Model Condensation

CEE 541. Structural Dynamics

Department of Civil and Environmental Engineering
Duke University
Henri P. Gavin
Fall, 2016

Problem Statement
Given a symmetric stiffness matrix K and a symmetric mass matrix M of a “full”
model with N coordinates and given a subset of n primary coordinates, p (1 ≤ pi ≤ N ,
i = 1, · · · , n); find condensed stiffness and mass matrices Kc and Mc such that the dynamic
characteristics of K and M are retained at the n condensed coordinates, p.

1 Static Condensation
Consider the partitioning of the full stiffness matrix equation
" # " # " #
fp Kpp Kps rp
= (1)
fs Ksp Kss N ×N
rs

where Kpp is n × n. Assume the external forces on the secondary coordinates are all zero
and that the displacements at the secondary coordinates are of no interest.
−1
fs = Ksp rp + Kss rs = 0 → rs = −Kss Ksp rp (2)
−1
fp = Kpp rp + Kps rs = Kpp rp − Ksp Kss Ksp rp (3)
−1
Kc = [Kpp − Kps Kss Ksp ] (4)

Since fs is assumed to be zero, rs may be computed from rp directly.

" # " #
rp In×n h i
= −1 rp (5)
rs −Kss Ksp N ×n
r = Tc rp (6)

Using the coordinate transformation matrix Tc ,

Kc = TcT K Tc (7)
" #" #
h
−1
i Kpp Kps In×n
= In×n −Kps Kss −1 (8)
Ksp Kpp −Kss Ksp

You can confirm that equation (8) simplifies to equation (4).

2 CEE 541. Structural Dynamics – Duke University – Fall 2016 – H.P. Gavin

2 Guyan Condensation
In the Guyan Condensation method, the static condensation coordinate transformation
is simply applied to the mass matrix,

Mc = TcT M Tc (9)

Stiffness and mass matrices condensed in this way preserve potential and kinetic energy.
1 T
V = r Kr (10)
2
1 T T 1
= rp Tc K Tc rp = rpT Kc rp (11)
2 2

1 T
T = ṙ M ṙ (12)
2
1 T T 1
= ṙp Tc M Tc ṙp = ṙpT Mc ṙp (13)
2 2
The P.E. and the K.E. of the condensed model is the same as that of the full model, but the
natural frequencies and mode shapes of the condensed model do not match those of the full
model. The dynamics characteristics of Kc and Mc are closer to those of K and M at the
lower frequencies.

CC BY-NC-ND H.P. Gavin October 18, 2018

Model Condensation 3

3 Dynamic Condensation
Consider the partitioning of the undamped and unforced second order equations of
motion " #" # " #" # " #
Mpp Mps r̈p Kpp Kps rp 0
+ = (14)
Msp Mss r̈s Ksp Kss rs 0
Assuming a solution
rp = r̄p cos ωt (15)
rs = r̄s cos ωt (16)
" #" # " #
Kpp − ω 2 Mpp Kps − ω 2 Mps r̄p 0
= (17)
Ksp − ω 2 Msp Kss − ω 2 Mss r̄s 0
Applying the static condensation equation directly to this partitioned eigenvalue problem,
h i h i h ih i−1 h i
Kc − ω 2 M c = Kpp − ω 2 Mpp − Kps − ω 2 Mps Kss − ω 2 Mss Ksp − ω 2 Msp
(18)
By defining a new coordinate transformation matrix
 
In×n
Tc =  h i−1 h i  , (19)
Kss − ω 2 Mss Ksp − ω 2 Msp
N ×n
h i
TcT K − ω2M Tc = TcT K Tc − ω 2
TcT M Tc (20)
So, the condensed stiffness and mass matrices are,
Kc = TcT K Tc (21)
Mc = TcT M Tc (22)
These condensed stiffness and mass matrices match the dynamics of the full model matrices
K and M exactly at a frequency ω.
Note:

• The frequency ω can be any frequency, e.g., the first mode frequency, a higher mode
frequency, or any intermediate frequency.
• If ω is set to zero, the dynamic condensation method reduces to Guyan condensation.
• A condensed stiffness matrix computed using dynamic condensation preserves the band-
edness of the full system stiffness matrix.
• This method may be applied recursively in order to compute modal frequencies.
– at iteration i, select a frequency ω (i) ;
– compute Tc using ω (i) ;
– compute Kc and Mc using Tc , K and M ;
– solve the eigenvalue problem for the smaller-dimensioned Kc and Mc in order to
calculate the modal frequency for the next iteration ω (i+1) ;
– increment the iteration counter, i = i + 1, and go to step 2.

CC BY-NC-ND H.P. Gavin October 18, 2018

4 CEE 541. Structural Dynamics – Duke University – Fall 2016 – H.P. Gavin

4 Modal Condensation
Given a matrix of n of the modal vectors from the full model (n < N )
 
| | |
R̄ =  r̄1 r̄2 · · · r̄n  (23)


| | | N ×n

and the n corresponding natural frequencies

ω12
 

ω22 
Ω2 = 
 
..  (24)

 . 

ωn2
find symmetric n × n matrices Kc and Mc that have the same modal vectors and frequencies.
Note that modal vectors in R̄ and the modal frequencies in 2 may be any subset of the modal
vectors and frequencies, not only the first n. Assume that the modal vectors are mass
normalized;
T
R̄N ×N MN ×N R̄N ×N = IN ×N (25)
T
R̄N ×N KN ×N R̄N ×N = Ω2N ×N (26)
Define a square modal matrix R̄p as the rows of the full modal matrix R̄ at n the primary
coordinates and n selected modes (not necessarily the first n modes). Columns of the matrix
R̄s are the modal vectors at the n selected modes
q at the remaining (secondary) coordinates.
(The matrix R̄p may be scaled by a factor N/n in order to match the condensed modal
vectors.) Now the problem is to find Kc and Mc such that
R̄pT Mc R̄p = In×n (27)
R̄pT Kc R̄p = Ω2n×n (28)

Because the matrix R̄p may be ill-conditioned, using the regularized pseudo-inverse helps
with numerical conditioning of Kc and Mc ;
h i−1
R̄p+ = R̄p R̄pT + β diag(R̄p R̄pT ) R̄p (29)

where the regularization factor β can be small (≈ 10−4 ). Using this pseudo-inverse, the
condensed mass and stiffness matrices can be computed to meet the desired goal of matching
the dynamics at the primary coordinates and the selected modes.
Mc = R̄p+ In×n R̄p+T (N/n) (30)
Kc = R̄p+ Ω2n×n R̄p+T (N/n) (31)
Responses in the n selected modes of the N − n secondary coordinates may be assembled
from the primary coordinate responses,
rs (t) = R̄s R̄p+ rp (t) (32)

CC BY-NC-ND H.P. Gavin October 18, 2018

Model Condensation 5

5 Modal Assurance Criteria

In order to assess the performance of model condensation methods, the modal vectors
of the full model and condensed model can be compared. Orthogonality of the full-model
modal vectors with the condensed-model modal vectors indicates that the condensed model
captures the dynamics of the full model.
Defining R̄p as the rows of the full modal matrix R̄ at n the primary coordinates and
the first n modes. and R̃p as the eigenvectors of Mc and Kc , the modal assurance criteria is

MAC = |R̄pT R̃| (33)

Ideally, the MAC is close to identity.

