Massachusetts Institute of Technology

Department of Electrical Engineering and Computer Science

∗
6.242, Fall 2004: MODEL REDUCTION

Balanced Truncation1
This lecture introduces balanced truncation for LTI systems: an important projection
model reduction method which delivers high quality reduced models by making an extra
effort in choosing the projection subspaces.

5.1 The classical balanced truncation algorithm

This section describes the basic algorithm for balanced truncation of continuous time
state space models.

5.1.1 Motivation: removal of (almost) uncontrollable/unobservable modes

Removing an unobservable or an uncontrollable mode is an easy way of reducing the
dimension of the state vector in a state space model. For example, system

ẋ1 = −x1 + x2 + f,
ẋ2 = −2x2 ,
y = x1 + x2 ,

can be replaced by
ẋ = −x + f, y = x
∗ c
A. Megretski, 2004
1
Version of September 27, 2004
 2

without changing its transfer function. In this state space model, with
   
−1 1 1  
A= , B= , C = 1 1 , D = 0,
0 −2 0

the controllability matrix

 
  1 0
Mc = B AB =
−1 0

satisfies pMc = 0 for p = [0 1], and hence the variable px = x2 represents an uncontrollable
mode. The removal of such mode can be viewed as a canonical projection model reduction

A 7→ Â = UAV, B 7→ B̂ = UB, C 7→ Ĉ = CV,

where the columns of V form a basis in the column range of Mc (which is the same as the
null space of p), and U can be selected quite arbitrarily subject to the usual constraint
UV = I.
Strictly speaking, the example above cannot even be considered as “model reduction”,
as the orders of the original and the projected systems are both equal to 1. A more
interesting situation is represented by the perturbed model

ẋ1 = −x1 + x2 + f,
ẋ2 = −2x2 + ǫf,
y = x1 + x2 ,

(same A, C, D but a modified B), where ǫ > 0 is a parameter. Intuitively, one can expect
that, when ǫ > 0 is small enough,

ẋ = −x + f, y = x

is still a good reduced model. This expectation can be related to the fact that x2 is
difficult to control by f when ǫ > 0 is small. One can say that the mode x2 = px, which
corresponds to the left (row) eigenvector of the A matrix (pA = −2p in this case), is
almost uncontrollable, which can be seen directly from the transfer function
s+2+ǫ 1+ǫ ǫ
G(s) = = − ,
(s + 2)(s + 1) s+1 s+2

which has a small coefficient at (s + 2)−1 in its partial fraction expansion.

 3

One can attempt to introduce a measure of importance of an LTI system’s mode

(understood as a pole a of system’s transfer function G(s)) as something related to the
absolute value of the coefficient with which (s − a)−1 enters the partial fraction expansion
of G. However, it is rarely a good idea to base a model reduction algorithm solely on
removal of “unimportant” system modes. For example, both modes a = 1−ǫ and a = 1+ǫ
of the LTI system with transfer function
1 1
H(s) = + ,
s+1−ǫ s+1+ǫ
where ǫ > 0 is small, are equally important, and none can be removed without introducing
a huge model reduction error, despite the fact that a very good reduced model Ĥ is given
by
2
Ĥ(s) = .
s+1
Balanced truncation is based on introducing a special joint measure of controllability
and observability for every vector in the state space of an LTI system. Then, the reduced
model is obtained by removing those components of the state vector which have the lowest
importance factor in terms of this measure.

5.1.2 Observability measures

Consider a state space model

ẋ(t) = Ax(t) + Bf (t), y(t) = Cx(t) + Df (t), (5.1)

where A is an n-by-n Hurwitz matrix (all eigenvalues have negative real part). When
f (t) ≡ 0 for t ≥ 0, the value of the output y(t) at a given moment t is uniquely defined
by x(0), and converges to zero exponentially as t → +∞. Hence the integral
Z ∞
Eo = |y(t)|2dt,
0

measuring the “observable output energy” accumulated in the initial state, is a function
of x(0), i.e. Eo = Eo (x(0)). Moreover, since y(t) is a linear function of x(0), Eo will be a
quadratic form with respect of x(0), i.e.

