18
$\begingroup$

Given a vector space $V$ equipped with two norms $|\cdot|$ and $||\cdot||$ which are equivalent on a subspace $W$ which is $||\cdot||$-dense in $V$, are the two norms necessarily equivalent?

The statement seems like a relatively straightforward thing to show, but I can't manage it. Having played around with a few proof strategies and not getting anywhere, I'm starting to think that that the statement isn't true, but I'm not yet familiar with many normed spaces and can't think of a counter-example. Any help or hints would be much appreciated.

EDIT: To clear things up, in the book that I took this question from (Linear Analysis by Béla Bollobás) a normed space is defined to be a real or complex vector space, so I think that the intention is for $V$ to be over $\mathbb{R}$ or $\mathbb{C}$.

$\endgroup$
  • $\begingroup$ Those things are tricky very often. I bet that the statement is false but with unnatural counter-examples. Let me think a bit about it $\endgroup$ – Giuseppe Negro Sep 3 '13 at 10:34
  • $\begingroup$ This is an interesting similar question which has a negative answer. $\endgroup$ – Giuseppe Negro Sep 3 '13 at 10:36
  • $\begingroup$ Just to make sure (cf. Mark Bennet's answer), we're dealing with vector spaces over $\mathbb{R}$ or $\mathbb{C}$? $\endgroup$ – Daniel Fischer Sep 3 '13 at 11:43
  • $\begingroup$ @DanielFischer having just checked the book that this question came from (Linear Analysis by Bela Bollobas) I'm fairly sure that this is the intention since a normed space is defined to be a vector space over $\mathbb{R}$ or $\mathbb{C}$. $\endgroup$ – Tom Oldfield Sep 3 '13 at 11:47
16
$\begingroup$

I submit this counterexample which, in my opinion, proves that the statement is false.

In the vein of this MathOverflow post by Gerald Edgar, let $V$ denote the real vector space of all polynomials of one variable and let $$\lVert P\rVert=\max_{x\in[0, 1]} \lvert P(x)\rvert,\qquad \forall P\in V.$$ Moreover, let $$W=\{a_0+a_2x^2+a_4x^4+\dots+a_{2k}x^{2k}\ :\ a_j\in \mathbb{R}\ k\in \mathbb{N}\}. $$ This is a dense subspace of $V$ (cfr. linked post).

Now consider the following linear operator: $$T(x^n)=\begin{cases} x^n & n\ \text{even} \\ nx^n& n\ \text{odd}\end{cases}$$ Its peculiarities are that:

  1. $T\equiv I$ (identity) on the dense subspace $W$;
  2. $T$ is not bounded.
  3. $T$ is $1:1$.

Define $$\lvert P \rvert=\lVert T(P)\rVert.$$ Since $T$ is linear and $1:1$, this defines a norm on $V$. Moreover, this norms agrees with $\lVert\cdot\rVert$ on $W$. Nevertheless, the two norms are not equivalent, because this would imply boundedness of the operator $T$.

$\endgroup$
  • $\begingroup$ This is interesting because it seems to use no choice at all. But of course, this is because the space $V$ has countable algebraic dimension... By the way, something strange happened: I wanted to upvote your answer, but I was told that this is not possible because I cannot upvote... my own post (!) $\endgroup$ – Etienne Sep 3 '13 at 13:36
  • $\begingroup$ Thanks a lot, this is exactly the kind of thing I was looking for! $\endgroup$ – Tom Oldfield Sep 3 '13 at 19:22
10
$\begingroup$

$\mathbb Q(\sqrt 2)$ has two norms which are equivalent on $\mathbb Q$ derived from the two embeddings into $\mathbb R$

$\endgroup$
  • $\begingroup$ Two embeddings? $\endgroup$ – Vishal Gupta Sep 3 '13 at 10:50
  • $\begingroup$ @Vishal: You can send $a+b\sqrt 2 \in \mathbb Q(\sqrt 2)$ to $a+b\sqrt 2\in \mathbb R$ or to $a-b\sqrt 2 \in \mathbb R$ $\endgroup$ – Mark Bennet Sep 3 '13 at 10:55
  • $\begingroup$ Oh! I thought you were referring to two embeddings of $\mathbb{Q}$ into $\mathbb{R}$. $\endgroup$ – Vishal Gupta Sep 3 '13 at 10:58
  • $\begingroup$ I find this example a bit unsatisfactory because $\mathbb{Q}$ is not a real vector space. I think that the OP intended this when speaking of a "vector space" with "norms". $\endgroup$ – Giuseppe Negro Sep 3 '13 at 11:01
  • 6
    $\begingroup$ @GiuseppeNegro Well I gave an example for the question as asked. And it does show that if there is any chance of proving something like this for a real vector space it will have to involve particular properties of the real numbers in some essential way - you can't just prove it from the properties of norms and vector spaces. $\endgroup$ – Mark Bennet Sep 3 '13 at 11:07
7
$\begingroup$

This is not true in general; and in fact this is "never" true.

