Rudimentary reductions revisited
y
Vivek Gore Eric Allender
Department of Computer Science Department of Computer Science
Rutgers University Rutgers University
New Brunswick, NJ 08903 New Brunswick, NJ 08903
April 23, 1997
Abstract: We show that log-b ounded rudimentary reductions de ned and studied by Jones in
0
1975 characterize Dlogtime-uniform AC .
1 Intro duction
In the rst part of the oft-cited pap er [Jon75], Jones intro duced logspace reductions as a to ol for study-
ing the relative complexity of problems in P. In the second part of [Jon75], Jones intro duced a re-
stricted version of logspace reducibility, called log-b ounded rudimentary reductions. The motivation
for intro ducing this restriction came from 1 the desire to have a to ol for talking ab out the structure
of very small complexity classes such as DSPACElog n, and 2 interest in generalizing the notion
of \rudimentary relations," which at that time was the ob ject of a considerable amount of attention
[Smu61, Sal73,Wra78, Wil79, PD80]. In [Jon75], Jones went on to showthatanumb er of problems
were complete for various complexity classes under log-b ounded rudimentary reductions. Although
logspace reducibility has proved to b e a very useful notion in complexity theory in the intervening years,
log-b ounded rudimentary reducibility has b een mentioned explicitly only seldom in subsequentwork {
although, as we demonstrate, it has b een considered implicitly many times, under di erent names.
A great deal of insight ab out the complexityofvarious problems has b een gained by the study of
uniform circuit complexity; many complexity classes have b een characterized in terms of the compu-
tational p ower of families of circuits fC : n 2 Ng where C takes inputs of length n, byvarying
n n
the typ es of gates and allowable fan-in on the circuits, and imp osing a \uniformity condition" which
sp eci es that the de nition of C in terms of n b e \easy" in some sense. Unfortunately, the question of
n
Supp orted in part by NSF grant CCR-9000045.
y
Supp orted in part by a DIMACS research assistantship.
DIMACS is a co op erative pro ject of Rutgers University, Princeton University,AT&T Bell Lab oratories and Bellcore.
DIMACS is an NSF Science and Technology Center, funded under contract STC-88-09648; and also receives supp ort from
the New Jersey Commission on Science and Technology. 1
which uniformity condition to use has remained somewhat controversial. For example, although much
work on uniform circuit complexity uses logspace uniformity, it is not clear that this is appropriate when
0
de ning sub classes of DSPACElog n, suchasAC , the class of languages accepted by p olynomial-size,
constant-depth circuits of unb ounded fan-in AND and OR gates.
0
Comp eting de nitions of \uniform AC "were prop osed by Immerman [Imm87, Imm89] and Buss
[Bus87]. Then it was shown by Barrington, Immerman, and Straubing, that these de nitions in fact
0
coincide [BIS90]; they show that Dlogtime-uniform AC may b e de ned alternatively in terms of rst-
order logic, O 1 time on a CRAM, inductive de nitions with O 1 inductive depth, and Sipser's log-time
0
hierarchy [Sip83]. Other characterizations of Dlogtime-uniform AC are found in [Clo90]. Since some of
these notions were de ned entirely indep endently of each other and indep endently of considerations of
uniform circuit complexity, the fact that these notions coincide is taken as evidence for the \correctness"
0
of the Dlogtime-uniformity condition when cho osing a de nition of uniform AC .
In this pap er, we add to this b o dy of evidence by showing that Jones' logspace rudimentary reductions
0
provide yet another characterization of Dlogtime-uniform AC . Throughout the rest of this pap er,
0 0
all references to AC will denote Dlogtime-uniformAC .
2 Preliminaries
De nition1 The following de nitions are due to Jones [Jon75].
th
Let b e an alphab et. The ternary predicate x; i; a is true i a is the i symbol of x where
x 2 .
For a p ositiveinteger c and a space b ound S , de ne the predicates P , P as follows:
c;S c;S
P x; u; v ; w , uv 6= w ^juv w jc S jxj. P x; u; v ; w , uv = w ^juv w j c S jxj,
c;S c;S
The class of S -b ounded rudimentary predicates, RUD is the class of predicates built from
S
P for every c via logical connectives, explicit transformations, and S -b ounded ; :;P , and
c;S c;S
existential and universal quanti cation. See [Jon75] for a complete de nition.
