>[!quote] In a Nutshell
>Eigenvectors are the directions that are only scaled when a given [[Maps and Induced Structures - Functions, Pushforwards and Pullbacks|transformation]] is applied to it. The scaling is denoted the eigenvalues. Intuitively, computing the eigenvalues is decomposing the complex problem of a transformations into its simpler, linear components. This can be achieved by changing the basis accordingly.
---
>[!info] Special Eigenvalue Problem
>For a given [[Matrices|matrix]] $\mathbf{A} \in \mathbb{C}^{n\times n}$ and $n \in \mathbb{N}$, $\lambda \in \mathbb{C}$ is an eigenvalue if there is a $x \in \mathbb{C}^{n} \setminus \{0\}$, for which$\mathbf{A}x=\lambda x.$
- $x$ is eigenvector for eigenvalue $\lambda$
- Eigenvectors are not unique
- The polynomial $\det(\mathbf{A-\lambda I})$ is called *characteristic polynomial* (degree $n$)
- If $x$ is eigenvector for eigenvalue $\lambda$, then $p(\mathbf{A})x=p(\lambda)x$ for any polynomial $p$
- For [[Matrices|symmetric matrices]] $\mathbf{A}\in \mathbb{R}^{n \times n}$
- All eigenvalues are real, eigenvectors can be chosen as real
- Eigenvectors of different eigenvalues are orthogonal
- For every eigenvalue of multiplicity $m$ there are $m$ linear independent eigenvectors
- There is an orthogonal basis $\{\mathbf{u}_{1}, ...,\mathbf{u}_{n} \}$ of $\mathbb{R}^n$ of eigenvectors to the eigenvalues $\lambda_1,...,\lambda_n$ $\mathbf{Au}_{j}=\lambda_{j}\mathbf{u}_{j}, \qquad \langle\mathbf{u_{j},\mathbf{u}_{i}} \rangle=\delta_{i,j}=\cases{1 \quad \text{if } i=j\\0 \quad \text{else}}$
---
#### Orthonormal Basis of Eigenvectors
For every real [[Matrices|symmetric matrix]] $\mathbf{A}\in \mathbb{R}^{n \times n}$, there exists an orthonormal basis of eigenvectors to the eigenvalues $\lambda_{i}$. Collecting these eigenvectors in a matrix $\mathbf{U}$, one obtains a similarity transformation that brings $\mathbf{A}$ to diagonal form:$\mathbf{U^{T}AU}= \Lambda$
---
#### Computation
Computing the eigenvalues using the characteristic polynomial is infeasible, because computation of this polynomial is prone to rounding errors and the following computation of roots is bad- [[Condition Number|conditioned]]. Therefore, use iterative algorithms, depending on the goal:
- Compute spectral radius (largest eigenvalue) only
- Direct [[Power Iteration|power iteration]]
- Compute specific eigenvalue
- Inverse [[Power Iteration|power iteration]]
- Compute all eigenvalues
- [[QR-Algorithm|QR-algorithm]]