Implicit Functions and Diffeomorphisms

without C 1

by

Leonid Hurwicz and Marcel K. Richter

Discussion Paper No. 279, March 1995

Center for Economic Research

Department of Economics
University of Minnesota
Minneapolis, MN 55455
 Implicit Functions and Diffeomorphisms
without C 1 *

by

Leonid Hurwicz and Marcel K. Richter

University of Minnesota

ABSTRACT. We prove implicit and inverse function theorems

for non-C 1 functions, and characterize non-C 1 diffeomorphisms.

I Introduction

With K representing either the real field lR or the complex field C, suppose
l(x,fI) = 0 for some function I: X x Y -+ K Ic , where X ~ K n and Y ~ KIc.
We seek a solution, an "implicit function," ¢; on some neighborhood of x such
that I( x, ¢;( x)) = 0 holds locally.
Cauchy, working with complex scalars, originally assumed that I was analytic
[10], and Dini's formulation for real scalars assumed that I was C 1 [15]. Similar
smoothness assumptions have formed the backbone of most proofs since then.
But what if I is only differentiable, rather than Cl - and only at (x, fI)? We
will show that a solution ¢; still exists.
We present two theorems that significantly weaken the C 1 assumption - to
either differentiability at (x, fI), or to partial differentiability with respect to y
at (x, fI), together with some continuity conditions on I. In the former case we
obtain differentiability of all solutions at x; in the latter case we obtain continuity
of all solutions at x. In neither case need there be a unique solution. However,
when I is differentiable in a neighborhood of (x, fI) and Iy is surjective there,
then there is a unique solution and it is differentiable.
The Differentiable Implicit Function Theorem yields a General Inverse Func-
tion Theorem, which in turn characterizes diffeomorphisms without continuous
differentiability.
Because implicit and inverse function theorems playa role in several parts of
mathematics, there are many applications.

* We are indebted to Professor Wayne H. Richter, University of Minnesota, for valuable

comments on an earlier version.
 II General Implicit Function Theorems

Let K represent either the real field R or the complex field C. We denote
by II . II any norm on the finite dimensional Banach space Kq over K. For any
zeKq and real Q ~ 0, we denote by Ba(z) the closed ball in Kq centered at z
with II . II-radius Q.
We will need to bound the norm on K n+k in terms of the norms on Kn and
Kk. Since we have not specified any relationship between these norms, we note
that the values II (x, 0) II of elements (x, 0) e KnH define a norm on Kn. Since
all norms on a finite dimensional real vector space are equivalent, there exists a
real /31 > 1 such that:
II(x, 0)11 ~ /3dlxll for all xeKn. (Ia)
Analogously, there exists a real /32 > 1 such that:
11(0, y)II ~ /3211yll for all yeKk. (Ib)

If X is a subset of Kq, we denote by idx the identity function on X.

By differentiability we will always mean Frechet differentiability. If 1 : X ->
Y for some open set X ~ Kq and some set Y ~ Kq, then we say that 1 :
X -> I(X) is a diffeomorphism if 1 is injective and both 1 and its inverse
1- 1 : I(X) -> X are differentiable. And we say that 1 is a local diffeomorphism
at x eX, if there is some nonempty open neighborhood U ~ X of x such that
I: U -> I(U) is a diffeomorphism. If a local diffeomorphism and its inverse are
r times differentiable, then we say it is a Dr diffeomorphism.
Where the classical proofs of implicit or inverse function theorems have often
made use of contraction mapping theorems, we will use Brouwer's fixed point
theorem for finite dimensional Banach spaces. (Although usually stated for the
real case K = R, it holds for K =C as well, since every finite dimensional
complex Banach space is isomorphic (hence homeomorphic) to en for some n,
which in turn is homeomorphic to R2n.)
We will also use the finite dimensional Banach space version of another the-
orem of Brouwer, which follows immediately from the result [6] for real scalars
by the isomorphism and homeomorphism just mentioned:
(Brouwer's Theorem on Invariance of Domain.) If U is an open
subset of K n and 9 : U -> K n is a continuous injection, then g(U)
is an open subset of K n , hence 9 is an open mapping, so g-l 1S
continuous and 9 is a homeomorphism of U onto I(U).

2
 Our first theorem assumes differentiability of f at a point, and yields differ-
entiability of solutions at the point.

Theorem 1. Differentiable Implicit Function Theorem. With K = lR

or K = C, let X be an open subset of K n and let Y be an open subset of Kk.
Suppose (x,y)eX x Y and f: X x Y -+ Kk. Suppose that:
f(x,fj) =0; (2a)
f(x, . ) is continuous on Y, for all xeX; (2b)
f( ., . ) is differentiable at (x, fj); (2c)
fy(x,fj) is surjective; i.e., (2d)
BP(x, fj)
BYI

det

Then:
a) There exists an open neighborhood Xo x Yo ~ K n x Kk of (x, y) and a
function ¢ : Xo -+ Yo such that:
f(x, ¢(x)) = 0 for all xeXo (3a)
¢(x)=fj. (3b)
b) For all sets X o, Yo, and all functions ¢ : Xo -+ Yo satisfying part (a), the
function ¢ is differentiable at x, with:
(4)
c) If (2b, c, d) are replaced by these stronger conditions:
f is differentiable on X x Y (2e)
fy is surjective on X x Y, (21)
then a neighborhood Xo x Yo of (x, fj) as in part (a) can be chosen so that:
i) there is a unique function ¢ : Xo -+ Yo satisfying part (a);
ii) ¢ is differentiable on X o, with:
¢'(x) = -(fy(x, ¢(X)))-l f:r(x, ¢(x)). (5)
d) If f is C r on X x Y for some r ~ 1, then there is a neighborhood Xl x YI
of (x, fj) and a unique function ¢ : Xl -+ YI satisfying (3a, b), and ¢ is C r on
Xl·

3
 Remark 1. We cannot totally eliminate the continuity hypothesis (2b).
Consider the function 1 : ]R 2 -+ ]R defined by :
X + y, if x + y =I 0 (6)
I(x, y) = { x2 + y2, otherwise.
The local continuity hypothesis of Theorem 1 fails, since 1 is discontinuous at
all (x, y) =I (0,0) with x + y = O. But all other hypotheses hold: 1(0,0) = 0,
1 is differentiable at (0,0), and ly(O,O) = 1. Here one cannot solve for y as a
function of x =I 0, since no (x, y) in ]R2, other than (0,0), satisfies I(x, y) = O.

Remark 2. Theorem 1 replaces the classical C 1 hypotheses of Cauchy's

(complex) theorem and Dini's (real) theorem(l) by the weaker assumptions (2b, c).
Its conclusions (a,b) are weaker than the classical versions, as the implicit func-
tions if> are not necessarily unique or C 1 • The conclusion (d) is part of the
classical Implicit Function Theorems.

Proof of Theorem 1. (a) (Existence.) To prove the existence assertions

(3), it suffices to prove:

3, ,,>0 36 bo 'rI8 ;S~6~0 'rIx 1I"'1I~6 3y lIyll~,,6

(7)
(i+X,Y+Y)EX & l(i+x,Y+y)=O.
For if we let, > 0 and 6> 0 be as in (7), and if we define
Xo = Interior(B;S(i)) (8)
then (7) guarantees existence of a point YEB,,6(fi) such that:
I(i + x, Y + y) = 0 (9a)
x=o ::::} y=O (9b)
lIyll ;£ ,lIxll· (9c)
So we can use the Axiom of Choice to define a function if> on X with if> ( x) =y
and with the properties (3).(2)
We break the proof of (7) into parts corresponding to its quantifiers.

(1) See the historical notes in Section IV.

