10
$\begingroup$

Let $(X,\| \cdot \|_{X}),(Y,\| \cdot\|_{Y})$ be Banach spaces, $X$ not reflexive, and let $T:X \rightarrow Y$ be a compact operator that is injective. Further let $|\cdot|$ be another norm on $X$ that is weaker than $\|\cdot\|_{X}$, i.e. there exists a constant $C>0$ such that $|x| \leq C \cdot \|x\|$ for all $x \in X$.
I want to show that:

For every $\varepsilon> 0$ there exists a constant $C_{\varepsilon}$ such that $\|Tx\|_{Y} \leq \varepsilon \cdot \|x\|_{X}+C_{\varepsilon}\cdot|x|$ for all $x \in X$.

I am struggling to prove this. In Compact operator norm estimate I found a proof for the case that $X$ is reflexive and $T$ not necessarily injective, but I cannot adapt this properly. Can anyone help me here?

$\endgroup$

1 Answer 1

10
$\begingroup$

The claim is false. We will construct a counterexample with $X = Y = (C[0,1], \| \cdot \|_{\infty})$ and $|\cdot | = \| \cdot \|_{L_{1}}$. This is the original counterexample constructed for this question and it is inspired by a generalisation of a hint to Exercise 6.13 in Functional Analysis, Sobolev Spaces and Partial Differential Equations by Brezis. To outline the initial construction, we take a countable dense sequence $(t_{k})_{k\in\mathbb{N}}$ in $[0,1]$ and a sequence $(g_{k})_{k\in\mathbb{N}}$ of suitable functions with (almost) disjoint supports. We then define a bounded linear operator $T$ on $C[0,1]$ given by \begin{equation} Tf = \sum_{k=1}^{\infty} 2^{1-k} f(t_{k}) g_{k} . \tag{$\ast$} \end{equation} It is shown the linear operator $T$ is injective and compact, while it is also shown there is no $C > 0$ such that \begin{equation} \|Tf\|_{\infty} \leq \tfrac{1}{2} \|f\|_{\infty} + C \|f\|_{L_{1}} \end{equation} for all $f\in C[0,1]$. After that, it is briefly mentioned how we can take $Y$ as a sequence space instead of $Y = C[0,1]$ and modify the counterexample to that setting. In particular, the case where $Y = c_{0}$ has a proof that is perhaps a bit simpler than the original argument.


Let $(t_{k})_{k\in\mathbb{N}}$ be a sequence that is dense in $[0,1]$. Also let $(g_{k})_{k\in\mathbb{N}}$ be a sequence in $C[0,1]$ such that $\|g_{k}\|_{\infty} = 1$ for all $k\in\mathbb{N}$ and \begin{equation} \{ t\in [0,1] : g_{j}(t) \neq 0 \} \cap \{ t\in [0,1] : g_{k}(t) \neq 0 \} = \emptyset \tag{1} \end{equation} for all $j,k\in \mathbb{N}$ with $j \neq k$. An example of such a sequence is $(g_{k})_{k\in\mathbb{N}}$, where $g_{k} \colon [0,1] \to \mathbb{R}$ is defined by $g_{k}(t) := \max\{1 - 2^{k+1}|t - \tfrac{3}{2^{k+1}}| , 0\}$ for each $k\in\mathbb{N}$. There is a helpful comment to this answer by Just a user that points out some of the geometric properties of these $g_{k}$ functions that may help the reader follow along with the proof. See here for the comment.

Define $T\colon C[0,1] \to C[0,1]$ by \begin{equation} Tf := \sum_{k=1}^{\infty} 2^{1-k} f(t_{k}) g_{k}. \end{equation} It is straightforward to check that $T$ is a bounded linear operator on $(C[0,1], \| \cdot \|_{\infty})$. It can in fact be shown directly using $(1)$ that $\|T\| = 1$, but we will not need to make use of this. We will show below that $T$ is injective and compact.

