Theoretical

Computer
Science
An Introduction

Markus Bläser
Universität des Saarlandes
Draft—February 10, 2015 and forever
2

c Markus Bläser 2007–2015
 Part I

Computability

3
4

c Markus Bläser 2007–2015
0 Prelude: Sets, relations, words

0.1 Sets
A set is a “collection of objects”. While this is a rather naive definition that
has some pitfalls, it is sufficient for our needs here. Axiomatic set theory is
beyond the scope of this lecture. A set A1 that contains the numbers 2, 3,
and 5 is denoted by A1 = {2, 3, 5}. 2, 3, and 5 are the elements of A. We
write 2 ∈ A1 to indicate that 2 is an element of A1 and 4 ∈ / A1 to indicate
that 4 is not. If a set A contains only a finite number of elements, then we
call it finite and the number of elements is the size of A, which is denoted
by |A|.
Infinite sets cannot be denoted like this, we denote them for instance as
{0, 1, 2, 3, . . . } and hope for your intuition to fill in the remaining numbers.
The last set, as you already guessed, is the set of natural numbers. We will
also use the symbol N for it. Note that 0 ∈ N.
A set B is a subset of A if every element of B is also an element of A.
In this case, we write B ⊆ A. There is one distinguished set that contains
no elements, the empty set; we denote it by ∅. The empty set is a subset of
every other set.
There are various operations on sets. The union A ∪ B of two sets A and
B is the set that contains all elements that are in A or in B. The intersection
A ∩ B is the set that contains all elements that are simultaneously in A and
in B. For instance, the union of {2, 3, 5} and {2, 4} is {2, 3, 4, 5}, their
intersection is {2}. The cartesian product A × B is the set of ordered tuples
{(a, b) | a ∈ A, b ∈ B}. The cartesian product of {2, 3, 5} and {2, 4} consists
of the six tuples {(2, 2), (2, 4), (3, 2), (3, 4), (5, 2), (5, 4)}.

0.2 Relations
Let A and B be two sets. A (binary) relation on A and B is a subset
R ⊆ A × B. If (a, b) ∈ R, then we will say that a and b stand in relation R.
Instead of (a, b) ∈ R, we will sometimes write aRb. While this looks weird
when the relation is called R, let us have a look a the following example.

Example 0.1 R1 = {(1, 1), (1, 2), (1, 3), (2, 2), (2, 3), (3, 3)} ⊆ {1, 2, 3} ×
{1, 2, 3} is a relation on {1, 2, 3}.

(a, b) ∈ R1 means that a is less than or equal to b. In this case, a ≤ b

looks much better than (a, b) ∈≤.

5
6 0. Prelude: Sets, relations, words

Consider a relation R on a set A and itself, i.e., R ⊆ A × A. R is called

reflexive if for all a ∈ A, (a, a) ∈ R. R1 is reflexive, since (1, 1), (2, 2), (3, 3) ∈
R1 .
R is called symmetric if (a, b) ∈ R implies (b, a) ∈ R, too, for all (a, b) ∈
A × A. The relation R1 is not symmetric, since for instance (1, 2) ∈ R1 but
(2, 1) ∈
/ R1 .
R is called antisymmetric if (a, b) ∈ R and a 6= b implies (b, a) ∈ / R.
R1 is antisymmetric, since we do not consider tuples of the form (a, a).
There are relations that are neither symmetric nor antisymmetric. R2 =
{(1, 2), (1, 3), (3, 1)} is such an example. It is not symmetric, since (1, 2) ∈ R2
but not (2, 1) ∈ R2 . It is not antisymmetric, because (1, 3) and (3, 1) are
both in R2 .
A relation R is called transitive if for all a, b, c ∈ A, (a, b) ∈ R and
(b, c) ∈ R implies (a, c) ∈ R. R1 is transitive. The only pairs that we have
to check are (1, 2) and (2, 3), since these are the only with three different
elements and one element in common. But (1, 3) is in R1 , too.
A relation that is reflexive, antisymmetric, and transitive is called partial
order. A partial order R is called total order if for all a, b ∈ A with a 6= b,
either (a, b) ∈ R or (b, a) ∈ R. For instance, R1 is a partial order, it is even
a total order. Relations R that are orders, well, order the elements in A.
If (a, b) ∈ R, then we also say that a is smaller than b with respect to R.
Elements a such that there does not exist any b 6= a with (b, a) ∈ R are
called minimal ; elements a such that there does not exist any b 6= a with
(a, b) ∈ R are called maximal. With respect to R1 , 1 is a minimal and 3 is a
maximal element. If an order is total than it has at most one minmal and at
most one maximal element. (The relation ≤ on N does not have a maximal
element for instance.)

Exercise 0.1 Show that if R is an order, then R is acyclic, i.e., there does
not exist a sequence a1 , . . . , ai of pairwise distincts elements such that

(a1 , a2 ), (a2 , a3 ), . . . , (ai−1 , ai ), (ai , a1 ) ∈ R.

Another important type of relations are equivalence relations. These are

relations that are reflexive, symmetric and transitive. R1 is not an equiv-
alence relation, it is not symmetric. Neither is R2 since it is not reflexive,
among other things.

Example 0.2 R3 = {(1, 1), (2, 2), (3, 3)} is the equality relation on {1, 2, 3}.

R3 is an equivalence relation. It is easy to check that it is reflexive and

it is trivially transitive and symmetric since there are no (a, b) ∈ R3 with
a 6= b. If R is an equivalence relation, then the equivalence class [a]R of an
element a ∈ A is {b ∈ A | (a, b) ∈ R}. If R is clear from the context, we will
often write [a] for short.

c Markus Bläser 2007–2015
0.3. Functions 7

Exercise 0.2 Proof the following: The equivalence classes of an equivalence

relation R on A are a partition of A, i.e., every a ∈ A belongs to exactly one
equivalence class of R.

0.3 Functions
A relation f on A × B is called a partial function if for all a ∈ A there is at
most one b ∈ B such that (a, b) ∈ f . If such a b exists, we also write f (a) = b.
If such a b does not exist, we say that f (a) is undefined. Occasionally, we
write f (a) = undefined, although this not quite correct, since it suggests
that “undefined” is an element of B. dom f denotes the set of all a such that
f (a) is defined. f is called total if dom f = A.
A total function f is called injective if for all a, a0 ∈ A, f (a) = f (a0 )
implies a = a0 . The image im f of a function is the set {f (a) | a ∈ A}. f is
called surjective if im f = B.
A total function f is called bijective if it is injective and surjective. In
this case, for every a ∈ A there is exactly one b ∈ B such that f (a) = b. In
this case, we can define the inverse f −1 of f , which is a function B → A:
f −1 (b) = a if f (a) = b.

0.4 Words
Let Σ be a finite nonempty set. In the context of words, Σ is usually called
an alphabet. The elements of Σ are called symbols or letters. A (finite)
word w over Σ is a finite sequence of elements from Σ, i.e, it is a function
w : {1, . . . , `} → Σ for some ` ∈ N. ` is called the length of w. The length
of w will also be denoted by |w|. There is one distinguished word of length
0, the empty word. We will usually denote the empty word by ε. Formally
it is the sequence ∅ → Σ.

Example 0.3 Let u : {1, 2, 3} → {a, b, c} be defined by u(1) = a, u(2) = b

and u(3) = a. u is a word over the alphabet {a, b, c} of length 3.

Let w : {1, . . . , n} → Σ be some word. Often, we will write w in a

compact form as w(1)w(2) . . . w(n) and instead of w(i), we write wi . Thus
we can write w in an even more compact form as w1 w2 . . . wn . wi is also
called the ith symbol of w. The word u from the example above can be
writen as aba.
The set of all words of length n over Σ is denoted by Σn . This usually
denotes the set of all n-tuples with entries from Σ, too, but this is fine, since
there
S∞is a natural bijection between sequences of length n and n-tuples. The
set n=0 Σn of all finite words is denoted by Σ∗ .
One important operation is concatenation. Informally, it is the word
that we get when juxtaposing two words. Formally it is defined as follows.

c Markus Bläser 2007–2015
8 0. Prelude: Sets, relations, words

Let w : {1, . . . , `} → Σ and x : {1, . . . , k} → Σ be two words. Then the

concatenation of w and x is the function

{1, . . . , ` + k} → Σ
(
w(i) if 1 ≤ i ≤ `
i 7→
x(i − `) if ` + 1 ≤ i ≤ ` + k.

We denote the concatenation of w and x by wx. Let v = ca. Then the

concatenation of u from the example above and v is abaca. Concatenation
is associative but not commutative. The empty word is a neutral element,
i.e., εw = wε = w for all words w.

Exercise 0.3 Let Σ be an alphabet. Prove the following:

1. The relation “the length of a is smaller than or equal to the length of

b” is a partial order on Σ∗ . If |Σ| = 1, then the order is even total.

2. The relation “the length of a equals the length of b” is an equivalence

relation.

Exercise 0.4 Let Σ be an alphabet and let R be a total order on Σ. The

lexicographic order ≤R n
lex with respect to R on Σ is defined as follows:

u1 u2 . . . un ≤R
lex v1 v2 . . . vn if ui Rvi where i = min{1 ≤ j ≤ n | uj 6= vj }
or i does not exist.

Show that ≤R
lex is indeed a total order.

c Markus Bläser 2007–2015
1 Introduction

During the last year you learnt what a computer can do: how to model a
problem, how to develop an algorithm, how to implement it and how to test
your program. In this lecture, you will learn what a computer cannot do.
Not just because you are missing some particular software or you are using
the wrong operating system; we will reason about problems that a computer
cannot solve no matter what.
One such task is verification: Our input is a computer program P and
an input/output specification S.1 S describes the desired input/output be-
haviour. We shall decide whether the program fulfills the specification or
not. Can this task be automated? That means, is there a computer pro-
gram V that given P and S returns 1, if P fulfills the specification and 0
otherwise? V should do this correctly for all possible pairs P and S.
Let us consider a (seemingly) easier task. Given a program P , decide
whether P returns the value 0 on all inputs or not. That means, is there a
program Z that given P as an input returns 1 if P returns the value 0 on all
inputs and returns 0 if there is an input on which P does not return 0. How
hard is this task? Does such a program Z exists? The following program
indicates that this task is very hard (and one of the goals of the first part of
this lecture is to show that it is impossible in this general setting).2
Program 1 expects four natural numbers as inputs that are initially stored
in the variables x0 , . . . , x3 . We do not specify its semantic formally at this
point, but I am sure that you understand what Program 1 does. This pro-
gram returns 1 on some input if and only if there are natural numbers x0 > 2
and x1 , x2 , x3 ≥ 1 such that xx1 0 + xx2 0 = xx3 0 . The famous Fermat’s last theo-
rem states that such four numbers do not exist. It took almost four hundred
years until a valid proof of this conjecture was given.

Excursus: Fermat’s last theorem

Pierre de Fermat (born 1607/08 in Beaumont de Lomagne, died 1665 in Castres,
France) was a French lawyer who pursued mathematics as a “hobby”. Nevertheless,
he is regarded as one of the greatest mathematicians. Fermat’s last theorem states
that the Diophantine equation xn + y n = z n does not have any integer solution
1
This is just at an intuitive level. We will formalize this soon.
2
Now it is—especially since you are studying at Saarland University—a legitimate
question why we should think about verification at all, when it is impossible. Well you
should think about verification, because it is impossible. This is the real fun! While
verification is impossible in general, it is still possible to solve interesting special cases or
verify one particular program.

9
10 1. Introduction

Program 1 Fermat
Input: x0 , . . . , x3
1: if x0 ≤ 2 then
2: return 0
3: fi
4: if x1 = 0 or x2 = 0 or x3 = 0 then
5: return 0
6: fi
x x x
7: if x1 0 + x2 0 = x3 0 then
8: return 1
9: else
10: return 0
11: fi

for n ≥ 3 and x, y, z > 0. The case n = 1 is of course trivial and for n = 2, the
equation is fulfilled by all Pythagorean triples. Although it was always called Fer-
mat’s last theorem, it was an unproven conjecture written by Fermat in the margin
of his copy of the Ancient Greek text Arithmetica by Diophantus. This note was
discovered posthumously. The last part of this note became famous: “[...] cuius
rei demonstrationem mirabilem sane detexi. Hanc marginis exiguitas non caperet”
(I have discovered a truly marvelous proof of this proposition. This margin is too
narrow to contain it.) Fermat’s last theorem was finally proven in 1995 by Andrew
Wiles with the help of Richard Taylor.

Further reading: Simon Singh, Fermats letzter Satz, dtv, 2000.

Exercise 1.1 Goldbach’s conjecture states that every even number ≥ 4 is

the sum of two primes. Show that if you were able to write the program Z,
then you could prove or disprove Goldbach’s conjecture.

c Markus Bläser 2007–2015
2 WHILE and FOR programs

We want to prove mathematical statements about programs. Specifications

of modern imperative programming languages like C++ or JAVA fill sev-
eral hundred pages and do typically not specify everything. (You might
remember some painful experiences.) As a first step, we will define a simple
programming language called WHILE and specify its semantic completely
in a mathematical model such that we can prove theorems about these pro-
grams. Then we will convince ourselves that C++ or JAVA programs cannot
compute more than WHILE programs.

2.1 Syntax
Let us start with defining the syntax of WHILE programs. WHILE programs
are strings over some alphabet. This alphabet contains

variables: x0 , x1 , x2 , . . .
constants: 0, 1, 2, . . .
key words: while, do, od
other symbols: :=, 6=, ; , +, −

Note that every variable is a symbol on its own and so is every constant.
Also “:=” is treated as one symbol, we just write it like this in reminiscence of
certain programming languages. (No programming language discrimination
is intended. Please do not send any complaint emails like last year.)

Definition 2.1 WHILE programs are defined inductively:

1. A simple statement is of the form

xi := xj + xk or
xi := xj − xk or
xi := c,

where i, j, k ∈ N and c ∈ N.

2. A WHILE program P is either a simple statement or of the form

(a) while xi 6= 0 do P1 od or
(b) P1 ; P2

11
12 2. WHILE and FOR programs

for some i ∈ N and WHILE programs P1 and P2 .

We call the set of all WHILE programs W. We give this set a little more
structure. The set of all WHILE programs that consist of only one simple
statement is called W0 . We inductively define the sets Wn as follows:

Wn = Wn−1 ∪ {P | ∃P1 ∈ Wn−1 , xi ∈ X such that P = while xi 6= 0 do P1 od or

∃P1 ∈ Wj , P2 ∈ Wk with j + k ≤ n − 1 such that P = P1 ; P2 }

Exercise 2.1 Show the following: A WHILE program P is in Wn if and

only if it was built by applying a rule from 2. in Definition 2.1 at most n
times.

Before we explain the semantics of WHILE programs, we first define

another simple language, the FOR language. It uses the same elements as
WHILE programs, we only have a different type of loop.

Definition 2.2 FOR programs are defined inductively:

1. A simple statement is of the form

xi := xj + xk or
xi := xj − xk or
xi := c,

where i, j, k ∈ N and c ∈ N.

2. A FOR program P is either a simple statement or it is of the form

(a) for xi do P1 od or
(b) P1 ; P2

for some i ∈ N and FOR programs P1 and P2 .

The set of all FOR programs is denoted by F. We define the subset Fn

in the same manner as we did for W. FOR programs differ from WHILE
programs just by having a different type of loop.

2.2 Semantics
A program P gets a number of inputs α0 , . . . , αs−1 ∈ N. The input is stored
in the variables x0 , . . . , xs−1 . The output of P is the content of x0 after the
execution of the program. The set X = {x0 , x1 , x2 , . . . } of possible variables
is infinite, but each WHILE or FOR program P always uses a finite number
of variables. Let ` = `(P ) denote the largest index of a variable in P . We
always assume that ` ≥ s − 1. A state is a vector S ∈ N`+1 . It describes the

c Markus Bläser 2007–2015
2.2. Semantics 13

content of the variables in P . If α0 , . . . , αs−1 is our input, then the initial

state S0 will be (α0 , . . . , αs−1 , 0, . . . , 0).
While this looks fine, there is one small annoying inaccuracy. If P =
P1 ; P2 and P1 uses variables x0 , x3 , and x7 and P2 uses x1 , x2 , and x9 , then
a state of P1 is a vector of length 8 but states of P2 and P have length
10. So a state of P1 is formally not a state of P . But it becomes a state
of P if we append two more entries to the vector, filled with 0. The “right”
mathematical object to model the states of a WHILE program are sequences
of natural numbers with finite support, i.e., functions S : N → N such that
there is an `0 ∈ N such that S(n) = 0 for all n ≥ `0 . Since only the values
of S up to `0 are interesting, we will just write down the values up to `0 and
often, we treat S as a finite vector in N`0 +1 .
Given a state S and a program P ∈ Wn , we will now describe what
happens when running P starting in S. This is done inductively, i.e., we
assume that we already know how programs behave that are “smaller” than
P where “smaller” means that they are built by less applications of the second
rule in Definition 2.1 (or Definition 2.2), i.e., these programs are in Wn−1 .
We now define a function Φ that defines the semantics of WHILE/FOR
programs. Let ΦP (S) denote the state that is reached after running P on
state S = (σ0 , . . . , σ` ). ΦP will be a partial function, i.e., ΦP (S) might be
undefined (in the case where P does not halt on S). WHILE programs may
not halt; at an intuitive level—we did not define the semantics so far—
1: x1 := 1;
2: while x1 6= 0 do
3: x1 := 1;
4: od

is such a program.1 The while loop never terminates and there is no state
that can be reached since there is no “after the execution of the program”.

1. If P is a simple statement then


(σ0 , . . . , σi−1 , σj + σk , σi+1 , . . . , σ` )
 if P is xi := xj + xk
ΦP (S) = (σ0 , . . . , σi−1 , max{σj − σk , 0}, σi+1 , . . . , σ` ) if P is xi := xj − xk

(σ0 , . . . , σi−1 , c, σi+1 , . . . , σ` ) if P is xi := c


2. (a) Assume P equals while xi 6= 0 do P1 od. Let r be the smallest

(r)
r ∈ N such that ΦP1 (S) is either undefined, which means that P1

1
The assignment within the loop is necessary because the empty program is no valid
WHILE program. There is no particular reason for this, we just defined it like this.

c Markus Bläser 2007–2015
14 2. WHILE and FOR programs

(r)
does not terminate, or the ith position in ΦP1 (S) equals 0.2 Then
(
(r) (r)
ΦP1 (S) if r exists and ΦP1 (S) is defined
ΦP (S) =
undefined otherwise
(b) If P is P1 ; P2 for WHILE programs P1 and P2 , then
(
ΦP2 (ΦP1 (S)) if ΦP1 (S) and ΦP2 (ΦP1 (S)) are both defined
ΦP (S) =
undefined otherwise

Lemma 2.3 For every program P , ΦP is well-defined.3

Proof overview: The proof will be done by structural induction. Structural
induction means that we assume that the statement is true for “simpler”
programs and we show that the statement is valid for P , too. If P ∈ Wn ,
then simpler will mean that the programs are in Wn−1 . In this way, structural
induction becomes ordinary induction, since we now can just do induction
on n.
Proof. Induction base: If P ∈ W0 , then ΦP (S) is defined for every S
and it is only defined once; thus it is well-defined. This shows the induction
basis.
Induction step: Let P ∈ Wn \ Wn−1 for n > 0. We assume that ΦQ is
well-defined for all programs Q ∈ Wn−1 . Since n > 0, P is either while
xi 6= 0 do P1 od or P1 ; P2 for P1 , P2 ∈ Wn−1 . By the induction hypothesis,
ΦP1 and ΦP2 are well-defined. If P is of the former form, then ΦP (S) is also
well-defined for every S by the rules in 2.(a). In the case that P = P1 ; P2
there is a caveat: There could be another decomposition of P = P10 ; P20 with
P1 6= P10 . We have to show that in both cases, ΦP will be the same function,
i.e., that for all S, ΦP2 (ΦP1 (S)) = ΦP20 (ΦP10 (S)) whenever both sides are
defined or both sides are undefined.
By Exercise 2.2, P can be uniquely written as Q1 ; Q2 ; . . . ; Qk where each
Qκ is either a simple statement or a while loop. This means that there are
indices i and j such that P1 = Q1 ; . . . ; Qi and P10 = Q1 ; . . . ; Qj . W.l.o.g.
assume that i > j. This means that
ΦP2 (ΦP1 (S)) = ΦQi+1 ;...;Qk (ΦQ1 ;...;Qi (S))
= ΦQi+1 ;...;Qk (ΦQj+1 ;...;Qi ((ΦQ1 ;...;Qj (S))))
= ΦQj+1 ;...;Qk (ΦQ1 ;...;Qj (S))
= ΦP20 (ΦP10 (S))
2
Here f (r) denotes the r-fold composition of functions, that is, f (0) (x) = x and
f (x) = f (f (r−1) (x)) for all x.
(r)
3
What does well-defined mean? Since we did an inductive definition, we have to ensure
that (1) each ΦP (S) is defined or we explicitly say that the function value is undefined
(since undefined means that P on S runs forever) and (2) we do not assign ΦP (S) different
values at different places.

c Markus Bläser 2007–2015
2.2. Semantics 15

whenever all arising terms are defined. If one of them is not defined, then
both ΦP2 (ΦP1 (S)) and ΦP20 (ΦP10 (S)) are not defined, which is exactly what
we want. In the equation above, we used the induction hypothesis two times.

Exercise 2.2 Every WHILE program P can be written as P = P1 ; P2 ; . . . ; Pk

such that each Pκ , 1 ≤ κ ≤ k, is either a simple statement or a while loop.4
This decomposition of P is unique.

Remark 2.4 We could have circumvented the problem above by requiring in

rule 2.(b) that if P = P1 ; P2 , then P1 is either a simple statement or a while
loop. This avoids the ambiguity above. Or we could have used brackets, i.e.,
rule 2.(b) would have been replaced by P = [P1 ; P2 ].

The semantics of the simple statements is as we expect it: xi = xj + xk

takes the values of xj and xk , adds them and stores the result in xi . In case of
subtraction, if the result is negative, we will set it to 0 instead, since we can
only store values from N. This operation is also called modified difference.
The syntax of a while loop is again as expected: We execute P1 as long
as the value of xi does not equal 0. If the loop does not terminate, the result
is undefined.
If P is the concatenation of P1 and P2 , we first execute P1 on S and then
P2 on the state produced by P1 provided that P1 halted.
The semantics of FOR programs is defined in the same manner. The
semantics of the simple statements stays the same, we only have to modify
the interpretation of for loops.

2. (a) Assume P equals for xi do P1 od for some FOR program P1 .

Then
(σ )
ΦP (S) = ΦP1i (S).

(b) If P is P1 ; P2 for FOR programs P1 and P2 , then

ΦP (S) = ΦP2 (ΦP1 (S)).

A for loop executes P1 σi times, where σi is the value of xi before the

execution of the for loop. This means that changing the value of xi during
the execution of the for loop does not have any effect.5 In particular, for
loops always terminate.
This means that we can simplify the interpretation of the concatenation
of FOR programs since we do not have to deal with undefined values of Φ.
4
We call a program a while loop if it is of the form while xi 6= 0 do Q od for some
WHILE program Q.
5
For loops in C++, for instance, work differently; they are merely while loops.

c Markus Bläser 2007–2015
16 2. WHILE and FOR programs

Exercise 2.3 Show that every FOR loop can be simulated by a WHILE loop.
(Simulation here means that for every FOR program P of the form for xi
do P1 od we can find a WHILE program Q such that ΦP = ΦQ , i.e., both
programs compute the same function.)

2.3 Computable functions and sets

Definition 2.5 Let P be a WHILE program. The function ϕP : Ns → N
computed by P is defined by

first entry of ΦP ((α1 , . . . , αs , 0, . . . , 0))

ϕP (α1 , . . . , αs ) = if ΦP ((α1 , . . . , αs , 0, . . . , 0)) is defined

undefined otherwise


for all (α1 , . . . , αs ) ∈ Ns .

Definition 2.6 1. A partial function f : Ns → N is WHILE computable
if there is a WHILE program P such that f = ϕP .
2. f is FOR computable if there is a FOR program P such that f = ϕP .
3. The class of all WHILE computable functions is denoted by R.
4. The class of all FOR computable functions is denoted by PR.
The acronyms R and PR stand for recursive and primitive recursive.
These are the first names that were used for these two classes of functions
and we also use them throughout this lecture.
By Exercise 2.3, PR ⊆ R. Since for loops always terminate, all functions
in PR are total. On the other hand, there are functions that are not total but
WHILE computable. For instance, we already saw a program that computes
the function N → N that is defined nowhere. There are also total functions
in R \ PR. We will show their existence later.
Most of the time, we will talk about subsets L ⊆ N. We also call such
sets languages. For any such L, we define its characteristic function to be
the following function:
χL : N → {0, 1}
(
1 if x ∈ L,
x 7→
0 otherwise.
Since {0, 1} is a subset of N, we can view χL as a function N → N.
Definition 2.7 1. A language L ⊆ N is called recursive or decidable if
χL ∈ R.
2. The set of all recursive languages is denoted by REC.

c Markus Bläser 2007–2015
3 Syntactic sugar

The languages WHILE and FOR consist of five constructs. We now want to
convince ourselves that we are still able to compute every function Ns → N
that JAVA or C++ could compute. (And no, we are not able to display
fancy graphics.)

3.1 Variable names

WHILE and FOR offers only boring variable names like x17 . But of course
we can replace them by more fancy ones. We can always revert to ordinary
WHILE/FOR programs by replacing the new fancy name by some xi that
has not been used so far.

3.2 Assignments
WHILE does not contain assignments of the form xi := xj . But this can be
easily simulated by
1: xk := 0;
2: xi := xj + xk
where xk is a new variable.

3.3 Procedures and functions

Let h be a WHILE or FOR computable function Nt → N, respectively, and
let P be a WHILE or FOR program for h. We can enrich WHILE or FOR
by statements of the form xi = h(xj1 , . . . , xjt ) and can always revert back to
ordinary WHILE or FOR as shown in Program 2.
Above, ` is the largest index of a variable used in the current program
and m is the largest index of a variable in P . P̂ is the program obtained
from P by replacing every variable xi by xi+`+1 . Basically, we replace every
occurence of h by a program that computes h and avoid any interferences
by choosing new variables. This does not give a very short program but we
do not care.

Exercise 3.1 The evaluation strategy above is call by value. How do you
implement call by reference?

17
18 3. Syntactic sugar

Program 2 Simulation of a function call

1: x`+1 = xj1 ;
.
2: ..
3: x`+t = xjt ;
4: x`+t+1 = 0;
.
5: ..
6: x`+m+1 = 0;
7: P̂ ;
8: xi = x`+1

3.4 Arithmetic operations

WHILE and FOR only offers addition and (modified) subtraction as arith-
metic operations. But this is enough to do everything else. As an example,
let us implement the multiplication xi := xj · xk :
1: xi := 0;
2: for xj do
3: xi := xi + xk
4: od
It is easy to see that the program above is correct. Since a FOR loop can
be simulated by a WHILE loop, we can perform multiplications in WHILE,
too.
Exercise 3.2 Show that integer division does not increase the power of FOR
programs, i.e., we can simulate computing the quotient xi := xj /xk and the
remainder xi := xj rem xk in ordinary FOR.
Exercise 3.3 Show that we even could replace the simple statements xi :=
xj + xk , xi := xj − xk , and xi := c by the three statements xi ++, xi −−,
and xi := 0. xi ++ increases the content of xi by 1. xi −− decreases it by 1
except when the content of xi equals 0, then it is left unchanged.

3.5 Arrays
Built-in types like int in C++ or JAVA can only store a limited amount of
information. But we can get as many variables of this type as we want by us-
ing dynamic arrays (or memalloc, if you prefer that). WHILE programs only
have a finite number of variables but each of them can store an arbitrarily
large amount of information. In this way, we can simulate arrays.
Lemma 3.1 There are FOR computable functions h., .i : N2 → N and πi :
N → N, i = 1, 2, such that
πi (hx1 , x2 i) = xi

c Markus Bläser 2007–2015
3.5. Arrays 19

for all x1 , x2 ∈ N and i = 1, 2.

Proof. We define hx1 , x2 i = 21 (x1 + x2 )(x1 + x2 + 1) + x2 . Note that either

x1 + x2 or x1 + x2 + 1 is even, therefore, hx1 , x2 i is a natural number. h., .i
is obviously FOR computable by the results in Section 3.4 and Exercise 3.3.
Let p = hx1 , x2 i be given. We want to reconstruct x1 and x2 from p. We
have
1 1
(z + 1)(z + 2) − z(z + 1) = z + 1.
2 2
Therefore, the largest z such that 12 z(z +1) ≤ p is x1 +x2 , since x2 ≤ x1 +x2 .
From this sum z = x1 + x2 , we can reconstruct x2 as p − 12 z(z + 1). Finally,
from the z and x2 we can reconstruct x1 as z − x2 . In Exercise 3.4, we
construct FOR programs for π1 and π2 .

Exercise 3.4 Construct FOR programs for the projections πi , i = 1, 2.

Remark 3.2 h., .i is also surjective.

While you can forget about the rest of the construction of arrays once
you believe that we can simulate them, this pairing functions is essential for
later chapters and you should not forget about its properties.

Excursus: Gauß’s formula for consecutive integer sums

Karl Friedrich Gauß (1777–1855) was a German mathematician, astronomer, and
geodesist. If he had lived today, he would also have been a computer scientist for
sure.
One hears the following anecdote quite often:1 At the age of seven, Gauß
attended the Volksschule. To have some spare time, the teacher Büttner told the
pupils to sum up the numbers from 1 to 100. Gauß quickly got the answer 5050 by
recognizing
P100 that 1 + 100 = 101, 2 + 99 = 101, . . . 50 + 51 = 101, in other words,
1
i=1 i = 2 100 · 101. Poor Büttner.
·
Gauß was also one of the main reasons against introducing the Euro.

Now if we want to store a0 , . . . , an−1 ∈ N in one number, we store them

as han−1 , han−2 , . . . , ha2 , ha1 , a0 ii . . .ii. Assume that A is the variable where
we want to store our array and n contains the number of elements. (We
are already using the convention from Section 3.1.) Initially, all entries are
zero. It is easy to check that h0, . . . , h0, h0, 0ii . . .i = 0. Thus initially, we
set A := 0. We want to extract the ith element from A and can do this as
shown in Program 3. This program is not a WHILE/FOR program not even
if you apply all the syntactic sugar mentioned so far. But I am sure that you
1
In this lecture notes, you occasionally find anecdotes about mathematicians and com-
puter scientists. I cannot guarantee that they are true but they are entertaining. This
particular one can be found on Wikipedia, which indicates that it is false.

c Markus Bläser 2007–2015
20 3. Syntactic sugar

Program 3 Extracting the ith element from A

Output: A[i] is returned in x0
1: x0 := A;
2: for n − i − 1 do
3: x0 := π2 (x0 )
4: od;
5: if i 6= 0 then
6: x0 := π1 (x0 )
7: fi

can convert it into a true WHILE/FOR program. The idea of the algorithm
above is simple. If we want to extract ai for i > 0, then we will get it via

π1 ◦ π2 ◦ · · · ◦ π2 (A).
| {z }
n−i−1

If i = 0, then we will get ai via

π2 ◦ · · · ◦ π2 (A).
| {z }
n−1

Exactly this is implemented in the program above.

Now assume that we want to store the value of b in the ith place of A.
The idea is basically the same as before, but now we also have to restore the
array. Program 4 shows how to achieve this.

Program 4 Simulating A[i] := b

1: x0 := A;
2: x1 := 0;
3: for n − i − 1 do
4: x1 := hπ1 (x0 ), x1 i;
5: x0 := π2 (x0 )
6: od;
7: if i 6= 0 then
8: x0 := hb, π2 (x0 )i
9: else
10: x0 := b
11: fi
12: for n − i − 1 do
13: x0 := hπ1 (x1 ), x0 i;
14: x1 := π2 (x1 )
15: od
16: A := x0 ;

c Markus Bläser 2007–2015
3.6. Further exercises 21

In the first for loop, we find the position of the element that we want
to change as we did before. In addition, we also store the elements that we re-
move from x0 in x1 . Assume that, for instance, x0 equals hai , hai−1 , . . . , ha2 , ha1 , a0 ii . . .ii
and x1 equals hai+1 , hai+2 , . . . , han−2 , han−1 , 0ii . . .ii. We replace ai by b. If
i = 0, then ai is the remaining element in x0 and we can just overwrite it
with b. If i > 0, then we throw ai away by computing π2 (x0 ) and replace it
by b via the pairing function. Note the 0 appended to the end of x1 . This
makes a case distinction redundant when reconstructing A in the second for
loop; we can always extract the next element from x1 by applying π1 . (One
could also make this convention for A.) In this second loop, we insert the
elements that were removed from x0 back into x0 . We extract them one by
one from x1 and add them to x0 by using the pairing function.

Exercise 3.5 A stack is a data structure that stores some objects, here our
objects will be natural numbers. We can either push a number onto the stack.
This operation stores the number in the stack. Or we can pop an element
from the stack. This removes the element from the stack that was the last to
be pushed onto the stack among all elements still in the stack. If the stack is
empty and we want to pop an element from the stack, this will result in an
error. So it works like a stack of plates where you can only either remove the
top plate or put another plate on the top.2 There is usually also a function
isempty that allows you to check whether a stack is empty or not.
1. How do you store a stack of natural numbers in one natural number?

2. Use the pairing function to implement the push operation in WHILE.

3. Use the projections to implement the pop operation in WHILE.

4. Implement the isempty function.

3.6 Further exercises

Exercise 3.6 Show how to simulate the if-then-else statement in FOR.

Exercise 3.7 The following two statements are also useful. Explain how to
simulate them in simple WHILE.
1. Input: v1 , . . . , vs declares v1 , . . . , vs as the input variables.

2. return x leaves the current program immediately and the value of x is

the output of the program.

Here is another function that we could use as a pairing functions. It is

injective but not surjective.
2
Dexterous people might do different things but computer scientist usually do not.

c Markus Bläser 2007–2015
22 3. Syntactic sugar

Exercise 3.8 Let k be some constant. Let p1 , . . . , pk be different prime num-

bers.

1. Show that the mapping g given by

Nk → N
(x1 , . . . , xk ) 7→ px1 1 · px2 2 · · · pxk k

is an injective mapping

2. Show that g is FOR computable.

3. Show that there is a FOR program that decides whether a given y ∈ N

is in im g.

4. Show that there is a FOR program that given y ∈ im g computes the

unique x1 , . . . , xk such that g(x1 , . . . , xk ) = y.

c Markus Bläser 2007–2015
A Ackermann function

Chapters that are numbered with latin characters instead of numbers are for
your personal entertainment only. They are not an official part of the lecture,
in particular, not relevant for any exams. But reading them does not hurt
either . . .

It is clear that there are functions that are WHILE computable but not
FOR computable, since FOR programs can only compute total functions but
WHILE programs can compute partial ones. Are there total functions that
are WHILE computable but not FOR computable? I.e. are WHILE loops
more powerful than FOR loops? The answer is affirmative.

A.1 Definition
The Ackermann function is a WHILE computable but not FOR computable
total function, which was first published in 1928 by Wilhelm Ackermann, a
student of David Hilbert. The so called Ackermann-Péter-Function, which
was defined later (1955) by Rózsa Péter and Raphael Robinson has only two
variables (instead of three).
The Ackermann function is the simplest example of a well defined total
function that is WHILE computable but not FOR computable, providing
a counterexample to the belief in the early 1900s that every WHILE com-
putable function was also FOR computable. (At that time, the two concepts
were called recursive and primitive recursive.) It grows faster than an expo-
nential function, or even a multiple exponential function. In fact, it grows
faster than most people (including me) can even imagine.
The Ackermann function is defined recursively for non-negative integers
x, y by
a(0, y) = y + 1
a(x, 0) = a(x − 1, 1) for x > 0
a(x, y) = a(x − 1, a(x, y − 1)) for x, y > 0

Lemma A.1 a is a total function, i.e, a(x, y) is defined for all x, y ∈ N.

Proof. We prove by induction on x that a(x, y) is defined for all x, y ∈ N.

Starting with the induction base x = 0 gives us a(0, y) = y + 1 for all y.
The induction step x − 1 → x again is a proof by induction, now on y. We
start with y = 0. By definition, a(x, 0) = a(x − 1, 1). The right-hand side
is defined by the induction hypothesis for x − 1. For the induction step

23
24 A. Ackermann function

y − 1 → y note that a(x, y) = a(x − 1, a(x, y − 1)) by definition. The right-

hand side is defined by the induction hypotheses for x − 1 and also y − 1.

Lemma A.2 a is WHILE computable.

Proof. We prove this by constructing a WHILE program that given x

and y, computes a(x, y). We use a stack S for the computation. Program 5
will compute the value of a with the top two elements of the stack as the
arguments. It first pushes the two inputs x and y on the stack. Then it
uses the recursive rules of a to compute the value of a. This might push
new elements on the stack, but since a is total, the stack will eventually be
consisting of a sole element, the requested value.

A.2 Some closed formulas

In this section, we keep x fixed and consider a as a function in y. For small
x, we can express a(x, .) in closed form. For x = 1, we get

a(1, y) = a(0, a(1, y − 1))

= a(2, y − 1) + 1
= a(0, a(1, y − 2)) + 1
= a(1, y − 2) + 2
..
.
= a(1, 0) + y
= a(0, 1) + y
= y + 2.

For x = 2, we have

a(2, y) = a(1, a(2, y − 1))

= a(2, y − 1) + 2
..
.
= a(2, 0) + 2y
= a(1, 1) + 2y
= 2y + 3.

c Markus Bläser 2007–2015
A.3. Some useful facts 25

Finally, for x = 3, we obtain

a(3, y) = a(2, a(3, y − 1))

= 2 · a(3, y − 1) + 3
= 2 · a(2, a(3, y − 2)) + 3
= 2 · (2 · a(3, y − 2) + 3) + 3

= 22 · a(3, y − 2) + 32
= 22 · (2 · a(3, y − 3) + 3) + 3 · (1 + 2)
= 23 · a(3, y − 3) + 3 · (1 + 2 + 22 )
..
.
y−1
X
y
= 2 · a(3, 0) + 3 · 2i
k=0
y y
= 2 · a(3, 0) + 3 · (2 − 1)
= 2y · a(2, 1) + 3 · 2y − 3
= 2y · 5 + 3 · 24 − 3
= 2y+3 − 3.

2
..
.
Exercise A.1 Show that a(4, y) = |22{z } −3.
y+3

A.3 Some useful facts

In this section, we show some monotonicity properties of a.

Lemma A.3 The function value is strictly greater then its second argument,
i.e., ∀x, y ∈ N
y < a(x, y).

Proof. Again, we show this by induction on x. For x = 0, we just use

the definition of a which yields

a(0, y) = y + 1 > y.

The induction step x − 1 → x is again shown by an inner induction on

y. So we start with y = 0:

a(x, 0) = a(x − 1, 1) > 1 > 0,

c Markus Bläser 2007–2015
26 A. Ackermann function

where the first inequality follows from the induction hypothesis for x. The
induction step for the inner induction is shown as follows:

a(x, y) = a(x − 1, a(x, y − 1)) > a(x, y − 1) > y − 1. (A.1)

a(x, y) = a(x − 1, a(x, y − 1)) is strictly greater than a(x, y − 1) because of

the induction hypothesis for x. But this is even strictly greater than y − 1
because of the induction hypothesis for y.
We need to prove that a(x, y) > y. But in (A.1), there stands a “>”
twice. Thus (A.1) even implies a(x, y) > y.

Lemma A.4 The Ackermann function is strictly monotonic increasing in

the second argument, i.e., ∀x, y ∈ N

a(x, y) < a(x, y + 1).

Proof. We consider two cases. For x = 0 we have a(0, y) = y + 1 which

is less than y + 2 = a(0, y + 1). So we get a(0, y) < a(0, y + 1).
For x > 0 we see from Lemma A.3 that a(x, y) < a(x − 1, a(x, y)) which
equals a(x, y + 1) by the definition of the Ackermann function.

Lemma A.5 For all x, y ∈ N, a(x, y + 1) ≤ a(x + 1, y).

Proof. By induction on y. For y = 0 the equation a(x, 1) = a(x + 1, 0)

follows from the definition of a.
Consider the induction step y − 1 → y. By Lemma A.3, we know that
y < a(x, y). Thus y + 1 ≤ a(x, y) ≤ a(x + 1, y − 1) because of the induction
hypothesis. Lemma A.4 allows us to plug the inequality y+1 ≤ a(x+1, y−1)
into the second argument yielding

a(x, y + 1) ≤ a(x, a(x + 1, y − 1)) = a(x + 1, y).

Lemma A.6 The Ackermann function is strictly monotonic increasing in

the first argument as well, i.e., ∀x, y ∈ N,

a(x, y) < a(x + 1, y).

Proof. Using Lemma A.4 first and then Lemma A.5, we obtain

a(x, y) < a(x, y + 1) ≤ a(x + 1, y),

c Markus Bläser 2007–2015
A.4. The Ackermann function is not FOR computable 27

A.4 The Ackermann function is not FOR computable

Let P be a FOR program that uses the variables x0 , . . . , x` . Assume that
these variables are initialized with the values v0 . . . v` and that after the
execution of P , the variables contain the values v00 . . . v`0 . We now define a
function fP (n) that essentially measures the size of the output that a FOR
program can produce (in terms of the size of the input). fP is defined via
( ` 
`
)
X  X
fP (n) = max vi0  vi ≤ n .


i=0 i=0

In other words, fP (n) bounds the sum of the values of x0 , . . . , x` after the
execution of P in terms of the sum of the values of x0 , . . . , x` before the
execution of P .

Lemma A.7 For every FOR program P there is a k such that

fP (n) < a(k, n).

Proof. By structural induction. We can assume w.l.o.g. that the simple

instructions of the FOR program are either xi ++ or xi −− or xi := 0. (This
was an exercise.)
Furthermore, we assume that for every FOR loop for xi do Q od in
P , xi does not appear in Q. If it does, we can replace the loop first by
xk := xi ; for xk do P1 od where xk is an unused variable. And of course,
since we have only a restricted set of simple statements, we have to replace
the assignment xk := xi by something else, too. For the moment, let P̂ be
the resulting program. Let xi be a variable of P . After the execution of
P̂ , the value of xi is the same as the value of xi after the execution of P .
Therefore, fP (n) ≤ fP̂ (n) and it suffices to bound fP̂ (n).
Let P = xi ++ or P = xi −− be a FOR program. If the sum of the values
is bounded by n before the execution of P , then it is bounded by n + 1 after
the execution. Thus
fP (n) ≤ n + 1 < a(1, n).

If P = xi := 0;, then the size of the ouput cannot be larger than the size of
the output. Hence
fP (n) ≤ n < a(0, n),

too. This finishs the induction base.

For the induction step, let P = P1 ; P2 be a FOR program. From the
induction hypothesis we know that fPi (n) < a(ki , n) for constants ki and
i ∈ {1, 2}. After the execution of P1 , the sum of the values of the variables
is bounded by a(k1 , n) which is also a bound on the sum of the values of the

c Markus Bläser 2007–2015
28 A. Ackermann function

variables before the execution of P2 . Altogether,

fP (n) = fP2 (fP1 (n))

< a(k2 , a(k1 , n))
< a(k3 , a(k3 , n)) by monotonicity with k3 = max{k1 , k2 }
< a(k3 , a(k3 + 1, n))
= a(k3 + 1, n + 1) per definition
< a(k3 + 2, n).

Let P = for xi do P1 od be a FOR program. Recall that xi does not occur

in P1 . From the induction hypothesis, we get fP1 (n) < a(k1 , n). Fix a tuple
ν0 , . . . , ν` for which the maximum is attained in the definition of xi . Let
m := νi be the value of xi in this tuple. We distinguish three cases:

m = 0: In this case, fP (n) = n < a(0, n).

m = 1: Here, fP (n) ≤ fP1 (n) < a(k1 , n).

m > 1: Here we have,

fP (n) = fP1 ◦ . . . ◦ fP1 (n − m) + m

| {z }
m times
< a(k1 , fP1 ◦ . . . ◦ fP1 (n − m)) + m
| {z }
m−1 times
≤ a(k1 , fP1 ◦ . . . ◦ fP1 (n − m)) + m − 1
| {z }
m−1 times
..
.
≤ a(k1 , a(k1 , . . . a(k1 , n − m)))
< a(k1 , a(k1 , . . ., a(k1 + 1, n − m)))
| {z }
m times
< a(k1 + 1, n).

Theorem A.8 The Ackerman function a is not FOR computable.

Proof. Assume that a was FOR computable, then â(k) = a(k, k) is FOR
computable as well. Let P be a FOR program for â. Lemma A.7 tells us
that there is a k such that

â(k) ≤ fP (k) < a(k, k) = â(k),

a contradiction. This proves that a is not FOR computable.

c Markus Bläser 2007–2015
A.4. The Ackermann function is not FOR computable 29

Program 5 Ackermann function

Input: x, y
1: push(S, x);
2: push(S, y);
3: while size(S) > 1 do
4: y := pop(S);
5: x := pop(S);
6: if x = 0 then
7: push(S, y + 1);
8: else
9: if y = 0 then
10: push(S, x − 1);
11: push(S, 1);
12: else
13: push(S, x − 1);
14: push(S, x);
15: push(S, y − 1);
16: fi
17: fi
18: od
19: x0 := pop(S);

c Markus Bläser 2007–2015
4 Gödel numberings

We will deal with two fundamental questions that we will use in the following
chapters. The first one is: How many WHILE programs are there? And the
second one is: How can we feed a WHILE program into another WHILE
program as an input?
There are certainly infinitely many WHILE programs: x0 := 0 is one,
x0 := 0; x0 := 0 is another, x0 := 0; x0 := 0; x0 := 0 is a third one; I guess
you are getting the idea. But there are different “sorts of infinite”.

Definition 4.1 A set S is countable if there exists an injective1 function

S → N. If there is a bijection S → N, then S is called countably infinite.

Exercise 4.1 If S is countable, then S is finite or countably infinite.

(Remark: Let f : S → N an injective function. If S is not finite, then its
image under f is infinite. Let I = f (S). We can now construct a bijection
g by letting g(0) = min{n | n ∈ I}, g(1) = min{n | n ∈ I \ {g(0)} }, and
so on. Since I is infinite, g(i) is defined for every i ∈ N. This construction
bases on the fact that the natural numbers are well-ordered: every nonempty
subset has a minimal element. For larger sets, this is not so clear any more.
But also beyond the scope of this lecture . . . )

Recall that the pairing function h., .i is a bijection from N×N → N. Thus
N × N is countable, too.

Exercise 4.2 1. Prove that Q is countable.

2. Prove that if A and B are countable, so is A × B.

We will show that the set of all WHILE programs is countably infinite.
By assigning each WHILE program a natural number in a unique way, we
can feed them into other WHILE programs, too. For just proving that the
set of all WHILE programs is countable, we can use any injective function.
But for the purpose of feeding WHILE programs into WHILE programs, this
function should also have some nice properties.
We will construct an injective function göd : W → N, that is, different
WHILE programs get different numbers. But this is not enough, we also
need the following two properties:
1
Recall that injective includes that the function is total

30
 31

1. There is a WHILE program C that computes the characteristic func-

tion of im göd, and

2. There is a WHILE program U that given i ∈ im göd and x ∈ N com-

putes ϕP (x) where P = göd−1 (i).

Furthermore, given i ∈ im göd it is “easy” to find the WHILE program P

with göd(P ) = i. “Easy” is not formally specified here. We could extend the
WHILE language and add a print command that can print fixed string and
the content of variables. Then it is quite easy (at least if you implemented
a stack in WHILE) to print P if i is given.
Such a mapping göd is usually called Gödel numbering, named after the
mathematician Kurt Gödel. The term “Gödel numbering” is used for many
different things, so be careful. (Often, the existence of U is not included in
the definition.)

