正交补与极小化问题

目录

1 正交补

设\(U\)是\(V\)的子集,则\(U\)的正交补(记为\(U^{\perp}\))是由\(V\)中与\(U\)的每个向量都正交的那些向量组成的集合:

\begin{equation} \label{eq:1} U^{\perp} = \{v\in V: \forall~u\in U, \langle v,u \rangle = 0 \} \end{equation}

若\(U\)是\(\mathbf{R}^{3}\)中的直线,则\(U^{\perp}\)是垂直于\(U\)且包含原点的平面。若\(U\)是\(\mathbf{R}^{3}\)中的平面,则\(U^{\perp}\)是垂直于\(U\)且包含原点的直线。

  1. 若\(U\)是\(V\)的子集,则\(U^{\perp}\)是\(V\)的子空间。
  2. \(\{0\}^{\perp} = V\)
  3. \(V^{\perp} = \{0\}\)
  4. 若\(U\)是\(V\)的子集,则\(U\cap U^{\perp} = \{0\}\)
  5. 若\(U\)和\(W\)均为\(V\)的子集且\(U\subset W\),则\(W^{\perp} \subset U^{\perp}\)
  1. 设\(U\)是\(V\)的子集,则对每个\(u\in U\)均有\( \langle 0,u \rangle = 0 \),于是\(0 \in U^{\perp}\). 设\(v,w\in U^{\perp}\),若\(u\in U\),则:\[ \langle v+w,u \rangle = \langle v,u \rangle + \langle w,u \rangle = 0\] 因此\(v+w\in U^{\perp}\),所以在加法下\(U^{\perp}\)是封闭的。

类似的对于\(v\in U^{\perp} \),有\( \langle \lambda v,u \rangle = \lambda \langle v,u \rangle = 0 \)这说明\(U^{\perp}\)在标量乘法下是封闭的。

所以\(U^{\perp}\)是\(V\)的子空间。

  1. \(\forall v\in V\), 均有:\(\langle v,0 \rangle = 0\) ,所以\( \{0\}^{\perp} = V \)
  2. 假设\(v\in V^{\perp}\),则\( \langle v,v \rangle = 0 \),则\(v=0\),所以\(V^{\perp} = 0\)
  3. 假设\(v\in U\cap U^{\perp}\),则有\( \langle v,v \rangle = 0\),则\(v = 0\),所以\(U \cap U^{\perp} = \{0\}\)
  4. 设\(U,W\)均为\(V\)的子集,则对于\(v\in W^{\perp}\),说明对于\(\forall~u\in W \),都有\( \langle v,u \rangle = 0 \),这表明对于每个\(u\in U\),都有\( \langle v,u \rangle = 0 \),所以\( v \in U^{\perp} \),所以有\(W^{\perp} \subset U^{\perp}\)

若\(U,W\)均为\(V\)的子空间,并且\(V\)中的每个元素都可以唯一的写成\(U\)中的一个向量与\(W\)中的一个向量的和,则\(V\)是\(U\)和\(W\)的直和,记为\(V = U\oplus W \)

设\(U\)是\(V\)的有限维子空间,则\(V = U \oplus U^{\perp}\)

首先证明: \[V = U + U^{\perp}\] 假设\(v\in V\),并设\(e_{1},\ldots ,e_{m}\)是\(U\)的规范正交基,则:

\begin{equation} \label{eq:2} v = \underbrace{\langle v,e_{1} \rangle e_{1} + \ldots + \langle v,e_{m} \rangle e_{m}}_{u} +\underbrace{ v -\langle v,e_{1} \rangle e_{1} - \ldots - \langle v,e_{m} \rangle e_{m}}_{w} \end{equation}

显然\(u\in U\),因为\(e_{1},\ldots ,e_{m}\)是\(U\)的一个规范正交基,所以对每个\(j=1,\ldots ,m\)均有:

\begin{equation} \label{eq:3} \langle w,e_{j} \rangle = \langle v,e_{j} \rangle - \langle v,e_{j} \rangle = 0 \end{equation}

