A while back we featured a post about why learning mathematics can be hard for programmers, and I claimed a major issue was not understanding the basic methods of proof (the lingua franca between intuition and rigorous mathematics). I boiled these down to the “basic four,” direct implication, contrapositive, contradiction, and induction. But in mathematics there is an ever growing supply of proof methods. There are books written about the “probabilistic method,” and I recently went to a lecture where the “linear algebra method” was displayed. There has been recent talk of a “quantum method” for proving theorems unrelated to quantum mechanics, and many more.

So in continuing our series of methods of proof, we’ll move up to some of the more advanced methods of proof. And in keeping with the spirit of the series, we’ll spend most of our time discussing the structural form of the proofs. This time, diagonalization.

## Diagonalization

Perhaps one of the most famous methods of proof after the basic four is proof by diagonalization. Why do they call it diagonalization? Because the idea behind diagonalization is to write out a table that describes how a collection of objects behaves, and then to manipulate the “diagonal” of that table to get a new object that you can prove isn’t in the table.

The simplest and most famous example of this is the proof that there is no bijection between the natural numbers and the real numbers. We defined injections, and surjections and bijections, in two earlier posts in this series, but for new readers a bijection is just a one-to-one mapping between two collections of things. For example, one can construct a bijection between all positive integers and all *even* positive integers by mapping to . If there is a bijection between two (perhaps infinite) sets, then we say they have the same size or *cardinality*. And so to say there is no bijection between the natural numbers and the real numbers is to say that one of these two sets (the real numbers) is somehow “larger” than the other, despite both being infinite in size. It’s deep, it used to be very controversial, and it made the method of diagonalization famous. Let’s see how it works.

**Theorem: **There is no bijection from the natural numbers to the real numbers .

*Proof.* Suppose to the contrary (i.e., we’re about to do proof by contradiction) that there is a bijection . That is, you give me a positive integer and I will spit out , with the property that different give different , and every real number is hit by some natural number (this is just what it means to be a one-to-one mapping).

First let me just do some setup. I claim that all we need to do is show that there is no bijection between and the real numbers *between 0 and 1*. In particular, I claim there is a bijection from to all real numbers, so if there is a bijection from then we could combine the two bijections. To show there is a bijection from , I can first make a bijection from the open interval to the interval by mapping to . With a little bit of extra work (read, messy details) you can extend this to all real numbers. Here’s a sketch: make a bijection from to by doubling; then make a bijection from to all real numbers by using the part to get , and use the part to get by subtracting 1 (almost! To be super rigorous you also have to argue that the missing number 1 doesn’t change the cardinality, or else write down a more complicated bijection; still, the idea should be clear).

Okay, setup is done. We just have to show there is no bijection between and the natural numbers.

The reason I did all that setup is so that I can use the fact that every real number in has an infinite binary decimal expansion whose only nonzero digits are after the decimal point. And so I’ll write down the expansion of as a row in a table (an infinite row), and below it I’ll write down the expansion of , below that , and so on, and the decimal points will line up. The table looks like this.

The ‘s above are either 0 or 1. I need to be a bit more detailed in my table, so I’ll index the digits of by , the digits of by , and so on. This makes the table look like this

It’s a bit harder to read, but trust me the notation is helpful.

Now by the assumption that is a bijection, I’m assuming that *every* real number shows up as a number in this table, and no real number shows up twice. So if I could construct a number that I can prove is not in the table, I will arrive at a contradiction: the table couldn’t have had all real numbers to begin with! And that will prove there is no bijection between the natural numbers and the real numbers.

Here’s how I’ll come up with such a number (this is the diagonalization part). It starts with 0., and it’s first digit after the decimal is . That is, we flip the bit to get the first digit of . The second digit is , the third is , and so on. In general, digit is .

Now we show that isn’t in the table. If it were, then it would have to be for some , i.e. be the -th row in the table. Moreover, by the way we built the table, the -th digit of would be . But we *defined * so that it’s -th digit was actually . This is very embarrassing for (it’s a contradiction!). So isn’t in the table.

It’s the kind of proof that blows your mind the first time you see it, because it says that *there is more than one kind of infinity*. Not something you think about every day, right?

## The Halting Problem

The second example we’ll show of a proof by diagonalization is the Halting Theorem, proved originally by Alan Turing, which says that there are some problems that computers can’t solve, even if given unbounded space and time to perform their computations. The formal mathematical model is called a Turing machine, but for simplicity you can think of “Turing machines” and “algorithms described in words” as the same thing. Or if you want it can be “programs written in programming language X.” So we’ll use the three words “Turing machine,” “algorithm,” and “program” interchangeably.

The proof works by actually defining a problem and proving it can’t be solved. The problem is called *the halting problem*, and it is the problem of deciding: given a program and an input to that program, will ever stop running when given as input? What I mean by “decide” is that any program that claims to solve the halting problem is itself required to halt for every possible input with the correct answer. A “halting problem solver” can’t loop infinitely!

So first we’ll give the standard proof that the halting problem can’t be solved, and then we’ll inspect the form of the proof more closely to see why it’s considered a diagonalization argument.

