Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
SIAM J. OPTIM.
c
2007 Society for Industrial and Applied Mathematics
Vol. 18, No. 2, pp. 666–690
ITERATIVE SOLUTION OF AUGMENTED SYSTEMS ARISING IN
INTERIOR METHODS
∗
ANDERS FORSGREN
†
, PHILIP E. GILL
‡
, AND JOSHUA D. GRIFFIN
§
Abstract. Iterative methods are proposed for certain augmented systems of linear equations
that arise in interior methods for general nonlinear optimization. Interior methods define a sequence
of KKT equations that represent the symmetrized (but indefinite) equations associated with New-
ton’s method for a point satisfying the perturbed optimality conditions. These equations involve both
the primal and dual variables and become increasingly ill-conditioned as the optimization proceeds.
In this context, an iterative linear solver must not only handle the ill-conditioning but also detect the
occurrence of KKT matrices with the wrong matrix inertia. A one-parameter family of equivalent
linear equations is formulated that includes the KKT system as a special case. The discussion focuses
on a particular system from this family, known as the “doubly augmented system,” that is positive
definite with respect to both the primal and dual variables. This property means that a standard
preconditioned conjugate-gradient method involving both primal and dual variables will either termi-
nate successfully or detect if the KKT matrix has the wrong inertia. Constraint preconditioning is a
well-known technique for preconditioning the conjugate-gradient method on augmented systems. A
family of constraint preconditioners is proposed that provably eliminates the inherent ill-conditioning
in the augmented system. A considerable benefit of combining constraint preconditioning with the
doubly augmented system is that the preconditioner need not be applied exactly. Two particular
“active-set” constraint preconditioners are formulated that involve only a subset of the rows of the
augmented system and thereby may be applied with considerably less work. Finally, some numerical
experiments illustrate the numerical performance of the proposed preconditioners and highlight some
theoretical properties of the preconditioned matrices.
Key words. large-scale nonlinear programming, nonconvex optimization, interior methods,
augmented systems, KKT systems, iterative methods, conjugate-gradient method, constraint pre-
conditioning
AMS subject classifications. 49J20, 49J15, 49M37, 49D37, 65F05, 65K05, 90C30
DOI. 10.1137/060650210
1. Introduction. This paper concerns the formulation and analysis of precon-
ditioned iterative methods for the solution of augmented systems of the form
(1.1)
H −A
T
AG
x
1
x
2
=
b
1
b
2
,
with A an m×n matrix, H symmetric, and G symmetric positive semidefinite. These
equations arise in a wide variety of scientific and engineering applications, where they
are known by a number of different names, including “augmented systems,” “saddle-
point systems,” “KKT systems,” and “equilibrium systems.” (The bibliography of the
survey by Benzi, Golub, and Liesen [3] contains 513 related articles.) The main focus
∗
Received by the editors January 17, 2006; accepted for publication (in revised form) March 5,
2007; published electronically August 22, 2007. The research of the second and third authors was
supported by National Science Foundation grants DMS-9973276, CCF-0082100, and DMS-0511766.
http://www.siam.org/journals/siopt/18-2/65021.html
†
Optimization and Systems Theory, Department of Mathematics, Royal Institute of Technology,
SE-100 44 Stockholm, Sweden (andersf@kth.se). The research of this author was supported by the
Swedish Research Council (VR).
‡
Department of Mathematics, University of California, San Diego, La Jolla, CA 92093-0112
(pgill@ucsd.edu).
§
Sandia National Laboratories, Livermore, CA 94551-9217 (jgriffi@sandia.gov). Part of this work
was carried out during the Spring of 2003 while this author was visiting KTH with financial support
from the G¨oran Gustafsson Foundation.
666
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
ITERATIVE SOLUTION OF AUGMENTED SYSTEMS 667
of this paper will be on the solution of augmented systems arising in interior methods
for general constrained optimization, in which case (1.1) is the system associated with
Newton’s method for finding values of the primal and dual variables that satisfy the
perturbed KKT optimality conditions (see, e.g., Wright [49] and Forsgren, Gill, and
Wright [15]). In this context H is the Hessian of the Lagrangian, A is the constraint
Jacobian, and G is diagonal.
