Lihong Feng
Max Planck Institute for Dynamics of Complex Technical Systems
Computational Methods in Systems and Control Theory
MAX-PLANCK-INSTITUT
DYNAMIK KOMPLEXER
TECHNISCHER SYSTEME
MAGDEBURG
Idea of MOR
Original large
ODE or PDE
Numerical simulation
of the full model
MOR
Much Faster
Reduced small
ODE or PDE
Numerical simulation
of the small model
Moment-matching MOR
Original linear system
CdX / dt GX Bu(t )
y (t ) LX (t )
Expanding around s0
Laplace transform
X Rn
sCX ( s) GX ( s) BU ( s)
Y ( s) LX ( s)
X ( s ) ( sC G ) B [( s0C G ) 1 C ]i ( s0C G ) 1 B( s s0 ) i
1
i 0
q
[( s0C G ) 1 C ]i ( s0C G ) 1 B( s s0 ) q
i 0
Projection matrix V:
X VZ
y (t ) LX (t )
Z R r , r n
CVdZ / dt GVZ Bu (t )
y (t ) LVZ (t )
Galerkin Projection
V Te 0
e CVdZ / dt GVZ Bu (t )
V T CVdZ / dt V T GVZ V T Bu (t )
y (t ) LVZ (t )
Define: C V T CV , G V T GV , B V T B, L LV
C dZ / dt G Z B u (t )
y (t ) L Z (t )
Max Planck Institute Magdeburg
i 0
1
i
1
Moments: M i L[( s0C G ) C ] ( s0C G ) B, i 0,1,...
If:
Then: H (s )
M i M i , i 0, , j
H ( s)
Overview
Linearization MOR.
Qudratic MOR.
Bilinearization MOR.
Variational analysis MOR.
Trajectory piece-wise linear MOR.
Other MOR methods for nonlinear systems.
Conclusions.
References.
Linearization MOR
Original large ODE
CdX / dt f ( X ) Bu(t )
Linearization: approximate
f ( X ) by a linear function
y (t ) LX (t )
f (X )
~
f ( X ) AX
M i ( A1C )i r , i 0,1,...
Max Planck Institute Magdeburg
CdX / dt f ( X 0 ) D f ( X X 0 ) Bu(t )
~
y (t ) LX (t )
u (t )
CdX / dt AX [ B, f ( X 0 ) D f X 0 ]
1
~
y (t ) LX (t )
V orthogonalization{r , M 1r , M 2 r , M j r}
~
r A1 B , M i [( s0C A) 1 C ]i r , i 0,1, .
03.02.2012
Example
1
i(t)
2
C
3
C
N-3
N-2
N-1
g ( x1 ) g ( x1 x2 )
g(x x ) g(x x ) 1
1
2
2
3
dX
u (t ),
dt g ( xk 1 xk ) g ( xk xk 1 )
g ( xn 1 xn )
y (t ) LX (t )
g ( x) e40x x 1
g ' ' (0) 2
x
2!
g (0) g ' (0) x 41x
82 x1 42 x2
41x 82 x 41x 1
1
2
3
dX
u (t ),
dt 41xk 1 82 xk 41xk 1
41xn 1 41xn
y (t ) LX (t )
8
Example
0.025
voltage at node 1
0.02
0.015
+ q=10
__ full simulation
0.01
0.005
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
Example
0.025
voltage at node 1
0.02
0.015
+ q=10
-- q=20
__full simulation
0.01
0.005
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
10
Example
0.025
voltage at node 1
0.02
0.015
+ q=10
-- q=20
0.01
...linear, no reduction
__ full simulation
0.005
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
11
Quadratic MOR
Approximate f ( X ) by a quadratic polynomial g ( X )
CdX / dt f ( X ) Bu(t )
y (t ) LX (t )
1
f ( X ) f ( X 0 ) D f ( X X 0 ) ( X X 0 )T H f ( X 0 )( X X 0 )
2
1
f ( X 0 ) D f ( X X 0 ) ( X X 0 )T H f ( X 0 )( X X 0 )
2
f (X )
~
CdX / dt AX X T WX B u (t )
~
y (t ) LX (t )
X VZ , Z R q , q n
g(X )
~
V T CXdZ / dt V T AVZ Z T V TWVZ V T B u (t )
y (t ) LVZ (t )
V orthogonalization{r , M 1r , M 2 r , M j r}
~
r ( s0C A) 1 B , M i [( s0C A) 1 C ]i r , i 0,1, .
