CALIFORNIA INSTITUTE OF TECHNOLOGY

Division of the Humanities and Social Sciences

Notes on the Implicit Function Theorem

KC Border
v. 2013.08.27::14.08

1 Implicit Function Theorems

The Implicit Function Theorem is a basic tool for analyzing extrema of differ-
entiable functions.

Definition 1 An equation of the form

f (x, p) = y (1)
implicitly defines x as a function of p on a domain P if there is a function
ξ on P for which f (ξ(p), p) = y for all p ∈ P . It is traditional to assume that
y = 0, but not essential.
The use of zero in the above equation serves to simplify notation. The
condition f (x, p) = y is equivalent to g(x, p) = 0 where g(x, p) = f (x, p) − y,
and this transformation of the problem is common in practice.
The implicit function theorem gives conditions under which it is possible to
solve for x as a function of p in the neighborhood of a known solution (x̄, p̄).
There are actually many implicit function theorems. If you make stronger as-
sumptions, you can derive stronger conclusions. In each of the theorems that
follows we are given a subset X of Rn , a metric space P (of parameters), a func-
tion f from X × P into Rn , and a point (x̄, p̄) in the interior of X × P such that
Dx f (x̄, p̄) exists and is invertible. Each asserts the existence( of neighborhoods
)
U of x̄ and W of p̄ and a function ξ : W → U such that f ξ(p), p = f (x̄, p̄)
for all p ∈ W . They differ in whether ξ is uniquely defined (in U ) and how
smooth it is. The following table serves as a guide to the theorems. For ease of
reference, each theorem is stated as a standalone result.

Theorem Hypotheses Conclusion

( ) ( )
All f is continuous on X × P f ξ(p), p = f x̄, p̄ for all p in W
Dx f (x̄, p̄) is invertible ξ(p̄) = x̄
2 ξ is continuous at p̄
3 Dx f is continuous on X × P ξ is unique in U
ξ is continuous on W
4 Df (x̄, p̄) (wrt x, p) exists ξ is differentiable at p̄
5 Df (wrt x, p) exists on X × P ξ is unique in U
Dx f is continuous on X × P ξ is differentiable on W
1 f is C k on X × P ξ is unique in U
ξ is C k on W

1
KC Border Notes on the Implicit Function Theorem 2

The first result is due to Halkin [13, Theorem B].

Theorem 2 (Implicit Function Theorem 0) Let X be a subset of Rn , let

P be a metric space, and let f : X × P → Rn be continuous. Suppose the
derivative Dx f of f with respect to x exists at a point and that Dx f (x̄, p̄) is
invertible. Let
ȳ = f (x̄, p̄).
Then for any neighborhood U of x̄, there is a neighborhood W of p̄ and a
function ξ : W → U such that:
a. ξ(p̄) = x̄.
( )
b. f ξ(p), p = ȳ for all p ∈ W .
c. ξ is continuous at the point p̄.

However, it may be that ξ is neither continuous nor uniquely defined on any

neighborhood of p̄. There are two ways to strengthen the hypotheses and derive
a stronger conclusion. One is to assume the derivative with respect to x exists
and is continuous on X × P . The other is to make P a subset of a Euclidean
space and assume that f has a derivative with respect to (x, p) at the single
point (x̄, p̄).
Taking the first approach allows us to conclude that the function ξ is uniquely
defined and moreover continuous. The following result is Theorem 9.3 in Loomis
and Sternberg [19, pp. 230–231].

Theorem 3 (Implicit Function Theorem 1a) Let X be an open subset Rn ,

let P be a metric space, and let f : X × P → Rn be continuous. Suppose the
derivative Dx f of f with respect to x exists at each point (x, p) and is continuous
on X × P . Assume that Dx f (x̄, p̄) is invertible. Let

ȳ = f (x̄, p̄).

Then there are neighborhoods U ⊂ X and W ⊂ P of x̄ and p̄, and a function

ξ : W → U such that:
a. f (ξ(p); p) = ȳ for all p ∈ W .

b. For each p ∈ W , ξ(p) is the unique solution to (1) lying in U . In particu-

lar, then
ξ(p̄) = x̄.

c. ξ is continuous on W .

The next result, also due to Halkin [13, Theorem E] takes the second ap-
proach. It concludes that ξ is differentiable at a single point. Related results
may be found in Hurwicz and Richter [14, Theorem 1], Leach [16, 17], Nijen-
huis [21], and Nikaidô [22, Theorem 5.6, p. 81].

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 3

Theorem 4 (Implicit Function Theorem 1b) Let X be a subset of Rn , let

P be an open subset of Rm , and let f : X × P → Rn be continuous. Suppose the
derivative Df of f with respect to (x, p) exists at (x̄, p̄). Write Df (x̄, p̄) = (T, S),
where T : Rn → Rn and S : Rm → Rm , so that Df (x̄, p̄)(h, z) = T h + Sz.
Assume T is invertible. Let
ȳ = f (x̄, p̄).
Then there is a neighborhood W of p̄ and a function ξ : W → X satisfying
a. ξ(p̄) = x̄.
( )
b. f ξ(p), p = ȳ for all p ∈ W .
c. ξ is differentiable (hence continuous) at p̄, and

Dξ(p̄) = −T −1 ◦ S.

The following result is Theorem 9.4 in Loomis and Sternberg [19, p. 231].
It strengthens the hypotheses of both Theorems 3 and 4. In return we get
differentiability of ξ on W .

