HANDOUT FIVE: VECTOR FIELDS

PETE L. CLARK
1. Introduction to Vector Fields in the plane and in space
We have already studied several kinds of functions of several variables: vector
valued functions of a scalar variable i.e., parameterized curves in the plane and
in space; scalar-valued functions of two or more variables (the case of two variables
z = f(x, y) giving the graph of a surface in space; and functions from the plane
to space, parameterized surfaces. Here we consider functions V from the plane to
itself and from space to itself, which are called vector elds.
A vector eld in the plane is given by a pair of functions of two variables, V (x, y) =
(P(x, y), Q(x, y). We picture it as follows: at each point (x
0
, y
0
) in the plane, we
get a vector (P(x
0
, y
0
), Q(x
0
, y
0
)) at that point. So in all every point in the plane
(or every point in a certain region of the plane) has a vector attached to it: overall
we get a eld of vectors.
A similar story holds in space: a vector eld gives a vector at every point of
space, so to describe it we need three functions of x, y and z: V (x, y, z) =
(P(x, y, z), Q(x, y, z, ), R(x, y, z)).
Example:
F(x, y, z) =
(x, y, z)
||(x, y, z)||
3
=

x
(x
2
+y
2
+z
2
)
3/2
,
y
(x
2
+y
2
+z
2
)
3/2
,
z
(x
2
+y
2
+z
2
)
3/2

.
This vector eld gives, up to a multiplicative constant, the force felt by a particle at
any point due to gravitational attraction to a mass centered at the origin, i.e., it is
a reformulation of Newtons inverse square law. That is, at any point, the vector
at that point points in the direction of the origin, and its magnitude is inversely
proportional to the square of the distance from that point to the origin
1
. Note also
that the vector eld is not dened at the origin, nor could it be extended to the
origin in a continuous manner, for two reasons: one the one hand the magnitude
of the vector eld approaches innity, and on the other hand therer are vectors in
the eld arbitrarily close to the origin pointing in every direction, so the direction
of the vector eld cannot be continuously extended to the origin either. In such
a situation namely, when there is a point P
0
= (x
0
, y
0
, z
0
) at which a vector
eld is not dened and could not be dened in a continuous way P
0
is called a
singularity of the vector eld.
We saw on p. 13 of handout four that if f(x, y, z) = (x
2
+ y
2
+ z
2
)
1/2
, then
1
Notice that the norm of the numerator is || (x, y, z)|| = d, the distance to the origin, so the
norm of the entire expression is
d
d
3
=
1
d
2
.
1
2 PETE L. CLARK
V = (f). We will see in the next unit that the existence of this function f leads
to a law of conservation of energy for a particle travelling through a gravita-
tional eld: at any point, the sum of the kinetic energy 1/2mv
2
and the potential
energy f is a constant.
Other examples of vector elds modelling elds are V
2
= (0, 0, mg), the constant
downward vector eld modelling the constant force due to gravity of a particle close
to the surface of the earth, and V
3
=
(x,y,z)
||(x,y,z)||
3
the same as Newtons law except
without a minus sign, which models the situation in which we have a positive
charge at the origin and want the force felt by a positively charged particle at a
given point of space: this is a repulsive force, and that it also is inversely propor-
tional to the square of the distance is known as Coulombs Law.
On the other hand, when trying to picture a vector eld in the plane or in space,
it is convenient to have a dierent physical interpretation in mind: we think of V
as the velocity eld of some sort of uid: that is, at any point P
0
, the eld V (P
0
)
gives the velocity vector for the ow: i.e., it tells the uid which way to go and how
fast.
Viewing a vector eld V (x, y) or V (x, y, z) as a velocity eld sets up a fundamen-
tal geometric problem: given a velocity eld, place a particle at a certain initial
point P
0
and release it: what is its trajectory? That is, we want to nd a curve
r(t) = (x(t), y(t), z(t)) such that r(0) = P
0
and with the property that for all t
(1) r

(t) = v(t) = V (r(t)).