Excursus: Kurt Gödel

Kurt Gödel (born 1906 in Brno, Austria–Hungary (now Czech Republic) and died
1978 in Princeton, USA) was a mathematician and logician.
He is best known for his incompleteness theorem which roughly says that in any
self-consistent recursive axiomatic system powerful enough to describe the arith-
metics of N, there are theorems that are true but cannot be deduced from the
axioms. (More one this later.) This destroyed David Hilbert’s dream that every-
thing in mathematics can be proven from a correctly-chosen finite system of axioms
In Princeton, Kurt Gödel became a close friend of Albert Einstein. Gödel, to
put it nicely, was a rather complicated person.

We define a concrete Gödel numbering göd for WHILE programs that we

will use throughout this lecture. göd will be defined inductively. We start
with the simple statements:

1. The statement xi := xj + xk is encoded by h0, hi, hj, kiii.

2. The statement xi := xj − xk is encoded by h1, hi, hj, kiii.

3. The statement xi := c is encoded by h2, hi, cii.

The while loop and the concatenation of programs is encoded as follows:

1. If P = while xi 6= 0 do P1 od, then göd(P ) = h3, hi, göd(P1 )ii.

2. If P = P1 ; P2 where P1 is a while loop or a simple statement, then

göd(P ) = h4, hgöd(P1 ), göd(P2 )ii.

Lemma 4.2 göd is well-defined.

c Markus Bläser 2007–2015
32 4. Gödel numberings

Proof. The proof is again by structural induction.

Induction base: Obviously, göd(P ) is well-defined for all P ∈ W0 .
Induction step: Now assume that P ∈ Wn for some n. Then either P =
while xi 6= 0 do P1 od or P = P1 ; P2 for some P1 , P2 ∈ Wn−1 . In the latter
case, P1 is a while loop or a simple statement. By the induction hypothesis,
göd(P1 ) and göd(P2 ) are both well-defined. Thus göd(P ) = h3, hi, göd(P1 )ii
or göd(P ) = h4, hgöd(P1 ), göd(P2 )ii, respectively, are both well-defined, too.
For the latter case, note that the decomposition into P1 ; P2 is unique by
Exercise 2.2.

Lemma 4.3 göd is injective.

Proof. We show the statement that for all n, göd(P ) = göd(Q) implies
P = Q for all P, Q ∈ Wn . From this, the assertion of the lemma follows.
Induction base: The statement is clear for all programs in W0 . This shows
the induction base.
Induction step: Now assume that göd(P ) = göd(Q) and assume that P ∈
Wn \ Wn−1 for some n > 0 and Q ∈ Wn . Since n > 0, göd(P ) is either
h3, hi, göd(P1 )ii or h4, hgöd(P1 ), göd(P2 )ii for some programs P1 , P2 ∈ Wn−1 .
We only treat the case göd(P ) = h3, hi, göd(P1 )ii, the other case is an exer-
cise. göd(P ) = göd(Q) in particular implies that π1 (göd(P )) = π1 (göd(Q)).
This shows that göd(Q) = h3, hj, göd(Q1 )ii. But also π2 (göd(P )) = π2 (göd(Q)),
i.e, hi, göd(P1 )i = hj, göd(Q1 )i. But since h., .i is a bijection, this means that
i = j and göd(P1 ) = göd(Q1 ). By the induction hypothesis, P1 = Q1 . Thus
P = Q.

Corollary 4.4 The set of all WHILE programs W is countable.

For the rest of this lecture, we denote the image of göd by G.

Exercise 4.3 Fill in the missing part in the proof of Lemma 4.3.

Excursus: Programming systems

The mapping göd associates with every i ∈ G a function ϕgöd−1 (i) . Instead of
ϕgöd−1 (i) we write ϕi . If we associate with every i ∈
/ G a fixed function ϕi , say
the function that is undefined everywhere, we get an infinite sequence of functions
(ϕi )i∈N .

Definition 4.5 1. A sequence (ψi )i∈N is called a programming system if the

set of all ψi is precisely R, the set of all WHILE computable functions.
2. It is universal, if the programming system has a universal program, i.e, there
is an index u such that ψu (hj, xi) = ψj (x) for all j, x ∈ N.

c Markus Bläser 2007–2015
 33

3. A universal programming system is called acceptable if there is an index c

such that ψψc (hj,ki) = ψj ◦ ψk .

The sequence (ϕi )i∈N of WHILE computable functions is an acceptable pro-

gramming system. It is certainly a programming system and we will see soon that
it is universal by constructing a universal program U . To see that it is acceptable,
note that h4, hk, jii is the Gödel number of ϕj ◦ ϕk , if j, k ∈ G. If j or k is not in
G, then ϕc just outputs the index of the function that is undefined everywhere. We
can check whether j, k ∈ G once we constructed the program C.
There are other programming systems and we will see some of them. Everything
that we prove in the remainder of this part is valid for every acceptable programming
system. For instance, the set of all functions computed by JAVA programs is an
acceptable programming systems. (We have to identify ASCII texts with natural
numbers somehow.)

c Markus Bläser 2007–2015
5 Diagonalization

In this chapter we will answer the question whether there is a function N → N

that is not computable by a WHILE program.

5.1 Proof by “counting”

Basically, we will show that the set of all total functions N → N is not
countable. Even the set of all functions N → {0, 1} is not countable.

Theorem 5.1 The set of all total functions N → {0, 1} is not countable.

Proof overview: The proof will use a technique that is called Cantor’s
diagonal argument. We assume that the set of all total functions N → {0, 1},
call it F , is countable. Then there is a bijection n between F and N, i.e.,
each function in f ∈ F gets a “number” n(f ). We construct a total function
c : N → {0, 1} that differs from every f ∈ F on the input n(f ). c ∈ F by
construction, on the other hand it differs from every f ∈ F on some input.
This is a contradiction.

Proof. Assume that F is countable and let n : F → N be a bijection.

Let fi be the function in F that is mapped onto i by n, i.e., n(fi ) = i for all
i ∈ N.
We arrange the values of the functions fi in a 2-dimensional tabular. The
ith row contains the values of fi and the jth column contains the values of
all functions on input j. This means that the entry in position (i, j) contains
the value fi (j) (see Figure 5.1).

0 1 2 3 ...
0 f0 (0) f0 (1) f0 (2) f0 (3) ...
1 f1 (0) f1 (1) f1 (2) f1 (3) ...
2 f2 (0) f2 (1) f2 (2) f2 (3) ...
3 f3 (0) f3 (1) f3 (2) f3 (3) ...
.. .. .. .. .. ..
. . . . . .

Figure 5.1: The diagonalization scheme

34
5.1. Proof by “counting” 35

Now we define the function c as follows: c(i) = 1 − fi (i) for all i ∈ N. In

other words,
(
0 if fi (i) = 1,
c(i) =
1 if fi (i) = 0.

c differs from the fi ’s in the entries that are underlined in the table in
Figure 5.1. Clearly c ∈ F . But this means that there is an index i0 such
that fi0 = c, since {fi | i ∈ N} = F . In particular,

fi0 (i0 ) = c(i0 ) = 1 − fi0 (i0 ).

But this is a contradiction since fi0 (i0 ) is a natural number and the equation
x = 1 − x has no integral solution.

Corollary 5.2 There is a total function N → {0, 1} that is not WHILE

computable.

Proof. The proof is by contradiction: If every function from F , the set

of all total functions N → {0, 1} was WHILE computable, then the image of
the mapping given by P 7→ ϕP would contain F . But this means that there
is an injective mapping i1 from F to some subset of W. Since the set W
of all WHILE programs is countable, there is an injective mapping i2 from
W → N. The composition i2 ◦ i1 is an injective mapping from F to N. This
means that F is countable, a contradiction.

Since the characteristic function of a subset of N is a function N → {0, 1},

we get the following corollary.

Corollary 5.3 There is a subset of N that is not recursive.

Excursus: Cantor’s diagonal argument

Georg F. L. P. Cantor (born 1845 in St. Petersburg, died 1918 in Halle), was a
German mathematician. He is known as one of the creator of set theory.
We saw two accomplishments of Cantor in this lecture: The proof that N2
is again countable and the diagonal argument that show that the set of all total
functions N → {0, 1} is not countable.
The following conversation between Dedekind and Cantor is reported:
Dedekind: “Eine Menge stelle ich mir vor wie einen Sack mit Dingen darin.”
Cantor (in theatric pose): “Eine Menge stelle ich mir vor wie einen Abgrund.”
Cantor should be right.

c Markus Bläser 2007–2015
36 5. Diagonalization

5.2 Explicit construction

The proof of the existence of a characteristic function that is not WHILE
computable was indirect and used the fact that there are more characteristic
functions than WHILE programs. Basically the same proof also yields a di-
rect construction of a characteristic function that is not WHILE computable.
This construction will be explicit, i.e., we can precisely say how the function
looks like. The function c that we will construct has the property that for
all i ∈ G, c(i) is defined iff ϕP (i) is not defined, where P = göd(i). In other
words, c(i) is defined whenever the WHILE program with Gödel number i
does not halt on i. This c is at least “semi-natural”. We will see soon how
to prove that natural tasks, like verification, are not computable.

Overview over alternative proof of Corollary 5.2: We will use the same
diagonalization scheme as in Figure 5.1. The construction becomes explicit,
since we do not use a hypothetical enumeration of all characteristic functions
but an enumeration of all WHILE programs that we already constructed.

Alternative proof of Corollary 5.2. We constructed an injective mapping

göd from the set of all WHILE programs to N in Chapter 4. We now define
a sequence of functions f0 , f1 , f2 , . . . by
(
ϕP (j) if i encodes the program P , i.e., göd(P ) = i
fi (j) =
0 if i does not encode any WHILE program
for all i and j. That means, if i encodes a WHILE program, i.e., there is a
program P such that göd(P ) = i, then fi will be the function ϕ computed
by P . If i does not encode a WHILE program, then fi will be a dummy
function; here we choose the function that is zero everywhere.
Now define the function c by
(
1 if fn (n) = 0 or is undefined
c(n) =
0 otherwise
for all n ∈ N. There is no WHILE program P that can compute c, because
if göd(P ) = i, then c differs from ϕP at the input i.

Remark 5.4 c essentially is the characteristic function of the set of all

Gödel numbers i such that göd−1 (i) either does not halt on i or returns
0.

Excursus: Programming systems II

Corollary 5.2 holds for all programming systems. All that we used is that there is
a mapping i → ψi such that the image of this mapping is R.

c Markus Bläser 2007–2015
5.3. Further exercises 37

5.3 Further exercises

Exercise 5.1 Show that for any nonempty set S, there is no bijection be-
tween S and its power set P(S).

c Markus Bläser 2007–2015
6 A universal WHILE program

In this chapter, we will construct the WHILE programs C and U for our
function göd. Assume that we are given an index g ∈ im göd, i.e., a valid
encoding g of a WHILE program P and an m ∈ N. U now has to simulate
P on input m with only a fixed number of variables. The program P has a
fixed number of variables, too, but since U has to be capable of simulating
every WHILE program, there is no a priori bound on the number of variables
in P . Thus U will use an array X to store the values of the variables of P .
Luckily, we do already know how to simulate arrays in WHILE (and even
FOR). Let ` be the largest index of a variable that occurs in P . Then an
array of length ` + 1 is sufficient to store all the values. It is not too hard to
extract this number ` given g. But since any upper bound on ` is fine too,
we just use an array of length g in U . g is an upper bound on ` because of
Exercise 6.1 (and the way we constructed göd).

Exercise 6.1 Show that for all j, k ∈ N, hj, ki ≥ max{j, k}.

A simple statement is encoded as h0, hi, hj, kiii (addition), h1, hi, hj, kiii
(subtraction), or h2, hi, cii (initialization with constant). Using π1 , we can
project onto the first component of these nested pairs and find out whether
the statement is an addition, subtraction, or initialization with a constant.
The result that we get by application of π2 then gives us the information
about the variables and/or constants involved. Program 6 shows how to
perform the addition. X stores the array that we need to simulate. When
we plug this routine into U , we might have to rename variables.

Program 6 Subroutine for addition

Input: hi, hj, kii stored in variable x2
1: x3 = π1 (x2 );
2: x4 = π1 (π2 (x2 ));
3: x5 = π2 (π2 (x2 ));
4: X[x3 ] := X[x4 ] + X[x5 ]

Exercise 6.2 Write the corresponding programs for subtraction and initial-
ization with a constant.

More problematic are while loops and concatenated programs. We will

use a stack S to keep track of the program flow. Luckily, we know from
Exercise 3.5 how to implement a stack in WHILE (and even FOR).

38
 39

Program 7 is our universal WHILE program. (Look at the code for a

while. Then try to imagine you had to write a universal C++ program.
Next, thank your professor.) There are five important variables in it whose
meaning we explain below:
X: the array that stores the values of the variables in the pro-
gram P := göd−1 (g).
S: the stack that stores (encodings of) pieces of P to be executed
later
cur : a variable that stores the piece of P that we are right now
simulating
term: is 0 if our simulation terminated and 1 otherwise.
type: stores the type (0 to 4) of the current statement that we
are simulating
In lines 1 to 5, we initalize X and store the input m in X[0]. Our current
program that we work on is g. We will exit the while loop if we finish the
simulation of P . cur will always be of the form
• h0, hi, hj, kiii (addition)
• h1, hi, hj, kiii (subtraction)
• h2, hi, cii (initialization)
• h3, hi, göd(P1 )ii (while loop)
• h4, hgöd(P1 ), göd(P2 )ii (concatenation)
In line 6, we set type := π1 (cur ). This value is between 0 to 4. If type ∈
{0, 1, 2}1 , then we just have to simulate an addition, subtraction, or initital-
ization. This is easy. The next two cases are far more interesting.
If type = 3, then cur = h3, hi, göd(P1 )ii. So we have to simulate a while
loop while xi 6= 0 do P1 od. In line 18, we check whether the condition
xi 6= 0 is fulfilled. If not, then we do not enter the while loop. If yes, then
we do the following: First we simulate P1 once and then we simulate the
while loop again. Therefore, we push cur (which equals h3, hi, göd(P1 )ii)
and π2 (π2 (cur )) (which equals göd(P1 )) on the stack.
If type = 4, then cur = h4, hgöd(P1 ), göd(P2 )ii. We push π2 (π2 (cur ))
(which is göd(P2 )) on the stack and then π1 (π2 (cur )) (which is göd(P1 )). In
this way, we will first execute P1 on top of the stack and then P2 .
At the end of the while loop, we check whether the stack is empty. If yes,
then our simulation has finished and we just have to copy X[0], the output
of the simulation, into x0 . If the stack is not empty, we pop the next piece
of program from the stack, store it into cur , and go on with the simulation.
1
That is not 100% correct; what we mean is that the value of type is in {0, 1, 2}. We
will use this sloppy notation often in the remainder of this chapter.

c Markus Bläser 2007–2015
40 6. A universal WHILE program

Program 7 Universal WHILE program U

Input: Gödel number g of P , input m of P
Output: ϕP (m) if P terminates on m. Otherwise U does not terminate.
1: X := 0; {Sets all entries of X to 0}
2: X[0] := m; {Stores input for simulation}
3: S := 0;
4: term := 1;
5: cur := g;
6: while term 6= 0 do
7: type := π1 (cur )
8: if type = 0 then
9: simulate addition (as in Program 6).
10: fi
11: if type = 1 then
12: simulate subtraction.
13: fi
14: if type = 2 then
15: simulate initialization with constant.
16: fi
17: if type = 3 then
18: i := π1 (π2 (cur ));
19: if X[i] 6= 0 then
20: push(cur , S); {push the loop once again}
21: push(π2 (π2 (cur )), S) {push the body of the loop}
22: fi
23: fi
24: if type = 4 then
25: push(π2 (π2 (cur )), S); {push two parts onto stack in reverse order}
26: push(π1 (π2 (cur )), S)
27: fi
28: if isempty(S) = 0 then
29: cur := pop(S)
30: else
31: term := 0
32: fi
33: od
34: x0 := X[0];

c Markus Bläser 2007–2015
6.1. Further exercises 41

The key in proving the correctness of the universal WHILE program is

to show the following lemma. The subsequent Theorem 6.2 is a special case
of it.
Lemma 6.1 Let T be the state that corresponds to the content of X, let σ be
the content of stack S, and let P = göd−1 (cur) before line 6 at the beginning
of the while loop. Let T 0 be the state that corresponds to the content of X,
when the content of S after line 26 at the end of the while loop is again σ
for the first time. Then T 0 = ΦP (T ), if this event occurs. If σ is never again
the content of S, i.e., U does not terminate, then ΦP (T ) is not defined, i.e.,
P does not terminate when started in state T .
Exercise 6.3 Prove Lemma 6.1. You can use—ta dah!—structural induc-
tion.
Theorem 6.2 There is a WHILE program U that given g ∈ im göd and m ∈
N, computes ϕgöd−1 (g) (m) if ϕgöd−1 (g) (m) is defined and does not terminate
otherwise.
Corollary 6.3 (Kleene normal form) Let f be a WHILE computable func-
tion. Then there are FOR program P1 , P2 , and P3 such that the program
P1 ; while x1 6= 0 do P2 od; P3
computes f
Proof. Let P be some WHILE program for f and let g = göd(P ). Our
universal program U is in Kleene normal form. (Recall that arrays, stacks,
and the projections all could be realized by FOR programs.) Instead of
giving U the Gödel number g as an input, we hardwire it into the program,
i.e, in line 5, g is now a constant and not a variable.

Exercise 6.4 Modify the program U in such a way that we get the program
C that checks whether a given g is in im göd. Can you achieve that C is a
FOR program?

6.1 Further exercises

Exercise 6.5 Is there a universal FOR program, i.e., a FOR program U0
that given a Gödel number i of a FOR program P and an input x computes
ϕP (x)?
Exercise 6.6 Show that the following function is FOR computable:
(
1 if program göd−1 (i) halts on x after ≤ t steps
hi, hx, tii 7→
0 otherwise
“After ≤ t steps” here means after the execution of ≤ t simple statements.
Note that by definition, the body of a while loop is never empty.

c Markus Bläser 2007–2015
7 The halting problem

The halting problem H is the following problem:

H = {hg, mi | g ∈ im göd and göd−1 (g) halts on m}.

This is a natural problem. The special halting problem H0 is the following

special case:

H0 = {g | g ∈ im göd and göd−1 (g) halts on g}.

Here we want to know whether a WHILE program halts on its own Gödel
number. While this is not as natural as the regular halting problem, it is a
little easier to prove that it is not decidable. In the next chapter, we formally
show that it is indeed a special case of the halting problem and develop a
general method to show that problems are not decidable.

7.1 The halting problem is not decidable

Theorem 7.1 H0 ∈
/ REC, i.e., the special halting problem is not decidable.

Proof. The proof is by contradiction. Assume that there is a WHILE

program P0 that decides H0 , i.e., ϕP0 = χH0 . In particular, P0 always
terminates. Consider the following program P :
1: P0 ;
2: if x0 = 1 then
3: x1 := 1;
4: while x1 6= 0 do
5: x1 := 1
6: od
7: fi

What does P do? It first runs P0 . If P0 returns 0, i.e., x0 = 0 after running

P0 , then P will terminate. If P0 returns 1, then P enters an infinite loop
and does not terminate. (Note that P0 either returns 0 or 1.)
Now assume that P terminates on input göd(P ). In this case, P0 returns
0 on göd(P ). But this means, that P does not terminate on göd(P ), a
contradiction.
If P does not terminates on göd(P ), then P0 returns 1 on göd(P ). But
this means that P terminates on göd(P ), again a contradiction.

42
7.2. Recursively enumerable languages 43

Since P either terminates on göd(P ) or does not, this case distinction is

exhaustive, and therefore, P0 cannot exist.

Excursus: Alan M. Turing and the halting problem

Alan M. Turing (born 1912, died 1954) was a British mathematician and cryptog-
rapher. He is one of the parents of Theoretical Computer Science.
He studied the halting problem (for Turing machines, a model that we will see
soon and that is equivalent to WHILE programs) to show that Hilbert’s Entschei-
dungsproblem is not decidable in 1936. (This was done indepently by Alonzo
Church whom we will meet later.) The Entscheidungsproblem is the problem in
symbolic logic to decide whether a given first-order statement is universally valid
or not. It was posed in this form by D. Hilbert in 1928 in the context of Hilbert’s
program.
During World War II, Turing contributed to the British efforts of breaking the
German ciphers. He died like snow white.

Further reading:
– The Turing Archive. www.turingarchive.org
– Andrew Hodges. Alan Turing: The Enigma, Walker Publishing Company, 2000.

7.2 Recursively enumerable languages

Definition 7.2 1. A language L ⊆ N is called recursively enumerable if
there is a WHILE program P such that

(a) for all x ∈ L, ϕP (x) = 1 and

(b) for all x ∈
/ L, ϕP (x) = 0 or ϕP (x) is undefined.

2. The set of all recursively enumerable languages is denoted by RE.

Remark 7.3 In condition 1.(b) of the definition above, we can always as-
sume that ϕP (x) is undefined. We can modify P in such a way that whenever
it returns 0, then it enters an infinite loop. Thus on x ∈ L, P halts (and
outputs 1), on x ∈/ L, P does not halt.

Theorem 7.4 The halting problem and the special halting problem are re-
cursively enumerable.

Proof. Let hg, mi be the given input. First we use the WHILE program
C to check whether g ∈ im göd. If not, then we enter an infinite loop. If
yes, then we simulate g on m using the univeral WHILE program U . It is
easy to see that this program terminates if and only if g encodes a WHILE
program and göd−1 (g) halts on m. If it terminates, then we return 1.

c Markus Bläser 2007–2015
44 7. The halting problem

Remark 7.5 The set that corresponds to the characteristic function c con-
structed in the alternative proof of Corollary 5.2 is not recursively enumer-
able, since we diagonalized against all WHILE programs not only those that
compute total functions.

Theorem 7.6 The following two statements are equivalent:

1. L ∈ REC.

2. L, L̄ ∈ RE.

Proof. For the “⇒” direction, note that L ∈ REC implies L̄ ∈ REC and
that REC ⊆ RE.
For the other direction, note that there are WHILE programs P and P̄
that halt on all m ∈ L and m ∈ / L. So either P or P̄ halts on a given m.
The problem is that we do not know which. If we run P first then it might
not halt on m ∈ L̄ and we never have a chance to run P̄ on m.
The trick is to run P and P̄ in parallel. To achieve this, we modify
our universal WHILE program U . In the while loop of U , we will simulate
one step of P and one step of P̄ . (We need two stacks S1 , S2 to do this,
two instances cur 1 , cur 2 of the variable cur , etc.) Eventually, one of the
programs P or P̄ will halt. Then we know whether m ∈ L or not.

Corollary 7.7 H̄0 ∈

/ RE.

Proof. We know that H0 is in RE but not in REC. By Theorem 7.6, H̄0

is not in RE.

Exercise 7.1 Show that the following three statements are equivalent:1

1. L ∈ RE.

2. There is a WHILE program P such that L = im ϕP .

3. L = ∅ or there is a FOR program P such that L = im ϕP .

1
This explains the name recursively enumerable: There is a WHILE computable func-
tion, here ϕP , that enumerates L, that means, if we compute ϕP (0), ϕP (1), ϕP (2), . . . , we
eventually enumerate all elements of L.

c Markus Bläser 2007–2015
8 Reductions

Let us come back to the verification problem: Does a given program match a
certain specification? One very general approach to model this is the follow-
ing: Given two encodings i, j ∈ im göd, do the WHILE programs göd−1 (i)
and göd−1 (j) compute the same function, i.e., is ϕgöd−1 (i) = ϕgöd−1 (j) . The
index i is the program that we want to verify, the index j is the specification
that it has to match. So let

V = {hi, ji | ϕgöd−1 (i) = ϕgöd−1 (j) }.

One can of course complain that a WHILE program is a very powerful spec-
ification. So we will also investigate the following (somewhat artifical but
undeniably simple) special case:

V0 = {i | ϕgöd−1 (i) (x) = 0 for all x ∈ N}.

So i ∈ V0 means that the WHILE program göd−1 (i) outputs a 0 on every

input (and in particular halts on every input).
Another relevant and basic problem is the termination problem: Does a
WHILE program halt on every input:

T = {i | ϕgöd−1 (i) is total}.

We will prove that none of these languages is decidable, i.e, V, V0 , T ∈

/
REC. So there is no hope for a general and automated way to predict/prove
semantic properties of WHILE programs and henceforth of computer pro-
grams in general.
We use a concept called reductions, a prominent concept in computer
science. Assume someone gives you a JAVA procedure that finds the mini-
mum in a list of integers. You can use this as a subroutine to quickly write
a JAVA program—though not the most efficient one—that sorts a list. We
will use the same principle but “in the reverse direction”. Assume there is
a WHILE program that decides V0 , say. Using this WHILE program, we
will construct a new WHILE program that decides the halting problem H0 .
Since the later does not exists, there former cannot exists either. Thus V0 is
not in REC, too.

8.1 Many-one reductions

Definition 8.1 Let L, L0 ⊆ N be two languages.

45
46 8. Reductions

1. A WHILE computable total function f : N → N is called a many-one

reduction from L to L0 if

for all x ∈ N: x ∈ L ⇐⇒ f (x) ∈ L0 .

2. If such an f exists, then we say that L is recursively many-one re-

ducible to L0 . We write L ≤ L0 in this case.

Example 8.2 If P were a program for H (we already know that it does not
exist), then the following P0 would be a program for H0 :
1: x0 := hx0 , x0 i;
2: P

The first line prepares the input for P . H0 expects a Gödel number, say g,
and the first line forms the pair hg, gi, the input for H. We have g ∈ H0
iff hg, gi ∈ H Questions about membership in H0 reduce to questions about
membership in H.
The second important property of reductions is that they should be “easy”:
If g is the Gödelnumber of P , then h4, hgöd(x0 := hx0 , x0 i), gii is the Gödel
number of P0 . Note that göd(x0 := hx0 , x0 i) is a constant that can be hard-
wired into the reduction.
The arguments above show H0 ≤ H.

The next lemma shows the importance of reductions. If L ≤ L0 , then

the fact that L0 is easy (that means, is in RE or REC) implies that L is easy,
too. The contraposition of this is: If L is hard (that is, is not in RE or REC)
then L0 is also hard.

Lemma 8.3 Let L, L0 ⊆ N and L ≤ L0 . Then the following statements hold:

1. If L0 ∈ RE, then L ∈ RE.

2. If L0 ∈ REC, then L ∈ REC.

Proof. Let f be a many one reduction from L to L0 . We prove the first

statement. Since L0 ∈ RE, there is a program P 0 such that ϕP 0 (m) = 1 for
all m ∈ L0 and ϕP 0 (m) is undefined for all m ∈
/ L0 . Consider the following
program P :
1: x0 := f (x0 )
2: P0

It computes ϕP = ϕP 0 ◦ f . We have

x ∈ L ⇒ f (x) ∈ L0 ⇒ ϕP 0 (f (x)) = 1 ⇒ ϕP (x) = 1

c Markus Bläser 2007–2015
8.1. Many-one reductions 47

and
x∈ / L0 ⇒ ϕP 0 (f (x)) is undefined ⇒ ϕP (x) is undefined.
/ L ⇒ f (x) ∈
Thus L ∈ RE (as witnessed by P ). The second statement is shown in the
same fashion, we just set ϕP 0 (m) = 0 if m ∈ / L0 and have to replace “is
undefined” in the second series of implications by “= 0”.

Corollary 8.4 Let L, L0 ⊆ N and L ≤ L0 . Then the following statements

hold:
/ RE, then L0 ∈
1. If L ∈ / RE.
/ REC, then L0 ∈
2. If L ∈ / REC.
Proof. These statement are the contrapositions of the statements of the
previous lemma.

Where is the subroutine idea gone...?

Assume there is a many one reduction f from H0 to V0 . From a

(hypothetical) WHILE program P that decides V0 , we get a program
Q for H0 as follows:
1: x0 := f (x0 );
2: P ;

So a many-one reduction is a very special kind of subroutine use: We

only use the subroutine (P ) once and the output of our program (Q)
has to be the output of the subroutine.
Why do we use many-one reductions? Because it is sufficient for
our needs and because we get finer results. For instance the first
statement of Lemma 8.3 is not true for Turing reductions (which is
a fancy name for arbitrary subroutine use).

Lemma 8.5 ≤ is a transitive relation.

Proof. Assume that L ≤ L0 and L0 ≤ L00 . Let f and g be the correspond-
ing reductions. Since f and g are WHILE computable, g ◦ f is, too. If P
and Q compute f and g, respectively, then P ; Q computes g ◦ f .
We claim that g ◦ f is a many-one reduction from L to L0 . We have
x ∈ L ⇐⇒ f (x) ∈ L0 ⇐⇒ g(f (x)) ∈ L00
for all x ∈ N by the definition of many-one reduction. This completes the
proof.

Exercise 8.1 Show the following: If L is many-one reducible to L0 , then L̄

is many-one reducible to L̄0 . (Hint: Just have a look at Figure 8.1.)

c Markus Bläser 2007–2015
48 8. Reductions

N N
L0
L
f

Figure 8.1: A reduction f maps the elements of L (also called “yes”-instances)

to a subset of the elements of L0 . The elements not in L (“no”-instances) are
mapped to a subset of the elements of N that are not in L0 . In short, “yes”-
instances go to “yes”-instances and “no”-instances to “no”-instances.

8.2 Termination and Verification

We show that neither V0 nor V nor V̄0 nor T are decidable.

Lemma 8.6 H0 ≤ V0 .

Proof. For a given input i, consider the following WHILE program Qi ,

where P := göd−1 (i):
1: x0 := i;
2: P ;
3: x0 := 0;

Note that Qi completely ignores its input. If P halts on its own Gödel
number i, then Qi always outputs 0, i.e., ϕQi (x) = 0 for all x. If P does not
halt on i, then Qi never halts, that is, ϕQi is the function that is nowhere
defined. In other words,

göd(P ) ∈ H0 ⇐⇒ gödTM (Qi ) ∈ V0 .

Thus the mapping f that maps

(
göd(Qi ) if i ∈ im göd
i 7→
i otherwise

is the desired reduction. (The case i ∈/ im göd is only of technical nature. If

i∈/ im göd, then i ∈
/ H0 and i ∈
/ V0 .)
We only have to convince ourselves that f is indeed WHILE computable.
The Gödel numbers of the three parts of Qi :
• h2, h0, iii.

c Markus Bläser 2007–2015
8.2. Termination and Verification 49

• i

• h2, h0, 0ii.

Therefore, the reductions is given by

i 7→ h4, hh2, h0, iii, h4, hi, h2, h0, 0iiiiii (8.1)

Note that the right hand side is not a Gödel number if i is not. (Therefore,
we do not exactly compute f but something similar which does the job, too.)
The above mapping is WHILE computable since the pairing function is. So
the reduction H0 ≤ V0 has an easy explicit description.
But wait, is it really this easy? Well, almost. Note that when concate-
nating two programs with Gödel numbers j and k, then the corresponding
Gödel number is h4, hj, kii. But this is only defined if the program corre-
sponding to j is a simple statement or a while loop. So if P above is not a
simple statement or a while loop, this reduction is not correct.
There are two ways to solve this. Either we can parse P and restructure
the whole program. This can be done in using a stack like the programs C
and U do. Or we can wrap P into a while loop that is executed exactly once:
1: xi := 1;
2: while xi 6= 0 do
3: xi := 0;
4: P
5: od

Note that the variable xi cannot appear in P since i is the Gödel number
of P . This works because our paring function increases monotonically and
every Gödel number contains the indices of the variables occuring in the
corresponding program. In the new program, we now either concatenate
programs where the first statement is a simple statement or a while loop. So
the final reduction is given by an expression similar to (8.1), more precisely
by
i 7→ c(h2, h0, iii, c(h2, hi, 1ii, c(h3, hi, c(h2, hi, 0ii, i)ii)))
where c(a, b) := h4, ha, bii. Thus f is WHILE computable.1

Exercise 8.2 Show that there is a WHILE program that given two encodings
i, j ∈ im göd, constructs the Gödel number of a WHILE program that is the
concatenation of göd−1 (i) and göd−1 (j). In the next chapter, we will see a
more formal way to do this.

Since V0 is a special case of V , V0 can be reduced to V .

1
Often, in our reductions, we are just concatenating the given (encoding of a) WHILE
program with some other encodings. This can usually be accomplished by a WHILE
computable total function.

c Markus Bläser 2007–2015
50 8. Reductions

Lemma 8.7 V0 ≤ V .

Proof. The reduction is given by i 7→ hi, ei where e is the Gödel number

of a WHILE program the constant function 0.
Verification is even harder than the halting problem, since V̄0 is not
recursively enumerable.

Lemma 8.8 H̄0 ≤ V0 .

Proof overview: When we showed H0 ≤ V0 , we constructed a WHILE

program Qi that simulated P := göd−1 (i) on i and outputted 0 if P halted.
Now the progra, Qi would have to output 0 if P does not halt. This straight
forward approach obviously does not work here.
A parameter that we did not use are the inputs of Qi . We will simulate
P only for a finite number of steps—the larger the input, the longer the
simulation.

Proof. For a given i ∈ im göd, consider the following WHILE program

Ki (using syntactic sugar at a maximum level):
1: Simulate P := göd −1(i) on i for x0 steps.
2: If P does not stop within x0 steps, then output 0
3: Otherwise output 1.

If i ∈
/ H0 , then in step 2 of Ki , P does not halt on i for any value of x0 .
Thus Ki always outputs 0. If i ∈ H0 , then there is a t ∈ N such that M
halts within t steps on i. Thus Ki will output 1 for every input ≥ t.
Thus the mapping
(
göd(Ki ) if i ∈ im göd,
i 7→
y otherwise,

where y is some element in V0 , is a many-one reduction from H̄0 to V0 . Note

that elements i ∈
/ im göd are in H̄0 . Like before, it is easy to see that this
mapping is WHILE computable.

Lemma 8.9 V0 ≤ T .

Proof. For a given i ∈ im göd, construct the following WHILE program

Ji :
1: Simulate P := göd−1
TM (i) on x0 .
2: If P does not output 0, then enter an infinite loop.
3: Otherwise halt.

c Markus Bläser 2007–2015
8.2. Termination and Verification 51

If P does not compute the function that is constantly 0, then either P

does not halt on some input or halts and outputs something else. In either
case, Ji does not terminate on this input.
Thus the mapping that maps
(
göd(Ji ) if i ∈ im göd
i 7→
i otherwise

is a many-one reduction from V0 to T .

Summarizing the result of this section, we get the following theorem. For
the complements of the mentioned languages, note that by Exercise 8.1, if,
say, H0 ≤ V0 then H̄0 ≤ V̄0 , too.

Theorem 8.10 V , V0 , and T are not recursively enumerable nor are their
complements.

How to construct a many-one reduction . . .

To reduce L to L0 :

1. First try to find for every i ∈ L an f (i) that is in L0 and for

every i ∈/ L, find an f (i) that is not in L0 .
If L is defined in terms of properties of functions computed by
WHILE programs (this is the “normal” case), try to map a given
encoding i that has (not) the property of L to an encoding f (i)
that has (not) the property of L0 .

2. Give a formal proof that your mapping f has indeed the reduc-
tion property.

3. "Prove" that f is WHILE computable. (A formal proof of this

might be tedious. A quick argument is usually sufficient.)

c Markus Bläser 2007–2015
9 More on reductions

From now on, if i ∈ im göd, we will use ϕi as a synonym for ϕgöd−1 (i) ; this
is only done to simplify notations as a reward for working through eight
chapters so far.

9.1 S-m-n Theorem

Theorem 9.1 (S-m-n Theorem) For every m, n ≥ 1, there is a FOR
computable function Snm : Nm+1 → N such that for all g ∈ im göd, y ∈ Nm ,
and z ∈ Nn
ϕm+n
g (y, z) = ϕnSnm (g,y) (z). 1
/ im göd, then Snm (g, y) ∈
Furthermore, if g ∈ / im göd.

Proof overview: The statement of the theorem looks complicated at a first

glance, but it just states the following simple thing: Given a Gödel number
of a program that expects n + m inputs and a number y ∈ Nm , we can
compute the Gödel number of a program that specializes the first m inputs
to y. While the statement of the theorem is quite simple, it is often very
useful.

Proof. Let Pg = göd−1 (g) the program that is given by g. Pg expects

m + n inputs. Given a y = (η0 , . . . , ηm−1 ), we now have to construct a
program Qg,y that depends on g and y and fulfills

ϕPg (y, z) = ϕQg,y (z) for all z ∈ Nn .

The following program Qg,y achieves this:

1: xm+n−1 := xn−1 ;
.
2: ..
3: xm := x0 ;
1
The superscripts n + m and n indicate the number of inputs to the program. Note
that the same program can have any number of inputs since we made the convention that
the inputs stand in the first variables at the beginning. So far, this number was always
clear from the context. Since it is important here, we added the extra superscript.
Strictly speaking ϕm+n
g expects one vector of length m + n. y is a vector of length m and
z is a vector of length n. If we glue them together, we get a vector of length m + n. For
the sake of simplicity, we write ϕm+n
g (y, z) instead of formally forming a vector of length
m + n out of y and z and then plugging this vector into ϕm+n g .

52
9.2. Reductions via the S-m-n Theorem 53

4: xm−1 := ηm−1 ;
.
5: ..
6: x0 := η0 ;
7: Pg

This program first copies the input z, which stands in the variables x0 , . . . , xn−1
into the variables xm , . . . , xm+n−1 . Then it stores y into x0 , . . . , xm−1 . The
values of y are hardwired into Qg,y . Then we run Pg on the input (x, y).
Thus Qg,y computes ϕPg (y, z) but only the entries from z are considered as
inputs.
The function
Snm : (g, y) 7→ göd(Qg,y )
is FOR computable. (We saw how to show this in the last chapter.) The
constructions in the last chapter were built in such a way that we automat-
ically get that if g is not a Gödel number, then (g, y) is not mapped to a
Gödel number either. This mapping above is the desired mapping, since

ϕm+n
g (y, z) = ϕm+n n n n
Pg (y, z) = ϕQg,y (z) = ϕgöd(Qg,y ) (z) = ϕSnm (g,y) (z)

for all y ∈ Nm and z ∈ Nn .

9.2 Reductions via the S-m-n Theorem

The S-m-n Theorem can be used to prove that a language is reducible to
another one. We here give an alternative proof of H̄0 ≤ V0 .
Alternative proof of Lemma 8.8. Consider the function f : N2 → N
defined by
(
/ im göd or göd−1 (g) does not halt on g after ≤ m steps
0 if g ∈
f (g, m) =
1 otherwise

The function f is WHILE computable, we can use C and the clocked version
of the universal Turing machine U . Let e be a Gödel number of f . By the
S-m-n Theorem,
f (g, m) = ϕ2e (g, m) = ϕS11 (e,g) (m)
for all g, m ∈ N. But by construction,

/ im gödTM or göd−1
g ∈ H̄0 ⇐⇒ g ∈ TM (g) does not halt on g
⇐⇒ f (g, m) = 0 for all N
⇐⇒ S11 (e, g) ∈ V0

Thus S11 (e, .) is the desired many one reduction.

c Markus Bläser 2007–2015
54 9. More on reductions

9.3 More problems

Here is another example, more of pedagogical value. Let c ∈ N. Let

Dc = {g | g ∈ im göd and | dom ϕgöd−1 (g) | ≥ c}

be the set of all encodings of Turing machines that compute a function that
is defined for at least c different arguments. Here is potential application:
As the last assignment of the Programmierung 2 lecture, you have to deliver
a program. You still need one point to be qualified for the exam. The TA
claims that your program does not halt on any input and you get no points
for your program. We will show that D1 ∈ / REC. This is good for you, since
it means that the TA will not be able to algorithmically verify his claim.
On the other hand, we will show that D1 ∈ RE, which is again good for
you, since it means that if your program halts on at least one input, you can
algorithmically find this input and maybe get the missing point. . .

Theorem 9.2 For every c ≥ 1, H0 ≤ Dc .

Proof. Consider the following function f : N2 → N defined by

(
0 if i ∈ im göd and ϕi (i) is defined,
f (i, x) =
undef. otherwise.

f is WHILE computable: Since f (i, x) is undefined if ϕi (i) is undefined, we

can just simulate göd−1 (i) on i and return 0 if göd−1 (i) halts.
Since f is WHILE computable, there is an index e ∈ im göd such that
f = ϕe . By the S-m-n Theorem,

f (i, x) = ϕ2e (i, x) = ϕS11 (e,i) (x) for all i and x.

Let i ∈ H0 . Then f (i, x) is defined for all x by construction, i.e., the function
x 7→ f (i, x) is total and in particular, its domain has at least c elements.
Thus S11 (e, i) ∈ Dc . Let i ∈ / H0 . Then f (i, x) is not defined for all x by
1
construction. Thus S1 (e, i) ∈ / Dc . The function i 7→ S11 (e, i) is recursive by
the S-m-n theorem (note that e is just a fixed number), thus it is the desired
reduction.

Theorem 9.3 For every c, Dc ∈ RE.

Proof. It is sufficient to prove that Dc = dom ϕP for some WHILE

program P .

c Markus Bläser 2007–2015
9.3. More problems 55

The golden rule of confusion

If something is not complicated enough,
invent many names for the same thing.
(In our case, history is responsible for this)

For a language L ⊆ N, the following statements are equivalent or just mean

the same thing by definition:
L ∈ REC, L is decidable, L is recursive, the characteristic function χL is
WHILE computable.
Also the following statements are equivalent or mean the same: L ∈ RE,
L is recursively enumerable, χ0L is WHILE computable (do not overlook
the prime!).
For a function f : N → N, the following statements are equivalent or
mean the same: f is recursive (but f can be partial!) and f is WHILE
computable.

c Markus Bläser 2007–2015
10 Rice’s Theorem

We saw numerous proofs that certain languages are not decidable. Rice’s
Theorem states that any language L is not decidable if it is defined in se-
mantic terms. This means that whether i ∈ im göd is in L only depends on
ϕi , the function computed by the machine göd−1 (i).

10.1 Recursion Theorem

Theorem 10.1 (Recursion Theorem) For every WHILE computable func-
tion f : Nn+1 → N, there is a g ∈ im göd such that

ϕng (z) = f (g, z) for all z ∈ Nn .

Proof overview: Let f (g, z) = ϕe (g, z). Now the S-m-n Theorem states
that f (g, z) = ϕSn1 (e,g) (z) for all z. If we now set g = e, then we are almost
there: we have e on the left-hand side and Sn1 (e, e) on the right-hand side
which is “almost the same”. If we now replace g, the first argument of f , by
something of the form Sn1 (y, y), then basically the same argument gives the
desired result.

Proof. The function h defined by

h(y, z) := f (Sn1 (y, y), z) for all y ∈ N, z ∈ Nn

is WHILE computable. Let e be a Gödel number for h, i.e., h = ϕe . The

S-m-n Theorem implies that

ϕn+1
e (y, z) = ϕnSn1 (e,y) (z) for all z ∈ Nn .

If we now set y = e and g = Sn1 (e, e), we get

f (g, z) = f (Sn1 (e, e), z) = h(e, z) = ϕn+1

e (e, z) = ϕnSn1 (e,e) (z) = ϕng (z).

Remark 10.2 Given the index e of h, we can compute g by a WHILE ma-

chine.

Exercise 10.1 Show that there is a Gödel number j with dom ϕj = {j}.

56
10.1. Recursion Theorem 57

10.1.1 Self reference

Consider the function s : N2 → N given by s(y, z) = y for all y, z ∈ N. This
function is certainly WHILE computable. By the Recursion Theorem, there
is an index g such that

ϕg (z) = s(g, z) = g for all z ∈ N.

This means that the WHILE program given by the Gödel number g computes
the constant function with value g and in this sense outputs its own source
code.

10.1.2 The halting problem again

Here is an alternative proof that the halting problem H is not decidable:
Assume that H is decidable and P is a WHILE program that decides H.
Then following function
(
0 if ϕe (x) is undefined
f (e, x) =
undefined otherwise

is Turing computable since we can check whether ϕe (x) is defined by envoking

P . By the Recursion Theorem, there is an index e0 such that

ϕe0 (x) = f (e0 , x).

Assume that ϕe0 (e0 ) is defined. Then f (e0 , e0 ) = ϕe0 (e0 ) is undefined by
construction, a contradiction. But if ϕe0 were undefined, then f (e0 , e0 ) =
ϕe0 (e0 ) = 0, a contradiction again. Thus H cannot be decidable.

10.1.3 Code minimization

Finally, consider the following language

Min = {g | for all g 0 with ϕg = ϕg0 , g ≤ g 0 }.

This is the set of all minimal WHILE programs (“shortest source codes”) in
the sense that for every g ∈ Min, whenever g 0 computes the same functions
as g, then g ≤ g 0 .

Theorem 10.3 Min ∈

/ RE.

Proof. The proof is by contradiction. Assume that Min ∈ RE. Then,

since Min 6= ∅ there is a WHILE computable total function h such that
Min = im h. The function

f : (g, w) 7→ f (g, w) = ϕk (w) with k = h(j) and j = min{i | g < h(i)}

c Markus Bläser 2007–2015
58 10. Rice’s Theorem

is WHILE computable since k can be determined by a WHILE program: We

successively compute h(0), h(1), h(2), . . . until we hit a j such that g < h(j).
Such a j exists, since Min is infinite.
By the recursion theorem, there is a Gödel number e such that

ϕe (w) = f (e, w) = ϕk (w) for all w ∈ N.

By the construction of f , e < k. But we also have k ∈ Min = im h. This is

a contradiction, as ϕe = ϕk and e < k implies k ∈
/ Min.

10.2 Fixed Point Theorem

A fixed point of a function f : N → N is a z0 ∈ {0, 1}∗ such that f (z0 ) = z0 .
Not every function has a fixed point, z 7→ z + 1 is an example. But every
Turing computable total function f has a semantic fixed point in the sense
that z0 and f (z0 ) are the Gödel numbers of Turing machines that compute
the same function.

Theorem 10.4 (Fixed Point Theorem) For all Turing computable total
functions f : N → N with im f ⊆ im göd and for all n ∈ N \ {0} there is an
e ∈ im göd such that
ϕnf(e) = ϕne .

Proof. Let g(z, y) = ϕnf(z) (y) for all z ∈ N, y ∈ Nn . g is Turing com-

putable since f is Turing computable and total.
By the Recursion Theorem, there is an e ∈ im gödTM such that

ϕne (y) = g(e, y) = ϕnf(e) (y) for all y ∈ Nn .

10.3 Rice’s Theorem

Definition 10.5 (Index set) A language I ⊆ im göd is an index set if

for all i, j ∈ im göd: i ∈ I and ϕi = ϕj =⇒ j ∈ I.

An index set I is nontrivial if, in addition, I 6= ∅ and I 6= im göd.

Exercise 10.2 Show that I is an index set if and only if there is a set F of
WHILE computable functions such that I = {i ∈ im göd | ϕi ∈ F }.

If an index set contains a Gödel number i, then it contains all Gödel

numbers j of Turing machines that compute the same function as göd−1 (i).
In this sense, the index sets are defined by semantic properties, i.e., properties
that only depend on ϕi .

c Markus Bläser 2007–2015
10.3. Rice’s Theorem 59

Example 10.6 The following languages are index sets:

1. V0 = {i ∈ im göd | ϕi (x) = 0 for all x ∈ N}.
2. T = {i ∈ im göd | ϕi is total}
3. Dc = {i ∈ im göd | | dom ϕi | ≥ c} for any c ∈ N,
4. Mon = {i ∈ im göd | ϕi is monotone}.
All of them are nontrivial except D0 .
Example 10.7 The following sets are not index sets:
1. H0 , the special halting problem,
2. H, the halting problem,
3. N1 = {g ∈ im göd | g ≤ 10000}.
H0 is not an index set, since we constructed a Gödel number j with
dom ϕj = {j} in Exercise 10.1. Thus j ∈ H0 but any other Gödel number k
that computes the same function does not halt on its own Gödel number k.
H is not an index set since it consists of pairs of the form hg, mi, which
might not be Gödel numbers. But this is of course not a formal argument,
by chance all of the pairs could be Gödel numbers. We leave it as an exercise
to construct a pair hg, mi ∈/ im göd with g ∈ im göd.
N1 is not an index set since for any function f , there are arbitrarily large
WHILE programs, i.e., with arbitrarily large Gödel numbers, that compute
f.
Theorem 10.8 (Rice’s Theorem) Every nontrivial index set is not de-
cidable.
Proof. Let I be a nontrivial index set. Since I is nontrivial, there are
Gödel numbers i and j such thats i ∈ I but j ∈/ I. If I were decidable, then
the function h : N → N defined by
(
i if x ∈/I
h(x) =
j if x ∈ I
would be Turing computable. Since im h ⊆ im göd, there is a Gödel number
e by the Fixed Point Theorem such that
ϕe = ϕh(e) .
If e ∈ I, then ϕe = ϕj . But since I is an index set and j ∈
/ I, we get e ∈
/ I, a
contradiction. If e ∈
/ I, then ϕe = ϕi . But since I is an index set and i ∈ I,
we get e ∈ I, a contradiction again. Thus I cannot be decidable.
Rice’s Theorem essentially says that every nontrivial semantic property
of Turing machines is undecidable!

c Markus Bläser 2007–2015
60 10. Rice’s Theorem

10.4 Further exercises

There are nontrivial index sets that are recursively enumerable, for instance,
Dc . Others are not, like T or V0 . Here is a criterion that is useful to prove
that an index set is not in RE.

