Académique Documents
Professionnel Documents
Culture Documents
4
Westin Seattle Hotel, Seattle, Washington, USA
June 11-13, 2008
Abstract— In this paper, minimum gain of an operator is In this paper, the minimum gain of a system has been
introduced. Moreover, some of its properties are presented. It studied. Although it has been showed that the minimum
is proved that the minimum gain of a strictly proper, stable, gain is not a norm on space of operators, a new stability
LTI system is zero. It is also shown that the minimum gain of
an operator fails to satisfy the triangular inequality. Finally, condition has been derived for feedback systems based on
the so-called large gain theorem is stated and a new stability the minimum gains of the open-loop systems.
condition for feedback interconnection of nonlinear systems The paper is organized as follows. In Section II we in-
is derived. troduce the notation and present some preliminary results.
Index Terms— Nonlinear systems, induced operator In Section III, the minimum gain of an operator is defined
norms, continuous-time systems, discrete-time systems, small- and some of its properties are derived. In Section IV, the
gain theorem large gain theorem is stated. An example is also provided
to illustrate the usage of the theorem.
I. I NTRODUCTION
II. N OTATION AND P RELIMINARIES
ne of the well-accepted and widely-used methods
O to study stability of systems is the input-output
approach. It was initiated by Popov, Zames, and Sandberg,
A. Notation
Let R and C denote the fields of real and complex
in the 1960s [8] [11] [9]. So far, it has been a fruitful area numbers, respectively. RHP stands for the portion of the
which has resulted in many of the recent developments complex plane with positive real part. Rn denotes the space
in control theory, such as robust control and small-gain of n×1 real vectors. For M ∈ Rn×m , M T is the transpose
based nonlinear stabilization techniques. The input-output of M ; σ(M ) is the minimum singular value of M; σ(M )
stability theory considers systems as mappings from an is a singular value of M; σ(M ) is the maximum singular
input space of functions into an output space. In this theory, value of M. The Euclidean norm in Rn is denoted by k · k.
the well-behaved input and output signals are considered In×n denotes the n×n identity matrix. Let Bp (c, r) denote
as members of input and output spaces. Therefore, if the the open ball with center c and radius r with norm p, i.e.
“well-behaved” inputs produce well-behaved outputs, the Bp (c, r) := {x| kx − ckp < r}. Lrp denotes Lebesgue p-
system is called stable. space of r-vector valued functions on [0, ∞], with norm
R∞ 1/p
The main contribution of the input-output stability the- defined as kf kp := 0 kf (t)kp for 1 ≤ p < ∞ and
ory in control theory is through the well-known small-gain kf k∞ := ess supt∈R kf (t)k. Usually r is a finite integer;
theorem. In this context, the most notable contributions we drop r and write Lp instead of Lrp , if no confusion will
have also been made by Zames and Sandberg, e.g. [11] arise. Similarly, let ℓp denote the vector space of discrete-
[9]. The small gain theorem says that the feedback loop time signals with norm k · kp . Let Xp denote either Lp or
will be stable if the loop gain is less than one. This simple ℓp and X denote Xp for any 0 < p ≤ ∞. To distinguish
rule has been a basis for numerous stabilization techniques among various norm notation, we indicate the space as a
such as nonlinear H∞ control [3]. subscript for the norm, such as k·kRn or k·kXp . Whenever
Stability of systems, in its various forms, still continues the space is not mentioned, norms with t argument denote
to inspire researchers. Motivated by the classical small Euclidean norm at t and without t denote the Xp norm
gain theorem, ”nonlinear gain” small gain theorems are where Xp is a general space or can clearly be understood
discussed in literature as [5] [10] [4]. The notion of non- from the content. Let Tτ denote the truncation operator:
uniform in time robust global asymptotic output stability for f (t) , 0 ≤ t < ∞, Tr f (t) = f (t) on [0, τ ], and
introduced in [6] for a wide class of systems. A small-gain zero otherwise. We also denote the truncation of f (t) by
theorem for a wide class of feedback systems is proposed fT (t) := Tr f (t).
