故事 · 一个家族,各管一摊
Story · One Family, Different Jobs
想象统计学是一个工具箱,里面不是只有一把锤子。问「100 个零件里合格几个」,那是二项的活;
问「这条产线一天出几次故障」,那是泊松;问「下一次报修还要等多久」,找指数;
问「这批轴的直径分布」,多半是正态;要是「随便摇一个 0 到 9 的数」,那就是均匀。
用错成员,再精密的计算也会算出荒谬的结论 —— 比如用正态去算「不良品个数小于零的概率」,纯属南辕北辙。
更有意思的是,这些成员之间有血缘:二项试验做得够多就长得像正态,事件够稀有就长得像泊松。
认全这张家谱,建模时才能一眼挑对那张「形状」。
Think of statistics as a toolbox — not just one hammer. "How many of 100 parts pass?" is Binomial work.
"How many faults per day on this line?" → Poisson. "How long until the next service call?" → Exponential.
"What's the diameter spread on this batch of shafts?" → almost always Normal. "Pull a random integer 0–9?" → Uniform.
Use the wrong member and even precise math gives nonsense — like asking a Normal model for "the probability of a negative defect count".
And the members are related by blood: enough Binomial trials look Normal, rare-enough events look Poisson.
Once you know the family tree, picking the right shape becomes a glance, not a guess.
1 每位成员长什么样、管什么事 What each member looks like, and what it covers
continuous2 家谱:成员之间的近似关系 Family Tree: Limit Approximations Between Members
箭头表示极限近似:二项当 n 大时趋于正态、当 p 小时趋于泊松;泊松事件密集时也趋于正态。 当前高亮的成员,会点亮与它相关的连线。 Arrows show limit approximations: Binomial → Normal when n is large, Binomial → Poisson when p is small, Poisson → Normal when events are dense. The highlighted member lights up the edges that involve it.
3 现实里的分布选型 Picking the Right Distribution in Practice
连续 vs 离散:能取小数、量出来的(尺寸、时间)多是连续型;数出来的整数(个数、次数)是离散型。这是选分布的第一道分叉。
Continuous vs discrete: anything you can measure (size, time) is usually continuous; anything you count (units, occurrences) is discrete. That's the first fork in the road.
看问题在问什么:问「成败几个」→二项;问「单位时间几次」→泊松;问「等多久」→指数;问「测量分布」→正态。先听清问题,再挑分布。
Read the question carefully: "how many pass/fail?" → Binomial. "How many per unit time?" → Poisson. "How long until?" → Exponential. "Measurement spread?" → Normal. The question picks the shape.
用关系简化计算:二项 n 大、p 不极端时可用正态近似算概率;p 极小、n 极大时用泊松近似。近似让手算可行。
Use the kinship to simplify: Binomial with large n and moderate p can be approximated by Normal. With tiny p and huge n, use the Poisson approximation. Approximations make hand calculation tractable.
建模第一步:拿到数据先判断它属于哪类现象,选对分布,后面的控制图类型、能力指数公式、假设检验才会成立。
Modeling step zero: classify the phenomenon first. Get the distribution right, and the control-chart type, capability formula, and hypothesis test all line up correctly. Get it wrong and everything downstream is invalid.
一句话In One Line
分布不是一堆要背的公式,而是现实现象的「形状字典」。每种随机现象都有一张对应的形状,
选对了,均值、方差、概率、控制图、能力指数全都顺理成章;选错了,整套分析从地基就歪了。
六西格玛 Measure 与 Analyze 阶段最常踩的坑,就是不分青红皂白地把一切都当正态。
记住分类的钥匙:是连续还是离散?在问个数、次数、等待,还是测量?
答清这两问,再借助家族成员之间的近似关系,你就能在真实数据面前,迅速挑出那张对的形状。
Distributions aren't a stack of formulas to memorize — they're a shape dictionary for real-world randomness. Each phenomenon has one shape.
Pick the right one and mean, variance, probabilities, control charts, and capability indices all click into place. Pick the wrong one and the whole analysis is crooked from the foundation up.
The classic Measure/Analyze pitfall is treating everything as Normal by default.
Remember the two keys: continuous or discrete? counting, occurrence, waiting, or measurement?
Answer those two and use the family relations — and you'll pick the right shape on sight.
常见误用Common Mistakes
把计数数据(不良品个数)当正态去算概率。计数型该用二项或泊松,正态只适合连续测量。
Treating count data (defect counts) as Normal when computing probabilities. Counts belong to Binomial or Poisson — Normal is for continuous measurement.
忽视分布有界,用正态算出「负的等待时间」「超过 100% 的合格率」。等待时间用指数、比例用二项,它们天然不会越界。
Ignoring natural bounds — using Normal to derive "negative waiting time" or "yield above 100%". Use Exponential for waiting time, Binomial for proportions — they respect the bounds by construction.
滥用近似,p 接近 0.5 还硬用泊松近似二项。近似有条件:泊松近似要 p 极小、n 极大;正态近似要 np 与 n(1-p) 都够大。
Abusing approximations — forcing a Poisson approximation on Binomial when p is near 0.5. Approximations have conditions: Poisson needs very small p and very large n; Normal needs both np and n(1-p) to be large enough.