18.

02 LECTURE NOTES, SPRING 2013

BJORN POONEN
These are an approximation of what was covered in lecture. (Please clear your browsers
cache before reloading this le to make sure you are getting the current version.)
1. February 5
1.1. Vectors. A vector v in R
3
is an ordered triple of real numbers. Example: 2, 3, 5).
(You can have vectors in R
2
or R
95786
too, if you want.) Geometrically, a vector is an arrow
with a length and a direction; its position does not matter.
The standard basis vectors for R
3
:
i := 1, 0, 0)
j := 0, 1, 0)
k := 0, 0, 1)
These are vectors in the directions of the three axes. Also, 0 := 0, 0, 0).
If P is a point in space, then the position vector P is the vector pointing from (0, 0, 0) to
P.
The length (or magnitude) of v = a, b, c) is [v[ :=

a
2
+b
2
+c
2
. This formula can be
explained by using the Pythagorean theorem twice.
A unit vector is a vector of length 1.
Addition: 3, 1) + 1, 4) = 4, 5). Subtraction: 3, 1) 1, 4) = 2, 3). Geometrically:
parallelogram law for +, triangle for . Important: If A and B are two points, and A and
B are their position vectors, then the vector from A to B is BA, because this is what you
have to add to A to get to B.
Scalar multiplication: 10 3, 1) = 30, 10). Scalar means number (one uses this word
when one wants to emphasize that it is not a vector). Scalar multiplication is a scalar times
a vector, and the result is a vector. Geometrically: cv has the same (or opposite) direction
as v, but possibly a dierent length. Two vectors v and w are parallel if one of them is a
scalar multiple of the other.
Question 1.1. How do you write a nonzero vector v such as 3, 4) as a (positive) scalar
times a unit vector?
1
Answer: The scalar is the length, and the unit vector is the original vector divided by its
length:
v = [v[
length
v
[v[
unit vector
.
In our example,
3, 4) = 5
length
_
3
5
,
4
5
_
unit vector
.
The unit vector is giving the direction of v.
Question 1.2. Does the zero vector 0 := 0, 0, 0) have a direction? It is best to say that it has
every direction, and hence to say that it is parallel to every other vector, and perpendicular
to every other vector.
Another example involving addition and scalar multiplication:
2 i + 3 j + 5 k
linear combination of i, j, k
= 2, 3, 5).
In general, a linear combination of vectors v
1
, . . . , v
n
is any vector obtained by multiplying
the vectors by (possibly dierent) scalars and adding the results, i.e., an expression of the
form c
1
v
1
+ +c
2
v
n
for some scalars c
1
, . . . , c
n
.
Question 1.3. Suppose that M is the midpoint of segment AB. In terms of the position
vectors A and B, what is the position vector M?
To get to M, start at A and go halfway from A to B. The vector from A to B is BA,
so the vector from A to M is
1
2
(BA). Thus
M = A+
1
2
(BA) =
A+B
2
.
Question 1.4. Prove that the midpoints of the sides of a space quadrilateral form a paral-
lelogram.
How should one approach a problem like this?
Give variable names to the objects given in the problem. Let A, B, C, D be the vertices
of the quadrilateral. Let A

be the midpoint of AB, let B

be the midpoint of BC, let C

be
the midpoint of CD, and let D

be the midpoint of DA. Let A, B, C, D, A

, B

, C

, D

be the
corresponding position vectors.
Write down known equations relating the variables. We know that
A

=
A+B
2
, B

=
B+C
2
, C

=
C+D
2
, D

=
D+A
2
.
2
See if these equations imply the desired conclusion. We compute
B

=
CA
2
C

=
CA
2
.
Thus segments B

and C

are parallel and have the same length. This means that
A

is a parallelogram.
1.2. Some advice for success in 18.02.
Read the Stellar website now, especially Information.
Do the reading assignments before lecture!
Come to oce hours! (Oce hours generally consist of a small group of students
discussing additional examples not covered in lecture or recitation, asking questions,
getting started on dicult homework problems, etc. The recitation leaders for this
class are some of the best math faculty and grad students worldwide, and oce hours
are your best chance to learn from them.)
Homework:
Its long and has dicult problems, so start now! When I created the homework
assignments this January, I labelled each problem with the date when you should
be able to start it.
Work together in groups! Its OK if other people tell you how to solve a problem,
but dont be looking at their solution as you write your own.
Do what it takes (come to oce hours, discuss problems with others) so that
when you hand in an assignment you are pretty sure that it is complete and
correct.
1.3. Dot product. Do we multiply vectors coordinate-wise? No! Why not? This does not
give a notion with useful geometric meaning. Instead:
Dot product (also called scalar product or inner product):
2, 3, 5)
vector
7, 8, 9)
vector
= (2)(7) + (3)(8) + (5)(9)
= 14 + 24 + 45
= 83
scalar
.
Important special case: a a = [a[
2
.
Theorem 1.5 (Geometric interpretation of the dot product). If is the angle between
nonzero vectors a and b, then
a b = [a[[b[ cos .
3
(If a or b is 0, then can be taken to be any real number, and the formula still holds, with
both sides being 0.)
Why?
Proof (=explanation). Let c = a b. Let a = [a[, b = [b[, c = [c[. Then
c
2
= (a b) (a b)
= a a a b b a +b b
= a
2
+b
2
2a b.
On the other hand, the law of cosines says
c
2
= a
2
+b
2
2ab cos .
Comparing shows that a b = ab cos .
Corollary 1.6. Vectors a and b are perpendicular a b = 0.
Denition 1.7 (Scalar component of a vector). If a is a vector, and b is a nonzero vector,
then the scalar component of a in the direction of b measures the amount of a that is in
the direction of b, i.e., what the x-coordinate of a would be in a new coordinate system in
which the x-axis were in the direction of b. It is dened by the following formulas:
comp
b
a := a
b
[b[
= [a[ cos .
The equality of the last two expressions is explained as follows:
a
b
[b[
=
1
[b[
(a b) =
1
[b[
([a[[b[ cos ) = [a[ cos .
Geometric interpretation: If a and b are position vectors, then comp
b
a is the length of the
projection of a onto the line containing b, except that if the endpoint of the projection is on
the other side of b, then comp
b
a is negative.
Example 1.8. If a = 2, 3, 5) and b = j, then comp
b
a = a j = 3.
1.4. Matrices.
Denition 1.9. An m n matrix is a rectangular array of numbers with m rows and n
columns.
Example:
A :=
_
_
_
_
_
3 5
4 0 0
6 7 9
1 2 3
_
_
_
_
_
4
is a 4 3 matrix. (The dimensions are always given in the order height width.)
The notation a
ij
means the number in the i
th
row and j
th
column. In the example above,
a
32
= 7.
An m 1 matrix is also called a column vector. For example,
_
_
_
2
3
5
_
_
_
is the same as the
vector 2, 3, 5).
1.5. Determinants. To each square matrix A is associated a number called the determinant.
det
_
a
_
:= a
= length of segment in the real line R determined by a
det
_
a b
c d
_
:= ad bc
= area of parallelogram in R
2
spanned by a, b) and c, d)
(the sign is + if c, d) is counterclockwise from a, b))
det
_
_
_
a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3
_
_
_
:= a
1
b
2
c
3
+a
2
b
3
c
1
+a
3
b
1
c
2
c
1
b
2
a
3
c
2
b
3
a
1
c
3
b
1
a
2
= volume of parallelepiped in R
3
spanned by the rows a, b, c
(the sign is + if a, b, c agree with the right hand rule)
(One could also use the columns instead of the rows.)
Alternative notation for determinant: [A[ =

a b
c d

. (This is a scalar, not a matrix!)

2. February 7
Laplace expansion (along the rst row) for a 3 3 determinant:

a
1
a
2
a
3
b
1
b
2
b
3
c
1
c
2
c
3

= +a
1

b
2
b
3
c
2
c
3

a
2

b
1
b
3
c
1
c
3

+a
3

b
1
b
2
c
1
c
2

.
The general rule leading to the formula above is this:
(1) Move your nger along the entries in a row.
(2) At each position, compute the minor, dened as the smaller determinant obtained by
crossing out the row and the column through your nger; then multiply the minor by
5
the number you are pointing at, and adjust the sign according to the checkerboard
pattern
+ +
+
+ +
(the pattern always starts with + in the upper left corner).
(3) Add up the results.
There is a similar expansion for a determinant of any size, computed along any row or
column.
Properties of determinants:
D-1: Interchanging two rows changes the sign of det A.
D-2: If one of the rows is all 0, then det A = 0.
D-3: Multiplying an entire row by a scalar c multiples det A by c.
D-4: Adding a multiple of a row to another row does not change det A.
These properties can all be interpreted geometrically. The same properties hold for column
operations.
Example 2.1. Let
A :=
_
2 3
5 7
_
.
If we add 10 times the rst row to the second row (and leave the rst row unchanged), we
get a new matrix
B :=
_
2 3
25 37
_
.
Property D-4 says that det B = det A.
Question 2.2. Suppose that A is a 3 3 matrix such that det A = 5. Doubling every entry
of A gives a matrix 2A. What is det(2A)?
Solution: Each time we multiply a row by 2, the determinant gets multiplied by 2. We
need to do this three times to double the whole matrix A, so the determinant gets multiplied
by 2 2 2 = 8. Thus det(2A) = 8 det(A) = 40.
2.1. Cross product.
6
Denition 2.3 (Cross product, also called vector product). The cross product of vectors a
and b in R
3
is the vector
a b :=

i j k
a
1
a
2
a
3
b
1
b
2
b
3

:= +

a
2
a
3
b
2
b
3

a
1
a
3
b
1
b
3

j +

a
1
a
2
b
1
b
2

k
= (a
2
b
3
a
3
b
2
) i (a
1
b
3
a
3
b
1
) j + (a
1
b
2
a
2
b
1
) k.
Geometrically: If a and b are nonzero vectors forming an angle , then ab is the vector
determined by the following three conditions:
It is perpendicular to a and b.
Its length is [a[[b[ sin (which equals the area of the parallelogram spanned by a and
b).
Its direction is given by the right hand rule: if you point the ngers of your right hand
in the direction of a so that bending your ngers makes them point in the direction
of b, then your thumb shows the direction of a b.
Example 2.4.
i k = j.
(It is a vector perpendicular to i and k, and its length is (1)(1) sin

2
= 1, so it must be j or
j. The right hand rule says it is j.)
For any vectors a and b,
a a = 0, b a = a b.
(We havent explained why the determinant and cross product have the geometric inter-
pretations claimed, but we might explain this later on.)
2.2. Matrix times a column vector.
2.2.1. Example and denition.
Example 2.5.
_
1 2 3
2 3 5
_
matrix
_
_
_
10
1
100
_
_
_
column vector
=
_
312
523
_
column vector
because
1, 2, 3) 10, 1, 100) = 312
2, 3, 5) 10, 1, 100) = 523.
The example above illustrates the following general denition:
7
Denition 2.6 (Matrix times a column vector). If A is a matrix and x is a column vector,
then
Ax :=
_
_
_
(rst row of A) x
(second row of A) x
.
.
.
_
_
_
(Ax is dened only if the width of A equals the height of x.)
2.2.2. Recovering columns of a matrix from matrix-vector multiplications. It is possible to
recover the columns of a matrix A just from knowing what A times each standard basis
vector (i, j, etc.) is.
Example 2.7. Recovering the rst column:
_
6
2
_
=
_
6 7 8
2 3 5
_
_
_
_
1
0
0
_
_
_
.
In general:
(rst column of A) = A
_
_
_
1
0
0
_
_
_
, (second column of A) = A
_
_
_
0
1
0
_
_
_
, and so on.
2.2.3. Matrix-vector multiplication as a linear combination of the columns.
Example 2.8.
_
6 7 8
2 3 5
_
_
_
_
x
y
z
_
_
_
=
_
6x + 7y + 8z
2x + 3y + 5z
_
= x
_
6
2
_
+y
_
7
3
_
+z
_
8
5
_
.
In general: Ax = (some linear combination of the columns of A) .
2.3. Matrices as linear transformations.
2.3.1. Going from a matrix to a linear transformation. Given a number, say 3, we can create
a function f(x) := 3x that takes as input a number x and outputs the number 3x.
Higher-dimensional analogue: Given a matrix A, we can create a function f(x) := Ax that
takes as input a column vector x and outputs the column vector Ax. Functions arising in
this way are called linear transformations.
Example 2.9. The 2 3 matrix
_
6 7 8
2 3 5
_
gives rise to the linear transformation
f
_
_
_
x
y
z
_
_
_
=
_
6x + 7y + 8z
2x + 3y + 5z
_
.
8
One writes f : R
3
R
2
since the possible inputs are the vectors in R
3
, and every output is
a vector in R
2
. What does this function do to j? Answer: f(j) = f
_
_
_
0
1
0
_
_
_
=
_
7
3
_
.
Remark 2.10. The reason that matrix-vector multiplication is dened the way it is is so that
linear transformations can be written as Ax!
Class on February 8 was cancelled because of winter storm Nemo.
3. February 12
3.1. Matrices as linear transformations, continued.
3.1.1. Going from a linear transformation to a matrix. Given a linear transformation, how
do we reconstruct the matrix that gives rise to it?
Answer 1: If we know a formula for the function, we can just read o the entries of the
matrix. For example, given the function f
_
_
_
x
y
z
_
_
_
=
_
6x + 7y
2x + 5z
_
, the matrix is
_
6 7 0
2 0 5
_
.
Answer 2: If we know only values of the linear transformation f, then we can recover the
matrix one column at a time: the rst column is f
_
_
_
1
0
0
_
_
_
, the second column is f
_
_
_
0
1
0
_
_
_
, and
so on.
Question 3.1. Given , there is a 2 2 matrix R whose associated linear transformation
rotates each vector in R
2
counterclockwise by the angle . What is it?
The rotation maps 1, 0) to cos , sin ) and 0, 1) to sin , cos ). Thus
(rst column of R) = R
_
1
0
_
=
_
cos
sin
_
(second column of R) = R
_
0
1
_
=
_
sin
cos
_
,
so
R =
_
cos sin
sin cos
_
.
3.1.2. 22 matrices as linear transformations of the plane. A linear transformation f : R
2

R
2
can be visualized by drawing a subset of the input plane R
2
, evaluating f at every point
of the subset, and plotting each value in the output plane R
2
.
9
Example 3.2. Let A =
_
3 0
0 3
_
. Let f be the associated linear transformation, so f(x, y) =
(3x, 3y). Then f maps a smiley face (say centered at the origin, of radius 1) to a larger smiley
face. In particular, f maps the left ear at (1, 0) to (3, 0). (Such a linear transformation is
called a dilation.)
Question 3.3. What is a matrix A whose associated linear transformation f sends the smiley
face to a fat smiley face, the same height as before but twice as wide?
Solution: The rst column is f
_
1
0
_
=
_
2
0
_
and the second column is f
_
0
1
_
=
_
0
1
_
, so
A =
_
2 0
0 1
_
.
3.1.3. Area scaling factor.
Example 3.4. What does the dilation given by
_
3 0
0 3
_
do to area?
(Hint: Consider what it does to a 1 1 square.)
Answer: It multiplies area by 9. Notice that det
_
3 0
0 3
_
= 9 too.
In general, for any 2 2 matrix A, the associated linear transformation has
area scaling factor = [ det A[.
(The area scaling factor is always nonnegative, while det A could be negative, so it is neces-
sary to take the absolute value.)
Why? If A =
_
a b
c d
_
, then the square spanned by
_
1
0
_
and
_
0
1
_
is transformed into
the parallelogram spanned by
_
a
c
_
and
_
b
d
_
; the square has area 1 and the parallelogram
has area [ det A[.
3.2. Matrix operations.
Addition, subtraction, and scalar multiplication are dened entrywise, just as for vectors.
(For addition and subtraction, the two matrices have to have the same dimensions.)
Transpose:
_
2 3 5
7 11 13
_
T
=
_
_
_
2 7
3 11
5 13
_
_
_
.
Each row of the original matrix corresponds to a column of the transpose matrix. The
ij-entry of A equals the ji-entry of A
T
.
10
Multiplication: If A is mn and B is n p, then AB is the mp matrix whose ij-entry
is the dot product
(i
th
row of A) (j
th
column of B)
Example 3.5. Multiplying a 2 3 matrix by a 3 3 matrix gives a 2 3 matrix:
_
2 3 7
0 5 1
_
_
_
_
10 0 9
1 2 0
0 1 0
_
_
_
=
_
23 13 18
5 11 0
_
.
For example, the 2, 1 entry is computed as follows:
0, 5, 1)
2nd row
10, 1, 0)
1st column
= 0(10) + 5(1) + 1(0) = 5
2,1 entry
.
Matrix multiplication corresponds to composition of the linear transformations. More
explicitly, if f : R
n
R
m
corresponds to a matrix A, and g: R
p
R
n
corresponds to a
matrix B, then the composition f g (i.e., the function f(g(x))) corresponds to AB. Thats
why matrix multiplication is dened the way it is.
Warning 3.6. Even when AB and BA both make sense, they might be unequal. (In other
words, matrix multiplication is not commutative.)
3.3. Identity matrix. The identity function on R
3
is f(x, y, z) := (x, y, z). In terms of
column vectors:
f
_
_
_
x
y
z
_
_
_
:=
_
_
_
1x + 0y + 0z
0x + 1y + 0z
0x + 0y + 1z
_
_
_
.
The corresponding matrix
I :=
_
_
_
1 0 0
0 1 0
0 0 1
_
_
_
is called the 3 3 identity matrix. (You can guess what the nn identity matrix looks like.)
It acts like the number 1:
IA = A and AI = A
whenever the multiplication makes sense.
3.4. Inverse matrices.
11
3.4.1. Motivation: solving linear systems. To solve 3x = 5, multiply both sides by 3
1
.
Similarly, one way to solve
2x
1
+ 3x
2
= 4
4x
1
+ 5x
2
= 6,
is to rewrite as
_
2 3
4 5
__
x
1
x
2
_
=
_
4
6
_
,
which has the shape Ax = b, and left multiply both sides by A
1
to get x = A
1
b.
3.4.2. Denition.
Denition 3.7. The inverse of an n n matrix A is another n n matrix A
1
such that
AA
1
= I and A
1
A = I.
It exists if and only if det A ,= 0.
If A corresponds to the linear transformation f, then A
1
corresponds to the inverse
function f
1
(if it exists).
3.4.3. Calculation.
(1) For each i, j, calculate the i, j minor, i.e., the determinant obtained by crossing out
the row and column containing a
ij
. These minors are the entries of a matrix.
(2) Change the signs by the checkerboard pattern to get the cofactor matrix.
(3) Take the transpose to get the adjoint matrix.
(4) Multiply by the scalar
1
det A
.
Example 3.8. Suppose that A =
_
_
_
0 2 3
0 1 1
2 3 5
_
_
_
. Cross out row 1 and column 2 and take the
determinant to get the 1, 2 minor:

