[Lecture Notes in Computer Science] Foundations of Computation Theory Volume 158 || A tight...
Transcript of [Lecture Notes in Computer Science] Foundations of Computation Theory Volume 158 || A tight...
A TIGHT ~(!o~log n)-BOUND ON THE TIME FOR PARALLEL,,RAM'S
TO COMPUTE NONDEGENERATED BOOLEAN FUNCTIONS
Hans-Ulrich Simon
Institut f~r angewandte Mathematik und Informatik
der Universit~t des Saarlandes
D-66OO Saarbr~cken
W.-Germany
Abstract:
A function f:{0,1} n ~ {0,1} is said to depend on dimension i iff there
exists an input vector x such that f(x) differs from f(xi), where x i
agrees with x in every dimension except i. In this case x is said to be
critical for f with respect to i. f is called nondeqenerated iff it de-
pends on all n dimensions.
The main result of this paper is that for each nondegenerated function
f: {O,I} n ~ {0,7} there exists an input vector x which is critical with
respect to at least ~(log n) dimensions. A function achieving this
bound is presented.
Together with earlier results from Cook,Dwork [2] and Reischuk [3] we
can conclude that a parallel RAM requires at least Q(loglog n) steps to
compute f.
1. Notations and Main Theorem
Let us define a PRAM (= Parallel RAM) to consist of a collection of
processors which compute synchronously in parallel and which communi-
cate with a common global random access memory. At each step each pro-
cessor can read from one global memory cell, do some computing and
write into one global memory cell. Any number of processors can read a
given global memory cell at once, but we allow at most one processor to
attempt to write into a given memory cell in one step. At the beginning
of the computation of a function f(xl,...,x n) the values Xl,...,x n are
stored in the global memory cells CI,...,C n. At the end of the computa-
tion f(x I .... ,x n) has to be stored in C I (compare the definitions in
[I],[2] and [3]).
Let B = {0,1}. For each Boolean function f:B n ~ B and for each input
vector x let c(f,x) denote the number of dimensions i such that x is
440
critical for f with respect to i. Let c(f) := max {c(f,x) I x£Bn}. In
[2] and [3] it is shown that a PRAM requires at least Q(log(c(f)))
steps to compute f.
F n denotes the set of nondegenerated Boolean functions of n variables.
Let Cn:= min {c(f) I f6Fn}.
Theorem: Let n~2.
I I I < log(n) + 2 log(n) - ~ loglog(n) + ~ < c n
Corollary:
Let f6F n. Then a PRAM requires at least ~(loglog n) steps to com-
pute f.
2. Proof of the Theorem
The upper bound for c n is shown by the following
Example I : n
For x=(x I .... ,x n) 6 B n let v(x):= ~ x.2 n-i. For n=m+2 m let i=I 1
• "" ) := Yv(x) fn(Xl, ..,Xm,Y O, .,Y2m i
fn has the following properties:
(i) f 6 F n n
(ii) C(fn ) = m+1
(iii) A PRAM (with m processors only) can compute fn in ~ log m] + 2
steps.
To prove the lower bound for c n some additional notations are required.
Let G = (V,E) be an undirected graph. The minimum degre ~ md(G) is de-
fined by md(G):= min {degree(v) I v6V}. Any not empty,finite sequence
p=(v I .... ,v r) of vertices such that
Vi£[1:r-1]: {vi,vi+ I} E E
is called a path in G. p is called a cycle iff in addition v1=v r. For
each function f:E ~ B, labeling the edges of G by O. or I, the weight
Wf(p) of p with respect to f is defined by
r-1 Wf(p) := ~ f({vi,vi+1}) , where E denotes the integer sum.
i=I
441
Let E n := {{x,y}l x,y6B n and x,y differ in exactly one dimension} .
The undirected graph Cn:=(Bn,En) is called the n-dimensional cube. For
every function f: B n ~ B let Af denote a function from E n to B defined
by F = ~ I if f(x)~f(y)
Af ({x,y})
L O if f(x)=f(y)
Example I (continued):
= I YO if x1=O f3 (xl,Yo,Yl) [ Yl if xi=I
and Af3, regarded as functions which label the vertices and the edges
of C3, can be drawn as
fO" 1
0 0
0
t
0 0
I I 0
0
The following Lemmata are obvious but important.
Lemma I:
Let G=(V,E) be a not empty,partial subgraph of C n, i.e. ~ ~ V c B n,
E c E . Then IVI ~ 2 md(G)" " n
Lemma 2 :
Let f be a function from B n to B and let c be a cycle in C . Then the n
weight of c with respect to Af is even.
Let f6F n be arbitrary but fixed. From now on we regard the vertices and
the edges of C n as labeled by f and ~f. Let c:=c(f).
