Iteration, Fixed points

Paul Seidel

18.01 Lecture Notes, Fall 2011

Take a function f (x).

Definition. A fixed point is a point x such that

f (x) = x .

Graphically, these are exactly those points where the graph of f , whose equation
is y = f (x), crosses the diagonal, whose equation is y = x. You can often solve
for them exactly:
Example. To determine the fixed points of the function f (x) = x3 , we solve

x3 = x
⇒ x3 − x = 0
⇒ x(x2 − 1) = 0

so the fixed points are x = 0 and x = −1, x = +1.

Example. The function f (x) = x2 + x + 1 has no fixed points.
x 1
√
Example. The function f (x) = 2 + x has a fixed point at x = 2.
Example. The function f (x) = cos(x) has a fixed point, as one can see by
looking at the graph. However, there’s no simple way to write down what that
fixed point is, since the equation cos(x) = x has no solution is terms of known
numbers (roots, etc.)

Why are we interested in them, and why are they called fixed? This comes from
the problem of iterating the function f , which means feeding the output of that
function back into itself as an input. Here’s the idea summarized in a picture:

/ f

1
What this actually means is that we start with some x0 (the starting value) and
get x1 , x2 , . . . by applying f once, twice, and so on. In formulae:

x1 = f (x0 ),
x2 = f (x1 ) = f (f (x0 )),
x3 = f (x2 ) = f (f (x1 )) = f (f (f (x0 ))),
...

We say that x1 , x2 , . . . are obtained by iterating f . This yields a sequence of

numbers xn defined by starting with x0 and applying

xn = f (xn−1 ) (1)

over and over. It’s actually the same as having a loop in a computer program
containing the command x := f(x) (but that would not be a proper math-
ematical way to say it). The behaviour of the xn can be quite complicated in
general, but there’s a simple special case:
Fact. If x0 is a fixed point of f , then the iteration consists of x0 = x1 = x2 =
···.

However, in applications, how often would it happen that the starting point
is exactly a fixed point? The question is whether fixed points can help us to
understand the more general behaviour of iteration. As we’ll see, calculus does
help with that. Before doing the theory, let’s investigate some examples.
√
Example. Take f (x) = x2 + x1 . We know that 2 is a fixed point, but let’s
instead start with x0 = 1 and iterate:
1 3
x1 = +1= ,
2 2
3 2 17
x2 = + = ,
4 3 12
17 12
x3 = + ≈ 1.41421569,
24 17
...
√
We observe
√ experimentally that the xn converge to 2. They never get exactly
equal to 2, but they get closer and closer. The same is true for other starting
values, there’s nothing special about 1!
Example. We are modelling the population of rabbits (in thousands) year-by-
year on an island. Each year, the population of rabbits tends to increase over
the previous year at a constant rate, until growth is limited by available re-
sources. Ultimately, if the population becomes too big, it can even decrease due
to exhaustion of resources. We model this by iterating a function

f (x) = ax − bx2

2
for some numbers a, b > 0. Actually, a > 0 is the rate of growth without any con-
straints, while the −bx2 term accounts for the limits imposed by resources. Let’s
suppose for now that a = 2.5, b = 1.5. We start with a very small population of
rabbits, x0 = 0.01 (just like they did in Australia). Then,

x0 = 0.01
x1 = 0.02485
x2 = 0.06119871625
x3 = 0.147378866319028
x4 = 0.335866370441347
x5 = 0.670456597913201
x6 = 1.001873420254987
x7 = 0.999058025317329
x8 = 1.000469656366881
x9 = 0.999764840950905
x10 = 1.000117496574880
...

As n gets larger, it seems that xn converges to 1. Again, it turns out that the
limiting value 1 is a fixed point.

Given these examples, one might think that the problem is rather easy after all:

Idea. (This turns out to be wrong!) Start with an arbitrary x0 and iterate f .
As n grows larger, xn always converges to some fixed point of f .

