Norms
Definitions
Type | Definition |
---|---|
Vector Norms | A vector norm on \(\mathbb{R}^n\) is a function \(f: \ \mathbb{R}^n \rightarrow \mathbb{R}\) that satisfies the following three properties:
|
Matrix Norms | A function \(f: \ \mathbb{R}^{m \times n} \rightarrow \mathbb{R}\) is a matrix norm if the following holds:
|
Common Norms
Norm | Vector | Matrix |
---|---|---|
1-norm | \(\| \mathbf{x} \|_1 = \| x_1 \| + \cdots + \| x_n \|\) | |
2-norm (\(l_2\)) | \(\| \mathbf{x} \|_2 = \left( \| x_1 \|^2 + \cdots + \| x_n \|^2 \right)^{1/2} = \left(\mathbf{x}^T \mathbf{x} \right)^{1/2}\) | If \(\mathbf{A} \in \mathbb{R}^{m \times n}\), then there exists a unit 2-norm n-vector \(\mathbf{z}\) such that: \(\mathbf{A}^T \mathbf{A} \mathbf{z} = \mu^2 \mathbf{z}\), where \(\mu = \|\mathbf{A}\|_2\). This implies that \(\|\mathbf{A}\|^2_2\) is a zero of \(p(\lambda) = \text{det} \left( \mathbf{A}^T \mathbf{A} - \lambda \mathbf{I} \right)\). In particular: \(\|\mathbf{A}\|_2 = \sqrt{\lambda_{max} \left( \mathbf{A}^T \mathbf{A} \right)}\). |
\(p\)-norm | \(\| \mathbf{x} \|_p = \left( \| x_1 \|^p + \cdots + \| x_n \|^p \right)^{1/p}, \quad p \geq 1\) | \(\|\mathbf{A}\|_p = \sup_{x \neq 0} \frac{\|\mathbf{A} \mathbf{x}\|_p}{\|\mathbf{x}\|_p}\) |
Infinity norm | \(\| \mathbf{x} \|_{\infty} = \max_{1 \leq i \leq n} \| x_i \|\) | |
Frobenius norm | \(\|\mathbf{A}\|_F = \sqrt{\sum^m_{i = 1} \sum^n_{j = 1} \|A_{ij}\|^2}\) |
Properties
Property | Vector | Matrix |
---|---|---|
Cauchy-Schwarz Inequality | \(\| \mathbf{x}^T \mathbf{y} \| \leq \| \mathbf{x}\|_2 \| \mathbf{y} \|_2\) | |
Norm Preservation | 2-norm is preserved under orthogonal transformation. Let \(\mathbf{Q} \in \mathbb{R}^{n \times n}\) be orthogonal matrix and \(\mathbf{x} \in \mathbb{R}^n\). Then: \(\| \mathbf{Q} \mathbf{x} \|^2_2 = \left( \mathbf{Q} \mathbf{x}\right)^T \left( \mathbf{Q} \mathbf{x}\right) = \mathbf{x}^T \mathbf{x} = \| \mathbf{x} \|^2_2\) |
|
Absolute and Relative Errors | Let \(\hat{\mathbf{x}}\) be an approximation of \(\mathbf{x} \in \mathbb{R}^n\). For a given vector norm \(\| \cdot \|\), the absolute error in \(\hat{\mathbf{x}}\) is defined as: \(\boldsymbol{\epsilon}_{abs} = \|\hat{\mathbf{x}} - \mathbf{x} \|\). If \(\mathbf{x} \neq \mathbf{0}\), then the relative error in \(\hat{\mathbf{x}}\) is defined as: \(\boldsymbol{\epsilon}_{rel} = \frac{\| \hat{\mathbf{x}} - \mathbf{x} \|}{\| \mathbf{x}\|}\). |
|
Sequence Convergence | A sequence \(\left\{ \mathbf{x}^{(k)} \right\}\) of \(n\)-vectors converges to \(\mathbf{x}\) if: \(\lim_{k \rightarrow \infty} \|\mathbf{x}^{(k)} - \mathbf{x} \| = 0.\) |