Exercise 10.3 Let I be a recursively enumerable index set. Show that for
all g ∈ I, there is an e ∈ I with dom ϕe ⊆ dom ϕg and dom |ϕe | is finite.

c Markus Bläser 2007–2015
11 Gödel’s incompleteness theorem

Loosely speaking, Gödel’s incompleteness theorem states that there are for-
mulas that are true but we cannot prove that they are true. Formulas here
means quantified arithmetic formulas, i.e., we have formulas with existen-
tial and universal quantifiers over the natural numbers with addition and
multiplication as our operations. “We cannot prove” means that there is no
effective way to show that the formula is true.

11.1 Arithmetic terms and formulas

Definition 11.1 Let V = {x0 , x1 , x2 , . . . } be a set of variables.1 Arithmetic
terms over V are defined inductively:

1. Every n ∈ N is an arithmetic term.

2. Every x ∈ V is an arithmetic term.

3. If s and t are arithmetic term, then (s + t) and (s · t) are arithmetic

terms, too.
(These are words over the infinite alphabet N ∪ V ∪ {(, ), +, ·}, not
polynomials or something like that.)

Definition 11.2 Arithmetic formulas are defined inductively:

1. If s and t are terms, then (s = t) is an arithmetic formula.

2. If F and G are arithmetic formulas, then ¬F , (F ∨ G), and (F ∧ G)

are arithmetic formulas.

3. If x is a variable and F is an arithmetic formula, then ∃xF and ∀xF

are arithmetic formulas.

Let F and G be formulas. We define the fact that G is a subformula of

F inductively: If G = F , then G is a subformula of F . If F = ¬F1 and G
is a subformula of F1 , then G is also a subformula of F . In the same way,
if F = (F1 ∨ F2 ) or F = (F1 ∧ F2 ) or F = ∃F1 or F = ∀xF1 and G is a
subformula of F1 or F2 , then G is also a subformula of F .
Let x be a variable and F be a formula. The occurences of x in F are
these position in F that contain the symbol x. An occurence of x in the
1
As usual, we will use other names for variables, too.

61
62 11. Gödel’s incompleteness theorem

formula F is bounded if this occurence is contained in a subformula of F of

the form ∃xG or ∀xG. An ocurrence that is not bounded it is called free.
F (x/n) denotes the formula that we get if we replace every free occurence
of x in F by n ∈ N.
A mapping a : V → N is called an assignment. We extend a to the set
of arithmetic terms in the obvious way:

a(n) = n for all n ∈ N,

a(s + t) = a(s) + a(t) for all terms s and t,
a(s · t) = a(s)a(t) for all terms s and t

Definition 11.3 We define true formulas inductively:

1. If s and t are terms, then (s = t) is true if a(s) = a(t) for all assign-
ments a.

2. F = ¬F1 is a true formula, if F1 is not a true formula

3. F = (F1 ∨ F2 ) is a true formula if F1 or F2 are true formulas.

4. F = (F1 ∧ F2 ) is a true formulas if F1 and F2 are true formulas.

5. F = ∃xF1 is a true formula if there is an n ∈ N such that F1 (x/n) is

a true formula.

6. F = ∀xF1 is a true formula if for all n ∈ N, F1 (x/n) is a true formula.

A formula that is not true is called false.

We define a function e that is an injective mapping from the set of all

arithmetic terms and formulas to N. It is defined inductively, in the same
manner we defined the mapping göd:

1. e(n) = h0, ni for all n ∈ N.

2. e(xi ) = h1, ii for all i ∈ N.

3. e(s + t) = h2, he(s), e(t)ii for all terms s and t.

4. e(s · t) = h3, he(s), e(t)ii for all terms s and t.

5. e(s = t) = h4, he(s), e(t)ii for all terms s and t.

6. e(¬F ) = h5, e(F )i for all formulas F .

7. e(F ∨ G) = h6, he(F ), e(G)ii for all formulas F and G.

8. e(F ∧ G) = h7, he(F ), e(G)ii for all formulas F and G.

9. e(∃xi F ) = h8, hi, e(F )ii for all formulas F .

c Markus Bläser 2007–2015
11.2. Computability and representability 63

10. e(∀xi F ) = h9, hi, e(F )ii for all formulas F .

It is easy to see that the set im e is decidable. Like for göd, the concrete
construction is not so important.

Definition 11.4 The set of all encodings of true formulas is denoted by T .

11.2 Computability and representability

In this section we establish a link between formulas and WHILE programs.
If F is a formula and y1 , . . . , yk are exactly these variables that occur free
in in F , then we indicate this by writing F (y1 , . . . , yk ). In this context,
instead of writing F (y1 /n1 , . . . , yk /nk ), we often just write F (n1 , . . . , nk ) for
the formula in which every free occurence of yκ is replaced by nκ , 1 ≤ κ ≤ k.

Definition 11.5 A function f : Nk → N is called arithmetically repre-

sentable if there is an arithmetic formula F (y1 , . . . , yk , z) such that

f (n1 , . . . , nk ) = s ⇐⇒ F (n1 , . . . , nk , s) is true

for all n1 , . . . , nk , s ∈ N.2 In the same way, we define arithmetical repre-

sentability for funcions f : Nk → Nm .

Example 11.6 1. The addition function is arithmetically representable

by
(z = (y1 + y2 )).
In the same way, we can represent the multiplication function.

2. The modified difference function (y1 , y2 ) 7→ max{y2 − y1 , 0} is arith-

metically representable by

((y1 + z) = y2 ) ∨ ((y1 > y2 ) ∧ (z = 0))).

Above, (y1 > y2) shorthands

∃h(y1 = y2 + h + 1).

3. Division with remainder is arithmetically representable, too: y1 DIV y2

is represented by

∃r((r < y2 ) ∧ (y1 = z · y2 + r))

and y1 MOD y2 is represented by

∃q((r < y2 ) ∧ (y1 = q · y2 + r)).

2
If f (n1 , . . . , nk ) is undefined, then F (n1 , . . . , nk , s) is not true for all s ∈ N.

c Markus Bläser 2007–2015
64 11. Gödel’s incompleteness theorem

11.2.1 Chinese remaindering and storing many number in one

Our goal is to show that every WHILE computable function is arithmetically
representable. One ingredient in this construction is a method of storing
many natural numbers in one natural number. We saw such a method when
we constructed dynamic arrays in WHILE programs. To access the elements
of the arrays we used a FOR loop to repeatedly apply one of the two “inverse”
functions π1 and π2 of the pairing function h., .i. Arithmetic formulas have
only a fixed number of variables, too, but do not have FOR loops. Therefore
we here construct another method for storing many values in one.

Theorem 11.7 (Chinese remainder theorem—concrete version) Let

n1 , . . . , nt be pairwise coprime, i.e., gcd(ni , nj ) = 1 for i 6= j. Then the map-
ping

πn1 ,...,nt : {0, . . . , n1 · · · nt − 1} → {0, . . . , n1 − 1} × · · · × {0, . . . , nt − 1}

m 7→ (m mod n1 , . . . , m mod nt )

is a bijection.3

Proof. The proof is by induction on t.

Induction base: : Assume that t = 2.4 Since n1 and n2 are coprime, there
are integers c1 and c2 such that 1 = c1 n1 + c2 n2 .5 Then

c1 n1 mod n2 = 1 and c2 n2 mod n1 = 1.

Since the sets {0, . . . , n1 n2 − 1} and {0, . . . , n1 − 1} × {0, . . . , n2 − 1} have

cardinality n1 n2 , it is sufficient to show that the mapping is surjective. Let
(a1 , a2 ) ∈ {0, . . . , n1 − 1} × {0, . . . , n2 − 1} be given. Consider a = a1 c2 n2 +
a2 c1 n1 . We have

a mod n1 = a1 and a mod n2 = a2

since 1 = c1 n1 + c2 n2 . a might not be in {0, 1, . . . , n1 n2 − 1}, but there is

an integer i such that a0 = a + in1 n2 is. Since n1 , n2 |in1 n2 ,

a0 mod n1 = a1 and a0 mod n2 = a2 ,

too.
3
i mod j here denotes the unique integer r ∈ {0, 1, . . . , j − 1} such that i = qj + r for
some q.
4
We could also assume that t = 1. Then the induction base would be trivial. But it
turns out that we would have to treat the case t = 2 in the induction step, so we can to
it right away.
5
We can get such integers, the so-called cofactors, via the extended Euclidian algorithm
for computing gcds.

c Markus Bläser 2007–2015
11.2. Computability and representability 65

Induction step: : Let N = n2 · · · nt with t > 2. n1 and N are coprime. By

the induction hypothesis, the mappings πn2 ,...,nt and πn1 ,N are bijections.
We have (m mod N ) mod ni = i mod ni for all 2 ≤ i ≤ t, since ni |N . Thus

πn1 ,...,nt (m) = (m1 , πn2 ,...,nt (m2 ))

where πn1 ,N = (m1 , m2 ). Since both mappings above are bijections, their
“composition” πn1 ,...,nt is a bijection, too.

Lemma 11.8 The number 1 + i · s!, 1 ≤ i ≤ s, are pairwise coprime.

Proof. Assume there are i < j and a prime number p such that p|(1+i·s!)
and p|(1 + j · s!). Thus p|((j − i) · s!). Since 0 ≤ j − i ≤ s and p is prime,
p|s!. From this p|1 follows.

Lemma 11.9 For all numbers a1 , . . . , ak , there are numbers A and S and
a formula M (x, u, v, w) such that M (x, κ, A, S) is true if and only if we
substitute aκ for x.

Proof. Consider

M (x, u, v, w) = (x = v mod (1 + uw)) ∧ (v < 1 + uw).

Now set s = max{a1 , . . . , ak , k}. By Lemma 11.8, the numbers 1 + iS,

1 ≤ i ≤ s are pairwise coprime. By the Chinese remainder theorem, there is
an A such that
aκ = A mod (1 + κS)
for 1 ≤ κ ≤ k, since aκ ≤ 1+κS by definition. Thus the formula M (aκ , κ, A, S)
is true. The second part of M ensures that no other value fulfills M (x, κ, A, S).

11.2.2 Main result

Theorem 11.10 Every WHILE computable function is arithmetically rep-
resentable.

Proof. We show by structural induction that for every WHILE program

P then function ΦP : N` → N` , where ` is the largest index of a variable
in P , is arithmetically representable by a formula FP (y0 , . . . , yk , z1 , . . . , zk ).
From this statement, the statement of the theorem follows, since

F (y0 , . . . , ys , z) = ∃a1 . . . ∃a` FP (y0 , . . . , ys , 0, . . . , 0, z, a1 , . . . , a` )

represents the function Ns+1 → N computed by P .

c Markus Bläser 2007–2015
66 11. Gödel’s incompleteness theorem

Induction base: If P = xi := xj + xk , then we set

^
FP (y0 , . . . , yk , z0 , . . . , zk ) = (zi = yj + yk ) ∧ zm = ym .
m6=i

If P = xi := xj − xk , then FP looks similar, we just replace the (zi =

yj + yk ) part by the formula for the modified difference from Example 11.6.
If P = xi := c, then we replace the (zi = yj + yk ) part by (zi = c).
Induction step: We first consider the case P = P1 ; P2 . By the induction
hypothesis, the functions ΦPi : N` → N` 6 are arithmetically representable
by formulas FPi , i = 1, 2. We have ΦP = ΦP2 ◦ ΦP1 . The formula

FP = ∃a0 . . . ∃a` FP1 (y0 , . . . , y` , a0 , . . . , a` ) ∧ FP2 (a0 , . . . , a` , z0 , . . . , z` )

represents ΦP ; the variables a0 , . . . , a` “connect” the two formulas in such a

way that the output of ΦP1 becomes the input of Φp2 .
It remains the case P = while xi 6= 0 do P1 od. Let FP1 be a formula
that represents ΦP1 This is more complicated, since we have to “connect” a
formula FP1 for an unknown number of times. We will use the formula M
from Lemma 11.9.

FP =∃A0 ∃S0 . . . ∃A` ∃S` ∃t

M (y0 , 0, A0 , S0 ) ∧ · · · ∧ M (y` , 0, A` , S` )∧
M (z0 , t, A0 , S0 ) ∧ · · · ∧ M (z` , t, A` , S` )∧
∀τ ∃v((τ ≥ t) ∨ (M (v, τ, Ai , Si ) ∧ (v > 0)))∧
M (0, t, Ai , Si )∧
∀τ ∃a0 . . . ∃a` ∃b0 . . . ∃b`
FP1 (a0 , . . . , a` , b0 , . . . , b` )∧
M (a0 , τ, A0 , S0 ) ∧ · · · ∧ M (a` , τ, A` , S` )∧
M (b0 , τ + 1, A0 , S0 ) ∧ · · · ∧ M (b` , τ + 1, A` , S` )



∨ (τ ≥ t)

The variable t denotes the number of times the while loop is executed. The
variables Ai and Si store values that encode the values that xi attains after
each execution of P1 in the while loop. The second line of the definition
of FP ensures that before the first execution, the value of the variable xλ
is yλ , 0 ≤ λ ≤ `. The third line ensures that after the tth execution, the
value of the variable xλ is zλ , 0 ≤ λ ≤ `. The fourth and fifth line ensure
that the first time that xi contains the value 0 is after the tth execution of
the while loop. The remainder of the formula ensures that the values that
x0 , . . . , x` have after the (τ + 1)th execution of the while loop are precisely
6
P1 or P2 might not contain the variable x` . We pad ΦPi to a function N` → N` in the
obvious way.

c Markus Bläser 2007–2015
11.3. Proof systems 67

the values that we get if we run P1 with x0 , . . . , x` containing the values after
τ th execution. Note that the formula M is satisfied by at most one value
for fixed τ , Aλ , and Sλ . This ensure consistency, i.e, even if we Aλ and Sλ
do not contain the values from Lemma 11.9, if the formula FP is satisfied,
then the values stored in Aλ and Sλ , 0 ≤ λ ≤ `, correspond to an execution
of the WHILE program P .

Remark 11.11 Furthermore, there is a WHILE program that given göd(P ),

computes the encoding of a formula presenting ϕP .

Lemma 11.12 If T ∈ RE, then T ∈ REC.

Proof. Let f be a total WHILE computable functions such that im f = T .

A WHILE program P that decides T first checks whether a given input x ∈
im e. Let e(F ) = x. P successively computes f (0), f (1), f (2), . . . , f (i), . . .
until either e(F ) = f (i) for e(¬F ) = f (i). In the first case, P outputs 1,
in the second, 0. Since either F or ¬F is true, P halts on all inputs and
therefore decides T .

Theorem 11.13 T ∈
/ RE.

Proof. Let L ∈ RE \ REC and F (y, z) be a formula that represents χ0L .

Let x = e(F ). It is easy to construct a WHILE program that given x and
an n ∈ N, computes an encoding en of the formula F (n, 1). Since

χ0L (n) = 1 ⇐⇒ F (n, 1) is true ⇐⇒ en ∈ T ,

the mapping n 7→ en is a many-one reduction from L to T .

11.3 Proof systems

What is a proof? This is a subtle question whose answer is beyond the scope
of this chapter. But here we only need two properties that without any doubt
are properties that every proof system should have: The first is that the set
of all (encodings of) correct proofs should be decidable, that is, there is a
WHILE program that can check whether a given proof is true. The second
one is that there should be a total WHILE computable mapping that assigns
to each proof the formula that is proven by this proof. Technically, proofs
are finite words over some alphabet of finite size. Thus we can view them as
natural numbers by using any “easy” injective mapping into N.

Definition 11.14 A proof system for a set L ⊆ N is a tuple (P, F ) such

that

c Markus Bläser 2007–2015
68 11. Gödel’s incompleteness theorem

1. P ⊆ N is decidable and

2. F : P → L is a total WHILE computable function.

We think of P of the set of (encodings of) proofs for the elements of L.

The mapping F assigns each proof p ∈ P the element of L that is proved by
p.

Definition 11.15 A proof system (P, F ) for L is complete if F is surjective.

Theorem 11.16 There is no complete proof system for the set of all true
arithmetic formulas T .

Proof. Assume there would be a complete proof system (P, F ). The

mapping (
F (p) if p ∈ P
f : p 7→
undef. otherwise
is WHILE computable. By construction im f = T . But this contradicts
Theorem 11.13.

c Markus Bläser 2007–2015
12 Turing machines

Turing machines are another model for computability. They were intro-
duced by Alan Turing in the 1930s to give a mathematical definition of an
algorithm. When Turing invented his machines, real computers were still
to be built. Turing machines do not directly model any real computers
or programming languages. They are abstract devices that model abstract
computational procedures. The intention of Alan Turing was to give a for-
mal definition of “intuitively computable”; rather than modeling computers,
he modeled mathematicians. We will see soon that Turing machines and
WHILE programs essentially compute the same functions.

Why Turing machines?

Turing machines are the model for computations that you find in
the textbooks. In my opinion, WHILE programs are easier to un-
derstand; it usually takes some time to get familiar with Turing ma-
chines.
I hope that at the end of this part you will see that it does not really matter
whether one uses Turing machines or WHILE programs. All we need is a
Gödel numbering, a universal Turing machine/WHILE program, and the
ability to compute a Gödel number of the composition of two programs
from the two individual Gödel numbers, i.e. an acceptable programming
system. In theory, computer scientist are modest people.

12.1 Definition
A Turing machine M has a finite control and a number, say k, of tapes. The
finite control is in one of the states from a set of states Q. Each of the tapes
consists of an infinite number of cells and each cell can store one symbol
from a finite alphabet Γ, the tape alphabet. (Here, finite alphabet is just a
fancy word for finite set.) Γ contains one distinguished symbol, the blank
. Each tape is two-sided infinite.1 That is, we can formally model it as a
function T : Z → Γ and T (i) denotes the content of the ith cell. Each tape
has a head that resides on one cell. The head can be moved back and forth
on the tape, in a cell by cell manner. Only the content of the cells on which
the heads currently reside can be read by M . In one step,
1
In some textbooks, the tapes are only one-sided infinite. As we will see soon, this
does not make any difference.

69
70 12. Turing machines

··· 0 1 1 0 0 0 1 1 ···

··· 0 0 1 1 0 0 1  ···

..

··· 1 0       ···

Figure 12.1: A Turing machine

1. M reads the content of the cells on which its heads reside,

2. then M may change the content of these cells.

3. M moves each head either one cell to the left, not at all, or one cell to
the right.

4. Finally, it changes its state.

The behaviour of M is described by a transition function

δ : Q × Γk → Q × Γk × {L, S, R}k .

δ can be a partial function. δ(q, γ1 , . . . , γk ) = (q 0 , γ10 , . . . , γk0 , r1 , . . . , rk )

means that if M is in state q and reads the symbols γ1 , . . . , γk on the tapes
1, . . . , k, then it will enter state q 0 , replace the symbol γ1 by γ10 on the first
tape, γ2 by γ20 on the second tape, etc., and move the heads as given by
r1 , . . . , rk . (L stands for “left”, S for “stay”, and R for “right”. If the head
stands in position i of the tape, then “left” means that the head moves to po-
sition i−1 and “right” means that it moves to position i+1.) If δ(q, γ1 , . . . , γk )
is undefined, then M halts. Figure 12.1 is a schematic of a Turing machine.
I do not know whether it is really helpful.

Definition 12.1 A k-tape Turing machine M is described by a tuple (Q, Σ, Γ, δ, q0 )

where:

c Markus Bläser 2007–2015
12.1. Definition 71

1. Q is a finite set, the set of states.

2. Σ is a finite alphabet, the input alphabet.

3. Γ is a finite alphabet, the tape alphabet. There is a distiguished symbol

 ∈ Γ, the blank. We have Σ ⊆ Γ \ {}.

4. δ : Q × Γk → Q × Γk × {L, S, R}k is the transition function.

5. q0 ∈ Q is the start state.

In the beginning, all tapes are filled with blanks. The only exception is
the first tape; here the input is stored. The input of a Turing machine is a
string w ∈ Σ∗ where Σ ⊆ Γ \ {} is the input alphabet. It is initially stored
in the cells 0, . . . , |w| − 1 of the first tape. All heads stand on the cell with
number 0 of the corresponding tape. The Turing machine starts in its start
state q0 and may now perform one step after another as described by the
transition function δ.

Example 12.2 Let us consider a first example: We will construct a 1-tape

Turing machine INC that increases a given number in binary by 1. We
assume that the head stands on the bit of lowest order and that in the end,
the head will stop there again. The lowest order bit stands on the left and
the highest order bit on the right. What does INC do? If the lowest order bit
is a 0, then INC replaces it by a 1 and is done. If the lowest order bit is 1,
then INC replaces it by a 0. This creates a carry and INC goes one step to
the right and repeats this process until it finds a 0 or a . The latter case
occurs if we add 1 to a number of the form 111 . . . 1.
INC has three states add, back, and stop. The state add is the start state.
The input alphabet is Σ = {0, 1} and the tape alphabet is Γ = {0, 1, }. The
transition function is given by the following table:

0 1 
add (back, 1, L) (add, 0, R) (back, 1, L)
back (back, 0, L) (back, 1, L) (stop,, R)
stop — — —

Above, “—” stands for undefined. In the state add, INC goes to the right
replacing every 1 by a 0 until it finds the first 0 or . This 0 or  is then
replaced by a 1 and INC enters the state back. In the state back, INC goes
to the left leaving the content of the cells unchanged until it finds the first .
It goes one step to the right and is done.
Instead of a table, a transition diagram is often more understandable.
Figure 12.2 show this diagram for the Turing machine of this exercise. The
states are drawn as circles and an arrow from q to q 0 with the label “α; β, r”

c Markus Bläser 2007–2015
72 12. Turing machines

1;1,L
1;0,R 0;0,L

0;1,L
1,L R
add back stop

Figure 12.2: The transition diagram of the Turing machine INC from Ex-
ample 12.2.

means that if the Turing machine is in state q and reads α, then it goes to
state q 0 , writes β, and moves its head as given by r ∈ {L, S, R}.

12.2 Configurations and computations

While computing, a Turing machine can change its state, change the posi-
tions of the heads, and the content of the tapes. The state together with the
positions of the heads and the content of the tapes is called a configuration.
While a tape of a Turing machine is potentially infinite, within any finite
number of steps, the machine can visit only a finite number of cells. Only
these cells can contain symbols other than . The only exeption is the first
tape on which the input stands in unvisited cells at the beginning. So instead
of modeling a tape as a function Z → Γ (which is the same as a two-sided
infinite word over Γ), we just store the relevant parts of the tape, i.e., the
cells that have already been visited and—in the case of the first tape—the
cells were the input is written.
Formally, a configuration C of a k-tape Turing machine is an element
(q, (p1 , x1 ), . . . (pk , xk )) ∈ Q × (N × Γ∗ )k such that 1 ≤ pκ ≤ |xκ | for all
1 ≤ κ ≤ k. q ∈ Q is the current state of M . x1 , . . . , xk is the content of the
cells visited so far of the tapes 1, . . . , k. pκ denotes the position of the head
of the tape κ, 1 ≤ κ ≤ k. We store the position relatively, i.e., pκ denotes
the position within xκ but not necessarily the absolute position on the tape.
The start configuration of M = (Q, Σ, Γ, δ, q0 ) with input w is the config-
uration (q0 , (1, w), (1, ), . . . , (1, )). The input w stands on the first tape
and the head is on the first symbol of it. On all other tapes, only one cell has
been visited so far (the one the head is residing on), and this cell necessarily
contains a . We usually denote the start configuration by SCM (w).
Let C = (q, (p1 , x1 ), . . . (pk , xk )) and C 0 = (q 0 , (p01 , x01 ), . . . (p0k , x0k )) be
two configurations and let xκ = uκ ακ vκ with |uκ | = pκ − 1 and ακ ∈ Γ for
1 ≤ κ ≤ k. In other words, ακ is the symbol of the cell the head is residing
on. Then C 0 is called a successor of C if C 0 is reached from C by one step of

c Markus Bläser 2007–2015
12.2. Configurations and computations 73

M . Formally this means that if δ(q, α1 , . . . , αk ) = (q 0 , β1 , . . . , βk , r1 , . . . , rk ),

then we have for all 1 ≤ κ ≤ k,

x0κ = uκ βκ vκ

and 
pκ − 1 if rκ = L,

0
pκ = pκ if rκ = S,

pκ + 1 if rκ = R,


unless pκ = 1 and rκ = L or pκ = |xκ | and rκ = R. In the latter two cases,

M is visiting a new cell. In these cases, we have to extend xκ by one symbol.
If pκ = 1 and rκ = L, then
x0κ = βκ vκ
and
p0κ = 1.

If pκ = |xκ | and rκ = R, then

x0κ = uκ βκ 

and
p0κ = |xκ | + 1.
We denote the fact that C 0 is a successor of C by C `M C 0 . Note that by
construction, each configuration has at most one successor. We denote the
reflexive and transitive closure of the relation `M by `∗M , i.e., C `∗M C 0 iff
there are configurations C1 , . . . , C` for some ` such that C `M C1 `M . . . `M
C` `M C 0 . If M is clear from the context, we will often omit the subscript
M.
A configuration that has no successor is called a halting configuration.
A Turing machine M halts on input w iff SCM (w) `∗M Ct for some halting
configuration Ct . (Note again that if it exists, then Ct is unique.) Otherwise
M does not halt on w. If M halts on w and Ct is a halting configuration,
we call a sequence SCM (w) `M C1 `M C2 `M . . . `M Ct a computation of
M on w. If M does not halt on w, then the corresponding computation is
infinite.
Assume that SCM (w) `∗M Ct and Ct = (q, (p1 , x1 ), . . . , (pk , xk )) is a
halting configuration. Let i ≤ p1 be the largest index such that x1 (i) = .
If such an index does not exist, we set i = 0. In the same way, let j ≥ p1
be the smallest index such that x1 (j) = . If such an index does not exist,
then j = |x1 | + 1. Let y = x1 (i + 1)x1 (i + 2) . . . x1 (j − 1). In other words, y
is the word that the head of tape 1 is standing on. y is called the output of
M on w.

c Markus Bläser 2007–2015
74 12. Turing machines

12.3 Functions and languages

A Turing machine M = (Q, Σ, Γ, δ, q0 ) computes a (partial) function ϕM :
Σ∗ → (Γ \ {})∗ defined by
(
the output of M on w if M halts on w,
ϕM (w) =
undefined otherwise.

Definition 12.3 A function f : Σ∗ → Σ∗ is Turing computable, if f = ϕM

for some Turing machine M = (Q, Σ, Γ, δ, q0 ).

We also want to define decidable languages. We could call a language L ⊆

Σ∗ decidable if its characteristic function Σ∗ → {0, 1} is Turing computable.
But this has the problem that 0 or 1 might not be elements of Σ. So we
either have to put 0 and 1 into Σ or we have to identify two symbols of Σ
with 0 and 1. While this works, there is a more elegant way (and this is the
one you usually will find in books, too): A Turing machine is now described
by a 6-tuple (Q, Σ, Γ, δ, q0 , Qacc ).2 Qacc ⊆ Q is called the set of accepting
states. A halting configuration (q, (p1 , x1 ), . . . , (pk , xk )) is called an accepting
configuration if q ∈ Qacc . Otherwise it is called a rejecting configuration.

Definition 12.4 Let L ⊆ Σ∗ be a language.

1. A Turing machine M = (Q, Σ, Γ, δ, q0 , Qacc ) recognizes a language L ⊆

Σ∗ , if for all w ∈ L, the computation of M on w ends in an accepting
configuration and for all w ∈ / L, the computation does not end in an
accepting configuration (i.e., it either ends in a rejecting configuration
or M does not halt on w).

2. M decides L, if in addition, M halts on all w ∈

/ L.

3. If M is a Turing machine then we denote by L(M ) the language rec-

ognized by M .

Thus halting in an accepting configuration means outputting 1 and halt-

ing in a rejecting configuration means outputting 0.

2
So if we want to compute a function, a 5-tuple is sufficient to describe the Turing
machine. If we want to decide or recognize languages, then we take a 6-tuple. We could
also always take a 6-tuple and simply ignore the accepting states if we compute functions.

c Markus Bläser 2007–2015
13 Examples, tricks, and
syntactic sugar

In the beginning, we will do low level descriptions of the Turing machine,

that is, we will write down the transition functions explicitly.
Understanding a Turing machines is like learning to code. In the begin-
ning, even writing a function consisting of ten lines of code is hard. Once
you got experienced, it is sufficient to specify the overall structure of your
program and then filling the functions with code is an easy tasks. The same
is true for Turing machines; though we will not fill in the technical details
after specifying the overall structure since we do not want to sell Turing
machines.

13.1 More Turing machines

Here are some example Turing machines that do some simple tasks. We will
need them later on.
The Turing machine ERASE in Figure 13.1 erases the cells to the right
of the head until the first blank is found.
The machine COPY in Figure 13.2 copies the content of the first tape to
the second tape provided that the second tape is empty. This is done in the
state copy. Once the first blank on the first tape is reached, the copy process
is finished and the Turing machine moves the heads back to the left-most
symbol.
Figure 13.3 shows the Turing machine COMPARE. In the state zero?,
it moves its head to the right until either a 1 or a  is found. In the first
case, the content is not zero. In the second case, the content is zero. In both
states backn and backy, we go back to the left until we find the first blank.
We use two different states for the same thing since we also have to store in

1; R
0; R

S
erase stop

Figure 13.1: The Turing machine ERASE

75
76 13. Examples, tricks, and syntactic sugar

0;0,L|0;0,L
0;0,L|1;1,L
1;1,R| 1,R 1;1,L|0;0,L
0;0,R| 0,R 1;1,L|1;1,L

L| L R| R
copy back stop

Figure 13.2: The Turing machine COPY

1;1,L
0;0,L

R
backn no
0;0,R ,L
;1
1

1;1,L
0;0,L
L

R
backy yes

Figure 13.3: The Turing machine COMPARE

the state whether the content is zero or not. The Turing machine stops in
the state yes or no.

Exercise 13.1 Construct a Turing machine DEC that decreases the content
of the tape by 1 if the content is > 0.

13.2 Some techniques and tricks

13.2.1 Concatenation
Although Turing machine are very unstructured objects per se, it is quite
useful to think about them in a structured way. For instance, assume we
want to test whether the content of the tapes represents some string in
binary that is either zero or one. We first run the Turing machine DEC
of Exercise 13.1. After that the content of the tape has been decreased, it
is zero if and only if it was zero or one before. Thus we can now run the
Turing machine COMPARE. This new “concatenated” Turing machine can
easily be constructed from DEC and COMPARE: First rename the states

c Markus Bläser 2007–2015
13.2. Some techniques and tricks 77

yes
stop
DEC COMPARE
no

Figure 13.4: The concatenated Turing machine. The triangle to the left
of DEC indicates that the starting state of DEC is the starting state of the
new Turing machine. The arrow labeled with stop means that whenever
DEC wants to enter the state stop of DEC, it enters the starting state of
COMPARE instead. The lines labeled with yes and no mean that yes and
no are the two halting states of COMPARE. Note that this concatenation
works well in particular because the Turing machines move the head back to
the first cell of the tape.

yes
stop
DEC COMPARE
no

Figure 13.5: The counting Turing machine. DEC is executed; after this,
the content of the tape is compared with zero. If the state no is entered,
the machine enters again the starting state of DEC and decrease the content
again. If the content of the tape is zero, then the Turing machine stops in
the state yes.

of COMPARE, then make a new Turing machine by throwing the states

of both together. On the states of DEC, the Turing machine behaves like
DEC, with one exception: Whenever DEC wants to enter stop, it enters
the starting state of COMPARE instead. On the states of COMPARE, the
Turing machine behaves like COMPARE. Figures 13.4 shows a schematic
drawing that we will use in the following.

13.2.2 Loops

If you concatenate Turing machines “with themselves”, you get loops. For
instance, if you want to have a counter on some tape that is decreased by
one until zero is reached, we can easily achieve this by concatenating the
machines DEC and COMPARE as depicted in Figure 13.5.

c Markus Bläser 2007–2015
78 13. Examples, tricks, and syntactic sugar

13.2.3 Marking of cells

All the machines INC, COPY, COMPARE go back to the left end of the
string on the tape. They can find this end because it is marked by a .
What if a Turing machine wants to find a position somewhere in the middle
of the string. In this case, it can “leave a marker” there. For this, we enlarge
the tape alphabet and add for each γ ∈ Γ a new symbol γ̄. The Turing
machine can replace the current symbol γ by the symbol γ̄ and move its
head somewhere else. It can find this position by going back and scanning
for a symbol that is not in the original alphabet Γ. It then replaces this
symbol γ̄ by γ and continues its computation. (Leaving more than one
marker in this way per tape could mean trouble!)

13.2.4 Storing information in the state

Have another look at the Turing machine COMPARE. After it reached the
first 1 or the first  it has to go back to the beginning of the string. In
both cases, COMPARE has to do the same thing: going back! But it has
to remember whether it found a 1 or a . Therefore, we need two states for
going back, backn and backy. One usually says that “the Turing machine
goes back and stores in its state whether it found a 1 or a ”. If there is
more to store, say an element from a set S, then it is more convenient to
take the cartesian product {q} × S as states. In our example, we could e.g.
also have used the elements from {back} × {1, } as states.

13.2.5 Parallel execution

Let M = (Q, Σ, Γ, δ, q0 ) and M 0 = (Q0 , Σ, Γ, δ 0 , q00 ) be two Turing machines
with k and k 0 tapes. We can construct a Turing machine N with k + k 0 tapes
which simulates M and M 0 in parallel as follows: N has states Q × Q0 and
starting state (q0 , q00 ). The transition function of N ,
0 0 0
∆ : (Q × Q0 ) × Γk+k → (Q × Q0 ) × Γk+k × {L, S, R}k+k ,

is defined by

∆((q, q 0 ), γ1 , . . . , γk+k0 ) = ((p, p0 ), α1 , . . . , αk , α10 , . . . , αk0 0 , r1 , . . . , rk , r10 , . . . , rk0 0 )

if

δ(q, γ1 , . . . , γk ) = (p, α1 , . . . , αk , r1 , . . . , rk ) and

0 0 0 0 0 0 0 0
δ(q , γk+1 , . . . , γk+k 0 ) = (p , α1 , . . . , αk 0 , r1 , . . . , rk 0 )

are both defined. If only one is defined, say δ(q, γ1 , . . . , γk ), then

∆((q, q 0 ), γ1 , . . . , γk+k0 ) = ((p, q 0 ), α1 , . . . , αk , γk+1 , . . . , γk+k0 , r1 , . . . , rk , S, . . . , S)

c Markus Bläser 2007–2015
13.3. Syntactic sugar 79

The other case is defined symmetrically. If both are undefined, then ∆((q, q 0 ), γ1 , . . . , γk+k0 )
is undefined, too. On the first k tapes, N behaves like M , on the other k 0
tapes, N behaves like M 0 . If one machine stops, then N does not move its
head on the corresponding tapes anymore and just writes the symbols that
it reads all the time. If the second machine stops, too, then N stops. (Of
course, since N gets its input on the first tape, M 0 is simulated on the empty
tape. If we want to simulate M and M 0 on the same input, then we have to
copy the input from the first tape to the (k + 1)th tape, using for instance
COPY, before N starts with the simulation).
Here is one application: take M to be any machine and M 0 is the machine
from Figure 13.5. Modify the function ∆ such that M 0 is simulated normally,
but M only executes a step when M 0 changes it state from the state no to its
start state. In this way, M executes as many steps as given by the counter
in the beginning. We will need this construction later on.

13.3 Syntactic sugar

For more complex Turing machines, describing them by transition diagrams
is a boring task. So after some training, we will go on by describing Turing
machines by sentences in natural language. Whenever you formulate such a
sentence, you should carefully think how a Turing machine actually would
do the thing that you are describing. Here is a description of the machine
INC:

Input: x ∈ {0, 1}∗ , viewed as a number in binary.

Output: x increased by 1

1. Go to the right and replace every 1 by a 0 until you reach the

first 0 or .

2. Write a 1 and go back to the right to the beginning of the string.

Once we got even more experienced with Turing machines, we could even
write:

Input: x ∈ {0, 1}∗ , viewed as a number in binary.

Output: x increased by 1

1. Increase the content of the tape by 1.

c Markus Bläser 2007–2015
80 13. Examples, tricks, and syntactic sugar

Warning!!! Although the example above suggests it, the sentence

“The Turing machine produces the desired ouput” is in general not an ad-
equate description of a Turing machine. Above, it is, because the Turing
machine INC does such a simple job. More complex jobs, require more de-
tailed descriptions.

If in doubt . . .

. . . whether your description of a Turing machine is o.k. always ask

yourself the following question: Given the description, can I write a
C program that gets k char* as input and does the same.
(JAVA is also fine.)

13.4 Further exercises

Exercise 13.2 Instead of a two-sided infinite tape, you can also find Turing
machines with a one-sided infinite tape in the literature. Such a tape can be
modelled by a function T : N → Γ. There is a distinguished symbol & that
marks the end of each tape. Initially, every tape is filled with blanks except
the 0th cell, which is filled with a &. The first tape contains the input x in
the cells 1, 2, . . . , |x|. Every Turing machine with one-sided infinite tapes has
to obey the following rules: If it does not read a &, it cannot write & on this
tape. If it reads a &, it has to write a & on this tape and must not move its
head to the left on this tape. In this way, it can never leave the tape to the
left.
Show that every Turing machine with two-sided infinite tapes can be sim-
ulated by a Turing machine with one-sided infinite tapes. Try not to increase
the number of tapes!

c Markus Bläser 2007–2015
14 Church–Turing thesis

14.1 WHILE versus Turing computability

In this chapter, we want to show that
WHILE computable
equals
Turing computable.
But there is of course a problem. WHILE programs compute functions
Ns → N whereas Turing machines compute functions Σ∗ → Σ∗ . To make
things a little easier, we can restrict ourselves to functions N → N, since we
can use a pairing function. For Turing machines, we use the input alphabet
Σ = {0, 1}.

14.1.1 N versus {0, 1}∗

We have to identify natural numbers with words over {0, 1} and vice versa.
For y ∈ N, let bin(y) ∈ {0, 1}∗ denote the binary expansion of y without
any leading zeros. (In particular, 0 ∈ N is represented by the empty word ε
and not by 0.) The function bin : N → {0, 1}∗ is an injective mapping—two
different numbers have different binary expansions—but it is not surjective,
since we do not cover strings with leading zeros. Not bad, but we want
to have a bijection between N and {0, 1}∗ . Consider the following mapping
{0, 1}∗ → N: Append a 1 to a given x ∈ {0, 1}∗ . This is an injective mapping
from {0, 1}∗ to the subset of all binary expansions without leading zeros of
some natural number. Since we do not have leading zeros, the function that
maps such a binary expansion to the corresponding natural number is also
injective. The combination of both gives an injective mapping {0, 1}∗ → N.
It is also surjective? No, the smallest number that we get in the image is 1,
by appending 1 to the empty word ε. So here is the next attempt:
1. Append a 1 to the word x ∈ {0, 1}∗ .
2. View this string 1x as some binary expansion. Let n be the corre-
sponding number, i.e, bin(n) = 1x.
3. Subtract 1 from n.
We call the mapping {0, 1}∗ → N that we get this way cod. More compactly,
we can write cod(x) = bin−1 (1x) − 1. (Note that we can write bin−1 , since
1x is in the image of bin.)

81
82 14. Church–Turing thesis

Exercise 14.1 Show that cod is indeed bijective.

14.1.2 N → N versus {0, 1}∗ → {0, 1}∗

Now, if f : N → N is a function, then fˆ : {0, 1}∗ → {0, 1}∗ is defined as
follows:
fˆ(x) = cod−1 (f (cod(x))) for all x ∈ {0, 1}∗ .
Or, in other words, the following diagram commutes:

f
N → N
−1
↓ cod ↓ cod−1
fˆ
{0, 1}∗ → {0, 1}∗

Conversely, if g : {0, 1}∗ → {0, 1}∗ , then ĝ : N → N is defined by

ĝ(n) = cod(g(cod−1 (n)) for all n ∈ N.

In other words,
g
{0, 1}∗ → {0, 1}∗
↓ cod ↓ cod
ĝ
N → N

Exercise 14.2 Show the following: For every f : N → N and g : {0, 1}∗ →
ˆ
{0, 1}∗ , fˆ = f and ĝˆ = g. (“.̂ is self-inverse.”)

Remark 14.1 The mapping cod is not too natural. For instance, with f :
N → N, we could associate the mapping bin(x) 7→ bin(f (x)). But if cod is a
bijection, then we have the nice property that .̂ is self-inverse.

Exercise 14.3 Show that cod and cod−1 are functions that are easy to com-
pute. In particular:

1. Write a WHILE program that, given an n ∈ N, computes the symbols

of cod−1 (n) and stores them in an array.

2. Construct1 a Turing machine that given x ∈ {0, 1}∗ , writes cod(x)

many 1’s on the first tape.

So the only reason why a WHILE program or Turing machine cannot compute
cod or cod−1 is that they cannot directly store elements from {0, 1}∗ or N,
respectively.
1
This is quite funny. While both WHILE programs and Turing machines are mathe-
matical objects, we write WHILE programs but construct Turing machines.

c Markus Bläser 2007–2015
14.2. GOTO programs 83

14.1.3 Pairing functions

We also need a pairing function for {0, 1}∗ , i.e., an injective mapping [., .] :
{0, 1}∗ × {0, 1}∗ → {0, 1}∗ . If we take our pairing function h., .i : N × N → N,
then of course,
(x, y) 7→ cod−1 (hcod(x), cod(y)i)
is a bijection {0, 1}∗ × {0, 1}∗ → {0, 1}∗ .
But there are nicer ways of defining a pairing function for {0, 1}∗ . The
easiest way would be to concatenate the strings x and y. But then we do
not know where x ends and y starts. We could use some special symbol #
and write it between x and y, but then we would have to enlarge the input
alphabet (which is not a tragedy, but we do not want to do it here). There
is another way: Let β(x) = x1 0x2 0 . . . x|x| 0, i.e, we insert a 0 after every
symbol of x. From β(x)11y, we can get x and y back, since the first pair
11 marks the end of β(x). The length of this encoding is 2|x| + |y| + 2. We
can get a shorter one by mapping (x, y) to β(bin(|x|))11xy. By scanning for
the first pair 11, we can divide the string into β(bin(|x|)) and xy. From the
first, we can reconstruct |x|. Once we know this, we can get x and y from x.
The length of this encoding is ≤ |x| + |y| + 2 log |x| + 2.

Exercise 14.4 Try to get even shorter encodings in this way. What is the
shortest that you can get?

14.2 GOTO programs

It turns out that it is useful to introduce some intermediate concept, GOTO
programs. GOTO programs have variables and the same simple statements
as WHILE programs but instead of a while loop, there is a goto statement.
Furthermore, all the lines are numbered.
Formally, a GOTO program is a sequence (1, s1 ), (2, s2 ), . . . , (m, sm ) where
each sµ is a statement of the form

1. xi = xj + xk or

2. xi = xj − xk or

3. xi := c or

4. if xi 6= 0 then goto λ

The semantics of the first three statements is the same as for WHILE
programs. After the µth statement is executed, the program goes on with
the (µ+1)th statement. The only execption is the conditional jump if xi 6= 0
then goto λ. If the content of xi is zero, then we go on with the (µ + 1)th
statement, otherwise, we go on with statement λ. If we ever reach a line that

c Markus Bläser 2007–2015
84 14. Church–Turing thesis

does not exist—either by jumping to a nonexisting line or by executing the

last statement sm and going to line m + 1—the program stops. The content
of x0 is the value computed by the program. As for WHILE programs, the
input is stored in the first s variables. The function Ns → N computed by a
GOTO program P is denoted by—surprise—ϕP .

Exercise 14.5 Show how to simulate an unconditional jump in GOTO.

(An unconditional jump, we denote it by goto λ, always jumps to line λ no
matter what.)

Exercise 14.6 Give a precise mathematical formulation of the semantics of

GOTO programs. A state should consist of a natural number, which stores the
current line to be executed, and a tuple/sequence with finite support of natural
numbers, which stores the content of the variables. Construct a function ΦP
that maps a state (i, V ) to the state that is reached after executing line i.

Every while loop can be simulated by a goto statement. If should be

fairly obvious that
while xi 6= 0 do
P
od
is simulated by
1: if xi 6= 0 then goto 3
2: goto 5
3: P
4: goto 1
5: . . .
The use of the labels is a little bit sloppy. The program P in general has more
than one line, so the label of the statement goto 1 is usually larger. Further-
more, we do not write tuples but lines and separate labels and statements
by “:”. We get the following theorem.

Lemma 14.2 For every WHILE program P there is a GOTO program Q

with ϕP = ϕQ .

14.3 Turing machines can simulate GOTO programs

Lemma 14.3 Let f : N → N. If f is GOTO computable, then fˆ is Turing
computable.

Proof. Assume that f is GOTO computable. Let P = (1, s1 ), . . . , (m, sm )

be a GOTO program computing f . It is fairly easy to see we can restrict
the simple statements of GOTO programs to xi ++, xi −−, and xi := 0.

c Markus Bläser 2007–2015
14.3. Turing machines can simulate GOTO programs 85

Assume that P uses variables x0 , . . . , x` . Our Turing machine uses ` + 1

tapes. Each tape stores the content of one of the registers in binary. The
input for the Turing machine is cod−1 (n). For the simulation, it is easier to
have bin(n) on the tape. We get it by appending a 1 to cod−1 (n) and then
subtracting 1 by using the Turing machine DEC.
We will use the Turing machines INC, DEC, ERASE, and COMPARE
(see Section 13) to construct a Turing machine M that simulates P , more
precisely, computes fˆ.
For each instruction (µ, sµ ), we have a state qµ . The invariant of the
simulation will be that whenever the Turing machine is in one of these states,
the content of the tapes correspond to the content of the registers before
executing the instruction sµ and all heads stand on the left-most symbol
that is not a blank. (The lowest order bit is standing on the left.)
Figure 14.1 shows an example of the construction for the program
1: if x0 6= 0 then goto 3
2: x0 ++
3: . . .
The arrow from the state q0 to the box with the label COMPARE means
that in q0 , M does nothing (i.e., writes the symbol that it reads and does not
move its head) and enters the starting state of the machine COMPARE. The
two arrows leaving this box with the labels yes and no mean that from the
states yes and no of COMPARE, we go to the states q3 and q2 ,respectively.
The Turing machine COMPARE is only a 1-tape Turing machine. It can be
easily extended it to an (` + 1)-tape machine that only works on the tape
corresponding to x0 . The same has to be done for the machine INC and
so on. From the example it should be clear how the general construction
works. For each instruction sµ , M goes from qµ to a copy of one of the
Turing machines that simulates the instruction xi ++, xi −−, xi := 0 or if
xi 6= 0 then goto λ. From the halting state(s) of these machines, M then
goes to qµ+1 , the only exception being the conditional jump.
It should be clear from the construction that the simulation is correct. To
formally prove the correctness, it is sufficient to show the following statement:
Claim. Assume that P is in state (µ, V ) and that ΦP (µ, V ) = (µ0 , V 0 ). If
M is in state qµ , the content of the tapes are bin(V (λ)), 0 ≤ λ ≤ `, and
the heads are standing on the lowest order bits of bin(V (λ)), then the next
state from q1 , . . . , qm that M will enter will be qµ0 . At this point, the content
of the tapes are bin(V 0 (λ)), 1 ≤ λ ≤ `, and the heads are standing on the
lowest order bits of bin(V (λ)).
From this claim, the correctness of the simulation follows immediately.
When M stops, we have to transform the binary expansion b = bin(n) on
tape 1 back into cod−1 (n), which is easy.