Eo (x(0)) = x(0)′ Wo x(0)

for some symmetric matrix Wo .

 4

The quadratic form Eo = E0 (x(0)) can be used as a measure of observability defined

on the state space of system (5.1). In particular, Eo (x(0)) = 0 whenever
Mo x(0) = [C; CA; . . . ; CAn−1 ]x(0) = 0.
The positive semidefinite matrix Wo of the quadratic form Eo is called the observability
Gramian2 . Since, by the definition, Eo (x(0)) ≥ 0 for all x(0), the matrix Wo is always
positive semidefinite. Indeed, Wo > 0 whenever the pair (C, A) is observable.
It is important to notice that the word “system state” actually includes two different
meanings. One meaning is that of a primal state, which, for a general model (5.1), is a
column n-vector. For example, for a state space model
ẋ1 = −x1 + f, ẋ2 = −x2 + f, y = x1 + x2 , (5.2)
a primal state is a particular column vector value of x(t) ∈ R2 , such as x(−1.2) = [1; −7].
Such column vector values are more precisely referred to as the primal states of (5.2).
On the other hand, one frequently refers to (5.2) as a “system with two states” (despite
the fact that the set R2 has an infinite number of elements), and, in this context, x1 = x1 (t)
and x2 = x2 (t) can be referred to as the two states of the system. Let us call such states
the dual states. For a general model (5.1), a dual state is a particular linear combination
xp = px(t) of the scalar components of x = x(t), defined by a row n-vector p. For example,
in (5.2), the dual states x1 = x1 (t) and x2 = x2 (t) are defined by row vectors p = [1 0]
and p = [0 1] respectively.
Therefore, it is natural to ask for a definition of an observability measure of a given
dual state of (5.1). It is defined as
E o (p) = inf Eo (x0 ) for p 6= 0,
x0 : px0 =1

i.e. as the minimal output energy which can be observed for t ≥ 0 when px(0) = 1. Note
that infimum over an empty set equals plus infinity, hence E o (0) = ∞. When the pair
(C, A) is observable, and hence Wo > 0 is invertible, the dual observability measure is
given by
1
E o (p) = for p 6= 0.
pWo−1 p′
The following theorem is frequently utilized for computing Wo numerically.

Theorem 5.1 Wo is the unique solution of the Lyapunov equation

Wo A + A′ Wo + C ′ C = 0. (5.3)
2
Whether this should be spelled as “Grammian” or “Gramian”, is unclear
 5

Proof Since system (5.1) is time invariant, the identity

Z ∞
′
x(0) Wo x(0) = |Cx(τ )|2 dτ
0

implies Z ∞
′
x(t) Wo x(t) = |Cx(τ )|2 dτ.
t
Differentiating the second identity with respect to t at t = 0 yields
2x(0)′ Wo Ax(0) = −|Cx(0)|2
for all x(0) ∈ Rn . Comparing the coefficients on both sides of the quadratic identity
yields (5.3).
Finding a numerical solution of (5.3) is not easy when n is about 104 and larger. In
such situation, Theorem 5.1 can be used as a basis for finding an approximation of Wo .
It is important to understand that the observability measure alone should not be the
only numerical test for choosing which states to eliminate in a model reduction procedure.
Instead, a combination of observability and a controllability measures, to be introduced
in the next subsection, should be used.

