A Clari cation Concerning the L Hierarchy
Eric Allender
Department of Computer Science, Rutgers University
P.O. Box 1179, Piscataway, NJ 08855-117 9,USA
e-mail: [email protected]
Octob er 29, 1997
Abstract
1 0
In [AO96], it is stated without proof that if NC L = AC L, then the
L hierarchy col lapses to some level. This note provides a proof of that claim.
The reader is referredto[AO96] for al l background, de nitions, and motivation.
Al l of the circuit complexity classes mentioned in this note arelogtime-uniform
[BIS90 ], unless otherwise speci ed.
1 The Pro of
The pro of consists of establishing the following fact.
1
Fact 1.1 The set of languages NC DET has a complete set under logspace
many-one reducibility.
This fact is sucient to establish the claim made in [AO96]. To see this,
1 0
assume that NC DET = AC DET. By Claim 1.1, there is a complete set
:L
:
L
0 0 L
, this complete set is in for AC DET. Since AC DET is equal to L
0
some xed level of this hierarchy, and thus AC DET collapses to this level.
It remains for us to establish Fact 1.1.
Supp orted in part by NSF grant CCR-9509603. 1
First, let's come up with a de nition of a \canonical" way for a log-time
1 2
machine to sp ecify a circuit. Let M b e a clo cked 3-tap e Turing machine
running for c log n time for some c.
Let us say that M is k-good on length n if for all p of length k log n, M n; p
is a string in the set
fb; ORACLE; 2; AND; 2; OR; 1; NOT; i; INPUTg
where jbj + jpj k log n, and i n, and such that if jpj = k log n, then M n; p=
i; INPUT for some i, and for each pre x q of p,ifM n; q = j; INPUT, then
i = j . Intuitively, p is an enco ding of a path from the output gate to a gate
1
g in the NC circuit, and M , on input n; p, is pro ducing as output the fan-in
of the g , and the typ e of gate that g is. The other conditions are merely for
technical convenience. The depth of the circuit is k log n.
Note that { for a suitable enco ding of clo cked Turing machines { the following
0
language is in uniform AC , for each c and k : fM; n : M is a clo cked Turing
machine running in c log n time, and M is k -go o d on length ng. This do es
dep end somewhat on the enco ding of Turing machines. However, note that
M n p i
for any reasonable enco ding of Turings machines M , the language f1 0 1 0 b
0
: the i-th output bit of M n; pisbg is in Dlogtime-uniform AC since it is
in Dlogtime. Checking if a circuit is 2-go o d can b e expressed as a rst-order
0 0
sentence over a uniform AC predicate, and thus it is in Dlogtime-uniform AC .
De ne the circuit C as follows: If M is a clo cked Turing machine
M ;c;k ;n
rnning in time c log n and M is k -go o d on length n, then this is the circuit
with gates having lab els of the form p, where the typ e and fan-in of gate p is
given by M n; p. If gate p has fan-in b, then for all strings q of length jbj that
lexicographically precede b, the gates that are input to p are the gates pq . If
M is not a clo cked Turing machine that is k -go o d on length n, then the circuit
is a circuit that trivially accepts the empty set.
I claim that the set C = fM; x : M is a clo cked Turing machine running in
5 log n time, and M is 2-go o d on length jxj, and the circuit C accepts
M ;c;k ;jxj
3
x where the oracle gates give the middle bit of the function DET applied to
1
their inputs, and jM j log log jxjg is complete for NC L.
1
We need to showthatCisinNC L, and that it is hard. Neither seems
completely trivial.
1
A clo cked TM running in time c log n is a machine that, on input n; p
1. computes c log n which is just c times jnj.
2. starts a counter that will allow it to execute only c log n steps.
Actually, steps 1 and 2 can b e b egun simultaneously; there are a numb er of programming
tricks with Turing machines that one can use. The p oint is, a there is some purely syntactic
part of the Turing machine description that we will call the \clo ck", b this \clo ck" enforces
a run-time on the Turing machine, and c every Turing machine is equivalent to a \clo cked"
Turing machine of comparable complexity.
2 0 1
Note that Dlogtime-uniform AC and NC have circuits that are Dlogtime-uniform even
with this additional restriction that the uniformity machine have three tap es [BIS90].
