Other Complexity Classes and Measures: Chapter 29 of The Forthcoming CRC Handbook On Algorithms and Theory of Computation

Download as pdf or txt
Download as pdf or txt
You are on page 1of 39

Other Complexity Classes and Measures

Chapter 29 of the forthcoming CRC Handbook on Algorithms and Theory of Computation


Eric Allender
1
Rutgers University
Michael C. Loui
2
University of Illinois at Urbana-Champaign
Kenneth W. Regan
3
State University of New York at Bualo
1 Introduction
In the previous two chapters, we have
introduced the basic complexity classes,
summarized the known relationships between these classes, and
seen how reducibility and completeness can be used to establish tight links between natural
computational problems and complexity classes.
Some natural problems seem not to be complete for any of the complexity classes we have seen
so far. For example, consider the problem of taking as input a graph G and a number k, and
deciding whether k is exactly the length of the shortest traveling salespersons tour. This is clearly
related to the TSP problem discussed in Chapter 28, Section 3, but in contrast to TSP, it seems
not to belong to NP, and also seems not to belong to co-NP.
To classify and understand this and other problems, we will introduce a few more complexity
classes. We cannot discuss all of the classes that have been studiedthere are further pointers to
the literature at the end of this chapter. Our goal is to describe some of the most important classes,
such as those dened by probabilistic and interactive computation.
1
Supported by the National Science Foundation under Grant CCR-9509603. Portions of this work were performed
while a visiting scholar at the Institute of Mathematical Sciences, Madras, India.
2
Supported by the National Science Foundation under Grant CCR-9315696.
3
Supported by the National Science Foundation under Grant CCR-9409104.
1
A common theme is that the new classes arise from the interaction of complexity theory with
other elds, such as randomized algorithms, formal logic, combinatorial optimization, and ma-
trix algebra. Complexity theory provides a common formal language for analyzing computational
performance in these areas. Other examples can be found in other chapters of this Handbook.
2 The Polynomial Hierarchy
Recall from Chapter 27, Section 2.8 that PSPACE is equal to the class of languages that can
be recognized in polynomial time on an alternating Turing machine, and that NP corresponds to
polynomial time on a nondeterministic Turing machine, which is just an alternating Turing machine
that uses only existential states. Thus, in some sense, NP sits near the very bottom of PSPACE,
and as we allow more use of the power of alternation, we slowly climb up toward PSPACE.
Many natural and important problems reside near the bottom of PSPACE in this sense, but
are neither known nor believed to be in NP. (We shall see some examples later in this chapter.)
Most of these problems can be accepted quickly by alternating Turing machines that make only
two or three alternations between existential and universal states. This observation motivates the
denition in the next paragraph.
With reference to Section 2.4 of Chapter 24, dene a k-alternating Turing machine to
be a machine such that on every computation path, the number of changes from an existential
state to universal state, or from a universal state to an existential state, is at most k 1. Thus,
a nondeterministic Turing machine, which stays in existential states, is a 1-alternating Turing
machine.
It turns out that the class of languages recognized in polynomial time by 2-alternating Turing
machines is precisely NP
SAT
. This is a manifestation of something more general, and it leads us
to the following denitions.
Let ( be a class of languages. Dene
NP
C
=

AC
NP
A
,

P
0
=

P
0
= P;
and for k 0, dene
2
q
q
q

d
d
d

&
&
&
&
&
&
d
d
d

PSPACE
PH
P

P
1
= NP

P
1
= co-NP

P
2

P
2
Figure 1: The polynomial hierarchy.

P
k+1
= NP

P
k
,

P
k+1
= co-

P
k+1
.
Observe that

P
1
= NP
P
= NP, because each of polynomially many queries to an oracle
language in P can be answered directly by a (nondeterministic) Turing machine in polynomial
time. Consequently,

P
1
= co-NP. For each k,

P
k

P
k+1
, and

P
k

P
k+1
, but these inclusions
are not known to be strict. See Figure 1.
The classes

P
k
and

P
k
constitute the polynomial hierarchy. Dene
PH =

k0

P
k
.
It is straightforward to prove that PH PSPACE, but it is not known whether the inclusion is strict.
In fact, if PH = PSPACE, then the polynomial hierarchy collapses to some level, i.e., PH =

P
m
for
some m.
3
We have already hinted that the levels of the polynomial hierarchy correspond to k-alternating
Turing machines. The next theorem makes this correspondence explicit, and also gives us a third
equivalent characterization.
Theorem 2.1 For any language A, the following are equivalent:
1. A

P
k
.
2. A is decided in polynomial time by a k-alternating Turing machine that starts in an existential
state.
3. There exists a language B P and a polynomial p such that for all x, x A if and only if
(y
1
: [y
1
[ p([x[))(y
2
: [y
2
[ p([x[)) (Qy
k
: [y
k
[ p([x[))[(x, y
1
, . . . , y
k
) B],
where the quantier Q is if k is odd, if k is even.
In Section 8 of Chapter 28, we discussed some of the startling consequences that would follow
if NP were included in P/poly, but observed that this inclusion was not known to imply P = NP.
It is known, however, that if NP P/poly, then PH collapses to its second level,

P
2
[Karp and
Lipton, 1982]. It is generally considered likely that PH does not collapse to any level, and hence
that all of its levels are distinct. Hence this result is considered strong evidence that NP is not a
subset of P/poly.
Also inside the polynomial hierarchy is the important class BPP of problems that can be solved
eciently and reliably by probabilistic algorithms, to which we now turn.
3 Probabilistic Complexity Classes
Since the 1970s, with the development of randomized algorithms for computational problems (see
Chapter 15), complexity theorists have placed randomized algorithms on a rm intellectual foun-
dation. In this section, we outline some basic concepts in this area.
A probabilistic Turing machine M can be formalized as a nondeterministic Turing machine
with exactly two choices at each step. During a computation, M chooses each possible next step
with independent probability 1/2. Intuitively, at each step, M ips a fair coin to decide what to
do next. The probability of a computation path of t steps is 1/2
t
. The probability that M accepts
4
an input string x, denoted by p
M
(x), is the sum of the probabilities of the accepting computation
paths.
Throughout this section, we consider only machines whose time complexity t(n) is time-con-
structible. Without loss of generality, we may assume that every computation path of such a
machine halts in exactly t steps.
Let A be a language. A probabilistic Turing machine M decides A with
for all x A for all x , A
unbounded two-sided error if p
M
(x) > 1/2 p
M
(x) 1/2
bounded two-sided error if p
M
(x) > 1/2 + p
M
(x) < 1/2
for some constant
one-sided error if p
M
(x) > 1/2 p
M
(x) = 0
Many practical and important probabilistic algorithms make one-sided errors. For example,
in the Solovay-Strassen primality testing algorithm of Chapter 15 (on randomized algorithms),
when the input x is a prime number, the algorithm always says prime; when x is composite,
the algorithm usually says composite, but may occasionally say prime. Using the denitions
above, this means that the Solovay-Strassen algorithm is a one-sided error algorithm for the set
A of composite numbers. It also is a bounded two-sided error algorithm for A, the set of prime
numbers.
These three kinds of errors suggest three complexity classes:
PP is the class of languages decided by probabilistic Turing machines of polynomial time
complexity with unbounded two-sided error.
BPP is the class of languages decided by probabilistic Turing machines of polynomial time
complexity with bounded two-sided error.
RP is the class of languages decided by probabilistic Turing machines of polynomial time
complexity with one-sided error.
In the literature, RP is also called R.
A probabilistic Turing machine M is a PP-machine (respectively, a BPP-machine, an RP-
machine) if M has polynomial time complexity, and M decides with two-sided error (bounded
two-sided error, one-sided error).
5
Through repeated Bernoulli trials, we can make the error probabilities of BPP-machines and RP-
machines arbitrarily small, as stated in the following theorem. (Among other things, this theorem
implies that RP BPP.)
Theorem 3.1 If L BPP, then for every polynomial q(n), there exists a BPP-machine M such
that p
M
(x) > 1 1/2
q(n)
for every x L, and p
M
(x) < 1/2
q(n)
for every x , L.
If L RP, then for every polynomial q(n), there exists an RP-machine M such that p
M
(x) >
1 1/2
q(n)
for every x in L.
It is important to note just how minuscule the probability of error is (provided that the coin
ips are truly random). If the probability of error is less than 1/2
5000
, then it is less likely that the
algorithm produces an incorrect answer than that the computer will be struck by a meteor. An
algorithm whose probability of error is 1/2
5000
is essentially as good as an algorithm that makes
no errors. For this reason, many computer scientists consider BPP to be the class of practically
feasible computational problems.
Next, we dene a class of problems that have probabilistic algorithms that make no errors.
Dene
ZPP = RP co-RP.
The letter Z in ZPP is for zero probability of error, as we now demonstrate. Suppose A ZPP.
Here is an algorithm that checks membership in A. Let M be an RP-machine that decides A, and
let M

be an RP-machine that decides A. For an input string x, alternately run M and M

on x,
repeatedly, until a computation path of one machine accepts x. If M accepts x, then accept x; if
M

accepts x, then reject x. This algorithm works correctly because when an RP-machine accepts
its input, it does not make a mistake. This algorithm might not terminate, but with very high
probability, the algorithm terminates after a few iterations.
The next theorem expresses some known relationships between probabilistic complexity classes
and other complexity classes, such as classes in the polynomial hierarchy (see Section 2).
Theorem 3.2
(a) P ZPP RP BPP PP PSPACE.
6
(b) RP NP PP.
(c) BPP

