Complexity of Algorithms

Lecture Notes, Spring 1999

Peter Gács
Boston University

and

László Lovász
Yale University

1
Contents
0 Introduction and Preliminaries 1
0.1 The subject of complexity theory . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
0.2 Some notation and definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1 Models of Computation 3
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Finite automata . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3 The Turing machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 The Random Access Machine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.5 Boolean functions and Boolean circuits . . . . . . . . . . . . . . . . . . . . . . . . 24

2 Algorithmic decidability 31
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
2.2 Recursive and recursively enumerable languages . . . . . . . . . . . . . . . . . . . 33
2.3 Other undecidable problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
2.4 Computability in logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3 Computation with resource bounds 50

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.2 Time and space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
3.3 Polynomial time I: Algorithms in arithmetic . . . . . . . . . . . . . . . . . . . . . 52
3.4 Polynomial time II: Graph algorithms . . . . . . . . . . . . . . . . . . . . . . . . 57
3.5 Polynomial space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62

4 General theorems on space and time complexity 65

4.1 Space versus time . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

5 Non-deterministic algorithms 72
5.1 Non-deterministic Turing machines . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2 Witnesses and the complexity of non-deterministic algorithms . . . . . . . . . . . 74
5.3 General results on nondeterministic complexity classes . . . . . . . . . . . . . . . 76
5.4 Examples of languages in NP . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.5 NP-completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
5.6 Further NP-complete problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89

6 Randomized algorithms 99
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2 Verifying a polynomial identity . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.3 Prime testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
6.4 Randomized complexity classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108

2
7 Information complexity: the complexity-theoretic notion of randomness 112
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.2 Information complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
7.3 The notion of a random sequence . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
7.4 Kolmogorov complexity and data compression . . . . . . . . . . . . . . . . . . . . 119

8 Pseudo-random numbers 124

8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.2 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
8.3 Classical methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
8.4 The notion of a psuedorandom number generator . . . . . . . . . . . . . . . . . . 127
8.5 One-way functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.6 Discrete square roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133

9 Parallel algorithms 135

9.1 Parallel random access machines . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
9.2 The class NC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138

10 Decision trees 143

10.1 Algorithms using decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
10.2 The notion of decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
10.3 Nondeterministic decision trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
10.4 Lower bounds on the depth of decision trees . . . . . . . . . . . . . . . . . . . . . 151

11 Communication complexity 155

11.1 Communication matrix and protocol-tree . . . . . . . . . . . . . . . . . . . . . . 155
11.2 Some protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
11.3 Non-deterministic communication complexity . . . . . . . . . . . . . . . . . . . . 160
11.4 Randomized protocols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164

12 The complexity of algebraic computations 166

13 Circuit complexity 167

13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
13.2 Lower bound for the Majority Function . . . . . . . . . . . . . . . . . . . . . . . 168
13.3 Monotone circuits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170

14 An application of complexity: cryptography 172

14.1 A classical problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
14.2 A simple complexity-theoretic model . . . . . . . . . . . . . . . . . . . . . . . . . 172
14.3 Public-key cryptography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
14.4 The Rivest-Shamir-Adleman code . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

0
0 Introduction and Preliminaries
0.1 The subject of complexity theory
The need to be able to measure the complexity of a problem, algorithm or structure, and
to obtain bounds and quantitive relations for complexity arises in more and more sciences:
besides computer science, the traditional branches of mathematics, statistical physics, biology,
medicine, social sciences and engineering are also confronted more and more frequently with this
problem. In the approach taken by computer science, complexity is measured by the quantity
of computational resources (time, storage, program, communication) used up by a particualr
task. These notes deal with the foundations of this theory.
Computation theory can basically be divided into three parts of different character. First,
the exact notions of algorithm, time, storage capacity, etc. must be introduced. For this, dif-
ferent mathematical machine models must be defined, and the time and storage needs of the
computations performed on these need to be clarified (this is generally measured as a function
of the size of input). By limiting the available resources, the range of solvable problems gets
narrower; this is how we arrive at different complexity classes. The most fundamental com-
plexity classes provide an important classification of problems arising in practice, but (perhaps
more surprisingly) even for those arising in classical areas of mathematics; this classification
reflects the practical and theoretical difficulty of problems quite well. The relationship between
different machine models also belongs to this first part of computation theory.
Second, one must determine the resource need of the most important algorithms in various
areas of mathematics, and give efficient algorithms to prove that certain important problems
belong to certain complexity classes. In these notes, we do not strive for completeness in
the investigation of concrete algorithms and problems; this is the task of the corresponding
fields of mathematics (combinatorics, operations research, numerical analysis, number theory).
Nevertheless, a large number of concrete algorithms will be described and analyzed to illustrate
certain notions and methods, and to establish the complexity of certain problems.
Third, one must find methods to prove “negative results”, i.e. for the proof that some
problems are actually unsolvable under certain resource restrictions. Often, these questions can
be formulated by asking whether certain complexity classes are different or empty. This problem
area includes the question whether a problem is algorithmically solvable at all; this question can
today be considered classical, and there are many important results concerining it; in particular,
the decidability or undecidablity of most concrete problems of interest is known.
The majority of algorithmic problems occurring in practice is, however, such that algorithmic
solvability itself is not in question, the question is only what resources must be used for the
solution. Such investigations, addressed to lower bounds, are very difficult and are still in their
infancy. In these notes, we can only give a taste of this sort of results. In particular, we
discuss complexity notions like communication complexity or decision tree complexity, where
by focusing only on one type of rather special resource, we can give a more complete analysis
of basic complexity classes.
It is, finally, worth noting that if a problem turns out to be “difficult” to solve, this is not
necessarily a negative result. More and more areas (random number generation, communication
protocols, cryptography, data protection) need problems and structures that are guaranteed to

1
be complex. These are important areas for the application of complexity theory; from among
them, we will deal with random number generation and cryptography, the theory of secret
communication.

0.2 Some notation and definitions

A finite set of symbols will sometimes be called an alphabet. A finite sequence formed from some
elements of an alphabet Σ is called a word. The empty word will also be considered a word,
and will be denoted by ∅. The set of words of length n over Σ is denoted by Σn , the set of all
words (including the empty word) over Σ is denoted by Σ∗ . A subset of Σ∗ , i.e. , an arbitrary
set of words, is called a language.
Note that the empty language is also denoted by ∅ but it is different, from the language {∅}
containing only the empty word.
Let us define some orderings of the set of words. Suppose that an ordering of the elements
of Σ is given. In the lexicographic ordering of the elements of Σ∗ , a word α precedes a word β if
either α is a prefix (beginning segment) of β or the first letter which is different in the two words
is smaller in α. (E.g., 35244 precedes 35344 which precedes 353447.) The lexicographic ordering
does not order all words in a single sequence: for example, every word beginning with 0 precedes
the word 1 over the alphabet {0, 1}. The increasing order is therefore often preferred: here,
shorter words precede longer ones and words of the same length are ordered lexicographically.
This is the ordering of {0, 1}∗ we get when we write up the natural numbers in the binary
number system.
The set of real numbers will be denoted by R, the set of integers by Z and the set of rational
numbers (fractions) by Q. The sign of the set of non-negative real (integer, rational) numbers
is R+ (Z+ , Q+ ). When the base of a logarithm will not be indicated it will be understood to
be 2.
Let f and g be two real (or even complex) functions defined over the natural numbers. We
write
f = O(g)
if there is a constant c > 0 such that for all n large enough we have |f (n)| ≤ c|g(n)|. We write
f = o(g)
if f is 0 only at a finite number of places and f (n)/g(n) → 0 if n → ∞. We will also use
sometimes an inverse of the big O notation: we write
f = Ω(g)
if g = O(f ). The notation
f = Θ(g)
means that both f = O(g) and g = O(f ) hold, i.e. there are constants c1 , c2 > 0 such that
for all n large enough we have c1 g(n) ≤ f (n) ≤ c2 g(n). We will also use this notation within
formulas. Thus,
(n + 1)2 = n2 + O(n)

2
means that (n + 1)2 can be written in the form n2 + R(n) where R(n) = O(n2 ). Keep in mind
that in this kind of formula, the equality sign is not symmetrical. Thus, O(n) = O(nn ) but
O(n2 ) = O(n). When such formulas become too complex it is better to go back to some more
explicit notation.

0.1 Exercise Is it true that 1 + 2 + · · · + n = O(n3 )? Can you make this statement sharper?
♦

1 Models of Computation
1.1 Introduction
In this section, we will treat the concept of “computation” or algorithm. This concept is
fundamental for our subject, but we will not define it formally. Rather, we consider it an
intuitive notion, which is amenable to various kinds of formalization (and thus, investigation
from a mathematical point of view).
An algorithm means a mathematical procedure serving for a computation or construction
(the computation of some function), and which can be carried out mechanically, without think-
ing. This is not really a definition, but one of the purposes of this course is to demonstrate that
a general agreement can be achieved on these matters. (This agreement is often formulated
as Church’s thesis.) A program in the Pascal (or any other) programming language is a good
example of an algorithm specification. Since the “mechanical” nature of an algorithm is its most
important feature, instead of the notion of algorithm, we will introduce various concepts of a
mathematical machine.
Mathematical machines compute some output from some input. The input and output can
be a word (finite sequence) over a fixed alphabet. Mathematical machines are very much like
the real computers the reader knows but somewhat idealized: we omit some inessential features
(e.g. hardware bugs), and add an infinitely expandable memory.
Here is a typical problem we often solve on the computer: Given a list of names, sort them
in alphabetical order. The input is a string consisting of names separated by commas: Bob,
Charlie, Alice. The output is also a string: Alice, Bob, Charlie. The problem is to compute a
function assigning to each string of names its alphabetically ordered copy.
In general, a typical algorithmic problem has infinitely many instances, whci then have
arbitrarily large size. Therefore we must consider either an infinite family of finite computers of
growing size, or some idealized infinite computer. The latter approach has the advantage that
it avoids the questions of what infinite families are allowed.
Historically, the first pure infinite model of computation was the Turing machine, intro-
duced by the English mathematician Turing in 1936, thus before the invention of programable
computers. The essence of this model is a central part that is bounded (with a structure inde-
pendent of the input) and an infinite storage (memory). (More exactly, the memory is an infinite
one-dimensional array of cells. The control is a finite automaton capable of making arbitrary
local changes to the scanned memory cell and of gradually changing the scanned position.) On
Turing machines, all computations can be carried out that could ever be carried out on any
other mathematical machine-models. This machine notion is used mainly in theoretical inves-

3
tigations. It is less appropriate for the definition of concrete algorithms since its description is
awkward, and mainly since it differs from existing computers in several important aspects.
The most important weakness of the Turing machine in comparison real computers is that its
memory is not accessible immediately: in order to read a distant memory cell, all intermediate
cells must also be read. This is remedied by the Random Access Machine (RAM). The RAM can
reach an arbitrary memory cell in a single step. It can be considered a simplified model of real
world computers along with the abstraction that it has unbounded memory and the capability
to store arbitrarily large integers in each of its memory cells. The RAM can be programmed in
an arbitrary programming language. For the description of algorithms, it is practical to use the
RAM since this is closest to real program writing. But we will see that the Turing machine and
the RAM are equivalent from many points of view; what is most important, the same functions
are computable on Turing machines and the RAM.
Despite their seeming theoretical limitations, we will consider logic circuits as a model of
computation, too. A given logic circuit allows only a given size of input. In this way, it can solve
only a finite number of problems; it will be, however, evident, that for a fixed input size, every
function is computable by a logical circuit. If we restrict the computation time, however, then
the difference between problems pertaining to logic circuits and to Turing-machines or the RAM
will not be that essential. Since the structure and work of logic circuits is the most transparent
and tractable, they play very important role in theoretical investigations (especially in the proof
of lower bounds on complexity).
If a clock and memory registers are added to a logic circuit we arrive at the interconnected
finite automata that form the typical hardware components of today’s computers.
Let us note that a fixed finite automaton, when used on inputs of arbitrary size, can compute
only very primitive functions, and is not an adequate computation model.
One of the simplest models for an infinite machine is to connect an infinite number of similar
automata into an array. This way we get a cellular automaton.
The key notion used in discussing machine models is simulation. This notion will not be
defined in full generality, since it refers also to machines or languages not even invented yet.
But its meaning will be clear. We will say that machine M simulates machine N if the internal
states and transitions of N can be traced by machine M in such a way that from the same
inputs, M computes the same outputs as N .

4
1.2 Finite automata
A finite automaton is a very simple and very general computing device. All we assume that if
it gets an input, then it changes its internal state and issues an output. More exactly, a finite
automaton has

— an input alphabet, which is a finite set Σ,

— an output alphabet, which is another finite set Σ , and

— a set Γ of internal states, which is also finite.

To describe a finite automaton, we need to specify, for every input a ∈ Σ and state s ∈ Γ,
the output α(a, s) ∈ Σ and the new state ω(a, s) ∈ Γ. To make the behavior of the automata
well-defined, we specify a starting state START.
At the beginning of a computation, the automaton is in state s0 = START. The input to
the computation is given in the form of a string a1 a2 . . . an ∈ Σ∗ . The first input letter a1 takes
the automaton to state s1 = ω(a1 , s0 ); the next input letter takes it into state s2 = ω(a2 , s1 )
etc. The result of the computation is the string b1 b2 . . . bn , where bk = α(ak , sk−1 ) is the output
at the k-th step.
Thus a finite automaton can be described as a 6-tuple Σ, Σ , Γ, α, ω, s0
, where Σ, Σ , Γ are
finite sets, α : Σ × Γ → Σ and ω : Σ × Γ → Γ are arbitrary mappings, and START ∈ Γ.
Remarks. 1. There are many possible variants of this notion, which are essentially equivalent.
Often the output alphabet and the output signal are omitted. In this case, the result of the
computation is read off from the state of the automaton at the end of the computation.
In the case of automata with output, it is often convenient to assume that Σ contains the
blank symbol ∗; in other words, we allow that the automaton does not give an output at certain
steps.
2. Your favorite PC can be modelled by a finite automaton where the input alphabet
consists of all possible keystrokes, and the output alphabet consists of all texts that it can write
on the screen following a keystroke (we ignore the mouse, ports, floppy drives etc.) Note that the
number of states is more than astronomical (if you have 1 GB of disk space, than this automaton
10
has something like 210 states). At the cost of allowing so many states, we could model almost
anything as a finite automaton. We’ll be interested in automata where the number of states is
much smaller - usually we assume it remains bounded while the size of the input is unbounded.
Every finite automaton can be described by a directed graph. The nodes of this graph are
the elements of Γ, and there is an edge labelled (a, b) from state s to state s if α(a, s) = b
and ω(a, s) = s . The computation performed by the automaton, given an input a1 a2 . . . an ,
corresponds to a directed path in this graph starting at node START, where the first labels of
the edges on this path are a1 , a2 , . . . , an . The second labels of the edges give the result of the
computation (figure 1.1).

(1.1) Example Let us construct an automaton that corrects quotation marks in a text in the
following sense: it reads a text character-by-character, and whenever it sees a quotation like
” . . . ”, it replaces it by “. . .”. All the automaton has to remember is whether it has seen an even

5
 (b,x) (a,x)
(c,x) (c,y)
(a,x)
(a,y) (b,y) (b,y)
START (c,x)

aabcabc yyxyxyx

Figure 1.1: A finite automaton

(’’,‘‘)

... ...
(z,z) (a,a) (a,a) (z,z)
START (’’,’’) OPEN

Figure 1.2: An automaton correcting quotation marks

or an odd number of ” symbols. So it will have two states: START and OPEN (i.e., quotation
is open). The input alphabet consists of whatever characters the text uses, including ”. The
output alphabet is the same, except that instead of ” we have two symbols “ and ”. Reading
any character other than ”, the automaton outputs the same symbol and does not change its
state. Reading ”, it outputs “ if it is in state START and outputs ” if it is in state OPEN; and
it changes its state (figure 1.2). ♦

1.1 Exercise Construct a finite automaton with a bounded number of states that receives
two integers in binary and outputs their sum. The automaton gets alternatingly one bit of each
number, starting from the right. If we get past the first bit of one of the inputs numbers, a special
symbol • is passed to the automaton instead of a bit; the input stops when two consecutive •
symbols are occur. ♦

1.2 Exercise Construct a finite automaton with as few states as possible that receives the
digits of an integer in decimal notation, starting from the left, and the last output is YES if the
number is divisible by 7 and NO if it is not. ♦

1.3 Exercise (a) For a fixed positive integer n, construct a finite automaton that reads a word
of length 2n, and its last output is YES if the first half of the word is the same as the second
half, and NO otherwise. (b) Prove that the automaton must have at least 2n states. ♦

6
1.4 Exercise Prove that there is no finite automaton that, for an input in {0, 1}∗ starting
with a “1”, would decide if this binary number is a prime. ♦

7
1.3 The Turing machine
1.3.1 The notion of a Turing machine
Informally, a Turing machine is a finite automaton equipped with an unbounded memory. This
memory is given in the form of one or more tapes, which are infinite in both directions. The tapes
are divided into an infinite number of cells in both directions. Every tape has a distinguished
starting cell which we will also call the 0th cell. On every cell of every tape, a symbol from a
finite alphabet Σ can be written. With the exception of finitely many cells, this symbol must
be a special symbol ∗ of the alphabet, denoting the “empty cell”.
To access the information on the tapes, we supply each tape by a read-write head. At every
step, this sits on a field of the tape.
The read-write heads are connected to a control unit, which is a finite automaton. Its possible
states form a finite set Γ. There is a distinguished starting state “START” and a halting state
“STOP”. Initially, the control unit is in the “START” state, and the heads sit on the starting
cells of the tapes. In every step, each head reads the symbol in the given cell of the tape, and
sends it to the control unit. Depending on these symbols and on its own state, the control unit
carries out three things:

— it sends a symbol to each head to overwrite the symbol on the tape (in particular, it can
give the direction to leave it unchanged);

— it sends one of the commands “MOVE RIGHT”, “MOVE LEFT” or “STAY” to each
head;

— it makes a transition into a new state (this may be the same as the old one);

Of course, the heads carry out these commands, which completes one step of the computa-
tion. The machine halts when the control unit reaches the “STOP” state.
While the above informal description uses some engineering jargon, it is not difficult to
translate it into purely mathematical terms. For our purposes, a Turing machine is completely
specified by the following data: T = k, Σ, Γ, α, β, γ
, where k ≥ 1 is a natural number, Σ and
Γ are finite sets, ∗ ∈ Σ ST ART, ST OP ∈ Γ, and α, β, γ are arbitrary mappings:

α: Γ × Σk → Γ,
β: Γ × Σk → Σk ,
γ: Γ × Σk → {−1, 0, 1}k .

Here α specifiess the new state, β gives the symbols to be written on the tape and γ specifies
how the heads move.
In what follows we fix the alphabet Σ and assume that it contains, besides the blank symbol
∗, at least two further symbols, say 0 and 1 (in most cases, it would be sufficient to confine
ourselves to these two symbols).

8
 Under the input of a Turing machine, we mean the k words initially written on the tapes.
We always assume that these are written on the tapes starting at the 0 field. Thus, the input
of a k-tape Turing machine is an ordered k-tuple, each element of which is a word in Σ∗ . Most
frequently, we write a non-empty word only on the first tape for input. If we say that the input
is a word x then we understand that the input is the k-tuple (x, ∅, . . . , ∅).
The output of the machine is an ordered k-tuple consisting of the words on the tapes.
Frequently, however, we are really interested only in one word, the rest is “garbage”. If we say
that the output is a single word and don’t specify which, then we understand the word on the
last tape.
It is practical to assume that the input words do not contain the symbol ∗. Otherwise, it
would not be possible to know where is the end of the input: a simple problem like “find out the
length of the input” would not be solvable: no matter how far the head has moved, it could not
know whether the input has already ended. We denote the alphabet Σ \ {∗} by Σ0 . (Another
solution would be to reserve a symbol for signalling “end of input” instead.) We also assume
that during its work, the Turing machine reads its whole input; with this, we exclude only trivial
cases.
Turing machines are defined in many different, but from all important points of view equiv-
alent, ways in different books. Often, tapes are infinite only in one direction; their number can
virtually always be restricted to two and in many respects even to one; we could assume that
besides the symbol ∗ (which in this case we identify with 0) the alphabet contains only the
symbol 1; about some tapes, we could stipulate that the machine can only read from them or
can only write onto them (but at least one tape must be both readable and writable) etc. The
equivalence of these variants (from the point of view of the computations performable on them)
can be verified with more or less work but without any greater difficulty. In this direction, we
will prove only as much as we need, but this should give a sufficient familiarity with the tools
of such simulations.

1.5 Exercise Construct a Turing machine that computes the following functions:
(a) x1 . . . xm → xm . . . x1 .

(b) x1 . . . xm → x1 . . . xm x1 . . . xm .

(c) x1 . . . xm → x1 x1 . . . xm xm .

(d) for an input of length m consisting of all 1’s, the binary form of m; for all other inputs,
for all other inputs, it never halts.
♦

1.6 Exercise Assume that we have two Turing machines, computing the functions f : Σ∗0 → Σ∗0
and g : Σ∗0 → Σ∗0 . Construct a Turing machine computing the function f ◦ g. ♦

1.7 Exercise Construct a Turing machine that makes 2|x| steps for each input x. ♦

1.8 Exercise Construct a Turing machine that on input x, halts in finitely many steps if and
only if the symbol 0 occurs in x. ♦

9
 3 . 1 4 1 5 9 * * * * * * *

C O M P U T E * P I * A N D

1 + 1 / 4 + 1 / 9 + 1 / 1 6

CU

Figure 1.3: A Turing maching with three tapes

1.3.2 Universal Turing machines

Based on the preceding, we can notice a significant difference between Turing machines and real
computers: for the computation of each function, we constructed a separate Turing machine,
while on real program-controlled computers, it is enough to write an appropriate program. We
will now show that Turing machines can also be operated this way: a Turing machine can be
constructed on which, using suitable “programs”, everything is computable that is computable
on any Turing machine. Such Turing machines are interesting not just because they are more
like programable computers but they will also play an important role in many proofs.
Let T = k + 1, Σ, ΓT , αT , βT , γT
and S = k, Σ, ΓS , αS , βS , γS
be two Turing machines
(k ≥ 1). Let p ∈ Σ∗0 . We say that T simulates S with program p if for arbitrary words
x1 , . . . , xk ∈ Σ∗0 , machine T halts in finitely many steps on input (x1 , . . . , xk , p) if and only if S
halts on input (x1 , . . . , xk ) and if at the time of the stop, the first k tapes of T each have the
same content as the tapes of S.
We say that a (k + 1)-tape Turing machine is universal (with respect to k-tape Turing
machines) if for every k-tape Turing machine S over Σ, there is a word (program) p with which
T simulates S.

(1.1) Theorem For every number k ≥ 1 and every alphabet Σ there is a (k + 1)-tape universal
Turing machine.

Proof The basic idea of the construction of a universal Turing machine is that on tape k + 1,
we write a table describing the work of the Turing machine S to be simulated. Besides this, the
universal Turing machine T writes it up for itself, which state of the simulated machine S is

10
currently in (even if there is only a finite number of states, the fixed machine T must simulate
all machines S, so it “cannot keep in its head” the states of S). In each step, on the basis of
this, and the symbols read on the other tapes, it looks up in the table the state that S makes
the transition into, what it writes on the tapes and what moves the heads make.
First, we give the construction using k + 2 tapes. For the sake of simplicity, assume that Σ
contains the symbols “0”, “1”, and “–1”. Let S = k, Σ, ΓS , αS , βS , γS
be an arbitrary k-tape
Turing machine. We identify each element of the state set ΓS \ {STOP} with a word of length
r over the alphabet Σ∗0 . Let the “code” of a given position of machine S be the following word:

gh1 . . . hk αS (g, h1 , . . . , hk )βS (g, h1 , . . . , hk )γS (g, h1 , . . . , hk )

where g ∈ ΓS is the given state of the control unit, and h1 , . . . , hk ∈ Σ are the symbols read by
each head. We concatenate all such words in arbitrary order and obtain so the word aS . This
is what we write on tape k + 1; while on tape k + 2, we write a state of machine S, initially the
name of the START state.
Further, we construct the Turing machine T  which simulates one step or S as follows. On
tape k + 1, it looks up the entry corresponding to the state remembered on tape k + 2 and the
symbols read by the first k heads, then it reads from there what is to be done: it writes the new
state on tape k + 2, then it lets its first k heads write the appropriate symbol and move in the
appropriate direction.
For the sake of completeness, we also define machine T  formally, but we also make
some concession to simplicity in that we do this only for the case k = 1. Thus, the ma-
chine has three heads. Besides the obligatory “START” and “STOP” states, let it also
have states NOMATCH-ON, NOMATCH-BACK-1, NOMATCH-BACK-2, MATCH-BACK,
WRITE, MOVE and AGAIN. Let h(i) denote the symbol read by the i-th head (1 ≤ i ≤ 3).
We describe the functions α, β, γ by the table in Figure 1.4 (wherever we do not specify a new
state the control unit stays in the old one).
In the typical run in Figure 1.5, the numbers on the left refer to lines in the above program.
The three tapes are separated by triple vertical lines, and the head positions are shown by
underscores.
Now return to the proof of Theorem 1.1. We can get rid of the (k + 2)-nd tape easily: its
contents (which is always just r cells) will be placed on cells −1, −2, . . . , −r. It seems, however,
that we still need two heads on this tape: one moves on its positive half, and one on the negative
half (they don’t need to cross over). We solve this by doubling each cell: the original symbol
stays in its left half, and in its right half there is a 1 if the corresonding head would just be
there (the other right half cells stay empty). It is easy to describe how a head must move on
this tape in order to be able to simulate the movement of both original heads.

1.9 Exercise Write a simulation of a Turing machine with a doubly infinite tape by a Turing
machine with a tape that is infinite only in one direction. ♦

1.10 Exercise Show that if we simulate a k-tape machine on the (k + 1)-tape universal Turing
machine, then on an arbitrary input, the number of steps increases only by a multiplicative
factor proportional to the length of the simulating program. ♦

11
START:
1: if h(2) = h(3) = ∗ then 2 and 3 moves right;
2: if h(2), h(3) = ∗ and h(2) = h(3) then “NOMATCH-ON” and 2,3 move right;
8: if h(3) = ∗ and h(2) = h(1) then “NOMATCH-BACK-1” and 2 moves right, 3 moves
left;
9: if h(3) = ∗ and h(2) = h(1) then “MATCH-BACK”, 2 moves right and 3 moves left;
18: if h(3) = ∗ and h(2) = ∗ then “STOP”;
NOMATCH-ON:
3: if h(3) = ∗ then 2 and 3 move right;
4: if h(3) = ∗ then “NOMATCH-BACK-1” and 2 moves right, 3 moves left;
NOMATCH-BACK-1:
5: if h(3) = ∗ then 3 moves left, 2 moves right;
6: if h(3) = ∗ then “NOMATCH-BACK-2”, 2 moves right;
NOMATCH-BACK-2:
7: “START”, 2 and 3 moves right;
MATCH-BACK:
10: if h(3) = ∗ then 3 moves left;
11: if h(3) = ∗ then “WRITE-STATE” and 3 moves right;
WRITE:
12: if h(3) = ∗ then 3 writes the symbol h(2) and 2,3 moves right;
13: if h(3) = ∗ then “MOVE”, head 1 writes h(2), 2 moves right and 3 moves left;
MOVE:
14: “AGAIN”, head 1 moves h(2);
AGAIN:
15: if h(2) = ∗ and h(3) = ∗ then 2 and 3 move left;
16: if h(2) = ∗ but h(3) = ∗ then 2 moves left;
17: if h(2) = h(3) = ∗ then “START”, and 2,3 move right.

Figure 1.4: A universal Turing machine

12
line Tape 3 Tape 2 Tape 1
1 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
2 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
3 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
4 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
5 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
6 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
7 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
1 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
8 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
9 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
10 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
11 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
12 ∗010∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
13 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗11∗
14 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗
15 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗
16 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗
17 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗
1 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗
18 ∗111∗ ∗ 000 0 000 0 0 010 0 000 0 0 010 1 111 0 1 ∗ ∗01∗

Figure 1.5: Example run of the universal Turing machine

13
 simulates 5th cell simulates 7th cell
of first tape of second tape

? ?
q q q H1 s5 t5 s6 t6 H2 s7 t7 q q q

6 6

simulated simulated
head 1 head 2

Figure 1.6: One tape simulating two tapes

1.11 Exercise Let T and S be two one-tape Turing machines. We say that T simulates the
work of S by program p (here p ∈ Σ∗0 ) if for all words x ∈ Σ∗0 , machine T halts on input p ∗ x
in a finite number of steps if and only if S halts on input x and at halting, we find the same
content on the tape of T as on the tape of S. Prove that there is a one-tape Turing machine T
that can simulate the work of every other one-tape Turing machine in this sense. ♦

1.3.3 More tapes versus one tape

Our next theorem shows that, in some sense, it is not essential, how many tapes a Turing
machine has.

(1.2) Theorem For every k-tape Turing machine S there is a one-tape Turing machine T
which replaces S in the following sense: for every word x ∈ Σ∗0 , machine S halts in finitely many
steps on input x if and only if T halts on input x, and at halt, the same is written on the last
tape of S as on the tape of T . Further, if S makes N steps then T makes O(N 2 ) steps.

Proof We must store the content of the tapes of S on the single tape of T . For this, first we
“stretch” the input written on the tape of T : we copy the symbol found on the i-th cell onto
the (2ki)-th cell. This can be done as follows: first, starting from the last symbol and stepping
right, we copy every symbol right by 2k positions. In the meantime, we write ∗ on positions
1, 2, . . . , 2k − 1. Then starting from the last symbol, it moves every symbol in the last block of
nonblanks 2k positions to right, etc.
Now, position 2ki + 2j − 2 (1 ≤ j ≤ k) will correspond to the i-th cell of tape j, and position
2k + 2j − 1 will hold a 1 or ∗ depending on whether the corresponding head of S, at the step
corresponding to the computation of S, is scanning that cell or not. Also, let us mark by a 0
the first even-numbered cell of the empty ends of the tapes. Thus, we assigned a configuration
of T to each configuration of the computation of S.

14
 Now we show how T can simulate the steps of S. First of all, T “keeps in its head” which
state S is in. It also knows what is the remainder of the number of the cell modulo 2k scanned
by its own head. Starting from right, let the head now make a pass over the whole tape. By
the time it reaches the end it knows what are the symbols read by the heads of S at this step.
From here, it can compute what will be the new state of S what will its heads write and wich
direction they will move. Starting backwards, for each 1 found in an odd cell, it can rewrite
correspondingly the cell before it, and can move the 1 by 2k positions to the left or right if
needed. (If in the meantime, it would pass beyond the beginning or ending 0, then it would
move that also by 2k positions in the appropriate direction.)
When the simulation of the computation of S is finished, the result must still be “com-
pressed”: the content of cell 2ki must be copied to cell i. This can be done similarly to the
initial “stretching”.
Obviously, the above described machine T will compute the same thing as S. The number of
steps is made up of three parts: the times of “stretching”, the simulation and “compression”. Let
M be the number of cells on machine T which will ever be scanned by the machine; obviously,
M = O(N ). The “stretching” and “compression” need time O(M 2 ). The simulation of one step
of S needs O(M ) steps, so the simulation needs O(M N ) steps. All together, this is still only
O(N 2 ) steps.
As we have seen, the simulation of a k-tape Turing machine by a 1-tape Turing machine
is not completely satisfactory: the number of steps increases quadratically. This is not just a
weakness of the specific construction we have described; there are computational tasks that can
be solved on a 2-tape Turing machine in some N steps but any 1-tape Turing machine needs
N 2 steps to solve them. We describe a simple example of such a task.
A palindrome is a word (say, over the alphabet {0, 1}) that does not change when reversed;
i.e., x1 . . . xn is a palindrome iff xi = xn−i+1 for all i. Let us analyze the task of recognizing a
palindrome.

(1.3) Theorem (a) There exists a 2-tape Turing machine that decides whether the input word
x ∈ {0, 1}n is a palindrome in O(n) steps. (b) Every one-tape Turing machine that decides
whether the input word x ∈ {0, 1}n is a palindrome has to make Ω(n2 ) steps in the worst case.

Proof Part (a) is easy: for example, we can copy the input on the second tape in n steps, then
move the first head to the beginning of the input in n further steps (leave the second head at the
end of the word), and compare x1 with xn , x2 with xn−1 , etc., in another n steps. Altogether,
this takes only 3n steps.
Part (b) is more difficult to prove. Consider any one-tape Turing machine that recognizes
palindromes. To be specific, say it ends up with writing a “1” on the starting field of the tape
if the input word is a palindrome, and a “0” if it is not. We are going to argue that for every
n, on some input of length n, the machine will have to make Ω(n2 ) moves.
It will be convenient to assume that n is divisible by 3 (the argument is very similar in the
general case). Let k = n/3. We restrict the inputs to words in which the middle third is all 0,
i.e., to words of the form x1 . . . xk 0 . . . 0x2k+1 . . . xn . (If we can show that already among such
words, there is one for which the machine must work for Ω(n2 ) time, we are done.)

15
 Fix any j such that k ≤ j ≤ 2k. Call the dividing line between fields j and j + 1 of the
tape the cut after j. Let us imagine that we have a little deamon sitting on this, and recording
the state of the central unit any time the head passes between these fields. At the end of the
computation, we get a sequence g1 g2 . . . gt of elements of Γ (the length t of the sequence may
be different for different inputs), the j-log of the given input. The key to proof is the following
observation.
Lemma. Let x = x1 . . . xk 0 . . . 0xk . . . x1 and y = y1 . . . yk 0 . . . 0yk . . . y1 be two different palin-
dromes and k ≤ j ≤ 2k. Then their j-logs are different.
Proof of the lemma. Suppose that the j-logs of x and y are the same, say g1 . . . gt . Consider
the input z = x1 . . . xk 0 . . . 0yk . . . y1 . Note that in this input, all the xi are to the left from the
cut and all the yi are to the right.
We show that the machine will conclude that z is a palindrome, which is a contradiction.
What happens when we start the machine with input z? For a while, the head will move
on the fields left from the cut, and hence the computation will proceed exactly as with input
x. When the head first reaches field j + 1, then it is in state g1 by the j-log of x. Next, the
head will spend some time to the right from the cut. This part of the computation will be
indentical with the corresponding part of the computation with input y: it starts in the same
state as the corresponding part of the computation of y does, and reads the same characters
from the tape, until the head moves back to field j again. We can follow the computation on
input z similarly, and see that the portion of the computation during its m-th stay to the left
of the cut is identical with the corresponding portion of the computation with input x, and the
portion of the computation during its m-th stay to the right of the cut is identical with the
corresponding portion of the computation with input y. Since the computation with input x
ends with writing a “1” on the starting field, the computation with input z ends in the same
way. This is a contradiction.
Now we return to the proof of the theorem. For a given m, the number of different j-logs of
length less than m is at most
|Γ|m − 1
1 + |Γ| + |Γ|2 + . . . + |Γ|m−1 = < 2|Γ|m−1 .
|Γ| − 1
This is true for any choice of j; hence the number of palindromes whose j-log for some j has
length less than m is at most
2(k + 1)|Γ|m−1 .
There are 2k palindromes of the type considered, and so the number of palindromes for whose
j-logs have length at least m for all j is at least
2k − 2(k + 1)|Γ|m−1 . (1.4)
So if we choose m so that this number is positive, then there will be a palindrome for which the
j-log has length at least m for all j. This implies that the deamons record at least (k + 1)m
moves, so the computation takes at least (k + 1)(m + 1) steps.
It is easy to check that the choice m = n/(6 log |Γ|) makes (1.4) positive, and so we have
found an input for which the computation takes at least (k + 1)m > n2 /(6 log |Γ|) steps.

16
1.12 Exercise In the simulation of k-tape machines by one-tape machines given above the
finite control of the simulating machine T was somewhat bigger than that of the simulated
machine S: moreover, the number of states of the simulating machine depends on k. Prove that
this is not necessary: there is a one-tape machine that can simulate arbitrary k-tape machines.
♦
∗ (1.13)Exercise Show that every k-tape Turing machine can be simulated by a two-tape one
in such a way that if on some input, the k-tape machine makes N steps then the two-tape one
makes at most O(N log N ).
[Hint: Rather than moving the simulated heads, move the simulated tapes! (Hennie-Stearns)]
♦

1.14 Exercise Two-dimensional tape.

(a) Define the notion of a Turing machine with a two-dimensional tape.
(b) Show that a two-tape Turing machine can simulate a Turing machine with a two-
dimensional tape. [Hint: Store on tape 1, with each symbol of the two-dimensional tape,
the coordinates of its original position.]
(c) Estimate the efficiency of the above simulation.
♦
∗ (1.15) Exercise Let f : Σ∗0 → Σ∗0 be a function. An online Turing machine contains, besides
the usual tapes, two extra tapes. The input tape is readable only in one direction, the output
tape is writeable only in one direction. An online Turing machine T computes function f if in a
single run, for each n, after receiving n symbols x1 , . . . , xn , it writes f (x1 . . . xn ) on the output
tape, terminated by a blank.
Find a problem that can be solved more efficiently on an online Turing machinw with a
two-dimensional working tape than with a one-dimensional working tape.
√
[Hint: On a two-dimensional tape, any one of n bits can be accessed in n steps. To exploit
this, let the input represent a sequence of operations on a “database”: insertions and queries,
and let f be the interpretation of these operations.] ♦

1.16 Exercise Tree tape.

(a) Define the notion of a Turing machine with a tree-like tape.
(b) Show that a two-tape Turing machine can simulate a Turing machine with a tree-like tape.
[Hint: Store on tape 1, with each symbol of the two-dimensional tape, an arbitrary number
identifying its original position and the numbers identifying its parent and children.]
(c) Estimate the efficiency of the above simulation.
(d) Find a problem which can be solved more efficiently with a tree-like tape than with any
finite-dimensional tape.
♦

17
1.4 The Random Access Machine
Trying to design Turing machines for different tasks, one notices that a Turing machine spends
a lot of its time by just sending its read-write heads from one end of the tape to the other.
One might design tricks to avoid some of this, but following this line we would drift farther
and farther away from real-life computers, which have a “random-access” memory, i.e., which
can access any field of their memory in one step. So one would like to modify the way we have
equipped Turing machines with memory so that we can reach an arbitrary memory cell in a
single step.
Of course, the machine has to know which cell to access, and hence we have to assigne
addresses to the cells. We want to retain the feature that the memory is unbounded; hence
we allow arbitrary integers as addresses. The address of the cell to access must itself be stored
somewhere; therefore, we allow arbitrary integers to be stored in each cell (rather than just a
single element of a fintie alphabet, as in the case of Turing machines).
Finally, we make the model more similar to everyday machines by making it programmable
(we could also say that we define the analogue of a universal Turing machine). This way we get
the notion of a Random Access Machine or RAM machine.
Now let us be more precise. The memory of a Random Access Machine is a doubly infinite
sequence . . . x[−1], x[0], x[1], . . . of memory registers. Each register can store an arbitrary integer.
At any given time, only finitely many of the numbers stored in memory are different from 0.
The program store is a (one-way) infinite sequence of registers called lines. We write here
a program of some finite length, in a certain programming language similar to the assembly
language of real machines. It is enough, for example, to permit the following statements:
x[i]:=0; x[i]:=x[i]+1; x[i]:=x[i]-1;
x[i]:=x[i]+x[j]; x[i]:=x[i]-x[j];
x[i]:=x[x[j]]; x[x[i]]:=x[j];
IF x[i]≤ 0 THEN GOTO p.

Here, i and j are the addresses of memory registers (i.e. arbitrary integers), p is the address
of some program line (i.e. an arbitrary natural number). The instruction before the last one
guarantees the possibility of immediate access. With it, the memory behaves as an array in a
conventional programming language like Pascal. The exact set of basic instructions is important
only to the extent that they should be sufficiently simple to implement, expressive enough to
make the desired computations possible, and their number be finite. For example, it would be
sufficient to allow the values −1, −2, −3 for i, j. We could also omit the operations of addition
and subtraction from among the elementary ones, since a program can be written for them. On
the other hand, we could also include multiplication, etc.
The input of the Random Access Machine is a finite sequence of natural numbers written
into the memory registers x[0], x[1], . . .. The Random Access Machine carries out an arbitrary
finite program. It stops when it arrives at a program line with no instruction in it. The output
is defined as the content of the registers x[i] after the program stops.
It is easy to write RAM subroutines for simple tasks that repeatedly occur in programs
solving more difficult things. Several of these are given as exercises. Here we discuss three tasks
that we need later on in this chapter.

18
(1.1) Example [Value assignment] Let i and j be two integers. Then the assignment
x[i]:=j
can be realized by the RAM program
x[i]:=0 ⎫
x[i]:=x[i]+1; ⎪
⎪
⎬
.
. j times
. ⎪
⎪
x[i]:=x[i]+1; ⎭
if j is positive, and
x[i]:=0 ⎫
x[i]:=x[i]-1; ⎪
⎪
⎬
.
. |j| times
. ⎪
⎪
x[i]:=x[i]-1; ⎭
if j is negative. ♦
(1.2) Example [Addition of a constant] Let i and j be two integers. Then the statement
x[i]:=x[i]+j
can be realized in the same way as in the previous example, just omitting the first row. ♦
(1.3) Example [Multiple branching] Let p0 , p1 , . . . , pr be indices of program rows, and suppose
that we know that the contents of register i satisfies 0 ≤ x[i] ≤ r. Then the statement
GOTO px[i]
can be realized by the RAM program
IF x[i]≤0 THEN GOTO p0 ;
x[i]:=x[i]-1:
IF x[i]≤0 THEN GOTO p1 ;
x[i]:=x[i]-1:
.
.
.
IF x[i]≤0 THEN GOTO pr .
(Attention must be paid when including this last program segment in a program, since it changes
the content of xi. If we need to preserve the content of x[i], but have a “scratch” register, say
x[−1], then we can do
x[-1]:=x[i];
IF x[-1]≤0 THEN GOTO p0 ;
x[-1]:=x[-1]-1:
IF x[-1]≤0 THEN GOTO p1 ;
x[-1]:=x[-1]-1:
.
.
.
IF x[-1]≤0 THEN GOTO pr .
If we don’t have a scratch register than we have to make room for one; since we won’t have
to go into such details, we leave it to the exercises. ♦

19
 Now we show that the RAM and Turing machines can compute essentially the same func-
tions, and their running times do not differ too much either. Let us consider (for simplicity) a
1-tape Turing machine, with alphabet {0, 1, 2}, where (deviating from earlier conventions but
more practically here) let 0 stand for the blank space symbol.
Every input x1 . . . xn of the Turing machine (which is a 1–2 sequence) can be interpreted as
an input of the RAM in two different ways: we can write the numbers n, x1 , . . . , xn into the
registers x[0], . . . , x[n], or we could assign to the sequence x1 . . . xn a single natural number by
replacing the 2’s with 0 and prefixing a 1. The output of the Turing machine can be interpreted
similarly to the output of the RAM.
We will consider the first interpretation first.

(1.4) Theorem For every (multitape) Turing machine over the alphabet {0, 1, 2}, one can
construct a program on the Random Access Machine with the following properties. It computes
for all inputs the same outputs as the Turing machine and if the Turing machine makes N steps
then the Random Access Machine makes O(N ) steps with numbers of O(log N ) digits.

Proof Let T = 1, {0, 1, 2}, Γ, α, β, γ

. Let Γ = {1, . . . , r}, where 1 = START and r = STOP.
During the simulation of the computation of the Turing machine, in register 2i of the RAM we
will find the same number (0,1 or 2) as in the i-th cell of the Turing machine. Register x[1] will
remember where is the head on the tape, and the state of the control unit will be determined
by where we are in the program.
Our program will be composed of parts Qij simulating the action of the Turing machine
when in state i and reading symbol j (1 ≤ i ≤ r − 1, 0 ≤ j ≤ 2) and lines Pi that jump to Qi,j
if the Turing machine is in state i and reads symbol j. Both are easy to realize. Pi is simply
GOTO Qi,x[1] ;
for 1 ≤ i ≤ i − 1; the program part Pr consists of a single empty program line. The program
parts Qij are only a bit more complicated:
x[x[1]]:= β(i, j);
x[1]:=x[1]+2γ(i, j);
GOTO Pα(i,j) ;
The program itself looks as follows.
x[1]:=0;
P1
P2
.
.
.
Pr
Q1,0
.
.
.
Qr−1,2
With this, we have described the simulation of the Turing machine by the RAM. To analyze
the number of steps and the size of the number used, it is enough to note that in N steps, the

20
Turing machine can write anything in at most O(N ) registers, so in each step of the Turing
machine we work with numbers of length O(log N ).
Another interpretation of the input of the Turing machine is, as mentioned above, to view
the input as a single natural number, and to enter it into the RAM as such. This number a
is thus in register x[0]. In this case, what we can do is to compute the digits of a with the
help of a simple program, write these (deleting the 1 in the first position) into the registers
x[0], . . . , x[n − 1], and apply the construction described in Theorem 1.4.

(1.5) Remark In the proof of Theorem 1.4, we did not use the instruction x[i] := x[i] +
x[j]; this instruction is needed when computing the digits of the input. Even this could be
accomplished without the addition operation if we dropped the restriction on the number of
steps. But if we allow arbitrary numbers as inputs to the RAM then, without this instruction
the running time the number of steps obtained would be exponential even for very simple
problems. Let us e.g. consider the problem that the content a of register x[1] must be added
to the content b of register x[0]. This is easy to carry out on the RAM in a bounded number
of steps. But if we exclude the instruction x[i] := x[i] + x[j] then the time it needs is at least
min{|a|, |b|}. ♦

Let a program be given now for the RAM. We can interpret its input and output each as
a word in {0, 1, −, #}∗ (denoting all occurring integers in binary, if needed with a sign, and
separating them by #). In this sense, the following theorem holds.

(1.6) Theorem For every Random Access Machine program there is a Turing machine com-
puting for each input the same output. If the Random Access Machine has running time N
then the Turing machine runs in O(N 2 ) steps.

Proof We will simulate the computation of the RAM by a four-tape Turing machine. We
write on the first tape the content of registers x[i] (in binary, and with sign if it is negative). We
could represent the content of all registers (representing, say, the content 0 by the symbol “*”).
This would cause a problem, however, because of the immediate (“random”) access feature of
the RAM. More exactly, the RAM can write even into the register with number 2N using only
one step with an integer of N bits. Of course, then the content of the overwhelming majority
of the registers with smaller indices remains 0 during the whole computation; it is not practical
to keep the content of these on the tape since then the tape will be very long, and it will take
exponential time for the head to walk to the place where it must write. Therefore we will store
on the tape of the Turing machine only the content of those registers into which the RAM
actually writes. Of course, then we must also record the number of the register in question.
What we will do therefore is that whenever the RAM writes a number y into a register x[z],
the Turing machine simulates this by writing the string ##y#z to the end of its first tape. (It
never rewrites this tape.) If the RAM reads the content of some register x[z] then on the first
tape of the Turing machine, starting from the back, the head looks up the first string of form
##u#z; this value u shows what was written in the z-th register the last time. If it does not
find such a string then it treats x[z] as 0.
Each instruction of the “programming language” of the RAM is easy to simulate by an
appropriate Turing machine using only the three other tapes. Our Turing machine will be a

21
“supermachine” in which a set of states corresponds to every program line. These states form
a Turing machine which carries out the instruction in question, and then it brings the heads
to the end of the first tape (to its last nonempty cell) and to cell 0 of the other tapes. The
STOP state of each such Turing machine is identified with the START state of the Turing
machine corresponding to the next line. (In case of the conditional jump, if x[i] ≤ 0 holds,
the “supermachine” goes into the starting state of the Turing machine corresponding to line
p.) The START of the Turing machine corresponding to line 0 will also be the START of the
supermachine. Besides this, there will be yet another STOP state: this corresponds to the
empty program line.
It is easy to see that the Turing machine thus constructed simulates the work of the RAM
step-by-step. It carries out most program lines in a number of steps proportional to the number
of digits of the numbers occurring in it, i.e. to the running time of the RAM spent on it. The
exception is readout, for wich possibly the whole tape must be searched. Since the length of the
tape is N , the total number of steps is O(N 2 ).