Theorem 5 (Semiclassical Implicit Function Theorem) Let X ×P be an

open subset of Rn × Rm , and let f : X × P → Rn be differentiable. Suppose
the derivative Dx f of f with respect to x is continuous on X × P . Assume that
Dx f (x̄, p̄) is invertible. Let
ȳ = f (x̄, p̄).
Then there are neighborhoods U ⊂ X and W ⊂ P of x̄ and p̄ on which
equation (1) uniquely defines x as a function of p. That is, there is a function
ξ : W → U such that:
a. f (ξ(p); p) = ȳ for all p ∈ W .
b. For each p ∈ W , ξ(p) is the unique solution to (1) lying in U . In particu-
lar, then
ξ(p̄) = x̄.

c. ξ is differentiable on W , and
   −1  
∂ξ 1 ∂ξ 1 ∂f 1 ∂f 1 ∂f 1 ∂f 1
··· ··· ···
 ∂p1 ∂pm   ∂x1 ∂xn   ∂p1 ∂pm 
     
 .. ..  = − .. ..   .. .. .
 . .   . .   . . 
 ∂ξ n ∂ξ n   ∂f n ∂f n   ∂f n ∂f n 
··· ··· ···
∂p1 ∂pm ∂x1 ∂xn ∂p1 ∂pm

The classical version maybe found, for instance, in Apostol [3, Theorem 7-6,
p. 146], Rudin [24, Theorem 9.28, p. 224], or Spivak [26, Theorem 2-12, p. 41].
Some of these have the weaker statement that there is a unique function ξ within
the class of continuous functions satisfying both ξ(p̄) = x̄ and f (ξ(p); p) = 0

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 4

for all p. Dieudonné [8, Theorem 10.2.3, p. 272] points out that the C k case
follows from the formula for Dξ and the fact that the mapping from invertible
linear transformations to their inverses, A 7→ A−1 , is C ∞ . (See Marsden [20,
Lemma 2, p. 231].)

Classical Implicit Function Theorem Let X × P be an open subset of

Rn × Rm , and let f : X × P → Rn be C k , for k ⩾ 1. Assume that Dx f (x̄, p̄) is
invertible. Let
ȳ = f (x̄, p̄).
Then there are neighborhoods U ⊂ X and W ⊂ P of x̄ and p̄ on which
equation (1) uniquely defines x as a function of p. That is, there is a function
ξ : W → U such that:
a. f (ξ(p); p) = ȳ for all p ∈ W .

b. For each p ∈ W , ξ(p) is the unique solution to (1) lying in U . In particu-

lar, then
ξ(p̄) = x̄.

c. ξ is C k on W , and
   −1  
∂ξ 1 ∂ξ 1 ∂f 1 ∂f 1 ∂f 1 ∂f 1
··· ··· ···
 ∂p1 ∂pm   ∂x1 ∂xn   ∂p1 ∂pm 
     
 .. ..  = − .. ..   .. .. .
 . .   . .   . . 
 ∂ξ n ∂ξ n   ∂f n ∂f n   ∂f n ∂f n 
··· ··· ···
∂p1 ∂pm ∂x1 ∂xn ∂p1 ∂pm

As a bonus, let me throw in the following result, which is inspired by Apos-

tol [4, Theorem 7.21].

Theorem 6 (Lipschitz Implicit Function Theorem) Let P be a compact

metric space and let f : R × P → R be continuous and assume that there are
real numbers 0 < m < M such that for each p

f (x, p) − f (y, p)
m⩽ ⩽ M.
x−y
( )
Then there is a unique function ξ : P → R satisfying f ξ(p), p = 0. Moreover,
ξ is continuous.

An interesting extension of this result to Banach spaces and functions with

compact range may be found in Warga [28].

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 5

1.1 Proofs of Implicit Function Theorems

The proofs given here are based on fixed point arguments and are adapted from
Halkin [12, 13], Rudin [24, pp. 220–227], Loomis and Sternberg [19, pp. 229–
231], Marsden [20, pp. 230–237], and Dieudonné [8, pp. 265—273]. Another
sort of proof, which is explicitly finite dimensional, of the classical case may be
found in Apostol [3, p. 146] or Spivak [26, p. 41].
The first step is to show that for each p (at least in a neighborhood of p̄)
there is a zero of the function f (x, p) − ȳ, where ȳ = f (x̄, p̄). As is often the
case, the problem of finding a zero of f − ȳ is best converted to the problem
of finding a fixed point of some other function. The obvious choice is to find a
fixed point of πX − (f − ȳ) (where πX (x, p) = x), but the obvious choice is not
clever enough in this case. Let

T = Dx f (x̄, p̄).
( )
Define φ : X × P → Rn by φ = πX − T −1 f − ȳ . That is,
( )
φ(x, p) = x − T −1 f (x, p) − ȳ . (2)
( )
Note that φ(x, p) = x if and only if T −1 f (x, p)− ȳ = 0. But the invertibility of
T −1 guarantees that this happens if and only if f (x, p) = ȳ. Thus the problem
of finding a zero of f (·, p) − ȳ is equivalent to that of finding a fixed point of
φ(·, p). Note also that
φ(x̄, p̄) = x̄. (3)
Observe that φ is continuous and also has a derivative Dx φ with respect to x
whenever f does. In fact,

Dx φ(x, p) = I − T −1 Dx f (x, p).

In particular, at (x̄, p̄), we get

Dx φ(x̄, p̄) = I − T −1 T = 0. (4)

That is, Dx φ(x̄, p̄) is the zero transformation.

Recall that for a linear transformation A, its operator norm ∥A∥ is defined
by ∥A∥ = sup|x|⩽1 |Ax|, and satisfies |Ax| ⩽ ∥A∥ · |x| for all x. If A is invertible,
then ∥A−1 ∥ > 0.

