Multivariable

Vector-Valued
Functions
Contents
Contents 2
31.1 Multivariable Vector-Valued Functions . . . . . . . . . . . 3
31.2 Derivatives . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
31.3 Linear Approximations & Differentials . . . . . . . . . . . 34
31.4 The Chain Rule . . . . . . . . . . . . . . . . . . . . . . . . 42
31.2 Optional: Limit Definition of the Derivative for Multivari-
able Vector-Valued Functions . . . . . . . . . . . . . . . . 55
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 3

31.1 Multivariable Vector-Valued

Functions

In Calculus I, we studied functions of the form y  f ( x ) , for example

f ( x )  x 2 . Such functions, which have a single variable for input and a
real number for output, are called single-variable real-valued functions.
The graph of such a function is a curve in the x y-plane, as seen for
example in Figure 1 of this section, where the function is f ( x )  x 2 .
4

-2 -1 1 2

-1

-2

Figure 1: Graph of f ( x )  x 2

If we use the standard symbol R to denote the set of real numbers,

then functions of this type can be denoted as f : R → R, where the name
of the function is “ f ” and where the arrow indicates that the function
takes input in R and gives output in R. This arrow notation, which you
might not have used up till now, is widely used in advanced mathematics.
In Calculus II, we added a new type of function to our repertoire,
including functions of the form z  f ( x, y ) , for example f ( x, y )  x 2 − y 2 ,
and of the form w  g ( x, y, z ) , for example g ( x, y, z )  5x y 2 z 3 . Such
functions, which have more than one variable for input and a real number
for output, are called multivariable real-valued functions. The graph of
a function of the form z  f ( x, y ) is a surface in R3 , as seen for example
in Figure 2 of this section, where the function is f ( x, y )  x 2 − y 2 .
If we use the standard symbol Rn to denote n-dimensional space with
real coordinates, then functions of this type can be denoted as f : Rn → R,
where, as before, the name of the function is “ f ” and where the arrow
indicates that the function takes input in Rn and gives output in R. For
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 4

Figure 2: Graph of f ( x, y )  x 2 − y 2

example, the function f ( x, y )  x 2 − y 2 is a function f : R2 → R, and the

function g ( x, y, z )  5x y 2 z 3 is a function g : R3 → R.
Yet a third type of function was discussed earlierf in Vector Calculus,
f (t )
g
where we considered functions of the form r ( t )  g ( t ) , for example
" #
f g f (t ) f cos t g
r (t )  t2
3t+1
, and of the form s ( t )  g (t ) , for example s ( t )  sin t .
h (t ) 2t
Such functions, which have one variable for input and a vector for output,
are called single-variable vector-valued functions. The functions f ( t )
and g ( t ) , which are the component functions of r ( t ) , are each a single-
variable real-valued function, and similarly for the component functions
of s ( t ) .
Graphing single-variable vector-valued functions is different from
graphing single-variable real-valued and multivariable real-valued func-
tions. We graph functions of the form y  f ( x ) and z  f ( x, y ) by having
a separate axis for each of the variables, both the “input” variables and the
“output” variable. By contrast, for graphing single-variable vector-valued
functions we do not given an axis for the “input” variable, for example
f cos t g
the variable t in the function s ( t )  sin t , but rather we have axes only
2t
for the “output” variables (in this example three axes). So, the picture
we see of a single-variable vector-valued function is technically not the
graph of the function, but rather the image of the function. See Figure 3
f cos t g
of this section for the image of s ( t )  sin t .
2t
Single-variable vector-valued functions can be denoted as r : R → Rn ,
where, as before, the name of the function is “r” and where the arrow
indicates that the function takes
f input
g in R and gives output in Rn . For
example, the function r ( t )  3t+1
t2 is a function r : R → R2 , and the
f cos t g
function s ( t )  sin t is a function s : R → R3 .
2t
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 5

1.0
0.5
0.0
-0.5
-1.0

20

10

1.0
0.5
0.0
-0.5
0
-1.0

Figure 3: Graph of s ( t )

We now want to introduce a new type of function that includes, and

generalizes, all three of the previous types of functions.

The Definition of Multivariable Vector-Valued Functions

The new type of function we consider, called multivariable vector-
valued functions, are functions of the form F : Rn → Rm , where n and m
are positive integers.
There is an unfortunate ambiguity in the way we think about functions
of the form F : Rn → Rm because there is an ambiguity in the way we
think about each of Rn and Rm . Specifically, we can think of each of Rn
and Rm as consisting of either points or vectors; these two points of view
are each useful in different situations. Hence, a function of the form
F : Rn → Rm could be thought of as taking points to points, or points to
vectors, or vectors to points, or vectors to vectors. Again, these different
points of view are each useful in different situations. Most often we will
think of a function of the form F : Rn → Rm as taking points to vectors,
though we will feel free to think of such a function in one of the other
ways as needed.
Because we will most often think of a function of the form F : Rn → Rm
as taking points to vectors, we will write such functions with that point
of view in mind.
For example, a multivariable vector-valued function G : R2 → R3
 x+y 
might be given by the formula G x, y 


3x 2 . Then, for example,
  2x+sin y
3
we see that G (3, 0)  27 . Observe that the “input” of this function is
6  x+y 
the point ( x, y ) , and the “output” is the vector 3x 2 .
2x+sin y
We note that if we wanted to write multivariable vector-valued
functions absolutely properly, we would write G (( x, y )) rather than


G x, y , where the inner parentheses in the notation G (( x, y )) refer to
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 6

the point ( x, y ) , and the outer parentheses in G (( x, y )) are the standard

parentheses used in the notation of functions. However, whereas it
would be proper to write G (( x, y )) , we will stick with the more standard


abbreviation G x, y .
In general, we have the following definition.

Multivariable Vector-Valued Functions

1. A multivariable vector-valued function is a function F : Rn → Rm

for some positive integers n and m.

2. A function F : Rn → Rm has the form

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  ,
..
 f ( x ,x. ,...,x ) 
m 1 2 n

where f1 , f2 , . . . , f m : R → R are multivariable real-valued func-

n

tions. The functions f1 , f2 , . . . , f m are called the component

functions of F.

3. A function F : R2 → R3 has the form

" P ( x,y ) #
F x, y  .


Q ( x,y )
R ( x, y )

We note that for a multivariable vector-valued function F : Rn → Rm ,

the two numbers n and m can be either equal or unequal. For example,
we could have functions F : R2 → R2 and G : R3 → R3 . The type of
function where n  m is particularly useful, and will be viewed at times
as taking points to points and at other times as taking points to vectors.
A single variable is a number in R, and multiple variables are points
in Rn for some positive integer n. Observe that we can, in particular, use
the value n  1 when we consider Rn , where R1 is just another name
for R. If we look functions of the form F : Rn → Rm in the special case
when n  1, then we obtain functions of the form F : R → Rm , which
are just single-variable vector-valued functions. Hence single-variable
vector-valued functions are a special case of multivariable vector-valued
functions.
Similarly, if we look functions of the form F : Rn → Rm in the special
case when m  1, then we obtain functions of the form F : Rn → R, which
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 7

are just multivariable real-valued functions. Hence multivariable real-

valued functions are also a special case of multivariable vector-valued
functions.
Clearly, when we use both n  1 and m  1, then functions of the
form functions of the form F : Rn → Rm become simply F : R → R,
which are single-variable real-valued functions. Hence, multivariable
vector-valued functions include all the previous three types of functions
we have seen.
Of course, just because multivariable vector-valued functions include
all the previous three types of functions does not alone make multi-
variable vector-valued functions interesting. The reason we study such
functions is because they in fact arise is many places in mathematics and
its applications, including diverse fields such as physics and economics.
Finally, we need one slight modification of our definition of multivari-
able vector-valued functions.
 Consider the function G : R2 → R3 given
 x+y
by the formula G x, y 


3x 2 . Observe that this formula works for
2x+e y
every possible point ( x, y ) in R2 . On the other hand,
" suppose we wanted
3x y
#
√
to define a function by the formula H x, y 


y−x . Observe that

2x+5y
H x, y is not defined for all ( x, y ) . Rather, it is defined only when
y − x ≥ 0, meaning when y ≥ x. Let A be the region of the x y-plane
that consists of all points ( x, y ) that satisfy y ≥ x; that is, the region A
consists of all points ( x, y ) that are on or above the line y  x. Hence, it
would not be proper to write H : R2 → R3 , because the R2 in this notation
would imply that all of R2 can be used as input. Rather, we would write
H : A → R3 .
In general, multivariable vector-valued functions have the form
F : D → Rm , where D is a subset of Rn . The set D is called the do-
main of the function F. In a course at this level we assume the domain
of a function F : D → Rm is the largest possible subset of Rn that allows
the formula for F to be defined. In more advanced mathematics courses,
the term “domain” is used slightly differently.

Visualizing Multivariable Vector-Valued Functions

As mentioned above, and as seen in Figure 1 and Figure 2 of this section,
we can graph functions of the form y  f ( x ) and z  f ( x, y ) . As seen
in Figure 3 of this section,
" # we can also draw the image of functions
f (t )
of the form s ( t )  g (t ) . The question of visualizing other types of
h (t )
multivariable vector-valued functions is trickier.
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 8

One type of multivariable vector-valued functions that we can

" visual-
# P ( x,y )
ize are functions F : R2 → R3 , which have the form F x, y 


Q ( x,y ) ,
R ( x, y )
where P ( x, y ) , Q ( x, y ) and R ( x, y ) are multivariable real-valued func-
tions P, "Q, R #: R2 → R. Whereas the image of a function of the form
f (t )
s (t )  g (t ) is a curve in R3 , the image of a function of the form
h ("t ) #
P ( x,y )
 Q ( x,y ) is, in general, a surface in R3 , although this surface


F x, y
R ( x,y )
might have self-intersections. The reason the image of such a function is
a surface (except in unusual cases) is that the two independent variables
x and y give the image of the function two degrees of freedom, that is,
two independent directions of motion. (An exception to that would be,


for example, if F x, y is a constant function, in which case its image
would be a single point.)
For example, we see in Figure 4 of this section" the image of the
x+3y
#
function F : R2 → R3 given by the formula F x, y 


2x−y ; the surface
x 2 −y 2
has been drawn slightly transparently, to help with visualization.

Figure 4: Graph of F

One important point to keep in mind regarding the images of such

functions is that, in contrast to the graphs of functions of the form
y  f ( x ) and z  f ( x, y ) , the images of functions F : R2 → R3 might
have self-intersection. For example, see Figure 5 of this section for the
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 9

 xy 
R2 R3 given by the formula G x, y 


image of the function G : → x2 .
y

Figure 5: Graph of G

One thing we can do to help us visualize multivariable vector-valued

functions is to use what we know about the images of single-variable
vector-valued functions. The idea is that for any function F : Rn → Rm ,
we can select any one of the coordinates in Rn , and fix the values of
all the other coordinates, and then think of the resulting function as a
function R → Rm .
For example, consider the function F : R2 → R3 given by the formula
x+3y
" #
F x, y 


2x−y . Let c be a real number. We can then make a new
x 2 −y 2
function F xc : R → R3 by holding the variable y in F x,y to the fixed


t+3c
valued c, obtaining the formula F xc ( t )  F ( t, c )  2t−c
2 2
. For example,
  t −c
t+6
we have F x2 ( t )  2t−2 . We can then plot the image of this new single-
t 2 −4
variable vector-valued function. The image of this function will be a curve
in R3 , and this curve will sit inside the image of the original function
F. Similarly, we can make another new single-variable vector-valued
function F dy : R → R3 by holding the variable x in F x, y to the fixed


 d+3t

valued d, obtaining the formula F dy ( t )  F ( d, t )  2d−t .
d 2 −t 2
These single-variable vector-valued functions obtained by holding all
but one of the variables fixed are called parameter curves.
See Figure 6 of this section for the image of the function F together
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 10

with the images of six parameter curves, which are F x0 , F x1 and F x2 in

red, and F 0y , F 1y and F 2y in blue.