0 1
2 5

= 0 2 = 2. Do the same for each i, j to get

all nine minors; put them in a matrix:
matrix of minors:
_
_
_
2 2 2
1 6 4
1 0 0
_
_
_
.
Change signs to get
cofactor matrix:
_
_
_
2 2 2
1 6 4
1 0 0
_
_
_
.
12
Take the transpose:
adjoint: adj A =
_
_
_
2 1 1
2 6 0
2 4 0
_
_
_
.
We nd det A = 2 (e.g., by Laplace expansion along rst column). Multiply by
1
det A
:
inverse matrix: A
1
=
1
2
_
_
_
2 1 1
2 6 0
2 4 0
_
_
_
=
_
_
_
1 1/2 1/2
1 3 0
1 2 0
_
_
_
.
3.5. Equations of planes.
Question 3.9. What does the set of vectors perpendicular to 1, 2, 3) look like?
Solution: Its a plane through the origin. Its equation is
1, 2, 3) x, y, z) = 0,
which is
x + 2y + 3z = 0.
The vector n := 1, 2, 3) is called a normal vector to the plane. (Normal is another word for
perpendicular.)
Question 3.10. What is the plane with normal vector 1, 2, 3) passing through (4, 5, 6)?
Solution: A point (x, y, z) lies on this plane if the vector from (4, 5, 6) to (x, y, z) (not the
position vector of (x, y, z)!) is perpendicular to 1, 2, 3), so its equation is
1, 2, 3) (x, y, z) 4, 5, 6)) = 0,
which is
(x 4) + 2(y 5) + 3(z 6) = 0.
4. February 14
Question 4.1. (Followup question) What is the distance from (2, 3, 5) to that plane?
Solution: If we choose any point on the plane, say (4, 5, 6), and form the vector
v := 2, 3, 5) 4, 5, 6) = 2, 2, 1)
between the two points, then the desired distance is not the length of v, because the straight
line path from (2, 3, 5) to (4, 5, 6) is not the shortest path from (2, 3, 5) to the plane. Instead
we want the amount of v in the direction parallel to the normal vector n := 1, 2, 3),
13
taking the absolute value if necessary. Thus the desired distance is the absolute value of the
scalar component of v in the direction of n. That scalar component is
comp
n
v = v
n
[n[
=
2, 2, 1) 1, 2, 3)

1
2
+ 2
2
+ 3
2
=
2 4 3

14
=
9

14
,
so the distance is 9/

14.
Question 4.2. Are the vector 5, 1, 1) and the plane x + 2y + 3z = 6
(1) parallel,
(2) perpendicular,
(3) both,
(4) or neither?
Hint: 5, 1, 1) 1, 2, 3) = 0.
Answer: The vector 5, 1, 1) is perpendicular not to the plane, but to a normal vector
of the plane. So it is parallel to the plane. The vectors perpendicular to the plane are the
scalar multiples of the normal vector, so 5, 1, 1) is not like this. So the answer is (1).
4.1. Square systems of linear equations.
4.1.1. Homogeneous systems. The system
x + 2y + 3z = 0
8x + 4z = 0
7x + 6y + 5z = 0
is of the form Ax = 0, where A is a 3 3 matrix, and x =
_
_
_
x
y
z
_
_
_
. It is called homogeneous
because the right sides are all 0.
Solving Ax = 0 for any square A:
If det A ,= 0, then x = 0 is the only solution. (Proof: Multiply both sides by A
1
.)
If det A = 0, then there are innitely many solutions.
Geometric interpretation, in the 3 3 case: For any 3 3 matrix A, the solution set to
Ax = 0 is the intersection of 3 planes passing through 0 (assuming that no row of A is all
0).
14
If det A ,= 0, the intersection is only 0 (this is what usually happens).
If det A = 0, the intersection is either a line or a plane (through 0).
4.1.2. General systems. A general square system has the form Ax = b.
For any square A:
If det A ,= 0, then x = A
1
b is the only solution. (Proof: Multiply both sides by
A
1
.)
If det A = 0, there are either innitely many solutions or no solutions.
If A is 3 3, the solution set is still an intersection of 3 planes, but they may be shifted
away from 0.
4.2. Lines. Two ways to describe lines in R
3
:
intersection of two planes
parametric equation (to be discussed today)
Think of the trajectory of an airplane moving at constant velocity. Let r
0
be the position
vector of the airplane at time t = 0. Let v be the velocity.
Where is the airplane at time t = 1? Answer: r
0
+v.
At time t = 2? Answer: r
0
+ 2v.
In general, at time t it is at
r(t) := r
0
+tv.
This is a vector-valued function of t. Each real number t gives one point on the line, and as
t varies, these points trace out the whole line.
Example 4.3. Line L through (1, 2, 3) and (4, 1, 3)? Use initial position vector r
0
:= 1, 2, 3)
and velocity vector v := 4, 1, 3)1, 2, 3) = 3, 1, 0) so that at time t = 1 the point reaches
(4, 1, 3). So L is given by
r = 1, 2, 3) +t3, 1, 0).
Parametric equations of L in terms of coordinates:
x = 1 + 3t, y = 2 t, z = 3.
Question 4.4. The lines
x = 1 + 3t, y = 2 t, z = 3
and
x = 2t, y = 1 +t, z = 1 +t
(1) are the same,
(2) are parallel,
(3) intersect in one point, or
(4) are skew (i.e., do not intersect, but are not parallel either)?
15
Answer: The velocity vectors 3, 1, 0) and 2, 1, 1) are not scalar multiples of each other,
so the lines are not the same, and are not parallel. They intersect if the system
1 + 3t = 2s
2 t = 1 +s
3 = 1 +s.
is solvable. Why did we use a dierent variable on the right side? Imagine airplanes moving
along the lines. If we used the same t on both sides, a solution would be a time when
both airplanes are at the same place. If we use t on the left and s on the right, a solution
would mean that airplane #1 at some time is at the same point as airplane #2 at a dierent
time, meaning that their paths still cross, and this is what were trying to test! OK, lets
now solve the system. The last equation implies s = 2. The rst equation then implies
t = 1. These values make all three equations true. So the lines intersect at the point
r
1
(1) = r
2
(2) = 4, 1, 3).
4.3. Parametric equations of curves. As t ranges through all real numbers,
x = 2 cos t, y = sin t
describes
1) A circle
2) An ellipse
3) A line of slope 1/2
4) A point.
Answer: Without the 2, it would be the unit circle. The 2 stretches it in the x-direction, to
make an ellipse. It is given by the implicit equation
x
2
+ 4y
2
= 4.
To nd the implicit equation, one must eliminate t from the parametric equations; how to
do this depends on the shape of the parametric equations, and may require some guesswork.
In this problem, we know that sin
2
t +cos
2
t = 1, and this can be rewritten as y
2
+(x/2)
2
= 1
in which t does not appear, which is equivalent to x
2
+ 4y
2
= 4. The parametric equations
and the implicit equation are completely dierent ways of describing the same curve: the
parametric equations tell you the curve one point at a time, one point for each value of t;
the implicit equation gives you a test for when a given point in R
2
lies on the curve.
Slope of tangent line to this ellipse at a given time? Use the chain rule
dy
dt
known
=
dy
dx
???
dx
dt
known
16
to get
y

=
dy
dx
=
dy/dt
dx/dt
=
cos t
2 sin t
=
1
2
cot t.
The following will be explained in recitation on Monday.
What is
d
2
y
dx
2
? Its
d
2
y
dx
2
=
dy

dx
=
dy

/dt
dx/dt
=
1
2
csc
2
t
2 sin t
=
1
4 sin
3
t
.
5. February 15
5.1. Parametric equations of curves: an example.
Example 5.1. Did Exercise 1I-4 about the roll of tape centered at (0, 0) and with end
initially at (a, 0). Parameter: the radian measure of the amount of tape unwound so far.
Length of tape unwound so far is a. Position of point where the unwound tape meets the
roll: (a cos , a sin ). Final answer:
r = a cos , a sin ) +acos , sin )
= a(1 +) cos , a(1 +) sin ).
5.2. Derivative of a vector-valued function.
5.2.1. Denition and physical interpretation.
Denition 5.2. The derivative of a vector-valued function r(t) is
r

(t) := lim
h0
r(t +h) r(t)
h
.
If r(t) is the position of a moving particle at time t, then r

(t) is its velocity at time t.

5.2.2. Calculating the derivative. Derivatives can be calculated coordinate-wise:
Theorem 5.3. If
r(t) = f(t) i +g(t) j,
then
r

(t) = f

(t) i +g

(t) j.
This holds because all the ingredients used in the denition of derivative (vector subtrac-
tion, scalar multiplication by 1/h, and limits) can be calculated coordinate-wise. Its the
same in 3D.
Advice: Use the denition to understand the derivative physically, but use the theorem
to calculate it.
17
5.3. Integration of vector-valued functions.
Similarly, the denite integral
_
b
a
r(t) dt := lim
t0
n

i=1
r(t

i
) t
can be computed coordinate-wise: if
r(t) = f(t)i +g(t)j,
then
_
b
a
r(t) dt =
__
b
a
f(t) dt
_
i +
__
b
a
g(t) dt
_
j.
Also, there is a fundamental theorem of calculus for vector-valued functions: if r(t) = R

(t),
then
_
b
a
r(t) dt
integral of a rate of change
= R(b) R(a)
total change
.
It follows from the scalar function version.
5.4. Acceleration. Recall: If r(t) is the position of an object at time t, then its velocity at
time t is v(t) := r

(t). Next, its acceleration at time t is a(t) := v

(t) = r

(t). All of these

are vector-valued functions.
On the other hand, its speed is [v(t)[; this is a scalar function whose values are nonnegative.
Question 5.4. What does it mean if a particles position vector r(t) satises
d|r|
dt
= 0 for all
t?
Answer: The length of the position vector is constant, so the particle is staying on a sphere
centered at the origin.
5.5. Arc length. Let s(t) be the scalar function that gives the arc length of the trajectory
from a starting point r(a) to a variable end point r(t). Imagine that dt is a tiny amount of
time. If it is really tiny, then we can imagine that the speed [v[ doesnt change much during
that time interval from time t to time t +dt, so the distance traveled is approximately [v[ dt.
Imagine that ds means the change in s(t) during the same time interval. Then ds is the
distance traveled, so
ds = [v[ dt.
(The reason this explanation is in magenta is that ds and dt are not actually numbers. It
would be more correct to say that s [v[ t, and then to divide by t and take a limit.)
In any case, mathematically what this means is that
ds
dt
= [v[.
18
So, by the fundamental theorem of calculus,
s(t) =
_
t
a
[v(u)[ du.
(The u in the integration can be thought of as a variable number increasing from a up to t;
we couldnt just use t, because the name t is already being used, as the upper limit of the
integration.)
Think: total distance traveled equals sum of distance traveled over each time subinterval
between a and t.
5.6. Unit tangent vector. Suppose that v(t) ,= 0. Then v(t) is a tangent vector to the
curve.
Denition 5.5. The unit tangent vector at r(t) is the vector
T :=
v(t)
[v(t)[
.
Then
dr
dt
= v = T[v[ = T
ds
dt
.
One writes dr = Tds.
5.7. Foci of an ellipse. (This section is just to help make sense of Keplers rst law, and
to help with one of the homework problems.)
One way to write down an ellipse is to write an equation
x
2
a
2
+
y
2
b
2
= 1.
Here a is half the width, and b is half the height.
But if you were to ask an ancient Greek what an ellipse is, the answer would be:
Fix two points P and Q and x a number greater than PQ. Then the locus
(possible positions) of a point R such that PR +RQ = is an ellipse.
The points P and Q are the foci (plural of focus) of the ellipse.
Another property of the foci: Inside an elliptical room with mirror walls, if you place a
light at one focus, the rays will reect and meet again at the other focus. (Same for sound.)
5.8. Keplers second law. In the early 1600s, Johannes Kepler noticed that Tycho Brahes
data on planetary motion was consistent with three laws:
(1) The orbit of a planet is an ellipse with the sun at one focus.
(2) A planet moves in a plane containing the sun, and the line segment connecting the
sun to the planet sweeps out area at a constant rate.
(3) The square of the period of revolution of a planet about the sun is proportional to
the cube of the major semiaxis of its elliptical orbit.
19
Another law: Gravitational force F is central: In mathematical terms, the planets
acceleration vector a is always parallel to the vector from the sun to the planet.
Theorem 5.6. Keplers second law is equivalent to the law saying that the acceleration is
central.
Proof. Let the origin be where the sun is. Let r = r(t) be the position vector of the planet.
Let A = A(t) be the area swept out from some starting time until t.
Between time t and t + t,
A Area(triangle) =
1
2
[r r[.
Divide by t to get a rate, and take the limit as t 0:
dA
dt
=
1
2
[r
dr
dt
[ =
1
2
[r v[.
Suppose that Keplers second law holds:
dA
dt
is constant. Then [r v[ is constant. But
the direction of r v is also constant since it is perpendicular to the plane of motion (and
by continuity cannot suddenly switch to the opposite direction). Thus r v is constant. So
d
dt
(r v) = 0.
On the other hand, by a rule for dierentiating a cross product (from the textbook reading
for today),
d
dt
(r v) =
dr
dt
v +r
dv
dt
(important: keep the r and v in order)
= v v +r a
= r a.
So
r a = 0.
This means that a is parallel to r.
The converse, that the acceleration being central implies Keplers second law, can be
proved by reversing the steps of the previous paragraph.
6. February 21
Midterm #1 on Tues. Feb. 26, 11am-12 in 50-340 (upstairs in Walker Memorial, not here!)
It covers everything up to what we do today (level curves, partial derivatives, tangent plane,
tangent plane approximation) but not Fridays topic (max/min). You dont need to know
Keplers laws. Practice midterms are available on Stellar; ignore the max/min problems in
them for now.
20
Homework #3 due Thurs. Mar. 7 at 11 A.M. in room 2-255. But start the problems
labelled Feb. 21 now (such problems may appear on the midterm), and come to oce hours
if you arent sure that youre doing them correctly.
6.1. Graphs and level curves of two-variable functions. Three ways to depict a 2-
variable function f(x, y):
(1) Map of its values: At many points (x, y) in the plane, write the value f(x, y). For
example, if f(x, y) := y(y + 1)/2 x + 10, then the values at integer points are
30 29 28 27 26 25 24 23 22 21 20
25 24 23 22 21 20 19 18 17 16 15
21 20 19 18 17 16 15 14 13 12 11
18 17 16 15 14 13 12 11 10 9 8
16 15 14 13 12 11 10 9 8 7 6
15 14 13 12 11 10 9 8 7 6 5
15 14 13 12 11 10 9 8 7 6 5
16 15 14 13 12 11 10 9 8 7 6
18 17 16 15 14 13 12 11 10 9 8
21 20 19 18 17 16 15 14 13 12 11
25 24 23 22 21 20 19 18 17 16 15
(the colored entry is f(0, 0)).
(2) Graph: Above each point (x, y) in the plane, plot a point in R
3
whose z-coordinate is
the value f(x, y). Taken together, these points form a surface in R
3
called the graph
of f. It is the set of points in space satisfying the equation z = f(x, y).
(3) Level curves: For each number h, the level curve at height h is the set of points (x, y)
in the xy-plane such that f(x, y) = h. In the example above, each level curve is a
parabola.
Question 6.1. Consider f(x, y) = x
2
+y
2
. Its graph is a paraboloid. Draw level curves for
equally spaced values of h; these are circles in R
2
centered at (0, 0), namely x
2
+ y
2
= h.
Then as one goes out, are the circles
(1) getting closer together
(2) occurring at equally spaced radii
(3) getting farther apart?
Answer: Getting closer together. The farther out you are, the steeper the paraboloid is,
so the shorter you have to go horizontally to get a xed increase in height.
6.2. Partial derivatives.
21
6.2.1. Introduction via an example.
Here is a map showing values of a function f(x, y) at integer points, with the colored value
at (0, 0):
20 19 18 17 16 15
16 16 16 16 16 16
12 13 14 15 16 17
8 10 12 14 16 18
4 7 10 13 16 19
If one starts at (0, 0) and moves to the right, the value increases by 2 for each increase of x
by 1; thus the rate of change, denoted
f
x
(0, 0), equals 2. (Note:
f
x
is often pronounced
partial f partial x.) Similarly,
f
y
(0, 0) = 3, and
f
x
(2, 1) = 3. Since
f
x
has potentially
dierent values at dierent points, it is itself a function dened on the plane.
6.2.2. Denition.
Denition 6.2. The partial derivative of f(x, y) with respect to x is a function
f
x
whose
value at (x
0
, y
0
) is
the rate of change of f(x, y) when x is varying near x = x
0
and y is held constant at
the value y
0
, or,
more precisely,
f
x
(x
0
, y
0
) := lim
h0
f(x
0
+h, y
0
) f(x
0
, y
0
)
h
(if the limit exists).
Other notations:
f
x
(x
0
, y
0
),
_
f
x
_
0
.
The 0 subscript on the derivative means evaluate me at (x
0
, y
0
).
What is the dierence between
df
dx
and
f
x
? The rst notation is used if f is a function
f(x) of x alone. The second notation is used if f is a function of several variables but we
are measuring the rate of change of f arising from a change in only the variable x.
Dene
f
y
similarly.
6.2.3. How to compute
f
x
.
View y as a constant and dierentiate with respect to x. (And then evaluate at (x
0
, y
0
) if
desired.) Example: If f(x, y) = x
3
y
5
, then f
x
= 3x
2
y
5
and f
x
(2, 1) = 12.
Question 6.3. Let f(x, y) = x
y
for (x, y) in the half-plane x > 0. What is
f
y
?
Possible answers:
22
(1) x
y
ln x
(2) yx
y1
(3) 0
(4) None of the above
Answer: (1). Computing
f
y
is like computing
d
dy
2
y
=
d
dy
e
y ln 2
= e
y ln 2
ln 2 = 2
y
ln 2
except with a constant x in place of the 2. So the answer is x
y
ln x.
More notation: to calculate the second partial derivative
f
xy
= (f
x
)
y
=

