NUMERICAL LINEAR ALGEBRA WITH APPLICATIONS

Numer. Linear Algebra Appl. 2008; 15:809–835

Published online 21 August 2008 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/nla.604

On solving periodic Riccati equations

A. Varga∗, †
German Aerospace Center, DLR-Oberpfaffenhofen, Institute of Robotics and Mechatronics,
D-82234 Wessling, Germany

SUMMARY
Numerically reliable algorithms to compute the periodic non-negative definite stabilizing solutions of the
periodic differential Riccati equation (PRDE) and discrete-time periodic Riccati equation (DPRE) are
proposed. For the numerical solution of PRDEs, a new multiple shooting-type algorithm is developed to
compute the periodic solutions in an arbitrary number of time moments within one period by employing
suitable discretizations of the continuous-time problems. In contrast to single shooting periodic generator
methods, the multiple shooting-type methods have the main advantage of being able to address problems
with larger periods. Three methods are discussed to solve DPREs. Two of the methods represent extensions
of the periodic QZ algorithm to non-square periodic pairs, whereas the third method represents an extension
of a quotient-product swapping and collapsing ‘fast’ algorithm. All proposed approaches are completely
general, being applicable to periodic Riccati equations with time-varying dimensions as well as with
singular weighting matrices. Copyright q 2008 John Wiley & Sons, Ltd.

Received 17 March 2008; Accepted 18 March 2008

KEY WORDS: periodic systems; periodic Riccati equation; periodic Lyapunov equation; periodic deadbeat
control; numerical methods

1. INTRODUCTION

In this paper we consider the numerical solution of the periodic Riccati differential equation
(PRDE):

− Ẋ (t) = AT (t)X (t)+ X (t)A(t)+C T (t)C(t)− X (t)B(t)B T (t)X (t) (1)

∗ Correspondence to: A. Varga, German Aerospace Center, DLR-Oberpfaffenhofen, Institute of Robotics and
Mechatronics, D-82234 Wessling, Germany.
†
E-mail: [email protected]

Contract/grant sponsor: Swedish Strategic Research Foundation

Copyright q 2008 John Wiley & Sons, Ltd.

810 A. VARGA

where A(t), B(t), C(t) are n ×n, n ×m, p ×n, and m ×m real T -periodic matrices (i.e. A(t + T ) =
A(t), B(t + T ) = B(t), C(t + T ) = C(t), ∀t) and integrable over [0, T ]. The PRDE plays a major
role in solving the optimal periodic linear-quadratic (PLQ) control problem [1] for the linear
continuous-time periodic system
ẋ(t) = A(t)x(t)+ B(t)u(t)
(2)
y(t) = C(t)x(t)
with the associated quadratic cost functional

  
1 ∞ Q(t) S(t) x(t)
J= [x T (t) u T (t)] dt (3)
2 0 S T (t) R(t) u(t)

where Q(t) = C T (t)C(t), S(t) is an n ×m T -periodic matrix, and R(t) = R T (t) > 0 is an
m ×m T -periodic matrix. Usually it is assumed that Q(t)− S T (t)R −1 (t)S(t)
0. Minimizing J
to find the optimal control u ∗ (t) can be recast into solving, for the stabilizing positive-
semidefinite T -periodic solution X (t), a PRDE of form (1) with A(t), B(t), and C(t) replaced
 = A(t)− B(t)R −1 (t)S(t), 
by A(t) 
B(t) and C(t), respectively, where   are defined
B(t) and C(t)
  T −1
by the full-rank Cholesky-like factorizations B(t) B (t) = B(t)R (t)B (t) and C
T T (t)C(t)
 =
−1 ∗
Q(t)− S (t)R (t)S(t). The optimal periodic state-feedback control law u (t) = F(t)x(t) with
T

F(t) = −R −1 (t)(X (t)B(t)+ S(t))T (4)

minimizes the performance index (3) and ensures that A(t)+ B(t)F(t) is asymptotically stable.
We will also be concerned with the numerical solution of the discrete-time periodic Riccati
equation (DPRE):

X k = CkT Ck + ATk X k+1 Ak −(ATk X k+1 Bk + Sk )(Rk + BkT X k+1 Bk )−1 (ATk X k+1 Bk + Sk )T (5)

where Ak ∈ Rn k+1 ×n k , Bk ∈ Rn k+1 ×m k , Ck ∈ R pk ×n k , Rk = RkT ∈ Rm k ×m k , and Sk ∈ Rn k ×m k are

N -periodic matrices (i.e. Ak+N = Ak , Bk+N = Bk , Ck+N = Ck , Rk+N = Rk , Sk+N = Sk , ∀k).
Equation (5) arises, for example, when solving the PLQ optimal control problem [1] for the linear
discrete-time periodic system
xk+1 = Ak xk + Bk u k
(6)
yk = Ck xk
by minimizing the quadratic cost functional
  
1 ∞ Qk Sk xk
J= [xk u k ] T
T T
(7)
2 k=0 Sk Rk uk

where Q k = CkT Ck and usually it is assumed that

 
Q k Sk

0
SkT Rk

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 811

Provided a non-negative N -periodic stabilizing solution X k of the DPRE (5) is known, the periodic
state-feedback matrix Fk in the optimal control law u ∗k = Fk xk , which minimizes the performance
index (7), results as

Fk = −(BkT X k+1 Bk + Rk )−1 (ATk X k+1 Bk + Sk )T (8)

The dual equations

Ẋ (t) = A(t)X (t)+ X (t)AT (t)+ B(t)B T (t)− X (t)C T (t)C(t)X (t) (9)

and

X k+1 = Bk BkT + Ak X k ATk −(Ak X k CkT + Sk )(Rk +Ck X k CkT )−1 (Ak X k CkT + Sk )T (10)

arise in the optimal filtering problems [1, 2]. A particular case of PRDE corresponding to the case
when B(t) ≡ 0 is the periodic Lyapunov differential equation (PLDE):

− Ẋ (t) = AT (t)X (t)+ X (t)A(t)+ Q(t) (11)

where Q(·) is symmetrical and T -periodic. This equation together with its dual

Ẋ (t) = A(t)X (t)+ X (t)AT (t)+ Q(t) (12)

have many important applications in the analysis of reachability/observability of linear continuous-

time periodic systems [3], in solving periodic stabilization problems [4], in computing Hankel-
and H2 -norms of periodic systems [2, 5], or in solving PRDEs by employing Newtons’
method [1].
This paper deals with the efficient and numerically reliable computation of the unique symmetric
stabilizing periodic solutions of the PRDE and DPRE. For the numerical solution of the dual
equations (9) and (10), the same methods can be applied to appropriate dual reformulations. The
paper relies on the preliminary results reported in [5] for solving PRDEs and in [6] for solving
DPREs.
To compute the T -periodic solution X (t) of Equation (1), traditionally the periodic generator
method is used (see, for example, [7]). Essentially, this single shooting-type method computes
an initial condition matrix X 0 (also called a periodic generator) such that the integration of the
ordinary differential equation (ODE) (1) with the initial value X (0) = X 0 leads to a T -periodic
solution X (t) satisfying X (T ) = X 0 . Unfortunately, this approach is known to be highly sensitive
due to the ill-conditioning of the underlying Hamiltonian-ODE problem to be solved (due to the
intrinsic unstable dynamics), especially in the case of large periods. Therefore, in this paper we
propose a numerically enhanced multiple shooting-type approach based on employing discretization
techniques. The main appeal of this approach is that the periodic solution X (t) is computed
simultaneously in many time moments within one period, so that the numerical difficulties related
to numerical integrations for large periods and/or unstable dynamics are highly alleviated. The
solution for intermediate values of time t can be computed using special numerical integration
by initializing the solution in the nearest time moment. The key numerical element for solving