**Theorem: **The halting program cannot be solved by Turing machines.

*Proof. *Suppose to the contrary that is a program that solves the halting problem. We’ll use as a black box to come up with a new program I’ll call meta-, defined in pseudo-python as follows.

[code language=”python”]

def metaT(P):

run T on (P,P)

if T says that P halts:

loop infinitely

else:

halt and output "success!"

[/code]

In words, meta- accepts as input the source code of a program , and then uses to tell if halts (when given its own source code as input). Based on the result, it behaves the *opposite* of ; if halts then meta- loops infinitely and vice versa. It’s a little meta, right?

Now let’s do something crazy: let’s run meta- on itself! That is, run

[code language=”python”]

metaT(metaT)

[/code]

So meta. The question is what is the output of this call? The meta- program uses to determine whether meta- halts when given itself as input. So let’s say that the answer to this question is “yes, it does halt.” Then by the definition of meta-, the program proceeds to loop forever. But this is a problem, because it means that `metaT(metaT)`

(which is the original thing we ran) actually does not halt, contradicting ‘s answer! Likewise, if says that `metaT(metaT)`

should loop infinitely, that will cause meta- to halt, a contradiction. So cannot be correct, and the halting problem can’t be solved.

This theorem is deep because it says that you can’t possibly write a program to which can always detect bugs in other programs. Infinite loops are just one special kind of bug.

But let’s take a closer look and see why this is a proof by diagonalization. The first thing we need to convince ourselves is that the set of all programs is countable (that is, there is a bijection from to the set of all programs). This shouldn’t be so hard to see: you can list all programs in lexicographic order, since the set of all strings is countable, and then throw out any that are not syntactically valid programs. Likewise, the set of all inputs, really just all strings, is countable.

The second thing we need to convince ourselves of is that a *problem* corresponds to an infinite binary string. To do this, we’ll restrict our attention to problems with yes/no answers, that is where the goal of the program is to output a single bit corresponding to yes or no for a given input. Then if we list all possible inputs in increasing lexicographic order, a problem can be represented by the infinite list of bits that are the correct outputs to each input.

For example, if the problem is to determine whether a given binary input string corresponds to an even number, the representation might look like this:

`010101010101010101...`

Of course this all depends on the details of how one encodes inputs, but the point is that if you wanted to you could nail all this down precisely. More importantly for us we can represent the halting problem as an infinite *table* of bits. If the columns of the table are all programs (in lex order), and the rows of the table correspond to inputs (in lex order), then the table would have at entry a 1 if halts and a 0 otherwise.

here is 1 if halts and 0 otherwise. The table encodes the answers to the halting problem for all possible inputs.

Now we assume for contradiction sake that some program solves the halting problem, i.e. that every entry of the table is computable. Now we’ll construct the answers output by meta- by flipping each bit of the diagonal of the table. The point is that meta- corresponds to some *row* of the table, because there is some input string that is interpreted as the source code of meta-. Then we argue that the entry of the table for contradicts its definition, and we’re done!

So these are two of the most high-profile uses of the method of diagonalization. It’s a great tool for your proving repertoire.

Until next time!

I never liked this proof, because “meta-T” in the input is given itself without any input!, then how could we expect an algorithm (or whatever ) would tell us, if it will halt or not, if we are not giving any input to test it !!?, where is the surprise?

We can invent infinite trivial examples of it “ohh tell me if a number is bigger than five”, the computer ask “for which number?” so we say “ohh we don’t tell, try to guess! for any X! is it bigger than five?” WTF!! of course it can’t tell you before you input any finite input, “if we don’t tell the number then no Turing machine can solve it” it’s kind of a joke.

I don’t understand your confusion. The meta-T program is defined to have a single input (P), and it calls T(P,P) where T is the supposed halting-problem-solver. Since P has a single input, it makes sense to use meta-T in the place of P because both P and meta-T have a single input.

There is no type error here… all inputs are accounted for.

Hello, ok I was not clear, here I go again: When the text says “P”, I think it means P(x), with the x input included, so we are talking about P as P(x) a program with a fixed finite input, then it’s a concrete instance of P, but having different x would lead to a different consequences(about halting or not halting), then when going to T(program, input) it calls T(P(x), P(x)), again including the input!, finally you call metaT with metaT, also without input, metaT(metaT(?)), so if your answer is “the source includes the input” then it must include P and it must even include X, so again we are talking about a concrete instance and not in general, but if we are talking in general, i.e. we lack of the specification of the input, then I think therein lies the impossibility of T to guess what would end happening, and then there is no big surprise and the theorem is trivial.

Think of P as the source code of a program. That is, it’s just a finite string. So T takes as input two string, the first one is supposed to represent a program and the second one can be any string. So T knows both the program P and its input x (represented as strings), and it needs to tell if the function call P(x) will return or loop infinitely. So T is always given a “concrete instance.” Then meta-T is just a program we’re defining, we give it a particular string as input (that happens to be the source code of meta-T), and the contradiction comes when we ask whether meta-T will loop infinitely or not.

In your first proof, what if N is 0.000 = 0 or 0.111… = 1? It can’t be, but you didn’t say why 😉

A very good and subtle point.

