A Financial Engineering Problem

Goal:
How to use LP to model multi‐period cash
management in financial firms?

@2020 New York University Tandon

2/24/2020 123
School of Engineering
 Problem Setting:
A firm, named Firm X, must determine investment
strategy for the firm during the next 3 years. At
present (time 0), $100K is available for investment.
Five investments named A, B, C, D and E are
available, with the cash flow per $1 investment as
follows:
year 0 1 2 3
A ‐$1 +$0.50 +$1 $0
B $0 ‐$1 +$0.50 +$1
C ‐$1 +$1.2 $0 $0
D ‐$1 $0 $0 +$1.9
E $0 $0 ‐$1 +$1.5

@2020 New York University Tandon

2/24/2020 124
School of Engineering
 Hypotheses:
1. The maximum investment for each
category is $75K.
2. The firm is NOT allowed to borrow money
to invest.
3. The interest rate for uninvested money is
8%.

Objective:
Maximize cash on hand at the end of year 3.

@2020 New York University Tandon

2/24/2020 125
School of Engineering
 Solution
Unknown Variables:
x1  $ amount invested in investment A;
x2  $ amount invested in investment B;

x5  $ amount invested in investment E;
St  $ amount invested in guaranteed 8%-return at time t.

Objective function:
P  x2  1.9 x4  1.5x5  1.08S2
@2020 New York University Tandon
2/24/2020 126
School of Engineering
 Solution (cont’d)
Constraints:
0  xi  75K , i  1, 2, ,5;
0  St , t  0,1, 2;
x1  x3  x4  S0  100 K ; (year 0)
0.5 x1  1.2 x3  1.08S0  x2  S1 ; (year 1)
x1  0.5 x2  1.08S1  x5  S 2 (year 2).
(cash available) (cash invested)
@2020 New York University Tandon
2/24/2020 127
School of Engineering
 Solution (cont’d)
Maximizing solution:
P  218,500; with
*

x  60, 000; x2  30, 000; x4  40, 000;

1
* * *

x  75, 000;
5
*

x  S0  S  S 2  0.
3
* *
1
* *

@2020 New York University Tandon

2/24/2020 128
School of Engineering
 Exercise 1

Use the simplex method to find a basic and feasible solution

to the following system of linear inequalities:
2x1  3x2  2x3  3,

  x1  x2  x3  5,
x  0, i 1,2,3
i

@2020 New York University Tandon

2/24/2020 129
School of Engineering
 Exercise 2

Use the two-phase method to solve the following LP problem:

min P  4 x1  2 x2  8 x3
2 x1  x2  3 x3  30,

 x1  2 x2  4 x3  40,
 x  0, i  1, 2,3
 i

@2020 New York University Tandon

2/24/2020 130
School of Engineering
 Exercise 3
State the following LP problem in standard form and solve it
using the simplex method:
min P  2 x1  x2  5 x3 subject to
 x1  2 x2  x3  8,
3 x  2 x  18,
 1 2

2 x1  x2  2 x3  4
 xi  0

@2020 New York University Tandon

2/24/2020 131
School of Engineering
 Duality
Goals:
• Reveal the Relationships between the Primal and
Dual LP Problems.
• Introduce the concept of “Shadow Price” and its
importance via solving the dual LP problem.
• Introduce the Dual Simplex Method (if time
permits)

@2020 New York University Tandon

2/24/2020 132
School of Engineering
 Primal and Dual LP Problems
Primal LP Problem ("canonical form"):
min P  cT x
subject to Ax  b
x  0.

D ual LP P rob l e m : Goals:

m ax Pd  b T y • Make explicit the effects of changes in the
constraints on the cost
subject to A T y  c
• Develop the “interior-point” software via
y  0. combination of primal and dual info
@2020 New York University Tandon
2/24/2020 133
School of Engineering
 Dual of a LP in Standard Form
First, notice that a linear program in standard form
can be rewritten as:

min P  c x T

subject to Ax  b
 Ax  b
x  0.
@2020 New York University Tandon
2/24/2020 134
School of Engineering
 Dual of a LP in Standard Form (cont’d)
Then, its “canonical” dual LP is:

max Pd  u T b  vT b
subject to AT u  AT v  c
u0 Or,
v  0.
max Pd  bT 
subject to A   c T

u v  free variable!

