实变函数3

\(L^p\) Space

\(L^p\)空间

(Definition):对可测集\(E\)

  1. \(1\leq p < \infty\),定义\(||f||_{L^p(E)} := (\int_{|E|} |f|^p dx)^{1/p}\)\(f\)\(L^p\)范数

  2. \(p=\infty\),定义\(||f||_{L^{\infty}(E)} := \inf\{M:\exists F \subset E, m(F)=0, \sup_{E\setminus F} |f| \} \leq M\)

    也就是说,\(|f|\)几乎处处小于等于\(||f||_{L^{\infty}(E)}\)

  3. \(1\leq p < \infty\),定义\(E\)上的\(L^p\)可积函数空间为\(L^p(E) := \{f:E \to [-\infty, +\infty],\;\int_E |f|^p < \infty,\; f\;measurable\}\)

  4. \(p=\infty\),定义\(E\)上的\(L^{\infty}\)可积函数空间为\(L^{\infty}(E):=\{f:E \to [-\infty, +\infty], ||f||_{L^{\infty}(E)}<\infty\}\)

不难知道,对\(1 \leq p \leq +\infty\)\(L^p\)是一个线性空间,除了线性空间外,\(L^p\)空间还有一些优美的性质

如果定义\(L^p\)空间上的距离函数\(d(f,g) := ||f-g||_{L^p}\),那么\((L^p, d)\)构成度量空间


  • (Minkowski):设\(E\)可测,那么\(||f+g||_{L^p} \leq ||f||_{L^p} + ||g||_{L^p}\)
  • (Holder's inequality):定义\(p\in[1,\infty]\)的共轭指标为\(p'\),满足\(1/p+1/p' = 1\)\(||fg||_{L^1} \leq ||f||_{L^p} \cdot ||g||_{L^{p'}}\)

上述证明非常经典,这里省略

(Corollary,加权的Holder's inequality):令\(1 \leq p_1 \leq p \leq p_2 \leq \infty\)\(1/p=\theta/p_1+(1-\theta)/p_2\),那么\(||f||_{L^p} \leq ||f||_{L^{p_1}}^{\theta} \cdot ||f||_{L^{p_2}}^{1-\theta}\)

(Corollary 1):如果\(f\in L^{p_1}\)\(f\in L^{p_2}\)\(p_1 \leq p \leq p_2\),那么\(f \in L^p\)

(Corollary 2):如果\(f\in L^p\)\(m(E) <\infty\),那么\(f \in L^{q}, \forall 1 \leq q \leq p\)


根据Minkowski不等式,容易验证度量空间的性质

\(L^p\)是一个完备度量空间,即\(Cauchy\)列一定收敛

证明:设\(\{f_n\} \subset L^p\)为Cauchy列,那么存在子列\(\{f_{n_k}\}\)使得\(||f_{n_{k+1}} - f_{n_k}|| \leq 1/2^k\),令\(g_m = \sum_{k=1}^m |f_{n_{k+1}} - f_{n_k}|\)\(g = \lim_{m \to \infty} g_m\)

不难知道\(||g_m||_{L^p} < 1\),根据Fatou's Lemma,我们有\(\int_E g^p dx \leq 1\),也就是说\(g\)几乎处处有限

\(g<\infty\)时,我们可以定义出\(\lim f_{n_k}\)(绝对收敛),因此定义\(f=\begin{cases} \lim f_{n_k}, & g<\infty\\ 0, &g=\infty\end{cases}\),我们证明\(||f-f_n||_{L^p} \to 0\)

\(1 \leq p< \infty\),考虑Fatou's Lemma,\(\int |f-f_n|^p =\int\lim|f_{n_k}-f_{n}|^p\leq \lim\inf \int |f_{n_k}- f_n|^p\)\(f=0\)的点是一个零测集,无需考虑),而后者被Cauchy列的性质控制。特别的,\(p=\infty\)时,\(|f-f_n|\)\(|f_{n_k}-f_n|\)的极限,也被Cauchy列性质控制,\(\square\)

\(1<p<\infty\),如果\(\{f_n\}_{n=1}^{\infty}\subset L^p\)有界,那么存在子列\(\{f_{n_k}\}\),使得\(f_{n_k} \overset{w,L^p}{\to} f\)

其中以\(L^p\)弱收敛至\(f\)是指,对任意\(g \in L^{p'}\),有\(\lim_{k} \int f_{n_k} g = \int fg\)

留到泛函来写


另一个重要的考虑对象是\(L^p\)空间的稠密子集

\(1\leq p<\infty\),那么\(C_c(\mathbb{R}^d)\)\(L^p\)中稠密

证明::我们可以选择适合的\(M\),记\(E = (\{f(x) \geq M\} \cup \{|x| \geq M\})^c\),那么\(E\)需要满足\(\int_{E^c} |f|^p dm \leq \epsilon/4\)

根据Lusin TH的推论,存在\(g \in C_c(B(0,M))\)使得\(m\{E\cap(f \neq g)\} < \frac{\epsilon}{2\cdot (2M)^p}\),并且\(\sup |g| \leq \sup |f|\),特别的,在\(E\)\(\sup |g| \leq M\) \[ \begin{aligned} \int|f-g|^p dm &= \int_{E}|f-g|^p dm + \int_{E^c} |f-g|^p dm \\ &\overset{Minkowski}{\leq} \int_{E \cap \{f \neq g\}} |f-g|^p dm + \int_{E^c} |f|^p dm \\ &\leq (2M)^p \cdot \frac{\epsilon}{2\cdot (2M)^p} + \frac{\epsilon}{2} =\epsilon \end{aligned} \] 由上,\(\square\)

(Corollary):对\(1 \leq p < \infty\)\(f\in L^p\),那么\(\lim_{h\to 0} \int |f(x+h)-f(x)|^p dx = 0\)

