故事 · 统计学的魔法
Story · The Magic Trick of Statistics
中心极限定理(CLT)说:取一个任意形状的母体(均值 μ、标准差 σ),每次抽 n 个样本算出均值 x̄,重复很多次 ——
这堆 x̄ 的分布会趋近正态 x̄ ~ N(μ, σ²/n),无论母体本身多奇怪。它的标准差(叫标准误 SE = σ/√n)随 n 增大而收窄。
这正是为什么平均值如此好用,也是为什么大样本更可靠。但注意分母是 √n 而非 n:想把误差减半,需要把样本量扩到 4 倍 —— 回报递减。
The Central Limit Theorem says: take any parent distribution (mean μ, std-dev σ) — any shape at all — draw n samples, compute the mean x̄, repeat many times.
The pile of x̄ values converges to x̄ ~ N(μ, σ²/n) no matter how ugly the parent looked. Its std-dev — called the standard error, SE = σ/√n — shrinks as n grows.
That's why averages are so useful, and why larger samples earn more trust. But notice the denominator is √n, not n: to cut the error in half you need 4× the sample size — diminishing returns.
1 母体分布:故意挑个不正态的 Parent Distribution: Picked to Look Non-Normal
均匀Uniform2 样本均值 x̄ 的分布:逐渐变成钟形 Distribution of the Sample Mean x̄: Drifting to a Bell
点"开始反复抽样",每次从上面奇怪的母体抽 n 个数算均值,落进这张直方图。看它如何越堆越像正态钟形,并随 n 增大越收越窄 —— 红色曲线是理论正态 N(μ, σ²/n)。 Hit "Start Resampling" and each cycle draws n numbers from that weird parent above, averages them, and drops the result into this histogram. Watch it pile up into a bell and, as n grows, tighten around μ — the red curve is the theoretical normal N(μ, σ²/n).
3 现实里的中心极限 CLT in the Real World
抽样分布:民调、抽检的均值即便母体偏态,也能用正态近似 —— CLT 是推断统计能成立的根基。
Sampling distributions: poll results and inspection means can be approximated as normal even when the parent is skewed — CLT is the bedrock of inferential statistics.
标准误:估计精度由 SE=σ/√n 决定,是置信区间和误差棒的来源。
Standard error: precision of an estimate is governed by SE = σ/√n — it's the source of every confidence interval and error bar.
样本量规划:想要的精度反推 n,因 √n 关系,提升精度的成本越来越高。
Sample-size planning: back-solve n from your target precision — because of the √n relationship, each extra digit of precision costs disproportionately more.
控制图:子组均值 X̄ 图天然趋正态,控制限按 σ/√n 收窄,正是 SPC 灵敏的原因。
Control charts: subgroup-mean X̄ charts inherit normality from CLT, and the control limits tighten at σ/√n — that's what makes SPC so sensitive.
一句话In One Line
CLT 给了我们一张"通行证":即使不知道母体长什么样,也能用正态来对均值做推断 —— 这是 t 检验、ANOVA、控制图能用的底层依据。
它还给了一把精度的尺子:误差以 1/√n 收窄。所以多采点确实更准,但永远别幻想"样本量随便加加就万无一失"——
因为是 √n 不是 n,精度翻倍的代价是四倍成本。
CLT hands us a free pass: even without knowing what the parent looks like, we can use the normal distribution to reason about its mean — that's the foundation of the t-test, ANOVA, and every control chart.
It also hands us a ruler for precision: error shrinks at 1/√n. So yes, more data really is more precise — but never daydream that "just throw in more samples" guarantees the answer.
The denominator is √n, not n: doubling precision costs four times the budget.
常见误用Common Mistakes
以为 CLT 让原始数据变正态。变正态的是均值的分布,单个数据该偏还偏。
Thinking CLT makes the raw data normal. Only the distribution of the mean turns normal — individual observations stay as skewed as ever.
n 很小就套正态。母体越偏,需要的 n 越大;小样本可能仍不够。
Reaching for the normal approximation at tiny n. The more skewed the parent, the larger the n you need — small samples may still fall short.
把 SE 当成总体标准差 σ。SE=σ/√n 是均值的波动,远小于 σ 本身。
Confusing SE with the population σ. SE = σ/√n is the jitter of the mean — far smaller than σ itself.