References
[1] Guyan RJ. Reduction of stiffness and mass matrices. AIAA J 1965;3:380

[2] Paz, M. Dynamic condensation. AIAA J 1984;22(5):724-727.

http://dx.doi.org/10.2514/3.48498

[3] Friswell, M.I., Garvey, S.D., and Penny, J.E.T., Model Reduction using Dynamic and
Iterated IRS Techniques, Journal of Sound and Vibration (1995) 186(2),311-323.

CC BY-NC-ND H.P. Gavin October 18, 2018

 6 CEE 541. Structural Dynamics – Duke University – Fall 2016 – H.P. Gavin

6 ModelCondensation.m
1 % ModelCondensation .m
2 % t e s t m a t r i x c o n d e n s a t i o n methods f o r dynamic model r e d u c t i o n
3 % RJ Guyan , R e d u c t i o n o f S t i f f n e s s and Mass M a t r i c e s , AIAA J , v3 , n2 , 1965 , p380
4 % M Paz , Dynamic c o n d e n s a t i o n . AIAA J , v22 , n5 , 1984 , pp724 −727
5
6 % March 4 2010 , J u l y 2 2015
7
8 newSystem = 1
9 i f newSystem % − mass and s t i f f n e s s m a t r i c e s c o r r e s p o n d t o a s e r i e s o f s p r i n g s
10
11 N = 100; % t o t a l number o f c o o r d i n a t e s
12
13 % d i a g o n a l mass m a t r i x
14 M = eye( N ); diag (1+0.2* rand(1 , N ));
15 % t r i −d i a g o n a l s t i f f n e s s m a t r i x
16 ki = N ˆ2 * ones (1 , N ); % (1+0.3∗ randn ( 1 ,N) ) ;
17 K = diag ([ ki (1: N -1)+ ki (2: N ) , ki ( N )]) - diag ( ki (2: N ) ,1) - diag ( ki (2: N ) , -1);
18 p = [ 5:10:100 ]; % primary c o o r d i n a t e s
19 s = zeros (1 , N ); s ( p ) = 1; s = find (˜ s ); % secondary coordinates
20 n = length ( p ); % number o f primary c o o r d i n a t e s
21
22 [X , Omega2 ] = eigs ( K , M , n , 0 ); % solve the f u l l e i g e n v a l u e prob
23
24 [ Omega2 , idx ] = sort ( diag ( Omega2 )); X = X (: , idx ); % s o r t t o i n c r e a s i n g nat ’ l f r q
25
26 end % newSystem
27
28 % Guyan Condensation and Dynamic Condensation − Guyan 1965 and Paz 1984 −−−−−
29
30 % p i c k t h e s m a l l e s t e i g e n −v a l u e ( n a t u r a l f r e u e n c y s q u a r e d )
31 w2 = min( Omega2 ) % dynamic c o n d e n s a t i o n ( Paz )
32 %w2 = 0 . 0 ; % Guyan condensation
33
34 D = K - w2 * M ;
35
36 T = zeros (N , n );
37
38 T (p ,1: n ) = eye( n ); % coordinate transformation
39 T (s ,1: n ) = -D (s , s ) \ D (s , p );
40
41 Mc_d = T ’ * M * T ;
42
43 Kc_d = T ’ * K * T ;
44
45 [ Xd , w2_d ] = eigs ( Kc_d , Mc_d , n , 0 );
46
47 [ w2_d , idx ] = sort ( diag ( w2_d )); Xd = Xd (: , idx ); % s o r t t o i n c r e a s i n g nat ’ l f r q
48
49 ms = diag ( diag ( Xd ’* Mc_d * Xd )); % modal mass
50 Xd = Xd / sqrt ( ms ); % mass n o r m a l i z e d
51
52 % Modal Condensation −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
53
54 % ”n” s e l e c t e d modes t o match f o r modal c o n d e n s a t i o n
55
56 %m = [ 1 2 3 6 7 9 12 15 16 17 ] ;
57 m = [1: n ];
58
59 Xp = X (p , m ) * sqrt ( N / n ); % ”m” mode s h a p e s a t t h e primary c o o r d i n a t e s
60
61 beta = 0e -4; % regularization factor
62
63 Xp_pi = ( Xp * Xp ’ + beta * diag ( diag ( Xp * Xp ’)) ) \ Xp ; % pseudo−i n v e r s e
64
65 Mc_m = Xp_pi * eye( length ( m )) * Xp_pi ’ * N / n ;
66
67 Kc_m = Xp_pi * diag ( Omega2 ( m )) * Xp_pi ’ * N / n ;

CC BY-NC-ND H.P. Gavin October 18, 2018

 Model Condensation 7

68
69 [ Xm , w2_m ] = eigs ( Kc_m , Mc_m , n , 0 );
70
71 [ w2_m , idx ] = sort ( diag ( w2_m )); Xd = Xd (: , idx ); % s o r t t o i n c r e a s i n g nat ’ l f r q
72
73 ms = diag ( diag ( Xm ’* Mc_m * Xm )); % modal mass
74 Xm = Xm / sqrt ( ms ); % mass n o r m a l i z e d
75
76 % I t e r a t e d Improved Reduced System ( IIRS ) − F r i s e w e l l 1998 −−−−−−−−−−−−−−−−
77
78 T = zeros (N , n );
79 To = -K (s , s ) \ K (s , p ); % s t a t i c condensation coordinate transformation
80 Ti = To ;
81 T (p ,1: n ) = eye( n ); % coordinate transformation
82 f o r iter = 1:7
83 T (s ,1: n ) = Ti ;
84 Mc_r = T ’* M * T ;
85 Kc_r = T ’* K * T ;
86 Ti = To + K (s , s ) \ [ M (s , p ) , M (s , s ) ] * T * ( Mc_r \ Kc_r );
87 end
88
89 [ Xr , w2_r ] = eigs ( Kc_r , Mc_r , n , 0 );
90
91 [ w2_r , idx ] = sort ( diag ( w2_r )); Xr = Xr (: , idx ); % s o r t t o i n c r e a s i n g nat ’ l f r q
92
93 ms = diag ( diag ( Xr ’* Mc_r * Xr )); % modal mass
94 Xr = Xr / sqrt ( ms ); % mass n o r m a l i z e d
95
96 %

