You are on page 1of 13

MATH 163, Spring 2020

SCRIPT 19: Differentiation in Rn

Throughout this sheet, let A be an open subset of Rn .

Definition 19.1. A linear transformation ϕ : Rn → Rm is a function such that for any x, y ∈ Rn


and for any λ ∈ R,

(a) ϕ(x + y) = ϕ(x) + ϕ(y), and

(b) ϕ(λx) = λϕ(x).

That is, ϕ is a linear transformation if it respects the two operations in Definition 14.2.

Lemma 19.2. Let ϕ : Rn → Rm be a linear transformation. Then ϕ(0) = 0.

Proof. ϕ(0) can be rewritten as ϕ(0 · 0) since 0 · x = 0 by Script 18. As such, by Script 19.1,
ϕ(0 · 0) = 0 · ϕ(0) = 0, also by Script 18.

Exercise 19.3. We denote x ∈ R2 by x = (x, y). Determine whether the following functions are
linear transformations:

(a) ϕ : R2 → R, ϕ(x, y) = x + y.

Proof. Taking any two vectors x, y ∈ Rn , the additivity of the transformation holds as ϕ(x)+
ϕ(y) = x1 + x2 + y1 + y2 = ϕ(x + y). Also, the second condition of multiplicativity holds, as
ϕ(λx) + ϕ(λy) = λx1 + λx2 = λ(x1 + x2 ) = λϕ(x). Thus, this is a linear transformation.

(b) ϕ : R2 → R2 , ϕ(x, y) = (x, y + 1).

Proof. This is not a linear transformation, as the linearity condition does not hold: ϕ(x+y) =
x1 + y1 + x2 + y2 + 1 6= x1 + y1 + x2 + y2 + 2 = ϕ(x) + ϕ(y).

(c) ϕ : R2 → R3 , ϕ(x, y) = (3x − y, x + 2y, 0).

Proof. This is a linear transformation. Let a = (x, y), b = (m, n) and let λ ∈ R. As such:

ϕ(λa + b) = ϕ(λx + m, λy + n) = (3λx + 3m − λy − n, λx + m + 2λy + 2n, 0)

= λ(3x − y, x + 2y, 0) + (3m − n, m + 2n, 0) = λϕ(a) + ϕ(b)


.

(d) ϕ : R2 → R3 , ϕ(x, y) = (x2 , x + y, x + y 3 ).

1
Proof. This is not a linear transformation. This is because ϕ(λx, λy) = (λ2 x2 , λx + λy, λx +
λ2 y 3 ), and λϕ(x, y) = (λx2 , λx + λy, λx + λy 2 ). As such, since ϕ(λx, λy) 6= λ(ϕ(x, y)), then,
by Script 19.1, this is not a linear transformation.

Question 19.4. a) Let ϕ : R → R be a linear transformation. What does the graph of ϕ look
like?

Proof. The graph of ϕ(x) is a line passing through the origin, as ϕ(x) = cx for all x ∈ R, for
some c ∈ R.

b) Let ϕ : R2 → R be a linear transformation. What does the graph of ϕ look like?

Proof. The graph of ϕ(x) would be a plane passing through the origin, as ϕ(x) = ax1 + bx2
for some a, b ∈ R and x ∈ R2 .

Exercise 19.5. (a) Let ϕ : Rn → Rm and ψ : Rm → R` be linear transformations. Prove that


ψ ◦ ϕ is also a linear transformation.

Proof. Since ϕ and ψ are linear transformations, then ψ ◦ ϕ(x + y) = ψ(ϕ(x) + ϕ(y)) =
ψ(ϕ(x)) + ψ(ϕ(y)). Also, similarly: ψ ◦ ϕ(λx) = ψ(λϕ(x)) = λψ(ϕ(x)). Thus, ψ ◦ ϕ is a linear
transformation, as it satisfies both conditions.

(b) Let ϕ, ψ : Rn → Rm be linear transformations and let λ ∈ R. Prove that ϕ + ψ and λϕ are
linear transformations.

Proof. Let x, y ∈ Rn . Since ϕ and ψ are linear transformations, then ϕ(x) + ϕ(y) = ϕ(x + y),
and similarly for ψ. Since (ϕ + ψ)(x) = ϕ(x) + ψ(x) (similarly for y), then (ϕ + ψ)(x + y) =
ϕ(x + y) + ψ(x + y).
Now, (ϕ + ψ)(λx) = ϕ(λx) + ψ(λx) = λϕ(x) + λψ(x) = λ(ϕ + ψ)(x).
As such, this is a linear transformation.

