Markov and Chebyshev Inequalities

How likely is a nonnegative random variable to be far above its own average? Suppose a factory's daily output has mean 100 units. Without knowing anything else about the distribution, can we say anything about the chance of producing more than 500 units on a given day? Remarkably, yes. The bound requires no knowledge of the shape of the distribution at all, only its mean. This is the power of tail inequalities: they trade precision for universality. They give guarantees that hold for every distribution satisfying a mild moment condition. This article develops the two most fundamental tail bounds in probability. Markov's inequality bounds the tail of a nonnegative random variable using only its mean. Chebyshev's inequality sharpens this by using the variance to control deviations from the mean in both directions. These inequalities are typically weak for any single well-understood distribution. Their strength lies elsewhere: they apply universally, need only first or second moments. Crucially, they become the engine behind the proof of the Weak Law of Large Numbers, one of the cornerstones of probability theory. The big picture is this: precise distributional knowledge (exact densities, moment generating functions, and so on) is often unavailable or unnecessary. Markov and Chebyshev let us convert coarse information — an expectation, a variance — into quantitative control over tail probabilities. This is the first step in a hierarchy of increasingly refined tail bounds. These include Chernoff bounds, sub-Gaussian concentration, and martingale inequalities. Such tools pervade modern probability, statistics, and theoretical computer science. This article states and proves both inequalities with full rigor. It builds geometric intuition with figures and works through two asymmetric numerical examples. It shows via explicit counterexamples why each hypothesis (nonnegativity, finite variance) cannot be dropped. It closes by connecting these humble inequalities to their historical origins. It also traces their vast modern footprint, from randomized algorithms to statistical estimation.

Key takeaways

Markov's inequality bounds $P(X \ge a)$ for nonnegative $X$ using only $E[X]$ , for any $a>0$ .
Chebyshev's inequality bounds $P(|X - \mu| \ge k)$ using the variance $\sigma^2 = \operatorname{Var}(X)$ .
Chebyshev is derived from Markov by applying it to the nonnegative variable $(X-\mu)^2$ .
Both bounds are distribution-free: they use only moments, not the full law of $X$ .
The bounds can be very loose for specific distributions but are tight in the worst case over all distributions with the given moments.
Key idea: turning a mean or variance into a guaranteed cap on tail probability, with no further assumptions on the distribution.

Where this lives today

Markov's and Chebyshev's inequalities are foundational throughout probability, statistics, and computer science precisely because they require so little. In statistics, Chebyshev's inequality underlies elementary proofs of the consistency of the sample mean as an estimator, connecting directly to the Weak Law of Large Numbers: as sample size $n$ grows, the variance of the sample mean shrinks like $\sigma^2/n$ , and Chebyshev shows the sample mean converges in probability to the true mean. In randomized algorithms and theoretical computer science, Markov and Chebyshev bounds (and their refinements, the Chernoff–Hoeffding bounds) are used to certify that randomized procedures succeed with high probability, for instance in analyzing hashing schemes, streaming algorithms, and load-balancing. In statistical mechanics and information theory, moment-based tail bounds justify concentration-of-measure phenomena central to the behavior of large systems of particles or symbols. In quality control and risk management, Chebyshev-type bounds provide distribution-free guarantees on the probability of extreme deviations when only a mean and variance can be reliably estimated from data. More broadly, these two inequalities are the entry point into the modern theory of Concentration Inequalities, where higher moments or moment generating functions yield exponentially sharper tail control at the cost of stronger assumptions.

:::quiz[If $X \ge 0$ has $E[X] = 20$ , what is the tightest bound Markov's inequality gives for $P(X \ge 50)$ ?]

0.5
0.4 ✓
2.5
0.04

:::faq[Why does Markov's inequality require only $E[X]$ but Chebyshev's requires both mean and variance?] Chebyshev's inequality is not a new, independent result — it is Markov's inequality applied to the transformed variable $(X-\mu)^2$ . Since that transformed variable's expectation is by definition the variance of $X$ , the variance enters only because it is literally $E[(X-\mu)^2]$ , the mean of the squared-deviation variable. In other words, Chebyshev needs "the mean of a certain nonnegative variable," and that mean happens to be named the variance when the nonnegative variable is a squared deviation. Any inequality of Markov type on a transformed nonnegative variable will require the mean of that transform to exist and be finite.

Ask the tutor: How does Chebyshev's inequality combine with independence to prove the Weak Law of Large Numbers step by step?

(\Omega, \mathcal{F}, P)

Intuition

Markov's inequality formalizes a simple accounting fact: if a nonnegative quantity has small average, it cannot too often be very large, because large values would inflate the average beyond what is observed. Concretely, if $X\ge 0$ takes the value $a$ or more on a set of probability $p$ , then even ignoring all other values,

E[X] \ge a \cdot P(X \ge a),

since the contribution from $\{X \ge a\}$ alone is at least $a\cdot P(X\ge a)$ , and the rest of $X$ contributes a nonnegative amount. Rearranging gives Markov's inequality directly. This is the entire idea; the proof below simply makes it precise using indicator functions. Chebyshev's inequality is Markov's inequality one level removed: instead of bounding how often $X$ itself is large, we bound how often the squared deviation $(X-\mu)^2$ is large, which controls how often $X$ strays from $\mu$ in either direction. Since $(X-\mu)^2$ is nonnegative, Markov applies immediately once we recognize that

\{|X-\mu| \ge k\} = \{(X-\mu)^2 \ge k^2\}.