5.1.3 Controllability measures

Since A is a Hurwitz matrix, every input signal f = f (t) of finite energy, i.e. such that
Z ∞
2
kf k = |f (t)|2 dt < ∞,
−∞

corresponds to a unique initial condition x(0) in (5.1) for which the corresponding solution
x = x(t) satisfies x(t) → 0 as t → −∞. This solution is given by
Z ∞
x(t) = eAτ Bf (t − τ )dτ,
0

where eM denotes the matrix exponent of a square matrix M. One can say that input
f = f (t) drives the system state from x(−∞) = 0 to x(0) = X(f (·)).
Let p be a 1-by-n row vector, so that the product px(t) is a dual state of (5.1) – a
linear combination of components of the state space vector. The (dual) controllability
measure E c = E c (p) is defined as the maximal value of |px(0)|2 which can be achieved by
using an input f = f (t) of unit energy:
E c (p) = max{|pX(f (·))|2 : kf k ≤ 1}.
 6

Accordingly, the primal controllability measure Ec = Ec (x0 ) is defined for a column

vector x0 ∈ Rn as
Ec (x0 ) = inf E c (p).
p: px0 =1

The following statement describes some basic properties of these controllability mea-
sures.
Theorem 5.2 Assuming that A is an n-by-n Hurwitz matrix.
(a) E c (p) = pWc p′ is a quadratic form with the coefficient matrix
Z ∞
′
Wc = eAt BB ′ eA t dt.
0

(b) Wc is the unique solution of the Lyapunov equation

AWc + Wc A′ = −BB ′ . (5.4)

(c) A given state x0 ∈ Rn is reachable from zero if and only if Ec (x0 ) > 0 or, equiva-
lently, the equation Wc p′ = x0 has a solution p′ . In this case Ec (x0 ) = px0 is the
minimum of kf (·)k2 subject to X(f (·)) = x0 .
Proof To prove (a), note that
Z
max −∞∞ g(t)′ f (t)dt = kgk,
kf k≤1

hence Z ∞
c
E (p) = |peAt B|2 dt = pWc p′ .
0
Statement (b) is actually a re-wording of Theorem 5.1, with C replaced by B ′ , A
replaced by A′ , Wo replaced by Wc , and x(0) replaced by p.
To prove (c), consider first the case when equation Wc p′ = x0 has no solution p. Then
there exists a row vector p0 such that p0 Wc = 0 but p0 x0 6= 0. Here the equality means
that |p0 X(f (·))|2 equals zero for every finite energy signal f = f (t). Since, from the
inequality, |p0 x0 |2 > 0, the state x0 is not reachable from zero.
Now assume that x0 = Wc p′ for some p. Then kf k2 ≥ pWc p′ = px0 whenever x0 =
X(f (·)). On the other hand, for
 ′ −A′ t ′
Be p , t ≤ 0,
f (t) =
0, t > 0,
we have kf k2 = px0 and x0 = X(f (·)).
 7

When the pair (A, B) is controllable, and, hence, Wc > 0, the primal controllability
measure Ec = Ec (x0 ) can be expressed as
1
Ec (x0 ) = for x0 6= 0.
x′0 Wc−1 x0

5.1.4 Joint controllability and observability measures

The joint controllability and observability measures are defined as products of the corre-
sponding controllability and observability measures:

Eoc (x0 ) = Eo (x0 )Ec (x0 ), E oc (p) = E o (p)E c (p).

For controllable and observable systems Wc and Wo are positive definite, and the formulae
can be simplified to

x′0 Wo x0 oc pWc p′
Eoc (x0 ) = (x0 =
6 0), E (p) = (p 6= 0).
x′0 Wc−1 x0 pWo−1 p′

For model reduction purposes, we are interested in finding a subspace of primal state
vectors for which the minimum of the joint controllability and observability measure
over all non-zero elements is maximal. A basis in this subspace will yield columns of
a projection matrix V . Similarly, we are interested in finding a subspace of dual state
vectors for which the minimum of the joint controllability and observability measure over
all non-zero elements is maximal. A basis in this subspace will yield rows of a projection
matrix V .
The following theorem can be used in finding such V and U.

Theorem 5.3 Let Wc = Lc L′c and Wo = L′o Lo be the Choleski decompositions of the
controllability and observability Gramians. Let

ρ1 ≥ · · · ≥ ρr > ρr+1 ≥ · · · ≥ ρn ≥ 0

be the ordered eigenvenvectors of L′c Wo Lc (possibly repeated). Let ψ1 , . . . , ψr be the corre-

sponding first r normalized eigenvectors of L′c Wo Lc , i.e.

