2
Language

You taught me language; and my profit on't
Is, I know how to curse. The red plague rid you
For learning me your language!
--Caliban (The Tempest 1.2.363-365)

'Tis but thy name that is my enemy;
Thou art thyself, though not a Montague.
What's Montague? It is nor hand, nor foot,
Nor arm, nor face, nor any other part
Belonging to a man. O, be some other name!
What's in a name? That which we call a rose
By any other word would smell as sweet...
--Juliet (Romeo and Juliet 2.2.38-44)

Lewis Carroll: Mathematics in Literature

In 1851, three years before Boole published The Laws of Thought, Charles Lutwidge Dodgson entered Oxford University. As an undergraduate, Dodgson distinguished himself in mathematics. After taking his degree, he spent the rest of his life at Oxford. As a mathematician he was mediocre. He wrote prolifically on mathematics, but his conservatism went against the grain of the mathematical innovation of the nineteenth century. His mathematical texts were never significant and are read now only because of historical interest. One point of interest in them, however, is the fascination with notation that they display. In his 1861 pamphlet The Formulae of Plane Trigonometry, for example, he introduced new symbols for the trigonometric functions. These were pictograms that attempted to express the nature of the functions, rather than the textual abbreviations (`sin', `cos', `tan', `sec', &c.) that were, and still are, the convention. Dodgson interested himself in logic and must have read and been influenced by the work of Peacock and later that of Boole.

Dodgson enjoyed the company of little girls. It feels likely to me that a factor in this tendency was an appreciation of the native state of the human mind, of the child's natural sense of wonder, the same sense revealed in To the Lighthouse. Dodgson enjoyed amusing his child friends with games of logic and wordplay. A favourite child friend of Dodgson's was Alice Liddell, the daughter of the dean of his college at Oxford. On July 4, 1862, Dodgson and a colleague accompanied her and her two sisters on a rowing trip up the River Isis near Oxford. On the way, Dodgson made up a story to amuse the girls. Alice later prevailed upon Dodgson to write down the tale. Dodgson is better known by his pseudonym, Lewis Carroll, for this work, which became Alice's Adventures in Wonderland (1865). The conundrums and absurdities of Wonderland, and of the looking-glass world in Through the Looking Glass (1871), are metaphor for a child's everyday encounters with the universe.

A full commentary on the works of Lewis Carroll, or even just on Alice's Adventures in Wonderland, would be a heavy tome. So we must restrict ourselves here to a few (arbitrarily selected) points illustrating the (omni)presence of nineteenth-century abstract algebraic thought in Carroll's work. We have seen that by the time Carroll began writing, the English algebraists had removed inherent meaning from algebraic symbols. Alice's fall down the rabbit hole in the beginning of Alice's Adventures in Wonderland is packed with images of symbols as arbitrarily defined markers, labels, and mechanically applied rules of inference. Just before she sees the white rabbit, she peeps at the book that her sister is reading, and thinks, "What is the use of a book without pictures or conversations?"[26] From within Alice's system of understanding, that of a seven-year-old girl in Victorian England, such a book is useless; she can't get any meaning out of it. As Alice falls, she passes "cupboards and bookshelves" filled with curiosities. Among them are "maps and pictures hung upon pegs". A map, of course, is an arbitrary encoding of a geographical area, just as an algebraic symbol is an arbitrary encoding. Carroll returns to the subject of maps in other writings. In Through the Looking Glass, there is the idea of a map of unit scale, so large that it covers exactly the country that it represents. In mathematical terms, Carroll has taken the one-to-one correspondence of map and terrain and made it an identity function. This transformation is a perfectly sound one in the domain of abstract mathematics, but when connected with physical reality it becomes ludicrously impractical. In The Hunting of the Snark, Carroll returns to the idea of an absurd map and makes the connexion with abstract algebra even more explicit. The following passage is from the second fit, a characterisation of the Bellman, the leader of the Snark hunting expedition:

He had bought a large map representing the sea,
Without the least vestige of land:
And the crew were much pleased when they found it to be
A map they could all understand.

"What's the good of Mercator's North Poles and Equators,
Tropics, Zones, and Meridian Lines?"
So the Bellman would cry: and the crew would reply
"They are merely conventional signs!

"Other maps are such shapes, with their islands and capes!
But we've got our brave captain to thank"
(So the crew would protest) that he's bought us the best--
A perfect and absolute blank!"[27]

Note Carroll's careful choice of words. The map represents the sea. The specification that the Bellman bought the map adds comedy. Objectively, he would have been just as well off with a blank sheet of paper, or with no paper at all. But the act of buying the map has made the blank sheet a map in the mind of the Bellman and his crew. Map symbols such as latitudes and longitudes are indeed arbitrary divisions, "conventional signs" like the symbols of abstract algebra. There are many other instances of play with labellings and encodings in Carroll's works, but I shall leave you with the excuse that they are best left for the reader to discover. That is, after all, the play of it.

The Hunting of the Snark also contains an instance of the falling into superfluity that I have mentioned as a consequence of the Information Age:

Then the Butcher contrived an ingenious plan
For making a separate sally;
And had fixed on a spot unfrequented by man,
A dismal and desolate valley.

But the very same plan to the Beaver occurred:
It had chosen the very same place:
Yet neither betrayed, by a sign or a word,
The disgust that appeared in his face.[28]

This is a metaphor for what had already begun in Dodgson's time. The Butcher has specialised; he has found his niche, "a spot unfrequented by man, / A dismal and desolate valley". When scholars Butcher and Beaver finally learn of each other's work, they find that they have been duplicating each other's efforts, but they conceal their disappointment. Maybe they can publish a joint paper on their method of hunting the Snark. If they do, one would hope that their proof would be mathematically rigorous, and not in the Bellman's style. For according to the Bellman, all that he has to do in order to prove a proposition is to state it three times in different ways. How many papers have you read (or written, at three o'clock in the morning on the night before it was due) that try to use this method to fudge and to escape logical rigour? "Just the place for a Snark! I have said it thrice: / What I tell you three times is true."[29] Unfortunately, the Butcher, who is a bit dim, puts his faith in the Bellman's misleading sophistry:

"'Tis the voice of the Jubjub!" he suddenly cried.
(This man, that they used to call "Dunce.")
"As the Bellman would tell you," he added with pride,
"I have uttered that sentiment once.

"`Tis the note of the Jubjub! Keep count, I entreat.
You will find I have told it you twice.
'Tis the song of the Jubjub! The proof is complete.
If only I've stated it thrice."[30]

This reference to the fall into superfluity is especially potent, embedded as it is among references to abstract algebra. For, as already noted, abstract algebra seems to have come over the entire Cambridge mathematical community at once.

Mock judicial proceedings are fertile ground for Dodgson. Legal language is natural language trying to behave as a formal system. Systems of law are fundamentally imperfect attempts to capture moral feelings of right and wrong in black marks on paper. The Snark hunting expedition has brought along a Barrister, "to arrange [the crew's] disputes". The sixth fit of The Hunting of the Snark comprises a dream of the Barrister in which the Snark is "defending a pig / On the charge of deserting its sty." The dream contains Carroll's indictment (so to speak) of legal finagling and the too-serious application of legal systems:

The indictment had never been clearly expressed,
And it seemed that the Snark had begun,
And had spoken three hours, before any one guessed
What the pig was supposed to have done.[31]

Alice's introduction to legal systems occurs in chapter eleven of Alice's Adventures in Wonderland. From her external point of view, it seems that the legal system encompasses not only laws, but also, more potently, behaviour in the court--or should I say, courtly behaviour. She identifies the judge "because of his great wig", a symbol that she recognises as signifying judgeship. The trial of the knave is an occasion for redefinition of terms, wordplay, and comic misinterpretation.

A fascination with symbolisation is a feature occurring in the lives of many scientists and mathematicians, especially in their childhoods. To take one contemporary example: The physicist Richard Feynman wrote in his autobiographical work Surely You're Joking, Mr. Feynman! of his invention of new symbols that seemed more reasonable to him when doing calculus. Since nobody else could understand this private scheme of abbreviation, Feynman eventually discarded it and began to use conventional notation. When I was learning calculus, in the eleventh grade in a suburban public high school, my own play with symbols was the source of some woe for my teacher. I could not bring myself to use the symbol `=' in substitutions of variables. If a new variable u were being introduced, I thought, it felt uncomfortable to treat it just like one of the pre-existing parameters of the problem by using `='. There should be some clue or mnemonic in the notation that would signify that such a variable was being defined, not merely expressed. So I borrowed the `:=' assignment symbol from the computer programming language Pascal and used it in substitions of variables in my calculus proofs. I also put slashes through my zeroes, as some computer character sets do, because I disliked the orthographical ambiguity between the numeric symbol `0' and the letter `O'. I was naïvely striving for a universal semeiotics, a set of symbols that would apply in all contexts.

Dodgson composed the first stanza of "Jabberwocky", the famous mirror-image nonsense poem in Through the Looking Glass, in 1855, when he was twenty-three:

Twas bryllyg, and the slythy toves
Did gyre and gymble in the wabe:
All mimsy were the borogoves;
and the mome raths outgrabe.

Significantly, he first conceived of this not as nonsense in everyday Modern English, but as Anglo-Saxon nonsense. But being nonsense, how can nonsense be in any particular language? Often we can recognise a foreign language by its sound, even if we do not know the language. Natural languages have distinctive sounds. Douglas Hofstadter presents some fascinating discussions of this idea of spirits or essences of particular languages in his collection of essays Metamagical Themas. This first stanza of "Jabberwocky" is easily read in Old English, with some syntactic translations, whereas portions of the rest of the poem, conceived of as Modern English nonsense, produce phonetic difficulties. Did you ever read a Tom Swift or Hardy Boys book in which somebody (Biff Hooper, perhaps) "chortled"? When Dodgson used this word in "Jabberwocky", it was just another of his many nonsense terms. But, ironically, the popularity of this particular piece of nonsense, along with the context in which the word occurred ("He chortled in his joy") has caused "chortled" to become a bona fide term in the English language, meaning a kind of laughing. The Anglo-Saxon appears briefly also in the seventh chapter of Through the Looking Glass, with the Anglo-Saxon messenger, Carroll's spoof on Anglo-Saxon scholarship. In the eighth chapter, the Anglo-Saxon is replaced, following historical developments, with the chivalric encounter of the White Knight and the Red Knight. Before having ado with the Red Knight, the White Knight mentions "the Rules of Battle". While they fight, Alice is left wondering what the Rules of Battle are, and trying to determine them empirically. Chivalry is a system foreign to Alice, so she has no intuition for its rules.

Puns have been around since prehistory. Punning comes as naturally to humans as telling jokes. But it was Dodgson who elevated wordplay from the status of a mere curiosity, a basis for cheap one-liners, and based works of literature on it. Dodgson's was the first literature to rely heavily on portmanteau words, words that use the human tendency toward analogy in phonetics, spelling, visual presentation, historical and cultural referents--every possible container for information--to overload single terms with multiple meanings. The portmanteau word would have entered literature without the work of Dodgson; after the new algebra of the nineteenth century and its parent philosophy of language, it was inevitable that what has come to be termed "semeiotics", the study of symbols, would become more explicit in literature just as it had in the sciences. James Joyce invented the concept independently of Dodgson, and when he began to publish his experimentation with the technique, people told him that it reminded them of the work of Lewis Carroll, whom he had never read at that point.[32] Finnegans Wake, Joyce's masterpiece of multiple semantics, contains several references to Lewis Carroll, among them an acknowledgement of Dodgson's priority: "To tell how your mead of, mard, is made of. All old Dadgerson's dodges one conning one's copying and that's what wonderland's wanderlad'll flaunt to the fair."[33] [Atherton 1960] contains a discussion of this and other references to Carroll in Finnegans Wake and of the influence in general of Carroll on Joyce. Dodgson often played his game of "doublets" with his child friends, in which the objective is to transform one word into another by making a series of words in which one letter is changed at each step in the series. The first sentence in the above quotation is Joyce's variant of doublets. The phrase "Dadgerson's dodges" needs no explaining.

