2
$\begingroup$

Let $\lVert \cdot \rVert_*:\mathbb{R}^2 \to \mathbb{R}, (x,y) \rightarrow \sqrt{x^2+2xy+3y^2}$ be a norm.

How can I find to constants $k,K \in \mathbb{R}^{>0}$ so that the following equivalence is given:$$k\lVert (x,y) \rVert_2 \leq \Vert (x,y) \rVert_* \leq K\Vert (x,y) \rVert_2.$$

My idea:

$$ k\lVert (x,y) \rVert_2 \leq \Vert (x,y) \rVert_* \leq K\Vert (x,y) \rVert_2 \\ \Leftrightarrow k\sqrt{|x|^2+|y|^2} \leq \sqrt{x^2+2xy+3y^2} \leq K\sqrt{|x|^2+|y|^2} $$

Now how can I find $k,K$? The problem is these constants must be minimal. Can someone give me a tip?

Thx

$\endgroup$
  • 1
    $\begingroup$ Hint. The unit disk for the $*$ norm is an ellipse. You want to find inscribed and circumscribed circles. Can you draw a picture? $\endgroup$ – Ethan Bolker Oct 16 '17 at 15:12
1
$\begingroup$

Here's a version that's more explicitly geometric, but whose underlying mathematics resemble Roberto's matrix diagonalization.

Rewrite the norm in rotated coordinates $(x', y')$, where $x = x' \cos \theta - y' \sin \theta$ and $y = x' \sin \theta + y' \cos \theta$. We'll choose $\theta$ at our convenience—specifically, we'll choose it such that the coordinate axes of the rotated coordinate system align with the axes of the unit ellipse $||(x, y)||_* = 1$, thus making the $xy$ terms in the norm vanish. In this case, \begin{align*}|| (x, y)||_*^2 =& x'^2 \cos^2 \theta \tag{$x^2$} - 2x' y' \sin \theta \cos \theta + y'^2 \sin^2 \theta \\ &+ 2 (x'^2 - y'^2) \sin \theta \cos \theta + 2x' y' (\cos^2 \theta - \sin^2 \theta) \tag{$2xy$} \\ &+ 3x^2 \sin^2 \theta + 6x' y' \sin \theta \cos \theta + 3y^2 \cos^2 \theta \tag{$3y^2$} \end{align*}

The coefficient of $2x' y'$ is $4 \sin \theta \cos \theta + 2 \cos^2 \theta - 2 \sin^2 \theta = 2 \sin (2\theta) + 2 \cos (2\theta)$. So we can make this term disappear by choosing $\theta = -\pi/8$, for which $$\begin{align*} \sin^2 \theta &= \frac{2 - \sqrt{2}}{4} \\ \cos^2 \theta &= \frac{2 + \sqrt{2}}{4} \\ \sin \theta \cos \theta &= -\frac{\sqrt{2}}{4}\end{align*}$$ We thus have $$ \begin{align*} || (x', y')||_*^2 &= x'^2 (\cos^2 \theta + 2 \sin \theta \cos \theta + 3 \sin^2 \theta) + y'^2 (\sin^2 \theta - 2 \sin \theta \cos \theta + 3 \cos^2 \theta) \\ &= (2 - \sqrt{2}) x'^2 + (2 + \sqrt{2}) y'^2\\ \end{align*}$$

Meanwhile, the ordinary Euclidean norm, invariant under rotation, is $$|| (x', y')||_2^2 = x'^2 + y'^2.$$ The path from here should be evident.

$\endgroup$
1
$\begingroup$

(How to make simple things difficult...)

By Young's inequality, for every $a>0$ you have that $$ 2|xy| = 2(a|x|)(|y|/a) \leq a^2 x^2 + \frac{y^2}{a^2} $$ hence, for every $a, b > 0$, $$ - a^2 x^2 - \frac{y^2}{a^2} \leq 2xy \leq b^2 x^2 + \frac{y^2}{b^2}\,. $$ These inequalities give $$ (1-a^2) x^2 + \left(3 - \frac{1}{a^2}\right) y^2 \leq x^2 + 2xy + 3y^2 \leq (1+b^2) x^2 + \left(3+\frac{1}{b^2}\right)y^2. $$ In order to get multiples of the Euclidean norm, you can choose $a,b> 0$ such that $$ 1-a^2 = 3 - \frac{1}{a^2}\,, \qquad 1+b^2 = 3 + \frac{1}{b^2}\,, $$ hence $$ a^2 = \sqrt{2}-1, \qquad b^2 = \sqrt{2}+1. $$ Substituting we finally get $$ (2-\sqrt{2})(x^2+y^2) \leq x^2 + 2xy + 3y^2 \leq (2+\sqrt{2}) (x^2+y^2). $$

$\endgroup$
0
$\begingroup$

Note that $$x^2 + 2xy + 3y^2 = \begin{pmatrix} x \\ y \end{pmatrix}^T \begin{pmatrix} 1 & 1 \\ 1 & 3 \end{pmatrix} \begin{pmatrix} x \\ y \end{pmatrix}.$$ The eigenvalues of the matrix have sum 4 and product 2, so they are both stricly positive; let us call them $\lambda_1$ and $\lambda_2$. Since the matrix is symmetric, there exists $Q$ orthogonal such that: $$\begin{pmatrix} 1 & 1 \\ 1 & 3 \end{pmatrix} = Q^T \begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2 \end{pmatrix} Q.$$ Now let us note: $$\begin{pmatrix} q_1 \\ q_2 \end{pmatrix}= Q \begin{pmatrix} x \\ y \end{pmatrix}.$$ Then you can see that $\| \mathbf q \|_2 = \| \mathbf x\|_2$, and $$x^2 + 2xy + 3y^2 = \begin{pmatrix} q_1 \\ q_2 \end{pmatrix}^T \begin{pmatrix} \lambda_1 & 0 \\ 0 & \lambda_2 \end{pmatrix} \begin{pmatrix} q_1 \\ q_2 \end{pmatrix} = \lambda_1 q_1^2 + \lambda_2 q_2^2,$$ from where it is easy to conclude.

Another way to proceed, adapted specifically to the example, is to notice that, on the one hand, $$ x^2 + 2xy + 3y^2 \leq 2 x^2 + 4y^2 \leq 4(x^2 + y^2),$$ and on the other: $$ x^2 + 2xy + 3y^2 = \left(\sqrt{a}x + \frac{y}{\sqrt{a}}\right)^2 + (1-a)x^2 + \left(3-\frac{1}{a}\right)y^2 \geq \min\left(1-a,3-\frac{1}{a}\right) (x^2 + y^2) \quad \forall a \in \mathbb R_+.$$ Any $a$ such that $1/3 < a < 1$ will guarantee that the $\min$ is strictly positive.

$\endgroup$

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.