DIFFERENTIABILITY AND THE DERIVATIVE

PAUL BRESSLER

1. Differentiablility
1.1. Linear approximation. In what follows, A and B are finite dimensional vector spaces, U ⊆ A
is an open subset.
Let a ∈ U . A map f : U → B is said to admit linear approximation at (near) a if there exists a
linear function L : A → B such that the “error” function
 : U − a = {u − a | u ∈ U } → B
determined by
f (a + h) = f (a) + L(h) + (h)
satisfies
1
(1) lim (h) = 0
h→~0 khk

for some choice of norms on A and B.

Exercise: Show that validity of condition (1) does not depend on the choice of norms on A and
B. 
If f admits linear approximation at a we also say that f is differentiable at a.
If, for all a ∈ U , f is differentiable at a we say that f is differentiable.
Example 1.1.
(1) Suppose that f is a constant function, i.e. there is a b ∈ B such that f (u) = b for all u ∈ A.
Then,
f (a + h) = f (a) + 0 + 0,
i.e. L = 0 and  = 0. Clearly,  = 0 satisfied the condition (1). Thus we conclude that a
constant function is differentiable.
(2) Suppose given b ∈ B and a linear map L : A → B. For u ∈ A let f (u) = b + L(u). For
a, h ∈ A
f (a + h) = (b + L(a)) + L(h) + 0,
i.e.  = 0. Clearly,  = 0 satisfied the condition (1). Thus we conclude that f is differen-
tiable.

Exercise: Show that

(1) Let f (x) = c0 +c1 x+c2 x2 +· · ·+cd xd , where ci ∈ R, i = 1, . . . , d. Find a linear approximation
of f near x = 0.
(2) any polynomial function of one variable is differentiable;
1
(3) the function f : R \ {0} → R given by f (x) = is differentiable.
x

1
2 P.BRESSLER

1.2. Uniqueness of linear approximation. We have already seen that the constant function
with value ~0 ∈ B admits a linear approximation with L = 0 and  = 0.
Lema 1.2. Suppose that ~0 = ~0 + L(h) + (h) is a linear approximation of the constant function
with value ~0 ∈ B. Then, L = 0 and, hence,  = 0.
Proof. Since L(h) = −(h), it follows that
1 1
lim L(h) = − lim L(h) = 0.
h→~0 khk h→~0 khk

Let ~0 6= u ∈ A, t > 0. Let h = tu. Then,

1 1
lim L(tu) = lim L(u) = 0,
t→0 ktuk t→0 kuk

which implies that L(u) = 0. Therefore, L = 0. 

Proposition 1.3. Suppose that f is differentiable at a and
f (a + h) = f (a) + L1 (h) + 1 (h)
and
f (a + h) = f (a) + L2 (h) + 2 (h)
are linear approximations of f near a. Then, L1 = L2 and, hence, 1 = 2 .
Proof. Taking the difference of the two linear approximations we find that
0 = 0 + (L1 − L2 )(h) + (1 − 2 )(h).
Since
1 1 1
(1 − 2 )(h) = 1 (h) − 2 (h)
khk khk khk
and
1 1
lim 1 (h) = lim 2 (h) = 0
h→~0 khk h→0 khk
~

it follows that
1
lim (1 − 2 )(h) = 0.
h→~0 khk
Therefore,
0 = 0 + (L1 − L2 )(h) + (1 − 2 )(h).
is a linear approximation of the the constant function with value ~0 ∈ B, hence L1 − L2 = 0. 

2. Derivative
2.1. The derivative at a point. Suppose that A and B are finite dimensional vector spaces,
U ⊆ A is an open subset and a ∈ U .
If f : U → B is a function differentiable at a, the linear approximation
f (a + h) = f (a) + L(h) + (h)
is uniquely determined by the condition (1).
The linear function L is called the derivative of f at a and, henceforth, will be denoted by Da f .
Thus,
f (a + h) = f (a) + (Df )a (h) + (h)
where Da f ∈ Hom(A, B) and  satisfies (1).
Lema 2.1. If f is differentiable at a, then f is continuous at a.
 DIFFERENTIABILITY AND THE DERIVATIVE 3

Proof. Since f is differentiable at a, for h ∈ A “small”

f (a + h) = f (a) + (Df )a (h) + (h)
where
• (Df )a is linear and therefore satisfies lim (Df )a (h) = 0
h→~0
1
• lim (h) = 0, hence
h→~0 khk

1 1
lim k(h)k = lim khk (h) = lim khk · lim (h) = 0.
h→~0 h→~0 khk h→~0 h→~0 khk

It follows that lim f (a + h) = f (a), i.e. f is continuous at a.

