Chapter 2.

Functions of several variables

This material is covered in Thomas (chapter 14 in the 11th edition, or chapter 11 in the 10th
edition). On the web site http://www.maths.tcd.ie/

richardt/2E1 there will be a

more detailed explanation of the sections of those chapters we are covering. We start with planes
(and more on lines), which are actually earlier in the book (section 12.5 in the 11th edition, 10.3
in the 10th) as we need to be comfortable manipulating these. It also helps develop an intuition
for working with concepts in space (R
3
).
2.1 Planes in space. The basic fact is that a plane in R
3
has an equation
x + y + z = c
where (, , ) is a (xed) nonzero vector in R
3
perpendicular to the plane and c is a constant.
We can think of equations of the above type, linear equations in 3 variables, as the simplest
kind of equation. Recall that in 2 variables lines are the most basic shapes, and graphs of linear
functions y = mx+c are lines. We are used to thinking of m as the slope and c as the y-intercept
(so that (0, c) is a point on the line). The slope m gives us the direction of the line, and to
distinguish one specic line among all (parallel) lines with the same slope we need to know one
point on the line.
When we look at 3 variables, it is a little simpler to see that we get a plane as the solutions of
a linear equation when we take the case where c = 0, or look at a homogeneous equation
x + y + z = 0.
We can recognise the left hand side as the dot product (, , )(x, y, z) and then we can interpret
the equation that the dot product is 0 in terms of perpendicular vectors: (, , ) (x, y, z).
If we now think geometrically, we see that the points (x, y, z) whose positions vectors are
perpendicular to the xed vector (, , ) are those points which lie in the plane through the
origin (0, 0, 0) perpendicular to (, , ). So the equation x + y + z = 0 does describe a
plane through the origin.
On the other hand, starting with any plane through the origin, we can take a nonzero vector
(, , ) perpendicular to the plane and the equations x + y + z = 0 will describe the plane
we began with.
Before explaining why we get all planes (not necessarily through the origin) by allowing c
to be nonzero, we consider the question of how we might describe the orientation of a plane in
space. For lines in the plane we use the slope usually (though there are vertical lines x = a that
have no slope). We could perhaps think about the slope of a plane as the slope of the steepest line
up the plane, but this will not tell us the orientation of our plane in space. We can move the plane
around (say rotating it around a vertical direction), keeping the same largest slope, but changing
the way the plane is oriented in space.
If you think about it for a moment, you will see that the method of picking a vector perpen-
dicular to the plane (which we sometimes call a normal vector to the plane) works well. You
could maybe construct a little model, like plasterers use, or a at board with a stick handle per-
pendicular to the board (attached by glue or a screw to one side of the board). If you want to
2 200607 Mathematics 2E1 (Timoney)
move the board around parallel to itself you have to keep the stick handle pointing in a constant
direction. (To be more precise, you could also turn the handle around the other way, so that it
will reverse direction completely, and still have the board in the same plane.)
You should see that picking a vector perpendicular to it allows you to say which way a plane
is tilted in space. The normal vector is not unique multiplying the vector by a positive or
negative (but not zero) scalar will not change the orientation of a perpendicular plane.
Now if we think about all the planes perpendicular to a given (or xed) normal vector, then
we can see that we could pick out one of these many planes by knowing one point on the plane.
Say p is a point on our plane, n is a normal vector. How do we come up with an equation
that is satised by all the points on the plane? And of course we want those points that satisfy
the equation to be on our plane. We have the solution in case the given point p = 0 is the origin.
If q is on the plane, then n q = 0.
In general, the position vectors of points q on the plane through p perpendicular to n would
not have to be perpendicular to n. However, the difference vector q p is represented by the
arrow (lying in the plane) from p to q, and so must be perpendicular to n. This is be a sketch of
a plane, the 3 axes and an example of the p and q.
Functions of several variables 3
It should be clear that if any point q R
3
satises q p perpendicular to n, then q is in the
plane. So the plane has equation
n (q p) = 0,
or
n q = n p.
If we write c = n p (which is a constant depending on the chosen point p and the chosen normal
n), then the equation has the form
n q = c.
Writing n = (, , ) and q = (x, y, z) for the arbitrary point on the plane we get the equation
x + y + z = c,
as claimed.
2.2 Example. Find parametric equations for the line of intersection of the two planes
2x 3y + 4z = 0
x + y + 2z = 3
4 200607 Mathematics 2E1 (Timoney)
We can write the parametric equations of the line if we know one point a on the line and a
(nonzero) vector b parallel to the line.
The normal vectors to the planes are (2, 3, 4) and (1, 1, 2). Thus the line is perpendicular
to both vectors, or parallel to their cross product. We can take
b = (2, 3, 4) (1, 1, 2).
To nd a point a on the line, we could (for instance) look for the point on the line where
z = 0, which we would nd by solving 3 simultaneous linear equations (or the system of 3
equations) got by taking the given equations for the two planes and the extra equation z = 0.
2.3 Cartesian equations of lines. If we start with any line in space R
3
, there are many planes
that contain the line. If you take one plane containing the line and rotate it around the line, you
get all the other planes containing that line.
So if you want to represent the line as the intersection of two planes, there are many possi-
bilities. (What this means is to kind of reverse the example above, where we started with two
planes and found parametric equations for the line where they intersect.) There is however a way
to pick two planes in a standard way. Starting with parametric equations for a line
x = a
1
+ b
1
t
y = a
2
+ b
2
t
z = a
3
+ b
3
t
we can solve each of the three equations for the parameter t in terms of all the rest of the quanti-
ties. We get
t =
x a
1
b
1
, t =
y a
2
b
2
, t =
z a
3
b
3
.
(At least we can do this if none of b
1
, b
2
or b
3
is zero.) If (x, y, z) is a point on the line, there
must be one value of t that gives the point (x, y, z) from the parametric equations, and so the 3
values for t must coincide. We get
x a
1
b
1
=
y a
2
b
2
=
z a
3
b
3
.
These equations are called cartesian equations for the line.
You might wonder what sort of equation we have with two equalities in it? Well it is two
equations
x a
1
b
1
=
y a
2
b
2
and
y a
2
b
2
=
z a
3
b
3
.
Each of these two equations is a linear equation, and so the equation of a plane. So we are
representing the line as the intersection of two particular planes when we write the cartesian
equations.
The rst plane
x a
1
b
1
=
y a
2
b
2
Functions of several variables 5
could be rewritten
1
b
1
x
1
b
2
y + 0z =
a
1
b
1

a
2
b
2
so that it looks like x + y + z = c. We see that there is no z term in the equation, or the
normal vector (1/b
1
, 1/b
2
, 0) is horizontal. This means that the plane is parallel to the z-axis
(or is the vertical plane that contains the line we started with).
We could similarly gure out that the second plane
y a
2
b
2
=
z a
3
b
3
is the plane parallel to the x-axis that contains our line.
Now, there is an aspect of this that is a bit untidy. From the cartesian equations we can see
that
x a
1
b
1
=
z a
3
b
3
follows immediately, and this is the equation of the plane parallel to the y-axis that contains our
line. It seems unsatisfactory that we should pick out the two planes parallel to the z-axis and the
x-axis that contains our line, while discriminating against the y-axis in a rather arbitrary way. So
although the cartesian equations reduce to representing our line as the intersection of two planes,
we could take the two planes to be any two of the planes containing the line and parallel to the
x, y and z-axes.
2.4 Example. Suppose we are given two vectors A = (3, 4, 5) and B = (1, 1, 2) that are not
parallel. Then when we take linear combinations
x = tA+ sB (for all t, s R)
we end up with all the points in a plane through the origin, the plane through the origin parallel
to both vectors A and B. (To see this, work graphically. tA is A scaled by an amount t, sB is
B scaled by s, and then we nd tA + sB by completing a parallelogram in the plane of A and
B, and then taking the diagonal of the parallelogram from the origin.)
Let us ask for the equation of that plane.
We need a vector perpendicular to the plane and one such vector is the cross product
AB = (3, 4, 5) (1, 1, 2)
= det
_
_
i j k
3 4 5
1 1 2
_
_
= i det
_
4 5
1 2
_
j det
_
3 5
1 2
_
+kdet
_
2 4
1 1
_
= 13i j 6k
= (13, 1, 6)
Since (0, 0, 0) is a point on the plane, the equation must be
13x y 6z = 0.
6 200607 Mathematics 2E1 (Timoney)
2.5 Remark. It may be instructive to relate the way we deal with planes in R
3
to what would
happen if we dealt with lines in R
2
in the same way.
We know that we usually write the equation of a line in the plane R
2
as y = mx+c. However,
there are vertical lines x = a that have no slope and cant be written in the form y = mx + c.
What we could do, if we wanted to deal with lines that might be vertical , is to rewrite y = mx+c
in the form mx + y = c. This way we can write the equation of every line in the form
x + y = c
where (, ) = (m, 1) for lines that have a slope m, and (, ) = (1, 0) for vertical lines. We
can see then that the vector (, ) is always perpendicular to the line. [The slope of (, ) =
(m, 1) is 1/(m) and the product of the slope m and 1/m is 1, showing that the vector is
perpendicular. In the case of a vertical line, you can see that (1, 0) is a horizontal vector.]
The solutions of a single linear equation x + y = c in R
2
form a line, while the solutions
of a single linear equation x + y + z = c in R
3
form a plane. [We have to rule out the case
where the coefcients on the left are all 0. 0 = c wont be a line or plane.]
2.6 Functions of two variables. We now proceed to the main topic of this chapter, functions of
2 or more variables. We will deal with scalar valued functions of two variables rst.
In the previous chapter we dealt with vector valued functions (or R
n
-valued functions) of a
single variable and we could picture these as curves in R
n
.
You might recall that a function f : A B from a set A (called the domain of the function)
to a set B is a rule that assigns one and only one value in B to each element a A. So there is
one value f(a), which is an element of B, for each a A. The name for the set B, where the
values are found, is less used. We might call it the target set (or the fancier co-domain), as the
word range is used for the set of actual values of f.
In 1E1 you learned about calculus of functions of one real variable, scalar valued functions
y = f(x) where the domain was either all of R or a subset of R. [Often the domain was
an interval or the set R with one bad point removed.] We are so used to seeing examples of
functions given by a formula (like f(x) = 1/x or f(x) = x
2
or f(x) = x cos x + x
2
+ e
x
) that
we can forget that there does not have to be a simple formula to say what f(x) is. There does
have to be a specication of what f(x) is for each x in the domain and that specication cannot
be ambiguous. Only one value f(x) for each x in the domain.
Typically we picture functions of one variable by looking at their graph, and then we see a
curve y = f(x) in R
2
. The graph of a function has one special property no vertical line can
cross it more than once. The graph y =

