Derivations and notes on Hotelling’s Lemma

Envelope Theorem

A very useful way to formulate the main ideas of the envelope theorem is to

take the maximization problem

max U (x, θ)
x

Where U (x, θ) is the function to be maximized, x is the choice variable, θ

is an exogenous parameter. Just for the simple version of the theorem, assume

that U (x, θ) is differentiable in both x and θ. Let x∗ (θ) ∈ arg max U (x, θ) be a
x

solution to the maximization problem for a given value of the parameter θ, and

of course for different values of θ the optimal solution x∗ can be different.

We define the value function to be the maximum level of U (x, θ). So V (θ) =

U (x∗ (θ), θ) = max U (x, θ)

x

The confusion arises with some of the notations that are commonly used. In

a strict sense, V (θ) is a function only of θ, not x, because x∗ (θ) is endogenously

determined.
0 ∂U (x∗ (θ),θ)
So the envelope thorem states that V (θ) = ∂θ = Uθ (x∗ (θ), θ)
∂U (x∗ (θ),θ) ∂x∗ (θ) ∂U (x∗ (θ),θ)
The intuition is that d
dθ (U (x∗ (θ), θ)) = ∂x ∂θ + ∂θ ; how-
∗
∂U (x (θ),θ)
ever, at the optimal level, ∂x = 0 implied by the FOC, therefore d
dθ (U (x∗ (θ), θ)) =
∂U (x∗ (θ),θ)
∂θ

1
 In words, it tells you that any change in the parameter θ only impacts

the value function directly, the indirect effect through fixing the optimal x∗ is

second-order, because at the optimum level the gains from changing the choice

variable is negligible.

Hotelling’s Lemma

For this part we assume that the production function f (·) uses I inputs to

produce one output, and is strictly concave, so the FOC approach is valid.

Let’s say for a given level of input prices, π(p) = max pf (x) − w.x =
x
∗ ∗ ∗
pf (x (p)) − w.x (p); where x (p) is the unique solution to the maximization

problem given output price p.

I n
∂x∗
0
o
∗ ∗ i (p)
∂ ∂
[pf (x∗ (p)) − w.x∗ (p)]
P
So π (p) = ∂xi [pf (x (p)) − w.x (p)] ∂p + ∂p
i=1
We know that for each i, ∂
∂xi [pf (x∗ (p)) − w.x∗ (p)] = 0 based on the FOC,

so we only have the second part of the expression:

0 ∂
π (p) = [pf (x∗ (p)) − w.x∗ (p)] = f (x∗ (p)) = y(w, p)
∂p

In words, the associated change in profit due to a change in price equals the

level of supply. Again, the idea is that with a slight increase in output price,

the firm makes slightly more money on all the units of output it was already

selling, but it does not change the amount of output it produces.

2
Shepherd’s Lemma

This applies to the cost minimization problem, when the firm has already de-

cided to produce at least amount y. The cost function is:

c(w, y) = min w.x

x

s. to f (x) ≥ y

Let’s say x∗ (w, y) is the solution to the cost minimization problem, so

c(w, y) = w.x∗ (w, y). Then,

I
∂c(w, y) ∂ X
= wi .x∗i (w, y) = x∗i (w, y)
∂wi ∂wi i=1

∂c(w,y)
So the Shephard’s lemma says ∂wi = x∗i (w, y), in words, the change in

cost associated with a change in input price equals the factor demand for that

input. Again the intuition is that the firm has to pay a little bit more for all

the units of good i it was using, but at the optimal level, it won’t fix the level

of x∗i .

Comparative Statics of Factor Demands

(These derivations use some Matrix algebra, which you will not be tested on in

the exam. These derivations are therefore optional to learn)

Consider the profit maximization problem with two inputs:

max p.f (x) − w.x

x

Let x(p, w) be the factor demand arising from a concave production function

3
f (·), which means in 
the profit-maximization
 problem, the SOC holds, and the
∂2f ∂2f
∂x21 ∂x1 ∂x2 
Hessian matrix, H =   is negative semi-definite.