Many of the benefits associated with the methods discussed in this paper derive
from formulating the interior method so that the diagonal G is positive definite.We
begin by presenting results for G positive definite and consider the treatment of sys-
tems with positive semidefinite and singular G in section 5. Throughout, for the case
where G is positive definite, we denote G by D and rewrite (1.1) as an equivalent
symmetric system Bx = b, where
(1.2) B =
H −A
T
−A −D
and b =
b
1
−b
2
,
with D positive definite and diagonal. We will refer to this symmetric system as the
KKT system. (It is possible to symmetrize (1.1) in a number of different ways. The
format (1.2) will simplify the linear algebra in later sections.) When D is nonsingular,
it is well known that the augmented system is equivalent to the two smaller systems
(1.3) (H + A
T
D
−1
A)x
1
= b
1
+ A
T
D
−1
b
2
and x
2
= D
−1
(b
2
− Ax
1
),
where the system for x
1
is known as the condensed system. It is less well known that
another equivalent system is the doubly augmented system
(1.4)
H +2A
T
D
−1
AA
T
AD
x
1
x
2
=
b
1
+2A
T
D
−1
b
2
b
2
,
which has been proposed for use with direct factorization methods by Forsgren and
Gill [16]. In this paper we investigate the properties of preconditioned iterative meth-
ods applied to system (1.2) directly or to the equivalent systems (1.3) and (1.4).
If the underlying optimization problem is not convex, the matrix H may be in-
definite. The KKT matrix B of (1.2) is said to have correct inertia if the matrix
H + A
T
D
−1
A is positive definite. This definition is based on the properties of the un-
derlying optimization problem. Broadly speaking, the KKT system has correct inertia
if the problem is locally convex (for further details see, e.g., Forsgren and Gill [16],
Forsgren [18], and Griffin [32]). If the KKT matrix has correct inertia, then systems
(1.2)–(1.4) have a common unique solution (see section 2).
1.1. Properties of the KKT system. The main issues associated with using
iterative methods to solve KKT systems are (i) termination control, (ii) inertia con-
trol, and (iii) inherent ill-conditioning. The first of these issues is common to other
applications where the linear system represents a linearization of some underlying non-
linear system of equations. Issues (ii) and (iii), however, are unique to optimization
and will be the principal topics of this paper.
In the context of interior methods, the KKT system (1.2) is solved as part of
a two-level iterative scheme. At the outer level, nonlinear equations that define the
first-order optimality conditions are parameterized by a small positive quantity μ.
The idea is that the solution of the parameterized equations should approach the
solution of the optimization problem as μ → 0. At the inner level, equations (1.2)
represent the symmetrized Newton equations associated with finding a zero of the
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
668 ANDERS FORSGREN, PHILIP E. GILL, AND JOSHUA D. GRIFFIN
perturbed optimality conditions for a given value of μ. Although systems (1.2)–(1.4)
have identical solutions, an iterative method will generally produce a different se-
quence of iterates in each case (see section 3 for a discussion of the equivalence of
iterative solvers in this context). An iterative method applied to the augmented sys-
tem (1.2) or the doubly augmented system (1.4) treats x
1
and x
2
as independent
variables, which is appropriate in the optimization context because x
1
and x
2
are as-
sociated with independent quantities in the perturbed optimality conditions (i.e., the
primal and dual variables). In contrast, an iterative solver for the condensed system
(1.3) will generate approximations to x
1
only, with the variables x
2
being defined as
x
2
= D
−1
(b
2
− Ax
1
). This becomes an important issue when an approximate solution
is obtained by truncating the iterations of the linear solver. During the early outer
iterations, it is usually inefficient to solve the KKT system accurately, and it is better
to accept an inexact solution that gives a residual norm that is less than some factor
of the norm of the right-hand side (see, e.g., Dembo, Eisenstat, and Steihaug [7]).
For the condensed system, the residual for the second block of equations will be zero
regardless of the accuracy of x
1
, which implies that termination must be based on the
accuracy of x
1
alone. It is particularly important for the solver to place equal weight
on x
1
and x
2
when system (1.2) is being solved in conjunction with a primal-dual
trust-region method (see Gertz and Gill [20] and Griffin [32]). The conjugate-gradient
version of this method exploits the property that the norms of the (x
1
,x
2
) iterates
increase monotonically (see Steihaug [44]). This property does not hold for (x
1
,x
2
)
iterates generated for the condensed system.
