故事 · 伯努利和那个「该不该再赌一把」的问题 Story · Bernoulli and the "Should I Bet Again?" Question
17 世纪末,瑞士数学世家伯努利家族的 雅各布·伯努利(Jacob Bernoulli)在思考一个赌徒的老问题: 一件只有「成」与「败」两种结果的事,重复做很多次,结果到底有没有规律? 他把这种最简单的随机试验抽象出来 —— 一次试验、两种结果、成功概率固定为 p —— 后人称它伯努利试验。 更重要的是他证明了大数定律:只要次数足够多,实际成功的比例就会无限逼近真实的 p。 这听上去理所当然,却是概率论史上第一次用严格的数学保证了「频率会收敛到概率」。 今天我们说「抽样越多越可靠」「合格率长期会稳定在某个值」,根子都在伯努利这块小积木上。 In the late 17th century, Jacob Bernoulli of the famous Swiss mathematical family was chewing on an old gambler's question: if a thing has only two outcomes — win or lose — and you repeat it many times, is there any underlying regularity? He abstracted out the simplest possible random trial — one shot, two outcomes, fixed success probability p — and posterity named it the Bernoulli trial. More importantly, he proved the Law of Large Numbers: with enough trials, the observed success rate converges arbitrarily close to the true p. That sounds obvious today, but it was the first time mathematics rigorously guaranteed that "frequency converges to probability." Every time someone says "bigger samples are more reliable" or "long-run yield will stabilize", the root is this little brick.

1 抛硬币:单次随机,长期收敛到 p Flip the Coin: Random Per Trial, Converging to p in the Long Run

点「抛 1 次」开始click "Flip Once" to start

上方是当前这一次的结果(成功/失败),下方记录累计成功比例随次数的变化。 起初它上蹿下跳,但抛得越多,就越紧地贴住理论线 p —— 这就是大数定律。 Top: the result of the current flip (success/failure). Bottom: the running success rate over time. It bounces around at first, but the more you flip, the more tightly it hugs the theoretical p line — that is the Law of Large Numbers.

2 现实里的伯努利试验 Bernoulli Trials in the Real World

二元结果:合格/不良、通过/未通过、点击/未点击、到货/缺货……凡是「非此即彼」的一次性判定,都是一次伯努利试验。 Binary outcomes: pass/fail, click/no-click, in-stock/out-of-stock, delivered/missed — any single, either-or judgment is one Bernoulli trial.
成功概率 p:制程的合格率、设备的可用率、转化率,本质都是某个 p。估计 p 是测量与分析阶段的常见任务。 The success probability p: process yield, equipment availability, conversion rate — each is ultimately some p. Estimating p is bread-and-butter work in Measure and Analyze.
二项的基石:把 n 次独立同 p 的伯努利试验加起来,数成功几次,就是二项分布。伯努利是「1 次」,二项是「n 次的总和」。 Foundation of the Binomial: sum n independent Bernoulli trials with the same p and count successes — that's the Binomial distribution. Bernoulli is "one shot"; Binomial is "the total of n shots".
大数定律:样本越大,观察到的合格率越接近真实合格率。这是抽样检验、过程能力估计「为什么要够多样本」的理论依据。 Law of Large Numbers: the bigger the sample, the closer the observed yield to the true yield. This is the theoretical reason acceptance sampling and capability studies demand "enough data".
一句话In One Line
伯努利试验之所以重要,是因为它是几乎一切计数型质量度量的原子。 一件「合格还是不良」的判定就是一次伯努利试验,一批 n 件的不良数就是 n 次伯努利之和(二项分布), 单位时间内的稀有缺陷数则在 p 极小、n 极大时趋于泊松。 它还给了我们一个朴素却深刻的承诺 —— 大数定律:频率会收敛到概率。 所以六西格玛才敢用「抽样估计的合格率」去推断整批、整条产线的真实水平。 但别忘了前提:每次试验要独立、p 要稳定。若制程在漂移、批次间 p 不同,单纯的伯努利模型就会失真。 The Bernoulli trial matters because it is the atom of nearly every count-based quality metric. A single pass/fail judgment is one Bernoulli trial; the defect count across a batch of n is the sum of n Bernoulli trials (the Binomial); the count of rare events per unit time tends to Poisson when p is tiny and n is huge. It also hands us a plain but profound promise — the Law of Large Numbers: frequency converges to probability. That's why Six Sigma can confidently use a sample-estimated yield to infer the true yield of the whole batch or the whole line. But don't forget the prerequisites: trials must be independent, and p must be stable. If the process is drifting or p varies between batches, a plain Bernoulli model breaks down.
常见误用Common Mistakes
赌徒谬误:以为连出几次失败后,下一次「该轮到成功了」。每次伯努利试验都独立,过去结果不改变下一次的 p。 The gambler's fallacy: believing that after a few failures in a row, success is "due" next. Every Bernoulli trial is independent — past outcomes never change the next trial's p.
样本太小就下结论:抛 5 次出 4 次正面,就断言硬币有偏。小样本波动大,大数定律只在次数足够多时才显现,要够多样本。 Drawing conclusions from tiny samples: 4 heads in 5 flips and calling the coin biased. Small samples swing wildly. The Law of Large Numbers only kicks in with enough trials — collect more data first.
把 p 不稳定的过程硬套伯努利(制程在漂移、批次 p 各异)。伯努利要求各次独立且 p 恒定,先用控制图确认过程稳定再建模。 Forcing a Bernoulli model on a process where p isn't stable (drifting process, different p per batch). Bernoulli requires independent trials with a constant p — use a control chart to confirm the process is stable before modeling.

伯努利试验