Figure 6: Graph of F with some parameter curves

It might seem uninteresting to draw the parameter curves for the

above function F, because we already saw the actual image of the this
function, which was a surface. However, parameter curves become more
useful for functions that we cannot easily visualize. For example,
" consider x+3y+4z
#
the function H : R3 → R3 given by the formula H x, y, z 


2x−y+z .
x 2 −y 2 −z 2
We cannot visualize the image of this function, because it would be some
sort of folded-over version of R3 sitting inside of R3 . However, to get some
sort of visual idea of the function, we can still graph parameter curves for
this function. For example, we can make a new function H xcd : R → R3


by holding the variables y and z in H x, y, z to the fixed  valued c and
t+3c+4d
d, obtaining the formula H xcd ( t )  H ( t, c, d )  2t−c+d . For example,
 t+11  t 2 −c 2 −d 2
we have H x12 ( t )  2t−1 . Even though we cannot visualize the original
t 2 −5
function directly, we can still plot the image of this parameter curve. See
Figure 7 of this section for the images of a few of the parameter curves
for H, including H x12 . In this figure, the curves in red are obtained by
holding y and z fixed; the curves in blue are obtained by holding z and
z fixed; and the curves in green are obtained by holding x and y fixed.
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 11

-10 0 10

10

-10

10

-10

Figure 7: Graph of some parameter curves of H

Limits of Multivariable Vector-Valued Functions

Limits are the essential idea that make calculus distinct from precalculus,
and limits are present—either explicitly or implicitly—in everything we
do in calculus. For example, the definition of derivatives and partial
derivatives are given directly in terms of limits. However, because
limits are somewhat tricky to work with, various convenient rules for
calculating derivatives have been found, for example the Product Rule
and the Quotient Rule. Of course, these rules are proved using limits, so
whenever these rules are used, limits are implicitly used as well, though
we do not have to deal with the limits explicitly in such situations.
In Calculus I, we saw limits of single-variable real-valued functions,
that is, limits of the form lim f ( x ) . In Calculus II, we saw limits of multi-
x→c
variable real-valued functions, that is, limits of the form lim f ( x, y ) .
( x, y ) → ( a,b )
In our current context, where we are interested in limits of multivariable
vector-valued functions, there is really very little that is new, similarly to
what we saw when we discussed limits of single-variable vector-valued
functions.
For limits of multivariable vector-valued functions, let us start with the
# case of a function G : R → R , which has the form G x, y 
particular 2 3

"P ( x,y )
Q ( x,y ) , where where P, Q, R : R2 → R are the component functions of
R ( x, y )
G. Let ( a, b ) be a point in R2 . We want to find


lim G x, y . It turns
( x, y ) → ( a,b )
out (though it takes some work to prove it rigorously, which we will not
do), that to compute the limit of G, we simply compute the limits of each
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 12

of the component functions separately, and since those are multivariable

real-valued functions, we compute those limits just as we did previously.
That is, we have

 lim P ( x, y ) 
 ( x,y )→( a,b ) 

 lim Q ( x, y ) 
lim G x, y   ( x, y ) → ( a,b )  ,
( x,y ) → ( a,b )
 ( x,ylim R ( x, y ) 
 
) → ( a,b ) 

provided that all three limits lim P ( x, y ) , and lim Q ( x, y )

( x,y ) → ( a,b ) ( x,y ) → ( a,b )
and lim R ( x, y ) exist. If even one of these three limits (of P ( x, y ) ,
( x, y ) → ( a,b )


and Q ( x, y ) and R ( x, y ) ) does not exist, then lim G x, y does not
( x, y ) → ( a,b )
exist.
The above limit can be written without using name of the function G,
by simply writing

 lim P ( x, y ) 
 P ( x, y )   ( x, y )→( a,b ) 
lim Q ( x, y )   ( x, ylim
   Q ( x, y )  ,
( x, y ) → ( a,b )  ) → ( a,b ) 
 R ( x, y )   lim R ( x, y ) 
 ( x, y )→( a,b ) 

again, provided all three limits exist.

Example 1
 3x+2y 
R2 R3 be given by the formula F x, y 


Let F : → x2 y3 . Find

2xe y
lim F x, y , if the limit exists.
( x,y ) → (1,0)
SOLUTION The component functions of F are the three multivariable
real-valued functions P, Q, R : R2 → R of F given by the formulas
P ( x, y )  3x + 2y, and Q ( x, y )  x 2 y 3 and R ( x, y )  2xe y . We
observing that each of these component functions is continuous, and
hence we see that

lim P ( x, y )  lim (3x + 2y )  3 · 1 + 2 · 0  3

( x,y ) → (1,0) ( x, y ) → (1,0)

lim Q ( x, y )  lim x 2 y 3  12 · 03  0
( x, y ) → (1,0) ( x,y ) → (1,0)

lim R ( x, y )  lim 2xe y  2 · 1 · e 0  2.

( x,y ) → (1,0) ( x,y ) → (1,0)
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 13

We deduce that lim F x, y exists, and that
( x, y ) → (1,0)

 lim (3x + 2y ) 
( x, y )→(1,0)  3

 2 3 
lim F x, y   ( x,ylim x y   0 .
( x, y ) → (1,0) ) → ( 1,0 )   
 ( x,ylim

2xe  2
y 
) → (1,0) 

More generally, we have the following definition.

Multivariable Vector-Valued Functions: Limits

Let F : Rn → Rm be a function, and let ( a 1 , a 2 , . . . , a n ) ∈ Rn . Suppose
that F is given by the formula

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  ,
..
 f ( x ,x. ,...,x ) 
m 1 2 n

where f1 , f2 , . . . , f m : Rn → R are the component functions of F.

1. The limit of F ( x1 , x2 , . . ., x n ) as ( x 1 , x2 , . . . , x n ) goes to

( a1 , a2 , . . . , a n ) is given by

lim F ( x1 , x2 , . . ., x n )
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
 lim f1 ( x 1 ,x 2 ,...,x n ) 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 
 lim f2 ( x 1 ,x2 ,...,x n ) 

x 1 ,x2 ,...,x n ) → ( a 1 ,a 2 ,...,a n )
  (  ,
 .
.. 
lim f m ( x 1 ,x2 ,...,x n ) 
 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 

provided that all of the limits of the form

lim f i ( x1 , x 2 , . . . , x n ) exist. If even one of
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
the limits lim f i ( x1 , x 2 , . . . , x n ) does not exist,
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
then lim F ( x1 , x 2 , . . ., x n ) does not exist.
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )

2. The above limit can be written without using name of the function
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 14

F by simply writing

 f1 ( x1 ,x2 ,...,x n ) 
 f2 ( x1 ,x2 ,...,x n ) 
lim ..
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )  . 
 fm ( x1 ,x2 ,...,x n ) 
 lim f1 ( x 1 ,x 2 ,...,x n ) 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 
 lim f2 ( x 1 ,x 2 ,...,x n ) 

x 1 ,x 2 ,...,x n ) → ( a 1 ,a 2 ,...,a n )
  (  ,
 .
.. 
lim f m ( x 1 ,x2 ,...,x n ) 
 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 

again, provided that all of the limits of the form

lim f i ( x1 , x 2 , . . . , x n ) exist.
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )

Example 2
 sin(x2 +y2 ) 
Find lim  x22 +y22  , if the limit exists.
( x, y ) → (0,0)  x 2 −y2 
 x +y 
sin ( x 2 +y 2 )
SOLUTION First, we consider the limit lim 2 2 . Using
( x,y ) → (0,0) x +y
polar coordinates, we know r 2  x 2 + y 2 . Clearly, as ( x, y ) gets closer
and closer to (0, 0) , then r gets closer and closer to 0. Hence, using
l’Hôpital’s Rule, we see that
sin ( x 2 + y 2 ) sin ( r 2 )
lim  lim
( x,y ) → (0,0) x2 + y2 r→0 r2
2r cos ( r 2 )
 lim
r→0 2r
 lim cos ( r 2 )  cos (02 )  1.
r→0
x 2 −y 2
Next, we consider the limit lim 2 2. Let us see what happens
( x,y ) → (0,0) x +y
as ( x, y ) approaches (0, 0) from different directions. First, suppose
that ( x, y ) approaches (0, 0) along the x-axis, which means that we set
y  0 and then take the limit, which yields
x2 − y2 x 2 − 02
lim  lim
( x,0) → (0,0) x2 + y2 ( x,0) → (0,0) x 2 + 02
 lim 1  1.
( x,0) → (0,0)
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 15

Second, suppose that ( x, y ) approaches (0, 0) along the y-axis, which

means that we set x  0 and then take the limit, which yields

x2 − y2 02 − y 2
lim  lim
(0, y ) → (0,0) x2 + y2 (0, y ) → (0,0) 02 + y 2

 lim − 1  −1.
( x,0) → (0,0)

Because we obtained different numbers as ( x, y ) approached (0, 0)

x 2 −y 2
from different directions, we deduce that lim 2 2 does not exist.
( x,y ) → (0,0) x +y
 sin(x2 +y2 ) 
We conclude that lim  x22 +y22  does not exist, because the
( x, y ) → (0,0)  x 2 −y2 
 x +y 
limit of one of the component functions does not exist.

Finally, we note that the notion of continuity works for multivariable

vector-valued functions just as it does for single-variable vector-valued
functions, again by looking at one component at a time. For example,
" # P ( x,y )
consider a function G : R2 → R3 , which has the form G x, y 


Q ( x,y ) .
R ( x,y )


Then the function G x, y is continuous if and only if all three of P ( x, y ) ,
Q ( x, y ) and R ( x, y ) are continuous, where the continuity of all three is
as discussed for multivariable real-valued functions.

Example 3
5x+y 2
 
R2 R3 be given by the formula F x, y 


Let F : → sin ( x y ) . Is F
xe y
continuous?
SOLUTION To figure out if F is continuous, we need to figure
out if each of the component functions P ( x, y )  5x + y 2 , and
Q ( x, y )  sin ( x y ) and R ( x, y )  xe y are continuous. We know that
polynomials in each of x and y are continuous, so each of 5x and y 2 are
continuous, and we we know that sums and products of continuous
functions are continuous, and hence P ( x, y )  5x + y 2 and x y are both
continuous. We know that sin x and e y are continuous, and we know
that compositions of continuous functions are continuous, and hence
Q ( x, y )  sin ( x y ) and R ( x, y )  xe y are continuous. We conclude
that F is continuous.
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 16

SUMMARY

Multivariable Vector-Valued Functions

1. A multivariable vector-valued function is a function F : Rn → Rm

for some positive integers n and m.

2. A function F : Rn → Rm has the form

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  ,
..
 f ( x ,x. ,...,x ) 
m 1 2 n

where f1 , f2 , . . . , f m : R → R are multivariable real-valued func-

n

tions. The functions f1 , f2 , . . . , f m are called the component

functions of F.

3. A function F : R2 → R3 has the form

" P ( x,y ) #
F x, y  .


Q ( x,y )
R ( x, y )

Multivariable Vector-Valued Functions: Limits

Let F : Rn → Rm be a function, and let ( a 1 , a 2 , . . . , a n ) ∈ Rn . Suppose
that F is given by the formula

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  ,
..
 f ( x ,x. ,...,x ) 
m 1 2 n

where f1 , f2 , . . . , f m : Rn → R are the component functions of F.

