Interior Point Methods for Linear Programming: Just Call Newton, Lagrange, and Fiacco and McCormick! Author(s): Roy Marsten, Radhika Subramanian, Matthew Saltzman, Irvin Lustig, David Shanno Source: Interfaces, Vol. 20, No. 4, The Practice of Mathematical Programming (Jul. - Aug., 1990), pp. 105-116 Published by: INFORMS Stable URL: http://www.jstor.org/stable/25061374 . Accessed: 15/05/2011 04:03 Your use of the JSTOR archive indicates your acceptance of JSTOR's Terms and Conditions of Use, available at . http://www.jstor.org/page/info/about/policies/terms.jsp. JSTOR's Terms and Conditions of Use provides, in part, that unless you have obtained prior permission, you may not download an entire issue of a journal or multiple copies of articles, and you may use content in the JSTOR archive only for your personal, non-commercial use. Please contact the publisher regarding any further use of this work. Publisher contact information may be obtained at . http://www.jstor.org/action/showPublisher?publisherCode=informs. . Each copy of any part of a JSTOR transmission must contain the same copyright notice that appears on the screen or printed page of such transmission. JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms of scholarship. For more information about JSTOR, please contact
[email protected].
INFORMS is collaborating with JSTOR to digitize, preserve and extend access to Interfaces.
http://www.jstor.org
for Linear
Point Methods
Interior
Just Call Newton,
Programming: and Fiacco
Lagrange,
and McCormick! Roy Marsten
School of Industrial and Systems Engineering Georgia Institute of Technology Atlanta, GA 30332
Radhika
Subramanian
School of Industrial and Systems Engineering Georgia Institute of Technology
Matthew
Saltzman
of Mathematical
Department
Sciences
Clemson University, Clemson, SC 29631
Irvin Lustig
Deptartment of Civil Engineering and Operations Research
Princeton University, Princeton, NJ 08544
David
Shanno
RUTCOR, Rutgers University New Brunswick, NJ 08903
are the right way
Interior point methods programs.
are also much
They
to convert
a minimization
constrained
minimization.
to convert
a minimization
with
minimizations This
with
what field
re
heavily
involved
to pick a time to pause things to a wider audience. forum,
in
INTERFACES
20: 4 July-August
In
is ideal for
Copyright ? 1990, The Institute of Management Sciences 0091-2102/90/2004/0105$01.25 This paper was refereed.
1990
the
"new
era
gramming" landmark paper.
it difficult
and explain terfaces, as an informal
to
are
is really a dispatch from the battle rather than a scholarly paper.
with
but the rapid these methods, implementing constant excitement have progress and made
programs
In 1984, Narendra gan
report will be very up to date (as of late March 1990) but brief and to the
This
been
told us how
Newton
and inequalities. Voila!
equations
a status could not be complete without port on the new interior point methods.
have
into an un
told us how into a se constraints
Linear
Interfaces issue on the current state of the art in linear programming
point. We
Lagrange
inequality
minimizations.
and
told us how
constraints equality Fiacco and McCormick
minimizations.
unconstrained
motivate,
with
quence of unconstrained solve
to derive,
easier
than they at first appeared.
understand
to solve large linear
Karmarkar
of mathematical
[1984] be pro
of his the publication Shortly thereafter his em
ployer, AT&T, invited the professional to mathematical community programming as representa roll over and die. Speaking we took this as tives of this community, rather a challenge. to make dramatic PROGRAMMING?LINEAR TUTORIAL
(pp. 105-116)
of us proceeded to the improvements Some
MARSTEN, SUBRAMANIAN, SALTZMAN for example, Robert Bixby simplex method; with CPlex, John Forrest with OSL, and Michael Saunders with the new MINOS. Many others worked Karmarkar's method, peared to be coming field, into our classical mization.
on bringing at first ap which out of left completely for opti important re
The first and most
sult in this effort was which
framework
et al. [1986], an equivalence be Gill
demonstrated
tween Karmarkar's Newton
method
and Shanno
[1988],
Marsten,
and Shanno
[1989].
and projected
barrier methods.
Newton Recall Newton's
[1986] Interfaces focus on the algorithms,
xk+1
xk-f(xk)/f'(xk)
e =
10"8.
we
have n equations = 0, where
Suppose variables: f(x)
results. This
dXj
r1
Jacobian at the point x, J(x), is de as the matrix with (/, j) component
The
(X)'
and Newton's
method
looks
like
for unconstrained
optimization. for op Lagrange's [1788] method timization with equality constraints. Third,
Second,
and McCormick's
method
for optimization constraints. Using these show how
[1968] barrier with inequality
three pieces, we to construct the primal-dual
and also
the most
successful
See, for example, putationally. al. [1989] and Lustig, Marsten, [1989].