P
2

P
2
.
(d) PH P
PP
.
(e) TC
0
PP.
(Note that the last inclusion is strict! TC
0
is not known to be dierent from NP, but it is a proper
subset of PP.) Figure 2 illustrates many of these relationships. PP is not considered to be a feasible
class because it contains NP.
Even though it is not clear that there is a good physical source of randomness that can be
used to execute probabilistic algorithms and obtain the desired low error bounds, pseudo-random
generators are often used and seem to work well. There is currently great interest in de-randomizing
probabilistic algorithms, but that topic is beyond the scope of this chapter. There is a simple sense
in which a probabilistic algorithm can be de-randomized, however. If an algorithm has very small
error probability (in particular, if it has error probability a little less than 1/2
n
), then there is one
sequence of coin ips that gives the right answer on all inputs of length n, and this sequence can
be hard-wired into the algorithm to yield a deterministic (but non-uniform) circuit family. More
formally:
Theorem 3.3 BPP P/poly.
There is another important way in which BPP, RP, and ZPP dier from PP (as well as from
NP and all of the other complexity classes we have discussed thus far): BPP, RP, and ZPP are not
known to have any complete languages. Intuitively, BPP is believed to lack complete sets because
there is no computable way to weed out those polynomial-time probabilistic Turing machines that
are not BPP-machines from those that are. The same goes for RP and ZPPa more detailed
discussion of this point may be found in [Sipser, 1982] and [Ambos-Spies, 1986]. To be sure, if
these classes equal P then trivially they have complete languages. Recent work [Impagliazzo and
Wigderson, 1997] proves that a highly plausible hardness assertion for languages in exponential
time implies P = BPP.
Log-space analogues of these probabilistic classes have also been studied, of which the most
important is RL, dened by probabilistic TMs with one-sided error that run in log space and may
7
&
&
&
&
&
&

d
d
d
d
d
d

NP
PP PH
P
PP
BPP
P
RP
ZPP
PSPACE
Figure 2: Probabilistic complexity classes.
use polynomially many random bits in any computation. An important problem in RL that is not
known to be in L is that of whether there is a path from node s to node t in an undirected graph,
or much the same thing, whether an undirected graph is connected.
4 Formal Logic and Complexity Classes
There is a surprisingly close connection between important complexity classes and natural notions
that arise in the study of formal logic. This connection has led to important applications of
complexity theory to logic, and vice-versa. Below, we present some basic notions from formal logic,
and then we show some of the connections between logic and complexity theory.
Descriptive complexity refers to the ability to describe and characterize individual problems
and whole complexity classes by certain kinds of formulas in formal logic. These descriptions
do not depend on an underlying machine modelthey are machine-independent. Furthermore,
computational problems can be described in terms of their native data structures, rather than
under ad hoc string encodings.
8
A relational structure consists of a set V (called the universe), a tuple E
1
, . . . , E
k
of relations
on V , and a tuple c
1
, . . . , c

of elements of V (k, 0). Its type is given by the tuple (a


1
, . . . , a
k
)
of arities of the respective relations, together with . In this chapter, V is always nite. For
example, directed graphs G = (V, E) are relational structures with the one binary relation E, and
their type has k = 1, a
1
= 2, and = 0, the last since there are no distinguished vertices. For
another example, instances of the Graph Accessibility Problem (GAP) consist of a directed
graph G = (V, E) along with two distinguished vertices s, t V (see Chapter 28, Section 5), so
they have = 2.
An ordinary binary string x can be regarded as a structure (V, X, ), where is a total order
on V that sequences the bits, and for all i (1 i [x[), x
i
= 1 if and only if X(u
i
) holds. Here
u
i
is the ith element of V under the total order, and x
i
is the ith bit of x. It is often desirable to
regard the ordering as xed, and focus attention on the single unary relation X() as the essence
of the string.
4.1 Systems of Logic
For our purposes, a system of logic (or logic language) L consists of the following:
1. A tuple (E
1
, . . . , E
k
) of relation symbols, with corresponding arities a
1
, . . . , a
k
1, and a
tuple (c
1
, . . . , c

) of constant symbols (k, 0). These symbols constitute the vocabulary of


L, and can be identied with the corresponding type of relational structures.
2. Optionally, a further nite collection of relation and constant symbols whose interpretations
are xed in all universes V under consideration. By default this collection contains the symbol
=, which is interpreted as the equality relation on V .
3. An unbounded supply of variable symbols u, v, w, . . . ranging over elements of V , and option-
ally, an unbounded supply of variable relation symbols R
1
, R
2
, R
3
, . . ., each with an associated
arity and ranging over relations on V .
4. A complete set of Boolean connectives, for which we use , , , , and , and the
quantiers , . Additional kinds of operators for building up formulas are discussed later.
9
The well-formed formulas of L, and the free, bound, positive, and negative occurrences of
symbols in a formula, are dened in the usual inductive manner. A sentence is a formula with
no free variables. A formula, or a whole system, is called rst-order if it has no relation variables
R
i
; otherwise it is second-order.
Just as machines of a particular type dene complexity classes, so also do logical formulas of a
particular type dene important classes of languages. The most common nomenclature for these
classes begins with a prex such as FO or F
1
for rst-order systems, and SO or F
2
for second-order.
SO denotes systems whose second-order formulas are restricted to the form (R
1
)(R
2
) . . . (R
k
)
with rst-order. After this prex, in parentheses, we list the vocabulary, and any extra xed-
interpretation symbols or additions to formulas. For instance, SO(Graphs, ) stands for the
second-order existential theory of graphs whose nodes are labeled and ordered. (The predicate
= is always available in the logics we study, and thus it is not explicitly listed with the other
xed-interpretation symbols such as .)
The xed-interpretation symbols deserve special mention. Many authorities treat them as part
of the vocabulary. A nite universe V may without loss of generality be identied with the set
1, . . . , n, where n N. Important xed-interpretation symbols for these sets, besides = and ,
are Suc, +, and , respectively standing for the successor, addition, and multiplication relations.
(Here +(i, j, k) stands for i + j = k, etc.) Insofar as they deal with the numeric coding of V and
do not depend on any structures that are being built on V , such xed-interpretation symbols are
commonly called numerical predicates.
4.2 Languages, Logics, and Complexity Classes
Let us see how a logical formula describes a language, just as a Turing machine or a program does.
A formal inductive denition of the following key notion, and much further information on systems
of logic, may be found in the standard text [Enderton, 1972].
Denition 4.1. Let be a sentence in a system L with vocabulary . A relational structure
of type satises (or models) , written [= , if becomes a true statement about when the
elements of are substituted for the corresponding vocabulary symbols of . The language of
is L

= : [= .
10
We say that describes L

, or describes the property of belonging to L

. Finally, given a
system L of vocabulary , L itself stands for the class of structures of type that are described
by formulas in L. If is the vocabulary Strings of binary strings, then L

is a language in the
familiar sense of a subset of 0, 1

, and systems L over dene ordinary classes of languages.


Thus dening sets of structures over generalizes the notion of dening languages over an alphabet.
For example, the formula (u)X(u) over binary strings describes the language 1

, while
(v, w)[v ,= w E(v, w)] denes complete (loop-free) graphs. The formula
Undir = (v, w)[E(v, w) E(w, v)] (u)E(u, u)
describes the property of being an undirected simple graph, treating an undirected edge as a pair of
directed edges, and ruling out self-loops. Given unary relation symbols X
1
, . . . , X
k
, the formula
Uniq
X
1
,...,X
k
= (v)[

1ik
X
i
(v)

1i<jk
(X
i
(v) X
j
(v))]
expresses that every element v is assigned exactly one i such that X
i
(v) holds. Given an arbitrary
nite alphabet = c
1
, . . . , c
k
, the vocabulary X
1
, . . . , X
k
, together with this formula, enables
us to dene languages of strings over . (Since the presence of Uniq does not aect any of the
syntactic characterizations that follow, we may now regard Strings as a vocabulary over any .)
Given a unary relation symbol R and the numerical predicate Suc on V , the formula
Alts
R
= (s, t)(u, v)[Suc(u, s) Suc(t, u) R(s) R(t) (Suc(u, v) (R(u) R(v))]
says that R is true of the rst element s, false of the last element t, and alternates true and false
in-between. This requires [V [ to be even. The following examples are used again below.
(1) The regular language (10)

is described by the rst-order formula


1
= Alts
X
.
(2) (11)

is described by the second-order formula


2
= (R)(u)[X(u) Alts
R
].
(3) Graph Three-Colorability:

3
= Undir (R
1
, R
2
, R
3
)[Uniq
R
1
,R
2
,R
3
(v, w)(E(v, w)

1i3
R
i
(v) R
i
(w))].
(4) GAP (i.e., s-t connectivity for directed graphs):

4
= (R)(u, v)[R(s) R(t) (R(u) E(u, v) R(v))].
11
Formula
4
says that there is no set R V that is closed under the edge relation and contains
s but doesnt contain t, and this is equivalent to the existence of a path from s to t. Much trickier
is the fact that deleting Uniq
R
1
,R
2
,R
3
from
3
leaves a formula that still denes exactly the set
of undirected 3-colorable graphs. This fact hints at the delicacy of complexity issues in logic.
Much of this study originated in research on database systems, because data base query lan-
guages correspond to logics. First-order logic is notoriously limited in expressive power, and this
limitation has motivated the study of extensions of rst-order logic, such as the following rst-order
operators.
Denition 4.2.
(a) Transitive closure (TC): Let be a formula in which the rst-order variables u
1
, . . . , u
k
and
v
1
, . . . , v
k
occur freely, and regard as implicitly dening a binary relation S on V
k
. That
is, S is the set of pairs (u, v) such that (u, v) holds. Then TC
(u
1
,...,u
k
,v
1
,...,v
k
)
is a formula,
and its semantics is the reexive-transitive closure of S.
(b) Least xed point (LFP): Let be a formula with free rst-order variables u
1
, . . . , u
k
and a free
k-ary relation symbol R that occurs only positively in . In this case, for any relational struc-
ture and S V
k
, the mapping f

(S) = (e
1
, . . . , e
k
) : [= (S, e
1
, . . . , e
k
) is monotone.
That is, if S T, then for every tuple of domain elements (e
1
, . . . , e
k
), if (R, u
1
, . . . , u
k
)
evaluates to true when R is set to S and each u
i
is set to e
i
, then also evaluates to true
when R is set to T, because R appears positively. Thus the mapping f

has a least xed


point in V
k
. Then LFP
(R,u
1
,...,u
k
)
is a formula, and its semantics is the least xed point of
f

, i.e., the smallest S such that f

(S) = S.
(c) Partial xed point (PFP): Even if f

above is not monotone, PFP


(R,u
1
,...,u
k
)
is a formula
whose semantics is the rst xed point found in the sequence f

(), f

(f

()), . . ., if it exists,
otherwise.
The rst-order variables u
1
, . . . , u
k
remain free in these formulas. The relation symbol R is bound
in LFP
(R,u
1
,...,u
k
)
, but since this formula is xing R uniquely rather than quantifying over it, the
formula LFP
(R,u
1
,...,u
k
)
is still regarded as rst-order (provided is rst-order).
12
A somewhat less natural but still useful operation is the deterministic transitive closure
operator. We write DTC for the restriction of (a) above to cases where the implicitly dened
binary relation S is a partial function. The DTC restriction is enforcible syntactically by replacing
any (sub)-formula to which TC is applied by

= (w
1
, . . . , w
k
)[


k
i=1
v
i
= w
i
], where

is the result of replacing each v


i
in by w
i
, 1 i k.
For example, s-t connectivity is denable by the FO(TC) and FO(LFP) formulas

4
= (u, v)[u = s v = t TC
(u,v)
E(u, v)],

4
= (u, v)[u = s v = t LFP
(R,u,v)
],
where = (u = v E(u, v) (w)[R(u, w) R(w, v)]). To understand how

4
works, starting
with S as the empty binary relation and substituting the current S for R at each turn, the rst
iteration yields S = (u, v) : u = v E(u, v) , the second iteration gives pairs of vertices connected
by a path of length at most 2, then 4, . . . , and the xed-point is the reexive-transitive closure E

of E. Then

4
is read as if it were (u, v)(u = s v = t E

(u, v)), or more simply, as if it were


E

(s, t).
Note however, that writing DTC. . . in place of TC. . . in

4
changes the property dened by
restricting it to directed graphs in which each non-sink vertex has out-degree 1. It is not known
whether s-t connectivity can be expressed using the DTC operator. This question is equivalent to
whether L = NL.
4.3 Logical Characterizations of Complexity Classes
As discussed by [Fagin, 1993], there is a uniform encoding method Enc such that for any vocabulary
and (nite) relational structure of type , Enc() is a standard string encoding of . For
instance with = Graphs, an n-vertex graph becomes the size-n
2
binary string that lists the
entries of its adjacency matrix in row-major order. Thus one can say that a language L

over any
vocabulary belongs to a complexity class ( if the string language Enc(L

) = Enc() : [=
is in (.
The following theorems of the form ( = L all hold in the following strong sense: for every
vocabulary and L()-formula , Enc(L

) (; and for every language A (, there is a L(Strings)-


formula such that L

= A. Although going to strings via Enc may seem counter to the motivation
13
expressed in the rst paragraph of this whole section, the generality and strength of these results
has a powerful impact in the desired direction: they dene the right notion of complexity class (
for any vocabulary . Hence we omit the vocabulary in the following statements.
Theorem 4.1
(a) PSPACE = FO(PFP, ).
(b) PH = SO.
(c) (Fagins Theorem) NP = SO.
(d) P = FO(LFP, ).
(e) NL = FO(TC, ).
(f ) L = FO(DTC, ).
(g) AC
0
= FO(+, ).
One other result should be mentioned with the above. Dene the spectrum of a formula by
S

= n : for some with n elements, [= . The paper [Jones and Selman, 1974] proved
that a language A belongs to NE if and only if there is a vocabulary and a sentence FO()
such that A = S

(identifying numbers and strings). Thus spectra characterize NE.


The ordering is needed in results (a), (d), (e), and (f). The paper [Chandra and Harel, 1982]
proved that FO(LFP) without cannot even dene (11)

(and their proof works also for FO(PFP)).


Put another way, without an (ad-hoc) ordering on the full database, one cannot express queries
of the kind Is the number of widgets in Toledo even? even in the powerful system of rst-order
logic with PFP. Note that, as a consequence of what we know about complexity classes, it follows
that FO(PFP, ) is more expressive than FO(TC, ). This result is an example of an application
of complexity theory to logic. In contrast, when the ordering is not present, it is much easier to
show directly that FO(PFP) is more powerful than FO(TC) than to use the tools of complexity
theory. Furthermore, the hypotheses FO(LFP) = FO(PFP) and FO(LFP, ) = FO(PFP, ) are
both equivalent to P = PSPACE [Abiteboul and Vianu, 1995]. This shows how logic can apply to
complexity theory.
14
4.4 A Short Digression: Logic and Formal Languages
There are two more logical characterizations that seem at rst to have little to do with complexity
theory. Characterizations such as these have been important in circuit complexity, but those
considerations are beyond the scope of this chapter.
Let SF stand for the class of star-free regular languages, which are dened by regular expressions
without Kleene stars, but with as an atom and complementation () as an operator. For example,
(10)

SF via the equivalent expression [( )(00 + 11)( ) + 0( ) + ( )1].


A formula is monadic if each of its relation symbols is unary. A second-order system is monadic
if every relation variable is unary. Let mSO denote the monadic second-order formulas. The formula

2
above denes (11)

in mSO(Suc). The following results are specic to the vocabulary of strings.


Theorem 4.2
(a) REG = mSO(Strings, ) = mSO(Strings, Suc).
(b) SF = FO(Strings, ).
Theorem 4.2, combined with Theorem 4.1 (b) and (c), shows that SO is much more expressive
than mSO, and SO() is similarly more expressive than mSO(). A seemingly smaller change
to mSO also results in a leap of expressiveness from the regular languages to the level of NP. The
paper [Lynch, 1982] showed that if we consider mSO(+) instead of mSO() (for strings), then
the resulting class contains NTIME[n], and hence contains many NP-complete languages, such as
Graph Three-Colorability.
5 Interactive Models and Complexity Classes
5.1 Interactive Proofs
In Section 2.2 of Chapter 27, we characterized NP as the set of languages whose membership proofs
can be checked quickly, by a deterministic Turing machine M of polynomial time complexity. A
dierent notion of proof involves interaction between two parties, a prover P and a verier V ,
who exchange messages. In an interactive proof system, the prover is an all-powerful machine,
with unlimited computational resources, analogous to a teacher. The verier is a computationally
15
limited machine, analogous to a student. Interactive proof systems are also called Arthur-Merlin
games: the wizard Merlin corresponds to P, and the impatient Arthur corresponds to V .
Formally, an interactive proof system comprises the following:
A read-only input tape on which an input string x is written.
A prover P, whose behavior is not restricted.
A verier V , which is a probabilistic Turing machine augmented with the capability to send
and receive messages. The running time of V is bounded by a polynomial in [x[.
A tape on which V writes messages to send to P, and a tape on which P writes messages to
send to V . The length of every message is bounded by a polynomial in [x[.
A computation of an interactive proof system (P, V ) proceeds in rounds, as follows. For j = 1, 2, . . .,
in round j, V performs some steps, writes a message m
j
, and temporarily stops. Then P reads m
j
and responds with a message m