1.17 Exercise Write a program for the RAM that for a given positive number a

(a) determines the largest number m with 2m ≤ a;

(b) computes its base 2 representation;

1.18 Exercise Let p(x) = a0 + a1 x + · · · + an xn be a polynomial with integer coefficients

a0 , . . . , an . Write a RAM program computing the coefficients of the polynomial (p(x))2
from those of p(x). Estimate the running time of your program in terms of n and K =
max{|a0 |, . . . , |an |}. ♦

1.19 Exercise Prove that if a RAM is not allowed to use the instruction x[i] := x[i] + x[j],
then adding the content a of x[1] to the content b of x[2] takes at least min{|a|, |b|} steps.
♦

1.20 Exercise Since the RAM is a single machine the problem of universality cannot be stated
in exactly the same way as for Turing machines: in some sense, this single RAM is universal.
However, the following “self-simulation” property of the RAM comes close. For a RAM program
p and input x, let R(p, x) be the output of the RAM. Let p, x
be the input of the RAM that
we obtain by writing the symbols of p one-by-one into registers 1,2,. . ., followed by a symbol #
and then by registers containing the original sequence x. Prove that there is a RAM program
u such that for all RAM programs p and inputs x we have R(u, p, x
) = R(p, x). ♦

1.21 Exercise [Pointer Machine.] After having seen finite-dimensional tapes and a tree tape,
we may want to consider a machine with a more general directed graph its storage medium.
Each cell c has a fixed number of edges, numbered 1, . . . , r, leaving it. When the head scans a
certain cell it can move to any of the cells λ(c, i) (i = 1, . . . , r) reachable from it along outgoing
edges. Since it seems impossible to agree on the best graph, we introduce a new kind of
elementary operation: to change the structure of the storage graph locally, around the scanning

22
head. Arbitrary transformations can be achieved by applying the following three operations
repeatedly (and ignoring nodes that become isolated): λ(c, i) := New, where New is a new
node; λ(c, i) := λ(λ(c, j)) and λ(λ(c, i)) := λ(c, j). A machine with this storage structure and
these three operations added to the usual Turing machine operations will be called a Pointer
Machine.
Let us call RAM’ the RAM from which the operations of addition and subtraction are
omitted, only the operation x[i] := x[i] + 1 is left. Prove that the Pointer Machine is equivalent
to RAM’, in the following sense.
For every Pointer Machine there is a RAM’ program computing for each input the same
output. If the Pointer Machine has running time N then the RAM’ runs in O(N ) steps.
For every RAM’ program there is a Pointer Machine computing for each input the same
output. If the RAM’ has running time N then the Pointer Machine runs in O(N ) steps.
Find out what Remark 1.5 says for this simulation. ♦

23
1.5 Boolean functions and Boolean circuits
A Boolean function is a mapping f : {0, 1}n → {0, 1}. The values 0,1 are sometimes identified
with the values False, True and the variables in f (x1 , . . . , xn ) are sometimes called Boolean
variables, Boolean variables or bits. In many algorithmic problems, there are n input Boolean
variables and one output bit. For example: given a graph G with N nodes, suppose we want to
decide whether it has a Hamiltonian cycle. In this case, the graph can be described with N2
Boolean variables: the nodes are numbered from 1 to N and xij (1 ≤ i < j ≤ N ) is 1 if i and
j are connected and 0 if they are not. The value of the function f (x12 , x13 , . . . , xn−1,n ) is 1 if
there is a Hamilton cycle in G and 0 if there is not. Our problem is the computation of the
value of this (implicitly given) Boolean function.
There are only four one-variable Boolean functions: the identically 0, identically 1, the
identity and the negation: x → x = 1 − x. We also use the notation ¬x. There are 16 Boolean
functions with 2 variables (because there are 24 mappings of {0, 1}2 into {0, 1}). We describe
only some of these two-variable Boolean functions: the operation of conjunction (logical AND):

1 if x = y = 1,
x∧y =
0 otherwise,
this can also be considered the common, or mod 2 multiplication, the operation of disjunction
(logical OR)

0 if x = y = 0,
x∨y =
1 otherwise,
the binary addition (logical exclusive OR of XOR)
x ⊕ y = x + y mod 2.
Among Boolean functions with several variables, one has the logical AND, OR and XOR defined
in the natural way. A more interesting function is MAJORITY, which is defined as follows:

1, if at least n/2 of the variables is 1;
MAJORITY(x1 , . . . , xn ) =
0, otherwise.
These bit-operations are connected by a number of useful identities. All three operations
AND, OR and XOR are associative and commutative. There are several distributivity proper-
ties:
x ∧ (y ∨ z) = (x ∧ y) ∨ (x ∧ z)
x ∨ (y ∧ z) = (x ∨ y) ∧ (x ∨ z)
and
x ∧ (y ⊕ z) = (x ∧ y) ⊕ (x ∧ z).
The DeMorgan identities connect negation with conjunction and disjunction:
x ∧ y = x ∨ y,
x ∨ y = x ∧ y.
Expressions composed using the operations of negation, conjunction and disjunction are called
Boolean polynomials.

24
(1.1) Lemma Every Boolean function is expressible as a Boolean polynomial.

Proof Let a1 , . . . , an ∈ {0, 1}. Let


xi if ai = 1,
zi =
xi if ai = 0,
and Ea1 ,...,an (x1 , . . . , xn ) = z1 ∧ · · · ∧ zn . Notice that Ea1 ,...,an (x1 , . . . , xn ) = 1 holds if and only
if (x1 , . . . , xn ) = (a1 , . . . , an ). Hence

f (x1 , . . . , xn ) = Ea1 ,...,an (x1 , . . . , xn ).

f (a1 ,...,an )=1

The Boolean polynomial constructed in the above proof has a special form. A Boolean
polynomial consisting of a single (negated or unnegated) variable is called a literal. We call an
elementary conjunction a Boolean polynomial in which variables and negated variables are joined
by the operation “∧”. (As a degenerate case, the constant 1 is also an elementary conjunction,
namely the empty one.) A Boolean polynomial is a disjunctive normal form if it consists of
elementary conjunctions, joined by the operation “∨”. We allow also the empty disjunction,
when the disjunctive normal form has no components. The Boolean function defined by such a
normal form is identically 0. In general, let us call a Boolean polynomial satisfiable if it is not
identically 0. As we see, a nontrivial disjunctive normal form is always satisfiable.
By a disjunctive k-normal form, we understand a disjunctive normal form in which every
conjunction contains at most k literals.

(1.2) Example Here is an important example of a Boolean function expressed by disjunctive

normal form: the selection function. Borrowing the notation from the programming language
C, we define it as

y if x = 1,
x?y : z =
z if x = 0.
It can be expressed as x?y : z = (x ∧ y) ∨ (¬x ∧ z). It is possible to construct the disjunctive
normal form of an arbitrary Boolean function by the repeated application of this example. ♦

Interchanging the role of the operations “∧” and “∨”, we can define the elementary disjunc-
tion and conjunctive normal form. The empty conjunction is also allowed, it is the constant 1.
In general, let us call a Boolean polynomial a tautology if it is identically 1.
We found that all Boolean functions can be expressed by a disjunctive normal form. From the
disjunctive normal form, we can obtain a conjunctive normal form, applying the distributivity
property repeatedly. We have seen that this is a way to decide whether the polynomial is a
tautology. Similarly, an algorithm to decide whether a polynomial is satisfiable is to bring it to
a disjunctive normal form. Both algorithms can take very long time.
In general, one and the same Boolean function can be expressed in many ways as a Boolean
polynomial. Given such an expression, it is easy to compute the value of the function. However,
most Boolean functions can be expressed only by very large Boolean polynomials; this may even
be so for Boolean functions that can be computed fast.

25
 0 
x=0 Q
s
Q x = x NOR x = 1
NOR
0 3



 0 
XXX
z Q
x NOR y = 0 s
Q x⇒y=1
NOR NOR
y=1 :
 0 3



Figure 1.7: A NOR circuit computing x ⇒ y, with assignment on edges

(1.3) Example [Majority Function] Let f (x1 , . . . , xn ) = 1 if and only if at least half of the
variables are 1. ♦

One reason why a computation might be much faster than the size of the Boolean polynomial
is that the size of a Boolean polynomial does not reflect the possibility of reusing partial results.
This deficiency is corrected by the following more general formalism.
Let G be a directed graph with numbered nodes that does not contain any directed cycle
(i.e. is acyclic). The sources, i.e. the nodes without incoming edges, are called input nodes. We
assign a literal (a variable or its negation) to each input node.
The sinks of the graph, i.e. those of its nodes without outgoing edges, will be called output
nodes. (In what follows, we will deal most frequently with circuits that have a single output
node.)
To each node v of the graph that is not a source, i.e. which has some degree d = d+ (v) > 0,
a “gate” is given, i.e. a Boolean function Fv : {0, 1}d → {0, 1}. The incoming edges of the
node are numbered in some increasing order and the variables of the function Fv are made to
correspond to them in this order. Such a graph is called a circuit.
The size of the circuit is the number of gates; its depth is the maximal length of paths leading
from input nodes to output nodes.
Every circuit H determines a Boolean function. We assign to each input node the value of
the assigned literal. This is the input assignment, or input of the computation. From this, we
can compute to each node v a value x(v) ∈ {0, 1}: if the start nodes u1 , . . . , ud of the incoming
edges have already received a value then v receives the value Fv (x(u1 ), . . . , x(ud )). The values
at the sinks give the output of the computation. We will say about the function defined this
way that it is computed by the circuit H.

1.22 Exercise Prove that in the above definition, the circuit computes a unique output for
every possible input assignment. ♦

(1.4) Example A NOR circuit computing x ⇒ y. We use the formulas

x ⇒ y = ¬(¬x NOR y), ¬x = x NOR x.

26
If the states of the input lines of the circuit are x and y then the state of the output line is
x ⇒ y. The assignment can be computed in 3 stages, since the longest path has 3 edges. See
Figure 1.7. ♦

(1.5) Example For a natural number n we can construct a circuit that will simultaneously
compute all the functions Ea1 ,...,an (x1 , . . . , xn ) (as defined above in the proof of Lemma 1.1) for
all values of the vector (a1 , . . . , an ). This circuit is called the decoder circuit since it has the
following behavior: for each input x1 , . . . , xn only one output node, namely Ex1 ,...,xn will be
true. If the output nodes are consecutively numbered then we can say that the circuit decodes
the binary representation of a number k into the k-th position in the output. This is similar
to addressing into a memory and is indeed the way a “random access” memory is addressed.
Suppose that a decoder circuit is given for n. To obtain one for n + 1, we split each output
y = Ea1 ,...,an (x1 , . . . , xn ) in two, and form the new nodes
Ea1 ,...,an ,1 (x1 , . . . , xn+1 ) = y ∧ xn+1 ,
Ea1 ,...,an ,0 (x1 , . . . , xn+1 ) = y ∧ ¬xn+1 ,
using a new copy of the input xn+1 and its negation. ♦

Of course, every Boolean function is computable by a trivial (depth 1) circuit in which a

single (possibly very complicated) gate computes the output immediately from the input. The
notion of circuits is interesting if we restrict the gates to some simple operations (AND, OR,
exclusive OR, implication, negation, etc.). If each gate is a conjunction, disjunction or negation
then using the DeMorgan rules, we can push the negations back to the inputs which, as literals,
can be negated variables anyway. If all gates are disjunctions or conjunctions then the circuit
is called Boolean.
The in-degree of the nodes is is called fan-in. This is often restricted to 2 or to some fixed
maximum. Sometimes, bounds are also imposed on the out-degree, or fan-out. This means that
a partial result cannot be “freely” distributed to an arbitrary number of places.
Let f : {0, 1}n → {0, 1} be an arbitrary Boolean function and let
f (x1 , . . . , xn ) = E1 ∨ · · · ∨ EN
be its representation by a disjunctive normal form. This representation corresponds to a depth
2 circuit in the following manner: let its input points correspond to the variables x1 , . . . , xn and
the negated variables x1 , . . . , xn . To every elementary conjunction Ei , let there correspond a
vertex into wich edges run from the input points belonging to the literals occurring in Ei , and
which computes the conjunction of these. Finally, edges lead from these vertices into the output
point t which computes their disjunction. Note that this circuit has large fan-in and fan-out.
We can consider each Boolean circuit as an algorithm serving to compute some Boolean
function. It can be seen immediately, however, that circuits “can do” less than e.g. Turing
machines: a circuit can deal only with inputs and outputs of a given size. It is also clear that
(since the graph is acyclic) the number of computation steps is bounded. If, however, we fix
the length of the input and the number of steps then by an appropriate circuit, we can already
simulate the work of every Turing machine computing a single bit. We can express this also
by saying that every Boolean function computable by a Turing machine in a certain number of
steps is also computable by a suitable, not too big, Boolean circuit.

27
(1.6) Theorem For every Turing machine T and every pair n, N ≥ 1 of numbers there is a
Boolean circuit with n inputs, depth O(N ), indegree at most 2, that on an input (x1 , . . . , xn ) ∈
{0, 1}n computes 1 if and only if after N steps of the Turing machine T , on the 0’th cell of the
first tape, there is a 1.

(Without the restrictions on the size and depth of the Boolean circuit, the statement would be
trivial since every Boolean function can be expressed by a Boolean circuit.)

Proof Let us be given a Turing machine T = k, Σ, α, β, γ

and n, N ≥ 1. For simplicity, let
us assume k = 1. Let us construct a directed graph with vertices v[t, g, p] and w[t, p, h] where
0 ≤ t ≤ N , g ∈ Γ, h ∈ Σ and −N ≤ p ≤ N . An edge runs into every point v[t + 1, g, p] and
w[t + 1, p, h] from the points v[r, g  , p + ε] and w[r, p + ε, h ] (g  ∈ Γ, h ∈ Σ, ε ∈ {−1, 0, 1}). Let
us take n input points s0 , . . . , sn−1 and draw an edge from si into the points w[0, i, h] (h ∈ Σ).
Let the output point be w[N, 0, 1].
In the vertices of the graph, the logical values computed during the evaluation of the Boolean
circuit (which we will denote, for simplicity, just like the corresponding vertex) describe a
computation of the machine T as follows: the value of vertex v[t, g, p] is true if after step t,
the control unit is in state g and the head scans the p-th cell of the tape. The value of vertex
w[t, p, h] is true if after step t, the p-th cell of the tape holds symbol h.
Certain ones among these logical values are given. The machine is initially in the state
START, and the head starts from cell 0:

v[0, g, p] = 1 ⇔ g = ST ART ∧ p = 0,

further we write the input onto cells 0, . . . , n − 1 of the tape:

w[0, p, h] = 1 ⇔ ((p < 0 ∨ p ≥ n) ∧ h = ∗) ∨ (0 ≤ p ≤ n − 1 ∧ h = xp ).

The rules of the Turing machine tell how to compute the logical values corresponding to the
rest of the vertices:

v[t + 1, g, p] = 1 ⇔ ∃g  ∈ Γ, ∃h ∈ Σ : α(g  , h ) = g ∧ v[t, g  , p − γ(g  , h )] = 1

∧w[t, p − γ(g  , h ), h ] = 1.
w[t + 1, p, h] = 1 ⇔ (∃g  ∈ Γ, ∃h ∈ Σ : v[t, g  , p] = 1 ∧ w[t, p, h ] = 1 ∧ β(g  , h ) = h)
∨(w[t, p, h] = 1 ∧ ∀g  ∈ Γ : w[t, g  , p] = 0).

It can be seen that these recursions can be taken as logical functions which turn the graph into
a Boolean circuit computing the desired functions. The size of the circuit will be O(N 2 ), its
depth O(N ). Since the in-degree of each point is at most 3|Σ| · |Γ| = O(1), we can transform
the circuit into a Boolean circuit of similar size and depth.

(1.7) Remark Our construction of a universal Turing machine in Theorem 1.1 is inefficient
and unrealistic. For most commonly used transition functions α, β, γ, a table is a very inefficient
way to store the description. A Boolean circuit (with a Boolean vector output) is often a vastly

28
more economical representation. It is possible to construct a universal one-tape Turing machine
V1 taking advantage of such a representation. The beginning of the tape of this machine would
not list the table of the transition function of the simulated machine, but would rather describe
the Boolean circuit computing it, along with a specific state of this circuit. Each stage of the
simulation would first simulate the Boolean circuit to find the values of the functions α, β, γ
and then proceed as before. ♦

1.23 Exercise Consider that x1 x0 is the binary representation of an integer x = 2x1 + x0

and similarly, y1 y0 is a binary representation of a number y. Let f (x0 , x1 , y0 , y1 , z0 , z1 ) be the
Boolean formula which is true if and only if z1 z0 is the binary representation of the number
x + y mod 4.
Express this formula using only conjunction, disjunction and negation. ♦

1.24 Exercise Convert into disjunctive normal form the following Boolean functions.

(a) x + y + z mod 2

(b) x + y + z + t mod 2

1.25 Exercise Convert into conjunctive normal form the formula (x ∧ y ∧ z) ⇒ (u ∧ v). ♦

1.26 Exercise Prove that for every Boolean circuit of size N , there is a Boolean circuit of size
at most N 2 with indegree 2, computing the same Boolean function. ♦

1.27 Exercise Prove that for every Boolean circuit of size N and indegree 2 there is a Boolean
circuit of size O(N ) and indegree at most 2 computing the same Boolean function. ♦

1.28 Exercise Prove that the Boolean polynomials are in one-to-one correspondence with
those Boolean circuits that are trees. ♦

1.29 Exercise Monotonic Boolean functions. A Boolean function is monotonic if its value
does not decrease whenever any of the variables is increased. Prove that for every Boolean
circuit computing a monotonic Boolean function there is another one that computes the same
function and uses only nonnegated variables and constants as inputs. ♦

1.30 Exercise Universal circuit. For each n, construct a Boolean circuit whose gates have
indegree ≤ 2, which has size O(2n ) with 2n + n inputs and which is universal in the following
sense: that for all binary strings p of length 2n and binary string x of length n, the output of
the circuit with input xp is the value, with argument x, of the Boolean function whose table is
given by p. [Hint: use the decoder circuit of Example 1.5.] ♦

1.31 Exercise Circuit size. The gates of the Boolean circuits in this exercise are assumed to
have indegree ≤ 2.

29
 (a) Prove the existence of a constant c such that for all n, there is a Boolean function such
that each Boolean circuit computing it has size at least c · 2n /n. [Hint: count the number
of circuits of size k.]
∗ (b) For a Boolean function f with n inputs, show that the size of the Boolean circuit needed
for its implementation is O(2n /n).

30
2 Algorithmic decidability
2.1 Introduction
In this chapter, we study the question: which problems can be solved by any algorithm (or
computing device) at all?
Until the 1930’s, it was the consensus among mathematicians — mostly not spelled out
precisely — that every mathematical question that can be formulated precisely, can also be
solved. This statement has two interpretations. We can talk about a single yes-or-no question
(say: is every planar graph 4-colorable? is every even integer larger than 2 the sum of two
primes?), and then the decision means that it can be proved or disproved from the axioms of
set theory (which were, and still are, generally accepted as the axioms of mathematics). This
belief was destroyed by the the Austrian mathematician Kurt Gödel, who published a famous
result in 1931. According to this, there are perfectly well formulated mathematical questions
that cannot be answered from the axioms of set theory.
Now one could think that this is a weakness of this particular system of axioms: perhaps by
adding some generally accepted axioms (which had been overlooked) one could get a new system
that would allow us to decide the truth of every well-formulated mathematical statement. Gödel
proved that this hope was also vain: no matter how we extend the axiom system of set theory
(allowing even infinitely many axioms, subject to some reasonable restrictions: no contradiction
should be derivable and that it should be possible to decide about a statement whether it is an
axiom or not), still there remain unsolvable problems.
The second meaning of the question of decidability is when we are concerned with a family
of questions and are looking for an algorithm that decides each of them. In 1936, Church
formulated a family of problems for which he could prove that they are not decidable by any
algorithm. For this statement to make sense, the mathematical notion of an algorithm had to
be created. Church used tools from logic, the notion of recursive functions, to formalize the
notion of algorithmic solvability.
Similarly as in connection with Gödel’s Theorem, it seems quite possible that one could
define algorithmic solvability in a different way, or extend the arsenal of algorithms with new
tools, allowing the solution of new problems. In the same year when Church published his work,
Turing created the notion of a Turing machine. Nowadays we call something algorithmically
computable if it can be computed by some Turing machine. But it turned out that Church’s
original model is equivalent to the Turing machine in the sense the same computational problems
can be solved by them. We have seen in the previous chapter that the same holds for the Random
Access Machine. Many other computational models have been proposed (some are quite different
from the Turing machine, RAM, or any real-life computer, like quantum computing or DNA
computing), but nobody found a machine model that could solve more computational problems
than the Turing machine.
Church in fact anticipated this by formulating the so-called Church Thesis, according to
which every “calculation” can be formalized in the system he gave. Today we state this hy-
pothesis in the form that all functions computable on any computing device are computable
on a Turing machine. As a consequence of this thesis (if we accept it) we can simply speak
of computable functions without referring to the specific type of machine on which they are

31
computable.
(One could perhaps make one exception from the Church Thesis for algorithms using ran-
domness. These can solve algorithmically unsolvable computational problems so that the answer
is correct with large probability. See Chapter 7 on Information Complexity.)

32
2.2 Recursive and recursively enumerable languages
Let Σ be a finite alphabet that contains the symbol “∗”. We will allow as input for a Turing
machine words that do not contain this special symbol: only letters from Σ0 = Σ \ {∗}.
We call a function f : Σ∗0 → Σ∗0 recursive or computable if there exists a Turing machine
that for any input x ∈ Σ∗0 will stop after finite time with f (x) written on its first tape. (We
have seen in the previous section that we can assume without loss of generality that the Turing
machine has only one tape.)
The notions of recursive, as well as that of “recursively enumerable” and “partial recursive”
defined below can be easily extended, in a unique way, to functions and sets over some countable
sets different from Σ∗0 , like the set of natural numbers, the set N ∗ of finite strings of natural
numbers, etc. The extension goes with help of some standard coding of, e.g., the set of natural
numbers by elements of Σ∗0 . Therefore even though we define these notions only over Σ∗0 , we
sometimes use them in connection with function defined over other domains. This is a bit sloppy
but does not lead to any confusion.
We call a language L recursive if its characteristic function

1 if x ∈ L,
fL (x) =
0 otherwise,

is recursive. Instead of saying that a language L is recursive, we can also say that the property
defining L is decidable. If a Turing machine calculates this function then we say that it decides
the language. It is obvious that every finite language is recursive. Also if a language is recursive
then its complement is also recursive.

(2.1) Remark It is obvious that there is a continuum of languages (and so uncountably many)
but only countably many Turing machines. So there must exist non-recursive languages. We
will see some concrete languages that are non-recursive. ♦

We call the language L recursively enumerable if L = ∅ or there exists a recursive function

f such that the range of f is L. This means that we can enumerate the elements of L: L =
{f (w0 ), f (w1 ), . . .}, when Σ∗0 = {w0 , w1 , . . .}. Here, the elements of L do not necessarily occur
in increasing order and repetition is also allowed.
We give an alternative definiton of recursively enumerable languages through the following
lemma.

(2.2) Lemma A language L is recursively enumerable iff there is a Turing machine T such

that if we write x on the first tape of T the machine stops iff x ∈ L.

Proof Let L be recursively enumerable. We can assume that it is nonempty. Let L be the
range of f . We prepare a Turing machine which on input x calculates f (y) in increasing order
of y ∈ Σ∗0 and it stops whenever it finds a y such that f (y) = x.
On the other hand, let us assume that L contains the set of words on which T stops. We
can assume that L is not empty and a ∈ L. We construct a Turing machine √ T02 that, when the
natural number i is its input it simulates T on input x which is the (i −  i) -th word of Σ∗0 ,

33
for i steps. If the simulated T stops then T0 ouputs x. Since every word of Σ∗0 will occur for
infinitely many values of i the range of T0 will be L.
√
There is nothing really tricky about the function (i −  i)2 ; all we need is that for i =
0, 1, 2, . . . its value assumes every non-negative integer infinitely many times. The technique
used in this proof, that of simulating infinitely many computations by a single one, is sometimes
called “dovetailing”.
Now we study the relationship between recursive and recursively enumerable languages.

(2.3) Lemma Every recursive language is recursively enumerable.

Proof This is clear if the language L is empty. We can change the Turing machine that decides
f to output the input if the intended output is 1, and to output some arbitrary fixed a ∈ L if
the intended output is 0.
The next theorem characterizes the relation of recursively enumerable and recursive lan-
guages.

(2.4) Theorem A language L is is recursive iff both languages L and Σ∗0 \ L are recursively
enumerable.

Proof If L is recursive then its complement is also recursive, and by the previous lemma, it
is recursively enumerable.
On the other hand, let us assume that both L and its complement are recursively enumerable.
We can construct two machines that enumerate them, and a third one simulating both that
detects if one of them lists x. Sooner or later this happens and then we know where x belongs.

Now, we will show that there are languages that are recursively enumerable (and hance
rather explicit), but not recursive.
Let T be a Turing machine with k tapes, and let LT be the set of those words x ∈ Σ∗0 words
for which T stops when we write x on all of its tapes.

(2.5) Theorem If T is a universal Turing machine with k + 1 tapes then LT is recursively

enumerable, but it is not recursive.

Proof The first statement follows from Lemma 2.3. We prove the second statement, for
simplicity, for k = 1.
Let us assume, by way of contradiction, that LT is recursive. Then Σ∗0 \ LT would be
recursively enumerable, so there would exist a 1-tape Turing machine T1 that on input x would
stop iff x ∈ LT . The machine T1 can be simulated on T by writing an appropriate program p
on the second tape of T . Then writing p on both tapes of T , it would stop if T1 would stop
because of the simulation. The machine T1 was defined, on the other hand, to stop on p if and
only if T does not stop with input p on both tapes (i.e. when p ∈ LT ). This is a contradiction.

This proof uses the so called diagonalization technique originating from set theory (where
it was used by Cantor to show that the set of all real numbers is not countable). The technique

34
forms the basis of many proofs in logic, set-theory and complexity theory. We will see more of
these in what follows.
There is a number of variants of the previous result, asserting the undecidability of similar
problems.
Let T be a Turing machine. The halting problem for T is the problem to decide, for all
possible inputs x, whether T halts on x. Thus, the decidability of the halting problem of T
means the decidability of the set of those x for which T halts. We can also speak about the
halting problem in general, which means that a pair (T, x) is given where T is a Turing machine
(given by its transition table) and x is an input.

(2.6) Theorem There is a 1-tape Turing machine whose halting problem is undecidable.

Proof Suppose that the halting problem is decidable for all one-tape Turing machines. Let
T be a 2-tape universal Turing machine and let us construct a 1-tape machine T0 similarly to
the proof of Theorem (0.2) (with k = 2), with the difference that at the start, we write the
i-th letter of word x not only in cell 4i but also in cell 4i − 2. Then on an input x, machine
T0 will simulate the work of T , when the latter starts with x on both of its tapes. Since it is
undecidable whether T halts for a given input (x, x), it is also undecidable about T0 whether it
halts on a given input x.
The above proof, however simple it is, is the prototype of a great number of undecidability
proofs. These proceed by taking any problem P1 known to be undecidable (in this case, mem-
bership in LT ) and showing that it can be reduced to the problem P2 at hand (in this case, the
halting problem of T0 ). The reduction shows that if P2 is decidable then so is P1 . But since we
already know that P1 is undecidable, we conclude that P2 is undecidable as well. The reduction
of a problem to some seemingly unrelated problem is, of course, often very tricky.
A description of a Turing machine is the listing of the sets Σ, Γ (where, as before, the
elements of Γ are coded by words over the set Σ0 ), and the table of the functions α, β, γ.

(2.7) Corollary It is algorithmically undecidable whether a Turing machine (given by its

description) halts on the empty input.

Proof Let T be a Turing machine whose halting problem is undecidable. We show that its
halting problem can be reduced to the general halting problem on the empty input. Indeed, for
each input x, we can construct a Turing machine Tx which, when started with an empty input,
writes x on the input tape and then simulates T . If we could decide whether Tx halts then we
could decide whether T halts on x.

(2.8) Corollary It is algorithmically undecidable whether for a one-tape Turing machine T

(given by its description), the set LT is empty.

Proof For a given machine S, let us construct a machine T that does the following: it first
erases everything from the tape and then turns into the machine S. The description of T can
obviously be easily constructed from the description of S. Thus, if S halts on the empty input
in finitely many steps then T halts on all inputs in finitely many steps, hence LT = Σ∗0 is not

35
empty. If S works for infinite time on the empty input then T works infinitely long on all inputs,
and thus LT is empty. Therefore if we could decide whether LT is empty we could also decide
whether S halts on the empty input, which is undecidable.
Obviously, just as its emptyness, we cannot decide any other property P of of LT either if
the empty language has it and Σ∗0 has not, or vice versa. Even a more general negative result is
true. We call a property of a language trivial if either all languages have it or none.

(2.9) Rice’s Theorem For any non-trivial language-property P , it is undecidable whether

the language LT of an arbitrary Turing machine T (given by its description) has this property.

Thus, it is undecidable on the basis of the description of T whether LT is finite, regular,

contains a given word, etc.

Proof We can assume that the empty language does not have property P (otherwise, we can
consider the negation of P ). Let T1 be a Turing machine for which LT1 has property P . For a
given Turing machine S, let us make a machine T as follows: for input x, first it simulates S on
the empty input. When the simulated S stops it simulates T1 on input x. Thus, if S does not
halt on the empty input then T does not halt on any input, so LT is the empty language. If S
halts on the empty input then T halts on exactly the same inputs as T1 , and thus LT = LT1 .
Thus if we could decide whether LT has property P we could also decide whether S halts on
empty input.

36
2.3 Other undecidable problems
First we mention a problem of geometrical nature. A prototile, or domino is a square shape
and has a natural number written on each side. A tile is an exact copy of some prototile. (To
avoid trivial solutions, let us require that the copy must be positioned in the same orientation as
the prototile, without rotation.) A kit is a finite set of prototiles, one of which is a distinguished
“initial domino”. Given a kit K, a tiling of whole plane with K (if it exists) assigns to each
position with integer coordinates a tile which is a copy of a prototile in K, in such a way that

— neighbor dominoes have the same number on their adjacent sides;

— the initial domino occurs.

It is easy to give a kit of dominoes with which the plane can be tiled (e.g. a single square
that has the same number on each side) and also a kit with which this is impossible (e.g., a
single square that has a different number on each side). It is, however, a surprising fact that
the tiling problem is algorithmically undecidable!
For the exact formulation, let us describe each kit by a word over Σ0 = {0, 1, +}, e.g. in such
a way that we write up the numbers written on the sides of the prototiles in binary, separated
by the symbol “+”, beginning at the top side, clockwise, then we join the so obtained number
4-tuples starting with the initial domino. (The details of the coding are not interesting.) Let
LTLNG [resp. LNTLNG ] the set of codes of those kits with which the plane is tileable [resp. not
tileable].

(2.1) Theorem The tiling problem is undecidable, i.e. the language LTLNG is not recursive.

Accepting, for the moment, this statement, according to Theorem 2.4, either the tiling or
the nontiling kits must form a language that is not recursively enumerable. Which one? For
the first look, we might think that LTLNG is recursive: the fact that the plane is tileable by a
kit can be proved by supplying the tiling. This is, however, not a finite proof, an actually the
truth is just the opposite:

(2.2) Theorem The language LNTLNG is recursively enumerable.

Taken together with Theorem 2.1, we see that LTLNG can not even be recursively enumerable.
In the proof of Theorem 2.2, the following lemma will play important role.

(2.3) Lemma The plane is tileable by a kit if an only if for all n, the square (2n + 1) × (2n + 1)
is tileable by it with the initial domino is in its center.

Proof The “only if” part of the statement is trivial. For the proof of the “if” part, consider a
sequence N1 , N2 , . . . of tilings of squares such that they all have odd sidelength and their side-
length converges to infinity. We will construct a tiling of the plane. Without loss of generality,
we can assume that the center of each square is at the origin. Let us consider first the 3 × 3
square centered at the origin. This is tiled by the kit somehow in each Ni . Since it can only be
tiled in finitely many ways, there is an infinite number of Ni ’s in which it is tiled in the same

37
way. With an appropriate thinning of the sequence Ni we can assume that this square is tiled
in the same way in each Ni . These nine tiles can already be fixed.
Proceeding, assume that the sequence has been thinned out in such a way that every re-
maining Ni tiles the square (2k + 1) × (2k + 1) centered at the origin in the same way, and we
have fixed these (2k + 1)2 tiles. Then in the remaining tilings Ni , the square (2k + 3) × (2k + 3)
centered at the origin is tiled only in a finite number of ways, and therefore one of theses tilings
occurs an infinite number of times. If we keep only these tilings Ni then every remaining tiling
tiles the square (2k +3)×(2k +3) centered at the origin in the same way, and this tiling contains
the tiles fixed previously. Now we can fix the new tiles on the edge of the bigger square.
Every tile covering some integer vertex unit square of the plane will be fixed sooner or later,
i.e. we have obtained a tiling of the whole plane. Since the condition imposed on the covering is
“local”, i.e. it refers only to two tiles dominoes, the tiles will be correctly matched in the final
tiling, too.

Proof of Theorem 2.2 Let us construct a Turing machine doing the following. For a word
x ∈ Σ∗0 , it first of all decides whether it codes a kit (this is easy); if not then it goes into an
infinite cycle. If yes, then with this set, it tries to tile one after the other the squares 1 × 1, 2 × 2,
3 × 3, etc. For each concrete square, it is decidable in a finite number of steps, whether it is
tileable, since the sides can only be numbered in finitely many ways by the numbers occurring
in the kit, and it is easy to verify whether among the tilings obtained this way there is one for
which every tile comes from the given kit. If the machine finds a square not tileable by the
given kit then it halts.
It is obvious that if x ∈ LTLNG , i.e. x either does not code a kit or codes a kit which tiles the
plane then this Turing machine does not stop. On the other hand, if x ∈ LNTLNG , i.e. x codes a
kit that does not tile the plane then according to Lemma 2.3, for a large enough k already the
square k × k is not tileable either, and therefore the Turing machine stops after finitely many
steps. Thus, according to Lemma 2.2, the language LNTLNG is recursively enumerable.

Proof of Theorem 2.1 Let T = k, Σ, α, β, γ

be an arbitrary Turing machine; we will
construct from it (using its description) a kit K which can tile the plane if and only if T does
not start on the empty input. This is, however, undecidable due to Corollary 2.7, so it is also
undecidable whether the constructed kit can tile the plane.
In defining the kit, we will write symbols, rather than numbers on the sides of the tiles; these
are easily changed to numbers. For simplicity, assume k = 1. It is also convenient to assume
(and achievable by a trivial modification of T ) that the machine T is in the starting state only
before the first step.
Let us subdivide the plane into unit squares whose centers are points with integer coordi-
nates. Assume that T does not halt on empty input. Then from the machine’s computation,
let us construct a tiling as follows (see Figure 2.1): if the content of the p-th cell of the machine
after step q is symbol h then let us write symbol h on the top side of the square with center
point (p, q) an on the bottom side of the square with center point (p, q + 1). If after step q, the
head scans cell p and the control unit is in state g then let us write the symbol g on the top
side of the square with center point (p, q) an on the bottom side of the square with center point

38
 ∗ 1 2 1 ∗g1
g1 g1

∗ 1 2 ∗g2 ∗
∗ 1 2 ∗g2 ∗
g2 g2

∗ 1 ∗g1 ∗ ∗
∗ 1 ∗g1 ∗ ∗
g1 g1

∗ ∗START ∗ ∗ ∗
∗ ∗START ∗ ∗ ∗

N N N P P P P P P P
∗ START∗ ∗ ∗ ∗

Figure 2.1: Tiling resulting from a particular computation

(p, q + 1). If the head, in step q, moves right [left], say from cell p − 1 [p + 1] to cell p and is in
state g after the move then let us write symbols −1, g [1, g] on the left [right] side of the square
with center (p, q) and on the right [left] side of the cell with center (p − 1, q) [(p + 1, q)]. Let us
write symbol “N” on the vertical edges of the squares in the bottom row if the edge is to the
left of the origin and the letter “P” if the edge is to the right of the origin. Let us reflect the
tiling with respect to the x axis, reversing also the order of the labels on each edge. Figure 2.1
shows the construction for the simple Turing machine which steps right on the empty tape and
writes on the tape 1’s and 2’s alternatingly.
Let us determine what kind of tiles occur in the tiling obtained. In the upper half-plane,
there are basically four kinds. If q > 0 and after step q, the head rests on position p then the
square with center (p, q) comes from one of the prototiles shown in Figure 2.2. If q > 0 and
after step q − 1 the head scans position p then the square with center (p, q) comes from one of
the prototiles on Figure 2.3a-b. If q > 0 and the head is not at position p either after step q − 1
or after step q then the square with position (p, q) has simply the form of Figure 2.3c. Finally,
the squares of the bottom line are shown in Figure 2.4. We obtain the tiles occurring in the
lower half-plane by reflecting the above ones across the horizontal axis and reversing the order
of the labels on each edge.
Now, figures 2.2–2.4 can be constructed from the description of the Turing machine; we thus
arrive at a kit KT whose initial domino is the middle domino of Figure 2.4. The above reasoning
shows that if T runs an infinite number of steps on empty input then the plane can be tiled with
this kit. Conversely, if the plane can be tiled with the kit KT then the initial domino covers
(say) point (0, 0); to the left and right of this, only the two other dominos of Figure 2.4 can
stand. Moving row-by-row from here we can see that the covering is unique and corresponds to

39
 h h g  h
1g  −1g 
hg hg hg
α(g, h) = g  α(g, h) = g  α(g, h) = g 
β(g, h) = h β(g, h) = h β(g, h) = h
γ(g, h) = 1 γ(g, h) = 0 γ(g, h) = −1

Figure 2.2: Tiles for currently scanned cell

hg hg h
1g −1g
h h h
a) b) c)

Figure 2.3: Tiles for previously scanned or unscanned cell

a computation of machine T on empty input.

The only not completely obvious case is when a tile of the form shown in the first two
examples of Figure 2.4 occurs. We must show that it can occur indeed only if the head moves
onto the corresponging tape square. The symbols ±1 in front of g exclude all other possibilities.
Since we have covered the whole plane, this computation is infinite.

(2.4) Remark The tiling problem is undecidable even if we do not distinguish an initial
domino. But the proof of this is much harder. ♦

We mention some more algorithmically undecidable problems without showing the proof of
undecidability. The proof is in each case a complicated encoding of the halting problem into
the problem at hand.
In 1900, Hilbert formulated 23 problems that he considered then the most exciting in
mathematics. These problems had a great effect on the development of the mathematics of the
century. (It is interesting to note that Hilbert thought: some of his problems will resist science
for centuries; until today, essentially all of them are solved.) One of these problems was the
following:

∗ ∗START ∗

N N N P P P
∗ START∗ ∗

Figure 2.4: Tiles for the bottom line

40
(2.5) Diophantine equation Given a polynomial p(x1 , . . . , xn ) with integer coefficients and
n variables, decide whether the equation p = 0 has integer solutions. ♦

(An equation is called Diophantine if we are looking for its integer solutions.)
In Hilbert’s time, the notion of algorithms was not formalized but he thought that a uni-
versally acceptable and always executable procedure could eventually be found that decides
for every Diophantine equation whether it is solvable. After the clarification of the notion of
algorithms and the finding of the first algorithmically undecidable problems, it became more
probable that this problem is algorithmically undecidable. Davis, Robinson and Myhill
reduced this conjecture to a specific problem of number theory which was eventually solved by
Mat’iyasevich in 1970. It was found therefore that the problem of solvability of Diophantine
equations is algorithmically undecidable.
We mention also an important problem of algebra. Let us be given n symbols: a1 , . . . , an .
The free group generated from these symbols is the set of all finite words formed from the
symbols a1 , . . . , an , a−1 −1 −1
1 , . . . , an in which the symbols ai and ai never follow each other (in any
order). We multiply two such words by writing them after each other and repeatedly erasing any
pair of the form ai a−1 −1
i or ai ai whenever they occur. It takes some, but not difficult, reasoning
to show that the multiplication defined this way is associative. We also permit the empty word,
this will be the unit element of the group. If we reverse a word and change all symbols ai in it
to a−1
i (and vice versa) then we obtain the inverse of the word. In this very simple structure,
the following problem is algorithmically undecidable.

(2.6) Word problem of groups In the free group generated by the symbols a1 , . . . , an , we
are given n + 1 words: α1 , . . . , αn and β. Is β in the subgroup generated by α1 , . . . , αn ? ♦

Finally, a problem from the field of topology. Let e1 , . . . , en be the unit vectors of the n-
dimensional Euclidean space. The convex hull of the points 0, e1 , . . . , en is called the standard
simplex. The faces of this simplex are the convex hulls of subsets of the set {0, e1 , . . . , en }.
A polyhedron is the union of an arbitrary set of faces of the standard simplex. Here is a
fundamental topological problem concerning a polyhedron P :

(2.7) Contractability of polyhedra Can a given polyhedron be contracted to a single point

(continuously, staying within itself)? ♦

We define this more precisely, as follows: we mark a point p in the polyhedron first and want
to move each point of the polyhedron in such a way within the polyhedron (say, from time 0 to
time 1) that it will finally slide into point p and during this, the polyhedron “is not teared”. Let
F (x, t) denote the position of point x at time t for 0 ≤ t ≤ 1. The mapping F : P × [0, 1] → P
is thus continuous in both of its variables together, having F (x, 0) = x and F (x, 1) = p for all
x. If there is such an F then we say that P is “contractable”. For example, a triangle, taken
with the area inside it, is contractable. The perimeter of the triangle (the union of the three
sides without the interior) is not contractable. (In general, we could say that a polyhedron is
contractable if no matter how a thin circular rubber band is tied on it, it is possible to slide this
rubber band to a single point.) The property of contractability turns out to be algorithmically
undecidable.

41
2.4 Computability in logic
2.4.1 Godel’s incompleteness theorem
Mathematicians have long held the conviction that a mathematical proof, when written out in
all detail, can be checked unambiguously. Aristotle made an attempt to formalize the rules of
deduction but the correct formalism was found only by Frege and Russell at the end of the
ninetieth century. It was championed as a sufficient foundation of mathematics by Hilbert.
We try to give an overview of the most important results concerning decidability in logic.
Mathematics deals with sentences, statements about some mathematical objects. All sen-
tences will be strings in some finite alphabet. We will always assume that the set of sentences
(sometimes also called a language) is decidable: it should be easy to distinguish (formally)
meaningful sentences from nonsense. Let us also agree that there is an algorithm computing
from each sentence ϕ, an other sentence ψ called its negation.

(2.1) Example Let L1 be the language consisting of all expressions of the form “l(a, b)” and
“l (a, b)” where a, b are natural numbers (in their usual, decimal representation). The sentences
l(a, b) and l (a, b) are each other’s negations. ♦

A proof of some sentence T is a finite string P that is proposed as an argument that T is

true. A formal system, or theory F is an algorithm to decide, for any pairs (P, T ) of strings
whether P is an acceptable proof T . A sentence T for which there is a proof in F is called a
theorem of the theory F.

(2.2) Example Here is a simple theory T1 based on the language L1 of the above Example
2.1. Let us call axioms all “l(a, b)” where b = a + 1. A proof is a sequence S1 , . . . , Sn of
sentences with the following property. If Si is in the sequence then either it is an axiom or there
are j, k < i and integers a, b, c such that Sj =“l(a, b)”, Sk =“l(b, c)” and Si = l(a, c). This theory
has a proof for all formulas of the form l(a, b) where a < b. ♦

A theory is called consistent if for no sentence can both it and its negation be a theorem.
Inconsistent theories are uninteresting, but sometimes we do not know whether a theory is
consistent.
A sentence S is called undecidable in a theory T if neither S nor its negation is a theorem
in T . A consistent theory is complete if it has no undecidable sentences.
The toy theory of Example 2.2 is incomplete since it will have no proof of either l(5, 3) nor
l (5, 3). But it is easy to make it complete e.g. by adding as axioms all formulas of the form
l (a, b) where a, b are natural numbers and a ≥ b.
Incompleteness simply means that the theory formulates only certain properties of a kind of
system: other properties depend exactly on which system we are considering. Completeness is
therefore not always even a desireable goal with certain theories. It is, however, if the goal of
our theory is to describe a certain system as completely as we can. We may want e.g. to have
a complete theory of the set of natural numbers in which all true sentences have proofs. Also,
complete theories have a desirable algorithmic property, as shown by the theorem below: this
shows that if there are no (logically) undecidable sentences in a theory then the truth of all
sentences (with respect to that theory) is algorithmically decidable.

42
(2.3) Theorem If a theory T is complete then there is an algorithm that for each sentence S
finds in T a proof either for S or for the negation of S.

Proof The algorithm starts enumerating all possible finite strings P and checking whether P
is a proof for S or a proof for the negation of S. Sooner or later, one of the proofs must turn
up, since it exists. Consistency implies that if one turns up the other does not exist.
Suppose that we want to develop a complete a theory of natural numbers. Since all sentences
about strings, tables, etc. can be encoded into sentences about natural numbers this theory must
express all statements about such things as well. In this way, in the language of natural numbers,
one can even speak about Turing machines, and about when a Turing machine halts.
Let L be some fixed r.e. set of integers that is not recursive. An arithmetical theory T is
called minimally adequate if for numbers n, the theory contains a sentence ϕn expressing the
statement “n ∈ L”; moreover, this statement is a theorem in T if and only if it is true.
It is reasonable to expect that a theory of natural numbers with a goal of completeness be
minimally adequate, i.e. that it should provide proofs for at least those facts that are verifiable
anyway directly by computation, as “n ∈ L” indeed is. (In the next subsection, we will describe
a minimally adequate theory.) Now we are in a position to prove one of the most famous theo-
rems of mathematics which has not ceased to exert its fascination on people with philosophical
interests:

(2.4) Gödel’s incompleteness theorem Every minimally adequate theory is incomplete.

Proof If the theory were complete then, according to Theorem 2.3 it would give a procedure
to decide all sentences of the form n ∈ L, which is impossible.

(2.5) Remark Looking more closely into the last proof, we see that for any adequate theory T
there is a natural number n such that though the sentence “n ∈ L” is expressible in T and true
but is not provable in T . There are other, more interesting sentences that are not provable, if
only the theory T is assumed strong enough: Gödel proved that the assertion of the consistency
of T is among these. This so-called Second Incompleteness Theorem of Gödel is beyond our
scope. ♦

(2.6) Remark Historically, Gödel’s theorems preceded the notion of computability by 3-4
years. ♦

2.4.2 First-order logic

Formulas Let us develop the formal system found most adequate to describe mathematics.
A first-order language uses the following symbols:

— An infinite supply of variables: x, y, z, x1 , x2 , . . ., to denote elements of the universe (the

set of objects) to which the language refers.

— Some function symbols like f, g, h, +, ·, f1 , f2 , . . . , where each function symbol has a prop-
erty called “arity” specifying the number of arguments of the function it will represent. A

43
 function of arity 0 is called a constant. It refers to some fixed element of the universe.
Some functions, like +, · are used in infix notation.

— Some predicate symbols like <, >, ⊂, ⊃, P, Q, R, P1 , P2 , . . ., also of different arities. A

predicate symbol with arity 0 is also called a propositional symbol. Some predicate
symbols, like <, are used with infix notation. The equality “=” is a distinguished
predicate symbol.

— Logical connectives: ¬, ∨, ∧, ⇒, ⇔, . . ..

— Quantifiers: ∀, ∃.

— Punctuation: (,).

A term is obtained by taking some constants and variables and applying function symbols
to them a finite number of times: e.g. (x + 2) + y or f (f (x, y), g(c)) are terms (here, 2 is a
constant).
An atomic formula has the form P (t1 , . . . , tk ) where P is a predicate symbol and ti are
terms: e.g. x + y < (x · x) + 1 is an atomic formula.
A formula is formed from atomic formulas by applying repeatedly the Boolean operations
and the adding of prefixes of the form ∀x and ∃x: e.g. ∀x(x < y) ⇒ ∃zg(c, z) or x = x ∨ y = y
are formulas. In the formula ∃y(∀x(F ) ⇒ G), the subformula F is called the scope of the
x-quantifier. An occurrence of a variable x in a formula is said to be bound if it is in the
scope of an x-quantifier; otherwise the occurrence is said to be free. A formula with no free
(occurrences of) variables is said to be a sentence; sentences make formulas which under any
given “interpretation” of the language, are either true of false.
Let us say that a term t is substitutable for variable x in formula A if no variable y occurs
in t for which some free occurrence of x in A is in the scope of some quantifier of y. If t is
substitutable for x in A then we write A[t/x] for the result of substituting t into every free
occurrence of x in A: e.g. if A = (x < 3 − x) and t = (y 2 ) then A[t/x] = (y 2 < 3 − y 2 ).
From now on, all our formal systems are some language of first-order logic, so they only
differ in what function symbols and predicate symbols are present.
There are natural ways to give interpretation to all terms and formulas of a first-order
language in such a way that under such an interpretation, all sentences become true or false.
This interpretation introduces a set called the universe and assigns functions and predicates
over this universe to the functions (and constants) and predicates of the language.

(2.7) Example Consider the language with the constants c0 , c1 and the two-argument function
symbol f . In one interpretation, the universe is the set of natural numbers, c0 = 0, c1 = 1,
f (a, b) = a + b. In another interpretation, the universe is {0, 1}, c0 = 0, c1 = 1, f (a, b) = a · b.
There are certain sentences that are true in both of these interpretations but not in all possible
ones: such is ∀x∀y f (x, y) = f (y, x). ♦

For a given theory T , an interpretation of its language is called a model of T if the ax-
ioms (and thus all theorems) of the theory are true in it. In the above Example 2.7, both
interpretations are models of the theory T1 defined by the single axiom ∀x∀y f (x, y) = f (y, x).

44
 It has been recognized long ago that the proof checking algorithm can be made independent
of the theory: theories are different only in their axioms. This algorithm is exactly what we mean
by “pure logical reasoning”; for first order logic, it was first formalized in the book Principia
Mathematica by Russell and Whitehead at the beginning of the 20th century. We will outline
one such algorithm at the end of the present subsection. Gödel proved in 1930 that if B implies
T in all interpretations of the sentences then there is a proof of the Principia Mathematica type
for it. The following theorem is a consequence.

(2.8) Gödel’s completeness theorem Let P be the set of all pairs (B, T ) where B is a finite
set of sentences and T is a sentence that is true in all interpretations in which the elements of
B are true. The set P is recursively enumerable.

Tarski proved that the algebraic theory of real numbers (and with it, all Euclidean geometry)
is complete. This is in contrast to the theories of natural numbers, among which the minimally
adequate ones are incomplete. (In the algebraic theory of real numbers, we cannot speak of an
“arbitrary integer”, only of an “arbitrary real number”.) Theorem 2.3 implies that there is an
algorithm to decide the truth of an arbitrary algebraic sentence on real numbers. The known
algorithms for doing this take a very long time, but are improving.

Proofs A proof is a sequence F1 , . . . , Fn of formulas in which each formula is either an axiom

or is obtained from previous formulas in the sequence using one of the rules given below. In
these rules, A, B, C are arbitrary formulas, and x is an arbitrary variable.
There is an infinite number of formulas that we will require to be part of the set of axioms
of each theory: these are therefore called logical axioms. These will not all necessarily be
sentences: they may contain free variables. To give the axioms, some more notions must be
defined.
Let F (X1 , . . . , Xn ) be a Boolean formula of the variables X1 , . . . , Xn , with the property that
it gives the value 1 for all possible substitutions of 0 or 1 into X1 , . . . , Xn . Let ϕ1 , . . . , ϕn be
arbitrary formulas. Formulas of the kind F (ϕ1 , . . . , ϕn ) are called tautologies.
The logical axioms of our system consist of the following groups:

Tautologies: All tautologies are axioms.

Equality axioms: Let t1 , . . . , tn , u1 , . . . , un be terms, f a function symbol and P a predicate

symbol, of arity n. Then

(t1 = u1 ∧ · · · ∧ tn = un ) ⇒ f (t1 , . . . , tn ) = f (u1 , . . . , un ),

(t1 = u1 ∧ · · · ∧ tn = un ) ⇒ (P (t1 , . . . , tn ) ⇔ P (u1 , . . . , un ))

are axioms.

The definition of ∃: For all formulas A and variables x, the formula ∃x A ⇔ ¬∀x ¬A is an
axiom.

Specialization: If term t is substitutable for variable x in formula A then ∀x A ⇒ A[t/x] is an

axiom.

45
 The system has two rules:

Modus ponens: From A ⇒ B and B ⇒ C, we can derive A ⇒ C.