1. The limit of F ( x1 , x2 , . . ., x n ) as ( x 1 , x2 , . . . , x n ) goes to

MULTIVARIABLE VECTOR-VALUED FUNCTIONS 17

( a1 , a2 , . . . , a n ) is given by

lim F ( x1 , x2 , . . ., x n )
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
 lim f1 ( x 1 ,x 2 ,...,x n ) 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 
 lim f2 ( x 1 ,x2 ,...,x n ) 

x 1 ,x2 ,...,x n ) → ( a 1 ,a 2 ,...,a n )
  (  ,
 .
.. 
lim f m ( x 1 ,x2 ,...,x n ) 
 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 

provided that all of the limits of the form

lim f i ( x1 , x 2 , . . . , x n ) exist. If even one of
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
the limits lim f i ( x1 , x 2 , . . . , x n ) does not exist,
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )
then lim F ( x1 , x 2 , . . ., x n ) does not exist.
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )

2. The above limit can be written without using name of the function
F by simply writing

 f1 ( x1 ,x2 ,...,x n ) 
 f2 ( x1 ,x2 ,...,x n ) 
lim ..
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )  . 
 fm ( x1 ,x2 ,...,x n ) 
 lim f1 ( x 1 ,x 2 ,...,x n ) 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 
 lim f2 ( x 1 ,x 2 ,...,x n ) 

x 1 ,x 2 ,...,x n ) → ( a 1 ,a 2 ,...,a n )
  (  ,
 .
.. 
lim f m ( x 1 ,x2 ,...,x n ) 
 
 ( x1 ,x2 ,...,x n )→( a1 ,a2 ,...,a n ) 

again, provided that all of the limits of the form

lim f i ( x1 , x 2 , . . . , x n ) exist.
( x1 ,x2 ,...,x n ) → ( a1 ,a2 ,...,a n )

EXAMPLES
MULTIVARIABLE VECTOR-VALUED FUNCTIONS 18

Example 4
√ 
9−x 2 −y 2
R2 R2 be given by the formula H x, y 


Let H : → . Find
ln ( x−1)
and sketch the domain of H.
SOLUTION Thep domain of H is the set of all points ( x, y ) in R2
for which both 9 − x 2 − y 2 and ln ( x − 1) are defined. That is, the
domain is the set of all points ( x, y ) in R2 for which both inequalities
9 − x 2 − y 2 ≥ 0 and x − 1 > 0 hold, which is the same as x 2 + y 2 ≤ 32
and x > 1. The solution of x 2 + y 2 ≤ 32 is the interior of the circle of
radius 3 centered at the origin together with the circle, shown below
in blue; the solution of x > 1 is the half-plane that is to the right of
the vertical line x  1, not including the line, shown below in red. The
domain of H is the intersection of these two regions, which is shown
below in purple, where the part of the boundary of the region that is
solid is included in the region, and the part of the boundary that is
dashed is not included in the region.

EXERCISES

Basic Exercises
 MULTIVARIABLE VECTOR-VALUED FUNCTIONS 19

1–3 Find and sketch the domain of each function. 2. Let G : R2 → R3 be given by the formula
" ln( x+y ) #
√
G x, y 


x2 y .
1. Let F : R2
→ R2
be given by the formula sin ( x y )

 √x−y 2 
F x, y  √ .
9−x
DERIVATIVES 20

31.2 Derivatives

In Calculus I, we saw how to take the derivative of single-variable real-

valued functions. For example, if f : R → R is given by the formula
f ( x )  x 2 , then the derivative of this function is given by the formula
f 0 ( x )  2x. If we do not need the formula for the function, but only the
name of the function, then the original function is f , and the derivative
is the function f 0. We also used the Leibniz notation for the derivative
df dy
of this function, which is written or . Whatever notation for
dx dx
the derivative is used, the main point is that we start with a function
f : R → R and then arrive at a new function, the derivative of f , which
is useful in the study of the original function.
Of course, the way we compute f 0 is via the limit definition, which is
given by the formula
f (x + h ) − f (x )
f 0 ( x )  lim ,
h→0 h
for those values of x for which the limit exists. In practice, once we prove
various computational rules such as the Product Rule, the Quotient Rule
and the Chain Rule (proved using the limit definition of derivatives), we
don’t subsequently need to use the rather cumbersome limit definition
to find derivatives in practice.
In Calculus II, we saw how to take the partial derivatives of multivari-
able real-valued functions. For example, if f : R2 → R is given by the
formula f ( x, y )  3x 2 y, then the partial derivatives of this function are
given by the formulas f x ( x, y )  6x y and f y ( x, y )  3x 2 . We also used
the Leibniz notation for the partial derivative of this function, which are
∂f ∂z ∂f ∂z
written ∂x or ∂x , and ∂y or ∂y .
The important point to note here is that in contrast to what we saw
for single-variable real-valued functions, for multivariable real-valued
functions we start with one function, for example f : R2 → R, and we
obtained not one single thing that is the derivative of f , but rather two
partial derivatives, each of which is also a function R2 → R.
Recall that the partial derivatives of a multivariable real-valued
function are also defined by a limit definition, though in practice we
compute them by holding all but one of the variables as if they were
constants, and then using the rules for computing derivatives of single-
variable real-valued functions.
DERIVATIVES 21

We previously saw how to take the derivative of single-variable

vector-valued functions.
f 2 g For example, if r : R → R2 is given by the
formula r ( t )  3t+1
t , then the derivative of this function is given by
the formula r ( t )  3 . We can view the derivative in this case in two
0  2t 
ways, either as a single vector-valued function, or as a vector containing
a number of derivatives of single-variable real-valued functions, namely,
the derivatives of the component functions of r.
We now turn to the derivative of the new type of function we are con-
sidering in this chapter, namely, multivariable vector-valued functions.

The Derivative of Multivariable Vector-Valued Func-

tions
Suppose, for example, that G : R2 → R3 is given by the formula G x, y 


5x+y
 
3x 2 . Rather than attempting to take the derivative via limits, we
2x+e y
now use partial derivatives, taking our inspiration from the way we
found derivatives for both multivariable real-valued functions and single-
variable vector-valued functions. The function G is defined
" by
# three
P ( x,y )
component functions, which we can write as G x, y 


Q ( x,y ) , where
R ( x, y )
P ( x, y )  5x + y and Q ( x, y )  3x 2 and R ( x, y )  2x + e y . Each of these
three functions has two partial derivatives, so that we obtain a total of
six partial derivatives, which are
∂P ∂P
5 and 1
∂x ∂y
∂Q ∂Q
 6x and 0 (1)
∂x ∂y
∂R ∂R
2 and  e y.
∂x ∂y
It would be possible simply to deal with each of these six partial deriva-
tives separately, but it turns out that we can arrange them in a particularly
convenient and useful way, by making " use
# of matrices.
P ( x,y )
Observe that in the notation Q ( x,y ) , the “input” variables x and y
R ( x, y )
in each of P, Q and R are written horizontally, in a row, whereas the
“output” variables P, Q and R themselves are written vertically, in a
column. We will arrange our six partial derivatives by maintaining the
original horizontal and vertical arrangement. That is, for the function
P ( x, y ) , which is in the top column of the output, we will find its two
partial derivatives, which are ∂P ∂x
and ∂P
∂y
, and write them horizontally, in
DERIVATIVES 22

that order. That will be the first row in our matrix. Next, for the function
Q ( x, y ) , which is in the second column of the output, we will find its two
∂Q ∂Q
partial derivatives, which are ∂x and ∂y , and write them horizontally, in
that order, forming the second row of our matrix. Finally, for the function
R ( x, y ) , which is in the third column of the output, we will find its two
partial derivatives, which are ∂R ∂x
and ∂R
∂y
, and write them horizontally, in
that order, forming the third row of our matrix. The matrix of partial
derivatives is then
 ∂P ∂P 
 ∂Q∂x ∂y 
∂Q 

 ∂x
 ∂R  .
∂y 
∂R
 ∂x ∂y 
This matrix is called the derivative of G, and is denoted DG; if we need
to specify the input variables, we would write DG ( x, y ) , which could
also be written as DG ( p ) , where p denotes a point in R2 .

Example 1
5x+y
 
R2 R3 be given by the formula G x, y 


Let G : → 3x 2 . Find the
2x+e y
derivative of G.
SOLUTION Using the partial derivatives we computed in Equation (1)
of this section, we see that
 ∂P ∂P 
 5 1 
∂x ∂y 

 ∂Q ∂Q 

DG x, y   ∂x  6x 0  .
 
∂y 
 ∂R ∂R 
  2 e y 
 ∂x ∂y 


Now let us look at the general case.

Multivariable Vector-Valued Functions:

Derivatives

1. Let F : Rn → Rm be a multivariable vector-valued function. Sup-

pose that F ( x 1 , x2 , . . . , x n ) is given by the formula

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  .
..
 f ( x ,x. ,...,x ) 
m 1 2 n
DERIVATIVES 23

The derivative (also called the Jacobian matrix) of F is the m × n

matrix
 ∂ f1 ∂ f1
···
∂ f1 
 ∂x ∂f
1 ∂x 2
∂ f2
∂x n 
∂ f2 

 ∂x21 ∂x 2
··· ∂x n 
DF ( x 1 , x2 , . . ., x n )   . .. .. ..  .
 (2)
 .. . . . 
 ∂ fm ∂ fm ∂ fm 
 ∂x1 ∂x 2
··· ∂x n 


2. If the point ( x1 , x 2 , . . ., x n ) is abbreviated by p, then the derivative

of F is also denoted DF ( p ) ; it is also written DF p or F 0 ( p ) . When
only the name of the derivative is needed, without listing the
variables, it is written DF.
" P ( x, y ) #
3. If F x, y 


Q ( x, y ) , then
R ( x,y )

 ∂P ∂P 
∂x ∂y 

 ∂Q ∂Q 

DF x, y   ∂x  .
∂y 
 ∂R ∂R
 ∂x ∂y 

Of course, not every multivariable vector-valued function has a

derivative, because the components of such a function are multivariable
real-valued functions, and not every multivariable real-valued function
has partial derivatives. For example, the function F : R2 → R2 given
f 5x+| y| g
by the formula F x, y  3x 2 y does not have all possible partial


derivatives when y  0.
It might seem reasonable to call a multivariable vector-valued function
“differentiable” if all its components have all their partial derivatives,
but we cannot use that term this way, because it has another meaning,
as mentioned briefly in an Optional Section of this chapter, where the
derivative of multivariable vector-valued function via limits is discussed.
Observe that if F : Rn → Rm is a multivariable vector-valued function,
then DF is an m × n matrix, where each entry in this matrix is a multi-
variable real-valued function. For example, for the function G : R2 → R3
discussed in Example 1 of this section, we saw that the derivative is a
3 × 2 matrix.
The obvious question now arises as to why we have arranged the
partial derivatives into a matrix in this way, as opposed to arranging
the partial derivatives in some other way, or simply treating the partial
DERIVATIVES 24

derivatives separately. The reason to arrange the partial derivatives in a

matrix in this particular way is that doing so will allow us to use matrix
multiplication, and other aspects of matrices, to provide some formulas
that are very nice analogs of what we saw in Calculus I for single-variable
real-valued functions. Most immediately, we will see examples of the
use of matrix multiplication in this context in Section 31.3, where we
discuss linear approximations, and in Section 31.4, where we discuss the
Chain Rule.
While the notion of the derivative of a multivariable vector-valued
function in general is new, we observe that for single-variable real-valued,
multivariable real-valued and single-variable vector-valued functions,
our matrix approach is really just another way of writing the familiar
derivatives of those previous types of functions.
First, suppose f : R → R is a single-variable real-valued function.
The derivative of f at a number x is just a single number f 0 ( x ) . On the
other hand, we can think of R as R1 , and so we can think of f : R1 → R1
as a multivariable vector-valued function. Viewed that way, the function
f has a derivative which is a 1 × 1 matrix, namely, the matrix D f ( x ) 
 [ f ( x ) ]. Of course, a 1 × 1 matrix is a silly thing, and we will just
 ∂f 
∂x
take the normal derivatives of single-variable real-valued functions, but
it is important to note that other than the extra square brackets, there is no
difference between studying the function f : R → R as a single-variable
real-valued function and the function f : R1 → R1 as a multivariable
vector-valued function.
Next, suppose f : Rn → R is a multivariable real-valued function.
When we first encountered such functions, we simply looked at each of
the n partial derivatives separately. On the other hand, we can think of
f : Rn → R1 as a multivariable vector-valued function. Viewed that way,
the function f has a derivative which is a 1 × n matrix, which is a row
matrix, namely, the matrix

∂f ∂f ∂f
 
D f ( x1 , x 2 , . . ., x n )  ... .
∂x1 ∂x2 ∂x n

Hence, we see that viewing a f : Rn → R as a multivariable vector-valued

function simply assembles the partial derivatives as a row matrix. In
subsequent sections, we will define and make use of the gradient of such
a function, which is just the transpose of D f ( x1 , x2 , . . ., x n ) , and hence is
a column vector.
Finally, suppose r : R → Rn is a single-variable vector-valued function.
Suppose that this function is written as a column vector via component
 f1 ( t ) 
functions as r ( t )   ...  . Then the derivative of this function is the
 fn ( t ) 
DERIVATIVES 25

 f10 ( t ) 
column vector r 0 ( t )   ...  . Once again, we can think of we can think
 f 0 ( t ) 
n
1 n
of r : R → R as a multivariable vector-valued function. Viewed that
way, the function f has a derivative which is an n × 1 matrix, namely,
 ∂∂tf1
  f 0 ( t ) 
   1 .  . Of course, an n × 1 matrix is the
the matrix Dr ( t )   ...
  0 .. 