这个推论通过取\(f\)的一个紧支集函数逼近

来证明

\(L^p\)范数的刻画

\(1 \leq p \leq \infty\)\(f\in L^p\),那么\(||f||_{L^p} = \sup\{\int fg :||g||_{L^{p'}} \leq 1\}\)

如果\(1\leq p < \infty\),上确界可以取得,即\(||f||_{L^p} = \max\{\int fg :||g||_{L^{p'}} \leq 1\}\)

证明:对\(1 \leq p \leq \infty\),根据Holder's inequality,\(\int fg \leq ||f||_{L^p} \cdot ||g||_{L^{p'}} \leq ||f||_{L^p}\)

\(1 \leq p < \infty\)时,令\(g = |f|^{p-2} f / ||f||_{L^p}^{p-1}\),那么\(||g||_{L^{p'}} = 1\)并且\(\int fg = ||f||_{L^p}\)

\(p=\infty\)时,设\(M < ||f||_{L^\infty}\),那么\(m(A) = m(\{|f|> M \}))>0\),令\(g = \frac{1_A(x) \cdot \text{sgn}(x)}{m(A)}\),则\(||g||_{L^1} = 1\),而\(\int fg \geq M\)

(Minkowski不等式):设\(1 \leq p \leq \infty\)\(f(x,y)\)\(\mathbb{R}^m \times \mathbb{R}^n\)上的可测函数,那么 \[ ||\;||f(x,y)||_{L^1(\mathbb{R}^n_y)}||_{L^p(\mathbb{R}_x^m)} \leq ||\;||f(x,y)||_{L^p(\mathbb{R}^m_x)}||_{L^1(\mathbb{R}_y^n)} \]

证明:不妨考虑\(f\geq 0\)\(1 < p < \infty\)并且\(||\;||f(x,y)||_{L^1(\mathbb{R}^m_x)}||_{L^p(\mathbb{R}_y^n)} < +\infty\)


  • (Fubini-Tonelli) 设\(f(x,y)\)是定义域在\(\mathbb{R}^{k+l}\)上的函数,如果\(f\)是非负可测(\(L^1\))函数,那么

    • 对几乎所有的\(x \in \mathbb{R}^k\)\(f|_x(y) = f(x,y)\)\(\mathbb{R}^l\)上的非负可测(\(L^1\))函数
    • \(F(x) = \int f(x,y) \text{d}y\),则\(F(x)\)\(\mathbb{R}^k\)上的非负可测(\(L^1\))函数,且\(\int F(x) dx = \int(\int f(x,y) dy)dx = \int f(x,y) dxdy\)
  • 这个定理的一个推论是如果\(E_1 \subset \mathbb{R}^k\)可测,\(E_2 \subset \mathbb{R}^l\)可测,那么\(E_1 \times E_2 \subset \mathbb{R}^{k+l}\)可测,并且\(m(E_1 \times E_2) = m(E_1 ) \cdot m(E_2)\)

    (考虑对\(1_{E_1}, 1_{E_2}\)运用Fubini定理)


\(F(x) = \int |f(x,y)|dy\),那么对\(\varphi(x) \in L^{p'}\),有\(|\int F(x)\varphi(x) dx| \leq \int F(x)|\varphi(x)|dx\)

\(=\int_x (\int_y |f(x,y)| dy) dx\overset{Fubini}{=}\int_y \int_x |f(x,y)|\cdot |\varphi(x)|dxdy \overset{Holder}{\leq} \int_y ||f|_y(x,y)||_{L^p} \cdot ||\varphi||_{L^{p'}}\)

注意到\(||F||_{L^p} = \sup \{ \int F\varphi dx:||\varphi||_{L^{p'}} \leq 1\}\),因此\(||F||_{L^p} \leq \int_y ||f|_y(x,y)||_{L^p}\)\(\square\)

(Corollary):上述不等式中,\(1\)可以更改为\(1 \leq q \leq p\)\(q\)

(Hardy不等式):设\(1<p<\infty\)\(f\in L^p((0,\infty))\),记\(F(x) = \frac{1}{x} \int_0^x f(t) dt\),那么\(||F||_{L^p} \leq \frac{p}{p-1} ||f||_{L^p}\)

注意到\(F(x) = \int_0^1 f(xt) dt\)之后,对\(t,x\)使用Minkowski不等式

\(L^p\)空间

(Definition):对于测度空间\((X,M,\mu)\)

  1. \(1 \leq p < \infty\)时,定义\(L^p(X, \mu):=\{f:||f||_{L^p(X,\mu)} < \infty\}\)

  2. \(p=\infty\)时,定义\(L^p(X, \mu) := \{f:\exists M \in (0,\infty), \exists E\subset X, \mu(E)=0, \forall x \in X \setminus E, |f(x)|\leq M\}\)

  3. 对可测函数\(f\),定义\(a_f:(0,\infty) \to [0,\infty]\)满足\(a_f(t) = \mu(\{|f(x)|>t\})\)\(f\)分布函数

  4. \(1\leq p < \infty\)时,对可测函数\(f\),定义\(||f||_{L^{p,\infty}(X, \mu)} = \sup_{t>0} t(a_f(t)^{1/p})\)

    定义弱\(L^p\)空间为\(L^{p,\infty}(X,\mu):= \{ f:||f||_{L^{p,\infty}(X,\mu) < \infty} \}\)

  5. \(p=\infty\)时,弱\(L^{\infty}\)空间就指\(L^{\infty}\)空间

\(1 \leq p \leq \infty\)\(L^{p}(X, \mu) \subset L^{p,\infty}(X, \mu)\)

证明:我们只需要证明\(\forall t > 0, a_f(t) t^p \leq ||f||_{L^p}^p\)

注意到\(a_f(t) = \int_{|f|>t} 1d\mu \leq \int_{|f|>t} \frac{|f(x)|^p}{t^p} d \mu \leq t^{-p} \int_X f|^p d \mu = t^{-p} ||f||_{L^p}^p\)\(\square\)

