4
$\begingroup$

Let $H$ be a infinite dimensional, separable Hilbert space over $\mathbb{C}$

Let $B$ a subset of $H$ such that $B$ is linearly independent and such that exists a homeomorphism $f : [0,1] \to B$ between the unit interval $[0,1]$ and $B$

I would like to know if is it true that:

For every fixed $x \in (0,1)$ $$ \overline{ \operatorname{span} f([0,x]) } \cap \overline{ \operatorname{span} f((x,1]) } = \operatorname{span} f(x) $$

$\endgroup$
12
  • $\begingroup$ When you say homeomorphism, are you using the norm topology on H or the weak topology? $\endgroup$
    – Yemon Choi
    Dec 21, 2017 at 13:43
  • 1
    $\begingroup$ @YemonChoi Without specification, I'm pretty sure it means norm topology. $\endgroup$
    – YCor
    Dec 21, 2017 at 13:47
  • $\begingroup$ @YemonChoi the norm topoly on $H$ $\endgroup$
    – Matey Math
    Dec 21, 2017 at 13:48
  • 1
    $\begingroup$ @hänsel $B$ is not countable $\endgroup$
    – Matey Math
    Dec 21, 2017 at 14:04
  • 2
    $\begingroup$ Interestingly, if the set B was a Hamel basis for H (I know this is not the assumption here), then no such map can exist, because a Hamel basis for H can never be compact. See, e.g., Theorem 3.7 in people.math.ethz.ch/~halorenz/publications/pdf/metal.pdf $\endgroup$ Dec 21, 2017 at 14:57

2 Answers 2

9
$\begingroup$

It's not true.


Edit: thinking twice there's a simpler solution.

On $L^2([0,1])$, define $f(x)(t)=1+t^{1/2x}$ for $x\in\mathopen]0,1]$ and $f(0)=1$. It's free (because the $t\mapsto t^x$, $x\ge 0$, form a free family). It depends continuously on $x$ (check at $x=0$). For $x\le 1/2$ it contains all (restrictions of) polynomials hence spans a dense subspace. So for $x=1/2$ the span intersection condition does not hold.


Previous solution (in the lines of my initial comment).

In $L^2([-1,1])$, define, for $0\le t\le 1/2$, $f(t)=\mathbb{1}_{[-1,2t]}$.

It's clearly algebraically free. The closure of its span is $\mathbf{C}\mathbb{1}_{[-1,0]}\oplus L^2([0,1])$ (indeed it contains, for $s<t$, $f(t/2)-f(s/2)=\mathbb{1}_{[s,t]}$, etc).

Let $K$ be a Cantor subset in $[0,1]$ of measure $1/2$. There exists an increasing (hence measurable) map $g:[0,1]\to K$ such that for every $t$, the measure of $[0,g(t)]\cap K$ is equal to $t/2$. Define, for $t\in [1/2,1]$, $$f(t)=\mathbb{1}_{[-1,1]}+\mathbb{1}_{[0,g(2t-1)]\cap K}.$$ Then $f$ is continuous on $[0,1]$.

Clearly it does not satisfy the span intersection condition at $x=1/2$.

What remains is to show that $(f(t))_{0\le t\le 1}$ is free. Consider a nontrivial linear relation: it can be written as $$\sum_{i=1}^na_if(t_i)=\sum_{j=1}^mb_jf(u_j),\quad\text{ with }0\le t_1<\dots<t_n\le 1/2,\; 1/2<u_1<\dots<u_m\le 1,$$ with all $a_i,b_j$ nonzero and $\max(m,n)\ge 1$. Since $(f(t))_{0\le t\le 1/2}$ is free, $m\ge 1$; write the linear dependence assumption as the equality in $L^2([-1,1])$: $$\sum_{i=1}^na_if(t_i)-\sum_{j=1}^{m-1}b_jf(u_j)=b_mf(u_m).$$ The left-hand term is piecewise constant on $[g(u_{m-1}),1]$ (or $[0,1]$ if $m=1$) while the right-hand term is not: if it were, $\mathbb{1}_L$ would be (almost everywhere) equal to a piecewise constant function, where $L=[g(u_{m-1}),1]\cap K$; since $L$ has positive mesure and its support has empty interior, this is not possible. This thus have a contradiction; accordingly $(f(t))_{0\le t\le 1}$ is indeed free.