 ∂ fn
  fn (t ) 
∂t
same as a column vector, and so we see that viewing a r : R → Rn as
a multivariable vector-valued function just gives us the derivative we
already knew, except for thinking of it as a matrix rather than a column
vector.
All told, if F : Rn → Rm and n , 1 and m , 1, then the derivative
DF that we have currently defined is something genuinely new, and if
either n  1 or m  1, or both, then DF is just a slightly different way of
writing the derivatives we are already familiar with. Hence, what we are
considering at present incorporates as special cases everything we have
seen up till now regarding derivatives.
Finally, we note that for a function of the form F : Rn → Rm , the
columns of the derivative matrix of F are tangent vectors to parameter
curves of the function, as defined in Section 31.1.

Basic Properties of the Derivative

In Calculus I, we saw a number of convenient rules for finding the
derivatives of single-variable real-valued functions, including the Product
Rule, the Quotient Rule and the Chain Rule. Some, but not all, of
these rules have analogs for multivariable vector-valued functions. For
example, because we cannot divide vectors, there is no analog of the
Quotient Rule for multivariable vector-valued functions. We can take
the dot product of multivariable vector-valued functions, just as we did
for single-variable vector-valued functions, and there is a version of
the Product Rule for multivariable vector-valued functions, though it
is a bit more complicated than the version we saw for single-variable
vector-valued functions, and we will not state it. There is a very nice
version of the Chain Rule in our present context, which is discussed in
Section 31.4. For now, we state the analogs of the most basic properties
of derivatives of single-variable real-valued functions.

Basic Rules for Derivatives

Let F , G : Rn → Rm be functions, let c be a real number, and let p be a
point in Rn . Suppose that F and G are differentiable. Then
DERIVATIVES 26

1. D ( F + G )( p )  DF ( p ) + DG ( p ) ;

2. D ( F − G )( p )  DF ( p ) − DG ( p ) ;

3. D ( cF )( p )  cDF ( p ) .

The Jacobian Determinant

Of the many things that can be done with matrices, recall determinants
of square matrices, which are very useful in a variety of situations. In
our present content, observe that while in general the derivative of a
multivariable vector-valued function F : Rn → Rm is not a square matrix
when n , m, it will be a square matrix when n  m, and in that case we
can take the determinant of the derivative.
Suppose we have a multivariable vector-valued function F : Rn → Rn .
The Jacobian determinant of F, is defined to be det DF. If p is a point
in Rn , the Jacobian determinant of F at p is denoted det DF ( p ) .

Example 2
 
3x 2 y
Let F : R2 → R2 be given by the formula F x, y 


5x+y 3
. Find the
Jacobian determinant of F.
 
6x y 3x 2



SOLUTION We compute DF x, y 5 3y 2
, and hence
det DF x, y  6x y · 3y 2 − 3x 2 · 5  18x y 3 − 15x 2 .

We mention that there is some slight confusion regarding the term

“Jacobian” in this context. Suppose F : Rn → Rn is a multivariable vector-
valued function. In some books, the derivative DF is called the “Jacobian
matrix” of F, which sounds very similar to the “Jacobian determinant”
of F. Even worse, some books use the word “Jacobian” without the
associated word “matrix” or “determinant,” and it is then unclear which
of the two uses of the word “Jacobian” is meant. To avoid confusion,
we will not use the word “Jacobian” by itself, and will say “Jacobian
determinant” when we mean that, and will say “derivative” when we
mean that.
There is also some confusion regardingf theg notation for the Jacobian
determinant. First, rather than writing det ac db , some people use the old
DERIVATIVES 27

f P ( x,y ) g
a b . Specifically, suppose we have F x, y 


fashioned notation c d Q ( x,y ) .
∂P ∂P
" #
∂x ∂y
Then DF x, y 


∂Q ∂Q , and whereas we would then write the Jacobian
∂x ∂y
∂P ∂P
" #
∂x ∂y
determinant of F as det DF x, y  det


∂Q ∂Q , the people who use
∂x ∂y
the alternative notation for determinants would write the Jacobian
∂P ∂P ∂P ∂P
∂x ∂y ∂x ∂y
determinant of F as ∂Q ∂Q . Because writing ∂Q ∂Q by hand can look
∂x ∂y ∂x ∂y
∂P ∂P ∂P ∂P
" # " #
∂x ∂y ∂x ∂y
very similar to ∂Q ∂Q , we recommend using the notation det ∂Q ∂Q ,
∂x ∂y ∂x ∂y
which is unambiguous. f P ( x,y ) g
Additionally, rather than writing F as Q ( x,y ) , where x and y are
the independent variables, it is sometimes convenient to write x and y
as the dependent variables, and u and v as the independent variables,
and one then writes xf  g ( u,g v ) and y  h ( u, v ) , which we would write
g ( u,v )
as T ( u, v )  xy  h ( u,v ) . The Jacobian determinant of T is then
 
 ∂g ∂g 
∂u ∂v
det ∂h ∂h . Sometimes, however, functions are not given names, so that
∂u ∂v
rather than writing x  g ( u, v ) , we simply think of x as a function of
u and v, which we could write x  x ( u, v ) , and similarly for y. The
Jacobian determinant of x and y as functions of u and v would then be
written, without function names and with the old fashioned notation for
∂x ∂x
∂u ∂v
the determinant, as ∂y ∂y . Moreover, instead of writing the Jacobian
∂u ∂v
determinant in this situation in either of the above notations, some people
∂ ( x,y )
use the old-fashioned notation ∂ ( u,v ) , which does not even need the name
of the function. Similar notation is used with more variables.
We summarize these ideas as follows.

The Jacobian Determinant

1. Let F : Rn → Rn be a function. The determinant of the derivative

of F is called the Jacobian determinant (or just the Jacobian) of
the function, and is denoted

det DF ( x1 , x2 , . . ., x n ) ,

or similarly if a different notation for the derivative is used.

 xx12 
2. If the vector  ..  is abbreviated by p, then the Jacobian determinant
 x. 
n
of F is also denoted det DF ( p ) . When only the name of the
DERIVATIVES 28

Jacobian determinant is needed, without listing the variables, it

is written det DF.
f P ( u,v ) g
3. If F ( u, v )  Q ( u,v ) , then the Jacobian determinant of F is some-
∂x ∂x
∂u ∂v ∂ ( x,y )
times denoted ∂y ∂y or ∂ ( u,v )
.
∂u ∂v

Geometric Meaning of the Jacobian Determinant

2 2 given by the formula H x, y 


f ax+bHy :g R → R be the linear map
Let
cx+dy . The two basis vectors e1  0 and e2  1 in R form a square,
1 0 2

as seen in Figure 1 of this section. We can then look at the parallelogram

formed by the two vectors H ( e1 ) and H (fe2 ) , which
g is seen in Figure 2 for
2x+y
the specific example where H x, y  x+y ; this new parallelogram


is the image result of applying the function H to the original square.
In general, we ask what is the relation between the area of the original
square, which is 1, and the area of the new parallelogram.

Figure 1: Square formed by the vectors e1 and e2

Figure 2: Parallelogram formed by the vectors F ( e1 ) and F ( e2 )

f g
Observe that H ( e1 )  [ ac ] and H ( e2 )  db . As we saw in our
discussion of the cross product, the area of the area off thegparallelogram
formed by the two vectors H ( e1 ) and H ( e2 ) is | det ac db |. Hence, we
see that applying the linear map H to the original square results in
DERIVATIVES 29

f g f g
multiplying the area by | det ac db |. We note that the matrix ac db should
look familiar in this context, because it is precisely DH.
More generally, let F : Rn → Rn be a function, and let p be a point in
Rn . Of course, the function F need not be a linear map. However, as we
will see in more detail in Section 31.3, the matrix DF ( p ) is the matrix of
the linear map that best approximates F at the point p. Then, analogously
to what we saw above for the linear map H, it turns out more generally
that if we were to take a small region A in Rn that is near p, then the
area of F ( A ) is approximately equal to DF ( p ) times the area of A. The
smaller the region, and the closer to p, the better the approximation. The
proof of this fact is not simple, and it requires certain hypotheses on the
function F, but the inutive idea is simply that we use the matrix DF ( p ) ,
and that matrix multiplication is a linear map.
This geometric way of thinking about DF ( p ) will be used when we
discuss the change of variable formula for double and triple integrals.

SUMMARY

Multivariable Vector-Valued Functions:

Derivatives

1. Let F : Rn → Rm be a multivariable vector-valued function. Sup-

pose that F ( x 1 , x2 , . . . , x n ) is given by the formula

 f1 ( x1 ,x2 ,...,x n ) 
F ( x1 , x2 , . . ., x n )  
 f2 ( x1 ,x2 ,...,x n )  .
..
 f ( x ,x. ,...,x ) 
m 1 2 n

The derivative (also called the Jacobian matrix) of F is the m × n

matrix
 ∂ f1 ∂ f1
···
∂ f1 
 ∂x ∂f
1 ∂x 2
∂ f2
∂x n 
∂ f2 

 ∂x21 ∂x 2
··· ∂x n 
DF ( x 1 , x2 , . . ., x n )   . .. .. ..  .

 .. . . . 
 ∂ fm ∂ fm ∂ fm 
 ∂x1 ∂x 2
··· ∂x n 

DERIVATIVES 30

2. If the point ( x1 , x 2 , . . ., x n ) is abbreviated by p, then the derivative

of F is also denoted DF ( p ) ; it is also written DF p or F 0 ( p ) . When
only the name of the derivative is needed, without listing the
variables, it is written DF.
" P ( x, y ) #
3. If F x, y 


Q ( x, y ) , then
R ( x,y )

 ∂P ∂P 
∂x ∂y 

 ∂Q ∂Q 

DF x, y   ∂x  .
∂y 
 ∂R ∂R
 ∂x ∂y 

Basic Rules for Derivatives

Let F , G : Rn→ Rm be functions, let c be a real number, and let p be a
point in Rn . Suppose that F and G are differentiable. Then

1. D ( F + G )( p )  DF ( p ) + DG ( p ) ;

2. D ( F − G )( p )  DF ( p ) − DG ( p ) ;

3. D ( cF )( p )  cDF ( p ) .

The Jacobian Determinant

1. Let F : Rn → Rn be a function. The determinant of the derivative

of F is called the Jacobian determinant (or just the Jacobian) of
the function, and is denoted

det DF ( x1 , x2 , . . ., x n ) ,

or similarly if a different notation for the derivative is used.

 xx12 
2. If the vector  ..  is abbreviated by p, then the Jacobian determinant
 x. 
n
of F is also denoted det DF ( p ) . When only the name of the
Jacobian determinant is needed, without listing the variables, it
is written det DF.
DERIVATIVES 31

f P ( u,v ) g
3. If F ( u, v )  Q ( u,v ) , then the Jacobian determinant of F is some-
∂x ∂x
∂u ∂v ∂ ( x,y )
times denoted ∂y ∂y or ∂ ( u,v )
.
∂u ∂v

EXAMPLES

Example 3
" 3x+2y−z #
R3 R3 be given by the formula H x, y, z 


Let H : → x 2 +y 3 . Find
4xz 2
the Jacobian determinant of H.
 3 2 −1 



SOLUTION We compute DH x, y, z 2x 3y 2 0 . Hence, by
4z 2 0 8xz
expanding along the bottom row of this matrix, we deduce that
det DH x, y, z  4z 2 · (0 − (−3y 2 )) − 0 · (0 − (−2x )) + 8xz · (9y 2 − 4x ) 


12y 2 z 2 + 72x y 2 z − 32x 2 z.

