Chapter 3

Small Gain Theorem and

Integral Quadratic Constraints
This chapter presents a version of the small gain theorem for L2 gains, and then proceeds
to introduce the general technique of working with quadratic dissipation inequalities.

3.1

Small Gain Theorems for L2 Gains

There exist many versions of the small gain theorem, distinguished by the type of gain
and the type of interconnection they apply to. One of the simplest versions applies to the
unit feedback interconnection on Figure 3.1, and claims that, assuming well-posedness, if
system has L2 gain < 1, then the unity feedback interconnection system (input
w, output y) has L2 gain not larger than /(1 ).
v

-+
f 6

Figure 3.1: Unity feedback interconnection

The small gain statement applies to DT and CT systems, and to both asymptotic
and fixed state versions of L2 gain. Of course, if the assumption about is in terms of
an asymptotic L2 gain then the conclusion about is also in terms of an asymptotic
L2 gain.
25

26

3.1.1

6.245 (Fall 2011) Small Gain

The Fixed State Small L2 Gain Theorem

Here is a formal presentation of a one version of the small gain theorem (the rest are
formulated and proven in a similar way).
m,m
Theorem 3.1. If system SDT
(X) has fixed state L2 gain < 1, and the unity
m,m

feedback system SDT (X) is well defined then the fixed state L2 gain of is not
larger than /(1 ).
1,1
(R) is the one step delay times a system defined by
Example 3.1. When SDT

(w, x0 ) = {y : y(0) = x0 , y(t + 1) = aw(t) t Z+ },

the unity feedback system is well defined for all values of parameter a R, and represents
the difference equation
y(t + 1) = ay(t) + v(t), y(0) = x0 .
The fixed state L2 gain of is = |a|. It can be shown (using the results on computing L2
gains of LTI models, to be discussed in the next chapter) that the fixed state L2 gain of is
= 1/(1 |a|) for |a| < 1, and = for |a| 1, which shows that the closed loop L2 gain
bound from Theorem 3.1 is exact in this case.

There are several ways to approach the proof of Theorem 3.1. To stress the generalizations to be discussed later in the chapter, we start with the following statement about
specific quadratic forms.
Lemma 3.1. For every h [0, 1) there exists d 0 such that
h2
|w y|2 |y|2 d(h2 |w|2 |y|2 ) 0
(1 h)2

w, y Rm .

(3.1)

Proof of Lemma 3.1

The left hand side of the inequality in (3.1) is a quadratic form = (x) with respect to the
vector x = [w; y] R2m :


 2
h2
h
2 I

dh
I

m
m
2
2
(1h)
(1h)



(x) = x Qx where Q =
h2
h2
Im
(1h)2 Im
d 1 + (1h)
2
(Im denotes the m-by-m identity matrix, and the prime means Hermitean conjugation, which
is the same as transposition for real vectors or matrices, and the elements of Rn are interpreted as

6.245 (Fall 2011) Small Gain

27

n-by-1 real matrices). By construction, (3.1) is equivalent to Q 0 being a positive semidefinite

matrix. Using the fact that 2-by-2 block matrix


a b
Q=
b c
is positive semidefinite if
c = c > 0, a = a bc1 b ,
the statement follows by using d = 1/(1 h).

Proof of Theorem 3.1

Consider arbitrary h (, 1), T Z+ , and y (w, x0 ) (which also means y (w + y, x0 )),
substitute y = y(t), w = w(t) into the inequality from (3.1), and then take a sum from t = 0 to
t = T , to get
T
X

t=0

T
X

 2

2 |w(t) y(t)|2 |y(t)|2 d
h
h |w(t)|2 |y(t)|2 ,

(3.2)

t=0

= h/(1 h). Since the fixed state L2 gain of is less than h, the sum on the right side
where h
of (3.2) is bounded from below by a constant c = c(x0 ) depending on the boundary condition x0
only. Since d 0, this means that the sum on the left side of (3.2) is bounded from below by the
constant c = dc(x0 ) which depends on the boundary condition x0 only. Since w y = v, this
= h/(1 h) for every h (, 1).
means that the fixed state L2 gain of is not larger than h

As h , we conclude that the fixed state L2 gain of is not larger than /(1 ).

