ARE211, Fall 2004

LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

C ONTENTS

6. Foundations of Comparative Statics 1

6.1. The envelope theorem for unconstrained maximization 1

6.2. The envelope theorem for constrained maximization 2

6.3. Application of the envelope theorem for constrained maximization 8

6. F OUNDATIONS OF C OMPARATIVE S TATICS

6.1. The envelope theorem for unconstrained maximization

In economics, we’re often interested in a function which has two arguments; the second is a function of
the first. As we’ve discussed in a previous lecture (CALCULUS2), economists typically deal with this
by invoking the (unfortunately named) concept of the total derivative: if f (b, x) = f (b, x(b)), then total
derivative of f x.r.t b is d f /db = f b (b, x(b)) + f x (b, x(b))x 0 (b), where b and x are here scalars. When b
changes in this case, there is a change in f due to two factors: first b changes, also, x changes as b changes.

In this lecture, we’ll consider the case in which the second argument is a special kind of function of b; x ∗ (·)
is the value of x that maximises (or minimizes) f for each value of b. The function f (b, x ∗ (b)) is then called
the value function.

Example π(p, q ∗ (p)); for each p, pick the q that maximizes profits for that p; call this function q ∗ (p). Now
ask how profit adjust as price changes and the producer adjusts quantity.

Other examples of value functions in economics are the expenditure function and the indirect utility function.

1
2 LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

Answer is given by the envelope theorem which says that in this case, d f /db = ∂ f /∂b. (i.e., you’ve learnt
to tell the difference between d f and ∂ f ; now you find that in this case, there isn’t any difference.)

The envelope theorem: Varian: Given f : R 2 → R1 (differentiable) and a function x∗ : R1 → R1 (differen-

tiable) defined by the condition that for each b, x ∗ (b) maximizes f (b, ·). Then the total derivative of the
function f (b, x∗ (b)) with respect to b is d f (·, x ∗ (·))/db = ∂ f (·, x∗ (·))/∂b.

Mathematical proof is trivial

d f (b, x∗ (b)) dx∗ (b)
= fb (b, x∗ (b)) + fx (b, x∗ (b))
db db
Necessary condition for x to maximize f (b, ·) is that f x (b, x) = 0; this is how x∗ (b) is defined; hence
fx (b, x∗ (b)) = 0 by definition of x∗ (b).

The picture is much more important; note that in the picture, the golden rule is broken, for display purposes:
the first component of the function is pictured on the horizontal axis.

• The line in the domain x∗ (b) has the property that vertically above points on this line, the function
f (b, ·) is maximized in the x direction.
• Now as you move out along the line x∗ (b) there are in general two contributors to the change in f ;
f changes because b changes AND because x changes.
• In this particular case, f doesn’t change when x changes, but it does change when b changes.
• It’s worth noting that there is nothing special about moving out along the line x ∗ (b): if you start out
at (b, x∗ (b)) and move in any direction whatsoever, the only thing that matters is the movement in
the b direction.
• Compare this picture to the general case in which we move out at the same angle, where the change
in f due to x is very substantial;
• In the general case, the x effect really makes a difference.

6.2. The envelope theorem for constrained maximization

The point of the above version of the envelope theorem is that when you are maximizing unconstrainedly a
function f (b, ·), where b is a parameter, then the rate of change in f as b changes does not depend on the
rate at which x∗ (·) moves with b. (In class, I did the theorem for the case in which x ∗ was a scalar; it is also
obviously true when x∗ is a vector.)
 ARE211, Fall 2004 3

x∗ (b)

(b3 , x∗ (b3 ))

(b2 , x∗ (b2 ))
(b1 , x∗ (b1 ))
x

PSfrag replacements
db
df

dx
∂f ∂f
df = ∗
∂b (b, x (b))db + (b, x∗ (b))dx
|∂x {z }
=0

When you evaluate d f , the dx term has no effect because it is multiplied by zero

F IGURE 1. The envelope theorem and the differential

An analogous result holds when you solve the problem:

maximize f (b, x) subject to the constraint that h(b, x) = 0. (1)

Let x∗ (b) denote the solution to (1) and let M(b) denote the maximized value of f given b.