h→~0

2.2. The derivative function. Suppose that f : U → B is differentiable (i.e. differentiable at
every point of U ). The assignment a 7→ (Df )a defines the derivative (function) of f , denoted
Df : U 7→ Hom(A, B).
Example 2.2.
(1) Suppose that f is a constant function, i.e. there is a b ∈ B such that f (u) = b for all u ∈ A.
Then,
f (a + h) = f (a) + 0 + 0,
is a linear approximation. Thus, we conclude that a constant function is differentiable and
the derivative of a constant function is the constant function with value 0 ∈ Hom(A, B).
(2) Suppose given b ∈ B and a linear map L : A → B. For u ∈ A let f (u) = b + L(u). For
a, h ∈ A
f (a + h) = (b + L(a)) + L(h) + 0,
is a linear approximation.
i.e.  = 0. Clearly,  = 0 satisfies the condition (1). Thus we conclude that f is differen-
tiable. Thus, we conclude that a function of the form “constant plus linear” is differentiable
and the derivative of u 7→ b + L(u) is the constant function with value L ∈ Hom(A, B).
(3) Suppose that A, B and C are finite dimensional vector space and β : A×B → C is a bilinear
map. For a, s ∈ A, b, t ∈ B,
β(a + s, b + t) = β(a, b) + (β(s, b) + β(a, t)) + β(s, t),
with A × B → C : (s, t) 7→ (β(s, b) + β(a, t)) linear. Thus, β is differentiable if and only if
it satisfies
1
lim β(s, t) = 0,
(s,t)→0 k(s, t)k

where k(s, t)k = max{ksk, ktk}.

For s 6= 0, t 6= 0,
1 1 1
06 kβ(s, t)k = kβ(s, t)k · k(s, t)k 6 kβ(s, t)k · k(s, t)k.
k(s, t)k k(s, t)k2 ksk · ktk
1
Since kβ(s, t)k is bounded, it follows that
ksk · ktk
1
lim kβ(s, t)k · k(s, t)k = 0,
(s,t)→0 ksk · ktk

therefore,
1
lim β(s, t) = 0.
(s,t)→0 k(s, t)k
4 P.BRESSLER

Thus, bilinear maps are differentiable, with

(Dβ)(a,b) (s, t) = β(s, b) + β(a, t).

Exercise: Generalize the above result to multilinear maps. 

(4) Let V be a finite dimensional vector space. Recall that GL(V ) is an open subset of End(V ),
hence it it makes sense to talk about differentiability of maps defined on GL(V ). Let
ι : GL(V ) → GL(V ) : f 7→ f −1
denote the inversion map, i.e. ι(f ) = f −1 .
Let f ∈ GL(V ), h ∈ End(V ) such that f + h ∈ GL(V ). Then,
(a + h)−1 = a−1 − a−1 ha−1 + (a−1 h)2 (1 + a−1 h)a−1
with h 7→ −a−1 ha−1 linear. Thus, ι is differentiable if and only if it satisfies
1
lim (a−1 h)2 (1 + a−1 h)a−1 = 0.
h→0 khk

Since
1 1
0< k(a−1 h)2 (1 + a−1 h)a−1 k ≤ ka−1 k2 khk2 k1 + a−1 hkka−1 k = khkka−1 k3 k1 + a−1 hk
khk khk
and
lim k1 + a−1 hk = 1,
h→0
it follows that
1
lim (a−1 h)2 (1 + a−1 h)a−1 = 0.
h→0 khk
Thus, ι is differentiable and
(Dι)a (h) = −a−1 ha−1 .
2.3. Algebra of linear approximations (basic rules of calculus).
Lema 2.3. Suppose that A, B, C, D are finite dimensional vector spaces, U ⊆ A, V ⊆ B are open
subset, f : U → C, g : V → D.
If f is differentiable at a ∈ U and g is differentiable at b ∈ V , then f × g : U × V → C × D is
differentiable at (a, b) ∈ U × V and
(D(f × g))(a,b) = (Df )a × (Dg)b .
Proof. Since
f (a + s) = f (a) + (Df )a (s) + f (s)
and
g(a + t) = g(a) + (Dg)a (t) + g (t),
if follows that
(f × g)(a + s, b + t) = (f (a + s), g(b + t)) = (f (a), g(b)) + ((Df )a (s), (Dg)b (t)) + (f (s), g (t)).
For s 6= 0, t 6= 0
k(f (s), g (t))k kf (s)k kg (t)k kf (s)k kg (t)k
06 6 + 6 +
k(s, t)k k(s, t)k k(s, t)k ksk ktk
Since
1 1
lim f (s) = 0, lim g (t) = 0
s→0 ksk t→0 ktk
it follows that
1
lim (f (s), g (t)) = 0.
(s,t)→0 k(s, t)k

 DIFFERENTIABILITY AND THE DERIVATIVE 5

Proposition 2.4 (Chain rule). Suppose that A, B and are finite dimensional vector spaces, U ⊆ A,
V ⊆ B are open subsets, a ∈ U , f : U → B satisfies f (U ) ⊂ V , g : V → C.
If f is differentiable at a and g is differentiable at f (a), then the composition g ◦ f : U → C is
differentiable at f (a) and
(D(g ◦ f ))a = (Dg))f (a) ◦ (Df )a .
Proof. Since f is continuous at a, for h sufficiently small, f (a+h) ∈ V so that g(f (a+h)) is defined.
Applying g to
f (a + h) = f (a) + (Df )a (h) + f (h)
we obtain
g(f (a + h)) = g(f (a)) + (Dg)f (a) ((Df )a (h) + f (h)) + g ((Df )a (h) + f (h))
= g(f (a)) + (Dg)f (a) ((Df )a (h)) + ((Dg)f (a) (f (h)) + g ((Df )a (h) + f (h))).
1 1
Since k(Dg)f (a) (f (h))k 6 k(Dg)f (a) k · kf (h)k, it follows that
khk khk
1 1
lim k(Dg)f (a) (f (h))k 6 lim k(Dg)f (a) k · kf (h)k,
h→0 khk h→0 khk
therefore
1
lim (Dg)f (a) (f (h)) = 0.
h→0 khk