Figure

Markov's inequality: mass beyond

a

is capped by area under the density

Intuition (continued): the geometric meaning of Chebyshev

Chebyshev's inequality says the probability mass sitting more than $k$ away from the mean is squeezed by $\sigma^2/k^2$ : as $k$ grows relative to $\sigma$ , the allowed tail mass shrinks quadratically. Setting $k = c\sigma$ shows the bound depends only on the number of standard deviations $c$ , matching the informal "how many sigmas away" language used throughout statistics.

Figure

Chebyshev's inequality: two-sided tail mass beyond

k

from

\mu

X

a > 0

X

Y = (X - \mu)^2

Worked examples

Example 1 (Markov, asymmetric numbers — factory output). A factory's daily output $X$ (in units) is a nonnegative random variable with $E[X] = 137$ . Management wants a guaranteed upper bound on the probability that output exceeds $823$ units on a given day, without assuming any particular distribution for $X$ (perhaps the historical data is too sparse to fit a model reliably). By Markov's inequality with $a = 823$ :

P(X \ge 823) \le \frac{137}{823} \approx 0.16647\ldots

So management can be certain — regardless of the true distribution of $X$ , as long as $X \ge 0$ and $E[X]=137$ — that on at most about $16.65\%$ of days will output reach or exceed $823$ units. Note this required no variance, no shape assumption, and no independence assumption; it used only the mean and nonnegativity. This is both the strength (universality) and the weakness (looseness) of the bound: if $X$ is known to be, say, exactly Normal with mean $137$ and small standard deviation, the true probability could be astronomically smaller than $0.1665$ .

Example 2 (Chebyshev, asymmetric numbers — exam scores). Suppose exam scores $X$ across a large population have mean $\mu = 71.4$ and variance $\sigma^2 = 58.3$ (so $\sigma \approx 7.635$ ). We want a guaranteed bound on the probability that a randomly selected student's score deviates from the mean by $18$ points or more in either direction, i.e., $P(|X - 71.4| \ge 18)$ . Apply Chebyshev's inequality with $k = 18$ :

P(|X - 71.4| \ge 18) \le \frac{58.3}{18^2} = \frac{58.3}{324} \approx 0.17994\ldots

So at most about $18.0\%$ of students score below $53.4$ or above $89.4$ , regardless of the shape of the score distribution (it need not be Normal, need not be symmetric — only finite mean and variance are used). As a cross-check using the "number of standard deviations" form, $k = 18 = c\sigma$ gives $c = 18/7.635 \approx 2.358$ , so

P(|X-\mu| \ge 2.358\,\sigma) \le \frac{1}{(2.358)^2} \approx 0.17994,

matching the direct computation (as it must, since both are the same inequality rewritten). This illustrates that Chebyshev's guarantee at "just over $2.36$ standard deviations" is far weaker than the familiar Normal-distribution heuristic (where such a deviation would have probability under $2\%$ ) — a reminder that Chebyshev is a worst-case, distribution-free bound, not a distributional approximation.

X \ge 0

Scope, edge cases & common traps

Situation	What happens
$X \ge 0$ but $E[X] = \infty$	Markov's bound $E[X]/a$ is trivially $\infty \ge$ anything; still true but vacuous.
$a \le 0$ in Markov	Statement is not used this way; $P(X\ge a)$ can be $1$ while bound is meaningless or undefined since inequality was derived for $a>0$ .
$\sigma^2 = 0$	$X = \mu$ almost surely; Chebyshev gives $P(\|X-\mu\|\ge k) \le 0$ for every $k>0$ , forcing equality with the true value $0$ — the bound is tight.
$k \le \sigma$ (few "sigmas")	Bound $\sigma^2/k^2 \ge 1$ , so the inequality is true but vacuous (probabilities are always $\le 1$ ).
Distribution symmetric & bounded, e.g. $X = \mu \pm k$ each w.p. $1/2$	Chebyshev is tight: $P(\|X-\mu\|\ge k) = 1 = \sigma^2/k^2$ exactly, showing the constant cannot be improved in general.
Applying Markov to a signed variable directly	Invalid — see counterexample above; must apply to a nonnegative transform such as $\|X\|$ or $(X-\mu)^2$ .

Common trapThe classic error

A very common mistake is to apply Markov's inequality directly to a random variable that is not known to be nonnegative — for instance, applying it to $X - \mu$ itself instead of to $(X-\mu)^2$ or $|X-\mu|$ . The inequality $P(X - \mu \ge k) \le E[X-\mu]/k$ is meaningless because $E[X-\mu] = 0$ , and even if it were applied validly, negativity of $X-\mu$ on part of the sample space destroys the proof's pointwise step. The correct route to a two-sided deviation bound is always through a genuinely nonnegative quantity — squared deviation for Chebyshev, or $|X|$ more generally — never through the signed variable directly.

