Consider the two dimensional sphere $\mathbb{S}^2$ and let $p, q \in \mathbb{S}^2$. Let $\text{exp}_{p}$ and $\text{exp}_{q}$ be the exponential maps on $\mathbb{S}^2$ at points $p$ and $q$ respectively. I am interested in the map $\psi := \text{exp}_{p}^{-1} \circ \text {exp}_{q}$ defined on the unit disc $\mathbb{D} \subset \mathbb{R}^2$. I expect that if $p$ and $q$ are nearby points, then the map $\psi$ is close to the identity map. My question is: Is there a way to quantify this closeness? More precisely, is it possible to write the Taylor series expansion of $\psi = (\psi_1, \psi_2)$ around the origin as \begin{align*} \psi_1(x,y) &= \psi_1(0) + \frac{\partial \psi_1}{\partial x}(0)~x + \frac{\partial \psi_1}{\partial y}(0)~ y + \ldots \\ \psi_2(x,y) &= \psi_2(0) + \frac{\partial \psi_2}{\partial x}(0)~x + \frac{\partial \psi_2}{\partial y}(0)~ y + \ldots \end{align*} and show that the only significant coefficients in the above expansion are $\frac{\partial \psi_1}{\partial x}(0) \approx 1$, $\frac{\partial \psi_2}{\partial y}(0) \approx 1$ and all other partial derivatives $\frac{\partial^{m+n} \psi_i}{\partial^{m} x \partial^{n} y}(0) \approx 0$, for all other choices of $m,n$ and any $i = 1,2$.

Edit: After reading the comments/ answers of user35593 and Gro-Tsen , I realized that the following comments are essential regarding the above question:

  1. First is regrding the comment of user35593 (please read his/her comment below): The question makes sense only when we specify a certain way of identifying the tangent spaces at the points $p$ and $q$. Consider the rotation $R :\mathbb{R}^3 \rightarrow \mathbb{R}^3$ which maps $p$ to $q$, then it also maps $T_{p}(\mathbb{S}^2)$ to $T_{q}(\mathbb{S}^2)$ and it gives a way of identifying the two tangent spaces. The above question I asked is for the tangent spaces identified in this manner.

  2. Second is regarding the answer of Gro-Tsen : One possible interpretation of ``$\frac{\partial^{m+n} \psi_i}{\partial^{m} x \partial^{n} y}(0) \approx 0$'' which I am interested in is: Given $\epsilon >0$, is it possible to find $\delta >0$ such that whenever $d_{\mathbb{S}^2}(p,q)\leq \delta$ we have \begin{align*} \left| \frac{\partial^{m+n} \psi_i}{\partial^{m} x \partial^{n} y}(0) \right| \leq \epsilon \end{align*} whenever $m+n \leq 100$ (except for the cases of $\frac{\partial \psi_1}{\partial x} (0)$ and $\frac{\partial \psi_2}{\partial y} (0)$ which I expect to be $\approx 1$).


