changed 3 years ago
Linked with GitHub

垂直子空間

Creative Commons License
This work by Jephian Lin is licensed under a Creative Commons Attribution 4.0 International License.

\(\newcommand{\trans}{^\top} \newcommand{\adj}{^{\rm adj}} \newcommand{\cof}{^{\rm cof}} \newcommand{\inp}[2]{\left\langle#1,#2\right\rangle} \newcommand{\dunion}{\mathbin{\dot\cup}} \newcommand{\bzero}{\mathbf{0}} \newcommand{\bone}{\mathbf{1}} \newcommand{\ba}{\mathbf{a}} \newcommand{\bb}{\mathbf{b}} \newcommand{\bc}{\mathbf{c}} \newcommand{\bd}{\mathbf{d}} \newcommand{\be}{\mathbf{e}} \newcommand{\bh}{\mathbf{h}} \newcommand{\bp}{\mathbf{p}} \newcommand{\bq}{\mathbf{q}} \newcommand{\br}{\mathbf{r}} \newcommand{\bx}{\mathbf{x}} \newcommand{\by}{\mathbf{y}} \newcommand{\bz}{\mathbf{z}} \newcommand{\bu}{\mathbf{u}} \newcommand{\bv}{\mathbf{v}} \newcommand{\bw}{\mathbf{w}} \newcommand{\tr}{\operatorname{tr}} \newcommand{\nul}{\operatorname{null}} \newcommand{\rank}{\operatorname{rank}} %\newcommand{\ker}{\operatorname{ker}} \newcommand{\range}{\operatorname{range}} \newcommand{\Col}{\operatorname{Col}} \newcommand{\Row}{\operatorname{Row}} \newcommand{\spec}{\operatorname{spec}} \newcommand{\vspan}{\operatorname{span}} \newcommand{\Vol}{\operatorname{Vol}} \newcommand{\sgn}{\operatorname{sgn}} \newcommand{\idmap}{\operatorname{id}} \newcommand{\am}{\operatorname{am}} \newcommand{\gm}{\operatorname{gm}} \newcommand{\mult}{\operatorname{mult}} \newcommand{\iner}{\operatorname{iner}}\)

from lingeo import random_int_list, random_good_matrix, kernel_matrix

Main idea

Let \(U\) and \(V\) be two subspaces under the same inner product space.
We say \(U\) and \(V\) are orthogonal if \(\langle {\bf u}, {\bf v} \rangle = 0\) for any \({\bf u}\in U\) and \({\bf v}\in V\).
Similarly, we say a collection of subspaces \(\{V_1,\ldots,V_k\}\) is orthogonal if they are pairwisely orthogonal.

If \(\{V_1,\ldots,V_k\}\) is orthogonal, then they are necssarily independent.
Therefore, we have the direct sum \(V_1 \oplus \cdots \oplus V_k\).
Suppose \(V = V_1 \oplus \cdots \oplus V_k\).
Then every vector \({\bf v}\in V\) can be uniquely written as \({\bf v} = {\bf v}_1 + \cdots + {\bf v}_k\) with \({\bf v}_i\in V_i\) for each \(i = 1,\ldots,k\).

Let \(A\) be an \(m\times n\) matrix.
We have seen several cases of mutually orthogonal subspaces related to \(A\).
With the new terminology, we may safely say:

  1. The subspaces \(\operatorname{Row}(A)\) and \(\operatorname{ker}(A)\) are orthogonal, and \(\mathbb{R}^n = \operatorname{Row}(A) \oplus \operatorname{ker}(A)\).
  2. The subspaces \(\operatorname{Col}(A)\) and \(\operatorname{ker}(A^\top)\) are orthogonal, and \(\mathbb{R}^m = \operatorname{Col}(A) \oplus \operatorname{ker}(A^\top)\).

Suppose \(V\) is a subspace in \(\mathbb{R}^n\).
We also have \(\mathbb{R}^n = V \oplus V^\perp\).

Side stories

  • projection matrix

Experiments

Exercise 1

執行以下程式碼。
已知 \(R\)\(A\) 的最簡階梯形式矩陣。

### code
set_random_seed(0)
print_ans = False
m,n,r = 2,4,2
A = random_good_matrix(m,n,r)
R = A.rref()
H = kernel_matrix(R)
c = random_int_list(2, r=3)
b = c[0]*R[0] + c[1]*H.transpose()[0]

print("A =")
show(A)
print("R =")
show(R)
print("b = ", b)

if print_ans:
    r = c[0]*R[0]
    h = c[1]*H.transpose()[0]
    print("r =", r)
    print("h =", h)
    print("|b|^2 =", b.norm()**2)
    print("|r|^2 + |h|^2 = %s + %s = %s"%(r.norm()**2, h.norm()**2, r.norm()**2 + h.norm()**2))
Exercise 1(a)

\({\bf b}\) 寫成 \({\bf r} + {\bf h}\)

其中 \({\bf r}\in\operatorname{Row}(A)\)\({\bf h}\in\operatorname{ker}(A)\)

