故事 · 高斯、误差与一颗找不回的小行星
Origin Story · Gauss, Errors, and a Lost Asteroid
1801 年，天文学家发现了小行星谷神星，却在它转到太阳背后前只测到几个零散位置，眼看就要永远丢失。
24 岁的德国数学家 高斯（Gauss）挺身而出：他假设每次观测的误差服从一条对称的钟形曲线 ——
小误差最常见、大误差越来越罕见 —— 用这条曲线把噪声里的真值「挤」了出来，成功预言谷神星会在哪里重现。它真的回来了。
这条曲线后来被叫做高斯分布，也就是正态分布。它的伟大之处在于：当大量微小、独立的随机因素叠加在一起，
结果几乎总会收敛成它（这是中心极限定理，下一页的高尔顿板会让你亲眼看到）。所以身高、误差、零件尺寸才都长成同一副模样。
In 1801, astronomers spotted the dwarf planet Ceres but lost track of it as it swung behind the Sun, with only a handful of scattered position measurements in hand.
A 24-year-old German mathematician — Carl Friedrich Gauss — stepped up. He assumed each observation's error followed a symmetric bell-shaped curve:
small errors common, large ones rare. Using that curve he squeezed the true position out of the noise and predicted exactly where Ceres would reappear. It did, on schedule.
That curve became known as the Gaussian, the normal distribution. Its power: whenever many small, independent random factors stack up,
the sum almost inevitably collapses onto this shape (that's the central limit theorem — the Galton board on the next page makes it visible).
That's why heights, errors, and part dimensions all wind up looking like the same bell.
1 钟形曲线：μ 移位置，σ 调胖瘦 The Bell Curve: μ Slides, σ Stretches
68-95-99.72 真实样本：直方图自己长成钟形 Real Samples: The Histogram Grows Into a Bell on Its Own
按当前 μ、σ 随机撒 600 个样本，统计每个区间落了几个。样本量越大，直方图就越贴近那条理论钟形 —— 理论与现实在这里握手。点左边的按钮再撒一批试试。 Draw 600 random samples from the current μ and σ and count how many land in each bin. The larger the sample, the closer the histogram hugs the theoretical bell — this is where theory and reality shake hands. Hit the button on the left to draw another batch.
3 现实里的正态分布 The Normal Distribution in the Wild
正态假设：t 检验、ANOVA、控制图、过程能力，绝大多数经典统计方法都默认数据近似正态。用之前先验正态性（正态概率图、检验）。
The normality assumption: t-tests, ANOVA, control charts, capability indices — almost every classical statistical tool quietly assumes the data is roughly normal. Always check normality first (normal probability plot, formal test) before you use them.
68-95-99.7：一眼判断数据是否异常的速算尺。某零件超出 ±3σ，几乎可断定不是随机波动，而是出了特殊原因，该查。
The 68-95-99.7 rule: a back-of-the-envelope ruler for spotting outliers. A part landing outside ±3σ is almost certainly not random jitter — it's a special cause and worth investigating.
过程能力的基础：Cp、Cpk、Pp、Ppk 全部用 σ 衡量制程的「胖瘦」与「居中」程度。正态是它们成立的前提。
The basis of process capability: Cp, Cpk, Pp, and Ppk all use σ to gauge how lean and how well-centered a process is. Normality is the precondition that keeps them valid.
自然界的钟形：身高、体重、血压、测量误差、考试成绩……大量独立小因素叠加的结果，几乎都鼓成钟形。这不是巧合，是中心极限定理。
Bells in nature: heights, weights, blood pressure, measurement errors, exam scores — anything that's the sum of many small independent factors tends to bulge into a bell. Not a coincidence — that's the central limit theorem at work.
一句话In One Line
正态分布是六西格玛的公用坐标系。它把「质量好不好」翻译成一句可量化的话：
数据离中心 μ 有几个 σ。「六个 σ」意味着规格界限离均值有 6 倍标准差那么远 ——
即便制程中心长期偏移 1.5σ，缺陷率也只有百万分之 3.4。
整套 DMAIC 的度量、控制图、能力指数，本质都是在数「几个 σ」。
但务必记住它的前提：数据得近似正态。偏态、有界、计数型数据硬套正态，结论会失真 —— 那是下一站要讲的「分布家族」。
The normal distribution is Six Sigma's shared coordinate system. It turns "is quality good?" into a single quantifiable question:
how many σ does the data sit from μ? "Six sigma" literally means the spec limits are six standard deviations from the mean —
so even if the process center drifts 1.5σ over the long run, the defect rate is still only 3.4 per million.
DMAIC's metrics, control charts, and capability indices are all, at heart, counting σ.
Just remember the precondition: the data has to be approximately normal. Forcing skewed, bounded, or count data into a normal model warps the conclusions — that's what the "distribution family" page tackles next.
常见误用Common Mistakes
默认一切数据都正态，不做检验直接套 t 检验、Cpk。先用正态概率图或正态性检验确认，偏态数据要先变换或换方法。
Assuming every dataset is normal and reaching straight for a t-test or Cpk. Confirm with a normal probability plot or a normality test first; transform skewed data or pick a method that doesn't assume normality.
把 σ 当成「越小越好」的孤立指标。σ 只描述离散程度，还得结合中心 μ 是否对准目标才能判断制程好坏。
Treating σ as a "smaller is better" standalone metric. Sigma only describes spread — judge process health together with how well μ is centered on the target.
用 68-95-99.7 去套明显有界或计数的数据（如不良品个数、等待时间）。那些该用泊松、指数等分布，正态法则不适用。
Applying the 68-95-99.7 rule to clearly bounded or count data like defect counts or waiting times. Reach for Poisson, exponential, or another appropriate distribution — the normal rule doesn't apply.