Lecture 8 - The Probabilistic Method
约 2028 个字 预计阅读时间 7 分钟
- Speaker: Prof. Alexander Scott
It can be very difficult to construct mathematical objects without embedding some sort of regular structure.
Random or typical objects often have desirable properties that are difficult to construct explicitly.
Example 1: Tournaments
A tournament is an orientation of a complete graph: every edge between \(x\) and \(y\) is assigned a direction (towards \(x\) or towards \(y\)).
A tournament has Property \(P_{k}\) if for every set of \(k\) players, there is someone who beats all of them. For example, the cyclic tournament on three vertices has Property \(P_{1}\). For large \(k\), tournaments with Property \(P_{k}\) are difficult to construct!
Theorem
If \(n \geq k^{2} 2^{k + 1}\), then some tournament with \(n\) vertices has Property \(P_k\).
Idea:
- Consider a random tournament.
- Show that with positive probability it has Property \(P_{k}\).
- Deduce that some tournament must have this property.
Proof
Let \(T\) be a random tournament on \(n\) vertices, where each edge between \(x\) and \(y\) is directed independently with equal probability towards \(x\) or \(y\). For each set \(S\) of \(k\) vertices, let \(B_{S}\) be the event that no vertex beats every vertex in \(S\) (we say that \(S\) is then a bad set).
Then
Here we have used the inequality \(1 + x \leq e^{x}\), which holds for all real numbers \(x\) (and is extremely useful).
We deduce that the expected number of bad \(k\)-sets \(B\) is at most
Suppose this expectation is strictly less than \(1\). If on average we have less than \(1\) bad set, then there must be some tournament where we have less than \(1\) bad set (the minimum is at most the average).
Thus, it is sufficient to show that
Recall that \(n \geq k^{2} 2^{k + 1}\); then
It is straightforward to see that this is less than \(1\). \(\square\)
Let's note:
- We did not directly show that \(T\) exists: we showed that a random tournament had a strictly positive probability of having the required properties.
- We needed to perform some calculations!
We can consider random structures, even when our original problem does not mention randomness.All the tools of Probability Theory are now at our disposal!
Example 2: Coding Theory
Coding Theory addresses the problem of transmitting information through a binary channel: in other words, we aim to send information as a sequence of \(0\)s and \(1\)s.
A code is a collection of binary strings, one for each type of information we aim to transmit (for example, one string for each letter of our alphabet).
Sometimes the channel is noisy, in which case we require our strings to be highly distinct (we need an error-correcting code). Even without noise, important questions arise. For example, how quickly can we transmit information through a channel?
A prefix of a binary string is an initial segment. For example, \(010\) is a prefix of \(0101\) but not of \(1010\).
A set \(F\) of binary strings is prefix-free if no string in \(F\) is a prefix of another. A fundamental theorem, the Kraft-McMillan Inequality, applies to prefix-free codes:
Theorem (Kraft-McMillan Inequality)
Let \(F\) be a prefix-free set of binary strings, and suppose that \(F\) contains \(N_{i}\) strings of length \(i\) for each \(i\). Then
Proof
Consider a random (infinite) sequence \(B = b_{1} b_{2} \dots\). For a string \(C\) of length \(k\),
Thus, the expected number of strings from \(F\) that occur as a prefix of \(B\) is
On the other hand, we can never have more than one string from \(F\) as prefixes simultaneously, so we deduce that
(the average is at most the maximum). \(\square\)
Example 3: Max Cut
In the first example, the linearity of expectation is straightforward: if \(X = \sum X_{i}\), then
It also applies to variance if the random variables are independent.
In the Max Cut problem, we are given a graph \(G\) and aim to divide its vertices into two classes \(V_{1}, V_{2}\) so that as many edges as possible have ends in both classes.
The Max Cut problem is known to be a challenging algorithmic problem. (It is NP-hard; in fact, it is NP-hard even to find a good approximate solution!)
A theorem provides a simple bound for the Max Cut problem:
Theorem
For every graph \(G\), there exists a partition \(V(G) = V_{1} \cup V_{2}\) such that at least half the edges of \(G\) have one end in each class.
Proof
Consider a random partition. For each edge \(e = (x, y)\), we define a random variable \(X_{e}\) by setting \(X_{e} = 1\) if \(e\) has one end in each set and \(X_{e} = 0\) otherwise. Let
We aim to show that there exists some partition in which \(X \geq \frac{e(G)}{2}\).
Observe that
Thus, by linearity of expectation
It follows that there exists some partition for which at least half the edges have one end in each class. \(\square\)
It is possible to derandomise this argument to obtain a very fast algorithm.
Example 4: Independent Sets
In the alteration method, we generate a random structure and then modify it to achieve the desired properties.
For example: an independent set in a graph is a set of vertices, none of which are joined by edges.