[由李昌諺同學提供]
答:
seed(0) 時執行程式碼得到 \[A = \begin{bmatrix} 1 & 0 & 3 & 5 \\ 3 & 1 & 4 & 10 \end{bmatrix}, R = \begin{bmatrix} 1 & 0 & 3 & 5 \\ 0 & 1 & -5 & -5 \end{bmatrix}, {\bf b}=(-5,5,-5,-10). \]

\(A\) 的各列向量為 \({\bf r}_1,{\bf r}_2\) \(\operatorname{Row}(A)=\operatorname{span}(\{{\bf r}_1,{\bf r}_2\})\)

\(R\)\(A\) 的最簡階梯式,可知 \({x}_3,{x}_4\) 為自由變數。 \(x_3=1 , x_4=0\) 時可得到 \({\bf h}_1 = (-3,5,1,0).\) \(x_3=0 , x_4=1\) 時可得到 \({\bf h}_2 = (-5,5,0,1).\) 所以 \(\operatorname{ker}(A)=\operatorname{span}(\{{\bf h}_1,{\bf h}_2\})\)

假設 \({\bf r}=c_1{\bf r}_1 + c_2{\bf r}_2\)\({\bf h}=c_3{\bf h}_1 + c_4{\bf h}_2\) 使得 \({\bf b}={\bf r}+{\bf h}\) 可以把它轉換成矩陣形式 \[\begin{bmatrix} 1 & 3 & -3 & -5 \\ 0 & 1 & 5 & 5 \\ 3 & 4 & 1 & 0 \\ 5 & 10 & 0 & 1 \\ \end{bmatrix} \begin{bmatrix} c_1 \\ c_2 \\ c_3 \\ c_4 \\ \end{bmatrix}= \begin{bmatrix} -5 \\ 5 \\ -5 \\ -10 \\ \end{bmatrix}. \] 用高斯消去法計算增廣矩陣 \[\left[\begin{array}{cccc|c} 1 & 3 & -3 & -5 & -5\\ 0 & 1 & 5 & 5 & 5\\ 3 & 4 & 1 & 0 & -5\\ 5 & 10 & 0 & 1 & -10\\ \end{array}\right] \rightarrow \cdots \rightarrow \left[\begin{array}{cccc|c} 1 & 3 & -3 & -5 & -5\\ 0 & 1 & 5 & 5 & 5\\ 0 & 0 & 35 & 40 & 35\\ 0 & 0 & 0 & \frac{37}{7} & 0\\ \end{array}\right] \] 得到 \(c_1 = -2,c_2 = 0,c_3 = 1,c_4 = 0.\)

經過計算,得 \({\bf r}=c_1{\bf r}_1 + c_2{\bf r}_2=(-2,0,-6,-10)\in\operatorname{Row}(A)\)\({\bf h}=c_3{\bf h}_1 + c_4{\bf h}_2=(-3,5,1,0)\in\operatorname{ker}(A)\) 使得 \({\bf b}={\bf r}+{\bf h}\)

[由蔡程亦同學提供]
答:
seed=2
\[A = \begin{bmatrix} 1 & 2 & -4 & -3 \\ 1 & 2 & -4 & -2 \end{bmatrix}, R = \begin{bmatrix} 1 & 2 & -4 & 0 \\ 0 & 0 & 0 & 1 \end{bmatrix}, {\bf b}=(3,1,-4,0). \]

\(A\) 的各列向量為 \({\bf r}_1,{\bf r}_2\) \(\operatorname{Row}(A)=\operatorname{span}(\{{\bf r}_1,{\bf r}_2\})\)

\(R\) 可知 \({x}_2,{x}_3\) 為自由變數。 \(x_2=1 , x_3=0\) 可得到 \({\bf h}_1 = (-2,1,0,0).\) \(x_2=0 , x_3=1\) 可得到 \({\bf h}_2 = (4,0,1,0).\) 所以 \(\operatorname{ker}(A)=\operatorname{span}(\{{\bf h}_1,{\bf h}_2\})\) 解方程式\(x(1,2,-4,-3)+y(1,2,-4,-2)+z(-2,1,0,0)+w(4,0,1,0)=(3,1,-4,0)\) 計算得 \(x=-2,y=3,z=-1,w=0\) \({\bf r}=(1,2,-4,0),{\bf h}=(2,-1,0,0)\) 使得 \({\bf b}={\bf r}+{\bf h}\)

Exercise 1(b)

證驗 \({\bf r}\)\({\bf h}\) 垂直﹐
而且 \(\|{\bf b}\|^2 = \|{\bf r}\|^2 + \|{\bf h}\|^2\)

