p

u
t
s
)

a
r
e

s
e
t

b
y

p
l
a
n
n
i
n
g

b
o
a
r
d
.
T
h
e

p
l
a
n
n
i
n
g

b
o
a
r
d
'
s

j
o
b

i
s

t
o

m
i
n
i
m
i
z
e

t
h
e

t
o
t
a
l
c
o
s
t

o
f

p
r
o
c
u
r
i
n
g

t
h
e

s
e
t

a
m
o
u
n
t
s

o
f

i
n
p
u
t
s

w
h
i
l
e

p
r
o
v
i
d
i
n
g
t
h
e

f
a
r
m
e
r

w
i
t
h

f
l
o
o
r

o
n

t
h
e

u
n
i
t
p
r
i
c
e

o
f

e
a
c
h

o
f

h
i
s

c
r
o
p
s

(
o
u
t
p
u
t
s
)
,
 
S
1

 
f
o
r

w
h
e
a
t

a
n
d
 
S
2

 
f
o
r

b
a
r
l
e
y
.

T
h
i
s

c
o
r
r
e
s
p
o
n
d
s

t
o

t
h
e

f
o
l
l
o
w
i
n
g

l
i
n
e
a
r

p
r
o
g
r
a
m
m
i
n
g

p
r
o
b
l
e
m
:
M(
i m
ni
i n
mi
i m
zi
ez
:  e
A
yt
h
A
e
 
+t
 o
Ft
a
yl
F
c
 o
+s
 t
P
yo
f
P

t
h
e

m
e
a
n
s

o
f

p
r
o
d
u
c
t
i
o
n

a
s

t
h
e

"
o
b
j
e
c
t
i
v
e

f
u
n
 c
t
i
o
n
"
)
Sy(
ut
A
bh
 
j e
e+
c f
t Fa
1r
t m
oe
y
: Fr
 
+m
 u
Ps
1
t
y
r
P
e
 c
≥e
 i
Sv
e
1

n
o

l
e
s
s

t
h
a
n
 
S
1
 
f
o
r

h
i
 s

w
h
e
a
t
)
y(
t
A
h
 e
+
 f
Fa
2r
 m
yer
F
 m
+u
 s
Pt
2
yre
Pc
 e
≥i
 v
Se
2
n
o

l
e
s
s

t
h
a
n
 
S
2
 
f
o
r

h
 i
s

b
a
r
l
e
y
)
(
yp
Ar
 i
c
≥e
s
0
,c
 a
n
yn
Fo
 t
≥
b
e
0
,n
 e
yg
P
a
 ti
≥v
e
0)
.

W
h
i
c
h

i
n

m
a
 t
r
i
x

f
o
r
m

b
e
c
o
m
e
s
:

Minimize: 

Subject to: 

The primal
problem
deals with
physical
quantities.
With all
inputs
available in
limited
quantities,
and
assuming
the unit
prices of
all outputs
is known,
what
quantities
of outputs
to produce
so as to
maximize
total
revenue?
The dual
problem
deals with
economic
values.
With floor
guarantees
on all
output unit
prices, and
assuming
the
available
quantity of
all inputs is
known,
what input
unit pricing
scheme to
set so as
to
minimize
total
expenditur
e?

To each
variable in
the primal
space
correspond
s an
inequality
to satisfy in
the dual
space,
both
indexed by
output
type. To
each
inequality
to satisfy in
the primal
space
correspond
sa
variable in
the dual
space,
both
indexed by
input type.