y
f
x
=

y
_

x
f
_
=

2
f
y x
,
rst take the x-derivative of f, and then take the y-derivative of the result. The other second
partial derivatives are f
xx
, f
yx
, and f
yy
. For example,
f
xx
=

2
f
x
2
.
For most functions arising in practice, including any function for which all the second partial
derivatives are continuous, f
xy
= f
yx
. So usually you dont have to worry about the order in
which you take derivatives.
6.3. The tangent plane (to the graph of a 2-variable function). (Below, numbers are
colored green to distinguish them from variables.)
Recall: To nd the tangent line at a point (x
0
, y
0
) on the curve y = f(x):
(1) compute the function f

(x)
(2) plug in x = x
0
to get a number f

(x
0
) (its going to be the slope)
(3) write down the equation of the line through (x
0
, y
0
) with slope f

(x
0
) (point-slope
form):
y y
0
= f

(x
0
)(x x
0
)
equation of tangent line
.
Equivalently:
y = f(x
0
) +
_
df
dx
_
0
(x x
0
)
. .
call this (x)
.
Then (x) is a linear polynomial (of the form mx + b) such that (x
0
) = f(x
0
) and

(x
0
) =
f

(x
0
) (same value and same derivative at x
0
).
If x is close to x
0
, then
f(x)
height of graph of f
(x)
height of tangent line
23
so
f(x) f(x
0
) +
_
df
dx
_
0
(x x
0
)
tangent line approximation
.
If we move f(x
0
) to the left side, this can be rewritten as
f
_
df
dx
_
0
x.
Now let f(x, y) be a 2-variable function. Let (x
0
, y
0
, z
0
) be a point on the graph of f(x, y)
(so z
0
= f(x
0
, y
0
)). To nd the tangent plane there:
(1) compute the functions
f
x
and
f
y
(2) plug in x = x
0
and y = y
0
to get two numbers
_
f
x
_
0
and
_
f
y
_
0
.
(3) write down the equation
z z
0
=
_
f
x
_
0
(x x
0
) +
_
f
y
_
0
(y y
0
)
Note: This plane is rigged so as to pass through (x
0
, y
0
, z
0
). It is also rigged to have the
same rate of increase as the graph of f(x, y) as one moves to the right or up starting from
(x
0
, y
0
).
Question 6.4. What function is the tangent plane the graph of?
Answer: Solve for z to get the equation of the plane in the form z = (x, y); this gives
(x, y) = f(x
0
, y
0
) +
_
f
x
_
0
(x x
0
) +
_
f
y
_
0
(y y
0
).
(We also replaced z
0
by f(x
0
, y
0
).)
The following example was not done in class.
Question 6.5. Suppose that f(x, y) = xy
2
5y. What is the tangent plane to the graph of
f at (2, 3, 3)?
24
Solution: We compute
f
x
= y
2
f
y
= 2xy 5
_
f
x
_
0
= 9
_
f
y
_
0
= 7,
z
0
= f(2, 3) = 3
so the tangent plane is
z 3 = 9(x 2) + 7(y 3).
6.4. The tangent plane approximation. If (x, y) is close to (x
0
, y
0
),
height of graph height of the tangent plane
f(x, y) f(x
0
, y
0
) +
_
f
x
_
0
(x x
0
) +
_
f
y
_
0
(y y
0
).
Equivalent formulation:
f
_
f
x
_
0
x +
_
f
y
_
0
y
f
x
x +f
y
y
(the partial derivatives f
x
and f
y
should be evaluated at (x
0
, y
0
)) The formula above tells
you approximately how much f changes in response to changes in x and y.
The approximation can be expected to be reasonably good if (x, y) is close to (x
0
, y
0
),
and the functions f
x
and f
y
are continuous in a neighborhood of (x
0
, y
0
), so that they dont
change too suddenly.
Question 6.6. A point P in R
2
is near (12, 5), but its coordinates could be o by as much
as 0.1 each. How far is P from the origin? Estimate the maximum error.
There was not enough time in class to complete the solution below.
Solution: Let f(x, y) =
_
x
2
+y
2
. Then f
x
=
x

x
2
+y
2
and f
y
=
y

x
2
+y
2
, so
f(x, y) f(12, 5) +f
x
(12, 5) x +f
y
(12, 5) y
13
12
13
x +
5
13
y,
which is 13 with maximum error of absolute value

12
13
(0.1) +
5
13
(0.1) 0.13.
25
(The worst error occurs when (x, y) = (0.1, 0.1) or (x, y) = (0.1, 0.1).)
MIDTERM 1 COVERS UP TO HERE
7. February 22
7.1. Maximum and minimum.
7.1.1. Critical points.
Denition 7.1. A critical point for f(x, y) is a point (a, b) such that f
x
(a, b) = 0 and
f
y
(a, b) = 0.
Question 7.2. Which of the following is true, if f(x, y) has a critical point at (a, b)?
(1) f(x, y) is a constant function
(2) The tangent plane to the graph of f(x, y) at (a, b, f(a, b)) is horizontal.
(3) There is no tangent plane to the graph of f(x, y) at (a, b, f(a, b)).
Answer: (2).
The function f(x, y) doesnt have to be constant; for example x
2
+y
2
has a critical point
at (0, 0), but it isnt constant.
Tangent plane at the point above (a, b):
z f(a, b) = f
x
(a, b)(x a) +f
y
(a, b)(y b),
If (a, b) is a critical point, then f
x
(a, b) and f
y
(a, b) are 0, so the tangent plane is z f(a, b) =
0, which is horizontal (its height z is constant).
Question 7.3. What are the critical points of f(x, y) := x
4
+y
3
3y?
Solution: Compute the partial derivatives, and set them equal to 0:
4x
3
= 0
3y
2
3 = 0.
Solving this system gives x = 0 and y = 1. So the critical points are (0, 1) and (0, 1).
7.1.2. Global and local extrema.
Denition 7.4. The function f(x, y) has a global maximum at the point (a, b) if f(x, y)
f(a, b) for all points (x, y) in the domain of f (i.e., for all points (x, y) where f(x, y) is
dened).
Global minimum is similar.
Example 7.5. If f(x, y) = (x y)
2
+ 5, then every point along the line x = y is a global
minimum.
26
Denition 7.6. The function f(x, y) has a local maximum at the point (a, b) if f(x, y)
f(a, b) at all points (x, y) suciently near (a, b) (but maybe there are other points farther
away from (a, b) where the value of f is even larger).
Local minimum is similar.
Every global max is automatically a local max.
Drew a level curve diagram, waved the chalk over it, and asked the students to yell
whenever it crossed a local min or local max.
Theorem 7.7. Every local max (or local min) is a critical point, assuming that the partial
derivatives exist at the point being tested.
Proof. The point has to be a local max for the partial functions obtained by plugging a
number into one of the variables, so by one-variable calculus, the derivatives of the partial
functions are 0 there if they exist.
Question 7.8. True or false: Every critical point of a function f(x, y) is either a local min
or a local max. (Hint: What happens for functions of 1 variable?)
Answer: False. Here are two counterexamples.
(1) f(x, y) := x
3
has a critical point at (0, 0), but it is not a local max (because there
are nearby points to the right where f(x, y) > 0), and not a local min (because there
are nearby points to the left where f(x, y) < 0).
27 8 1 0 1 8 27
27 8 1 0 1 8 27
27 8 1 0 1 8 27
27 8 1 0 1 8 27
27 8 1 0 1 8 27
27 8 1 0 1 8 27
27 8 1 0 1 8 27
(2) f(x, y) := x
2
y
2
has a critical point at (0, 0), but it is neither a local max nor a
local min, because there are nearby points (on one axis or the other) where the value
is larger or smaller than 0.
0 5 8 9 8 5 0
5 0 3 4 3 0 5
8 3 0 1 0 3 8
9 4 1 0 1 4 9
8 3 0 1 0 3 8
5 0 3 4 3 0 5
0 5 8 9 8 5 0
27
7.1.3. Solving unconstrained max/min problems.
As a warmup for the general method, consider the following 1-variable examples:
The global min of f(x) := x
2
on R occurs at x = 0 where the derivative f

(x) = 2x
becomes zero.
The function f(x) := [x[ on R has a global min at x = 0, where f

(x) is undened.
Consider the function f(x) := x
3
3x on R. Its derivative f

(x) = 3x
2
3 becomes
0 at x = 1. But neither x = 1 nor x = 1 is a global max; in fact, f(x) + as
x +, so a global max does not exist.
Consider f(x) := x
3
3x again, but this time restricted to the interval [10, 10]. As
before, f

(x) = 0 if and only if x = 1. We have f(1) = 2 and f(1) = 2, but

the global max on [10, 10] actually occurs on the boundary of the interval, where
f(10) = 970.
General method for solving a max/min problem (without constraint equations):
1. Identify the function f(x, y) to be maximized.
2. Identify the domain 1 on which f is to be maximized.
Easy case: only constraint inequalities or no constraints, so that the dimension
of 1 equals the number of variables, which for this function is 2. (Example:
constraint inequalities x, y > 0 dening the rst quadrant.)
Hard case: constraint equation(s), so that the dimension of 1 is less than the
number of variables. (Example: constraint equation x
2
+ y
2
= 1 dening the
unit circle, which is of dimension only 1.)
Assume that we are in the easy case. (In the hard case, a more sophisticated
method is needed: Lagrange multipliers.)
3. Check all of the following to nd potential maxima:
A. critical points in 1 (points where f
x
= 0 and f
y
= 0 simultaneously)
B. points in 1 where f
x
or f
y
is undened
C. behavior at (what happens to f as (x, y) in 1?)
D. boundary behavior (if there are constraint inequalities) this may lead to
another, lower-dimensional max/min problem.
The global max (a, b) is to be found among these points, so evaluate f at these
points to nd the maximum value. (Warning: If the values of f become larger and
larger as (x, y) approaches the boundary or , then the global max does not exist.)
The method also works for nding max/min of functions in 3 variables on a 3-dimensional
region, and so on.
Terminology: the global maximum is the location (a, b), but the maximum value is f(a, b).
In any max/min problem, you will need to determine which is being asked for.
Example 7.9. Find the point on the surface xyz
2
= 2 closest to the origin.
28
Solution: There is a constraint equation (3 variables, but only 2-dimensional domain), but
fortunately we can eliminate z by solving for z.
Exploit symmetry: The surface lies in the regions where xy > 0, which means that x, y > 0
or x, y < 0. Because of the symmetries
(x, y, z) (x, y, z)
(x, y, z) (x, y, z)
that preserve the equation of the surface, its enough to consider the parts with x, y > 0 and
z > 0: this is the set of points of the form
_
x, y,
_
2
xy
_
for x, y > 0.
1. What function do we want to minimize? Shortcut: The point where distance is minimized
is the same as the point where distance
2
is minimized; lets use distance
2
since it has a
simpler formula, namely
f(x, y) := x
2
+y
2
+
2
xy
.
2. What is the region 1? Since we are considering x, y > 0 only, 1 is the rst quadrant.
(Its 2-dimensional, so we dont need Lagrange multipliers.)
3. A. Critical points: solve
f
x
= 2x
2
x
2
y
= 0
f
y
= 2y
2
xy
2
= 0.
These lead to
x
3
y = 1
xy
3
= 1,
which imply x
3
y = xy
3
, and we may divide by xy to get x
2
= y
2
, so x = y or x = y.
Since x, y > 0, we must have x = y. Then x
4
= 1, and x > 0 so x = 1. Thus the only
critical point is (x, y) = (1, 1).
B. Points in 1 where f
x
or f
y
is undened: none. (The points where x = 0 or y = 0 are
not part of 1.)
C. Behavior as (x, y) : as x or y grows, the function f(x, y) tends to + (because
of the x
2
and y
2
terms in f(x, y)), so its not approaching a minimum out there.
D. Boundary behavior: As x 0 (from the right, while y is bounded), the function
f(x, y) tends to + (because of the
2
xy
term in f(x, y)), so its not approaching a
minimum there. Same if y 0.
29
Conclusion: f(x, y) is minimized at (x, y) = (1, 1), and z =

2 there, so the point is

(1, 1,

2). The other symmetric points are (1, 1,

2) and (1, 1,

2) and (1, 1,

2).
(And the minimum value of the distance is
_
D(1, 1) =

1 + 1 + 2 = 2.)
The following good question came from a student:
Question 7.10. The point (1, 1,

2) we computed seems to be the point with minimum

height, closest to the xy-plane (among those on the piece of the surface above 1), but how
do we know that there isnt another point on the surface that is even closer to the origin?
Answer: The function we minimized was not the z-coordinate of a point on the surface,
but the squared distance to the origin, so the point we computed is truly the one closest to
the origin.
Another way of saying this: we found the lowest point on the graph z = f(x, y) above 1,
but this graph is dierent from the piece of the original surface, which is given by z =
_
2
xy
.
If we had wanted the point closest to the xy-plane, we would have instead minimized the
z-coordinate, which is given by the function
_
2
xy
; in that case, the minimum does not exist
since
_
2
xy
can be made arbitrarily close to 0 by taking x and y to be large positive numbers.
8. February 28
8.1. Least squares interpolation. Problem: Given data points (x
1
, y
1
), . . . , (x
n
, y
n
) that
approximately lie on an unknown line y = ax +b, nd the line.
What are the unknowns here? a and b!
Given a candidate line y = ax + b, how do we measure how good of an approximation it
is? For each input x
i
, the line predicts an output of ax
i
+b, but the actual output was y
i
, so
the error in the prediction is [y
i
(ax
i
+ b)[. Then the total error from all the data points
would be
n

i=1
[y
i
(ax
i
+b)[,
and we want to nd a, b that make this small. Because of the absolute values, the partial
derivatives of this function do not exist everywhere, which complicates the minimization
problem, so instead we try to minimize the sum of the squares of the errors.
Denition 8.1. The least squares line (the best line) is the one for which
D :=
n

i=1
(y
i
(ax
i
+b))
2
is minimum.
30
This D is a function D(a, b).
The minimum occurs where
D
a
and
D
b
are both 0. Instead of expanding out D, use the
chain rule:
D
a
=
n

i=1
2(y
i
(ax
i
+b))(x
i
) = 0
D
b
=
n

i=1
2(y
i
(ax
i
+b))(1) = 0.
This is a system of two linear equations in a and b (remember: the x
i
and y
i
are given
numbers). Solving for a and b gives the best line.
The same method can be used to approximate data points by the graph of an unknown
quadratic function y = ax
2
+bx +c.
How do you know what kind of function to use? Maybe the source of the data suggests
that a particular shape of function is the right answer. Or maybe the data themselves look
as if they can be tted with a parabola, or . . . .
If you suspect that a function y = cx
d
is best, plot ln x versus ln y so that the relationship
is linear again:
ln y = C +d ln x,
where C := ln c and d are constants to be solved for. In other words, nd the line that best
approximates the points (ln x
i
, ln y
i
); this gives C and d.
8.2. Second derivative test for two-variable functions. Examples of critical point be-
havior at (0, 0):
x
2
+y
2
has local min
x
2
y
2
has local max
x
2
y
2
has saddle point
x
3
has none of the above
Example 8.2. f(x, y) = x
2
+ 6xy + 11y
2
. Complete the square: its (x + 3y)
2
+ 2y
2
, so it
has a local min at (0, 0).
More generally: consider f(x, y) =
1
2
Ax
2
+Bxy +
1
2
Cy
2
. (Reason for the
1
2
: it makes
f
xx
= A, f
xy
= f
yx
= B, f
yy
= C
at (0, 0).) Lets suppose that A ,= 0. Completing the square rewrites the function as
f(x, y) =
A
2
_
x +
B
A
y
_
2
+
_
AC B
2
2A
_
y
2
.
31
Case Conclusion
AC B
2
> 0 and A > 0 local min
AC B
2
> 0 and A < 0 local max
AC B
2
< 0 saddle point
AC B
2
= 0 inconclusive
Inconclusive means that it could be anything: a local min, a local max, a saddle point, or
something weirder.
The formula AC B
2
can be remembered as a determinant:
AC B
2
=

A B
B C

f
xx
f
xy
f
yx
f
yy

evaluated at (a, b).

