RANDOM SPARSE UNARY PREDICATES 1 Introduction.

Let n be a positive integer, 0 p 1. The random unary predicate U n,p is a probability space over predicates U on [n] = {1, . . . , n} with the probabilities determined by Pr[U (x)] = p, 1 x n and the events U (x) being mutually independent over 1 x n. Informally, we think of ipping a coin for each x to determine if U (x) holds, the coin coming up heads with probability p. We shall examine the rst order language < [n], , U > with equality, a unary predicate U and a binary predicate . Examples of sentences in this language are: A : x U (x) B : x U (x) y y < x C : x,y U (x) U (y) z [x < z z < y] (>, , < are natuarally denable from and equality.) For any such sentence S we have the probability
modified:1994-01-29

Pr[Un,p |= S] While the use of unary predicates is natural for logicians there are two other equivalent formulations that will prove useful. We may think of U as a subset of [n] and speak about i U rather than U (i). Second we may associate with U a sequence of zeroes and ones where the i-th term is one if U (i) and zero if U (i). Thus we may talk of starting at i and going to the next one. We shall use all three formulations interchangably. Ehrenfeucht [??] showed that for any constant p and any sentence S in this language lim Pr[Un,p |= S]
n

revision:1994-01-28

exists. In the case of sentences A and C the limiting probability is one. But sentence B eectively states 1 U , hence its limiting probability is p. We get around these edge eects with a new language, consisting of equality, a unary predicate U , and a ternary predicate C. We consider C as a built in predicate on [n] with C(x, y, z) holding if and only if either x < y < z 1

432

or y < z < x or z < x < y. Thinking of [n] as a cycle, with 1 coming directly after n, C(x, y, z) is the relation that x to y to z goes in a clockwise direction. For any sentence S in this new language we can again dene Pr[Un,p |= S] only in this case Ehrenfeuchts results give a Zero-One Law: for any constant p and sentence S
n

lim Pr[Un,p |= S] = 0 or 1

modified:1994-01-29

We shall call the rst language the linear language and the second language the circular language. As a general guide, the circular language will tend to Zero-One Laws while the linear language, because of edge eects, will tend to limit laws. We shall not restrict ourselves to p constant but rather consider p = p(n) as a function of n. We have in mind the Evolution of Random Graphs as rst developed by Erds and Rnyi. Here as p = p(n) evolves from zero o e to one the unary predicate evolves from holding for no x to holding for all x. Analogously (but without formal denition) we have threshold functions for various properties. For example, p(n) = n 1 is a threshold property for A. When p(n) n1 almost surely A fails while when p(n) n 1 almost surely A holds. In Shelah,Spencer [??] we showed that when p = n with (0, 1), irrational then a Zero-One Law held for the random graph G(n, p) and in Luczak, Spencer [??] we found a near characterization of those p = p(n) for which the Zero-One Law held. The situation with random unary predicates turns out to be somewhat simpler. Let us say p = p(n) satises the Zero-One Law for circular unary predicates if for every sentence S in the circular language
n

lim Pr[Un,p(n) |= S] = 0 or 1

Here is our main result. Theorem 1. Let p = p(n) be such that p(n) [0, 1] for all n and either p(n) n1

revision:1994-01-28

or for some positive integer k n k or for all >0 n p(n) and n 2 1 p(n)
1

p(n)

n k+1

432

or for some positive integer k n k or 1 p(n) n1 Then p(n) satises the Zero-One Law for circular unary predicates. Inversely if p(n) falls into none of the above categories then it does not satisfy the Zero-One Law for circular unary predicates. The inverse part is relatively simple. Let A k be the sentence that there exist k consecutive elements x1 , . . . , xk U . (x, y are consecutive if for no z is C(x, z, y). For k = 2 this is example C. ) Then Pr[A k ] is (for a given n) a monotone function of p. When p(n) cn 1/k and c a positive constant the probability Pr[Ak ] approaches a limit strictly between zero and one. (Roughly speaking, n1/k is a threshold function for Ak .) Thus for p(n) to satisfy the Zero-One law we must have p(n) n 1/k or p(n) 1/k . Further (replacing U with U ), the same holds with p(n) replaced n by 1 p(n). For p(n) to fall between these cracks it must be in one of the above ve categories.
1

1 p(n)

n k+1

modified:1994-01-29

Remark. Dolan [??] has shown that p(n) satises the Zero-One Law for linear unary predicates if and only if p(n) n 1 or n1 p(n) n1/2 1 or n1 1/2 . For n1/2 or 1 p(n) n 1 p(n) n p(n) = o(1) he considered the following property: D : x U (x)[U (x+1)U (x+2)]y [U (y)[U (y+1)U (y+2)]y < x]U (x+1) (Addition is not in our language but we write x + 1 as shorthand for that z for which x < z but there is no w with x < w < z.) In our zero-one formulation D basically states that the rst time we have 11 comes before the rst time we have 101. This actually has limiting probability .5. This example illustrates that limiting probability for linear unary predicates can depend on edge eects and not just edge eects looking at U on a xed size set 1, . . . , k or n, n 1, . . . , n k. We defer our results for linear unary predicates to section 4.

revision:1994-01-28

When p(n) n1 the Zero-One Law is trivially satises since almost surely there is no x for which U (x). Also, if p(n) satises the Zero-One Law 3

432

so does 1 p(n). Suppose p = p(n) satises p(n) n and 1 p(n) n for all > 0. We show in a section 3 that for every t there is a sequence A1 AR with the property that for any sentence A of quantier depth t either all models < [u], C, U > that contain A 1 AR as a subsequence satisfy A or no such models satisfy A. (< [u], C, U > contains A 1 AR as a subsequence if for some 1 j u for all 1 i R we have U (j + i) if and only if xi = 1, with j + i dened modulo u.) For p(n) in this range < [u], C, U > almost surely contains any such xed sequence A 1 AR as a subsequence and hence the Zero-One Law is satised. This leaves us with only one case in Theorem 1, and that will be the object of the next section.

