Professional Documents
Culture Documents
Abstract-A coterie under a ground set lr consists of subsets many other coteries, which are not vote-assignable. The main
(called quorums) of U such that any pair of quorums intersect motivation for generalizing the “majority” idea to general
each other. “Nondominated (ND)” coteries are of particular coteries was to have a large number of choices in implementing
interest, sinc’e they are “optimal” in some sense. By assigning
a Boolean variable to each element in U , we represent a family distributed mutual exclusion, in order to select the most
of subsets of U by a Boolean function of these variables. We suitable one for the application in hand. Consider, for example,
.. characterize the ND coteries as exactly those families which can a replicated database, in which the copies of each data item
be represented by positive, self-dual functions. In this Boolean are stored at different sites interconnected by a communication
- framework, we prove that any function representing an ND
. coterie can be decomposed into copies of the three-majority network. When the sites and/or the communication links are
function, an$ this decomposition is representable as a binary subject to failure, the availability of a data item is defined
tree. We also show that the class of ND coteries proposed by as the probability that all sites storing a quorum of copies
Agrawal and El Abbadi is related to a special case of the above are operating and can communicate with each other (so that
binary decomposition, and that the composition proposed by the data item can be accessed for reading or writing). It is
Neilsen and Mizuno is closely related to the classical Ashenhurst
decomposition of Boolean functions. A number of other results known that, for a general underlying network, the coterie
are also obtained. The compactness of the proofs of most of these with the highest availability may not be vote-assignable [25].
results indicates the suitability of Boolean algebra for the analysis Therefore, we need to study other more general coteries. Also,
of coteries. if the underlying network has some special structure (e.g.,
. Index Terms- Boolean algebra, coteries, decomposition of mesh or hypercube), it is conceivable that the complexity of
Boolean functions, distributed systems, mutual exclusion, self- mutual exclusion protocols depends on the adopted coterie. In
dual Boolean functions. such a case, it would be advantageous to have a variety of
coteries as candidates for implementation.
I. INTRODUCI‘ION AND DEFINITIONS Consider a ground set U , consisting of n elements. Any
subset of these elements can be represented by an n-vector,
X = (x~,x~,...,x,) E {O,l},, where zi is 1 (0) if
A. History and Motivation
the ith element is (not) in the subset. In this paper, we
threshold function with the input weights equal to 1 and the ii) [minimality] There are no G ,H E C such that G c H ,
threshold value equal to 2) one of whose inputs is a single and
variable. Another class of ND coteries has recently been iii) [intersection property] VG,H E C: G n H # 4. 0
identified by Neilsen and Mizuno [24]. In terms of our Boolean According to the above definition, the empty family contain-
representation, it turns out that the decomposition operation ing no subset is a coterie. The subsets belonging to a coterie
used to describe this class is a special case of the classical are called quorums.
Boolean function decomposition due to Ashenhurst [2], studied Domination [12]: Let C and D be coteries. C dominates D
extensively in switching theory and other fields (see Remark if C # D and VG E D 3H E C : H G. A coterie C is
5.3). nondominated (ND) if no coterie dominates C. 0
A vote-assignable coterie has the desirable property that To see that ND coteries are particularly useful in practical
it can be represented by a set of n weights, where n is, as applications, let coterie D be dominated by C. Then if a
before, the number of elements in the underlying set. The partition (that results from a network partition failure) contains
only known general way of describing an arbitrary ND coterie, a quorum belonging to D, then it must contain a quorum
however, was to list all its quorums. This is not, in general, belonging to C.
space-efficient, since a coterie can have an exponential number Bicoterie [ll]: An ordered pair B = ( P ,Q ) , where P and
(in terms of n) of quorums. A composition of copies of the Q are families of subsets of U, is a bicoterie under U if
basic majority function that realizes a given positive, self-dual i) Both P and Q satisfy conditions i) and ii) for a coterie
function (e.g., the B-decomposition mentioned above), on the (with C = P,Q).
other hand, often provides a compact method of describing an ii) V G E P V H E Q : G ~ H # ~ .
ND coterie. A bicoterie B = ( P ,Q ) is called a write-read coterie
Some concurrency control schemes for a replicated database (wr-coterie, for short), if it satisfies
system (e.g., [3], [7], [ l l ] , [15]) make use of a set of iii) VG,H E P: G n H # 4 (i.e., P is a coterie). 0
“read quorums” and a set of “write quorums” such that In [ 3 ] , families P and Q in a bicoterie B = ( P , Q )
each write quorum intersects each write quorum and each
are said to be “complementary” to each other. A write-read
read quorum but the read quorums need not intersect each
coterie B = ( P ,Q ) models “write quorums” by P and “read
other. To model such a system, we introduce “bicoteries” and
quorums” by Q.
“write-read coteries,” and investigate their properties as well.
