7
$\begingroup$

Note: Whenever I say "dual space" in this question, I mean the algebraic dual space.

Consider an infinite-dimensional vector space $V$. Since it is infinite-dimensional, its double-dual $V^{**}$ is strictly larger than $V$. On the other hand, there's a natural injection $\iota:V\to V^{**}$, so by identifying $v$ with $\iota(v)$, we get $V\subsetneq V^{**}$.

Of course $V^{**}$ again is an infinite dimensional vector space, so taking its double-dual gives again a larger vector space. So if we define $$V_0=V, V_{n+1}=V_n^{**}$$ and do the identification as above, we get an infinite sequence of vector spaces $$V_0\subsetneq V_1\subsetneq V_2\subsetneq\dots$$ We can now define the limit space as $$V_\infty = \bigcup_{n=0}^\infty V_n$$ It is not hard to verify that this is also a vector space: For every finite set of vectors in $V_\infty$ you find a $V_n$ which contains all of them, and thus also their linear combination, which therefore also is in $V_\infty$.

Now obviously also for the dual spaces we have that $V_{n+1}^*$ is the double dual of $V_n^*$, therefore we can also define the limit space of the duals: $$W_\infty = \bigcup_{n=0}^\infty V_n^*$$

Now my question is: Is $W_\infty=V_\infty^*$?

It seems intuitive that it should be, but on the other hand, each $V_n$ (except for $V_0$) is also the dual of $V_{n-1}^*$, and thus the same intuition gives that $V_\infty=W_\infty^*$. But both together cannot hold, since otherwise the double-dual of $V_\infty$ would be again $V_\infty$, which cannot be since $V_\infty$ is infinite-dimensional. So clearly the intuition doesn't help here. Indeed, that argument seems to indicate that both are not the same; however I can't see how to make that rigorous.

Now clearly $W_\infty\subseteq V_\infty^*$, so assuming they are not equal, is it possible to explicitly construct an element of $V_\infty^*\setminus W_\infty$?

$\endgroup$
  • 1
    $\begingroup$ The double dual is not always strictly larger, even in infinite dimensions. There are a wealth of spaces, called reflexive spaces, where they are naturally isomorphic. $L^p$ for $1<p<\infty$ are classic examples. In Hilbert spaces the situation is even cleaner, since $H,H^*,H^{**}$ are all naturally isomorphic. There are also irreflexive spaces where they are not naturally isomorphic but they are "non-naturally" isomorphic (i.e. they are linearly isometric but not under the canonical embedding). $\endgroup$ – Ian Oct 4 '14 at 22:17
  • $\begingroup$ (Cont.) As I recall the process you've described will actually terminate for this reason after a small number of steps. Unfortunately I can't find a reference. $\endgroup$ – Ian Oct 4 '14 at 22:17
  • $\begingroup$ Oh, sorry, I think you are talking about the algebraic dual space in this discussion, rather than the continuous dual space of a topological vector space. It might be helpful to readers to make that explicit. $\endgroup$ – Ian Oct 4 '14 at 22:24
  • $\begingroup$ As for continuous duals, I think it's quite the opposite of what you are stating, Ian: either $X=X^{**}$ or the chain will go on forever. Every space is a closed subspace of its own continuous double dual and a closed subspace of a reflexive space is reflexive... $\endgroup$ – wisefool Oct 4 '14 at 22:26
  • $\begingroup$ @Ian: Thanks, I indeed meant the algebraic dual space. I'll add a note to this effect. Of course, the situation with continuous duals might also be interesting. $\endgroup$ – celtschk Oct 4 '14 at 22:33
1
$\begingroup$

I believe this is a possible answer for your question. Please check it.

First of all, since we want $V_i^*\subset V_{k}^*$, $i\leq k$, we must understand how to evaluate $f\in V_i^*$ on elements of $V_k$.

Let $k=i+1$. Since $V_{i+1}=(V_i^*)^*$ then $f(x):=x(f)$, for every $x\in V_{i+1}$. Let us suppose that we understand how to evaluate $f\in V_i^*$ on elements of $V_k$ $($thus we may assume that $f\in V_k^*)$ then $f(x):=x(f)$ for every $x\in V_{k+1}$.

Remark 1: Thus, if $f\in V_i^*$, $x\in V_k$ and $k>i$ then $f(x)=x(f)$. $($Am I right here?$)$

Assuming this is right, let us construct two sequences $(x_{2n+2})_{n\in\mathbb{N}}$ and $(y_{2n+1})_{n\in\mathbb{N}}$ .

Since $V_{2n+1}^*\setminus V_{2n}^*\neq \emptyset$ then exists $x_{2n+2}\in (V_{2n+1}^*)^*=V_{2n+2}$ such that

  1. $x_{2n+2}|_{V_{2n}^*}\equiv 0$ and
  2. for some $y_{2n+1}\in V_{2n+1}^*\setminus V_{2n}^*$ we have $x_{2n+2}(y_{2n+1})=1$.

Let us prove that $(x_{2n+2})_{n\in\mathbb{N}}$ is linear independent.

Assume $g=\sum_{i=1}^ta_ix_{2n_i+2}=0$ and $n_1<n_2<\ldots<n_t$.

Now, $0=g(y_{2n_1+1})=\sum_{i=1}^ta_ix_{2n_i+2}(y_{2n_1+1})$.

If $n_i>n_1$ then $y_{2n_1+1}\in V_{2n_1+1}^*\subset V_{2n_i}^*$ and $x_{2n_i+2}(y_{2n_1+1})=0$, by construction.

Notice that $x_{2n_1+2}(y_{2n_1+1})=1$, by construction. Thus, $a_1=0$. Analogously, we obtain $a_2=\ldots=a_t=0$. Thus, $(x_{2n+2})_{n\in\mathbb{N}}$ is linear independent.

Now, let $f\in W_\infty$, thus $f\in V_{2n}^*$ for some $n$.

Thus, if $k\geq n$ then $x_{2k+2}\in V_{2k+2}$ and $2k+2>2n$. Therefore, by remark 1, $f(x_{2k+2})=x_{2k+2}(f)$.

Next, by construction, $x_{2k+2}(f)=0$, since $V_{2n}^*\subset V_{2k}^*$. Therefore if $k\geq n$ then $f(x_{2k+2})=0$.

Remark 2:Thus, for every $f\in W_\infty$, exists $n$ such that for $k\geq n$, $f(x_{2k+2})=0$.

Now since $(x_{2n+2})_{n\in\mathbb{N}}$ is linear independent, we can extend this set to a basis of $V_{\infty}$. Let $g\in V_{\infty}^*$ be such that $g(x_{2n+2})=1$ for every $n$. Thus, $g\in V_{\infty}^*\setminus W_{\infty}$, by remark 2.

$\endgroup$
  • $\begingroup$ Very nice. As far as I can tell, your argumentation is correct. $\endgroup$ – celtschk Oct 5 '14 at 20:46

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.