Ivan Shishkin, Rye (1878)

Discussions

On the localization of eigenvalues : Gershgorin and Hadamard

0 messages

Solution

Solution by pietro.pende · EN

  1. Let λ∈Sp⁡(U)\lambda\in\operatorname{Sp}(U). By definition of the spectrum, there exists an eigenvector

    Z=(z1⋮zn)∈Cn∖{0}Z= \begin{pmatrix} z_1\\ \vdots\\ z_n \end{pmatrix} \in\mathbb{C}^n\setminus\{0\}

    such that

    UZ=λZ.UZ=\lambda Z.

    Since Z≠0Z\neq0, the set

    {∣zx∣  |  x∈⟦1,n⟧}\left\{|z_x|\;\middle|\;x\in\llbracket 1,n\rrbracket\right\}

    is a non-empty finite subset of R\mathbb{R}. It therefore has a maximum. Choose

    i∈⟦1,n⟧i\in\llbracket 1,n\rrbracket

    such that

    ∣zi∣=max⁡x∈⟦1,n⟧∣zx∣.|z_i| = \max_{x\in\llbracket 1,n\rrbracket}|z_x|.

    Since Z≠0Z\neq0, this maximum is strictly positive, and consequently

    zi≠0.z_i\neq0.

    The equality UZ=λZUZ=\lambda Z gives, by considering its ii-th component,

    ∑j=1nuijzj=λzi.\sum_{j=1}^n u_{ij}z_j=\lambda z_i.

    Isolating the term corresponding to j=ij=i, we obtain

    (λ−uii)zi=∑1⩽j⩽nj≠iuijzj.(\lambda-u_{ii})z_i = \sum_{\substack{1\leqslant j\leqslant n\\j\neq i}}u_{ij}z_j.

    Since zi≠0z_i\neq0, we can divide by ziz_i, whence

    λ−uii=∑1⩽j⩽nj≠iuijzjzi.\lambda-u_{ii} = \sum_{\substack{1\leqslant j\leqslant n\\j\neq i}} u_{ij}\frac{z_j}{z_i}.

    Taking absolute values and using the triangle inequality, we obtain

    ∣λ−uii∣≤∑1⩽j⩽nj≠i∣uij∣∣zj∣∣zi∣.|\lambda-u_{ii}| \leq \sum_{\substack{1\leqslant j\leqslant n\\j\neq i}} |u_{ij}| \frac{|z_j|}{|z_i|}.

    By the definition of ii,

    ∀j∈⟦1,n⟧,∣zj∣⩽∣zi∣,\forall j\in\llbracket 1,n\rrbracket,\qquad |z_j|\leqslant|z_i|,

    and hence

    ∀j∈⟦1,n⟧,∣zj∣∣zi∣⩽1.\forall j\in\llbracket 1,n\rrbracket,\qquad \frac{|z_j|}{|z_i|}\leqslant1.

    It follows that

    ∣λ−uii∣⩽∑1⩽j⩽nj≠i∣uij∣=Λi.|\lambda-u_{ii}| \leqslant \sum_{\substack{1\leqslant j\leqslant n\\j\neq i}}|u_{ij}| = \Lambda_i.

    By the definition of DiD_i, we therefore have

    λ∈Di.\lambda\in D_i.

    Thus, every eigenvalue of UU belongs to one of the disks DiD_i. Consequently,

    Sp⁡(U)⊂⋃i=1nDi  .\operatorname{Sp}(U) \subset \bigcup_{i=1}^nD_i \; .

  2. Suppose that

    ∀i∈⟦1,n⟧,∣uii∣>∑1⩽j⩽nj≠i∣uij∣.\forall i\in\llbracket1,n\rrbracket,\qquad |u_{ii}|> \sum_{\substack{1\leqslant j\leqslant n\\j\neq i}}|u_{ij}|.

    For every i∈⟦1,n⟧i\in\llbracket1,n\rrbracket, we then have

    ∣uii∣>Λi.|u_{ii}|>\Lambda_i.

    Consequently,

    0∉Di.0\notin D_i.

    Indeed, if 0∈Di0\in D_i, then, by the definition of DiD_i,

    ∣uii∣⩽Λi,|u_{ii}|\leqslant \Lambda_i,

    which would contradict the hypothesis.

    Thus,

    0∉⋃i=1nDi.0\notin\bigcup_{i=1}^nD_i.

    Now, by the first question,

    Sp⁡(U)⊂⋃i=1nDi.\operatorname{Sp}(U) \subset \bigcup_{i=1}^nD_i.

    We deduce that

    0∉Sp⁡(U).0\notin\operatorname{Sp}(U).

    By definition of the spectrum,

    0∉Sp⁡(U)⟺ker⁡(U)={0}.0\notin\operatorname{Sp}(U) \quad\Longleftrightarrow\quad \ker(U)=\{0\}.

    The linear map associated with UU is therefore injective. Since it is an endomorphism of the finite-dimensional vector space Cn\mathbb C^n, it is then bijective. Consequently, UU is invertible.

    Thus,

    U∈GLn(C)  .U\in GL_n(\mathbb C)\; .

No messages yet.