Before leaving the subject of Lewis Carroll, I must remark upon the presence in his work of a certain number that had no distinction in his day but recently has become well known because of its appearance in the work of another author. Whenever Carroll needed some arbitrary number, it seems, he chose forty-two. In chapter twelve of Alice's Adventures in Wonderland, the King, during the trial of the Knave, produces "Rule forty-two. All persons more than a mile high to leave the court."[34] The Baker in The Hunting of the Snark came with forty-two boxes. In his preface to The Hunting of the Snark, Carroll explains that the Bellman "would only refer to his Naval Code"--this is the system within which the Bellman exists. The helmsman is prevented from saying anything to the Bellman by "Rule 42 of the Code, `No one shall speak to the Man at the Helm,'" which the Bellman has symmetrically amended with the words `and the man at the helm shall speak to no one.' The "man at the helm" is, by definition, the helmsman. But it is revealed in a footnote that the office of helmsman was not a permanent assignment. The implication is that the helmsman, when he stopped being helmsman, would be able to speak to the Bellman. But then the helmsman would not be the helmsman anymore, so he wouldn't be able to speak as the helmsman. (This play with self-reference and the essence of self is another favourite of Dodgson's; Alice spends much thought wondering whether she can consider herself as her self, after changing so much.)

Forty-two would have remained an insignificant feature had it not been for the "Hitchhikers" series of novels by Douglas Adams, beginning with The Hitchhiker's Guide to the Galaxy, in which it is revealed that the planet earth is an artificial computing device, constructed in order to find the ultimate question of "life, the universe, and everything", the answer to which has already been found to be forty-two. It seems likely that Adams borrowed this particular answer from Lewis Carroll. The earth is destroyed for a petty motivation just before the moment of output, and as a result a replacement planet has to be constructed and the entire computation restarted. Dodgson would have loved this idea. So would Charles Babbage, a nineteenth-century pioneer of automatic digital computing. If only the makers of the earth had built in a checkpointing mechanism!

Slang: The Art of Architecture

Many subcultures have their own argot, formed by the play with analogy that the human mind seems unable to resist. Sometimes when I am discussing with my advisor some changes I want to make to a program, he asks me what effects my proposed alterations would have upon other facets of the system, and I respond with, "Oh, all this is orthogonal." What I mean is that the changes I want to make to some parameters of the system will have no effect upon some other parameters; the two sets of parameters are independent of each other. In linear algebra, when two vectors are completely independent of each other, that is, when they have no shared components pointing in a common direction, they are said to be "orthogonal". This style of generalisation of a term based upon an analogy constructed from its semantics recalls the Principle of the Permanence of Equivalent Forms.

A common word overheard in the halls of the high school that I attended was "lunshin". This was part of a local dialect of what Cran, McCrum, and MacNeil in The Story of English term "Black English". "Lunshin", as it was used, described a person who was temporarily being foolish. My guess is that it was derived from the idiom of the same meaning, "out to lunch". The new word "shody" (from "shorty", as in "Hey, shorty!") was used as a masculine personal pronoun. The dipthong "ya", a blurring of the Southern "you-all", often took the place of "you", in the singular as well as the plural case. Conjugation of the verb "to be" was simplified, leaving much to context and resulting in sentences such as "Shody lunshin" (He is being foolish), "Ya be lunshin" (You are being foolish, or, You are foolish), "Shody been smacked upside the head" (He has been hit on the head, or, He had been hit on the head), and "I been at the sto'" (I was [imperfect] at the store). In pronunciation, there is a tendency to elide final consonants and consonants directly followed by other consonants, sometimes gliding over them, sometimes omitting them altogether. Thus "lu(n)shi(n)" (discussed above), "ai(n)' go' do da(t)" (ain't gonna do that), and "Lo' Boa(t)" (Love Boat, a psychedelic drug).

When I left the linguistically fertile ground of that high school and arrived at university I was struck with a linguistic commentary on the recreational habits and ribaldry of many contemporary college students: the plethora of words and phrases meaning "vomit" and/or "to vomit" (Middle English, from Latin "vomitus", the past participle of "vomere"). At most American colleges and universities, a weekend cannot pass without seeing multitudes puke, yuke, yak, yag, gak, splag, zool, barf, woff, rauf, ralph, boot, heave, hurl, lose it, let 'er rip, uneat, wear their lunch, have dinner for the second time, toss their cookies, 'gurgitate, throw up, shove up, upchuck, chuck Cheerios, toss tacos, make macaroni, paint the pavement, shout at their shoes, feed the fish, drain the main, swab the deck, blow lunch, blow chow, blow chunks, spew chunks, blow groceries, upset the shopping cart, clean house, worship the porcelain, pray on the porcelain altar, pray to the porcelain prince, drive the porcelain bus, talk to the big white telephone, talk to Huey, talk to Ralph, scream Buick, do the yellow-belly dance, fill the gut bucket, or deliver a street pizza, in a cascade of technicolor yawns, 3-D yawns, chunky waterfalls, pigtracts (from "pig extract"), liquid laughs, and reverse peristalsis. In addition there are the more polite "disgorge" (Middle French "desgorger"), the primary meaning of which has been broadened from that of an oral discharge to any kind of mass discharge, and the oh-so-euphemistic British "pass up". These people are intoxicated, inebriated, drunk, tipsy, silly, happy, buzzed, tanked, loaded, leasing liquor, bombed, blitzed, hammered, rocked, smashed, totalled, liquidated, sloshed, gunned, whaled, wasted, plastered, polluted, pissed, haunted, pie-eyed, tight, gone, gonzo, gonged, out-of-it, done-like-dinner, smokin', toasted, fried, baked, sauced, fricasséed, zambezeed, zinged, zombied, zooed, wazooed, fucked up, and shitfaced (or just plain 'faced). This particular lexical variety is probably not what President Frank Rhodes has in mind when he speaks of the great "diversity" of "the Cornell community". But, there it is.

* * *

The fascination with language and meaning evidenced in many of the works discussed in this thesis is typical of computer science. Computer science is an outgrowth of mathematics, and its roots lie in nineteenth-century symbol-pushing. Much of the creativity in computer science lies in finding ways of representing information within prescribed schemes. Many computer scientists, uncomfortable with the freedom, intractability, and continual evolution of natural languages as eighteenth-century natural scientists were uncomfortable with the proposal of arbitrary taxonomies and as Hamilton was uncomfortable with the arbitrariness of the rules of abstract algebra, try to live in a make-believe universe of static languages. Like Lily Briscoe in To the Lighthouse, they wish to ascend to some one perfect representation and are frustrated with their inability to find such a representation. This type of person, grasping for a mandated system, wishing to be imprisoned, is like the child who protests when someone drives even one mile per hour above the speed limit or crosses the double yellow line to pass a standing bus.

David Gries, a professor in the computer science department at Cornell University, seems to me a more moderate case of this type of behaviour. He latches onto lexicons and manuals of style in an attempt to reify English usage, and once he has found what he would like to be the standard, he suggests--vigorously--that other people consider adopting it. One of the distinctions that he espouses is between the semantics of "which" and those of "that". Gries, armed with supporting references, believes that, regardless of most spoken usage, "that" should introduce an essential qualifier, as in "the house that was built by Jack", whereas "which" should introduce an accidental qualifier, as in "that house, which was built by Jack". Some other members of the department also hold this view, so that when I once gave my advisor a paper of mine to review, I found it annotated with remarks on such "errors" when he gave it back to me. Some people, it seems, would like to be able to treat English more like a computer language, data in a standard format to be chewed up and spit out by some compiler program. It is true that we need common context in language, but at the same time, the range of expression should not be so restricted that individual styles are hampered. In my opinion, Gries goes too far toward the extreme of standardisation. In 1989, Gries sent this electronic message to the department:


[An associate of mine in the department] found an author that recommends the use of "they" in the singular for "he or she". Based on that recommendation, I looked up "they" in the Oxford English Dictionary and found the following.

they: Often used in reference to a singular noun made universal by "every", "any", "no", etc. or applicable to one of either sex (= "he or she").

The OED cites several examples of the latter use--i.e. of a singular noun applicable to one of either sex. Here are two of them.

	From 1759: If a person is born of a gloomy temper, they cannot help it.
	From 1866: Now, nobody does anything well that they cannot help doing.

Similarly, the OED says that "them" is
	often used for "him or her", referring to a singular person whose sex is not stated.

	From 1874: Whenever anyone was ill, she brewed them a drink.

Similarly, the OED says that "their" is
	often used instead of "his or hers" when the gender is inclusive or
uncertain.

	From 1845: Every human being must do something with their senses.

From: 1898: It's enough to drive anyone out of their senses.


That's enough for me. From now on I use "they" for "he or she", "them" for "him or her", and "their" for "his or hers".[38]
This illustrates Gries's acute awareness of the system of language that he uses to express his thoughts, a system that most people take for granted.

Language Theory

We have seen some formal systems of expression--the propositional and predicate calculi, the list world, and Peano Arithmetic--and we have noted the coexisting duality of expression and meaning. On one level these formal systems are only methods of shunting symbols. And yet, when these symbols are given interpretations, there are correspondences between their forms and combinations and the forms and combinations of ideas that they represent. As Lavoisier wrote, a language is as important as the ideas expressed by it. We have studied the ideas of formal logic and described formal languages in which these ideas can be expressed. Now we must give more attention to the theory of languages.

We will begin with some definitions. An alphabet is a finite set of symbols. "Symbol" here is an undefined term; anything can function as a symbol. We can do Peano arithmetic by forming strings of `s's or by queueing up people at a coach stop. Chess can be played using plastic pieces on a table, actors in a field that has been squared off, characters in Through the Looking-Glass, or even by post, with no physical board or pieces at all. An alphabet is conventionally denoted by the Greek letter `'. A string over an alphabet is a finite sequence of symbols from the alphabet. A symbol by itself is a string whose length is one. The length of a string is the number of occurrences of symbols in the string. The length of a string x is denoted by `|x|'. The null string is the special string that contains no symbols (i.e., its length is zero). It is represented by `'. String variables are usually represented by the letters u, v, w, x, y, z, and symbol variables are usually represented by a, b, c, .... The concatenation operation joins two strings. It is similar to append in the lists world and to . in Peano arithmetic. Concatenation, like multiplication in arithmetic, is symbolised either by juxtaposing two variables and omitting the operator, as in `xy', or by the operator `.', as in `x.y'. As nil is the identity element for append in the lists world, is the identity element for . in the strings world. A substring is a string that is part of another string, that is, string x is a substring of string y if and only if there exist strings u and v such that y=uxv. u or v or both may be the null string, in which cases, respectively, x is said to be a prefix of y, a suffix of y, or simply y. A string x superscripted with a natural number n represents x repeated n times: a5 = aaaaa. So a superscript n is shorthand for n-1 concatenation operations. (n-1 concatenations, not n of them, because to repeat a string only once does not require a concatenation; it requires only writing out the string.) This was the origin of our meta-notation for strings of `s's in Peano arithmetic. The superscript 0 is defined, following Peacock's Principle of the Permanence of Equivalent Forms, as producing , the identity element for ..