Generalization: If the variable x does not occur free in A then from A ⇒ B we can derive
A ⇒ ∀x B.

(2.9) Remark The generalization rule says that if we can derive a statement B containing
the variable x without using any properties of x in our assumptions then it is true for arbitrary
values of x. It does not say that B ⇒ ∀xB is true. ♦

For the system above, the following stronger form of Gödel’s completeness theorem holds.

(2.10) Theorem Suppose that B is a set of sentences and T is a sentence that is true in all
interpretations in which the elements of B are true. Then there is a proof in the proof system
P of T from the axioms of B.

A simple theory of arithmetic; Church’s Theorem This theory N contains two con-
stants, 0 and 1, the function symbols +, · and the predicate symbol <. There is only a finite
number of simple nonlogical axioms (all of them without quantifier).

¬(x + 1 = 0).
1 + x = 1 + y ⇒ x = y.
x+0 = x.
x + (1 + y) = 1 + (x + y).
x·0 = 0.
x · (1 + y) = (x · y) + x.
¬(x < 0).
x < (1 + y) ⇔ x < y ∨ x = y.
x < y ∨ x = y ∨ y < x.

(2.11) Theorem The theory N is minimally adequate. Thus, there is a minimally adequate
consistent theory of arithmetic with a finite system of axioms.

This fact implies the following theorem of Church, showing that the problem of logical
provability is algorithmically undecidable.

(2.12) Undecidability Theorem of Predicate Calculus The set P of all sentences that
can be proven without any axioms, is undecidable.

46
Proof Let N be a finite system of axioms of a minimally adequate consistent theory of arith-
metic, and let N be the sentence obtained by taking the conjunction of all these axioms and
applying universal quantification. Let us remember the definition of “minimally adequate”: we
used there a nonrecursive r.e. set L of natural numbers. In arithmetic, we can write up a
formula Q(n) saying N ⇒ (n ∈ L). There is a proof for “n ∈ L” in N if and only if there is
a proof for Q(n) in P If we had a decision procedure for P we could decide, Q(n); since we
cannot, there is no decision procedure for P.

47
2.1 Exercise Prove that a function is recursive if and only if its graph { (x, f (x)) : x ∈ Σ∗0 } is
recursively enumerable. ♦

2.2 Exercise
(a) Prove that a language is recursively enumerable if and only if it can be enumerated without
repetition by some Turing machine.

(b) Prove that a language is recursive if and only if it can be enumerated in increasing order
by some Turing machine.
♦

In the exercises below, we will sometimes use the following notion. A function f defined on
a subset of Σ∗0 is called partial recursive (abbreviated as p.r.) if there exists a Turing machine
that for any input x ∈ Σ∗0 will stop after finite time if and only if f (x) is defined and in this
case, it will have f (x) written on its first tape.

2.3 Exercise Let us call two Turing machines equivalent if for all inputs, they give the same
outputs. Let the function f : Σ∗0 → {0, 1} be 1 if p, q are codes of equivalent Turing machines
and 0 otherwise. Prove that f is undecidable. ♦

2.4 Exercise Inseparability Theorem. Let U be a one-tape Turing machine simulating the
universal two-tape Turing machine. Let u (x) be 0 if the first symbol of the value computed
on input x is 0, and 1 if U halts but this first symbol is not 0. Then u is a partial recursive
function, defined for those x on which U halts. Prove that there is no computable total function
which is an extension of the function u (x). In particular, the two disjoint r.e. sets defined by
the conditions u = 0 and u = 1 cannot be enclosed into disjoint recursive sets. ♦

2.5 Exercise Nonrecursive function with recursive graph. Give a p.r. function f that is not
extendable to a recursive function, and whose graph is recursive. Hint: use the running time of
the universal Turing machine. ♦

2.6 Exercise Construct an undecidable, recursively enumerable set B of pairs of natural num-
bers with the property that for all x, the set { y : (x, y) ∈ B } is decidable, and at the same
time, for all y, the set { x : (x, y) ∈ B } is decidable. ♦

2.7 Exercise Let #E denote the number of elements of the set E. Construct an undecidable
set S of natural numbers such that
1
lim #(S ∩ {0, 1, . . . , n}) = 0.
n→∞ n

Can you construct an undecidable set for which the same limit is 1? ♦

2.8 Exercise A rooted tree is a set of “nodes” in wich each node has some “children”, the
single “root” node has no parent and each other node has a unique parent. A path is a sequence
of nodes in which each node is the parent of the next one. Suppose that each node has only
finitely many children and the tree is infinite. Prove that then the tree has an infinite path. ♦

48
2.9 Exercise Consider a Turing machine T which we allow now to be used in the following
nonstandard manner: in the initial configuration, it is not required that the number of nonblank
symbols be finite. Suppose that T halts for all possible initial configurations of the tape. Prove
that then there is an n such that for all initial configurations, on all tapes, the heads of T stay
within distance n of the origin. ♦

2.10 Exercise Show that there is a kit of dominoes with the property that it tiles the plane
but does not tile it periodically. ♦

2.11 Exercise Let T be a one-tape Turing machines that never overwrites a nonblank symbol
by a blank one. Let the partial function fT (n) be defined if T , started with the empty tape,
will ever write a nonblank symbol in cell n; in this case, let it be the first such symbol. Prove
that there is a T for which fT (n) cannot be extended to a recursive function. ♦
∗ (2.12) Exercise Show that there is a kit of dominoes with the property that it tiles the plane
but does not tile it recursively.
[Hint: Take the Turing machine of Exercise 2.11. Use the kit assigned to it by the proof
of the tiling problem. Again, we will only consider the prototiles associated with the upper
half-plane. We turn each of these prototiles into several others by writing a second tape symbol
on both the top edge and the bottom edge of each prototile P in the following way. If the tape
symbol of both the top and the bottom of P is ∗ or both are different from ∗ then for all symbols
h in Σ0 , we make a new prototile Ph by adding add h to both the top and the bottom of P . If
the bottom of P has ∗ and the top has a nonblank tape symbol h then we make a new prototile
P  by adding h to both the top and the bottom. The new kit for the upper half-plane consists
of all prototiles of the form Ph and P  .] ♦

2.13 Exercise Let us consider the following modifications of the tiling problem.
— In P1 , tiles are allowed to be rotated 180 degrees.

— In P2 , flipping around a vertical axis is allowed.

— In P3 , flipping around the main diagonal axis is allowed.

Prove that there is always a tiling for P1 , the problem P2 is decidable and problem P3 is
undecidable. ♦

2.14 Exercise Show that the following modification of the tiling problem is also undecidable.
We use tiles marked on the corners instead of the sides and all tiles meeting in a corner must
have the same mark. ♦

2.15 Exercise Our proof of Gödel’s theorem does not seem to give a specific sentence ϕT
undecidable for a given minimally adequate theory T . Show that such a sentence can be con-
structed, if the language L used in the definition of “minimally adequate” is obtained by any
standard coding from the nonrecursive r.e. set constructed in the proof of the undecidablity of
the halting problem. ♦

49
3 Computation with resource bounds
3.1 Introduction
The algorithmic solvability of some problems can be very far from their practical solvability.
There are algorithmically solvable problems that cannot be solved, for an input of a given size,
in fewer than exponentially or doubly exponentially many steps (see Chapter 4). Complexity
theory, a major branch of the theory of algorithms, investigates the solvability of individual
problems under certain resource restrictions. The most important resources are time and space
(storage).
We define these notions in terms of the Turing machine model of computation. This defini-
tion is suitable for theoretical study; in describing algorithms, using the RAM is more convenient,
and it also approximates reality better. It follows, however, from our constructions in section 1.3
that from the point of view of the most important types of resource restrictions (e.g. polynomial
time and space) it does not matter, which machine model is used in the definition.
This leads os to the definition of various complexity classes: classes of problems solvable
within given time bounds, depending on the size of the input. Every positive function of
the input size defines such a class, but some of them are particularly important. The most
central complexity class is polynomial time. Many algorithms important in practice run in
polynomial time (in short, are polynomial). Polynomial algorithms are often very interesting
mathematically, since they are built on deeper insight into the mathematical structure of the
problems, and often use strong mathematical tools.
We restrict the computational tasks to yes-or-no problems; this is not too much of a restric-
tion, and pays off in what we gain simplicity of presentation. Note that the task of computing
any output can be broken down to computing its bits in any reasonable binary representation.
Most of this chapter is spent on illustrating how certain computational tasks can be solved
within given resource contraints. We start with the most important case, and show that most
of the basic everyday computational tasks can be solved in polynomial time. These basic
tasks include tasks in number theory (arithmetic operations, greatest common divisor, modular
arithmetic) linear algebra (Gaussian elimination) and graph theory. (We cannot in any sense
survey all the basic algorithms, especially in graph theory; we’ll restrict ourselves to a few that
will be needed later.
Polynomial space is a much more general class than polynomial time (i.e., a much less restric-
tive resource constraint). The most important computational problems solvable in polynomail
space (but most probably not in polynomial time) are games like chess or GO. We give a detailed
description of this connection.
We end with a briefer discussion of other typical complexity classes.

3.2 Time and space

Let us fix some finite alphabet Σ, inluding the blank symbol ∗ and let Σ0 = Σ \ {∗}. In this
chapter, when a Turing machine is used for computation, we assume that it has an input tape
and output tape and k ≥ 1 work tapes. At start, there is a word in Σ∗0 written on the input
tape.

50
 The time demand of a Turing machine T is a function timeT (n) defined as the maximum of
the number of steps taken by T over all possible inputs of length n. We assume timeT (n) ≥ n
(the machine must read the input; this is not necessarily so but we exclude only trivial cases
with this assumption). It may happen that timeT (n) = ∞.
Similarly, the function spaceT (n) is defined as the maximum number, over all inputs of
length n, of all cells on all tapes to which the machine writes. (This way, the cells occupied
by the input are not counted in the space requirement.) We assume that spaceT (n) ≥ 1 (the
machine has some output).
A Turing machine T is called polynomial , if there is a polynomial f (n) such that timeT (n) =
O(f (n)). This is equivalent to saying that there is a constant c such that the time demand of T
is O(nc ). We can define exponential Turing machines similarly (for which the time demand is
c
O(2n ) for some c > 0), and also Turing machines working in polynomial and exponential space.
Now we consider a yes-or-no problem. This can be formalized as the task of deciding whether
the input word x belongs to a fixed language LıΣ∗0 .
We say that a language L ∈ Σ∗0 has time complexity at most f (n), if it can be decided by
a Turing machine with time demand at most f (n). We denote by DTIME(f (n)) the class of
languages whose time complexity is at most f (n). (The letter “D” indicates that we consider here
only deterministic algorithms; later, we will also consider algorithms that are “nondeterministic”
or use randomness). We denote by PTIME, or simply by P , the class of all languages decidable
by a polynomial Turing machine. We define similarly when a laguage has space complexity at
most f (n), and also the language classes DSPACE(f (n)) and PSPACE (polynomial space).

(3.1) Remark It would be tempting to define the time complexity of a language L as the
optimum time of a Turing machine that decides the language. Note that we were more careful
above, and only defined when the time complexity is at most f (n). The reason is that there
may not be a best algorithm (Turing machine) solving a given problem: some algorithms may
work better for smaller instances, some others on larger, some others on even larger etc. Section
3.3 contains a theorem that provides such examples. ♦

(3.2) Remark When we say that the multiplication of two numbers of size n can be performed
in time n2 then we actually find an upper bound on the complexity of a function (multiplica-
tion of two numbers represented by the input strings) rather than a language. The classes
DTIME(f (n)), DSPACE(f (n)), etc. are defined as classes of languages; corresponding classes
of functions can also be defined.
Sometimes, it is easy to give a trivial lower bound on the complexity of a function. Consider
e.g. the function is x · y where x and y are numbers in binary notation. Its computation requires
at least |x| + |y| steps, since this is the length of the output. Lower bounds on the complexity
of languages are never this trivial, since the output of the computation deciding the language
is a single bit. ♦

How to define time on the RAM machine? The number of steps of the Random Access
Machine is not the best measure of the “time it takes to work”. One could (mis)use the fact that
the instructions operate on natural numbers of arbitrary size, and develop computational tricks
that work in this model but use such huge integers that to turn them into practical computations

51
would be impossible. For example, we can simulate vector addition by the addition of two very
large natural numbers.
Therefore, we prefer to characterize the running time of RAM algorithms by two numbers,
and say that “the machine makes at most n steps on numbers with at most k bits”. Similarly,
the space requirement is best characterized by saying that “the machine stores most n numbers
with at most k bits”.
If we want a single number to characterize the running time of a RAM computation, we can
count as the time of a step not one unit but the number of bits of the integers occurring in it
(both register addresses and their contents). Since the number of bits of an integer is essentially
base two logarithm of its absolute value, it is also usual to call this model logarithmic cost
RAM.)
In arithmetical and algebraic algorithms, it is sometimes convenient to count the arithmeti-
cal operations; on a Random Access Machine, this corresponds to extending the set of basic
operations of the programming language to include the subtraction, multiplication, division
(with remainder) and comparison of integers, and counting the number of steps instead of the
running time. If we perform only a polynomial number of operations (in terms of the length of
the input) on numbers with at most a polynomial number of digits, then our algorithm will be
polynomial in the logarithmic cost model.

3.3 Polynomial time I: Algorithms in arithmetic

3.3.1 Arithmetic operations
All basic arithmetic operations are polynomial: addition, subtraction, multiplication and divi-
sion of integers with remainder. (Recall that the length of an integer n as input is the number
of its bits, i.e., log2 n + O(1)). We learn polynomial time algorithms for all these operations
in elementary school (linear time algorithms in the case of addition and subtraction, quadratic
time algorithms in the case of multiplication and division). We also count the comparison of
two numbers as a trivial but basic arithmetic operation, and this can also be done in polynomial
(linear) time.
A less trivial polynomial time arithmetic algorithm the Euclidean algorithm, computing the
greatest common divisor of two numbers.
Euclidean Algorithm. We are given two natural numbers, a and b. Select one that is not
larger than the other, let this be a (say). If a = 0 then the greatest common divisor of a and b
is gcd(a, b) = b. If a > 0 then let us divide b by a, with remainder, and let r be the remainder.
Then gcd(a, b) = gcd(a, r), and it is enough therefore to determine the greatest common divisor
of a and r. Since r < a, tis recurrence will terminate in a finite number of iterations and we get
the greatest common divisor of a and b.
Notice that strictly speaking, the algorithm given above is not a program for the Random
Access Machine. It is a recursive program, and even as such it is given somewhat informally.
But we know that such an informal program can be translated into a formal one, and a recursive
program can be translated into a machine-language program (most compilers can do that).

(3.1) Lemma The Euclidean algorithm takes polynomial time. More exactly, it carries out of
O(log a + log b) arithmetical operations carried out on input (a, b).

52
Proof Since 0 ≤ r < a ≤ b, the Euclidean algorithm will terminate sooner or later. Let us
see that it terminates in polynomial time. Notice that b ≥ a + r > 2r and thus r < b/2. Hence
ar < ab/2. Therefore after log(ab) iterations, the product of the two numbers will be smaller
than 1, hence one of them will be 0, i.e. the algorithm terminates. Each iteration consist of
elementary arithmetic operations, and can be carried out in polynomial time.
It is an important feature of the Euclidean algorithm not only gives the value of the greatest
common divisor, but also delivers integers p, q such that gcd(a, b) = pa + qb. For this, we simply
maintain such a form for all numbers computed during the algorithm. If a = p1 a + q1 b and
b = p2 a + q2 b and we divide, say, b by a with remainder: b = ha + r then

r = (p2 − hp1 )a + (q2 − hp2 )b,

and thus we obtain the representation of the new number r in the form p a + q  b.

(3.2) Remark The Euclidean algorithm is sometimes given by the following iteration: if a = 0
then we are done. If a > b then let us switch the numbers. If 0 < a ≤ b then let b := b − a.
Mathematically, essentially the same thing happens (Euclid’s original algorithm was closer to
this), this algorithm is not polynomial : even the computation of gcd(1, b) requires b iterations,
which is exponentially large in terms of the number log b + O(1) of digits of the input. ♦

The operations of addition, subtraction, multiplication can be carried out in polynomial

times also in the ring of remainder classes modulo an integer m. We represent the remainder
classes by the smallest nonnegative remainder. We carry out the operation on these as on
integers; at the end, another division by m, with remainder, is necessary.
If m is a prime number then we can also carry out the division in the field of the residue
classes modulo m, in polynomial time. This is different from division with remainder! It means
that given integers a, b and m, where 0 ≤ a, b ≤ m − 1 and b = 0, we can compute an integer x
with 0 ≤ x < m such that
bx ≡ a (mod m).
(Such an x is sometimes denoted by a/b (mod m).)
The solution is to apply the Euclidean algorithm to compute the greatest common divisor
of the numbers b, m. Of course, we know in advance that the result is 1. But as remarked, we
also obtain integers p and q such that bp + mq = 1. In other words, bp ≡ 1 (mod m), and thus
b(ap) ≡ a (mod m). So the quotient x we are looking for is the remainder of the product ap
after dividing by m.
We mention yet another application of the Euclidean algorithm. Suppose that a certain
integer x is unknown to us but we know its remainders x1 , . . . , xk with respect to the moduli
m1 , . . . , mk which are all relatively prime to each other. The Chinese Remainder Theorem says
that these remainders uniquely determine the remainder of x modulo the product m = m1 · · · mk .
But how can we compute this remainder?
It suffices to deal with the case k = 2 since for general k, the algorithm follows from this
by mathematical induction. We are looking for an integer x such that x ≡ x1 (mod m1 ) and
x ≡ x2 (mod m2 ) (we also want that 0 ≤ x ≤ m1 m2 − 1, but this we can achieve by dividing
with remainder at the end).

53
 In other words, we are looking for integers x, q1 and q2 such that x = x1 + q1 m1 and
x = x2 + q2 m2 . Subtracting, we get x2 − x1 = q1 m1 − q2 m2 . This equation does not determine
the numbers q1 and q2 uniquely, but this is not important. We can find, using the Euclidean
algorithm, numbers q1 and q2 such that

x2 − x1 = q1 m1 − q2 m2 ,

and compute x = x1 + q1 m1 = x2 + q2 m2 . Then x ≡ x1 (mod m1 ) and x ≡ x2 (mod m2 ), as

desired.
Next, we discuss the operation of exponentiation. Since even to write down the number 2n ,
we need an exponential number of digits (in terms of the length of the input as the number
of binary digits of n), so of course, this number is not computable in polynomial time. The
situation changes, however, if we want to carry out the exponentiation modulo m: then ab is
also a residue class modulo m, and hence it can be represented by log m + O(1) bits. We will
show that it can be not only represented polynomially but also computed in polynomial time.

(3.3) Lemma Let a, b and m be three natural numbers. Then ab (mod m) can be computed
in polynomial time, or more exactly, with O(log b) arithmetical operations, carried out on natural
numbers with O(log m + log a) digits.

Algorithm Let us write b in binary:

b = 2r1 + · · · + 2rk ,

where 0 ≤ r1 < · · · < rk . It is obvious that rk ≤ log b and therefore k ≤ log b. Now, the numbers
a2 (mod m) for 0 ≤ t ≤ log b are easily obtained by repeated squaring, and then we multiply
t

those k together that make up ab . Of course, we carry out all operations modulo m, i.e., after
each multiplication, we also perform a division with remainder by m.
a
(3.4) Remark It is not known whether a! mod m or b mod m can be computed in polynomial
time. ♦

3.3.2 Gaussian elimination

The basic operations of linear algebra are polynomial: addition and inner product of vectors,
multiplication and inversion of matrices, the computation of determinants. However, these facts
are non-trivial in the last two cases, so we will deal with them in detail.
Let A = (aij ) be an arbitrary n × n matrix consisting of integers.
Let us verify, first of all, that the polynomial computation of det(A) is not inherently im-
possible, in the sense that the result can be written down with polynomially many bits. Let
K = max |aij |, then to write down the matrix A we need obviously at least L = n2 + log K bits.
On the other hand, the definition of determinants gives

| det(A)| ≤ n!K n ,

54
hence det(A) can be written down using

log(n!K n ) + O(1) ≤ n(log n + log K) + O(1)

bits. This is polynomial in L.

Linear algebra gives a formula for each element of det(A−1 ) as the quotient of two sub-
determinants of A. This shows that A−1 can also be written down with polynomially many
bits.
The usual procedure to compute the determinant is Gaussian elimination. We can view
this as the transformation of the matrix into a lower triangular matrix with column operations.
These transformations do not change the determinant, and in the final triangular matrix, the
computation of the determinant is trivial: we just multiply the diagonal elements to obtain
it. It is also easy to obtain the inverse matrix from this form; we will not deal with this issue
separately.
Gaussian elimination. Suppose that for all i such that 1 ≤ i ≤ t, we have achieved already
that in the i’th row, only the first i entries hold a nonzero element. Pick a nonzero element
from the last n − t columns (stop if if there is no such element). Call this element the pivot
element of this stage. Rearrange the rows and columns so that this element gets into position
(t + 1, t + 1). Subtract column t + 1, multiplied by at+1,i /at+1,t+1 , from column i column for all
i = t + 2, . . . , n, in order to get 0’s in the elements (t + 1, t + 2), . . . , (t + 1, n). These subtractions
do not change value of the determinant and the rearrangement changes at most the sign, which
is easy to keep track of.
Since one iteration of the Gaussian elimination uses O(n2 ) arithmetic operations and n
iterations must be performed, this procedure uses O(n3 ) arithmetic operations. But the problem
is that we must also divide, and not with remainder. This does not cause a problem over a finite
field, but it does in the case of the rational field. We assumed that the elements of the original
matrix are integers; but during the run of the algorithm, matrices also occur that consist of
rational numbers. In what form should these matrix elements be stored? The natural answer
is that as pairs of integers (whose quotient is the rational number).
Do we require that the fractions be in simplified form, i.e., that their numerator and denom-
inator be relatively prime to each other? We could do so; then we have to simplify each matrix
element after each iteration, for which we would have to perform the Euclidean algorithm. This
can be performed in polynomial time, but it is a lot of extra work, and it is desirable to avoid
it. (Of course, we also have to show that in the simplified form, the occurring numerators and
denominators have only polynomially many digits. This will follow from the discussions below.)
We could also choose not to require that the matrix elements be in simplified form. Then
we define the sum and product of two rational numbers a/b and c/d by the following formulas:
(ad + bc)/(bd) and (ac)/(bd). With this convention, the problem is that the numerators and
denominators occurring in the course of the algorithm can become very large (have a nonpoly-
nomial number of digits)!
Fortunately, we can give a procedure that stores the fractions in partially simplified form,
and avoids both the simplification and the excessive growth of the number of digits. For this, let
us analyze a little the matrices occurring during Gaussian elimination. We can assume that the
pivot elements are, as they come, in positions (1, 1), . . . , (n, n), i.e., we do not have to permute

55
 (k)
the rows and columns. Let (aij ) (1 ≤ i, j ≤ n) be the matrix obtained after k iterations. Let
us denote the elements in the main diagonal of the final matrix, for simplicity, by d1 , . . . , dn
(n)
(thus, di = aii ). Let D(k) denote the submatrix determined by the first k rows and columns of
(k)
matrix A, and let Dij , for k + 1 ≤ i, j ≤ n, denote the submatrix determined by the first k rows
(k) (k)
and the ith row and the first k columns and the jth column. Let dij = det(Dij ). Obviously,
(k−1)
det(D(k) ) = dkk .

(3.5) Lemma
(k)
(k) dij
aij = .
det(D(k) )

(k)
Proof If we compute det(Dij ) using Gaussian elimination, then in its main diagonal, we
(k)
obtain the elements d1 , . . . , dk , aij . Thus

(k) (k)
dij = d1 · · · dk · aij .

Similarly,
det(D(k) ) = d1 · · · dk .
Dividing these two equations by each other, we obtain the lemma.
By this lemma, every number occurring in the Gaussian elimination can be represented as a
fraction both the numerator and the denominator of which is a determinant of some submatrix
of the original A matrix. In this way, a polynomial number of digits is certainly enough to
represent all the fractions obtained.
However, it is not necessary to compute the simplifications of all fractions obtained in the
process. By the definition of Gaussian elimination we have that
(k) (k)
(k+1) (k) ai,k+1 ak+1,j
aij = aij − (k)
ak+1,k+1

and hence
(k) (k) (k) (k)
(k+1) dij dk+1,k+1 − di,k+1 dk+1,j
dij = (k−1)
.
dk,k
(k)
This formula can be considered as a recurrence for computing the numbers dij . Since the left-
hand side is an integer, the division can be carried out exactly. Using the above considerations,
we find that the number of digits in the quotient is polynomial in terms of the size of the input.
There are at least two further possibilities to remedy the problem of the fractions occurring
in Gaussian elimination.
We can approximate the numbers by binary “decimals” of limited accuracy (as it seems
natural from the point of view of computer implementation), allowing, say, p bits after the
binary “decimal point”. Then the result is only an approximation, but since the determinant
is an integer, it is enough to compute it with an error smaller than 1/2. Using the methods of

56
numerical analysis, it can be determined how large p must be chosen to make the error in the
end result smaller than 1/2. It turns out that a polynomial number of digits are enough, and
this leads to a polynomial algorithm.
The third possibility is based on the remark that if m > | det(A)| then it is enough to
determine the value of det(A) modulo m. If m is a prime number then computing modulo m,
we don’t have to use fractions at all. Since we know that | det(A)| < n!K n it is enough to choose
for m a prime number greater than n!K n .
It is, however, not quite easy to select such a large prime (see the section on randomized
algorithms). An easier method is to choose m as the product of different small primes: m = 2 ·
3 · · · pk where for k we can be choose, e.g., the total number of bits occurring in the representation
of A. Then it is easy to compute the remainder of det(A) modulo pi for all pi , using Gaussian
elimination in the field of residue classes modulo pi . Then we can compute the remainder of
det(A) modulo m using the Chinese Remainder Theorem. (Since k is small we can afford to
find the first k primes simply by brute force. But the cost of this computation must be judged
differently anyway since the same primes can then be used for the computation of arbitrarily
many determinants.)

(3.6) Remark The modular method is successfully applicable in a number of other cases. One
way to look at this method is to consider it as an encoding of the integers in a way different from
the binary (or decimal) number system: we code the integer n by its remainder after division
by the primes 2,3, etc. This is an infinite number of bits, but if we know in advance that no
number occurring in the computation is larger than N then it is enough to consider the first
k primes whose product is larger than N . In this encoding, the arithmetic operations can be
performed very simply, and even in parallel for the different primes. Comparison by magnitude
is, however, awkward. ♦

3.4 Polynomial time II: Graph algorithms

The most important algorithms of graph theory are polynomial. It is impossible to survey
this topic in this course; graph algorithms can fill several books. We restrict ourselves to a
brief discussion of a few examples that provide particularly important insight into some basic
techniques. We’ll group our examples around two algorithmic methods: searching a graph and
augmenting paths.

3.4.1 How is a graph given?

3.4.2 Searching a graph
a. General search and connectivity testing. Perhaps the most fundamental question to
ask about a graph is whether or not it is connected. If not, we often want to find its connected
components.
These tasks are trivial to do in polynomial time, in a variety of ways. The reason why we
describe a solution is because its elements will be useful later.
Let G be a graph. Select any node r. Build up a tree T as follows. at the beginning, T
consists of just r. At any stage, we look if there is an edge between the nodes of T and the rest

57
of the node. If there is such an edge, we add it to T , together with its endnode outside T . If
not, then we know that G is disconnected and T is a spanning tree of a connected component.
We can delete the nodes of T grom the graph and proceed recursively.
It is costumary to call the nodes of T labelled. The usual way to look for an edge between a
labelled and an unlabelled node is to look at the labelled nodes and investigate the edges going
out of them. A simple but important observation is that if at one stage we find that none of
the edges of a given labelled node goes to an unlabelled node, then we don’t need to investigate
this node at any other later stage (since only an unlabelled node can become labelled, not vice
versa). Therefore we can mark this node as taken care of, or scanned. At any time, we only
need to look at those nodes that are labelled but not scanned.
This general labelling-scanning procedure is called searching the graph.
b. Breadth-First-Search and shortest paths. Specializing the order in which we investigate
edges out of the labelled nodes, we get special search algorithms that are good for solving
different kinds of graph problems. Perhaps the simplest such specialization is depth-first-search.
The problem is to find a shortest path in a graph from a distinguished node s (the source) to
a distinguished node t (the sink). The solution of this problem is very simple once we embed
the problem in the larger family of problems:we ask for the shortest path from the source s to
every other node of the graph. Proceed recursively: label the neighbors of the source s with a
1. Label with k those unlabelled vertices which are neighbors of vertices of label k − 1. Then
the label of each node is its distance from s. (This method is perhaps the simplest example of
a technique called dynamic programming.
This idea reappears in the more general problem of finding a shortest path from s to t in a
graph with weighted edges. (The weight cij on the edge {i, j} is interpreted as the length of the
edge.) Again we embed this problem in the larger family of problems which asks for a shortest
(minimum weight) path from s to every other node of the graph.
Dijkstra’s algorithm recognizes that a path between two vertices may be of minimum weight
even if other paths have fewer edges. So if we begin at the source, among all neighbors j of
s, find one such that the edge sj has minimum weight csj . We can confidently assert that the
shortest path in the graph from s to this neighbor has length csj , but we are not sure about
the other neighbors of s. So label this one neighbor csj . You may think of s as labelled with
a 0. In the course of the algorithm we maintain a set T of vertices, each of which we have
confidently labelled with the minimum weight of a path from s to it. Call this label d(s, j), the
distance from s to the node j. These minimum weight paths pass only through vertices already
in T . At each step of the algorithm, we consider vertices on the frontier of T (vertices not in T
which are neighbors of vertices in T ). Consider edges ij between vertices i in T and vertices j
in this frontier. An upper bound for the minimum weight of a path from s to j is the smallest
of the numbers d(s, i) + cij where i is in T . Find i ∈ T and j which minimize d(s, i) + cij . We
can confidently label such a node j with d(s, j) = d(s, i) + cij and add j to T for the following
reason: Any path from s to j must leave the set T at a node i then pass through a node j 
of the frontier. Hence this competing path from s to j has length at least d(s, i ) + ci j  . By
our choice of i and j, the path from s to i in T then immediately to j has length no greater
d(s, i ) + ci j  (figure 3.1).
c. Depth-First-Search, 2-connected components, and 3-connected components.
d. Depth-First-Search and planarity.

58
Figure 3.1: Dijkstra’s shortest path algorithm.

59
3.4.3 Maximum bipartite matching and alternating paths
Let G be a graph; we want to find a maximum size matching in G (recall that a matching is
subset of the edges having mutually no endpoint in common). A matching is perfect if it mmets
every node. We describe a polynomial time algorithm that finds a maximum size matching in
a bipartite graph.
Assume that we have a matching M ; how can we see whether M is maximum? One answer
to this question is provided by the following simple criterion due to Berge. An alternating path
(with respect to M ) is a path in which every other edge belongs to M (the path may start with
an edge of M or with an edge not in M ). An alterning path is augmenting if it starts and ends
with an edge not in M , and moreover, its endpoints are not incindent with any edge of M . (An
augmenting path has necessarily odd length.)

(3.1) Lemma A matching in G is maximum if and only if there is no augmenting path with
respect to it.

Proof It is obvious that if we have an augmenting path P with respect to M , then we can
improve M : we can replace in M the edges of P ∩ M by the edges of P \ M , to get a larger
matching.
Conversely, if M is not maximum, and there exists a larger matching M  , then consider the
connected components of M ∪ M  . These are either alternating paths or alternating circuits
with respect to M . At least one such component must contain more edges of M  than of M :
this coponent is an augmenting path with respect to M .
This fact is not very deep since it relates the existence of something (a larger matching) to
the existence of something else (an augmenting path). But it does give a motivation for most
of the known polynomial time matching algorithms: one can look for an augmenting path by
building up a search tree.
Given a bipartite graph G = (A, B), we want to find a maximal matching M . We use the
notion of alternating trees. Given a matching M , we call a vertex of the graph exposed (by M )
if it is not covered by M . An alternating tree is a subgraph of G which contains exactly one
exposed node r; for which every node at odd distance from r has degree 2 in the tree; such that
along any path from r every other edge belongs to the matching; and whose endpoints are all
at even distance from r.
The vertices of the tree at even distance from r (including r itself) are called outer nodes,
and the vertices at odd distance from r are called inner nodes.
An alternating forest is a forest such that every component is an alternating tree, and every
exposed node of the graph is contained in the forest.
If we have a matching M and would like to extend it, we take an exposed vertex r and
try to match it to a neighbour; but the neighbor may already be in the matching, so we leave
this edge out; this creates a new exposed node which we try to match etc. This leads to the
notion of alternating paths. Searching for alternating paths from an exposed node r leads to
the construction of an alternating tree.
Now we describe the algorithm. We start with the empty matching and the alternating
forest consisting of all nodes and no edges. At every stage, either we can increase the size of the

60
 e e

u v u v

Figure 3.2: Extending a bipartite graph matching

v v

u u
p
p

Figure 3.3: Extending an alternating forest

matching (deleting some edges but adding more) or we can extend the forest already constructed
(and keeping the matching unchanged), or we are stuck and the matching is optimal.
Suppose that we already have a candidate matching M and alternating forest F with respect
to M .
Case 1. If there is an edge e connecting two outer nodes of the forest, then these nodes
must belong to different trees since the graph is bipartite. Let u and v be the exposed nodes
of the two trees. We have an alternating path from u to v, and switching the matching edges
along the path we get a new matching M  with one more edge than M . The forest consisting
of all exposed points of M  is alternating with respect to M  . (See figure 3.2).
Case 2. Otherwise, if there is an edge connecting an outer node p of the forest to a vertex
u of V (G) \ V (F ), this vertex u cannot be exposed from the definition of an alternating forest.
Hence it is the endpoint of an edge of the matching, say uv. Vertex v is also in V (G) \ V (F )
(otherwise u would have to be in F ), and so we can extend F by adding the edges pu and uv.
(See figure 3.3).
Case 3. If neither of the cases above applies, the outer points must be connected to inner

61
points only. A simple but crucial counting argument will show that the matching must be
optimal. (We will use this counting argument several times during this lecture).
There is one exposed vertex in every component of F , and none outside F , so the total
number of exposed points of the matching is equal to the number of components of F . Each
outer point of the trees of F is matched to an inner point, which is its predecessor in the tree,
except for the exposed vertex of each tree. Thus

#outer points = #inner points + #exposed points.

Let D be the set of outer vertices and A be the set of inner vertices of F . The neighbor set of
D in the graph is included in A, the number of exposed points of M is exactly |D| − |A|.
Let M  be any matching of the graph. M  can only match points of D to points of A, and
must therefore it must leave at least |D|−|A| points exposed. Hence M is a maximum matching.
Thus we have a valid algorithm for finding a maximum matching in a bipartite graph. At
every step we must consider at most n2 edges, and since the size of M or of F increases at
each step, there are at most n2 steps, hence a O(n4 ) algorithm. Simple considerations enable
to reduce the running time to O(n3 ). It is possible, but quite difficult, to reduce it further to
O(n2.5 ).

3.5 Polynomial space

Obviously, all polynomial-time algorithms require only polynomial space, but polynomial space
is significantly more general: many exponential time algorithms need only polynomial space.
(It is clear that using polynomial space, we terminate in exponential time or never: if the same
configuration (state of the CU, position of heads, contents of tape) is repeated, then we are in
cycle).
For example, the space requirement of the trivial graph coloring algorithm treated in is
polynomial (even linear): if we survey all colorings in lexicographic order then it is sufficient
to keep track of which coloring is currently checked and for which edges has it already been
checked whether they connect points with the same color. In a similar way, we could carry out
in polynomial space every brute force algorithm that tries to find in a graph a Hamilton cycle,
the largest clique, etc.
The most typical example for a polynomial-space algorithm is finding the optimal step in a
2-person game like chess, by searching through all possible continuations.
Let us describe a general model for a “chess-like” game.
We assume that every given position of a game can be described by a word x of length n
over some finite alphabet Σ (typically, telling the position of the pieces on the board, the name
of the player whose turn it is and possibly some more information, e.g. in the case of chess
whether the king has moved already, etc.). An initial configuration is distinguished. We assume
that the two players take turns and have a way to tell, for two given positions whether it is legal
to move from one into the other, by an algorithm taking polynomial time1 If there is no legal
move in a position then it is a terminal position and a certain algorithm decides in polynomial
time, who won. We assume that the game is always finished in nc steps.
1
It would be sufficient to assume polynomial space but it makes the game rather boring if it takes more than
polynomial time to decide whether a move is legal or not.

62
 A position is winning for the player whose turn it is if there is a strategy that, starting from
this position, leads to the victory of this player no matter what the other player does. A position
is losing if it no matter what move the player makes, his opponent has a winning strategy. Since
the game is finite, every position is either a winning or losing position. Note that
— — every legal move from a losing position leads to a winning position;
— — there is a move from any winning position that leads to a losing position.
We can visualize a game as a huge rooted tree. The root is labelled by the starting position.
The children of the root are labelled by the positions that can be reached from the starting
position by a legal move. Similarly, each node i of the tree is labelled by a position xi in the
game, and its children are labelled by the positions that can be reached from xi by a legal move.
The leaves of the tree are the terminal positions.
Note that the same position may occur more than once as a label, if there are different
sequences of legal moves leading to that position. A node of the tree corresponds to a subgame,
i.e., to a sequence x0 , x1 , . . . , xk of positions such that a legal step leads from each position xi
into xi+1 . (We could identify nodes of the tree that correspond to the same position, and work
on the resulting graph, which is not a tree any more. However, it is moree transparent to have
the whole tree around.)
We describe an algorithm that decides about each position in such a game whether it is a
winning or losing position. The algorithm works in polynomial space but exponential time.
Algorithm to find an optimal move. Roughly speaking, we search the game-tree in a
depth-first-search manner, and label positions “winning’ and “loosing”.
At a given step, the machine analyzes all possible continuations of some partial game
x0 , x1 , . . . , xk . It will always be maintained that among all continuations of x0 , x1 , . . . , xi
(0 ≤ i ≤ k) those that come before xi+1 (with respect to the lexicographical ordering of the
words of length n) are “bad steps”, i.e. they are either illegal moves or lead to a winning po-
sition (the position is winning for the other player, so to go there is a bad move!). In other
words, if there are “good” moves at all, the algorithm finds the lexicographically first good
move. The algorithm maintains the last legal continuation y of xk it has studied (or that no
such continuation was found yet). All the legal continuations that come before y lead to winning
positions.
The algorithm looks at the words of length n following y in lexicographical order, to see
whether they are legal continuations of xk . If it finds one, this is xk+1 and goes on to examine
the one longer partial game obtained this way. If it does not find such a move, and it did not
find legal move from this position, that this position is terminal, and the algorithm marks it
“winning” or “loosing” according to the game rules. If there was at least one legal move from xk ,
then we know that all legal moves from this position lead to winning positions, so this position
is loosing. The algorithm marks it so, and removes xk from the partial game. It also marks
xk−1 as winning (all you have to do to win is to move to xk ), and removes it from the partial
game.
Eventually, the root gets labelled, and that tells is who wins in this game.
3.1 Exercise The Fibonacci numbers are defined by the following recurrence: F0 = 0, F1 = 1,
Fk = Fk−1 + Fk−2 for k > 1. Let 1 ≤ a ≤ b and let Fk denote the greatest Fibonacci number not

63
greater than b. Prove that the Euclidean algorithm, when applied to the pair (a, b), terminates
in at most k steps. How many steps does the algorithm take when applied to (Fk , Fk−1 )? ♦

3.2 Exercise (a) Show that for computing a/b (mod m), we don’t need that m is a prime;
it suffices to assume that the greatest common divisor of b and m is 1.
(b) More generally, show that given integers a, b and m, we can find find in polynomial time
an integer x such that bx ≡ a (mod m), or conclude that no such integer exists. ♦

3.3 Exercise Show that if A is a square matrix consisting of integers, then to write down
det(A) we need at most as many bits as to write up A. [Hint: If a1 , . . . , an are the row vectors
of A then | det(A)| ≤ |a1 | · · · |an | (this so-called “Hadamard Inequality” is analogous to the
statement that the area of a parallelogram is smaller than the product of the lengths of its
sides).] ♦

3.4 Exercise Let A be an n × n matrix with 2’s in the main diagonal, 1’s below it and 0’s
above it. Show that if Gaussian elimination without simplification is applied to this matrix then
after k iterations, each of the denominators will be 22 −1 . ♦
k

64
4 General theorems on space and time complexity
If for a language L, there is a Turing machine deciding L for which for all large enough n the
relation timeT (n) ≤ f (n) holds then there is also a Turing machine recognizing L for which
this inequality holds for all n. Indeed, for small values of n we assign the task of deciding the
language to the control unit.
It can be expected that for the price of further complicating the machine, the time demands
can be decreased. The next theorem shows the machine can indeed be accelerated by an arbitrary
constant factor, at least if its time need is large enough (the time spent on reading the input
cannot be saved).

(4.1) Linear Speedup Theorem For every Turing machine and c > 0 there is a Turing
machine S over the same alphabet wich decides the same language an for which timeS (n) ≤
c · timeT (n) + n.

Proof For simplicity, let us also assume that T has a single work tape (the proof would be
similar for k tapes). We can assume that c = 1/p where p is an integer.
Let the Turing machine S have an input-tape, 2p − 1 “starting” tapes and 2p − 1 further
work tapes. Let us number these each from 1 − p to p − 1. Let the index of cell j of (start- or
work) tape i be the number j(2p − 1) + i. The start- or work cell with index t will correspond
to cell t on the input resp. worktape of machine T . Let S also have an output tape.
Machine S begins its work by copying every letter of input x from its input tape to the cell
with the corresponding index on its starting tapes, then moves every head back to cell 0. From
then on, it ignores the “real” input tape.
Every further step of machine S will correspond p consecutive steps of machine T . After pk
steps of machine T , let the scanning head of the input tape and the work tape rest on cells t
and s respectively. We will plan machine S in such a way that in this case, each cell of each
start- resp. worktape of S holds the same symbol as the corresponding cell of the corresponding
tape of T , and the heads rest on the starting-tape cells with indices t − p + 1, . . . , t + p − 1 and
the work-tape cells with indices s − p + 1, . . . , s + p − 1. We assume that the control unit of
machine S “knows” also which head scans the cell corresponding to t resp. s. It knows further
what is the state of the control unit of T .
Since the control unit of S sees not only what is read by T ’s control unit at the present
moment on its input- and worktape but also the cells at a distance at most p − 1 from these, it
can compute where T ’s heads will step and what they will write in the next p steps. Say, after
p steps, the heads of T will be in positions t + i and s + j (where, say, i, j > 0). Obviously,
i, j < p. Notice that in the meanwhile, the “work head” could change the symbols written on
the work tape only in the interval [s − p + 1, s + p − 1].
Let now the control unit of S do the following: compute and remember what will be the
state of T ’s control unit p steps later. Remember wich heads rest on the cells corresponding to
the positions (t + i) and (s + j). Let it rewrite the symbols on the work tape according to the
configuration p steps later (this is possible since there is a head on each work cell with indices
in the interval [s − p + 1, s + p − 1]). Finally, move the start heads with indices in the interval
[t − p + 1, t − p + i] and the work heads with indices in the interval [s − p + 1, s − p + j] one

65
step right; in this way, the indices occupied by them will fill the interval [t + p, t + p + i − 1]
resp. [s + p, s + p + i − 1] which, together with the heads that stayed in their place, gives interval
[t + i − p + 1, t + i + p − 1] resp. [s + j − p + 1, s + j + p − 1].
If during the p steps under consideration, T writes on the output tape (either 0 or 1) and
stops, then let S do this, too. Thus, we constructed a machine S that (apart from the initial
copying) makes only a pth of the number of steps of T and decides the same language.

∗ (4.1)Exercise For every Turing machine T and c > 0, one can find a Turing machine S
with the same number of tapes that decides the same language and for which timeS (n) ≤
c · timeT (n) + n (here, we allow the extension of the alphabet.♦

4.2 Exercise Formulate and prove the analogue of the above problem for storage in place of
time. ♦

It is trivial that the storage demand of a k-tape Turing machine is at most k times its
time demand (since in one step, at most k cells will be written). Therefore if we have L ∈
DTIME(f (n)) for a language then there is a constant k (depending on the language) that
L ∈ DSPACE(k · f (n)). (If extending the alphabet is allowed and f (n) > n then DSPACE(k ·
f (n)) = DSPACE(f (n)) and thus it follows that DTIME(f (n)) ⊂ DSPACE(f (n)).) On the
other hand, the time demand is not greater than an exponential function of the space demand
(since exactly the same memory configuration, taking into account also the positions of the
heads and the state of the control unit, cannot occur more than once without getting into a
cycle). Computing more precisely, the number of different memory configurations is at most
c · f (n)k mf (n) where m is the size of the alphabet.
Since according to the above, the time complexity of a language does not depend on a
constant factor, and in this upper bound the numbers c, k, m are constants, it follows that if
f (n) > log n and L ∈ DSPACE(f (n)) then L ∈ DTIME((m + 1)f (n) ).
A recursive language can have arbitrarily large time (and, due to the above inequality, also
space-) complexity. More precisely:

(4.2) Theorem For every recursive function f (n) there is a recursive language L that is not
an element of DTIME(f (n)).

Proof The proof is similar to the proof of the fact that the halting problem is undecidable.
We can assume f (n) > n. Let T be the 2-tape universal Turing machine constructed in Chapter
1, and let L consist of all words x for which it is true that having x as input on both of its tape,
T halts in at most f (|x|)4 steps. L is obviously recursive.
Let us now assume that L ∈ DTIME(f (n)). Then there is a Turing machine (with some
k > 0 tapes) deciding L in time f (n). From this we can construct a 1-tape Turing machine
deciding L in time cf (n)2 (e.g. in such a way that it stops and writes 0 or 1 as its decision on a
certain cell). Since for large enough n we have cf (n)2 < f (n)3 , and the words shorter than this
can be recognized by the control unit directly, we can also make a 1-tape Turing machine that
always stops in time f (n)3 . Let us modify this machine in such a way that if a word x is in L
then it runs forever, while if x ∈ Σ∗0 \ L then it stops. This machine be S can be simulated on

66
T by some program p in such a way that T halts with input (x, p) if and only if S halts with
input x; moreover, it halts in these cases within |p|f (|x|)3 steps.
There are two cases. If p ∈ L then—according to the definition of L—starting with input p
on both tapes, machine T will stop. Since the program simulates S it follows that S halts with
input p. This is, however, impossible, since S does not halt at all for inputs from L.
On the other hand, if p ∈ L then—according to the construction of S—starting with p on
its first tape, this machine halts in time |p|f (|p|)3 < f (|p|)4 . Thus, T also halts in time f (|p|)4 .
But then p ∈ L by the definition of the language L.
This contradiction shows that L ∈ DTIME(f (n)).
There is also a different way to look at the above result. For some fixed universal two-tape
Turing machine U and an arbitrary function t(n) > 0, the t-bounded halting problem asks,
for n and all inputs p, x of maximum length n, whether the above machine U halts in t(n)
steps. (Similar questions can be asked about storage.) This problem seems decidable in t(n)
steps, though this is true only with some qualification: for this, the function t(n) must itself be
computable in t(n) steps (see the definition of “fully time-constructible” below). We can also
expect a result similar to the undecidability of the halting problem, saying that the t-bounded
halting problem cannot be decided in time “much less” than t(n). How much less is “much less”
here depends on some results on the complexity of simulation between Turing machines.
We call a function f : Z+ → Z+ fully time-constructible if there is a multitape Turing
machine that for each input of length n using exactly f (n) time steps. The meaning of this
strange definition is that with fully time-constructable functions, it is easy to bound the running
time of Turing machines: If there is a Turing machine making exactly f (n) steps on each input
of length n then we can build this into any other Turing machine as a clock: their tapes, except
the work tapes, are different, and the combined Turing machine carries out in each step the
work of both machines.
Obviously, every fully time-constructible function is recursive. On the other hands, it is easy
to see that n2 , 2n , n! and every “reasonable” function is fully time-constructible. The lemma
below guarantees the existence of many completely time-constructable functions.
Let us call a function f : Z+ → Z+ well-computable if there is a Turing machine computing
f (n) in time O(f (n)). (Here, we write n and f (n) in unary notation: the number n is given
by a sequence 1 . . . 1 of length n and we want as output a sequence 1 . . . 1 of length f (n). The
results would not be changed, however, if n and f (n) were represented e.g. in binary notation.)
Now the following lemma is easy to prove:

(4.3) Lemma
(a) To every well-computable function f (n), there is a fully time-constructible function g(n)
such that f (n) ≤ g(n) ≤ const · f (n).
(b) For every fully time-constructible function g(n) there is a well-computable function f (n)
with g(n) ≤ f (n) ≤ const · g(n).
(c) For every recursive function f there is a fully time-constructible function g with f ≤ g.

This lemma allows us to use, in most cases, fully time-constructible and well-computable
functions interchangeably. Following the custom, we will use the former.

67
(4.4) Theorem If f (n) is fully time-constructible and g(n) log g(n) = o(f (n)) then there is a
language in DTIME(f (n)) that does not belong to DTIME(g(n)).

This says that the time complexities of recursive languages are “sufficiently dense”. Analo-
gous, but easier, results hold for storage complexities.

4.3 Exercise Prove the above theorem, and the following, closely related statement. Let
t (n) log t (n) = o(t(n)). Then the t(n)-bounded halting problem cannot be decided on a two-
tape Turing machine in time t (n). ♦

4.4 Exercise Show that if S(n) is any function and S  (n) = o(S(n)) then the S(n) space-
bounded halting problem cannot be solved in time S  (n). ♦

The full time-constructibility of the function f plays very important role in the last theorem.
If we drop it then there can be an arbitrarily large “gap” below f (n) which contains the time-
complexity of no language at all.

(4.5) Gap Theorem For every recursive function ϕ(n) ≥ n there is a recursive function f (n)
such that
DTIME(ϕ(f (n))) = DTIME(f (n)).

Thus, there is a recursive function f with

DTIME(f (n)2 ) = DTIME(f (n)),
moreover, there is even one with
f (n)
DTIME(22 ) = DTIME(f (n)).

Proof Let us fix a 2-tape universal Turing machine. Denote τ (x, y) the time needed for T
compute from input x on the first tape and y on the second tape. (This can also be infinite.)

