为何交换极限与积分如此困难
我们想知道何时 lim integral f_n = integral lim f_n——即何时可以把极限挪进积分号内。对黎曼积分,这几乎总需要一致收敛这一沉重的假设。著名的警示是逃逸的凸包:在 [0, ∞) 上令 f_n 在 [n, n+1] 上为 1、其余处为 0。则 f_n 在每一点都趋于 0,可对所有 n 都有 integral f_n = 1。于是 lim integral f_n = 1 而 integral lim f_n = 0。质量漏到无穷远去了。任何定理都必须有一个假设来排除这种情况。
单调收敛与法图
[[monotone-convergence-theorem|单调收敛定理]](MCT)。 若 0 <= f_1 <= f_2 <= … 可测且 f_n 逐点递增趋于 f,则 integral f_n 递增趋于 integral f。这个 “递增” 假设正是阻止逃逸凸包的关键:不断上涨的潮水无法把质量漏走。MCT 是驱动非负积分可加性的引擎,并经由它驱动全部线性性。
[[fatou-lemma|法图引理]]。 对任意非负可测的 f_n(不假设单调、不假设收敛),integral(liminf f_n) <= liminf integral f_n。它是单向的安全网:在逃逸凸包的例子里 liminf f_n = 0,故左端为 0,而 liminf integral f_n = 1——不等式 0 <= 1 成立,恰在质量逃逸处出现严格的损失。法图引理由对 g_n = inf_{k>=n} f_k 应用 MCT 得到,这些 g_n 递增趋于 liminf f_n。
Escaping bump: f_n = 1 on [n, n+1], 0 elsewhere on [0, infinity)
pointwise: for each fixed x, f_n(x) = 0 once n > x => f_n -> 0
integrals: integral f_n = 1 for every n
MCT? NO -- f_n is not increasing (the bump moves), so MCT does not apply.
Fatou: liminf f_n = 0
integral(liminf f_n) = 0 <= liminf integral f_n = 1. OK (strict).
Dominated? would need g >= |f_n| all n with integral g < infinity;
the smallest such g is 1 on [0, infinity), integral = infinity.
NO dominating majorant exists -> DCT does not apply either.控制收敛定理
[[dominated-convergence|控制收敛定理]](DCT)。 设 f_n 几乎处处趋于 f,且存在单个固定的可积控制函数 g——即对所有 n 有 |f_n| <= g,且 integral g < ∞。则 f 可积,且 lim integral f_n = integral f。更进一步,integral |f_n - f| -> 0。这是实际中最常动用的定理:不需要单调性、不需要一致收敛——只要逐点收敛,再由一个可积的 “盖子” g 把它压住即可。
Proof of DCT from Fatou (sketch):
Hypotheses: f_n -> f a.e., |f_n| <= g, integral g < infinity.
Step 1. g + f_n >= 0. Apply Fatou to (g + f_n):
integral(g + f) <= liminf integral(g + f_n)
=> integral g + integral f <= integral g + liminf integral f_n
=> integral f <= liminf integral f_n.
Step 2. g - f_n >= 0. Apply Fatou to (g - f_n):
integral(g - f) <= liminf integral(g - f_n)
=> integral g - integral f <= integral g - limsup integral f_n
=> limsup integral f_n <= integral f.
Combine: limsup integral f_n <= integral f <= liminf integral f_n,
forcing lim integral f_n = integral f. QED