3
Any bit of the DETERMINANT can b e reduced to the middle bit. If I want the lower-
k
n
order bit of fx, this is the middle-bit of some function GapL function gx de ned as f x2
for some k . 2
1
First, let's show that it's in NC L. Let m b e xed. We'll describ e the
circuit accepting C for inputs of length m. First, given input M; x, where
0
jxj = n, the circuit will evaluate the AC predicate to check that M runs for
5 log n time and is 2-go o d on length jxj;thus let's assume that M is go o d, and
let's concentrate on length n.For each string p, there will b e circuitry evaluating
M n; p. The output of the circuit our circuit accepting C , of course, is the value
of gate in circuit C on input x. Recall that the empty string is the
M ;c;k ;n
name of the output gate of C . Here is the circuitry that will evaluate
M ;c;k ;n
0 0
any given gate p. With NC circuitry, with \free" calls to the AC predicates
that are computed only once, given M n; p we can compute the \typ e" of gate
p. If the gate is of typ e INPUT, then a sub circuit of depth log n can compute
the input bit i to which gate p is connected. If the gate is of any other typ e,
then a sub circuit of depth log b can compute the fan-in b of gate p. That is,
near the \top" of this circuit, there will b e gates checking if the fan-in is 2, and
attempting to compute the AND of the inputs of gate p; near the b ottom of
4
the circuit, there will b e gates checking if the fan-in is n , and attempting to
4
compute DETthe values of the input gates pq , etc. The total circuit depth
required to evaluate a gate of fan-in d is O log d+ depth of its inputs. This is
1
all that's required in order to show that this is in NC L.
Now, let's show hardness.
1
For this, we need to show that anything accepted byNC L circuits is
accepted by circuits of the form C for some k -good Turing machine M ,
M ;c;k ;n
for some c and k . If wehave this, then a standard \padding" reduction will
show that our set C is complete. What we need is that a log-time machine,
given a path p, can compute the typ e and fan-in of the gate that is reached by
following path p from the output gate. By \following a path p", I mean that
at each oracle gate of fan-in b, log b bits are used to determine the input wire
from the oracle gate that is followed.
1
Let A b e accepted by logspace-uniform NC L circuits C . Note that there
n
1
is a very uniform NC L family of circuits recognizing the language fn; p; t; i
: g is the gate reached by following path p in C , and either i is 0 and the gate
n
has typ e t,ori logb and t is the i-th bit of the binary representation of the
fan-in of gate g .g That is, the L oracle can b e used to determine the name
4
Actually sp ecifying the lab eling M uses will b e a bit messy. The circuit has a very regular
structure:
Use OR gates to guess the typ e of the gate g .
Use ANDs to
{ check that the guess is correct, and
{ simulate the gate.
The circuits of typ e 1 are all similar. To do part 2, we a Use OR gates to guess
the next bit of the fan-in b of g b use AND gates to check that this bit is correct
Once wehave the entire fan-in, wehave a gate that actually simulates the gate of the
original circuit, and then we rep eat the pro cess for the inputs to gate g.
Lo oking at a path name, it is not to o hard to compute what typ e of gate one is at, and
what the fan-in needs to b e. However, it will b e a bit messy to describ e. 3
of the gate reached by following a given path, and to determine the output of
the uniformity machine for C .Now it is not hard to use oracle gates of this
n
form to build a new circuit family recognizing A, and having the prop erty that
the circuits are sp eci ed bya k -go o d machine M . The details are left to the
interested reader.
This completes the sketch of the pro of.
The language C we build is complete under pro jections which is more re-
strictive than b eing complete under many-one reductions.
The same construction can b e carried out for other functions f in place of
1
DET. However, if f is not at least hard for NC under pro jections, then the
1
language C we build won't necessarily b e hard for NC f . Furthermore, if we
1
want to simulate logspace-uniform NC f , then this construction will require
that f b e hard for logspace.
Acknowledgment: Thanks to Meena Maha jan and V. Vinay for p ointing
out that the claim made in [AO96] requires pro of.
References
[AO96] Allender and M. Ogihara. Relationships among PL, L, and the deter-
minant. RAIRO - Theoretical Information and Application, 30 1996,
1{21.
[BIS90] D. A. Mix Barrington, N. Immerman, and H. Straubing. On unifor-
1
mity within NC . Journal of Computer and System Sciences, 41:274{
306, 1990. 4