Prerequisites

1. Language acceptance by Deterministic Finite Automata (DFAs)
2. Language acceptance by Non-deterministic Finite Automata (NFAs)

Introduction

Assuming that the reader has the understanding of DFAs and NFAs, we pose the following question to them.

Are there languages that are accepted by NFAs but not by DFAs?

Another way to think about this question is the following.

Are there languages that truly use the power of non-determinism.

As it turns out, the answer to the questions above is no. This is extremely surprising to understand that non-determinism does not add much power to the finite automata with respect to computability.

We shall now show that given any non-deterministic finite automaton that accepts a language $L$, we can construct a deterministic automaton that also accepts $L$. Recall that when we defined NFAs, we spoke of multiple possibilities of transition from the current state upon reading a letter of the input, and $\varepsilon$ transitions. These gave the finite state machines the power of non-determinism. In particular, we said that the machine splits into multiple copies when it encounters a situation of multiple transition possibilities.

How does one encode the notion of "multiple copies of a finite state machine running in parallel"?

One way to encode this is through the power-set construction. Let us illustrate that using an example.

Now let us stare at the transition table of this NFA $N_1 = (Q = \{q_1, q_2, q_3\}, \Sigma = \{0,1\}, \delta, q_1, \{q_1\})$ and try to construct a DFA $D_1 = (Q', \Sigma, \delta', S, F)$ that accepts the same language as $N_1$.

	$0$	$1$	$\varepsilon$
$q_1$		$q_2$	$q_3$
$q_2$	$\{q_2, q_3\}$	$q_3$
$q_3$	$q_1$

Note that we need to keep track of the transitions across multiple copies of the machine, deterministically. Towards this the following idea could work.

Can we create a new finite state automaton whose states are labeled by subsets of $\{q_1, q_2, q_3\}$.

Let $Q'$ denote the power-set of $\{q_1, q_2, q_3\}$. Let $\Sigma'$ be the same as $\Sigma = \{0,1\}$. Now we shall the define the transitions as follows.

For a subset $R$ of $\{q_1, q_2, q_3\}$ and a letter $a$ of the alphabet, we could possibly define the transition function $\delta'$ as follows.

$$\delta'(R,a) = \{q \in Q \mid q\in \cup_{r\in R}\delta(r,a) \}.$$

For example, $\delta'(\{q_2\}, 0) = \{q_2, q_3\}$, $\delta'(\{q_1\}, 0) = \emptyset$, and so on. But this does not take care of $\varepsilon$-transitions. Recall that $\varepsilon$-transitions create copies of the machine without reading a letter of the input. For any subset $R$, let $E(R)$ be defined as follows.

$E(R) = \{ R\cup$ states $q$ that are reachable from some $r\in R$ via one or more $\varepsilon$ transitions$\}$.

Thus, we modify the above definition of $\delta'(R,a)$ as follows.

[ \delta'(R,a) = {q \in Q \mid q\in \cup_{r\in R}E(\delta(r,a)) }. ]

Now the start state of the machine $D_1$ is the set of all the states that are reachable from $q_1$ (the start state of $N_1$) via $\varepsilon$ transitions. That is, $E({q_1})$. Final states $F'$ of the machine $D_1$ are going to be the set of states in $Q'$ that contain an accepting state of $N_1$.

We show this pictorially as follows.

$N_1$">

We can also remove the states $\{1\}$ and $\{1,2\}$ as they do not have any incoming edges to simplify the DFA.

Is power-set construction the only way to construct a DFA given an NFA?

No, there are more conversion algorithms. Kleene's algorithm is yet another way to convert a NFA into a DFA.

Given a NFA over $k$ states, the DFA constructed through the power-set construction could have $2^k$ many states.

Things to ponder:

1. Given a NFA over $k$ states, can we construct a DFA with much fewer states than $2^k$.
2. Are there any languages for which there is a NFA with few states and for every equivalent DFA has a lot of states.

Related topics

1. Language acceptance by Deterministic Finite Automata
2. Language acceptance by Non-Deterministic Finite Automata
3. Converting a Regular Expression to NFA