Chapter 14

Exponential Function

The function et .

The exponential function is denoted mathematically by et and in Matlab by

exp(t). This function is the solution to the world’s simplest, and perhaps most
important, differential equation,
ẏ = ky
This equation is the basis for any mathematical model describing the time evolution
of a quantity with a rate of production that is proportional to the quantity itself.
Such models include populations, investments, feedback, and radioactivity. We
are using t for the independent variable, y for the dependent variable, k for the
proportionality constant, and
dy
ẏ =
dt
for the rate of growth, or derivative, with respect to t. We are looking for a function
that is proportional to its own derivative.
Let’s start by examining the function
y = 2t
We know what 2t means if t is an integer, 2t is the t-th power of 2.
2−1 = 1/2, 20 = 1, 21 = 1, 22 = 4, ...
We also know what 2t means if t = p/q is a rational number, the ratio of two
integers, 2p/q is the q-th root of the p-th power of 2.
√
21/2 = 2 = 1.4142...,

Copyright °c 2009 Cleve Moler

Matlab° R
is a registered trademark of The MathWorks, Inc.TM
August 8, 2009

1
2 Chapter 14. Exponential Function

3.5

2.5

1.5

0.5

0
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Figure 14.1. The blue curve is the graph of y = 2t . The green curve is
the graph of the rate of growth, ẏ = dy/dt.

0.9

0.85

0.8

0.75

0.7

0.65

0.6

0.55

0.5
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

Figure 14.2. The ratio, ẏ/y.

√
25/3 =
3
25 = 3.1748...,
√
2355/113 =
113
2355 = 8.8250...
In principal, for floating point arithmetic, this is all we need to know. All floating
point numbers are ratios of two integers. We do not have to be concerned yet about
the definition of 2t for irrational t. If Matlab can compute powers and roots, we
can plot the graph of 2t , the blue curve in figure 14.1
What is the derivative of 2t ? Maybe you have never considered this question,
or don’t remember the answer. (Be careful, it is not t2t−1 .) We can plot the
graph of the approximate derivative, using a step size of something like 0.0001. The
 3

following code produces figure 14.1, the graphs of both y = 2t and its approximate
derivative, ẏ.
t = 0:.01:2;
h = .00001;
y = 2.^t;
ydot = (2.^(t+h) - 2.^t)/h;
plot(t,[y; ydot])
The graph of the derivative has the same shape as the graph of the original
function. Let’s look at their ratio, ẏ(t)/y(t).
plot(t,ydot./y)
We see that the ratio of the derivative to the function, shown in figure 14.2, has a
constant value, ẏ/y = 0.6932..., that does not depend upon t .
Now, if you are following along with a live Matlab, repeat the preceeding
calculations with y = 3t instead of y = 2t . You should find that the ratio is again
independent of t. This time ẏ/y = 1.0986.... Better yet, experiment with expgui.
If we take any value a and look at y = at , we find that, numerically at least,
the ratio ẏ/y is constant. In other words, ẏ is proportional to y. If a = 2, the
proportionality constant is less than one. If a = 3, the proportionality constant is
greater than one. Can we find an a so that ẏ/y is actually equal to one? If so, we
have found a function that is equal to its own derivative.
The approximate derivative of the function y(t) = at is
at+h − at
ẏ(t) =
h
This can be factored and written
ah − 1 t
ẏ(t) = a
h
So the ratio of the derivative to the function is
ẏ(t) ah − 1
=
y(t) h
The ratio depends upon h, but not upon t. If we want the ratio to be equal to 1,
we need to find a so that
ah − 1
=1
h
Solving this equation for a, we find
a = (1 + h)1/h
The approximate derivative becomes more accurate as h goes to zero, so we are
interested in the value of
(1 + h)1/h
4 Chapter 14. Exponential Function

as h approaches zero. This involves taking numbers very close to 1 and raising
them to very large powers. The surprising fact is that this limiting process defines
a number that turns out to be one of the most important quantities in mathematics
e = lim (1 + h)1/h
h→0
Here is the beginning and end of a table of values generated by repeatedly cutting
h in half.
format long
format compact
h = 1;
while h > 2*eps
h = h/2;
e = (1 + h)^(1/h);
disp([h e])
end

