Professional Documents
Culture Documents
Spaces and Norms in Robust Control PDF
Spaces and Norms in Robust Control PDF
(3.5)
In control theory the 1-, 2- and oo-norm are most important since they have
obvious physical interpretations:
n
Ilxlll = L lXii, resource. (3.6)
i=1
n
IIxll2 = L IXil2= ";x*x , energy (3.7)
i=1
Ilxli oo = m?X lXii, peak. (3.8)
•
In (3.7), x* denotes the complex conjugate transpose of x. Notice that the
2-norm IIxl12 is the usual Euclidean length of the complex vector x. All norms
on en are equivalent norms which means that if II . lin and II . 11,6 are norms
on en, then there exists a pair CI, C2 > 0 so that
"Ix E en. (3.9)
In particular, "Ix E en:
IIxll2 :5 Ilxlll :5 v'nllxlb (3.10)
Ilxli oo :5 IIxll2 :5 v'nllxll oo (3.11)
Ilxli oo :5 Ilxlh :5 nllxll oo . (3.12)
Now let us consider the space H = e mxn , namely the space of m x n
complex matrices. e mxn is also a linear space. Matrix p-norms on e mxn are
defined in terms of the p-norms for vectors on en:
IIAxll p e mxn .
IIAllp = sup -11-11-'
xECn,x#O X p
VA E (3.13)
Notice that the matrix p-norms are induced norms. They are induced by the
corresponding p-norms on vectors. One can think of IIAllp as the maximum
gain of the matrix A measured by the p-norm ratio of vectors before and
after multiplication by A. In general matrix p-norms are difficult to compute.
However, for p = 1,2, or 00 there exist simple algorithms to compute IIAllp
exactly. If A = [aij] E e mxn we have
m
The matrix 2-norm and F-norm are invariant under multiplication by unitary
or orthogonal matrices. Assume that Q*Q = I and Z* Z = I for Q E c mxm
and Z E c nxn . Then
In modern control theory singular values have been used to extend the clas-
sical frequency response Bode plot to multivariable systems. Consider the
input-output relation
where G(s) is a transfer function matrix. How do we then evaluate the fre-
quency response G(jw)? An obvious choice would be to pick one of the in-
duced matrix norms introduced previously. The 1-, 2- and <Xl-norm all have
potential engineering applications. However, the control theory for using them
in design or analysis is only well-developed for the 2-norm. Thus let us eval-
uate the frequency response of G(s) at the frequency w by
16 3. Spaces and Norms in Robust Control Theory
(3.26)
Letting 0 ::; w ::; 00 we may compute the matrix 2-norm for every w to
obtain a upper bound for the "gain" of the transfer matrix G(s). However,
we would like to have a lower bound on G(jw) as well. This lower bound can
be obtained with the minimum singular value given by
Thus if we measure the "gain" of the system G(s) as the 2-norm ratio of the
input and output, then the maximum and minimum singular values of G(jw)
will constitute upper and lower bounds on this gain. In fact, we may assess
the system "gain" even more detailed using the singular value decomposition
given below. Let us, however, first introduce the following important lemma
which establishes the relation between the singular values and the eigenvalues
of a complex matrix.
Lemma 3.1 (Singular values and Eigenvalues). The singular values of
a complex matrix A E c mxn , denoted ai(A), are the k largest nonnegative
square roots of the eigenvalues of A*A where k = min{n,m}. Thus
i = 1,2,··· , k. (3.28)
It is usually assumed that the singular values are ordered such that ai ~ ai+!.
Thus
_ IIAull2
a(A) = 0'1 (A) = uECn,U,tO
sup -11-11- = IIAII2
U 2
(3.29)
![(A)
.
= ak(A) = uEcn,u,to
mf
IIAul12 = II A _111-21 '
-11-11- if A-I exists. (3.30)
U 2
The ratio between the maximum and minimum singular value is denoted the
condition number K,:
a(A) (3.31 )
K,(A) = ![(A) .
n
G= Y~U* (3.32)
= LYi(J'iUi (3.34)
i=l
where
~k: diag((J'1,(J'2,'" , (J'k).