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
812 A. VARGA

the discrete-time problem is the ordered periodic real Schur form (PRSF) of a symplectic matrix
product together with the computation of the corresponding periodic invariant bases of the stable
periodic invariant subspace [8, 9].
The solution of DPRE for constant dimensions has been considered by several authors [8–10],
but DPREs with time-varying dimensions have been considered only recently in [11]. A standard
assumption in all proposed algorithms is the invertibility of Rk . The methods proposed in [8, 9]
rely on the periodic QZ decomposition to compute orthogonal bases for suitable periodic deflating
subspaces from which the solution results. The main numerical element for these methods is the
computation of an ordered periodic QZ decomposition of a pair of periodic square matrices with
constant dimensions. Algorithms to perform this decomposition have been proposed in [8, 9], but
presently there is no robust numerical software implementing these methods for pairs of real
periodic matrices [12].
The methods proposed in [10, 11] belong to the family of ‘fast’ methods, by reducing the
problem to a single discrete-time algebraic Riccati equation (DARE) satisfied say by X 1 = X N +1 .
This DARE can then be solved using standard deflating subspace methods. The rest of the solution
X k for k = N , N −1, . . . , 2, can be obtained by a convergent direct iteration on (5) [13]. The
quotient-product swapping and collapsing method of [10, 14] has been reinterpreted to perform
locally orthogonal compressions on an extended regular pencil to obtain a subpencil whose stable
deflating subspace generates the stabilizing solution at a certain time moment [15]. This method
performs substantially less operations than the periodic QZ-decomposition-based approach, and
this justifies the term ‘fast’ used to label it [15]. The method of [11] performs essentially the same
reduction of the extended pencil by using structure preserving non-orthogonal block compressions.
In the resulting final subpencil, the ‘matrices’ of a standard problem can be identified, which define
the solution at a certain time moment. To compute the solution, a structure-preserving iterative
doubling algorithm with quadratic convergence is employed in [11], but any alternative approach
for standard systems can be employed as well.
In this paper, three extensions of the above-described methods are proposed, which can address
the most general case of time-varying dimensions and singular weighting matrices Rk . DPREs
with possibly time-varying dimensions are to be expected when solving spectral and inner–outer
factorization problems for periodic systems using deflation-based approaches similar to those
proposed in [16] for standard systems. Problems with singular Rk arise when solving, for example,
deadbeat control problems via an PLQ-optimization-based approach [17]. All three proposed
methods are able to address problems with varying number of null characteristic multipliers of
Ak . The first two methods extend the periodic QZ algorithm to handle non-square periodic pairs
with time-varying dimensions. To apply the standard periodic QZ algorithm, a certain periodic
pair is preprocessed by first deflating its infinite characteristic multipliers and then isolating the
core finite characteristic multipliers. The first method performs an explicit finite–infinite separation
of characteristic spectra, whereas the second method employs dedicated deflation techniques to
separate the trivial zero and infinite characteristic multipliers originating from the time-varying
dimensions. The third method performs structure exploiting orthogonal transformations to reduce
an extended regular pencil to a regular subpencil, whose deflating subspaces produce the solution of
DPRE at a fixed time moment. This method extends the quotient-product swapping and collapsing
approach of [10] along the lines of the reduction technique employed in [18]. The main advantage
of this approach is that its implementation is straightforward using available robust numerical
software. For the ‘fast’ method, reliable software implementation is already available in a recently
developed Periodic Systems Toolbox [19].

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 813

2. SOLUTION OF PRDE

In this section we address the computation of the periodic stabilizing solution of the PRDE (1)
and of the periodic solution of the PLDE (11).

2.1. Existence results

First, we introduce some basic concepts for continuous-time periodic systems. Let
A (t,
) denote
the transition matrix corresponding to A(t) satisfying

*
A (t,
)
= A(t)
A (t,
),
A (
,
) = I (13)
*t

For a T -periodic A(t),  A (
) :=
A (
+ T,
) is called the monodromy matrix corresponding to
A(t), and its eigenvalues, which are independent of
, are called characteristic multipliers. We say
that A(t) is a stable periodic matrix if all its characteristic multipliers have magnitudes less than 1.
To provide the main existence result, we recall the notions of reachability/stabilizability and
observability/detectability of periodic systems (see, for example, [1]). A characteristic multiplier
 of A(t) is said to be unreachable if TA (0)x = x, x  = 0, imply that B T (t)
TA (0, t)x = 0, almost
everywhere for t ∈ [0, T ]. A characteristic multiplier  of A(t) is said to be reachable if it is not
unreachable. System (2) is said to be stabilizable if all characteristic multipliers  with ||
1 are
reachable. Similarly, a characteristic multiplier  of A(t) is said to be unobservable if  A (0)x = x
and x  = 0 imply that C(t)
A (t, 0)x = 0, almost everywhere for t ∈ [0, T ]. A characteristic multiplier
 of A(t) is said to be observable if it is not unobservable. System (2) is said to be detectable if
all characteristic multipliers  with ||
1 are observable.
A solution X (t) of (1) is called a stabilizing solution if A(t)− B(t)B T (t)X (t) is stable. The
following result from [20] gives necessary and sufficient conditions for the existence of a unique
symmetric positive-semidefinite stabilizing solution:

Theorem 1
The PRDE (1) admits a unique T -periodic stabilizing solution X (t) = X T (t)
0 if and only if the
pair (A(t), B(t)) is stabilizable and the pair (A(t), C(t)) is detectable.

A less restrictive assumption guarantees the existence of a (possibly non-unique) stabilizing

symmetric positive-semidefinite stabilizing solution [1]:

Theorem 2
The PRDE (1) admits a T -periodic stabilizing solution X (t) = X T (t)
0 if and only if the pair
(A(t), B(t)) is stabilizable and no unit-modulus characteristic multiplier of A(t) is unobservable.

This theorem is applicable in the important case C(t) ≡ 0.

If B(t) ≡ 0, the PRDE becomes a PLDE for which the existence result has a particularly simple
form:

Theorem 3
The PLDE (11) admits a unique T -periodic solution X (t) if and only if A(t) does not have
reciprocal characteristic multipliers.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
814 A. VARGA

In what follows, we will tacitly assume that the appropriate existence conditions are always
fulfilled when solving particular PRDEs.

2.2. Periodic generator method

Let H (t) be the Hamiltonian matrix corresponding to the PRDE (1):
 
A(t) −B(t)B T (t)
H (t) =
−C T (t)C(t) −AT (t)

We have J H (t)+ H T (t)J = 0, where

 
O I
J=
−I O
thus, H (t) is indeed a Hamiltonian matrix. Theorems 1 and 2 guarantee that the monodromy matrix
 H (0) is symplectic (i.e. TH (0)J  H (0) = J ), with 2n eigenvalues appearing in pairs of the form
(i , 1/i ), for i = 1, . . . , n. This property is the key aspect of the following approach to solve the
PRDE (1) (see [7]):
1. Compute the symplectic monodromy matrix  H (0).
2. Compute orthogonal Z to reduce  H (0) to an ordered real Schur form (RSF) such that
 
11 12
Z  H (0)Z =
T
(14)
O 22
where 11 has n eigenvalues inside the unit circle and 22 has n eigenvalues outside the
unit circle.
3. Partition Z into n ×n blocks
 
Z 11 Z 12
Z=
Z 21 Z 22
and integrate from t = 0 to t = T the linear matrix differential equation:
 
Z 11
Ṡ(t) = H (t)S(t), S(0) =
Z 21
From the conformably partitioned solution
 
S1 (t)
S(t) =
S2 (t)

compute X (t) = S2 (t)S1−1 (t).

Alternatively, it is possible to determine the periodic generator X 0 := X (0) by solving the
algebraic Riccati equation [3]:
21 +22 X 0 − X 0 11 − X 0 12 X 0 = 0

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 815

where
 
11 12
 H (0) =
21 22
Then, the solution on [0, T ] can be computed by numerical integration of the PRDE (9).
Note that the same approaches can be employed to solve the dual PRDE (9) by defining
 
−AT (t) B(t)B T (t)
H (t) =
C T (t)C(t) A(t)
To solve the PLDE (11), we determine the periodic generator X 0 := X (0) by solving the standard
discrete-time Lyapunov equation satisfied by X (T )(= X (0)) [3]
X 0 = TA (0)X 0  A (0)+ W (T, 0) (15)
where

tf
W (t f , t) :=
TA (
, t)Q(
)
A (
, t) d
(16)
t
and then integrate (11) backward in time from t = T to t = 0.
The periodic generator approach to solve PRDEs is potentially numerically unreliable because it
involves the numerical integration of one or two ODEs with unstable dynamics: the first to compute
 H (0) and the second to compute S(t). Therefore, for large periods, this approach will almost
certainly fail because of the uncontrolled accumulation of roundoff errors. The second approach
will encounter similar difficulties for large periods because of the need to compute  H (0).

2.3. Multiple shooting method to solve PRDE

To alleviate the numerical difficulties related to the periodic generator method, we propose an
alternative approach that relies on determining the transition matrix
H (T, 0) in a product form
(also recommended in [21])

H (T, 0) =
H (T, T −) . . .
H (2, )
H (, 0) (17)
where  = T /N for a suitably chosen integer period N . We denote Hk :=
H (k, (k −1)) for
k = 1, 2, . . . , which is obviously an N -periodic matrix. Using the algorithm of [8], we can determine
an orthogonal N -periodic matrix Z k to reduce Hk to an ordered PRSF such that
 
Jk;11 Jk;12
Z k+1 Hk Z k = (18)
O Jk;22
where 11 := J N ;11 . . . J2;11 J1;11 has n eigenvalues inside the unit circle and 22 := J N ;22 . . .
J2;22 J1;22 has n eigenvalues outside the unit circle. Since  H (0) = H N . . . H2 H1 , it follows that
Z 1T  H (0)Z 1 is in the ordered RSF (14), where both 11 and 22 are in RSF, and Z is simply
Z 1 . If we partition Z k into n ×n blocks as
 
Z k;11 Z k;12
Zk = (19)
Z k;21 Z k;22

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
816 A. VARGA

we obtain the solution of PRDE at t = (k −1) as

−1
X k := X ((k −1)) = Z k;21 Z k;11

Some computational aspects are relevant for a robust implementation of this approach. To
compute Hk =
H (k, (k −1)), we have to integrate the ODE

Ẏ (t) = H (t)Y (t), Y ((k −1)) = I

for t ∈ [k, (k −1)] to obtain Hk = Y (k). Since each matrix Hk is symplectic, it is impor-
tant to employ numerical integrators that are able to guarantee this property. Note that standard
methods (even the simple explicit Euler method) do not ensure that Hk will be symplectic. Among
methods able to compute symplectic solutions are the Gauss–Legendre (diagonal Padé approxi-
mants) methods, which belong to the class of symplectic Runge–Kutta methods [7, 22]. MATLAB
software for symplectic integration is freely available [23].
We can also apply the fast algorithm for DPRE (see the following section) by observing that
the solution X (t) at two successive time steps (k −1) and k is related as [3]