3.2

Integral Quadratic Constraints

Given a system S S m,k (X) and a quadratic form : Rm+k R, let us say that
S satisfies the fixed state integral quadratic constraint (IQC) defined by if for every
x0 X there exists a constant c = c(x0 ) such that the inequality
T
X
t=0

([w(t); y(t)])dt c

(3.3)

is satisfied for all w m and y S(w, x0 ).

The proof of Theorem 3.1 follows a certain path. First, the assumptions and the
conclusion to be proven are viewed as IQCs: the original L2 gain bound means that
satisfies the IQC defined by
1 ([w; y]) = h2 |w|2 |y|2

28

6.245 (Fall 2011) Small Gain

while the objective is to prove that satisfies the IQC defined by

0 ([w; y]) = r|w y|2 |y|2
for r as small as h/(1 h).
The second step is pure manipulation with parameter-dependent quadratic forms: we
consider the function : R2m R2 R defined by
([w; y], [r; d]) = 0 ([w; y]) d1 ([w; y]) = r|w y|2 |y|2 d(h2 |w|2 |y|2 )

(3.4)

(this = (v, x) is a quadratic form with respect to its first argument v = [w; y] R2m ,
and an affine function of its second argument x = [r; d] R2 ), and try to find the value of
the second argument x = [r; d] such that d 0, and the resulting quadratic form of the
first argument v R2m is positive semidefinite. For every r, d satisfying these conditions,

r is an upper bound for the L2 gain of .

Since we want the smallest possible upper bound for the closed loop L2 gain, the
objective is to minimize r subject to the constraints. In the proof of Theorem 3.1 , this
is done analytically (in fact, the algebra behind the optimization is hidden, and the
explicit optimizer r = h/(1 h), d = 1/(1 h) is given). Similar analytical proofs can
be given to many similar statements concerning stability of systems described in terms of
IQC. An alternative, once a specific value of h is given, would be to solve the optimization
task
r min subject to d 0, (v, [r, d]) 0 v
numerically, using appropriate optimization software, to be discussed later in this chapter.
Of course, this approach is not limited to the fixed state IQC in discrete time. Similar
definitions are available in the asymptotic format (where the constant c in (3.3) is
allowed to depend on w and y, but not on T , i.e. c = c(x0 , w, y)), and for CT models
(where the sums from t = 0 to t = T are replaced by integrals from t = 0 to t = T ).

3.2.1

Example: Passive Feedback

The following classical statement on feedback system stability is another example of using
IQCs. For variety, it will be presented in an asymptotic continuous time setting.
m,m
Let us call system SCT
(X) passive if
Z T
inf
y(t) w(t)dt >
(3.5)
T 0

for every w L , x0 X, and y (w, x0 ). The terminology is motivated by the

situation when the components of y are currents, and the corresponding components of

6.245 (Fall 2011) Small Gain

v

29
w

-f
6

Figure 3.2: Interconnection for Theorem 3.2

w are voltages, at all ports of a passive circuit. Then the integral in (3.5) represents the
energy supplied into the circuit on the time interval [0, T ], which natuarally is bounded
from below by minus the initial value of energy stored within the circuit.
It turns out that, when well-posed, a feedback interconnection with a passive circuit
subsystem , as shown on Figure 3.2, has L2 gain not larger than 1.
m,m
Theorem 3.2. If system SCT
(X) is passive, and the system

+ (v, x0 ) = {w : v w (w, x0 )}
is well defined, then the L2 gain of + isnot larger than 1
Proof of Theorem 3.2
By assumption, the asymptotic IQC defined by
1 ([w; y]) = w y
is satisfied for . Since v = w + y, the objective is to prove that satisfies the IQC defined by
0 ([w; y]) = |w + y|2 |w|2 .
Indeed, the latter IQC is certified by the fact that the quadratic form = 0 21 , i.e.
([w; y]) = |w + y|2 |w|2 2w y = |y|2
is positive semidefinite.