The Main Case.

Here we let k be a positive integer and assume n k p(n) n k+1

revision:1994-01-28

Our object is to show that p(n) satises the Zero-One Law for circular unary predicates. We shall let t be a xed, though arbitrary large, positive integer. We shall examine the equivalence class under the t-move Ehrenfeucht game of the circular model. For the most part, however, we shall examine linear models. We dene (as Ehrenfeucht did) an equivalence class on models M =< n, , U >, two models M, M being equivalent if they satisfy the same depth t sentences or, equivalently, if the t-move Ehrenfeucht game on M, M is won by the Duplicator. The addition of models (with M on [n], M on [n ] we dene M +M on [n+n ]) yields an addition of equivalence classes. We shall denote the equivalence classes by x, y, . . . and the sum by x + y. Results from the beautiful theory of these classes are given in Section 3. Let us consider a random unary function U dened on all positive integers 1, 2, . . . and with Pr[U (i)] = p for all i, these events mutually independent. (In the end only the values of U (i) for 1 i n will count but allowing U to be dened over all positive integers allows for a ctitious play that shall simplify the analysis.) Now for any starting point i examine i, i + 1, . . . until reaching the rst j (perhaps i itself) for which U (j). Call [i, j] the 1-interval of i. (With probability one there will be such a j; ctitious play allows us to postpone the analysis of those negligible cases when no j is found before j > n.) What are the possible Ehrenfeucht values of < [i, j], , U >? The model must have a series of zeroes (i.e., U ) followed 4

432

modified:1994-01-29

revision:1994-01-28

by one one (i.e., U ). There is an s (s = 3 t will do) so that all such models with at least s zeroes have the same Ehrenfeuct value. We can write these values as a1 , . . . as and b (ai having i 1 zeroes, b having s zeroes). Call this value the 1-value of i. The probability of the 1-value being any particular ai is p while the probability of it being b is 1. (All asymptotics are as p 0.) We let E1 denote this set of possible 1-values and we split E1 = P1 T1 with P1 = {b} and T1 = {a1 , . . . , as }. The 1-values in T1 we call 1-transient, the 1-value in P1 we call 1-persistent. Now (with an eye toward induction) we dene the 2-interval of i = i 0 . Take the 1-interval of i, say [i0 , i1 ). Then take the 1-interval of i1 , say [i1 , i2 ). Continue until reaching a 1-interval [i u , iu+1 ) whose 1-value is 1-transient. (Of course, this could happen with the very rst interval.) We call [i, i u+1 ) the 2-interval of i. Now we describe the possible 2-values for this 2-interval. In terms of Ehrenfeucht value we can write the interval as b + b + . . . + b + a i where there are u (possibly zero) bs. Any b + . . . + b with at least s addends b has (see 3.4) the same value, call it B. Let jb denote the sum of j bs. We dene the transient 2-values T 2 as those of the form jb + ai with 0 j < s and the persistent 2-values P2 as those of the form B + ai . For example, let t = 5 and s = 35 = 243. Then i has 2-value 6b + a22 if, starting at i, six times there are at least 243 zeroes before a one and after the sixth one there are 21 zeroes and then a one. The 2-value is B + a 5 if at least 243 times there are at least 243 zeroes before the next one and the rst time two ones appear less than 243 apart they are exactly 5 apart. What are the probabilities for i = i0 having any particular 2-value? The rst 1-interval [i0 , i1 ) has distribution for 1-value as previously discussed: p for each a i and 1 for b. Having determined the rst 1-interval the values starting at i1 have not yet been examined. Hence the 1-value of the second 1-interval will be independent of the 1-value of the rst and, in general, the sequence of 1-values will be of mutually independent values. Then the transient 2values jb + ai each have probability p while the persistent 2-values B + a i will each have probability 1 + o(1). We let P2 denote the set of persistent s 2-values, T2 the set of transient 2-values and E2 = P2 T2 the set of 2-values. The 3-value will contain all the notions of the general case. Beginning at i = i0 take its 2-interval [i0 , i1 ). Then take successive 2-intervals [i1 , i2 ), . . . , [iu1 , iu ) until reaching an interval [iu , iu+1 ) whose 2-value is transient. The 3-interval for i is then [i, i u+1 ). Let x1 , . . . , xu , yu+1 be the 2-values for the successive intervals. Fromthe procedure all x i P2 while yu+1 T2 . Now consider (see 3.1)the Ehrenfeucht equivalence classes (again with respect to a t-move game) over P 2 . (A is the set of strings 5

