Sie sind auf Seite 1von 2

Math464 - HW 3 Solutions 1

Linear Optimization (Spring 2016)


Brief Solutions to Homework 3
1. We first prove that f is convex when each fi is convex.

(a) Since the fi s are convex, [0, 1], and for x, y Rn , we have

fi ( x + (1 )y ) fi (x) + (1 )fi (y), i = 1, . . . , m. (1)

For f (x) = m
P
i=1 fi (x), we get

m
X
f ( x + (1 )y ) = fi ( x + (1 )y )
i=1
Xm
fi (x) + (1 )fi (y), (by (1))
i=1
m
! m
!
X X
= fi (x) + (1 ) fi (y) = f (x) + (1 )f (y),
i=1 i=1

which gives the definition of convexity of f .


(b) Now consider the case where each fi is PL convex, i.e.,

fi (x) = max cTij x + dij , i = 1, . . . , m.



j=1,...,mi

Hence, we get
m
X m
X
cTiji x + diji

f (x) = fi (x) = max
ji =1,...,mi
i=1 i=1
!
X 
(cTij1 x + dij1 ) + + (cTijm x + dijm )

= max
j1 =1,...,m1 , ..., jm =1,...,mm
i=1,...,m
!T !
m
X m
X
= max cij1 + + cijm x+ dij1 + + dijm ,
j1 =1,...,m1 , ..., jm =1,...,mm
i=1 i=1

showing that f is also PL convex.

Alternatively, we could use induction to prove both the above results. It may also be
a bit more straightforward as far as notation goes. Consider the PL convex case. The
main step is to show the result for m = 2, i.e., for f (x) = f1 (x) + f2 (x), where

f1 (x) = max (cTi x + di ), and f2 (x) = max (gjT x + hj ).


i=1,...,p j=1,...,q
Math464 - HW 3 Solutions 2

Hence we get
f (x) = max (cTi x + di ) + max (gjT x + hj )
i=1,...,p j=1,...,q
 T
ci x + di + gjT x + hj

= max
i=1,...,p, j=1,...,q

(ci + gj )T x + (di + hj ) ,
 
= max
i=1,...,p, j=1,...,q

which is PL convex.
Pm We could now assume the result holds for m, and look at f (x) =
P m+1
i=1 fi (x) = i=1 fi (x) + fm+1 (x), which is the sum of two convex functions by the
induction assumption. By the above result for m = 2, if follows that the result holds
for m + 1 as well.
2. The piecewise linear (PL) convex function f (x) can be described as follows.

x + 1 : x1
f (x) = 0 : 1x2
2x 4 : x2

Note that dT x is a scalar, and hence f (dT x) will be a PL convex function of the
form f (dT x) = maxi=1,2,3 (ei (dT x) + fi ), where e1 = 1, f1 = 1, e2 = 0, f2 = 0, and
e3 = 2, f3 = 4. The LP can be written as follows.
min cT x + z
s.t. z dT x + 1
z 0
z 2dT x 4
Ax b
3.
min 2x1 + 3z2
s.t. z2 x2 10
z2 (x2 10)
x1 + 2 + x2 5
x1 + 2 x2 5
(x1 + 2) + x2 5
(x1 + 2) x2 5
P 1
4. The LP for minimizing Tt=0 |at | is as follows:
PT 1
min t=0 zt (total fuel consumed)
s.t. zt at , t = 0, . . . , T 1 (abs. value)
zt at , t = 0, . . . , T 1 (abs. value)
xt+1 = xt + vt , t = 0, . . . , T 1 (displacement)
vt+1 = vt + at , t = 0, . . . , T 1 (velocity)
x0 = 0, v0 = 0 (initial conditions)
xT = 1, vT = 0 (final conditions)
For minimizing the maximum thrust, replace the objective function by min z, and the
absolute value constraints by z at and z at for t = 0, . . . , T 1.

Das könnte Ihnen auch gefallen