A language over an alphabet is a set of strings over that alphabet. Since languages are sets, all the operations of set theory apply to languages. Here are some of them and descriptions of what strings they generate for two languages L0 and L1 over an alphabet :
OPERATION GENERATES
All strings over that are not in L0
L0L1 All strings that are in either or both of L0 and L1
L0L1 All strings that are in both L0 and L1

The concatenation of two languages L0 and L1, denoted "L0L1", is the set of all possible string concatenations formed by taking the first string from L0 and the second from L1. The concatenation of a set with itself can be abbreviated using a superscript, as the concatenation of a string with itself can be abbreviated using a superscript. So L2 is LL, L3 is LLL, and so on. L0 is . (This is not the same as the null string ; rather, it is the set whose only member is .) There is also the `*' operation, called Kleene closure. (For your information, "Kleene" is pronounced with two long `e' sounds, like "Kleenee".) Kleene closure is like superscripting, except that the number of set product operations is a wildcard; any string that can be produced by concatenating any number of strings in L is in L*. So L* = Li. For example, {a, b, c}* = {, a, b, c, aa, ab, ac, ba, bb, bc, ca, cb, cc, aaa, aab, ... }

Regular expressions are descriptions of languages. The class of languages that can be described by regular expressions is called the regular languages. I have never found any good reason for using the word "regular" to describe these languages. They can be generated by rules, but so can many languages. Perhaps the justification is that the regular languages can be generated by very simple rules, in which case a better term than "regular" would have been "simple". In any case, the term is entrenched. Regular expressions are defined inductively. (L(R) is the language represented by regular expression R.)

Although it does not add any more expressive power to regular expressions, it is notationally convenient to define another operation on regular expressions, positive closure. If r is a regular expression, then r+ is a shorthand for rr*. In other terms, L(r+) is L(r *)-; the resultant string must contain at least one string from the generating set. * and + (positive closure) have higher precedence than ., which has higher precedence than + (union). As usual, unneeded parentheses can be quietly dropped. Here are some sample regular expressions and their meanings in English:
EXPRESSION MEANING
(0+1)* all strings of 0's and 1's
0*(0+1)*1* all strings of 0's and 1's
0*+1* all strings consisting only of 0's or only of 1's
0*1* all strings over {0, 1} that do not contain the substring 10
(a+b+c)+bab all strings of a's, b's, and c's of length at least four that have the suffix bab.
(0+)(10)*(1+) all strings of 0's and 1's that contain neither 00 nor 11 as a substring
(aa)* all strings of a's that are of even length
a*b(a+b)* all strings of a's and b's that contain at least one b
(a+b)*b(a+b)* all strings of a's and b's that contain at least one b

As the first two examples above demonstrate, there is more than one way to skin a cat: Always reduce your regular expressions to simple terms that make the effect of the expression most apparent. Also, note the difference in meaning between (0+1)*, 0*+1*, and 0*1*. These forms have different effects, and beginning students often confuse them. We have now seen four formal languages. Here is a description of similar operations and identities in them:
LANGUAGE OPERATION IDENTITY
propositional calculus false
true
lists append nil
Peano arithmetic + 0
. s(0)
regular expressions+
.

EXERCISES

0. Two regular expressions denoting the set of all strings of a's and b's that contain at least one b are given above. There is a third simple regular expression that denotes this set. What is it? How is it related to each of the other two, and it what ways does it differ from each of them? Your answer should not be in terms of the operations *, +, and . involved in each of these regular expressions, but in terms of the different ways in which the subexpressions in each of the regular expressions correspond to the parts of the strings in the set that is being described.

Which of the following statements are true? Explain.

Describe the following regular expressions in English. Do not simply read off the regular expression mechanically substituting English phrases for regular expression operations. Rather, give the restriction(s) upon the strings that are implemented by the given regular expressions. So, for example, in the table above, the meaning of (0+)(10)*(1+) is given not as "either a 0 or the null string, followed by 10 any number of times, possibly followed by a 1" but as "all strings of 0's and 1's that contain neither 00 nor 11 as a substring". This says what the regular expression does, not how it looks.

Give regular expressions for the following sets of strings over a, b:

Languages are fundamentally connected with machines and algorithms. We shall see that problem instances can be encoded or viewed as strings, and a problem can be viewed as a question of deciding whether any given string is in a particular language. This decision is more difficult for some languages than for others, and for some it is impossible. For now let us examine the history and theory behind some computing machines.

Babbage: The First Hacker

Charles Babbage, a co-founder of the Cambridge Analytical Society with George Peacock and John Herschel, turned away from work in abstract algebra to devote himself to designing and constructing automatic, mechanical computers. Babbage was singularly devoted to this project, regularly working ten hours or more each day. When offered the Lucasian professorship at Cambridge, the prestiguous chair once occupied by Isaac Newton, his first impulse was to turn down the offer because of the time that the duties of the professorship would take away from his work. He was convinced to accept the appointment, but eventually resigned in order to work more on his "calculating engines". Luckily, Babbage possessed a sizable inheritance which he used to fund much of his work. In his own time, Babbage became regarded by many as a crackpot, a relentless peddler of impractical ideas. Babbage recognised many of the principles of computer architechture and programming. However, the value of his work has been noted only in retrospect, after its rediscovery, a century later, by the mathematicians and engineers who constructed the first electronic digital computers.

In Babbage's time many mathematical tables, such as tables of logarithms and of astronomical constants, were calculated by human computers using the method of differences. The method is best illustrated by example, and we choose the same one used by Menabrea in his "Sketch of the Analytical Engine". Suppose one wants to calculate a table of the squares. One starts off with a second difference of 2, a first difference of 1, and the first of the squares, 0:

0 1 2

The squares will be generated in the first column of the table. The second column, an intermediate result, is a sequence of successive odd integers. To derive the second row from the first, we first write down the constant second difference, 2. Then we add the second difference of the first row, 2, to the first difference of the first row, 1. The result, 3, is the first difference of the second row. Finally, the first difference of the first row, 1, is added to the first square, 0, to produce the second square, 1:

0 1 2

1 3 2

Note the pattern in which each successive row is generated from the row above it. The new entry in the first column is the sum of the first and second columns of the previous row. Similarly, the new entry in the second column is the sum of the second and third columns of the previous row. Thus sums are flowing from the rightmost column of the table, where the constant 2 is injected at each step, to the leftmost column of the table, where the squares are being output. The next row is (4, 5, 2). The Difference Engine, as it existed on paper, was capable of managing seven such columns. That is, it was restricted to the sequences whose seventh difference is 0. Note that this calculation is based upon the simple algebraic relationship (n+1)2 = n2+(2n+1). If the rows of the table are numbered starting from 0, then the nth row, in general, contains the triple (n2, 2n+1, 2). If the nth row is known, then the (n+1)st row can be easily computed. This is a simple iterative algorithm.

Babbage's childhood reveals in his personality the germ of what was to become his life's occupation. When offered a toy some of the workings of which were concealed, the young Babbage would demand what was inside it. If the answer were not sufficient, the toy would be broken open, and Babbage would see for himself. Thus the complex rules of operation of the toy would be resolved into a synthesis of the simpler rules governing its parts. Babbage ascribes this account of his behaviour to his mother; he himself must have been too young to remember it. Even at that stage of his development, he felt the urge to seek the hidden harmony. Like many other computer scientists, Babbage enjoyed devising and foiling security systems and codes. He records an episode at school during a period when he was sneaking out of his room at night to study when his roommate devised alarm systems to wake him if Babbage attempted to open the door of the room in sneaking out. In some sense Babbage's childhood lasted his entire life. He always retained his boyhood fascinations. In his memoirs he recounts an expedition into the crater of Vesuvius, where his native guides refused to follow him, in which he timed the interval between small eruptions of lava and, wanting to see a pool of molten lava, approached the site of the eruptions during the short intervals of inactivity. What confidence in inductive truth! Such a reckless, living Englishman!

Babbage's preoccupation with ways in which information can be conveyed led to his later inventions. The complications of his mechanical drawings, for example, caused him to develop a new system of mechanical notation. He later described this as "one of the most important additions I have made to human knowledge".[39] His computing machines, like all computing devices, were physical encodings of arithmetical (or algebraic) principles, as Hammurabi's Code was the embodiment of legal principles. Babbage introduced punched cards as media for the storage of computer instructions. The method had been used previously to control automatic looms that wove patterns, just as the punched roll of paper in a player piano causes the automatic reproduction of a piece of music. Ada Augusta, Countess of Lovelace, observed with an intimate sense of the beauty of the mathematics of the Analytical Engine, "We may say most aptly, that the Analytical Engine weaves algebraical patterns just as the Jacquard loom weaves flowers and leaves."[40]

It is said generally that Babbage was ahead of his time. In particular, his ideas for automatic computing machinery were ahead of the state of technology. The construction of Babbage's first machine, the Difference Engine, required the prior fabrication of a collection of many new tools. It was said by some that the 17,000 spent on the Difference Engine by the British government was well disposed even without the completion of the project, because of the advances in the state of the art of mechanical engineering that accompanied the development of the Difference Engine. (This same argument is now employed to justify spending by the Department of Defense in the United States.) Babbage's nineteenth-century universe was still the mechanistic one of Newton, so he never could have conceived a physical implementation of a computing device as anything but a mechanism, an arrangement of gears, springs, and levers. The developments in the study of electromagnetism which were to lead to the genesis of electronics did not occur until the latter part of the nineteenth century.

Babbage was the first hacker. He worked empirically, concerning himself more often with getting the best performance out of a physical implementation of a mathematical system than with specifying that system in the abstract. In this respect he was a prime example of the English scientist, the gentleman of independent means twiddling knobs on his apparatus and stumbling upon something neat every once in a while. I mean "hacker" in the old sense of the word. During the Golden Age a "hacker" was simply someone who hacked out computer programs, using the same method, very much founded upon observation, as Babbage. A hacker throws a program together, often introducing several bugs in the process, and then runs it on test data in order to find and to eradicate the bugs. A modern mathematician, in the spirit of the Continental mathematicians of the nineteenth century, would advocate developing the program simultaneously with a mathematical proof of its correctness. Hackers, using the engineering approach, do not bother with such preliminaries, but jump right into what they consider the real substance of the problem.[41] Babbage must have been aware of the significance of his machines in terms of abstract algebraic properties, having been so active in the Analytical Society and having written privately on the subject as early as 1821. Yet all his published writings refer to his machines in arithmetical terms. Ada Lovelace, in her explications of Babbage's calculating engines, refers to abstract algebra, but Babbage, as a hacker, seems to have taken the pragmatic, real-world, arithmetical view of their functioning.