1 x (by the square root we always mean the positive

square root) is an example. We can infer from the fact that we are not dealing with complex
numbers that we must have the domain A = x R : x 1 = (, 1], or perhaps we could
have some reason to restrict the domain to a smaller set than the numbers 1.
In the following graphs, we see the graph y =

1 x and an annotated version showing

that points x in the domain, drawn along the horizontal axis, have values y = f(x) =

1 x
represented by the height of the point on the graph above x, or the point y on the vertical axis.
You can think of the rule behind the function as associating y to x, or sending x to y, and the
arrow represents this.
Functions of several variables 7
By a (scalar-valued) function of two variables, we mean a function f : R
2
R, or a function
f : A R where A R
2
. Despite the fact that such a function need not be dened by a simple
formula (we only have to have a rule for nding the values f(x, y) at all point (x, y) A, a rule
that is not ambiguous and produces one value f(x, y)), we will usually consider functions that
are specied by a simple formula.
Examples are
(i) f(x, y) = 1 + 2x 3y
(ii) f(x, y) =
x + y
1 x y
2
(iii) f(x, y) =
_
1 x
2
y
2
(iv) f(x, y) = cos x cos y
Technically we should say not only what the rule is for nding the values f(x, y), we should
also say what the domain is. However, what we usually do is take the domain to be either the
whole of R
2
, or the largest subset of R
2
where the given formula makes sense. That means
excluding from the domain points (x, y) where we would be dividing by zero in the formula, or
points where we would end up taking the square root of something negative, or points where we
would be taking the logarithm ln of something that is not strictly positive.
In the example f(x, y) =
_
1 x
2
y
2
, we need to restrict the domain to those (x, y)
where 1 x
2
y
2
0. This set is reasonably easy to describe, if we rearrange the inequality as
x
2
+ y
2
1. It means
(x, y) R
2
: (dist((x, y), (0, 0)))
2
1
and that is the same as the points of distance at most 1 from the origin in the plane R
2
. This set
is then the closed unit disc, the unit circle together with its inside disc.
Though we have used f(x, y) above, we could use any other two names for the two variables.
For example f(x
1
, x
2
) can be handier at times.
2.7 Graphs of functions of two variables. For a function of two variables, one way to picture
it is via its graph. When you remember that graphs of a function of one variable (ordinary scalar
8 200607 Mathematics 2E1 (Timoney)
valued function of one variable) are in two dimensions (the plane), you should not be surprised
that a graph of a function f(x, y) will live in 3-dimensional space R
3
.
We already need a plan (think of the horizontal plane or the oor) to picture the points (x, y)
in the domain of the function, and then what we do for the graph is plot one point (x, y, z) with
z = f(x, y) the height or altitude of the point above (x, y) in the horizontal plane.
For example if f(4, 5) = 3, one of the points we would plot on the graph would be the point
(4, 5, 3) that is 3 units above the point (4, 3) in the horizontal plane:
When we plot all the points (x, y, z) with z = f(x, y) we end up with one point plotted above
each point (x, y) in the horizontal plane and then what we get is a surface in space.
For example in the case f(x, y) = 1 + 2x 3y (a very simple formula as the terms are all
constant or linear) the graph z = f(x, y) is
z = 1 + 2x 3y
and we can recognise that as the equation of a plane
2x + 3y + z = 1
which has (2, 3, 1) = 2i + 3j + k as a normal vector. Perhaps a way to draw it is to notice
that the plane crosses the axes in (1/2, 0, 0), (0, 1/3, 0) and (0, 0, 1). There is just one plane
Functions of several variables 9
through 3 points, unless the points are in a line. Here is a plot of that plane drawn by the program
Mathematica:
In the case f(x, y) = cos x cos y, we can try to piece together what the graph z = f(x, y)
looks like by looking at sections of the graph where x or y is constant. For example if we
freeze y = 0, we are looking at the section of the whole graph above the x-axis, where we see
z = f(x, 0) = cos x, a standard cosine wave starting at z = 1 when x = 0.
If we look at what happens when x = 0, which is along the y-axis, we get z = f(0, y) = cos y
the same shape again. If, instead of taking a section with x = 0 we take a section where
x = x
0
is some other constant we get the graph z = cos x
0
cos y. This is a modied version of
the regular cosine wave z = cos y, which now oscillates between z = cos x
0
and z = cos x
0
.
Looking at sections in the perpendicular direction, where y = y
0
is constant we get a similar
shape z = f(x, y
0
) = cos x cos y
0
= (cos y
0
) cos x.
The whole graph z = f(x, y) is a sort of landscape. If we think of the x-axis as pointing East
and the y-axis as pointing North (along the horizontal oor), we have a landscape that always
looks like a cosine wave if we follow it either East-West or North-South. Perhaps it takes some
intuition to see what it is really like and so here is a picture drawn by Mathematica.
10 200607 Mathematics 2E1 (Timoney)
You can see peaks and valleys. The highest points happen when cos x = sin y = 1, or when
cos x = sin y = 1, and there we get z = 1. The lowest points happen when cos x = 1 and
sin y = 1, or when cos x = 1 and sin y = 1. At these lowest points we get z = 1.
This may be a good place to explain that the altitude z = f(x, y) of the points on a graph can
be negative as well as positive. Negative altitude points happen when the graph dips below the
oor level.
Next we look at z = x
2
+ y
2
, a somewhat easier example to gure out. This is the graph for
f(x, y) = x
2
+y
2
, and if you look to see what happens when y = 0 we get z = f(x, 0) = x
2
+0,
or z = x
2
. This is a familiar parabola graph of x
2
. For the section along the y-axis, where x = 0
we get z = f(0, y) = y
2
and this is again the same shape.
If you take a section of this graph in some other direction from the origin, you get still the
same shape because
f(x, y) = x
2
+ y
2
= (dist((x, y), (0, 0)))
2
.
When you piece this information together, you can see that this is what the graph looks like.
This is a graph that is rotationally symmetric around the z-axis and looks like a standard
parabola (x
2
graph) when you take any vertical section through the origin. It is called a paraboloid
Functions of several variables 11
and in fact this is the shape for a reector in a standard torch (or car head-lamps when they used
to be just round). So it is a shape that has some applications.
2.8 Partial derivatives. We move on now to calculus for functions of two variables. There
are a number of different concepts to absorb, but the simplest to explain is the idea of a partial
derivative.
Start with a function z = f(x, y) (which we can visualise in terms of its graph) and a point
(x
0
, y
0
). To get the partial derivative with respect to x of f(x, y) at the point (x
0
, y
0
) we consider
the function of a single variable x that we get by xing y to be y
0
. That is the function x
f(x, y
0
), or z = f(x, y
0
).
In terms of the graph z = f(x, y) of the whole function of two variables, we can see the
values z = f(x, y
0
) as those values along the line y = y
0
parallel to the x-axis in the (horizontal)
x-y-plane. That means that the graph of the function of one variable z = f(x, y
0
) can be got
from the whole graph of the function of two variables by taking a section at y = y
0
.
Here is an attempt at a picture showing what the section of z = cos x cos y at y = 0.9 looks
like:
One is a three dimensional picture cutting away the part of the graph z = f(x, y) where y < 0.9
and the other is a simple graph of the prole.
What we do now is take the derivative with respect to x of this f(x, y
0
) and evaluate it at
x = x
0
. This gives the partial derivative at (x
0
, y
0
), the partial derivative with respect to x.
While we denote regular derivatives either by a d/dx of a prime notation, we denote the partial
derivative by /x. So the partial derivative evaluated at (x
0
, y
0
) is denoted
f
x
[
(x
0
,y
0
)
,
z
x
[
(x
0
,y
0
)
, or sometimes f
x
(x
0
, y
0
).
(We will mostly use the f/x notation rather than the f
x
notation, but the f
x
notation is also
used quite often.) We could summarise the denition as
f
x
[
(x
0
,y
0
)
=
d
dx
[
x=x
0
f(x, y
0
)
12 200607 Mathematics 2E1 (Timoney)
If we want to see graphically what we are getting, recall that the ordinary derivative (of a
function of one variable) is the slope of the tangent line to the graph at the point. So if for
example, we were taking f(x, y) = cos x cos y, (x
0
, y
0
) = (0.5, 0.9) we would be calculating the
slope of this line
We can calculate this slope (or this derivative) exactly:
f(x, y
0
) = cos x cos y
0
= cos x cos 0.9
d
dx
f(x, y
0
) = sin x sin y
0
= sin x sin 0.9
d
dx
[
x=x
0
f(x, y
0
) = sin x
0
cos y
0
= sin 0.5 cos 0.9 = 0.298016
f
x
[
(x
0
,y
0
)
= 0.298016
Looking at the calculation, you can see that what we do is replace y by the constant y
0
and
then take the derivative d/dx with respect to x (and nally put x = x
0
). We usually do it in a
slightly quicker way. When taking the partial /x with respect to x, we treat the variable y as
a constant. (We dont have to go so far as to replace y by a specic number y
0
.) If we repeat the
above calculation this way, it would look like this:
f(x, y) = cos x cos y
f
x
= sin x cos y
f
x
[
(x
0
,y
0
)
= sin x
0
cos y
0
= sin 0.5 cos 0.9
= 0.298016.
Functions of several variables 13
Summary: To get the partial derivative with respect to x, take the derivative with respect to
x while treating the variable y as though it was a constant.
Nowpartial derivatives with respect to y are similar, except we freeze x = x
0
and differentiate
with respect to a variable y this time.
f
y
[
(x
0
,y
0
)
=
d
dy
[
y=y
0
f(x
0
, y)
Or, we can summarise by saying: To get the partial derivative with respect to y, take the deriva-
tive with respect to y treating the variable x as though it was a constant.
Graphically it means taking a section though our graph z = f(x, y) in the x-direction, and
then nding the slope of that.
If we take our example again f(x, y) = cos x cos y, (x
0
, y
0
) = (0.5, 0.9), and nd the partial
with respect to y this time, we nd
f(x, y) = cos x cos y
f
y
= cos x(sin y)
= cos x sin y
f
x
[
(x
0
,y
0
)
= cos x
0
sin y
0
= cos 0.5 sin 0.9
= 0.687434.
2.9 Examples. It is necessary to become accustomed to taking partial derivatives. The rules are
as for derivatives of functions of one variable, and so it is a question of getting practice in these
again.
You need to remember the rules chain rule, product rule, perhaps the quotient rule (or
else know how to treat
u
v
= uv
1
as the product of u times the reciprocal of v), and the basic
differentiation formulae for cos t, sin t, e
t
, ln t, tan t, sec t and so forth.
Exercise: Look at the exercises in Thomas Calculus (14.3 the 11th edition, or 11.3 in the
10th) entitled Calculating rst order Partial derivatives and try them for yourself. The book has
the answers to the odd numbered ones so you can see how you are doing.
2.10 Directional derivatives. With partial derivatives we have concentrated on (vertical) sec-
tions in the directions parallel to the two axes (the x-axis for /x and the y-axis for /y), but
there seems to be no compelling reason to take only the slope of our graph z = f(x, y) in these
two directions.
Why should we not consider all possible directions through (x
0
, y
0
)? The slope we get is
called a directional derivative.
To explain it, we should nd a good way to describe a direction. We know vectors in the
plane have a direction and a magnitude. As we are only interested in a direction, we can stick to
unit vectors (magnitude = 1) u to describe a direction. Going through (x
0
, y
0
) in the direction u
we have a line and we know how to describe this line with parametric equations:
(x, y) = (x
0
, y
0
) + tu
14 200607 Mathematics 2E1 (Timoney)
At t = 0 we have our point (x
0
, y
0
) and we are heading in the direction specied by u. If we
now look at the values of f(x, y) along this line, we get the function
t f((x
0
, y
0
) + tu)
of the parameter (single variable) t. The graph of this function of t is what we would get by
taking a (vertical) section through the graph z = f(x, y). It is the section through (x
0
, y
0
)
in the direction u. We are going to take the slope of this section at the point (x
0
, y
0
) we are
concentrating on. We can see that the slope is the same
1
as the slope of the function of t, that is
the same as the derivative with respect to t at t = 0:
d
dt
[
t=0
f((x
0
, y
0
) + tu)
We dene the directional derivative D
u
f(x
0
, y
0
) of f in the direction u at (x
0
, y
0
) to be
D
u
f(x
0
, y
0
) =
d
dt
[
t=0
f((x
0
, y
0
) + tu)
We can notice that if we take u = i = (1, 0) = the unit vector in the direction of the positive
x-axis, then
D
u
f(x
0
, y
0
) = D
i
f(x
0
, y
0
)
=
d
dt
[
t=0
f((x
0
, y
0
) + ti)
=
d
dt
[
t=0
f((x
0
+ t, y
0
))
=
d
dx
[
x=x
0
f(x, y
0
)
=
f
x
[
(x
0
,y
0
)
Similarly for u = j = (0, 1) the directional derivative D
j
f is the same as f/y.
We can compute what D
u
f(x
0
, y
0
) for the example we have used above, that is f(x, y) =
cos x cos y and (x
0
, y
0
) = (0.5, 0.9). Well work it out for an arbitrary unit vector u = (u
1
, u
2
).
D
u
f(x
0
, y
0
) =
d
dt
[
t=0
f((x
0
, y
0
) + tu)
=
d
dt
[
t=0
f(x
0
+ tu
1
, y
0
+ tu
2
)
=
d
dt
cos(x
0
+ tu
1
) cos(y
0
+ tu
2
)
= (u
1
sin(x
0
+ tu
1
) cos(y
0
+ tu
2
) cos(x
0
+ tu
1
)u
2
sin(y
0
+ tu
2
)) [
t=0
= u
1
sin x
0
cos y
0
u
2
cos x
0
sin y
0
= sin x
0
cos y
0
u
1
cos x
0
sin y
0
u
2
= 0.298016u
1
0.687434u
2
.
1
For this to be true we are relying on the fact that u is a unit vector if we allowed non unit vectors, the scale
along the t axis would not correspond to the distance scale in the horizontal (x, y) plane, and this would affect the
calculation of the slope.
Functions of several variables 15
Youll notice that the same numbers are coming in as we had for the partial derivatives. Of
course this was just one example of a very simple kind, and the fact that the directional derivative
involves the same numbers we got from the partial derivatives might be a uke.
If you try more examples though, you will begin to see that we seem to always have this
D
u
f(x
0
, y
0
) =
_
f
x
[
(x
0
,y
0
)
_
u
1
+
_
f
y
[
(x
0
,y
0
)
_
u
2
(1)
Or we might be better to use the other notation for partials
D
u
f(x
0
, y
0
) = f
x
(x
0
, y
0
)u
1
+ f
y
(x
0
, y
0
)u
2
or to just write the formula at an arbitrary (x, y) (without trying to write down that both sides
have to be evaluated at the same point):
D
u
f =
f
x
u
1
+
f
y
u
2
.
Anyhow, it turns out that this formula is usually true. Not quite always, but usually in exam-
ples we can easily compute with. We will explain a little more carefully later when it is a valid
formula.
The upshot of all this is that the values of the partial derivatives contain the information
needed to nd all the directional derivatives.
2.11 Tangent plane to a graph. In order to explain what it is that makes the formula for direc-
tional derivatives given above true, and the circumstances where we can be sure it is true, we
explain about the tangent plane to a graph z = f(x, y) at a point (x
0
, y
0
, z
0
) on the graph.
Just as in one variable there are some graphs with no tangent line, or functions that are not
differentiable at certain points, there are cases of graphs z = f(x, y) that do not have a tangent
plane. Before we can make that more clear, we will need to say what we mean by a tangent
plane.
What we will do rst is work out what the equation for the tangent plane should be, assuming
there is a tangent plane. Then we will be able to say what the denition of a tangent plane is.
So we look for the tangent plane to a graph z = f(x, y) at a point (x
0
, y
0
, z
0
) on the graph.
One simple thing that we need in practice is that if the point is on the graph then z
0
= f(x
0
, y
0
).
(We use this to work out the z-coordinate if given the x- and y-coordinates.)
Then we consider the sections of the graph in the x- and y-directions, the same sections
we looked at when we were describing partial derivatives. Rather than considering the graphs
z = f(x, y
0
) (where we x y = y
0
) and z = f(x
0
, y) (where we x x = x
0
) as graphs in a
vertical plane we look at them as curves in space. We can parametrise them as
x = t
y = y
0
z = f(t, y
0
)
16 200607 Mathematics 2E1 (Timoney)
and
x = x
0
y = t
z = f(x
0
, t
0
).
Look at the rst of these
(x, y, z) = (t, y
0
, f(t, y
0
))
which passes by our point (x
0
, y
0
, z
0
) at t = x
0
. It will have a tangent vector at that point
d
dt
[
t=x
0
(t, y
0
, f(t, y
0
)) =
_
1, 0,
d
dt
f(t, y
0
)
_
[
t=x
0
=
_
1, 0,
f
x
[
(x
0
,y
0
)
_
Whatever the tangent plane to the graph z = f(x, y) is, it seems reasonable to believe that
this tangent vector should be in the plane (or parallel to it).
Similarly working with the curve we get along the y-direction, we have
(x, y, z) = (x
0
, t, f(x
0
, t))
and the tangent vector to that (at t = y
0
) will come out as
d
dt
[
t=y
0
(x
0
, t, f(x
0
, t)) =
_
0, 1,
d
dt
f(x
0
, t)
_
[
t=y
0
=
_
0, 1,
f
y
[
(x
0
,y
0
)
_
.
This vector should be in the plane also, for the same reason.
So our tangent plane should be parallel to both the vectors
A = (1, 0, f
x
(x
0
, y
0
)) and B = (0, 1, f
y
(x
0
, y
0
)).
From what we did in Example 2.4 above we know then that A B should be normal to our
tangent plane. We can work that out:
AB = det
_
_
i j k
1 0 f
x
(x
0
, y
0
)
0 1 f
y
(x
0
, y
0
)
_
_
= i det
_
0 f
x
(x
0
, y
0
)
1 f
y
(x
0
, y
0
)
_
j det
_
1 f
x
(x
0
, y
0
)
0 f
y
(x
0
, y
0
)
_
+kdet
_
1 0
0 1
_
= f
x
(x
0
, y
0
)i f
y
(x
0
, y
0
)j + 6k
= (f
x
(x
0
, y
0
), f
y
(x
0
, y
0
), 1)
We can then write down that the equation of the plane through (x
0
, y
0
, z
0
) perpendicular to that
vector should be
f
x
(x
0
, y
0
)(x x
0
) f
y
(x
0
, y
0
) + (z z
0
) = 0.
Rearranging it to have fewer minus signs, we get
z = z
0
+ f
x
(x
0
, y
0
)(x x
0
) + f
y
(x
0
, y
0
).
Recalling that z
0
= f(x
0
, y
0
), we nd that the tangent plane should have this equation
Functions of several variables 17
z = f(x
0
, y
0
) +
_
f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
)
_
(y y
0
)
(if there is a tangent plane).
2.12 Linear approximation. We turn now to the denition of what it means for z = f(x, y) to
have a tangent plane at (x
0
, y
0
, z
0
). The idea is fairly simple the graph z = f(x, y) should be
close to the graph of the plane near the point (x
0
, y
0
).
We should be able to say that there is a plane
z = z
0
+ a
1
(x x
0
) + a
2
(y y
0
)
(or graph of a function with only constant and linear terms) through (x
0
, y
0
, z
0
) so that it is true
that
f(x, y)