432

modified:1994-01-29

revision:1994-01-28

over alphabet A.) Let be the equivalence class for the string x 1 xu , then the 3-value of i is dened as the pair = y u+1 . We let E3 be the set of all such pairs and we call persistent (and place it in P 3 ) if is a persistent state (as dened in 3.2)in P 2 ; otherwise we call transient and place it in T3 . If x1 xu and x1 xu are equivalent as strings in P2 then x1 + . . . + xu and x1 + . . . + xu have (as shown in 3.4) the same Ehrenfeucht value. So the 3-value of i determines the Ehrenfeucht value of the 3-interval of i though possibly it has more information. What are the probabilities for the 3-value of i? Again we get a string of 2-values z 1 z2 . . . whose values are mutually independent and we stop when we hit a transient 2-value. We shall see (in the course of the full induction argument) that the probability of having 3-value is c for persistent and c p for transient . Now let us dene k-interval and k-value, including the split into persistent and transient k-values by induction on k. Suppose E k , Pk , Tk have been dened. Beginning at i = i0 let [i0 , i1 ) be the k-interval and then take succesive k-intervals [i1 , i2 ), . . . , [iu1 , iu ) until reaching a k-interval [iu , iu+1 ) with transient k-value. Then [i, iu+1 ) is the k + 1-interval of i. (Incidentally, suppose U (i). Then [i, i + 1) is the 1-interval of i which is transient. But then [i, i + 1) is the 2-interval of i and is transient. For all k [i, i + 1) is the k-interval of i and is transient.) Let x 1 , . . . , xu , yu+1 be the succesive k-values of the intervals. Let be the equivalence class of x 1 xu in Pk . Then i has k + 1-value = yu+1 . This value is persistent if is persistent and transient if is transient. This denes E k+1 , Pk+1 , Tk+1 , completing the induction. Our construction has assured that the k-value of i determines the Ehrenfeucht value of the k-interval of i, though it may have even more information. Now let us x i and look at the distribution of its k-value V k . We show, by induction on k, that for every persistent Pr[V k = ] = c + o(1) while for every transient Pr[V k = ] = (c + o(1))p. Here each c is a positive constant. Assume the result for k and set p = Pr[V k = ] for all Ek . Let p be the probability that V k is transient so that p cp, c a positive constant. Let x 1 , . . . , xu , yu+1 be the successive kvalues of the k-intervals beginning at i, stopping at the rst transient value. We can assume these values are taken independently from the inductively dened distribution on Ek . The distribution of the rst transient value is the conditional distribution of V k given that V k is transient so the probability that it is some transient y is dy + o(1) where dy = cy / cy , the sum over all transient y . Note all dy are positive constants. The key to the argument is the distribution for the Ehrenfeucht equiva6

432

modified:1994-01-29

lence class for the nite sequence x 1 xu Pk . Let M be the set of all such equivalence classes. Let Lu be the event that precisely u persistent xs are found and then a transient y. Then Pr[L u ] = (1 p )u p precisely. For Pk let p+ = p /(1 p ), the conditional probability that V k = given that V k is persistent. Note that (as p 0) p+ p c Conditioning on Lu the x1 , . . . , xu are mutually independent with distributions given by the p+ . Dene on M a Markov Chain (see 3.3) with transition probability p+ from and to + . We let M (p) denote this Markov Chain. Observe that the set of states M is independent of p and the nonzeroness of the transition probabilities is independent of p (0, 1) though the actual transition probabilities do depend on p. There is a particular state O representing the null sequence. Let f (u, ) be the probability of being at state at time u, beginning at O at time zero. Then f (u, ) is precisely the conditional distribution for given L u . But therefore, letting W denote the Ehrenfeucht equivalence class,

Pr[W = ] =
u=0

f (u, )(1 p )u p

modified:1994-01-29

Let be the Markov Chain on the same set with transition probability c from to + and let f o (u, ) be the probability of going from O to in u steps under M o . Observe that M o is the limit of M (p) as p 0 in that taking the limit of any (1-step) transition probability in M (p) as p gives the transition probability in M o . Now we need some Markov Chain asymptotics. Assume is transient. We claim (recall p cp)

Mo

Pr[W = ] c
u=0

f o (u, ) p

revision:1994-01-28

and that the interior sum converges. In general the probability of remaining in a transient state drops exponentially in u so there exist constants K, so that f o (u, ) < K(1 )u for all u giving the convergence. Moreover there exists 1 , 2 , K1 so that for all 0 < p < 1 we bound uniformly f (u, ) < K1 (1 2 )u for all u. Pick 3 1 so that for 0 < p < 3 we have p 2cp. For any positive we nd U so that for 0 < p < 3
u=U f (u, )(1 p )u p < K1 (1 p u=U 2) up

2cK1
2

(1

2)

<

432

For any xed 0 u < U we have lim p0 f (u, ) = f o (u, ) so that lim
0u<U

p0

f (u, )(1 p )u p cf o (u, ) = p 0u<U

With U suciently large this may be made within /2 of c f o (u, ). 0 But this holds for arbitrarily small, giving the claimed asymptotics of Pr[W = ].

