http://uu.diva-portal.

org




This is an author produced version of a paper presented at 3
rd
IFAC
Workshop on Distributed Estimation and Control in Networked Systems
(Netsys 2012), September 14-15, Santa Barbara, California, USA. This
paper has been peer-reviewed but may not include the final publisher
proof-corrections or pagination.

Citation for the paper:
Nygren, Johannes, Pelckmans, Kristiaan
A cooperative decentralized PI control strategy: discrete-time analysis
and nonlinear feedback"
In: 3rd IFAC Workshop on Distributed Estimation and Control in
Networked Systems, Santa Barbara, CA, 2012

A Cooperative Decentralized PI Control Strategy:
Discrete-time Analysis and Nonlinear Feedback
Johannes Nygren

Kristiaan Pelckmans

Division of Systems and Control, Department of Information Technology,

Uppsala University, Uppsala, Sweden.
Abstract: This paper discusses an extension of a PI control strategy towards the control of a large
m m-MIMO system. This strategy is fully decentralized, it requires only the tuning of m different
controllers, while we only allow for neighboring controllers to exchange error signals. This makes it a
strong candidate for an implementation on a decentralized, low-power and high performance Wireless
Sensor Network (WSN). The main idea is to feed locally observed control errors (feedback) not only
into the local control law, but also in a xed proportionate way into neighboring controllers. The analysis
concerns convergence to a set point. The analysis is essentially based on a conversion of the PI control
law into a discrete-time gradient descent scheme. As an interesting byproduct, this analysis indicates
how to deal with quantization functions and nonlinear effects in the feedback signals.
The problem of designing and analyzing control strategies us-
ing WSNs has been investigated in some depth during the last
decade, see e.g. Bemporad et al. (2010) for a survey. For a per-
spective of automatic control, see e.g. Hespanha et al. (2007).
For a discussion of issues of communication and design, see
e.g. Fischione et al. (2010). A prominent theoretical theme
until date concerns stability of the resulting control law, e.g.
Elia and Mitter (2001), while studies of communication issues
often involves techniques of quantization (as in Brockett and
Liberzon (2000)) and (stochastic) delay (see e.g. Cloosterman
et al. (2010)).
In practice, the most prominent way is to use non-cooperative
PI(D) controllers implemented on different parts of the MIMO
system, see e.g. Palmor et al. (1995). It is indicated that the
performance of this approach can be improved considerably by
optimization of the loop-pairing of input-output signals which
are fed with one univariate control law, see e.g.

Astr om and
H agglund (2006). This approach is represented in Figure (1).
This approach has the advantage of being simple to implement,
and to be fully decentralized. However, the rigid structure
might not be appropriate for control of strongly interconnected
systems.
At the other end of the spectrum, one can nd multivariate con-
trol laws which control the MIMO system based on techniques
of optimal control design or Model Predictive Control (MPC).
Such approaches are in general computationally demanding,
while they require a centralized node deciding on the control us-
ing all information which is gathered fromthe individual nodes.
Despite such strong requirements, this approach is most preva-
lent in modern distributed control as in Maciejowski (1989).
We are interested in control strategies which strike a good
trade-off between the centralized and decentralized extremes.
That is, we desire not to be bound by the univariate nature
of the decentralized approach, while the application areas in
mind (in WSNs) offer neither the computational resources nor
the centralized capabilities which are in general required for