  • 4
    $\begingroup$ Have you tried writing down explicit formulae? Letting $\mathbb{S}^2 = \{x^2+y^2+z^2=1\} \subseteq \mathbb{R}^3$, we can assume $q=(0,0,1)$ and $p=(\sin δ,0,\cos δ)$, then $\exp_p(u,v) = (\sin r\,\cos θ, \sin r\,\sin θ, \cos r)$ where $(r,θ)$ are the polar coordinates of $(u,v)$, and $ψ$ takes $(u,v)$ to $(u',v')$ where $(\sin r'\,\sin θ', \cos r', \sin r'\,\cos θ') = (-\sin δ\,\cos r + \cos δ\,\sin r\,\cos θ, \sin r\,\sin θ, \cos δ\,\cos r + \sin δ\,\sin r\,\cos θ)$. At least the map $(r,θ)\mapsto (r',θ')$ looks reasonably explicit. $\endgroup$
    – Gro-Tsen
    Apr 18, 2020 at 9:58
  • 2
    $\begingroup$ Your map maps the tangent space of q to the tangent space of p. You need to specify how you identify the tangent spaces with $\mathbb{R}^2$, i.e. what bases you choose in order to arrive at a map $\mathbb{R}^2\rightarrow \mathbb{R}^2$. You could for example consider the geodesic from p to q. and identify the tangent vectors to the geodesic at p and q with each other. $\endgroup$
    – user35593
    Apr 18, 2020 at 10:48
  • 2
    $\begingroup$ What should be true is that for fixed $p$, and $q$ nearby, the maps $F_q = \exp_q^{-1} \exp_p$, considered as a map on (a ball in) $\Bbb R^2$ by using a fixed trivialization of $TS^2$ near $p$. These should vary smoothly in $q$, with $F_p$ given by the identity. Since $D^n F_p = 0$ for $n>1$, and the maps $F_p$ are smooth in $p$, the Taylor approximation should give the existence of $C_n$ so that $|D^n F_q| \leq C_n |p-q|$ for $q$ near to $p$. This seems like the sort of smallness estimate you want. Gro-Tsen's answer implies that $C_n \to \infty$, though. $\endgroup$
    – mme
    Apr 18, 2020 at 11:39
  • 1
    $\begingroup$ You can argue that your map and therefore also the partial derivatives at 0 depend smoothly on $p, q$. For $p=q$ you get the identity. Then the coefficients are exactly 1 resp. 0. For $p\noteq$ they are of order $d(p,q)$ away from the identity case. $\endgroup$
    – user35593
    Apr 18, 2020 at 14:49
  • 1
    $\begingroup$ It probably is easiest to consider, given a curve $c$ such that $c(0) = p$, the map $$ \psi_t = \exp_p^{-1}\circ\exp_{c(t)}: T_{c(t)}\mathbb{S} \rightarrow T_p\mathbb{S}. $$ In particular, its derivative with respect to $t$ at $t = 0$ is a essentially a Jacobi field and satisfies the Jacobi equation, which here is just $$ J'' + J = 0. $$ $\endgroup$
    – Deane Yang
    Apr 18, 2020 at 19:21

2 Answers 2


I believe the following proves that the partial derivatives of $\psi$ at the origin cannot be bounded by a constant (so they certainly cannot be $\approx 0$ for any reasonable meaning of this symbol).

Assume on the contrary that (for some fixed $p,q$ distinct and not antipodal), the partial derivatives of $\psi := \exp_p^{-1} \circ \exp_q$ (defined in some neighborhood of the origin) are all bounded by a constant. Then, by summing the Taylor series expansion of $\psi$ at $0$ we see that $\psi$ extends to a real-analytic function $\psi\colon\mathbb{R}^2 \to \mathbb{R}^2$, which by analytic extension must still satisfy $\exp_p \circ \psi = \exp_q$. Let me argue why this is impossible.

We can assume w.l.o.g. that the coordinates $(u,v)$ on the tangent plane to $\mathbb{S}^2$ at $q$ were chosen so that $\exp_q$ maps the axis $(u,0)$ to the great circle connecting $q$ and $p$, and more precisely, if $0<\delta<\pi$ is the distance between $q$ and $p$ on $\mathbb{S}^2$, that $\exp_q$ takes $(\delta,0)$ to $p$. Furthermore, we can similarly assume on the coordinates $(u',v')$ of the tangent plane at $p$ that $\exp_p$ maps the axis $(u',0)$ to the same great circle and takes $(-\delta,0)$ to $q$. Then $\exp_q(u,0)$ is the point obtained by traveling a distance $u$ on $\mathbb{S}^2$ starting from $q$ in the direction of $p$, and $\exp_p(u,0)$ is the point obtained by traveling a distance $u$ on $\mathbb{S}^2$ starting from $p$ in the direction opposite to $p$, thus $\psi(u,0) = (u-\delta,0)$ for $u$ in the neighborhood of $0$, hence everywhere by analytic extension.

