18. Nonregular Languages

One of the interesting properties of the class of \(\mathbf{REGULAR}\) languages is that we have good tools for determining when languages are not in the class.

Pumping Lemma

We saw in the last lecture that DFAs (or deterministic finite automata) can be defined as Turing machines with a read-only input tape that can only be traversed left-to-right and no work tape. It can also be defined more directly as an abstract machine that can be described using a transition diagram like this one:

DFA transition diagram

More precisely, a DFA \(M\) is defined by:

  • a set \(Q = \{1,2,\ldots,m\}\) of states;
  • a subset $F \subseteq Q$ of these states that are the accepting states; and
  • a transition function \(\delta : Q \times \{0,1\} \to Q\) that indicate which state is reached from a given state when the next symbol of the input string read by \(M\) is the one that labels the edge.

The start state of \(M\) is by convention the first state \(1\). The DFA \(M\) accepts a string \(x \in \{0,1\}^*\) if and only if it ends up in one of the accepting states in \(F\) after reading all of the symbols of \(x\).

This transition diagram view of DFAs lead to a simple observation: when the DFA \(M\) has \(m\) states, then every string \(x\) of length \(|x| > m\) is associated with a walk on the transition diagram of \(M\) that must visit some state more than once. In other words, the walk must include a cycle. This simple observation may seem almost trivial, but it implies that all regular languages have a useful pumping property.

Pumping Lemma. For every regular language \(L\), there is a number \(p\) such that for any string \(s \in L\) of length at least \(p\), we can write \(s = xyz\) where

  • \(\lvert y \rvert > 0\),
  • \(\lvert xy \rvert \le p\), and
  • For each \(k \ge 0\), \(x y^k z \in L\).

Proof. Let \(M = (m, \delta, F)\) be a DFA that recognizes \(L\). We will show that \(L\) satisfies the pumping property for the value \(p = m\).

Fix any string \(s\) of length \(|s| \ge m\) that is accepted by \(M\). Let \(q_0, q_1, \ldots, q_p \in [m]\) denote the states of \(M\) reached after reading the first \(i\) symbols of \(s\) for \(i = 0, 1, \ldots, p\).

Since \(M\) has only \(m\) distinct states, by the pigeonhole principle, there must be two indices \(0 \le i < j \le m\) for which \(q_i = q_j\). Define \(x = s_1\cdots s_i\), \(y = s_{i+1}\cdots s_j\), and \(z = s_{j+1} \cdots s_{|s|}\), so that \(s = xyz\).

By construction, \(s = xyz\). Since \(j > i\), the string \(y\) has length at least 1. And since \(j \le m\), \(|xy| \le m = p\). Finally, we note that the string \(y\) induces a path from \(q_i\) back to itself, so for any \(k \ge 0\) the string \(xy^k\) ends up at state \(q_i\) in \(M\). So \(xy^kz\) is accepted by \(M\).

A language \(L\) that satisfies the condition of the lemma is said to satisfy the pumping property.

A value \(p\) for which \(L\) satisfies the pumping property is known as a pumping length of \(L\).

Using the Pumping Lemma

When we use the Pumping Lemma to show that a given language is not regular, we usually want to use its contrapositive form:

Pumping Lemma. (Contrapositive form) Let \(L\) be a language that satisfies the following property: For every positive integer \(p\), there exists a string \(s \in L\) of length \(\lvert s\rvert \ge p\) such that for every decomposition \(s=xyz\) with \(|y| > 0\) and \(|xy| \le p\), there exists a value \(i \ge 0\) where \(xy^iz \notin L\). Then \(L\) is not a regular language.

The Pumping Lemma can be used to show that even some very simple languages are not regular.

Proposition. The language \(L = \{ 0^n 1^n : n \ge 0 \}\) is not regular.

Proof. For any \(p \ge 1\), choose \(s = 0^p1^p\). This string is in \(L\) and has length \(|s| \ge p\).

Fix any decomposition \(s = xyz\) with \(|xy| \le p\) and \(|y| > 0\). In any such decomposition, \(y = 0^\ell\) for some \(\ell \ge 1\). But then \(xy^2z = 0^{p+\ell}1^p \notin L\).

Hence, by the (contrapositive form of the) pumping lemma, \(L\) is not regular.

The language above can be thought of as an “equality” language, in that it contains the strings of the form $0^*1^*$ with an equal number of 0s and 1s. The analogous “less than” and “greater than” languages are also not regular.

Proposition.  The language \(L_< = \{ 0^m 1^n : n > m \ge 0 \}\) is not regular.

