AML702 Applied Computational Methods
I
I
T Lecture 03
Dc Approximations, Errors and
E Their Analysis
L
H
I
Approximations and Errors
• Approximation is unavoidable in mathematical
I modeling the real world phenomena
I • Approximation leads to errors
T • Estimating the errors in computation is
Dc necessary for reliability of computed results
E
L
H
I
2
Accuracy and Precision
Accuracy : It refers to how closely the measured or computed value
matches with the true value.
I Precision: It refers to how closely the computed or measured
I values agree with each other in repeated computation or
measurement.
T
Consider fig. 4.1 from Chapra,
as shown here
Dc
E
(a) inaccurate and imprecise,
L
(b) accurate and imprecise,
H (c) inaccurate and precise,
I (d) accurate and precise.
Errors in Computations
Error: The deviation from an expected or true
value. True value may not be known always!
I
I E.g. Area of a rectangle A= LxB=12x10=120 cm2
T But by measurement LmxBm=11.9x10.1
The computed value Ac= 120.19 cm2
D Error , E=|A- Ac |=0.19. This error is known as
E Absolute Error. When it is compared with the actual
L value or true value, it is known as relative error,
H Er = |A- Ac |/A=0.19/120.
I This is expressed in percentage as
Er = |A- Ac |/A*100=0.19/120*100 = 0.15%.
4
Floating Points and Machine Accuracy
Floating Point Representation
I These numbers are represented by a sign bit s, an
exact integer exponent ε, an exact positive integer
I
mantissa, M. (B – base, E – bias)
T
D
E Machine Accuracy, εm
L It refers to the smallest floating point number which
H added to the floating point number 1.0 produces a
I floating point result different from 1.0.
A typical 32 bit computer with base 2, has εm = 3x10-8
5
Machine Representation
Representing Real numbers on digital computers
I Computers have finite amount of memory. Consider
representing π= 3.141592653589793…………….
I
Or the repeating fraction 1/3 = 0.333333333333333……
T
Any such number x ∈ℜ needs infinite digits for accurate
D representation.
E x = ±(1.d1d 2d 3d 4..............d1) × 2e
L Where di are binary digits having values 0 or 1 and e is
H an integer exponent. The mantissa can be expressed
I as d1 d 2 d 3 d 4
1.d1d 2d 3d 4.......... = 1 + + 2 + 3 + 4 .........
2 2 2 2
E.g. The binary number x is represented as:
6 x = −(1.11010.........) × 21 = −(1. 12 + 14 + 0 + 161 .........) × 2 = −3.62510
Floating Point Representation
Floating point representation requires finite digits
I Corresponding to a real number x its digital
counterpart will be fl(x) with some k bits
I
T fl ( x) = ±(1.d 1d 2d 3d 4..............d 1) × 2e
D Now storing this in memory will require k bits which is an
E approximation for storing x with infinite bits. It is
L interesting to know the accuracy of such a
H representation. Alternatively we wish to determine the
I relative error as: | fl ( x) − x |
| x|
The IEEE standard gives the confidence by estimating
1
this error to be η = 2 × 2−k where k is the number of
7
digits (bits) used in the representation
Floating Point and Machine Precision
The IEEE standard gives the confidence by estimating
this error to be η = 1 × 2−k where k is the number of
I 2
digits (bits) used in the representation.
I
This η is known as machine precision or rounding unit.
T
The floating point word in common double precison
D systems like MATLAB is by default, has 64 bits. Out of
E them, 52 bits are used for the mantissa (or fraction)
while the rest store the sign and the exponent.
L
Therefore the machine precision for MATLAB is
H
I η = 2−53 ≈ 2.2×10−16.
(here k=52)
Check eps = 2.2204 × 10−16 in MATLAB
8
Floating point ranges
The exponent range is -1022 to 1023.
I (11 bits including 1 bit for sign)
I • The largest possible number MATLAB can
T store has
• +1.111111…111 X 21023 = (2-2-52)X 21023
D
• This yields approximately 21024 = 1.7997 X 10308
E
L • The smallest possible number MATLAB can
H store with full precision has
I • +1.00000…00000 X 2-1022
• This yields 2-1022 = 2.2251 X 10-308
Real Vs Floating point numbers
Comparing floating point numbers to real numbers.
Property Real numbers Floating point
I numbers
I
T Range Infinite Finite
D . Precision Infinite Finite
E
Existence Real Subset of real
L numbers
H
I
10
Roundoff Errors
• Finite-precision causes roundoff errors in
I numerical computation
I • Roundoff errors accumulate slowly
T • Subtracting nearly equal numbers leads to
D severe loss of precision. A similar loss of
E precision occurs when two numbers
L separated by large differences in magnitude
H are added.
I • Roundoff errors are unavoidable, good
algorithms can minimize their effect
Truncation Errors
I
I
T
D
E
L
H
I
Total Numerical Error
• The total numerical error is the sum of the
truncation and roundoff errors.
I • The truncation error generally increases as the
I step size increases, while the roundoff error
T decreases as the step size increases - this leads
to a point of diminishing returns for step size.
D
E
L
H
I
