Grassmann流形 — 用于子空间比较【理论】

SIHENG HUANG

已于 2022-07-07 17:19:11 修改

阅读量3.7k

点赞数 1

CC 4.0 BY-SA版权

文章标签：线性代数拓扑学

于 2022-07-07 16:51:01 首次发布

本文链接：https://blog.csdn.net/Alphonse_Huang/article/details/125662391

本文介绍了Grassmann流形的概念，它用于比较不同维度的子空间。内容包括子空间的定义，Grassmann流形的详细描述，以及在该流形中测量距离的方法，如投影距离、Binet-Cauchy距离、最大相关距离、最小相关距离和Procrustes距离。此外，文章还提到了主要角度在这些距离度量中的关键作用。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Grassmann流形 — 用于子空间比较【理论】

1. 子空间

1.1 向量空间/线性空间

向量空间的定义

若一组向量组成的集合 $V$ 在数域 $P$ 上满足加法及乘法运算，则这组向量组成的集合 $V$ 被称为一个向量空间。

1.2 向量子空间/线性子空间

向量子空间定义

若集合 $W$ 是数域 $P$ 上向量空间 $V$ 的一个非空子集，且满足加法及乘法运算，则 $W$ 为 $V$ 的一个向量子空间。

2. Grassmann Manifolds

Mainfolds
Manifolds Described By Wikipedia

Definition

The Grassmannian manifold $\mathcal{G}(m, D)$ refers to the $D - d i m e n s i o n a l$ formed by all $m - d i m e n s i o n a l$ embedded into a $D - d i m e n s i o n a l$ real(or complex) Euclidean space.

An element of $\mathcal{G}(m, D)$ can be represented by an orthonormal matrix $Y$ of size $D$ by $m$ such that $Y ^{'}Y = I_m$ , where $I_m$ is the $m$ by $m$ identity matrix.

For example, $Y$ can be the $m$ basis vectors of a set of pictures in $R^D$ .

However, the matrix representation of a point in $\mathcal{G}(m, D)$ is not unique:

Two matrices $Y_1$ and $Y_2$ are considered the same if and only if $span(Y_1) = span(Y_2)$ , where $s p a n (Y)$ denotes the subspace spanned by the column vectors of $Y$ .
$span(Y_1) = span(Y_2)$ if and only if $Y_1R_1 = Y_2R_2$ for some $R_1, R_2 ∈ O(m)$ .

With this understanding, we will often use the notation $Y$ when we actually mean its equivalence class $s p a n (Y)$ , and use $Y_1 = Y_2$ when we mean $span(Y_1) = span(Y_2)$ , for simplicity.

Formally, the Riemannian distance between two subspaces is the length of the shortest geodesic connecting the two points on the Grassmann manifold.

However, there is a more intuitive and computationally efficient way of defining the distances using the principal angles.

3. Principle Angle

与定义向量之间的角度类似，也可以定义空间（或子空间）之间的角度。在这种情况下，这些角称为Principle Angles between Subspaces(PABS).

Traditionally, PABS are introduced and used via their cosines. The tangents of PABS have attracted relatively less attention, but are important for analysis of convergence of subspace iterations for eigenvalue problems.

Recall