(2) We could avoid the Axiom of Choice. For example, we could use the lexicographic
ordering based on Y1, ••• ,Yk to choose among multiple candidates Y with minimwn norm in
each of the compact sets (J(x + x, . »-1(0) n B,,6(Y)

4
 i) To choose 'Y, first note that (2d) implies that fy(x, y) has an inverse T =
(fy(x, i)))-l. So we can define:
M = max{IITfx(x, y) . xii: IIxll = I}, (10)
and then choose any TJ > 1 and define:
'Y = TJmax{M, I}, (11)
so:
'Y> 1. (12)

ii) The choice of "6 is delicate, and will involve several steps. First note
that, since f is differentiable at (x,y), we have: for all (x,y)eKn x Kk with
(x+x,y+y)eX,
f(x + x, y + y) = f(x, y) + fx(x, y) . x + fy(x, y) . y + r(x, y), (13)
where r(x, . ) is o(lI(x, y)II), i.e.,

IIr(x, y)1I _ 0 for all (x + x, y + y)eX with 0 i-1I(x, y)lI- O. (14)

lI(x, y)1I
Note that the function r(x, . ) defined in (13) is continuous for each x.
Now for any given x, the problem of finding y with (x+x, y+y)eX and with
f(x+x, y+y) = 0 can be stated in each of these equivalent forms: (x+x, y+y)eX
and
f(x+x,y+y)=O (15a)
fx(x, y) . x + fy(x, y) . y + r(x, y) = 0 (by (13) and (2a)) (15b)
r(x, y) = - fx(x, y) . x - fy(x, y). y (rearranging) (15c)
Tr(x, y) = -Tfx(x, y) . x - Tfy(x, y) . y (applying T) (15d)
Tr(x, y) = -Tfx(x, y) . x - y (by definition of T) (15e)
y = -Tr(x, y) - Tfx(x, y) . x (rearranging). (151)
In (iii) below, for every x near 0, we will apply a fixed point theorem to find a
solution y of (15). To prepare for the fixed point theorem, we will bound the
two terms on the right hand side of (151).

ii.1) To bound the first term on the right in (151), we note that, since r(x, y)
is o(x, y) by (14), so is Tr(x, y), and therefore:

IITr(x, y)1I _ 0 for all (x + x, y + y)eX with 0 i-1I(x, y)lI- O.

lI(x, y)1I
hence (16)

5
 'r/c e>O 36 g>o [Bg(i, fj) ~ X
& 'r/(x, y) lI(x,Y)II~gIlTr(x, y)11 ~ cll(x, y)lI].
Choosing c = 1 - 1/T}, we obtain from (16) a 6 > 0 such that:

Bg(i, fj) ~ X
and
(17)
1
'r/(x,y) lI(x,Y)II~gIlTr(x,y)11 ~ (1- ;J)II(x,Y)II.

Using /31 and /32 from (1) we now define:

- 1,
6= 6, (18)
/31 + /32,
so:

6>6>0 (19a)
B({31+{32"Y)6(i, fj) ~ x. (19b)
Then:
1
'r/6 6~6~0 'r/(x, y) lI(x,Y)II~6 IITr(x, y)11 ~ (1 - ;Jh6. (20)

ii.2) To bound the second term in the right hand side of (15!), choose any 6
with 8 f; 6 f; O. Then:

'r/x IIxll~6 IITfx(i, fj) . xii ~ Mllxll (by (10))

~M6
(21)
1
~
-T}
-,6 (by (11)).

iii) Now we complete the proof of (7) by picking any 6 with 8 f; 6 f; 0

and any xeKn with Ilxll ~ 6; we must find a yeK k with Ilyll ~ ,6 such that
(i+x,fj+y)eX and f(i+x,fj+y) = o. With (15!) in mind, we define the
function F : Kk -l- Kk by:

F(y) = -Tr(x, y) - Tfx(i, fj) . x, (22)

6
which is continuous in y because r(x, .) is (by (13) and (2b)). Then for all y
with Ilyll ~ ,8 we have:
lI(x, y)1I = lI(x, 0) + (0, y)1I
~ lI(x, 0)11 + 11(0, y)1I
(23)
~ ,8111xll + ,8211yll (by (1))

so by (19b):
(x + x, y+ y)EB({31+{32'"'()C(X, y) ~ X, (24)

and
IIF(y)1I = 11- Tr(x, y) - Tlx(x, y) . xII
~ IITr(x, y)1I + IITlx(x, y) . xII
1 1 (25)
~ (1 - -),8 + -,8 (by (20) and (21))
- TJ TJ
= ,8.
Thus F is a continuous function from the ball B,",(c(O) into itself. By Brouwer's
Fixed Point Theorem, F must have a fixed point F(y) = YEB,",(c(O), and by (24)
and (151) this is equivalent to lIyll ~ ,8 and I(x + x, y + y) O. =
(b) (Differentiability at x.) To prove differentiability of ¢ at x, and the
formula (4), consider any function ¢ as in (3). Since 1 is differentiable at (x, y),
we can suppose it has some Lipschitz constant i' > 0 at (x, y). Without loss of
generality, suppose that:
x =0 and y = 0, (26a)
hence:
1(0,0) = 0 (by (2a))
(26b)
¢(O) = 0 (by (3b)).
We must show that:
1I¢(x) - ¢(O) + (fy(O, 0))-1 Ix(O, O)xll
(27)
IIxll
Since 1 is differentiable at (0,0),
I(x, ¢(x)) = 1(0,0) + Ix(O, O)x + ly(O, O)¢(x) + o(lI(x, ¢(x))ll), (28)

7
so by (26) and (3a) we have:
Il/x(O, O)x + ly(O, O)</>(x) II O.
..."..,.--:--:-:-:-:---+l (29)
lI(x, </>(x)) II lI(x,¢,(x»II- 0
With (31 and (32 as in (1), we have:
lI(x, </>(x))11 ~ lI(x, 0)11 + 11(0, </>(x)) II
~ (3111xll + (3211</>(x)1I (30)
~ «(31 + (321)llxll·
So (29) implies:
Il/x(O, O)x + ly(O, O)</>(x) II
(31)
Ilxll
Then because (fy(O,O))-l is linear,
lI(fy(O, 0))-1 Ix(O, O)x + </>(x)1I
Ilxll IIxll-o l 0, (32)

and this reformulation of (27) verifies the existence and the claimed value of
</>'(0).
(c.ii) (Differentiability on X o.) Let </> : Xo --;. Yo be any implicit function
whose existence was established in part (a). With the stronger hypotheses
(2e, f), differentiability of </> on all of Xo can be established by a proof anal-
ogous to that just given in (b) for differentiability at i. Thus let </> be any
function satisfying (3a, b) as in in part (a). Let xeXo and y = </>(x). Then by
(2e, f) 1 satisfies all the conditions (2a, b, c, d) with (x, ii) replaced by (x, y), so,
as in part (b), </> is differentiable at x and </>'(x) = -(fy(x,y))-l/x(x,Y).

(c.i) (Uniqueness.) To prove the uniqueness of </>, define G : X x Y --;. X X Kk

by:
G(x, y) = (x, I(x, y)). (33)
By hypothesis (2e), G is differentiable on X x Y, and from hypothesis (2f) we
see that the derivative G' (x, y) is surjective at all (x, y) e X x Y. So by part (c.ii)
of Theorem 4 (the General Inverse Function Theorem )(3) applied to G, in some
=
open neighborhood Xl x Y1 ~ X X Kk ofG(x, y) (x, 0) there is a unique inverse
function H : Xl x Y1 --;. X X Y, and H is a local diffeomorphism at (x,O). In
particular for each xeX 1 there is a unique yeY1 satisfying G(x, y) (x, 0), i.e.,=
(3) There is no circularity in our reasoning here. Our prooflater of part (c) of Theorem 4
will depend only on parts (a) and (b) (but not (c)) of the present theorem.