X k = (X k+1 Hk,12 − Hk,22 )−1 (Hk,21 − X k+1 Hk,11 ) (20)

where Hk is partitioned into n ×n blocks

 
Hk,11 Hk,12
Hk =
Hk,21 Hk,22

Rewriting the above expression of X k , we obtain

   
I I
Hk = (Hk,11 + Hk,12 X k )
Xk X k+1

which, with Mk := Hk and L k := I , is in form (24), which underlies the methods discussed in the
following section for solving DPRE. If Z 1 , computed by the fast algorithm, is partitioned into form
−1
(19), then X 1 = Z 1,21 Z 1,11 . The rest of the solution X k for k = N , . . . , 2 is computed iteratively
using (20), with X N +1 := X 1 . The main advantage of this method is its ease of implementation,
since no sophisticated methods (such as algorithms to compute ordered PRSF) are involved.
To compute the values of the solution X (t) between two discretization moments t0 = (k −1)
and t f = k, special ODE solvers as those proposed in [24, 25] can be used to integrate (1) in
backward time with X (t f ) = X (k) or (9) in forward time with X (t0 ) = X ((k −1)). A distinctive
feature of solvers discussed in [25] is their capability to preserve the positivity of the numerical
solution of differential Riccati equations.
The computational approach involves as preprocessing and postprocessing steps the numerical
integration of the underlying PRDE on small time intervals. These parts of computations are
usually the most computer intensive and we assumed tacitly that existing standard or symplectic
integration techniques can be used to perform them. Fortunately, it appears that these computations

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 817

are ‘embarrassingly’ parallelizable, which makes the multiple shooting approach very appealing
on parallel machines.

2.4. Multiple shooting method to solve PLDE

To solve the PLDE (11) using a multiple shooting-type approach, we observe that the solution
X (t) at time moments t and t − is related as [3]
X (t −) =
TA (t, t −)X (t)
A (t, t −)+ W (t, t −)
where W (t, t −) is defined in (16). Thus, the solution at successive time moments (k −1) and
k satisfies
X k = FkT X k+1 Fk + Wk (21)
where Fk :=
A (k, (k −1)) and Wk := W (k, (k −1)). By imposing X N +1 = X 1 , the N -coupled
equations in (21) for k = 1, . . . , N represent a discrete-time backward periodic Lyapunov equation.
To solve the periodic discrete-time Lyapunov equation (21), a numerically reliable method,
such as that proposed in [26], can be used. This method is based on computing the PRSF of
the periodic matrix Fk . For the computation of PRSF, numerically stable algorithms have been
proposed in [8, 9] and robust numerical software implementations are available in the recently
developed Periodic Systems Toolbox for MATLAB [19]. For the solution of the periodic Lyapunov
equation (21), robust numerical software is also available in this toolbox. A similar approach can
be used to solve the dual PLDE (12).
The computation of Fk and Wk for k = 1, . . . , N in (21) can be done using numerical integration of
appropriate ODEs. To compute Fk , the matrix differential equation (13) must be integrated from
=
(k −1) to k using appropriate methods for ODEs. Since the time step  can be chosen arbitrarily
small, the numerical integration even for unstable A(t) will not raise any numerical difficulties.
To compute Wk observe that for given t f , Y (t) := W (t f , t) in (16) satisfies the Lyapunov
differential equation:
−Ẏ (t) = A(t)Y (t)+Y (t)AT (t)+ Q(t), Y (t f ) = 0 (22)
Thus, Wk can be computed as Wk = Y ((k −1)) by integrating the above equation backward in
time from t f = k to t0 = (k −1). To integrate the PLDE (22) it is important to use methods
that preserve the symmetry of the solution and, if appropriate, also its positive definiteness. For
example, methods that are able to preserve positive definiteness have been proposed in [25].

3. SOLUTION OF DPRE

In this section we address the computation of the periodic stabilizing solution X k of the DPRE
(5). Note, however, that any method to solve the DPRE (5) can also be employed to solve the dual
equation (10) by solving the DPRE (5) with the replacements:
Ak ← ATN −k , Bk ← C NT −k , Ck ← B NT −k , Sk ← S NT −k , Rk ← R N −k
From the computed solution X k of the DPRE, the solution of the dual equation is recovered with
the replacements:
X k ← X N −k+2 , k = 2, . . . , N

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
818 A. VARGA

3.1. Existence results

First, we introduce some basic concepts for discrete-time periodic systems that are suitable for
problems with time-varying dimensions [27]. The transition matrix between time moments i and
j is defined as
A ( j, i) := A j−1 . . . Ai+1 Ai , with
A (i, i) := Ini . For an N -periodic Ak ,  A ( j) :=

A ( j + N , j) is called the monodromy matrix at time j corresponding to Ak and its eigenvalues
are called characteristic multipliers at time j. The spectrum (
A ( j + N , j)) contains at least
n j −n zero elements, where n := mink {n k }. The rest of n eigenvalues are independent of j and
form the core characteristic multipliers [28]. We say that Ak is a stable periodic matrix if all its
characteristic multipliers have magnitudes less than 1.
The notions of reachability/stabilizability and observability/detectability of periodic systems
can be defined in a very similar manner to the continuous-time case [27]. A characteristic multiplier
 of Ak is said to be unreachable at time
if TA (
)x = x and x  = 0 imply that B Tj−1
TA (
, j)x = 0,
for j =
, . . . ,
+ N −1. A characteristic multiplier  of Ak is said to be reachable if it is not
unreachable. System (6) is said to be stabilizable if all characteristic multipliers  with ||
1
are reachable. Similarly, a characteristic multiplier  of Ak is said to be unobservable at time

if  A (
)x = x and x  = 0 imply that C j
A ( j,
)x = 0, for j =
, . . . ,
+ N −1. A characteristic
multiplier  of Ak is said to be observable if it is not unobservable. System (6) is said to be
detectable if all characteristic multipliers  with ||
1 are observable.
A solution X k of (5) is called a stabilizing solution if the corresponding feedback (8) ensures that
Ak + Bk Fk is stable. The following results give necessary and sufficient conditions for the existence
of symmetric positive-semidefinite stabilizing solutions for problems with constant dimensions.
Since the proofs in [1], based on a standard lifting technique, cover formally the case of time-
varying dimensions, these results also apply to this more general case. The following result is the
discrete-time counterpart of Theorem 1 [13].

Theorem 4
The DPRE (5) admits a unique N -periodic stabilizing solution X k = X kT
0 if and only if the pair
(Ak , Bk ) is stabilizable and the pair (Ak , Ck ) is detectable.

A less restrictive assumption guarantees the existence of a (possibly non-unique) stabilizing

symmetric positive-semidefinite stabilizing solution [1]:

Theorem 5
The DPRE (5) admits an N -periodic stabilizing solution X k = X kT
0 if and only if the pair (Ak , Bk )
is stabilizable and no unit-modulus characteristic multiplier of Ak is unobservable.

Note that this theorem is also applicable in the important case Ck = 0, k = 1, . . . , N .

3.2. Extended periodic QZ algorithm

Consider the periodic matrix pairs:
⎡ ⎤ ⎡ ⎤
Ak O Bk In k+1 O O
⎢ ⎥ ⎢ ⎥
Mk = ⎢
⎣ −Q k In k −Sk ⎥
⎦, Lk = ⎢
⎣ O ATk O⎥
⎦ (23)
SkT O Rk O −BkT O

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 819

where Mk ∈ R(n k +n k+1 +m k )×(2n k +m k ) and L k ∈ R(n k +n k+1 +m k )×(2n k+1 +m k+1 ) . With these pairs, the
DPRE (5) together with (8) can be expressed as
⎡ ⎤ ⎡ ⎤
In k In k+1
⎢ ⎥ ⎢ ⎥
Mk ⎣ X k ⎦ = L k ⎣ X k+1 ⎦ k (24)
Fk Fk+1
for k := Ak + Bk Fk ∈ Rn k+1 ×n k . This alternative form of the DPRE leads to a straightforward
extension of the methods of [9] to compute the stabilizing periodic solution X k :
1. Compute the N -periodic matrices Vk and Z k such that for k = 1, . . . , N

