min max optimization over compact sets

85 Views Asked by At

Suppose I have an infinite set $\mathcal{A}$. All I know about $\mathcal{A}$ is that elements of $\mathcal{A}$ have bounded $l_2$ norms. Suppose I have a constant $b$. Can I solve the following problem. (Even approximate solutions are okay)

$$\min_{x \in \mathrm{R}^d} \sup_{a \in \mathcal{A}}\left|a^Tx-b\right|^2$$

EDIT: The objective of this problem is not to find an exact solution but tell something about $||x^*||_2$. Providing an upper bound for it or similar.

EDIT: So this is a follow up. Let's assume $\mathcal{A}$ is compact. Then we can say the inner supremum is always attained at some $a^*$, so now we are left with $$\min_x (a^{*^T}x - b)^2$$, which can be solved as $x^* = (a^*a^{*^T})^\dagger ab$ where we used the pseudo inverse. Can we now say anything about $||x^*||_2$

1

There are 1 best solutions below

0
On BEST ANSWER

First of all, since you seem to be interested in where the minimum occurs instead of what the minimum is, squaring makes no difference at all. Since squaring is a strictly increasing function on positive real numbers, $$\min_{x \in \Bbb R^d} \sup_{a \in \mathcal A}\left|a^Tx-b\right| = \sqrt{\min_{x \in \Bbb R^d} \sup_{a \in \mathcal A}\left|a^Tx-b\right|^2}$$ and occurs at the same $a, x$.

But $\left|a^Tx-b\right|= \pm a^Tx \pm b$ for the appropriate sign choices. If $a^Tx > b$, then increasing $a^Tx$ will increase $\left|a^Tx-b\right|$. If $a^Tx < b$, then decreasing $a^Tx$ will increase $\left|a^Tx-b\right|$. Thus the suprema of $\left|a^Tx-b\right|$ occur at extrema of $a^Tx$. But extrema of $a^Tx$ are also extrema of $a^T(mx)$ for $0 \ne m \in \Bbb R$. Letting $$a^*(x) = \arg\max_{a \in \overline{\mathcal A}}\left|a^Tx-b\right|,$$ we see that $a^*(mx) = a^*(x)$. However, for any $x$, we can choose $m = \frac b{[a^*(x)]^Tx}$ unless $[a^*(x)]^Tx = 0$. And for that choice of $m$, $$[a^*(mx)]^Tmx - b = 0$$

If $0\ne x \in \mathcal A$, then you can choose a constant $c$ such that $|x^T(cx) - b| > b$, so $|a^*(x)^Tx\ne 0$ for such $x$. This means that there are guaranteed to be $x$ with $a^*(x)^Tx - b = 0$.

If $M = \sup \{\|a\| : a \in \mathcal A\}$, then for any $x \notin \mathcal A^\perp, x_1 = \frac b{[a^*(x)]^Tx}x$ gives the minimum possible value of $0$ for $\sup_{a \in \mathcal A}\left|a^Tx_1-b\right|$. Thus $x^*$ can be $x_1$. But note that if we limit $\|x\| = 1$ while moving it towards $\mathcal A^\perp, [a^*(x)]^Tx$ approaches $0$. Thus $\|x_1\|$ can be arbitrarily large.

Thus if $\operatorname{span}\mathcal A \ne \Bbb R^d$, then $\mathcal A^\perp \ne \{0\}$, and there is no bound on the size $\|x^*\|$. I don't think it is bounded even when $\operatorname{span}\mathcal A = \Bbb R^d$, but I haven't worked that out yet.