Team Presentation
Kevin Cai, Josh Michel, Dan Warner
March 9, 2015
Theorem 3.10
If a function f : X Rn Rm is such that, for i = 1, ..., m and j = 1, ..., n, all
in a neighborhood of a in X, then f is differentiable at a.
1.1
fi
xj
exist and are continuous
Proof
We start by proving the theorem for f : X R2 R, then extending that proof to functions f : X Rn
R, which finally allows us to prove it for any f : X Rn Rm .
Proof for f : X R2 R
In order to prove that continuity of each partial derivative at a point implies differentiability at that
point, we must prove, given that fx1 and fx2 are continuous at a point a X, that
f (x1 , x2 ) h(x1 , x2 )
lim
=0
xa
kx ak
where h(x1 , x2 ) is the best linear approximation of f at a.
We begin the proof with the expression
f (x1 , x2 ) f (a1 , a2 ) = [f (x1 , x2 ) f (a1 , x2 )] + [f (a1 , x2 ) f (a1 , a2 )].
Note that we can now apply the Mean Value Theorem to the bracketed expressions on the right by keeping
all but one variable constant:
f (x1 , x2 ) f (a1 , x2 ) = fx1 (c1 , x2 )(x1 a1 ), for some c1 between a1 and x1
f (a1 , x2 ) f (a1 , a2 ) = fx2 (a1 , c2 )(x2 a2 ), for some c2 between a2 and x2
Therefore the original expression becomes
f (x1 , x2 ) f (a1 , a2 ) = fx1 (c1 , x2 )(x1 a1 ) + fx2 (a1 , c2 )(x2 a2 ).
(1)
Recall that we defined the best linear approximation of a scalar-valued function of two variables at
a = (a1 , a2 ) to be
h(x1 , x2 ) = f (a1 , a2 ) + fx1 (a1 , a2 )(x1 a1 ) + fx2 (a1 , a2 )(x2 a2 )
Therefore we have, by a little rearrangement and careful use of Triangle Inequality,
f (x1 , x2 ) h(x1 , x2 ) = f (x1 , x2 ) f (a1 , a2 ) fx1 (a1 , a2 )(x1 a1 ) fx2 (a1 , a2 )(x2 a2 )
= fx1 (c1 , x2 )(x1 a1 ) + fx2 (a1 , c2 )(x2 a2 ) fx1 (a1 , a2 )(x1 a1 ) fx2 (a1 , a2 )(x2 a2 )
fx1 (c1 , x2 )(x1 a1 ) fx1 (a1 , a2 )(x1 a1 ) + fx2 (a1 , c2 )(x2 a2 ) fx2 (a1 , a2 )(x2 a2 )
= fx1 (c1 , x2 ) fx1 (a1 , a2 ) x1 a1 + fx2 (a1 , c2 ) fx2 (a1 , a2 ) x2 a2
Note that, for i = 1, 2,
1
kx ak = [(x1 a1 )2 + (x2 a2 )2 ] 2 xi ai
Therefore,
f (x1 , x2 ) h(x1 , x2 ) fx1 (c1 , x2 ) fx1 (a1 , a2 ) x1 a1 + fx2 (a1 , c2 ) fx2 (a1 , a2 ) x2 a2
fx1 (c1 , x2 ) fx1 (a1 , a2 ) + fx2 (a1 , c2 ) fx2 (a1 , a2 ) kx ak
From this, we get
f (x1 , x2 ) h(x1 , x2 )
kx ak
fx1 (c1 , x2 ) fx1 (a1 , a2 ) + fx2 (a1 , c2 ) fx2 (a1 , a2 )
(2)
Since ci is between ai and xi for i = 1, 2, then as x approaches a, ci gets squeezed towards ai , i.e.
f
x a ci ai . Since we are given that all partial derivatives x
are continuous at a, we have for the
i
right side of (2), by the definition of continuity:
lim fx1 (c1 , x2 ) fx1 (a1 , a2 ) + fx2 (a1 , c2 ) fx2 (a1 , a2 ) = fx1 (a1 , a2 ) fx1 (a1 , a2 ) + fx2 (a1 , a2 ) fx2 (a1 , a2 )
xa
=0
By taking the limit of the left side of (2) as well, we then arrive at the conclusion that f (x1 , x2 ) is differentiable
at a (note that since it must be nonnegative because of the absolute value signs, the limit of the left side is
precisely equal to the limit of the right side as x a, i.e. 0):
f (x1 , x2 ) h(x1 , x2 )
=0
lim
xa
kx ak
We can now extend this result to scalar-valued functions in general.