Even more generally:
Second derivative test: Suppose that f(x, y) has a critical point at (a, b) (i.e., f
x
(a, b) =
0 and f
y
(a, b) = 0), and that all second derivatives exist and are continuous in a neighborhood
of (a, b). Dene
A := f
xx
(a, b), B := f
xy
(a, b), C := f
yy
(a, b).
Then the type of the critical point is given by the table above. This works even when A = 0.
Warning 8.3. Its not OK to use the second derivative test when there is a constraint equa-
tion!
8.2.1. An example.
Question 8.4. What kind of critical point does
f(x, y) := xy(x y) = x
2
y xy
2
have at (0, 0)?
Calculate all the partial derivatives up to second order:
f
x
= 2xy y
2
f
y
= x
2
2xy
f
xx
= 2y
f
xy
= f
yx
= 2x 2y
f
yy
= 2x.
32
Since f
x
and f
y
are 0 at (0, 0), there is a critical point there. Next,
A := f
xx
(0, 0) = 0
B := f
xy
(0, 0) = 0
C := f
yy
(0, 0) = 0.
Since AC B
2
= 0, the second derivative test is inconclusive.
So what next? The lines x = 0, y = 0 and x y = 0 divide the plane into six regions
on which f(x, y) is alternately positive or negative. So f(x, y) has neither a local min nor a
local max at (0, 0). In fact, it has what is called a monkey saddle: there are three negative
regions, two for the legs and one for the tail. (A monkey was drawn.)
Challenge: Find a surface that has an octopus saddle!
9. March 1
9.1. More on the tangent plane approximation.
Question 9.1. What is the linear polynomial in the coecients a, b that best approximates
the x-coordinate of the solution to the system
ax +by = 1
2x +y = 0
for (a, b) near (7, 3)?
Solution: The variables that are changing are a and b, and the function f(a, b) to be
approximated is the x-coordinate of the solution. We can nd f(a, b) explicitly by solving
the system: solve the second equation for y to get y = 2x, and plug back into the rst
equation to get ax +b(2x) = 1, so x = 1/(a 2b). Thus
f(a, b) :=
1
a 2b
f
a
=
1
(a 2b)
2
f
b
=
2
(a 2b)
2
The linear polynomial that best approximates this for (a, b) near (7, 3) is
f(7, 3) +f
a
(7, 3)(a 7) + f
b
(7, 3)(b 3) = 1 + (1)(a 7) + 2(b 3)
= 2 a + 2b.
33
So, for example, the solution x to the system
7.1x + 3.2y = 1
2x +y = 0
is going to be approximately
1 + (1)(0.1) + 2(0.2) = 1.3.
Approximation can be done for functions of more than two variables too.
Question 9.2. What is the best linear polynomial approximation to a 3-variable function
f(x, y, z) for inputs near a starting point (x
0
, y
0
, z
0
)?
As one moves from (x
0
, y
0
, z
0
) to (x, y, z), the approximation formula says that f is
caused by x, y, z, each magnied by the corresponding partial derivative:
f
change in f
(f
x
)
0
x + (f
y
)
0
y + (f
z
)
0
z.
Thus
f(x, y, z) = f(x
0
, y
0
, z
0
)
starting value
+ f
change in f
f(x, y, z) f(x
0
, y
0
, z
0
) + (f
x
)
0
(x x
0
) + (f
y
)
0
(y y
0
) + (f
z
)
0
(z z
0
).
9.2. Dierentials. Recall: if u = f(x), then one writes du = f

(x) dx. Here du and dx are

dierentials: they are not numbers, vectors, or matrices, but instead a new kind of object.
If f(x, y) is a 2-variable function, its total dierential is
df := f
x
dx +f
y
dy.
It helps us remember two formulas:
(1) The approximation formula
f f
x
x +f
y
y.
(2) The chain rule: if x = x(t) and y = y(t), then we divide by dt to get
df
dt
= f
x
dx
dt
+f
y
dy
dt
.
This can also be written as
df
dt
=
f
x
dx
dt
+
f
y
dy
dt
.
34
More complicated version: Suppose that Q = Q(u, v, w) where u = u(x, y), v = v(x, y),
and w = w(x, y). (Drew dependency diagram as in book.) Then
Q
x
=
Q
u
u
x
+
Q
v
v
x
+
Q
w
w
x
and
Q
y
is similar.
9.3. Using the chain rule to derive the product and quotient rules.
Example 9.3 (Product rule). If f = uv where u = u(x) and v = v(x), then
df
dx
= f
u
du
dx
+f
v
dv
dx
.
In other words,
d
dx
(uv) = vu

+uv

.
The quotient rule can also be obtained this way.
9.4. Review: denitions of cos and sin. What is the meaning of cos and sin ?
Denition 9.4. Draw a unit circle centered at the origin. Let P be the point reached by
going units counterclockwise from (1, 0). (If is negative, this means going clockwise.)
Then
cos := x-coordinate of P
sin := y-coordinate of P.
Example 9.5. cos = (x-coordinate of (1, 0)) = 1.
9.5. Review: polar coordinates. A point P in the plane can be specied in rectangular
coordinates x, y or in polar coordinates r, . Here r means the distance OP to the origin,
and means the angle measured counterclockwise from the positive x-axis to the ray

OP
(the value of is not completely determined by P, since one can 2 as many times as one
wants; to make unique, require it to be in a certain range, usually [0, 2) or (, ]).
Then
P = x, y) = rcos , sin ),
and we get the conversion formulas
x = r cos
y = r sin .
for going from r, to x, y. To go the other way, from x, y to r, , use
r =
_
x
2
+y
2
35
and then solve for in one of the two equations above, making sure that it species a point
in the correct quadrant so that the other equation is not o by a sign.
Question 9.6. At a time when a particle is at (4, 3) and has velocity vector 3, 1) (in
rectangular coordinates), what is
dr
dt
?
Solution: View r = r(x, y) and x = x(t), y = y(t). Namely, start with r =
_
x
2
+y
2
, and
apply the chain rule
dr
dt
=
r
x
dx
dt
+
r
y
dy
dt
.
Here
r
x
=
1
2
(x
2
+y
2
)
1/2
(2x) =
x
_
x
2
+y
2
r
y
=
1
2
(x
2
+y
2
)
1/2
(2y) =
y
_
x
2
+y
2
,
which at (4, 3) are 4/5 and 3/5. Also, at the given time
dx
dt
,
dy
dt
) = 3, 1), so
dr
dt
=
r
x
dx
dt
+
r
y
dy
dt
=
4
5
(3) +
3
5
(1)
=
9
5
.
Alternative solution (avoiding square roots): Start with
r
2
= x
2
+y
2
.
Apply
d
dt
:
2r
dr
dt
= 2x
dx
dt
+ 2y
dy
dt
.
At the given time, r = 5, so this becomes
2(5)
dr
dt
= 2(4)(3) + 2(3)(1),
which again leads to
dr
dt
=
9
5
.
36
10. March 5
10.1. Gradient.
Denition 10.1. The gradient of a scalar function f(x, y, z) is the vector-valued function
f :=
_
f
x
,
f
y
,
f
z
_
.
It is kind of like a derivative. Similar denition for functions in any number of variables.
Example 10.2. Let f(x, y, z) = x
2
y + 7z, and let P = (2, 3, 5). What is f(P)?
Solution: f = 2xy, x
2
, 7), so its value at (2, 3, 5) is 12, 4, 7).
We can restate certain theorems and denitions in terms of :
Chain rule: If f = f(x, y, z) and r(t) = (x(t), y(t), z(t)), then
d
dt
f(r(t)) =
f
x
dx
dt
+
f
y
dy
dt
+
f
z
dz
dt
= f(r(t)) r

(t).
Critical points of f(x, y): These are points P where f(P) = 0.
Geometric property of :
Theorem 10.3. If r(t) is any parametric curve on which f is constant, then at each time
t, the vector f(r(t)) is perpendicular to the tangent vector (velocity vector) r

(t).
Proof. For some constant c,
f(r(t)) = c.
Take
d
dt
:
f(r(t)) r

(t) = 0.
So f(r(t)) is perpendicular to r

(t).
In 2D, this says that f(P) is perpendicular to the level curve f(x, y) = c through P.
In 3D, this says that f(P) is perpendicular to the level surface f(x, y, z) = c through P.
Proof: f(P) is perpendicular to the tangent vector to any curve in the level surface.
Examples 10.4.
If f(x, y, z) = 2x + 3y + 5z, then at every point P in R
3
, the vector f = 2, 3, 5) is
perpendicular to the level surface 2x + 3y + 5z = c through P (which is a plane).
If f(x, y) = x
2
+ y
2
, then f = 2x, 2y), which points out from the origin, perpen-
dicular to the level curves (which are circles).
37
10.2. Tangent plane. Before we talked about the tangent plane to a graph z = f(x, y).
Now we can talk about the tangent plane to almost any surface:
Denition 10.5. The tangent plane to the surface f(x, y, z) = c at P = (x
0
, y
0
, z
0
) is the
plane with normal vector f(P) through P:
f
x
(P)(x x
0
) +
f
y
(P)(y y
0
) +
f
z
(P)(z z
0
) = 0.
(This makes sense only if f(P) ,= 0.)
Example 10.6. If f(x, y, z) = x
2
+y
2
z
2
then f = 2x, 2y, 2z), so the cone x
2
+y
2
z
2
=
0 has a tangent plane at each point except (0, 0, 0). At P = (3, 4, 5), the tangent plane is
6(x 3) + 8(y 4) 10(z 5) = 0.
10.3. Directional derivatives. Recall: Given f(x, y), the partial derivatives f
x
and f
y
measure the rate of change of f as one moves in the direction of i or j.
More generally:
Denition 10.7. Let u be any unit vector. The directional derivative of f(x) in the direction
u is
D
u
f(x) :=
d
ds
f(x +su)

s=0
= lim
s0
f(x +su) f(x)
s
.
This is because r(s) := x +su describes the position of a point starting at x and moving
at speed 1 in the direction u.
Why s, and not t? Answer: speed is 1, so distance traveled s equals time t.
Alternative notation:
df
ds

u
(P). Sometimes also
df
ds

P
or just
df
ds
(this presumes that P and
u are understood).
Question 10.8. If f(x, y) := x
2
+y
2
and x = 6, 8), in which directions u is the directional
derivative maximum?
Possible answers:
(1) i
(2) j
(3) (3/5, 4/5)
(4) (3/5, 4/5)
(5) (4/5, 3/5) or (4/5, 3/5).
(6) None of these.
38
Theorem 10.9 (Formula for calculating directional derivatives).
D
u
f(x) = f(x) u.
Proof.
D
u
f(x) =
d
ds
f(x +su)

s=0
= f(x +su)
d
ds
(x +su)

s=0
= f(x) u.