(c) Let πI : Rm → Rk be the projection function from Definition 14.33. Prove that πI is a linear
transformation.

Proof. For any x, y ∈ Rm , πI (x + y) = (x1 + y1 , x2 + y2 , . . . , xk + yk ) = (x1 , x2 , . . . , xk ) +


(y1 , y2 , . . . , yk ) = πI (x) + πI (y).
Also, π(λx) = (λx1 , . . . , λxk ) = λ(x1 , . . . , xk ) = λπI (x).
As such, πI is a linear transformation.

For any vector x ∈ Rn , let (x)j = πj (x). That is, if x = (x1 , x2 , . . . , xn ), then (x)j = xj .

2
Definition 19.6. The jth standard basis vector in Rn is the vector ej defined by
(
1 if k = j
(ej )k =
0 if k 6= j.

For example, the standard basis vectors for R3 are e1 = (1, 0, 0), e2 = (0, 1, 0), and e3 = (0, 0, 1).
Notice that if x = (x1 , . . . , xn ) ∈ Rn , then x = x1 e1 + x2 e2 + · · · + xn en .
Definition 19.7. For any linear transformation ϕ : Rn → Rm , we denote by [ϕ]ij the ith component
of the vector ϕ(ej ); i.e., [ϕ]ij = ϕi (ej ).
Exercise 19.8. (a) Let ϕ : Rn → Rm be a linear transformation and let x ∈ Rn . Find a formula
for ϕ(x) in terms of the [ϕ]ij , the components of x, and the standard basis vectors in Rm .

Proof. The formula can be written as follows:


m X
X n
[ϕ]ij xj ei .
i=1 j=1

(b) For 1 ≤ i ≤ m and 1 ≤ j ≤ n, let Aij ∈ R. Prove that there is a unique linear transformation
ϕ : Rn → Rm such that [ϕ]ij = Aij , for all i, j.
Pm Pn
Proof. From part a, let ϕ(x) = i=1 j=1 Aij xj ei .
Let there exist a ϕ0 (x) such that ϕ0 (x) 6= ϕ(x) for some x ∈ Rn . As such, this would imply
0
that [ϕij ] 6= [ϕij ] for some i, j, which would be a contradiction as they should both be equal
to Aij . As such, ϕ(x) is unique.

Definition 19.9. We define an m × n matrix M to be an array of scalars:


 
a11 a12 · · · a1n
 a21 a22 · · · a2n 
M = {aij } =  .
 
.. .. .. 
 .. . . . 
am1 am2 ··· amn
So aij denotes the scalar in row i, column j of the matrix. For every linear transformation
ϕ : Rn → Rm there is a corresponding m × n matrix {[ϕ]ij }. We denote {[ϕ]ij } by [ϕ]. Also, by
Exercise 19.8, given a matrix of scalars, there is a unique linear transformation ϕ : Rn → Rm that
corresponds to it.
Exercise 19.10. a) Let ϕ : R3 → R2 be given by ϕ(x, y, z) = (3x + 2y − z, 4x − 5y + 2z). Write
down the matrix [ϕ].

Proof.  
3 2 −1
4 −5 2

3
b) What is the linear transformation that corresponds to the matrix
 
−2 3
 4 6 ?
1 0

Proof. The linear transformation is ϕ(x, y) = (−2x + 3y, 4x + 6y, x).

Theorem 19.11. Let ϕ : Rn → Rm be a linear transformation. Then there is a constant Mϕ ∈ R


such that for all x ∈ Rn , we have kϕ(x)k ≤ Mϕ kxk.

Proof. Using the additive properties and multiplicative of ϕ(x), then kϕ(x)k = kϕ(x1 e1 + . . . +
xn en )k = kx1 ϕ(e1 ) + . . . + xn ϕ(en )k. From the triangle inequality, this means that kϕ(x)k ≤
x1 kϕ(e1 )k + . . . + xn kϕ(en )k. As such, there must exist some vector a such that ai ≥ kϕ(ei )k, and
kϕ(x)k ≤ |x1 |a1 + . . . + |xn |an ≤ kxkkak. Then, let Mϕ = kak.

Corollary 19.12. Any linear transformation ϕ : Rn → Rm is uniformly continuous.

