A New Polynomial-Time Algorithm for Linear Programming

COMBINATOR1CA4 (4) (1984) 373--395

A N E W POLYNOMIAL-TIME ALGORITHM
FOR LINEAR P R O G R A M M I N G

N. K A R M A R K A R

Received 20 August 1984
Revised 9 November 1984

We present a new polynomial-time algorithm for linear programming. In the worst case,
the algorithm requires O(tf'SL) arithmetic operations on O(L) bit numbers, where n is the number of
variables and L is the number of bits in the input. The running,time of this algorithm is better than
the ellipsoid algorithm by a factor of O(n~'~). We prove that given a polytope P and a strictly in-
terior point a E P, there is a projective transformation of the space that maps P, a to P', a' having
the following property. The ratio of the radius of the smallest sphere with center a', containing
P ' to the radius of the largest sphere with center a' contained in P ' is O(n). The algorithm consists
of repeated application of such projective transformations each followed by optimization over an
inscribed sphere to create a sequence of points which converges to the optimal solution in poly-
nomial time.

1. Informal outline

In this Section we give an informal discussion of the main ideas involved in
the algorithm. A more rigorous description will be given in the next Section.

1.1. Problem definition

We shall reduce the general linear programming problem over the rationals
to the lollowing particular case:

minimize crx
Ax=O
subject to ~ ~'xi = 1
(x=>O.
Here x = ( x l . . . . , x.)TER ", cEZ", A E Z " ~ ' .

This is a substantially revised version of the paper presented at the Symposium on Theory
of Computing, Washington D. C., April 1984.
AMS subject classification (1980): 90 C 05

374 N. KARMARKAR

Let 12 denote the subspace {xlAx=0}. Let A denote the simplex A =
= {xlx >=O, .~xi = 1} and let H be the polytope defined by
1I = ONA.
Thus the problem we consider is
minimize crx
subject to xE/7.

1.2. Optimization over a sphere

If we replace the constraint xEA by a constraint of the form x~S, where S
is a sphere, then it is trivial to see that the optimum can be found by solving a
linear system of equations.
The intersection of a sphere and an affine space is a lower dimensional
sphere inside that affine space. Hence the problem becomes:
min c'rx, subject to xES" (a sphere)
where c' was obtained by projecting c orthogonally onto t2.
But this problem is trivial: From the center of the sphere, take a step along
the direction - c ' , of length equal to the radius of the sphere.

1.3. Bounds on the objective function

Let a0 be strictly interior point in the polytope P. Suppose we draw an ellip-
soid E with center a0, that is contained in the polyt ope and solve the optimization prob-
lem over the restricted region E instead of P. How good a solution do we get as com-
pared to the optimal solution? To derive a bound, we create another ellipsoid E'
by magnifying E by a sufficiently large factor v so that E' contains P.
E c P c E', E" = rE.

Let rE, A , fr, denote the minimum values of objective function f ( x ) on E, P, and
E' respectively.
f ( a o ) - f ~ <----f(ao)--fe ~--f(ao)-fw = vtf(ao)-fe].
The last equation follows from the linearity off(x).
f(a0) - f ~ 1
f(ao)-fp- v
~-~ ~ (1_1).
f(ao) - f e
Thus by going from a0 to the point, say a', that minimizesf(x) over E we come closer
to the minimum value of the objective function by a factor I - v -1. We can repeat
the same process with a' as the center. The rate of convergence of this method depends
on v, the smaller the value of v, the faster the convergence.

A NEW POLYNOMIAL-TIME ALGORITHM FOR LINEAR P R O G R A M M I N G 375

1.4. Projective transformations

Projective transformations of R" are described by formulas of the form
Cx+d
x~ frx+g
where C6R *X", d, f6R", g~R and the matrix

is non-singular.
We shall be interested in projective transformations of the hyperplane
2:= { IZx, = 1}.

It is easy to see that these transformations are described by
Cx
X ~" eTfx

where e=(1, 1. . . . . 1)r and C is a non-singular n×n matrix.
For each interior point a of A there is a unique projective transformation
T(a, ao) of the hyperplane 2: fixing all vertices of the simplex A and moving a =
= (al .... , a,) r to the center a0=(1/n)e. This transformation is given by
t _ x_i / a f .
x~ = Z xJaj
J

Observe that the corresponding matrix C is now diagonal: C =
,1
= d i a g [ 1 .... 7,,)" The transformation C(a, a0) has the following properties :

1. T is one-to-one and maps the simplex onto itself. Its inverse is given by
a i XPt
Xi = ,~ adxj i= 1,2,...,n.
J

2. Each facet of the simplex given by xi=0 is mapped onto the corresponding
facet xi=0.
1
3. Image of the point x = a is the center of the simplex given by X-' ~ - - .
I
n
4. Let Ai denote the ith column of A. Then the system of equations
z~ Aixi = 0
i
now becomes
.~ Aiaix~
t
=0
Z a,x;
i

376 N. K A R M A R K A R

or
Z A;x; = o
i
where
A~ = aiAi.
We denote the affine subspace {x'lA'x'=0} by f2".
5. If a~ f2 then the center of the simplex aoC f2'.
Let B(ao, r) be the largest ball with center a0, the same as the center of the
simplex, and radius r that is contained in the simplex. Similarly let B(ao, R) be
the smallest ball containing A. It is easy to show that R / r = n - 1 .
B(a0, r) ~ A ~ B(ao, R)
B(a0, r)O~2' =c A (q (2" ~ B(a0, R)N f2'.
But A fqf2' is the image 17' of the polytope H=Af~f2. The intersection of a ball
and an affine subspace is a ball of lower dimension in that subspace and has the
same radius if the center of the original ball lies in the affine subspace, which it
does in our case. Hence we get
R
B'(ao, r ) ~ H ' ~ B ' ( a o , R), -r- = n - 1
which proves that v =dimension of the space can always be achieved by this method.
(Note: A is an n - 1 dimensional simplex.)