As we’ve already warned, that can’t be true. After all, there are functions with
no fixed points at all. And even if we have a fixed point and we start close to
it, convergence is not the only possible behaviour.
Example. Consider the same rabbit problem as before, but now with faster
growth: a = 3.25, b = 1.5. We look for fixed points
13 3
x = f (x) = x − x2
4 2

3
and find that the only one with x > 0 is at x = 3/2. Let’s try it out:

x0 = 0.01
x1 = 0.03235
x2 = 0.103567716250000
x3 = 0.320505670038639
x4 = 0.887557600835202
x5 = 1.702924460513897
x6 = 1.184576919345345
x7 = 1.745041271103810
x8 = 1.103630574303982
x9 = 1.759798699680136
x10 = 1.074008578866596
x11 = 1.760286240097869
x12 = 1.073018809701235
x13 = 1.760257082570031
...

Problem! Instead of arriving at the fixed point 3/2, the values seem to jump
between ≈ 1.76 and ≈ 1.07. Maybe that’s because we didn’t start close enough
to the fixed point?
x0 = 1.49
x1 = 1.51235
x2 = 1.48433371625
x3 = 1.519214706017698
x4 = 1.475427810086859
x5 = 1.529809548615736
x6 = 1.461405150447320
x7 = 1.546009218322859
x8 = 1.439313204840405
x9 = 1.570334163289379
x10 = 1.404661954099851
x11 = 1.605538542881098
x12 = 1.351369245348429
x13 = 1.652651791472020
x14 = 1.274231406500550
x15 = 1.705753555158233
x16 = 1.179306267861831
...
Still not working! In fact, what we again observe is an oscillation in the popu-
lation. This is exciting because we did not build it into our equation (1), it is a
phenomenon that arises unexpectedly from that equation.

To understand the difference between the two last-mentioned examples, we need

calculus. Specifically, we will use linear approximation for (1) near a fixed point.

4
Let f (x) = x and suppose that xn−1 is close to x. Then

xn = f (xn−1 ) ≈ f (x) + f 0 (x)(xn−1 − x) = x + f 0 (x)(xn−1 − x)

(note that here x is the point we are approximating around, unlike the usual
situation) and hence

|xn − x| ≈ |f 0 (x)| · |xn−1 − x|.

Conclusion. (This is now correct!) If |f 0 (x)| < 1 and xn−1 is close to x, then
xn will be even closer to x (by a factor of |f 0 (x)| approximately) and hence as
n gets larger, we will get convergence to x.

We can now introduce the official names. A fixed point x is called attracting if
starting with some number sufficiently close to x and iterating it always leads
to convergence to x. Our conclusion is that:

If a fixed point has |f 0 (x)| < 1, it is attracting.

On the other hand, a fixed point that pushes away nearby values is called
repelling. One can show by a similar analysis that:

If a fixed point has |f 0 (x)| > 1, it is repelling.

The borderline situation f 0 (x) = ±1 needs to be looked at on a case-by-case

basis (may be attracting, repelling, or show mixed behaviour). Now we can
review our previous examples in this light:
√ √
Examples. f (x) = x2 + x1 has a fixed point at x = 2, and f 0 ( 2) = 0. So this
is very very attracting.

f (x) = 52 x − 23 x2 has a fixed point at x = 1, and f 0 (1) = 5/2 − 1 = −1/2, so it

is attracting (the negative sign means that we reach the fixed point by bouncing
around from left to right). By the way, there’s another fixed point at x = 0, but
there f 0 (0) = 5/2, so that one is repelling.
0
f (x) = 13 3 2
4 x − 2 x has a fixed point at x = 3/2, and f (3/2) =
13
4 − 9
2 = −5/4,
so it is repelling.

As we can see from the last two examples, changing a parameter (in this case a)
can have the effect of changing the situation from attracting to repelling. Some
of the phenomena you hear referred to as “tipping points” are of this kind.