Remark. The rough notion here is that the probability of having a transient k + 1-value is dominated by having few persistent k-intervals and then a transient k-interval. The transient 2-intervals all had at most s persistent 1-intervals. The situation changes with 3-intervals. Recall Ba i consisted of at least s ones each preceeded by at least s zeroes and then two ones i apart. Consider an arbitrarily long grouping of 2-intervals of 2-value Ba i but, say, with none of the form Ba3 , i.e., 1001 not appearing and then, say, follow the last one, say Ba1 , with a one so that the 3- interval ends 111. For every u there is a cu p probability of this being the 3-interval with u such 2-intervals and cu > 0 but all such 3-intervals would be considered transient since a persistent sequence in P 2 must surely contain every value in P 2 .
modified:1994-01-29

Now suppose is persistent. Again we have the precise formula

Pr[W = ] =
u=0

f (u, )(1 p )u p

only this time it is the tail of the sum that dominates. As is persistent there is a limiting probability L = limu f o (u, ) with L > 0 and furthermore the M (p) approach M o in the sense that L = lim lim f (u, )
p0 u

revision:1994-01-28

We claim Pr[W = ] = L + o(1) For any > 0 there exist and U so that for p L < f (u, ) < L + 8 and u U we have

432

Then, as

u=0

L(1 p )u p = L,

| Pr[W = ] L|
u=U

(1 p )u p + (L + 1)
0u<U

(1 p )u p

For xed U the second sum is o(1) (as p 0) while the rst sum is less than so the entire expression is less than 2 for p suciently small. As was arbitrary this gives the claim. Recall that the k + 1-value of the full k + 1-interval is a pair consisting of the Ehrenfeucht value W just discussed and the k-value of the rst transient type yu+1 . The transient types value has a limiting distribution which is independent of W , for conditional on any L u the distribution on yu+1 is the same. All possible y Tk have a limiting probability dy (0, 1). Hence the probability of a k + 1-value being = y is simply the product of the probabilities and hence approaches a constant if , and hence , is persistent and is cp if , and hence is transient. This completes the inductive argument for the limiting probabilities of the k-values of the k-intervals. We now let L = Lk be the length of the k-interval of i and nd bounds on the distribution of L. A simple induction shows that if the sequence 1 1 of k ones appears after i then the k-interval of i ends with this sequence or possibly before. Thus we get the crude bound
modified:1994-01-29

Pr[L > ka] < (1 pk )a so that asymptotically P r[L > pk ] < ec where c is a positive constant. In fact, this gives the correct order of magnitude, L is (speaking roughly) almost always on the order of p k . We claim that there are positive constants t , ct so that Pr[Lt >
tp t

] > ct

revision:1994-01-28

The argument is by induction, for t = 1 the random variable L 1 is simply the number of trials until a success which occurs with probability p and the distribution is easily computable. Assume this true for t and let (as previously shown) et p be the asymptotic probability that a t-interval will be transient. Pick ft positive with ft et < .5. With probability at least .5, the rst = ft p1 t-intervals after i will be persistent. Conditioning 9

432

modified:1994-01-29

on an interval being persistent is conditioning on an event that holds with probability 1 o(1) so that each of these t-intervals will have length at least t with probability at least c o(1). As the lengths are independent tp t with conditional probability at least .99 at least c t /2 of the intervals have length at least t pt . Thus with probability at least, say .4 the total length Lt+1 is at least ct t pt /2 which is t+1 p(t+1) for an appropriate constant t+1 , completing the induction. Up to now the relation between p and n, the number of integers, has not appeared. Recall that p 0 and n so that np k but npk+1 0. Now begin at i = i0 = 1 and generate the k-interval [i0 , i1 ). Then generate the k-interval [i1 , i2 ) beginning at i1 and continue. (We do this with k xed. Even if one of the intervals is transient we simply continue with kintervals. Again we imagine continuing forever through the integers.) Let N be that maximal u for which iu 1 n, so that we have split [n] into N k-intervals plus some excess. As each sequence of k ones denitely will end a k-interval N is at least the number of disjoint subintervals of k ones. Simple expectation and variance calculations show that N > .99np k almost surely. On the other side set, with foresight, c = 4c 1 1 . If N < cnpk k k then the sum of the lengths of the rst cnp k k-intervals would be less than n. But these lengths are independent identically distributed variables and each length is at least k pk with probability at least ck so that almost surely at least ck cnpk /2 of them would have length at least k pk and thus their total length would be at least (cck k /2)n > n. That is, almost surely C1 npk < N < C2 npk where C1 , C2 are absolute constants. Let 1 , . . . , N be the k-values of the k-intervals generated by this procedure. Now we make two claims about this procedure. We rst claim that almost surely none of the i are transient. Each i has probability cp of being transient so the probability that some i , 1 i C2 npk is transient is at most (cp)C2 npk = (npk+1 ) = o(1). And almost surely N < C2 npk , proving the claim. Let A1 AR be any xed sequence of elements of Pk . The second claim is that almost surely A1 Ar appears as a subsequence of the sequence, more precisely that almost surely there exists i with 1 i N R so that i+j = Aj for 1 j R. (For technical reasons we want the subsequence not to start with 1 .) As each i has a positive probability of being any particular x Pk and the i are independent and C1 npk almost surely 10