On the other hand, if $(u,v)$ lies on the circle $C$ with radius $\pi$ around the origin then $\exp_q$ takes $(u,v)$ to the antipode $\tilde q$ of $q$. But the inverse image of $\tilde q$ by $\exp_p$ is discrete (since $\exp_p$ is a diffeomorphism outside of circles of radius $k\pi$ around the origin which are mapped to either $p$ or its antipode $\tilde p$, and we are assuming $p,q,\tilde p,\tilde q$ distinct); and $\psi$ must map $C$ (which is connected) inside this inverse image: so $\psi$ must be constant on $C$. But this contradicts the fact that $\psi(\pi,0) = (\pi-\delta,0)$ and $\psi(-\pi,0) = (-\pi-\delta,0)$ (as per previous paragraph) are not equal.

  • $\begingroup$ Nice answer! Though I only realized after the comment of user35593 that the way we identify the tangent spaces is very important and the question makes sense only if we identify it appropriately - like your observation above shows. Hence I have edited the question. The expectation is $\psi \approx \text{Id}$ only when $T_{p}(\mathbb{S}^2)$ and $T_{q}(\mathbb{S}^2)$ are identified like I have mentioned in the edited version. This is not the way you have identified the tangent spaces though. $\endgroup$
    – April
    Apr 18, 2020 at 12:08
  • $\begingroup$ It is the way I identified the tangent spaces. However, now that you've edited the question to clarify what you mean by $\approx 0$, and the order of the quanrifiers, this doesn't answer it. But it's not related to the identification of the tangent spaces. $\endgroup$
    – Gro-Tsen
    Apr 18, 2020 at 14:12
  • $\begingroup$ The derivatives being bounded would not necesssrily imply the Taylor series to coverge unless you mean bounded by a common constant in which case we would only have convergence on a square. $\endgroup$
    – user35593
    Apr 18, 2020 at 14:52

Not a solution but too long for a comment.

Note that the map depends only on the distance $d(p,q)$ of $p,q$, i.e. we have $\phi(t)\colon \mathbb{R}^2\rightarrow \mathbb{R}^2$ where $t=d(p,q)$. Note further that $\phi(t_1+t_2)=\phi(t_1)\circ \phi(t_2)$, i.e. $\phi(t)$ is a one-parameter group. Let $D=d\phi(t)/dt|_{t=0}$. Note that then $d\phi(t)/dt=D(\phi(t))$. Hence $\phi(T)(u)$ can be found by solving the ODE $f'(t)=D(f(t))$ on the interval $[0,T]$ with initial condition $f(0)=u$. Let us now find $D$ for our case of the sphere.

The exponential and its inverse on the sphere are $exp(v)=\begin{pmatrix} sin(|v|)\frac{v}{|v|}\\ cos(|v|) \end{pmatrix}$ and $log \left(\begin{pmatrix} x\\ y\\z \end{pmatrix}\right)=\frac{arccos(z)}{\sqrt{1-z^2}}\begin{pmatrix} x\\ y \end{pmatrix}$. Applying and infinitesimal rotation by $\alpha$ after the exponential map yields $$ \begin{pmatrix} sin(|v|)\frac{v_1}{|v|}-\alpha \cdot cos(|v|)\\sin(|v|)\frac{v_2}{|v|} \\ cos(|v|)+\alpha \cdot sin(|v|)\frac{v_1}{|v|} \end{pmatrix} $$ Now applying the logarithm using $$ d\frac{arccos(z)}{\sqrt{1-z^2}}/dz=\frac{arccos(z)z-\sqrt{1-z^2}}{\sqrt{1-z^2}^3} $$ yields $$ v+\alpha \cdot sin(|v|)\frac{v_1}{|v|} \frac{(|v|cos(|v|)-sin(|v|))}{sin(|v|)^3} sin(|v|)\frac{v}{|v|} -\begin{pmatrix} \alpha \frac{cos(|v|)}{sin(|v|)}|v|\\0\end{pmatrix}\\ =v+\alpha \cdot \frac{(|v|cos(|v|)-sin(|v|))v_1}{sin(|v|)|v|^2} v -\begin{pmatrix} \alpha \frac{cos(|v|)}{sin(|v|)}|v|\\0\end{pmatrix} $$ Hence $$ D(v)=\frac{(|v|cos(|v|)-sin(|v|))v_1}{sin(|v|)|v|^2} v-\begin{pmatrix} \frac{cos(|v|)}{sin(|v|)}|v|\\0\end{pmatrix}. $$


Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.