level-set 水平集,拆分树,合并树,持续同调

Level-set(水平集)方法。这一方法由香港中文大学的Michael Yu Wang提出,认为实体的边界可由更高一维的函数的等高面来描述,在每次迭代时通过求解Jacobi扩散方程确定边界的演化过程。


1. Level-set(水平集), sub level-set (中文??), super level-set (超水平集????)
一个标量方程的某个真值a的逆象f-1(a)称为level set,表示了方程值等于a的所有空间点。
根据空间点的方程值是否大于等于a和小于等于a,
                满足f-1(a)的元素:         level_set, 水平集,本质就是等高线  ,即f(x) = a对应部分
                满足f-1(<=a)的元素:      sub-level set,本质就是低于等高线的部分,即f(x)<=a对应部分
                满足 f-1(>=a)的元素:     super-level set,本质就是高于等高线的部分,即f(x)>=a对应的部分
       
    因此包含极小值的等高线:内部的是 sub-level set, 外部是super-level set, 等高线本身是level set

2. Join tree(连接树)split tree(拆分树)

考虑横切(sweep) 函数f(x,y):Consider a sweep of the input function f(x,y) in increasing order of function value. The nature of topological change to the sub-level sets of f(x,y) when the sweep encounters a vertex determines the vertex type. 

       (1)点是规则点(regular): 下水平集的拓扑不发生改变(都在局部极小值的上升线上), The topology of the sub-level sets do not change. 

     (2)点是极小值(minimum):如果产生一个新的下水平集, A new sub-level set component is created. The scalar function shown in Figure 2(a) has three minima A,B, and C. 

     (3)点是划分鞍点(split saddle):如果两个下水平集合成一个。 Two sub-level set components merge into one. This is equivalent to one of the components being destroyed. The point D in Figure 2(a) represents one such saddle where the sub-level set components created at minima A and C merge into one. 

level-set 水平集,拆分树,合并树,持续同调

  拆分树(Split tree)  跟踪了下水平集(sub level set)的连接特性。在split tree中的每个节点都对应到一个极小值或者一个鞍点。拆分树可以追踪sub-level sets连通性的变化,其包含了两种关键顶点:
      (1)一个是minimum,表示创建了一个新的sub-level set组件,
      (2)另一个是split saddle,表示两个sub-level sets组件合并成一个。

(The split tree tracks the changes in the connectivity of sub-level sets of the input scalar function. Nodes of the split tree correspond to the set of minima and split saddles of f(x,y). Figure 2(b) shows the split tree corresponding to the input function shown in Figure 2(a). )

   合并树(Join Tree)定义类似Split Tree, 可以看成倒过来的拆分树(Split Tree), 它跟踪了上水平集(Super-Level set)的连接关系,合并树中的节点包含最大值和鞍点,最大值点处产生一个上升水平集,鞍点是上水平集合并的地方。

(The join tree of f(x) is defined similarly, and tracks the connectivity of the super-level sets of f(x,y). Nodes of the join tree correspond to the set of maxima and join saddles of f(x,y). A maximum is a vertex where a new super-level set component is created during the sweep of the input in decreasing order of function value, and a join saddle is a vertex where two super-level set components merge into one.)

  临界点:非规则点的极大值、极小值、鞍点,均称谓临界点。(A vertex that is not regular, such as a minimum or a saddle, is called critical. )

 Regular vertices are often inserted into the join/split tree as degree-2 nodes to obtain an augmented join tree / augmented split tree. 

  轮廓树(Contoure tree)  ===  Split tree + Join Tree. (轮廓树 = 拆分树  +  合并树)

3. 拓扑持续同调 (Topological Persistence)

   产生者:产生一个新拓扑组件的关键点,在Split Tree中为极小值,在Join Tree中为极大值。

   销毁者: 一个拓扑组成消失的关键点,不论哪种树,都是鞍点。

( A critical point is called a creator if a new component is created and a destroyer otherwise.)

   持久值:相邻生产者和销毁者之间的标量差值( It turns out that one can pair up each creator v1 uniquely with a destroyer v2 that destroys the component created at v1. The persistence value of v1 and v2 is defined as f(v2)− f(v1), which is intuitively the lifetime of the feature created at v1, and is thus a measure of the importance of v1 and v2. ),可以理解程v1点的生命值。

    持久性图,或者持续同调表(persistence diagram)是建立在关键点上的,其将一系列的点展示在一个2D平面上,点的x和y坐标分别表示了对应拓扑不见的开始和消亡的时间。

level-set 水平集,拆分树,合并树,持续同调

          第一刀,轮廓线只有一条!

     level-set 水平集,拆分树,合并树,持续同调

           第二刀,轮廓线变成2条!!!拓扑结构变化(一条曲线变成了两条),

持续同调表(Persistent Diagram)
Bottleneck Distance和
Wasserstein Distance。
持续同调(PH)为我们提供了一个寻找不需要降维就能精要的刻画数据全貌的方法。

==================================================================================

亏格:就是孔洞的个数。

水平集——那些我膜拜过的牛人2(https://blog.****.net/yutianxin123/article/details/69802364