j
, which V reads in round j +1. An interactive proof system (P, V )
accepts an input string x if the probability of acceptance by V satises p
V
(x) > 1/2.
In an interactive proof system, a prover can convince the verier about the truth of a statement
without exhibiting an entire proof, as the following example illustrates.
Example: Consider the graph non-isomorphism problem: the input consists of two graphs
G and H, and the decision is yes if and only if G is not isomorphic to H. Although there
is a short proof that two graphs are isomorphic (namely: the proof consists of the isomorphism
mapping G onto H), nobody has found a general way of proving that two graphs are not isomorphic
that is signicantly shorter than listing all n! permutations and showing that each fails to be an
isomorphism. (That is, the graph non-isomorphism problem is in co-NP, but is not known to be in
NP.) In contrast, the verier V in an interactive proof system is able to take statistical evidence
into account, and determine beyond all reasonable doubt that two graphs are non-isomorphic,
using the following protocol.
In each round, V randomly chooses either G or H with equal probability; if V chooses G, then
V computes a random permutation G

of G, presents G

to P, and asks P whether G

came from
G or from H (and similarly if V chooses H). If P gave an erroneous answer on the rst round,
and G is isomorphic to H, then after k subsequent rounds, the probability that P answers all the
16
subsequent queries correctly is 1/2
k
. (To see this, it is important to understand that the prover P
does not see the coins that V ips in making its random choices; P sees only the graphs G

and
H

that V sends as messages.) V accepts the interaction with P as proof that G and H are
non-isomorphic if P is able to pick the correct graph for 100 consecutive rounds. Note that V has
ample grounds to accept this as a convincing demonstration: if the graphs are indeed isomorphic,
the prover P would have to have an incredible streak of luck to fool V .
The complexity class IP comprises the languages A for which there exists a verier V and an
such that
there exists a prover

P such that for all x in A, the interactive proof system (

P, V ) accepts
x with probability greater than 1/2 +; and
for every prover P and every x , A, the interactive proof system (P, V ) rejects x with
probability greater than 1/2 +.
By substituting random choices for existential choices in the proof that ATIME(t) DSPACE(t)
(Theorem 2.8 in Chapter 27), it is straightforward to show that IP PSPACE. It was originally
believed likely that IP was a small subclass of PSPACE. Evidence supporting this belief was the
construction in [Fortnow and Sipser, 1988] of an oracle language B for which co-NP
B
IP
B
,= ,
so that IP
B
is strictly included in PSPACE
B
. Using a proof technique that does not relativize,
however, [Shamir, 1992] (building on the work of [Lund et al., 1992]) proved that in fact, IP and
PSPACE are the same class.
Theorem 5.1 IP = PSPACE.
If NP is a proper subset of PSPACE, as is widely believed, then Theorem 5.1 says that interactive
proof systems can decide a larger class of languages than NP.
5.2 Probabilistically Checkable Proofs
In an interactive proof system, the verier does not need a complete conventional proof to become
convinced about the membership of a word in a language, but uses random choices to query parts
of a proof that the prover may know. This interpretation inspired another notion of proof: a
17
proof consists of a (potentially) large amount of information that the verier need only inspect in
a few places in order to become convinced. The following denition makes this idea more precise.
A language L has a probabilistically checkable proof if there exists an oracle BPP-machine
M such that
for all x L, there exists an oracle language B
x
such that M
Bx
accepts x.
for all x , L, and for every language B, machine M
B
rejects x.
Intuitively, the oracle language B
x
represents a proof of membership of x in L. Notice that B
x
can be nite since the length of each possible query during a computation of M
Bx
on x is bounded
by the running time of M. The oracle language takes the role of the prover in an interactive
proof systembut in contrast to an interactive proof system, the prover cannot change strategy
adaptively in response to the questions that the verier poses. This change results in a potentially
stronger system, since a machine M that has bounded error probability relative to all languages B
might not have bounded error probability relative to some adaptive prover. Although this change
to the proof system framework may seem modest, it leads to a characterization of a class that seems
to be much larger than PSPACE.
Theorem 5.2 A has a probabilistically checkable proof if and only if A NEXP.
Although the notion of probabilistically checkable proofs seems to lead us away from feasible
complexity classes, by considering natural restrictions on how the proof is accessed, we can obtain
important insights into familiar complexity classes.
Let PCP(r(n), q(n)) denote the class of languages with probabilistically checkable proofs in
which the probabilistic oracle Turing machine M makes O(r(n)) random binary choices, and queries
its oracle O(q(n)) times. (For this denition, we assume that M has either one or two choices for
each step.) It follows from the denitions that BPP = PCP(n
O(1)
, 0), and NP = PCP(0, n
O(1)
).
Theorem 5.3 NP = PCP(log n, 1).
Theorem 5.3 asserts that for every language L in NP, a proof that x L can be encoded so
that the verier can be convinced of the correctness of the proof (or detect an incorrect proof) by
using only O(log n) random choices, and inspecting only a constant number of bits of the proof!
18
This surprising characterization of NP has important applications to the complexity of nding
approximate solutions to optimization problems, as discussed in Section 6.2 below.
6 Classifying the Complexity of Functions
Up to now, we have considered only the complexity of decision problems. (Recall that a decision
problem is a problem in which, for every input, the output is either yes or no.) Most of the
functions that we actually compute are functions that produce more than one bit of output. For
example, instead of merely deciding whether a graph has a clique of size m, we often want to nd
a clique. Problems in NP are naturally associated with this kind of search problem.
Of course, any function f can be analyzed in terms of a decision problem in a straightforward
way by considering the decision problem A
f
that takes as input x and i, and answers yes if the
i
th
bit of f(x) is 1. But there are other ways of formulating functions as decision problems, and
sometimes it is instructive to study the complexity of functions directly instead of their associated
decision problems. In this section and the sections that follow, we will discuss some of the more
useful classications.
The most important class of functions is the class that we can compute quickly.
FP is the set of functions computable in polynomial time by deterministic Turing machines.
In an analogous way, we dene FL, FNC
k
, etc., to be the set of functions computable by de-
terministic log-space machines, by NC
k
circuits, etc. We also dene FPSPACE to be the class
of functions f computable by deterministic machines in polynomial space, such that also [f(x)[ is
bounded by a polynomial in [x[. This restriction is essential because a machine that uses polynomial
space could run for an exponential number of steps, producing an exponentially long output.
To study functions that appear to be dicult to compute, we again use the notions of re-
ducibility and completeness. Analogous to Cook reducibility to oracle languages, we consider Cook
reducibility to a function given as an oracle. For a function f whose length [f(x)[ is bounded by
a polynomial in [x[, we say that a language A is Cook reducible to f if there is a polynomial-time
oracle Turing machine M that accepts A, where the oracle is accessed as follows: M writes a string
y on the query tape, and in the next step y is replaced by f(y). As usual, we let P
f
and FP
f
denote
the class of languages and functions computable in polynomial time with oracle f, respectively.
19
Let ( be a class of functions. When ( is at least as big as FP, then we will use Cook reducibility
to dene completeness. That is, a function f is (-complete, if f is in ( and ( FP
f
. When we
are discussing smaller classes ( (where polynomial-time is too powerful to give a meaningful notion
of reducibility), then when we say that a function f is (-complete, it refers to completeness under
AC
0
-Turing reducibility, which was dened in Chapter 28, Section 6. In this chapter, we consider
only these two variants of Turing reducibility. There are many other ways to reduce one function
to another, just as there are many kinds of reductions between languages.
We use these notions to study optimization problems in Sections 6.1 and 6.2 and counting
problems in Section 7.
6.1 Optimization Classes
Given an optimization (minimization) problem, we most often study the following associated deci-
sion problem:
Is the optimal value at most k?
Alternatively, we could formulate the decision problem as the following:
Is the optimal value exactly k?
For example, consider the Traveling Salesperson problem (TSP) again. TSP asks whether
the length of the optimal tour is at most d
0
. Dene Exact TSP to be the decision problem that
asks whether the length of the optimal tour is exactly d
0
. It is not clear that Exact TSP is in
NP or in co-NP, but Exact TSP can be expressed as the intersection of TSP and its complement
TSP: the length of the optimal tour is d
0
if there is a tour whose length is at most d
0
, and no tour
whose length is at most d
0
1. Similar remarks apply to the optimization problem Max Clique:
given an undirected graph G, nd the maximum size of a clique in G.
Exact versions of many optimization problems can be expressed as the intersection of a language
in NP and a language in co-NP. This observation motivates the denition of a new complexity class:
D
P
is the class of languages A such that A = A
1
A
2
for some languages A
1
in NP and A
2
in co-NP.
The letter D in D
P
means dierence: A D
P
if and only if A is the dierence of two languages,
i.e., A = A
1
A
3
for some languages A
1
and A
3
in NP.
20
Not only is Exact TSP in D
P
, but also Exact TSP is D
P
-complete. Exact versions of
many other NP-complete problems, including Clique, are also D
P
-complete [Papadimitriou and
Yannakakis, 1984].
Although it is not known whether D
P
is contained in NP, it is straightforward to prove that
NP D
P
P
NP