@2020 New York University Tandon
2/24/2020 135
School of Engineering
 An LP with Mixed Constraints
What is the “canonical” dual of the following LP:

min P  c x T

subject to A1 x  b1
A2 x  b2
A3 x  b3
x  0.
@2020 New York University Tandon
2/24/2020 136
School of Engineering
 An LP with Mixed Constraints (cont’d)

The dual LP problem is:

max Pd  b1T y1  b2T y2  b3T y3

subject to A1T y1  A2T y2  A3T y3  c
y1  0, y2  0, y3 unrestricted.

@2020 New York University Tandon

2/24/2020 137
School of Engineering
 Theorem 1 of Weak Duality
(on the standard form)
Consider the primal and dual problems:

Primal LP Problem Dual LP Problem

min P  c x T
max Pd  b T

subject to Ax  b subject to AT   c
x0  free variable
Then
P  c x  Pd  b 
T T
for any feasible x, .
@2020 New York University Tandon
2/24/2020 138
School of Engineering
Relation of primal and dual values:

Dual values Primal values P

@2020 New York University Tandon

2/24/2020 139
School of Engineering
Comments
1. If the dual is unbounded, then the primal is
infeasible.

2. If there is a pair of feasible solutions  x ,  

such that cT x  bT , then x and  are
optimal for their respective problems.

@2020 New York University Tandon

2/24/2020 140
School of Engineering
 An Example
Consider the primal problem

min P  x1
subject to x1  1
 x1  1
x1  0.
What is its dual problem? Unbounded?
@2020 New York University Tandon
2/24/2020 141
School of Engineering
 An Example (cont’d)
The dual problem is

max Pd  1   2
subject to 1   2  1
1 ,  2  0.

@2020 New York University Tandon

2/24/2020 142
School of Engineering
 Theorem 2 of Strong Duality

Primal LP Problem Dual LP Problem

min P  cT x max Pd  bT 
subject to Ax  b subject to A   c
T

x0  free variable

If one problem has an optimal solution, then
the other problem also has an optimal solution. Plus

P  cT x  Pd  bT  for optimal x, .
@2020 New York University Tandon
2/24/2020 143
School of Engineering
 Sketch of Proof
Let x* be an optimal basic feasible solution for the primal
x*  col  xB xN  .
Noting that A   B N  and c  col  cB cN  , then
xB  B 1b and the reduced costs cTN  cBT B 1 N  0.

It is direct to check that *  ( B 1 )T cB is feasible and optimal

for the dual and has the same optimal primal value,
as shown below.

@2020 New York University Tandon

2/24/2020 144
School of Engineering
 Sketch of Proof (cont’d)
*  ( B 1 )T cB is feasible because

 
T
1 T 1 1
A *  A ( B ) cB  ( B A) cB  I B N
T T T
cB


 col cB (B 1 N )T cB   col(cB cN ) : c.

*  ( B 1 )T cB is optimal because
Pd  bT *  bT ( B 1 )T cB  ( B 1b)T cB  cT x*  P* .
 
xTB

@2020 New York University Tandon

2/24/2020 145
School of Engineering
 An Equivalence Result
From the proof, we obtain:

Primal optimality condition cTN  cBT B 1 N  0

 The dual feasibility condition

 
T
1
A *  c, with *  B
T
cB

@2020 New York University Tandon

2/24/2020 146
School of Engineering
 Complementary Slackness

Let x* be the optimal primal (in standard form) and

 * the optimal dual. Then, it holds:
x*T  c  AT  *   0.