Yl -+ Ym : The m columns of Y.
u~ -+ u~ : The n rows of U* .
This is known as the singular-value decomposition (SVD) of the matrix
G.
An interpretation for the SVD of a real matrix A is as follows. Any real
matrix A, looked at geometrically, maps a unit radius hyper-sphere into a
hyper-ellipsoid. The singular values (J'i(A) give the lengths of the principal
axis of the ellipsoid. The singular vectors Yi give the mutually orthogonal
directions of these major axes and the singular vectors Ui are mapped into
the Yi vectors with gain (J'i, that is, AUi = (J'iYi·
Example 3.1 (SVD of a Real Matrix). This example is taken from [MZ89].
Let A be given by
1.5 1.5
0.5 0.5
0 0
-0.5 -().5
-I -1
-1.5 -1.5
-2 -2
-2 -I 0 2 -2 -I 0 2
1
a(G) = ,q:(G-l) (3.40)
1
,q:(G) = a(G-l) (3.41)
properties of where
This illustrates that if the input vector u(jw) is in the direction of Uj then the
gain of the system is precisely O"j and the output vector y(jw) is then precisely
in the direction of Yj. The sets {Ul, U2, ... ,un} and {Yl, Y2, ... ,Ym} are not
surprisingly known as the input respectively output principal directions. The
singular values O"i are also known as the principal or directional gains of the
system matrix G.
Thus when G(s) is a transfer function matrix we can plot the singular
values O"i(G(jW)) for i = 1,··· ,k as functions of frequency w. These curves
are the multivariable generalization of the S1SO amplitude-ratio Bode plot.
For multivariable systems the amplification of the input vector sinusoid ue jwt
depends on the direction of the complex vector U as illustrated above. The
amplification is at least Q.(G(jw)) and at most a(G(jw)). The condition num-
ber K,(G(jw)), see (3.31), plotted versus frequency w outlines the system gain
sensitivity to the direction of the input vector. If K,(G(jw)) » 1 the gain of
the transfer function matrix will vary considerably with the input direction
and G(s) is said to be ill-conditioned. Conversely, if K,(G(jw)) ~ 1, Vw ~ 0,
the gain of the transfer matrix will be insensitive to the input direction and
the system is said to be well-conditioned. A well-conditioned multi variable
system behaves much like a single-variable system and controller design for
well-conditioned system are fairly straightforward. For ill-conditioned sys-
tems, however, much more care has to be taken in both design and analysis.
The norm concept introduced in Section 3.2 for constant vectors and matrices
can also be extended to time- and frequency domain functions. Such norms
will be denoted operator norms.
The reader may verify that the above operator norms fulfill the conditions
given in Section 3.1. Notice that we use the notation Ilfll for the operator
norm and the notation IIf(t)1I for the norm of the value of f at time t. Apart
from these norms, the following measure is often used in control theory:
Ilfllp = lim
T-,>oo
~ IT J2(t)dt.
2T_T
(3.54)
Ilfllp is known as the power or RMS value of f. However, IIfllp is not a true
norm since property (ii) in Section 3.1 is not fulfilled. II flip is often denoted
a semi-norm since it fulfill the remaining three properties. The power norm
is useful since it is often defined where the 2-norm is not, eg for signals with
limited but persistent power.
Equivalently, we may also define operator norms for scalar frequency do-
main functions f(s). Here only the 2- and 00 norm are of practical interest:
IIfl12 = .2.-1
271"
00
-00
f*(jw)f(jw)w.,; (3.55)
11f112 = [ : J2(t)dt = 1
271" 100
-00 f*(jw)f(jw)dw = IIfll2 (3.57)
such that the operator 2 norm of a time-domain function and its Laplace
transform is the same. This fact is one of the main reasons for using the 2
norm in control theory. Note that a equivalent statement does not hold for
the 00 norm.