\(1 \leq p < \infty\)\(f \in L^p(\mathbb{R^d})\),那么\(||f||_{L^p}^p = \int_0^{\infty} p t^{p-1} a_f(t) dt\)

证明:\(\int_0^{\infty} p t^{p-1} a_f(t) dt = \int_0^{\infty} pt^{p-1} \int_{\mathbb{R}^d} 1_{|f|>t} dxdt \overset{Fubini}{=} \int_{\mathbb{R}^d} \int_0^{\infty} pt^{p-1} 1_{f|>t} dt dx = \int_{\mathbb{R}^d} \int_0^{|f(x)|} pt^{p-1} dt dx=\int_{\mathbb{R}^d} |f(x)|^p dx\)\(\square\)

微分

这里的部分对应的是Rudin的书的Chapter 7

Hardy-Littlewood极大函数

(Definition):

  1. 定义\(L_{loc}^p(E) := \{f \in L^p(K), \forall K compact \subset E\}\),称该集合中的元素为\(L^p\)局部可积函数
  2. \(f \in L_{loc}^1(\mathbb{R}^d)\),称\(Mf(x) = \sup_{r>0} \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)| dy\),这个函数称为Hardy-Littlewood极大函数

如果\(f \in L_{loc}^1(\mathbb{R}^d)\),那么\(Mf\)下半连续

证明:即说明\(\{Mf > \alpha\}\)是一个开集,假设\(Mf(x_0) > \alpha\),那么存在\(r_0\),使得\(\frac{1}{|B(x_0,r_0)|}\int_{B(x_0,r_0)} |f(y)| dy > \alpha\)

\(r_0\)稍稍变小,存在\(r_1>r_0\),使得\(\frac{1}{|B(x_0,r_1)|}\int_{B(x_0,r_0)} |f(y)| dy > \alpha\)

此时\(\forall x \in B(x_0, r_1-r_0)\),有\(\frac{1}{|B(x,r_1)|}\int_{B(x,r_1)} |f(y)| dy \overset{B(x_0,r_1) \subset B(x,r_1)}{\geq} \frac{1}{|B(x_0,r_1)|}\int_{B(x_0,r_0)} |f(y)|> \alpha\)

从而\(B(x_0,r_1-r_0) \subset \{Mf > \alpha\}\),这就证明了\(Mf\)下半连续,\(\square\)


下面,我们将花较长的篇幅运用以下定理来得到极大函数的性质

(Definiton):对算子\(M\),称其为半线性算子,如果(1)\(||M(\alpha f)|| = |\alpha|\cdot ||Mf||\),(2)\(||M(f+g)|| \leq ||Mf||+||Mg||\)

不难验证,极大函数是一个半线性算子

  • \(T\)是一个半线性算子,如果\(||Tf||_{L^{p,\infty}} \leq C_p ||f||_{L^{p}}\)并且\(||Tf||_{L^{q,\infty}} \leq C_q ||f||_{L^{q}}\)

    那么,对\(1 \leq p < r < q \leq +\infty\)\(r\),都有\(||Tf||_{L^r} \leq C_r ||f||_{L^r}\)

    也就是说,如果\(T\)的两个端点都弱\(L^p\)控制,那么在之间的区间都被\(L^p\)控制

对于\(M\)来说,我们将证明其在\(1,\infty\)这两点的弱控制性,从而得到其对\(1<p<\infty\)的控制,特别的,我们将说明\(p=1\)时没有控制性

(\(Mf\)\(\infty\)的情形):如果\(f\in L^{\infty}\),那么\(Mf \in L^{\infty}\)并且\(||Mf||_{L^{\infty}} \leq ||f||_{L^{\infty}}\)

证明略

(\(Mf\)\(1\)的情形):如果\(f \in L^1\),那么\(Mf \in L^{1,\infty}\),并且\(||M_f||_{L^{1,\infty}} \leq 3^d ||f||_{L^1}\),其中\(d\)为空间的维数

证明:固定\(\epsilon > 0\)\(Mf\)是下半连续的,因此\(\{Mf(x) > \epsilon\}\)是开集,不难证明\(m(\{Mf(x) > \epsilon\}) < +\infty\),因此\(m(\{Mf(x)>\epsilon\}) = \sup \{m(K):K\subset \{Mf(x)>\epsilon\} \;compact\}\)

任取\(K \subset \{Mf(x)>\epsilon\}\),对\(x \in K\),存在\(r_x\),使得\(|B(x,r_x)| < \epsilon^{-1} \int_{B(x,r_x)} |f(y)| dy\),根据\(K\)的紧性,存在\(\{B(x,r_x)\}\)的一个有限子覆盖,再根据下面的引理


  • (Vitali covering lemma):设\(W \subset \bigcup_{i=1}^N B(x_0,r_i)\),那么存在\(I \subset \{1,2,...,N\}\),使得
    1. \(B(x_i,r_i) \cap B(x_j,r_j) = \emptyset, \forall i \neq j\)
    2. \(W \subset \bigcup_{i \in I} B(x_i,3r_i)\)

证明:我们依次取出还没有和已经被选择出的球相交的球中半径最大的球即可,1很容易保证,对于2,注意到半径更大的球扩大3倍之后可以轻易把原本的球覆盖即可,\(\square\)


我们可以取出互不相交的球\(\{B(x_{i}, r_i)\}_{i=0}^n\),并且扩大三倍后覆盖全空间

那么,\(m(K) \leq 3^d \sum_{i=0}^n |B(x,r_x)| < 3^d\epsilon^{-1} \int_{B(x,r_x)} |f(y)| dy \leq 3^d \epsilon^{-1} ||f||_{L^1(\mathbb{R^d})}\),整理之后,就可以得到所述结论,\(\square\)

