(

5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law for random graphs
with order and random binary functions
Saharon Shelah

Institute of Mathematics, The Hebrew University

Department of Mathematics, Rutgers University
Department of Mathematics, University of WI, Madison
done: August 1993, October 31, 1993
last corrections introduced June 5, 1996
printed: October 6, 2003

The research partially supported by the United States-Israel Binational Science

Foundation and NSF under grant #144-EF67; Publication no 548.
0
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 1
0 Introduction
Let G
<
(n, p) denote the usual random graph G(n, p) on a totally ordered set
of n vertices. (We naturally think of the vertex set as 1, . . . , n with the usual
<). We will x p =
1
2
for deniteness. Let L
<
denote the rst order language
with predicates equality (x = y), adjacency (x y) and less than (x < y).
For any sentence A in L
<
let f(n) = f
A
(n) denote the probability that the
random G
<
(n, p) has property A. It is known Compton, Henson and Shelah
[CHSh245] that there are A for which f(n) does not converge. Here we show
what is called a very weak zero-one law (from [Sh 463]):
Theorem 0.1 For every A in language L
<
lim
n
(f
A
(n + 1) f
A
(n)) = 0
Note, as an extreme example, that this implies the nonexistence of a sentence
A holding with probability 1 o(1) when n is even and with probability o(1)
when n is odd (as in Kaufman, Shelah [KfSh201]).
In 2 we give the proof, based on a circuit complexity result. In 3 we
prove that result, which is very close to the now classic theorem that parity
cannot be given by an AC
0
circuit. In 4 we give a very weak zero-one
law for random two-place functions. The proof is very similar, the random
function theorem being perhaps of more interest to logicians, the random
graph theorem to discrete mathematicians.
The reader should thank Joel Spencer who totally rewrote the paper
(using the computer science jargon rather than the logicians one), and with
some revisions up to the restatement in the proof of 2.1 but with 3.1, this
is the version presented here. We thank the referee for comments on the
exposition, and we thank Tomasz Luczak and Joel Spencer for reminding me
this problem on G
<
(n, p) in summer 93.
On a work continuing this of Boppana and Spencer see 3.2(5).
1 The Proof.
Let G be a xed graph on the ordered set 1, . . . , 2n+1. For a property A and
for i = n, n+1 let g(i) = g
G,A
(i) denote the probability that G
S
satises A
where S is chosen uniformly from all subsets of 1, . . . , 2n+1 of size precisely
i. We shall show
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 2
Theorem 1.1 g(n + 1) g(n) = o(1)
More precisely, given A and > 0 there exists n
0
so that for any G as above
with n n
0
we have [g(n + 1) g(n)[ < .
We rst show that Thm.0.1 follows from Thm.1.1. The idea is that a
random G
<
(i, p) on i = n or n + 1 vertices is created by rst taking a
random G
<
(2n+1, p) and then restricting to a random set S of size i. Thus
(xing A) f(n), f(n + 1) are the averages of g
G
(n), g
G
(n + 1) over all G. By
Thm.1.1 we have g
G,A
(n) g
G,A
(n + 1) = o(1) for all G and therefore their
averages are only o(1) apart.
Now we show Thm.1.1. Fix G, A as above. Let P(S) be the Boolean value
of the statement that G[
S
satises A. For 1 x 2n + 1 let z
x
denote the
Boolean value of x S so that P(S) is a Boolean function of z
1
, . . . , z
2n+1
.
We claim this function has a particularly simple form. Any A can be built up
from primitives x = y, x < y, x y by , and, critically,
x
. As G is xed
the primitives have values true or false. Let , be themselves. Consider

x
W(x) where for each 1 x 2n + 1 we let W(x) on G[
S
is given by
W

(x). Then
x
W(x) has the interpretation
xS
W(x) which is expressed
as
2n+1
x=1
(z
x
W

