Professional Documents
Culture Documents
net/publication/228572901
The central path visits all the vertices of the Klee–Minty cube
CITATIONS READS
20 139
4 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Tamás Terlaky on 02 June 2014.
Abstract
The Klee-Minty cube is a well-known worst case example for which the simplex method
takes an exponential number of iterations as the algorithm visits all the 2n vertices of the
n-dimensional cube. While such behavior is excluded by polynomial interior point meth-
ods, we show that, by adding an exponential number of redundant inequalities, the central
path can be bent along the edges of the Klee-Minty cube. More precisely, for an arbitrarily
small δ, the central path takes 2n − 2 turns as it passes through the δ-neighborhood of all
the vertices of the Klee-Minty cube in the same order as the simplex method does.
1 Introduction
While the simplex method, introduced by Dantzig [1] in 1947, works very well in practice
for linear optimization problems, Klee and Minty [5] gave an example in 1972 for which the
simplex method takes an exponential number of iterations. More precisely, they considered a
maximization problem over an n-dimensional squashed cube and proved that a variant of the
simplex method visits all its 2n vertices; that is, the time complexity is not polynomial for the
worst case, as 2n −1 iterations are necessary for this n-dimensional linear optimization problem.
The pivot rule used in the Klee-Minty example was the most negative reduced cost but variants
of the Klee-Minty n-cube showing an exponential running time exist for most pivot rules; see [9]
and the references therein. The Klee-Minty worst-case example partially stimulated the search
for a polynomial algorithm and, in 1979, Khachiyan’s [4] ellipsoid method proved that linear
programming is indeed polynomially solvable. In 1984, Karmarkar [3] proposed a more efficient
polynomial algorithm that sparked the research on polynomial interior point methods. In short,
while the simplex method goes along the edges of the polyhedron corresponding to the feasible
region, interior point methods pass through the interior of this polyhedron. Starting at the
analytic center, most interior point methods follow the so-called central path and converge to
the analytic center of the optimal face; see e.g., [7, 12, 11].
1
2 The central path visits all the vertices of the Klee-Minty cube
In this paper, following the Klee-Minty approach, we show that, by carefully adding an
exponential number of redundant constraints to the Klee-Minty n-cube, the central path can
be bent along its edges. In other words, we give an example where, for an arbitrarily small
δ, starting from the δ-neighborhood of a vertex adjacent to the optimal solution, the central
path takes 2n − 2 turns as, before converging to the optimal solution, it passes through the
δ-neighborhood of all the vertices of the Klee-Minty cube in the same order as the simplex
method does.
Before stating the main result in Section 2 and giving its proof in Section 3, we illustrate
the bending of the central path in the 2 and 3 dimensional cases. Fig. 1 and Fig. 2 show the
trajectory of the central path starting from the highest vertex and converging to the origin
after visiting each vertex of the Klee-Minty cube. The redundant constraints correspond
to hyperplanes parallel to the facets of the cube containing the origin. More precisely, in
dimension 2, the redundant inequality 16 + x1 ≥ 0 is added 15 360 times and the redundant
inequality 16 + x2 − 14 x1 ≥ 0 is added 40 960 times. Starting from the highest vertex and with
δ = 0.1, the central path visits the δ-neighborhood of each vertex of the Klee-Minty cube in
the same order as the simplex algorithm does before converging to the optimal solution; that
is, the origin. In dimension 3, the redundant inequality 48 + x1 ≥ 0 (resp. 48 + x2 − 41 x1 ≥ 0
and 48 + x3 − 14 x2 ≥ 0) is added 161 280 (resp. 552 960, and 1 474 560 times).
v {2}
v {1,2}
χh
{1}
χh (v2 )
v {1}
v∅
Figure 1: Central path nearing all the vertices of the Klee-Minty 2-cube.
v {3}
v {2,3}
χh v {1,3}
v {1,2,3}
{1}
χh (v3 ) v {2}
v∅ v {1,2}
v {1}
Figure 2: Central path nearing all the vertices of the Klee-Minty 3-cube.
