This page does not represent the most current semester of this course; it is present merely as an archive.

MCS discusses induction in chapter 5. This text is intended to supplement, not replace, that text.

1 The idea, programmer’s version

Suppose I have a loop that iterates over some list. How do I know what the results will be after the loop is over? The most common reasoning goes like this:

  1. I know what’s true before the loop starts
  2. I know it stays true each pass through the loop
    1. because it starts true
    2. and one step of the loop does not make it false
  3. So I know it must be true at the end

This outline may make intuitive sense, but we need to formalize it to make it into a proof strategy.

  1. We need some starting place. We call that the base case. It can be as complicated as we need, though often it will be quite simple.

  2. We need some loop-like way of getting from one truth to the next. We call that the inductive step. It generally works as follows

    1. we assume it starts true; this assumption is called the inductive hypothesis
    2. we use that assumption to prove that it will still be true one step later
  3. We conclude it must always be true. This step is not a proof, it’s an axiom called the principle of induction.

2 The idea, formal logic version

The principle of induction is an axiom or proof rule that looks like:

\begin{aligned} &P(0)\\ &P(n) \rightarrow P(n+1)\\ \therefore\;&\forall x \in \mathbb N \;.\; P(x) \end{aligned}

Thus, to use it, we first prove P(0); then prove P(n) \rightarrow P(n+1) by assuming P(n) and proving P(n+1); then state that by the principle of induction, \forall x \in \mathbb N \;.\; P(x).

In this approach first prove P(0) is called the base case; then prove P(n) \rightarrow P(n+1) is called the inductive step; and by assuming P(n) is called the inductive hypothesis.

What if your base case is more involved than P(0)? The formal answer is define a different P

Let’s try proving the following by induction:

All Fibonacci numbers are positive

A fairly high-level proof might look like

We proceed by induction.

Base case
The first two Fibonacci numbers are both 1, a positive number.
Inductive step
Assume that the kth and (k+1)th Fibonacci number are both positive. Then the (k+2)th Fibonacci number must also be positive because it is the sum of two positive numbers.

By the principle of induction, it follows that all Fibonacci numbers are positive.

All well and good. But how do we make this fit the formal variation? We define a special P(n) to mean the (n+1)th and (n+2)th Fibonacci number are both positive. With this special predicate, our base case becomes just P(0) (i.e., the 1st and 2nd are positive) and our inductive step becomes a proof that P(n) \rightarrow P(n+1).

Although we can convert arbitrary induction into a formal variation, we won’t in this class or anywhere else outside of formal logic.

3 Examples

Consider this code:

Java Python
double babylonian(double x) {
    double y = 1;
    for(int i=0; i<20; i+=1) {
        y = (y + x/y)/2;
    }
    return y;
}
def babylonian(x):
    y = 1
    for i in range(20):
      y = (y + x/y)/2
    return y

How could we verify that the end result is between 1 and x?

Initially, y is between 1 and x (in particular, it is 1). Each pass through the loop y is updated to be the average of two values: y and x/y. We know one of those (y) is between 1 and x; but is the other?

Consider x/y. We proceed by cases.

Case 1: x < 1
In this case, y is between 1 and x means xy ≤ 1. Because xy, x/y ≤ 1. Because y ≤ 1, xx/y. Thus, xx/y ≤ 1, meaning it is between 1 and x.
Case 2: x ≥ 1
In this case, y is between 1 and x means 1 ≤ yx. Because yx, 1 ≤ x/y. Because 1 ≤ y, x/yx. Thus, 1 ≤ x/yx, meaning it is between 1 and x.

Because x/y is between 1 and x in both cases, it is between them in general.

Thus, the loop replaces y with the average of two numbers, both between 1 and x, so it keeps y between 1 and x.

Because we start between 1 and x and that does not change, y ends up (and thus the function returns a value that is) between 1 and x.

Incidentally, the functions actually return \sqrt{x}, though proving that is beyond the scope of this course.