. ._____...il-_....---... -"......-..

"-
on Decision and Control
FP-14-2 4 ~ 4 0 Honolulu, Hawaii December 1990

Frequency Response of Linear Time Periodic Systems *

Norman M. Wereley and Steven R. Hall

Dept. of Aeronautics and Astronautics
Massachusetts Institute of Technology
77 Massachusetts Ave., Rm 33-105
Cambridge, MA 02139

Abstract
Let A ( t )be an n x n matrix whose elements are piecewise continuous
A frequency response notion comparable to the classical Bode gain functions of time with a finite number of discontinuities in t E [O.T].
and phase response for linear time invariant (LTI) systems has not Then, the n x n matrix @(t,r)is the state transition matriz [4]that
been developed for linear time periodic (LTP) systems. In this paper, satisfies the differential equation
fundamental input and output signal spaces are identified that lead
to a one-to-one map and a linear operator (transfer function). The d
,@(t,.) = A ( t ) @ ( t , ~ )@(t.t)
; =I (3)
LTP frequency response, including a characterization of gain, phase
and their directional properties, is then presented in terms familiar to
Since A ( . ) is periodic, then e(.,.)
is also periodic, so that
the multivariable LTI control theory.
sP(t f N T , r t N T ) = @ ( t , T ) (4)
Introduction This leads to the well known Floquet Theorem.
The fundamental notion behind Bode plots for LTI systems is that a
complex exponential (sinusoidal) test input signal at a given frequency Theorem 1 (Floquet) Consider the LTP state space model ( I ) ,
is mapped by the LTI transfer function into a complex exponential then:
output signal of the same frequency, but with possibly different am- S t a t e transition matrix. The state transition matrix of (1) can
plitude and phase. In contrast, if a complex exponential is input to a1zoay.s be expressed as
an LTP system, possibly several (or an infinite number of) harmonics
will appear in the output signal, each with possibly different gain and
phase. As a result, the notion of a transfer function for LTP systems
has been elusive. where P(t,to) is a nonsingular n x n T-periodic naatrir and Q is
This one-to-many map for complex exponentials is well understood U constant matrix.
and has been one motivation for the development of the describing
function (DF) [ l ]or equivalently, the harmonic balance technique [2] Similarity transformation. The state transformation
as applied to LTP systems. Both the DF and harmonic balance ap-
proaches enforce the one-to-one map, so that higher harmonics can be 4 4 = P ( t >t o ) w ( t ) (6)
neglected. However, this can lead to grossly inaccurate results. Several transforms x ( t ) into a periodically time varying system of cwrdi-
authors [1,3] have dealt with this problem by including as many har- nates, v ( t ) ) such that the dynamics matrix in the new state s p c e
monics as influence the fundamental harmonic. However, an infinite is time invariant:
number of harmonics may influence the fundamental. To date, no com-
parable notion t o the LTI transfer function has been presented for LTP + ( t ) = Q w ( t ) t B(t)u(t),
systems. Therefore, in this paper, a systematic approach is taken to y ( t ) = C(t)v(t)+ D(t)u(t). (7)
identify the linear operator that explicitly describes the input-output
relationship between signal spaces of fundamental importance to LTP where B ( t ) = P - l ( t , t o ) B ( t ) and C ( t ) = C ( t ) P ( t , t o ) .
systems.
Stability. An LTP system is stable iff all eigenvalues of the
+
monodromy matrix, @(to T, t o ) , lie on the open unit disk,
Linear Time Periodic Systems
{@(to t T , t o ) l E Do, (8)
LTP systems are described by a state space model of the form
where D o= ( 2 ; 121 < 1).
k ( t ) = A ( t ) z ( t )t B ( t ) u ( t )
Proof: The proof is straightforward. For details, see [ 5 ] . 0
Y ( t ) = C(t14t)f D(t)u(t) (1)
It will be assumed that to = 0 in the sequel, such that P ( t , t o ) =
where z ( t ) E R", y ( t ) E R", u ( t ) E R", and A ( t ) , B ( t ) , C ( t ) ,and P ( t ) . The Floquet stability theorem has played a pervasive role in thc
D ( t ) are real-valued matrices of appropriate dimension. A(.),B(.), analysis of LTP systems [2,G], and will be referred to repeatedly in the
C(.), and D ( . )are periodic with period T . That is, sequel.