432

revision:1994-01-28

revision:1994-01-28

this xed sequence will appear in the rst C 1 npk s. And almost surely N > C1 npk , proving the claim. We have a third claim that is somewhat technical. For any 1 j k let 1 , . . . , u denote the j-values of the successive j-intervals starting at one, where u is the last such interval that is in P j . We know that almost surely 1 u is persistent in Pj . We claim further that almost surely 2 3 u is persistent in Pj . It suces to show this for any particular j as there are only a nite number of them. For any integer A we have u 1 A almost surely and the probability that 2 A+1 is transient goes to zero with A so almost surely 2 u is persistent. Let us call [b, c) a super k-interval (for a given U ) if it is a k-interval and further for every 1 j k letting 1 , . . . , u be the successive j-values of the j-intervals beginning at b and stopping with the last persistent value - that then 2 3 u is persistent in Pj . So almost surely the k-interval [1, i 1 ) is a super k-interval. We shall show, for an appropriate sequence A 1 , . . . , AR , that all U satisfying the above three claims give models < n, C, U > which have the same Ehrenfeucht value. We rst need some glue. Call [a, b) an incomplete k-interval (with respect to some xed arbitrary U ) if the k-interval beginning at a is not completed by b 1. Suppose [a, b) is an incomplete k-interval and [b, c) is a persistent super k-interval. We claim [a, c) is a persistent k-interval. The argument is by induction on k. For k = 1, [a, b) must consist of just zeroes while [b, c) consists of at least s zeroes followed by a one. But then so does [a, c). Assume the result for k and let [a, b) be an incomplete k + 1-interval and [b, c) be a persistent k + 1-interval. We split [a, b) into a (possibly empty) sequence x1 , x2 , . . . , xr of persistent k-intervals followed by (possibly null) incomplete k-interval [a+ , b) with value, say, y. We split [b, c) (renumbering for convenience) into a sequence xr+1 , . . . , xs , ys+1 of k-intervals, all persistent except the last which is transient. Then, by induction, y+x r+1 is a persistent k-interval with some value xr+1 . Then [a, c) splits into k-intervals with values x1 , . . . , xr , xr+1 , xr+2 , . . . , xs , ys+1 . By the super-persistency xr+2 xs is persistent in Pk and hence (see 3.2) so is x1 xr xr+1 xr+2 xs and therefore [a, c) is a persistent k + 1- interval. Now let < [n], C, U > be any model that meets the three claims above, all of which hold almost surely for p in this range. We set i = i 0 = 1 and nd successive k-intervals [i0 , i1 ), [i1 , i2 ), . . . until [iu1 , iu ) and then U on [iu , n] gives an incomplete k-interval. By the third claim [1, i 1 ) is superpersistent and so the interval [iu , n][1, i1 ) (going around the corner) is k-persistent. Hence we have split [n] (now thinking of it as a cycle with 1 following n ) 11

432

modified:1994-01-29

into k-persistent intervals with k-values x 1 , x2 , . . . , xu . The k-value for x1 may be dierent from that for [1, i1 ) but the others have remained the same. This sequence contains the sequence A 1 AR described in 3.5 . But this implies (see 3.6) that the Ehrenfeuct value is determined, completing the proof.

3
3.1

Background.
The Ehrenfeucht Game.

Let A be a xed nite alphabet (in application A is P k or {0, 1}) and t a xed positive integer. We consider the space A of nite sequences a 1 au of elements of A. We can associate with each sequence a model < [u], , f > where f : [u] A is given by f (i) = ai . For completeness we describe the t-round Ehrenfeucht Games on sequences a 1 au and a1 au . There are two players, Spoiler and Duplicator. On each round the Spoiler rst selects one term from either sequencs and then the Duplicator chooses a term from the other sequence. Let i1 , . . . it be the indices of the terms chosen from the rst sequence, iq in the q-th round and let i1 , . . . it denote the corresponding indices in the second sequence. For Duplicator to win he must rst assure that aiq = aiq for each q, i.e. that he selects each round the same letter as Spoiler did. Second he must assure that for all a, b
modified:1994-01-29

ia < ib ia < ib and ia = ib ia = ib (It is a foolish strategy for Spoiler to pick an already selected term since Duplicator will simply pick its already selected counterpart but this possiblity comes in in the Recursion discussed later.) This is a perfect information game so some player will win. Two sequences are called equivalent if Duplicator wins. Ehrenfeucht showed that this is an equivalence class and that two sequences are equivalent if their models have the same truth value on all sentences of quantier depth at most t. We let M denote the set of equivalence classes which is known to be a nite set. A forms a semigroup under concatenation, denoted +, and this operation lters to an operation, also denoted +, on M . We use x, y, . . . to denote elements of M : x + y their sum; O is the equivalence class of the null sequence which acts as identity. We associate a A with the sequence a of length one and its equivalence class (which contains only it),, also called a. We let jx denote x + . . . + x with j summands. From analysis of the Ehrenfeucht game (see 3.4) it is 12

432

revision:1994-01-28

known that there exists s (for deniteness we may take s = 3 t ) so that: jx = kx for all j, k s, x M Example. With A = {0, 1} we naturally associate sequences such as 101 with < {1, 2, 3}, , f > with f (1) = 1, f (2) = 0, f (3) = 1. The addition of 101 and 1101 is their concatenation (in that order) 1011101. The rst order language has as atomic formulas x y, x = y and f (x) = a for each a A. The sentence x y z f (x) = 1 f (y) = 0 f (z) = 1 x < y y < z is satised by 01110001 but not by 000111000 so these are in dierent equivalence classes with t = 3. We could also write that 101 appears as consecutive terms with x y z f (x) = 1f (y) = 0f (z) = 1x < yy < z w [(x < ww < y)(y < ww < z)] Informally we would just say x f (x) = f (x + 1) = f (x + 2) = 1 but the quantier depth is four.