0.500000000000000 2.250000000000000
0.250000000000000 2.441406250000000
0.125000000000000 2.565784513950348
0.062500000000000 2.637928497366600
0.031250000000000 2.676990129378183
0.015625000000000 2.697344952565099
... ...
0.000000000000014 2.718281828459026
0.000000000000007 2.718281828459036
0.000000000000004 2.718281828459040
0.000000000000002 2.718281828459043
0.000000000000001 2.718281828459044
0.000000000000000 2.718281828459045
The last line of output involves a value of h that is not zero, but is so small that it
prints as a string of zeros. We are actually computing
51
(1 + 2−51 )2
which is
1
(1 + )2251799813685248
2251799813685248
The result gives us the numerical value of e correct to 16 significant decimal digits.
It’s easy to remember the repeating pattern of the first 10 significant digits.
e = 2.718281828...
Let’s derive a more useful representation of the exponential function. Start
by putting t back in the picture.
et = ( lim (1 + h)1/h )t
h→0

= lim (1 + h)t/h
h→0
 5

Here is the Binomial Theorem.

n(n − 1) n−2 2 n(n − 1)(n − 2) n−3 3
(a + b)n = an + nan−1 b + a b + a b + ...
2! 3!
If n is an integer, this terminates after n+1 terms with bn . But if n is not an integer,
the expansion is an infinite series. Apply the binonimial theorem with a = 1, b = h
and n = t/h.
(t/h)(t/h − 1) 2 (t/h)(t/h − 1)(t/h − 2) 3
(1 + h)t/h = 1 + (t/h)h + h + h + ...
2! 3!
t(t − h) t(t − h)(t − 2h)
= 1+t+ + + ...
2! 3!
Now let h go to zero. We get the power series for the exponential function.
t2 t3 tn
et = 1 + t + + + ... + + ...
2! 3! n!
This series is a rigorous mathematical definition that applies to any t, positive or
negative, rational or irrational, real or complex. The n + 1-st term is tn /n!. As
n increases, the tn in the numerator is eventually overwhelmed by the n! in the
denominator, so the terms go to zero fast enough that the infinite series converges.
It is almost possible to use this power series for actual computation of et . Here
is an experimental Matlab program.
function s = expex(t)
% EXPEX Experimental version of EXP(T)
s = 1;
term = 1;
n = 0;
r = 0;
while r ~= s
r = s;
n = n + 1;
term = (t/n)*term;
s = s + term;
end
Notice that there are no powers or factorials. Each term is obtained from the
previous one using the fact that
tn t tn−1
=
n! n (n − 1)!
The potentially infinite loop is terminated when r == s, that is when the floating
point values of two successive partial sums are equal.
There are “only” two things wrong with this program – its speed and its
accuracy. The terms in the series increase as long as |t/n| ≥ 1, then decrease after
n reaches the point where |t/n| < 1. So if |t| is not too large, say |t| < 2, everything
is OK; only a few terms are required and the sum is computed accurately. But
6 Chapter 14. Exponential Function