A(t + N T ) = A ( t ) (2)
Fundamental Signal Spaces
and similarly for B ( . ) ,C(.), and D ( . ) . (In the sequel, the period
will be assumed to be T throughout, hence thc term T-periodic). For LTI systems the test input signal of interest is the complcs vs
The state space model ( 1 ) will be denoted by the &tuple S = ponential,
[ A ( l ) , B ( t ) , C ( t ) , D ( t ) and
] , the sct of all such systems with n finite u ( t ) = uOeSt, s E C,
will be denoted by Pmxm.The state space model S is strictly proper which is inappropriate for LTP systems because it results in ;I OIIV [ ( I
iff D ( t )= 0 . inany map [i]. However, Floquet theory can be viewed as the S(YIN.IIt ~ ~ i .
'Presented at The 29fh IEEE C'onfewzce on Decision and Conttd; Honolulu, signals that. increase geonzetrically from period to period. 'l'Iii> l~,.i,l:,
Hawaii: December 5-7, 1990 to the concept of a geometrically periodic test signal for L'I'I' S ~ ~ I R ~ I I I ~ ~

CH2917-3/90/0000-3650$1.OO @ 1990 IEEE 3650

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.
 Definition 2 (Geometrically periodic signals) A geometrically Proof: The proof is omitted for brevity. For details, see [5]. 0
periodic (GP) signal, u ( t ) , with fundamental frequency, w p , and cor- Note that the steady state response is a one-to-one map from G P
responding period T , has the property input signals t o GP output signals. Thus, the GP steady state output
response is analogous to the LTI transfer function.
u ( t + NT) = zNu(t), (10)
Definition 5 ( L T P integral operator representation) The
where z E C . 0 steady state output response can be expressed as

Since the LTP system maps a single harmonic into many harmonics,
then it makes sense that the test signal should include all harmonics as
well. Moreover, the G P signal can be expressed as a complex Fourier where the integml operator kernel G ( z ; t , r ) is defined QS

series of the periodic portion of the GP signal, modulated by a complex G ( z ; t , r ) = C ( t ) @ ( T , O ) [ z l -@(T,O)]-l@(T,r)B(r)

exponential signal, which leads to the concept of the exponentially
modulated periodic test signal. 0; t<r
%C(t)E(t); t = r (17)
Definition 3 (Exponentially modulated periodic signal) A
C(t)@(t,T)E(r); t > r
(complex) exponentially modulated periodic (EAfP) signal can be ex- The integml operator defined by (16-17) will be denoted by G ( z ) SO
pressed as the complex Fourier series of a periodic signal of frequency, that (16) can be expressed in a more compact form,
wpr modulated b y a complex exponential signal,
y ( t ) = G(z)u(t). (18)
u(t)= u,esnt (11) G ( z ) is the integral o p e r a t o r transfer function. 0
nEZ
Note that G ( z ; t , r ) has been defined, for t = T, as the average of
where t 2 0,sn = s t jnwpr and s E C. 0
the integral operator kernel across the discontinuity. Since this is a set
of measure zero, it does not affect the value of the integral. However,
A strong analogy will be proposed at this point. GP (or EAIP) G ( z ; t , t ) is chosen such that correct answers are obtained in certain
signals are to LTP systems, what complex exponentials (sinusoids) are limiting procedures not discussed here [5]. The integral operator G ( z )
to LTI systems. One difficulty introduced by the EMP signal is the plays a similar role in the study of LTP systems to the transfer function
infinite number of Fourier coefficients required to describe an arbitrary matrix in the LTI control theory.
periodic signal. This is a drawback that is inherent to the analysis of The state space model appropriate to the description of LTP systems
LTP systems and cannot be avoided. Clearly, bounded GP (or EMP) evolving from period to period i s defined.
signals are Lz[O,T] signals, and have bounded energy. In addition, an
orthonormal basis for G P signals in &[O, T ] consists of sinusoids that Definition 6 (Integral operator state space model) A n L T P
are harmonics of the fundamental or pump frequency, wp. system (1) can be represented b y an integral operator state space
model of the form