97
98 % D i s p l a y R e s u l t s −−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−−
99
100 epsPlots = 0; i f epsPlots , formatPlot (1); e l s e formatPlot (0); end
101
102 f p r i n t f ( ’ Natural Frequencies \ n ’)
103 fprintf (’ \ t \ tDynamic \ t \ tModal \ t \ tIIRS \ n ’)
104 f p r i n t f ( ’ Full \ t \ tCondensation \ tCondensation \ tCondensation \ n ’)
105 format
106 f o r i =1:10
107 f p r i n t f ( ’ %8.6 f \ t %8.6 f \ t %8.6 f \ t %8.6 f \ n ’ , ...
108 sqrt ( Omega2 ( i ))/2/ pi , ...
109 sqrt ( w2_d ( i ))/2/ pi , ...
110 sqrt ( w2_m ( i ))/2/ pi , ...
111 sqrt ( w2_r ( i ))/2/ pi )
112 end
113
114 % match s i g n s o f mode s h a p e s and s c a l e mode v e c t o r s t o f u l l model v a l u e s
115 Xd = Xd * diag ( sign ( X (N ,1: n ) .* Xd (n ,1: n ) ) ) * sqrt ( N / n );
116 Xm = Xm * diag ( sign ( X (N ,1: n ) .* Xm (n ,1: n ) ) ) * sqrt ( N / n );
117 Xr = Xr * diag ( sign ( X (N ,1: n ) .* Xr (n ,1: n ) ) ) * sqrt ( N / n );
118
119 % f i r s t ”n” mode s h a p e s a t t h e primary c o o r d i n a t e s
120 format bank
121 a c t u a l _ m o d e _ s h a p e s = Xp
122 d y n a m i c _ m o d e _ s h a p e s = Xd
123 m o d a l _ c o n d e n s a t i o n _ m o d e _ s h a p e s = Xm
124 i i r s _ c o n d e n s a t i o n _ m o d e _ s h a p e s = Xr
125
126 % e r r o r o f condensed mode s h a p e s
127 err_d = ( Xd - Xp )./ Xp ;
128 err_m = ( Xm - Xp )./ Xp ;
129 err_r = ( Xr - Xp )./ Xp ;
130
131 % modal a s s u r a n c e c r i t e r i a ”MAC”
132 MACpd = abs( Xp ’ * Xd ); % − e y e ( n ) ;
133 MACpm = abs( Xp ’ * Xm ); % − e y e ( n ) ;
134 MACpr = abs( Xp ’ * Xr ); % − e y e ( n ) ;

CC BY-NC-ND H.P. Gavin October 18, 2018

 8 CEE 541. Structural Dynamics – Duke University – Fall 2016 – H.P. Gavin

7 Results
1 Dynamic Modal IIRS
2 Full Condensation Condensation Condensation
3 0.248754 0.248754 0.248754 0.248756
4 0.746200 0.751695 0.746200 0.746553
5 1.243465 1.273221 1.243465 1.247959
6 1.740425 1.825566 1.740425 1.766281
7 2.236961 2.420037 2.236961 2.334217
8
9 Mc_d =
10
11 5.06 1.66 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00
12 1.66 6.72 1.66 0.00 0.00 0.00 0.00 0.00 0.00 0.00
13 0.00 1.66 6.72 1.66 0.00 0.00 0.00 0.00 0.00 0.00
14 0.00 0.00 1.66 6.72 1.66 0.00 0.00 0.00 0.00 0.00
15 0.00 0.00 0.00 1.66 6.72 1.66 0.00 0.00 0.00 0.00
16 0.00 0.00 0.00 0.00 1.66 6.72 1.66 0.00 0.00 0.00
17 0.00 0.00 0.00 0.00 0.00 1.66 6.72 1.66 0.00 0.00
18 0.00 0.00 0.00 0.00 0.00 0.00 1.66 6.72 1.66 0.00
19 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.66 6.72 1.66
20 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 1.66 8.89
21
22 Mc_m =
23
24 10.11 -0.12 0.12 -0.12 0.13 -0.13 0.14 -0.15 0.17 -0.20
25 -0.12 10.12 -0.12 0.12 -0.13 0.14 -0.14 0.16 -0.17 0.20
26 0.12 -0.12 10.12 -0.13 0.13 -0.14 0.15 -0.16 0.18 -0.21
27 -0.12 0.12 -0.13 10.13 -0.14 0.14 -0.15 0.17 -0.19 0.22
28 0.13 -0.13 0.13 -0.14 10.14 -0.15 0.16 -0.18 0.20 -0.24
29 -0.13 0.14 -0.14 0.14 -0.15 10.16 -0.18 0.20 -0.23 0.28
30 0.14 -0.14 0.15 -0.15 0.16 -0.18 10.19 -0.22 0.26 -0.33
31 -0.15 0.16 -0.16 0.17 -0.18 0.20 -0.22 10.25 -0.31 0.42
32 0.17 -0.17 0.18 -0.19 0.20 -0.23 0.26 -0.31 10.40 -0.59
33 -0.20 0.20 -0.21 0.22 -0.24 0.28 -0.33 0.42 -0.59 11.07
34
35 Mc_r =
36
37 5.04 1.31 2.22 -2.49 1.26 -0.32 -0.05 0.09 -0.05 0.02
38 1.31 4.63 2.58 0.57 -1.01 0.31 0.09 -0.14 0.07 0.00
39 2.22 2.58 7.38 0.47 0.81 -0.68 0.22 -0.04 0.04 -0.09
40 -2.49 0.57 0.47 9.31 0.04 0.58 -0.60 0.38 -0.25 0.24
41 1.26 -1.01 0.81 0.04 9.61 -0.06 0.58 -0.66 0.47 -0.38
42 -0.32 0.31 -0.68 0.58 -0.06 10.02 -0.41 0.65 -0.61 0.45
43 -0.05 0.09 0.22 -0.60 0.58 -0.41 10.64 -0.75 0.63 -0.48
44 0.09 -0.14 -0.04 0.38 -0.66 0.65 -0.75 11.03 -0.86 0.50
45 -0.05 0.07 0.04 -0.25 0.47 -0.61 0.63 -0.86 11.18 -0.91
46 0.02 0.00 -0.09 0.24 -0.38 0.45 -0.48 0.50 -0.91 11.33
47
48
49 round( Kc_d )
50
51 3000 -1000 0 0 0 0 0 0 0 0
52 -1000 2000 -1000 0 0 0 0 0 0 0
53 0 -1000 2000 -1000 0 0 0 0 0 0
54 0 0 -1000 2000 -1000 0 0 0 0 0
55 0 0 0 -1000 2000 -1000 0 0 0 0
56 0 0 0 0 -1000 2000 -1000 0 0 0
57 0 0 0 0 0 -1000 2000 -1000 0 0
58 0 0 0 0 0 0 -1000 2000 -1000 0
59 0 0 0 0 0 0 0 -1000 2000 -1000
60 0 0 0 0 0 0 0 0 -1000 1000
61
62 round( Kc_m )
63
64 5343 -2563 796 -427 289 -222 184 -159 134 -60
65 -2563 3576 -2195 658 -361 252 -199 168 -139 62
66 796 -2195 3437 -2128 622 -340 238 -188 151 -66
67 -427 658 -2128 3401 -2107 609 -331 231 -173 74