Such a curve is called an integral curve and Equation (1) is called the ow equa-
tion: it is a system of ordinary dierential equations in the variables x, y and z:
see Section 11.1 of your text.
We now give some examples of vector elds in the plane.
Example 1: V (x, y) = (x, y). That is, at any point in the plane, the vector at
that point is that point; otherwise put, it points in the same direction (i.e., away
from the origin) and has the same magnitude. Suppose we interpret this eld as
a velocity eld and want to nd the trajectories. As we just said that at every
point a particle is pushed radially outward, the trajectories should be straight lines
emanating outward from the origin. In terms of equations we have
x

(t) = P(x, y) = x.
y

(t) = Q(x, y) = y.
Here (quite luckily) each of the two equations involves only one variable at a time,
so we just need to solve the dierential equation df/dt = f both times: the solution
to this is f(t) = Ce
t
, so x(t) = C
1
e
t
, y(t) = C
2
e
t
. Indeed x(0) = C
1
e
0
= C
1
and y(0) = C
2
e
0
= C
2
, so the integral curve passing through the point (x
0
, y
0
)
is r(t) = (x
0
e
t
, y
0
e
t
). This is a straight line since y/x =
y0e
t
x0e
t
= y
0
/x
0
. Note the
special point (x
0
, y
0
) = (0, 0), at which the vector eld is zero. If we start at a
point and the velocity vector is zero, then we never leave! For this reason a point
(x
0
, y
0
) where F(x
0
, y
0
) = (0, 0) is often called a stationary point.
HANDOUT FIVE: VECTOR FIELDS 3
Example 2: V (x, y) = (x, y). This is exactly the opposite vector eld. Ge-
ometrically, at any nonzero point we are getting pushed back towards the origin, so
we expect the integral curves to be straight lines converging to the origin. Indeed,
the system is (x

, y

) = (x, y) which boils down to dx/dt = x, dy/dt = y,

with solutions x(t) = C
1
e
t
, y(t) = C
2
e
t
, and again we have C
1
= x
0
, C
2
= y
0
and y(t)/x(t) = y
0
/x
0
a constant. But this time
lim
t
r(t) = lim
t
(x
0
e
t
, y
0
e
t
) = (0, 0),
so all the trajectories approach the origin (the one that starts at the origin doesnt
go anywhere).
Example 3: V
L
(x, y) = (y, x). Notice that (y, x) (x, y) = yx + xy = 0,
so the velocity vector at a point P
0
= (x
0
, y
0
) is perpendicular to the line segment
OP
0
indeed, testing the point (1, 0) V (1, 0) = (0, 1), it is 90 degrees to the
left of OP
0
. It thus seems plausible that the trajectories will be counterclockwise
circles centered at the origin. This time the system of dierential equations we get
is
x

(t) = y(t), y

(t) = x(t)
so that x = y

= (x

= x

, and also y

= y. A solution to this is
x(t) = Rcos t, y(t) = Rsin t. So the trajectories are counterclockwise circles of
radius R.
Example 4: V
R
(x, y) = (y, x). Again (y, x) (x, y) = 0 so the velocity vec-
tor at a point P
0
is perpendicular to the radial vector OP
0
, but this time (1, 0)
V (1, 0) = (0, 1), which turns 90 degrees to the right. Going back through the
previous argument, we see that indeed the trajectories are clockwise circles.
Example 5: Put r =

x
2
+y
2
, and consider the vector eld V
a
(x, y) = (
y
r
a
,
x
r
a
),
where a is some constant. Since V
a
(x, y) = V
L
(x, y)/r
a
, we are just rescaling the
vector eld of Example 3 so that at each point the vector eld points 90 degrees
to the left of the radial vector OP
0
but has norm ||(y, x)/r
a
|| = ||(y, x)||/r
a
=
r/r
a
= r
1a
.
2. Divergence and Curl
Let V (x, y, z) = (P(x, y, z), Q(x, y, z), R(x, y, z)) be a vector eld in space. We
view a vector eld F(x, y) in the plane as a special case of this, with R 0. Here
are two dierent ways to, roughly speaking, take a derivative of a vector eld:
The divergence:
Div(V ) = V = (

