DFA minimisation using the Myhill-Nerode theorem

Johanna Hogberg

Lars Larsson

Abstract
The Myhill-Nerode theorem is an important characterisation of
regular languages, and it also has many practical implications.
In this chapter, we introduce the theorem an present its proof.
One consequence of the theorem is an algorithm for minimising
DFAs that is outlined in the latter part of this paper. To clarify
how the algorithm works, we conclude with an example of its
application.

Introduction

This chapter gives an introduction to DFA minimisation using the MyhillNerode theorem. DFA minimisation is an important topic because it can be
applied both theoretically and practically, in for instance compilers. Minimising a DFA increases its efficiency by reducing its amount of states and it also
enables us to determine if two DFAs are equivalent.
The chapter is structured as follows. Following this introduction in Section 2,
some relevant background information and definitions are stated, which will be
used in the subsequent sections. In Section 3, the Myhill-Nerode theorem is
stated and its proof presented. Section 4 contains the algorithm for DFA minimisation that uses the Myhill-Nerode theorem. An example showing how such
a minimisation might be conducted in a specific case is found in Example 4.2.

Background information and definitions

The reader is advised to skim through this section to recall, at least at an

intuitive level, the terms that will be used in this chapter. Following these
intuitive descriptions, more formal definitions are provided.
A deterministic finite automaton, or DFA for short, may be considered a kind
of very basic computational device. It processes strings of symbols of a certain
alphabet and yields a result that is said to be either accepting or rejecting.
Normally, an accepting result means that the string is part of a recognised
language. The language is then defined as all strings that bring the DFA into
an accepting state.

a
q1

q2

q3

q4

Figure 1: DFA accepting the language an , n > 2

Definition 2.1 (DFA) A deterministic finite automaton (DFA) A is a quintuple A = (S, , , s, F ) consisting of:

a
a
a
a
a

finite set of states S,

finite set of symbols , called the alphabet,
transition function : S 7 S,
starting state s S, and
set of accepting states F S.

We extend the definition of to the domain of by letting

1. (q, ) = q, where denotes the empty string, and
2. (q, aw) = ((q, a), w), where a and w .
Finally, A is said to accept the string w if (q0 , w) F .
DFA are commonly represented by state diagrams, such as in Figure 1 and
Figure 2. The states that are accepting have a double border, the initial state
is the state with the angle on the left and the arrows or transitions signify what
state should be entered after reading the symbol on the arrow.
There are some limitations as to what kind of language the DFA can recognise.
The number of symbols in the alphabet as well as the number of states in the
DFA has to be finite. This means that while it is entirely possible to construct
a DFA that accepts strings that for instance look like an , n > 2 (let the starting
state of the DFA be followed by two rejecting states and finally by a state that
accepts any number of as, see Figure 1) it is not possible to recognise a language
such as an bn , n N (the symbol a repeated n times, followed by the exact same
number of bs) because that would require infinitely many states to keep track of
the number of as to ensure that the same number of bs are read. See Figure 2
for an example of how a DFA accepting the language an bn , n 2 might be
constructed. Note the obvious difficulty that arises if n may be arbitrarily large
as well as the difference between this language and an .
The class of languages that can be recognised by a DFA are called regular
languages or Type-3 languages in the Chomsky hierarchy. Every regular language can be recognised by a read-only Turing machine, described by regular
expression, generated by a regular grammar, and defined in a so-called monadic
second-order logic.
As far as this chapter is concerned, the formal definition including syntax for
regular languages has no useful meaning. Instead, it is sufficient to define
regular languages as follows.
2

a
q00

q10

q20

b
a

q11

q21

b
a

a, b

qbin

a, b

q22

a, b
Figure 2: DFA accepting the language an bn , n 2
Definition 2.2 (regular language) The class of languages that can be recognised by a deterministic finite automaton.
The Myhill-Nerode theorem that will provide the theoretical background that
allows us to minimise a DFA requires the following definitions that are related
to equivalence relations.
Definition 2.3 (equivalence relation) An equivalence relation, , is a binary relation on a set if it is
Reflexive (a a)
Symmetric (if a b then b a)
Transitive (if a b and b c then a c)
Definition 2.4 (equivalence classes, index) An equivalence relation on
a set S imposes a partition S/ on S, such that elements e, e0 S are in the
same part of S/ if e e0 . The parts of S/ are called equivalence classes, and
we write [e] to denote the unique equivalence class of S/ to which e belongs.
The index of equivalence relation is |S/|.
Definition 2.5 (right invariant) An equivalence relation on strings of symbols from some alphabet is said to be right invariant if x, y with x y
and all w we have that xw yw.
This last definition states that an equivalence relation has the right invariant
property if two equivalent strings (x and y) that are in the language still are
equivalent if a third string (w) is appended to the right of both of them.
3