CC BY-NC-ND H.P. Gavin October 18, 2018

 Model Condensation 9

68 289 -361 622 -2107 3389 -2100 604 -325 215 -87
69 -222 252 -340 609 -2100 3384 -2096 598 -307 113
70 184 -199 238 -331 604 -2096 3381 -2089 573 -174
71 -159 168 -188 231 -325 598 -2089 3365 -2042 364
72 134 -139 151 -173 215 -307 573 -2042 3252 -1597
73 -60 62 -66 74 -87 113 -174 364 -1597 1341
74
75 round({ Kc_r )
76
77 3484 -1789 834 -424 -22 185 -144 63 -15 1
78 -1789 3548 -3089 1465 -383 -198 284 -177 70 -15
79 834 -3089 5795 -4709 1984 -519 -105 210 -126 33
80 -424 1465 -4709 7017 -4995 2023 -634 46 71 -26
81 -22 -383 1984 -4995 6603 -4413 1702 -583 144 -28
82 185 -198 -519 2023 -4413 5669 -3643 1292 -473 121
83 -144 284 -105 -634 1702 -3643 4802 -2969 912 -237
84 63 -177 210 46 -583 1292 -2969 4080 -2397 446
85 -15 70 -126 71 144 -473 912 -2397 3473 -1660
86 1 -15 33 -26 -28 121 -237 446 -1660 1364

0.8

0.6

0.4

0.2

-0.2

-0.4

-0.6

-0.8
0 20 40 60 80 100

CC BY-NC-ND H.P. Gavin October 18, 2018

10 CEE 541. Structural Dynamics – Duke University – Fall 2016 – H.P. Gavin

% error
12
10

dynamic
8
6
4
2
0
-2
0 20 40 60 80 100

0.013
0.0125
0.012
0.0115
modal

0.011
0.0105
0.01
0.0095
0.009
0.0085
0.008
0 20 40 60 80 100

50
40
30
20
iirs

10
0
-10
-20
0 20 40 60 80 100

log|MAC|, true :: dynamic

0
-1
-2
-3
-4
-5
-6
-7
-8 10
9
8
7 10
6 9
8
5 7
condensed 4 6
3 5
2 4 true
3
1 2
1

CC BY-NC-ND H.P. Gavin October 18, 2018

Model Condensation 11

log|MAC|, true :: modal

0
-1
-2
-3
-4
-5
-6
-7
-8 10
9
8
7 10
6 9
8
5 7
condensed 4 6
3 5
2 4 true
3
1 2
1

CC BY-NC-ND H.P. Gavin October 18, 2018

Journal of Sound and Vibration (1995) 186(2), 311–323

MODEL REDUCTION USING DYNAMIC AND

ITERATED IRS TECHNIQUES
M. I. F
Department of Mechanical Engineering, University of Wales Swansea, Singleton Park,
Swansea SA2 8PP, Wales

S. D. G  J. E. T. P
Department of Mechanical and Electrical Engineering, Aston University, Birmingham B4 7ET,
England

(Received 20 January 1994, and in final form 11 October 1994)

Static or Guyan reduction is widely used to reduce the number of degrees of freedom in
a finite element model but it is exact only at zero frequency. The Improved Reduced System
(IRS) method makes some allowance for the inertia terms and produces a reduced model
which more accurately estimates the modal model of the full system. In this paper the IRS
method is extended by obtaining the equivalent transformation based on dynamic rather than
static reduction. An iterative algorithm, based on the IRS method, is also described. On
convergence this algorithm provides a reduced model which reproduces a subset of the modal
model of the full system. The iterative version of the IRS method based on dynamic reduction
is also described.
7 1995 Academic Press Limited

1. INTRODUCTION
Model reduction, whereby the number of degrees of freedom in a model is reduced, is applied
to large finite element models to give faster computation of the natural frequencies and mode
shapes of a structure. Model reduction also has a role to play in experimental modal analysis,
since the reduced mass and stiffness matrices may also be used to compare the analytical
and experimental models by using orthogonality checks. The transformation inherent in the
model reduction schemes may also be used to expand the measured mode shapes to the full
size of the finite element model, and these mode shapes may then be used in test analysis
correlation or model updating exercises.
One of the oldest and most popular reduction methods is static or Guyan reduction [1].
In this process the inertia terms associated with the discarded degrees of freedom are
neglected. However, while exact for a static model, when applied to a dynamic model the
reduced model generated is not exact and often lacks the required accuracy. O’Callaghan
[2] proposed a modified method which he called the Improved Reduced System (IRS)
method. In this approach an extra term is added to the static reduction transformation to
make some allowance for the inertia forces. This extra term allows the modal vectors of
interest in the full model to be approximated more accurately but relies on the statically
reduced model.
In this paper the IRS method is extended in two ways: by using the transformation from
dynamic reduction instead of static reduction as the basic transformation; and by
introducing an iterative scheme in which the corrective term is generated iteratively using
311
0022–460X/95/370311+13 $12.00/0 7 1995 Academic Press Limited
312 . .  ET AL .
the current best estimate of the reduced model. The convergence of the natural frequencies
of the reduced model to those of the full model is examined. The method is demonstrated
by using a 15-degree-of-freedom discrete mass–spring system, a 90-degree-of-freedom model
of a clamped slotted plate and a 96-degree-of-freedom plane frame.

