1

Structured Doubly Coprime Factorizations

and Youla Parameterizations
for a Class of Distributed Control Problems
Serban Sabau] , Cristian Oara? , Sean Warnick and Ali Jadbabaie[

I. Introduction
Technological advances in various disciplines such as computer vision, ranging sensors, wireless communications
and estimation and detection algorithms have made autonomous vehicles that have the ability to perform almost
any maneuver on the road without any human assistance a reality. Such maneuvering can be safely performed on
public roadways with traffic present and includes emergency stoping, steering and turning, navigating, lateral and
garage parking. Sensor data and other information is collected and processed in order to facilitate lane keeping,
traffic signs recognition and awareness, object tracking in vehicles operating range (pedestrians or other vehicles),
etc. and then used to determine accordingly the vehicles desired trajectory.
In particular, vehicle platooning is important for traveling at high speeds on the motorway. Given the technological
platforms described above that are already available to the automobile industry, it would be useful to design and
implement a feasible control scheme along with suitable control algorithms, in order to tackle the undesired and
potentially risky oscillatory phenomena inherent to the dynamics of vehicle platooning (the socalled Forrester effect,
or bullwhip effect [5]) while maximizing the throughput of highway traffic by reducing the accordion effect on the
behavior of the platoon.

A. Motivation
We consider an array of n autonomous, selfdriving vehicles moving in one dimension, all traveling in the same
(positive) direction of a highway with a single line of traffic. The key issue of the performance of the string of
vehicles is maintaining a constant spacing distance between any two successive vehicles, whilst maintaining the
velocity imposed by the leader vehicle (first car in the string).
]S
erban

Sabau is with the Electrical and Computer Engineering Dept., Stevens Institute of Technology. email: ssabau@stevens.edu

? Cristian

Oara is with the Faculty of the Automatic Control Dept., University Politehnica Bucharest. email: cris-

tian.oara@acse.pub.ro The work of Cristian Oara has been supported by a grant of the Romanian National Authority for Scientific
Research, CNCS - UEFISCDI, project identification number PN-II-ID-PCE-2011-3-0235
Sean Warnick is with the Information and Decision Algorithms Laboratories and the Faculty of the Computer Science
Department at Brigham Young University. email: sean.warnick@gmail.com.
[ Ali

Jadbabaie is with the Faculty of the Electrical and Systems Engineering Dept., University of Pennsylvannia. email:

jadbabai@seas.upenn.edu.

DRAFT

Platoon motion control has been a problem of interest for the control engineering community for about fifty years
(see for example [1]) as supported by the abundance of scientific publications on the topic. The platoon motion
control problem is easy to solve (using standard design methods) provided that each vehicle in the platoon has
instant access at all times at the relative coordinates and at the instantaneous speed of all other vehicles. Even with
todays ubiquitous communication technology this is not a realistic scenario.
Considering the above, the control engineering community has considered of interest only those decentralized
control schemes which use exclusively information that can be measured with sensors placed onboard each vehicle,
specifically the relative distance from each vehicle to its predecessor in the platoon. This is typically a distance
ranging between 20 and 250 yards, which can be measured using a low cost but highly reliable optic (lidar) or
radar (electromagnetic) ranging sensor mounted on each vehicle. However, a difficulty dubbed string instability
has been documented quite early ([2], [3]) for the control schemes with long strings of vehicles that use only local
information. A rigorous definition of string instability is given in [3], but intuitively it can be described as the
phenomenon of amplification of a disturbance to the leader vehicle, as the disturbance propagates toward the back
of the platoon. String instability is obviously undesirable, since it can ultimately lead to vehicles in the string
crashing into each other. Equally important, lack of a strong attenuation of a disturbance to any car in the string
propagating towards the back the string [4, Section IV] (i.e. a high sensitivity to disturbances), even if it does not
necessarily lead instability, it produces another highly unwanted effect: the socalled accordion effect. The accordion
effect greatly disrupts the throughput of the highway traffic.
B. Contribution
We provide a novel distributed control scheme for platoon motion control that provides a decentralized implementation of the leader information controller [4]. The leader information controller has been known to provide good
string stability performance along with low sensitivity to disturbances [4]. In our design, in order to regulate the
spacing between itself and its predecessor in the string, every vehicle (with the exception of the leader vehicle) uses
only two pieces of information from its predecessor. Firstly, the relative distance from itself to its predecessor, which
can be measured using an on-board laser ranging sensor and secondly, a low data rate, digital radio signal sent from
its predecessor. We provide a generalization of the concept of leader information controllers that is capable o dealing
with nonhomogeneous strings of vehicles and with the case of having non identical controllers for each vehicle
in the string. We also provide the Youla parameterization of all leader information controllers associated with a given
platoon of vehicles, by imposing a particular sparsity constraint on the Youla parameter. Our parameterization is
based on a doubly coprime factorization of the aggregated model of the platoon that features a special structure.
The sparsity constraint imposed on the Youla parameter preserves the structure of the doubly coprime factorizations
and consequently the structure of the resulted stabilizing controllers. The tuning of the local controllers can be done
by choosing H 2 or a H (or mixed) cost functionals, represents the usual metrics in control theory to evaluate
reliability and performance. Our Youla parameterization is amenable to incorporating such costs in the controller
design method.
Recent results on the socalled indirect leader broadcast scheme from [14] dealing with homogeneous strings of
vehicles, present certain similarities with our leader information controller with the important distinction that in our
leader information controller scheme we broadcast the control signal of the predecessor vehicle instead of a local
DRAFT

estimate of the leaders state (as in [14]). The control signal is basically generated on board of the predecessor vehicle,
hence there is no need to estimate it. The fact that information free of noise is broadcasted (with sometimedelay
induced by the wireless communication systems) has profound implications in terms of the performance of the closed
loop. Furthermore, the leader information control scheme introduced here can be adapted as to compensate for the
feedforward timedelay induced by the wireless communication broadcasting of the predecessors control signal, as
will be explained in the sequel in full detail.
In [8] a control scheme for homogeneous strings of vehicles, very similar to the leader information controller is
presented, in which the predecessors control signal is broadcasted. However, because the scheme is implemented
taking into account a Pade approximation of the feedforward timedelay induced by the wireless communication,
the resulting controller from [8, Section V] will never be (in general) a leader information controller, as we argue in
the sequel. The argument is related to the way the broadcasting delays propagate through the string of vehicles
(phenomenon which is studied in detail in [14]).
II. Preliminaries
A. Notation
Most of the notation we use in the sequel is fairly standard in the systems and control literature. the Laplace
transform and inverse Laplace transform operator are denoted by L and L1 . The Laplace teansform complex variable
is s C and Laplace transforms will be typically denoted by an upper case letter, that is L{u(t)} = U(s) and L1 {U(s)} =
u(t).
n

The number of vehicles in the platoon

de f

x = y

x is by definition equal to y

R(s)

Set of all realrational transfer functions.

pq

TFM

Set of p q matrices having all entries in R(s)

S
Transfer function matrix, or, equivalently, p,q R(s)pq

the open left halfplane. Stability region for TFMs.

