18.950 Handout 4. Inverse and Implicit Function Theorems.
Theorem 1 (Inverse Function Theorem). Suppose U ⊂ Rn is open,
f : U → Rn is C 1 , x0 ∈ U and dfx0 is invertible. Then there exists a
neighborhood V of x0 in U and a neighborhood W of f (x0 ) in Rn such that
f has a C 1 inverse g = f −1 : W → V. (Thus f (g(y)) = y for all y ∈ W
and g(f (x)) = x for all x ∈ V .) Moreover,
dgy = (dfg(y) )−1 for all y ∈ W
and g is smooth whenever f is smooth.
Remark. The theorem says that a continuously differentiable function f
between regions in Rn is locally invertible near points where its differential
is invertible.
Proof. Without loss of generality, we may assume that x0 = 0, f (x0 ) = 0
and dfx0 = I. (Otherwise, replace f with f�(x) = dfx−1 0
(f (x + x0 ) − f (x0 )).
Note that if the theorem holds with f , 0, 0, I and a function g� in place of f
�
x0 , f (x0 ), dfx0 and g respectively, then it is easily verified that the theorem
as stated holds with g(y) = x0 + g�(dfx−1 0
(y − f (x0 ))).)
Since dfx is continuous in x at x0 (see Exercise 1), there exists a number
r > 0 such that
1
x ∈ B r (0) =⇒ �dfx − I� ≤ .
2
(Recall that for a linear transformation A : Rn → Rm we define the norm
of A by �A� = sup{|v|≤1} |A(v)|.) Fix y ∈ Br/2 (0). Define a function φ by
φ(x) = x − f (x) + y.
Note that dφx = I − dfx and hence
�dφx � ≤ 1/2 if x ∈ B r (0).
Thus
1�
d
|φ(x)| ≤ |φ(x) − y | + |y | = | φ(tx)dt| + |y |
0 dt
� 1 � 1
= | dφtx · xdt| + |y | ≤ �dφtx �|x|dt + |y |
0 0
≤ r/2 + r/2 = r (1)
1
whenever x ∈ B r (0). i.e. φ is a map from B r (0) into itself. For any x, z ∈
B r (0),
�� 1 �
� d �
|φ(z) − φ(x)| =
�
� φ(x + t(z − x))dt��
0 dt
� 1
≤ |dφx+t(z−x) · (z − x)|dt
0
� 1
≤ �dφx+t(z−x) �|z − x|dt
0
1
≤ |z − x|.
2
Thus φ : B r (0) → B r (0) is a contraction, and hence φ has a unique fixed
point xy ∈ B r (0). i.e. there is a unique point xy ∈ B r (0) with f (xy ) = y. In
fact xy ∈ Br (0) since 2r > |y | = |f (xy )| ≥ |xy | − |xy − f (xy )| ≥ |xy | − 21 |xy | =
1 −1 (W ) ∩ B (0). Note then that V is open.
2 |xy |. Set W = Br/2 (0) and V = f r
Define g : W → V by g(y) = xy . Then f (g(y)) = y for all y ∈ W and
g(f (x)) = x for all x ∈ V.
Next we show that g is differentiable, with dgy = (dfg(y) )−1 . First note
that with ψ : Br (0) → Rn defined by ψ(x) = x − f (x), we have that for
x1 , x2 ∈ Br (0),
|x1 − x2 | − |f (x1 ) − f (x2 )| ≤ |(x1 − x2 ) − (f (x1 ) − f (x2 ))|
≤ |ψ(x1 ) − ψ(x2 )|
1
≤ |x1 − x2 |
2
where the last inequality follows by estimating as in (1), using dψx = I −dfx .
Hence
1
|x1 − x2 | ≤ |f (x1 ) − f (x2 )|
2
for any x1 , x2 ∈ Br (0), which implies
|g(y1 ) − g(y2 )| ≤ 2|y1 − y2 | (2)
for any y1 , y2 ∈ W = Br/2 (0). In particular, g is continuous.
2
Now fix y ∈ W , and let A = dfg(y) . Since W is open, there exists δ > 0
such that y + k ∈ W if k ∈ Bδ (0). Let h = g(y + k) − g(y). Then k =
y + k − y = f (g(y + k)) − f (g(y)) = f (g(y) + h) − f (g(y)) and hence, for
k ∈ Bδ (0) \ {0},
|g(y + k) − g(y) − A−1 k | |A−1 (Ah − k)| |h|
=
|k| |h| |k |
−1
�A �|k − Ah| |h|
≤
|h| |k|
−1
�A �|f (g(y) + h) − f (g(y)) − Ah|
≤ 2 (3)
|h|
where the last estimate follows from (2). Note that since g(y+k) = g(y) =⇒
f (g(y + k)) = f (g(y)) =⇒ y + k = y =⇒ k = 0, we have that h = � 0
if k �= 0. Sice A = dfg(y) , it follows from the definition of differentiability
of f that the right hand side of (3) tends to 0 as h → 0, and hence, since
|h| ≤ 2|k | by (2), it follows that
|g(y + k) − g(y) − A−1 k |
lim = 0.
k→0 |k |
i.e. g is differentiable at y and
dgy = (dfg(y) )−1 . (4)
Finally, note that the function y �→ dgy is the composition of the function
y �→ dfg(y) and matrix inversion A �→ A−1 . Matrix inversion is a smooth map
of the entries, and the function y �→ dfg(y) is continuous since g is continuous
and f is C 1 . Hence we conclude that y �→ dgy is continuous; i.e. that g is
C 1 . Repeatedly differentiating (4) shows that g is smooth if f is smooth.