(x)). For convenience we can be redundant and replace

x
W(x) by
2n+1
x=1
(z
x
W

(x)). For example

x

y
x y becomes

x
[z
x

yx
z
x
]
Thus P(S) can be built up from z
1
, . . . , z
2n+1
be means of the standard
, , and , over (at most) 2n + 1 inputs. That is (see 3) P(S) can be
expressed by an AC
0
circuit over z
1
, . . . , z
2n+1
(of course with the number
of levels bounded by the length d
A
of the sequence A (can get less) and the
number of nodes bounded by d
A
n
d
A
). Now g(i), for i = n, n + 1, is the
probability P holds when a randomly chosen set of precisely i of the zs are
set to True. From Thm.2.1 below g(n + 1) g(n) = o(1) giving Thm. 1.1
and hence Thm. 0.1.
2 AC
0
Functions
We consider Boolean functions of z
1
, . . . , z
m
. (In our application m = 2n+1.)
The functions z
i
, z
i
, called literals, are the level 0 functions. A level i + 1
function is the or of polynomially many level i functions. An AC
0
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 3
function is a level d function for any constant d. By standard technical
means we can express any AC
0
function in a levelled form so that the level
i + 1 functions used are either all s of level i functions or all s of level i
functions and the choice alternates with i (at most doubling the number of
levels). It is a classic result of circuit complexity that parity is not an AC
0
function. Let C be an AC
0
function. For 0 i m let f(i) = f
C
(i) denote
the probability C holds when precisely i of the z
j
are set to True and these
i are chosen randomly.
Theorem 2.1 f(n + 1) f(n) = o(1)
Called a restriction balanced if [i : (i) = 0[ = [i : (i) = 1[.
Now more fully the theorem says
() for every , d, t there is n
,d,t
satisfying: if n n
,d,t
and C is an AC
0
Boolean circuit of z
1
, . . . , z
2n+1
of level d with n
t
nodes then
[f
C
(n + 1) f
C
(n)[ < .
This statement is proved by induction on d.
We choose the following
(i) c
0
= (n4)t > 0
(ii) =
1
2
,

=
1
2
1+
(iii) k is such that k t
(iv) we choose k

inductively on k such that k

large enough.
(v) c
1
a large enough real
(vi) n
0
is large enough
For a node x of the circuit C let
x
be the set of nodes which fans into it;
(without loss of generality in the level 1 we have only OR).
First we assume d > 2. Note

1
drawing as below a balanced restriction with domain with n elements,
with probability 1 /3 we have: in C
1
= C

, every node of the

level 1 (i.e. for which
x
is a set of atoms) satises [
x
[ c
0
(lnn).
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 4
[Why? Choose randomly a set u
0
of [n/2] pairwise disjoint pairs of numbers
among 1, . . . , 2n + 1, and then for each i, j u decide with probability
half that (i) = 0, (j) = 1 and with probability half that (i) = 1, (j) = 0
(independently for disjoint pairs). This certainly gives a balanced .
Now if x is a node of C of the level 1, the probability that does not
decide the truth value which the node compute is (
1
4
)
|Yx|
. Note: after
drawing u, if
x
contains a pair from u the probability is zero, we only
increase compared to drawing just a restriction. So the probability that for
some x of the level 1 of C, [
x
[ (ln4)t(lnn) + 1 and the truth value is not
computed, is [C[ (
1
4
)
(n4)t(nn)+1
1/2, so there is
0
for which for any
such x the truth value is computed.]
Next, we say that a restriction

extends a restriction if

(i) ,= (i) (i) = . Now

2
Choosing randomly a restriction
1
as below we have:
1
is a balanced
restriction extending such that [i : i 1, . . . , 2n + 1, (i) = [
2[n

] + 1 and with probability 1 /3 for every node y of C of the

level 1 we have, [
y
[ k.
[Why? We draw a set u
1
of (2n+1[dom(
0
)[(2[n

0
]+1))/2 pairs from i :

