FOUNDATIONS OF OPTIMIZATION: IE6001

Shadow Prices

Napat Rujeerapaiboon
Semester I, AY2019/2020
 Example 1 (perturbed)

Assume that p1 , the availability of X, is not precisely known.

max x1 + x2 objective function

s.t. 2x1 + x2 ≤ p1 constraint on availability of X
x1 + 3x2 ≤ 18 constraint on availability of Y
x1 ≤ 4 constraint on demand of A
x1 , x2 ≥ 0 non-negativity constraints

−min −x1 − x2
s.t. 2x1 + x2 + x3 = p1
x1 + 3x2 + x4 = 18
x1 + x5 = 4
x1 , x2 , x3 , x4 , x5 ≥ 0
 Example 1 (perturbed)

The value function v (p1 ) expresses the optimal value of the LP

as a function of the unknown availability parameter p1 .

v (p1 ) = min −x1 − x2

s.t. 2x1 + x2 + x3 = p1
x1 + 3x2 + x4 = 18
x1 + x5 = 4
x1 , x2 , x3 , x4 , x5 ≥ 0
 Example 1 (perturbed)
x2

Optimal Solution with basic variables x 1 , x 2, x

5
and nonbasic variables x 3 , x 4

Availability of X:
p1 = 11

Optimal Value:
v(11) = -8

x1
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
+1
 Example 1 (perturbed)
x2

Degenerate Optimal Solution

I = {1,2,3}, I = {1,2,4}, I = {1,2,5} are all optimal!

Availability of X:
2
p1 = 12 3

Optimal Value:
2 2
v(12 3 ) = -8 3

x1
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
+1
 Example 1 (perturbed)
x2

Optimal Solution with basic variables x 1 , x 2, x

3
and nonbasic variables x 4 , x 5

Availability of X:
2
p1 > 12 3

Optimal Value:
2
v(p1) = -8 3

x1
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
+1
 Example 1 (perturbed)
x2

Degenerate Optimal Solution

I = {1,2,5}, I = {2,3,5}, I = {2,4,5} are all optimal!

Availability of X:
p1 = 6

Optimal Value:
v(6) = -6

x1
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
+1
 Example 1 (perturbed)
x2

Optimal Solution with basic variables x 2 , x 4, x

5
and nonbasic variables x 1 , x 3

Availability of X:
0 > p1 < 6

Optimal Value:
0 > v(p1) > -6

x1
-9
-8
-7
-6
-5
-4
-3
-2
-1
0
+1
 Example 1 (perturbed)

Note: v (p1 ) is non-increasing, convex and piecewise linear.

v(p1)

2
0 6 12 3

Problem p1
infeasible
for p1 < 0

-6

2
-8 3
 Perturbation

Let p ∈ Rm denote a general RHS and define the

value function v (p) : Rm → R by:
n o
v (p) = min c T x | A x = p; x ≥ 0

Solving the original LP (the reference problem)

n o
min x0 = c T x | A x = b, x ≥ 0

thus computes v (b).

 Shadow Prices

Suppose we have solved the reference problem

n o
min x0 = c T x | Ax = b, x ≥ 0

and found an optimal basis matrix B satisfying

xB = B −1 b ≥ 0 (Feasibility)

and

r = cN − N T (B −1 )T cB ≥ 0 (Optimality).
 Shadow Prices (cont)

Definition: The vector of shadow prices Π ∈ Rm is defined as

Π = (B −1 )T cB ,

where B = B(I) is an optimal basis.

Note that there can be more than one optimal basis

⇒ The shadow prices need not be unique.

The shadow Prices give information about the sensitivity of the

value function v (p) at p = b.
 Local Behaviour of Value Function

Theorem: v (p) = v (b) + ΠT (p − b) for all p ∈ Rm with

B −1 p ≥ 0.

Proof:
• If B −1 p ≥ 0, then B remains the optimal basis for

min{c T x : Ax = p, x ≥ 0}

since r is not affected by changing b to p.

• Thus, we find

v (p) = cBT B −1 p
= cBT B −1 b + cBT B −1 (p − b)
= v (b) + ΠT (p − b) 
 Global Behaviour of Value Function

Theorem: v (p) ≥ v (b) + ΠT (p − b) for all p ∈ Rm .

Proof:
v (p) = minx≥0;Ax=p c T x


= minx≥0;Ax=p c T x − ΠT (Ax − p)


≥ minx≥0 c T x − ΠT (Ax − p)


= minx≥0 (c T − ΠT A)x + ΠT p


= ΠT p + minx≥0 (c T − ΠT A)x

| {z }
≥0 see next slide!
 Global Behaviour of Value Function

 

xB
cBT B −1
 T
c − ΠT A x cBT cNT
  
= | − [B | N]
xN
   
 T T
 xB T
 −1
 xB
= cB | cN − cB I | B N
xN xN

= cBT xB − cBT xB + cNT − cBT B −1 N xN

= r T xN

≥ 0 (as r ≥ 0, and xN ≥ 0)
 Global Behaviour of Value Function

Thus, we find

v (p) ≥ ΠT p + minx≥0 (c T − ΠT A)x



≥ ΠT p
= ΠT b + ΠT (p − b)
= cBT B −1 b + ΠT (p − b)
= v (b) + ΠT (p − b)


 Shadow Prices in Example 1
Note: Π1 is the shadow price for the budget of ingredient X.

v(p1)
p1 = b1

2
0 6 11 12 3

p1

-6 1
1
2
-8 3

At p1 = b1 = 11, the optimal costs change by Π1 = − 25

if the availability of X increases by 1.
 Interpretation

• Assume the company can buy a ”small” additional amount

of ingredient X, at price µ1 per unit.

• Is it worthwhile to buy additional units of X?

• Yes if µ1 + Π1 < 0 (overall cost decreases);

• No if µ1 + Π1 > 0 (overall cost increases).

⇒ Therefore, −Π1 is the maximum price one should pay for

one additional unit of X!
 Evaluation of Shadow Prices

Sometimes shadow prices can be read off the final tableau.

Lemma: Suppose row t is initially a “≤-constraint” and a slack

variable xs had been added. Then, Πt = βs , where βs is the
objective coefficient of xs in the final (optimal) tableau.
Proof:
• If xs is nonbasic in the final tableau, then

βs = −rs = −(cN − N T B −T cb )T es
= −cs + ΠT Nes = 0 + ΠT et = Πt .

• If xs is basic in the final tableau, then

βs = 0 = cs = esT cB = esT B T Π = etT Π = Πt . 

 Evaluation of Shadow Prices (cont)

Sometimes shadow prices can be read off the final tableau.

Lemma: Suppose row t is initially a “≥-constraint” and a

surplus variable xs had been added. Then, Πt = −βs , where βs
is the objective coefficient of xs in the final tableau.
 Example 1 (revisited)

The final tableau for Example 1 is:

BV x0 x1 x2 x3 x4 x5 RHS
x0 1 0 0 − 25 − 15 0 −8
x2 0 0 1 − 15 2
5 0 5
x5 0 0 0 − 35 1
5 1 1
3
x1 0 1 0 5 − 15 0 3

• The constraint on the availability of X was standardised by

introducing the slack variable x3 .
• The shadow price Π1 for that constraint thus coincides with
the coefficient of x3 in the objective row of the above
tableau ⇒ Π1 = − 52 .