(\(Mf\)\(1\)的反例情形):如果\(f \in L^1\)并且\(f\)不几乎处处为\(0\),那么\(Mf \notin L^1\)

证明:由于\(f\)不几乎处处为0,因此存在\(a>0\),使得\(\int_{B(0,a)} |f(y)| dy = c>0\)

对于\(|x|\geq a\),都有\(Mf(x) \geq \frac{1}{B(x,2|x|)} \int_{B(x,2|x|)} |f(y)| dy \geq \frac{V_d}{|x|^d} \int_{B(0,a)} |f(y)| dy\),其中\(V_d\)是半径为\(1\)的超球的体积

于是\(\int_{\mathbb{R}^d} Mf(x) dx \geq \int_{|x| \geq a} Mf(x) dx \geq c \cdot V_d \cdot \int_{|x| \geq a} \frac{1}{|x|^d} dx = +\infty\)

Lebesgue微分定理

(Definition):对\(f \in L_{loc}^1(\mathbb{R}^d)\),称\(x\)\(f\)Lebesgue点,如果\(\lim_{r \to 0^+} \frac{1}{|B(x,r)|} \int_{B(x,r)} |f(y)-f(x)| dy = 0\)

如果\(f\)\(x\)连续,那么\(x\)\(f\)的Lebesgue点

证明是简单的

(Lebesgue point theorem,微分定理):设\(f \in L_{loc}^1(\mathbb{R}^d)\),那么几乎所有\(x \in \mathbb{R}^d\)都是\(f\)的Lebesgue点

证明:不妨设\(f \in L^1\),否则考虑\(f_n = f 1_{\overline{B(0,n)}} \in L^1\)

\(M_r f(x) = \frac{1}{B(x,r)} \int_{B(x,r)} |f(y) - f(x)| dy\)\(f^*(x) = \lim_{r \to 0^+} f_r(x)\),我们证明\(f^* \overset{a.e}{ \to } 0\)

任取\(\epsilon > 0\),存在\(g \in C_c(\mathbb{R}^d)\),使得\(||f-g||_{L^1} < \epsilon\),记\(h = f-g\),注意到\(M_r f = M_r(g + h) \leq M_rg + M_rh\),再注意到\(g^* = 0\),我们有\(f^* \leq h^*\)。展开\(M_r h\),得到\(M_r h \leq M h + |h|\),于是\(f^* \leq h^* \leq Mh + |h|\)

那么,任给\(\lambda > 0\)\(\{f^* > \lambda\} \subset \{Mh > \lambda/2\}\cup \{h > \lambda / 2\}\),于是$m({f^* > }) m({Mh > /2}) + m({h > / 2})||h||_{L^1} (3^d+1) (3^d+1) $

\(\epsilon \to 0\),再令$ \(,我们得到\)f^* 0\(,\)$

下面这个定理揭示了Lebesgue点具有和连续点相近的性质

\(x \in \mathbb{R}^d\),设\(\{E_n\}_{n=1}^\infty\)为一族Borel集,并且\(x \in E_n \subset B(x,r_n)\),其中\(\lim_{n \to \infty} r_n = 0\)

此外,存在\(\alpha > 0\),使得\(m(E_n(x)) \geq \alpha |B(x,r_n)|\),如果\(f \in L_{loc}^1(\mathbb{R}^d)\),那么\(f(x) = \lim \frac{1}{m(E_n(x))} \int _{E_n(x)} f dm\)对所有的Lebesgue点\(x\)成立(从而几乎处处成立)

证明:\(\frac{1}{m(E_n(x))} \int_{E_n(x)} |f(y)-f(x)| dy \leq \frac{1}{\alpha m(B(x,r_n))} \int_{B(x,r_n)} |f(y)-f(x)| dy\to 0\),从而\(f(x) = \lim \frac{1}{m(E_n(x))} \int _{E_n(x)} f dm\)\(\square\)

接下来,我们定义出导数的概念

\(f \in L^1\),令\(F = \int_{(-\infty, x)} f dm\),那么\(F'(x)=f(x)\),对任何Lebesgue点\(x\)(从而\(F'(x)=f(x)\)几乎处处成立)

证明:令\(\delta_n = 1/n\)\(f(x) = \lim_{n \to \infty} \frac{1}{m([x,x+\delta_n])} \int_{[x,x+\delta_n]} f dm =\lim \frac{1}{\delta_n} (F(x+\delta_n) - F(x))\),于是\(F_+'(x) = f(x)\),类似的有\(F'_-(x) = f(x)\),于是\(F'(x)=f(x)\)\(\square\)

杂谈

Vitali Covering Theorem

在之前,我们已经介绍了Vitali的一个覆盖定理

  • \(W \subset \bigcup_{i=1}^N B(x_0,r_i)\),那么存在\(I \subset \{1,2,...,N\}\),使得
    1. \(B(x_i,r_i) \cap B(x_j,r_j) = \emptyset, \forall i \neq j\)
    2. \(W \subset \bigcup_{i \in I} B(x_i,3r_i)\)

为了接下来的讨论,我们需要对这个覆盖定理作一些推广

(Vitali Covering Theorem):设\(\mathcal{F}\)是一个闭球族,满足\(\sup_{B \in \mathcal{F}}\{r(B)\} < \infty\),那么存在\(\mathcal{G} \subset \mathcal{F}\),使得

  • \(\bigcup_{B \in \mathcal{F}} B \subset \bigcup_{B \in \mathcal{G}} 5B\)
  • \(\mathcal{G}\)中闭球两两不交

证明:不妨设\(R = \sup\{r(B)\}\),记\(\mathcal{F}_k = \{B \in \mathcal{F}: \frac{R}{2^k} < r(B) \leq \frac{R}{2^{k-1}}\}\)