Finally, we can introduce the following 2 transfer function on a stable
transfer functions G(s):
IIG(s)lh-l2 = 271"
1 JOO G*(jw)G(jw)dw
-00 (3.58)
i:
initial conditions are zero, then the output y(t) is given by
where a.,/3 can be any combination of 1,2,00 and P. The induced norm of 9
is thus the maximum gain of the system if we measure the input u using the
operator norm a. and the output y using the operator norm /3.
Using the 4 operator norms defined above we may thus construct 16 dif-
ferent induced norms for measuring the gain of the system with impulse
response g(t). However, some of these induced norms are trivially 0 or 00.
Let for example the input u is measured by the 00 norm and the output y
by the 2 norm. It is then clear that the induced norm Ilg1100-+2 = 00 since
u(t) -1+ 0 for t ~ 00. Let us consider the case where both the input u and
the output yare measured using the 2 norm 2 • Then by Parceval
=
JJ~oo y*(jw)y(jw)dw
sup
u(jw)#O JJ~oo u*(jw)u(jw)dw (3.63)
VJ::'oo u(jw)*IG(jw)l2u(jw))dw
= sup
u(jw)¥O VJ::'oo u*(jw)u(jw)dw (3.65)
= IIG(s)IIHoo . (3.67)
The transfer function 11.00 norm thus measures the maximum possible
increase in signal energy (as measured by the operator 2-norm) for all possible
inputs for which the 2-norm exists. We may say that the transfer function
11.00 norm is induced by the operator 2-norm. The 11.00 norm measures the
maximum gain of the system for a class of input signals characterized by their
operator 2-norm.
Unlike the 11.00 norm, the 11.2 transfer function norm is not induced by
any operator norm. In this sense, it corresponds to the Frobenius for matrices
which is not induced by any vector norm. Assume that the input u(t) = c5(t),
a unit impulse. Then the output is given by y(t) = g(t) such that lIyl12 = IIg1l2'
But by Parceval we have that IIgll2 = IIG(s)IIH2' Consequently, the transfer
function 11.2 norm measures the operator 2-norm of the output given a specific
input signal, namely a unit impulse.
Another interpretation of the 11.2 goes as follows. If u(t) is unit variance
white noise then the output power norm IIYllp = IIG(s)IIH2' Thus the trans-
fer function 11.2 also measures the RMS value of the output y(t) given unit
variance white noise on the input u(t). All quadratic norm control schemes
like Linear Quadratic optimal control and Kalman-Bucy optimal filtering
minimizes the 11.2 norm of a closed loop transfer function.
Remark 3.1. Throughout this section we have emphasized that the transfer
function G(s) must be stable. Does this then mean that these methods cannot
be applied to unstable plants? The answer is no! In control design analysis
the transfer function norm measures are applied to closed loop systems. Then
it is clearly a reasonable assumption that the system is stable.
In Table 3.1 we have shown induced system gains for all combinations of
the 2, 00 and power.
lIy II
00 IIG(s) 111£2 IIgll1 00
11/111,1 = I: = I: t
I: I: J
II/(t)1I1dt Ih(t)ldt (3.68)
I: I: mrx
1I/1h,2 = II/(t)1I 2 dt = IT(t)/(t)dt (3.69)
11/112,00 = I: II/(t)lI~dt =
n
I: (mrx Ih(t)lf dt (3.73)
The operator norm 11/11"",8 thus denotes the norm we get by using the
vector f3 norm at each time instant t and the scalar operator norm a over
time.
The power II/l1p naturally extends to
(3.77)
where III (t) 112 is the usual Euclidean length of the vector 1(t).
Furthermore, for transfer function matrices the norms (3.58).and (3.59)
extend to
i:
24 3. Spaces and Norms in Robust Control Theory
i:
the output vector y(t) is given by the convolution integral
(3.81)
where ai, a2, /31, /32 can be any combination of 1, 2, 00, P. The induced opera-
tor norm IIg(t)II(C>1,C>2)-+(i3d3 2 ) thus measures the maximum possible "gain" of
the system if the input is measured by IlullC>1,C>2 and the output by lIylli3t.i32'
The introduced concept of induced operator norms thus provides 100 ways
of measuring the gain of a multivariable system. There are well motivated
reasons for applying several of these induced norms in control theory, see
Example 3.2. Others are trivially zero or 00.