8
satisfying f(x, y) = o. Since f(x,4>(x)) = 0 for the function 4> established in
part (a), this proves the uniqueness of 4> on Xl·

(d) (C r properties.) The uniqueness and C r assertions are part ofthe classical
Implicit Function Theorem. I

Even as simple a function as this:

f(x,y)=y-(x)~ (34)
fails to satisfy the differentiability hypothesis of the preceding theorem. And,
while it clearly admits the unique implicit function 4>(x) = (x)!, that solution is
not differentiable at O. Indeed, the same would be true of any function
f(x, y) =y - t/J(x), (35)
with t/J nondifferentiable. Such examples motivate our next implicit function the-
orem, which in the spirit of Goursat [19] has a weaker differentiability hypothesis
(no differentiability with respect to x) but a stronger continuity hypothesis (joint
in x and y). The conclusion of our second theorem will be weaker than our first,
asserting merely continuity of solutions rather than differentiability at the initial
point.

Theorem 2. Continuous Implicit Function Theorem. With K = ~

or K = C, let X be an open subset of K n and let Y be an open subset of Kk.
Suppose (x, y) E X X Y and f : X x Y -+ Kk. Suppose that:
f(x, y) = 0; (36a)
f( ., . ) is continuous on X x Y; (36b)
f( x, . ) is differentiable at y; (36c)
fy(x,y) is surjective; i.e., (36d)
ap(x,y)
aYI

det

Then:

9
 a) There exists an open neighborhood Xo x Yo ~ K n x Kk of (x, y) and a
function ¢ : Xo -+ Yo such that:
f(x, ¢(x)) = 0 for all xeXo (37a)
¢(x) = y. (37b)

b) The neighborhood Xo x Yo in (a) can be chosen so that there exists at

least one function ¢ : Xo -+ Yo satisfying (37), and so that every such function
is continuous at x.
c) (Goursat.) If assumption (36c) is replaced by this stronger condition:
f(x, . ) is CIon Y for all xeX, (36e)
then there is a neighborhood XIX Y1 of (x, y) and a unique function ¢ : Xl -+ Y1
satisfying (37 a, b), and ¢ is continuous on Xl.

Remark 3. As W. H. Young showed [35], the existence result (a) for the
case k = 1 only requires that f be continuous separately in x and y, rather than
jointly. See the historical remarks below.
Remark 4. Theorem 2(a,b) replaces the Cl hypothesis of Goursat's result
[19]C4) by the weaker assumptions (36b, c). Its conclusions in (a) and (b) are
weaker than in Goursat [19], as the implicit functions ¢ are not necessarily
unique or continuous. Here is an example for K = lR. in which the uniqueness
and continuity conclusions of part (c) cannot be obtained under the non-C 1
hypotheses of parts (a) and (b): Let f(x, y) = x - g(y), where

g(y) = { (y/2) + y2 sin(l/y), if y =I 0 (38)

0, otherwise.

Proof of Theorem 2. Without loss of generality, we assume that:

x=O and y = 0, (39)
and we choose t> 0 such that:
Bt"(O) ~ X. (40)

(4) See the historical notes in Section IV, and the comparisons with results ofW. H. Young
and of Bliss.

10
 (a) (Existence.) Without loss of generality, we assume that:(5)
fy (0,0) = idKk. (41)
This is justified since fy(O,O) is invertible (36d), so we could define the func-
tion g(x, y) = f(x, (fy(O, O))-ly), which would have all the properties (36) with
gy(O, 0) the identity on Kk, and 9 would admit an implicit function ¢ if and only
if f admits an implicit function ¢ = (fy(O, 0))-1¢:
g(x,¢(x)) = 0 <=> f(x,¢(x)) = o. (42)

Now define F : X x Y -+ Kk by:

F(x, y) = y - f(x, y). (43)
By hypothesis (36c) and by (39) and (41),
f(O, y) = f(O, 0) + fy(O, O)y + R(y)
(44)
= y+R(y)
for some function R(y) of y that is o(llyID:
IIR(y)1I
(45)
lIyll
Then:
IIF(O, y)1I Ily - f(O, y)11 (by (43) and (44))
lIyll IIYII (46)
IIR(y)11
(by (45)),
Ilyll
so there exists a 1 > 0 such that:(6)
V, O<'"Y~i Vy "Y"~'"Y IIF(O, y)1I < ,. (47)
We next show that:
(48)

(5) Recall that idKk is the identity function on Kk. Cf. p. 2.

(6) Our proof of (47) is based on the differentiability hypothesis (36c). Alternatively,
we could obtain (47) by simply assuming that F(iC, .) has a Lipschitz constant less than 1 at
y = o. That property is a weakened version of the hypotheses of Goursat's lemma [19) (p. 185,
=
§1, for k 1, n =1, and p. 188, §3, for arbitrary k), which sets the stage for his application of
what is now known as the Contraction Mapping Theorem. But our continuity proof in part (b)
will still use the differentiability hypothesis (36c).

11
First define the correspondence C: {c:eK : 0 ~ c: ~ t} - X X Y:
C(c:) = {(x, y) : IIxll ~ c: & Ilyll ~ 'Y}, (49)
and the corresponding maximum function:
M(c:) = max{IIF(x, y)1I : (x, y)eC(c:)}. (50)
Since F is continuous by hypothesis (36b), and since C(c:) is clearly a continuous
correspondence ,(7) it follows from the Maximum Theorem(8) that M(·) is a
continuous function. So (48) follows from (47).
Thus for every 'Y ~ l' and every x with IIxll ~ C:-y, the function F(x, . ) carries
B-y(O) into B-y(O). It is also continuous by (36b), so by Brouwer's Fixed Point
Theorem there exists a yeB-y(O) such that
F(x, y) = y, (51)
i.e.,
f(x, y) = O. (52)
If x =
:f:. 0, then we define ¢(x) y, choosing any such y;(9) if x = 0 we define
¢(O) = o. Then (37) holds, with Xo = Be(O) and Yo = B-y(O).
We have thus shown:
For every 'Y > 0 with 'Y ~ 1', and every c: ~
0 with c: ~ C:-y,
there exists a function ¢ : X-y -+ Y-y such that (37) holds (53)
with Xo = Be(O) and Yo = B-y(O).
(b) (Continuity.) We now prove that a neighborhood Xo x Yo of (x, y) can
be chosen so that not only does there exist a function ¢ : Xo -+ Yo satisfying
(37), but every such function is continuous at x = O.
We can choose 8> 0 with 8 ~ l' such that:
IIR(y)1I < lIyll for all yeBo(O) (using (45)), (54a)
and by (44) we can also ensure that:
y :f:. 0 => f(O, y) :f:. 0 for all yeBo(O). (54b)

(7) I.e .• both upper and lower hemicontinuous. Cf. Berge [1]. [2], where "semicontinuity"
is used for what we are calling "hemicontinuity."
(8) Cf. [12. p. 889. Remark]. [13. p. 19. 1.8{4}]. [1. p. 122]. [2. p. 116].
(9) As in footnote 2. the Axiom of Choice is not needed.

12
Defining'Y = 8, Xo = B.-y(O) and Yo = B-y(O), it follows from (53) that there
exists a function ¢ : Xo --... Yo satisfying (37).
Now let ¢ be any function from Xo into Yo satisfying (37). Suppose that
¢ is not continuous at O. Then there is a sequence Xl, X2, ... converging to 0,
but with ¢(Xi) not converging to ¢(O) = O. So there is a subsequence, whose
elements we again denote by Xi, for which
¢( Xi) --,------+
0-+00
ii:l O. (55)
By (54b) that implies:
f(O, ii) :I 0, (56)
so by the continuity hypothesis (36b) we have f(x, y) :I 0 for all (x, y) in some
neighborhood of (0, ii). But defining Yj = ¢(Xj) we also have f(xj, Yj) = 0 by the
solution property (37 a), and (Xj, Yj) --,------+
0-+00
(0, ii) by (55). This contradiction
completes the proof of continuity.