Proof. Let  > 0, and let δ ∈ R such that δ = |Mϕ|+1 where Mϕ is the same as from Script 19.11.
As such, since kx−yk < δ, by the definition of uniform continuity, then kx−yk(|Mϕ |+1) < . Since
kϕ(x) − ϕ(y)k = kϕ(x − y)k because ϕ is a linear transformation, so kϕ(x) − ϕ(y)k = kϕ(x − y)k <
kx − yk(|Mϕ + 1| <  by Script 19.11. Therefore, since for all  > 0, there exists a delta such that
kϕ(x) − ϕ(y)k < , then ϕ is uniformly continuous.

kϕ(h)k
Lemma 19.13. Let ϕ : Rn → Rm be a linear transformation. If lim = 0, then ϕ is the
h→0 khk
zero transformation, i.e. ϕ(x) = 0, for all x.

Proof. If kxk =
6 0, then let λ ∈ R. As such,

kxk kλϕ(x)k kϕ(λx)k


kϕ(x)k = kϕ(x)k = kxk = kxk
kxk kλxk kλxk

Also, since limλ→0 λx = 0, then limλ→0 kϕ(λx)k


λx = 0 since limh→0 kϕ(h)k
h = 0. As such, limλ→0 kxk kϕ(λx)k
λx =
0. Therefore, kϕ(x)k = 0. So, by Script 18.9, ϕ(x) = 0.
If kxk = 0, then, by Script 18.9, x = 0, and, by Script 19.2, ϕ(x) = 0.

For the remainder of this sheet, we assume that A is an open subset of Rn .

Definition 19.14. A function f : A → Rm is differentiable at a point a ∈ A if there exists a linear


transformation ϕ : Rn → Rm such that

kf (a + h) − f (a) − ϕ(h)k
lim = 0.
h→0 khk

When such a linear transformation ϕ exists, it is called the (total) derivative of f at a and is
denoted by Df (a).

4
Remark 19.15. For every a ∈ A at which f is differentiable, Df (a) : Rn → Rm is a linear
transformation defined on all of Rn . In particular, Df (a)(x) is the derivative of f at a ∈ A,
evaluated at x ∈ Rn .

Proposition 19.16. The derivative at a ∈ A of a function f : A → Rm is unique. That is, if ϕ


and ψ are two linear transformations that satisfy the limit of Definition 19.14, then ϕ = ψ. So,
Df (a) is well-defined.

Proof. If both ϕ and ψ satisfy Script 19.14, then

kf (a + h) − f (a) − ϕ(h)k kf (a + h) − f (a) − ψ(h)k


lim = 0 = lim
h→0 khk h→0 khk

As such, for all  > 0 there must exist a δ1 , δ2 > 0 such that if h < δ1 , δ2 , then:

1.
kf (a + h) − f (a) − ϕ(h)k 
lim <
h→0 khk 2
2.
kf (a + h) − f (a) − ψ(h)k 
lim <
h→0 khk 2

As such, by Script 18.9c, then k kψ(h)−ϕ(h)k


khk k = 0. By Script 19.5b, this means that ψ − ϕ is
a linear transformation, and so, by Script 19.13, ψ − ϕ = 0. As such, ψ = ϕ.

Exercise 19.17. (a) Let f : A → Rm be a constant function, and a ∈ A. Then Df (a) = 0.


Note that here 0 represents the zero transformation.

Proof. Since f is a constant function, then, for all a ∈ A, f (a) = x where x is some
constant in Rm . By Script 19.14, f is differentiable if there exists an LT ϕ such that
limh→0 kf (a+h)−f
khk
(a)−ϕ(h)k
= 0. Since f (a+h = f (a), then ϕ must be such that limh→0 kϕ(h)k
khk =
0. By Script 19.13, this means that ϕ exists and is the 0 transformation. As such, by Script
19.14, Df (a) = ϕ = 0.

(b) Let f : Rn → Rm be a linear transformation, and a ∈ Rn . Then Df (a) = f .

Proof. Let  > 0 and let δ = . As such, WTS that for any  there must exist a δ > 0 such
that if h ∈ A and 0 < kh − 0k < δ then k kf (a+h)−f
khk
(a)−f (h)k
k.
As such, if 0 < khk < δ then

kf (a + h) − f (a) − f (h)k f (a) − f (a) + f (h) − f (h)


k k−0=k k−0
khk khk

Therefore, k0k < δ = . So, D(f (a)) = f .