Hk,11 Hk,12
Vk Mk Z k =
O Hk,22

Tk,11 Tk,12
Vk L k Z k+1 =
O Tk,22
where Hk,11 ∈ Rn k+1 ×n k , Tk,11 ∈ Rn k+1 ×n k+1 is invertible, and all eigenvalues of T −1 H (1)
11 11
are in the unit circle.
2. Partition the rows of Z k to conform with the column partitioning of matrix Mk and partition
the columns of Z k to conform with the column partitioning of Vk Mk Z k
⎡ ⎤
Z k,11 Z k,12
⎢ ⎥
Z k = ⎣ Z k,21 Z k,22 ⎦
Z k,31 Z k,32
−1 −1
If all Z k,11 are invertible, compute X k = Z k,21 Z k,11 and Fk = Z k,31 Z k,11 .
Note that in the case of time-varying dimensions, the direct application of the periodic QZ
algorithm at Step 1 is not possible. It is shown in what follows how to overcome this difficulty with
recently developed algorithms to reduce periodic matrix pairs to orthogonally similar canonical
forms.
The computation of Vk and Z k at Step 1 can be performed in three main steps. In the first
step, a finite–infinite spectral separation is performed by using the recently developed algorithm to
compute Kronecker-like forms of periodic pairs [29]. A specialization of this algorithm to perform
only the finite–infinite spectral separation is given in Appendix A. By applying this algorithm to
the dual periodic pair ( M k , 
L k ), where
k = M NT −k+1 ,
M k = 1, . . . , N , 
L k = L TN −k , k = 1, . . . , N −1, 
L N = L TN
and performing suitable row and column permutations (see [29] for details), orthogonal matrices
Vk1 and Z k1 can be determined such that
 
f f,∞
Mk Mk
Vk Mk Z k =
1 1
O Mk∞
 
f f,∞
Lk Lk
Vk L k Z k+1 =
1 1
O L∞k

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
820 A. VARGA

where L k and Mk∞ are non-singular matrices. The pair (Mk , L k ) contains the finite charac-
f f f

teristic multipliers, whereas the pair (Mk∞ , L ∞ k ) contains the infinite characteristic multipliers.
Note that the finite characteristic multipliers at time k are the eigenvalues of the product
(L k+N −1 )−1 Mk+N −1 . . . (L k )−1 Mk and the nilpotent matrix product (Mk∞ )−1 L ∞ ∞ −1
f f f f
k . . . (Mk+N −1 )
∞
L k+N −1 characterizes the infinite characteristic multipliers. The problem is not solvable if the
above separation cannot be performed because the associated lifted pencil (see Section 3.4) is not
regular. Equivalently, the resulting periodic Kronecker-like form contains parts that correspond to
a left or right Kronecker structure.
In the second step, orthogonal matrices Vk2 and Z k2 are determined such that
⎡ f f
⎤
Mk,11 Mk,12
Z k2 = ⎣ ⎦
f
Vk2 Mk
f
O Mk,22
⎡ f f
⎤
L k,11 L k,11
=⎣ ⎦
f
Vk2 L k 2
Z k+1
f
O L k,22

f f
where the pair (Mk,22 , L k,22 ) has constant dimensions and is in a periodic generalized Hessenberg
form [9], L k,11 ∈ R(n k+1 −n)×(n k+1 −n) is upper triangular and Mk,11 ∈ R(n k+1 −n)×(n k −n) is upper
f f

f f
trapezoidal, with n = min n k . The pair (Mk,11 , L k,11 ) has only characteristic multipliers equal to
zero. This reduction can be performed by extending the generalized periodic Hessenberg reduction
procedures of [8, 9] to the case of non-constant dimensions similarly as done in [30] for the periodic
Hessenberg form.
Finally, orthogonal matrices Vk3 and Z k3 are computed such that
 
f Mks Mks,u
Vk3 Mk,22 Z k3 =
O Mku
 
f L sk L s,u
k
Vk3 L k,22 Z k+1
3
=
O L uk

where the pair (Mks , L sk ) has only stable characteristic multipliers, and the pair (Mku , L uk ) has only
unstable characteristic multipliers. Because of the symplectic nature of the eigenvalue problem,
the non-zero eigenvalues must appear in reciprocal pairs. For this step, the algorithms of [8, 9] to
compute and reorder the periodic QZ decomposition can be used.
The final transformation matrix Z k is obtained as

Z k = Z k1 diag(I, Z k2 ) diag(I, Z k3 )

To compute X k and Fk , the accumulation of the left transformations Vk is not necessary.

Each computational step of the above algorithm is numerically stable and has a computational
complexity of O(N (2n +m)3 ), where n and m are the maximum problem dimensions for n k and
m k , respectively.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 821

3.3. Periodic QZ algorithm with infinite structure deflation

An approach similar to that proposed for standard systems in [31] can be used to deflate trivial
infinite characteristic multipliers first. For this purpose, orthogonal Vk1 are chosen such that
⎡ ⎤ ⎡O⎤
Bk
⎢ ⎥ ⎢O⎥
Vk1 ⎣ −Sk ⎦ = ⎢
⎣ ⎦
⎥
Rk Rk

where all R k ∈ Rm k ×m k are non-singular. Note that this is a necessary condition for the existence
of a solution of the DPRE. Let us define the reduced periodic pair (M k , L k ) from
   
Mk O Lk O
Vk Mk =
1
, Vk L k =
1
∗ Rk ∗ O

where M k ∈ R(n k +n k+1 )×2n k , L k ∈ R(n k +n k+1 )×2n k+1 .

Then the N -periodic orthogonal transformation matrices Vk2 and Z k2 can be computed (see
Appendix B) such that for k = 1, . . . , N the resulting matrices can be partitioned as follows:
⎡ 0 ⎤
Mk ∗ ∗
⎢ ⎥
Vk2 M k Z k2 = ⎢⎣ O Mk
c
∗ ⎥⎦
∞
O O Mk
⎡ 0 ⎤
Lk ∗ ∗
⎢ ⎥
Vk2 L k Z k+1
2
=⎢ ⎣ O Lk
c
∗ ⎥⎦
∞
O O Lk

where Mk0 ∈ R(n k+1 −n)×(n k −n) , Mkc ∈ R2n×2n , Mk∞ ∈ R(n k −n)×(n k −n) , L 0k ∈ R(n k+1 −n)×(n k+1 −n) , L ck ∈
(n k −n)×(n k+1 −n)
R2n×2n , L ∞
k ∈R , and Mk∞ and L 0k are invertible. The periodic pair (Mk0 , L 0k ) contains
the trivial zero characteristic values of the pair (M k , L k ), whereas the periodic pair (Mk∞ , L ∞ k )
contains the trivial infinite characteristic values. At each time moment, the numbers of zero and
infinite characteristic values are equal. As shown in the previous subsection, their number depends,
however, on the starting time moment used to define the period.
The periodic pair (Mkc , L ck ) contains the 2n core characteristic values of the original pair (M k , L k )
and the characteristic value spectrum is symmetric with respect to the unit circle. We can employ
the standard periodic QZ algorithm for constant dimensions [8, 9] to separate the stable and unstable
characteristic value spectra. We obtain with appropriate orthogonal transformation matrices Vk3
and Z k3
 s   s 
Mk ∗ Lk ∗
Vk Mk Z k =
3 c 3
, Vk L k Z k+1 =
3 c 3
(25)
O Mku O L sk

where Mks ∈ Rn×n , L sk ∈ Rn×n is invertible, and ((L s )−1 M s (1)) ⊂ C− .

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
822 A. VARGA

Finally, let us partition Z k := Z k2 diag(In k −n , Z k3 , In k −n ) such that its first n k columns correspond
to the stable spectrum (i.e. the trivial zero characteristic values and the stable core characteristic
values) of the pair (M k , L k )
 
Z k,11 Z k,12
Zk =
Z k,21 Z k,22

and Z k,11 is square. Provided all Z k,11 are invertible, the solution of the DPRE is given by
−1
X k = Z k,21 Z k,11

The state-feedback matrix Fk results from (8).

As for the previous method, all deflation steps as well as the QZ-decomposition step are
numerically stable and have a computational complexity of O(N (2n +m)3 ). The main advantage
of this second approach is that both in the initial deflation step and in the second separation step
no numerical rank decisions are necessary. In the case of problems with repeated characteristic
values at infinity (e.g. when solving DPRE for a system with pure delays), the final separation step
can be still enhanced by performing first the finite–infinite structure separation using the algorithm
of Appendix A. It is well known even in the case of constant systems that this approach allows
generally to separate finite–infinite structures more reliably in the presence of multiple infinite
eigenvalues than a pure QZ-method-based reordering.

3.4. Fast algorithm

This algorithm can be explained by defining the associated lifted pencil to the periodic pair
(Mk , L k )
⎡ ⎤
M1 −L 1 O ... O
⎢ ⎥
⎢ O M2 −L 2 ... O ⎥
⎢ ⎥
⎢ . . . . . ⎥
⎢
H − zT := ⎢ . . . . . . . . . ⎥ (26)
. ⎥
⎢ ⎥
⎢ ⎥
⎣ O M N −1 −L N −1 ⎦
−z L N O ... O MN
N
of order  = i=1 (2n i +m i ). Under the usual assumptions for the existence of the solution of the
DPRE, this pencil is regular and has no eigenvalues on the unit circle.
The proposed algorithm implicitly constructs a stable deflating subspace of the pencil (26) by
employing structure exploiting reductions similar to that employed in [18] for computing periodic
systems’ zeros. The method can be seen as an extension of the swapping and collapsing approach
proposed in [10, 14] of quotient-products (see [15]). The basic reduction is performed as follows.
Consider the (n 1 +2n 2 +n 3 +m 1 +m 2 )th order orthogonal transformation matrix U1 compressing
the rows of the matrix
   
−L 1 R1
to
M2 O

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 823

where R1 is a non-singular matrix of order 2n 2 +m 2 . Applying U1 to the first two block rows of
H − zT we obtain for the non-zero elements
   
M1 −L 1 O M1 R1 − L1
U1 =
O M2 −L 2 M2 O − L2

which defines the new matrices M 2 and 

L 2 with n 1 +n 3 +m 1 rows.
Then construct the (n 1 +2n i+1 +n i+2 +m 1 +m i+1 )th-order orthogonal transformations Ui for
i = 2, . . . , N −1 such that
   
i −
M Li O i
M Ri − Li
Ui =
O Mi+1 −L i+1 i+1 O −
M L i+1
where Ri are invertible matrices of order 2n i+1 +m i+1 . This recursively defines the new matrices
Mi+1 and 
L i+1 with n 1 +n i+2 +m 1 rows.
Applying the transformations Ui successively to the ith and (i +1)th block rows of the trans-
formed pencil H − zT , the reduced pencil
⎡ ⎤
M1 R 1 − L1 O O
⎢ ⎥
⎢ .. ⎥
⎢ 
M2 O R2 . O ⎥
⎢ ⎥
⎢ ⎥
H − zT = ⎢ ⎢ .
.. .
.. .
.. . . . − ⎥ (27)
L −2 ⎥
⎢ N ⎥
⎢ ⎥
⎢M  . . . R N −1 ⎥
⎣ N −1 − z L N −1 O O ⎦
 