1.5. Invariant potential functions
The algorithm creates a sequence of points x (°), x (1), ..., x (k), ... having dec-
reasing values of the objective function. In the k th step, the point x (k) is brought
into the center by a projective transformation. Then we optimize the objective
function over the intersection of the inscribed ball and the affine subspace to find
the next point x (k+l). Based on the results of the previous Sections, we expect the
objective function to be reduced by a factor of (1 - n - 0 at least, in each step, but
there is one more technicality we have to deal with. The set of linear functions is
not invariant under a projective transformation, but ratios of linear functions are
transformed into ratios of linear functions. With every linear objective function
/(x) we associate a "potential function" g(x) expressed in terms of logarithms of
ratios of linear functions, in order to measure the progress of the algorithm.

It has the following properties:
1. Any desired amount of reduction in the value o f / ( x ) can be achieved by sufficient
reduction in the value of g(x).
2. g(x) is mapped into a function of the same.form by projective transformations
T(a, a0) described in 1.4.
3. Optimization of g(×) in each step can be done approximately by optimizing a
linear function (but a different linear function in different steps).

A NEW POLYNOMIAL-TIME A L G O R I T H M FOR LINEAR P R O G R A M M I N G 377

1.6. Complexity of the algorithm
The value of the objective function is reduced by a constant factor in O(n)
steps. As in the ellipsoid algorithm we define
L = log (1 + ID,..~!)+log (1 +~)
where Dm=x = max {Idet(X)lIX is a square submatrix of constraint matrix A}
: max {[c;I, Ib,[ i = l .... , n}.
Clearly, L is not greater than the length of the input.
During the course of the algorithm, we periodically check for optimality by
converting the interior point solution to an extreme point solution without degrading
the objective function value, and testing the extreme point for optimality. The inter-
val between two successive checks can be adjusted so that time spent between checks
is comparable to tinae spent in checking for optimality. Any two distinct values of
the objective function on the vertices of the polytope have to differ by at least 2 -kz
for some constant k [3]. Therefore, in at most O(nL) steps we reach a situation in
which any extreme point better than the current interior point has to be an exact
optimal solution.
In each step, we have to solve a linear system of equations which takes O(n 3)
arithmetic operations in the worst case. However, if we solve the equations at each
step by modifying the solution to the previous step, we can s a v e a factor of 1/n.
This method is described in Section 6.
For each arithmetic operation, we need a precision of O(L) bits. Note that this
much precision is required for any method that solves the equations A x = b if
det (A) :O(2L), e.g., the complexity of gaussian elimination is O(n3L In L lnln L),
although it is normally stated as O(n~). This factor L must appear in the conlplexity
of any algorithm for linear programming, since representing the output can require
O(L) bits/variable in the worst case.
The overall complexity of our algorithm is O(n3-SL"~In L Inln L) as compared
to O(n6L 2 In L lnln L) for the ellipsoid algorithm.

2. Main algorithm
2.1. Problem definition
In order to focus on the ideas which are new and the most important in our
algorithm, we make several simplifying assumptions. In Section 5 we will remove
all these restrictions by showing how a linear programming problem in standard
form can be transformed into our canonical form.

Canonical form of the problem
Input: ACZ '~×', c~Z".
Let g2={xlAx=0}, A={xlx=>0, ~ ' x i = l } and ao=e/n.
Assumptions: crx=>0 for every x~ ~2NA; ao is a feasible starting point, i.e. A a o - 0 ;
crao#O.
Output:either Xo~QNA such that crxo=0 or a proof that min{erxlxCflAA}>0,

378 N. KARMARKAR

2.2. Comments

1. Note that the feasible region is the intersection of an affine space with a
simplex rather than the positive orthant. Initially it takes one projective transfor-
mation to bring a linear programming problem to this form.
2. The linear system of equations defining f2 is homogeneous i.e., the right
hand side is zero. The projective transformation that transforms positive orthant
into simplex also transforms inhomogeneous system of equations into an homo-
geneous one.
3. The target minimum value of the objective function is zero. This will be
justified by combining primal and dual into a single problem.

2.3. Description of the Algorithm

The algorithm creates a sequence of points x (°), x (1), ..., x (~) by these steps:
Step 0. Initialize
x (°) : center of the simplex
Step 1. Compute the next point in the sequence
x(k+ 1) .~_ ~D(x(k)).

Step 2. Check for infeasibility
Step 3. Check for optimality
go to Step 1.
Now we describe the steps in more detail.
Step 1. The function b=q~(a) is defined by the following sequence of operations.
Let D = d i a g {al, a2 . . . . . a,} be the diagonal matrix whose fh diagonal
entry is a~.
1. Let B = ~
i.e., augment the matrix AD with a row of all l's. This guarantees that Ker B ~ 2;.
2. Compute the orthogonal projection of Dc into the null space of B.
cp = [ I - BT (BBr)-IB] Dc.
^ Cp
3. c - ~ 7 I , i.e. ~ is the unit vector in the direction of %.
4. b'=a0-~tr~, i.e. take a step of length ctr in the direction ~, where r is
the radius of the largest inscribed sphere
1
r=
t/n(n- 1)
and ~E(O, l) is a parameter which can be set equal to 1/4.