Proof. For any \(p \ge 1\), choose \(s = 0^p1^{p+1}\). This string is in \(L\) and has length \(|s| \ge p\).

Fix any decomposition \(s = xyz\) with \(|xy| \le p\) and \(|y| > 0\). In any such decomposition, \(y = 0^\ell\) for some \(\ell \ge 1\). But then \(xy^2z = 0^{p+\ell}1^{p+1} \notin L\).

Hence, by the pumping lemma, \(L\) is not regular.

Proposition. The language \(L_> = \{ 0^m 1^n : m > n \ge 0 \}\) is not regular.

Proof. For any \(p \ge 1\), choose \(s = 0^{p+1}1^p\). This string is in \(L\) and has length \(|s| \ge p\).

Fix any decomposition \(s = xyz\) with \(|xy| \le p\) and \(|y| > 0\). In any such decomposition, \(y = 0^\ell\) for some \(\ell \ge 1\). But then \(xy^0z = xz = 0^{p+1-\ell}1^{p} \notin L\).

Hence, by the pumping lemma, \(L\) is not regular.

The Pumping Lemma can also be used to show that some simple unary languages are not regular.

Proposition. The language \(L = \{ 0^{2^n} : n \ge 0 \}\) is not regular.

Proof. For any \(p \ge 1\), choose \(s = 0^{2^p}\). This string is in \(L\) and has length \(|s| \ge p\).

Fix any decomposition \(s = xyz\) with \(|xy| \le p\) and \(|y| > 0\). In any such decomposition, \(y = 0^\ell\) for some \(\ell\) in the range \(1, \ldots, p\). But then \(xy^2z = 0^{2^p + \ell}\) is a string of length between \(2^p + 1 > 2^p\) and \(2^p + p < 2^p + 2^p = 2^{p+1}\), so \(xy^2z \notin L\).

Hence, by the pumping lemma, \(L\) is not regular.

Myhill-Nerode Theorem

The pumping lemma is a useful tool for showing that some languages are non-regular. But one important warning about it is that the converse of the lemma is not true: there are non-regular languages that satisfy the pumping property. So it does not characterize the set of regular languages.

There is, however, a stronger result that does characterize the class of regular languages. To introduce this tool, we first need a new definition.

Definition. The strings \(x,y \in \{0,1\}^*\) are \(L\)-equivalent for some language \(L \subseteq \{0,1\}^*\), denoted \(x \equiv_L y\), if for every string \(z \in \{0,1\}^*\), we have that \(xz \in L \Leftrightarrow yz \in L\).

(As you can verify directly, the relation \(\equiv_L\) does define an equivalence relation, thus justifying the name and symbol we use for it.)

The index of a language \(L \subseteq \{0,1\}^*\) is the minimum cardinality of a set \(X \subseteq \{0,1\}^*\) such that every string \(y \in \{0,1\}^*\) is \(L\)-equivalent to some string \(x \in X\).

As it turns out, the index of a language provides a characterization of regular languages, in the following sense.

Myhill-Nerode Theorem. The language \(L \subseteq \{0,1\}^*\) is regular if and only if it has finite index.

Proof. \((\Rightarrow)\) Let \(L\) be any regular language. Let \(M\) be a DFA with \(m\) states that decides \(L\). Without loss of generality, assume that each state of \(M\) is reachable by some string. Form \(X\) by taking one string that causes \(M\) to halt at each of its states. Then every string \(y \in \{0,1\}^*\) is \(L\)-equivalent to the string \(x \in X\) associated with the state in which \(M\) halts on input \(y\), since \(M\) then halts on the same state in inputs \(xz\) and \(yz\) for any string \(z \in \{0,1\}^*\).

\((\Leftarrow)\) Let \(X = \{x_1,\ldots,x_m\}\) be a set of minimal cardinality that defines the index of \(L\). We can create a DFA \(M\) by making one state for each of these strings and defining the transition \(\delta(x_i,\sigma)\) to be the state for the string \(x_j\) that satisfies \(x_j \equiv_L x_i \sigma\) for each \(\sigma \in \{0,1\}\). Set the accepting states of \(M\) to be the states corresponding to the strings in \(X \cap L\). Then \(M\) decides \(L\) (exercise: verify this claim!) and so \(L\) is a regular language.

Note that the proof established even more than what was required by the statement of the Myhill-Nerode Theorem: it shows that the index of a regular language \(L\) determines the state complexity of \(L\) — the minimum number of states of any DFA that decides it.


Eric Blais ©2024 — Last edited on Mar. 19, 2024

results matching ""

    No results matching ""