3.2
modified:1994-01-29

Persistent and Transient.

Denition and Theorem. We call x M persistent if y z x + y + z = x y z z + y + x = x p s y p + y + s = x (1) (2) (3)

These three properties are equivalent. We call x transient if it is not persistent. Proof of Equivalence. (3) (1) : Take z = s, regardless of y. Then x + y + z = (p + y + s) + y + s = p + (y + s + y) + s = x (1) (3) : Let Rx = {x + v : v M }. We rst claim there exists u M with |Rx + u| = 1, i.e., all x + y + u the same. Otherwise take u M with |Rx + u| minimal and say v, w Rx + u. As Rx + u Rx we write

revision:1994-01-28

13

432

v = x + u1 , w = x + u2 . From(1), with y = u1 , we have x = v + u3 and thus w = v + u4 with u4 = u3 + u2 . Then w + su4 = v + (s + 1)u4 = v + su4 Adding su4 to R+u sends v, w to the same element so |R+u+su 4 | < |R+u|, contradicting the minimality. Now say R x + u = {u5 }. Again by (1) there exists u6 with u5 + u6 = x. Then Rx + (u + u6 ) = {x} so that (3) holds with p = x, s = u + u6 . By reversing addition (noting that (3) is selfdual while the dual of (1) is (2)) these arguments give that (3) and (2) are equivalent, completing the proof. Let x be persistent and consider v = x+y. Let z be such that x+w+z = x for all w. Then for all w v+w+(z+y) = (x+(y+w)+z)+y = x+y = v and hence v is persistent. Dually,if x is persistent y + x is persistent. Together If x is persistent then w1 + x + w2 is persistent for any w1 , w2 M . From (1) the relation x R u dened by v (x + v = u) is an equivalence relation on the set of persistent x M . We let R x denote the R -class containing x so that Rx = {x + v : v M }
modified:1994-01-29

From(2) the relation x L u dened by v (v + x = u) is also an equivalence relation on the set of persistent x M . We let L x denote the L -class containing x so that Lx = {v + x : v M }. Let x be persistent and let p, s (by (3)) be such that p + z + s = x for all x. Setting z = O, x = p + s. Thus for all z x + z + x = (p + s) + z + (p + s) = p + (s + z + p) + s = x Let Rx , Ly be equivalence classes under R , L respectively. Then x + y Rx Ly . Let z Rx Ly . Then there exist a, b with x = z + a and y = b + z so that x + y = z + (a + b) + z. But as z is persistent the above argument (with z as x and a + b as z) gives z + (a + b) + z = z. Thus

revision:1994-01-28

Rx Ly = {x + y} for all persistent x, y

Remarks. Let A = {0.1}. A sequence a1 au is transient if and only if there is a sentence Q of quantier depth at most t so that a 1 au fails Q 14

432

but there is an extension to a1 au au+1 av which satises Q such that all further extensions a1 av av+1 aw also satisfy Q. For example, with t = 4, let Q be the existence of a block 101. If a sequence does not satisfy Q then the extension given by adding 101 does satisfy Q and all further extensions will satisfy Q. Thus for a 1 au to be persistent for t = 4 it must contain 101 and indeed all blocks of length three. We think of property (3) of persistency as indicating that a persistent sequence is characterized by p, its prex, and s, its sux. There are properties such as x f (x) = 1 y y < x that depend on the left side of the sequence, in this case the value f (1). There are other properties such as x f (x) = 1 y x < y which depend on the right side of the sequence. There will be sequences with values p, s for the left and right side respectively so that the Ehrenfeucht value of the sequence is now determined, regardless of what is placed in the middle.

modified:1994-01-29

Remarks. Certain sentences Q have the property that if any a 1 au satises Q then all extensions a 1 au au+1 av satisfy Q. The sentence that the rst term of the sequence is 1 has this property; the sentence that the last term of the sequence is 1 does not have this property. Call such properties unrighteous, as they (roughly) do not depend on the right hand side of the sequence. Sequences with Ehrenfeucht value in a given R x all have the same truth value for all unrighteous properties. Sequences with Ehrenfeucht value in a given Lx would all have the same truth value for all (correspondingly dened) unleftuous properties.