A(s)

Subset of R(s) whose entries have poles in

A(s)pq

Set of p q stable TFMs

A(s)

Set of stable TFMs, or, equivalently,

The plant is a TFM with strictly proper entries.

The LTI controller is an element of R(s)mm

H(G, K)

The TFM from [T wT ]T to [yT uT ]T in Figure 1

P(s) ? u(t)

the time response with zero initial conditions of a (LTI) system P(s) with input u(t)

R(s)

S
p,q

A(s)pq

Remark II.1. When the time argument (t) or the frequency argument (s) can be inferred from the context it is omitted.
Table I contains notation for certain structured lower triangular matrices which will be used in the sequel. We
further adopt the following notation:

DRAFT

TABLE I
Notation for structured lower triangular matrices



D d1 , d2 , . . . , dn



B b1 , b2 , . . . , bn

b1

b1

..



R r1 , r2 , . . . , rn

d1
0
0
..
.
0

0
d2
0
..
.
0

0
b2
b2
..
.
0
0
r1
r2
r3
..
.
rn

...
...
...

0
0
b3
..
.
0
0
0
r2
r3
..
.
rn

...
...
...
..
.
...

0
0
d3
..
.
0

...
...
...
0
0
r3
..
.
rn

0
0
0
..
.
dn
0
0
0
..
.

bn1
bn1

...
0

...
0

...
0

..
...
.

. . . rn

0
0
0
..
.
0
bn

B. The Dynamical Model of Vehicles

Assumption II.2. (Vehicle Dynamical Model) Each of the n individual vehicle is given by a Linear and Time Invariant, finite
dimensional transfer function Gk (s), 1 k n, whose minimal, strictly proper realization has no unstable hidden modes (see
e.g. [11] for fundamental notions in LTI systems such as realizations, minimality, hidden modes, etc. ).
Assumption II.2 is the most common assumption throughout theoretical control literature when dealing with
platoons of vehicles. It is considered to be an accurate mathematical representation of vehicles higherorder dynamics
for longitudinal control (see [7],[6, Section 2.1.2] and the references within). Provided the car is equipped with an
Electrohydraulic Braking and Throttle actuation system then this actuating system can incorporate an inputoutput
linearization of the breaking system and by doing that, it can incorporate the nonlinear lower order dynamics of the
braking system [6].
Under these circumstances one instance of feedback linearization could lead for example to a simple continuous
time, point mass model ([7, (1)/pp. 4270]) with first order actuator dynamics whose transfer function from the controls
uk to the vehicles position yk is:

Gk (s) =

1
mk s2 (k s + 1)

for

1kn

(1)

Theoretical and experimental results related to vehicle dynamical models such as (1) can be found in [7] for a Citroen
C4s. The model in (1) does not hold for certain special situations such as emergency breaking or high speed tight
turns but it can effectively describe under regular conditions, vehicles in highway traffic.
Assumption II.3. (Linear and Time Invariant Control) We assume that the control law is given by a linear and timeinvariant
operator.

DRAFT

+
+

Fig. 1.

Standard unity feedback configuration

C. Standard Unity Feedback Configuration

Under the working assumptions II.2 and II.3, we will focus on the standard feedback configuration of Fig. 1, where
G(s) is an LTI plant and K(s) is an LTI controller that are finite dimensional and operate in continuous time. Here, w
and are the input disturbance and sensor noise, respectively. In addition, u and y are the control and measurable
output vectors, respectively. For general information about Linear Feedback Control Systems we refer to [13], [10],
[11].
If the TFM H(G, K) is in A(s)2n2n (i.e. it is stable) we say that K is a stabilizing controller of G, or equivalently that
K stabilizes G. If a stabilizing controller of G exists, we say that G is stabilizable.

D. Coprime and Doubly Coprime Factorization for LTI Systems ( [10])

A right coprime factorization (RCF) of G over is a fractional representation of the form G = NM1 , with N Ann
and M Ann , and for which there exist X Ann and Y Ann satisfying YM + XN = In ([10, Ch. 4, Corollary 17]),
e 1 N,
e
where In is the identity matrix. Analogously, a left coprime factorization (LCF) of G (over ) is defined by G = M
e+ N
e Ann . Due to the natural interpretation
e Ann and M
e Ann , satisfying M
eY
eX
e = I for X
e Ann and Y
with N
e and M factors are sometimes called the
of the coprime factorizations as fractional representations, the invertible M
denominator TFMs of the coprime factorization.


e is called a
e N,
e X, Y, X,
eY
Definition II.4. [10, Ch.4, Remark pp. 79] ( DCF) A collection of eight stable TFMs M, N, M,
e and M are invertible, yield the following factorizations:
doubly coprime factorization (DCF) of G over if M
e 1 N
e = NM1
G=M
and satisfy the following equality (Bezouts identity):

N
e
e M
e X

e
Y
X Y

M
= I2n .

(2)

To avoid excessive terminology, we refer to doubly coprime factorizations over simply as doubly coprime
factorizations (DCFs).


e be a DCF of G. Any stabilizing controller can
e N,
e X, Y, X,
eY
Theorem II.5. (Youla) [10, Ch.5, Theorem 1] Let M, N, M,
be written as:
1
KQ = YQ
XQ

(3)
DRAFT

eQ are defined as:

eQ , YQ and Y
for some Q in Ann , where XQ , X
XQ
eQ
X
YQ
eQ
Y

de f

e
X + QM

(4)

de f

e + MQ
X

(5)

e
Y QN

(6)

e NQ
Y

(7)

de f

de f

It also holds that KQ stabilizes G for any Q in Ann .



eQ in Theorem II.5 is also a DCF of G:
e N,
e XQ , YQ , X
eQ , Y
Remark II.6. The following identity shows that M, N, M,

(X

e + MQ) M
e
e
N
M

Q Ann
= I2n ,

(Y QN)
e NQ)
e (X + QM)
e (Y
N

(8)

E. The Closed Loop Map H(G, K)

Lemma II.7. ([10, (7)/ pp.101]) For any stabilizing controller KQ from (3) in Theorem II.5, it holds that

y (I + GKQ )1

=
u K (I + GK )1
Q
Q

(I + GKQ )1 G

(I + KQ G)1 w

(9)

and furthermore [10, (32)/ pp.107]