= z
0
+ a
1
(x x
0
) + a
2
(y y
0
)
is a good approximation when we keep (x, y) not too far from (x
0
, y
0
). Unless we are a bit more
precise about what good means, this will still be ambiguous. We want to insist on something that
cant be true of two different planes and is still often achievable.
What we do is insist that the error in the above approximation should be small compared to
the distance from (x, y) to (x
0
, y
0
) (when that distance is itself small).
We make the denition that a function z = f(x, y) is differentiable at (x
0
, y
0
) if it is possible
to nd constants a
1
and a
2
so that
lim
(x,y)(x
0
,y
0
)
f(x, y) (f(x
0
, y
0
) + a
1
(x x
0
) + a
2
(y y
0
))
dist((x, y), (x
0
, y
0
))
= 0.
We could rewrite that
lim
(x,y)(x
0
,y
0
)
f(x, y) (f(x
0
, y
0
) + a
1
(x x
0
) + a
2
(y y
0
))
_
(x x
0
)
2
+ (y y
0
)
2
= 0. (2)
Sometimes the term total derivative is used for this rather than derivative.
We are being a bit cavalier here in bringing in limits of functions of two variables with no
explanation. The book by Thomas has a lot more than we are going to do about this. One thing
to point out is that when we let (x, y) approach (x
0
, y
0
), we never actually take (x, y) to be equal
to (x
0
, y
0
). That means we dont end up having to worry about interpreting 0/0 in the above
limit. Another important thing is that (x, y) does not just approach (x
0
, y
0
) from one direction,
but from all directions at once. The interpretation of the limit being 0 is that the thing we are
taking the limit of will be close to 0 as long as we take (x, y) within some small distance of
(x
0
, y
0
). But we must not have to make the small distance be 0, and we should never consider
(x, y) = (x
0
, y
0
).
18 200607 Mathematics 2E1 (Timoney)
We claim now that the number a
1
in the denition must always be the partial derivative with
respect to x at (x
0
, y
0
) (and the number a
2
must be the partial with respect to y). For this we go
back to rst principles, and write the partial with respect to x as a limit
f
x
[
(x
0
,y
0
)
=
d
dx
[
x=x
0
f(x, y
0
) = lim
xx
0
f(x, y
0
) f(x
0
, y
0
)
x x
0
What we are going to prove is that, if we know (2), then the partial derivative with respect to
x must exist at (x
0
, y
)
and must be the number a
1
. We do this by looking at the difference
f(x, y
0
) f(x
0
, y
0
)
x x
0
a
1
=
f(x, y
0
) f(x
0
, y
0
) a
1
(x x
0
)
x x
0
=
f(x, y
0
) (f(x
0
, y
0
) + a
1
(x x
0
))
x x
0
and its absolute value