Integral Operator Approach

Two closely related representations of an LTP system will be in-
where * k E Cn,and 'lik E LT[O,T] defined over the subinterval k a8
troduced. The first representation is based on an integral operator
approach. 'iik = u ( t - hT), k
E 2, t > kT. (20)
T h e o r e m 4 (LTP system response t o G P signals) Consider the
Ilere: A : C" -+ C", B : Ly[O,T] -+ C", C : Cn -+ LT[O,T],
state space model of an asy9~iptoticallystable LTP system in (1). If the
input to the LTP system is a G P signal, where i is not an eigenvalue
fJ : Ly[O,T] -+ LT[O,T]. The operators are defined below:
of the monodromy matrax, then the state respnse consists of a GP A = qi(T,O),
steady state response
Blik = L T @ ( t , r ) E ( r ) u ( r ) d r ,

z s s ( t ) = @i(t,o)[zr- @(t,o)l-llT@ ( t , r p ( r ) u ( r ) d r c = C(t)qi(t,O),

+Jd@(t,r)E(r)u(r)dr: t > 7- (12)

The integnd operator state space model will be denoted by the 4-tuple
and Q transient response of the form S = [A, B, C, D] = [@(T,O),B, C, D]. 0

Using the above definition, the integral operator can be expressed in

z t , ( t ) = @(t,O)
{ (0 - [zl- @(LO)]-' l T @ ( t , r ) B ( r ) u ( r ) d r(13)
} a compa,ct form.
L e m m a 7 (Integral o p e r a t o r transfer function) The integml
which vanishes as t --+ W. Also, the GP s t e a d y state o u t p u t re- opemtor transfer function, G(z ) , explicitly describes the relationship
sponse is given b y
btween the input, u(t)E L y [ O , T ] and the output, y ( t ) E L;[O,T], as

yss(l) = C ( t ) @ ( t , O ) [ d- @(f,O)]-' I T @ ( t . r ) B ( r ) u ( r ) d r G(i) = C ( 2 1 - A ) B t b (22)

+C(t) Jdl r)d+

@(t,T)B(T)u( ~ D ( t ) u ( t ) ; t > T (14)
Proof: The above expression follows directly from Definition 6.
Although the notation, A, has been introduced for the monodromy
0

matrix, the notation @(T,0) will be retained for clarity.

and the transient o u t p u t response is gzven by The integral operator approach was useful in the development of
a generalized Nyquist criterion for LTP systems [5,8], and is useful
primarily in an analytical context. However, the integral operator
approach does not lead to useful numerical procedures, so that alter-
(is) natives must be developed.

3651

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.
 Harmonic Balance Approach Expanding (1) in terms of these Fourier series':
The time periodic parametric excitation associated with most phys-
ical systems can be expressed by a sum of sinusoids of relatively low
harmonic number. From an engineering standpoint, a sensible numer-
ical procedure would exploit this tendency, so that an approach based
on harmonic balance is developed [2]. Harmonic balance offers two ad-
vantages: (1)low frequency contributions to the parametric excitation
0 =
nEZ { yn -
mEZ
Cnqmxm-
mEZ 1
D n - m ~ mcant.

Now, the complex exponentials, {ej'"+* 1 n E Z } , form an orthonormal

can be more easily captured than with a time discretization approach, basis in Lz[O,T], so that the terms enclosed by braces must vanish.
(2) time periodic parametric excitation and EMP signals are naturally This procedure is referred to as the principle of harmonic balance.
described by the complex Fourier series. Here, the frequency response Hence, the two equations below hold V n E 2:
(or the steady state forced response) of the LTP state space model (1)
to an EMP signal (11) is determined. snxn = An-mxm t Bn-mum,
mEZ mEZ