(I + GKQ )1

K (I + GK )1
Q
Q

(I + GKQ )1 G I NXQ
=

(I + KQ G)1 MXQ

(I + GKQ )1

K (I + GK )1
Q
Q

(I + GKQ )1 G
=

(I + K G)1

NYQ

MY

(10)

eQ N)
e
(I X

(11)

or equivalently

eQ M
e
Y
eQ M
e
X

eQ N
e
Y

F. The Platoon Motion Problem

We denote with n the number of cars in the platoon and with Gk , for 1 k n the linear and timeinvariant
dynamical model of each vehicle. More precisely, Gk is the transfer function from the controls uk (t) of each vehicle
to its position yk (t) on the roadway.
We denote with yo (t) the time evolution of the leader vehicles position. Note that yo (t) can be looked as the
reference for the entire platoon. The goal is for every car in the platoon to follow the leader while keeping
the prescribed intervehicle spacing. Note that is the constant intervehicle spacing policy i.e. the desired distance
separation between two consecutive vehicles (it is always a prespecified positive constant). We can write the time
evolution for the position of each vehicle as [4, (1)/ pp. 1836]


yk (t) = Gk (s) ? wk (t) + uk (t) k, for

1kn

(12)

where wk (t) are additive input disturbances to the kth vehicle.

Assumption II.8. (Additive Disturbances on the Plants Input) We make the assumption that while in motion, each vehicle
is affected by an additive disturbance to the input.
DRAFT

We denote the performance variables zk (t), (with 1 k n) to be the vehicle interspacing distances:

zk (t) = yk1 (t) yk (t) , for

1kn

(13)

Remark II.9. For illustrative simplicity and without any loss of generality we assume that = 0. We explicitly mention here,
there is no loss of generality in considering different vehicles with different lengths, which can be included in the spacing policies
as needed. These details are omitted here to improve the clarity of the exposition.
Definition II.10. Equation (13) represents the socalled constant spacing policy, however in practice a spacing policy that is
proportional with the speed is required, hence we denote
de f

H(s) = hs + 1,

h0

(14)

Remark II.11. We remark here that for h = 0 in (14) the dynamic timeheadway becomes a constant vehicle interspacing
policy and equal to one.
We also assume the standard conditions (zero error initial conditions [4, Section II])in order to obtain the
following time domain equations of the interspacing distances

z1 = yo HG1 ? (u1 + w1 )

zk+1 = Gk ? (uk + wk ) HGk+1 ? (uk+1 + wk+1 ),

de f

Definition II.12. Denote with z =


z1
de f

the vector of the controls, and with w =

z2

...


w1

w2

(15a)

1 k (n 1)

(15b)

T


de f
the vector of the interspacing distances, with u = u1 u2 . . .
T
the vector of the vehicles additive input disturbances.
. . . wn

zn

T
un

Definition II.13. Define the following n n TFM

de f

T = O

..
.

...

...

1
..
.

H
..
.

...

...

..
.

(16)

de f

Definition II.14. Denote with G1 = D(G1 , G2 , . . . , Gn ) the aggregated, multivariable model of the platoon. Next, denote with
de f

G = TD(G1 , G2 , . . . , Gn ) the coupled, multivariable TFM from the controls vector u to the error signals vector z.

G. The Closed Loop Maps of Interest

We rewrite (15) in matrix form, minding the meaning of G from Definition II.14 in order to obtain
DRAFT

z1


z2


z3 =

..
.


zn

HG1
1

G
O
1

O ? yo O

.
..
.
.
.

O
O

...

HG2

...

G2
..
.

HG3
..
.

...

...

...

(u1 + w1 )


O (u2 + w2 )


O ? (u3 + w3 )

..
..
.
.


(un + wn )
HGn
O

(17)

We further denote
de f

V1n =


1

T
(18)

such that (17) together with the controller equation u(t) = K ? z(t) we get [4, (10)/pp.1838]
z = V1n ? yo G ? w G ? u

(19a)

u=K?z

(19b)

1) The Transfer Function from yo and w to z: Plug (19b) in (19a) in order to get

1

1
z = I + GK V1n ? yo I + GK G ? w

(20)

which is similar to [4, (13)/ pp. 1839], the transfer function from the leaders coordinate yo and the disturbances w
respectively, to the interspacing errors z.
Proposition II.15. For any stabilizing controller KQ from (3) in Theorem II.5, it holds that

1

1
z = I + GKQ V1n ? yo I + GKQ G ? w

(21)

can be expressed as
eQ N
eQ MV
e n ? yo Y
e ?w
z=Y
1

(22)

2) The Transfer Function from yo and w to u: In order to minimize the control effort we might also be interested in
the transfer function from the reference trajectory yo or the disturbances w to the controls u. By plugging this time
(19a) in (19b) we get


u = KQ V1n ? yo G ? w G ? u
which yields
u = (I + KQ G)1 KQ V1n ? yo (I + KQ G)1 KQ G ? w

(23)

hence the following holds

Proposition II.16. For any stabilizing controller KQ from (3) in Theorem II.5, it holds that
u = KQ (I + GKQ )1 V1n ? yo KQ (I + GKQ )1 G ? w

(24)

eQ MV
e n ? yo X
eQ N
e?w
u=X
1

(25)

or equivalently

DRAFT

wk1
rk1

Fig. 2.

Kk1 (s)

wk

yk1

Gk1 (s)

rk

Kk (s)

Gk (s)

yk

Predecessor Following Control Scheme

H. The Predecessor Following Control Scheme

This control policy has been widely studied in the control literature. The control scheme is depicted in Figure II-G2
and it essentially implies that we constrain the controller K(s) in (19b) to be diagonal. For an extensive analysis of
this type of control schemes we refer to [5] and the references within. The fundamental drawback of the predecessor
following control scheme is that for any LTI controller it always leads to undesired effects such as string instability.

I. The Leader Information Control Scheme

In [4, (11)/ pp. 1838] homogeneous strings of vehicles are studied and particular attention is payed to the control
laws of the form uk = Kk (yo yk ), for 1 k n, where (yo yk ) = z1 + z2 + + zk is the relative distance of the kth
vehicle to the leader vehicle. The intuition behind it being the fact that the leader vehicles trajectory, yo (t) is basically
the reference for the entire string of vehicles. Writing such control policies in a compact form yields the following
relation

u1


u2


u3 =

..
.


un

K1

...

K2

...

O
..
.

O
..
.

K3
..
.

...
..
.

...

O
z1


O
(z1 + z2 )

O
(z1 + z2 + z3 )

..
..
.
.

Kn
(z1 + z2 + z3 + + zn1 + zn )

(26)

If we rewrite the controllers equation such that the vector of the measurements (i.e. the controllers input) is the
vector z then the expression of the controller becomes K = R(K1 , K2 , . . . , Kn ), or more explicitly

u1


u2


u3 =

..
.


un

K1

...

K2

K2

...

K3
..
.

K3
..
.

K3
..
.

...
..
.

Kn

Kn

Kn

...

O

..
.