x
,

y
,

z
) (P, Q, R) =
P
x
+
Q
y
+
R
z
.
The curl:
curl(V ) = V =

R
y

Q
z
,
P
z

R
x
,
Q
x

P
y

.
Note well that the divergence of a vector eld is a scalar-valued function of three
variables, whereas the curl of a vector eld is another vector eld.
4 PETE L. CLARK
Clearly it is no harder to compute the divergence or the curl of a vector eld than
the gradient of a function: either way were just taking some partial derivatives.
On the other hand, the gradient has a useful geometric interpretation as the path
of steepest ascent. It would be nice to have some similar geometric and/or physical
intuition for the divergence and the curl. Lets try to see what they are by looking
at some of our examples from the last section.
Before we do this, however, we note the special case of the curl of a planar vector
eld V (x, y) = (P(x, y), Q(x, y), 0): the only terms in the denition of the curl in
which we are not either dierentiating the z-component R which is zero, or with
respect to z which is 0 is the last, so we get
curl(P(x, y), Q(x, y), 0) = (0, 0,
Q
x

P
y
) = (
Q
x

P
y
)

k.
Thus, although the curl of a planar vector eld is technically a vector-valued func-
tion, it always points in the same direction, so it is useful to think of the scalar
curl
Q
x

P
y
and remember that its direction is always perpendicular to the plane.
Example 1: V (x, y) = (x, y). Then Div(V ) =
P
x
+
Q
y
=

x
(x) +

y
(y) = 2.
Also curl(V ) =
Q
x

P
y
=

x
(y)

y
(x) = 0. Thus, for this radially outward
vector eld, the divergence is positive at every point and the curl is zero.
Example 2: V (x, y) = (x, y). Then Div(V ) =

x
(x) +

y
(y) = 2, whereas
curl(V ) =

x
(y)

y
(x) = 0. This radially inward vector eld has constant
negative divergence and zero curl.
Example 3: V
L
(x, y) = (y, x). Now Div(V ) =

x
(y) +

y
(x) = 0, whereas
curl(V
L
) =

x
(x)

y
(y) = 2. This purely rotational vector eld has no diver-
gence but positive curl (and note that we are rotating in a positive direction).
Example 4: V
R
(x, y) = (y, x) = V
L
(x, y). Indeed Div(V ) = Div(V ) and
curl(V ) = curl(V ) for any vector eld V , as you are invited to check, so this
clockwise rotational vector eld has zero divergence and negative curl.
From these examples it seems that divergence is positive or negative according
to whether there is an outward or inward ow at a point, and the (scalar part of
the) curl is positive or negative according to whether there is a counterclockwise or
clockwise rotation about the point. This is the right idea, but we need to be careful
that we understand why the picture works for all points of these vector elds rather
than just at the origin. Indeed, we are saying that a point has positive divergence
if it is overall a source for a uid: it has more uid owing out than owing
in. This is clear at the origin for Example 1; why is it true at some other point
(x
0
, y
0
)? The answer is that the ow is purely radial, and the magnitude of the ow
is increasing with the distance from the origin. That is, consider the ray joining the
origin to (x
0
, y
0
). In between the origin and (x
0
, y
0
) the speed is smaller than it
is at (x
0
, y
0
), and beyond (x
0
, y
0
) the speed is larger than it is at that point. This
does mean that uid is owing out faster than it is owing in, so the divergence is
HANDOUT FIVE: VECTOR FIELDS 5
positive.
Divergence as ux density: We can argue more generally that divergence at a
point represents a net ow as follows: suppose we have a uid owing in three-
dimensional space and we draw an imaginary box around the uid at a certain
point P
0
= (x
0
, y
0
, z
0
). Then we have the notion of the ux through the surface of
the box, which is the total amount of uid leaving the box minus the total amount
of uid entering the box. Now imagine we do this with a variable box of volume V :
as V 0 we are zooming in on the net ow about the point P
0
. Indeed, it makes
sense to dene the ux density as the limit as V 0 of the ux through a region
of surface area A divided by the surface area V .
We claim that the ux density at P
0
is exactly the divergence at P
0
. Indeed
our box has six faces, so we need to compute the ux through each face, which is
approximately (F n)S, where n is the outward normal vector for the face and
S is the surface area. Suppose our point (x
0
, y
0
, z
0
) is at the bottom-left corner
of the box, and the sides of the box have length x, y and z. Then the surface
area of the top and bottom faces is xy, the outward normal for the top face is