Given f and x, we maximize D

u
f(x) by choosing u to point in the direction of f(x).
And if u is so chosen, then D
u
f(x) = [f(x)[, by the geometric interpretation of the dot
product of two vectors forming a zero angle. Hence:
Geometric interpretation of f:
direction of f = direction in which f is increasing the fastest (perpendicular to
level curve/surface)
length of f = directional derivative of f in that direction
Example 10.10. Back to f(x, y) := x
2
+ y
2
and P = (6, 8). We have f(x, y) = 2x, 2y),
which at P is 12, 16), which is 20 times the unit vector u := 3/5, 4/5). So the function is
increasing the fastest in the direction of 3/5, 4/5) (and its rate of increase in that direction
is 20). So the answer to Question 10.8 is (3).
Question 10.11. In Question 10.8, in which directions is the directional derivative 0?
Solution: In the direction v := 4/5, 3/5) (tangent to the level curve), we have
f(x) v = 0,
so the directional derivative D
v
f(P) is 0. The same thing happens in the direction v =
4/5, 3/5). So the answer to Question 10.11 is (5).
11. March 7
11.1. Constraint equations vs. constraint inequalities.
(1) Find the minimum value of f(x, y) := 2x
2
7xy subject to the constraint inequalities
x 1, y 1, and x +y 4.
This is a standard max/min problem with boundaries to check.
39
(2) Find the minimum value of f(x, y, z) := x
2
+y
2
+z
2
subject to the constraint equation
3x + 5y +z = 9.
Eliminate z to get an equivalent problem: Find the minimum value of F(x, y) :=
x
2
+y
2
+ (9 3x 5y)
2
with no constraint equation.
(3) Find the minimum value of f(x, y, z) := x
2
+ y
2
subject to the constraint equation
e
x+y
= xy + 2.
When there is a constraint equation (and you cant or dont want to eliminate it),
you need Lagrange multipliers!
11.2. Lagrange multipliers.
Lagrange multipliers a method for nding max/min of f(x, y) when x and y are required
to satisfy a constraint g(x, y) = c (the hard case):
1. Identify the function f(x, y) to be maximized (or minimized).
2. Identify the constraint equation g(x, y) = c and constraint inequalities that dene
the domain 1 of f. Usually the presence of the constraint equation means that the
dimension of 1 is one less than the number of variables (but constraint inequalities
do not reduce the dimension).
3. Compute f and g.
4. Solve the system
g = c
f = g
in (x, y, ) to nd the possible pairs (x, y) (we dont care about the value of , so
a good strategy is to eliminate it).
5. Check also
A. points on g = c where g = 0
B. points on g = c where f or g is not dierentiable
C. behavior at (what happens to f as (x, y) approaches along the constraint
curve g = c?)
D. boundary points (if there are constraint inequalities)
The global max (a, b) is to be found among these points from steps 5 and 5, so
evaluate f at these points to nd the maximum value.
(Warning: If the values of f become larger and larger as (x, y) approaches the
boundary or , then the global max does not exist.)
Sample problem: Find the highest point on the intersection of the cylinder x
2
+ y
2
= 13
and the plane 2x + 3y z = 8 in R
3
.
40
Steps 1 and 2: Identify the function f(x, y) to be maximized/minimized, and the constraint
equation g(x, y) = c. (Also nd any constraint inequalities: these will be useful later, to
determine what boundaries need to be checked.)
We want to maximize z = 2x + 3y 8 subject to the constraint x
2
+ y
2
= 13. So take
f(x, y) := 2x + 3y 8 and g(x, y) := x
2
+y
2
. No constraint inequalities.
Step 3: Compute f and g.
We get f = 2, 3) and g = 2x, 2y).
Step 4: Solve the system
g = c
f = g
in (x, y, ) to nd the possible pairs (x, y) (we dont care about the value of , so a good
strategy is to eliminate it).
The system to be solved is
x
2
+y
2
= 13
2 = (2x)
3 = (2y).
Multiply the second equation by y and the third equation by x and equate to get
2y = 3x.
Solve for y and substitute into the rst equation:
x
2
+
_
3
2
x
_
2
= 13
13
4
x
2
= 13
x
2
= 4
x = 2.
So we get (x, y) = (2, 3) or (2, 3).
Step 5: Check also
A. points on g = c where g = 0
No such points, since g = 0 only at (0, 0), which is not on x
2
+y
2
= 13.
B. points on g = c where f or g is not dierentiable
No such points.
C. behavior at (what happens to f as (x, y) approaches along the constraint curve
g = c?)
Not applicable the constraint curve is bounded.
41
D. boundary points (if there are constraint inequalities)
Not applicable.
So the only points to check are (2, 3) and (2, 3). We have f(2, 3) = 5, and f(2, 3) =
21, so the maximum is at (x, y) = (2, 3). There, z = f(2, 3) = 5. So the highest point is
(2, 3, 5).
Why does this work? Why should f be a multiple of g at a maximum (when g ,= 0)?
Answer: First, the vector g is perpendicular to the level curve g = c. Claim: f is
perpendicular to g = c too. Reason: If not, then the directional derivative of f in one of
two directions along g = c would be positive, so one could increase f by moving in that
direction, meaning that we werent really at a maximum.
Since f and g are both perpendicular to the level curve, f must be a multiple of g
(if g ,= 0).
Lagrange multipliers apply also to max/min of functions of more than 2 variables.
(Lagrange multipliers also can be used when there is more than one constraint this is
discussed in the textbook, but we wont study such problems in this course.)
The second derivative test cannot be used when there is a constraint equation.
Sample problem 2: Find the point on the surface x
2
+y
2
= (z 1)
3
closest to the origin.
Solution: (We are going to make some mistakes below in red, and then correct them in
green.) We want to minimize f(x, y, z) := x
2
+y
2
+z
2
subject to the constraint g = 0 where
g(x, y, z) := x
2
+y
2
(z 1)
3
. We need to solve the Lagrange multiplier system
x
2
+y
2
(z 1)
3
= 0
2x, 2y, 2z) = 2x, 2y, 3(z 1)
2
),
which, when written out in components, says
x
2
+y
2
(z 1)
3
= 0
2x = (2x)
2y = (2y)
2z = (3(z 1)
2
).
Either the second or third equation implies = 1. Substituting = 1 in the last equation
leads to
2z = 3(z 1)
2
3(z 1)
2
+ 2z = 0
3z
2
4z + 3 = 0
42
but 4
2
4(3)(3) < 0, so there are no solutions.
Oops, we divided by either x or y to obtain = 1, so that was valid only if x ,= 0 or
y ,= 0. Thus to nish nding all solutions to the Lagrange multiplier system, We also need
to consider the case in which x = y = 0. Then the constraint equation implies z = 1, and
substituting this into the last of the four equations in the Lagrange multiplier system says
2 = (0)
which is impossible. So there is no minimum.
Oops, we forgot to check points where g = 0. This says
2x, 2y, 3(z 1)
2
) = 0, 0, 0),
which leads to (x, y, z) = (0, 0, 1). (The other parts of Step 5 do not give anything additional.)
By the geometry, there has to be a minimum somewhere, and (0, 0, 1) is the only candidate,
so (0, 0, 1) is the closest point.
The surface could be sketched by considering the slices obtained by intersecting with
horizontal planes z = c.
Make sure that you read the new notes on non-independent variables.
12. March 8
Topic for today: non-independent variables (partial derivatives of functions of variables
related by constraint equations).
12.1. Constraint equations and dimension. Part of the specication of a function is its
domain, the set of inputs on which it is being considered. Usually, each constraint equation
reduces the dimension of the domain by 1. This leads to the following:
Rule of thumb: Usually, if a domain dened by e constraint equations in n
variables, it will be (n e)-dimensional.
Example 12.1. The constraint equation
x + 2y + 3z = 5
denes a 2-dimensional domain (a plane).
Example 12.2. The constraint equations
x
2
+y
2
+z
2
= 100
x + 2y + 3z = 5
dene a 1-dimensional domain (a circle, arising as the intersection of a sphere and a plane).
43
Warning 12.3. It is not always true that each constraint equation reduces the dimension
by 1. (See the new notes on non-independent variables.)
Constraint inequalities usually do not aect the dimension.
12.2. Independent variables.
Example 12.4. On the upper half of the sphere x
2
+ y
2
+ z
2
= 1 in R
3
one can express z
in terms of x, y, namely
z =
_
1 x
2
y
2
.
Here we think of x, y as independent variables, and z depends on x, y. On the lower half one
would use
z =
_
1 x
2
y
2
.
instead.
Dimension can be understood as the number of independent variables.
Rule of thumb, rewritten: If a domain is dened by e constraint equations in
n variables, then
dimension = # independent variables
usually
= n e.
In this situation, then one can locally choose n e variables to be the inde-
pendent ones so that each of the remaining e variables can be expressed as a
function of the n e independent variables.
The precise mathematical statement along these lines is called the implicit function theorem;
it is discussed in more advanced math courses.
12.3. Derivatives.
Question 12.5. Consider the function f(x, y) := xy where (x, y) is constrained to lie on
the line 2x +y = 7. What is
df
dx
at the point (3, 1)?
(We wrote
df
dx
instead of
f
x
because there is only one independent variable: y depends
on x.)
Incorrect solution: The derivative of xy with respect to x is y, whose value at (3, 1) is 1.
What makes this wrong? It is true that if f(x, y) := xy on R
2
, then
f
x
= y. But the
denition of partial derivative assumes that it makes sense to hold y constant while varying
x, which is impossible if (x, y) is required to satisfy the constraint 2x+y = 7. If x is varying,
then y is not constant.
44
Correct solution 1 (elimination of dependent variable): How many independent variables?
Answer: 2 1 = 1. To compute
df
dx
, we need x to be the independent variable. Now use the
constraint equation to eliminate the dependent variable y and express everything in terms
of x:
y = 7 2x
f = x(7 2x) = 7x 2x
2
df
dx
= 7 4x,
and the value of
df
dx
at (x, y) = (3, 1) is 7 4(3) = 5.
In Correct Solution 1, we were lucky that it was easy to solve for y in terms of x. In more
complicated situations, this might not be possible, but one can still determine how quickly y
changes as x changes, by taking the dierential of the constraint equation. To see how this
works, lets solve the same problem again.
Correct solution 2 (dierentials): If f = xy is viewed as a function on R
2
, the denition
of df gives
(1) df = y dx +x dy.
This expresses how f changes as x and y change.
If f is restricted to a function on the domain dened by the constraint equation, then (1)
still holds, but now any change in x causes a change in y, so dx and dy are related. To nd
the relation, take the dierential of the constraint equation 2x +y = 7; this gives
2 dx +dy = 0
so
dy = 2 dx
(which makes sense since (x, y) is constrained to lie on the line 2x +y = 7 of slope 2). To
compute
df
dx
, we want to consider f as a function of the independent variable x alone, so
we should express df in terms of dx alone. To eliminate the dy term, substitute dy = 2 dx
into (1) to get
df = y dx +x(2 dx)
= (y 2x) dx.
This means that
df
dx
= y 2x.
At (3, 1), this is 1 2(3) = 5.
45
12.4. Partial derivatives. Now lets consider what happens in a question like Question 12.5
when there is more than one independent variable.
Question 12.6. Consider the function f(x, y, z) := x+y +x
2
z where (x, y, z) is constrained
to lie on the surface xyz = 6. What is
f
x
at the point (1, 2, 3)?
Answer: In the presence of the constraint equation xyz = 6, the notation
f
x
is meaning-
less, so the question does not make sense!
Here is why: Usually
f
x
means the rate of change of f as x varies while all the other
variables are held constant. But we cant hold both y and z constant while varying x, if we
want the constraint equation xyz = 6 to remain true.
Conclusions:
1. We are allowed to talk about partial derivatives of f only if f is expressed as a function
of independent variables (independence guarantees that we can vary one variable while
holding the others constant).
2. If f is initially expressed in terms of variables satisfying constraint equations, we must
choose some of the variables to be the independent ones, and view f and all other variables
as functions of the independent variables before talking about partial derivatives of f. The
notation for the partial derivatives must indicate which variables are being used as the
independent ones.
The notational convention is that all the independent variables are listed at the bottom
of the partial derivative notation, with the variables being held constant listed as subscripts
outside parentheses:
Denition 12.7. The notation
_
f
x
_
y
means that we are viewing f as a function of independent variables x and y, and measuring
the rate of change of f as x varies while holding y constant.
Question 12.8. Suppose that f(x, y, z) := x + y + x
2
z, where x, y, z are constrained to lie
on the surface xyz = 6. What is
_
f
x
_
y
at the point (1, 2, 3)?
Correct solution 1 (elimination of dependent variable): We can use the constraint equation
to eliminate z and express f in terms of independent variables x and y:
f = x +y +x
2
_
6
xy
_
= x +y +
6x
y
.
46
Then
_
f
x
_
y
= 1 +
6
y
,
so
_
f
x
_
y
at (1, 2, 3) is
1 +
6
2
= 4.
12.5. Partial derivatives and dierentials. Now lets solve the same problem using dif-
ferentials.
Correct solution 2 (dierentials): To compute
_
f
x
_
y
, in which the independent variables
are x and y, we need to express df in the form
df =? dx+? dy,
where each ? represents a function; then
_
f
x
_
y
is the rst ? (and
_
f
y
_
x
is the second ?).
But f is initially given as a function of dependent variables x, y, z. If f = x + y + x
2
z is
viewed as a function on R
3
, the denition of df gives
(2) df = (1 + 2xz) dx +dy +x
2
dz.
If f is restricted to a function on the domain dened by the constraint equation, then (2)
still holds, but taking the dierential of the constraint equation xyz = 6 gives a relation
between dx, dy, dz:
(3) yz dx +xz dy +xy dz = 0.
Because we want df in terms of dx and dy only, we solve (3) for dz,
xy dz = yz dx xz dy
dz =
z
x
dx
z
y
dy,
and substitute into (2):
df = (1 + 2xz) dx +dy +x
2
dz
= (1 + 2xz) dx +dy +x
2
_

z
x
dx
z
y
dy
_
= (1 + 2xz) dx +dy xz dx
x
2
z
y
dy
= (1 +xz) dx +
_
1
x
2
z
y
_
dy.
This means that
_
f
x
_
y
= 1 +xz,
47
and the value of
_
f
x
_
y
at (1, 2, 3) is 1 + 1(3) = 4.
12.6. Proving rules concerning partial derivatives. There are many rules relating dif-
ferent partial derivatives, but they all follow from the method of dierentials, so there is no
need to memorize the rules. The purpose of this section is not to list rules to be memorized,
but to give practice in using the method of dierentials.
Problem 12.9. The cyclic rule states that if variables x, y, z are related by a constraint
equation such that (as expected) any two of the variables may be taken as the independent
variables, then
_
x
y
_
z
_
y
z
_
x
_
z
x
_
y
= 1.
Prove this rule.
Proof. Let f(x, y, z) = 0 be the constraint equation, where f is a function that makes sense
on R
3
. Taking the dierential of the constraint equation gives
f
x
dx +f
y
dy +f
z
dz = 0,
where f
x
, f
y
, f
z
are the partial derivatives of f viewed as a function on R
3
(or at least a
3-dimensional part of R
3
). Solving for dx gives
(4) dx =
f
y
f
x
dy
f
z
f
x
dz,
which means that when x is viewed as a function of independent variables y, z, then
_
x
y
_
z
=
f
y
f
x
,
the coecient of dy in (4). A similar argument shows that
_
y
z
_
x
=
f
z
f
y
_
z
x
_
y
=
f
x
f
z
,
and multiplying all three gives
_
x
y
_
z
_
y
z
_
x
_
z
x
_
y
=
_

f
y
f
x
__

f
z
f
y
__

f
x
f
z
_
= 1.
48
13. March 12
13.1. Proving rules concerning partial derivatives, continued. Another example is
the two-Jacobian rule. To state it, we need a denition:
Denition 13.1. If u = u(x, y) and v = v(x, y), then the Jacobian of (u, v) with respect to
(x, y) is the function
(u, v)
(x, y)
:= det
_
u
x
u
y
v
x
v
y
_
= u
x
v
y
v
x
u
y
.
Here u
x
means
_
u
x
_
y
, and so on.
Two-Jacobian rule: If u, v, w, x, y are related by constraint equations such that any two of
the variables may be taken as the independent variables, then
_
u
v
_
w
=
(u, w)/(x, y)
(v, w)/(x, y)
.
(The right side could also be written out as
u
x
w
y
w
x
u
y
v
x
w
y
w
x
v
y
,
a ratio of determinants.)
How is the two-Jacobian rule used? It says that if one knows the partial derivatives of all
the variables with respect to independent variables x, y, then one can calculate the partial
derivatives using any other variables as the independent ones.
The two-Jacobian rule can be proved using dierentials.
13.2. Non-independent variables, continued: gradient in terms of new variables.
Problem 13.2. Suppose that u = u(x, y), and v = v(x, y). If u is viewed as a function of x
and v, what is
u :=
__
u
x
_
v
,
_
u
v
_
x
_
in terms of u
x
, u
y
, v
x
, v
y
?
Solution: We need to write
du =? dx+? dv.
Given:
du = u
x
dx +u
y
dy
dv = v
x
dx +v
y
dy.
49
Eliminate dy by solving for it in the second equation, and substituting it into the rst:
du = u
x
dx +u
y
_
dv v
x
dx
v
y
_
=
u
x
v
y
u
y
v
x
v
y
dx +
u
y
v
y
dv.
Thus the gradient of u viewed as function of x and v is

x,v
u =
_
u
x
v
y
u
y
v
x
v
y
,
u
y
v
y
_
.
13.3. Partial dierential equations.
Ordinary dierential equation (ODE): An equation involving derivatives of an unknown
function of one variable.
Example:
d
2
f
dx
2
5
df
dx
+ 6f = 0.
Here, f(x) = e
2x
is one solution.
Partial dierential equation (PDE): An equation involving partial derivatives of an unknown
function of several variables.
Example:
f
t
= t
2

2
f
x
2
+y
f
z
+f.
(You are supposed to solve for f = f(x, y, z, t).)
PDEs are generally very hard to solve. E.g., Navier-Stokes equations governing uid ow
there is a $1,000,000 prize for understanding the solutions.
Example 13.3 (An easy PDE). Find all f = f(x, y) satisfying

2
f
x
2
= 0.
(You dont get $1,000,000 for solving this one.)
Solution:

x
_
f
x
_
= 0,
so the function
f
x
is constant along each horizontal line; i.e.,
f
x
= g(y) for some function
g. So along each horizontal line, f grows at a constant rate (depending on the line). Thus
f(x, y) = g(y)x +h(y) for some functions g and h.
Conversely, given any functions g(y) and h(y), the function f(x, y) = g(y)x + h(y) is a
solution.
To have a unique solution, more information about f would need to be given.
50
Example 13.4 (Laplace equation). 1-dimensional case: Imagine an insulated metal rod with
ends held at dierent temperatures. Let x = position and w(x) = steady-state (equilibrium)
temperature at position x.
Physical heuristic: Once equilibrium is reached, the temperature of each tiny bit of rod
should equal the average of the temperatures of the neighboring bits, so w(x) should be a
linear polynomial ax +b. So
d
2
w
dx
2
= 0. (If there were a point where the graph were concave
up, the temperature would be increasing there; the reverse for concave down.)
2-dimensional case: Now imagine a metal plate with each point on the boundary held at
some temperature.

