Skip to main content

The Probabilistic Method


The Notion of Probability

Definition

Let Ω\Omega be a finite set of outcomes of some sequence of trials, so that all these outcomes are equally likely. Let AΩA \subseteq \Omega. Then Ω\Omega is called a sample space, and AA is called an event. The ratio

P(A)=AΩP(A)=\frac{|A|}{|\Omega|}

is called the probability of AA.

Proposition

Let A1,A2,,AnA_1, A_2, \cdots, A_n be events from the same sample space. Then

P(A1A2An)P(A1)+P(A2)++P(An).P\left(A_1 \cup A_2 \cup \cdots \cup A_n\right) \leq P\left(A_1\right)+P\left(A_2\right)+\cdots+P\left(A_n\right) .

Proof

We simply have to show that

A1AnA1++An.\left|A_1 \cup \cdots \cup A_n\right| \leq\left|A_1\right|+\cdots+\left|A_n\right| .

This is true as the left-hand side counts each element of the sample space that is part of at least one of the AiA_i exactly once, while the right-hand side counts each element of the sample space that is part of at least one of the AiA_i at least once.

Non-constructive Proofs

Theorem

For all positive integers k3k \geq 3, the inequality R(k,k)>R(k, k)> 2k/22^{k / 2} holds.

Proof

Let G=KnG=K_n, and let us color each edge of GG red or blue as follows. For each edge, flip a coin. If we get a head, we color that edge red, otherwise we color that edge blue. This way each edge will be red with probability one half, and blue with probability one half. We are going to show that the probability pp that we get no monochromatic KkK_k-subgraphs in GG this way is more than zero. On the other hand, p=FΩp=\frac{|F|}{|\Omega|}, the number of favorable outcomes divided by the number of all outcomes, where Ω\Omega is the set of all possible 2-colorings of the edges of a complete graph on nn vertices. So p>0p>0 implies that there is at least one favorable outcome, that is, there is at least one KnK_n with 2-colored edges that does not contain any monochromatic KkK_k-subgraphs.

Instead of proving that p>0p>0, we will prove that 1p<11-p<1, which is an equivalent statement. Note that 1p1-p is the probability that we get at least one monochromatic subgraph in our randomly colored graph G=KnG=K_n. The number of ways to 2-color the edges of a given KkK_k-subgraph of KnK_n is clearly 2(k2)2^{\left(\begin{array}{c}k \\ 2\end{array}\right)} as there are two choices for the color of each edge. Out of all these colorings, only two will be monochromatic, one with all edges red, and one with all edges blue. Therefore, the probability that a randomly chosen KkK_k-subgraph is monochromatic is

22(k2)=21(k2)\frac{2}{2^{\left(\begin{array}{c} k \\ 2 \end{array}\right)}}=2^{1-\left(\begin{array}{c} k \\ 2 \end{array}\right)}

The graph KnK_n has (nk)\left(\begin{array}{l}n \\ k\end{array}\right) subgraphs that are isomorphic to KkK_k. Each of them has the same chance to be monochromatic. On the other hand, the probability that at least one of them is monochromatic is at most the sum of these (nk)\left(\begin{array}{l}n \\ k\end{array}\right) individual probabilities, by the proposition In other words, if ASA_S denotes the event that the KkK_k-subgraph SS of GG has monochromatic edges, then

P(SAS)SP(AS)=(nk)21(k2)P\left(\bigcup_S A_S\right) \leq \sum_S P\left(A_S\right)=\left(\begin{array}{l} n \\ k \end{array}\right) 2^{1-\left(\begin{array}{l} k \\ 2 \end{array}\right)}

where SS ranges through all KkK_k-subgraphs of GG. Now let us assume, in accordance with our criterion, that n2k/2n \leq 2^{k / 2}. Then the last term can be bounded as follows.

(nk)21(k2)<nkk!21(k2)22k2/2k!2(k2)=22k/2k!<1,\left(\begin{array}{l} n \\ k \end{array}\right) 2^{1-\left(\begin{array}{l} k \\ 2 \end{array}\right)}<\frac{n^k}{k !} \cdot 2^{1-\left(\begin{array}{l} k \\ 2 \end{array}\right)} \leq \frac{2 \cdot 2^{k^2 / 2}}{k ! 2^{\left(\begin{array}{l} k \\ 2 \end{array}\right)}}=2 \cdot \frac{2^{k / 2}}{k !}<1,

for all k3k \geq 3. The last inequality is very easy to prove, for example by induction.

Conditional Probability

Definition

If AA and BB are two events from the same sample space Ω\Omega, and P(AB)=P(A)P(B)P(A \cap B)=P(A) \cdot P(B), then AA and BB are called independent events. Otherwise they are called dependent.

Definition

Let AA and BB be events from the same sample space, and assume P(B)>0P(B)>0. Let

P(AB)=P(AB)P(B)P(A \mid B)=\frac{P(A \cap B)}{P(B)}

Then P(AB)P(A \mid B) is called a conditional probability, and is read "the probability of AA given BB ".

That is, P(AB)P(A \mid B) is the probability of AA given that BB occurs. The following proposition is now immediate from the definitions.

Proposition

The events AA and BB are independent if and only if P(AB)=P(A)P(A \mid B)=P(A) holds

Bayes' Theorem

Let AA and BB be mutually exclusive events so that AB=ΩA \cup B = \Omega. Let CC be any event. Then

P(C)=P(CA)P(A)+P(CB)P(B).P(C) = P(C \mid A) \cdot P(A) + P(C \mid B) \cdot P(B).

Proof

As AA and BB are mutually exclusive, ACA \cap C and BCB \cap C are disjoint, and since AB=ΩA \cup B = \Omega, their union is exactly CC. Therefore,

P(C)=P(CA)+P(CB),P(C) = P(C \cap A) + P(C \cap B),

and the proof follows as the first (resp. second) member of the right-hand side agrees with the first (resp. second) member of the right-hand side of the Bayes' Theorem.