0
$\begingroup$

My intuition is that if a convex function and a concave function are close together, then they both must be close to flat. Is this intuition right, in infinite dimensional spaces?

More precisely, let $K$ be a non-empty convex subset of $[-1, 1]^ω$, let $f : K \to \mathbb{R}$ be a convex function, and let $g : K \to \mathbb{R}$ be a concave function. Suppose that $0 ≤ g - f ≤ 1$. I want to know whether there exists a linear function $λ : \mathbb{R}^ω \to \mathbb{R}$ and $a, b ∈ \mathbb{R}$ such that, for all $x ∈ K$, $$ λ(x) + a \; \leq f \; \leq g \; \leq \; λ(x) + b $$

(This is kind of a follow-up to this question.)

$\endgroup$
0
$\begingroup$

I think I figured it out (with help from a "similar question" in the sidebar). The answer is yes, as long as $K$ has an internal point.

The Separating Hyperplane Theorem says that if $A$ and $B$ are disjoint convex subsets of a vector space, and at least one of them has an internal point, then they can be properly separated by a nonzero linear functional.

Let $$\begin{aligned} A & = \{ (x, y) ∣ x ∈ K, \ y ≥ f(x) + 1 \} \\ B & = \{ (x, y) ∣ x ∈ K, \ y ≤ g(x) - 1 \} \end{aligned}$$ Then $A$ and $B$ are disjoint convex subsets of $\mathbb{R}^ω × \mathbb{R}$. Furthermore, $A$ has an internal point. (If $x$ is an internal point of $K$, then $(x, f(x) + 2)$ is an internal point of $A$.) So the theorem applies.

Let $h : \mathbb{R}^ω × \mathbb{R} \to \mathbb{R}$ be a nonzero linear functional that separates $A$ and $B$: for some $α$, $h(A) < α < h(B)$. We can rewrite this with a little algebra: there is a linear function $λ : \mathbb{R}^ω \to \mathbb{R}$ and a number $c ∈ \mathbb{R}$ such that for $(x, y) ∈ A$ and $(x, y') ∈ B$, $$y' < λ(x) + c < y$$ In particular, for each $x ∈ K$, $(x, f(x) + 1) ∈ A$ and $(x, g(x) - 1) ∈ B$, so we have $$ λ(x) + c - 1 < f(x) ≤ g(x) < λ(x) + c + 1 $$

$\endgroup$

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.