HANKEL SINGULAR VALUES OF FLEXIBLE

STRUCTURES IN DISCRETE TIME

K.B. Lim
NASA Langley Research Center
W. Gawronskiy
Jet Propulsion Laboratory, California Institute of Technology

Abstract
Approximate analytical expressions for controllability and observability grammian matrices
and Hankel singular values of discrete LTI exible structures are derived. The diagonal domi-
nance property of the discrete grammians is shown which results in the invariance of the principal
directions. The approximate discrete Hankel singular values converge to the continuous formula
with increased sampling rate while the controllability and observability grammians go to zero
and innity respectively. The approximate formula are accurate up to frequencies close to the
Nyquist.

1 Introduction
It is well known that degrees of controllability and observability for linear systems are conveniently
captured by the singular values of grammians. These singular values have a wide range of applica-
tions from system identication and model reduction to actuator and sensor placement for eective
control and sensing conguration. Although the physical interpretation and approximating formula
has been investigated in detail in the past for continuous systems (see for example [1]-[9]), there is
a signicant lack of results for discrete systems although the results are expected to be analogous
to the continuous case. This need for results in the discrete domain is painfully clear, for example,
when a control engineer is faced with the task of analysis and design of controllers for a large order
model of a discrete system.

Research Engineer, Guidance & Control Branch, Flight Dynamics & Control Division, MS 161, Hampton, Virginia
23681, k.b.lim@larc.nasa.gov
y Technical Sta, Communications Ground Systems, MS 144-201, Pasadena, California, 91109,
wodek@gdansk.jpl.nasa.gov

1
 In this paper, analytical expressions for controllability and observability grammian matrices
and Hankel singular values of discrete LTI exible structures are derived. Results based on two
types of models for discrete exible structures are given: discretization of continuous systems
via sampling and zero-order-hold (ZOH) and implicitly discrete models. The rst type of model
is typically obtained by analytical means while the second type typically arises from system
identication. The sampled/ZOH model is a parameterized model which allows direct comparisons
to continuous singular values when the sampling rate is varied. Derivations of the approximate
singular value formulas are given only for the rst type of model and the results based on the
second type of parameterization are summarized as corollaries. For the class of exible structures
with small damping and distinct frequencies, the above formulae are signicantly simplied. The
approach is complementary to the earlier results on continuous time exible structures reported
in [5, 6, 7]. Similar to the continuous case, the diagonal dominance property of the discrete
grammians for small damping is shown. As a result, the approximate invariance of principal
controllability and observability directions also hold for discrete time exible structures. The
dependence of the grammians on the sampling time and in particular their deviation from the
corresponding continuous grammian is investigated. In particular, it is shown that the approximate
discrete Hankel singular values formula converges to the approximate continuous formula with
increased sampling rate while the controllability and observability grammians go to zero and innity
respectively. It is shown by numerical examples that the approximate formula for singular values
of discrete controlability and observability grammians and Hankel singular values are accurate up
to frequencies close to the Nyquist frequency. Two levels of damping are assumed to evaluate the
eect of violating the assumption of a lightly damped exible structure.

2 Flexible Structure
2.1 Continuous Time
Let the triple (A; B; C ) denote a modal state-space representation of a exible structure with n
structural modes. Following earlier denitions [6, 7, 8, 9, 10] dene the modal state vector, x, of
dimension n2  1, where n2 = 2n, such that
 T
x = _1 !11


_n !nn (1)

2
then the modal state equations take the form
2
B1 3

x_ = diag(A1;


; An)x + 64 ... 75 u (2)

Bn
h i
y = C1


Cn x (3)

where
" # " #
2i !i !i bi
Ai = !i 0 ; Bi = 0
h i
Ci = cri !i cdi
1
(4)

and i = 1; . . . ; n, bi = iT E , cdi = F i and cri = G i. Notice that for small damping

0 < i  1 (5)

the above choice of the state vector gives the approximately normal state matrix and hence
approximately orthogonal eigenvectors. For exible structures with distinct natural frequencies,
the steady-state controllability and observability grammians asymptotically (as  ! 0) approach
2-by-2 block diagonal matrices as given in [6, 7, 10]
!2
^2 ^
ci = 4ii! ; oi2 = 4i! ; i4 = 4ii!i
2
2
(6)
i i i i i i
where

ij2 = bi bTj (7)

^i2 = !12 cTdicdi + cTricri (8)
i
are the modal grammian coecients [8, 9].
2.2 Discrete Time
Two dierent forms of parameterizations of the discrete exible structures are considered. The
rst form is used in the detailed derivations in the remaining sections and the results based on the
second form of parameterization are given as corollaries without details.

