6 Common Probability Inequalities

1 Boole's Inequality

By disjoint additivity we can compute P(A1∪A2)=P(A1)+P(A2)−P(A1∪A2).
Actually we have the general Inclusion-Exclusion Principle: (1.1)P(⋃i=1nAi)=Σ1−Σ2+Σ3−Σ4+⋯+(−1)n−1Σn, where Σk=∑1≤i1<⋯<ik≤nP(Ai1⋯Aik), which can be proved by induction and definition. However we can have a neater result when it comes to inequality:

Theorem (Union Bound, Boole's Inequality)

A1,⋯,An∈F are a collection of events on a probability space (Ω,F,P). Then P(⋃i=1nAi)≤∑i=1nP(Ai).

This can be proved by induction on n and (1.1), so is omitted.

2 Cauchy-Schwarz Inequality

Theorem (Cauchy-Schwarz Inequality)

Let X,Y be two RVs on the same probability space. Then (E[XY])2≤E[X2]E[Y2], i.e. (Cov(X,Y))2≤Var(X)Var(Y).

3 Concentration Inequalities

We often want to estimate the tail, i.e. P(X≥c) or P(|X−μ|≥c). This is important for proving convergence results, bounding failure probabilities, or probabilistic bounds on runtimes.

3.1 Markov's Inequality

Theorem (Markov's Inequality)

Let X be a non-negative RV with E[|X|]<∞. Then for any constant c>0, P(X≥c)≤E[X]c.

Theorem (Generalized Markov's Inequality)

Let X:Ω→R be an arbitrary RV. Then for all constants c>0,k>0, P(|X|≥c)≤E[|X|k]ck.

3.2 Chebyshev's Inequality

Theorem (Chebyshev's Inequality)

For all RVs X with E[X]=μ<∞ and for all constants c>0, P(|X−μ|≥c)≤Var(X)c2.

3.3 Chernoff Inequalities

Theorem (Chernoff Inequalities)

A one-parameter family of bounds derived as follows:

  1. ∀t>0,c∈R, P(X≥c)≤mint>0MX(t)etc.
  2. ∀t<0,c∈R, P(X≤c)≤mint<0MX(t)etc.

3.4 Hoeffding's Inequality

This inequality is specific for bounded RVs.

Theorem (Hoeffding's Inequality)

Let X1,⋯,Xn be independent RVs with E[Xi]=μi<∞, and P(ai≤Xi≤bi)=1 for some constants ai,bi∈R. Let Sn=X1+⋯+Xn, then ∀ε>0, P(|Sn−E[Sn]|≥ε)≤2exp⁡[−2ε2∑i=1n(bi−ai)2].
Equivalently, P(|Snn−E[Sn]n|≥ε)≤2exp⁡[−2ε2n2∑i=1n(bi−ai)2].

Hoeffding's bound only depends on the range of Xi, not its distribution over [ai,bi]. Using more information about Xi can lead to a sharper bound, like Chernoff.

Given the above two lemmas, we can prove Hoeffding's inequality.
First, since E[expt(Sn−∑i=1nμi)]=E[∏i=1net(Xi−μi)]=∏i=1nE[et(Xi−μi)]≤∏i=1net22(bi−ai2)2=et22∑i=1n(bi−ai2)2,
("≤" by lemma 2) so Sn is sub-Gaussian with variance proxy σ2=∑i=1n(bi−ai)24.
Next, let X=Sn and apply lemma 1, we complete the proof.