# Show that $d$ is direction of closed convex set.

This exercise is from Bazaraa Linear programming book.

And I don't see the point of the exercise since the definition of direction is already given by hypothesis.

Am I missing something or the exercise it's not well stated?

Let $S$ be a closed convex set in $\mathbb R^n$ and let $x\in S$ . Suppose that $d$ is a nonzero vector in $\mathbb R^n$ and that $x +\lambda d\in S$ for all $\lambda\ge 0$. Show that $d$ is a direction of $S.$

• What is the definition of a vector being a "direction of a set" that is given to you? Feb 3, 2018 at 4:29
• @астонвіллаолофмэллбэрг $d\in\mathbb R^n$ nonzero is a direction of $S$ if for all $x\in S$, $x+ld\in S$ for all $l\ge 0$ Feb 3, 2018 at 4:30
• In the question you have, $x$ is one point of $S$. The definition demands that $x + \lambda d \in S$ for all $x \in S$, so you have to use the fact that this holds true for one point, and that $S$ is closed convex, to show that it holds true for all $x \in S$. Feb 3, 2018 at 4:33
• @астонвіллаолофмэллбэрг oh Feb 3, 2018 at 4:39
• @астонвіллаолофмэллбэрг I understand now. What can I do now? Do you think that by contradiction it's a good idea? Feb 3, 2018 at 4:40

You need to show that if $y \in S$ then $y+\lambda d \in S$ for all $\lambda >0$.

Note that for $t \in [0,1]$ we have $(1-t) y + t(x+\lambda d) \in S$.

Pick $\mu >0$, and let $t_n = {1 \over n} \mu$ and let $\lambda = n$ in the above equation (pick $n$ large enough so that $t_n \in (0,1]$).

This gives $(1-t_n) y + t_n x + \mu d \in S$ for all $n$.

Let $n \to \infty$ and use the fact that $S$ is closed to get the desired result.

Excuse the crude diagram:

To illustrate why closure is specified, consider the set $S=\mathbb{R} \times (0,\infty) \cup (0,0)$. Then with $d=(1,0)$ we see that if $x \in \mathbb{R} \times (0,\infty)$ then $x+ \lambda d \in S$ for all $\lambda$ but $(0,0)+ \lambda d \notin S$ for any $\lambda \neq 0$.

• Yes that's exactly what I need to prove, I wonder if a proof by contradiction is a good idea? Feb 3, 2018 at 5:23
• No. ${}{}{}{}{}{}$ Feb 3, 2018 at 5:23
• You really should try to draw a picture to see where the point $(1-t_n) y + t_n x + \mu d \in S$ lies in the context of $y,x,d$. Think of a triangle formed by the points $y,x,x+\lambda d$ and let $\lambda$ get larger and larger. Then the triangle gets more and more like an infinite rectangle formed by the points $x,x+\lambda d, y, y+ \lambda d$. Then closure gives us the desired result. Feb 3, 2018 at 5:53
• @Anneliset.: I added a small diagram that might help. Feb 7, 2018 at 1:37
• Alright, thank you so much 😌 Feb 7, 2018 at 2:06