In this great question by Nathaniel Johnston, and in its answers, we can learn the following remarkable inequality: For all $v,w \in \mathbb{R}^n$ we have \begin{align*} \|v^2\| \, \|w^2\| - \langle v^2, w^2 \rangle \le \|v\|^2 \|w\|^2 - \langle v,w \rangle^2; \quad (*) \end{align*} here, $\langle\cdot,\cdot\rangle$ denotes the standard inner product on $\mathbb{R}^n$, $\|\cdot\|$ denotes the Euclidean norm and $v^2,w^2 \in \mathbb{R}^n$ denote the elementwise squares of $v$ and $w$. Both sides of $(*)$ are nonnegative by the Cauchy-Schwarz inequality, and the LHS gives a non-zero bound for the right RHS, in general.

What strikes me is that the RHS and the LHS of $(*)$ have different (linear) symmetry groups: the RHS does not change if we apply any orthogonal matrix $U \in \mathbb{R}^{n \times n}$ to both $v$ and $w$, while this is not true for the LHS. Hence, we can strengthen $(*)$ to \begin{align*} \sup_{U^*U = I}\Big(\|(Uv)^2\| \, \|(Uw)^2\| - \langle (Uv)^2, (Uw)^2 \rangle\Big) \le \|v\|^2 \|w\|^2 - \langle v,w \rangle^2. \quad (**) \end{align*} Unfortunately, I have no idea how to evaluate the LHS of $(**)$.

Question. Can we explicitely evaluate the LHS of $(**)$? Or, more generally, is there a version of $(*)$ for which both sides are invariant under multiplying both $v$ and $w$ by (identical) orthogonal matrices?

Admittedly, this question is a bit vague since it might depend on one's perspective which expressions one considers to be "explicit" and which inequalities one considers to be a "version" of $(*)$. Nevertheless, I'm wondering whether some people share my intuition that there should be a more symmetric version of $(*)$.

Edit. Maybe it is worthwhile to add the following motivating example: If we choose $n = 2$ and $v = (1,1)/\sqrt{2}$, $w = (1,-1)/\sqrt{2}$, then those vectors are orthogonal and the RHS of $(*)$ equals $1$, while the LHS of $(*)$ vanishes since $v^2$ and $w^2$ are linearly dependent. However, the LHS of $(**)$ is also equal to $1$; to see this, choose $U = \frac{1}{\sqrt{2}} \begin{pmatrix} 1 & -1 \\ 1 & 1 \end{pmatrix} $ and observe that $Uv = (0,1)$, $Uw = (1,0)$.


$(**)$ is always equality. By homogenuity we may suppose that $\|v\|=\|w\|=1$, then $Uv,Uw$ are any two unit vectors with prescribed inner product $\langle Uv,Uw\rangle =\langle u,v\rangle$. I claim that you may find two two-dimensional vectors with prescribed given inner product such that equality in $(*)$ holds. Then you may make $n$-dimensional examples by adding zero coordinates. Well, just look at GH from MO's proof of (*) and look what we need for equality. If $n=2$, it is sufficient that $F(v,w):=v_1w_1-v_2w_2=0$. Since $F((v_1,v_2),(w_1,w_2))=-F((w_2,w_1),(v_2,v_1))$, we may rotate the vector $(v_1,v_2)$ until it becomes equal $(w_2,w_1)$, at this moment the vector $(w_1,w_2)$ becomes equal to $(v_2,v_1)$ (since the inner product and orientation define the pair of unit vectors upto rotation). By continuity $F$ attains zero value at some intermediate point.

  • $\begingroup$ Very nice idea to introduce the intermediate value theorem! Do you think we can conclude from your proof that for $v,w$ in general position, there are exactly $\binom n2$ “essentially different” extremal matrices, one for each pair of coordinates? (By “essentially different” I mean up to permutations and/or sign changes of rows of $U$.) $\endgroup$
    – Wolfgang
    Nov 19 '18 at 9:37

This is not an answer, rather a heuristic argument.

I have done some numerical experiments for pairs of randomly chosen $v,w$, where I have used products of random Householder matrices for the $U$'s. They suggest that the LHS can come arbitrarily close to the RHS. So there should be a set of extremal unitary matrices $U$ for which equality is attained. For $n=2$ this $U$ seems unique up to sign (but the equation for it is already rather messy for the general case of $v,w$), and for $n\geqslant3$ there seems to be a whole bunch of them. (a continuum? or rather a finite number of isolated matrices?)

By homogeneity, we can assume wlog $\|v\|= \|w\|=1$. Then it would be an interesting question to what those extremal unitary matrices, which are defined only by the pair $\{v,w\}\in\mathbb R^n\times\mathbb R^n$, correspond geometrically.

  • $\begingroup$ Is the unicity phenomenon you noticed anyhow related to the fact that for $ n=2 $ one can establish an isomorphism from $ \mathbb{R}^{n} $ to an algebraically closed field? $\endgroup$ Nov 18 '18 at 21:35
  • $\begingroup$ @SylvainJULIEN I do not see why algebraic closedness should play a role here. But sure if $u,v$ for $n=2$ have, say, algebraic coordinates, the entries of $U$ will be algebraic as well. (Less sure for $n>2$.) $\endgroup$
    – Wolfgang
    Nov 18 '18 at 21:40
  • $\begingroup$ My utterly vague idea was to draw a parallel with the fact that in the quaternions division algebra, a given element has "a whole bunch" of square roots. $\endgroup$ Nov 18 '18 at 21:49
  • 4
    $\begingroup$ If you know the answer for $n=2$, can't you get the answer in general by a unitary transformation which takes the two-dimensional space generated by $v$ and $w$ to the two-dimensional space generated by $e_1$ and $e_2$? $\endgroup$
    – Will Sawin
    Nov 18 '18 at 22:39

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.