A NEW POLYNOMIAL-TIME ALGORITHM FOR LINEAR PROGRAMMING 379

5. Apply inverse projective transformation to b"

Db"
h~--~
erDb , •
Return b.
Step 2. Check for infeasibility.
We define a "potential" function by
cTx
f(x) = Z ln--
i Xt

We expect certain improvement 6 in the potential function at each step. The value
of 3 depends on the choice of parameter ~ in Step 1.4• For example if ~ - 1/4 then
6=1/8. If we don't observe the expected improvement i.e. if f(x(k+l))>f(x(k))--3
then we stop and conclude that the minimum value of the objective function must
be strictly positive. When the canonical form of the problem was obtained by trans-
formation on the standard linear program, then this situation corresponds to the
case that the original problem does not have a finite optimum i.e. it is either infea-
sible or unbounded•
Step 3. Check for optimality.
This check is carried out periodically• It involves going from the current
interior point to an extreme point without increasing the value of the objective
function and testing the extreme point for optimality• This is done only when the
time spent since the last check exceeds the time required for checking•

3. Underlying conceptual algorithm

In this Section we show how the computations in step 2 represent an underlying
conceptual sequence of operations. We will also introduce more notation and state
theorems about the performance of the algorithm. All proofs will be given in Sec-
tion 4. (A note on notation : In choosing variable names we use the same philosophy
as in the design of a large program: Some variables are local to a theorem while
others are global, some theorems are stated in terms of "formal" parameters and
in order to apply them to the algorithm, one has to substitute actual values for the
formal parameters e.g•, a result about one step of the algorilhm may be stated in
terms of input point a and output point b, and can be applied to the k th step of the
algorithm by substituting a = x_ (k) and b : x ( k + I) •)
•

3.1. The top level

Theorem 1. In O(n(q+log n)) steps the algorithm finds a feasible point x such that
cTx
_ _ :
-~ 2--q •
e Ta 0

We associate the objective function c r x with a "potential function" f ( x )

380 N. KARMARKAR

given by

f(x) y
7 ~ x; J"
Theorem 2. Either (i) f(xtk+l))~f(x(k))--6 or
(ii) the minimum value o f objective function is strictly positive, where 6 is a
constant and depends on the choice o f the value o f the parameter ~ o f Step 1, Sub-
step 4.
A particular choice that works: If ~ = 1/4, then 6~1/8.

3.2. The basic step

One step of the algorittun is of the form b=~0(a) which consists of a sequ-
ence of three steps:
1. Perform the projective transformation T(a, a0) of the simplex A that
maps input point a to thecenter a0.
2. Optimize (approximately) the transformed objective function over an
inscribed sphere to find a point b'.
3. Apply the inverse of T to point b' to obtain output b.
We describe each of the steps in detail.

3.3. Projective transformation T(a, ao)

Let a = (al, a2 . . . . , a,)
Let D = d i a g {at, a2 . . . . , a,} be the diagonal matrix with diagonal entries
al, a2 . . . . . a,. Then T(a, ao) is given by
D-1X
x'- where e=(1,1,.. 1).
eTD -ix "'

Its inverse is givenby
Dx"
X -- eTDx .

L e t f ' be the transformed potential function defined by
f ( x ) = f ' (T(x)).
Then
c'Ty
f'(Y) = Z In ~ - ~ Inaj
J Yj J
where c ' = Dc. Let f2' be the transformed affine space f2.
A x = O ¢~, AD x" =0.

Thus 12' is the null-space of AD.

A NEW POLYNOMIAL-TIME A L G O R I T H M FOR LINEAR P R O G R A M M I N G 381

3.4. Optimization over a ball

Let r be the radius of the largest ball with center a0 that can be inscribed in
the simplex A. Then
1
t-
en(n-i)

We optimize over a smaller ball B(ao, c~r) 0 < ~ < 1 for two reasons:
1. It allows optimization of f '(y) to be approximated very closely by opti-
mization of a linear function.
2. If we wish to perform arithmetic operations approximately rather than
by exact rational arithmetic, it gives us a margin to absorb round-off errors without
going outside the simplex.
One choice of the value of ~ that works is ~ = 1/4 and corresponds to 3 > 1/8
in Theorem 2. We restrict the affine space ~ ' = {Yl ADy=0} by the equation ~'y~=
= 1. Let O" be the resulting at'fine space. Let B be the matrix obtained by adding
a row of all l's to AD. Then any displacement in ~ " is in the null space of B,
i.e., for u E ~ " ~ ? ' = u + K e r B .
We are interested in optimizing f ' ( y ) over B(a0, ~r)NO". First, we prove
the existence of a point that achieves a constant reduction in the potential function.
Theorem3. There exists a point b'EB(a0, c o ' ) n ~ " such that f ' ( b ' ) = < f ' ( a 0 ) - 3 ,
where 3 is a constant.
Then we prove that minimization of f '(x) can be approximated by minimi-
zation of the linear function c'rx.
Theorem 4. Let b" be the point that minimizes c ' r x over B(ao, e r ) N f U . Then
f ' ( b ' ) ~ = f ' ( a o ) - 6 where 3 is a positive constant depending on ~. For c~=1/4 we
may take 6 = 1/8.

