Optimization I; Chapter 3 56

Chapter 3 Quadratic Programming

3.1 Constrained quadratic programming problems
A special case of the NLP arises when the objective functional f is quadratic
and the constraints h, g are linear in x lR
n
. Such an NLP is called a Quadratic
Programming (QP) problem. Its general form is
minimize f(x) :=
1
2
x
T
Bx x
T
b (3.1a)
over x lR
n
subject to A
1
x = c (3.1b)
A
2
x d , (3.1c)
where B lR
nn
is symmetric, A
1
lR
mn
, A
2
lR
pn
, and b lR
n
, c
lR
m
, d lR
p
.
As we shall see in this chapter, the QP (3.1a)-(3.1c) can be solved iteratively
by active set strategies or interior point methods where each iteration requires
the solution of an equality constrained QP problem.
3.2 Equality constrained quadratic programming
If only equality constraints are imposed, the QP (3.1a)-(3.1c) reduces to
minimize f(x) :=
1
2
x
T
Bx x
T
b (3.2a)
over x lR
n
subject to Ax = c , (3.2b)
where A lR
mn
, m n. For the time being we assume that A has full row
rank m.
The KKT conditions for the solution x

lR
n
of the QP (3.2a),(3.2b) give rise
to the following linear system
_
B A
T
A 0
_
. .
=: K
_
x

_
=
_
b
c
_
, (3.3)
where

lR
m
is the associated Lagrange multiplier.
We denote by Z lR
n(nm)
the matrix whose columns span KerA, i.e., AZ = 0.
Optimization I; Chapter 3 57
Denition 3.1 KKT matrix and reduced Hessian
The matrix K in (3.3) is called the KKT matrix and the matrix Z
T
BZ is
referred to as the reduced Hessian.
Lemma 3.2 Existence and uniqueness
Assume that A lR
mn
has full row rank m n and that the reduced Hessian
Z
T
BZ is positive denite. Then, the KKT matrix K is nonsingular. Hence,
the KKT system (3.3) has a unique solution (x

).
Proof: The proof is left as an exercise.

Under the conditions of the previous lemma, it follows that the second order
sucient optimality conditions are satised so that x

is a strict local minimizer

of the QP (3.2a),(3.2b). A direct argument shows that x

is in fact a global
minimizer.
Theorem 3.3 Global minimizer
Let the assumptions of Lemma 3.2 be satised and let (x

) be the unique
solution of the KKT system (3.3). Then, x

is the unique global solution of the

QP (3.2a),(3.2b).
Proof: Let x T be a feasible point, i.e., Ax = c, and p := x

x. Then,
Ap = 0. Substituting x = x

p into the objective functional, we get

f(x) =
1
2
(x

p)
T
B(x

p) (x

p)
T
b =
=
1
2
p
T
Bp p
T
Bx

+ p
T
b + f(x

) .
Now, (3.3) implies Bx

= b A
T

. Observing Ap = 0, we have
p
T
Bx

= p
T
(b A
T

) = p
T
b (Ap)
T

. .
= 0
,
whence
f(x) =
1
2
p
T
Bp + f(x

) .
In view of p Ker A, we can write p = Zu , u lR
nm
, and hence,
f(x) =
1
2
u
T
Z
T
BZu + f(x

) .
Since Z
T
BZ is positive denite, we deduce f(x) > f(x

). Consequently, x

is
the unique global minimizer of the QP (3.2a),(3.2b).

Optimization I; Chapter 3 58
3.3 Direct solution of the KKT system
As far as the direct solution of the KKT system (3.3) is concerned, we distinguish
between symmetric factorization and the range-space and null-space approach.
3.3.1 Symmetric indenite factorization
A possible way to solve the KKT system (3.3) is to provide a symmetric fac-
torization of the KKT matrix according to
P
T
KP = LDL
T
, (3.4)
where P is an appropriately chosen permutation matrix, L is lower triangular
with diag(L) = I, and D is block diagonal.
Based on (3.4), the KKT system (3.3) is solved as follows:
solve Ly = P
T
_
b
c
_
, (3.5a)
solve D y = y , (3.5b)
solve L
T
y = y , (3.5c)
set
_
x