(c) (Uniqueness and continuity.) If f(x, .) is also Cion Y for all xeX, then
the local uniqueness and continuity of the implicit function are part of Goursat's
result [19]. I

Remark 5. While some of the hypotheses of the Continuous and the Dif-
ferentiable Implicit Function Theorems are the same, in total they are noncom-
parable. The differentiable version has stronger differentiability hypotheses, but
weaker continuity properties. While the function f in (34) satisfies the hypothe-
ses of the Continuous Theorem but not those of the Differentiable Theorem, the
reverse is true for the function
f(x, y) = {Y + x2 + y2, if X is ~ational (57)
y, otherwise.
Here f(x, y) is not continuous in X at (0, y) for y :I 0, so it violates (36b) for
(x, ii) = (0,0); but it is differentiable at (0,0), and satisfies the other hypotheses
of the Differentiable Theorem.

We have obtained implicit functions under two different hypotheses about the
independent variable x. In our first theorem f(x, y) was differentiable jointly in
X and y at (0,0); and we concluded that the implicit function ¢(x) was differ-

entiable at X =
O. In our second theorem f(x, y) was only differentiable with
=
respect to y at (0,0), and f(x, y) was continuous with respect to X at x 0; and
=
we concluded that the implicit function ¢(x) was continuous at x O. Now we

13
show that these two results can be combined, allowing two types of independent
variables - those satisfying differentiability hypotheses yield differentiability
conclusions for the implicit function, and those satisfying continuity hypothesis
yield continuity conclusions. We are interested in solving f( v, w, Y) = 0 for Y as
a function ¢ of (v, w): f(v, w, ¢(v, w» = o. The role played by :x in Theorem 1
is played here by v, and the role of:x in Theorem 2 is played here by w.

Theorem 3. General Implicit Function Theorem. With K = ~ or

K = C, let V, W, and Y be open subsets of Km, KP, and Kk, respectively,
where m + P > OJ we allow m = 0 or p = 0, in which case V or W, respectively,
is empty. Suppose (v, iii, y)e V x W x Y and f : V x W x Y -+ Kk. Suppose
that:
f(v, iii, y) = OJ (57a)
f(v, ., .) is continuous on W x Y, for all veVj (57b)
f( . , iii, . ) is differentiable at (v, y)j (57c)
fy (v, iii, y) is surjective j i.e., (57d)
8j1(v, iii, y) 8j1(v, iii, y)
8y! 8Yk
det ;f0.
8fk(V, iii, y) 8fk(V, iii, y)
8y! 8Yk
Then:
a) There exist open neighborhoods Va of v and Yo of y such that, for every
ve Va there exists an open neighborhood Wv of iii and a function ¢(v, .) : Wv x Yo
such that:
f(v, w, ¢(v, w» = 0 for all weWv (58a)
¢(v, iii) = y. (58b)
b) The neighborhood Yo in (a) can be chosen so that every function ¢ satis-
fying (58) also satisfies:
¢(v, . ) is continuous at iii, for all ve Va (59a)
¢( . , iii) is differentiable at v, and: (59b)
¢v(v, iii) = - (fy(v, iii, y»-! fv(v, iii, y).
c) If f( ., ., .) is continuous on V x W x Y, then there is a neighborhood
Va x Wo x Yo and a function ¢ satisfying (58) that is continuous at (v, iii, y).

14
 d) (Goursat [19], [20].) If f(·, ., .) is continuous on V x W x Y, and
if f(v,w,·) is C 1 on Y for all (V,W)EV x W, then there is a neighborhood
V1 x W 1 X Y1 of (ii, w, y) and a unique function </> : V1 X W 1 ..... Y1 satisfying
(58a, b), and </> is continuous on V1 x W1 .

Remark 6. This theorem is a generalization of the Differentiability and the

Continuous Implicit Function Theorems: when w is absent, the hypotheses and
conclusions are those of the Differentiable Implicit Function Theorem, and when
v is absent, they are those of the Continuous Implicit Function Theorem.

Proof of Theorem 3. Without loss of generality, we simplify by assuming

that:
ii = ° and w= ° and y= 0. (60)
Also without loss of generality, we assume that:
fy(O, 0, 0) = idKk. (61)
This is justified since fy(O, 0, 0) is invertible (57d), so we could define the function
g( v, w, y) = f( v, w, (fy (0,0, O))-ly), which would have all the properties (57)
with gy(O, 0, 0) the identity on Kk, and 9 would admit an implicit function </> if
and only if f admits an implicit function ¢ = (fy(O,O,O))-l</>:
g(v,w,</>(v,w)) = ° ¢} f(v,w,¢(v,w)) = 0. (62)

(a) (Existence.) Define F : V x W x Y by:

F(v, w, y) = y - f(v, w, V). (63)
By hypothesis (57c),
f(v, 0, y) = f(O, 0, 0) + fv(O, 0, O)v + fy(O, 0, O)y + R(v, y)
= y + fv(O, 0, O)v + R(v, y) (by (61), (57a) and (60)) (64)
for some function R( v, y) of (v, y) that is o(lI( v, y)ll):
R(v, y)
lI(v, y)1I O;tIl(v,y)lI ..... o I 0.
(65)

By (63) and (64) we have:

IIF( v, 0, y)11 = Ily - f( v, 0, y)1I
= IIR( v, y) + fv (0,0, O)vll
(66)
< IIR(v,y)11
= II(v, y)11 II( v, y)1I + IIfv (0,0, O)lIlIvll,

15
so by (65) there exists a, > 0 such that, for allll(v, y)1I < , we have
1
V(v, y) lI(v,y)II~-Y IIF(v, 0, y)1I < 2' + IIfv(O, 0, O)lIllvll· (67)
Now define:

eo = { ~IIJ.(;'O'O)II' if IIfv (0,0,0)11 "# 0

(68)

2" otherwise,
so eo > O. Now let
e = min{eo,,}. (69)
Then:

"Iv IIvll<e Vy lI(v,y)II;;;-y IIF(v, 0, y)1I <

1
2' + 2'
1
=,. (70)
Without loss of generality, we may suppose that the norm on V x Y is the
maximum norm, and then (70) and (69) imply:
(71)
Next we will show that this inequality holds also for all w near O. Since
F(v, 0, y) is defined for IIvll < e & lIyll ~ " openness of W ensures there is some
b> 0 for which F(v, w, y) is defined for allllxll < e & IIwll ~ b & lIyll ~ ,.
Then paralleling the proof of Theorem 2,(l0) we define:
C(c) = {(w, y) : Ilwll ~ c & lIyll ~ ,}, (72)
for all C with 0 ~ c ~ b. Clearly C( . ) is a continuous correspondence, so by
the continuity hypothesis (57b) and the Maximum Theorem, for every v with
IIvll < e, the function
Mv(c) = max{IIF(v,w,y)ll: IIwll ~ c & lIyll ~,} (73)
is continuous for 0 ~ c ~ b. So by (71) for every v with IIvll < e, there exists a
Cv > 0 with Cv ~ b such that:
(74)
Thus for every v with IIvll < e and every w with Ilwll ~ cv , the func-
tion F(v, w, .) carries B-y(O) into B-y(O). It is also continuous by (57b), so by
Brouwer's Fixed Point Theorem there exists a yeB-y(O) such that
F(v, w, y) = y, (75)
(10) cr. (48). (49), and (50).

16
l.e.,
I(v,w,y) = o. (76)
°
If w ::j: 0, then we define ¢(v, w) = y, for any such y;(ll) if w = x = we define
¢(O) = 0. Then (58) holds with Vo = Be (0), Wv = Bo. (0), and Yv = B-y(O).