3.3

IQCs and Semidefinite Programming

One convenient way of setting up an optimization problem (or program) is by using the
format
Lx min subject to x Q, Ax = B,
(3.6)
x

30

6.245 (Fall 2011) Small Gain

where Q is typically a relatively simple subset Q of Rm , A is a real m-by-n matrix, L is a

1-by-m row vector, and B Rn . The task is usually to figure out whether the program
is feasible, i.e. if an element x Q such that Ax = B does exist, and, if feasible, finding
an element x Q such that Ax = B and Lx is minimal (or at least close to minimal in
some sense).
For example, a linear program can be defined in the form (3.6) with Q being the
m
with non-negative
positive quadrant Q = Rm
+ , i.e. the subset of vectors x R
coefficients.
Let us fix a particular way of re-arranging the entries of m-by-1 column vectors,
where m = k(k + 1)/2, to form real symmetric k-by-k matrices. For example, we can have



x1
x1 x2

x2

=
.
x2 x3
x3
Let Sk+ be the set of all vectors x Rm for which (x) is a positive semidefinite matrix.
When Q = Sk+ for some k, the task (3.6) is called a semidefinite program.
This definition is general enough, but it turns out to be both convenient and more
efficient to view Q as the set of vectors

x1

x = ... , xi Qi ,
xr
where each of the Qi is, for some d = di , one of the following:
(a) Rd ;
(b) Rd+ ;
(c) Sd+ ;
(d) the set

(e) the set

x0
Lk = . . . Rk+1 :

xk

x0
Lk = . . . Rk+1 :

xk

x0 0, x20 x21 + + x2k

x0 0, x1 0, x0 x1 x22 + + x2k

6.245 (Fall 2011) Small Gain

31

For example, the task of finding real numbers r, d such that d 0 , the quadratic form
in (3.4) is positive definitene, and r is as small as possible (something motivated by the
small gain theorem), can be viewed as a semidefinite program with

r
2

d : r R, d R+ , s S .
Q=

s
In the resulting task (3.6) we have n = d(d + 1)/2 with d = 2,

 


r
2
s
r
s
r

dh
1
2
1

B = 0, A d = ( (s)) =
r
r1+d
s2 s3
s
(so that Ax = B simply states that is positive semidefinite), and

r

L d = r.
s

3.3.1

The spot Toolbox

There are many tools for solving semidefinite programs numerically. Currently, withing
the MATLAB environment, one of the most commonly used free semidefinite program
solvers is SeDuMi. Since representing semidefinite programs in the SeDuMi-prescribed
format can be time-consuming and error-prone, additional tools were developed to help
with problem description (in particular, the free YALMIP toolbox appears to be popular).
This section describes the use of an alternative MATLAB toolbox, spot, which will be
made available from the 6.245 class locker. Here is the code for solving the semidefinite
program associated with the small gain theorem. It returns the best upper bound for
the closed loop L2 gain, when stability can be establishes (subject to the well-posedness
assumption), or [].
function g=spotsample0(h)
r=msspoly(r);
%
d=msspoly(d);
%
s=msspoly(s,[3,1]);
%
pr = mssprog;
%
pr.free = r;
%

symbolic [r]
symbolic [d]
symbolic [s1;s2;s3]
initialize program
register r as free