Kn

z1

z2

z3

..
.

zn

(27)

Remark II.17. An excellent analysis of this control policy can be found in [4, (11)/ pp. 1838] for the scenario when all vehicles
are considered identical, all controllers are taken to be identical and a constant interspacing policy is considered. The TFM of
the type (27) controllers can be retrieved from [4, (12)/ pp. 1838], by taking Kp = 0 (control without predecessor information).
One of the nice features of this type of control policies is the excellent sensitivity to disturbances, specifically the fact that the
closedloop TFM from w to z is lower bidiagonal. This essentially means that any disturbance to the kth car in the string, will
DRAFT

10

only be felt by the car behind it (the (k + 1)th car). Another nice feature, is the particular manner in which the disturbances
to the leader (or yo (t)) propagate through the platoon, and the fact that a type (27) controller can achieve string stability.
In this work we introduce a generalization of the control law from [4, (11)/ pp. 1838], which we also dub leader
information controllers, due to the fact that our generalization inherits the closedloop features of the type (27) controller
from [4, (11)/ pp. 1838]. Our definitions allows us to handle the case of heterogeneous strings of vehicles (in which
the dynamical models of the vehicles differ up to a unimodular factor), and nonconstant time headways H(s)
Definition II.18. We say a controller KQ (s) Rnn (s) to be a leader information controller, if KQ internally stabilizes
G from Definition II.14 and the TFM of the closedloop map from the leader trajectory yo (t) to the error vector z(t), namely

1
In + GKQ
is a diagonal TFM.
Remark II.19. The TFM (In + GKQ )1 is relevant in platoon motion control, because it is the TFM through which the
leader trajectory yo (t) influences the error vector z(t). The conceptual connections between the class of controllers introduced in
Definition II.18 and controllers having the form in (27) will be made clear in the sequel.
III. Main Result

Definition III.1. A square TFM (s) A(s)

is said to be unimodular if it is proper, stable, invertible and its inverse 1 (s)
is proper and stable.
Proposition III.2. The n n TFM H(s)1 T(s) with H(s) from Definition II.10 and T(s) as defined in (16) is a unimodular
TFM.
The following key assumption will be in effect throughout this entire Section.
Assumption III.3. We assume throughout this section, that the dynamical model of each vehicle is equal to a strictly proper
de f

TFM Go (s) up to an unimodular factor k (s). More precisely, for any 1 k n it holds that Gk = k Go , for some k A(s)
(with 1
A(s)).
k
Remark III.4. Assumption III.3 allows us to model the different masses and the different actuating time constants corresponding
to different types vehicles in the platoon (e.g. heavy vehicles versus automobiles), according to the LTI model of the vehicles
from (1).
In particular, for the model in (1) with k 0
Gk (s) =

1
mk s2 (k s

+ 1)

for

1kn

(28)

Go (s) would be the double integrator 1/s2 .

Assumption III.5. We assume throughout this section that the TFM of Go (s) from Assumption III.3 also contains a conveniently
designed Pade rational approximation, which is the same for all n vehicles in the platoon. This assumption is made as to take
into account an actuation time delay es which is considered to be the same for all n vehicles in the platoon. We remark here
that the delay is known from vehicle specifications and it can be verified through model validation techniques.
Remark III.6. As a consequence of Assumption III.5 Go (s) is always a strictly proper TFM. We further remark here that the
Pade approximation will introduce nonminimal phase zeros in Go (s).
DRAFT

11


eo, N
eo , Xo , Yo ,
Definition III.7. We compute a DCF (2) of the scalar rational function Go (s) and denote it with Mo , No , M

eo . Such a DCF can be computed using standard methods as in [12]. Note that all eight factors of the DCF are scalar
eo , Y
X
eo and No being strictly proper (since Go is strictly proper).
rational functions, with N
de f

Proposition III.8. Denote as before with G1 = D(G1 , G2 , . . . , Gn ) the aggregated model of the platoon. Note that in view of
de f

Assumption III.3 it follows that G1 = D(1 , 2 , . . . , n )Go . Denote the diagonal, unimodular matrix = D(1 , 2 , . . . , n )
and its inverse with 1 , respectively. Then the following expression is a DCF of G1

N
e1

Y
1

e 1 X
e1
M


e1
X1 Y

M1
= I2n .

(29)



e1 are given by
e1, N
e1 , X1 , Y1 , X
e1 , Y
where M1 , N1 , M

N
e1

Y
1

X
e1

Y
e1

e 1 de f
M

=
X
1

M1 de f
=

N
1

N
eo

Y
o

X
eo 1

Y
eo In

e o In
M

X I

(30a)

Mo 1

NI

(30b)

o n

o n

Proof: The proof is an immediate consequence of [10, Theorem 43/pp. 75].

Proposition III.9. Given the time headway H from Definition II.10, the TFM T as defined in (16), the plant G = TG1 as
defined in Definition II.14 with G1 as defined in Proposition III.8 and the DCF (30) of G1 from Proposition III.8 then we can


e satisfying G = M
e N,
e X, Y, X,
eY
e 1 N
e = NM1 and also:
construct a DCF of G denoted M, N, M,

N
e M
e X
e M

(31)
= I2n

e
Y
X
Y
N


e
e N,
e X, Y, X,
eY
where the doubly coprime factors are given by M, N, M,

N
e

X
e

Y
e

e de f
M

=
X

M de f
=

N
eo T

H1 Y T
o

1 T1 X
eo

eo In
Y

1
H X o In
e o In
M

(32a)

1 T1 HMo

HN I

(32b)

o n

Proof: The proof follows by Proposition III.8 and the inspection of (32).
The Youla parameterization represents in Linear Control Systems the core theoretical result for controller design
(H 2 or H optimal control, robust control etc.) The following Theorem provides the Youla parameterization of all
leader information controllers (as defined in Definition II.18) associated with a given platoon of vehicles.
Theorem III.10. Given the TFM G(s) from Proposition III.9 the Youla parameterization for the set of all leader information
stabilizing controllers (as in Definition II.18) can be obtained from the DCF (32) by restricting the Youla parameters Q Ann
de f

de f

to be diagonal, Q = D(Q1 , Q2 , . . . , Qn ), with Qk A(s), for any 1 k n. Given a diagonal, Q = D(Q1 , Q2 , . . . , Qn ), with
Qk A(s), for any 1 k n then any leader information controller KQ is given by a left coprime factorization of the form


 


 
de f
eo Q T
eoQ
(33)
Y Q XQ =
H1 Yo In N
H1 Xo In + M

DRAFT

12

where YQ and XQ have the following expressions

eo )1
(Yo Q1 HN
O

(H1 Y + Q N
e
eo )2
(Yo Q2 HN
o
2 o )1

eo )2
O
(H1 Yo + Q3 N
YQ =

..
..

.
.