所以\(w\)正交于\(\mathrm{span}(e_{1},\ldots ,e_{m})\)中的每个向量。也就是说\(w\in U^{\perp}\),于是\(v = u + w\)其中\(u\in U,w\in U^{\perp}\) 另外因为\(U\cap U^{\perp} = \{0\}\),所以\[V= U \oplus U^{\perp}\]

设\(V\)是有限维的且\(U\)是\(V\)的子空间,则\(\dim U^{\perp} = \dim V - \dim U\)

\(U\)是\(V\)的有限维子空间,则\(U= (U^{\perp})^{\perp}\)

首先证明\(U\subset (U^{\perp})^{\perp}\)。设\(u\in U\),则对每个\(v\in U^{\perp}\),均有\( \langle v,u \rangle = 0 \).因为 \(u\)正交与\(U^{\perp}\)中的向量,所以\(u\in (U^{\perp})^{\perp}\)

然后我证明另一个方向。设\(v\in (U^{\perp})^{\perp}\)。设\(v\in (U^{\perp})^{\perp}\)令\(v=u+w\),其中\(u\in U\),且\(w\in U^{\perp}\),从而\(v-u = w\in U^{\perp}\)。因为\(v\in (U^{\perp})^{\perp}\)且\(u\in (U^{\perp})^{\perp}\),所以\(v-u\in (U^{\perp})^{\perp}\),所以\(v-u\in U^{\perp} \cap (U^{\perp})^{\perp}\)。这表明\(v-u\)与自身正交。从而\(v-u=0\),即\(v=u\),于是\(v\in U\),因此\( (U^{\perp})^{\perp} \subset U\)

设\(U\)是\(V\)的有限维子空间。定义\(V\)到\(U\)上的正交投影为如下算子\(P_{U} \in \mathcal{L}(V)\):对\(v\in V\),将其写成\(v= u+w\),其中\(u\in U,w\in U^{\perp}\),则\(P_{U}v = u\)

直和分解\(V = U \oplus U^{\perp}\)表明每个\(v\in V\)可唯一的写成\(v = u + w\),其中\(u\in U\)且\(w\in U^{\perp}\),于是\(P_{U}v\)定义合理。

设\(x\in V\),\(x\neq 0\)且\(U = \mathrm{span}(x)\),证明对每个\(x\in V\)均有:

\begin{equation} \label{eq:4} P_{U}v = \frac{ \langle v,x \rangle }{ \| x \|^{2} } x \end{equation}

已知\(\forall~v\in V\)都可以写为:

\begin{equation} \label{eq:5} v= \frac{ \langle v,x \rangle }{ \| x \|^{2} } x + (v - \frac{ \langle v,x \rangle }{ \| x \|^{2} } x) \end{equation}

上式第一项和第二项是互相垂直的。第一项属于\(\mathrm{span}(x)\),第二项正交与\(x\),从而第二项属于\(U^{\perp}\)。所以\(P_{U}v\)等于上式右端第一项。

接下来给出几条正交投影\(P_{U}\)的性质: 设\(U\)是\(V\)的有限维子空间且\(v\in V\),则:

  1. \(P_{U}\in \mathcal{L}(V) \)
  2. \(\forall ~u, P_{U}u = u\)
  3. \(\forall ~w\in U^{\perp},P_{U}w = 0\)
  4. \(\mathrm{range} P_{U} = U\)
  5. \(\mathrm{null}P_{U} = U^{\perp}\)
  6. \(v-P_{U}v \in U^{\perp} \)
  7. \((P_{U})^{2} = P_{U}\)
  8. \( \| (P_{U})v \| \leq \| v \| \)
  9. 对\(U\)的每个规范正交基\(e_{1},\ldots ,e_{n}\)均有\(P_{U}v = \langle v,e_{1} \rangle e_{1} + \ldots + \langle v,e_{m} \rangle e_{m} \)

为了证明\(P_{U}\)是\(V\)上的线性映射,设\(v_{1},v_{2}\in V\),设:

\begin{equation} \label{eq:6} v_{1} = u_{1} + w_{1}, v_{2} = u_{2} + w_{2} \end{equation}