Domination [ 111: Bicoterie B = ( P ,Q ) dominates bicoterie
They are closely related to “complementary quorum set” and
B’ = (R, S) if
“antiquorum set” defined in [3].
In the rest of this section we define terms that are used in i) VG E R 3H E P : H C G,
this paper. Section I1 reviews, in the terminology of Boolean ii) VG E S 3H E Q: H C G, and
algebra, a number of known results as well as those that iii) ( P , Q ) # ( R , S ) ,i.e., P # R or Q # S.
follow easily from definitions. In Section 111, we investigate A bicoterie B is said to be nondominated (ND) if no
the relation among coteries, bicoteries, and write-read coteries. bicoterie dominates B. Similarly, a wr-coterie B is said to
Our discussion of coterie decomposition begins in Section IV- be nondominated (ND) if no wr-coterie dominates B. 0
A, and our main theorem mentioned above will be proved A nondominated bicoterie is called a “quorum agreement”
in Section IV-B. In Section IV-C, we consider the “nonre- in [3].
dundantly B-decomposable” coteries, as a generalization of In this paper, we define the Boolean value 0 to be less than
the class of coteries consisting of structured quorums [l]. the Boolean value 1, i.e., 0 < 1. For two Boolean vectors
Section IV-D considers decomposition of general coteries X = (51, z2,. . . ,zn)and Y = (yl,y2, .. . ,y,), we write
and its relevance to ND coteries. In Section V, we con- X 5 Y if 2; 5 yi for all i = 1 , 2 , . . . , n.
sider another kind of decomposition of ND coteries as a Positive Boolean function: A Boolean function f is positive
generalization of the method proposed in [24], and discuss (or monotone increasing) if V X , Y E {0, I}n: X 5 Y
its relevance to the Ashenhurst decomposition. Section VI =+ f ( X ) i f ( Y ) - 0
focuses on vote-assignable coteries and their generalization Minimal vectors: Let f be a positive Boolean function.
to multidimensional vote-assignable coteries introduced in X E { O , l } n is a minimal vector of f if f ( X ) = 1 and
[8]. Preliminary versions of this paper appeared before as a [(Y 5 X ) A (Y # X ) ] + f ( Y ) = 0. The set of all
technical report [16] and also in conference proceedings [17]. minimal vectors (or the subsets they represent according to
In this paper, we often omit the Boolean AND operator our convention given below) of function f will be denoted by
A, and simply concatenate variables to be ANDed together. M inS et (f). 0
E and c will denote set inclusion and proper set inclusion, Let C be a family of distinct nonempty subsets of U
respectively. satisfying the minimality condition. In this paper, we do
not distinguish subset G E C from its characteristic vector,
(91, g2,.. . ,gn), where for i = 1 , 2 , . , n,
+ .
B. Definitions
Coterie [12]: Let U = {u1,u2,. . . , un}be a ground set and
C be a family of subsets of U. C is a coterie under U if 1 if ui E G ,
i) G E C + G # 4, si = 0 if U , 4 G.
IBARAKI AND KAh4EDA: THEORY OF COTERIES: MUTUAL EXCLUSION IN DISTRIBUTED SYSTEMS 781
f C ( 4 = { 0 otherwise.
(This relationship between fc and
GEC u,EG
fg
implies that
It is easy to see that M i n S e t ( f c ) = C holds if C satisfies M i n S e t ( f g ) represents the set of all minimal transversals
the minimality condition. [12] of C.) 0
Remark 1.1: It is known (e.g., [23]) that any positive Property 1.3: Let fc be the Boolean function representing
function f can be represented by a unique disjunctive form a family C of nonempty subsets of U , which satisfies the
without complemented literals such that each term is the prime minimality condition. Then the dual of fc is given by
implicant corresponding to a minimal vector. This form is
0 if X 2 G for some G E C
called the minimum sum expression for f . For fc, it has the
following form.
f 3 X )= { 1 otherwise
f C ( X )= v {
GEC u,EG
A %I. Property 1.4: If f is positive, then so is f d . 0
0
ii) n = 2 : None.
A
iii) n = 3: There is only one such function f = xyvyzvzz,
where U = { x , y , z } . 0
We shall call the function
M(z,y,z) = xy v yz v zx
fl I2
the basic majority function. By the above remark, this function
Fig. 1. Decomposition of f . using the binary operator hfz.
represents the simplest nontrivial ND coterie. Theorem 4.1
implies that any function composed of basic majority functions
(without using constant 0 or 1) is positive and self-dual. A n
interesting question would be: what class of positive, self- f
dual boolean functions can be composed from copies of the n
basic majority function? To answer this question, we define the
decomposable, positive, self-dual function f on U as follows.
1) f = 2 , where z E U , or
2) f = M(fl,f2,f3), where fl, f 2 and f3 are decompos-
able functions on U .