2. THE STANDARD IMPROVED REDUCED SYSTEM (IRS) METHOD

Possibly the most popular and certainly the simplest reduction method is static reduction,
introduced by Guyan [1]. The state and force vectors, x and f, and the mass and stiffness
matrices, M and K, are split into sub vectors and matrices relating to the master degrees of
freedom, which are retained, and the slave degrees of freedom, which are eliminated. If no
force is applied to the slave degrees of freedom and the damping is negligible, the equation
of motion of the structure becomes

$ %6 7 $ %6 7 6 7
Mmm Mms ẍm K Kms xm f
+ mm = m . (1)
Msm Mss ẍs Ksm Kss xs 0

The subscripts m and s relate to the master and slave co-ordinates respectively. Neglecting
the inertia terms for the second set of equations gives
Ksm xm+Kss xs=0, (2)
which may be used to eliminate the slave degrees of freedom so that

67$ %
xm I
= 4xm 5=Ts xm , (3)
xs −K−1
ss Ksm

where Ts denotes the static transformation between the full state vector and the master
co-ordinates. The reduced mass and stiffness matrices are then given by
MR=TTS MTS , KR=TTS KTS , (4)
where MR and KR are the reduced mass and stiffness matrices. Note that any frequency
response functions generated by the reduced matrices in equations (4) are exact only at zero
frequency. As the excitation frequency increases the inertia terms neglected in equation (1)
become more significant.
O’Callaghan [2] improved the static reduction method by introducing a technique known
as the Improved Reduced System (IRS) method. The method perturbs the transformation
from the static case by including the inertia terms as pseudo-static forces. Obviously, it is
impossible to emulate the behaviour of a full system with a reduced model and every
reduction transformation sacrifices accuracy for speed in some way. O’Callaghan’s
technique [2] results in a reduced system which matches the low frequency resonances of the
full system better than static reduction. However, the IRS reduced stiffness matrix will be
stiffer than the Guyan reduced matrix and the reduced mass matrix is less suitable for
othogonality checks than the reduced mass matrix from Guyan reduction [3]. Gordis [3]
generated the transformation for the standard IRS method by using a binomial series
expansion in frequency. This derivation is summarized here for comparison with the
derivation of the dynamic IRS transformation. From equation (1), for sinusodial excitation
with frequency v,
[Kss−v 2Mss ]xs=−[Ksm−v 2Msm ]xm . (5)
   313
Rearranging this equation and using the binomial theorem gives

xs=−[Kss−v 2Mss ]−1 [Ksm−v 2Msm ]xm=−K−1 2 −1 −1

ss [I−v Mss Kss ] [Ksm−v 2Msm ]xm
=−K−1 2 −1 4 2
ss [I+v Mss Kss +o(v )][Ksm−v Msm ]xm

=−K−1 2 −1 4
ss [Ksm+v (Mss Kss Ksm−Msm )+o(v )]xm , (6)
where o(v 4 ) denotes an error of order v 4. The object is to improve the natural frequency
and mode shape estimation from the reduced model. The reduced model based on static
reduction, to first order in v 2, satisfies
v 2MR xm=KR xm , or v 2xm=M−1
R KR xm , (7)
where MR and KR are the reduced mass and stiffness matrices obtained from static reduction,
and v and xm are a natural frequency of the reduced model and the associated eigenvector
at the master co-ordinates. Equation (6) becomes, upon ignoring the terms in v 4 and higher
powers
xs=[−K−1 −1 −1 −1
ss Ksm+Kss (Msm−Mss Kss Ksm )MR KR ]xm . (8)

Equation (8) defines a transformation that generates the slave co-ordinates from the
master co-ordinates. Although only strictly correct when the co-ordinate vector is a mode
shape, it may be applied as a general transformation, TIRS , which may be conveniently
written as [2]

TIRS=Ts+SMTs M−1
R KR , (9)

where

$ %
0 0
S= .
0 K−1
ss

The reduced mass and stiffness matrices obtained by using the IRS method are then

MIRS=TTIRS MTIRS , KIRS=TTIRS KTIRS . (10)

Although equation (9) is a convenient form in which to express the IRS transformation, in
practice it is inefficient to compute the transformation in this way. The transformation may
be generated by using an expression similar to equation (8).

3. A DYNAMIC IRS PROCEDURE

The transformation in the standard IRS method may be viewed as a perturbation on the
transformation from the static reduction method. A similar perturbation may be obtained
for dynamic reduction. Static reduction is correct only at zero frequency. The transformation
for the dynamic reduction method is correct at a given frequency V and the transformation
equivalent to equation (3) is given by

67$ %
xm I
= 4xm 5=Td xm . (11)
xs −(Kss−V 2Mss )−1(Ksm−V 2Msm )

A dynamic IRS method may be formulated in the same way as the standard IRS method
314 . .  ET AL .
except that the slave degrees of freedom are written as a power series in (v 2−V 2 ). Thus, one
obtains, in a manner similar to that for equation (6),
xs=−[Dss−(v 2−V 2 )Mss ]−1 [Dsm−(v 2−V 2 )Msm ]xm
=−D−1 2 2 −1 −1
ss [I−(v −V )Mss Dss ] [Dsm−(v 2−V 2 )Msm ]xm
=−D−1 2 2 −1 2 2 2 2 2
ss [I+(v −V )Mss Dss +o((v −V ) )][Dsm−(v −V )Msm ]xm

=−D−1 2 2 −1 2 2 2
ss [Dsm+(v −V )(Mss Dss Dsm−Msm )+o((v −V ) ]xm , (12)
2 2
where Dss=Kss−V Mss and Dsm=Ksm−V Msm . The reduced model based on the dynamic
reduction, to first order in (v 2−V 2 ), satisfies
(v 2−V 2 )MR xm=DR xm , or (v 2−V 2 )xm=M−1
R DR xm , (13)
2
where DR=KR−V MR . Equation (13) applies only at the natural frequencies and mode
shapes (compare with equation (7)). The dynamic IRS (DIRS) transformation is therefore
TDIRS=Td+Sd MTd M−1
R DR , (14)
where

$ %
0 0
Sd= .
0 D−1
ss

The reduced mass and stiffness matrices are given by expressions similar to equations (10).