_
= P y . (3.5d)
3.3.2 Range-space approach
The range-space approach applies, if B lR
nn
is symmetric positive denite.
Block Gauss elimination of the primal variable x

leads to the Schur complement

system
AB
1
A
T

= AB
1
b c (3.6)
with the Schur complement S lR
mm
given by S := AB
1
A
T
. The range-
space approach is particularly eective, if
B is well conditioned and easily invertible (e.g., B is diagonal or block-
diagonal),
B
1
is known explicitly (e.g., by means of a quasi-Newton updating for-
mula),
the number m of equality constraints is small.
Optimization I; Chapter 3 59
3.3.3 Null-space approach
The null-space approach does not require regularity of B and thus has a wider
range of applicability than the range-space approach.
We assume that A lR
mn
has full row rank m and that Z
T
BZ is positive
denite, where Z lR
n(nm)
is the matrix whose columns span Ker A which
can be computed by QR factorization (cf. Chapter 2.4).
We partition the vector x

according to
x

= Y w
Y
+ Zw
Z
, (3.7)
where Y lR
nm
is such that [Y Z] lR
nn
is nonsingular and w
Y
lR
m
, w
Z

lR
nm
.
Substituting (3.7) into the second equation of (3.3), we obtain
Ax

= AY w
Y
+ AZ
..
= 0
w
Z
= c , (3.8)
i.e., Y w
Y
is a particular solution of Ax = c.
Since A lR
mn
has rank m and [Y Z] lR
nn
is nonsingular, the product
matrix A[Y Z] = [AY 0] lR
mm
is nonsingular. Hence, w
Y
is well determined
by (3.8).
On the other hand, substituting (3.7) into the rst equation of (3.3), we get
BY w
Y
+ BZw
Z
+ A
T

= b .
Multiplying by Z
T
and observing Z
T
A
T
= (AZ)
T
= 0 yields
Z
T
BZw
Z
= Z
T
b Z
T
BY w
Y
. (3.9)
The reduced KKT system (3.9) can be solved by a Cholesky factorization of the
reduced Hessian Z
T
BZ lR
(nm)(nm)
. Once w
Y
and w
Z
have been computed
as the solutions of (3.8) and (3.9), x

is obtained according to (3.7).

Finally, the Lagrange multiplier turns out to be the solution of the linear system
arising from the multiplication of the rst equation in (3.7) by Y
T
:
(AY )
T

= Y
T
b Y
T
Bx

. (3.10)
3.4 Iterative solution of the KKT system
If the direct solution of the KKT system (3.3) is computationally too costly,
the alternative is to use an iterative method. An iterative solver can be ap-
plied either to the entire KKT system or, as in the range-space and null-space
approach, use the special structure of the KKT matrix.
Optimization I; Chapter 3 60
3.4.1 Krylov methods
The KKT matrix K lR
(n+m)(n+m)
is indenite. In fact, if A has full row rank
m, K has n positive and m negative eigenvalues. Therefore, for the iterative
solution of (3.3) Krylov subspace methods like GMRES (Generalized Minimum
RESidual) and QMR (Quasi Minimum Residual) are appropriate candidates.
3.4.2 Transforming range-space iterations
We assume B lR
nn
to be symmetric positive denite and suppose that

B is
some symmetric positive denite and easily invertible approximation of B such
that

B
1
B I.
We choose K
L
lR
(n+m)(n+m)
as the lower triangular block matrix
K
L
=
_
I 0
A

B
1
I
_
, (3.11)
which gives rise to the regular splitting
K
L
K =
_

B A
T
0

S
_
. .
=: M
1

_

B(I

B
1
B) 0
A(I

B
1
B) 0
_
. .
=: M
2
0
, (3.12)
where

S lR
mm
is given by

S := A

B
1
A
T
. (3.13)
We set
:= (x, )
T
, := (b, c)
T
.
Given an iterate
(0)
lR
n+m
, we compute
(k)
, k lN, by means of the
transforming range-space iterations

(k+1)
=
(k)
+ M
1
1
K
L
( K
(k)
) = (3.14)
= (I M
1
1
K
L
K)
(k)
+ M
1
1
K
L
, k 0 .
The transforming range-space iteration (3.14) will be implemented as follows:
d
(k)
= (d
(k)
1
, d
(k)
2
)
T
:= K
(k)
, (3.15a)
K
L
d
(k)
= (d
(k)
1
, A

B
1
d
(k)
1
+ d
(k)
2
)
T
, (3.15b)
M
1

(k)
= K
L
d
(k)
, (3.15c)