P
2

P
2
.
Thus, D
P
lies between the rst two levels of the polynomial hierarchy.
We have characterized the complexity of computing the optimal value of an instance of an
optimization problem, but we have not yet characterized the complexity of computing the optimal
solution itself. An optimization algorithm produces not only a yes or no answer, but also,
when feasible solutions exist, an optimal solution.
First, for a maximization problem, suppose that we have a subroutine that solves the decision
problem Is the optimal value at least k? Sometimes, with repeated calls to the subroutine, we
can construct an optimal solution. For example, suppose subroutine S solves the Clique problem;
for an input graph G and integer k, the subroutine outputs yes if G has a clique of k (or more)
vertices. To construct the largest clique in an input graph, rst, determine the size K of the largest
clique by binary search on 1, . . . , n with log
2
n calls to S. Next, for each vertex v, in sequence,
determine whether deleting v produces a graph whose largest clique has size K by calling S. If so,
then delete v and continue with the remaining graph. If not, then look for a clique of size K 1
among the neighbors of v.
The method outlined in the last paragraph uses S in the same way as an oracle Turing machine
queries an oracle language in NP. With this observation, we dene the following classes:
FP
NP
is the set of functions computable in polynomial time by deterministic oracle Turing
machines with oracle languages in NP.
FP
NP[log n]
is the set of functions computable in polynomial time by deterministic oracle
Turing machines with oracle languages in NP that make O(log n) queries during computations
on inputs of length n
FP
NP
and FP
NP[log n]
contain many well-studied optimization problems (see [Krentel, 1988]).
The problem of producing the optimal tour in the Traveling Salesperson problem is FP
NP
-
21
complete. The problem of determining the size of the largest clique subgraph in a graph is
FP
NP[log n]
-complete.
6.2 Approximability and Complexity
As discussed in Chapter 34 (on approximation algorithms), because polynomial-time algorithms
for NP-complete optimization problems are unlikely to exist, we ask whether a polynomial-time
algorithm can produce a feasible solution that is close to optimal.
Fix an optimization problem with a positive integer-valued objective function g. For each
problem instance x, let OPT(x) be the optimal value, that is, g(z), where z is a feasible solution
to x that achieves the best possible value of g. Let M be a deterministic Turing machine that on
input x produces as output a feasible solution M(x) for . We say M is an -approximation
algorithm if for all x,
[g(M(x)) OPT(x)[
maxg(M(x)), OPT(x)
.
(This denition handles both minimization and maximization problems.) The problem has a
polynomial-time approximation scheme if for every xed , there is a polynomial-time -
approximation algorithm. Although the running time is polynomial in [x[, the time could be
exponential in 1/.
Several NP-complete problems, including Knapsack, have polynomial-time approximation
schemes. It is natural to ask whether all NP-complete optimization problems have polynomial-
time approximation schemes. We dene an important class of optimization problems, MAX-SNP,
whose complete problems apparently do not.
First, we dene a reducibility between optimization problems that preserves the quality of
solutions. Let
1
and
2
be optimization problems with objective functions g
1
and g
2
, respectively.
An L-reduction from
1
to
2
is dened by a pair of polynomial-time computable functions f
and f

that satisfy the following properties:


1. if x is an instance of
1
with optimal value OPT(x), then f(x) is an instance of
2
whose
optimal value satises OPT(f(x)) c OPT(x) for some constant c
2. if z is a feasible solution of f(x), then f

(z) is a feasible solution of x, such that


[OPT(x) g
1
(f

(z))[ c

[OPT(f(x)) g
2
(z)[
22
for some constant c

.
The second property implies that if z is an optimal solution to f(x), then f

(z) is an optimal solution


to x. From the denitions, it follows that if there is an L-reduction from
1
to
2
, and there is
a polynomial-time approximation scheme for
2
, then there is a polynomial-time approximation
scheme for
1
.
To dene MAX-SNP, it will help to recall the characterization of NP as SO in Section 4.3. This
characterization says that for any A in NP, there is a rst-order formula such that x A if and
only if
S
1
. . . S
l
(x, S
1
, . . . , S
l
).
For many important NP-complete problems, it is sucient to consider having only a single second-
order variable S, and to consider formulas having only universal quantiers. Thus, for such a
language A we have a quantier-free formula such that x A if and only if
Su
1
. . . u
k
(S, u
1
, . . . , u
k
).
Now dene MAX-SNP
0
to be the class of optimization problems mapping input x to
max
S
[(y
1
, . . . , y
k
) : (S, y
1
, . . . , y
k
)[.
For example, we can express in this form the Max Cut problem, the problem of nding the largest
cut in an input graph G = (V, E) with vertex set V and edge set E. A set of vertices S is the
optimal solution if it maximizes
[(v, w) : E(v, w) S(v) S(w)[.
That is, the optimal solution S maximizes the number of edges (v, w) between vertices v in S and
vertices w in V S.
Dene MAX-SNP to be the class of all optimization problems that are L-reducible to a problem
in MAX-SNP
0
. MAX-SNP contains many natural optimization problems. Max Cut is MAX-SNP-
complete, and Max Clique is MAX-SNP-hard, under L-reductions.
A surprising connection between the existence of probabilistically checkable proofs (Section 5.2)
and the existence of approximation algorithms comes out in the next major theorem.
23
Theorem 6.1 If there is a polynomial-time approximation scheme for some MAX-SNP-hard prob-
lem, then P = NP.
In particular, unless P = NP, there is no polynomial-time approximation scheme for Max Cut
or Max Clique. To prove this theorem, all we need to do is show its statement for a particular
problem that is MAX-SNP-complete under L-reductions. However, we prefer to show the idea of
the proof for the Max Clique problem, which although MAX-SNP-hard is not known to belong to
MAX-SNP. It gives a strikingly dierent kind of reduction from an arbitrary language A in NP to
Clique over the reduction from A to SAT to Clique in Chapter 28, Section 4, and its discovery
by [Feige et al., 1991, Feige et al., 1996] stimulated the whole area.
Proof. Let A NP. By Theorem 5.3, namely NP = PCP[O(log n), O(1)], there is a probabilistic
oracle Turing machine M constrained to use r(n) = O(log n) random bits and make at most a
constant number of queries in any computation path, such that:
for all x A, there exists an oracle language B
x
such that Prob
s{ 0,1 }
r(n) [M
Bx
(x, s) = 1] >
3/4;
for all x / A, and for every language B, Prob
s{ 0,1 }
r(n) [M
B
(x, s) = 1] < 1/4.
Now dene a transcript of M on input x to consist of a string s 0, 1
r(n)
together with
a sequence of pairs (w
i
, a
i
), where w
i
is an oracle query and a
i
0, 1 is a possible yes/no
answer. In addition, a transcript must be valid: for all i, 0 i < , on input x with random
bits s, having made queries w
1
, . . . , w
i
to its (unspecied) oracle and received answers a
1
, . . . , a
i
,
machine M writes w
i+1
as its next query string. Thus a transcript provides enough information
to determine a full computation path of M on input x, and the transcript is accepting if and only
if this computation path accepts. Finally, call two transcripts consistent if whenever a string w
appears as w
i
in one transcript and w
j
in the other, the corresponding answer bits a
i
and a
j
are the same.
Construction: Let G
x
be the undirected graph whose node set V
x
is the set of all accepting
transcripts, and whose edges connect pairs of transcripts that are consistent.
Complexity: Since r(n) + = O(log n), there are only polynomially many transcripts, and since
consistency is easy to check, G
x
is constructed in polynomial time.
24
Correctness: If x A, then take the oracle B
x
specied above and let C be the set of accepting
transcripts whose answer bits are given by B
x
. These transcripts are consistent with each other,
and there are at least (3/4)2
r(n)
such accepting transcripts, so C forms a clique of size at least
(3/4)2
r(n)
in G
x
. Now suppose x / A, and suppose C

