1

Properties of the Singular Value Decomposition

A good reference on numerical linear algebra is

G. H. Golub and C. F. van Loan, Matrix Computations, The Johns

Hopkins University Press, 1983.

Preliminary definitions:

Hermitian: Consider x ∈Cn . Then we define the vector " x Hermitian"

by x H := x T . That is, x H is the complex conjugate transpose of x .
Similarly, for a matrix A ∈Cm × n , we define A H ∈Cn× m by A T . We say
that A ∈Cn× n is a Hermitian matrix if A = A H .

Euclidean inner product: Given x, y ∈Cn . Let the elements1 of x and y

 x1   y1 
x  y 
be denoted x =   and y =   .
2 2

 M  M
x  y 
 n  n
Then the Euclidean inner product is defined as

x, y := x H y
= x1 y1 + x2 y2 +K +xn yn

Euclidean vector norm: Let " •,• " denote the Euclidean inner
product. Then the vector norm associated with this inner product is
given by

x 2 := x, x
n .
= ∑ x1
2

i =1

We often omit the "2" subscript when we are discussing the

Euclidean norm (or "2-norm") exclusively.

1 Sometimes we use subscripts to denote the elements of a vector, and

sometimes to denote different members of a set of vectors. The meaning will be
clear from context.
 2

Euclidean matrix norm: Given A ∈Cm × n . Then the matrix norm induced
by the Euclidean vector norm is given by:

Av 2
A 2 := max v ≠ 0
v2

= λ max ( A H A)

where λ max ( A H A) denotes the largest eigenvalue of the matrix A H A . (It

is a fact that all the eigenvalues of a matrix having the form A H A are
all real and nonnegative.)

Orthogonality: Two vectors x, y ∈Cn are orthogonal if x, y = 0 .

Orthonormal Set: A collection of vectors {x1, x2 ,K , xm } ∈Cn is said to be

an orthonormal set if

0, i ≠ j
xi , x j = 
1, i = j

(Hence xi = 1, ∀i .)

Orthogonal Complement: Consider a subspace X ⊆ Cn . The orthogonal

complement of X , denoted X ⊥ , is defined as

X ⊥ := {x ∈Cn : x, y = 0 ∀ y ∈X} .

That is, every vector in X ⊥ is orthogonal to every vector in X .

Unitary Matrix: A matrix U ∈Cn× n is unitary if U HU = UU H = In .

F a c t : If U is a unitary matrix, then the columns of U form an

orthonormal basis (ONB) for Cn .

Proof of Fact: Denote the columns of U as U = [u1 u2 L un ] . Then

 3

u1H 
 H
u
U HU =  2 [u1 u2 L un ]
 M
 H
un 
u1H u1 u1H u2 L u1H  1 0 L 0
 H 
u2 u1 u2H u2 L u2H  0 1 L 0
= = 
 M M O M  M M O M
 H  
H
un u1 un u2 L unH  0 0 L 1 

Singular Value Decomposition:

Consider M ∈Cm × n . Then there exist unitary matrices

U = [u1 u2 K um ]
V = [v1 v2 K vn ]

such that

 Σ  H
 U  0 V , m ≥ n

A=  

U[ Σ 0]V H , m ≤ n

where

σ 1 0 L 0
0 σ L 0
Σ= , p = min( m,n)
2

M M O M
0 0 σ p 
 L

and
σ1 ≥ σ 2 ≥K ≥ σ p ≥ 0 .

Terminology: We refer to σ i as the i 'th singular value, to ui as the i ' t h

left singular vector, and to vi as the i 'th right singular vector.
 4

Properties of Singular Values and Vectors:

(1) Each singular value and associated singular vectors satisfy

Avi = σ i ui , i = 1,K , p

(2) The largest singular value, σ max := σ1 , satisfies

Av 2
σ max := max v ≠ 0
v2
= A2

(3) The smallest singular value, σ min := σ p , satisfies

Av 2
σ min := min v ≠ 0
v2

If A is square and invertible, then A−1 2 = 1σ .

min

(4) Suppose that σ1 ≥ σ 2 ≥K ≥ σ r > σ r +1 =K = σ p = 0 . Then

rank( A) = r

(5) Suppose that rank( A) = r . Then

R( A) = span{u1 ,u2 ,K ,ur } ,

where R( A) denotes the range, or columnspace, of A .