Notice that there is a difference between (1) and the familiar specification, i.e., max f (b, x) subject to
g(x) = b. But the familiar specification is a special case of the current one. To see this, let h(b, x) = b − g(x).
The current specification allows for more general comparative statics than we have seen before. In our
original specification i.e., max f (b, x) subject to g(x) = b, we learnt how to do comparative statics w.r.t.
4 LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

b, but not with respect to the other parameters of g(·). When we write the problem in the current form,
we can do comparative statics w.r.t. any parameter of either the objective or the constraint. For example,
suppose that our problem is max u(x) s.t. p · x = y. Our analysis in the preceding section taught us how to
do comparative statics w.r.t. y but not w.r.t. the components of p. We are about to see how to do comparative


statics w.r.t. these components as well.

∂ f (b,x∗ (b))
+ λ∗ (b) ∂h(b,x
dM(b) ∗ (b))
The constrained version of the envelope theorem says that db = ∂b ∂b . (Notice that
∂ f (b,x∗ (b))
for the special case in which f does not depend on b, so that ∂b = 0, and h(b, x ∗ ) = b − g(x∗ ), so that
∂h(b,x∗ (b)) ∂ f (b,x (b))
+ λ∗ (b) ∂h(b,x
∗ ∗ (b))
db = λ (b), which
dM(b)
∂b = 1, the expression db = ∂b ∂b reduces to simply dM(b) ∗

is the old familiar result: λ∗ (·) measures the rate at which the objective function increases as the constraint
is relaxed.) As in the unconstrained version of the envelope theorem, the total derivative of M w.r.t. b does
involve the partial derivatives of f w.r.t. the elements of the x ∗ vector, but these terms disappear in the
dM(b)
expression for db .

The striking difference between the unconstrained and the constrained theorems is that in the former case, the
∂ f (b,x∗ (b))
movement in the x direction didn’t matter because the ∂xi ’s were zero. In the present case, the gradient
of f isn’t zero, and yet the movement in the x direction still doesn’t matter. So while the unconstrained
theorem is very easy to explain intuitively, the constrained theorem is by no means so.

Fig. 2 illustrates the theorem, for the simplest case in which there is only one constraint, h(b, x ∗ ) = b−g(x∗ ),
To present the result in its sharpest form, the figure depicts a linear optimization problem, i.e., the level sets
of both the objective f and the constraint h are affine functions. In the upper panel, we start out at an
optimum x∗ (b̄) on the constraint set, i.e., the level set of h associated with b̄. (Note that for this linear case,
the optimum is not unique). In the right panel, the level sets of the objective function have a different slope
from that of the constraint function, so we start out at an arbitrary (non-optimal) point x˜∗ on the constraint
set. Now consider what happens when b changes to b 0 , and the constraint line moves outwards. It is critical
to my story that the new constraint line is parallel to the old one: this must be the case because g is affine,
so its gradient cannot change direction (or length) with x. In the upper panel, the change in the objective
function doesn’t depend on how x∗ moves, provided that x∗ moves to the new, parallel constraint line. In
the right panel, different directions of movement from the old constraint line to the new result in different
changes in the objective function. This is a graphical depiction of the mathematical result that when you
start out at a constrained optimum, and shift to the new, parallel constraint line, the change in the value of
the objective depends on the partials of both f and h w.r.t. b but not on the partials of either f or h w.r.t. the
components of x∗ .
 ARE211, Fall 2004 5

f,h f,h

x2 x2
L.S. of f at b0 (shift of dx0 )
L.S. of f at b̄
L.S. of f at b0 (shift of dx)
dx0
L.S. of f at b0 (shift of dx)
L.S. of f at b̄ dx
x(b̄) L.S. of h at b0 x̃
L.S. of h at b̄
x1
L.S. of h at b̄ L.S. of h at b0 x1
F IGURE 2. The envelope theorem: constrained optimization when Hessian is zero

