8

6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR
m 3
SAHARON SHELAH AND MOR DORON
Abstract. We consider the notion of bounded m-ary patch-width dened in
[4], and its very close relative m-constructibility dened below. We show that
the notions of m-constructibility all coincide for m 3, while 1-constructibility
is a weaker notion. The same holds for bounded m-ary patch-width. The case
m = 2 is left open.
1. Introduction
1.1. Background. Our interest in this subject started from investigating spectra
of monadic sentences, so let us begin with a short description of spectra. Let be a
sentence in (a fragment of) second order logic (SOL). The spectrum of is the set
n N : has a model of size n. In 1952 Scholz dened the notion of spectrum
and asked for a characterization of all spectra of rst order (FO) sentences. In [1]
Asser asked if the complement of a FO spectrum is itself a FO spectrum.
Denition 1.1. A set A N is eventually periodic if for some n, p N, for all
m > n, m A i m+p A.
In [3] Durand, Fagin and Loescher showed that the spectrum of a FO sentence
in a vocabulary with nitely many unary relation symbols and one function symbol
is eventually periodic. In [5] Gurevich and Shelah generalized this for spectrum
of monadic second order (MSO) sentence in the same vocabulary. Inspired by
[5] Fisher and Makowsky in [4] showed that the spectrum of a CMSO sentence (a
monadic sentence with counting quantiers) is eventually periodic provided that all
its models have bounded patch-width. The notion of patch-width of structures (usu-
ally graphs) is a complexity measure on structures, generalizing clique-width. Their
proof remains valid if we consider m-ary patch-width, i.e. we allow m-ary relations
as auxiliary relations. In [6] Shelah generalized the proof of [5] and showed eventual
periodicity for a MSO sentence provided that all its models are constructible by
recursion using operations that preserve monadic theory (see denitions below).
1.2. summation of results. The above results on eventual periodicity led us to
ask: What are the relations between the dierent notions for which we have eventual
periodicity of MSO spectra? In other words do we have three dierent results, or
are they all equivalent? We give an answer here. In [2] Courcelle proved (using
somewhat dierent notations) that a class of structures is constructible i it is
monadicly interpretable in trees, thus implying that two of the results coincide. We
give a proof of Courcelles result more coherent with our denition, which we use
The Author would like to thank the Israel Science Foundation for partial support of this
research (Grant no. 242/03).
Publication no. 865 in Saharon Shelahs list.
1
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


2 SAHARON SHELAH AND MOR DORON
later on. We prove that the notions of bounded m-ary patch-width is very close
to m-constructibility (constructibility where we allow m-ary relations as auxiliary
relations) (see lemmas 2.9 and 2.10). Next we show that for m 3 a class of
modes is contained in a m-constructible class i it is contained in a 3-constructible
class (see Theorem 3.7). The same holds for classes of bounded m-ary patch-width.
Finally we show that in the above theorem we can not replace 3-constructible by
1-constructible. That is there exists a 3-constructible class which is not contained
in any 1-constructible class. We give a specic example (see 4.1). The case m = 2
is left open.
2. Preliminary definitions and previous results
Notation 2.1.
(1) Let be a nite relational vocabulary.
(2) For R let n(R) be the number of places of R. We say that R is n(R)-ary
or n(R) place. We allow n(R) = 0 i.e. the interpretation of R is in T, F.
We call nice if R n(R) > 0.
(3) For k , let
k
be P
1
, ..., P
k
with P
1
, ..., P
k
unary predicates.
(4) A k-colored -structure is a
k
-structure in which the interpretation of the
P
i
s is a partition of the set of elements of the model (but some P
i
s may
be empty).
(5) A k-const -structure is a
k
-structure in which every predicate P
i
is inter-
preted by a singleton. We denote such a structure by (M, a
1
, ..., a
k
) where
M is a -structure and a
1
, ..., a
k
M.
Denition 2.2.
(1) A monadic second order (MSO) formula in vocabulary is a second order
formula in which every second order quantier quanties an unary relation
symbol. The notion of quantier depth extends, naturally to MSO formulas.
(2) Let M be a -structure, and q a natural number. The monadic q-theory of
M, Th
MSO
q
(M), is the set of all sentences of quantier depth q that hold
in M.
(3) Let M be a -structure, and n, q natural numbers. Let a = (a
1
, ..., a
n
)
n
[M[. The q-type of a in M, tp
q
( a, M), is the set of all formulas
, of quantier depth q in free variables x
1
, ..., x
n
, such that: M [=
[a
1
, ..., a
n
]. If q = 0 we sometimes write tp
qf
( a.M).
(4) The notion of a q-type extends to MSO logic. We write tp
MSO
q
( a, M) for the
set of MSO formulas , of quantier depth q in free variables x
1
, ..., x
n
,
such that: M [= [a
1
, ..., a
n
].
(5) The set of all formally possible q-types in a vocabulary and in variables
x
1
, ..., x
n
), will be denoted be TP
q
(x
1
, ..., x
n
), ), and similarly TP
MSO
q
(x
1
, ..., x
n
), ).
We may write TP
MSO
q
(n, ) instead of TP
MSO
q
(x
1
, ..., x
n
), ).
Denition 2.3 (Patch-width).
(1) Let be a nice vocabulary, M a -structure, k a natural number, and P
a nite set of k-colored -structures. We say that M have patch-width at
most k (with respect to P) and denote pwd
P
(M) k, if M is the -redact
of a k-colored -structure which is in the closer of P under the operations:
(i) disjoint union - .,
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 3
(ii) recoloring -
ij
(change all the elements with color P
i
to color P
j
)
and
(iii) modications -
R,B
(redene the relation R by the quantier free
formula B in vocabulary
k
).
A class K of -structures is a PW(k)-class, if for some nite set of k-
colored -structures P the elements of K are all the -redacts of structures
of patch-width at most k with respect to P. We say K is of bounded patch-
width (BPW) if it is a PW(k)-class for some k N.
(2) In the denition above we may instead of k-colored -structures, talk about