(6) Suppose that rank( A) = r . Then

N( A) = span{vr +1 ,vr + 2 ,K ,vn } ,

where N( A) denotes the (right) nullspace of A .

(7) Suppose that rank( A) = r . Then

R ⊥ ( A) = span{ur +1 ,ur + 2 ,K ,um } ,

where R ⊥ ( A) denotes the orthogonal complement of R( A) .

 5

(8) Suppose that rank( A) = r . Then

N ⊥ ( A) = span{v1 ,v2 ,K ,vr } ,

where N ⊥ ( A) denotes the orthogonal complement of N( A) .

(9) Suppose that rank( A) = r . Then

r
A = ∑ σ i ui viH
i =1

= Ur Σ r V rH

where

Ur := [u1 u2 K ur ] , V r := [v1 v2 K vr ] , Σ r := diag(σ1 σ 2 K σr ) .

(10) If A ∈Cn× n is invertible, then

n
A−1 = ∑ σ1i vi uiH
i =1

Definition: The pseudoinverse of A , denoted A # , is given by

A # = V r Σ r−1Ur .

The pseudoinverse has the following properties:

• If m = n and rank( A) = n , then A # = A−1 .

• If m > n , and rank( A) = n , then A is left invertible, and A # = A-L .
• If m < n , and rank( A) = m , then A is right invertible, and A # = A− R .
• Note that A # is well defined even if rank( A) < min( m,n) .

Many results for linear systems have the form "if such and such a
matrix has full rank, then such and such a property holds"
(paraphrased from Golub and Van Loan). Such results are n a i v e , in
that they neglect the the fact that a matrix generated from physical
data is almost always full rank. The more important question is n o t
"does the matrix have full rank?", but rather "how close is the matrix
 6

to one which does not have full rank?". The SVD is a very useful tool
in making this concept precise.

Consider A ∈Cm × n , and that A has full rank:

rank( A) = p:= min(m,n) .

Suppose that we perturb the elements of A , yielding  = A + ∆A . We

wish to know how large the error ∆A can become before  loses
rank.

Proposition:

(a) Suppose that ∆A 2 < σ min ( A) . Then

rank( A + ∆A) = p .

(b) There exists a matrix ∆A , with ∆A 2 = σ min ( A) such that

rank( A + ∆A) < p .

Proof:

(a) Using the triangle inequality (and dropping the subscript):

σ min ( A + ∆A):= min v =1 ( A + ∆A)v

≥ min v =1 { Av − ∆Av }
≥ min v =1 Av − max v =1 ∆Av
≥ σ min (A) − σ max (∆A)
>0
⇒ rank( A + ∆A) = p

p
(b) Let A = ∑ σ iui viH , and consider ∆A = − σ pu p v pH (where σ p = σ min ). It is
i =1
p−1
easy to see that A + ∆A = ∑ σ iui viH , and thus that rank( A + ∆A) = p − 1 .
i =1
 7

Suppose that rank( A) = p , but A has very small singular values. Then
A is "close" to a singular matrix in the sense that there exists a small
perturbation ∆A to the elements of A that causes  to lose rank.
Indeed, such a matrix should possibly be treated in applications as
though it were singular.

In practice, people do not look always look for small singular values
to determine distance to singularity. It is often more useful to
compute the r a t i o between the maximum and minimum singular
values.

