Help in probabilistic proof of the Caro-Wei theorem

208 Views Asked by At

While studying Turan's theorem, I would like to ask about some parts of this proof. I am weak at probability, so unless I spread and open everything up and formalize, I would not be able to understand. I have checked some posts in MSE too but either they asked different things or I just could not understand them.

I want to know why every graph $G$ of $n$ vertices has a clique of size $\geq \sum_{v\in V(G)}\frac{1}{n-d(v)}$. So, here is my understanding so far:

In a few notes I have checked, they start by considering a random permutation $p$ on $V=V(G)$, and consider the set $Q_p$, which is the set of vertices adjacent to all vertices on their left in $p$. I can see why this is a clique. Next, consider the random variable $X=|Q_p|$. Here is my confusion: (Question 1) Does this mean the sample space is $S=\{Q_p\mid p \text{ is an permutation of }V\}$, or just $S=\{p\mid p \text{ is a permutation of }V\}$?

From that definition of $X$, I really thought it can only mean the first sample space, but now I am guessing that it is actually the second one. That means the random variable $X$ assigns any $p$ to $|Q_p|$, am I right?

And so, if $v$ is any vertex, consider the random variable $X_v$, defined such that $X_v(p)=1$ if $v\in Q_p$ and $0$ otherwise. Then, $X=\sum_{v\in V(G)}X_v$. (Question 2: If I use the first $S$ instead and then define $X_v(Q_p)$ in the same way, then this summation still holds (whether it is useful or not), right?).

From here, I simply need to use linearity of expectation, but I need the quantity $E[X_v]=P(X_v=1)$. To get this, I try to count everything, which is

$$P(X_v=1)=P(\{p\mid X_v(p)=1\})=\frac{|\{p\mid X_v(p)=1\}|}{n!}.$$

The numerator is the same as the number of $p$ such that $v\in Q_p$, or eqv such that $v$ appears before any of its non-neighbours (there are $n-d(v)-1$ of them without $v$ itself). To count this, we first put $v$ and all non-neighbors of $v$ in the $n$ boxes (done in $\binom{n}{n-d(v)}$ ways), permute them except $v$ since it has to appear first (done in $(n-d(v)-1)!$ ways), and permute the rest in $d(v)!$ ways. Thus,

$$P(X_v=1)=\frac{\binom{n}{n-d(v)}(n-d(v)-1)!d(v)!}{n!}=\frac{\frac{1}{(n-d(v))!}(n-d(v)-1)!}{1}=\frac{1}{n-d(v)}.$$

And we are done. Is this a correct understanding? If I use the first sample space instead, can this result be obtained too? I am having trouble counting $Q_p$ if I use the first $S$.

Lastly, without counting everything, is there any other way to get the $\frac{1}{n-d(v)}$ quicker? Thank you.

EDIT: The rest follows from first moment method.

2

There are 2 best solutions below

17
On BEST ANSWER

The correct sample space is in fact the set of all permutations of $V$. That's why the solution begins "consider a random permutation of $V$". If the solution were written a bit more carefully, it would begin "consider a uniformly random permutation of $V$", which tells you that you are sampling uniformly from the set of all permutations.

Using a random $Q_p$ gets you into trouble. If your sample space is "subsets $Q_p$ of $V$", then you're not sampling uniformly. So how are you sampling? You don't know. Then you can't compute any probabilities.

As for the quick way to see that $\Pr[X_v = 1] = \frac1{n - d(v)}$, let $S \subseteq V$ consist of $v$, together with all vertices not adjacent to $v$. We have $X_v = 1$ if $v$ is the leftmost vertex of $S$ in the permutation (no matter where vertices of $V-S$ go).

A uniformly random permutation of $V$ induces a uniformly random permutation of $S$. The probability that $v$ is the first vertex in this uniformly random permutation of $S$ is $\frac1{|S|}$, by symmetry: any of the $|S|$ vertices in $S$ is equally likely to be first. So $\Pr[X_v = 1] = \frac1{|S|} = \frac1{n-d(v)}$.

0
On

I like to think of Caro Wei's theorem just as the expected result of the greedy algorithm to get an independent set.

You start out with an empty set and consider each vertex once in a specific order, if there are no neighbours of the vertex in the set then you insert the vertex, otherwise you don't.

If the order in which you process the vertices is random then vertex $v$ will be inside the set with probability at least $\frac{1}{d(v)+1}$ (because if $v$ is the first element among the $d(v)+1$ vertices consisting of itself and its $d(v)$ neighbours then $v$ will definitely be inserted). Lineality of expectation yields the desired bound for the expected value of the final set.

This way of looking at it also shows clear avenues which could be followed to improve the results.