L′c Wo Lc ψi = ρi ψi , |ψi |2 = 1, ψi′ ψk = 0 for i 6= k.

1/2
Let σi = ρi .
 8

(a) ρ1 ≥ · · · ≥ ρr are also the eigenvalues of Lo Wc L′o , and the corresponding normalized
row eigenvectors can be defined by
φi = σi−1 ψi′ L′c L′o (i = 1, . . . , r).

(b) The set of all linear combinations of vectors Lc ψi is the only r-dimensional linear
subspace V in Rn such that Eoc (v) ≥ ρr for every v ∈ V.
(c) The set of all linear combinations of row vectors φi Lo is the only r-dimensional
linear subspace U of row n-vectors such that E oc (u) ≥ ρk for every u ∈ U.
(d) UV = Ir , where
−1/2
 
σ1 φ1
..
h i
−1/2 −1/2
V = Lc ψ1 σ1 . . . ψr σr , U =  Lo .
 
.
−1/2
σr φr

The proof of the theorem is obtained by inspection.

5.1.5 Balanced truncation

The projection model reduction algorithm which uses the projection matrices U, V de-
scribed in the previous subsection is called balanced truncation. The “balancing” termi-
nology comes from the following trivial observation.
Theorem 5.4 Assume that system (5.1) is both controllable and observable. Then Wc =
Lc L′c > 0 and Wo = L′o Lo > 0 are positive definite. Moreover, if Ψ is the orthogonal
matrix which columns are the ordered eigenvectors of L′c Wo Lc , and Σ is the diagonal
matrix with numbers σi on the diagonal, then the linear state transformation x = Sz,
with S = Lc ΨΣ−1/2 , yields an equivalent state space model with the coefficient matrices
Ā = S −1 AS, B̄ = S −1 B, C̄ = CS,
for which both the controllability and observability Gramians equal Σ.
State state space models for which Wo = Wc are equal diagonal matrices with ordered
positive diagonal entries σk are called balanced. The numbers σk are called the (Han-
kel) singular numbers of the corresponding system. The (canonical) method of balanced
truncation is based on removing the states of a balanced realization which correspond to
singular numbers below a certain threshold. Indeed, a practical implementation does not
have to involve a calculation of a complete balanced model: only the projection matrices
U, V are necessary.
 9

5.1.6 Example
Let
1 1
G(s) = + ,
s+1−ǫ s+1+ǫ
where ǫ > 0 is a small parameter. A state space model is
   
−1 + ǫ 0 1  
A= , B= , C = 1 1 , D = 0.
0 −1 − ǫ 1

The controllability Gramians are given by

1
 
1 1−ǫ
1
Wo = Wc = 1 .
2 1 1+ǫ

The Hankel singular numbers are the eigenvalues of Wo = Wc , and equal

√
1 1 ± 1 − ǫ2 + ǫ4
σ1,2 = .
2 1 − ǫ2
The corresponding eigenvectors are
 
1
v1,2 = 1 .
2σ1,2 − 1−ǫ

The dominant eigenvector defines

" #
√1 h i
V ≈ 2
, U≈ √1 √1 .
√1 2 2
2

The resulting reduced system is approximately given by

√ √
 ≈ −1, B̂ ≈ 2, Ĉ =≈ 2.

5.2 Properties of balanced truncation

In this section, basic properties of reduced models obtained via balanced truncation (exact
or approximate) are discussed.
 10

5.2.1 Approximate Gramians

The two most expensive phases of numerical calculations associated with the canonical
method of balanced truncation are finding the Gramians Wo , Wc , and calculating the
dominant eigenvectors ψi of L′l Wo Lc . for systems of large dimensions (more than 104
states) finding the Gramians exactly becomes difficult.
As a viable alternative, lower and upper bounds of the Gramians can be used to
provide provably reliable results. Here by lower bounds of Gramians Wo , Wc we mean
positive semidefinite matrices Wo− , Wc− for which the inequalities