∂2f ∂2f
∂x1 ∂x2 ∂x22
Now, the FOC of the profit maximization problem gives us:

∂f (x1 , x2 )
p = w1
∂x1
∂f (x1 , x2 )
p = w2
∂x2

Taking the derivative of these two above w.r. to w1 we get

∂ 2 f ∂x1 ∂ 2 f ∂x2 1
2 + =
∂x1 ∂w1 ∂x1 ∂x2 ∂w1 p
∂ 2 f ∂x1 ∂ 2 f ∂x2
+ =0
∂x1 ∂x2 ∂w1 ∂x22 ∂w1

Similarly, taking the derivatives of the two FOCs w.r. to w2 we get

∂ 2 f ∂x1 ∂ 2 f ∂x2
+ =0
∂x21 ∂w2 ∂x1 ∂x2 ∂w2
∂ 2 f ∂x1 ∂ 2 f ∂x2 1
+ 2 =
∂x1 ∂x2 ∂w2 ∂x2 ∂w2 p

Writing all these in a matrix form, we get

    
∂2f ∂2f ∂x1 ∂x1 1
∂x21 ∂x1 ∂x2   ∂w1 ∂w1  p 0
=

 2 2
 
∂ f ∂ f ∂x2 ∂x2 1
∂x1 ∂x2 ∂x22 ∂w1 ∂w2 0 p

This means,

     
∂2f ∂2f ∂x1 ∂x1
∂x21 ∂x1 ∂x2   ∂w1 ∂w1  1 0
 p.  =

 
∂2f ∂2f ∂x2 ∂x2
∂x1 ∂x2 ∂x22 ∂w1 ∂w2 0 1

4
  
∂x1 ∂x1
 ∂w1 ∂w1 
This means p.   is the inverse of the hessian matrix. As the Hes-
∂x2 ∂x2
∂w1 ∂w2
sian is negative semi-definite, this matrix is also negative semi-definite. Which
∂xi
implies that ∂wi ≤0

Assuming a strictly concave production function, how can we first prove that

y(p, w) is increasing in p? (that is, law of supply holds)

When the production function is strictly concave, the associated cost func-

tion is strictly convex in y, which means cyy (w, y) > 0.

∂
This is the same as saying ∂y (cy (w, y)) > 0, that is, the marginal cost

function is strictly increasing in y.

We know from childhood that at the profit-maximizing output level y ∗ ,

marginal cost equals price. So p = cy (w, y ∗ ).

What happens when p ↑? At the new optimal y ∗ , we must still have p =

marginal cost, so for a higher p, the higher marginal cost will be attained at a

higher y ∗ (this follows from the upward sloping MC curve in this case).

That simply means, as p goes up, y(p, w) = f (x1 (p, w), x2 (p, w)) also goes

up. Law of supply holds in this case.

Now, going back to the FOCs:

∂f (x1 , x2 )
p = w1
∂x1
∂f (x1 , x2 )
p = w2
∂x2

Dividing one equation by the other:

∂f (x1 ,x2 )
∂x1 w1
∂f (x1 ,x2 )
=
w2
∂x2

5
 Taking the derivatives w.r. to p of this equation gives us:

∂f ∂ 2 f ∂x1 ∂ 2 f ∂x2 ∂f ∂ 2 f ∂x2 ∂ 2 f ∂x1

   
+ − + =0
∂x2 ∂x21 ∂p ∂x1 ∂x2 ∂p ∂x1 ∂x22 ∂p ∂x1 ∂x2 ∂p

Collecting terms, we get

 <0 >0
  >0 <0

z }| { z }| { z }| { z }| {
 ∂f ∂ 2 f ∂f ∂ 2 f 
∂x1   ∂f ∂ 2 f ∂f ∂ 2 f 
 ∂x2
− + − =0
 
∂p  ∂x2 ∂x21 ∂x1 ∂x1 ∂x2  ∂p  ∂x2 ∂x1 ∂x2 ∂x1 ∂x22 
 
| {z }
>0
| {z }
<0

∂2f
The assumption here is that ∂x1 ∂x2 > 0. With this assumption, the sum

of the two parts above can only be 0 is one part is positive and the other part
∂x1 ∂x2
is negative. That in turn can only happen if ∂p and ∂p both have the same
∂x1 ∂x2
sign. As we know that y(p, w) is increasing in p, this means both ∂p and ∂p

must be positive.
∂2f
What does the assumption ∂x1 ∂x2 > 0 mean? It means we have assumed

that the two inputs are complements, in that the marginal productivity of xi
∂2f
is increasing in the amount used of x−i , which means ∂x1 ∂x2 > 0. What this

assumption does is guarantees that the optimal way to produce more output is

to choose more of both inputs.