IEOR E4007

G. Iyengar December 17th, 2015

IEOR E4007 Final Exam

Instructions : 200pts

1. There are 4 questions in all.

2. You have 3 hours to do the exam.

3. The exam is a closed book; but you are allowed to bring in one standard sheet of
paper as a cheat sheet.

4. Calculators are allowed but no computers.

Please sign the following declaration and return with your answerbook.

I declare that I did not receive nor did I give any assistance to anyone else
during this exam.

———————————————————————–

1
1. Short answer problems Total points: 70
(a) Find the global maximum and minimum of the function 10
f (x, y) = 2 + 2x + 2y − x2 − y 2 ,
on the set D = {(x, y) : x + y = 9, x ≥ 0, y ≥ 0} by representing the problem as a
constrained optimization problem in one variable.
Solution: Since f (x, y) = 2 + 2(x + y) − (x + y)2 + 2xy, it follows that

max/min(x,y)∈D f (x, y) ≡ max/min0≤x≤9 {2x(9 − x)}

Since g(x) = 2x(9−x) is a differentiable concave function of x, the global maximum

is achieved at the point where g 0 (x) = 2(9 − x) − 2x = 0, i.e. x = 9/2. Therefore,
the constrained maximum is achieved at the point (x, y) = ( 29 , 92 )
Since g(x) is a differentiable concave function, the global minimum is achieved
at the boundary. Since g(x) = 0 at x = 0 and x = 9, it follows that both the
boundary points are global minima. Therefore, the constrained minima are given
by (x, y) ∈ {(9, 0), (0, 9)}.

(b) Compute the step implied by the Hessian-based method (also known as the Newton 5
method) for the function f (x, y) = 2 + 2x + 2y − x2 − y 2 at the point (x, y) = (1, 2)
Solution: The gradient is given by
   
2 − 2x 0
∇f = =
2 − 2y −2

The Hessian is given by

∇2 f = −2I
Thus, the step  
2 −10
δNewton = [∇ f ] ∇f =
1

(c) Consider the constrained optimization problem

min x2 + y 2 ,
s.t. (x − 1)3 − y 2 ≥ 0.
The global minimizer of this optimization problem is (x, y) = (1, 0).
i. Show that this is not a KKT point for this problem. 10
Solution: Since the constraint g(x) is active at the point x̄ = (1, 0), it follows
that the KKT conditions are given by

∇f (x̄) = u∇g(x̄), u ≥ 0.

2
 Since ∇g(x̄) = 0 and ∇f (x̄) 6= 0, these equations do not have a solution, i.e.
x̄ is not a KKT point.
ii. How is it the case that a global maximum is not a KKT point? 5
Solution: A local maximum is a KKT point if a constraint qualification is
met. Since the constraints are not linear, we would need the gradients of
the active constraints to be linearly independent. This is also violated since
∇g(x̄) = 0. Thus, there is no guarantee that the global maximum will be a
KKT point.
(d) For what values of (α, β), if any, is the following optimization problem convex? 10
Give a short justification.
minx∈R2 c>x − ln(1 − b> x), 
2−α 0
s.t. x0 x ≥ 2,
0 −1 + 2α
b> x ≤ 1,
3βd> x = 1.

Solution: The objective is convex. The feasible set is convex provided the quadratic
constraint defines a convex set. This would be the case, if the matrix Q =
diag(2−α, −1+2α) is negative semidefinite. This will never be the case; therefore,
the problem is never convex.
(e) Reformulate the following nonlinear knapsack problem into integer linear program. 15
max 11 ln(x1 + 1) + 8 ln(x2 + 1) + 6 ln(x3 + 1) + 4 ln(x4 + 1),
s.t. 7x1 + 5x2 + 4x3 + 3x4 ≤ 14,
xj ≥ 0 and an integer, ∀j = 1, . . . , 4.

Solution: Let ki = b W ci
c, where W = 14 and c = [7, 5, 4, 3]. For j = 0, . . . , ki ,
i = 1, . . . , 4, define φij = pi ln(1 + j), where p = [11, 8, 6, 4]. Then the above
nonlinear integer program is equivalent to the linear integer program
P4 Pki
max φij yij
Pi=1
4 Pj=0
ki
s.t.
Pi=1 j=0 jci yij ≤ W,
ki
j=0 yij = 1, i = 1, . . . , 4,
yij ∈ {0, 1}, j = 0, . . . , ki , i = 1, . . . , 4.