Proof for f : X Rn R
We follow the logic of the proof for f : X R2 R. In order to prove that continuity of each partial
derivative at a point implies differentiability at that point, we must prove, given that fxi is continuous at a
point a X for i = 1, ..., n, that
f (x) h(x)
=0
lim
xa
kx ak
where h(x) is the best linear approximation of f at a.
We begin the proof with the expression
f (x1 , x2 , ..., xn ) f (a1 , a2 , ..., an ) = [f (x1 , x2 , ..., xn ) f (a1 , x2 , ..., xn )]
+ [f (a1 , x2 , ..., xn ) f (a1 , a2 , ..., xn )] + ...
+ [f (a1 , a2 , ..., ai1 , xi , xi+1 ..., xn ) f (a1 , a2 , ..., ai1 , ai , xi+1 ..., xn )]
+ ... + [f (a1 , a2 , ..., an1 , xn ) f (a1 , a2 , ..., an1 , an )].
Note that we can now apply the Mean Value Theorem to the bracketed expressions on the right by keeping
all but one variable constant:
f (a1 , a2 , ..., ai1 , xi , xi+1 , ..., xn ) f (a1 , a2 , ..., ai1 , ai , xi+1 ..., xn ) = fxi (a1 , a2 , ..., ci , xi+1 , ..., xn )(xi ai ),
for some ci between ai and xi , for i = 1, ..., n
Therefore the original expression becomes
f (x1 , x2 , ..., xn ) f (a1 , a2 , ..., an ) =
n
X
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn )(xi ai ).
i=1
Since, for a scalar-valued function of n variables, the best linear approximation at a = (a1 , ..., an ) is
h(x) = f (a) + f (x a) = f (a) +
n
X
fxi (a)(xi ai ),
i=1
we get that, by a little rearrangement and careful use of Triangle Inequality,
n
X
fxi (a)(xi ai )
f (x) h(x) = |f (x) f (a)
i=1
n
n
X
X
=
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn )(xi ai )
fxi (a)(xi ai )
i=1
i=1
n
X
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn )(xi ai ) fxi (a)(xi ai )
i=1
n
X
=
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn ) fxi (a) xi ai
i=1
Since, for i = 1, ..., n,
kx ak =
X
n
(xi ai )2
i=1
21
xi a i |
(3)
we get
n
X
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn ) fxi (a) kx ak
f (x) h(x)
i=1
from which it follows that
f (x) h(x)
kx ak
n
X
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn ) fxi (a)
(4)
i=1
Since ci is between ai and xi for i = 1, ..., n, then as x approaches a, ci gets squeezed towards ai , i.e.
f
x a ci ai . Since we are given that all partial derivatives x
are continuous at a, we have for the
i
right side of (4), by the definition of continuity:
lim
xa
n
n
X
X
fxi (a) fxi (a)
fxi (a1 , a2 , ..., ci , xi+1 , ..., xn ) fxi (a) =
i=1
i=1
=0
Therefore, for the limit of the left side of (4):
f (x) h(x)
=0
lim
xa
kx ak
Proof for f : X Rn Rm
Now that we have proved Theorem 3.10 for scalar-valued functions, the proof for vector-valued functions
fi
becomes quite simple. What we need to prove is, given that, for i = 1, ..., m and j = 1, ..., n, all x
exist
j
and are continuous in a neighborhood of a, that
lim
xa
kf (x) f (a) Df (a)(x a)k
=0
kx ak
We start by expressing the numerator in terms of its component functions:
f (x) f (a) Df (a)(x a) = (H1 , H2 , ..., Hm ), where Hi = fi (x) fi (a) fi (a)(x a)
Note that in each Hi , we have already established that fi is differentiable at a by the previous proof for
scalar-valued functions. We also note that
(|H1 | + ... + |Hm |)2 = |H1 |2 + ... + |Hm |2 + 2|H1 ||H2 | + 2|H1 ||H3 | + ... + 2|Hm1 ||Hm |
|H1 |2 + ... + |Hm |2
p
|H1 | + ... + |Hm | |H1 |2 + ... + |Hm |2
We can now express the function in the limit as
kf (x) f (a) Df (a)(x a)k
k(H1 , H2 , ..., Hm )k
=
kx ak
kx ak
1
2 2
)
(H12 + H22 + ... + Hm
kx ak
|H1 | + |H2 | + ... + |Hm |
kx ak
|H2 |
|Hm |
|H1 |
=
+
+ ... +
kx ak kx ak
kx ak
From the proof of 3.10 for scalar-valued functions, we know that
lim
xa
|Hi |
|fi (x) fi (a) fi (a)(x a)|
= lim
=0
kx ak xa
kx ak
because all partial derivatives exist and are continuous at a. Therefore,
lim
xa
kf (x) f (a) Df (a)(x a)k
=0
kx ak