+
structures where
+
, [
+
[ = k and every relation in
+
is
at most m-ary. We then talk about m-ary patch-width, where the rest of
the denition remains unchanged. Note that the notions of patch-width and
unary patch- width are close but not identical as in the former we demand
that the colors are disjoint.
In [4] it is proved that:
Theorem 2.4. Let be a MSO() sentence, and suppose Mod() is contained in
some class of bounded m-ary patch-width. Then spec() is eventually periodic.
Denition 2.5 (Addition operations).
(1) For k, k
1
, k
2
N, let S
,k,k1,k2
be the set of all addition operations of a
k
1
-const -structure with a k
2
-const -structure, resulting in a k-const -
structure. Formally each s S
,k,k1,k2
consists of:
(i) Sets A
l
= A
s
l
1, ..., k
l
for l 1, 2.
(ii) For l 1, 2, a 1-1 function g
l
= g
s
l
from A
l
to 1, ..., k such that:
Im(g
1
) Im(g
2
) = 1, ..., k.
(iii) For l 1, 2 a set B
l
1, ..., k
l

2
, and a set B 1, ..., k
1

1, ..., k
2
.
(iv) For each R with n(R) = n and each w
l
1, ..., n for l 1, 2,
a function f
R,w1,w2
= f
s
R,w1,w2
with range T, F, and domain: triplets
of the form (p, q
1
, q
2
) where:
p TP
0
(x
1
, ..., x
n
), ) were is a vocabulary with k
1
+ k
2
in-
dividual constants and two unary predicates,
For l 1, 2, q
l
TP
0
(x
i
: i w
l
), ).
(2) Let k, k
1
, k
2
N and s S
,k,k1,k2
. Let (M
l
, a
l
1
, ..., a
l
k
l
) be k
l
-const -
structure for l 1, 2. The addition (M
1
, a
1
1
, ..., a
1
k1
)
s
(M
2
, a
2
1
, ..., a
2
k2
) is
dened whenever:
([M
1
[ [M
2
[) (a
1
1
, ..., a
1
k1
a
2
1
, ..., a
2
k2
) and
For l 1, 2: a
l
i
= a
l
j
(i, j) B
l
and a
1
i
= a
2
j
(i, j) B,
to be the k-const -structure (M, b
1
, ..., b
k
) dened by:
(i) [M[ = ([M
1
[ a
1
1
, ..., a
1
k1
) ([M
2
[ a
2
1
, ..., a
2
k2
) a
l
i
: l 1, 2, i
A
l
).
(ii) For each l 1, 2 and i A
l
, a
l
i
= b
g
l
(i)
.
(iii) For all R with n(R) = n and x = (x
1
, ..., x
n
)
n
[M[, let w
l
= i :
x
i
[M
l
[ for l 1, 2. Let p be the quantier free type of x in the
model with a
1
1
, ..., a
1
k1
a
2
1
, ..., a
2
k2
as constants, and [M
1
[, [M
2
[ as
unary predicates. For l 1, 2 let q
l
= tp
qf
(x
i
: i w
l
), M
l
). Now
the value of R
M
( x) is dened to be f
s
R,w1,w2
(p, q
1
, q
2
).
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


4 SAHARON SHELAH AND MOR DORON
(3) For technical reasons we would like to allow empty structures. i.e. let

t
:= R : n(R) = 0, and X
t
. Now Null
X
is the
t
-structure
with [Null
X
[ = and R
NullX
= True R X. Then if s S
,k,k1,0
,
and M is a
k1
- structure then M
s
Null
X
is a well dened
k
-structure.
Furthermore for any -structure M, M .Null

is dened and equal to M.

