20

Let H be a separable complex Hilbert space. Let A be a dense sub-space of H. Is it possible to find a complete orthonormal system for H that is contained in A?

user25640
  • 1,574
  • 11
  • 21
  • 2
    But, interestingly, this is not true for the nonseparable case. A dense linear subspace need not contain a maximal orthonormal system. – GEdgar Sep 23 '12 at 13:05
  • 1
    @GEdgar: I was just thinking about this (as you can see from the ramblings in my answer). Could you describe a counterexample? – Nate Eldredge Sep 23 '12 at 13:13

3 Answers3

24

nonseparable counterexample (but the original question specifies "separable")

Pointed out to me in newsgroup sci.math.research by Robert Israel in 2000.

See Dixmier, "Sur les bases orthonormales dans les espaces prehilbertiens", Acta Sci. Math. Szeged 15 (1953) 29-30.

Let $H_1$ be an infinite-dimensional separable Hilbert space. There is a linearly independent set $\{x_b: b \in B\}$ in $H_1$, indexed by a set $B$ of cardinality $c$ (e.g. in $L^2([0,1])$ you could take $x_b$ for $0 < b < 1$ as the step function $x_b(t) = 0$ for $t < b$, $1$ for $t \ge b$). Let $H_2$ be the non-separable Hilbert space with orthonormal basis $\{y_b: b \in B\}$ indexed by $B$. Let $K$ be the linear span of the vectors $\{x_b + y_b: b \in B\}$ in the direct sum $H_1 \oplus H_2$. Then there is no orthonormal set in $K$ whose span is dense in $K$. In fact:

1) Any orthonormal set in $K$ is countable.

Proof: Suppose $\{ z_a: a \in A \}$ is an orthonormal set in $K$. We can write each $z_a$ as a finite sum: $$z_a = \sum_{i=1}^n c_i (x_{b_i} + y_{b_i}),$$ where $\sum_i c_i x_{b_i} \ne 0$ since the $x_b$ are linearly independent. Let $\{ w_j: j \in \mathbb N \}$ be an orthonormal basis of $H_1$. For each $j$, we can have $\langle w_j, z_a\rangle \ne 0$ for at most countably many $a$'s, so there are only countably many $a$'s for which some $\langle w_j, z_a\rangle \ne 0$, and thus at most countably many $a$'s for which any $\langle w, z_a\rangle \ne 0$ with $w \in H_1$. But since $$\left\langle \sum_{i=1}^n c_i x_{b_i}, z_a\right\rangle = \left\|\sum_{i=1}^n c_i x_{b_i}\right\|^2 > 0,$$ this says that $A$ is at most countable.

2) The span of any countable orthonormal set is separable, and therefore is not dense in $K$.

GEdgar
  • 111,679
  • Nice answer. You should maybe clarify that $H_1$ is considered as the subspace $H_1 \oplus {0} \leq H_1 \oplus H_2$ in step 1). Also, you write "... there are at only countably many ...", i.e. the "at" should be removed. – PhoemueX Oct 20 '14 at 20:55
  • Isn't that precisely the construction of a preHilbert space that admits no ONB? – C-star-W-star Oct 21 '14 at 16:23
  • May you explain more about given $B$ of cardinality $c$, how can we find a non-separable Hilbert space with orthonormal basis ${y_b: b \in B}$ indexed by $B$? Thanks! – John Oct 17 '15 at 09:48
  • 1
    Given a set $B$, the space $l^2(B)$ of all functions $x : B \to \mathbb C$ such that $\sum_{b \in B} |x(b)|^2 < +\infty$ is your Hilbert space. – GEdgar Oct 17 '15 at 14:51
  • Sorry I couldn't follow the argument. Why $\overline{K}$ is not separable (since as you had shown, a separable space allows uncountably many linearly independent vectors)? In the last step, we know $\langle w, z_z\rangle\not=0$ for at most countably many $a$'s, but this is only true for a fixed $w$. But if we take $\sum c_i x_{b_i}$ as $w$, it's not fixed when $a$ varies? – Just a user Oct 18 '21 at 02:35
  • @Justauser ... The distance between any two of the vectors $x_b+y_b$ is $\ge$ the distance between the two vectors $y_b$ which is $\sqrt{2}$. So the set ${x_b+y_b : b \in B}$ is not separable; $K$ is metrizable and has a subset which is not separable, so this tells us $K$ itself is not separable. – GEdgar Oct 18 '21 at 10:27
13

Yes.

The proof is essentially the same as finding an orthonormal basis for $H$ itself. As a subspace of a separable metric space, $A$ has a countable dense subset $\{x_n\}$. Apply the Gram-Schmidt algorithm to $\{x_n\}$ to get an orthonormal set $\{e_n\}$. By construction, the linear span of $\{e_n\}$ contains every $x_n$, so it is dense in $A$ and hence also in $H$, making $\{e_n\}$ a complete orthonormal system.

Edit: I'd be interested in seeing an answer which addresses the non-separable case. I thought about using my argument with a transfinite induction, but it doesn't work. Suppose we have an orthonormal subset of $A$, $\{e_i\}_{i < \alpha}$ for some ordinal $\alpha$. Let $E_\alpha$ be the closed span of $\{e_i\}_{i <\alpha}$; if $E_\alpha \ne H$ then we can find some vector $x \in A \cap E_\alpha^c$. The natural choice for the next basis vector $e_\alpha$ would be $e_\alpha = x - P_{E_\alpha} x$, as this will be orthogonal to all $\{e_i\}_{i < \alpha}$; but as $A$ is not closed, $E_\alpha$ need not be contained in $A$, so we could get $e_\alpha \notin A$ and this breaks down.

Of course the problem can happen only when $\alpha$ is a limit ordinal, so this seems to suggest that maybe there is something special about the separable case, and perhaps the statement is false in general. But I couldn't think of a counterexample offhand.

Nate Eldredge
  • 97,710
  • The Gram-Schmidt algorithm I was taught only works on countable subsets where all the $x_n$ are linearly independent. But you can create such a linearly independent dense subset ${y_n}$ from ${x_n}$ by omitting all $x_n$ that are linearly dependent. The span of both subsets then remains the same. – The Coding Wombat Dec 02 '19 at 12:03
  • 3
    @TheCodingWombat: Sure. Or, if you apply Gram-Schmidt on a linearly dependent set, it spits out a bunch of zeros which you can just ignore. – Nate Eldredge Dec 02 '19 at 13:37
3

Let $\{e_n\}_{n=1}^{+\infty}$ a countable orthonormal basis for $H$. Since $A$ is dense in $H$, for each $n,k\geq 1$, let $a_{n,k}\in A$ such that $\lVert e_n-a_{n,k}\rVert\leq k^{—1}$. We apply Gram-Schmidt process to get an orthonormal family $\{b_{nk}\}\subset A$ (as $A$ is a subspace) which generates $\operatorname{Span}(a_{n,k},k,n\geq 1\}$. Then $\{b_{n,k}\}$ is an orthonormal basis.

Davide Giraudo
  • 172,925