O
O

eo)
(H1 Xo + Q1 M

O
XQ =

..

...

...

eo )3
(Yo Q3 HN
..
.

...
..
.

O
..
.

...

eo )n
(Yo Qn HN

...

eo)
(H1 Xo + Q2 M

...

O
..
.

eo)
(H1 Xo + Q3 M
..
.

...
..
.

O
..
.

...

eo)
(H1 Xo + Qn M

(34a)

(34b)

Proof: Given the DCF in Proposition III.9 it can be seen that (In + GKQ )1 is diagonal if and only if (In +
eQ M
e o In )1 is diagonal (M
e o is not identically zero). Since (In + GKQ )1 = Y
e (see Lemma II.7) it follows
GKQ )1 (M
eQ =
e o In )1 is diagonal if and only if YQ is diagonal. But from Proposition III.9 it follows that Y
that (In + GKQ )1 (M
eQ is diagonal if and only if the Youla parameter Q(s) is
eo In + (HNo )Q and since neither No nor H are zero, clearly Y
Y
a diagonal TFM.
Corollary III.11. Any of the leader information controllers parameterized in Theorem

K1
u1
u1

O
O
O
.
.
.
O
O

O
2 1
O
O
...
O
O u2

u2

1
u3

O
.
.
.
O
O
u
O

2
3

1 O
3

? . +H
.. = H
O
O
O
1
3 . . .
O
O ..
.

..
..
..
.
.

.
..

.
.
.
un1

.
.
.
.
.
.
. un1

O
O
O
.
.
.

O
u
n1
n
n
O
n

III.10 can be factored as

O

...

K2

...

K3

O
..
.

O
..
.

...
..
.

O
..
.

...

..

O

?

Kn

z1
z2
z3
..
.
zn1

zn
(35)

where for any 1 k n


1
de f
1
e
eo)
H1 Kk = H1 1
(H1 Xo + Qk M
k H Y o Qk No

(36)


1
de f
e
eo)
Kk = 1
(Xo + Qk HM
k Y o Qk H No

(37)

or equivalently

Proof: It follows by multiplication to the left of the coprime factorization (33) of KQ with the diagonal TFM




1
1
1 
1
1
eo , 1 H1 H1 Yo Q2 N
eo , . . . , 1
eo
D 1
H1 Yo Q1 N
H1 Yo Qn N
n H
2
1 H

DRAFT

13

y3

G3

u3

K3

y2

z3

u2

G2

K2

G1

u1

y0

z1

K1

+
1
2
3

Fig. 3.

y1

z2

1
1
2

Leader Information Control

u1
u2
u3
..
.
un1
un

O


1
2 1



O

=

O


..

.


O

...

...

2
1
3

...

O
..
.

1
3
4

...
..
.

O
..
.

...

1
n n1

..
.

O


O


O
?

O

..
.


O

u1
u2
u3
..
.
un1
un

K1

...

K2

...

K3

O
..
.

O
..
.

O
..
.

...
..
.

O
..
.

...

..

O

?

O


O

Kn

z1
z2
z3
..
.
zn1
zn

Remark III.12. A decentralized implementation of the leader information controller from Theorem III.10, according to (35) is
presented in Figure 3 for a three vehicle platoon. The scheme for any two consecutive vehicles in the platoon (k 2) is depicted
in Figure 4. To make the graphics more readable we illustrate the case in which H(s) = 1. The implementation for H(s) = hs + 1
with h > 0 should become straightforward from equation (35) (simply add a cascaded H1 (s) filter on each uk and zk+1 branch).

yk+1

Gk+1

uk+1

Kk+1

zk+1

yk

Gk

+
uk+1

k+1 k

uk

uk

Kk

zk

yk1

1
k1
k

uk1

Wired/Lidar
Wireless
Fig. 4.

Leader Information Control

DRAFT

14

A. Description of the Control Scheme from Figure 3 and Figure 4

The blocks Gk (with 1 k n) represent the dynamical models of the vehicles and each control signal uk is
fed in the Electrohydraulic Braking and Throttle actuation system on board the kthe vehicle. The zk measurement,
represents the distance to the preceding vehicle which is measured using ranging sensors on board the kthe vehicle.
The control signal uk1 produced on board the (k 1)th vehicle is broadcasted (e.g. using wireless communications)
k1 and H1 Kk are the linear filters introduced in Corollary III.11, and they
to the kth vehicle. The blocks H1 1
k
can be implemented using digital computers (digital controllers) on board the kth vehicle. The controllers which
k1 and H1 Kk use the signals uk1 and zk , respectively (available on board the kth
implement the filters H1 1
k
car) in order to produce the control signal uk , which is fed into the Electrohydraulic Braking and Throttle actuation
system on board the kthe vehicle
Remark III.13. The particular type of structure featured by the controller in (35) has been investigated in [15], [16] on the
basis of the socalled signal structure of a LTI network as introduced by Goncalves & Warnick in [17]. One of the examples
given in [15], [16] is a ring network of LTI systems, while the controller in (35) features a line structure (in fact a ring
with the link between agents 1 and n broken). The analysis is made from the perspective of establishing the connections between
all the left coprime factorizations associated with a certain TFM and all possible signal structures associated with the same
TFM.
Remark III.14. The so called indirect leader broadcast scheme from [14] studied for homogeneous strings of vehicles presents
certain similarities with our leader information controller from Theorem III.10, with the important distinction that in our leader
information controller from Corollary III.11 we broadcast the control signal of the predecessor vehicle instead of an estimate
of the leaders state (as in [14]). The control signal is basically generated on board of the predecessor vehicle, hence there is
no need to estimate it and the fact that information free of noise is broadcasted (with some timedelay , as explained later)
has profound implications in terms of the performance of the closed loop. Furthermore, the leader information control scheme
from Corollary III.11 can be adapted as to compensate for the feedforward timedelay induced by the wireless communication
broadcasting of the predecessors control signal, as will be explained later in full detail.
Remark III.15. In [8] a control scheme very similar to the leader information controller is presented, in which the predecessors
control signal is broadcasted. However, because the scheme is implemented taking into account a Pade approximation of the
feedforward timedelay induced by the wireless communication, the resulting controller from [8, Section V] will never be (in
general) a leader information controller, as we argue in the sequel. The argument is related to the way the broadcasting
delays propagate through the string of vehicles (phenomenon which is studied in detail in [14]).
Remark III.16. For an analysis of the important advantages offered by the leader information controller scheme both with
respect to string stability and sensitivity to disturbances we refer to [4].

DRAFT

15

B. An Additional Perspective on Definition II.18

For the current Subsection III-B only, let us assume (for simplicity) that H(s) from Definition II.10 is equal to 1




de f
and denote e
T = B b1 , b2 , . . . , bn where b1 = b2 = = bn = 1. Note that e
T1 = R b1 , b2 , . . . , bn We remark that by


premultiplying (35) to the left with D 1 , 2 , . . . , n we get

1 ? u1

? u

2
2

3 ? u3

..

n1 ? un1

?u
n
n

...