in [7]. In [2], it is shown that for an open loop unstable Let U := X and Y := X denote input and output signal
system which is closed loop stable the gain must exceed spaces, respectively. A nonlinear time-varying system can
one. be thought of as a possibly unbounded operator H : Dh →
H2 x
Y where Dh ⊆ U. The action of H on any u ∈ Dh is
denoted by Hu. A system H is called to be stable if Dh =
2
U. For an operator H : U → Y , let γp (H) stand for the
1
induced norm (gain) of the operator defined as -3 -2 -1
1 2 3 x
kHukT -1
γp (H) := sup (1) -2
0 6= u ∈ U kukT
3441
3
3442
4
u1 e1 y1
H1 Moreover, substituting (36) and (35) in (30) and (31),
+
-
respectively,
y2 e2 + u2
H2
+ ν(H1 )
ky1T k ≤ (ku1T k + ν(H2 ) ku2T k)
Fig. 4: The feedback system ν(H1 )ν(H2 ) − 1
(37)
and
ν(H2 )
e1 = u 1 − y2 ky2T k ≤ (ν(H1 ) ku1T k + ku2T k)
ν(H1 )ν(H2 ) − 1
e2 = u 2 + y1 (38)
(25)
y 1 = H1 e 1
Hence, the norms of ke1T k, k12T k, ky1T k and ky2T k are
y 2 = H2 e 2
bounded. If, in addition, u1 , u2 ∈ X , then (35-38) must
Theorem 4.1: Consider the feedback interconnection also be satisfied if T approaches ∞. Therefore,
described by (25) and shown in Fig. 4. Then, if 1 <
ν(H1 )ν(H2 ) < ∞, the feedback system is input-output- 1
ke1 k ≤ (ku1 k + ν(H2 ) ku2 k) (39)
stable. ν(H1 )ν(H2 ) − 1
Proof: To show stability of the feedback intercon-
1
nection, we must show that u1 , u2 ∈ X imply that e1 , e2 , ke2 k ≤ (ν(H1 ) ku1 k + ku2 k) (40)
ν(H1 )ν(H2 ) − 1
y1 and y2 are also in X . According to the definition of ν,
we have ν(H1 )
ky1 k ≤ (ku1 k + ν(H2 ) ku2 k) (41)
ν(H1 ) ke1T k ≤ ky1T k (26) ν(H1 )ν(H2 ) − 1
ν(H2 ) ke2T k ≤ ky2T k (27)
ν(H2 )
ky2 k ≤ (ν(H1 ) ku1 k + ku2 k) (42)
On the other hand, ν(H1 )ν(H2 ) − 1
y1T = e2T − u2T (28) Consequently, e1 , e2 , y1 and y2 are also in X .
y2T = u1T − e1T (29) Example 4.1: Let H1 be the convolution operator de-
fined by (8) where g(t) is the impulse response of
Thus,
s+1
ky1T k ≤ ke2T k + ku2T k (30) G(s) = k
s−1
ky2T k ≤ ke1T k + ku1T k (31)
where k ∈ R. Let H2 be a memoryless nonlinearity
Substituting (26) and (27) in (30) and (31), respectively, depicted in Fig. 1. As shown in Example 3.2, ν(H1 /k) = 1
which implies that ν(H1 ) = |k|. On the other hand, we
ν(H1 ) ke1T k ≤ ke2T k + ku2T k (32) have ν(H2 ) = 0.5. Consequently ν(H1 )ν(H2 ) = 0.5|k|.
ν(H2 ) ke2T k ≤ ke1T k + ku1T k (33) The large gain theorem, namely Theorem 4.1, guarantees
that the feedback system is stable if |k| > 2.
Using (27) and (31), Equation (32) implies that
ν(H2 )ν(H1 ) ke1T k ≤ ν(H2 ) ke2T k + ν(H2 ) ku2T k
V. C ONCLUSION
≤ ky2T k + ν(H2 ) ku2T k
≤ ke1T k + ku1T k + ν(H2 ) ku2T k The minimum gain of an operator as well as some of
(34) its properties are introduced. These properties are useful
in the computation of the minimum gain of a system. For
Since ν(H1 )ν(H2 ) > 1, instance, it is shown that the minimum gains of strictly
1 proper, stable, LTI systems are zero. When it comes to the
ke1T k ≤ (ku1T k + ν(H2 ) ku2T k) metric properties, the minimum gain of an operator fails to
ν(H1 )ν(H2 ) − 1
(35) satisfy the triangular inequality which implies that it is not
a metric or a norm in the space of operators. Finally, the
Similarly, so-called large gain theorem is stated and proved. This
1 theorem implies a new stability condition for feedback
ke2T k ≤ (ν(H1 ) ku1T k + ku2T k)
ν(H1 )ν(H2 ) − 1 interconnection of nonlinear systems. An example is pro-
(36) vided to illustrate the derived stability condition.
