7
$\begingroup$

Is there a (simple) proof for Nelson's theorem that a symmetric operator is essentially selfadjoint if it contains a dense subset of analytic vectors?

$\endgroup$
2
$\begingroup$

Analytic vectors are used to construct continuous unitary group on a dense subset, it extends to the entire space and the generator is the self-adjoint closure of the original operator, see p.200-202 in Reed and Simon.

$\endgroup$
3
+200
$\begingroup$

Let $A$ be a symmetric operator in a Hilbert space $\mathcal H$. For $f \in \mathcal C^\infty(A)$ let $\mathcal L_f$ be the vector space spanned by $\{A^k f | k=0, 1,... \}$ and $\mathcal H_f := \overline{\mathcal L_f} $.

Nussbaum's "Vector of Uniqueness" Lemma: If for non-real $\lambda$, $\mathcal D(A)$ contains a dense set of vectors $u$ such that $\mathcal R(A | \mathcal L_u-\lambda)$ is dense in $\mathcal H_u$ then $\mathcal R(A-\lambda)$ is dense in $\mathcal H$ ($A$ is essentially self-adjoint).
Proof: There are vectors in $\mathcal R(A | \mathcal L_u-\lambda)$ arbitrarily close to $u$ and $\mathcal R(A | \mathcal L_u-\lambda) \subseteq \mathcal R(A-\lambda)$. $\blacksquare$

$f$ is called an analytic vector for $A$ if $ \sum_{n=0}^\infty \frac{\Vert A^nf \Vert s^n}{ n!} \lt \infty $ for some $s\gt 0$ (depending on $f$) $(\Leftarrow \Vert A^nf \Vert = O(n!) )$.

Nelson's Theorem: If $\mathcal D(A)$ contains a dense set of analytic vectors then $A$ is essentially self-adjoint.
Proof: We prove a stronger theorem of Nussbaum with the weaker condition "quasi-analytic" vector: $\sum_{n=1}^\infty \frac{1}{\Vert A^nf \Vert ^{\frac{1}{n}}} = \infty$ ($\Leftarrow\Vert A^nf \Vert = O( n^n)$). W.l.o.g. $\Vert f \Vert=1$.
On an orthonormal basis $\{e_n\}$ constructed from $\{A^nf \}$ by the Gram-Schmidt process, $A | \mathcal L_f$ is a Jacobi matrix operator

$$ \begin{bmatrix} a_0 & b_0 & 0 & 0 & \cdots \\ b_0 & a_1 & b_1 & 0 & \cdots\\ 0 & b_1 & a_2 & b_2 & \cdots \\ 0 & 0 & b_2 & a_3 & \cdots \\ \vdots & \vdots & \vdots & \vdots & \ddots \\ \end{bmatrix}. $$ ($Ae_m \in \text{span}(e_0,...e_{m+1})$ so $(e_n, Ae_m) = 0$ for $n\gt m+1$. Since $A$ is symmetric, the matrix must be tridiagonal. The off diagonal elements can be made positive real by choice of phase of the $e_n$ .)

Then $b_0b_1...b_{n-1}=(e_n, A^n f)\le \Vert A^nf \Vert$. Using Carleman's inequality: $$\infty = \sum_{n=1}^\infty\frac{1}{\Vert A^nf \Vert ^{\frac{1}{n}}} \le \sum_{n=1}^\infty \frac{1}{b_0 b_1...b_{n-1}}^\frac{1}{n} \le e\sum_{n=0}^\infty \frac{1}{b_n}$$ By Carleman's test $\mathcal A | \mathcal L_f$ is essentially self-adjoint and by Nussbaum's lemma so is $A$. $\blacksquare$

Carleman's test: A Jacobi matrix operator $J$ is essentially self-adjoint if $\sum_{n=0}^\infty \frac{1}{b_n} =\infty$.
Proof: Suppose a vector $\{p_k\}$ is orthogonal to the range of $J-\lambda$ for non-real $\lambda$. Then: \begin{align} \lambda p_0 = a_0 p_0 +b_0 p_1 \\ \lambda p_n = b_{n-1}p_{n-1} + a_np_n +b_n p_{n+1} && n=1, 2, ... \end{align} Multiplying this and its conjugate by $\bar p_n$ and $p_n$ respectively and subtracting leads to: \begin{align} \frac{b_0(p_1 \bar p_0 -\bar p_1 p_0)} {\lambda - \bar \lambda} = |p_0|^2 \\ \frac{b_n(p_{n+1} \bar p_n -\bar p_{n+1} p_n )}{\lambda - \bar \lambda} = \frac{b_{n-1}(p_n \bar p_{n-1} -\bar p_n p_{n-1})}{\lambda - \bar \lambda}+ |p_n|^2 && n=1, 2, ... \end{align} Note that the LHSs are $ \ge |p_0|^2$. Using the Cauchy-Schwarz inequality: $$\infty = |p_0|^2\sum_{n=0}^\infty \frac{1}{b_n} \le \frac{1}{\lambda - \bar\lambda}\sum_{n=0}^\infty (p_{n+1} \bar p_n -\bar p_{n+1} p_n) \le \frac{1}{|\text{Im}(\lambda)|}\sum_{n=0}^\infty |p_n|^2$$ $\blacksquare$

$\endgroup$
  • $\begingroup$ Just wanted to let you know that I stumbled on your post and I found it really useful. Thank you $\endgroup$ – Giuseppe Negro May 3 '16 at 10:15

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.