(k+1)
=
(k)
+
(k)
. (3.15d)
Optimization I; Chapter 3 61
3.4.3 Transforming null-space iterations
We assume that x lR
n
and R
m
admit the decomposition
x = (x
1
, x
2
)
T
, x
1
lR
m
1
, x
2
lR
nm
1
, (3.16a)
= (
1
,
2
)
T
,
1
lR
m
1
,
2
lR
mm
1
, (3.16b)
and that A lR
mn
and B lR
nn
can be partitioned by means of
A =
_
A
11
A
12
A
21
A
22
_
, B =
_
B
11
B
12
B
21
B
22
_
, (3.17)
where A
11
, B
11
lR
m
1
m
1
with nonsingular A
11
.
Partitioning the right-hand side in (3.3) accordingly, the KKT system takes the
form
_
_
_
_
_
_
B
11
B
12
[ A
T
11
A
T
21
B
21
B
22
[ A
T
12
A
T
22
[
A
11
A
12
[ 0 0
A
21
A
22
[ 0 0
_
_
_
_
_
_
_
_
_
_
_
_
x

1
x

2
_
_
_
_
_
_
=
_
_
_
_
_
_
b
1
b
2

c
1
c
2
_
_
_
_
_
_
. (3.18)
We rearrange (3.18) by exchanging the second and third rows and columns
_
_
_
_
_
_
B
11
A
T
11
[ B
12
A
T
21
A
11
0 [ A
12
0
[
B
21
A
T
12
[ B
22
A
T
22
A
21
0 [ A
22
0
_
_
_
_
_
_
_
_
_
_
_
_
x

2
_
_
_
_
_
_
=
_
_
_
_
_
_
b
1
c
1

b
2
c
2
_
_
_
_
_
_
. (3.19)
Observing B
12
= B
T
21
, in block form (3.19) can be written as follows
_
/ B
T
B T
_
. .
=: K
_

2
_
. .
=:

=
_

1

2
_
. .
=:
, (3.20)
where

i
:= (x

i
,

i
)
T
,
i
:= (b
i
, c
i
)
T
, 1 i 2.
We note that block Gauss elimination in (3.20) leads to the Schur complement
system
_
/ B
T
0 o
__

2
_
=
_

1

2
B
1

1
_
. (3.21)
The Schur complement o is given by
o = T B/
1
B
T
, (3.22)
Optimization I; Chapter 3 62
where
/
1
=
_
0 A
1
11
A
T
11
A
T
11
B
11
A
1
11
_
. (3.23)
We assume that

A
11
lR
m
1
m
1
is an easily invertible approximation of A
11
and
dene

/ =
_
B
11

A
T
11

A
11
0
_
. (3.24)
We remark that the inverse of

/ is given as in (3.23) with A
1
11
, A
T
11
replaced
by

A
1
11
and

A
T
11
, respectively.
We introduce the right transform
K
R
=
_
I

/
1
B
T
0 I
_
, (3.25)
which gives rise to the regular splitting
KK
R
=
_

/ 0
B

o
_
. .
=: M
1

_
(I /

/
1
)

/ (I +/

/
1
)B
T
0 0
_
. .
=: M
2
0
, (3.26)
where

o := T B

/
1
B
T
. (3.27)
Given a start iterate
(0)
= (
(0)
1
,
(0)
2
)
T
, we solve the KKT system (3.20) by
the transforming null-space iterations

(k+1)
=
(k)
+ K
R
M
1
1
( K
(k)
) = (3.28)
= (I K
R
M
1
1
K)
(k)
+ K
R
M
1
1
, k 0 .
3.5 Active set strategies for convex QP problems
3.5.1 Primal active set strategies
We consider the constrained QP problem
minimize f(x) :=
1
2
x
T
Bx x
T
b (3.29a)
over x lR
n
subject to Cx = c (3.29b)
Ax d , (3.29c)
Optimization I; Chapter 3 63
where B lR
nn
is symmetric positive denite, C lR
mn
, A lR
pn
, and
b lR
n
, c lR
m
, d lR
p
.
We write the matrices A and C in the form
A =
_
_
a
1

a
p
_
_
, a
i
lR
n
, C =
_
_
c
1

c
m
_
_
, c
i
lR
n
. (3.30)
The inequality constraints (3.29c) can be equivalently stated as
a
T
i
x d
i
, 1 i m . (3.31)
The primal active set strategy is an iterative procedure:
Given a feasible iterate x
()
, 0, we determine an active set
1
ac
(x
()
) 1, ..., p (3.32)
and consider the corresponding constraints as equality constraints, whereas the
remaining inequality constraints are disregarded. Setting
p = x
()
x , b
()
= Bx
()
b , (3.33)
we nd
f(x) = f(x
()
p) =
1
2
p
T
Bp (b
()
)
T
p + g ,
where g :=
1
2
(x
()
)
T
Bx
()
b
T
x
()
.
The equality constrained QP problem to be solved at the (+1)-st iteration
step is then:
minimize
1
2
p
T
Bp (b
()
)
T
p (3.34a)
over p lR
n
subject to Cp = 0 (3.34b)
a
T
i
p = 0 , i 1
ac
(x
()
) , (3.34c)
We denote the solution of (3.34a)-(3.34c) by p
()
. The new iterate x
(+1)
is then
obtained according to
x
(+1)
= x
()