3443
5
A PPENDIX I For p = ∞,
P ROOF OF LEMMA 3.1
kHxkL∞ = sup kΦ(t, x(t))k ≥ sup µp kx(t)k
t t
The proofs for the continuous-time and discrete-time p (48)
cases are the same and only the first one is given here. We = µp sup kx(t)k = µp kxkL∞
t
use the following property of the smallest singular value
Equations (47) and (48) imply that µp is a lower bound
of matrices (e.g. [12] pp. 21):
for ν(H). This completes the proof.
kM xk
σ(M ) = min kM xk = min (43)
kxk=1 x6=0 kxk A PPENDIX III
P ROOF OF LEMMA 3.3
Let M = U ΣV T be the SVD of M , where V = T
Let x̂(t) = [x̂1 (t) x̂2 (t) · · · x̂n (t)] ,
[v1 , v2 , · · · , vn ] ∈ Rn×n , U, Σ ∈ Rn×n . It is well-known (
that vn is the minimizer of (43), e.g. [12]. Let x ∈ L2 , we sin(ωt) k = i,
have x̂k (t) =
0 otherwise.
Z ∞
2 2
kM xk = kM x(t)k2 dt where i corresponds to the strictly proper row in G(s) and
Z0 ∞ ω ≥ π. Let
2
≥ σ(M )2 kx(t)k2 dt (44) jωk π
0 x(t) := x̂(t) − x̂ t − (49)
Z ∞ π ω
2 2
= σ(M )2 kx(t)k2 dt = σ(M )2 kxk where ⌊r⌋ denotes the floor function of a real number r,
0
which is the largest integer less than or equal to r, namely
which shows that σ(M ) is a lower bound for ν(H). ∀r ∈ R ; ⌊r⌋ := sup{n ∈ Z|n ≤ r}. It is trivial that
To show that it is the greatest lower bound, let x(t) =
vn −t 0
kvn k e . We have
0
.
Z ∞
vn −t
2
Z ∞
−t
2
..
2
e
dt = 1/ 0 ≤ t ≤ ωπ ,
kxk = e
dt =
2 sin(ωt) ith row
kvn k
0 0 x(t) =
..
(45)
.
0
and
2
∞
vn −t
Z
t > ωπ .
0
2
kM xk =
M kvn k e
dt
0
Z ∞ and j ω k π
e−2t
= kM vn k2 dt kx(t)k = x̂i (t) − x̂i t −
kvn k2 π ω
0
(46)
Z ∞
2 e
−2t Thus,
= kσ(M )vn k 2 dt
0 kvn k kxkL∞ = sup | sin(ωt)| = 1 (50)
Z ∞ t
2 2
= kσ(M )k e−2t dt = 1/2 kσ(M )k Z ⌊ω π
π ⌋ω
0 kxk2L2 = | sin(ωt)|2 dt
0
Equations (45) and (46) imply that ν(H) is equal to σ(M ) ⌊ ω ⌋ π
for some input. This completes the proof. sin (2ωt) π ω
= 1/2
t− (51)
2ω
0
A PPENDIX II
jωk π sin 2π⌊ ωπ ⌋
= −
P ROOF OF LEMMA 3.2 π 2ω 4ω
Z ⌊ ωπ ⌋ ωπ
Let x ∈ Lp , for p 6= ∞, kxkL1 = | sin(ωt)| dt (52)
Z ∞ Z ∞ 0
p
kHxkpLp = kΦ(t, x(t))k dt ≥ µpp kx(t)kp dt of |x̂(t)| depicted in
To calculate (52), consider the graph
0Z
∞
0 Fig. 5. The number of peaks is ωπ . Moreover,
p p
= µpp kx(t)k dt = µpp kxkLp Z ωπ
0
2
S= sin (ωt) dt = (53)
(47) 0 ω
3444
6
1
of G(s) at s0 . Since all poles of Y (s) are in LHP, y(t) is
a decaying signal. On the other hand, u(t) is an unstable
S signal, rising by time. If we truncate both u(t) and y(t)
0 1
at T , which is chosen sufficiently large, the corresponding
x
gain of the system will be small. By increasing T , the gain
can be decreased as much as desired. Therefore, ν(H) = 0.
