Why does the correction matter more for small samples?

The factor n/(n − 1) is larger when n is small and approaches one as n grows.

Does it fix the sample mean?

No. It adjusts the variance calculation around the sample mean.

Is normality required for unbiased sample variance?

No. Independent, identically distributed observations with finite variance suffice for the usual unbiasedness result. Some further inference procedures require additional assumptions.

AI Glossary / Computer vision and multimodal AI / Intermediate

Bessel's correction

Definition

Bessel's correction replaces division by n with division by n − 1 when estimating a population variance from n observations using their sample mean. Under independent, identically distributed sampling with finite variance, this makes the usual sample-variance estimator unbiased.

What it means

The sample mean is calculated from the same observations whose spread is being measured. It tends to sit closer to those observations than the unknown population mean would, so averaging squared deviations from it by dividing by n underestimates population variance on average.

Dividing by n − 1 compensates for that effect. The correction concerns the expected result over repeated samples; it does not make the variance from one small sample equal to the true population variance.

In image processing, it can appear when estimating noise variance at a pixel from repeated captures of an unchanged scene. The statistical assumptions still matter: motion, changing light or correlated noise can make the observed variation represent more than the intended random noise.

How it works

First calculate the sample mean. Subtract it from each observation, square the deviations and sum them. Divide the sum by n − 1, provided n is greater than one.

The deviations from the sample mean sum to zero. Once n − 1 deviations are known, the last is constrained by that relationship, providing the familiar degrees-of-freedom explanation. More formally, the expected sum of squared deviations is (n − 1) times the population variance under the usual assumptions.

Taking the square root gives the conventional sample standard deviation, but the square root is nonlinear. Bessel's correction makes variance unbiased; it does not generally make the resulting standard deviation unbiased. Other settings, such as weighted data, correlated observations or fitted regression models, require their own denominator and assumption analysis.

Three practical examples

A camera test estimates per-pixel variation from repeated aligned captures taken under stable conditions.
A quality analyst estimates process variability from a sample of independent measurements rather than a complete population listing.
A data-science team checks a software function's degrees-of-freedom setting before comparing variance calculations across tools.

Think of it this way

Think of measuring spread around a center that was itself chosen to fit the same data. That fitted center makes the data look a little tighter than they would around an independently known population center. The correction allows for that fitted quantity. The analogy is limited: n − 1 follows a precise expectation calculation, not a general rule that every estimated number always costs exactly one observation in every model.

Worked example

Take observations 2, 4 and 6. Their sample mean is 4, and the squared deviations are 4, 0 and 4, summing to 8. Dividing by n = 3 gives approximately 2.667. Applying Bessel's correction gives 8/(3 − 1) = 4.

The conventional sample standard deviation is therefore the square root of 4, which is 2. The correction increased the variance estimate by the factor n/(n − 1) = 3/2.

If these three numbers are the entire finite population whose descriptive variance is wanted, division by three can be appropriate instead. The intended quantity determines the calculation. Neither result is an arithmetic mistake merely because the denominators differ.

Illustrative case study

Fictional example: Quietframe Lab captures a stack of images of a fixed gray target to characterize repeatability. For each pixel, engineers calculate the mean and corrected sample variance across frames.

They also inspect the sequence for flickering illumination and slight camera movement. A high variance near an edge may reflect misalignment rather than sensor noise, so they do not interpret every variance value as an isolated hardware property. They record exposure settings and the number of frames used.

The lab reports the estimate and its assumptions rather than calling it exact. This fictional case illustrates a measurement workflow; it does not establish a particular camera's noise performance.

Common misconceptions and mistakes

Unbiased does not mean error-free or least variable. A corrected variance estimate can still vary substantially between small samples. The correction also does not remove outliers, measurement bias or dependence between observations.

It is incorrect to say that dividing by n − 1 makes the usual standard deviation exactly unbiased in general. It is also incorrect to use n − 1 automatically when describing the variance of a complete finite list as the population of interest. With only one observation and an unknown mean, the usual corrected sample variance is undefined because the denominator is zero.

Business use

The correction matters when reports, monitoring thresholds or model inputs rely on estimated variability. Different software defaults can create avoidable discrepancies if one calculation divides by n and another by n − 1.

State whether the goal is descriptive variance or estimation of a broader population variance. Record sample size and sampling conditions. For image stacks, check alignment and stable capture conditions before interpreting variability as noise. Where observations are weighted or dependent, use a method justified for that setting rather than applying the familiar correction without examining its assumptions.

Check your understanding

A corrected sample variance is unbiased. Does taking its square root automatically produce an unbiased estimate of population standard deviation?

Answer: No. The square-root transformation is nonlinear, so unbiasedness does not generally carry through.

Related terms

Sample variance
Population variance
Standard deviation
Degrees of freedom
Unbiased estimator
independent sampling
noise estimation
quantisation noise

Learn next

Calculate variance for the same short dataset using both denominators and explain what each result estimates. Then repeat the calculation after adding more observations. Check the degrees-of-freedom option in the software you use, and distinguish the variance estimator's bias from its uncertainty across samples.

Degrees of freedom

Sources

NIST/SEMATECH, Measures of Scale, sample variance and standard deviation formulas: https://www.itl.nist.gov/div898/handbook/eda/section3/eda356.htm Benhamou, A few properties of sample variance, research discussion of the independent-sample result: https://arxiv.org/abs/1809.03774 The three-number calculation and image-stack case are original illustrations. They do not validate a particular dataset's independence assumptions.

Previous: Bernoulli trial Next: Best bin first