De nition2 A function f : ! is S -b ounded rudimentary i the predicate R de ned by
Rx; i; a , f x;i;aisinRUD .
S
th
That is, f is S -b ounded rudimentary i the predicate \a is the i symbol of f x" is in RUD . The
S
cS jxj
fact that jij c S jxj for some c and all x implies that jf xj2 .
Note that, as de ned ab ove, RUD is a set of predicates;however we can just as easily view RUD as
S S
a set of languages. A language L will b e said to b e in RUD i the characteristic function of L is S -
S
b ounded rudimentary. Alternatively, each predicate Q 2 RUD over alphab et de nes a language L
S 2
over alphab et [f; g where \," is any symb ol not in de ned by L = fx; y ;:::;y : Qx; y ;:::;y g.
1 m 1 m
These alternative de nitions are easily seen to b e equivalent. For more formal arguments along these
lines, see [Wra78].
In this pap er as in [Jon75] we are interested primarily in RUD , the class obtained when the space
log
0
b ound S is logarithmic. In Section 3, we show that RUD =AC .
log
The pro ofs in this pap er will not make use of the circuit formalism;thus detailed de nitions of the
Dlogtime-uniformity conditions are omitted here. They may b e found in [BIS90, BCGR90]. Instead,
0
we will use twocharacterizations of AC that were shown to b e equivalent in [BIS90] see also [BCGR90],
one in terms of alternating Turing machines, and one in terms of rst-order logic.
We use the mo del of alternating Turing machine used by Ruzzo [Ruz81] in which access to the input
is provided via a sp ecial tap e whichwe call the input address tape onto which an address i maybe
th
written in binary, following which in unit time the i input symbol is available. This convention is
necessary to allow machines with sublinear running times to have access to all of the input. Consult
[BIS90, BCGR90] for examples illustrating how these machines compute.
Now let timelog n denote the class of languages accepted by alternating Turing machines running
k
in O log n time, b eginning in an existential con guration, and making at most k 1 alternations b etween
existential and universal con gurations on any computation path. Let LH the logtime hierarchyof
S
[Sip83] denote timelog n.
k
k
Immerman [Imm87, Imm89] has studied the expressivepower of rst-order logic in de ning languages;
he calls the resulting class FO. Due to space limitations, we refer the reader to [Imm87, Imm89] for the
relevant de nitions.
0
Theorem 3 [BIS90] AC =LH=FO.
Two minor complications arise from the statement of Theorem 3:
0
1. Alternating Turing machines can accept input over any alphab et , whereas FO and AC are
classes of languages over the alphab et f0; 1g.Thus we simply assume a binary enco ding of any
other alphab et. Alternatively, the FO and circuit frameworks can b e mo di ed to allow arbitrary
alphab ets.
0
2. It is clear what it means for a function to b e computed byanAC circuit, but it is less clear how
a function can b e sp eci ed in the LH or FO settings. To resolve this diculty,we use a standard
1
technique that has b een used since at least [Jon75]. For a function f with p olynomial growth rate ,
def
th
let A c; i; z = the i symbol of f z isc. Then wesay that f is in LH or FOifA is in LH
f f
1
That is, there is some p olynomial p such that for all x; jf xj pjxj. 3
0
or FO. It has b een noticed by many authors e.g., [BIS90] that f is computed byAC circuits
0 2
i A 2 AC .
f
0 0
Note that the comp osition of two functions in AC is also a function in AC .
0
3 Equivalence of AC and RUD
log
0
Theorem 4 AC =RUD .
log
Pro of.
This containment is mentioned in Prop osition 5 of [Vol84] where the converse containmentis
left as an op en question. We present a complete pro of here.
The pro of pro ceeds by induction on the de nition of RUD , showing that each language in RUD
log log
is in LH. The input to an ATM is assumed to have b oth ends marked by the sp ecial symbol $.
Consider the inductive de nition of RUD :
log
1. x; i; a and : x; i; a are in RUD .
log
Let L = fx; i; a : x; i; ag. Consider an alternating TM M for L:
On input $x; i; a$,
M existentially guesses the p ositions of all the commas there are only a constantnumber of
them.
M universally veri es that that the guessed p ositions contain the appropriate markers and
the other p ositions contain 0's or 1's.
Using the marker p ositions, M calculates jxj and jij.Itwas observed in [Bus87] that this can
b e done in logarithmic time. M calculates logjxj from jxj and veri es that jijlogjxj.