In his Ninth Bridgewater Treatise, Babbage compares the operations of the universe to the operations of one of his calculating engines. The Bridgewater Treatises had been the legacy of Francis Henry Egerton, Earl of Bridgewater, who left 8000 to support the production of treatises "on the Power, Wisdom, and Goodness of God, as manifested in the Creation"--that is, to commission popular expository works supporting the Argument from Design. There were eight of these Bridgewater Treatises, and Babbage took the name "Bridgewater Treatise" for his own work, presumably to heighten its popularity. The Ninth Bridgewater Treatise was a reply to William Whewell, who wrote in the first of the Bridgewater Treatises that one could not approach the divine through the "deductive", or applied, sciences. "We have no reason whatever", wrote Whewell, "to expect from their speculations any help, when we ascend to the first cause and supreme ruler of the universe."[42] Babbage and Whewell had become friends at Cambridge and were both interested in the mathematical sciences, but they had turned in opposite directions. Babbage had given up abstract mathematics in favour of implementing mathematics with his calculating engines, and Whewell, possessing a broad knowledge of many scientific fields but not the insight to make the same kinds of discoveries as his fellows, turned toward the history and philosophy of science. How could Babbage's masses of cogs and springs, clearly the work of humans and not a direct result of God, approach the wonderful functioning of God's system of celestial mechanics, and his animal and plant creatures? This was Whewell's reasoning. Babbage refrained from blunt rebuttal of Whewell, because he agreed with the general idea that one could approach the divine through the investigation of Nature. He wrote in his preface:

One of the chief defects of the Treatises above referred to appears to me to arise from their not pursuing the argument to a sufficient extent. When a multitude of apparently unconnected facts is traced up to some common principle, we feel spontaneously an admiration for him who has explained to us the connexion; and if, advancing another stage in the investigation, he prove that other facts, apparently at variance with that principle, are not merely no exceptions, but are themselves inevitable consequences of its application, our admiration of the principle, and our respect for its discoverer, are still further enhanced.[43]

This passage displays perhaps a tinge of egotism, concentrating as it does upon the discoverer of a scientific principle instead of on the principle itself, but the sentiment is clear enough. Babbage was a devotee of Laplace's idea of a clockwork universe. Given a set of physical laws and an initial condition for the universe, one could, if given enough time for the requisite calculations, apply the laws iteratively and predict the state of the universe at any particular time in the future. This spirit of determinism was "in the air"; it was not particular to Laplace. Any reasonably intelligent child can come up with it, and it has continued to have a great influence on modern science. Einstein, for example, refused to accept the nondeterminism of quantum mechanics, insisting that "God does not play dice with the universe." Babbage took this clockwork idea and made it literal: his machines were mechanical constructions of gears and springs, demonstrating the production of apparently complex behaviour--the apparent harmony--by simple, mechanical, deterministic laws--the hidden harmony.

In one chapter of his treatise, Babbage details the construction of a clockwork machine that produces a series of numbers as its output. He conceives of a hierarchy of levels of controlling rules. This hierarchy of rules in Babbage's computing system is similar to the hierarchies of laws in legal systems. In the United States, for example, the highest law is the Constitution. It determines what other types of laws may be made, and what precedences they are to have, and so on. Similarly, the highest rule in Babbage's machine is the rule that determines what rule will determine the machine's output. So, for example (after Babbage's construction), the highest rule might say "Use rule A until the number 420 is produced. Then switch to rule B." Rules A and B are as follows:

A Begin with the number 0, and increase by 1 on each step.

B Begin with a count of 421. On each step, produce the square of the current value of the count and then increase the count by 1.

Babbage's construction uses a more complex set of rules, but the idea of rules supervising other rules in the operation of a machine is the same. After seeing the first few hundred numbers produced by the machine, any reasonable observer would be fairly confident that the sequence being produced is a simple succession: each number is always one greater than the previous number in the sequence. They would have perceived the apparent harmony. But on the 421st step, the observer would be surprised, and their world-view sent flying topsy-turvy, by the appearance of 177,241 instead of the expected 421. Babbage suggested such a hierarchy of controlling rules as a basis for miracles. In this scheme, miracles were not instances of divine intervention in the universe, but effects of higher-order rules preprogrammed into the universe through divine foreknowledge. Obviously, the universe is more than a collection of cogs and flywheels charged with the purpose of calculating a list of numbers. But the analogy between the physical laws of the universe and the computational rules of Babbage's calculating engines is a good one. And given the basis of science in mathematics, it is not difficult to conceive of the universe as a mathematical system. Babbage inverted Whewell's hierarchy of the natural sciences. As the German mathematician Leopold Kronecker said in the late nineteenth century, "God made the integers; all the rest is the work of man."

Babbage himself is at least partly culpable for his fall into superfluity. He was so often excited by an idea or a discovery that he plunged into work on its implementation without enough preliminary study. As a result of the British government's cancellation of funding for the Difference Engine, Babbage was forced out of the domain of implementation and into one of contemplation, and in only a few months he had come up with a much more general computing device, the Analytical Engine. As noted above, the Difference Engine imposed a linear flow of information, evidenced in the notation above by a right-to-left flow of sums through the columns of differences. The Analytical Engine was described by Babbage as the Difference Engine "eating its own tail", in an echo of Kekulé's dream-inspiration for the structure of the benzene ring. Information from the left could be channelled back into the machine on the right. Such a machine was capable of all the operations of mathematical analysis, not merely of calculating tables based upon the method of differences. If Babbage had disciplined himself to spend more time thinking his ideas through before beginning construction of his machines, he could have gone to the government for funding on the Analytical Engine and followed the design through to its implementation. As it was, Babbage would repeatedly come up with a scheme, invest much time and money in bringing that scheme to implementation, and then conceive a simpler scheme and abandon all the work that had gone before. One activity upon which he wasted a great amount of labour was the optimisation of the mechanical process of performing carries during additions. He noted that although most of the work of addition, that of adding the corresponding digits of the two addends, can be performed in parallel, adding each pair of corresponding digits simultaneously, the carrying from one place to the next was inherently sequential, because carries on the right of a sum can induce more carries further to the left. He had many ideas about mechanisms for anticipating the occurrence of a carry, and for storing the carries of a running sum and adding them into the result as a final step. He spent all his time hacking and hardly ever saw a project through to a conclusion. Thus Babbage is another example of a scientist uncomfortable with imperfect expression. Because he could not bring himself to accept inefficiency when he could sense the existence of a better way of building a machine, he flitted from one scheme to another and never stuck with his work. His work was both a product and a victim of his compulsion to seek a perfect representation for his ideas.

Finite Automata[44]

Mathematicians have constructed abstract models of computation that do away with the complicating implementation details of real-world computers. These types of machines are equivalent to classes of real-world computers. "Equivalent" here means able to solve the same classes of problems. Because of this equivalence, mathematicians can prove theorems about abstract models of computation and know that their results apply to all real-world computers. One wonders what Babbage, the mechanical implementor, would have thought of such abstract machines, that exist only on paper.

A deterministic finite automaton (DFA) is a simple theoretical computer. It is given an input string and always starts operating in a designated initial state. Each character in the input causes the DFA to make a state transition. The DFA can use its internal state to remember information about the prefix of the input that it has seen so far. Together, the states and the transitions between them are the DFA's control. The DFA accepts the input if it is in a final state when the end of the input is reached. Otherwise the DFA rejects the input. The DFA can pass through a final state, when it is not at the end of the string, without accepting. A simple example of a DFA is a combination lock. Suppose a lock requires a 4-digit combination to be entered, one digit at a time. Whenever an incorrect digit is encountered, the lock returns to its initial configuration and the entire combination must be re-entered from the beginning. Only when the entire combination has been entered does the lock reach a final state and open. If another number is entered after the correct combination, the lock resets to its initial state. The transition diagram for one such lock with alphabet 0, 1, 2, 3, 4 looks like this. What is the simple regular expression for the language that this DFA accepts? Since each transition is labelled with the input symbol that causes the transition, at any point while the DFA is operating the path that it has followed through its control is labelled with the prefix of the input that has been scanned so far. Formally, a DFA consists of the following five objects:

A DFA can also be represented in tabular form: The only trickiness in this automaton is transitions such as the one from SAW_0303 back to SAW_030 on symbol 0. 03030 is not a substring of the combination, but its suffix 030 is. So it is possible that although the prefix 03 was garbage, the suffix 030 will be followed by 34 and will complete the combination. If the transition were to return the automaton to its initial state, such a case would be handled incorrectly. The rule being followed here is that when a new symbol is seen in a state that is labelled with a string, concatenate this latest symbol onto the right-hand end of the label and move to the state labelled with the longest prefix of the resulting string. Instead of having the lock reset when an erroneous combination is entered one may wish to have the lock enter a state in which it refuses to accept any more input. This can be accomplished by adding a dead state and having the automaton move into it whenever an incorrect symbol is encountered. A dead state is a nonfinal state whose outgoing transitions all point back to itself. Thus, if an automaton enters a dead state, it can never move out of the dead state, so it will never accept the input that it is reading. Since the formal definition of a DFA requires that d be defined for every combination of state and symbol, transitions to a dead state should exist at all points at which the automaton determines that it definitely will not accept its input.

Constructing an automaton is one way of specifying an algorithm. The state of an automaton is analogous to the state of the algorithm--the values of all its variables, including the program counter, a special variable that points to the next instruction to be executed. States should be labelled with the information that they represent, as in the diagram above. The transition function maps pairs of a state and a symbol to new states. It is a functional representation of the transition arrows in the diagram above.

A DFA is finite because it can remember only a bounded amount of information. This is because the DFA's only mechanism for memory is its state. Since the DFA has to make a state transition for each symbol in its input, and since the maximum-length noncircular chain of transitions among n states is of length n-1, if the length of the input string is greater than or equal to the number of states in the DFA then it must be the case that the DFA has traversed at least one cycle among its states and has therefore lost some of the information conveyed by the input. To answer a specific question about an input, such as "Does this string have a suffix that matches the combination?", not all the information in the string must be preserved. The question is deciding what information to throw away.

A DFA is deterministic because there is exactly one next state when the DFA sees a particular symbol in a particular state. The machine has no choice of states to enter. When it sees a symbol a in a state q, it must go to state (q, a). Thus a deterministic machine will always traverse the same path through its set of states when confronted with the same input. (A nondeterministic machine has a choice of states to enter when it sees a particular symbol in a particular state.)

Finite automata have applications in the sense that many programs are concrete implementations of the abstract concept of a finite automaton. However, in most cases the link between the two concepts is obscured by differences in structure between the program and the automaton. There are some applications, though, in which the design of a program has been based upon the concept of a finite automaton. Many text editors use variants of regular expressions to specify patterns when searching for strings within a piece of text, and the search is accomplished by constructing a finite automaton that recognises the specified pattern. Lexical scanning, the first phase of compilation, uses a finite automaton to recognise reserved words, numbers, identifiers, and other categories of strings within text that is being compiled.

EXERCISES

Construct deterministic finite automata that accept the following sets of strings:

Describe (in plain English!) the languages accepted by these automata:

* * *

The nondeterministic finite automaton (NFA) is a generalisation of the deterministic finite automaton. Nondeterminism is a relaxation of constraints. A simple nondeterministic algorithm is the problem of finding all the items on a shopping list inside a supermarket. The items are located at various points along various aisles, and each of these locations must be visited. A path containing all these points must be constructed. But there are many possible paths, and the choice of one particular path is left to the shopper. Similarly, an NFA has a choice of transitions out of any particular state on any particular symbol. The formal definition of an NFA is the same as that of a DFA except that the transition function maps to 2Q, not to Q. The set 2Q is called the power set of Q. Its elements are the subsets of Q. The power set of 0, 1, 2, for example, is , 0, 1, 2, 0, 1, 0, 2, 1, 2, 0, 1, 2. The notation for power sets is in the spirit of Peacock; it arises from the fact that for any set S, |2S| = 2. Since the transition function in an NFA maps into the power set of the set of states, the result is a set of states instead of a single state. From this set the machine chooses a next state. Because of the nondeterminism, there may be paths labelled with the input string that end on nonfinal states as well as ones that do end on final states. The path is chosen in such a way that if there is any path that terminates on a final state, it will be the path selected. If there is more than one such path terminating on a final state, the particular choice of path does not matter.

One way of looking at this is that the NFA magically guesses the correct path, if one exists, and chooses its next state accordingly. When one is solving a maze in a puzzle book, the problem is made easier by one's ability to see the whole maze at once: it is all laid out on the page. But if a real copy of such a maze were made and a person were placed into it, the problem would be more difficult. The person would not be able to refer to the complete structure of the maze in deciding which way to turn at a branch in a passageway. They would have to choose blindly and explore. If they explored all of the maze, then they could produce a map that would allow them to perform the "magic" of making the correct choice at all branching points.

