SECTION
4: ROUNDOFF AND
TRUNCATION ERRORS
MAE 4020/5020 – Numerical Methods with MATLAB
2 Definitions of Error
K. Webb MAE 4020/5020
True Error
3
Absolute error – the difference between an
approximation and the true value
Relative error – the true error as a percentage of the
true value
Both definitions require knowledge of the true value!
If we had that, why would we be approximating?
K. Webb MAE 4020/5020
Approximating the Error
4
Since we don’t know the true value, we can only
approximate the error
Often, approximations are made iteratively
Approximate the error as the change in the
approximate value from one iteration to the next
K. Webb MAE 4020/5020
Relative Approximate Error
5
We don’t know the true value, so we can’t calculate
the true error – approximate the error
Relative approximate error – an approximation of
the error relative to the approximation itself
K. Webb MAE 4020/5020
Stopping Criterion
6
For iterative approximations, continue to iterate
until the relative approximate error magnitude is
less than a specified stopping criterion
For accuracy to at least significant figures set the
stopping criterion to
K. Webb MAE 4020/5020
7 Roundoff Error
K. Webb MAE 4020/5020
Roundoff Errors
8
Roundoff errors occur due to the way in which
computers represent numerical values
Computer representation of numerical values is
limited in terms of:
Magnitude – there are upper and lower bounds on the
magnitude of numbers that can be represented
Precision – not all numbers can be represented exactly
Certain types of mathematical manipulations are
more susceptible to roundoff error than others
K. Webb MAE 4020/5020
Number Systems – Decimal
9
We are accustomed to the decimal number system
A base‐10 number system
Ten digits: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9
Each digit represents an integer power of 10
7281.36
6 x 10‐2
3 x 10‐1
1 x 10‐0
8 x 101
2 x 102
7 x 103
K. Webb MAE 4020/5020
Binary Number System
10
Computers represent numbers in binary format
A base‐2 number system
Two digits: 0, 1
Easy to store binary values in computer hardware – an on or off switch – a high or low
voltage
One digit is a bit – eight bits is a byte
Each bit represents an integer power of 2
10110101
1 x 20 = 1
0 x 21 = 0
1 x 22 = 4
0 x 23 = 0 = (181)10
1 x 24 = 16
1 x 25 = 32
0 x 26 = 0
1 x 27 = 128
K. Webb MAE 4020/5020
IEEE Double‐Precision Format
11
By default, MATLAB uses double‐precision floating
point to store numeric values ‐ double
64‐bit binary word
Signed
Sign bit
Exponent – (e) Mantissa – (f)
11 bits 52 bits
K. Webb MAE 4020/5020
IEEE Double‐Precision Format
12
Mantissa
Only the fractional portion of the mantissa stored
Bit to the left of the binary point assumed to be 1
Normalized numbers
Really a 53‐bit value
Exponent
11‐bit signed integer value: ‐1022 … 1023
Two special cases:
e = 0x000 (i.e. all zeros): zero if f = 0, subnormal #’s if f 0
e = 0x7FF (i.e. all ones): ∞ if f = 0, NaN if f 0
K. Webb MAE 4020/5020
Normalized numbers
13
Leading zeros are removed
Most significant digit (must be a 1 in binary) moved to
the left of the binary point
53rd bit of the mantissa (always 1) needn’t be stored
Maximum mantissa
1.111……11
Not stored
K. Webb MAE 4020/5020
Subnormal Numbers
14
If , always, then the smallest number that
could be represented is: 2 .
If we allow for , then the most significant bit
is somewhere to the right of the binary point
Leading zeros – not normalized … subnormal
Allows for smaller numbers, filling in the hole around
zero
Subnormal numbers represented by setting the
exponent to zero
Smallest subnormal number:
2 2
K. Webb MAE 4020/5020
Doubles – Range
15
Maximum value:
1 2 2 .
Minimum normal value:
2 .
Minimum subnormal value:
2 2
Precision – machine epsilon
.
K. Webb MAE 4020/5020
Roundoff Error – Mathematical Operations
16
Certain types of mathematical operations are more
susceptible to roundoff errors:
Subtractive cancellation – subtracting of two nearly‐
equal numbers results in a loss of significant digits
Large computations – even if the roundoff error from a
single operation is small, the cumulative error from
many operations may be significant
Adding large and small numbers – as in an infinite
series
Inner products – (i.e. dot product) very common
operation – solution of linear systems of equations
K. Webb MAE 4020/5020
17 Truncation Error
K. Webb MAE 4020/5020
Truncation Errors
18
Errors that result from the use of an approximation
in place of an exact mathematical procedure
E.g. numerical integration, or the approximation
derivatives with finite‐difference approximations
To understand how truncation errors arise, and to
gain an understanding of their magnitudes, we’ll
make use of the Taylor Series
K. Webb MAE 4020/5020
Taylor Series
19
Taylor’s Theorem – any smooth (i.e., continuously
differentiable) function can be approximated as a
polynomial
Taylor Series
This infinite series is an equality
An exact representation of any smooth function as a
polynomial
An infinite‐order polynomial – impractical
K. Webb MAE 4020/5020
Taylor Series Approximation
20
Can approximate a function as a polynomial by truncating
the Taylor series after a finite number of terms
⋯
2! !
where is the step size
Chapra
K. Webb MAE 4020/5020
Taylor Series Truncation Error
21
Can account for error by lumping the and
higher‐order terms into a single term,
⋯
2! !
is the error associated with truncating after terms
1 !
is some (unknown) value of between and
K. Webb MAE 4020/5020
Derivative Mean‐Value Theorem
22
If and are continuous on , then
there is a point on this interval, , where is the
slope of the line joining and
Chapra
K. Webb MAE 4020/5020
Truncation Error – Dependence on Step Size
23
1 !