(b) (Continuity with respect to w.) To prove part (59a), we note that the
function defined by i( w, y) = 1(0, w, y) satisfies the hypotheses of the Continu-
ous Implicit Function Theorem, and so by part (b) of that theorem we immedi-
ately obtain property (59a).
(Differentiability with respect to v.) To prove (59b), we note that the function
j defined by j( v, y) = I( v, 0, y) satisfies the hypotheses of the Differentiable
Implicit Function theorem, and so by part (b) of that theorem we immediately
obtain property (59b).

(c) (Continuity with respect to (v, w).) Part (c) follows immediately from
part (b) of the Continuous Implicit Function Theorem, since if we define x =
(v, w), then all the hypotheses of that theorem apply.

(d) (Uniqueness and local continuity.) See part (c) of Theorem 2. I

III A General Inverse Function Theorem

As usual, it is straightforward to prove an inverse function theorem from an

implicit function.

Theorem 4. General Inverse Function Theorem. Let K = IR or

K = C, and let X be an open subset of K n , and let 9 : X -+ K n be continuous
on X and differentiable at xeX. Let y = g(x). Suppose g'(x) is surjective, i.e.,
rank(g'(x)) = n. Then:
a) There exists an open neighborhood Y of fJ and an injection h : Y -+ X
such that:
go h = idy (77a)
h(y) = x. (77b)

(11) As in footnote 2, the Axiom of Choice is not needed.

17
 b) Every function h satisfying (77) is differentiable at g(x), with:
h'(g(x» = (g'(x»-l. (78)
c) If the differentiability of g at x is strengthened to differentiability on X,
and surjectivity of g'(x) is strengthened to surjectivity of g' on X, then:
i) the neighborhood Y in part (a) can be chosen so that h is also
differentiable and surjective on Y;
ii) the neighborhood Xo can also be chosen so that gfxo: Xo -+ g(Xo)
is a diffeomorphism (in particular, the inverse function h is
unique) and:
h'(g(x» = (g'(X»-l for all xeXo. (79)
d) If g is C r on X for some r ~ 1, then there is an open neighborhood Y of
ii, and there is a unique function h : Y -+ X mapping Y onto an open set, and
satisfying the conditions (77a,b), and it is cr.

Proof. We will prove the General Inverse Function Theorem from the Dif-
ferentiable Implicit Function Theorem, reversing the roles of x and y. Define
f(x, y) = g(x) - y (80)
for all x eX, so f is differentiable at (x, ii).

(a) (Existence.) To prove part (a), note that because f:z:(x, fI) =
g'(x) is
surjective, part (a) of the Differentiable Implicit Function Theorem implies that
there exists an open neighborhood Y of g(x) and a function tf; : Y -+ X such
that:
f(tf;(y),y) =0 for all yeY (81a)
tf;(ii) = x. (81b)
=
So in view of (80), h tf; satisfies (77). And h is an injection, since if h(y) = h(ij)
= =
then y g(h(y» g(h(ij» =fj by (77a).

(b) (Differentiability at x.) To prove part (b), we note that part (b) of the
Differentiable Implicit Function Theorem implies that any tf; satisfying (81) is dif-
ferentiable at fI = g( x). Then by (77) the Chain Rule implies that g' (h(Y) )h' (Y) =
the identity on Y, so h = tf; also satisfies (78).

(c.i) (Differentiability on Y.) Assume now that g'(x) exists and is surjective
for all x eX. The same argument used in (b) to prove differentiability of any

18
right inverse of 9 at g( x) now shows that h is differentiable at all y e Y, hence
continuous on Y. To see that, let h : Y -+ X be as in part (a), and let xeX and
ii = g(5:). Since 9 is differentiable on X, it is continuous on X; and since g'(5:)
is surjective, part (a) applies with (x, ii) replacing (x, y) so, as in part (b), h is
differentiable at ii and h'(ii) = (g'(5:»-l.

(c.ii) (Uniqueness and diffeomorphism.) Since any right inverse h of 9 is

also injective on Y (by part (a» and continuous (by part (c.i», it follows from
Brouwer's theorem on Invariance of Domain that h is a homeomorphism from Y
to the open neighborhood h(Y) of x. So Xo = h(Y) is an open neighborhood of
x, and gfxo: Xo -+ g(Xo) is a homeomorphism.
It follows that 9 fxo has a unique inverse. Indeed, for any yeg(Xo) we
have g(h(y» =
y because h is a right inverse of g. Applying h to each side
yields h(g(h(y») = h(y). Thus, since h is a homeomorphism onto the open
neighborhood Xo of x, we have h(g(x» = =
x for all xeXo, so hog idxo . Then
h is a left inverse as well as a right inverse of g.
By (c.i) the inverse h of 9 is differentiable, so 9 fxo is a diffeomorphism,
and (79) holds.

(d) (C r properties.) The uniqueness and C r assertions are part of the classical
Inverse Function Theorem. I

Diffeomorphism Corollary. Let K = ~ or K = C, let X be an open

subset of K n , let 9 : X -+ K n , and let the integer r ~ 1. Then 9 is a local Dr
diffeomorphism(12) at xeX if and only if there is some neighborhood of x on
which 9 is Dr and 9' is surjective.

Proof. The necessity of local Dr differentiability and surjectivity is obvious.

The sufficiency is contained in part (c.ii) of the General Inverse Function Theo-
rem when r =
1. For r> 1 the proof follows by induction, paralleling the usual
proofs of the classical Inverse Function Theorem. (13) I

(12) cr. page 2.

(13) cr. [14, p. 272, (10.2.3)).

19
 Remarks 7. a) There are examples(14) in which no inverse h in part (a) of
the General Inverse Function Theorem is unique or continuous. In such examples,
with f as defined by (80), none of the functions ¢ in Theorems l(a,b), 2(a,b),
and 3( a, b) can be unique or continuous.

b) The continuity hypothesis on h cannot be dropped altogether, as shown

by instances based on example (6) above.

c) The uniqueness and differentiability properties that follow from the ad-
ditional hypotheses in part (c) are new results in the real case K = JR. When
K = C, however, the local differentiability assumption in (c) is only apparently
weaker than the classical assumptions when K = JR, it is only apparently weaker
when K = C, since then the local differentiability hypothesis of (c )(15) implies, by
Goursat's Theorem [11, p. 100] on analytic functions, that 9 is actually analytic
- in which case the classical hypotheses and conclusions of part (d) hold.

d) As an example of a diffeomorphism that is not C 1 , yet satisfies our hy-

potheses in part (c.ii) of Theorem 4 and the Corollary, consider:

g(x) = {x + x 2
sin(1/x), for -1 < x < 1
(82)
0, for x = o.

e) The theorems have useful applications, since the classical Implicit Function
Theorem is imbedded in many parts of mathematics, including differentiable
manifolds and optimization theory. The weaker hypotheses of the theorems
above admit the possibility of extending existing results to these and other areas.
For example, replacing the classical Implicit Function Theorem by the gen-
eral version (Theorem 1) allows a proof of the Lagrange Multiplier Theorem that
weakens the classical C 1 hypothesis to just continuity of the constraint functions
and their differentiability at the extremum [23]. As another example, compar-
ative statics question~ for equilibrium systems(16) can now be addressed with
weaker assumptions.

(14) E.g., (38) or [14, p. 273, Problem 2].

(15) But not those of parts (a) and (b).
(16) Cf. [33, Part I).

20
 IV Some Historical Background and Comparisons

We originally developed the general theorems to solve an applied problem

[23]. However, attempts to ascertain their novelty led to a historical study, and
we present a few of the highlights that may be of interest to some readers.
We sketch the development of the implicit function theorem, with primary
emphasis on the smoothness assumptions underlying the proofs. For the early
period we have been guided by the information in Osgood [30], with Cauchy the
earliest mentioned. We have not tried to examine the history prior to Cauchy's
work. We use the notation of the previous sections.