2
w
x
2
+

2
w
y
2
= 0 .
Notation:
:=
_

x
,

y
_

2
:=
:=

2
x
2
+

2
y
2
(Laplace operator)
Then the Laplace equation can be rewritten as
2
w = 0.
Example 13.5 (Heat equation). Same rod, not in equilibrium yet. Now w = w(x, t). Then
w
t
= k

2
w
x
2
.
for some constant k (depending on how well the rod conducts heat). E.g., if at a given time,
the temperature is concave up at a point, then that point will be warmed up by its neighbors.
In 2 dimensions:
w
t
= k
_

2
w
x
2
+

2
w
y
2
_
Example 13.6 (Wave equation). Imagine a rubber band whose shape is the graph of a
function w(x, t) (viewed as function of x at a xed time t). Then

2
w
t
2
= c
2

2
w
x
2
for a constant c, which represents the speed at which waves travel. E.g., if at a given time,
the graph is concave up at a point, then that piece of rubber band will be accelerated upward
by the pull of the neighboring pieces.
For this class, you are not required to memorize the named PDEs above, but you should
understand what it means for a function to be a solution to a PDE, and you should know
51
how to nd all the solutions f(x, y) to simple PDEs such as

2
f
x
2
= 0. The named PDEs
were presented just to give a few examples of how they come up in physics.
14. March 14
14.1. Double integrals. Let R be a region in R
2
, cut into tiny regions R
1
, . . . , R
n
. Choose
(x
1
, y
1
) in R
1
, . . . , (x
n
, y
n
) in R
n
. Then

R
f(x, y) dA f(x
1
, y
1
) Area(R
1
) + +f(x
n
, y
n
) Area(R
n
).
Its a number. (The actual denition involves a limit as the maximum size of the R
i
goes to
0.)
If f 0 everywhere on R, then

R
f(x, y) dA can be interpreted as the volume under the
graph of f (above R).
14.1.1. Computing double integrals as iterated integrals.
Suppose that R is a rectangle [a, b] [c, d]. Partition [a, b] into tiny subintervals, so R gets
sliced into thin rectangles, and the volume above it gets sliced into slabs. Then

R
f(x, y) dA =
_
x=b
x=a
__
y=d
y=c
f(x, y) dy
_
dx
=:
_
b
a
_
d
c
f(x, y) dy dx.
Think of the inner integral (with x treated as a constant) as the area of a slab; multiplying
it by the width dx of a slab gives the volume of the slab and we sum these (integrate)
to get the total volume. (In class, we formed a 5 4 grid of students, and integrated the
height function.)
Similarly,

R
f(x, y) dA =
_
d
c
_
b
a
f(x, y) dx dy.
14.1.2. Non-rectangular regions.
Problem 14.1. Let R be the bounded region between y x = 2 and y = x
2
. Find

R
(2x + 4y) dA.
Solution: Well compute it as

(2x + 4y) dy dx. What are the limits of integration?

Step 1: Sketch the region. Solving y x = 2 and y = x
2
shows that the line and the
parabola intersect at (1, 1) and (2, 4).
Step 2: The outer integral goes from the smallest x-coordinate of a point in R to the largest
x-coordinate.
52
The smallest x-coordinate is 1 and the largest x-coordinate is 2. So the outer integral
will look like
_
x=2
x=1
Step 3: Hold x xed, and increase y; look at the y-values where the line enters and leaves
R usually these depend on x. The inner integral goes from the smaller y-value to the
larger y-value.
It enters at y = x
2
and leaves at y = x + 2, so the iterated integral is
_
x=2
x=1
_
y=x+2
y=x
2
(2x + 4y) dy dx.
Then how do you evaluate the iterated integral?
Step 4: Evaluate the inner integral rst, treating x as constant. The result will be a
function of x.
Its
_
y=x+2
y=x
2
(2x + 4y) dy = 2xy + 2y
2

y=x+2
y=x
2
= (2x(x + 2) + 2(x + 2)
2
) (2x
3
+ 2x
4
)
= 2x
4
2x
3
+ 4x
2
+ 12x + 8.
Step 5: Then evaluate the outer integral, which is just a denite integral of a 1-variable
function of x.
Answer:
_
x=2
x=1
_
y=x+2
y=x
2
(2x + 4y) dy dx =
_
2
1
(2x
4
2x
3
+ 4x
2
+ 12x + 8) dx
=
333
10
.
14.1.3. Dividing the region into pieces. Sometimes to compute a double integral, the region
needs to divided into two or more pieces. The integral is then the sum of the integrals over
the pieces.
Example: rectangle with a smaller rectangle removed from its center.
14.1.4. Volume between two surfaces. If f(x, y) and g(x, y) are functions on a region R and
f(x, y) g(x, y) for all (x, y) in R, then the volume of the solid between the graphs of f
and g is

R
(f(x, y) g(x, y)) dA.
Problem (p. 960, #44): Set up an iterated integral that computes the volume of the region
bounded by the surfaces z = x
2
+ 3y
2
and z = 4 y
2
.
53
Solution: The intersection of the two surfaces is dened by the system
z = x
2
+ 3y
2
z = 4 y
2
.
The projection to the xy-plane of this is obtained by eliminating z:
x
2
+ 3y
2
= 4 y
2
.
So R is the region bounded by the ellipse
x
2
+ 4y
2
= 4.
Inside this ellipse the function 4 y
2
is larger than x
2
+ 3y
2
(as you can see by comparing
their values at (0, 0)). So we want

R
((4 y
2
) (x
2
+ 3y
2
)) dA =
_
1
1
_

44y
2

44y
2
(4 x
2
4y
2
) dx dy.
(Note: it is OK to write
_
1
1
instead of
_
y=1
y=1
, since the variable is determined by the
matching dierential: the integrals are nested, so the leftmost
_
corresponds to the rightmost
dierential dy, and so on.)
15. March 15
15.1. Double integrals, continued. Whats wrong with the following?
_
5
2
_
x+7
x+3
(x
2
+ 2y) dx dy
Since the outer dierential on the right is dy, this would mean
_
y=5
y=2
_
x=x+7
x=x+3
(x
2
+ 2y) dx dy,
but the inner limits make no sense. In the inner integral, y is constant, and the range for x
could depend on y, but the range for x cannot be expressed in terms of x!
15.2. Area in polar coordinates. Question: The area of the region described by r
[3, 3.2], [/4, /4 + 0.1] is closest to x/100 for which integer x?
Answer: 6.
Why? In general, the region with polar coordinates in [r, r + r] and [, + ] is
approximately a rectangle with sides r and r (the latter is the length of an arc of radius
r and measure , by denition of radian). So its area is approximately r r , which in
our case is 3(0.2)(0.1) = 0.06.
To remember:
dA = dx dy = r dr d.
54
15.3. Double integrals in polar coordinates.
Problem 15.1. Re-express
_

2
1
_

2y
2

2y
2
x
_
x
2
+y
2
dx dy
as an iterated integral in polar coordinates.
Solution: Substitute
x = r cos
y = r sin
dx dy = r dr d
but also re-compute the limits of integration.
The inequalities

_
2 y
2
x
_
2 y
2
are equivalent to
x
2
2 y
2
,
and to
x
2
+y
2
2.
Thus the region is dened by
x
2
+y
2
2
1 y

2.
In the circle of radius

2 centered at (0, 0), our region is the upper segment obtained by

cutting the circle with the chord from (1, 1) to (1, 1). (Draw it!)
So [/4, 3/4], and the upper limit for r is

2, but what is the lower limit for r? The

inequality y 1 becomes r sin 1 in polar coordinates, so r 1/ sin .
Final answer:
_
3/4
/4
_

2
1/ sin
(r cos )r r dr d.
15.4. Applications of double integrals.
15.4.1. Average value. Warmup: The average of numbers x
1
, . . . , x
n
is
x
1
+ +x
n
n
.
Denition 15.2. The average value of f(x, y) on a region R is

R
f dA
Area(R)
55
15.4.2. Mass and centroid. For a metal plate in the xy-plane,
mass = (mass per unit area
. .
density
)(area)
But if its density is not constant, = (x, y) (in g/cm
2
, say), then each piece of area must
be multiplied by the density there.
Denition 15.3. The mass of a 2-dimensional object is
m :=

R
(x, y) dA
. .
dm
.
The x-coordinate of the centroid is the average of the x-coordinates of the points in the
region, weighted by density. Same for the y-coordinate. So:
Denition 15.4. The centroid of a 2-dimensional object is the point ( x, y) where
x :=

R
x dm
m
=

R
x(x, y) dA

R
(x, y) dA
y :=

R
y dm
m
=

R
y(x, y) dA

R
(x, y) dA
.
(If density is not specied, assume (x, y) 1.) The centroid is also called the center of mass
or the center of gravity.
Sometimes symmetry gives a shortcut. For example, the centroid of an equilateral triangle
(of constant density) has reectional symmetry in each of its three altitudes, so the centroid
must lie on all three altitudes, hence at the center. Another example: A parallelogram has
180

rotational symmetry around the point where the two diagonals meet, so that point
must be the centroid.
15.4.3. Moment of inertia. The moment of inertia of an object with respect to an axis mea-
sures how dicult it is to rotate it (around that axis).
For a point mass:
I = (distance to axis)
2
m
In general (since dierent pieces are at dierent distances to the axis):
I =

R
(distance to axis)
2
dm.
Special case: The polar moment of inertia of an object in the xy-plane is its moment of
inertia around the z-axis. The distance from (x, y, z) to the nearest point on the z-axis
(namely, (0, 0, z)) is
_
x
2
+y
2
, so
(distance to axis)
2
= x
2
+y
2
in this special case.
56
Example 15.5. Polar moment of inertia of a triangle R of constant density 1 with vertices
at (3, 0) and (0, 2)?
Answer: Since density is 1, we have dm = dA. The polar moment of inertia is

R
(x
2
+y
2
) dA =
_
2
0
_
33y/2
(33y/2)
(x
2
+y
2
) dx dy =
(We didnt have time to explain how to get the limits of integration, but this is the same as
in earlier examples: draw the region, nd the range for the outer integration variable y, and
nd the range for x for each xed value of y.)
16. March 19
16.1. Applications of double integrals, continued.
16.1.1. Volume of revolution.
Theorem 16.1 (First theorem of Pappus). A plane region R lies on one side of an axis, and
is then rotated 360

around the axis. Let A = Area(R), and let D be the distance traveled by
the centroid of R. Then the solid of revolution has volume V = AD.
Why is this true? Set up a coordinate system so that the axis is the y-axis, and R is in the
half-plane x 0. A little rectangle of area dA sweeps out a ring of volume approximately
2x A since the ring can be cut into approximate rectangular parallelepipeds with base
dA and heights summing to the length of the ring (2x). Summing over all rectangles and
taking a limit as their size goes to 0 yields
V =

R
2x dA = (2A)

R
x dA
A
= A(2 x) = AD.
16.2. Review: gradient and directional derivative. Recall the geometric interpretation
of f:
direction of f = direction in which f is increasing the fastest (perpendicular to
level curve/surface)
length of f = directional derivative of f in that direction
Also recall the formula for computing a directional derivative (starting at x and moving
in the direction u):
D
u
f(x) = f(x) u.
Example 16.2. Suppose that f(x, y) = x
2
+ 5y.
What is (f)(6, 1) and what does it say geometrically? Well, f = 2x, 5), so (f)(6, 1) =
12, 5). This has length

12
2
+ 5
2
= 13 and direction

12
13
,
5
13
_
. This says that f is increasing
57
the fastest in the direction of

12
13
,
5
13
_
, and that the directional derivative (rate of change) in
that direction is 13.
We have f(6, 1) = 41. What is the nearest point to (6, 1) where f takes the value 41.26,
approximately? If from (6, 1) we move a distance in the direction of fastest increase, the
value of f increases by 13, so we need 13 = 0.26, so = 0.02. The point reached from
(6, 1) by moving 0.02 units in the direction of

12
13
,
5
13
_
is
6, 1) + 0.02
_
12
13
,
5
13
_
6, 1) + 0.02 1, 1/2) = 6.02, 1.01).
(The approximation of 5/13 by 1/2 was pretty crude, but good enough for blackboard work.)
What is the equation of the level curve through (6, 1)?
The value of f at (6, 1) is 41, so the level curve has equation x
2
+ 5y = 41.
What is the slope of the level curve at (6, 1)?
Solution 1: It is perpendicular to f, so the slope is 12/5.
Solution 2: The level curve is y =
1
5
x
2
+
41
5
, i.e., the graph of (x) :=
1
5
x
2
+
41
5
, and

(x) =
2
5
x, so the slope is

(6) =
12
5
.
In which direction(s) u is the directional derivative of f at (6, 1) equal to 56/5?
If u = a, b), then we need a
2
+b
2
= 1 (so that u is a unit vector), and (f)(6, 1)u = 56/5,
which says 12a+5b = 56/5. Solving this system (many algebra details skipped) leads to two
possibilities
_
3
5
,
4
5
_
,
_
837
845
,
116
845
_
.
In which direction(s) u is the directional derivative of f at (6, 1) equal to 14?
Solution: None, because the maximum directional derivative was 13!
In which direction(s) u is the directional derivative of f at (6, 1) equal to 13?
Solution: Since (f)(6, 1) had length 13, the only unit vector u that dots with it to give
13 is the unit vector in the opposite direction, namely
12
13
,
5
13
).
What does the graph of f look like?
The biggest slide ever.
58
16.3. Review: dierentials. When solving problems involving non-independent variables,
one usually needs to take the dierential of each constraint equation.
Problem 16.3. What is the result of taking the dierential of both sides of
ln w +e
uv
= 7 ?
Solution: Let L := ln w +e
uv
(the left side). Then
L
u
= e
uv
v
L
v
= e
uv
u
L
w
=
1
w
,
so by denition,
dL = ve
uv
du +ue
uv
dv +
1
w
dw.
Similarly (but much more easily), applying d to the right side gives
0 du + 0 dv + 0 dw = 0.
Thus the answer is the equation
ve
uv
du +ue
uv
dv +
1
w
dw = 0.
16.4. Review: non-independent variables.
Problem 16.4. Express
_
u
t
_
x
in terms of partial derivatives with respect to independent
variables u and v.
Solution using dierentials: For independent variables u and v, everything else is a function
of u and v, and
dt := t
u
du +t
v
dv
dx := x
u
du +x
v
dv.
To get
_
u
t
_
x
, re-express du in terms of dt and dx:
du =? dt+? dx
and take the coecient of dt. Lecture actually ended here, because time was up. To do
this, pretend that dt and dx are known and du and dv are unknowns, and solve for du by
59
eliminating dv. Namely, multiply the rst equation by x
v
and the second equation by t
v
,
and subtract:
x
v
dt = x
v
t
u
du +x
v
t
v
dv
t
v
dx = t
v
x
u
du +t
v
x
v
dv
x
v
dt t
v
dx = (x
v
t
u
t
v
x
u
) du
du =
x
v
x
v
t
u
x
u
t
v
dt + (who cares) dx
so
_
u
t
_
x
=
x
v
t
u
x
v
t
v
x
u
.
If the problem gave the function t(u, v) explicitly, then t
u
and t
v
could be computed as
explicit functions of u and v. If, moreover, the problem asked for the value of
_
u
t
_
x
at a
certain numerical point like (t, u, v, x) = (2, 3, 5, 7), then those could be plugged in at the
end.
MIDTERM #2 COVERS EVERYTHING UP TO HERE
17. March 22
17.1. Some quick clarications.
Question 17.1. Which of the following need to be checked when maximizing f(x, y) subject
to g(x, y) = c using Lagrange multipliers?
(1) solutions to the system
g(x, y) = c
f = g
(2) points where x = 0 or y = 0
(3) points on g = c where f = 0
(4) points on g = c where g = 0
(5) points where f does not exist
(6) points where g does not exist
(7) behavior of f as (x, y) along g = c
(8) boundary behavior
Answer: All of them except (2) and (3). (The only reason for handling x = 0 or y = 0 as
separate cases would be if you wanted to divide by x or y while solving.)
Question 17.2. What are the boundaries of the following domains?
60
The disk x
2
+y
2
1.
Answer: The circle x
2
+y
2
= 1.
The rst quadrant, where x 0 and y 0.
Answer: The nonnegative parts of the x- and y-axes.
The interval consisting of numbers x such that 3 x 5.
Answer: The two points x = 3 and x = 5.
The circle x
2
+y
2
= 1.
Answer: No constraint inequalities, so no boundary!
The arc of the circle x
2
+y
2
= 1 given by 0 /4.
Answer: The points (1, 0) (where = 0) and (

2/2,

2/2) (where = /4).