Antoine Deza, Eissa Nematollahi, Reza Peyghami and Tamás Terlaky 3
We consider the following Klee-Minty variant where ε is a small positive factor by which the
unit cube [0, 1]n is squashed.
min xn
subject to 0 ≤ x1 ≤ 1
ε xk−1 ≤ xk ≤ 1 − ε xk−1 for k = 2, . . . , n.
We denote this linear optimization problem by KM. This minimization problem has 2n
constraints, n variables and the feasible region is an n-dimensional cube denoted by C. Some
variants of the simplex method take 2n − 1 iterations to solve KM as they visit all the vertices
ordered by the decreasing value of the last coordinate xn starting from v {n} = (0, . . . , 0, 1) till
the optimal value x∗n = 0 is reached at the origin v ∅ . If an interior point method is used to
solve KM, the central path starts from the analytic center χ of C and converges to the origin
as it is shown in Fig. 3.
v {2}
v {1,2}
χ
v {1}
v∅
While adding a set h of redundant inequalities does not change the feasible region of
KM, the analytic center χh and the central path are affected by the addition of redundant
constraints. We consider redundant inequalities induced by hyperplanes parallel to the n facets
of C containing the origin. To ease the analysis, we consider that all redundant hyperplanes
are put at the same distance d to the corresponding parallel facet of C. The constraint parallel
to H1 : x1 = 0 is added h1 times and the constraint parallel to Hk : xk = εxk−1 is added
hk times for k = 2, . . . , n. By abuse of notation, the set h is denoted by the integer-vector
h = (h1 , . . . , hn ). With these notations, the redundant linear optimization problem KMh is
defined by
4 The central path visits all the vertices of the Klee-Minty cube
min xn
subject to 0 ≤ x1 ≤ 1
ε xk−1 ≤ xk ≤ 1 − ε xk−1 for k = 2, . . . , n
0 ≤ d + x1 repeated h1 times
ε x 1 ≤ d + x2 repeated h2 times
.. ..
. .
ε xn−1 ≤ d + xn repeated hn times.
To give a flavor of the main result, we first present Lemma 2.1 stating that, by adding εd+1
n−1 δ
times the redundant inequality εxn−1 ≤ d + xn to the original KM formulation, the analytic
center χh can be pushed arbitrarily close to the vertex v {n} = (0, . . . , 0, 1). To warranty,
without loss of generality, that h is integer valued, we assume that both ε−1 and δ−1 are
positive integers.
While Lemma 2.1 sets the starting point of the central path in the δ-neighborhood of v {n} ,
Proposition 2.2 states that, for a careful choice of d and h, the central path of the cube takes
2n − 2 turns before converging to the origin as it passes through the δ-neighborhood of all the
2n vertices of the Klee-Minty n-cube.
n
2 −1
Proposition 2.2. Given ε ≤ 14 , δ < εn−1 , choose integer d ≥ n2n+1 and h = 4nd δ ( ε ,...,
n
2 −2 k−1 n−1
εk
, . . . , 2 εn ), then for each vertex v S of the Klee-Minty n-cube, there is a point χh (vnS )
of the central path satisfying |χh (vnS ) − v S |∞ ≤ δ.
s 1 = x1
sk = xk − εxk−1 for k = 2, . . . , n
s̄1 = 1 − x1
s̄k = 1 − εxk−1 − xk for k = 2, . . . , n
s̃1 = d + s1 repeated h1 times
.. ..
. .
s̃n = d + sn repeated hn times.
Antoine Deza, Eissa Nematollahi, Reza Peyghami and Tamás Terlaky 5
The optimality conditions (the gradient is equal to zero at optimality) for this concave maxi-
mization problem give
1 ε 1 ε hk hk+1 ε
σkh
− h
σk+1
− σ̄kh
− h
σ̄k+1
+ σ̃kh h −
σ̃k+1
=0 for k = 1, . . . , n − 1
1 1 hn
h − σ̄ h + σ̃ h =0 (1)
σn n n
σkh > 0, σ̄kh > 0, σ̃kh >0 for k = 1, . . . , n,
where
σ1h = ξ1h
h
σkh = ξkh − εξk−1 for k = 2, . . . , n
σ̄1h = 1 − ξ1h
σ̄kh = 1 − εξk−1
h − ξkh for k = 2, . . . , n
σ̃kh =d+ σkh for k = 1, . . . , n.