1. Claim There is a recursive function h such that for all n > 0 and all x, y ∈ Σ∗0 , if |x|, |y| ≤ n
then either τ (x, y) ≤ h(n) or τ (x, y) ≥ (ϕ(h(n)))3 .
If the function
ψ(n) = max{ τ (x, y) : |x|, |y| ≤ n, τ (x, y) is finite }
was recursive this would satisfy the conditions trivially. This function is, however, not recursive
(exercise: prove it!). We introduce therefore the following “constructive version”: for a given n,
let us start from the time bound t = n + 1. Let us arrange all pairs (x, y) ∈ (Σ∗0 )2 , |x|, |y| ≤ n
in a queue. Take the first element (x, y) of the queue and run the machine with this input. If
it stops within time t then throw out the pair (x, y). If it stops in s steps where t < s ≤ ϕ(t)3
then let t := s and throw out the pair (x, y) again. (Here, we used that ϕ(n) is recursive.) If
the machine does not stop even after ϕ(t)3 steps then stop it and place the pair (x, y) to the
end of the queue. If we have passed the queue without throwing out any pair then let us stop,
with h(n) := t. This function clearly has the property formulated in the Claim.

68
2. We will show that with the function h(n) defined above,

DTIME(h(n)) = DTIME(ϕ(h(n))).

For this, consider an arbitrary language L ∈ DTIME(ϕ(h(n))) (containment in the other direc-
tion is trivial). To this, a Turing machine can thus be given that decides L in time ϕ(h(n)).
Therefore a one-tape Turing machine can be given that decides L in time ϕ(h(n))2 . This latter
Turing machine can be simulated on the given universal Turing machine T with some program
p on its second tape, in time, |p| · ϕ(h(n)). Thus, if n is large enough then T works on all
inputs (y, p) (|y| ≤ n) for at most ϕ(h(n))3 steps. But then, due to the definition of h(n),
it works on each such input at most h(n) steps. Thus, this machine decides, with the given
program (which we can also put into the control unit, if we want) the language L in time h(n),
i.e. L ∈ DTIME(h(n)).
As a consequence of the theorem, we see that there is a recursive function f (n) with

DTIME((m + 1)f (n) ) = DTIME(f (n)),

and thus

DTIME(f (n)) = DSPACE(f (n)).

For a given problem, there is often no “best” algorithm: the following surprising theorem is
true.

(4.6) Speed-up Theorem For every recursive function g(n) there is a recursive language L
such that for every Turing machine T deciding L there is a Turing machine S deciding L with
g(timeS (n)) < timeT (n).

The Linear Speedup Theorem applies to every language; this theorem states only the exis-
tence of an arbitrarily “speedable” language. In general, for an arbitrary language, better than
linear speed-up cannot be expected.

Proof The essence of the proof is that as we allow more complicated machines we can “hard-
wire” more information into the control unit. Thus, the machine needs to work only with longer
inputs “on their own merit”, and we want to construct the language in such a way that this
should be easier and easier. It will not be enough, however, to hard-wire only the membership
or non-membership of “short” words in L, we will need more information about them.
Without loss of generality, we can assume that g(n) > n and that g is a fully time-
constructable function. Let us define a function h with the recursion

h(0) = 1, h(n) = (g(h(n − 1)))3 .

It is easy to see that h(n) is a monotonically increasing (in fact, very fast increasing), fully
time-constructable function. Fix a universal Turing machine T0 with, say, two tapes. Let
τ (x, y) denote the time spent by T0 working on input (x, y) (this can also be infinite). Let us
call the pair (x, y) “fast” if |y| ≤ |x| and τ (x, y) ≤ h(|x| − |y|).

69
 Let (x1 , x2 , . . .) be an ordering of the words e.g. in increasing order; we will select a word yi
for certain indices i as follows. For each index i = 1, 2, . . . in turn, we check whether there is a
word y not selected yet that makes (xi , y) fast; if there are such words let yi be a shortest one
among these. Let L consist of all words xi for which yi exists and the Turing machine T0 halts
on input (xi , yi ) with the word “0” on its first tape. (These are the words not accepted by T0
with program yi .)
First we convince ourselves that L is recursive, moreover, for all natural numbers k the
question x ∈ L is decidable in h(n − k) steps (where n = |x|) if n is large enough. We can decide
the membership of xi if we decide whether yi exists, find yi (if it exists), and run the Turing
machine T0 on input (xi , yi ) for time h(|xi | − |yi |).
This last step itself is already too much if |yi | ≤ k; therefore we make a list of all pairs (xi , yi )
with |yi | ≤ k (this is a finite list), and put this into the control unit. This begins therefore by
checking whether the given word x is in this list as the first element of a pair, and if it is,
it accepts x (beyond the reading of x, this is only one step!). Suppose that xi is not in the
list. Then yi , if it exists, is longer than k. We can try all inputs (x, y) with k < |y| ≤ |x| for
“fastness” and this needs only (m2 n + 1)h(n − k − 1) (including the computation of h(|x| − |y|)).
The function h(n) grows so fast that this is less than h(n − k). Now we have yi and also see
whether T0 accepts the pair (xi , yi ).
Second, we show that if a program y accepts the language L on on the machine T0 (i.e. stops
for all Σ∗0 writing 1 or 0 on its first tape according to whether x is in the language L) then y
cannot be equal to any of the selected words yi . This follows by the usual “diagonal” reasoning:
if yi = y then let us see whether xi is in the language L. If yes then T0 must give result “1” for
the pair (xi , yi ) (since y = yi decides L). But then according to the definition of L, we did not
put xi into it. Conversely, if xi ∈ L then it was left out since T0 answers “1” on input (xi , yi );
but then xi ∈ L since the program y = yi decides L. We get a contradiction in both cases.
Third, we convince ourselves that if program y decides L on the machine T0 then (x, y) can
be “fast” only for finitely many words x. Let namely (x, y) be “fast”, where x = xi . Since y
was available at the selection of yi (it was not selected earlier) therefore we would have had to
choose some yi for this i and the actually selected yi could not be longer than y. Thus, if x
differs from all words xj with |yj | ≤ |y| then (x, y) is not “fast”.
Finally, consider an arbitrary Turing machine T deciding L. To this, we can make a one-
tape Turing machine T1 which also decides L and has timeT1 (n) ≤ (timeT (n))2 . Since the
machine T0 is universal, T0 simulates T1 by some program y in such a way that (let us be
generous) τ (x, y) ≤ (timeT (|x|))3 for all sufficiently long words x. According to what was
proved above, however, we have τ (x, y) ≥ h(|x| − |y|) for all but finitely many x, and thus
timeT (n) ≥ (h(n − |y|))1/3 .
Thus, for the above constructed Turing machine S deciding L in h(n − |y| − 1) steps, we
have

timeT (n) ≥ (h(n − |y|))1/3 ≥ g(h(n − |y| − 1)) ≥ g(timeS (n)).

The most important conclusion to be drawn from the speed-up theorem is that even though
it is convenient to talk about the computational complexity of a certain language L, rigorous

70
statements concerning complexity generally don’t refer to a single function t(n) as the complex-
ity, but only give upper bounds t (n) (by constructing a Turing machine deciding the language
in time t (n)) or lower bounds t (n) (showing that no Turing machine can make the decision in
time t (n) for all n).

4.1 Space versus time

Above, some general theorems were stated with respect to complexity measures. It was shown
that there are languages requiring a large amount of time to decide them. Analogous theorems
can be proved for the storage complexity. It is natural to ask about the relation of these
two complexity measures. There are some very simple relations mentioned in the text before
Theorem 4.2.
There is a variety of natural and interesting questions about the trade-off between storage
and time. Let us first mention the well-know practical problem that the work of most computers
can be speeded up significantly by adding memory. The relation here is not really between the
storage and time complexity of computations, only between slower and faster memory. Possibly,
between random-access memory versus the memory on disks, which is closer to the serial-access
model of Turing machines.
There are some examples of real storage-time trade-off in practice. Suppose that during
a computation, the values of a small but complex Boolean function will be used repeatedly.
Then, on a random-access machine, it is worth computing these values once for all inputs and
use table look-up later. Similarly, if a certain field of our records in a data base is often used for
lookup then it is worth computing a table facilitating this kind of search (inverting). All these
examples fall into the following category. We know some problem P and an algorithm A that
solves it. Another algorithm A is also known that solves P in less time and more storage than
A. But generally, we don’t have any proof that with the smaller amount of time really more
storage is needed to solve P . Moreover, when a lower bound is known on the time complexity of
some function, we have generally no better estimate of the storage complexity than the trivial
one mentioned above (and vice versa).

71
5 Non-deterministic algorithms
5.1 Non-deterministic Turing machines
A non-deterministic Turing machine differs from a deterministic one only in that in every po-
sition, the state of the control unit and the symbols scanned by the heads permit more than
one possible action. To each state g ∈ Γ and symbols h1 , . . . , hk a set of “legal actions” is
given where a legal action is a (2k + 1)-tuple consisting of a new state g  ∈ Γ, new symbols
h1 , . . . , hk and moves j1 , . . . , jk ∈ {−1, 0, 1}. More exactly, a non-deterministic Turing machine
is an ordered 4-tuple T = (k, Σ, Γ, Φ) where k ≥ 1 is a natural number, Σ and Γ are finite sets,
∗ ∈ Σ, START, STOP ∈ Γ (so far, everything is as with a deterministic Turing machine) and
Φ ⊂ (Γ × Σk ) × (Γ × Σk × {−1, 0, 1}k )
is an arbitrary relation. A legal computation of the machine is a sequence of steps where in each
step (just as with the deterministic Turing machine) the control unit enters a new state, the
heads write new letters on the tapes and move at most one step left or right. The steps must
satisfy the following conditions: if the state of the control unit was g ∈ Γ before the step and
the heads read on the tapes the symbols h1 , . . . , hk ∈ Σ then for the new state g  , the newly
written symbols h1 , . . . , hk and the steps ε1 , . . . , εk ∈ {−1, 0, 1} we have
(g, h1 , . . . , hk , g  , h1 , . . . , hk , ε1 , . . . , εk ) ∈ Φ.
A non-deterministic Turing machine can have therefore several legal computations for the same
input.
We say that the non-deterministic Turing machine T accepts word x ∈ Σ∗0 in time t if
whenever we write x on the first tape and the empty word on the other tapes, the machine has
a legal computation consisting of t steps, with this input, which at its halting has in position 0
of the first tape the symbol “1”. (There may be other legal computations that last much longer
or maybe don’t even stop, or reject the word.)
We say that a non-deterministic Turing machine T recognizes a language L if L consists
exactly of those words accepted by T (in arbitarily long finite time). If, in addition to this, the
machine accepts all words x ∈ L in time f (|x|) (where f : Z+ → Z+ ), then we say that the
machine recognizes L in time f (n) (recognizability in storage f (n) is defined similarly). The
class of languages recognizable by a non-deterministic Turing machine in time f (n) is denoted
by NTIME(f (n)).
Unlike deterministic classes, the non-deterministic recognizability of a language L does not
mean that the complementary language Σ∗0 \ L is recognizable (we will see below that each re-
cursively enumerable but not recursive language is an example for this). Therefore we introduce
the classes co−NTIME(f (n)): a language L belongs to a class co−NTIME(f (n)) if and only if
the complementary language Σ∗0 \ L belongs to NTIME(f (n)).
The notion of acceptance in storage s, and the classes NSPACE(f (n)), co−NSPACE(f (n))
are defined analogously.

(5.1) Remark
1. The deterministic Turing machines can be considered, of course, as special non-
deterministic Turing machines.

72
 2. A non-deterministic Turing machine can make several kinds of step in a situation; we did
not assume any probability distribution on these, we cannot therefore speak about the
probability of some computation. If we did this then we would speak of randomized, or
probabilistic, Turing machines, which are the object of a later section. In contrast to non-
deterministic Turing machines, randomized Turing machines model computing processes
that are practically important.

(5.2) Theorem Languages recognizable by non-deterministic Turing machines are exactly the
recursively enumerable languages.

Proof Assume first that language L is recursively enumerable. Then, there is a Turing machine
T that halts in finitely many steps on input x if and only if x ∈ L. Let us modify T in such a
way that when before stops it writes the symbol 1 onto field 0 of the first tape. Obviously, this
modified T has a legal computation accepting x if and only if x ∈ L.
Conversely, assume that L is recognizable by a non-deterministic Turing machine T ; we
show that L is recursively enumerable. We can assume that L is nonempty and let a ∈ L.
Let the set L# consist of all finite legal computations of the Turing machine T . Each element
of L# contains, in an appropriate encoding, of a sequence of configurations, or instantaneous
descriptions, as they follow in time. Each configuration shows the internal state and the symbol
found in each tape cell at the given instant, as well as the positions of the tape heads. The
set L# is obviously recursive since given two configurations, it can be decided whether the
second one can be obtained in one computation step of T from the first one. Let S be a Turing
machine that for an input y decides whether it is in L# and if yes then whether it describes a
legal computation accepting some word x. The range of values of the recursive function defined
by S is obviously just L.

73
5.2 Witnesses and the complexity of non-deterministic algorithms
Let us fix a finite alphabet Σ0 and consider a language L over it. Let us investigate first, what
it really means if L is recognizable within some time by a non-deterministic Turing machine.
We will show that this is connected with how easy it is to “prove” for a word that it is in L.
Let f and g be two functions with g(n) ≥ n. We say that the language L0 ∈ DTIME(g(n))
is a witness of length f (n) and time g(n) for language L if we have x ∈ L if and only if there is
a word y ∈ Σ∗0 with |y| ≤ f (|x|) and x&y ∈ L0 . (Here, & is a new symbol serving the separation
of the words x and y.)

(5.1) Theorem

(a) Every language L ∈ NTIME(f (n)) has a witness of length O(f (n)) and time O(n).

(b) If language L has a witness of length f (n) and time g(n) then L is in NTIME(g(f (n) +
n + 1)).

Proof
(a): Let T be the nondeterministic Turing machine recognizing the language L in time f (n)
with, say, two tapes. Following the pattern of the proof of Theorem 5.2, let us assign to each
word x in L the description of a legal computation of T accepting x in time f (|x|). It is not
difficult to make a Turing machine deciding about a string of length N in O(N ) steps whether
it is the description of a legal computation and if yes then whether this computation accepts
the word x. Thus, the witness is composed of the pairs x&y where y is a legal computation
accepting x.
(b) Let L0 be a witness of L with length f (n) and time g(n), and consider a deterministic
Turing machine S deciding L0 in time g(n). Let us construct a non-deterministic Turing machine
T doing the following. If x is written on its first tape then it first computes (deterministically)
the value of f (|x|) and writes this many 1’s on the second tape. Then it writes symbol & at the
end of x and makes a transition into its only state in which its behavior is nondeterministic.
While staying in this state it writes a word y of lengt at most f (|x|) after the word x&. This
happens as follows: while it reads a 1 on the second tape it has |Σ0 | + 1 legal moves: either it
writes some symbol of the alphabet on the first tape, moves right on the first tape and left on
the second tape or it writes nothing and makes a transition into state START2.
From state START2, on the first tape, the machine moves the head on the starting cell,
erases the second tape and then proceeds to work as the Turing machine S.
This machine T has an accepting legal computation if and only if there is a word y ∈ Σ∗0
of length at most f (|x|) for which S accepts word x&y, i.e. if x ∈ L. The running time of this
computation is obviously at most O(f (|x|)) + g(|x| + 1 + f (|x|)) = O(g(|x| + 1 + f (x))).

(5.2) Corollary For an arbitrary language L ⊂ Σ∗0 , the following properties are equivalent:

— L is recognizable on a non-deterministic Turing machine in polynomial time.

— L has a witness of polynomial length and time.

74
(5.3) Remark We mention it without proof (even without exact formulation) that these prop-
erties are also equivalent to the following: one can give a definition of L in the formal axiom
system of set theory such that, for a word x ∈ L, the statement “x ∈ L” can be proved from
the axioms of set theory in a number of steps polynomial in |x|. ♦

We denote the class of languages having the property stated in Corollary 5.2 by NP. The
languages L for which Σ∗0 \ L is in NP form the class co−NP. As we mentioned earlier, with
these classes of languages, what is easy is not the solution of the recognition problem of the
language, only the verification of the witnesses for the soluton. We will see later that these
classes are fundamental: they contain a large part of the algorithmic problems important from
the point of view of practice.
Many important languages are given by their witnesses—more precisely, by the language
L0 and function f (n) in our definition of witnesses (we will see many examples for this later).
In such cases we are asking whether a given word x is in L (i.e. , whether there is a y with
|y| ≤ f (n) and x&y ∈ L0 ). Without danger of confusion, the word y itself will also be called
the witness word , or simply witness, belonging to x in the witness language L. Very often,
we are not only interested whether a witness word exists but would also like to produce one.
This problem can be called the search problem belonging to the language L. There can be, of
course, several search problems belonging to a language. A search problem can make sense even
if the corresponding decision problem is trivial. For example, every natural number has a prime
decomposition but this is not easy to find.
Since every deterministic Turing machine can be considered non-deterministic it is obvious
that

DTIME(f (n)) ⊂ NTIME(f (n)).

In analogy with the fact that there is a recursively enumerable but not recursive language (i.e. ,
without limits on time or storage, the non-deterministic Turing machines are “stronger”), we
would expect that the above inclusion is strict. This is proved, however, only in very special
cases (e.g., in case of linear functions f , by Paul, Pippenger, Trotter and Szemeredi).
Later, we will treat the most important special case, the relation of the classes P and NP in
detail.

75
5.3 General results on nondeterministic complexity classes
The following simple relations connect the nondeterministic time- and space complexity classes:

(5.1) Theorem Let f be a well-computable function. Then

(a) NTIME(f (n)) ⊂ DSPACE(f (n))

 cf (n) ).
(b) NSPACE(f (n)) ⊂ c>0 DTIME(2

Proof
(a): The essence of the construction is that all legal computations of a nondeterministic
Turing machine can be tried out one after the other using only as much space as needed for one
such legal computation; above this, we need some extra space to keep track of where we are in
the trieout of of the cases.
More exactly, this can be described as follows: Let T be a non-deterministic Turing machine
recognizing language L in time f (n). As mentioned, we can assume that all legal computations
of T take at most f (n) steps where n is the length of the input. Let us modify the work of
T in such a way that (for some input x) it will choose first always the lexicographically first
action (we fix some ordering of Σ and Γ, this makes the actions lexicographically ordered). We
give the new (deterministic) machine called S an extra “bookkeeping” tape on which it writes
up which legal action it has chosen. If the present legal computation of T does not end with
the acceptance of x then machine S must not stop but must look up, on its bookkeeping tape,
the last action (say, this is the j-th one) which it can change to a lexicographically larger legal
one. Let it perform a legal computation of T in such a way that up to the j-th step it performs
the steps recorded on the bookkeeping tape, in the j-th step it performs the lexicographically
next legal action, and after it, the lexicographically first one (and, of course, it rewrites the
bookkeeping tape accordingly).
The modified, deterministic Turing machine S tries out all legal computations of the original
machine T and uses only as much storage as the original machine (which is at most f (n)), plus
the space used on the bookkeeping tape (which is again only O(f (n))).
(b): Let T = k, Σ, Γ, Φ
be a non-deterministic Turing machine recognizing L with storage
f (n). We can assume that T has only one tape. We want to try out all legal computations of T .
Some care is needed since a legal computation of T can last as long as 2f (n) steps, so there can
f (n)
even be 22 legal computations; we do not have time for checking this many computations.
To better organize the checking, we illustrate the situation by a graph as follows. Let us
fix the length n of the inputs. By configuration of the machine, we understand a triple (g, p, h)
where g ∈ Γ, −f (n) ≤ p ≤ f (n) and h ∈ Σ2f (n)+1 . The state g is the state of the control
unit at the given moment, the number p says where is the head and h specifies the symbols
on the tape (since we are interested in computations whose storage need is at most f (n) it
is sufficient to consider 2f (n) + 1 cells). It can be seen that number of configurations is at
most |Γ|(2f (n) + 1)m2f (n)+1 = 2O(f (n)) . Every configuration can be coded by a word of length
O(f (n)) over Σ.
Prepare a directed graph G whose vertices are the configurations; we draw an edge from
vertex u to vertex v if the machine has a legal action leading from configuration u to configuration

76
v. Add a vertex v0 and draw an edge to v0 from every configuration in which the machine is in
state STOP and has 1 on cell 0 of its tape. Denote ux the starting configuration corresponding
to input x. Word x is in L if and only if in this directed graph, a directed path leads from ux
to v0 .
On the RAM, we can construct the graph G in time 2O(f (n)) and (e.g. using breadth-first
search) we can decide in time O(|V (G)|) = 2O(f (n)) whether it contains a directed path from ux
to v0 . Since the RAM can be simulated by Turing machines in quadratic time, the time bound
remains 2O(f (n)) also on the Turing machine.
The following interesting theorem shows that the storage requirement is not essentially
decreased if we allow non-deterministic Turing machines.

(5.2) Theorem [Savitch’s Theorem] If f (n) is a well-computable function and f (n) ≥ log n
then

NSPACE(f (n)) ⊂ DSPACE(f (n)2 ).

Proof Let T = 1, Σ, Γ, Φ
be a non-deterministic Turing machine recognizing L with storage
f (n). Let us fix the length n of inputs. Consider the above graph G; we want to decide whether
it contains a directed path leading from ux to v0 . Now, of course, we do not want to construct
this whole graph since it is very big. We will therefore view it as given by a certain “oracle”.
Here, this means that about any two vertices, we can decide in a single step whether they
are connected by an edge. More exactly, this can be formulated as follows. Let us extend
the definition of Turing machines. An Turing machine with oracle (for G) is a special kind
of machine with three extra tapes reserved for the “oracle”. The machine has a special state
ORACLE. When it is in this state then in a single step, it writes onto the third oracle-tape a 1
or 0 depending on whether the words written onto the first and second oracle tapes are names
of graph vertices (configurations) connected by an edge, and enters the state START. In every
other state, it behaves like an ordinary Turing machine. When the machine enters the state
ORACLE we say it asks a question from the oracle. The question is, of course, given by the
pair of strings written onto the first two oracle tapes, and the answer comes on the third one.

(5.3) Lemma Suppose that a directed graph G is given on the set of of words of length t.
Then there is a Turing machine with an oracle for G which for given vertices u, v and natural
number q decides, using storage at most O(qt), whether there is a path of length at most 2q
from u to v.

Proof The Turing machine to be constructed will have two tapes besides the three oracle-
tapes. At start, the first tape contains the pair (u, q), the second one the pair (v, q). The work
of the machine proceeds in stages. At the beginning of some intermediate stage, both tapes will
contain a few pairs (x, r) where x is the name of a vertex and t ≤ q is a natural number.
Let (x, r) and (y, s) be the last pair on the two tapes. In the present stage, the machine asks
the question wether there is a path of length at most min{2r , 2s } from x to y. If min{r, s} = 0
then the answer can be read off immediately from an oracle-tape. If min{r, s} ≥ 1 then let
m = min{r, s} − 1. We write a pair (w, m) to the end of the first tape and determine recursively

77
whether there is a path of length at most 2m from w to y. If there is one then we write (w, m)
to the end of the second tape, erase it from the end of the first tape and determine whether
there is a path of length at most 2m from x to w. If there is one then we erase (w, m) from the
end of the second tape: we know that there is a path of length at most min{2r , 2s } from x to
y. If there is no path of length at most 2m either between x and w or between w and y then we
try the lexicographically next w. If we have tried all w’s then we know that there is no path of
length min{2r , 2s } between x and y.
It is easy to see that the second elements of the pairs are decreasing from left to right on
both tapes, so at most q pairs will ever get on each tape. One pair requires O(t + log q) symbols.
The storage thus used is only O(q log q + qt). This finishes the proof of the lemma.
Returning to the proof of the theorem, note that the question whether there is an edge
between two vertices of the graph G can be decided easily without the help of additional storage;
we might as well consider this decision as an oracle. The Lemma is therefore applicable with
values t, q = O(f (n)), and we obtain that it can be decided with at most tq + q log q = O(f (n)2 )
storage whether from a given vertex ux there is a directed path into v0 , i.e. whether the word x
is in L.
As we noted, the class PSPACE of languages decidable on a deterministic Turing machine in
polynomial storage is very important. It seems natural to introduce the class NPSPACE which
is the class of languages recognizable on a non-deterministic Turing machine with polynomial
storage. But the following corrollary of Savitch’s theorem shows that this would not lead to any
new notion:

(5.4) Corollary PSPACE = NPSPACE.

78
5.4 Examples of languages in NP
5.4.1 Graphs
In this section, by a graph we mean a simple graph: an undirected graph without multiple edges
or loops. n is always the number of nodes. Unless we say otherwise, we assume that the graph
2
is described by its adjacency matrix, which we consider as a string in {0, 1}n . In this way, a
graph property can be considered a language over {0, 1}. We can thus ask whether a certain
graph property is in NP. (Notice that describing a graph in one of the other usual ways, e.g. by
giving a list of neighbors for each node, would not affect the membership of graph properties in
NP. It is namely easy to compute these representations from each other in polynomial time.)
The following graph properties are in NP.
n
(5.1) Graph-connectivity Certificate: a set of 2 paths, one for each pair of nodes. ♦

(5.2) Graph non-connectivity Certificate: a proper subset of the set of nodes that is not
connected by any edge to the rest of the nodes. ♦

(5.3) Graph planarity The natural witness is a concrete diagram, though some analysis is
needed to see that in case such a diagram exists then one exists in which the coordinates of
every node are integers whose number of bits is polynomial in n.
It is interesting to remark the fact known in graph theory that this can be realized using
single straight-line segments for the edges and thus, it is enough to specify the coordinates of the
nodes. We must be careful, however, since the coordinates of the nodes used in the drawing may
have too many bits, violating the requirement on the length of the witness. (It can be proved
that every planar graph can be drawn in the plane in such a way that each edge is a straigh-line
segment and the coordinates of every node are integers whose number of bits is polynomial in
n.)
It is possible, however, to give a purely combinatorial way of drawing the graph. Let G
be a graph with n nodes and m edges which we assume for simplicity to be connected. After
drawing it in the plane, the edges partition the plane into domains which we call “countries”
(the unbounded domain is also a country). We need the following fact, called Euler’s formula:

(5.4) Theorem A connected planar graph with n nodes and m edges has n + m − 2 countries.

Thus to specify the drawing we give a set of m − n + 2 country names and for every country,
we specify the sequence of edges forming its boundary. In this case, it is enough to check whether
every edge is in exactly two boundaries.
The fact that the existence of such a set of edge sequences is a necessary condition of
planarity follows from Euler’s formula. The sufficiency of this condition requires somewhat
harder tools from topology; we will not go into these details. (Specifying a set of edge sequences
as country boundaries amounts to defining a two-dimensional surface with the graph drawn
onto it. A theorem of topology says that if a connected graph drawn on that surface satisfies
Euler’s formula then the surface is topologically equivalent (homeomorphic) to the plane.) ♦

(5.5) Non-planarity Let us review the following facts.

79
 1. Let K5 be the graph obtained by connecting five nodes in every possible way. This graph
is also called a “complete pentagon”. Let K33 be the 6-node bipartite graph containing
two sets A, B of three nodes each, with every possible edge between A and B. This graph
is also called “three houses, three wells” after a certain puzzle with a similar name. It is
easy to see that K5 and K33 are nonplanar.

2. Given a graph G, we say that a graph G is a subdivision of G if it is obtained from G

by replacing each edge of G with arbitrarily long non-intersecting paths. It is easy to see
that if G is nonplanar then any subdivision is nonplanar.

3. If a graph is nonplanar then, obviously, every graph containing it is also nonplanar.

The following fundamental theorem of graph theory says that the nonplanar graphs are just
the ones obtained by the above operations:

(5.6) Kuratowski’s Theorem A graph is nonplanar if and only if it contains a subgraph

that is a topological version of either K5 or K33 .

If the graph is nonplanar then the subgraph whose existence is stated by Kuratowski’s
Theorem can serve as a certificate for this. ♦

(5.7) Existence of perfect matching A certificate is the perfect matching itself. ♦

(5.8) Non-existence of a perfect matching Witnesses for the non-existence in case of

bipartite graphs are based on a fundamental theorem. Let G be a bipartite graph G consisting
with bipartition classes U and W . Recall the following theorem:

(5.9) Theorem Frobenius’s Theorem A bipartite graph G has a perfect matching if and only
if |U | = |W | and for any k, any k nodes in W have at least k neighbors in U .

Hence, if in some bipartite graph there is no perfect matching then this can be certified
either by noting that U and W have different cardinality, or by a subset of W violating the
conditions of the theorem.
Now let G be an arbitrary graph. If there is a perfect matching then it is easy to see that
for any k, if we delete any k nodes, there remain at most k connected components of odd size.
The following fundamental (and deeper) theorem says that this condition is not only necessary
for the existence of a perfect matching but also sufficient.

(5.10) Theorem Tutte’s Theorem A graph G has a perfect matching if and only if for any k,
if we delete any k nodes, there remain at most k connected components of odd size.

This way, if there is no perfect matching in the graph then this can be certified by a set of
nodes whose deletion creates too many odd components. ♦

A Hamiltonian cycle of a graph is a cycle going through each node exactly once.

(5.11) Existence of a Hamiltonian cycle

Certificate: the Hamitonian cycle itself. ♦

80
 A coloring of a graph is an assignment of some symbol called “color” to each node in such
a way that neighboring nodes get different colors.

(5.12) Colorability with three colors If a graph can be colored with three colors the
coloring itself is a certificate. Of course, this is valid for any number of colors. ♦

Most of the properties listed above, up to (and including) the non-existence of a perfect
matching, can be solved in polynomial time (i.e., they are in P). For the Hamiltonian cycle
problem and the three-colorability problem, no polynomial algorithm is known (we return to
this later).

5.4.2 Arithmetic and algebra

To show that many fundamental problems in arithmetic and algebra also belong to the class
NP, we recall that every natural number can be considered a word in {0, 1}∗ (representing the
number in binary). We start with the problem of deciding whether a natural number is a prime.

(5.13) Compositeness of an integer Certificate of compositeness: a proper divisor. ♦

(5.14) Primality It is significantly more difficult to find witnesses for primality. We use the
following fundamental theorem of number theory:

(5.15) Theorem An integer n ≥ 2 is prime if and only if there is a natural number a such
that an−1 ≡ 1 (mod n) but am ≡ 1 (mod n) for any m such that 1 ≤ m < n − 1.

(This theorem says that there is a so-called “primitive root” a for n, whose powers run
through all non-0 residues mod n.)
With this theorem in mind,we would like to use the number a to be the witness for the
primality of n. Since, obviously, only the remainder of the number a after division by n is
significant here, there will also be a witness a with 1 ≤ a < n. In this way, the restriction on the
length of the witness is satisfied: a does not have more bits than n itself. Let k be the number
of bits of n.
As we have seen in chapter 3, we can also check the condition

an−1 ≡ 1 (mod n) (5.16)

in polynomial time. It is, however, a much harder question how to verify the further conditions:

am ≡ 1 (mod n) (1 ≤ m < n − 1). (5.17)

We have seen that we can do this for each specific m, but it seems that we must do this n − 2
times, i.e. exponentially many times in terms of k. We can use, however, the (easy) number-
theoretical fact that if (5.16) holds then the smallest m = m0 violating (5.17) (if there is any)
is a divisor of n − 1. It is also easy to see that then (5.17) is violated by every multiple of m0
smaller than n − 1. Thus, if the prime factor decomposition of n − 1 is n − 1 = pr11 · · · prt t then

81
(5.17) is violated by some m = (n − 1)/pi . It is enough therefore to verify that for all i with
1 ≤ i ≤ t,

a(n−1)/pi ≡ 1 (mod n).

Now, it is obvious that t ≤ k and therefore we have to check (5.17) for at most k values which
can be done in the way described before, in polynomial total time.
There is, however, another difficulty: how are we to compute the prime decomposition of
n − 1? This, in itself, is a harder problem than to decide whether n is a prime. We can,
however, add the prime decomposition of n − 1 to the “witness”; this consists therefore, besides
the number a, of the numbers p1 , r1 , . . . , pt , rt (it is easy to see that this is at most 3k bits).
Now only the problem remains to check whether this is a prime decomposition indeed, i.e. that
n−1 = pr11 · · · prt t (this is easy) and that p1 , . . . , pt are indeed primes. We can do this recursively.
We still have to check that this recursion gives witnesses of polynomial length and it can be
decided in polynomial time that these are witnesses. Let L(k) denote the maximum length of
the witnesses in case of numbers n of k bits. Then, according to the above recursion,

t
L(k) ≤ 3k + L(ki )
i=1

where ki is the number of bits of the prime pi . Since p1 · · · pt ≤ n − 1 < n it follows easily that

k1 + · · · + kt < k + t.

Also obviously ki < k. Using this, it follows from the above recursion that L(k) ≤ 3k 2 . In fact,
the inequality L(k) ≤ 3(k − 1)2 is easier to prove. This is namely obvious for k = 1, 2 and if we
know that it holds for all numbers less than k then

t 
t
L(k) ≤ 3k + L(ki ) ≤ 3k + 3(ki − 1)2
i=1 i=1

t
≤ 3k + 3(k − 2) ki − 1 ≤ 3k + 3(k − 2) · (k − 1) ≤ 3(k − 1)2 .
i=1

We can prove similarly that it is decidable about a string in polynomial time whether it is a
certificate defined in the above way. ♦

Usually we are not satisfied with knowing whether a given number n is a prime or not, but
if it is not a prime then we might also want to find one of its proper divisors. (If we can solve
this problem then repeating it, we can find the complete prime decomposition.) This is not a
decision problem, but it is not difficult to reformulate it as a decision problem:

(5.18) Existence of a bounded divisor Given two natural numbers n and k; does n have
a proper divisor not greater than k?
It is clear that this problem is in NP: the certificate is the divisor. ♦

The complementary language is also in NP:

82
(5.19) Nonexistence of a bounded divisor This is the set of all pairs (n, k) such that
every proper divisor of n is greater than k. A certificate for this is the prime decomposition of
n, together with a certificateof the primality of every prime factor.
It is not known whether the problem of compositeness (even less, the existence of a bounded
divisor) is in P. Extending the notion of algorithms and using random numbers, it is decidable
in polynomial time about a number whether it is a prime (see the section on randomized
algorithms). At the same time, the corresponding search problem (the search for a proper
divisor), or, equivalently, deciding the existence of bounded divisors, is significantly harder; for
this, a polynomial algorithm was not yet found even when the use of random numbers is allowed.
♦

Now we turn to some basic questions in algebra. A notion analogous for primality of a
positive integer is irreducibility of a polynomial (for simplicity, with a single variable, and with
rational coefficients). A polynomial is reducible if it can be written as the product of two
non-constant polynomials with rational coefficients.

(5.20) Reducibility of a polynomial over the rational field

Certificate: a proper divisor; but some remarks are in order.
Let f be the polynomial. To prove that this problem is in NP we must convince ourselves
that the number of bits necessary for writing down a proper divisor can be bounded by a
polynomial of the number of bits in the representation of f . (We omit the proof of this here.)
It can also be shown that this language is in P. ♦

5.4.3 Systems of linear inequalities

A system Ax ≤ b of linear inequalities (where A is an integer matrix with m rows and n columns
and b is a column vector of m elements) can be considered a word over the alphabet consisting
of the symbols “0”, “1”, “,” and “;” when e.g. we represent its elements in the binary number
system, write the matrix row after row, placing a comma after each number and a semicolon
after each row. The following properties of systems of linear inequalities are in NP:

(5.21) Existence of solution The solution offers itself as an obvious witness of solvability
but we must be careful: we must be convinced that if a system of linear equations with integer
coefficients has a solution then it has a solution among rational numbers, moreover, even a
solution in which the numerators and denominators have only a polynomial number of bits.
These facts follow from the elements of the theory of linear programming. ♦

(5.22) Nonexistence of solution Witnesses for the non-existence of solution can be found
using the following fundamental theorem known from linear programming:

(5.23) Farkas Lemma The system Ax ≤ b of inequalities is unsolvable if and only if the
following system of inequalities is solvable: y T A = 0, y T b = −1, y ≥ 0.

In words, this lemma says that a system of linear inequalities is unsolvable if and only if
a contradiction can be obtained by a linear combination of the inequalities with nonnegative
coefficients.

83
 Using this, a solution of the other system of inequalities given in the lemma (the nonnegative
coefficients) is a witness of the nonexistence of a solution. ♦

(5.24) Existence of an integer solution The solution itself is a witness but we need
some reasoning again to limit the size of witnesses, which is more difficult here. ♦

It is not known wheter the non-existence of an integer solution is in NP, i.e., if this fact can
be certified by a polynomial length and polynomial time verifiable certificate.
It is important to note that the fundamental problem of linear programming, i.e. looking
for the optimum of a linear object function under linear conditions, can be easily reduced to
the problem of solvability of systems of linear inequalities. Similarly, the search for optimal
solutions can be reduced to the decision of the existence of integer solutions.
For a long time, it was unknown whether the problem of solvability of systems of linear
inequalities is in P (the well-known simplex method is not polynomial). The first polynomial
algorithm for this problem was the ellipsoid method of L. G. Khachian (relying on work by
Yudin and Nemirovskii).
The running time of this method led, however, to a very high-degree polynomial; it could
not therefore compete in practice with the simplex method which, though is exponential in the
worst case, is on average (as shown by experience) much faster than the ellipsoid method.
Several polynomial-time linear programming algorithms have been found since; among these,
Karmarkar’s method can compete with the simplex method even in practice.
No polynomial algorithm is known for solving systems of linear inequalities in integers; one
cannot even hope to find such an algorithm (see the notion of NP-completeness below).
Reviewing the above list of examples, the following observatons can be made.

— For many properties that are in NP, their negation (i.e. the complement of the corre-
sponding language) is also in NP. This fact is, however, generally non-trivial; in various
branches of mathematics, often the most fundamental theorems assert this for certain
languages.

— It is often the case that if some property (language) turns out to be in NP ∩ co−NP
then sooner or later it also turns out to be in P. This happened, for example, with the
existence of perfect matchings, planarity, the solution of systems of linear inequalities.
Research is very intensive on prime testing. If NP is considered an analog of “recursively
enumerable” and P an analog of “recursive” then we can expect that this is always the
case. However, there is no proof for this; moreover, this cannot really be expected to be
true in full generality.

— With other NP problems, their solution in polynomial time seems hopeless, they are very
hard to handle (Hamiltonian cycle, graph coloring, and integer solution of a system of
linear inequalities). We cannot prove that these are not in P (we don’t know whether
P = NP holds); but still, one can prove a certain exact property of these problems that
shows that they are hard. We will turn to this later.

— There are many problems in NP with a naturally corresponding search problem and with
the property that if we can solve the decision problem then we can also solve (in a natural

84
 manner) the search problem. E.g., if we can decide whether there is a perfect matching in
a certain graph then we can search for perfect matching in polynomial time in the following
way: we delete edges from the graph as long as a perfect matching still remains in it. When
we get stuck, the remaining graph must be a perfect matching. Using similar simple tricks,
the search problem corresponding to the existence of Hamiltonian cycles, colorability with
3 colors, etc. can be reduced to the decision problem. This is, however, not always so.
E.g., our ability to decide in polynomial time (at least, in some sense) whether a number
is a prime was not applicable to the problem of finding a proper divisor.

— A number of NP-problems has a related optimization problem which is easier to state,

even if it is not an NP-problem by its form. For example, instead of the general matching
problem, it is easier to say that the problem is to find out the maximum size of a matching
in the graph. In case of the coloring problem, we may want to look for the chromatic
number, the smallest number of colors with which the graph is colorable. The solvability
of a set of linear inequalities is intimately connected with the problem of finding a solution
that maximizes a certain linear form: this is the problem of linear programming. Several
other examples come later. If there is a polynomial algorithm solving the optimization
problem then it automatically solves the associated NP problem. If there is a polynomial
algorithm solving the NP-problem then, using binary search, it provides a polynomial
algorithm to solve the associated optimization problem.

There are, of course, interesting problems (languages) also in other non-deterministic com-
plexity classes. The non-deterministic exponential time (NEXPTIME) class can be defined as
c
the union of the classes NTIME(2n ) for all c > 0. We can formulate an example in connection
with Ramsey’s Theorem. Let G be a graph; the Ramsey number R(G) belonging to G is the
smallest N > 0 for which it is the case that no matter how we color the edges of the N -node
complete graph with two colors, some color contains a copy of G. (Ramsey’s Theorem says that
such a finite number exists, which is non-trivial.) Let L consist of the pairs (G, N ) for which
R(G) > N . The size of the input (G, N ) (if G is described, say, by its adjacency matrix) is
O(|V (G)|2 + log N ).
Now, L is in NEXPTIME since the fact (G, N ) ∈ L is witnessed by a coloring of the complete
graph on N nodes in which no homogenously colored copy of G; this property can be checked in
time O(N |V (G)| ) which is exponential in the size of the input (but not worse). On the other hand,
deterministically, we know no better algorithm to decide (G, N ) ∈ L than a double exponential
one. The trivial algoritm, which is, unfortunately, the best known, goes over all colorings of the
edges of the N -node complete graph, and the number of these is 2N (N −1)/2 .

5.5 NP-completeness
We say that a language L1 ⊂ Σ∗1 is polynomially reducible to a language L2 ⊂ Σ∗2 if there is
a function f : Σ∗1 → Σ∗2 computable in polynomial time such that for all words x ∈ Σ∗1 we have

x ∈ L1 ⇔ x ∈ L2 .

It is easy to verify from the definition that this relation is transitive:

85
(5.1) Proposition If L1 is polynomially reducible to L2 and L2 is polynomially reducible to
L3 then L1 is polynomially reducible to L3 .

The membership of a language in P can also be expressed by saying that it is polynomially

reducible to the language {0, 1}.

(5.2) Proposition If a language is in P then every language is in P that is polynomially

reducible to it. If a language is in NP then every language is in NP that it polynomially
reducible to it.

We call a language NP-complete if it belongs to NP and every language in NP is polyno-

mially reducible to it. These are thus the “hardest” languages in NP. The word “completeness”
suggests that the solution of the decision problem of a complete language contains, in some
sense, the solution to the decision problem of all other NP languages. If we could show about
even a single NP-complete language that it is in P then P = NP would follow. The following
observation is also obvious.

(5.3) Proposition If an NP-complete language L1 is polynomially reducible to a language L2

in NP then L2 is also NP-complete.

It is not obvious at all that NP-complete languages exist. Our first goal is to give an NP-
complete language; later (by polynomial reduction, using 5.3) we will prove the NP-completeness
of many other problems.
A Boolean polynomial is called satisfiable if the Boolean function defined by it is not
identically 0.

(5.4) Satisfiability Problem For a given Boolean polynomial f , decide whether it is sat-
isfiable. We consider the problem, in general, in the case when the Boolean polynomial is a
conjunctive normal form. ♦

We can consider each conjunctive normal form as a word over the alphabet consisting of the
symbols “x”, “0”, “1”, “+”, “¬”, “∧” and “∨” (we write the indices of the variables in binary
number system, e.g. x6 = x110). Let SAT denote the language formed from the satisfiable
conjunctive normal forms.
The following theorem is one of the central results in complexity theory.

(5.5) Theorem Cook–Levin Theorem. The language SAT is NP-complete.

Proof Let L be an arbitrary language in NP. Then there is a non-deterministic Turing machine
T = k, Σ, Γ, Φ
and there are integers c, c1 > 0 such that T recognizes L in time c1 · nc . We
can assume k = 1. Let us consider an arbitrary word h1 · · · hn ∈ Σ∗ . Let N = c1 · nc . Let us
introduce the following variables:
x[n, g] (0 ≤ n ≤ N, g ∈ Γ),
y[n, p] (0 ≤ n ≤ N, −N ≤ p ≤ N ),
z[n, p, h] (0 ≤ n ≤ N, −N ≤ p ≤ N, h ∈ Σ).

86
 If a legal computation of the machine T is given then let us assign to these variables the
following values: x[n, g] is true if after the n-th step, the control unit is in state g; y[n, p] is true
if after the n-th step, the head is on the p-th tape cell; z[n, p, h] is true if after the n-the step,
the p-th tape cell contains symbol h. The variables x, y, z obviously determine the computation
of the Turing machine.
However, not each possible system of values assigned to the variables will correspond to
a computation of the Turing machine. One can easily write up logical relations among the
variables that, when taken together, express the fact that this is a legal computation accepting
h1 · · · hn . We must require that the control unit be in some state in each step:

x[n, g] (0 ≤ n ≤ N );
g∈Γ

and it should not be in two states:

¬x[n, g] ∨ ¬x[n, g  ] (g, g  ∈ Γ, 0 ≤ n ≤ N ).
We can require, similarly, that the head should be only in one position in each step and there
should be one and only one symbol in each tape cell. We write that initially, the machine is in
state START and at the end of the computation, in state STOP, and the head starts from cell
0:
x[0, START] = 1, x[N, STOP] = 1, y[0, 0] = 1;
and, similarly, that the tape contains initially the input h1 · · · hn and finally the symbol 1 on
cell 0:
z[0, i − 1, hi ] = 1 (1 ≤ i ≤ n)
z[0, i − 1, ∗] = 1 (i < 0 or i > n)
z[N, 0, 1] = 1.

We must further express the computation rules of the machine, i.e., that for all g, g  ∈ Γ,
h, h ∈ Σ, ε ∈ {−1, 0, 1} and −N ≤ p ≤ N we have
(x[n, g] ∧ y[n, p] ∧ z[n, p, h]) ⇒ ¬(x[n + 1, g  ] ∧ y[n + 1, p + ε] ∧ z[n + 1, p, h ])
and that where there is no head the tape content does not change:
¬y[n, p] ⇒ (z[n, p, h] ⇔ z[n + 1, p, h]).
For the sake of clarity, the the last two formulas are not in conjunctive normal form but it is
easy to bring them to such form. Joining all these relations by the sign “∧” we get a conjunctive
normal form that is satisfiable if and only if the Turing machine T has a computation of at most
N steps accepting h1 · · · hn . It easy to verify that for given h1 , . . . , hn , the described construction
of a formula can be carried out in polynomial time.
It will be useful to prove the NP-completeness of some special cases of the satisfiability
problem. A conjunctive normal form is called a k-form if in each of its components, at most k
literals occur. Let k-SAT denote the language made up by the satisfiable k-forms. Let further
SAT-k denote the language consisting of those satisfiable conjunctive normal forms in which
each variable occurs in at most k elementary disjunctions.

87
(5.6) Theorem The language k-SAT is NP-complete.

Proof Let B be a Boolean circuit with inputs x1 , . . . , xn (a conjunctive normal form is a

special case of this). We will find a 3-normal form that is satisfiable if and only if the function
computed by B is not identically 0. Let us introduce a new variable yi for each node i of the
circuit. The meaning of these variables is that in a satisfying assignment, these are the values
computed by the corresponding nodes. Let us write up all the restrictions for yi . For each input
node i, with node variable yi and input variable xi we write
yi ⇔ xi (1 ≤ i ≤ n).
If yi is the variable for an ∧ node with inputs yj and yk then we write
y i ≡ yj ∧ yk .
If yi is the variable for a ∨ node with inputs yj and yk then we write
yi ≡ yj ∨ yk .
If yi is the variable for a ¬ node with input yj then we write
yi ≡ ¬yj .
Finally, if yi is the output node then we add the clause
yi .
Each of these statements involves only three variables and is therefore expressible as a 3-normal
form. The conjunction of all these is satisfiable if and only if B is satisfiable.
it is natural to wonder at this point why have we considered just the 3-satisfiability problem.
The problems 4-SAT, 5-SAT, etc. are harder than 3-SAT therefore these are, of course, also NP-
complete. The theorem below shows, on the other hand, that the problem 2-SAT is already
not NP-complete (at least if P = NP). (This illustrates the fact that often a little modification
of the conditions of a problem leads from a polynomially solvable problem to an NP-complete
one.)
(5.7) Theorem The language 2-SAT is in P.

Proof Let B be a 2-normal form on the variables x1 , . . . , xn . Let us use the convention that
the variables xi are also written as x1i and the negated variables xi are also written as new
symbols x0i . Let us construct a directed graph G on the set V (G) = {x1 , . . . , xn , x1 , . . . , xn } in
the following way: we connect node xεi to node xδj if x1−εi ∨ xδj is an elementary disjunction in
B. (This disjunction is equivalent to xεi ⇒ xδj .) Let us notice that then in this graph, there is
also an edge from x1−δ
j to x1−ε
i . In this directed graph, let us consider the strongly connected
components; these are the classes of nodes obtained when we group two nodes in one class
whenever there is a directed path between them.
(5.8) Lemma The formula B is satisfiable if and only if none of the strongly connected com-
ponents of G contains both a variable and its negation.
The theorem follows from this lemma since it is easy to find in polynomial time the strongly
connected components of a directed graph.

88
Proof of Lemma 5.8 Let us note first that if an assignment of values satisfies formula B
and xεi is “true” in this assignment then every xδj is “true” to which an edge leads from xεi :
otherwise, the elementary disjunction x1−ε i ∨ xδj would not be satisfied. It follows from this that
the nodes of a strongly connected component are either all “true” or none of them. But then,
a variable and its negation cannot simultaneously be present in a component.
Conversely, let us assume that no strongly connected component contains both a variable
and its negation. Consider a variable xi . According to the condition, there cannot be directed
paths in both directions between x0i and x1i . Let us assume there is no such directed path
in either direction. Let us then draw a new edge from x1i to x0i . This will not violate our
assumption that no connected component contains both a node and its negation. If namely
such a connected components should arise then it would contain the new edge, but then both
x1i and x0i would belong to this component and therefore there would be a path from x0i to x1i .
But then this path would also be in the original graph, which is impossible.
Repeating this procedure, we can draw in new edges (moreover, always from a variable to
its negation) in such a way that in the obtained graph, between each variable and its negation,
there will be a directed path in exactly one direction. Let now be xi = 1 if a directed path
leads from x0i to x1i and 0 if not. We claim that this assignment satisfies all disjunctions. Let
us namely consider an elementary disjunction, say, xi ∨ xj . If both of its members were false
then—according to the definition—there were a directed path from x1i to x0i and from x1j to x0j .
Further, according to the definition of the graph, there is an edge from x0i to x1j and from x0j to
x1i . But then, x0i and x1i are in a strongly connected components, which is a contradiction.

(5.9) Theorem The language SAT-3 is NP-complete.

Proof Let B be a Boolean formula of the variables x1 , . . . , xn . For each variable xj , replace
the i-th occurrence of xj in B, with new variable yji : let the new formula be B  . For each j,
assuming there are m occurrences of xj in B, form the conjunction

Cj = (yj1 ⇒ yj2 ) ∧ (yj2 ⇒ yj3 ) ∧ · · · ∧ (yjm ⇒ yj1 ).

