# MS&E 211 Linear and Nonlinear Optimization Prof.

Yinyu Ye

Fall 2006 Oct 24, 2006

Homework Assignment 2: Sample Solution
Problem 1. Assume producing x1 poster frames, x2 deluxe poster frames, x3 wood picture frames and x4 metal picture frames will maximize the proﬁt of Oasis. The problem can be formulated as follows: 1. Maximize 2x1 + 4x2 + x3 + x4

subject to x1 + 3x2 + x4 ≤ 4000 2x1 + x2 ≤ 3000 x2 + 4x3 + x4 ≤ 3000 xi ≥ 0 i = 1, 2, 3, 4

2. Solve it by the simplex method. First write it as the standard form:

Minimize

−2x1 − 4x2 − x3 − x4

subject to x1 + 3x2 + x4 + x5 = 4000 2x1 + x2 + x6 = 3000 x2 + 4x3 + x4 + x7 = 3000 xi ≥ 0 i = 1, . . . , 7

We use (4 3 3) to represent (4000 3000 3000) temporally in iterations for the simplicity, and choose the highlighted numbers as the pivot elements in iterations. Its initial simplex tableau is: x1 B 5 6 7 -2 1 2 0 x2 -4 3 1 1 x3 -1 0 0 4 x4 -1 1 0 1 x5 0 1 0 0 x6 0 0 1 0 x7 0 0 0 1 0 4 3 3

Iteration 1: x1 B 2 6 7 Iteration 2: x1 B 2 6 3 Iteration 3: x1 B 2 1 3 0 0 1 0 x2 0 1 0 0 x3 0 0 0 1 x4
7 20 2 5 −1 5 3 20 2 −3 1 3 5 3 1 −3

x2 0 1 0 0

x3 -1 0 0 4

x4
1 3 1 3

x5
4 3 1 3

x6 0 0 1 0

x7 0 0 0 1
16 3 4 3 5 3 5 3

−1 3
2 3

−1 3 −1 3

x2 0 1 0 0

x3 0 0 0 1

x4
1 2 1 3

x5
5 4 1 3

x6 0 0 1 0

x7
1 4

−3 4 1 3 5 3 1 − 12

0 0
1 4

−1 3
1 6

−1 3
1 − 12

23 4 4 3 5 3 5 12

x5
11 10 2 5 −1 5 1 − 10

x6
9 20 −1 5 3 5 1 20

x7
1 4 13 2

0 0
1 4

1 1
1 2

The optimal solution: (x1 , x2 , x3 , x4 ) = 1000 ∗ (1, 1, 1 , 0) = (1000, 1000, 500, 0). The 2 optimal value is 1000 ∗
13 2

= 6500.

9 3. The dual optimal solution(shadow prices) is y = ( 11 , 20 , 1 ), which can be read directly 10 4

from the ﬁnal table. 4. In the optimal solution not any metal picture frame is produced. Assume now we increase its price to p. In our simplex tableau we use p to replace the original price 1. The order that we conduct the operations of the simplex method doesn’t matter so we can just focus on the last step. Force x4 to enter the basis, which requires −p +
4 3

+

1 6

3 20

< 0, i.e.,p > 27/20. (p = 27/20 would be ﬁne too. In that case two

vertices are optimal. Think why?) Problem 2.

1. First formulate the Phase-I problem for it: Minimize y1 + y2

subject to x1 + 2x2 + x3 + 3x4 + y1 = 3 x1 + x2 + 2x3 + 4x4 + y2 = 5 x, y ≥ 0. Then apply the simplex method to it. The initial simplex tableau for this phase-I problem is: 0 y1 y2 Its canonical form is: -2 y1 y2 1 1 -3 2 1 -3 1 2 -7 3 4 0 1 0 0 0 1 -8 3 5 1 1 0 2 1 0 1 2 0 3 4 1 1 0 1 0 1 0 3 5

