The eigenvalue equation describes how a matrix transforms certain special vectors — scaling them without changing their direction. These appear throughout data analysis: in PCA, graph clustering, spectral decomposition, and more.
$A \in \mathbb{R}^{n \times n}$ is square. $\lambda \in \mathbb{R}$ is the eigenvalue (scalar). $\vec{v} \in \mathbb{R}^n$ is the eigenvector. The eigenvector is not changed in direction by the linear transformation $A$ — only its magnitude is scaled by $\lambda$.
Rearranging the eigenvalue equation: $(A - \lambda I)\vec{v} = \vec{0}$. For non-trivial $\vec{v} \neq \vec{0}$ to exist, the matrix $A - \lambda I$ must be singular, i.e. its determinant is zero. This gives the characteristic equation. For a $2 \times 2$ matrix $A = \begin{bmatrix}a&b\\c&d\end{bmatrix}$:
These two facts are the most powerful shortcuts for finding unknown eigenvalues. If you know some eigenvalues and the trace/det, you can solve for the rest without finding the characteristic polynomial.
The eigenvectors are preserved for $A^{-1}$, $A + cI$, and $A^k$ — only the eigenvalues change. For $A^T$, the eigenvalues are the same but the eigenvectors are generally different.
This is the most tested fact on every CISC 271 exam. For a triangular matrix, you can read the eigenvalues directly off the diagonal without any calculation. This applies even if the matrix has complex off-diagonal entries.
where $V = [\vec{v}_1 \; \vec{v}_2 \; \cdots \; \vec{v}_n]$ is the matrix of eigenvectors (columns) and $\Lambda = \text{diag}(\lambda_1, \ldots, \lambda_n)$. This holds if and only if $V$ is invertible (eigenvectors are linearly independent). A sufficient condition: all eigenvalues are distinct.
Similar matrices share the same eigenvalues (but generally different eigenvectors). Their traces and determinants are also equal.
Proof of orthogonality: for distinct $\lambda_1 \neq \lambda_2$, consider $(B\vec{v}_1) \cdot \vec{v}_2$. Compute this two ways: (1) $= \lambda_1(\vec{v}_1 \cdot \vec{v}_2)$ and (2) $= \lambda_2(\vec{v}_1 \cdot \vec{v}_2)$ (using $B^T = B$). Since $\lambda_1 \neq \lambda_2$, we must have $\vec{v}_1 \cdot \vec{v}_2 = 0$.
For any real symmetric matrix $B$: $Q$ is orthogonal (columns are unit eigenvectors of $B$), $\Lambda = \text{diag}(\lambda_1, \ldots, \lambda_n)$ with real eigenvalues. Since $Q^{-1} = Q^T$, this is a special case of diagonalization. Can also be written as a sum of rank-1 matrices:
This outer-product expansion is the foundation of PCA and the SVD.
The quadratic form $\vec{u}^T B \vec{u}$ is the energy test. This can be proven using the spectral decomposition: $\vec{u}^T B \vec{u} = \sum_j \lambda_j \alpha_j^2$ where $\alpha_j = \vec{q}_j^T \vec{u}$.
Any matrix of the form $M^TM$ is symmetric positive semidefinite. It is positive definite (PD) if and only if the columns of $M$ are linearly independent (i.e. $M$ is full rank). This is because $\vec{u}^T(M^TM)\vec{u} = \|M\vec{u}\|^2 \geq 0$, and equals zero iff $M\vec{u} = \vec{0}$, iff $\vec{u}$ is in the null space of $M$.
The vector $\vec{w}$ is always an eigenvector with eigenvalue $\lambda = 1 + \|\vec{w}\|^2/c$. Any vector orthogonal to $\vec{w}$ is an eigenvector with eigenvalue $\lambda = 1$. This follows from: $A_3\vec{w} = \vec{w} + \frac{1}{c}\vec{w}(\vec{w}^T\vec{w}) = (1 + \|\vec{w}\|^2/c)\vec{w}$.
For diagonalizable $A = V\Lambda V^{-1}$: $A^2 = V\Lambda^2 V^{-1}$, etc. The eigenvalues are raised to the $k$th power; eigenvectors are unchanged. This implies a "square root" matrix $A^{1/2} = V\Lambda^{1/2}V^{-1}$ exists when all eigenvalues are non-negative.