Visual Tools
Calculators
Tables
Mathematical Keyboard
Converters
Other Tools

How to Use Markov's Inequality

When all you know about a nonnegative random variable is its mean, Markov's inequality still caps the chance of a large value: the probability of reaching aa is at most the mean divided by aa.

Recognize the task

A nonnegative quantity (waiting time, income, count, file size) with a known mean and no distribution, and a question like "at most how likely is it to exceed aa?" or "show that P(X≥a)P(X \geq a) is small". Test: is X≥0X \geq 0 always, and is aa larger than the mean? Then Markov gives a useful bound.
The inequality is stated in the Markov's inequality lesson.

Steps

StepOn a waiting time with mean 10 minutes, P(X≥30)P(X \geq 30)
1Check that the variable is nonnegative. Markov fails for variables that can be negative. Shift or use another bound in that case.a waiting time is ≥0\geq 0 ✓
2Find the mean. From the problem, or as on the expected value page.E[X]=10E[X] = 10
3Divide the mean by the threshold. Markov's inequality: P(X≥a)≤E[X]aP(X \geq a) \leq \frac{E[X]}{a} for every a>0a > 0.P(X≥30)≤1030=13P(X \geq 30) \leq \frac{10}{30} = \frac{1}{3}
4Judge the bound. It is an upper limit, not the probability. If a≤E[X]a \leq E[X] the bound is 11 or more and says nothing.at most a one-in-three chance

Worked examples

Example 1 · a direct bound

mean only
X≥0X \geq 0, E[X]=50E[X] = 50
P(X≥200)≤50200P(X \geq 200) \leq \frac{50}{200}
mean over threshold
P(X≥200)≤14P(X \geq 200) \leq \frac{1}{4}
at most 25%

Example 2 · a lower bound for the complement

flip it
X≥0X \geq 0, E[X]=20E[X] = 20, find a bound on P(X<50)P(X < 50)
P(X≥50)≤2050=25P(X \geq 50) \leq \frac{20}{50} = \frac{2}{5}
Markov on the upper tail
P(X<50)≥1−25=35P(X < 50) \geq 1 - \frac{2}{5} = \frac{3}{5}
complement rule

Example 3 · how loose the bound can be

compare
exponential wait, mean 10: P(X≥30)≤13P(X \geq 30) \leq \frac{1}{3}
Markov
exact: P(X≥30)=e−30/10=e−3P(X \geq 30) = e^{-30/10} = e^{-3}
exponential tail
e−3≈0.0498e^{-3} \approx 0.0498, far below 0.33330.3333
true, but crude
See the bound against real distributions in the Markov inequality visualizer.

See it happen

If more than a fraction E[X]a\frac{E[X]}{a} of the probability sat at aa or above, those values alone would push the mean above E[X]E[X], because XX can never be negative to pull it back down. So the tail is capped by the mean. The bound is exact only for a variable that sits at 00 and aa and nowhere else, which is why it is usually loose.

Common mistakes

Applying it to a variable that can be negative

✗ temperature change with E[X]=1E[X] = 1: P(X≥5)≤15P(X \geq 5) \leq \frac{1}{5}
✓ Markov needs X≥0X \geq 0; with negative values the bound can fail

Reading the bound as the probability

✗ P(X≥30)=13P(X \geq 30) = \frac{1}{3}
✓ P(X≥30)≤13P(X \geq 30) \leq \frac{1}{3}; the true value may be much smaller

A threshold at or below the mean

✗ E[X]=10E[X] = 10: P(X≥5)≤2P(X \geq 5) \leq 2, so the method failed
✓ the bound is true but useless; Markov helps only for a>E[X]a > E[X]