3.3

The Markov Chain.

revision:1994-01-28

Now consider a probability distribution over A, selecting each a with nonzero probability pa . This naturally induces a distribution over A u , the sequences of length u, assuming each element is chosen independently. This then leads to a distribution over the equivalence classes M . For all u 0, x M let Pu (x) be the probability that a random string a 1 au is in class x. On M we dene a Markov Chain, for each x the transition probability from x to x + a being pa . In Markov Chain theory the states x M are split into persistent and transient, a state x is persistent if and only if it lies in a minimal closed set. We claim Markov Chain persistency is precisely persistency as dened by (1), (2), (3). If C is closed and x C then R x C and Rx is itself closed. If x satises (1) then Ru = Rx for all u = x + y Rx so x is Markov Chain 15

432

persistent. Conversely if x is Markov Chain persistent then R x must be minimal closed so Ru = Rx for all u = x + y Rx and so x satises (1). The Markov Chain M restricted to a minimal closed set R x is aperiodic since x + sa Rx and (x + sa) + a = x + sa. Hence from Markov Chain theory when x is persistent lim u Pu (x) exists. A random walk on M , beginning at O, will with probability one eventually reach a minimal closed set Rx and then it must stay in Rx forever. Let P [Rx ] denote the probability that Rx is the closed state reached.

3.4

Recursion.

Again let A be a nite alphabet, M the set of equivalence classes of A and now specify some B M . As B is also a nite set we can dene equivalence classes (with respect to the same constant t) on B, let M + denote the set of such classes. Now let b1 bu and b1 bu be equivalent sequences of B. We claim that b1 + . . . + b u = b 1 + . . . + b u as elements of M . Let s1 , . . . , su , s1 , . . . , su be specic elements of A in the repective bi or bi classes. It suces to give a strategy for Duplicator with models s1 + . . . + su and s1 + . . . + su . Suppose Spoiler picks an element x in, say, some si . In the game on B we know Duplicator has a winning reply to bi of some bi . Now Duplicator will pick some x in si . To decide the appropriate x in si to pick Duplicator considers a subgame on s i and si . As these are equivalent Duplicator will be able to nd such x for the at most t times that he is required to. This general recursion includes the previous statement that for all j, k s and any x M we have jx = kx. Here B = {x} and this says that Duplicator can win the t-move Ehrenfeucht game between a sequence of j xs and a sequence of k xs; that is, that < [j], > and < [k], > are equivalent a basic result on Ehrenfeucht games. In our argument we apply it inductively with A = Pk . We know, inductively, that all k-intervals having the same k-value x Pk have the same Ehrenfeucht value. Now the k + 1-interval of i is associated with a sequence x1 xu Pk and a tail yu+1 Tk . We call two such k + 1-intervals equivalent if x 1 xu and x1 xu are equivalent in Pk and yu+1 = yu +1 . Now x1 + . . . + xu = x1 + . . . + xu and so the k + 1-intervals have equal Ehrenfeucht value.

revision:1994-01-28

modified:1994-01-29

16

432

3.5

Cycles.

modified:1994-01-29

Again let M be the set of equivalence classes on A. Now consider cycles a1 au (thinking of a1 following au ) with ai A and consider equivalence classes under the (t + 1)-move Ehrenfeucht game. Here we must preserve the ternary clockwise predicate C(x, y, z). Any rst move a i reduces the cycle to a linear structure ai au a1 ai1 of the form < [u], , f > with an Ehrenfeucht value x = xi . Two cycles are equivalent if they yield the same set of values xi M . For every persistent x M let (by (3)) p = p x , q = sx be such that x = px +y+sx for all y M . Let Px and Sx be xed sequences (i.e., elements of A) for these equivalence classes and let R x be the sequence consisting of Sx in reverse order followed by Px . If the cycle a1 au contains Rx as a subsequence then selecting ai as the rst element of Px gives a linear structure beginning with Px and ending with Sx , hence of value px +y +sx = x. Let R A be a specic sequence given by the concatenation of the above Rx for all persistent x M . Then we claim R is a universal sequence in the sense that all a1 au A (for any u) that contain R as a subsequence are equivalent. For every persistent x M there is an a i so that ai ai1 has value x. Conversely every ai belongs to at most one of the Rx creating R (maybe none if ai isnt part of R) and so there will be an R x not containing that ai . Then in ai ai1 the subsequence Rx will appear as an interval. Hence the value of ai ai1 can be written w1 + x + w2 , which is persistent. That is, the values of ai ai1 are precisely the persistent x and hence the class of a1 au in the circular t + 1-Ehrenfeucht game is determined.

3.6

Recursion on Cycles.

revision:1994-01-28

Again let A be a nite alphabet, M the set of equivalence classes in A and specify some B M . Suppose a cycle a 1 au on A may be decomposed into intervals s1 , . . . , sr with Ehrenfeucht values b1 br . Then the Ehrenfeucht value of the cycle b1 br determines the Ehrenfeucht value of a 1 au . The argument is the same as for recursion on intervals. Let a 1 au and a1 au be decomposed into s1 sr and s1 sr with Ehrenfeucht values b1 br and b1 br . Spoiler picks x in some si . In the game on cycles over B Duplicator can respond bi to bi . Then Duplicator picks an x si so that he can win the subgame on si and si . We apply this is 2 with A = {0, 1} and B = P k . Here the Pk

17

432

may have more information than the Ehrenfeucht value but this only helps Duplicator.

The Linear Model.