f,h f,h

x2 x2

PSfrag replacements

dx’ dx
dx
dx’
L.S. of f at b0 (shift of dx) L.S. of f at b̄ L.S. of f at b0 (shift of dx)
L.S. of f at b̄ 0
x(b̄) L.S. of f at b0 (shift of dx ) x̃ L.S. of f at b0 (shift of dx0 )
L.S. of h at b̄ L.S. of h at b0 L.S. of h at b̄ L.S. of h at b0

x1 x1
F IGURE 3. The envelope theorem: constrained optimization when Hessian is nonzero

When either the objective or the constraint are nonlinear, i.e., when you add some curvature to the level sets,
it’s a little less easy to see from pictures what’s going on. Indeed the important implication of the envelope
theorem seems to be false, since differences in the direction of movement dx do make a difference. However,
as the two panels of Fig. 3 illustrate, the difference is much smaller when you start out at an optimum x ∗ (b̄)
(as in the left panel of the figure) than when you start out at some arbitrary point x˜∗ (as in the right panel
of the figure). In both panels, you get to a higher level set when you move to the new constraint line in the
direction dx of the gradient than if you move in a different direction, such as dx 0 . In the left panel, however,
the difference between the level sets you reach is only second order, i.e., due to the Hessian term, but in the
right panel the difference is first order (i.e., due to the first order term in the Taylor expansion).

As we’ve noted, Fig. 3 seems to contradict the envelope theorem, which says that the only things that matter
are (a) how far out the constraint moves when b changes; (b) how rapidly f increases as you move out to
the new constraint. What doesn’t matter, according to the envelope theorem, is the direction in which you
6 LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

move in order to get to the new constraint line. The reconciliation of this paradox (at least for the case when
only one constraint is binding—when two or more are binding we have to tell a more complicated story) is
that the envelope theorem is only telling you about the first order Taylor approximation to the change in f
when b changes. In Fig. 2, there are no second order effects, i.e., the Hessian is zero. So the first term in the
Taylor expansion tells the whole story: movements in the x ∗ direction don’t matter at all when you start out
at an optimum. In the left panel of Fig. 3, there is no first order effect because we start out at an optimum,
only second order effects. Since the envelope theorem just gives you a first-order approximation to the true
change in f when b changes, the differences we see in the left panel of Fig. 3 evaporate. In the right panel
of Fig. 3, we don’t start from an optimum and there are both first and second order effects.

Now for the formalism. We’ll do the general NPP, i.e., with m inequality constraints. Once again, we’ll see
that the direction in which the solution vector x ∗ (·) moves as b moves has no effect on the total derivative of
f w.r.t. b.

The envelope theorem for constrained maximization: Consider f : R × R n → R1 (differentiable) and

λ ∗ (b)) satisfy the KKT conditions for
h : R1 × Rn → Rm (differentiable). For b ∈ R, suppose that (x ∗ (b),λ
the nonlinear programming problem:

maxx f (b, x) such that h j (b, x) ≥ 0, for j = 1, ..., m.

Then the total derivative of the function f (b, x ∗ (b)) with respect to b is
d f (b, x∗ (b)) ∂ f (b, x∗ (b)) m
∂h j (b, x∗ (b))
db
=
∂b
+ ∑ λ∗j (b) ∂b
j=1

Please note that the KKT for this more general version of the NPP is that

∇ f (b, x∗ (b)) = − ∑mj=1 λ∗j (b)∇h j (b, x∗ (b)), with λ∗j (b) ≥ 0 for all j, and h j (b, x∗ (b)) > 0 =⇒ λ∗j (b) = 0.

In words, in this setting the gradient of f belongs to the nonpositive cone defined by the gradients of the
constraints that are satisfied with equality at x ∗ (b)). To reconcile this with our usual treatment, note that
for the special case in which h j (b, x) = b j − g j (x), we have ∇h j = −∇g j , so that in this special case,
the nonnegative cone defined by any subset of the gradients of g j ’s is the nonpositive cone defined the
corresponding subset of the gradients of h j ’s.