Proof of Theorem 2: Let X, P , and f : X × P → Rn be as in the hypotheses of

Theorem 2.
In order to apply a fixed point argument, we must first find a subset of X
that is mapped into itself. By the definition of differentiability and (4) we can
choose r > 0 so that
φ(x, p̄) − φ(x̄, p̄) 1
⩽ for all x ∈ B̄r (x̄).
|x − x̄| 2

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 6

Noting that φ(x̄, p̄) = x̄ and rearranging, it follows that

r
φ(x, p̄) − x̄ ⩽ for all x ∈ B̄r (x̄).
2
For each p set m(p) = maxx φ(x, p) − φ(x, p̄) as x runs over the compact
set B̄r (x̄). Since φ is continuous (and B̄r (x̄) is a fixed set), the Maximum
Theorem ?? implies that m is continuous. Since m(p̄) = 0, there is some ε > 0
such that |m(p)| < 2r for all p ∈ Bε (p̄). That is,
r
φ(x, p) − φ(x, p̄) < for all x ∈ B̄r (x̄), p ∈ Bε (p̄).
2
For each p ∈ Bε (p̄), the function φ maps B̄r (x̄) into itself, for

φ(x, p) − x̄ ⩽ φ(x, p) − φ(x, p̄) + φ(x, p̄) − x̄

r r
< +
2 2
= r.

That is, φ(x, p) ∈ B̄r (x̄). Since φ is continuous and B̄r (x̄) is compact and
convex, by the Brouwer Fixed Point Theorem (e.g., [5, Corollary 6.6, p. 29]),
there is some x ∈ B̄r (x̄) satisfying φ(x, p) = x, or in other words f (x, p) = 0.
We have just proven parts (a) and (b) of Theorem 2. That is, for every
neighborhood X of x̄, there is a neighborhood W = B̄ε(r) (p̄) of p̄ and a function
( )
ξ from B̄ε(r) (p̄) into B̄r (x̄) ⊂ X satisfying ξ(p̄) = x̄ and f ξ(p), p = 0 for all
p ∈ W . (Halkin actually breaks this part out as Theorem A.)
We can use the above result to construct a ξ that is continuous at p̄. Start
with a given neighborhood U of x̄. Construct a sequence of r1 > r2 > · · · > 0
satisfying lim rn = 0 and for each n consider the neighborhood Un = U ∩Brn (x̄).
From the argument above there is a neighborhood ( Wn of) p̄ and a function ξn
from Wn into Un ⊂ U satisfying ξn (p̄) = x̄ and f ξn (p), p = 0 for all p ∈ Wn .
Without loss of generality we may assume Wn ⊃ Wn+1 (otherwise replace Wn+1
with Wn ∩ Wn+1 ), so set W = W1 . Define ξ : W → U by ξ(p) (= ξn (p) ) for
p ∈ Wn \ Wn+1 . Then ξ is continuous at p̄, satisfies ξ(p̄) = x̄, and f ξ(p), p = 0
for all p ∈ W .

Note that the above proof used in an essential way the compactness of B̄r (x̄),
which relies on the finite dimensionality of Rn . The compactness was used first
to show that m(p) is finite, and second to apply the Brouwer fixed point theorem.
Theorem 3 adds to the hypotheses of Theorem 2. It assumes that Dx f exists
everywhere on X × P and is continuous. The conclusion is that there are some
neighborhoods U of x̄ and W of p̄ (and a continuous
) function ξ : W → U such
that ξ(p) is the unique solution to f ξ(p), p = 0 lying in U . It is the uniqueness
of ξ(p) that puts a restriction on U . If U is too large, say U = X, then the
solution need not be unique. (On the other hand, it is easy to show, as does
Dieudonné [8, pp. 270–271], there is at most one continuous ξ, provided U is
connected.) The argument we use here, which resembles that of Loomis and

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 7

Sternberg, duplicates some of the proof of Theorem 2, but we do not actually

need to assume that the domain of f lies in the finite dimensional space Rn ×Rm ,
any Banach spaces will do, and the proof need not change. This means that
we cannot use Brouwer’s theorem, since closed balls are not compact in general
Banach spaces. Instead, we will be able to use the Mean Value Theorem and
the Contraction Mapping Theorem.

Proof of Theorem 3: Let X, P , and let f : X × P → Rn obey the hypotheses of

Theorem 3. Set T = Dx f (x̄, p̄), and recall that T is invertible.
Again we must find a suitable subset of X so that each φ(·, p) maps this set
into itself. Now we use the hypothesis that Dx f (and hence Dx φ) exists and
is continuous on X × P to deduce that there is a neighborhood B̄r (x̄) × W1 of
(x̄, p̄) on which the operator norm ∥Dx φ∥ is strictly less than 12 . Set U = B̄r (x̄).
Since φ(x̄, p̄) = x̄ and since φ is continuous (as f is), we can now choose W so
that p̄ ∈ W , W ⊂ W1 , and p ∈ W implies
r
φ(x̄, p) − x̄ < .
2
We now show that for each p ∈ W , the mapping x 7→ φ(x, p) is a contraction
that maps B̄r (x̄) into itself. To see this, note that the Mean Value Theorem (or
Taylor’s Theorem) implies

φ(x, p) − φ(y, p) = Dx φ(z, p)(x − y),

for some z lying on the segment between x and y. If x and y lie in B̄r (x̄), then
z too must lie in B̄r (x̄), so ∥Dx φ(z, p)∥ < 12 . It follows that

1
φ(x, p) − φ(y, p) < |x − y| for all x, y ∈ B̄r (x̄), p ∈ Bε (p̄), (5)
2
so φ(·, p) is a contraction on B̄r (x̄) with contraction constant 21 .
To see that B̄r (x̄) is mapped into itself, let (x, p) belong to B̄r (x̄) × W and
observe that

φ(x, p) − x̄ ⩽ φ(x, p) − φ(x̄, p) + φ(x̄, p) − x̄

1 r
< x − x̄ +
2 2
< r.

