故事 · 棣莫弗、拉普拉斯与钟形的诞生
Origin Story · De Moivre, Laplace, and the Birth of the Bell
1733 年，旅居伦敦、靠帮赌徒算概率为生的法国数学家 Abraham de Moivre 在研究"抛 n 次硬币正面次数"时，
为了避开庞大的组合数计算，推导出一个光滑的近似公式 —— 那正是正态钟形曲线的雏形（当时还没有"正态"这个名字）。
半个多世纪后，Pierre-Simon Laplace 把它从 p = 0.5 推广到任意 p，写进了他的概率论巨著。
于是人类第一次意识到：大量独立随机事件叠加，结果会自动汇聚成钟形。这条定理是中心极限定理的前身，
也是今天"n 大时用正态近似二项"的源头。正态不是被发明的，它是从抛硬币里被发现的。
In 1733, the French mathematician Abraham de Moivre — then living in London and earning a living computing odds for gamblers — was studying the number of heads in n coin tosses.
To dodge the brutal combinatorial sums, he derived a smooth approximation formula — and that formula was the first sketch of the normal bell curve (the word "normal" didn't exist yet).
Half a century later, Pierre-Simon Laplace extended it from p = 0.5 to any p and recorded it in his great treatise on probability.
For the first time, humanity grasped a deep truth: add up enough independent random events and the result spontaneously settles into a bell. This theorem is the direct ancestor of the Central Limit Theorem,
and the wellspring of today's "use the normal to approximate the binomial when n is large." The normal wasn't invented — it was discovered, hiding inside coin tosses.
1 n 越大，正态曲线越严丝合缝地盖住柱群 The larger n, the tighter the normal curve hugs the bars
还很离散2 近似误差随 n 增大而衰减 The approximation error decays as n grows
纵轴是二项柱顶与正态曲线的最大偏差，横轴是 n。误差大约以 1/√n 的速度衰减 —— 这也是下一页 √n 法则的同一根脉络。 The vertical axis is the maximum gap between bar tops and the normal curve; the horizontal axis is n. The error fades roughly like 1/√n — the same root cause behind the √n rule on the next page.
3 现实里的棣莫弗-拉普拉斯 De Moivre–Laplace in the Real World
正态的起源：钟形曲线第一次登场，就是为了近似抛硬币的二项分布，而非凭空假设。
Origin of the normal: the bell curve made its first appearance specifically to approximate the binomial distribution of coin tosses — not as some convenient assumption.
二项近似：n 大时用正态算 B(n, p) 的尾概率，免去逐项累加组合数，工程上极常用。
Binomial approximation: for large n, use the normal to compute tail probabilities of B(n, p) — no more grinding through combinatorial sums. A workhorse in engineering practice.
CLT 前身：它是中心极限定理的特例与历史先声，铺平了"万物趋正态"的道路。
Ancestor of the CLT: it is the historical forerunner and a special case of the Central Limit Theorem, paving the way for the idea that "almost everything tends toward normal."
连续近似离散：用连续的正态去逼近离散的计数（配合 ±0.5 连续性校正），是属性数据建模的常规手段。
Continuous approximating discrete: using the continuous normal to approximate a discrete count (with a ±0.5 continuity correction) is the standard tool for modeling attribute data.
一句话In One Line
这条定理告诉我们一个深刻的事实：正态钟形不是人为规定的，而是大量独立成败叠加后自然涌现的形状。
二项分布只要 n 够大、p 不太极端，就会自动逼近 N(np, np(1−p))。
它的价值有两层：实用上，让我们能用连续的正态去近似离散的二项，省掉组合数；思想上，它是
中心极限定理的第一声雷 —— 下一页我们会看到，这个"趋向正态"的魔法远不止硬币，而是几乎对任何母体都成立。
但要记住前提：n 要足够大、p 不能太靠近 0 或 1，否则曲线根本贴不上去。
This theorem hands us a deep truth: the normal bell isn't a human convention — it is the shape that emerges, on its own, when many independent success-or-fail events stack up.
As long as n is large enough and p isn't too extreme, the binomial drifts automatically toward N(np, np(1−p)).
Its value lives on two levels: practically, it lets us approximate the discrete binomial with the continuous normal and skip the combinatorial grind; conceptually, it is the
first thunderclap of the Central Limit Theorem — on the next page you will see that this "drift toward normal" magic isn't limited to coins, but holds for almost any population.
Just remember the precondition: n must be large enough, and p mustn't sit too close to 0 or 1 — otherwise the curve simply will not lay down on the bars.
常见误用Common Mistakes
n 很小或 p 很极端还硬用正态近似。经验法则 np 与 n(1−p) 都 ≥ 5 时近似才靠谱，p 极端时应直接用二项或泊松。
Forcing the normal approximation when n is small or p is extreme. The rule of thumb is np ≥ 5 and n(1−p) ≥ 5; when p is near 0 or 1, use the exact binomial or a Poisson approximation instead.
用连续正态算离散概率却忘了连续性校正。算 P(X ≤ k) 时用 k+0.5、算 P(X ≥ k) 时用 k−0.5，否则系统性偏差。
Using the continuous normal for a discrete probability and forgetting the continuity correction. For P(X ≤ k) use k + 0.5; for P(X ≥ k) use k − 0.5 — otherwise you pick up a systematic bias.
把"近似"当成"相等"。它始终是近似，对尾部小概率事件误差可能放大，关键决策仍要核对精确值。
Treating "approximation" as "equality". It is always an approximation; for small tail probabilities the relative error can blow up. Cross-check against the exact value before any high-stakes decision.