M existentially guesses the bits of i and writes them on a worktap e, and then universally
checks that for each j i, the guessed value for p osition jxj +1+ j is correct.
th
M again uses marker p ositions to read a, and check that the i bit of x is a.
2. If c is a p ositiveinteger then b oth P x; u; v ; w and P x; u; v ; w are in RUD where
c;log c;log log
P x; u; v ; w , uv = w ^juv w j c logjxj,
c;log
P x; u; v ; w , uv 6= w ^juv w j c logjxj.
c;log
2
Again, for equivalence to hold, the output of the circuit must b e interpreted in sucha way as to allow an \endmarker,"
0
or else all functions f in AC must have jf xj = jf y j if jxj = jy j, which is clearly not the case for some functions in LH
and FO. 4
Consider an ATM M that b ehaves as follows:
On input $x; u; v ; w $,
M rst guesses and veri es the p ositions of all the markers as in the previous case, and uses
this to compute c log jxj.
Using the marker p ositions, determine whether juv w j c logjxj, and if so, read eachof
u; v ; w from the input, as in the previous case, and verify whether uv = w or uv 6= w for
P .
c;log
3. If Qx; y ;:::;y and Rx; y ;:::;y are in RUD , then so are
1 m 1 m log
Qx; y ;:::;y ^ Rx; y ;:::;y and Qx; y ;:::;y _ Rx; y ;:::;y .
1 m 1 m 1 m 1 m
This case is trivial, using universal and existential branching, resp ectively.
4. If Qx; z ;:::;z isinRUD and eachof ;:::; is either a string in or one of the variables
1 m log 1 m
y ;:::;y , then the predicate R de ned by
1 r
Rx; y ;:::;y , Qx; ;:::; isinRUD .
1 r 1 m log
S
Inductively, Q can b e accepted by an alternating TM M in Timelog n. Note that r and m
k
k
are constants here and ;:::; dep end only on y ;:::;y and not on x. Therefore, the conversion
1 m 1 r
of y ;:::;y into ;:::; only requires a nite amount of information that can b e supplied to
1 r 1 m
an ATM in its nite control.
Consider a function f that converts strings of the form x; y ;:::;y into x; ;:::; .Ifwe can
1 r 1 m
show that f can b e computed in LH, then we can use the fact that LH functions are closed under
comp osition to conclude the pro of for this case, using the inductivehyp othesis. That is, we need
th
to show that the language A = fciz : the i symbol of f z iscg is in LH. Note that, for
f
well-formed inputs, z will b e of the form x; y ;:::;y .
1 r
Consider an ATM M that b ehaves as follows:
On input $ciz $
M guesses all the marker p ositions in $ciz $aswell as in z itself there are a constant
numb er of these and veri es them as in the previous cases.
M uses the marker p ositions to compute jxj and jij, and rejects if jij is to o long. Otherwise,
as in the previous cases, M guesses the contents of the c, i, and y ;:::;y elds of the input,
1 r
and checks that the guesses are correct. 5
th
If i jxj, accept i the i bit of x is c. Otherwise, compute ;:::; and accept i c is in
1 m
p osition i jxj of ;:::; . It is easy to verify that there is ample time to p erform these
1 m
op erations.
0
5. If c is a p ositiveinteger, and Qx; y ;:::;y isinRUD , then the predicates R and R de ned
1 m+1 log
as follows are also in RUD
log
Rx; y ;:::;y , 9z [jz jc logjxj ^ Qx; y ;:::;y ;z]
1 m 1 m
0
R x; y ;:::;y , 8z [jz j c logjxj Qx; y ;:::;y ;z]
1 m 1 m
0 0
Inductively, Q is accepted by some LH machine M .We construct a machine M accepting R .On
0
input $x; y ;:::;y $;M starts o by getting the marker p ositions as in the previous cases. Then
1 m
it universally veri es that for all z with jz jc logjxj, M accepts $x; y ;:::;y ;z$. To see that
1 m
this is p ossible, note that M consults its input only at the end of a computation branch. As so on
0 th
as M go es into input mo de and has j written on its index tap e, M either consults the j symbol
of its input, or reads the symb ol in p osition j jx; y ;:::;y j of z .
1 m
A similar argument, using existential branching, shows that R is in LH.