(Of course, yj1 ⇒ yj2 = ¬yj1 ∨ yj2 .) The formula B  ∧ C1 ∧ · · · ∧ Cn contains at most 3 occurrences
of each variable, is a conjunctive normal form if B is, and is satisfiable obviously if and only if
B is.

5.6 Further NP-complete problems

5.6.1 NP-complete problems for graphs
One might think that NP-complete problems are of logical character. In what follows, we will
show the NP-completeness of a number of important “everyday” combinatorial, algebraic, etc.
problems. When we show about a problem that it is NP-complete, then it follows that it is not
in P unless P = NP. Therefore we can consider the NP-completeness of a language as a proof
of its undecidability in polynomial time.
Let us formulate a fundamental combinatorial problem:

89
(5.1) Blocking Set Problem: Given a system {A1 , . . . , Am } of finite sets and a natural
number k, is there a set with at most k elements intersecting every Ai ? ♦

We have met a special case of this problem, the Blocking Set Problem for the edges of a
bipartite graph. This special case was polynomial time solvable. In contrast to this, we prove:

(5.2) Theorem The Blocking Set Problem is NP-complete.

Proof We reduce 3-SAT to this problem. For a given conjunctive 3-normal form B we con-
struct a system of sets as follows: let the underlying set be the set {x1 , . . . , xn , x1 , . . . , xn } of the
variable symbols occurring in B and their negations. For each clause of B, let us take the set
of literals occurring in it; let us further take the sets {xi , xi }. The elements of this set system
can be blocked with at most n nodes if and only if the normal form is satisfiable.
The Blocking Set Problem remains NP-complete even if we impose various restrictions on
the set system. It can be seen from the above construction that the Blocking Set Problem is
NP-complete even for a system of sets with at most three elements. (We will see a little later
that this holds even if the system contains only two-element sets, i.e., the edges of a graph.) If
we reduce the language SAT first to the language SAT-3 according to Theorem 5.9 and apply to
this the above construction then we obtain a set system for which each element of the underlying
set is in at most 4 sets.
With a little care, we can show that the Blocking Set Problem remains NP-complete even
for set-systems in which each element is contained in at most 3 sets. Indeed, it is easy to reduce
the Satisfiablity Problem to the case when the input is a conjunctive normal form in which every
variable occurs at least once negated and at least one unnegated; then the construction above
gives such a set-system.
We cannot go further than this: if each element is in at most 2 sets then the Blocking Set
Problem is solvable in polynomial time. In fact, it is easy to reduce this special case of the
blocking set problem to the matching problem.
It is easy to see that the following problem is equivalent to the Blocking Set Problem (only
the roles of “elements” and “subsets” must be interchanged):

(5.3) Covering problem: Given a system {A1 , . . . , Am } of subsets of a finite set S and a
natural number k. Can k sets be selected in such a way that their union is the whole set S? ♦

According to the discussion above, this problem is NP-complete already even when each of
the given subsets has at most 3 elements. has only 2 elements, the problem becomes polynomially
solvable, as the following exercise shows:

5.1 Exercise Prove that the covering problem, if every set in the set system is restricted to
have at most 2 elements, is reducible to the following matching problem: given a graph G and
a natural number k, is there a matching of size k in G? ♦

For set systems, the following pair of problems is also important:

90
(5.4) k-partition problem: Given a system {A1 , . . . , Am } of subsets of a finite set V and a
natural number k. Can a subsystem of k sets {Ai1 , . . . , Aik } be selected that gives a partition
of the underlying set (i.e. consists of disjoint sets whose union is the whole set V )? ♦

(5.5) Partition problem: Given a system {A1 , . . . , Am } of subsets of a finite set S. Can a
subsystem (of any size) be selected that gives a partition of the underlying set? ♦

If all the Ai are of the same size, then of course the number of sets in a partition is uniquely
determined, and so the two problems are equivalent.

(5.6) Theorem The k-partition problem and the partition problem are NP-complete.

Proof We reduce the the Covering Problem with sets having at most 3 elements to the k-
partition problem. Thus we are given a system of sets with at most 3 elements and a natural
number k. We want to decide whether k of these given sets can be selected in such a way that
their union is the whole S. Let us expand the system by adding all subsets of the given sets
(it is here that we exploit the fact that the given sets are bounded: from this, the number of
sets grows as most 23 = 9-fold). Obviously, if k sets of the original system cover S then k
appropriate sets of the expanded system provide a partition of S, and vice versa. In this way,
we have found that the k-partition problem is NP-complete.
Second, we reduce the k-partition problem to the partition problem. Let U be a k-element
set disjoint from S. Let our new underlying set be S ∪ U , and let the sets of our new set system
be the sets of form Ai ∪ {u} where u ∈ U . Obviously, if from this new set system, some sets can
be selected that form a partition of the underlying set then the number of these is k and the
parts falling in S give a partition of S into k sets. Conversely, every partition of S into k sets
Ai provides a partition of the set S ∪ U into sets from the new set system. Thus, the partition
problem is NP-complete.
If the given sets have two elements then the Set Partition Problem is just the perfect matching
problem and can therefore be solved in polynomial time. The argument above shows that, on
the other hand, the Set Partition Problem for sets with at most 3 elements is NP-complete.
Next we treat a fundamental graph-theoretic problem, the coloring problem. We have seen
that the problem of coloring with two colors is solvable in polynomial time. On the other hand:

(5.7) Theorem The coloring of graphs with three colors is an NP-complete problem.

Proof Let us be given a 3-form B; we construct a graph G for it that is colorable with three
colors if and only if B is satisfiable.
For the nodes of the graph G, we first take the literals, and we connect each variable with
its negation. We take two more nodes, u and v, and connect them with each other, further
we connect u with all unnegated and negated variables. Finally, we take a pentagon for each
elementary disjunction zi1 ∨ zi2 ∨ zi3 ; we connect two neighboring vertices of the pentagon with
v, and its three other vertices with zi1 , zi2 and zi3 . We claim that the graph G thus constructed
is colorable with three colors if and only if B is satisfiable (Figure 5.1).
The following remark, which is very easily verified, plays a key role in the proof: if for some
clause zi1 ∨ zi2 ∨ zi3 , the nodes zi1 , zi2 , zi3 and v are colored with three colors then this coloring

91
 x1 x x3 x4
2
u v

x1 x x3 x4
2

Figure 5.1: The graph whose 3-coloring is equivalent to satisfying the expression
(x1 ∨ x2 ∨ x4 ) ∧ (x1 ∨ x2 ∨ x3 )

can be extended to the pentagon as a legal coloring if and only if the colors of zi1 , zi2 , zi3 and v
are not identical.
Let us first assume that B is satisfiable, and let us consider the corresponding value assign-
ment. Let us color red those (negated or unnegated) variables that are “true”, and blue the
others. Let us color u yellow and v blue. Since every elementary disjunction must contain a red
node, this coloring can be legally extended to the nodes of the pentagons.
Conversely, let us assume that the graph G is colorable with three colors and let us consider
a “legal” coloring with red, yellow and blue. We can assume that the node v is blue and the
node u is yellow. Then the nodes corresponding to the variables can only be blue and red, and
between each variable and its negation, one is red and the other one is blue. Then the fact that
the pentagons are also colored implies that each elementary disjunction contains a red node.
But this also means that taking the red nodes as “true”, we get a value assignment satisfying
B.
It follows easily from the previous theorem that for every number k ≥ 3 the k-colorability
of graphs is NP-complete.
The following is another very basic graph theory problem. A set S of nodes of a graph is
independent, if no edge connects two of them.

(5.8) Independent node set problem: Given a graph G and a natural number k, are there
k independent nodes in G? ♦

(5.9) Theorem The independent node set problem is NP-complete.

Proof We reduce to this problem the problem of coloring with 3 colors. Let G be an arbitrary
graph with n nodes and let us construct the graph H as follows: Take three disjoint copies G1 ,
G2 , G3 of G and connect the corresponding nodes of the three copies. Let H be the graph
obtained, this has thus 3n nodes.

92
 We claim that there are n independent nodes in H if and only if G is colorable with three
colors. Indeed, if G is colorable with three colors, say, with red, blue and yellow, then the nodes
in G1 corresponding to the red nodes, the nodes in G2 corresponding to the blue nodes and the
nodes in G3 corresponding to the yellow nodes are independent even if taken together in H, and
their number is n. The converse can be proved similarly.
In the set system constructed in the proof of Theorem 5.2 there were sets of at most three
elements, for the reason that we reduced the 3-SAT problem to the Blocking Set Problem. Since
the 2-SAT problem is in P, we could expect that the Blocking Set Problem for two-element sets
is in P. We note that this case is especially interesting since the issue here is the blocking of the
edges of graphs. We can notice that the nodes outside a blocking set are independent (there
is no edge among them). The converse is true in the following sense: if an independent set is
maximal (no other node can be added to it while preserving independence) then its complement
is a blocking set for the edges. Our search for a minimum Blocking set can therefore be replaced
with a search for a maximum independent set, which is also a fundamental graph-theoretical
problem. Formulating it as a yes-no question:

(5.10) Remark The independent vertex set problem (and similarly, the Blocking set problem)
are only NP-complete if k is part of the input. It is namely obvious that if we fix k (e.g., k = 137)
then for a graph of n nodes it can be decided in polynomial time (in the given example, in time
O(n137 )) whether it has k independent nodes. The situation is different with colorability, where
already the colorability with 3 colors is NP-complete. ♦

From the NP-completeness of the Independent Set Problem, we get the NP-completeness
of two other basic graph-theory problems for free. First, notice that the complement of an
independent set of nodes is a blocking set for the family of edges, and vice versa. Hence we get

(5.11) Corollary The Blocking Set Problem for the family of edges of a graph is NP-complete.

(Recall that in the special case when the graph is bipartite, then the minimum size of a
blocking set is equal to the size of a maximum matching, and therefore it can be computed in
polynomial time.)
Another easy transformation is to look at the complementary graph G of G (this is the graph
on the same set of nodes, with “adjacent” and “non-adjacent” interchanged). An independent
set in G corresponds to a clique (complete subgraph) in G and vice versa. Thus the problem of
finding a k-element independent set is (trivially) reduced to the problem of finding a k-element
clique.

(5.12) Corollary The problem of deciding whether a graph has a clique of size k is NP-
complete.

Very many other important combinatorial and graph-theoretical problems are NP-complete:
— Does a given graph have a Hamiltonial circuit?

— Can we cover the nodes with disjoint triangles (for “2-angles”, this is the matching prob-
lem!),

93
 — Does there exist a family of k node-disjoint paths connecting k given pairs of nodes?

The book “Computers and Intractability” by Garey and Johnson (Freeman, 1979) lists NP-
complete problems by the hundreds.

5.6.2 NP-complete problems in arithmetic and algebra

A number of NP-complete problems is known also outside combinatorics. The most important
one among these is the following. In fact, the NP-completeness of this problem was observed
(informally, without an exact definition or proof) by Edmonds several years before the Cook–
Levin Theorem.

(5.13) Linear Diophantine inequalities Given a system Ax ≤ b of linear inequalities with

integer coefficients, decide whether it has a solution in integers. ♦

(Recall that the epithet “Diophantine” indicates that we are looking for the solution among
integers.)

(5.14) Theorem The solvability of a Diophantine system of linear inequalities is an NP-

complete problem.

Proof Let a 3-form B be given over the variables x1 , . . . , xn . Let us write up the following
inequalities:

0 ≤ xi ≤ 1 for all i,
xi1 + xi2 + xi3 ≥ 1 if xi1 ∨ xi2 ∨ xi3 is in B,
xi1 + xi2 + (1 − xi3 ) ≥ 1 if xi1 ∨ xi2 ∨ xi3 is in B,
xi1 + (1 − xi2 ) + (1 − xi3 ) ≥ 1 if xi1 ∨ xi2 ∨ xi3 is in B,
(1 − xi1 ) + (1 − xi2 ) + (1 − xi3 ) ≥ 1 if xi1 ∨ xi2 ∨ xi3 is in B.

The solutions of this system of inequalities are obviously exactly the value assignments satisfying
B, and so we have reduced the problem 3-SAT to the problem of solvability in integers of systems
of linear inequalities.
We mention that already a very special case of this problem is NP-complete:

(5.15) Subset sum problem: Given natural numbers a1 , . . . , am and b. Does the set
{a1 , . . . , am } have a subset whose sum is b? (The empty sum is 0 by definition.) ♦

(5.16) Theorem The subset sum problem is NP-complete.

94
Proof We reduce the set-partition problem to the subset sum problem. Let {A1 , . . . , Am } be
a family of subsets of the set S = {0, . . . , n − 1}, we want to decide whether it has a subfamily

giving a partition of S. Let q = m + 1 and let us assign a number ai = j∈Ai q j to each set Ai .
Further, let b = 1 + q + · · · + q n−1 . We claim that Ai1 ∪ · · · ∪ Aik is a partition of the set S if
and only if
ai1 + · · · + aik = b.
The “only if” is trivial. Conversely, assume ai1 + · · · + aik = b. Let dj be the number of those
sets Air that contain the element j (0 ≤ j ≤ n − 1). Then

ai1 + · · · + aik = dj q j .
j

Since the representation of the integer b with respect to the number base q is unique, it follow
that dj = 1, i.e., Ai1 ∪ · · · ∪ Aik is a partition of S.
This last problem illustrates nicely that the way we encode numbers can significantly influ-
ence the complexity of a problem. Let us assume that each number ai is encoded in such a way
that it requires ai bits (e.g., with a sequence 1 · · · 1 of length ai ). In short, we say that we use
the unary notation. The length of the input will increase this way, and therefore the number
of steps of the algorithms will be smaller in comparison with the length of the input.

(5.17) Theorem In unary notation, the subset sum problem is polynomially solvable.

(The general problem of solving linear inequalities in integers is NP-complete even under
unary notation; this is shown by the proof of Theorem 5.14 where only coefficients with absolute
value at most 2 are used.)

Proof For every p with 1 ≤ p ≤ m, we determine the set Tp of those natural numbers t that
can be represented in the form ai1 + · · · + aik , where 1 ≤ i1 ≤ · · · ≤ ik ≤ p. This can be done
using the following trivial recursion:
T0 = {0}, Tp+1 = Tp ∪ { t + ap+1 : t ∈ Tp }.
If Tm is found then we must only check whether b ∈ Tp holds.
We must see yet that this simple algorithm is polynomial. This follows immediately from

the observation that Tp ⊂ {0, . . . , i ai } and thus the size of the sets Tp is polynomial in the

size of the input, which is now i ai .
The method of this proof, that of keeping the results of recursive calls to avoid recomputation
later, is called dynamic programming.

(5.18) Remark A function f is called NP-hard if every problem in NP can be reduced to it

in the sense that if we add the computation of the value of the function f to the instructions of
the Random Access Machine (and thus consider it a single step) then every problem in NP can
be solved in polynomial time (the problem itself need not be in NP).
An NP-hard function may or may not be 01-valued (i.e., the characteristic function of a
language). The characteristic function of every NP-complete language is NP-hard, but there

95
are languages with NP-hard characteristic functions which are not in NP, and so are strictly
harder than any problem in NP (e.g., to decide about a position of the GO game on an n × n
board, who can win).
There are many important NP-hard functions whose values are not 0 or 1. If there is an
optimization problem associated with an NP-problem, like in many important discrete optimiza-
tion problems of operations research, then in case the problem is NP-complete the associated
optimization problem is NP-hard. Some examples:
— the famous Traveling Salesman Problem: a non-negative “cost” is assigned to each edge of
a graph, and we want to find Hamilton cycle with minimum cost (the cost of a Hamilton
cycle is the sum of costs of its edges);

— the Steiner problem (under the previous conditions, find a connected graph containing
given vertices with minimum cost);

— the knapsack problem (the optimization problem associated with a more general version
of the subset sum problem);

— a large fraction of scheduling problems.

Many enumeration problems are also NP-hard (e.g., to determine the number of all perfect
matchings, Hamilton circuits or legal colorings). ♦

(5.19) Remark Most NP problems occurring in practice turn out to be either NP-complete or
in P. Nobody succeeded yet to put either into P or among the NP-complete ones the following
problems:
PRIMALITY. Is a given natural number a prime?
BOUNDED DIVISOR. Does a given natural number n have a proper divisor not greater
than k?
GRAPH ISOMORPHISM. Are two given graph isomorphic?
The primality problem is probably in P (this is proved using an old number-theoretical
conjecture, the so-called Generalized Riemann Hypothesis; see the section on “Randomized
algorithms”). For the problems of bounded divisor and isomorphism, it is rather expected that
they are neither in P nor NP-complete. ♦

(5.20) Remark When a problem turns out to be NP-complete we cannot hope to find for
it such an efficient, polynomial algorithm such as e.g. for the matching problem. Since such
problems can be very important in practice we cannot give them up because of such a negative
result. Around an NP-complete problem, a mass of partial results of various types is born:
special classes for which it is polynomially solvable; algorithms that are exponential in the
worst case but are fairly well usable for not too large inputs, or for problems occuring in
practice (whether or not we are able to describe the special structure of “real word” problems
that make them easy); “heuristics” (approximation algorithms) that do not give exact solution
but (provably or in practice) give good approximation.
It is, however, sometimes just the complexity of the problems that can be utilized: see the
section on cryptography. ♦

96
5.2 Exercise A quantified Boolean expression is a Boolean expression in which the quantifiers
∀x and ∃x can also be used. Prove that the problem of deciding about a given quantified Boolean
expression whether it is true is in PSPACE. ♦

5.3 Exercise Let f be a length-preserving one-to-one function over binary strings computable
in polynomial time. We define the language L of those strings y for which there is an x with
f (x) = y such that the first bit of x is 1. Prove that L is in NP ∩ co−NP. ♦

5.4 Exercise We say that a quantified Boolean formula is in class Fk if all of its quantifiers
are in front and the number of alternations between existential and universal quantifiers is at
most k. Let Lk be the set of true closed formulas in Fk . Prove that if P = NP then for all k we
have Lk ∈ P. [Hint: induction on k.] ♦

5.5 Exercise Show that the following general matching problem is in both NP and co-NP:
Given a graph G and a natural number k, does there exist a k-edge matching in G? ♦

5.6 Exercise Give a polynomial algorithm to decide whether a disjunctive normal form is
satisfiable. ♦

5.7 Exercise Show that the Satisfiablity Problem can be reduced to the special case when
each variable occurs at least once unnegated and at least one negated. ♦

5.8 Exercise Given a graph G and a variable xv for each vertex v of G. Write up a conjunctive
normal form that is true if and only if the values of the variables give a legal coloring of the
graph G with 2 colors. (I.e. the normal form is satisfiable if and only bif the graph is colorable
with 2 colors.) ♦

5.9 Exercise Given a graph G and three colors, 1,2 and 3. Let us introduce, to each vertex v
and color i a logical value x[v, i]. Write up a conjunctive normal form B for the variables x[v, i]
which is satisfiable if and only if G can be colored with 3 colors.
[Hint: Let B be such that it is true if and only if there is a coloring of the vertices with the
given 3 colors for which x[v, i] is true if and only if vertex v has color i.] ♦

5.10 Exercise Define the language 3-SAT-3 and show that it is NP-complete. ♦

5.11 Exercise Define the language SAT-2 and show that it is in P. ♦

5.12 Exercise Prove that it is also NP-complete to decide whether in a given 2n-vertex graph,
there is an n-element independent set. ♦

5.13 Exercise In the GRAPH EMBEDDING PROBLEM, what is given is a pair (G1 , G2 )
of graphs. We ask whether G2 has a subgraph isomorphic to G1 . Prove that this problem is
NP-complete. ♦

5.14 Exercise Prove that it is also NP-complete to decide whether the chromatic number of
a graph G (the smallest number of colors with which its vertices can be colored) is equal to the
number of elements of its largest complete subgraph. ♦

97
5.15 Exercise Prove that if a system of sets is such that every element of the (finite) underlying
set belongs to at most two sets of the system, then the Blocking Set Problem with for this system
is polynomial time solvable.
[Hint: reduce it to the general matching problem.] ♦

5.16 Exercise Prove that for “hypergraphs”, already the problem of coloring with two colors
is NP-complete: Given a system {A1 , . . . , An } of subsets of a finite set. Can the nodes of S be
colored with two colors in such a way that each Ai contains nodes of both colors? ♦

5.17 Exercise An instance of the problem of 0-1 Integer Programming is defined as follows.
The input of the problem is the arrays of integers aij , bi for i = 1, . . . , m, j = 1, . . . , n. The task
is to see if the set of equations

n
aij xj = bi (i = 1, . . . , m)
j=1

is satisfiable with xj = 0, 1. The Subset Sum Problem is a special case with m = 1.

Make an immediate reduction of the 0-1 Integer Programming problem to the Subset Sum
Problem. ♦

5.18 Exercise The SUM PARTITION PROBLEM is the following. Given a set A =

{a1 , . . . , an } of integers, find a subset B of A such that i∈B ai = i∈B ai . Prove that this
problem is NP-complete. [Hint: use the NP-completeness of the subset sum problem.] ♦

5.19 Exercise Consider the following problem: given a finite set of integers in binary form,
one has to partition this set into three subsets such that the sum of numbers within each subset
is the same. Prove that this problem is NP-complete. You can use the fact that the problem of
partitioning into two sets is known to be NP-complete. ♦

5.20 Exercise The bounded tiling problem B is the following language. Its words have the form
T &n&s. Here, the string T represents a set of tile types and n represents a natural number.
The string s represents a sequence of 4n − 4 tiles. The string T &n&s belongs to B if and only
if there is a tiling of an n × n square with tiles whose type is in T in such a way that the tiles
on the boundary of the square are given by s (starting, say, at the lower left corner and going
counterclockwise). Prove that the language B is NP-complete. ♦

5.21 Exercise Consider the following tiling problem T . Given is a finite set of tile types with
a distinguished initial tile I among them and a number n in binary. It is to decide whether an
n × n square can be tiled tiles of these types, when the lower left corner must be the initial tile.
Prove that if T is solvable in time polynomial in n then NEXPTIME=EXPTIME. ♦

98
6 Randomized algorithms
6.1 Introduction
We cited Church’s Thesis in Section: every ”algorithm” (in the heuristic meaining of the word)
is realizable on a Turing machine. It turned out that other models of computation were able to
solve exactly the same class of problems.
But there is an extension of the notion of an algorithm that is more powerful than a Turing
machine, and still realizable in the real world. This is the notion of a randomized algorithm:
we permit ”coin tossing”, i.e., the have access to a random number generator. Such machines
will be able to solve problems that the Turing machine cannot solve (we will formulate and
prove this in an exact sense in a later section); furthermore, such machines can solve some
problems more efficiently than Turing machines. We start with a discussion of such examples.
The simplest example of such an application of randomization is checking an algebraic identity;
the most important is quite certainly testing whether an integer is a prime.
Since in this way, we obtain a new, stronger mathematical notion of a machine, corresponding
randomized complexity classes can also be introduced. Some of the most important ones will
be treated at the end of the Chapter.

6.2 Verifying a polynomial identity

Let f (x1 , . . . , xn ) be a rational polynomial with n variables that has degree at most k in each of
its variables. We would like to decide whether f is identically 0 (as a function of n variables).
We know from classical algebra that a polynomial is identically 0 if and only if, after “opening its
parentheses”, all terms “cancel”. This criterion is, however, not always useful. It is conceivable,
e.g., that the polynomial is given in a parenthesized form an the opening of the parentheses
leads to exponentially many terms as in

(x1 + y1 )(x2 + y2 ) · · · (xn + yn ) + 1.

It would also be good to say something about polynomials in whose definition not only the basic
algebraic operations occur but also some other ones, like the computation of a determinant
(which is a polynomial itself but is often computed, as we have seen, in some special way).
The basic idea is that we write random numbers in place of the variables and compute the
value of the polynomial. If this is not 0 then, naturally, the polynomial cannot be identically 0.
If the computed value is 0 then though it can happen that the polynomial is not identically 0,
but “hitting” one of its roots has very small probability; therefore in this case we can conclude
that the polynomial is identically 0; the probability that we make a mistake is small.
If we could give real values to the variables, chosen according to the uniform distribution
e.g. in the interval [0, 1], then the probability of error would be 0. We must in reality, however,
compute with discrete values; therefore we assume that the values of the variables are chosen
from among the integers of the interval [0, N − 1], independently and according to the uniform
distribution. In this case, the probability of error will not be 0 but it will be “small” if N is
large enough. This is the meaning of the following fundamental result:

99
(6.1) Theorem Schwartz’s Lemma If f is a not identically 0 polynomial in n variables with de-
gree at most k, and the values ξi (i = 1, . . . , n) are chosen in the interval [0, N −1] independently
of each other according to the uniform distribution then
k
Prob{f (ξ1 , . . . , ξn ) = 0} ≤ .
N
(The degree of a polynomial in several variables is defined as the largest degree of its terms
(monomials); the degree of a monomial is the sum of the exponents of the variables in it.)

Proof We prove the assertion by induction on n. The statement is true for n = 1 since a
polynomial in one variable of degree k can have at most k roots. Let n > 1 and let us arrange
f according to the powers of x1 :

f = f0 + f1 x1 + f2 x21 + · · · + ft xt1 ,

where f0 , . . . , ft are polynomials of the variables x2 , . . . , xn , the term ft is not identically 0, and
t ≤ k. Now,

Prob{f (ξ1 , . . . , ξn ) = 0}
≤ Prob{f (ξ1 , . . . , ξn ) = 0 | ft (ξ2 , . . . , ξn ) = 0}Prob{ft (ξ2 , . . . , ξn ) = 0}
+ Prob{f (ξ1 , . . . , ξn ) = 0 | ft (ξ2 , . . . , ξn ) = 0}Prob{ft (ξ2 , . . . , ξn ) = 0}
≤ Prob{ft (ξ2 , . . . , ξn ) = 0} + Prob{f (ξ1 , . . . , ξn ) = 0 | ft (ξ2 , . . . , ξn ) = 0}.

Here can estimate the first term by the induction hypothesis, using that the degree of ft is at
most k − t; thus the first term is at most (k − t)/N . The second term is at most t/N (since ξ1
is independent of the variables ξ2 , . . . , ξn , therefore no matter how the latter are fixed in such
a way that ft = 0 ( and therefore f as a polynomial of x1 is not identically 0), the probability
that ξ1 is its root is at most t/N ). Hence
k−t t k
Prob{f (ξ1 , . . . , ξn ) = 0} ≤ + ≤ .
N N N

This suggests the following a randomized algorithm to decide whether a polynomial f is

identically 0:

(6.2) Algorithm We compute f (ξ1 , . . . , ξn ) with integer values ξi chosen randomly and inde-
pendently of each other according to the uniform distribution in the interval [0, 2kn]. If we don’t
get the value 0 we stop: f is not identically 0. If we get 0 value we repeat the computation. If
we get 0 value 100 times we stop and declare that f is identically 0.

If f is identically 0 then this algorithm will determine this. If f is not identically 0 then in
every separate iteration—according to Schwartz’s Lemma—the probability that the result is 0
is less than 1/2. With 100 experiments repeated independently of each other, the probability
that this occurs every time, i.e., that the algorithm asserts erroneously that f is identically 0,
is less than 2−100 .

100
 Two things are needed for us to be able to actually carry out this algorithm: on the one hand,
we must be able to generate random numbers (here, we assume that this can be implemented,
and even in time polynomial in the number of bits of the integers to be generated), on the other
hand, we must be able to evaluate f in polynomial time (the size of the input is the length of
the “definition” of f ; this definition can be, e.g., an expression containing multiplications and
additions with parentheses, but also something entirely different, e.g., a determinant form).
As a surprising example for the application of the method we present a matching algorithm.
(We have already treated the matching problem in Chapters 3 and 5). Let G be a bipartite
graph with the edge set E(G) whose edges run between sets A and B, A = {a1 , . . . , an },
B = {b1 , . . . , bn }. Let us assign to each edge ai bj a variable xij . Let us construct the n × n
matrix M as follows:

xij if ai bj ∈ E(G),
mij =
0 otherwise.
The determinant of this graph is closely connected with the matchings of the graph G as Dénes
Kőnig noticed while analyzing a work of Frobenius:
(6.3) Theorem There is a perfect matching in the bipartite graph G if and only if det(M ) is
not identically 0.

Proof Consider a term in the expansion of the determinant:

±m1π(1) m2π(2) · · · mnπ(n) ,
where π is a permutation of the numbers 1, . . . , n. For this not to be 0, we need that ai and bπ(i)
be connected for all i; in other words, that {a1 bπ(1) , . . . , an bπ(n) } be a perfect matching in G. In
this way, if there is no perfect matching in G then the determinant is identically 0. If there are
perfect matchings in G then to each one of them a nonzero expansion term corresponds. Since
these terms do not cancel each other (any two of them contain at least two different variables),
the determinant is not identically 0.
Since det(M ) is a polynomial of the elements of the matrix M that is computable in poly-
nomial time (e.g. by Gaussian elimination) this theorem offers a polynomial-time randomized
algorithm for the matching problem in bipartite graphs. We mentioned it before that there
is also a polynomial-time deterministic algorithm for this problem (the “Hungarian method”).
One advantage of the algorithm treated here is that it is very easy to program (determinant-
computation can generally be found in the program library). If we use “fast” matrix mul-
tiplication methods then this randomized algorithm is a little faster than the fastest known
deterministic one: it can be completed in time O(n2.4 ) instead of O(n2.5 ). Its main advantage
is, however, that it is well suitable to parallelization, as we will see in a later section.
In non-bipartite graphs, it can also be decided by a similar but slightly more complicated
method whether there is a perfect matching. Let V = {v1 , . . . , vn } be the vertex set of the graph
G. Assign again to each edge vi vj (where i < j) a variable xij and construct an asymmetric
n × n matrix T = (tij ) as follows:
⎧
⎪
⎨ xij if vi vj ∈ E(G) and i < j,
tij = −xij if vi vj ∈ E(G) and i > j,
⎪
⎩ 0 otherwise.

101
The following analogue of the above cited Frobenius-Kőnig theorem comes from Tutte and we
formulate it here without proof:

(6.4) Theorem There is a perfect matching in the graph G if and only if det(T ) is not iden-
tically 0.

This theorem offers, similarly to the case of the bipartite graph, a randomized algorithm for
deciding whether there is a perfect matching in G.

102
6.3 Prime testing
Let m be an odd natural number, we want to decide whether it is a prime. We have seen in the
previous chapter that this problem is in NP ∩ co−NP. The witnesses described there did not
lead, however, (at least for the time being) to a polynomial-time prime test. We will therefore
give first a new, more complicated NP description of compositeness.

(6.1) Theorem “Little” Fermat Theorem If m is a prime then am−1 − 1 is divisible by m for
all natural numbers 1 ≤ a ≤ m − 1.

If—for a given m—the integer am−1 − 1 is divisible by m then we say that a satisfies the
Fermat condition.
The Fermat condition, when required for all integers 1 ≤ a ≤ m − 1, also characterizes
primes:

(6.2) Lemma An integer m > 0 is a prime if and only all integers 1 ≤ a ≤ m − 1 satisfy the
Fermat condition.

Indeed, if m is composite, then we can choose for a an integer not relatively prime to m,
and then am−1 − 1 is obviously not divisible by m.
A further nice feature of the Fermat condition is that it can be checked in polynomial time
for given m and a. This was discussed in chapter 5.
Of course, we cannot check the Fermat condition for every a: this would take exponential
time. The question is therefore to which a’s should we apply it?
We could just try a = 2. This is in fact not a bad test, and it reveals the non-primality of
many (in a sense, most) composite numbers, but it may fail. For example, 561 = 3 · 11 · 17 is
not a prime, but 561 | 2560 − 1. Any other specific choice of a would have similar shortcoming.
The next idea is to select as random a and check Fermat condition. If m is a prime, then
a will of course satisfy it. Suppose that m is not prime, then at least those a’s not relatively
prime to m will violate it. Unfortunately, the number of such a’s may be miniscule compared
with the number of all choices for a, and so the probability that our random choice will pick
one is negligible. (In fact, we can compute the greatest common divisor of a and m right away:
if we can find an a not relatively prime to m, then this will yield a proper divisor of m.)
So we need to use a’s relatively prime to m. Unfortunately, there are composite numbers
m (the so-called pseudo-primes) for which the Fermat condition is satisfied for all a relatively
prime to m; for such numbers, it will be especially difficult to find an integer a violating the
condition. (Such a pseudo-prime is e.g. 561 = 3 · 11 · 17.)
But at least of m is not a pseudo-prime, then the random choice for a works. This is
guaranteed by the following lemma.

(6.3) Lemma If m is not a prime and not a pseudo-prime then at most half of the integers a,
1 ≤ a ≤ m − 1 relatively prime to m satisfies the Fermat condition.

Note that none of the non-relatively-prime a’s satisfy the Fermat condition.

103
Proof Since m is not a pseudoprime, there is at least one b relatively prime to m such that
bm−1 − 1 is not divisible by m. Now if a is “bad”, i.e., am−1 − 1 is divisible by m, then ab mod
m is “good”:
(ab)m−1 − 1 = (am−1 − 1)bm−1 + bm−1 − 1,
and here the first term is divisible by m but the second is not.
Hence for every a that is “bad”, we find another a (namely ab mod m) that is “good”. It is
easy to see that different a’s yield different “good” numbers. Thus at least half of the numbers
a must be good.
Thus, if m is not a pseudo-prime then the following randomized prime test works: check
whether a randomly chosen integer 1 ≤ a ≤ m − 1 satisfies the Fermat condition. If not then
we know that m is not a prime. If yes then repeat the procedure. If we found 100 times,
independently of each other, that the Fermat condition is satisfied then we say that m is a
prime. It can still happen that m is composite, but if the probability that we picked an integer
a satisfying the condition is less than 1/2 at every step, and hence the probability that this
occurs 100 times in a row is less than 2−100 .
Unfortunately, this method fails for pseudoprimes (it finds them prime with large proba-
bility). It turns out that one can modify the Fermat condition just a little to overcome this
difficulty. Let us write the number m − 1 in the form 2k M where M is odd. We say that a
satisfies the Miller condition if at least one of the numbers

aM − 1, aM + 1, a2M + 1, a4M + 1, . . . , a2
k−1 M
+1

is divisible by m. Note that the product of these numbers is just am−1 − 1. Hence every number
satisfying the Miller condition also satisfies the Fermat condition.
If m is a prime then it divides this product, and hence it divides one of these factors, i.e.,
every a satisfies the Miller condition. If m is composite then, however, it could happen that
some a satisfies the Fermat condition but not the Miller condition (m can be a divisor of a
product without being the divisor of any of its factors).
Thus the Miller condition provides a potentially stronger primality test than the Fermat
condition. The question is: how much stronger?
We will need some fundamental facts about pseudoprimes.

(6.4) Lemma Every pseudoprime m is

(a) Odd

(b) Squarefree (is not divisible by any square).

Proof (a) If m > 2 is even then a = m−1 will violate the Fermat condition, since (m−1)m−1 ≡
−1 ≡ 1 (mod m).
(b) Assume that p2 | m; let k be the largest exponent for which pk | m. Then a = m/p − 1
violates the Fermat condition since the last two terms of the binomial expansion of (m/p−1)m−1
are −(m − 1)(m/p) + 1 ≡ m/p + 1 ≡ 1 (mod pk ) (all earlier terms are divisible by pk ) and if
an integer is not divisible by pk then it is not divisible by m either.

104
(6.5) Lemma Let m = p1 p2 · · · pt where the pi ’s are different primes. The relation ak ≡ 1
(mod m) holds for all a relatively prime to m if and only if pi − 1 divides k for all i with
1 ≤ i ≤ t.

Proof If pi − 1 divides k for all i with 1 ≤ i ≤ t then ak − 1 is divisible by pi according to

the little Fermat Theorem and then it is also divisible by m. Conversely, suppose that ak ≡ 1
(mod m) for all a relatively prime to m. If e.g. p1 − 1 would not divide k then let g be a
primitive root modulo p1 (the existence of primitive roots was stated in Chapter 5) According
to the Chinese Remainder Theorem, there is a residue class h modulo m with h ≡ g (mod p1 )
and h ≡ 1 (mod pi ) for all i ≥ 2. Then (h, m) = 1 and p1  | hk − 1, so m  | hk − 1.

(6.6) Corollary The number m is a pseudoprime if and only if m = p1 p2 · · · pt where the pi ’s

are different primes, t ≥ 2, and (pi − 1) divides (m − 1) for all i with 1 ≤ i ≤ t.

(6.7) Remark This is how one can show about the above example 561 that it is a pseudoprime.
♦

Now we can prove the main fact that makes the Miller test better than the Fermat test.

(6.8) Theorem If m is a composite number then at least half of the numbers 1, . . . , m − 1

violate the Miller condition.

Proof Since we have already seen that the lemma holds for non-pseudoprimes, in what follows
we can assume that m is a pseudoprime. Let p1 · · · pt (t ≥ 2) be the prime decomposition of m.
By the above, these primes are all odd and distinct, and we have (pi − 1) | (m − 1) = 2k M for
all i with 1 ≤ i ≤ t.
Let l be the largest exponent with the property that none of the numbers pi − 1 divides 2l M .
Since the numbers pi − 1 are even while M is odd, such an exponent exists (e.g. 0) and clearly
0 ≤ l < k. Further, by the definition of l, there is a j for which pj − 1 divides 2l+1 M . Therefore
pj − 1 divides 2s M for all s with l < s ≤ k, and hence pj divides a2 M − 1 for all primitive
s

residue classes a. Consequently pj cannot divide a2 M + 1 which is larger by 2, and hence m

s

does not divide a2 M + 1 either. If therefore a is a residue class that is a Miller accomplice
s

then m must already be a divisor of one of the remainder classes aM − 1, aM + 1, a2M + 1, . . .,

a2 M + 1, a2 M + 1. Hence for each such a, the number m divides either the product of the
l−1 l

first l + 1, which is (a2 M − 1), or the last one, (a2 M + 1). Let us call the primitive residue class
l l

a modulo m an “accomplice of the first kind” if a2 M ≡ 1 (mod m) and an “accomplice of the

l

second kind” if a2 M ≡ −1 (mod m).

l

Let us estimate first the number of accomplices of the first kind. Consider an index i with
1 ≤ i ≤ t. Since pi − 1 does not divide the exponent 2l M , Lemma 6.5 implies that there is
a number a not divisible by pi for which a2 M − 1 is not divisible by pi . The reasoning of
l

Lemma 6.3 shows that then at most half of the mod pi residue classes will satisfy the Fermat
comdition belonging to the above exponent, i.e. such that a2 M − 1 is divisible by pi . According
l

to the Chinese Remainder Theorem, there is a one-to-one correspondence between the primitive
residue classes with respect to the product p1 · · · pt as modulus and the t-tuples of primitive

105
residue classes modulo the primes p1 , . . . , pt . Thus, modulo p1 · · · pt , at most a 2t -th fraction of
the primitive residue classes is such that every pi divides (a2 M − 1). Therefore at most a 2t -th
l

fraction of the mod m primitive residue classes are accomplices of the first kind.
It is easy to see that the product of two accomplices of the second kind is one of the first
kind. Hence multiplying all accomplices of the second kind by a fixed one of the second kind,
we obtain accomplices of the first kind, and thus the number of accomplices of the second kind
is at least as large as the number of accomplices of the first kind. (If there is no accomplice of
the second kind to multiply with then the situation is even better: zero is certainly not greater
than the number of accomplices of the first kind.) Hence even the two kinds together make up
at most a 2t−1 -th part of the primitive residue classes, and so (due to t ≥ 2) at most a half.

(6.9) Lemma For a given m and a, it is decidable in polynomial time whether a satisfies the
Miller condition.

For this, it is enough to recall from Chapter 3 that the remainder of ab modulo c is computable
in polynomial time. Based on these three lemmas, the following randomized algorithm, called
the Miller–Rabin test, can be given for prime testing:

(6.10) Algorithm We choose a number between 1 and m − 1 randomly and check whether it
satisfies the Miller condition. If it does not then m is composite. If it does then we choose a
new a. If the Miller condition is satisfied 100 times consecutively then we declare that m is a
prime. ♦

If m is a prime then the algorithm will certainly assert this. If m is composite then the
number a chosen randomly violates the Miller condition with probability 1/2. After hundred
independent experiments the probability will therefore be at most 2−100 that the Miller condition
is not violated even once, i.e., that the algorithm asserts that m is a prime.

(6.11) Remarks
1. If m is found composite by the algorithm then — interestingly enough — we see this not
from its finding a divisor but from the fact that one of the residues violates the Miller
condition. If at the same time, the residue a does not violate the Fermat condition, then
m cannot be relatively prime to each of the numbers aM − 1, aM + 1, a2M + 1, a4m + 1, · · ·,
a2 M + 1, therefore computing its greatest common divisors with each, one of them will
k−1

be a proper divisor of m. No polynomial algorithm (either deterministic or randomized)

is known for finding a factorization in the case when the Fermat condition is also violated.
This problem is significantly more difficult also in practice than the testing of primality. We
will see in the section on cryptography that this empirical fact has important applications.
2. For a given m, we can try to find an integer a violating the Miller condition not by
random choice but by trying out the numbers 1,2, etc. It is not known how small is
the first such integer if m is composite. Using, however, a hundred year old conjecture
of analytic number theory, the so-called Generalized Riemann Hypothesis, (which is too
technical to be formulated here) one can show that it is not greater than log m. Thus, this
deterministic prime test works in polynomial time if the Generalized Riemann Hypothesis
is true.

106
♦

We can use the prime testing algorithm learned above to look for a prime number with
n digits (say, in the binary number system). Choose namely a number k randomly from the
interval [2n−1 , 2n − 1] and check whether it is a prime, say, with an error probability of at most
2−100 /n. If it is, we stop. If it is not we choose a new number k. Now, it follows from the
theory of prime numbers that in this interval, not only there is a prime number but the number
of primes is rather large: asymptotically (log e)2n−1 /n, i.e., a randomly chosen n-digit number
will be a prime with probability cca. 1.44/n. Repeating therefore this experiment O(n) times
we find a prime number with very large probability.
We can choose a random prime similarly from any sufficiently long interval, e.g. from the
interval [1, 2n ].

107
6.4 Randomized complexity classes
In the previous subsections, we treated algorithms that used random numbers. Now we define
a class of problems solvable by such algorithms.
First we define the corresponding machine. Let T = (k, Σ, Γ, Φ) be a non-deterministic
Turing machine and let us be given a probability distribution for every g ∈ Γ, h1 , . . . , hk ∈ Σ
on the set

{ (g  , h1 , . . . , hk , ε1 , . . . , εk ) : (g, h1 , . . . , hk , g  , h1 , . . . , hk , ε1 , . . . , εk ) ∈ Φ }.

(It is useful to assume that the probabilities of events are rational numbers, since then events
with such probabilities are easy to generate, provided that we can generate mutually indepen-
dent bits.) A non-deterministic Turing machine together with these distributions is called a
randomized Turing machine.
Every legal computation of a randomized Turing machine has some probability. We say that
a randomized Turing machine weakly decides (or, decides in the Monte-Carlo sense) a
language L if for all inputs x ∈ Σ∗ , it stops with probability at least 3/4 in such a way that in
case of x ∈ L it writes 1 on the result tape, and in case of x ∈ L, it writes 0 on the result tape.
Shortly: the probability that it gives a wrong answer is at most 1/4.
In our examples, we used randomized algorithms in a stronger sense: they could err only in
one direction. We say that a randomized Turing machine accepts a language L if for all inputs
x, it always rejects the word x in case of x ∈ L, and if x ∈ L then the probability is at least 1/2
that it accepts the word x.
We say that a randomized Turing machine strongly decides (or, decides in the Las
Vegas sense) a language L if it gives a correct answer for each word x ∈ Σ∗ with probability
1. (Every single computation of finite length has positive probability and so the 0-probability
exception cannot be that the machine stops with a wrong answer, only that it works for an
infinite time.)
In case of a randomized Turing machine, for each input, we can distunguish the number of
steps in the longest computation and the expected number of steps. The class of all languages
that are weakly decidable on a randomized Turing machine in polynomial expected time is
denoted by BPP (Bounded Probability Polynomial). The class of languages that can be accepted
on a randomized Turing machine in polynomial expected time is denoted by RP (Random
Polynomial). The class of all languages that can be strongly decided on a randomized Turing
machine in polynomial expected time is denoted by ΔRP. Obviously, BPP ⊃ RP ⊃ ΔRP ⊃ P.
The constant 3/4 in the definition of weak decidability is arbitrary: we could say here any
number smaller than 1 but greater than 1/2 without changing the definition of the class BPP
(it cannot be 1/2: with this probability, we can give a correct answer by coin-tossing). If
namely the machine gives a correct answer with probability 1/2 < c < 1 then let us repeat the
computation t times on input x and accept as answer the one given more often. It is easy to
see from the Law of Large Numbers that the probability that this answer is wrong is less than
ct1 where c1 is a constant smaller than 1 depending only on c. For sufficiently large t this can be
made arbitrarily small and this changes the expected number of steps only by a constant factor.
It can be similarly seen that the constant 1/2 in the definition of acceptance can be replaced
with an arbitrary positive number smaller than 1.

108
 Finally, we note that instead of the expected number of steps in the definition of the classes
BPP and RP, we could also consider the largest number of steps; this would still not change
the classes. Obviously, if the largest number of steps is polynomial, then so is the expected
number of steps. Conversely, if the expected number of steps is polynomial, say, at most |x|d ,
then according to Markov’s Inequality, the probability that a computation lasts a longer time
than 8|x|d is at most 1/8. We can therefore build in a counter that stops the machine after 8|x|d
steps, and writes 0 on the result tape. This increases the probability of error by at most 1/8.
The same is, however, not known for the class ΔRP: the restriction of the longest running
time would lead here already to a deterministic algorithm, and it is not known whether ΔRP is
equal to P (moreover, this is rather expected not to be the case; there are examples for problems
solvable by polynomial Las Vegas algorithms for which no polynomial deterministic algorithm
is known).

(6.1) Remark A Turing machine using randomness could also be defined in a different way:
we could consider a deterministic Turing machine which has, besides the usual (input-, work-
and result-) tapes also a tape on whose every cell a bit (say, 0 or 1) is written that is selected
randomly with probability 1/2. The bits written on the different cells are mutually independent.
The machine itself works deterministically but its computation depends, of course, on chance
(on the symbols written on the random tape). It is easy to see that such a deterministic Turing
machine fitted with a random tape and the non-deterministic Turing machine fitted with a
probability distribution can replace each other in all definitions.
We could also define a randomized Random Access Machine: this would have an extra cell
w in which there is always a 0 or 1 with probability 1/2. We have to add the instruction y := w
to the programming language. Every time this is executed a new random bit occurs in the cell
w that is completely independent of the previous bits. Again, it is not difficult to see that this
does not bring any significant difference. ♦

It can be seen that every language in RP is also in NP. It is trivial that the classes BPP and
ΔRP are closed with respect to the taking of complement: they contain, together with every
language L the language Σ∗ \ L. The definition of the class RP is not such and it is not known
whether this class is closed with respect to complement. It is therefore worth defining the class
co−RP: A language L is in co−RP if Σ∗ \ L is in RP.
“Witnesses” provided a useful characterization of the class NP. An analogous theorem holds
also for the class RP.

(6.2) Theorem A language L is in RP if and only if there is a language L ∈ P and a

polynomial f (n) such that

(a) L = { x ∈ Σ∗ : y ∈ Σf (|x|) x&y ∈ L } and

(b) if x ∈ L then at least half of the words y of length f (|x|) are such that x&y ∈ L .

Proof Similar to the proof of the corresponding theorem on NP.

The connection of the classes RP and ΔRP is closer than it could be expected on the basis
of the analogy to the classes NP and P:

109
(6.3) Theorem The following properties are equivalent for a language L:
(i) L ∈ ΔRP;
(ii) L ∈ RP ∩ co−RP;
(iii) There is a randomized Turing machine with polynomial (worst-case) running time that
can write, besides the symbols “0” and “1”, also the words “I GIVE UP”; the answers
“0” and “1” are never wrong, i.e., in case of x ∈ L the result is “1” or “I GIVE UP”, and
in case of x ∈ L it is “0” or “I GIVE UP”. The probability of the answer “I GIVE UP” is
at most 1/2.

Proof It is obvious that (i) implies (ii). It can also be easily seen that (ii) implies (iii). Let us
submit x to a randomized Turing machine that accepts L in polynomial time and also to one
that accepts Σ∗ \ L in polynomial time. If the two give opposite answers then the answer of the
first machine is correct. If they give identical answers then we “give it up”. In this case, one of
them made an error and therefore this has a probability at most 1/2.
Finally, to see that (iii) implies (i) we just have to modify the Turing machine T0 given in
(iii) in such a way that instead of the answer “I GIVE IT UP”, it should start again. If on
input x, the number of steps of T0 is τ and the probability of giving it up is p then on this same
input, the expected number of steps of the modified machine is
∞
 τ
pt−1 (1 − p)tτ = ≤ 2τ.
t=1
1−p

We have seen in the previous subsection that the “language” of composite numbers is in RP.
Even more is true: Adleman and Huang have shown that this language is also in ΔRP. For our
other important example, the not identically 0 polynomials, it is only known that they are in
RP. Among the algebraic (mainly group-theoretical) problems, there are many that are in RP
or ΔRP but no polynomial algorithm is known for their solution.

(6.4) Remark The algorithms that use randomization should not be confused with the algo-
rithms whose performance (e.g., the expected value of their number of steps) is being examined
for random inputs. Here we did not assume any probability distribution on the set of inputs,
but considered the worst case. The investigation of the behavior of algorithms on random inputs
coming from a certain distribution is an important but difficult area, still in its infancy, that we
will not treat here. ♦

6.1 Exercise Suppose that some experiment has some probability p of success. Prove that in
n3 experiments,
 it is possible to compute an approximation p̂ of p such that the probability of
|p − p̂| > p(1 − p)/n is at most 1/n. [Hint: Use Tshebysheff’s Inequality.] ♦

6.2 Exercise We want to compute a real quantity a. Suppose that we have a randomized
algorithm that computes an approximation A (which is a random variable) such that the prob-
ability that |A − a| > 1 is at most 1/20. Show that by calling the algorithm t times, you can
compute an approximation B such that the probability that |B − a| > 1 is at most 2−t . ♦

110
6.3 Exercise Suppose that somebody gives you three n × n matrices A, B, C (of integers of
maximimum length l) and claims C = AB. You are too busy to verify this claim exactly and
do the following. You choose a random vector x of length n whose entries are integers chosen
uniformly from some interval [0, . . . , N − 1], and check A(Bx) = Cx. If this is true you accept
the claim otherwise you reject it.

