The Chebyshev's theorem calculator above answers a question that sounds impossible: how much of a data set must lie close to its mean when you know nothing at all about the shape of the distribution. The answer is that at least 1 minus 1 divided by k squared of the data lies within k standard deviations of the mean, and it holds for every distribution with a finite variance — skewed, bimodal, discrete, or something nobody has a name for.
Arb Digital publishes this because that guarantee is genuinely useful and routinely forgotten. Most people reach for the empirical rule, which says roughly 68, 95 and 99.7 percent within one, two and three standard deviations. But the empirical rule is a statement about normal distributions specifically, and a great deal of real data — session durations, order values, response times, income — is not remotely normal. Chebyshev's inequality is what you have left when the normal assumption fails, and it never fails.
What This Chebyshev's Theorem Calculator Does
It evaluates the inequality in both directions. Give it a k and it returns the minimum proportion guaranteed to lie within k standard deviations, along with the maximum proportion that can lie outside. Give it a target proportion and it returns the smallest k that guarantees it. If you also supply a mean and a standard deviation, it converts the bound into an actual interval in your data's units, and a sample size turns the proportion into a guaranteed count of observations.
The bars compare the Chebyshev bound against what a normal distribution actually achieves at the same k. That comparison is the point of the page. At two standard deviations Chebyshev guarantees 75 percent while a normal distribution delivers 95.4 percent; at three, Chebyshev guarantees 88.9 percent against a normal 99.7 percent. The gap is not an error — it is the cost of a guarantee that must hold for the worst distribution imaginable, not the friendliest one.
That is exactly the boundary with our empirical rule calculator. The empirical rule assumes normality and gives tight, accurate figures when that assumption holds. Chebyshev assumes nothing and gives loose figures that are always true. Use the empirical rule when you have evidence the data are roughly normal; use this page when you do not, or when you need a bound that cannot be wrong.
How to Use It
- Enter a value of k above 1. Below that the theorem is vacuous, and the tool will say so rather than reporting a negative proportion.
- Or set a target proportion and read the required k from the fourth grid item instead.
- Add your mean and standard deviation to see the bound expressed as an interval in real units.
- Add the number of observations to turn the proportion into a guaranteed minimum count.
- Compare the bars to see how much tighter the normal figures are, and treat the difference as the value of knowing the shape of your data.
The Formula and How It Is Calculated
Chebyshev's inequality states that the proportion of values further than k standard deviations from the mean is at most 1 divided by k squared. Turned around, the proportion within k standard deviations is at least 1 − 1/k². Wolfram MathWorld's entry on the Chebyshev inequality gives the formal statement and its proof.
Work the default through. With k = 2, 1 divided by 4 is 0.25, so at least 75 percent of any data set lies within two standard deviations of its mean, and at most 25 percent lies outside. With a mean of 100 and a standard deviation of 15, that interval is 100 ± 30, or 70 to 130. Across 200 observations, at least 150 of them must fall in that range. Not approximately, not usually — at least.
The inverse direction solves 1 − 1/k² = p for k, giving k = 1 ÷ √(1 − p). For a 90 percent guarantee that is 1 ÷ √0.1 = 3.162 standard deviations. For 95 percent it is 4.472, and for 99 percent it is 10. Those numbers are strikingly larger than the familiar normal-distribution figures of 1.96 and 2.576, and the difference is worth internalising: without a distributional assumption, a 95 percent guarantee costs you more than four standard deviations rather than two.
Why the Bound Holds, in One Argument
The proof is short enough to follow without notation. Variance is the average of the squared distances from the mean. Now suppose some fraction f of the data sits at least k standard deviations away. Every one of those points contributes at least k² times the variance to that average, purely because its squared distance is at least (kσ)². Even if every other point sits exactly on the mean and contributes nothing, the average squared distance is already at least f × k² × σ².
But that average is the variance, which is σ². So f × k² × σ² cannot exceed σ², which forces f to be at most 1/k². That is the whole theorem. Notice what the argument never used: nothing about symmetry, nothing about the shape of the tails, nothing about whether the data are continuous. It only used the definition of variance, which is why the result applies to every distribution that has one.
The argument also shows why the bound is tight. It is achieved exactly by the worst case the proof imagined: a distribution with a fraction 1/k² of its mass sitting precisely at k standard deviations and the rest sitting on the mean. No inequality can be improved without excluding that distribution, which is why any tighter figure must come with an assumption attached.
