Next: Computer Arithmetic.
Up : Main Previous:Computer Representation of Numbers
Numerical Errors:
Numerical errors arise during computations due to round-off errors and truncation errors.
Round-off Errors:
Round-off error occurs because computers use fixed number of bits and hence fixed number of binary digits to represent numbers. In a numerical computation round-off errors are introduced at every stage of computation. Hence though an individual round-off error due to a given number at a given numerical step may be small but the cumulative effect can be significant.
When the number of bits required for representing a number are less then the number is usually rounded to fit the available number of bits. This is done either by chopping or by symmetric rounding.
Chopping: Rounding a number by chopping amounts to
dropping the extra digits. Here the given number is truncated.
Suppose that we are using a computer with a fixed word length of
four digits. Then the truncated representation of the number will be
. The digits
will be dropped. Now to evaluate the error
due to chopping let us consider the normalized representation of
the given number
i.e.
chopping error in representing
.
So in general if a number is the true value of a given
number and
is the normalized form of the
rounded (chopped) number
and
is the
normalized form of the chopping error then
Since
, the chopping error
Symmetric Round-off Error :
In the symmetric round-off method the last retained significant
digit is rounded up by 1 if the first discarded digit is greater
or equal to 5.In other words, if in
is such that
then the last digit in
is raised by 1
before chopping
. For example let
be two given numbers to be rounded to
five digit numbers. The normalized form x and y are
and
. On rounding
these numbers to five digits we get
and
respectively. Now w.r.t
here
In either case error
.
Truncation Errors:
Often an approximation is used in place of an exact mathematical
procedure. For instance consider the Taylor series expansion of
say i.e.
Practically we cannot use all of the infinite number of terms in the series for computing the sine of angle x. We usually terminate the process after a certain number of terms. The error that results due to such a termination or truncation is called as 'truncation error'.
Usually in evaluating logarithms, exponentials, trigonometric
functions, hyperbolic functions etc. an infinite series of the
form
is replaced by a
finite series
. Thus a truncation
error of
is introduced
in the computation.
For example let us consider evaluation of exponential function
using first three terms at
Truncation Error
Absolute and Relative Errors:
Absolute Error: Suppose that and
denote the true
and approximate values of a datum then the error incurred on
approximating
by
is given by
and the absolute error i.e. magnitude of the error is
given by
Relative Error: Relative Error or normalized
error in representing a true datum
by an
approximate value
is defined by
and
Sometimes is defined by
If
and
then
Machine Epsilon: Let us assume that we have a decimal computer system.
We know that we would encounter round-off error when a number is represented in floating-point form. The relative round-off error due to chopping is defined by
Here we know that
i.e. maximum relative round-off error due to chopping is given by
. We know that the value of 'd' i.e the length of
mantissa is machine dependent. Hence the maximum relative
round-off error due to chopping is also known as machine epsilon
. Similarly , maximum relative
round-off error due to symmetric rounding is given by
Machine-Epsilon
for symmetric rounding is given
by,
It is important to note that the machine epsilon represents upper bound for the round-off error due to floating point representation.
For a computer system with binary representation the machine epsilon due to chopping and symmetric rounding are given by
respectively.
Eg: Assume that our binary machine has 24-bit mantissa. Then
. Say that our system
can represent a q decimal digit mantissa.
Then,
i.e
that our machine can store numbers with seven
significant decimal digits.
Next: Computer Arithmetic. Up : Main Previous:Computer Representation of Numbers