The important attributes of the addition operations are the following:
Theorem 2.6. Let k, k
1
, k
2
N. Then:
(1) S
,k,k1,k2
is nite.
(2) The addition theorem:
Let M, M
t
be k
1
-const -structures such that Th
q
MSO
(M) = Th
q
MSO
(M
t
),
and N, N
t
be k
2
-const -structures such that Th
q
MSO
(N) = Th
q
MSO
(N
t
),
and s S
,k,k1,k2
. Assume that the additions M
s
N and M
t

s
N
t
are
dened. Then
Th
q
MSO
(M
s
N) = Th
q
MSO
(M
t

s
N
t
).
Denition 2.7 (Constructibility). A class K of -structures is (m

, k

)-constructible,
if there exists: A nite relational vocabulary
+
, a nite set of structures P,
and a nite set of addition operations S such that:
(i) Every relation in
+
is at most m

-ary.
(ii) Every structure in P is a k-const
+
-structure for some k k

.
(iii) Every operation in S is in S

+
,k,k1,k2
for some k, k
1
, k
2
k

.
(iv) The elements of K are all the -redacts of structures in the closer of P
under the operations in S.
We say that K is m

-constructible if it is (m

, k

)- constructible for some k

, and
that it is constructible if it is m

-constructible for some m

.
In [6] it is proved that:
Theorem 2.8. Let be a MSO() sentence, and suppose Mod() is contained in
some m-constructible class. Then spec() is eventually periodic.
This is a generalization of 2.4 as we have:
Lemma 2.9. Let be a nice vocabulary, and K be a m-ary PW(k)-class of -
structures. Then K is a (m, 0)-constructible class.
Proof. First note that the disjoint union operation of
+
-structures is in S

+
,0,0,0
.
As for the recoloring and the modication operations, those are unary operations,
so we look at the operation s S

+
,0,0,0
that acts as recoloring or modication on
its left operand. So M
s
Null

is the desired recoloring or modication of M.

In the addition operations we allow omitting marked elements, and the universe
of the the two operands is not necessarily disjoint. This is not allowed in the
operations of patch-width. It turns out though that these are the only essential
dierences between the two types of operations as suggested by the following:
Lemma 2.10. Let K be a (m, 0)-constructible class such that the vocabulary
+
associated with K is nice. Then K is of bounded m-ary patch-width.
Proof. K is (m, 0)-constructible so we have a vocabulary
+
and sets S and P.
Now the set of atomic structures for the patch-width denition will be the same P.
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 5
The vocabulary of the patch-width denition will be:
+
R
t
: R
+
P
1
, P
2
.
P
1
, P
2
are new unary relation symbols. We now have to show for each operation
in S how to simulate it by operations of patch-width. Let s S and let M
1
, M
2
be
+
-structures. Denote by M
t
1
, M
t
2
the trivial extensions to the new vocabulary.
We will now describe a series of patch-width operations on M
t
1
, M
t
2
resulting in a
structure M

such that M

+

= M
1

s
M
2
, this will complete the proof. First
color all the elements of M
t
l
by P
l
for l 1, 2. Next for each R
+
redene R
t
to be the same as R, do this for both M
t
1
, M
t
2
. Now take the disjoint union of the
to resulting structures. Finally we have to redene the relations of
+
of the our
disjoint union to be as in M
1

s
M
2
. Let R
+
be n-ary and let w
1
, w
2
1, ..., n
satisfy w
1
w
2
= 1, ..., n. Let p be the quantier free type in the vocabulary with
two unary relations S
1
, S
2
saying that for i n and l 1, 2, x
i
S
l
i i w
l
.
Now dene:

R,w1,w2
(x
1
, ...x
n
) :=

iw1
P
1
(x
i
)

iw2
P
2
(x
i
) [

q
l
TP0(xi:iw
l
),
+
)
f
s
R,w
1
,w
2
(p,q1,q2)=T
q
t
1
q
t
2
].
Where q
t
l
is the disjunction of all the formulas in q
l
where we replace every relation
R
+
be R
t
. Now redene the relation R using the modication
R,B
for the
formula:
B(x
1
, ..., x
n
) :=

w1,w21,...,n]
w1 w2=1,...,n]

R,w1,w2
(x
1
, ..., x
n
).
Do this for all R
+
and we are done.
Notation 2.11 (Trees).
(1) The vocabulary of trees,
trees
, is , c
rt
.
(2) The vocabulary of k-trees,
ktrees
, is , c
rt
P
1
, ..., P
k
i.e (
trees
)
k
.
(3) A tree T is a
trees
-structure in which:
For every t [T[ the set s [T[ : s
T
t is linearly ordered by
T
.
For all x [T[, c
rt
T

T
x.
(4) A k-tree T is a
ktrees
-structure, such that T[
trees
is a tree.
(5) A 2-tree T is directed binary (DB) if (c
T
rt
, P
T
1
, P
T
2
) is a partition of [T[,
and each non maximal element of T has exactly two immediate successors
one in P
T
1
and the other in P
T
2
. For k 2, a k-tree T is DB if ([T[;
T
, c
T
rt
, P
T
1
, P
T
2
) is.
Denition 2.12 (Monadic interpretation).
(1) We call c a monadic k-interpretation scheme for a vocabulary if c consists
of:
Natural numbers k
1
= k
c
1
and k
1
= k
c
2
both less then or equal to k.
For every l k
1
a monadic
k2trees
-formula
c
=,l
(x).
For every R n-place relation, and every
1,...,n]
0, ..., k
1
a
monadic
k2trees
-formula: =
c
R,
(x
1
, ..., x
n
).
(2) Let c be a monadic k-interpretation scheme for a vocabulary , and T a
k
c
2
-tree. The interpretation of T by c denoted by T
[c]
is the -model M
dened by:
[M[ = (t, l) [T[ 0, ..., k
1
: T [=
=,l
(t)
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