This work was supported in by the Swedish Foundation for Strategic Re-
search (SSF), project number RIT08-0065.
u
1
(k)
z
1
(k) G
11
(q) r
1
(k)
u
2
(k)
z
2
(k)
G
22
(q)
r
2
(k)
G
12
(q)
G
21
(q)
PI
1
(q)
PI
2
(q)
Fig. 1. The traditional design of a decentralized PID control
strategy for a 2 2 MIMO system, with (u
1
, z
1
) and
(u
2
, z
2
) paired.
centralized control. The approach adopted here is to take an
array of m (P)I control laws and apply them to the m input-
output signal pairs of the MIMO system at hand. Then we nd
that a simple modication which allows neighboring controllers
to exchange feedback may give us a main boost in performance.
This approach is given in Fig. (2). The main contribution of this
paper is to provide convergence conditions of the control errors,
for a properly chosen diffusion matrix D.
The techniques that we use to analyze such PI rules are based
on a reduction of the control law into a gradient descent rule as
in Nygren et al. (2012). This is achieved by the construction of
a certain matrix Q
m
. The benet of doing so is that it gives
us access of the wealth of techniques available for gradient
descent. Consequently, the result is quite general, and covers
many interesting cases, including dealing with Wiener systems,
quantization functions and the design of energy-efcient con-
trol laws. The other benet of this technique is that the result
concerns the total control error. It is noteworthy that a bound
of this performance measure implies (BIBO) stability as well.
Finally, we like to point out that this approach makes it possible
to compare tuning of PI control laws with techniques of setting
the gain in techniques of stochastic approximation, adaptive
ltering and optimization.
This paper is organized as follows. The following section
details the proposed control strategy, and exemplies a few
u
1
(k)
z
1
(k) G
11
(q) r
1
(k)
u
2
(k)
z
2
(k)
G
22
(q)
r
2
(k)
G
12
(q)
G
21
(q)
PI(q)
PI(q)
d
12
d
21
Fig. 2. An example of a collaborative decentralized PI con-
trol strategy, where feedback (errors) are distributed
to neighboring controllers as well. That is, errors are
rescaled by a factor d
ij
if going from controller i to
neighboring j. To keep the gure as simple as possible,
we assume here that d
11
= d
22
= 1, and that f(x) = x.
interesting cases. Section 3 gives a discrete-time analysis by a
reduction to a gradient descent scheme. The technical material
on which this result hinges is presented in Appendix A and B.
Section 4 gives conclusions and points to further work. The
following notational conventions are used. We denote scalars
as lowercase letters, vectors are denoted as lower-case boldface
letters and matrices are denoted as upper-case, boldface letters.
A vector (semi-) norm is denoted as x
Q
=
1
2
x
T
(Q
T
+Q)x,
for a vector x and a matrix Q such that Q
T
+ Q is positive
(semi)denite (i.e. Q
T
+Q 0).
1. PRELIMINARIES
1.1 Setting
Consider the following linear MIMO system with m inputs and
m outputs:
_

_
z
1
(k)
z
2
(k)
.
.
.
z
m
(k)
_

_
=
_

_
G
11
(q) G
12
(q) . . . G
1m
(q)
G
21
(q) G
22
(q) . . . G
2m
(q)
.
.
.
.
.
.
.
.
.
G
m1
(q) G
m2
(q) . . . G
mm
(q)
_

_
_

_
u
1
(k)
u
2
(k)
.
.
.
u
m
(k)
_

_
,
(1)
where u
i
(k) is the ith input (manipulated variable) at time
k N, z
i
(k) is the ith output (process variable), and G
ij
(q) is
a discrete time transfer operator describing the inuence of the
jth input on the ith output. This paper considers the problem
of controlling the m outputs to their respective reference values
{r
i
}
m
i=1
which are assumed to be constant with respect to time.
Let D R
mm
be a xed diffusion matrix, such that D
ij
=
d
ij
R for all i, j = 1, . . . , m. Also let u
i
(0), u
i
(1), = 0
for all i = 1, . . . , m (implying that the MIMO system (1)
starts from rest). The discrete-time fDI control law is given
as follows.

i
(k) = r
i
z
i
(k), (2)
u
i
(k + 1) = u
i
(k) +
m

j=1
d
ij
f(
j
(k)), (3)
where r
i
is the reference signal (set-point) and
i
(k) is the
control error of the ith output signal. The gain at time k is
denoted , which is chosen to be larger than zero for all k.
We see that the control input i depends on a weighting of the
control errors
j
(k), j = 1, . . . , m, which is determined by the
weighting coefcients d
ij
. The function f : R R can, for
example, describe quantization of the feedback signal.
Now assume that all subsystems G
ij
(q) are FIR of nite order
n 1. Let h
ij
= [h
ij
(0), h
ij
(1), . . . , h
ij
(n)]
T
be a (n + 1)-
dimensional column vector containing the Markov parameters
h
ij
(0), . . . , h
ij
(n) of G
ij
(q). Note that h
ij
(n) = 0 for all i and
j since the order of the subsystems is n 1. The reason for
including this zero in h
ij
will be apparent later.
Let u
k
= [u
1
(k), . . . u
1
(k n), u
2
(k), . . . , u
m
(k n)]
T
be a
m(n + 1)-dimensional column vector where each input from
time k n to k are stacked on each other. Then the ith output
at time k of (1) can be written concisely as
z
i
(k) = [h
T
i1
, h
T
i2
, . . . , h
T
im
]u
k
h
T
i
u
k
. (4)
Let r = [r
1
, r
2
, . . . , r
m
]
T
and z
k
= [z
1
(k), z
2
(k), . . . , z
m
(k)]
T
.
Then the control law (2) and (3) can be written concisely as