Example 4
 
6x+y 2
R2 R2 be given by the formula F x, y 


Let F : → 3x 2 +2y
. Find and
plot all points ( x, y ) in R2 for which det DF x, y  0.


f g
6 2y
We compute DF x, y  6x . Hence det DF x, y 



SOLUTION 2

6·2−6x·2y  12−12x y. Therefore det DF x, y  0 yields 12−12x y  0,
which is the same as x y  1, which in turn is the equivalent to y  x1 .
Hence, plotting all points for which det DF x, y  0 is the same as


drawing the graph of y  x1 , which is shown below.
DERIVATIVES 32

COMMON MISTAKES

The derivative of a multivariable vector-valued function of the form

F : Rn → Rm is an m × n matrix. A very common mistake when finding
the derivative of such a function is to make an n × m matrix rather than
an m × n matrix. Additionally, even if an m × n matrix is correctly used,
a common mistake is to arrange the partial derivatives in the wrong
order within the matrix. The only correct way to form the derivative of a
multivariable vector-valued function is by placing the partial derivatives
as stated in the text, namely, by having each row in the matrix correspond
to a component function (which yields m rows), and having the partial
derivatives in each row correspond to the variables in F in the given
order (which yields n columns).

EXERCISES

Basic Exercises
 DERIVATIVES 33

1–3 Find the derivative of each function. 4. Let F : R2 → R2 be given by the formula
x2 y
F x, y 


3x−y 2
.
1. Let F : R3 R2
→  be given  by the formula
3x 2 +yz
F x, y, z  xz 3 +2y 5 . 5. Let G : R3 → R3 be given by the formula


 3x+2y+z 
G x, y, z 


e 5x+y .
yz
2. Let G : R2 → R4 be given by the formula
 e x +sin y 
 ln( x y ) 
G x, y   √x  .

6. Let H : R4 → R4 be given by the formula
 x 3 y   4x−3y+2z−w 
H x, y, z, w   3y+zw

 .
4zw
w3
 
3. Let H : R4 → R3 be
 given by the formula
x+2y+3z+4w
H x, y, z, w 


xy .
6zw
7. Let K : R2 →f R2 be given
g by the formula
2 2
K x, y  x −12x−y


xy . Find and plot all points
( x, y ) in R for which det DK x, y  0.
2

4–6 Find the Jacobian determinant of each function.
LINEAR APPROXIMATIONS & DIFFERENTIALS 34

31.3 Linear Approximations & Differentials

In Calculus I, we saw the formula for the tangent line for a single-variable
real-valued function y  f ( x ) at a point x  a, which is

L ( x )  f ( a ) + f 0 ( a )( x − a ) . (1)

This formula was used to find approximate values for functions that are
difficult to compute. Similarly, in Calculus II, we saw the formula for the
tangent plane for a multivariable real-valued function z  f ( x, y ) at a
point ( x, y )  ( a, b ) , which is

L ( x, y )  f ( a, b ) + f x ( a, b )( x − a ) + f y ( a, b )( y − b ) . (2)

We now want the analog for multivariable vector-valued functions

of the above formulas for the tangent line and the tangent plane. In
contrast to tangent lines and tangent planes, which are easy to visualize,
in the case of multivariable vector-valued functions there is no nice
visualization of the analog of tangent lines and tangent planes. However,
while we have no nice visualization, it is nonetheless possible to find a
formula that is analogous to the formulas in Equation (1) and Equation (2)
of this section, and that can similarly be used to approximate the original
function.

The Linear Approximation of a Multivariable Vector-

Valued Function
In the following discussion, we note that a point Rn can also be thought of
as corresponding to a vector from the origin to the point. For example, if
p  ( a, b ) is a point in R2 , we can also think of this point as corresponding
to the vector [ ba ]. In principle, the name of this vector should not be p,
because points and vectors are theoretically different things, but because
the point ( a, b ) gives rise to the vector [ ba ], and the other way around, we
will avoid cumbersome notation and use the notation p for both ( a, b )
and [ ba ]. Whether p means the point or the vector can be understood
from the context.
We now return to our question of the linear approximation of a
Multivariable Vector-Valued Function. The key to finding analogs in
our current context of Equation (1) and Equation (2) of this section is to
LINEAR APPROXIMATIONS & DIFFERENTIALS 35

observe in Equation (2) that the expression f x ( a, b )( x − a ) + f y ( a, b )( y − b )

can be thought off as ag product of the row matrix [ f x ( a,b ) f y ( a,b ) ] and the
x−a
column matrix y−b . If we write p  ( a, b ) , then the row matrix
[ f x ( a,b ) ff y ( a,bg) ] is just D f ( p ) . If we write v  ( x, y ) , then the column
x−a
vector y−b is just v − p, where we think of v and p here as vectors. We
can then rewrite Equation (2) as L ( v )  f ( p ) + D f ( p )( v − p ) . This last
formula looks completely analogous to Equation (1), and, even better, it
holds for all multivariable vector-valued functions, as we now state.

Linear Approximations for Multivariable

Vector-Valued Functions
Let F : Rn → Rm be a function, and let p be a vector in Rn . The linear
approximation of F at p is the function L : Rn → Rm given by the
formula
L ( v )  F ( p ) + DF ( p )( v − p ) . (3)

Example 1
 √ 
y+ x
R2 R2 be given by the formula F x, y 


Let F : → √
x+ 3 y
. Find the
linear approximation of F at p  (9, 8) .
1
1
" √
#
2 x
We compute DF x, y  . Then F (9, 8) 

 11 
SOLUTION 1 1
11
2
2y 3
1 
1
and DF (9, 8)  6
1 1 , and we deduce that
12

" # " #!
x 9
L x, y  F (9, 8) + DF (9, 8)


−
y 8
1 1
x + y + 23
" # " #" # " #
11 1 x−9
 + 6
1  6
1 .
11 1 12 y−8 x + 12 y + 43

The value of the linear approximation is, as its name suggests, in

calculating the approximate value of a function that is hard to calcu-
late directly. Suppose we have a multivariable vector-valued function
F : Rn → Rm , and we wish to find an approximate value for F ( v ) for some
point v in Rn , due to the fact that F ( v ) is hard to calculate directly. As
was the case with linear approximations for single-variable real-valued
LINEAR APPROXIMATIONS & DIFFERENTIALS 36

and multivariable real-valued functions, here too our ability to make use
of Equation (3) to approximate F ( v ) depends upon finding a convenient
point p such that on the one hand p is very close to v (the closer p is to v
the better the approximation will be), and such that on the other hand it
is easy to compute both F ( p ) and DF ( p ) . Of course, in practice, it is not
always possible to find such a point p, but it is possible in enough cases
that this very easy to use method of approximation is of value.

Example 2
 √ 
y+ x
R2 R2 be given by the formula F x, y 


Let F : → √
x+ 3 y
. Use the
linear approximation to compute an approximate value for F (9.1, 7.8) .
SOLUTION We saw in Example 1 of this section that the linear ap-
proximation of F at p  (9, 8) is
1
" # " #" #
11 1 x−9
L x, y  + 6 .


1
11 1 12 y−8

Then
1
" # " #" # " #
11 1 9.1 − 9 1 649
F (9.1, 7.8) ≈ L (9.1, 7.8)  + 6 1  .
11 1 12 7.8 − 8 60 665

The Differential of a Multivariable Vector-Valued Func-

tion
In Calculus I, we saw an alternative way of formulating the idea of linear
approximation, which was via differentials. Consider first a single-
variable real-valued function y  f ( x ) at a point x  a. Suppose we are
able to compute f ( a ) , but we then want to change the value of x by ∆x,
where ∆x is used to denote some small change in the value of x. That
is, we want to find f ( a + ∆x ) . The question then arise, if we change
the value of x by ∆x, how much will the value of y change. We let this
change in y be denoted ∆y, so that we have ∆y  f ( a + ∆x ) − f ( a ) . If
∆x is close to zero, we saw that

∆y  f ( a + ∆x ) − f ( a )
≈ L ( a + ∆x ) − f ( a )  f ( a ) + f 0 ( a )(( a + ∆x ) − a ) − f ( a ) (4)
 f ( a ) ∆x.
0
LINEAR APPROXIMATIONS & DIFFERENTIALS 37

We now rewrite Equation (4) in a different notation. We let dx  ∆x, and

we let
dy  f 0 ( a ) dx.
With this new notation, we can rewrite Equation (4) as ∆y ≈ dy. Hence,
if we want to find an approximate value for ∆y, we can compute dy,
which is often much easier to compute.
Similarly, in Calculus II, we saw the same idea in the context of
multivariable real-valued functions, where we want to approximate the
function z  f ( x, y ) at the point ( x, y )  ( a, b ) . In this case, we change
the value of x by ∆x, and the value of y by ∆y, and we want to compute
the change in z, which is defined to be ∆z  f ( a + ∆x, b + ∆y ) − f ( a, b ) .
As before, if ∆x and ∆y are close to zero, then it is seen that ∆z ≈
f x ( a, b ) ∆x + f y ( a, b ) ∆y. In this case, we let dx  ∆x and dy  ∆y, and
we let
∂z ∂z
dz  f x ( a, b ) dx + f y ( a, b ) dy  dx + dy,
∂x ∂y
which is the same as dz  L ( a + ∆x, b + ∆y ) − L ( a, b ) , where we observe
that L ( a, b )  f ( a, b ) . It is then seen that ∆z ≈ dz.
We now want the analog for multivariable vector-valued functions of
the above formulas for differentials. Let F : Rn → Rm be a function, and
let p be a point in Rn . We want to approximate the function z  F ( v )
at the point v  p in Rn . In this case, we change the value of v by
∆v, and we want to compute the change in z, which is defined to be
∆z  F ( p + ∆v ) − F ( p ) . As before, if ∆v is close to the origin, then we
use Equation (3) of this section to obtain

∆z  F ( p + ∆v ) − F ( p )
≈ L ( p + ∆v ) − F ( p )
(5)
 F ( p ) + DF ( p )(( p + ∆v ) − p ) − F ( p )
 DF ( p ) ∆v.

We let dv  ∆v, and we let

dz  DF ( p ) dv, (6)

which, by Equation (5), is the same as dz  L ( p + ∆v ) − L ( p ) , where we

observe that L ( p )  F ( p ) . Once again, it is then seen that ∆z ≈ dz.
When working with multivariable vector-valued functions, it is com-
mon not to assign a single letter, such as “z,” to denote the value of F ( v ) .
As such, we can rewrite ∆z  F ( p+∆v ) −F ( p ) as ∆F ( p )  F ( p+∆v ) −F ( p ) ,
and we can rewrite Equation (6) as

dF ( p )  DF ( p ) dv.
LINEAR APPROXIMATIONS & DIFFERENTIALS 38

If ∆v is close to the origin, then ∆F ( p ) ≈ dF ( p ) . Hence, if we want to

find an approximate value for ∆F ( p ) , we can compute dF ( p ) , which is
often much easier to compute.
The above ideas can be summarized as follows.

Differentials for Multivariable Vector-Valued

Functions
Let F : Rn → Rm be a function, and let p be a point in Rn . Let z  F ( v ) .

1. Let ∆v be a point in Rn that is close to the origin. The change in F

at p, denoted ∆z or ∆F ( p ) , is defined by

∆z  ∆F ( p )  F ( p + ∆v ) − F ( p )

2. Let dv  ∆v. The differential of F at p, denoted dz or dF ( p ) , is

defined by
dz  dF ( p )  DF ( p ) dv.

3. If ∆v is close to the origin, then dF ( p ) can be used as an approxi-

mation to ∆F ( p )

Example 3
 √ 
y+ x
Let F : R2 → R2 be given by the formula F x, y 


√
x+ 3 y
. Find dF ( p ) .