(a) How large must N be chosen to make the probability of false acceptance smaller than
0.01?

(b) Compare the time complexity the probabilistic algorithm to the one of the deterministic
algorithm computing AB.

6.4 Exercise Show that if m is a pseudoprime then the Miller–Rabin test not only discovers
this with large probability but it can also be used to find a decomposition of m into two factors.
♦

6.5 Exercise Show that the Turing machine equipped with a random tape and the non-
deterministic Turing machine equipped with a probability distribution are equivalent: if some
language is accepted in polynomial time by the one then it is also accepted by the other one. ♦

6.6 Exercise Formulate what it means that a randomized RAM accepts a certain language
in polynomial time and show that this is equivalent to the fact that some randomized Turing
machine accepts it. ♦

6.7 Exercise Let us call a Boolean formula with n variables robust if it is either unsatisfiable
or has at least 2n /n2 satisfying assignments. Give a probabilistic polynomial algorithm to decide
the satisfiability of robust formulas. ♦

111
7 Information complexity: the complexity-theoretic notion of
randomness
7.1 Introduction
The mathematical foundation of probability theory appears among the famous problems of
Hilbert formulated in 1900 (mentioned before). Von Mises made an important attempt in 1919
to define the randomness of a 0-1 sequence. his attempt can be sketched as follows. We require
that the frequency of 0’s and 1’s be approximately the same. This is clearly not enough, but
we can require the same to hold also if we select every other number of the sequence. more
generally, we can require the same for all subsequences obtained by selecting indices from an
arithmetic progression. This approach, however, did not prove sufficiently fruitful.
In 1931 Kolmogorov initiated another approach, using measure theory. His theory was
very successful from the point of view of probability theory, and it is the basis of the rigorous
development of probability theory in any textbook today.
However, this standard approach fails to capture some important aspects For example, in
probability theory based on measure theory, we cannot speak of the randomness of a single 0-1
sequence, only of the probability of a set of sequences. At the same time, in an everyday sense, it
is ”obvious” that the sequence ”Head, Head, Head,...” cannot be the result of coin tossing. In the
1960’s Kolmogorov and independently Chaitin revived the idea of von Mises, using complexity-
theoretic tools. They defined the information complexity (information content) of a sequence;
then (roughly speaking) random sequences are those whose information content is as large as
possible. The importance of these results goes beyond the foundation of probabity theory;
it contributes to the clarification of the basic notions in several fields like data compression,
information theory and statistics.
In this chapter we introduce the notion of information complexity first. Then we discuss the
notion of an informatically random sequence, and show that such sequences behave like ”usual”
random sequences: they obeu the Laws of Large Numbers. Finally, we discuss the problem of
optimal encoding of various structures.

7.2 Information complexity

Fix an alphabet Σ. Let Σ0 = Σ \ {∗} and consider a two-tape universal Turing machine over Σ.
It will be convenient to identify Σ0 with the set {0, 1, . . . , m − 1}. Consider a 2-tape, universal
Turing machine T over Σ. We say that the word (program) q over Σ prints word x if writing q
on the second tape (the program tape) of T and leaving the first tape empty, the machine stops
in finitely many steps with the word x on its first tape (the data tape).
Let us note right away that every word is printable on T . There is namely a one-tape
(perhaps large, but rather trivial) Turing machine Sx that, when started with the empty tape,
writes the word x onto it and halts. This Turing machine can be simulated by a program qx
that, in this way, prints x.
The information complexity (also called Kolmogorov complexity) of a word x ∈ Σ∗0 we
mean the length of the shortest word (program) that makes T print the word x. We denote the
complexity of the word x by KT (x).

112
 We can also consider the program printing x as a “code” of the word x where the Turing
machine T performs the decoding. This kind of code will be called a Kolmogorov code. For
the time being, we make no assumptions about how much time this decoding (or encoding,
finding the appropriate program) can take.
We would like the complexity to be a characteristic property of the word x and to depend on
the machine T as little as possible. It is, unfortunately, easy to make a Turing machine that is
obviously “clumsy”. For example, it uses only every second letter of each program and “skips”
the intermediate letters. Such a machine can be universal, but every word will be defined twice
as complex as on the machine without this strange behavior.
We show that if we impose some—rather simple—conditions on the machine T then it will
no longer be essential which universal Turing machine is used for the definition of information
complexity. Roughly speaking, it is enough to assume that every input of a computation per-
formable on T can also be submitted as part of the program. To make this more exact, we
assume that there is a word (say, DATA) for which the following holds:
(a) Every one-tape Turing machine can be simulated by a program that does not contain the
word DATA as a subword;
(b) If the machine is started so that its program tape contains a word of the form xDATAy
where the word x does not contain the subword DATA, then the machine halts, if and
only if it halts when started with y written on the data tape and x on the program tape,
and in fact with the same output on the data tape,
It is easy to see that every universal Turing machine can be modified to satisfy the assump-
tions (a) and (b). In what follows, we will always assume that our universal Turing machine has
these properties.
(7.1) Lemma There is a constant cT (depending only on T ) such that KT (x) ≤ |x| + cT .

Proof T is universal, therefore the (trivial) one-tape Turing machine that does nothing (stops
immediately) can be simulated on it by a program p0 (not containing the word DATA). But
then, for every word x ∈ Σ∗0 , the program p0 DATAx will print the word x and stop. Thus the
constant cT = |p0 | + 4 satisfies the conditions.
In what follows we assume, to be specific, that cT ≤ 100.
(7.2) Remark We had to be a little careful since we did not want to restrict what symbols
can occur in the word x. In BASIC, for example, the instruction PRINT "x" is not good for
printing words x that contain the symbol ”. We are interested in knowing how concisely the
word x can be coded in the given alphabet, and we do not allow therefore the extension of
the alphabet. ♦
We prove now the basic theorem showing that the complexity (under the above conditions)
does not depend too much on the given machine.
(7.3) Invariance Theorem Let T and S be universal Turing machines satisfying the condi-
tions (a), (b). Then there is a constant cT S such that for every word x we have |KT (x)−KS (x)| ≤
cT S .

113
Proof We can simulate the two-tape Turing machine S by a one-tape Turing machine S0 in
such a way that if on S, a program q prints a word x then writing q on the single tape of S0 , it
also stops in finitely many steps, with x printed on its tape. Further, we can simulate the work
of Turing machine S0 on T by a program pS0 that does not contain the subword DATA.
Now let x be an arbitrary word from Σ∗0 and let qx be a shortest program printing x on
S. Consider the program pS0 DATAqx on T : this obviously prints x and has length only |qx | +
|pS0 | + 4. The inequality in the other direction is obtained similarly.
On the basis of this lemma, we will not restrict generality if we consider T fixed and do not
indicate the index T .
Unfortunately, the following theorem shows that the optimal code cannot be found algorith-
mically.

(7.4) Theorem The function K(x) is not recursive.

Proof The essence of the proof is a classical logical paradox, the so-called typewriter-paradox.
(This can be formulated simply as follows: let n be the smallest number that cannot be defined
with fewer than 100 symbols. We have just defined n with fewer than 100 symbols!)
Assume, by way of contradiction, that K(x) is computable. Let c be a natural number to be
chosen appropriately. Arrange the elements of Σ∗0 in increasing order, and let x(k) denote the
k-th word according to this ordering. Let x0 be the first word with K(x0 ) ≥ c. Assuming that
our language can be programmed in the language Pascal let us consider the following simple
program.

var k: integer;
function x(k : integer): integer;
..
.
function Kolm(k : integer): integer;
..
.
begin
k := 0;
while Kolm(k) < c do k := k + 1;
print(x(k));
end.

(The dotted parts stand for subroutines computing the given functions. The first is easy and
could be explicitely included. The second is hypothetical, based on the assumption that K(x)
is computable.)
This program obviously prints x0 . When determining its length we must take into account
the subroutines for the computation of the functions x(k) and Kolm(k) = K(x(k)) (where x(k)
is the k-th string); but this is a constant (independent of c). Thus even taken together, the
number of all these symbols is only log c + O(1). If we take c large enough this program consists
of fewer than c symbols and prints x0 , which is a contradiction.
As a simple application of the theorem, we get a new proof for the undecidability of the
halting problem. To this end, let’s ask the following question: Why is it not possible to compute

114
K(x) as follows? Take all words y in increasing order and check whether T prints x when started
with y on its program tape. Return the first y for which this happens; its length is K(x).
We know that something must be wrong here, since K(x) is not computable. The only
trouble with this algorithm is that T may never halt with some y. If the halting problem were
decidable, we could “weed out” in advance the programs on which T would work forever, and
not even try these. Thus we could compute K(x).
Thus the halting problem is not decidable.
In contrast to Theorem 7.4, we show that the complexity K(x) can be very well approximated
on the average.
For this, we must first make it precise what we mean by “on the average”. Assume that the
input words come from some probability distribution; in other words, every word x ∈ Σ∗0 has a
probability p(x). Thus

p(x) ≥ 0, p(x) = 1.
x∈Σ∗0

We assume that p(x) is computable, i.e., each p(x) is a rational number whose numerator and
denominator are computable from x. A simple example of a computable probability distribution
is p(xk ) = 2−k where xk is the k-th word in size order, or p(x) = (m + 1)−|x|−1 where m is the
alphabet size.

(7.5) Remark There is a more general notion of a computable probability distribution that
does not restrict probabilities to rational numbers; for example, {e−1 , 1 − e−1 } could also be
considered a computable probability distribution. Without going into details we remark that
our theorems would also hold for this more general class. ♦

(7.6) Theorem For every computable probability distribution there is an algorithm computing
a Kolmogorov code f (x) for every word x such that the expectation of |f (x)| − K(x) is finite.

Proof For simplicity of presentation, assume that p(x) > 0 for every word x. Let x1 , x2 , . . .
be an ordering of the words in Σ∗0 for which p(x1 ) ≥ p(x2 ) ≥ · · ·, and the words with equal
probability are, say, in increasing order (since each word has positive probability, for every x
there are only a finite number of words with probability at least p(x), and hence this is indeed
a single sequence).

(7.7) Lemma (a) Given a word x, the index i for which x = xi is computable.
(b) Given a natural number i, the word xi is computable.

Proof (a) Let y1 , y2 , . . . be all words arranged in increasing order. Given a words x, it is easy
to find the index j for which x = yj . Next, find the first k ≥ j for which

p(y1 ) + · · · + p(yk ) > 1 − p(yj ). (7.8)

Since the left-hand side converges to 1 while the right-hand side is less than 1, this will occur
sooner or later.

115
 Clearly each of the remaining words yk+1 , yk+2 , . . . has probability less than p(yj ), and
hence to determine the index of x = yj it suffices to order the finite set {y1 , . . . , yk } according
to decreasing p, and find the index of yj among them.
(b) Given an index i, we can compute the indices of y1 , y2 , . . . using (a) and wait until i
shows up.
Returning to the proof of the theorem, the program of the algorithm in the above lemma,
together with the number i, provides a Kolmogorov code f (xi ) for the word xi . We show that
this code satisfies the requirements of the theorem. Obviously, |f (x)| ≥ K(x). Furthermore,
the expected value of |f (x)| − K(x) is
∞

p(xi )(|f (xi )| − K(xi )).
i=1

We want to show that this sum is finite. Since its terms are non-negative, it suffices to show
that it partial sums remain bounded, i.e., that

N
p(xi )(|f (xi )| − K(xi )) < C
i=1

for some C independent of N . To the end, write


N
p(xi )(|f (xi )| − K(xi )) (7.9)
i=1

N 
N
= p(xi )(|f (xi )| − logm i) + p(xi )(logm i − K(xi )). (7.10)
i=1 i=1

We claim that both sums remain bounded. The difference |f (xi )| − logm i is just the length of
the program computing xi without the length of the parameter i, and hence it is an absolute
constant C. Thus the first sum in (7.9) is at most C.
To estimate the second term in (7.9), we use the following simple but useful principle. Let
a1 ≥ a2 ≥ . . . ≥ am be a decreasing sequence and let b1 , . . . , bm be an arbitrary sequence of real
numbers. Let b∗1 ≥ . . . ≥ b∗m be the sequence b ordered decreasingly, and let b∗∗ ∗∗
1 ≤ . . . ≤ bm be
the sequence b ordered increasingly. Then
  
ai b∗∗
i ≤ ai bi ≤ ai b∗i .
i i i

Let (z1 , z2 , . . . , ) be an ordering of the words so that K(z1 ) ≤ K(z2 ) ≤ . . . (we can’t compute
this ordering, but we don’t have to compute it). Then by the above principle,

N 
N
p(xi )K(xi ) ≥ p(xi )K(zi ).
i=1 i=1

The number of words x with K(x) = k is at most mk , and hence the number of words x with
K(x) ≤ k is at most 1 + m + . . . mk < mk+1 . This is the same as saying that
i ≤ mK(zi )+1 ,

116
and hence

K(zi ) ≥ logm i − 1.

Thus

N 
N 
N
p(xi )(logm i − K(xi )) ≤ p(xi )(logm i − K(zi )) ≤ p(xi ) = 1.
i=1 i=1 i=1

This proves the theorem.

7.3 The notion of a random sequence

In this section, we assume that Σ0 = {0, 1}, i.e., we will consider only the complexity of 0-1
sequences. Roughly speaking, we want to consider a sequence random if there is no “regularity”
in it. Here, we want to be as general as possible and consider any kind of regularity that would
enable a more economical coding of the sequence (so that the complexity of the sequence would
be small).

(7.1) Remark Note that this is not the only possible idea of regularity. One might consider
a 0-1-sequence regular if the number of 0’s in it is about the same as the number of 1’s. This
kind of regularity is compatible with (in fact implied by) randomness: we should really consider
only regularities that are shared only by a small minority of the sequences. ♦

Let us estimate first the complexity of “average” 0-1 sequences.

(7.2) Lemma The number of 0-1 sequences x of length n with K(x) ≤ n − k is less than
2n−k+1 .

Proof The number of “codewords” of length at most n−k is at most 1+2+· · ·+2n−k < 2n−k+1 ,
hence only fewer than 2n−k+1 strings x can have such a code.

(7.3) Corollary The complexity of 99% of the n-digit 0-1 sequences is greater than n − 7. If
we choose a 0-1 sequence of length n randomly then |K(x) − n| ≤ 100 with probability 1 − 2100 .

Another corollary of this simple lemma is that it shows, in a certain sense, a “counterex-
ample” to Church’s Thesis, as we noted in the introduction to the section on randomized com-
putation. Consider the following problem: For a given n, construct a 0-1 sequence of length n
whose Kolmogorov complexity is greater than n/2. According to the exercise mentioned after
Theorem 7.4, this problem is algorithmically undecidable. On the other hand, the above lemma
shows that with large probability, a randomly chosen sequence is appropriate.
According to Theorem 7.4, it is algorithmically impossible to find the best code. There are,
however, some easily recognizable properties indicating about a word that it is codable more
efficiently than its length. The next lemma shows such a property:

117
(7.4) Lemma If the number of 1’s in a 0 − 1-sequence x of length n is k then
 
n
K(x) ≤ log2 + log2 n + log2 k + O(1).
k

Let k = pn (0 < p < 1), then this can be estimated as

K(x) ≤ (−p log p − (1 − p) log(1 − p))n + O(log n).

In particular, if k > (1/2 + ε)n or k < (1/2 − ε)n then

K(x) ≤ cn + O(log n)

where c = −(1/2 + ε) · log(1/2 + ε) − (1/2 − ε) · log(1/2 − ε) is a positive constant smaller than

1 and depending only on ε.

Proof x can be described as the “lexicographically t-th sequence among all sequences of length
 
n containing exactly k 1’s”. Since the number of sequences of
length n containing k 1’s is nk ,
the description of the numbers t, n and k needs only log2 nk + 2 log2 n + 2 log2 k bits. (Here,
the factor 2 is due to the need to separate the three pieces of information from each other; we
leave it to the reader to find the trick.) The program choosing the appropriate sequence needs
only a constant number of bits.
The estimate of the binomial coefficient is done by a method familiar from probability theory.

On the basis of the above, one can consider either |x| − K(x) or |x|/K(x) as a measure
of the randomness (or, rather, non-randomness) of the word x. the larger are these numbers,
the smaller is K(x) relative to |x|, which means that x has more “regularity” and so it is less
random.
In case of infinite sequences, a sharper difference can be made: we can define whether a given
sequence is random. Several definitions are possible; we introduce here the simplest version.
Let x be an infinite 0–1 sequence, and let xn denote its starting segment formed by the first n
elements. We call the sequence x informatically random if K(xn )/n → 1 when n → ∞.
It can be shown that every informatically weakly random sequence satisfies the laws of large
numbers. We consider here only the simplest such result. Let an denote the number of 1’s in
the string xn , then the previous lemma immediately implies the following theorem:

(7.5) Theorem If x is informatically random then an /n → 1/2 (n → ∞).

The question arises whether the definition of an algorithmically random sequence is not too
strict, whether there are any algorithmically random infinite sequences at all. Let us show that
not only there are such sequences but that almost all sequences have this property:

(7.6) Theorem Let the elements of an infinite 0-1 sequence x be 0’s or 1’s, independently
from each other, with probability 1/2. Then x is algorithmically random with probability 1.

118
Proof For a fixed > 0, let S be the set of all those finite sequences y for which K(y) <
(1 − )|y|, and let An denote the event xn ∈ S. Then by lemma 7.2,

Prob(An ) ≤ 2−n < 2(1−)n + 12−n = 21−n ,
y∈S

and hence the sum ∞ k=1 Prob(An ) is convergent. But then, the Borel-Cantelli Lemma in
probability theory implies that with probability 1, only finitely many of the events Ak occur.
But this just means that K(xn )/n → ∞.

7.4 Kolmogorov complexity and data compression

Let L ⊂ Σ∗0 be a recursive language and suppose that we want to find a short program, “code”,
only for the words in L. For each word x in L, we are thus looking for a program f (x) ∈ {0, 1}∗
printing it. We call the function f : L → Σ∗ a Kolmogorov code of L. The conciseness of
the code is the function
η(n) = max{ |f (x)| : x ∈ L, |x| ≤ n }.
We can easily get a lower bound on the conciseness of any Kolmogorov code of any language.
Let Ln denote the set of words of L of length at most n. Then obviously,
η(n) ≥ log2 |Ln |.
We call this estimate the information theoretical lower bound.
This lower bound is sharp (to within an additive constant). We can code every word x in L
simply by telling its serial number in the increasing ordering. If the word x of length n is the
t-th element then this requires log2 t ≤ log2 |Ln | bits, plus a constant number of additional bits
(the program for taking the elements of Σ∗ in lexicographic order, checking their membership
in L and printing the t-th one).
We arrive at more interesting questions if we stipulate that the code from the word and,
conversely, the word from the code should be polynomially computable. In other words: we are
looking for a language L and two polynomially computable functions:
f : L → L , g : L → L
with g ◦ f = idL for which, for every x in L the code |f (x)| is “short” compared to |x|. Such
a pair of functions is called a polynomial-time code. (Instead of the polynomial time bound
we could, of course, consider other complexity restrictions.)
We present some examples when a polynomial-time code approaches the information-
theoretical bound.

(7.1) Example In the proof of Lemma 7.4, for the coding of the 0-1 sequences of length n
with exactly m 1’s, we used the simple coding in which the code of a sequence is the number
giving its place in the lexicographic ordering. We will show that this coding is polynomial.
Let us view each 0-1 sequence as the obvious code of a subset of the n-element set {n −
1, n − 2, . . . , 0}. Each such set can be written as {a1 , . . . , am } with a1 > a2 > · · · > am . Then
the set {b1 , . . . , bm } precedes the set {a1 , . . . , am } lexicografically if and only if there is an i such

119
that bi < ai while aj = bj holds for all j < i. Let {a1 , . . . , am }, be the lexicographically
 ai 
t-th set.
Then the number of subsets {b1 , . . . , bn } with this property is exactly m−i+1 . Summing this
for all i we find that
     
a1 a2 am
t=1+ + + ··· + . (7.2)
m m−1 1

Given a1 , . . . , am , the value of t is easily computable in time polynomial in n. Conversely,

n
if t < m is given then t is easy to write in the above form: first we find, using binary
search,
 a2 
the greatest natural number a1 with am1 ≤ t − 1, then the greatest number a2 with
a1 
m−1 ≤ t − 1 − m , etc. We do this for m steps. The numbers obtained this way satisfy
a1 +1 a1   a1 
a1 > a2 >· · ·; indeed, according
a1 
to the definition of a1 we have m = m + m−1 > t−1 and
a1
therefore m−1 > t − 1 − m implying a1 > a2 . It follows similarly that a2 > a3 > . . . > am ≥ 0
and that there is no “remainder” after m steps, i.e., that 7.2 holds. It can therefore be determined
in polynomial time which subset is lexicographically the t-th. ♦

(7.3) Example Consider trees, given by their adjacency matrices (but any other “reasonable”
representation would also do). In such representations, the vertices of the tree have a given
order, which we can also express saying that the vertices of the tree are labeled by numbers
from 0 to (n − 1). We consider two trees equal if whenever the nodes i, j are connected in the
first one they are also connected in the second one and vice versa (so, if we renumber the nodes
of the tree then we may arrive at a different tree). Such trees are called labeled trees. Let
us first see what does the information-theoretical lower bound give us, i.e., how many trees are
there. The following classical result, called Cayley’s Theorem, applies here:

(7.4) Theorem The number of n-node labeled trees is nn−2 .

Consequently, by the information-theoretical lower bound, for any encoding of trees some
n-node tree needs a code with length at least log(nn−2 ) = (n − 2) log n. But can this lower
bound be achieved by a polynomial-time computable code?

(a) Coding trees by their adjacency matrices takes n2 bits.

(b) We fare better if we specify each tree by enumerating its edges. Then we must give a
“name” to each vertex; since there are n vertices we can give to each one a 0-1 sequence
of length log n as its name. We specify each edge by its two endnodes. In this way, the
enumeration of the edges takes cca. 2(n − 1) log2 n bits.

(c) We can save a factor of 2 in (b) if we distinguish a root in the tree, say the node 0, and
we specify the tree by the sequence (α(1), . . . , α(n − 1)) in which α(i) is the first interior
node on the path from node i to the root (the “father” of i). This is (n − 1)log n bits,
which is already nearly optimal.

(d) There is, however, a procedure, the so-called Prüfer code, that sets up a bijection between
the n-node labeled trees and the sequences of length n − 2 of the numbers 0, . . . , n − 1.
(Thereby it also proves Cayley’s theorem). Each such sequence can be considered the

120
 expression of a natural number in the base n number system; in this way, we order a
“serial number” between 0 and nn−2 to the n-node labeled trees. Expressing these serial
numbers in the base two number system, we get a coding in which the code of each number
has length at most (n − 2) log n.

The Prüfer code can be considered a refinement of the procedure (c). The idea is that we
order the edges [i, α(i)] not by the magnitude of i but a little differently. Let us define the
permutation (i1 , . . . , in ) as follows: let i1 be the smallest endnode (leaf) of the tree; if i1 , . . . , ik
are already defined then let ik+1 be the smallest endnode of the graph remaining after deleting
the nodes i1 , . . . , ik . (We do not consider the root 0 an endnode.) Let in = 0. With the ik ’s
thus defined, let us consider the sequence (α(i1 ), . . . , α(in−1 )). The last element of this is 0 (the
“father” of the node in−1 can namely be only in ), it is therefore not interesting. We call the
remaining sequence (α(i1 ), . . . , α(in−2 )) the Prüfer code of the tree.

(7.5) Claim The Prüfer code of a tree determines the tree.

For this, it is enough to see that the Prüfer code determines the sequence i1 , . . . , in ; then we
know all the edges of the tree (the pairs [i, α(i)]).
The node i1 is the smallest endnode of the tree; hence to determine i1 , it is enough to figure
out the endnodes from the Prüfer code. But this is obvious: the endnodes are exactly those
that are not the “fathers” of other nodes, i.e., the ones that do not occur among the numbers
α(i1 ), . . . , α(in−1 ), 0. The node i1 is therefore uniquely determined.
Assume that we know already that the Prüfer code uniquely determines i1 , . . . , ik−1 . It
follows similarly to the above that ik is the smallest number not occurring neither among
i1 , . . . , ik−1 nor among α(ik ), . . . , α(in−1 ). So, ik is also uniquely determined.

(7.6) Claim Every sequence (b1 , . . . , bn−2 ), where 1 ≤ bi ≤ n, occurs as the Prüfer code of
some tree.

Using the idea of the proof above, let bn−1 = 0 and let us define the permutation i1 , . . . , in by
the recursion that ik is the smallest number not occurring neither among i1 , . . . , ik−1 nor among
bk , . . . , bn−1 , where (1 ≤ k ≤ n − 1); and let in = 0. Connect ik with bk for all 1 ≤ k ≤ n − 1
and let γ(ik ) = bk . In this way, we obtain a graph G with n − 1 edges on the nodes 1, . . . , n.
This graph is connected, since for every i the γ(i) comes later in the sequence i1 , . . . , in than i
and therefore the sequence i, γ(i), γ(γ(i)), . . . is a path connecting i to the node 0. But then G
is a connected graph with n − 1 edges, therefore it is a tree. That the sequence (b1 , . . . , bn−2 ) is
the Prüfer code of G is obvious from the construction. ♦

(7.7) Remark An exact correspondence like the Prüfer code has other advantages besides
optimal Kolmogorov coding. Suppose that our task is to write a program for a randomized
Turing machine that outputs a random labeled tree of size n in such a way that all trees occur
with the same probability. The Prüfer code gives an efficient algorithm for this. We just have
to generate randomly a sequence b1 , . . . , bn−2 , which is easy, and then decode from it the tree
by the above algorithm. ♦

121
(7.8) Example Consider now the unlabeled trees. These can be defined as the equivalence
classes of labeled trees where two labeled trees are considered equivalent if they are isomorphic,
i.e., by a suitable relabeling, they become the same labeled tree. We assume that we represent
each equivalence class by one of its elements, i.e., by a labeled tree (it is not interesting now,
by which one). Since each labeled tree can be labeled in at most n! ways (its labelings are
not necessarily all different as labeled trees!) therefore the number of unlabeled trees is at
least nn−2 /n! ≤ 2n−2 . (According to a difficult result of George Pólya, the number of n-node
unlabeled trees is asymptotically c1 cn2 n3/2 where c1 and c2 are constants defined in a certain
complicated way.) The information-theoretical lower bound is therefore at least n − 2.
On the other hand, we can use the following coding procedure. Consider an n-node tree F .
Walk through F by the “depth-first search” rule: Let x0 be the node labeled 0 and define the
nodes x1 , x2 , . . . as follows: if xi has a neighbor that does not occur yet in the sequence then let
xi+1 be the smallest one among these. If it has not and xi = x0 then let xi+1 be the neighbor
of xi on the path leading from xi to x0 . Finally, if xi = x0 and every neighbor of x0 occured
already in the sequence then we stop.
It is easy to see that for the sequence thus defined, every edge occurs among the pairs
[xi , xi+1 ], moreover, it occurs once in both directions. It follows that the length of the sequence
is exactly 2n − 1. Let now εi = 1 if xi+1 is farther from the root than xi and εi = 0 otherwise.
It is easy to understand that the sequence ε0 ε1 · · · ε2n−3 determines the tree uniquely; passing
trough the sequence, we can draw the graph and construct the sequence x1 , . . . , xi of nodes
step-for-step. In step (i + 1), if εi = 1 then we take a new node (this will be xi+1 ) and connect
it with xi ; if εi = 0 then let xi+1 be the neighbor of xi in the “direction” of x0 . ♦

(7.9) Remarks

1. With this coding, the code assigned to a tree depends on the labeling but it does not
determine it uniquely (it only determines the unlabeled tree uniquely).

2. The coding is not bijective: not every 0-1 sequence will be the code of an unlabeled tree.
We can notice that

(a) There are as many 1’s as 0’s in each tree;

(b) In every starting segment of every code, there are at least as many 1’s as 0’s

(the difference between the number of 1’s and the number of 0’s among the first i numbers
gives the distance of the node xi from the node 0). It is easy to see that for each 0-1
sequence having the properties (a) − (b), there is a labeled tree whose code it is. It is not
sure, however, that this tree, as an unlabeled tree, is given with just this labeling (this
depends on which unlabeled trees are represented by which of their labelings). Therefore
the code does not even use all the words with properties (a) − (b).

3. The number of 0-1 sequences


having properties (a) − (b) is, according to a well-known

1 2n−2
combinatorial theorem, n n−1 (the so-called catalan number). We can formulate a tree
notion to which the sequences with properties (a) − (b) correspond exactly: these are the
rooted planar trees, which are drawn without intersection into the plane in such a way
that their distinguished vertex—their root—is on the left edge of the page. This drawing

122
 defines an ordering among the “sons” (neighbors farther from the root) “from the top to
the bottom”; the drawing is characterized by these orderings. The above described coding
can also be done in rooted planar trees and creates a bijection between them and the
sequences with the properties (a) − (b).

123
8 Pseudo-random numbers
8.1 Introduction
We have seen that various important algorithms use random numbers (or, equivalently, inde-
pendent random bits). But how do we get such bits?
One possible source is from outside the computer. We could obtain “real” random sequences,
say, from radioactive decay. In most cases, however, this would not work: our computers are
very fast and we have no physical device giving the equivalent of unbiased coin-tosses at this
rate.
Thus we have to resort to generating our random bits by the computer. However, a long
sequence generated by a short program is never random, according to the notion of randomness
introduced using information complexity. Thus we are forced to use algorithms that generate
random-looking sequences; but, as Von Neumann (one of the first mathematicians to propose
the use of these) put it, everybody using them is inevitably “in the state of sin”. In this chapter,
we will understand the kind of protection we can get against the graver consequences of this
sin.
There are other reasons besides practical ones to study pseudorandom number generators.
We often want to repeat some computation for various reasons, including error checking. In this
case, if our source of random numbers was really random, then the only way to use the same
random numbers again is to store them, using a lot of space. With pseudorandom numbers, this
is not the case: we only have to store the “seed”, which is much shorter. Another, and more
important, reason is that there are applications where what we want is only that the sequence
should “look random” to somebody who does not know how it was generated. The collection
of these applications is called cryptography, to be treated in a later chapter.
The way a pseudo-random bit generator works is that it turns a short random string called
the “seed” into a longer pseudorandom string. We require that it works in polynomial time.
The resulting string has to “look” random: and the important fact is that this can be defined
exactly. Roughly speaking, there should be no polynomial time algorithm that distinguishes it
from a truly random sequence. Another feature, often easier to verify, is that no algorithm can
predict any of its bits from the previous bits. We prove the equivalence of these two conditions.

8.2 Introduction
We have seen that various important algorithms use random numbers (or, equivalently, inde-
pendent random bits). But how do we get such bits?
One possible source is from outside the computer. We could obtain “real” random sequences,
say, from radioactive decay. In most cases, however, this would not work: our computers are
very fast and we have no physical device giving the equivalent of unbiased coin-tosses at this
rate.
Thus we have to resort to generating our random bits by the computer. However, a long
sequence generated by a short program is never random, according to the notion of randomness
introduced using information complexity. Thus we are forced to use algorithms that generate
random-looking sequences; but, as Von Neumann (one of the first mathematicians to propose
the use of these) put it, everybody using them is inevitably “in the state of sin”. In this chapter,

124
we will understand the kind of protection we can get against the graver consequences of this
sin.
There are other reasons besides practical ones to study pseudorandom number generators.
We often want to repeat some computation for various reasons, including error checking. In this
case, if our source of random numbers was really random, then the only way to use the same
random numbers again is to store them, using a lot of space. With pseudorandom numbers, this
is not the case: we only have to store the “seed”, which is much shorter. Another, and more
important, reason is that there are applications where what we want is only that the sequence
should “look random” to somebody who does not know how it was generated. The collection
of these applications is called cryptography, to be treated in a later chapter.
The way a pseudo-random bit generator works is that it turns a short random string called
the “seed” into a longer pseudorandom string. We require that it works in polynomial time.
The resulting string has to “look” random: and the important fact is that this can be defined
exactly. Roughly speaking, there should be no polynomial time algorithm that distinguishes it
from a truly random sequence. Another feature, often easier to verify, is that no algorithm can
predict any of its bits from the previous bits. We prove the equivalence of these two conditions.
But how do we design such a generator? Various ad hoc methods that produce random-
looking sequences (like taking the bits in the binary representation of a root of a given equation)
turn out to produce strings that do not pass the strict criteria we impose. A general method
to obtain such sequences is based on one-way functions: functions that are easy to evaluate
but difficult to invert. While the existence of such functions is not proved (it would imply that
P is different from NP), there are several candidates, that are secure at least against current
techniques.

8.3 Classical methods

There are several classical methods that generate a “random-looking” sequence of bits. None of
these meets the strict standards to be formulated in the next section; but due to their simplicity
and efficiency, they (especially linear congruential generators, example 2 below) can be used
well in practice. There is a large amount of practical information about the best choice of the
parameters; we don;’t go into this here, but refer to Volume 2 of Knuth’s book.
(8.1) Example Shift registers are defined as follows. Let f : {0, 1}n → {0, 1} be a func-
tion that is easy to compute. Starting with a seed of n bits a0 , a1 , . . . , an−1 , we compute bits
an , an+1 , an+2 , . . . recursively, by
ak = f (ak−1 , ak−2 , . . . , ak−n ).
The name “shift register” comes from the fact that we only need to store n + 1 bits: after
storing f (a0 , . . . , an−1 ) in an , we don’t need a0 any more, and we can shift a1 to a0 , a2 to a1 ,
etc. The most important special case is when f is a linear function over the 2-element field, and
we’ll restrict ourselves to this case.
Looking at particular instances, the bits generated by a linear shift register look random, at
least for a while. Of course, the sequence a0 .a1 , . . . , will eventually have some n-tuple repeated,
and then it will be periodic from then on; but this need not happen sooner than a2n , and indeed
one can select the (linear) function f so that the period of the sequence is as large as 2n .

125
 The problem is that the sequence has more hidden structure than just perodicity. Indeed,
let
f (x0 , dots, xn−1 ) = b0 x0 + b1 x1 + . . . bn−1 xn−1
(where bi ∈ {0, 1}). Assume that we we don’t know the coefficients b0 , . . . , bn−1 , but observe the
first n bits an , . . . , a2n−1 of the output sequence. Then we have the following system of linear
equations to determine the bi :

b0 a0 + b1 a1 + . . . bn−1 an−1 = an
b0 a1 + b1 a2 + . . . bn−1 an = an+1
..
.
b0 an−1 + b1 an + . . . bn a2n−2 = a2n−1

Here are n equations to determine these n unknowns (the equations are over the 2-element field).
Once we have the bi , we can predict all the remaining elements of the sequence a2n , a2n+1 , . . .
It may happen, of course, that this system is not uniquely solvable, because the equations are
dependent. For example, we might start with the seed 00 . . . 0, in which case the equations are
meaningless. But it can be shown that for a random choice of the seed, the equations determine
the coefficients bi with positive probability. So after seeing the first 2n elements of the sequence,
the rest “does not look random” for an observer who is willing to perform a relatively simple
(polynomial time) computation. ♦

(8.2) Example The most important pseudorandom number generators in practice are linear
congruential generators. Such a generator is given by three parameters a, b and m, which are
positive integers. Starting with a seed X0 , which is an integer in the range 0 ≤ X0 ≤ m − 1,
the generator recursively computes integers X1 , X2 , . . . by

Xi = aXi−1 + b (mod m).

One might use all the Xi or extract, say, the middle bit of each, and output this sequence.
It turns out that the output of these generators can also be predicted by a polynomial-time
computation, after observing a polynomial number of output bits. The algorithms to do so
are much more involved, however, and due to their fastness and simplicity, linear congruential
generators are good for most practical applications. ♦
√
(8.3) Example As a third example, let us look at the binary expansion of, say, 5:
√
5 = 10.001111000110111 . . .

This sequence looks rather random. Of course, we cannot use the same number all the time;
√ √
but we can pick, say, an n-bit integer a as our “seed”, and output the bits of a −  a.
Unfortunately, this method turns out to be “breakable” by rather advanced (but polynomial-
time) methods from algorithmic number theory. ♦

126
8.4 The notion of a psuedorandom number generator
In general, a pseudo-random bit generator transforms a short, truly random sequence s (the
“seed”) into a longer sequence g(s) that still “looks” random. The success of using g(s) in place of
a random sequence depends on how severely the randomness of g(s) is tested by the application.
If the application has the ability to test all possible seeds that might have generated g(s) then it
finds the true seed and not much randomness remains. For this, however, the application may
have to run too long. We would like to call g a pseudo-random bit generator if no applications
running only in polynomial time can distinguish g(s) from truly random strings.
To make the defition precise, we need some preparation. We say that a function f : ZZ+ → IR
is negligible, if nk f (n) → 0 as n → ∞ for each fixed k. In other words, f tends to 0 faster than
the reciprocal of any polynomial. It will be convenient to denote this (analogously to the “big-O”
notation), by

f (n) = NEGL(n).

Note that a polynomial multiple of a negligible function is still negligible; thus

nr NEGL(n) = NEGL(n)

for each fixed r.

Consider a polynomial time computable function G : {0, 1}∗ → {0, 1}∗ , where we assume
that |G(x)| depends only on |x| and |x| ≤ |G(x)| < |x|c for some constant c. We call such a
function a generator. Let A be a randomized polynomial time algorithm (Turing machine) that
accepts any 0-1 string x as input and computes a bit A(x) from it. (We will interpret 0 as an
output as “not random”, and 1 as an output as “random”.) Fix an n ≥ 1. Let x be chosen
uniformly from {0, 1}n and let y be chosen uniformly from {0, 1}N , where N = |G(x)|. We flip
a coin, and depending on its result, we either feed G(x) or y to A. We say that A is successful
if either G(x) was fed to A and it output 0 or y was fed and the output is 1.
The generator G is called a (safe) random number generator if for every randomized poly-
nomial time algorithm A that takes a 0-1 string x as input and computes a bit A(x) from it,
the probability that A is successful is at most 1/2 + NEGL(n).
Condition (b) says that G(x) passes every “reasonable” test (any test computable in ran-
domized polynomial time) in the sense that the probability that such a test recognizes that
G(x) is not truely random is only negligibly larger than 1/2 (which can of course be achieved
by guessing randomly). The probability in (b) is over the random choice of x and y, over the
coin flip determining which one is fed to A, and over the internal coin flips of A.
This requirement is so strong that it is unknown whether safe random number generators
exist at all (if they do, then P = NP; see the exercises). But we will see in the next section that
they exist under some complexity-theoretic assumptions.
Condition (b) is very general and it is difficult to verify. The following theorem of Yao
provides a way to establish that a function is a safe random number generator that is often
more convenient. What it says is that every bit of G(x) is highly unpredictable from the
previous bits, as long as the prediction algorithm does not use too much time.

127
 We say that a generator g is unpredictible if the following holds. Let n ≥ 1 and let x be
chosen uniformly from {0, 1}n . Let g(x) = G1 G2 . . . GN . For every randomized polynomial time
algorithm A that accepts a string x ∈ {0, 1}i (i ≤ N ) as input and computes a bit from it,
  1
max P A(G1 . . . Gi ) = Gi+1 = + NEGL(n). (8.1)
i 2
Informally: we try to use A to predict each bit of g1 . . . gN from the previous bits. Then
at any stage, the probability that we suceed is only negligibly larger than 1/2 (which we could
achive just by guessing randomly).

(8.2) Theorem [A. Yao] A generator g is a safe random number generator if and only if it is
unpredictible.

Proof I. Suppose that g is no unpredictable. Then there is a randomized polynomial time

algorithm A, a constant k > 0, there are infinitely many values of n, and for each such value a
certain i < N , such that
  1 1
P A(G1 . . . Gi ) = Gi+1 > + k.
2 n
(where x ∈ {0, 1}n is a uniformly chosen random string and g(x) = G1 . . . GN ).
But then we can perform the following randomness test B on strings y = y1 . . . yN : if
A(y1 . . . yi ) = yi+1 we declare y “non-random”, else we declare it “random”. Then if we give B
either a truly random string R1 . . . RN or g(x) (each with probability 1/2), the probability of
success is
1   1  
P A(R1 . . . Ri ) = Ri+1 + P A(G1 . . . Gi ) = Gi+1
2   2
11 1 1 1 1 1
≥ + + k = + k.
22 2 2 n 2 2n
Since this is non-negligibly larger than 1/2, the generator is not safe.
II. Assume that there exists an algorithm A that, for infinitely many values of n, distinguishes
the pseudorandom sequence g(x) = G1 . . . GN from a truly random sequence r = R1 . . . RN with
a success probability that is at least 1/2 + n−k for some constant k > 0. We want to show that
in this case we can predict some bit of G1 . . . GN .
The success probability of A is
1 1
P(A(r) = 1) + P(A(g(x)) = 0)
2 2
(since success means that A accepts r, if r is given to it, but rejects g(x), if g(x) is given to it).
In other words, this probability is
1 1 
+ P(A(r) = 1) − P(A(g(x)) = 1) .
2 2
So it follows that
2
P(A(r) = 1) − P(A(g(x)) = 1) > . (8.3)
nc

128
 The trick is to consider the mixed sequences

y i = G1 . . . Gi Ri+1 . . . RN

and subject them to A. We have y 0 = r and y N = g(x), and hence

2
P(A(y 0 ) = 1) − P(A(y N ) = 1) > .
nk
Hence it follows that there is an index i, 1 ≤ i ≤ N , such that
2
P(A(y i−1 ) = 1) − P(A(y i ) = 1) > . (8.4)
N nk
We are going to show that this i is predictible with success probability non-negligibly larger
than 1/2. The prediction algorithm is simple: we guess that

X = Ri+1 ⊕ A(y i )

is the value of Gi+1 (this value is indeed computed by a randomized polynomial time algorithm
from G1 , . . . , Gi ).
For simplicity of notation, write G = Gi , R = Ri , and

A0 = A(G1 . . . Gi−1 0Ri+1 . . . RN ), A1 = A(G1 . . . Gi−1 1Ri+1 . . . RN ).

Note that G, X, R, A0 and A1 are random bits. These are not independent in general, but R
is independent of the rest. Moreover, it is easy to see that we can write

X = R(1 − A1 ) + (1 − R)A0

Hence G = X (in other words, our prediction is successful) is equivalent to saying that

G(R(1 − A1 ) + (1 − R)A0 ) + (1 − G)(1 − R(1 − A1 ) − (1 − R)A0 ) = 1,

or

GR(1 − A1 ) + G(1 − R)A0 + (1 − G)

− (1 − G)R(1 − A1 ) − (1 − G)(1 − R)A0 = 1.

Since the left hand side is always 0 or 1, the probability that this happens is the expectation of
the left hand side. But this is

E GR(1 − A1 ) + G(1 − R)A0 + (1 − G) − (1 − G)R(1 − A1 )

− (1 − G)(1 − R)A0

= E(GR(1 − A1 )) + E(G(1 − R)A0 )

+ E(1 − G) − E((1 − G)R(1 − A1 ))
− E((1 − G)(1 − R)A0 ).

129
Since R is independent of the other variables, we can pull out E(R) = 1/2, and get
1 
E(G(1 − A1 )) + E(GA0 ) + 2E(1 − G) − E((1 − G)(1 − A1 )) − E((1 − G)A0 )
2
1
= E(1 + 2GA0 − 2GA1 − A0 + A1 ).
2

We can also express (8.4) in terms of these bits. In fact, A(y i ) = RA1 + (1 − R)A0 , and
hence
1
P(A(y i ) = 1) = E(A0 + A1 ).
2
Similarly, A(y i+1 ) = GA1 + (1 − G)A0 , and hence

P(A(y i+1 ) = 1) = E(GA1 + (1 − G)A0 ).

Thus (8.4) implies that

1 1 2
E(A0 + A1 − 2GA1 − 2(1 − G)A0 ) = E(A1 − A0 − 2GA1 + 2GA0 ) > .
2 2 N nk
But then
1 1 1 1 1
P(G = X) = E(1 + 2GA0 − 2GA1 − A0 + A1 ) > + > + k+c ,
2 2 N nk 2 n
which shows that our prediction is non-negligibly better than the trivial.
The existence of pseudo-random bit generators can be derived from some (unproved) com-
plexity theoretic assumptions that are nevertheless rather plausible. These are discussed in the
next section.

8.5 One-way functions

A one-way function is a function that is easy to compute but difficult to invert. The exact
definition is as follows.

(8.1) Definition A function f : {0, 1}∗ → {0, 1}∗ is called a one-way function if

(a) there is a constant c ≥ 1 such that |x|1/c < |f (x)| < |x|c ;

(b) f (x) is polynomial time computable;

(c) for every randomized polynomial time algorithm A that computes 0 − 1 strings from 0 − 1
strings, and for a string y randomly and uniformly chosen from {0, 1}n ,

P(f (A(f (y))) = f (y)) = N EG(n). (8.2)

130
 The third condition needs explanation. We can read it as follows: if we pick a random
string y of length n, compute f (y), and then try to use A to compute a pre-image of f (y), the
probability that we succeed is negligible. Note that we don’t assume that f is invertible, so we
cannot simply write A(f (y)) = y.
But why don’t we write simply
P(f (A(z)) = z) = N EG(n) (8.3)
for a uniformly chosen z? The point is that since f may not be onto, it could be the case that
most strings z are not in the range of f . then the above probability would be small, even if
in the cases when z is in the range, a pre-image could always be easily computed. Thus (8.2)
concentrates on the cases when a pre-image exists, and stipulates that even these are hard.
A one-way permutation is a one-way function that is one-to-one and satisfies |f (x)| = |x| for
every x. It is clear that under this assumption, (8.3) is equivalent to (8.2).

(8.4) Theorem Let g be a safe random number generator, and assume that |g(x)| ≥ 2|x|.
Then g is one-way.

Proof Suppose that g is not one-way, then there exists a constant k > 0, a randomized
polynomial time algorithm A, and infinitely many values n such that for a string y randomly
and uniformly chosen from {0, 1}n ,
1
P(g(A(g(y))) = g(y)) > .
nk
Now consider the following randomness test B: we declare a string z ∈ {0, 1}N “non-random” if
g(A(z)) = z, and “random” otherwise. If we give B either a truly random string r = R1 . . . RN
or g(x) (each with probability 1/2), the probability of success is
1 1
P(g(A(r)) = r) + P(g(A(g(x))) = g(x))
2 2
The first term is very close to 1/2; indeed, the total number of strings in the range of g is 2n ,
so the probability that r is one of these os 2n−N < 2−n (and of course even if r is in the range
of g, it may not be equal to g(A(r)), which would help us but we don’t have to use it). The
second term is at least 1/nk by assumption. thus the probability of success is at least
 
1 1 1 1 1 1
1− n + k
> + k,
2 2 2n 2 4n
which is non-negligibly larger than 1/2.
In the reverse direction we describe how to use a one-way permutation f to construct a safe
random number generator. For this, we need the following lemma due to Goldreich and Levin.
For two 0-1 strings u = u1 . . . un and v = v1 . . . vn , define b(u, v) = u1 v1 ⊕ . . . ⊕ un vn .

(8.5) Lemma Let f be a one-way permutation. Then for every randomized polynomial time
algorithm A, and for x, y chosen randomly and uniformly from {0, 1}n ,
  1
P A(x, y) = b(f −1 x, y) = + NEGL(n).
2

131
 Informally, selecting a random subset of the bits of f −1 , their sum modulo 2 cannot be
computed with success probability non-negligibly larger than 1/2. the proof of this lemma is
quite involved, and we have to state it without proof.

(8.6) Theorem Suppose that a one-way permutation exists. Then there exists a safe random
number generator.

Proof We describe a random number generator. Let f be a one-way permutation and k > 0 a
constant. We stretch a random sequence x = (x1 , . . . , x2n )of length 2n (the parity of the length
is even for convenience) to a pseudorandom sequence of length N = nk as follows. Compute the
strings
y1t . . . ynt = f t (x1 , . . . , xn )
for t = 1, . . . , N (here f t is the function f iterated t times), and let
gt = y1t xn+1 ⊕ . . . ⊕ ynt x2n .
We show that g(x) = gN gN −1 . . . g1 defines a safe random number generator (note the trick of
writing the bits in the reverse order!). Using theorem 8.2, it suffices to show that for every 1 ≤
i ≤ N , and for every randomized polynomial time algorithm A computing a bit from gN . . . gi+1 ,
the probability that this bit is gi is only negligibly larger than 1/2. We can be generous and
give the algorithm not only these bits but all the strings f t (x1 , . . . , xn ) for t ≥ i + 1 and also the
string xn+1 . . . x2n (from which gN . . . gi+1 is easily computed). But then, we don’t have to give
f i+2 (x1 , . . . , xn ) . . . , f N (x1 , . . . , xn ); these are easily computed from f i+1 (x1 , . . . , xn ). Thus we
may assume that the algorithm gets z = f i+1 (x1 , . . . , xn ) and u = xn+1 . . . x2n , and has to guess
gi = b(f −1 (z), u). But by lemma 8.5, the probability of succeeding is only negligibly larger than
1/2, as claimed.

8.5.1 Candidates for one-way functions

Number theory provides several candidates of one-way functions. The length of inputs and
outputs will not be exactly n, only polynomial in n.

The factoring problem. Let x represent a pair of primes of length n (say, along with a proof
of their primality). Let f (n, x) be their product. Many special cases of this problem are
solvable in polynomial time but the still, a large fraction of the instances remains difficult.

The discrete logarithm problem. Given a prime number p, a primitive root g for p and a
positive integer i < p, we output p, g and y = g i mod p. The inversion problem for this is
called the discrete logarithm problem since given p, g, y, what we are looking for is i which
is also known as the index, of discrete logarithm, of y with respect to p.

The discrete square root problem. Given positive integers m and x < m, the function
outputs m and y = x2 mod m. The inversion problem is to find a number x with x2 ≡ x
(mod m). This is solvable in polynomial time by a probabilistic algorithm if m is a prime
but is considered difficult in the general case.

132
8.6 Discrete square roots
In this section we discuss the number theoretic algorithm to extract square roots.
We call the integers 0, 1, . . . , p − 1 residues (modulo p). Let p be an odd prime. We say that
y is a square root of x (modulo p), if

y 2 ≡ x (mod p).

If x has a square root then it is called a quadratic residue.