Example 3.2 (Induced Norms). Consider a multivariable control system with
the purpose of minimizing the consumption of a given resource e(t), eg fuel.
It will then be reasonable to measure e using the norm
(3.82)
Furthermore, assume that the disturbances d(t) on the system can be well
i:
described by their energy. Then the norm
where s is the impulse response matrix of the transfer function S(s) from
d(s) to e(s). •
3.4 Banach and Hilbert Spaces 25
where both input and output are measured by II . 112,2, The II . 112,2 norm
measures the energy of the signal and the induced operator 2 norm thus
measures the maximum possible increase in signal energy for all possible
inputs u(t) for which the 2-norm is defined. As for scalar systems, we have
the very important relationship:
IIgll(2,2)--+(2,2) = IIG(s)lI1i oo = sup a(G(jw)) . (3.86)
wER
The following lemma recapitulate our result.
Lemma 3.3 (Transfer Function Hoo Norm). LetG(s) be a (stable) mul-
tivariable transfer function matrix. If both input u(t) and output yet) are
measured by the 2 norm 11,112,2, then the maximum gain of the system for all
inputs u(t) with lIu1l2,2 ~ 00 is given by the transfer function Hoo norm:
lIy1l2,2
IIG(s)ll1ioo
= sup -II-11- .
u,to u 2,2
(3.87)
In the following we will define some spaces whose elements are time- or
frequency domain functions which are useful in modern control theory. Let
us first, however, introduce the concept of Banach and Hilbert spaces.
Let (H, II . II) be a normed space. Having defined a norm II . II we can assess
convergence in H. A sequence Un}, n = 1,2,··· in H converges to I E H
and I is the limit of the sequence if the sequence of real positive numbers
111- Inll converges to zero. If such I exists, then the sequence is convergent.
A sequence Un}, n = 1,2,··· in H is called a Cauchy sequence if for
each f > 0 there exists a natural number no E N such that for n, m 2:
no we have Illn - 1m II ~ f. Intuitively, the elements in a Cauchy sequence
eventually cluster around each other. They are "trying to converge" . Clearly,
every convergent sequence is a Cauchy sequence. However, in an arbitrary
normed space not every Cauchy sequence is convergent. A normed space
(H, II . II) is said to be complete if every Cauchy sequence is convergent. A
pre-Hilbert space (H, < .,. » is said to be complete if it is complete with
respect to the norm induced by the scalar product < .,. >.
A complete normed space is called a Banach space and a complete pre-
Hilbert space is denoted a Hilbert space.
Example 3.3. cn and Rn are Banach spaces under the norms 11·111, II . 112
and II . 1100 defined by (3.6)-(3.8). •
Example 3.4. cn and Rn are Hilbert spaces under the scalar product
m
•
3.5 Lebesgue and Hardy Spaces
In Section 3.2 we only considered the space c mxn , i.e spaces whose elements
are constant matrices (or constant vectors as special cases). However, the
concepts introduced above apply also to linear spaces whose elements are
time domain or frequency domain functions (or operators). Two such spaces
are important in robust control theory, namely Lebesgue and Hardy spaces.
In 1i00 control theory, 1i stands for Hardy.
3.5 Lebesgue and Hardy Spaces 27
i:
to be square-Lebesgue integrable:
such that the operator 2 norm Ilf1l2,2 exists, see Equation (3.72). The set
of all such signals is a Banach space under the norm 111112,2. This space is
called the Lebesgue space .c 2(R, Rn). The first argument gives the domain of
('D(f)), or input space, and the second gives the range of (R(f)), or output
space. Quite often in the literature these arguments are dropped for brevity.
Thus, which version of the Lebesgue space is meant must be determined form
the context. To avoid any confusing we will keep the arguments here.
The subspace of .c 2 (R, R n) for which f (t) = 0, Vi < 0 (causal time func-
tions) is the Hardy space 1l2(R, Rn) under the norm IlfI12,2.