Wo ≥ Wo− , Wc ≥ Wc−

are guaranteed. The definition of upper bounds will be more strict: by upper bounds of
Gramians Wo , Wc defined by the Lyapunov equalities

Wo A + A′ Wo = −C ′ C, AWc + Wc A′ = −BB ′ ,

where A is a Hurwitz matrix, we mean solutions Wo+ , Wc+ of the corresponding Lyapunov
inequalities
Wo+ A + A′ Wo+ ≤ −C ′ C, AWc+ + Wc+ A′ ≤ −BB ′ .
These inequalities imply that Wo+ ≥ Wo and Wc+ ≥ Wc , but the inverse implication is not
always true.
The following simple observation can be used to produce lower bounds of the Gramians.

Theorem 5.5 Let A be an n-by-n Hurwitz matrix. Let F be an n-by-m matrix. For
s ∈ C with Re(s) > 0 define
p
a = a(s) = (sIn − A)−1 (s̄In + A), b = b(s) = 2Re(s)(sIn − A)−1 B.

Then

(a) a is a Schur matrix (all eigenvalues strictly inside the unit disc);

(b) an n-by-n matrix P is a solution of the “continuous time” Lyapunov equation

AP + P A′ = −BB ′

if and only if it is a solution of the “discrete time” Lyapunov equation

P = aP a′ + bb′ ;
 11

(c) the matrix P from (b) is the limit

P = lim Pk ,
k→∞

where the symmetric matrices P0 ≤ P1 ≤ P2 ≤ · · · ≤ P are defined by

P0 = 0, Pk+1 = a(sk )Pk a(sk )′ + b(sk )b(sk )′ ,

and {sk } is a sequence of complex numbers contained in a compact subset of the

open right half plane.

The theorem reduces finding a lower bound of a solution of a Lyapunov equation to

solving systems of linear equations (used to produce b(sk ) and the products a(sk )Pk a(sk )′ .
Calculation of an upper bound of a Gramian could be more tricky. One approach to
finding such upper bounds relies on having a valid energy function for A available.
Indeed, assume that Q = Q′ satisfies

AQ + QA′ < 0.

If Wc− is a good lower bound of the controllability Gramian Wc , defined by

AWc + Wc A′ = −BB ′ ,

then
AWc− + Wc− A′ ≈ −BB ′ ,
and hence
A(Wc− + ǫQ) + (Wc− + ǫQ)A′ ≤ −BB ′
for some ǫ > 0 (which will, hopefully, be small enough). Then Wc− and Wc− + ǫQ are a
lower and an upper bound for the controllability Gramian Wc .

5.2.2 Lower bounds for model reduction error

A major benefit of doing balanced truncation is given by a lower bound of the error of
arbitrary reduced models (not only those produced via balanced truncation).

Theorem 5.6 Let Wo− = Fo′ Fo and Wc− Fc Fc′ be lower bounds of the observability and
controllability Gramians Wo , Wc of a stable LTI model G = G(s). Let

σ1− ≥ σ2− ≥ · · · ≥ 0
 12

be the ordered singular numbers of Fo Fc . Then σk− is a lower bound for the k-th Hankel
singular number σk = σk (G) of the system, and

kG − Ĝk∞ ≥ σk−

for every system Ĝ of order less than k.

Proof Let Zk denote the subspace spanned by the k dominant eigenvectors of Fc′ Wo− Fc ,
i.e.
|Fo Fc z| ≥ σk− |z| ∀ z ∈ Zk .
Since Wc ≥ Fc Fc′ , every vector Fc z lies in the range of Wc , and q ′ Fc z ≤ |z|2 whenever
Fc z = Wc q. Hence every state x(0) = Fc z can be reached from x(−∞) = 0 using a minimal
energy input f = fz (t) (depending linearly on z) of energy not exceeding |z|2 . On the
other hand, every state x(0) = Fc z with z ∈ Zk will produce at least |Fo Fc z|2 ≥ (σk− )2 |z|2
of output energy. Since Ĝ is a linear system of order less than k, there exists at least one
non-zero z ∈ Zk for which the input f = fz (t) produces a zero state x̂(0) = 0 at zero time.
Then, assuming the input is zero for t > 0, the error output energy is at least (σk− )2 |z|2 .
Since the testing input energy is not larger than |z|2 > 0, this yields an energy gain of
(σk− )2 , which means that kG − Ĝk∞ ≥ σk− .