...

...

O
..
.

O
..
.

3
..
.

...
..
.

O
..
.

...

n1

O


O


O
?

O

..
.

O

u1
u2
u3
..
.
un1
un

1 K1


O


O


+
O


..


.


O

...

2 K2

...

3 K3

O
..
.

O
..
.

...
..
.

O
..
.

...

n Kn

..

z1
z2
z3
..
.
zn1

zn
(38)

where Kk are as in (37) for any 1 k n and is further equivalent to



1  
 

u= e
T D 1 , 2 , . . . , n
D 1 , 2 , . . . , n D K1 , K2 , . . . , Kn ? z

(39)


1
 
 

u = D 1 , 2 , . . . , n e
T1 D 1 , 2 , . . . , n D K1 , K2 , . . . , Kn ? z

(40)



1

u = D 1 , 2 , . . . , n e
T1 D 1 K1 , 2 K2 , . . . , n Kn ? z

(41)

and equivalently that

and equivalently that

and now consider that all the diagonal entries of the (diagonal) TFM of the Youla parameter are equal, that is Qk = Q
in the definition of Kk (s) from (37) for all 1 k n, for some stable TFM Q (s).



1

1

1

e
e
e o ), 1 Yo Q N
eo (Xo + Q M
e o ), . . . , 1
eo) ? z
u = R 1
(Xo + Q M
(Xo + Q M
n Y o Q No
2
1 Yo Q No

(42)

which is a type (27) controller.

C. Tuning the Leader Information Controller for Performance
The Youla parameterization from Theorem III.11 represents the theoretical result that can be used for a systematic
design of leader information controllers for a platoon of vehicles. The parameterization in Theorem III.11 can be
coupled with H or H 2 optimal design methods from the standard case ([13]) to achieve the desired closedloop
performance.
The following result is a Corollary of Theorem III.11
Corollary III.17. Given the DCF of G from Proposition III.9 then for any of the leader information controllers parameterized
in Theorem III.11, the transfer function from the exogenous signals yo and w to the internal loop signals of interest z and u is
given by
e NQ)MV
e NQ)N
e n ? yo (Y
e ?w
z = (Y
1

(43a)
DRAFT

16

e + MQ)M
eV
en ? yo (X
e + MQ)N
e?w
u = (X
1

(43b)

de f

where Q = D(Q1 , Q2 , . . . , Qn ), with Qk A(s), for any 1 k n is the diagonal Youla parameter.
Proof: It follows from Theorem III.10 and Proposition II.15 and Proposition II.16.
There are many ways to tune the leader information controller, depending on the performance objective that needs
to be achieved. One possible performance objective could be to avoid any amplification of yo and of the additive
disturbances w along the platoon. Another could be to avoid the effect of disturbance wk to the k-th vehicle to be
amplified at zk+1 we could solve for Qk and Qk+1 (look at the k-th column on the right hand side of (43a)


(Y
eo Qk+1 HNo )N
eo k
zk+1

min
= Q ,QminA(s)
zk
eo Qk HNo )N
eo k
k k+1
H(Y

(44)

and then plug the optimal Qk , Qk+1 back into Corollary III.11 to obtain the expressions for Kk and Kk+1 . If Qk is already
fixed in (44) then we can solve (44) for Qk+1 .
The H cost from (44) is just one example of performance criterion.It is common in the literature because it is related
with the so-called string instability phenomenon. Alternatively, the cost function that attenuates the amplification of
the control signal along the platoon such as kuk+1 /uk k could be employed. In [8] a mixed criterion is used that
minimizes the H norm of the TFM from uk to both uk+1 and zk . We note however, that for the case of homogeneous
strings of vehicles (as it is the case in [8]), the feedforward term 1
can always be made equal to 1 for any leader
k+1 k
information controller.

DRAFT

17

yk+1

Gk+1

uk+1

Kk+1

yk

zk+1

uk

Gk

+
uk+1

uk

1
k+1 k

es

zk

Kk

yk1

k1
1
k

es

uk1

Wired/Lidar
Wireless
Fig. 5.

u1
u2
u3
..
.
un1
un

Leader Information Control with Feedforward Time Delay

O


1 es
2 1



O

=
O


..


.


O

...

...

1
2 es
3

...

O
..
.

1
3 es
4
..
.

...
..
.

O
..
.

...

s
1
n n1 e

O


O


O
?

O

..
.

O

u1
u2
u3
..
.
un1
un

K1

...

K2

...

K3

O
..
.

O
..
.

O
..
.

...
..
.

O
..
.

...

..

O

?

Kn

z1
z2
z3
..
.
zn1

IV. Dealing with feedforward timedelays

In this section we look at the scenario encountered in practice when we must deal with a non zero time delay on
the feedforward links uk , for any 2 k (n 1). This is due to the fact that the feedforward link is implemented
via a wireless communication system whose physical limitations imply that there will always exist a time delay es
when broadcasting the uk signal, for any 2 k (n 1) as pictured in Figure 5.
We point out here that if any of the leader information controllers parameterized in TheoremIII.10 and Corollary III.11 is implemented using a wireless channel with a nonzero time delay es for the feedforward link (as in
Figure 5 and the subsequent equation), then the equivalent TFM from z to u of the controller that results from that
implementation no longer satisfies Definition II.18. Furthermore, the effects of the timedelays could fundamentally
alter the closed loop performance.

DRAFT

zn

18

yk+1

uk+1

Gk+1

yk

zk+1

Kk+1 Rk+1

uk

Gk

+
uk+1

es

uk

zk

Kk Rk

es

yk1

uk1

Wired/Lidar
Wireless
Fig. 6.

u1
u2
u3
..
.
un1
un

...

es

...

...

O
..
.

O
..
.

es
..
.

...
..
.

O
..
.

...

es

O


O


O
?

O

..
.

O

u1
u2
u3
..
.
un1
un

K1 R1


O


O


+
O

.
.
.


O

...

K2 R2

...

K3 R3

O
..
.

O
..
.

...
..
.

O
..
.

...

K n Rn

..

z1
z2
z3
..
.
zn1
zn

A. Feedback Path Compensation for Communication Delays

We will prove next that any nonzero time delay on the feedforward link uk , for any 2 k n, cannot be
compensated by any LTI filter on the feedback link Kk+1 with 1 k n 1 such as to yield a leader information
controller. (We consider the a delay on the control signal un , although the last car in the platoon need not broadcast
to anybody.) For illustrative simplicity we will consider H(s) = 1 and we will consider that all plants to be identical,
i.e. k = 1, for all 1 k n. We add on each feedback path Kk (s), 1 k n a LTI filter compensation factor Rk (s).
Note that the controllers equation becomes

u1
u2
u3
..
.
un1
un

1


es


O


=
O


..


