The four vector norms that play signicant roles in the compressed sensing framework are the $\iota_0$ , the $\iota _1$ , $\iota_2$, and $\iota_\infty$ norms, denoted by $\|x\|_0$, $\|x\|_1$, $\|x\|_2$ and $\|x\|_\infty$ respectively.
Given a vector $x\in R^m$.
$\|x\|_0$ is equal tothe number of the non-zero elements in the vector $x$.
$\|x\|_1=\sum_{i=1}^{m}|x_i|$.
$\|x\|_2=\sqrt{x_1^2+x_2^2+...+x_m^2}$.
$\|x\|_\infty=\max_i |x_i|$.
The vector norm $\|x\|_p$ for $p=1, 2, 3, ...$ is defined as
$\|x\|_p=(\sum_{i=1}^m |x_i|^p )^\frac{1}{p}.$
The $p$-norm of vector $x$ is implemented as Norm[$x$, $p$], with the 2-norm being returned by Norm[$x$][1].
These norms have natural generalizations to matrices, inheriting many appealing properties from
the vector case. In particular, there is a parallel duality structure.
For two rectangular matrices $X\in R ^{m \times n}$, $X\in R ^{m \times n}$ .
Define the inner product as $:= \sum_{i=1}^m \sum_{j=1}^n X_{ij}Y_{ij}= \sqrt{Tr(X^T Y)}.$
The norm associated with this inner product is called the Frobenius (or Hilbert-Schmidt) norm
$\|X\|_F$ . The Frobenius norm is also equal to the Euclidean, or $\iota_2$, norm of the vector of singular
values, i.e.,
$\|X\|_F:= \sqrt{}=\sqrt{X^T X}=({\sum_{i=1}^r \delta_i^2})^{\frac{1}{2}}.$
The operator norm (or induced 2-norm) of a matrix is equal to its largest singular value (i.e., the $\iota_1$ norm of the singular values):
$\|X\|:= \delta_ 1(X).$
The nuclear norm of a matrix is equal to the sum of its singular values, i.e.,
$\|X\|_*:= \sum_{i=1}^r {\delta_i(X)}$,
and is alternatively known by several other names including the Schatten 1-norm, the Ky Fan r-norm, and the trace class norm. Since the singular values are all positive, the nuclear norm isequal to the $\iota_1$ norm of the vector of singular values. These three norms are related by the followinginequalities which hold for any matrix $X$ of rank at most $r$:
$\|X\| \leq \|X\|_F \leq \|X\|_{*} \leq \sqrt{r}\|X\|_F \leq r\|X\|.$
Table 1: A dictionary relating the concepts of cardinality and rank minimization[2].
parsimony concept
cardinality
rank
Hilbert Space norm
Euclidean
Frobenius
sparsity inducing norm
$\iota_1$
nuclear
dual norm
$\iota_{\infty}$
operator
norm additivity
disjoint support
orthogonal row and column spaces
convex optimization
linear programming
semidenite programming
The rank of a matrix $X$ is equals to the number of the singular values, and the singular values are all non-zeros.
The $\iota_0$ norm of a vector is the number of the non-zero elements in the vector.
In compressed sensing (in sparse representation), the original objective function is:
$\min \|x\|_0$ s.t. $Ax=b$.
In low rank representation, the original objective function is:
$\min rank(Z)$ s.t. $X=XZ$.
I'm curious to the relation ship between sparse representation and low rank representation.
References:
[2]Recht, Benjamin, Maryam Fazel, and Pablo A. Parrilo. "Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization."SIAM review52, no. 3 (2010): 471-501.
Apendix:
Vector and matrix normsSyntax
n = norm(X,2)
n = norm(X)
n = norm(X,1)
n = norm(X,Inf)
n = norm(X,'fro')
n = norm(v,p)
n = norm(v,Inf)
n = norm(v,-Inf)
Description
The norm function calculates several different types of matrix and vector norms. If the input is a vector or a matrix:
n = norm(X,2) returns the 2-norm of X.
n = norm(X) is the same as n = norm(X,2).
n = norm(X,1) returns the 1-norm of X.
n = norm(X,Inf) returns the infinity norm of X.
n = norm(X,'fro') returns the Frobenius norm of X.
In addition, when the input is a vector v:
n = norm(v,p) returns the p-norm of v. The p-norm is sum(abs(v).^p)^(1/p).
n = norm(v,Inf) returns the largest element of abs(v).
n = norm(v,-Inf) returns the smallest element of abs(v).
转载本文请联系原作者获取授权,同时请注明本文来自甄亮利科学网博客。
链接地址:http://blog.sciencenet.cn/blog-621576-652662.html
上一篇:人工智能与机器学习中一些问题的初步认识
下一篇:机器学习课程推荐