It remains to show that

1
lim (g ((Df )a (h) + f (h)) = 0.
h→0 khk

Let  > 0.
1 kg (s)k 
• Since lim g (s) = 0, there exists r > 0 such that, if ksk < r, then <
s→0 ksk ksk 1 + k(Df )a k
ksk
or, equivalently, kg (s)k < .
1 + k(Df )a k
• Since lim ((Df )a (h) + f (h)) = 0, there exists δ > 0 such that, if khk < δ, then k(Df )a (h) +
h→0
f (h)k < r and kf (h)k 6 khk.
Since
kg ((Df )a (h) + f (h))k k(Df )a (h) + f (h)k
<
khk khk(1 + k(Df )a k)
k(Df )a (h)k + kf (h)k k(Df )a kkhk + khk
6 6 = ,
khk(1 + k(Df )a k) khk(1 + k(Df )a k)
it follows that
1
lim (g ((Df )a (h) + f (h)) = 0.
h→0 khk


Corollary 2.5. Suppose that A, B and C are finite dimensional vector spaces, U ⊆ A is an open
subset, a ∈ U , f : U → B g : U → C.
The maps f and g are differentiable at a if and only if the map (f, g) : U → B ×C is differentiable
at a and
(D(f, g))a = ((Df )a , (Dg)a ).
Proof.
⇒ The map (f, g) is equal to the composition
v7→(v,v) f ×g
U −−−−→ U × U −−→ B × C.
6 P.BRESSLER

⇐ The map f is equal to the composition

(f,g) B pr
U −−→ B × C −−→ B.
The map g is equal to the composition
(f,g) C pr
U −−→ B × C −−→ C.

Corollary 2.6 (Sum rule). Suppose that A and B are finite dimensional vector spaces, U ⊆ A is
an open subset, a ∈ U , f, g : U → B.
If f and g are differentiable at a, then f + g is differentiable at a and
(D(f + g))a = (Df )a + (Dg)a .
Proof. The map f + g is equal to the composition
v7→(v,v) f ×g +
U −−−−→ U × U −−→ B × B −
→ B.
+
The addition map B × B −
→ B is linear.

Corollary 2.7 (Leibniz rule). Suppose that A and B are finite dimensional vector spaces, U ⊆ A
is an open subset, a ∈ U , f : U → R, g : U → B.
If f and g are differentiable at a, then g · f is differentiable at a and
(D(f · g))a = (Df )a · g(a) + f (a) · (Dg)a .
Proof. The map f · g is equal to the composition
v7→(v,v) f ×g µ
U −−−−→ U × U −−→ R × B →
− B.
µ
The scalar multiplication map R × B →
− B is bilinear.

Example 2.8. Let V be a finite dimensional vector space. Recall that GL(V ) is an open subset of
End(V ), hence it it makes sense to talk about differentiability of maps defined on GL(V ). Let
ι : GL(V ) → GL(V ) : f 7→ f −1
denote the inversion map, i.e. ι(f ) = f −1 .
We would like to show that ι is differentiable, i.e. find a map GL(V ) 3 f 7→ Lf ∈ Hom(End(V ), End(V ))
such that, for f ∈ GL(V ) and h ∈ End(V ) small
ι(f + h) = ι(f ) + Lf (h) + f (h),
where f (h) satisfies the condition (1). Since the latter condition determines Lf uniquely, it is
important to make the right guess.
If, in fact, the map ι were differentiable, we could calculate its derivative using the rules of
calculus. We are going to do just that and use the result as our guess of Lf .
By definition, ι satisfies ι ◦ Id = 1, i.e., for f ∈ GL(V ), ι(f ) ◦ f = 1. The derivative of the left
hand side is given by
D(ι ◦ Id) = (Dι) · Id + ι ◦ (DId) = (Dι) · Id + ι · Id,
i.e.
(D(ι ◦ Id))f (h) = (Dι)f (h) · f + ι(f ) · h = (Dι)f (h) · f + f −1 · h.
Since D1 = 0 it follows that
(D(ι ◦ Id))f (h) = (Dι)f (h) · f + ι(f ) · h = (Dι)f (h) · f + f −1 · h = 0.
Solving for (Dι)f (h) we obtain
(Dι)f (h) = −f −1 · h · f −1 .
 DIFFERENTIABILITY AND THE DERIVATIVE 7

So, we will take this result as a guess.

Exercise: Show that
(h) = (f + h)−1 − f −1 + f −1 · h · f −1
satisfies (1). 

Departamento de Matemáticas, Universidad de Los Andes

Email address: [email protected]