Max Planck Institute Magdeburg
12
Example
1
g ( x1 ) g ( x1 x2 )
i(t)
g(x x ) g(x x ) 1
1
2
2
3
dX
u (t ),
dt g ( xk 1 xk ) g ( xk xk 1 )
g ( xn 1 xn )
y (t ) LX (t )
2
C
3
C
N-3
N-2
N-1
g ( x) e40x x 1
g ' ' (0) 2
g ( x) g (0) g ' (0) x
x
2!
g (0) g ' (0) x 41x 800 x 2
82 x1 42 x2
41x 82 x 41x
2
2
800
(
x
x
)
800
(
x
x
)
1
2
3
1
2
2
3
1
dX
u (t ),
2
2
dt 41xk 1 82 xk 41xk 1 800 ( xk 1 xk ) 800 ( xk xk 1 )
0
41xn 1 41xn
800 ( xn 1 xn ) 2
y (t ) LX (t )
13
Example
1
i(t)
2
C
3
C
82 x1 42 x2
800 x 800 ( x1 x2 )
41x 82 x 41x
2
2
800
(
x
x
)
800
(
x
x
)
1
2
3
1
2
2
3
dX
u (t ),
2
2
41
x
82
x
41
x
dt k 1
k
k 1
800 ( xk 1 xk ) 800 ( xk xk 1 )
0
2
41
x
41
x
800 ( xn 1 xn )
n 1
n
2
1
y (t ) LX (t )
~
CdX / dt AX X T WX B u (t )
~
y (t ) LX (t )
N-3
N-2
N-1
14
Example
W
W 1 R nn
W i R n n
1600 800 0 0
800 0 0
800
0
0
0 0
0 0 0
0 0
800 800
0
800
0
800
0
800 800
0
0
0
0
0
0 0
i 1, i, i 1
i(t)
2
C
3
C
N-3
N-2
N-1
0
0
0
0
0 i 1
i
i 1
W n R nn
0
0 0 0
0 0
0 0 0
800 800
0 0 800 800
X TW 1 X
T
T
i
X WX X W X
X TW n X
Z T V T W 1VZ
T T
T
T
i
Z V WVZ Z V W VZ
Z T V T W nVZ
15
Example
0.025
0.02
voltage at node 1
0.015
linear, no reduction
__ full simulation
0.01
0.005
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
16
Example
0.018
0.016
voltage at node 1
0.014
0.012
0.01
__ full simulation
0.008
0.006
0.004
0.002
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
17
Example
0.018
0.016
voltage at node 1
0.014
0.012
quadratic, no reduction
0.01
__ full simulation
0.008
0.006
0.004
0.002
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
18
Bilinearization MOR
Approximate f ( X ) by quadratic polynomial g ( X ), but written into Kronecker
product
Taylor series expansion:
dX / dt f ( X ) Bu(t )
y (t ) LX (t )
1
f ( X ) f ( X 0 ) D f ( X X 0 ) ( X X 0 )T H f ( X 0 )( X X 0 )
2
1
f ( X 0 ) D f ( X X 0 ) ( X X 0 )T H f ( X 0 )( X X 0 )
2
f ( X 0 ) A1 X A2 X X
f (X )
dX / dt A X N X u (t ) B u (t )
y (t ) L X (t )
g(X )
X R N , N n2
X VZ , Z Rq , q n
dZ / dt V T AVZ V T N VZu(t ) V T B u (t )
y (t ) LVZ (t )
19
Bilinearization MOR
dX / dt f ( X ) Bu(t )
y (t ) LX (t )
A
A 1
0
dX / dt A X N X u (t ) B u (t )
Carleman bilinearization
A2
A1 I I A1
y (t ) L X (t )
0
0
N
B
B
0
Kronecker product
X
X X
B
B
0
L L 0
a11B a1n B
A B
a B a B
mn
m1
Carleman bilinearization:
[1] W.J. Rugh, Nolinear System Theory, The John Hopkins University Press, Boltimore, 1981.
[2] S. Sastry, Nonlinear Systems: Analysis, Stability and Control, Springer, New York, 1999.