Finally we describe an algorithm to minimize c'rx, over B(a0, e'r).

Algorithm A.
1. Project c' orthogonally hIto the null space o f B:

c, = [ I - Br(BB r) -1B]c'.
2. Normalize cv:
Cp

: lop---L
3. Take a step o f length ~tr along the direction - c v

h p _~ a o - ~ r c p ,

Theorem 5. The point b" returned by algorithm A minimizes c ' r x over B(a0, ~r)OfY'.

382 N. KARMARKAR

4. Analysis of the algorithm

Proof of Theorem 5. Let z be any point such that zEB(ao, ~r)7) Q"

b', z E Q" ~ B (b" - z) = 0

=~ B T ( B B r ) - a B ( b " - Z) = 0

= (c' -- ce)T(b" -- Z) = 0

=~ c ' T ( b ' - - Z ) : c ~ ( b ' - - z )

ceT(b' - Z) : IceTIcer[ao-- o~r~e - z]

= [cpr[{~.er(ao-- z) -- ~r}

~-er(ao-z) --<--[~:pllao--Z[ --<--~r since zEB(ao, ~r)

c~(b'-z) <_- 0

c ' r ( b " - z) <= 0

c ' T h ' ~ c tT z, for all zEB(ao, a r ) N Q " , l

Proof of Theorem 3. Let x* be the point in ( 2 " N S where the m i n i m u m of c t T x
is achieved. Clearly x* ~[B(ao, ~r).
Let b" be the point where the line segment aox* intersects the b o u n d a r y
of the sphere B(ao,~r). h ' = ( 1 - 2 ) a o + 2 x * for some 2E[0, 1]. By definition,
b ' E B ( a o , ~ r ) and ao, x*EQ" imply b'EgT'.
Hence h'EB(ao, ~r)Og2"

c ' r b ' = (1 - 2 ) c ' r a o + A c ' r x * = (1--2)c'Tao

cPTao ]
c'rb ' 1- 2
and
b~ = (1 - 5~)aoj + 2x*.
Hence
c ' r ao b~ (1 - )0 a. r + 2x,*-
f'(ao)-f'(b') = Z In = ~ In
c ' r b ' aoj i (1 - 2)aoj

=Zln[14 2 x~.]
j l--2 aojJ"


Now we use the following inequality: If P~_->O then /-/(1 +P3=>I + z ~ P t . There-
i
fore, z~ In (1 + P 3 =>ln (1 + z~ P~).
i

f f ( a o ) - f ' ( b ' ) => In lq 1--2 aoi • x = In 1+

h" = ( 1 - 2 ) a o + 2 X *
b'-ao = 2(x*-ao)
c{r = ]b'-aol = 2[x*-aol -<- XR
where R is the radius of the smallest circumscribing sphere and R/r=n- 1.
r
)~ -->_cx -- - -c~
R n--1
2n n~
14 1~-2 >- 1 ' na/(n -1) = 12r - - >- 1 +
-- -I- 1 - c ~ / ( n - 1) n-l--~
f ' ( a o ) - f ' ( b ' ) ->_ ln(1 + c 0.
Taking 6 = I n ( 1 +c0, f'(h')<=f'(ao)-3. I
Proof of Theorem 4. First we prove a few lemmas.
X2
Lemma 4.1. If Ix1-<__13<1 then [ln(l+x)-xl~--
2(1 - / ~ )
Proof. A simple exercise in calculus. I

Lemma 4.2. Let fl=ct n Thenforany xEB(ao,~r), In "'J ~ f12
• aoj 2 (1 - fl) "
Proof. xEB(ao, ~r) implies ]x-ao/2_-<~2r~, Therefore

y
7" ~ aoj / (l/n) 2 n(n--1)
Xj -- aoj =
i-<13 for j 1. . . . , n
I aoj
Iln(iNxj--aollaoj = 2(1-I 13) ~ aoi )' by Lemma 4.1
ao.~ J
xi--ao,]< txi--ao,l'
Zln = Z l n [ l + xj-a°J]
xj
j j aoj aoj J
~[lnXJ x,--aoy<= ------'-szl (xi--aoji' - fl~
• ao~. aoj 2(I--13__#~ aoj. / 2(1--13)
but
xy--aoj_ 1 [~xj--~aoj]=-~ol/[1--1]=O
J aoj -- aoj " j

~- 2(1-13)" I

384 N. KARMARKAR

Now we prove Theorem 4.
Definef(x)=n In c'Tx Let bm be the point in B (ao, czr)A f2" where f(x)
CpT ao •
achieves its minimum value.
f(ao) - f ( b ' ) = f(ao) -fib.,) +f(h,,,) - f ( b ' )
= [f(a0)-f(b,~)]
+ [f(b,.)-(f(ao) +f(b,.))]
-[f(b')-(f(ao) +f(b'))]
+ [f(b,,) -f(b')]
By Theorem 3,
(1) f(ao)-f(b,,,) --> In (I +a).
For xEB(ao, :¢r)(q f2",
f(x) - (f(ao) +f(x)) = Z In - - - -
cPTx In e ' r a ° - n ~ ' In e'r--''~x
j Xj J ao j cp T ao