6 SAHARON SHELAH AND MOR DORON
For every R n-place relation:
R
M
= ((t
i
, l
i
) : i n)
n
[M[ : T [=
R,(li:in)
(t
1
, ..., t
n
)
(3) A -model M is monadicly k-interpretable in trees if for some c a monadic
k-interpretation scheme for , and some k
c
2
-tree, T, we have: T
[c]
= M.
We denote the class of all the -structures monadicly k-interpretable in
trees by K
mo
,k
.
(4) For c a monadic k-interpretation scheme for we denote by K
mo
c
the class
of all -structures M such that for some k
c
2
-tree, T, we have: T
[c]
= M.
K
mo,db
c
is the same as K
mo
c
only we demand that T is directed binary.
(5) We say that c has the leaf property if k
c
1
= 0 and for every k
c
2
-tree T, and
every t [T[: T [=
c
0,=
[t] implies that t is a maximal element in T.
Without loss of generality we may assume that k
c
1
= 0. This is because of the
following:
Lemma 2.13. For every c a monadic k-interpretation scheme for a vocabulary ,
there exists c
t
be a monadic (k + 2)-interpretation scheme for , such that:
k
c

1
= 0.
k
c

2
= k
c
2
+ 2.
For every k
c
2
-tree T, there exists a k
c

2
-tree T
t
, such that: T
[c]
= T
t[c

]
.
Hence K
mo
c
K
mo
c
.
Proof. Let s
1
and s
2
be the two new unary predicates, and let T be a k
c
2
-tree.
Dene T
t
as follows: [T
t
[ = [T[ ([T[ 0, ..., k
c
1
), s
T

1
= [T[, s
T

2
= [T[ 0, ..., k
c
1
,
and if t
1
is the immediate successor of t
2
in T then dene, t
1
<
T

(t
1
, 0) <
T

(t
1
, 1) <
T

...(t
1
, k
c
1
) <
T

t
2
. Now dene:

=,0
(x) := s
2
(x)

l<k
c
1
(y)[s
1
(y) (
l
(x, y)] (
c
=,l
(y))
s1
Where
l
(x, y) is a formula stating that there are exactly l elements between x
and y and all of them are in s
2
, and (
c
=,l
(y))
s1
is the formula
c
=,l
(y) relativized
to s
1
i.e we replace every quantier of the form x or x by x s
1
or x s
1
respectively. It should be clear that T [=
c
=,l
[t] i T
t
[=
c

=,0
[(t, l)]. The relations
are dealt with in a similar way.
Lemma 2.14. Let K be a (m

, k

)-constructible class of -models. Then there

exists a natural number k

such that K K
mo
,k
. Moreover for some monadic
k

-interpretation scheme c with the leaf property, we have K K

mo,db
c
.
We will not go into detail here especially as a similar result was proved by
Courcelle in [2]. We do however give a sketch of a proof containing some denitions
that will be useful later.
Sketch. Suppose P and S are the nite sets of structures and operations generating
K, and
+
the vocabulary associated with K (see 2.7). Now with every M K we
can associate a tree which represents the construction of M from the structures in
P. Formally we dene:
Denition 2.15. We say that the pair (T, M) with T = T;
T
, c
T
rt
, S
T
1
, S
T
2
) a DB
tree and M= M
t
: t T), is a full representation of M K when:
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 7
(1) Every M
t
is a k
t
-const
+
-structure for some k
t
k

.
(2) For every t T
T
-maximal, M
t
P.
(3) The -redact of M
c
T
rt
is M.
(4) For every t, a non-maximal element of T, let s
1
, s
2
be its immediate succes-
sors with s
l
S
T
l
. Then M
t
= M
s1

s
M
s2
for some s S

+
,ks
1
,ks
2
,kt
S.
Denition 2.16.
(1) Let

be the vocabulary
k2trees
with the following unary predicates:
(a) S
1
and S
2
.
(b) P
k
for k k

.
(c) Q
s
for s S.
(d) R
N
for N P.
k
2
is the total number of unary predicates in

, i.e k
2
= [P[ +[S[ +k

+2.
(2) A

-structure T is a representation of M K, if we can nd M = M

t
:
t [T[) such that:
(a) (([T[,
T
, c
T
rt
, S
T
1
, S
T
2
), M) is a full representation of M.
(b) P
T
k
: k k

) is a partition of [T[. If t P
T
k
, then k
t
= k i.e. M
t
is a
k-const
+
-structure. We write k
T
(t) = k i t P
T
k
.
(c) Q
T
s
: s S) R
T
N
: N P) is a partition of [T[.
(d) For every t T
T
-maximal, t R
T
Mt
.
(e) For every t T non-maximal, let s
1
, s
2
be its immediate successors
with s
l
S
T
l
. Suppose M
t
= M
s1

s
M
s2
for some s S

+
,ks
1
,ks
2
,kt

S. Then t Q
T
s
.
Note that:
Observation 2.17.
(1) Every M K has a full representation, and hence a representation.
(2) If M
l
K are represented by T
l
for l 1, 2, and T
1

= T
2
. Then M
1

=
M
2
.
Now dene: k
1
= max[N[ : N P, k
2
is the number of unary predicates in

(see 2.16(1)), and let k

= maxk
1
, k
2
. We can dene a k

-interpretation
scheme c with k
c
1
= k
1
and k
c
2
= k
2
such that for all M K, and T a representation
of M we have M