We shall show in the next section that every ND coterie
is decomposable. Thus the basic majority function is a “uni-
versal function” for positive, self-dual boolean functions, just
like NAND and NOR functions are universal functions for all
boolean functions. In fact, it turns out that every ND coterie
has a more specific form of decomposition, which we call Fig. 2. Illustration for Example 4.2.
“binary decomposition” (B-decomposition, for short), as will
be discussed in the next section.
If all functions become single variables, then the generated
finite tree is called the binary decomposition tree for f .
B. B-Decomposition of ND Coteries
Example 4.2: Let f l and f2 be defined by
Define the function on two variables,
Mz(y,2 ) = M ( z ,9.2 )
based on the basic majority function. We consider composition The decomposition tree for the function f defined by
of f from f l and f z (or decomposition of f into fl and f 2 )
of the form. f = Mw(fl,f2) = ( u y v x y v u u v v x ) w v ( u x v u v y v v x y )
is shown in Fig. 2. 0
Remark 4.2: Different trees may represent the same ND
where f l and f2 are positive, self-dual functions. Note that coterie. For example, a tree consisting of the root labeled y
f i and J2 may depend on 2 . By Theorem 4.1, f is positive and two child nodes labeled z and y represents an ND coterie
and self-dual. By considering MinSet( f ) , i.e., the family of { {y}}, which is also representable by a one-node tree labeled
subsets that f represents, we obtain the following result. Y. 0
Theorem 4.2: Let C1 and C, be ND coteries under U . The We say that a positive function f [or the ND coterie
following family C is also an ND coterie. M inset (f )] has a binary decomposition (B-decomposition,
for short) or f is binary decomposable (B-decomposable, for
C 1M i n i m d ( { G U {z}lG E Cl U ( 7 2 )
short) if
U {GI U GzlGi E Ci,G2 E C 2 } ) . 1) f is a single variable, or
2) f = Mx( fl , fz) for some variable z, where both f l and
Proof: Let f l = fc, and f2 = fc,. Then C =
f2 have B-decompositions.
MinSet( f ) , where f is given by (4). The rest follows from
point i) given just after Theorem 2.4. 0 To prove that every ND coterie is B-decomposable, we start
The above decomposition (4) can be conveniently repre- with some lemmas.
sented by a binary tree, as illustrated in Fig. 1, where the label Lemma 4.1: Let a nonconstant function g be positive and
of each internal node indicates the subscript x of the binary dual-major (resp., dual-minor). Then g can be represented as
operator, M,, used for decomposition. By recursively applying
this decomposition to the generated f,’s, we may eventually
end up with functions of single variables. A single-variable
function is represented by a leaf node labeled by the variable. for some IC > 0, where each f i is positive and self-dual.
IBARAKI AND KAMEDA: THEORY OF COTERIES: MUTUAL EXCLUSION IN DISTRIBUTED SYSTEMS
785
Proof: We only prove the case where g is dual-major. Proof: Take any .r E I' and decompose f with respect
The proof for the dual-minor case can be obtained by taking to r a5 in (3) of Section 111, i.e.,
the dual of the dual-major case. Observe the following.
f =rf,=lVf,=o=zg d vg.
i) Since g is dual-major, at least one of g ( X ) and g ( X )
equals 1 for any vector X [by Property 1.5ii)l. where g = f,,oand 9 is dual-minor. By Lemma 4.1, g d can
ii) Since g is positive, g ( Y ) = 1 implies g ( X ) = 1 for any be represented as
>
X Y . (Or equivalently, g ( Y )= 0 implies g ( X ) = 0
for any X 5 Y . ) gCi= fl k f2 v " ' vfk
Given a dual-major g, we first present a formal procedure and g as
for constructing, for an arbitrary Y E M i n S e t ( g ) ,a positive,
self-dual function fir, satisfying f i F ( Y ) = 1 and f y 5 g. The g= f:AftA..'A f t =f l A f 2 A . . . A f k
following procedure constructs f r ~in terms of its true domain
for some self-dual functions, f l . f 2 , . . . f k , on U - {r}. 0
and false domain, T and F, respectively [i.e., f l , ( X ) = 1
The next lemma shows that f can be composed from func-
(resp., 0) for all X E T (resp., F ) ] .
tions f l . f 2 . f . . . f k , using only the basic majority function.
. 1) Let T := { X l X 2 Y } and F := ( Z ( g ( 2 )= O}. (This Lemma 4.3: Consider the decomposition
makes sure that fl,(Y) = 1 and fu 5 g hold initially.)
-
2) If 3 2 4 TUF such that 2 E T , then F := F U ( X J X 5 f = v
(fl f'2 v. v
' ' fk)n v (fl A f2 A " ' A fk)
2).(Since it is self-dual, f u ( X ) must satisfy fl,(z) in Lemma 4.2. This can be represented as
= 1 +*fl,(Z) = 0.) Repeat this step as long as it is
applicable.
3) If TUF = (0, l}n,then stop. Otherwise, choose a vector
Z 4 T U F (T U F contains neither 2 nor Z),and let and
T := TU { X l X 2 2 ) . Return to step 2.