Obviously, 0 has only one square root modulo p: if y 2 ≡ 0 (mod p), then p|y 2 , and since
p is a prime, this implies that p|y. For every other integer x, if y is a square root of x, then so
is p − y = −y (mod p). There are no further square roots: indeed, if z 2 ≡ x for some residue
z, then p|y 2 − z 2 = (y − z)(y + z) and so either p|y − z or p|y + z. Thus z ≡ y or z ≡ −y as
claimed.
This implies that not every integer has a square root modulo p: squaring maps the non-zero
residues onto a subset of size (p − 1)/2, and the other (p − 1)/2 have no square root.
The following lemma provides an easy way to decide if a residue has a square root.

(8.1) Lemma A residue x has a square root if and only if

x(p−1)/2 ≡ 1 (mod p). (8.2)

Proof The “only if” part is easy: if x has a square root y, then

x(p−1)/2 ≡ y p−1 ≡ 1 (mod p)

by Fermat’s “Little” Theorem. Conversely, the polynomial x(p−1)/2 − 1 has degree (p − 1)/2,
and hence it has at most (p − 1)/2 “roots” modulo p (this can be proved just like the well-know
theorem that a polynomial of degree n has at most n real roots). Since all quadratic residues
are roots of x(p−1)/2 − 1, none of the quadratic non-residues can be.
But how to find this square root? For some primes, this is easy:

(8.3) Lemma Assume that p ≡ 3 (mod 4). Then for every quadratic residue x, x(p+1)/4 is a
square root of x.

Indeed,
 2
x(p+1)/4 = x(p+1)/2 = x · x(p−1)/2 ≡ x (mod p).

The case when p ≡ 1 (mod 4) is more difficult, and the solution uses randomization. In
fact, randomization is only needed in the following auxiliary algorithm:

(8.4) Lemma Let p be an odd prime. Then we can find a quadratic non-residue modulo p in
randomized polynomial time.

133
 This can be done by selecting a random residue z = 0, and then testing (using lemma 8.1
whether it is a quadratic residue. If not, we try another z. Since the chance of hitting one is
1/2, we find one in an average of two trials.
One could, of course, try to avoid randomization by testing the residues 2, 3, 5, 7, . . . to see
if they have a square root. Sooner or later we will find a quadratic non-residue. However, it
is not known whether the smallest quadratic non-residue will be found in polynomial time this
way. It is conjectured that one never has to try more than O(log2 p) numbers this way.
Now let us return to the problem of finding the square root of a residue x, in the case when
p is a prime satisfying p ≡ 1 (mod 4). We can write p − 1 = 2k q, where q is odd and k ≥ 2.
We start with finding a quadratic non-residue z. The trick is to find an even power z 2t such
that xq z 2t ≡ 1 (mod p). Then we can take y = x(q+1)/2 z t (mod p). Indeed,

y 2 ≡ xq+1 z 2t ≡ x (mod p).

To construct such a power of z, we consruct for all j ≤ k − 1 an integer tj > 0 such that

x2 q z 2
j j+1 t
j
≡1 (mod p). (8.5)

For j = 0, this is just what we need. For j = k − 1, we can take tk−1 = q:

x2 z2 = x(p−1)/2 z p−1 ≡ 1
k−1 q kq
(mod p),

since x is a quadratic residue and z p−1 ≡ 1 (mod p) by Fermat’s “Little” theorem. This
suggests that we construct the number tj “backwards” for j = k − 2, k − 3, . . .
Suppose that we have tj , j > 0, and we want to construct tj−1 . We know that
   

p x2 q z 2 − 1 = x2 z2 x2 z2
j j+1 t j−1 q jt j−1 q jt
j j
−1 j
+1

We test which of the two factors is a multiple of p. If it is the first, we can simply take tj−1 = tj .
So suppose that it is the second. Now take

tj−1 = tj + 2k−j−1 q.

Then
x2 z2 = x2 z2
j t +2k−1 q
= x2 z2 z (p−1)/2 ≡ (−1)(−1) = 1,
j−1 q jt j−1 q j−1 q jt
j−1 j j

since z is a quadratic non-residue.

This completes the description of the algorithm.
Exercises
1. Show that squaring an integer is not a random number generator.
2. For a string x, let rev(x) denote the reverse string. Show that if g(s) is a safe random
number generator, then so is rev(g(s)).
3. If P = NP then no one-way function exists.
4. Somebody proposes the following random number generator: it takes an integer x with
n bits as the seed, and outputs x3 /10n . Show that this random number generator is not safe.

134
9 Parallel algorithms
New technology makes it more urgent to develop the mathematical foundations of parallel
computation. In spite of the energetic research done, the search for a canonical model of
parallel computation has not settled on a model that would strike the same balance between
theory and practice as the Random Access Machine. The main problem is the modelling of the
communication between different processors and subprograms: this can happen on immediate
channels, along paths fixed in advance, “radio broadcast” like, etc.
A similar question that can be modelled in different ways is the synchronization of the clocks
of the different processors: this can happen with some common signals, or not even at all.
In this section, we treat only one model, the so-called parallel Random Access Machine,
which has been elaborated most from a complexity-theoretic point of view. Results achieved for
this special case expose, however, some fundamental questions of the parallellizability of com-
putations. The presented algorithms can be considered, on the other hand, as programs written
in some high-level language: they must be implemented according to the specific technological
solutions.

9.1 Parallel random access machines

The most investigated mathematical model of machines performing parallel computation is the
parallel Random Access Machine (PRAM). This consists of some fixed number p of identical
Random Access Machines (processors). The program store of the machines is common and
they also have a common memory consisting, say, of the cells x[i] (where i runs through the
integers). It will be convenient to assume (though it would not be absolutely necessary) that
each processor owns an infinite number of program cells u[i]. Beyond this, every processor has
a separate memory cell v containing the serial number of the processor. The processor can read
its own name v and can read and write its own cells x, y, u[i] as well as the common memory
cells x[i]. In other words, to the instructions allowed for the Random Access Machine, we must
add the instructions

u[i] := 0; u[i] := u[i] + 1; u[i] := u[i] − 1; u[i] := u[j];

u[i] := u[i] + u[j]; u[i] := u[i] − u[j]; u[i] := u[u[j]]; u[u[i]] := u[j];
u[i] := x[u[j]]; x[u[i]] := u[j]; if u[i] ≤ 0 then goto p;

We write the input into the cells x[1], x[2], . . .. In addition to the input and the common
program, we must also specify how many processors will be used; we can write this into the cell
x[−1]. The processors carry out the program in parallel but in lockstep. (Since they can refer
to their own name they will not necessarily compute the same thing.) We use a logarithmic
cost function: the cost of writing or reading an integer k from a memory cell x[t] is the total
number of digits in k and t, i.e., approximately log2 |k| + log2 |t|. The next step begins after
each processor has finished the previous step. The machine stops when each processor arrives
at a program line in which there is no instruction. The output is the content of the cells x[i].
An important question to decide is how to regulate the use of the common memory. What
happens if several processors want to write to or read from the same memory cell? We referred to

135
this problem already in connection with the definition of the Parallel Pointer Machine. Several
conventions exist for the avoidance of these conflicts. We mention four of these:

— Two processors must not read from or write to the same cell. We call this the exclusive-
read, exclusive-write (EREW) model. We could also call it completely conflict-free.
This must be understood in such a way that it is the responsibility of programmer to
prevent attempts of simultaneous access to the same cell. If such an attempt occurs the
machine signals program error.

— Maybe the most natural model is the one in which we permit many processors to read
the same cell at the same time but when they want to write this way, this is considered
a program error. This is called the concurrent-read, exclusive-write (CREW) model,
and could also be called half conflict-free.

— Several processors can read from the same cell and write to the same cell but only if they
want to write the same thing. (The machine signals a program error only if two processors
want to write different numbers into the same cell). We call this model concurrent-read,
concurrent-write (CRCW); it can also be called conflict-limiting.

— Many processors can read from the same cell or write to the same cell. If several ones
want to write into the same cell the processor with the smalles serial number succeeds: this
model is called priority concurrent-read, concurrent-write (P-CRCW), or shortly,
the priority model.

9.1 Exercise

(a) Prove that one can determine which one of two 0-1-strings of length n is lexicographically
larger, using n processors, in O(1) steps on the priority model and in O(log n) steps on
the conflict-free model.
∗ (b) Show that on the completely conflict-free model, this actually requires Ω(log n) steps.
∗ (c) How many steps are needed on the other two models?

9.2 Exercise Show that the sum of two 0-1-sequences of length at most n, as binary numbers,
can be computed with n2 processors in O(1) steps on the priority model. ♦

9.3 Exercise

(a) Show that the sum of n 0-1-sequences of length at most n as binary numbers can be
computed, using n3 processors, in O(log n) steps on the priority model.
∗ (b) Show that n2 processors are also sufficient for this.
∗ (c) Perform the same on the completely conflict-free model.

136
 It is obvious that the above models are stronger and stronger since they permit more and
more. It can be shown, however, that—at least if the number of processors is not too great—the
computations we can do on the strongest one, the priority model, are not much faster than the
ones performable on the conflict-free model. The following lemma is concerned with such a
statement.

(9.1) Lemma For every program P, there is a program Q such that if P computes some
output from some input with p processors in time t on the priority model then Q computes on
the conflict-free model the same with O(p2 ) processors in time O(t(log p)2 ).

(9.2) Remark On the PRAM machines, it is necessary to specify the number of processors
not only since the computation depends on this but also since this is—besides the time and the
storage—an important complexity measure of the computation. If it is not restricted then we
can solve very difficult problems very fast. We can decide, e.g., the 3-colorability of a graph if,
for each coloring of the set of vertices and each edge of the graph, we make a processor that
checks whether in the given coloring, the endpoints of the given edge have different colors. The
results must be summarized yet, of course, but on the conflict-limiting machine, this can be
done in a single step. ♦

Proof A separate processor of the conflict-free machine will correspond to every processor
of the priority machine. These are called supervisor processors. Further, every supervisor
processor will have p subordinate processors. One step of the priority machine computation
will be simulated by a stage of the computation of the conflict-free machine.
The basic idea of the construction is that whatever is in the priority machine after a given
step of the computation in a given cell z should be contained, in the corresponding stage of
the computation of the conflict-free machine, in each of the cells with addresses 2pz, 2pz +
1, . . . , 2pz + p − 1. If in a step of the priority machine, processor i must read or write cell z then
in the corresponding stage of the conflict-free machine, the corresponding supervisor processor
will read or write the cell with address 2pz + i. This will certainly avoid all conflicts since the
different processors use different cells modulo p.
We must make sure, however, that by the end of the stage, the conflict-free machine writes
into each cell 2pz, 2pz + 1, . . . , 2pz + p − 1 whatever the priority rule would write into z in the
corresponding step of the priority machine. For this, we insert a phase consisting of O(log p)
auxiliary steps accomplishing this to the end of each stage.
First, each supervisor processor i that in the present stage has written into cell 2pz +i, writes
a 1 into cell 2pz + p + i. Then, in what is called the “first step” of the phase, it looks whether
there is a 1 in cell 2pz + p + i − 1. If yes, it goes to sleep for the rest of the phase. Otherwise, it
writes a 1 there and “wakes” a subordinate. In general, at the beginning of step k, processor i
will have at most 2k−1 subordinates awake (including, possibly, itself); these (at least the ones
that are awake) will examine the corresponding cells 2pz + p + i − 2k−1 , ..., 2pz + p + i − (2k − 1).
The ones that find a 1 go to sleep. Each of the others writes a 1, wakes a new subordinate,
sends it 2k−1 steps left while itself goes 2k steps left. Whichever subordinate gets below 2pz + p
goes to sleep; if a supervisor i leaves does this it knows already that it has “won”.
It is easy to convince ourselves that if in the corresponding step of the priority machine,
several processors wanted to write into cell z then the corresponding supervisor and subordinate

137
processors cannot get into conflict while moving in the interval [2pz + p, 2pz + 2p − 1]. It can be
seen namely that in the k-th step, if a supervisor processor i is active then the active processors
j ≤ i and their subordinates have written 1 into each of the 2k−1 positions downwards starting
with 2pz + p + i that are still ≥ 2pz + p. If a supervisor processor or its subordinates started to
the right from them and reaches a cell ≤ i in the k-th step it will necessarily step into one of
these 1’s and go to sleep, before it could get into conflict with the i-th supervisor processor or
its subordinates. This also shows that always a single supervisor will win, namely the one with
the smallest number.
The winner still has the job to see to it that what it wrote into the cell 2pz + i will be written
into each cell of interval [2pz, 2pz + p − 1]. This is easy to do by a procedure very similar to the
previous one: the processor writes the desired value into cell 2pz, then it wakes a subordinate;
the two of them write the desired value into the cells 2pz + 1 and 2pz + 2 then they wake one
subordinate each, etc. When they all have passed 2pz + p − 1 the phase has ended and the next
simulation stage can start.
We leave to the reader to plan the waking of the subordinates.
Each of the above “steps” requires the performance of several program instructions but it is
easy to see that only a bounded number is needed, whose cost is, even in case of the logarithmic-
cost model, only O(log p + log z). In this way, the time elapsing between two simulating stages
is only O(log p(log p + log z)). Since the simulated step of the priority machine also takes at
least log z units of time the running time is thereby increased only O((log z)2 )-fold.
In what follows if we do not say otherwise we use the conflict-free (EREW) model. According
to the previous lemma, we could have agreed on one of the other models.
It is easy to convince ourselves that the following statement holds.

(9.3) Proposition If a computation can be performed with p processors in t steps with num-
bers of at most s bits then for all q < p, it can be performed with q processors in O(tp/q) steps
with numbers of at most O(s + log(p/q)) bits. In particular, it can be performed on a sequencial
Random Access Machine in O(tp) steps with numbers of length O(s + log p).

The fundamental question of the complexity theory of parallel algorithms is just the opposite
of this: given is a sequential algorithm with time N and we would like to implement it on p
processors in “essentially” N/p (say, in O(N/p)) steps. Next, we will overview some complexity
classes motivated by this question.
Randomization is, as we will see, an even more important tool in the case of parallel com-
putations than in the sequential case. The randomized parallel Random Access Machine
differs from the above introduced parallel Random Access Machine only in that each processor
has an extra cell in which, with probability 1/2, there is always 0 or an 1. If the processor reads
this bit then a new random bit occurs in the cell. The random bits are completely independent
(both within one processor and between different processors).

9.2 The class NC

We say that a program for the parallel Random Access Machine is an NC-program if there are
constants c1 , c2 > 0 such that for all inputs x the program computes conflict-free with O(|x|c1 )

138
processors in time O((log |x|)c2 ). (According to Lemma 9.1, it would not change this definition
if we used e.g. the priority model instead.)
The class NC of languages consists of those languages L ⊂ {0, 1}∗ whose characteristic
function can be computed by an NC-program.

(9.1) Remark The goal of the introduction of the class NC is not to model practically imple-
mentable parallel computations. In practice, we can generally use much more than logarithmic
time but (at least in the foreseeable future) only on much fewer than polynomially many pro-
cessors. The goal of the notion is to describe those problems solvable with a polynomial number
of operations, with the additional property that these operations are maximally parallelizable
(in case of an input of size n, on the completely conflict-free machine, log n steps are needed
even to let all input bits have an effect on the output). ♦

Obviously, N C ⊂ P . It is not known whether equality holds here but the answer is probably
no.
We define the randomized NC, or RNC, class of languages on the pattern of the class BPP.
This consists of those languages L for which there is a number c > 0 and a program computing,
on each input x ∈ {0, 1}∗ , on the randomized P RAM machine, with O(|x|c ) processors (say,
in a completely conflict-free manner), in time O(log |x|c ), either a 0 or an 1. If x ∈ L then the
probability of the result 0 is smaller than 1/4, if x ∈ L then the probability of the result 1 is
smaller than 1/4.
Around the class NC, a complexity theory can be built similar to the one around the class
P. The NC-reduction of a language to another language can be defined and, e.g. inside the
class P, it can be shown that there are languages that are P-complete, i.e. to which every other
language in P is NC-reducible. We will not deal with the details of this; rather, we confine
ourselves to some important examples.

(9.2) Proposition The adjacency-matrices of graphs containing a triangle form a language in

NC.

Proof The NC-algorithm is essentially trivial. Originally, let x[0] = 0. First, we determine the
number n of points of the graph. Then we instruct the processor with serial number i + jn + kn2
to check whether the point triple (i, j, k) forms a triangle. If no then the processor halts. If yes
then it writes a 1 into the 0’th common cell and halts. Whether we use the conflict-limiting or
the priority model, we have x[0] = 1 at the end of the computation if and only if the graph has
a triangle. (Notice that this algorithm makes O(1) steps.)
Our next example is less trivial, moreover, at the first sight, it is surprising: the connec-
tivity of graphs. The usual algorithms (breadth-first or depth-first search) are namely strongly
sequential: every step depends on the result of the earlier steps. For the parallelization, we use
a trick similar to the one we used earlier for the proof of Savitch’s theorem.

(9.3) Proposition The adjacency matrices of connected graphs form a language in NC.

139
Proof We will describe the algorithm on the conflict-limiting model. Again, we instruct the
processor with serial number i + jn + kn2 to watch the triple (i, j, k). If it sees two edges in
the triple then it inserts the third one. (If several processors want to insert the same edge then
they all want to write the same thing into the same cell and this is permitted.) If we repeat
this t times then, obviously, exactly those pairs of points will be connected whose distance in
the original graph is at most 2t . In this way, repeating O(log n) times, we obtain a complete
graph if and only if the original graph was connected.
Clearly, it can be similarly decided whether in a given graph, there is a path connecting
two given points, moreover, even the distance of two points can be determined by a suitable
modification of the above algorithm.

9.4 Exercise Give an NC algorithm that in a given graph, computes the distance of two
points. ♦

(9.4) Proposition The product of two matrices (in particular, the scalar product of two vec-
tors), and the k-th power of an n × n matrix (k ≤ n) is NC-computable.

Proof We can compute the scalar product of two vectors as follows: we multiply—parallelly—
their corresponding elements; then we group the products obtained this way in pairs and form
the sums; then we group these sums in pairs and form the sums, etc. Now, we can also compute
the product of two matrices since each element of the product is the scalar product of two
vectors, and these can be computed parallelly. Now the k-th power of an n × n matrix can be
computed on the pattern of ab (mod c) (Lemma 3.3).
The next algorithm is maybe the most important tool of the theory of parallel computations.

(9.5) Csánky’s Theorem The determinant of an arbitrary integer matrix can be computed
by an NC algorithm. Consequently, the invertible matrices form an NC-language.

Proof We present an algorithm proposed by Chistov. The idea is now to try to represent
the determinant by a suitable matrix power-series. Let B be an n × n matrix and let Bk denote
the k × k submatrix in its left upper corner. Assume first that these submatrices Bk are not
singular, i.e., that their determinants are not 0. Then B is invertible and according to the
known formula for the inverse, we have

(B −1 )nn = det Bn−1 / det B

where (B −1 )nn denotes the element standing in the right lower corner of the matrix B −1 . Hence
det Bn−1
det B = .
(B −1 )nn
Continuing this, we obtain
1
det B = −1 .
(B −1 )nn · (Bn−1 )n−1,n−1 · · · (B1−1 )11

140
Let us write B in the form B = I − A where I = In is the n × n unit matrix. Assuming, for a
moment, that the elements of A are small enough, the following series expansion holds:

Bk−1 = Ik + Ak + A2k + · · · ,

which gives

(Bk−1 )kk = 1 + (Ak )kk + (A2k )kk + · · · .

Hence
1 1
=
(Bk−1 )kk 1 + (Ak )kk + (A2k )kk + · · ·
= 1 − [(Ak )kk + (A2k )kk + · · ·] + [(Ak )kk + (A2k )kk + · · ·]2 − · · · ,

and hence

n
det B = (1 − [(Ak )kk + (A2k )kk + · · ·] + [(Ak )kk + (A2k )kk + · · ·]2 − · · ·).
k=1

We cannot, of course, compute these infinite series composed of infinite series. We claim,
however, that it is enough to compute only n terms from each series. More exactly, let us
substitute tA in place of A where t is a real variable. For small enough t, the matrices Ik − tAk
are certainly not singular and the above series expansions hold. We gain, however, more. After
substitution, the formula looks as follows:

n
det(I − tA) = (1 − [t(Ak )kk + t2 (A2k )kk + · · ·] + [t(Ak )kk + t2 (A2k )kk + · · ·]2 − · · ·).
k=1

Now comes the decisive idea: the left-hand side is a polynomial of t of degree at most n, hence
from the power series on the right-hand side, it is enough to compute only the terms of degree
at most n. In this way, det(I − tA) consists of the terms of degree at most n of the following
polynomial:

n 
n 
n
F (t) = [ (− tm (Am j
k )kk ) ].
k=1 j=0 m=1

Now, however complicated the formula defining F (t) may seem, it can be computed easily in
the NC sense. Deleting from it the terms of degree higher than n, we get a polynomial identical
to det(I − tA). Also, as a polynomial identity, our identity holds for all values of t, not only
for the small ones, and no nonsingularity assumptions are needed. Substituting t = 1 here, we
obtain det B.
Using Theorem 6.3 with random substitutions, we arrive at the following important appli-
cation:

(9.6) Corollary The adjacency matrices of the graphs with complete matchings form a lan-
guage in RNC.

141
 No combinatorial proof (i.e. one avoiding the use of Csánky’s theorem) is known for this
fact. It must be noted that the algorithm only determines whether the graph has a complete
matching but it does not give the matching if it exists. This, significantly harder, problem can
also be solved in the RNC sense (by an algorithm of Karp, Upfal and Wigderson).

9.5 Exercise Consider the following problem. Given a Boolean circuit and its input, compute
its output. Prove that if this problem is in NC then P=NC. ♦

142
10 Decision trees
The logical framework of many algorithms can be described by a tree: we start from the root
and in every internal node, the result of a certain “test” determines which way we continue.
E.g., most sorting algorithms make comparisons between certain pairs of elements and continue
the work according to the result of the comparison. We assume that the tests performed in
such computations contain all the necessary information about the input, i.e., when we arrive
at a leaf of the tree, all that is left is to read off the output from the leaf. The complexity of
the tree gives some information about the complexity of the algorithm; for example, the depth
of the tree (the number of edges in the longest path leaving the root) tells us how many tests
must be performed in the worst case during the computation. We can describe, of course, every
algorithm by a trivial tree of depth 1 (the test performed in the root is the computation of the
end result). This algorithmic scheme makes sense only if we restrict the kind of tests allowed in
the nodes.
We will see that decision trees not only give a graphical representation of the structure of
some algorithms but are also suitable for proving lower bounds on their depth. Such a lower
bound can be interpreted as saying that the problem cannot be solved (for the worst input) in
fewer steps than some given number, if we assume that information on the input is available
only by the permissible tests (for example, in sorting we can only compare the given numbers
with each other and cannot perform e.g. arithmetic operations on them).

10.1 Algorithms using decision trees

Consider some simple examples.

10.1.1 Binary search

Perhaps the simplest situation in which decision trees are used is binary search. We want to
compute an integer k, about which at the beginning we only that it lies, say, in the interval
[1, N ]. We have an algorithm that, given any integer m, 1 ≤ m ≤ N , can decide whether a ≤ m
is true. Then by calling this algorithm log2 N  times, we can determine a. We have used this
method when we showed that factoring of an integer can be reduced to the problem of finding
a bounded divisor.
We can describe this algorithm by a rooted binary tree: every node will correspond to an
interval [u, v] ⊆ [1, N ]. The root corresponds to the interval [1, N ], and each node corresponds
to the interval of integers that are still possible values for a if we arrive at the node. The
leaves correspond to one-element intervals, i.e., the possible values of a. For an internal node
corresponding to the interval [u, v], we select w = (u + v)/2 and test if a ≤ w. Depending on
the outcome of this test, we proceed to one of the children of the node, which correspond to the
intervals [u, w] and [w + 1, v].

10.1.2 Finding a false coin with a one-armed scale

We are given n coins looking identical from the outside. We know that each must weigh 1g;
but we also know that there is a false one among them that is lighter than the rest. We have

143
a one-armed scale; we can measure with it the weight of an arbitrary subset of the coins. How
many measurements are enough to decide which coin is false?
The solution is simple: with one measurement, we can decide about an arbitrary set of
coins whether the false one is among them. If we put n/2 coins on the scale, then after one
measurement, we have to find the false coin only among at most n/2 ones. This recursion
ends in log2 n steps.
We can characterize the algorithm by a rooted binary tree. Every vertex v corresponds to a
set Xv of coins; arriving into this vertex we already know that the false coin is to be found into
this set. (The root corresponds to the original set, and the leaves to the 1-element sets.) For
every internal node v, we divide the set Xv into two parts, with numbers of elements |Xv |/2
and |Xv |/2. These correspond to the children of v. Measuring the first one we learn which
one contains the false coin.

10.1.3 Finding a false coin with a two-armed scale

Again, we are given n outwardly identical coins. We know that there is a false one among them
that is lighter than the rest. This time we have a two-armed scale but without weights. On this,
we can find out which one of two (disjoint) sets of coins is lighter, or whether they are equal.
How many measurements suffice to decide which coin is false?
Here is a solution. One measurement consists of putting the same number of coins into each
pan. If one side is lighter then the false coin is in that pan. If the two sides have equal weight
then the false coin is among the ones left out. It is most practical to put n/3 coins into both
pans; then after one measurement, the fals coin must be found only among at most n/3 coins.
This recursion terminates in log3 n steps.
Since one measurement has 3 possible outcomes, the algorithm can be characterized by a
rooted tree in which each internal node has 3 children. Every node v corresponds to a set Xv
of coins; arriving into this node we already know that the false coin is to be found in this set.
(As above, the root corresponds to the original set and the leaves to the one-element sets.)
For each internal node v, we divide the set Xv into three parts, with |Xv |/3, |Xv |/3 and
|Xv | − 2|Xv |/3 elements. These correspond to the children of v. Comparing the two first ones
we can find out which one of the three contains the false coin.

10.1.4 Sorting
Given are n elements that are ordered in some way (unknown to us). We know a procedure to
decide the order of two elements; this is called a comparison and considered an elementary
step. We would like to determine the complete ordering using as few comparisons as possible.
Many algorithms are know for this basic problem of data processing; we treat this question only
to the depth necessary
n
for the illustration of decision trees.
Obviously, 2 comparisons are enough: with these, we can learn about every pair of ele-
ments, wich which one in the pair is greater, and this determines the order. These comparisons
are not, however, independent: often, we can infer the order of certain pairs using transitivity.

Indeed, it is enough to make nk=1 log2 k ∼ n log2 n comparisons. Here is the simplest way
to see this: suppose that we already determined the ordering of the first n − 1 elements. Then

144
already only the n-th element must be “inserted”, which can obviously be done with log2 n
comparisons.
This algorithm, as well as any other sorting algorithm working with comparisons, can be
represented by a binary tree. The root corresponds to the first comparison; depending on its
result, the algorithm branches into one of the children of the root. Here, we make another
comparison, etc. Every leaf corresponds to a complete ordering.

(10.1) Remark In the above sorting algorithm, we only counted the comparisons. For a real
implementation, one should also take into account the other operations, e.g. the movement of
data, etc. From this point of view, the above algorithm is not good since every insertion may
require the movement of all elements placed earlier and this may cause Ω(n2 ) extra steps. There
exist, however, sorting algorithms requiring altogether only O(n log n) steps. ♦

10.1.5 Convex hull

The determination of the convex hull of n planar points is as basic among the geometrical
algorithms as sorting for data processing. The points are given by their coordinates: p1 =
(x1 , y1 ), . . . , pn = (xn , yn ). We assume, for simplicity, that the points are in general position,
i.e. no 3 of them is on one straight line. We want to determine those indices i0 , . . . , ik−1 , ik = i0
for which pi0 , . . . , pik−1 , pik are the vertices of the convex hull of the given point set, in this order
along the convex hull (starting anticlockwise, say, from the point with the smallest abscissa).
The idea of “insertion” gives a simple algorithm here, too. Sort the elements by their xi
coordinates; this can be done in time O(n log n). Suppose that p1 , . . . , pn are already indexed
in this order. Delete the point pn and determine the convex hull of the points p1 , . . . , pn−1 : let
this be the sequence of points pj0 , . . . , pjm−1 , pjm where j0 = jm = 1.
Now, the addition of pn consists of deleting the arc of the polygon pj0 , . . . , pjm “visible”
from pn and replacing it with the point pn . Let us determine the first and last elements of the
sequence pj0 , . . . , pjm visible from pn , let these be pja and pjb . Then the convex hull sought for
is pj0 , . . . , pja , pn , pjb , pjm (Figure 10.1).
How to determine whether some vertex pjs is visible from pn ? The point pn−1 is evidently
among the vertices of the polygon and is visible from pn ; let jt = n − 1. If s < t then, obviously,
pjs is visible from pn if and only if pn is below the line pjs pjs+1 . Similarly, if s > t then pjs is
visible from pn if and only if pn is above the line pjs pjs−1 . In this way, it can be decided about
every pjs in O(1) steps whether it is visible from pn .
Using this, we can determine a and b in O(log n) steps and we can perform the “insertion”
of the point pn . This recursion gives an algorithm with O(n log n) steps.
It is worth separating here the steps in which we do computations with the coordinates of the
points, from the other steps (of combinatorial character). We do not know namely, how large are
the coordinates of the points, whether multiple-precision computation is needed, etc. Analysing
the described algorithm, we can see that the coordinates needed to be taken into account only
in two ways: at the sorting, when we had to make comparisons among the abscissas, and at
deciding whether point pn was above or below the straight line determined by the points pi and
pj . The last one can be also formulated by saying that we must determine the orientation of
the triangle pi pj pk . This can be done in several ways using the tools of analytic geometry.

145
 p
jb

p
n

p
ja
Figure 10.1: A simple convex hull algorithm

The above algorithm can again be described by a binary decision tree: each of its nodes
corresponds either to the comparison of the abscissas of two given points or to the determination
of the orientation of a triangle given by three points. The algorithm gives a tree of depth
O(n log n). (Many other algorithms looking for the convex hull lead to a decision tree of similar
depth.)

10.2 The notion of decision trees

To formalize the notion of a decision tree let us be given the set A of possible inputs, the set
B of possible outputs and a set Φ of functions defined on A with values in {1, . . . , d}, the test-
functions. A decision tree is a rooted tree whose internal nodes (including the root) have d
children (the tree is d-regular), its leaves are labelled with the elements of B, the other nodes
with the functions of Φ. We assume that for every vertex, the edges leaving it are numbered in
some order.
Every decision tree determines a function f : A → B. Let namely a ∈ A. Starting from the
root, we walk down to a leaf as follows. If we are in an internal node v then we compute the
test function assigned to v at the place a; if its value is i then we step further to the i-th child
of node v. In this way, we arrive at a leaf w; the value of f (a) is the label of w.
The question is that for a given function f , what is the decision tree with minimum depth
computing it.
In the simplest case, we want to compute a Boolean function f (x1 , . . . , xn ) and every test
that can be made in the vertices of the decision tree is the reading in of the value of one of the
variables. In this case, we call the decision tree simple. Every simple decision tree is binary

146
 x
2

1 0

x x
3 3
1 0 1 0

1 x x 0
4 1
1 0 1 0

1 0 1 0

Figure 10.2: A simple decision tree

(2-regular), the internal nodes are indexed with the variables, the leaves with 0 and 1.
The decision tree corresponding to binary search over the interval [1, 2n ] can be considered
as simple, if we regard the consecutive comparisons as asking for the next bit of the unknown
number a. The decision tree for sorting is not simple: there, the tests (comparisons) are not
independent since the ordering is transitive. We denote by D(f ) the minimal depth of a simple
decision tree computing a Boolean function f .
(10.1) Example Consider the Boolean function
f (x1 , x2 , x3 , x4 ) = (x1 ∨ x2 ) ∧ (x2 ∨ x3 ) ∧ (x3 ∨ x4 ).
This is computed by the simple decision tree in Figure 10.2. This shows that D(f ) ≤ 3. It
is easy to see that we cannot compute this funtion by a decision tree of depth 2, and hence
D(f ) = 3. ♦
Every decision tree can also be considered a two-person “twenty questions”-like game. One
player (Xavier) thinks of an element a ∈ A, and it is the task of the other player (Yvette) to
determine the value of f (a). For this, she can pose questions to Xavier. Her questions cannot be,
however, arbitrary, she can only ask the value of some test function in Φ. How many questions
do suffice for her to compute the answer? Yvette’s strategy corresponds to a decision tree, and
Xavier plays optimally if with his answers, he drives Yvette to the leaf farthest away from the
root. (Xavier can “cheat, as long as he is not caught”—i.e., he can change his mind about the
element a ∈ A as long as the new one still makes all his previous answers correct. In case of a
simple decision tree, Xavier has no such worry at all.)

10.3 Nondeterministic decision trees

The idea learned in Chapter 5, nondeterminism, helps in other complexity-theoretic investi-
gations, too. In the decision-tree model, the same idea can be formulated as follows (we will

147
only consider the case of simple decision trees). Let f : {0, 1}n → {0, 1} be the function to
be computed. Two numbers characterize the nondeterministic decision-tree complexity (simi-
larly to having two complexity classe for non-deterministic polynomial time, namely NP and
co-NP). For every input x, let D(f, x) denote the minimum number of those variables whose
value already determines the value of f (x). Let

D0 (f ) = max{D(f, x) : f (x) = 0}, D1 (f ) = max{D(f, x) : f (x) = 1}.

In other words, D0 (f ) is the smallest number with the property that for all inputs x with
f (x) = 0, we can test D0 (f ) variables in such a way that knowing these, the value of the
function is already determined (it may depend on x which variables we will test). The number
D1 (f ) can be characterized similarly. Obviously,

D(f ) ≥ max{D0 (f ), D1 (f )}.

It can be seen from the examples below that equality does not necessarily hold here.

(10.1) Example Assign a Boolean variable xe to each edge e of the complete graph Kn . Then
every assignment corresponds to an n-point graph (we connect n
with edges those pairs whose
assigned value is 1). Let f be the Boolean function with 2 variables whose value is 1 if in
the graph corresponding to the input, the degree of every node is at least one and 0 otherwise
(i.e. if there is an isolated point). Then D0 (f ) ≤ n − 1 since if there is an isolated point in the
graph it is enough to know about the n − 1 edges leaving it that they are not in the graph. It
is also easy to see that we cannot infer an isolated point from the adjacency or nonadjacency of
n − 2 pairs, and thus

D0 (f ) = n − 1.

Similarly, if there are no isolated points in a graph then this can be proved by the existence of
n − 1 edges (it is enough to know one edge leaving each node and one of the edges even covers
2 nodes). If the input graph is an (n − 1)-star then fewer than n − 1 edges are not enough.
Therefore

D1 (f ) = n − 1.

Thus, whichever is the case, we can know the answer after n − 1 lucky questions. On the other
hand, if we want to decide which one is the case then we cannot know in advance which edges
to ask; it can be shown that the situation is as bad as it can be, namely
 
n
D(f ) = .
2

We return to the proof of this in the next section (exercise 10.8). ♦

(10.2) Example Let now G be an arbitrary but fixed n-point graph an let us assign a variable
to each of its vertices. An assignment of the variables corresponds to a subset of the vertices.

148
Let the value of the function f be 0 if this set is independent in the graph and 1 otherwise. This
property can also be simply expressed by a Boolean formula:

f (x1 , . . . , xn ) = (xi ∧ xj ).
ij∈E(G)

If the value of this Boolean function is 1 then this will be found out already from testing 2
vertices, but of course not from testing a single point, i.e.

D1 (f ) = 2.

On the other hand, if after testing certain points we are sure that the set is independent then
the vertices that we did not ask must form an independent set. Thus

D0 (f ) ≥ n − α

where α is the maximum number of independent points in the graph. It can also be proved (see
Theorem 10.6) that if n is a prime and a cyclic permutation of the points of the graph maps
the graph onto itself, and the graph has some edges but is not complete, then

D(f ) = n.

We see therefore that D(f ) can be substantially larger than the maximum of D0 (f ) and
D1 (f ), moreover, it can be that D1 (f ) = 2 and D(f ) = n. However, the following beautiful
relation holds:

(10.3) Theorem
D(f ) ≤ D0 (f )D1 (f ).

Proof We use induction over the number n of variables. If n = 1 then the inequality is trivial.
Let (say) f (0, . . . , 0) = 0; then k ≤ D0 (f ) variables can be chosen such that fixing their
values to 0, the function is 0 independently of the other variables. We can assume that the first
k variables have this property.
Next, consider the following decision algorithm. We ask the value of the first k variables; let
the obtained answers be a1 , . . . , ak . Fixing these, we obtain a Boolean function

g(xk+1 , . . . , xn ) = f (a1 , . . . , ak , xk+1 , . . . , xn ).

Obviously, D0 (g) ≤ D0 (f ) and D1 (g) ≤ D1 (f ). We claim that the latter inequality can be
strenghened:

D1 (g) ≤ D1 (f ) − 1.

Consider an input (ak+1 , . . . , an ) of g with g(ak+1 , . . . , an ) = 1. Together with the bits a1 , . . . , ak ,

this gives an input of the Boolean function f for which f (a1 , . . . , an ) = 1. According to the defi-
nition of the quantity D1 (f ), one can choose m ≤ D1 (f ) variables, say, xi1 , . . . , xim of f such that
fixing them at the value ai , the value of f becomes 1 independently of the other variables. One

149
of the first k variables must occur among these m variables; otherwise, f (0, . . . , 0, ak+1 , . . . , an )
would have to be 0 (due to the fixing of the first k variables) but would also have to be 1 (due
to the fixing of xi1 , . . . , xim ), which is a contradiction. Thus, in the function g, at the position
ak+1 , . . . , ak , only m − 1 variables must be fixed to obtain the identically 1 function. From this,
the claim follows. From the induction hypothesis,

D(g) ≤ D0 (g)D1 (g) ≤ D0 (f )(D1 (f ) − 1)

and hence

D(f ) ≤ k + D(g) ≤ D0 (f ) + D(g) ≤ D0 (f )D1 (f ).

In Example 10.2, we could define the function by a disjunctive 2-normal form and D1 (f ) = 2
was true. This is not an accidental coincidence:

(10.4) Proposition If f is expressible by a disjunctive k-normal form then D1 (f ) ≤ k. If f is

expressible by a conjunctive k-normal form then D0 (f ) ≤ k.

Proof It is enough to prove the first assertion. Let (a1 , . . . , an ) be an input for which the
value of the function is 1. Then there is an elementary conjunction in the disjunctive normal
form whose value is 1. If we fix the variables occurring in this conjunction then the value of the
function will be 1 independently of the values of the other variables.
For monotone functions, the connection expressed in the previous proposition is even tighter:

(10.5) Proposition A monotonic Boolean function is expressible by a disjunctive [conjunctive]

k-normal form if and only if D1 (f ) ≤ k [D0 (f ) ≤ k].

Proof According to Proposition 10.4, it is sufficient to see that if D1 (f ) = k then f is ex-

pressible by a disjunctive k-normalform. Let {xi1 , . . . , xim } be a subset of the variables minimal
with respect to containment, that can be fixed in such a way as to make the obtained function
is identically 1. (Such a function is called a minterm.) Notice that then we had to fix every
variable xij necessarily to 1: due to the monotonicity, this fixing gives the identically 1 function,
and if a variable could also be fixed to 0 then it would not have to be fixed to begin with.
We will show that m ≤ k. Let us namely assign the value 1 to the variables xi1 , . . . , xim and
0 to the others. According to the foregoing, the value of the function is 1. By the definition of
the quantity D1 (f ), we can fix in this assignment k values in such a way as to make the obtained
function identically 1. By the above remarks, we can assume that we only fix 1’s, i.e. we only
fix some of the variables xi1 , . . . , xim . But then due to the minimality of the set {xi1 , . . . , xim },
we had to fix all of them, and hence m ≤ k.

Let us prepare for every minterm S the elementary conjunction ES = xi ∈S xi and take the
disjunction of these. By what was said above, we obtain a disjunctive k-normal form this way.
It can be verified trivially that this defines the function f .

150
10.4 Lower bounds on the depth of decision trees
We mentioned that decision trees as computation models have the merit that non-trivial lower
bounds can be given for their depth. First we mention, however, a non-trivial lower bound also
called information-theoretic estimate.
(10.1) Lemma If the range of f has t elements then the depth of every decision tree of degree
d computing f is at least logd t.

Proof A d-regular rooted tree of depth h has at most dh leaves. Since every element of the
range of f must occur as a label of a leaf it follows that t ≥ dh .
As an application, let us take an arbitrary sorting algorithm. The input of this is a per-
mutation a1 , . . . , an of the elements 1, 2, . . . , n, its output is the same, while the test functions
compare two elements:

1 if ai < aj
ϕij (a1 , . . . , an ) =
0 otherwise.
Since there are n! possible outputs, the depth of any binary decision tree computing the complete
order is at least log n! ∼ n log n. The sorting algorithm mentioned in the introduction makes at
most log n + log(n − 1) + · · · + log 1 ∼ n log n comparisons.
This bound is often very weak; if e.g. only a single bit must be computed then it says nothing.
Another simple trick for proving lower bounds is the following observation.
(10.2) Lemma Assume that there is an input a ∈ A such that no matter how we choose k test
functions, say, ϕ1 , . . . , ϕk , there is an a ∈ A for which f (a ) = f (a) but ϕi (a ) = ϕi (a) holds for
all 1 ≤ i ≤ k. Then the depth of every decision tree computing f is greater than k.
For application, let us see how many comparisons suffice to find the largest one of n elements.
We have seen (championship by elimination) that n−1 comparisons are enough for this. Lemma
10.1 gives only log n for lower bound; but we can apply Lemma 10.2 as follows. Let a =
(a1 , . . . , an ) be an arbitrary permutation, and consider k < n − 1 comparison tests. The pairs
(i, j) for which ai and aj will be compared form a graph G over the underlying set {1, . . . , n}.
Since it has fewer than n − 1 edges this graph falls into two disconnected parts, G1 and G2 .
Without loss of generality, let G1 contain the maximal element and let p denote its number of
vertices. Let a = (a1 , . . . an ) be the permutation containing the numbers 1, . . . , p in the positions
corresponding to the vertices of G1 and the numbers p + 1, . . . , n in those corresponding to the
vertices of G2 ; the order of the numbers within both sets must be the same as in the original
permutation. Then the maximal element is in different places in a and in a but the given k
tests give the same result for both permutations.
In what follows we show estimates for the depth of some more special decision trees, applying,
however, some more interesting methods. First we mention a result which gives a lower bound
of unusual character for the depth of decision trees.
(10.3) Theorem Let f : {0, 1}n → {0, 1} be an arbitrary Boolean function. Let N denote the
number of those substitutions making the value of the function “1” and let 2k be the largest
power of 2 dividing N . Then the depth of any decision tree computing f is at least n − k.

151
Proof Consider an arbitrary decision tree of depth d that computes the function f , and a leaf
of this tree. Here, m ≤ d variables are fixed, therefore there are at least 2n−m inputs leading
to this leaf. All of these correspond to the same function value, therefore the number of inputs
leading to this leaf and giving the function value “1” is either 0 or or 2n−m . This number is
therefore divisible by 2n−d . Since this holds for all leaves, the number of inputs giving the value
“1” is divisible by 2n−d and hence k ≥ n − d.
We call a Boolean function f of n variables evasive if it cannot be computed by a decision
tree of length smaller than n. It follows from Theorem 10.3 that if a Boolean function has an
odd number of substitutions making it “1” then the function is evasive.
We obtain another important class of evasive functions by symmetry-conditions. A Boolean
function is called symmetric if every permutation of its variables leaves its value unchanged.
E.g., the functions x1 + · · · + xn , x1 ∨ · · · ∨ xn and x1 ∧ · · · ∧ xn are symmetric. A Boolean
function is symmetric if and only if its value depends only on how many of its variables are 0
or 1.

(10.4) Proposition Every non-constant symmetric Boolean function is evasive.

Proof Let f : {0, 1}n → {0, 1} be the Boolean function in question. Since f is not constant,
there is a j with 1 ≤ j ≤ n such that if j − 1 variables have value 1 then the function’s value is
0 but if j variables are 1 then the function’s value is 1 (or the other way around).
Using this, we can propose the following strategy to Xavier. Xavier thinks of a 0-1-sequence
of length n and Yvette can ask the value of each of the xi . Xavier answers 1 on the first j − 1
questions and 0 on every following question. Thus after n − 1 questions, Yvette cannot know
whether the number of 1’s is j − 1 or j, i.e. she cannot know the value of the function.
Symmetric Boolean functions are very special; the following class is significantly more gen-
eral. A Boolean function of n variables is called weakly symmetric if for all pairs xi , xj of
variables, there is a permutation of the variables that takes xi into xj but does not change the
value of the function. E.g. the function

(x1 ∧ x2 ) ∨ (x2 ∧ x3 ) ∨ · · · ∨ (xn−1 ∧ xn ) ∨ (xn ∧ x1 )

is weakly symmetric but not symmetric. The question below (the so-called generalized Aandera-
Rosenberg-Karp conjecture) is open:

(10.5) Conjecture If a non-constant monotone Boolean function is weakly symmetric then it

is evasive.

We show that this conjecture is true in an important special case.

(10.6) Theorem If a non-constant monotone Boolean function is weakly symmetric and the
number of its variables is a prime number then it is evasive.

152
Proof Let p be the number of variables (emphasizing that this number is a prime). We use
the group-theoretic result that if a prime p divides the order of a group, then the group has
an element of order p. In our case, those permutations of the variables that leave the value of
the function invariant form a group, and from the week symmetry it follows that the order of
this group is divisible by p. Thus the group has an element of order p. This means that with a
suitable labeling of the variables, the substitution x1 → x2 → · · · → xp → x1 does not change
the value of the function.
Now consider the number

M= f (x1 , . . . , xp )(p − 1)x1 +···+xp , (10.7)
where the summation extends to all systems of values (x1 , . . . , xp ) ∈ {0, 1}p . By an argument
that is quite similar to the proof of Theorem 10.3, we can see that if f is non-evasive, then M
is divisible by p.
It follows that in the definition of M , if in some term, not all the values x1 , . . . , xp are the
same, then p identical terms can be made from it by cyclic substitution. The contribution of
such terms is therefore divisible by n. Since the function is not constant and is monotone,
it follows that f (0, . . . , 0) = 0 and f (1, . . . , 1) = 1, from which it can be seen that M gives
remainder (−1)p modulo p.
We get important examples of weakly symmetric Boolean functions taking any graph prop-
erty. Consider an arbitrary property of graphs, e.g. planarity; we only assume that if a graph
has this property then every graph isomorphic with it also has it. We can specify a graph
with n points by fixing its vertices (let these be 1, . . . , n), and for all pairs i, j ⊂ {1, . . . , n},
we introduce a Boolean variable xij with value 1 if i and j are connected and 0 if they are  
not. In this way, the planarity of n-point graph can be considered a Boolean function with n2
variables. Now, this Boolean function is weakly symmetric: for every two pairs, say, {i, j} and
{u, v}, there is a permutation of the vertices taking i into u and j into v. This permutation also
induces a permutation on the set of point pairs that takes the first pair into the second one and
does not change the planarity property.
A graph property is called trivial if either every graph has it or no one has it. A graph
property is monotone if whenever a graph has it each of its subgraphs has it. For most graph
properties that we investigate (connecivity, the existence of a Hamiltonian circuit, the existence
of complete matching, colorability, etc.) either the property itself or its negation is monotonic.
The Aandera-Rosenberg-Karp conjecture, in its original form, concerns graph properties:

(10.8) Conjecture Every non-trivial monotonic graph property is evasive, i.e., every decision
tree that decides
n
such a graph property and that can only test whether two nodes are connected,
has depth 2 .

This conjecture is proved for a number of graph properties: for a general property, what is
known is only that the tree has depth Ω(n2 ) (Rivest and Vuillemin) and that the theorem is
true if the number of points is a prime power (Kahn, Saks and Sturtevant). The analogous
conjecture is also proved for bipartite graphs (Yao).

10.1 Exercise Prove that fewer measurements do not suffice in either problem 10.1.2 or prob-
lem 10.1.3. ♦

153
10.2 Exercise Show that to pick the median of 2n + 1 numbers,
(a) at least 2n comparisons are needed;
(b)* O(n) comparisons suffice. ♦

10.3 Exercise Show that the problem of sorting n real numbers can be reduced in a linear
number of steps to the problem of determining the convex hull of n planar points. ♦

10.4 Exercise Show that the second phase of the above algorithm, i.e. the determination of
the convex hull of the points p1 , . . . , pi for i = 2, . . . , n, can be performed in O(n) steps provided
that the points are already sorted by their x coordinates. ♦

10.5 Exercise Give an example showing that in Proposition 10.5, the condition of monotonic-
ity cannot be omitted. ♦

10.6 Exercise Given an n-variable Boolean function f , construct the following polynomial:

Ψf (t) = f (x1 , . . . , xn )tx1 +···+xn where the summation extends to all systems of values
(x1 , . . . , xn ) ∈ {0, 1}n . Prove that if f can be computed by a decision tree of depth d, then
Ψf (t) is divisible by (t + 1)n−d . ♦

10.7 Exercise Prove that the connectedness of a graph is a evasive property. ♦

10.8 Exercise
(a) Prove that if n is even then on n fixed points, the number of graphs not containing
isolated points is odd.
(b) If n is even then the graph property that in an n-point graph there is no isolated point,
is evasive.
(c)* This statement holds also for odd n. ♦

10.9 Exercise A tournament is a complete graph each of whose edges is directed. Each
tournament can be described by n2 bits saying how the individual edges  
of the graph are
directed. In this way, every property of tournamentse can be considered an n2 -variable Goolean
function. Prove that the tournament property that there is a 0-degree vertex is evasive. ♦

10.10 Exercise (a) If we allow a polynomial of degree n2 as test function then a decision tree
of depth 1 can be given to decide whether n numbers are different.
(b) If we allow degree n polynomials as test functions then a depth n decision tree can be
given to decide whether n numbers are different. ♦

10.11 Exercise Given are 2n different real numbers: x1 , . . . , xn , y1 , . . . , yn . We want to decide

whether it is true that ordering them by magnitude, there is a xj between every pair of yi ’s.
Prove that this needs Ω(n log n) comparisons. ♦

154
11 Communication complexity
With many algorithmic and data processing problems, the main difficulty is the transport of
information between different processors. Here, we will discuss a model which—in the simplest
case of 2 participating processors—attempts to characterise the part of complexity due to the
moving of data.
Let us be given thus two processors, and assume that each of them knows only part of the
input. Their task is to compute something from this; we will only consider the case when this
something is a single bit, i.e., they want to determine some property of the (whole) input. We
abstract from the time- and other cost incurred by the local computation of the processors; we
consider therefore only the communication between them. We would like to achieve that they
solve their task having to communicate as few bits as possible. Looking from the outside, we
will see that one processor sends a bit ε1 to the other one; then one of them (maybe the other
one, maybe the same one) sends a bit ε2 , and so on. At the end, both processors must “know”
the bit to be computed.
To make it more graphic, instead of the two processors, we will speak of two players, Alice
and Bob. Imagine that Alice is in Europe and Bob in New Zeeland; then the assumption that
the cost of communication dwarfs the cost of local computations is rather realistic.
What is the algorithm in the area of algorithmic complexity is the protocol in the area of
communication complexity. This means that we prescribe for each player, for each stage of the
game where his/her input is x and bits ε1 , . . . , εk were sent so far (including who sent them)
whether the next turn is his/her (this can only depend on the messages ε1 , . . . , εk and not on x;
it must namely be also known to the other player to avoid conflicts), and if yes then—depending
on these—what bit must be sent. Each player knows this protocol, including the “meaning” of
the messages of the other player (in case of what inputs could the other one have sent it). We
assume that both players obey the protocol.
It is easy to give a trivial protocol: Let Alice send Bob the part of the input known to her.
Then Bob can already compute the end result and communicate it to Alice using a single bit.
We will see that this can be, in general, far from the optimum. We will also see that in the area
of communication complexity, some notions can be formed that are similar to those in the area
of algorithmic complexity, and these are often easier to handle.