p
()
,

[0, 1] , (3.35)
where

is chosen such that x

(+1)
stays feasible.
In particular, for i 1
ac
(x
()
) we have
a
T
i
x
(+1)
= a
T
i
x
()

a
T
i
p
()
= a
T
i
x
()
d
i
.
Optimization I; Chapter 3 64
On the other hand, if a
T
i
p
()
0 for some i / 1
ac
(x
()
), it follows that
a
T
i
x
(+1)
= a
T
i
x
()

a
T
i
p
()
a
T
i
x
()
d
i
.
Finally, if a
T
i
p
()
< 0 for i / 1
ac
(x
()
), we have
a
T
i
x
(+1)
= a
T
i
x
()

a
T
i
p
()
d
i

a
T
i
x
()
d
i
a
T
i
p
()
.
Consequently, in order to guarantee feasibility, we choose

:= min (1, min

i / I
ax
(x
()
)
a
T
i
x
()
d
i
a
T
i
p
()
) . (3.36)
a
T
i
p
()
<0
We dene
1
b
(p
()
) := i / 1
ac
(x
()
) [ a
T
i
p
()
< 0 , min
i / I
ax
(x
()
)
a
T
i
x
()d
i
a
T
i
p
()
< 1 . (3.37)
Clearly,
a
T
i
x
(+1)
= a
T
i
_
x
()

p
()
_
= d
i
, i 1
b
(x
()
) .
Therefore, 1
b
(p
()
) is referred to as the set of blocking constraints.
We specify 1
ac
(x
(+1)
) by adding to 1
ac
(x
()
) the most restrictive blocking con-
straint:
1
ac
(x
(+1)
) := (3.38)
1
ac
(x
()
) j 1
b
(x
()
[
a
T
j
x
()
d
j
a
T
j
p
()
= min
i / I
ac
(x
()
)
a
T
i
x
()
d
i
a
T
i
p
()
.
a
T
i
p
()
<0
Further information with respect to a proper specication of the set of active
constraints is provided by systematically checking the KKT conditions:
Assume that p
()
= 0 is the solution of the QP problem (3.34a)-(3.34c). Since
p
()
satises the KKT conditions associated with that QP problem, there exist
Lagrange multipliers
()
lR
m
and
()
i
, i 1
ac
(x
()
), such that
b
()
= (Bx
()
b) =
m

i=1

()
i
c
i

iI
ac
(x
()
)

()
i
a
i
. (3.39)
If we set

()
i
:= 0 , i 1, ..., p 1
ac
(x
()
) ,
Optimization I; Chapter 3 65
it is clear that x
()
,
()
, and
()
satisfy the rst KKT condition with respect
to the original QP problem (3.29a)-(3.29c).
Since x
()
is feasible, the second and third KKT conditions also hold true.
We check the fourth KKT condition in terms of the sign of the multiplier
()
:
If

()
i
0 , i 1
ac
(x
()
) ,
the fourth KKT condition holds true. Consequently, x
()
is a strict local mini-
mum, since B is symmetric positive denite.
On the other hand, if there exists j 1
ac
(x
()
) such that

()
j
< 0 ,
we remove that constraint from the active set. We show that this strategy
produces a direction p in the subsequent iteration step that is feasible with
respect to the dropped constraint:
Theorem 3.4 Feasible and descent direction for active set strategy
We assume that x lR
n
satises the KKT conditions for the QP problem
(3.34a)-(3.34c), i.e., in particular (3.39) holds true. Further, assume that the
constraint gradients c
i
, 1 i m, a
i
, i 1
ac
( x) are linearly independent. Fi-
nally, suppose there is j 1
ac
( x) such that
j
< 0.
Let p is the solution of the QP problem
minimize
1
2
p
T
Bp

b
T
p (3.40a)
over p lR
n
subject to Cp = 0 (3.40b)
a
T
i
p = 0 , i 1
ac
( x) j . (3.40c)
where

b := B x b.
Then, p is a feasible direction for the constraint j, i.e.,
a
T
j
p 0 .
Moreover, if p satises the second order sucient optimality conditions for
(3.40a)-(3.40c), then a
T
j
p < 0, i.e., p is a descent direction.
Proof. Since p is a solution of the QP problem (3.40a)-(3.40c), there exist
Lagrange multipliers

i
, 1 i m, and
i
, i 1
ac
( x), i ,= j, such that
Bp

b =
m

i=1

i
c
i

iI
ac
( x),i=j

i
a
i
. (3.41)
Optimization I; Chapter 3 66
Let Z be the null space basis of the matrix
_
[c
i
]
T
1im
[a
i
]
T
iI
ac
( x),i=j