5.2.3 Upper bonds for balanced truncation errors

The result from the previous subsection states that, for a stable LTI system G, no method
can produce a reduced model Ĝ of order less than k such that the H-Infinity error kG−Ĝk∞
is less than the k-th singular number σk = σk (G). The statement is easy to apply, since
lower bounds σk− of σk can be computed by using lower bounds of system Gramians.
The following theorem gives an upper bound of model reduction error for the exact
implementation of the balanced truncation method.

Theorem 5.7 Let σ1 > σ2 > · · · > σh be the ordered set of different Hankel singular
numbers of a stable LTI system G. Let Ĝ be the reduced model obtained by removing the
states corresponding to singular numbers not larger than σk from a balanced realization of
G. Then Ĝ is stable, and satisfies
h
X
kG − Ĝk∞ ≤ 2 σi .
i=k
 13

The utility of Theorem 5.7 in practical calculations of H-Infinity norms of model re-
duction errors is questionable: an exact calculation of the H-Infinity norm is possible at
about the same cost, and the upper bound itself can be quite conservative. Neverthe-
less, the theorem provides an important reassuring insight into the potential of balanced
truncation: since the singular numbers of exponentially stable LTI systems decay expo-
nentially, the upper bound of Theorem 5.7 is not expected to be much larger than the
lower bound.
For example, for a system with singular numbers σi = 2−i , a kth order reduced model
cannot have quality better than 2−k−1, and exact balanced truncation is guaranteed to
provide quality of at least 2−k+1 .
The proof of Theorem 5.7 is based on estimating the quality of balanced truncation
in the case when only the states of a balanced realization corresponding to the smallest
Hankel singular number are removed, which is done in the following technical lemma.

Lemma 5.1 Let W = W ′ > 0 be a positive definite symmetric n-by-n matrix satisfying
the Lyapunov equalities

W A + A′ W = −C ′ C, AW + W A′ = −BB ′ . (5.5)

Assume that W, A, B, C can be partitioned as

       ′ 
P 0 A11 A12 B1 ′ C1
W = , A= , B= , C = ,
0 σIr A21 A22 B2 C2′

where A22 is an r-by-r matrix, and matrices B2 , C2′ have r rows. Then

(a) transfer matrices

G(s) = C(sIn − A)−1 B, G1 (s) = C1 (sIn−r − A11 )−1 B1

are stable;

(b) the Lyapunov equalities

P A11 + A′11 P = −C1′ C1 , A11 P + P A′11 = −B1 B1′

are satisfied;

