General norm sparsification

The material in this lecture is taken from Sparsifying sums of norms.

Suppose that $N_1,\ldots,N_m : \R^n \to \R_+$ are semi-norms on $\R^n$, and define the semi-norm given by their $\ell_p$ sum, for $p > 1$:

$$\begin{equation}\label{eq:ns} N(x) \mathrel{\mathop:}=\left(N_1(x)^p + \cdots + N_m(x)^p\right)^{1/p}\,.\end{equation} $$

Our goal will be to sparification where the terms are $f_i(x) \mathrel{\mathop:}=N_i(x)^p$ in our usual notation, and $F(x) = N(x)^p$.

A semi-norm $N$ is nonnegative and satisfies $N(\lambda x) = |\lambda| N(x)$ and $N(x+y) \leqslant N(x)+N(y)$ for all $\lambda\in \R$, $x,y \in \R^n$, though possibly $N(x)=0$ for $x \neq 0$.

Note that, equivalently, a semi-norm is a homogeneous function: $N(\lambda x) = |\lambda| N(x)$ such that $N$ is also convex.

Examples

Importance scores