Fig. 5: |x̂(t)|
A PPENDIX V
P ROOF OF LEMMA 3.5
Consequently, Let y(t) := Hu(t), which implies that u(t) = H −1 y(t).
jω k 2
jω k Therefore
kxkL1 = S= (54) kyT k kyT k 1
π π ω ν(H) = inf = inf = inf ku k
u∈U kuT k u∈D h kuT k u∈D h T
To calculate the norm of the output kyk, we can first find kyT k
the response of the system to input x̂(t), namely ŷ(t),
and 1 1
= =
then obtain the output using y(t) = ŷ(t) − ŷ t − ωπ ωπ kuT k kH −1 y T k
implied by the linearity property of the system and (49). sup sup (61)
u∈Dh kyT k u∈Dh kyT k
If we let ω → ∞, the response of the system to x̂(t) 1 1
approaches to zero. The reason is that the amplitude of = =
kH −1 yT k γ(H −1 )
all elements of the i-th row of G(s) approaches to zero sup
at high frequencies. Therefore, limw→∞ kŷ(t)k = 0 and y∈Rh kyT k
consequently R EFERENCES
lim kyk = 0 (55)
w→∞ [1] J. A. Ball and J. W. Helton, “Interconnection of nonlinear
On the other hand, (51) and (54) imply causal systems,” IEEE Trans. on Automatic Control, vol. 34,
pp. 1132 – 1140, 1989
2 [2] T.T. Georgiou , M. Khammash, and A. Megretski, “On a
lim kxkL2 = 1/2, lim kxkL1 = (56) large-gain theorem,” Systems and Control Letters, vol. 32,
w→∞ w→∞ π
pp. 231 – 234, 1997
Equations (50), (55) and (56) imply [3] J. W. Hilton and M. R. James, Extending H∞ Control to
Nonlinear Systems, SIAM, 1999.
ν1 (H) = 0, ν2 (H) = 0, ν∞ (H) = 0 (57) [4] B. Ingalls and E. D. Sontag, “A small-gain theorem with
applications to input/output systems, incremental stability,
A PPENDIX IV detectability and interconnection,” Journal of the Franklin
Institute, vol. 339, pp. 211 – 229, 2002.
P ROOF OF LEMMA 3.4
[5] Z. P. Jiang, A. Teel and L. Praly, “Small-gain theorem for ISS
The proofs for the continuous-time and discrete-time systems and applications,” Mathematics of Control, Signals,
cases are the same and only the first one is given here. and Systems, vol. 7, pp. 95 – 120, 1994.
[6] I. Karafyllis, “The Non-Uniform in Time Small-Gain The-
Let s0 be the RHP zero of G(s), namely there exists w
orem for a Wide Class of Control Systems with Outputs,”
such that G(s0 )w = 0. If σ0 + iω0 = s0 ∈ C, trivially s∗0 European Journal of Control, vol. 10, no. 4, pp. 307–323,
is also a RHP zero of G(s). Let 2004.
( [7] I. Karafyllis and Z. P. Jiang, “A small-gain theorem for a
w es0 t if s0 ∈ R, wide class of feedback systems with control applications,”
u(t) = σ0 t
(58) SIAM J. Control and Optimization, in press, 2007.
w e sin (ω0 t) if s0 ∈ C.
[8] V. M. Popov, “Absolute stability of nonlinear systems of
Consequently, automatic control,” Automation Remote Control, vol. 22, pp.
( 857 - 875, 1962.
1 [9] I. W. Sandberg, “On the L2 -boundedness of solutions of
w· s−s0 if s0 ∈ R.
U (s) = ω0 (59) nonlinear functional equations,” Bell System Tech. Journal,
w· (s−σ0 )2 +ω02
if s0 ∈ C, vol. 43, pt. 11, pp. 1581 – 1599, July 1964.
[10] A. R. Teel, “A nonlinear small gain theorem for the analysis
We have of control systems with saturation,” IEEE Transactions on
(
G(s) · w · 1
if s0 ∈ R. Automatic Control, vol. 41, no. 9, pp. 1256–1270, 1989.
s−s0
Y (s) = ω0 (60) [11] G. Zames, “On the input-output stability for time-varying
G(s) · w · (s−σ0 )2 +ω02
if s0 ∈ C, nonlinear feedback systems, Part I, conditions derived using
concepts of loop gain, conicity and positivity,” IEEE Trans.
Since G(s) is assumed to be stable, Y (s) is a stable signal. on Automat. Contr., vol. 11, no. 2, pp. 228 – 238, 1966.
It is important to note that Y (s) does not have a pole at [12] K. Zhou, Essentials of Robust Control, Prentice Hall, Upper
s0 . The reason is that the pole at s0 is canceled by the zero Saddle River, 1998.
3445