We choose x4 as the entering basic variable. Pivoting on element 3 in column x4 , we get:
1 3 1 3 5 3 2 3 5 −3 2 −3 1 3 2 3

0 1 0

x4 y2 Pivoting on
2 3

7 3 1 3 4 −3

0 -1 0 1 1 1

−1 3

in column x3 , we get: 0 x4 x3
1 2

0
3 2

0 0 0 1 1 0

1 1 -2

1 −1 2
3 2

0
1 2 3 2

−1 2

−5 2

3 From this tableau, we get a feasible solution (0, 0, 2 , 1 , 0, 0). The reduced cost vector 2

(0, 0, 0, 0, 1, 1) ≥ 0 means this solution is optimal. And the optimal value is 0. So Phase-I procedure terminates here with a BFS to the original problem. Phase-II:

First we check r = c − AT (A−1 )T cB . Here c = (2, 3, 2, 2), AB =  B r = (2, 5, 0, 0). It’s canonical simplex tableau is:

3 1 4 2

 , thus we ﬁnd

2 x4 x3
1 2 1 −2

5
3 2

0 0 1

0 1 0

-4
1 2 3 2

−5 2

The reduced cost (2, 5, 0, 0) ≥ 0 implies this basic feasible solution is already optimal.
3 So, we get the ﬁnal solution for the original problem is (0, 0, 2 , 1 ). And the optimal 2

value is 4.

2. Change b from 

3

5 the initial tableau:

 to 

11 5

 . Then apply the simplex method to Phase-I, we get

-2 y1 y2 1 1

-3 2 1

-3 -7 1 2 3 4

0 1 0

0 0 1

-16 11 5

After 2 iterations, we get the ﬁnal tableau for the Phase-I problem: 1 y1 x2 1 0 1 3 -3 2 5 -5 4 0 1 0 3 -2 1 -1 1 5

-1 0

The optimal value is positive, so the original problem is infeasible and the algorithm terminates. The shadow price vector to this problem, i.e. an infeasibility certiﬁcate for the original feasible set is
 −1 T     1  1 2    1  . =  

0 1

0

−2

Remark: The original system doesn’t have a feasible solution. This system has a form {x : Ax = b, x ≥ 0}. For this type of system, Farkas’ Lemma (Lecture03, Page 6) has the following equivalent version: Let A ∈ Rm×n , b ∈ Rm . The system {x : Ax = b, x ≥ 0} has a feasible solution if and only if that AT y ≤ 0, bT y > 0 has no solution. A vector y, with AT y ≤ 0, bT y > 0, is called an infeasible certiﬁcate for the system. So the infeasible certiﬁcate (1, −2) for this system is given as above.

Problem 3 The LP model can be written as:

Maximize π T x − w · v T s subject to Ax − s = 0 x≤q x≥0 where decision variables x ∈ Rn and s ∈ Rm , and q is the given quantity limit. The dual of the problem is Minimize qT y

subject to AT p + y ≥ π p=w·v y≥0 By complementary slackness conditions: xi (aT p + yi − πi ) = 0 i yj (qj − xj ) = 0 Case 1) 0 < xj < qj : xj < qj , yj (qj − xj ) = 0 ⇒ yj = 0 xj > 0, xj (aT p + yj − πj ) = 0 ⇒ aT p + yj − πj = 0 j j ⇒ aT p = πj j Case 2) xj = 0 : xj < qj , yj (qj − xj ) = 0 ⇒ yj = 0 ⇒ a T p ≥ πj j Case 3) xj = qj : xj > 0, xj (aT p + yj − πj ) = 0 ⇒ aT p + yj − πj = 0 j j ⇒ yj = πj − aT p ≥ 0 j

In summary:

0 < xj < qj , =⇒ aT p = πj j xj = 0, xj = q j , =⇒ aT p ≥ πj j =⇒ aT p ≤ πj j

Where p is the dual variable vector associated with constraints Ax − s = 0 and y is the dual vector associated with constraints x ≤ q. p ≥ 0 is implied by v ≥ 0. We can see from above that the prices are fair. The interpretation of the dual is to ﬁnd state prices p and slacks y between the parimutuel prices AT p and price limit π such that the total diﬀerence q T y = q T π − q T AT p is minimized. The complementary conditions imply the relationships of primal and dual. Most of these conditions are the same as those in our robust model. 2) Orders honored: (100, 50, 100, a, 50). 0 ≤ a ≤ 100 The expected proﬁt: 22.5. The actual proﬁt: −30 − 0.05a. (The company lost money!) Problem 4 1. The linear program actually describes the min-cut problem. Suppose ui (i = A, B) is meant to be 1 if i is in the cut with S and 0 otherwise. Each of the y variables is meant to be 1 if the corresponding edge contributes to the cut capacity. Then these constraints guarantee these variables behave exactly as they should. For example, the ﬁrst constraint forces SA to be in the cut if A is not with S. The third constraint implies that AB contributes to the cut capacity only if A is with S and B is with T. Note here we assume uA (or uB ) is either 0 or 1 for the simplicity of the discussion. Their ranges can actually be extended to any real value. 2. The dual is:

Maximize subject to

fSA + fSB fSA ≤ 3 fSB ≤ 2 fAB ≤ 1

fBA ≤ 1 fAT ≤ 1 fBT ≤ 3 fSA + fBA − fAB − fAT = 0 fSB + fAB − fBA − fBT = 0 f ≥0 let fij be the ﬂow through edge (i, j). The constraints describe all edge capacity constraints and ﬂow conservations. The objective function is the total ﬂow through the network. So it is a max-ﬂow problem. 3. Any feasible (s, t)−cut is a feasible solution to the min-cut formulation, so its capacity is an upper bound of the value of the min-cut LP formulation. Similarly the value of any feasible ﬂow pattern is a lower bound of the value of the max-ﬂow LP formulation. By weak duality, the former is always no less than the latter. Problem 5 According to the deﬁnition of the concavity, we need to prove for any b1 , b2 ∈ [10, 30], 1 1 0 ≤ α ≤ 1, we always have: z(αb1 + (1 − α)b2 ) ≥ αz(b1 ) + (1 − α)z(b2 ) 1 1 1 1 . Suppose an optimal feasible solution to the original linear program with parameter b1 = b1 is 1 x = (x1 , x2 , x3 ). Similarly suppose an optimal feasible solution to the original linear program with parameter b1 = b2 is y = (y1 , y2 , y3 ). 1 Let w = αx + (1 − α)y. Then w1 + 3w2 + 6w3 = α(x1 + 3x2 + 6x3 ) + (1 − α)(y1 + 3y2 + 6y3 ) ≤ αb1 + (1 − α)b2 . 1 1 Similarly, it is easy to check: 2w1 + 15w3 ≤ 30 8w2 + 9w3 ≤ 30 w1 , w 2 , w 3 ≥ 0 Therefore, w is a feasible solution to the linear program:

Maximize

5x1 + 15x2 + 30x3

subject to x1 + 3x2 + 6x3 ≤ αb1 + (1 − α)b2 1 1 2x1 + 15x3 ≤ 30 8x2 + 9x3 ≤ 30 x1 , x 2 , x 3 ≥ 0 So, by the deﬁnition of z(b1 ), we know: 5w1 + 15w2 + 30w3 ≤ z(αb1 + (1 − α)b2 ). 1 1 And 5w1 + 15w2 + 30w3 = α(5x1 + 15x2 + 30x3 ) + (1 − α)(5y1 + 15y2 + 30y3 ) = αz(b1 ) + (1 − α)z(b2 ) 1 1 Thus, according to the deﬁnition, we know z(b1 ) is concave. Remark: This proof can be generalized: z(b) is concave for b if z(b) is deﬁned as follows:

z(b) := Maximize

cT x

subject to Ax = b x≥0