14

Can someone give me an example of noncomplete inner product space $H$, its closed linear subspace of $H_0$ and element $x\in H$ such that there is no orthogonal projection of $x$ on $H_0$. In other words I need to construct a counterexample to theorem about orthogonal projection when inner product space is not complete.

Norbert
  • 58,398

2 Answers2

12

As I did not succeed to come with an example by myself, I tried searching google books for "not complete" "orthogonal projection".

I found a counterexample in the book Linear Operator Theory in Engineering and Science By Arch W. Naylor, George R. Sell, pages 289 and 302. (BTW this was the first ever book on functional analysis I was reading as a student. Now I wish didn't stop somewhere halfway back then - it seems that it contains a lot of interesting things and I remember the style of the book as very readable.)

I will reproduce a brief version of their example below, but it is probably better if you have a look in the book. I've kept the notation from the book, so $X$ is the noncomplete space (your $H$) and $M$ is the subspace (your $H_0$).


Suppose that $H$ is any Hilbert space and $X$ is a dense subspace of $H$ that is not closed. Let $z\in H\setminus X$. We define $$M=\{y\in X; \langle y,z \rangle=0\}.$$ This set $M$ is a closed subspace of $X$.

Now choose $x_0\in X\setminus M$. We will show that $x_0$ has no orthogonal projection on $M$.

Suppose that $y_0\in M$ is the orthogonal projection of $x_0$. This means $y_0-x_0$ is orthogonal to $M$ and thus it is a scalar multiple of $z$. This would imply $z\in X$, a contradiction.


So now we only ask whether it is possible to find $H$, $M$, $z$, $x_0$ as above.

The choice from the book I mentioned is the following: $H=\ell_2$,
$X=$ the set of all sequences that have only finitely many nonzero terms
$z=(\frac1{k^2})_{k=0}^\infty$
$x_0$ can be chosen arbitrary such that $\langle x_0,z \rangle \ne 0$.

  • Oh, you've posted this while I was writing... I wasn't notified. The example I produced is an explicit version of yours. – t.b. Nov 16 '11 at 04:06
  • Well, +1 anyway, it's much simpler in the abstract. – t.b. Nov 16 '11 at 04:12
  • Thanks, this is exactly what I needed. – Norbert Nov 16 '11 at 09:05
  • 'This set $M$ is a closed subspace of $H$.' Here you meant to say closed subspace of $X$, right? Also, could you please explain why, in the last paragraph, '$y_0-x_0$ is orthogonal to $M$' would imply it is a scalar multiple of $z$? – lychtalent Jun 30 '22 at 10:30
  • @lychtalent I have edited (so the post now says it is a closed subspace of $X$.) About orthogonality to $M$: W.l.o.g., let us assume $|z|=1$. Let $v\in X$ be any vector such that $v\perp M$. Let $c=\langle v,z\rangle$. Then for $v'=v-cz$ we have both $v'\in M$ and $v'\perp M$. Now $v'\perp v'$ implies $v'=0$. So we have $v-cz=0$ and $v=cz$. – Martin Sleziak Jun 30 '22 at 11:13
  • Thanks for the quick respond! However, I don't get why $v'=v-cz$ is an element of $M$. If I introduce the set $M'={y\in H;\langle y,z \rangle=0}$, then I see that $v'\in M'$, and clearly $M\subset M'$. I suspect that $X$ being a dense subset of $H$ would be used, but have not figured out how. – lychtalent Jun 30 '22 at 11:27
  • 1
    @lychtalent You're right again - and this shows that I wasn't careful enough when looking into this problem. However, your comments suggests that this could be finished if we know (or can prove) that for any vector $x$ and any subspace $M$ we have $x\perp M$ $\implies$ $x\perp\overline M$. I would suggest that we might continue this in chat - at least if you know how to use chat. – Martin Sleziak Jun 30 '22 at 11:55
10

Let $H$ be the inner product space consisting of $\ell^2$-sequences with finite support, let $\lambda = 2^{-1/2}$ and put $$ z = \sum_{n=1}^\infty \;\lambda^n \,e_n \in \ell^2 \smallsetminus H $$ Then $\langle z, z \rangle = \sum_{n=1}^\infty \lambda^{2n} = \sum_{n=1}^{\infty} 2^{-n} = 1$.

The subspace $H_0 = \{y \in H\,:\,\langle z, y \rangle = 0\}$ is closed in $H$ because $\langle z, \cdot\rangle: H \to \mathbb{R}$ is continuous.

The projection of $x = e_1$ to $H_0$ should be $$ y = e_1 - \langle z,e_1\rangle\, z = e_1 - \lambda z = \lambda^2 e_1 - \sum_{n=2}^\infty \lambda^{n+1}e_n \in \ell^2 \smallsetminus H_0. $$ For $k \geq 2$ put $$ z_k = \sum_{n=2}^k \;\lambda^{n} \,e_n + \lambda^{k+1} \frac{1}{1-\lambda} e_{k+1} \in H_0. $$ Then $y_k = \lambda^2 e_1-\lambda z_k \in H_0$ because $$ \langle y_k, z\rangle =\lambda^2 - \sum_{n=2}^{k}\,\lambda^{2n+1}-\frac{\lambda^{2k+2}}{1+\lambda} = \lambda^2 - \lambda^2 \sum_{n=1}^\infty \lambda^{2n} = 0. $$ On the other hand, we have $y_k \to y$ in $\ell^2$, so $$ \|e_1 - y\| \leq d(e_1,H_0) \leq \lim_{k\to\infty} \|e_1-y_k\| = \|e_1-y\| $$ and we're done because $y \in \overline{H}_0$ in $\ell^2$ is the only point realizing $d(e_1,\overline{H}_0)$ in $\ell^2$, thus there can be no point in $H_0$ minimizing the distance to $e_1$ because $y \notin H_0$.

t.b.
  • 80,986