We saw in Example 1 of this section that DF x, y 

SOLUTION
1
1
" √
#
2 x f
dx
g
1 1
2
. If we write dv  dy , then
2y 3

 √1 1  " dx #  2√1 x dx + dy 
2 x
dz  dF ( p )  DF x, y dv   1   dx + 1 dy  .


1 
2  dy
   2 
2y 3   2y 3 

Is important to note that using differentials to approximate certain

values is not really any different from using the linear approximation;
it is just a different way of writing it, and it emphasizes a different way
of thinking about it: with the linear approximation, we think of the
tangent line as being close to the original function at the point of tangency,
LINEAR APPROXIMATIONS & DIFFERENTIALS 39

whereas with differentials, we think of approximating a change in the

value of the “output” variable when we make a small change in the value
of the “input” variable.

SUMMARY

Linear Approximations for Multivariable

Vector-Valued Functions
Let F : Rn → Rm be a function, and let p be a vector in Rn . The linear
approximation of F at p is the function L : Rn → Rm given by the
formula
L ( v )  F ( p ) + DF ( p )( v − p ) .

Differentials for Multivariable Vector-Valued

Functions
Let F : Rn → Rm be a function, and let p be a point in Rn . Let z  F ( v ) .

1. Let ∆v be a point in Rn that is close to the origin. The change in F

at p, denoted ∆z or ∆F ( p ) , is defined by

∆z  ∆F ( p )  F ( p + ∆v ) − F ( p )

2. Let dv  ∆v. The differential of F at p, denoted dz or dF ( p ) , is

defined by
dz  dF ( p )  DF ( p ) dv.

3. If ∆v is close to the origin, then dF ( p ) can be used as an approxi-

mation to ∆F ( p )
LINEAR APPROXIMATIONS & DIFFERENTIALS 40

EXAMPLES

Example 4
Let F : R2 → R2 be a function. Suppose that F (1, 2) 
5
7 and that
DF (1, 2)  34 15 . Estimate the value of F (1.02, 1.97) .
 

SOLUTION Using Equation (3) of this section, we see that

" # " #!
1.02 1
F (1.02, 1.97) ≈ L (1.02, 1.97)  F (1, 2) + DF (1, 2) −
1.97 2
" # " #" # " #
5 3 1 0.02 5.03
 +  .
7 4 5 −0.03 6.93

COMMON MISTAKES

There is a question regarding the exact meaning of the symbols dx,

dy, dz, etc. Some people think of these symbols as denoting an “infinites-
imally small” changes in these variables, whatever infinitesimally small
might mean. From a historical perspective, the idea of infinitesimally
small changes in variables was indeed used in the early period of the
development of Calculus, and for some people that approach is still
intuitively appealing. However, it is important to note that in the modern
approach to calculus (now over 150 years old), there are no such things
as infinitesimally small real numbers (as can be seen in a course on real
analysis). Hence, it would be a mistake to think that dx, dy, dz, etc., are
actually infinitesimally small numbers, even though thinking about them
that way might be convenient or intuitively appealing. Unfortunately, it
takes more advanced mathematics than is available to us here to give
a rigorous definition of what such things really are, though there is
indeed such a definition. At the same time, there now exists a completely
rigorous way to define infinitesimally small numbers, not as part of the
real numbers (which is not possible) but as some other kind of number
 LINEAR APPROXIMATIONS & DIFFERENTIALS 41

that can be added on to the real numbers; the study of that approach,
which would take us very far afield, is called non-standard analysis.

EXERCISES

Basic Exercises

1–2 For each function, use the linear approximation 2. Let G : R3 → R2 be given by the formula
x+2y+3z
f g
to compute the requested approximation. G x, y z  e √3 x yz . Compute an approximate

1. Let F : R2 → R2 be value for G (1.1, 0.9, −1.1) .

 given by the formula
√
xy
F x, y 


y . Compute an approximate value
x
for F (1.9, 18.2) .
THE CHAIN RULE 42

31.4 The Chain Rule

In Calculus I, we saw a number of methods for finding the derivatives of

single-variable real-valued functions, for example the Product Rule, the
Quotient Rule and the Chain Rule. Whereas the Product Rule does not
have a simple analog for multivariable vector-valued functions, and the
Quotient Rule has no analog at all for such functions, it turns out that
the Chain Rule has a very nice analog for multivariable vector-valued
functions, as we will now see. There are two (equivalent) ways to state the
Chain Rule for multivariable vector-valued functions, one approach using
matrices and matrix multiplication, and the other not using matrices. We
will start with the matrix formulation, because the formula for the Chain
Rule for multivariable vector-valued functions in the matrix approach
looks just like the Chain Rule formula from Calculus I, and because it
shows the power of matrices; the matrix approach to the Chain Rule is
the “right way to do it.” We will then show the non-matrix formulation,
which, while not as aesthetically appealing as the matrix approach, is
useful in practice.
The version of the Chain Rule from Calculus I is used to find the
derivatives of functions such as h : R → R given by the formula h ( x ) 
sin ( x 2 + 7) . The idea is to decompose this type of function into the
composition of two simpler functions, where one of the simpler functions
is “inside” the other function. For example, for the function h ( x ) 
sin ( x 2 + 7) , we define new functions f : R → R and g : R → R given
by the formulas f ( x )  sin x and g ( x )  x 2 + 7, and we then see that
h ( x )  f ( g ( x )) . Therefore, in order to find the derivative of h ( x ) , we
need to find the derivative of f ( g ( x )) , and the latter type of expression
is precisely the subject of the Chain Rule, which says that

[ f ( g ( x )) ]0  f 0 ( g ( x )) g 0 ( x ) .

For example, we see that [sin ( x 2 + 7) ]0  cos ( x 2 + 7) · 2x.

We now want to rewrite the formula for the Chain Rule for single-
variable real-valued functions using a slightly different notation, which
you might or might not have seen previously. Whether or not you have
already seen this notation, it is very important to become familiar with
this notation, because it is used extensively in advanced mathematics.
THE CHAIN RULE 43

Composition of Functions
Rather than starting with a single function, for example h : R → R
given by the formula h ( x )  sin ( x 2 + 7) , and then decomposing it by
writing h ( x )  f ( g ( x )) , where f : R → R and g : R → R are given by the
formulas f ( x )  sin x and g ( x )  x 2 + 7, we now want to start with two
functions such as f and g and then combine them.
For example, let f : R → R and g : R → R be given by the formulas
f ( x )  x 2 and g ( x )  x + 3. We then want to construct a new function,
which will be given by the formula f ( g ( x ))  ( x + 3) 2 . It would be very
convenient to have a name for this combined function. We could call this
function by a new letter, for example writing k : R → R as k ( x )  ( x + 3) 2 ,
but calling the function that results from combining f and g by an
arbitrary name such as k is confusing. It would be much better to give
the new function a name that reflects that it is made up of f and g, and
that is what we now state.
Let f : R → R and g : R → R be functions. The composition of f and
g is the function f ◦ g : R → R given by the formula

( f ◦ g )( x )  f ( g ( x )) . (1)

For example, let f : R → R and g : R → R be given by the formulas

f ( x )  x 2 and g ( x )  x + 3. Then ( f ◦ g )( x )  f ( g ( x ))  ( x + 3) 2 , and
( g ◦ f )( x )  g ( f ( x ))  x 2 + 3. Observe that f ◦ g is not the same as g ◦ f .
The reader who is encountering the notation f ◦ g for the first time
might find it necessary to get used to the fact that it is “backwards” from
what might be expected, because f ◦ g means doing g first and then f
even though we generally read from left to right in English. Think of
“◦” as meaning “following.” We will stick with the “◦” notation in spite
of any slight confusion it might cause at first, because it is extremely
widespread, and because the reader will find that it works well once she
is used to it.
It is important to look carefully at the definition of composition. The
notation “ f ◦ g” is the name of one single-variable real-valued function,
which we constructed out of the two functions f and g. In Equation (1),
we say how to compute this function, which means, we say what this
function does to each real number x. Specifically, this equation says
that the value of the function f ◦ g at the number x, which is denoted
( f ◦ g )( x ) , is computed by doing f ( g ( x )) .
Observe that the parentheses in the expression “ ( f ◦ g )( x ) ” need to
be used exactly as written. It would not be correct to write “ f ◦ g ( x ) ,”
because ◦ is an operation that combines two functions, whereas “ f ( x ) ”
is not a function but a single number.
THE CHAIN RULE 44

Composition of functions works for multivariable vector-valued

functions just as it does for single-variable real-valued functions, with
one caveat, which is that the “output” of the first function has to be the
“input” of the second function in order to form a composition of two
functions.
For example, if F : R3 → R4 and G : R2 → R3 are functions then we
can form the composition F ◦ G : R2 → R4 , because the “output” of G
and the “input” of F are both R3 , but we cannot form the composition
G ◦ F, because the “output” of F is not the same as the “input” of G.
More generally, we have the following definition.

Composition of Functions
Let G : Rn → Rk and F : Rk → Rm be functions. The composition of F
and G is the function F ◦ G : Rn → Rm given by the formula

( F ◦ G )( p )  F ( G ( p )) . (2)

Example 1
let F : R3 → R4 and G : R2 → R3 be given by the formulas
" 5x+y #  xy 
F x, y, z  3z G x, y  .



2xz
and x+2y
y−z x−y

Find the formula for ( F ◦ G ) x, y .
SOLUTION The function F ◦ G : R2 → R4 is given by the formula

 5( x y )+( x+2y )   x+2y+5x y 

3 ( x−y ) 3x−2y
( F ◦ G ) x, y   2( x y )( x−y )    2x 2 y−2x y 2  .


 ( x+2y )−( x−y )   3y 

Now that we have the notion of composition of functions, we can

restate the the Chain Rule for single-variable real-valued functions as

[ ( f ◦ g )( x ) ]0  f 0 ( g ( x )) g 0 ( x ) . (3)

For example, suppose we want to find the derivative of the function

p : R → R be given by the formula p ( x )  cos ( x 2 ) . We then define
THE CHAIN RULE 45

functions h : R → R and k : R → R by the formulas h ( x )  cos x and

k ( x )  x 2 , and we observe that p  h ◦ k. We then use the Chain Rule to
compute

[cos ( x 2 ) ]0  p 0 ( x )  [ ( h ◦ k )( x ) ]0  h 0 ( k ( x )) k 0 ( x )  − sin ( x 2 ) · 2x.

Of course, in practice, when someone wants to compute [cos ( x 2 ) ]0, she

would not given names to the functions h and k, and would simply write
[cos ( x 2 ) ]0  − sin ( x 2 ) · 2x, but we wanted to emphasize that the Chain
Rule is really about the composition of functions.
We now turn to the analog of Equation (3) for multivariable vector-
valued functions.

The Chain Rule via Matrix Multiplication

Let G : Rn → Rk and F : Rk → Rm be functions. We want to find a
formula for D ( F ◦ G )( p ) that is analogous to Equation (3) of this section.
If we examine the right-hand side of Equation (3), we see that the
reason we can multiply f 0 ( g ( x )) and g 0 ( x ) is that for each real number
x, the quantities f 0 ( g ( x )) and g 0 ( x ) are also real numbers, and we know
that real numbers can be multiplied. The analogs for multivariable
vector-valued functions of f 0 ( g ( x )) and g 0 ( x ) would be DF ( G ( p )) and
DG ( p ) . What are these two quantities? They are matrices, and, more
specifically, we know that DF ( G ( p )) is an m × k matrix and DG ( p ) is a
k × n matrix. Recalling the definition of matrix multiplication, we see
from the dimensions of the matrices DF ( G ( p )) and DG ( p ) that they
can in fact be multiplied. Indeed, not only can these two matrices be
multiplied, but taking their product is exactly what we need for the
Chain Rule for multivariable vector-valued functions, which we now
state.

The Chain Rule via Matrix Multiplication

Let G : Rn → Rk and F : Rk → Rm be functions, and let p be a point in
Rn . Suppose that F and G are differentiable. Then

D ( F ◦ G )( p )  DF ( G ( p )) DG ( p ) , (4)

where the multiplication is matrix multiplication.