4. ITERATED IRS TECHNIQUES

The transformation (9) relies on the reduced mass and stiffness matrices obtained from
static reduction. Once the transformation has been computed, an improved estimate of these
reduced matrices is available from equations (10). These improved estimates could be used
in the definition of the IRS transformation, equation (9), to give a more acurate
transformation. The transformation for the first iteration is given by equation (9), and for
subsequent iterations is
TIRS,i+1=Ts+SMTIRS,i M−1
IRS,i KIRS,i , (15)
where the subscript i denotes the ith iteration. In equation (15) the transformation TIRS,i is
the current IRS transformation and MIRS,i and KIRS,i are the associated reduced mass and
stiffness matrices given by equations (10). A new transformation, TIRS,i+1 , is obtained which
then becomes the current IRS transformation for the next iteration. Blair et al. [4] considered
an iterated IRS technique, but the second term of equation (15) is changed to give
TIRS,i+1=Ts+SMTs M−1
IRS,i KIRS,i . (16)
Notice that the static transformation is retained in the second term, whereas the updated
transformation is used in equation (15). The algorithm defined by (16) converges to yield
reduced mass and stiffness matrices that do not reproduce the eigenvalues of the full system.
The dynamic IRS method given in the previous section may easily be extended to an iterative
scheme similar to (15).
A full proof that the iterated IRS method converges is extremely difficult and will
not be given here. Convergence may be implied by considering the accuracy of the
series expansion given by equation (6). The generation of the IRS method requires the
approximation of the natural frequencies of the full system by using the reduced model,
   315
equation (7). At each iteration the reduced model will become more accurate and experience
has shown that the iterated IRS method converges monotonically from the statically reduced
model to the SEREP model (defined in reference [5]). The speed of convergence will depend
on the choice of master degrees of freedom, and a poor choice will lead to a very slow rate
of convergence. Choosing master degrees of freedom for which a low frequency mode is not
observable leads to difficulties in SEREP reduction. In the iterated IRS method the modes
that are not observable are simply not included in the reduced model which will converge
to the lowest observable modes.
It will now be proved that when the iterative scheme given by equation (15) converges,
the reduced model reproduces the lower observable eigenvalues and their associated reduced
eigenvectors of the full system. On convergence the transformation is

$ %
I
TIRS= , (17)
tIRS

where
tIRS=−K−1 −1 −1
ss Ksm+Kss (Msm+Mss tIRS )MIRS KIRS . (18)
Assume that l is an eigenvalue with its associated eigenvector fm of the converged, reduced
system defined by MIRS and KIRS . Let fs be the eigenvector at the slave co-ordinates estimated
by using the converged transformation TIRS , so that

$ %
fm
f= =TIRS fm . (19)
fs

Then we have to show that l is an eigenvalue of the full system with eigenvector f. Since
fm is an eigenvector of the reduced system

IRS KIRS fm=lfm ,

M−1 (20)
and equations (18), (19) and (20) may be combined to give
fs=tIRS fm=[−K−1 −1
ss Ksm+lKss (Msm+Mss tIRS )]fm , (21)
or, by premultiplying by Kss and using equation (19),
Kss fs=−Ksm fm+lMsm fm+lMss fs . (22)
Rearranging gives

6 7 6 7
fm fm
l[Msm Mss ] =[Ksm Kss ] . (23)
fs fs

Since l is an eigenvalue of the reduced system with eigenvector fm ,

6 7
fm
lTTIRS MTIRSfm=l[I tTIRS ]M
fs

6 7
fm
=[I tTIRS ]K =TTIRS KTIRS fm . (24)
fs
316 . .  ET AL .
Multiplying out expression (24) and using equation (23) or equivalently equation (22)
produces

6 7 6 7
fm fm
l[Mmm Mms ] =[Kmm Kms ] . (25)
fs fs

Combining equations (23) and (25) gives lMf=Kf and shows that l is an eigenvalue of
the full system with associated eigenvalue f.
One consequence of the convergence to the eigenvalues and eigenvectors of the full model
is that, on convergence, the transformation obtained from the iterated IRS method is exactly
the same as that produced by the SEREP method [5]. This convergence to the SEREP
transformation is dependent on all the lower modes being observable, based on the choice
of the master degrees of freedom. If the lower modes are not observable then the SEREP
transformation is not defined. In this case a SEREP transformation may be obtained by
choosing only the lowest observable modes, and then the iterated IRS transformation will
converge to this redefined SEREP transformation.

5. NUMERICAL CONSIDERATIONS
We will now consider some practical issues relating to the implementation of the iterated
IRS method. Of fundamental importance are the computational requirements of the
method. In general, determining whether the iterated IRS requires more computation than
the SEREP reduction is extremely difficult, particularly since eigenvector extraction and the
iterated IRS method are both iterative procedures, and the number of iterations will depend
on the properties of the full system and the choice of master degrees of freedom. For example,
subspace iteration will require many iterations if the highest eigenvalue in the subspace is
close to the lowest discarded eigenvalue. A poor choice of master degrees of freedom will
require many iterations before the IRS method converges.
Although a full analysis of the computational requirements of the IRS method is difficult,
a comparison can be made with static reduction. Static reduction is often implemented in
commerical finite element codes, and has thus shown the computational benefits of reduction
compared to a direct eigensolution. Even this limited comparison is difficult, due to the
uncertainty in the implementation of the methods and the type of system modelled. To make
the comparison possible, we will assume the mass and stiffness matrices are fully populated.
Let n and N represent the number of master and slave degrees of freedom respectively. In
general, nWN and static reduction, performed one degree of freedom at a time, requires
approximately 23 N 3 floating point multiplications. Calculating the transformation matrix
involves the LU decomposition of Kss and requires an extra 13 N 3 multiplications. The
transformation must be calculated to determine the mode shapes at all the degrees of
freedom. The iterated IRS method requires the computation of equation (15) at every
iteration. The calculation involves only the lower part of the transformation given by
tIRS,i+1=−K−1 −1 −1
ss Ksm+Kss (Msm+Mss tIRS,i )MIRS,i KIRS,i (26)
where

$ %
I
TIRS,i= .
tIRS,i

The first part of equation (26) is the static transformation. Before the iterated IRS method
begins, the matrices K−1 −1
ss Msm and Kss Mss are calculated. If these matrices are calculated at
the same time as the static reduction transformation, this requires, for nWN, approximately
   317

Figure 1. The discrete 15-degree-of-freedom example.

N 3 extra multiplications. For every iteration the reduced mass and stiffness matrices are
calculated by using a total of n(n+1)N 2 multiplications. The other products in
equation (26) require, for nWN, approximately nN 2 multiplications. Thus the iterated
IRS method requires approximately N 3 extra multiplications initially and n(n+2)N 2
multiplications per iteration. The computational requirement per iteration is relatively small,
and overall the iterated IRS will require an extra computational effort similar to that of the
original static reduction.