MN − z L N O O ... O
is obtained, which is orthogonally similar to H − zT . Since matrices Ri are non-singular, the
N − z 
regular subpencil M L N of order 2n 1 +m 1 will contain all finite eigenvalues of the original
pencil. To check the regularity of the extended pencil, the reciprocal condition numbers of the
upper triangular matrices Ri can be cheaply estimated to detect possible rank losses.
To compute the solution X 1 , orthogonals V1 and Z 1 are determined such that
 
H11 − zT11 H12 − zT12
N − z 
V1 ( M L N )Z 1 =
O H22 − zT22
where H11 − zT11 has only finite and stable eigenvalues. Partition Z 1 conformably into two block
columns and three block rows
⎡ ⎤
Z 1,11 Z 1,12
⎢ ⎥
Z 1 = ⎣ Z 1,21 Z 1,22 ⎦
Z 1,31 Z 1,32

such that Z 1,11 ∈ Rn 1 ×n 1 . Provided Z 1,11 is invertible, the solution X 1 and feedback F1 can be
−1 −1
computed as X 1 = Z 1,21 Z 1,11 and F1 = Z 1,31 Z 1,11 . The rest of the solution X k and the corresponding
feedback gain Fk for k = N , . . . , 2 are computed iteratively with (5) and (8). Since the process to
iterate directly on Equation (5) is convergent [13], this iteration produces virtually no errors.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
824 A. VARGA

To estimate the computational effort of the fast approach, let us assume constant dimensions
n and m for matrices Ak ∈ Rn×n and Bk ∈ Rn×m . The reduction of the pencil H − zT can be
done by computing successively N −1 QR decompositions of (4n +2m)×(2n +m) matrices and
applying the transformation to two sub-blocks of the same dimensions. The reduction step thus
has a computational complexity of O((N −1)(2n +m)3 ). Since the last step, the computation of
stable deflating subspace of the reduced pencil M N − z 
L N , has a complexity of O((2n +m)3 ), it
follows that the overall computational complexity of the proposed approach corresponds to what
is expected for a satisfactory algorithm for periodic systems. Although the periodic QZ-algorithm-
based approach has the same computational complexity, the fast approach is substantially more
efficient not only because it requires less number of operations but also because it is rich in the
so-called level 3 BLAS operations. This justifies the term fast algorithms introduced in [15].
Since the main reduction consists of successive QR-decompositions, it can be shown [32] that
the matrices of the computed reduced pencil H −T satisfy

U X − X 2  M f (2n +m) X 2 , X = H, T

where U is the matrix of accumulated left orthogonal transformations,  M is the relative machine
precision, and f (2n +m) is a quantity of order 2n +m. The subsequent computational step is
performed using the algorithm of [33] for a preliminary finite–infinite spectrum separation in
combination with the deflating subspace method to solve DARE of [31]. Thus, this step is also based
exclusively on orthogonal transformations and therefore is numerically stable as well. Overall, it
is thus guaranteed that the computed solution is exact for a slightly perturbed extended pencil. It
follows that the proposed algorithm to solve the DPRE is numerically backward stable in a certain
sense.
Since the structure of the perturbed extended pencil is not preserved in the reduction, we
cannot say, however, that the computed zeros are exact for slightly perturbed original data (i.e. the
algorithm is not strongly stable). In spite of this weaker type of stability, the proposed algorithm is
guaranteed to be numerically reliable, computationally efficient, and straightforward to implement
using available standard software.
Regarding the accuracy of the computed solution, it is to be expected that the extended periodic
QZ algorithm produces in general more accurate results than the fast algorithm. This claim is
supported by recent results on structured condition numbers for invariant subspaces [34], which
show that invariant subspaces of block cyclic matrices can be much more sensitive to unstructured
perturbations (as those introduced by the fast approach) than to structured perturbations.

3.5. Application to deadbeat control

An interesting application of the proposed techniques addresses the case of singular weighting
matrices Rk . Note that this case has not been addressed in any of the existing methods [9–11, 14].
The PLQ-optimization-based approach proposed in [17] for deadbeat control can be easily extended
to periodic systems. To this end, a PLQ optimization problem with Rk = 0, Sk = 0, and Q k = CkT Ck
is solved, where Ck ∈ Rm k ×n k are chosen such that the periodic system (Ak , Bk , Ck ) has no finite
zeros. For standard systems the choice of Ck is straightforward and involves the computation of
the Kalman controllability form [17]. The same approach can be used for periodic systems by
computing the periodic Kalman reachability form ( A k , 
Bk ) of the periodic pair (Ak , Bk ) using the
algorithm proposed in [35]. For a completely reachable pair (Ak , Bk ), this algorithm computes

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 825

orthogonal N -periodic transformation matrices Z k such that

k = Z k+1
A T
Ak Z k , 
Bk = Z k+1
T
Bk

where each matrix [  k ] is in a staircase form

Bk A
⎡ ⎤
Ak;1,0 Ak;1,1 Ak;1,2 ... Ak;1,

⎢ ⎥
⎢ O Ak;2,1 Ak;2,2 ... Ak;2,
⎥
[ k ] = ⎢
Bk | A ⎢ .
⎥
⎢ . .. .. .. .. ⎥ ⎥
⎣ . . . . . ⎦
O O O Ak;
,
−1 Ak;
,

(i) (i) (i) (i−1)
(i)
with Ak;i,i ∈ Rk+1 ×k , i = 1, . . . ,
; Ak;i,i−1 ∈ Rk+1 ×k ; and rankAk;i,i−1 = k+1 , i = 1, . . . ,
.
Ck can be chosen in the form

Ck = diag(Ck,1 , . . . , Ck,
−1 , I(
) )Z kT

k

(i−1) (i) (i−1)

(k −k+1 )×k
where Ck,i−1 ∈ R contains in each line a single non-zero element (e.g. set to 1) in
the position corresponding to a linearly dependent column of Ak;i,i−1 . For instance, in the simple
case when all sub-diagonal blocks are invertible (see Example 4 in the following section), Ck can
be simply chosen as

Ck = [O . . . O I(
) ]Z kT
k

4. NUMERICAL EXPERIMENTS

In this section several examples are considered to illustrate the capabilities of the proposed algo-
rithms. For the solution of DPRE computational results are reported only for the fast algorithm for
which reliable numerical implementations are available in the recently developed Periodic Systems
Toolbox for MATLAB [19]. To assess the accuracy of the results, the residuals (assuming Sk = 0)

rk = X k − Q k − ATk X k+1 (Ak + Bk Fk ) F

can be computed, from which a total residual can be defined as

 1/2
N
Residual = 2
rj (28)
j=1

Example 1
To illustrate the capabilities of the multiple shooting approach, we consider the numerical solution
to the PLDE (12), where we consider
 
0 1
A(t) =
−10 cos(t)−1 −24−10 sin(t)

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
826 A. VARGA

and we choose Q(t) such that the exact solution of (12) is

 
1+cos(t) 0
X (t) =
0 1+sin(t)

The period of the problem is T = 2. The characteristic multipliers of A(t) are e−0.046T and
e−23.95T ; thus, the problem is moderately stiff and moderately ill-conditioned.
We computed the unique periodic solution X (t) of (12) using sampling periods  = T /N , for
N = 1, 16, 64, 128, 256. The value of N = 1 corresponds to the single shooting method. For the
discretization of the continuous-time problem and the solution of the resulting discretized periodic
Lyapunov equations, tools available in the Periodic Systems Toolbox [19] have been used. In
Table I we present accuracy results obtained using three MATLAB ODE solvers: the Dormand–
Prince Runge–Kutta (4,5) code ode45, the non-stiff variable order Adams–Bashforth–Moulton
solver ode113, and the numerical differentiation-formulas-based stiff solver ode15s all with
both relative and absolute tolerances set to 10−8 . The accuracy of the solution is evaluated as
X (tk )− X (tk ) 2 in each point tk = (k −1), for k = 1, . . . , N .
Although the accuracy of the multiple shooting method is always better than that of the single
shooting approach, the accuracy gains obtained for the Runge–Kutta method are remarkable. Note
that the more accurate multistep methods implemented in ode113 and ode15s achieve their
limiting accuracy already for N = 64 and 16, respectively. We expect that similar conclusions can
also be drawn for the solution of PRDEs.

Example 2
This is example 2 of [9] and has a singular monodromy matrix. We solved for this problem the
deadbeat control problem by choosing R1 = R2 = R3 = 0 and
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
64 −8 8 1 −1 −15 1 0 −3
1 ⎢ ⎥ 1 ⎢ ⎥ 1 ⎢ ⎥
Q 1 = ⎣ −8 1 −1⎦ , Q 2 = ⎣ −1 1 15 ⎦ , Q 3 = ⎣ 0 0 0 ⎦
66 227 10
8 −1 1 −15 15 225 −3 0 9
The resulting exact solution is

F1 = [30 −20 2], F2 = 39

1
[−118 59 −74], F3 = [8 −48 0]

The computed floating-point solution had a total residual of 2.53×10−10 .