3
2.2.1 Sampled/ZOH Model
Consider a continuous exible structure as dened by the block diagonal modal state space repre-
sentation in Section 2.1 and sampled at the outputs with period T and with a ZOH at the inputs.
The state equation is given in this case by
~ (k) + Bu
x(k + 1) = Ax ~ (k) (9)
~ (k) + Du
y(k) = Cx ~ (k) (10)
where C~ = C , D~ = D while the discrete system matrices, A~ and B~ are given by
A~ = e AT = blk-diag(A~ 1 (T ); . . . ; A~ n (T )) (11)
Z (k+1)T
B~ = eA((k+1)T  ) dB
kT
= blk-diag(M~ 1 ; . . . ; M~ n)B (12)
where M~ i = A~i()d, and by denoting the damped frequency of the continuous structure as
RT
o
q
!di = !i 1 i2, the ith block of A~ is
i !i T
"
sin(! T ) !i sin(!di T )
#

A~ (T ) = e
i !i
+!di cos(!didiT ) (13)
i !di !i sin(!d i T ) sin(!di T )
i !i
cos(!d i T )
+!d i

2.2.2 Implicitly Discrete Model

In general, the state matrix of a discrete time model of a exible structure may be fully populated.
The following denes a similarity transformation to block diagonalize the state matrix:
Lemma 1 Let the quadruple (Az ; Bz ; Cz ; Dz) denote the discrete state space matrices of a exible
structure. Let (zi; vi ) denote the ith eigenvalue and eigenvector pair of Az . The state transformation
matrix  
V = Re(v1 ) Im(v1 ) . . . Re(vn ) Im(vn ) (14)
block diagonalizes the state equations as in (9) and (10) where

A~ = blk-diag(A~1(T ); . . . ; A~n(T )) (15)

B~ = V 1Bz ; C~ = Cz V; D~ = Dz (16)
and " #
Re(zi ) Im(zi )
A~i = Im( (17)
zi) Re(zi)

4
For a lightly damped exible structure, its ith discrete eigenvalue lies just inside the unit circle and
can be written as zi = e(  +j )T where i > 0. The A~i matrix in (17) then becomes
i i

" #

A~i = cos( i T ) sin( iT ) T

sin( iT ) cos( iT ) e (18)
i

Since the above discrete eigenvalue is related to the eigenvalue of the corresponding sampled
continuous signal the following analogy holds: i $ i !i , i $ !di .
2.3 Small Damping Approximation
Assuming that the sampling rate is suciently fast such that the sampling theorem is satised (see
for example p.111 in [11]), i.e., !i  T for all i, one obtains from (5)

i!iT  1 (19)

The 2-by-2 block matrix A~i in (13) can be approximated by

A~i(T ) 
= i (T )e i !i T (20)

where i (T ) is an orthogonal matrix of the form

" #

i (T ) = cos( !diT ) sin(!diT ) (21)

sin(!di T ) cos(!di T )
Note that (20) and (21) are analogous to (18). Using (20), the denite integral in (12) reduces to

B~ 
= blk-diag(M1; . . . ; Mn )B (22)

where " #

Mi = !12 ab i abi (23)

i i i
and

ai = !i sin(!i T ) + O(i) (24)

bi = !i(1 cos(!iT )) + O(i) (25)

5
3 Controllability Grammian
3.1 Denition
For the time interval (ko T; k1T ), the discrete time controllability grammian, Wc (ko; k1), is dened
in terms of the state transition matrix, , and input matrix, B~ ,
kX
1 1
Wc (ko ; k1 ) = (k1; k + 1)B~ B~ T T (k1; k + 1)
k=ko
= Pc(k1 ko)PcT (k1 ko) (26)

where Pc (k1 ko) id the discrete time controllability matrix

h i
Pc(k1 ko) = B~ A~B~ . . . A~k1 ko 1
B~ (27)

It can be shown that the above grammian satises the following equation
~ c (ko ; k1)A~T + B~ B~ T = Wc (ko ; k1)
AW
+(k1; ko)B~ B~ T T (k1; ko) (28)

For asymptotically stable linear systems, the last term in (28) vanishes as k1 ! 1. This leads to
the steady-state discrete time controllability grammian, Wc1 , which satises the following Sylvester
equation
~ c1 A~T + B~ B~ T = Wc1
AW (29)
3.2 Closed-Form Solution
By taking advantage of the 2 by 2 block diagonal form of the state matrix in (11), the Sylvester
equation in (29) can be written as a set of 2-by-2 Sylvester equations
A~ i [Wc1 ]ij A~ Tj + [B~ B~ T ]ij = [Wc1 ]ij (30)

where i; j = 1; . . . ; n, and
~ i = A~i (T )
A (31)
[B~ B~ T ]ij = B~i Bi BjT B~j (32)

and [Wc1 ]ij is the (i; j )th 2 by 2 block of [Wc1 ]. For small damping, (30) can be approximated
by
e i !i T i [Wc1 ]ij jT e j !j T [Wc1 ]ij = [B~ B~ T ]ij (33)