T h e o r e m 8 ( L T P system response t o EMP signals) Consider ~n = Cn-mxm t Dn--mum. (32)

the L T P state space model i n (l), which is assumed to be asymptoti- mEZ mEZ
cally stable. The T-periodic matrix B ( t ) can be expanded i n a complex Although the above equations are a concise representation of the
Fourier series, input-output relationship between the Fourier coefficients of the input
B ( r )= B , e J l W p T (23) and output signals, manipulating summations can be tedious. There-
IEZ fore, a Toeplitz form notation will be utilized.
and similarly for C ( t ) and D ( t ) . If the input to the L T P system above Definition 9 (Harmonic state space model) The system of equa-
is an E M P signal (11), then the state response consists of an EhJP tions (32) can be expressed as the infinite dimensional matriz equation,
steady s t a t e forced response,
sx = ( A - N ) X t B u ,
z s s ( t )= Pn...i ( s J - Q)-' B ~ - ~ u ~ e ' " ~ , (24) Y = cxtvu. (33)
n,l,mEZ
Define doubly infinite vectors representing the harmonics of the state,
and a transient s t a t e response,
[
XT = . .,X T , , X T I , X : 0 T , I T , x:, . . .1 3 (34)
ztr(t)

which vanishes as t
= +(t,O)
{ €0 -
I,mEZ
- Q)-'Bi-mUm

CO. The steady state o u t p u t response is an

--f
I 3 (25) and simiIarEy for the control signal, U , and output signal, Y . The T -
periodic dynamics matrix, A(t),is expressed in terms of its harmon-
ics, {A,ln E Z } , as a doubly infinite block Toeplitz matriz called a
Toeplitz f o r m [lo],
EMP signal of the form

(26) A= 7 (35)
and the transient o u t p u t response is given by

with a similar definition for t3 in terms of { B n l n E Z } , C in terms

Proof: The proof is omitted for brevity. For details, see [5] 0 of {C,ln E Z } , and V in terms of { D J n E 2). Finally, define the
Thus, an LTP system induces a one-to-one map from EMP inputs modulatzon frequency matrix,
to EMP outputs at steady state. Hence, the LTP frequency response N = blkdiag{jnwpI}; V n E 2. (36)
notion can be stated: at steady state, an asymptotically stable L T P
This infinite dimensional matrix equation (33) is called the harmonic
system maps an E M P input signal, to an E M P output signal of the
state space model, denoted by the 4-tuple S = [(A-N), B, C, V]. 0
same frequency, but with possibly different amplitude and phase (as
long as s is not an eigenvalue of Q). The amplitude and phase refers The harmonic state space model is similar to Toeplitz systems [ll].
to the amplitude and phase of all the harmonics in the input and output In fact, if the modulation frequency matrix is set to zero, N = 0. then
signals. This spatial dependency will be quantified in the sequel. the harmonic state space model is identical to the Toeplitz system.
The steady state response can also be obtained by direct application Unfortunately, the Toeplitz system results are not directly applicable
of harmonic balance. Consider the LTP state space model (1). The to the LTP case since the harmonic state space model is qunsi-Toqdi/:,
dynamics matrix can be expanded in a complex Fourier series Also, the harmonic state space model has more structure tliaii t l t r
general Toeplitz system. The Toeplitz forms in the harmonic s i ; ~ t r
A(t)= AmejmWpt, (28) space model are Hermitian, and since the harmonics associatrd \villi
mEZ
a given system matrix generally grow s m d as the harmonic tiiiiiilwr
and similarly for B(t),C ( t ) , D ( t ) . An EMP test signal, u ( t ) , implies grows large, the Toeplitz forms are also effectively bandcd. Iloiwwr.
that the steady state response is also an EMP signal, this additional structure is not exploited by the Toeplitz systctii rosiilis
in [Ill.
L e m m a 10 (Harmonic transfer functions) II%r linriiiollir
transfer function (HTF),@s), describes the inyul-oct/plt/ 11 / < t f i o t i -
ship between the harmonics of the input signal, and ilroct of //It , ~ 1 J / 1 ~ t , i
signal, such that
and that the steady state output, y ( t ) , is an EMP signal, Y = C(S)U, (.IT\
where
O(s) = C [sZ- ( A -NI]-' 6 t P. (3s)

