Let $X$ be a Euclidean space (i.e. a vector space over $\mathbb{R}$ or $\mathbb{C}$ with an inner product, $\textbf{not necessarily complete}$). Let $S$ be a subspace of $X$ with its orthogonal complement $S^\perp$ equal to $\{0\}$. Must we have $\overline{S} = X?$

This is certainly true whenever $\overline{S}$ is complete in $X$, since we can write $X = \overline{S} \oplus S^\perp$. In particular it is true whenever $X$ is a Hilbert space.

But in the general case that $X$ is Euclidean...?

  • $\begingroup$ Your question doesn't make much sense? every euclidean space is a Hilbert space $\endgroup$ – JustDroppedIn May 12 at 11:33
  • $\begingroup$ The definition of a Euclidean space used in my course is that it is an inner product space - not necessarily complete. $\endgroup$ – Baran Karakus May 14 at 8:17
  • $\begingroup$ Yes but you can prove its completeness easily, you only need that $\mathbb{R}$ is complete (and this is the so called "least upper bound axiom") $\endgroup$ – JustDroppedIn May 14 at 21:03
  • 1
    $\begingroup$ This is not true. Take the space C[0, 1] of continuous functions onto the reals from [0, 1], equipped with the L2 norm, which is induced from an inner product. This space is incomplete but is a Euclidean space (i.e. an inner product space). $\endgroup$ – Baran Karakus May 18 at 17:34
  • $\begingroup$ This is not a Euclidean space, since it is not finite dimensional; the definition of a Euclidean space specifically refers to a finite dimensional space. $\endgroup$ – JustDroppedIn May 19 at 12:11

Equip the space $X = C_c^\infty(\mathbb{R})$ of smooth compactly supported test functions with the inner product arising from $L^2(\mathbb{R})$.

Let $S = \{f \in X: \int_0^1 f(s) ds = 0\}$. I claim that $S^\perp = \{0\}$ and that $S$ is not dense.

Firstly, if $g \in S^\perp$ then $\operatorname{supp}(g) \subseteq [0,1]$. Indeed, otherwise there is some interval $[a,b]$ disjoint from $[0,1]$ such that either $g>\varepsilon$ or $g < -\varepsilon$ on $[a,b]$ for some $\varepsilon > 0$. Then, a smooth probability density function $f$ with support in $[a,b]$ lies in $S$ and has either $\int_{\mathbb{R}} fg > \varepsilon > 0$ or $\int_{\mathbb{R}} fg < - \varepsilon <0$.

As a result, if $g \neq 0$ then $g$ is not constant on $[0,1]$ so that there exist $x,y \in [0,1]$ such that $g(x) \neq g(y)$. To see that this cannot happen, let $f$ be a smooth probability density function with support in $[0,1]$ and define $f_x^\lambda(y) = \lambda^{-1}f(\lambda^{-1}(x-y))$. For $\lambda$ sufficiently small $f_x^\lambda$ and $f_y^\lambda$ have disjoint support so that $f_x^\lambda - f_y^\lambda \in S$. Hence $0 = \int_0^1 (f_x^\lambda - f_y^\lambda) g$. However, $f_x^\lambda$ is a mollifier centered at $x$ and so by standard properties of mollification $\int_0^1 f_x^\lambda g \to g(x)$ as $\lambda \to 0$ which implies that $g(x) - g(y) = 0$, a contradiction. Hence $g = 0$ and so $S^\perp = 0$.

Next I show that $S$ is not dense in $X$. Note that if $f_n \in S$ and $f_n \to f$ in $L^2(\mathbb{R})$-norm then $\int_0^1 f = \int_0^1 (f-f_n) = \int_{\mathbb{R}} 1_{[0,1]} (f-f_n) \to 0$ as $n \to \infty$ by Cauchy-Schwarz and so $\int_0^1 f = 0$. Hence, $S$ is a proper closed subset of $X$ with $S^\perp = \{0\}$.

  • $\begingroup$ I don't see that $S$ is closed in $X$. Let $x^{(k)}$ be the vector whose first component is $1$, whose next $k$ components are $-\frac1k$, and whose remaining components are all $0$. Let $y$ be the vector whose first component is $1$ and whose other components are all $0$. All these vectors are in $X$. The difference $y-x^{(k)}$ has $k$ non-zero components, all equal to $-\frac1k$, so its $\ell^2$ norm is $1/\sqrt k$, which approaches $0$ as $k\to\infty$. So $x^{(k)}\to y$. But all of the $x^{(k)}$ are in $S$ and $y$ isn't. $\endgroup$ – Andreas Blass May 23 at 1:35
  • $\begingroup$ @AndreasBlass Shoot, thank you for pointing this out. I was quite tired and apparently very very careless when writing this. Of course, the idea of your example can easily be used to show that that $S$ was dense in the corresponding $X$. I've now updated the answer to contain the example I was really hoping to avoid when writing my previous answer. Hopefully this time I'm not missing something stupid. $\endgroup$ – Rhys Steele May 23 at 8:11

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.