8
$\begingroup$

In an $n$-dimensional Euclidean space, for any unit vector $$\boldsymbol{y}=(y_1,y_2,\cdots, y_n)\in\mathbb{S}^{n-1}=\{\boldsymbol{x}\in\mathbb{R}^n:\|\boldsymbol{x}\|_2=1\},$$ we can express $\boldsymbol{y}$ in the spherical coordinate system via $\phi(\boldsymbol{y})=(\phi_1(\boldsymbol{y}), \phi_2(\boldsymbol{y}), \ldots, \phi_{n-1}(\boldsymbol{y}))^{\mathrm{T}}\in\mathbb{R}^{n-1}$ s.t. $$ \begin{aligned} y_1 &=\cos \phi_1 \\ y_2 &=\sin \phi_1 \cos \phi_2 \\ y_3 &=\sin \phi_1 \sin \phi_2 \cos \phi_3 \\ & ~\,\,\vdots \\ y_{n-1} &=\sin \phi_1 \cdots \sin \phi_{n-2} \cos \phi_{n-1} \\ y_{n} &=\sin \phi_1 \cdots \sin \phi_{n-2} \sin \phi_{n-1}, \end{aligned} $$ where $0 \leq \phi_{n-1}<2 \pi$, and $0 \leq \phi_{i}\le \pi$, $\forall\,i=1,2,\ldots,n-2$.

My question is, if the Euclidean distance is no more than the spherical distance, i.e., if one has $$ \boxed{\|\boldsymbol{x}-\boldsymbol{y}\|_2 \leq\|\phi(\boldsymbol{x})-\phi(\boldsymbol{y})\|_2}. $$ I have been thinking this for several days, but am still unable to prove it. What is easy to show is that $\|\boldsymbol{x}-\boldsymbol{y}\|_2 \leq\|\phi(\boldsymbol{x})-\phi(\boldsymbol{y})\|_1$, which simply follows from the fact that the Euclidean distance is the shortest between any two points, and $|\phi_i(\boldsymbol{x})-\phi_i(\boldsymbol{y})|$ is exactly the length of the arc used to align $\boldsymbol{x}$ and $\boldsymbol{y}$ along the $i$-th spherical coordinate, which is longer than the corresponding chord.

If this is not true, then does it hold for any two close vectors $\boldsymbol{x}$ and $\boldsymbol{y}$ in the sense that $\underset{i=1,2,\ldots,n}{\max}|\phi_i(\boldsymbol{x})-\phi_i(\boldsymbol{y})|\le\delta$ for some small $\delta$?

$\endgroup$
6
  • $\begingroup$ Just checking, by $\|v\|_1$ do you mean $|v_1|+\dots +|v_n|$? $\endgroup$ Commented Sep 27, 2022 at 9:06
  • $\begingroup$ @CalvinKhor Yes, I do mean this. $\endgroup$ Commented Sep 27, 2022 at 9:20
  • $\begingroup$ That's an usual definition of "spherical distance." $\endgroup$ Commented Oct 28, 2022 at 6:49
  • 1
    $\begingroup$ @jwguan It's not a matter about naming. The point is that even for $S^1$ the definition is not natural, e.g. $φ(\cos θ,-\sin θ)=2\mathrm π-θ$ for small $θ$ and $\|φ(\cos θ,-\sin θ)-φ(1,0)\|=2\mathrm π-θ$, but the distance should be more naturally defined to be $θ$. $\endgroup$ Commented Oct 28, 2022 at 9:35
  • 1
    $\begingroup$ "and $|\phi_i(\mathbf x)-\phi_i(\mathbf y)|$ is exactly the length of the arc" - No, the arc is generally shorter. E.g. near the north pole, two points at the same latitude can have very different longitudes, but the arc length is small. $\endgroup$ Commented Nov 1, 2022 at 22:17

3 Answers 3

Reset to default
6
+50
$\begingroup$

Yes, the Euclidean distance is no more than "the spherical distance" as defined in the question.

Christophe Leuridan's answer illustrates how this can be proved by "the mean value inequality for vector-to-vector maps", such as proposition 2.7.6 of Lecture Notes on Multivariable Calculus. This answer provides a more elementary proof.

Let $n\ge2$, $\boldsymbol x, \boldsymbol y\in \Bbb S^{n-1}$. Suppose $\phi(\boldsymbol x)=(\alpha_1, \cdots, \alpha_n)$ and $\phi(\boldsymbol y)=(\beta_1,\cdots, \beta_n)$.
What we need to prove is the following claim, where the LHS of the inequality is $( \|\boldsymbol{x}-\boldsymbol{y}\|_2)^2$ while the RHS is $(\|\phi(\boldsymbol{x})-\phi(\boldsymbol{y})\|_2)^2$.

