Chebyshev's Inequality

The inequality

For any random variable X with finite mean μ = E[X] and finite variance σ² = Var(X), and any real number k > 0:

P(|X − μ| ≥ kσ) ≤ 1/k²

Equivalently, setting t = kσ:

P(|X − μ| ≥ t) ≤ σ²/t²

No assumptions about the distribution's shape — discrete, continuous, skewed, multimodal, fat-tailed — the bound holds. This universality is what makes Chebyshev's inequality the workhorse of probability theory.

Concrete numbers

k	Chebyshev bound 1/k²	Normal P(|Z| ≥ k)	Looseness factor
1	≤ 100% (vacuous)	31.7%	—
2	≤ 25%	4.6%	5.5×
3	≤ 11.1%	0.27%	41×
4	≤ 6.25%	0.0063%	990×
5	≤ 4%	0.000057%	70,000×
10	≤ 1%	10⁻²³	10²¹×

Chebyshev is loose for nice (light-tailed) distributions and tight for adversarial ones. The middle column is the "tight" bound for a Normal; Chebyshev pays for not knowing the shape.

Proof sketch

Apply Markov's inequality to the non-negative random variable Y = (X − μ)² with threshold a = (kσ)²:

P((X − μ)² ≥ (kσ)²) ≤ E[(X − μ)²] / (kσ)²
                    = σ² / (k²σ²)
                    = 1/k²

The event {(X − μ)² ≥ (kσ)²} is identical to {|X − μ| ≥ kσ}, giving Chebyshev. Two lines, one application of Markov. The same recipe — apply Markov to a transformed non-negative quantity — produces Chernoff bounds (apply to e^{tX}), Cantelli's one-sided bound (apply with a translation), and Bernstein's inequality.

Worked example: investment returns

Suppose a stock has monthly return X with μ = 1% and σ = 4%. What's the maximum probability of a monthly loss exceeding 7%?

A loss of 7% means X ≤ −7%.
|X − μ| = |X − 1%| ≥ 8% when X ≤ −7%.
8% / σ = 8% / 4% = 2 standard deviations away.

P(|X − μ| ≥ 2σ) ≤ 1/2² = 25%

So the probability of a ≥ 7% monthly loss is at most 25%, regardless of return distribution.

If returns are Normal the actual probability would be far lower (~2%). But for skewed, heavy-tailed financial data, Chebyshev's 25% is the only universally valid worst-case bound — and that conservatism is exactly why risk managers care.

Weak law of large numbers — in three lines

Let X₁, …, Xₙ be i.i.d. with mean μ and variance σ². Define the sample mean X̄ₙ = (1/n)·ΣXᵢ. Then:

E[X̄ₙ] = μ                 (linearity)
Var(X̄ₙ) = σ²/n             (i.i.d. variances add)

Chebyshev: P(|X̄ₙ − μ| ≥ ε) ≤ Var(X̄ₙ)/ε² = σ²/(nε²) → 0 as n → ∞.

So X̄ₙ → μ in probability — the weak law of large numbers. Bernoulli's 1713 proof took pages of combinatorics; Chebyshev's 1867 inequality collapsed it to a paragraph and revealed why the result is essentially about variance shrinking under averaging.

Markov vs Chebyshev vs Chernoff vs Hoeffding

Bound	Form	Requires	Tail decay	Use when
Markov	P(X ≥ a) ≤ E[X]/a	X ≥ 0, finite mean	1/a (linear)	only mean known, X non-negative
Chebyshev	P(|X−μ| ≥ kσ) ≤ 1/k²	finite variance	1/k² (polynomial)	variance known, distribution arbitrary
Cantelli (one-sided)	P(X−μ ≥ kσ) ≤ 1/(1+k²)	finite variance	1/k² but ~2× tighter	one-sided question
Chernoff	P(X ≥ a) ≤ e^{−tx}·M(t)	MGF exists in a neighborhood	e^{−Ω(a)} (exponential)	MGF available; sums of indep RVs
Hoeffding	P(X̄−μ ≥ ε) ≤ e^{−2nε²/(b−a)²}	bounded support [a, b]	e^{−Ω(nε²)} (exponential)	i.i.d. bounded; e.g. Bernoulli
Bernstein	P(X̄−μ ≥ ε) ≤ e^{−nε²/(2σ²+2bε/3)}	bounded + variance	between Chebyshev and Hoeffding	small variance, large support

The hierarchy: Markov uses one moment, Chebyshev uses two, Chernoff/Hoeffding use all of them via the MGF. More information → tighter bound. Chebyshev's polynomial decay is dramatically beaten by Chernoff's exponential decay when the MGF exists.

