You are on page 1of 11

1

9.1 Sources of error


• Error classification
– Modeling error
– User error
– Bugs
– Discretization error
– Truncation or rounding error
– Manipulation error
• Error tests
– Ill conditioning and Hilbert matrix Hij=1/(I+j-1)
– Reality checks

R.T. Haftka EML5526 Finite Element Analysis University of Florida


2

9.2 Ill-conditioning
• Sensitivity to data

 k1  k1   u1   P  k1u1  k1u 2  P
 k     or
 1 k1  k 2  u 2   0   k1u1  ( k1  k 2 )u 2  o

R.T. Haftka EML5526 Finite Element Analysis University of Florida


3

Example: Equations with Hilbert Matrix


M tl b
Matlab
function [[h,rhs]=hilbert(n)
, ] ( ) • Order 3
%calculates the hilbert matrix of >> [h,rhs]=hilbert(3)
%order n and the right hand side
h=
%for a solution of all ones
1.0000 0.5000 0.3333
for i=1:n
0.5000 0.3333 0.2500
for j=1:n
j 1:n
0.3333 0.2500 0.2000
h(i,j)=1/(i+j-1);
rhs =
end
1 8333 1
1.8333 1.0833
0833 0
0.7833
7833
end
rhs=sum(h,1);
end
R.T. Haftka EML5526 Finite Element Analysis University of Florida
4

Ill conditioning
h,rhs]=hilbert(10); sol=(h\rhs')‘
Columns 1 through 9
1.0000 1.0000 1.0000 1.0000 0.9999 1.0003 0.9995 1.0005 0.9997
[h,rhs]=hilbert(12); sol=(h\rhs')'
Warning: Matrix is close to singular or badly scaled
scaled.
Results may be inaccurate. RCOND = 2.458252e-017.
Columns 1 through 9
1.0000 1.0000 0.9999 1.0009 0.9931 1.0300 0.9168 1.1500 0.8246
[h,rhs]=hilbert(13); sol=(h\rhs')'
Warning: Matrix is close to singular or badly scaled.
Results may be inaccurate. RCOND = 1.158544e-019.
Columns 1 through 9
1 0000
1.0000 0
0.9997
9997 1
1.0117
0117 0
0.8071
8071 2
2.7244
7244 -8
8.3249
3249 33
33.4625
4625 -74
74.1604
1604 117
117.9244
9244

R.T. Haftka EML5526 Finite Element Analysis University of Florida


5

How to induce ill-conditioning


. •Large variations in stiffness

R.T. Haftka EML5526 Finite Element Analysis University of Florida


6

Another example


k 1  c 2  k  cs   u1   P  c  cos 
    

 
Where
 k  cs k 1  s 2   v1   0  i 
s  sin

• The slopes of the lines of the two equations


ddv1 1  c 2 dv1
d cs
 and 
du1 cs du1 1  s 2
• Problems for large alpha
alpha.

R.T. Haftka EML5526 Finite Element Analysis University of Florida


7

Application to slanted boundaries


.

R.T. Haftka EML5526 Finite Element Analysis University of Florida


8
9.3 The condition number
 max
• Definition C (K ) 
 min

• Interpretation
I t t ti as vibration
ib ti frequencies
f i
• Scale the matrix to unit diagonal

Ks   I D  0 where Ks   SKS & Sii  1


Kii
• Use transformation D=S-1Ds

K    K 11 K 22 .... K nn Ds   0


• Stiff region acts as a large “mass” reducing lowest
frequency without changing much highest one.

R.T. Haftka EML5526 Finite Element Analysis University of Florida


9

Interpretation
• Eigenvalue expansion (Vi, orthonormal eigenvectors)
n n
1
K  i 1
 iV iV i T
and K 1
 
i 1 i
V iV i T

• Consider perturbation in loading


K (d  d '))  r  Vk
Kd '   kVk d '  K 1 Vk  Vk / k
• If
r  V j then d  Vj / j
• Then j d'
 
d k
• Worst case when j=n, k=1: error amplified by C(K).
However, very unlikely that r looks like high eigenvector
• Haftka, R.T., “Stiffness-Matrix Condition Number and Shape Sensitivity Errors,” AIAA
Journal, Vol. 28, No. 7, pp. 1322-1324, 1990
R.T. Haftka EML5526 Finite Element Analysis University of Florida
10

Hilbert matrix
[h,rhs]=hilbert(10); error=(h\rhs')'-ones(1,10)
Columns 1 throughg 7
-1.5172e-009 1.3060e-007 -2.7734e-006 2.5151e-005 -1.1972e-004
3.2854e-004 -5.3822e-004
>> cond(h)
1.6025e+013

>> [h,rhs]=hilbert(11);
[h h ] hilb t(11) error=(h\rhs')'-ones(1,11)
(h\ h ')' (1 11)
Columns 1 through 7
-1.0099e-008 1.0692e-006 -2.7992e-005 3.1534e-004 -1.8906e-003
6.6835e-003 -1.4621e-002
>> cond(h)
5.2260e+014

R.T. Haftka EML5526 Finite Element Analysis University of Florida


11
Loss of accuracy
•Number
Number of digits lost
dacc  d  dloss dloss  log10 C  K 
•Estimate of condition number
2 m1
 hmax 
C K   b 
2m/n
N els
 hmin 
where
b  a positive number
hmax , hmax  maximum and minimum node spacing in the mesh
N els  number of elements in FE model
2m  differential equation order (e.g. 2 for plane stress, 4 for beam)
n  dimensionality of the FE mesh (e.g., 1 for beam, 2 for plane stress)

R.T. Haftka EML5526 Finite Element Analysis University of Florida

You might also like