.
The second order necessary optimality conditions imply that
Z
T
BZ
is positive semidenite. Since p has the form p = Zp
Z
for some vector p
Z
, we
deduce
p
T
Bp 0 .
Since we have assumed that (3.39) holds true (with b
()
, x
()
,
()
i
,
()
i
replaced
by

b, x,

i
,
i
), subtraction of (3.39) from (3.41) yields
Bp =
m

i=1
(

i
)c
i

iI
ac
( x),i=j
(
i

i
)a
i
+
j
a
j
. (3.42)
Forming the inner product with p and observing c
T
i
p = 0, 1 i m, and
a
T
i
p = 0, i 1
ac
( x), i ,= j, we nd
p
T
Bp =
j
a
T
j
p . (3.43)
Since
j
< 0, we must have a
T
j
p 0.
If the second order sucient optimality conditions are satised , it follows from
(3.43) that
a
T
j
p = 0 p
T
Bp = p
T
Z
Z
T
BZp
Z
= 0 p
Z
= 0 ,
which implies p = 0. Due to the linear independence of the constraint gradients,
then (3.42) gives us
j
= 0, which is a contradiction. Hence, we must have
p
T
Bp > 0, whence a
T
j
p < 0.

Corollary 3.5 Strictly decreasing objective functional

Suppose that p
()
,= 0 is a solution of the quadratic subprogram (3.34a)-(3.34c)
and satises the second order sucient optimality conditions. Then, the ob-
jective functional of the original QP problem is strictly decreasing along the
direction p
()
.
Proof. Let us denote by Z the null space basis matrix associated with (3.34a)-
(3.34c). Then Z
T
BZ is positive denite, and we nd that p
()
is the unique
global minimizer of (3.34a)-(3.34c).
On the other hand, p = 0 is also a feasible point for (3.34a)-(3.34c). Conse-
quently, the value of the objective functional at p = 0 must be larger, i.e.,
1
2
(p
()
)
T
Bp
()
(b
()
)
T
p
()
< 0 .
Optimization I; Chapter 3 67
Since (p
()
)
T
Bp
()
0 and

[0, 1], we obtain

1
2

(p
()
)
T
Bp
()
(b
()
)
T
p
()
< 0 .
It follows that
1
2
(x
()

p
()
)
T
)B(x
)

p
()
) b
T
(x
()
+

p
()
) =
=
1
2
(x
()
)
T
Bx
()
b
T
x
()
+

_
1
2

(p
()
)
T
Bp
()
(b
()
)
T
p
()

<
<
1
2
(x
()
)
T
Bx
()
b
T
x
()
.

As far as the specication of the active set is concerned, after having com-
puted the Lagrange multipliers
()
,
()
, if p
()
= 0, usually the most negative
multiplier
()
i
is removed from the active set. This leads us to the following
algorithm:
Primal active set strategy
Step 1: Compute a feasible starting point x
(0)
and determine the set 1
ac
(x
(0)
)
of active inequality constraints.
Step 2: For 0, proceed as follows:
Compute p
()
as the solution of (3.34a)-(3.34c).
Case 1: If p
()
= 0, compute the multipliers
()
i
, 1 i m,
()
i
, i 1
ac
(x
()
)
that satisfy (3.39).
If
()
i
0, i 1
ac
(x
()
), then stop the algorithm.
The solution is x

= x
()
.
Otherwise, determine j 1
ac
(x
()
) such that

()
j
= min
iI
ac
(x
()
)

()
i
.
Set x
(+1)
= x
()
and 1
ac
(x
(+1)
) := 1
ac
(x
()
) j.
Case 2: If p
()
,= 0, compute

and set
x
(+1)
:= x
()