The are several dimensions we could use for making historical comparisons.
We could discuss assumptions: real scalars or complex scalars; the number of
x and y variables; the smoothness assumptions on f with respect to x or y or
(x, y); smoothness assumptions at (x, y) or in a neighborhood of the point. We
could discuss conclusions: existence of implicit function, uniqueness of implicit
functions, and smoothness of implicit functions. We could discuss methods of
proof: Cauchy's calculus of residues, Cauchy's Calcul des limites, differential
equations, or fixed point theorems.
We will give a brief chronological history, highlighting major differences
among the various contributions, comparing our three theorems above with the
major earlier results.

1831. Cauchy. Cauchy's results [10] are presumably the earliest rigorous
existence proofs of the Implicit Function Theorem. Here the underlying scalar
field K is that of the complex numbers. He mentions the general case in the
= =
introduction, and gives a detailed analysis of the special case k 1, n 1. The
function f is assumed to be represented by a power series - i.e., to be analytic.
A unique implicit function ¢ is obtained, and it is shown to be analytic.
Cauchy's tools were his calculus of residues and his Calculus of Limits. Their
=
flavor may be experienced by considering the simple case where k 1 and n 1, =
so that both x and yare one-dimensional. Then a presentation of Cauchy's
approach, in modern language looks roughly like the following.
Assuming f is analytic and not identically zero, for any given x, the gener-
alized Argument Principle implies:(17)

1
~
7rZ
1
C
fy(x, y)
y f(
x, Y
-f..
) dy = ~O'rar,
r=l
(83)

(17) cr. [11, p. 124, Theorem 3.6]

21
where the a r are zeros of f(x, . ) and the ct r are their multiplicities, and where C
is a closed rectifiable curve not passing through any root a r . Since the roots of
an analytic function that is not identically zero are isolated, one can pick a small
enough circle C in the complex plane about the root al of f(x, . ) so that there
are no other zeros in the circle. In particular, Cauchy considers the case that
the root al has multiplicity 1.c 18 ) Then denoting al by ¢(x), (83) becomes:(19)

_1 1 fy(X'Y)d - ¢( )
21ri c y f(x, y) y - x. (84)

In view of the continuity of f and its derivatives, Rouche's Theorem(20) implies

that for all nearby x the number of zero's (counting multiplicities) ofthe function
f(x, .) remains the same; so there is a unique local solution function ¢, and it
can be calculated by the integral formula (84).
Cauchy further verifies that ¢ is (locally) an analytic function - that it is
represented by its Maclaurin series. First he applies the calculus of residues to
calculate derivatives for the Maclaurin series. Without loss of generality assum-
ing x = 0 and ¢(x) = 0:(21)

¢(m)(o) = ~
21rZ c
1y [ a: (fy(X, y))
ax f(x, y) z=o
] dy

1
= -2 .1
1rZ
yAm(y)dy (defining Am)
C
(85)
= ~(21ri
21rZ
Res(yAm(Y); 0)) (by the Residue Theorem)
1 dm - 1
= (m _ 1)!;~ dym-1 «ym)yAm(y)),
where the last equality follows (22) since yAm(Y) has a pole of order m. Then he
shows that the series converges to ¢.
Of course this is an anachronistic account of his methods: the generalized
Argument Principle, Rouche's Theorem, etc., had perhaps not yet crystallized

(18) When the usual assumption is made that J,,(x, fi) is surjective, it follows that the
unique root within the circle is a simple root.
(19) Cf. [10, p. 76 (52)]. Cauchy expresses y in polar coordinates, so his integrand has
additional factor y.
(20) Cf. [11, p. 125, Theorem 3.8].

(21) Cf. [10, p. 83, equation (97)].

(22) Cf. [11, p. 113, Proposition 2.4].

22
as named theorems, and Cauchy established them in the context of his proof. (23)

1852. Cauchy. In his memoir [9], Cauchy used an alternative approach for
proving the Implicit Function Theorem, namely existence theorems for differen-
tial equations. Again the underlying scalar field is that of the complex numbers,
and the function f is analytic ("synectic"). The implicit function ¢ is unique
and analytic. It is explicitly assumed that fy(x, y) is nonsingular.
Although he mentions more general cases, formal proofs are given only for the
case where x is one-dimensional but y can be multi-dimensional (i.e., n 1 and =
k ~ 1). In this situation Cauchy is able to apply existence theorems for ordinary
differential equations to establish existence, uniqueness, and differentiability of
the implicit function ¢.
We can illustrate the logic of his approach again for the simple case f(x, y) =
owith both x and yone-dimensional. Assume(24) that f is C 1 and thatfy(x, y) =f.
O. In that case, the right hand side of the ordinary differential equation
dy fx(x, y)
(86)
dx fy(x, y)
with the initial condition f( x, y) = 0 is well defined and has a unique local
solution, say y = =
¢(x), satisfying y ¢(x). It can be shown that the function
¢ is the unique function locally satisfying f(x, ¢(x)) = 0 with y = ¢(x); i.e.,
it is the implicit function for the given function f. Furthermore, its derivative
satisfies the relation
¢'(x) = fx(x, ¢(x)). (87)
fy(x, ¢(x))
Thus we have obtained the standard Implicit Function Theorem (for this simple
case), but at the cost again of assuming the given function f to be C 1 . Hence
it could have been used to obtain Dini's result for a system involving a single
independent variable. Perhaps an analogous approach using existence theorems
for partial differential equations could be used when there are several independent
variables, i.e., when n ~ l.
Theorem 1 above weakens the assumptions of both Cauchy papers by re-
ducing the smoothness requirements on f (from analytic in a neighborhood to
differentiable at a point). Although Theorem 1 obtains existence of an implicit

(23) A lucid exposition of the methods he used in [10) for establishing (84) is contained
in [27).
(24) This logic applies both to the real scalars and also to the complex scalars (in which
case C1 is equivalent to analytic).

23
function and asserts that all implicit functions are differentiable, it cannot claim
uniqueness of the implicit function,

1875. Briot and Bouquet. (25) Briot and Bouquet in their [5] are clearly
aware of and following Cauchy's work in the theory of complex variables. The
scalars are complex, and I is analytic. In Section 211, p. 336, Theorem III, they
state and prove the theorem for the special case k =
1, n =
1; in Section 212,
p. 337, Theorem IV, they state and sketch an argument for the more general
case k ~ 1, but still with n = l.
They use solution methods for ordinary differential equations to establish ex-
istence of an analytic implicit function. The uniqueness of the implicit function
is implied by their results on ordinary differential equations (an explicit unique-
ness argument is given in Section 209, p. 332, for the case of a single ordinary
differential equation).

1877-78. Dini. In his lecture notes [15], Dini states the Implicit Function
Theorem in the form found in most textbooks today. The scalars are real, and I
is C 1 (or cr for some r ~ 1). He solves for a unique implicit function ¢, which
he proves is C 1 (respectively, cr).
At that time, of course, Brouwer's Fixed Point Theorem [7] was not available. (26)
Instead, Dini used the Intermediate Value Theorem together with a C 1 assump-
tion. To indicate the essence of his approach, consider the simple case where
both x and yare one-dimensional and there is only a single given function I and
a single equation I(x, y) = O.
In this context, Dini assumes that, in some neighborhood of (x, y), the
function I is C 1 and furthermore, that the partial derivative Iy(x, y) f O.
His proof of the existence of an implicit function y =
¢( x) then uses two
propositions of differential calculus: the Intermediate Value Theorem and Tay-
lor's Formula with Remainder (Extended Theorem of the Mean) in the form
I(x + h, Y + k) =I(x, y) + hlx(x + Oh, y + Ok) + kly(x + Oh, y + Ok), with
0<0<l.
Thus even to establish existence, this proof uses the existence of the deriva-
tives of I with respect to x and y in a neighborhood of (x, y). Furthermore it
uses the continuity of the partially to show that, in a sufficiently small neigh-
borhood, it retains the same sign as at (x, y), and the continuity of Ix is also
used.
(25) (We have not had access to Briot and Bouquet's first edition of 1859. (Osgood [30)
mistakenly dates the 2nd edition as 1873).)
(26) The paper itself is dated July, 1910.