Definition (Condition Number): Consider A ∈Cm × n and suppose that

rank( A) = p = min(m,n) . Then the (Euclidean) condition number of A is
defined as
σ max ( A)
κ ( A):= .
σ min ( A)
###

Suppose rank( A) = p , but that κ ( A) >> 1. It follows that A is "almost

rank deficient". In fact, the process of calculating an inverse 2 for A
may not be numerically robust, and calculations involving A−1 may be
prone to error. Let's explore this idea.

Consider the linear system of equations

Cn , m ≥ n, rank( A) = n .
Cm×n , b ∈C
Ax = b, A ∈C

Suppose that we are given the data for A and b and need to solve for
x . Let R( A) denote the range of A . If b ∈R
R( A) , then we may find x
from
x = A# b .

In reality, the elements of A and b will be corrupted by errors.

These may arise due to uncertainty in the methods used to generate
the data. Errors also arise due to numerical roundoff in the computer
representation of the data. We would like to know how these errors
affect the accuracy of the solution vector x .

2 If A is not square, then we calculate A # which will be equal to the left or

right inverse, whichever is appropriate.
 8

Let
Â:= A + ∆A
b̂:= b + ∆b
x̂:= x + ∆x

so that Âx̂ = b̂ . Our next result relates the errors ∆A and ∆b to errors
in the computed value of x .

Cm×n has rank n , and that

Proposition: Consider A ∈C

∆A
≤ α < 1.
σ min ( A)

∆A ∆b
Suppose that ∆A and ∆b satisfy the bounds ≤ δ and ≤δ,
A b
where δ is a constant. Then

∆x 2δ
≤ κ ( A) .
x 1− α
###

The proof is obtained as a sequence of homework exercises in Golub

and van Loan.

Note that we can, in principle, calculate A−1 by successively solving

1  0  0 
0  1  0 
the above linear system for b =  ,  ,K ,  .
 M  M  M
0  0  1 
    

It follows from this result that if κ ( A) >> 1, then small relative errors
in A and b m a y result in large relative errors in the computed value
of x . One should be cautioned, however, that this error estimate is
only an upper bound. Hence, the computed answer is not g u a r a n t e e d
to be incorrect. However, there are nonpathological examples for
which the upper bound on the error is achieved! Hence, if no
additional information is available, the answer to any calculation
involving the inverse of a matrix that is ill-conditioned (i.e., κ ( A) >> 1)
should be viewed dubiously.
 9

1 100 
Example: Consider A =  . MATLAB says that rank( A) = 2 :
0 1 
»A=[1 100;0 1];
»rank(A)
ans =
2
»cond(A)
ans =
1.0002e+04

The condition number looks pretty large; however, for the purposes
of solving linear systems of two equations in two unknowns, this
matrix is not particularly ill-conditioned with respect to numerical
roundoff error in the computer computations. Suppose, on the other
hand, that A and b are constructed from physical data. How reliable
are calculations based upon A−1?

For example, are we sure that the zero in the (2,1) element of A is
really zero?

Suppose that the (2,1) element of A is perturbed:

 1 100 
 =   ,
0.009 1 
and consider the product A−1  ; if  = A , this product will equal the
identity matrix. Let us see what the error is for our example:

»inv(A)*Ahat
ans =
0.1 0.0
0.009 1.0

We see that a small error in one of the elements of A results in a

large error in the product A−1 Â !

On the other hand, consider an error of the same magnitude in the

(1,1) element:

1.009 100 
 = 
 0 1 
 10

»inv(A)*Ahat
ans =
1.009 0
0 1.0

In this case, the error in the product A−1 Â is also small.

What to do with an "almost rank deficient" matrix:

Suppose we conclude that κ ( A) is so large that A should be viewed as

rank deficient. How do we find an approximation to A ? An obvious
approach is to attempt to approximate A by a matrix of lower rank.

Approach 1: Motivated by the discussion above, let's use the

information contained in the SVD and try to find a matrix of l o w e r
rank that is "close" to A .

