Chebyshev's Inequality: A Distribution-Free Bound on Spread

The 68-95-99.7 rule tells you exactly how much of a normal distribution lies within 1, 2, or 3 standard deviations of the mean — but it only works if the data is actually normal. Chebyshev's inequality answers a harder question: without assuming anything at all about the shape of a distribution, how much can you still guarantee? The answer turns out to be a real, usable number — just a more conservative one.

The Formula

For a random variable with any distribution, finite mean μ, and finite standard deviation σ, and for any constant k > 1:

Equivalently, phrased as a guaranteed minimum rather than a maximum:

No assumption about symmetry, unimodality, or shape appears anywhere in this statement — it holds for a wildly skewed distribution, a distribution with several separate peaks, or a distribution that's mostly one single spike with a thin long tail. That generality is exactly the tradeoff: the bound is real and always safe to use, but it's far looser than what a specific distribution like the normal would actually give you.

Worked Example

Problem: A factory's output for some measured quantity has mean μ = 500g and standard deviation σ = 5g. Nothing is known about the shape of the distribution — it might not be normal at all. What is the guaranteed minimum proportion of output within 490g–510g (that is, within k = 2 standard deviations of the mean)?

At least 75% of output is guaranteed to fall between 490g and 510g, regardless of the true shape of the distribution. Compare that to a normal distribution's 95% for the same 2-standard-deviation window (from the 68-95-99.7 rule) — Chebyshev's bound is real and universally valid, but considerably more conservative, precisely because it has to hold even for distributions far stranger than a bell curve.

Is the Bound Ever Actually Tight?

It's reasonable to wonder whether 1/k² is just a loose, worst-case number that no real distribution comes close to. It isn't — the bound is exactly achievable. Consider a three-point distribution with mean 0 and standard deviation σ: a value of −kσ with probability 1/(2k²), a value of +kσ with probability 1/(2k²), and a value of 0 with probability 1 − 1/k². For k = 2, that's ±2σ each with probability 0.125, and 0 with probability 0.75 (0.125 + 0.125 + 0.75 = 1, confirming a valid distribution).

Checking the variance: E[(X − 0)²] = 0.125 × (2σ)² + 0.125 × (2σ)² + 0.75 × 0² = 0.25 × 4σ² = σ² — exactly matching the stated standard deviation σ, as required. And P(|X| ≥ 2σ) here is exactly 0.125 + 0.125 = 0.25, which is exactly 1/2² = 1/4 — the Chebyshev bound holding with equality, not just as a safe overestimate. This confirms the inequality is the tightest possible guarantee that can be made without further information about shape; no smaller bound could be claimed to hold for every distribution.

Where the Formula Comes From

Chebyshev's inequality follows directly from a simpler, more general result called Markov's inequality: for any non-negative random variable Y with finite mean, and any a > 0, P(Y ≥ a) ≤ E[Y]/a. The derivation applies this to Y = (X − μ)², which is always non-negative:

The final step is just algebra: the event (X − μ)² ≥ k²σ² is exactly the same event as |X − μ| ≥ kσ (squaring preserves the inequality for non-negative quantities on both sides), so P(|X − μ| ≥ kσ) ≤ 1/k² follows immediately. This is a common pattern in probability theory — a specific, well-known inequality turning out to be a particular application of a more general one, rather than needing its own separate proof machinery.

The One-Sided Version: Cantelli's Inequality

Chebyshev's inequality bounds deviation in either direction at once. Often only one direction matters — for instance, "how much of the distribution could plausibly be at least kσ above the mean," with no interest in the lower tail. For this one-sided question, a sharper, specifically-derived bound applies, known as Cantelli's inequality:

For k = 2, this gives 1/(1 + 4) = 1/5 = 0.20 — meaningfully sharper than simply halving the two-sided Chebyshev bound (which would naively suggest 0.25/2 = 0.125, an even smaller and, in this case, an incorrect number to use, since Chebyshev's two-sided bound cannot be split evenly between the two tails without further assumptions). Cantelli's inequality is the correct tool specifically when only one tail is of interest and no symmetry can be assumed.

Real-World Uses

Manufacturing and quality control: exactly as in the worked example above, Chebyshev's inequality gives a guaranteed minimum yield-within-tolerance figure even when a process's output distribution hasn't been confirmed to be normal — useful for a conservative worst-case commitment before enough data has accumulated to check the actual shape.

Proving the Law of Large Numbers: the standard textbook proof of the weak law of large numbers applies Chebyshev's inequality directly to the sample mean, whose variance shrinks as σ²/n — plugging this into the inequality shows the probability of the sample mean straying far from the true mean goes to zero as sample size n grows, which is precisely the guarantee the law makes.

Concentration bounds in algorithm analysis and machine learning: whenever a rigorous, distribution-free guarantee is needed about how tightly a random quantity clusters around its expected value — without being able to assume normality — Chebyshev's inequality (or one of its refinements, like the Chernoff bound) is a standard starting tool.

Common Mistakes

Treating the bound as an estimate of the actual proportion, rather than a guaranteed minimum. The worked example's 75% is a floor, not a prediction — the real proportion within 2 standard deviations could be 80%, 95%, or even 100%, depending on the true (unknown) shape. Chebyshev's inequality never claims to say what the proportion actually is, only what it cannot fall below.

Using it when the distribution's shape is already known. If data is confirmed to be approximately normal, the 68-95-99.7 empirical rule gives a far more precise and still-accurate figure (95% vs. Chebyshev's 75% at k = 2) — Chebyshev's inequality is a fallback for when shape is genuinely unknown or confirmed non-normal, not a general-purpose replacement for it.

Applying it with k ≤ 1. The inequality only produces a meaningful (non-trivial) bound for k > 1 — at k = 1 it gives P(|X − μ| ≥ σ) ≤ 1, which is true of any probability but says nothing useful, since every probability is automatically ≤ 1 anyway.