The following lemma states that, for hn large enough relatively to the other hk values, the
analytic center χh is pushed to the neighborhood of the vertex v {n} = (0, . . . , 0, 1).
1
Lemma 3.1. Given δ ≤ ε ≤ 4 and h1 , . . . , hn−1 , we have
h {n} d+1 1 1 hk 2
|χ − v |∞ ≤ δ for hn ≥ where = max + .
ηn ηn 1≤k≤n−1 2εn−k d δ
Proof. The analytic center χh = (ξ1h , . . . , ξnh ) is the solution of (1). Let us consider the n-th
equation of (1). Since σ̃nh ≤ d + 1, we have: hn ≤ d+1 σ̄h
. Thus, as hn ≥ d+1 h
ηn , we have σ̄n ≤ ηn ,
n
which implies σ̄nh ≤ 2δ . Let us then consider the (n − 1)-th equation. We have
1 1 ε ε hn−1 hn ε ε hn−1 hn ε
h
= h
+ h
+ h− h + h ≥ h− h + h.
σn−1 σ̄n−1 σn σ̄n σ̃n−1 σ̃n σ̄n σ̃n−1 σ̃n
1 2ε hn−1 2
h
≥ − ≥ ,
σn−1 ηn d δ
6 The central path visits all the vertices of the Klee-Minty cube
h
i.e., σn−1 ≤ 2δ . The first n − 1 equations of (1) can be rewritten as
!
1 hk 1 1 hk+1 ε
h
+ h = h +ε h
+ h + h for k = 1, . . . , n − 1.
σk σ̃k σ̄k σk+1 σ̃k+1 σ̄k+1
For k ≤ n − 2, forward substitutions for the k-th, (k + 1)-th,. . . (n − 1)-th equations give
n−1
1 hk 1 X εj−k εn−k hn εn−k εn−k
h
+ h
= h
+ 2 h
+ h + + h ,
σk σ̃k σ̄k σ̄ j σn σ̃nh σ̄n
j=k+1
which implies
1 εn−k hk hn εn−k
≥ − + .
σkh σ̄nh σ̃kh σ̃nh
Since σ̄nh ≤ ηn , σ̃kh ≥ d and σ̃nh ≤ d + 1, this implies
1 2εn−k hk 2
h
≥ − ≥ ,
σk ηn d δ
2εn−1
−1
d 0 0 0 . . . 0 d+1
1 −ε
d+1 d 0 0 ... 0 0
−1 2ε −ε 2
0 . . . 0 0
d d+1 d
−1 −ε 2ε 2 −ε 3
A = d
d(d+1) d+1 d ... 0 0 .
. .. .. .. .. .. ..
..
. . . . . .
−1 −ε −ε 2 −ε 3 −ε n−2
d d(d+1) d(d+1) d(d+1) . . . d 0
−1 −ε −ε 2 −ε 3 2ε n−2 −ε n−1
d d(d+1) d(d+1) d(d+1) . . . d+1 d
δε(d+1)
Proof. Multiplying both sides of Ah ≥ b by 4n , we have
d(1 − ε) ≥ 2n − 1 + ε
d(1 − ε) ≥ 2n − 2 + ε
d(1 − ε) ≥ 2n+1 − 5 + ε
d(1 − ε) ≥ 2n − 2k + k−1
n j−1 + ε
P
j=1 2 − 2 for k = 4, . . . , n,
Antoine Deza, Eissa Nematollahi, Reza Peyghami and Tamás Terlaky 7
Proof. The first 2 inequalities are direct reformulations of thePfirst 2 inequalities of the
system Ah ≥ b. For k = 2, . . . , n − 1, the inequality: ℓk εk−1 − k−1j=1 uj ε
j−1 ≥ 4n can be
δ
2hk εk−1 hk+1 εk h1 1 Pk−1 j−1 4n
rewritten as: d+1 − d − d − d(d+1) j=2 hj ε ≥ δ.