E: = ,W,(fL. F t + l ) , for L = 2.3,... .k - 1,
Clearly, the above procedure terminates, since T always
gets enlarged in step 3. The fact that the resulting f1- is where Fk fk.
indeed self-dual follows from the property of T and F that, Proof: Easy if one treats the two cases, x = 0 and .7: = 1,
for any vector X , exactly one of X and X belongs to T separately. 17
(and the other belongs to F). We now prove this property. Theorem 4.3: Any ND coterie is B-decomposable.
Since T U F = (0, l}n upon termination, it suffices to show Proof: By Lemmas 4.2 and 4.3, a self-dual function f
that neither T nor F contains both X and X during the has a B-decomposition into smaller self-dual functions (fz},
construction. That this is true initially, i.e., for the sets T and F which do not depend on the variable II: used for decomposition.
of step 1 above, can be shown as follows. Since X 2 Y ( # 0) Apply this decomposition recursively to all fi until only single
* obviously means X 2 Y , X E T implies X 4 T . Also, variables result. 0
X E F implies g ( X ) = 0 and hence g(X) = 1, i.e., X 4 F, Corollary 4.1: Any ND coterie is decomposable. 0
since g is dual-major. Next, consider the case where F is Now, B-decomposition discussed so far defines a single-
augmented in step 2. The resulting set F does not contain output circuit made up of copies of the basic majority gate,
both X and X for any X , since for any new vector X 5 Z whose inputs are variables from U (no constants) and whose
in F , X E T holds because X 2 2.Finally, in step 3, the output is f (Fig. 3). To check whether a given subset S C: U
resulting set T has this property, since for any new vector contains a quorum of the corresponding ND coterie, we can
X > Z in T , X 4 T holds because X 5 2 and 2 4 T (if simply feed (the vector representing) S into this circuit. If
2 E T then Z E F by step 2). The positivity of f i - is obvious the output is 1, then S contains a quorum of the ND coterie;
from the construction. otherwise, it does not. In this sense, the circuit is a compact
I Applying the above procedure to every minimal vector Y representation of an ND coterie, which often requires less
of g, we obtain space than a list of all the quorums of the ND coterie.
Remark 4.4: B-decomposition can be generalized if we
= l/ fY. substitute the basic majority function of (4.2) by any positive
YEMmSet(g) +
self-dual ( k 1)-ary function F and represent
0
Remark 4.3: In the above construction, many f1 can be
redundant. It is usually possible to make k in ( 2 ) fairly small. by a node that is labeled by z and has k ( > 2) child nodes,
However, the true minimization is left open. 0 f l . f;... . . . f r c . In particular, if F represents a vote-assignable
Lemma 4.2: Let f be a positive, self-dual function on I J . ND coterie (see Section VI for definition), in which each f z
Then f can be decomposed as follows. has weight 1 and r has weight k - I , it is given by
YPl,YP2 or PlP2,
f = (wv f 2 ) Y v U1f2,
v ((fl)w=O v f 2 ) Y v ( ( f l ) w = O A f 2 ) and fz depends on more than one variable, which is different
= [ ( ( f l ) w = l v f 2 ) Y v ( ( f l ) w = l A f2)l.I from w.Then, as noted in the proof of the if-part of Lemma
v ((fl)w=O v f2)Y v ((fl)w=O A f 2 ) . 4.4, either y or iii (but not others) can be the root. This
corresponds to the case 2 in the statement of the lemma. 0
Define g by
Example 4.3: Consider the positive, self-dual function f of
gd = ( ( f l ) w = l v f2)y v ((fl)w=l A f2)> Example 4.2,
i.e., the coefficient of w in the expression for f . Since tu f = 1Ln: v lU3-y v 1I1UZ v llxy v uuw v uvy v 1L1UY.
is the root (i.e., there is a decomposition of type (7), f =
(91V g2)w V (91 A g2)), gd must be of the form The above lemma (condition 2) asserts that either one of U and
z (and no other variable) can be the root of a decomposition
gd = 91 v g2 tree. 0
where g1 and g2 are positive, functions defined On 3Recall from Remark 4.1 that the positive, self-dual functions of up to three
two disjoint sets of variables. Now, as noted in Remark 1.1, variables are very simple.
a positive function has a unique disjunctive form, called the 4Note that I; 2 2 follows from the assumptions on f.
IBARAKI AND KAMEDA: THEORY OF COTERIES: MUTUAL EXCLUSION IN DISTRIBUTED SYSTEMS 787
TABLE I
GENERATE I O rHr EXAMPLEI N Fig. 5
APPLYING
- w}. { r.