(f) Reformulate the following robust mean-variance problem as a second-order cone 15

program:
maxx min{µ> > > >
1 x − λx Q1 x, µ2 x − λx Q2 x},

where Q1 = V1> V1 , Q2 = V2> V2 , and λ1 , λ2 > 0.

3
 Solution: The optimization problem is equivalent to the following problem

maxx,β β
s.t. µ> >
1 x − λx Q1 x − β ≥ 0,
µ2 x − λx> Q2 x − β ≥ 0.
>

Now, each of the two quadratic constraints can be reformulated as second-order

cone constraint as follows.
maxx,β,σ1 ,σ2 β
s.t. µ> 1 x − λσ1 − β ≥ 0,
µ>
2 x − λσ 2 − β ≥ 0,
2V1 x
≤ 1 + σ1
1 − σ1 2
 
2V2 x
≤ 1 + σ2
1 − σ2 2
σ1 , σ2 ≥ 0.

2. Active set for a QP Total points: 30

Consider the following inequality constrained QP
min x21 + x22 − 2x1 − 2x2
s.t. x1 + 4x2 ≤ 10,
4x1 + x2 ≤ 10,
x≥0
 >
(a) Show that x̄ = 2 2 is a critical point for the equality constrained QP constructed 10
by keeping only the constraints active at x̄.
Solution: Since 4x̄1 + x̄2 = x̄1 + 4x̄2 = 10, it follows that the relevant equality
constrained QP is given by

min ∇f (x̄)> y + y12 + y22

s.t. y1 + 4y2 = 0,
4y1 + y2 = 0,

where ∇f (x̄) = 21. The constraints imply that y = 0.

The KKT conditions are given by
       
1 2y1 1 4
2 + = u1 + u2 .
1 2y2 4 1

u1 = u2 = 25 is the unique solution for this system of equations. Therefore, the

point x̄ is a critical point for the equality constrained problem.

4
 (b) Is the point x̄ a critical point for the inequality constrained QP? Give a short 5
explanation.
Solution: Both the dual variables u1 and u2 do not have the correct sign. There-
fore, x̄ is not a critical point for the inequality constrained QP.

(c) The constraint 4x1 + x2 ≤ 10 is active at the point x̄ = 21. Construct an equality 15
constrained QP that excludes this constraint. Compute the improving direction
and the next iterate.
Solution: The equality constrained QP is given by

min ∇f (x̄)> y + y12 + y22

s.t. y1 + 4y2 = 0.

The KKT conditions are given by

     
1 2y1 1
2 + =u .
1 2y2 4

Thus, y1 = 21 (u − 2) and y2 = 21 (4u − 2). Substituting back into the constraint

y1 + 4y2 = 8.5u − 5 = 0, we get u = 10/17. Therefore, y1 = − 12 17
3
and y2 = 17 .
It is easy to check that we can take
 22 the full step, i.e. λ = 1, is feasible. Therefore,
37
the next iterate is given by x = 17 17 .

3. Driver scheduling problem Total points: 40

The Columbia University provides night transportation in the Morningside Heights area
starting from 7pm up until 2am. Over this period, the shuttle service uses several drivers
but only one works at any given time. If a shift begins at or before 9pm, a regular driver
can be obtained for a 4-hr shift at the cost of $60 for the shift. Otherwise part-time
student drivers will have to be used. The student drivers are willing to work 3-hr shifts
for $37.5 and and 2-hr shifts for $34. We are interested in designing the cheapest shift
schedule that covers all the hours.
(a) Argue that the problem of computing the minimum cost driver shift schedule can 15
be modeled as a shortest path problem on an appropriate directed graph. Sketch
the directed graph and label the corresponding arc lengths.
Solution: The node set is V = {7, 8, 9, 10, 11, 12, 1, 2}, i.e. one node for each hour
that the service needs to cover. Each arc in the graph represent a possible shift
and the “length” of the arc is the cost of the shift:

(i) Regular driver: (7, 11), (8, 12), (9, 1) with cost 60

(ii) Three hour shifts (7, 10), (8, 11), (9, 12), (10, 1), (11, 2) with cost 37.5

5
 (iii) Two hour shifts: (7, 9), (8, 10), (9, 11), (10, 12), (11, 1), (12, 2) with cost 34

A feasible driver schedule corresponds to a path from node 7 to node 2. Thus, the
minimum cost driver shift corresponds to the shortest path from node 7 to node 2.