2k+1 n
ℓk εk−1 ≥ for k = 1, . . . , n − 1.
δ
4n
Proof. The proof is by induction on k. We have ℓ1 ≥ δ and the result follows by using
uk ≥ ℓk in Lemma 3.3.
1
Corollary 3.5. Given δ ≤ ε ≤ 4 and a positive integer h satisfying the last n − 1 inequalities
of Ah ≥ b, we have
2hn εn−1 h1 2
|χh − v {n} |∞ ≤ δ for ≥ + .
d+1 d δ
Proof. Corollary
3.4 implies ℓk ≥ 0 for k = 1, . . . , n − 1. Hence, in Lemma 3.1 we have
1 1 h1 2
ηn = 2εn−1 d + δ which gives the result.
The central path of KMh can be defined as the set of analytic centers χh (α) = (xh1 , . . . , xhn−1 , α)
of the intersection of the hyperplane Hα : xn = α with the feasible region of KMh where
0 ≤ α ≤ ξnh , see [7]. These intersections are called the α-level sets and χh (α) is the solution of
the following system
h ε
(
1
sh
− shε − s̄1h − s̄hε + hs̃hk − s̃k+1
h =0 for k = 1, . . . , n − 1
k k+1 k k+1 k k+1 (2)
shk > 0, s̄hk > 0, s̃hk > 0 for k = 1, . . . , n − 1,
8 The central path visits all the vertices of the Klee-Minty cube
where
sh1 = xh1
shk = xhk − εxhk−1 for k = 2, . . . , n − 1
shn = α − εxn−1
s̄h1 = 1 − xh1
s̄hk = 1 − εxhk−1 − xhk for k = 2, . . . , n − 1
s̄hn = 1−α− εxhn−1
s̃hk = d + shk for k = 1, . . . , n.
In the rest of the paper, we assume that δ ≤ εn−1 and that a positive integer h satisfying
Ah ≥ b is given. Corollary 3.5 implies that 1 − δ ≤ ξnh and therefore we can consider the
α-level set for α ≤ 1 − εn−1 as it implies α ≤ ξnh .
Lemma 3.6. Given ε ≤ 41 , δ < εn−1 , integer d ≥ n2n+1 and a positive integer h satisfying
Ah ≥ b; for 0 ≤ α ≤ 1 − εn−1 and k 6∈ {1, n}, if the k-th coordinate xhk of the analytic center
χh (α) satisfies
h i t1 = 1
2
xhk ∈ εk−1 − tk εk−1 δ, 1 − εk−1 + tk εk−1 δ , where t2 = 1 − 4n−1
tk+1 = tk − n1 for k = 2, . . . , n − 1,
Proof. Assume to the contrary that the statement is false, i.e., xhk̂ < 1−ε for k̂ = 1, . . . , k−1.
Considering the first equation of (2) and successively using xh1 < 1 − ε, δ < ε and Lemma 3.3,
we have
ε ε 1 1 h1 h2 ε 1 h1 h2 ε 1 h1 h2 ε 4n − 1
h
+ h = h − h + h − h ≥− + − ≥− + − ≥ ,
s2 s̄2 s1 s̄1 s̃1 s̃2 ε d+1 d δ d+1 d δ
ε ε 1 1 hk−1 hk ε 1 1 hk−1 hk ε 2k n
+ = − + − ≥ − + − ≥ ℓ k−1 ≥ ,
shk s̄hk shk−1 s̄hk−1 s̃hk−1 s̃hk εk−2 1 − εk−2 − ε d + 1 d εk−2 δ
Antoine Deza, Eissa Nematollahi, Reza Peyghami and Tamás Terlaky 9
εk−1 δ εk−1 δ
xhk ≤ εxhk−1 + < ε(εk−2
− t k−1 εk−2
δ) + ≤ εk−1 − tk εk−1 δ,
2k−1 n 2k−1 n
or
εk−1 δ εk−1 δ
xhk ≥ 1 − εxhk−1 − > 1 − ε(εk−2
− t k−1 εk−2
δ) − ≥ 1 − εk−1 + tk εk−1 δ.