其中\(u_{1},u_{2}\in U\),\(w_{1},w_{2}\in U^{\perp}\),则\(P_{U}v_{1} = u_{1},P_{U}v_{2} = u_{2}\),从而: \[v_{1} + v_{2} = u_{1} + u_{2} + w_{1} + w_{2}\] 其中\(u_{1} + u_{2}\in U\),且\(w_{1} + w_{2}\in U^{\perp}\),所以\(P_{U}(v_{1}+ v_{2}) = u_{1} + u_{2} = P_{U}v_{1} + P_{U}v_{2} \) 类似的有,设\(\lambda\in \mathbf{F}\),若\(v= u + w\),其中\(u\in U\)且\(w\in U^{\perp}\),则\(\lambda v = \lambda u + \lambda w\),其中\(\lambda u\in U, \lambda w\in U^{\perp}\),于是\(P_{U}(\lambda v) = \lambda u = \lambda P_{U}v\),因此\(P_{U}\)是\(V\)到\(V\)的线性映射。

设\(u\in U\),则\(u = u + 0\),则其中\(u\in U,0\in U^{\perp}\),所以\(P_{U}u = u\)

设\(w\in U^{\perp}\),则\(w = 0 + w\),则其中\(0\in U,w\in U^{\perp}\),所以\(P_{U}w = 0\)

由\(P_{U}\)的定义可知\(\mathrm{range}(P_{U})\subset U\)。由上面第二步可知\(U\subset \mathrm{range}P_{U}\),于是\( \mathrm{range}P_{U} = U\)

因为\(U^{\perp}\subset \mathrm{null}P_{U}\)。另外,\(\forall~v\in \mathrm{null}P_{u}\)则\(v= 0+ v\),其中\(0\in U,v\in U^{\perp}\),因此\(\mathrm{null}P_{U}\subset U^{\perp}\)

若\(v = u + w\),其中\(u\in U\),且\(w\in U^{\perp}\),则: \[v- P_{U}v= v-u = w\in U^{\perp} \]

若\(v = u+w\),其中\(u\in U\),且\(w\in U^{\perp}\),则:

\begin{equation} \label{eq:7} (P_{U})^{2}v = P_{U}(P_{U}v) = P_{U}u= u= P_{U}v \end{equation}

所以\(P_{U}^{2} = P_{U}\)

若\(v = u+w\),其中\(u\in U\),且\(w\in U^{\perp}\),则: \[ \| P_{U}v \|^{2} = \| u \|^{2} \leq \| u \|^{2} + \| w \|^{2} = \| v \|^{2} \]

2 极小化问题

经常会遇到这样的问题:给定\(V\)的子空间\(U\)和点\(v\in V\),求点\(u\in U\)使得\( \| v-u \| \)最小。通过正交投影可以完美解决这个问题。

设\(U\)是\(V\)的最小子空间,\(v\in V\),且\(u\in U\),则:

\begin{equation} \label{eq:8} \| v - P_{U}v \| \leq \| v - u \| \end{equation}

当且仅当\(P_{U}v = u\)时等号成立。

我们有:

\begin{eqnarray} \label{eq:9} \| v- P_{U}v \|^{2} &\leq& \| v- P_{U}v \|^{2} + \| P_{U}v - u \|^{2} &=& \| v - P_{U}v + P_{U}v - u \|^{2} \\ &=& \| v - u \|^{2} \end{eqnarray}

上式的第一个不等号成立是因为\( \| P_{U}v - u \|^{2} \)是一个非负实数,第二个等号成立是因为勾股定理,第三个等式成立是简单的消元计算。把上式两端开平方即可。

上式的证明表明\(P_{U}v\)是\(U\)中离\(v\)最近的点。之前我们又有\(P_{U}v = \langle v,e_{1} \rangle e_{1} + \ldots + \langle v,e_{m} \rangle \),其中\(e_{1},\ldots ,e_{m}\)是\(U\)的规范正交基。