5
Exercise 19.18. Let f : R → R and a ∈ R. Show that f is differentiable at a in the sense of
Definition 19.14 if and only if f is differentiable at a in the sense of Definition 10.1. Show that if
f is differentiable at a (in either sense), then Df (a)(x) = f 0 (a)x. Are these two uses of the word
‘differentiable’ consistent?
(→) For the forward direction: Let f be differentiable in the sense of Script 19.14. As such,
there must exist a linear transformation ϕ : R → R such that limh→0 |f (a+h)−f (a)−ϕ(h)
|h| = 0. Also,
ϕ must be of the form ϕ(x) = cx for some constant c ∈ R, due to the 1x1 matrix representation of
ϕ. Therefore, limh→0 |f (a+h)−f
|h|
(a)−ch
= 0.
Let  > 0. This means that there must exist a δ > 0 such that if 0 < khk < δ, then

f (a + h) − f (a) − ch
| |<
h
and
f (a + h) − f (a)
| − c| < 
h
This proves that limh→0 |f (a+h)−f |h|
(a)|
= c, and so f 0 (a) = c. Furthermore, since ϕ = Df (a),
then Df (a)(x) = ϕ(x) = cx = f 0 (a)x.
(←) For the reverse direction: Let f be differentiable in the sense of Script 12.1. This means
that limh→0 |f (a+h)−f
|h|
(a)|
= f 0 (a). Let ϕ(x) = f 0 (a)(x). Since f 0 (a) is a constant, then ϕ is an LT.
Now, let  > 0. Then, there must exist a δ > 0 such that if 0 < |h| < δ then
f (a + h) − f (a)
| − f 0 (a)| < 
h
and
f (a + h) − f (a) − f 0 (a)h f (a + h) − f (a) − ϕ(h)
| |=| |<
h h
|f (a+h)−f (a)−ϕ(h)
This allows for and satisfies the fact that limh→0 |h| = 0. As such, since ϕ =
Df (a), then Df (a)(x) = ϕ(x) = f 0 (a)(x).
Theorem 19.19. No proof in class
If f : A → Rm is differentiable at a ∈ A, then f is continuous at a.
Theorem 19.20. No proof in class
(a) Let f, g : A → Rm be differentiable at a ∈ A. Then f + g is differentiable at a, and

D(f + g)(a) = Df (a) + Dg(a).

(b) Let f : A → Rm be differentiable at a ∈ A, and let λ ∈ R. Then λf is differentiable at a, and

D(λf )(a) = λDf (a).

Theorem 19.21 (Chain Rule). No proof in class


Let f : A → Rm and g : B → R` such that f (A) ⊂ B ⊂ Rm . Suppose f is differentiable at a ∈ A
and g is differentiable at f (a). Then g ◦ f is differentiable at a, and

D(g ◦ f )(a) = Dg(f (a)) ◦ Df (a).

6
Remark 19.22. Definition 19.14 can be hard to work with. We now introduce partial derivatives
which provide us with an easy way to compute derivatives, when we know that a function is
differentiable. (See Theorems 19.28 and 19.33 below.) We begin by considering real-valued functions
f : A → R.
Definition 19.23. Let f : A → R, and let a ∈ A. For any 1 ≤ j ≤ n, the limit
f (a + hej ) − f (a)
∂j f (a) = lim ,
h→0 h
if it exists, is called the jth partial derivative of f at a.
Question 19.24. Let f : R2 → R. What is the geometric interpretation of ∂1 f (a) and ∂2 f (a)?
Proof. Let f (x, y) → R, and let a = (x0 , y0 ). As such, ∂1 f (a) = ∂f ∂f
∂x (a), and ∂2 f (a) = ∂y (a). Let
us focus on ∂1 f (a).
Imagine any graph(f ) ∈ R3 .
This means that y0 is held constant while x0 is varied. As such, imagine a plane parallel to the
xz plane, located at y = y0 , intersecting f . As you move in the +x and −x directions on this plane
intersecting f , the slope of x changes. In other words, the partial derivative ∂1 (a) is the slope of
the tangent line to the intersection of graph(f ) with the y = y0 plane, and similarly for ∂2 and the
x = x0 plane. Also, ∂1 f (a) measures the rate of change of f with respect to x as you move along
the plane y = y0 . 0