24
 The first printed version of Dini's implicit function theorem is in Genocchi's
[17], as edited by Peano. (The German translation is [18].) A clear formulation
of Dini's theorem and a proof is also found in [31], with a helpful diagram.
Theorem 1 weakens Dini's assumptions by allowing both real and complex
scalars, and by weakening the smoothness requirements on f. Although Theo-
rem 1 asserts that all the implicit functions are differentiable at (x, ii), it cannot
claim uniqueness of the implicit functions.

1893. Jordan. In Sections 91-95 of [24] the scalars are real, and the results
do not seem to go beyond those of Dini. The general case k ~ 1, n ~ 1 is
considered, f is C 1 , and a unique implicit function ¢ is obtained. Jordan notes
that its partial derivatives exist, and his discussion makes clear that they are
continuous.
In Section 191, page 178, the scalars are complex. The case k = 1, n ~ 1 is
considered in detail; the possibility of an extension to k ~ 1 is mentioned. The
function f is analytic ("synectic"), and a unique implicit function ¢ is obtained,
and it is shown to be analytic, with the usual formula for its partial derivatives.
= =
A clear proof along the same lines, for k 1, n 1 is also to be found in [31,
pp. 345-3467].
Our Theorem 1 weakens the analytic hypothesis on f to differentiability at
(x, ii), while retaining existence of an implicit function and differentiability of
every implicit function at (x, y). It cannot, however, guarantee uniqueness or
analyticity of the implicit function.

1899. Lindelof. In [26] the scalars are complex. The general case k ~
1, n ~ 1 is considered. The function f is assumed analytic, an implicit function
¢ is obtained, and it is shown to be analytic.
The proof is based on power series expansions of f and the implicit function
¢, using Cauchy's Calcul des limites, in contrast to Jordan (1893), which reduced
the complex case to the real case.
Our Theorem 1 weakens Lindelof's assumption that f is analytic to differ-
entiabilityat (x, ii) and local continuity, while retaining existence of an implicit
function. While it obtains differentiability of every implicit function at (x, ii), it
cannot claim uniqueness.

1901. Osgood. This contains a clear statement of an implicit function

theorem for complex scalars, with k ~ 1, n ~ 1, and with an explicit statement
of uniqueness of the implicit function. We have also found this to be a very
useful source for historical information; see especially pp. 19 (footnote 30) and
103 (footnote 247).

25
 1903. Goursat. In [19] the scalars are real, and n ~ 1 and k ~ 1. Goursat
makes no differentiability assumption with respect to x. He assumes that !
is continuous in a neighborhood of (x, y), and he retains the assumptions that
!(x, . ) is C 1 for x near O. He obtains a unique implicit function, and shows that
it is continuous.
Goursat's proof used a fixed point theorem, what we would now call a con-
traction mapping theorem, which he proved using Picard's method of successive
approximations. Defining
g(x, y) =y - y - (fy(x, y))-l !(x, y) (88)
he shows that g(x, y) = y has a unique solution for y = ¢(x) that is a continuous
function of x. (27)
Theorem 2 weakens Goursat's assumptions by allowing both real and com-
plex scalars, and by weakening the smoothness requirements on!. Although
Theorem 2 obtains existence of an implicit function and asserts that all implicit
functions are continuous, it cannot claim uniqueness of the implicit functions.
The Goursat result (dropping differentiability of ! with respect to x, and
assuming continuity of ! near (x, y)) is also found in Bliss's published lecture
[3, pp. 8-9] on the "Fundamental Existence Theorems." Although Bliss cites
Goursat's paper, his method of proof is based on Taylor's Theorem.
In [22] Graves notes that existence, uniqueness, and continuity can be ob-
tained without assuming any differentiability of ! with respect to the variable
x. His remark (p. 139) is in the midst of his proof of his Theorem 2 (p. 138).

1909. W. H. Young. There are several theorems to consider in Young's

[35]. In all of them, the scalars are the reals.
Young's Theorem 9 modifies Dini's result, replacing the C r differentiability
assumption on ! with an assumption that! is differentiable of order r for some
r ~ 2. It asserts the existence of a unique implicit function ¢, which is also
differentiable of order r. Neither our Theorem 1 nor Dini's theorem require
second order differentiability for existence of an implicit function.
Our Theorem 1 uses weaker assumptions than Young's Theorem 9 in allowing
both real and complex scalars, and it uses weaker smoothness requirements on
! (differentiability at (x, y), instead of twice differentiable at (x, y), hence C 1
in a neighborhood). Although our Theorem 1 obtains existence of an implicit

(27) He notes that the solution is unique provided that 9 is Lipschitz continuous with con-
stant K < I, observing that this automatically holds when fy is continuous in a neighborhood
of (x,y). (See also footnote 6, p. 11 above.)

26
function and asserts that all implicit functions are differentiable at x, it cannot
claim uniqueness of the implicit functions.
Young's Theorem 10 has a weaker assumption on f than Dini's theorem, since
it requires a lower order differentiability of f with respect to y (differentiability
in a neighborhood of(x, fi), rather than C l ). However it strengthen's Dini's as-
sumption that fy(x, fi) is nonsingular (hence nonsingular in some neighborhood,
since f is C l for Dini), by imposing nonsingularity conditions on certain prin-
cipal minors (in some neighborhood). Young then concludes that there exists a
unique implicit function ¢, and it is differentiable at (x, fi).
Young remarks that only the uniqueness of ¢ is affected if one weakens the
assumptions of his Theorem 10 by only requiring that the Jacobian and principal
minor conditions at (x, fi), rather than throughout some neighborhoodJ28)
Our Theorem 2 uses weaker assumptions than in Young's Theorem 10 in al-
lowing both real and complex scalars, in only requiring continuity of f locally
and differentiability at (x, fi) (rather than differentiability locally), in not requir-
ing Young's conditions on principal minors of the Jacobian in a neighborhood (it
only requires that the Jacobian itself be nonsingular at (x, fi)). Our Theorem 2,
however, cannot claim uniqueness of the implicit functions.
Young claims (in his Corollary 3 of Theorem 10) that Dini's result is a corol-
lary of his Theorem 10, although that is not clear to us.

In neither Theorem 9 or Theorem 10 does Young consider a situation in

which f is not differentiable with respect to x, in contrast to Goursat's result
and our Theorem 2. However, his Theorem 5, as extended in his section 10,
contains a special case that does not require differentiability. For the special case
k = 1, n ~ 1, Assuming only that fy(x, fi) is nonzero, and that f is continuous
separately in each argument, he obtains in parts (1) and (2) the existence of
an implicit function ¢. (While there need not be a unique implicit function, he
notes that there exists an upper semi continuous one and a lower semi continuous
one.)
The basic idea of Young's proof is similar in spirit to what others have done.
Because it yields a better result for the special case k = 1, n ~ 1 than either his
general Theorems 9 or 10, or our Theorem 2, we present here a proof sketch. (We
weaken his continuity assumption with respect to x by only requiring continuity
at x.)

Because f(x,fi) = 0 and fy(x,fi) :f. 0, there exists a iJ such that

f(x, iJ) > 0 and f(x, -iJ) < O. Because f( ., iJ) is continuous at x, then
(28) Page 420, paragraph 16.