$\endgroup$
2
  • $\begingroup$ Sorry for the "duplication" - your answer was posted just as I was finishing up mine. $\endgroup$
    – Yemon Choi
    Dec 21, 2017 at 16:13
  • $\begingroup$ @Ycor thanks for your nice counterexample $\endgroup$
    – Matey Math
    Dec 21, 2017 at 16:20
4
$\begingroup$

In a different direction from Yves Cornulier's suggestion: how about something based on Cauchy kernels as elements of Hardy space? I'm writing this off the top of my head so I might not be taking the most efficient approach.

That is: let ${\mathbb D}$ be the open unit disc in complex plane and let $$ H= \left\{ f : {\mathbb D} \to {\mathbb C} \hbox{ analytic} \colon \sum_{n\geq 0} |a_n|^2 < \infty \right\} $$ where in the above, $f$ has Taylor series $\sum_{n\geq 1} a_nz^n$. Give $H$ the usual norm and inner product, so that (with the obvious abuse of notation) $$ \langle f, g\rangle = \sum_{n\geq 0} a_n \overline{b_n} $$

Take $\displaystyle f_t(z) = \frac{1}{1-tz}$ for $-1/2\leq t \leq 1/2$. Note that $f_t$ has Taylor series $\sum_{n\geq 0} t^n z^n$.

A routine estimate of norms shows that $F: t \mapsto f_t$ is continuous from $[-1/2,1/2]$ to $H$; it is clearly injective. Then $B=F([-1/2,1/2])$ is compact (continuous image of a compact space) and $F:[-1/2,1/2]\to B$ is a homeomorphism (continuous bijection between two compact Hausdorff spaces).

To see that $B$ is linearly independent, note that every $g\in \operatorname{span}(B)$ is a rational function of $z$ with simple poles, all outside $\overline{\mathbb D}$, and the representation of such a function as a linear combination of things in $B$ is unique.

Claim: $F([-1/2,0])$ has dense linear span in $H$.

Proof of claim. Consider the "perp" of this set, i.e. the set of all $h\in H$ such that $\langle h,f_t\rangle = 0$ for all $-1/2\leq t\leq 0$. By the definition of our inner product: if $h(z) = \sum_{n\geq 0} c_n z^n$ then $$ h(t) = \sum_{n\geq 0} c_n t^n=0 \quad\hbox{for all $-1/2\leq t\leq 0$} $$ so that the zero-set of the function $h$ has an accumulation point in the interior of the disc. This forces $h=0$ by standard properties of analytic functions. (Alternatively: just calculate the Taylor coefficients $c_n$ by differentiation to get $c_n=0$ for all $n$.)

Thus $F([-1/2,0])$ has trivial perp, which by basic Hilbert space theory implies it has dense linear span. QED

A similar argument shows that $F([0,1/2])$ also has dense linear span.

So $\overline{\operatorname{span}F([-1/2,0])} \cap \overline{\operatorname{span}F([0,1/2])} = H$, which is quite a lot larger than $\operatorname{span}\{ F(0)\} = {\mathbb C}$.

Remark. The family of functions $f_w(z) = (1- \overline{w}z)^{-1}$, for $w\in {\mathbb D}$ are useful to keep in mind when thinking of how "linearly independent with dense linear span" is a much weaker condition than "forms a (Schauder) basis", when dealing with Banach spaces. They are also the prototypical examples of reproducing kernels for Hilbert spaces of functions.

$\endgroup$
1
  • $\begingroup$ @YCor's answer beat me to it while I was typing, so I think his answer should be the one that's accepted $\endgroup$
    – Yemon Choi
    Dec 21, 2017 at 16:11

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.