\mu}(B_{\mu})\right),C_{\mu}\right\rrangle_{\mu}^{\rm FR}, which together with ( 5 ) implies ( 10 ). Definition 2 Let \mathcal{G} be a regular metric on P^{\infty}_{+}(M) , given by ( 5 ) for some map \phi_{\mu}^{\mathcal{G}}:S^{\infty}_{0}(M)\to S^{\infty}_{0}(M) . Then for \alpha\in{\mathbb{R}} we define the \alpha -connection of \mathcal{G} by \nabla^{({\mathcal{G}},\alpha)}_{A}B:=\nabla^{({\rm m})}_{A}B-\frac{\alpha+1}{2}{\mathcal{K}}^{\mathcal{G}}(A,B). (11) Observe that according to ( 11 ) we have {\mathcal{K}}^{\mathcal{G}}(A,B)=\nabla^{({\rm m})}_{A}B-\nabla^{({\mathcal{G}},1)}_{A}B , whence ( 11 ) is equivalent to \nabla^{({\mathcal{G}},\alpha)}_{A}B\;=\;\frac{1-\alpha}{2}\;\nabla^{({\rm m})}_{A}B+\frac{1+\alpha}{2}\;\nabla^{({\mathcal{G}},1)}_{A}B. We now collect the following properties of these \alpha -connections. Theorem 4.1 1. Equation ( 11 ) defines indeed a connection of P^{\infty}_{+}(M) . 2. The torsion of \nabla^{({\mathcal{G}},\alpha)} is given by {\rm Tor}^{({\mathcal{G}},\alpha)}(A,B)\;=\;\frac{\alpha+1}{2}\;{\rm Tor}^{({\mathcal{G}},1)}(A,B)\;=\;\frac{\alpha+1}{2}\;\big({\mathcal{K}}^{\mathcal{G}}(B,A)-{\mathcal{K}}^{\mathcal{G}}(A,B)\big). (12) That is, apart from \nabla^{(-1)}=\nabla^{({\rm m})} , either all or none of the \nabla^{(\alpha)} , \alpha\neq-1 , have torsion. 3. \nabla^{({\mathcal{G}},\alpha)} and \nabla^{({\mathcal{G}},-\alpha)} are conjugate w.r.t. {\mathcal{G}} , i.e., we have the identity A\;{\mathcal{G}}(B,C)={\mathcal{G}}\left(\nabla^{({\mathcal{G}},\alpha)}_{A}B,C\right)+{\mathcal{G}}\left(B,\nabla^{({\mathcal{G}},-\alpha)}_{A}C\right) for all vector fields A,B,C on P^{\infty}_{+}(M) . In particular, \nabla^{({\mathcal{G}},0)} is compatible with {\mathcal{G}} , and the connection \nabla^{({\mathcal{G}},1)} has vanishing curvature. 4. The Levi-Civita connection of {\mathcal{G}} is given as \nabla^{LC}_{A}B=\nabla^{({\rm m})}_{A}B-\frac{1}{2}{\mathcal{K}}^{\mathcal{G}}(A,B)-\frac{1}{2}{\mathcal{K}}^{\mathcal{G}}(B,A)+\frac{1}{2}{\mathcal{D}}^{\mathcal{G}}(A,B), (13) where {\mathcal{D}}^{\mathcal{G}} is the symmetric tensor determined by the identity {\mathcal{G}}({\mathcal{D}^{\mathcal{G}}}(A,B),C)\;=\;{\mathcal{A}}^{\mathcal{G}}(C;A,B). (14) Proof Since {\mathcal{A}}^{\mathcal{G}} is tensorial, ( 9 ) immediately implies that {\mathcal{K}}^{\mathcal{G}} is tensorial in both entries as well. From this,