Another view is that whenever there is a choice to be made the NFA makes one copy of itself for each possible choice and makes a different choice in each copy. Going back to the case of the maze explorer, suppose they have the ability to duplicate themselves. At each branch, they can continue in one direction and send a duplicate in every other direction. Then the duplicates can duplicate themselves when they arrive at branch points, and so on. When the explorer (or any of their duplicates) comes to a dead end, they die. In this way, it is guaranteed that some copy of the explorer will find a way out of the maze, if one exists. This is similar to the "many-worlds" philosophy of quantum mechanics, in which every time there is a nondeterministic choice to be made, new copies of the universe arise in which the outcome is different. The advantages of NFA's over DFA's are that they usually have fewer states and that they are often easier to construct and to understand. They do have the disadvantage of being more difficult to simulate on a deterministic computer.

You may remember a regular expressions exercise in which you were given two simple regular expressions denoting the set of all strings over a, b containing at least 1 `b' and were asked to find a third simple regular expression denoting this set. The three expressions are a*b(a+b)*, (a+b)*ba*, and (a+b)*b(a+b)*. The middle term in each of these expressions is `b'. The expressions differ in terms of which `b' in each accepted string is denoted by the middle `b' term. The `b' term in the first expression must correspond to the first `b' in the string, since the first term in this expression generates only strings of `a's. Similarly, the `b' term in the second expression must correspond to the last `b' in the string. The third expression is more interesting. Its middle term can correspond to any `b' in the string. This correspondence is nondeterministic; there is a choice involved. Here is an NFA whose construction follows the form of this regular expression. Of the two other expressions, one corresponds to a DFA and the other corresponds to an NFA. Which is which? Try to construct these two automata. It may seem that both expressions must correspond to deterministic machines, since the `b' term in each of them corresponds to a specific `b' in each accepted string. But remember that a regular expression can be read from right to left as well as from left to right, whereas a finite automaton always works on its input from left to right.

Does the addition of nondeterminism add any computing power to finite automata? What we mean by "power" in this context is not the ability to do computations faster, but the ability to do computations at all. Are there any problems that can be solved, in any amount of time, by an NFA but cannot be solved by a DFA? The answer is no. This can be proven by showing, for any given NFA, how to construct a DFA to simulate it. This process is called the subset construction. We have seen that the names of the states of an automaton are arbitrary and that the states can have labels that suggest the information that they represent. Just as we can label a state "I have just seen the suffix 030" or "I have seen a `b'", we can label the states in this new DFA "The NFA that I am simulating would currently be in one of the following states: ...". So whenever the NFA is in some state q, the DFA is in a state whose label contains the name `q'. The trick here is noticing the correspondence between states and names of states. Obviously we cannot pack more than one state of the original NFA into a single state of the new DFA. But we can pack the names of several states in the NFA into the name of a single state in the DFA. We construct a DFA whose states are labelled with subsets of the names of the states of the original NFA. Suppose the NFA has states Q, alphabet , transition function , initial state q0, and final states F. Then the resultant DFA has states Q', the same alphabet , transition function ', initial state q0', and final states F', defined interms of the original (Q, , , q0, F) as follows:

An example will illustrate the construction. Here is an NFA that accepts the language of the regular expression (0+1)*00. We wish to transform this NFA into a DFA. The start state of the new machine will be [INIT]. (INIT, 0) in the NFA is the set of states INIT, ZERO_1, so '([INIT], 0) in the DFA must be the single state labelled [INIT, ZERO_1]. (INIT, 1) is INIT, so '([INIT], 1) must be [INIT]. Now we have referenced another new state in the DFA, so we need to compute its outgoing transitions. A good way to keep track of the work that still has to be done is to make a list of all the new states that are referenced. When you begin, the list should contain the single state q0', the initial state of the new machine. Each time a previously unreferenced state is referenced, add it to the list. Work in a loop, taking states off the list and computing their outgoing transitions. Since the new states are all members of the power set of the set of old states, the most there can be of them is 2. If you start with an eight-state NFA, that's only 256 states! (Fortunately, the actual number of states produced by the subset construction is usually much less than this exponential upper bound.) We continue in the construction by computing the outgoing transitions of the new state [INIT, ZERO_1]. (ZERO_1, 0) = ZERO_2, and we have already seen that (INIT, 0) = INIT, ZERO_1. So d'([INIT, ZERO_1], 0) is the state labelled with ZERO_2INIT, ZERO_1, [INIT, ZERO_1, ZERO_2]. The final construction is this.

Another extension to finite automata is the allowance of -moves. An -move is a transition that consumes no symbol. Since concatenating to a string does not change the string, any number of transitions labelled may occur in a path without affecting the string of symbols that labels it. The formal definition of an NFA with -moves is the same as that for an NFA without -moves except that d maps from QX() instead of from QX. This minor alteration in the formal definition is what allows transitions to be labelled with . -moves make it easier to build automata from regular expressions. It is easy to define recursively a way to build up an NFA with -moves that accepts exactly the language of a given regular expression. This process is described in the figure.

The subset construction works on NFA's with -moves with two slight modifications. These modifications make use of the operation -CLOSURE. For a state q, -CLOSURE(q) is the set of all states that are reachable from q by traversal of a path labelled . This path may consist of zero, one, or more transitions. Thus q is always an element of -CLOSURE(q). In the subset construction on an NFA with -moves the new initial state is the state whose label is the names of the sets in the -CLOSURE of the old initial state. So if the old initial state q0 has -moves to q1 and to q2, then the new initial state is [q0, q1, q2]. Instead of the union of all the next states of each state in the label, the new next state is the union of the -CLOSUREs of all these next states.

You have seen that DFA's, NFA's without -moves, and NFA's with -moves are all equivalent. You have also seen that there is an NFA with -moves (and therefore an NFA without -moves and a DFA) for any regular expression. This raises the question of whether or not finite automata and regular expressions are equivalent in power, of whether or not there is a regular expression for any finite automaton. We prove that this is so by showing how to construct a regular expression whose language is the same as that of a given DFA.

Number the states of the DFA from 1 to n, and let q1 be the initial state. Let R[kij] denote the set of strings that take the DFA from state qi to state qj without passing through any state numbered higher than k. To "pass through" a state is to enter it and then leave it; therefore the path may begin and/or end on a state numbered higher than k without passing through any state numbered higher than k. Since there are only n states in the machine, R[nij]denotes all the strings that label paths from qi to qj. R[kij] for any (i, j, k) can be defined recursively. Any string that takes the DFA from state qi to state qj without passing through any state numbered higher than k falls into one of two cases. It may take the machine from qi to qj without passing through k. In this case, the string is in R[k-1 i j]. Or it may cause the machine to pass through k one or more times. Such a string can be analysed into several parts. The first part will take the machine from qi to qk without passing through qk or any other state numbered higher than k-1. This substring is in R[k-1 i k]. The second part can continue through qk and then loop back to qk zero or more times. On each circuit from qk back to qk, the machine will not pass through qk; it will only begin and end on qk. So this substring is in (R[k-1 k k])*. The third part will take the machine out of qk for the final time and continue to qj. This substring is in R[k-1 k j]. So R[kij]can be defined as the union of the two cases:

R[kij] = R[k-1 i j] R[k-1 i k](R[k-1 k k])*R>k-1 k j]

The base case is R[0ij],the set of strings that takes the automaton from state qi to state qj without passing through any state. If qi and qj are not neighbours, R[0ij]must be empty, because the machine would have to pass through some intermediate state in order to get to from qi to qj. If qi and qj are neighbours, then R[0ij]is the set of all symbols that label transitions from qi to qj. If i = j, then R[0ij] contains as well as all these symbols.

R[0ij] = i!=j -> a | (qi, a) = qj

[] i=j -> a | (qi, a) = qj

Regular expressions can be constructed easily from these sets. The regular expression for the base case is the union of the symbols that label the transitions from qi to qj (and also , if i=j). A DFA accepts a string if and only if that string takes it from its initial state q1 to one of its final states. So the language of a DFA is [qjF] R[n1j]. If r[kij] is the regular expression whose language is the set R[kij], then the regular expression for the language of the DFA is r[n1 f0] + r[n1 f1] + ... + r[n1 fm], where the fi are the final states.

In the combination-lock DFA from earlier in this discussion, number INIT 1, SAW_0 2, SAW_03 3, SAW_030 4, SAW_0303 5, and SAW_03034 6. Since the only final state is SAW_03034, which has been assigned the number 6, a regular expression corresponding to this automaton is r[61 6]. We can solve for this value recursively. Only a small portion of the derivation is presented here; enduring the entire process would be nothing but tedium. For our purposes, the process of constructing a regular expression from a DFA is less important than the theoretical equivalence that the existence of such a construction algorithm establishes between finite automata and regular expressions. Here, then, is a short, representative example. Do not bother to peruse it in its entirety; it is too tedious to be of much value other than as a reference.

r[12 2] = r[02 1](r[01 1])*r[01 2]
= (0+1+2+4)(+1+2+3+4)*0
= (0+1+2+4)(1+2+3+4)*0
r[12 3] = r[02 3] + r[02 1](r[01 1])*r[01 3]
= 3 + r[02 1](r[01 1])*
= 3
r[13 3] = r[03 3] + r[03 1](r[11 1])*r[01 3]
= + r[03 1](r[11 1])*
=
r[13 4] = r[03 4] + r[03 1](r[01 1])*r[01 4]
= 0 + (r[01 1])*r[01 4]
= 0
r[14 2] = r[04 2] + r[04 1](r[01 1])*r[01 2]
= + (0+1+2+4)(+1+2+3+4)*0
= (0+1+2+4)(1+2+3+4)*0
r[14 4] = r[04 4] + r[04 1](r[01 1])*r[01 4]
= + (r[01 1])*r[01 4]
=
r[23 3] = r[13 3] + r[13 2](r[12 2])*r[12 3]
= + (r[12 2])*r[12 3]
=
r[23 4] = r[13 4] + r[13 2](r[12 2])*r[12 4]
= 0 + (r[12 2])*r[12 4]
= 0
r[24 3] = r[14 3] + r[14 2](r[12 2])*r[12 3]
= + (0+1+2+4)(1+2+3+4)*0((0+1+2+4)(1+2+3+4)*0)*3
= (0+1+2+4)(1+2+3+4)*0((0+1+2+4)(1+2+3+4)*0)*3
r[24 4] = r[14 4] + r[14 2](r[12 2])*r[12 4]
= + r[14 2](r[12 2])*
=
r[34 4] = r[24 4] + r[24 3](r[23 3])*r[23 4]
= + (0+1+2+4)(1+2+3+4)*0((0+1+2+4)(1+2+3+4)*0)*3*0
= + (0+1+2+4)(1+2+3+4)*0((0+1+2+4)(1+2+3+4)*0)*30

You can see why this process is almost never done by hand! It is easy to write a program to implement the algorithm, though.

This completes a demonstration of the equivalence of all three types of automata that we have seen and regular expressions. Here is a diagram of the reductions that we have shown between them. The arrows are labelled with the name of the proof that was used. They can be read as "can be simulated by".

By now you have noticed that there is more than one finite automaton that can accept any given regular language. Some automata are needlessly complex Rube Goldberg contraptions, with spaghetti-like transition diagrams. Is there any sure-fire way of finding the simplest automaton that will solve a given problem? And how is the hand-waving nebulosity of "simplest" made formal? It is reasonable to define "simplest" as "having the fewest possible states". An algorithm for finding a minimum-state DFA equivalent to any given DFA will be given. When combined with the algorithms already presented for conversions between regular expressions and various types of finite automata, this forms a method of finding the minimum-state DFA accepting any regular language specified in any of these formalisms.