k
and the outward normal for the bottom face is

k. So the ux along the top face

is approximately F

k = Q(x
0
, y
0
, z
0
+ z)xy.
2
The ux through the bottom
face is F (

k) = Q(x
0
, y
0
, z
0
)yz, so the net ux through the top and bottom
faces is the dierence of these, or:
Q(x
0
, y
0
, z
0
+ z)xy Q(x
0
, y
0
, z
0
)xy =
xy (Q(x
0
, y
0
, z
0
+ z) Q(x
0
, y
0
, z
0
)) = (xyz)

Q(x
0
, y
0
, z
0
+ z) Q(x
0
, y
0
, z
0
)
z

.
But (xyz) is the volume of the box, and as z goes to zero the other factor ap-
proaches
Q
z
, so we get lim
xyz0
(ux through top and bottom faces)/(volume
of the box) is /dz. Now we have also to compute the same limit using the left
and right faces, getting
P
x
and through the front and back faces, getting
Q
y
.
Now we have accounted for all six faces in the surface, so the total ux density is
P
x
+
Q
y
+
R
z
= Div(V ), and weve shown that the ux density is equal to the
divergence!
Now that we have justied our phyiscal intuition about the divergence, we in-
troduce the terminology that a vector eld whose divergence is identically zero is
incompressible.
Here is the corresponding physical intuition for the curl of a vector eld: a curl
measures the rotation of a vector eld about a point in the following sense. First
consider the case of a planar vector eld: if we stuck a paddlewheel in the uid at
the point P
0
, then it will turn in the direction of the curl: i.e., counterclockwise if
the curl is positive and clockwise if the curl is negative. It is understood that the
axis of rotation is the z-axis. If we now have a three-dimensional vector eld, the
2
Its only approximately this because the x and y-coordinates are varying along this face, and we
are assuming they are constant. The exact value of the ux is given by a surface integral, which
we will see later in the course, and this argument will reappear in the form of the Divergence
Theorem.
6 PETE L. CLARK
curl captures the net rotation, which will have some axis: in other words, the axis
is such that if we orient the paddlewheel in the direction of curl(F) it will turn with
maximum speed; if we orient it perpendicular to curl(F) it will not turn at all, and
in general it will turn with a speed of || curl(F)|| sin , where is the angle between
the axis of the paddlewheel and curl(F) (this comes from the cross product formula
||v w|| = ||v|||||w|| sin ).
In line with this interpretation, we say that a vector eld F with curl(F) 0
is irrotational.
It is worth asking how one could prove this statement: how does one give a rigor-
ous argument about a paddlewheel? (Indeed, when I took multivariable calculus,
more than ten years ago now, I regarded this statement about the curl measuring
rotation with great suspicion.) But it will turn out that at the very end of the
course we will be in a position to understand rigorously this geometric interpre-
tation of the curl, as well as to revisit the interpretation of the divergence as ux
density. Indeed, these statements are the geometry behind the two most important
results in the course, Stokes Theorem (for the curl) and the Divergence Theorem
(for the divergence, of course!).
We want to give another example to show that a vector eld being irrotational
is actually quite subtle: we really cannot tell just by looking at a rough sketch of
it. For instance, on p. 484 of your text, it is pointed out that just because the
trajectories of a vector eld are circles, it does not necessarily mean that there is
nonzero curl. This is because the curl measures the tendency to rotate locally about
that point, not the tendency for all trajectories to rotate about some central axis.
(The curl is dened in terms of partial derivatives at a point, so is a statement
about very small neighborhoods about that point; it could not see the long-term
behavior of rotation.) However from the arrows drawn in for the vector eld in
Figure 9.42b one cannot conclude whether the vector eld is irrotational or not.
We see explore this by considering the family of vector elds from the last section.
Example: For any number a, we put F
a
(x, y) = (
y
r
a
,
x
r
a
), where r =

x
2
+y
2
.
Note that all of these vector elds dier from the a = 0 case just by rescaling:
in particular all the trajectories are circles. We computed in the last section that
when a = 0 the curl was identically equal to 2. Lets see what happens in general:
the scalar part of the curl is
Q
x