General rule of thumb: To nd the boundary, take one of the constraint inequalities and
change it to =; this gives one piece of the boundary. Then do this for each constraint
inequality to get all the pieces of the boundary.
In particular, if there are no constraint inequalities, there is no boundary.
17.2. Change of variables in double integrals. Recall how to do change of variables
(substitution) in one-variable integrals; e.g.,
_
1
1/2
1

1 x
2
dx =
_
/2
/6
1
_
1 sin
2
u
(cos u du) .
where
x = sin u
dx = cos u du.
The boundary points were re-expressed in terms of u:
x = 1 corresponds to u = /2
x = 1/2 corresponds to u = /6.
Today: The 2-variable analogue.
17.2.1. Transformations. (x, y) = f(u, v)
Example (from long ago): f(u, v) = (2u, v) (i.e., x = 2u, y = v) stretches in the horizontal
direction. Mr. Smiley becomes wider.
The transformation x = u, y = v 3u
2
does something worse. (Mr. Smiley is not smiling
anymore.) But if you had to do an integral over this region, you could convert it into an
integral over the original circle.
One way to visualize f: Plug in various input points (u, v) and see where they get mapped
to.
Another way: Draw images of the lines u = c and v = c.
61
Important: From now on, f should give a one-to-one transformation from a region S in
the uv-plane to a region R in the xy-plane. One-to-one means that dierent points inside
S get mapped to dierent points inside R. (Without this condition, the uv-integral might
double-count some parts of the xy-integral.)
17.2.2. Steps for changing variables.
Goal: Re-express

R
f(x, y) dx dy
as a double integral in new variables u, v.
(1) Choose a transformation x = x(u, v), y = y(u, v) in order to make the integrand
simpler or the region simpler.
(2) Find the equations of the boundary curves of R in terms of x, y.
(3) Rewrite these in terms of u, v to nd the corresponding region in the uv-plane.
(4) Compute the Jacobian determinant
(x, y)
(u, v)
= det
_
x
u
x
v
y
u
y
v
_
(Note: This assumes that x, y are given as functions of u, v. If for some reason you
have u, v as functions of x, y, and dont feel like solving for x, y in terms of u, v, then
rst compute
(u, v)
(x, y)
= det
_
u
x
u
y
v
x
v
y
_
and then use the identity
(x, y)
(u, v)
=
_
(u, v)
(x, y)
_
1
to compute the Jacobian.)
(5) Take the absolute value of the Jacobian determinant to get the area scaling factor:

(x, y)
(u, v)

.
Area is always nonnegative!
(6) Substitute
x = x(u, v)
y = y(u, v)
dx dy =

(x, y)
(u, v)

du dv.
62
into the integral, and remember to use the equations of the boundary curves in terms
of u and v to nd the new region, and hence the new limits of integration.
Remark 17.3. In the case of the change of variable
x = r cos , y = r sin
the area scaling formula
dx dy =

(x, y)
(u, v)

du dv.
turns out to be
dx dy = r dr d.
Problem 17.4. Let R be the square with corners (1, 0) and (0, 1). Evaluate
I :=

R
sin
2
(x y)
x +y + 2
dx dy.
Solution:
Its often a good idea to choose u, v so that the sides of R correspond to u = constant,
v = constant.
In this problem: Try u = x +y, v = x y, so x = (u +v)/2, y = (u v)/2.
Boundary curves: u = 1, u = 1, v = 1, v = 1.
Jacobian:
(x, y)
(u, v)
= det
_
1/2 1/2
1/2 1/2
_
= 1/2
so
dx dy =

1
2

du dv =
1
2
du dv.
So
I =
_
1
1
_
1
1
sin
2
v
u + 2
1
2
du dv.
To nish, one should evaluate the inside integral (with variable u), and then the outside
integral (with variable v).
17.2.3. Why is the scale factor equal to

(x, y)
(u, v)

? The rectangle [u, u+du] [v, v +dv] maps

to (approximately) a parallelogram whose sides are given by the vectors x
u
du, y
u
du) and
x
v
dv, y
v
dv). In other words, as the input moves du in the u-direction, the x-coordinate of
the output changes by approximately x
u
du, and so on, because x
u
is the rate of change of x
with respect to u (as v is held constant).
63
The area of the parallelogram spanned by those two vectors (which we now write as
columns of a matrix) is the absolute value of
det
_
x
u
du x
v
dv
y
u
du y
v
dv
_
= det
_
x
u
x
v
y
u
y
v
_
du dv.
(Here we used twice that if one multiplies an entire column of a matrix by a number, the
determinant gets multiplied by that number.) This explains why
dx dy =

(x, y)
(u, v)

du dv.
18. April 2
18.1. Review: change of variable in double integrals.
Problem 18.1. Let R be the region in the rst quadrant inside the ellipse x
2
+ 9y
2
= 12
but outside the circle x
2
+y
2
= 4. Evaluate

R
xye
y
2
1 y
2
dx dy.
Solution: Since x
2
and y
2
appear in the equations of the boundary curves, and since y
2
appears in e
y
2
in the integrand, lets use the change of variables u = x
2
and v = y
2
, so
x =

u and y =

v. The original region R is described by x

2
+ 9y
2
12 and x
2
+ y
2
4
and x, y 0. The corresponding region S in the uv-plane is described by u + 9v 12 and
u +v 4 and u, v 0; this is a triangle with vertices at (4, 0), (12, 0), and (3, 1). We have
(u, v)
(x, y)
= det
_
u
x
u
y
v
x
v
y
_
= det
_
2x 0
0 2y
_
= 4xy,
so
(x, y)
(u, v)
=
1
4xy
,
which is positive, so its absolute value is also
1
4xy
, which is the area scaling factor:
dx dy =
1
4xy
du dv.
Thus the integral becomes

S
xye
y
2
1 y
2
1
4xy
du dv =
1
4

S
e
v
1 v
du dv.
(We re-expressed everything in terms of u and v.) At this point, we have the option of
doing the integration in the order dv du instead of du dv, but the order dv du would require
dividing the triangle into two triangles, so it is easier to keep v as the outer variable. The
64
range for v is [0, 1]. For each v, we get the lower and upper limits for u by solving u +v = 4
and u + 9v = 12 for u in terms of v. This leads to
1
4
_
v=1
v=0
_
u=129v
u=4v
e
v
1 v
du dv =
1
4
_
v=1
v=0
e
v
1 v
u

u=129v
u=4v
dv
=
1
4
_
v=1
v=0
e
v
1 v
(8 8v) dv
= 2
_
v=1
v=0
e
v
dv
= 2 e
v
[
v=1
v=0
= 2e 2.
18.2. Vector elds.
Denition 18.2. A vector eld is a function whose value at each point of a region is a vector.
It could be showing, for example,
the wind velocity at each point,
the velocity of a uid at each point,
the strength and direction of an electromagnetic force,
or the gradient of a function at each point.
Mathematically, a 2-dimensional vector eld has the form
F(x, y) = P(x, y), Q(x, y))
= P(x, y) i +Q(x, y) j,
where P and Q are functions.
Example 18.3. Consider F(x, y) = yi. Its values in the upper half plane are vectors
pointing to the left, and its values in the lower half plane are vectors pointing to the right.
(The students in class demonstrated this by forming one big vector eld with their hands.)
Example 18.4. The gradient eld of f(x, y) := x
2
+ y
2
is F(x, y) := 2xi + 2yj, which is
always pointing outward. Think of the direction of fastest increase, looking at a top view of
a paraboloid. The gradient eld is everywhere perpendicular to the level curves, which are
circles.
F is continuous P and Q are continuous.
F is dierentiable P and Q are dierentiable.
65
18.3. Line integrals. Let C be an oriented curve in R
2
. (Oriented means that one of the
two directions along the curve has been chosen.)
Three kinds of line integrals:
Line integral with respect to arc length:
_
C
f(x, y) ds. (Remember: s = distance
traveled = arc length.)
Line integral with respect to coordinate variables:
_
C
f(x, y) dx or
_
C
f(x, y) dy or
_
C
P(x, y) dx +Q(x, y) dy
Line integral of a vector eld:
_
C
F dr.
Which are vectors and which are scalars? They are all scalars!
18.3.1. What do these mean? Divide C into n pieces by labelling points P
0
, P
1
, . . . , P
n
in
order along C where P
0
and P
n
are the endpoints. Choose a sample point P

i
on the arc
from P
i1
to P
i
. Let s
i
be the arc length of that arc. Then
_
C
f(x, y) ds f(P

1
)s
1
+ +f(P

n
)s
n
.
Actual denition:
_
C
f(x, y) ds := lim
max s
i
0
(f(P

1
)s
1
+ +f(P

n
)s
n
) .
Similarly,
_
C
f(x, y) dx := lim
max s
i
0
(f(P

1
)(x(P
1
) x(P
0
)) + +f(P

n
)(x(P
n
) x(P
n1
)))
and
_
C
F dr := lim
max s
i
0
_
F(P

1
)

P
0
P
1
+ +F(P

n
)

P
n1
P
n
_
.
Example 18.5.
_
C
1 ds = arc length of C.
18.3.2. How do you compute these?
(1) Choose a parametrization of C, say (x(t), y(t)) for t [a, b]. Now x, y, s, r are all
functions of t.
(2) Make the substitution x = x(t) and y = y(t) in the integrand.
(3) Also substitute
dx = x

(t) dt
dy = y

(t) dt
ds =
_
dx
2
+dy
2
=
_
x

(t)
2
+y

(t)
2
dt
dr = dx, dy) = x

(t), y

(t)) dt.
(4) Change
_
C
to
_
b
a
.
66
(5) Evaluate the resulting 1-variable integral in t.
Problem 18.6. Let C be the upper half of the circle x
2
+y
2
= 4, oriented counterclockwise.
Let F(x, y) = yi. Compute
_
C
F dr.
Is the answer going to be positive or negative? Positive, because F and r form an angle
less than 90

.
Solution: Choose the parametrization (2 cos t, 2 sin t) for t [0, ]. Then
_
C
F dr =
_
C
y, 0) dx, dy)
=
_

0
2 sin t, 0) 2 sin t, 2 cos t) dt
=
_

0
4 sin
2
t dt
=
_

0
(2 2 cos 2t) dt
= 2.
18.3.3. Geometric interpretation. Recall that dr = Tds = dx, dy), where T is the unit
tangent vector. So if F = Pi +Qj, then
_
C
F dr =
_
C
(F T) ds
=
_
C
P dx +Qdy.
18.3.4. Applications. If C is a wire whose density per unit length is (x, y), then
mass m :=
_
C
dm =
_
C
ds
mass-weighted average

f :=
_
C
f dm
m
centroid := ( x, y)
moment of inertia I :=
_
C
(distance to axis)
2
dm.
All of these involve line integrals with respect to arc length.
Also, given a force eld that is constant (independent of position),
work :=

force

displacement := F r,
More generally, for an object moving along C, the work done by a not necessarily constant
force eld F is
W :=
_
C
F dr.
67
(Think of adding up the work done over each little piece of C.)
18.4. Adding and reversing paths. If the endpoint of path C
1
equals the start point of
C
2
, then concatenating gives a path C
1
+C
2
, and
_
C
1
+C
2
=
_
C
1
+
_
C
2
.
Let C be a piecewise dierentiable oriented curve. Then C denotes the same curve with
the reverse orientation. Intuitively, for each piece of C, the length ds stays the same, but
the changes dx, dy, and dr change sign, so
_
C
f ds =
_
C
f ds
_
C
P dx +Qdy =
_
C
P dx +Qdy
_
C
F dr =
_
C
F dr.
19. April 4
19.1. Fundamental theorem of calculus for line integrals.
1
st
FTC: If f(x) is a continuous function on [a, b], and
F(x) :=
_
x
a
f(t) dt
then F

(x) = f(x).
2
nd
FTC: If G(x) is such that G

(x) is continuous on an interval [a, b], then

_
b
a
G

(t) dt = G(b) G(a).

Its the 2
nd
FTC that is usually used each time you compute a denite integral.
FTC for line integrals: If C starts at A and ends at B, and f(x, y) is a function that is
continuously dierentiable at each point of C, then
_
C
f dr = f(B) f(A).
The left side is adding up the change in f along each piece of the curve, and the right side
is the total change in f.
68
Proof. Let r(t) = (x(t), y(t)) for t [a, b] be a parametrization of C. Let G(t) = f(r(t)).
Then
_
C
f dr =
_
C
_
f
x
,
f
y
_
dx, dy)
=
_
C
f
x
dx +
f
y
dy
=
_
b
a
_
f
x
dx
dt
+
f
y
dy
dt
_
dt
=
_
b
a
G

(t) dt (by the chain rule)

= G(b) G(a) (by the 2
nd
FTC)
= f(r(b)) f(r(a))
= f(B) f(A).
Consequences:

_
B
A
f dr is path independent: depends only on the endpoints A and B, not on the
path taken from A to B!
If C is a closed curve (starts and ends at the same point), then
_
C
f dr = 0. (There
is no dierence between
_
and
_
: the former is just a notation to help remind you
that the curve is closed.)
These are properties special to gradient vector elds.
Example 19.1. Let F = y i. Let C be the upper half of the unit circle, and let C

be
the lower half, both oriented from (1, 0) to (1, 0). The geometry shows that
_
C
F dr is
negative, but
_
C

F dr is positive. So line integrals of F are not path independent.

19.2. Equivalent conditions.
Theorem 19.2. For a continuous vector eld F = P i + Qj on a region D, the following
are equivalent:
(1)
_
C
F dr = 0 for every closed curve C inside D.
(2)
_
B
A
F dr is path independent for all points A and B in D.
(3) F is a gradient vector eld (i.e., F = f for some dierentiable function f(x, y), which
is then called a potential function for F the multivariable analogue of an antideriva-
tive).
(4) P dx + Qdy is an exact dierential (i.e., has the form f
x
dx + f
y
dy = df for some
f(x, y)).
If any of these hold, F is called conservative.
69
The theorem is saying that if any one of the four conditions holds for F, then all of the
other conditions hold too.
Proof. (1) = (2): Suppose that
_
C
F dr = 0 for every closed curve C. We need to prove
path independence. So suppose that C
1
and C
2
are two paths from A to B. Then C
1
C
2
is a closed loop, so
0 =
_
C
1
C
2
F dr =
_
C
1
F dr
_
C
2
F dr.
So
_
C
1
F dr =
_
C
2
F dr.
(2) = (3): Suppose that line integrals of F are independent of path. Choose a start
point (a, b) in D. Dene
f(x, y) :=
_
(x,y)
(a,b)
F dr.
(This makes sense since the value is independent of the path.) We will show that f =
F. First, lets calculate f
x
(x
0
, y
0
). Choose a path from (a, b) to (x
0
, y
0
) and continue it
horizontally to (x, y
0
). Then
_
(x,y
0
)
(a,b)
F dr =
_
(x
0
,y
0
)
(a,b)
F dr +
_
(x,y
0
)
(x
0
,y
0
)
F dr
f(x, y
0
) = f(x
0
, y
0
) +
_
t=x
t=x
0
F(t, y
0
) i dt (using r(t) = (t, y
0
) so dr = dt, 0) = i dt)
= f(x
0
, y
0
) +
_
t=x
t=x
0
P(t, y
0
) dt
f
x
(x
0
, y
0
) = P(x
0
, y
0
) (by the single-variable 1
st
FTC).
So f
x
= P. Similarly f
y
= Q. So f = F.
Class actually ended here; well nish the proof next time.
(3) = (1): Suppose that F = f. Then for any closed curve C starting and ending at
A, the FTC implies that
_
C
F dr = f(A) f(A) = 0.
(3) (4): These are just dierent ways of saying that P = f
x
and Q = f
y
for some
dierentiable f(x, y).
Using the implications proved above, one can get from the truth of any one of the four
conditions to the truth of any other.
20. April 5
Last time:
FTC for line integral of a gradient: For a path C from A to B, and continuously dierentiable
70
f,
_
C
f dr = f(B) f(A).
20.1. Conservative force elds in physics. In physics, if F is a force eld (e.g., gravita-
tional force), and F = V then V is called potential energy. If an object moves from A to
B along a path C,
1
2
mv v

B
A
=
_
mv

(t) v(t) dt =
_
ma(t) r

(t) dt =
_
C
F dr
FTC
= V [
B
A
increase in kinetic energy = work done by the force on the object = decrease in potential energy,
so
kinetic energy + potential energy is constant:
conservation of energy! Thats why F is called conservative.
20.2. Test for gradient eld. Recall: A continuous vector eld F(x, y) = P(x, y), Q(x, y))
on a region D is conservative if one of the following equivalent conditions holds:
(1)
_
C
F dr = 0 for every closed curve C.
(2)
_
B
A
F dr is path independent for all points A and B.
(3) F = f for some f.
(4) P dx +Qdy is exact (equal to df for some f).
Example 20.1. Let F = P i + Qj := 2xy, x
2
), so P(x, y) = 2xy and Q(x, y) = x
2
. Then
F = f, where f(x, y) := x
2
y. In other words, condition (3) holds for F. The theorem now
says that all four conditions hold for F.
For example, (4) says that P dx+Qdy is exact. This is true: it is df, where f(x, y) := x
2
y.
Suppose that C is the unit circle, traversed counterclockwise. Is
_
C
2xy, x
2
) dr = 0? Yes,
since (1) is true for F.
Problem, part (a): Suppose that F(x, y) = (6x
2
+8xy)i +(4x
2
+3y
2
)j. Is F conservative?
To help with this, we introduce a new condition, assuming that F is continuously dier-
entiable:
(5) Q
x
= P
y
.
Conditions (1)(4) imply (5):
Proof: If P = f
x
and Q = f
y
, then Q
x
= f
yx
= f
xy
= P
y
.
Conversely, (5) implies (1)(4) if D is simply connected. Well talk about this later on.
For now: R
2
is simply connected, as is any rectangle.
71
Back to the problem:
P = 6x
2
+ 8xy
P
y
= 8x
Q = 4x
2
+ 3y
2
Q
x
= 8x.
Since Q
x
= P
y
, F is conservative.
20.3. Finding the potential.
Problem, part (b): Find a function f such that f = F.
Solution 1: use line integrals (FTC backwards).
Choose a path C from (0, 0) to (a, b). Then FTC says
f(a, b) f(0, 0) =
_
C
F dr.
The f(0, 0) is a constant of integration (if f is a potential for F, then so is f + c for any
constant c). Lets take C = C
1
+ C
2
where C
1
goes from (0, 0) to (a, 0) and C
2
goes from
(a, 0) to (a, b). Parametrize C
1
by (t, 0) for t [0, a]:
_
C
1
F dr =
_
a
0
6t
2
, 4t
2
) dt, 0)
=
_
a
0
6t
2
dt
= 2t
3

a
0
= 2a
3
and parametrize C
2
by (a, t) for t [0, b]:
_
C
2
F dr =
_
b
0
6a
2
+ 8at, 4a
2
+ 3t
2
) 0, dt)
=
_
a
0
4a
2
+ 3t
2
dt
= 4a
2
t +t
3

b
0
= 4a
2
b +b
3
72
so
f(a, b) f(0, 0) =
_
C
F dr
=
_
C
1
F dr +
_
C
2
F dr
= 2a
3
+ 4a
2
b +b
3
,
so
f(x, y) = 2x
3
+ 4x
2
y +y
3
(+constant).
There are innitely many possible answers f(x, y), one potential function for each value of
the constant.
Solution 2: use antiderivatives.
We want f(x, y) such that
f
x
= 6x
2
+ 8xy
f
y
= 4x
2
+ 3y
2
.
Using the rst equation, treating y as constant, we get
f(x, y) = 2x
3
+ 4x
2
y +g(y)
for some constant g(y) that does not depend on x. Take
d
dy
: get
f
y
= 4x
2
+g