Suppose to the contrary that all . Then consider the seven numbers corresponding to setting any nonempty subset of the first three decimal digits to be nonzero, with the remaining digits all zero. Note these are the unique binary representations of these numbers. The pigeonhole principle says that at least one of these numbers has to occur after row 3 in the table, say in row j. But then entry by definition, a contradiction.

I think the argument is analogous for the other case.

I do not understand this argument. In particular, a binary representation that has only finitely many 1s is not unique, as you can swap the infinite sequence of zeros to ones and flip the position before from 1 to 0. (if there was such a position. Indeed, a non-zero real number is rational if and only if it has exactly two expansions in every base ≥ 2). Moreover, using the pigeonhole principle here somehow destroys the beauty of the argument. It can be fixed much easier by using at least three digits and avoiding the confusion with non-unique representations of real numbers.

Ah, I was wrong here: there are rational numbers that only have one decimal representation (like 1/7). But numbers which have a finite decimal representation (i.e., eventually there are only zeros) have two different such representations — and if I understand this correctly, then you are talking about such numbers, aren’t you?

Suppose f(1) = 0.1 and f(2) = 0.0001, and for all n>2, b(n)(n) = 0. Then it so happens that N=0.01111111… ie 1/4+1/8+1/16+… which =1/2 which = 0.1 which /was/ in the list, so no contradiction. It’s easy to avoid this with decimal digits — map odd digits to 2 and even digits to 1, eg. You could do the same using pairs of binary digits — [00,01] -> 10, [11,10] -> 01.

See the other comment in this thread for why b (n)(n) cannot be zero for all n > 2.

I don’t think that fixes it — all of {.001, .010, .011, .100, .101, .101 and .111} have b[n][n]=0 if placed in position n>3 so there’s no lack of pigeon holes. If you choose to represent the numbers in your list with recurring 1’s instead of recurring 0’s, you just risk producing N=0.010000… and having 0.001111111… already in the list.

No the pigeon hole is to get one of those numbers past row three. I.e. they can’t all be before 3

Hello. I think that a little introduction to Lisp where code and data is the same should be a good way to prepare the reader.

For example: the following program computes the length of the input.

(defun myT (input)

(cond ( (atom input) 1)

( (null input) 0 )

( (endp input) 0 )

( t (+ (myT (car input)) (myT (cdr input))))))

(myT ‘(* 10 20 30 40))

We apply myt to itself with:

(myt ‘(defun myT (input)

(cond ( (atom input) 1)

( (null input) 0 )

( (endp input) 0 )

( t (+ (myT (car input)) (myT (cdr input))))))

)

The result is 36.

The strong concept is that code is data

In Lisp you can eval a list as a program with eval, also you could use the program meta to count the number of atoms in a list, so a program is a list that you can eval, and there is nothing special in applying meta to meta since it gives the number of atoms in meta as a list that is 15. You can see this in action here:

(defun meta(input)

(cond ((null input) 0)

((atom input) 1)

(t (+ (meta (car input))

(meta (cdr input))))))

A program to sum three numbers:

(defparameter

suma3 ‘(loop repeat 3 sum (progn (princ “input a number:”) (read)) into total

finally (format t “The sum is ~a~%” total)))

CL-USER> (eval suma3)

input a number:10

input a number:10

input a number:40

The sum is 60

NIL

CL-USER> (meta suma3)

15

You can try with (meta (meta (meta (meta meta))))

Reblogged this on Elmar Klausmeier's Weblog and commented:

Very clear presentation on the uncountability of the real numbers, and the halting problem.

What I don’t understand is how the infinite construction works. When you are saying all the bits are flipped, are there more bits horizontally or vertically.

Put differently, consider 1/2, 1/4, 3/4, 1/8, 3/8, … Then go to something I remember from Calculus – for any real number and epsilon, there exists N such that there is a number on the list closer to the real number than epsilon.

0.111111…. == 1, but what is the infinity of the infinite sum? Are a countable number of elements enough, or do you need 2**(aleph-zero)? Zeno’s paradox, do infinite series every really add up to the convergent sum?

If you are worried about real numbers, try rewriting the argument to prove the following (easier) theorem: the set of all 0-1 sequences is uncountable. This is the core of the proof for the real numbers, and then to improve that proof to prove the real numbers are uncountable, you just have to show that the set of “collisions” you can get like 0.111… = 1 is only countable. In other words, nothing in the proof relies on sums of infinite series or even calculus.

I think I understood the halting theorem proof but where exactly did you require diagonalization? In other words, where did we use the fact that Turing machines are countable?

The set of inputs to a Turing machine is the set of all finite strings (which is countable). A TM is also encoded as a finite string. A TM is determined by its output for each input, the definition of T provides a 0-1 value for each finite string (each input turing machine), and the definition of metaT flips that bit. But metaT is itself a Turing machine, so it shows up in the list of inputs. This is the diagonalization.

At the end you say: ” Then we argue that the entry of the table for (\textup{meta-}T, \textup{meta-}T) contradicts its definition ”

So the contradiction is that the entry cannot be 0 or 1, because running meta-T with it’s own representation as input forever flips between 0 and 1, right?