Table I. Accuracy results for max X (tk )− X (tk ) 2 .

N ode45 ode113 ode15s

1 1.6×10−4 1.9×10−8 2.4×10−7

16 1.4×10−6 8.3×10−9 2.1×10−8
64 1.1×10−7 5.6×10−9 5.4×10−8
128 9.0×10−9 9.1×10−9 1.1×10−7
256 1.1×10−9 8.3×10−9 8.9×10−8

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 827

Example 3
This is an example with time-varying dimensions, defined by the 3-periodic matrices
⎡ ⎤
    2 −3
−3 2 9 6 −3 ⎢ ⎥
A1 = , A2 = , A3 = ⎣ 4 −15⎦
0 0 −4 4 −2
−2 9
⎡ ⎤
    0
1 0 ⎢ ⎥
B1 = , B2 = , B3 = ⎣ 1⎦
1 1
1
For this problem the deadbeat control problem has been solved by choosing R1 = R2 = R3 = 0 and
⎡ ⎤
1 0 0 ⎡ ⎤
⎢ ⎥ 1
−
1  
⎢ 1 1⎥ ⎢ ⎥ 1 0
⎢0 ⎥
− ⎥, Q =⎢ 2 2 ⎥ , Q3 =
Q1 = ⎢ 2 ⎣
⎢ 2 2⎥ 1 1 ⎦ 0 0
⎣ 1 1 ⎦ −
0 − 2 2
2 2
The resulting total residual of the computed solution is 2.1×10−12 .
For reference purposes, the exact solution of the DPRE and the corresponding periodic deadbeat
state-feedback are given as
⎡ ⎤
11 39
−3 − ⎡ ⎤
⎢ 2 2⎥ 2003 1007  
⎢ ⎥ −
⎢ 5 25 ⎥ ⎢ 22 22 ⎥ 23 −78
X1 = ⎢⎢ −3
⎥ , X2 = ⎢ ⎥ X3 =
⎢ 2 2 ⎥⎥
⎣ 1007 509 ⎦ −78 297
⎣ 39 25 ⎦ −
22 22
− 85
2 2
 
80 40 8 32
F1 = [6 −4 −22], F2 = − , F3 = −
33 33 5 5

Example 4
This example has been used in [36] to design periodic output feedback controllers for peri-
odic systems with relatively large periods. The discrete-time periodic system originates from a
continuous-time periodic model of a spacecraft pointing and attitude system described in [37]. This
system has state and input dimensions of n = 4, m = 1, respectively, and a period of T = 6073.8 s.
The discretized system for different sampling periods T /N has been used in [36] to design periodic
output feedback controllers for this system. The matrices of the discrete-time periodic system can
be computed explicitly for arbitrary values of N .
The deadbeat control problem has been solved for different sampling times by choosing Rk = 0,
Sk = 0, and Q k = CkT Ck with Ck = [0 0 0 1]Z kT , where Z k is the N -periodic orthogonal transfor-
mation matrix that brings the periodic pair (Ak , Bk ) into the periodic Kalman reachability form
k , 
(A Bk ). Note that in this form each  k is an upper
Bk has only the (1,1) entry non-zero and each A
Hessenberg matrix.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
828 A. VARGA

Table II. Computational results for satellite deadbeat control.

N 40 120 360 600

Residual 1.0×10−11 7.7×10−14 8.4×10−12 2.4×10−11

ITER 2 2 2 6
STDEG 2×10−155 5×10−324 1×10−322 2×10−322

In Table II we present the following results obtained for different values of N : the total residual
(28), ITER—the number of iterations performed on DPRE (5) to achieve the limiting accuracy,
and STDEG—the achieved closed-loop stability degree (i.e. the maximum modulus of closed-loop
characteristic multipliers). Note that N = 600 corresponds to a typical sampling period of about
10 s used to control satellites on low Earth orbits.

5. CONCLUDING REMARKS

The proposed algorithms to solve PRDEs, PLDEs, and DPREs are completely general, numerically
reliable, and computationally efficient. To solve the PRDE and PLDE, we proposed a multiple
shooting-type approach, which computes the periodic solutions in an arbitrary number of equidistant
time instants within one period, by employing a suitable discretization of the continuous-time
problems. In contrast to traditionally used single shooting periodic generator methods, the multiple
shooting method has the advantage of successfully tackling problems with large periods and/or
unstable dynamics. The proposed methods to solve DPREs, in contrast to existing methods, are
able to address problems with time-varying dimensions and/or with singular weighting matrices.
The new methods to solve PRDEs and DPREs can also be used to solve the more general periodic
Riccati equations, as those arising in periodic H∞ -synthesis problems [38–40].
For the solution of PLDEs using the multiple shooting approach and the solution of DPREs
using the fast algorithm, robust numerical implementations are available in the Periodic Systems
Toolbox [19]. The codes for solving DPREs underlie user-friendly software to solve PLQ control
and filtering-related problems. This software used in conjunction with software tools to compute
periodic Kalman forms allows one to solve periodic deadbeat control problems in the most general
setting (i.e. for controllable systems with time-varying dimensions).

APPENDIX A

We present a special version of the general algorithm proposed in [29] to compute periodic
Kronecker-like forms. The resulting algorithm is intended to efficiently perform the finite–infinite
spectral separation of the associated lifted regular pencil P(z) := H − zT defined in (26) by manip-
ulating exclusively the given periodic pair (Mk , L k ), where Mk ∈ Rk ×k and L k ∈ Rk ×k+1 . The
resulting reduced form of this pair is the required one in the proposed extended periodic QZ
algorithm to solve DPRE.
For N = 1, this algorithm corresponds to the finite–infinite structure separation procedure used
to determine the Kronecker structure of a linear pencil [41]. In the case of multiple infinite

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 829

eigenvalues, this approach is the numerically reliable alternative to the reordering approach via
the QZ algorithm.
The main computation is to determine orthogonal periodic transformation matrices Q k and Z k
such that
 ∞   ∞ 
Mk ∗ Lk ∗
Q k Mk Z k = f
, Q k L k Z k+1 = f
(A1)
O Mk O Lk

where
(a) Mk∞ is invertible and the product (Mk∞ )−1 L ∞ ∞ −1 ∞
k . . . (Mk+N −1 ) L k+N −1 is nilpotent and
f
(b) L k is non-singular.
Using the above orthogonal transformations to reduce the periodic pair (Mk , L k ) to the form
in (A1) is equivalent to computing P(z)  = QP(z)Z, where Q = diag(Q 1 , Q 2 , . . . , Q N ) and Z =

diag(Z 1 , Z 2 , . . . , Z N ). Since P(z) has the same cyclic structure as P(z), by using appropriate
permutation matrices 1 and 2 we can reorder its blocks such that
 ∞ 
P (z) ∗

1 P(z)2 = (A2)
O P f (z)

with each non-zero block having the same cyclic structure as P(z) in (26). For example, the
diagonal blocks have the form
⎡ ⎤
M1x −L 1x O ... O
⎢ ⎥
⎢ O M2x −L 2x ... O ⎥
⎢ ⎥
⎢ . . . . . ⎥
⎢
P (z) = ⎢ .
x . . . . . . . . ⎥
. ⎥
⎢ ⎥
⎢ O M Nx −1 −L xN −1 ⎥
⎣ ⎦
−z L xN O ... O M Nx

where the upper index x stands for ∞ or f . The reduction algorithm that we describe in this
appendix guarantees the above properties (a) and (b) of the reduced matrices. It follows that
pencils P ∞ (z) and P f (z) have only infinite and finite zeros, respectively [29]. Note that the simple
infinite eigenvalues of P ∞ (z) and P f (z) introduced by lifting play no role in counting zeros (the
multiplicities of infinite eigenvalues are excess one to the multiplicities of infinite zeros).
The proposed algorithm has two main steps, which we discuss subsequently.

A.1. Preliminary compression

In the first step we reduce the problem to an equivalent one, but with a special structure of matrices.
(1) (1)
Let Q k and Z k be orthogonal periodic matrices such that
   
(1) (1) (1)
Bk Ak (1) (1) (1)
O Ek
Mk := Q k Mk Z k = , L k := Q k L k Z k+1 = (A3)
Dk C k O O

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
830 A. VARGA

where for k = 1, . . . , N , E k ∈ Rn k+1 ×n k+1 is invertible, Ak ∈ Rn k+1 ×n k , Bk ∈ Rn k+1 ×m k , Ck ∈ R pk ×n k ,

Dk ∈ R pk ×m k , with m k := k −n k and pk := k −n k+1 . Owing to the regularity of the lifted pencil,
it follows that
 
Bk
Dk
must have a full column rank and [Dk Ck ] must have a full row rank.
The compression of each L k to a non-singular E k can be done by computing a full orthogonal
(1)
decomposition Q k L k Vk+1 = diag(E k , O) using either the singular-value decomposition or a rank-
revealing QR-decomposition followed by an RQ-decomposition. Finally, the form in (A3) is
obtained by applying column permutations with an appropriate permutation matrix k . For both
rank determination techniques, we can arrange that each resulting E k is upper triangular. Define
(1)
Z k = Vk k .