=-Z In xj
j aoj

I ~ In xj f12
(2~ If(x)- (f(a°)+f(x))l = aoj <= 2(i= D
by Lemwa 4.2.
Since f(x) depends on c'rx in a monotonically increasing manner, fix)
aud e'rx achieve their minimum values over B(a0, ~r) (3 Q" at the same point, viz b',
(3) f(b,,,) ~ f(b').
By (1), (2) and (3),
fl ~ ~2 ~2 H 2
f(ao) - f ( b ' ) >= In (1 + 00
(l-fl) - 2 n ( n _ 1) [1 _ ~/n__Z_~_]
Define
0¢2 ~2//
6=a
2 (n-l)
[
1--c~
V ]
f ( a o ) - f ( b ' ) => 6.
Observe that
~2
cx~ l
1. .lira 3(n)=~z-~-
... • If ~=~-, then lira 6 (n) => I
1-c¢
>1
2. If" n=>4, c~=1 then ~ = T o " |

Proof of Theorem 2. By Theorem 4, f " (b' ) = f '(a o)-6. Applying the inverse trans-
<
form T-l(a, ao) and substituting x(aJ=T-l(ao), we obtain x(k+l)=T-l(b'),
f(xC~+.)~f(x%-6. |


Proof of Theorem 1. For each k we have
f ( x (k+x)) ~_ f ( x ok))-- 6
f ( x (k)) _~ f ( x C°))-- k6
crx(k) In era° -
T xF" J aoj

In cr x(k)
n ~_ 2 In (x5 k)) -- ~ In a0j - k3.
crao j J

c r x tk) k
But x~k)~l, aoj= 1 , so lnc---~ao<-ln(n)--~'n Therefore there exists
a constant k" such that after m=k'(n(q+ln (n))) steps.
cTx
Crao =< 2 - L II

5. Transformation of a linear program to canonical form

Consider a linear programming problem in the standard form:
minimize cTx
~Ax_~ b
subject to t x _-> O.
There are many ways by which this problem can be converted to our canonical
form. We describe one such method.
Step 1. Combine primal and dual problems.
Ax=>h
AZu =< c
eTx--bTu = 0

x_~O u~O.
By duality theory, this combined problem is feasible if and only the original problem
has finite optimum solution.
Step 2. bltroduee slack variables.
Ax-y =b
Aru+v = c
crx-bru = 0
x_~O u~O y=>O v~O.

386 N. KARMARKAR

Step 3. Introduce an artificial variable to create an interior starthlg pohTt. Let Xo, Yo,
uo, vo be m i c t l y interior points in the positive orthant
minimize ~.
Ax-y+(b-Axo+Yo)• = b
JAru+v+(c-Aruo)~ = c
subject to ] e r x _ b r u + ( _ c r x o + b r v o ) X = 0
Ix_-->0 y=>0, u=>0, v:>0, )~=>0.
Note that x - x 0 , Y=Yo, u = u o , V=Vo, ;~=1 is a strictly interior feasible solution
which we can take as a starting point. The minimum value of X is zero if and only
if the problem in step 2 is feasible.
Step 3. Change o f Notation. For convenience of description we rewrite the problem
in Step 3 as
minimize c r x , c, xER"
subject to Ax = b, x => 0
x = a is a known strictly interior starting point, and the target value of the objective
function we are interested in is zero.
Step 4. A projective transformation o f the positive orthant into a simplex. Consider
the transformation x ' = T ( x ) where xER", x ' E R "+1 are defined by

xl/ai i = 1, 2, ..., n
x" -- X (xflaj) + 1
J

Xn+ 1 ~- 1 - X i.
i=1

Let P+ denote the positive orthant
P+ = {xER"lx =~ 0}
and A denote the simplex
n+I
A={xER"+IIx->0, z~x,=l}- i=l

The transformation T(x) has the following properties:
I. It maps P+ onto A.
2. It is one-to-one and the inverse transformation is given bv

a;x~ i = 1, 2 . . . . . n.
"k"i _ _ Xn+l

3. The inaage of point a is the center of the simplex.
4. Let Ai denote ith column of A. I f

~ Aix i = b
i=1


then
.~, Aialx; = b
Xn+l
or

Aiaix; - b x , + t 0.
i=l

We define the columns o f a matrix A" by these equations:
A ~ = a i A i, i= 1,...,n
A~+I : - b .
Then
n+l
Z A x; = O.
i=1

Let I2" denote this affine subspace, i.e. f2'={x'ER"+IIA'x'=0}. Thus we
get a system of homogeneous equations in the transformed space.
5. Since we are interested in target value of zero, we define Z to be the afflne
subspace corresponding to the zero set of the objective function in the original space
Z = {xER"IcTx = 0}.
Substituting
a i x~
Xi = . ,

Xn+ l
we get
n+l
ciaixi•
Y I
- O.
l=l Xn+l
Define c'ER "+1 by
c;=aici i=1,2 .... ,n
c L 1 = 0.
Then c T x = 0 implies c ' T x ' = 0 .
Therefore the image of Z is given by Z ' = {x'ER"+l[c'rx'=0}.
The transformed problem is
minimize crTx ~

A'x' = 0
subject to n+l
×' > 0, Z x'i = 1
i=l

The center of the simplex is a feasible starting point. Thus we get a problem
in the canonical form.
Application of our main algorithm to this problem either gives a solution
x' such that c ' T x ' = 0 or we conclude that minimum value of c ' r x ' is strictly posi-
tive. In the first case, the inverse image of x' gives an optimal solution to the original
problem. In the latter case we conclude lhat the original problem does not have
a finite optimum, i.e. it is either infeasible or unbounded.