( I ti. 11. J ) . { r . W . s} 1
llfT,,(gi 7 ry v .rg V yg
Fig. 5. A labeled complete binary tree used in the Agrawal-El Abbadi
construction. and
M,,, = -cy v y z v zx
complete binary tree, in which all nodes have distinct labels
from U . where T , y, and z are variables, and represent MTy(g)by a
Algorithm GENERATE: node with label ( r .y), having an edge to the node with label g,
Start with the set { T ) , where E t r is the root. T is and Mzyzby a leaf node with label ( T . y, 2 ) . Fig. 4(b) shows
unscanned. the resulting tree for the function f of Example 4.3. The bi-
If no set contains an unscanned nonleaf node, then halt. nary tree representation of a nonredundantly B-decomposable
The current family of sets is an ND coterie. Other- function f with this compact notation is unique. U
wise, choose an arbitrary set S containing an unscanned Remark 4.7: The number of labeled rooted trees on n
nonleaf node x, and go to step 3. nodes, in which each node has either 0 or 2 child nodes,
Let y and z be the child nodes of s. Generate from gives an upper bound on the number of nonredundantly B-
S the following three sets: S U {y}, S U {t}, and decomposable ND coteries, as different trees may represent
(5' U {g, z } ) - {x}. x is now scanned but y and z are the same Boolean function. Such trees exist only for odd
not. Return to step 2. 7 1 = 2 m + 1. and their number is given as follows (see problem
As an example, consider the complete binary tree given in 21 in Section 3.4.4 of [19]).
Fig. 5. Algorithm GENERATE generates sets S as illustrated in
Table I. Read it from left to right. The underlined labels indi-
cate those nodes which are unscanned and nonleaf. Therefore,
the 15 sets containing only labels without underlines constitute
the constructed coterie. This is asymptotically much smaller than the number of ND
In view of the definition of B-decomposition and Theorem coteries cited in Remark 2.1. 0
4.2, it is not difficult to see that the coterie constructed as
above is equal to M i n S e t ( f ) ,where f is the Boolean function D. R-Decompovtion of General Coteries
associated with the given complete binary tree. The condition The B-decomposition of ND coteries discussed in Section
that the tree be complete can be easily incorporated into our IV-B can be generalized to all other coteries with a slight
algorithm DECOMPOSE by testing in step 2 whether two modification. This will reveal an interesting structural relation
disjoint subsets of variables of f l and f 2 have the same between a coterie and the ND coteries that dominate it.
cardinality. Consider the basic majority function, k f ( x . y. z ) = xyVyzV
It is clear that algorithm GENERATE can be applied not Z T . It is easily seen that M ~ ~ (zy) .: A4(0. y, z ) = yz (AND
only to a complete binary tree but also to any binary tree with function) and l f ~ ( yz. ) = M ( 1 . y . 2 ) = y V z (OR function),
distinct node labels. Given such a tree as input, GENERATE which are conveniently represented by nodes with labels 0
will generate the nonredundantly B-decomposable ND coterie and 1, respectively. Clearly. y z is dual-minor and y V z is
associated with the tree. GENERATE can also be generalized to dual-major.
nonbinary trees with distinct node labels, in which the Boolean Now, consider B-decomposition with the modification that
function associated with each node is of type (6) stated in some nodes may now have labels 0 (but not 1). Since every
Remark 4.4. In this case, the family of sets generated from S node represents a positive, dual-minor function (recall that a
in step 3 of GENERATE should read self-dual function is also dual-minor), repeated applications
sU{y1).sU{y2).... . s U { y k } . (.TU { y ~p2.
. ' . ' . yk})-{:r}. of Theorem 4.lb) show that the function f it represents is
positive and dual-minor. i.e., ( ' - , 2 / l m S e t ( f ) is a coterie.
if the current node has label .r and k child nodes p1. y2 . . . . yk. Example 4 5: The tree shown in Fig. 6 represents
The generalization of this type was already mentioned in [l]. 0
Remark 4.6: Lemma 4.4 asserts that the nonredundant B-
decomposition is unique if we do not distinguish the two roots
which are possible only under the circumstance stated in that
IBARAKI AND KAMEDA: THEORY OF COTERIES: MUTUAL EXCLUSION IN DISTRIBUTED SYSTEMS 789
Lemma 5.1 implies that a positive function f is self-dual if Note that decomposition (2) is a special case of (17), where
and only if it has a representation h ( Y )is a single variable. Decomposition (17) is not universal,
since, for example, the ND coterie consisting of all the three-
element subsets of U = (u1, u2,. . . , us} does not have such
f ( X ) = pd(X’)x,-lxn v q(X’)Z,-l v q d ( X ’ ) x , v p ( X ’ ) , decomposition.