We will not prove Equation (4), but let us look at an example, to see
that it really works.
THE CHAIN RULE 46

Example 2
Let F : R3 → R2 and G : R2 → R3 be given by the formulas
f 2x y g  x−4y 
F x, y, z  G x, y  .



y−z and xy
x+y




Find each of DF ( G x, y ) and DG x, y and D ( F ◦ G ) x, y directly,
and verify the The Chain Rule via Matrix Multiplication for this
example.
SOLUTION The function F ◦ G : R2 → R2 is given by the formula

f 2( x−4y )( x y ) g f 2 y−8x y 2 g
( F ◦ G ) x, y  ( x y )−( x+y )  2xx y−x−y .

Next, we compute the derivatives of our three functions, which are

f g
2y 2x 0
DF x, y, z 


 10 −41 −1
DG x, y 


y x
1 1
f 4x y−8y 2 2x 2 −16x y g
D ( F ◦ G ) x, y  .


y−1 x−1

We then have
f g f g
2 ( x y ) 2 ( x−4y ) 0 2x y 2x−8y 0
DF ( G x, y )   .


0 1 −1 0 1 −1

Finally, we compute
f g1
−4

2x y 2x−8y 0
DF ( G x, y ) DG x, y 



y x
0 1 −1 1 1
f 2x y·1+(2x−8y ) ·y+0·1 2x y· (−4) + (2x−8y ) ·x+0·1
g
 0·1+1·y−1·1 0· (−4) +1·x−1·1
f 4x y−8y 2 2x 2 −16x y
g
 y−1 x−1
 D ( F ◦ G ) x, y .

Hence, we see that the The Chain Rule via Matrix Multiplication
works for this example.

The Chain Rule without Matrices

Whereas the nicest-looking, and most general, formulation of the Chain
Rule for multivariable vector-valued functions is via matrix multiplication,
THE CHAIN RULE 47

it is also possible to state and use specific cases of this Chain Rule without
matrices. This non-matrix formulation of the Chain Rule will be written
entirely using Leibniz notation for derivatives. To start, let us recall the
statement of the Chain Rule for single-variable real-valued functions in
Leibniz Notation, which is as follows.
Let f : R → R and g : R → R be functions. We write y  f ( u ) and
u  g ( x ) . Suppose that f and g are differentiable. Then
dy dy du
 . (5)
dx du dx
We make a few observations about this formulation of the Chain
Rule for single-variable real-valued functions. First, the formula in
Equation (5) is completely equivalent to the formula in Equation (3); only
the notation has changed. Second, observe that in Equation (5), we did
not make use of the names of the function f and g, but rather, all we
used was the names of the variables. Third, whereas it appears as if we
could prove Equation (5) by simply canceling the two appearances of
du in the right-hand side of the equation, such canceling, which useful
mnemonically, is not actually a valid thing to do. We will soon see why
it is not valid.
To see the non-matrix version of the Chain Rule for multivariable
vector-valued functions, we first look at various special cases, starting
with functions of the form f : R2 → R and r : R → R2 , and with the
composition f ◦ r. Let t be a real number. By Equation (4) of this section
we see that
D ( f ◦ r )( t )  D f ( r ( t )) Dr ( t ) . (6)

f x ( t ) g in the above equation, let us first write z 

To write out the matrices
f ( x, y ) and r ( t )  y ( t ) . Then Equation (6) can be rewritten as

∂z ∂z
 " dx #
dz
  
 dt
dy .
dt ∂x ∂y dt

Multiplying the matrices on the right-hand side of this equation gives

∂z dx ∂z dy
" #
dz
 
 + .
dt ∂x dt ∂y dt
Of course, using 1 × 1 matrices is silly, and we simply equate the entries
in these two matrices to obtain
dz ∂z dx ∂z dy
 + . (7)
dt ∂x dt ∂y dt
This last equation is the Chain Rule in the particular case of a function
of the form z  f ( x, y ) , where each of x and y are functions of a single
variable t.
THE CHAIN RULE 48

Observe that in the above version of the Chain Rule, we write single-
dy
variable derivatives dz dx
dt , dt and dt , whereas we write partial derivatives
∂z ∂z
∂x
and ∂y . It is important to keep track of which functions have one
variable, which then use the notation for single-variable derivatives,
and which functions have more than one variable, which then use the
notation for partial derivatives.

Example 3
√
Let z  x 2 + y 3 , and x  t and y  sin t. Find dz
dt .
SOLUTION First, we compute the various derivatives (some partial
and some regular) that appear in the right-hand side of Equation (7)
∂z ∂z
of this section, obtaining ∂x  2x, and ∂y  3y 2 , and dx
dt 
1
√ and
2 t
dy
dt  cos t. Next, using Equation (7), we compute

dz ∂z dx ∂z dy
 +
dt ∂x dt ∂y dt
1
 2x · √ + 3y 2 · cos t
2 t
√ 1
 2 t · √ + 3 (sin t ) 2 · cos t
2 t
 1 + 3 sin2 t cos t.

Observe in Example 3 of this section, that after using Equation (7),

which yielded an expression that contained all three variables x, y and t,
we then rewrote each of x and y in terms of t, so that the final answer
involves only the variable t. It would not have been proper to leave the
answer involving all of x, y and t, because we were asked to find dz dt ,
which means that we are thinking of z as a function of t, and hence this
derivative should be a function of t.
We will skip the details, but the same type of argument as above in
the case of w  f ( x, y, z ) , where each of x, y and z are functions of the
single variable t, yields the equation

dw ∂w dx ∂w dy ∂w dz
 + + .
dt ∂x dt ∂y dt ∂z dt

Clearly, the same idea holds for functions of more than three variables.
Next, we look at at functions of the form f : R2 → R and G : Rf2 → R2g ,
x ( s,t )
with the composition f ◦ G. We write z  f ( x, y ) and G x, y  y ( s,t ) .


THE CHAIN RULE 49

Then by Equation (4) of this section we see that

D ( f ◦ G ) x, y  D f ( G x, y ) DG x, y ,

which, when written as matrices, yields,

dy 
∂z ∂z
  dx
∂z ∂z
  
 ds  .
 ds 
∂s ∂t ∂x ∂y  dx dy 
dt dt 


Multiplying the matrices on the right-hand side of this equation gives

∂z ∂x ∂z ∂y ∂z ∂x ∂z ∂y
" #
∂z ∂z
 
 + + ,
∂s ∂t ∂x ∂s ∂y ∂s ∂x ∂t ∂y ∂t

and then equating corresponding entries of these matrices yields

∂z ∂z ∂x ∂z ∂y ∂z ∂z ∂x ∂z ∂y
 + and  + . (8)
∂s ∂x ∂s ∂y ∂s ∂t ∂x ∂t ∂y ∂t

This last equation is the Chain Rule in the particular case of a function
of the form z  f ( x, y ) , where each of x and y are functions of two
variables s and t.
We can now see that the wish of beginning calculus students to cancel
the two appearances of du in the right-hand side of Equation (5) is not
a valid thing to do. If such a cancellation were valid, then it would
plausibly also be valid to cancel the various instances of ∂x and ∂y in
either of the equations of Equation (8), but doing so in the left-hand
equation would yield ∂z ∂s
 ∂z + ∂z , which, upon cancelling, yields 1  1+1,
∂s ∂s
and that is clearly not possible. Hence, this type of canceling is not valid.
The same type of argument as above in the case of w  f ( x, y, z ) ,
where each of x, y and z are functions of two variables s and t, yields
the equations

∂w ∂w ∂x ∂w ∂y ∂w ∂z ∂w ∂w ∂x ∂w ∂y ∂w ∂z
 + + and  + + .
∂s ∂x ∂s ∂y ∂s ∂z ∂s ∂t ∂x ∂t ∂y ∂t ∂z ∂t

Clearly, the same idea holds for functions of more than three variables.

Example 4
∂z ∂z
Let z  5x 2 y, and x  3s + t and y  sin ( st ) . Find ∂s
and ∂t
.
SOLUTION First, we compute the various derivatives (some partial
and some regular) that appear in the right-hand side of Equation (8) of
∂z ∂z
this section, obtaining ∂x  10x y, and ∂y  5x 2 , and ∂x
∂s
 3, and ∂x
∂t
 1,
THE CHAIN RULE 50

∂y dy
and ∂s  t cos ( st ) and dt  s cos ( st ) . Next, using Equation (8), we
compute

∂z ∂z ∂x ∂z ∂y
 +
∂s ∂x ∂s ∂y ∂s
 10x y · 3 + 5x 2 · t cos ( st )
 10 (3s + t ) sin ( st ) · 3 + 5 (3s + t ) 2 · t cos ( st )
 30 (3s + t ) sin ( st ) + 5t (3s + t ) 2 cos ( st )

and
∂z ∂z ∂x ∂z ∂y
 +
∂t ∂x ∂t ∂y ∂t
 10x y · 1 + 5x 2 · s cos ( st )
 10 (3s + t ) sin ( st ) + 5s (3s + t ) 2 cos ( st ) .

Observe in Example 4 of this section, that similarly to Example 3 of

this section, after using Equation (8), which yielded an expression that
contained all three variables x, y, s and t, we then rewrote each of x and
y in terms of s and t, so that the final answer involves only the variable s
and t.
Formulas such as those in Equation (8) are very convenient to use in
practice, and, as we now mention, they can be worked out in any given
situation without the use of matrices. Let us look at the particular case of
a function of the form z  f ( x, y ) , where each of x and y are functions of
two variables s and t. We can write the relation between these variables
via a tree diagram, which, in this particular situation, is seen in Figure 1
of this section.

Figure 1: Tree diagram

∂z
Suppose we want to find the formula for . We then consider the
∂s
tree diagram in Figure 1, and we see that to go from z to s, there are two
routes, one of which is via x and the other via y. These two routes tell us
THE CHAIN RULE 51

that in the Chain Rule for this particular situation, we have two terms,
∂z ∂x ∂z ∂y
which are , representing the route via x, and , representing
∂x ∂s ∂y ∂s
the route via y. We then add these terms, and obtain the equation on
the left-hand side of Equation (8) of this section. Clearly, we can use the
analogous method in other cases.
Finally, we can write out the general version of the Chain Rule without
matrices as follows.

The Chain Rule Without Matrices

Let z  f ( x1 , . . . , x n ) , and where each of x 1 , . . . , x n is a function of
t1 , . . . , t m . Then
∂z ∂z ∂x1 ∂z ∂x2 ∂z ∂x n
 + +···+ , (9)
∂t i ∂x1 ∂t i ∂x 2 ∂t i ∂x n ∂t i
for each i  1, 2, . . . , m.

Ultimately, there is one Chain Rule for multivariable vector-valued

functions, whether we write it via matrices or via equations such as
Equation (8) of this section. Each of these formulations of the Chain Rule
is useful in certain situations.

SUMMARY

Composition of Functions
Let G : Rn → Rk and F : Rk → Rm be functions. The composition of F
and G is the function F ◦ G : Rn → Rm given by the formula

( F ◦ G )( p )  F ( G ( p )) .

The Chain Rule via Matrix Multiplication

THE CHAIN RULE 52

Let G : Rn → Rk and F : Rk → Rm be functions, and let p be a point in

Rn . Suppose that F and G are differentiable. Then

D ( F ◦ G )( p )  DF ( G ( p )) DG ( p ) ,

where the multiplication is matrix multiplication.

The Chain Rule Without Matrices

Let z  f ( x1 , . . . , x n ) , and where each of x 1 , . . . , x n is a function of
t1 , . . . , t m . Then
∂z ∂z ∂x1 ∂z ∂x2 ∂z ∂x n
 + +···+ ,
∂t i ∂x1 ∂t i ∂x 2 ∂t i ∂x n ∂t i
for each i  1, 2, . . . , m.

