A population is every unit you could ever measure. A sample is the handful you actually did. That's a difference in KIND, not just size — and it's exactly why sample standard deviation needs a correction that population standard deviation never does.
A population is the complete set of everything you actually care about — every unit ever made, every measurement that could ever be taken. A sample is the smaller subset you actually measured or collected. That single distinction is why the sample standard deviation formula divides by (n−1)instead of n — a detail called Bessel's correction that trips up almost everyone who first meets it, because it looks like it should just be a typo. It isn't. It exists because a sample's own mean is, by construction, the single number that minimizes squared deviations within that specific sample — which means the sample slightly under-represents how spread out the real population actually is. Dividing by (n−1) instead of n corrects for that built-in bias.
The population mean (μ) and population standard deviation (σ) are true, fixed numbers — they describe the actual, complete set of everything you care about, whether that set is 12 units or 12 million. In almost every real situation you can't ever know μ and σ exactly, because measuring the entire population is impractical or impossible (it might mean destructively testing every part you make). The sample mean (x̄) and sample standard deviation (s) are what you actually calculate from the units you did measure — and because they're built from only part of the picture, they are estimates, carrying their own uncertainty, that would come out slightly different if you happened to draw a different sample. That's exactly why statisticians insist on separate symbols for each: μ/σ for a true, fixed parameter you're trying to pin down; x̄/s for your best guess at it, computed from data.
Because x̄ is only an estimate of μ, computing spread the "obvious" way — measuring every sample point's distance from x̄, squaring, and dividing by n — has a subtle problem. The sample's own mean x̄ is, by construction, the single value that makes those squared distances as small as they can possibly be for that specific set of numbers. Any other candidate center, including the true (unknown) population mean μ, would produce squared distances that are equal or larger. So measuring spread around x̄ instead of the real μ systematically comes out a little too small — dividing by n underestimates the true population variance.
It's a basic property of the arithmetic mean: for any fixed set of numbers, the value that minimizes the sum of squared distances to all of them is that set's own mean — no other number, including the true population mean μ, can do better. So when you compute spread using x̄ instead of μ, you're measuring against the one point guaranteed to make the spread look as small as possible for that sample. Do this across every possible sample you could draw, and the shortfall is not random noise that cancels out — it's a consistent, one-directional bias, always pushing the estimate low. Dividing by (n−1) instead of n inflates the result by precisely the right amount to cancel that bias: not enough to overcorrect, not so little that the bias survives. That's the "n−1 degrees of freedom" you'll see in textbooks — one degree of freedom is "used up" by needing x̄ itself, calculated from that same data, before you can measure any deviations from it at all. It won't make any single sample's s² land exactly on σ² — no correction can guarantee that — but averaged over every sample you could ever draw, ÷(n−1) is unbiased and ÷n never is.
False — and the size comparison is the whole reason this stays confusing. Population and sample differ in kind, not headcount. A population can be small: all 12 gas turbines installed at one specific power plant is a complete population — measure every one of them, and their true mean and true standard deviation are exact, known numbers with zero estimation error, because nothing was left out. A sample can be enormous: a national poll of 50,000 voters is still only a sample of the full electorate, and it still carries real sampling error, no matter how large that 50,000 gets — unless it happens to equal every single voter. The defining question is never "how many?" — it's "did this capture every unit that exists, or only some of them?" That's precisely why statisticians reserve separate symbols for each: μ and σ for a population — a true, fixed parameter, sometimes literally unknowable in practice — and x̄ and s for a sample — your best calculated guess at it, always paired with an implicit "give or take." Swap the symbols, or treat a big sample as if it were a population, and you erase the fact that one of these is a settled truth and the other is still an estimate with its own uncertainty.
Explains the difference between a population (the complete set of everything you actually care about) and a sample (the smaller subset you actually measured), why population parameters (μ, σ) are true fixed values while sample statistics (x̄, s) are estimates with their own uncertainty, and why that distinction is exactly why sample standard deviation divides by (n−1) instead of n — Bessel's correction.
The word "population" sounds like it just means "lots of things," so it is natural to assume a population is simply a very large sample. It is not. A population is defined by completeness — it is every unit or every measurement that exists for whatever you are studying, however many or few that happens to be. A sample is defined by being a subset — some, but not all, of that complete set, gathered because measuring literally everything is impractical, expensive, or physically impossible (as with destructive testing). Population parameters (μ, σ) describe the complete set exactly; sample statistics (x̄, s) are calculated approximations of those parameters, and they carry sampling uncertainty that never fully disappears no matter how careful the measurement is.
Sample variance is defined as s² = Σ(xᵢ − x̄)² / (n − 1), not divided by n as the "average of squared deviations" pattern would suggest. The reason is that x̄ is calculated from the same data whose spread is being measured, and the arithmetic mean has a built-in mathematical property: it is the single value that minimizes the sum of squared deviations for its own dataset. Measuring deviations from x̄ instead of the true, unknown population mean μ therefore produces a sum of squared deviations that is systematically too small, on average, across every sample you could draw. Dividing by (n − 1) instead of n inflates the result by exactly the factor needed to cancel that bias, making s² an unbiased estimator of σ² — meaning that if you repeated the sampling process many times and averaged all the resulting s² values together, they would converge on the true σ², whereas averaging the ÷n version would converge on a value that is too low.
Quality control and destructive testing: when testing a bolt, weld, or material sample means destroying it, the full population can never be measured directly — every reported mean and standard deviation from a test batch is a sample statistic, x̄ and s, standing in for the unknowable true μ and σ of the entire production run. Statistical process control: control limits and capability indices (like Cpk) are built from sample statistics collected from a process, and using the wrong divisor (n instead of n−1) understates the true process variation, potentially passing a process that is actually less capable than it appears. Research and reported data: any dataset that is not literally every unit in existence — survey respondents, test specimens, sensor readings over a finite time window — is a sample, and its summary statistics should be read as estimates with attached uncertainty, not as the final, true answer.
No. Size alone never converts a sample into a population. What matters is completeness: did the million data points include every single unit that exists, or only a very large fraction of them? A sample of one million voters out of an electorate of 150 million is still a sample, and it still has real sampling error — just usually much smaller sampling error than a sample of 100 would have. A population, by definition, means nothing was left out.
Yes, and this is exactly what breaks the "population = big" intuition. If you care about all 12 gas turbines installed at one specific plant, and you measure every single one, that is a complete population of size 12. Its mean and standard deviation are the true μ and σ — exact numbers, not estimates — precisely because nothing was excluded, regardless of how small the total count is.
It falls directly out of the math of expected values: it can be shown that the expected value of Σ(xᵢ − x̄)² equals (n−1)·σ², not n·σ². Dividing by (n−1) is therefore the exact factor that makes the expected value of s² equal σ² itself — that is the formal definition of an unbiased estimator. Any other divisor would leave a systematic over- or under-correction behind.
No — unbiasedness is a statement about averages across many hypothetical samples, not a guarantee about any one sample. A single sample's s can still land above or below the true σ purely by chance, especially with small n. What (n−1) guarantees is that there is no systematic direction to that error: it does not consistently run low the way the ÷n version does.
If every unit in the population was genuinely measured — nothing held out, nothing estimated — divide by n. At that point there is no sample mean standing in for an unknown population mean; x̄ and μ are the same calculation on the same complete data, so the bias that (n−1) exists to correct never arises in the first place.
Try our STEM Learning Studio
More calculators, simulators, and guides for this discipline.