is a clique of size greater than (1/4)2


r(n)
in G
x
. Because the transcripts in C

are mutually consistent, there exists a single oracle B that


produces all the answer bits to queries in transcripts in C

. But then Prob


s
[M
B
(x, s) = 1] > 1/4,
contradicting the PCP condition on M.
Thus we have proved the statement of the theorem for Max Clique. The proof of the general
statement is similar. 2
Note that the cases x A and x / A in this proof lead to a (3/4,1/4) gap in the maximum
clique size of G
x
. If there were a polynomial-time algorithm guaranteed to determine within a
factor better than 3, then this algorithm could tell the 3/4 case apart from the 1/4 case, and
hence decide whether x A. Since G
x
can be constructed in polynomial time (in particular, G
x
has size at most 2
r(n)+
= n
O(1)
), P = NP would follow. Hence we can say that Clique is NP-hard
to approximate within a factor of 3. A long sequence of improvements to this basic construction
has pushed the hardness-of-approximation not only to any xed constant factor, but also to factors
that increase with n. Moreover, approximation-preserving reductions have extended this kind of
hardness result to many other optimization problems.
7 Counting
Two other important classes of functions deserve special mention:
#P is the class of functions f such that there exists a nondeterministic polynomial-time Turing
machine M with the property that f(x) is the number of accepting computation paths of M
on input x.
#L is the class of functions f such that there exists a nondeterministic log-space Turing
machine M with the property that f(x) is the number of accepting computation paths of M
on input x.
25
Some functions in #P are clearly at least as dicult to compute as some NP-complete problems
are to decide. For instance, consider the following problem.
Number of Satisfying Assignments to a 3CNF Formula (#3CNF)
Instance: A Boolean formula in conjunctive normal form with at most three variables per clause.
Output: The number of distinct assignments to the variables that cause the formula to evaluate
to true.
Note that #3CNF is in #P, and note also that the NP-complete problem of determining
whether x 3SAT is merely the question of whether #3CNF(x) = 0.
In apparent contrast to #P, all functions in #L can be computed by NC circuits.
Theorem 7.1 Relationships between counting classes.
FP #P FPSPACE
P
PP
= P
#P
(and thus also FP
NP
FP
#P
)
FL #L FNC
2
.
It is not surprising that #P and #L capture the complexity of various functions that involve
counting, but as the following examples illustrate, it sometimes is surprising which functions are
dicult to compute.
The proof of Cooks Theorem that appears in Chapter 28 also proves that #3CNF is complete
for #P, because it shows that for every nondeterministic polynomial-time machine M and every
input x, one can eciently construct a formula with the property that each accepting computation
of M on input x corresponds to a distinct satisfying assignment, and vice versa. Thus the number
of satisfying assignments equals the number of accepting computation paths. A reduction with this
property is called parsimonious.
Most NP-complete languages that one encounters in practice are known to be complete under
parsimonious reductions. (The reader may wish to check which of the reductions presented in Chap-
ter 34 are parsimonious.) For any such complete language, it is clear how to dene a corresponding
complete function in #P.
26
Similarly, for the Graph Accessibility Problem (GAP), which is complete for NL, we can
dene the function that counts the number of paths from the start vertex s to the terminal vertex t.
For reasons that will become clear soon, we consider two versions of this problem: one for general
directed graphs, and one for directed acyclic graphs. (The restriction of GAP to acyclic graphs
remains NL-complete.)
Number of Paths in a Graph (#Paths)
Instance: A directed graph on n vertices, with two vertices s and t.
Output: The number of simple paths from s to t. (A path is a simple path if it visits no vertex
more than once.)
Number of Paths in a Directed Acyclic Graph (#DAG-Paths)
Instance: A directed acyclic graph on n vertices, with two vertices s and t.
Output: The number of paths from s to t. (In an acyclic graph, all paths are simple.)
As one might expect, the problem #DAG-Paths is complete for #L, but it may come as a
surprise that #Paths is complete for #P, as shown by [Valiant, 1979]! That is, although it is easy
to decide whether there is a path between two vertices, it seems quite dicult to count the number
of distinct paths, unless the underlying graph is acyclic.
As another example of this phenomenon, consider the problem 2SAT, which is the same as
3SAT except that each clause has at most two literals. 2SAT is complete for NL, but the problem
of counting the number of satisfying assignments for these formulas is complete for #P.
A striking illustration of the relationship between #P and #L is provided by the following two
important problems from linear algebra.
Determinant
Instance: An integer matrix.
Output: The determinant of the matrix.
Recall that the determinant of a matrix M with entries M
i,j
is given by

sign()
n

i=1
M
i,(i)
,
27
where the sum is over all permutations on 1, . . . , n, and sign() is 1 if can be written as
the composition of an odd number of transpositions, and sign() is 1 otherwise.
Permanent
Instance: An integer matrix.
Output: The permanent of the matrix. The permanent of a matrix is given by

i=1
M
i,(i)
.
The reader is probably familiar with the determinant function, which can be computed eciently
by Gaussian elimination. The permanent may be less familiar, although its denition is formally
simpler. Nobody has ever found an ecient way to compute the permanent, however.
We need to introduce slight modication of our function classes to classify these problems,
however, because #L and #P consist of functions that take only non-negative values, whereas both
the permanent and determinant can be negative.
Dene GapL to be the class of functions that can be expressed as the dierence of two #L
functions, and dene GapP to be the dierence of two #P functions.
Theorem 7.2 (a) Permanent is complete for GapP.
(b) Determinant is complete for GapL
The class of problems that are AC
0
-Turing reducible to Determinant is one of the most
important subclasses of NC, and in fact contains most of the natural problems for which NC
algorithms are known.
8 Kolmogorov Complexity
Until now, we have considered only dynamic complexity measures, namely, the time and space
used by Turing machines. Kolmogorov complexity is a static complexity measure that captures
the diculty of describing a string. For example, the string consisting of three million zeroes
can be described with fewer than three million symbols (as in this sentence). In contrast, for a
28
string consisting of three million randomly generated bits, with high probability there is no shorter
description than the string itself.
Let U be a universal Turing machine (see Chapter 26, Section 2.2). Let denote the empty
string. The Kolmogorov complexity of a binary string y with respect to U, denoted by K
U
(y),
is the length of the shortest binary string i such that on input i, ), machine U outputs y. In
essence, i is a description of y, for it tells U how to generate y.
The next theorem states that dierent choices for the universal Turing machine aect the
denition of Kolmogorov complexity in only a small way.
Theorem 8.1 (Invariance Theorem) There exists a universal Turing machine U such that for
every universal Turing machine U

, there is a constant c such that for all y, K


U
(y) K
U
(y) +c.
Henceforth, let K be dened by the universal Turing machine of Theorem 8.1. For every integer
n and every binary string y of length n, because y can be described by giving itself explicitly,
K(y) n + c

for a constant c

. Call y incompressible if K(y) n. Since there are 2


n
binary
strings of length n, and only 2
n
1 possible shorter descriptions, there exists an incompressible
string for every length n.
Kolmogorov complexity gives a precise mathematical meaning to the intuitive notion of ran-
domness. If someone ips a coin fty times and it comes up heads each time, then intuitively,
the sequence of ips is not randomalthough from the standpoint of probability theory the all-
heads sequence is precisely as likely as any other sequence. Probability theory does not provide the
tools for calling one sequence more random than another; Kolmogorov complexity theory does.
Kolmogorov complexity provides a useful framework for presenting combinatorial arguments.
For example, when one wants to prove that an object with some property P exists, then it is
sucient to show that any object that does not have property P has a short description; thus any
incompressible (or random) object must have property P. This sort of argument has been useful
in proving lower bounds in complexity theory. For example, the paper [Dietzfelbinger et al., 1991]
uses Kolmogorov complexity to show that no Turing machine with a single worktape can compute
the transpose of a matrix in less than time (n
3/2
/