(13) Remark 5.2: In [24], Neilsen and Mizuno proposed a coterie
such that p ( X ’ ) and q(Xf) are positive functions satisfying join algorithm that constructs an ND coterie C3 over ground
set U = ( U - (20))U U’ from two ND coteries C1 and C2
P ( X ’ ) I dX’)I P “ W (14)
under disjoint ground sets U and U’,respectively, where xo
(i.e., p(x’) 5 qd(x’) 5 p’((~’).) is a designated element in U.That is, G3 E C3 holds if and
Note that (13) can be considered as a rule for constructing a only if
self-dual function f from smaller functions p(X’) and q(X’), G3 = G1 f o r some G1 E U with xo G I , or
where p ( X ’ ) and q(X’) canbe chosen independently as long (23 = (GI - (50))U G2 for some G ZE C2 and some
as (14) is satisfied. A special case of this decomposition is
GI E U with xo E G I .
f(X)= ~ ~ - 1 2V ,q(Xf)2,-1 V qd(X’)x, (15) They show, among other things, that C3 is an ND coterie if
(i.e., pd(X‘) = 1 and p ( X ’ ) = 0). In this case, condition (14) and only if so are C1 and C2.
is trivially satisfied for any positive function q(X’). Now, let F ( z 0 , Z ) [resp., h ( Y ) ]be the positive, self-dual
Remark 5.1: Relation (15) implies the following inequality: function associated with C1 (resp., C,) given in the following
(The number of positive, self-dual functions of up to n form:
variables) 2 (The number of positive functions of up to n - 2 F(zo,Z ) = g d ( Z ) z o v 9 ( Z ) ,
variables). where
However, this does not give any better lower bound on the
g d ( Z ) = F ( l . 2)and g ( 2 ) = F ( 0 ,2).
number of positive, self-dual functions than that given in
Remark 2.1. 0 Then it is not difficult to see that the ND coterie given by
Now take any positive, self-dual function f ( X ) and consider
f ( X ) of (17) is nothing but the above ND coterie C3. In this
its decomposition (13). It can be rewritten as follows. sense, (17) is a Boolean description of Neilsen-Mizuno join
algorithm. An advantage of the Boolean approach is that the
f ( X ) = pd(X’)xn-1z, v q(X’)s,-1 v qd(X’)zn v p ( X ’ ) NDness is an immediate consequence of the self-duality of f,
= pd(X’)(3&12, v q(X’)x,-1 v qqX’)x,) though a rather involved argument was necessary in [24] to
show it directly from the join algorithm.
v P ( X 0 [by (211 Moreover, the construction of C3 as above is valid even
= g d ( X ) h ( X )v g(X). if Y and 2 are not disjoint, as it can be considered as an
IBARAKl AND KAMEDA: THEORY OF COTERIES: MUTUAL EXCLUSION IN DISTRIBUTED SYSTEMS 791
algorithm to compute M inset ( f ) of general decomposition Therefore, (21), hence (20), is a decomposition of the form
(10). However, in this case, some generated sets may be (17). 0
included in other sets, and, therefore, operation M znzmnl(C3) Remark 5.3: Properties and algorithms related to A-
is necessary to obtain the corresponding ND coterie. 0 decomposition have been extensively studied in the literature
We point out here the relationship of decomposition (17) of Boolean functions (e.g., [2], [ 141). Interestingly, this subject
to the classical decomposition of Ashenhurst [2] (referred has also been studied in other areas under various names, such
to henceforth as A-decomposition), which can be applied to as “committees” in game theory [28], “clutters” in set systems
any Boolean function. A Boolean function f ( X ) of n ( 2 3) [ 5 ] , and “modules” in coherent systems of reliability theory
variables has an A-decomposition if it can be represented as [6]. Based on the result of Billera [5],Mohring [22] developed
an algorithm to test the existence of A-decomposition for a
f ( X ) = JYW),
2) (19) positive Boolean function (in Boolean terminology), which
+
where F is a function of n - IC 1 variables, h is a function runs in polynomial time in n and m = J M i n S e t ( f ) ) .The
of k (> 2) variables, and Y and Z are given by (18). The time complexity was further improved to O(n2m3)in [27]. 0
next lemma shows that decomposition (17) is equivalent to As a special case of (17), assume that
A-decomposition applied to the functions representing ND
g d ( Z ) = g ’ ( Z ’ ) V z and g ( Z ) = g ’ ( Z ’ ) A z (22)
coteries.
Lemma 5.2: Let a Boolean function f ( X ) that depends where z is a variable in 2, and 2‘ denotes the
on n ( 2 3) variables be positive and self-dual. Then f has vector of variables Z from which z is deleted, and
a decomposition of the form (17) if and only if it has an g ’ ( 2 ’ ) is positive and self-dual. Then (17) becomes
A-decomposition of the form (19).
Proof: Decomposition (17) is obviously a special case of
A-decomposition. Conversely, if f has an A-decomposition, it which is in the same form as (7). Therefore, nonredundantly
can be decomposed as follows. B-decomposable ND coteries in Section IV-C are contained
f ( Y ,2)= F ( l . Z ) h ( Y ) v F ( O , Z ) h ( Y ) . in the class of ND coteries which are obtained by recursively
(20)
applying decomposition (17) to all the generated functions (as
Since f depends on Y,there is an assignment Z = A such long as they contain more than three variables).
that F ( 1 , A ) = 1 and F ( 0 , A ) = 0. (If not, there is an
assignment such that F ( 0 ,A) = 1 and F(1.A) = 0. In this VI. VOTE-ASSIGNABLE
COTERIES
case, interchange the roles of h and 6 by redefining h as h.)