Thus φ(x, p) ∈ B̄r (x̄).

Since B̄r (x̄) is a closed subset of the complete metric space Rn , it is complete
itself, so the Contraction Mapping Theorem guarantees that there is a unique
fixed point of φ(·, p) in B̄r (x̄). In other words, for each p ∈ W there is a unique
point ξ(p) lying in U = B̄r (x̄) satisfying f (ξ(p), p) = 0.
It remains to show that ξ must be continuous on W . This follows from
a general result on parametric contraction mappings, presented as Lemma 7
below, which also appears in [19, Corollary 4, p. 230].

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 8

Note that the above proof nowhere uses the finite dimensionality of Rm , so
the theorem actually applies to a general Banach space.

Proof of Theorem 4: For this theorem, in addition to the hypotheses of The-

orem 2, we need P to be a subset of a Euclidean space (or more generally a
Banach space), so that it makes sense to partially differentiate with respect to
p. Now assume f is differentiable with respect to (x, p) at the point (x̄, p̄).
There is a neighborhood W of p̄ and a function ξ : W → X satisfying the
conclusions of Theorem 2. It turns out that under the added hypotheses, such
a function ξ is differentiable at p̄.
We start by showing that ξ is locally Lipschitz continuous at p̄. First set

∆(x, p) = f (x, p) − f (x̄, p̄) − T (x − x̄) − S(p − p̄).

Since Df exists at (x̄, p̄), there exists r > 0 such that Br (x̄) × Br (p̄) ⊂ X × W
and if |x − x̄| < r and |p − p̄| < r, then

∆(x, p) 1
< ,
|x − x̄| + |p − p̄| 2 ∥T −1 ∥
which in turn implies
1 1
T −1 ∆(x, p) < |x − x̄| + |p − p̄|.
2 2
Since ξ is continuous at p̄ and ξ(p̄) = x̄, there is some r ⩾ δ > 0 such that
|p − p̄| < δ implies |ξ(p) − x̄| < r. Thus
( ) 1 1
T −1 ∆ ξ(p), p < ξ(p) − x̄ + |p − p̄| for all p ∈ Bδ (p̄). (6)
2 2
( )
But f ξ(p), p − f (x̄, p̄) = 0 implies
( ) ( )
T −1 ∆ ξ(p), p = ξ(p) − x̄ + T −1 S(p − p̄) . (7)

Therefore, from the facts that |a + b| < c implies |a| < |b| + c, and ξ(p̄) = x̄,
equations (6) and (7) imply
1 1
ξ(p) − ξ(p̄) < T −1 S(p − p̄) + ξ(p) − ξ(p̄) + |p − p̄| for all p ∈ Bδ (p̄)
2 2
or, ( )
ξ(p) − ξ(p̄) < 2∥T −1 S∥ + 1 |p − p̄| for all p ∈ Bδ (p̄).
That is, ξ satisfies a local Lipschitz condition at p̄. For future use set M =
2∥T −1 S∥ + 1.
Now we are in a position to prove that −T −1 S is the differential of ξ at p̄.
Let ε > 0 be given. Choose 0 < r < δ so that |x − x̄| < r and |p − p̄| < r implies

∆(x, p) ε
< ,
|x − x̄| + |p − p̄| (M + 1) ∥T −1 ∥

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 9

so
( ) ( ) ε ( )
ξ(p)−ξ(p̄) +T −1 S(p−p̄) = T −1 ∆ ξ(p), p < |ξ(p)−ξ(p̄)|+|p−p̄| ⩽ ε|p−p̄|,
(M + 1)

for |p − p̄| < r, which shows that indeed −T −1 S is the differential of ξ at p̄.

Proof of Theorem 5: ************

Proof of Theorem 1: ************

Proof of Theorem 6: Let f satisfy the hypotheses of the theorem. Let C(P )
denote the set of continuous real functions on P . Then C(P ) is complete under
the uniform norm metric, ∥f − g∥ = supp |f (p) − g(p)| [2, Lemma 3.97, p. 124].
For each p define the function ψp : R → R by
1
ψp (x) = (x) −
f (x, p).
M
Note that ψp (x) = x if and only if f (x, p) = 0. If ψp has a unique fixed
point,
( then
) we shall have shown that there is a unique function ξ satisfying
f ξ(p), p = 0. It suffices to show that ψp is a contraction.
To see this, write
f (x, p) − f (y, p)
ψp (x) − ψp (y) = x−y−
( M )
1 f (x, p) − f (y, p)
= 1− (x − y).
M x−y
By hypothesis
f (x, p) − f (y, p)
0<m⩽ ⩽ M,
x−y
so ( m)
|ψp (x) − ψp (y)| ⩽ 1 − |x − y|.
M
This shows that ψp is a contraction with constant 1 − M m
< 1.
To see that ξ is actually continuous, define the function ψ : C(P ) → C(P )
via
1 ( )
ψg(p) = g(p) − f g(p), p .
M
(Since f is continuous, ψg is continuous whenever g is continuous.) The point-
(wise argument
) above is independent of p, so it also shows that |ψg(p)−ψh(p)| ⩽
1− M m
|g(p) − h(p)| for any functions g and h. Thus
( m)
∥ψg − ψh∥ ⩽ 1 − ∥g − h∥.
M
In other words ψ is a contraction on C(P ), so( it has )a unique fixed point ḡ in
C(P ), so ḡ is continuous. But ḡ also satisfies f ḡ(p), p , but since ξ(p) is unique
we have ξ = ḡ is continuous.

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 10

x
f (x, p) > 0
(x1 , p1 ) f (x, p) = 0

f′ (x3 , p3 )
(x2 , p2 )
f (x, p) < 0 p

Figure 1. Looking for implicit functions.