We first show that $T$ is injective. Let $f\in C[0,1]$ such that $Tf = 0$. Let $k\in \mathbb{N}$. Since $\|g_{k}\|_{\infty} = 1 > 0$ there is some $t\in [0,1]$ such that $g_{k}(t) \neq 0$. By $(1)$ we have $g_{j}(t) = 0$ for all $j\in\mathbb{N}\setminus\{k\}$. It follows that \begin{equation} 2^{1-k} f(t_{k}) g_{k}(t) = (Tf)(t) = 0. \end{equation} Since $2^{1-k} g_{k}(t) \neq 0$ we have $f(t_{k}) = 0$. As $k\in\mathbb{N}$ was arbitrary we see that $f(t_{k}) = 0$ for all $k\in\mathbb{N}$. But since $\{t_{k} : k\in\mathbb{N}\}$ is dense in $[0,1]$, it follows from the continuity of $f$ that $f(t) = 0$ for all $t\in [0,1]$. Hence $f = 0$. This shows that $T$ is injective.

We now show $T$ is compact. Let $n\in\mathbb{N}$. Define $T_{n} \colon C[0,1] \to C[0,1]$ by \begin{equation} T_{n}f := \sum_{k=1}^{n} 2^{1-k} f(t_{k}) g_{k}. \end{equation} Then $T_{n}$ is a finite rank linear operator on $C[0,1]$. Let $f\in C[0,1]$. Using $(1)$, we have \begin{equation} |T_{n}f - Tf| ={} \Big| \sum_{k=n+1}^{\infty} 2^{1-k} f(t_{k}) g_{k} \Big| \leq \sup \{2^{1-k} |f(t_{k})| : k\in \mathbb{N} \text{ and } k > n \} \leq 2^{-n} \|f\|_{\infty} . \tag{2} \end{equation} It follows from $(2)$ that $\|T_{n} - T\| \leq 2^{-n}$. This shows that $T$ is in the norm closure of the set of finite rank linear operators and is consequently compact.

We now show there does not exist $C > 0$ such that \begin{equation} \| Tf \|_{\infty} \leq \tfrac{1}{2} \|f\|_{\infty} + C \|f\|_{L_{1}} \tag{3} \end{equation} for all $f\in C[0,1]$, where $\|f\|_{L_{1}} = \int_{0}^{1} |f|$ for each $f\in C[0,1]$. Since $\|f\|_{L_{1}} \leq \|f\|_{\infty}$ for every $f\in C[0,1]$, showing that $(3)$ fails for every $C > 0$ would provide the desired counterexample. Let $C > 0$. Since $\|g_{0}\|_{\infty} = 1$ there is $s_{0}\in [0,1]$ such that $|g_{0}(s_{0})| = 1$. Take $f\in C[0,1]$ such that $\|f\|_{\infty} = 1$, $f(s_{0}) = 1$ and $\int_{0}^{1} |f| < \tfrac{1}{2C}$. An example of such a function is $f\colon [0,1] \to \mathbb{R}$ defined by $f(t) := \max\{1 - 4C|t - s_{0}|, 0\}$. We have \begin{equation} \tfrac{1}{2} \|f\|_{\infty} + C \|f\|_{L_{1}} < \tfrac{1}{2} + C \cdot \tfrac{1}{2C} = 1 \tag{4} \end{equation} while \begin{equation} |(Tf)(s_{0})| = |f(s_{0}) g_{0}(s_{0})| = 1 . \end{equation} Consequently, it follows that $\|Tf\|_{\infty} \geq 1$ and hence, using $(4)$, that \begin{equation} \|Tf\|_{\infty} \geq 1 > \tfrac{1}{2} \|f\|_{\infty} + C \|f\|_{L_{1}} . \end{equation} As $C > 0$ was arbitrary, we have a counterexample.