To prove this inclusion, let L 2 FO. Therefore, there exists a sentence in Prenex Normal Form
=Q j Q j :::Q j j ;:::;j
1 1 2 2 m m 1 m
such that 8x
x 2 L , x j=Q j Q j :::Q j j ;:::;j
1 1 2 2 m m 1 m
where each Q is a quanti er 9 or 8.
i
We will show that for each such and its asso ciated there is a RUD predicate R such that
log
for each string x,
x j=Q j Q j :::Q j j ;:::;j
1 1 2 2 m m 1 m
if and only if
0 0 0
Q y Q y :::Q y R x; y ;:::;y
1 2 m 1 m
1 2 m
0
where each Q ranges over all binary strings of length at most d log jxje+ 1 and is the same typ e existential
i
or universal as Q . By the closure prop erties of RUD , and by the fact that the length of each y
i log i
is b ounded, it follows that since R is in RUD , L is also in RUD , and the pro of is complete. It
log log
remains only to construct R .
The formula is built up from the primitive predicates =;<, X , and BI T , and from the constants
1 and n. Clearly, it will suce to show that each of these primitive predicates is \equivalent" in the 6
same sense to a relation in RUD .For example, we will show that there is a relation S in RUD such
log log
that, for anynumb ers j and j , the primitive predicate j 1 2 1 2 1 2 where y and y are binary strings of length at most d log jxje + 1 representing the numb ers j and j , 1 2 1 2 resp ectively. The constant 1 is explicitly de nable in RUD , and Jones [Jon75] has already observed that the log v relations u + v = w; u v = w; u = w and juj = w are all in RUD , so long as u; v ; w are constrained log to have length c logjxj for some c.Thus the constant n is de nable, since the relation jxj = j is equivalentto x; j; 0 _ x; j; 1 ^: x; j +1; 0 _ x; j +1; 1. It follows easily that the following set is in RUD : fx; y ;:::;y : for each i, jy j = d log jxje +1g. Therefore for the rest of this pro of, log 1 m i we will assume that all variables y refer to strings of exactly this length. This will result in a few i simpli cations. Nowwe consider each of the primitive predicates in turn. j = j $ 8i[ y ;i;0 , y ;i;0] 1 2 1 2 j 1 2 1;log 1 1;log 2 This says that the binary representation of j is uv and the binary representation j is uw such 1 2 that the most signi cant bits of v and w are 0 and 1 resp ectively. Note that the values taken by each of the variables u; v ; w have lengths logjxj. X j $ x; y ; 1. 1 1 BI T j ;j $ y ;y ; 1. 1 2 2 1 This completes the pro of. We remark that standard translational techniques maynow b e used to prove the following general- ization of Theorem 4. Note that for the case of S n=n, this is the well-known theorem of Wrathall [Wra78]. Corollary 5 For functions S such that the binary representation of S n can b e computed from n in S time O S n, RUD = timeS n. S k k 4 Reducibilit ies Jones de ned RUD as a rst step towards de ning a very restrictive notion of reducibility. Jones to ok log the familiar notion of a many-one reduction namely, f reduces A to B if x 2 A , f x 2 B , and imp osed the restriction that f b e log-b ounded rudimentary. There was one additional restriction that rud Jones imp osed in de ning his logspace-rudimentary reducibility denoted ; he required that there log 7 b e some constant c such that for all x, log jf xj = c logjxj. That is, the length of f x should b e equal to a p olynomial in jxj, and not just b ounded by one. In [Jon75], Jones comments that this nal restriction is stronger than the more natural restriction of simply having f have p olynomial growth rate, rud but explains that it \seems to b e necessary" in order to have the relation b e transitive i.e., in order log for the comp osition of two log-b ounded rudimentary functions to b e log-b ounded rudimentary. However by Theorem 4 we see that for functions f with p olynomial growth rate, f is log-b ounded rudimentary i f is in LH. Since the comp osition of two functions in LH is easily seen to b e in LH, it follows that Jones' rationale for imp osing the stronger length requirementwas unfounded. 