p
()
.
In case of blocking constraints, compute 1
ac
(x
(+1)
) according to (3.38).
There are several techniques to compute an initial feasible point x
(0)
T. A
common one requires the knowledge of some approximation x of a feasible point
which should not be too infeasible. It amounts to the solution of the following
Optimization I; Chapter 3 68
linear programming problem:
minimize e
T
z , (3.44a)
over (x, z) lR
n
lR
m+p
subject to c
T
i
x +
i
z
i
= c
i
, 1 i m , (3.44b)
a
T
i
x
m+i
z
m+i
d
i
, 1 i p , (3.44c)
z 0 , (3.44d)
where
e = (1, ..., 1)
T
,
i
=
_
sign(c
T
i
x c
i
) , 1 i m
1 , m + 1 i m+p
. (3.45)
A feasible starting point for this linear problem is given by
x
(0)
= x , z
(0)
i
=
_
[c
T
i
x c
i
[ , 1 i m
max(a
T
im
x d
im
, 0) , m + 1 i m+p
.(3.46)
Obviously, the optimal value of the linear programming subproblem is zero, and
any solution provides a feasible point for the original one.
Another technique introduces a measure of infeasibility in the objective func-
tional in terms of a penalty parameter > 0:
minimize
1
2
x
T
Bx x
T
b +t , (3.47a)
over (x, t) lR
n
lR
subject to c
T
i
x c
i
t , 1 i m , (3.47b)
(c
T
i
x c
i
) t , 1 i m , (3.47c)
a
T
i
x d
i
t , 1 i p , (3.47d)
t 0 . (3.47e)
For suciently large penalty parameter > 0, the solution of (3.47a)-(3.47e) is
(x, 0) with x solving the original quadratic programming problem.
3.5.2 Primal-dual active set strategies
We consider a primal-dual active set strategy which does not require feasibility
of the iterates. It is based on a Moreau-Yosida type approximation of the
indicator function of the convex set of inequality constraints
K := v lR
p
[ v
i
0 , 1 i p . (3.48)
The indicator function I
K
: K lR of K is given by
I
K
(v) :=
_
0 , v K
+ , v / K
. (3.49)
Optimization I; Chapter 3 69
The complementarity conditions
a
T
i
x

d
i
0 ,

i
0 ,

i
(a
T
i
x

d
i
) = 0 , 1 i p
can be equivalently stated as

I
K
(v

) , v

i
:= a
T
i
x

d
i
, 1 i p , (3.50)
where I
K
denotes the subdierential of the indicator function I
K
as given by
I
K
(v) I
K
(v) +
T
(w v) I
K
(w) , w lR
p
. (3.51)
Using the generalized Moreau-Yosida approximation of the indicator function
I
K
, (3.50) can be replaced by the computationally more feasible condition

=
_
v

+
1

P
K
(v

+
1

, (3.52)
where is an appropriately chosen positive constant and P
K
denotes the pro-
jection onto K as given by
P
K
(w) :=
_
w
i
, w
i
< 0
0 , w
i
0
, 1 i p .
Note that (3.52) can be equivalently written as

i
= max
_
0, a
T
i
x

d
i
+

_
. (3.53)
Now, given startiterates x
(0)
,
(0)
,
(0)
, the primal-dual active set strategy pro-
ceeds as follows: For 1 we determine the set 1
ac
(x
()
) of active constraints
according to
1
ac
(x
()
) := 1 i p [ a
T
i
x
()
d
i
+

()
i

> 0 . (3.54)
We dene
1
in
(x
()
) := 1, ..., p 1
ac
(x
()
) (3.55)
and set
p
()
:= card 1
ac
(x
()
) , (3.56)

(+1)
i
:= 0 , i 1
in
(x
()
) . (3.57)
We compute (x
(+1)
,
(+1)
,
(+1)
) lR
n
lR
m
lR
p
()
as the solution of the
KKT system associated with the equality constrained quadratic programming
Optimization I; Chapter 3 70
problem
minimize Q(x) :=
1
2
x
T
Bx x
T
b (3.58a)
over x lR
n
subject to c
T
i
x = c
i
, 1 i m , (3.58b)
a
T
i
x = d
i
, i 1
ac
(x
()
) . (3.58c)
Since feasibility is not required, any startiterate (x
(0)
,
(0)
,
(0)
) can be chosen.
The following results show that we can expect convergence of the primal-dual
active set strategy, provided the matrix B is symmetric positive denite.
Theorem 3.3 Reduction of the objective functional
Assume B lR
nn
to be symmetric, positive denite and refer to | |
E
:=
(()
T
B())
1/2
as the associated energy norm. Let x
()
, 0, be the iterates
generated by the primal-dual active set strategy. Then, for Q(x) :=
1
2
x
T
Bxx
T
b
and 1 there holds
Q(x
()
) Q(x
(1)
) = (3.59)
=
1
2
|x
()
x
(1)
|
2
E

iI
ac
(x
()

()
i
(d
i
a
T
i
x
(1)
) 0 .
i / I
ac
(x
(1)
)
Proof. Observing the KKT conditions for (3.58a)-(3.58c), we obtain
Q(x
()
) Q(x
(1)
) =
=
1
2
|x
()
x
(1)
|
2
E
+ (x
()
x
(1)
)
T
(Bx
()
b) =
=
1
2
|x
()
x
(1)
|
2
E

iI
ac
(x
()
)