Proof: The Lagrangian for this problem is

m
λ ∗ (b))
L(b, x∗ (b),λ = f (b, x ∗ (b)) + ∑ λ∗j (b)h j (b, x∗ (b))
j=1
 ARE211, Fall 2004 7

since λ∗j (b)h j (b, x∗ (b)) = 0 for all j, we have

M(b) ≡ f (b, x ∗ (b)) ≡ λ ∗ (b))

L(b, x∗ (b),λ

It follows that
dM(b, x∗ (b)) d f (b, x∗ (b)) λ∗ (b))
dL(b, x∗ (b),λ
= =
db db db
∂L(b, x (b),λ
∗ λ (b))
∗ n λ ∗ (b)) dx∗ i (b)
∂L(b, x∗ (b),λ m λ ∗ (b)) dλ∗j (b)
∂L(b, x∗ (b),λ
= + ∑ + ∑
∂b i=1 ∂xi db j=1 ∂λ j db
∂ f (b, x∗ (b)) m
∂h j (b, x∗ (b))
=
∂b
+ ∑ λ∗j (b) ∂b
j=1
!
n
∂ f (b, x∗ (b)) m
∂h j (b, x∗ (b)) dxi (b) m dλ∗j (b)
+ ∑ ∂xi
+ ∑ λ∗j (b) ∂xi db
+ ∑ h j (b, x∗ (b)) db
i=1 j=1 j=1

∂L(b,x∗ (b),λ∗ (b))

Since x∗ (b) satisfies the KKT conditions each of the ∂xi ’s is zero. That is, for each i, the term
in the large parentheses on the second line is zero. Moreover if h(b, x ∗ (b)) < 0, then λ∗j (·) is zero on a
dλ∗j (b) dλ∗j (b)
neighborhood of b and hence db is zero; Hence for all j h j (b, x∗ (b)) db is zero. Conclude that all
terms on the second line are zero. This proves
d f (b, x∗ (b)) ∂ f (b, x∗ (b)) m
∂h j (b, x∗ (b))
db
=
∂b
+ ∑ λ∗j (b) ∂b
j=1

Having emphasized that the constrained and the unconstrained versions of the envelope theorem are so
completely different, we now remark that they are really not so different after all. In fact, the so-called
“unconstrained version” is a special case of the “constrained version,” with a particularly simple constraint.
Specifically, assume that the first derivative of f (·, ·) is positive. Now define g(β, x) = β and consider the
constrained maximization problem

max f (β, x) such that g(β, x) ≤ b. (2)

β,x

In the solution to (2), β necessarily will equal b (since f increases with β) and x will maximize f (b, ·). Hence
this problem is equivalent to the unconstrained maximization problem we considered in the beginning of this
lecture. More formally, the KKT conditions for (2) require that ∇ f and ∇g are colinear; in this case, since
g doesn’t depend on x, a necessary condition for a solution at x ∗ (b) is that f x (b, x∗ ) = 0, just as in the
unconstrained version.
8 LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

6.3. Application of the envelope theorem for constrained maximization

PSfrag
Jacob Viner’s famous figure of the long-run and short-run replacements
average total
cost functions provides a nice example of the envelope theorem. Con-

Average Cost
sider a production function q = f (`, k). The LRATC curve assigns to
LRATC(·)
each q the average total cost associated with the cost-minimizing com- SRATC(·, k)
bination of labor and capital. That is,
w` + rk
LRATC(q) = min s.t. q ≤ f (`, k)
{`,k} q q q̄ Output

For each level of capital k, SRATCk (q) is the average cost of producing q F IGURE 4. LRATC & SRATC
using k and whatever is the required level of `. For each q, LRATC(q) =
SRATCk (q) at the level k that is optimal for that q. The graphs exhibit the well-known property that each
SRATC curve is tangent to the LRATC curve at the point where they agree (see Fig. 4). Because of this
relationship, the LRATC is referred to as the outer envelope of the SRATC curves. The “puzzle” here is
that one would expect the short-run curve to be steeper than the long-run curve at q̄, because when output
increases from q, capital is held constant in the short-run, but varies in the long run: we would expect that
by adjusting both inputs in response to an increase in q, the producer could reduce costs relative to the case
in which he is required to hold capital constant.