11.1 Communication matrix and protocol-tree

Let Alice’s possible inputs be a1 , . . . , an and Bob’s possible inputs b1 , . . . , bm (since the local
computation is free it is indifferent for us how these are coded). Let cij be the value to be
computed for inputs ai and bj . The matrix C = (cij )ni=1 m j=1 is called the communication
matrix of the problem in question. This matrix completely describes the problem: both players
know the whole matrix C. Alice knows the index i of a row of C, while Bob knows the index
j of a column of C. Their task is to determine the element cij . The trivial protocol is that
e.g. Alice sends Bob the number i; this means log n bits. (If m < n then it is better, of course,
to proceed the other way.)
Let us see first what a protocol means for this matrix. First of all, the protocol must
determine who starts. Suppose that Alice sends first a bit ε1 . This bit must be determined

155
by the index i known to Alice; in other words, the rows of C must be divided in two parts
according to ε1 = 0 or 1. The matrix C is thus decomposed into two submatrices, C0 and
C1 . This decomposition is determined by the protocol, therefore both players know it. Alice’s
message determines which one of C0 and C1 contains her row. From now on therefore the
problem has been narrowed down to the corresponding smaller matrix.
The next message decomposes C0 and C1 . If the sender is Bob then he divides the columns
into two classes; if it is Alice then she divides the rows again. It is not important that the second
message have the same “meaning”, i.e., that it divide the same rows [columns] in the matrices
C0 and C1 ; moreover, it is also possible that it subdivides the rows of C0 and the columns of
C2 (Alice’s message “0” means that “I have more to say”, and her message “1” that “it is your
turn”).
Proceeding this way, we see that the protocol corresponds to a decomposition of the matrix to
ever smaller submatrices. In each “turn”, every actual submatrix is divided into two submatrices
either by a horizontal or by a vertical split. We will call such a decomposition into submatrices
a guillotine-decomposition. (It is important to note that rows and columns of the matrix
can be divided into two parts in an arbitrary way; their original order plays no role.)
When does this protocol stop? If the players have narrowed down the possibilities to a
submatrix C  then this means that both know that the row or column of the other one belongs
to this submatrix. If from this, they can tell the result in all cases then either all elements of
this submatrix are 0 or all are 1.
In this way, the determination of communication complexity leads to the following combina-
torial problem: in how many turns can we decompose a given 0-1 matrix into matrices consisting
of all 0’s and all 1’s, if in each turn, every submatrix obtained so far can only be split in two,
horizontally or vertically? (If we obtain an all-0 or all-1 matrix earlier we stop splitting it. But
sometimes, it will be more useful to pretend that we keep splitting even this one: formally, we
agree that an all-0 matrix consisting of 0 rows cam be split from an all-0 matrix as well as from
an all-1 matrix.)
We can make the protocol even more graphic with the help of a binary tree. Every point of
the tree is a submatrix of C. The root is the matrix C, its left child is C0 and its right child
is C1 . The two children of every matrix are obtained by dividing its rows or columns into two
classes. The leaves of the tree are all-0 or all-1 matrices.
Following the protocol, the players move on this tree from the root to some leaf. If they are
in some node then whether its children arise by a horizontal or vertical split determines who
sends the next bit. The bit is 0 or 1 according to whether the row [column] of the sender is in
the left or right child of the node. If they arrive to a vertex then all elements of this matrix
are the same and this is the answer to the communication problem. The time requirement
of the protocol is the depth of this tree. The communication complexity of matrix C is the
smallest possible time requirement of all protocols solving it. We denote it by κ(C).
Note that if we split each matrix in each turn (i.e. if the tree is a complete binary tree) then
exactly half of its leaves is all-0 and half is all-1. This follows from the fact that we have split
all matrices of the penultimate “generation” into an all-0 matrix and an all-1 matrix. In this
way, if the depth of the tree is t then among its leaves, there are 2t−1 all-1 (and just as many
all-0). If we stop earlier on the branches where we arrive earlier at an all-0 or all-1 matrix it will
still be true that the number of all-1 leaves is at most 2t−1 since we could continue the branch

156
formall by making one of the split-off matrices “empty”.
This observation leads to a simple but important lower bound on the communication com-
plexity of the matrix C. Let rk(C) denote the rank of matrix C.

(11.1) Lemma
κ(C) ≥ 1 + log rk(C).

Proof Consider a protocol-tree of depth κ(C) and let L1 , . . . , LN be its leaves. These are
submatrices of C. Let Mi denote the matrix (having the same size as C) obtained by writing
0 into all elements of C not belonging to Li . By the previous remark, we see that there are at
most 2κ(C)−1 non-0 matrices Mi ; it is also easy to see that all of these have rank 1. Now,

C = M1 + M 2 + · · · + MN ,

and thus, using the well-known fact from linear algebra that the rank of the sum of matrices is
not greater than the sum of their rank,

rk(C) ≤ rk(M1 ) + · · · + rk(MN ) ≤ 2κ(C)−1 .

This implies the lemma.

(11.2) Corollary If the rows of matrix C are linearly independent then the trivial protocol is
optimal.

Consider a simple but important communication problem to which this result is applicable
and which will be an important example in several other aspects.

(11.3) Example Both Alice and Bob know some 0-1 sequence of length n; they want to decide
whether the two sequences are equal. ♦

The communication matrix belonging to the problem is obviously a 2n × 2n unit matrix.

Since its rank is 2n no protocol is better for this problem than the trivial (n + 1 bit) one.
By another, also simple reasoning, we can also show that almost this many bits must be
communicated not only for the worst input but for almost all inputs:

(11.4) Theorem Consider an arbitary communication protocoll deciding about two 0-1-
sequence of length n whether they are identical, and let h > 0. Then the number of sequences
a ∈ {0, 1}n for the protocol uses fewer than h bits on input (a, a) is at most 2h .

Proof For each input (a, b), let J(a, b) denote the “record” of the protocol, i.e. the 0-1-sequence
formed by the bits sent to each other. We claim that if a = b then J(a, a) = J(b, b); this implies
the theorem trivially since the number of h-length records is at most 2h .
Suppose that J(a, a) = J(b, b) and consider the record J(a, b). We show that this is equal
to J(a, a).
Suppose that this is not so, and let the i-th bit be the first one in which they differ. On the
inputs (a, a), (b, b) and (a, b) not only the first i − 1 bits are the same but also the direction

157
of communication. Alice namely cannot determine in the first i − 1 steps whether Bob has the
sequence a or b, and since the protocol determines for her whether it is her turn to send, it
determines this the same way for inputs (a, a) and (a, b). Similarly, the i-th bit will be sent in
the same direction on all three inputs, say, Alice sends it to Bob. But at this time, the inputs
(a, a) and (b, b) seem to Alice the same and therefore the i-th bit will also be the same, which
is a contradiction. Thus, J(a, b) = J(a, a).
The protocol terminates on input (a, b) by both players knowing that the two sequences are
different. But from Adel’s point of view, her own input as well as the communication are the
same as on input (a, a), and therefore the protocol comes to wrong conclusion on that input.
This contradiction proves that J(a, a) = J(b, b).
One of the main applications of communication complexity is that sometimes we can get a
lower bound on the number of steps of algorithms by estimating the amount of communication
between certain data parts. To illustrate this we give a solution for an earlier exercise. A
palindrome is a string with the property that it is equal to its reverse.

(11.5) Theorem Every 1-tape Turing machine needs Ω(n2 ) steps to decide about a sequence
of length 2n whether it is a palindrome.

Proof Consider an arbitrary 1-tape Turing machine deciding this question. Let us seat Alice
and Bob in such a way that Alice sees cells n, n − 1, . . . , 0, −1, . . . of the tape and Bob sees its
cells n + 1, n + 2, . . .; we show the structure of the Turing machine to both of them. At start,
both see therefore a string of length n and must decide whether these strings are equal (Alice’s
sequence is read in reverse order).
The work of the Turing machine offers a simple protocol to Alice and Bob: Alice mentally
runs the Turing machine as long as the scanning head is on her half of the tape, then she sends
a message to Bob: “the head moves over to you with this and this internal state”. Then Bob
runs it mentally as long as the head is in his half, and then he tells Alice the internal state with
which the head must return to Alice’s half, etc. So, if the head moves over k times from one
half to the other one then they send each other log |Γ| bits (where Γ is the set of states of the
machine). At the end, the Turing machine writes the answer into cell 0 and Alice will know
whether the word is a palindrome. For the price of 1 bit, she can let Bob also know this.
According to Theorem 11.4, we have therefore at most 2n/2 palindroms with k log |Γ| < n/2,
i.e. for most inputs, the head passed between the cells n and (n + 1) at least cn times, where
c = 1/(2 log |Γ|). This is still only Ω(n) steps but a similar reasoning shows that for all h ≥ 0,
with the exception of 2h · 2n/2 inputs, the machine passes between cells (n − h) and (n − h + 1)
at least cn times. For the sake of proving this, consider a palindrom α of length 2h and write
in front of it a sequence β of length n − h and behind it a sequence γ of length n − h. The
sequence obtained this way is a palindrome if and only if β = γ −1 where we denoted by γ −1 the
inversion of γ. By Theorem 11.4 and the above reasoning, for every α there are at most 2n/2
strings β for which on input βαβ −1 , the head passes between cells n − h and n − h + 1 fewer
than cn times. Since the number of α’s is 2h the assertion follows.
If we add up this estimate for all h with 0 ≤ h ≤ n/2 the number of exceptions is at most

2n/2 + 2 · 2n/2 + 4 · 2n/2 + · · · + 2n/2−1 · 2n/2 < 2n ,

158
hence there is an input on which the number of steps is at least (n/2) · (cn) = Ω(n2 ).

11.1 Exercise Show that the following communication problems cannot be solved with fewer
than the trivial number of bits (n + 1). The inputs of Alice and Bob are one subset each of an
n-element set, X and Y . They must decide whether

(a) X and Y are disjoint;

(b) |X ∩ Y | is even.

11.2 Some protocols

In our examples until now, even the smartest communication protocol could not outperform
the trivial one (when Alice sends the complete information to Bob). In this subsection, we
show a few protocols that solve their problems surprisingly cheaply by the tricky organization
of communication.

(11.1) Example Alice and Bob know a subtree each of a (previously fixed) tree T with n
nodes. Alice has subtree TA and Bob has subtree TB . They wan to decide whether the subtrees
have a common point.
The trivial protocol uses obviously log M bits where M is the number of subtrees; M can
even be greater than 2n−1 if e.g. T is a star. (For different subtrees, Alice’s message must be
different. If Alice gives the same message for polygons TA and TA and, say, TA ⊂ TA then TA
has a vertex v that is not in TA ; if Bob’s subtree consists of the single point v then he cannot
find the answer based on this message.)
Consider, however, the following protocol: Alice chooses a vertex x ∈ V (TA ) and sends it to
Bob (we reserve a special message for the case when TA is empty; in this case, they will be done
). If x is also a vertex of the tree TB then they are done (Bob has a special message for this
case). If not then Bob looks up the point of TB closest to x (let this be y) and sends it to Alice.
If this is in TA then Alice knows that the two trees have a common point; if y is not in the tree
TA then the two trees have no common points at all. This protocol uses only 1 + 2log(n + 1)
bits. ♦

11.2 Exercise Prove that in Example 11.1, any protocol requires at least log n bits. ♦
∗ (11.3) Exercise Refine the above protocol to use only log n + log log n + 1 bits. ♦

(11.2) Example Given is a graph G with n points. Alice knows a point set SA spanning a
complete subgraph and Bob knows an independent SB point set in the graph. They want to
decide whether the two subgraphs have a common point.
If Alice wants to give the complete information to Bob about the point set known to her
then log M bits would be needed, where M is the number of complete subgraphs. This can be,
however, even 2n/2 , i.e. (in the worst case) Alice must use Ω(n) bits. The situation is similar
with Bob.

159
 The following protocol is significantly more economical. Alice checks whether the set SA
has a vertex with degree at most n/2 − 1. If there is one then it sends to Bob a 1 and then the
name of such a vertex v. Then both of them know that Alice’s set consists only of v and some
of its neighbors, i.e. they reduced the problem to a graph with n/2 vertices.
If every node of SA has degree larger than n/2 − 1 then Alice sends to Bob only a 0. Then
Bob checks whether the set SB has a point with degree larger than n/2 − 1. If it has then it
sends Alice a 1 and the name of such a node w. Similarly to the foregoing, after this both of
them will know that besides w, the set SB can contain only points that are not neighbors of
w, and they thus again succeeded in reducing the problem to a graph with at most (n + 1)/2
vertices.
Finally, if every vertex of SB has degree at most n/2 − 1, Bob sends a 0 to Alice. After this,
they know that their sets are disjoint.
The above turn uses at most O(log n) bits and since it decreases the number of vertices
of the graph to half, it will be repeated at most log n times. Therefore the complete protocol
is only O((log n)2 ). More careful computation shows that the number of used bits is at most
log n(2 + log n)/2. ♦

11.3 Non-deterministic communication complexity

As with algorithms, the nondeterministic version plays an important role also with proto-
cols. This can be defined—in a fashion somewhat analogous to the notion of “witness”, or
“testimony”—in the following way. We want that for every input of Alice and Bob for which
the answer is 1, a “superior being” can reveal a short 0-1 sequence convincing both Alice and
Bob that the answer is indeed 1. They do not have to believe the revelation of the “supe-
rior being” but if they signal anything at all this can only be that on their part, they accept
the proof. This non-deterministic protocol consists therefore of certain possible “revelations”
x1 , . . . , xn ∈ {0, 1}∗ all of which are acceptable for certain inputs of Alice and Bob. For a given
pair of inputs, there is an xi acceptable for both of them if and only if for this pair of inputs,
the answer to the communication problem is 1. The length of the longest xi is the complexity
of the protocol. Finally, the nondeterministic communication complexity of matrix C is
the minimum complexity of all non-deterministic protocols applicable to it; we denote this by
κN D (C)

(11.1) Example Suppose that Alice and Bob know a polygon each in the plane, and they
want to decide whether the two polygons have a common point. If the superior being wants to
convince the players that their polygons are not disjoint she can do this by revealing a common
point. Both players can check that the revealed point indeed belongs to their polygon.
We can notice that in this example, the superior being can also easily prove the negative
answer: if the two polygons are disjoint then it is enough to reveal a straight line such that
Alice’s polygon is on its left side, Bob’s polygon is on its right side. (We do not discuss here
the exact number of bits in the inputs and the revelations.) ♦

(11.2) Example In Example 11.3, if the superior being wants to prove that the two strings are
different it is enough for her to declare: “Alice’s i-th bit is 0 while Bob’s is not.” This is—apart

160
from the textual part, which belongs to the protocol—only log n + 1 bits, i.e. much less than
the complexity of the optimal deterministic protocol.
We remark that even the superior being cannot give a proof that two words are equal in
fewer than n bits, as we will see right away. ♦

Let x be a possible revelation of the superior being and let Hx be the set of all possible
pairs (i, j) for which x “convinces” the players that cij = 1. We note that if (i1 , j1 ) ∈ Hx and
(i2 , j2 ) ∈ Hx then (i1 , j2 ) and (i2 , j1 ) also belong to Hx : since (i1 , j1 ) ∈ Hx , Alice, possessing i1 ,
accepts the revelation x; since (i2 , j2 ) ∈ Hx , Bob, possessing j2 , accepts the revelation x; thus,
when they have (i1 , j2 ) both accept x, hence (i1 , j2 ) ∈ Hx .
We can therefore also consider Hx as a submatrix of C consisting of all 1’s. The submatrices
belonging to the possible revelations of a nondeterministic protocol cover the 1’s of the matrix
C since the protocol must apply to all inputs with answer 1 (it is possible that a matrix element
belongs to several such submatrices). The 1’s of C can therefore be covered with at most 2κN D (C)
all-1 submatrices.
Conversely, if the 1’s of the matrix C can be covered with 2t all-1 submatrices then it is easy
to give a non-deterministic protocol of complexity t: the superior being reveals only the number
of the submatrix covering the given input pair. Both players verify whether their respective
input is a row or column of the revealed submatrix. If yes then they can be convinced that the
corresponding matrix element is 1. We have thus proved the following statement:

(11.3) Lemma κN D (C) is the smallest natural number t for which the 1’s of the matrix can
be covered with 2t all-1 submatrices.

In the negation of Example 11.2, the matrix C is the 2n ×2n unit matrix. Obviously, only the
1 × 1 submatrices of this are all-1, the covering of the 1’s requires therefore 2n such submatrices.
Thus, the non-deterministic complexity of this problem is also n.
Let κ(C) = s. Then C can be decomposed into 2s submatrices half of which are all-0 and
half are all-1. According to Lemma 11.3 the nondeterministic communication complexity of C
is therefore at most s − 1. Hence

κN D (C) ≤ κ(C) − 1.

Example 11.2 shows that there can be a big difference between the two quantities.
Let C denote the matrix obtained from C by changing all 1’s to 0 and all 0’s to 1. Obviously,
κ(C) = κ(C). Example 11.2 also shows that κN D (C) and κN D (C) can be very different. On
the basis of the previous remarks, we have

max{1 + κN D (C), 1 + κN D (C)} ≤ κ(C).

The following important theorem (Aho, Ullman and Yannakakis) shows that here, already,
the difference between the two sides of the inequality cannot be too great.

(11.4) Theorem
κ(C) ≤ (2 + κN D (C)) · (2 + κN D (C)).

161
 We will prove a sharper inequality. In case of an arbitrary 0-1 matrix C, let (C) denote the
largest number t for which C has a t × t submatrix in which—after a suitable rearrangement of
the rows and columns—there are all 1’s in the main diagonal and all 0’s everywhere above the
main diagonal. Obviously,

(C) ≤ rk(C),

and Lemma 11.3 implies

log (C) ≤ κN D (C).

The following inequality therefore implies theorem 11.4.

(11.5) Theorem
κ(C) ≤ 1 + log (C)(κN D (C) + 2).

Proof We use induction on log (C). If (C) ≤ 1 then the protocol is trivial. Let (C) > 1
and p = κN D (C). Then the 0’s of the matrix C can be covered with 2p all-0 submatrices, say,
M1 , . . . , M2p . We want to give a protocol that decides the communication problem with at most
(p + 2) log (C) bits. The protocol fixes the submatrices Mi , this is therefore known to the
players.
For every submatrix Mi , let us consider the matrix Ai formed by the rows of C intersecting
Mi and the matrix Bi formed by the columns of C intersecting Mi . The basis of the protocol is
the following, very easily verifiable, statement:

(11.6) Claim
(Ai ) + (Bi ) ≤ (C).

11.4 Exercise Prove this claim. ♦

Now, we can prescribe the following protocol:

Alice checks whether there is an index i for which Mi intersects her row and for which
(Ai ) ≤ 12 (C). If yes then she sends “1” and the index i to Bob and the first phase of protocol
has ended. If not then she sends “0”. Now, Bob checks whether there is an index i for which Mi
intersects his column and (Bi ) ≤ 12 (C). If yes then he sends a “1” and the index i to Alice.
Else he sends “0”. Now the first phase has ended in any case.
If either Alice or Bob find a suitable index in the first phas then by the communication of
at most p + 2 bits, they have restricted the problem to a matrix C  (= Ai or Bi ) for which
(C  ) ≤ 12 (C). Hence the theorem follows by induction.
If both players sent “0” in the first phase then they can finish the protocol: the answer is
“1”. Indeed, if there was a 0 in the intersection of Alice’s row and Bob’s column then this would
belong to some submatrix Mi . However, for these submatrices, we have on the one hand
1
(Ai ) > (C)
2

162
(since they did not suit Alice), on the other hand
1
(Bi ) > (C)
2
since they did not suit Bob. But this contradicts the above Claim.
It is interesting to formulate another corollary of the above theorem (compare it with Lemma
11.1):

(11.7) Corollary
κ(C) ≤ 1 + log(1 + rk(C))(2 + κN D (C)).

11.5 Exercise Show that in Theorems 11.4 and 11.5 and in Corollary 11.7, with more careful
planning, the factor (2 + κN D ) can be replaced with (1 + κN D ). ♦

To show the power of Theorems 11.4 and 11.5 consider the examples treated in Subsection
11.2. If C is the matrix corresponding to Example 11.1 (in which 1 means that the subtrees
are disjoint) then κN D (C) ≤ log n (it is sufficient to name a common vertex). It is also easy
to obtain that κN D (C) ≤ 1 + log(n − 1) (if the subtrees are disjoint then it is sufficient to
name an edge of the path connecting them, together with telling that after deleting it, which
component will contain TA and which one TB ). It can also be shown that the rank of C is 2n.
Therefore whichever of the theorems 11.4 and 11.5 we use, we get a protocol using O((log n)2 )
bits. This is much better than the trivial one but is not as good as the special protocol treated
in subsection 11.2.
Let now C be the matrix corresponding to Example 11.2. It is again true that κN D (C) ≤
log n, for the same reason as above. It can also be shown that the rank of C is exactly n.
From this it follows, by Theorem 11.5, that κ(C) = O((log n)2 ) which is (apart from a constant
factor) the best know result. It must be mentioned that what is know for the value of κN D (C),
is only the estimate κN D = O((log n)2 ) coming from the inequality κN D ≤ κ.

(11.8) Remark We can continue dissecting the analogy of algorithms and protocols a little
further. Let us be given a set H of (for simplicity, quadratic) 0-1 matrices. We say that
H ∈ Pcomm if the communication complexity of every matrix C ∈ H is not greater than a
polynomial of log log n where n is the number of rows of the matrix. (I.e., if the complexity is a
good deal smaller than the trivial 1 + log n.) We say that H ∈ NPcomm if the non-deterministic
communication complexity of every matrix C ∈ H is not greater than a polynomial of log log n.
We say that H ∈ co−NPcomm if the matrix set { :C ∈ H } is in NPcomm . Then Example 11.2
shows that

Pcomm = NPcomm ,

and Theorem 11.4 implies

Pcomm = NPcomm ∩ co−NPcomm .

163
11.4 Randomized protocols
In this part, we give an example showing that randomization can decrease the complexity of
protocols significantly. We consider again the problem whether the inputs of the two players
are identical. Both inputs are 0-1 sequences of length n, say x and y. We can also view these
as natural numbers between 0 and 2n − 1. As we have seen, the communication complexity of
this problem is n.
If the players are allowed to choose random numbers then the question can be settled much
easier, by the following protocol. The only change on the model is that both players have a
random number generator; these generate independent bits (it does not restrict generality if we
assume that the bits of the two players are independent of each other, too). The bit computed
by the two players will be a random variable; the protocol is good if this is equal to the “true”
value with probability at least 2/3.

(11.1) Protocol Alice chooses a random prime number p in the interval 1 ≤ p ≤ N and divides
x by p with remainder. Let the remainder by r; then Alice sends Bob the numbers p and r. Bob
checks whether y ≡ r (mod p). If not then he determines that x = y. If yes then he concludes
that x = y. ♦

First we note that this protocol uses only 2 log N bits since 1 ≤ r ≤ p ≤ N . The problem
is that it may be wrong; let us find out in what direction and with what probability. If x = y
then it gives always the right result. If x = y then it is conceivable that x and y give the same
remainder at division by p and so the protocol arrives at a wrong conclusion. This occurs if p
divides the difference d = |x − y|. Let p1 , . . . , pk be the prime divisors of d, then
d ≥ p1 · · · pk ≥ 2 · 3 · 5 · · · · · q,
where q is the k-th prime number.
(Now we will use some number-theoretical facts. For those who are unfamiliar with them but
feel the need for completeness we include a proof of some weaker but still satisfactory versions
of these facts in the next section.)
It is a known number-theoretical fact (see the next section for the proof) that for large
enough q we have, say,
3
2 · 3 · 5 · · · · · q > e 4 q > 2q
(Lovász has 2q−1 .)
Since d < 2n it follows from this that q < n and therefore k ≤ π(n) (where π(n) is the
number of primes up to n). Hence the probability that we have chosen a prime divisor of d can
be estimated as follows:
k π(n)
Prob(p | d) = ≤ .
π(N ) π(N )
Now, according to the prime number theorem, we have π(n) " n/ log n and so if we choose
N = cn then the above bound is asymptotically 1/c), i.e. it can be made arbitrarily small with
the choice of c. At the same time, the number of bits to be transmitted is only 2 log N = 2 log n+
constant.

164
(11.2) Remark The use of randomness does not help in every communication problem this
much. We have seen in one of the exercises that determining the disjointness or the parity of
the intersection of two sets behaves, from the point of view of deterministic protocols, as the
decision of the identity of 0-1 sequences. These problems behave, however, already differently
from the point of view of protocols that also allow randomization: Chor and Goldreich have
shown that Ω(n) bits are needed for the randomized computation of the parity of intersection,
and Kalyanasundaram and Schnittger proved similar lower bound for the randomized
communication complexity of the decision of disjointness of two sets. ♦

165
12 The complexity of algebraic computations
(To be written)

166
13 Circuit complexity
13.1 Introduction
A central, but extremely difficult problem in the theory of computation is to prove lower bounds
on the time and space complexity of various computational tasks. The key problem is whether
P = N P , but much simpler questions remain unanswered. The approach of classical logic,
trying to extend the methods that were used in Chapter 2 to prove undecidability results, seems
to fail badly.
Another, more promising approach to proving lower bounds on the computational com-
plexity of a problem is combinatorial. This approach focuses on the Boolean circuit model of
computation, and tries to analyze the (rather complex) flow of information through the steps
of the computation. We illustrate this method by a beautiful (but rather difficult) proof in this
spirit (it might underline the difficulty of these questions that this is perhaps the easiest proof
to tell in this area!).
We discuss two very simple functions, already introduced in Chapter 1: the majority function

1, if at least n/2 of the variables is 1;
MAJORITY(x1 , . . . , xn ) =
0, otherwise.
and the parity function or XOR function
PARITY(x1 , . . . , xn ) = x1 + x2 + . . . + xn (mod 2).
These functions are of course very easy to compute, but suppose we want to do it in parallel
in very little time? Instead of going into the complications of PRAMs, let us consider a more
general model of parallel computation, namely, Boolean circuits with small depth.
At this point, we have to make an important remark: if we restrict the fan-in (indegree) of
the Boolean circuit to 2, then it is easy to see that we need at least log n depth, just to get
an output node which is dependent on all the inputs (see exercise). However, we will allow
arbitrary fan-in and fan-out (this is analogous to the concurrent-read-concurrent-write model
of parallel computation).
We can recall now from Chapter 1 that every Boolean function can be computed by a
Boolean circuit of depth 2. However, it is easy to see that even for simple functions like the
majority function, the resulting circuit is exponentially large. On the other hand, if we have a
polynomial time algorithm to compute such a function, then this can be converted (again, as
described in Chapter 1) to a Boolean circuit of polynomial size. However. this circuit will have
large (typically linear, if you are careful, logarithmic) depth.
Can we simultaneously restrict the size to polynomial and the depth to less than logarithmic?
The answer is negative even for quite simple functions. In a series of increasingly stronger results,
Furst–Saxe–Sipser, Ajtai, Yao and Hastad proved that every constant-depth circuit computing
the parity function has exponential size, and the every polynomial-size circuit computing the
parity function has (essentially) logarithmic depth. Let us state the result in detail (the proof
is too complex to be given here).

(13.1) Theorem Every circuit with n input bits and depth d that computes the parity function
1/(d−1)
has at least 2(1/10)n gates.

167
 Not much later Razborov proved analogous results for the majority function. In fact, he
proved a stronger result by allowing circuits that may have parity gates, or XOR gates, in
addition to the usual AND, OR and NOT gates, where a parity gate computes the binary sum
of any number of bits. The proof, though not easy, can be reproduced here for the enjoyment
of the truely interested.

13.2 Lower bound for the Majority Function

Let us start with the exact statement of the theorem.

(13.1) Theorem If C is a circuit of depth d, with AND, OR, XOR, and NOT gates that
(1/2d) √
computes the majority function of n input bits, then C has at least 2n /10 n gates.

We would not really need all these types of gates (just like we don’t need both AND and OR
gates in a Boolean circuit). We can assume without loss of generality that the circuit uses only
NOT, XOR and OR gates, since these can be used to simulate AND gates within constant depth
and polynomial size (see exercise).
The idea of the proof is to introduce “approximations” of the gates used during the compu-
tation. Using the approximate gates, instead of the real gates, one computes an approximation
of the majority function. The quality of the approximation will be measured in terms of the
number of inputs on which the modified circuit differs from the original. The main point of
the approximation is to keep the computed function “simple” in some sense. We will show
that every “simple” function, and in particular the approximation we compute, differs from the
majority function on a significant fraction of the inputs. Since the approximation of each gate
has a limited effect on the function computed, we can conclude that many approximations had
to occur.
The proof will be easier to describe if we generalize the result to a family of closely related
functions, the k-threshold functions fk . The k-threshold function is 1 when at least k of the
inputs are 1. It is easy to see that if there is a circuit of size s that computes the majority
function of 2n − 1 elements in depth d, then for each k, 1 ≤ k ≤ n, there is a circuit of
depth d and size at most s that computes the k-threshold function on n elements (see exercise).
Therefore, any exponential lower bound for fk implies a similar bound for the majority function.
We shall consider k = (n + h + 1)/2 for an appropriate h.
Each Boolean function can be expressed as a polynomial over the two-element field GF (2).
In fact, such a representation can be computed following the computation described by the
circuit. If p1 and p2 are polynomials representing two functions, then p1 + p2 is the polynomial
corresponding to the XOR of the two functions. The polynomial p1 p2 corresponds to their AND,
which makes it easy to see that (p1 + 1)(p2 + 1) + 1 corresponds to their OR. The polynomial
1 − p corresponds to the negation of the function represented by the polynomial p.
The measure of simplicity of a Boolean function f for this proof is the degree of the poly-
nomial representing the function or for short, the degree of the function. Note that the inputs
have degree 1, i.e., they are very simple. But the degree may grow very fast as we follow the
circuit; in fact, since we do not restrict fan-in, a single OR gate can produce a function with
arbitrarily high degree!

168
 The trick is to show that these functions can be approximated by polynomials of low degree.
The following lemma will serve as the basis for the approximation.

(13.2) Lemma Let g1 , . . . , gm be Boolean functions of degree at most h. If r ≥ 1 and f =

∨m  n−r
i=1 gi , then there is a function f of degree at most rh that differs from f on at most 2
inputs.

Proof Let us go through the indices 1, 2, . . . , m one by one, and for each such number, flip a
coin. If we get HEAD, we select this number; else, we move on. Let I1 be the set of numbers
selected (so I1 is a random subset of {1, . . . , m}). We repeat this experiment r times, to get the
random subsets I1 , . . . , Ir . Let

fj = gi ,
i∈Ij

and consider f  = ∨rj=1 fj . We claim that the probability that f  satisfies the requirements of
the lemma is non-zero.
It is clear that the degree of the polynomial for f  is at most rh. Furthermore, consider
an input α; we claim that the probability that f  (α) = f (α) is at most 2−r . To see this,
consider two cases. If gi (α) = 0 for every i, then both f (α) = 0 and f  (α) = 0. On the other
hand, if there exists an index i for which gi (α) = 1, then f (α) = 1 and for each j, fj (α) = 0
independently with probability at most 1/2. Therefore, f  (α) = 0 with probability at most 2−r ,
and the expected number of inputs on which f  = f is at most 2n−r . Hence for at least one
particular choice of the sets Ij , the polynomial f  differs from f on at most 2n−r inputs.
Next we show that any function of low degree has to differ from the k-threshold function on
a significant fraction of the inputs.

(13.3) Lemma Let n/2 ≤ k ≤ n. Every polynomial n

with n variables of degree h = 2k − n − 1
differs from the k-threshold function on at least k inputs.

Proof Let g be a polynomial of degree h and let B denote the set of vectors where it differs
from fk . Let A denote the set of all 0-1 vectors of length n containing exactly k 1’s.

For each Boolean function f , consider the summation function fˆ(x) = y≤x f (y). It is
trivial to see that the summation function of the monomial xi1 . . . xir is 1 for the incidence
vector of the set {i1 , . . . , ir } and 0 on all other vectors. Hence it follows that f has degree at
most h if and only if fˆ vanishes on all vectors with more than h 1’s. In contrast to this, the
summation function of the k-threshold fk is 0 on all vectors with fewer than k 1’s, but 1 on all
vectors with exactly k 1’s.
Consider the matrix M = (mab ) whose rows are indexed by the members of A, whose
columns are indexed by the members of B, and

1, if a ≥ b,
mab =
0, otherwise.

We want to show that the

n
columns of this matrix generate the whole space GF (2)A . This will
imply that |B| ≥ |A| = k .

169
 Let a1 , a2 ∈ A and let a1 ∧ a2 denote their coordinatewise minimum. Then we have, by the
definition of B,
    
ma 2 b = 1= (fk (u) + g(u)) = fk (u) + g(u).
b≤a1 b≤a1 ∧a2 u≤a1 ∧a2 u≤a1 ∧a2 u≤a1 ∧a2
b∈B b∈B

The second term of this last expression is 0, since a1 ∧ a2 contains at least h + 1 1’s. The
first term is also 0 except if a1 = a2 . The columns of M therefore generate the unit vector
corresponding to the coordinate a1 , and so they generate the whole space.
It is easy now to complete the proof theorem 13.1. Assume that there is a circuit of size s
and depth d to compute the k-threshold function for inputs of size n. Now apply Lemma 13.2
with r = n1/(2d)  to approximate the OR gates in this circuit. The functions computed by the
gates at the ith level will be approximated by polynomials of degree at most ri . Therefore, each
resulting approximation pk of the k-threshold function will have degree at most rd . Lemma 13.3
implies that for k = (n + rd + 1)/2, the polynomial pk differs from the k-threshold function
n n
on at least k inputs. This shows that s2 n−r ≥ k . From this, routine calculations yield that
 
n r−n 2r
s≥ 2 >√ ,
k πn

which establishes the desired exponential lower bound.

13.3 Monotone circuits

Perhaps the deepest result on circuit complexity was obtained by Razborov in 1985. The main
point is that he does not make any assumption on the depth of the circuit; but, unfortunately,
he still has to make a rather strong restriction, namely monotonicity. A Boolean circuit is
monotone, if all of its input nodes are unnegated variables, and it has no NOT gates. Obviously,
such a circuit can compute only monotone functions; but it is not difficult to see that every
monotone function can be computed by a monotone circuit. The monotonicity of the function
is not a serious restriction; many interesting functions in NP (e.g. matching, clique, colorability

etc.) are monotone. For example, the k-clique function is defined as follows: it has n2 variables
xij , 1 ≤ i < j ≤ n, and its value is 1 if and only if the graph described by the particular setting
of the variables has a clique of size k. Since some of these are NP-complete, it follows that every
problem in NP can be reduced in polynomial time to the computation of a monotone function
in NP.
Razborov gave a superpolynomial lower bound on the monotone circuit complexity of the
clique problem, without any restriction on the depth. This result was strengthened by Andreev,
Alon and Boppana, to an exponential lower bound on the monotone circuit complexity of the
k-clique function.
Unfortunately, while, as we have seen, restricting the functions to monotone functions does
not mean a substantial restriction of generality, the restriction of the circuits to monotone
circuits does. É. Tardos constructed a family of monotone Boolean functions which can be
computed in polynomial time, but need an exponential number of gates if we use a monotone
circuit.

170
13.1 Exercise If a Boolean circuit has fan-in 2, and it computes a function in n variables that
depends on each of its variables, then its depth is at least log n. ♦

13.2 Exercise (a) Show that the negation of a variable can be computed by using only XOR
gates, if we can use the constant 1. (b) Show that the AND of n variables can be computed by
a circuit using NOT, OR and XOR gates, with constant depth and polynomial size. ♦

13.3 Exercise Suppose that there is a circuit of size s that computes the majority function of
2n − 1 elements in depth d. Show that for each k, 1 ≤ k ≤ n, there is a circuit of depth d and
size at most s that computes the k-threshold function on n elements. ♦

171
14 An application of complexity: cryptography
The complexity of a phenomenon can be the main obstacle of exploring it. Our book—we
hope—proves that complexity is not only an obstacle to research but also an important and
exciting subject. It goes, however, beyond this: it has applications where it is precisely the
complexity of a phenomenon that is exploited. We have discussed the problem of generating
pseudorandom numbers in Chapter 8. This chapter treats another such subject: cryptography,
i.e., the science of secret codes. It was the application of the results of complexity theory that
elevated secret codes beyond the well-known (military, intelligence) applications, and made them
one of the most important ingredient of computer security, electronic trade, internet etc.

14.1 A classical problem

Sender wants to send a message x to Receiver (where x is e.g. a 0-1-sequence of length n). The
goal is that when the message gets into the hands of any unauthorized third party, she should
not understand it. For this, we “code” the message, which means that instead of the message,
Sender sends a code y of it, from which the receiver can recompute the original message but the
unauthorized interceptor cannot. For this, we use a key d that is (say) also a 0-1-sequence of
length n. Only Sender and Receiver know this key.
Thus, Sender computes a “code” y = f (x, d) that is also a 0-1-sequence of length n. We
assume that for all d, f (·, d) is a bijective mapping of {0, 1}n to itself. Then f −1 (·, d) exists and
thus Receiver, knowing the key d, can reconstruct the message x. The simplest, frequently used
function f is f (x, d) = x ⊕ d (bitwise addition modulo 2).

(14.1) Remark This so-called “one-time pad” method is very safe. It was used during World
War II for communication between the American President and the British Prime Minister. Its
disadvantage is that it requires a very long key. It can be expensive to make sure that Sender
and Receiver both have such a common key; but note that the key can be sent at a safer time
and by a completely different method than the message. ♦

14.2 A simple complexity-theoretic model

Let us look at a problem now that has—apparently—nothing to do with the above one. From
a certain bank, we can withdraw money using an ATM. The client types his name or account
number (in practice, he inserts a card on which these data are stored) and a password. The
bank’s computer checks whether this is indeed the client’s password. If this checks out the
automaton hands out the desired amount of money. In theory, only the client knows this
password (it is not written on his card!), so if he takes care that nobody else can find it out,
then this system provides complete security.
The problem is that the bank must also know the password and therefore a bank employee
can abuse it. Can one design a system in which it is impossible to figure out the password,
even with the knowledge of the complete password-checking program? This seemingly self-
contradictory requirement is satisfiable!
Here is a solution (not a very efficient one, and certainly never used in practice, but one that
illustrates the idea how complexity theory enters this field). The client takes n nodes numbered

172
from 1 to n, draws in a random Hamiltonian circuit and then adds arbitrary additional edges.
He remembers the Hamiltonian circuit; this will be his password. He gives the whole graph to
the bank (without marking the Hamiltonian circuit in it).
If somebody shows up at the bank in the name of the client and gives a set of edges on the
n nodes as her password, the bank checks it whether it is a Hamiltonian circuit of the graph
stored there. If so, the password will be accepted; if not, it will be rejected.
Now it seems that we have made it easier to impersonate our client: the imposter does
not have to know the password (the particular Hamiltonian circuit); she can give any other
Hamiltonian circuit of the client’s graph. But note that even if she learns the graph, she must
still solve the problem of finding a Hamiltonian circuit in a graph. And this is NP-hard!

(14.1) Remarks

1. Instead of the Hamiltonian circuit problem, we could have based the system on any other
NP-complete problem.

2. We glossed over a difficult question: how many more edges should the client add to the
graph and how? The problem is that the NP-completeness of the Hamiltonian circuit
problem means only that its solution is hard in the worst case. We don’t know how to
construct one graph in wich there is a Hamiltonian circuit but it is hard to find.
It is a natural idea to try to generate the graph by random selection. If we chose it
randomly from among all n-point graphs then it can be shown that in it, with large
probability, it is easy to find a Hamiltonian circuit. If we chose a random one among all
n-point graphs with m edges then the situation is similar both with too large m and with
too small m. The case m = n log n at least seems hard. In some cases, one can show that
certain randomized constructions yield instances of NP-hard problems that are hard with
high probability (in the sense that if one could solve a random instance in polynomial time
with non-negligible probability, then we could solve all instances in randomized polynomial
time). These studies are beyond the scope of this book.

14.3 Public-key cryptography

In this section, we describe a system that improves on the methods of classical cryptography
in several points. Let us note first of all that the system intends to serve primarily civil rather
than military goals. For using electronic mail, in particular, if we use it for electronic commerce,
we must recreate some tools of traditional correspondence like envelope, signature, company
letterhead, etc.
The system has N ≥ 2 participants. Every participant has a public key ei (she will publish it
e.g. in a phone-book-like directory) and a secret key di known to herself only. There is, further,
a publicly known encoding/decoding function that computes from every message x and (secret
or public) key e a message f (x, e). (The message x and its code must come from some easily
specifiable set H; this can be e.g. {0, 1}n but can also be the set of residue classes modulo m.

173
We assume that the message itself contains the names of the sender and receiver also in “human
language”.) For every x ∈ H and every i with 1 ≤ i ≤ N , we must have
f (f (x, ei ), di ) = f (f (x, di ), ei ) = x. (14.1)
If participant i wants to send a message to j then she sends the message y = f (f (x, di ), ej )
instead. From this, j can compute the original message by the formula x = f (f (y, dj ), ei ).
For this system to be useable, trivially it must satisfy
(C1) f (x, ei ) can be computed efficiently from x and ei .
The security of the system will be guaranteed by
(C2) f (x, di ) cannot be computed efficiently even in the knowledge of x, ei and an arbitrary
number of dj1 , . . . , djh (jr = i).
By “efficient”, we mean polynomial time, but the system makes sense under other resource-
bounds too. A function f with the above properties will be called a trapdoor function.
Condition (C1) guarantees that if participant i sends a message to participant j then she
can encode it in polynomial time and the addressee can decode it in polynomial time. Condition
(C2) can be interpreted to say that if somebody encoded a message x with the public key of
a participant i and then she lost the original then no coalition of the participants can restore
the original (efficiently) if i is not among them. This condition provides the “security” of the
system. It implies, besides the classical requirement, a number of other security conditions.
(14.2) Only j can decode a message addressed to j.

Proof Assume that a band k1 , . . . , kr of unauthorized participants finds the message

f (f (x, di ), ej ), and knows even who sent it to whom. Suppose that they can compute x ef-
ficiently from this. Then k1 , . . . , kr and i together could compute x also from f (x, ej ). Let,
namely, z = f (x, ej ); then k1 , . . . , kr and i knows the message f (x, ej ) = f (f (z, di ), ej ) and
thus using the method of k1 , . . . , kj , can compute z. But from this, they can compute x by the
formula x = f (z, di ), which contradicts condition (C2).
The following can be verified by similar reasoning:
(14.3) Nobody can forge a message in the name of i, i.e. participant j receiving a message
that he can successfully decode using the public key of i (and his own private key), can be sure
that the message could have been sent only by i.
(14.4) j can prove to a third person (e.g. in a court of justice) that i has sent the given
message; in the process, the secret elements of the system (the keys di ) need not be revealed.
(14.5) j cannot change the message (and have it accepted e.g. in a court as coming from i)
or send it in the name of i to somebody else.
It is not at all clear, of course, whether trapdoor functions exists. Several such function have
been proposed; many of the proposed systems turned out insecure later on—the corresponding
complexity conditions were not true.) In the next subsection, we describe one such system that
is one of the earliest, and is most widely used (and of course, to our current knowledge, is
secure).

174
14.4 The Rivest-Shamir-Adleman code
In a simpler version of this system (in its abbreviated form, the RSA code), the “post office”
generates two n-digit prime numbers, p and q for itself, and computes the number m = pq. It
publishes this number (but the prime decomposition remains secret!). Then it generates, for
each subscriber, a number ei with 1 ≤ ei < m that is relatively prime to (p − 1) and (q − 1). (It
can do this by generating a random ei between 0 and (p−1)(q−1) and checking by the Euclidean
algorithm whether it is relatively prime to (p − 1)(q − 1). If it is not, it tries a new number. It
is easy to see that after an expected number of log n trials, it finds a good number ei with high
probability.) Then, using the Euclidean algorithm, it finds a number di with 1 ≤ di < m such
that

ei di ≡ 1 (mod (p − 1)(q − 1)).

(here (p − 1)(q − 1) = ϕ(m), the number of positive integers smaller than m and relatively prime
to it). The public key is the number ei , the secret key is the number di . The message x itself
is considered a natural number with 0 ≤ x < m (if it is longer then it will be cut into pieces).
The encoding function is defined by the formula

f (x, e) = xe (mod m) 0 ≤ f (x, e) < m.

The same formula serves for decoding, only with d in place of e.

The inverse relation between coding and decoding (formula 14.1) follows from the “little”
Fermat theorem. By definition, ei di = 1 + ϕ(m)r = 1 + r(p − 1)(q − 1) where r is a natural
number. Thus, if (x, p) = 1 then

f (f (x, ei ), di ) ≡ (xei )di = xei di = x(xp−1 )r(q−1) ≡ x (mod p).

On the other hand, if p|x then obviously

xei di ≡ 0 ≡ x (mod p).

Thus

xei di ≡ x (mod p)

holds for all x. It similarly follows that

xei di ≡ x (mod q),

and hence

xei di ≡ x (mod m).

Since both the first and the last number are between 0 and m − 1 it follows that they are equal,
i.e. f (f (x, ei ), di ) = x.
It is easy to check condition (C1): knowing x and ei and m, the remainder of xei after
division by m can be computed in polynomial time, as we have seen it in Chapter 3. Condition
(C2) holds only in the following, weaker form:

175
(C2’) f (x, di ) cannot be computed efficiently from the knowledge of x and ei .
This condition can be formulated to say that with respect to a composite modulus, extracting the
ei -th root cannot be accomplished in polynomial time without knowing the prime decomposition
of the modulus. We cannot prove this condition (even with the hypothesis P=NP) but at least
it seems true according to the present state of number theory.
Several objections can be raised against the above simple version of the RSA code. First of
all, the post office can decode every message, since it knows the numbers p, q and the secret keys
di . But even if we assume that this information will be destroyed after setting up the system,
unauthorized persons can still misuse the system. In fact, Every participant of the system can
solve any message sent to any other participant. (This does not contradict condition (C2’) since
participant j of the system knows, besides x and ei , also the key dj .)
Indeed, consider participant j and assume that she got her hands on the message z =
f (f (x, di ), ek ) sent to participant k. Let y = f (x, di ). Participant j solves the message not
meant for her as follows. She computes a factoring u · v of (ej dj − 1), where (u, ek ) = 1 while
every prime divisor of v also divides ek . To do this, she computes, by the Euclidean algorithm,
the greatest common divisor v1 of of ei and ej dj − 1, then the greatest commond divisor v2
of ek and (ej dj − 1)/v1 , then the greatest commond divisor v3 of (ej dj − 1)/(v1 v3 ), etc. This
process terminates in at most t = log(ej dj − 1) steps, i.e. vt = 1. Then v = v1 · · · vt and
u = (ej dj − 1)/v gives the desired factoring.
Notice that (ϕ(m), ek ) = 1 and therefore (ϕ(m), v) = 1. Since ϕ(m)|ej dj − 1 = uv, it follows
that ϕ(m)|u. Since (u, ek ) = 1, there are natural numbers s and t with sek = tu + 1. Then
z s ≡ y sek = y(y u )t ≡ y (mod m)
and hence
x ≡ y ei ≡ z ei s .
Thus, participant j can also compute x.

14.1 Exercise Show that even if all participants of the system are honest an outsider can
cause harm as follows. Assume that the outsider gets two versions of one and the same letter,
sent to two different participants, say f (f (x, di ), ej ) and f (f (x, di ), ek ) where (ej , ek ) = 1 (with
a little luck, this will be the case). Then he can reconstruct the text x. ♦

Now we descibe a better version of the RSA code. Every participant generates two n-digit
prime numbers, pi and qi and computes the number mi = pi qi . Then she generates for herself
a number ei with 1 ≤ ei < mi relatively prime to (pi − 1) and (qi − 1). With the help of the
Euclidean algorithm, she finds a number di with 1 ≤ di < mi for which
ei di ≡ 1 (mod (pi − 1)(qi − 1))
(here, (pi − 1)(qi − 1) = ϕ(mi ), the number of positive integers smaller than mi and relatively
prime to it). The public key consists of the pair (ei , mi ) and the secret key of the pair (di , mi ).
The message itself will be considered a natural number. If 0 ≤ x < mi then the encoding
function will be defined, as before, by the formula
f (x, ei , m) ≡ xei (mod mi ), 0 ≤ f (x, ei , mi ) < mi .

176
Since, however, different participants use different moduli, it will be practical to extend the
definition to a common domain, which can even be chosen to be the set of natural numbers. Let

x be written in a base mi notation: x = j xj mji , and compute the function by the formula

f (x, ei , mi ) = f (xj , ei , mi )mji .
j

We define the decoding function similarly, using di in place of ei .

For the simpler version it follows, similarly to what was said above, that these functions are
inverses of each other, that (C1) holds, and that it can also be conjectured that (C2) holds. In
this version, the “post office” holds no non-public information, and of course, each key dj has
no information on the other keys. Therefore the above mentioned errors do not occur.

177