We shall do this for the case m ≥ n , so that

Σ 
A = U  V H ,
0

where

U = [u1 K um ] ∈C
Cm×m , V = [v1 K vn ] ∈C
Cn×n , and Σ = diag(σ1 σ 2 K σ n ).

Suppose that

σ1 ≥K ≥ σ r > ε > σ r+1 ≥K σ n ≥ 0 ,

where ε is so small (or so much smaller than σ r ) that the last (n-r+1)
singular values are all effectively zero. Then we say that A has
effective rank equal to r .

Furthermore, let Reff ( A) and N eff ( A) denote the effective range and
effective nullspace of A , respectively. Then we can calculate bases
for these subspaces by choosing appropriate singular vectors:

Reff ( A):= span{u1, K , ur } and N eff ( A):= span{vr+1, K , vn } .

 11

Similarly,

⊥
Reff ( A):= span{ur+1, K , um } and N eff
⊥
( A):= span{v1, K , vr } .

That is, the effective range of A is spanned by the left singular

vectors corresponding to the nonnegligible singular values of A , the
effective nullspace is spanned by the right singular vectors
corresponding to the negligible left singular values, and so forth.

Define
r
Aeff = ∑ σ iui viH
i=1
= Ur Σ r VrH

where

Ur := [u1 u2 K ur ] , V r := [v1 v2 K vr ] , Σ r := diag(σ1 σ 2 K σr ) .

If have chosen wisely, then

σ
( )
κ Aeff = 1
σr
will not be "too large". Note that

b ∈R ( ) Reff ( A) .
R Aeff ⇔ b ∈R

Hence, if b ∈R ( )
R Aeff , then the equation Ax = b has a solution that is
#
robust against small errors in A . This solution is given by x = Aeff b.

Example(continued):

1 100 
Once again, consider A =   . Let's look at the SVD of A , and use
0 1 
the information from the singular values and vectors to construct a
rank 1 matrix, Aeff , that is "close" to A .

»[U,S,V] = svd(A)

U=
 12

1.0 -0.01
0.01 1.0

S=
100.01 0
0 0.01

V=
0.01 -1.0
1.0 0.01

Let's calculate basis vectors for Reff and N eff :

»Reff = U(:,1)
Reff =
1.0000
0.0100

The value of Reff makes sense, because both columns of A have much
larger entries in the first row than in the second.

»Neff = V(:,2)
Neff =
-1.0000
0.0100

Check to see that Neff "almost" gets multiplied by zero, and thus is in
the "effective" nullspace of A :

»A*Neff
ans =
-0.0001
0.01

Now, let's construct a rank 1 matrix that approximates A by deleting

the small singular value and associated singular vectors:

Aeff = σ1u1v1H

( ) ( )
Note that R Aeff = Reff ( A) and N Aeff = N eff ( A) .
 13

»Aeff = Reff*S(1,1)*V(:,1)'
Aeff =
0.9999 100.0000
0.0100 0.9999

»rank(Aeff)
ans =
1

( )
We see that, as expected, rank Aeff = 1 .

Approach 2: Suppose we try to apply this idea to a matrix obtained

from physical data. Unfortunately, it is sometimes difficult to relate
the singular values and vectors to physical quantities. An alternate
approach which may avoid this difficulty is to merely delete columns
of the matrix. For example, suppose that we merely zeroed the first
column of A :

0 100 
A1 =  
0 1 

It is easy to see that R( A1 ) ≅ Reff ( A) and N ( A1 ) ≅ N eff ( A) .

Shortly, we shall see how to apply these concepts to evaluate the

rank of the DC gain matrix of a plant we wish to control. Before doing
this, we need to review one more topic from numerical linear
algebra. This topic is one which can cause a great deal of headaches
in control applications.

Scaling and Choice of Units

There is a difficulty associated with using σ min ( A) and/or κ ( A) as

measures of distance to singularity. Namely, the size of these
quantities varies with scaling. For example, consider the equation

Cq , u ∈C
y = Au, y ∈C Cp,

where the elements of the vectors represent physical quantities.