2k−1 n 2k−1 n
This is impossible as xhk ∈ εk−1 − tk εk−1 δ, 1 − εk−1 + tk εk−1 δ .
Lemma 3.7. Given ε ≤ 14 , δ < εn−1 , d ≥ n2n+1 , a positive integer h satisfying Ah ≥ b and
t1 , . . . , tn as specified in Lemma 3.6; for 0 ≤ α ≤ 1−εn−1 and k 6∈ {1, n}, if the k-th coordinate
xhk of the analytic center χh (α) satisfies
h i
xhk ∈ tk εk−1 δ, 1 − tk εk−1 δ ,
1 k̂−1
then shk̂ ≥ ε δ for some k̂ smaller than or equal to k − 1.
2k̂ n
1 k̂−1
Proof. Assume to the contrary that the statement is false, i.e., shk̂ < ε δ for k̂ =
2k̂ n
1, . . . , k − 1. This implies
k̂−1
εk̂−1 δ X 1
xhk̂ < for k̂ = 1, . . . , k − 1.
4n 2j−2
j=1
Considering the (k − 1)-st equation of (2) and using s̄hk−1 = 1 − 2εxhk−2 − shk−1 , we have
or
k−2 k−1
εk−1 δ εk−1 δ X 1 εk−1 δ εk−1 δ X 1
xhk ≥ 1 − εxhk−1 − > 1 − − = 1 − ≥ 1 − tk εk−1 δ.
2k−1 n 4n 2j−2 2k−1 n 4n 2j−2
j=1 j=1
Case 1: vkS = 0
The inequality |xhk − vkS | > tk εk−1 δ implies xhk > tk εk−1 δ and, since εxhk ≤ xhk+1 ≤ 1 − εxhk , we
have
tk εk δ < xhk+1 < 1 − tk εk δ.
Since tk+1 < tk , this implies tk+1 εk δ < xhk+1 < 1 − tk+1 εk δ. This contradicts the inequality
S |≤t
|xhk+1 − vk+1 k S S
k+1 ε δ, where vk+1 = 0 or 1 because vk = 0.
Since εxhk ≤ xhk+1 ≤ 1 − εxhk , we have ε(vkS + tk εk−1 δ) < xhk+1 < 1 − ε(vkS + tk εk−1 δ). Since
tk+1 < tk , this implies εvkS + tk+1 εk δ < xhk+1 < 1 − εvkS − tk+1 εk δ. This contradicts the in-
S |≤t
equality |xhk+1 − vk+1 k S S
k+1 ε δ, where vk+1 = εvk or 1 − εvk .
S
or
εk δ εk δ
xhk+1 ≥ 1 − εxhk − > 1 − ε(εk−1
− t k εk−1
δ) − ≥ 1 − εk + tk+1 εk δ.
2k n 2k n
S |≤t
This contradicts the inequality |xhk+1 − vk+1 k S k S
k+1 ε δ, where vk+1 ≥ ε because vk > 0.
Subsubcase 2.2.2: εk−1 − tk εk−1 δ ≤ xhk < vkS − tk εk−1 δ (we have k 6= 1 since 0 < vkS < 1)
By Lemma 3.6, there is a k̂ smaller than or equal to k − 1 such that xhk̂ ≥ 1 − ε, which implies:
shk̂ ≥ 1 − 2ε. Considering the k̂-th equation of (2) and using shk̂ ≥ 1 − 2ε, we have
εk̂ εk̂−1 εk̂−1 εk̂ hk̂ εk̂−1 hk̂+1 εk̂ εk̂−1 εk̂−1 εk̂ hk̂ εk̂−1 hk̂+1 εk̂
= h − h − h + − h ≤ − − h + − ,
sh s s̄ s̄ s̃h s̃ 1 − 2ε ε s̄ d d+1
k̂+1 k̂ k̂ k̂+1 k̂ k̂+1 k̂+1
which implies
εk̂ hk̂ εk̂−1 hk̂+1 εk̂
≤ − = uk̂ εk̂−1 .
sh d d+1
k̂+1
The previous inequality, which corresponds to the case i = k̂, can be generalized to:
i
εi X
≤ uj εj−1 for i = k̂, . . . , k − 1,
shi+1
j=k̂
which clearly holds for k = k̂ + 1 and, for k > k̂ + 1, is obtained by multiplying the i-th
equation of (2) by εi−1 for i = k̂ + 1, . . . , k − 1 and using successively a similar argument.