larger values of t require more terms and the program requires more time. This is
not a very serious defect if t is real and positive. The series converges so rapidly
that the extra time is hardly noticeable.
However, if t is real and negative the computed result may be inaccurate. The
terms alternate in sign and cancel each other in the sum to produce a small value
for et . Take, for example, t = −20. The true value of e−20 is roughly 2 · 10−9 .
Unfortunately, the largest terms in the series are (−20)19 /19! and (−20)20 /20!,
which are opposite in sign and both of size 4 · 107 . There is 16 orders of magnitude
difference between the size of the largest terms and the size of the final sum. With
only 16 digits of accuracy, we lose everything. The computed value obtained from
expex(-20) is completely wrong.
For real, negative t it is possible to get an accurate result from the power
series by using the fact that
1
et =
e−t
For complex t, there is no such easy fix for the accuracy difficulties of the power
series.
In contrast to its more famous cousin, π, the actual numerical value of e is
not very important. It’s the exponential function
et
that’s important. In fact, Matlab doesn’t have the value of e built in. Nevertheless,
we can use
e = expex(1)
to compute an approximate value for e. Only seventeen terms are required to get
floating point accuracy.
e = 2.718281828459045
If we revise the code to compute the numerator and denominator separately using
integer arithmetic, we find a rational approximation to e with 17! in the denomina-
tor.
966858672404690
e=
355687428096000
After computing e, you could then use e^t, but exp(t) is preferable.
The logarithm is the inverse function of the exponential. If
y = et
then
loge (y) = t
The function loge (y) is known as the natural logarithm and is often denoted by ln y.
More generally, if
y = at
 7

then
loga (y) = t
The function log10 (y) is known as the common logarithm. Matlab uses log(y),
log10(y), and log2(y) for loge (y), log10 (y), and log2 (y).
The term exponential growth is often used informally to describe any kind of
rapid growth. Mathematically, the term refers to any time evolution, y(t), where
the rate of growth is proportional to the quantity itself.
ẏ = ky
The solution to this equation is determined for all t by specifying the value of y at
one particular t, usually t = 0.
y(0) = y0
Then
y(t) = y0 ekt
Suppose, at time t = 0, we have a million E. coli bacteria in a test tube under
ideal laboratory conditions. Twenty minutes later each bacterium has fissioned to
produce another one. So at t = 20, the population is two million. Every 20 minutes
the population doubles. At t = 40, it’s four million. At t = 60, it’s eight million.
And so on. The population, measured in millions of cells, y(t), is
y(t) = 2t/20
Let k = ln 2/20 = .0347. Then, with t measured in minutes and the population y(t)
measured in millions, we have
ẏ = ky, y(0) = 1
Consequently
y(t) = ekt
This is exponential growth, but it cannot go on forever. Eventually, the growth rate
is affected by the size of the container. Initially at least, the size of the population
is modelled by the exponential function.
Suppose, at time t = 0, you invest $1000 in a savings account that pays 5%
interest, compounded yearly. A year later, at t = 1, the bank adds 5% of $1000
to your account, giving you y(1) = 1050. Another year later you get 5% of 1050,
which is 52.50, giving y(2) = 1102.50. If y(0) = 1000 is your initial investment,
r = 0.05 is the yearly interest rate, t is measured in years, and h is the step size for
the compound interest calculation, we have
y(t + h) = y(t) + rhy(t)
What if the interest is compounded monthly instead of yearly? At the end of the
each month, you get .05/12 times your current balance added to your account. The
8 Chapter 14. Exponential Function

same equation applies, but now with h = 1/12 instead of h = 1. Rewrite the
equation as
y(t + h) − y(t)
= ry(t)
h
and let h tend to zero. We get

ẏ(t) = ry(t)

This defines interest compounded continuously. The evolution of your investment

is described by

y(t) = y(0)ert

Here is a Matlab program that tabulates the growth of $1000 invested at 5% over
a 20 year period , with interest compounded yearly, monthly, and continuously.

format bank
r = 0.05;
y0 = 1000;
for t = 0:20
y1 = (1+r)^t*y0;
y2 = (1+r/12)^(12*t)*y0;
y3 = exp(r*t)*y0;
disp([t y1 y2 y3])
end

The first few and last few lines of output are

t yearly monthly continuous

0 1000.00 1000.00 1000.00
1 1050.00 1051.16 1051.27
2 1102.50 1104.94 1105.17
3 1157.63 1161.47 1161.83
4 1215.51 1220.90 1221.40
5 1276.28 1283.36 1284.03
.. ....... ....... .......
16 2182.87 2221.85 2225.54
17 2292.02 2335.52 2339.65
18 2406.62 2455.01 2459.60
19 2526.95 2580.61 2585.71
20 2653.30 2712.64 2718.28

