Mathematics Stack Exchange is a question and answer site for people studying math at any level and professionals in related fields. Join them; it only takes a minute:

Sign up
Here's how it works:
  1. Anybody can ask a question
  2. Anybody can answer
  3. The best answers are voted up and rise to the top

Consider an inner product space $V$ and $v \in V$.

It seems that the only scalar invariants a vector can have under the isometry come from $\langle v, v \rangle$. For example $\langle v, v \rangle$, $2\langle v, v \rangle^2 - 3$ or $\sin(\langle v, v \rangle)$ for real spaces etc. But how can it be proved?

I'm interested not just in a proof as is, but in insights behind it. Various proofs and/or references would be nice.

share|cite|improve this question
What do you mean by $v^2$ ? – Belgi Oct 23 '12 at 11:31
up vote 1 down vote accepted

I suppose that by "the isometry" you mean the orthogonal group of all isometries. The orbits of that group are easily shown to be the sets of the $v$ with $\langle v,v\rangle=c$ for some constant $c$. Now if you have any invariant, its level sets must be unions of orbits, which for a polynomial invariant can only be a finite union for reasons of dimension. It can be seen algebraically that any polynomial function of the coordinates whose level sets are unions of sets of the form $\langle v,v\rangle=c$ is itself a polynomial in $\langle v,v\rangle$.

Added, since apparently not so obvious, an explanation why orbits are sets of the $v$ with $\langle v,v\rangle=c$ for some constant $c$. On one hand, since any isometry preserves the scalar product, it is clear that the value of $\langle v,v\rangle$ cannot vary along an orbit: $\langle v',v'\rangle=\langle g(v),g(v)\rangle=\langle v,v\rangle$ whenever there is some isometry $g$ mapping $v$ to $v'$. On the other hand if $\langle v,v\rangle=\langle v',v'\rangle=c>0$ we must find an isometry $g$ mapping $v$ to $v'$: note that $v/\sqrt c$ and $v'/\sqrt c$ are vectors of unit length, so each one can be (individually) extended to an orthonormal basis; then the linear map sending the orthonomal basis extended from $v$ to the one extended from $v'$ is such an isometry $g$.

share|cite|improve this answer
I'll need some time to understand it, but I suspected the answer would be something like that. If somebody knows a reference to a detailed derivation, you are welcome to share. – Yrogirg Oct 23 '12 at 17:27
"Now if you have any invariant, its level sets must be unions of orbits" Could you please elaborate this part? Is that a theorem or what? – Yrogirg Oct 25 '12 at 13:04
Being invariant means precisely being constant on orbits. Therefore the set where an invariant takes a specific value is necessarily a union of (zero or more) orbits. – Marc van Leeuwen Oct 25 '12 at 13:34
"The orbits of that group are easily shown to be the sets of the $v$ with $\langle v,v\rangle=c$ for some constant $c$." How do you do it? And how would you do it for cases when the group acts not just on $V$, but for example on $V \otimes V^*$ or something of that kind? Oh should I ask it as a separate question? – Yrogirg Oct 29 '12 at 8:28
@Yrogirg: See the added paragraph. The case where the group of isometries acts on $V\otimes V^*$ is another kettle of fish; the elements of that space are not vectors of $V$ so there is no such condition as $\langle v,v\rangle=c$ to begin with. Elements $V\otimes V^*$ of can be identified with linear maps $V\to V$ (endomorphisms), and even for the action of all of $GL_n(\Bbb R)$, the orbits are not described by polynomial invariants alone in this case. – Marc van Leeuwen Oct 30 '12 at 9:45

Let $u$ and $v$ be two vectors with $u^2 = v^2=L$. I will show that there is an isometry taking $u$ to $v$. Specifically, I will reflect in the hyperplane orthogonal to $u-v$. Here is the computation:

Let $w = u - v$. If $w=0$, then $u=v$ and we are done, so we may assume $w \neq 0$. Reflection in $w^{\perp}$ is given by the formula $$x \mapsto x - 2 \frac{\langle w,x \rangle}{\langle w,w \rangle} w.$$ Since $w \neq 0$, the denominator is not $0$.

We have $\langle u,w \rangle = \langle u,u \rangle-\langle u,v \rangle = L - \langle u,v \rangle$ and $\langle w,w \rangle - \langle u,u \rangle - 2 \langle u,v \rangle + \langle v,v \rangle = 2L - 2 \langle u, v \rangle$. So $\langle u,w \rangle/ \langle w,w \rangle = 1/2$. (Challenge: Explain this formula in terms of Euclidean geometry!) Plugging into the formula for the reflection, $$u \mapsto u-2 (1/2) w=u-(u-v) = v.$$ Similarly, $v \mapsto u$.

Let's spell out why this means that every invariant function must be of the form $f(\langle v, v \rangle)$ for some function $f$. Let $g: \mathbb{R}^n \to \mathbb{R}$ be invariant under isometries. Define $f(x) = g(\sqrt{L}, 0,0,0, \ldots, 0)$. Then for any vector $v$ with $v^2=L$, there is an isometry taking $v$ to $(\sqrt{L},0,0,\ldots,0)$ (by the above). So $g(v) = g(\sqrt{L},0,0,0\ldots,0) = f(L)$. In short, for any invariant $g$, there is a function $f:\mathbb{R}_{+} \to \mathbb{R}$ so that $g(v) = f(v^2)$ for all $v$.

share|cite|improve this answer
It's not the answer to the question. Why there is no other function $h : V \to \mathbb R$ so that if $h(v) = h(u) = L$ there exists an isometry $u \mapsto v$? For a simplicity I took real inner product spaces. – Yrogirg Oct 23 '12 at 17:26
As you say, there are lots of such functions: $\sqrt{v^2}$, $e^{v^2}$, etc. But they are all functions of $v^2$, as this argument shows. See if the last paragraph I just added helps. – David Speyer Oct 23 '12 at 23:42

Your Answer


By posting your answer, you agree to the privacy policy and terms of service.

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