()
i
a
T
i
(x
()
x
(1)
) .
For i 1
ac
(x
()
) we have a
T
i
x
()
= d
i
, whereas a
T
i
x
(1)
= d
i
for i 1
ac
(x
(1)
).
we thus get
Q(x
()
) Q(x
(1)
) =
=
1
2
|x
()
x
(1)
|
2
E

iI
ac
(x
()
)

()
i
(d
i
a
T
i
x
(1)
) .
i / I
ac
(x
(1)
)
But
()
i
0, i 1
ac
(x
()
) and a
T
i
x
(1)
d
i
, i / 1
ac
(x
(1)
) which gives the
assertion.

Corollary 3.4 Convergence of a subsequence to a local minimum

Optimization I; Chapter 3 71
Let x
()
, 0, be the iterates generated by the primal-dual active set strategy.
Then, there exist a subsequence lN

lN and x

lR
n
such that x
()
x

,
, lN

. Moreover, x

is a local minimizer of (3.29a)-(3.29c).

Proof. The sequence (x
()
)
lN
is bounded, since otherwise we would have
Q(x
()
) + as in contrast to the result of the previous theo-
rem. Consequently, there exist a subsequence lN

lN and x

lR
n
such
that x
()
x

, , lN

. Passing to the limit in the KKT system for

(3.58a)-(3.58c) shows that x

is a local minimizer.

The following result gives an a priori estimate in the energy norm.

Theorem 3.5 A priori error estimate in the energy norm
Let x
()
, 0, be the iterates generated by the primal-dual active set strategy
and lN

lN such that x
()
x

, , lN

. Then, there holds

|x
()
x

|
2
E
2

iI
ac
(x

i
(d
i
a
T
i
x
()
) . (3.60)
i / I
ac
(x
()
)
Proof. The proof is left as an exercise.
3.6 Interior-point methods
Interior-point methods are iterative schemes where the iterates approximate a
local minimum from inside the feasible set. For ease of exposition, we restrict
ourselves to inequality constrained quadratic programming problems of the form
minimize Q(x) :=
1
2
x
T
Bx x
T
b (3.61a)
over x lR
n
subject to Ax d, (3.61b)
where B lR
nn
is symmetric, positive semidenite, A = [a
i
]
1ip
lR
pn
, b
lR
n
, and d lR
p
.
We already know from Chapter 2 that the KKT conditions for (3.61a)-(3.61b)
can be stated as follows:
If x

lR
n
is a solution of (3.61a)-(3.61b), there exists a multiplier

lR
p
such that
Bx

+ A
T

b = 0 , (3.62a)
Ax

d 0 , (3.62b)

i
(Ax

d)
i
= 0 , 1 i p , (3.62c)

i
0 , 1 i p . (3.62d)
Optimization I; Chapter 3 72
By introducing the slack variable z := d Ax, the above conditions can be
equivalently formulated as follows
Bx

+ A
T

b = 0 , (3.63a)
Ax

+ z

d = 0 , (3.63b)

i
z
i
= 0 , 1 i p , (3.63c)
z

i
,

i
0 , 1 i p . (3.63d)
We can rewrite (3.63a)-(3.63d) as a constrained system of nonlinear equations.
We dene the nonlinear map
F(x, , z) =
_
_
Bx + A
T
b
Ax + z d
ZD

e
_
_
, z, 0 , (3.64)
where
Z := diag(z
1
, ..., z
p
) , D

:= diag(
1
, ...,
p
) , e := (1, ..., 1)
T
.
Given a feasible iterate (x, , z), we introduce a duality measure according to
:=
1
p
p

i=1
z
i

i
=
z
T

p
. (3.65)
Denition 3.6 Central path
The set of points (x

, z

) , > 0, satisfying
F(x

, z

) =
_
_
0
0
e
_
_
, z

> 0 (3.66)
is called the central path.
The idea is to apply Newtons method to (3.64) to compute (x

, z

) on
the central path, where [0, 1] is a parameter chosen by the algorithm. The
Newton increments (x, z, ) are the solution of the linear system
_
_
B A
T
0
A 0 I
0 Z D

