What does Bessel's Correction · Why n − 1, Not n · LSS.WIKI explain?

Why does the sample variance divide by n − 1 instead of n? Because using the sample mean in place of the true mean systematically underestimates the spread.

Where should I continue learning?

Continue through the related NEWMETH topic pages or open the corresponding workspace tool.

Bessel's Correction · Why n − 1, Not n

抽样控制台Sampling Console

样本量 nSample size n 4

每次从总体抓 n 个，重复 3000 次取平均估计 Each trial draws n points; we average over 3000 trials

真值 σ² 固定为 100（总体标准差 σ=10）。看两种分母的平均估计离它多远。 The true σ² is fixed at 100 (population σ = 10). Watch how far each denominator's mean estimate sits from that.

真值 σ²（总体方差） True σ² (population variance)

100

Σ(xᵢ − μ)² / N · 我们要瞄准它 Σ(xᵢ − μ)² / N · the target we're aiming for

÷ n（有偏，偏低） ÷ n (biased, runs low)

—

系统性低估真值 Systematically below the true value

÷ (n−1)（无偏，对准） ÷ (n − 1) (unbiased, on target)

—

贝塞尔校正后 After Bessel's correction

故事 · 一个德国天文学家的细账 Story · A German Astronomer's Careful Bookkeeping

这条校正以德国天文学家、数学家 Friedrich Bessel 命名。19 世纪测量天体位置时，人们只有少量观测样本，却想估出测量误差究竟有多大。Bessel 注意到一个微妙之处：用样本自己的平均当作「真中心」去算偏差，偏差总会偏小 —— 因为样本均值恰恰是让平方和最小的那个点，没有任何别的中心能让偏差更小。换句话说，你已经「花掉」了一个自由度（n 个数据，1 个被均值用掉，只剩 n − 1 个能自由变动），所以分母该是 n − 1 而非 n。这一改，估计才不偏不倚。一个世纪后，这成了六西格玛 Measure 阶段每天都在用的默认设置。 The correction is named after the German astronomer and mathematician Friedrich Bessel. In the 19th century, observers nailing down stellar positions had only a handful of measurements, yet they wanted to know how noisy their instruments really were. Bessel spotted something subtle: if you center the deviations on the sample's own average, the squared sum is always going to come out too small — because the sample mean is, by definition, the point that minimizes that sum. No other center can do less. Put differently, you've already "spent" one degree of freedom on the mean — out of n points, only n − 1 are free to vary. So the denominator should be n − 1, not n. That single swap turns the estimator unbiased. A century later, it became the default setting every Six Sigma Measure phase quietly relies on.

1 ÷n vs ÷(n−1) · 谁对准了真值 ÷n vs ÷(n − 1) · Which Lands on the True Value

n=4 · 校正最关键

蓝色虚线是真值 σ²=100。橙柱（÷n）明显矮于真值 —— 系统性低估；绿柱（÷n−1）几乎贴着虚线。把 n 调到 2、3，看橙柱缩得最狠；调到 40，两者趋同。 The dashed blue line is the true σ² = 100. The orange bar (÷n) sits clearly below it — systematic underestimate. The green bar (÷(n − 1)) hugs the line. Push n to 2 or 3 and watch the orange bar collapse hardest; push it to 40 and the two bars converge.

2 自由度：n 个偏差，只有 n−1 个能自由变 Degrees of Freedom: n Deviations, Only n − 1 Are Free

每个偏差 (xᵢ − x̄) 是一根竖条，正负相抵。因为它们必须加起来等于 0（均值的定义），所以最后一根不能自由取值 —— 被前 n−1 根锁死。这就是「丢了一个自由度」的几何含义。 Each deviation (xᵢ − x̄) is one vertical bar — positives cancel negatives. Because they must sum to zero (that's the definition of the mean), the last bar can't move freely — the first n − 1 lock it in place. That's the geometry of "losing one degree of freedom".

3 现实里的 n−1 n − 1 in the Real World

实验小样本：研发只跑 5、8 个样品估波动，÷n 会让你以为过程比实际更稳，÷(n−1) 才不冒进。 Small R&D samples: estimating spread off 5 or 8 prototypes, ÷n makes the process look tighter than it is. ÷(n − 1) keeps you honest.

软件默认：Excel 的 VAR.S、STDEV.S 用 n−1（样本）；VAR.P、STDEV.P 用 N（总体）。选错分母结果就偏。 Software defaults: Excel's VAR.S and STDEV.S use n − 1 (sample); VAR.P and STDEV.P use N (population). Pick the wrong one and your answer is biased.

无处不在的自由度：t 检验、ANOVA、回归里到处是 n−1、n−k —— 同一个逻辑：每估一个参数，就花掉一个自由度。 Degrees of freedom everywhere: t-tests, ANOVA, regression — n − 1 and n − k show up everywhere. Same logic: every parameter you estimate spends one degree of freedom.

总体已知时：若真知道全体每个值（普查 / 全数检验），分母就该用 N，不必校正 —— 此时没有「顶替真均值」的问题。 When the population is known: if you actually have every value (census or 100% inspection), divide by N — no correction needed. There's no "stand-in for μ" problem to fix.

一句话In One Line

n − 1 不是凑数，是为「我们不知道真均值」付的税。真方差是相对真均值 μ 算的，可现实只有样本均值 x̄ —— 而 x̄ 恰恰是让 Σ(xᵢ − x̄)² 最小的点，所以这堆平方和天生偏小。除以更小的 n − 1，正好把缺口补平，使估计的期望等于真值，这就是无偏。它和自由度是一枚硬币两面：n 个偏差被「和为零」约束锁住一个，真正能自由变动的只有 n − 1 个。 n 越小，校正越关键；n 越大，n−1 和 n 几乎没差别 —— 这也是为什么大样本里你常常看不出区别。 n − 1 isn't a fudge factor — it's the tax you pay for "not knowing the true mean". True variance is taken about μ, but in practice all you have is the sample mean x̄ — and x̄ is precisely the point that minimizes Σ(xᵢ − x̄)², so that sum of squares is built-in too small. Dividing by the smaller n − 1 instead of n closes the gap and makes the estimator's expectation equal the truth — that's what "unbiased" means. It's the flip side of degrees of freedom: the "sum to zero" constraint locks one deviation in place, so only n − 1 are free to move. Small n? The correction is everything. Large n? n − 1 and n become nearly identical — which is why on big samples you rarely notice the difference.

常见误用Common Mistakes

样本数据却除以 n（误用总体公式）。手上是样本就用 n−1，否则系统性低估波动与风险。 Dividing sample data by n (using the population formula by mistake). Sample data → divide by n − 1. Anything else systematically underestimates spread and risk.

以为标准差 s 也是无偏的。无偏的是方差 s²；开根号后的 s 略有偏（但仍是最常用估计）。 Assuming the standard deviation s is also unbiased. It's the variance s² that is unbiased; once you square-root it, s carries a small bias (but it's still the everyday estimator).

Excel 随手抓 VAR.P / STDEV.P 算样本。样本统计一律用带 .S 的函数（分母 n−1）。 Grabbing Excel's VAR.P / STDEV.P for sample data. For sample statistics, always reach for the .S functions (denominator n − 1).