20
Bilinearization MOR
dX / dt A X N X u (t ) B u (t )
How to compute V?
y (t ) L X (t )
Volterra series expression of bilinear system
According to the theory in [Rugh 1981], the output response of the bilinear system
can be expressed into Volterra series,
y (t ) y n (t )
n 1
( reg )
n
(t1 , t 2 , , t n ) L e
A t n
At n1
N e
At1
N e
hn( reg ) ( s1 , s2 , , sn ) LT ( sn I A) 1 N ( sn 1 I A) 1 N ( s2 I A) 1 N ( s1 I A) 1 B
(1) n cT ( I sn A1 ) 1 A1 N ( I sn 1 A1 ) 1 A1 N ( I s2 A1 ) 1 A1 N ( I s1 A1 ) 1 A1 B
Max Planck Institute Magdeburg
21
Bilinearization MOR
dX / dt A X N X u (t ) B u (t )
How to compute V?
y (t ) L X (t )
Laplace transform:
hn( reg ) ( s1 , s2 , , sn ) L( sn I A) 1 N ( sn 1 I A) 1 N ( s2 I A) 1 N ( s1 I A) 1 B
(1) n LT ( I sn A1 ) 1 A1 N ( I sn 1 A1 ) 1 A1 N ( I s2 A1 ) 1 A1 N ( I s1 A1 ) 1 A1 B
( I sn A1 ) 1 I A1sn Ai sni
( reg )
n
ln 1
l1 1
Multimoments:
22
Bilinearization MOR
dX / dt A X N X u (t ) B u (t )
How to compute V?
y (t ) L X (t )
( reg )
n
ln 1
l1 1
Multimoments:
range{V j } K q j { A1 , A1 NV j 1} { A1 NV j 1, A1 NV j 1 ,, A
q j
NV j 1}
Reduced model:
dZ / dt V T AVZ V T N VZu(t ) V T B u (t )
y (t ) LVZ (t )
23
Example
0.018
original system
quadratic MOR, q=20
quadratic, no reduction
bilinearization, q=20
0.012
range{V1} span{( A ) 1 B, , ( A ) 19 B}
range{V2 } span{( A ) 1 NV1 (:,1)}
0.01
0.008
0.006
0.004
0.002
0
0
4
6
time, u(t)=0, t<3, u(t)=1, t>=3;
10
24
Original system:
1
f ( X ) f ( X 0 ) D f ( X X 0 ) ( X X 0 )T H f ( X 0 )( X X )
2
f ( X 0 ) A1 X A2 X X A3 X X X
dX / dt f ( X ) Bu(t )
y (t ) LX (t )
~
dX / dt A1 X A2 X X B u~ (t )
y (t ) LX (t )
or
~
dX / dt A1 X A2 X X A3 X X X B u~ (t )
y (t ) LX (t )
Variational analysis:
~
dX / dt A1 X A2 X X B u~ (t )
y (t ) LX (t )
or
~
dX / dt A1 X A2 X X A3 X X X B u~ (t )
y (t ) LX (t )
X ( , t ) X 1 (t ) a 2 X 2 (t ) 3 X 3 (t )
25
X (t ) X 1 (t ) a 2 X 2 (t ) 3 X 3 (t )
d (X 1 2 X 2 3 X 3 ) / dt A1 (X 1 2 X 2 3 X 3 )
A2 [(X 1 2 X 2 3 X 3 ) (X 1 2 X 2 3 X 3 )]
~
A3 [(X 1 2 X 2 3 X 3 ) (X 1 2 X 2 3 X 3 ) (X 1 2 X 2 3 X 3 )] B u~ (t )
y (t ) LX (t )
~
dX 1 (t ) / dt A1 X 1 (t ) B u~ (t )
2 :
dX 2 (t ) / dt A1 X 2 (t ) A2 ( X1 X1 )
3 :
dX 3 (t ) / dt A1 X 3 (t ) A2 ( X 1 X 2 X 2 X 1) A3 ( X 1 X 1 X 1 )
26
X (t ) X 1 (t ) a 2 X 2 (t ) 3 X 3 (t )
~
dX 1 (t ) / dt A1 X 1 (t ) B u~ (t )
2 :
dX 2 (t ) / dt A1 X 2 (t ) A2 ( X1 X1 )
3 :
dX 3 (t ) / dt A1 X 3 (t ) A2 ( X 1 X 2 X 2 X 1) A3 ( X 1 X 1 X 1 )
X 1 V1Z1
~
~
V1 span{ A B ,, A1 q1 B}
X 2 V2 Z 2
X 3 V3 Z 3
V3 span{ A11[ A2 , A3 ], , A1 q2 [ A2 , A3 ]}
1
1
