# A Question about Orthogonal Operators

Let $X$ be a finite-dimensional inner product space, and $T$ a linear operator on $X$. Let $W$ be a subset of $X$ with the following property:

If $T$ preserves norms on $W$, then $T$ is orthogonal on $X$.

One might say that the set $W$ 'detects' orthogonality.

My question is: what is the smallest number of elements that $W$ may contain (as a function of dim$(X) = n$) ?

Quite simply, we can observe that $n + {n \choose 2}$ would form an upper bound for the minimal cardinality of $W$. However, I cannot think of a way to proof whether or not this upper bound is also a lower bound.

-
It's not quite clear what you're saying. Does that condition on $W$ hold for all $T$? –  lhf Mar 2 '11 at 1:41
The idea is that $W$ is "big" enough that if $T$ preserves norms on $W$, it is orthogonal on the entire space. –  Isaac Solomon Mar 2 '11 at 2:26

Suppose a set $W$ is a 'detector'. This means that for any matrix $T$, if $w'T'Tw = w'w$ for all $w \in W$ then $T$ must be orthogonal. This is a set of $|W|$ linear equations on $T'T$.

If $|W| < \binom{n+1}{2}$, then there is a non-zero symmetric matrix $A$ such that $w'Aw = 0$ for all $w \in W$. Note that $A$ has some non-zero eigenvalue.

Clearly $w'Iw = w'w$ for all $w \in W$, and so for all $\epsilon \in \mathbb{R}$, we have $w'(I+\epsilon A)w = w'w$ for all $w \in W$. Choosing $\epsilon > 0$ small enough, we get some matrix $P = I+\epsilon A$ which is positive semidefinite, and thus has a square root $T$ such that $P=T'T$.

Since $\epsilon \neq 0$, $P$ has some positive eigenvalue different than $1$. Thus $T$ has some eigenvalue which is not of unit magnitude, and so cannot be orthogonal.

Addendum: This argument shows that $W$ is a detector iff the outer products $\{ w \otimes w : w \in W\}$ generate the linear space of all symmetric matrices. An example is the set $$W = \{ e_i+e_j : 1 \leq i \leq j \leq n \}.$$

What follows answers the related question where $T$ preserves all inner products on $W$.

You can take any basis of $X$ as $W$: if $T$ preserves all inner products among a basis, it will preserve all inner products.

On the other hand, if you take less than $n$ elements, then consider the projection into their linear span. On $W$ it acts as the identity, so it preserves inner products. But its not even regular.

The same arguments show that $W$ detects orthogonality iff it spans $X$.

-

Edit: This answer gives a construction of $W$ which shows that it need not contain more than ${n+1\choose 2}$ elements (if $X$ is real), but does not answer the question posed, as Yuval's argument does.

The easiest way to construct a finite set $W\subset X$ which detects orthogonality is as follows:

Let $\{\alpha_1,\dots,\alpha_n\}$ be a basis for $X$. Let

$W=\{\alpha_j+c\alpha_k:j\leq k;c=0,1\text{ or }i\}.$

Then $|W|=n^2$. Suppose $T$ is a linear operator on $X$ such that $\|T\alpha\|=\|\alpha\|$ for every $\alpha\in W$. If $j\leq k$, then

$\langle T\alpha_j,T\alpha_k\rangle=\frac{1}{2}\left(\|T\alpha_j+T\alpha_k\|^2+i\|T\alpha_j+iT\alpha_k\|^2\right)-\left(\|T\alpha_j\|^2+\|T\alpha_k\|^2\right)$

$=\frac{1}{2}\left(\|T(\alpha_j+\alpha_k)\|^2+i\|T(\alpha_j+i\alpha_k)\|^2\right)-\left(\|T\alpha_j\|^2+\|T\alpha_k\|^2\right)$

$=\frac{1}{2}\left(\|\alpha_j+\alpha_k\|^2+i\|\alpha_j+i\alpha_k\|^2\right)-\left(\|\alpha_j\|^2+\|\alpha_k\|^2\right)$

$=\langle \alpha_j,\alpha_k\rangle.$

If $j>k$, the above calculation shows that $\langle T\alpha_k,T\alpha_j\rangle=\langle \alpha_k,\alpha_j\rangle$, hence $\langle T\alpha_j,T\alpha_k\rangle=\overline{\langle T\alpha_k,T\alpha_j\rangle}=\overline{\langle \alpha_k,\alpha_j\rangle}=\langle\alpha_j,\alpha_k\rangle.$ Thus $T$ preserves inner products, and $T$ is unitary. Notice that if $X$ is real, $W$ need not contain more than ${n+1\choose2}$ elements; simply choose the set above, without the $n\choose2$ elements $\alpha_j+i\alpha_k$, $j\leq k$.

-
Right, I misread the question. –  Yuval Filmus Mar 2 '11 at 2:25
Now I also have a real answer. –  Yuval Filmus Mar 2 '11 at 3:19
I tried to vote your answer up but I don't have the requisite '15 reputation.' Could you answer what is probably a dumb question about your solution? Why if $|W|<{n+1\choose2}$ must there exist the non-zero symmetric matrix $A$ such that $w'Aw=0$ for all $w\in W$? I understand that the system of $|W|$ equations $w'T'Tw=w'w$ in the ${n+1\choose2}$ variables $T'T$ is invertible only if $|W|\geq{n+1\choose2}$, I am just having trouble making the leap from there. I have been thinking about this problem for a while along with the original poster, which is why I'm interested. –  Nick Strehlke Mar 2 '11 at 6:31
The space of all symmetric matrices has dimension $\binom{n+1}{2}$. The solutions of the non-homogeneous system $w'Aw = w'w$ are $I+B$, where $B$ are the solutions for the corresponding homogeneous system $w'Bw = 0$. If $|W|<\binom{n+1}{2}$ then this system is underspecified and so has a non-trivial solution. –  Yuval Filmus Mar 2 '11 at 19:38
Right, I thought about that for a while after reading your solution, which is very elegant. Thanks for explaining that. –  Nick Strehlke Mar 2 '11 at 23:49