We don’t know , so we don’t know
We do know it’s proportional to , where is the
step size
Error is on the order of
If (first‐order approx.), halving the step size
will quarter the error
K. Webb MAE 4020/5020
24 Truncation Errors in Practice
• Discretizing equations
• Finite‐difference approximations
K. Webb MAE 4020/5020
Discretization of Equations
25
As engineers, many of the mathematical
expressions we are interested in are differential
equations
We know how to evaluate derivatives analytically
Need an approximation for the derivative operation in
order to solve numerically
Discretization – conversion of a continuous
function, e.g. differentiation, to a discrete
approximation for numerical evaluation
K. Webb MAE 4020/5020
Finite Difference Approximations
26
Recall the definition of a derivative
lim
→
Remove the limit to approximate this numerically
This is the forward difference approximation
Uses value at and forward one step at to
approximate the derivative at
K. Webb MAE 4020/5020
Discretizing Equations – Example
27
A free‐falling object (bungee jumper example from
the text) can be modeled as
where is velocity, is mass, is gravitational acceleration, and is a
lumped drag coefficient
This is a non‐linear ordinary differential equation
(ODE), which can be solved analytically to yield
tanh ∙
K. Webb MAE 4020/5020
Discretizing Equations – Example
28
To solve numerically instead, approximate the
derivative operation with a finite difference
Solving for and using to denote the time step
yields
We’ve transformed the differential equation to a
difference equation
An algebraic equation
Can be solved iteratively – using a loop
K. Webb MAE 4020/5020
Discretizing Equations – Example
29
The term in the square brackets is the original diff. eq., i.e. it is
The difference equation is a first‐order Taylor series
approximation
Where we know that the error is on the order of the step size
squared
Taylor series provides a relation between the step size and
the accuracy of the numerical solution to the diff. eqn.
K. Webb MAE 4020/5020
Finite Difference Methods
30
The preceding example showed
One method – forward difference – for numerically
approximating a derivative
Transformation of a differential equation to a difference
equation
How Taylor series can provide an understanding of the
error associated with an approximation
Now we’ll take a closer look at three finite
difference methods and how Taylor series can help
us understand the error associated with each
K. Webb MAE 4020/5020
Forward Difference
31
Can also derive the forward So, the error term is
difference approximation from
the Taylor Series
The forward difference,
Solving for
including error, is
We’ve already seen that Error of the forward
difference approximation is
on the order of the step
size
K. Webb MAE 4020/5020
Forward Difference
32
Value of the
function,
at and
forward one
step at
used to
approximate
the derivative
Chapra
at
K. Webb MAE 4020/5020
Backward Difference
33
Backward difference uses value of Again the error is on the order of
at and one step backward at the step size
to approximate the derivative at
The backward difference
This can also be developed by
expression, including error,
expanding the Taylor series backward becomes
Then solving for
Error of the backward
difference approximation is
on the order of the step size
K. Webb MAE 4020/5020
Backward Difference
34
Now use the
value of
at and
backward one
step at to
approximate
the derivative
at
Again, error is
Chapra
K. Webb MAE 4020/5020
Central Difference
35
Central difference uses value of Now, the remainder term is
one step backward at and ones
step ahead at to approximate
the derivative at The central difference expression,
including error, becomes
2
This can also be developed by 2
subtracting the backward Taylor
series from the forward series Error of the central difference
approximation is on the order of
Second‐order derivative terms
the step size squared
cancel, leaving
Central difference method is
2 more accurate than forward or
backward
Uses more information
K. Webb MAE 4020/5020
Central Difference
36
Now use the
value of
backward one
step at and
forward one step
at to
approximate the
derivative at
Reduced error:
Chapra
K. Webb MAE 4020/5020
37 Total Numerical Error
K. Webb MAE 4020/5020
Total Numerical Error
38
Total numerical error is the sum of roundoff and
truncation error
Roundoff error is largely out of your control, and, with
double precision arithmetic, it is not typically an issue
Truncation error can be a significant problem, but can
be reduced by decreasing step size
Reducing step size reduces truncation error, but
may also result in subtractive cancellation, thereby
increasing roundoff error
Choose step size to minimize total error
Or, more typically, to reduce truncation error to an
acceptable level
K. Webb MAE 4020/5020
Total Numerical Error
39
Reducing step size reduces truncation error, but may
also result in subtractive cancellation, thereby
increasing roundoff error
Could choose step size
to minimize total error
But, more typically,
reduce step size just
enough to reduce
truncation error to an
acceptable level
Chapra
K. Webb MAE 4020/5020
Central Difference Error Analysis
40
First derivative of a function in terms of the central
difference approximation is
The last term on the right is the truncation error
There is also roundoff error associated with each value
were represents a rounded value, and is the
corresponding roundoff error
K. Webb MAE 4020/5020
Central Difference Error Analysis
41
Substituting the expressions for the rounded values into
the expression for the true derivative yields
Giving a total error of
Truncation error
Roundoff error
Truncation error increases with step size
Roundoff error decreases with step size
K. Webb MAE 4020/5020