Gradients

Math 131 Multivariate Calculus

D Joyce, Spring 2014

∂f
Last time. Introduced partial derivatives like
∂x
of scalar-valued functions Rn → R, also called
scalar fields on Rn .

Total derivatives. We’ve seen what partial

derivatives of scalar-valued functions f : Rn → R
are and what they mean geometrically. But if these
are only partial derivatives, then what is the ‘total’
derivative? The answer will be, more or less, that
the partial derivatives, taken together, form the to-
tal derivative.
Figure 1: A Lyre of Ur
First, we’ll develop the concept of total derivative
for a scalar-valued function f : R2 → R, that is, a
scalar field on the plane R2 . We’ll find that that them. In that sense, the pair of these two slopes
total derivative is what we’ll call the gradient of will do what we want the ‘slope’ of the plane to be.
f , denoted ∇f . Next, we’ll slightly generalize that We’ll call the vector whose coordinates are these
to a scalar-valued function f : Rn → R defined partial derivatives the gradient of f , denoted ∇f ,
on n-space Rn . Finally we’ll generalize that to a or gradf .
vector-valued function f : Rn → Rm . The symbol ∇ is a nabla, and is pronounced
“del” even though it’s an upside down delta. The
The gradient of a function R2 → R. Let f be word nabla is a variant of nevel or nebel, an an-
a function R2 → R. The graph of this function, cient harp or lyre. One is shown in figure 1. See
z = f (x, y), is a surface in R3 . We would like the the wiki article on the Lyres of Ur, the oldest sur-
derivative of f to be the ‘slope’ of the tangent plane. viving stringed instruments.
But a plane doesn’t have a single slope; it slopes When n = 2, the gradient of the scalar field f :
differently in different directions. The plane tan- R2 → R is  
∂f ∂f
gent to this surface and passing through the point ∇f = , .
(a, b, f (a, b)) has the equation ∂x ∂y
∂f ∂f
z = f (a, b) + fx (a, b)(x − a) + fy (a, b)(y − b). Note that the partial derivatives and are
∂x ∂y
actually functions of two variables, R2 → R, there-
Here, fx (a, b) is the partial derivative of f evaluated fore, the gradient is also a function of two variables,
at (a, b), and it’s the slope in the x-direction. Like- but it’s a vector-valued function, ∇f : R2 → R2 ,
wise fy (a, b) is the slope in the y-direction. These
two slopes determine the plane, and we’ll see soon
 
∂f ∂f
how to compute slopes in other directions from ∇f (x, y) = (fx , fy ) = (x, y), (x, y) .
∂x ∂y

1
Thus, whereas f was a scaler field, its gradient ∇f
is a vector field.

Figure 3: Vector field ∇f (x, y)

both these directions, but there might not be a

tangent plane. For the ordinary functions we en-
counter, there is always a tangent plane, but you
can construct weird functions that don’t have them.
Figure 2: z = ye3x+4y The question becomes: how can you tell if a func-
tion has a tangent plane?
We know, if there is a tangent plane, then its
Example. Let f (x, y) = ye 3x+4y
. Its graph is equation will be
shown in figure 2. Note that there’s a deep val- z = f (a, b) + fx (a, b)(x − a) + fy (a, b)(y − b).
ley in the 4th quadrant of the xy-plane. Also, the
surface rises very quickly for positive values of y. So the question becomes: how can you tell if this
First, compute the partial derivatives of f . You’ll plane actually is a tangent plane? The answer is
find that ∂f∂x
= 3ye3x+4y , and, by the product rule, that the linear function
∂f
∂y
= e3x+4y + 4ye3x+4y . Therefore, the gradient of h(x, y) = f (a, b) + f (a, b)(x − a) + f (a, b)(y − b)
x y
f is
  is a good approximation to the function f (x, y) near
∂f ∂f
∇f (x, y) = ∂x
(x, y), ∂y
(x, y) (a, b). More precisely, the vertical distance between
f (x, y) and h(x, y) approaches 0 much faster than
= (3ye3x+4y , e3x+4y + 4ye3x+4y ).
the horizontal distance between (x, y) and (a, b).
The gradient if f is illustrated in figure 3. The That, in turn, can be described in terms of limits
vectors in this vector field point in the direction of ratios as follows:
of fastest ascent. In the 4th quadrant, they point f (x, y) − h(x, y)
lim = 0.
left meaning that the quickest way up out of that (x,y)→(a,b) k(x, y) − (a, b)k
deep valley is to decrease x. In the 1st and 2nd
quadrants, the up the surface is to increase y. (Note that this limit is a vector limit.)
That leads us to our definition of differentiability
for a function f : R2 → R. We’ll say f is differen-
Tangent planes and differentiability. Some-
tiable at (a, b) if, first, the partial derivatives fx and
times, the partial derivatives don’t tell the whole
fy exist, and, second, the linear function
story. There may be derivatives in both the x- and
y-directions, that is, there may be tangent lines in h(x, y) = f (a, b) + fx (a, b)(x − a) + fy (a, b)(y − b)

2
defined by the partial derivatives is a good approx- In summary, for scalar-valued functions f :
imation of f at (a, b) in the sense that Rn → R, the gradient
f (x, y) − h(x, y) 
∂f ∂f ∂f

lim = 0. ∇f = (fx1 , fx2 , . . . , fxn ) = , ,...,
(x,y)→(a,b) k(x, y) − (a, b)k ∂x1 ∂x2 ∂xn
Then, when a function is differentiable, we’ll take
is the derivative.
the gradient, ∇f , which is the vector of partial
derivatives, to be the derivative. Math 131 Home Page at
http://math.clarku.edu/~djoyce/ma131/
Most common functions are differentiable.
Partial derivatives are easy to compute, so it’s easy
to show the first condition of differentiability is sat-
isfied just by exhibiting the derivatives. We expect
the second condition to usually hold, and it does. A
useful theorem (which we won’t prove), says that
if the partial derivatives are continuous, then the
second condition holds, so the function is differen-
tiable. That means, in practice, that the partial
derivatives are enough.

Generalize to scalar-valued functions Rn →

R with n > 2. Given a scalar-valued function
f : Rn → R, we’ll say f is differentiable at a when
it’s partial derivatives fx1 (a), fx2 (a), . . . , fxn (a) ex-
ist and the linear function h(x) =
f (a)+fx1 (a)(x1 −a1 )+fx2 (a)(x2 −a2 )+· · ·+fxn (a)(xn −an )
is a good approximation of f near a in the sense
that
f (x) − h(x)
lim = 0.
x→a kx − ak
Since h(x) can be rewritten in terms of dot products
and the gradient as
h(x) = f (a) + ∇f (a) · (x − a),
that last limit can be also be written as
f (x) − [f (a) + ∇f (a) · (x − a)]
lim = 0.
x→a kx − ak
This last condition is analogous to the condition of
differentiability for ordinary functions f : R → R
that looks like
f (x) − [f (a) + f 0 (a)(x − a)]
lim = 0.
x→a (x − a)