In the rest of the chapter, the following two equivalence relations will be of
particular importance.
Definition 2.6 (equivalence relation for DFA:s) Let A be a DFA. For strings
x, y , let x A y iff the same state in the DFA is reached from the initial
state by reading the strings x and y.
Definition 2.7 (equivalence relation for languages) Let L be any language
over . For strings x, y , let x L y iff w we have that xw L
yw L.
Definition 2.8 (distinguishing string) Let A = (Q, , , q0 , F ) be a DFA.
The string w is a distinguishing string for states p, q Q if exactly one of
(p, w) and (q, w) is in F .
Intuitively, a distinguishing string s for states p and q in an automaton A is
string which is mapped to an accepting state from exactly one of p and q. In
other words, the existence of a distinguishing string for a pair of states p, q Q
proves that p and q are not equivalent.

Myhill-Nerode theorem

In this section, the Myhill-Nerode theorem [Ner58] and its proof are presented.
Theorem 3.1 Let L . Then the following statements are equivalent:
1. There is a DFA that accepts L (L is regular).
2. There is a right-invariant equivalence relation of finite index such that
L is the union of some of the equivalence classes of .
3. L is of finite index.
Proof
The proof that we give here is an adaption of that in [Vis05], which
uses the implication order (1) (2) (3) (1).
(1) (2). Since L is a regular language, there exits a DFA A = (Q, , , q0 , F )
that recognises L. Recall that A is a right-invariant relation of finite index.
It remains to observe that
[
L=
[x] ,
x:(q0 ,x)F

where x .
(2) (3). Assume that L is the union of some equivalence classes of a rightinvariant relation of finite index. Now, if (the partition imposed by) is a
refinement of (the partition imposed by) L , then L must be of finite index
since it has fewer equivalence classes than , which by assumption is of finite
index. As we shall see, is indeed a refinement of L .
Since is right-invariant, we have that for every pair x, y such that x y,
it must hold that (z) xz yz. Moreover, we have that L is the union of some
4

equivalence classes of , so if x y, then x L y L. Combining these

implications gives us
x y (z) xz yz (z) xz L yz L x L y.
We now know that x y implies x L y, and draw the conclusion that is a
refinement of L .
(3) (1). To prove that if L is of finite index, then L is regular, it suffices
to construct, for an arbitrary L , a DFA A which recognises L. The idea that
underlies the construction is to use the equivalence classes of L as states in
A. First, we choose x1 , . . . , xk as representatives for the k equivalence classes
of L , and then assemble the DFA A = (Q, , , q0 , F ), where
Q = {[x1 ], . . . , [xk ]},
([x], a) = [xa],
q0 = [], where denotes the empty string, and
F = {[x] | x L}.
It should be obvious that is well defined and that (q0 , x) F iff x L.
Please note that the Myhill-Nerode theorem lets us decide whether a language
L is regular, by determining if either of Condition 2 or Condition 3 is met.

DFA minimisation

Not only does the Myhill-Nerode Theorem provide an informative characterisation of the recognisable languages, but it also suggests an algorithm for minimising DFAs. Before we attempt to describe this algorithm, we wish to make
a few definitions that will shorten and simplify the later discussion.
Definition 4.1 (partition, block) Let S be a set. A (finite) partition of S
is a set P of mutually disjoint non-empty sets {S1 , . . . , Sk }, such that
Si S, where 1 i k, and
the union of the sets in the partition equals the whole of S. Formally,
k
[

Si = S .

i=1

The components S1 , . . . , Sk are called the blocks or parts of P . If s S, then

we denote by [s]P the unique block of P to which s belongs. When P is not
important, or follows from the context, we omit the subscript of [s]P .
In the proof of Theorem 3.1, it is established that if L is a regular language, and
the index of L is iL N, then it is both necessary and sufficient for a DFA to
have iL states in order to recognise L correctly. Now, let A = (Q, , , q0 , F ) be a
DFA recognising the regular language L(A). Furthermore, let [q], where q Q,
be the set of strings {x | x (q0 , x) = q}. We note that {[q] | q Q}
5