6
and equivalently by postmultiplying by the orthogonal matrix j one obtains
i i [Wc1 ]ij [Wc1 ]ij j j 1 = [B~ B~ T ]ij j j 1 (34)
where
i = e i !i T
(35)
After some manipulation, it can be shown (see Appendix A) that the solution for the steady state
discrete time controllability grammian for exible structures is given as follows:
Proposition 1 
ij2 [Qij ]11 [Qij ]21 
[Wc1 ]ij 
= 2 !i2 !j2 Re 1 ij + 2 ij (36)
where i; j = 1; . . . ; n and
[Qij ]11 = j fai aj + bi bj + j (bi aj ai bj )g (37)
[Qij ]21 = j f ai aj + bi bj + j (biaj + ai bj )g (38)
ij = ii j 1j (39)
ij = ii j 1 j (40)
" # " #

1 = 1 j 1 j
j 1 ; 2 = j 1 (41)
For the state space parameterized as in (15) to (17) the following results hold:
Corollary 1
Q ij ]11 Q ij ]21
 
1 [
[Wc1 ]ij = 2  Re 1  + 2  [
(42)
j ij ij

where i; j = 1; . . . ; n and
[Q ij ]11 = zj faij + dij + j (cij bij )g (43)
[Q ij ]21 = zj f aij + dij + j (cij + bij )g (44)
ij = izi j 1zj (45)
ij = izi j 1 zj (46)
i = e i T
(47)
In the above corollary, zi denotes the ith discrete eigenvalue while aij , bij , cij , and dij are the input
matrices dened by h i  
= ac ij dbij
 T
B~ B~ T = V 1 Bz BzT V ij (48)
ij ij ij

Note that Corollary 1 is an exact relationship.

7
3.3 Diagonal Dominance
The denominator scalars in (36) can be expanded as follows
   
ij = (1 i !i T )i (1 + j !j T )j + O(i2 ) (49)
ij (1 + i !i T )i (1 + j !j T )j + O(i2)
where i is the i-th discrete eigenvalue of the i-th 2-by-2 orthogonal matrix i . For the o-diagonal
block matrices where i 6= j
" #

ij

   i !i T (i + j ) + O(i )
=    ! T ( +  ) + O( 2 )
i j
2
ij i j i i i j i
" #

= i j
 (50)
i j

For the diagonal block matrices where i = j

" #
2i !i Ti + O(i2 )
 
ii =
ii i i i!iT (i + i ) + O(i2)

" #

= 2 jsin(
i !i Ti
!i T ) (51)
Figure 1 shows the undamped discrete eigenvalues and denominator scalars ij and ij in the
complex plane. For small damping, the eigenvalues lie close to the unit circle, i.e., multiplied by
the scalar i (see 35). Notice from gure 1 that if the system has distinct complex conjugate poles,
the vectors i and j will never be collinear if i 6= j so that ij 6= 0 and ij 6= 0. From (50) and
(51), note that only the denominator factor ii asymptotically approaches zero as the damping
ratio approaches zero. Since ai ; aj ; bi ; bj are constants, the numerator factors, [Qij ]11 and [Qij ]21,
in (36) are also constants. This means that the diagonal block matrices of the grammian, [Wc1 ]ii,
which contains the denominator factor ii , can be arbitrarily large as i ! 0 while the magnitude
of the o-diagonal block matrices, [Wc1 ]ij is xed. Thus the controllability grammian matrix for
discrete exible structures is diagonally dominanant. Consider only the block diagonal terms, for
i = j . (37) and (38) simplify to
[Qii ]11 = i (ai2 + bi2 ) 
= 2 !i2 i (1 cos(!i T )) (52)
[Qii ]21 = i (bi + jai )2 
= 2 !i (1 cos(!i T )) (53)
Using (52) and (53), the block diagonal grammian in (36) can be simplied to the form
" #

= ii (1 cos3 (!i T ))
2 1 1

[Wc1 ]ii  2 !i
i T
1
sin(1! T )
i
(54)
sin(! T )
i i T

Furthermore, only the diagonal elements of the block diagonal matrix are inversely proportional to
the damping so that the simplest approximation form can be written as follows
8
Proposition 2
[Wc1 ]ii 
= ci2 I2 2 (55)
2 2 (1 cos(!i T ))
where ci2 = 4  ii  !2 T
(56)
!
i i i

The rst term in (56) corresponds to the ith controllability grammian for the corresponding
continuous system. The term ii corresponds to the ith modal grammian for controllability.
Similarly for the state space parameterized as in (15) to (17), the diagonal dominance of [Wc1 ]ii
in (42) holds because it contains the denominator factor ii which can be arbitrarily large as i ! 0.
After some algebra, it follows that the block diagonal grammian in (42) can be expressed as
  