_
_
_
_
x

z
_
_
=
_
_
r
b
r
d
ZD

e + e
_
_
, (3.67)
where
r
b
:= Bx + A
T
b , r
d
:= Ax + z d .
The new iterate (x, , z) is then determined by means of
(x, , z) = (x, , z) + (x, , z) (3.68)
Optimization I; Chapter 3 73
with chosen such that (x, , z) stays feasible.
Since D

is a nonsingular diagonal matrix, the increment z in the slack variable

can be easily eliminated resulting in
_
B A
T
A D
1

Z
__
x

_
=
_
r
b
r
d
D
1

g
_
, (3.69)
where g := ZD

e +e.
Optimization I; Chapter 3 74
3.7 Logarithmic barrier functions
We consider the inequality constrained quadratic programming problem (3.62a)-
(3.62b). Algorithms based on barrier functions are iterative methods where the
iterates are forced to stay within the interior
T
int
:= x lR
n
[ a
T
i
x c
i
< 0 , 1 i p . (3.70)
Barrier functions have the following properties:
They are innite outside T
int
.
They are smooth within T
int
.
They approach as x approaches the boundary of T
int
.
Denition 3.7 Logarithmic barrier function
For the quadratic programming problem (3.62a)-(3.62b) the objective function-
als
B
()
(x) := Q(x)
p

i=1
log(d
i
a
T
i
x) , > 0 (3.71)
are called logarithmic barrier functions. The parameter is referred to as the
barrier parameter.
Theorem 3.8 Properties of the logarithmic barrier function
Assume that the set o of solutions of (3.62a)-(3.62b) is nonempty and bounded
and that the interior T
int
of the feasible set is nonempty. Let
k

lN
be a
decreasing sequence of barrier parameters with
k
0 as k . Then there
holds:
(i) For any > 0 the logarithmic barrier function B
()
(x) is convex in T
Int
and attains a minimizer x() on T
int
. Any local minimizer x() is also a global
minimizer of B
()
(x).
(ii) If x(
k
)
lN
is a sequence of minimizers, then there exists lN

lN such
that x(
k
) x

o, k lN

.
(iii) If Q

is the optimal value of the objective functional Q in (3.62a)-(3.62b),

then for any sequence x(
k
)
lN
of minimizers there holds
Q(x(
k
)) Q

, B
(
k
)
(x(
k
)) Q

as k .
Proof. We refer to M.H. Wright; Interior methods for constrained optimiza-
tion. Acta Numerica, 341-407, 1992.

Optimization I; Chapter 3 75
We will have a closer look at the relationship between a minimizer of B
()
(x)
and a point (x, ) satisfying the KKT conditions for (3.62a)-(3.62b).
If x() is a minimizer of B
()
(x), we obviously have

x
B
()
(x()) =
x
Q(x()) +
p

i=1

d
i
a
T
i
x()
a
i
= 0 . (3.72)
Denition 3.9 Perturbed complementarity
The vector z
()
lR
p
with components
z
()
i
:=

d
i
a
T
i
x()
, 1 i p (3.73)
is called perturbed or approximate complementarity.
The reason for the above denition will become obvious shortly.
Indeed, in terms of the perturbed complementarity, (3.72) can be equivalently
stated as

x
Q(x()) +
p

i=1
z
()
i
a
i
= 0 . (3.74)
We have to compare (3.74) with the rst of the KKT conditions for (3.62a)-
(3.62b) which is given by

x
L(x, ) =
x
Q(x) +
p

i=1

i
a
i
= 0 . (3.75)
Obviously, (3.75) looks very much the same as (3.74).
The other KKT conditions are as follows:
a
T
i
x d
i
0 , 1 i p , (3.76a)

i
0 , 1 i p , (3.76b)

i
(a
T
i
x d
i
) = 0 , 1 i p . (3.76c)
Apparently, (3.76a) and (3.76b) are satised by x = x() and = z
()
.
However, (3.76c) does not hold true, since it follows readily from (3.73) that
z
()
i
(d
i
a
T
i
x()) = > 0 , 1 i p . (3.77)
On the other hand, as 0 a minimizer x() and the associated z
()
come
closer and closer to satisfying (3.76c). This is reason why z
()
is called perturbed
(approximate) complementarity.
Optimization I; Chapter 3 76
Theorem 3.10 Further properties of the logarithmic barrier function
Assume that T
int
,= and that x

is a local solution of (3.62a)-(3.62b) with

multiplier

such that the KKT conditions are satised.

Suppose further that the LICQ, strict complementarity, and the second order
sucient optimality conditions fold true at (x

). Then there holds:

(i) For suciently small > 0 there exists a strict local minimizer x() of
B
()
(x) such that the function x() is continuously dierentiable in some neigh-
borhood of x

and x() x

as 0.
(ii) For the perturbed complementarity z
()
there holds:
z
()

as 0 . (3.78)
(iii) For suciently small > 0, the Hessian
xx
B
()
(x) is positive denite.
Proof. We refer to M.H. Wright; Interior methods for constrained optimiza-
tion. Acta Numerica, 341-407, 1992.