If the KKT matrix does not have the correct inertia, the solution of (1.2) is not
useful, and the optimization continues with an alternative technique based on either
implicitly or explicitly modifying the matrix H (see, e.g., Toint [45], Steihaug [44],
Gould et al. [30], Hager [33], and Griffin [32]). It is therefore important that the
iterative solver is able to detect if B does not have correct inertia.
As the perturbation parameter μ is reduced, the KKT systems become increas-
ingly ill-conditioned. The precise form of this ill-conditioning depends on the formu-
lation of the interior method, but a common feature is that some diagonal elements
of D are big and some are small. (It is almost always possible to formulate an interior
method that requires the solution of an unsymmetric system that does not exhibit
inevitable ill-conditioning as μ → 0. This unsymmetric system could be solved us-
ing an unsymmetric solver such as
GMRES or QMR. Unfortunately, this approach
is unsuitable for general KKT systems because an unsymmetric solver is unable to
determine if the KKT matrix has correct inertia.) In section 3 we consider a pre-
conditioned conjugate-gradient (PCG) method that provably removes the inherent
ill-conditioning. In particular, we define a one-parameter family of preconditioners
related to the class of so-called constraint preconditioners proposed by Keller, Gould,
and Wathen [34]. Several authors have used constraint preconditioners in conjunction
with the conjugate-gradient method to solve the indefinite KKT system (1.2) with
b
2
= 0 and D = 0 (see, e.g., Lukˇsan and Vlˇcek [36], Gould, Hribar, and Nocedal [29],
Perugia and Simoncini [40], and Bergamaschi, Gondzio, and Zilli [4]). Recently, Dol-
lar [12] and Dollar et al. [11] have proposed constraint preconditioners for system (1.2)
with no explicit inertial or diagonal condition on D, but a full row-rank requirement
on A and the assumption that b
2
=0.
Methods that require b
2
= 0 must perform an initial projection step that effec-
tively shifts the right-hand side to zero. The constraint preconditioner then forces the
x
1
iterates to lie in the null space of A. A disadvantage with this approach is that the
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
ITERATIVE SOLUTION OF AUGMENTED SYSTEMS 669
constraint preconditioner must be applied exactly if subsequent iterates are to lie in
the null space. This limits the ability to perform approximate solves with the precon-
ditioner, as is often required when the matrix A has a PDE-like structure that also
must be handled using an iterative solver (see, e.g., Saad [41], Notay [37], Simoncini
and Szyld [43], and Elman et al. [14]). In section 3 we consider preconditioners that
do not require the assumption that b
2
= 0, and hence do not require an accurate solve
with the preconditioner.
1.2. A PCG method for the KKT system. The goal of this paper is to for-
mulate iterative methods that not only provide termination control and inertia control,
but also eliminate the inevitable ill-conditioning associated with interior methods. All
these features are present in an algorithm based on applying a PCG method to the
doubly augmented system (1.4). This system is positive definite if the KKT matrix
has correct inertia, and gives equal weight to x
1
and x
2
for early terminations. As
preconditioner we use the constraint preconditioner
(1.5) P =
M +2A
T
D
−1
AA
T
AD
,
where M is an approximation of H such that M + A
T
D
−1
A is positive definite.
The equations Pv = r used to apply the preconditioner are solved by exploiting the
equivalence of the systems
M +2A
T
D
−1
AA
T
AD
v
1
v
2
=
r
1
r
2
,(1.6a)
M −A
T
−A −D
v
1
v
2
=
r
1
− 2A
T
D
−1
r
2
−r
2
, and(1.6b)
(M + A
T
D
−1
A)v
1
= r
1
− A
T
D
−1
r
2
,v
2
= D
−1
(r
2
− Av
1
)(1.6c)
(see section 3). This allows us to compute the solution of (1.6a) by solving either
(1.6b) or (1.6c). (The particular choice will depend on the relative efficiency of the
methods available to solve the condensed and augmented systems.)
We emphasize that the doubly augmented systems are never formed or factored
explicitly. The matrix associated with the doubly augmented equations (1.4) is used
only as an operator to define products of the form v = Bu. As mentioned above, the
equations (1.6a) that apply the preconditioner are solved using either (1.6b) or (1.6c).
An important property of the method is that these equations also may be solved using
an iterative method. (It is safe to use the augmented or condensed system for the
preconditioner equations Pv = r because the inertia of P is guaranteed by the choice
of M (see section 3).)