k
= r z
k
, (5)
u
k+1
= A
m
u
k
+ {DF(
k
) e
1
}, (6)
where F(
k
) = [f(
1
(k)), f(
2
(k)), . . . , f(
m
(k))]
T
and
e
1
= [1, 0, . . . , 0] R
m
. The symbol is the Kronecker
product operator, and the diffusion matrix D R
mm
equals
D = [d
ij
] R
mm
, (7)
where d
ij
are the weighting coefcients appearing in the fDI
control update (3). Figure 2 shows how a D-matrix (with ones
in the diagonal) is represented in a block-diagram of a 2 2
MIMO system with PI control and linear feedback. Here, is
embedded in the PI-blocks. Note that acts as a scaling factor
on D, and can therefore be embedded in Das well.
In a decentralized control setting, there are typically structural
restrictions on D, depending on how the different controllers
can communicate. If the control is completely decoupled, then
D is a diagonal matrix. If any controller can communicate to
any other controller, Dis dense.
The matrix A
m
R
m(n+1)m(n+1)
is a block-diagonal matrix
with m blocks. Such block, denoted as A R
(n+1)(n+1)
, is
given as
A =
_

_
1 0 . . . 0
1 0 . . . 0
0 1 0
.
.
.
.
.
.
.
.
.
0 1 0
_

_
. (8)
The structure of Ais elaborated on in Nygren et al. (2012).
We also assume that the reference signal r is feasible, i.e. that
there exist an input signal u

= ( 1
n+1
) R
m(n+1)
such that h
T
i
u

m
j=1

j
h
T
ij
1
n+1
= r
i
for all i. Note that
h
T
ij
1
n+1
is equal to the static gain of subsystem ij, i.e. G
ij
(1).
This means that the reference signal is feasible if there exist a
solution to the following equation
_

_
G
11
(1) . . . G
1m
(1)
.
.
.
.
.
.
G
m1
(1) . . . G
mm
(1)
_

_ G = r. (9)
Figure 3 gives a block diagram of the system, the fDI control
strategy and the nonlinear feedback loop. The m m matrix
G(q) = [G
ij
(q)] represents the system transfer operator in
(1). This block diagram also motivates the name of the fDI
controller from the ordering of the f, D and integrator block.
This paper will base its analysis on the fDI control strategy,
and the analysis requires that the static gain G is a symmetric
matrix. In most applications G is not symmetric, but Figure 3

f D

G(q)
F() DF()
u z
r
(q-1)
-
I
Fig. 3. Block diagram of the MIMO closed loop system. The
open loop system is of Hammerstein type, with a (possibly)
nonlinear function f followed by a linear part consisting
of three blocks.
indicates that one can substitute G(q) and Dwith, for example,
G
T
G(q) and DG
T
respectively. The equivalent closed loop
system will then have a symmetric static gain.
If one wishes to substitute the DI-control blocks with a more
complex controller C(q), one may factorize it as C(q) =
D

q1

C(q) and then merge

C(q) with G(q). If C(q) is of PI-
form, for example, the merged system

C(q)G(q) will preserve
its FIR form. In the 2 2 MIMO system in Figure 2,

C(q) is a
diagonal 2 2 matrix.
The main questions arising from this discussion are howto tune
the parameters of

C(q), howto symmetrize the static gain in the
best way (multiplying with G
T
is just one possible way), and
how to choose the gain and the diffusion matrix D subject
to some sparsity pattern which represents the possible con-
troller communication pathways. In the forthcoming analysis
we restrict ourselves to the problem of choosing under the
assumption that Dand G(q) are xed.
1.2 Special Cases
The setting given in the previous subsection is general enough
to include the following cases of particular interest:
Q: To reduce communication cost in the sensor/actuator net-
work, sensor measurements can be crudely quantized.
This can be realized by letting f() be a piecewise con-
stant, monotonic stair function:
f(x) = x
i
, x
i
x < x
i+1
, i Z . (10)
In Elia and Mitter (2001), the coarsest quantizer for stabi-
lization is characterized, and it turns out to be of logarith-
mic form. In this paper, we focus on convergence analysis
rather than stabilization.
C: The function f() can be any monotone function. For
example, if we want to avoid communication at small
control errors, it makes sense to choose a dead-zone
function with parameter > 0, dened as
f(x) =
_
x |x| >
0 |x| .
(11)
W: It is possible to replace f(
i
(k)) with r
i

f(z
i
(k)), where

f : R R. This is convenient if, for example, the

quantized output is communicated between the sensors
and actuators instead of the quantized control error. The
open loop system is then a Wiener system instead of a
Hammerstein system.
2. ANALYSIS
Given the setting as described in the previous section, we now
want to choose gains such that the m control errors become as
small as possible for a given diffusion matrix D. In this section,
we arrive at sufcient conditions on the gain and function f
for asymptotic convergence of the control errors. Now, consider
the vector (semi)norm x
2
Qm
=
1
2
x
T
Q
m
x
T
, where Q
m