An extended definition of the d function of a DFA will simplify the statement of the algorithm. We will extend d so that it takes a string as a parameter rather than only a single symbol. For any state q in the DFA's control and any string x over , `(q, x) will yield the state reached by the DFA after traversing the path labelled x that begins at q. In the three cases of the definition that follows, a is any symbol in the alphabet of the DFA, and x is any string over the alphabet of the DFA.

`(q, ) = q

`(q, a) = (q, a)

`(q, xa) = (`(q, x), a)

A string x is a witness for a pair of states qi, qj if and only if one of `(qi, x) and `(qj, x) is a final state and the other is not. It follows that is a witness for a pair consisting of a final state and a nonfinal state, and only for such pairs. Why does one use the term "witness"? Such a string bears witness to the fact that the two states do not always effect the same actions when given the same strings as inputs. Somewhere down the line, perhaps after many transitions, the machine can end up in a final state in one case but not in the other. This means that these two states are inherently distinct, that they must be separate in order to preserve the information that the DFA needs in order to decide whether or not some string is in its language. States in groups that have no witnesses to distinguish them from each other are equivalent. The members of each of these groups can be merged into a single state.

To find out which states, if any, are redundant, we can work backwards from the final states. We set up a triangular-shaped table that has one entry for every possible unordered pair of states. We begin by marking all pairs of a final state and a nonfinal state as non-equivalent. This is the base case. Now we loop through the table marking pairs that have witnesses, until we reach an iteration during which no marks are added. On each run through the table, for each pair of states (qi, qj) that is not already marked as having a witness, we examine (qi, a) and (qj, a) for every symbol a in the alphabet. If for some a the pair of states ((qi, a), (qj, a)) is marked as having a witness, then we mark (qi, qj) as having a witness. For if some string x is a witness for ((qi, a), (qj, a)), then ax is a witness for (qi, qj). In this way we make progress backwards along the transitions, beginning at the final states and continuing until we have covered the entire machine. Once this process is complete, we can construct a new DFA in which any and all groups of equivalent states have been merged into single states.

For example, suppose that we want to minimise this DFA. For convenience in illustrating the operation of the algorithm, we will mark entries in the table with the number of the iteration on which the two states were found to have a witness. We begin, then, by marking a 0 in all the spaces corresponding to a nonfinal state and a final state. Now, on the first iteration of the algorithm, we examine every space in the table that is still blank, trying to establish witnesses. The first pair that we look at on this iteration is (C, A). (C, 0) = D and (A, 0) = B. We go to the space in the table that corresponds to the pair (D, B) and see that this space has not yet been marked. Since we don't yet know that D and B are distinct, we can't use the 0 transition to distinguish C and A. Now we try the 1 transition. The situation here is more productive. On a 1, C goes to E and A goes to G. States E and G have been marked as distinct, so we now mark the space for the pair (C, A). Note that what we have just done is computed a witness for (C, A), namely, the string 1. (Remember that is a witness for (E, G), since E is a final state and G is not.) We continue in this manner, and at the end of our first pass through the table, the only pairs that have not yet been marked are (B, A), (D, A), (G, A), (D, B), (G, B), (F, C), and (G, D). During this first pass through the table, we have been unable to distinguish these pairs of states using either their 0 transitions or their 1 transitions. However, we did distinguish some pairs during this pass, and this new information may allow us to make more distinctions in our next pass. On the second iteration, we mark (B, A), (D, B), (G, B), (F, C), and (G, D). On the third, we mark (G, A). Going into the fourth iteration, the only unmarked space is (D, A). But states D and A both have the same outgoing transitions. Therefore, there is no way to distinguish them. Since no more spaces in the table can be marked, the algorithm terminates. The minimised DFA has states D and A merged. Here is how the table looks at the end of the algorithm:

B	2
C	1	1
D		2	1
E	0	0	0	0
F	1	1	2	1	0
G	3	2	1	2	0	1
	A	B	C	D	E	F

MORE EXERCISES

Turing and an Abstract Model of Computation

A century after Babbage, Alan Turing, with the backing of electronics, was instrumental in bringing the automatic computer to realisation. Like Babbage, Turing interested himself in the mathematical theory of computation and in the practical engineering needed to implement mathematics in a physical computer. In Turing, this engineering interest took the form of electronics. Had Babbage lived a century later than he did, he would have concerned himself with electrical engineering instead of mechanical engineering. Whereas Babbage's encoding of mathematics was mechanical, Turing's was electrical. Turing did experiment with a special-purpose mechanical computer designed to calculate zeroes of the zeta-function, a function describing the thinning out of the prime numbers. In another echo of Babbage, Turing never completed construction of the zeta-function machine, although his progress was interrupted by World War II instead of by a funding cut.

Just before the war Turing published a paper entitled "On Computable Numbers, with an Application to the Entsceidungsproblem". The Entscheidungsproblem, proposed by Hilbert, was the question of whether or not there was a mathematical procedure for ascertaining the decidability of any mathematical problem. That is, given some mathematical assertion, is there a way to decide whether or not that assertion is provable? Gödel had already proven the existence, within all mathematical systems of sufficient expressive power, of statements that are true but unprovable within the system. Thus the Entscheidungsproblem had great importance: if there were always a procedure for ascertaining whether or not an assertion were provable, then the unprovable statements of Gödel could be detected and pruned from mathematics, excluded from mathematical enquiry. In "On Computable Numbers" Turing showed that in the general case these statements could not be detected. The mathematical model of computation advanced by Turing in "On Computable Numbers", which became known as the "Turing Machine", has been a fundamental abstraction in the theory of computation. A Turing Machine consists of a finite control, corresponding to the program in a modern electronic computer or the sequences of punched cards that were to be fed into the Analytical Engine, and a semi-infinite tape, corresponding to a computer memory. The finite control is like a finite automaton. It can remember information by itself, but only a finite amount of information. The tape is the big difference between a Turing Machine and a finite automaton. Since the tape is semi-infinite (it has a left end, but is unbounded on the right), it can be used to store an unbounded amount of information. The tape is divided into cells. On each tape cell a symbol from an alphabet can be written. Remember that we think of "alphabet" in the abstract sense here, as an arbitrary collection of symbols. If we view each byte in the memory of an electronic computer as one cell, then the alphabet of the computer can be viewed as the integers from 0 to 255 inclusive (255=28-1). The only restriction upon it is that the number of symbols must be finite. The control, the program, is a collection of rules that describe the behaviour of the Turing Machine in any state. You can see why the control is restricted to being finite; an infinitely long program could not be specified explicitly. (Even if you spent all your life on it, you would make no progress in writing an infinitely long program. This would be a hacker's nightmare.) The control is always in one of a finite number of states, just as in a finite automaton. At any time during the operation of the Turing Machine, its configuration is fully described by the state of the finite control, the contents of the tape, and the position of the read/write head. Similarly, in an electronic stored-program computer, the configuration is fully described by the location of the instruction within the program that is about to be executed and the contents of the memory. Confronted with a symbol under the read/write head while in some state, the finite control replaces that symbol with another symbol (possibly the same symbol) and either moves the head left one cell, moves the head right one cell, or halts. Then the control enters a new state (unless the machine has halted). This simple construction of finite control and memory, reminiscent of the Analytical Engine, can compute any function that can be computed. Because of this equivalence of the Turing Machine to all other general-purpose computers, computer scientists can prove theorems about the simple Turing Machine and be assured that those theorems apply to all the complexities of real-world computing devices. Finite automata can be viewed as a special case of Turing Machines in which the control is restricted from writing on the tape and from moving the read/write head left.

Formally, a Turing Machine consists of seven objects (Q, , , , q0, B, F):

need not be defined for all possible pairs of a state and a tape symbol. In cases in which d is undefined, the machine halts. If it is in a final state when this occurs, it accepts. Otherwise, it rejects. Sometimes it does not matter whether the machine accepts its input or not, for for some problems the output is not in the form of an acceptance or rejection decision, but in the form of an output string left on the tape when the machine halts. The input alphabet is as in a finite automaton. But in a Turing Machine there is also a tape alphabet that contains the symbols that the Turing Machine can write onto its tape. This tape alphbet, , must contain at least one symbol that is not in . This symbol is the blank. When the Turing Machine begins running, the leftmost portion of its tape contains the input string, and the tape is padded on the right with an infinity of blanks. You can see why this special blank symbol is needed. If the input string were allowed to contain blanks, it would be impossible for the Turing Machine to find the end of the input. It could scan outward on its tape and find a million blanks, and it would still not be sure whether it had passed the end of its input, because there would still be the possibility that the next character might be non-blank.

A simple language that can be accepted by a Turing Machine and not by any finite automaton is {0n1n | n>=0}. (This differs from 0*1*, which can be accepted by a finite automaton. 0n1n is a subset of 0*1* in which the number of 0's must equal the number of 1's.) If this is not obvious to you immediately, try to construct a finite automaton that accepts it. What barrier do you run up against? A finite automaton cannot count indefinitely. Sooner or later it reaches the limit of its storage capacity, whatever that capacity is. The only way that a finite automaton can remember a piece of information is by changing its state, and there can be only finitely many states. For any finite automaton, the language {0n1n | n>=0} contains an infinity of instances that are beyond the storage capacity. Another interesting and simple example is the question of deciding whether a given number is prime. (The number can be given in any base. The simplest case to deal with is that of a unary number, in which any number n is represented by a string of 0's (or any convenient symbol) of length n).

Interestingly, "On Computable Numbers" came close to superfluity. When Turing released his paper, Alonzo Church, a mathematician at Princeton, had recently published his own solution to the Entscheidungsproblem. Turing's paper had great significance, though, because of the nature of the mathematical model of computation that it introduced. Church's approach had been completely abstract. But the Turing Machine bridges the gap between physical reality and abstract model. It is a mathematical entity, and yet one speaks of it in terms of familiar physical parts such as tapes and read/write heads. Babbage would have appreciated this closing of the gap between engineering and theory.

Many arguments about Turing Machines involve hand-waving because the mathematical rigour of detailing a Turing Machines's transition function is drudgery, and nobody likes to do it. This example will be mostly in English, although there will be some formal notation. A Turing Machine is given a binary number, in the conventional notation of zeroes and ones, stored on its tape with the least significant digit on the left and the most significant digit on the right. (Conventionally, numbers are written in the opposite direction, with the most significant digit on the left, but since the tape is bounded on the left, it is easier and makes for a clearer presentation to extend the number on the right instead of moving everything one cell to the right and placing the result of a carry on the left.) The Turing Machine must increment the number, leaving the result on the tape. In this problem it does not matter whether the Turing Machine accepts or rejects; we are interested only in the contents of the tape when the machine halts.

What is the information that the Turing machine should remember in its finite control? The only way that a previous stage of the computation can affect a later stage is through a carry out of one digit's place and into the next. So the finite control should remember whether or not it must add in a carry during the current operation. Two states suffice for this; they can be labelled "CARRY" and "NO_CARRY".

What should be the initial state? Since the machine must add 1 to its input, it should begin in CARRY.