Suppose that we change the units used to measure the elements of u

and y :
 14

ynew := D1−1y, unew := D2−1u

where D1 and D2 are diagonal matrices whose diagonal elements are

all real and positive. (The use of inverses is to conform with the
notation in Golub and Van Loan; it also suggests the common practice
of scaling signals by their nominal values.)

It is easy to show that ynew and unew satisfy

ynew = Anewunew , where Anew := D1−1 AD2 .

Hence changing units for the physical quantities related by A

corrsponds to scaling A by diagonal matrices. It is easy to show that
the singular values and condition number of Anew may be very
different than those of the original matrix. Indeed, let us consider
our example

1 100 
A= ,
0 1 
d1 0 
for which σ min (A) = 0.01 and κ (A) = 10,000 . If we choose D1 =   and
 0 1
1 0 
D2 =   , then
0 d2 

1 100 d2 
D1−1 AD2 = d1  .
0 1 

It follows that, by changing units, we can make the condition number

and singular values of Anew arbitrarily close to one. This means that
the matrix is no longer "almost singular"? Does it also mean that the
physical property that depended upon A being full rank is now
present r o b u s t l y ? Maybe and maybe not. It all depends if the units
of Anew are physically reasonable. The latter condition is a qualitative
one that depends upon good engineering sense. For example, if one
were studying deposition rates of a metal oxide film on a silicon
substrate, measuring speed in terms of microns/minute might be
very appropriate. If, in order to obtain a nice " A " matrix, we had to
change these units to furlongs/fortnight, then it would be very
difficult to interpret our answer.
 15

Unfortunately, finding a physically reaonable set of units with which

to work is sometimes problematic. In a MIMO feedback system, we
may be comparing very different types of signals, for example,
voltages, pressures, speeds, temperatures. Our choice of units will
dictate the conditioning of various matrices, such as the DC gain
matrix. To draw conclusions from the condition number, we must
nevertheless have units that make comparisons between different
physical quantities meaningful.
 16

Systems Interpretations
Consider the system

ẋ = Ax + Bu, Rn , u ∈R
x ∈R Rp
y = Cx , Rq
y ∈R
P( s ) = C( sI − A)−1 B

We will assume the system is controllable and observable and (for

simplicity) that p = q , and that A has stable eigenvalues. Consider the
problem of achieving zero steady state error to a step command. We
have seen that this problem is solvable precisely when the DC gain
matrix satisfies rank( P(0)) = q . What happens when the DC gain matrix
is a l m o s t rank deficient? Intuitively, this means that the command
tracking problem is "almost" unsolvable.

We shall suppose that in physically relevant units the DC gain matrix

satisfies:
(i) σ max ( P(0)) > 1
and
(i) σ min ( P(0)) << 1.

Together, (i) and (ii) imply

(iii) κ ( P(0)) >> 1.

1 10 
(For example, consider the matrix A =  , which has σ max = 100.01 ,
0 1 
σ min = 0.01, and κ ≅ 10000 .)

It follows from (i) that there does exist some control authority at DC
(i.e., the DC gain matrix is nonzero). However, (ii) implies that some
control inputs are relatively ineffective. Ineffective control inputs
can arise in two ways.

• If a column of P(0) has small magnitude, then the particular input

associated with that column has relatively little DC authority.
• If two or more columns of P(0) are almost linearly dependent,
then the control inputs associated with those columns are almost
redundant.
 17

Finally, it follows from (iii) that any calculations involving P(0)−1 are
sensitive to errors in the data for P(0) .