R
m(n+1)m(n+1)
. Eq. (6) becomes:
u

u
k+1

2
Qm
= u

u
k

2
A
T
m
QmAm
2(u

u
k
)
T
A
T
m
Q
m
{DF(
k
) e
1
}
+
2
DF(
k
) e
1

2
Qm
.
(12)
Let e
i
= [0, . . . , 0, 1, 0, . . . , 0]
T
R
m(n+1)
, where the 1 is
in the 1 + (n + 1)(i 1) entry. As such, the analysis will
hinge on the availability of a proper matrix Q
m
. We will use
the following properties of the matrix Q
m
, and show later the
existence of such a matrix.
B1: Q
m
is symmetric and positive semidenite.
B2: for all i we have that A
T
m
Q
m
e
i
= h
i
.
B3: Q
m
A
T
Q
m
Ais positive semidenite.
A necessary condition for Q
m
to satisfy B2 is that all h
ij
(n) =
0, which we assumed in the Preliminaries section. Appendix A
provides a proof that a matrix Q
m
which satises B1-B3 exist
in the SISO case, if and only if the static gain of the system is
positive. Together with the feasibility condition (9), we require
the static gain of the SISO system to be strictly positive (unless,
of course, the reference signal is zero). This positivity constraint
is natural since the gain of the fDI-controller is positive.
The case with negative static gain can be dealt with by using
negative gains. Appendix B includes an existence proof of Q
m
for the MIMO system (1). For existence of Q
m
in the MIMO
case it is necessary that G 0. Let d
i
= [d
i1
, . . . , d
im
]
T
equal
the ith row of D. Using property B2 and that r is feasible
we make the following derivation of the cross-term in equation
(12):
(u

u
k
)
T
A
T
m
Q
m
{DF(
k
) e
1
} =
(u

u
k
)
T
A
T
m
Q
m
{ e
1
d
1
F(
k
) + + e
m
d
m
F(
k
)}
=
1
(k) d
1
F(
k
) + +
m
(k) d
m
F(
k
)
=
T
k
DF(
k
).
(13)
Now let x
ij
, i, j = 1, . . . , m be the element of Q
m
in the
1 + (i 1)(n + 1)th row and 1 + (j 1)(n + 1)th column
(see Appendix B for details). Also let X = [x
ij
] R
mm
,
and let x
i
be the ith column of X. Because of B1, X 0 (see
Appendix B). Then the last term in (12) becomes
DF(
k
) e
1

2
Qm
=
_
_
m

j=1
d
j
F(
k
)x
j1
, . . .
_
_
{DF(
k
) e
1
}
T
=
_
_
m

j=1
d
j
F(
k
)x
j1
, . . . ,
m

j=1
d
j
F(
k
)x
jm
_
_
DF(
k
)
= F(
k
)
T
D
T
[x
1
, x
2
, . . . , x
m
]DF(
k
)
= F(
k
)
T
D
T
XDF(
k
). (14)
In the second equality relation in (14) we use the fact that the
elements of {DF(
k
) e
1
} between the 1 + (i 1)(n + 1)
entries (i = 1, . . . , m) are zero. Using (12) together with (13),
(14) and B3, we now arrive at the following expression
u

u
k+1

2
Qm
u

u
k

2
Qm
2
T
k
DF(
k
) +
2
F(
k
)
T
D
T
XDF(
k
).
(15)
This leads to the following lemma, which gives an interesting
bound of the control error norms.
Lemma 1. Given the system (1) where each G
ij
(q) is FIR of
order n 1 with static gain G 0, with the fDI-controller
(5) and (6), a feasible reference signal r and a matrix Q
m
which
satises conditions B1-B3, with an associated X, and assuming
that u
0
= 0, the control errors fromtime 0 to k will be bounded
as
k