0
(i) = pairwise disjoint and for each i, j u, decide with probability
1
2
that
1
(i) = 0,
1
(j) = 1 and with probability half that
1
(i) = 1,
1
(j) = 0.
For each node y C
1
of the level 1 the probability that the number of
y

y
not assigned a truth value by
1
is k + 1 is at most

[
y
[
k + 1

1
2n

0 +1

k+1
(c
0
lnn)
k+1
n

0
(k+1)
< n
t
.]
We now choose by induction on k a restriction
2,
such that

3
(a)
2,
0
=
1
,
2,

2,+1
, 2n + 1 (2[n

] + 1) = [dom
2,
[
(b) every y C of the level 2 there is a set w
y,
of k

atoms such
that: if z
y
, then [
z
w
y,
[ k .
Now for C
2,k
we can invert AND and OR (multiplying the size by a
constant c
1
) decreasing d by one thus carrying the induction step.
For = 0 let
2,0
=
1
. For + 1, for each y C of level 2 let = :
a restriction with domain w
y,
let

y
= z
y
: the truth value at z is still not computed under
2,
,
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 5
and try to choose by induction on i an atom z
y,,,i

y
z
y,,,j
: j < i,
such that dom(z
y,,,i
) is disjoint to

j<i
dom(z
y,,,j
) w
y,
. Let it be dened if
i < i
y,
.
Now
2,+1
will for each decide that make the truth value
computed in y true, or will leave only (k
+1
k

)/2
k

of the atoms in

i
domz
y,,,i
w
y,
undetermined (this is done as in the previous two stages).
But now by
1
+
2
, C

2,k
can be considered having d 1 levels
(because, as said above we can invert the AND and OR in level 1 and 2).
We have translate our problem to one with [n

k
], d 1,
k
(t +
1
),

3
instead n, d, t, (the t + is just for n
t+
> c
1
n
t
).
Also note: , c
1
does not depend on n. So we can use the induction
hypothesis. We still have to check the case d 2, we still are assuming level
1 consist of cases of OR, and for almost all random
1
(as in
1
) for every x
of level 1 we have [
x
[ c
0
lnn (so again changing n).
So as above we can add this assumption. Choose randomly a complete
restriction
0
with [i :
0
(i) = 1[ = n, and let
1
be gotten from
0
by
changing one zero to 1, so [i :
1
(i) = 1[ = n + 1.
Now the probability that C

0= 0 but C

1= 1 is small: it require that

for some node x of level 1 is made false in C

0 while there is no such x

for C

1, but if x() is such for C

0 it is made true then with probability

1
|Yx|
2n+1
1
c
0
lnn
n
the z
i
changed is not in
x()
. Contradiction, thus
nishing the proof.
3 Two Place Functions
Here we consider the random structure ([n], F
n
) where F
n
(x, y) is a random
function from [n] [n] to [n]. (We no longer have order. A typical sentence
would be
x

y
F(x, y) = x.): Again for any sentence A we dene f(n) = f
A
(n)
to be the probability A holds in the space of structures on [n] with uniform
distribution. Again it is known [CHSh245] that convergence fails, there are
A for which f(n) does not converge. Again our result is a very weak zero
one law.
Theorem 3.1 For every A
lim
n
f
A
(n + 1) f
A
(n) = 0
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 6
Again let m = 2n+1. Let F

(x, y, z) be a three-place function from [m][m]

[m] to [m]. For S [m] of cardinality i = n or n + 1 we dene F

S
, a partial
function from [S] [S] to [S] by setting F

S
(x, y) = F

(x, y, z) where z is the

minimal value for which F

(x, y, z) S. If there is no such z then F

S
(x, y)
is not dened. This occurs with probability (
mi
m
)
m
for any particular x, y so
the probability F