Claim: we have $$ \sum_{i=1}^{n-1}(\cos\alpha_i\prod_{j=1}^{i-1}\sin\alpha_j-\cos\beta_i\prod_{j=1}^{i-1}\sin\beta_j)^2 + (\prod_{j=1}^{n-1}\sin\alpha_j-\prod_{j=1}^{n-1}\sin\beta_j)^2 \le \sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2 $$

Proof. Do induction on $n$.

For $n=2$, $$\begin{aligned}\text{LHS}&=(\cos\alpha_1-\cos\beta_1)^2 + (\sin\alpha_1-\sin\beta_1)^2\\ &=2-2\cos(\alpha_1-\beta_1)=4\sin^2(\frac{\alpha_1-\beta_1}2)\le (\alpha_1-\beta_1)^2=\text{RHS.}\end{aligned}$$

For $n+1$, let $\displaystyle{A=\prod_{j=1}^{n-1}\sin\alpha_j}$, $\displaystyle{B=\prod_{j=1}^{n-1}\sin\beta_j}$. Note $|AB|\le1.$

$$\begin{aligned} \text{LHS } =&\sum_{i=1}^{n}(\cos\alpha_i\prod_{j=1}^{i-1}\sin\alpha_j-\cos\beta_i\prod_{j=1}^{i-1}\sin\beta_j)^2 + (\prod_{j=1}^{n}\sin\alpha_j-\prod_{j=1}^{n}\sin\beta_j)^2 \\ &\quad(\text{Apply the induction hypothesis.})\\ \le&\sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2-(A-B)^2 +(A\cos\alpha_n-B\cos\beta_n)^2+(A\sin\alpha_n-B\sin\beta_n)^2\\ =&\sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2 + (2-2(\cos\alpha_n\cos\beta_n+\sin\alpha_n\sin\beta_n))AB\\ =&\sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2 + (2-2\cos(\alpha_n-\beta_n))AB\\ \le&\sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2 + (2-2\cos(\alpha_n-\beta_n))\\ \le&\sum_{i=1}^{n-1}(\alpha_i-\beta_i)^2 + (\alpha_n-\beta_n)^2\\ =&\text{ RHS} \quad\Box \end{aligned}$$

Tracing the conditions for the equalities that appear in the proof above, we can verify immediately that $\|\boldsymbol{x}-\boldsymbol{y}\|_2=\|\phi(\boldsymbol{x})-\phi(\boldsymbol{y})\|_2$ iff $\boldsymbol x = \boldsymbol y$.

$\endgroup$
4
$\begingroup$

If I reformulate correctly, you express $y \in \mathbb{S}_{n-1}$ as a function $f$ of $\phi = (\phi_1,\ldots,\phi_{n-1}) \in \mathbb{R}^{n-1}$, and you ask whether this function $f$ is $1$-Lipschitz or not. Since $f$ is differentiable, it suffices to check that $||df(\phi)|| \le 1$ is everywhere. The operator norm $||df(\phi)||$ is the square root of the largest eigenvalue of $df(\phi)^\top \times df(\phi)$.
When $n=3$, the Jacobian matrix is
$$df(\phi) = \left( \begin{array}{cc} -\sin \phi_1 & 0 \\ \cos\phi_1 \cos\phi_2 & -\sin\phi_1 \sin\phi_2\\ \cos\phi_1 \sin\phi_2 & \sin\phi_1 \cos\phi_2 \end{array} \right)$$ One checks that $df(\phi)^\top \times df(\phi)$ is diagonal with diagonal entries $1$ and $\sin^2\phi_1$. My impression is that the same method works in higher dimensions, with heavier formulas.

$\endgroup$
1
$\begingroup$

The spherical distance between $x,x'\in\mathcal{S}^{n-1}$ is just $d(x,x')=\arccos(x^Tx')$. The Euclidean distance is $\|x-x'\|_2=\sqrt{2(1-x^Tx')}$. So both are just a function of $x^Tx'\in[-1,1]$. At $x^Tx'=1$, both expressions are zero. Moreover for $t\in(-1,1)$, the derivative of $t\mapsto \arccos(t)$ is smaller than the derivative of $t\mapsto \sqrt{2(1-t)}$. So, $\arccos(t)\geq \sqrt{2(1-t)}$ for all $t\in[-1,1]$. We conclude that $d(x,x')\geq \|x-x'\|_2$ for all $x,x'\in\mathcal{S}^{n-1}$.

$\endgroup$
1
  • 4
    $\begingroup$ Thanks for the answer. But it seems that your ``spherical distance'' is quite different from the one in the question. $\endgroup$ Commented Oct 28, 2022 at 9:06

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.