[由李昌諺同學提供]
答(seed(0)):
答:
計算 \[\begin{aligned} \langle {\bf r}, {\bf h}\rangle &=(-2,0,-6,10) \cdot (-3,5,1,0) \\ &= 6+0-6+0 \\ &= 0 \\ \end{aligned} \] \(\|{\bf b}\|^2 = 25+25+25+100 = 175\)
\(\|{\bf r}\|^2 = 4+0+36+100 = 140\)
\(\|{\bf h}\|^2 = 9+25+1+0 = 35\)
\(\|{\bf b}\|^2 = \|{\bf r}\|^2 + \|{\bf h}\|^2\)

[由蔡程亦同學提供]
seed=2時)
計算 \[\begin{aligned} \langle {\bf r}, {\bf h}\rangle &=(1,2,-4,0) \cdot (2,-1,0,0) \\ &= 0. \\ \end{aligned} \] 所以\({\bf r}\)\({\bf h}\) 垂直。 \(\|{\bf b}\|^2 = 26\)\(\|{\bf r}\|^2 = 21\)\(\|{\bf h}\|^2 = 5\) \(\|{\bf b}\|^2 = \|{\bf r}\|^2 + \|{\bf h}\|^2\)

Exercise 1©

因為每一個 中的向量都可以分解成 \({\bf r}\in\operatorname{Row}(A)\)\({\bf h}\in\operatorname{ker}(A)\) 中的向量相加。
說明對任何 \(m\times n\) 矩陣都有
\[\{A{\bf x}: {\bf x}\in\mathbb{R}^n \} = \{ A{\bf r} : {\bf r}\in\operatorname{Row}(A)\}. \]

Exercises

Exercise 2

\(S = \{V_1,\ldots,V_k\}\) 為一群子空間。
證明若 \(S\) 是垂直的集合﹐則 \(S\) 線性獨立。

Exercise 3

\(S = \{V_1, V_2, V_3\}\) 垂直。
\(V = V_1 \oplus V_2 \oplus V_3\)
\(P\)\(V\) 的投影矩陣、
\(P_1,P_2,P_3\) 分別為 \(V_1,V_2,V_3\) 的投影矩陣。

Exercise 3(a)

說明 \(P_1P_2 = P_2P_1\)

Exercise 3(b)

說明 \(P = P_1 + P_2 + P_3\)

Exercise 3©

\(V = \mathbb{R}^n\) 為全空間。
說明 \(I_n = P_1 + P_2 + P_3\)

Exercise 4

依照步驟證明以下敘述等價:

  1. \(P\) 是某個空間的投影矩陣。
  2. \(P\) 對稱、而且 \(P = P^2\)
Exercise 4(a)

證明若 \(P\) 是一個投影矩陣﹐
\(P\)\(\operatorname{Col}(P)\) 的投影矩陣。
因此如果 \({\bf u}\in\operatorname{Col}(P)\)\(P{\bf u} = {\bf u}\)
如果 \({\bf u}\in\operatorname{ker}(P^\top)\)\(P{\bf u} = {\bf 0}\)
同時每個向量都可以分成 \({\bf u} = P{\bf u} + (I - P){\bf u}\)
使得 \(P{\bf u}\in\operatorname{Col}(P)\)\((I - P){\bf u}\in\operatorname{ker}(P^\top)\)

藉由這些性質說明如果條件一成立則條件二成立。

Exercise 4(b)

\(P\) 對稱且 \(P = P^2\)
說明 \(\mathbb{R}^n = \operatorname{Col}(P) \oplus \operatorname{ker}(P)\)
如果 \({\bf u}\in\operatorname{Col}(P)\)\(P{\bf u} = {\bf u}\)
如果 \({\bf u}\in\operatorname{ker}(P)\)\(P{\bf u} = {\bf 0}\)

藉由這些性質說明如果條件二成立則條件一成立。

Exercise 5

證明若 \(V = V_1 \oplus \cdots \oplus V_k\)
則每一個向量 \({\bf v}\in V\) 都可以被寫成 \({\bf v} = {\bf v}_1 + \cdots + {\bf v}_k\)
使得對每一個 \(i = 1,\ldots,k\) 都有 \({\bf v}_i\in V_i\)
而且這種寫法唯一。

Exercise 6

利用垂直空間分解母空間的現象在其它向量空間也很常見。

Exercise 6(a)

\(U = \mathcal{M}_{n\times n}\) 為一向量空間﹐搭配內積 \(\langle A,B\rangle = \operatorname{tr}(B^\top A)\)
\(V\)\(U\) 中所有對稱矩陣的集合、
\(W\)\(U\) 中所有反對稱矩陣的集合。
說明 \(V\)\(W\) 垂直且 \(U = V \oplus W\)

Exercise 6(a)

\(U = \mathcal{P}_{d}\) 為一向量空間﹐搭配內積
\[\langle a_0 + a_1x + \cdots + a_dx^d, b_0 + b_1x + \cdots + b_dx^d \rangle = a_0b_0 + a_1b_1 + \cdots + a_db_d.\]
\(V\)\(U\) 中所有偶函數的集合、
\(W\)\(U\) 中所有奇函數的集合。
說明 \(V\)\(W\) 垂直且 \(U = V \oplus W\)

Select a repo