那么\(\mathcal{F} = \bigcup_{k=1}^{\infty} \mathcal{F}_k\)并且当\(j \neq k\)时,\(\mathcal{F}_j \cap \mathcal{F}_k = \emptyset\)

定义\(\mathcal{G}_1\)\(\mathcal{F}_1\)的极大两两不交的闭球族,根据Zorn's lemma,这个集合存在;之后,定义\(\mathcal{\widetilde{F}_1} := \{B \in \mathcal{F}_2 : \forall \widetilde{B} \in \mathcal{G}_1, \widetilde{B} \cap B = \emptyset\}\)\(\mathcal{G}_2\)\(\mathcal{\widetilde{F}_2}\)的极大两两不交的闭球族,依次类推,我们定义出\(\mathcal{\widetilde{F}_k},\mathcal{G}_k\)

最后,令\(\mathcal{G} = \bigcup_{k=1}^{\infty} \mathcal{G}_k\)即可

自然\(\mathcal{G}\)中元素两两不交,我们下面证明性质1,任取\(B \in \mathcal{F}\),假设\(B \in \mathcal{F}_k\)

如果\(B \in \mathcal{\widetilde{F}}_k\),那么存在\(\widetilde{B} \in \mathcal{G}_k\)使得\(B \cap \widetilde{B}\neq \emptyset\),如果\(B \in \widetilde{F}_k \setminus \mathcal{F}_k\),那么存在\(\widetilde{B} \in G_j(j<k)\)使得\(B \cap \widetilde{B}\neq \emptyset\),无论哪一种情况,存在\(\widetilde{B} \in \mathcal{G}\)使得\(B\cap \widetilde{B} \neq \emptyset\)\(r(B)<2r(\widetilde{B})\),此时\(B \subset 5\widetilde{B}\)\(\square\)

(Definition):设闭球族\(\mathcal{F}\)\(E\)的覆盖,如果\(\forall x \in E\)\(\inf_{x \in B \in \mathcal{F}} \{r(B)\} = 0\),那么我们称\(\mathcal{F}\)\(E\)的一个细覆盖

在细覆盖的条件下,我们可以给出Vitali构造的更强的性质:

设闭球族\(\mathcal{F}\)\(E\)的细覆盖,且\(\sup_{B \subset \mathcal{F}}{r(B)} < \infty\),那么存在两两不交的可数子族\(\mathcal{G}\),使得 \[ \forall B_1,B_2,...,B_k \in \mathcal{G},E \setminus (\bigcup_{1\leq i\leq k} B_i) \subset \bigcup_{B \in G \setminus \bigcup B_i} 5B \]

证明:\(\mathcal{G}\)的取法和Vitali定理的构造是一样的,固定\(B_1,...,B_k\)

\(x \in E \setminus (\bigcup_{1\leq i\leq k} B_i)\),其属于开集\((\bigcup_{i=1}^k B_i)^c\),因此存在\(\delta > 0\),使得\(B(x,\delta) \subset (\bigcup_{i=1}^k B_i)^c\)

由于\(\mathcal{F}\)\(E\)的细覆盖,因此存在\(B_x \in \mathcal{F}\),使得\(B_x \subset B(x,\delta)\),从而\(B_x \cap (\bigcup_{i=1}^k B_i) = \emptyset\)

根据\(\mathcal{G}\)的性质,存在\(\widetilde{B} \in \mathcal{G} \setminus \bigcup B_i\),使得\(B_x \subset 5\widetilde{B}\),这就证明了结论,\(\square\)

如果我们有细覆盖,实际上我们还能揭示\(\mathcal{G}\)作为一个子族,它的测度和\(\mathcal{F}\)的测度是相近的:

设闭球族\(\mathcal{F}\)\(E\)的细覆盖,且\(\sup_{B \subset \mathcal{F}}{r(B)} < \infty\)\(m^*(E) < \infty\)时,存在可数子族\(\mathcal{G} = \{B_k\}\)满足Vitali覆盖定理中的条件,并且\(m^*(E \setminus \bigcup_{k} B_k) = 0\)

证明:\(m^*(E) < \infty\),因此存在开集\(G\),使得\(m(G) < m^*(E) + 1\),根据细覆盖和开集的性质,\(\forall x \in E, \exists \delta > 0, \forall 0<\delta_0 < \delta, \exists B_x \in \mathcal{F}, B_x \subset B(x,\delta_0) \subset \mathcal{G}\)

这告诉我们,如果我们取\(\mathcal{F}\)中的子族\(\mathcal{H} = \{B\in \mathcal{F}:B \subset \mathcal{G}\}\),则\(\mathcal{H}\)也是\(E\)的一个细覆盖

我们取\(\mathcal{G}\)\(\mathcal{H}\)的Vitali子族,\(m^*(\bigcup_{B \in \mathcal{G}} B) \leq m^*(G)<\infty\)

因此,\(m^*(E \setminus \bigcup_{i=1}^k B_i) \leq m^*(\bigcup_{j=k+1}^{\infty} 5B_j) \leq 5^d \sum_{j=k+1}^{\infty} m(B_j) \to 0\)\(\square\)

单调函数的可导性

通过证明单调函数的不连续点集是可数的,我们可以证明

单调函数\(f\)几乎处处连续

下面,我们来证明比这个更强的一个性质

单调函数\(f\)几乎处处可导

定义\(D^{\pm} f(x) = \limsup_{\delta \to 0^{\pm}} \frac{f(x+\delta) - f(x)}{\delta}\)\(D^{\pm} f(x) = \liminf_{\delta \to 0^{\pm}} \frac{f(x+\delta) - f(x)}{\delta}\)

我们只需要证明:\(\{D^+f = D^-f = D_+f = D_- f \neq \infty\}^c\)是一个零测集即可,注意到\(D^+ f\geq D_+f, D^- f \geq D_- f\),那么只需要证明\(\{D_+ f < D^- f\}, \{D^+ f > D_- f\}\)以及四个导数相同但等于无限的集合为零测集即可,前两个集合的证明是类似的,我们只证明\(\{D^+ f > D_- f\}\)为零测集