27
 for all small enough :c we have both f(x +:c, y) > 0 and f(x +:c, -y) < O.
Then continuity of f(x+:c, . ) and the Intermediate Value Theorem imply
f(x + :c, y*) = 0 for some y* in the interval [-y, Yl. By continuity of
f(x +:c, . ), the set of such y* is closed, so one way to define the value of
an implicit function at :c is to chose the maximum such y*.

Other. Without aiming at completeness, we mention a few other treatments

of the implicit function theorem. Careful statements of the classical (Dini) ver-
sion are found in Bolza [4] and Caratheodory [8].
In the complex domain, Goursat [20, pp. 399 ff.], [21, pp. 233 ff.] and
Osgood [32, §§45, 105], [32, pp. 86-86]' and Markushevich [28, p. 109] all utilize
the Weierstrass Preparation Theorem [34, pp. 107-114].(29) This theorem deals
with cases where the equation f(:c, y) = 0 has roots y(:c) whose multiplicity m
may equal or exceed 1, and is applicable to cases where fy(x, y) = 0, and in
this respect has broader applicability than the results established in the present
paper. When m = 1, the Preparation Theorem can be used to prove the Implicit
Function Theorem for the case where fy(x, y) "# 0 and f is analytic. This is
shown most explicitly by Markushevich [28, p. 109, Theorem 3.11].
Evgrafov's proof [16] uses series expansions and Cauchy's integral formu-
lae, as well as the Frechet property of the differential. In addition to prov-
ing existence, uniqueness, and continuity, it also derives the formula </>' (:c) =
- fy(:C, y)f(:c, y).
The paper by Hildebrand and Graves [22] was perhaps the first to use Banach
space techniques to establish the Implicit Function Theorem through methods
applying to both the real and complex domains.
Nijenhuis [29], building on a notion of Leach [25], obtained inverse and im-
plicit functions in the real case, under an assumption ("strong differentiability")
weaker than C 1 , but stronger than Frechet differentiability.

(29) Published in 1886, but presented in lectures since 1860.

28
 References
1. C. Berge, Espaces Topologique, Dunod, Paris, 1959.
2. , Topological Spaces, The Macmillan Company, New York, 1963.
3. G. A. Bliss, Fundamental Existence Theorems, in Lectures on Mathemat-
ics, American Mathematical Society, 1913, 1-107, The Princeton Colloquium.
Lectures delivered September 15 to 17, 1909.
4. O. Bolza, Voriesungen iiber Variationsrechnung, Chelsea Publishing Com-
pany, New York, NY, no date, author's preface dated 1909, second edition.
5. C. A. A. Briot and C. Bouquet, Theorie des fonctions doublement periodiques
et, en particulier des fonctions elliptiques, Gauthiers-Villars, Paris, 1875, Deuxieme
Edition.
6.1. E. J. Brouwer, Beweis der Invarianz des n-dimensionalen Gebiets, Math.
Ann. 71 (1912), 305-313.
7. , Uber Abbildung von Mannigfaligkeiten, Math. Ann. 71 (1912),
97-115.
8. C. Caratheodory, Calculus of Variations and Partial Differential Equa-
tions of the First Order, Chelsea Publishing Company, New York, NY, 1982,
originally published as Variaiionsrechnung und Partielle Differeniialgleichungen
ersier Ordnung, B. G. Teubner, Berlin, 1935.
9. A. Cauchy, Memoir sur 1'application du Calcul intinitesimal a la determination
des fonctions implicites, in Oeuvres Completes d'Augustin Cauchy, vol. 11, 1899,
406-415, edited by L'Academie des Sciences; C.R., T. 34, p. 265 (23 fevrier 1852).
10. , Resume d'un Memoir sur la mecanique celeste et sur un nouveau
calcul appeJe calcul des limites, in Oeuvres Completes d'Augustin Cauchy, II,
vol. 12, Gauthier-Villars, Paris, 1916,48-112, edited by L'Academie des Sciences.
The part pertaining to series expansions was previously read at a meeting of the
Academie de Turin, 11 October, 1831.
11. J. B. Conway, Functions of One Complex Variable, Springer-Verlag, New
York, NY, 1978, Second Edition.
12. G. Debreu, A Social Equilibrium Existence Theorem, Proceedings of the
National Academy of Sciences 38 (1952), 886-893.
13. , Theory of Value, John Wiley and Sons, New York, 1959.
14. J. Dieudonne, Foundations of Modern Analysis, vol. 1, Academic Press,
New York, 1969, enlarged and corrected printing.
15. U. Dini, Analisi Intinitesimale, Litografia Gorani, Pisa, 1877-1878, lithographed
version of manuscript lecture notes.

29
 16. M. A. Evgrafov, Analytic Functions, W. B. Saunders Company, Philadel-
phia, 1966.
17. A. Genocchi, Calcolo Differenziale e Principii di Calcolo Integrale, Fratelli
Bocca, Turin, 1884, pubblicato con aggiunte dal Dr. Guiseppe Peano.
18. , Differentiairechnung und Grundziige der Integralrechnung, Teub-
ner, Leipzig, 1899, German translation by G. Bohlman and A. Schepp.
19. E. Goursat, Sur la tMorie des fonctions implicites, Bulletin de la Societe
Mathematique de France 31 (1903), 184-192.
20. , A Course in Mathematical Analysis, vol. I, Ginn and Company,
Boston, 1904, translated by E. R. Hedrick.
21. , A Course in Mathematical Analysis. Functions of a Complex
Variable, Ginn and Company, Boston, 1916, Vol. II, Part I, Functions of a Com-
plex Variable.
22. H. Hildebrand and L. M. Graves, Implicit Functions and Their Differen-
tials in General Analysis, Transactions of the American Mathematical Society
(1927).
23. L. Hurwicz and M. K. Richter, Optimization and Lagrange Multipliers:
Non-C! Constraints and "Minimal" Constraint Qualifications, Discussion Paper
No. 280, Department of Economics, University of Minnesota, 1995.
24. C. Jordan, Cours d'analyse de l'Ecole Poly technique, Gauthier-Villars et
Fils, Paris, 1893, Tome Premier, Calcul differential, deuxieme edition.
25. N. B. Leach, A Note on Inverse Function Theorems, Proceedings of the
American Mathematical Sociaty 12 (1961),694-697.
26. E. Lindelof, Demonstration eJementaire de l'existence des fonctions im-
plicites, Bulletin des Sciences Mathematique 23 (1899), 68-75, Deuxieme Serie.
27. E. Lindelof, Le Calcul des Residus et ses Applications it la TMorie des
Fonctions, Chelsea Publishing Company, New York, NY, 1947, preface dated
November 13, 1904.
28. A. I. Markushevich, Theory of Functions of a Complex Variable, vol. II,
PRH, 1965, revised English edition, translated and edited by R. A. Silverman.
29. A. Nijenhuis, Strong Derivatives and Inverse Mappings, American Math-
ematical Monthly 81 (1974),969-980.
30. W. F. Osgood, Analytische Funktionen komplexer Groessen, in Encyk-
lopaedie de Mathematischen Wissenschaften, vol. Zweiter Teil. Analysis, Teub-
ner, Leipzig, 1901-1921,1-114, originally published in "Heft 1," whose date was
27.XII.1901.

30
 31. W. F. Osgood, Lehrbuch der Funktionentheorie, vol. I, Teubner, Leipzig,
1907.
32. , Lehrbuch der Funktionentheorie, vol. 11.1, Teubner, Leipzig,
1929, 2nd edition.
33. P. A. Samuelson, Foundations of Economic Analysis, Harvard University
Press, Cambridge, MA, 1947.
34. K. Weierstrass, Abhandlungen aus der Fun ction enlehre , Julius Springer,
Berlin, 1886.
35. W. H. Young, On Implicit Functions and their Differentials, Proceedings
of the London Mathematical Society 7 (1909), 397-421, Second Series.

31