[Wc1 ]ii = 41i aii+i Tdii I22 + aii2biidii aii2+biidii
 
+ tan(1 i T) aii2biidii aii2biidii (57)
Furthermore, only the rst term in (57) is inversely proportional to damping so that the simplest
form of the approximation can be written as

Corollary 2
[Wc1 ]ii  a +d
= ii4 T ii I22 (58)
i

4 Observability Grammian
4.1 Denition
For the time interval (ko T; k1T ), the discrete time observability grammian, Wo (ko ; k1), is dened
by
Wo (ko ; k1 ) = PoT (k1 ko )Po (k1 ko ) (59)
where the discrete observability matrix is
C~
2 3

6
6 C A~
~ 7
7
Po(k1 ko) = 66 .. 7 (60)
4 . 7
5
C~ A~k1 ko 1

It can be shown that the above grammian satises

A~T Wo(ko; k1)A~ + C~T C~ = Wo(ko; k1)

+(k1; ko)T C~ T C~ (k1; ko) (61)

9
For asymtotically stable linear systems, the last term in (61) vanishes as k1 ! 1. This leads to
the steady-state discrete time observability grammian, Wo1 , which satises the Sylvester equation

A~T Wo1 A~ + C~T C~ = Wo1 (62)

4.2 Closed-Form Solution
~ B~ ) can be replaced by (A~T ; C~ T ) so that a set of 2-by-2
Analogous to the controllability case, (A;
Sylvester equations for the observability grammian satises

A~Ti [Wo1 ]ij A~j + [C~T C~ ]ij = [Wo1 ]ij (63)

where i; j = 1; . . . ; n, and [Wo1 ]ij is the (i; j )th 2-by-2 block of [Wo1 ]. With the same approach
as taken in Section 3.2, it can be shown (see Appendix B) after some algebra that the solution for
the steady state discrete time observability grammian for exible structures is given as:

Proposition 3 
1 [R ] [R ] 
[Wo1 ]ij 
= 2 Re 1 ij 22 + 2 ij 12 (64)
ij ij

where 
[Rij ]22 = j ij11 + ij22 j (ij21 ij12 ) (65)
n o
[Rij ]12 = j ij11 + ij22 j (ij21 + ij12) (66)

For the state space parameterized as in (15) to (17) analogous results hold. However, the
output matrix appears in a dierent form. The outer product of the output matrix for the (i; j )
block becomes " #

 
11  12
[C~ C~ ]ij = Ci Cj = 21 22
T T ij ij
(67)
ij ij

where

Ci = Cz [Re(vi); Im(vi)] (68)

ij11 = Re(vi)T CzT Cz Re(vj ) (69)
ij12 = Re(vi)T CzT Cz Im(vj ) (70)
ij21 = Im(vi )T CzT Cz Re(vj ) (71)
ij22 = Im(vi)T CzT Cz Im(vj ) (72)

10
Note the symmetry for i = j
ii21 = ii12 (73)
This dierent form of the state and output matrix leads to the following result for the (i; j ) block
of the observability grammian.

Corollary 3 
[R ] [R ] 
[Wo1 ]ij = 21 Re 1 ij 22 + 2 ij 12 (74)
j ij ij

where 
[R ij ]22 = zj ij11 + ij22 j (ij21 ij12 ) (75)
n o
[Rij ]12 = zj ij11 + ij22 j (ij21 + ij12) (76)

Note that the above corollary is an exact relationship and is very similar in form to the approxi-
mation in Proposition 3.
4.3 Diagonal Dominance
The diagonal dominance argument for the observability grammian is similar to the controllability
case. From (50) and (51), note that only the denominator factor ii asymptotically goes to zero as
the damping ratio approaches zero. Since the terms ijkl are xed constants, the numerator factors,
[Rij ]22 and [Rij ]12, in (64) will also be xed constants. This means that the diagonal block matrices
of the grammian, [Wo1 ]ii , which contains the denominator factor ii , can be made arbitrarily large
as i ! 0 while the o-diagonal block matrices, [Wc1 ]ij will not. This represents the diagonal
dominance property of the observability grammian for discrete exible structures. Therefore,
consider only the block diagonal terms. After some algebra, the block diagonal observability
grammian in (64) can be reduced to the form
  
[Wo1 ]ii = 41 iii !+i Tii I2 2 + ii2ii12ii ii112+ii ii22
11 22 11 22 12

 
1 2  12
ii
+ tan(!i T ) ii11 + ii22  11
ii +  22
ii (77)
2ii12
Furthermore, only the rst term in (77) is inversely proportional to damping so that the simplest
form of the approximation can be written as follows

Proposition 4
[Wo1 ]ii 
= oi2 I22 (78)