  1. 证明\(\{D^+ f > D_- f\}\)为零测集,注意到四个导数\(\geq 0\),我们只需要证明对\(r, s \in \mathbb{Q}\)\(\{D^+ f > r > s > D_- f\}\)是零测集即可,记该集合为\(E_{r,s}\)

由于\(f\)为几乎处处连续函数,因此为可测函数,从而\(E_{r,s}\)为可测集,那么我们可以取开集\(G\),使得\(m(G) < (1+\epsilon) m(E_{r,s})\)

\(\forall x \in E_{r,s}\)\(D_- f(x) < s\),根据下极限的性质,我们可以取出一列子列\(\{\delta_k\} \to 0\),使得\(\frac{f(x-\delta_k) - f(x)}{-\delta_k} < s\)并且\([x-\delta_k,x] \subset G\),记\(\mathcal{F}_x = \{[x-\delta_k, x]\}\)\(\mathcal{F}_1 = \bigcup_{x \in E_{r,s}} \mathcal{F}_x\)\(\mathcal{F}_1\)\(E_{r,s}\)的细覆盖,我们取其Vitali子族\(\mathcal{G}_1\)

\(E_{r,s}' = E_{r,s} \cap (\bigcup_{I \in \mathcal{G}_1} I)\),类似上面的讨论,对于\(y \in E_{r,s}'\),我们可以取一列子列\(\{\sigma_k\} \to 0\),使得\(\frac{f(y+\sigma_k) - f(y)}{\sigma_k} > r\)并且\([y,y+\sigma_k] \subset [x_n-\delta_n,x_n] \in \mathcal{G}_1\),同样的,这样的区间的全体构成细覆盖,取其Vitali子族\(\mathcal{G}_2\)

  1. 根据\(\mathcal{G}_1, \mathcal{G}_2\)的选取方式,有如下的性质
  1. \(m(\bigcup_{I \in \mathcal{G_1}}) = \sum_{i=1}^{\infty} \delta_i \leq m(G) < (1+\epsilon) m(E_{r,s})\)

  2. (细覆盖的选取)\(m(E_{r,s}') = m(E_{r,s})\)

  3. \(m(\bigcup_{I \in \mathcal{G_2}} I) = \sum_{i=1}^{\infty} \sigma_i \geq m(\bigcup_{I \in \mathcal{G_2}} I \cap E_{r,s}') =m(E_{r,s}') = m(E_{r,s})\)

  4. \(\forall [x_n - \delta_n,x_n] \in \mathcal{G}_1\)\(f(x_n) - f(x_n - \delta_n) < s \delta_n\)

  5. \(\forall [y_m,y_m + \sigma_m] \in \mathcal{G}_2\)\(f(y_m + \sigma_m) - f(y_m) > r \sigma_m\)

  6. \([y_m,y_m+\sigma_m] \subset [x_n-\delta_n,x_n]\)并且\([y_m,y_m+\sigma_m]\)不交

\(\sum_{m=1}^{\infty} f(y_m + \sigma_m) - f(y_m) \overset{5}{>} r\sum_{m=1}^{\infty} \sigma_m \overset{3}{\geq} rm(E_{r,s})\)

另一方面,\(\sum_{m=1}^{\infty} f(y_m + \sigma_m) - f(y_m) \overset{6}{\leq} \sum_{n=1}^{\infty} f(x_n) - f(x_n - \delta_n) \overset{4}{\leq} s \sum \delta_n \overset{1}{\leq} s(1+\epsilon) m(E_{r,s})\)

\(\epsilon \to 0\),我们得到\(rm(E_{r,s}) \leq sm(E_{r,s})\),矛盾,于是\(m(E_{r,s}) = 0\)

  1. 最后,我们还需要证明集合\(E = \{f' = \infty\}\)是零测集

考虑集合\(E_m = \{f' > m\}\),那么\(E = \liminf E_m\),对于\(x \in E_m\),存在\(c_x\),使得\(\frac{f(x \pm c_x) - f(x)}{\pm c_x} > m, \forall 0<c<c_x\),令\(\mathcal{H}_x = \{[x-c,x+c]:0<c<c_x\}\)\(\mathcal{H} = \bigcup \mathcal{H}_x\)

那么,\(\mathcal{H}\)是一个细覆盖,取其Vitali子族\(\mathcal{I} = \{[x_k-c_k,x_k+c_k]\}\)

\(\sum c_k \leq \frac{1}{m} \sum f(x_k + c_k) - f(x_k) \leq \frac{1}{m} (f(b)-f(a))\),令\(m \to \infty\),我们得到\(m(E)=0\)\(\square\)

积分与导数

\(f:[a,b] \to \mathbb{R}\)单调,\(-\infty < a < b < \infty\),那么\(f'\)可测,并且\(\int_{[a,b]} |f'| dx \leq |f(b)-f(a)|\)

\(I = \{x\in (a,b) : f'(x)\;exists\}\),那么\(m([a,b]) = m(I)\),补充定义\(f(x)=f(b), x>b\)

\(f_n(x) = n(f(x + \frac{1}{n}) - f(x))\)\(f' = \lim f_n\),而\(f_n\)可测,因此\(f'\)可测,不妨设\(f\)单增,那么\(f' \geq 0(a.e.)\)

根据Fatou's lemma,\(\int_{[a,b]} f'(x) dx = \int_{[a,b]} \lim f_n(x)dx \leq \liminf \int_{[a,b]} f_n(x)dx\)

\(= \liminf n (\int_{[b,b+1/n]} f dx-\int_{[a,a+1/n] } f dx)=f(b)- \liminf n\int_{[a,a+1/n]} f(a) dx = f(b)-f(a)\)\(\square\)