Exercise 14.7 Give a detailed description of the general construction.

c Markus Bläser 2007–2015
86 14. Church–Turing thesis

no
q1 COMPARE yes

stop
q2 INC

q3

Figure 14.1: The simulating Turing machine for the example GOTO pro-
gram.

Exercise 14.8 Prove the claim in the proof of Lemma 14.3

14.4 WHILE programs can simulate Turing machines

Lemma 14.4 Let g : {0, 1}∗ → {0, 1}∗ . If g is Turing computable, then ĝ
is WHILE computable.

Proof. Let M = (Q, {0, 1}, Γ, δ, q0 ) be a k-tape Turing machine that

computes g. By renaming states, we can assume that Q = {1, 2, . . . , q}. We
represent the symbols of Γ by the numbers 0, 1, . . . , s − 1. By Exercise 13.2,
we can assume that each tape is onesided infinite.
The content of each tape κ is stored in an array Aκ . Aκ [i] = j means
that the cell i contains the symbol that corresponds to j. Of course, Aκ
always stores only a finite amount of data. Recall that the arrays that we
created in WHILE are dynamic, so we can extend them whenever M visits
a new cell. The variable pκ contains the (absolute) position of the head.
In the beginning, we have to write cod−1 (x0 ) into A1 . We can do this
using the Turing machine constructed in Exercise 14.3.
The domain of δ is finite, thus we can hardwire the table of δ into our
WHILE program. Then it is easy to simulate one step of M , we just have to
update the corresponding cells of the array and adjust p1 , . . . , pk and change
the state. The simulation of one such step is embedded into a while loop of
the form

while δ(q, a1 , . . . , ak ) is defined do . . . od

c Markus Bläser 2007–2015
14.5. Church–Turing thesis 87

From the construction, it is clear that the simulation is correct.

Corollary 14.5 (Kleene normal form) For every WHILE computable func-
tion f , there are FOR programs P1 , P2 , P3 and a WHILE program of the form

P1 ; while xi 6= 0 do P2 od; P3

that computes f .

Proof. We convert a WHILE program into an equivalent GOTO program,

then into an equivalent Turing machine, and finally back into an equivalent
WHILE program. This WHILE program has only one while loop. All the
other things, like simulating an array etc. can be done by FOR programs.

14.5 Church–Turing thesis

From Lemmas 14.2, 14.3, and 14.4, we get the following result.

Theorem 14.6 Let f : N → N. Then the following three statements are

equivalent:

1. f is WHILE computable.

2. f is GOTO computable.

3. fˆ is Turing computable.

The Church–Turing thesis states that any notion of “intuitively com-

putable” is equivalent to Turing computable (or WHILE computable, . . . ).
The theorem above is one justification of the Church–Turing thesis. The
Church–Turing thesis is not a statement that you could prove. It is a state-
ment about the physical world we are living in. You can either accept the
Church–Turing thesis or reject it. So far, the Church–Turing thesis seems to
hold and it is widely accepted among computer scientists. Even a quantum
computer would not change this. Quantum computers cannot compute more
than Turing machines. Maybe they can do it faster, but this is another story.
But who knows, maybe some day a brilliant (crazy?) physicist will come up
with a device that decides the halting problem.

c Markus Bläser 2007–2015
B Primitive and µ-recursion

Historically, primitive recursive functions and µ-recursive functions are one

of the first concepts to capture “computability”.

B.1 Primitive recursion

We are considering functions Ns → N for any s ≥ 1.

Definition B.1 The set of all primitive recursive functions is defined in-
ductively as follows:
1. Every constant function is primitive recursive.
2. Every projection psi : Ns → N (mapping (a1 , . . . , as ) to ai ) is primitive
recursive.
3. The successor function suc : N → N defined by suc(n) = n + 1 is
primitive recursive.
4. If f : Ns → N and gi : Nt → N, 1 ≤ i ≤ s, are primitive recursive, then
their composition defined by
(a1 , . . . , at ) 7→ f (g1 (a1 , . . . , at ), . . . , gs (a1 , . . . , at ))
is primitive recursive.
5. If g : Ns → N and h : Ns+2 → N are primitive recursive, then the
function f : Ns+1 → N defined by
f (0, a1 , . . . , as ) = g(a1 , . . . , as )
f (n + 1, a1 , . . . , as ) = h(f (n, a1 , . . . , as ), n, a1 , . . . , as )
is primitive recursive. This scheme is called primitive recursion.

We want to show that primitive recursive functions are the same as FOR
computable functions. Therefore, we first look at some fundamental func-
tions that appear in FOR programs:

Example B.2 The function add(x, y) = x + y is primitive recursive. We

have
add(0, y) = y,
add(x + 1, y) = suc(add(x, y)).

88
B.1. Primitive recursion 89

Above, we did not write down the projections explicitly. The correct definition
looks like this:
add(0, y) = p11 (y),
add(x + 1, y) = suc(p31 (add(x, y), x, y)).
Since this looks somewhat confusing, we will omit the projections in the fol-
lowing.

Exercise B.1 Show that the function mult(x, y) = xy is primitive recursive.

Example B.3 The predecessor function defined by

(
n − 1 if n > 0
pred(n) =
0 otherwise
is primitive recursive. We have
pred(0) = 0,
pred(n + 1) = n.
This is a primitive recursion scheme.

Exercise B.2 Prove that the modified difference sub(x, y) = max{x − y, 0}

is primitive recursive.

B.1.1 Bounded maximization

Finally, we will show that the pairing function h., .i and “its inverses” π1 and
π2 are primitive recursive. Recall that hx, yi = 12 (x + y)(x + y + 1) + y. From
1 1
n(n + 1) = (n − 1)n + n
2 2
we get a primitive recursion scheme for 12 n(n + 1) and from this function, we
can easily get hx, yi.
For the inverse functions, we need bounded maximization: Let P be a
predicate on N and view P as a function P : N → {0, 1}. Assume that P (as
a function) is primitive recursive. We claim that
bounded-max -P (n) := max{x ≤ n | P (x) = 1}
is primitive recursive. (If no such x exists, i.e., the maximum is undefined,
we set bounded-max -P (n) = 0.) This can be seen as follows:
bounded-max -P (0) = 0
(
n+1 if P (n + 1) = 1
bounded-max -P (n + 1) =
bounded-max -P (n) otherwise
= (1 − P (n + 1)) · bounded-max -P (n) + P (n + 1) · (n + 1).

c Markus Bläser 2007–2015
90 B. Primitive and µ-recursion

In the same way, we can see that the bounded existential quantifier defined
by (
1 if there is an x ≤ n with P (x) = 1
bounded-∃-P (n) :=
0 otherwise
is primitive recursive:

bounded-∃-P (0) = P (0)

bounded-∃-P (n + 1) = P (n + 1) + bounded-∃-P (n)(1 − P (n + 1))

Above, P has only one argument. It is easy to see that for a predicate
with s arguments,

bounded-maxi -P (x1 , . . . , xs ) := max{x ≤ xi | P (x1 , . . . , xi−1 , x, xi+1 , . . . , xs ) = 1}

is primitive recursive. In the same way, we can define bounded-∃i -P and

show that it is primitive recursive.
With these quantifiers, we can easily invert the pairing function. Let
Q(x, y, z) be the predicate “hx, yi = z”. It is not too hard to see that this
predicate is primitive recursive. Now we can write:

π1 (z) = bounded-max1 -Q(x, bounded-∃2 -Q(x, y, z), z)

π2 (z) = bounded-max2 -Q(bounded-∃1 -Q(x, y, z), y, z)

It immediately follows that also forming larger “pairs” ha1 , . . . , as i and

the corresponding inverse functions π1 , . . . , πs are primitive recursive.

B.1.2 Simultaneous primitive recursion

Let gi : Ns → N and hi : Ns+t+1 → N, 1 ≤ i ≤ t, be primitive recursive. The
functions fi : Ns+1 → N, 1 ≤ i ≤ t, defined by the simultaneous primitive
recursion scheme

fi (0, a1 , . . . , as ) = gi (a1 , . . . , as ),
fi (n + 1, a1 , . . . , as ) = hi (f1 (n, a1 , . . . , as ), . . . , ft (n, a1 , . . . , as ), n, a1 , . . . , as ),

for i = 1, . . . , t, are primitive recursive. To see this, define f by

f (0, a) = hg1 (π1 (a), . . . , πs (a)), . . . , gt (π1 (a), . . . , πs (a))i

f (n + 1, a) = hh1 (π1 (f (n, π1 (a), . . . , πs (a))), . . . ,
πt (f (n, π1 (a), . . . , πs (a))), n, π1 (a), . . . , πs (a)), . . .
ht (π1 (f (n, π1 (a), . . . , πs (a))), . . . ,
πt (f (n, π1 (a), . . . , πs (a))), n, π1 (a), . . . , πs (a))i.

f is primitive recursive. By an easy induction on n, we can show that

πi (f (n, a)) = fi (n, π1 (a), . . . , πs (a)).

c Markus Bläser 2007–2015
B.1. Primitive recursion 91

We can rewrite this as

fi (n, a1 , . . . , as ) = πi (f (n, ha1 , . . . , as i)).

Thus each fi is primitive recursive.

B.1.3 Primitive recursion versus for loops

The next lemma shows that for all FOR programs, there are primitive re-
cursive functions that compute the values of the variables after executing
P.

Lemma B.4 Let P be a FOR program with s inputs. Let ` be the largest in-
dex of a variable in P . Then there are primitive recursive functions v0 , . . . , v` :
N`+1 → N such that

(v0 (a0 , . . . , a` ), . . . , v` (a0 , . . . , a` )) = ΦP (a0 , . . . , a` )

for all a0 , . . . , a` ∈ N.

Proof. The proof is by structural induction.

Induction base: If P is xi := xj + xk then each vλ is the projection on the
λth component, except for vi , which is xj + xk . Since modified subtraction
and constant functions are primitive recursive, too, we can cover the cases
xi := xj − xk and xi := c in the same way.
Induction step: If P = P1 ; P2 , then by the induction hypothesis, there are
primitive recursive functions vi,0 , . . . , vi,` , i = 1, 2, such that

(vi,0 (a0 , . . . , a` ), . . . , vi,` (a0 , . . . , a` )) = ΦPi (a0 , . . . , a` ), i = 1, 2.

Since ΦP = ΦP2 ◦ ΦP1 , we get that

vλ (a0 , . . . , a` ) = v2,λ (v1,0 (a0 , . . . , a` ), . . . , v1,` (a0 , . . . , a` ))

for all 0 ≤ λ ≤ ` and a0 , . . . , a` ∈ N. Thus v0 , . . . , v` are primitive recursive.

If P = for xi do P1 od, then by the induction hypothesis, there are
functions v1,0 , . . . , v1,` such that

(v1,0 (a0 , . . . , a` ), . . . , v1,` (a0 , . . . , a` )) = ΦP1 (a0 , . . . , a` ). (B.1)

Define u0 , . . . , u` by

uλ (0, a0 , . . . , a` ) = aλ
uλ (n + 1, a0 , . . . , a` ) = v1,λ (u0 (n, a0 , . . . , a` ), . . . , u` (n, a0 , . . . , a` ))

for 0 ≤ λ ≤ `. This is a simultaneous primitive recursion scheme. We claim

that
(n)
(u0 (n, a0 , . . . , a` ), . . . , u` (n, a0 , . . . , a` )) = ΦP1 (a0 , . . . , a` ) (B.2)

c Markus Bläser 2007–2015
92 B. Primitive and µ-recursion

for all n, a0 , . . . , a` ∈ N. This claim is shown by induction on n.

Induction base: The case n = 0 is clear, since both sides are (a0 , . . . , a` ) in
this case.
Induction step: We have

uλ (n + 1, a0 , . . . , a` ) = v1,λ (u0 (n, a0 , . . . , a` ), . . . , u` (n, a1 , . . . , a` ))

(n)
= v1,λ (ΦP1 (a0 , . . . , a` ))
(n+1)
= λth entry of ΦP1 (a0 , . . . , a` ).

The last equality follows from the induction hypothesis (B.1).

Altogether, this shows (B.2). We get vλ by vλ (a0 , . . . , an ) = uλ (ai , a0 , . . . , an ).

Lemma B.5 For every primitive recursive function f , there is a FOR pro-
gram P with ϕP = f .

Proof. This proof is again by structural induction.

Induction base: Constant functions, projections, and the successor function
are all FOR computable.
Induction step: If f is the composition of h and g1 , . . . , gs , then by the induc-
tion hypothesis, there are FOR programs P and Q1 , . . . , Qs that compute h
and g1 , . . . , gs . From this, we easily get a program that computes f .
If f is defined by

f (0, a1 , . . . , as ) = g(a1 , . . . , as ),
f (n + 1, a1 , . . . , as ) = h(f (n, a1 , . . . , as ), n, a1 , . . . , as ),

then there are programs P and Q that compute h and g, respectively. Now
the following program computes f (a0 , a1 , . . . , as ):
1: x0 := g(a1 , . . . , as );
2: for a0 do
3: x0 := h(x0 , a0 , a1 , . . . , as )
4: od

We saw how to simulate function calls of FOR computable functions.

Theorem B.6 A function f is primitive recursive iff it is FOR computable.

Proof. The “⇐”-direction is Lemma B.5. The function v0 of Lemma B.4

is the function that is computed by the program P . This show the other
direction.

c Markus Bläser 2007–2015
B.2. µ-recursion 93

B.2 µ-recursion
The µ-operator allows unbounded search.

Definition B.7 Let f : Ns+1 → N. The function µf : Ns → N is defined by

µf (a1 , . . . , as ) = min{n | f (n, a1 , . . . , as ) = 0 and

for all m < n, f (m, a1 , . . . , as ) is defined }.

Definition B.8 The set of all µ-recursive functions is defined inductively

as in Defition B.1 except that the set is closed under µ-recursion instead of
primitive recursion.

Theorem B.9 A function is µ-recursive iff it is WHILE computable.

Proof. This proof is just an “add-on” to the proof of Theorem B.6.

For the “⇒”-direction, we just have to consider one more case in the
proof of Lemma B.5. If f = µg for some µ-recursive function g : Ns+1 → N,
then we have to show that f is WHILE computable provided that g is. The
following program computes f :
1: n := 0;
2: while f (n, x0 , . . . , xs−1 ) 6= 0 do
3: n := n + 1
4: od
5: x0 := n
Thus program finds the first n such that f (n, x0 , . . . , xs−1 ) = 0. If no such
n exists, the while loop does not terminate. If one f (n, x0 , . . . , xs−1 ) is
undefined, the program does not terminate, too.
For the other direction, assume that we have a WHILE program P =
while xi 6= 0 do P1 od. The functions uλ (n, a0 , . . . , a` ) constructed in the
proof of Lemma B.4 is the content of the variable xλ after executing P1 n
times. Thus
µuλ (a0 , . . . , a` )
is the number of times the while loop is executed, and

uλ (µuλ (a0 , . . . , a` ), a0 , . . . , a` )

is the content of xλ after executing P .

Excursus: Programming systems III

We can also assign Gödel numbers to recursions schemes. We start by as-

signed numbers to the constant functions, projections, and successor function. For
instance, h0, hs, cii could stand for the function of arity s that has the value c

c Markus Bläser 2007–2015
94 B. Primitive and µ-recursion

everywhere, h1, hs, iii encodes psi and so on. Then we define Gödel numbers for
composition and primitive and µ-recursion. If we have a functions f or arity s
and s functions gi of arity t and i and j1 , . . . , js are their Gödel numbers, then
h3, t, s, i, j1 , . . . , js i is the Gödel number for their composition.
Let θi be the function that is computed by the recursion scheme with Gödel
number i. If i is not a valid Gödel number, then θi is some dummy function, for
instance, the function that is undefined everywhere. Then the sequence (θi )i∈N is
a programming system.
Once we have constructed a universal Turing machine and/or WHILE program,
we will see that it is also universal. It is clearly acceptable, since composition is
directly available in recursion schemes.

c Markus Bläser 2007–2015
15 A universal Turing machine

15.1 Gödel numberings for Turing machines

We want to find a mapping that maps Turing machines to strings in {0, 1}∗ .
Let M = (Q, {0, 1}, Γ, δ, q0 , Qacc ) be a k-tape Turing machine. We can as-
sume that Q = {0, 1, . . . , s} and Γ = {0, 1, . . . , `}. We assume that ` is the
blank.
We can encode a state q by bin(q) and a symbol γ by bin(γ). The fact
that
δ(q, γ1 , . . . , γk ) = (q 0 , γ10 , . . . , γk0 , r1 , . . . , rk )
can be encoded by

[bin(q), bin(γ1 ), . . . , bin(γk ), bin(q 0 ), bin(γ10 ), . . . , bin(γk0 ), r̂1 , . . . , r̂k ]

where 
00 if rκ = S

r̂κ = 10 if rκ = L

01 if rκ = R


[., .] denotes (one of) the pairing functions discussed in Section 14.1.3. It is
extended to larger tuples as expected:

[a1 , . . . , am ] := [a1 , [a2 , . . . , [am−1 , am ]]].

If δ(q, γ1 , . . . , γk ) is undefined, then we encode this by

[bin(q), bin(γ1 ), . . . , bin(γk ), bin(s + 1), ε, . . . , ε, ε, . . . , ε]

The second part of the tuple is a dummy value, the non-existing state s + 1
is used for saying that the value is undefined.
We construct a mapping gödTM from the set of all Turing machines to
{0, 1}∗ by building a large pair consisting of:

• bin(k), the number of tapes,

• bin(s + 1), the size of Q,

• bin(` + 1), the size of Γ,

• the encodings of δ(q, γ1 , . . . , γk ), q ∈ Q, and γ1 , . . . , γk ∈ Γ, in lexico-

graphic order,

95
96 15. A universal Turing machine

• bin(q0 ), the start state,

• bin(|Qacc |), the number of accepting states,

• bin(q), q ∈ Qacc , in ascending order.

If M is supposed to compute a function (instead of recognizing some language

L ⊆ {0, 1}∗ ), then we indicate this by not giving any accepting states. It is
clear that gödTM is an injective mapping.

Definition 15.1 A mapping g from the set of all Turing machines over the
input alphabet {0, 1} to {0, 1}∗ is called a Gödel numbering if

1. g is injective,

2. there is a Turing machine C that computes the characteristic function

of im g, and

3. there is a Turing machine U that given i ∈ im g and x ∈ {0, 1}∗ com-

putes ϕM (x) where M = g −1 (i).

Constructing the Turing machine CTM that computes the characteristic

function of im gödTM is rather easy. A i ∈ im gödTM is just a concatenation
of some tupels, we just have to check whether they have the right form.

15.2 A universal Turing machine

Finally, we construct the universal Turing machine UTM . UTM simulates a
given Turing machine M step by step. The biggest problem for constructing
a universal Turing machine UTM is that is has a fixed number of tapes, a
fixed working alphabet, and a fixed number of states. To simulate M , we
encode the symbols of the work alphabet of M in binary. We store all k
tapes of M on one tape of UTM , the second one, say. To do so, we build
blocks. If in some step, the ith cells of the k tapes of M contain the symbols
i1 , . . . , ik , then the ith block is

# bin(i1 )# bin(i2 )# . . . # bin(ik ).

To the right of this block, there is the block corresponding to the (i + 1)th
cells of M , to left the one corresponding to the (i − 1)th cells of M . Between
two such blocks, UTM writes $ as a separator. So the k tapes of M are
“interleaved”.
Of course, UTM has to bring its second tape into this form. In the
beginning, it initializes its second tape by writing the blocks

# bin(xj )# bin(`)# . . . # bin(`)

c Markus Bläser 2007–2015
15.2. A universal Turing machine 97

Input: i ∈ {0, 1}∗

Output: accept, if i ∈ im gödTM , reject otherwise

1. Extract the values k, s, and r from i.

2. From these values, CTM can compute the size and the number
of tuples encoding the transition function.

3. Test whether all these tupels are in ascending order, whether

they have the correct number of entries, and whether the entries
are all in the given bounds.

4. Finally check whether q0 and whether the accepting states are

between 0 and s and whether they are in ascending order and
every state is only listed once.

5. If one of these conditions is violated, then reject. Otherwise

accept.

Figure 15.1: The Turing machine CTM .

for j = 1, . . . , n on it where x = x1 x2 . . . xn denotes the input for M . Recall

the ` is the blank of M . Whenever M enters a cell that has not been visited
before on any tape, then UTM will enter a cell that contains a blank (of
UTM ). UTM then first creates a new block that consists solely of blanks (of
M ).
UTM has only one head on its second tape. But M has k heads on its k
tapes. UTM remembers that the head of the jth tape of M is standing on
the ith cell by replacing the jth # of the ith block by a ∗.
One can prove the correctness of UTM by induction on the number of
steps of M . This is not hard, but quite some slave work. We get the
following result.

Theorem 15.2 There is a Turing machine UTM that, given a pair [g, x]
with g ∈ im gödTM and x ∈ {0, 1}∗ , computes ϕgöd−1 (g) (x).
TM

Exercise 15.1 Show that the constructed Turing machine UTM is correct.

c Markus Bläser 2007–2015
98 15. A universal Turing machine

Input: [g, x] with g ∈ im gödTM and x ∈ {0, 1}∗

Output: ϕgöd−1 (g) (x)
TM

1. UTM copies the start state to the third tape.

2. UTM copies the input x from the first tape to the second tape
as described above.
It replaces all # of the first block by ∗.

3. UTM moves the head to the first symbol of the leftmost block.

4. While the transition function is not undefined

(a) UTM goes to the right.

(b) Whenever it finds a ∗, it copies the following number in
binary to the fourth tape.
(c) If UTM reaches the right end of tape 2, then it looks up
the tuple of δ that corresponds to the current state (on
tape 3) and the symbols copied to tape 4.
(d) UTM replaces the state on tape 3 by the new state.
(e) UTM goes to the left
(f) Whenever UTM finds a ∗, it updates the corresponding
cells and moves the ∗ to its new position.

5. If M := göd−1
TM (g) is supposed to compute a function, then
UTM copies the content of tape 2 that corresponds to the first
tape of M back to tape 1 and stops.

6. If M is supposed to decide a language, then M accepts if the

current state on tape 3 is in the list of accepting states of M ,
otherwise it rejects.

Figure 15.2: The universal Turing machine UTM

c Markus Bläser 2007–2015
C Kolmogorov Complexity

Kolmogorov complexity measures the “information” stored in a string x ∈

{0, 1}∗ . While the string 0n intuitively has low information content, since it
can be described by “a sequence of n zeros” and we need log n bits to write
n down, the string 01100010010001111010110100110101 does not seem to
have a short description. Komogorov complexity, like the halting problem,
is another natural problem that is not Turing computable.

C.1 Definition
Definition C.1 The Kolmororov complexity K(x) of a string x ∈ {0, 1}∗
is the length of a shortest string y ∈ {0, 1}∗ such that the universal Turing
machine UTM outputs x on input y.

This means, that we measure the number of bits that we need to produce
x. This can be seen as the ultimate compression task. The input y for UTM
is the compressed version of x.
Why not just take the length of a shortest encoding of a Turing machine
that outputs x on the empty word? The problem is that the encoding of
a Turing machine that outputs x on the empty word usually needs |x| + 1
states. Thus the trivial encoding of x would have length Θ(|x| log |x|), which
is not a desaster but also not very nice.

Lemma C.2 There is a constant u such that K(x) ≤ |x| + u for all x ∈
{0, 1}∗ .

Proof. Let e be the Gödel number of a Turing machine that computes

the identity. The universal Turing machine outputs x on input [e, x]. We
have |[e, x]| ≤ 2|e| + 2 + |x| (or ≤ |e| + |x| + 2 log |e| + 2 depending on our
pairing function).1 Setting u = 2|e| + 2 concludes the proof.

Example C.3 Consider the sequence zn = 0n , n ∈ N. The following Turing

machine outputs 0n on input bin(n):

1
Here is the place where we need that the length of a pair [a, b] can be bounded nicely
in terms of |a| and |b|.

99
100 C. Kolmogorov Complexity

Input: x = bin(n)

1. Count in binary to x.

2. Every time you increase the counter, output “0”.

Thus K(zn ) ≤ | bin(n)| + O(1) = log |zn | + O(1).

Exercise C.1 Let f : N → N be total, injective, and WHILE computable.

Let zn = 0f (n) . Show that K(zn ) ≤ log f −1 (|zn |) + O(1).

The exercise above shows that there are sequences of words whose Kol-
mogorov complexity grows very slowly. Are there sequences whose complex-
ity is close to the upper bound of Lemma C.2?

Pitfall

Kolmogorov complexity only measures the length of the input string

y for UTM . It does not say anything about the number of cells used
by UTM on input y!

Lemma C.4 For every natural number n ≥ 1, there is a word wn ∈ {0, 1}n
such that K(wn ) ≥ n.

Proof. There are 1 + 2 + 4 + . . . 2n−1 = 2n − 1 strings in {0, 1}∗ of length

< n. Thus there are at most 2n − 1 string with Kologorov complexity < n,
namely, the outputs of UTM on these 2n − 1 strings. There are 2n words in
{0, 1}n . Thus at least one of them has Kolmogorov complexity n.

Exercise C.2 Show that there are 2n (1 − 2−i ) words w in {0, 1}n such that
K(w) ≥ n − i.

We have defined Kolmogorov in terms of Turing machines. The definition

of K depends on our universal Turing machine UTM . But there are many
universal Turing machines. Does the choice of the Turing machine affect the
Kolmogorov complexity? The answer is “not very much”. For a universal
Turing machine V , let KV be the Kolmogorov complexity with respect to V .
In particular, the K that we used so far is KUTM .

Lemma C.5 Let V and V 0 be univeral Turing machines. Then there is a

constant cV 0 such that KV (x) ≤ KV 0 (x) + cV 0 for all x ∈ {0, 1}∗ .

c Markus Bläser 2007–2015
C.2. Kolmogorov random strings 101

Proof. Let y be a string of length KV 0 (x) such that V 0 on input y outputs

x. Let g 0 := gödTM (V 0 ). Then V outputs x on input [g 0 , y]. Thus

KV (x) ≤ |[g 0 , y]| ≤ 2|g 0 | + 2 + |y| = 2|g 0 | + 2 + KV 0 (x).

Setting cV 0 = 2|g 0 | + 2 concludes the proof.

Corollary C.6 For all universal Turing machines V and V 0 , there is a con-
stant cV,V 0 such that |KV (x) − KV 0 (x)| ≤ cV,V 0 for all x ∈ {0, 1}∗ .

Remark C.7 Above, the Turing machines V and V 0 could even work with
different Gödel numberings. We could even compare Turing machines with
any other acceptable programming system.

C.2 Kolmogorov random strings

Definition C.8 A string x ∈ {0, 1}∗ is Kolmogorov random if K(x) ≥ |x|.

Kolmogorov random means that a string is incompressible. It does not

have any “regular structure” that we could exploit to represent it by some
shorter string. By Lemma C.4, Kolmogorov random strings exist. Kol-
mogorov randomness only speaks about one string at a time. It does not
say anything about drawing strings at random. If we draw strings from
{0, 1}n uniformly at random, then we see every string with the same prob-
ability 2−n , in particular the string 0n and any Kolmogorov random string,
too.

Theorem C.9 Let A be a decidable language. Let an be the nth word in A

with respect to the lexicographic ordering. Then K(an ) ≤ log n + O(1).

Proof. Consider the following Turing machine:

Input: bin(n)

1. Enumerate {0, 1}∗ in the lexicographic order. Let x be the

current word.

2. Since A is decidable, we can check whether x is in A.

3. Repeat this process until n words in A have been found.

4. Print x.

c Markus Bläser 2007–2015
102 C. Kolmogorov Complexity

This Turing machine finds an on input bin(n). Let e be the Gödel number
of this machine. By construction, UTM on input [e, bin(n)] outputs an .

Remark C.10 If there is an  > 0 such that |A ∩ {0, 1}n | ≤ (2 − )n for all
n, then no long enough x in A can be Kolmogorov random.

C.3 Lower bounds for 1-tape Turing machines

In the exercises, we constructed a 1-tape Turing machine that on input x#
produces x#x, i.e., it copies x to the right of #. The number of steps of
such a Turing machine is  · n2 for some  > 0.
We can use Kolmogorov complexity to prove that this is best possible.
Let M be a Turing machine that on input x# produces x#x. Let q be the
number of states of M . Let y be a string of length n. Consider the string
y0n #. Let t be the number of steps that M computes on y0n #.
Consider the computation on y0n #. For each cell i with n + 1 ≤ i ≤ 2n,
we write down the sequence of states of M , when M moves its head from
cell i to cell i + 1.

Lemma C.11 If q1 , . . . , qs is such a sequence of states as decribed above,

then K(y) ≤ s · log q + log log q + log n + O(1).

Proof. Assume that q1 , . . . , qs is the sequence between the (n + j)th and

(n + j + 1)th cell. The following Turing machine N outputs y on input
q1 , . . . , qs and log j.

Input: q1 , . . . , qs , log j

1. Print 0n−j+1 # on the tape

2. For i = 1, . . . , s start on the leftmost 0 and simulate M starting

in state qi until M enters the cell left to the leftmost 0.

3. Remove all unneccessary 0’s and #’s.

N simulates the parts of the computation of M to the right of the (n+j)th

cell. When M moves from the (n + j)th cell to the right, the only thing that
determines its behaviour is the state. (Here we use that M has only one
tape!). Thus N will produce 0n−j+1 #y0n in step 2. In step 3, N just
discards all the 0’s and the #. Thus it produces y in the end.
The length of the input is s · log q + log log q for the states (we can use a
fixed length code with words of length q) plus log n for the cell number.

c Markus Bläser 2007–2015
C.4. Undecidability of the Kolmogorov Complexity 103

Lemma C.12 For each n + 1 ≤ i ≤ 2n, let si be the length of the sequence
of states between cell i and i + 1. There is an n + 1 ≤ j ≤ 2n such that
sj ≤ t/n, where t is the total number of steps of M on y0n #

Proof. Every state in a sequence

P2n of states corresponds to one step of the
Turing machine. Hence t ≥ i=n+1 si .
Now take a Kolmogorov random y, i.e., K(y) ≥ n. Let j as in Lemma C.12.
By Lemmas C.11 and C.12,

K(y) − log log q − log n − O(1) n2

t ≥ n · sj ≥ ≥ − log n − O(1).
log q log q

Theorem C.13 For every 1-tape Turing machine M that solves the copy
problem, there is an  > 0 such that M makes at least  · n2 steps in the
worst case on inputs of length n.

C.4 Undecidability of the Kolmogorov Complexity

The halting problem was our “starting” undecidable problem. All other re-
sults about non-computability followed by reduction. Kolmogorov complex-
ity is another natural problem for which it is easy to give a direct proof that
it is not Turing computable.

Theorem C.14 K is not Turing computable.

Proof. Assume on the contrary that there is a Turing machine M that

computes K. Let xn be the first word with respect to the lexicographic
order such that K(xn ) ≥ n. (xn exists, since Kolmogorov random words
have arbitrarily high complexity.)
We construct a Turing machine N that on input bin n, outputs xn as
follows:

Input: bin n

1. Enumerate {0, 1}∗ in lexicographic order. Let x be the current

string.

2. Compute K(x) using M .

3. If K(x) < n, then go on with the next x.

4. Otherwise, print x.

c Markus Bläser 2007–2015
104 C. Kolmogorov Complexity

By construction, N on input bin n produces the lexicographically first

string whose Kolmogorov complexity is at least n. Let e be an index of N .
On input [e, bin n], UTM prints xn . Hence

K(xn ) ≤ |[e, bin n]| ≤ log n + O(1).

This constradicts K(xn ) ≥ n for large enough n.

K is a function and not a language but we easily get a language out of
it.

Exercise C.3 Show that the language {[x, cod(k)] | K(x) = k} is not de-
cidable.

K is not harder then H, since we can show the following result.

Theorem C.15 If H ∈ REC, then K is Turing computable.

Proof. Let M be a Turing machine that decides H. The following Turing

machine N computes K(x):

Input: x ∈ {0, 1}∗

1. Enumerate {0, 1}∗ in lexicographic order. Let y be the current

string.

2. Use M to decide whether UTM halts on y.

3. If yes, simulate UTM on y. Let z be the output.

4. If z 6= x, then go on with the next y.

5. Return |z|.

N always terminates since in step 2, we check whether UTM will halt and
since there is always a string of length |x| + O(1) such that UTM outputs
x. By construction, N finds a shortest string y such that UTM on input y
outputs x. The length |y| of this is K(x).

c Markus Bläser 2007–2015
 Part II

Complexity

105
16 Turing machines and
complexity classes

Computability theory tries to separate problems that can be solved algorith-

mically from problems that cannot be solved algorithmically. Here, “can”
and “cannot” means that there exists or does not exist a Turing machine,
WHILE program, JAVA program, etc. that decides or recognizes the given
language.
While it is nice to know that there is a Turing machine, WHILE pro-
gram, or JAVA program that decides my problem, it does not help at all
if the running time is so large that I will not live long enough to see the
outcome. Complexity theory tries to separate problems that can be solved
in an acceptable amount of time (“feasible” or “tractable” problems) from
problems that cannot be solved in an acceptable amount of time (“infeasi-
ble” or “intracable” problems). Space comsumption is another resource that
we will investigate.

16.1 Deterministic complexity classes

Let M be a deterministic Turing machine and let x be an input. Assume
that M halts on x, i.e., there is a unique accepting or rejecting configuration
Ct such that SC(x) `∗M Ct . Then, by definition, there is a unique sequence

SC(x) `M C1 `M · · · `M Ct .

This sequence is called a computation of M on x. t is the number of steps

that M performs on input x. We denote this number t by TimeM (x). If M
does not halt on x, then the computation of M on x is infinite. In this case
TimeM (x) is infinite.
For an n ∈ N, we define the time complexity of M as

TimeM (n) = max{TimeM (x) | |x| = n}.

In other words, TimeM (n) measures the worst case behaviour of M on inputs
of length n. Let t : N → N be some function. A deterministic Turing machine
M is t time bounded if TimeM (n) ≤ t(n) for all n.
For a configuration C = (q, (p1 , x1 ), . . . (pk , xk )), Space(C) = max1≤κ≤k |xκ |
is the space used by the configuration. Occassionally, we will equip Turing
machines with an extra input tape. This input tape contains, guess what,
the input x of the Turing machine. This input tape is read-only, that is, the

106
16.1. Deterministic complexity classes 107

Turing machine can only read the symbols but not change them. (Techni-
cally, this is achieved by requiring that whenever the Turing machine reads
a symbol on the input tape it has to write the same symbol.) What is an
extra input tape good for? The space used on the input tape (that is, the
symbols occupied by the input) is not counted in the definition of Space(C).
In this way, we can talk about sublinear space complexity.

Example 16.1 Consider the language

L = {x ∈ {0, 1}∗ | the number of 0’s in x equals the number of 1’s}.

L can be recognized with space O(log n). We read the input and for every
0 that we encounter, we increase a binary counter on the work tape by one.
Then we read the input a second time and decrease the counter for every 1.
We accept if in the end, the counter on the work tape is zero. In every step,
we store number ≤ |x| on the work tape. This needs log n bits (on the work
tape).

Let M be a deterministic Turing machine and let x be an input. First

assume that M halts on x. Let

SC(x) `M C1 `M · · · `M Ct

be the computation of M on x. Then SpaceM (x) = max{Space(Cτ ) | 1 ≤

τ ≤ t}. If M does not halt on x, then we build the maximum over infinitely
many configurations. If the maximum does not exist, then SpaceM (x) = ∞.
For an n ∈ N, we define the space complexity of M as

SpaceM (n) = max{SpaceM (x) | |x| = n}.

In other words, SpaceM (n) measures the worst case behaviour of M on

inputs of length n. Let s : N → N be some function. A deterministic Turing
machine M is s space bounded if SpaceM (n) ≤ s(n) for all n.
A language L is deterministically t time decidable iff there is a deter-
ministic Turing machine M such that L = L(M ) and TimeM (n) ≤ t(n)
for all n. In the same way, a function f is deterministically computable in
time t(n) iff there is a deterministic Turing machine M that computes f
and TimeM (n) ≤ t(n) for all n. Note that a time bounded Turing machine
always halts.

Definition 16.2 Let t : N → N. Then

DTime(t) = {L | L is deterministically t time decidable},

DTimek (t) = {L | there is a t time bounded k-tape Turing machine M
with L = L(M )}.

c Markus Bläser 2007–2015
108 16. Turing machines and complexity classes

S
For a set of functions T , DTime(T ) = t∈T DTime(t). DTimek (T ) is defined
analogously.
The same is done for space complexity: A language L is deterministically
s space recognizable if there is a deterministic Turing machine M such that
L = L(M ) and SpaceM (n) ≤ s(n) for all n. Note that a space bounded
Turing machine might not halt on inputs that are not in L(M ). But we will
see in the next chapter that one can effectively detect when a space bounded
machine has entered an infinite loop. In the same way, a function f is deter-
ministically computable in space s if there is a deterministic Turing machine
M that computes f and SpaceM (n) ≤ s(n) for all n. We will see that for
space bounded computations, also sublinear functions s are meaningful. But
to speak of sublinear space complexity, the input should not be counted. We
will use a Turing machine M with an extra input tape.

Definition 16.3 Let s : N → N. Then

DSpace(s) = {L | L is deterministically s space recognizable},

DSpacek (s) = {L | there is a s space bounded k-tape Turing machine M
with L = L(M )}.

In the definition of DSpace(s), the Turing machines have an additional input

tape.
S
For a set of functions S, DSpace(S) = s∈S DSpace(s). DSpacek (S) is de-
fined analogously.

Exercise 16.1 Intuitively, it is clear that sublinear time is not very mean-
ingful here.1 Give a formal proof for this. In particular show: Let M
be a deterministic Turing machine. Assume that there is an n such that
M reads at most n − 1 symbols of the input x for each x with |x| = n.
Smare words ∗a1 , . . . , am with |ai | < n for all 1 ≤ i ≤ m such that
Then there
L(M ) = i=1 ai {0, 1} .

16.2 Nondeterministic complexity classes

In the following chapters, we will need nondeterministic Turing machines,
too. Instead of a function

δ : Q × Γk → Q × Γk × {L, S, R}k ,

the transition function is now a function

δ : Q × Γk → P(Q × Γk × {L, S, R}k ).

1
This however changes if we have random access to the input and we are content with
approximate results. Sublinear time algorithms are a very active area of research right
now.

c Markus Bläser 2007–2015
16.2. Nondeterministic complexity classes 109

1;1,R 0;1,R
0;0,R 1;0,R

0;1,R
S
ident invert stop
0;1,R

Figure 16.1: A nondeterministic 1-tape Turing machine. Whenever the Tur-

ing machine reads a 0, it may switch between the states ident and invert.
In the state ident, the machine does not change the bits read. In the state
invert, it exchanges 0’s with 1’s and vice versa.

If in a given state q, a Turing machine reads the symbols γ1 , . . . , γk on the

tapes, then it now has several possibilities of performing a step. Therefore,
a configuration C now has several successor configurations, one for each
possible step that the Turing machine can perform.
Let M be a nondeterministic Turing machine and x ∈ Σ∗ . We define a
(possibly infinite) rooted labeled tree T , the computation tree of M on x as
follows: The root is labeled with SC(x). As long as there is a node v that
is labeled with a configuration C that is not a halting configuration, we do
the following: Let C1 , . . . , C` be all configurations such that C `M Cλ for
1 ≤ λ ≤ `. Now v gets ` children labeled with C1 , . . . , C` . (Note that it is
possible that different nodes in T might have the same label.) A path from
the root to a leaf in T is called a computation path. It is accepting if the con-
figuration at the leaf is accepting, otherwise it is rejecting. There also might
be infinite computation paths; these are neither accepting nor rejecting. A
nondeterministic Turing machine accepts an input x iff the corresponding
computation tree has an accepting computation path. Note that if M is de-
terministic, then the computation tree is a path. A nondeterministic Turing
machine recognizes the language
L(M ) = {x ∈ Σ∗ |

Example 16.4 Figure 16.1 shows a nondeterministic Turing machine M .

In the state ident, M just goes to the right and leaves the content of the
tape unchanged. In the state invert, it goes to the right and replaces every 0
by a 1 and vice versa. Whenever it reads a 1, M may nondeterministically
choose to stay in its current state or to go the other state. M accepts if it
is in the state invert after reading the whole input. Figure 16.2 shows the
computation tree on input 010. It is quite easy to see that L(M ) = {x |
x contains at least one 0}.

TimeM (x) is the length of a shortest accepting computation path in the

computation tree of M on x, if such a path exists, and ∞ otherwise. We set
TimeM (n) = max{TimeM (x) | |x| = n, x ∈ L(M )}.

c Markus Bläser 2007–2015
110 16. Turing machines and complexity classes

 0 1 0  ident

 0 1 0  ident 1 1 0  invert

 0 1 0  ident 1 0 0  invert

 0 1 0  ident  0 1 1  invert 1 0 0  ident 1 0 1  invert

 0 1 1  stop 1 0 1  stop

Figure 16.2: The computation tree of the Turing machine from Figure 16.1
on the word 010. The nodes are labeled with the configurations, right to the
tape, there is the current state standing. The position of the head is marked
by the small triangle. There are four paths in the computation tree, two of
them are accepting (the state is stop).

If there is no x ∈ L(M ) with length n, then TimeM (n) = 0. Note that this
definition is somewhat different to the deterministic case, where we took
the maximum over all x of length n. Let t : N → N be some function. A
nondeterministic Turing machine M is weakly t time bounded if TimeM (n) ≤
t(n) for all n.

Exercise 16.2 Show that for any nondeterministic Turing machine M that
is weakly t time bounded there is an equivalent Turing machine M 0 (i.e.,
M (x) = M 0 (x) for all x) that is weakly O(t) time bounded such that for
every input x, the computation tree of M 0 on x is a binary tree.

If every computation of M on every x (and not only in L(M )) has length

at most t(|x|), then M is strongly t time bounded. Although strongly time
bounded seems to be stronger than weakly time bounded, we will see soon
that these two concepts lead to the same complexity classes for all “reason-
able” 2 time bounds.

Definition 16.5 Let t : N → N. Then

NTime(t) = {L | there is a weakly t time bounded nondeterministic

Turing machine M with L = L(M )},
NTimek (t) = {L | there is a weakly t time bounded nondeterministic
k-tape Turing machine M with L = L(M )}.
2
One definition of reasonable is the following: Pick your favourite book on algorithms
and open a random page. If you see a function N → N on this page, then it is reasonable,
maybe except for the inverse of the Ackermann function.

c Markus Bläser 2007–2015
16.3. An example 111

S
For a set of functions T , NTime(T ) = t∈T NTime(t). NTimek (T ) is defined
analogously.
Warning! I am fully aware of the fact that there does not exists a
physical realization of a nondeterministic Turing machine! (At least, I do
not know of any.) Nondeterministic Turing machine are not interesting per
se (at least not for an overwhelming majority of the world population), they
are interesting because they characterize important classes of problems. The
most important ones are the so-called NP-complete problems, a class which
we will encounter soon. The example in Section 16.3 gives a first impression.
For a nondeterministic Turing machine M and an input x ∈ L(M ), we
define space SpaceM (x) as follows: we take the minimum over all accepting
paths of the maximum of the space used by any configuration along this path
if such an accepting path exists, and ∞ otherwise. We set

SpaceM (n) = max{SpaceM (x) | |x| = n, x ∈ L(M )}.

If there is no x of length n in L(M ), then SpaceM (n) = 0. Let s : N → N

be some function. A nondeterministic Turing machine M is weakly s space
bounded if SpaceM (n) ≤ s(n) for all n. We define strongly s space bounded
in the same way as we did for strongly time bounded.

Definition 16.6 Let s : N → N. Then

NSpace(s) = {L | there is a weakly s space bounded nondeterministic

Turing machine M with L = L(M )},
NSpacek (s) = {L | there is a weakly s space bounded nondeterministic
k-tape Turing machine M with L = L(M )}.

In the case of NSpace(s), the Turing machines have an extra input tape.

16.3 An example
Consider the following arithmetic formula

x1 + 2x2 (1 − x1 ) + x3 .

We want to know whether we can assign the values 0 and 1 to the variables
in such a way that the formula evaluates to 1. Above x1 7→ 1, x2 7→ 0, and
x3 7→ 0 is such an assignment. The formula

x1 (1 − x1 )

does not have such an assignmet. We want to decide whether a given formula
has such an assignment or not.

c Markus Bläser 2007–2015
112 16. Turing machines and complexity classes

To make formulas accessible to Turing machines, we have to encode them

as binary strings. The actual way how we do this will not matter in the
following, as long as the encoding is “easily accessible”. Here, this means
that given an assignment, we can easily evaluate the formula F in time, say,
O(`3 ) 3 where ` is the length of (the encoding of) the formula.
The following excursus formalizes the problem, but I recommend to skip
it first.

Excursus: Formalization

Let X = {x1 , x2 , . . . } be a set of variables. Arithmetic formulas are defined

inductively:

1. Every x ∈ X and every z ∈ Z is an arithmetic formula.

2. If F and G are arithmetic formulas, then (F · G) and (F + G) are arithmetic

formulas.

An assignment is a map a : X → Z. If we replace every occurence of a variable

x by a(x) in a formula F , then F just describes an integer. We extend a to the set
of all arithmetic formulas inductively along the above definition:

1. a(z) = z for all z ∈ Z.

2. a(F · G) = a(F ) · a(G) and a(F + G) = a(F ) + a(G) for formulas F and G.

Since in every formula, only a finite number of variable occur, we usually restrict
assignments to the variables occurring in a given formula. An assignment is called
an S assignment for some S ⊆ Z, if im a ⊆ S.
We can encode arithmetic formulas as follows: For instance, we can encode the
variable xi by 0 bin(i) and a constant n by 1σ(z) bin(|z|) where σ(z) is 1 if z ≥ 0 and
0 otherwise. Then we define the encoding c inductively by c(F · G) = [0, c(F ), c(G)]
and c(F +G) = [1, c(F ), c(G)]. This is a very structured encoding, since it explicitly
stores the order in which operations are performed. Alternatively, we first encode
xi by the string x bin(i) and z by σ(z) bin(|z|). Now we can view our formula
as a string over the alphabet {(, ), +, ·, x, 0, 1}. To get a string over {0, 1}, we just
replace each of the seven symbols by a different binary string of fixed length. (Three
is sufficient, since 23 ≥ 7.) This is a rather unstructured encoding. Nevertheless,
both encodings allow us to evaluate the formula in time O(`3 ).