f(x, y
0
) f(x
0
, y
0
)
x x
0
a
1

=
[f(x, y
0
) (f(x
0
, y
0
) + a
1
(x x
0
))[
[x x
0
[
.
From (2) we know that this is small as long as
dist((x, y
0
), (x
0
, y
0
))
is small, which is the same as
[x x
0
[ small.
In this way we prove that
lim
xx
0
f(x, y
0
) f(x
0
, y
0
)
x x
0
= a
1
must be true if (2) is true. That is
f
x
[
(x
0
,y
0
)
= a
1
Similarly, it is possible to show that if (2) is true, then
f
y
[
(x
0
,y
0
)
= a
2
and in fact that for any unit vector u
D
u
f(x
0
, y
0
) = a
1
u
1
+ a
2
u
2
must be true if (2) holds.
This justies the equation (1) if we know that f(x, y) has a derivative at (x
0
, y
0
). In fact,
most functions that we can easily write down do have derivatives. At the worst they usually can
have some bad points where things go wrong, points (x
0
, y
0
) where there is no derivative.
Functions of several variables 19
To make a precise statement, here is the theorem (but we will not go any way towards proving
it). If the partial derivatives f
x
(x, y) = f/x and f
y
(x, y) = f/y exist at all points near
(x
0
, y
0
) and we have
lim
(x,y)(x
0
,y
0
)
f
x
(x, y) = f
x
(x
0
, y
0
) and lim
(x,y)(x
0
,y
0
)
f
y
(x, y) = f
x
(x
0
, y
0
)
then f is differentiable at (x
0
, y
0
).
With one more bit of terminology, we can state this more tidily. A function f(x, y) is called
continuous at the point (x
0
, y
0
) if it is true that
lim
(x,y)(x
0
,y
0
)
f(x, y) = f(x
0
, y
0
).
We wont discuss this much. (There is some discussion in the book by Thomas.) The summary
is that this is usually true, apart from maybe one or two bad points, of functions that can be easily
written down by a formula. In the practical real world, there can be problems because functions
that have to be considered arise from the real world rather than from nice tidy formulae. In these
cases, we could be faced with discontinuous functions.
The theorem stated above can be restated like this
Theorem. If a function z = f(x, y) has partial derivatives f/x and f/y that
are continuous functions at (x
0
, y
0
), then f has a derivative at (x
0
, y
0
).
This means (2) holds with
a
1
=
f
x
[
(x
0
,y
0
)
and a
2
=
f
y
[
(x
0
,y
0
)
The reason this is useful is that we can usually see that we are in the situation where the partial
derivatives are continuous. You can compute partial derivatives easily and then looking at the
formulae you get, you can usually see that you are in the situation where they give you continuous
functions.
As we are not going to go into this in any detail, we will just summarise by saying there are
theorems that justify the claim that most of the functions we encounter have derivatives, and then
the directional derivative formula (1) is valid.
Restating (2) in a less precise way, and using what we know about what a
1
and a
2
have to be
we get the linear approximation formula
f(x, y)

= f(x
0
, y
0
) +
_
f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
)
_
(y y
0
)
which is valid for (x, y) close to (x
0
, y
0
).
There is an example of working this out with numerical values for everything on the solutions
to Tutorial 7 (question 2). But the more important fact is that linear approximation is a way to
20 200607 Mathematics 2E1 (Timoney)
tackle complicated problems by reducing themto linear ones. We can solve most linear problems.
The theory of linear problems (linear algebra) is well enough understood that we know how to
do most things. The difculty is that many of the interesting problems we encounter in the real
world of applications are not linear. With linear approximation we can tackle these problems by
replacing them by linear problems (which we can usually solve pretty easily). However, we are
then left with the question of how to relate the solution to the approximate problem to the actual
nonlinear problem.
2.13 The gradient vector. We dene the gradient vector at (x
0
, y
0
) of a function (x, y) of two
variables as
f [
(x
0
,y
0
)
=
_
f
x
[
(x
0
,y
0
)
,
f
y
[
(x
0
,y
0
)
_
The symbol has a name nabla but we usually just pronounce it gradient. We can maybe
simplify the above denition by removing the evaluation at (x
0
, y
0
) and writing
f =
_
f
x
,
f
y
_
where we understand that both sides are to be evaluated at the same place.
Weve seen that the partial derivatives are crucial numbers for several purposes and the idea
of the gradient is to wrap up both numbers in one (vector) quantity. It is perhaps a short step, but
it is quite a useful step.
2.14 The gradient vector and directional derivatives. We can restate the formula (1) using the
gradient vector as
D
u
f(x
0
, y
0
) =
_
f [
(x
0
,y
0
)
_
u
Remembering how dot products work, we can write this as
D
u
f(x
0
, y
0
) =
_
_
f [
(x
0
,y
0
)
_
_
|u| cos
where is the angle between the gradient vector and the direction u. Since |u| = 1, we have
D
u
f(x
0
, y
0
) =
_
_
f [
(x
0
,y
0
)
_
_
cos
Since 1 cos 1, we can then say that

_
_
f [
(x
0
,y
0
)
_
_
D
u
f(x
0
, y
0
)
_
_
f [
(x
0
,y
0
)
_
_
Moreover, we can say that the largest directional derivative is the magnitude of the gradient
vector and it is achieved only when the direction u is the unit vector in the direction of the
gradient vector. That is the largest directional derivative is when
u =
f [
(x
0
,y
0
)
_
_
f [
(x
0
,y
0
)
.
_
_
The smallest possible directional derivative is minus the length of the gradient vector and that
happens only when u is the unit vector in the opposite direction to the gradient.
(All this supposes that the gradient vector is not the zero vector, which can happen. In that
case all the directional derivatives are 0.)
Functions of several variables 21
2.15 The Chain Rule for partial derivatives. We can now state one version of the chain rule
that involves functions of more than one variable.
First a quick refresher on the chain rule in one variable. It says that if y = f(x) and x = x(t)
so that in the end we can regard y as a function of t by y = f(x(t)), then
dy
dt
=
dy
dx
dx
dt
.
In other words it is ok to cancel the dxs in this Leibniz notation for derivatives.
For a function z = f(x, y) of two variables where each of x and y is itself dependent on t
by x = (t) and y = y(t), we can ask how the derivative dz/dt is related to the derivatives of the
functions that make up the composition z = f(x(t), y(t)). The formula is
dz
dt
=
f
x
dx
dt
+
f
y
dy
dt
So the simple rule of cancelling the dxs becomes less obvious for partial derivatives. Some-
how we end up with only one fraction dz/dt when we cancel both the partials x and y.
In vector form we can express the formula as
dz
dt
= (f)
_
dx
dt
,
dy
dt
_
So we can see the gradient multiplied (or dot producted with) the derivative vector (or velocity
vector) of the parametric curve
(x, y) = (x(t), y(t))
We wont prove this formula, but the way it is usually proved relies on linear approximation.
Fix on a value t = t
0
where we want to prove the chain rule formula. Then take (x
0
, y
0
) =
(x(t
0
), y(t
0
)) in the linear approximation formula
f(x, y)

= f(x
0
, y
0
) +
_
f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
)
_
(y y
0
).
Into this we put the linear approximation formulae in one variable for x(t) and y(t) around t = t
0
.
These say
x(t)

= x(t
0
) + x

(t
0
)(t t
0
)
y(t)

= y(t
0
) + y

(t
0
)(t t
0
)
(or that the graphs x = x(t) and y = y(t) are close to the graphs of their tangent lines at t
0
, if we
keep t near t
0
.
Combining the three linear approximations we get
f(x(t), y(t))

= f(x
0
, y
0
) +
_
f
x
[
(x
0
,y
0
)
_
x

(t
0
)(t t
0
) +
_
f
y
[
(x
0
,y
0
)
_
y

(t
0
)(t t
0
)
22 200607 Mathematics 2E1 (Timoney)
and this leads to a proof that
d
dt
[
t=t
0
f(x(t), y(t)) = lim
tt
0
f(x(t), y(t)) f(x(t
0
), y(t
0
))
t t
0
has the value it is supposed to have in the chain rule formula.
We will not go into details.
2.16 Level curves. We have concentrated so far on visualising function z = f(x, y) by their
graphs, which are surfaces (or look like landscapes) in space.
There is another way which we will now consider. It is to draw (on a two dimensional plane)
the curves
f(x, y) = c
for different values of c. These are called the level curves for the function f(x, y).
There are everyday examples where we see these curves. On the weather chart on the TV or
in the newspaper you see lines (called isobars) where the pressure is constant. So these are level
curves for the function f(x, y) = atmospheric pressure at point (x, y) on the earth. [Perhaps there
is a small problem as the earth is not a plane.] We also sometimes see charts of the temperature
today or tomorrow, usually coloured (so that red zones are hot, blue zones are cold). The dividing
curves between the different colours are curves where the temperature is constant (30 degrees,
25 degrees, or whatever). So they are level curves for the temperature function.
An example that is maybe better for us are the contour lines we see on certain maps. There
will be many such contour lines around mountains or hills, perhaps marked with 100 metres, 200
metres and so on. They are lines on the map of tracks you could in principle follow that would
be neither uphill nor downhill. They are lines of constant altitude, or level curves for the function
f(x, y) = altitude. Off shore they can represent negative altitudes (or depths of the sea bed).
This is exactly how our level curves for functions f(x, y) correspond to the graph z = f(x, y) of
a function of two variables.
We can maybe look to some of our examples to see how these level curves work out. Here is
the example z = 1 + 2x 3y we plotted rst done with contours:
Functions of several variables 23
It is maybe not so useful without indicating what the values of c is on each line. We could see by
hand that 1 + 2x 3y = c is a line y = (2/3)x + (c 1)/3 of slope 2/3 and that is what we see
in the picture various lines of that slope.
Here is the picture for z = cos x cos y, once with just the level curves, and then with shading.
The lighter shaded regions (like near the origin (x, y) = (0, 0)) are higher values.
For the example z = x
2
+ y
2
, it is quite easy to see that the level curves x
2
+ y
2
= c are
circles (of radius