(c) kG − G1 k∞ ≤ 2σ.
 14

Proof It is sufficient to consider the case when the dimension m of f = f (t) equals the
dimension k of y = y(t). (If m < k, add zero columns to B, if m > k, add zero rows to
C.) First note that re-writing (5.5) in terms of the blocks Aik , Bi , Ci yields
P A11 + A′11 P = −C1′ C1 , (5.6)
P A12 + σA′21 = −C1′ C2 , (5.7)
σ(A22 + A′22 ) = −C2′ C2 , (5.8)
A11 P + P A′11 = −B1 B1′ , (5.9)
σA12 + P A′21 = −B1 B2′ , (5.10)
σ(A22 + A′22 ) = −B2 B2′ . (5.11)
Note (5.8) together with (5.11) implies that C2′ = B2 θ for some unitary matrix θ. Also,
(5.6) and (5.9) prove (b).
To prove (a), note that for every complex eigenvector v 6= 0 of A, Av = sv for some
s ∈ C, multiplication of the first equation in (5.5) by v ′ on the left and by v on the right
yields
2Re(s)v ′ W v = −|Cv|2 .
Hence either Re(s) < 0 or Re(s) = 0 and Cv = 0. Hence all unstable modes of A are
unobservable, and G = G(s) has no unstable poles. The same proof applies to G1 , since
A11 satisfies similar Lyapunov equations.
To prove (c), consider the following state space model of the error system G − G1 :
ẋ1 = A11 x1 + A12 x2 + B1 f,
ẋ2 = A21 x1 + A22 x2 + B2 f,
ẋ3 = A11 x3 + B1 f,
e = C1 x1 + c2 x2 − C1 x3 .
It would be sufficient to find a positive definite quadratic form V (x) = x′ Hx such that
dV (x(t))
ψ(t) = 4σ 2 |f (t)|2 − |e(t)|2 − ≥0
dt
for all solutions of system equations. Indeed, such Lyapunov function V can be readily
presented, though there is no easy way to describe the intuitive meaning of its format:
V (x) = σ 2 (x1 + x3 )′ P −1 (x1 + x3 ) + (x1 − x3 )′ P (x1 − x3 ) + 2σ|x2 |2 .
To streamline the derivation, introduce the shortcut notation
z = x1 + x3 , ∆ = x1 − x3 , δ = C1 ∆, u = σ −1 B2′ x2 , q = B1′ P −1 z.
 15

The equations now take the form

˙
∆ = A11 ∆ + A12 x2 ,
ż = A11 z + A12 x2 + 2B1 f,
ẋ2 = A22 x2 + 0.5A21 (z + ∆) + B2 f,
e = C1 ∆ + σθ′ u.
We have
ψ = 4σ 2 |f |2 − |C1 ∆ + σθ′ u|2 − 2σ 2 z ′ P −1 (A11 z + A12 x2 + 2B1 f )
−2∆′ P (A11 ∆ + A12 x2 ) − 4σx′2 [A22 x2 + 0.5A21 (z + ∆) + B2 f ]
= 4σ 2 |f |2 − |C1 ∆ + σθ′ u|2 + σ 2 |q|2 − 4σ 2 q ′ f +
|δ|2 + 2σ 2 |u|2 − 4σ 2 u′f − 2z ′ [σ 2 P −1 A12 + σA′21 ]x2 − 2∆′ [P A12 + σA′21 ]x2
= 4σ 2 |f |2 − |C1 ∆ + σθ′ u|2 + σ 2 |q|2 − 4σ 2 q ′ f + |δ|2
+2σ 2 |u|2 − 4σ 2 u′f + 2σ 2 q ′ u + 2σδ ′ θ′ u
≥ 4σ 2 |f |2 − |C1 ∆ + σθ′ u|2 + |δ|2 + 2σ 2 |u|2 + 2σδ ′ θ′ u − σ 2 |u − 2f |2
= 0
(the first identity transformation utilized (5.6), (5.9), (5.8), the second identity used (5.7),
(5.10), the third (inequality) applied minimization with respect to q, the last (identity)
depended on θ being a unitary matrix.
It is important to realize that the lemma remains valid when the original Lyapunov
equalities are replaced by the corresponding Lyapunov inequalities (of course, the equal-
ities in (b) will get replaced by inequalities as well). Indeed, Lyapunov inequalities
W A + A′ W ≤ −C ′ C, AW + W A′ ≤ −BB ′
are equivalent to Lyapunov equalities
W A + A′ W = −C̃ ′ C̃, AW + W A′ = −B̃ B̃ ′ ,
where  
  C
B̃ = B Bδ , C̃ = ,
Cδ
and Bδ , Cδ are chosen appropriately. Note that G(s) is a left upper corner block in the
transfer matrix
G̃ = C̃(sIn − A)−1 B̃.
Since H-Infinity norm of a transfer matrix is not smaller than H-Infinity norm of its block,
applying Lemma 5.1 to the system defined by A, B̃, C̃ yields the stated generalization.