.


O

...

...

...
..
.

O
..
.

es
..
.

..

...

O
..
.
es

1
O K1 R1


O O


O O


O O

.. ..
. .


O
1

...

K2 R2

...

K3 R3

O
..
.

O
..
.

...
..
.

O
..
.

...

K n Rn

..

z1
z2
z3
..
.
zn1
zn

(45)

or equivalently
DRAFT

19

u1
u2
u3
..
.
un1
un

1


e1 (s)


e0 (s)


= (1) (s)
e


..


.

(n3) (s)
e

...

...

1 (s)

e0 (s)
..
.

e1 (s)
..
.

...
..
.

e(n4) (s)

e(n5) (s)

..

...

O
..
.
e1 (s)

O K1 R1

O O

O O

O O

.. ..
. .

1
O

...

K2 R2

...

K3 R3

O
..
.

O
..
.

...
..
.

O
..
.

...

K n Rn

..

z1
z2
z3
..
.
zn1
zn

(46)
Remark IV.1. Clearly, from (46) for any nonzero , the resulting TFM from z to u is not a leader information controller,
irrespective of the chosen LTI compensation factors Rk (s), with 1 k n. Furthermore, we remark that the timedelay propagates
through the controller along the string of vehicles and the effect depends on the length of the platoon. An in depth analysis of
how the feedforward communication delays es propagate through the platoon is made in [14].
Remark IV.2. We remind here the basic fact from control theory that the delay es (on any of the feedforward channels uk ,
with 1 k n) cannot be efficiently compensated by a series connection on the feedforward path, for example with a rational
approximation of the anticipative element es , that would cancel out the effect of the delay.
Remark IV.3. For the method proposed in [8], the communication delay es is firstly approximated by a suitably chosen
Pade approximation denoted with D(s). Next, the model containing this Pade approximation is ran through a H synthesis
routine that produces the optimal feedback and feedforward controllers (for a certain costfunction defined in [8]). As stated in
Remark IV.1, the LTI controller (resulted from [8, Section V]) placed on the feedback path cannot compensate for the es delay
such as to produce a leader information controller. Furthermore, according to Remark IV.2 the time delay Pade approximation
D(s) cannot simply be canceled out by a cascade connection with the feedforward controller (containing a D1 (s)). Consequently,
the H method from [8] will indeed return an internally stabilizing controller, that minimizes the corresponding H norm
based cost function, but it will not be a leader information controller. Reference [9] also contains an analysis that takes into
account the effect of sampling and of the time delay induced by the wireless broadcasting.

DRAFT

20

yk+1

Gk+1 e

uk+1

Kk+1 e

zk+1

yk

uk

Gk es

+
uk+1

uk

1
k+1 k

es

Kk es

zk

yk1

k1
1
k

es

uk1

Wired/Lidar
Wireless
Fig. 7.

Leader Information Control with Communication Delays Compensation

B. New Control Mechanism Using High Accuracy Synchronization

We propose a novel control mechanism for an efficient implementation of the leader information controllers from
the set parameterized in TheoremIII.10 and Corollary III.11. Our control scheme is designed such that it can cope
efficiently with the nonzero time delays es induced by the wireless channel.
In Figure 7 a non zero time delay is present on the feedforward links uk , for any 2 k (n 1), because the
feedforward link could be implemented via a wireless communication system whose physical limitations imply that
there will always exist a time delay es when at the receiver (in the (k + 1)th vehicle) of the uk signal, for any
2 k (n 1). We place a delay of exactly seconds on each of the sensor measurements zk , which in turn appears
as a es factor in the TFM Kk+1 (s), for any 1 k n 1. Having a delay es on both uk and zk+1 is equivalent with
having a es delay in the model of the plant Gk+1 , for any 1 k n 1. The argument for this is the controller
equation for the scheme in Figure 7, given in (47). This time delay, will be considered to be part of the plant model,
when designing the controller Kk+1 as specified in detail in the sequel.

u1
es ? u2
es ? u3
..
.
es ? un1
e

? un

1 es

2 1

1
= H

..

...

...

1
2 es
3

...

O
..
.

1
3 es
4

...
..
.

O
..
.

...

s
1
n n1 e

..
.

O


O


O
?

O

..
.


O

u1
u2
u3
..
.
un1
un

DRAFT

21

1
+H

K1

...

K2 es

...

...
..
.

O
O

O
..
.

...

Kn es

K3 e

O
..
.

O
..
.

O
..
.

..

z1
z2
z3
..
.
zn1
zn

(47)

Note that (47) is exactly the equation for the leader information controller from Corollary III.11 premultiplied with
the diagonal TFM D(1, es , es , . . . , es ). It is important to mention that the controller as given in (47) acts on the
es delayed version of the plant that the controller as given in (35) acts on in the statement of Corollary III.11.
C. Leader Information Controller in the Presence of Time Delays Controller Design Algorithm
Step 1 We include the es wireless communication time delay into the model of the vehicle, in a manner made
precise by the following assumptions.
Assumption IV.4. We assume from this point onward, that the dynamical model of each vehicle is equal to a strictly proper
de f

TFM Go (s) up to an unimodular factor k (s). More precisely, for any 1 k n it holds that Gk = k Go , for some k A(s)
(with 1
A(s)).
k
Assumption IV.5. We assume from this point onward that the TFM of Go (s) from Assumption III.3 also contains a conveniently
designed Pade rational approximation for the factor e(+)s , which is the same for all n vehicles in the platoon. This assumption
is made as to take into account for the actuation system time delay es (e.g. Electrohydraulic Braking and Throttle actuation
system) and for the nominal delay induced by the wireless communication system es ( assumed to be constant), which
are both considered to be the same for all n vehicles in the platoon. We remark here that the delay is known from vehicle
specifications and it can be verified through model validation techniques. The nominal value of the delay is also known from
the wireless communication systems technical specifications.
Step 2 Design offline the leader information controllers Kk (s), with 1 k n. For the plant satisfying Assumptions IV.4 andIV.5 we compute a DCF using Proposition III.9. Then, note that for any diagonal Youla parameter in
Theorem III.10 we obtain an internally stabilizing controller. If further performance is required, such as guarantees
related to string stability, then a supplemental design procedure as suggested in Subsection III-C must be performed
in order to compute an adequate diagonal Youla parameter. The controller that results by plugging in the obtained
diagonal Youla parameter into Theorem III.10 will admit a factorization of the form given in Corollary III.11.
Step 3 Each of the controllers Kk (s), with 1 k n obtained at Step 2 along with the k (1 k n) factors will be
used for the implementation of the feedback loop as in Figure 7, where a delay es is added on each measurement
zk+1 .