Then A. Vote-Assignable Coteries and Threshold Functions
f(Y,A) = h(Y). A Boolean function f ( X ) is called a thresholdfunction [23]
if there is a set of weights, 1 1 ~ 1 . .. . , w,, and a threshold T
and h is positive since f is positive. Next, consider assign-
such that f ( X ) = 1 if and only if E, w,x, 2 T . Here,
ments Y = B1 and Y = B2 such that B1 > B P , and
all weights and thresholds are nonnegative integers. Any
h(&) = l , h ( B z ) = 0, that is f(l31.Z) = F ( 1 . Z ) and threshold function with nonnegative weights is positive. The
f ( B 2 . Z ) = F ( 0 , Z ) . Since f is positive in Y , this implies
basic majority function is a threshold function with weights
that
w 1 = 1112 = 7113 = 1 and threshold T = 2.
characterized in a very simple way i n this framework. and [ 141 M. A. Harrison, Introduction :o Switching and Automata Theory. New
many properties have been proved and/or reproved from York: McGraw-Hill, 1965.
[ 151 M. Herlihy, “Dynamic quorum adjustment for partitioned data,” ACM
these characterizations. We have also discussed decomposi- Trans. Dutabase Syst.. vol. 12, pp. 170-194, June 1987.
tions of ND coteries extensively. Shannon decomposition (1) [ I h ] T. Iharaki and T. Kameda, “A theory of coteries,” Tech. Rep. CSS/LCCR
TR90-09, Lab. of Computer and Communications Research, Simon
and Ashenhurst decomposition (19) provided theoretical basis Fraser Univ., Aug. 1990.
for these coterie decompositions. 171 T. lbaraki and T. Kameda, “A boolean theory of coteries,” in Proc.
One open problem we are currently investigating is: to 3rd IEEE SvmD. Parallel and Distributed Processinn, Dallas, T X , Dec.
1991, pp. i56-157.
describe a given ND coterie, how should one design the 181 D. Kleitman and G. Markowsky, “On Dedekind’s problem: The number
corresponding network of basic majority functions‘? Although ’ of isotone Boolean functions,;’ Trans. AMS, vol: 214, pp. 373-390,
the lemmas in Section 1V-€3 give a construction of such a 1975.
191 D. E. Knuth, The Art of Computer Programming Vol. I : Fundamental
network, it is desirable to design an optimal network (e.g., Algorithm.\. Reading, MA: Addison-Wesley, 1973.
using the minimum number of copies of the basic majority 201 A.D. KorSunov, “Solution of Dedekind’s problem on the number of
monotone Boolean functions,” Soviet Math. Dokl., vol. 18, pp. 442-445,
function). This is reminiscent of the minimization of the 1 U77
number of NAND gates in a NAND-realization of a Boolean M. Maekawa. “A flalgorithm for mutual exclusion in decentralized
function. Furthermore, the network need not be a tree; this systems,“ ACM Trans. Comput. Syst., vol. 3, pp. 145-159, May 1985.
R.H. Mohring, “Algorithmic aspects of the substitution decomposition
may help reduce the network complexity further. in optimization over relations. set systems and Boolean functions,”
Most of these design optimization problems are likely to Algorithms and Software for Optimization, edited by C.L. Monma, Ann.
be NP-hard. However, as in the case of logic design of Operat. Res., vol. 4, pp. 195.- 225, 1985.
S. Muroga, Threshold Logic and Its Applications. New York: Wiley-
computer hardware, it should be possible to develop reasonable Interscience, 1971.
heuristics to generate “good” networks. M. I-. Neilsen and M. Mizuno, “Coterie join algorithm,” IEEE Trans.
Parallel Distributed Syst., vol. 3, no. 5 , pp. 582-590, Sept. 1992.
M. Ohradovic and P. Berman. “Voting as the optimal static pessimistic
scheme for managing replicated data,” in Proc. 9th Symp. Reliable
DiJtributcd .syst., 1990, pp. 126- 135.
ACKNOWLEDGMF.NT
U. N. Peled and B. Simeone, “Polynomial-time algorithm for regular
We have ’benefited from the comments given by the mony- set-covering arid threshold synthesis,” Discrete Appl. Math., vol. 12, pp.
57--69, 1985.
mous reviewers, and from the discussions with T. Eiter of K. G. Ramamurthy, “A new algorithm to find the smallest committee
Technical llniversity of Vienna, M. Yamashita of Hiroshima containing a given set of players,” Opsearch, vol. 25, pp. 49-56, 1988.