Exercise 1. Let L(Rn ; Rn ) be the set of linear transformations from Rn
into itself with the metric d(A, B) = �A − B�. (cf. Exercise 10 of handout
1.) Let U ⊂ Rn be open and f : U → Rn be a C 1 function. Show that the
map x �→ dfx is continuous as a map from U into L(Rn ; Rn ).
Exercise 2. Suppose g : [a, b] → Rn is continuous. Show that
�� b � � b
� �
� �≤
g(t)dt
|g(t)|dt
�
�
a a
3
where
�� | · |� denotes the Euclidean norm. You may use without proof that
� b � �b
� a h(t)dt� ≤ a |h(t)|dt for a scalar valued function h.
Exercise 3. Define f : R → R by f (x) = x2 + x2 sin x1 if x = � 0 and
f (0) = 0. Compute f � (x) for all x ∈ R. Show that f � (0) > 0, yet f is not
onetoone in any neighborhood of 0. This example shows that in the Inverse
Function Theorem, the hypothesis that f is C 1 cannot be weakened to the
hypothesis that f is differentiable.
Exercise 4. Define f : R2 → R2 by f (x, y) = (ex cos y, ex sin y). Show
that f is C 1 and that df(x,y) is invertible for all (x, y) ∈ R2 and yet f is
not a onetoone function globally. Why doesn’t this contradict the Inverse
Function Theorem?
Next we prove the Implicit Function Theorem. This theorem gives con
ditions under which one can solve, locally, a system of equations
fi (x, y) = 0, i = 1, 2, . . . n
where x ∈ Rm and y ∈ Rn , for y in terms of x. (Thus, y = (y1 , . . . , yn )
where y1 , . . . , yn are regarded as n unknowns, satisfying the n equations
fi (x, y) = 0, i = 1, . . . , n.) Geometrically, the set of solutions (x, y) to the
system of equations is the graph of a function y = g(x). Note that we have
from linear algebra that if for each i, the function fi is linear with constant
coefficients
� � in the variables yj , then whenever the (constant) n × n matrix
∂ fi
∂ yj is invertible, the system of equations is solvable for y in terms
1≤i,j≤n
of x. Implicit function theorem says that whenever fi are C 1 and this matrix
is invertible at a point (a, b), then the system is solvable for y in terms of x
locally in a neighborhood of (a, b).
We shall use the following notation: For an Rn valued function f (x, y) =
(f1 (x, y), f2 (x, y), . . . , fn (x, y)) in a domain U ⊂ Rm+n ≡ Rm × Rn , where
x ∈ Rm , y ∈ Rn , we shall � denote
� by dx f the partial differential represented
∂ fi
by the n × m matrix ∂ xj and by dy f the partial differential
1≤i≤n,1≤j≤m
� �
represented by the n × n matrix ∂∂ yfji .
1≤i,j≤n
Theorem 2 (Implicit Function Theorem). Let U ⊂ Rm+n ≡ Rm × Rn
be an open set, f : U → Rn a C 1 function, (a, b) ∈ U a point such that
f (a, b) = 0 and dy f |(a,b) invertible. Then there exists a neighborhood V of
4
(a, b) in U , a neighborhood W of a in Rm and a C 1 function g : W → Rn
such that
{(x, y) ∈ V : f (x, y) = 0} = {(x, g(x)) : x ∈ W } .
Moreover,
�
−1 �
dgx = − (dy f ) �
dx f |(x,g(x))
(x,g(x))
and g is smooth if f is smooth.
Proof. Define F : U → Rm+n by F (x, y) = (x, f (x, y)). Then F is C 1 in
U, F (a, b) = (a, 0) and det dF(a,b) = det dy f |(a,b) �= 0. Hence by the Inverse
Function Theorem, F has a C 1 inverse F −1 : W � → V for neighborhoods
V of (a, b) and
W of (a, 0) in R × R . Set W = {x ∈ Rm : (x, 0) ∈ W
� m n � }.
m
Then W is open in R . Note then that if x ∈ W , then (x, 0) ∈ W � so that
(x, 0) = F (x1 , y1 ) where (x1 , y1 ) ∈ V is uniquely determined by x. (In fact,
by the definition of F , x1 = x.) Define g : W → Rn by setting y1 = g(x).
Thus g(x) is defined by F −1 (x, 0) = (x, g(x)); i.e. by g(x) = π ◦ F −1 (x, 0)
where π : Rm × Rn → Rn is the projection map π(x, y) = y. Then {(x, y) ∈
V : f (x, y) = 0} = {(x, y) ∈ V : F (x, y) = (x, 0)} = {(x, g(x)) : x ∈ W }.
Since π is a smooth map and F −1 is C 1 , it follows that g is C 1 . The formula
for dgx follows by differentiating the identity
f (x, g(x)) ≡ 0 on W
using the chain rule. By repeatedly differentiating this identity, it follows
that g is smooth if f is smooth.
