I have the following two definitions in my notes:

The spectrum of an operator:

We define $\sigma(T)$, the spectrum of T, by,

$$\sigma(T):=\{\lambda\in\mathbb C: T-\lambda I\,\, \text{is not invertible}\}$$

The point spectrum of an operator:

The set of all eigenvalues of T is called the point spectrum of T and is denoted by $\sigma_p(T):$

$$\sigma_p(T):=\{\lambda \in \mathbb C :\lambda\, \text{ is an eigenvalue of $T$}\}$$

After everything I have done with eigenvalues, I would have thought that these definitions were equivalent. The standard eigenvalue problem is to find all $\lambda \in \mathbb C$ where,

$$Tx=\lambda x$$

For an operator $T$ and some vector in the space we are working in, $x$. From this it follows that,

$$(T-\lambda I) x = 0$$

Where the operator $T-\lambda I$ must be singular (for the eigenvalue problem).

Am I right in saying that these two definitions are equivalent? If they were, I am unsure as to why they are given two seperate definitions. And, if they are not, could somebody give me an example that differentiates the two definitions?



Suppose $T$ is a bounded operator on a Banach space $X$. $\lambda\in\rho(T)$ iff $T-\lambda I$ is a linear bijection. In that case, the inverse $(T-\lambda I)^{-1}$ is automatically continuous by the closed graph theorem.

There are three basic things that can stand in the way of $T-\lambda I$ being invertible.

  1. $T-\lambda I$ is not injective. Equivalently, $Tx=\lambda x$ for some $x\ne 0$, which means that $\lambda$ is an eigenvalue of $T$.

  2. The range of $T-\lambda I$ is not dense in $X$. Equivalently, there is a non-zero bounded linear functional $x^{\star}\in X^{\star}$ such that $x^{\star}((T-\lambda I)y)=0$ for all $y$; this, in turn is equivalent to $(T-\lambda I)^{\star}x^{\star}=0$, which means $\overline{\lambda}$ is an eigenvalue of $T^{\star}$.

  3. $T-\lambda I$ is injective and has dense, non-closed range. In this case, the inverse $T-\lambda I$ cannot be bounded, which gives the existence of a sequence $\{ y_n \}$ of unit vectors in the range of $\mathcal{R}(T-\lambda I)$ such that $\lim_n\|(T-\lambda I)y_n\|=\infty$. After renormalization, you obtain a sequence of unit vectors $\{ x_n \}$ such that $\lim_n\|(T-\lambda I)x_n\|=0$. So $\lambda$ is an approximate eigenvalue of $T$ in this case.

You have an example of (1); there are plenty of examples of eigenvalues.

Case (2) is peculiar to infinite-dimensional spaces. A simple example is the shift operator on $\ell^2(\mathbb{N})$: $$ U(a_0,a_1,a_2,\cdots) = (0,a_0,a_1,a_2,\cdots) $$ For $a=(a_0,a_1,a_2,\cdots)$, $\|Ua\|=\|a\|$. So $U$ is injective, and its range is a proper closed subspace of $\ell^2(\mathbb{N})$. You can't do this with a finite shift.

Case (3) is also peculiar to infinite-dimensional spaces. A nice example of this is the multiplication operator $(Mf)(x)=xf(x)$ on $L^{2}[0,1]$. Every $\lambda \in [0,1]$ is an approximate eigenvalue. The simplest case is $\lambda=0$. The functions $$ f_n = \sqrt{n}\chi_{[0,1/n]},\;\;\; n=1,2,3,\cdots, $$ are unit vectors because $$ \int_{0}^{1}|f_n|^2dx = \int_{0}^{1/n}ndx = 1. $$ And $$ \|Mf_n\|^2 = \int_{0}^{1}x^{2}|f_n|^2dx = \int_{0}^{1/n}n xdx= \frac{n}{2}x^{2}|_{0}^{1/n}= \frac{1}{n}\rightarrow 0. $$ So $0$ is an approximate eigenvalue of $M$. The range is dense because $M=M^{\star}$ means $M^{\star}$ does not have eigenvalue $0$ either. You can see how $\chi_{[\lambda-1/n,\lambda+1/n]}$ is very close to being an eigenfunction of $M$ with eigenvalue $\lambda$, if $\lambda \in [0,1]$, but it just can't quite get there.


For $T$ to have $\lambda$ as an eigenvalue, $T-\lambda I$ must be non-injective. For $\lambda$ to be in the spectrum of $T$, it must only be non-invertible. These are equivalent when $T$ is an operator on a finite-dimensional space, but not in general!

For example, let $T$ be the shift operator $(x_0,x_1,\dots) \mapsto (0,x_0,x_1,\dots)$ on your favorite subspace of $\Bbb{R}^\infty$.

Then $T-\lambda I$ maps $x=(x_0,x_1,\dots)$ to $(-\lambda x_0, x_0-\lambda x_1, x_1-\lambda x_2,\dots)$, which only vanishes when $x=0$. So the point spectrum of $T$ is empty. On the other hand, $T$ itself is not surjective and hence not invertible, so the spectrum of $T$ contains $0$.


Consider an example is the shift operator on ℓ2(N): U(a0,a1,a2,⋯)=(0,a0,a1,a2,⋯). The spectrum contains 0. Is there is spectrum contains any other elements?


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.