Math 0602339
Math 0602339
Math 0602339
net/publication/2125921
CITATION READS
1 47
1 author:
Leonid Vaserstein
Pennsylvania State University
93 PUBLICATIONS 1,927 CITATIONS
SEE PROFILE
All content following this page was uploaded by Leonid Vaserstein on 24 December 2012.
Abstra
t. It is well known that every l∞ linear approximation problem can be reduced to a
linear program. In this paper we show that conversely every linear program can be reduced
to an l∞ linear approximation problem.
It is well known that every l∞ linear approximation problem can be reduced to a linear
program. In this paper we show that conversely every linear program can be reduced to
an l∞ linear approximation problem.
Now we recall relevent definitions.
An affine function of variables x1 , . . . , xn is b0 + c1 x1 + · · · + cn xn where b0 , ci are
given numbers.
A linear constraint is any of the following constraints: f ≤ g, f ≥ g, f = g, where f, g
are affine functions.
A linear program is an optimization (maximization or minimization) of an affine func-
tion subject to a finite system of linear constraints.
An l∞ linear approximation problem, also known as (discrete) Chebyshev approxima-
tion problem or finding the least-absolute-deviation fit, is the problem of minimization of
the following function:
where fi are affine functions. This objective function is piece-wise linear and convex.
Given any Chebyshev approximation problem, here is a well-known reduction (Vaser-
stein, 2003) to a linear program with one additional variable t:
1
As usual, x ≥ 0 means that every entry of the column x is ≥ 0. Later we write y ≤ t
for a column y and a number t if every entry of y is ≤ t. We go even further in abusing
notation, denoting by y − t the column obtaining from y by subtracting t from every entry.
Similarly we denote by M + c the matrix obtained from M by adding a number c to every
entry.
This problem (1) (of finding an optimal strategy) is about finding a feasible solution
for a system of linear constraints. It can be written as the following linear program with
an additional variable t:
X
t → min, M x ≤ t, x ≥ 0, xi = 1. (2)
Now we find the largest entry c in the matrix M . If c = 0, then M = 0 and the
problem (1) is trivial (every mixed strategy x is optimal). So we assume that c > 0.
Adding the number c to every entry of the matrix M, we obtain a matrix M + c ≥ 0
(all entries ≥ 0). The linear program (2) is equivalent to
X
t → min, (M + c)x ≤ t, x ≥ 0, xi = 1 (3)
in the sense that these two programs have the same feasible solutions and the same optimal
solutions. The optimal value for (2) is 0 while the optimal value for (3) is c.
Now we can rewrite (3) as follows:
X
k(M + c)xk∞ → min, x ≥ 0, xi = 1 (4)
(M + c)x
c−x
k P k∞ → min . (5)
Px i + c − 1
− xi − c + 1
Note that the optimization problems (4) and (5) have the same optimal value c and
every optimal solution of (4) is optimal for (5). Conversely, for every
P x with a negative
entry, the objective function in (5) is > c. Also, for every x with xi 6= 1, the objective
function in (5) is > c. So every optimal solution for (5) is feasible and hence optimal for
(4).
Thus, we have reduced solving any symmetric matrix game with N × N payoff matrix
to a Chebyshev approximation problem (5) with 2N + 2 affine functions in N variables.
Remark. It is well known that every l1 linear approximation problem can be reduced
to a linear program. Our result implies that every l1 linear approximation problem can
be reduced to a l∞ linear approximation problem. I do not know whether the converse is
true.
2
Note that our reduction of the l1 linear approximation problem
m
X
|fi | → min (6)
i=1
where fi are affine functions in n variables, produces first the well-known linear program
(Vaserstein, 2003)
Xm
ti → min, −ti ≤ fi ≤ ti
i=1
with m + n variables and 2m linear constraints, then a symmetric game with the payoff
matrix of size (3m + 2n + 1) × (3m + 2n + 1), and finally a Chebyshev approximation
problem with 6m + 4n + 4 affine functions in 3m + 2n + 1 variables.
By comparison, an obvious direct reduction produces
References
Vaserstein, L. N. (2003), Introduction to Linear Programming, Prentice Hall. (There is a
Chinese translation by Mechanical Industry Publishing House ISBN: 7111173295. )