32

6.245 (Fall 2011) Small Gain

pr.pos = d;
% register d as positive
pr.psd = s;
% register s as semidefinite
S = mss_v2s(s);
% S=[s1 s2;s2 s3]
pr.eq = [r-d*(h^2) -r; -r r-1+d]-S;
% this is 0
pr.sedumi = r;
% minimize r
g=sqrt(pr({r}));
% get optimum
Here is a different code, which is mostly indented to save programmers brains by
skipping some algebra done by hand (it refers directly to quadratic form (3.4) instead of
its matrix). In this case, it turns out that it also generates a smaller size SeDuMi program.
function g=spotsample1(h)
r=msspoly(r);
% symbolic [r]
d=msspoly(d);
% symbolic [d]
s=msspoly(s,[3,1]);
% symbolic [s1;s2;s3]
w=msspoly(w);
% symbolic [r]
y=msspoly(y);
% symbolic [y]
z=[w;y];
pr = mssprog;
% initialize program
pr.free = r;
% register r as free
pr.pos = d;
% register d as positive
pr.psd = s;
% register s as semidefinite
S = mss_v2s(s);
% S=[s1 s2;s2 s3]
% equality constraint
pr.eq = r*(w-y)^2-y^2-d*(h^2*w^2-y^2)-z*S*z;
pr.sedumi = r;
% minimize r
g=sqrt(pr({r}));
% get optimum
Note the difference between registered variables r,d,s and unregistered variables w,y.
When handling equality constraints which involve unregistered variables, spot generates
equality constraints which are equivalent to the expression being zero for all possible
values of unregistered variables.
Example: Robustness to Structured Uncertainty
Consider the feedback interconnection from Figure 3.3, where v, w, q, y are signals from
m,m
Lm , M SCT
({0}) is the linear system defined by a real given m-by-m real matrix M0 :
M (w, 0) = {M0 w},

6.245 (Fall 2011) Small Gain

33

m,m
1,1
and SCT
(X) is the parallel connection of m systems i SCT
(Xi ) (i = 1, . . . , m),
in the sense that X = X1 Xm , and

y
q1
1

..

..
. , (x1 , . . . , xm ) = . : yi i (qi , xi ) .

qm
m

Assume furthermore that L2 gain of each of the subsystems i is known to be smaller

than 1. The objective is to establish L2 gain stability of the closed loop (as usually,
subject to the well-posedness assumption).
v

-f 6

Figure 3.3: M/ interconnection

by

Since the L2 gain bound for each i establishes that M satisfies the IQC defined
i ([w; y]) = |ei M w|2 |ei y|2

(i = 1, 2, . . . , m),

where ei is the 1-by-m row vector with all 0 elements except 1 at the i-th position.
On the other hand, the closed loop L2 gain is not larger than 0 if M satisfyies
the IQC defined by
0 ([w; y]) = r|w y|2 |y|2
with r = 2 .
The standrad approach to IQC analysis suggests searching for non-negative constants
d1 , . . . , dm and r R such that the quadratic form
= 0

m
X

d i i

i=1

is positive semidefinite. To simplify notation, it is useful to note that

m
X
i=1

di i ([w; y]) = (M w) D(M w) y Dy,

34
where

6.245 (Fall 2011) Small Gain

D=

d1
0

0
...
dm

is the diagonal matrix with entries di on the diagonal.

The following spot code implements minimization of r subject to the constraints.
function r=spotsample(M)
% function r=spotsample(M)
% M must be a real m-by-matrix
m=size(M,1);
% problem dimension
r=msspoly(r);
% [r]
w=msspoly(w,[m,1]);
% [w1;w2;...;wm]
y=msspoly(y,[m,1]);
% [y1;y2;...;ym]
d=msspoly(d,[m,1]);
% [d1;d2;...;dm]
q=msspoly(Q,nchoosek(2*m+1,2)); %[q1;...]
pr = mssprog;
% initialize program
pr.free = r;
% register r
pr.pos = d;
% register d
pr.psd = q;
% register q
D = diag(d);
% diagonal D
Q = mss_v2s(q);
% symmetric Q
z = [w;y];
v = w-y;
Mw = M*w;
pr.eq = r*v*v-y*y-Mw*D*Mw+y*D*y-z*Q*z;
pr.sedumi = r;
% minimize r
r=pr({r});
% L2 gain bound