The coefficients that

inequalities in the ess
The theorem states:

Suppose that x = (x1,
is primal feasible and
that y = (y1, y2, ... , ym
feasible. Let (w1, w2, 
denote the correspon
slack variables, and l
(z1, z2, ... , zn) denote
corresponding dual s
variables. Then x and
optimal for their respe
problems if and only

 xjzj = 0, for j = 1, 
and
 wiyi = 0, for i = 1,

So if the i-th slack va
primal is not zero, the
variable of the dual is
zero. Likewise, if the 
variable of the dual is
then the j-th variable
primal is equal to zer

This necessary cond

optimality conveys a
economic principle. I
form (when maximizi
is slack in a constrain
resource (i.e., there a
"leftovers"), then add
quantities of that reso
have no value. Likew
is slack in the dual (s
price non-negativity c
requirement , i.e., the
zero, then there mus
supplies (no "leftover

[edit]Theory
[edit]Existenceo
optimal solutio
Geometrically, the lin
constraints define the
region, which is
a convex polyhedron
function is a convex f
which implies that ev
minimum is a global
similarly, a linear fun
a concave function, w
implies that every loc
maximum is a global

Optimal solution need

for two reasons. First
constraints are incon
no feasible solution e
instance, the constra
and x ≤ 1 cannot be
jointly; in this case, w
the LP is infeasible. S
when the polytope is
in the direction of the
the objective function
gradient of the object
is the vector of the co
the objective function
optimal value is attain

[edit]Optimalver
(and rays) of po
Otherwise, if a feasib
exists and if the (linea
function is bounded,
optimum value is alw
attained on the bound
optimal level-set, by
the maximum
principle for convex
functions (alternative
the minimum principl
for concave functions
that linear functions a
convex and concave.
some problems have
optimal solutions: Fo
the problem of finding
solution to a system o
inequalities is a linea
programming problem
the objective function
function (that is, the c
function taking the va
everywhere): For this
problem with the zero
for its objective-funct
are two distinct soluti
every convex combin
solutions is a solution

The vertices of the po

also called basic feas
solutions. The reason
choice of name is as
Let d denote the num
variables. Then the fu
theorem of linear ineq
implies (for feasible p
that for every vertex 
feasible region, there
of d (or fewer) inequa
constraints from the L
that, when we treat
those d constraints a
the unique solution is
Thereby we can stud
vertices by means of
certain subsets of the
constraints (a discret
rather than the contin
solutions. This princip
underlies the simplex
algorithm for solving
programs.

[edit]Algorithm

A series of linear constra

variables produces a reg
possible values for those
Solvable problems will h
region in the shape of a 
polygon.
[edit]Simplex alg
of Dantzig
The simplex algorithm
developed by George
1947, solves LP prob
constructing a feasib
at a vertex of the poly
then walking along a
edges of the polytope
with non-decreasing
the objective function
optimum is reached.
practical problems, "s
occurs: Many pivots
with no increase in th
function.[2] In rare pra
problems, the usual v
the simplex algorithm
actually "cycle".[3] To
cycles, researchers d
new pivoting rules .[4]

In practice, the
simplex algorithm is q
efficient and can be g
to find the global opti
certain precautions
against cycling are ta
simplex algorithm ha
proved to solve "rand
problems efficiently, i
cubic number of step
(Borgwardt, Todd), w
similar to its behavior
practical problems[5]

However, the simplex

has poor worst-case
Klee and Minty const
family of linear progra
problems for which th
method takes a numb
exponential in the pro
[6]
 In fact, for some tim
not known whether th
programming problem
solvable in polynomia
time (complexity clas

[edit]Ellipsoid
al
following Khac
This long standing iss
resolved by Leonid K
1979 with the introdu
the ellipsoid method,
worst-case polynomi
algorithm for linear
programming. To sol
problem which has n
and can be encoded
bits, this algorithm
uses O(n4L) pseudo-
operations on numbe
with O(L) digits.
Khahiyan's algorithm
convergence analysis
(real-number) predec
notably the iterative
methods developed b
Shor and the approxi
algorithms by Arkadi
and D. Yudin.