partitions . We observe that for every q Q, and x, y [q], the equivalence

x L(A) y holds and draw the conclusion that the partition imposed by Q is a
refinement of the partition imposed by L(A) . This leads us to believe that if
the the number of states of an automaton A exceeds the number of equivalence
classes of L(A) , then the automaton can be reduced in size without altering
the language that it recognises.
A first attempt to minimise A could be to find, for each equivalence class C
in the partition of that follows from L(A) , the subset Q0 of Q such that
q0 Q0 [q 0 ] = C and merge the states of Q0 into a single state. The problem with
this approach is that to be absolutely sure that two states p and q can be merged,
we must examine (p, w) and (q, w) for every string in {w | |w| |Q|} .
A more efficient method is to start by assuming that all accepting states can
be merged into one, and all rejecting into another, and then refine this initial
partition until it is stable. By stable, we mean that if states p and q both belong
to block B, then for all , the state p goes to block B 0 on , if and only if
q goes to B 0 on . If the refinement process is allowed to continue long enough,
then the partition will eventually become a set of singletons, in which case it is
stable and termination guaranteed.
The minimisation algorithm that stems from the latter method is listed in
Algorithm 1 on page 10. Although it is obvious that a minimal automaton may
not contain unreachable states, the partitioning procedure alone will not purge
such states from the input automaton A. Instead, the removal of unreachable
states is done in RemoveUnreachableStates, while Collapse merges the
states of A in accordance with the partitioning produced by RefinePartition.
b
a
a

q2

a
q3

b
b

a
b

q1

q6

q4

b
q5

Figure 3: The DFA that is given as input to Algorithm 1 in Example 4.2.

Example 4.2 We shall now illustrate how RefinePartition of Algorithm 1
finds the coarsest stable partition with a running example. The input DFA
with alphabet = {a, b} is given in Figure 3, and the reader should note that
all of its states are reachable from the initial state q1 . The algorithm begins
with the trivial partition,
P0 = {I, II}, where I = {q1 , q3 , q4 , q5 }, and II = {q2 , q6 } ,
and determines, for each q Q and , the block [(q, )]P0 .
6

I:

[(q1 , a)] = [q2 ] = II

[(q3 , a)] = [q4 ] = I
[(q4 , a)] = [q4 ] = I
[(q5 , a)] = [q6 ] = II

[(q1 , b)] = [q5 ] = I

[(q3 , b)] = [q3 ] = I
[(q4 , b)] = [q3 ] = I
[(q5 , b)] = [q1 ] = I

II:

[(q2 , a)] = [q3 ] = I

[(q6 , a)] = [q5 ] = I

[(q2 , b)] = [q6 ] = II

[(q6 , b)] = [q2 ] = II

The partition P0 is not stable, as q1 and q5 goes to a rejecting state on a, while

q3 and q4 goes to an accepting state on the same input symbol. As the states in
II behave identically on both input symbols, only I has to be refined, yielding
the next partition in sequence
P1 = {I, II, III}, where I = {q1 , q5 }, II = {q2 , q6 }, and III = {q3 , q4 } .
P1 is also unstable, which the algorithm determines by examining the behaviour
of all q Q for all input symbols.
I:

[(q1 , a)] = [q2 ] = II

[(q5 , a)] = [q6 ] = II

[(q1 , b)] = [q5 ] = I

[(q5 , b)] = [q1 ] = I

II:

[(q2 , a)] = [q3 ] = III

[(q6 , a)] = [q5 ] = I

[(q2 , b)] = [q6 ] = II

[(q6 , b)] = [q2 ] = II

III:

[(q3 , a)] = [q4 ] = III

[(q4 , a)] = [q4 ] = III

[(q3 , b)] = [q3 ] = III

[(q4 , b)] = [q3 ] = III

In this iteration, the only states that belong to the same block and disagrees
are q2 and q6 of block II. After II is divided into {q2 } and {q6 }, partition
P2 = {I, II, III, IV},
where I = {q1 , q5 }, II = {q2 }, III = {q3 , q4 } and IV = {q6 }
remains. When deciding whether P2 is stable, the algorithm could in theory
take a shortcut and only examine block I, as it contains all states that were
mapped to block IIIP1 in the previous iteration, and is hence the only block
that could have been affected by the refinement of IIIP1 . Indeed, it turns out
that I has indeed become inconsistent.
I:

[(q1 , a)] = [q2 ] = II

[(q5 , a)] = [q6 ] = IV

[(q1 , b)] = [q5 ] = I

[(q5 , b)] = [q1 ] = I

Block I is consequently split, yielding

P3 = {I, II, III, IV, V },
where I = {q1 }, II = {q2 }, III = {q3 , q4 }, IV = {q6 } and V = {q5 } .