9*

388 N. KARMARKAR

6. Incremental computation of inverse

6.1. A modification to the main algorithm

The computational effort in each step of the main algorithm is dominated
by the computation of cv:
cv = [I--Br(BBr)-~B]Dc.

We have to find the inverse of (BB r) or solve a linear system of equations
of the form (BBr)u=v.

(1)

The only quantity that changes from step to step is the diagonal matrix D,
since D}k) =x} k). In order to take advantage of the computations done in previous
steps, we exploit the following facts about matrix inserve:
Let D, D' be n × n diagonal matrices.
1. If D and D" are "close" in some suitable norm, the inverse of AD'2A r
can be used in place of the inverse of AD2A r.
2. If D and D' differ in only one entry, the inverse of AD'2A r can be com-
puted in O(n') arithmetic operations, given the inverse of AD2A r.
(Note: Instead of finding the inverse, one could also devise an algorithm
based on other techniques such as LU decomposition. In practice, it may be better
to update the LU decomposition directly, but this will not change the theoretical
worst-case complexity of the algorithm.)
We define a diagonal matrix D "ok), a "working approximation" to D t*) in
step k, such that
1 [D','k']
(2) ~I_D~) I -<2 for i-- 1,2,...,n.

We use (A(D'(k))2Ar) -1 in place of (A(D(k))2Ar) -1. Initially, D'(°)=D(°)= 1 I.
n
We update D' by the following strategy:
1. D'(k+lJ=a(k)D "Ck), where a (k) is a scaling factor whose significance will
be explained in Section 6.3. For the purpose of this section take a (k)= 1.
2. For i = l t o n d o
if
'

then let D~[k+I)=D~(~+I); and update (A(D'(k+x))~Ar) -~ by rank-one modification.
Thus, this strategy maintains the property in equation (2). Now we describe
the rank-one modification procedure. Suppose a symmetric square matrix M is

A N E W P O L Y N O M I A L - T I M E A L G O R I T H M FOR L I N E A R P R O G R A M M I N G 389

modified by a rank-one matrix expressed as the outer product of two vectors u and
v. Then its inverse can be modified by the equation
(3) (M+uvr)-i - - 1 (M-Xu)(M-Xv) r
=M -- l + u r M - l v "
Given M -x, computation of (M+uvr) -x can be done in O(n 2) steps. In order to
apply equation (3), note that if D" and D" differ only in the ith entry then
(4) AD'2Ar=AD"ZAr+[D~ -D~Z] alaT
where a~ is the ith column of A. This is clearly a rank-one modification. If D' and
D" differ in l entries we can perform I successive rank-one updates in O(n21) time
to obtain the new inverse. All other operations required for computing cv, the
orthogonal projection of c, involve either multiplication of a vector by a matrix,
or linear combination of two vectors, hence can be performed in O(n 2) time.

6.2. Equivalence of the modified and original problems

In this Section we analyze the effect of the replacement of D (k) by D '(k).
Consider the following generalization of the optimization problem over an inscribed
sphere in the transformed space:
Problem P" :
Minimize c'rx }
(5) subject to Bx = 0 .
and x r Q x --<_~'r
(Taking Q = I and ~'=ct corresponds to the original problem.)
We are going to show that replacing D ~k) by D "tk~ corresponds to solving
the problem P , for appropriate choice of Q. Let h(x)=xrQx. Its gradient is given
by Vh(x)=2Qx.
At the optimal solution x to the problem P',
(6) c' = BTZ+# .2Qx
where the vector 7, and the scalar # are undetermined Lagrange multipliers
BQ-lc" = BQ-aBr~,+21~Bx = BQ-1BrZ
(7) k = (BQ-1BT)-IBQ-Xe ".
Substituting for k in equation (6), x, which we now denote by cv, is given (up to
a scale factor) by
(8) x = cv = [I-Q-XBT(BQ-1Br)-IB]Q-lc ".

This requires the inverse of BQ-IB r. Substituting B = 7 [A° 1 we get

[ADQ-1DA r ADQ-le]
(9) BQ-XB r = [ (ADO_Xe)r erQ_Xe | .

390 N. KARMARKAR

It is enough to know the inverse of the submatrix ADQ-aDA r, because then
(BQ-~Br) -1 can be computed in O(n 2) additional steps by the formula
(10) [M ~]-~_ 1 [(c-arM-xa)M-~+(M-~a)(M-la) T -M-Xa]
aT c-arM-la - (M-la) r 1

where M, a and c are m × m , m × l , 1)<I matrices, respectively. But we know
(AD'2Ar) -~ where D ' = D .E, and E is a diagonal "error matrix" such that E~E
E[1/2, 2]. Setting Q = E -2 we get

AD,2A r = ADEeDAr = ADQ-1DA r.

In the modified algorithm we keep (AD'eAr) -~ around, update it whenever
any entry in D' changes and use it in equation (8) to compute cp and thereby solve
the modified problem P'.
Now we relate the solution of problem P' to the main optimization problem.
Since Qu=Eff2E[]/2, 2]

(11) L x T x <=
2 xrQx ~ 2xrx.
Hence
B ao, ~--r ~ {xlxrQx ~ c~'r} ~ B(ao, 2~'r).