442
Observation:
For each vertex x of C n
on X.
there are c(f,x)~c edges labeled I incident
For every vertex x6B n let x~ denote the i'th component of x. Let
i611:n] an arbitrary but fixed dimension. Since f is nondegenerated
there exists an edge {x,x i} labeled 1. W.l.o.g.x.=O and x~=l.- For 1 1
9=O,1 let V9:= {y6Bnl Yi =9}" CJn denotes the (n-1)-dimensional subcube
induced by V.. We define subsets U.cV. in the following way: 3 3 ]
a) yEU O i f f t h e r e e x i s t p a t h s p j i n CJn s u c h t h a t
( i ) PO s t a r t s i n x a n d e n d s i n y .
i i ( i i ) P l s t a r t s i n x a n d e n d s i n g .
( i i i ) w A f ( p o ) = w ~ f ( p l ) = o .
b) y 6 u 1 i f f y l 6 U o .
G. d e n o t e s t h e s u b g r a p h o f C 9 i n d u c e d by U . . T h i s c o n s t r u c t i o n a n d t h e 3 n ]
following claim are illustrated in Figure I.
Figure I:
C ° n
J I Jn-2c+1 Y Y
.........
X
±3n-2c+I
0 " i " 0 ¥1
x
iJ 1
C I n
Claim:
I) x£U O and xiEU1
2) Vy6Uo: {y,yi} is labeled I .
3) md(Go) = md(G1) ~ n-2c+1
Proof of the claim:
I) The paths pO=(X) and p1=(x i) fulfil the conditions (i), (ii) and
(iii) above.
2) Let YEUo, yi6uI ' Po a path weighted O from x to y and Pl a path
443
weighted 0 from x i to yi. Let us assume for the sake of contradic-
tion that {y,yi} is labeled O. Then
po- (y,yi)-p~1" (xi,x) ,
where "." denotes the composition and " -I " the reversal of paths,
is a cycle weighted I in contradiction to Lemma 2. Thus {y,yi} is
labeled I.
3) Let y6U 0 and Di:= [1:n]~{i}. From the observation above and from the
fact that {y,y~} is labeled I it follows:
(I) There are at most c-I dimensions j£D• such that {y,yJ} is la- l
beled I. There are at most c-I dimensions j6D i such that
{yl,y 13} is labeled I, where y 13 agrees with y in every dimen-
sion except i and j.
(2) There are at most 2(c-1) dimensions j£D. such that {y,yJ} is l
labeled I or {yl,y 13} is labeled I.
(3) There are at least n-1-2(c-1)=n-2c+1 dimensions j£D. such that l
{y,yJ} is labeled 0 and {yi,yij} is labeled O.
Thus md(G O) = md(G I) Z n-2c+I
By the claim and Lemma I: iUoi = IUlJ Z 2 n-2c+I . Thus there are at
least 2 n-2c+I edges labeled I in dimension i. Summing over all dimen-
sions we observe that at least n2 n-2c+I edges are labeled I. On the
other hand this number cannot exceed c2 n-1 . Setting
n2 n-2c+I ~ c2 n-1
a straightforward computation shows that
I I I c > ~ log(n) - ~ loglog(n) + ~ ,
provided that n~2. This proves the Theorem.
3. Conclusions
What are the corresponding results in the case of nondegenerated func-
tions f: S1x...xS n ~ S, where SI,...,S n are arbitrary finite sets with
at least two elements ? The following example shows that there exist
nondegenerated functions which are computable by a single processor in
constant time.
Example 2:
For every n6rN let f~ be the function from [1:n]xB n to B given by n
fn(i,xl ..... x n) := x i
With a straightforward modification of this example one can show that
for every s>O there exists a family fs of functions of n variables n
such that Xlr...,x n
For every i611:n] at most O(n s) values can be assigned to x i- (I)
(2) A single processor can compute fa in constant time. n
On the other hand one can show by the main theorem of this paper and an
easy binary coding argument the
Corollary:
Let f be a family of nondegenerated functions of n variables n
xl,...,x n such that at most O(log n) values can be assigned to x i
every i611:n]. Then a PRAM requires at least Q(loglog n) steps to
compute fn"
for
Acknowledgements:
I am grateful to R~diger Reischuk for many helpful discussions and for
pointing me to the family fn in example I, and to Helmut Alt and Kurt
Mehlhorn for suggestions to simplify the original proof.
References:
[I] A.Borodin,J.Hopcroft. Routing and Merging on Parallel Models of
Computation. Proc. 14'th annual ACM, 5/1982. pp.338-344.
[2] S.Cook,C.Dwork. Bounds on the Time for Parallel RAM's to Compute
Simple Functions. Proc. 14'th annual ACM, 5/1982. pp.231-233.
[3] R.Reischuk. A Lower Time Bound for Parallel RAM's without Simul-
taneous Writes. IBM Research Report RJ3431 (40917), 3/1982.