Now we reach the most complex part of the design of a Turing Machine, the transition function. In this example the d function is rather simple. Here is a table for it:
STATE TAPE SYMBOL ACTION
NO_CARRY 0 (NO_CARRY, 0, right)
1 (NO_CARRY, 1, right)
B undefined; halt when this case is encountered
CARRY 0 (NO_CARRY, 1, right)
1 (CARRY, 0, right)
B (NO_CARRY, 1, right)

This function specifies that whenever the machine is in state NO_CARRY it should leave the contents of the tape unchanged. (Since there is no way to get out of NO_CARRY once this state has been entered, we could just as easily have omitted NO_CARRY altogether and had the machine halt in cases where it would have entered NO_CARRY.) When the machine is in CARRY and sees a 0 on the tape, it writes a 1, representing the original 0 with a carry added in, and enters NO_CARRY. When the machine is in CARRY and sees a 1 on the tape, there will be a carry into the next column, so it writes a 0 and stays in CARRY. If the machine sees a blank while it is in CARRY, this means that there is a carry out of the last column, and the number becomes one digit longer. In this case, the blank is replaced with a 1.

During the war Turing was involved in the secret effort of the Government Code and Cipher School to break the German "Enigma" code. The code was generated by a simple mechanical computer consisting of rotors and plugboards. Turing was a godsend for the British war effort, and this intense, prolonged intimacy with coding was beneficial in his later development of electronic stored-program computers. Turing and other members of the deciphering effort designed electromechanical devices to step through combinations and detect possible Enigma encodings. These machines, along with the Enigma device which they were designed to defeat, were a link between the mechanical computing "engines" of Babbage and the purely electrical machines that were to appear after the war.

At the end of the war, Turing joined the staff of the National Physical Laboratory, and wrote a proposal for the construction of an electronic computer. The proposed machine was dubbed ACE, an acronym for "Automatic Computing Engine", in recognition of Babbage's Analytical Engine. Other teams in Britain and in the United States were also at work on electronic stored-program computers. Unlike Babbage's machines, these would use the same memory for both data and instructions. Whereas Babbage's machines continually read their instructions from punched cards and used their internal memories only for representing the numbers to which the program was referring, stored-program computers take advantage of the fact that a program can be viewed as nothing more than a sequence of numbers. If a number, an instruction code, is assigned to each operation that the machine can perform, then a program, which is a sequence of operations, reduces to a numerical sequence. In fact, programs and data become interchangeable. This presents the delightful possibility of programs that rewrite themselves, or programs that write other programs.

When I was first learning assembly language, on an old 8-bit microprocessor with 16-bit pretensions, the Motorola 6809, one of my favourite tricks was this:

ENTRY: CMPX #$8603

...loop body... BNE ENTRY+1

In this instance, what I wanted was a loop in which a register, A (In case you know nothing about computer architecture, a register can be thought of as just a special kind of variable), would be loaded with the value 3 at the beginning of every iteration but the first, but would have its value preserved on the first iteration. So I hid the code for the "A := 3" instruction ("LDA #$03" in 6809 assembly language) inside the operand of a harmless instruction, CMPX#, which, in the context of this loop, had no significant effect. So, when the processor entered the loop and began the first iteration, it would execute the harmless instruction, skip over the "A := 3" that was hidden in the operand of that instruction, and therefore leave the contents of A undisturbed. But, the branch instruction at the bottom of the loop, which tells the processor what point to return to to begin the next iteration, points to a location one byte past the CMPX#. (ENTRY is the label of the CMPX# instruction, so ENTRY+1 points just past the code for the CMPX# in memory.) So now the machine treats what was previously an operand as an instruction code! It sees $86 (`$' indicates a base-sixteen number), and treats that as the code for the instruction LDA#, which takes a one-byte operand, in this case 3. Babbage's machine would never have been able to perform this sort of devilish trick. It used separate streams of cards for data and for instructions, and so could not interchange them. This is a small example of the sort of ingenious finagling that the programmers and operators of the first electronic stored-program computers puzzled over every day. By finding creative ways to represent information inside a computer memory, they could make their machines sing the music of the spheres.

The first one of these machines to operate was constructed at Manchester University. It used luminous dots written on cathode ray tubes (CRT's, merely black-and-white television screens) for storing its instructions and data. The only output mechanism was the primitive one of peeking at the CRT's and viewing their patterns of dots as the binary data inside the machine's memory. If you have ever played with graphics on a microcomputer then you can understand the simple delight of this: there on the screen was a window into the essence of the machine's computation. What appeared to the eye as random splotches of light and dark was known by the intellect as the carrier of all the information of the problem that the machine had been programmed to solve. One of the first things I did with my old 6809 system when I learned how to program its graphics hardware was to write a short routine that set the graphics hardware to display the program and the data that it was modifying. So on the screen, as light and dark spots, appeared the instructions and data of the very program that had issued the command that caused the screen to display it. This delightful little circularity conveys the beauty of encodings and ciphers.

Ironically for one so aware of the omnipresence of systems, Turing was maltreated by the system of English society. Turing was homosexual, and when he inadvertently revealed to the police that he had had homosexual intercourse he was arrested and charged with the crime of "gross indecency". He was compelled to submit to injections of oestrogen, a female sex hormone that was thought to be a remedy for his "disease". After he had served this sentence, when he seemed to have passed the nadir, he killed himself by eating an apple laced with cyanide. He left no explanation for his suicide, at least not explicitly. The explanation lies encoded somewehere within the story of Turing's life. Hugh Whitemore, author of the biographical play Breaking the Code, believes that Turing's suicide was the ultimate experiment in artificial intelligence. In his famous article "Computing Machinery and Intelligence" in the British philosophical journal Mind, Turing had discussed the idea of a deterministic machine simulating human behaviour. If, he said, the machine could not be distinguished objectively from a human, then one would have to credit it with intelligence. This raises the existential question of whether or not humans are biological automata, whether death is the condition of a machine halting. Can mind exist without body? If so, in what manner? Turing had been fascinated by this since his childhood. He could find his answer only by first-hand experience. If this was not the primary motivation for Turing's suicide, it must at least have been in his mind when he bit into the apple. Turing's death served a dual purpose: it was an escape from social systems as well as a way of possibly answering the mind-body question. Since the end of the war, Turing had been under pressure not only because of his homosexuality, but also because of his way of learning about the world. The academic system within which Turing was operating enforced a duality of pure science and engineering. The administrative entities within which Turing was placed were organised around this division between theory and application, so Turing was pressured, not so much by any individual administrators as by the administrative system, to be either a theorist or an implementor, but not both. This was contrary to his nature and must have been uncomfortable for him.

EXERCISE

Suppose that a Turing Machine constructor produces defective Turing Machines. In these machines, whenever the finite control calls for the head to move left, the head sticks where it is. What kind of languages is accepted by these machines? Why?

Extensions to Turing Machines

There are many bells and whistles that can be added to Turing Machines that simplify their construction by hiding the nitty-gritty details. However, none of them constitute an increase in computing power over the capability of the basic Turing Machine, because all these extensions can be simulated on a basic Turing Machine.

One useful extension is the multiple-track Turing Machine. On a Turing Machine with k tracks, k symbols at a time are under the read/write head. A basic Turing Machine with tape alphabet ' can be viewed as a multi-track Turing Machine with tape alphabet . The key is to pack a column of k symbols of the alphabet of the k-track Turing Machine into one symbol of the alphabet of a basic Turing Machine. The elements of ' will be k-vectors [a0, a1, ... ak-1] of the elements of . No changes or fancy encodings need to be made in states or transitions, because the k-track machine has only one read/write head and thus cannot treat symbols that are all in one column independently. There are examples in human languages of two symbols being treated as a single symbol. Although these symbols seldom occur in Modern English, Old English used the symbol "æ" to represent the vowel sound in "mat", for example. This is reminiscent of the encoding of names of states that was used in the subset construction.

A multi-tape Turing Machine is a generalisation of the multi-track Turing Machine. It has k tapes, each with its own independently mobile read/write head. Such a machine can be easily simulated by a 2k-track machine, and therefore by a basic machine. Each tape is implemented with two tracks. One of these tracks contains the contents of the tape. The other contains all blanks except for a single non-blank symbol. This symbol marks the position of the read/write head for the tape. The transition function manipulates the contents of the data track and moves this marker symbol along the marker track in accordance with the actions of the k-tape machine being simulated. (Yes, this argument is hand-waving. I warned you.)

A machine with a tape that extends infinitely in both directions can be simulated by a 2-track machine. There is a nice image for this simulation: the two-way infinite tape is snapped at its origin, and half of it is rotated in a semicircle to match up with the other half. There is one track for each half of the tape. But why stop at a recording surface (or a recording solid, or a recording hyper-solid) that is infinite only in one dimension? If the tape[45] is an infinite plane, an infinite solid, an infinite object in a space of any number of dimensions, then it still can be simulated using the one-dimensional, semi-infinite tape of the basic Turing Machine. This is because any such multi-dimensional tape that is divided into cells contains the same infinite number of cells as a one-dimensional, semi-infinite tape. (Specifically, they both have 0 cells.) They may be more immediately accessible in higher dimensions, but they are all still there and are accessible, given enough time for the read/write head to move, in a tape of any dimension. To understand why this is so, see the appendix on infinity. All that needs to be done is to define some arbitrary one-to-one mapping between the 0 cells of one tape and the 0 cells of the other and then implement that mapping in the finite control of the one-dimensional Turing Machine. You should take a moment to understand how this simulation can be combined with the previous one to simulate a machine with k two-way infinite tapes (or k multi-dimensional tapes) using a 4k-track machine.

One naïve objection to the statement that Turing Machines can do anything that real-world computers can do is "they can't do graphics". This is untrue. Of course if a basic Turing Machine were actually a physical construction it would not have the display hardware necessary to output graphics in a visual form. But a graphics display is only one physical way of regarding the abstract mathematical construct of points in a coördinate system. A Turing Machine is capable of computing such points. Therefore, minus the fancy displays, Turing Machines can do graphics. Think of a Cartesian (i.e. rectilinear) coördinate system. It can be represented using 1's and 0's on a two-dimensional recording surface of a Turing Machine. If a point can be one of several colours, then instead of just 1's and 0's we can have a symbol for each possible colour. In general, a Turing Machine can do graphics in any coördinate system (polar, cylindrical, or spherical, for example) and in any number of dimensions by simulating a tape of the desired number of dimensions that has its cells laid out according to the desired coördinate system.

As was the case with finite automata, nondeterminism does not add any computing power to the Turing Machine. A deterministic Turing Machine is capable of simulating every case in which a nondeterministic Turing Machine might find itself. The deterministic machine can allocate separate areas on its tape for each case. Whenever a nondeterministic choice is encountered, a new data area is created. When there is no more room in a data area, all the areas stored to the right of it on the tape are shifted right to make more room. There is the possibility that one of the simulations might not halt but instead enter an endless loop. We will see shortly that there is no algorithm to detect whether an algorithm can ever get into an endless loop, so this case cannot be anticipated. If the cases were simulated sequentially, with the simulating machine completing one case before embarking upon the next, then a nonterminating case would prevent the simulator from ever proceeding to the next case. Therefore, the simulator must be time-shared between all the cases of the simulation. The simulator is a loop that visits each case once on each pass. On each visit, one move is accomplished. With this arrangement, infinite loops cannot monopolise the simulator.

EXERCISES

In the problems that follow, remember not to get too nitty-gritty in your descriptions.