[edit]Interior
poi
methods, follow
Karmarkar
Khachiyan's algorithm
landmark importance
establishing the polyn
solvability of linear pr
The algorithm had litt
impact, as the simple
more efficient for all b
constructed families o
programs. However,
new lines of research
programming with the
development of interi
methods, which can
implemented as a pra
In contrast to the sim
algorithm, which finds
optimal solution by pr
along points on the b
a polytopal set, interi
methods move throug
interior of the feasible

In 1984, N. Karmarka
a new interior point p
method for linear
programming. Karma
algorithm not only im
Khachiyan's theoretic
case polynomial bou
(giving O(n3.5L)). Kar
claimed that his algor
exhibited dramatic pr
performance improve
the simplex method,
created great interes
point methods. Since
interior point methods
proposed and analyz
successful implemen
were based on affine
scaling variants of th
For both theoretical a
properties, barrier
function or path-follow
ds are the most comm
recently.

[edit]Other algor
General optimization
such as gradient des
also be used to solve
Though not specializ
these solvers may of
advantages, such as
advantage of problem
(e.g. if the problem
involves sparse matr
structured matrix like
Fourier transform). A
possible advantage is
solvers do not always
LP in canonical form.

In other cases, speci

linear programs are s
custom algorithms; s
example is the origin
proposed to solve the

[edit]Compariso
interior-point m
versus simplex
algorithms
The current opinion is
efficiency of good
implementations of s
based methods and i
methods are similar f
applications of linear
programming.[7] Howe
specific types of LP p
may be that one type
better than another (s
much better).

LP solvers are in wid

use for optimization o
problems in industry,
optimization of flow in
transportation networ

[edit]Open
prob
and recent wo
There are several op
problems in the theor
programming, the so
which would represen
fundamental breakthr
mathematics and pot
major advances in ou
solve large-scale line
programs.

 Does LP admit a
polynomial-time a
 Does LP admit a
polynomial algori
a strictly complem
solution?
 Does LP admit a
algorithm in the r
(unit cost) model
computation?
This closely related s
problems has been c
by Stephen Smale as
the 18 greatest unsol
problems of the 21st
Smale's words, the th
of the problem "is the
unsolved problem of
programming theory.
algorithms exist to so
programming in weak
polynomial time, such
the ellipsoid
methods and interior
techniques, no algori
yet been found that a
strongly polynomial-t
performance in the n
constraints and the n
variables. The develo
such algorithms woul
great theoretical inter
perhaps allow practic
solving large LPs as

Although the Hirsch
conjecture was recen
disproved for higher d
it still leaves the follo
questions open.

 Are there pivot ru

lead to polynomia
Simplex variants
 Do all polytopal g
polynomially-bou
diameter?

These questions rela

performance analysis
development of Simp
methods. The immen
efficiency of the Simp
algorithm in practice
exponential-time theo
performance hints tha
may be variations of
that run in polynomia
strongly polynomial ti
would be of great pra
theoretical significanc
whether any such va
particularly as an app
deciding if LP can be
strongly polynomial ti

The Simplex algorithm

variants fall in the fam
following algorithms,
because they solve li
programming problem
moving from vertex to
along edges of a poly
means that their theo
performance is limite
maximum number of
between any two ver
LP polytope. As a res
interested in knowing
maximum graph-theo
diameter of polytopal
has been proved that
polytopes have sube
diameter. The recent
the Hirsch conjecture
step to prove whethe
polytope has superpo
diameter. If any such
exist, then no edge-fo
variant can run in pol
time. Questions abou
diameter are of indep
mathematical interes

Simplex pivot method

primal (or dual) feasib
the other hand, criss-
methods do not prese
or dual) feasibility—th
visit primal feasible, d
or primal-and-dual in
bases in any order. P
methods of this type
studied since the 197
Essentially, these me
attempt to find the sh
path on the arrangem
polytope under the lin
programming problem
contrast to polytopal
graphs of arrangeme
polytopes are known
small diameter, allow
possibility of strongly
time criss-cross pivot
without resolving que
about the diameter o
polytopes.[9]

[edit]Integer
unknowns
If the unknown variab
required to be intege
problem is called an 
programming (IP) or 
linear programming
problem. In contrast t
programming, which
solved efficiently in th
case, integer program
problems are in many
situations (those with
variables) NP-hard. 0
programming or bin
programming (BIP)
special case of intege
programming where
are required to be 0 o
than arbitrary integer
problem is also class
hard, and in fact the d
version was one of K
NP-complete problem

