This question appeared on an old final exam and I am having difficulty completing it for practice.

Let $S_r$ and $S_l$ be defined on the hilbert space $l_2[0,\infty)\to l_2[0,\infty)$ as the following: $S_r (x_0, x_1, x_2, \dots) \mapsto (0,x_0, x_1, \dots)$ and $S_l (x_0, x_1, x_2,\dots) \mapsto (x_1, x_2, x_3,\dots)$. That is to say, $S_r$ is the right shift operator and $S_l$ is the left shift operator.

I fully understand how to describe the spectrum of each individually, however things become clouded once they are combined.

The question asks to:

Describe the spectrum of $S_r + S_l$

That is, $(S_r + S_l)(x_0, x_1, x_2,\dots) \mapsto (x_1, x_0+x_2, x_1+x_3,\dots)$

To begin I noted first that $S_r + S_l$ is self-adjoint which implies two things. First, the spectrum will necessarily be a subset of $\mathbb{R}$. Second, it implies that $R_\sigma(S_r+S_l)$ is empty.

Suppose that $\mu\in P_\sigma(S_r+S_l)$, then $\mu x = (S_r+S_l)x$ for some $x=(x_0,x_1,x_2,\dots)$.

Then $\mu(x_0,x_1,\dots) = (x_1,x_0+x_2, x_1+x_3, \dots, x_{n-1}+x_{n+1},\dots)$ and $x_1 = \mu x_0$, $x_{n+1} = \mu x_n - x_{n-1}$

So, for $x$ to be an eigenvector of $\mu$, we must have $x=x_0(1, \mu, \mu^2-1, \mu^3-2\mu, \mu^4-3\mu^2+1,\dots,\mu^n - (n-1)\mu^{n-2} + O(\mu^{n-4}+1),\dots) ~~~(\star)$

It follows that as $n\to\infty$ if $|\mu|>1$, then $|((S_l+S_r)x)_n|\to\infty$ and so no eigenvalues of $S_l+S_r$ are greater than 1.

For $\mu=1$ we must have $x_1 = x_0$, $x_2 = 0$, $x_3 = -x_0$, $x_{n+1} = x_n - x_{n-1}$

And we get that $x = x_0(1,1,0,-1,0,1,0,-1,\dots)$ which is only in $l_2[0,\infty)$ if $x_0 = 0$, so $1$ is not an eigenvalue of $S_l+S_r$.

In the case that $\mu=0$, by $(\star)$

$x = x_0(1,0,-1,0,1,\dots)$, and since we are in $l_2[0,\infty)$, $x_0=0$ and $x=0$, so $0$ is not an eigenvalue of $S_l+S_r$.

With this work so far, I was able to show as well that the operator norm is $2$ and the operator is definitely not compact (else $\pm 2$ would have been an eigenvalue)

It is at this point that I lose track of where to continue from here. With what argument can I show that for $|\mu|<1$ that they are or aren't eigenvalues? I expect but cannot find the right wording to show that there are in fact no eigenvalues.


There is a general method for finding sequences which satisfy linear recurrent relations with constant coefficients. In this case, we have the relation $$ x_{n+1}+x_{n-1} = \mu x_n \text{ for } n\geq 1. \text{ (*)} $$ Here is how we solve it: consider all the sequences (not nesessarily from $l^2$) which satisfy (*). They form a linear space (it's easy to check the axioms); let's denote it $M$.

What is $\dim M$? One can see that

$\forall u,v\in \mathbb{C}~ \exists y^{u,v}\in M: y^{u,v}_0 = u, y^{u,v}_1 = v$;

if $x,y\in M$, $x_0 = y_0, x_1 = y_1$, then $x = y$.

From this we can see that any $x\in M$ can be represented as a linear combination of two linearly independent sequences $y^{1,0}$ and $y^{0,1}$: $$ x = x_0\cdot y^{1,0} + x_1\cdot y^{0,1}, $$ so $\dim M = 2$.