A.2. Basic reduction

The goal of this reduction step is to separate the infinite (∞) and finite ( f ) zero structures of the
(1) (1)
reduced lifted pencil P (1) (z), which corresponds to the reduced periodic pair (Mk , L k ). In this
(2) (2)
step we compute orthogonal periodic matrices Q k and Z k such that
 ∞   ∞ 
(2) (2) (1) (2)
Mk ∗ (2) (2) (1) (2)
Lk ∗
Mk := Q k Mk Z k = f
, Tk := Q k L k Z k+1 = f
(A4)
O Mk O Lk
∞ ∞
where for k = 1, . . . , N , Mk∞ ∈ Rn k ×n k is in a staircase form (see (A5)) with invertible matrices
n∞ ∞
k ×n k+1 has the non-zero trailing part of full column rank (see (A6)),
on its main diagonal, L ∞ k ∈R
f f f f
Mk ∈ Rn k+1 ×n k , and L k ∈ Rn k+1 ×n k+1 is invertible. The following basic reduction algorithm can be
f f

used to identify and separate the finite and infinite structures of a compressed periodic pair:
Algorithm for FINITE–INFINITE separation.
(0) (2) (2)
For k = 1, . . . , N , set n ∞
f
k = 0, n k = n k , k = m k ; set Q k = Ik , Z k = Ik .
step i
1. For each k = 1, . . . , N , compute (e.g. by performing the QR-decomposition with column
pivoting on Dk ) an orthogonal matrix Wk and a permutation matrix k such that
⎡ ⎤
 
-------------

Bk,1 Bk,2 Ak
⎢ ⎥ Bk Ak
⎢D ⎥
⎣ k,1 Dk,2 Ck,1 ⎦ := diag(In k+1 , Wk ) D C diag(k , In k )
--------------- k k
O O Ck,2
(i) (i)
where Dk,1 ∈ R
k ×
k is invertible and upper triangular.
2. For each k = 1, . . . , N , compress the rows of
 
Bk,1
Dk,1

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 831

with orthogonal X k such that

       

---------
Bk,1 Bk,2 Ak

--------
Bk,11 Bk,12 Ak,1 E k,1 Ek
:= X k , := X k
O Bk,22 Ak,2 Dk,1 Dk,2 Ck,1 E k,2 O
(i) (i)
where Bk,11 ∈ R
k ×
k has a full row rank and is upper triangular and E k,2 ∈ Rn k+1 ×n k+1 is
invertible and upper triangular (see [42]).
3. For k = 1, . . . , N , determine orthogonal Uk to compress the rows of Bk,22 to an invertible
upper triangular matrix such that
 

Bk,22
Uk Bk,22 =
O
(i) (i−1) (i)
(i) (i−1) (i)
where Bk,22 ∈ Rk+1 ×(k −
k ) is invertible, and k+1 := k −
k ; compute orthogonal
Vk+1 such that Uk E k,2 Vk+1 is upper triangular.
4. For k = 1, . . . , N , form the transformation matrices
k = diag(I (i) ,Uk , I
Q (i) ) diag(X k , I p (i) ) diag(In k+1 , Wk )

p k k −
k k −
k


Z k = diag(k , In k ) diag(Im k , Vk )

and transform the submatrices and partition them as

(i) ⎡ ⎤

k
-------------------

-------------------

Bk,11 Bk,12 Ak,11 Ak,12

-------------------------------------⎥
k+1 ⎢
(i)
⎢ O 
Bk,22 Ak,21 Ak,22 ⎥
⎢ ⎥  
(i) ⎢- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ⎥ Bk Ak
n k+1 −k+1 ⎢ O O Ak,31 Ak,32 ⎥:= Q k 
Zk
(i)
⎣------------------------------------- ⎦
vpk −
k Dk Ck
O O Ck,21 Ck,22
(i) (i−1) (i) (i) (i)

k k −
k k n k −k

(i) ⎡ ⎤
------------------
------------------

k O O E k,11 E k,12
------------------------------------
⎢ ⎥
(i) ⎢O ⎥
k+1 ⎢
O E k,21 E k,22
⎥  
-
⎢ - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - ---- ⎥ O Ek
(i) ⎢O ⎥:= Q
n k+1 −k+1 O O E k,32 k 
Z k+1
⎣
------------------------------------ ⎦
(i) O O
pk −
k O O O O
(i) (i−1) (i) (i) (i)

k+1 k+1 −
k+1 k+1 n k+1 −k+1

where Bk,11 and  Bk,22 are invertible and upper triangular, and E k,21 and E k,32 are invertible
and upper triangular.
5. For k = 1, . . . , N , update Ak := Ak,32 , E k := E k,32 , Bk := Ak,31 , Ck := Ck,22 , Dk := Ck,21 .
(2) k )Q (2) , Z (2) := Z (2) diag(In ∞ , 
6. For k = 1, . . . , N , Q k := diag(In ∞
k
,Q k k k k
Z k ).
(i−1) (i) (i)
7. For k = 1, . . . , N , update n ∞ ∞ f f
k := n k +k , n k := n k −k , pk := pk −
k .

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
832 A. VARGA

(i)
8. If k = 0 for k = 1, . . . , K , then go to exit
9. i := i +1 go to step i;
(2) (2) (1) (2) (2) (2) (1) (2)
exit. Compute Mk := Q k Mk Z k , L k := Q k L k Z k+1 and partition them according to (A4).
The computation stops when all Bk and Dk have null number of columns. The resulting finite
f f f f
periodic pair is (Mk , L k ) := (Ak , E k ), with Mk ∈ Rn k+1 ×n k , and L k ∈ Rn k+1 ×n k+1 invertible and
f f f f

upper triangular. Since the associated lifted pencil has full rank for almost all finite values of z,
f f
the pair (Mk , L k ) can have only finite eigenvalues. These are the eigenvalues of (L f )−1 M f (k).
The compression at Step 2 can be done by performing a QR-decomposition of
 
Bk,1
Dk,1

which exploits the upper triangular shape of Dk,1 . This can be achieved by employing sequences
of Givens transformations to zero successive elements under the diagonal of Bk,1 . By starting from
below (i.e. zeroing first the diagonal element of Dk,1 ) the upper triangular structure of E k,2 is
automatically achieved. For details see [42].
The compression at Step 3 of Bk,22 to a full row rank matrix can be done simultaneously with
maintaining E k,2 upper triangular. This compression technique represents the main computational
step in determining the periodic controllability staircase form of periodic descriptor systems (see
[35] for more details).
(2) (2)
At the end of this algorithm we obtain globally the reduced matrices Mk and L k in the form
of (A4), where the periodic pair (Mk∞ , L ∞
k ) is in the following staircase form:
⎡ ∞ ∞ ∞ ∞ ⎤
Mk;1,1 Mk;1,2 ... Mk;1,
−1 Mk;1,

⎢ ∞
... ∞ ∞ ⎥
⎢ O Mk;2,2 Mk;2,
−1 Mk;2,
⎥
⎢ ⎥
∞ ⎢ .. .. .. .. .. ⎥
Mk = ⎢ . . . . . ⎥ (A5)
⎢ ⎥
⎢ ∞ ∞ ⎥
⎣ O O ... Mk;
−1,
−1 Mk;
−1,
⎦
∞
O O ... O Mk;
,

⎡ ⎤
O L∞
k;1,2 ... L∞
k;1,
−1 L∞
k;1,

⎢O ... L∞ L∞ ⎥
⎢ O k;2,
−1 k;2,
⎥
⎢ ⎥
⎢ ⎥
L k = ⎢ ...
∞ ..
.
..
.
..
.
..
. ⎥ (A6)
⎢ ⎥
⎢ ⎥
⎣O O ... O L∞
k;
−1,
⎦
O O ... O O
(i) (i) (i−1)
∞ ∈ R(k+1 +
k
where
is the number of steps performed by the algorithm, Mk;i,i )×k
is invertible
(i) (i) (i)
(i) (i) (i−1) (k+1 +
k )×k+1
(k+1 +
k = k ), and L ∞
k;i,i+1 ∈ R is
of full column rank. Since by construction
the associated lifted finite values of z, the pair (Mk∞ , L ∞
pencil P ∞ (z) is invertible for all k ) has
only an infinite structure.
(2) (1) (1) (2)
The final transformation matrices are Q k = Q k Q k and Z k = Z k Z k .