Let's now consider the implications of (i)-(iii) for the constant

command tracking problem. Recall that this control problem is
solvable only if P(0) is nonsingular. We developed two solutions to
this problem. First, we used state feedback and a constant gain
[ ]
−1
precompensator, N = C( − A + BK )−1 B :

r v u x y
N Σ B ( s I- A ) - 1 C

Second, we augmented the system with integrators, and fed back

both the plant and integrator states:

u x y
Σ Σ B Σ ( sI- A ) - 1 C Σ
-
-
r
Σ
K -

w e
KI I/ s
 18

In each case, closed loop stability implies that the steady state
response of the system output to a step command r(t) = r01(t) satisfies
yss = r0 . Note also that, in each case, yss = P(0)uss . It follows that the
steady state control signal must satisfy:

uss = P(0)−1 r0

( )
Since σ max P(0)−1 = 1
σ min ( P(0))
, it follows from (ii) that relatively
large control signals will be required to track certain step commands.
Large control signals are undesirable because they may saturate
control actuators or cause other undesirable nonlinear behavior.

Furthermore, (iii) tells us that small changes in the data of P(0) may
cause large changes in P(0)−1. It follows that small errors in P(0) may
cause the size of the control signal generated to force command
tracking to be much larger than that indicated from the nominal
value of P(0) . Hence, even if the control signals obtained by
examining the nominal model of P(0) are reasonable, the control
signals obtained from the true plant may not be.

It follows from our discussion of effective range and nullspace that if

r0 ∈RReff ( A), then the command r(t) = r01(t) can be tracked without
generating excessively large control signals.

Let's see an application of these ideas to data from a real system.

First, we shall describe the system, and then analyze its DC gain
matrix.
 19

Example: Reactive Ion Etching of Silicon Wafers

Current practice in the semiconductor and flat panel display

manufacturing industries uses little real-time feedback control. We
are engaged in a project that will incorporate real-time feedback
control into the process of Reactive Ion Etching (RIE).

The ultimate design goal is to reliably and reproducibly etch a

pattern on a wafer; this is one of many steps needed to produce a
semiconductor chip...

mask

SiO2

Si

mask

SiO2

Si

SiO2

Si

There are a number of characteristics of the etch we would like to

control. Among these are:

1. Sidewall angle: Are the sides of the etched feature vertical or

sloped?

2. Selectivity: What are the relative etch rates of SiO2 , Si , and the
photoresist mask?

3. Etch rate: Is this constant over the course of a single etch, and
from etch to etch for many etches?

4. Uniformity: Are etch properties consistent across the wafer?

 20

Wafer etching takes place inside a low pressure plasma chamber; a

rough diagram is shown below:

Gas
In let

Plasm a

Throt t le
Valve

RF Mat ching T urb o

Net work Pump

We have the following three actuators that we can use for control:

RF Power, Throttle angle, and CF4 Gas Flow Rate.

We also have sensors to measure three signals that we can feed back
and attempt to regulate:

Vbias , [ F ] , Pressure.

These actuators and sensors are related to the plasma, not the wafer.
Currently, we cannot sense features on the wafer surface in real
time. Hence, our control strategy is to use the control inputs to
regulate the plasma properties, which we c a n measure. The plasma
properties are only indirectly related to the wafer etch; however,
they do determine the environment in which the etch takes place. I t
 21

is our hypothesis that by better regulating the plasma environment,

we will achieve a higher yield of usable etched wafers.

We now describe the control inputs and sensed outputs in more

detail.

Inputs:

The flow input determines the rate at which CF4 gas enters the etch
chamber.

The RF power input has two effects: (i) it disassociates CF4 → CF3+ + F
(a charged ion and reactive fluorine radical whose density we denote
by [ F ]), and (ii) it sets up a bias voltage, Vbias , across the plasma. This
bias voltage accelerates the ions so that they bombard the wafer
surface. The physical energy thus imparted to the surface, combined
with the chemical reactions between [ F ] and Si are responsible for
etching the exposed portion of the wafer surface.

The throttle input determines the rate at which gases are exhausted
from the chamber.

Outputs:

The Vbias output is responsible for the physical component of etching.

The [ F ] output is responsible for the chemical component of etching.