Lemma 7 (Continuity of fixed points) Let φ : X × P → X be continuous

in p for each x, where X is a complete metric space under the metric d and P
is a metrizable space. Suppose that φ is a uniform contraction in x. That is,
there is some 0 ⩽ α < 1 such that
( )
d φ(x, p) − φ(y, p) ⩽ αd(x, y)

for all x and y in X and all p in P . Then (the mapping) ξ : P → X from p to

the unique fixed point of φ(·, p), defined by φ ξ(p), p = ξ(p), is continuous.

Proof : Fix a point p in P and let ε > 0 be given. Let ρ be a compatible metric
on P and using the continuity of φ(x, ·) on P , choose δ > 0 so that ρ(p, q) < δ
implies that ( ( ) ( ))
d φ ξ(p), p , φ ξ(p), q < (1 − α)ε.
So if ρ(p, q) < δ, then
( ) ( ( ) ( ))
d ξ(p), ξ(q) = d φ ξ(p), p , φ ξ(q), q
( ( ) ( )) ( ( ) ( ))
⩽ d φ ξ(p), p , φ ξ(p), q + d φ ξ(p), q , φ ξ(q), q
( )
< (1 − α)ε + αd ξ(p), ξ(q)

so ( )
(1 − α)d ξ(p), ξ(q) < (1 − α)ε
or ( )
d ξ(p), ξ(q) < ε,
which proves that ξ is continuous at p.

1.2 Examples
Figure 1 illustrates the Implicit Function Theorem for the special case n = m =
1, which is the only one I can draw. The figure is drawn sideways since we
are looking for x as a function of p. In this case, the requirement that the
differential with respect to x be invertible reduces to ∂f ∂x ̸= 0. That is, in the
diagram the gradient of f may not be horizontal. In the figure, you can see
that the points, (x1 , p1 ), (x2 , p2 ), and (x3 , p3 ), the differentials Dx f are zero.
At (x1 , p1 ) and (x2 , p2 ) there is no way to define x as a continuous function of

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 11

p locally. (Note however, that if we allowed a discontinuous function, we could

define x as a function of p in a neighborhood of p1 or p2 , but not uniquely.) At
the point (x3 , p3 ), we can uniquely define x as a function of p near p3 , but this
function is not differentiable.
Another example of the failure of the conclusion of the Classical Implicit
Function Theorem is provided by the function from Example ??.

Example 8 (Differential not invertible) Define f : R × R → R by

f (x, p) = −(x − p2 )(x − 2p2 ).

Consider the function implicitly defined by f (x, p) = 0. The function f is

zero along the parabolas x = p2 and x = 2p2 , and in particular f (0, 0) = 0.
See Figure 2 on page 18. The hypothesis of the Implicit Function Theorem is
not satisfied since ∂f∂x
(0,0)
= 0. The conclusion also fails. The problem here
is not that a smooth implicit function through (x, p) = (0, 0) fails to exist.
The problem is that it is not unique. There are four distinct continuously
differentiable implicitly defined functions. □

Example 9 (Lack of continuous differentiability) Consider again the func-

tion h(x) = x + 2x2 sin x12 from Example ??. Recall that h is differentiable
everywhere, but not continuously differentiable at zero. Furthermore, h(0) = 0,
h′ (0) = 1, but h is not monotone on any neighborhood of zero. Now consider
the function f (x, p) = h(x) − p. It satisfies f (0, 0) = 0 and ∂f∂x
(0,0)
̸= 0, but it
there is no unique implicitly defined function on any neighborhood, nor is there
any continuous implicitly defined function.
To see this, note that f (x, p) = 0 if and only if h(x) = p. So a unique
implicitly defined function exists only if h is invertible on some neighborhood of
zero. But this is not so, for given any ε > 0, there is some 0 < p < 2ε for which
there are 0 < x < x′ < ε satisfying ( h(x)
′
) = h(x ) = p. It is also easy to see that
no continuous function satisfies h ξ(p) = p either. □
If x is more than one dimensional there are subtler ways in which Dx f
may fail to be continuous. The next example is taken from Dieudonné [8,
Problem 10.2.1, p. 273].

Example 10 Define f : R2 → R2 by

f 1 (x, y) = x

and 



 y − x2 0 ⩽ x2 ⩽ y





f 2 (x, y) = y 2 − x2 y
 0 ⩽ y ⩽ x2

 x2





−f 2 (x, −y) y ⩽ 0.

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 12

Then f is everywhere differentiable on R2 , and Df (0, 0) is the identity mapping, Work out the details.

but Df is not continuous at the origin. Furthermore in every neighborhood of

the origin there are distinct points (x, y) and (x′ , y ′ ) with f (x, y) = f (x′ , y ′ ).
Thus f has no local inverse, so the equation f (x, y) − p = f (0, 0) does not
uniquely define a function. □

1.3 Implicit vs. inverse function theorems

In this section we discuss the relationship between the existence of a unique
implicitly defined function and the existence of an inverse function. These results
are quite standard and may be found, for instance, in Marsden [20, p. 234].
First we show how the implicit function theorem can be used to prove an
inverse function theorem. Suppose X ⊂ Rn and g : X → Rn . Let P be a
neighborhood of p̄ = g(x̄). Consider f : X × P → Rn defined by

f (x, p) = g(x) − p.

Then f (x, p) = 0 if and only if p = g(x). Thus if( there )is a unique implicitly
defined function ξ : P → X implicitly defined by f ξ(p), p = 0, it follows that g
is invertible and ξ = g −1 . Now compare the Jacobian matrix of f with respect
to x and observe that it is just the Jacobian matrix of g. Thus each of the
implicit function theorems has a corresponding inverse function theorem.
We could also proceed in the other direction, as is usually the case in text-
books. Let X × P be a subset of Rn × Rm , and let f : X × P → Rn , and
suppose f (x̄, p̄) = 0. Define a function g : X × P → Rn × P by
( )
g(x, p) = f (x, p), p .