5 Conclusion Wehave considered the log-b ounded rudimentary predicates de ned by Jones in [Jon75] and wehave 0 shown that they corresp ond exactly to the class of sets accepted by Dlogtime-uniform AC circuits. Thus Jones was probably the rst to study this complexity class, which has lo omed large in imp ortance in recentyears. We b elieve that this augments the already comp elling arguments of [BIS90] in favor of using the Dlogtime-uniformity condition when studying small circuit complexity classes. It is instructive to note some of the op en questions p osed by Jones in [Jon75]. He mentions on more than one o ccasion the question of whether or not RUD = DSPACES n, and he explicitly considers S the p ossibility that RUD = NSPACElog n [Jon75], Theorem 23. However, the results of [Ajt83] log 0 and [FSS84] refute this p ossibilityby showing that PARITY is in NSPACElog n AC . The strongest results along this line are those of [Has87] and [Yao85], showing that for any function S such that for all , S n=on , the PARITY language is in DSPACE1 RUD . Here, DSPACE1 denotes the class S of regular sets. However, any generalization of this result for larger functions S will require entirely di erent techniques and will havetotake uniformityinto account in some way, since the techniques of [Nep70] see also [Vol84] show that for each >0, NSPACElog n RUD , and it follows from n [Wra78] that RUD contains complete sets for each level of the p olynomial hierarchy.Thus the question n of whether or not RUD = DSPACEn remains an interesting op en problem. Note, however, that n an armative answer would imply that the p olynomial hierarchy is equal to PSPACE. References P 1 [Ajt83] M. Ajtai. -formulae on nite structures. Annals of Pure and AppliedLogic 24 1983 1 1{48. 1 [BIS90] D. A. Mix Barrington, N. Immerman, and H. Straubing. On uniformity within NC . Journal of Computer and System Sciences 41 1990 274{306. [Bus87] S. R. Buss. The Bo olean formula value problem is in ALOGTIME. In Proc. 19th ACM Symposium on Theory of Computing, 1987, pp. 123{131. 8 [BCGR90] S. Buss, S. Co ok, A. Gupta, and V. Ramachandran. An optimal parallel algorithm for formula evaluation. Submitted, 1990. [Clo90] P. Clote. Bounded arithmetic and computational complexity. In Proc. 5th Structurein Complexity Theory Conference, 1990, pp. 186{199. [FSS84] M. Furst, J. Saxe, and M. Sipser. Parity, circuits and the p olynomial time hierarchy. Math- ematical Systems Theory 17 1984 13{27. [Has87] J. Hastad. Computational limitations for small depth circuits . MIT Press, Cambridge, 1987. [Imm87] N. Immerman. Expressibility as a complexity measure: Results and directions. In Proc. 2nd Structure in Complexity Theory Conference, 1987, pp. 194{202. [Imm89] N. Immerman. Expressibility and parallel complexity. SIAM Journal on Computing 18 1989 625{638. [Jon75] N. D. Jones. Space-b ounded reducibility among combinatorial problems. Journal of Com- puter and System Sciences 11 1975 68{85. Corrigendum: Journal of Computer and System Sciences 15 1977 241. [Nep70] V.A. Nep omnjasci . Rudimentary predicates and Turing calculations. Soviet Math. Dokl. 11 1970 1462{1465. [PD80] J.B. Paris and C. Dimitracop oulos. Truth de nitions for formulae. In Logic and Algo- 0 rithmic: an International Symposium Held in Honour of Ernst Specker, Monographie no. 30 de l'Enseignement Math ematique, 1980, pp. 317{329. [Ruz81] W.L. Ruzzo. On uniform circuit complexity. Journal of Computer and System Sciences 21 1981 365{383. [Sal73] A. Salomaa. Formal Languages . Academic Press, 1973. [Sip83] M. Sipser. Borel sets and circuit complexity.InProc. 15th ACM Symposium on Theory of Computing, 1983, pp. 61{69. [Smu61] R. Smullyan. Theory of formal systems. In Annals of Math. Studies 47. Princeton University Press, 1961. [Vol84] H. Volger. The role of rudimentary relations in complexity theory. Bul letin de la Societ e Math ematique de France, series 2, numb er 16 1984 pp. 41{51. [Wil79] A. Wilkie. Applications of complexity theory to -de nability problems in arithmetic. 0 Lecture Notes in Mathematics vol. 834, 1979, pp. 363{369. [Wra78] C. Wrathall. Rudimentary predicates and relative computation. SIAM Journal on Computing 7 1978 194{209. [Yao85] A. Yao. Separating the p olynomial-time hierarchyby oracles. In Proc. 26th IEEE Symposium on Foundations of Computer Science, 1985, pp. 1{10. 9