Since the encoding does not matter in the following, we will not specify
it explicitly. We just assume that the encoding is reasonable. Since there
3
O(`3 ) can be easily achieved for reasonable encodings: A formula F of length ` has
at most ` arithmetic operations and the value of the formula in the end has at most `
bits (proof by induction). Addition and multiplication can be performed in time O(`2 ) by
the methods that you learn in school and we have ≤ ` of them. Using more sophisticated
methods and a better analysis, one can bring down the evaluation time to O(`1+ ) for any
 > 0.

c Markus Bläser 2007–2015
16.3. An example 113

is usually no danger of confusion, we will even write F for both the for-
mula itself (as a mathematical object) and its encoding (instead of c(F ) or
something like that). Let

AFSAT = {F | there is an {0, 1} assignment such that F evaluates to 1.}

How hard is it to decide whether a given formula F has a {0, 1} assignment

such that F evaluates to 1? I.e., how hard is it to decide whether F ∈
AFSAT? Assume that F has length n. Since F can have at most n variables,
there are 2n possible assignments. For each of them we can check in time
O(n3 ) whether F evaluates to 1 or not. Thus

AFSAT ∈ DTime(O(2n · n3 ))

A nondeterministic Turing machine can do the following. It first reads the

input and for each symbol it reads it has two options: Either write a 0 on
the second tape or a 1, see Figure 16.3. The computation tree has 2n paths.
At every leaf, the machine has written one string from {0, 1}n on the second
tape. We now interpret this string as an assignment to the variables, ignoring
some of the bits if there are fewer than n variables. The machine now just
(deterministically) evaluates the formula with respect to this assignment and
accepts if the outcome is 1 and rejects otherwise.
The machine is clearly O(n3 ) time bounded, since the length of each com-
putation path is dominated by the time needed for evaluating the formula.
It correctly decides whether F ∈ AFSAT, too: If there is a {0, 1} assignment
such that F evaluates to 1, then it will be generated along some computation
path and at the end of this path, the Turing machine will accept. If there is
no such assignment, then there cannot be any accepting computation path
at all. Hence
AFSAT ∈ NTime(O(n3 )).
The deterministic Turing machine tries all possible assignments, one after
another. It is an open problem whether there is a substantially more clever
way. A nondeterministic Turing machine can try them in parallel. Or we
can view it like this: If we are given an assignment, then we can easily check
whether F evaluates to 1 under this assignment.
Like AFSAT (which is more of pedagogical value), there are an abun-
dance of similar and very, very, very important problems (so called NP-
complete problems), that have the same property: To find a solution, we do
not know anything really better than trying all possible solution. But if we
get a potential solution, we can easily verify whether it is really a solution.
This is what give nondeterminism its right to exists . . . (and there might be
some other reasons, too)

Excursus: Complexity of WHILE programs

c Markus Bläser 2007–2015
114 16. Turing machines and complexity classes

1;1,R| 1,R
0;0,R| 0,R

S| S
gen stop

Figure 16.3: A nondeterminstic Turing machine that generates all strings

from {0, 1}n .

We proved that WHILE computable functions and Turing computable functions

are essentially the same. But what about time and space complexity. Are there
functions that can be computed much faster by WHILE programs than be Turing
machines or vice versa? The answers are “no” and “it depends”.
First of all, Turing machines get bit strings as inputs and WHILE programs get
natural numbers. The complexity of a Turing machine is measured in the length
of the input. So we should measure the running time of a WHILE program as a
function in | cod−1 (n)| = log n. The running time of a WHILE program on some
input i is the number of simple statements that are executed on i. (If a simple
statement is executed several times, then it is of course counted this many times.)
The running time on inputs of length ` is the maximum over the running times on
inputs i with | cod−1 (i)| = `. If S = (σ0 , . . . , σ` ) is a state of a WHILE program,
then the space used by this state is Space(S) := max{| cod−1 (σλ )| | 0 ≤ λ ≤ `}.
Let P be a simple statement. Assume we have a state S = (σ0 , . . . , σ` ) and let
S 0 = (σ00 , . . . σ`0 ) := ΦP (S 0 ). We claim that Space(S 0 ) ≤ 1 + Space(S), if Space(S) is
large enough. Assume that P = xi := xj + xk . Then σi0 = σj + σk and all other en-
tries are not changed. But then | cod−1 (σi0 )| ≤ 1 + max{| cod−1 (σj )|, | cod−1 (σk )|}.
The same is true if P is a subtraction. The case P = xi := c can only change
Space(S) is Space(S) ≤ | cod−1 (c)|. But every WHILE program contains only a
finite number of constants. This means the assigning constants does not have any
asymptotic effect. It follows by induction that by executing t simple statements,
we can increase the space consumption by at most t.
Now consider the simulation of a WHILE program by a Turing machine. We
first replaced the WHILE program by a GOTO programm but the number of simple
statements and the number of space used by the GOTO program is the essentially
same. In this simulation, the content of each variable is stored on a different tape.
Hence we do not need more space in this simulation than the GOTO program does.
To simulate a simple statement, we have just to add or subtract two numbers. This
can be done in time linear in the size of the operands. (We only used incrementation
and decrementation in our simulation, but it is easily extended to addition and
subtraction.) Thus we only get a quadratic slowdown when simulating WHILE
programs by Turing machines.
When we simulated Turing machines by WHILE programs, we stored the tapes
in array and then could easily do a step-by-step simulation. So the running time
of the Turing machines is multiplied by the time needed to manipulate the arrays.
If we use the ordinary pairing function h., .i, then the sizes can explode. But since
we only store elements a0 , . . . , as from a finite set {0, 1, . . . ,Pb}, say, we can do this
s
by interpreting a0 , . . . , as as the digits of a b-nary number i=0 ai bi . In this way,

c Markus Bläser 2007–2015
16.3. An example 115

we again only get a quadratic slowdown.

c Markus Bläser 2007–2015
17 Tape reduction, compression,
and acceleration

In this chapter, we further investigate the Turing machine model before we

come to results that also hold in other models of computation in the next
chapters

17.1 Tape reduction

Definition 17.1 A deterministic Turing machine M simulates a Turing
machine M 0 , if L(M ) = L(M 0 ) and for all inputs x, M halts on x iff M 0
halts on x.

Theorem 17.2 Every deterministic Turing machine can be simulated by a

deterministic 1-tape Turing machine.

Remark 17.3 The construction is quite similar to the universal Turing ma-
chine that we constructed in the first part of this lecture.

Proof. Let M a be k-tape Turing machine. We construct a 1-tape Turing

machine S that simulates M . S simulates one step of M by a sequence of
steps.
We have to store the content of all k tapes on one tape. We think of the
tape of S divided into 2k tracks. To this aim, we enlarge the tape alphabet
of S. The work alphabet of S is Γ0 = (Γ × {∗, −})k ∪ Σ ∪ {}. The (2κ − 1)th
component of a symbol of Γ0 stores the content of the κth tape of M . The
2κth component is used to mark the position of the head on the κth tape.
There will be exactly one ∗ on the 2κ track, this ∗ will mark the position of
the head. All other entries of the track will be filled with −’s. Figure 17.1
depicts this construction: One column on the righthand side of the figure is
one symbol on the tape of S. In particular, a  or − in such a column is not
the blank of S. The blank of S is just , one column just filled with one .
Figure 17.2 shows how the simulating machine S works. Let x ∈ Σ∗ be
the given input. First, S replaces the input x by the corresponding symbols
from (Γ × {∗, −})k . The first track contains x, all other odd tracks contain
blanks. On the even tracks, the ∗ is in the first position of each track.
One step of M is now simulated as follows: S always starts on the left-
most position of the tape visited so far and moves to the right until it reaches

116
17.1. Tape reduction 117

0 1 1 0 1 1
− ∗ −
0 0 1 0 0 1
− − −
..
..

1 0  1 0 
∗ − −

Figure 17.1: Lefthand side: The k tapes of the k-tape Turing machine M .
Righthand-side: The one and only tape of the simulating machine S. The
tape of S is divided into 2k tracks, two for each tape of M . The first track
of each such pair of tracks stores the content of the corresponding tape of
M , the second stores the position of the head which is marked by “∗”.

the first blank (of S)1 . On its way, S collects the k symbols under the heads of
M and stores them in its finite control. Once S has collected all the symbols,
it can simulate the transition of M . It changes the state accordingly and
now moves to the left until it reaches the first blank (of S). On its way back,
it makes the changes that M would make. It replaces the entries in the
components marked by a ∗ and moves the ∗ in the corresponding direction.
If M has not halted yet, S repeats the loop described above. If M halts,
S halts, too, and accepts iff M accepts.

Remark 17.4 The above construction also works for nondeterministic Tur-
ing machines. Whenever S has collected all the symbols and simulates the
actual transition of M , it chooses one possible transition nondeterministi-
cally.

Remark 17.5 If M has an additional input tape, then we can also equip S
with an additional input tape. If M has a sublinear space bound, then also
S has a sublinear space bound.

Remark 17.6 (Implementation details) The description of the simula-

tor in the above proof is rather high level. A more low level description,
i.e., the explicit transition function, usually does not provide any insights.
1
A blank of S indicates that we reached a position that M has not visited so far on
any of its tapes.

c Markus Bläser 2007–2015
118 17. Tape reduction, compression, and acceleration

Input: x ∈ Σ∗
Output: accept if x ∈ L(M ), reject otherwise

1. S replaces the input x by the corresponding symbol of Γ0 ,

i.e., x1 is replaced by (x1 , ∗, , ∗, . . . , , ∗) and each other xν is
replaced by (xκ , −, , −, . . . , , −).

2. S always stores the current state of M in its finite control. In

the beginning, this is the starting state of M .

3. As long as M does not halt, S repeats the following:

(a) S moves to the right until it reaches the first blank. On

its way to the right, S reads the symbols that the heads
of M are reading and stores them in its finite control.
(b) When S reaches the right end of the tape content, it has
gathered all the information to simulate one step of M . It
changes the internally stored state of M accordingly.
(c) S now moves to the left until it reaches the first blank. On
its way to the left, S replaces the entries in components
that are marked by a ∗ by the symbol that M would write
on the corresponding tape and moves the ∗ like M would
move the corresponding head.
If S has to move on of the markers ∗ to a cell that still
contains , the blank of S, then it first replaces this blank
by (, −, . . . , , −).

4. If M accepts, S accepts. Otherwise, S rejects.

Figure 17.2: The simulator S.

c Markus Bläser 2007–2015
17.1. Tape reduction 119

But we tacitely assume that you can write down—at least in principle—the
transition function of the 1-tape Turing machine constructed above.
Let’s convince ourselves that we can really do this: Consider the part of
S that collects the symbols that M would read. The states are of the form
{collect} × Q × (Γ ∪ {/})k . The first entry of a tuple (collect, q, γ1 , . . . , γk )
indicates that we are in a collection phase. (If the collection phase were the
only phase that uses tuple of the form Q × (Γ ∪ {/})k , then we could skip
this first component.) The second component stores the current state of M .
It shall not be changed during the collection phase. Finally, γκ stores the
symbol that is read by M on the κth tape. γκ = / indicates that the position
of the head on tape κ has not been found yet.
The transition function δ 0 of S (restricted to the states of the collect
phase) is now defined as follows:

δ 0 ((collect, q, γ1 , . . . , γk ), (η1 , . . . , η2k )) = ((collect, q, γ10 , . . . , γk0 ), (η1 , . . . , η2k ), R)

for all q ∈ Q, γ1 , . . . , γk ∈ Γ ∪ {/} where

(
0 γκ if η2κ = −
γκ =
η2κ−1 if η2κ = ∗

for 1 ≤ κ ≤ k. If the symbol η2κ on the 2κth track is ∗, then we found the
head on the κth tape and store η2κ−1 , the symbol that M reads on the κth
tape, in the state of S.

Definition 17.7 Let t, s : N → N. Then

DTimeSpace(t, s) = {L | there is a t time and s space bounded

Turing machine M with L = L(M )}.

DTimeSpacek (t, s) and NTimeSpacek (t, s), . . . are defined accordingly.

Theorem 17.8 For all t, s : N → N,

DTimeSpace(t, s) ⊆ DTimeSpace1 (O(ts), O(s)),

NTimeSpace(t, s) ⊆ NTimeSpace1 (O(ts), O(s)).

Proof. Let L ∈ DTimeSpace(t, s). Let M be a deterministic Turing

machine with L = L(M ). Assume that M has k tapes. The Turing machine
S in Theorem 17.2 simulates one step of M by O(s(|x|)) steps (where x is
the given input). M makes at most t(|x|) steps. Furthermore, S does not
use more twice the space M uses. (On each track, S does not use more space
then M on the corresponding tape. But on one tape M could use the cells
to the left of cell 0 and on the other to the right.)
The nondeterministic case follows in the same way by Remark 17.4.

c Markus Bläser 2007–2015
120 17. Tape reduction, compression, and acceleration

Corollary 17.9 For all t : N → N,

DTime(t) ⊆ DTime1 (O(t2 )),

NTime(t) ⊆ NTime1 (O(t2 )).

Proof. Let M be a t time bounded Turing machine. In t steps, a turing

machine can visit at most t cells. Thus SpaceM (n) ≤ t(n) for all n, and the
corollary follows from Theorem 17.8.

17.2 Tape compression

The aim of this and the next section is to show that we do not need to
take care of constant factors. Every Turing machine M can be simulated
by another one that uses only a constant fraction of the space used by M .
(Note that we will not even care about polynomial factors in the following.)

Theorem 17.10 For all 0 <  ≤ 1 and all s : N → N,

DSpace(s(n)) ⊆ DSpace1,E (ds(n)e)

NSpace(s(n)) ⊆ NSpace1,E (ds(n)e)

Proof overview: In the same way as a 64 bit architectur can store more
information in one memory cell than an 8 bit architectur, we enlarge the
tape alphabet to store several symbols in one symbol and then just simulate.

Proof. Let c = d1/e. Let M be a deterministic k-tape Turing machine

with work alphabet Γ. We simulate M by a deterministic k-tape Turing
machine with work alphabet Γ0 = Γc ∪ Σ ∪ {}. A block of c contiguous
cells of a tape of M are coded into one cell of S. Instead of s cells, S then
uses only ds/ce ≤ dse cells. S can simulate M step by step. S stores the
position of the head within a block of c cells in its state. If M moves his head
within such a block, then S does not move its head at all but just changes
the symbol.
If M does not have an extra input tape, then it first has to compress the
input of length n into dn/ce cells. But in this case, M can never use less
than space n. If M has an extra input tape, this step is not necessary.
If M is nondeterministic, the same simulation works.

Remark 17.11 (Implementation details) Again, let’s try to formalize a

part of the transition function δ 0 of S. The states of S are of the form

c Markus Bläser 2007–2015
17.3. Acceleration 121

Q × {1, . . . , c}. (q, i) means that M is in state q and its head is on the ith
symbol of the current block. Assume that δ(q, η) = (q 0 , η 0 , R). Then
(
0 ((q 0 , i + 1), (γ10 , . . . , γc0 ), S) if i < c
δ ((q, i), (γ1 , . . . , γc )) =
((q 0 , 1), (γ10 , . . . , γc0 ), R) if i = c
for all q ∈ Q, i ∈ {1, . . . , c}, and all (γ1 , . . . , γc ) with γi = η, where γj0 = γj
for j 6= i and γi0 = η 0 .
Exercise 17.1 Show the following “converse” of Theorem 17.10: For any
s space and t time bounded Turing machine M with input alphabet {0, 1},
there is a O(s) space and O(t) time bounded Turing machine that only uses
the work alphabet {0, 1, }.

17.3 Acceleration
Next, we prove a similar speed up for time. This simulation is a little more
complicated than the previous one.
Exercise 17.2 Show the following: For all k ≥ 2, all t : N → N, and all
0 <  ≤ 1,
DTimek (t(n)) ⊆ DTimek (n + (n + t(n)))
NTimek (t(n)) ⊆ NTimek (n + (n + t(n))).

Proof overview: Like in Theorem 17.10, we want to store several, say c,

cells into one. To get a speed up, the simulating machine S now has to
simulate c steps in one step. This is no problem if M stays within the c cells
of one block, since we just can precompute the outcome. Problematic is the
following case: During the c steps, the Turing machine M goes back and
forth between two cells that belong to different (but neighboured) blocks.
To overcome this problem, S always stores three blocks in its finite control:
The block B where the head of M is located and the blocks to the left and
the right of B. With these three blocks, S can simulate c steps of M in its
finite control. Then S updates the tape content. If M left the block B, then
S also has to update the blocks in its finite control.

If t(n) = ω(n), then we can speed the computation by any factor  in

Exercise 17.2. If t(n) = O(n), then we can get a running time of (1 + )n for
any  > 0.

What to measure?

Time and space consumption of Turing machines should only be

measured up to constant factors!

c Markus Bläser 2007–2015
122 17. Tape reduction, compression, and acceleration

17.4 Further exercises

Exercise 17.3 Prove the following. Let c be some constant. Every k-tape
Turing machine M can be simulated by a k-tape Turing machine S such that
(
n if n ≤ c
TimeS (n) =
TimeM (n) + c otherwise
(
0 if n ≤ c
SpaceS (n) =
SpaceM (n) otherwise

In other words, only the asymptotic behaviour matters.

c Markus Bläser 2007–2015
18 Space versus Time,
Nondeterminism versus Deter-
minism

18.1 Constructible functions

Definition 18.1 Let s, t : N → N.

1. t is time constructible if there is a O(t) time bounded deterministic

Turing machine M that computes the function 1n 7→ bin(t(n)).

2. s is space constructible if there is an O(s) space bounded deterministic

Turing machine M (with extra input tape) that computes the function
1n 7→ bin(s(n)).

Above, bin(n) denotes the binary representation of n.

Exercise 18.1 Show the following:

1. If t is time constructible, then there is a O(t) time bounded determin-

istic Turing machine that on input x writes 1t(|x|) on one of its tapes.

2. If s is space constructible, then there is a s space bounded deterministic

Turing machine (with extra input tape) that on input x writes 1s(|x|)
on one of its tapes.

Time and space constructible functions “behave well”. One examples for
this is the following result.

Lemma 18.2 Let t be time constructible and s be space constructible.

1. If L ∈ NTime(t) then there is a strongly O(t) time bounded nondeter-

mistic Turing machine N with L = L(N ).

2. If L ∈ NSpace(s), then there is a strongly O(s) space bounded nonde-

termistic Turing machine N with L = L(N ).

Proof. We start with the first statement: Let M be some weakly t time
bounded Turing machine with L(M ) = L. Consider the following turing
machine N :

123
124 18. Space versus time, nondeterminism versus determinism

Input: x

1. Construct bin(t(|x|)) on some extra tapes.

2. Simulate M step by step.

On the extra tape, count the number of simulated steps with a
binary counter.

3. When more than t(|x|) steps have been simulated, then stop
and reject.

4. If M halts earlier, then accept if M has accepted and reject

otherwise.

N is clearly O(t) time bounded, since counting to t(|x|) in binary can be

done in O(t(|x|)) time (amortized analysis!). If M accepts x, then there is an
accepting path whose length is at most t(|x|). This path will be simulated
by N and hence N will accept x. If M does not accept x, then all paths
in the computation tree are either infinite or rejecting. In both cases, the
corresponding path of N will be rejecting.
For the second part, let M be some weakly s space bounded Turing
machine with an extra input tape such that L(M 0 ) = L. Consider the
following Turing machine N :

Input: x

1. Mark 2s(|x|) cells with a new symbol  on each work tape (see
Exercise 18.1), s(|x|) to the left of cell 0 and s(|x|) to the right.

2. Simulate M on x pretending that each  is a .

3. When we read a real blank  during the simulation, then we

stop and reject.

N is clearly O(s) space bounded. If M accepts x, then there is an

accepting computation path on which M is s space bounded. When N
simulates this path, then N will never reach a  and hence will accept. If
M does not accept x, then N will not accept, too.
Most interesting functions are space and time constructible.

Exercise 18.2 Let a, b, c ∈ N and let f (n) = 2an · nb · logc (n).

c Markus Bläser 2007–2015
18.2. The configuration graph 125

1. If f ∈ ω(n), then f is time constructible.

2. f is also space constructible. This even holds if a, b, c are rational
numberers provided that f ∈ Ω(log n).

18.2 The configuration graph

Let M be a Turing machine. The set of all configurations of M together
with the relation `M can be interpreted as an infinite directed graph. The
node set of this graph is the set of all configurations. We denote this set by
Conf M . We denote this graph by CGM = (Conf M , `M ).1 This is an infinite
graph. But to decide whether a Turing machine accepts an input x, we just
have to find out whether we can reach an accepting configuration from the
starting configuration SCM (x). This task is undecidable in general but it
becomes feasible when the Turing machine is time or space bounded. For a
given space bound s, the relevant part of CGM is finite.

Lemma 18.3 Let M be an s space bounded Turing machine with s(n) ≥

log n for all n. There is a constant c (depending on M ) such that M on
input x can reach at most cs(|x|) configurations from SC(x).

Proof. Let M = (Q, Σ, Γ, δ, q0 , Qacc ) be a k-tape Turing machine. A

configuration of M is described by the current state, the content of the work
tapes and the position of the heads. There are |Q| states, |Γ|s(|x|) possible
contents of a tape and s(|x|) possible positions of the heads. Thus the number
of configurations is at most
 k
|Q| · |Γ|s(|x|) · s(|x|)k · (|x| + 2) (18.1)

If the Turing machine does not have an extra input tape, the last factor
|x| + 2 is not necessary. It is easy to see that (18.1) is bounded by cs(|x|)
for some constant c only depending on |Q|, |Γ|, and k. (To bound the last
factor |x| + 2, we need the assumption s(n) ≥ log n.)

Exercise 18.3 Give an upper bound for the constant c above.

Corollary 18.4 Let s(n) ≥ log n for all n. If a deterministic s space

bounded Turing machine halts on an input x, then it can perform at most
cs(|x|) steps on x, where c is the constant from Lemma 18.3

Proof. By contradiction: If M makes more steps, then the computation of

M would contain the same configuration twice by the pigeon hole principle.
Since M is deterministic, this means that the computation is infinite.
1
Note that the binary relation `M is nothing else than a set of pairs of elements from
Conf M , that is, directed edges.

c Markus Bläser 2007–2015
126 18. Space versus time, nondeterminism versus determinism

Corollary 18.5 Let s(n) ≥ log n be space constructible. Then DSpace(s) is

closed under complement, i.e., if L ∈ DSpace(s) so is L̄.

Proof. Let M be an s space bounded deterministic Turing machine for

L. We construct a deterministic Turing machine M̄ as follows: M̄ simulates
M step by step. If M halts, then M̄ halts, too, and accepts if M rejects and
vice versa. Problematic is the case when M does not halt. Here M̄ has to
halt and accept. M marks s(n) cells on an extra tape before starting the
simulation and uses it as a c-nary counter to count the number of steps of
M . If M makes more than cs(n) steps, then M̄ stops and accepts.

Remark 18.6 Corollary 18.5 is trivially true for deterministic time classes,
we just have to exchange accepting states with rejecting states.

Exercise 18.4 Show the following: Let M be a nondeterministic Turing

machine that is weakly s space bounded. If M accepts some input x, then
there is an accepting path in the computation tree of M on x that has length
at most cs(|x|) .

We will use the configuration graph to simulate space bounded Turing

machines by time bounded ones and nondeterministic ones by deterministic
ones. The following observation is crucial for the proof.

Observation 18.7 Let G be a graph. If there is a path of length ` from u

to v, then there is a node w such that there are paths from u to w and w to
v of length d`/2e and b`/2c, respectively.

Lemma 18.8 Let M be an s space bounded (deterministic or nondetermin-

istic) Turing machine where s(n) ≥ log n is space constructible. There is
a 2O(s(|x|)) time bounded deterministic Turing machine M1 and O(s2 (|x|))
space bounded deterministic Turing machine M2 that given x decides whether
an accepting configuration in CGM is reachable from SCM (x).

Proof. To achieve the time bound 2O(s(|x|)) , M1 simply generates the

whole graph of all configurations that use space s(|x|). We can enumerate
these configurations, because s is space constructible.2 We can enumerate
all edges in `M , since to find the successor of a configuration, we need time
O(max{|x|, s(|x|)}). Since 2s(|x|) ≥ |x|, this is within the required time
bound. By Lemma 18.3, it is now sufficient to check whether M1 could
2
To enumerate the configurations, we have to encode them as strings. This can be
done by giving the states numbers and writing all numbers in the configuration down in
binary. We separate the components of a configuration by using some new symbol. In
this way, the strings have length ≤ c · s(|x|) for some constant c. We can now enumerate
all strings of length c · s(|x|) for instance in lexicographical order and check whether it is
a valid configuration.

c Markus Bläser 2007–2015
18.3. Space versus time 127

reach an accepting configuration from SC(x). This can be done by your

favourite connectivity algorithm.
The real fun is to do this in space O(s2 ), i.e., to construct M2 . To achieve
this, define
(
1 if C 0 can be reached from C with ≤ ` steps,
R(C, C 0 , `) =
0 otherwise.
If we can compute R(C, C 0 , `) for every configuration C and C 0 and ` ≤
cs(n) in time 2O(s(|x|)) and space O(s2 (|x|)), then we are done. Here c is
the constant from Lemma 18.3. We enumerate all accepting configurations
C—there are at most cs(|n|) —and compute R(SC(x), C, cs(n) ). We accept iff
at least one of these values is one. Note that we can reuse the space when
computing R(SC(x), C, cs(n) ) for the next C.
We will compute R(C, C 0 , `) recursively. We use the identity
R(C, C 0 , `) = 1 ⇐⇒ there is a configuration C 00 such that
R(C, C 00 , d`/2e) = R(C 00 , C 0 , d`/2e) = 1
or C `M C 0 .
This suggests the following recursive approach. Enumerate all configurations
C 00 , one at a time. Then compute R(C, C 00 , d`/2e). If this value is 1, then
also compute R(C 00 , C 0 , d`/2e). If this is 1, then we are done. If one of the
two values is zero, then we try the next C 00 . If we tried all C 00 without
success, then M2 rejects.
Let S(`) denote the maximum space needed to compute R(C, C 00 , `) for
any C, C 00 . We have
S(`) ≤ O(s(|x|)) + S(d`/2e),
S(1) ≤ O(s(|x|)).
To see this note that we need O(s(|x|)) space to write down C 00 and then
we need S(d`/2e) space to compute R(C, C 00 , d`/2e) and R(C 00 , C 0 , d`/2e),
since we can use the same cells twice. Therefore S(`) = O(s(|x|) · log `). In
particular, S(cs(|x|) ) = O(s2 (|x|)).

18.3 Space versus time

As a first application, we show that a space bounded Turing machines can
be simulated by time bounded ones with an exponential loss.
Theorem 18.9 Let s(n) ≥ log(n) be space constructible. Then
DSpace(s) ⊆ NSpace(s) ⊆ DTime(2O(s) ).
Proof. The first inclusion is trivial. The second inclusion follows from
Lemma 18.8.

c Markus Bläser 2007–2015
128 18. Space versus time, nondeterminism versus determinism

18.4 Nondeterminism versus determinism

Theorem 18.10 Let t be time constructible. Then

NTime(t) ⊆ NSpace(t) ⊆ DTime(2O(t) ).

Proof. The first inclusion is trivial, since a t time bounded Turing

machine can use at most t(|x|) cells. The second inclusion follows from
Lemma 18.8.

Theorem 18.11 (Savitch) Let s(n) ≥ log n be space constructible. Then

NSpace(s) ⊆ DSpace(O(s2 )).

Proof. Again, this follows from Lemma 18.8.

c Markus Bläser 2007–2015
19 Space and time hierarchies

Hierarchies

Is more space more power? Is more time more power?

The answer is “yes” provided that the space and time bounds behave
well, that is, they shall be constructible.
In the case of time “more” means “somewhat more” and not just “more”
(see Theorem 17.8).

19.1 A technical lemma

Lemma 19.1 Let s1 , s2 , t1 , t2 : N → N with s1 = o(s2 ) and t1 = o(t2 ).
Assume that s2 (n) ≥ log n and t2 (n) ≥ (1 + )n for all n and some  > 0.
Let s2 be space constructible and t2 be time constructible.

1. There is a deterministic Turing machine C1 that is s2 space bounded

such that for every s1 space bounded 1-tape Turing machine M , L(C1 ) 6=
L(M ).

2. There is a deterministic Turing machine C2 that is t2 time bounded

such that for every t1 time bounded 1-tape Turing machine M , L(C2 ) 6=
L(M ).

Proof overview: The proof is by diagonalization. When we constructed a

function that is not WHILE (or Turing) computable, we just constructed a
function whose value on i was different from fi (i). Here the function that
we construct has to be computable in space s2 and time t2 , respectively,
which complicates the construction somewhat. We use the universal Turing
machine to compute the values fi (i) and then diagonalize.

Proof. Let g be the Gödel number of some 1-tape Turing machine M .

It is easy to modify the universal Turing machine in such a way that it can
simulate a t-time bounded 1-tape Turing machine in time O(|g| · t(n)) on
inputs of length n:

• We use one tape to simulate M . The ith symbol of the work alphabet
is represented by the string bin(i) and the symbols are separated by
#.

129
130 19. Space and time hierarchies

(Remember that the tape alphabet of M might be much larger than

the alphabet of U .)

• We do not need to mark the position of the head of M on this tape,

since we can simply use the head of U on this tape.

• The encoding g of M stands on a second tape of U , the current state

of M , stored as a number in binary, stands on the third tape of M .

• Now U simulates one step of M as follows. It looks in g for the entry

that corresponds to the current state of M and the current symbol of
M on the first tape. Then it replaces the state and the symbol and
moves the head accordingly. In one such step, we have to read g once,
hence simulating one step of M can be done by O(g) steps of U .
Note that U is O(|g| · s(n)) space bounded, when M is s space bounded. If
M has an extra input tape, then U also has an extra input tape. (But this
was even true for UTM .)
We use U to construct C1 :

Input: x ∈ {0, 1}∗ , interpreted as [g, y] with g ∈ im gödTM .

1. If x does not have the above form, then reject.

2. Mark s2 (|x|) symbols to the left and right of cell 0 on the first
tape.

3. Simulate M := göd−1
TM (g) on x on the first tape (using the
machine U ).

4. On an extra tape, count the number of simulated steps.

5. If the simulation ever leaves the marked cells, then stop and
reject.

6. If more than 3s2 (|x|) step are simulated, then stop and accept.
(We can count up to this value by marking s2 (|x|) cells and
counting in ternary.)

7. Accept, if M rejects. Otherwise, reject.

Now let M be a s1 space bounded 1-tape Turing machine. We claim that

L(M ) 6= L(C1 ). Let g be the Gödel number of M and let x = [g, z] for some
sufficiently long z.
First assume that x ∈ L(C1 ). We will show that in this case, x ∈ /
L(M ), which means that L(M ) 6= L(C1 ). If x ∈ L(C1 ), then C1 accepts

c Markus Bläser 2007–2015
19.2. Deterministic hierarchy theorems 131

x. This means that either M performed more than 3s2 (|x|) many steps or
M halts on x and rejects. In the second case, we are done. For the first
case, note that there is a constant c such that M cannot make more than
cs1 (|x|) · (s1 (|x|) + 2) · (|x| + 2) steps without entering an infinite loop.1 Thus
if 3s2 (|x|) > cs1 (|x|) · (s1 (|x|) + 2) · (|x| + 2) then we get that x ∈
/ L(M ). But
3 s 2 (|x|) >c s 1 (|x|) · (s1 (|x|) + 2) · (|x| + 2) is equivalent to

log 3 · s2 (|x|) > log c · s1 (|x|) + log(s1 (|x|) + 2) + log(|x| + 2)

This is fulfilled by assumption for all long enough x, i.e., for long enough z,
because s2 (|x|) ≥ log(|x|).
The second case is x ∈/ L(C1 ). We will show that now M accepts x. Note
that C1 always terminates. If C1 rejects y, then M ran out of space or M
halted and accepted. The second case, x ∈ L(M ) and we are done. We will
next show that the first case cannot happen. Since M is s1 space bounded,
the simulation via U needs space |g| · s1 (|x|). But |g| · s1 (|x|) ≤ s2 (|x|) for
sufficiently large |x|. Thus this case cannot happen.
C1 is s2 space bounded by construction. This proves the theorem.
The construction of C2 is similar, even easier. We do not have to check
whether M runs out of space. We do not need to count to 3s2 (|x|) to detect
infinite loops. Instead we count the number of steps made by C2 . If more
then t2 (|x|) step are made, then we stop and reject. In this way, C2 becomes
O(t2 ) time bounded. (To get down to t2 , use acceleration.) Since we can
simulate one step of M by |g| steps, the simulation of M takes |g| · t1 (|x|)
steps of C2 provided that M is t1 time bounded. This is less than t2 (|x|) if
z is long enough. The rest of the proof is similar.

19.2 Deterministic hierarchy theorems

Theorem 19.2 (Deterministic space hierarchy) Let s2 (n) ≥ log n be
space constructible and s1 (n) = o(s2 (n)). Then

DSpace(s1 ) ( DSpace(s2 ).

Proof. Consider C1 from Lemma 19.1. L(C1 ) ∈ DSpace(s2 ). There

is no s1 space bounded deterministic 1-tape Turing machine M such that
L(M ) = L(C1 ). But for every s1 space bounded deterministic k-tape Turing
machine N there is a s1 space bounded deterministic 1-tape Turing machine
N 0 with L(N 0 ) = L(N ) by Theorem 17.8. Thus there is also no s1 space
bounded deterministic k-tape Turing machine M such that L(M ) = L(C1 ).

1
We cannot bound |x| + 2 by ds1 (|x|) , since s1 might be sublogarithmic.

c Markus Bläser 2007–2015
132 19. Space and time hierarchies

Next, we do the same for time complexity classes. The result will not be
as nice as for space complexity, since we cannot simulate arbitrary determin-
istic Turing machines by 1-tape Turing machines without any slowdown.

Theorem 19.3 (Deterministic time hierarchy) Let t2 be time constructible

and t21 = o(t2 ). Then
DTime(t1 ) ( DTime(t2 ).

Proof. Consider C2 from Lemma 19.1. L(C2 ) ∈ DTime(t2 ). There

is no t21 time bounded deterministic 1-tape Turing machine M such that
L(M ) = L(C2 ). But for every t1 space bounded deterministic k-tape Turing
machine N there is a t21 time bounded deterministic 1-tape Turing machine
N 0 with L(N 0 ) = L(N ) by Theorem 17.8. Thus there is also no t1 time
bounded deterministic k-tape Turing machine M such that L(M ) = L(C1 ).

19.3 Remarks
The assumption t21 = o(t2 ) in the proof of the time hierarchy theorem is
needed, since we incurr a quadractic slowdown when simulating k-tape Tur-
ing machines by 1-tape Turing machines.
Hennie and Stearns showed the following theorem.

Theorem 19.4 (Hennie & Stearns) Every t time and s space bounded
deterministic k-tape Turing machine can be simulated by an O(t log t) time
bounded and O(s) space bounded deterministic 2-tape Turing machine.

We do not give a proof here. Using this theorem, we let C2 diagonalizes

against 2-tape Turing machines instead of 1-tape Turing machines. This
gives the following stronger version of the time hierarchy theorem.

Theorem 19.5 Let t2 be time constructible and t1 log t1 = o(t2 ). Then

DTime(t1 ) ( DTime(t2 ).

The answer to the following question is not known.2

Research Problem 19.1 Can the assumption t1 log t1 = o(t2 ) be further

weakened? In particular, can we get a better simulation of arbitrary deter-
ministirc Turing machines on Turing machines with a fixed number of tapes?

If the number of tapes is fixed, then one can obtain a tight time hierarchy.
Again we do not give a proof here.
2
If you can answer it, we should talk about your dissertation.

c Markus Bläser 2007–2015
19.3. Remarks 133

Theorem 19.6 (Fürer) Let k ≥ 2, t2 time constructible, and t1 = o(t2 ).

Then
DTimek (t1 ) ( DTimek (t2 ).

We conclude with pointing out that the assumption that s2 and t2 are
constructible are really necessary.

Theorem 19.7 (Borodin’s gap theorem) Let f be a total recursive func-

tion N → N with f (n) ≥ n for all n. Then there are total recursive functions
s, t : N → N with s(n) ≥ n and t(n) ≥ n for all n such that

DTime(f (t(n))) = DTime(t(n)),

DSpace(f (s(n))) = DSpace(s(n)).

Proof. We only construct t, the construction of s is similar. Let Tg (n) :=

Timegöd−1 (g) (n) be the maximum running time of the Turing machine with
TM
Gödel number g on inputs of length n.
We first show: For all n ∈ N, there is an m ∈ N such that for all Gödel
numbers g with cod(g) ≤ n,

Tg (n) ≤ f (m) ⇒ Tg (n) ≤ m.

Let m0 = n and mi+1 = f (mi ) + 1 für 1 ≤ i ≤ n + 1. The n + 2 intervals

[mi , f (mi )] are pairwise disjoint, because f (n) ≥ n for all n. Therefore,
there is an i0 such that Tg (n) ∈ / [mi0 , f (mi0 )] for all g with cod(g) ≤ n. Set
m = mi0 .
Let t(n) be the m defined above corresponding to n. t is recursive:
We can compute the intervals, since f is total and recursive. We can test
Tg (n) ∈ / [mi , f (mi )] by simulating göd−1 (g) on all inputs of length n. Since
each of these simulation can be stopped after f (mi ) steps, this is decidable.
Now let M be f (t(n)) time bounded, i.e., Tg (n) ≤ f (t(n)) for all n, where
g = gödTM (M ). By the construction of t, Tg (n) ≤ t(n) for all n ≥ cod(g).
Therefore, L(M ) ∈ DTime(t). Thus, DTime(f ◦ t) = DTime(t).
n
Set for instance g(n) = 2n (or 22 or . . . ) and think for a minute how
unnatural non-constructible time or space bounds are.

Excursus: Nondeterministic hierarchies

For nondeterministic space, we can use Savitch’s theorem to show the following:

NSpace(s1 ) ⊆ DSpace(s21 ) ( DSpace(s22 ) ⊆ NSpace(s22 )

for any functions s1 , s2 : N → N with s1 = o(s2 ), s1 and s2 space constructible, and

s1 (n) ≥ log n. It is even possible to show a tight hierarchy like in Theorem 19.2.
This uses the non-trivial—and unexpected—fact that NSpace(s) is closed under
complementation, the so-called Immerman–Szelepcsényi Theorem.

c Markus Bläser 2007–2015
134 19. Space and time hierarchies

For nondeterministic time, neither of the two approaches is known to work. But
one can get the following hierarchy result: For a function t : N → N, let t̃ be the
function defined by t̃(n) = t(n + 1). If t1 is time constructible and t̃1 = o(t2 ) then

NTime(t2 ) \ NTime(t1 ) 6= ∅.

The proof of this result is lengthy. Note that for polynomial functions or exponential
functions, t̃1 = O(t1 ). Thus we get a tight nondeterministic time hierarchy for these
functions.

c Markus Bläser 2007–2015
20 P and NP

We are looking for complexity classes, that are robust in the sense that
“reasonable” changes to the machine model should not change the class.
Furthermore, the classes should also characterize interesting problems.

Definition 20.1
[
P= DTime(O(ni ))
i∈N
[
NP = NTime(O(ni ))
i∈N

P (P stands for polynomial time) is the class of problems that are con-
sidered to be feasible or tractable. Frankly, an algorithm with running time
O(n1024 ) is not feasible in practice, but the definition above has been very
fruitful. If a natural problem turns out to be in P, then we usually will
have an algorithm whose running time has a low exponent. In this sense, P
contains all languages that we can decide quickly.
NP (NP stands for nondeterministic polynomial time and not for non-
polynomial time) on the other hand, is a class of languages that we would
like to decide quickly. There are thousands of interesting and important
problems in NP for which we do not know deterministic polynomial time
algorithms.
The class P is a robust class. A language that can be decided by a
deterministic Turing machine in polynomial time can be decided by a WHILE
program in polynomial time and vice versa. (This follows easily by inspecting
the simulations that we designed in the first part of the lecture. But read
the excursus in Chapter 16.) This is also true for NP, if we equip WHILE
programs with nondeterminism in a suitable way.
The question whether P = NP is one of the big open problems in com-
puter science. Most researchers believe that these classes are different, but
there is no valid proof so far. The best that we can show is
i
[ [
NP = NTime(O(ni )) ⊆ DTime(2O(n ) ) =: EXP,
i∈N i∈N

that is, nondeterministic polynomially time bounded Turing machines can

be simulated by deterministic poly-exponential time bounded ones.

135
136 20. P and NP

Excursus: The millenium prize problems

The question whether P equals NP is one of the seven millenium prize problems of
the Clay mathematics institute. (www.claymath.org). If you settle this question,
you get $1000000 (and become famous, at least as famous as a computer scientist
can become).
Gerhard Woeginger’s P-versus-NP webpage (www.win.tue.nl/∼gwoegi/P-versus-NP.htm)
keeps track of the outgrowth.

20.1 Problems in P
Here is one important problem in P. You may consult your favourite book
on algorithms for many other ones.
s-t-CONN is the problem whether a directed graph has a path from a
given source node s to a target node t:

s-t-CONN = {(G, s, t) | G is a directed graph

that has a directed path from s to t}.

(G, s, t) is an encoding of the graph G and the source and target nodes s
and t. A reasonable encoding would be the following: All nodes are rep-
resented by numbers 1, . . . , n, written down in binary. We encode an edge
by [bin(i), bin(j)]. We encode the whole graph by building a large pair that
consists of bin(n), bin(s), bin(t), and the encodings of all edges, using our
pairing function. Since we only talk about polynomial time computability,
the concrete encoding does not matter, and we will not specify the encoding
in the following.
We will also just write (G, s, t) or G and will not apply an encoding
function. You are now old enough to distinguish whether we mean the graph
G itself or its encoding.

Theorem 20.2 s-t-CONN ∈ P.

20.2 NP and certificates

Beside the definition of NP above, there is an equivalent one based on veri-
fiers. We call a Turing machine a polynomial time Turing machine if it is p
time bounded for some polynomial p.

Definition 20.3 A deterministic polynomial time Turing machine M is

called a polynomial time verifier for L ⊆ {0, 1}∗ , if there is a polynomial
p such that the following holds:
1. For all x ∈ L there is a c ∈ {0, 1}∗ with |c| ≤ p(|x|) such that M
accepts [x, c].

c Markus Bläser 2007–2015
20.3. Problems in NP 137

2. For all x ∈/ L and all c ∈ {0, 1}∗ , M on input [x, c] reads at most p(|x|)
bits of c and always rejects [x, c].
We denote the language L that M verifies by V (M ).

The string c serves as a certificate (or witness or proof ) that x is in L. A

language L is verifiable in polynomial time if each x in L has a polynomially
long proof that x ∈ L. For each x not in L no such proof exists.
Note that the language V (M ) that a verifier verifies is not the language
that it accepts as a “normal” Turing machine. L(M ) can be viewed as a
binary relation, the pairs of all (x, c) such that M accepts [x, c].

Theorem 20.4 L ∈ NP iff there is a polynomial time verifier for L.

Proof. We only prove the “=⇒”-direction. Since L is in NP there is a

nondeterministic Turing machine M whose time complexity is bounded by
some polynomial p such that L(M ) = L. We may assume w.l.o.g. that
in each step, M has at most two nondeterministic choices. We construct a
polynomial time verifier V for L. V has one more tape than M . Let [x, c]
be the input for V . On the additional tape, V marks p(|x|) cells. (This is
possible in polynomial time, since p is time constructible.) Then it copies
the first p(|x|) symbols of c into these marked cells. Now V simulates M
on x step by step. In each simulated step, it reads one of the bits of c on
the additional tape. If M has a nondeterministic choice, V uses the bit of c
read in this step to chose one of the two possibilites M has. (To this aim,
we order the tuples in the relation δ arbitrarily. If the bit read is 0, then we
take the choice which appears before the other one in this ordering. If the
bit read is 1, we take the other choice.) In this way, c specifies one path in
the computation tree T of M on x. Now if x ∈ L, then there is one path
in T of length at most p(|x|) that is accepting. Let c be the bit string that
corresponds to this path. Then V accepts [x, c]. If x ∈ / L, then no such path
exists and hence V will not accept [x, c] for any c. Cleary, the running time
of V is bounded by O(p(|x|)).

Exercise 20.1 Prove the other direction of Theorem 20.4

20.3 Problems in NP
There is an abundance of problems in NP. We here just cover the most basic
ones (most likely, even less).
A clique of a graph G = (V, E) is a subset C of V such that for all
u, v ∈ C with u 6= v, {u, v} ∈ E. A clique C is called a k-clique if |C| = k.
Clique is the following language:

Clique = {(G, k) | G is an undirected graph with a k-clique}.

c Markus Bläser 2007–2015
138 20. P and NP

A vertex cover of a graph G = (V, E) is a subset C of V such that for

each edge e ∈ E, e ∩ C 6= ∅. (Recall that edges of an undirected graph are
sets of size two. Thus this condition means that every edge is covered by at
least one vertex in C.) VC is the following problem:
VC = {(G, k) | G is an undirected graph
that has a vertex cover of size ≤ k}.
Subset-Sum is the following problem:
Subset-Sum = {(x1 , . . . , xn , b) | x1 , . . . , xn , b ∈ N and there is an
X
I ⊆ {1, . . . , n} with xi = b.}
i∈I

Let G = (V, E) be a graph and V = {v1 , . . . , vn }. G has a Hamiltonian

cycle if there is a permutation π such that for all 1 ≤ i < n, {vπ(i) , vπ(i+1) } ∈
E and {vπ(n) , vπ(1) } ∈ E, i.e., there is a cycle that visits each vertex of V
exactly once. HC is the following problem:
HC = {G | G has a Hamiltonian cycle}.
 
Next we consider a weighted complete graph G = (V, V2 , w) where
   
V V
2 denotes all subsets of V of size 2 and w : 2 → N assigns to each
edge a nonnegative weight. The P weight of a Hamiltonian cycle is the weight
of the edges contained in it, i.e., n−1
i=1 w({vπ(i) , vπ(i+1) }) + w({vπ(n) , vπ(1) }).
The traveling salesman problem is the following problem:
TSP = {(G, b) | G is a weighted graph with
a Hamiltonian cycle of weight ≤ b}.
You can think of a truck that has to deliver goods to different shops and we
want to know whether a short tour exists.
Let x1 , . . . , xn be Boolean variables, i.e., variables that can take values
0 and 1, interpreted as false and true. A literal is either a variable xi or
its negation x̄i . A clause is a disjunction of literals `1 ∨ · · · ∨ `k . k is the
length of the clause. A formula in conjunctive normal form (formula in
CNF for short) is a conjunction of clauses c1 ∧ · · · ∧ cm . An assignment a is
a mapping that assigns each variable a value in {0, 1}. Such an assignment
extends to literals, clauses, and formulas in the obvious way. A formula
is called satisfiable, if there is an assignment such that the formula attains
the value 1. Such an assignment is called a satisfying assignment. If all
assignments are satisfying, then the formula is called a tautology.
The satisfiability problem, the mother of all problems in NP, is the fol-
lowing problem:
SAT = {φ | φ is a satisfiable formula in CNF}.

c Markus Bläser 2007–2015
20.3. Problems in NP 139

A formula in CNF is in `-CNF, if all its clauses have length at most `.

`SAT is the following problem:

`SAT = {φ | φ is a satisfiable formula in `-CNF}.

Golden rule of nondeterminism

Nondeterminism is interesting
because it characterizes important problems.

We do not know any physical equivalent to nondeterminism. As far as I

know, nobody has been built a nondeterministic Turing machine. But NP
is an interesting class because it contains a lot of important problems.

Theorem 20.5 Clique, VC, Subset-Sum, HC, TSP, SAT, `SAT ∈ NP.

Proof. We show that all of the problems have a polynomial time verifier.
Let’s start with Clique. On input [x, y], a verifier M for Clique first checks
whether x is an encoding of the form (G, k). If not, M rejects. It now
interprets the string y as a list of k nodes of G, for instance such an encoding
could be bin(i1 )$ . . . $ bin(ik ) where i1 , . . . , ik are nodes of G. (Since y ∈
{0, 1}∗ , we would then map for instance 0 7→ 00, 1 7→ 01, and $ 7→ 11.) If
y is not of this form, then M rejects. If y has this form, then M checks
whether {ij , ih } is an edge of G, 1 ≤ j < h ≤ k. If yes, then M accepts,
otherwise it rejects.
We have to show that there is a y such that [x, y] ∈ L(M ) iff x = (G, k)
for some graph G that has a k-clique. Assume that x = (G, k) for some
graph G that has a k-clique. Then a list of the nodes that form a clique is
a proof that makes M accept. On the other hand, if G has no k-clique or x
is not a valid encoding, then no proof will make M accept.
For SAT and `SAT, an assignment to the variables that satisfies the
formula is a possible proof. For VC, a subset of the nodes of size less than
k that covers all edges is a possible proof. For Subset-Sum, it is the set of
indices I, for HC and TSP, it is the appropriate permutation. The rest of
the proof is now an easy exercise.

c Markus Bläser 2007–2015
140 20. P and NP

Decision versus Verification

Cum grano salis:

P is the class of languages L for which we can efficiently decide
“x ∈ L?”.
NP is the class of languages L for which we can efficiently verify
whether a proof for “x ∈ L!” is correct or not.

c Markus Bläser 2007–2015
21 Reduction and completeness

NP contains a lot of interesting problems for which we would like to have

efficient algorithms. But most researchers believe that the classes P and NP
do not coincide, that is, there is a language L ∈ NP such that L ∈ / P. But we
are far from having a proof for this. Instead, we try to identify the “hardest”
languages in NP, the so-called NP-complete languages. If we can show that
a problem is NP-complete, then this is a strong indication that it does not
have a deterministic polynomial time algorithm.