First, take for example $V=\mathcal C([0,1])$, the space of all continuous functions on $[0,1]$, and let $\Vert\;\Vert=\Vert\,\Vert_\infty$. Let also $W\subset V$ be the space of all polynomial functions on $[0,1]$. Then $W$ is dense in $(V,\Vert\;\Vert)$. Let $E\subset V$ be a linear subspace such that $E\oplus W=V$ (you can find this $E$ with some amount of "choice"). Then $E$ certainly has infinite dimension; so one can find (again with some amount of "choice") a norm $\vert \;\vert_E$ on $E$ which is not equivalent to (the restriction of) $\Vert\;\Vert$ to $E$. Now define the norm $\vert\;\vert$ on $V$ by $\vert e\oplus w\vert=\vert e\vert_E+\Vert w\Vert$ for every $(e,w)\in E\times W$. Then $\vert\;\vert$ are equal on $W$ but not equivalent.

This "construction" works in fact on any (infinite-dimensional) vector space $V$ with any given norm $\Vert\;\Vert$ on $V$. That is, the following holds true: For any infinite-dimensional vector space $V$ and any nor $\Vert\;\Vert$, one can find another norm $\vert \;\vert$ on $V$ such that $\vert\;\vert$ and $\Vert\; \Vert$ coincide on a dense (with respect to $\Vert\;\Vert$) linear subspace $W\subset V$ and yet are not equivalent.

To prove this, all what you need is to find a dense linear subspace $W\subset V$ such that the quotient space $V/W$ is infinite-dimensional; equivalently, for which you can write $V=E\oplus W$ with $\dim E=\infty$.

Now, it is "well known" that you can find a dense set $D\subset W$ (with respect to $\Vert\;\Vert$) which is also linearly independent; se e.g. here: Does there exist a linearly independent and dense subset? Take any sequence $(x_n)\subset D$ such that $\Vert x_n\Vert\to 0$. Then $D\setminus\{ x_n;\; n\in\mathbb N\}$ is again dense in $V$ (wrt $\Vert\;\Vert$). (Indeed, since the set $A=\{ 0\}\cup\{ x_n;\; n\in\mathbb N\}$ is countable, it has empty interior wrt $\Vert\;\Vert$ because nonempty open sets in a normed vector space are uncountable, and moreover $A$ is $\Vert\;\Vert$-closed in $V$; so $V\setminus A$ is dense and open in $V$ wrt $\Vert\;\Vert$, and hence $D\setminus A$ is also $\Vert\;\Vert$-dense in $V$). If you denote by $W$ the linear span of $D\setminus\{ x_n;\; n\in\mathbb N\}$ and by $E$ the linear span of $\{ x_n;\; n\in\mathbb N\}$, then everything is OK.

Note also that the "construction" does not contradict Giuseppe's statement since there is no reason for $\vert\;\vert$ to be complete; but perhaps it could be suitably modified?

$\endgroup$
  • $\begingroup$ This is a nice post. Actually I think that both the example I have written above and the one by Mark Bennett are, roughly speaking, versions of this general construction that you present here. $\endgroup$ – Giuseppe Negro Sep 3 '13 at 16:42
  • $\begingroup$ For your example, I completely agree. For the one by Mark Bennet, I cannot say (I would need more details...) $\endgroup$ – Etienne Sep 3 '13 at 17:11
1
$\begingroup$

WARNING: This answer contains a (probably fatal) error, see comments.

The statement is true with an additional assumption:

$V$ is complete with respect to $\lvert \cdot \rvert$ (recall that $W$ is dense with respect to the other norm $\lVert\cdot\rVert$).

Proof. By assumption, constants $C_1, C_2$ exist such that the following inequalities hold: \begin{equation}\tag{1} \begin{array}{ccc} \lvert y\rvert\le C_1\lVert y\rVert, & \lVert y\rVert\le C_2\lvert y\rvert & \forall y\in W. \end{array} \end{equation} So take $x\in V$. By assumption, we can find a sequence $y_n\in W$ converging to $x$ with respect to $\lVert\cdot\lVert$. Because of the first inequality in (1), $y_n$ is Cauchy with respect to both norms and so, using completeness, we see that a vector $y\in V$ exists such that $$\lvert y_n-y\rvert\to 0.$$ Actually, $x=y$ because $$\lvert x-y\rvert\le \lvert x-y_n\rvert+\lvert y_n-y\rvert\le C_1\lVert x-y_n\rVert + \lvert y_n-y\rvert, $$ and the right hand side of this inequality tends to $0$ as $n\to \infty$. Therefore, by letting $n\to \infty$ in the inequalities $$\lvert y_n\rvert\le C_1\lVert y_n\rVert, \qquad \lVert y_n\rVert\le C_2\lvert y_n\rvert$$ we see that the two norms are equivalent on the whole of $V$.

$\endgroup$
  • 1
    $\begingroup$ I had thought about this approach, but I'm not sure that the result $x=y$ follows directly, since if $x\not\in W$ there is no guarantee that $x-y_n \in W$ and so I don't think we can immediately apply $|x-y_n|\leq C_1 ||x-y_n||$... $\endgroup$ – Tom Oldfield Sep 3 '13 at 11:41
  • $\begingroup$ You are right, I am being circular. Moreover, now that you made me think at it, this is exactly the kind of problem that is mentioned in the linked MathOverflow page. Therefore I guess that my answer is completely useless. I'll wait a bit and then delete it. $\endgroup$ – Giuseppe Negro Sep 3 '13 at 12:31
  • $\begingroup$ No problem, thanks for trying! I've spent a couple of days thinking about it and haven't managed to come up with anything better, although I have done this exact same thing a number of times... $\endgroup$ – Tom Oldfield Sep 3 '13 at 12:33

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.