Remark:考虑Cantor函数,其几乎处处可导,导数为0,但是\(f(1)-f(0)=1\),这说明等号不一定能取到

求和与导数

\(\{f_n\}\)是一族单调函数,且\(\sum_{n=1}^{\infty} f_n(x)\)\([a,b]\)上收敛,那么\((\sum_{n=1}^{\infty} f_n(x))' = \sum_{n=1}^{\infty} f_n'(x)\)几乎处处成立

证明:记\(F = \sum_{n=1}^{\infty} f_n\),存在一个零测集\(E\),在\(E\)的补集上,\(\{f_n\},F\)可导

\(F = \sum_{n=1}^N f_n + R_N\),自然\(F' = \sum_{n=1}^N f_n' + R_N'\),我们证明\(R_N' \overset{a.e.}{\to} 0\)

首先,\(R_N' \overset{a.e.}{\geq} R_{N+1}'\)\(R_N'\)单调有界,因此\(R_N'\)收敛,记\(\lim R_N' = r \geq 0\)

根据Fatou's lemma,\(\int r dx \leq \liminf \int_{[a,b]} R_N' dx \leq \liminf (R_N(b)-R_N(a)) = 0\),从而\(r\overset{a.e.}{=}0\)\(\square\)

有界变差函数

(Definition):设\(f:[a,b] \to \mathbb{R}\)\(\mathcal{P}\)\([a,b]\)的一个分划,\(\mathcal{P}:a=x_0<x_1<...<x_n = b\)

  1. \(V(\mathcal{P};f)=\sum_{i=1}^n |f(x_i) - f(x_{i-1})|\)\(f\)关于\(\mathcal{P}\)变差
  2. \(V_a^b(f):= \sup_{\mathcal{P} \;on\;[a,b]}\{V(\mathcal{P};f)\}\)\(f\)\([a,b]\)上的全变差
  3. 如果\(V_a^b(f) < \infty\),则称\(f\)有界变差函数,简称\(BV\)函数,\([a,b]\)\(BV\)函数的全体为\(BV([a,b])\)
  4. 如果\(f \in BV([a,b])\),称\(V_a^x(f)\)\(f\)\([a,b]\)上的变差函数

Prop1:如果\(f \in BV([a,b])\),那么\(f\)\([a,b]\)上有界

考虑划分\(\mathcal{P} : a<x<b\)\(V(\mathcal{P};f) = |f(x)-f(a)|+|f(b)-f(x)| < \infty\),这就说明\(|f(x)| < \infty\)

Prop2:\(V_a^b(\alpha f + \beta g) \leq |\alpha| V_a^b (f) + |\beta| V_a^b(g)\)

Prop3:\(V_a^b(f) = V_a^c(f) + V_c^b(f), \forall c \in (a,b)\)

任何一个\([a,c],[c,b]\)的划分可以拼成一个\([a,b]\)的划分,任何一个\([a,b]\)的划分在中间添加\(c\)之后可以拆成两个子划分,通过这样的方式来说明即可

Prop4:单调函数是有界变差函数

Prop5:可导并且导数有界的函数是有界变差函数

(Jordan分解):\(f\in BV([a,b])\)当且仅当\(f = g-h\),其中\(g,h\)\([a,b]\)上单增

证明:\(\Leftarrow\)是比较容易的,下面考虑\(\Rightarrow\)

\(F(x) = V_a^x(f)\)\(g = \frac{1}{2}(F+f)\)\(h = \frac{1}{2}(F-f)\),此时\(f = g-h\)

\(a \leq x \leq y \leq b\)\(g(y)-g(x) = \frac{1}{2}(V_x^y(f) + f(y)-f(x)) \geq \frac{1}{2}(V_x^y(f) - |f(y)-f(x)|) \geq 0\)

因此\(g\)单增,\(h\)类似,\(\square\)

(Corollary):对于有界变差函数\(f\),其不连续点集至多可数,并且几乎处处可导,\(f' \in L^1\)

\(f \in BV([a,b])\),记\(F(x) = V_a^x(f)\),那么\(F'\overset{a.e.}{=}|f'|\)

证明:\(f,F\)都几乎处处可导,注意到对几乎所有的\(x\)

\(|f'(x)| = \lim_{h \to 0^+} \frac{|f(x+h)-f(x)|}{h} \leq \lim_{h \to 0^+} \frac{V_a^{x+h} - V_a^x}{h} = F'(x)\)

我们证明对\(E=\{x\in (a,b) : F',f' \;exists\;,F'>|f'|\}\)\(m^*(E)=0\)

这只需要证明\(E_k = \{x:F'-|f'|>1/k\}\)\(m^*(E_k) = 0\)

如果\(F'(x) - |f'(x)|>1/k\),那么将存在\(m\),对\(y,z \in (x-\frac{1}{m},x+\frac{1}{m})\)\(y<z\),有\(\frac{F(z)-F(y)}{z-y} - \frac{|f(z)-f(y)|}{z-y} > \frac{1}{2k}\),这一点运用导数的极限性质就可以证明,此时,设\(E_{k,m} =\{x : \forall y,z \in (x-\frac{1}{m},x+\frac{1}{m}),y<z, \frac{V_y^z(f) - |f(z)-f(y)|}{z-y} > \frac{1}{2k}\}\),那么\(E_k \subset \bigcup_m E_{k,m}\)

接下来,我们取划分\(\mathcal{P}\)满足\(\Delta(\mathcal{P}) < \frac{1}{m}\)并且\(V(f;\mathcal{P})>V_a^b(f) - \epsilon\)

此时,对于\(x_i,x_{i+1} \in \mathcal{P}\),有\(m^*(E_{k,m} \cap (x_{i-1},x_i)) \leq 2k(V_{x_{i-1}}^{x_{i}} - |f(x_i)-f(x_{i-1})|)\)