11
 ^2
oi2 = (ii4 +! Tii ) = 4i!  T1
11 22
(79)
i i i i

The rst term in (79) corresponds to the ith observability grammian for the corresponding contin-
uous system.
Similarly for the state space parameterized as in (15) to (17), the diagonal dominance of [Wc1 ]ii
in (74) holds because it contains the denominator factor ii which can be arbitrarily large as i ! 0.
After some algebra, it follows that the block diagonal grammian in (74) can be expressed as
11 22 
[Wo1 ]ii = 21 ii + ii1 I2 2
 i  i
i 
 !
+  i
1
2 (80)
 4i 2  2i cos(2 i T ) + 1
2
1
where

1 = ( ii11 + ii22 )(2i cos(2 i T) 1) + 2ii12  2i sin(2 i T) (81)

2 = ( ii11 + ii22 )2i sin(2 i T) 2ii12 (2i cos(2 i T) 1) (82)
Furthermore, only the rst term in (80) is inversely proportional to damping so that the simplest
form can be written as the approximation below.
Corollary 4
[Wo1 ]ii  11 + 22
= ii4 T ii I22 (83)
i

5 Hankel Singular Values

Due to the diagonal dominance property of the discrete controllability and observability grammian
for exible structures, the square of the ith Hankel singular value follows from Propositions 2 and
4:
Proposition 5

i4 
= oi2 ci2
= 1 cos( !iT ) b bT (cT c + !2cT c ) (84)
8!i i T 2 i i di di i ri ri
6 2

Similarly, for the state space parameterized as in (15) to (17), Corollaries 2 and 4 lead to the
approximate Hankel singular values
Corollary 5
i4  (a + d )(11 + 22)
= ii (4ii Tii)2 ii I22 (85)
i

12
 Let the factors of deviations of the singular values of the discrete grammians from the singular
values of the continuous grammians (as given by (6)) be dened by the following for the ith mode:

ci = 2(1 !cos( !iT )) (86)

2T
i
1
oi = T (87)

In the limit when the sampling period approaches zero, the singular values of the scaled discrete
grammians converge to continuous values while the discrete Hankel singular value approaches the
Hankel singular value of the continuous system [5, 6, 7] as follows:

Proposition 6

1 = 1
lim c i (88)
T !0 T
oi  T = 1 (89)
!2

lim i4 = ii^i (90)

T !0 4i!i
where ii2 and ^i2 are dened by (7) and (8).

Note that without the sampling period scaling factor, the discrete controllability grammian ap-
proaches zero while the discrete observability grammian approaches innity. This result is consistent
with the earlier and more general result involving principal component analysis (see Proposition
7, [4]). In addition, the above convergence of the discrete to continuous Hankel singular values for
exible structures is analogous to the more general result (Proposition 8, [4]) where the singular
values of the discrete Hankel matrix converges to the corresponding singular values of the grammians
for the balanced system. For the state space parameterized as in (15) to (17), the Hankel singular
value dependence on the inverse square of the sampling period in Corollary 5 cancels with the
numerator factor (aii + dii) which is proportional to square of the sampling period as indicated
by (12) and (48). Indeed, similar results hold for the above type of parameterization in that the
controllability and observability grammians go to zero and innity respectively, with decreasing
sampling period.
The relationship between the discrete Hankel matrix Po Pc [4, 12] and the approximate formula
for the Hankel singular values 2 given in (84) is given below.

13
Proposition 7 Dene the SVDs Po = UooVoT , and Pc = UccVcT , then
PoPc = R(Wo Wc) 21 S T 
= R 2ST (91)
where R = Uo VoT , S = Vc UcT , and RT R = I = S T S
For comparison purposes with respect to the singular values of the continuous grammians, the
factors ci  T1 and oi  T are used. This additional sampling period factor makes the singular values
of the discrete grammian physically consistent with continuous singular values. Figure 2 shows
the eect of sampling on the singular values of the observability and controllability grammians
and the Hankel singular values as compared to the corresponding continuous singular values. At
high sampling rates (for instance !NY Q  :2), the predicted discrete singular values are close
!i

to the corresponding continuous singular values. Both the controllability and Hankel singular
values decrease with slower sampling rate. The exact discrete singular values are expected to
drop signicantly in the neighborhood of Nyquist frequencies. This singularity near Nyquist is not
predicted by the approximate analytical formula. In particular, the observability factor remains
constant which is counter intuitive and hence this approximation appears to fail near the Nyquist
frequency.