X (t ) X 1 (t ) a 2 X 2 (t ) 3 X 3 (t )
X 1 2 X 2 3 X 3
V1Z1 2V2 Z 2 3V3 Z 3
X (t ) span{V1 ,V2 ,V3}
27
~
dX / dt A1 X A2 X X A3 X X X B u~ (t )
y (t ) LX (t )
~ u~(t )
dX / dt A1 X A2 X X A3 X X X B
y (t ) LX (t )
Compute V:
X (t ) VZ
~~
Reduced model: dZ / dt V T A1VZ V T A2VZ VZ V T A3VZ VZ VZ V T B
u (t )
y (t ) LVZ (t )
28
voltage at node 1
Example
0.018
0.016
0.014
0.012
0.01
0.008
0.006
0.004
V20 orth{ A2 }
0.002
0
0
4
6
time, u(t)=0, t<3; u(t)=1, t>=3
range{V } span{V1 , V2 }
10
29
Example
V for [Phillips 2000] method:
0.018
0.016
V20 A1 A2 (V1 V1 )
voltage at node 1
0.014
0.012
0.01
0.008
0.006
0.004
0.002
0
0
original
our method, q=9, relative error=0.0073
Phillips 2003 method, q=9, relative error=0.0076
2
4
6
8
time, u(t)=0, t<0.3; u(t)=1, t>=0.3
30
s 1
~ g ( X ) Bu ,
dX / dt w
i i
y (t ) LX (t )
i 0
y (t ) LX (t )
f (X )
g3 ( X )
g1 ( X )
g i ( X ) f ( X i ) Ai ( X X i ), i 0,1, , s 1
X2
Ai (a jk ),
X1
X0
g4 ( X )
a jk
f j ( X )
xk
Xi
g2 ( X )
31
dX / dt f ( X ) Bu(t )
dX / dt gi ( X ) Bu,
y (t ) LX (t )
y(t ) LX (t )
i 0
~ (X ) f (X ) w
~ ( X ) A ( X X ), i 0,1,, s 1
gi ( X ) w
i
i
i
i
i
~ AX w
~ ( f (X ) A X )
w
i i
i
i
i i
s 1
~ ~ T
~ AX B
dX / dt w
[u, wi ] ,
i i
How to compute V?
~
~
range{Vi } span{ A Bi , , Ai qi Bi }
1
i
range{V } span{V1 , , Vs 1}
i 0,1, , s 1
i 0
y(t ) LX (t )
~
B [ B, B0 ], B0 f ( X i ) AX i
32
~ ~ T
~ AX B
dX / dt w
[u, wi ] ,
i i
i 0
y(t ) LX (t )
Reduced model:
s 1
~ ~ T
~ V T A VZ V T B
dZ / dt w
[u, wi ] ,
i
i
i 0
y (t ) LVZ (t )
33
34
Conclusions
Krylov subspace based nonlinear MOR methods include linearization
methods, quadratic methods, bilinearization methods,
variational analysis methods, trajectory piece-wise linear (polynomial)
methods, etc..
Quadratic, bilinearization, variational analysis are suitable for weakly
nonlinear systems.
Tranjectory picec-wise linear (polynomial) performs better for strong
nonlinear systems.
Other methods like POD, reduced basis could be applied to MEMS
simulation.
35
References
Quadratic MOR:
[1] Yong Chen, Model order reduction for nonlinear systems, MIT master thesis, 1999.
Bilinearization MOR:
[2] J. R. Phillips, Projection Frameworks for Model Reduction of Weakly Nonlinear
Systems, Proc. IEEE/ACM DAC, pp. 184-189, 2000.
[3] Z. Bai, and D. Skoogh, A projection method for model reduction of bilinear
dynamical systems, Linear Algebra Appl., 415, pp. 406425, 2006.
[4] L. Feng and P. Benner, A Note on Projection Techniques for Model Order Reduction of
Bilinear Systems, AIP Conf. Proc. 936, pp. 208-211, 2007.
36
References
03.02.2012
37