Noticing that we could have initially permute shk̂+1 and s̄hk̂+1 , gives
k−1
εk−1 X
≤ uj εj−1 .
s̄hk
j=k̂
i.e.,
k−1
εk εk k−1
X
+ ≥ ℓk ε − uj εj−1 .
shk+1 s̄hk+1
j=k̂
k̂−1
εk εk 4n X 2k̂+1 n
+ ≥ + uj εj−1 ≥ ,
shk+1 s̄hk+1 δ
j=1
δ
12 The central path visits all the vertices of the Klee-Minty cube
εk δ εk δ
xhk+1 ≤ εxhk + < ε(vkS − tk εk−1 δ) + ≤ εvkS − tk+1 εk δ,
2k̂ n 2k̂ n
or
εk δ εk δ
xhk+1 ≥ 1 − εxhk − > 1 − ε(vkS − tk εk−1 δ) − ≥ 1 − εvkS + tk+1 εk δ.
2k̂ n 2k̂ n
S |≤t
This contradicts the inequality |xhk+1 − vk+1 k S S S
k+1 ε δ, where vk+1 = εvk or 1 − εvk .
Case 3: vkS = 1
The inequality |xhk − vkS | > tk εk−1 δ implies xhk < 1 − tk εk−1 δ.
εk δ εk δ 1
xhk+1 ≤ εxhk + < t k εk
δ + = (tk + k )εk δ,
2k n 2k n 2 n
or
εk δ εk δ 1
xhk+1 ≥ 1 − εxhk − > 1 − t k εk
δ − = 1 − (tk + k )εk δ.
2k n 2k n 2 n
This contradicts the inequality |xhk+1 −vk+1
S |≤t k S S
k+1 ε δ, where vk+1 = ε or 1−ε because vk = 1.
ε ε 1 1 h1 h2 ε 1 4n − 1
h
+ h = h − h + h − h ≥ − + ℓ1 ≥ .
s2 s̄2 s1 s̄1 s̃1 s̃2 δ δ
2εδ 2εδ
xh2 ≤ εxh1 + < ε(1 − δ) + ≤ ε − t2 εδ,
4n − 1 4n − 1
or
2εδ 2εk δ
xh2 ≥ 1 − εxh1 − > 1 − ε(1 − δ) − ≥ 1 − ε + t2 εδ.
4n − 1 4n − 1
This contradicts |xh2 − v2S | ≤ t2 εδ where v2S = ε or 1 − ε since v1S = 1.
Antoine Deza, Eissa Nematollahi, Reza Peyghami and Tamás Terlaky 13
Subsubcase 3.2.2: k 6= 1
εk̂−1 δ
By Lemma 3.7, there is a k̂ smaller than or equal to k − 1 such that shk̂ ≥ . Considering
2k̂ n
the k̂-th equation of (2), we have
εk̂ εk̂−1 εk̂−1 εk̂ hk̂ εk̂−1 hk̂+1 εk̂ 2k̂ n hk̂ εk̂−1 hk̂+1 ε
k̂
2k̂ n
= − − + − ≤ + − = uk̂ εk̂−1 + .
sh s h s̄ h s̄ h s̃ h s̃ h δ d d+1 δ
k̂+1 k̂ k̂ k̂+1 k̂ k̂+1
This inequality, which corresponds to the case i = k̂, can be generalized to:
i
εi X 2k̂ n
≤ uj εj−1 + for i = k̂, . . . , k − 1,
shi+1 δ
j=k̂
which clearly holds for k = k̂ + 1 and, for k > k̂ + 1, is obtained by multiplying the i-th
equation of (2) by εi−1 for i = k̂ + 1, . . . , k − 1 and using successively a similar argument.