DRAFT

22

yk+1

uk+1

Gk+1

zk+1

Kk+1

yk

uk

Gk

+
uk+1

1
k+1 k

es

Kk

zk

yk1

k1
1
k

es
GPS clk

uk1
GPS clk

Wired
Wireless
Fig. 8.

Leader Information Control with Synchronization Mechanism

zk+1

zk+1 es

uk es

es

uk

GPS clk

Fig. 9.

Description of the Block Components used in FIgure 8

Because the communication delay if not properly compensated can cause serious degradation of the closed loop
performance, in our scheme we consider that if possibleis not desirable to approximate the delay es on the
feedforward path with a linear filter that approximates es (such as a Pade approximation) to be incorporated both
in the Kk+1 (s) and Gk+1 (s)
Considering that Assumption IV.5 is enforced, for a sound implementation of the scheme in Figure 7 we will
use high accuracy GPS clock synchronization in order to evaluate very precisely the delay on each channel uk
(with 1 k n) and also to delay the sensor measurement zk+1 (t) accordingly. In practice, although the delay
induced by the wireless communication can be well evaluated (about 20 milliseconds), is actually a time varying
value, featuring relatively small variations around its nominal value. Our implementation can cope well with this
hypothesis.

D. TimeDelay Implementation using Synchronous Clocks

Assumption IV.6. We assume that all continuous time LTI filters involved in our design, will be implemented as discrete
time LTI filters, using for instance zero order hold elements and a sufficiently high sampling frequency. This is done in order
to ensure operable interfaces with the digital communications systems.
Each vehicle in the platoon is equipped with both the devices sketched in Figure 9. The following mechanism
described below for the (k + 1)th vehicle is implemented in each vehicle, for 2 k n. The device sketched in
DRAFT

23

Figure 9, placed in the (k + 1)th vehicle uses a Precision Time Protocol (e.g. PTP/IEEE-1588, Grandmaster with
Hardware Timestamping with GPS synchronizetion) or something similar or equivalent to detect accurately (for
each sample of the transmitted signal uk ) the delay seconds with which each sample of the signal uk (transmitted
from the kth vehicle) arrives at the receiver on the (k + 1)th vehicle. Each sample of the measurement signal zk+1
(measured with sensors onboard the (k + 1)th vehicle) is delayed by the device sketched in Figure 9, part (a) with
the exact same seconds (determined at the receiving of the current sample of uk ) and then it is used to compute
the sample of the control signal uk+1 .
The general scheme is presented in Figure 8, where the blocks Kk and k with 1 k n, have the same meaning
as those described in the algorithm from Subsection IV-C.
V. Conclusions
We have introduced a generalization of the concept of leader information controller for a non homogeneous platoon
of vehicles and we have provided the Youla parameterization of all the such stabilizing controllers. The nice feature
of the leader information controller scheme is that it allows for an implementation where the controller placed on
each vehicle uses only locally available information. We also presented a scheme that accurately compensates for the
time delay introduced by the wireless broadcasting of information, such as to preserve the leader information controller
scheme.
References
[1] W. S. Levine and M. Athans On the Optimal Error Regulation of a String of Moving Vehicles, IEEE Trans. Aut. Control, Vol.
11, No. 3, 1966. (pp. 355361)
[2] L. Peppard String Stability of Relative Motion PID Vehicle Control Systems, IEEE Trans. Aut. Control, Vol. 19, No. 5, 1974.
(pp. 579581)
[3] D. Swaroop and J. K. Hedrick String Stability of Interconnected Systems, IEEE Trans. Aut. Control, Vol.41, No.3, 1996.
(pp. 349357)
[4] P. Seiler, A. Pant and J. K. Hedrick Disturbance Propagation in Vehicle Strings, IEEE Trans. Aut. Control, Vol.49, No.10, 2004.
(pp. 18351841)
[5] R. H. Middleton and J. Braslavsky String Stability in Classes of Linear Time Invariant Formation Control with Limited
Communication Range, IEEE Trans. Aut. Control, Vol.55, No.7, 2010. (pp. 15191530)
[6] M.R.I. Nieuwenhuijze, String Stability Analysis of Bidirectional Adaptive Cruise Control, Master Thesis, Eindhoven University
of Technology, 2010.
[7] G. Naus, R. Vugts, J. Ploeg, vd Molengraft and M. Steinbuch, StringStable CACC Design and Experimental Validation: A
Frequency Domain Approach, IEEE Transactions on Vehicular Technology, Vol. 59, pp. 42684279, November 2010.
[8] J. Ploeg, D.P. Shukla, N.vd Wouw and H. Nijmeijer, Controller Synthesis for String Stability of Vehicle Platoons, IEEE
Transactions on Intelligent Transportation Systems, Vol. 15, pp. 854865, April 2014.
[9] S. Oncu, J. Ploeg, , N. vd Wouw and H. Nijmeijer, Cooperative Adaptive Cruise Control: Network Aware Analysis of String
Stability, IEEE Transactions on Intelligent Transportation Systems, Vol. 15, No.4, pp. 15271537, August 2014.
[10] M. Vidyasagar Control System Synthesis: A Factorization Approach, MIT Press, Signal Processing, Optimization, and Control
Series, 1985.
[11] B. Francis A Course in H Control Theory, Series Lecture Notes in Control and Information Sciences, New York: Springer-Verlag,
1987, vol. 88.
V. M. Lucic Doubly Coprime Factorization Revisited, IEEE Trans. Aut. Control, Vol.46, No.3, 2001. (pp. 457-459)
[12] C. Oara, A. Varga Minimal Degree Factorization of Rational Matrices, SIAM J. Matrix Anal Appl., Vol. 21, No.1, 1999. (pp.
245-277)

DRAFT

24

[13] K. Zhou, J.C. Doyle and K. Glover, Robust and Optimal Control, Upper Saddle River, NJ: Prentice Hall, 1996.
[14] A.A Peters, R.H. Middleton and O. Mason, Leader tracking in homogeneous vehicle platoons with broadcast delays
Automatica. Vol. 50, 2014. (pp. 6474)
[15] S. Sabau, C. Oara, S. Warnick, A. Jadbabaie A Novel Description of Linear and TimeInvariant Networks via Structured
Coprime Factorizations http://arxiv.org/abs/1303.2760
[16] S. Sabau, C. Oara, S. Warnick and A. Jadbabaie Structured coprime factorizations description of Linear and Time-Invariant
networks. Decision and Control (CDC), 2013 IEEE 52nd Annual Conference on. IEEE, 2013.
[17] J. Goncalves, S. Warnick Necessary and Sufficient Conditions for Dynamical Structure Reconstruction of LTI Networks IEEE
Trans. Autom. Control, vol. 53, pp. 16701674, 2008.

DRAFT