This is regarding a clarification in page 384 of a paper published in Annals of Statistics by Amari.

In page no. 384, he defines $$R_i(t)=\frac{\partial}{\partial \theta_i} D_{\alpha}(q(x,t),p(x,\theta))$$ where for $|\alpha|\neq 1$,

\begin{eqnarray*} D_{\alpha}(q(x,t),p(x,\theta))=\frac{8}{1-\alpha^2}\left[1-\int q(x,t)^{\frac{1+\alpha}{2}}p(x,\theta)^{\frac{1-\alpha}{2}}~dx\right] \end{eqnarray*}

The one I am struggling to understand is that how did the author obtain the differential equation (A.19): $$\ddot{R_i}(t)=-\frac{1-\alpha^2}{4}i(t)R(t)$$ where, I think, $i(t)=\int \frac{1}{q(x,t)}(\dot{q}(x,t))^2~dx$ from bottom of page 382.

When I derived manually I obtained (by allowing interchange of differentiation and integral) $$\ddot{R_i}(t)=\int p(x,\theta)^{-\frac{1+\alpha}{2}}\left[q(x,t)^{\frac{\alpha-1}{2}}\ddot{q}(x,t)+\frac{\alpha-1}{2}q(x,t)^{\frac{\alpha-3}{2}}\dot{q}(x,t))^2\right]\frac{\partial}{\partial \theta_i}(p(x,\theta)~dx.$$ I am wondering how can the right hand side of the above be written as a product of two integrals as in (A.19). As this is appeared in Annals of Statistics, I am not in a position to doubt this as well.

It would be of great help if anyone who knows differential/Riemannian geometry can clarify why should (A.16) be true under the mentioned circumstances.

NB: I asked the above question in math.stackexchange where I couldn't get any response so far.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.