6 Example
To validate the analytical formula, the exact and approximate grammians are computed for a former
NASA experimental structure called the Control-Structures Interaction Evolutionary Model (CEM)
and is described in more detail in [8, 9]. A total of eight air thrusters are selected along with three
displacement sensors. The structural model consists of n2 = 12 modes whose rst six modes
are suspension modes. The frequencies are closely spaced and lightly damped, which is a typical
phenomenon for this kind of structure. Case 1 assumes 1 % damping ratio while case 2 assumes
5 % damping ratio for all modes. Note that a exible structure with 5 % damping ratios for all
modes (case 2) will not usually be considered as lightly damped. This signicant level of damping
is used for the purpose of evaluating the level of the approximation errors in the singular value
formulas.
Figure 3 shows the comparisons between the exact (29) and the approximate singular values
of the controllability (54,55) and observability (77,78) grammians and Hankel singular values (84).
14
The three rows of plots in gure 3 correspond to the sampling rates of

q = fN Y Qmaxmax
f
i fi
= 100; :1; :0001 (92)
i i

where maxi fi = !212 . The rst two rows from gure 3 representing normalized sampling rates of
q = 100 and :1, show that the approximate formula predicts the singular values accurately, up to
frequencies near 90 % of Nyquist frequency. However, the last row of plots (q = :0001) show a near
singular condition represented by a large drop in the smallest singular value with increased errors
in the remaining singular values. However, the last row corresponds to frequencies very close to
Nyquist i.e., q = :0001.
Figure 4 shows RMS error plots of the approximate diagonal singular values for both types
of approximations as a function of sampling rate, 2  fN Y Q . Each error of the singular value is
normalized by the corresponding exact value. The gure shows that the approximate formula
predicts quite accurately down to Nyquist frequencies that are only 10 percent higher than the
fastest mode. The normalized RMS error is dominated by errors in the smallest singular values
consistent with gure 3.
To evaluate the eect of larger damping ratios (case 2) in the approximate formulas for the
singular values at dierent sampling frequencies, gure 5 shows the comparisons between the exact
and the approximate singular values of the controllability and observability grammians and Hankel
singular values. The three rows of plots in gure 5 corresponds to the sampling rates in case 1. As
in the lighter damping case, the approximate formula predicts the singular values accurately, up to
frequencies near 90 % of Nyquist frequency. The last row of plots similarly shows a near singular
condition represented by a large drop in the smallest singular value with increased errors in the
remaining singular values.
Figure 6 shows the approximate diagonal singular values as a function of sampling rate. Figure
6 shows that the approximate formula predicts quite consistently down to Nyquist frequencies that
are only 10 percent higher than the fastest mode. The normalized RMS error is again dominated by
errors in the smallest singular values consistent with gure 5. The RMS error signicantly increases
with the ve fold increase in damping ratios. However, it is noted that the damping ratios for case
2 are too large to be considered a lightly damped exible structure.

15
7 Conclusions
The results complement earlier work on continuous time exible structure. For exible structures
modeled in discrete time, analytical expressions for singular values of controllability and observ-
ability grammian matrices and Hankel singular values are derived and validated through numerical
examples. For the class of exible structures with small damping and distinct frequencies, the above
formulae are signicantly simplied. It is found that the approximate formula is quite accurate
up to near Nyquist frequencies. The discrete Hankel singular values converges to the approximate
continuous formula with increased sampling rate. The simple but accurate approximate formula
could provide useful physical insights in the selection of actuators and sensors, model reduction,
and controller designs for exible structures modeled in discrete time.

8 Acknowledgement
The authors would like to thank Dr. Jer-Nan Juang of NASA Langley Research Center for
discussions on the signicance of the resulting formulas in system identication of exible structures.

A Proof of Proposition 1
By modal decomposition of the orthogonal 2 by 2 matrix in (21)

i = Xi iXiH (93)

where
i = diag(i ; i ) = diag(ej!di T ; e j!di T ) (94)
" #
1
Xi = p 1 1 = Xjj j (95)
2
(34) can be decomposed, after premultiplying by XiH and postmultiplying by Xj to obtain

iiXiH [Wc1 ]ij Xj XiH [Wc1 ]ij Xj j j 1

= XiH [B~ B~ T ]ij Xj j j 1 (96)

It follows that the four elements of the 2-by-2 matrix, [Wc1 ]ij , satisfy
" #
ij [W~ c1 ]11
ij ij [W~ c1 ]12
ij
= Qij (97)
ij [W~ c1 ]21
ij ij [W~ c1 ]22
ij

16
where
[Wc1 ]ij = Xi [W~ c1 ]ij XjH (98)
2 3
~ ~
[W~ c1 ]ij = 4 [W~ c1 ]ij21 [W~ c1 ]ij22 5
11 12
(99)
[Wc1 ]ij [Wc1 ]ij
Qij = XiH [B~ B~ T ]ij Xj j j 1 (100)
and ij and ij are dened by (39) and (40). For small damping, (22) can be used to simplify the
outer product [B~ B~ T ]ij appearing in (100) to
[B~ B~ T ]ij 
= [M~ BB T M~ T ]ij = Mi [BB T ]ij M T (101)
where M~ = blk-diag(M1 ; . . . ; Mn ). Using the expression Mi in (23) and [BB T ]ij where
" #