Noticing that we could have initially permute shk̂+1 and s̄hk̂+1 , gives
k−1
εk−1 X j−1 2k̂ n
≤ u j ε + .
s̄hk δ
j=k̂
Using Lemma 3.3, Corollary 3.4, and uk ≥ ℓk , the previous inequality gives
k−1 k̂−1
εk εk X 2k̂ n 4n X 2k̂ n 2k̂+1 n 2k̂ n 2k̂ n
+ ≥ ℓk εk−1 − ui εi−1 − ≥ + ui εi−1 − ≥ − = ,
shk+1 s̄hk+1 δ δ
i=1
δ δ δ δ
i=k̂
Proposition 2.2 is a direct corollary of Proposition 3.8 since, for S 6= ∅ and S 6= {n}, we
have |xhn − vnS | = 0; implying |xhk − vkS | ≤ tk εk−1 δ for k = 1, . . . , n − 1. In other words,
|χh (vnS ) − v S |∞ ≤ δ. Furthermore, by Corollary 3.5 we have |χh − v {n} |∞ ≤ δ, and the central
path converges to the origin v ∅ .
14 The central path visits all the vertices of the Klee-Minty cube
• This result highlights that, although the central path is a smooth analytical curve in
the interior of the set of feasible solutions, it might be severely distracted by redundant
constraints. In particular, exponentially many redundant constraints inter-playing with
the geometry of the problem, may force the central path to take exponentially many and
arbitrarily sharp turns.
• Our example leads to an Ω(2n ) lower bound for the number of iterations needed for
central path-following
√ interior point methods. The theoretical iteration-complexity up-
per bound O( N L) = O(29n n4 ) as, for this example, the number of constraints N =
O(26n n2 ) and the bit-length
q of the input-data L = O(26n n3 ). Therefore, the Ω(2n ) lower
N
bound yields an Ω( 6 )
iteration-complexity lower bound. Using a different analysis,
ln2 N√
Todd and Ye [10] gave an Ω( 3 N L) iteration-complexity lower bound. In a subsequent
paper, Deza, Nematollahi and Terlaky [2] essentially closed the gap between the lower
and upper bounds.
References
[1] G. B. Dantzig: Maximization of a linear function of variables subject to linear inequalities.
In: T. C. Koopmans (ed.) Activity Analysis of Production and Allocation. John Wiley
(1951) 339–347.
[2] A. Deza, E. Nematollahi and T. Terlaky: How good are interior point methods? Klee-
Minty cubes tighten iteration-complexity bounds. AdvOL-Report 2004/20, McMaster Uni-
versity (2004).
[5] V. Klee and G. J. Minty: How good is the simplex algorithm? In: O. Shisha (ed.)
Inequalities III, Academic Press (1972) 159–175.
[6] N. Megiddo: Pathways to the optimal set in linear programming. In: N. Megiddo (ed.)
Progress in Mathematical Programming: Interior-Point and Related Methods, Springer-
Verlag (1988) 131–158; also in: Proceedings of the 7th Mathematical Programming Sym-
posium of Japan, Nagoya, Japan (1986) 1–35.
[7] C. Roos, T. Terlaky and J-Ph. Vial: Theory and Algorithms for Linear Optimization:
An Interior Point Approach. Wiley-Interscience Series in Discrete Mathematics and Op-
timization. John Wiley (1997).
[8] G. Sonnevend: An ”analytical centre” for polyhedrons and new classes of global al-
gorithms for linear (smooth, convex) programming. In: A. Prékopa, J. Szelezsán, and
B. Strazicky (eds.) System Modelling and Optimization: Proceedings of the 12th IFIP-
Conference, Budapest 1985. Lecture Notes in Control and Information Sciences 84
Springer Verlag (1986) 866–876.
[9] T. Terlaky and S. Zhang: Pivot rules for linear programming - a survey. Annals of Oper-
ations Research 46 (1993) 203–233.
[10] M. Todd and Y. Ye: A lower bound on the number of iterations of long-step and poly-
nomial interior-point linear programming algorithms. Annals of Operations Research 62
(1996) 233–252.
[12] Y. Ye: Interior-Point Algorithms: Theory and Analysis. Wiley-Interscience Series in Dis-
crete Mathematics and Optimization. John Wiley (1997).