We have already remarked in 1 that Zero-One Laws generally do not hold for the linear model < [n], , U > and that P. Dolan has characterized those p = p(n) for which they do. Our main object in this section is the following convergence result. Theorem 2. Let k be a positive integer, and S a rst order sentence. Then there is a constant c = ck,S so that for any p = p(n) satisfying n k we have
n
1

p(n)

n k+1

lim Pr[Un,p |= S] = c

Again we shall x the quantier depth t of S and consider Ehrenfeucht classes with respect to that t. For each P k let c be the constant dened in 2 as the limiting probability that a k-interval has k-value . Let M be the set of equivalence classes of P k , a Markov Chain as dened in 3, and for each R -class Rx let P [Rx ], as dened in 3, be the probability that a random sequence 1 2 eventually falls into Rx . In < [n], , U > let 1 N denote the sequence of k-values of the successive k-intervals, denoted [1, i 1 ), [i1 , i2 ), . . ., from 1. Set, with foresight, = 102 3t . We shall call U on [n] right nice if it satises two conditions. The rst is simply that all the 1 , . . . , N described above are persistent. We know from 2 that this holds almost surely. The second will be a particular universality condition. Let A1 AR be a specic sequence in P k with the property that for every Rx and Ly there exists a q so that A1 Aq Ly and Aq+1 AR Rx

modified:1994-01-29

revision:1994-01-28

(We can nd such a sequence for a particular choice of R x and Ly by taking specic sequences in Pk in those classes and concatenating them. The full sequence is achieved by concatenting these sequences for all choices of R x and Ly . Note that as some A1 Aq Ly the full sequence is persistent.) The second condition is that inside any interval [x, x + n] [1, n] there exist R consecutive k-intervals [i L , iL+1 ), . . . , [iL+R , iL+R+1 ) whose k-values 18

432

modified:1994-01-29

are, in order, precisely A1 , . . . , AR . We claim this condition holds almost surely. We can cover [1, n] with a nite number of intervals [y, y + 3 n] and it suces to show that almost always all of them contain such a sequence, so it suces to show that a xed [y, y + 3 n] has such a sequence. Generating the k-intervals from 1 almost surely a k-interval ends after y and before y + 6 n. Now we generate a random sequence 1 on an interval of length 6 n. But constants do not aect the analysis of 2 and almost surely A 1 AR appears. Now on < [n], , U > dene U r by U r (i) if and only if U (n + 1 r). U r is the sequence U in reverse order. Call U left nice if U r is right nice. Call U nice if it is right nice and left nice. As all four conditions hold almost surely, the random Un,p is almost surely nice. r r Let U be nice and let 1 N and 1 N r denote the sequences of kvalues for U and U r respectively and let Rx and Rxr denote their R -classes respectively. (Both exist since the sequences are persistent.) Claim. The values Rx and Rxr determine the Ehrenfeucht value of nice U . We rst show that Theorem 2 will follow from the Claim. Let R x , Rxr be any two R -classes. Let U be random and consider < [n], , U >. The sequence of k-values lies in Rx with probability P [Rx ] + o(1). The same holds for U r on [n]. But U r examines U on [(1 )n, n] so as < .5 the values of the R -classes are independent and so the joint probability of the values being Rx and Rxr respectively is P [Rx ]P [Rxr ] + o(1). Given the Claim < [n], , U > would then have a value v = v(R x , Rxr ). As P [Rx ]P [Rxr ] = P [Rx ] P [Rxr ] = 1 1 = 1

revision:1994-01-28

this would give a limiting distribution for the Ehrenfeucht value v on < [n], , U >. Now for the claim. Fix two models M =< [n], , U > and M =< [n ], , U >, both nice and both with the same values R x , Rxr . Consider the tmove Ehrenfeucht game. For the rst move suppose Spoiler picks m M . By symmetry suppose m n . Let [ir1 , ir ) be one of the k-intervals with, 2 say, .51n ir .52n. We allow Duplicator a free move and have him select ir . Let 1 N and 1 N be the sequences of k-values for M and M respectively. Let z be the class of 1 r . Since U is nice this sequence already contains A1 AR and hence is persistent so z Rx . Let z be the class of r+1 N . By the same argument z is persistent. In M inside of, say, [.5n, .51n] we nd the block A 1 AR . By the universality property we can split this block into a segment in L z and another in Rz . Adding more 19

432

to the left or right doesnt change the nature of this split. Thus there is an interval [ir 1 , ir ) so that 1 r Lz and r +1 N Rz . Spoiler plays ir in response to ir . The class of 1 r is z and z Rx . The class z of 1 r is in Lz and Rx . As z Lz Rx , z = z . Thus [1, ir ) under M and [1, ir ) under M have the same Ehrenfeucht value. Thus Duplicator can respond successfully to the at most t moves (including the initial move m) made in these intervals. Thus Spoiler may as well play the remaining t 1 moves on M1 =< [ir , n], , U > and M1 =< [ir , n ], , U >. These intervals have lengths n1 n and n1 n respectively. But now M and M are both nice 3 3 with respect to 1 = 3 - the sequence A1 AR still appears inside every interval of length n 1 n1 in M and 1 n1 in M . Hence we can apply the same argument for the second move - for convenience still looking at Ehrenfeucht values with respect to the t move game. After t moves we still have nice Mt , Mt with respect to t 102 so the arguments are still valid. But at the end of t rounds Duplicator has won.

revision:1994-01-28

modified:1994-01-29

20

432