Suppose g is invertible, that is, it is one-to-one and onto. Define ξ : P → X by

( )
ξ(p) = πx g −1 (0, p) .

Then ξ is the unique function implicitly defined by

( ( ) )
f ξ(p), p , p = 0

for all p ∈ P . Now let’s compare hypotheses. The standard Inverse Function
Theorem, e.g. [20, Theorem 7.1.1, p. 206], says that if g is continuously dif-
ferentiable and has a nonsingular Jacobian matrix at some point, then there
is a neighborhood of the point where g is invertible. The Jacobian matrix for

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 13

( )
g(x, p) = f (x, p), p above is
 
∂f 1 ∂f 1 ∂f 1 ∂f 1
 ∂x ··· ··· 
 1 ∂xn ∂p1 ∂pm 
 .. .. .. .. 
 . . . . 
 
 ∂f n ∂f n ∂f n ∂f n 
 ··· ··· 
 ∂x1 ∂xn ∂p1 ∂pm .
 
 
 
 0 ··· 0 1 0 
 
 .. .. .. 
 . . . 
0 ··· 0 0 1
Since this is block diagonal, it is easy to see that this Jacobian matrix is non-
singular at (x̄, p̄) if and only if the derivative Dx f (x̄, p̄), is invertible.

1.4 Global inversion

The inverse function theorems proven above are local results. Even if the Jaco-
bian matrix of a function never vanishes, it may be that the function does not
have an inverse everywhere. The following example is well known, see e.g., [20,
Example 7.1.2, p. 208]. Add a section on the
Gale–Nikaidô
Theorem.

Example 11 (A function without a global inverse) Define f : R2 → R2

via
f (x, y) = (ex cos y, ex sin y).
Then the Jacobian matrix is
( x )
e cos y −ex sin y
ex sin y ex cos y

which has determinant e2x (cos2 y + sin2 y) = e2x > 0 everywhere. Nonetheless,
f is not invertible since f (x, y) = f (x, y + 2π) for every x and y. □

2 Applications of the Implicit Function Theo-

rem
2.1 A fundamental lemma
A curve in Rn is simply a function from an interval of R into Rn , usually
assumed to be continuous.
Fundamental Lemma on Curves Let U be an open set in Rn and let
∗ ∗
g : U → R . Let x ∈ U satisfy g(x ) = 0, and suppose g is differentiable
m

at x∗ . Assume that g1 ′ (x∗ ), . . . , gm ′ (x∗ ) are linearly independent. Let v ∈ Rn

satisfy
gi ′ (x∗ ) · v = 0, i = 1, . . . , m.

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 14

Then there exists δ > 0 and a curve x̂ : (−δ, δ) → U satisfying:

1. x̂(0) = x∗ .
( )
2. g x̂(α) = 0 for all α ∈ (−δ, δ).
3. x̂ is differentiable at 0. Moreover, if g is C k on U , then x̂ is C k on (−δ, δ).
4. x̂′ (0) = v.

Proof : Since the gi ′ (x∗ )s are linearly independent, n ⩾ m, and without loss of
generality, we may assume the coordinates are numbered so that the m × m
matrix  
∂g 1 ∂g 1
···
 ∂x1 ∂xm 
 . .. 
 . 
 . . 
 ∂g m ∂g m 
···
∂x1 ∂xm
is invertible at x∗ .
Fix v satisfying gi ′ (x∗ ) · v = 0 for all i = 1, . . . , m. Rearranging terms we
have
∑m
∂g i (x∗ ) ∑ n
∂g i (x∗ )
· vj = − · vj i = 1, . . . , m,
j=1
∂xj j=m+1
∂xj

or in matrix terms
   
∂g 1 ∂g 1   ∂g 1 ∂g 1  
···  ...  vm+1
 ∂x1 ∂xm  v1  ∂xm+1 ∂xn 
 .  . 
 . ..  .  = −
 .. ..   .. 
 . ,
 . .  .  . . 
 ∂g m ∂g m   ∂g m ∂g m  vn
vm
··· ...
∂x1 ∂xm ∂xm+1 ∂xn
so
 −1  ∂g 1 ∂g 1

  ∂g 1 ∂g 1  
··· ...
v1     vm+1
  
∂x1 ∂xm ∂xm+1 ∂xn
 ..  
 .  = −  .. ..    .. ..   .. 
 . .
 . .   . . 
 ∂g m ∂g m   ∂g m ∂g m  vn
vm
··· ...
∂x1 ∂xm ∂xm+1 ∂xn

Observe that these conditions completely characterize v. That is, for any y ∈
Rn ,
( ′ ∗ )
gi (x ) · y = 0, i = 1, . . . , m, and yj = vj , j = m+1, . . . , n =⇒ y = v.
(8)
Define the C ∞ function f : Rm × R → Rn by

f (z, α) = (z1 , . . . , zm , x∗m+1 + αvm+1 , . . . , x∗n + αvn ).

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 15

Set z ∗ = (x∗1 , . . . , x∗m ) and note that f (z ∗ , 0) = x∗ . Since x∗ is an interior point

of U , there is a neighborhood W of z ∗ and an interval (−η, η) in R so that for
every z ∈ W and α ∈ (−η, η), the point f (z, α) belongs to U ⊂ Rn . Finally,
define h : W × (−η, η) → Rm by
( )
h(z, α) = g f (z, α) = g(z1 , . . . , zm , x∗m+1 + αvm+1 , . . . , x∗n + αvn ).