c) around the origin if c > 0. For c = 0 the level curve x

2
+ y
2
= 0 is just
the one point (0, 0), and for c < 0 the level curve is empty. Here is a (shaded) of some of the
contours from the Mathematica programme:
24 200607 Mathematics 2E1 (Timoney)
From the fact that the level curves are circles we could immediately realise that the function is
symmetrical around the origin, if we did not know that already. We can build the graph, which
we saw before was a paraboloid, by lifting up the level curves x
2
+ y
2
= c to height c above the
horizontal plane.
This may help you to see the relationship between the level curves and the graph. The map
example, where they correspond to tracks you can walk along the landscape without going up or
down at any stage, is also helpful.
Though a graph is maybe easier to visualise, it takes 3 dimensions. Maps with contour lines
showing (say) the Mourne mountains could be harder to interpret, but they are easier to make
than 3-dimensional models of the mountains.
2.17 Level curves and the gradient. Here we explain that the gradient vector at (x
0
, y
0
) of
f(x, y) will always be perpendicular to the level curve f(x, y) = c that goes through the point
(x
0
, y
0
). That is the one where c = f(x
0
, y
0
).
Let us start with an example, the simple linear example z = 1 + 2x 3y and the point
(x
0
, y
0
) = (7, 11). We pictured the level curves (lines) for this earlier. So in this case we are
dealing with the function f(x, y) = 1 + 2x 3y and we can work out the right value of c so
that the level curve f(x, y) = c passes through (x
0
, y
0
) = (7, 11). We need f(7, 11) = c, or
1 + 14 33 = c, c = 18. So the level curve is 1 + 2x 3y = 18 or 2x 3y = 19. As this
is a line we can get its slope by rewriting it as 2x + 19 = 3y, or y = (2/3)x + 19/3, so that the
slope is 2/3. If we nd the gradient of f we get
f =
_
f
x
,
f
y
_
= (2, 3).
So the slope of the gradient vector is 3/2 and the product of this with the slope of the level
curve (line) is (3/2)(2/3) = 1. This shows they are perpendicular (in this example).
Functions of several variables 25
We take now another example f(x, y) = x
2
+y
2
and the point (4, 1). We saw this example
before and we know the level curves are concentric circles centered at the origin. So the tangent
lines are perpendicular to the radius. In particular the tangent line to the circle through (4, 1) is
perpendicular to the vector (4, 1). But if we work out the gradient vector we get
f [
(4,1)
=
_
f
x
,
f
y
_
[
(4,1)
= (2x, 2y) [
(4,1)
= (8, 2).
As this gradient vector is a multiple of (4, 1) we see that it is indeed perpendicular to the tangent
line (to the level curve, or level circle in this example).
Weve seen that the claimed fact that the gradient is always perpendicular to the tangent line
to the level curve has worked out in two examples. Now we will indicate why it is always true.
The level curve f(x, y) = c through a point (x
0
, y
0
) can be described in a different way by
parametric equations
_
x = x(t)
y = y(t)
in such a way that (x(0), y(0)) = (x
0
, y
0
). We then have the equation
f(x(t), y(t)) = c
satised for all t. If we differentiate both sides with respect to t we get
d
dt
f(x(t), y(t)) = 0
f
x
dx
dt
+
f
y
dy
dt
= 0
by using the chain rule. When we put t = 0 this comes down to
_
f
x
[
(x
0
,y
0
)
,
f
y
[
(x
0
,y
0
)
_

_
dx
dt
[
t=0
,
dy
dt
[
t=0
_
= 0.
This means that the dot product of the gradient vector of f at the point (x
0
, y
0
) is perpendicular
to the velocity vector to the curve.
There are some details to ll in, but this is one way to prove that the gradient is perpendicular
to the level curve.
2.18 Example. Find the equation of the tangent line to
cos(2x) cos y =
1
2
at the point (/8, /4).
26 200607 Mathematics 2E1 (Timoney)
To do this we need the gradient vector to f(x, y) = cos x cos y at the point (/4, /4).
f [
(/8,/4)
=
_
f
x
,
f
y
_
[
(/8,/4)
= (2 sin(2x) cos y, cos(2x) sin y) [
(/8,/4)
= (2(1/

2)(1/

2), (1/

2)(1/

2)) = (1, 1/2)

A quick way to write down the equation of the line through (/4, /4) perpendicular to the
vector (1, 1/2) is this:
(1)
_
x

8
_
+
_

1
2
_
_
y

4
_
= 0.
2.19 Gradient is direction of steepest ascent. The word gradient vector is justied by the fact
that it relates to the direction of steepest ascent.
In a way we know this already because we know from 2.14 that the directional derivative
D
u
f(x
0
, y
0
) is largest (for a xed point (x
0
, y
0
)) when u is the unit vector in the direction of the
gradient f [
(x
0
,y
0
)
. We also know that the largest possible value of the directional derivative is
the magnitude of the gradient vector.
We can relate this to the graph z = f(x, y). We can thing of the graph as a landscape. Sup-
pose we are starting from the point on the graph where (x, y) = (x
0
, y
0
). The directional deriva-
tive D
u
f(x
0
, y
0
) gives us the slope we would encounter if we travelled from position (x
0
, y
0
) in
the direction u. We get the largest slope in the direction of the gradient, the smallest slope in the
opposite direction.
That means that the gradient vector is the most uphill direction.
The level curve direction is the direction perpendicular to the gradient. That is what we just
gured out above when we showed that the gradient vector is perpendicular to (the tangent line)
to the level curve. It also makes sense that if you travel along the level curve you will encounter
zero slope.
2.20 Functions of 3 variables and level surfaces. We look now at functions f(x, y, z) or
f(x
1
, x
2
, x
3
) of 3 variables.
A lot of what we have to say will be similar to what we did before for functions of two
variables, but there will also be some differences.
The rst thing that is the same (or hardly any different) is that the function is a well-dened
rule associated a single value f(x, y, z) to each point (x, y, z) in its domain. Usually, but not
necessarily, the functions we will consider will be specied by giving a formula to nd the
values. Examples could be
f(x, y, z) = x
2
+ y
2
+ z
2
f(x, y, z) = x cos(y + z) +
1
x
f(x
1
, x
2
, x
3
) =
x
1
x
2
x
3
_
x
2
1
+ x
2
2
+ x
2
3
+ 1
Functions of several variables 27
As with the earlier case of functions of two variables, a function f(x, y, z) will be dened for
points (x, y, z) in some domain, some subset of R
3
or for points all of R
3
if the domain is the
whole of R
3
. As with functions of two variables we will sometimes assume that the domain is as
big as it can be for the formula for f(x, y, z) to make sense. In the second of the three examples
above, we should exclude points with x = 0 from the domain as we dont want to be dividing by
zero. For the other two examples above we dont need any restrictions.
However, when we start to think of graphs, we nd a problem. A graph of a function of one
variable (a graph y = f(x)) we need to draw it in the plane (so we need two dimensions, one for
the domain and one more for the values). Similarly when we looked at functions of two variables
we needed 3 dimensions to t in the graph z = f(x, y).
A graph of a function f(x, y, z) would need 4 dimensions. It already takes 3 dimensional
space to keep track of the domain points (x, y, z) and we would need to nd a fourth dimension
to t in the values. So we cant look at it in reality. We can do it theoretically, but then we can
just work by analogy with what we are familiar with.
The level surface approach f(x, y, z) = c is available. It is harder perhaps to think about than
the level curves we had in the plane (for functions of two variables), but the level surfaces are in
space where we can see (or visualise) them.
For example in the case f(x, y, z) = x
2
+ y
2
+ z
2
, we can describe the level surfaces quite
easily. For c > 0, the equation f(x, y, z) = c means the distance squared from (x, y, z) to the
origin is c. So the points (x, y, z) satisfying f(x, y, z) = c are exactly the points on the sphere
of radius

c around the origin. [By the sphere we mean the surface. We generally use the word
ball for the inside of the sphere, or the surface together with the inside.] When c = 0 there is
only the one point (x, y, z) = (0, 0, 0) (the origin) on the level surface f(x, y, z) = c. We see
then that the level set is not always a surface but it typically is a surface if there are any points
at all on it. In this example, when c < 0 there are no points satisfying f(x, y, z) = c.
2.21 Derivatives of functions of 3 variables. Quite a number of the things we said about deriva-
tives of functions of two variables will also work (more or less in the same way) for functions of
3 variables. There are more variables and so more things to take care about, but most of the ideas
are quite similar.
Partial derivatives. The partial derivative of f(x, y, z) with respect to x is denoted
f
x

it is computed by differentiating f(x, y, z) with respect to x while treating y and z (the other
variables) as though they were constant. We could dene it more formally as
f
x
[
(x
0
,y
0
,z
0
)
=
d
dx
[
x=x
0
f(x, y
0
, z
0
) =
d
dt
[
t=0
f(x
0
+ t, y
0
, z
0
).
The other partial derivatives
f
y
and
f
z
work in a similar way.
There is some merit in the alternative notations for partial derivatives such as f
x
(x
0
, y
0
, z
0
)
for the value of f/x at the point (x
0
, y
0
, z
0
). We will generall try to stick to the f/x notation
even though it is maybe more clumsy at times.
Directional derivatives are again similar to the two variable case. For a unit vector u R
3
the directional derivative of f(x, y, z) at (x
0
, y
0
, z
0
) in the direction u is
D
u
f(x
0
, y
0
, z
0
) =
d
dt
[
t=0
f((x
0
, y
0
, z
0
) + tu)
28 200607 Mathematics 2E1 (Timoney)
and we compute it using partial derivatives
D
u
f(x
0
, y
0
, z
0
) =
_
f
x
[
(x
0
,y
0
,z
0
)
_
u
1
+
_
f
y
[
(x
0
,y
0
,z
0
)
_
u
2
+
_
f
z
[
(x
0
,y
0
,z
0
)
_
u
3
(where u
1
, u
2
and u
3
are the components of the unit vector u = (u
1
, u
2
, u
3
) = u
1
i +u
2
j +u
3
u
3
).
The justication of this formula for the directional derivative depends on the validity of the
linear approximation formula
f(x, y, z)

