3
$\begingroup$

I believe I found a complicated proof by bounding the spectral norm $||uu^T-vv^T||^2_2:=\max_{||x||=1}|(u^Tx)^2-(v^Tx)^2|$. Using the fact that $dist(x,y):=\sin|x-y|$ is a distance function over unit vectors, we may prove $$(v^Tx)^2-(u^Tx)^2=\sin^2(u,x)-\sin^2(v,x)=dist^2(u,x)-dist^2(v,x) =(dist(u,x)-dist(v,x))(dist(u,x)+dist(v,x)) \leq 2dist(u,v),$$ but the approximation factor is too large. Can you improve the bound using small changes to the above, or find another simple proof using trigo or Taylor's approx?

Note: This is a follow-up on a previous question that was stated wrong here. I hope it is correct this time.

$\endgroup$
2
  • 1
    $\begingroup$ I am a bit confused what you are asking for. I think I have an elementary but somewhat brute-force proof of the inequality in the title of your question, requiring only Cauchy-Schwarz. Would that be enough, or are you asking about the particular approach that you propose in the question? $\endgroup$
    – Yemon Choi
    Feb 26, 2022 at 15:56
  • $\begingroup$ Any answer will be welcomed! Thanks. $\endgroup$ Feb 28, 2022 at 1:14

4 Answers 4

7
$\begingroup$

This is a variation of Fedor Petrov's proof (found independently). First, it suffices to prove the statement when $x$ lies in $V:=\mathbb{R}u+\mathbb{R}v$. Indeed, we see this reduction readily by decomposing $x$ as $cy+z$, where $c\in[0,1]$ is a scalar, $y\in V$ is a unit vector, and $z\in V^\perp$. By this reduction, we can assume that $V=\mathbb{R}^2$, and $$\qquad x=(1,0),\qquad u=(\cos\alpha,\sin\alpha),\qquad v=(\cos\beta,\sin\beta).$$ So we need to prove that $$\cos^2\beta-\cos^2\alpha\leq|\sin(\alpha-\beta)|.$$ However, this is clear, because the left-hand side equals $\sin(\alpha+\beta)\sin(\alpha-\beta)$.

P.S. I upvoted the question and the other three answers, for fun and world peace.

$\endgroup$
4
  • 1
    $\begingroup$ Beautiful, just the answer I was looking for. I missed the fact that sin(α+β)sin(α−β) is the left hand side. $\endgroup$ Feb 28, 2022 at 8:02
  • 1
    $\begingroup$ I marked all your detailed answers. The key observation is the inequality of Fedor, but I found the prolog of GH from MO cleaner. The answer of OZAWA uses spectral norm which I hoped to avoid, but the proof is much shorter than mine. The answer of Yemon Choi use "simpler" algebra and is probably correct but I gave up in the middle. $\endgroup$ Mar 2, 2022 at 10:05
  • 1
    $\begingroup$ I just realized that I can accept only one answer. This is sad because in math there might be very different answers which are all nice, short and correct just as in this case. I upvoted all and marked GH since his/her answer is the closest to what I used. Maybe I should pass the mark to a different answer every week... $\endgroup$ Mar 2, 2022 at 10:10
  • $\begingroup$ @DanFeldman: I agree. BTW the answers by Yemon Choi and Fedor Petrov deserve more votes. Currently they only have 1 and 2 votes, respectively. $\endgroup$
    – GH from MO
    Mar 2, 2022 at 10:21
5
$\begingroup$

As is observed, since $$|\langle v,x\rangle|^2-|\langle u,x\rangle|^2 = \langle (vv^* -uu^*) x,x\rangle,$$ the value $$\sup_{\|x\|\le1} |\langle v,x\rangle|^2-|\langle u,x\rangle|^2$$ coincides with the spectral norm of the rank two hermitian matrix $vv^* -uu^*$ which has zero trace and two eigenvalues $\pm\lambda$. Since $$2\lambda^2=\operatorname{Tr}((vv^* -uu^*)^2)=2-2|\langle u,v\rangle|^2,$$ one has $\lambda=\sqrt{1-|\langle u,v\rangle|^2}$.