j=0

2
D
+F(
j
)
2
D
F(
j
)
2
D
T
XD
_

T
G +
k

j=0

j
F(
j
)
2
D
.
(16)
Proof. First we note that G 0 is equivalent to the existence
of a Q
m
satisfying B1-B3, according to Appendix B.
Unfolding the recursion (15) gives
u

u
k+1

2
Qm
u

u
0

2
Qm
2
k

j=0
_

T
j
DF(
j
)
_
+
k

j=0
_

2
F(
j
)
2
D
T
XD
_
.
(17)
Let q
p
be the pth column of Q
m
. To develop the rst term
of (17), we use the fact that q
T
p
u

m
j=1

j
1
T
h
ji
for all
p = 1 + (i 1)(n + 1) (due to B2) and q
T
p
u

= 0 for all
other p (see Appendix A). Since u
0
= 0, the rst term of (17)
becomes
u

2
Qm
= [
m

j=1

j
e
T
1
h
j1
, 0, . . . ]u

=
m

i=1

i
m

j=1

j
e
T
1
h
ji
=
T
G.
(18)
The cross terms in (17) are replaced by
2
T
j
DF(
j
) =

j
F(
j
)
2
D

2
D
F(
j
)
2
D
.
(19)
Equation (16) is then obtained by using (18) and (19) in (17)
and reorganizing the terms.
Lemma 1 states that the control errors will be bounded for all
k = 0, 1, . . . if is small enough and if
j
F(
j
)
2
D
is
bounded. The gain will be small enough to achieve bounded
control errors if F(
j
)
T
(DD
T
XD)F(
j
) 0 for all j, or
if (D
T
+D2D
T
XD) is positive semidenite. If F(
k
) =

k
, we also get asymptotic convergence, i.e. lim
k

k
= 0 if
this condition holds.
Now we are ready to state a main result of this paper, which
gives conditions for asymptotic convergence of the control
errors with respect to and a proper constraint on the function
f (which might represent quantization coarseness as in Elia and
Mitter (2001)).
Theorem 2. Given the system (1) where each G
ij
(q) is FIR of
order n 1 with static gain G 0, a matrix D R
mm
such that D
T
+D 0 and a function f : R R that satises
the following condition
c|| |f()| ||, R (20)
where 0 < c 1 is chosen such that

min

max
>
(1 c)
2
1 + c
2
, (21)
where
min
and
max
is the minimum and maximum eigen-
value of
1
2
(D+D
T
). Then there exist a

> 0 such that if the

fDI-controller (5) and (6) is applied, with a feasible reference
signal r, and assuming that u
0
= 0, we get asymptotic conver-
gence of the control errors, i.e. lim
k

k
= 0, for all gains
satisfying 0 <

.
Proof. Without loss of generality, we can assume that
max
=
1, since scales with D. By assumption we have that
min

0. Also denote the maximum eigenvalue of D
T
XDas
DXD
.
The main ideas behind the proof is to bound the different
quantities of (16) in Lemma 1. Due to (20), we have

k
F(
k
)
2
D

k
F(
k
)
2
(1 c)
2

2
, (22)
F(
k
)
2
D

min
F(
k
)
2

min
c
2

2
, (23)
and
F(
k
)
2
D
T
XD

DXD
F(
k
)
2

DXD

2
. (24)
Then, (16) can be rewritten as

min
(1 + c
2
) (1 c)
2

DXD
)
_
k

j=0

2

T
G.
(25)
Due to (21), the quantity
min
(1 + c
2
) (1 c)
2
is strictly
positive. Furthermore, since X 0,
DXD
is nonnegative.
Therefore, there exist some number

such that 0 <

implies that the quantity on the left hand side of (25) is strictly
positive. This means that

j=0

2
< , implying that

2
0 as k .
3. DISCUSSION
This paper proposed a design of cooperative PI control laws
for controlling a MIMO system. The key was to have cooper-
ation be encoded in a matrix D, and to work out a reduction
to a traditional gradient descent scheme. At this point, this
technique is only worked out theoretically for MIMO systems
where the components can be represented as FIRsystems, while
it remains open how to extend to the IIR cases as well.
REFERENCES
K.J.

