故事 · 伯努利与赌局里的数学
Origin Story · Bernoulli and the Math of the Gambling Table
十七世纪末,瑞士数学家 Jacob Bernoulli 把"一次成功或失败"的最简单随机试验称作伯努利试验,
再把 n 次独立重复的伯努利试验加起来 —— 就成了二项分布。它最早服务于赌局与保险:n 局里赢几局、n 个保单出几个险。
公式里的 C(n, k) 是组合数,回答"n 个位置里挑 k 个放成功,有多少种排法";后面的 pᵏ(1−p)ⁿ⁻ᵏ
是每一种排法的概率。今天它是抽样检验、不良品计数、A/B 测试的基石 —— 凡是数"计数型成败",背后都是它。
Late in the 17th century the Swiss mathematician Jacob Bernoulli named the simplest random experiment — one success-or-fail trial — a Bernoulli trial.
Sum up n independent repetitions of it and you have the binomial distribution. Its first clients were gamblers and insurers: how many of n games will I win, how many of n policies will pay out?
In the formula, C(n, k) is the binomial coefficient — "n choose k" — counting the ways to place k successes among n slots; the trailing pᵏ(1−p)ⁿ⁻ᵏ
gives the probability of each such arrangement. Today it is the bedrock of acceptance sampling, defect counting and A/B testing — anywhere you tally pass/fail outcomes, the binomial is underneath.
1 p 拉偏它、n 拉大它,柱子悄悄长成钟形 p Skews It, n Grows It — the Bars Quietly Form a Bell
对称symmetric2 方差 np(1−p):p=0.5 时最不确定 Variance np(1−p): Maximum Uncertainty at p = 0.5
固定 n,方差随 p 走出一条抛物线:p 越靠近 0.5,结果越难猜、波动越大;p 越靠近 0 或 1,几乎必然,波动趋零。 Hold n fixed and the variance traces a parabola in p: closer to 0.5 means the outcome is hardest to predict and the jitter is largest; closer to 0 or 1 means near-certainty, and the jitter collapses toward zero.
3 现实里的二项分布 The Binomial in the Wild
不良品计数:抽 n 件查出几件不良,是来料检验、批次放行判定的基本模型。
Defect counting: how many defects in n inspected pieces — the workhorse model for incoming inspection and batch-release decisions.
抽样检验:AQL 抽样计划用二项(或超几何)算"抽 n 件接受/拒收"的概率,制定 OC 曲线。
Acceptance sampling: AQL plans use the binomial (or hypergeometric) to compute accept/reject probabilities and draw the OC curve.
A/B 测试:n 个访客里几个转化,转化次数服从二项,是显著性检验的输入。
A/B testing: the conversion count out of n visitors follows a binomial — exactly the input significance tests need.
正态近似:n 大且 np、n(1−p) 都不太小时,可用正态近似二项,简化计算。
Normal approximation: when n is large and both np and n(1−p) are comfortably above ~5, swap in a normal approximation and the math gets easy.
一句话In One Line
二项分布回答的是离散世界里最常见的一问:n 次独立成败里,成功几次。
p 控制偏斜 —— 越靠 0.5 越对称,越偏两端越歪;n 控制平滑 ——
n 一大,根据中心极限定理,离散的柱群就趋近连续的正态钟形(这正是"正态近似二项"的来历)。
它和泊松、正态是亲戚:n 大、p 小时它趋近泊松,n 大、p 适中时它趋近正态。
掌握了它,离散计数的统计世界就有了第一块基石。
The binomial answers the most common question in the discrete world: how many successes in n independent trials.
p controls skew — closer to 0.5 means more symmetric, closer to the edges means more lopsided. n controls smoothness —
grow n and the central limit theorem drags those discrete bars toward a continuous normal bell (that's exactly the "normal approximation to the binomial").
It is a relative of the Poisson and the normal: large n, small p pushes it toward Poisson, large n, moderate p pushes it toward normal.
Master it and discrete counting has its first cornerstone.
常见误用Common Mistakes
试验之间并不独立还硬套二项。连续生产、有传染/批次相关时独立假设失效,结果会偏。
Forcing the binomial onto trials that aren't independent. Continuous production runs, contagion, or batch-to-batch correlation all break independence — the answer will be biased.
不放回抽样用二项。从有限批次不放回抽样应用超几何分布;批量远大于样本时才近似二项。
Using the binomial for sampling without replacement. When you draw without replacement from a finite lot you need the hypergeometric; the binomial is only a safe approximation when the lot is much larger than the sample.
p 在过程中会漂移却当常数。二项要求每次 p 相同;若过程不稳定,先用控制图确认 p 稳定。
Treating a drifting process p as a constant. The binomial assumes the same p on every trial — if the process isn't stable, confirm with a p-chart before you trust the math.