6. A DISCRETE MASS–SPRING EXAMPLE

We will begin by illustrating the application of the IRS method to a simple mass–spring
system. The system consists of 15 inertias joined together by 16 springs to form a chain, as
shown in Figure 1. The inertias at the ends of the chain are connected to ground by springs.
Each spring has a stiffness of 1 kN/m and each inertia has a mass of 1 kg. The full
15-degree-of-freedom model is reduced to one with five degrees of freedom by selecting the
response at the 1st, 4th, 7th, 11th and 15th inertias. The natural frequencies for the models
resulting from both the static reduced and the iterated IRS method, for ten iterations, are
given in Table 1. The first iteration is the standard IRS method. The table shows that static
reduction fails to reproduce any of the natural frequencies of the original analytical model.
Furthermore, because the static reduction ignores inertia terms the higher frequency modes
are the least accurate. The inertia terms are critical to the accurate estimation of the higher
natural frequencies. When using the iterated IRS method, the lower frequencies converge
most rapidly, although all frequencies will converge to the lower natural frequencies of the

T 1
Convergence of natural frequencies for the 15-dof example when using the iterated IRS
procedure
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iterations 1 2 3 4 5
Static
reduction 6·3227 13·2268 21·6133 29·8367 34·5388
1 6·1992 12·3437 18·5543 25·2854 31·2899
2 6·1991 12·3389 18·4018 24·6036 30·5049
3 6·1991 12·3386 18·3719 24·3860 30·2018
4 6·1991 12·3386 18·3637 24·2966 30·0573
5 6·1991 12·3386 18·3610 24·2542 29·9784
6 6·1991 12·3386 18·3600 24·2323 29·9309
7 6·1991 12·3386 18·3596 24·2204 29·9000
8 6·1991 12·3386 18·3594 24·2135 29·8788
9 6·1991 12·3386 18·3593 24·2095 29·8636
10 6·1991 12·3386 18·3593 24·2071 29·8523
Exact 6·1991 12·3386 18·3592 24·2030 29·8137
318 . .  ET AL .
T 2
Convergence of natural frequencies for the 15-dof example (when using equation (16) in the
iterative IRS procedure)
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iterations 1 2 3 4 5
Static
reduction 6·3227 13·2268 21·6133 29·8367 34·5388
1 6·1992 12·3437 18·5543 25·2854 31·2899
2 6·1996 12·3740 18·9494 26·2979 31·8706
3 6·1996 12·3716 18·8841 25·9335 31·6153
4 6·1996 12·3722 18·8988 26·0674 31·8077
5 6·1996 12·3720 18·8951 26·0113 31·6263
6 6·1996 12·3721 18·8963 26·0350 31·8112
7 6·1996 12·3720 18·8958 26·0243 31·6226
8 6·1996 12·3721 18·8960 26·0290 31·8189
9 6·1996 12·3720 18·8959 26·0270 31·6179
10 6·1996 12·3721 18·8960 26·0276 31·8275
Exact 6·1991 12·3386 18·3592 24·2030 29·8137

full model after sufficient iterations. The transformation matrix converges to that obtained
by the SEREP method [5].
In Table 2 is shown the effect of using the algorithm suggested by Blair et al. [4], equation
(12), to update the transformation matrix. The natural frequencies of the reduced model
converge, but not to the natural frequencies of the full model. The natural frequencies from
the standard IRS method are closer to the frequencies of the full model than the fully
converged frequencies when using this method.
In Table 3 is shown the effect of using a dynamic IRS iterative procedure, based on
equation (16), with a central frequency of 12 rad/s (V=12 rad/s). As expected, the natural

T 3
Convergence of natural frequencies for the 15-dof example (when using the dynamic IRS
iterative procedure, V=12 rad/s)
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iterations 1 2 3 4 5
Dynamic
reduction 7·2094 12·3418 19·8337 28·4467 33·8911
1 6·2029 12·3386 18·3947 24·7950 30·9329
2 6·1993 12·3386 18·3646 24·3947 30·3046
3 6·1992 12·3386 18·3604 24·2808 30·0798
4 6·1991 12·3386 18·3596 24·2384 29·9769
5 6·1991 12·3386 18·3593 24·2203 29·9214
6 6·1991 12·3386 18·3593 24·2119 29·8880
7 6·1991 12·3386 18·3592 24·2077 29·8663
8 6·1991 12·3386 18·3592 24·2056 29·8516
9 6·1991 12·3386 18·3592 24·2044 29·8414
10 6·1991 12·3386 18·3592 24·2038 29·8341
Exact 6·1991 12·3386 18·3592 24·2030 29·8137
   319
T 4
Convergence of natural frequencies for the 15-dof example when using the iterated IRS
procedure where one mode is not observable
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iterations 1 2 3 4 5
Static
reduction 3·4409 10·5616 19·0172 26·0810 34·5282
1 3·4240 10·2327 16·9988 23·8133 31·2802
2 3·4240 10·2320 16·9303 23·5304 30·4696
3 3·4240 10·2320 16·9218 23·4549 30·1272
4 3·4240 10·2320 16·9204 23·4290 29·9463
5 3·4240 10·2320 16·9201 23·4186 29·8394
6 3·4240 10·2320 16·9200 23·4140 29·7721
7 3·4240 10·2320 16·9200 23·4119 29·7280
8 3·4240 10·2320 16·9200 23·4108 29·6981
9 3·4240 10·2320 16·9200 23·4102 29·6776
10 3·4240 10·2320 16·9200 23·4100 29·6631
Exact 3·4240 10·0000 10·2320 16·9200 23·4096

frequency closest to 12 rads/s, that is the second frequency, converges most quickly. More
iterations would confirm that the natural frequencies do indeed converge to those of the full
model.
The discrete example will now be used to demonstrate the effect of an unobservable mode
in the frequency range of interest. Suppose that the stiffnesses k15 and k16 in the system shown
in Figure 1 are set to k15=0 and k16=100 Nm−1 . Thus mass number 15 is decoupled from
the other 14 masses. Furthermore, the stiffness k16 is chosen so that the local mode involving
mass 15 is the second natural frequency of the full system. Suppose that we choose the
response of the 1st, 4th, 7th, 11th and 14th masses as the master degrees of freedom. Now
the local mode involving mass 15 is not observable. The effect of applying the iterated IRS
method to this system is shown in Table 4. Notice that the method has converged to natural
frequencies of the full system, but the second mode has been missed out since it is not
observable.

7. EXAMPLES OF CONTINUOUS STRUCTURES

We now illustrate the application of the iterative IRS method to a discrete model of a
continuous structure. The structure considered is a plate with a slot on one side and clamped
along two other sides, as shown in Figure 2. The full model has 90 degrees of freedom and

