故事 · 一勺汤与一次大选 Origin Story · A Spoon of Soup and a US Election
厨师从不会喝完整锅汤来尝咸淡 —— 只要搅匀,舀一勺,这一勺就能代表整锅. 1936 年美国大选,《文学文摘》寄出 1000 万张问卷、收回 230 万份,预测兰登大胜; 而盖洛普只访问了几千名精心抽取的选民,却准确预测罗斯福连任. 《文摘》样本虽大却有偏(只问有电话、有车的富人),盖洛普样本虽小却有代表性. 这告诉我们:抽样的关键不在「多」,而在「准」与「随机」.一勺没搅匀的汤,舀一桶也没用. 这正是六西格玛 Measure 阶段的基石:用对的样本,推断看不见的总体. No chef drinks the entire pot to check the seasoning — stir well, take one spoonful, and that spoonful speaks for the whole pot. In the 1936 US presidential election, Literary Digest mailed out 10 million surveys and got 2.3 million back, predicting a Landon landslide; Gallup polled only a few thousand carefully sampled voters and correctly called Roosevelt's re-election. Literary Digest's sample was huge but biased (only people with phones and cars); Gallup's sample was small but representative. The lesson: good sampling is about being right and random, not about being large. A bucket of un-stirred soup is still un-stirred soup. This is the bedrock of the Six Sigma Measure phase: a good sample is how we infer an invisible population.

1 样本均值围着真值跳,n 越大跳得越小 Sample Means Dance Around the Truth — Bigger n, Smaller Dance

波动较大

灰色是总体分布(真相),蓝色虚线是真均值 μ.每抽一个样本,就得到一个橙色的样本均值 x̄ —— 它落在 μ 附近但不正好等于 μ,这就是抽样误差.把 n 调大,橙点会越来越挤在 μ 周围.

2 误差按 √n 收窄:想准一倍,得多花四倍样本 Error Shrinks With √n: Twice the Accuracy Costs 4× the Samples

标准误 = σ/√n,是条先陡后平的曲线.从 n=4 加到 n=16,误差减半;想再减半得加到 n=64. 抽样精度有「报酬递减」 —— 这就是为什么无限加样本量并不划算. SE = σ/√n is a steep-then-flat curve. Going from n=4 to n=16 halves the error; halving it again takes n=64. Sampling accuracy has diminishing returns — which is why endlessly piling on samples stops being worth it.

3 现实里的样本与总体 Sample & Population in the Real World

抽样检验:一批 10 万件不可能全检,按 AQL 抽几十件推断整批合格率,省成本又够准 —— 前提是随机、有代表性. Acceptance sampling: A 100k-unit batch can't be 100% inspected. Per AQL, you pull a few dozen and infer the batch's defect rate — cheap and accurate, as long as the pull is random and representative.
x̄ 估 μ · s 估 σ:希腊字母(μ、σ)是总体参数(真值、未知);拉丁字母(x̄、s)是样本统计量(算得出、估计值).分清楚,公式才不乱. x̄ for μ · s for σ: Greek letters (μ, σ) are population parameters — the true, unknown values. Latin letters (x̄, s) are sample statistics — what you actually compute. Keep them apart and your formulas stay sane.
抽样误差不可消除:只要不普查,每个样本都会和真值有差.它不是「错误」,而是抽样固有的随机性,只能靠加大 n 来缩小. Sampling error never goes away: short of a full census, every sample misses the true value by something. It's not a "mistake" — it's the built-in randomness of sampling, and only larger n can shrink it.
为何不普查:成本高、耗时长,破坏性检测(测灯泡寿命、做拉伸试验)更会把产品全毁掉.样本推断常是唯一可行的路. Why not a full census: cost, time, and — for destructive tests like bulb lifetime or tensile testing — the inconvenient fact that 100% inspection destroys 100% of the product. Sample-based inference is often the only path.
一句话In One Line
整个推断统计就建立在一句话上:用算得出的样本,去猜看不见的总体. 样本均值 x̄ 是对总体均值 μ 的估计,它本身是个随机变量 —— 每抽一次就抖一下,抖动幅度由标准误 σ/√n 度量. 关键洞见有两条:一是样本越大越准,但按 √n 报酬递减,盲目堆样本不经济; 二是代表性比数量更重要,一个有偏的大样本,远不如一个随机的小样本. 搅匀了再尝那一勺,才能代表整锅汤. All of inferential statistics rests on one line: use the sample you can compute to guess the population you can't see. The sample mean x̄ is an estimate of the population mean μ — itself a random variable that wobbles each draw, with the wobble measured by the standard error σ/√n. Two takeaways: first, bigger samples are more accurate, but √n gives diminishing returns — blindly piling on samples is uneconomical; second, representativeness beats raw size — a biased megasample is far worse than a small random one. Stir the pot, then taste — that spoonful is the only one that speaks for the whole soup.
常见误用Common Mistakes
以为样本均值 x̄ 就等于真均值 μ.x̄ 只是估计,必带抽样误差;要谈把握得给区间(置信区间). Treating the sample mean x̄ as if it equals the true mean μ. x̄ is just an estimate and always carries sampling error — if you want a confidence claim, give a confidence interval.
用有偏样本(便利抽样、自选)冒充随机.样本必须随机且有代表性,否则样本再大也系统性偏离真值. Passing off a biased sample (convenience or self-selected) as random. Samples must be random and representative — otherwise no amount of size will fix the systematic gap from the truth.
把总体参数 σ 和样本统计量 s 混用符号.μ/σ 是总体(未知真值),x̄/s 是样本(算得的估计),严格区分. Mixing population parameter σ with sample statistic s. μ/σ are the population truth (unknown); x̄/s are the sample estimates (computed). Keep the notation strict.

样本与总体