Prove a certain property of linear functionals, using the Hahn-Banach-Separation theorems

181 Views Asked by At

I've been trying to solve this question, with no luck so far:

Let $X$ be a real linear space, and $\{\|\cdot \|_i\}_{i=1}^{n}$ family of norms on $X$. Let $f$ be a linear functional on $X$ such that for every $x\in X$: $f(x)\leq \max \|x\|_i$.

Prove that there are $t_1,...,t_n\geq0$ such that $\sum t_i=1$ and for every $x\in X$: $f(x)\leq\sum t_i \|x\|_i$.

I've been told that the best way to look at this questions is by separating $A=\mathop{\rm conv}\{(\|x\|_1-f(x),\ldots,\|x\|_n-f(x)):x\in X\}\subset \mathbb R^n$ with some other set.

It didn't do me much good to try this my own. My observations are that $A$ is obviously convex, and every point in $\{(\|x\|_1-f(x),\ldots,\|x|\|_n-f(x)):x\in X\}\$$ has at least one non-negative coordinate. $A$ isn't necessarily closed since $f$ isn't necessarily bounded with one specific norm (therefore not continuous), nor is opened. Don't know what separation theorem would do, and more than that - even if I could separate, all I'd get is $\varphi(A)\geq \gamma \geq \varphi(B)$ which would give me information about $\varphi(a)$ and not on $f$.

I thought it might be possible to prove that every point in a has at least one non-negative coordinate (is this even true?) and seperate from $\{(b_1,...,b_n):\forall i\:: b_i<0\}$

I'd love some tips/hints/guide/any thoughts actually.

Thanks a lot!

1

There are 1 best solutions below

6
On BEST ANSWER

Feel free to read chunks of this answer until it gives you some new idea, then try to proceed on your own.

Notice that the thesis is the existence of $t_1,\dots,t_n\ge 0$ with $\sum t_i=1$ and such that $\sum t_i(\|x\|_i-f(x))\ge 0$.
Your idea of choosing $B:=\{(b_1,\dots,b_n):\forall i\ b_i<0\}$ is good: now $A$ and $B$ are two convex sets and if you can find a nontrivial functional $\phi$ which separates them (which means $\phi(a)\ge\phi(b)$ for any $a\in A$, $b\in B$) then you are done.
In fact, any nontrivial functional $\phi:\mathbb{R}^n\to\mathbb{R}$ has the form $\phi(x)=\sum_{i=1}^n t_i x_i$ (where not all the $t_i$'s are $0$), so in your case, since $0\in\overline{B}$, you get $\sum t_i(\|x\|_i-f(x))=\phi(a)\ge\phi(0)=0$ (where we have chosen $a:=(\|x\|_i-f(x))_i\in A$).
Moreover, for any $i$, the vector $b:=(0,\dots,0,-1,0,\dots,0)$ having $-1$ as the $i$-th coordinate lies in $\overline{B}$, while $0\in A$ (just choose $x=0$!); thus $-t_i=\phi(b)\le\phi(0)=0$ and you obtain $t_i\ge 0$. Finally by rescaling your functional by a positive constant you may suppose $\sum t_i=1$ and the property $\forall x\ \sum t_i(\|x\|_i-f(x))\ge 0$ is obviously still true (it is invariant under rescaling), so you are finished!

We are left to check the existence of a nontrivial functional $\phi$: here Hahn-Banach is your friend. You can verify the so-called first geometric form (separation of an open convex set from a convex set): $B$ is open and convex, while $A$ is convex. If we show that $A\cap B=\emptyset$, then all the hypotheses of this form of Hahn-Banach are satisfied and we are done.
Notice that we have not yet used the fact that $\|\cdot\|_i$ are norms. Suppose $y\in A\cap B$: as $y\in A$ there exist $\lambda_1,\dots,\lambda_m\ge 0$ such that $\sum_{j=1}^m \lambda_j=1$ and some $x_1,\dots,x_m\in X$ such that $y=\sum_{j=1}^m\lambda_j(\|x_j\|_i-f(x_j))_{i=1}^n$.
(By $(\|x_j\|_i-f(x_j))_{i=1}^n$ I mean the vector $(\|x_j\|_1-f(x_j),\dots,\|x_j\|_n-f(x_j))$ ).
So componentwise we have $$\forall i\ \sum_{j=1}^m\lambda_j(\|x_j\|_i-f(x_j))=y_i<0$$ (since $y\in B$). So $$\forall i\ f(\sum \lambda_j x_j)=\sum \lambda_j f(x_j)>\sum\lambda_j\|x_j\|_i\ge\|\sum \lambda_j x_j\|_i$$ (the last inequality comes from the properties of a norm) and if we put $x:=\sum \lambda_j x_j$ this chain of inequalities gives $\forall i\ f(x)>\|x\|_i$, which contradicts the hypothesis $f(x)\le\max\|x\|_i$. This contradiction proves that $A\cap B=\emptyset$.