3652

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.
Proof: Eliminating x from (33), in favor of Y and U, results in the Principal Gains and Directions
desired input-output relationship. 0
However, there are two problems associated with the harmonic trans- Directional properties of LTP systems can be generalized as any
fer function. First, it is not dear that the harmonic transfer function, property of the integral operator transfer function (harmonic transfer
which requires the inversion of a doubly infinite matrix, will always function) or GP (EMP) signals that exhibits a spatial dependency, in
exist. Second, the harmonic transfer function described above is a addition to the frequency dependency shared with scalar LTI trans-
doubly infinite matrix operator, which cannot be implemented on the fer functions and scalar sinusoidal signals. This spatial dependency
computer. The first problem will be dealt with, in general, by ap- is manifested in the infinite number of Fourier coefficients required
plication of the second Floquet result. The second problem will be to characterize EMP signals, as well as the multi-input multi-output
mitigated by truncating the HTF in order to implement analyses on nature of general LTP systems.
the computer. Clearly, the one-to-one map induced by GP signals, that is, the inte-
Let us consider some examples. Consider the LTP state space model gral operator transfer function, is linear since the underlying dynamics
with time invariant dynamics, S = [Q,B ( t ) , C ( t ) ,D ( t ) ] . Here, Q , is were linear. Thus, the singular value decomposition of the integral
a constant matrix, BO that A = blkdiag{Q}. Following the procedure operator transfer function will provide useful interpretations of do-
outlined above yields main and range spaces, and directions of maximal amplification for
this linear map. The singular values of the integral operator, and their
associated directions, are well defined [9], and are parameterized by z
on the unit circle in the z plane. However, the integral operator ap-
proach provides little insight into how to compute the singular values
The HTF is the terms in braces: since the integral operator is a functional operating on vector functions
(signals).
IC2
On the other hand, the harmonic balance approach involves the mul-
tiplication of complex matrices and vectors and can be easily be im-
Direct application of harmonic balance produces the same result as in plemented on the computer. The harmonic transfer function maps an
Theorem 8. EMP input into an EMP output'accordiug to
A large class of problems that are important in the study of LTP sys-
tems are those represented by LTI plants with input or output ampli-
C ( s ) : U -+ y ; C ( s ) E pmxm (45)
tude modulations. The HTF for both these cases simplifies markedly.
Consider the state space model of an amplitude modulated input to that is, the plant has m independent inputs and m independent out-
an LTI plant, S = [ Q , B ( t ) , C , D ( t ) ] .The input-output relationship puts. The input signal ~ ( tis )an EMP signal, and provided the internal
(39)can be further simplified, dynamics represented by A ( t ) are asymptotically stable (the eigenval-
ues of the monodromy matrix are on the unit disk, or the eigenvalues
of Q are in the LHP), the steady state output signal, y ( t ) , will also be
an EMP signal.
Hence, the HTF is the terms in braces: The harmonic transfer function is a complex matrix that changes
value with frequency, w , where the frequency range of interest is given
Bm,n(s) = C ( s m I - Q)-' Bm-n + Dm-n, (42) br
Instead, consider the case of an amplitude modula.ted output signal
from an LTI plant, S = [Q, B , C ( t ) , D ( t ) ] ,then the HTF is
that is, the unit circle in the z-plane is mapped to the imaginary axis
Em,n(s) = Cm-n ( s n l - ~ 1 - Bl + D ~ - , , . (43) in the fundamental strip in the s-plane. For any value of frequency,
w 6 Ro, the singular value decomposition (SVD) of the HTF function
In the latter two cases, no infinite sums need to be computed. can be computed,
Two different forms of the HTF were presented. The first form of the
HTF in Theorem 8, corresponds to the harmonic state space model S p(jw) = U(jw)C(jw)V*(jw), (47)
= [[Q-.+')E,, I?, D],where (Q-NI is a block diagonal quasi-Toeplitz
form. The second form of the HTF in Lemma 10, corresponds to the where the superscript * denotes the Hermitian or complex conjugate.
harmonic state space model S = [(A-N),8, E , D],where ( A - N ) a Each quantity in the SVD is parameterized by frequency.
full quasi-Toeplitz form. However, both representations are related by Since the harmonic transfer function is infinite dimensional, the sin-
a similarity transformation. gular values will be studied by examining the singular values of the
truncation of the HTF. Thus, N positive harmonics (as well as N
Lemma 11 ( A similarity transformation) Consider the LTP negative harmonics and the zeroth harnionic so that the HTF is sym-
d a t e s p c r model, S = [ A ( t ) , B ( t ) , C ( t ) , D ( t ) ] .Its harmonic state metric) are included and the truncated HTF is denoted by C ~ ( j w ) .
spct model, denoted by the 4-tuple SI=[(A-N), B,C,'D], where (A- The implication is that if enough harmonics are included in the trun-
.t') is U full quosi-Tocplitt form. The similarity tmnsformation given cated HTF, then correct answers will be obtained in SVD analyses.
by x = P v , where P is the Toeplitt form corresponding to the periodic Assume that P- d ~"
i v ) is invertible. Then:
+ +
I

