Visual Tools
Calculators
Tables
Mathematical Keyboard
Converters
Other Tools


Probability Inequalities






Probability Inequalities


In many probability problems, the exact distribution of a random variable is unknown or too complicated to work with directly.

Probability inequalities address this situation by providing guaranteed bounds on probabilities using limited information, such as an expected value or a variance. Instead of describing outcomes precisely, they restrict how extreme those outcomes can be.

These results are deliberately general. They apply across wide classes of random variables and make minimal assumptions, trading sharpness for reliability. For this reason, probability inequalities play a central role in theoretical analysis, estimation, and convergence arguments.

Key Terms

Expected Value— E[X]E[X], the quantity used in Markov's inequality
Variance— Var⁡(X)\operatorname{Var}(X), the quantity used in Chebyshev's inequality
Standard Deviation— σ\sigma, the spread measure in Chebyshev's bound
Random Variable— the object both inequalities constrain

See All Probability Definitions →


How This Section Is Organized


This page provides an overview of probability inequalities and how they fit into probability theory.

Each inequality is treated on its own page, where its assumptions, statement, and implications are presented in detail. The parent page focuses on connections, scope, and conceptual differences rather than formulas or proofs.

As new inequalities are added, they are integrated into the same structure, allowing this section to expand without changing its overall organization.

Inequality Pages:

• Markov's Inequality
• Chebyshev's Inequality
• (Future additions will appear here)


What Probability Inequalities Do

Waits at a counter average 10 minutes, and nothing else is known about them. How likely is a wait of 50 minutes or more?
Without the distribution there is no exact answer. Something can still be said.
Whatever the distribution, the chance cannot exceed 1/5.
waiting time with mean 10exact P(X ≥ 50)bound 10/50
exponentialabout 0.00670.2
uniform on 0 to 2000.2
0 with probability 0.8, 50 with probability 0.20.20.2
If waits of 5050 minutes or more had probability above 15\frac{1}{5}, they alone would push the average above 1010. So the bound holds for every non-negative waiting time with that average.
Partial information gives a guarantee: a bound, not an exact value.
They do not attempt to compute exact probabilities. Instead, they provide bounds that are guaranteed to hold whenever the stated assumptions are satisfied.
The definition
A probability inequality bounds the probability of an event using only partial information, such as a mean or a variance, and holds for every distribution that meets its assumptions.
Probability inequalities place limits on how likely certain events can be, without requiring full knowledge of the underlying distribution.
So P(X≥50)≤1050=15P(X \ge 50) \le \frac{10}{50} = \frac{1}{5} for any non-negative XX with mean 1010; this is Markov's inequality.
Inequalities are tools for reasoning when an exact calculation is impossible or unnecessary.
In practice, this means they can:
• bound tail probabilities
• control the chance of large deviations
• give worst-case guarantees
• remain valid even when the distribution is unknown
Probability inequalities are therefore tools for reasoning under uncertainty when precise calculation is not possible or not necessary.
A bound can be far from the truth.
For exponential waits the true chance is about 0.00670.0067, thirty times smaller than 15\frac{1}{5}. The bound is the worst case, reached only by the two-point distribution in the last row. Sharper bounds need more information, as set out in types of probability inequalities.
The rest of the page covers why inequalities matter, what they depend on, their types, the featured inequalities and how they are related, how they are used, and how they compare with exact distributions.

Why Inequalities Matter in Probability


Exact probability calculations are often unavailable or impractical.

In many situations, the full distribution of a random variable is unknown, difficult to compute, or unnecessary for the question being asked. Probability inequalities make it possible to reason in such cases by providing bounds that hold under broad conditions.

Because they rely on minimal assumptions, inequalities are used:
• to justify convergence results
• to control error and variability
• to obtain guarantees that remain valid across many models

For this reason, probability inequalities are foundational tools in both theoretical arguments and applied probability.

What Inequalities Depend On


Probability inequalities are built on a small set of basic quantities associated with random variables.

Most inequalities rely on:
• the expectation of a random variable
• measures of spread such as variance
• structural assumptions like non-negativity or boundedness

Different inequalities require different levels of information.
The fewer assumptions used, the more general the bound tends to be, but the less tight it becomes.

This shared dependence on simple characteristics explains why very different random variables can be constrained by the same inequality.
33.0026.5020.0013.507.00Mean = 20.0012.0015.0018.0020.0022.0025.0028.001234567
A dataset with its mean and squared deviations

The inputs to the inequalities are exactly what this picture shows: a mean and the spread of the values around it, measured by variance. A random variable with small spread cannot often land far from its mean, and that is the fact the bounds make precise. Move the points and watch the variance respond on the variance visualizer.