When the Loose Bound Is the Right Answer
A guarantee that is far from tight still beats an estimate that might be wrong, and there are three situations where that trade favours Chebyshev clearly. The first is heavily skewed data where the normal assumption is visibly false. Revenue per customer, time on page and support ticket resolution times all have long right tails, and applying the empirical rule to them understates the proportion of extreme values badly.
The second is small samples, where you cannot check the shape convincingly. With thirty observations you cannot distinguish a mildly heavy-tailed distribution from a normal one, so a bound that does not care is safer than an assumption you cannot test. The third is any setting where being wrong is expensive — capacity planning, service-level commitments, quality limits. A promise that holds for every distribution is a promise you can make in writing. Our standard deviation calculator and variance calculator produce the inputs, and the percentile calculator gives the observed proportions to compare against.
Comparing the Bound With What Your Data Actually Do
A useful diagnostic falls out of running both figures side by side. Compute the actual proportion of your data within two standard deviations and compare it against the 75 percent floor and the normal 95.4 percent.
If your observed figure is close to 95 percent, the data are behaving roughly normally and the empirical rule is a fair description. If it sits nearer 80 percent, the distribution has heavier tails than normal, and any analysis assuming normality will underestimate how often extreme values appear. If it is below 75 percent, something is wrong with the calculation itself — a mis-specified mean, a standard deviation computed on the wrong denominator, or a data-entry error — because no data set can violate the bound. That last case makes Chebyshev a genuine validity check on your own arithmetic, which is a use most textbooks never mention. Penn State's STAT 500 applied statistics course covers the descriptive measures this comparison relies on.
What the Theorem Does Not Give You
Three limits are worth stating plainly. First, the bound is two-sided and symmetric in distance, not in probability. It says at most 1/k² of the data lies outside the interval, but it says nothing about how that fraction splits between the low side and the high side. A distribution can put all its outlying mass above the mean and still satisfy the inequality.
Second, it says nothing about the mean or the standard deviation themselves. Both are inputs. If your standard deviation is computed from a small sample, it is an estimate with its own uncertainty, and the interval inherits that. Third, it is a bound on proportions within a data set, not a probability statement about a future observation, unless the data genuinely are a random sample from the population you care about. The related but distinct question of how sample means behave is handled by our central limit theorem calculator, and individual values from a known normal population by the normal distribution calculator.
Arb Digital analyses marketing data as it actually is — skewed, spiky and long-tailed — rather than as the textbook would prefer it.
Browse All Free Tools Talk To Our TeamCommon Mistakes to Avoid
- Using it with k of 1 or less — the bound is zero or negative there and guarantees nothing at all, which is a true statement rather than a useful one.
- Reading the bound as an estimate — 75 percent is a floor, not a prediction, and most real data sets comfortably exceed it.
- Applying the empirical rule to skewed data — 68, 95 and 99.7 are normal-distribution figures and can be badly wrong on long-tailed data.
- Assuming the outliers split evenly — the theorem bounds the total outside the interval, not how it divides between the two tails.
- Forgetting the standard deviation is an estimate — on a small sample the interval carries the uncertainty of the figure that defined it.
Related Free Tools From Arb Digital
Get the tighter normal figures with the empirical rule calculator, compute the spread with the standard deviation calculator, standardise a value with the z-score calculator, check where a value sits with the percentile calculator, or work with sample means using the central limit theorem calculator. The full free online tools hub lists every statistics tool we publish.
Frequently Asked Questions
That at least 1 minus 1 divided by k squared of any data set lies within k standard deviations of the mean. It holds for every distribution with a finite variance, whatever its shape.
At least 75 percent, for any distribution. A normal distribution actually delivers about 95.4 percent, so the bound is a floor rather than an estimate.
Because 1 divided by 1 squared is 1, so the bound becomes zero percent. The inequality is true but empty there, which is why it is only used above one standard deviation.
The empirical rule assumes the data are normally distributed and gives tight figures. Chebyshev assumes nothing about the shape and gives looser figures that are always correct.
About 4.47, using k equal to one divided by the square root of one minus the proportion. That is far more than the 1.96 a normal distribution would need.
No. If your data appear to violate it, the mean or the standard deviation has been computed incorrectly, which makes the theorem a useful check on your own arithmetic.
Both. It is a statement about any collection of numbers with a finite variance, so it applies equally to a full population and to the observations in a sample.
This page explains a statistical result for educational purposes. The bound is a guaranteed minimum rather than an estimate, and it depends on a mean and a standard deviation that have been calculated correctly.