Figure 2. The finite element model of the slotted plate, showing the reduced co-ordinates.
320 . .  ET AL .
T 5
Convergence of natural frequencies for the plate example when using the iterated IRS procedure
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iteration 1 2 3 4 5 6
Static
reduction 30·5886 108·257 147·269 225·203 328·732 441·383
1 30·5162 104·745 141·253 208·938 273·084 368·303
2 30·5162 104·740 141·233 208·587 265·959 361·732
3 30·5162 104·740 141·232 208·538 263·929 359·185
4 30·5162 104·740 141·232 208·527 263·193 357·846
5 30·5162 104·740 141·232 208·524 262·885 357·009
6 30·5162 104·740 141·232 208·523 262·744 356·426
7 30·5162 104·740 141·232 208·522 262·677 355·992
8 30·5162 104·740 141·232 208·522 262·642 355·654
9 30·5162 104·740 141·232 208·522 262·624 355·386
10 30·5162 104·740 141·232 208·522 262·615 355·169
Exact 30·5162 104·740 141·232 208·522 262·603 354·107

is reduced to the six degrees of freedom shown in Figure 2. This choice of co-ordinates
represents the best selection of co-ordinates for Guyan reduction on the basis of the relative
importance of the diagonal stiffness and inertia terms [6, 7]. The result of using the iterative
IRS method on this example is shown in Table 5. The convergence of the natural frequencies
of the reduced model to those of the full model is clear. As expected, the lower modes
converge more quickly than the higher modes. The result of applying the iterative dynamic
IRS method based on a frequency of 150 rad/s is shown in Table 6. The natural frequencies
close to 150 rad/s converge quicker than those frequencies further away.
Our third example concerns a two dimensional frame shown in Figure 3 and modelled
with 34 nodes. Each node has three degrees of freedom and two of the nodes are grounded,

T 6
Convergence of natural frequencies for the plate example (when using the dynamic IRS iterative
procedure, V=150 rad/s)
Natural frequency (rads/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iteration 1 2 3 4 5 6
Dynamic
reduction 58·5194 109·257 141·350 213·130 305·242 423·897
1 30·5561 104·750 141·232 208·566 266·898 364·963
2 30·5173 104·740 141·232 208·526 263·669 359·799
3 30·5162 104·740 141·232 208·523 262·949 357·851
4 30·5162 104·740 141·232 208·522 262·730 356·812
5 30·5162 104·740 141·232 208·522 262·653 356·145
6 30·5162 104·740 141·232 208·522 262·624 355·674
7 30·5162 104·740 141·232 208·522 262·612 355·326
8 30·5162 104·740 141·232 208·522 262·607 355·062
9 30·5162 104·740 141·232 208·522 262·605 354·859
10 30·5162 104·740 141·232 208·522 262·604 354·702
Exact 30·5162 104·740 141·232 208·522 262·603 354·107
   321

Figure 3. The frame example showing the node positions (+) and the reduced co-ordinates (arrows).

giving 96 degrees of freedom in the full model. This model is reduced to the 12 degrees of
freedom shown in Figure 3. This co-ordinate selection is not the optimum set for the
application of static reduction [7] but is satisfactory for our purposes. Applying the iterated
IRS method to this example gives the results shown in Table 7. Only the first six frequencies
are shown. The convergence is satisfactory, although not particularly rapid. A better choice
of the reduced co-ordinates would help to speed the convergence. Our first two examples
demonstrated the iterated, dynamic IRS method when the centre frequency, V, is relatively
small. What happens if a large centre frequency is used? In this example the 12th natural
frequency of the full model is at approximately 4455 rad/s. Suppose that we tried a centre
frequency of 3000 rad/s. The iterated DIRS method converges to the higher natural
frequencies of the full system and the first three frequencies are missed. One major problem
is the appearance of spurious natural frequencies in the reduced model. These spurious
frequencies are easily recognized as they are not constant. In the current example, the effect
is demonstrated in Figure 4. On a diagram such as this it is easy to pick out the true natural
frequencies, which accurately reproduce those of the full model. Notice that the natural
frequencies closest to V=3000 rad/s converge the fastest.

T 7
Convergence of natural frequencies for the frame example
Natural frequency (rad/s)
ZXXXXXXXXXXXXXXXXCXXXXXXXXXXXXXXXXV
Iterations 1 2 3 4 5 6
Static
reduction 471·776 1536·00 1712·52 1914·52 2041·44 2702·54
1 464·687 1170·34 1646·01 1690·56 1903·97 1922·29
2 464·679 1145·59 1638·26 1671·79 1886·98 1907·10
3 464·676 1139·89 1635·11 1667·04 1880·62 1905·43
4 464·675 1136·77 1633·07 1664·71 1877·08 1904·78
5 464·674 1134·78 1631·62 1663·35 1874·79 1904·43
6 464·674 1133·41 1630·54 1662·46 1873·20 1904·21
7 464·674 1132·41 1629·72 1661·85 1872·03 1904·06
8 464·673 1131·66 1629·08 1661·40 1871·14 1903·95
9 464·673 1131·08 1628·57 1661·07 1870·44 1903·87
10 464·673 1130·61 1628·15 1660·81 1869·88 1903·80
Exact 464·672 1127·42 1624·88 1659·24 1865·83 1903·45
322 . .  ET AL .

Figure 4. The convergence of the natural frequencies when using the dynamic IRS method with a large centre
frequency V=3000 rad/s.

8. CONCLUSIONS
An iterative method based on the Improved Reduced System (IRS) algorithm has been
presented. The natural frequencies of the reduced model converge to the lower frequencies
of the full model, with those of lower frequency converging more quickly. A dynamic
IRS method, based on dynamic rather than static reduction has been derived and tested.
The iterative version of this method has also been derived. For a small centre frequency,
the natural frequencies of the reduced model converge to the lower frequencies of the full
model. The natural frequency of the reduced model closest to the chosen centre frequency
for the reduction converges quickest. If the centre frequency is large then the iterated
dynamic IRS method misses the lower frequencies of the full model. Spurious frequencies
are introduced, although they are easily identified as they are not constant from one iteration
to the next.

REFERENCES
1. R. J. G 1965 American Institute of Aeronautics and Astronautics Journal 3(2), 380. Reduction
of stiffness and mass matrices.
2. J. C. O’C 1989 Proceedings of the 7th International Modal Analysis Conference, Las
Vegas, January 1989, 17–21. A procedure for an improved reduced system (IRS) model.
3. J. H. G 1992 Proceedings of the 10th International Modal Analysis Conference, San Diego,
California, February 1992, 471–479. An analysis of the improved reduced system (IRS) model
reduction procedure.
4. M. A. B, T. S. C and J. M. D 1991 Proceedings of the 9th International Modal
Analysis Conference, Florence, Italy, April 1991, 621–626. An iterative approach to a reduced mass
matrix.
5. J. C. O’C, P. A and R. R 1989 Proceedings of the 7th International Modal
Analysis Conference, Las Vegas, January 1989 29–37. System equivalent reduction expansion
process (SEREP).
   323
6. R. D. H and J. H. O 1975 Earthquake Engineering and Structural Dyanmics 3, 375–383.
Automatic masters for eigenvalue economisation.
7. J. E. T. P, M. I. F and S. D. G 1994 American Institute of Aeronautics and
Astronautics Journal 32,(2), 407–414. Automatic choice of measurement locations for dynamic
testing.