Astr om and T. H agglund. Advanced PID control. ISA-The Instrumenta-
tion, Systems, and Automation Society; Research Triangle Park, NC 27709,
2006.
A. Bemporad, M. Heemels, and M. Johansson. Networked Control Systems,
volume 406. Springer-Verlag New York Inc, 2010.
R.W. Brockett and D. Liberzon. Quantized feedback stabilization of linear
systems. IEEE Transactions on Automatic Control, 45(7):12791289, 2000.
MBG Cloosterman, L. Hetel, N. Van De Wouw, W. Heemels, J. Daafouz,
and H. Nijmeijer. Controller synthesis for networked control systems.
Automatica, 46(10):15841594, 2010.
N. Elia and S.K. Mitter. Stabilization of linear systems with limited informa-
tion. IEEE Transactions on Automatic Control, 46(9):13841400, 2001.
C. Fischione, P. Park, P. Di Marco, and K.H. Johansson. Design principles
of wireless sensor networks protocols for control applications. Wireless
Networking Based Control, pages 203238, 2010.
J.P. Hespanha, P. Naghshtabrizi, and Y. Xu. A survey of recent results in
networked control systems. Proceedings of the IEEE, 95(1):138162, 2007.
R.A. Horn and C.R. Johnson. Matrix analysis. Cambridge Univ Pr, 1990.
J.M. Maciejowski. Multivariable feedback design. Electronic Systems Engi-
neering Series, Wokingham, England, Addison-Wesley, 1989.
J. Nygren, K. Pelckmans, and B. Carlsson. A discrete-time PI control law with
nonlinear feedback. Submitted, 2012.
Z.J. Palmor, Y. Halevi, and N. Krasney. Automatic tuning of decentralized PID
controllers for TITO processes. Automatica, 31(7):10011010, 1995.
Appendix A. EXISTENCE PROOF IN THE SISO CASE
Here we give an existence proof of a Q
m
satisfying B1-B3
in the SISO case. In Appendix B, existence is proven for the
MIMO case.
Theorem 3. For a vector h = [h
0
, h
1
, . . . , h
n1
], there exist
a Q R
n+1n+1
such that Q is positive semidenite (B1),
A
T
Qe
1
= [h
T
, 0]
T
(B2) and QA
T
QAis positive semidef-
inite (B3), if and only if

n1
j=0
h
j
0. A is dened in (8).
The proof of this theorem is structured in three steps. First, a
row sum property of Q is derived. Then, an equivalent matrix
equation is derived using the Jordan formof A. Using the afore-
mentioned row sum property, the condition that

n1
j=0
h
j
0
becomes apparent. In the last step of the proof, a specic Q
m
which satises B1-B3 is derived using the equivalent matrix
relation from the second step.
The Q-matrix has the following structure (due to B2):
Q =
_

_
x h
0
x h
1
. . . h
n1
h
0
x
h
1

Q
.
.
.
h
n1
_

_
. (A.1)
where

Q = {q
ij
}
n
i,j=0
is a symmetric n n matrix, and x is a
scalar. As a rst step, we prove the following proposition.
Proposition 4. Denote q
p
as the pth row of Q. If Q
A
T
QA 0 (B3), the sum of all rows of Q (except for the
rst row) is zero, i.e. q
p
1 = 0 for all p = 2, . . . , n + 1. If
A
T
Qe
1
= [h
T
, 0]
T
(B2), then q
1
1 =

n1
j=1
h
j
.
Proof. The fact that q
1
1 =

n1
j=1
h
j
if B2 holds follows
from (A.1). Now observe that A1 = 1. This means that
1
T
(Q A
T
QA)1 = 0 regardless of the properties of Q. By
the Rayleigh-Ritz theorem (see Horn and Johnson (1990)), we
conclude that if B3 holds, the minimum eigenvalue of (Q
A
T
QA) is zero (with 1 as the corresponding eigenvector).
This gives the following equations q
1
1 q
1
1 q
2
1 = 0,
q
p
1 q
p+1
1 = 0 for all p = 2, . . . , n and q
p+1
1 = 0. These
equations imply that q
p
1 = 0 for all p = 2, . . . , n + 1.
Now we use the following Jordan form transformation on A:
B = T
1
AT, (A.2)
where
B =
_

_
1 0 . . . 0
0 0 . . . 0
0 1 0
.
.
.
.
.
.
.
.
.
0 1 0
_

_
. (A.3)
It turns out that
T =
_

_
1 0 . . . 0
1 1 0 . . . 0
1 1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. 0
1 1 . . . 1
_

_
. (A.4)
Now consider the following derivation
QA
T
QA = QT
T
B
T
T
T
QTBT
1

T
T
QTB
T
_
T
T
QT
_
B
MB
T
MB Z.
(A.5)
Since any matrix Q and its transformation T
T
QT has the
same inertia (same amount of negative, positive and zero eigen-
values) provided that T is invertible (which it is), we have that
Q 0 iff M 0. Furthermore, Z 0 iff QA
T
QA 0.
From calculations we get
M=
_

_
n1

j=0
h
j
0 . . . 0
0
.
.
.