Similarly, we will consider spaces whose elements are time functions with
finite operator 1- and 00 norms:
- .cdR, Rn) is the (Banach) space of absolute-Lebesgue integrable vector-
i: i: t
valued functions of time:
under the operator I-norm IlfI11,1, see Equation (3.68). 1l 1 (R,Rn) is the
subspace of .c 1(R,Rn) for which f(t) = 0, Vi < o.
- .coo(R, Rn) is the (Banach) space of bounded real vector-valued functions
of time:
supm?Xlfi(t)1 < 00 (3.97)
tER '
<0
iz(t) = { ~ "It
"It ~ 0
(3.99)
28 3. Spaces and Norms in Robust Control Theory
Notice that f2(t) is the ordinary step function. Now ft(t) belongs to all
three spaces 1l1(R,R), 1i2(R,R) and 1loo (R,R). Both its absolute and its
square are integrable from a to 00 and it has finite peak. In contrast, the
step function f2(t) is a member of 1loo(R, R) only, since neither of the 1- or
2-norms have finite values. ft(t) and f2(t) is shown in Figure 3.2. •
I.S I.S
1 1
O.S .. o.S .
3
~ 0 r;:l 0
-0.5. -0.5 ..
-I -I
-I.S -I.S
-s 0 S 10 -s 0 5 10
Time [sees] Time [sees]
As mentioned above, in both robust control theory and the older LQG the-
ory, only the operator 2 norm is used for time-domain signals. Only signals be-
longing to 1l2(R, Rn) is thus considered. The time domain spaces Hi (R, Rn)
and 1loo (R, R n) do have potential engineering applications but the control
theory for using them in design or analysis is not yet well-developed even
though it is an area of intensive research. Since 111 (R, Rn) can represent fuel
usage, and 1l oo (R, Rn) can treat the issue of actuator or sensor saturation,
there are strong motivations for developing such theories.
i:
Example 3.6. L2(R, Rn) is a Hilbert space under the scalar product
i:
The corresponding induced norm is the 2-norm since
•
3.5 Lebesgue and Hardy Spaces 29
Now consider the vector valued function f(jw) which is defined for all fre-
quencies -00 < w < 00 (that is, on the imaginary axis), takes values in C n
and is square-Lebesgue integrable on the imaginary axis. The space of all
such functions is a Hilbert space under the scalar product
1 /00 f(jw)*g(jw)dw.
< f,g >= -2 (3.lO2)
1f -00
It is denoted the Lebesgue space £2 (jR, C n ). The corresponding induced
norm IS
Next, 1-l2(C, cn) is the (Hardy) space of all functions f(s) which are analytic
in ~e(s) > 0 (i.e., no right half plane (RHP) poles), take values in C n and
satisfy the uniform square-integrability condition
1
sup -2
e
/0000 f(~ + jw)* f(~ + jw)dw <
1f_
00 . (3.104)
1-l2 (C, cn) is a Hilbert space under the scalar product (3.102) and with in-
duced norm (3.103).
Because of Parceval's theorem 1-l 2(C, cn) is just the set of Laplace trans-
forms of signals in 1-l2(R,Rn). Furthermore, if f(t) E 1-l 2(R,R n) then its
Laplace transform f(s) E 1-l2(C, cn).
Finally we will present 4 frequency domain spaces of matrix-valued func-
tions.
- £2(jR, c mxn ) is the space of all functions F(jw), defined on the imaginary
axis which take values in c mxn and are square-integrable on the imaginary
i:
axis. £2 (jR, c m x n) is a Hilbert space under the scalar product
Note that we have used subscript 1-l2 in (3.lO6). However, sin<:€ F(jw) is
only defined on the imaginary axis it is not a transfer function. Rather it
is a frequency response. On the other hand, the norm (3.lO6) is identical
to the transfer function 2-norm in (3.78). Thus the given notation.