prtion I J ~thr: Flmpet solution, P ( t ) , transforms the above harmonic the quantity C,(jw) is a m ( 2 N 1)x m(2N 1)complex matrix
htlltf: q c e 1nfJdd lo u hormonic state spnce model, denoted by the 4- consisting of the singular values of the truncated HTF:
/ I I ~ / I s2
: = [(Q- N ) , s , C , D ] , where ( 9 - N ) is n doubly infinite block
ijia!lurrel i i i d r i x . Ilcrc: u m a s ( ~ )= al(w) 2 Q Z ( W ) 2 *..2 ~ Z N + I ( W=)umtn(u). (48)
Q - N = P-*(A-N)P, The minimmi singular value umin(w) is the gain associated with
n = P-lB, the minimum amplification direction. However if more harmonics
c = CP. (44) are included in the analysis, a smaller minimum singular value
may be found. These singular values are known as the principal
I'l.f,t,fi lI,l l l l . l , i l i l h , M'l' [SI. 0 gains.
'/ ~ , i ~ ~ ~ i l ;~~,~I ;i ,~i.~, ~ y l ~ , ~ iliiis ~ ~ .virtiie
~ i ;tlio i i ~ i i that it is an algebraic
hijlijI;fit,ii,y ~ , I ; ~ I I ~ I , I , I I ;i Io~i i , itisl,wil til' ;I. t . i i i w varying one requiring the +
V,(jw) is a complex matrix m(2N 1 ) x m ( 2 N 1) whose col- +
i ~ (~) ll) ll, ,
~ ~ ~ ~ l is,f t it11 ~ ! i i I ~ ~ r ~ . i i i i ~I,IIV
i ~ . siitiil;lrii,y
~~l,~, f.ransformation {
umn vectors a ( " ) ( ~ )are
} the principal input directions or right
r(!liijirt*h I 1114 I i ~ I i i ~ i It ~. ~~ ~~ I I ~ ~ I I S' l~hOyI lIi ;h~ I'oriii~ 7'. singular vectors of p N ( j W ) and form a basis for the dontain spuce.

3653

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.
(c) +
U ~ ( j wis) a complex matrix m(2N 1) x m(2N 1) whose col- +
umn vectors { u ( " ) ( w ) } are the principal output directions or left
singular vectors of c ~ ( j w and ) form a basis for the range space.

This singular value analysis is carried out for w E 00, so that the
principal gains can be plotted versus frequency as a Bode gain plot
(that is, decibels versus frequency, w E 00). This principal gain di-
agmm is analogous in many respects to the singular value plot for
multivariable LTI systems, although the specific interpretation must
be carefully worked. Also, the principal gain diagram repeats itself in
the nth complementary strip in the s-plane, that is,
..
e -10
- 5 3
At each frequency it is assumed that the input to the asymptotically
stable system is a unit complex periodic signal of the form
-15
lL(q= unej(w+nwP)t. (50)
nEZ