Multidimensional and operator versions

Chebyshev generalizes to vectors. For a random vector X ∈ ℝᵈ with mean μ and covariance Σ:

P(||X − μ||₂ ≥ k·tr(Σ)^{1/2}) ≤ 1/k²
P((X − μ)ᵀΣ⁻¹(X − μ) ≥ k²) ≤ d/k²    (Mahalanobis form)

And to matrices: matrix Chebyshev bounds the spectral norm of a sum of random matrices in terms of operator-variance, with applications to randomized numerical linear algebra (random projections, sketching, sparse recovery).

Where Chebyshev shows up

• Weak law of large numbers. The three-line proof above. Strong law needs sharper tools (Borel-Cantelli) but the weak version is pure Chebyshev.
• Concentration of measure. Variance bounds suffice when you only have second moments; tighter inequalities (Talagrand, log-Sobolev) refine in specific cases.
• Confidence intervals without normality. For unknown distributions, Chebyshev gives X̄ ± kσ/√n contains μ with probability ≥ 1 − 1/k². At k = 4.47, ≥ 95% — wider than Normal's 1.96, but distribution-free.
• Random matrix theory. Bound on the deviation of empirical eigenvalues from the limiting Wigner distribution.
• Algorithm analysis. Randomized algorithms with high probability of success: bound the variance of the estimator and Chebyshev bounds the failure probability.
• Monte Carlo. A Monte Carlo estimator with variance σ²/n has confidence radius σ/(√n ε)·k from Chebyshev — guides sample-size choice when the distribution is unknown.
• Polynomial Chebyshev (different person, same name). Note: Pafnuty Chebyshev also gave us Chebyshev polynomials; they're unrelated to the inequality but share an inventor.

The one-sided sharpening: Cantelli's inequality

Chebyshev bounds both tails together. For just one tail, Cantelli does better:

P(X − μ ≥ kσ) ≤ 1/(1 + k²)

At k = 2 this is 1/5 = 20% versus Chebyshev's 25%; at k = 3 it's 1/10 = 10% versus 11.1%. The improvement is modest because Chebyshev was already nearly tight for symmetric two-tailed events. Use Cantelli whenever the question is one-sided ("what's the chance returns drop below x?") — same effort, tighter bound.

Common pitfalls

• Treating Chebyshev as tight. For well-behaved distributions (Normal, exponential, bounded) Chebyshev is comically loose. Use it when you genuinely don't know the shape.
• Forgetting finite variance is required. Cauchy distributions have no variance — Chebyshev says nothing about them. Use median-based bounds or Lévy distance instead.
• Confusing with Chebyshev polynomials. Different topic, same Pafnuty Lvovich Chebyshev. The polynomials are a basis for [−1, 1] used in approximation theory.
• Applying when k < 1. The bound 1/k² > 1 is vacuous for k ≤ 1. Only useful for k strictly greater than 1.
• Forgetting variance is for the right random variable. If you're bounding X̄ₙ, use Var(X̄ₙ) = σ²/n, not σ². This is the most common mistake when applying Chebyshev to sample means.
• Plug-in estimator confusion. Replacing σ² with the sample variance s² turns Chebyshev's exact bound into an approximate one with extra sampling variability — confidence intervals from sample variance need t-distribution corrections.

Historical note

Pafnuty Lvovich Chebyshev (1821–1894) proved the inequality in 1867 in the context of his work on the weak law of large numbers. His student Andrey Markov (yes, that Markov) generalized Chebyshev's proof to extract the simpler inequality P(X ≥ a) ≤ E[X]/a in 1884 — what we now call Markov's inequality. So historically, Chebyshev came first and Markov is the simpler special case, even though pedagogically we now teach Markov first as the "building block."

Irénée-Jules Bienaymé proved a special case earlier in 1853; some sources call it the Bienaymé-Chebyshev inequality. The Cantelli one-sided refinement is due to Francesco Cantelli, 1928.