Convex Optimization : some definitions

最新推荐文章于 2021-04-11 18:03:55 发布

tangwing

最新推荐文章于 2021-04-11 18:03:55 发布

阅读量478

点赞数

分类专栏： # Optimization

本文链接：https://blog.csdn.net/tangwing/article/details/86633842

版权

Optimization 专栏收录该内容

7 篇文章 0 订阅

订阅专栏

Linear relation: Ax

Affine relation: Ax+b

Affine Set

contains the line through any two distinct points in the set: $x=ax_1+(1-a)x_2$
can be expressed as the solution set of a system of linear equations

Convex Set

If two points are in the set, then the line segment as well.
Affine set with 0<a<1

Convex Combinations

Consider any number of points, with their coefficient being positive and summing up to 1

Convex Hull

All convex combinations of any points from the set.

Conic (non-negative) combination

Like convex combinations but without the constraint of "summing up to 1"
Formed with any points together with the origin.

Conic Hull

All conic combinations of any points from the set.

Hyperplane and halfspace

hyperplanes are affine and convex:
halfspaces are convex

Polyhedron/polytope

solution space of a system of inequalities like in a linear program
intersections of a finit set of hyperplanes and halfspaces

Operations that preserves covexity

S is convex set -> f(S) too, f^-1(S) as well. with f an affine function
scale, translation, projection...
perspective (projection) function and its inverse: $f:R^{n+1}\rightarrow R^{n}$ . $x^*=[\vec{x},t]\in R^{n+1}, f(x^*)=\{\vec{x}/t|t>0\}.$
linear fractional function: a fraction of linear functions. $f:R^{m}\rightarrow R^{n}. f(x)=\frac{Ax+b}{Cx+d}$ ,with the denominator >0.
can be seen as a streching of vision, or the projection of a 3d object onto a camera. x=[x1,x2]. If the object is convex then the result is convex too.

Generalized Inequalities

A proper cone is closed (contains boundary), solid (not empty not a ray) and pointed(not a line).

define $x\preceq _Ky \Leftrightarrow y-x\in K$ , reads "x less than y with respect to proper cone K". Remember K somehow define a 'positive' space, consider an orthant. The coordinates y-x must still in the same orthant.

componentwise inequality: $K=R_+^n$ , x<y == xi<yi
matrix inequality: $K=S_+^n, X\preceq Y \Leftrightarrow Y-X \in S_+^n$ ,(w.r.t positive semidefinite)
$\preceq_K is\ a\ generalized\ \leq$ , but the latter defines also a "linear ordering", while the former doesn't, where comes the difference between Minimum (less than anyone else) and Minimal(less than anyone else that are comparable).