黎曼度量不变性

文章基于A Riemannian Framework for Tensor Computing论文推导黎曼度量不变性，过程比较粗糙，以后准备加log-map和exp-map以及黎曼均值求解过程，该篇论文还涉及该黎曼框架上的求导，标准坐标系，以及梯度，牛顿梯度的修正。
geodesics on tensor manifold going through $\Sigma$ with tangent vector $W$:
$${\Gamma }_{(\Sigma ,W)}(t)=e^{(\frac{1}{t}W)}\Sigma e^{(\frac{1}{t}W{)}^T}$$
where $W,\Sigma \in$ symmetric spaces
geodesics on tensor manifold going through $I$ with tangent vector $W$:
$${\Gamma }_{(I,W)}(t)=e^{(\frac{t}{2}W)}{e}^{(\frac{t}{2}W{)}^T}=e^{tW}$$
Ps，$e^{tW},e^{\frac{t}{2}W}$ are symmetric matric. If transformed $W$ as $W=Udiag(w_i)U^T$, $e^{tW}=Udiag(t{e}^{w_i})U^T={\left(e^{tW}\right)}^T,e^{\frac{t}{2}W}=Udiag(\frac{t}{2}{e}^{w_i})U^T={\left(e^{\frac{t}{2}W}\right)}^T$
the eigenvector matric of symmetric matric can be transformed as unitary matrix by unitization.
$$\begin{array}{rl}\frac{d\Gamma (t)}{dt}=& \frac{d{e}^{tW}}{dt}\\ =& \frac{dUdiag(t{e}^{w_i})U^T}{dt}\\ =& U{\left[\frac{d}{dt}t{e}^{w_i}\right]}_{n\ast n}{U}^T\\ =& Udiag(w_i{e}^{(t{w}_i)})U^T\\ =& Udiag(e^{\frac{t{w}_i}{2}})diag(w_i)diag(e^{\frac{t{w}_i}{2}})U^T\\ =& Udiag(e^{\frac{t{w}_i}{2}})U^{T}Udiag(w_i)U^{T}Udiag(e^{\frac{t{w}_i}{2}})U^T\\ =& e^{(\frac{t}{2}W)}W{e}^{(\frac{t}{2}W)}\\ =& {\Gamma (t)}^{\frac{1}{2}}\ast W\end{array}$$
Ps, $\frac{d{e}^{At}}{dt}=A{e}^{At}$
We obtain geodesic starting from any other point of the manifold:
$${\Gamma }_{(\Sigma ,W)}(t)={\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}={\Sigma }^{\frac{1}{2}}{e}^{t{\Sigma }^{-\frac{1}{2}}W{\Sigma }^{-\frac{1}{2}}}{\Sigma }^{\frac{1}{2}}$$
$\because {\Sigma }^{-\frac{1}{2}}\in$ symmetric spaces, $\therefore {\Sigma }^{-\frac{1}{2}}W{\Sigma }^{-\frac{1}{2}}\in$ symmetric spaces $\therefore {\Sigma }^{-\frac{1}{2}}\ast W=U_{1}diag(d_i)U_1^T$
$$\begin{array}{rl}\frac{d{\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W}(t)}{dt}=& U_{1}diag(d_i{e}^{(t{d}_i)})U_1^T\\ =& U_{1}diag(e^{\frac{t{d}_i}{2}})diag(d_i)diag(e^{\frac{t{d}_i}{2}})U_1^T\\ =& U_{1}diag(e^{\frac{t{d}_i}{2}})U_1^T{U}_{1}diag(d_i)U_1^T{U}_{1}diag(e^{\frac{t{d}_i}{2}})U_1\\ =& e^{(\frac{t}{2}{\Sigma }^{-\frac{1}{2}}\ast W)}{\Sigma }^{-\frac{1}{2}}\ast W{e}^{(\frac{t}{2}{\Sigma }^{-\frac{1}{2}}\ast W)}\\ =& {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W))}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\end{array}$$
$$\begin{array}{rl}\frac{d{\Gamma }_{(\Sigma ,W)}(t)}{dt}=& \frac{d\left({\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(t)\right)}{dt}\\ =& {\Sigma }^{\frac{1}{2}}\frac{d{\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(t)}{dt}{\Sigma }^{\frac{1}{2}}\\ =& {\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\end{array}$$
the norm at ${\Gamma }_{(\Sigma ,W)}(t)$ is
$$\begin{array}{rl}{\Vert {\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_{{\Gamma }_{(\Sigma ,W)}(t)}^2=& {\Vert {\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_{{\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(t)}^2\\ =& {\Vert {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_{{\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(t)}^2\\ =& {\Vert {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_{{\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast I}^2\\ =& {\Vert {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_I^2\\ =& {\Vert {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_2^2\end{array}$$
the length of the curve between time 0 and time 1 is:
$$\begin{array}{rl}L=& {\int }_0^1{\Vert \frac{d{\Gamma }_{(\Sigma ,W)}(t)}{dt}\Vert }_{{\Gamma }_{(\Sigma ,W)}(t)}^{2}dt\\ =& {\int }_0^1{\Vert {\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}^{\frac{1}{2}}(t)\ast {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_{{\Gamma }_{(\Sigma ,W)}(t)}^{2}dt\\ =& {\int }_0^1{\Vert {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_2^{2}dt\\ =& {\Vert {\Sigma }^{-\frac{1}{2}}\ast W\Vert }_2^2\end{array}$$
when ${\Gamma }_{(\Sigma ,W)}(1)=P$
$$\begin{array}{rl}{\Gamma }_{(\Sigma ,W)}(1)=& {\Sigma }^{\frac{1}{2}}\ast {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(1)=P\\ {\Gamma }_{(I,{\Sigma }^{-\frac{1}{2}}\ast W)}(1)=& {\Sigma }^{-\frac{1}{2}}\ast P=e^{{\Sigma }^{-\frac{1}{2}}\ast W}\\ W=& {\Sigma }^{\frac{1}{2}}\ast log({\Sigma }^{-\frac{1}{2}}\ast P)\end{array}$$
therefore, the length $L$:
$$\begin{array}{rl}L=& {\Vert {\Sigma }^{-\frac{1}{2}}\ast {\Sigma }^{\frac{1}{2}}\ast log({\Sigma }^{-\frac{1}{2}}\ast P)\Vert }_2^2\\ L=& {\Vert log({\Sigma }^{-\frac{1}{2}}\ast P)\Vert }_2^2\end{array}$$