Introduction: Engineering problems and computational methods

20 cm x 30 cm rectangular plate
Initially at temperature f(x,y). Then, insulated at x=0, zero
temperature at y=0, and convective heat transfer at x=20 and
at y=30
Find temperature variation with time at the center

Experimental: Advantages: No equations, Arbitrary Shape

Issues: Size, Material, Time, f(x,y)

( )
cos( i x )sin ( j y )
T ( x, y, t ) = cij e
T 2T 2T i2 + 2j t
= +
Analytical : t x 2
y 2

i =1 j =1
Adv: Fast, Adaptable, Accurate; Issues: Ideal conditions
Introduction: Engineering problems and computational methods
Numerical Method: Reduce to algebraic equations
E.g.,

i,j+1
i-1,j i,j i+1,j
i,j-1

t + t
T Ti , j Ti , j
t

=
t t
Ti +1, j Ti , j Ti , j Ti 1, j
Ti +1, j 2Ti , j + Ti 1, j
2T x x
= =
x 2
x x 2
Introduction: Engineering problems and computational methods

Results in a set of linear (or nonlinear, if is a function of

temperature) equations

Adv: Arbitrary shape, initial and boundary conditions; Usually

fast.
Issues: Convergence, Accuracy, Efficiency
 Error Analysis: Round-off and Truncation errors
Backward and Forward error analysis
All computational results have errors: Computers have finite
storage length and computations cannot be carried out
infinitely (We do not consider Model error and Data error)

Round-off Error: Due to finite storage length. E.g., with 8

significant digits: 2 is stored as 1.4142136 (actually as
0.14142136x101), an error of about 3x106 percent.

Truncation Error: Due to finite steps in computation. E.g., exp(2)

is computed using the infinite series, truncated after a finite
number of terms. True value is 7.389056098930650. 1 term
1; 2 terms3, 35, 46.333, 5-7, 6-7.356,,15-7.3890561.
 Error Analysis: Round-off and Truncation errors
Backward and Forward error analysis
Round-off Error: Considering decimal storage, real numbers are
stored as mantissa (m) x 10 power (p) with 0.1 m <1. If mantissa
rounded-off at the tth digit, maximum error = 0.5x10t.
Maximum relative error = 5x10t called the round-off unit, u.

Definitions:
True Error (e) = True Value Approximate Value
Approximate Error () = Current Approximation Previous Approximation
(e.g., x=x/2+1/x, x=1,1.5, 1.4166667,1.4142157,1.4142136)
True Relative Error (er) = (True Value Approximate Value) / True Value
Approximate Relative Error (r) = (Current Approximation Previous
Approximation) / Current Approximation
 Error Propagation

y is a function of n independent variables, x1,x2,,xn, i.e. vector x

y=f(x)
Due to error in the variables (xi = xi ~xi ), there is error in y

y = f (~
~ x)
f n
Error: y = y y = f ( x ) f ( x ) = xi
~ ~
i =1 xi ~
xi

(Neglecting higher order terms in the Taylors series)

 Error Propagation
2h
g= 2
t
Height measurement accurate to a mm, time to 100th of sec.
Error in g, if h=100.000 m and t=4.52 s?
g 2 g 4h
= 2; = 3
h t t t
~
2 4h
g ~ 2 h ~ 3 t
t t
= 0.0978933h 4.33156t
Max value 0.0434 (Relative error is more meaningful, roughly
0.4%.)
 Error Propagation and Condition Number of a problem
Condition number: Ratio of the relative change in the function,
f(x), to the relative change in the variable for a small
perturbation x in the variable x.
f ( x + x ) f (x )
f (x ) xf (x )
CP = or, as x 0,
x f (x )
x

Well-Conditioned (<1) and Ill-conditioned (>1)

E.g.: Hyperbola xyy+1=0. Measure x, compute y=f(x)=1/(1x).
CP = |x/(1-x)| Ill-conditioned for x>0.5 (x=0.8, y=5; x=0.808,y=5.2083)
Similarly, y = 1 + x 1 for small x ;
CP = x close to but <1. (x=0.01, y=.004988; x=0.0101,y=.005037)
(
2 1+ x 1+ x )
 Error Analysis: Forward and Backward error analysis

Forward Error: requires the T.V.

(Recall that True Error (e) = True Value Approximate Value)
E.g.: 2 = 1.414213562373095. With 4 digit accuracy (u=5x104),
approx. value = 1.414. True error, e = 2.134x104. True relative
error = 0.015%.
Backward Error: determines the change necessary in the input
variable (or data) to explain the final error in the output (or
result) obtained through errors introduced at various stages of
calculation.
E.g.: 1.4142=1.999396. Error=6.04x104, Relative error = 0.03%.
y = 1 + x 1 with x=0.001616 gives y = 1.002 1 =
0.0009995, which is exact for x=0.002 (error=0.000384, 24%)
 Condition Number of an algorithm
Condition number: Exact input data, x; Corresponding
function value y; Algorithm A operating on x produces yA on a
machine with precision unit u. Input data, xA, which would
produce y, if exact computations are done. Condition number
of the algorithm x xA
x
CA =
u
Well-Conditioned (<1) and Ill-conditioned (>1)
y = 1 + x 1 for x =0.001616, we get xA=0.002.
0.001616 0.002
0.001616
CA = 4
= 475
5 10
x x=0.001616, y=0.0008076,
y = 1+ x 1 =
1+ x +1 xA=0.00161585, CA=0.18
 Error Analysis: Vector Norms
What if the computed result is a vector?

[A]{x} = {b}
Error vector
{ } = {x} {~x }

Or, relative error vector

x ~
x
{ r } =
x
Magnitude: In terms of the Norm of the vector
 Error Analysis: Vector Norms
Some properties of the vector norms are ( is a scalar):
x = 0 , only if x is a null vector, otherwise x > 0

x = x

x1 + x2 x1 + x2

Magnitude: The Lp norm of an n-dimensional vector, {x}, i.e.,

(x1,x2,,xn) T, is given by

x p
( p p p
= x1 + x2 + x3 + ... + xn )
p 1/ p
p 1
 Common Vector Norms
p=2, Euclidean norm: (length) x 2 = x12 + x22 + ... + xn2

n
p=1, (total distance ) x 1 = xi
i =1

p=, Maximum norm x

= max xi
i