P
y
. Let us compute
dr
dx
and
dr
dy
in advance. We
have
dr
dx
=
2x
2

x
2
+y
2
=
x
r
,
and similarly
dr
dy
=
y
r
.
So
Q
x
= d(x/r
a
)/dx = 1/r
a
+x(d(r
a
/dx) = r
a
+x(a)r
a1
dr/dx) =
HANDOUT FIVE: VECTOR FIELDS 7
r
a
a(x)r
a1
(x/r) = r
a
ax
2
r
a2
=
r
2
ax
2
r
a+2
.
Similarly

P
y
= d(y/r
a
)/dy = d(y/r
a
)/dy =
r
2
2ax
2
r
a1
.
So the scalar part of the curl is
2(r
2
a(x
2
+y
2
))
r
a+2
=
(2 a)(x
2
+y
2
)
r
a+2
.
That is, when a < 2 the curl is positive in particular this covers the case a = 0
when a > 2 the curl is negative: in other words, in this case, despite the fact that
any given particle travels around in a counterclockwise circle, a paddlewheel nailed
to any point will spin clockwise. Most of all, at the special value a = 2 we get an
irrotational vector eld,
F

(x, y) =

y
x
2
+y
2
,
x
x
2
+y
2

.
This very special vector eld will come up again several times in the course.
Example: If f = f(x, y, z) is a function of three variables and F = (f), then
curl(F) = 0. That is, gradient vector elds are irrotational. We leave this calcula-
tion as an exercise (Exercise 29 in Section 9.7).
3. The turning operators L and R
In this section we expose a little secret about vector elds: for planar vector
elds, the curl and the divergence can be understood in terms of each other just by
turning the vector eld. For this, we introduce the following two simple operators:
L(x, y) = (y, x), R(x, y) = (y, x).
As we saw above, for any vector in the plane v = (x, y), Lv just gives us the vector
which is rotated 90 degrees to the left, whereas Rv is the vector rotated 90 degrees
to the right.
L and R can also be applied to vector elds:
L : (P(x, y), Q(x, y) (Q(x, y), P(x, y)).
R : (P(x, y), Q(x, y) (Q(x, y), P(x, y)).
Geometrically, this just means that L(F) is obtained from F just by spinning each
vector 90 degrees to the left, and similarly R(F) is obtained by F just by spinning
each vector 90 degrees to the right. Note that this is not the same as spinning the en-
tire plane 90 degrees: for instance, the vector elds F
1
(x, y) and F
2
(y, x) = (y, x)
are both symmetric about all rotations through the origin, but L(F
1
) = F
2
, and
R(F
2
) = F
1
.
Recall now that F
1
(x, y) = (x, y) has constant divergence 2 and constant curl
0, whereas F
2
(x, y) = (y, x) has constant divergence 0 and constant curl 2. But
L(F
1
) = F
2
. This is an instance of the following simple but useful fact: For any
planar vector eld F = (P(x, y), Q(x, y)),
curl(L(F)) = Div(F)
8 PETE L. CLARK
Div(R(F)) = curl(F).
This works in general for the same reason as the above example: F = (P, Q) implies
L(F) = (Q, P)
, so curl(L(F)) =

x
(P)

