LECTURE 8

Directional Derivatives and the Gradient

→
In this lecture we specialize to the case where f : Rn R is a real-valued function of several variables. For
such a function the differential Df reduces to an 1 × n matrix, or equivalently an n-dimensional vector. In
fact we have



Df = ∂f
∂x1
∂f
∂x2
· ·· ∂f
∂xn
=
∂f ∂f
,
∂x1 ∂x2
, ··· ,
∂f
∂xn
≡ ∇f

so Df can be identified with the gradient of f .

We’ll come back to the gradient in a minute. But first let me introduce the notion of directional
derivatives.

Definition 8.1. Let f be a function from Rn toR, and let u be a unit vector in Rn (i.e, a vector of length
1). Then the directional derivative of f in the direction u at the point x is the limit

f (x + tu ) − f (x )


Duf (x) ≡ dtd f (x + tu) 



≡ lim
→0 t
t =0 t

The directional derivative of f : Rn → R along the direction u at the point x is interpretable as the rate of
change in f as one moves away from the point x in the direction of u.

Remark 8.2. We restrict u to be a unit vector because most often we’re interested only in how a function
changes when we move in different directions. Since, we care only about the direction of u and not its
magnitude; we simply fix its magnitude to be 1.

 
Example 8.3. Compute the rate of change of f : (x, y, z) → x2 yz in the direction u = √13 √13 − √13 at
, ,

the point (1,1,0).

1
 8. DIRECTIONAL DERIVATIVES AND THE GRADIENT 2
We need to compute



Duf(x) = dtd f (1, 1, 0) + t
3 3
√1 , √1 , − √1
3


 t=0
= f 1 + √ t, 1 + √ t, 0 − √ t 
d 1 1 1
dt 3 3 3

 2


t=0

=
d
dt
1+ √
t
y\1 + √
t
− √ 
t
3 3 3






t=0

1
= 2 1+ √
t
√ 1 + √ − √  t t

3 3 3 3

2


t=0

+ 1+ √
t
√ − √ 
1 t

3 3 3

2

t=0
+ 1+ √
t
1+ √
t
− √1 
3 3 3

 t=0

= (2) √1 (1)(0)
3


1
+ (1)2 √ (0)

3 1 
+ (1)2 (1) − √
3
1
= −√
3

Below we give a theorem that makes computations such as the one above a lot simpler.

Theorem 8.4. If f : →
Rn R is differentiable then all directional derivatives exist and, moreover, the
directional derivative of f in the direction u at the point x is given by

∇f (x) · u
Proof. Let γ : R → Rn be the function
γ(t) = x + tu

so that

γ1 (t) = x1 + tu1
γ2 (t) = x2 + tu2
..
.
γn (t) = xn + tun

and

x u
f ( + t ) = f (γ (t))
 8. DIRECTIONAL DERIVATIVES AND THE GRADIENT 3
By the chain rule we have


D x d
x u
u f ( ) = f ( + t )
dt


 t=0
(f ◦ γ )
d
=
dt t=0
D
= (f ◦ γ )|t=0
= Df (γ (0)) Dγ (0)
 dγ1 (0) 

∂f  dt

 
dγ2
∂f ∂f
= (γ (0)) (γ (0)) ( γ (0)) 
(0)

∂x1 ∂x2
···
∂xn 
dt
.
.


.
dγn

 
(0)
dt

u1

∂f  
x  
∂f ∂f u2
=
∂x1
( x)
∂x2
(x ) ···
∂xn
( 
 ) .
.


.

un
∂f ∂f ∂f
= (x) u1 + ( x ) u2 + · · · + ( x) un
∂x1 ∂x2 ∂xn

∂f ∂f ∂f

= ( x) , (x ), · · · , ( x) · (u1 , u2 , . . . , un )
∂x1 ∂x2 ∂xn
= ∇ f ( x) · u

Example 8.5. Let’s return to the preceding example and use our spanking new  formula to compute the
directional derivative of f (x, y, z ) = x2 yz along the direction u = √13 , √13 , − √13 at the point (1, 1, 0).


∇f (1, 1, 0) = ∂f ∂f ∂f
, , 
∂x ∂y ∂z
  (1 1 0) , ,

2xyz, x2 z, x2 y (1 1 0) , ,

= (0, 0, 1)
So
 
f ((1, 1, 0) + tu)
d
dt
= ∇ f (1, 1, 0) · u
t=0


= (0, 0, 1) √
1, 1, √ −√
1
3 3 3
·

= 1 −√
3

The gradient f not only makes the computation of directional derivatives easier, it also makes it easy to
∇

identify the direction in which a function increases most rapidly:

Theorem 8.6. Let f : Rn → R be a differentiable function and assume that ∇f(x) = 0. Then the direction
of ∇f (x) coincides with the direction in which f (x) is increasing most rapidly.
Proof. We want to determine the direction u in which a directional derivative


Duf (x) = dtd f (x + tu) 



t =0
 8. DIRECTIONAL DERIVATIVES AND THE GRADIENT 4
is maximized. Using the preceding theorem we have
Duf (x) = f (x) u ∇ ·

= f (x) u cos(θ) ∇

where θ is the angle between f (x) and u. Since u is, by definition, a unit vector u = 1 and so
∇


f (x + tu)
d
= ∇f (x) cos(θ)


dt =0
t

The right hand side is obviously maximized when θ = 0; i.e. when u points in the same direction as ∇f (x).
Remark 8.7. Another way of phrasing the result of this theorem is that, when one imagines the graph of
f as a surface with hilltops and valleys, the direction of the ∇f (x) corresponds to the direction uphill at
the point x.

Here is another application of the gradient.

Theorem 8.8. Let f : Rn → R be a differentiable function and let x 0 be a point on the level surface
S = {x ∈ Rn | f (x) = k}
∇
Then f (x0 ) is normal to the surface S at the point x0 in the following sense: if v is the tangent vector at
t = 0 to any curve γ (t) that lies within S and satisfies γ (t) = 0, then v f (x0 ) = 0. ·∇
Proof.Let γ (t) be such a curve. Since γ (t) lies in S for all t we must have
f (γ(t)) = k
Therefore,
 
0=
d
dt
◦
(f γ )
t=0
=D D
f (γ (0) γ (0)
 dγ1 (0) 

∂f  dt

 
dγ2

= ( x0 ) ∂f
( x0 ) ∂f
( x )
(0)

∂x1 ∂x2
···
∂xn
0 

dt
.
.


.
dγn

 
(0)
dt

v1

∂f  
x0 )
∂f
( x0 )
∂f 
( x0 ) 
v2
..


=
∂x1
(
∂x2
···
∂xn  . 
vn
= ∇f (x0 ) · v

Because the gradient of f at the point x0 is perpendicular to the tangent vector at x0 to any curve γ (t) that
lives in a level surface S = {x ∈ Rn | f (x) = k} it is reasonable to define the plane tangent to the surface
S at the point x0 in terms of the gradient.
Definition 8.9. Let f : Rn → R be a diffetentiable function and let S be a surface in Rn of the form
S= x{ ∈ R | f
n (x) = k}, the the tangent plane to S at the point x0 is defined by the equation
∇f (x0 ) · (x − x0 ) = 0