@2020 New York University Tandon

2/24/2020 147
School of Engineering
 An Example
Verify the strong duality theorem on the LP:

min P   x1  2 x2
subject to  2 x1  x2  2
 x1  2 x2  7
x1 3
x1 , x2  0.
@2020 New York University Tandon
2/24/2020 148
School of Engineering
 A Practical Example
A baker produces two types of cakes:
x1  elaborate  and x2  simple  .

max P  24 x1  14x2
subject to 3x1  2 x2  1200 (basic ingredients)
4x1  x2  1000 (fancier ingredients)
2 x1  x2  700 (labor)
x1  0, x2  0.
@2020 New York University Tandon
2/24/2020 149
School of Engineering
 A Practical Example (cont’d)
Its dual LP problem is:

min Pd  12001  1000 2  700 3

subject to 31  4 2  2 3  24 (elaborate cake)
21   2   3  14 (simple cake)
1  0,  2  0,  3  0.

@2020 New York University Tandon

2/24/2020 150
School of Engineering
Answers
P  8880, x1  160, x2  360
Pd  8880, 1  6.4,  2  1.2,  3  0.

1) For this problem, there are 20 hours of excess labor available,

as at the optimal solution 700  2 x1*  x2*  20. So,
additional labor has no value to the baker   3  0  .

@2020 New York University Tandon

2/24/2020 151
School of Engineering
Answers
P  8880, x1  160, x2  360
Pd  8880, 1  6.4,  2  1.2,  3  0.

2) In case the baker wants to purchase additional

quantities of the ingredients, each pound of basic
ingredients will be worth 1  6.4$ in profit,
while each pound of fancy ingredients will be
worth  2  1.2$. For these reasons, the dual variables
are called shadow prices.
@2020 New York University Tandon
2/24/2020 152
School of Engineering
Answers
P  8880, x1  160, x2  360
Pd  8880, 1  6.4,  2  1.2,  3  0.

3) The dual problem can also be understood as the "price"

a company should pay to take over the baker's business.

@2020 New York University Tandon

2/24/2020 153
School of Engineering
 Sensitivity
Goal: Understand the effects of parameter
variations on the optimal solution.

@2020 New York University Tandon

2/24/2020 154
School of Engineering
 Key Observations
• The current basis is feasible if
1
B b  0.

xB

• It is optimal if
1
c  c B N  0.
T
N
T
B

@2020 New York University Tandon

2/24/2020 155
School of Engineering
 An Example
Consider the linear programming problem
min P   x1  2 x2
subject to  2 x1  x2  2,
 x1  2 x2  7,
x1  3,
x1 , x2  0.

@2020 New York University Tandon

2/24/2020 156
School of Engineering
From Lecture 3, the last step in applying the
Simplex Method yields (Optimal solution):

Basic rhs
x1 x2 x3 x4 x5
-P 0 0 0 1 2 13

x2 0 1 0 1/2 1/2 5
x1 1 0 0 0 1 3
x3 0 0 1 -1/2 3/2 3

@2020 New York University Tandon

2/24/2020 157
School of Engineering
 Current Basis: xB   x2 , x1 , x3  , with
T

1 2 1  0 0.5 0.5 
  1  
B  2 1 0  , B   0 0 1 ,
0 1 0  1 0.5 1.5 
 
0 0  0.5 0.5 

N  1  , B 1 N   ,
0  0 1 
0 1   0.5 1.5 
 
cB   2  1 0  , cN   0 0  .
T T

@2020 New York University Tandon

2/24/2020 158
School of Engineering
 Optimal primal variables:
xB   x2 , x1 , x3   B b   5 3 3 , 1
T T

Reduced costs:
c  c B N  1 2   (0 0)
1
T T T T
N B

Optimal dual variables:

*   B 
1 T
cB   0  1  2  .
T

@2020 New York University Tandon

2/24/2020 159
School of Engineering
 An Example of Sensitivity Analysis
Consider the perturbed linear programming problem
min P   x1  2 x2
subject to  2 x1  x2  2,
 x1  2 x2  7  ,
x1  3,
x1 , x2  0.
Show that the optimal basis remains unchanged as long as
   10, 6 .