M
0
_

_
, (A.6)
where the (k, l)-entry

M
k,l
=

n1
j=k1

i=l1
q
ij
. The rst
row and column of M have zero entries due to Proposition 4.
To make M 0 (and hence Q 0), we see from (A.6) that B1
holds only if

n1
j=0
h
j
0.
Furthermore, we have that
Z =
_

_
0 0 . . . 0
0
.
.
.

Z
0
_

_
, (A.7)
where

Z
k,l
= 2

n1
j=k1
q
j,l1
q
k1,l1
.
The last step of the proof is to give an example of a Qsuch that
B1-B3 is satised. If we construct a Q for which Z 0, we
get that Q A
T
QA 0 (B3) and, since the lower principal
n n submatrix of B have n zero eigenvalues, M 0 (which
implies B1). Therefore, we construct the following Q:
Q =
_

_
x h
0
x h
1
h
2
. . . h
n1
h
0
x x h
0
+ a
0
b
1
h
1
h
2
. . . h
n1
h
1
b
1
h
1
a
1
b
2
0 . . . 0
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
h
n2
h
n2
0 . . . b
n2
a
n2
b
n1
h
n1
h
n1
0 . . . 0 b
n1
a
n1
_

_
,
(A.8)
where
_

_
a
0
=
n1

i=1
h
i
+
n1

k=1
|
n1

i=k
h
i
|
a
j
= |
n1

i=j
h
i
| + 2
n1

k=j+1
|
n1

i=k
h
i
|, j = 1, . . . , n 2
a
n1
= |h
n1
|,
(A.9)
and
b
j
=
n1

k=j

n1

i=k
h
i

, j = 1, . . . , n 1. (A.10)
We then see that

Z becomes

Z =
_

_
g(x)
n1

j=1
h
j
. . . h
n1

n1

j=1
h
j

n1

j=1
h
j

0 . . .
.
.
. 0
.
.
.
.
.
.
.
.
.
h
n1
0 . . . |h
n1
|
_

_
, (A.11)
where
g(x) = x
n1

j=0
h
j

n1

k=1

n1

i=k
h
i

. (A.12)
By Gerzgorins disc theorem (see Horn and Johnson (1990)),
we see that

Z 0 if x =

n1
j=0
h
j
+2

n1
k=1

n1
i=k
h
i

. Since
Z 0, then B1 holds (since M 0) and also, QA
T
QA 0
(B3) holds. B2 holds by construction (A.1).
Appendix B. THE MIMO CASE
In this appendix, an existence proof of a Q
m
which satises B1-
B3 in the MIMO case is given. The formal statement, which is
to be proven, is given as the following theorem.
Theorem 5. For given m(n+1)-vectors h
i
= [h
T
i1
, . . . , h
im
]
T
,
h
ij
= [h
ij
(0), . . . , h
ij
(n1), 0]
T
and a m-dimensional matrix
G {

n1
k=0
h
ij
(k)}
i,j
, there exist a Q
m
R
m(n+1)m(n+1)
such that Q
m
is positive semidenite (B1), A
T
m
Q
m
e
i
= h
i
for
all i (B2) and Q
m
A
T
m
Q
m
A
m
is positive semidenite (B3),
if and only if G 0 (which also means that G = G
T
).
The structure of the proof is the same as for the SISO case
in Appendix A. First, we note that the m(n + 1)-dimensional
matrix Q
m
can be divided into m
2
sub-blocks, which we
denote Q
ij
R
n+1n+1
, i, j = 1, . . . , m. The sub-block Q
ij
,
which is located at block-rowi and block-column j of Q
m
, has
the following structure (due to B2):
Q
ij
=
_

_
x
ij
h
ij
(0) x
ij
. . . h
ij
(n 1)
h
ji
(0) x
ij
h
ji
(1)

Q
ij
.
.
.
h
ji
(n 1)
_

_
. (B.1)
Also note that Q
ij
= Q
T
ji
. B1 implies that {ve
1
}
T
Q
m
{v
e
1
} = v
T
Xv 0 for any vector v, hence X is positive
semidenite.
As a rst step of the proof, we prove the following proposition
(which is a generalization of Proposition 4).
Proposition 6. Denote q
p
ij
as the pth row of Q
ij
. If Q
m

