C l e v e s C o r n e r

Floating points
IEEE Standard unifies arithmetic model
by Cleve Moler

f you look carefully at the definition of fundamental

where f is the fraction or mantissa and e is the exponent. The

arithmetic operations like addition and multiplication,

fraction must satisfy

you soon encounter the mathematical abstraction known

as the real numbers. But actual computation with real numbers

is not very practical because it involves limits and infinities.
Instead, MATLAB and most other technical computing

environments use floating-point arithmetic, which involves a

finite set of numbers with finite precision. This leads to
phenomena like roundoff error, underflow, and overflow. Most
of the time, MATLAB can be effectively used without worrying
about these details, but every once in a while, it pays to know
something about the properties and limitations of floatingpoint numbers.

What is the
output?

0f<1
and must be representable in binary using at most 52 bits. In
52

other words, 2 f must be an integer in the interval

Twenty years ago, the situation was far more complicated

e = 1 c

The exponent must be an integer in the interval

-1022 e 1023
The finiteness of f is a limitation on precision. The
dont meet these limitations must be approximated by ones
that do.

number system. Some were binary; some were decimal. There

was even a Russian computer that used trinary arithmetic.
Among the binary computers, some used 2 as the base; others
used 8 or 16. And everybody had a different precision.
In 1985, the IEEE Standards Board and the American

Double precision floating-numbers can be stored in a

National Standards Institute adopted the ANSI/IEEE Standard

64-bit word, with 52 bits for f , 11 bits for e, and 1 bit for the

754-1985 for Binary Floating-Point Arithmetic. This was the

sign of the number. The sign of e is accommodated by storing

culmination of almost a decade of work by a 92-person

e+1023, which is between 1 and 211-2. The two extreme values

working group of mathematicians, computer scientists and

for the exponent field, 0 and 211-1, are reserved for exceptional

engineers from universities, computer manufacturers, and

floating-point numbers, which we will describe later.

microprocessor companies.
All computers designed in the last 15 or so years use IEEE
floating-point arithmetic. This doesnt mean that they all get
exactly the same results, because there is some flexibility within
the standard. But it does mean that we now have a machineindependent model of how floating-point arithmetic behaves.

The picture above shows the distribution of the positive

MATLAB uses the IEEE double precision format. There is also a

numbers in a toy floating-point system with only three bits

single precision format which saves space but isnt much faster

each for f and e. Between 2e and 2e+1 the numbers are equally

on modern machines. And, there is an extended precision

spaced with an increment of 2e-3. As e increases, the spacing

format, which is optional and therefore is one of the reasons

increases. The spacing of the numbers between 1 and 2 in our

for lack of uniformity among different machines.

toy system is 2-3, or 8. In the full IEEE system, this spacing is

2-52. MATLAB calls this quantity eps, which is short for

Most floating point numbers are normalized. This means

they can be expressed as

machine epsilon.

x = (1 + f ) 2

eps = 2^(52)

b = a 1
c = b + b + b

0 252f < 253

finiteness of e is a limitation on range. Any numbers that

than it is today. Each computer had its own floating-point

a = 4/3

C l e v e s C o r n e r ( c o n t i n u e d )

Before the IEEE standard, different machines had different

of the vector

values of eps.
0:0.1:1

The approximate decimal value of eps is 2.2204 10-16.

Either eps/2 or eps can be called the roundoff level. The

is exactly equal to 1, but if you form this vector yourself by

maximum relative error incurred when the result of a single

repeated additions of 0.1, you will miss hitting the final 1

arithmetic operation is rounded to the nearest floating-point

exactly.

number is eps/2. The maximum relative spacing between

Another example is provided by the MATLAB code

numbers is eps. In either case, you can say that the roundoff

segment in the margin on the previous page. With exact

level is about 16 decimal digits.

computation, e would be 0. But in floating-point, the

A very important example occurs with the simple MATLAB

statement

computed e is not 0. It turns out that the only roundoff error

occurs in the first statement. The value stored in a cannot be
4

exactly 3, except on that Russian trinary computer. The

t = 0.1

value stored in b is close to 3, but its last bit is 0. The value

The value stored in t is not exactly 0.1 because expressing

1

stored in c is not exactly equal to 1 because the additions are

the decimal fraction 10 in binary requires an infinite series.

done without any error. So the value stored in e is not 0. In

In fact,

fact, e is equal to eps. Before the IEEE standard, this code was

10 = 12 + 12 + 02 + 02 + 12 +12 + 02
4

10 + 0 211 + 1 212 +

After the first term, the sequence of coefficients 1, 0, 0, 1 is

repeated infinitely often. The floating-point number nearest

used as a quick way to estimate the roundoff level on various

computers.
The roundoff level eps is sometimes called floating-point
zero, but thats a misnomer. There are many floating-point

0.1 is obtained by rounding this series to 53 terms, including

numbers much smaller than eps. The smallest positive

rounding the last four coefficients to binary 1010. Grouping

normalized floating-point number has f = 0 and e = -1022.