[BB ]ij = Bi Bj = 0 0
T T  ij 0
(102)
where ij2 is dened by (7), the expression in (101) can be expanded to
" #
2
T
= ! 2ij! 2 ab iaaj ab ibbj
[B~ B~ ]ij  (103)
i j i j i j

Using (94,95,103), Qij in (100) can be approximated as

" #
 
= 2! 2 ! 2 [[Q ij ]11 [Qij ]21
2
Qij  ij
Qij ]21 [Qij ]11 (104)
i j

where [Qij ]11 and [Qij ]21 are dened in (37) and (38). From (97), the 2 by 2 matrix [W~ c1 ]ij can
be written as
Qij ]11 1 [Qij ]21
" 1 #
  2 [
[W~ c1 ]ij = ij
2 !i !j
2 2

1
ij

[Qij ]21
1
ij
 [Qij ]11 (105)
ij 
ij

Finally, (36) is obtained from (98) and (105).

B Proof of Proposition 3
With the same approach as taken in the proof of Proposition 1, it can be shown that
" #
ij [W~ o1 ]11
ij ij [W~ o1 ]12
ij
= Rij (106)
ij [W~ o1 ]21
ij ij [W~ o1 ]22
ij

where
[Wo1 ]ij = Xi [W~ o1 ]ij XjH (107)
2 3
[ ~
W ] 11
[ ~
W ] 12
[W~ o1 ]ij = 4 ~ o1 ij21 ~ o1 ij22 5 (108)
[Wo1 ]ij [Wo1 ]ij
Rij = XiH [C~T C~ ]ij Xj j j 1 (109)

17
where ij and ij are given in (39) and (40). The output matrix product, [C~ T C~ ]ij can be written
as

[C~ T C~ ]ij = CTi Cj

" #
cTricrj c c
1 T
!j ri dj
= 1 cT c 1 T c c
!i di rj !i !j di dj
" #

= ij11 ij12 (110)

ij21 ij22
Note that for i = j , ii21 = ii12. For the special case of rate sensors only,

ij12 = ij21 = ij22 = 0; ij11 = cTricrj (111)

while for the case of displacement sensors only,

ij12 = ij21 = ij11 = 0; ij22 = c!di!cdj

T
(112)
i j

It can be shown that Rij in (109) can be approximated as

" #

= 2 [[R ij ]22 [Rij ]12
Rij  1 (113)
Rij ]12 [Rij ]22
where [Rij ]22 and [Rij ]12 are given by (65) and (66). Finally, after some algebra, [Wo1 ]ij given by
(64) is obtained.

References
[1] Brockett, R.W., Finite Dimensional Linear Systems, John Wiley & Sons, Inc, New York, 1970,
Chapter 2.

[2] Kwakernaak, H., and Sivan, R., Linear Optimal Control Systems, John Wiley & Sons, New
York, 1972, Chapter 1.

[3] Skelton, R.E., Dynamic Systems Control, John Wiley & Sons, Inc, New York, 1988, Chapter
5.

[4] Moore, B.C., \Principal Component Analysis in Linear Systems: Controllability, Observability,
and Model Reduction," IEEE Transactions on Automatic Control, vol. AC-26, No.1, 1981,
pp.17-32.

18
 [5] Gregory, C.Z., Jr., \Reduction of Large Flexible Spacecraft Models Using Internal Balancing
Theory," Journal of Guidance, Control, and Dynamics, vol.7, No.6, 1984, pp.725-732.

[6] Williams, T., \Closed Form Grammians and Model Reduction for Flexible Space Structures,"
IEEE Transactions on Automatic Control, vol. AC-35, 1990, pp.379-382.

[7] Gawronski, W., and Williams, T., \Model Reduction for Flexible Space Structures," Journal
of Guidance, Control, and Dynamics, vol.14, No.1, 1991, pp.68-76.

[8] Lim, K.B., and Gawronski, W., \Actuator and Sensor Placement for Control of Flexible
Structures," Control and Dynamics Systems, ed. C.T. Leondes, vol.57, Academic Press, San
Diego, 1993, pp.109-152.

[9] Gawronski, W., and Lim, K.B., \Balanced Actuator and Sensor Placement for Flexible
Structures," AIAA Paper # 95-3259, August 1995.

[10] Gawronski, W., Balanced Control of Flexible Structures, Springer-Verlag, London, 1995.

[11] Franklin, G.F., Powell, J.D., and Workman, M.L., Digital Control of Dynamic Systems,
Addison-Wesley Publishing Co., Reading, MA, 1992.

[12] Juang, J-N., Applied System Identication, Prentice Hall, Englewood Clis, New Jersey, 1994,
Chapter 5.