The inequalities are grouped by which of these inputs they use.

Types of Probability Inequalities


Probability inequalities can be grouped according to the kind of information they use and the type of bounds they provide.

Common high-level categories include:
• inequalities based only on expectation
• inequalities that incorporate variance or higher moments
• tail bounds that control extreme deviations
• concentration-type inequalities that sharpen bounds under stronger assumptions

These categories are not rigid.
They reflect increasing levels of available information, with stronger assumptions generally leading to tighter bounds.

Fact: From 1 in 40 to 2 in a billion
For at least 600600 heads in 10001000 fair tosses, Chebyshev's inequality allows up to 2501002=0.025\frac{250}{100^2} = 0.025. Hoeffding's inequality, which also uses that the tosses are independent and each is 00 or 11, gives e−20≈2×10−9e^{-20} \approx 2 \times 10^{-9}; the exact chance is about 1.4×10−101.4 \times 10^{-10}.

The distance between the first categories and the last one can be very large. Take 10001000 tosses of a fair coin and the event of at least 600600 heads, 100100 above the expected 500500. A variance-based bound sees only the spread of the total, a variance of 1000⋅14=2501000 \cdot \frac{1}{4} = 250. A concentration inequality also uses how the total is built: as a sum of many independent tosses, each adding either 00 or 11.

The categories above can be lined up with the kind of information each one consumes and where each is typically most useful.
DownloadPNG2×PDF
Layout
CategoryInformation usedTypical use
Expectation-basedE[X] only (sometimes plus non-negativity)broadest applicability; coarsest bounds (e.g., Markov)
Variance / higher momentsEX] together with [Var(X) or higher momentssharper deviation bounds around the mean (e.g., Chebyshev)
Tail boundsmoment information or partial distribution shapecontrolling probability of extreme deviations
Concentration-typeindependence plus boundedness or sub-Gaussianityvery tight bounds for sums and averages of many variables
Types of Probability InequalitiesLearn Math Class

Featured Inequalities


This section highlights core probability inequalities that are used throughout probability theory.
Each inequality has its own page with assumptions, statements, and typical uses.

Markov's Inequality
• Applies to non-negative random variables
• Uses expectation only
• Provides very general, often loose bounds

Pitfall: Using Markov on a variable that can be negative
Applying Markov to X=±10X = \pm 10 with equal chances, where E[X]=0E[X] = 0, gives P(X≥5)≤05=0P(X \geq 5) \leq \frac{0}{5} = 0. In fact P(X≥5)=12P(X \geq 5) = \frac{1}{2}; the bound holds only for X≥0X \geq 0.

In symbols, Markov's inequality states that P(X≥a)≤E[X]aP(X \geq a) \leq \frac{E[X]}{a} for every threshold a>0a > 0. The only inputs are the mean and the threshold, so the bound can be written down as soon as E[X]E[X] is known. A mean of 22, for example, gives P(X≥10)≤210=0.2P(X \geq 10) \leq \frac{2}{10} = 0.2 for any variable that meets the condition in the first bullet above.



Chebyshev's Inequality
• Uses both expectation and variance
• Gives tighter bounds than Markov under stronger assumptions
• Central for reasoning about deviations from the mean


xPDFμ=10μ-aμ+a24681012141618
Chebyshev bound for a normal variable, mean 10, variance 4, deviation 3

The shaded tails are the actual probability of straying at least 3 from the mean and the bound is variance over 9, which sits well above them: the inequality is valid but loose because it uses only the variance. The Markov visualizer shows the even looser bound that uses only the mean. Change the deviation and compare bound and truth on the Chebyshev inequality visualizer.

How the featured inequalities relate is set out next.

Relationship Between Common Inequalities


Probability inequalities form a hierarchy based on the assumptions they require.

• Markov's inequality uses only non-negativity and expectation, making it broadly applicable but often loose.
• Chebyshev's inequality adds information about variance, which tightens the bound while reducing generality.

This illustrates a general trade-off:
• fewer assumptions → wider applicability
• more assumptions → sharper bounds

Many other inequalities fit into this pattern, refining earlier ones by incorporating additional structure or moments.