Turing Machine Encodings and the Universal Turing Machine

Turing machines can be encoded as strings. States and symbols can be assigned numbers. Since symbols can be assigned numbers, and a machine with tape alphabet 0, 1, B can represent any number (in unary, using 1 as a delimiter between numbers) using these symbols on its tape, a Turing Machine with tape alphabet 0, 1, B can simulate any Turing Machine. Therefore the following discussion can and will be restricted to such machines. A template for the encoding of a transition could be

<state>1<symbol>1<nonfinal/final><state>1<symbol>1lt;left/right>

where the "<state>" and "<symbol>" templates represent state and symbol numbers in unary, "<nonfinal/final>" represents either 0, if the following state code designates a nonfinal state, or 1, if it represents a final state, and "<left/right>" represents either 0 for left or 1 for right. There are many other encodings that would work as well; this is an arbitrary choice. The 1's in the template are delimiters. Since we have already used a supermarket example, in describing nondeterminism, we shall use another here. The plastic sticks that are placed between different customers' purchases at a supermarket checkout serve the same function as the 1's that we are placing between individual codes in this Turing Machine encoding scheme. They mark the end of one order (or the end of one code) and the beginning of the next. A transition function can be encoded as a concatenation of transition encodings. Note that in order to encode a Turing Machine, all one must do is encode its transition function, since the input alphabet and the tape alphabet are fixed as 0, 1 and 0, 1, B, respectively, and the states are defined implicitly by their occurrences in the transition function. The initial state can be designated by assigning it the special code of zero. (In unary, this code would be represented by , the string of 0's of length zero.) This encoding of a Turing Machine M can be denoted e(M), viewing e as an encoding function. So that e can be defined for all strings of 1's and 0's, we can define any string of 1's and 0's that does not match the above format to be an encoding of the Turing Machine that has no states. We now have a method of putting all the information about a Turing Machine into a string. And strings can be used as data for Turing Machines. This suggests interesting possibilities.

The Universal Turing Machine takes as its input a Turing Machine encoding and another string. It simulates the specified machine with this string as its input. The universal machine uses three tapes: one to store the encoding of the machine being simulated, one to store the current state of the simulation, and one to store the contents of the tape of the machine being simulated. Note that the tape that stores the encoding is read-only, because the encoding is never modified. On each step of the simulation, the universal machine looks up the proper move for the current combination of state and tape symbol, and alters the state and the tape contents of the simulation accordingly. A universal Turing Machine, compared with a special-purpose Turing Machine, is like a general-purpose, programmable computer in comparison with a dedicated system. A common type of dedicated system in the real world is the word processor. Such a computer can edit pieces of text, but it cannot be programmed to do anything else. A general-purpose computer can simulate any special-purpose computer. In particular, it can simulate a word processor by executing a text-editing program. We can denote a Turing Machine M when run on an input string x as M[x]. Call the Universal Turing Machine U. Then for any Turing Machine M and any string x, U[<e(M), x>] yields the same result as M[x] does.

The Halting Problem: An Undecidable Problem

Note that in some cases a Turing Machine may neither accept nor reject, but instead may loop infinitely. A finite automaton cannot loop infinitely, because eventually it must reach the end of its input. But giving the Turing Machine the power to rewrite its input also gives it the capability of looping infinitely. It is desirable to produce Turing Machines that halt on all inputs, just as it is desirable to produce computer programs that can never enter infinite loops. A Turing Machine that halts on all possible inputs is called a total Turing Machine. Turing Machines that are not total may halt on some of their inputs, but waiting for a nontotal Turing machine to halt is a never-ending task; one can never be sure if the machine is still doing useful computation and might halt even within the next few transitions or whether it has entered an infinite loop. (You may have had the experience of sitting in front of a computer that seemed to be taking a long time to finish and wondering whether or not you should interrupt it and see whether it had entered an infinite loop.) If a language is accepted by some total Turing Machine, the language is termed recursive. (The meaning of the word "recursive" in this domain should not be confused with its meaning in the domain of program implementation strategies; although recursive sets and recursive programs do bear a close relation, it can be confusing for beginning students.) One might think that every language is recursive, that is, that every well-defined problem can be solved by a computer. In this case, obtaining the solution would simply be a matter of constructing a correct algorithm. But there are many interesting problems that cannot be encoded as recursive languages. These problems are undecidable.

A problem A is reducible to another problem B if and only if there is a recursive function that maps instances of A to instances of B. The statement "A is reducible to B" is denoted A <= B. This `<=' notation arises from the notion of a hierarchy of problems; problems higher up in the hierarchy are more difficult than problems lower down, and any given problem is reducible to problems that are as hard or harder. Thus `<=' defines a classification of problems in terms of difficulty.

We want to know whether it is possible to design an algorithm that takes as its input another algorithm and the input to that algorithm, and decides whether or not the given algorithm will ever halt after it has been set running on the given input. In Turing Machine terms, we want a machine H that, when given <e(M), x> for any Turing Machine M and any string x, will accept if and only if M would eventually halt (i.e., accept or reject) if run on x, and reject otherwise. A good way to see if such an algorithm can exist is not to go through all the trouble of trying to construct it but instead to assume that it exists and then to see if its existence implies a contradiction. If its existence does imply a contradiction, then it cannot exist, and we need not bother to try to construct it. This method of proof by contradiction is called indirect proof. It may be familiar to you from geometry. So suppose H exists. Then we can construct from H a machine M* that behaves as follows. When run on a string over 0, 1 that is e(M) for some M (remember that e is defined for all strings over 0, 1), M* simulates H[<e(M), e(M)>] and

* halts iff M[e(M)] does not halt

* loops endlessly iff M[e(M)] halts

This simulation of H decides whether or not M halts when given its own encoding as input. In a mind with an intuition for paradox, this raises the question of what happens when M* is given its own encoding as input. Think it through for yourself using the definition of M*, then read on.

What we have is a contradiction. M*[e(M*)] halts if and only if it does not halt, and does not halt if and only if it halts. (Therefore, the moon is made of green cheese.) So we must have made some false assumption. Note that L(M*) is reducible to L(H). The recursive function that maps elements of L(M*) to elements of L(H) is implemented in the above-mentioned modifications to the transition function of H. We made no assumption in the construction of M* from H; that process is perfectly well-defined. The only assumption that has been made is that H exists; the only way that M* cannot exist is if H, the machine on which it is based, also cannot exist. Therefore there is no algorithm for the halting problem; the question of whether or not any given algorithm halts on any given input is provably undecidable.

EXERCISES

A Taste of Gödel's Incompleteness Theorem

"Logic dictated that logic did not apply." This was the statement used by Mister Spock in a Star Trek episode to give logical justification to an action that seemed the product of illogic[46]. It is also an informal statement of a logical truth first proven by Kurt Gödel in 1931. At that time mathematics had been trying to set its own house in order ever since the days of the Cambridge Analytical Society. There had been too many loopholes that had arisen through lack of rigorous demonstrations of mathematical truths, and it had taken about a century to clean them up. By the early twentieth century mathematicians were searching for the pièce de résistance of mathematical rigour, a proof of the consistency and completeness of mathematics. Consistency of a mathematical system means that it is impossible to use the axioms and inference rules of the system to prove anything that is false. Inconsistent mathematical systems are not of much use, because mathematics is boring when, like the Bellman, one can prove anything that one wants to prove. Completeness means that it is possible to prove every true proposition that can be expressed within the syntax and semantics of the system. Obviously, completeness is a desirable property, for when one is trying to come up with a proof of a conjecture that seems true, one would like to be assured that one is not on a wild goose chase. But Gödel showed that any mathematical system with enough expressive power to encompass number theory is necessarily incomplete or inconsistent. This proof is one of the most famous and one of the most valuable demonstrations of the old saw that you can't have your cake and eat it too.

Gödel accomplished his proof by encoding mathematical symbols using numbers, in a scheme called Gödel numbering, and then using this encoding to construct a mathematical proposition that, when decoded, expresses a mathematical equivalent of the English sentence "This statement cannot be proven". The coding is what allows the self-reference of such a statement; it can talk about itself on two levels, within the code and outside the code. The proposition must be either true or false. If it's true, then, as it declares, it can't be proven. In this case the mathematical system that generated this statement is incomplete, since the statement is an example of a proposition that is true and yet has no proof within the system. If it's false, then it can be proven, which means the system is inconsistent, since the statement is an example of a false proposition that has a proof within the system.

We can construct a Turing Machine E to enumerate theorems in any formal system by successively applying inference rules to statements already known to be true. Eventually, any given theorem will be enumerated by such a machine. (Of course, most of the theorems will not be interesting, but the interesting ones will all be there. This is like the old story about the monkeys at the typewriters eventually producing the complete works of Shakespeare, or the situation in Borges's Library of Babel.) We can use E to create a machine called G that has the following behaviour: G simulates E. If and when this simulation of E produces a theorem that states that G does not halt, G halts. Think about the implications of this construction for the consistency and completeness of the system whose theorems are enumerated by E, then read on.

If G halts, then the formal system whose theorems it has been enumerating is inconsistent, because the statement that G does not halt is provable but untrue. If G does not halt, then the system is incomplete, because it generates no proof that G does not halt. An adversary to this argument might say, "Well, if my mathematical system can't express the fact that G never halts, then I'll just add that fact explicitly, as an axiom". This would be like trying to plug one small hole in a disintegrating dyke: another leak will automatically spring up. For in adding this axiom to the system, the adversary has changed the system, so the original enumeration machine no longer applies. We now need a slightly altered machine, E', to enumerate theorems in this new system. We can use this new enumeration machine to construct a machine G', and we will find that the new system is still incomplete, because it fails to include as a theorem the true statement that G' never halts. Thus, whenever the adversary attempts to wriggle out from under the effects of the Incompleteness Theorem, he is easily caught again. You should be aware that this is only a quick, hand-waving argument. The complete, mathematically rigorous Incompleteness Theorem is a gem, an opus.

Gödel's Incompleteness Theorem is the capstone to the wave of mathematisation begun by Newton. It shows that there is a limit; mathematics cannot encompass all truth. This has been reassuring to humanists. It has reassured me of the purpose in my nonmathematical work. If, as the western scientific tradition suggests, the universe really is nothing but a mathematical system, then anything that happens, including all thought, must be expressible in terms of number theory. Can it be so? This is what Turing was trying to decide when he bit into the apple. If this question is ever decided, a very fundamental harmony will have been revealed. But humans will not be able to say that they have found the key to all truth. For if the universe is simply a mathematical system, then it contains truths that cannot be reached using logic. What might such unprovable truths be? The enlightenments of Zen, perhaps? When I first studied formal logic I was filled with the same joy that filled the Newtonian mathematisers: everything could be reduced to symbol manipulation, I thought; it was feasible that eventually there would be no need for the messy ambiguities of literature and philosophy. How narrow-minded and ignorant that sentiment was! I thought that I had a logical reason to abandon the humanities when, in actuality, logic was saying exactly the opposite.


NEXT CHAPTER: 3. Conclusion
[26][Carroll 1936] p. 17.
[27][Carroll 1936] p. 761.
[28][Carroll 1936] p. 769.
[29][Carroll 1936] p. 757.
[30][Carroll 1936] p. 770.
[31][Carroll 1936] p. 774.
[32][Atherton 1960]
[33][Joyce 1939] p. 374.
[34][Carroll 1936] p. 124.
[38]David Gries, in an electronic message to the Cornell University Department of Computer Science dated March 9, 1989. Used by permission.
[39][Morrison] p. 143.
[40][Morrison] p. 252. Ada's notes on Menabrea.
[41]The Golden Age ended sometime in the early 1980's, and now the media have popularised the word "hacker" as a term for one who maliciously enters and damages computer systems to which they do not have access. This is unfortunate.
[42]William Whewell, Astronomy and General Physics, Considered with Reference to Natural Theology 1/e. London: William Pickering, 1833. p. 334.
[43][Babbage 1838]
[44]The proofs in this section appear in [Hopcroft & Ullman 1979].
[45]I use the word "tape" here, for lack of a concise substitute, to designate a recording material of any number of dimensions.
[46]The episode was "The Galileo Seven".