The Pressure output determines (among other things) the mean free
path between collisions in the chamber. The longer the mean free
path, the more energy the ions have when they impact the wafer
surface, and the greater the physical component of the etching
process.

Modelling:

Using small signal step response data together with black box system
identification techniques, we obtained a linear model of the system
at the nominal operating condition:

Throttle: 10% open

CF4 Flow: 15 sccm
 22

RF power: 800 watts

Vbias : 400 volts

Pressure: 10 millitorr
[ F ]: 5

The DC gain matrix of the resulting transfer function model is:

13.2 −6.77 0.37 

P(0) = −1.0 0.973 0 
 
0.123 0.0682 0.011

To aid in analyzing this matrix, we normalize each input and output

by its nominal value. This yields a normalized DC gain matrix:

0.0025 0 0  13.2 −6.77 0.37  10.0 0 0 


PN (0) = 0 0.1 0   −1.0 0.973 0  0 15.0 0 
   
0 0 0.02  0.123 0.0682 0.011 0 0 800 

0.3300 −0.2539 0.7400 

= −1.0000 1.4595 0.0 
 
0.2460 0.2046 1.7600 

0.3 −0.3 0.7 

≅ −1.0 1.5 0.0 
 
0.2 0.2 1.8 

Note that both CF4 flow and Throttle angle each primarily affect
pressure. Specifically, if we open the exhaust throttle, then the
steady state pressure will decrease. On the other hand, if we increase
the rate of CF4 flow, then the steady state pressure will i n c r e a s e .
These effects are plausible, because flow affects the rate at which
gases enter the chamber, and throttle affects the rate at which gases
leave the chamber. Power has no steady state effect on pressure, but
it does affect both Vbias and [ F ] relatively more strongly than do flow
and throttle.

»[U,S,V] = svd(P_N)
 23

U= % left singular vectors

0.4302 -0.0407 -0.9018
-0.3499 0.9134 -0.2082
0.8321 0.4051 0.3787

S= % singular values
1.9663 0 0
0 1.7846 0
0 0 0.0046

V= % right singular vectors

0.3543 -0.4635 0.8122
-0.2287 0.7992 0.5558
0.9067 0.3827 -0.1771

»kappa = cond(P_N)

kappa =
425.0254

The large condition number is consistent with the fact that the
throttle and flow inputs are almost redundant, and thus do not yield
two independent degrees of control authority. As we have noted,
both these inputs primarily affect pressure.

As a consequence of the poorly conditioned DC gain matrix, we

cannot use our three actuators to independently control all three
outputs unless we can tolerate very large control signals without
saturating the actuators. As it happens, both the throttle and the
mass flowmeter regulating CF4 flow into the chamber are quite
nonlinear, and operate fairly close to saturation. These actuators
certainly cannot tolerate large control signals.

An approach to the problem of two redundant inputs is to keep one

of them fixed at its nominal value, and use only the other for the
purpose of feedback control. Engineering insight must be used to
decide which input to keep, and which to ignore. In this case, we
keep flow fixed because the mass flowmeter saturates easily. (In
particular, the flow variable cannot be negative -- CF4 cannot be
sucked out of the chamber using the flowmeter!)
 24

With flow removed as an actuator, the DC gain matrix becomes:

»DCnew1 = DCnew(:,[1 3])

DCnew1 =
0.3300 0.7400
-1.0000 0
0.2460 1.7600

We now have a control problem with two inputs and three outputs.
Becasue we cannot independently control all three outputs, we must
choose two of them. (More generally, we can choose any two
independent linear combinations of the three outputs.) The physics
of the etch process suggests that Vbias and [ F ] are relatively more
important than Pressure (although this is a debatable point). Deleting
Pressure from the DC gain matrix yields

»DCnew2 = DCnew1([1 3],:)

DCnew2 =
0.3300 0.7400
0.2460 1.7600

»svd(DCnew2)
ans =
1.9423
0.2053

The condition number is now ≅ 10, and the control problem is

reasonably well-conditioned.