Proposition 19.25. Let f : A → R. Let a ∈ A and let η > 0 be such that B(a, η) ⊂ A. Define
g : (−η, η) → R by g(h) = f (a + hej ). Fix h0 ∈ (−η, η). Then ∂j f exists at a + h0 ej if and only if
g 0 exists at h0 . Moreover, if they both exist, then ∂j f (a + h0 ej ) = g 0 (h0 ).
(→) For the forward direction: Let 1 ≤ j ≤ n. Now, since h0 ∈ (−η, η) then a+h0 ej ∈ B(a, η) ⊂
f (a+h0 ej +hej )−f (a+h0 ej
A. By Script 19.23, ∂j f (a + h0 ej ) = limh→0 h = limh→0 g(h0 +h)−g(h)
h . As
0
such, by Scirpt 12.1, g (ho ) = ∂j f (a + h0 ej .
(←) For the reverse direction: By Script 12.1,
g(h0 + h) − g(h0 ) f (a + (h0 + h)ej ) − f (a + h0 ej ) f (a + h0 ej + hej ) − f (a + h0 ej )
g 0 (h0 ) = lim = lim = lim
h→0 h h→0 h h→0 h
Therefore, by Script 19.32, g 0 (h0 ) = ∂j f (a + h0 ej ).
Remark 19.26. In particular Proposition 19.25 says that if ∂j f (a) exists if, and only if, g 0 (0)
exists. Moreover if they exist, they are equal.
Exercise 19.27. Use Proposition 19.25 to compute the following partial derivatives:
(a) Let f : R3 → R, f (x) = kxk2 . Let a ∈ R3 . Find ∂3 f (a).

Proof. Let g(h) = f (a + he3 ). By Script 19.26, if g 0 (0) exists, then it must be equal to
∂3 f (a).

g(h) = ka + he3 k2 = ka1 e1 + a2 e2 + (a3 + h)e3 k2 = a21 + a22 + a23 + 2a3 h + h2

As such, g 0 (h) = 2a3 + 2h, so g 0 (0) = 2a3 = ∂3 f (a).

7
(b) Let G : R2 → R be given by
 xy
 if x 6= (0, 0)
G(x, y) = x2 + y2
0 if x = (0, 0).

Let (a, b) ∈ R2 . Find ∂1 G(a, b) and ∂2 G(a, b).

Proof. By Script 19.25, g10 (h) = ∂1 G((a, b) + he1 and g20 (h) = ∂2 G((a, b) + he2 . As such,
g10 (0) = ∂1 G(a, b) and g20 (0) = ∂2 G(a, b).
As such, ∂1 G(a, b) = limh→0 g1 (h)−g
h
1 (0)
and ∂2 G(a, b) = limh→0 g2 (h)−g2 (0)
h . Also, g1 (h) =
(a+h)(b)
G(a + h, b) = (a+b)2 +b2 and g2 (h) = G(a, b + h) = a(a)(b+h)
2 +(b+h)2 .

Through tedious algebra, the final results are:

b3 − a2 b
∂1 G(a, b) =
a4 + 2a3 b2 + b4
a3 − ab2
∂2 G(a, b) =
a4 + 2a3 b2 + b4
.

1/3 1/5
(c) Let H : R2 → R be defined by H(x) = x1 x2 . Find ∂1 H(0) and ∂2 H(0).

Proof. Let x1 = x2 = 0. Let f (h) = H(0 + he1 and let g(h) = H(0 + he2 . By Script 19.25,
f 0 (h) = ∂1 H(0 + he1 , and similarly for g 0 (h). As such, f 0 (0) = ∂1 H(0), and g 0 (0) = ∂2 H(0).
Since f (h) = (0 + h)1/3 01/5 = 0, and g(h) = 01/3 (0 + h)1/5 = 0, so f (0) = g(0) = 0. Since 0
is a constant, then f 0 (0) = g 0 (0) = 0. Therefore, ∂1 H(0) = 0 and ∂2 H(0) = 0.

Theorem 19.28. If f : A → R is differentiable at a ∈ A, then the jth partial derivative ∂j f (a)


exists for every 1 ≤ j ≤ n.
Moreover,

[Df (a)] = [ ∂1 f (a) ∂2 f (a) · · · ∂n f (a) ].


