Can a variance be smaller than the standard deviation?

Yes, whenever the standard deviation is between 0 and 1, because squaring a fraction makes it smaller. A standard deviation of 0.5 corresponds to a variance of 0.25. That pairing surprises people who think the variance is always the larger, more serious number. Compare them only after you remember the units. One is in squared units and the other is not, so the inequality is not a ranking of importance.

Should I ever add standard deviations instead of variances?

Add them only when you are building a hard bound you refuse to treat as independent noise. Label that sum as a bound. For independent contributions, add the variances and take the square root once. Adding the standard deviations of 3 and 5 to get 8 overstates the spread of a sum, which is the square root of 34, about 5.83. Keep the unit, the count, and the name of the formula in the same sentence as any figure you copy into a report.

Does a variance of zero mean every value matched the mean?

Yes. A sum of squared gaps is zero only when every gap is zero, so every included value equals the mean. It does not mean the mean was zero. A list of 7, 7, and 7 has variance 0 and mean 7. A row you forgot to type can change that. If the missing row would have differed, the variance of the full list is not the one you computed.

Why not publish the variance to the public and skip the root?

Readers add and compare in the original unit. A variance of 4 square minutes is easy to hear as 4 minutes. The standard deviation of 2 minutes matches the way they already talk about the readings. Publish the root for the audience, and keep the variance in the appendix if a later calculation needs the squares. The translation is cheap, and the misunderstanding is not.

How do I recover a variance from a published standard deviation?

Square it, and record whether that standard deviation used n or n minus 1 so the variance inherits the right label. Squaring 2 gives the population variance 4 on the eight-reading example. Squaring the sample standard deviation of about 2.14 gives about 4.57, the sample variance. If the source never said which divisor it used, you cannot recover a variance you should pool. Ask, or treat the figure as unlabeled.

Variance Is the Squared Language of Spread

The square root is a translation

From the eight readings 2, 4, 4, 4, 5, 5, 7, and 9, the squared gaps from the mean of 5 add to 32. Divide by 8 and the population variance is 4. The standard deviation is the square root of 4, which is 2. Both numbers describe the same spread, and they are not interchangeable in a sentence. A variance of 4 does not mean a typical reading sits 4 units from the mean. It means the average squared gap, using the population divisor, is 4, and the root of that average is 2 units, back in the original measure.

Speak the 2 to a colleague who will compare the figure with the readings. Keep the 4 if the next operation is algebraic. Squared units are awkward in a meeting and useful in a sum. The Variance Calculator returns the squared figure and leaves the square root as a separate step, which is the right split when you are about to pool or add. If you only needed a readable distance, take the root before the number leaves your notes, and label the unit. Minutes in, minutes out, after the root and not before.

Why standard deviations do not average

Two machines have standard deviations of 3 and 5 in the same unit. The average of those two numbers is 4, and it does not describe a combined process. Square them first: the variances are 9 and 25. If the machines are independent and you have a reason to treat them as equal contributors, the average variance is 17, and the combined standard deviation is the square root of 17, about 4.12. The naive average of 4 understated the spread. The error is small here and much larger when one standard deviation is near zero and the other is not.

Adding the standard deviations themselves, 3 plus 5 equals 8, is a worse shortcut. Independent noise does not stack edge to edge like two rulers. It stacks in the squares. Use 8 only if you are adding a worst-case bound, and call it a bound, not a standard deviation. For a tolerance stack on independent sources, add the variances and then take one square root at the end. On this pair, the root of 9 plus 25 is the root of 34, about 5.83, which is the standard deviation of a sum, a different question from the average of the two variances.

Sample variance beside the population figure

Divide the same sum of squares, 32, by 7 instead of 8 and the sample variance is about 4.57. The sample standard deviation is the square root, about 2.14. People sometimes square a published standard deviation to recover a variance and forget which divisor produced it. Squaring 2 gives 4, the population variance. Squaring 2.14 gives about 4.58, close to 32 divided by 7. If you square the wrong one, every later pooled test inherits the mistake. Store the sum of squares and the divisor so the variance can be rebuilt without guessing.

The gap between 4 and about 4.57 is the entire content of the n versus n minus 1 choice, seen in squared units. It looks more dramatic as a variance than as a standard deviation, because the square root compresses the ratio. Do not switch units in the middle of an argument to make a process look stable. If a vendor quotes a variance, ask whether the divisor was n or n minus 1 before you compare it with yours. A 14 percent gap in the variance, 4.57 against 4, is about a 7 percent gap after the root, and either can be the whole difference between two quotes.

A budget error in squared money

Suppose a weekly cost has a variance of 25 square dollars around its mean, using whatever divisor you already declared. The standard deviation is 5 dollars. A manager who reads the variance as 25 dollars of typical error will pad the budget by 25 instead of by a small multiple of 5. On a week whose mean cost is 80 dollars, a 5-dollar standard deviation is a modest wobble, and a 25-dollar pad is a different plan. Translate to dollars before the number reaches a budget line. Leave the 25 in the working papers where the squares still have a job.

If two independent cost lines have variances 25 and 9, the variance of the total is 34, and the standard deviation of the total is about 5.83 dollars, not 5 plus 3. A contingency built by adding the two standard deviations sets aside 8 dollars of typical combined movement, which is a looser bound than the root-sum-square. Say which one you used. For a single week that must not surprise you, a bound can be right. For a year of independent weeks, the root-sum-square matches the way independent variances accumulate, and the sum of the standard deviations will overstate the pile.

When the squared figure should stay

Keep the variance when you are pooling groups that you believe share one spread. A group of 10 with sample variance 4 and a group of 6 with sample variance 16 pool by degrees of freedom: 9 times 4 plus 5 times 16, divided by 14. That is 36 plus 80, which is 116, divided by 14, about 8.29. You cannot get that number by averaging the standard deviations of 2 and 4. The weights live on the variances. Convert to a standard deviation only after the weighted average exists, if you still need a figure in the original unit.

Keep the variance, too, when you are comparing a model with a residual. A residual variance of 9 against a total variance of 36 means a quarter of the squared variation is still unexplained, a statement that is natural in squared units and clumsy after two separate square roots. Report that fraction, or report both variances. If a reader needs the leftover in original units, the residual standard deviation is 3. Give them 3 and also the fraction, so the share of the problem that remains is not lost in the translation back from squares.

A checklist before you publish a variance

Write the unit as square units, write n, and write the divisor. For the eight readings, a clean line is that the population variance is 4 square units on n equals 8, from a sum of squares of 32. The matching standard deviation is 2 units. Anyone can test the square root. Anyone can test the division. A variance published as 4, with no unit and no n, invites the 25-dollar misreading and the divisor mix-up at the same time. The extra words are the whole safety device. Keep the unit, the count, and the name of the formula in the same sentence as any figure you copy into a report.

Then ask whether the mean is a sensible center for squares. A single spike dominates a variance even more than it dominates a mean, because the gap is squared. On a list whose gaps include a 13 and several 2s, the 13 contributes 169 to the sum of squares and each 2 contributes 4. One point can own the variance. If that point is a typo, the variance is a typo. Check the largest absolute gap before you quote the square. If the point is real, say so, and consider a resistant spread alongside the variance so the spike is visible as itself.

Open the worksheet

The matching page is the Variance Calculator. It opens with a worked example, names the formula, and keeps the steps next to the result. Use it when you want the arithmetic on your own numbers rather than the argument on this page.