$\endgroup$
2
  • $\begingroup$ Thanks, my proof using eigenvalues is much longer. Can you please remind me the sufficient conditions for a matrix to have $\pm \lambda$ eigenvalues as assumed here? $\endgroup$ Mar 2, 2022 at 9:59
  • $\begingroup$ The matrix has rank two and zero trace. $\endgroup$ Mar 3, 2022 at 0:03
5
$\begingroup$

Denote by $\alpha$, $\beta$, $\gamma$ the angles between $v$ and $x$, $u$ and $x$, $v$ and $u$ respectively (so, $\alpha,\beta,\gamma\in [0,\pi]$). Then $\alpha,\beta,\gamma$ are three planar angles of a trihedral angle (or three sides of a spherical triangle if you prefer) and they satisfy $|\beta-\alpha|\leqslant \gamma\leqslant \alpha+\beta$ and $\alpha+\beta+\gamma\leqslant 2\pi$.

Your inequality reads as $$\cos^2\alpha-\cos^2\beta\leqslant \sin \gamma.$$ We have $$ \cos^2\alpha-\cos^2\beta=\frac{1+\cos 2\alpha}2-\frac{1+\cos 2\beta}2= \frac{\cos 2\alpha-\cos 2\beta}2=\sin(\beta-\alpha)\sin(\alpha+\beta)\\ \leqslant \min(\sin|\beta-\alpha|,|\sin(\alpha+\beta)|)\leqslant \sin \gamma, $$ the last inequality follows from the above inequalities for $\alpha,\beta,\gamma$: $\gamma$ belongs to the segment between $|\beta-\alpha|$ and $\min(\alpha+\beta,2\pi-\alpha-\beta)$, and the minimal value of sine on this segment is realized on one of endpoints.

$\endgroup$
2
  • $\begingroup$ Do we really need the first paragraph? I mean, the inequalities and constraint on the range of the angles are not really necessary for the set of equations below, right? $\endgroup$ May 31, 2022 at 9:32
  • $\begingroup$ No, we explicitly use the bounds from the first paragraph. $\endgroup$ May 31, 2022 at 9:50
2
$\begingroup$

First reduction

After a rotation (or a selection of coordinates) we may WLOG assume that u and v are vectors in ${\bf R}^3$ and that $x=(1,0,0)^\top$. So the desired inequality is equivalent to $$ v_1^2-u_1^2 \leq \sqrt{ 1- (u^\top v)^2 } $$ and it suffices to prove that $$ (u_1^2-v_1^2)^2 \leq 1 - (u^\top v)^2 $$ which is equivalent to $$ 1 \geq (u_1^2-v_1^2)^2 + (u^\top v)^2 $$

Second reduction

Since the RHS increases if we replace each entry of $u$ or $v$ by its modulus, and this change still preserves the constraint that $u$ and $v$ are unit vectors, we may WLOG assume that all entries of these vectors are in $[0,1]$. (This is not essential, but it makes some formulas more symmetrical)

To simplify some formulas, let $a=u_1$ and $b=v_1$.

Then, by the Cauchy–Schwarz inequality (in ${\bf R}^2$) and the assumption that $u$ and $v$ are unit vectors, $$ u_2v_2+u_3v_3 \leq \sqrt{1-a^2}\sqrt{1-b^2} $$ and so $$ (u^\top v)^2 =(ab+u_2v_2+u_3v_3)^2 \leq (ab+\sqrt{1-a^2}\sqrt{1-b^2})^2 $$

Hence it suffices to prove that $$ \displaystyle 1 \geq (a^2-b^2)^2 + (ab+\sqrt{1-a^2}\sqrt{1-b})^2 $$ subject to the constraints $0\leq a\leq 1, 4\leq b\leq 1$.

Final step

Expanding, this is equivalent to $$ 1\geq a^4 -2a^2b^2 + b^4 + a^2b^2 + 2ab\sqrt{1-a^2}\sqrt{1-b^2} + (1-a^2)(1-b^2) $$ which rearranges to $$ a^2(1-a^2) + b^2(1-b^2) \geq 2 ab\sqrt{1-a^2}\sqrt{1-b^2} $$ This now follows by applying the AM-GM inequality to $a\sqrt{1-a^2}$ and $b\sqrt{1-b^2}$.

Afterthought

The use of Cauchy–Schwarz in the second stage could probably be replaced with a more geometric argument, but I don't quite have the right intuition.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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