How can we find a nice basis for $M$? For $\lambda\in \mathbb{C}$, let $p^\lambda$ denote the sequence of its powers, $(p^\lambda)_n = \lambda^n$. It's easy to see that $p^\lambda\in M$ if and only if $\lambda^2 + 1 = \mu\lambda$. For $|\mu|<2$, this equation has two roots: $$ \lambda_1 = \frac{\mu +\sqrt{4-\mu^2}i}{2}; \lambda_2 = \lambda_1^* = \lambda_1^{-1}. $$ So $p^{\lambda_1}$ and $p^{\lambda_2}$ form a basis for $M$. Now, for $x$ to satisfy $(S_l+S_r)x = \mu x$, one more condition is needed in addition to (*): $x_1 = \mu x_0$. So for $x = c_1 p^{\lambda_1} + c_2 p^{\lambda_2}$ we have $$ \begin{cases} c_1+c_2 = x_0\\ c_1\lambda_1 + c_2\lambda_2 = \mu x_0 \end{cases} $$ from which we get $$ c_1 = \frac c2x_0, c_2 = \frac{c^*}2x_0, \text{ where }c = 1 - \frac{\mu}{\sqrt{4-\mu^2}}i, $$ so we can write $$ x_n = c_1\lambda_1^n + c_2\lambda_2^n = \Re(c \lambda_1^n) x_0. $$ Since $|\lambda_1| = 1$, the sequence $z_n = \Re(c \lambda_1^n)$ would converge to $0$ only if it was constant $0$, but $z_0 = 1$, so $z\not\in l^2[0,\infty)$. This means that $x \in l^2[0,\infty)$ only if $x_0= 0$, i.e., the values from $(-2, 2)$ are not eigenvalues.

Note, though, that the sequence $z$ is bounded. So when you say that if $|\mu|\geq 1$, and $x$ satisfies $(S_l+S_r)x = \mu x$, $x_0\neq 0$, then $|x_n|\to \infty$, it's not true: if $1<|\mu|<2$, then $x$ is bounded. The problem is that the term which you said is $O(\mu^4+1)$ is a sum of many terms with coefficients which depend on $n$ and have different signs, so you cannot say so easily how it behaves when $n\to \infty$.

So you need a new way to show that when $|\mu|\geq 2$, and $x$ satisfies $(S_l+S_r)x = \mu x$, then either all $x_n$ are equal to $0$ or $|x_n|\to \infty$ as $n \to \infty$.

If $\mu = \pm 2$, then $x = x_0(1,\pm 2, 3, \pm 4, \dots)$;

And if $|\mu|>2$, then you can apply the same method as before and see that $|x_n| = |x_0|O(\left(\frac{|\mu|+\sqrt{\mu^2-4}}{2}\right)^n)$.

(You could also use the same method to find the solutions for $\mu = \pm 2$, but with a little modification: in these cases the equation on $\lambda$ has a single root of multiplicity $2$. And one can show that the equation has a root $\lambda$ of multiplicity $m$, then the sequences $n^k\lambda^n$, where $k = 0, \dots, m-1$, lie in $M$.)


Even though this question is almost three years old: for anyone that stumbles across this problem like me, I want to give a simple argument on why the operator given by the sum of the right shift $S_r$ and left shift $S_l$ on $\ell_2(\mathbb N)$ does not have any eigenvalues. I will need a few well known results which can for example be looked up in Chapter VII.§1 of the book "Introduction to Hilbert Space" by Berberian (1976).

  • All eigenvalues of a bounded operator $T$ are bound by $\Vert T\Vert$.
  • All eigenvalues of a self-adjoint operator are real.
  • $\mu\in\mathbb C$ is eigenvalue of $S_l$ if and only if $|\mu|<1$.

As stated in the original post, $S_l+S_r$ is self-adjoint and $\Vert S_l+S_r\Vert\leq \Vert S_l\Vert+\Vert S_r\Vert=2$ so if that operator has eigenvalues, those have to be elements of the real interval $[-2,2]$. Let $\lambda\in[-2,2]$ be given and assume there exists $x\in\ell_2(\mathbb N)$ such that $(S_l+S_r)x=\lambda x$ (our goal will be to show that $x$ has to be the zero sequence). Applying $S_l$ to this equation from the left gives

$$ 0=(S_l^2+S_lS_r-\lambda S_l)x=(S_l^2-\lambda S_l+\operatorname{id})x $$

since $S_lS_r$ obviously is the identity $\operatorname{id}$ on $\ell_2(\mathbb N)$. Looking at the related equation $t^2-\lambda t+1=0$ yields the solutions

$$ t_\pm=\frac{\lambda}2\pm i\sqrt{1-\frac{\lambda^2}4} $$

and thus we have

$$ ( S_l-t_+\operatorname{id} )( S_l-t_-\operatorname{id} ) x=0. $$

Note that $|t_\pm|=1$ since $\lambda\in[-2,2]$ so $t_\pm$ can not be an eigenvalue of $S_l$ as stated at the beginning. This implies $( S_l-t_-\operatorname{id} ) x=0$ and further $ x=0$ so $\lambda$ is not an eigenvalue of $S_l+S_r$. In total we showed that $S_l+S_r$ does not have any eigenvalues.


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.