Let $(\mathcal{H},\langle \cdot, \cdot \rangle)$ be an arbitrary Hilbert space. Can one construct two independent and identically distributed random elements $X,Y:(\Omega,\mathbb{F},P)\to (\mathcal{H},\langle \cdot, \cdot \rangle)$ with $\text{supp}(X)\not = \{0\}$, such that $$ \langle X(\omega),Y(\omega) \rangle =0 $$ for almost all $\omega\in\Omega$, i.e. $X$ and $Y$ are almost surely orthogonal.


I have shown that this can not be done for separable Hilbert spaces $\mathcal{H}$, but is it possible to make such a construction in non-separable Hilbert spaces?

  • $\begingroup$ I'm not sure what you mean by a random function taking values in a Hilbert space. Usually people mean the Bochner integrable functions on a probability space. And their range is, by definition, inside a separable subspace almost surely. $\endgroup$ – Stephen Montgomery-Smith Nov 8 '16 at 17:54
  • $\begingroup$ @StephenMontgomery-Smith - $X$ and $Y$ are random Borel elements in $\mathcal{H}$ (not random functions, but random variables if you will, such that each $\omega$ maps to $X(\omega)\in\mathcal{H}$). We don't assume that $X$ and $Y$ are bochner integrable. This would simply lead the the arguments for separable Hilbert spaces, in which case I have already shown non-existence. I would like to make an argument that almost surely orthogonal random i.i.d. elements are not in general degenerate, but they are if $\mathcal{H}$ is separable (or the random elements are Bochner-measurable) they are. $\endgroup$ – Martin Nov 8 '16 at 19:23
  • $\begingroup$ What precisely do you mean by "random element?" I assume the map $\omega \mapsto X(\omega)$ is measurable, but with respect to which topology on $\mathcal H$? The norm topology or the weak topology? (I think the Borel sets generated can be different for non-separable spaces.) $\endgroup$ – Stephen Montgomery-Smith Nov 8 '16 at 21:08
  • $\begingroup$ Measurable with respect to the sigma algebra generated by the norm topology. Sorry if this was not explicitly written in the question. $\endgroup$ – Martin Nov 8 '16 at 21:18

Counterexample. This counterexample is inspired by class notes by D.J.H. Garling from the 1980s.

Let $\kappa$ be a measurable cardinal https://en.wikipedia.org/wiki/Measurable_cardinal which is also a limit ordinal. Thus there is a $\{0,1\}$ valued measure on $\kappa$ in which every subset is measurable.

Let $\mathcal H$ be the Hilbert space whose basis is of cardinality $\kappa$, and let $(e_{\alpha})_{\alpha \in \kappa}$ be a basis. Let $X,Y:\kappa \to \mathcal H$ be the functions $X(\alpha) = e_{\alpha}$ and $Y(\alpha) = e_{\alpha'}$, where $\alpha'$ denotes the successor ordinal of $\alpha$.

Clearly $\langle X,Y\rangle = 0$ everywhere. It remains to show that $X$ and $Y$ are independent. But this follows because any two subsets of $\kappa$ are independent (since their measures can only take the values $0$ or $1$).

It does use measurable cardinals, whose existence cannot be proved. But most likely, if their existence can be disproved, then probably the same proof will show ZF is inconsistent.

  • 1
    $\begingroup$ As an aside, if $X$ and $X'$ are independent copies of $X$ defined on the first and second coordinates of the product space $\kappa \times \kappa$, then $\langle X,Y\rangle$ is non-measurable! This is essentially the point D.J.H. Garling was trying to make in his notes - separability is no mere technicality. $\endgroup$ – Stephen Montgomery-Smith Nov 10 '16 at 20:01

Let $\mathcal{A} = \mathbb{R}^{[0,1]}$ be the set of all functions from $[0,1]$ to $\mathbb R$, not necessarily continuous, equipped with the product Gaussian measure. Let $\mathcal H$ be $L^2(\mathcal{A})$. $\mathcal H$ is not separable.

Let $x$ be uniformly distributed on $[0,1]$ and let $X_x$ be the function that takes $a \in \mathcal A$ to $a(x)$. The marginal distribution of each $X$ is a Gaussian, so $X_x \in \mathcal H$. Furthermore, $X_x$ is independent of $X_y$ unless $x=y$, which happens with probability zero.


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.