It follows that if x ∈ Rn , then
n
X
Df (a)(x) = xj ∂j f (a).
j=1

(Hint: For the first part, use the Chain Rule.)


kf (a+h)−f (a)−Df (a)(h)k
Proof. Since f is differentiable at a, then limh→0 khk = 0, where Df (a) is an LT.
kf (a+gej )−f (a)−Df (a)(gej )k f (a=gej )−f (a)
Let h = gej . Now, for limg→0 kgej k = 0. This means that limg→0 ( g −
gDf (a)(ej ) f (a+gej )−f (a)
g ) = 0, and so limg→0 g = Df (a)(ej ) = ∂j f (a).
As such, [Df (a)] = 1[Df (a)e1 , . . . , Df (a)e
Pn ] = [∂1 f (a), . . . , ∂n f (a)].
If x ∈ Rn , then Df (a)(x) = [Df (a)]x = nj=1 xj ∂j f (a).

8
Exercise 19.29. Prove that G and H as in Exercise 19.27 b) and c) are not differentiable at 0.
Deduce that the converse to Theorem 19.28 is false. (See Remark 19.35 and Theorem 19.36 below.)

Proof. G. By Script 19.28 and Script 19.27, Df (a))(x) = x1 ∂1 f (a) + x2 ∂2 f (a) = 0. Now,
consider all h such that x = y. Then, limh→0 kG(0+h)−G(0)−ϕ(h)k
khk = limx→0 kG(x,x)k

| 2x|
=
x2
| |
limx→0 2x2
√ = limx→0 √1 . This limit does not exist, so G is not differentiable.
| 2x 2 2x

H. By Script 19.28 and Script 19.27, Df (a))(x) = x1 ∂1 f (a) + x2 ∂2 f (a) = 0. Now, consider
1/3 1/5
all h such that x = y. Then, limh→0 kH(0+h)−H(0)−ϕ(h)k
khk = limx→0 kH(x,x)k

| 2x
| = limx→0 |x |√2x|
x |
.
We know that this limit does not exist, and so H is not differentiable.
√ √
Converse. Let  = 2, and assume that there exists a δ > 0 such that if 0 < |x| < δ then √2x17/15 < 2.
Let y = min( 2δ ), 215/7
1
, and so there must exist some x < y such that 1/x ∈ N. Then,
√ 7/15 √ 7/15 15/7·7/15 √
| 2x | < | 2y | and, as such, |√2x17/15 | > |√2y17/15 | ≥ 2 √2 = 2. This, however, is a
contradiction, and so no such δ exists, and the limit is not equal to 0.

Exercise 19.30. Suppose that f, g : A → R are differentiable at a ∈ A.

(a) Compute ∂j (f + g)(a) in terms of ∂j f (a) and ∂j g(a).

(f +g)(a+hej )−(f +g)(a) f (a+hej )−f (a)


Proof. By Script 19.23, ∂j (f +g)(a) = limh→0 h = limh→0 h +
g(a+hej )−g(a)
limh→0 h . Then, by Script 19.23, ∂j (f + g)(a) = ∂j f (a) + ∂j g(a).

(b) Compute ∂j (f g)(a) in terms of f (a), g(a), ∂j f (a) and ∂j g(a).

(f g)(a+hej )−(f g)(a)


Proof. By Script 19.23, ∂j (f g)(a) = limh→0 h
f (a+hej )g(a+hej )−f (a)g(a+hej )+f (a)g(a+hej )−f (a)g(a)
= limh→0 h
f (a+hej )−f (a) g(a+hej )−g(a)
= (limh→0 g(a + hej ))(limh→0 h ) + (limh→0 f (a))(limh→0 h )
As such, by Script 19.23, ∂j (f g)(a) = g(a)∂j f (a) + f (a)∂j g(a).

(c) Compute ∂j ( f1 )(a) in terms of ∂j f (a) and f (a). What assumption(s) do you need to make?

Proof. It must be assumed that f (a) 6= 0. As such, there must exist a δ > 0 such that if
0 < |h|δ then f (a + hej ) 6= 0. To this end:

1 1
1 f (a + hej ) − f (a) f (a) − f (a + hej ) f (a + hej ) − f (a)
∂j (a) = lim = lim = lim
f h→0 h h→0 h · f (a + hej )f (a) h→0 h

−1 −1 −1 −∂j f (a)
As such, limh→0 f (a+hej )f (a) = ∂j f (a) · f (a) · limh→0 f (a+hej ) = ∂j f (a) f (a)−1 1 = f (a)2
.
f (a)

9
(d) Compute ∂j ( fg )(a) in terms of f (a), g(a), ∂j f (a), and ∂j g(a). What assumption(s) do you
need to make?