where l l ~ ~ l l Then
= 1. ~ assuming that the system is in steady state, -291 -0.5 0 0.5 7
the resulting output will also be a complex periodic signal,
FREQUENCY IN RADIANSBEC
Figure 1: Principal gain diagram for lossy Mathieu equation example.
The gain of the LTP system at a given frequency is characterized by the
where principal gains (singular values) of the harmonic transfer function. The
b N l h = ((d~V(jw)uN((~. (52) principal gain diagram is a graph of these singular values over the frequency
Consider a maximum direction analysis at a specified frequency range w E (-wp/2,wp/2]. The shaded circle on the maximum principal
gain locus corresponds to the frequency of the maximum direction analysis
w E 00.The SVD produces a maximum singular value, # ( U ) , with
shown in Figure 2.
a corresponding complex valued right singular vector, ?(U), with
+
m ( 2 N 1) elements, and left singular vector, E ( w ) with m(2N 1) +-
elements. The maximum right and left singular vectors correspond The Lossy Mathieu Equation
to the first columns of V ~ ( j wand
) U,(jw), respectively. The block
vector elements of *(U) can be expressed in polar form as Here, the time periodic dynamics of the Lossy Mathieu equation [7]
are considered. Defining the state vector as
(53)
Here, the o denotes the Schur product or an element by element vec- .'=[e e], (57)
tor multiplication. The quantities 6, and qn are parameterized by
leads t o the system matrices:
frequency, although this frequency dependence will be implicit to sim-
plify notation. Also, n denotes the block vector of ? ( w ) corresponding 0
to the nth harmonic. Each block vector element of G(w) contribute; a A(t) =
different amplitude and phase. The block vector elements of E ( w ) can
also be written in polar form as

(54) c = [ l 11
Again, the quantities 2, and 6, depend implicitly on frequency.
At steady state, the input and output signals that correspond to The parameter values selected for this example are wp = 2, C = 0.2, and
the direction of maximum amplification can be reconstructed from the /3 = 0.2. For most of the analyses performed in this section, truncation
information contained in the maximum direction analysis. First, the of the harmonic transfer function t o N = 10 harmonics was sufficient
to obtain reasonable results.
input signal corresponding to the maximum amplification direction can
be reconstructed from the quantities ii, and qnfor all n E 2 as below: The principal gain diagram of this system is shown in Figure 1.
There are an infinite number of principal gain loci corresponding to the
E(t)= 6, 0 , j A e j ( W + n w P ) t . infinite dimensional domain and range spaces of the harmonic transfer
(55) function, and only a small number are shown in this diagram. The
nEZ
largest gains occur for the zeroth, and first harmonics, which can be
The steady state output signal corresponding to the maximum ampli- concluded by examination of the singular vectors associated with each
fication direction, @(t), can be reconstructed from a, iin, and 4,,: gain plot.
A maximum amplification analysis is performed for w = wp/4 =
g(t> = 3 6, 0 &nej(w+nwp)t. (56) 0.5rad/sec as indicated by the shaded circle in Figure 1. The max-
nEZ imal input (principal) direction is given by the right singular vector
associated with the maximum singular value,
Finally, the amplitude and phase of any periodic input to the LTP
system at a frequency w E Ro can be expressed as a linear combina- r - TT--
. . -
tion of the principal input directions (right singular vect,ors) at that
frequency. The amplitude and phase of the corresponding output can 0.8227 - j0.1088
be predicted at steady state by the same linear combination of the c(t)= &I2 +
0.3497 j0.0572 1
principal output directions (left singular vectors) weighted by their +
0.3319 j0.2692 ejwpt
corresponding principal gains (singular values).

3654

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.
 a ) MAXIMUM INPUT DIRECTION
It is instructive to express the input direction in polar form, I
2

d
“ 0
2
- -2
I I
0 2 4 6 8 10 12 14 18 18 20
TIME IN INTEGER FUNOAMENTAL PERIODS

Note that all of the harmonics enter with a specific phase. The max- b) MAXIMUM OUTPUT DIRECTION
i
imum principal gain is the maximum singular value, 8 = 1.1798. The
maximum output (principal) direction is given by the left singular vec-
tor associated with the maximum singular value,

I I
0.5594 - j0.562G e-jwpt 0 2 4 6 8 10 12 14 18 18 20
+
ij(t) = 1.1798ejtI2 0.1664 j0.0646 1 . TIME IN INTEGER FUNOAMENTAL PERIODS
0.2647 - j0.4809 ejwpt
c) ZERO INITIAL CONDITION RESPONSE TO MAXIMUM INPUT DIRECTION
I