y
(Q) =
P
x
+
Q
y
= Div(F).Similarly, R(F) =
(Q, P), so
Div(R(F)) =

x
(Q) +

y
(P) =
Q
x

P
y
= curl(F).
Nothing like this works in space, since there are innitely many dierent possible
axes for a rotation.
This observation is certainly not very deep: there is no more content than the
fact that (x, y) (y, x) = 0. Nevertheless it will prove useful late in the day: it will
help us to understand the relationship between Greens Theorem, Stokes Theorem
and the Divergence Theorem.
4. Extra: Unit tangent vector fields on the sphere and the torus
If we have a surface S in space, then it makes sense to consider vector elds
F(x, y, z) dened on the surface. Indeed, the ux of a vector eld through a closed
surface is obtained by adding up the contribution of the component of the vector
eld which is normal to the surface at every point, a construction which is called
a surface integral and will be studied later in the course.
But consider now exactly the opposite kind of vector eld on a surface: namely a
vector eld which is tangent to the surface at every point. For instance F(x, y) =
(y, x) is a tangent vector eld to the unit circle at every point.
Suppose we look for tangent vector elds on a surface satisfying the additional
condition that every vector is a unit vector at any point: call this a unit tan-
gent vector eld on the surface S. If we have a tangent vector eld that is merely
nonzero at every point, then we can just divide by the norm to get a unit tan-
gent eld: for instance, we can renormalize the above example to get F
T
= (
y
r
,
x
r
),
where as usual r =

x
2
+y
2
. So F
T
is a unit tangent vector eld on the unit circle.
But the circle is a curve, not a surface: what about a unit tangent vector eld
on the unit sphere? We can visualize what we are asking for as follows: suppose
that the sphere is hairy: at each point we have a one inch hair emanating from
that point. We want to comb the hair on the sphere, meaning we want the hair
to lie at against the surface of the sphere, and we want to do this in a continuous
manner no parting of the hair!
Heres a very simple vector eld on the sphere: at each point, we go north
with unit speed, where north means the direction of the shortest path from our
starting point to the north pole
3
At rst it seems like this will give a combing,
but there are two problem points: at the north pole, where it is no longer possible
3
Such a path will be an arc of a great circle, which is obtained by slicing the sphere through
the unique plane containing our starting point P, the north pole N and the center of the earth.
HANDOUT FIVE: VECTOR FIELDS 9
to go north, and also at the south pole when all directions are north according to
the way we have dened it (because all the meridian lines of the sphere run through
the north and the south pole): we are o by two points from combing the sphere.
If you think a bit, you can reduce the number of problem points to one, because it
you remove just the north pole from the sphere, you can unfold what remains into a
surface which looks like the plane: this is the so-called stereographic projection,
which one way of passing from a map on the surface of the sphere (i.e., a globe) to a
at map suitable for a textbook: there is a lot of distortion and the angles change,
but one can get from a nonvanishing tangent eld on the plane to a nonvanishing
tangent eld on the sphere minus the north pole in this way. And there are plenty
of nonvanishing vector elds in the plane, e.g. the constant vector eld F(x, y) =

i.
But no matter which vector eld you choose, you will nd yourself in trouble when
it comes to extending it to the north pole of the sphere.
In fact it is a theorem that you will never succeed: the No Combing Theo-
rem says that there just does not exist a continuously varying unit tangent vector
eld on the unit sphere.
4
The no-combing theorem has real-life consequences: for
instance, it implies the result that at any given time, there is at least one point on
the surface of the earth where there is no wind blowing! Less prosaically, the fact
that any unit vector eld on the surface of the sphere must have a singularity will
come up whenever one tries to study the global
5
behavior of dierential equations
on the surface of the earth.
The situation is much dierent on the torus, which, recall, is given parametri-
cally as
R(u, v) = ((R +a cos u) cos v, (R +a cos u) sin v, a sin u),
with 0 < a < R. Here if we take the tangent vectors in the u and v directions, we
get
T
u
= (R a sin u) cos v, (R a sin u) sin v, a cos u)
T
v
= ((R +a cos u) sin v, (R +a) cos v, 0)
In fact both T
u
and T
v
are nonvanishing for any value of u and v, so after dividing
them by their norms, we nd that we not only have one unit tangent vector eld on
the torus, we have two unit tangent vector elds which are moreover perpendicular
at every point: T
u
T
v
= 0.
Thus the sphere and the torus are qualitatively dierent, as shown in the completely
dierent behavior of tangent vector elds on them. The branch of mathematics that
studies qualitative dierences in surfaces (and other geometric objects) via the
behavior of vector elds is called dierential topology.
4
In more innocent times this was called the hairy ball theorem, but it is hard to say this
with a straight face.
5
We see in this example where the word global comes from!