In section 4 we formulate and analyze two variants of the preconditioner (1.5)
that exploit the asymptotic behavior of the elements of D. The use of these so-called
active-set preconditioners may require significantly less work when the underlying
optimization problem has more constraints than variables. In section 5, we consider
the case where G is positive semidefinite and singular. Finally, in section 6, we present
some numerical examples illustrating the properties of the proposed preconditioners.
1.3. Notation and assumptions. Unless explicitly indicated otherwise, ·
denotes the vector two-norm or its subordinate matrix norm. The inertia of a real
symmetric matrix A, denoted by In(A), is the integer triple (a
+
,a
−
,a
0
) giving the
number of positive, negative, and zero eigenvalues of A. The spectrum of a (possi-
bly unsymmetric) matrix A is denoted by eig(A). As the analysis concerns matrices
Copyright © by SIAM. Unauthorized reproduction of this article is prohibited.
670 ANDERS FORSGREN, PHILIP E. GILL, AND JOSHUA D. GRIFFIN
with only real eigenvalues, eig(A) is regarded as an ordered set, with the least (i.e.,
“leftmost”) eigenvalue, denoted by eig
min
(A), appearing first. The quantity σ
k
(A)
denotes the kth largest singular value of A. Given a positive-definite A, the unique
positive-definite X such that X
2
= A is denoted by A
1/2
. Given vectors x
1
and x
2
,
the column vector consisting of the elements of x
1
augmented by the elements of x
2
is denoted by (x
1
,x
2
).
When μ is a positive scalar such that μ → 0, the notation p = O
μ
means that
there exists a constant K such that |p|≤Kμ for all μ sufficiently small. For a positive
p, p = Ω(1/μ) implies that there exists a constant K such that 1/p ≤ Kμ for all
μ sufficiently small. In particular, p = O
1
means that |p| is bounded, and, for a
positive p, p = Ω(1) means that p is bounded away from zero. For a positive p, the
notation p = Θ
1
is used for the case where both p = O
1
and p = Ω(1), so that p
remains bounded and is bounded away from zero as μ → 0.
As discussed in section 1.1, we are concerned with solving a sequence of systems of
the form (1.2), where the matrices A, H, and D depend implicitly on μ. In particular,
A and H are first and second derivatives evaluated at a point depending on μ, and
D is an explicit function of μ. The notation defined above allows us to characterize
the properties of H, A, and D in terms of their behavior as μ → 0. Throughout the
analysis, it is assumed that the following properties hold:
(A
1
) H and A are both O
1
.
(A
2
) The row indices of A may be partitioned into disjoint subsets S, M, and B
such that d
ii
= O
μ
for i ∈S, d
ii
= Θ
1
for i ∈M, and d
ii
= Ω(1/μ) for
i ∈B.
(A
3
)IfA
S
is the matrix of rows of A with indices in S and r = rank(A
S
), then r
remains constant as μ → 0 and σ
r
(A
S
) = Θ(1).
The second assumption reflects the fact that for μ sufficiently small, some diagonal
elements of D are “small,” some are “medium,” and some are “big.”
It is often the case in practice that the equations and variables corresponding to
unit rows of A are eliminated directly from the KKT system. This elimination creates
no additional nonzero elements and provides a smaller “partially condensed” system
with an Ω(1/μ) diagonal term added to H. It will be shown that preconditioners for
both the full and partially condensed KKT systems depend on the eigenvalues of the
same matrix (see Lemmas 3.4 and 3.5). It follows that our analysis also applies to
preconditioners defined for the partially condensed system.
2. A parameterized system of linear equations. In this section, it is shown
how the indefinite KKT system (1.2) may be embedded in a family of equivalent
linear systems, parameterized by a scalar ν. This parameterization facilitates the
simultaneous analysis of the three systems (1.2)–(1.4).
Definition 2.1 (the parameterized system). Let ν denote a scalar. Associ-
ated with the KKT equations Bx = b of (1.2), we define the parameterized equations
B(ν)x = b(ν), with
B(ν)=
H +(1+ν)A
T
D
−1
AνA
T
νA νD
and b(ν)=
b
1
+(1+ν)A
T
D
−1
b
2
νb
2
,
where H is symmetric and D is positive definite and diagonal.
The following proposition states the equivalence of the KKT system (1.2) and the
parameterized system of Definition 2.1.
Proposition 2.2 (equivalence of the parameterized systems). Let ν denote a
scalar parameter. If ν =0, then the system Bx = b of (1.2) and the system B(ν)x =