= f(x
0
, y
0
, z
0
) +
_
f
x
[
(x
0
,y
0
,z
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
,z
0
)
_
(y y
0
) +
_
f
z
[
(x
0
,y
0
,z
0
)
_
(z z
0
)
To be more accurate, what we need is that f(x, y, z) should have a derivative at the point
(x
0
, y
0
, z
0
), which means that there should be constants a
1
, a
2
and a
3
so that
lim
(x,y,z)(x
0
,y
0
,z
0
)
f(x, y, z) (f(x
0
, y
0
, z
0
) + a
1
(x x
0
) + a
2
(y y
0
) + a
3
(z z
0
)
dist((x, y, z), (x
0
, y
0
, z
0
))
= 0
As in the two variable case we will not go into the denition of a limit in space, but roughly
the limit above being 0 means that if (x, y, z) is close enough to (x
0
, y
0
, z
0
) but (x, y, z) ,=
(x
0
, y
0
, z
0
), then the fraction above is small.
If the function has a derivative, then the numbers a
1
, a
2
and a
3
have to be equal to the partial
derivatives f/y, f/y and f/z evaluated at (x
0
, y
0
, z
0
). Moreover the formula in the box
above for directional derivatives will be valid.
How do we know a function has a derivative? Well, as in the 2 variable case there is a theorem
that guarantees that most of the functions we can easily write down have derivatives (at least at
most points). The theorem says that if the 3 partial derivatives exist and are continuous functions
at (x
0
, y
0
, z
0
), then the function has a derivative. We wont go into continuous functions much,
but a function g(x, y, z) is called continuous at a point (x
0
, y
0
, z
0
) of its domain if
lim
(x,y,z)(x
0
,y
0
,z
0
)
g(x, y, z) = g(x
0
, y
0
, z
0
).
The function is called continuous (with no mention of a specic point) if it is continuous at each
point of its domain. There are theorems about limits of products, quotients and such combina-
tions of functions that allow one to be sure that most formulae dene continuous functions (apart
perhaps from problem points).
Then there is the gradient vector
f [
(x
0
,y
0
,z
0
)
=
_
f
x
[
(x
0
,y
0
,z
0
)
,
f
y
[
(x
0
,y
0
,z
0
)
,
f
z
[
(x
0
,y
0
,z
0
)
_
Functions of several variables 29
or
f =
_
f
x
,
f
y
,
f
z
_
The formula in the box above for the directional derivative can be restated with the gradient
vector
D
u
f(x
0
, y
0
, z
0
) =
_
f [
(x
0
,y
0
,z
0
)
_
u
So, as before in 2 variables, we can say that the directional derivative is
D
u
f(x
0
, y
0
, z
0
) =
_
_
f [
(x
0
,y
0
,z
0
)
_
_
cos
where is the angle between the gradient vector and the direction (unit vector) u. From this we
can say that the directional derivative is largest when u is the unit vector in the direction of the
gradient, that is when
u =
1
_
_
f [
(x
0
,y
0
,z
0
)
_
_
f [
(x
0
,y
0
,z
0
)
.
And the largest possible value of the directional derivative is the magnitude of the gradient vector.
The smallest possible directional derivative occurs when u is in the exact opposite direction to
the gradient vector ( = and cos = 1) and this smallest directional derivative is minus the
magnitude of the gradient vector.
Another way to put that is that the gradient vector is a vector in the direction of steepest in-
crease of the function and the length of the gradient vector is the rate of increase in that direction.
An application of this that arises fairly frequently is when the function represents potential
energy. If a charged particle is subject to an electric eld it would have a potential energy
V (x, y, z) that depends on its position in the eld. If left unchecked by any other inuence,
the particle would tend to go in the direction of steepest decrease of potential energy. That
means that the force acting on the particle would be F = V (and would depend on position
(x, y, z)). The same idea would apply if the particle has some mass and is in a gravitational eld
(caused by the earth, the moon or some other body). The gravitational force acting on it would be
minus the gradient (thus in the direction of free fall). If a gas like the atmosphere had a variable
pressure then the gas will tend to ll in the low pressure areas. So it will be inuenced to go
towards the low pressure areas in the direction of fastest decrease of pressure.
A familiar example of this is evident in the weather charts we see every day. The curves of
constant pressure (isobars) are shown, but the wind will be almost perpendicular to those lines.
It will maybe not quite be perpendicular as there is also an inuence of other forces like coreolis
forces arising from the rotation of the earth. But this is a seemingly two dimensional example
because the weather charts are two dimensional. In fact the atmosphere itself is three dimensional
and the air can move vertically as well as horizontally (though this is more likely to be noticed
in a plane at higher altitudes than the ground).
An important thing to mention is the property that the gradient is perpendicular to the level
surface. This ts with the fact that the directional derivative is zero in the directions perpendicular
to the level surface.
30 200607 Mathematics 2E1 (Timoney)
We will go on to give some examples of this, but we wont really explain it further. A
reasonably good explanation could be given by considering curves p(t) = (x(t), y(t), z(t)) that
lie entirely in a level surface f(x, y, z) = c and satisfy p(0) = (x
0
, y
0
, z
0
). Then differentiate the
equation
f(x(t), y(t), z(t)) = c
with respect to t using the chain rule. This is similar to an explanation we outlined in 2.17 above
for level curves of functions of two variables. A difference is that when we are dealing with a
level surface there are many possible curves on the surface that pass through the point (x
0
, y
0
, z
0
)
and we should look at all of them to see all possible tangent directions. (For each direction in the
tangent plane to the surface you can expect to nd a curve in the surface that is tangent to that
direction as it passes (x
0
, y
0
, z
0
).)
We did not say yet what the chain rule is in this case, but it is like the two variable version
with one extra term:
d
dt
f(x(t), y(t), z(t)) =
f
x
dx
dt
+
f
y
dy
dt
+
f
z
dz
dt
2.22 Examples. (i) Find the equation of the tangent plane to the surface
x
4
+ y
2
+ z
6
= 26
at the point (2, 3, 1).
We can see that we are dealing with a level surface to the function f(x, y, z) = x
4
+y
2
+z
6
,
the level surface f(x, y, z) = c where c = 26. According to what we said above the normal
vector to the level surface is f [
(2,3,1)
and so we can set about computing that gradient
vector.
f
x
= 4x
3
f
y
= 2y
f
z
= 6z
5
Evaluating these at the point (2, 3, 1) we get our normal vector
f [
(2,3,1)
= (32, 6, 6)
So the equation of the plane through (2, 3, 1) perpendicular to that is
32(x 2) + 6(y 3) + 6(z 1) = 0
or 32x + 6y + 6z = 88.
Functions of several variables 31
(ii) We can treat a graph z = f(x, y) of a function f(x, y) of 2 variables as a level surface
g(x, y, z) = 0 of the function g(x, y, z) = z f(x, y) of 3 variables.
This allows us to calculate the tangent plane to the graph in a different way to the way we
did previously. Say we want the tangent plane at the point in the graph where (x, y) =
(x
0
, y
0
). We will need to notice before we nish that the z coordinate at that point on the
graph is z
0
= f(x
0
, y
0
). We have
g =
_
g
x
,
g
y
,
g
z
_
=
_

f
x
,
f
y
, 1
_
and we should evaluate this at (x
0
, y
0
, z
0
) to get the normal vector to the tangent plane. This
gives the equation to the tangent plane to the graph is
_

f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_

f
y
[
(x
0
,y
0
)
_
(y y
0
) + (z z
0
) = 0.
We can rearrange that as
z = z
0
+
_
f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
)
_
(y y
0
)
and this should look familiar.
2.23 Higher order partial derivatives. Now that we know about partial derivatives we can
observe that it is possible to take partial derivatives of partial derivatives. (This is like taking
second derivatives of functions of one variable.) For example, say f(x, y, z) = x
2
y +yz +5xyz
We can compute
f
x
= 2xy + 5yz

z
_
f
x
_
= 5y
f
z
= 5xy

x
_
f
z
_
= 5y
We use the notation

2
f
xz
for

x
_
f
z
_
(or sometimes the notation f
zx
). We can do something similar for third order partials (partials of
second order partials) like

3
f
yxz
=

y
_

2
f
xz
_
.
32 200607 Mathematics 2E1 (Timoney)
There is really just one important thing about higher order partials, apart from what they are.
That is that the order (usually) does not matter. So for instance

2
f
xz
=

2
f
zx
You will see that in the example where we computed them we did get the same answer, and there
is a theorem that says we will usually get the same answer independent of the order we take the
partials. Of course it will matter which partials you take and how many times you take the partial
with respect to any give variable.
To state the theorem a little carefully (and explain what the word usually means), the the-
orem says that if all the second order partial derivatives of (say) f(x, y, z) are continuous func-
tions, then the order does not matter. So if all of these are continuous

x
_
f
x
_
,

y
_
f
y
_
,

z
_
f
z
_
,

x
_
f
y
_
,

y
_
f
x
_
, . . .
then the mixed partials dont depend on the order:

x
_
f
y
_
=

y
_
f
x
_
,

x
_
f
z
_
=

z
_
f
x
_
,

y
_
f
y
_
=

z
_
f
x
_
Typically we can be condent that the theorem about mixed partials will apply when we are
dealing with functions given by a formula. The second order partial derivatives will be easy
enough to work out using the various rules for differentiation, but even without working them
out we know we will end up with a formula for the partials and the second order partials. As
most formulae give continuous functions, we should be able to apply the theorem. (The only
catch might be cases where there are some bad points for the formulae we get for the second
order partial derivatives, bad points that were not really bad for the function we began with.)
We use the notation