= T
[c]
. Note that indeed T is a DB k
c
2
-tree. We will not specify
all the formulas of c as they tend to be very long and complicated, but do note
that all the information about M can be decoded from the representation of M
using monadic formulas. Finally by an argument very close to that of 2.13 we may
assume that c has the leaf property.
3. Equivalence of m-ary patch-width for m 3
We come now to the main part of our result. Basically what we do here is
proving the reverse inclusion of 2.14. It turns out that in our constructible class we
only need 3-ary relations as auxiliary relations, thus we can replace constructible
by 3-constructible. It follows that a class K is contained in a constructible class, i
it is contained in a 3-constructible class, and similarly for m-ary path-width. We
start with an investigation of directed binary trees that will be useful later.
Notation 3.1. Let T be a DB k-tree. Let n N and x
1
, ..., x
n
T be xed maximal
elements of T.
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


8 SAHARON SHELAH AND MOR DORON
(1) For x, y T denote by x y the minimal element z with z x, y.
(2) For x, y T with x y denote [x, y) := z Tx z < y and similarly
(x, y), (x, y] and [x, y].
(3) Dene Y := x
1
, ..., x
n
x
i
x
j
: i, j n c
T
rt
. Note that [Y [ 2n.
(4) For any non-maximal x T let F
R
(x) T (resp. F
L
(x) T) be the unique
immediate successor of x which is in P
T
1
(resp. P
T
2
).
(5) For y, y
t
Y with y < y
t
dene,
T
3
y,y
:= [y, y
t
]
z T : (s (y, y
t
))F
R
(s) y
t
F
L
(s) z
z T : (s (y, y
t
))F
L
(s) y
t
F
R
(s) z
(6) Let T
R
= c
T
rt
t T : F
R
(c
T
rt
) t, and similarly T
L
.
Lemma 3.2. Let R
R
(y, y
t
) and R
L
(y, y
t
) be binary relations meaning F
R
(y) y
t
and F
L
(y) y
t
respectively. The type tp
MSO
q
((x
1
, ..., x
n
), T) is computable from
the structure Y ;
T
, R
R
, R
L
), the types tp
MSO
q
((y, y
t
), T[
T
3
y,y

) : y, y
t
Y, y <
y
t
, (y, y
t
) Y = , and the types tp
MSO
q
(c
T
rt
, T[
TL
), tp
MSO
q
(c
T
rt
, T[
TR
).
Proof. Without going into detail note that from the sets T
3
y,y
with y, y
t
as above,
T
L
and T
R
, we can choose a decomposition of [T[, in which only the elements of Y
belong to more then one set. Hence we can reconstruct the structure T with the
elements of Y as marked elements from the reduced structures: T[
T
3
y,y

with y, y
t
as marked elements, T[
TL
and T[
TL
with C
T
rt
as marked element, in a way that the
q theory of the resulting structure depends only on the q theory of the operands.
The structure Y ;
T
, R
R
, R
L
) determines the order of the construction.
Claim 3.3. Let k

be a natural number, and c a monadic k

-interpretation scheme
with the leaf property for a vocabulary . Then there exists a natural number k

,
and a (3, k

)-constructible class of structures, K, such that: K

mo,db
c
K.
Proof. Let q

be the maximal quantier rank of the formulas

Q,0
: Q .
Dene the vocabulary
+
to consist of:
.

k2trees
.
Two 3-place relations R
R
and R
L
.
For each t TP
MSO
q
(2,
k2trees
), a 3-place relation R
3
t
.
For each t TP
MSO
q
(2,
k2trees
), a 2-place relation R
2
t
.
For each t TP
MSO
q
(1,
k2trees
) two 0-place relations R
R
t
and R
L
t
.
Before we dene the set of addition operations S, and the set P, let us dene:
Denition 3.4. A
+
-structure, T, is called a correct k
2
-tree if:
For each Q , Q
T
= .
T[
k2trees
is a DB k
2
-tree.
For each x
1
, x
2
, x
3
maximal elements of [T[, let y = x
1
x
2
and y
t
= y x
3
then we have, R
T
R
(x
1
, x
2
, x
3
) F
R
(y) y
t
, and similarly for R
L
.
For each t TP
MSO
q
(2,
k2trees
), and x
1
, x
2
, x
3
maximal elements of
[T[, let y = x
1
x
2
and y
t
= y x
3
then we have, (R
3
t
)
T
(x
1
, x
2
, x
3
)
tp
MSO
q
((y, y
t
), T[
T
3
y,y