In polar form,

I
0 2 4 6 8 10 12 14 18 18 20
TIME IN INEGER FUNDAMENTAL PERIODS

Figure 2: Maximum direction analysis for lossy Mathieu equation exam-

ple. The directions of maximum amplification are the input and output
Note that the direction associated with the maximum principal gain directions associated with the largest principal gain (singular value) of the
consists predominantly of three harmonics, and that each harmonic harmonic transfer function at a particular frequency.
has a corresponding associated phase change. Thus, even for single-
input single-output LTP systems, phase is a directional property of
the harmonic transfer function.
[3] W. Szemplibska-Stupnicka, “The generalized harmonic balance
The maximum input direction is shown in Figure 2a, and the cor-
method for determining the combination resonance in the para-
responding output direction is shown in Figure 2b, both as predicted metric dynamic systems,” Journal of Sound and Vibration, vol. 58,
from the singular value analysis. The input signal in Figure 2a was
no. 3, pp. 347-361,1978.
used as the input to a simulation of the LTP system with zero ini-
tial conditions. The resulting output signal is shown in Figure 2c. [4] L. A. Zadeh and C. A. Desoer, Linear System Theory: The State
At steady state, the simulated system output and the output as pre- Space Approach. McGraw-Hill, 1963.
dicted from the singular value analysis coincide, illustrating the LTP [5] N. M. Wereley, Analysis and Control of Linear Periodically Time
interpretation of the Bode frequency response. Varying Systems. PhD thesis, Dept. of Aeronautics and Astro-
nautics, M.I.T., 1990. To appear Fall 1990.
Summary [GI P. Friedmann, C. E. Hammond, and Tze-Hsin Woo, “Efficient
numerical treatment of periodic systems with application to sta-
A frequency response for linear time periodic systems was developed bility problems,” Inter. J. for Numerical Methods in Engineering,
that exploited the one-to-one map induced by geometrically periodic vol. 11, no. 7, pp. 1117-1136, 1977.
signals. This map was described by an integral operator, based on the [7] J . A. Richards, Analysis of Periodically Time Varying Sys-
GP test input, and a generalized harmonic balance approach, based tems. Communications and Control Engineering Series, Springer-
on an EMP input. The singular values or principal gains of the LTP Verlag, 1983.
operator were discussed and the LTP principal gain diagram described.
Directional properties of the LTP operator were discussed, and notions [8] S. R. Hall and N. M. Wereley, “Generalized Nyquist stability cri-
of the domain and range spaces are presented. terion for linear time periodic systems,” in Proceedings of the 1990
The framework of linear operators described in this paper have lead American Control Conference, pp. 1518-1525, 1990.
to the development of a comprehensive open loop analysis theory for [9]J. A. Cochran, The Analysis of Linear Integml Equations.
linear time periodic systems, including a characterization of poles, McGraw-Hill, 1972.
transmission zeroes and their directional properties [5]. A general- [lo] U. Grenander and G. Szego, Toepiitz Forms. Chelsea Publishing
ized Nyquist criterion has also been developed using this framework, Co., second (textually unaltered) ed., 1958. Reprinted 1984.
as well as application of stability robustness and small gain notions
to LTP systems. This linear operator framework has also led to the [ll] J. E. Wall Jr., Control and Estimation for Large-scale Systems
development of a comprehensive frequency domain interpretation for Having Spatial Symmetry. PhD thesis, Dept. of Electrical Engi-
LTP systems. neering and Computer Science, M.I.T., 1978.

References
[l] A. Leonhard, “The describing function method a.pplied for the
investigation of parametric oscillations,” in Proceedings of the 2nd
IFAC World Congress, (Basle, Switzerhnd), pp. 21-28, 19G3.
12) J. Dugundji and J. H. Wendell, “Some analysis methods for ro-
tating systems with periodic coefficients,” AlAA Journal, vol. 21,
pp. 890-897, June 1983.

3655

Authorized licensed use limited to: Peking University. Downloaded on December 05,2021 at 07:15:54 UTC from IEEE Xplore. Restrictions apply.