2
f
y
2
for what we get by taking the partial derivative with respect to y two times, and similarly for the
other variables.
2.24 Maxima and minima (general remarks). There are quite a few practical reasons to want
to nd the maximum or minimum value of something. For example to get the most power out of
some machine or do build something at the least possible cost.
Typically there are parameters to be adjusted, changed or designed into things, and the end
result may depend on the choices of the parameters in some rather complicated way. But, from
a mathematical points of view, we can assume that the way the end result depends on the pa-
rameters has been gured out somehow and we are then presented with a problem of nding
the largest (or smallest) value of some function of the parameters. Typically there will be some
restriction on the feasible choices of the parameters, but as long as there is any scope to vary the
parameters we have something to do to nd the optimum choice of the parameters.
Functions of several variables 33
Now typically there may be very many variables. For example an engine management sys-
tem would have various things it could control like the air/fuel ratio, the ignition timing and the
optimum might depend on the ambient conditions such as the load, engine revs, engine oil tem-
perature, outside temperature and humidity, speed, acceleration and other things. An aeroplane
has many more things to manage (aps, rudder, thrust on several engines, climate control in the
cabin and so on). A manufacturing process may have many different choices.
These real world problems are potentially very complicated and would require a lot of back-
ground. Just to illustrate the ideas that would be needed to solve such a problem we examine
the case where there are either 2 or three parameters. So we consider the question of nding the
largest value of f(x, y) for (x, y) in some subset of R
2
. (Or alternatively to nd the largest value
of f(x, y, z) for (x, y, z) in some subset of R
3
.)
Last year you found out about how calculus helps to deal with similar problems where we
have only one variable, that is nding the largest value of a function f(x). It may then not be
surprising that calculus of several variables (partial derivatives) will be relevant when we have
more variables.
To begin with we state a theorem that says that some such problems have a solution. Perhaps
it should be obvious in a practical problem that there is bound to be an optimum. At least it
is obvious if the problem is sensible. Mathematically we can have a theoretical problem that
may or may not have a solution and the theorem can help then. Also, even if our problem is
supposed to come forma practical problemwhere we are sure there should be a solution, having a
mathematical reason that there is an optimum can provide a check that our mathematical problem
is a realistic reection of the true problem.
To state the theorem we need to talk about special properties the domain of our function may
have called closed and bounded. In language that maybe seems more precise than it is, a set in
R
2
or R
3
is called a closed set if it includes all of its boundary points or edge points. We could
help explain this with some fairly typical examples that are too complicated. The rst quadrant
S = (x, y) R
2
: x > 0 and y > 0
without the bounding axes included is not closed, but if we include the edges we do get a closed
set. So
T = (x, y) R
2
: x 0 and y 0
is closed. Here are some pictures (where the axes are supposed to be dotted lines to indicate
they are not included on the left set, while the axes are solid on the right to indicate that they are
included).
34 200607 Mathematics 2E1 (Timoney)
Another example is an disc without its circumference included (which is not closed)
S = (x, y) R
2
: (x 1)
1
+ (y 2)
2
< 5
and the version including the circle, which we would call the closed disc and is a closed set:
T = (x, y) R
2
: (x 1)
1
+ (y 2)
2
5
(These discs have centre (1, 2) and radius 5. You might like to sketch them using a dotted circle
with the inside shaded to show S and a solid circle with the inside shaded to draw T.)
In general then a set in the plane R
2
is closed if it includes all its boundary curve. We can
also make similar examples in space such as the positive orthant
T = (x, y, z) R
3
: x 0, y 0 and z 0.
As we are including the bounding quarter planes in this set, T is closed. We could also look at
closed balls (the inside of a sphere plus its bounding sphere) as examples of closed sets.
Next we explain what bounded sets are. A set S R
2
is called bounded if it does not
stretch innitely far away from the origin. To make it more precise we say that S is bounded if
there is a nite radius R > 0 so that S is contained in the disc of radius R around the origin.
A subset S of R
3
is called bounded if there is a ball of nite radius that contains S.
So the quadrants are not bounded, but the discs are.
The promised theorem that guarantees the existence of largest and smallest values is this:
Theorem. If S is a closed and bounded set in R
2
or R
3
and f : S Ris a continuous
function, then f has a largest and a smallest value in S.
To spell it out more carefully, there are points p
m
S and p
M
S so that
f(p
m
) f(x) f(p
M
) holds for all x S.
The point x = p
M
is then a point where the value of f is greatest. Or to be even more careful,
there is no other point in S where f has a strictly larger value.
Functions of several variables 35
This theorem may not seem like much, but you can think of it like this. Finding the largest
value of a function can be like nding a needle in a haystack in the sense that it can be very
difcult. But knowing that there is denitely a needle to be found in the haystack is a help. If
you just search furiously till you get tired and nd nothing, you dont know whether there is any
needle or not.
Actually, in our case we can use the information that there is a maximum (or minimum),
together with a method of narrowing down the possible points where the maximum could be, to
provide a way to nd the maximum. This is where calculus comes in.
2.25 Critical Points. For a function f(x) of one variable, a critical point is a point where f

(x) =
0. For a function f(x, y) of two variables a critical point is a point (x
0
, y
0
) where both partial
derivatives are zero:
f
x
= 0 and
f
y
= 0.
For a function f(x, y, z) of three variables a critical point is a point (x
0
, y
0
, z
0
) where all
three partial derivatives are zero:
f
x
= 0,
f
y
= 0 and
f
z
= 0.
The relationship of critical points to nding maximum points (or minimum points) is that
max and min points have to be critical points if they are not at the boundary of the domain.
We call points interior points of a set if are not at the boundary of the set. (Expressing it more
carefully a point (x
0
, y
0
) S R
2
is called an interior point of S if all (x, y) within some
positive distance of (x
0
, y
0
) are in S.
The result is that if a maximum occurs at an interior point of the domain (and the function
in question has partial derivatives) then the maximum occurs at a critical point. [This result is
actually quite easy to prove based on the analogous result in one variable. When we take partial
derivatives we are differentiating only in one variable at a time.]
Since the number of equations to be satised (simultaneously) at a critical point is the same
as the number of variables (2 equations for the case f(x, y), and 3 equations for the case of
f(x, y, z)) typically the number of critical points will be quite small. Usually a nite number
anyhow. So knowing that the maximum is at a critical points reduces the number of points to be
looked at in a dramatic way.
There is a problem, though. We saw just above that a continuous function on a closed and
bounded set will have a largest value but there is a problem. The largest value may not be at an
interior point it could also be on the boundary. What we have so far is a way of narrowing
down the location of the maximum if it is in the interior, not on the boundary. Well need a
similar method to deal with the boundary case. That will come a little later and is called Lagrange
multipliers.
First we will look at an example where there is no boundary
2.26 Example. (Least square t example.) Assume an experiment that is supposed to produce
points on a straight line y = mx + c produces instead a number of data points
(x
1
, y
1
), (x
2
, y
2
), . . . , (x
n
, y
n
)
36 200607 Mathematics 2E1 (Timoney)
In the real world, except my the purest uke, if there are more than 2 points, there will not be any
one line that goes through all the data points. The question is to nd the line that best ts the
data. One interpretation of best is the line that minimizes the sum of the squares of the vertical
discrepancies of the points from the line. So we choose the slope m and the y-intercept c so as
to minimize
f(m, c) =
n

i=1
(y
i
(mx
i
+ c))
2
(The picture is supposed to illustrate what we are trying to do. The blobs are the data points and
we are trying to choose the line to minimise the sum of the squares of the vertical distances from
the points to the line.)
As far as we know now, the right choice of m and c could be any numbers. So we cannot
use the theorem guaranteeing that there is an answer to our problem. (The set of possible points
(m, c) appears to be all of R
2
and this is not a bounded set of possibilities.) So we will instead
take the view that this is a sensible problem and that there is bound to be an answer.
If this is true, then the minimum will certainly be at an interior point of the allowable (m, c).
Hence it will be at a critical point, a point where both partial derivatives f/m and f/c are
Functions of several variables 37
0.
2
We now calculate the partial derivatives and set them equal to 0.
f
m
=
n

i=1
2(y
i
(mx
i
+ c))(x
i
)
= 2
n

i=1
(x
i
y
i
+ mx
2
i
+ cx
i
)
f
c
=
n

i=1
2(y
i
(mx
i
+ c))(1)
= 2
n

i=1
(y
i
+ mx
i
+ c)
= 2
_
n

i=1
(y
i
+ mx
i
) +
n

i=1
c
_
= 2
_
n

i=1
(y
i
+ mx
i
) + nc
_
From the equation
f
c
= 0 we nd
c =
1
n
n

i=1
(y
i
+ mx
i
)
or
c =
1
n
_
n

i=1
y
i
m
n

i=1
x
i
_
and when we put that into the equation
f
m
= 0 we get an equation for m only:
n

i=1
(x
i
y
i
+ mx
2
i
+ cx
i
) = 0

i=1
(x
i
y
i
) + m
n

i=1
x
2
i
+ c
n

i=1
x
i
= 0

i=1
(x
i
y
i
) + m
n

i=1
x
2
i
+
1
n
_
n

i=1
y
i
m
n

i=1
x
i
_
n

i=1
x
i
= 0

i=1
(x
i
y
i
) + m
n

i=1
x
2
i
+
1
n
_
n

i=1
y
i
__
n

i=1
x
i
_

m
n
_
n

i=1
x
i
_
2
= 0
2
It is perhaps messy that the variables for our problem are called m and c rather than x and y, but on the other
hand we are used to using m for a slope and c for a y-intercept.
38 200607 Mathematics 2E1 (Timoney)
If we collect all the terms with m to the left and put the remainder to the right we get
m
_
_
n

i=1
x
2
i

1
n
_
n

i=1
x
i
_
2
_
_
=
n

i=1
(x
i
y
i
)
1
n
_
n

i=1
x
i
__
n

i=1
y
i
_
This leads to the only choice for m as
m =

n
i=1
(x
i
y
i
)
1
n
(

n
i=1
x
i
) (

n
i=1
y
i
)

n
i=1
x
2
i

1
n
(

n
i=1
x
i
)
2
and we already had c in terms of m and the data points:
c =
1
n
_
n

i=1
y
i
m
n

i=1
x
i
_
When you apply the methid of least squares to nd the best t line to your data points, what
you need is these two formulae. The application of our methods from several variables calculus
is one way to arrive at these formulae, but there are others. (This method is given in an exercise
in Thomas in section 14.7 of the 11th edition.)
2.27 Lagrange multipliers. We now give the promised method of Lagrange multipliers. It is
aimed at nding the maximum or minimum values of a function f(x, y) subject to a constraint
g(x, y) = 0.
In fact it does not really do that directly. What it does is provide an analogue of a critical point
and narrows down the possible max and minimum points to those that solve several equations.
Typically these equations have just a few solutions, and so we are left with only a few candidates
for the max and min points.
It is not true any longer that the maximum points have to be at a critical point of f. That
result relied on being at an interior point of the allowed (x, y). Here we are restricted to points
on the curve g(x, y) = 0 and there will be no interior points of the curve in the plane.
What can say instead is that at the maximum (or minimum) point the gradient of f will be a
multiple of the gradient of g. We usually call this multiple by the name (Greek letter lambda).
The way to say that in equations is
f = g.
As this is a vector equation
_
f
x
,
f
y
_
=
_
g
x
,
g
y
_
it comes down to two scalar equations
f
x
=
g
x
f
y
=
g
y
Functions of several variables 39
We now have 3 unknowns x, y and but we also have 3 equations that they should satisfy
because we have g(x, y)) = 0 as well as the two equations above. So we typically get just a few
solutions.
That was Lagrange multipliers in 2 dimensions. For 3 dimensions we have a similar thing.
The maximum value of a function f(x, y, z) of 3 variables subject to a constraint g(x, y, z) = 0
will occur at a place where there is a value of that satises
f
x
=
g
x
f
y
=
g
y
f
z
=
g
z
g(x, y, z) = 0
Now we have 4 equations for the 4 unknowns x, y, z and the Lagrange multiplier .
2.28 Example. Find the largest and smallest values of f(x, y) = xy on the ellipse
x
2
3
+
y
2
5
= 1
Technically the ellipse (just the curve without the inside) is a closed set because it includes
all of its boundary. (In fact it is its own boundary, which is a bit confusing perhaps.) The ellipse
is also a bounded set and the function f(x, y) is continuous (everywhere and so on that set). That
means there must be a solution to our problem both a largest and a smallest value of f(x, y)
on the ellipse must exist.
To narrow down where the maximum and minimum occurs, we use Lagrange multipliers. We
take g(x, y) = x
2
/3 +y
2
/5 1 so that the ellipse is g(x, y) = 0. The maximum (and minimum)
must occur at points that solve
f = g and g(x, y) = 0
for some . We work out what f = g says and we get
(y, x) = (2x/3, 2y/5)
So y = 2x/3 and x = 2y/5. If we put the rst of these in the second we get
x =
4
2
15
x
or
x
_
1
4
2
15
_
= 0
Thus there are two cases to look at x = 0 or
2
= 15/4 (which means =