Taking ~'=~/2, and taking intersection with f2"

(12) B(a0, ~r/4)0 f2" c= {xlxTQx ~ ~'r, Bx = O} c B(ao, ar)N g2".

Because of the first inclusion,

(13) rain {f'(x)[xrQx ~ c(r, Bx = 0}
rain {f'(x)lxEB(a0, c~r/4)f-) f2"} ~f'(a,,) - l n (1 +c~/4).

The last inequality follows from Theorem 3. Thus for the modified algorithm, the
claim of Theorem 3 is modified as follows

(]4) f ' ( b ' ) <=f'(a0)-ln (I +a/4).

Because of the second inclusion, Lemma 4.2 continues to be valid and we can appro-
ximate minimization of f "(x) by minimization of a linear function. The claim (as in
Theorem 2) about one step of the algorithm f ( x (k+t))~f(x(k))-6 also remains
valid where 6 is redefined as

(15) 6= In(l+e/4) 2(1--fl)"

This affects the number of steps by only a constant factor and the algorithm still
works correctly.


6.3. Performance of the modified algorithm

In this Subsection we show that the total number of rank-one updating ope-
rations in m steps of the modified algorithm is O(mJ/n). Since each rank-one modi-
fication requires 0@21 arithmetic operations, the average work per step is O(n 2"5)
as compared to O(n 3) in the simpler form of the algorithm.
In each step lh'-a0]_=c~r. Substituting h'=T(x(k+t)); a 0 = n - l e and T ( x ) =
D-1X I
- - r - -

eTD-lx " 1/n(n - 1)
[ x}k+i)lx}k) 1.] 2 < ~_2
n] = n ( n - l ) '
j
Let
a(k) 1 j~ (k+l) .(k)
(16) : -- Xj /Xj ,
El
Then
f i]2<=
Let
..(k+l)
(17/ A!k) - .tl
X!k) ~(k) •

Therefore • [A~ k ) - 1]2_-<fla. Recall that D (k) = Diag {x(~'), x~ k). . . . . x(,k)}, D(k+l)=
i
Diag {x}k+l), ,,(~,+11 x,(k+l)}. D"(k) is updated in two stages.
First we scale D "(k) by a factor a (k) defined in equation (16)
(18) D ' ( k + l ) _-- a(g)D'(k).

Then for each entry i such that
D,(k+ 11]~

we r e s e t D ; I ( g + I ) = D ! / k + l ) .
Define discrepancy between D uo and D '(k) as

(19) 5}k) = In f D;i('' "l
t D}p 1"

Just after an updating operation for index i,
6} k) = 0.

Just before an updating operation for index i,

392 N. KA.RMARKAR

Between two successive updates, say at steps kx and k~,

~}~')-,~1 ~'> = " Z ' [z}~+'- ,~?>1.
k:k I

Let A~ k) be the change in the discrepancy at step k.

(20) z16}~) = 6}~+x)_ 6}~).

Then

(21) k•l
k~k I

Now we analyze the change in discrepancy in step k, assuming that no updating
operation was performed for index i.

D{,(k+l> D}p> a(~)x}k>
Af}k) = &}k+x)_6[~) = In --.ntk+l) D~ff) = In ~ = - l n A [ k)

(22) IzJ6}~)I = IlnzJ}~>
1.
Let nL= the number of updating operations corresponding to index i in m
n

steps of the algorithm and N = Z nL, be the total number of updating operations
1=1
in m steps.
By equations (20) and (21),

k=l

(23) In ~ N ~ Z ~ Iln Atk)[.
i=1 k=l

We use equation (17) to bound the R.H.S. of equation (23)

Z [A?>- 1]' ~_ t~' =, IA?) - 11 ~-
i

=~ [In A}a - (A/k>- 1)] ~_ (A[k)-- 1)3
2-0 ---~ by Lemma 4.1

Z [A[~)- 1]' _ p' =, Z [a~~>- if ~_ ¢n~.
l i


Hence
Iln A}~)I = Z ]lnA[ k) - (A}~)- 1)-4-(A!k)-- 1)1
i=l i=l

-~ ~ [lnA}~)-(A}*'-I)I÷ ~ [A}~)- ll
i=i i=1

< y [Aif)- 1]2

2 (l

(24) ~ IlnA}k)] -< m "2(i-- B) ~-gn-/~ = O (m Vn-).
i=I k:l

From equations (22) and (24)
(25) N : O(ml/n).

7. Concluding remarks
7.1. Global analysis of optimization algorithms
While theoretical methods for analyzing local convergence of non-linear
programming and other geometric algorithms are well-developed the state-of-the-art
of global convergence analysis is rather unsatisfactory. The algorithmic and analy-
tical techniques introduced in this paper may turn out to be valuable in designing
geometric algorithms with provably good global convergence properties. Our method
can be thought of as a steepest descent method with respect to a particular metric
space over a simplex defined in terms of "cross-ratio", a projective invariant. The
global nature of our result was made possible because any (strictly interior) point in
the feasible region can be mapped to any other such point by a transformation that
preserves affine spaces as well as the metric. This metric can be easily generalized to
arbitrary convex sets with "well-behaved" boundary and to intersections of such con-
vex sets. This metric effectively transforms the feasible region so that the boundary of
the region is at infinite distance from the interior points. Furthermore, this tians-
formation is independent of the objective function being optimized. Contrast this
with the penalty function methods which require an ad hoc mixture of objective
function and penalty function. It is not clear a priori in what proportion the two
functions should be mixed and the right proportion depends on both the objective
function and the feasible region.

