1
$\begingroup$

In Stephen Boyd & Lieven Vandenberghe's Convex Optimization book we have following set

$$\{x\in \mathbb R^2_+ \mid x_1x_2\geq1\}$$

Now the analytical expression for the supporting hyperplane is given as follows $$\frac{x_1}{t^2}+x_2=\frac{2}{t}.$$ I know that the supporting hyperplane at the boundary point $x_0$ is $a^Tx=a^Tx_0$. (The boundary of the above set is $(t,1/t)$). Further, I know that at $x_0$ the supporting hyperplane is trangent to the set. The tangent to the set at any boundary point is $-\frac{1}{t^2}$. I want to know how we obtain the following analytical expression for the supporting hyperplane $$\frac{x_1}{t^2}+x_2=\frac{2}{t}$$ Any help in this regard will be much appreciated. Thanks in advance.

$\endgroup$
2
$\begingroup$

Consider the set $C$ to be the epigraph of $f$: $$C = \{(x,r) \mid f(x)\leq r\}.$$ An outer normal vector at $(t,f(t))$ is given by $$\big(f'(t),-1\big)$$ and the supporting hyperplane is described by $$0=\langle (f'(t),-1),(x,f(x))-(t,f(t))\rangle.$$ Now if your case, $f(x)=1/x$ and so $f'(x)=-1/x^2$.

Consider the point $(t,1/t)$ fixed, and in the boundary of $C$ for your choice. Then $(x_1,x_2)=(x,f(x))$ is in the graph, where $x=x_1$ and $x_2=1/x_1$. In this case, \begin{align*}0&=\langle (f'(t),-1),(x,f(x))-(t,f(t))\rangle\\ &= \langle (-1/t^2,-1),(x_1,x_2)-(t,1/t)\rangle\\ &=(-1/t^2)(x_1-t)+(-1)(x_2-1/t),\\ &=\frac{-x_1}{t^2}+\frac{1}{t}-x_2+\frac{1}{t}, \end{align*} which turns in what you have.

$\endgroup$
  • 1
    $\begingroup$ Thank you so much for your answer. I think your answer is a bit difficult to understand therefore I try to explain my own reasoning below. Pick a point on the boundary $(t,1/t)$. The tangent to the boundary is $(1,-1/t^2)$. The normal vector associated with the hyperplane must be normal to the tangent vector. Let the normal vector to be $a^T=[a_1 a_2]$ then we must have $<a^T,(1,-1/t^2)>=0$ which results in $a_1=a_2/t^2$. Now if we pick $a_2=1$ then we know that product of $a^T$ with the reference point $(t,1/t)$ is equal to the product of $a^T$ with $(x_1,x_2)$. Therefore ... $\endgroup$ – Frank Moses May 15 '18 at 6:02
  • 1
    $\begingroup$ $$a^T[x_1 x_2]^T=a^T[t 1/t]^T$$ $$x_1/t^2+x_2=2/t$$. Is this explanation wrong? $\endgroup$ – Frank Moses May 15 '18 at 6:08
  • $\begingroup$ Please let me know if my reasoning is wrong. Thank you so much. $\endgroup$ – Frank Moses May 15 '18 at 6:12
  • 2
    $\begingroup$ Looks fine to me! $\endgroup$ – max_zorn May 15 '18 at 6:45

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.