19
 List of Figures
Figure 1 Eigenvalues of exible structure

Figure 2 Eect of sampling: deviations from continuous singular values

Figure 3 Exact and approximate singular values of grammians for CEM structure; i = :01; rst
row: q = 100, second row: q = :1, third row: q = :0001

Figure 4 RMS error in approximate singular values for CEM structure; i = :01; Wc error: Eq.54
(solid), Eq.55 (dash); Wo error: Eq.77 (solid), Eq.78 (dash); HSV error: Eq.84 (solid)

Figure 5 Exact and approximate singular values of grammians for CEM structure; i = :05; rst
row: q = 100, second row: q = :1, third row: q = :0001

Figure 6 RMS error in approximate singular values for CEM structure; i = :05; Wc error: Eq.54
(solid), Eq.55 (dash); Wo error: Eq.77 (solid), Eq.78 (dash); HSV error: Eq.84 (solid)

20
 Im

λi ρij
λj
ω iT ωjT
Re
µ ij

λ∗i

Figure 1: Eigenvalues of exible structure

21
 oi * T
1

0.8

1
_
0.6 ci* T

o i* c i
0.4

0.2

0
0 0.5 1 1.5 2
f_i/f_NYQ

Figure 2: Eect of sampling: deviations from continuous singular values

22
 σ (Wc ) σ (Wo) HSV
+ = exact, o = Eq 54, x = Eq 55 + = exact, o = Eq 77, x = Eq 78 + = exact, o = Eq 84
0 6 2
10 10 10

−1 q=100 10
4 q=100 1
10 q=100
10
2 0
−2
10 10
10
0 −1
10 10
−3
10 −2 −2
10 10
−4
10 10
−4 −3
10
−5 −6 −4
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30

2 4 2
10 10 10
2 1
1 q=.1 10 q=.1 10 q=.1
10
0 0
0
10 10
10
−2 −1
10 10
−1
10 −4 −2
10 10
−2
10 10
−6 −3
10
−3 −8 −4
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30

2 5 2
10 10 10

q=.0001 q=.0001 0 q=.0001

0 0 10
10 10
−2
10
−2 −5
10 10
−4
10
−4 −10
10 10 −6
10

−6 −15 −8
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30
state number state number state number

Figure 3: Exact and approximate singular values of grammians for CEM structure; i = :01; rst row:
q = 100, second row: q = :1, third row: q = :0001

23
 5
10

RMS sv(Wc) error

0
10

−4 −2 0 2
10 10 10 10
5
10
RMS sv(Wo) error

0
10

−4 −2 0 2
10 10 10 10
5
10
RMS HSV error

0
10

−4 −2 0 2
10 10 10 10
(f_NYQ − f_MAX)/f_MAX

Figure 4: RMS error in approximate singular values for CEM structure; i = :01; Wc error: Eq.54 (solid),
Eq.55 (dash); Wo error: Eq.77 (solid), Eq.78 (dash); HSV error: Eq.84 (solid)

24
 σ (Wc ) σ (Wo) HSV
+ = exact, o = Eq 54, x = Eq 55 + = exact, o = Eq 77, x = Eq 78 + = exact, o = Eq 84
−1 6 2
10 10 10
1
−2 q=100 10
4 q=100 10 q=100
10
0
2 10
−3
10
10 10
−1
0
10 −2
10
−4 10
−2
10 −3
10
−5
10 10
−4
−4
10
−6 −6 −5
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30

1 4 2
10 10 10
1
2 10
0 q=.1 10 q=.1 q=.1
10
0
0 10
−1
10
10 10
−1
−2
10 −2
10
−2 10
−4
10 −3
10
−3
10 10
−6
−4
10
−4 −8 −5
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30

1 5 2
10 10 10
0
10 0
q=.0001 q=.0001 10 q=.0001
0
−1 10
10 −2
10
−2
10 −5 −4
−3
10 10
10
−6
−4 10
10 −10
10
−8
−5 10
10
−6 −15 −10
10 10 10
0 10 20 30 0 10 20 30 0 10 20 30
state number state number state number

Figure 5: Exact and approximate singular values of grammians for CEM structure; i = :05; rst row:
q = 100, second row: q = :1, third row: q = :0001

25
 5
10

RMS sv(Wc) error

0
10

−4 −2 0 2
10 10 10 10
5
10
RMS sv(Wo) error

0
10

−4 −2 0 2
10 10 10 10
5
10
RMS HSV error

0
10

−4 −2 0 2
10 10 10 10
(f_NYQ − f_MAX)/f_MAX

Figure 6: RMS error in approximate singular values for CEM structure; i = :05; Wc error: Eq.54 (solid),
Eq.55 (dash); Wo error: Eq.77 (solid), Eq.78 (dash); HSV error: Eq.84 (solid)

26