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 833

APPENDIX B

Here, we propose an algorithm to reduce a periodic pair (Mk , L k ), where Mk ∈ R(n k +n k+1 )×2n k ,
L k ∈ R(n k +n k+1 )×2n k+1 , to a compressed form, where the parts corresponding to trivial characteristic
values in the origin and at infinity, originated from the presence of time-varying dimensions, are
separated. Let us assume that n 1 = n := min n k . The following procedure determines orthogonal
periodic transformation matrices Q k and Z k to achieve this separation using a sequence of QR-
and RQ-decompositions:
Algorithm for separation of trivial zero and infinite spectra.
Set Z 1 = In 1 and compute orthogonal Q N (e.g. using the QR-decomposition) such that
 

L N ,1
TN = Q N L N =
O
where L N ,1 is square and upper triangular.
For k = N , N −1, . . . , 2
1. Compute orthogonal Z k (e.g. using the RQ-decomposition) such that
⎧
⎪ [O Mk,2 ] if n k > n k+1
⎪
⎨ 
Hk := Q k Mk Z k = Mk,1
⎪
⎪ if n k n k+1
⎩ 
Mk,2
k,2 is square and upper triangular.
where M
2. Compute orthogonal Q k−1 (e.g. using the QR-decomposition) such that
⎧
 
⎪ [ L k−1,1 L k−1,2 ] if n k−1 n k
⎪
⎨  
Tk−1 = Q k−1 L k−1 Z k = 
L k−1,1
⎪
⎪ if n k−1 > n k
⎩
O
where 
L k−1,1 is square and upper triangular.
Compute H1 = Q 1 M1 .
The resulting matrices Hk and Tk can be partitioned as follows:
⎡ ⎤
Hk,11 Hk,12 Hk,13
⎢ ⎥
Hk := Q k Mk Z k = ⎣ O Hk,22 Hk,23 ⎦
O O Hk,33
⎡ ⎤
Tk,11 Tk,12 Tk,13
⎢ ⎥
Tk := Q k L k Z k+1 = ⎣ O Tk,22 Tk,23 ⎦
O O Tk,33
where Hk,11 ∈ R(n k+1 −n)×(n k −n) , Hk,22 ∈ R2n×2n , Hk,33 ∈ R(n k −n)×(n k −n) , Tk,11 ∈ R(n k+1 −n)×(n k+1 −n) ,
Tk,22 ∈ R2n×2n , and Tk,33 ∈ R(n k −n)×(n k+1 −n) . Because the lifted pencil associated with the pair

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
834 A. VARGA

(Mk , L k ) is regular, it follows immediately that the upper triangular periodic matrices Hk,33 and
Tk,11 are invertible. The periodic pair (Hk,11 , Tk,11 ) contains the trivial zero characteristic values
of the original pair (Mk , L k ), whereas the periodic pair (Hk,33 , Tk,33 ) contains the trivial infinite
characteristic values. The numbers of zero and infinite characteristic values are equal but depend
on the starting time moment used to define the period. The periodic pair (Hk,22 , Tk,22 ) contains
the 2n core characteristic values of the original pair (Mk , L k ). A more refined version of the
above algorithm can be devised such that the periodic pair (Hk,22 , Tk,22 ) results in a generalized
periodic Hessenberg form (see [8, 9]).

ACKNOWLEDGEMENTS
This work was partially supported by the Swedish Strategic Research Foundation Grant ‘Matrix Pencil
Computations in Computer-Aided Control System Design: Theory, Algorithms and Software Tools.’ The
author thanks the anonymous reviewers for their very detailed and careful review, which substantially
improved the presentation of the results.

REFERENCES
1. Bittanti S. The periodic Riccati equation. In The Riccati Equation, Bittanti S, Laub AJ, Willems JC (eds).
Springer: Berlin, 1991; 127–162.
2. Colaneri P. Periodic control systems: theoretical aspects. Proceedings of IFAC Workshop on Periodic Systems,
Yokohama, Japan, 2005.
3. Brockett RW. Finite Dimensional Linear Systems. Wiley: New York, 1970.
4. De Nicolao G, Strada S. On the use of reachability Gramians for the stabilization of linear periodic systems.
Automatica 1997; 33:729–732.
5. Varga A. On solving periodic differential matrix equations with applications to periodic system norms computation.
Proceedings of CDC’05, Seville, Spain, 2005.
6. Varga A. On solving discrete-time periodic Riccati equations. Proceedings of IFAC 2005 World Congress, Prague,
Czech Republic, 2005.
7. Hench JJ, Kenney CS, Laub AJ. Methods for the numerical integration of Hamiltonian systems. Circuits, Systems
and Signal Processing 1994; 13:695–732.
8. Bojanczyk AW, Golub G, Van Dooren P. The periodic Schur decomposition. Algorithms and applications. In
Proceedings SPIE Conference, San Diego, vol. 1770, Luk FT (ed.), July 1992; 31–42.
9. Hench JJ, Laub AJ. Numerical solution of the discrete-time periodic Riccati equation. IEEE Transactions on
Automatic Control 1994; 39:1197–1210.
10. Benner P, Byers R, Mayo R, Quintana-Orti ES, Hernandez V. Parallel algorithms for LQ optimal control of
discrete-time periodic linear systems. Journal of Parallel and Distributed Computing 2002; 62:306–325.
11. Chu EKW, Fan HY, Lin WW, Wang CS. Structure-preserving algorithm for periodic discrete-time algebraic
Riccati equations. International Journal of Control 2004; 77:767–788.
12. Kressner D. Numerical Methods for General and Structured Eigenvalue Problems. Lecture Notes in Computational
Science and Engineering, vol. 46. Springer: Berlin, 2005.
13. Bittanti S, Colaneri P, De Nicolao G. The difference periodic Riccati equation for the periodic prediction problem.
IEEE Transactions on Automatic Control 1988; 33:706–712.
14. Benner P, Byers R. Evaluating products of matrix pencils and collapsing matrix products. Numerical Linear
Algebra with Applications 2001; 8:357–380.
15. Van Dooren P. Two point boundary value and periodic eigenvalue problems. Proceedings of the CACSD’99
Symposium, Kohala Coast, HI, 1999.
16. Oară C, Varga A. The general inner-outer factorization problem for discrete-time systems. Proceedings of the
ECC’99, Karlsruhe, Germany, 1999.
17. Emami-Naeini A, Franklin GF. Deadbeat control and tracking of discrete-time systems. IEEE Transactions on
Automatic Control 1982; 27:176–181.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla
 PERIODIC RICCATI EQUATIONS 835

18. Varga A, Van Dooren P. Computing the zeros of periodic descriptor systems. Systems and Control Letters 2003;
50:371–381.
19. Varga A. A periodic systems toolbox for MATLAB. Proceedings of IFAC 2005 World Congress, Prague, Czech
Republic, 2005.
20. Bittanti S, Colaneri P, Guardabassi G. Analysis of the periodic Lyapunov and Riccati equations via canonical
decomposition. SIAM Journal on Control and Optimization 1986; 24:1138–1149.
21. Hench JJ. Private communication, 1995.
22. Hairer E, Lubich C, Wanner G. Geometric Numerical Integration: Structure-Preserving Algorithms for Ordinary
Differential Equations. Springer: Berlin, 2002.
23. Hairer E, Hairer M. GniCodes—MATLAB programs for geometric numerical integration. In Frontiers in Numerical
Analysis Durham 2002, Blowey J et al. (eds). Springer: Berlin, 2003; 199–240.
24. Dieci L. Numerical integration of the differential Riccati equation and some related issues. SIAM Journal on
Numerical Analysis 1992; 29:781–815.
25. Dieci L, Eirola T. Positive definiteness in the numerical solution of Riccati differential equations. Numerische
Mathematik 1994; 67:303–313.
26. Varga A. Periodic Lyapunov equations: some applications and new algorithms. International Journal of Control
1997; 67:69–87.
27. Bittanti S, Colaneri P. Analysis of discrete-time linear periodic systems. In Digital Control and Signal Processing
Systems and Techniques, Control and Dynamics Systems, vol. 78, Leondes CT (ed.). Academic Press: New York,
1996; 313–339.
28. Grasselli OM, Longhi S. Pole-placement for nonreachable periodic discrete-time systems. Mathematics of Control
Signals and Systems 1991; 4:439–455.
29. Varga A. Computation of Kronecker-like forms of periodic matrix pairs. Proceedings of MTNS’04, Leuven,
Belgium, 2004.
30. Varga A. Balancing related methods for minimal realization of periodic systems. Systems and Control Letters
1999; 36:339–349.
31. Van Dooren P. A generalized eigenvalue approach for solving Riccati equations. SIAM Journal on Scientific and
Statistical Computing 1981; 2:121–135.
32. Golub GH, Van Loan CF. Matrix Computations. John Hopkins University Press: Baltimore, 1989.
33. Misra P, Van Dooren P, Varga A. Computation of structural invariants of generalized state-space systems.
Automatica 1994; 30:1921–1936.
34. Byers R, Kressner D. Structured condition numbers for invariant subspaces. SIAM Journal on Matrix Analysis
and Applications 2006; 28:326–347.
35. Varga A. Computation of Kalman decompositions of periodic systems. European Journal of Control 2004; 10:1–8.
36. Varga A, Pieters S. Gradient-based approach to solve optimal periodic output feedback control problems.
Automatica 1998; 34:477–481.
37. Pittelkau ME. Optimal periodic control for spacecraft pointing and attitude determination. Journal of Guidance,
Control, and Dynamics 1993; 16:1078–1084.
38. Colaneri P. Continuous-time periodic systems in H2 and H∞ . Part I: Theoretical aspects. Kybernetika 2000;
36:211–242.
39. Colaneri P. Continuous-time periodic systems in H2 and H∞ . Part II: State feedback problems. Kybernetika
2000; 36:329–350.
40. Sagfors MF, Toivonen HT, Lennartson B. State-space solution to the periodic multirate H∞ control problem: a
lifting approach. IEEE Transactions on Automatic Control 2000; 45:2345–2350.
41. Van Dooren P. The computation of Kronecker’s canonical form of a singular pencil. Linear Algebra and its
Applications 1979; 27:103–141.
42. Oară C, Van Dooren P. An improved algorithm for the computation of structural invariants of a system pencil
and related geometric aspects. Systems and Control Letters 1997; 30:39–48.

Copyright q 2008 John Wiley & Sons, Ltd. Numer. Linear Algebra Appl. 2008; 15:809–835
DOI: 10.1002/nla