) = t.
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 9
For each t TP
MSO
q
(2,
k2trees
), and x
1
, x
2
maximal elements of [T[, let
y = x
1
x
2
then we have, (R
2
t
)
T
(x
1
, x
2
) tp
MSO
q
((c
T
rt
, y), T[
T
3
c
T
rt
,y
) = t.
For each t TP
MSO
q
(1,
k2trees
), (R
R
t
)
T
= T i tp
MSO
q
(c
T
rt
, T[
TR
) = t,
and similarly for R
L
t
.
Note that every DB k
2
-tree can be uniquely extended to a correct DB k
2
-tree.
Now dene Our P to consist of all singleton correct models (models with one
element) of the vocabulary
+
, plus all the Null
+
-structures (see denition 2.5(5))
.
We now turn to the denition of the operations in S. Let u be a possible color
of a singleton k
2
-tree. Formally u P
3
, ..., P
k2
. We dene the operation
u
on
DB k
2
-trees as the addition of two trees with root of color u. Formally Let T
1
, T
2
be DB k
2
-trees dene T = T
1

u
T
2
by:
[T[ = [T
1
[ [T
2
[ c.
c is the root of T i.e. c
T
rt
= c and t [T[, c <
T
x.
c has color u i.e. for all i 3, c P
T
i
i i u.
c
T1
rt
P
T
1
and c
T2
rt
P
T
2
.
The rest of the relations on T
1
and T
2
remain unchanged.
Note that indeed T
1

u
T
2
is a DB k
2
-tree whenever T
1
and T
2
are, and hence
u
extends uniquely to an operation on correct k
2
-trees.
Now for l 1, 2 let A
l
be a
+
structure such that there exists a correct k
2
-tree
with [A
l
[ [T
l
[, T
l
[
]A
l
]
= A
l
, and every element of A
l
is maximal in T
l
. Dene an
operation s
u
on such structures by: A
1

su
A
2
= (T
1

u
T
2
)[
]A1]]A2]
. It is easy to
verify that
su
is well dened and indeed belongs to S

+
,0,0,0
. We now have:
Lemma 3.5. For every correct k
2
-tree, T and every set A [T[ of maximal el-
ements, the restriction T[
A
is in the closer of P under the operations s
u
: u
1, ..., k
2
.
Proof. First it is obvious that we can construct T from P using the operations

u
: u 3, ..., k
2
. Now use the same construction only replace in each step the
operation T
1

u
T
2
by the operation T
1
[
A

u
T
2
[
A
.
The last thing we need now is to decode the relations in the correct structure
into the relations in our vocabulary . For this we use:
Lemma 3.6. There exist s

+
,0,0,0
such that For every correct k
2
-tree, T and
every set A [T[ of maximal elements, the structure A
t
= T[
A

s
Null

satises
for each Q with n(Q) = n,
() Q
A

= (x
1
, ..., x
n
)
n
A : T [=
Q,0
(x
1
, ..., x
n
).
Proof. Let Q be an n-place relation symbol, and w
1
, w
2
1, ..., n. We should
dene f
s3
Q,w1,w2
in such a way that () will hold. As we have k
s

1
= k
s

2
= k
s

= 0
and we are only interested in Null

as the right operand, the only relevant case is

w
1
= 1, ..., n and w
2
= . In order to have () We need to dene a function:
f
s

Q
: p : p is a quantier free type of n variables in vocabulary
+
T, F
such that for all (x
1
, ..., x
n
)
n
A, f
Q
(tp
qf
((x
1
, ..., x
n
), A
t
)) = T i T [=
Q,0
(x
1
, ..., x
n
).
Recall that by lemma 3.2 the value of T [=
Q,0
(x
1
, ..., x
n
), is determined by Y ;
T
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


10 SAHARON SHELAH AND MOR DORON
, R
R
, R
L
), the types tp
MSO
q
((y, y
t
), T[
T
3
y,y

) : y, y
t
Y, y < y
t
, (y, y
t
) Y = , and
the types tp
MSO
q
(c
T
rt
, T[
TL
), tp
MSO
q
(c
T
rt
, T[
TR
) (see 3.2 and notation 3.1). But as T
is correct these all are determined by p so we are done.
We can now conclude the proof of lemma 3.3. Dene S = s
u
: u 1, ..., k
2

, and let K be the constructible class of -structures dened by P and S. Let

M be a -structure in K
mo,db
c
. So we have M

= T
[c]
1
for some DB k
2
-tree T
1
.
Let T
2
be the correct extension of T
1
. Let A = x [T
2
[ : T
1
[=
=,0
(x), and
A = T
2
[
A

s
Null

. From lemma 3.5 we have that T

2
[
A
is in the closer of P under
the operations in S and hence so is A. From lemma 3.6 and the denition of T
[c]
1
,
we have that A[

= T
[c]
1

= M, so M K as desired.

From lemmas 3.3 and 2.14 we conclude our main:

Theorem 3.7. Let K be a class of -structures. Then K is contained in a m-
constructible class for some m N i K is contained in a 3-constructible class.
The same holds for patch-width:
Corollary 3.8. Let be a nice vocabulary and K a class of -structures. Then
K is contained in a class of bounded m-ary patch-width for some m N i K is
contained in a class of bounded 3-ary patch-width.
Proof. Assume K K
t
for some K
t
of bounded m-ary patch-width. By lemma 2.9 K
t
is (m, 0)-constructible. By theorem 3.7 K
t
is contained in some 3-constructible K
tt
.
Notice that that the set S dened in the proof of 3.3 satises that S S

