## Saturday, January 07, 2006

### The 0-1 Law

To start off the year 2006, I want to talk about the 0-1 law, which is an extremely important and elegant topic in finite model theory. A logic L (over a fixed vocabulary σ) is said to admit the 0-1 law if, for any sentence f of L, the fraction of σ-structures with universe {0,...,n-1} that satisfy the sentence f converges to either 0 or 1 as n approaches infinity. We denote this fraction by μn(f) and the limit, if exists, by μ(f). [Of course, this means that μ and μn ranges between 0 and 1.] In other words, every sentence f is almost surely false or almost surely true. For simplicity, let us restrict our attention to undirected loopless graphs, i.e., σ contains only one binary relation E which is both symmetric and irreflexive. Below is perhaps the most fundamental theorem about the 0-1 law.

Theorem (Fagin 1976, Glebskii et al. 1969): First-order logic admits the 0-1 law

Most finite model theorists equate the admission of 0-1 law with the inability to do "counting". Consider the following properties EVEN = { G : G has even number of vertices }. We see that μn(EVEN) = 1 iff n is even, and hence μ(EVEN) does not exist. I am yet to see a property whose asymptotic probability μ converges to something other than 0 or 1, or diverges, but has nothing to do with counting. Consider the following properties:

1. μ(BIPARTITE) = 0
2. μ(HAMILTONIAN) = 1
3. μ(CONNECTED) = 1
4. μ(TREE) = 0

On the other hand, if you admit constant or function symbols, it is possible to have first-order properties whose asymptotic probability converges to something other than 0 or 1. So, it is crucial to assume that our vocabulary σ be purely relational.

What is the use of 0-1 law in finite model theory apart from the amusement of some theoreticians? [I will have to confess many proofs about the 0-1 law, including Fagin's proof of the above theorem, are usually of great beauty.] Well, a primary concern of finite model theory is to determine how expressive a logic is over a given class of finite structures. By proving that a logic admits the 0-1 law, we show that any property whose asymptotic probability does not converge to either 0 or 1 is not expressible in the logic. For example, an immediate corollary of the above theorem is that first-order logic cannot express EVEN.

Let me briefly outline the proof ideas of the above theorem which recurs in 0-1 law proofs. We only need to show that, for each natural number k, there exists a property Pk of graphs such that

1. μ(Pk) = 1, and
2. Any two models of Pk cannot be distinguished by a first-order sentence of quantifier rank k. [Recall that the quantifier rank of a first-order sentence is the maximum depth of any nesting of quantifiers of the sentence.]

Why is this sufficient? Let f be a first-order sentence of quantifier rank k. There are two cases. First, there exists a model of Pk that is also a model of f. In this case, it is not hard to show that any model of Pk is also a model of f, which implies that μ(f) ≥ μ(Pk) = 1. The second case is that all models of Pk are not models of f, i.e., they are models of "not f". Therefore, μ("not f") ≥ μ(Pk) = 1. That is, we have μ(f) = 0.

The property Pk is commonly referred to as an extension axiom. Note that Pk does not have to be L-definable (where L is the logic that is to admit 0-1 law).

Two excellent expositions of the basics of 0-1 laws include Leonid Libkin's Elements of Finite Model Theory chapter 12, and James Lynch DIMACS 1995-1996 Tutorial notes on "Random Finite Models".

Andy D said...

Again, thanks for some great posts!

Can you tell us anything about quantitative strengthenings of 0-1 laws? What are the typical convergence rates of the probabilities as model size grows?

Also, for places/logics where the 0-1 law fails, are there any remnants of the phenomenon?

twidjaja said...

Great questions!

> Can you tell us anything about
> quantitative strengthenings of 0-1 laws?
> What are the typical convergence rates
> of the probabilities as model size
> grows?

If I understand your question correctly, the most well-known quantitative strengthening of 0-1 laws (for first-order logic) is probably Spencer-Shelah theorem.
Let me first fix some notations before stating the theorem. The random graph model (G_n,p(n)), where G_n is the set of all graphs with vertices {0,...,n-1}, and p(n) is a function from G_n to [0,1], is basically the probability space that you get by picking each edge with probability p(n). You can check that we used the model (G_n,1/2) in the posting. Random graph models of the form (G_n,c), where c is a constant, is called uniform random graph models. However, graph theorists like Joel Spencer mostly deal with non-uniform probability model. Consider the function p_c(n) = 1/n^c, where c is a real number. [This means that one deals with "sparse" graphs.] With respect to the model (G_n,p_c(n)), when does first-order logic admit the 0-1 law? Spencer-Shelah theorem says that it admits the 0-1 law iff c is irrational. This is perhaps the deepest theorem wrt 0-1 law, which is covered in full detail in Spencer's book "The Strange Logic of Random Graphs".

> Also, for places/logics where the 0-1
> law fails, are there any remnants of the
> phenomenon?

0-1 law is actually only one of the many properties that logicians concern. It is the strongest, and the most useful, of such properties by the way. In the case when a logic fails to have the 0-1 law, logicians consider some other properties like:
1. Convergence. A logic has the convergence property if mu(f) always exists for each sentence f in the logic.
2. Slow convergence rate (if 1 fails). A logic has the slow convergence rate property if mu_{n+1}(f)-mu_{n}(f) converges.

Lynch's tutorial is probably the best source of information on the issues above.

Andy D said...

Thanks!

Spencer-Shelah sounds fascinating, although that's not the kind of result I was thinking about.

I wanted to know, for the (G_n,1/2) model, if you fix a first-order graph property f with mu(f) = 0 (/1), how fast does the probability mu_{n}(f) of satisfying it approach 0 (/1) as n increases? Is it always exponentially, or slower in some cases?

twidjaja said...

I checked with my supervisor and the calculation in his book (chapter 12). It is exponentially fast.

Anonymous said...

Great post, very useful indeed.