Compound interest actually qualifies as exponential growth, although with modest

interest rates, most people would not use that term.
Let’s borrow money to buy a car. We’ll take out a $20,000 car loan at 10%
per year interest, make monthly payments, and plan to pay off the loan in 3 years.
What is our monthly payment, p? Each monthly transaction adds interest to our
 9

current balance and subtracts the monthly payment.

y(t + h) = y(t) + rhy(t) − p

= (1 + rh)y(t) − p

Apply this repeatedly for two, three, then n months.

y(t + 2h) = (1 + rh)y(t + h) − p

= (1 + rh)2 y(t) − ((1 + rh) + 1)p
y(t + 3h) = (1 + rh)3 y(t) − ((1 + rh)2 + (1 + rh) + 1)p
y(t + nh) = (1 + rh)n y(0) − ((1 + rh)n−1 + ... + (1 + rh) + 1)p
= (1 + rh)n y(0) − ((1 + rh)n − 1)/(1 + rh − 1)p

Solve for p

p = (1 + rh)n /((1 + rh)n − 1)rhy0

Use Matlab to evaluate this for our car loan.

y0 = 20000
r = .10
h = 1/12
n = 36
p = (1+r*h)^n/((1+r*h)^n-1)*r*h*y0

We find the monthly payment would be

p = 645.34

If we didn’t have to pay interest on the loan and just made 36 monthly payments,
they would be

y0/n

= 555.56

It’s hard to think about continuous compounding for a loan because we would have
to figure out how to make infinitely many infinitely small payments.

Exercises

14.1 e cubed. The value of e3 is close to 20. How close? What is the percentage
error?

14.2 expgui.
(a) With expgui, the graph of y = at , the blue line, always intercepts the y-axis at
y = 1. Where does the graph of dy/dx, the green line, intercept the y-axis?
10 Chapter 14. Exponential Function

(b) What happens if you replace plot by semilogy in expgui?

14.3 Computing e.
(a) If we try to compute (1+h)1/h for small values of h that are inverse powers of 10,
it doesn’t work very well. Since inverse powers of 10 cannot be represented exactly
as binary floating point numbers, the portion of h that effectively gets added to 1
is different than the value involved in the computation of 1/h. That’s why we used
inverse powers of 2 in the computation shown in the text. Try this:
format long
format compact
h = 1;
while h > 1.e-15
h = h/10;
e = (1 + h)^(1/h);
disp([h e])
end
How close do you get to computing the correct value of e?
(b) Now try this instead:
format long
format compact
h = 1;
while h > 1.e-15
h = h/10;
e = (1 + h)^(1/(1+h-1));
disp([h e])
end
How well does this work? Why?

14.4 expex. Modify expex by inserting

disp([term s])
as the last statement inside the while loop. Change the output you see at the
command line.
format compact
format long
Explain what you see when you try expex(t) for various real values of t.
expex(.001)
expex(-.001)
expex(.1)
expex(-.1)
expex(1)
expex(-1)
 11

Try some imaginary values of t.

expex(.1i)
expex(i)
expex(i*pi/3)
expex(i*pi)
expex(2*i*pi)
Increase the width of the output window, change the output format and try larger
values of t.
format long e
expex(10)
expex(-10)
expex(10*pi*i)

14.5 Instrument expex. Investigate both the cost and the accuracy of expex. Modify
expex so that it returns both the sum s and the number of terms required n. Assess
the relative error by comparing the result from expex(t) with the result from the
built-in function exp(t).
relerr = abs((exp(t) - expex(t))/exp(t))

Make a table showing that the number of terms required increases and the relative
error detioriates for large t, particularly negative t.

14.6 Rational expex. Make a version of expex that does not do any divisions, but
returns two large integers that form the numerator and denominator of a rational
approximation to e. Reproduce the result
966858672404690
e=
355687428096000