30 3. Spaces and Norms in Robust Control Theory
- 1i2(C, c mxn ) is the space of all functions F(s) which are analytical in
~e(s) > 0 (no RHP poles), take values in c mxn and satisfy the uniform
square-integrability condition
sup 21
~>o 7r
1 00
-00
tr {F(~ + jw)* F(~ + jw)}dw < 00 . (3.107)
1i2(C, c mxn ) is a Hilbert space under the scalar product (3.105) and with
i:
induced norm:
- .coo(jR, c mxn ) is the space of all functions F(jw), defined on the imagi-
nary axis, which take values in c mxn and are bounded on the imaginary
axis:
sup IIF(jw)112 = sup a(F(jw)) < 00 (3.109)
wER wER
- 1ioo (C, c mxn ) is the space of all functions F(s) which are analytic in
~e(s) > 0 (no RHP poles), take values in c mxn and are bounded in the
RHP:
sup IIF(s)lb = sup a(F(jw)) < 00 (3.111)
iRe(s»O wER
1ioo (C, c mxn ) is a Banach space under the transfer function 00 norm:
IIF(s)ll1ioo = sup a(F(jw)) . (3.112)
wER
3.6 Summary
Sets, norms and spaces frequently used in robust control theory were dis-
cussed. It was illustrated how the familiar vector p-norms induce correspond-
ing matrix norms. Of particular interest are the matrix 1, 2 and 00 norms
partly because they have obvious physical interpretations and partly because
they can be effectively computed. In "classical" robust control theory the
main emphasis has been put on the matrix 2-norm. However, recently also
the 1- and oo-norm have attracted considerable attention. It is well known
that the matrix 2-norm equals the maximum singular value. In Section 3.2.1
a thorough introduction to singular values and the singular value decomposi-
tion were given. The singular value concept has played a key role in modern
control theory for a number of reasons, in particular because it enables us to
extend the classical Bode plot to multivariable systems. A well known prob-
lem with multivariable systems is that the "gain" of a transfer function matrix
is not unique: it depends on the direction of the input vector. The maximum
(minimum) singular value measures the maximum (minimum) possible gain
of a transfer matrix frequency response G(jw) in terms of the 2-norm of the
input vector before and after multiplication by G(jw). Thus the gain of G(jw)
is bounded by its maximum and minimum singular value as the input vector
varies over all possible directions. By plotting the singular values a(G(jw»
and Q.(G(jw» for each frequency w we obtain a singular value Bode plot; the
multivariable generalization of the classical magnitude Bode plot.
In Section 3.3 we considered operator norms and induced operator norms.
In particular, we showed that if the input and output of a stable multi variable
system are measured by the operator 2-norm 11·112,2, then the induced norm
is the transfer function 11.00 norm.
Next, in Section 3.4 the scalar product < .,. > was used to define pre-
Hilbert spaces. Completeness of normed spaces was discussed. A Banach
space is a complete normed space and a Hilbert space is a complete pre-
Hilbert space.
In Section 3.5 it was discussed how the concept of Banach and Hilbert
spaces cover not only spaces of constant vectors and matrices like c mxn , but
also spaces whose elements are time domain or frequency domain functions.
Two classes of Banach and Hilbert spaces are of particular interest in modern
control theory. In fact, they have been given independent names, viz Lebesgue
32 3. Spaces and Norms in Robust Control Theory
and Hardy spaces. In the famous 1ioe and 1i2 control theory, 1i stands for
Hardy.
Using Lebesgue and Hardy spaces we can neatly classify some time domain
and frequency domain functions. In Section'3.5 several different Lebesgue and
Hardy spaces are introduced. For example, .coe(jR, c mxn ) is the Lebesgue
space of all multivariable frequency responses F(jw) for which the singular
value upper bound across frequency exists suPwa(F(jw)) < 00. The related
Hardy space 1i oe (C, c mxn ) is the space of all multivariable stable transfer
function matrices F(s) for which the same condition suPwa(F(jw)) < 00
holds. A central result in 1ioe control theory is that if G(s) E 1i oe (C, c mxn )
and u(t) E 1i2(R, R n), then y(t) E 1i2(R, Rm) and the induced operator
2-norm ofu(t) on y(t) equals the 1ioe norm IIG(s)lltioo'