The primal-dual
has been
com
Marsten and method
[1986], by Megiddo developed and Yoshise Mizuno, Kojima, [1988], Monteiro and Adler [1989], McShane, and Shanno Monma, [1989], Choi,
INTERFACES20:4 106
et
=
xk+i
xk
if we
Or, dx
interior point method. is The primal-dual of the many the most elegant theoretically
Shanno
=
is, in fact,
consists of three crucial point methods building blocks. First, Newton's [1687] method for solving nonlinear equations
variants
x
and
fined
Fiacco
in n
article. We
the whole subject has much clearer and sim just recently gotten foundation for interior pler. The theoretical
and hence
a
=
their imple and their performance, rather
mentation, than on complexity a good time because
will
for finding of a single variable: method
zero of a function = 0. Given an initial estimate x? we f(x) compute a sequence of trial solutions
example;
lent Hooker
and Lustig,
for k = 0, 1, 2, ... , stopping when \f(xk)\ < ewhere e is some stopping tolerance; for
A great deal of theoretical and computa tional work has been done since the excel
will
Monma,
=
__ [/(x*)]"1/^*).
let
xk+1
-
xk
vector and move the displacement to the other side the Jacobian matrix
denote
J(x")dx
=
Newton's unconstrained well:
(1)
-f(xk). method
can be applied
to the
problem as = g(x) take/(x) g'(x) to search for a method
minimization
to minimize
and use Newton's
can zero of f(x). Each step of the method a quadratic be interpreted as constructing of g(x) and stepping di approximation
INTERIORPOINT METHODS to the minimum
rectly
approximation. If g is a real valued
L(x, y) by solving the system of in (n + m) variables: (n + m) equations
of n vari
function of g will
of n equations
system
following variables:
the unconstrained
function
a local minimum
ables,
and then minimize
of this
satisfy in n
the
?k = ?Lix)_ZvM(x) dXj dxj() ?1y'dxj( =
for
j
1,
. . ., n
1,
. . ., m.
'
= o
and
f:w=o i
for
In this case
the Newton
iteration
( 1 ) looks
like
equations
Newton's
-Vg(xk),
component
x > 0. The idea of conditions: negativity the barrier function approach is to start from a point in the strict interior of the in
??(x).
dXjdXj If g is convex,
then any local minimizer is also a global minimizer. If x* is a local minimizer of g(x); that is, Vg(x*) = 0, and if V2g(x*) method
has
will
sufficiently
full rank,
then Newton's
to x* if started converge close to x*.
(x? > 0 for all /) and construct a equalities barrier that prevents any variable from = 0). For exam reaching the boundary (xf ple, adding "?log (x7)" to the objective function
Of
Lagrange constrained
discovered
how
to transform
optimization problem, with into an unconstrained constraints,
equality
To solve
problem. minimize
the problem
= 0 subject to gi(x) i
=
1,
. . ., m
form a Lagrangian
f{x)-
a
course
0. is on
if the constrained
optimum = some the boundary is, 0, which (that xj is always true for linear programming), then
prevent us from solution is to use a barrier
the barrier will
reaching
it. The
that balances the contribution of parameter the true objective function against that of the barrier
function
2m 0
to
subject f(x) is replaced by the family
of unconstrained
7 1
on the positive ?x.Let x(/x) be the mini
is parameterized
barrier parameter mizer of B(x|??).
Fiacco and McCormick that x*, the con x(p) [1968] as \x 0. The set of strained minimizer, show
minimizers
x(??) is called
the "central
trajectory."
As
a simple
example,
consider
the
XX
0,
=
. . . , n.
1,
for;
we
In practice,
do not have
0.
on barrier methods
1955],
Carroll
Primal-Dual with
constrained
Interior Newton's
minimization
(x1)-Mlog(x2) a dB = OXi
dB a= ox2
Xi(M)
2(x1 + 2(x2 + =
X2(M)
for un
and
the Lagran for converting into unconstrained
linear programming. the primal-dual Consider
(x2+l)2-Mlog
Point Method
and barrier methods
>0:
?
and Huard
method
gian constrained
l)2 +
of Frisch
[1961],
The unconstrained minimum would be at ? is ( 1, ?1), but the constrained minimum at the origin (x*, x%) = (0, 0). For any \i
+
to compute
?i. x(ijl) very accurately before reducing extended the ear Fiacco and McCormick
Armed
?(x|M)-(x1
-
M_Q
[1964].
to >
0.
x;
dx;
The
x2
->
as ^
dg(;,U)_d/(*) dx;
[1954,
(Xl+ l)2 + (x2+ l)2
>
x*
in n variables:
equations
lier work
problem minimize
subject
->
xk(nk)
?k.
1 and go to 2.
In step 2, we can find x{?i) by using to solve the system of n Newton's method
2 = log(zy)
B(x\ix)=f(x)-?
/j,k+1
+ -Vl
bTy
subject
to
ATy
c
0.
+ 2M
-* 0. which (0, 0) as ?x approaches In general we can't get a closed-form solution for the central trajectory, but can use
the following algorithm: = 0. 1. Choose ju0> 0, set k
INTERFACES20:4 108
can be handled by La The equations and the nonnegativity grange's method Fiacco and McCormick's conditions by barrier method. strained
functions
by Newton's
The resulting uncon can then be optimized
method.
Suppose
that we
INTERIORPOINT METHODS first introduce two Lagrangians:
2 ATy +
and then form
the barriers
= cTx Lp(x, y)
-
?i2
= 7 1
log (Xj)
-
b)
yT(Ax
n
=
Ld(x, y, z) -
bTy +
zxT(ATy +
?i2
= 7 1
dP
c).
dD
all of the partial derivatives them to zero, we get just three sets of equations (because of duplication).
for
bility,
=
1,
slackness.
mentary
ATy? + z?-c the primal
That
is, p
we will
start p at some positive let it approach zero. If we
matrix
let diag { } denote with the listed elements
=
value
and
a diagonal on its
A 0 0 0 AT 1 Z 0 X
diag
=