Interpretation of skewness tensor in information geometry

65 Views Asked by At

In many textbooks about the information geometry (for instance, "Information Geometry with Applications" by Amari), authors emphasize that statistical manifold is described by tripple $\{\mathcal{M},G,T\}$, where $\mathcal{M}$ is manifold, $G$ is metric on it and $T$ is third-order tensor, which is called skewness tensor. The authors also emphasize the difference between this structure and the Riemannian structure $\{\mathcal{M},G\}$.

My question: the tensor $T_{ijk}$ is symmetric, whereas in physics I know only one third-order tensor, torsion tensor, which is anti-symmetric. How should I interpret the third order symmetric tensor? What is the most suitable physical analogy for this tensor?

Additional info: statistical manifolds appear during the study of distance between probability measures. For family of probability measures $p=p(x,\xi)$, where $x$ corresponds to random variable and $\xi$ corresponds to parameters ($\xi$ is vector!), there is a unique metric tensor, defined by $$g_{ij}=\int dx\,p(x,\xi)\frac{\partial\ln p(x,\xi)}{\partial \xi_i}\frac{\partial\ln p(x,\xi)}{\partial \xi_j},$$ and there is a unique skewness tensor defined by $$T_{ijk}=\int dx\,p(x,\xi)\frac{\partial\ln p(x,\xi)}{\partial \xi_i}\frac{\partial\ln p(x,\xi)}{\partial \xi_j}\frac{\partial\ln p(x,\xi)}{\partial \xi_k}.$$

In his book, Amari introduces the third-order tensor via the concept of dual connections. Let two vectors, $A$ and $B$. Let $\langle A,B\rangle$ is the scalar product of this two vectors. Let two symmetric affince connections $\Gamma$ and $\Gamma^*$. Let $\Pi$ and $\Pi^*$ are parallel transport operations w.r.t such connections. So, we have $$\langle A,B\rangle =\langle \Pi A, \Pi^*B\rangle.$$ Next, the skewness tensor is $$T_{ijk}=\Gamma_{ijk}^*-\Gamma_{ijk}.$$

Note: I understand that this question may very specific and there are not lot of details, but it corresponds to my current understanding.