+
,0,0,0
so K
tt
is in fact (3, 0)-constructible. Notice further that in the proof of 3.3 as we
do not need null structures in the construction, hence we may replace
+
by a
nice vocabulary. So by lemma 2.10 K
tt
is a bounded 3-ary patch-width class as
desired.
4. A counter example for the unary case
It turns out that we can not replace the number 3 in theorem 3.7 by 1. This is
because of the following:
Theorem 4.1. There exists a nice vocabulary , and a class of -structures K,
contained in some 3-constructible class, that is not contained in any 1-constructible
class.
Proof. Let = R with n(R) = 4. Set p N be large enough (to be dened later).
Let T be a tree. For x, y T Dene:
x
T
y = x y = the <
T
minimal z T such that z
T
x and z
T
x.
d
T
(x, y) = d(x, y) = min[S[ : S T, x, x y S, S is dense in (T, <
T
).
d
T
p
(x, y) = d
p
(x, y) = d(x, y) (mod p).
Let q : 0, ..., p 1
2
0, 1 be some function that will be dened later. We
now dene c a 0-interpretation scheme for :
k
c
1
= k
c
2
= 0.

c
=,0
(x) = yy > x i.e. the elements of the interpreted structure are the
leafs of the tree.

c
R,0
(x
1
, x
2
, x
3
, x
4
) = q(d
p
(x
1
, x
2
), d
p
(x
3
, x
4
)) = 0.
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 11
We have to show that
R,0
is indeed a monadic formula in
trees
. Note that there
exists a monadic formula
dp=0
(x, y) such that for any tree T, T [=
dp=0
(x, y) i
d
T
l
(x, y) = 0.
dp=0
(x, y) will say that there exists a set X such that:
x, x y X,
if z, z
t
X and z < z
tt
< z
t
then z
tt
X,
if z
t
is the immediate successor in X of z X, then there exist exactly
p 1 elements (of T) between them.
similarly we have formulas
dp=i
(x, y) for 0 < i < p. Now dene:

R,0
(x
1
, x
2
, x
3
, x
4
) =

n1,n20,...,p1]
n1n2 (mod p)

dp=n1
(x
1
, x
2
)
dp=n2
(x
3
, x
4
).
This gives us c as desired. Dene K = K
mo,db
c
. By 3.3 K is contained in a 3-
constructible class (in fact in a 3-ary BPW class).
For each n N let M
n
= (
n
2, ) i.e. M
n
is the complete binary tree of depth
n, and N
n
= M
[c]
n
. Let K
t
be a constructible class of -structures, so
+
=
k
for
some k N. Towards contradiction assume that N
n
K
t
for all n N. Let P be
the set of atomic structures associates with K
t
. w.l.o.g. we may assume that P
consists of singleton structures only. Otherwise increase k by max[M[ : M P
and construct each M P from singletons of distinct colors. Now let K K
t
, and
let (T, M) be a full representation of K (see 2.15). Assume K

= N
n
for some n. So
we have a 1-1 function f, from
n
2 to the leafs of T, as every
n
2 corresponds to
a unique element a K under the isomorphisms, and for every element of a K
there exist a unique t a leaf of T such that a = [M
t
[. Dene f() = t. Note
that f is not onto, as some of the leafs of T may be omitted during the creation
process. For each t T let A
t
= f
1
(s) : s
T
s range(f). So A
t

n
2.
For each A
t
let a = a

= [M
f()
[. a

is an element of M
t
, so A
t
is divided
into 2
k
parts according to the color of a

in M
t
, (more formally according to the
type tp

+
\
qf
(a

, M
t
). We therefore have B
t
A
t
such that [B
t
[
]At]
2
k
, and all the
elements of f
1
(B
t
) have the same color. Now dene:
C
t
= d
Nn
p
(, ) : , B
t
0, ..., p 1.
We have
]At]
2
k
[B
t
[ 2
]Ct]
. For the right-hand inequality use induction on [C
t
[.
Hence we conclude
[A
t
[ [B
t
[ 2
k
2
]Ct]+k
.
Now note that if C
t
,= 0, ..., p1, then C
t
n
n
p
| and hence [A
t
[ 2
]Ct]+k

2
n
n
p
+k
. We now consider two cases:
Case 1 There exist s T with two immediate successors t
1
, t
2
T such that:
[A
t1
[, [A
t2
[ > 2
n
n
p
+k
.
According to what we saw above we have C
1
= C
2
= 0, ..., p 1. So for
l 1, 2 we have (
t
l
,i
,
t
l
,i
: i 0, ..., p 1) such that:
()
t
l
,i
,
t
l
,i
: i 0, ..., p 1 all have the same color in M
t
l
.
() d
Nn
p
(
t
l
,i
,
t
l
,i
) = i for all i < p.
Denote by m the number of quantier free types of couples in the vocabulary
(actually in our case m = 2
(2
4
)
). Note that m does not depend on p. So
for each l 1, 2, 0, ..., p 1 is deviled into m parts according to the type
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