1 3 . Shapley, “On committees,” in New Methods of Thought and
University, M. L. Neilsen and M. Mizuno of Kansas State Uni- Procedures, F. Zwycky and A.G. Wilson, Eds. New York: Springer,
versity, Y. Crama of Limburg University and H. Nagamochi 1Y67.
of Kyoto University. In particular, a comment by T. Eiter has D. Skeen, “A quorum-based commit protocol,’’ in Proc. 6th Berkeley
Workdiop Distributed Data Management and Computer Networks, Feb.
led to Theorem 4.4, and Remark 4.6 was provided by one of 1982, pp. 69-80.
the anonymous reviewers. R. H . Thomas, “A majority cnnsensus approach to concurrency control,”
ACM 7rans. Database Syst., vol. 4, pp. 180-209, June 1979.
R. 0. Winder. “Threshold logic asymptotes,” IEEE Trans. Comput., vol.
C-19, pp. 349-353, 1970.
REFERENCES S. Ydjima and 7’. Iharaki, “A lower bound o n the number of threshold
functions.“ IEEE Trans. Electron. Comput.. vol. EC-14, pp. 926-929,
[ I ] D. Agrawal and A. El Ahbadi, “Efficient solution t o the distributed 1065.
mutual exclusion problem,” in Proc. 8th ACM Symp. Principles of 1331 Yu. A. ZUKV,“Asymptotics of the logarithm of the number of threshold
Distributed Computing, Edmonton, Aug. 1989, pp. 193-200. functions of the algebra of logic,” Solwt Math. Dokl., vol. 39, pp.
[2] R. L. Ashenhurst, “The decomposition of switching functions,” in Proc. 512-513, I Y X Y .
Int. Symp. Theory of Switching, Harvard Univ., 1957, pp. 74-116.
131 D. Barbara and H. Garcia-Molina, “Mutual exclusion in partitioned
distributed systems,” Distributed Comput., vol. 1, pp. 119- 132, 1986.
[4] P. Bertolazzi and A. Sassano, “An O( rnri) time algorithm for regular
set-covering problems,” Theoret. Comput. Sci., vol. 54, pp. 237--247.
1987.
(51 I,. J. Billera, “On the composition and decomposition of clutters,” J . Toshihide Ibaraki (S’63-M’68) received the B E.,
Combinatorial Theory, vol. 11, pp. 234-245. 1971. M E , dnd Dr E degrees from Kyoto University,
[6] Z. W. Birnbaum and J. D. Esary, “Modules of coherent \ystrms,” SIAM Kyoto, Japan, in 1963, 1965, and 1970, respectively
J . Appl. Math., vol. 13, pp. 444-462. 1965.
Since 1969, he has been with the Department
[7] S. Y. Cheung, M. Ahamad, and M. H. Ammar, “Optimizing vote and
quorum assignments for reading and writing replicated data,” IEEE of Applied Mathematics and Physics, Kyoto Uni-
versity, except for two years from 1983 to 1985,
Trans. Knowledge Data Eng., to he published.
[8] -, “Multi-dimensional voting: A general method for implementing during which time he was with the Department
synchronization in distributed systems,” Tech. Rep. GlT-ICS-89/35, of Computer and Intormdtion Sciences, Toyohashi
School of Inform. Comput. Sci., Georgia Institute of Technology. 1989. University of Technology Currently he is a Pro-
191 Y. Crama, “Dualization of regular Boolean functions,” Discrete 4ppl. fessor of Applied Mathematics and Physics. He
Math., vol. 16, pp. 79-85, 1987. hds held a number of visiting appointments with
[ l o ] T. Eiter and G. Gottlob, “Identifying the minimal trdnsversals of a University of Illinois, University uf Waterloo, Simon Fraser University, and
hypergraph and related problems,” ’Tech. Rep. CU-TR9 1 16, Technischt:
I
Rutgers University His research interest includes conihtnatorial optimization,
Universitat Wien, 1991. computational complexity, datahdse systems, dnd dlgorithms.
[ 11J A. Fu, “Enhancing concurrency and availability for database systems.” Dr Ihdraki IS a member of the Association for Computing Machinery,
Ph.D. dissertation. Simon Fraser Univ., 1990. the Mathemdtical Programming Society, the Operations Research Society of
[I21 H. Garcia-Molina and D. Barbara, “How tu assign votes in a distributed Jdpan, the Institute of Electronics, Informdtion and Communication Engineers
system,” J . ACM, vol. 32, pp. 841-860. Oct. 1985. of Japdn, the Information Processing Society of Japan, the Japan Association
[I31 H. Gifford, “Weighted voting for replicated data,“ in pro^,. 7rh 4CM for Automatic Control Engineers, and the Jdpdnese Society for Artificial
Symp. Operat. Syst., Dec. 1979, pp. 150.- 162. Intelligence
794 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 4, NO. 7, JULY 1993