I have been struggling to analyse the spectrum of the adjoint of a bounded linear operator on a hilbert space.
Throughout the internet I have found vague references that $\sigma(T^*) = \sigma(T)$ but I fail to be able to prove that.
What I have succeeded in doing is to prove that $\rho(T^*) = \overline{\rho(T)}$ (where $\rho$ is the resolvent $\rho(T) = \mathbb{C} \backslash \sigma(T)$).
Thus imho $\sigma(T^*)$ should be equal to $\overline{\sigma(T)}$ and not the above...

  • $\begingroup$ This isn't true for Hilbert spaces (see msteve's answer). It is however true for Banach spaces. $\endgroup$ Feb 7 '17 at 10:05
  • 1
    $\begingroup$ @enthdegree How can that be, considering that every Hilbert space is a Banach space? $\endgroup$
    – iolo
    Feb 7 '17 at 18:39
  • 3
    $\begingroup$ Because adjoint operators in (complex) Hilbert spaces are constructed using an inner product $(\cdot | \cdot):\mathfrak{H}\times \mathfrak{H}^\ast \to \mathbb{C}$, which is sesquilinear. On the other hand the composition-with-dual used to construct Banach space adjoint operators, $\langle \cdot,\cdot\rangle:\mathfrak{X}\times \mathfrak{X}^\ast \to \mathbb{F}$, is an outright bilinear form. $\endgroup$ Feb 8 '17 at 1:31
  • 1
    $\begingroup$ To expand on @enthdegree's comment, these two different notions of adjoint can be reconciled: the Hilbert space adjoint is the Banach space adjoint using the natural identification of the dual of a Hilbert space with its conjugate space. This identification arises via an antilinear map, and hence the introduction of the conjugate. $\endgroup$
    – J. Loreaux
    Feb 9 '17 at 18:08

You're correct that $\sigma(T^*) = \overline{\sigma(T)}$, where the overline denotes complex conjugation (and not closure as a subset of $\mathbb{C}$ - recall that the resolvent set of a bounded operator is open, and hence the spectrum is closed). In order to conclude that $\sigma(T^*) = \sigma(T)$, you need to impose the stronger condition that the eigenvalues be real (by specifying that $T$ be self-adjoint, e.g.).

  • $\begingroup$ If the hilbert space is real, then the equality holds, right? Or in the real case these things are not well defined? $\endgroup$
    – Filburt
    Jan 26 '18 at 1:59

The critical point is in complex Hilbert space you use a sesquilinear inner product to make adjoints but in Banach space you use a bilinear form. So in the Banach situation the conjugates disappear.

To reduce confusion I'll replace the two senses of adjoint with $T^T$ for Banach adjoint and $T^H$ for Hilbert adjoint. We borrow the rest of our notation from Pedersen:

Adjoint $T^{H}$ for $T$ in (complex) Hilbert space $\mathfrak{H}$ is constructed through a sesqui-linear inner product $(\cdot|\cdot):\mathfrak{H}\times \mathfrak{H}^\ast→\mathbb{C}$ as follows: $$T^H:\mathfrak{H}^\ast\to \mathfrak{H}^\ast; \qquad (\cdot|T^H\phi)\triangleq (T \cdot |\phi) \in \mathfrak{H}^\ast $$

On the other hand, in a Banach space $\mathfrak{X}$, adjoint $T^T$ of $T$ is constructed using a bi-linear composition-with-dual operator $\langle \cdot,\cdot\rangle :\mathfrak{X}\times \mathfrak{X}^\ast\to \mathbb{F}$ in an identical looking way:

$$T^T:\mathfrak{X}^\ast\to \mathfrak{X}^\ast; \qquad \langle \cdot, T^T\phi\rangle\triangleq\langle T\cdot ,\phi\rangle \in \mathfrak{X}^\ast$$

Spectrum is: $$\sigma(T)=\left\{ \lambda : \lambda I - T \not\in \mathbf{GL} \right\}.$$

$\mathbf{GL}$ is the general linear group, 'operators with bounded inverse.'

With this notation, it is true that $\sigma(T)=\sigma(T^T)=\overline{\sigma(T^H)}$.

We see $\sigma(T)=\overline{\sigma(T^H)}$ by computing: \begin{align} \sigma(T^H)&=\{\lambda: \lambda I^H - T^H \not\in \mathbf{GL}(\mathfrak{H^\ast})\} \\ &= \{\overline\lambda: \phi\mapsto\phi((\lambda I - T)(\cdot)) \not\in\mathbf{GL}(\mathfrak{H}^\ast)\}\\ &= \{\overline\lambda: (\lambda I - T) \not\in\mathbf{GL}(\mathfrak{H})\} \tag{A} \\ &= \overline{\sigma(T)} \end{align} where $(\text{A})$ follows because, letting $S^H=\phi \mapsto \phi((\lambda I-T)(\cdot))$, if $(\lambda I -T)\in\mathbf{GL}(\mathfrak{H})$, then by computation $\phi\mapsto \phi((\lambda I-T)^{-1}(\cdot))\in \mathbf{GL}(\mathfrak{H}^\ast)$ is an inverse of $(S^H)^{-1}$ so $S^H \in\mathbf{GL}(\mathfrak{H}^\ast)$. On the other hand $S^H\in\mathbf{GL}(\mathfrak{H}^\ast)\Rightarrow (S^H)^{-1}\in \mathbf{GL}(\mathfrak{H}^\ast)$ and by computation, $((S^H)^{-1})^H\in \mathbf{B}(\mathfrak{H})$ must behave as an inverse for $\lambda I - T.$ (We have just shown the complements of the two sets are subsets of one another)

The Banach part follows by the same argument, replacing $(\cdot|\cdot)$ with $\langle \cdot,\cdot\rangle$, $\mathfrak{H}$ with $\mathfrak{X}$, $\cdot^H$ with $\cdot ^T$, and getting rid of conjugation.

This is part of E 4.1.11 in Analysis Now by Pedersen.

  • $\begingroup$ you make the notation more complicated, not simpler. $\endgroup$
    – Rüdiger
    Feb 9 '17 at 22:50
  • 1
    $\begingroup$ Unfortunately you need precisely all of it to make sense of the question $\endgroup$ Feb 10 '17 at 0:13
  • $\begingroup$ your explanation is so clear! thank you! $\endgroup$
    – allen i
    Apr 20 '20 at 18:26

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.