S
is not always dened is at most i
2
(
mi
m
)
m
= o(1).
We generate a random three-place F

and then consider F

S
with S a
random set of size i = n or n+1. Conditioning on F

S
being always dened it
then has the distribution of a random two-place function on i points. Thus
Pr[A] over [n], F
n
is within o(1) of Pr[A] when F
n
= F

S
is chosen in this
manner. Thus, as in 2, it suces to show for any F

and A that, letting

g(i) denote the probability F

S
satises A with S a uniformly chosen i-set,
g(n + 1) g(n) = o(1). Again x F

and A and let z

x
be the Boolean value
of x S for 1 x 2n+1. In A replace the ternary relation F(a, b) = c by

y<d
z
F(a,b,y)
& z
F(a,b,d)
. (For d = 1 this is simply True.) As in 2 replace

x
P(x) by
x
(z
x
P

(x)) where P

(x) has been inductively dened as the

replacement of P(x). Then the statement that F

S
satises A becomes a
Boolean function of the z
1
, . . . , z
m
, as before it is an AC
0
function, and by
2 we have g(n + 1) g(n) = o(1).

The following discussion is directed mainly for logicians but may be of
interest for CS-oriented readers as well.
Discussion 3.2 (1) Note that the results of [Sh463] cannot be gotten in
this way as the proof here use high symmetry. The problem there was:
let p = p
i
: i N) be a sequence of probabilities such that

i
p
i
< .
Let G(n, p) be the random graph with set of nodes [n] = 1, . . . , n and
the edges drawn independently, and for i ,= j the probability of i, j
being an edge is p
|ij|
.
The very weak 0-1 law was proved for this context in [Sh463] (earlier
on this context (probability depending on distance) was introduced and
investigated in Luczak and Shelah [LuSh435]). Now drawing G(2n +
1, p) and then restricting ourselves to a random S 1, . . . , 2n + 1
with n, and with n + 1 elements, fail as G(2n + 1, p)
S
does not have
the same distribution as G([S[, p).
(2) We may want to phrase the result generally;
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 7
One way: just say that M
n
, M
n+1
can be gotten as above : draw
a model on [2n + 1] = 1, . . . , 2n + 1 (i.e. with this universe),
then choose randomly subsets P

n
with n+ elements and restrict
yourself to it.
(3) Two random linear order satises the very weak 0-1 zero law (mean:
take two random functions from [n] to [0, 1]
R
). The proof should be
clear.
(4) All this is for xed probabilities; we then can allow probabilities de-
pending on n e.g. we may consider G
<
(n, p
n
) is the model with set of
elements 1, . . . , n, the order relation and we draw edges with edge
probability p
n
depending on n. This call for estimating two number
(for rst order sentence):

n
= [Prob(G
<
(n, p
n+1
) [= ) Prob(G
<
(n, p
n
) [= )[

n
= [Prob(G
<
(n, p
n+1
) [= ) Prob(G
<
(n + 1, p
n+1
[= )[
As for
n
the question is how much does the proof here depend on hav-
ing the probability
1
2
. Direct inspection on the proof show it does not
at all (this just inuence on determining the specic Boolean function
with 2n+1 variables) so we know that
n
converge to zero.
As for
n
, clearly the question is how fast p
n
change.
(5) As said in [Sh463] we can also consider lim(Prob
n+h(n)
(M
n+h(n)
[= )
Prob
n
(M
n
[= )) = 0, i.e. characterize the function h for which this
holds but this was not dealt with there. Hopefully there is a threshold
phenomena. Probably this family of problems will appeal to mathe-
maticians with an analytic background.
Another problem, closer to my heart, is to understand the model the-
ory: in some sense rst order formulas cannot express too much, but
can we nd a more direct statement fullling this?

