Visual Tools
Calculators
Tables
Mathematical Keyboard
Converters
Other Tools

How to Use Chebyshev's Inequality

Chebyshev's inequality bounds how much probability can sit far from the mean, using only the mean and the standard deviation. At least 1−1k21 - \frac{1}{k^2} of the probability lies within kk standard deviations of the mean, for any distribution.

Recognize the task

A question about staying near the mean or straying far from it, with only μ\mu and σ\sigma known: "at least what fraction lies between 60 and 80?", "bound P(∣X−μ∣≥c)P(|X - \mu| \geq c)", "how many samples make the average reliable?". Test: is the interval centered at the mean, or can you shrink it to one that is? Then count standard deviations.
The inequality is stated in the Chebyshev's inequality lesson.

Steps

StepOn scores with μ=70\mu = 70, σ=5\sigma = 5, P(60<X<80)P(60 < X < 80)
1Find the mean and the standard deviation. From the problem or from the standard deviation page.μ=70, σ=5\mu = 70,\ \sigma = 5
2Write the interval as mean plus or minus a distance. If the interval is not symmetric, use the largest symmetric interval inside it; the bound still holds for the bigger one.(60,80)=70±10(60, 80) = 70 \pm 10
3Express the distance in standard deviations. k=distanceσk = \frac{\text{distance}}{\sigma}. The bound needs k>1k > 1 to say anything.k=105=2k = \frac{10}{5} = 2
4Apply the bound or its complement. Chebyshev's inequality: P(∣X−μ∣≥kσ)≤1k2P(|X - \mu| \geq k\sigma) \leq \frac{1}{k^2}, so the probability of staying inside is at least 1−1k21 - \frac{1}{k^2}.P(60<X<80)≥1−14=34P(60 < X < 80) \geq 1 - \frac{1}{4} = \frac{3}{4}

Worked examples

Example 1 · a fair die, bound versus exact

variance form
μ=3.5\mu = 3.5, σ2=3512\sigma^2 = \frac{35}{12}, bound P(∣X−3.5∣≥2.5)P(|X - 3.5| \geq 2.5)
P≤σ2c2=35/126.25=715≈0.4667P \leq \frac{\sigma^2}{c^2} = \frac{35/12}{6.25} = \frac{7}{15} \approx 0.4667
distance c=2.5c = 2.5
exact: P(X∈{1,6})=13P(X \in \{1, 6\}) = \frac{1}{3}, below the bound
bound holds, not tight

Example 2 · an interval that is not centered

shrink first
μ=50, σ=4\mu = 50,\ \sigma = 4, bound P(42<X<62)P(42 < X < 62)
4242 is 8 below, 6262 is 12 above
(42,58)=50±8⊂(42,62)(42, 58) = 50 \pm 8 \subset (42, 62), k=2k = 2
largest centered interval inside
P(42<X<62)≥1−14=34P(42 < X < 62) \geq 1 - \frac{1}{4} = \frac{3}{4}
bigger interval, at least as likely

Example 3 · how reliable an average is

sample mean
Xˉ\bar{X} = share of heads in 100 flips, μ=0.5\mu = 0.5
Var⁡(Xˉ)=0.25100=0.0025\operatorname{Var}(\bar{X}) = \frac{0.25}{100} = 0.0025
variance of one flip over nn
P(∣Xˉ−0.5∣≥0.1)≤0.00250.01=0.25P(|\bar{X} - 0.5| \geq 0.1) \leq \frac{0.0025}{0.01} = 0.25
variance form, c=0.1c = 0.1
Compare the bound with real distributions in the Chebyshev inequality visualizer.

See it happen

Apply Markov's reasoning to the squared distance (X−μ)2(X - \mu)^2, which is never negative and has mean σ2\sigma^2. Far values have huge squared distances, so only a little probability can sit there before the average squared distance would exceed σ2\sigma^2. That leaves at most 1k2\frac{1}{k^2} of the probability beyond kk standard deviations: at most 25%25\% beyond 2σ2\sigma, at most about 11%11\% beyond 3σ3\sigma.

Common mistakes

Using k of 1 or less

✗ k=1k = 1: at least 1−1=01 - 1 = 0 within one standard deviation, so the method failed
✓ the bound is true but empty; Chebyshev helps only for k>1k > 1

Dividing by the standard deviation squared the wrong way

✗ P(∣X−μ∣≥c)≤σc2P(|X - \mu| \geq c) \leq \frac{\sigma}{c^2}
✓ P(∣X−μ∣≥c)≤σ2c2P(|X - \mu| \geq c) \leq \frac{\sigma^2}{c^2}, or 1k2\frac{1}{k^2} with k=cσk = \frac{c}{\sigma}

Using an uncentered interval as it is

✗ (42,62)(42, 62) with μ=50\mu = 50: k=202⋅4=2.5k = \frac{20}{2 \cdot 4} = 2.5
✓ shrink to the centered (42,58)(42, 58): k=2k = 2

Confusing the bound with the normal rule

✗ within 2σ2\sigma: exactly 95%95\% for any distribution
✓ Chebyshev guarantees only 75%75\%; about 95%95\% holds for the normal