最终有\(m^*(E_{k,m}) \leq \sum m^*(E_{k,m} \cap (x_{i-1},x_i)) \leq 2k(V_a^b - V(f;\mathcal{P})) \leq 2k\epsilon\),令\(\epsilon \to 0\),得到\(m^*(E_{k,m}) = 0\),一路回推即可证明结论,\(\square\)

这里设出\(E_{k,m}\)主要是为了得到一致性

(Corollary):\(f\in BV([a,b])\),则\(\int_a^b |f'| dx \leq V_a^b(f)\)

绝对连续函数

(Definition):1. 设\(f:[a,b] \to \mathbb{R}\),其中\(-\infty <a < b < +\infty\),如果对任意\(\epsilon > 0\),存在\(\delta > 0\),使得\([a,b]\)中任意有限个互不相交的开区间\((x_i,y_i)\),有\(\sum (y_i - x_i) < \delta\)都有\(\sum |f(y_i) - f(x_i)| < \epsilon\),则称\(f\)\([a,b]\)上绝对连续

  1. 绝对连续函数的全体记为\(AC([a,b])\)

Prop1:\(AC([a,b])\)构成线性空间

Prop2:\(f\)绝对连续,则\(f\)连续

Prop3:如果\(f \in L^1([a,b])\),那么\(F = \int_a f \in AC([a,b])\)

根据积分的绝对连续性证明

Prop4:如果\(f \in L^1([a,b])\),那么\(\int_a^x f' \in AC([a,b])\)

\(f \in AC([a,b])\),则\(f \in BV([a,b])\)

存在\(\delta>0\),使得对\([a,b]\)中有限个互不相交的开区间\((x_i,y_i)\),如果\(\sum (y_i-x_i) < \delta\),那么\(\sum |f(y_i) - f(x_i)| \leq 1\)

考虑\(a,a+\delta,a+2\delta,...,b\),将这些间隔点插入到任意分划\(\mathcal{P}\)中,得到\(\mathcal{P}_0\),自然\(V(f;\mathcal{P}) \leq V(f;\mathcal{P}_0)\),对于\(\mathcal{P}_0\)来说,对于\(a+i\delta\)\(a+(i+1)\delta\)之间的贡献可以被\(1\)所限制,从而\(V(f;\mathcal{P}) \leq \lceil \frac{b-a}{\delta} \rceil\)

如果\(f \in AC([a,b])\),那么\(m(f(Z))=0\),其中\(Z\)为零测集

任取\(\epsilon\),设\(\delta\)是绝对连续中关于\(\epsilon\)得到的参数,对于\(Z\),我们取开集\(G \supset Z\)使得\(m(G)<\delta\)

由于\(m(G)<\delta\),不妨设\(G = \bigcup (a_k,b_k)\),那么\(\sum |f(c_k)-f(d_k)| < \epsilon\),其中\(c_k,d_k \in (a_k,b_k)\),我们取\(c_k,d_k\)分别为\((a_k,b_k)\)\(f\)的最小值点和最大值点,我们就证明了\(m(f(G)) < \epsilon\)

\(f \in AC([a,b])\),那么\(f\)几乎处处可导,并且\(f' \in L^1\)

如果\(f'=0\)几乎处处成立,那么\(f\)是常数

证明:\(f \in BV([a,b])\),从而几乎处处可导,并且\(f' \in L^1\)

\(E = \{x \in [a,b]:f'(x)=0\}\),任取\(c \in (a,b)\),记\(E_c = (a,c) \cap E\),固定\(\epsilon > 0\),对于\(y \in E_c\),存在\(h\),使得\([y-h,y] \subset (a,c)\)并且对任意\(x \in [y-h,y]\),有\(\frac{|f(y)-f(x)|}{y-x} < \frac{\epsilon}{2(c-a)}\),此时\(\mathcal{F}:=\{[x,y]:y \in E_c, \frac{|f(y)-f(x)|}{y-x} < \frac{\epsilon}{2(c-a)}\}\)构成\(E_c\)的细覆盖

根据Vitali覆盖定理,我们可以取出\(\mathcal{F}\)中的互不相交的闭区间\(\{[x_i,y_i]\}\)并且\(\sum y_i - x_i = m(E_c)=c-a\)

\(\delta\)为关于\(f,\epsilon/2\)的决定连续性的常数,设\(N\)满足\(\sum^N y_i - x_i > c - a - \delta\)

不妨设\(y_1<x_2,y_2<x_3,...,y_{N-1}<x_N\),补充\(y_0= a,x_{N+1}=c\),那么\(\sum_{i=0}^N |x_{i+1}-y_i| < \delta\)

于是\(|f(c)-f(a)|\leq \sum_{i=0}^N |f(x_{i+1})-f(y_i)| + \sum_{i=1}^N |f(y_i)-f(x_i)|\leq \frac{\epsilon}{2} + \frac{\epsilon}{2} = \epsilon\),令\(\epsilon \to 0\),得到\(f(c)=f(a)\)\(\square\)

(微积分基本定理):如果\(f \in AC([a,b])\),那么\(f(x)=f(a)+\int_{[a,x)} f' dm, \forall x \in [a,b]\)

证明:\(f \in AC([a,b])\),因此\(f' \in L^1\),从而\(F=\int_a^x f' dm \in AC([a,b])\)

并且\(F'=f'\)几乎处处成立(微分定理),因此\(f-F\)是常数,通过代入可以得到常数为\(f(a)\)\(\square\)

\(f \in BV([a,b])\),则

  1. \(f \in AC([a,b])\)当且仅当\(V_a^x(f) \in AC([a,b ])\)

  2. \(f \in AC([a,b])\)当且仅当\(\int_a^b |f'| dm = V_a^b(f)\)

  3. \(f \in AC([a,b])\)当且仅当\(f\)将零测集映射到零测集

在绝对连续函数上的积分有和数分中的积分类似的技巧