@2020 New York University Tandon

2/24/2020 160
School of Engineering
 A Short Proof
 Observe that the optimality condition remains unchanged
for this special type of perturbations.

 The feasibility condition requires that

B (b  b)  0, with b   0  0 
1 T

Or equivalently,
 5   0.5 
   
B 1b   B 1b   3    0 
 3   0.5 
   
implying the claimed range for  .
@2020 New York University Tandon
2/24/2020 161
School of Engineering
 An Example of Sensitivity Analysis
Validate the above claim by solving the perturbed LP problem
min P   x1  2 x2
subject to  2 x1  x2  2,
 x1  2 x2  7  ,
x1  3,
x1 , x2  0.
for different perturbations :   4, 8, 1, 4.

@2020 New York University Tandon

2/24/2020 162
School of Engineering
 The Primal‐Dual Interior‐Point Method
for LP
Motivation: fast polynomial‐time method for
large LP problems

Interior‐Point Method by N. Karmarkar (1984):

Search for optimal solution through interior points,

instead of extreme points.

@2020 New York University Tandon

2/24/2020 163
School of Engineering
 Primal‐Dual Problems
The Primal Problem (P):
min z  c x T

subject to Ax  b, x  0. (A: full row-rank)

The Dual Problem (D):

max w  bT y
subject to AT y  s  c, s  0  slack variables  .

@2020 New York University Tandon

2/24/2020 164
School of Engineering
 Complementary Slackness Conditions

For optimal solution x to (P) and optimal solution y to (D),

the "complementary slackness" condition must hold:
xT  c  AT y   0. Duality Gap :
c T x  bT y
Or , equivalently,
x j s j  0, j  1, , n.

@2020 New York University Tandon

2/24/2020 165
School of Engineering
 Key Idea of the PD Interior Method

For positive   0, approximate the optimal solution

with x    , y    , s    satisfying:
 Ax  b, x  0
 T
 PD   A y  s  c, s  0
 x s   , j  1, , n.
 j j

Remark 1:   0 gives interior points x  0, s  0.

Remark 2 :   0 gives the optimal solution!
@2020 New York University Tandon
2/24/2020 166
School of Engineering
 Duality Gap

c x  b y  x s  n .
T T T

The above duality gap measures the closeness

to the optimal solutions.

@2020 New York University Tandon

2/24/2020 167
School of Engineering
 Algorithm
Given estimates x  0, y and s  0, with
Ax  b and A y  s  c, but x j s j   ,
T

find new estimates x  x, y  y, s  s

such that
A  x  x   b  Ax  0 (*)
Similarly, A y  s  0. * *
T

@2020 New York University Tandon

2/24/2020 168
School of Engineering
 Algorithm (cont’d)
In addition, the nonlinear complementary
slackness
x j  x j  s j  s j   
is approximated by linear equations:
s j x j  x j s j    x j s j ***
Thus, these three equations determine the
directions x, y, s.
@2020 New York University Tandon
2/24/2020 169
School of Engineering
 Update Parameter Law

k 1  k , 0    1.

@2020 New York University Tandon

2/24/2020 170
School of Engineering
Comment
In practice, we often adopt:
 x  ,    x  x,

 y  ,    y  y,

 s  ,    s  s,
where  is a step length, chosen to ensure that
x, s are positive (i.e. interior points).

More to come at the end of NL programming:

Interior-point methods for convex programming
@2020 New York University Tandon
2/24/2020 171
School of Engineering
 Exercise
Can you solve the following problem using both the
Simplex and the Interior‐Point methods?

min P  2 x1  3 x2  2 x3  x4  x5
3 x1  3 x2  4 x3  2 x4  x5  0,

subject to  x1  x2  x3  3 x4  x5  2,
 x  0, i  1, ,5.
 i
@2020 New York University Tandon
2/24/2020 172
School of Engineering