15/2).
40 200607 Mathematics 2E1 (Timoney)
If x = 0, then we must have g(x, y) = 0 and that leaves us with y
2
/5 = 1, or y =

5.
Then (y, x) = (2x/3, 2y/5) comes to (y, 0) = (0, 2y/3) and there is no that can solve this.
So x = 0 is not an option.
If =

15/2, then y = 2x/3 = (

15/3)x = (
_
5/3)x. So then g(x, y) = 0 gives
x
2
3
+
5x
2
15
= 1
x
2
_
1
3
+
1
3
_
= 1
x
2
=
3
2
x =
_
3
2
and y = 2x/3 = (
_
5/3)
_
3/2 =
_
5/2.
We are left with 4 possible points (
_
3/2,
_
5/2), (
_
3/2,
_
5/2), (
_
3/2,
_
5/2) and
(
_
3/2,
_
5/2). The values of f(x, y) = xy at these 4 points are

15/2,

15/2,

15/2
and

15/2. Thus the largest possible value has to be

15/2 (which occurs at the two points
(
_
3/2,
_
5/2) and (
_
3/2,
_
5/2)). The smallest value is

15/2 (and it occurs at the

points (
_
3/2,
_
5/2)).
2.29 Second derivative test. We now turn to the generalisation to functions of two variables of
the familiar second derivative test concerning functions of one variable. To get ourselves straight
let us recall that one variable test. Weve mentioned above that a maximum point of a function
f(x) at an interior point c of the domain will be a critical point (that is a point where f

(c) = 0).
However, you may recall that this is true not just of absolute maximum points but also of local
maximum points (which are ofter called relative maximum points. Similarly local minimum
points are critical points if they are not at an end point.
Suppose x = c is a critical point of f(x). The second derivative test says
If f

(c) > 0, then x = c is a local minimum point;

If f

(c) < 0, then x = c is a local maximum point;

But if f

(c) = 0, then no conclusion can be reached without further information.

When we go to 2 variables, we still have the idea of local maximum and local minimum
points. If you think of a graph z = f(x, y) as a landscape, then a local maximum point (x
0
, y
0
)
corresponds to the top of a hill on the graph. It also means f(x
0
, y
0
) f(x, y) holds for (x, y)
within some(possibly small) distance of (x
0
, y
0
). A local maximum point does not have to cor-
respond to the top of the highest mountain that would correspond to the overall largest value
and we can call this an absolute maximum point.
We can also describe local minimum points as corresponding to the bottom of a valley on the
graph. What we can say is that both local max points and local min points will be critical points
Functions of several variables 41
of f(x, y) if they are at interior points of the domain. So it is not only absolute max and absolute
min points, but also local max and local min points that must be critical points. In one variable
there are critical points that are neither (local) max nor min points and there are more critical
points where the second derivative test does not tell you what is going on.
In two variables, there are not just peaks and valleys on the graph that are critical points, there
are also mountain pass points. If you think of a road across a mountain range, it will usually go
over a pass (the lowest point where you can cross the mountain range). As you drive across the
pass, there will be a highest point you reach on the road, but as you look to the left and right the
mountains will rise on either side of you. So if you imagine that the mountain range is stretching
from north to south and the road is heading due east, you see that along the easterly road you
are at your highest point when you cross the pass, but if you imagine walking in the northerly
direction perpendicular to the road you would be at your lowest point as you cross the road. This
ensure that both the partial derivatives with respect to the x and y directions are 0. In the x
direction (the east direction) we think we are at a local maximum, but in the northerly direction
(the y direction) we think we are at a local minimum. If we look all around us we see that we are
neither at the top of a hill nor at the bottom of a valley. Nevertheless we are at a critical point.
These critical points are called saddle points (though maybe they could be called mountain
pass points) because if you think of a horse saddle you sit down into it when you think about the
front to back direction (so it looks like a local minimum in that direction), but you are astride
the horse with your legs down each side (so looking only from left to right you are at a local
maximum).
The second derivative test in two variables applies to critical points (x
0
, y
0
) where f/x =
f/y = 0 and one way to explain it is to talk in terms of the matrix of second order partial
derivatives
_

2
f
x
2

2
f
xy

2
f
xy

2
f
y
2
_
evaluated at the critical point. We want to talk about the matrix being strictly positive, or
strictly negative and the way we express this is in terms of both eigenvalues being positive or
both being negative.
If both eigenvalues are strictly positive then (x
0
, y
0
) is a local minimum point;
If both eigenvalues are strictly negative then (x
0
, y
0
) is a local maximum point;
If one eigenvalue is strictly positive and the other is strictly negative then (x
0
, y
0
) is a
saddle point;
But if either eigenvalue is zero then no conclusion can be reached without further informa-
tion.
Rather than computing eigenvalues we can decribe this using the determinant of the matrix
=
_

2
f
x
2

2
f
y
2

_

2
f
xy
_
2
_
[
(x
0
,y
0
)
= f
xx
(x
0
, y
0
)f
yy
(x
0
, y
0
) (f
xy
(x
0
, y
0
))
2
42 200607 Mathematics 2E1 (Timoney)
using the fact that the determinant is the product of the eigenvalues. (We see then that we should
have > 0 in the case where both eigenvalues are strictly positive and in the case where both
eigenvalues are strictly negative. From < 0 we nd that the two eigenvalues have opposite
signs.
Suppose (x
0
, y
0
) is a critical point of f(x, y) and is as above.
If > 0 and

2
f
x
2
[
(x
0
,y
0
)
> 0 then (x
0
, y
0
) is a local minimum point;
If > 0 and

2
f
x
2
[
(x
0
,y
0
)
< 0 then (x
0
, y
0
) is a local maximum point;
If < 0 then (x
0
, y
0
) is a saddle point;
But if = 0 then no conclusion can be reached without further information.
2.30 Taylors theorem. We turn now to Taylors theorem in two variables, which can be used to
prove that the second derivative test is right.
Recall for a moment what Taylors theorem says for functions of one variable. We can begin
by saying that it makes precise the linear approximation formula
f(x)

= f(x
0
) + f

(x
0
)(x x
0
) (for x near x
0
).
Or at least it provides a way of guring out how accurate or inaccurate the formula is in a given
case (where the function f and the point x
0
are specied). The theorem says
f(x) = f(x
0
) + f

(x
0
)(x x
0
) +
1
2!
f

(c)(x x
0
)
2
for some c between x
0
and x (provided f

(x) exists for x between x

0
and x).
In fact this is just a baby version of Taylors theorem, dealing with the remainder term
in linear approximation. There are also versions dealing with quadratic approcimation, cubic
approximation, and so on. In general
f(x) = f(x
0
) + f

(x
0
)(x x
0
) +
1
2!
f

(x
0
)(x x
0
)
2
+ +
1
n!
f
(n)
(x
0
)(x x
0
)
n
+
1
(n + 1)!
f
(n+1)
(c)(x x
0
)
n+1
(where as before c is between x
0
and x).
We will content ourselves with stating the two variable version of the baby theorem though
there are versions of Taylors theorem for functions of any number of variables and up to any
degree n. The notation gets a bit complicated and that is why we stick to the simplest case.
Recall the linear approximation formula
f(x, y)

= f(x
0
, y
0
) +
_
f
x
[
(x
0
,y
0
)
_
(x x
0
) +
_
f
y
[
(x
0
,y
0
)
_
(y y
0
)
Functions of several variables 43
which we rewrite using the subscript notation for derivatives:
f(x, y)

= f(x
0
, y
0
) + f
x
(x
0
, y
0
)(x x
0
) + f
y
(x
0
, y
0
)(y y
0
)
(valid for (x, y) close to (x
0
, y
0
)).
The quadratic approximation formula would be
f(x, y)

= f(x
0
, y
0
) + f
x
(x
0
, y
0
)(x x
0
) + f
y
(x
0
, y
0
)(y y
0
)
+
1
2!
_
f
xx
(x
0
, y
0
)(x x
0
)
2
+ f
yy
(x
0
, y
0
)(y y
0
)
2
+ 2f
xy
(x
0
, y
0
)(x x
0
)(y y
0
)
_
and Taylors theorem gives an equation, rather than an approximation, where the second order
partials are evaluated at some point c on the line joining (x
0
, y
0
) and (x, y):
f(x, y) = f(x
0
, y
0
) + f
x
(x
0
, y
0
)(x x
0
) + f
y
(x
0
, y
0
)(y y
0
)
+
1
2!
_
f
xx
(c)(x x
0
)
2
+ f
yy
(c)(y y
0
)
2
+ 2f
xy
(c)(x x
0
)(y y
0
)
_
One way to use this is to estimate the size of the quadratic terms in the above formula (with-
out being able to know c exactly) so as to get a worse case scenario for the error in the linear
approximation formula.
Another way is to prove that the second derivative test works. If (x
0
, y
0
) is a critical point,
then there are no linear terms (or the coefcients f
x
(x
0
, y
0
) = f
y
(x
0
, y
0
) = 0). If we can be
sure that the quadratic terms denitely add to something positive then we can conclude that
f(x, y) f(x
0
, y
0
). The assumption for the local min case of the second derivative test makes
it possible to deduce that the quadratic terms add to something positive if c is close enough to
(x
0
, y
0
). This is guranteed if (x, y) is close to (x
0
, y
0
). Thats roughly the idea anyhow.
Richard M. Timoney (January 9, 2007)