The algorithm would need another iteration to notice that the partition does
not change and that it is done, but we can establish this by a simple observation.
7

The only blocks that could have been affected by the last refinement are I, II
and IV, but these are all singletons and cannot be refined further in any case.
It appears that the original DFA was close to minimal, as the minimised version contains only one state less. If we analyse the automaton, we see that it
recognises the language L, given by
L = L1 L2 , where
L1 = {w | |w|a is even}, and
L2 = {w | w = uav, where |u|b is even, and |u|a is odd.} ,
where |w|a denotes the number of occurences of a in w. In the light of this,
states q1 , q2 , q5 and q6 all remember a unique aspect of the the input string w
read so far. State q1 remembers that w has an even number of as and bs, q2
that it has and odd number of as and bs, while q5 and q6 covers the remaining
two combinations. As these four states are all both useful and distinct, no two
of the can be merged. States q3 and q4 are both accepting and remember that
w can be divided in the manner required by the definition of L2 . Since q3 and
q4 are equivalent in all aspects that matter, they can be merged.
4.1

Distinguishing strings

One could easily modify Algorithm 1 to generate, for each (reachable) state q
of the input automaton A, a string which distinguishes it from every state q 0 6
[q]L(A) : With every state q, we associate a string sq . When the algorithm
first enters RefinePartition, all associated strings are empty. Every time
a block B is found to be unstable with respect to some input symbol a and
is consequently partitioned into smaller blocks, the symbol which forced the
split is appended to sq for all q that belonged to B. Eventually, the algorithm
terminates by outputting the DFA A0 = (Q0 , , 0 , q00 , F 0 ), the states of which
are the equivalence classes of L(A) . It then holds for every pair of states
[q1 ], [q2 ] Q0 , that sq , where q [q1 ][q2 ], is a distinguishing string for [q1 ], [q2 ].
As a final observation, we note that the set {qs |q Q} is prefixed-closed and
thus forms a tree over .

Further reading

In 1971, Hopcroft presented a minimisation algorithm for DFAs that runs

in time O(n log n), where n is the number of states of the input algorithm.
Hopcrofts algorithm is similar to the one presented in this chapter, but uses
the so-called process the smaller half-strategy to avoid unnecessary comparisons [Hop71].
During the fall semester of 2005, a course covering this topic and the background
topics were given at the department of Computer Science at the University of
Illinois by Prof. Viswanathan. The lecture notes from that course are quite
extensive and may be suitable as further reading [Vis05].

References
[Hop71] J. E. Hopcroft. An n log n algorithm for minimizing states in a finite
automaton. Theory of Machines and Computations, 1971.
[Ner58] A. Nerode. Linear automaton transformations. In In Proc. of the
American Mathematical Society 9, pages 541544, 1958.
[Vis05] Professor Mahesh Viswanathan.
Lecture notes.
CS475:
Automata, Formal Languages, and Computational Complexity
http://www.cs.uiuc.edu/class/fa05/cs475/Lectures/, 2005.

Input: a deterministic finite automaton A

Output: the minimal DFA recognising L(A)
01: procedure Minimise (DFA A)
02:
A0 := RemoveUnreachableStates (A) ;
03:
(Q, , , q0 , F ) := A0 ;
04:
P := RefinePartition (A0 , {Q \ F, F }) ;
05:
return Collapse (A0 , P ) ;
01: procedure RemoveUnreachableStates (DFA (Q, , , q0 , F ))
02:
Q0 := q0 ;
03:
while q Q0 such that q is not marked
04:
choose q Q0 such that q is not marked ;
05:
mark q ;
06:
for each a
07:
if (q, a) 6 Q0
08:
add (q, a) to Q0 ;
09:
0 (q, a) := (q, a) ;
10:
11:
return (Q0 , , 0 , q0 , F Q0 ) ;
01: procedure RefinePartition (DFA A = (Q, , , q0 , F ), Partition P )
02:
P 0 := ;
03:
for each q in Q
04:
obsq := ;
05:
for each symbol a in
06:
add (a, [(q, a)]P ) to obsq ;
07:
add q to [obsq ]P 0 ;
08:
09:
if P 0 is equal to P
10:
return P 0 ;
11:
else
12:
return RefinePartition (A, P 0 ) ;
01: procedure Collapse (DFA A = (Q, , , q0 , F ), Partition P )
02:
for each q Q
03:
add [q]P to Q0 ;
04:
for each a
05:
0 ([q]P , a) := [(q, a)]P ;
06:
07:
for each q F
08:
add [q]P to F 0 ;
09:
10:
return (Q0 , , 0 , [q0 ]P , F 0 ) ;

Algorithm 1: A minimisation algorithm for DFAs.

10