The mathematician behind it

Origins · 1 minAndrey Andreyevich Markov and how this was discovered

Andrey Andreyevich Markov

1856–1922 · Imperial Academy of Sciences, St. Petersburg

Markov was a Russian mathematician trained under Pafnuty Chebyshev at St. Petersburg University, and he became one of the leading figures of the St. Petersburg school of probability. He is best remembered today for Markov chains, sequences of dependent random variables satisfying the memoryless property, which he introduced partly to demonstrate that the law of large numbers could hold even without independence. The inequality bearing his name appears as an elementary tool within his broader development of moment-based probabilistic reasoning, building on and generalizing bounds his teacher Chebyshev had used.

History

- 1821 — earlier moment-based tail arguments anticipated in the work of Irénée-Jules Bienaymé on error propagation.

- 1867 — Pafnuty Chebyshev publishes the inequality now bearing his name, using it to give a fully rigorous proof of the weak law of large numbers.

- 1884 — Andrey Markov, Chebyshev's student, generalizes and systematizes the underlying moment-inequality technique in his probabilistic writings.

- 1906 — Markov introduces Markov chains, extending probabilistic limit theorems to dependent sequences.

- 20th century — the Markov–Chebyshev inequalities become standard textbook tools and the starting point for sharper concentration inequalities (Chernoff, Hoeffding, Bernstein).

Why is Markov's inequality often called "weak" or "loose"?

Markov's inequality uses only the mean of a nonnegative random variable, discarding all other distributional information (variance, shape, tail decay). For distributions with light tails — such as the Normal or exponential with small parameters — the true tail probability can be many orders of magnitude smaller than Markov's bound. The inequality is tight only for a very particular two-point distribution (mass at $0$ and at $a$ ), which is rarely the distribution encountered in practice. Its value lies not in numerical sharpness for known distributions, but in its complete universality: it holds for every nonnegative random variable with a finite mean, no matter how little else is known about it, which makes it indispensable as a building block for further results.

Does Chebyshev's inequality assume the distribution is symmetric or bell-shaped?

No. Chebyshev's inequality makes no assumption about symmetry, unimodality, or any particular shape; it only requires a finite mean and finite variance. This is precisely its power and its cost: it applies to any distribution with two finite moments, including highly skewed or multimodal ones, but as a result cannot exploit shape information the way Normal-based rules (like the empirical 68–95–99.7 rule) do. For a Normal distribution specifically, the true tail probability at two standard deviations is about $4.6\%$ , far below Chebyshev's guaranteed bound of $25\%$ — the gap is the price paid for distribution-free validity.

Can Chebyshev's inequality ever be an equality, or is it always strict?

It can be an exact equality for specific distributions. If $X$ takes the values $\mu - k$ and $\mu + k$ each with probability $1/2$ , then $\operatorname{Var}(X) = k^2$ , and $P(|X-\mu|\ge k) = 1$ , which equals $\sigma^2/k^2 = k^2/k^2 = 1$ exactly. This shows the constant in Chebyshev's inequality cannot be improved without further assumptions: there exist distributions making the bound tight, so no universal distribution-free bound with a smaller constant could be correct.

How do Markov and Chebyshev relate to the Weak Law of Large Numbers?

The Weak Law of Large Numbers states that the sample mean of i.i.d. random variables converges in probability to the true mean as sample size grows. The standard elementary proof (assuming finite variance) applies Chebyshev's inequality to the sample mean $\bar X_n$ , whose variance is $\sigma^2/n$ by properties of Variance of Sums of Random Variables under independence. This gives $P(|\bar X_n - \mu| \ge \varepsilon) \le \sigma^2/(n\varepsilon^2) \to 0$ as $n \to \infty$ , which is exactly the statement of convergence in probability. Chebyshev's inequality is thus the direct mechanism, not just a conceptual analogy, behind this foundational limit theorem.

Is there a one-sided version of Chebyshev's inequality that gives a tighter bound in a single direction?

Yes — this is known as the Cantelli inequality (one-sided Chebyshev), which states that for a random variable with mean $\mu$ and variance $\sigma^2$ , and any $k>0$ ,

P(X - \mu \ge k) \le \frac{\sigma^2}{\sigma^2+k^2}.

This bound is always at least as sharp as applying two-sided Chebyshev and halving, and is proved by a similar Markov-style argument applied to $(X-\mu+t)^2$ for an optimally chosen shift $t$ . It illustrates that the basic Markov–Chebyshev technique is flexible and admits refinements beyond the article's core two results.

Do these inequalities require independence of any random variables?

No. Both Markov's and Chebyshev's inequalities, as stated here, are single-variable statements requiring no independence assumption whatsoever; they hold for any single random variable satisfying the stated moment conditions. Independence becomes relevant only when these inequalities are applied to sums or averages of several random variables (as in the Weak Law of Large Numbers), where independence is used separately, via Variance of Sums of Random Variables, to compute the variance of the sum before Chebyshev is invoked.