Fact: The tenth power beats the variance
For XX exponential with mean 11 and a=10a = 10, Markov gives 0.10.1 and Chebyshev 181≈0.012\frac{1}{81} \approx 0.012. Markov applied to X10X^{10}, with E[X10]=10!E[X^{10}] = 10!, gives 10!1010≈0.00036\frac{10!}{10^{10}} \approx 0.00036, about 3434 times smaller; the true tail is e−10≈0.000045e^{-10} \approx 0.000045.

The same step works with any power. Because XkX^k is never negative when X≥0X \geq 0, Markov's inequality applied to XkX^k gives P(X≥a)≤E[Xk]akP(X \geq a) \leq \frac{E[X^k]}{a^k}, a bound that falls like 1ak\frac{1}{a^k} instead of 1a\frac{1}{a}. Chebyshev's inequality is the case k=2k = 2, applied to the deviation X−μX - \mu, and each further moment that is known adds one more bound to choose from.
For X exponential with mean 1 and variance 1: true tail e to the minus a, Markov bound 1/a, Chebyshev bound 1/(a-1) squaredaP(X ≥ a)1234560.250.50.751Markov 1/aChebyshevtrue e^(−a)X exponential: mean 1, variance 1Chebyshev wins past a ≈ 2.6both stay above the true tailMore information, a tighter bound, still not exact
Markov and Chebyshev against the true tail

Take X exponential with mean 1 and variance 1, so the true tail is P(X ≥ a) = e⁻ᵃ. Markov, using only the mean, gives 1/a. Chebyshev, using the variance too, gives 1/(a − 1)² for a > 1. Past a ≈ 2.6 Chebyshev is the smaller bound, but both stay well above the true curve: each is a guarantee, not an estimate.

Neither bound uses the exponential shape; only the true curve does.
Each added assumption rules out distributions, so the worst case a bound must cover can only get smaller.

How Inequalities Are Used


Probability inequalities are used to control uncertainty when exact calculations are unavailable or unnecessary.

Typical uses include:
• bounding tail probabilities
• estimating how far a random variable can deviate from a reference value
• proving convergence results without specifying distributions
• comparing variability across different models

Because they provide guarantees that hold under broad conditions, inequalities are often used as intermediate tools. They establish limits first, and more precise results are built on top of them when additional information becomes available.

Inequalities vs Exact Distributions


Probability inequalities do not describe the full behavior of a random variable.

They provide bounds that must hold, but they do not capture how probability is distributed within those bounds. As a result, the estimates they give may be conservative or loose.

Pitfall: Reading the bound as the probability
Taking Markov's P(X≥4)≤14P(X \geq 4) \leq \frac{1}{4}, for an exponential XX with mean 11, as P(X≥4)≈0.25P(X \geq 4) \approx 0.25 overstates the chance about 14 times. The exact value is e−4≈0.018e^{-4} \approx 0.018.

When the exact distribution of a random variable is known and manageable, working directly with that distribution is usually preferable. Inequalities become most valuable when such information is missing, incomplete, or too costly to compute.

In this sense, inequalities complement exact methods rather than replace them.

Featured Inequalities at a Glance

The table below collects the inequalities introduced above into a single reference, lining each one up with its structural requirement, the information it consumes, the formal statement, what it actually bounds, and how tight it tends to be. As new inequalities are added to this section, they will slot into the same comparison.

Common Pitfalls

Each of the two mistakes below is marked where it happens in the sections above; here each is worked through in full, with the wrong and the right working side by side.
DownloadPNG2×PDF
1Using Markov on a variable that can be negative
Markov's inequality is applied to a variable that takes negative values as well as positive ones. The mean is computed correctly and put into E[X]a\frac{E[X]}{a}, but the condition X≥0X \geq 0 is never checked.
Wrong turn
Error here
Let XX be −10-10 or 1010, each with probability 12\frac{1}{2}.
E[X]=(−10)⋅12+10⋅12=0E[X] = (-10) \cdot \frac{1}{2} + 10 \cdot \frac{1}{2} = 0
P(X≥5)≤E[X]5=05=0P(X \geq 5) \leq \frac{E[X]}{5} = \frac{0}{5} = 0
Right path
XX can be negative, so Markov's inequality does not apply to it. Directly, P(X≥5)=P(X=10)=12P(X \geq 5) = P(X = 10) = \frac{1}{2}. A valid Markov bound comes from the shifted variable X+10X + 10, which is never negative and has mean 1010:
P(X≥5)=P(X+10≥15)P(X \geq 5) = P(X + 10 \geq 15)