21.1 Polynomial time reductions

Definition 21.1 Let L, L0 ⊆ Σ∗ .

1. A function f : Σ∗ → Σ∗ is called a many-one polynomial time reduc-

tion from L to L0 if f is polynomial time computable and

for all x ∈ Σ∗ : x ∈ L ⇐⇒ f (x) ∈ L0 .

2. L is (many-one) polynomial time reducible to L0 if there is a many-one

polynomial time reduction from L to L0 . We denote this by L ≤P L0 .

Compared to recursive many-one reductions, the function f now shall be

polynomial time computable. The reason is that f shall preserve polynomial
time computability.

Lemma 21.2 If L ≤P L0 and L0 ∈ P, then L ∈ P.

Proof. Let f be a polynomial time reduction from L to L0 . Let M 0

be a polynomial time deterministic Turing machine with L0 = L(M 0 ). We
construct a polynomial time deterministic Turing machine M for L as follows:
On input x, M first computes f (x) and then simulates M 0 on f (x). M
accepts if M 0 accepts, and rejects otherwise.
First of all, L(M ) = L, because x ∈ L ⇐⇒ f (x) ∈ L0 . It remains to
show that M is indeed polynomial time bounded. To see this, assume that
f is p(n) time computable and M 0 is q(n) time bounded for polynomials p
and q. Since f is p(n) time computable, |f (x)| ≤ p(|x|) for all x ∈ Σ∗ . Thus
M is q(p(n)) time bounded which is again a polynomial.

Lemma 21.3 ≤P is a transitive relation.

141
142 21. Reduction and completeness

Proof. Let L ≤P L0 and L0 ≤P L00 . Let f and g be corresponding

reductions. We have to show that L ≤P L00 . We claim that g ◦ f is a
polynomial time reduction from L to L00 .
First of all, we have for all x ∈ Σ∗ :

x ∈ L ⇐⇒ f (x) ∈ L0 ⇐⇒ g(f (x)) ∈ L00 .

It remains to show that g ◦ f is polynomial time computable. This is shown

as in the proof of Lemma 21.2.

Polynomial time many one reductions versus

recursive many one reductions

A many one reduction f from L to L0 has the following property:

x ∈ L ⇐⇒ f (x) ∈ L0 for all x ∈ {0, 1}∗ .

Recursive many one reduction:

• f is Turing computable.

• f is total.

Polynomial time many one reduction:

• f is polynomial time computable. (This implies that f is total.)

Important properties of recursive many one reducibility:

• ≤ is transitive.

• If L ≤ L0 and L0 ∈ REC (or RE), then L ∈ REC (or RE)

Important properties of polynomial time many one reductions:

• ≤P is transitive.

• If L ≤P L0 and L0 ∈ P, then L ∈ P.

21.2 NP-complete problems

Definition 21.4 1. A language L is NP-hard if for all L0 ∈ NP, L0 ≤P L.
2. L is called NP-complete, if L is NP-hard and L ∈ NP.

Lemma 21.5 If L is NP-hard and L ∈ P, then P = NP.

c Markus Bläser 2007–2015
21.2. NP-complete problems 143

Proof. Let L0 ∈ NP. Since L is NP-hard, L0 ≤P L. By Lemma 21.2,

L0 ∈ P.
How can we show that a language is NP-hard? Once we have identified
one NP-hard language, the following lemma provides a way to do so.

Lemma 21.6 If L is NP-hard and L ≤P L0 , then L0 is NP-hard.

Proof. Let L00 ∈ NP. Since L is NP-hard, L00 ≤P L. Since ≤P is transitive,

L00 ≤P L0 . Thus L0 is NP-hard, too.
It is the famous Cook–Karp–Levin theorem that provides a first NP-
complete problem. We defer the proof of it to the next chapters.

Theorem 21.7 (Cook–Karp–Levin) 1 SAT is NP-complete.

Excursus: Cook, Karp, or Levin?

In his original paper, Steve Cook did not talk about satisfiability at all, he always
talked about tautologies and showed that this problem was NP-complete. This
problem is co-NP-complete and it is a big open question whether it is also NP-
complete. So how can Steve Cook talk about NP and tautologies? The reason
is that he uses a coarser kind of reductions, Turing reductions, instead of many-
one reductions. But essentially all his Turing reductions are many-one. This was
pointed out by Richard Karp who also showed that many other problems are NP-
complete under many-one reductions.
Leonid Levin, a poor guy from the former Soviet Union, invented at the same
time as Steve Cook and Richard Karp a similar theory of NP-completeness. Since
the cold war was really cold at this time, western scientists became aware of his
findings more than a decade later. (He also did not get a Turing award.)

It is rather easy to show that our problem AFSAT is NP-complete.

Lemma 21.8 SAT ≤ AFSAT.

Proof. Let φ be a Boolean formula in CNF with n variables. We construct

an arithmetic formula Fφ such that every satisfying assignment a ∈ {0, 1}n of
φ is an assignment such that a(F ) = 1 and every non-satisfying assignment
is an assignment such that a(F ) = 0. (Note that we interpret 0 and 1 as
Boolean values and as integers.)
We construct this formula along the structure of formulas in CNF. Let `
be a literal. If ` = x, then F` = x. If ` = x̄, then F` = 1 − x.
If c = `1 ∨ · · · ∨ `k is a clause, then Fc = 1 − (1 − F`1 ) · · · (1 − F`k ). Fc
evaluates to 1 iff one of the F`i evaluates to 1.
1
This theorem is usually called Cook’s theorem. Less conservative authors call it Cook–
Levin theorem. The name is Cook–Karp–Levin theorem is my creation, use it at your own
risk.

c Markus Bläser 2007–2015
144 21. Reduction and completeness

Finally, if φ = c1 ∧ · · · ∧ cm is a conjunction of clauses, then Fφ =

Fc1 · · · Fcm . Fφ evaluates to 1 if all Fci evaluate to 1.
Thus, φ 7→ Fφ is the desired reduction. It is easy to see that this mapping
is polynomial time computable.
Let’s start to show that the other problems introduced in this chapter
are all NP-complete.

Lemma 21.9 For all ` ≥ 3, `SAT is NP-complete. 2

Proof. We show that SAT ≤ `SAT. It suffices to show this for ` = 3. Let
φ be a formula in CNF. We have to map φ to a formula ψ in 3-CNF such
that φ is satisfiable iff ψ is satisfiable.
We replace each clause c of length > 3 of φ by a bunch of new clauses.
Let c = `1 ∨ · · · ∨ `k with literals `κ . Let y1 , . . . , yk−3 be new variables. (We
need new variables for each clause.) We replace c by

(`1 ∨`2 ∨y1 )∧(ȳ1 ∨`3 ∨y2 )∧· · ·∧(ȳk−4 ∨`k−2 ∨yk−3 )∧(ȳk−3 ∨`k−1 ∨`k ) (21.1)

If we do this for every clause of φ, we get the formula ψ. This transformation

is obviously polynomial time computable.
It remains to show that φ is satisfiable iff ψ is satisfiable. Let a be
some satisfying assignment for φ. We extend this assignment to a satisfying
assignment of ψ. Since a satisfies φ, for a given clause c, there is one literal,
say `i such that a(`i ) = 1. If we assign to all yj with j < i−1 the value 1 and
to all other yj the value 0, then all clauses in (21.1) are satisfied. Thus we
found a satisfying assignment for ψ. On the other hand, if ψ is satisfiable,
then any satisfying assignment b that satisfies all the clauses in (21.1) has
to set one `i to 1, since the yj ’s can only satisfy at most k − 3 clauses but
(21.1) contains k − 2 clauses. Thus the restriction of b to the variables of φ
satisfies c and henceforth φ.

Exercise 21.1 Astonishingly, 2SAT ∈ P.

1. Given a formula φ in 2-CNF over variables x1 , . . . , xn , we construct a

directed graph G = (V, E) as follows: V = {x1 , . . . , xn , x̄1 , . . . , x̄n } is
the set of all literals and for each clause `1 ∨ `2 we add the two edges
(`¯1 , `2 ) and (`¯2 , `1 ) to E. Show the following: φ is satisfiable iff for all
1 ≤ ν ≤ n, there is no directed cycle that contains xν and x̄ν .

2. Conclude that 2SAT ∈ P.

2
Our proof of the Cook-Karp-Levin theorem will actually show that 3SAT is NP-
complete. It is nevertheless very instructive to see the reduction from SAT to 3SAT.

c Markus Bläser 2007–2015
21.2. NP-complete problems 145

Reducing SAT to `SAT was not too hard. (Or at least, it does not look
too unreasonable that one can find such a reduction.) Reducing SAT or `SAT
to Clique, for instance, looks much harder, since these problems seem to be
completely unrelated. First such reductions look like art, but nowadays it
has become routine work (with some exceptions) and there is a huge toolbox
available.

Lemma 21.10 `SAT ≤P Clique.

Proof. Let φ be a formula in 3-CNF. We may assume that each clause of

φ has exactly three literals by possibly repeating some literals. Let

φ = (`1,1 ∨ `1,2 ∨ `1,3 ) ∧ · · · ∧ (`m,1 ∨ `m,2 ∨ `m,3 )

We have to construct a pair (G, k) such that G = (V, E) has a k-clique iff φ
is satisfiable. We set V = {(1, 1), (1, 2), (1, 3), . . . , (m, 1), (m, 2), (m, 3)}, one
node for each literal of a clause. E is the set of all {(i, s), (j, t)} such that
i 6= j and `i,s 6= `¯j,t . In other words, there is no edge {(i, s), (j, t)} iff `i,s
and `j,t cannot be simultaneously set to 1 (because one is the negation of
the other). Finally, we set k = m.
If φ is satisfiable, then there is a satisfying assignment for φ, i.e., an
assignment that assigns to at least one literal of each clause the value 1. Let
`1,s1 , . . . , `m,sm be these literals. Then (1, s1 ), . . . , (m, sm ) form a clique of
size m in G.
Conversely, if G has a clique of size k, then it is of the form (1, s1 ), . . . , (m, sm ),
because there is no edge between (i, s) and (i, t) for s 6= t. Then we can set
all the literals `1,s1 , . . . , `m,sm to 1 and hence φ is satisfiable.
The mapping φ 7→ (G, k), is obviously polynomial time computable.

Lemma 21.11 Clique ≤ VC.

 
V
Proof. For a graph G = (V, E), let Ḡ = (V, 2 \ E) be its complement,
i.e, e is an edge of G iff e is not an edge of Ḡ.
Let C be a clique of G. We will show below that V \ C is an vertex cover
of Ḡ. Conversely, if D is a vertex cover of Ḡ, then V \ D is a clique of G.
In particular, G has a clique of size at least k iff Ḡ has an vertex cover of
size at most n − k. Thus (G, k) 7→ (Ḡ, n − k) is the desired reduction. This
reduction is of course polynomial time computable.
If C is a clique of G, then there are no edges between nodes of C in Ḡ.
Thus the nodes of V \ C cover all edges of Ḡ, since every edge in Ḡ has at
least one node not in C. Conversely, if D is a vertex cover of Ḡ, then there
are no edges between then nodes of V \ D, because otherwise D would not
be a vertex cover. Thus V \ D is a clique in G.

c Markus Bläser 2007–2015
146 21. Reduction and completeness

Lemma 21.12 3SAT ≤P Subset-Sum.

The proof of this lemma is more complicated; we defer it to the next

chapter.

Exercise 21.2 Consider the following dynamic programming approach to

Subset-Sum. Let x1 , . . . , xn , b be the given instance. We define a predicate
P (ν, β) for 1 ≤ ν ≤ n and 0 ≤ β ≤ b by
( P
1 if there is an I ⊆ {1, . . . , ν} with i∈I xi = β
P (ν, β) =
0 otherwise

1. Show that P (ν, β) = P (ν − 1, β) ∨ P (ν − 1, β − xν ).

2. Design an algorithm with running time O(nb) for Subset-Sum.

3. Does this show that P = NP? (See also Chapter ??.)

Lemma 21.13 3SAT ≤P HC.

This proof is again deferred to the next section.

Lemma 21.14 HC ≤P TSP.

Proof. Let G = (V, E)bean input of HC. An input of TSP is a weighted

complete graph H = (V, V2 , w). We assign to edges from G weight 1 and
to “nonedges” weight 2, i.e,
(
1 if e ∈ E
w(e) =
2 if e ∈
/E

By construction, G has a Hamiltonian cycle iff H has a Hamiltonian cycle

of weight n.

Theorem 21.15 Clique, VC, Subset-Sum, HC, TSP, SAT, 3SAT, and
AFSAT are NP-complete.

The proof of the theorem follows from the lemmas in this chapter.

c Markus Bläser 2007–2015
21.2. NP-complete problems 147

AFSAT

SAT HC TSP

3SAT Subset-Sum

Clique VC

Figure 21.1: The reduction scheme. An arrow from A to B means that we

proved A ≤P B.

c Markus Bläser 2007–2015
22 More reductions

Only the result on Subset-Sum was discussed in class. The results on HC are
not relevant for the exam.
In this chapter, we construct the two missing reductions from the last
chapter. They are more complicated than the ones in the last chapter, but
now you should be old enough to understand them. When you see such re-
ductions for the first time, they look like complicated magic, but constructing
them has become a routine job, with some notable exceptions.

22.1 Subset-Sum
We first start with the proof of Lemma 21.12. As Exercise 21.2 suggests, the
instances created by the reduction will use large numbers, that is, numbers
whose size is exponential in the number of clauses of the Boolean formula
(or equivalently, the length of the binary representation will we polynomial
in m).
Proof of Lemma 21.12. Let φ be a formula in 3-CNF. We have to con-
struct an instance of Subset-Sum, P i.e., numbers a1 , . . . , at , b such that there
is a subset I ⊆ {1, . . . , t} with i∈I ai = b iff φ is satisfiable.
Let x1 , . . . , xn be the variables of φ. Let c1 , . . . , cm be the clauses of φ.
For each literal ` we will construct a number a(`) as follows: The number
a(`) is of the form a0 (`) + 10n · a1 (`). The first part a0 (`) is the variable part,
the second part a1 (`) is the clause part. For a variable xν , let cµ1 , . . . , cµsν
be the clauses in which it appears positively, or in other words, cµ1 , . . . , cµsν
are the clauses that contain the literal xν . Then

a(xν ) = 10ν−1 + 10n (10µ1 −1 + · · · + 10µsν −1 ).

For a literal x̄ν , let cµ̄1 , . . . , cµ̄s̄ν are the clauses that contain the literal x̄ν .
Then
a(x̄ν ) = 10ν−1 + 10n (10µ̄1 −1 + · · · + 10µ̄s̄ν −1 ).
Choosing a(xν ) indicates that we set xν to 1. Choosing a(x̄ν ) indicates that
we set x̄ν to 1, i.e., xν to 0. Of course we can set xν either to 1 or to 0. This
means that we shall only be able to select one of a(xν ) and a(x̄ν ). Thus in
the “target number” b = b0 + 10n b1 , be set b0 = 1 + 10 + · · · + 10n−1 .
The numbers a(xν ) and a(x̄ν ) have digits 0 or 1. For each position 10i ,
there are at most 3 numbers that have digit 1 at position 10i . In the variable
part, this is clear, since only a(xi ) and a(x̄i ) have a 1 in position 10i . In

148
22.2. Hamiltonian Cycle 149

the clause part, this is due to the fact that each clause consists of at most
three literals. Since our base 10 is larger than 3, in the sum of any subset
of a(xν ), a(x̄ν ), 1 ≤ ν ≤ n, no carry can occur. (We could have chosen a
smaller base but 10 is so convenient.) This means that any sum that yields
b0 in the lower n digits either contains a(xν ) or a(x̄ν ), 1 ≤ ν ≤ n. This
ensures consistency, that means, we can read off a corresponding assignment
from the chosen numbers.
Finally, we have to ensure that the assignment is also satisfying. This
is done by choosing b1 properly. Each clause should be satisfied, so a first
try would be to set b1 = 1 + 10 + · · · + 10m−1 . But a clause cµ could be
satisfied by two or three literals, in this case the digit of 10n−1+µ is 2 or
3. The problem is that we do not know in advance whether it is 1, 2, or 3.
Therefore, we set b1 = 3(1 + 10 + · · · + 10m−1 ) and introduce “filler numbers”
cµ,1 = cµ,2 = 10n−1+µ , 1 ≤ µ ≤ m. We can use these filler numbers to reach
the digit 3 in position 10n−1+µ . But to reach 3, at least one 1 has to come
from an a(xν ); thus the clause is satisfied if we reach 3.
Overall, the considerations above show that φ has a satisfying assignment
iff a subset of a(xν ), a(x̄ν ), 1 ≤ ν ≤ n, and cµ,1 , cµ,2 , 1 ≤ µ ≤ m, sums up to
b. Thus the reduction above is a polynomial time many one reduction from
3SAT to Subset-Sum.

22.2 Hamiltonian Cycle

In order to prove Lemma 21.13, we introduce an intermediate problem, di-
rected Hamiltonian cycle. Here we consider directed graphs G = (V, E). The
edges are now ordered pairs, i.e., elements of V × V . We say that (u, v) is
an edge from u to v. Let V = {v1 , . . . , vn }. Now a (directed) Hamiltonian
cycle is a permutation such that (vπ(i) , vπ(i+1) ) ∈ E for all 1 ≤ i < n and
(vπ(n) , vπ(1) ) ∈ E. That is, it is a cycle that visist each node exactly once
and all the edges in the cycle have to point in the same direction.
Dir-HC is a generalization of HC. To each undirected graph H corre-
sponds a directed one, G, in a natural way: Each undirected edge {u, v} is
replaced by two directed ones, (u, v) and (v, u). Any Hamiltonian cycle in
G induces a Hamiltonian cycle in H in a natural way. Given H, G can be
computed easily. Thus HC ≤P Dir-HC. But we can also show the converse.

Lemma 22.1 Dir-HC ≤P HC.

Proof. Let G = (V, E) be a directed graph. We construct a undirected

graph G0 = (V 0 , E 0 ) such that G has a Hamiltonian cycle iff G0 has a Hamil-
tonian cycle.
G0 is obtained from G as follows: For every node v ∈ V , we introduce

c Markus Bläser 2007–2015
150 22. More reductions

u1 w1 u1,out w1,in

u2 w2 u2,out w2,in
.. v .. .. vin v vout ..
. . . .

us wt us,out wt,in

Figure 22.1: The gadget for the reduction: u1 , . . . , us are the nodes such that
there is an edge (ui , v) ∈ E, that is, an edge entering v. The nodes w1 , . . . , wt
are the nodes such that there is an edge (v, wj ) ∈ E, that is, an edge leaving
v. The two lists u1 , . . . , us and v1 , . . . , vt need not be disjoint. The righthand
side show the gadget. Every node is replaced by three nodes, v, vin , vout .
For every directed edge (x, y), we add the undirected edge {xout , yin }.

a1 a2

b1 b2

c1 c2

Figure 22.2: The gadget for the reduction of 3SAT to Dir-HC.

three nodes vin , v, vout and connect vin with v and vout with v.1 Then for
every directed edge (x, y) ∈ E, we add the undirected edge {xout , yin } to E 0 .
Figure 22.1 shows this construction.
Given G, we can construct G0 in polynomial time. Thus what remains to
show is the follow: G has a Hamiltonian cycle iff G0 has a Hamiltonian cycle.
Assume that G has a Hamiltonian cycle C. Then we get a Hamiltonian cycle
of G0 as follows. For every edge (x, y) of C, we take the edge {xout , yin }.
Furthermore We add the {vin , v} and {v, vout }. This gives a Hamiltonian
cycle of G0 .
For the converse direction, observe that every node in a Hamiltonian cycle
is incident with two edges. Since every node v is only incident with two edges
in G0 , the edges {vin , v} and {v, vout } have to be in a Hamiltonian cycle in
G0 . The other edges of the Hamiltonian cycle in G0 induce a Hamiltonian
cycle in G.
For 3SAT, we need a more complicated gadget.
1
Such a thing is usually called a gadget. We replace a node or edge (or something like
this) by a small graph (or something like this). The term gadget is used in an informal
way, there is no formal definition of a gadget.

c Markus Bläser 2007–2015
22.2. Hamiltonian Cycle 151

a1 a2 a1 a2 a1 a2

b1 b2 b1 b2 b1 b2

c1 c2 c1 c2 c1 c2

Figure 22.3: The paths connecting the nodes in S.

a1 a2 a1 a2 a1 a2

b1 b2 b1 b2 b1 b2

c1 c2 c1 c2 c1 c2

Figure 22.4: No matter how we connect a1 with b2 , there are always inner
nodes left that are not covered and cannot be covered by other paths.

Lemma 22.2 Let G be the graph in Figure 22.2.

1. For every nonempty subset S ⊆ {(a1 , a2 ), (b1 , b2 ), (c1 , c2 )}, there are
node disjoint path from s to t for all (s, t) ∈ S such that all inner
nodes of G lie on one of these path.

2. For any other subset T ⊆ {a1 , b1 , c1 } × {a2 , b2 , c2 }, no such paths exist.

Proof. We start with the first part: Figure 22.3 shows these paths in the
case of one, two, or three pairs. Only the number of pairs in S matters, since
the structure of the gadget G is invariant under simultaneous cyclic shifts of
the nodes a1 , b1 , c1 and a2 , b2 , c2 .
For the second part consider any other pair. Since the gadget is invariant
under cyclic shifts, it is enough to consider the pair (a1 , b2 ) and the pair
(a2 , b1 ). Figure 22.4 shows all possibilities how to connect a1 with b2 . In
each case, inner nodes are not covered and it is not possible to cover all of
them with other paths. The other pair is an exercise.

Exercise 22.1 Draw the corresponding figures for the pair (a2 , b1 ).

Lemma 22.3 3SAT ≤P Dir-HC.

Proof. For this proof, we have to do the following: Given a formula φ

in 3-CNF, we have to map it to a directed graph G (depending on φ) such
that φ is satisfiable iff G has a Hamiltonian cycle. Every variable of φ will
be represented by a node, every clause will be represented by a gadget from
Figure 22.2.
Let x1 , . . . , xn be the variables of φ and c1 , . . . , cm be its clauses. We
call the nodes representing the variables x1 , . . . , xn , too. There will be two
paths from xi to xi+1 for each 1 ≤ i < n and two paths from xn to x1 . One

c Markus Bläser 2007–2015
152 22. More reductions

corresponds to the fact that xi is set to 1, the other one corresponds to the
case that xi is set to 0. Let Cj be the gadget that represents cj . Assume
that xi occurs positively in clauses cj1 , . . . , cjs and negatively in the clauses
ck1 , . . . , ckt . Then an edge goes from xi to Cj1 . If it is the first literal in cj1 ,
then this edge goes to the node a1 , if it is the second, then it enters through
b1 , and if it is the third, it uses c1 . Then there is an edge from Cj1 to Cj2 .
It leaves Cj1 through the node corresponding to the entry node. I.e., if we
entered Cj1 through a1 , we also leave to through a2 , and so on. Finally,
the edge leaving Cjs goes to xi+1 . The second paths is constructed in the
same manner and goes through Ck1 , . . . , Ckt . Every clause gadget appears
on one, two, or three path, depending on the number of its literals. Finally,
we remove all the ai , bi , and ci nodes, i = 1, 2. For each such node, if there
is one edge going into it and a second one leaving it, then we replace these
two edges by one edge going from the start node of the first edge to the end
node of the second edge. When such a node is only incident with one edge,
we remove it and its edge completely.
Figure 22.5 shows an example. x2 is the first literal of c3 , the third of c5 ,
and the first of c8 . x̄1 is the second literal of c2 .
Let G be the graph constructed from φ. G can certainly constructed in
polynomial time. So it remains to show that φ has a satisfying assignment
iff G has a Hamiltonian cycle.
For the “⇒”-direction, let a be a satisfying assignment of φ. We construct
a Hamiltonian cycle as follows. If a(xi ) = 1, we use all the edges of the paths
to xi+1 that contain the clauses in which xi occurs positively. In the other
case, we use the other path. Since a is a satisfying assignment, at least one
of the inner node of Ci that were right of a1 , b1 , or c1 is incident with one of
these edges. And by construction, also the corresponding inner nodes that
were left of a2 , b2 , or c2 are. By the first part of Lemma 22.2, we can connect
the corresponding pairs such that all inner nodes of the gadget lie on a path.
This gives a Hamiltonian cycle of G.
For the converse direction, let H be a Hamiltonian cycle of G. By the
second part of Lemma 22.2, when the cycle enters a clause gadget through
the inner node that was right to a1 , it leaves it through the inner node left
to a2 and so forth. This means that the next variable node that the cycle
visits after xi is xi+1 . Since only one edge can leave xi , the cycle either goes
through the path with positive occurences of xi or through the path with
negative occurences of xi . In the first case, we set xi to 1, in the second to 0.
Since H is a Hamiltonian cycle, it goes through each clause gadget at least
once. Hence this assignment will be a satisfying assignment.

Corollary 22.4 3SAT ≤P HC.

c Markus Bläser 2007–2015
22.2. Hamiltonian Cycle 153

C3 C5 C8

a1 a2 a1 a2 a1 a2

b1 b2 b1 b2 b1 b2

c1 c2 c1 c2 c1 c2

x2 x3

a1 a2

b1 b2

c1 c2

C2

C3 C5 C8

x2 x3

C2

Figure 22.5: An example. The variable x2 appear positively in the clauses

c3 , c5 , and c8 in the first, third, and first position. It appears negatively in
c2 in the second position. The other ai , bi , and ci nodes of the gadges lie on
paths between other variables. Next, all the ai , bi , and ci nodes, i = 1, 2 are
removed and the two edges incident to them are replaced by one.

c Markus Bläser 2007–2015
D Proof of the Cook–Karp–Levin
theorem

Reducing 3SAT to Subset-Sum, for instance, was a hard job, because the
problems look totally different. To show that SAT is NP-hard, we have to
reduce any language in NP to SAT. The only thing that we know about such
an L is that there is a polynomially time bounded nondeterministic Turing
machine M with L(M ) = M . Thus we have to reduce the question whether
a Turing machine M accepts a word x to the question whether some formula
in CNF is satisfiable. (This makes 3SAT ≤P Subset-Sum look like a picnic.)
The general reduction scheme looks as follows:

Turing machines
↓
oblivious Turing machines
↓
Boolean circuits
↓
CSAT
↓
SAT

As an intermediate concept, we introduce Boolean circuits. We show

that Boolean circuits can simulate Turing machines. To do so, we have to
first make the Turing machine oblivious, that means, that on all inputs of a
specific length, the Turing machine moves its heads in the same way. Once
we know that Boolean circuits can simulate Turing machine, it is rather
easy to show that the circuit satisfiability problem CSAT (given a circuit C,
does it have a satisfying assignment?) is NP-hard. Lastly, we show that
CSAT ≤P SAT.

D.1 Boolean functions and circuits

We interpret the value 0 as Boolean false and 1 as Boolean true. A function
{0, 1}n → {0, 1}m is called a Boolean function. n is its arity, also called the
input size, and m is its output size.
A Boolean circuit C with n inputs and m outputs is an acyclic digraph
with ≥ n nodes of indegree zero and m nodes of outdegree zero. Each node
has either indegree zero, one or two. If its indegree is zero, then it is labeled

154
D.1. Boolean functions and circuits 155

with x1 , . . . , xn or 0 or 1. Such a node is called an input node. If a node has

indegree one, then it is labeled with ¬. Such a node computes the Boolean
negation. If a node has indegree two, it is labeled with ∨ or ∧ and the
node computes the Boolean or or Boolean and, respectively. The nodes with
outdegree zero are ordered so that we can speak about the first output bit,
the second output bit etc. The nodes in a Boolean circuit are sometimes
called gates, the edges are called wires.
The depth of a node v of C is the length of a longest path from a node
of indegree zero to v. (The length of a path is the number of edges in
it.) The depth of v is denoted by depth(v). The depth of C is defined as
depth(C) = max{depth(v) | v is a node of C}. The size of C is the number
of nodes in it and is denoted by size(C).
Such a Boolean circuit C computes a Boolean function {0, 1}n → {0, 1}m
as follows. Let ξ ∈ {0, 1}n be a given input. With each node, we associate a
value val(v, ξ) ∈ {0, 1} computed at it. If v is an input node, then val(v, ξ) =
ξi , if v is labeled with xi . If v is labeled with 0 or 1, then val(v, ξ) is 0 or
1, respectively. This defines the values for all nodes of depth 0. Assume
that the value of all nodes of depth d are known. Then we compute val(v, ξ)
of a node v of depth d + 1 as follows: If v is labeled with ¬ and u is the
predecessor of v, then val(v, ξ) = ¬ val(u, ξ). If v is labeled with ∨ or ∧
and u1 , u2 are the predecessors of v, then val(v, ξ) = val(u1 , ξ) ∨ val(u2 , ξ)
or val(v, ξ) = val(u1 , ξ) ∧ val(u2 , ξ). For each node v, this defines a function
{0, 1}n → {0, 1} computed at v by ξ 7→ val(v, ξ). Let g1 , . . . , gm be the
functions computed at the output nodes (in this order). Then C computes a
function {0, 1}n → {0, 1}m defined by ξ 7→ g1 (ξ)g2 (ξ) . . . gm (ξ). We denote
this function by C(ξ).
The labels are taken from {¬, ∨, ∧}. This set is also called standard basis.
This standard is known to be complete, that is, for any Boolean function
f : {0, 1}n → {0, 1}m , there is Boolean circuit (over the standard basis) that
computes it. For instance, the CNF of a function directly defines a circuit
for it. (Note that we can simulate one Boolean and or or of arity n by n − 1
Boolean and or or of arity 2.)
Boolean circuits can be viewed as a model of parallel computation, since
a node can compute its value as soon as it knows the value of its predecessor.
Thus, the depth of a circuits can be seen as the time taken by the circuit
to compute the result. Its size measures the “hardware” needed to built the
circuit.

Exercise D.1 Every Boolean function f : {0, 1}n → {0, 1} can be computed
by a Boolean circuit of size 2O(n) . (Remark: This can be sharpened to (1 +
) · 2n /n for any  > 0. The latter bound is tight: For any  > 0 and any
large enough n, there is a Boolean function f : {0, 1}n → {0, 1} such that
every circuit computing f has size (1 − )2n /n.)

c Markus Bläser 2007–2015
156 D. Proof of the Cook–Karp–Levin theorem

D.2 Uniform families of circuits

There is a fundamental difference between circuits and Turing machines.
Turing machines compute functions with variable input length, e.g., {0, 1}∗ →
{0, 1}∗ . Boolean circuits only compute a function of fixed size {0, 1}n →
{0, 1}m . To overcome the problem that circuits compute functions of fixed
length, we will introduce families of circuits.
In the following, we will only look at Boolean circuits with one output
node, i.e., circuits that decide languages. Most of the concepts and results
presented in the remainder of this chapter also work for circuits with more
output nodes, that is, circuits that compute functions.

Definition D.1 1. A sequence C = (Cn ) of Boolean circuits such that

Ci has i inputs is called a family of Boolean circuits.

2. C is s size bounded and d depth bounded if size(Ci ) ≤ s(i) and depth(Ci ) ≤

d(i) for all i.

3. C computes the function {0, 1}∗ → {0, 1} given by x 7→ C|x| (x). Since
we can interpret this as a characteristic function, we also say that C
decides a language. We write L(C) for this language.

Families of Boolean circuits can decide nonrecursive languages, in fact

any L ⊆ {0, 1}∗ is decided by a familiy of Boolean circuits. To exclude such
phenomena, we put some restrictions on the families.

Definition D.2 1. A family of circuits is called s space and t time con-

structible, if there is an s space bounded and t time bounded determin-
istic Turing machine that given input 1n writes down an encoding of
Cn .

2. A family of circuits C is called polynomial time uniform if is is con-

structible in time polynomial in n.

Polynomial time uniform families of circuits always have polynomially

bounded size.

D.3 Simulating Turing machines by families of cir-

cuits
If we want to simulate Turing machines by circuits, there is a problem. For
two different inputs of the same length n, a Turing machine can do completely
different things, on the one input, it could move to the left, on the other it
could move to the right. But the same two inputs are fed into the circuit Cn
which is static and essentially does the same on all inputs. How can such a

c Markus Bläser 2007–2015
D.3. Simulating Turing machines by families of circuits 157

poor circuit simulate the behaviour of the Turing machine on all inputs of
length n. The idea is to tame the Turing machine.

Definition D.3 A Turing machine is called oblivious if the movement of

the heads are the same for all inputs of length n. (In particular, it performs
the same number of steps on each input of length n.)

Lemma D.4 Let t be time constructible. For every t time bounded deter-
ministic Turing machine M , there is an oblivious O(t2 ) time bounded 1-tape
deterministic Turing machine S with L(M ) = L(S).

Proof. First, we replace the Turing machine M by a Turing machine

that uses a one-sided infinite tape. S is basically the construction from
Lemma 17.2. Since t is time constructible, it is also space constructible. On
input x, S first marks t(|x|) cells and then simulates M as described in the
proof of Lemma 17.2.
To simulate one step of M , it makes a sweep over all the marked t(|x|)
cells and not just those visisted by M so far. Furthermore, S halts after
exactly simulating t(|x|) steps of M . If M halted before, then S just performs
some dummy steps that do not change anything. In this way, S becomes
oblivious.

Lemma D.5 Let M be a polynomial time bounded oblivious 1-tape deter-

ministic Turing machine with input alphabet {0, 1}. Then there a polynomial
time uniform family of circuits C with L(M ) = L(C).

Proof. Let M = (Q, Σ, Γ, δ, q0 , Qacc ). Let M be t time bounded. We can

assume that the set of states of M is a subset of {0, 1}d for some constant
d. 0 . . . 0 shall be the start state. Since a circuit can only deal with values
0 and 1, we will represent the tape alphabet by words from {0, 1}c for some
constant c. 0 will be mapped to 0 . . . 0 and 1 will be mapped to 1 . . . 1. (This
choice is fairly arbitrary and does not really matter.)
We first built a circuit D which gets an input from {0, 1}d ×{0, 1}c , a state
of M and an encoding of a symbol from the tape alphabet, and produces
as output again an element from {0, 1}d × {0, 1}c . If we feed (the encoding
of) (q, γ) in to D, we get the output (q 0 , γ 0 ) where δ(q, γ) = (q 0 , γ 0 , r) for
some r ∈ {L, S, R}, that is, D computes the transition function (except the
direction).
The circuit Cn now works in layers. Each layer consists of d edges that
carry the state of M and of t(n) “packets” of c edges, one packet for each
cell that M potentially can visit. Between the ith and the (i + 1)th layer we
place a copy of D; the d edges that correspond to the state and the c edges
that correspond to the cell from which M reads a symbol in step i. Since
M is oblivious, this cell is independent of the input. Into the first layer, we

c Markus Bläser 2007–2015
158 D. Proof of the Cook–Karp–Levin theorem

feed the input bits into these packets of edges that correspond to the n cells
that contain the input. In all other packets, we feed constants that encode
the blank, say 0 . . . 01. In the edges that correspond to the state we feed
the constants that that encode that start state, that is, 0 . . . 0. After the
last layer, we feed the edges that carry the state into a small circuit E that
outputs 1, iff the encoded state is accepting and 0 otherwise. See Figure D.1
for a sketch of Cn .
On input 1n , a Turing machine N can construct Cn as follows: Cn has a
very regular structure, so N constructs it layer by layer. The circuit D can
be “hard-wired” 1 into N , since the size of D is finite. The only problem is to
find out where to place D. But since M is oblivious, it suffices to simulate
M on 1n . This simulation also gives us the number of layers that Cn has,
namely the number of steps that M performs.
Since M is polynomial time bounded, the family (Cn ) is polynomial time
uniform.

D.4 The proof

Before we show that SAT is NP-hard, we show that a more general problem
is NP-hard, the circuit satisfiability problem CSAT which is the following
problem: Given (an encoding of) a Boolean circuit C, decide whether there
is a Boolean vector ξ with C(ξ) = 1.

Theorem D.6 CSAT is NP-hard

Proof. Let L ∈ NP and let M be a polynomial time verifier for it. We can
assume that M is oblivious. Let p be the polynomial that bounds the length
of the certificates. We can also assume that all certificates y such that M
accepts [x, y] have length exactly p(|x|). To do so, we can for instance replace
each 0 of y by 00 and each 1 by 11 and pad the certificate by appending 01.
This doubles the length of the certificates, which is fine.
We saw in Lemma D.5, that for any oblivious polynomial time bounded
Turing machine, there is a polynomial time uniform family of polynomial
size circuits Ci that decides the same language.
Now our reduction works as follows: Since for each x of length n, all
interesting certificates (certificates such that M might accept [x, y]) have
the same length, all interesting pairs [x, y] have the same length `(n), which
depends only on n. Given x, we construct C`(|x|) . Then we construct a
circuit with n + p(n) inputs that given x and y computes [x, y] and use its
output as the input to C`(|x|) Finally, we specialize the inputs belonging to
the symbols of the first part of the input to x. Our reduction simply maps
x to this circuit.
1
This means that there is a “subroutine” in N that prints D on the tape

c Markus Bläser 2007–2015
D.4. The proof 159

0 1 0 

0 x1 x2 x3 0 1

..

Figure D.1: The circuit Cn that simulates the Turing machine on inputs of
length n. At the top, the tape of the Turing machine is shown. The input
is 010. The Turing machine moves its head two times to the right during its
first two steps. The states are a subset of {0, 1}3 and the symbols of the tape
alphabet Γ are encoded as words from {0, 1}2 . Since 0 ∈ Γ is encoded by 00
and 1 ∈ Γ is encoded by 11, we just can duplicate the input node xν and
feed the two edges into D. A blank is represented by 01. There are “edges”
at the bottom that do not end in any nodes. They actually do not appear in
Cn , we have just drawn them to depict the regular structure of the layers.

c Markus Bläser 2007–2015
160 D. Proof of the Cook–Karp–Levin theorem

By construction, this new circuit C 0 has an input y with C 0 (y) = 1 if

there is a y such that M ([x, y]) = 1. Thus the mapping x 7→ C 0 is a many
one reduction of L to CSAT. It is also polynomial time, since C can be
constructed in polynomial time.

Theorem D.7 SAT is NP-hard.

Proof. It is sufficient to show CSAT ≤ SAT. Let C be the given input

circuit and let s be its size. Note that we cannot just expand a circuit into
an equivalent formula, since this may result in an expontential blow up.
Instead, we introduce new input variables a1 , . . . , as . The idea is that
aj is the output of the jth gate of C. We will construct a formula by
constructing a formula in CNF for each gate. The final formula will be the
conjunction of these small formulas.
Let gate j be an input gate. We can assume that each variable in C is
assigned to exactly one input gate. If the gate is labeled with a variable, we
do nothing. aj models this variable. If j is labeled with a constant, then
we add the clause aj or ¬aj . This forces aj to be 1 or 0, respectively, in a
satisfying assignment.
If gate j is the negation of gate i, our formula contains the expression

(ai ∨ aj ) ∧ (¬ai ∨ ¬aj ).

These formula is satisfied iff ai and aj get different values.

If gate j is the conjunction of gates h and i, then we add the expression

(¬aj ∨ ai ) ∧ (¬aj ∨ ah ) ∧ (aj ∨ ¬ai ∨ ¬ah ). (D.1)

In a satisfying assignment, either aj is 0 and at least one of ai and ah is 0

or all three are 1.
There is a similar expression for Boolean or.
Finally, we have the additional expression as , where s is the output gate.
This forces the output to be 1.
Now if there is an input such that C(x) = 1, then by construction,
we get a satisfying assignment by giving the variables a1 , . . . , as the values
computed at the corresponding gates. Conversely, any satisfying assignment
also gives an assigment to the input variables of C such that C evaluates to
one.

Remark D.8 The formula constructed in the proof above is actually in 3-

CNF. So we directly get that 3SAT is NP-complete.

Exercise D.2 Construct a similar expression as in (D.1) for the Boolean

or.

c Markus Bläser 2007–2015
 Part III

Formal languages

161
23 Finite automata and regular lan-
guages

23.1 The Chomsky hierachy

In the 1950s, Noam Chomsky started to formalize (natural) languages by
generative grammars, that is, a set of rules that describes how to gener-
ate sentences. While the purpose of Noam Chomsky was to study natural
languages, his ideas turned out to be very useful in computer science. For in-
stance, programming languages are often described by so-called context-free
grammars.
Chomsky studied four types of rules that led to four different types of
languages, usually called type-0, type-1, type-2, and type-3. We will formally
define what a grammar is when we come to type-2 languages (also called
context-free languages) since grammars are natural for type-2 languages.
For type-3 languages (also called regular languages), finite automata are the
more natural model. But once we have characterized type-2 languages in
terms of grammars, we will do so for type-3, too.

Excursus: Noam Chomsky

Noam Chomsky (born 1928 in Philadelphia, USA) is a linguist. In the 1960s, he
became known outside of the scientific community for his pretty radical political
views.
In computer science, he is mainly known for the study of the power of formal
grammars. The so-called Chomsky hierachy contains four classes of languages that
can be generated by four different kinds of grammars.

23.2 Finite automata

Finite automata describe systems that have only a finite number of states.
Consider the following toy example, a coffee vending machine. This machine
sells coffee for 40 cents.1 A potential customer can perform the following
actions: He can insert coins with the values 10 cents or 20 cents. Once he
inserted 40 cents or more, the machine brews a coffee. If the customer has not
inserted 40 cents so far, he can press the “Money back” button. The machine
keeps any overpaid money. The machine has 5 states. The four states 0c,
1
No, this machine ist not on campus.

162
23.2. Finite automata 163

20c

1 1

2 2

2
10c 30c
B
1

1
2
B

0c brew

Figure 23.1: A finite automaton that models the coffee vending machine. A
label of 1 or 2 on the edge means that the customer has inserted 10 or 20
cents, respectively. The label B means that the “Money back” button was
pressed.

10c, 20c, and 30c correspond to the amount of money inserted so far, the
state brew is entered if at least 40 cents have been inserted. Of course, the
machine starts in the state 0c. This is indicated by the triangle on the left
side of the circle. Figure 23.1 shows a diagram of the automaton COFFEE.
An arc from one state to another means that if the customer performs the
action the edge is labeled with, then the automaton will change the state
accordingly. Once the state brew is reached, the machine is supposed to
brew a coffee. A clever coffee machine would then go back to the start state
but we leave our machine as it is for now.

Exercise 23.1 1. Modify the coffee automaton such that it gives change
back. The amount of change should be indicated by the state that the
automaton ends in.

2. Modify the coffee automaton such that the customer has the choice
between several types of coffees.

Nowadays, finite automata still have several applications:

• Finite automata (and variants thereof) are used to verify systems that
have only a finite number of states.

• Finite automata can be used for string matching, see for instance Chap-
ter 32 in “Introduction to Algorithms” by Cormen, Leiserson, Rivest,
and Stein.

c Markus Bläser 2007–2015
164 23. Finite automata and regular languages

• Finite automata can be used as a tokenizer to preprocess the source

code of a computer program.

Basically, a finite automaton is a deterministic 1-tape Turing machine

that is highly restricted:

• The Turing machine M cannot change symbols on the tape.

• It can only move the head to the right.2

In other words, M has no memory and can read each symbol of the input
once and that’s it. “No memory” means no memory on the tape, M can store
a finite amount of information in its states. This is—surprise!—the reason
why M is called a finite automaton.
After throwing away all the things that are not necessary, this is what is
left of the Turing machine:

Definition 23.1 A finite automaton is described by a 5-tuple (Q, Σ, δ, q0 , Qacc ):

1. Q is the set of states.

2. Σ is the input alphabet.

3. δ : Q × Σ → Q is the transition function.

4. q0 ∈ Q is the start state.

5. Qacc ⊆ Q is the set of accepting states.

23.2.1 Computations and regular languages

A finite automaton M = (Q, Σ, δ, q0 , Qacc ) starts in state q0 . Then it reads
the first symbol w1 of its input w = w1 . . . wn and moves from q0 to the
state δ(q0 , w1 ). It continues with reading the second symbol w2 and so on.
We already defined configurations and computations for Turing machines.
We could use this definition for finite automata, too. But the only thing
of a configuration that remains relevant for a finite automaton is the state.
So a computation is essentially just a sequence of states. To get a clear
presentation, let us make this explicit:

Definition 23.2 Let M = (Q, Σ, δ, q0 , Qacc ) be a finite automaton. Let w ∈

Σ∗ .

1. A sequence s0 , . . . , sn is called a computation of M on w if

(a) s0 = q0 ,
(b) for all 0 ≤ ν < n: δ(sν , wν+1 ) = sν+1 ,
2
This renders the first condition useless, but we keep the first one for aesthetic reasons.

c Markus Bläser 2007–2015
23.2. Finite automata 165

2. The computation is called an accepting computation if in addition,

sn ∈ Qacc . Otherwise the computation is called a rejecting computa-
tion.

Remark 23.3 1. Since δ is a function, a computation, if it exists, is

always unique.

2. A computation of M on w need not exist, since δ(rν , wν+1 ) might be

undefined. We will treat such an unfinished computation also as a re-
jecting computation. In Lemma 23.5 below, we show that we can always
assume that δ is total and that there are no unfinished computations.

Definition 23.4 1. A finite automaton M = (Q, Σ, δ, q0 , Qacc ) accepts a

∗
word w ∈ Σ if there is an accepting computation of M on w. Other-
wise, we say that M rejects w.

2. L(M ) = {w ∈ Σ∗ | M accepts w} is the language that is recognized by

M.

3. A ⊆ Σ∗ is a regular language if there is a finite automaton M such

that A = L(M ).

4. REG denotes the set of all regular languages.3

Consider our automaton COFFEE. We did not specify the accepting

states of COFFEE, the natural choice is of course just to have one accept-
ing state, namely, brew. Then COFFEE accepts a word w ∈ {1, 2, B}∗ if
the corresponding sequence of actions results in a coffee for the customer.
For instance, 22, 112, 1224 , and 12B12B12B22 are accepted by COFFEE
whereas 1, 12, 12B12B are not. L(COFFEE) is the set of all words that
result in a coffee for the customer.
It will be useful to extend the transition function δ : Q × Σ → Q to words
over Σ. (δ ∗ will play the role of `∗ .) We inductively define the extended
transition function δ ∗ : Q × Σ∗ → Q. We first defined δ ∗ for words of length
0, then of length 1, length 2, and so on. For q ∈ Q, w ∈ Σ∗ , and σ ∈ Σ, we
3
For REG, we are careful and do not fix the alphabet size to {0, 1}. For REC, working
solely over {0, 1} was no problem, since one can encode larger alphabets. Finite automata
are weak and might not be able to decode the encoding. Therefore, the definition of REG
depends on the alphabet Σ. There are two solutions: Either instead of REG we define
for each Σ the class REGΣ of regular languages over Σ. Or we take a large Σ0 , like all
symbols occuring in this script. Then the input alphabet of every finite automaton in this
script is a subset of Σ0 and we can consider this automaton as an automaton over Σ0 .
Any solution is fine.
4
Not all customers are smart . . .

c Markus Bläser 2007–2015
166 23. Finite automata and regular languages

0
1 2 1
1 0

0
1

1
0 3

Figure 23.2: The finite automaton M1 . The double circle around state 3
indicates that 3 is an accepting state.

set

δ ∗ (q, ε) = q
(
δ(δ ∗ (q, w), σ) if δ ∗ (q, w) is defined
δ ∗ (q, wσ) =
undefined otherwise

The first line basically states that if the automaton reads the empty word
(which it cannot do), then it would stay in its actual state. Next, we get
δ ∗ (q, σ) = δ(δ ∗ (q, ε), σ) = δ(q, σ) for all q ∈ Q, σ ∈ Σ. So for words of length
1, δ and δ ∗ coincide, which is what we want. Intuitively, δ ∗ (q, w) is the state
that the automaton reaches if it starts in state q and then reads w.