Observe that h possesses the same degree of differentiability as g, since h is the

composition of g with the C ∞ function f .
i ∂gi
Then for j = 1, . . . , m, we have = ∂h ∂zj (z, 0) = ∂xj (x), where z = (x1 , . . . , xm ).
Therefore the m-vectors
 
∂hi (z ∗ ; 0)
 
 ∂z1 
 .. 
 .  , i = 1, . . . , m
 
 ∂hi (z ∗ ; 0) 
∂zm
are linearly independent.
But h(z ∗ , 0) = 0, so by the Implicit Function Theorem 4 there is an inter-
val (−δ, δ) ⊂ (−η, η) about 0, a neighborhood V ⊂ W of z ∗ and a function
ζ : (−δ, δ) → V such that
ζ(0) = z ∗ ,
( )
h ζ(α), α = 0, for all α ∈ (−δ, δ),
and ζ is differentiable at 0. Moreover, by Implicit Function Theorem 1, if g is
C k on U , then h is C k , so ζ is C k on (−δ, δ).
Define the curve x̂ : (−δ, δ) → U by
( ) ( )
x̂(α) = f ζ(α), α = ζ1 (α), . . . , ζm (α), x∗m+1 + αvm+1 , . . . , x∗n + αvn . (9)

Then x̂(0) = x∗ , ( )
g x̂(α) = 0 for all α ∈ (−δ, δ),
and x̂ is differentiable at 0, and if g is C k , then x̂ is C k . So by the Chain Rule,

gi ′ (x∗ ) · x̂′ (0) = 0, i = 1, . . . , m.

Now by construction (9), x̂′j (0) = vj , for j = m+1, . . . , n. Thus (8) implies
x̂′ (0) = v.

2.2 A note on comparative statics

“Comparative statics” analysis tells us how equilibrium values of endogenous
variables x1 , . . . , xn (the things we want to solve for) change as a function of the
exogenous parameters p1 , . . . , pm . (As such it is hardly unique to economics.)

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 16

Typically we can write the equilibrium conditions of our model as the zero of a
system of equations in the endogenous variables and the exogenous parameters:

F 1 (x1 , . . . , xn ; p1 , . . . , pm ) = 0
.. (10)
.
n
F (x1 , . . . , xn ; p1 , . . . , pm ) = 0

This implicitly defines x as a function of p, which we will explicitly denote

x = ξ(p), or
( )
(x1 , . . . , xn ) = ξ1 (p1 , . . . , pm ), . . . , ξn (p1 , . . . , pm ) .

This explicit function, if it exists, satisfies the implicit definition

( )
F ξ(p); p = 0 (11)

for at least a rectangle of values of p. The Implicit Function Theorem tells that
such an explicit function exists whenever it is possible to solve for all its partial
derivatives. ( )
Setting G(p) = F ξ(p); p , and differentiating Gi with respect to pj , yields,
by equation (11),
∑ ∂F i ∂ξ k ∂F i
+ =0 (12)
∂xk ∂pj ∂pj
k

for each i = 1, . . . , n, j = 1, . . . , m. In matrix terms we have

    
∂F 1 ∂F 1 ∂ξ 1 ∂ξ 1 ∂F 1 ∂F 1
··· ··· ···
 ∂x1 ∂xn  ∂p1 ∂pm   ∂p1 ∂pm 
    
 .. ..  .. .. + .. ..  = 0.
 . .  . .   . . 
 ∂F n ∂F n  ∂ξ n ∂ξ n   ∂F n ∂F n 
··· ··· ···
∂x1 ∂xn ∂p1 ∂pm ∂p1 ∂pm
(13)
 
∂F 1 ∂F 1
···
 ∂x1 ∂xn 
 . .. 
Provided 
 . . 
.  has an inverse (the hypothesis of the Implicit
 ∂F n ∂F n 
···
∂x1 ∂xn
Function Theorem) we can solve this:
   −1  
∂ξ 1 ∂ξ 1 ∂F 1 ∂F 1 ∂F 1 ∂F 1
··· ··· ···
 ∂p1 ∂pm   ∂x1 ∂xn   ∂p1 ∂pm 
     
 .. ..  = − .. ..   .. .. 
 . .   . .   . . 
 ∂ξ n ∂ξ n   ∂F n ∂F n   ∂F n ∂F n 
··· ··· ···
∂p1 ∂pm ∂x1 ∂xn ∂p1 ∂pm
(14)

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 17

The old-fashioned derivation (see, e.g., Samuelson [25, pp. 10–14]) of this
same result runs like this: “Totally differentiate” the ith row of equation (10)
to get
∑ ∂F i ∑ ∂F i
dxk + dpℓ = 0 (15)
∂xk ∂pℓ
k ℓ

for all i. Now set all dpℓ s equal to zero except pj , and divide by dpj to get
∑ ∂F i dxk ∂F i
+ =0 (16)
∂xk dpj ∂pj
k

for all i and j, which is equivalent to equation (12). For further information on
total differentials and how to manipulate them, see [3, Chapter 6].
Using Cramer’s Rule (e.g. [4, pp. 93–94]), we see then that

∂F 1 ∂F 1 ∂F 1 ∂F 1 ∂F 1
··· ···
∂x1 ∂xi−1 ∂pj ∂xi+1 ∂xn
.. .. .. .. ..
. . . . .
∂F n ∂F n ∂F n ∂F n ∂F n
··· ···
dxi ∂ξ i ∂x1 ∂xi−1 ∂pj ∂xi+1 ∂xn
= =− . (17)
dpj ∂pj ∂F 1 ∂F 1
···
∂x1 ∂xn
.. ..
. .
∂F n ∂F n
···
∂x1 ∂xn
 
∂F 1 ∂F 1
···
 ∂x1 ∂xn 
 . .. 