the resulting terms together four at a time expresses the

The largest floating-point number has f a little less than 1 and

approximation as a base 16, or hexadecimal, series. So the

e = 1023. MATLAB calls these numbers realmin and realmax

resulting value of t is actually

Together with eps, they characterize the standard system.

t = (1 + 9 16 + 9 162 + 9 163 + + 9 1612 + 10 1613) 2-4

The MATLAB command

format hex

causes t to be printed as

Name

Binary

Decimal

eps

2^(-52)

2.2204e-16

realmin

2^(-1022)

2.2251e-308

realmax

(2-eps)*2^1023

1.7977e+308

When any computation tries to produce a value larger

than realmax, it is said to overflow. The result is an

3fb999999999999a

The first three characters, 3fb, give the hexadecimal

representation of the biased exponent, e+1023, when e is -4.
The other 13 characters are the hex representation of the
fraction f.
So, the value stored in t is very close to, but not exactly

exceptional floating-point value called Inf, or infinity. It is

represented by taking f = 0 and e = 1024 and satisfies relations
like 1/Inf = 0 and Inf+Inf = Inf.
When any computation tries to produce a value smaller
than realmin, it is said to underflow. This involves one of the
optional, and controversial, aspects of the IEEE standard.

equal to, 0.1. The distinction is occasionally important. For

Many, but not all, machines allow exceptional denormal or

example, the quantity

subnormal floating-point numbers in the interval between

realmin and eps*realmin. The smallest positive subnormal

0.3/0.1

number is about 0.494e-323. Any results smaller than this are

is not exactly equal to 3 because the actual numerator is a

set to zero. On machines without subnormals, any result less

little less than 0.3 and the actual denominator is a little greater

than realmin is set to zero. The subnormal numbers fill in

than 0.1.
Ten steps of length t are not precisely the same as one step
of length 1. MATLAB is careful to arrange that the last element

the gap you can see in our toy system between zero and the
smallest positive number. They do provide an elegant model
for handling underflow, but their practical importance for

MATLAB style computation is very rare.

When any computation tries to produce a value that is
undefined even in the real number system, the result is an

MATLAB notices the tiny value of U(2,2) and prints a

message warning that the matrix is close to singular. It then
computes the ratio of two roundoff errors

exceptional value known as Not-a-Number, or NaN. Examples

include 0/0 and Inf-Inf.

x(2)

MATLAB uses the floating-point system to handle integers.

Mathematically, the numbers 3 and 3.0 are the same, but

c(2)/U(2,2)

16

This value is substituted back into the first equation to give

many programming languages would use different

representations for the two. MATLAB does not distinguish
between them. We like to use the term flint to describe a

x(1)

(11 - x(2))/10

-0.5

floating-point number whose value is an integer. Floating-

The singular equations are consistent. There are an infinite

point operations on flints do not introduce any roundoff

number of solutions. The details of the roundoff error

error, as long as the results are not too large. Addition,

determine which particular solution happens to be computed.

subtraction and multiplication of flints produce the exact flint

result, if it is not larger than 2^53. Division and square root
involving flints also produce a flint when the result is an
integer. For example, sqrt(363/3) produces 11, with no
roundoff error.
As an example of how roundoff error effects matrix
computations, consider the two-by-two set of linear equations
10x1 +

x2 = 11

3x1 + 0.3x2 = 3.3

The obvious solution is x1 = 1, x2 = 1. But the MATLAB

statements
A = [10
b = [11
x = A\b

1; 3
3.3]'

0.3]

produce
x =
-0.5000
16.0000

Why? Well, the equations are singular. The second

equation is just 0.3 times the first. But the floating-point

Our final example plots a seventh degree polynomial.

x = 0.988:.0001:1.012;
y = x.^7-7*x.^6+21*x.^5-35*x.^4+35*x.^3-
21*x.^2+7*x-1;
plot(x,y)

representation of the matrix A is not exactly singular because

A(2,2) is not exactly 0.3.

Gaussian elimination transforms the equations to the

upper triangular system
U*x = c

where

But the resulting plot doesnt look anything like a polynomial.

It isnt smooth. You are seeing roundoff error in action. The
y-axis scale factor is tiny, 10-14. The tiny values of y are being
computed by taking sums and differences of numbers as large
as 35 1.0124. There is severe subtractive cancellation. The
example was contrived by using the Symbolic Toolbox to

U(2,2)= 0.3 - 3*(0.1)

= -5.5551e-17

Cleve Moler is chairman

to be near x = 1. If the values of y are computed instead by

and co-founder of
The MathWorks.

and
c(2)

expand (x - 1)7 and carefully choosing the range for the x-axis

y = (x-1).^7;
= 3.3 - 33*(0.1)
= -4.4409e-16

then a smooth (but very flat) plot results.

His e-mail address is

[email protected].