If only some of the un

variables are required
integers, then the pro
called a mixed integ
programming (MIP)
These are generally a
hard.
There are however s
important subclasses
MIP problems that ar
solvable, most notab
where the constraint
is totally unimodular 
right-hand sides of th
constraints are intege

Advanced algorithms
integer linear program

 cutting-plane me
 branch and boun
 branch and cut
 branch and price
 if the problem ha
extra structure, it
possible to apply
column generatio

Such integer-program
algorithms are discus
Padberg and in Beas

[edit]Integral Li
Programs
A linear program in re
variables is said to be
it has at least one op
solution which is inte
Likewise, a
polyhedron 

to be integral if for a
feasible objective fun
the linear
program 

optimum x * with inte
coordinates. As obse
Edmonds and Giles i
can equivalently say
polyhedron is integra
bounded feasible inte
objective function c, t
optimal value of the l
progam 
is an integer.

Integral linear progra

central importance in
polyhedral aspect
of combinatorial
optimization since the
an alternate characte
problem. Specifically
problem, the convex
solutions is an integra
polyhedron; if this po
has a nice/compact d
then we can efficientl
optimal feasible solut
any linear objective. C
if we can prove that a
programming relaxat
integral, then it is the
description of the con
feasible (integral) sol

Note that terminology

consistent throughou
literature, so one sho
careful to distinguish
following two concep

 in an integer line
program, describ
previous section,
are forcibly const
be integers, and
problem is NP-ha
general,
 in an integral line
program, describ
section, variables
constrained to be
but rather one ha
somehow that th
continuous probl
has an integral o
(assuming c is in
this optimal value
found efficiently s
polynomial-size l
programs can be
polynomial time.

One common way of

that a polyhedron is i
show that it is totally
There are other gene
including the integer
decomposition prope
dual integrality. Othe
well-known integral L
the matching polytop
polyhedra, submodul
polyhedra, and the in
of 2 generalized poly
polymatroids --- e.g.
Schrijver 2003.

A bounded integral p
sometimes called a c
lattice polytope, parti
two dimensions.

[edit]Solvers an
scripting
(programming
languages
Free, opensource:
MINTO| here [11] Mixe
Optimizer (an integer
programming solver
branch and bound alg
has publicly available
code but not open so

Proprietary:
[edit]See also

 Mathematical pro
 Nonlinear progra
 Convex program
 Simplex algorithm
solve LP problem
 Quadratic progra
superset of linea
programming
 Leonid Kantorovi
the founders of li
programming
 Shadow price
 MPS file format
 nl file format
 MIP example, job
problem
 Linear-fractional
programming (LF
 Oriented matroid

 see also the "Ext

section below

[edit]Reference

1. ^ Vazirani (2001, p

2. ^ Dantzig and Tha

3. ^ Padberg.

4. ^ Bland; Fukuda a

also Murty, Dantzig

Padberg, Papadim

5. ^ Dantzig and Tha

6. ^ Dantzig and Tha

and Steiglitz, Murty

7. ^ Beasley, Todd.

8. ^ For solving netw

in transportation ne

specialized implem

simplex algorithm c

improve its efficien

and Thapa.
9. ^ See Fukuda and

10. ^ http://www.mose

11. ^ http://coral.ie.leh

wnload.html

12. ^ name="Linear pr

at CenterSpace

Software">http://ww

et/landing.php?id=

[edit]Further re
A reader may consid
beginning with Nering
Tucker, with the first
Dantzig and Thapa, o
Williams.