≤1015=23\leq \frac{10}{15} = \frac{2}{3}
Why it happensThe formula uses only E[X]E[X] and aa, so nothing in it shows the condition. Negative values cancel positive ones in the mean, and a small mean then gives a bound that is too small.
How to checkLook at the smallest value XX can take before using the formula. A bound of 00 for an event that clearly can happen, such as X=10X = 10 here, is a sign that the condition was broken.
2Reading the bound as the probability
The inequality gives an upper limit on a probability, and that limit is reported as the probability itself.
Start
XX is exponential with mean 11.
Wrong turn
Error here
Markov's inequality gives
P(X≥4)≤E[X]4=14P(X \geq 4) \leq \frac{E[X]}{4} = \frac{1}{4}
And the answer is given as P(X≥4)≈0.25P(X \geq 4) \approx 0.25.
Right path
The bound only says that the probability is at most 0.250.25. The exponential distribution gives the exact value
P(X≥4)=e−4≈0.018P(X \geq 4) = e^{-4} \approx 0.018
Which is about one time in 5555. Chebyshev's inequality, which also uses the variance 11, gives 1(4−1)2=19≈0.11\frac{1}{(4 - 1)^2} = \frac{1}{9} \approx 0.11: closer, but still about six times too large.
Why it happensThe bound is the only number the inequality produces, and it comes out of a formula just as an exact probability would. The sign ≤\leq in front of it is easy to read past.
How to checkThe same bound holds for every non-negative variable with mean 11. One of them is X=0X = 0 with probability 34\frac{3}{4} and X=4X = 4 with probability 14\frac{1}{4}, where P(X≥4)=14P(X \geq 4) = \frac{1}{4} exactly; the exponential, with the same mean, has 0.0180.018. A number shared by such different variables cannot be the probability for each of them.
Common pitfalls with probability inequalities·/probability/inequalitiesLearn Math Class
Both slips misuse what Markov's inequality gives. It holds only for a variable that is never negative, and it gives an upper bound, not the probability itself. Looking at the smallest value XX can take, and writing ≤\le in the answer, catches both.

Skills You Should Learn

Each skill opens its step-by-step method page.
Random variable methods

Probability Inequalities FAQ

What's the difference between Markov's and Chebyshev's inequality?

+
Markov uses only non-negativity and the expectation, so it applies almost anywhere but gives loose bounds. Chebyshev adds the variance, which tightens the bound at the cost of needing more information. That is the general pattern: each additional assumption buys a sharper bound and narrows the range of variables it covers.Read more →

Do you need to know the distribution to use a probability inequality?

+
No — that is precisely why they exist. An inequality uses a few summary facts, typically the expectation and sometimes the variance, plus structural conditions like non-negativity. From those alone it guarantees a bound holding for every distribution that meets the conditions. You trade precision for the freedom of not needing the full distribution.Read more →

Why do probability inequalities often give such loose bounds?

+
Because they must hold for every distribution satisfying the assumptions, including the worst one. The bound is set by that worst case, so for any particular well-behaved distribution it will be conservative. A bound can even come out above 1 and say nothing at all — technically true, practically vacuous. That is the price of not needing the distribution.Read more →

Does a tighter bound always mean a better inequality?

+
No — tightness and applicability pull against each other. A sharper bound is bought with stronger assumptions, and if those assumptions fail the bound is not merely loose but invalid. The right choice depends on what you actually know: use the tightest inequality whose conditions you can genuinely verify, not the tightest one available.Read more →

How much sharper is a bound that uses independence?

+
It can be many orders of magnitude sharper. For at least 600600 heads in 10001000 fair tosses, Chebyshev's inequality sees only the variance 250250 and allows up to 2501002=0.025\frac{250}{100^2} = 0.025. Hoeffding's inequality also uses that the tosses are independent and each adds 00 or 11, and gives e−20≈2×10−9e^{-20} \approx 2 \times 10^{-9}. The exact chance is about 1.4×10−101.4 \times 10^{-10}.Read more →

Can Markov's inequality be applied to higher powers of X?

+
Yes, and that is often how sharper bounds are built. For X≥0X \ge 0, XkX^k is never negative, so P(X≥a)≤E[Xk]akP(X \ge a) \le \frac{E[X^k]}{a^k}, which falls like 1ak\frac{1}{a^k}. Chebyshev's inequality is the case k=2k = 2 applied to X−μX - \mu. For an exponential XX with mean 11 and a=10a = 10, Markov gives 0.10.1, Chebyshev about 0.0120.012, and the tenth power 10!1010≈0.00036\frac{10!}{10^{10}} \approx 0.00036.Read more →