log n).
29
9 Research Issues and Summary
As stated in the introduction to Chapter 27, the goals of complexity theory are (i) to ascertain the
amount of computational resources required to solve important computational problems, and (ii)
to classify problems according to their diculty. The preceding two chapters have explained how
complexity theory has devised a classication scheme in order to meet the second goal. The present
chapter has presented a few of the additional notions of complexity that have been devised in order
to capture more problems in this scheme. Progress toward the rst goal (proving lower bounds)
depends on knowing that levels in this classication scheme are in fact distinct. Thus the core
research questions in complexity theory are expressed in terms of separating complexity classes:
Is L dierent from NL?
Is P dierent from RP or BPP?
Is P dierent from NP?
Is NP dierent from PSPACE?
Motivated by these questions, much current research is devoted to eorts to understand the power
of nondeterminism, randomization, and interaction. In these studies, researchers have gone well
beyond the theory presented in Chapters 27 through this chapter:
beyond Turing machines and Boolean circuits, to restricted and specialized models in which
nontrivial lower bounds on complexity can be proved;
beyond deterministic reducibilities, to nondeterministic and probabilistic reducibilities, and
rened versions of the reducibilities considered here;
beyond worst case complexity, to average case complexity.
We have illustrated how recent research in complexity theory has had direct applications to
other areas of computer science and mathematics. Probabilistically checkable proofs were used to
show that obtaining approximate solutions to some optimization problems is as dicult as solving
them exactly. Complexity theory provides new tools for studying questions in nite model theory,
a branch of mathematical logic. Fundamental questions in complexity theory are intimately linked
30
to practical questions about the use of cryptography for computer security, such as the existence
of one-way functions and the strength of public key cryptosystems.
This last point illustrates the urgent practical need for progress in computational complexity
theory. Many popular cryptographic systems in current use are based on unproven assumptions
about the diculty of computing certain functions (such as the factoring and discrete logarithm
problems; see Chapters 3842 of this Handbook for more background on cryptography). All of these
systems are thus based on wishful thinking and conjecture. The need to resolve these open questions
and replace conjecture with mathematical certainty should be self-evident. (In the brief history of
complexity theory, we have learned that many popular conjectures turn out to be incorrect.)
With precisely dened models and mathematically rigorous proofs, research in complexity the-
ory will continue to provide sound insights into the diculty of solving real computational problems.
10 Dening Terms
Descriptive complexity: The study of classes of languages described by formulas in certain
systems of logic.
Incompressible string: A string whose Kolmogorov complexity equals its length, so that it
has no shorter encodings.
Interactive proof system: A protocol in which one or more provers try to convince another
party called the verier that the prover(s) possess certain true knowledge, such as the mem-
bership of a string x in a given language, often with the goal of revealing no further details
about this knowledge. The prover(s) and verier are formally dened as probabilistic Tur-
ing machines with special interaction tapes for exchanging messages.
Kolmogorov complexity: The minimum number of bits into which a string can be compressed
without losing information. This is dened with respect to a xed but universal decompression
scheme, given by a universal Turing machine.
L-reduction A Karp reduction that preserves approximation properties of optimization problems.
Optimization problem: A computational problem in which the object is not to decide some
yes/no property, as with a decision problem, but to nd the best solution in those yes cases
31
where a solution exists.
Polynomial hierarchy: The collection of classes of languages accepted by k-alternating Turing
machines, over all k 0 and with initial state existential or universal. The bottom level
(k = 0) is the class P, and the next level (k = 1) comprises NP and co-NP.
Polynomial time approximation scheme (PTAS): A meta-algorithm that for every > 0
produces a polynomial time -approximation algorithm for a given optimization problem.
Probabilistic Turing machine: A Turing machine in which some transitions are random choices
among nitely many alternatives.
Probabilistically checkable proof: An interactive proof system in which provers follow a xed
strategy, one not aected by any messages from the verier. The provers strategy for a given
instance x of a decision problem can be represented by a nite oracle language B
x
, which
constitutes a proof of the correct answer for x.
Relational structure: The counterpart in formal logic of a data structure or class instance in
the object-oriented sense. Examples are strings, directed graphs, and undirected graphs. Sets
of relational structures generalize the notion of languages as sets of strings.
References
[Abiteboul and Vianu, 1991] S. Abiteboul and V. Vianu. Datalog extensions for database queries
and updates. J. Comp. Sys. Sci., 43:62124, 1991.
[Abiteboul and Vianu, 1995] S. Abiteboul and V. Vianu. Computing with rst-order logic. J.
Comp. Sys. Sci., 50:309335, 1995.
[Adleman, 1978] L. Adleman. Two theorems on random polynomial time. In Proc. 19th Annual
IEEE Symposium on Foundations of Computer Science, pages 7583, 1978.
[Allender, 1997] E. Allender. The permanent requires large uniform threshold circuits. Technical
Report TR 97-51, DIMACS, September 1997. Submitted for publication.
[Alvarez and Jenner, 1993] C. Alvarez and B. Jenner. A very hard log-space counting class. Theor.
Comp. Sci., 107:330, 1993.
32
[Ambos-Spies, 1986] K. Ambos-Spies. A note on complete problems for complexity classes. Inf.
Proc. Lett., 23:227230, 1986.
[Arora and Safra, 1992] S. Arora and S. Safra. Probabilistic checking of proofs. In Proc. 33rd
Annual IEEE Symposium on Foundations of Computer Science, pages 213, 1992.
[Arora et al., 1992] S. Arora, C. Lund, R. Motwani, M. Sudan, and M. Szegedy. Proof verication
and hardness of approximation problems. In Proc. 33rd Annual IEEE Symposium on Foundations
of Computer Science, pages 1423, 1992.
[Babai and Moran, 1988] L. Babai and S. Moran. Arthur-Merlin games: A randomized proof sys-
tem, and a hierarchy of complexity classes. J. Comp. Sys. Sci., 36:254276, 1988.
[Babai et al., 1991] L. Babai, L. Fortnow, and C. Lund. Nondeterministic exponential time has
two-prover interactive protocols. Computational Complexity, 1:340, 1991. Addendum in vol. 2
of same journal.
[Balcazar et al., 1990] J. Balcazar, J. Daz, and J. Gabarr o. Structural Complexity I,II. Springer
Verlag, 1990. Part I published in 1988.
[Barrington et al., 1990] D. Mix Barrington, N. Immerman, and H. Straubing. On uniformity
within NC
1
. J. Comp. Sys. Sci., 41:274306, 1990.
[Bovet and Crescenzi, 1994] D. Bovet and P. Crescenzi. Introduction to the Theory of Complexity.
Prentice Hall International (UK) Limited, Hertfordshire, U.K., 1994.
[B uchi, 1960] J. B uchi. Weak second-order arithmetic and nite automata. Zeitschrift f ur Mathe-
matische Logik und Grundlagen der Mathematik, 6:6692, 1960.
[Chandra and Harel, 1982] A. Chandra and D. Harel. Structure and complexity of relational
queries. J. Comp. Sys. Sci., 25:99128, 1982.
[Chandra et al., 1981] A. Chandra, D. Kozen, and L. Stockmeyer. Alternation. J. Assn. Comp.
Mach., 28:114133, 1981.
[Dietzfelbinger et al., 1991] M. Dietzfelbinger, W. Maass, and G. Schnitger. The complexity of
matrix transposition on one-tape o-line Turing machines. Theor. Comp. Sci., 82:113129, 1991.
33
[Downey and Fellows, 1995] R. Downey and M. Fellows. Fixed-parameter tractability and com-
pleteness I: Basic theory. SIAM J. Comput., 24:873921, 1995.
[Enderton, 1972] H.B. Enderton. A mathematical introduction to logic. Academic Press, New York,
1972.
[Fagin, 1974] R. Fagin. Generalized rst-order spectra and polynomial-time recognizable sets. In
R. Karp, editor, Complexity of Computation: Proceedings of a Symposium in Applied Mathemat-
ics of the American Mathematical Society and the Society for Industrial and Applied Mathemat-
ics, Vol. VII, pages 4373. SIAM-AMS, 1974.
[Fagin, 1993] R. Fagin. Finite model theorya personal perspective. Theor. Comp. Sci., 116:331,
1993.
[Feige et al., 1991] U. Feige, S. Goldwasser, L. Lovasz, Safra S, and M. Szegedy. Approximat-
ing clique is almost NP-complete. In Proc. 32nd Annual IEEE Symposium on Foundations of
Computer Science, pages 212, 1991.
[Feige et al., 1996] U. Feige, S. Goldwasser, L. Lov asz, Safra S, and M. Szegedy. Interactive proofs
and the hardness of approximating cliques. Journal of the ACM, 43:268292, 1996.
[Fortnow and Sipser, 1988] L. Fortnow and M. Sipser. Are there interactive protocols for co-NP
languages? Inf. Proc. Lett., 28:249251, 1988.
[Gill, 1977] J. Gill. Computational complexity of probabilistic Turing machines. SIAM J. Comput.,
6:675695, 1977.
[Goldwasser et al., 1989] S. Goldwasser, S. Micali, and C. Racko. The knowledge complexity of
interactive proof systems. SIAM J. Comput., 18:186208, 1989.
