The chain rule
Based on lecture notes by James McKernan
Theorem 1 (Chain Rule). Let U Rn and let V Rm be two open subsets.
Let f : U V and g : V Rp be two functions. If f is differentiable at p and
g is differentiable at q = f (p), then g f : U Rp is differentiable at p, with
derivative:
D(g f )(p) = (D(g)(q))(D(f )(p)).
It is interesting to untwist this result in specific cases. Suppose we are given
f : R R2
and
g : R2 R.
So f (x) = (f1 (x), f2 (x)) and g = g(y, z). Then
df1
g
g
dx (p)
Df (p) = df
and
Dg(q) = ( (q),
(q)).
2
y
z
dx (p)
So
d(g f )
g
df1
g
df2
= D(g f )(p) = Dg(q)Df (p) =
(q)
(p) +
(q)
(p).
dx
y
dx
z
dx
Example 2. Suppose that f (x) = (x2 , x3 ) and g(y, z) = yz. If we apply the chain
rule, we get
D(g f )(x) = z(2x) + y(3x2 ) = 5x4 .
On the other hand (g f )(x) = x5 , and of course
dx5
= 5x4 .
dx
In general, if f = f (x1 , . . . , xn ) and g = g(y1 , . . . , yn ) then the (i, k) entry of
D(g f )(p), that is
(g f )i
xk
is given by the dot product of the ith row of Dg(q) and the kth column of Df (p),
m
X gi
(g f )i
fj
=
(q)
(p).
xk
y
x
j
k
j=1
If y = f (x) and z = g(y) then we get
m
X zi yj
zi
=
.
xk
yj xk
j=1
We can use the chain rule to prove some of the simple rules for derivatives.
Suppose that we have
f : Rn Rm
and
g : Rn Rm .
Suppose that f and g are differentiable at p. What about f + g? Well there is
a function
a : R2m Rm ,
which sends (~u, ~v ) Rm Rm to the sum ~u+~v . In coordinates (u1 , u2 , . . . , um , v1 , v2 , . . . , vm ),
a(u1 , u2 , . . . , um , v1 , v2 , . . . , vm ) = (u1 + v1 , u2 + v2 , . . . , um + vm ).
1
Now a is differentiable (it is a polynomial, linear even). There is function
h : Rn R2m ,
which sends q to (f (q), g(q)). The composition a h : Rn Rm is the function
we want to differentiate, it sends p to f (p) + g(p). The chain rule says that that
the function is differentiable at p and
D(f + g)(p) = Df (p) + Dg(p).
Now suppose that m = 1. Instead of a, consider the function
m : R2 R,
given by m(x, y) = xy. Then m is differentiable, with derivative
Dm(x, y) = (y, x).
So the chain rule says the composition of h and m, namely the function which sends
p to the product f (p)g(p) is differentiable and the derivative satisfies the usual rule
D(f g)(p) = g(p)D(f )(p) + f (p)D(g)(p).
Here is another example of the chain rule, suppose
x = r cos
y = r sin .
Then
f
f x f y
=
+
r
x r
y r
f
f
cos +
sin .
=
x
y
Similarly,
f
f x f y
=
+
x
y
f
f
= r sin +
r cos .
x
y
We can rewrite this as
=
cos
r sin
sin
r cos
x
Now the determinant of
cos
r sin
sin
r cos
is
r(cos2 + sin2 ) = r.
So if r 6= 0, then we can invert the matrix above and we get
1 r cos sin
x
r
=
r r sin cos
y
We now turn to a proof of the chain rule. We will need:
Lemma 3. Let A Rn be an open subset and let f : A Rm be a function.
If f is differentiable at p, then there is a > 0 such that if kq pk < , then
kf (q) f (p)k < (K + 1)kq pk,
where K is the Frobenius norm of Df (p).
Proof. As f is differentiable at p, there is a constant > 0 such that if kq pk < ,
then
kf (q) f (p) Df (p)(q p)k
< 1.
kq pk
Hence
kf (q) f (p) Df (p)(q p)k < kq pk.
But then
kf (q) f (p)k = kf (q) f (p) Df (p)(q p) + Df (p)(q p)k
kf (q) f (p) Df (p)(q p)k + kDf (p)(q p)k
k(q p)k + Kk(q p)k
= (K + 1)k(q p)k,
Proof of (1). Lets fix some notation. We want the derivative at p. Let q = f (p).
Let p0 be a point in U (which we imagine is close to p). Finally, let q 0 = f (p0 ) (so
if p0 is close to p, then we expect q 0 to be close to q).
The trick is to carefully define an auxiliary function G : V Rp ,
( 0
g(q )g(q)Dg(q)(q 0 q)
if q 0 6= q
0
kq 0 qk
G(q ) =
~0
if q 0 = q.
Then G is continuous at q = f (p), as g is differentiable at q. Now,
(g f )(p0 ) (g f )(p) Dg(q)Df (p)(p0 p)
kp0 pk
0
Dg(q)(f (p ) f (p)) Dg(q)(q 0 q) + g(q 0 ) g(q) Dg(q)Df (p)(p0 p)
=
kp0 pk
f (p0 ) f (p) Df (p)(p0 p) g(q 0 ) g(q) Dg(q)(q 0 q)
= Dg(q)
+
kp0 pk
kp0 pk
f (p0 ) f (p) Df (p)(p0 p)
kf (p0 ) f (p)k
= Dg(q)
+ G(f (p0 ))
.
0
kp pk
kp0 pk
As p0 approaches p, note that
f (p0 ) f (p) Df (p)(p0 p)
,
kp0 pk
and G(p0 ) both approach zero and
kf (p0 ) f (p)k
K + 1.
kp0 pk
So then
(g f )(p0 ) (g f )(p) Dg(q)Df (p)(p0 p)
,
kp0 pk
approaches zero as well, which is what we want.