23.2.2 How to design a finite automaton

How does one design a finite automaton? In the case of the coffee vending
machines, the states were already implicit in the description of the problem
and we just had to extract them. In other cases, this might be not this clear.
Assume we want to design an automaton M1 = (Q, {0, 1}, δ, q0 , Qacc ) that
accepts exactly the strings that contain three 0’s in a row, i.e, L(M1 ) shall
be L1 = {w ∈ {0, 1}∗ | 000 is a subword of w}.
To put ourselves into the position of a finite automaton, which has only
a finite amount of memory and can only look at one symbol at a time, think
of a word with a quadrillion of symbols, too much to look at all of them at
once and too much to remember them all. Now we have to scan them one
by one. What we have to remember is the number of 0’s that we saw after
the last 1.
We have four states, 0, 1, 2, and 3. These states count the number of 0’s.
If we are in state i and we read another 0, then we go to state i + 1. Once
we reached state 3, we know that the word is in L1 . So Qacc = {3} and once
we enter 3, we will never leave it. Whenever we read a 1 in the states 0, 1,
or 2, we have to go back to state 0.

c Markus Bläser 2007–2015
23.3. Closure properties, part I 167

Exercise 23.2 The automaton M1 basically searches for the string 000 in
the word w. Design a similar automaton that searches for the sequence 01011.
Can you devise an algorithm that given any sequence s, constructs an au-
tomaton that searches for s in a given word w?

23.3 Closure properties, part I

To get a deeper understanding of regular languages, let’s try to prove some
closure properties. We start with complementation: Given L ∈ REG is
L̄ = Σ∗ \ L again regular?5 We have an automaton M = (Q, Σ, δ, q0 , Qacc )
for L and we have to construct an automaton M̄ for L̄. The first idea is to
exchange the accepting states and the rejecting states, that is, Q \ Qacc will
be the set of accepting states of M̄ . If w ∈ L(M ), then M̄ will indeed reject
w, since the accepting computation of M on w is turned into a rejecting
one of M̄ . If w ∈/ L(M ), there is a problem: A rejecting computation of
M on w is turned into an accepting computation of M̄ on w. This is fine.
But w ∈ / L(M ) can also mean that there is an unfinished computation of
M on w. But then the computation of M̄ on w will also be unfinished and
therefore w ∈ / L(M̄ ). The next lemma shows how to make the transition
function a total function. Once we have done this, there are no unfinished
computations any more and our construction above will work.

Lemma 23.5 Let M = (Q, Σ, δ, q0 , Qacc ) be a finite automaton. Then there

is a finite automaton M 0 = (Q0 , Σ, δ 0 , q0 , Qacc ) such that δ 0 is a total function
and L(M ) = L(M 0 ).

Proof overview: We add an extra dead-lock state to Q. Whenever δ is not

defined, M 0 enters the dead lock state instead and can never leave this state.

Proof. Define Q0 = Q ∪ {dead-lock}, where dead-lock ∈ / Q. For q ∈ Q

and σ ∈ Σ, we set
(
δ(q, σ) if δ(q, σ) is defined,
δ 0 (q, σ) =
dead-lock otherwise, and
δ 0 (dead-lock, σ) = dead-lock.

δ 0 is obviously total. If w ∈ Σ∗ is accepted by M , then the computation of

M 0 on w is the same as the one of M on w. Thus M 0 also accepts w. If w is
rejected by M , then there is a rejecting computation of M on w or there is
no finished computation of M at all on w. In the first case, the computation
We always assume that the universe Σ∗ of L is known, so it is clear what the comple-
5

ment is.

c Markus Bläser 2007–2015
168 23. Finite automata and regular languages

is a rejecting computation of M 0 on w, too. In the latter case, δ ∗ (q0 , w) is

undefined. But this means that M 0 will enter dead-lock, which it cannot
leave. Thus there is a rejecting computation of M 0 on w.
Next we try to show that REG is closed under intersection and union, i.e,
if A, B ⊆ Σ∗ are regular languages, so are A ∪ B and A ∩ B. The construc-
tion that we will use is called product automaton. We already defined the
product of Turing machines. The product construction for finite automata
is essentially the same, it is just simpler because finite automata are sim-
pler. The product construction also works for the set difference A \ B. This
in particular implies that REG is closed under complementation, since Σ∗ is
regular. If M1 and M2 with states Q1 and Q2 are automata with L(M1 ) = A
and L(M2 ) = B, then the product automaton will have states Q1 × Q2 . This
automaton can simulate M1 and M2 in parallel. We use the first component
of the tuples to simulate M1 and the second to simulate M2 .

Lemma 23.6 Let M1 = (Q1 , Σ, δ1 , q0,1 , Qacc,1 ) and M2 = (Q2 , Σ, δ2 , q0,2 , Qacc,2 )
be two finite automata such that δ1 and δ2 are total functions. Then the tran-
sition function ∆ defined by

∆ : (Q1 × Q2 ) × Σ → Q1 × Q2
((q1 , q2 ), σ) 7→ (δ1 (q1 , σ), δ2 (q2 , σ))

fulfills
∆∗ ((q1 , q2 ), w) = (δ1∗ (q1 , w), δ2∗ (q2 , w))
for all q1 ∈ Q1 , q2 ∈ Q2 , and w ∈ Σ∗ .

Proof. The proof is by induction on |w|. We have ∆∗ ((q1 , q2 ), ε) =

(q1 , q2 ).
For the induction step, let w = w0 σ. We have

∆∗ ((q1 , q2 ), w) = ∆(∆∗ ((q1 , q2 ), w0 ), σ)

= ∆((δ1∗ (q1 , w0 ), δ2∗ (q2 , w0 )), σ)
= (δ1 (δ1∗ (q1 , w0 ), σ), δ2 (δ2∗ (q2 , w0 ), σ))
= (δ1∗ (q1 , w), δ2∗ (q2 , w)).

The second equality follows from the induction hypothesis.

Theorem 23.7 REG is closed under intersection, union, and set difference,
i.e, if A, B ⊆ Σ∗ are regular languages, then A ∩ B, A ∪ B, and A \ B are
regular, too.

Proof. Let M1 = (Q1 , Σ, δ1 , q0,1 , Qacc,1 ) and M2 = (Q2 , Σ, δ2 , q0,2 , Qacc,2 )

be finite automata with L(M1 ) = A and L(M2 ) = B. We may assume that

c Markus Bläser 2007–2015
23.3. Closure properties, part I 169

δ1 and δ2 are total functions. Let M = (Q1 × Q2 , Σ, ∆, (q0,1 , q0,2 ), F ) where

∆ is the function as defined in Lemma 23.6.
By defining the set of accepting states F appropriately, M will recognize
A ∩ B, A ∪ B, or A \ B. For A ∩ B, we set F = Qacc,1 × Qacc,2 . By
Lemma 23.6, ∆∗ ((q0,1 , q0,2 ), w) = (δ ∗ (q0,1 , w), δ ∗ (q0,2 , w)). We have w ∈
A ∩ B iff δ ∗ (q0,1 , w) ∈ Qacc,1 and δ ∗ (q0,2 , w)) ∈ Qacc,2 . Thus, the choice for
F above is the right one. For A ∪ B, we set F = Q1 × Qacc,2 ∪ Qacc,1 × Q2 .
For A \ B, we set F = Qacc,1 × (Q2 \ Qacc,2 ).

c Markus Bläser 2007–2015
24 Nondeterministic finite au-
tomata

In the last chapter, we showed that REG is closed under the operations
complementation, union, and intersection. In this chapter, among other
things, we want to show the closure under concatenation and Kleene closure.

Definition 24.1 Let A, B ⊆ Σ∗ .

1. The concatenation of A and B is

AB = {wx | w ∈ A, x ∈ B}.

2. The Kleene closure of A is

A∗ = {x1 x2 . . . xm | m ≥ 0 and xµ ∈ A, 1 ≤ µ ≤ m}

The concatenation AB of A and B is the set of all words that consist of

one word from A concatenated with one word from B. The Kleene closure
is the set of all words that we get when we concatenate an arbitrary number
of words (this includes zero words) from A.

Example 24.2 1. {aa, ab}{a, bb} = {aaa, aabb, aba, abbb}.

2. {aa, b}∗ = {ε, aa, b, aaaa, aab, baa, bb, aaaaaa, aaaab, . . . } is the set of
all words in which a’s always occur in pairs.

Exercise 24.1 Prove the following:

1. ∅A = A∅ = ∅ for all A ⊆ Σ∗ .

2. ∅∗ = {ε} and {ε}∗ = {ε}.

Exercise 24.2 Prove the following:

1. If A, B ⊆ Σ∗ are finite, then |AB| ≤ |A| · |B|.

2. If A 6= ∅ and A 6= {ε}, then A∗ is infinite.

170
24.1. Nondeterminism 171

24.1 Nondeterminism
When we showed that REG is closed under union or intersection, we took
two automata for A, B ∈ REG and constructed another automaton out of
these two automata for A ∩ B or A ∪ B. Here is an attempt for AB: w ∈ AB
if there are x ∈ A and y ∈ B such that w = xy. So we could first run A on x
and then B on y. The problem is that we do not know when we leave x and
enter y. The event that A enters an accepting state is not enough; during
the computation on x, A can enter and leave accepting states several times.
For instance, let A = {x ∈ {0, 1}∗ | the number of 0’s in x is even} and
B = {y ∈ {0, 1}∗ | the number of 1’s in y is odd}. How does an automata
for AB look like? In a first part, we have to count the 0’s modulo 2. At
some point, we have to switch and count the 1’s modulo 2. Figure 24.1 shows
an automaton for AB. The part consisting of the states 0-even and 0-odd
counts the 0’s modulo 2. From the state 0-even, we can go to the second part
of the automaton consisting of the states 1-even and 1-odd. This part counts
the number of 1’s modulo 2. The state 0-even is left by two arrows that are
labeled with 0 and two arrows that are labeled with 1. So this automaton is
nondeterministic. We introduced nondeterminism in the context of Turing
machines and the concept is the same for automata (which are restricted
Turing machines): The automaton can choose which of the transitions it
will make. The automaton accepts a word if there is a sequence of choices
such that the automaton ends in an accepting state. Among other things,
we use nondeterminism here to construct a nondeterministic automaton for
AB. The amazing thing is, that, in constrast to general Turing machines, we
know how to simulate a nondeterministic finite automaton by a deterministic
one (without any time loss).
Another way to introduce nondeterminism are ε-transitions. These are
arrows in the transition diagram that are labeled with ε. This means that the
automaton may choose to make the ε-transition without reading a symbol
of the input. Figure 24.2 shows an automaton with ε-transitions for AB.

24.1.1 Formal definition

Recall that for a set S, P(S) denotes the power set of S, that is the set of
all subsets of S. For an alphabet Σ, Σε denotes the set Σ ∪ {ε}.

Definition 24.3 A nondeterministic finite automaton is described by a 5-

tuple M = (Q, Σ, δ, q0 , Qacc ):

1. Q is the set of states.

2. Σ is the input alphabet.

3. δ : Q × Σε → P(Q) is the transition function.

c Markus Bläser 2007–2015
172 24. Nondeterministic finite automata

1 1

0
0−even o−odd
0

0 0

1
0
1
1−even 1−odd
1

Figure 24.1: A nondeterministic finite automaton for AB.

1 1

0
0−even o−odd
0

0 0

1
1−even 1−odd
1

Figure 24.2: A nondeterministic finite automaton with ε-transitions for AB.

The silver rule of nondeterminism

Always remember the golden rule of nondeterminism.

Nondeterminism is interesting because it is useful tool: We can model

different choices. Again, we do not know how to build an nondeterministic
finite automaton.

c Markus Bläser 2007–2015
24.1. Nondeterminism 173

4. q0 ∈ Q is the start state.

5. Qacc ⊆ Q is the set of accepting states.

If δ is only a function Q × Σ → P(Q), then M is called a nondeterministic
finite automaton without ε-transitions.

To distinguish between nondeterministic finite automata and the finite

automata of the last chapter, we call the latter ones deterministic finite au-
tomata. Deterministic finite automata are a special case of nondeterministic
finite automata. Formally, this is not exactly true. But from the transition
function δ : Q × Σ → Q of a deterministic finite automata, we get a tran-
sition function Q × Σ → P(Q) by (q, σ) 7→ {δ(q, σ)} for all q ∈ Q, σ ∈ Σ.
This means that the transition function does not map to states but to sets
consisting of a single state.

24.1.2 Computations and computation trees

A nondeterministic finite automaton starts in q0 . Then it has several choices.
There is the possibility to make an ε-transition and enter a new state without
reading a symbol. Or it may read a symbol and go to one of several states.
On one word w, there may be plenty of computations now.

Definition 24.4 Let M = (Q, Σ, δ, q0 , Qacc ) be nondeterministic finite au-

tomaton. Let w ∈ Σ∗ , w = w1 . . . wn .
1. s0 , s1 , . . . , sm ∈ Q is called a computation of M on w if we can write
w = u1 u2 . . . um with uµ ∈ Σε such that

(a) s0 = q0 .
(b) for all 0 ≤ µ < m, sµ+1 ∈ δ(sµ , uµ+1 ).

2. The computation above is called an accepting computation if sm ∈

Qacc . Otherwise, it is called a rejecting computation.

Definition 24.5 1. A nondeterministic finite automaton M accepts a

word w if there is an accepting computation of M on w. Otherwise,
M rejects w.

2. L(M ) = {w ∈ Σ∗ | M accepts w} is the language recognized by M .

We do not define nondeterministic regular languages since we will show

below that for every nondeterministic finite automaton there is a determin-
istic one that recognizes the same language; a really surprising result.
Let M = (Q, Σ, δ, q0 , Qacc ). Again, we would also like to extend the
transition function δ as we did for deterministic finite automata. But this is
not as easy. We first define the ε-closure of the transition function. For q ∈ Q

c Markus Bläser 2007–2015
174 24. Nondeterministic finite automata

and σ ∈ Σ, δ (ε) (q, σ) denotes all states that we can reach from q by making
an arbitrary number of ε-transitions and then one transition that is labeled
with σ. (And we are not allowed to make any ε-transitions afterwards.)
Formally,
δ (ε) : Q × Σ → P(Q)
(q, σ) 7→ {r | there are k ≥ 0 and s0 = q, s1 , . . . , sk such that
sκ+1 ∈ δ(sκ , ε), 0 ≤ κ < k, and r ∈ δ(sk , σ)}.

For a subset R ⊆ Q of the states, R(ε) denotes all the states in Q from which
we can reach a state in R just by ε-transitions. Formally,
R(ε) = {r ∈ Q | there are k ≥ 0 and s0 = r, s1 , . . . sk such that
sκ+1 ∈ δ(sκ , ε), 0 ≤ κ < k, and sk ∈ R.}.

Lemma 24.6 If M = (Q, Σ, δ, q0 , Qacc ) is a nondeterministic finite automa-

(ε)
ton, then M 0 = (Q, Σ, δ (ε) , q0 , Qacc ) is a nondeterministic finite automaton
without ε-transitions such that L(M ) = L(M 0 ).

Proof. M 0 obviously does not have ε-transitions. Assume that M accepts

w. Then we can write w = u1 u2 . . . um with uµ ∈ Σε , and there are states
s0 , s1 , . . . , sm with s0 = q0 , δ(sµ , uµ ) = sµ+1 for 0 ≤ µ < m, and sm ∈ Qacc .
Let i1 , . . . , in be indices such that uiν ∈ Σ for 1 ≤ ν ≤ n, i.e., uiν is wν , the
νth symbols of w. Let i0 = 0. Then δ (ε) (siν , uiν+1 ) = siν+1 for 0 ≤ ν < n
(ε) (ε)
and sin ∈ Qacc by the construction of δ (ε) and Qacc . Thus M 0 accepts w,
too.
Conversely, if M 0 accepts w = w1 w2 . . . wn with wν ∈ Σ for 1 ≤ ν ≤ n,
then M also accepts w: There are states s0 , . . . , sn such that s0 = q0 and
(ε)
δ (ε) (sν , wν+1 ) = sν+1 for 0 ≤ ν < n, and sn ∈ Qacc . For all 0 ≤ ν < n,
there are states sν,0 = sν , sν,1 , . . . , sν,jν such that δ(sν,h , ) = sν,h+1 , 0 ≤
h < jν , and δ(sν,jν , wν+1 ) = sν+1 by the construction of δ (ε) . And there
are states sn,0 = sn , sn,1 , . . . , sn,jn such that δ(sn,h , ) = sn,h+1 , 0 ≤ h < jn ,
(ε)
and sn,jn ∈ Qacc by the construction of Qacc . All these intermediate states
together form an accepting computation of M on w.

Remark 24.7 All finished computations of a finite automaton without ε-

transitions on a given input have the same length.

For a finite automaton M = (Q, Σ, δ, q0 , Qacc ), the extended transition

function δ ∗ : Q × Σ∗ → P(Q) is defined inductively as follows:
δ ∗ (q, ε) = {q} for all q ∈ Q,
δ ∗ (q, wσ) = (ε) (r, σ) for all q ∈ Q, w ∈ Σ∗ , σ ∈ Σ.
S
r∈δ ∗ (q,w) δ

δ ∗ (q, x) are all the states that we can reach if we start from q, read x, and
do not allow any ε-transition after reading the last symbol of x.

c Markus Bläser 2007–2015
24.1. Nondeterminism 175

0−even

0−odd 1−even

0−odd 1−odd

0−even 1−odd

0−even 1−odd 1−even

0−even 1−odd 1−even 1−odd

Figure 24.3: The computation tree of the automaton from Figure 24.1 on the
word 01011. The root is the start state 0-even. Then the automaton reads a
0. It has two possibilities: either it moves to 0-odd or to 1-even. These two
possibilities are represented by the two children. There are four different
computations on 01011, two of them are accepting. These two accepting
configurations correspond to splitting 01011 either as ε ∈ A and 01011 ∈ B
or 0101 ∈ A and 1 ∈ B.

As in the case of general Turing machines, it is sometimes more easier to

represent the computations of a nondeterministic finite automaton M by a
computation tree. We only do this for ε-transition free automata here. The
computation tree of M on a word w is a rooted tree whose nodes are labeled.
The root is labeled with q0 . For all states s ∈ δ(q0 , w1 ), we add one child to
q0 . This child is labeled with s. Consider one such child and let its label be
s0 . For each state r ∈ δ(s0 , w2 ), we add one child to s0 with label r and so
forth. The depth of the tree is bounded by n. The states of every level i of
the computation tree are precisely δ ∗ (q0 , w≤i ) where w≤i is the prefix of w of
length i. M accepts w if and only if there is a path of length n from the root
to some leaf that is labeled by an accepting state in the computation tree.
Figure 24.3 shows the computation tree of the automaton from Figure 24.1
on the word 01011.

c Markus Bläser 2007–2015
176 24. Nondeterministic finite automata

24.2 Determinism versus nondeterminism

Theorem 24.8 Let M = (Q, Σ, δ, q0 , Qacc ) be a nondeterministic finite au-
tomaton. Then there is a deterministic finite automaton M̂ = (Q̂, Σ, δ̂, q̂0 , Q̂acc )
such that L(M ) = L(M̂ ).

Proof overview: By Lemma 24.6, we can assume that M does not have
any ε-transitions. Look at the computation tree of M on some word w of
length n. Each level i of the tree contains all the states that M can reach
after reading the first i symbols of w. If the nth level contains a state from
Qacc , then M accepts w. A deterministic automaton has to keep track of all
the states that are in one level. But it has to store them in one state. The
solution is to take P(Q), the power set of Q, as Q̂, the set of states of M̂ .

Proof. By Lemma 24.6, we can assume that M does not have any ε-
transitions. We set Q̂ = P(Q) and q̂0 = {q0 }. We define the transition
function δ̂ as follows:
[
δ̂(R, σ) = δ(r, σ) for all R ∈ Q̂ and σ ∈ Σ.
r∈R

(Note that δ is a function Q̂ × Σ → Q̂. It does map into the power set of Q
but not into the power set of Q̂. Thus M̂ is deterministic!) Finally, we set
Q̂acc = {R ∈ Q̂ | R ∩ Qacc 6= ∅}.
We now show by induction in the length of w ∈ Σ∗ that for all R ⊆ Q
and w ∈ Σ∗ , [
δ̂ ∗ (R, w) = δ ∗ (r, w).
r∈R

For w = , δ̂ ∗ (R, w) = R = δ ∗ (R, w). Now let w = xσ with σ ∈ Σ. We have

δ̂ ∗ (R, w) = δ̂(δ̂ ∗ (R, x), σ)

[
= δ̂( δ ∗ (r, x), σ)
r∈R
[
= δ(s, σ)
s∈ r∈R δ ∗ (r,x)
S
[
= δ ∗ (r, w)
r∈R

Above, the second equality follows from the induction hypothesis and the
third equality follows from the definition of δ̂. M̂ accepts w iff δ̂ ∗ (q0 , w) ∈
Q̂acc . M accepts w iff δ ∗ ({q0 }, w) ∩ Qacc 6= ∅. From the definition of Q̂acc it
follows that M̂ accepts w iff M accepts w. Thus L(M ) = L(M 0 ).

c Markus Bläser 2007–2015
24.3. Closure properties, part II 177

0
1

1 1
0 0 0
q0 q1 q2 q3

Figure 24.4: A nondeterministic finite automaton for S3 .

Exercise 24.3 Why does this approach not work for general nondetermin-
istic Turing machines?

Somehow one feels cheated when seeing the subset construction, but it is
correct. The deterministic finite automaton pays for being deterministic by a
huge increase in the number of states. If the nondeterministic automaton has
n states, the deterministic automaton has 2n states and there are examples
where this is (almost) neccessary. Here is one such example.

Example 24.9 For n ∈ N, consider the language

Sn = {w ∈ {0, 1}∗ | the nth symbol from the end is 0}.

There is a nondeterministic finite automaton for Sn that has n + 1 states,

see Figure 24.4 for n = 3. A deterministic finite automaton could store the
last n symbols that it has seen, i.e., Q = {0, 1}n . (Yes, states can also just be
strings. We just have to use a finite set.) The transition function is defined
by δ(σ1 . . . σn , τ ) = σ2 . . . σn τ for all σ1 . . . σn ∈ Q and τ ∈ Σ. The start
state is 1n and the accepting states are all states of the form 0σ2 . . . σn . This
automaton has 2n states. We will see soon that this is neccessary.

24.3 Closure properties, part II

Theorem 24.10 REG is closed under concatenation and Kleene closure,
i.e., for all A, B ∈ REG, we have AB, A∗ ∈ REG.

Proof. Let M = (Q, Σ, δ, q0 , Qacc ) be a deterministic finite automaton

for A and M 0 = (Q0 , Σ, δ 0 , q00 , Q0acc ) be a deterministic finite automaton for
B. W.l.o.g we can assume that Q ∩ Q0 = ∅.
For AB, we construct an automaton N1 = (Q ∪ Q0 , Σ, γ1 , q0 , Q0acc ). We
connect each accepting state of M by an ε-transition to the start state q00 of

c Markus Bläser 2007–2015
178 24. Nondeterministic finite automata

M 0 . Formally,

δ(q, σ)

 if q ∈ Q and σ 6= ε,
δ 0 (q, σ)

if q ∈ Q0 and σ 6= ε,
γ1 (q, σ) =


q00 if q ∈ Qacc and σ = ε,

undef. otherwise.

It is clear from the construction that L(N1 ) = AB.

For A∗ , we construct an automaton N2 = (Q ∪ {r}, Σ, γ2 , r, Qacc ∪ {r})
where r ∈ / Q. The first idea is to insert an ε-transition from every accepting
state to the start state. In this way, whenever we enter an accepting state,
N2 can either stop or add another word from A. This automaton recognizes
A+ = A∗ \ {ε}. To add ε to the language, one idea would be to make the
start state an accepting state, too. While this adds ε to the language, this
does not work (cf. Exercise 24.4). Instead, we add a new start state r that
is an accepting state. This adds ε to the language. From r, there is an
ε-transition to the old start state q0 of M . Formally,

δ(q, σ) if q ∈ Q and σ 6= ε,

γ2 (q, σ) = q0 if q ∈ {r} ∪ Qacc and σ = ε,

undef. otherwise.


It is clear from the construction that L(N2 ) = A∗ .

Exercise 24.4 Give an example that in the proof above, we need the extra
state r in N2 , i.e., show that it is not correct to create an ε-transition from
each accepting state to the start state and make the start state an accepting
state, too.

24.4 Further exercises

Exercise 24.5 For a word w = w1 w2 . . . wn ∈ Σ∗ , let wrev denote the word
wn wn−1 . . . w1 . For a language L ⊆ Σ∗ , Lrev = {wrev | w ∈ L}. Show that if
L is regular, so is Lrev .

c Markus Bläser 2007–2015
25 Regular expressions

25.1 Formal definition

25.1.1 Syntax
Definition 25.1 Let Σ be an alphabet and assume that the symbols “(”, “)”,
∅”, “ε”, “+”, and “ ∗ ” do not belong to Σ. Regular expressions over Σ are
defined inductively:

1. The symbols ∅ and ε are regular expressions.

2. For each σ ∈ Σ, σ is a regular expression.

3. If E and F are regular expressions, then (E + F ), (EF ) and (E ∗ ) are

regular expressions.

Above ∅ and ε are symbols that will represent the empty set and the set
{ε}, but they are not the empty set or the empty word themselves. But since
these underlined symbols usually look awkward, we will write ∅ instead of ∅
and ε instead of ε. It is usually clear from the context whether we mean the
symbols for the empty set and the empty word or the objects themselves.

25.1.2 Semantics
Definition 25.2 Let E be a regular expression. The language L(E) denoted
by E is defined inductively:

1. If E = ∅, then L(E) = ∅.
If E = ε, then L(E) = {ε}.

2. If E = σ for some σ ∈ Σ, then L(E) = {σ}.

3. If E = (E1 + E2 ), then L(E) = L(E1 ) ∪ L(E2 ).

If E = (E1 E2 ), then L(E) = L(E1 )L(E2 ).
If E = (E1∗ ), then L(E) = L(E1 )∗ .

The symbol ∅ represents the empty set and the symbol ε represents the
set that contains solely the empty word. A symbol σ ∈ Σ represents the
set that contains the symbol itself. These three cases form the basis of the
definition. Next come the cases where E is composed of smaller expressions.
The operator “+” corresponds to the union of the corresponding languages,
the concatenation of the expression corresponds to the concatenation of the

179
180 25. Regular expressions

corresponding languages and the “ ∗ ”-operator stand for the Kleene closure.
Union, concatenation, and Kleene closure are also called the regular opera-
tions.

Excursus: Regular expressions in Unix

Unix uses regular expressions though with a different notation and some extensions
(which are just syntactic sugar):
• The dot “.” can mean any one symbol. This can be replaced by σ1 + σ2 +
· · · + σ` , if Σ = {σ1 , . . . , σ` }. But the dot is very handy, since the alphabet
in Unix is pretty large.
• [τ1 τ2 . . . τk ] is the union of the k symbols, i.e, in our notation it is τ1 + τ2 +
· · · + τk .
• Since the ASCII symbols are ordered, one can also write something like [a-z],
which means all symbols between a and z.
(Warning: this is not really true any more since the order nowadays depends
on the locale. So maybe fancy characters like ä could occur between a and
z. Therefore, shortcuts like [:lower:] (all lower case letters) have been
introduced.)
• The union of two expressions is denoted by |.
• E∗ is the Kleene closure of E, E+ stands for E + := EE ∗ (that is, the
concatenation of any positive number of words from L(E)), and E{m} stands
for EE . . . E}.
| {z
m times

For instance, elements of programming languages, so-called tokens are usually

elements from a regular language.1 Consider our language WHILE and assume that
we have to write it as an ASCII text. So while is not a symbol anymore but it is
composed of the five letters while. The regular expression [Ww][Hh][Ii][Ll][Ee]
describes all possibilites to write the keyword while. We assume that WHILE
is not a case-sensitive language. The expression x0|x[1-9][0-9]* describes the
variable names that we are allowed to use. (We forbid leading zeros here.)
The tool lex performs the so-called lexical analysis. Basically, it gives us the
tokens from the ASCII text. We give lex a list of regular expressions together with
an action to be performed and then lex scans the source code for occurences of
elements of the regular languages. Whenever one is found, it executes the corre-
sponding action (like inserting a new entry in the list of used variables, etc.).
The tools grep or egrep work with regular expressions, too. For instance,
egrep ’x0|x[1-9][0-9]*’ while.prg gives you all occurences of variables in the
source code while.prg.

1
But notice that the languages of all valid source codes are usually not regular. They
are usually context-free, a concept that we will meet after Christmas. The use of “usually
context-free” is questionable here, in particular, Bodo disagrees. The set of source codes
of a “pure” and “simple” programming language like WHILE is context-free, the set of an
overloaded one like C++ or JAVA usually is not.

c Markus Bläser 2007–2015
25.2. Algebraic laws 181

25.2 Algebraic laws

Like addition and multiplication over, say, the integers fulfill several algebraic
laws, the regular operations fulfill an abundance of such laws, too. If we write
17 + 4 = 21, then this means that the natural numbers on the left-hand side
and the right-hand side are the same; though the words on both side are not:
the left-hand side is the concatenation of the four symbols 1, 7, +, and 4, the
right hand side is the concatenation of 2 and 1. In the same way, if we write
E = F for two regular expressions, then this means that L(E) = L(F ), i.e.,
the two expressions describe the same language. Like x + y = y + x holds
for all natural numbers x and y, we can formulate such laws for regular
expressions.

Theorem 25.3 For all regular expressions E, F , and G, we have

1. E + F = F + E (commutativity law for union),

2. (E + F) + G = E + (F + G) (associativity law for union),

3. (EF)G = E(FG) (associativity law for concatenation),

4. ∅ + E = E + ∅ = E (∅ is an identity for union),

5. εE = Eε = E (ε is the identity for concatenation),

6. ∅E = E∅ = ∅ (∅ is an annihilator for concatenation),

7. E + E = E (union is idempotent),

8. (E + F)G = (EG) + (FG) (right distributive law),

9. E(F + G) = (EF) + (EG) (left distributive law),

10. (E ∗ )∗ = E ∗ ,

11. ∅∗ = ε,

12. ε∗ = ε.

Proof. We only prove the first, sixth, and tenth part. The rest is left as
an exercise.
For the first part, we use the fact that the union of sets is commutative:
L(E + F ) = L(E) ∪ L(F ) = L(F ) ∪ L(E) = L(F + E).
Part 6: For any two languages A and B, AB is the set of all words w = ab
with a ∈ A and b ∈ B. If one of A and B is empty, then no such word w
exists, thus AB = ∅ in this case.
Part 10: Let L := L(E). L∗ ⊆ (L∗ )∗ is clear, since (L∗ )∗ is the set of
all words that we get by concatenating an arbitrary number of words from

c Markus Bläser 2007–2015
182 25. Regular expressions

L∗ , in particular of one word. To show that (L∗ )∗ ⊆ L∗ , let x1 , . . . , xn ∈

L∗ . This means that we can write every xν as yν,1 . . . yν,jν with yν,i ∈ L
for 1 ≤ i ≤ jν , 1 ≤ ν ≤ n. This means that we can write x1 . . . xn =
y1,1 . . . y1,j1 y2,1 . . . yn−1,jn−1 yn,1 . . . yn,jn ∈ L∗ .

Exercise 25.1 1. Prove the remaining parts of Theorem 25.3.

2. Construct two regular expressions E and F with EF 6= F E.

Exercise 25.2 Prove the following identities:

1. (E + F )∗ = (E ∗ F ∗ )∗ .

2. ε + EE ∗ = E ∗ .

3. (ε + E)∗ = E ∗ .

25.2.1 Precedence of operators

In the definition of regular expressions, whenever we built new expression
out of two other, we placed brackets around the expression. Since these
brackets often make the expressions hard to read, we are allowed to omit
them. But then we have to make conventions about the order of precedence
of operators and how to get back the completely parenthesized expressions.

• Kleene closure ∗ has the highest precedence. That is, it applies to

the smallest sequence of symbols to its left that form a valid regular
expression. We put brackets around this expression and the star.

• Next comes concatenation. All concatenations that are adjacent are

grouped together. Since concatenation is associative, it does not matter
in which order we group consecutive concatenations. But to get a
deterministic procedure, we will always group them from the left to
the right.

• Finally, we group all unions. Again, the order does not matter, but we
will do it from the left to the right by convention.

This is just like the precedence of exponentiation over multiplication over

addition in arithmetic expressions.

Example 25.4 Consider the expression 01∗ 0 + 0 + ε. It is transformed back

into ((((0(1∗ ))0) + 0) + ε).

c Markus Bläser 2007–2015
25.3. Regular expressions characterize regular languages 183

25.3 Regular expressions characterize regular lan-

guages
As the name suggests, regular expression characterize—surprise, surprise!—
exactly the regular languages. This means that for every regular expression
E, L(E) is regular. And conversely, if L is regular, then we can find a regular
expression E such that L(E) = L.

Theorem 25.5 If E is a regular expression, then L(E) ∈ REG.

Proof overview: The proof is done by structural induction. We first

prove the statement of the theorem for the simple regular expressions (Defi-
nition 25.1, 1. and 2.). This is the induction basis. In the induction step, we
have prove the statement for the expressions E1 + E2 , E1 E2 , and E1∗ and the
induction hypothesis is that L(E1 ) and L(E2 ) are regular. We can reduce
structural induction to “regular” induction by viewing it as induction in the
number of applications of the operators “+”, concatenation, and “ ∗ ”.

Proof. Induction base: We have to construct automata that accept the

languages ∅, {ε}, and {σ}. But this is easy (see also Exercise 25.3).
Induction step: We are given E = E1 + E2 , E = E1 E2 , or E = E1∗ and
we know that L(E1 ) and L(E2 ) are regular. We have to show that L(E) =
L(E1 ) ∪ L(E2 ) or L(E) = L(E1 )L(E2 ) or L(E) = L(E1 )∗ , respectively, is
regular. But this is clear, since we already showed that REG is closed under
union, concatenation and Kleene closure.

Exercise 25.3 Construct finite automata that accept the languages ∅, {ε},
and {σ} for σ ∈ Σ.

Theorem 25.6 For every deterministic finite automaton M = (Q, Σ, δ, q0 , Qacc ),

there is a regular expression E such that L(M ) = L(E).

Proof overview: We assume that Q = {1, 2, . . . , n}. Assume that 1 is the

start state and j1 , . . . , j` are the accepting states. For each pair i, j of states,
we will inductively define expressions Ei,j k , 0 ≤ k ≤ n, such that in the end,

L(Ei,j ) is exactly the set of all strings w such that δ ∗ (i, w) = j, i.e, if M starts
n

in i and reads w, then it ends in j. But then L(M ) = L(E1,j n + · · · + E n ).

1 1,j`
k
What does the superscript k mean? L(Ei,j ) will be the set of all words such
that δ ∗ (i, w) = j and for each prefix w0 of w with w0 6= ε and w0 6= w,
δ ∗ (i, w0 ) ≤ k, i.e, if M starts in i and reads w, then it ends in j and during
this computation, M only entered states from the set {1, . . . , k} (but the i in
the beginning and the j in the end do not count). We will start with k = 0
and then go on by induction.

c Markus Bläser 2007–2015
184 25. Regular expressions

k such that
Proof. By induction on k, we construct expressions Ei,j
k ) = {w | δ ∗ (i, w) = j and for each prefix w 0 of w
L(Ei,j
(25.1)
with w0 6= ε and w0 6= w, δ ∗ (i, w0 ) ≤ k}
for all 1 ≤ i, j ≤ n and 0 ≤ k ≤ n.
Induction base: If k = 0, then M is not allowed to enter any intermediate
state when going from i to j. If i 6= j, then this means that M can only take
an arc directly from i to j. Let σ1 , . . . , σt be the labels of the arcs from i to j.
Then Ei,j 0 = σ + σ + · · · + σ with the convention that this means E 0 = ∅
1 2 t i,j
if t = 0, i.e., there are no direct arcs from i to j. If i = j, then let σ1 , . . . , σt
be the labels of all arcs from i to itself. Now we set Ei,i 0 = ε+σ +···+σ .
1 t
It is clear from the construction that Ei,j 0 fulfills (25.1).

Induction step: Assume that we have found regular expressions such that
k+1
(25.1) holds for some k. We have to construct the expressions Ei,j . A path
that goes from i to j and all states in between are from {1, . . . , k + 1} can go
from i to j with only going through states from {1, . . . , k}. For this, we know
already an regular expression, namely Ei,k k . Or, when going from i to j, we

go through k + 1 exactly once. But we can view this as going from i to k + 1

and all intermediate states are from {1, . . . , k} and then going from k +1 to j
and again, all intermediate states are from {1, . . . , k}. For the first part, we
have already designed a regular expression, namely Ei,k+1 k , but also for the
k
second part we have one, namely Ek+1,j . Their concatenation, Ei,k+1 k k
Ek+1,j ,
describes all words such that the automaton started in i, goes only through
states {1, . . . , k +1} and only once through k +1, and ends in j. In general, if
we go from i to j and all intermediate states are from {1, . . . , k + 1}, then we
go from i to k+1 and all intermediate states are from {1, . . . , k}, after that we
may go several times from k + 1 to k + 1 and all intermediate states are from
{1, . . . , k} and finally we go from k + 1 to j and all intermediate states are
from {1, . . . , k}. The corresponding expression is Ei,k+1 k k
(Ek+1,k+1 )∗ Ek+1,j
k .
k+1 k k k ∗ k
Altogether, we have Ei,j = Ei,j + Ei,k+1 (Ek+1,k+1 ) Ek+1,j . It is clear that
k+1
L(Ei,k ) is a subset of the language on the right-hand side of (25.1). But also
the converse is true: Let x1 , . . . , xs be all prefixes of w (sorted by increasing
length) such that δ ∗ (i, xi ) = k + 1, 1 ≤ i ≤ s, and define yi by xi+1 = xi yi .
Then δ ∗ (i, x1 ) = k+1, δ ∗ (k+1, yi ) = k+1 for 1 ≤ i < s, and δ ∗ (k+1, ys ) = j
k+1
where ys fulfills xs ys = w. But this means that w ∈ L(Ei,j ). This proves
(25.1).
From (25.1), it follows that L(M ) = L(E1,j n + · · · + E n ) where Q
1 1,j` acc =
{j1 , . . . , j` }.

Remark 25.7 The algorithm above does essentially the same as the Floyd–
k+1
Warshall algorithm for computing all-pair shortest paths. If we replace Ei,j =
k k k ∗ k k k k k
Ei,j + Ei,k+1 (Ek+1,k+1 ) Ek+1,j by di,j = min{di,j , di,k+1 + dk+1,j }, we can
compute the shortest distances between all pairs of nodes.

c Markus Bläser 2007–2015
26 The pumping lemma

To show that a language L is regular, it is sufficient to give a deterministic

finite automaton M with L = L(M ) or a nondeterministic one or a regular
expression. But how do we show that a language L is not regular? Are there
non-regular languages at all? Well there are since there are undecidable
languages, but are there simple non-regular languages? Here is one example:

A = {0n 1n | n ∈ N}.

If there were a finite automaton for A, then it would have to keep track of
the number of 0’s that it read so far and compare it with the number of
1’s. But in a finite number of states, you can only store a finite amount of
information. But M potentially has to be able to store an arbitrarily large
amount of information, namely n. (Warning! Never ever write this in an
exam! This is just an intuition. Maybe there is a way other than counting
to check whether the input is of the form 0n 1n —there is not, but you have
to give a formal proof.)

26.1 The pumping lemma

Lemma 26.1 (Pumping lemma) Let L be a regular language. Then there
is an n > 0 such that for all words u, v, w with uvw ∈ L and |v| ≥ n, there are
words x, y, z with v = xyz and |y| > 0 such that for all i ∈ N, uxy i zw ∈ L.

Proof overview: We choose n to be the number of states of a finite au-

tomaton M that recognizes L. In the part of the computation of M on v, at
least one state has to repeat by the pigeon hole principle. This means that
we have discovered a loop in the computation of the automaton. By going
through this loop i times instead of one time, we get the words uxy i zw.
Below, there is a automaton with uvw on the input tape. In the second
line, there are the states of M when moving from one cell to the next. By
the pigeon hole principle, two of them are equal, say, qi = qj with i < j. The
word vi+1 . . . vj between them can be “pumped”.

... um v1 v2 ... v n w1 ...

q0 q1 q2 ... qn−1 qn

185
186 26. The pumping lemma

Proof. Since L is regular, it is accepted by a deterministic finite automa-

ton M = (Q, Σ, δ, q0 , Qacc ). Let n = |Q|. Let u, v, w be words such that
uvw ∈ L and |v| ≥ n. Let |u| = m, |v| = n0 ≥ n, and |w| = `. Since
uvw ∈ L, there is an accepting computation s0 , s1 , . . . , sm+n0 +` of M on
uvw, i.e., s0 = q0 , for all 0 ≤ j ≤ m + n0 + `

δ(sj , uj+1 )
 if 0 ≤ j < m,
sj+1 = δ(sj , vj+1−m ) if m ≤ j < m + n0 ,

δ(sj , wj+1−m−n0 ) if m + n0 ≤ j < m + n0 + `,


and sm+n0 +` ∈ Qacc . Since sm , . . . , sm+n0 are more than n states, there are
indices m ≤ j1 < j2 ≤ m + n0 such that sj1 = sj2 by the pigeon principle.
Let v = xyz with |x| = j1 − m and |y| = j2 − j1 > 0. Then δ ∗ (q0 , ux) = sj1 ,
δ ∗ (sj1 , y) = sj2 , and δ ∗ (sj2 , zw) ∈ Qacc . Since sj1 = sj2 ,

δ ∗ (q0 , uxy i zw) = δ ∗ (sj1 , y i zw)

= δ ∗ (sj1 , y i−1 zw)
..
.
= δ ∗ (sj1 , yzw)
= δ ∗ (sj1 , zw)
∈ Qacc

Thus uxy i zw ∈ L for all i ∈ N.

26.2 How to apply the pumping lemma

To show that a language L is not regular, one can show that L does not
fulfill the condition of the pumping lemma since the contraposition of the
statement of the pumping lemma says that if L does not fulfill the condition
then it is not regular.

Example 26.2 Let us show that A = {0n 1n | n ∈ N} is not regular. The

contraposition of the pumping lemma says that if for all n there are words
u, v, w with uvw ∈ A and |v| ≥ n such that for all words x, y, z with xyz = v
and |y| > 0, there is an i ∈ N such that uxy i zw ∈
/ A, then A is not regular.
Let n be given. We set u = ε, v = 0n , and w = 1n . Obviously, uvw ∈ A
and |v| ≥ n. Let v = xyz with |y| > 0. Since v = 0n , x = 0r , y = 0s ,
and z = 0t with r + s + t = n and s > 0. We have uxy i zw = 0n+(i−1)s 1n .
Except for i = 1, this is not a word in A. (Even setting i = 0, i.e., “pumping
down” gives a contradiction here.) Thus A does not fufill the condition of the
pumping lemma and thus is not regular.

c Markus Bläser 2007–2015
26.3. How to decide properties of regular languages 187

Example 26.3 Closure properties are sometimes useful to simplify proofs

of non-regularity: Consider

B = {x ∈ {0, 1}∗ | the number of 0’s equals the number of 1’s in x}.

We have A = B ∩ L(0∗ 1∗ ). If B were regular, so would be A, a contradiction.

The pumping lemma game

Proving that a language L is not regular via the pumping lemma can
be considered as a game between you and an adversary, for instance
your professor.

1. Your professor picks an n ∈ N \ {0}.

2. You pick words u, v, w such that uvw ∈ L and |v| ≥ n.

3. Your professor picks words x, y, z such that v = xyz and |y| >
0.

4. You pick an i ∈ N.

Now it comes to the showdown: You win if uxy i zw ∈ / L. Your

i
professor wins if uxy zw ∈ L. If you have a winning strategy, i.e., no
matter what your professor picks, you can always make your choices
such that you win, then L is not regular.
(If L is indeed not regular, this is one of the rare chances to win against
your professor.)

The iron pumping lemma rule

The condition of the pumping lemma is only necessary.

To show that a language L is not regular you can show that it does not
satisfy the condition of the pumping lemma. But you cannot prove that L
is regular by showing that L fulfills the condition of the pumping lemma.
There are non-regular languages that fulfill the condition of the pumping
lemma.

26.3 How to decide properties of regular languages

Not part of the lecture

Which properties can we decide about regular languages? Can we decide

whether a regular language is empty? This question does not seem to be

c Markus Bläser 2007–2015
188 26. The pumping lemma

meaningful since a regular language is empty if it is empty. So the true

question is: Given a representation of a regular language, for instance a
deterministic finite automaton M , is L(M ) empty? Here are the problems
that we will consider:

Word problem: Given M and w ∈ Σ∗ , is w ∈ L(M )?

Emptiness problem: Given M , is L(M ) = ∅?

Finiteness problem: Given M , is |L(M )| < ∞?

Equivalence problem: Given M1 and M2 , is L(M1 ) = L(M2 )?

The running time of the decision procedures depends on the representa-

tion of the regular language. We always assume that the regular language
is given by a deterministic finite automaton. Of course, if we are given a
nondeterministic finite automaton, we can transform it into an equivalent
deterministic one, but this new automaton is much larger than the nonde-
terministic one. Nevertheless, some of the decision procedures that we give
even work for nondeterministic automata without any changes.

26.3.1 The word problem

Testing whether w ∈ L(M ) is easy. We just compute δ ∗ (q0 , w) and test
whether it is in Qacc . This takes |w| steps if the automaton is deterministic.
If M is nondeterministic, then we just compute a list of states that can be
reached when reading a longer and longer prefix of w.

26.3.2 Testing emptiness

If a given deterministic finite automaton M has n states, then L(M ) contains
a word of length < n, provided that L(M ) is non-empty. This follows from
the pumping lemma. Assume we have a word w ∈ L(M ) with |w| ≥ n. Then
the pumping lemma says that there must be a shorter word in L(M ), namely
the one that we get when we set i = 0. Thus in order to decide whether
L(M ) is empty or not, we “just” have to check all words of length < n. While
n −1
|Σ| = |Σ|
Pn−1 ν
these are finitely many, there are still a lot, namely ν=0 |Σ|−1 . A
much faster algorithm is obtained by viewing the transition diagram as a
directed graph and then checking whether there is a path from the start
state to an accepting state.

26.3.3 Testing finiteness

L(M ) is infinite iff L(M ) contains a word of length between n and 2n − 1
where n is the number of states of M : If L(M ) is infinite, then it contains
words that are arbitrarily long. By the pumping lemma, as long as the word

c Markus Bläser 2007–2015
26.4. Further exercises 189

has length ≥ 2n, we can shorten it by an (unknown) amount between 1 and

n. Thus we can bring it down to a length between n and 2n − 1. On the
other hand, if the language has a word with length between n and 2n − 1,
then we can pump this word and thus L(M ) is infinite. Again we get a
faster algorithm by using graphs algorithms to check whether a node in the
transition diagram lies on any cycle and then checking whether there is a
path from the start state to some accepting state that contains a node on a
cycle.

26.3.4 Testing equivalence

When is L(M1 ) = L(M2 )? This is equivalent to L(M1 ) \ L(M2 ) = ∅ and
L(M2 ) \ L(M1 ) = ∅. Thus we just have to construct the appropriate product
automata and can reduce the problem to emptiness testing.