Another way to present the rst problem for our case is: close (or
at least narrow) the analytic gap between [CHSh245] and the present
paper.
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 8
After this work, Boppana and Spencer [BS], continuing the present
paper and [CHSh245], address the problem and completely solve it.
More specically they proved the following.
For every sentence A there exists a number t so that
m(n) = O(nln
t
n) implies
lim
n
f
A
(n + m(n)) f
A
(n) = 0.
And
For every number t there exists a sentence A and a func-
tion m(n) = O(nln
t
n) so that f
A
(n + m(n)) f
A
(n) does
not approach zero.
Together we could say: a function m(n) has the property that for all A
and all m

(n) m(n) we have f

A
(n + m

(n)) f
A
(n) 0 if and only
if m(n) = o(nln
t
n) for all t.
For improving the bound from this side they have used Hastad switch-
ing lemma [Hastad] (see [AS], 11.2, Lemma 2.1).
(6) If we use logic stronger than rst order , it cannot be too strong (on
monadic logic see [KfSh201]), but we may allow quantication over
subsets of size k
n
, e.g. log(n) there are two issues:
(A) when for both n and n + 1 we quantify over subsets of size k
n
,
we should just increase M by having the set [n]
kn
as a set of
extra elements, so in (*), P is chosen as a random subset of
1, 2, 3, . . . , 2n, 2n + 1 with n or n + 1 elements but the model
has about (2n+1)
kn
elements; this require stronger theorem, still
true (up to very near to exponentiation)
(B) if k
n
,= k
n+1
we need to show it does not matter, we may choose
to round k
n
= log
2
(n) so only for rare n the value change so we
weaken a little the theorem or we may look at sentences for which
this does not matter .
Maybe more naturally, together with choosing randomly /
n
we
choose a number k
n
, and the probability of k
n
= k
n
+ i if i
[k
n
/2, k
n
/2] being 1/k
n
.
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 9
And we ask for p

n
=: Prob(/
n
[= where the monadic quan-
tier is interpreted as varying on set with k
n
elements) for
sentence (the point of the distribution of k
n
is just that for n,
n + 1 they dier a little). E.g. if for a random graph on n (prob-
ability 0.5) we ask on the property the size of maximal clique of
size at most [log
2
n]
2
is even it satises the very weak zero one
law
Of course we know much more on this, still it shows that this old result
(more exactly - a weakened version) can be put in our framework.
References
[A] M. Ajtai, A
1
1
formula on nite structures, Annals of Pure and
Applied Logic 24(1983):149.
[AS] N. Alon and J. Spencer, The Probabilistic Method, John Wiley
(New York), 1992.
[BS] R.B. Boppana and J.H. Spencer (1995), Smoothness laws for ran-
dom ordered graphs, to appear in Logic and Random Structures:
DIMACS Workshop, November 57, 1995, DIMACS series in dis-
crete mathematics and theoretical computer science, American
Mathematical Society, Providence, Rhode Island.
[Hastad] J. Hastad, Almost optimal lower bounds for small depth circuits,
in S. Micali, ed. Advances in Computer Research, Vol. 5, JAI
Press, Greenwich CT, 143-170
[KfSh201] M. Kaufman and S. Shelah, On random models of nite power
and monadic logic, J. Symb. Logic 54(1985):285293.
[CHSh245] K. Compton, C.S. Henson and S. Shelah, Non-convergence, Un-
decidability and Intractability in asymptotic problems, Annals of
Pure and Applied Logic 36(1987):207224.
[LuSh435] T. Luczak and S. Shelah, Convergence in homogeneous random
graphs, Random Structures and Algorithm, 6 (1965):371391.
(
5
4
8
)


r
e
v
i
s
i
o
n
:
1
9
9
6
-
0
6
-
1
1







m
o
d
i
f
i
e
d
:
1
9
9
6
-
0
6
-
1
5


Very weak zero one law. . . , Sh 548 October 6, 2003 10
[Sh463] S. Shelah, On the very weak 0 1 law for random graphs with
order, J. of Logic and Computation, 6 (1966):137159.