A
T
m
Q
m
A
m
0 (B3), the sum of the rows of Q
ij
, i, j =
1, . . . , m (except for the rst row) is zero, i.e. q
p
ij
1 = 0 for
all p = 2, . . . , n + 1. If A
T
m
Q
m
e
i
= h
i
(B2), then q
1
ij
1 =

n1
k=1
h
ij
(k).
Proof. The fact that q
1
ij
1 =

n1
k=1
h
ij
(k) if B2 holds follows
from (B.1). Now observe that A
m
{e
i
1
n+1
} = {e
i
1
n+1
},
where e
i
is a m-dimensional zero vector except for a 1 at the
ith element. This means that {e
i
1
n+1
}
T
(QA
T
QA){e
i

1
n+1
} = 0 for all i = 1, . . . , m. By the Rayleigh-Ritz theorem,
we conclude that if B3 holds, (Q
m
A
T
m
Q
m
A
m
){e
i

1
n+1
} = 0 for all i. This implies that q
1
ij
1q
1
ij
1q
2
ij
1 = 0,
q
p
ij
1 q
p+1
ij
1 = 0 for p = 2, . . . , n and q
n+1
ij
1 = 0. These
equations imply that q
p
ij
1 = 0 for all p = 2, . . . , n + 1.
Proposition 6 also tells us that if both B2 and B3 holds, then

n1
k=0
h
ij
(k) =

n1
k=0
h
ji
(k) and hence, Gis symmetric. This
is because Q
ij
= Q
ji
for all i, j, implying that the sum of all
elements of Q
ij
(which is

n1
k=0
h
ij
(k)) is equal to the sum of
all elements of Q
ji
.
In the second step of the proof, we make the the following
transformation
Qm A
T
m
QmAm = Qm T
T
m
B
T
m
T
T
m
QmTmBmT
1
m

T
T
m
QmTm B
T
m
_
T
T
m
QmTm
_
Bm
Mm B
T
m
MmBm Zm.
(B.2)
B
m
is the Jordan form of A
m
, and the transformation matrix
T
m
is block-diagonal with m blocks of the form (A.4). All
sub-blocks M
ij
of M
m
have the same structure as (A.6). This
means that {ve
1
}
T
M
m
{ve
1
} = G, and hence, G 0 is
a necessary condition for B1 to hold. Also, all sub-blocks Z
ij
of Z
m
have the same structure as (A.7).
Now choose Q
m
such that

Q
ij
=
_

_
x
ij
h
ij
(0) +
n1

k=1
h
ji
(k) . . . h
ij
(n 1)
h
ji
(1) 0 . . .
.
.
.
h
ji
(n 1) 0 . . .
_

_
(B.3)
if i = j. Also choose x
ij
= 0 for i = j. Then choose Q
ii
such
that it has the same structure as (A.8), but with the following
parameters:
_

_
a
0
=
n1

k=1
h
ii
(k) +
m

j=1
_
n1

l=1
|
n1

k=l
h
ji
(k)|
_
ap =
m

j=1
_
|
n1

k=p
h
ji
(k)| + 2
n1

l=p+1
|
n1

k=l
h
ji
(k)|
_
,
a
n1
=
m

j=1
|h
ji
(n 1)|,
(B.4)
and b
p
=

m
j=1

n1
l=p

n1
k=l
h
ji
(k)

for all j = 1, . . . , n.
Denote the ij-entry of G as G
ij
. Then, the sub-blocks Z
ij
of
Z
m
have the same structure as (A.7) with

Z
ij
=
_

_
G
ij
. . . h
ij
(n 1)

n1

k=1
h
ij
(k) 0 . . .
.
.
.
hn 0 . . .
_

_
, (B.5)
for i = j, and

Z
ii
=
_

_
g
ii
(x
ii
)
n1

k=1
h
ii
(k) . . . h
ii
(n 1)

n1

k=1
h
ii
(k)
m

j=1
|
n1

k=1
h
ji
(k)| . . .
.
.
.
.
.
.
h
ii
(n 1) 0 . . .
m

j=1
|h
ji
(n 1)|
_

_
,
(B.6)
where g
ii
(x
ii
) = x
ii
G
ii

m
j=1
_

n1
l=1

n1
k=l
h
ji
(k)

_
.
By Gerzgorins disc theorem, we see that Z
m
0 if
x
ii
=
m

j=1
_
G
ji
+ 2
n1

l=1

n1

k=l
h
ji
(k)

_
. (B.7)