Proof. It must be assumed that g(a) 6= 0 as such, there must exist a δ > 0 such that if
0 < |h|δ then g(a + hej ) 6= 0. To this end:

f f
f g (a + hej ) − g (a) f (a + hej )( g1 (a + hej ) − g1 (a)) 1
g (a)(f (a + hej ) − f (a))
∂j (a) = lim = lim + lim
g h→0 h h→0 h h→0 h
1 1
= lim f (a + hej ) · ∂j (a) + lim (a) · ∂j f (a)
h→0 g h→0 g

∂j f (a)g(a)−∂j g(a)f (a)


As such, since f is continuous, then ∂( fg )(a) = g(a)2
.

Results (a), (b), and (d) are known as the Sum Rule, the Product Rule, and the Quotient Rule for
partial derivatives, respectively.
(Hint: Use Proposition 19.25.)

Exercise 19.31. For each of the following, find all given partial derivatives. No justification is
required.

a) f (x, y) = x2 y + 5x − 6ey x.

Proof. (a) ∂x f = 2xf + 5 − 6ey


(b) ∂y f = x2 − 6ey x

b) f (x, y, z) = cos(zexy ) sin(z 2 )x.

Proof. (a) ∂x f = sin(z 2 )(cos(zexy − xyzexy sin(zexy ))


(b) ∂y f = − sin 9z 2 )x2 zexy sin(zexy )
(c) ∂z f = x(2z cos(z 2 ) cos(zexy ) − exy sin(z 2 ) sin(zexy))

We now consider functions f : A → Rm .

Proposition 19.32. Let f : A → Rm and let a ∈ A. Then f is differentiable at a if, and only if,
fi is differentiable at a for 1 ≤ i ≤ m. Moreover, if f is differentiable at a, then, for 1 ≤ i ≤ m,

(Df (a))i = Dfi (a).

Hint: For the forward direction use the projection function and the Chain Rule.

10
Proof. (→) For the forward direction: Let f be differentiable at a. As such, by Script 18.37,
fx = πi ◦ f . By Script 19.5, πi is a linear transformation, so by Script 19.17, Dπi (a) = πi . So πi
is differentiable, and, by the Chain Rule, Dfi (a) must exist. As such, Dfi (a) = Dπi (f (a))◦Df (a) =
πi ◦ Df (a) = (Df (a))i .
(←) For the reverse direction: Let fi be differentiable at a for 1 ≤ i ≤ M . By Script 18.37,
f (a) = (fi (a), . . . , fm (a)), and so, by Script 19.6, it is the same as fa (a)ei + · · · + fM (a)em . Now,
since fi is differentiable at a, and all ei are constant, then, by Script 19.20, f is differentiable at
a. As such, Df (a) = (Df1 (a), . . . , Dfm (a)).

Theorem 19.33. Suppose f : A → Rm is differentiable at a ∈ A. Then for every 1 ≤ i ≤ m and


every 1 ≤ j ≤ n, the jth partial derivatives ∂j fi exist and

[Df (a)] = {∂j fi (a)}.

Proof. By Script 19.32, each fi is differentiable at a, and so:


 
Df1 (a)
Df (a) = 
 .. 
. 
Dfm (a)
As such, by Script 19.28, for each fi , then the jth partial derivative exists, and so [Dfi (a)] =
[∂1 fi (a) . . . ∂n f1 (a)]. As such, by Script 19.9,
 
∂1 f1 (a) . . . ∂n f1 (a)
[Df (a)] = {∂j fi (a)} = 
 .. .. .. 
. . . 
∂1 fm (a) . . . ∂n fm (a)

Exercise 19.34. Assuming that f : R2 → R3 given by f (x, y) = (x cos y, x2 y, ey ) is differentiable,


write down [Df (a, b)].

Proof.
[Df (a)] = {(cos(y), 2xy, 0), (−x sin(y), x2 , ey )}
This can also be written as a matrix.
 
cos(y) 2xy 0
[Df (a)] =
−x sin(y) x2 ey

Remark 19.35. We would like a converse to Theorem 19.33. However, Exercise 19.29 shows that
the existence of partials does not always imply that a function is differentiable. The next theorem
(which we will not prove) provides a partial converse.

Theorem 19.36. No proof in class Let a ∈ A and f : A → Rm . For every 1 ≤ j ≤ n and


1 ≤ i ≤ m, suppose that ∂j fi exists at a and is continuous at a. Then f is differentiable at a.