The Schöning–Seidel–I don’t know who version

of the pumping lemma

In many textbooks or lecture notes of other lecturers you often find

the following version of the pumping lemma:
Let L be a regular language. Then there is an n > 0 such that for all
words t with t ∈ L and |t| ≥ n, there are words x, y, z with t = xyz,
|xy| ≤ n, and |y| > 0 such that for all i ∈ N, xy i z ∈ L.
Our version is more general than this version, since we can set u = ,
v the prefix of length n of t, and w the rest of t.
Our version is much cooler, too. Let L = {0i 10n 1n | i ≥ 10, n ∈ N}.
In the version above, we cannot rule out that y is a substring of
the first 10 zeroes and then pumping does not help. There are
ways to work around this, but why do you want to make your life
unneccesarily complicated. We can just choose u = 0i 1, v = 0n , and
w = 1n and are done.

26.4 Further exercises

Exercise 26.1 Show that the following languages are not regular by using
the pumping lemma.

1. L = {0n 1m | n > m},

2. L = {0p | p is prime},

3. L = {ai bj ck | i + j = k}.

c Markus Bläser 2007–2015
190 26. The pumping lemma

Exercise 26.2 Show that no infinite subset of {0n 1n | n ∈ N} is regular.

Exercise 26.3 A set U ⊆ N is called ultimately periodic if there are n0 ∈ N

and p > 0 such that for all n ≥ n0 , n ∈ U iff n + p ∈ U . p is called the
period of U .
Show the following: L ⊆ {0}∗ is regular iff E = {e ∈ N | 0e ∈ L} is ultimately
periodic.

c Markus Bläser 2007–2015
E The Myhill-Nerode theorem
and minimal automata

Recall that an equivalence relation is a relation that is reflexive, symmetric,

and transitive. We here consider equivalence relations on Σ∗ . Let R be such
an equivalence relation. For x ∈ Σ∗ , [x]R denotes the equivalence class of x,
i.e., the set of all y ∈ Σ∗ such that xRy. The equivalence classes of R form
a partition of Σ∗ , that means, they are pairwise disjoint and their union is
Σ∗ . We call a relation right invariant (with respect to concatenation) if for
all x, y ∈ Σ∗ ,
xRy =⇒ xzRyz for all z ∈ Σ∗ .
The index of an equivalence relation R is the number of equivalence classes
of R and is denoted by index(R). (If the number of equivalence classes is
not finite, then index(R) is infinite.)

Definition E.1 (Automaton relation) Let M = (Q, Σ, δ, q0 , Qacc ) be a

deterministic finite automaton. The relation ≡M is defined on Σ∗ by

x ≡M y : ⇐⇒ δ ∗ (q0 , x) = δ ∗ (q0 , y).

Lemma E.2 For every deterministic finite automaton M , ≡M is an equiv-

alence relation that is right invariant and has a finite index.

Proof. ≡M is an equivalence relation because = on the set Q is an

equivalence relation.
It is right invariant, because δ ∗ (q0 , xz) = δ ∗ (δ ∗ (q0 , x), z). If x ≡M y, then
δ ∗ (q0 , x) = δ ∗ (q0 , y) and

δ ∗ (q0 , xz) = δ ∗ (δ ∗ (q0 , x), z) = δ ∗ (δ ∗ (q0 , y), z) = δ ∗ (q0 , yz)

and therefore, xz ≡M yz.

Finally, the index of ≡M is bounded from above by |Q|, thus it is finite.

Remark E.3 If all states of M are reachable from the start state, i.e., for
all q ∈ Q, there is an x ∈ Σ∗ such that δ ∗ (q0 , x) = q, then index(≡M ) = |Q|.

Let L ⊆ Σ∗ , and let M be a deterministic finite automaton for it. ≡M

defines a relation on Σ∗ that of course depends on M . If we take two different
deterministic finite automata M1 and M2 for L, then the relations might

191
192 E. The Myhill-Nerode theorem and minimal automata

be different. Next, we define a relation ∼L on Σ∗ that is independent of

the chosen automaton. As we will see, every other relation ≡M will be a
refinement of it, i.e, every equivalence class of ≡M is contained in a class of
∼L(M ) . The relation ∼L is even defined for languages that are not regular.

Definition E.4 (Myhill-Nerode relation) Let L ⊆ Σ∗ . The Myhill-Nerode

relation ∼L is defined on Σ∗ by

x ∼L y : ⇐⇒ [for all z ∈ Σ∗ : xz ∈ L ⇐⇒ yz ∈ L].

Lemma E.5 For every L ⊆ Σ∗ , ∼L is an equivalence relation that is right

invariant.

Proof. ∼L is an equivalence relation, since ⇐⇒ is an equivalence relation.

To see that it is right invariant, let x ∼L y. We have to show that
xw ∼L yw for all w ∈ Σ∗ . xw ∼L yw means that for all z ∈ Σ∗ , xwz ∈
L ⇐⇒ ywz ∈ L. But this is clear since x ∼L y means that for all z 0 ∈ Σ∗ ,
xz 0 ∈ L ⇐⇒ yz 0 ∈ L, in particular for z 0 = wz.

E.1 The Myhill-Nerode theorem

Theorem E.6 (Myhill-Nerode) Let L ⊆ Σ∗ . The following three state-
ments are equivalent:

1. L is regular.

2. L is the union of some equivalence classes of a right invariant equiva-

lence relation with finite index.

3. ∼L has finite index.

Proof. 1. =⇒ 2.: If L is regular, then there is a deterministic finite

automaton M = (Q, Σ, δ, q0 , Qacc ) with L(M ) = L. The relation ≡M is an
equivalence relation that is right invariant. Its index is ≤ |Q| and hence
finite. By definition,

L = {x ∈ Σ∗ | δ ∗ (q0 , x) ∈ Qacc }
[
= [x]M .
x: δ ∗ (q0 ,x)∈Qacc

Although the union is over infinitely many words x, only finitely many dis-
tinct equivalence classes appear in it.
2. =⇒ 3.: Let R be a right invariant equivalence relation with finite index
such that L = [x1 ]R ∪ · · · ∪ [xt ]R . We show that R is a refinement of ∼L , that
is, every equivalence class C of R is a subset of some equivalence class C 0 of

c Markus Bläser 2007–2015
E.1. The Myhill-Nerode theorem 193

∼L . This implies index(∼L ) ≤ index(R). Since index(R) is finite, index(∼L )

is finite, too.
Let x, y ∈ Σ∗ with xRy. If we can show that x ∼L y, then we are done
since this means that any two elements from an equivalence class of R are
in the same equivalence class of ∼L . Hence every equivalence class of R is
contained in an equivalence class of ∼L . Since R is right invariant,

xzRyz for all z ∈ Σ∗ . (E.1)

Since L = [x1 ]R ∪ · · · ∪ [xt ]R , R-equivalent words are either both in L or both

not in L. Hence (E.1) implies

xz ∈ L ⇐⇒ yz ∈ L for all z ∈ Σ∗ .

Thus x ∼L y.
3. =⇒ 1.: Given ∼L , we construct a deterministic finite automaton M =
(Q, Σ, δ, q0 , Qacc ) with L = L(M ). We set

• Q = the set of equivalence classes of ∼L ,

• δ([x]∼L , σ) = [xσ]∼L for all σ ∈ Σ,

• q0 = [ε]∼L ,

• Qacc = {[x]∼L | x ∈ L}.

δ is defined in terms of representatives, thus we have to check that it is well-

defined, that means, if x ∼L y, then xσ ∼L yσ. But this follows immediately
from the right invariance of ∼L .
It remains to verify that L(M ) = L. An easy proof by induction shows
that δ ∗ ([ε]∼L , x) = [x]∼L . Thus,
[
L(M ) = {x | δ ∗ ([ε]∼L , x) ∈ Qacc } = [x]∼L = L
x∈L

since the words in an equivalence classes of ∼L are either all in L or all not
in L.

Exercise E.1 Show that δ ∗ ([ε]∼L , x) = [x]∼L for all x ∈ Σ∗ in the “3. =⇒
1.”-part of the proof of the Myhill–Nerode theorem.

Example E.7 Let A = {0n 1n | n ∈ N}. We have 0i 6∼A 0j for i 6= j,

since 0i 1i ∈ A but 0j 1i ∈
/ A. Thus [0i ]∼A for i ∈ N are pairwise distinct
equivalence classes. Thus the index of ∼A is infinite and A is not regular.

c Markus Bläser 2007–2015
194 E. The Myhill-Nerode theorem and minimal automata

Myhill-Nerode theorem versus Pumping lemma

Both results are tools to show that a language is not regular.

Pumping lemma: often easy to apply but does not always work

Myhill-Nerode theorem: always works but often it is quite some

work to determine the equivalence classes of ∼L . Keep in mind that
to show that ∼L has infinite index it is sufficient to find an infinite
number of equivalence classes—we do not have to find all of them.

E.2 The minimal automaton

Let M = (Q, Σ, δ, q0 , Qacc ) and M 0 = (Q0 , Σ, δ, q00 , Q0acc ) be deterministic
finite automata such that δ and δ 0 are total. M and M 0 are isomorphic if
there is a bijection i : Q → Q0 sucht that

1. for all q ∈ Q and σ ∈ Σ,

i(δ(q, σ)) = δ 0 (i(q), σ),

2. i(q0 ) = q00 ,

3. i(Qacc ) = Q0acc .

Such a mapping i is called an isomorphism. The first condition says that the
following diagram commutes:

Together with the second and third condition, this means that two isomor-
phic automata are the same up to renaming the states.

Theorem E.8 Let L ⊆ Σ∗ be a regular language.

1. Any deterministic finite automaton M 0 = (Q0 , Σ, δ 0 , q00 , Q0acc ) such that

δ 0 is a total function and L(M 0 ) = L has at least index(∼L ) states.

2. Every deterministic finite automaton with total transition function and

index(∼L ) many states that recognizes L is isomorpic to the automaton
M constructed in the “3. =⇒ 1.”-part of the proof of the Myhill–Nerode
theorem.

c Markus Bläser 2007–2015
E.2. The minimal automaton 195

Proof. Part 1: If we combine the “1. =⇒ 2.”- and the “2. =⇒ 3.”-part
of the proof of the Myhill-Nerode theorem, we see that the relation ≡M 0 is
a refinement of ∼L . Thus |Q0 | ≥ index(≡M 0 ) ≥ index(∼L ).
Part 2: Assume now that |Q0 | = |Q|. This means that index(∼L ) =
index(≡M 0 ). Since ≡M 0 is a refinement of ∼L , this means that the equiv-
alence classes of both relations are the same and hence both equivalence
relations are the same. In particular, we can just simply write [x] for the
equivalence class of x in any of the two relations. Let

b : Q0 → Q
q 0 7→ [x] where x is chosen such that (δ 0 )∗ (q00 , x) = q 0

Since |Q0 | = |Q|, every state of M 0 is reachable. Thus b is defined for

every q 0 ∈ Q0 . Second, we have to check that b is well-defined: If y fulfills
(δ 0 )∗ (q00 , y) = q 0 , too, then x ≡M 0 y and hence x ∼L y. Since ≡M 0 is a
refinement of ∼L , b is certainly surjective, and because |Q0 | = |Q|, it is a
bijection, too.
To show that M 0 is the same automaton as M (up to renaming of the
states), we have to show that

1. δ(q, σ) = b(δ 0 (b−1 (q), σ)) for all q ∈ Q, σ ∈ Σ, and

2. b(Q0acc ) = Qacc .

For the first statement, let q = [x], and let b−1 (q) = q 0 . Then b(δ 0 (q 0 , σ)) =
[xσ] by the definition of b. For the second statement, let q 0 ∈ Q0acc and let
(δ 0 )∗ (q00 , x) = q 0 . Then b(q 0 ) = [x]. x ∈ L(M 0 ) = L and thus [x] ∈ Qacc .
This argument can be reversed, and thus we are done.

Example E.9 Consider the language L = L(0∗ 10∗ 10∗ ). L is the language
of all words in {0, 1}∗ that contain exactly two 1’s. We claim that ∼L has
four equivalence classes:

Ai = {x ∈ {0, 1}∗ | the number of 1’s in x equals i} for i = 0, 1, 2

∗
A>2 = {x ∈ {0, 1} | the number of 1’s in x is > 2}

If the number of 1’s in a word x equals i ≤ 2, then xz ∈ L iff the number

of 1’s in z equals 2 − i. If a word x has > 2 1’s, then xz ∈ / L for any
z ∈ {0, 1}∗ . Thus A0 , A1 ,A2 , and A>2 are indeed the equivalence classes of
L. 0, 1, 11, and 111 are representatives of these classes. The corresponding
minimal automaton is shown in Figure E.1.

c Markus Bläser 2007–2015
196 E. The Myhill-Nerode theorem and minimal automata

1
0 0 0 0

1 1 1
[0] [1] [11] [111]

Figure E.1: The minimal automaton for L(0∗ 10∗ 10∗ ). It has four states that
count the number of 1’s.

E.3 An algorithm for minimizing deterministic finite

automata
This section was not treated in class and is not relevant for the exams.
How do we actually find the minimal automaton for some language L?
Let M = (Q, Σ, δ, q0 , Qacc ) be any finite deterministic automaton with L =
L(M ). We can assume that every state is reachable from q0 . We know
that ≡M is a refinement of ∼L . The equivalence classes of ≡M correspond
to the states of M . Thus we have to identify these groups of states that
form one equivalence class of ∼L . Let A and A0 be equivalence classes of
≡M and assume that q and q 0 are the correspoding states, i.e., x ∈ A ⇐⇒
δ(q0 , x) = q and x ∈ A0 ⇐⇒ δ(q0 , x) = q 0 . A and A0 are subsets of the same
equivalence class of ∼L iff for all z ∈ Σ∗ , δ ∗ (q, z) ∈ Qacc iff δ ∗ (q 0 , z) ∈ Qacc . A
proof or witness that two states are not equivalent is a z with δ ∗ (q, z) ∈ Qacc
and δ ∗ (q 0 , z) ∈
/ Qacc . If there is such a z, then there is such a z that is short.
∗ δ ∗ (q 0 , z) ∈
Lemma E.10 If there is a z with
|Q|

δ (q, z) ∈ Q∗ acc and / Qacc , then
there is a word z with |z | ≤ 2 such that δ (q, z ) ∈ Qacc and δ ∗ (q 0 , z 0 ) ∈
0 0 0 /
Qacc or vice versa.

Proof. Let s0 , s1 , . . . , st be the computation of M when it starts in q = s0

and reads z. Let s00 , s01 , . . . , s0t be the computation of M when it starts in
q 0 = s00 and reads z. If t ≤ |Q|


2 , then we are done. Otherwise, there are
0 0 |Q|


indices i and j, i < j, such that {si , si } = {sj , sj } since 2 is the number
of unordered pairs with elements from Q. If si = sj and s0i = s0j , then, as in
the proof of the pumping lemma, we can shorten the two computations by
leaving out the states si+1 , . . . , sj and s0i+1 , . . . , s0j . The corresponding word
is z 0 = z1 . . . zi zj+1 . . . zt . If si = s0j and s0i = sj , then s0 , . . . , si , s0j+1 , . . . , s0t
and s00 , . . . , s0i , sj , . . . , st are the computations of M on ẑ when started in
q = s0 and q 0 = s00 . But now δ ∗ (q, z 0 ) ∈ / Qacc and δ ∗ (q 0 , z 0 ) ∈ Qacc . If z 0 is
still longer than |Q|


2 , we can repeat the process.
To decide whether two states are
equivalent, we “just” have to check
whether for all z ∈ Σ∗ with |z| ≤ |Q|2 , δ ∗ (q, z) ∈ Q ∗ 0
acc iff δ (q , z) ∈ Qacc . If
q and q 0 are equivalent, then we can remove one of them, say q, and all arcs

c Markus Bläser 2007–2015
E.4. Further exercises 197

in the transition diagram that come into q now point to q 0 instead (i.e., if
δ(p, σ) = q, then δ(p, σ) = q 0 in the new automaton). The new automaton
has one state less, and we can go on until we do not find a pair of equivalent
states.
But there is a much faster algorithm. Basically, when we have such a z
that proves that q and q 0 are not equivalent, then all the pairs we go through
when reading z are not equivalent, too. Algorithm 1 constructs these pairs
backwards, starting from these pairs that have exactly one state in Qacc and
one state not in Qacc .

Algorithm 1 Minimization of deterministic finite automata

Input: a deterministic finite automaton M = (Q, Σ, δ, q0 , Qacc )
Output: the minimal deterministic finite automaton M̂ with L(M ) =
L(M̂ )
1: Initialize a Boolean table T of all unordered pairs from Q and set
T ({q, q 0 }) = 1 for all pairs {q, q 0 }.
2: Initialize a list L of unordered pairs from Q.
3: For each pair {q, q 0 } such that q ∈ Qacc and q 0 ∈ / Qacc or vice versa, set
0 0
T ({q, q }) = 0 and add {q, q } to L.
4: while L is not empty do
5: remove a pair {q, q 0 } from L.
6: for all pairs {p, p0 } with {δ(p, σ), δ(p0 , σ)} = {q, q 0 } for some σ ∈ Σ, if
T ({p, p0 }) = 1, then set T ({p, p0 }) = 0 and add {p, p0 } to L.
7: od
8: while there is a pair {q, q 0 } with T ({q, q 0 }) = 1 do
9: remove q from the set of states and set δ(p, σ) = q 0 for all p ∈ Q and
σ ∈ Σ with δ(p, σ) = q.
10: od
11: return the constructed automaton

Exercise E.2 1. Show that Algorithm 1 sets T (q, q 0 ) = 0 iff there is a z

with δ (q, z) ∈ Qacc and δ ∗ (q 0 , z) ∈
∗ / Qacc or vice versa.

2. Show that Algorithm 1 constructs indeed the minimal deterministic fi-

nite automaton equivalent to M .

E.4 Further exercises

Exercise E.3 Let Ln = {x ∈ {0, 1}∗ | the n-last symbol of x is a 1}.

1. Show that index(Ln ) ≥ 2n . Conclude that the power set construction

that transforms a nondeterministic finite automaton into a determin-
istic one is essentially optimal.

c Markus Bläser 2007–2015
198 E. The Myhill-Nerode theorem and minimal automata

2. Give a regular expression for Ln that consists of a linear number of

symbols.

c Markus Bläser 2007–2015
27 Grammars

When Noam Chomsky invented grammars, he wanted to study sentences

in natural languages. He wanted to formulate rules like a Satz in German
consists of a Subjekt followed by a Prädikat and then maybe followed by an
Objekt. A Subjekt consists of an Artikel and a Nomen. An Artikel can be
der, die, or das. A lot of words can be a Nomen, examples are Hund, Katze,
and Maus.1

Definition 27.1 A grammar G is described by a 4-tuple (V, Σ, P, S).

1. V is a finite set, the set of variables or nonterminals.

2. Σ is a finite set, the set of terminal symbols. We have V ∩ Σ = ∅.

3. P is a finite subset of (V ∪ Σ)+ × (V ∪ Σ)∗ , the set of productions.

4. S ∈ V is the start variable.

Convention 27.2 If (u, v) ∈ P is a production, we will often write u → v

instead of (u, v).

In the example above, Satz would be the start variable. Subjekt, Prädikat,
. . . would be variables. The letters “d”, “e”, “r”, . . . are terminal symbols.

Definition 27.3 1. A grammar G = (V, Σ, P, S) defines a relation ⇒G

on (V ∪ Σ)∗ as follows: u ⇒G v if we can write u = xyz and v = xy 0 z
and there is a production y → y 0 ∈ P . We say that v is derivable
from u in one step. v is derivable from u if u ⇒∗G v, where ⇒∗G is the
reflexive and transitive closure of ⇒G .

2. A word u ∈ (V ∪ Σ)∗ is called a sentence if S ⇒∗G u.

3. A sequence of sentences w0 , . . . , wt ∈ (V ∪ Σ)∗ with w0 = S, wτ ⇒G

wτ +1 for 0 ≤ τ < t, and wt = u is called a derivation of u. (A
derivation can be considered as a witness or proof that S ⇒∗G u.)

4. The language generated by G is L(G) = {u ∈ Σ∗ | S ⇒∗G u}. (Note

that words in L(G) do not contain any variables.)
1
Yes, I know, this is oversimplified and not complete.

199
200 27. Grammars

Example 27.4 Let G1 = ({S}, {0, 1}, P1 , S) where P1 consists of the pro-
ductions

S→ε
S → 0S1

Exercise 27.1 Prove by induction on i that 0i S1i is the only sentence of

length 2i + 1 with S ⇒∗G1 0i S1i . Conclude that L(G1 ) = {0i 1i | i ∈ N}.

Syntactic sugar 27.5 If P contains the productions u → v1 , . . . , u → vt

(with the same left-hand sides u), then we will often write u → v1 | · · · | vt .

Example 27.6 Let G2 = ({W, V, N, N + , Z, Z + }, {a, b, . . . , z, 0, 1, . . . , 9, :, =, 6=, ;,

+, −}, P2 , W ) where P2 consists of the productions

W → V :=V +V | V :=V −V | V :=N |

while V 6= 0 do W od |
W ;W
V → xN +
N + → Z | Z +N
N → Z | ZN
Z → 0 | 1 | ··· | 9
Z+ → 1 | 2 | · · · | 9

It is quite easy (but a little tedious) to see that L(G2 ) is the set of all WHILE
programs (now over a finite alphabet).2 From N + , we can derive all decimal
representations of natural numbers (without leading zeros). From V , we can
derive all variable names. From W , we can derive all WHILE programs.
The first three productions produce the simple statements, the other two the
while loop and the concatenation.

Example 27.7 Let G3 = ({S, E, Z}, {0, 1, 2}, P3 , S) where P3 is given by

S → 0EZ | 0SEZ
ZE → EZ
0E → 01
1E → 11
1Z → 12
2Z → 22
2
Because of the simple structure of WHILE programs, we do not even need whitespaces
to separate the elements. Feel free to insert them if you like.

c Markus Bläser 2007–2015
27.1. The Chomsky hierachy 201

Exercise 27.2 1. Prove by induction on n that S ⇒∗G3 0n 1n 2n for all

n ≥ 1.

2. Show that whenever S ⇒∗G3 w, then the number of 0’s in w equals the
number of E’s plus 1’s in w and it also equals the number of Z’s plus
2’s in w.

3. Show that whenever one uses the rule 1Z → 12 and there is an E to

the right of the 2 created, then one cannot derive a word from Σ∗ from
the resulting sentence.

4. Conclude that L(G3 ) = {0n 1n 2n | n ≥ 1}.

27.1 The Chomsky hierachy

Definition 27.8 Let G = (V, Σ, P, S) be a grammar.

1. Every grammar G is a type-0 grammar.

2. G is a type-1 grammar if |u| ≤ |v| for every production u → v ∈ P .

The only exception is the production S → ε. If S → ε ∈ P , then S
does not appear in the right-hand side of any production of P .

3. G is a type-2 grammar if it is type-1 and in addition, the left-hand

side of every production is an element from V .

4. G is a type-3 grammar if it is type-2 and in addition, the right-hand

side of every production is of the form ΣV ∪ Σ.

Definition 27.9 Let i ∈ {0, 1, 2, 3}. A language L ⊆ Σ∗ is a type-i lan-

guage if there is a type-i grammar G with L = L(G).

The grammar in the Example 27.6 is a type-2 grammar. Type-2 gram-

mars are also called context-free grammars and type-2 languages are called
context-free languages. The idea behind this name is that we can replace a
variable A in a sentence regardless of the context it is standing in.
Type-1 grammars are also called context-sensitive grammars and type-
1 languages are called context-sensitive languages. Here rules of the form
xAy → w are possible and we can replace A only if it stands in the context
xAy.3
Type-3 grammars are also called right-linear grammars: “linear”, because
the derivation trees (to be defined in the next chapter) degenerate essentially
3
The name context-sensitive is not too well chosen, type-0 grammars have the same
property, too, since they are more general. Nevertheless, the term context-sensitive is
reserved for type-1 grammars and languages. But the important property of type-1 gram-
mars is that they cannot shorten sentences by applying a production.

c Markus Bläser 2007–2015
202 27. Grammars

to a linear chain, and “right” because the variable stand at the right-hand
end of the productions. Type-3 grammars are called regular grammars, too.
Theorem 27.12 explains this: type-3 languages are exactly the regular lan-
guages. The grammar that we get from the grammar in Example 27.6 by
only taking the variables {N, N + , Z, Z + }, these productions that use the
variables {N, N + , Z, Z + }, and the start symbol N + generates the digital
representations without leading zeros of all natural numbers. It is “almost”
right linear. The variables Z and Z + are just place holders for a bunch of
terminals. The grammar gets right linear if we replace the productions of
the type N → ZN and N + → Z + N by productions of the form N → 0N ,
N → 1N , etc.

Definition 27.10 1. The set of all type-2 languages is denoted by CFL.

2. The set of all type-1 languages is denoted by CSL.

By definition, the set of all type-3 languages is a subset of all type-2

languages. This inclusion is strict, since {0n 1n | n ∈ N} is context-free but
not regular. The grammar G1 in Example 27.4 is an “almost context-free”
grammar for this language. The only problem is that we can derive ε from the
start symbol S and S is standing on the righthand side of some productions.
But for context-free grammars, this is not a real problem. The grammar

S → ε | S0
S 0 → 01 | 0S 0 1

is a type-2 grammar for {0n 1n | n ∈ N}. We will later see a general way
to get rid of productions of the form A → ε in an “almost context-free”
grammar. (Note that this is not possible for context-sensitive grammars!)

Syntactic sugar 27.11 When we write down a context-free grammar, we

often only write down the productions in the following. Then the following
conventions apply:

1. The symbols on the lefthand side are the variables.

2. All other symbols are terminals.

3. The lefthand side of the first productions is the start variable.

In the same way, type-2 languages are a subset of the type-1 languages.
The language {0n 1n 2n | n ≥ 1} is context-sensitive, as shown in Exam-
ple 27.7, but we will see soon that it is not context-free. Hence this inclusion
is also strict.
The set of all type-0 languages equals RE—a fact that we will not prove
here. On the other hand, CSL ⊆ REC: Given a string w ∈ Σ∗ , we can

c Markus Bläser 2007–2015
27.2. Type-3 languages 203

generate all derivations for words of length |w|, because once we reached a
sentence of length > |w| in the derivation, we can stop, since productions of
context-sensitive grammars can never shorten a sentence. Thus the type-1
languages are a strict subset of the type-0 languages.

27.2 Type-3 languages

Theorem 27.12 Let L ⊆ Σ∗ . L is a type-3 language iff L is regular.

Proof. “⇒”: Let G = (V, Σ, P, S) be a type-3 grammar with L(G) = L.

We will construct a nondeterministic finite automaton M with L(M ) =
L. We first assume that ε ∈ / L(G). Let F ∈ / V . We set M = (V ∪
{F }, Σ, δ, S, {F }) where
(
{B | A → σB ∈ P } if A → σ ∈
/P
δ(A, σ) =
{B | A → σB ∈ P } ∪ {F } otherwise.

Let w = w1 w2 . . . wn ∈ Σ∗ \ {ε}. We have

w ∈ L(G) ⇐⇒ there are variables V1 , . . . , Vn−1 ∈ V with

S ⇒G w1 V1 ⇒G w1 w2 V2 ⇒G . . . ⇒G w1 w2 . . . wn−1 Vn−1 ⇒G w1 w2 . . . wn−1 wn
⇐⇒ there are states V1 , . . . , Vn−1 ∈ V with
V1 ∈ δ(S, w1 ), V2 ∈ δ(V1 , w2 ), . . . , Vn−1 ∈ δ(Vn−2 , wn−1 ), F ∈ δ(Vn−1 , wn )
⇐⇒ w ∈ L(M ).

If ε ∈ L(G), then we first construct an automaton M for L(G) \ {ε} first.4

Then we add a new start state that is also an end state and add an ε-
transition to the old start state of M . This automaton recognizes L(M ) ∪
{ε} = L(G).
“⇐”: is shown in Exercise 27.3.

Exercise 27.3 Let M = (Q, Σ, δ, q0 , Qacc ) be a deterministic finite automa-

ton. Let G = (Q, Σ, P, q0 ) where the set of productions contains a production

q → σq 0 for all q, q 0 ∈ Q and σ ∈ Σ with δ(q, σ) = q 0

and in addition

q→σ for all q ∈ Q, q 0 ∈ Qacc , and σ ∈ Σ with δ(q, σ) = q 0 .

If q0 ∈ Qacc , we will add the production q0 → ε, too. Show that L(G) =

L(M ).
4
This is possible, since we get a type-3 grammar for L(G) \ {ε} by removing the
production S → ε.

c Markus Bläser 2007–2015
204 27. Grammars

Type-3 languages (regular languages)

Let L ⊆ Σ∗ . The following statements are equivalent:

• There is a right-linear grammar G with L(G) = L.

• There is a deterministic finite automaton M with L(M ) = L.

• There is a nondeterministic finite automaton M with L(M ) =

L.

• There is a regular expression E with L(E) = L.

• ∼L has finite index.

Although all these concepts describe regular languages, they have different
properties: nondeterministic finite automata often have much fewer states
than deterministic ones for the same language. Deciding whether two de-
terministic finite automata recognize the same language is easy whereas
this is a hard problem for regular expressions (we will see this later on),
etc.

Exercise 27.4 A grammar G = (V, Σ, P, S) is called left-linear if it is type-

2 and the right-hand sides of all productions are of the form V Σ ∪ Σ. Let
L ⊆ Σ∗ . Show that there is a left-linear grammar G with L(G) = L iff L is
regular.

c Markus Bläser 2007–2015
28 Context-free grammars

By definition, context-free grammars have at most one ε-production, namely,

S → ε, and if this production is present, then S is not allowed to appear
on the righthand side of any production. It is often convenient to allow
arbitrary ε-productions. For instance, the grammar

S → ε | 0S1

looks much nicer than the one given by

S0 → ε | S
S → 01 | 0S1

For context-free grammars (but not for context-sensitive grammars!), we

can allow arbitrary ε-productions. So from now on, we allow arbitrary ε-
productions in context-free grammars. We will see that we can always find
an equivalent grammar that has at most one ε-production of the form S → ε.

28.1 Derivation trees and ambiguity

Consider the grammar G given by

E → E ∗ E | E + E | (E) | x

It generates all arithmetic expressions with the operations ∗ and + over the
variable1 x.2 A word w ∈ Σ∗ is in L(G) if S ⇒∗ w.3 A derivation is a
witness for the fact that S ⇒∗ w, i.e., a sequence of sentences such that
S ⇒ w1 ⇒ w2 ⇒ . . . ⇒ wt ⇒ w. Usually, a word w has many derivations.
Here are two examples for the word x + x ∗ x in the example above:

E ⇒ E+E ⇒ x+E ⇒ x+E∗E ⇒ x+x∗E ⇒ x+x∗x (28.1)

E ⇒ E + E ⇒ E + E ∗ E ⇒ E + E ∗ x ⇒ E + x ∗ x ⇒ x + x ∗ x (28.2)
1
This is a variable in the expression. Do not confuse this with a variable of the grammar.
2
Yes, I know, arithmetic expressions over one variable without any constants are not
really exciting. But you can replace x by your favourite collection of variables and con-
stants.
3
Whenever the grammar is clear from the context, we will write S ⇒∗ w instead of
S ⇒∗G w.4
4
This could be a record number of footnotes per page.

205
206 28. Context-free grammars

In the first derivation, we always replace the leftmost variable. Such deriva-
tions are called leftmost derivations. In the second derivation, we always
replace the rightmost variable. Such derivations are called, guess what, right-
most derivations. Although the two derivations are different, they are not
“really different”, they correspond to the same derivation tree.

Definition 28.1 Let G = (V, Σ, P, S) be a context-free grammar.

1. A derivation tree (or parse tree) is an ordered tree with a node labeling
such that:

(a) The root is labeled with S.

(b) All leaves are labeled with an element from V ∪ Σ or with ε. In
the latter case, the leaf is the only child of its parent.
(c) All interior nodes are labeled with an element of V . If A is this
label and the labels of the children are x1 , x2 , . . . , xt (in this order)
then A → x1 x2 . . . xt ∈ P .5

2. The yield (or leaf-word or front or . . . )6 of a parse tree is the concate-

nation of the labels of the leaves (in the order induced by the ordering
of the vertices in the tree).

Figure 28.1 shows the derivation tree that corresponds to the two deriva-
tions (28.1) and (28.2). The leftmost derivation (28.1) is obtained by doing
a depth-first search and visiting the children from left to right, the rightmost
derivation (28.2) is obtained by visiting them from right to left. In general,
each derivation tree corresponds to exactly one left derivation and exactly
one right derivation.
But there is another derivation tree for x+x∗x. It is shown in Figure 28.2.
Having several derivation trees for the same word is in general a bad thing.

Definition 28.2 1. A context-free grammar is called ambiguous if there

is a word that has two or more derivation trees. Otherwise the grammar
is called unambiguous.

2. A context-free language is called unambiguous if there is an unam-

biguous grammar that generates the language. Otherwise it is called
inherently ambiguous.

The derivation tree in Figure 28.2 is unnatural, because it does not re-
spect the usual precedence of the operators “∗” and “+”. But there is an
5
If t > 1, then every xτ ∈ V ∪ Σ. If t = 1, then x1 = ε is possible. In this case, A → ε
is a production of P .
6
There are far too many names for this.

c Markus Bläser 2007–2015
28.1. Derivation trees and ambiguity 207

E + E

x E ∗ E

x x

Figure 28.1: A derivation tree for x + x ∗ x.

E ∗ E

E + E x

x x

Figure 28.2: Another derivation tree for x + x ∗ x.

c Markus Bläser 2007–2015
208 28. Context-free grammars

unambiguous grammar:

E →T |T +E
T →F |F ∗T
F → x | (E)

It is by no means obvious that the grammar above is unambiguous and this

fact requires a formal proof. The proof is rather tedious and can be shown
as follows:
• There is only one derivation tree for T + T + · · · + T .

• There is only one derivation tree for deriving F ∗ F ∗ · · · ∗ F from T .

(This means that the root of the tree is labeled by T .) In turn, there
is only one derivation tree for deriving e1 ∗ e2 ∗ · · · ∗ et from T where
each eτ is either x or (E).

• Now we prove by the number of pairs of brackets that there is at most

one derivation tree for every sentence: Take a pair of matching brackets
in a sentence s. The subsentence s0 between these brackets has to be
derivable from E, since F → (E) is the only rule that introduces
brackets and the chosen pair of brackets was a matching one. By
the induction hypothesis, there is only one derivation tree for s0 . If
we replace the sentence (s0 ) by F in s, then again by the induction
hypothesis, there is only one derivation tree for this sentence.
Note that simply respecting the precedence of operators is not enough for
making the grammar unambiguous: If we had taken the rule T → F | T ∗ T ,
then the grammar still would be ambiguous.
There are context-free languages that are inherently ambiguous.
Theorem 28.3 (without proof ) The language {0n 1n 2m 3m | n, m ≥ 1} ∪
{0n 1m 2m 3n | n, m ≥ 1} is context-free and inherently ambiguous.

Exercise 28.1 Show that the language from Theorem 28.3 is context-free.

28.2 Elimination of useless variables

This section was not treated in class and is not relevant for the exams.
Assume you created a big context-free grammar for some, say, program-
ming language. Can you find out whether everything is really needed in your
grammar or are there some artifacts in it, i.e., variables that you do not need
any more because you later changed some things somewhere else and now
these variables do not occur in any derivation of a terminal word?

Definition 28.4 Let G = (V, Σ, P, S) be a context-free grammar.

c Markus Bläser 2007–2015
28.2. Elimination of useless variables 209

1. A variable A is generating if there is a word w ∈ Σ∗ such that A ⇒∗G w.

2. A variable A is reachable if there are words x, y ∈ (V ∪ Σ)∗ such that
S ⇒∗G xAy.
3. A variable A ∈ V is useful if it is reachable and generating. Otherwise,
A is useless.
Theorem 28.5 Let G = (V, Σ, P, S) be a context-free grammar with L(G) 6=
∅. Let H = (W, Σ, Q, S) be the grammar that is obtained as follows:
1. Remove all variables A ∈ V that are not generating and all productions
that contain A. Let the resulting grammar be G0 = (V 0 , Σ, P 0 , S).
2. Remove all variables A ∈ V 0 that are not reachable in G0 and all pro-
ductions that contain A.
Then L(H) = L(G) and H does not contain any useless variables.
Proof. We start with some simple observation: If S ⇒∗ uAv ⇒∗ w with
u, v ∈ (V ∪ Σ)∗ and w ∈ Σ∗ , then A is both reachable and generating. The
fact that A is reachable is obvious. But A is also generating, since we can
derive some substring of w from it.
Next we show that H contains no useless variables:
Let A ∈ W . Since A survived the first step (as it is in W , it has to be in
V 0 ), there is a w ∈ Σ∗ such that A ⇒∗G w. Every variable in a corresponding
derivation is also generating since we can derive a subword of w from it
(cf. the observation above) and therefore, every variable in the derivation
survives the first step and is in V 0 . Thus A ⇒∗G0 w and A is also generating
in G0 .
A variable A ∈ W is certainly reachable in G0 since it survived the second
step. This means that there are u, v ∈ (V 0 ∪Σ)∗ such that S ⇒∗G0 uAv. (Note
that S is still in G0 , since L(G) 6= ∅.) But every variable in a correspoding
derivation is also reachable in G0 . Hence all variables in the derivation are in
H and therefore, S ⇒∗H uAv. We know that every variable in the derivation
is generating in G0 , thus S ⇒∗H uAv ⇒∗G0 w for some w ∈ Σ∗ . But this
means that every variable in a derivation corresponding to uAv ⇒∗G0 w is
reachable from S in G0 . Hence all of them are in H, too, and we have
S ⇒∗H uAv ⇒∗H w. Thus A is generating and reachable, hence it is useful.
It remains to show that L(G) = L(H):
“⊇”: Is obvious since we only remove variables from G.
“⊆”: If S ⇒∗G w for some w ∈ Σ∗ , then every variable in a derivation of w
is both reachable and generating in G by the observation at the beginning
of the proof. So all variables in the derivation survive the first step and are
in G0 . But since all variables are in G0 , we still have S ⇒∗G0 w and thus all
variables in the derivation survive the second step. Therefore S ⇒∗H w and
w ∈ L(H).

c Markus Bläser 2007–2015
210 28. Context-free grammars

Example 28.6 Consider the following grammar

S → AB | 0
A→0

We cannot derive any terminal word from B, hence we remove the production
S → AB. Now we cannot derive any sentence of the form xAy from S, hence
we remove the rule A → 0, too. If we had reversed the order of the two steps,
then we would not have removed anything in the first step and only the rule
S → AB in the second step. The production A → 0 would not have been
removed.

Theorem 28.5 provides a way to eliminate useless symbols once we can

determine the generating and reachable variables. Algorithms 2 and 3 solve
these two tasks.

Exercise 28.2 Show that the Algorithms 2 and 3 are indeed correct.

Algorithm 2 Determining the generating variables

Input: A context-free grammar G = (V, Σ, P, S)
Output: The set V0 of all variables that are generating.
1: Add all A ∈ V to V0 for which there is a production A → u with u ∈ Σ∗ .

2: while there is a production A → a1 a2 . . . at such that A ∈

/ V0 and all aτ
that are in V are in V0 do
3: Add A to V0 .
4: od
5: Return V0 .

Algorithm 3 Determining the reachable variables

Input: A context-free grammar G = (V, Σ, P, S)
Output: The set V1 of all variables that are reachable
1: Add all A ∈ V to V1 for which there is a production S → xAy for some
x, y ∈ (V ∪ Σ)∗ .
2: while there is a production A → a1 a2 . . . at such that A ∈ V1 and some
aτ is in V \ V1 do
3: Add all aτ ∈ V \ V1 to V1 , 1 ≤ τ ≤ t.
4: od
5: Return V1 .

c Markus Bläser 2007–2015
29 Chomsky normal form

In this chapter, we show a formal form for context-free grammars, the so-
called Chomsky normal form. On the way, we also see how to eliminate
ε-productions.

29.1 Elimination of ε-productions

Definition 29.1 Let G = (V, Σ, P, S) be a context-free grammar. A ∈ V is
called nullable if A ⇒∗G ε.

Theorem 29.2 Let G = (V, Σ, P, S) be a context-free grammar. Let H =

(V, Σ, Q, S) be the grammar that is generated as follows:

1. Replace every production A → a1 a2 . . . ak by all 2` productions, one for

each possibility to leave an aiλ out where ai1 , . . . , ai` are all nullable
variables among a1 , a2 , . . . , ak .

2. Remove every ε-production. (This removes also an ε-production that

we might have introduced in the first step when a1 , . . . , ak are all nul-
lable.)

We have L(G) \ {ε} = L(H).

Proof. We show the following:

For all A ∈ V and u ∈ (V ∪ Σ)∗ : A ⇒∗H u ⇐⇒ A ⇒∗G u and u 6= ε.

From this statement, the claim of the theorem follows:

“=⇒”: The proof is by induction on the length of the derivation:
Induction base: If A → u ∈ Q, then u 6= ε by construction. There is a
production A → a1 a2 . . . at , each aτ ∈ V ∪ Σ, and indices j1 , . . . , j` such
that the concatenation of all aτ with τ ∈ / {j1 , . . . , j` } is u and all aτ with
∗
τ ∈ {j1 , . . . , j` } are nullable. Thus A ⇒G u.
Induction step: If A ⇒∗H u, then A ⇒∗H w ⇒H u. This means that w = xBz
such that B → y ∈ Q and u = xyz. As in the proof of the induction base,
we can show that B ⇒∗G y holds. By the induction hypothesis, A ⇒∗G xBz.
Altogether, A ⇒∗G xyz = u.
“⇐=”: Is left as an exercise.

Exercise 29.1 Show the “⇐=”-direction of the proof of Theorem 29.2.

211
212 29. Chomsky normal form

Theorem 29.2 provides a way to eliminate ε-productions once we can

determine the nullable variables. Algorithm 4 solves the latter task.
Exercise 29.2 Show that Algorithm 4 is indeed correct.

Algorithm 4 Determining the nullable variables

Input: A context-free grammar G = (V, Σ, P, S)
Output: The set V0 of all variables that are nullable.
1: Add all A ∈ V to V0 for which there is a production A → ε ∈ P .
2: while there is a production A → a1 a2 . . . at such that A is not in V0 and
all aτ are nullable do
3: Add A to V0 .
4: od
5: Return V0 .

29.2 Elimination of chain productions

Definition 29.3 Let G = (V, Σ, P, S) be a context-free grammar. A produc-
tion of the form A → B with A, B ∈ V is called a chain production.
Like ε-productions, chain productions are useful for getting compact
grammars; E → T | E ∗ T is an example. On the other hand, like ε-
productions, chain productions are not desirable, because they do not gen-
erate anything really new. But again, there is a way to get rid of chain
productions.
First of all, we can immediately remove all productions of the form A →
A. We build a directed graph H = (V, E). There is an edge (A, B) ∈ E
if there is a chain rule A → B ∈ P . (Recall that productions are tuples,
therefore we can also write E = P ∩ (V × V ).) If H has a directed cycle,
then there are productions Bτ → Bτ +1 ∈ P , 1 ≤ τ < t, and Bt → B1 ∈ P .
But this means that the variables B1 , . . . , Bt are interchangable. Whenever
we have a sentence that contains a variable Bi we can replace this by any
Bj by using the chain productions.
Exercise 29.3 Let G = (V, Σ, P, S) be a context-free grammar and let H =
(V, P ∩ V × V ). Assume that there is a directed cycle in H consisting of
nodes B1 , B2 , . . . , Bt , t ≥ 2. Let G0 = (V 0 , Σ, P 0 , S) be the grammar that
we obtain by replacing all occurrences of a Bi by B1 , removing the variables
B2 , . . . , Bt from V and the production of the form B1 → B1 . Show the
following: L(G) = L(G0 ).
If we apply the construction above several times, we obtain a grammar,
call it again G = (V, Σ, P, S), such that the corresponding graph H = (V, P ∩
V × V ) is acyclic.

c Markus Bläser 2007–2015
29.3. The Chomsky normal form 213

Theorem 29.4 Let G = (V, Σ, P, S) be a context free grammar such that

the graph H = (V, P ∩ V × V ) is acyclic. Then there is a grammar G0 =
(V, Σ, P, S) without chain productions with L(G) = L(G0 ).

Proof. The proof is by induction on the number of chain productions in

P (or the number of edges in H).
Induction base: If there are no chain productions, then there is nothing to
prove.
Induction step: Since H is acyclic and contains at least one edge, there must
be one variable A that has indegree ≥ 1 but outdegree 0. Let B1 , . . . , Bt
all variables such that Bτ → A ∈ P . Let A → u1 , . . . , A → u` ∈ P all
productions with lefthand side A. Since A has outdegree 0 in H, uλ ∈ / V
for all 1 ≤ λ ≤ `. We remove the productions Bτ → A and replace them by
Bτ → uλ for 1 ≤ τ ≤ t and 1 ≤ λ ≤ `. Let G00 be the resulting grammar.
Since we removed at least one chain production and did not introduce any
new ones, G00 has at least one chain production less than G. By the induction
hypothesis, there is a grammar G0 without any chain productions such that
L(G00 ) = L(G0 ). Hence we are done if we can show that L(G00 ) = L(G):
If S ⇒∗G w for some w ∈ Σ∗ and the production Bτ → A is used in the
corresponding derivation, then eventually, a production A → uλ has to be
used, too, since w ∈ Σ∗ . Hence we can use the production Bτ → uλ directly
and get a derivation in G0 . Conversely, if S ⇒∗G0 w and the production
Bτ → uλ is used in a corresponding derivation, then we can replace this step
by two steps that use the productions Bτ → A and A → uλ .

29.3 The Chomsky normal form

Definition 29.5 A context-free grammar G = (V, Σ, P, S) is in Chomsky
normal form if all its productions are either of the form

A → BC

or

A→σ

with A, B, C ∈ V and σ ∈ Σ.

Theorem 29.6 For every context-free grammar G = (V, Σ, P, S) with L(G) 6=

∅ there is a context-free grammar G0 = (V 0 , Σ, P 0 , S) in Chomsky normal
form with L(G0 ) = L(G) \ {ε}.

Proof. By the result of the previous sections, we can assume that G does
not contain any ε-productions and chain rules. Thereafter, L(G) does not

c Markus Bläser 2007–2015
214 29. Chomsky normal form

contain the empty word anymore. For every σ ∈ Σ, we introduce a new

variable Tσ , add the new production Tσ → σ, and replace every occurence
of σ in the productions in P by Tσ except in productions of the form A → σ
(since this would introduce new chain productions, but A → σ has already
the “right” form for Chomsky normal form).
Thereafter, every productions is either of the form A → σ or A →
A1 A2 . . . At where A1 , A2 , . . . , At are all variables and t ≥ 2. Hence we are
almost there except that the righthand sides might have too many variables.
We can overcome this problem by introducing new variables C2 , . . . , Ct−1
and replacing the production A → A1 A2 . . . At by

A → A1 C 2
C 2 → A2 C 3
...
Ct−2 → At−2 Ct−1
Ct−1 → At−1 At

The resulting grammar G0 is obviously in Chomsky normal form. It is easy

to see that L(G0 ) = L(G) \ {ε} (see Exercise 29.4).

Exercise 29.4 Prove that the grammar G0 constructed in the proof of The-
orem 29.6 indeed fulfills L(G0 ) = L(G) \ {ε} (and even L(G0 ) = L(G), since
we assumed that G does not contain any ε-productions).

Exercise 29.5 Let G be a context-free grammar and let H be the grammar

in Chomsky normal form constructed in this section. If G has p productions,
how many productions can H have?

29.4 Further exercises

A context-free grammar G = (V, Σ, P, S) is in Greibach normal form if for
every A → v ∈ P , v ∈ ΣV ∗ .

Exercise 29.6 Show that for every context-free grammar G, there is a context-
free grammar H in Greibach normal form such that L(G) \ {ε} = L(H).
(Hint: First convert into Chomsky normal form.)

c Markus Bläser 2007–2015