3
$\begingroup$

Let $X$ be a normed space (not necessarily complete) and $\{w_n\}_{n\in\mathbb N}\subset X$ be a linearly independent set. Let $\{x^*_n\}_{n\in\mathbb N}$ be a set of linear functionals on $X$ with the property that $x_i^*(w_j)=\delta_{ij}$. These linear functional are not necessarily bounded.

My question is the following: Is there any linear transformation (an analogue to Gramm-Schmidt procedure) that creates a new linearly independent set $\{u_n\}_{n\in\mathbb N}$, such that $$ \mathrm{span}\,\{u_n\}_{n\in\mathbb N}=\mathrm{span}\,\{w_n\}_{n\in\mathbb N} \subset X, $$ and for which the existence of a bounded dual set $\{u^*_n\}_{n\in\mathbb N}\subset X^*$, i.e., $u_i^*(u_j)=\delta_{ij}$, is guaranteed?

$\endgroup$

1 Answer 1

5
$\begingroup$

Sure. We can ensure that ${\rm span}(u_1, \ldots, u_k) = {\rm span}(w_1, \ldots, w_k)$ and $\|u_k^*\| = 1$ for each $k$. The construction goes by induction on $k$. For $k = 1$ let $u_1 = \frac{1}{\|w_1\|}w_1$ and use Hahn-Banach to find $u_1^* \in X^*$ such that $\|u_1^*\| = u_1^*(u_1) = 1$. Inductively suppose $u_1, \ldots, u_k$ and $u_1^*, \ldots, u_k^*$ are chosen so that ${\rm span}(u_1, \ldots, u_k) = {\rm span}(w_1, \ldots, w_k)$ and so that $\|u_i^*\| = 1$ and $u_i^*(u_j) = \delta_{ij}$ for $1 \leq i,j \leq k$. Then we can find $u_{k+1} \in \bigcap_{i=1}^k {\rm ker}\, u_i^*$ such that ${\rm span}(u_1, \ldots, u_{k+1}) = {\rm span}(w_1, \ldots, w_{k+1})$; an explicit formula is $$u_{k+1} = w_{k+1} - \sum_{i=1}^k u_i^*(w_{k+1})u_i.$$ Define a linear functional $\tilde{u}_{k+1}^*$ on this span such that $\tilde{u}_{k+1}(u_i) = \delta_{i,k+1}$ for $1 \leq i \leq k+1$; by scaling $u_{k+1}$ we can ensure that $\|\tilde{u}_{k+1}^*\| = 1$. Then use Hahn-Banach to extend $\tilde{u}_{k+1}^*$ to $X$ without increasing its norm.

$\endgroup$
6
  • $\begingroup$ Thnx! - Although, I was expecting $u^*_i(u_j)=\delta_{ij}$, for all $i,j$. $\endgroup$
    – smyrlis
    Jan 6, 2014 at 4:21
  • $\begingroup$ Yes, my construction achieves $u_i^*(u_j) = \delta_{ij}$ for all $i, j$. Each $u_{k+1}$ is chosen to lie in the intersection of the kernels of all $u_i^*$ for $i \leq k$, so we get $u_i^*(u_j) = 0$ for $i < j$; beyond that, we only have to worry about $u_{k+1}^*(u_j)$ for $j \leq k+1$, and the construction explicitly handles that. $\endgroup$
    – Nik Weaver
    Jan 6, 2014 at 4:53
  • $\begingroup$ If you scale $u_{k+1}$ to ensure $\left\|\tilde u_{k+1}^\ast\right\|=1$, you lose the property that $\left\|u_{k+1}\right\|=1$, don't you? $\endgroup$
    – 0xbadf00d
    Jan 18, 2021 at 14:33
  • $\begingroup$ Correct. I believe you can achieve $\|u_k\| = \|u_k^*\| = 1$ in finite dimensions --- there's a name for this kind of basis which I forget --- but it's probably not possible in infinite dimensions. $\endgroup$
    – Nik Weaver
    Jan 18, 2021 at 14:53
  • $\begingroup$ @NikWeaver It is called "Auerbach basis". I'd be interested in whether we can obtain the result in finite dimension using your approach. I only know a proof using the determinant. $\endgroup$
    – 0xbadf00d
    Jan 18, 2021 at 18:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.