11
Remark 19.37. In light of this theorem, it is now easily seen that the function in Exercise 19.34
is indeed differentiable.

We now give a more accessible version of the chain rule. First we need the following technical
lemma.

Lemma 19.38. For any linear transformations ψ : Rn → Rm and ϕ : Rm → R` , for every 1 ≤ j ≤ n


and every 1 ≤ k ≤ `,
m
X
[ϕ ◦ ψ]kj = [ϕ]ki [ψ]ij .
i=1

Proof. Let x ∈ Rn , and let y = ψ(x) ∈ Rm and let = ϕ(y) ∈ P R` . This means that (ϕ ◦ ψ)(x) = z.
This means that [ψ] is an
Pm m × n matrix, and, as such, yj = nk=1 [ψ]jk xk . Furthermore, [ϕ] is an
` × m matrix, and zi = j=1 [ψ]ij yj . Finally, [ϕ ◦ ψ] is an n` matrix, and so:
m
X n
X
zi = [ϕ]ij [ψ]jk xk
j=1 k=1
n X
X m
zi = ( [ϕij ][ψjk ])xk
k=1 j=1
n
X
zi = [ϕ ◦ ψ]ik xk
k=1
Pm
As such, [ϕ ◦ ψ]ik = j=1 [ϕ]ij [ψ]jk

Theorem 19.39 (Chain Rule in Coordinates). Let f : A → Rm , let f (A) ⊂ B ⊂ Rm , and let
g : B → R` . Suppose f is differentiable at a ∈ A and g is differentiable at f (a). Then, for
1 ≤ k ≤ ` and 1 ≤ j ≤ n,
m
X
∂j (g ◦ f )k (a) = ∂i gk (f (a))∂j fi (a).
i=1

Proof. Let k be such that 1 ≤ k ≤ ` and let j be such that 1 ≤ j ≤ n. Then, ∂j (g ◦ f )k (a) =
[D(g ◦ f )k (a)]j by Script 19.28. This means that ∂j (g ◦ f )k (a) = [D(πk ◦ g ◦ f )(a)]j = [D(gk ◦
f )(a)]j = [Dgk (f (a)) ◦ Df (a)]j , with the last equality because of Script 19.21. Since this, then,
has dimensions Rn → R, then, by applying Script 19.38 inside the 1 × j matrix:

[Dgk (f (a)) ◦ Df (a)]j = [Dgk ((a)) ◦ Df (a)]1j


X
= [Dgk (f (a))]1i [Df (a)]ij
1≤i≤m
X
= ∂i gk (f (a))∂j fi (a)
1≤i≤m

12
Exercise 19.40. a) Define f : R2 −→ R3 by f (x, y) = (x2 , x + y, xy 2 ) and g : R3 −→ R2 by
g(x, y, z) = (x, x + y + 2z). Let (a, b) ∈ R2 . Find [Df (a, b)], [Dg(f (a, b)] and [D(g ◦ f )(a, b)].

Proof. By Script 19.28,


 
2a 0
[Df (a, b)] =  1 1 
2
b 2ab
Also,

Dg1 (x2 , x + y, xy 2 )
   
1 0 0
[Dg(f )] = =
Dg2 (x2 , x + y, xy 2 ) 1 1 2
Finally, by Script 19.39,
 
2a 0
D(g ◦ f )(a, b) =
2b2 + 2a + 1 4ab + 1

b) Define f : R2 −→ R by f (x, y) = exy and define g : R −→ R2 by g(t) = (3t2 , t). Let (a, b) ∈ R2 .
Find [Df (a, b)], [Dg(f (a, b)] and [D(g ◦ f )(a, b)].

Proof. By Script 19.28,

beab aeab
   
[Df (a, b)] = ∂1 f (b) ∂2 f (a, b) =

Also,
     
Dg1 (t) ∂1 g1 (t) 6t
[Dg(t)] = = =
Dg2 (t) ∂2 g2 (t) 1
This also means that:

Dg1 (eab ) 6eab


     
ab ∂1 g1 (f (a, b))
[Dg(f (a, b))] = Dg(e ) = = =
Dg2 (eab ) ∂2 g2 (f (a, b)) 1

Finally, by Script 19.39,

6be2ab 6ae2ab
 
[D(g ◦ f )(a, b)] =
beab aeab

13

You might also like