(b) Solve the shortest path problem to calculate the minimum cost schedule and the 15
corresponding cost. In order to get full credit for this part you need to show all the
computations.
Solution: Note that the graph is directed and acyclic. All arcs are from nodes i
to node j > i. For such a graph we can use a shortest path algorithm where the
states are the nodes and the number of stages is just 1. This is akin to computing
the shortest path on a capital budgeting graph.
Let V (k) denote the shortest path from node 7 to node k. Then V (7) = 0, and
for all k 6= 9,
V (k) = min {V (i) + cik }
{i:cik 6=0}

and the preceding the node in the optimal path is given by

π(k) = argmin{V (i) + cik }

{i:cik 6=0}

For this particular graph, the computation proceeds as follows:

V (8) = +∞ π(8) = ∅
V (9) = V (7) + c7,9 = 34 π(9) = 7
V (10) = min{V (7) + c7,10 , V (8) + c8,10 }
= min{37.5, +∞} = 37.5 π(10) = 7
V (11) = min{V (7) + c7,11 , V (8) + c8,11 , V (9) + c9,11 }
= min{60, +∞, 68} = 60 π(11) = 7
V (12) = min{V (8) + c8,12 , V (9) + c9,12 , V (10) + c10,12 }
= min{∞, 71.5, 71.5} = 71.5 π(12) = {9, 10}
V (1) = min{V (9) + c9,1 , V (10) + c10,1 , V (11) + c11,1 }
= min{94, 75, 94} = 75 π(1) = 10
V (2) = min{V (11) + c11,2 , V (12) + c12,2 }
= min{97.5, 105.5} = 97.5 π(2) = 11

The cost of the min-cost schedule is 97.5 and the schedule itself is given by
{(7, 11), (11, 2)}

(c) Can you read off the solution for the min-cost schedule from 7pm-1am from the 5
computations in part (b) ? If so, what is it?
Solution: Since V (j) denotes the shortest path from node 7 to all other nodes, the
value of V (1) = 75 gives the cost of the min-cost schedule from 7 → 1, and optimal

6
 schedule can be computed by the π values. The optimal schedule is (7, 10), (10, 1).

(d) Can you read off the solution for the min-cost schedule from 9pm-2am from the 5
solution for part (b) ? If so, what is it?
Solution: Since V (j) denotes the shortest paths from node 7 to node j, we cannot
read off the min-cost schedule for 9 → 2 from the computations in part (b).

4. Dark pool execution problem Total points: 60

A trader wants to sell B = 20 blocks of shares on n = 5 possible dark pool venues.
In a dark pool, the volume of shares available for trading are not known before the trade
is placed. Suppose that all the venues are independent of each other. Let

Fj (x) = P(S̃j ≤ x), x ∈ {0, 1, . . . , ∞}

denote the CDF of the random supply S̃j ≥ 0 available in dark pool j.
Suppose the trader submits an order k blocks of shares to dark pool j. Then depending
on the supply S̃j , a random amount x̃ ≤ k will get executed. In this case, we assume
that the revenue is given by
Xx̃
vj (`),
`=0

i.e. vj (`) is the price at which the `-th block is sold in dark pool j, conditioned on that
the supply is high enough. We assume that vj (0) = 0. If a block of shares is not sold
on a venue j, it is returned back to the seller, i.e. the seller does not obtain any revenue
for it.
Let rj (k) denote the expected revenue from submitting an order of k blocks to venue j.
Suppose we formulate the trader’s optimization problem as follows:

max r1 (x1 ) + r2 (x2 ) + . . . + r5 (x5 ),

s.t. x1 + x2 + x3 + x4 + x5 ≤ 20,
xi ∈ Z+ .