[Hartmanis, 1989] J. Hartmanis, editor. Computational Complexity Theory. American Mathemat-
ical Society, 1989.
[Hartmanis, 1994] J. Hartmanis. On computational complexity and the nature of computer science.
Comm. Assn. Comp. Mach., 37:3743, 1994.
34
[Immerman, 1989] N. Immerman. Descriptive and computational complexity. In J. Hartmanis,
editor, Computational Complexity Theory, volume 38 of Proc. Symp. in Applied Math., pages
7591. American Mathematical Society, 1989.
[Impagliazzo and Wigderson, 1997] R. Impagliazzo and A. Wigderson. P = BPP if E requires
exponential circuits: Derandomizing the XOR Lemma. In Proc. 29th Annual ACM Symposium
on the Theory of Computing, pages 220229, 1997.
[Jones and Selman, 1974] N. Jones and A. Selman. Turing machines and the spectra of rst-order
formulas. J. Assn. Comp. Mach., 39:139150, 1974.
[Karp and Lipton, 1982] R. Karp and R. Lipton. Turing machines that take advice.
LEnseignement Mathematique, 28:191210, 1982.
[Krentel, 1988] M. Krentel. The complexity of optimization problems. J. Comp. Sys. Sci., 36:490
509, 1988.
[Lautemann, 1983] C. Lautemann. BPP and the polynomial hierarchy. Inf. Proc. Lett., 17:215217,
1983.
[Li and Vit anyi, 1993] M. Li and P. Vit anyi. An Introduction to Kolmogorov Complexity and its
Applications. Springer Verlag, 1993.
[Lindell, 1994] S. Lindell. How to dene exponentiation from addition and multiplication in rst-
order logic on nite structures, 1994. Unpublished manuscript. The main result will appear in a
forthcoming text by N. Immerman, Descriptive and Computational Complexity, in the Springer
Verlag Graduate Texts in Computer Science series.
[Lund et al., 1992] C. Lund, L. Fortnow, H. Karlo, and N. Nisan. Algebraic methods for interac-
tive proof systems. J. Assn. Comp. Mach., 39:859868, 1992.
[Lutz, 1997] J. Lutz. The quantitative structure of exponential time. In L. Hemaspaandra and
A. Selman, editors, Complexity Theory Retrospective II, pages 225260. Springer Verlag, 1997.
[Lynch, 1982] J. Lynch. Complexity classes and theories of nite models. Math. Sys. Thy., 15:127
144, 1982.
35
[McNaughton and Papert, 1971] R. McNaughton and S. Papert. Counter-Free Automata. MIT
Press, Cambridge, MA, 1971.
[Papadimitriou and Yannakakis, 1984] C. Papadimitriou and M. Yannakakis. The complexity of
facets (and some facets of complexity). J. Comp. Sys. Sci., 28:244259, 1984.
[Papadimitriou, 1994] C. Papadimitriou. Computational Complexity. Addison-Wesley, Reading,
Mass., 1994.
[Sch utzenberger, 1965] M.P. Sch utzenberger. On nite monoids having only trivial subgroups.
Inform. and Control, 8:190194, 1965.
[Shamir, 1992] A. Shamir. IP = PSPACE. J. Assn. Comp. Mach., 39:869877, 1992.
[Sipser, 1982] M. Sipser. On relativization and the existence of complete sets. In Proc. 9th Annual
International Conference on Automata, Languages, and Programming, volume 140 of Lect. Notes
in Comp. Sci., pages 523531. Springer Verlag, 1982.
[Sipser, 1983] M. Sipser. Borel sets and circuit complexity. In Proc. 15th Annual ACM Symposium
on the Theory of Computing, pages 6169, 1983.
[Sipser, 1992] M. Sipser. The history and status of the P versus NP question. In Proc. 24th Annual
ACM Symposium on the Theory of Computing, pages 603618, 1992.
[Stearns, 1990] R. Stearns. Juris Hartmanis: the beginnings of computational complexity. In
A. Selman, editor, Complexity Theory Retrospective, pages 518. Springer Verlag, New York,
1990.
[Stockmeyer, 1976] L. Stockmeyer. The polynomial time hierarchy. Theor. Comp. Sci., 3:122,
1976.
[Stockmeyer, 1987] L. Stockmeyer. Classifying the computational complexity of problems. J. Symb.
Logic, 52:143, 1987.
[Toda, 1991] S. Toda. PP is as hard as the polynomial-time hierarchy. SIAM J. Comput., 20:865
877, 1991.
36
[Valiant, 1979] L. Valiant. The complexity of computing the permanent. Theor. Comp. Sci., 8:189
201, 1979.
[van Leeuwen, 1990] J. van Leeuwen, editor. Handbook of Theoretical Computer Science, volume A.
Elsevier and MIT Press, 1990.
[Vinay, 1991] V. Vinay. Counting auxiliary pushdown automata and semi-unbounded arithmetic
circuits. In Proc. 6th Annual IEEE Conference on Structure in Complexity Theory, pages 270
284, 1991.
[Wagner and Wechsung, 1986] K. Wagner and G. Wechsung. Computational Complexity. D. Reidel,
1986.
[Wang, 1997] J. Wang. Average-case computational complexity theory. In L. Hemaspaandra and
A. Selman, editors, Complexity Theory Retrospective II, pages 295328. Springer Verlag, 1997.
[Wrathall, 1976] C. Wrathall. Complete sets and the polynomial-time hierarchy. Theor. Comp.
Sci., 3:2333, 1976.
Further Information
Primary sources for major theorems presented in this chapter include: Theorem 2.1 [Stock-
meyer, 1976, Wrathall, 1976, Chandra et al., 1981]; Theorem 3.2(a,b) [Gill, 1977], (c) [Sipser, 1983,
Lautemann, 1983], (d) [Toda, 1991], (e) [Allender, 1997]; Theorem 3.3 [Adleman, 1978]; Theo-
rem 4.1(a) [Abiteboul and Vianu, 1991], (b) [Stockmeyer, 1976], (c) [Fagin, 1974], (d,e,f) [Immer-
man, 1989], (g) ([Lindell, 1994], cf. [Barrington et al., 1990]); Theorem 4.2(a) [B uchi, 1960], (b)
[Sch utzenberger, 1965, McNaughton and Papert, 1971]; Theorem 5.1 [Shamir, 1992]; Theorem 5.2
[Babai et al., 1991]; Theorem 5.3 [Arora and Safra, 1992]; Theorem 6.1 [Arora et al., 1992]; Theo-
rem 7.2(a) [Valiant, 1979], (b) [Vinay, 1991]. The operators in Denition 4.2 are from [Immerman,
1989] and [Abiteboul and Vianu, 1991]. Interactive proof systems were dened by [Goldwasser et
al., 1989], and in the Arthur-Merlin formulation, by [Babai and Moran, 1988]. A large com-
pendium of optimization problems and hardness results collected by P. Crescenzi and V. Kann is
available at:
http://www.nada.kth.se/~viggo/index-en.html
37
The class #P was introduced by [Valiant, 1979], and #L by [Alvarez and Jenner, 1993]. The book
[Li and Vit anyi, 1993] gives a far-reaching and comprehensive scholarly treatment of Kolmogorov
complexity, with many applications, as well as the source of Theorem 8.1.
Three contemporary textbooks on complexity theory are [Balcazar et al., 1990], [Bovet and
Crescenzi, 1994], and [Papadimitriou, 1994]. [Wagner and Wechsung, 1986] is an exhaustive survey
of complexity theory that covers work published before 1986. Another perspective of some of the
issues covered in these three chapters may be found in the survey [Stockmeyer, 1987].
A good general reference is the Handbook of Theoretical Computer Science [van Leeuwen, 1990],
volume A. The following chapters in the Handbook are particularly relevant: Machine models and
simulations, by P. van Emde Boas, pp. 166; A catalog of complexity classes, by D. S. Johnson,
pp. 67161; Machine-independent complexity theory, by J. I. Seiferas, pp. 163186; Kolmogorov
complexity and its applications, by M. Li and P. M. B. Vitanyi, pp. 187254; and The complexity
of nite functions, by R. B. Boppana and M. Sipser, pp. 757804, which covers circuit complexity.
A collection of articles edited by Hartmanis [Hartmanis, 1989] includes an overview of complexity
theory, and chapters on sparse complete languages, on relativizations, on interactive proof systems,
and on applications of complexity theory to cryptography. For historical perspectives on complexity
theory, see [Hartmanis, 1994], [Sipser, 1992], and [Stearns, 1990].
There are many areas of complexity theory that we have not been able to cover in these chapters.
Some of them cross-pollinate with other elds of computer science and are reected in other chapters
of this Handbook. Three others are average-case complexity, resource-bounded measure theory, and
parameterized complexity. Recent surveys on the rst two are [Wang, 1997, Lutz, 1997], while the
third stems from [Downey and Fellows, 1995] and has its own Web site, currently maintained by
M. Hallett at:
http://csr.uvic.ca/home/mhallett/W.hier/compendium.html
Surveys and lecture notes on complexity theory that can be obtained via WWW are maintained
by A. Czumaj and M. Kutylowski at:
http://www.uni-paderborn.de/fachbereich/AG/agmadh/WWW/english/scripts.html
38
As usual with the WWW, these links are subject to change. A good stem page to begin searches
is the site for SIGACT, the ACM Special Interest Group on Algorithms and Computation Theory:
http://sigact.acm.org/
This has a pointer to a Virtual Address Book that indexes the personal Web pages of over
1,000 computer scientists, including all three authors of these chapters. Many of these pages have
downloadable papers and links to further research resources.
Research papers on complexity theory are presented at several annual conferences, including
the annual ACM Symposium on Theory of Computing; the annual International Colloquium on
Automata, Languages, and Programming, sponsored by the European Association for Theoretical
Computer Science (EATCS); and the annual Symposium on Foundations of Computer Science,
sponsored by the IEEE. The annual Conference on Computational Complexity (formerly Struc-
ture in Complexity Theory), also sponsored by the IEEE, is entirely devoted to complexity theory.
Research articles on complexity theory regularly appear in the following journals, among others:
Chicago Journal on Theoretical Computer Science, Computational Complexity, Information and
Computation, Journal of the ACM, Journal of Computer and System Sciences, Mathematical Sys-
tems Theory, SIAM Journal on Computing, and Theoretical Computer Science. Each issue of ACM
SIGACT News and Bulletin of the EATCS contains a column on complexity theory.
39

You might also like