EXAMPLES

Example 5
The height and width of a rectangle are each changing as a function of
time. Suppose that at a certain instant of time, the height is 3 inches
and is increasing at a rate of 2 inches per second, and the width is 5
inches and is decreasing at a rate of 1 inch per second. Find the rate of
change of the area of the rectangle at that instant of time.
SOLUTION Let h denote the height of the rectangle and let w denote
the width of the rectangle. The area of the rectangle as a function
of height and width is the function A : R2 → R given by the formula
A ( h, w )  hw. Then ∂A
∂h
∂A
 w and ∂w  h, and hence at the given instant
∂A ∂A
of time we have ∂h  5 and ∂w  3.
Let t denote the time. We think of h and w as functions of t. The
information in the problem says that at the given instant of time, we
have dhdt  2 and dt  −1.
dw
 THE CHAIN RULE 53

By Equation (7) of this section we then see that at the given instant
of time, we have
dA ∂A dh ∂A dw
 +  5 · 2 + 3 · (−1)  7.
dt ∂h dt ∂w dt

COMMON MISTAKES

When using the Chain Rule, whether in the matrix formulation or

the non-matrix formulation, the final answer should be expressed in
the variables in the domain of F ◦ G. A common mistake is to mix
different types of variables. For example, when using formulas such
∂z ∂y ∂z ∂y
as ∂z
∂r
∂z ∂x
 ∂x ∂r
+ ∂y ∂r
and ∂z
∂s
∂z ∂x
 ∂x ∂s
+ ∂y ∂s
, the final answer should be
expressed in terms of the variables r and s only, with all instances of x
and y expressed in terms of r and s.

EXERCISES

Basic Exercises

1–3 For each pair of functions F : R2 → R4 and

R3 → R2 , find the formula for ( F ◦ G ) x, y, z .


G:
4–6 For each pair of functions F : R3 → R2 and
 2x−y 
G : R2 → R3 , find each of DF ( G x, y ) and DG x, y



 x y2 
f g
1. F x, y  and G x, y, z  5x+3z



z−y .

 xy 2  and D ( F ◦ G ) x, y directly, and verify the The Chain
Rule via Matrix Multiplication for these functions.
" 5x y #
x−y
f sin( x+3) g
2. F x, y  and G x, y, z 



4y cos ( y−z ) . f 2x−3y+z g  5x

4. F x, y, z  and G x, y 



6x x y+yz
8y .
y−x
" ey # f 3x y g
sin (3x )
3. F x, y  and G x, y, z 


 x+3 
. x yz
f g
5. F x, y, z  and G x, y 



x+2y 5yz x−y+z y+1 .
xy x+y
 THE CHAIN RULE 54

 y−x 
e 2x+z
f g
6. F x, y, z  and G x, y  13. The height and radius of a cylinder are each



y 2 +xz
4x .
xy
changing as a function of time. Suppose that at a
certain instant of time, the height is 3 inches and is
For each set of functions, find dz decreasing at a rate of 5 inches per second, and the
7–9 dt .
radius is 2 inches and is increasing at a rate of 2
inches per second. Find the rate of change of the
7. z  x 4 + 5y, and x  3t + 1 and y  t 2 . volume of the cylinder at that instant of time. Is the
volume increasing or decreasing at that instant of
8. z  3x y 2 , and x  tan t and y  t 3 .
time?
9. z  sin ( x y ) , and x  2t 5 and y  t + 2.
14. The length, width and height of a box are each
changing as a function of time. Suppose that at a
∂z ∂z certain instant of time, the length is 4 inches and is
10–12 For each set of functions, find ∂s
and ∂t
.
increasing at a rate of 2 inches per second, the
width is 3 inches and is decreasing at a rate of 6
10. z  3x + y 2 , and x  s + 2t and y  s − t.
inches per second, and the height is 2 inches and is
11. z  x 2 y, and x  s 2 t 3 and y  2t. increasing at a rate of 1 inch per second. Find the
rate of change of the volume of the box at that
12. z  cos (2x + 3y ) , and x  s 3 and y  st. instant of time. Is the volume increasing or
decreasing at that instant of time?
 Optional Section: Limit Definition of the Derivative for Multivariable Vector-Valued
Functions 55

Optional: Limit Definition of the Derivative

for Multivariable Vector-Valued Functions

In Section 31.2 we saw the definition of the derivative of multivari-

able vector-valued functions. Specifically, if we have a differentiable
function F : Rn → Rm , then the derivative of this function is the matrix
DF, which has the various partial derivatives of the components of
F as the entries in the matrix. One the one hand, this definition of
the derivative is very convenient to compute, and is very useful in
solving a variety of mathematical problems. On the other hand, this
definition is not entirely satisfactory from a conceptual point of view,
because it is not defined as a limit analogous to the limit definition of
the derivative of single-variable real-valued functions. That definition,
as we recall, is that if f : R → R is a function, and a is a real number,
then the derivative of f at a is defined to be

f (a + h ) − f (a )
f 0 ( a )  lim , (1)
h→0 h
provided that the limit exists. Recall too that partial derivatives are
defined via similar limits.
The question arises, can the derivative of multivariable vector-
valued functions also be defined via some sort of limit? Of course,
it would be very troubling if such a limit definition existed but it
were not equivalent to the definition of the derivative of multivariable
vector-valued functions that we have already seen. We will now see
that a limit definition of the derivative of multivariable vector-valued
functions is indeed possible, and it yields the same matrix we are
already familiar with, though we will skip the rigorous details (which
are not trivial).
Let F : Rn → Rm be a multivariable vector-valued function, and let
p be a point in Rn . If we want to find the derivative of F at p via a limit,
the first thing one might attempt is an exact copy of the right-hand
side of Equation (1), but with the single-variable real-valued function
f replaced by the multivariable vector-valued function F, and with the
 Optional Section: Limit Definition of the Derivative for Multivariable Vector-Valued
Functions 56

real numbers a and h and 0 replaced by p and k and 0 in Rn , which are

now thought of as vectors. Such a substitution would yield
F (p + k ) − F (p )
lim .
k→0 k
There is nothing wrong with the numerator of the above fraction,
because we can add and subtract vectors in Rn , but we have a very
serious problem with the denominator, because we cannot divide by
vectors in Rn . So, the above limit simply makes no sense.
To solve our problem, we rearrange Equation (1) into a form that is
more useful to our purposes. First, we bring f 0 ( a ) to the left-hand side
of the equation, and bring it into the limit, which is possible because
f 0 ( a ) does not depend upon h, and we obtain
" #
f (a + h ) − f (a )
lim − f 0 ( a )  0.
h→0 h

Next, we take common denominator, yielding

f ( a + h ) − [ f ( a ) + f 0 ( a ) h]
lim  0.
h→0 h
Finally, we observe that whether or not this fraction goes to zero as h
goes to zero does not change if we modify it by a plus or minus sign,
and hence we deduce that the limit will not change if we introduce an
absolute value in the denominator, obtaining
f ( a + h ) − [ f ( a ) + f 0 ( a ) h]
lim  0. (2)
h→0 |h|
Equation (2) is equivalent to Equation (1). Hence, we could use
Equation (2) to define the derivative of the function f at the point a.
That is, the function f is differentiable at a if there exists a number,
denoted f 0 ( a ) , such that Equation (2) holds. Of course, using Equa-
tion (2) instead of Equation (1) is much less convenient and intuitively
meaningful if we are interested only in the derivatives of single-variable
real-valued functions, but, unlike Equation (1), it turns out that Equa-
tion (2) can be generalized to multivariable vector-valued functions, as
we now see.
As before, let F : Rn → Rm be a multivariable vector-valued function,
and let p be a point in Rn . We now want to consider the analog of
Equation (2), but, again with the single-variable real-valued function f
replaced by the multivariable vector-valued function F, with the real
numbers a and h and 0 replaced by p and k and 0 in Rn (again thought
 Optional Section: Limit Definition of the Derivative for Multivariable Vector-Valued
Functions 57

of as vectors), and with the absolute value of h, denoted |h|, replace by

the length of k, denoted |k|. Such a substitution would yield
F ( p + k ) − [F ( p ) + DF ( p ) k]
lim  0. (3)
k→0 |k|
In contrast with our previous, and unsuccessful, attempt to formu-
late the analog of Equation (1) for the function F, Equation (3) makes
good sense, for the following reasons. First, we note that DF ( p ) is an
m × n matrix, and that k, which is a column vector, can be thought of
an an n × 1 matrix, so that it makes sense to take the product DF ( p ) k,
which is a column vector in Rm . Second, we note that because F ( p + k )
and F ( p ) are vectors in Rm , the addition and subtraction in the nu-
merator of Equation (3) make sense. Finally, we observe that whereas
k is a vector, and hence we cannot divide by it, there is no problem
dividing by the number |k|. Putting these observations together, we
see that Equation (3) is a reasonable criterion we could impose upon
the function F.
Not only does Equation (3) makes sense, but it can be used to define
the derivative of F via limits. That is, we say that the function F is
differentiable at p if there exists an m × n matrix A such that
F ( p + k ) − [F ( p ) + A k]
lim  0.
k→0 |k|
The above definition, while nicely analogous to the limit definition of
the single-variable real-valued functions, raises a number of questions.
Suppose that the function F in the above definition is differentiable
at p. Is the matrix A the same as the matrix of partial derivatives
DF ( p ) that was defined in Equation (2) of Section 1? Indeed, if F
is differentiable, does that guarantee that all the components of F
have all their partial derivatives, which is needed to form the matrix
DF ( p ) ? And, conversely, if all the components of F have all their
partial derivatives, does that guarantee that F is differentiable in the
sense of the limit in Equation (3)?
It turns out that things work out reasonably nicely, as we will now
state, though without giving proofs. First, if F is differentiable in the
sense of the limit in Equation (3), then it is always the case that all
the components of F have all their partial derivatives, and, further,
the matrix A is precisely the matrix of partial derivatives DF ( p ) that
was defined in Equation (2) of Section 1. For the converse, we need a
slight additional hypothesis, namely, that all the components of F have
continuous partial derivatives on some neighborhood of p; if that is
true, then F is differentiable in the sense of the limit in Equation (3).
 Optional Section: Limit Definition of the Derivative for Multivariable Vector-Valued
Functions 58

We will not go into further theoretical details, but let us look at the
relation of the limit definition and the matrix of partial derivatives via
an example.
x2
 
R2 R3 is given by the formula F x, y 


Suppose that F : → 3y .
5x y
Then
2x 0 
DF x, y   0 3  .


5y 5x 
x
Let p  y and k  [ st ]. Then the left-hand side of Equation (3)
becomes
F ( p + k ) − [F ( p ) + DF ( p ) k]
lim
k→0 |k|
 ( x + s ) 2   x 2  2x 0  " #
 3 ( y + t )  − *.  3y  +  0 3  s +/
      t
5 ( x + s )( y + t )  , 5x y  5y 5x 
 limf g
-
√
s → 0 s2 + t2
[t] 0
 x 2 + 2xs + s 2 − x 2 − 2xs − 0 
3y + 3t − 3y − 0 − 3t
 
 
5x y + 5xt + 5ys + 5st − 5x y − 5yt − 5x y 
 limf g √
[ st ]→ 00 s2 + t2
 √ s 2   q s 
 s 2 +t 2   1+ t 22 
s
 lim  0   lim  0  . (4)
 
s→0  5st  s→0  5t 
t→0  √ t→0  q
 s 2 +t 2 
  1+ st 22 
To complete this calculation, we observe that no matter what non-zero
values of s and t we substitute in the expression q 1 2 , the denominator
1+ t 2
s
will always be at least 1, and therefore the whole fraction will be no
greater than 1. It then follows that
s 1
lim q  lim s q  0,
s→0 t2 s→0 t2
t→0 1+ s2 t→0 1+ s2

and similarly
5t 1
lim q  lim 5t q  0.
s→0 t2 s→0 t2
t→0 1+ s2 t→0 1+ s2
We deduce that the last expression in Equation (4) equals the zero
vector, which is what we wanted to verify.
 Optional Section: Limit Definition of the Derivative for Multivariable Vector-Valued
Functions 59

The above example is not a proof that the limit definition of DF ( p )

given in Equation (3) is the same as the matrix of partial derivatives
given in Equation (2), but the example shows that the equivalence of
these definitions, which can be proved, is at least plausible.