(a) Write an expression for rj (k) in terms of the functions Fj (·) and vj (·). 15
Solution: The revenue we get from selling the k-th block is vj (k), and the prob-
ability that we can indeed sell this block is given by P(S˜j ≥ k) = 1 − Fj (k). So
the total expected revenue we get from selling k blocks in venue j is
k
X

rj (k) = vj (`) 1 − Fj (`) .
`=0

7
 An alternative expression is as follows. The probability gj (`; k) that exactly ` ≤ k
blocks get executed on venue j is

Fj (` + 1) − Fj (`), 0 ≤ ` < k
gj (`; k) =
1 − Fj (k), ` = k.

Thus, an alternative expression for fj (k) is as foll ows:

k
X `
X 
rj (k) = gj (`; k) vj (γ)
`=0 γ=0

This expression simplifies to the expression above.

(b) We want to solve this problem using the recursion that we used to solve the capital 10
budgeting problem. Let

Vj (s) = max rj (xk ) + rj+1 (xk+1 ) + . . . + r5 (x5 ),

s.t. xj + xj+1 + . . . + x5 ≤ s,
xi ∈ Z+ .

Write out the recursion for this problem relating the function Vj (s) to either the
function Vj+1 (·) or Vj−1 (·). Be sure to record the optimal solution πj (·) in stage j
and state s; you will need it later.
Solution: Vj+1 (s) is the maximum revenue we get from selling s blocks through
venue j + 1, ..., 5. Now suppose we want to include venue k in our consideration.
To decide how many blocks in s should we allocate to venue j, we use the following
recursion:
Vj (s) = max {rj (x) + Vj+1 (s − x)}.
0≤x≤s

The optimal solution

πj (s) = argmax{rj (x) + Vj+1 (s − x)}.

0≤x≤s

(c) Carefully describe how you will initialize your recursion. 5

Solution: To initialize the recursion, we first compute V5 (s), 0 ≤ s ≤ B. This is
straightforward since our only choice is to put everything into venue 5, we have
V5 (s) = r5 (s).

(d) Suppose now you have all the functions Vk (s) for k = 1, . . . , 5 and s = 0, . . . , B. 10
What is the optimal order that you would send to each dark pool?

8
 Solution: The optimal order x ∈ Z5 can be constructed as follows:
s←B
for j = 1 : 5 do
xj ← πj (s)
s ← s − xj
end for
(e) The trader’s optimization appears knapsack-like. Will the recursion 5
Uj (s) = max{rj (1) + Uj (s − 1), Uj−1 (s)}
work for this problem? Give an explanation justifying your answer.
Solution: The increase in revenue by allocating one additional block to venue k
depends on the current number of blocks assigned to venue k. In fact, the revenue is
decreasing as additional blocks are added to the venue, i.e. we have a diminishing
return scenario. The value function Uj (s) is unable to keep track of the diminishing
return; therefore, we cannot use to solve a recursion based on this value function
to solve the optimization problem.
(f) The dark pool problem, as stated and formulated, implicitly assumes that the orders 15
to all dark pools will be put in simultaneously, ignoring that the block may not be
completely executed.
Suppose instead the trader first submits an order to the first dark pool, observes the
execution, and then proceeds to put the order into the second dark pool, and so on.
Describe a dynamic program for computing the optimal policy for this stochastic
problem.
[Hint: Carefully examine dynamic programming recursion above. You should be
able to re-use most of it.]
Solution: Let Vj (s) denote the optimal expected revenue from executing s blocks
over the dark pools m = j, j + 1, . . . , 5.
Suppose we execute x blocks in dark pool j. Then, the number of blocks y executed
will be 0 ≤ y ≤ x with probability gj (y; x), where gj (y; x) is defined in the solution
of part (a), and the remaining inventory s − y would have be routed to dark pools
m ≥ k + 1. Hence, the value of action x in dark pool k is given by
x
( y ! )
X X
gj (y; x) vk (`) + Vk+1 (s − y)
y=0 `=0

Thus, the value function

( x ( y ! ))
X X
Vk (s) = max gj (y; x) vk (`) + Vk+1 (s − y) ,
0≤x≤s
y=0 `=0

9
and the optimal policy order size
( x ( y ! ))
X X
πk (s) = argmax gj (y; x) vk (`) + Vk+1 (s − y) ,
0≤x≤s
y=0 `=0

10