 Dmitris Alevras a
W. Padberg, Line
Optimization and
Extensions: Prob
Extensions, Univ
Springer-Verlag,
(Problems from P
with solutions.)
 A. Bachem and W
Kern. Linear Prog
Duality: An Introd
Oriented Matroid
Universitext. Spr
Verlag, 1992.
(Combinatorial)
 J. E. Beasley,
editor. Advances
and Integer Prog
 Oxford Science,
(Collection of sur
 Mark de Berg, M
Kreveld, Mark Ov
and Otfried Schw
(2000). Computa
Geometry (2nd re
ed.). Springer-Ve
3-540-65620-0. C
Linear Programm
pp. 63–94. Descr
randomized half-
intersection algor
linear programmi
 R. G. Bland, New
pivoting rules for
method, Math. O
(1977) 103–107.
 Karl-Heinz Borgw
Simplex Algorithm
Probabilistic Ana
Algorithms and
Combinatorics, V
Springer-Verlag,
(Average behavio
random problems
 Richard W. Cottle
Basic George B.
Stanford Busines
Stanford Univers
Stanford, Californ
(Selected papers
B. Dantzig)
 George B. Dantz
Mukund N. Thap
1997. Linear prog
 1: Introduction. S
Verlag.
 George B. Dantz
Mukund N. Thap
2003. Linear Pro
2: Theory and Ex
Springer-Verlag.
(Comprehensive
e.g. pivoting and
point algorithms,
problems, decom
following Dantzig
Wolfe and Bende
introducing stoch
programming.)
 Edmonds, J. and
"A min-max relat
submodular func
graphs," Ann. Dis
Math., v1, pp. 18
 Fukuda, Komei; T
Tamás (1997). "C
methods: A fresh
pivot
algorithms". Math
Programming:
Series  B(Amster
Holland
Publishing Co.) 7
from the 16th Inte
Symposium on
Mathematical Pro
held in Lausanne
pp. 369—395. do
i:10.1016/S0025-
 5610(97)00062-2
R1464775.
 Michael R. Garey
S.
Johnson (1979). 
and Intractability:
the Theory of NP
Completeness. W
Freeman. ISBN 0
1045-5. A6: MP1
PROGRAMMING
(computer scienc
complexity theory
 Bernd Gärtner, J
Matoušek (2006)
ding and Using L
Programming, Be
Springer. ISBN 3
30697-8 (elemen
introduction for
mathematicians a
computer scientis
 Kattta G. Murty, L
Programming, W
(comprehensive
classical approac
 Evar D. Nering a
W. Tucker, 1993
Programs and Re
Problems, Acade
(elementary)
 M. Padberg, Line
Optimization and
Extensions, Seco
Springer-Verlag,
(carefully written
 primal and dual s
algorithms and p
algorithms, with a
introduction to in
programming ---
the traveling sale
problem for Odys
 Christos H.
Papadimitriou an
Steiglitz, Combin
Optimization: Alg
and Complexity,
republication with
preface, Dover. (
science)
 Cornelis Roos, T
Terlaky, Jean-Ph
Vial, Interior Poin
for Linear Optimi
Second Edition, S
Verlag, 2006. (G
level)
 Alexander Schrijv
(2003). Combina
optimization: poly
efficiency. Spring
 Alexander Schrijv
of Linear and Inte
Programming. Jo
sons, 1998, ISBN
98232-6 (mathem
 Michael J. Todd
2002). "The man
linear
programming". M
l
 Programming 91
survey, from the
International Sym
Mathematical
Programming.)
 Robert J. Vander
Programming: Fo
and Extensions,
International Ser
Operations Rese
Management Sci
114, Springer Ve
2008. ISBN 978-
74387-5. (An on-
edition was forme
available. Vande
still contains exte
materials.)
 Vazirani, Vijay
V. (2001). Appro
Algorithms. Sprin
Verlag. ISBN 3-5
8. (Computer sci
 H. P. Williams, M
Building in Mathe
Programming, Th
Edition, 1990. (M
 Stephen J. Wrigh
1997, Primal-Dua
Point Methods, S
(Graduate level)
 Yinyu Ye, 1997, 
Point Algorithms:
and Analysis, Wi
(Advanced gradu
 Ziegler, Günter M
1–3 and 6–7 in L
Polytopes, Spring
New York, 1994.

[edit]External l

 Guidance on For
problems