12 SAHARON SHELAH AND MOR DORON
tp
qf
((
t
l
,i
,
t
l
,i
), M
t
l
). We claim that we can (a priori) choose p (large enough)
and q in such a way that we can nd i
1
, i
2
, j
1
, j
2
such that for each l 1, 2:
(
t
l
,i
l
,
t
l
,i
l
) and (
t
l
,j
l
,
t
l
,j
l
) have the same quantier free type in vocabulary in
M
t
l
, and on the other hand: q(i
1
, j
1
) ,= q(i
2
, j
2
). This is of course a contradiction
as the quantier free type of (
t
l
,i
l
,
t
l
,i
l
) and (
t
l
,j
l
,
t
l
,j
l
) in vocabulary
k
in M
t
l
determines the value of R(
t
l
,i
l
,
t
l
,i
l
,
t
l
,j
l
,
t
l
,j
l
) in M
s
and hance in M
c
T
rt
. But this
value is true i q(i
l
, j
l
) = 0 in contradiction with q(i
1
, j
1
) ,= q(i
2
, j
2
). Why can we
choose p and q as desired? For a given p the number of functions from 0, ..., p1
2
to 0, 1 such that we can not choose as above (i.e. functions that respects some
partition of 0, ..., p 1 into m parts is the number of partitions m
p
m
p
, time the
number of functions that respect that partition 2
mm
, or 2
2p log(m)+m
2
. The total
number of functions is 2
p
2
. So if we choose (a priori) p such that p
2
> 2p log(m)+m
2
we can choose a function q as desired.
Assume now that the assumption of Case 1 does not hold. Assume also that
we have chosen n large enough such that 2

n
p
k
> 4. In this case we can nd
t
0
, t
1
, ..., t
d
T such that :
d 5.
t
0
= c
T
rt
.
t
d
is a leaf of T.
For 0 i < d, t
i+1
is an immediate successor in T, of t
1
.
For 0 i < d, denote by s
i+1
the immediate successor of t
i
dierent from
t
i+1
, then [A
si+1
[ 2

n
p
k
.
Note that for any 0 < i d:

0<ji
A
sj
and A
ti
is a partition of A
c
T
rt
, and that
[A
c
T
rt
[ = 2
n
. So we can nd 0 < i

d such that [

0<ji

A
sj
[, [A
t
i

[ > 2

n
p
k
.
We proceed similarly to Case 1. As there we can nd (
t
i
,i
,
t
i
,i
A
t
i

:
i 0, ..., p 1) that satisfy () and () above, and the same for (
i
,
i
: i
0, ..., p 1) where
i
,
i


0<ji
A
sj
. Again let m denote the number of
quantier free types of couples in the vocabulary . This time we want to choose
p and q in such a way that we can nd: i, j
1
, j
2
such that: (
t
i
,j1
,
t
i
,j1
) and
(
t
i
,j2
,
t
i
,j2
) have the same quantier free type in vocabulary in M
t
i

, and on
the other hand: q(i, j
1
) ,= q(i, j
2
). Again this is a contradiction as the quantier free
type of (
t
i
,j
l
,
t
i
,j
l
) for l 1, 2 determines the value of R(
t
i
,j
l
,
t
i
,j
l
,
i
,
i
) in
M
c
T
rt
. Again this value is true i q(i, j
l
) = 0 in contradiction with q(i, j
1
) ,= q(i, j
2
).
Why can we choose p and q as desired? For a given p the number of functions
from 0, ..., p 1
2
to 0, 1 such that we can not choose as above is the number
of partitions m
p
, times the number of functions that respect that partition 2
mp
,
or 2
p log(m)+mp
. So if we choose p such that p
2
> p log(m) +m p we can choose a
function q as desired. Note that the function we used for the second case will also
work for the rst case so we can use one denition of q. In both cases we get a
contradiction and the proof is complete.

References
[1] G unter Asser. Das Repr asentantenproblem im Pr adikatenkalk ul der ersten Stufe mit Identit at.
Z. Math. Logik Grundlagen Math., 1:252263, 1955.
8
6
5


r
e
v
i
s
i
o
n
:
2
0
0
6
-
0
7
-
1
9







m
o
d
i
f
i
e
d
:
2
0
0
6
-
0
7
-
1
9


BOUNDED m-ARY PATCH-WIDTH ARE EQUIVALENT FOR m 3 13
[2] Bruno Courcelle. The monadic second order logic of graphs. VI. On several representations of
graphs by relational structures. Discrete Appl. Math., 54(2-3):117149, 1994. Ecient algo-
rithms and partial k-trees.
[3] Arnaud Durand, Ronald Fagin, and Bernd Loescher. Spectra with only unary function symbols.
In Computer science logic (Aarhus, 1997), volume 1414 of Lecture Notes in Comput. Sci.,
pages 189202. Springer, Berlin, 1998.
[4] E. Fischer and J. A. Makowsky. On spectra of sentences of monadic second order logic with
counting. J. Symbolic Logic, 69(3):617640, 2004.
[5] Yury Gurevich and Saharon Shelah. Spectra of monadic second-order formulas with one nary
function. LiCS03. IEEE, 2003.
[6] Saharon Shelah. Spectra of monadic second order sentences. Sci. Math. Jpn., 59(2):351355,
2004. Special issue on set theory and algebraic model theory.
The Hebrew university of Jerusalem
Einstein Institute of Mathematics
Edmond Safra Campus
Givat Ram
Jerusalem 91904, Israel
E-mail address: shelah@math.huji.ac.il
E-mail address: mord@math.huji.ac.il