Here are some additional comments. Upon further inspection, it came to my attention that instead of taking the codomain of $T$ as $C[0,1]$ as done in $(\ast )$, we could have instead modified the definition of $T$ and instead mapped into a sequence space. Perhaps the simplest example of this is the following. Take a dense sequence $(t_{k})_{k\in\mathbb{N}}$ in $[0,1]$ and define $S\colon C[0,1] \to c_{0}$ by \begin{equation} Sf := (2^{1-k} f(t_{k}))_{k\in\mathbb{N}} . \end{equation} The advantage here is that we no longer need to find a suitable sequence of functions $(g_{k})_{k\in\mathbb{N}}$ as in the definition of $(\ast )$. As in the above proof, the map $S$ is an injective compact operator and there is no $C > 0$ such that \begin{equation} \|Sf\|_{c_{0}} \leq \tfrac{1}{2} \|f\|_{\infty} + C \|f\|_{L_{1}} \end{equation} for all $f\in C[0,1]$. We can also take $\ell_{\infty}$ in place of $c_{0}$ in the above definition of $S$ and all the details are exactly the same as in the case with $c_{0}$.

Another important example of this generalisation is to take $p\in [1, \infty )$ as well as a dense sequence $(t_{k})_{k\in\mathbb{N}}$ in $[0,1]$ and define $S\colon C[0,1] \to \ell_{p}$ by \begin{equation} Sf := \big( 2^{-\tfrac{k}{p} } f(t_{k}) \big)_{k\in\mathbb{N}} . \end{equation} As done in the other cases, it is shown similarly that $S$ is an injective compact operator. This time, however, it is simpler to show that there is no $C > 0$ such that \begin{equation} \|Sf\|_{\ell_{p}} \leq \tfrac{1}{4} \|f\|_{\infty} + C \|f\|_{L_{1}} \tag{5} \end{equation} for all $f\in C[0,1]$. We show $(5)$ in detail. Let $C > 0$. Take $f\in C[0,1]$ such that $f(t_{1}) = 1$, $\|f\|_{\infty} = 1$ and $\int_{0}^{1} |f| < \tfrac{1}{4C}$. We have \begin{equation} \tfrac{1}{4} \|f\|_{\infty} + C \|f\|_{L_{1}} < \tfrac{1}{4} + C \cdot \tfrac{1}{4C} = \tfrac{1}{2} \tag{6} \end{equation} and \begin{equation} \| Sf \|_{\ell_{p}}^{p} = \sum_{k=1}^{\infty} |2^{-\tfrac{k}{p} } f(t_{k}) |^{p} \geq |2^{-\tfrac{1}{p} } f(t_{1})|^{p} = 2^{-1} \geq (2^{-1})^{p} , \tag{7} \end{equation} so it follows by combining $(6)$ and $(7)$ that \begin{equation} \| Sf \|_{\ell_{p}} \geq \tfrac{1}{2} > \tfrac{1}{4} \|f\|_{\infty} + C \|f\|_{L_{1}} . \end{equation} This shows there is no $C > 0$ such that $(5)$ holds for all $f\in C[0,1]$. Note that by taking $p\in (1, \infty )$ we can find a counterexample where the space $Y$ in the original question is reflexive, and by taking $p = 2$ we can even take the space $Y$ in the original question as a Hilbert space.

Finally, the reader is encouraged to take a look at this related result. Note that the linked result is not applicable to this question because we cannot typically find a suitable Banach space $(Y, \|\cdot\|_{Y})$ to apply the result.

$\endgroup$
4
  • 1
    $\begingroup$ Very nice! One small note: I think the indices $k \in \mathbb{N}_{0}$ should start at 2, but that doesn't matter for the argument. Thx again. $\endgroup$
    – benny
    Commented 17 hours ago
  • 1
    $\begingroup$ This is very nice! $\endgroup$
    – jr01
    Commented 17 hours ago
  • $\begingroup$ Thanks for the comments. There was a problem with the examples of the $g_{k}$ functions that has now been fixed. Let me know if anything else does not look right. $\endgroup$ Commented 13 hours ago
  • 1
    $\begingroup$ Good answer! Learned a lot. Just to point out the obvious that geometrically the (non-zero part of) the graph of $f(t):=\max\{1-a|t-b|, 0\}$ for sufficiently large $a$ forms the two legs of the isosceles triangle with base $[b-\frac{1}{a}, b+\frac{1}{a}]$, and apex $(b, 1)$, which makes it easier to follow the argument. $\endgroup$ Commented 10 hours ago

You must log in to answer this question.

Start asking to get answers

Find the answer to your question by asking.

Ask question

Explore related questions

See similar questions with these tags.