Or, letting ∆ denote the determinant of  .. 
. , and letting ∆i,j
 ∂F n ∂F n 
···
∂x1 ∂xn
denote the determinant of the matrix formed by deleting its i-th row and j-th
column, we have

∂ξ i ∑
n
∂F k ∆k,i
= − (−1)i+k . (18)
∂pj ∂pj ∆
k=1

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 18

f >0 f >0

f <0

f <0

Figure 2. f (x, y) = −(y − x2 )(y − 2x2 ).

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 19

References
[1] S. N. Afriat. 1971. Theory of maxima and the method of Lagrange. SIAM
Journal of Applied Mathematics 20:343–357.
http://www.jstor.org/stable/2099955
[2] C. D. Aliprantis and K. C. Border. 2006. Infinite dimensional analysis: A
hitchhiker’s guide, 3d. ed. Berlin: SpringerVerlag.
[3] T. M. Apostol. 1957. Mathematical analysis: A modern approach to ad-
vanced calculus. Addison-Wesley series in mathematics. Reading, Mas-
sachusetts: Addison Wesley.

[4] . 1969. Calculus, 2d. ed., volume 2. Waltham, Massachusetts:

Blaisdell.

[5] K. C. Border. 1985. Fixed point theorems with applications to economics

and game theory. New York: Cambridge University Press.
[6] J.-P. Crouzeix. 1981. Continuity and differentiability properties of quasi-
convex functions on Rn . In S. Schaible and W. Ziemba, eds., Generalized
Concavity in Optimization and Economics, pages 109–130. NY: Academic
Press.
[7] J.-P. Crouzeix and J. A. Ferland. 1982. Criteria for quasi-convexity and
pseudo-convexity: Relationships and comparisons. Mathematical Program-
ming 23:193–205. DOI: 10.1007/BF01583788

[8] J. Dieudonné. 1969. Foundations of modern analysis. Number 10-I in

Pure and Applied Mathematics. New York: Academic Press. Volume 1 of
Treatise on Analysis.
[9] E. E. Diewert, M. Avriel, and I. Zang. 1981. Nine kinds of quasiconcavity
and concavity. Journal of Economic Theory 25(3):397–420.
DOI: 10.1016/0022-0531(81)90039-9

[10] D. Gale and H. Nikaidô. 1965. Jacobian matrix and global univalence of
mappings. Mathematische Annalen 159(2):81–93.
DOI: 10.1007/BF01360282
[11] W. Ginsberg. 1973. Concavity and quasiconcavity in economics. Journal
of Economic Theory 6(6):596–605. DOI: 10.1016/0022-0531(73)90080-X
[12] H. Halkin. 1972. Necessary conditions for optimal control problems with
infinite horizons. Discussion Paper 7210, CORE.
[13] . 1974. Implicit functions and optimization problems without con-
tinuous differentiability of the data. SIAM Journal on Control 12(2):229–
236. DOI: 10.1137/0312017

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 20

[14] L. Hurwicz and M. K. Richter. 2003. Implicit functions and diffeomor-

phisms without C 1 . Advances in Mathematical Economics 5:65–96.

[15] F. John. 1948. Extremum problems with inequalities as subsidiary con-

ditions. In K. O. Friedrichs, O. E. Neugebauer, and J. J. Stoker, eds.,
Studies and Essays: Courant Anniversary Volume, pages 187–204. New
York: Interscience.

[16] E. B. Leach. 1961. A note on inverse function theorems. Proceedings of

the American Mathematical Society 12:694–697.
http://www.jstor.org/stable/2034858
[17] . 1963. On a related function theorem. Proceedings of the American
Mathematical Society 14:687–689.
http://www.jstor.org/stable/2034971

[18] A. Leroux. 1984. Other determinantal conditions for concavity and quasi-
concavity. Journal of Mathematical Economics 13(1):43–49.
DOI: 10.1016/0304-4068(84)90023-5

[19] L. H. Loomis and S. Sternberg. 1968. Advanced calculus. Reading, Mas-

sachusetts: Addison–Wesley.

[20] J. E. Marsden. 1974. Elementary classical analysis. San Francisco: W. H.

Freeman and Company.

[21] A. Nijenhuis. 1974. Strong derivatives and inverse mappings. American

Mathematical Monthly 81(9):969–980.

[22] H. Nikaidô. 1968. Convex structures and economic theory. Mathematics

in Science and Engineering. New York: Academic Press.

[23] K. Otani. 1983. A characterization of quasi-concave functions. Journal of

Economic Theory 31(1):194–196. DOI: 10.1016/0022-0531(83)90030-3

[24] W. Rudin. 1976. Principles of mathematical analysis, 3d. ed. International

Series in Pure and Applied Mathematics. New York: McGraw Hill.

[25] P. A. Samuelson. 1965. Foundations of economic analysis. New York:

Athenaeum. Reprint of the 1947 edition published by Harvard University
Press.
[26] M. Spivak. 1965. Calculus on manifolds. Mathematics Monograph Series.
New York: Benjamin.
[27] H. Uzawa. 1958. The Kuhn–Tucker conditions in concave programming.
In K. J. Arrow, L. Hurwicz, and H. Uzawa, eds., Studies in Linear and
Non-linear Programming, number 2 in Stanford Mathematical Studies in
the Social Sciences, chapter 3, pages 32–37. Stanford, California: Stanford
University Press.

v. 2013.08.27::14.08
KC Border Notes on the Implicit Function Theorem 21

[28] J. Warga. 1978. An implicit function theorem without differentiability.

Proceedings of the American Mathematical Society 69(1):65–69.
http://www.jstor.org/stable/2043190.pdf

v. 2013.08.27::14.08