Theorem
Let \(G\) have \(n\) vertices and average degree \(d\). Then \(G\) contains an independent set of size at least \(\frac{n}{2d}\).
Proof
We generate a set in two steps. First, let \(S\) be a random subset of \(V(G)\) obtained by including each vertex independently with probability \(p\). Then, let \(T\) be obtained from \(S\) by deleting one end from each edge.
The expected size of \(S\) is
The expected number of edges contained in \(S\) is
Thus, on average, the number of vertices remaining is
Setting \(p = \frac{1}{d}\) yields
\(\square\)
Broader Horizons: Random Graphs
A highly important and fascinating example is provided by random graphs. The theory of random graphs was initially developed in the 1960s, but it has since grown into a significant and influential area of research, with connections to numerous fields.
In the \(G(n, p)\) model, we consider an \(n\)-vertex graph in which each edge is present independently with probability \(p\). Thus, \(G(n, 0)\) has no edges, \(G(n, 1)\) is the complete graph, and \(G(n, \frac{1}{2})\) has (on average) approximately half the edges.
Random graphs are used to model numerous real-world processes, from social networks to epidemics. A deep theory exists regarding the various changes that a typical random graph undergoes as \(p\) increases from \(0\) to \(1\).
What can we say about the typical structure of a graph in \(G(n, p)\), where \(p = p(n)\) depends on \(n\)?
Let us consider triangles. How large must \(p\) be for triangles to appear in \(G(n, p)\)? The expected number of triangles is given by
Thus, if \(p \ll \frac{1}{n}\), then the expected number of triangles tends to \(0\). It follows (for example, by Markov's Inequality) that the probability that \(G\) contains a triangle tends to \(0\).
On the other hand, if \(p \gg \frac{1}{n}\), then the expected number of triangles tends to infinity. Does this imply that the probability of obtaining a triangle tends to \(1\)?
To address this, we need to consider the variance.
Idea:
- Consider a random graph \(G(n, p)\), and let \(X\) denote the number of triangles
- Calculate the mean and variance of \(X\)
- Use Chebyshev's Inequality to show that it is highly unlikely that \(X = 0\)
For each triple of vertices \(A = \{x, y, z\}\), we define a random variable \(X_{A}\) as follows:
We know that \(\mathrm{E}(X) \sim \frac{n^{3}p^{3}}{6}\). Let us calculate its variance \(\sigma^{2}\). We have
where the sum is taken over all pairs of \(A\) and \(B\).
If \(A\) and \(B\) are disjoint, then \(X_{A}\) and \(X_{B}\) are independent. Thus, \(\text{cov}(A, B) = 0\). In fact, if \(|A \cup B| = 1\), then they remain independent.
Thus
If \(|A \cup B| = 2, then\)
and if \(|A \cup B| = 3\), then \(A = B\),
Thus
We now apply Chebyshev's Inequality:
By setting \(t\) to different values, one can determine how large \(p\) must be for the expected number of triangles to meet certain conditions.
We have demonstrated:
- If \(np \rightarrow 0\), then \(P(G \ \text{contains a triangle}) \rightarrow 0\)
- If \(np \rightarrow \infty\), then \(P(G \ \text{contains a triangle}) \rightarrow 1\)
We say that \(p(n) = \frac{1}{n}\) is a threshold function for the presence of triangles in \(G\).
We observe the same pattern for other graphs: there exists a threshold \(p_{H}\) such that if \(p \ll p_{H}\), then it is highly unlikely that \(G(n, p)\) contains a copy of \(H\); but if \(p \gg p_{H}\), then \(G(n, p)\) will almost certainly contain copies. This is an example of a phase transition.
For many graphs \(H\), the threshold for \(G\) to contain a copy of \(H\) occurs around the point where the expected number of copies of \(H\) becomes large.
Finally, let us briefly discuss martingale methods. A martingale is (informally) a sequence of random variables \(X_{0}, X_{1}, \dots\) where \(\mathrm{E}(X_{i + 1} | X_{i}) = X_{i}\): at each step, the expected value remains unchanged. (This can be likened to making a sequence of fair bets in a casino.)
Martingales are highly useful in analysing various types of random graph processes. For example, let \(\chi(G)\) denote the chromatic number of \(G\). Estimating the average value of \(\chi(G)\) for a random graph is challenging, but it is known to be approximately
The key idea is as follows:
- Reveal the graph one vertex at a time: let \(X_{i}\) denote the expected chromatic number of \(G\) given the information provided by the first \(i\) vertices.
- The sequence \((X_{i})_{i = 0}^{n}\) forms a martingale!
- Now apply a martingale inequality (for example, the Hoeffding Inequality, a relative of the Chernoff Inequality for binomial random variables) to show that \(X\) is likely to be close to its mean.
The result follows elegantly!