7.2. Comments on the factor "L" in running time
Since representing the output of a linear programming problem requires
O(L) bits per variable in the worst case, the factor L must appear in the worst-case
complexity of any algorithm for linear programming. In practice L is much smaller
than n. Therefore, whenever it is possible to replace a factor of n by L, one should
always do so if one is interested in an algorithm which is efficient in practice.

394 N. KARMARKAR

7.3. Performance in practice

Each step of the algorithm requires optimization of a linear function over an
ellipsoid or equivalently, solution of a system of linear equations of the type
(ADAr)x=b, where A is a fixed matrix and D is a diagonal matrix with positive
entries which changes by small amount from step to step. We devise a method based
on successive rank-one modifications and prove a worst-case bound of O(n 2"5)
arithmetic operations per step.
In practice, the matrices encountered are sparse, hence more efficient methods
for solving the above problem are possible. Another feature of the algorithm which
can lead to computational saving is that it is not necessary to find the exact solution
to the optimization problem stated above. Here it is useful to distinguish between
two types of approximate solutions. Let x0 be the exact solution and x an approxi-
mate solution and let cTx be the objective function. A strong approximation (or
approximation in the solution space) requires that x be close to x0 in some suitable
norm. A weak approximation (or approximation in objective function space)requires
that c r x be close to crx0. A weak approximation is sufficient for our method, and
is easier to achieve numerically. The optimization problem in each step is a good
candidate for applying iterative methods to obtain approximate solutions. The
worst-case bound is based on finding an exact solution.
The number of steps of the algorithm depends on the "R/r" ratio, i.e. the
ratio of radius of the sphere circumscribing the polytope to the radius of the inscribed
sphere. We prove an O(n) upper bound on this ratio. However it is likely to be smaller
than n in typical problems of praclieal interest and also in the worst-case if the cir-
cumscribing sphere is chosen to include only that portion of the polytope having
values of objective function better than the current value, and after computing lhe
direction of one step (as described in section 2) its length is chosen to be ~r" ralher
than ar where r' is the largest possible step length without going outside the feasible
region. The possibility of improvement in R/r ratio gives rise to a number of research
problems for further investigation, such as probabilistic analysis of the problem,
experimental measurements on problems arising in practical applications, analysis
of special cases of practical interest such as "box" type of constraints and finally,
the most difficult oi all, obtaining a better worst-case bound.

7.4. Numerical stability and round-off errors

In the ellipsoid algorithm, the round-off errors in numerical computation
accumulate from step to step. The amount of precision required in arithmetic ope-
rations grows with the number of steps. In our algorithm, errors do not accumulate.
On the contrary, the algorithm is self-correcting in the sense that the round-off
error made in one step gets compensated for by future steps. If we allow 1% error
in computation of the vector x (~ in each step, we can compensate for the errors
by running the algorithm 2 % more steps.


7.5. Multiple columns and parallel computation

In the simplex method, the current solution is modified by introducing a non-
zero coefficient for one of the columns in the constraint matrix. Our method allows
the current solution to be modified by introducing several columns at once. In the
algorithm described in this paper, all columns are "active", but a variation in which
a suitably chosen subset of columns is active at one time is possible. It seems that
on a highly parallel architecture a method that uses many columns at once may be
preferable to the simplex method.

7.6, Other areas for further research

Since many polynomial-time solvable combinatorial problems are special
cases o f linear programming, this work raises the possibility o f applying the new
techniques to such special cases. Problems solved by the column generation method
require a special variant of the algorithm.
Acknowledgement. The author wishes to thank Arjen K. Lenstra, Mark Manasse,
Joan Plumstead, Alon ltai, L~iszl6 Babai, Mike Garey and Ron Graham for pointing
out errors and making suggestions for improvement.

References

[1] H. S, M. COXETER,Introduction to Geometry, Wiley (1961).
[2] G. B. DANTZIG, Linear Programming and Extensions, Princeton University Press, Princeton,
NJ (t963).
[3] M. GR6rSCHEL, L. LOVASZand A. ScHmJvzg, The Ellipsoid Method and its Consequences in
Combinatorial Optimization, Combinatorica 1 (1981), 169--197.
[4] L. G. KHACHIVAr~,A polynomial Algorithm in Linear Programming, Doklady Akademii Nauk
SSSR 244:S (1979), 1093--1096, translated in Sov&t Mathematics Doklady 20:1 (1979),
191--194.
[5] V. KLEEand G. L. MINTV,How good is the simplex algorithm? in Inequalities 111, (ed. O. Shisha)
Academic Press, New York, 1972, 159-179.
[6] O. VEBLENand J. W. YOUNG,Projective Geomeo:v, I--2, Blaisdell, New York, (1938).
[7] R. J. WAL~Ea, Algebraic Curves, Princeton University Press (1950).

N. Karmarkar
A T & T Bell Laboratories
Murray Hill, N J 07974
U.S.A.

A New Polynomial-Time Algorithm for Linear Programming

More Related Content

What's hot (20)

Similar to A New Polynomial-Time Algorithm for Linear Programming (20)

More from SSA KPI (20)

Recently uploaded (20)

A New Polynomial-Time Algorithm for Linear Programming