The envelope theorem resolves the puzzle. To see the relationship between this theorem and Fig. 4 we have
to move into input space. See Fig. 5 which, note, is very similar to Fig. 3. Along the long-run average cost
curve, we have solved the cost-minization problem, so that the constrained envelope theorem holds. For
example, at the point (q, LRATC(q)) in Fig. 4, the input mix (k, `) is chosen to be at the point where the
isoquant corresponding to q is tangent to the iso-cost line. Now suppose as in the upper panel of Fig. 5,
you move from (k, `) to the new isoquant line corresponding to q + dq. To a first order approximation, the
new isoquant will necessarily be parallel to the old (because, again, the gradient is constant in a first order
approximation). As the figure illustrates, it doesn’t matter much which way you combine inputs in order to
get to the new isoquant: to a first order approximation, all input mixes result in the same cost increment.
In particular, if you produce the additional output entirely by increasing labor (as you do along the SRATC
curve), then the increment in your cost is, to a first order approximation the same as if you had increased
both inputs in the optimal proportions (as you do along the LRATC curve). In other words, the slopes of
your LRATC and SRATC curves are the same at q.
 ARE211, Fall 2004 9

Isocost line at q + dq (new optimal input mix)

PSfrag replacements
Isocost line at q + dq (increase only `)
q
q̄ Isoquant for q + dq
Average Cost increase only `
Output
optimal input adjustment
SRATC(·, k)
(k, `)
LRATC(·)
Isoquant for q Isocost line at q

k
k
`

Isocost line at q̄ + dq (new optimal input mix)

Isocost line at q̄ + dq (increase only `)
increase only `
Isoquant for q̄ + dq
(k, `0 )
Isocost line at q̄
line at q̄ + dq (new optimal input mix)
optimal input adjustment
Isocost line at q̄ + dq (increase only `) ¯
(k̄, `)
Isoquant for q̄ Isoquant for q̄
Isoquant for q̄ + dq
adjust both inputs Isocost line at q̄

k
k

F IGURE 5. The envelope theorem applied to Viner’s cost diagram.

Now suppose you are producing a higher level of output, q̄. In the bottom panel, we compare the case in
which we are producing with the original, now suboptimal level of capital k and labor ` 0 > `,¯ against the
¯ that would be optimal for this new higher level q̄. This situation is represented by the
level of inputs (k̄, `)
¯ to
lower panel of Fig. 5. As the figure indicates, when you move from the isoquant corresponding to ( k̄, `)
10 LECTURE #23: THU, DEC 2, 2004 PRINT DATE: DECEMBER 17, 2004 (COMPSTAT1)

the new isoquant line corresponding to q̄ + dq, it matters a lot whether you are starting from the optimal
¯ or starting from your original capital level k, which requires an input mix of (k, ` 0 ). There
input mix (k̄, `)
are two things to notice:

¯ you are on a lower isocost line than when you produce with
(1) Notice that when you produce with ( k̄, `)
(k, `0 ). This observation corresponds to the fact that in the Average Cost diagram (Fig. 4) above the
point q̄ the level of the short run average cost curve is higher than the level of the long run average
cost curve.
(2) Now observe that if you are initially producing q̄ and then you shift to producing q̄ + dq, if your
starting point is the suboptimal mix (k, ` 0 ), and you are allowed only to increase ` because k is fixed,
¯ and
then the increase in costs is much larger than if your starting point were the optimal mix ( k̄, `),
you adjust both inputs. This observation corresponds to the fact that in the Average Cost diagram
(Fig. 4) above the point q̄ the slope of the short run average cost curve is steeper than the slope of
the long run average cost curve.