(y).
Comparing with the second requirement shows that g

(y) = 3y
2
, so g(y) = y
3
+ c. Substi-
tuting this back in yields
f(x, y) = 2x
3
+ 4x
2
y +y
3
+c.
Whichever method was used, it is easy to check the answer: just compute f and make
sure that it equals F.
21. April 9
21.1. 2D Curl.
Denition 21.1. The (scalar) curl of a 2-dimensional vector eld F = Pi +Qj is
curl F :=

y
P Q

= Q
x
P
y
.
It is a scalar function of (x, y).
So the following are equivalent:
73
(5) Q
x
= P
y
(6) curl F = 0.
Roughly, failure of conservativeness is measured by curl F.
21.1.1. Physical interpretations of curl. If F is the velocity eld of a uid, and you nail
the center of a tiny paddlewheel at a point (see p. 1068 in the textbook, or V4.3 in the
supplementary notes), then
curl F = 2(angular velocity of paddlewheel).
Examples 21.2.
F = 2, 3) (constant ow) has curl F = 0
F = x, y) (expansion) has curl F = 0
F = y, x) (rotation with angular velocity 1) has curl F = 2
If instead F is a force eld, curl F is torque exerted on a tiny dumbbell (rotational analogue
of force). An analogy:

force
mass
=

acceleration :=
d
dt
(

velocity)
torque
moment of inertia
= angular acceleration :=
d
dt
(angular velocity).
21.2. Greens theorem. Suppose C is a simple closed curve bounding a region R. (Simple
means that C does not cross itself; for example, a gure-eight is a closed curve, but not a
simple closed curve.) Assume that C is positively-oriented: this means that as you walk along
the curve, the region is to the left.
Greens theorem: If F is a continuously dierentiable vector eld, then
_
C
F dr =

R
curl FdA
Equivalent formulation: For any continuously dierentiable functions P and Q,
_
C
P dx +Qdy =

R
(Q
x
P
y
) dA.
Sketch of proof. Two comments:
Its really two identities in one:
_
C
P dx =

R
P
y
dA and
_
C
Qdy =

R
Q
x
dA.
Lets prove the rst one (the proof of the other one is similar).
74
Additivity: If its true for R
1
and R
2
, its true for R := R
1
R
2
since if C
1
, C
2
, C are
the curves bounding R
1
, R
2
, R, then
_
C
=
_
C
1
+
_
C
2

R
=

R
1
+

R
2
.
(The overlapping parts of C
1
and C
2
are in opposite directions, so those parts of the
integral cancel.)
Case 1: R is vertically simple, i.e., the area between two graphs y = g
1
(x) and y = g
2
(x)
for x [a, b].
Then C = C
1
+ C
2
+ C
3
+ C
4
: the bottom, right, top, and left boundaries. To compute
the line integrals on these four curves, use the following parametrizations:
C
1
: t, g
1
(t)) for t from a to b
C
2
: b, t) for t from g
1
(b) to g
2
(b)
C
3
: t, g
2
(t)) for t from b to a
C
4
: a, t) for t from g
2
(a) to g
1
(a).
Substituting these in yields
_
C
1
P dx =
_
b
a
P(t, g
1
(t)) dt =
_
b
a
P(x, g
1
(x)) dx (its OK to rename t as x)
_
C
2
P dx = 0 (since x = b, dx = 0)
_
C
3
P dx =
_
a
b
P(t, g
2
(t)) dt =
_
b
a
P(x, g
2
(x)) dx
_
C
4
P dx = 0 (same as for C
2
),
so
_
C
P dx =
_
b
a
P(x, g
1
(x)) dx
_
b
a
P(x, g
2
(x)) dx
=
_
b
a
(P(x, g
2
(x)) P(x, g
1
(x))) dx
=
_
b
a
_
_
g
2
(x)
g
1
(x)
P
y
dy
_
dx
=

R
P
y
dA.
75
Case 2: R can be sliced with vertical lines into nitely many vertically simple pieces. Use
Case 1 and additivity.
Case 3: Any region R. Approximate R better and better with regions as in Case 2 and
take the limit of both sides.
21.3. Review: using line integrals.
Question 21.3. In line integrals, how do you know which of ds, dx, dr to use?
It depends on what you want the line integral to represent. Some guidelines:
If integrating a vector eld, usually dr is what you want. For example, work done by
a force eld F is
_
C
F dr.
For mass, average value, centroid, or moment of inertia of a wire, use dm = ds.
(Here ds is the length of a small piece of wire, is the mass per unit length, and dm
is the mass of the small piece of wire.)
22. April 11
22.1. Applications of Greens theorem.
22.1.1. Computing area.
Example 22.1. Taking P = 0 and Q = x in Greens theorem yields
_
C
x dy =

R
(1 0) dA = Area(R).
Similarly, taking P = y and Q = 0 yields

_
C
y dx =

R
(0 (1)) dA = Area(R).
In the 19th century, a mechanical device called a planometer was invented that could compute
such line integrals, so that areas could be computed by just having the arm of the device
trace the boundary.
22.1.2. Vector elds with curl equal to 0 everywhere. If C is a simple closed curve, and F is a
vector eld such that curl F = 0 not only on C but also everywhere in the region R bounded
by C, then
_
C
F dr = 0.
Proof:
_
C
F dr =

R
curl FdA = 0.
(The sign is there in case C is not positively-oriented.)
76
In fact, this works even if C is a closed curve that is not a simple closed curve. For example,
a gure-eight can be decomposed into two simple closed curves. (Even more general closed
curves can be approximated by nite unions of simple closed curves.)
This explains why if curl F = 0 on a region without holes (like R
2
or a rectangle), then F
is conservative. In terms of our numbering of conditions, this is (6) =(1).
22.2. Flux. Recall:
_
C
F dr =
_
C
F Tds
where T is the unit tangent vector at each point of C. Rotate T by 90

clockwise to get the

(outward) unit normal vector n (at each point of C).
Denition 22.2. The ux of F across a plane curve C is
_
C
F nds.
22.2.1. Physical interpretation of ux. Suppose that the velocity of a 2-dimensional uid of
unit density is given by a constant vector eld F.
How much uid crosses a tiny line segment of length ds perpendicular to F in one second?
A rectangles worth of uid crosses the segment, so the answer is [F[ ds.
What if the segment is a piece of C, but F is not perpendicular to C? Then the uid
that crosses the segment in one second is a parallelogram of height F n and base ds, so the
answer is F nds.
How much uid crosses the whole closed curve C per second? Answer:
_
C
F nds. Thats
ux! This works even if F is not constant, but just continuous (which means that F is almost
constant in small regions).
Note: Flux is measuring the rate of outward ow. At places along C where F is pointing
inwards, the contribution to ux will be negative.
22.2.2. Computing ux. Rotating
dr = Tds = dx, dy)
90

clockwise yields
nds = dy, dx).
77
So if F = Mi +Nj, then
ux =
_
C
F nds
=
_
C
M, N) dy, dx)
=
_
C
N dx +M dy,
which can be computed as usual, by choosing a parametrization of C, and so on.
22.3. Divergence.
Denition 22.3. The (2-dimensional) divergence of F = Mi +Nj is
div F := M
x
+N
y
.
Vector or scalar? Scalar.
Another way to remember the denition of div F:
div F = F
=
_

x
,

y
_
M, N)
=
M
x
+
N
y
.
In contrast,
f :=
_
f
x
,
f
y
_
.
22.3.1. Physical interpretation of divergence. If F is the velocity eld of an incompressible
uid, then div F measures how much uid is being added to the system per unit area and
per unit time.
Example 22.4. Suppose that F = xi+yj. At every point, there is more uid going out than
coming in, so every point is a source: div F > 0. (In fact, div F = 1 + 1 = 2 everywhere.)
Example 22.5. Suppose that F = xi yj. This time, div F = 1 + (1) = 0: the uid is
just owing along, with nothing added or removed.
22.4. Greens theorem for ux. Suppose as usual that C is a positively-oriented simple
closed curve enclosing a region R.
The original Greens theorem: If F = Pi +Qj, then
_
C
F dr =

R
curl FdA
_
C
P dx +Qdy =

R
(Q
x
P
y
) dA.
78
Greens theorem for ux (also known as Greens theorem in normal form): If F = Mi + Nj,
then
_
C
F nds =

R
div FdA
_
C
N dx +M dy =

R
(M
x
+N
y
) dA.
Proof of Greens theorem for ux. In the original Greens theorem, take P = N and Q =
M. Done!
Physically, Greens theorem for ux makes a lot of sense: For an incompressible uid, if
you add up all the uid that is being added to the system inside the region R, the same
amount must be owing outward across C (by conservation of mass, it has to go somewhere!)
Problem 22.6. Verify Greens theorem for ux when F = xi + yj and C is the circle of
radius r centered at (0, 0).
Solution: The ux of F across C is
_
C
F nds =
_
C
[F[ ds
=
_
C
r ds
= r
_
C
ds
= r(2r)
= 2r
2
.
On the other hand, div F = 2 at every point of the interior R of the circle, so

R
div FdA =

R
2 dA = 2r
2
.
Greens theorem for ux says that these two quantities should be equal, and they are!
23. April 12
23.1. Greens theorem for ux, continued.
Problem 23.1. What is the ux of F := xi yj across the circle C of radius r centered at
(0, 0)?
79
Solution 1: The ux is
_
C
F nds =
_
C
x, y) dy, dx)
=
_
C
y dx +x dy.
Normally one would have to parametrize C to compute this line integral, but in this case,
there is a shortcut, because y dx + x dy is exact, namely df where f = xy. Another way of
thinking about this: the integral is the same as
_
C
y, x) dx, dy) =
_
C
y, x) dr,
but y, x) is a conservative vector eld (its the gradient of xy), so the integral is 0.
Solution 2: Use Greens theorem for ux: the ux is
_
C
F nds =

R
div FdA = 0
since
div F = M
x
+N
y
= 1 + (1) = 0
at every point of R.
Solution 3: Use symmetry: the value of F n at (x, y) is the opposite of F n at (y, x), so
_
C
F nds = 0.
23.2. Extended Greens theorem. Given a region R, the notation R is sometimes used
to denote its boundary, positively oriented. If R consists of several curves C
1
, C
2
, . . . , C
n
,
then
_
R
F dr =

R
curl FdA
is still true, but the left side is now
_
C
1
F dr + +
_
Cn
F dr.
Just make sure that each C
i
is positively-oriented, so that R lies to the left as one is traversing
the curve.
Why is this true? Cut R into smaller regions, apply the original Greens theorem to each
piece, and add the results.
80
23.3. Simply connected regions. Suppose that D is a connected region in R
2
. Lets say
that a curve C inside D is shrinkable in D if it is possible to continuously shrink C to a
point within D.
Example 23.2. If D is R
2
with the unit disk removed, then a circle of radius 2 centered at
the origin is contained in D but is not shrinkable in D. (To shrink it, the curve would have
to enter the unit disk, at least temporarily, and that is not allowed.)
Denition 23.3. Call D simply connected if every closed curve C in D is shrinkable in D.
Examples of simply connected regions:
a solid rectangle
a disk
a solid triangle
the whole plane R
2
a weird spiral-shaped region drawn on the board
23.4. Test for gradient, again. Let F be a continuous vector eld on a region D.
6 conditions:
(1)
_
C
F dr = 0 for every closed curve C inside D
(2)
_
B
A
F dr is path independent
(3) F = f for some f
(4) P dx +Qdy = df for some f
(5) Q
x
= P
y
at every point of D
(6) curl F = 0 at every point of D.
1,2,3,4 below are equivalent (if satised, F is conservative).
5,6 are equivalent (by denition of curl)
1,2,3,4 imply 5,6.
Fake proof that 6 1: Suppose that curl F = 0 at every point of D. Then for any (simple)
closed curve C in D, Greens theorem gives
_
C
F dr =

interior of C
curl FdA = 0,
since the integrand is 0.
Problem: If the interior of C is not entirely contained in D, curl F might not even be
dened on that interior, so the invocation of Greens theorem is not valid.
This is why 6 1 works only when D is simply connected.
Flowchart for testing whether F is a gradient:
81
Is curl F = 0 everywhere in the region? If no, then F is not a gradient. If yes, check: is
the domain simply connected? If yes, then F is a gradient. If no, further testing is required.
If you think it is a gradient, maybe you can prove it by guessing f such that f = F.
If you think it is not a gradient, maybe you can prove it by nding a closed curve C such
that
_
C
F dr ,= 0.
Example 23.4. Is F =
y
x
2
+y
2
i +
x
x
2
+y
2
j conservative on its domain?
Solution: It turns out that curl F = 0. But the region on which F is dened is the
punctured plane, which is not simply connected. In fact, if C is the unit circle, parametrized
by x = cos t and y = sin t for t [0, 2], then
_
C
F dr =
_
2
0
sin t, cos t) sin t, cos t) dt
=
_
2
0
(sin
2
t + cos
2
t) dt
=
_
2
0
1 dt
= 2.
If F really were a gradient, the integral would have been 0. Thus F is not a gradient.
On the other hand, if F is considered only on the region x > 0, then it turns out that
F = f where f(x, y) = tan
1
(y/x). (This is the of polar coordinates.)
MIDTERM #3 COVERS EVERYTHING UP TO HERE
24. April 18
24.1. Review: parametrizing a line segment.
Problem 24.1. Find a parametrization of the line segment from (2, 3) to (9, 4).
Solution 1: Imagine a particle moving at constant speed along the segment from t = 0 to
t = 1. The initial position vector is 2, 3), and the velocity vector is
9, 4) 2, 3) = 7, 1),
so the parametrization of the line is
r(t) := 2, 3) +t7, 1)
= 2 + 7t, 3 +t).
Answer: 2 + 7t, 3 +t) for t [0, 1].
82
Solution 2: Use the x-coordinate as the parameter: t = x. Next, nd y in terms of x: the
line through (2, 3) to (9, 4) in point-slope form is
y 3 =
4 3
9 2
(x 2),
so
y =
1
7
(x 2) + 3 =
x + 19
7
.
Thus
r(t) := t,
t + 19
7
)
for t [2, 9] is a parametrization. (Warning: If the line segment had been vertical, then we
could not have used x as the parameter, but we could have used y as the parameter instead.)
24.2. Review: parametrizing an arc of a circle.
Problem 24.2. Find a parametrization of the right half of the circle of radius 2 centered at
(5, 3).
Solution 1: The unit circle centered at (0, 0) has parametrization
cos t, sin t).
Multiplying each position vector by 2 gives a parametrization of the circle of radius 2 centered
at (0, 0):
2 cos t, 2 sin t).
Adding 5, 3) to each position vector gives a parametrization of the circle of radius 2 centered
at (5, 3):
r(t) := 5 + 2 cos t, 3 + 2 sin t),
and we want this for t [/2, /2] to get only the right half.
Solution 2: Can we use the x-coordinate as the parameter? Not directly, because for most
values of x there are two points on the curve, so y cannot be expressed as a function of x.
If we cut the semicircle in half, then we could use a dierent parametrization on each half
using the x-coordinate as parameter.
But its easier to use the y-coordinate as the parameter: t = y. To nd x in terms of y,
write the equation of the circle
(distance from (x, y) to (5, 3)) = 2
_
(x 5)
2
+ (y 3)
2
= 2
(x 5)
2
+ (y 3)
2
= 4
83
and solve for x:
(x 5)
2
= 4 (y 3)
2
x 5 = +
_
4 (y 3)
2
(x 5 is positive on the right half of the circle, so we took the plus sign in the square root)
x =
_
4 (y 3)
2
+ 5.
Thus the parametrization is
r(t) :=
_
4 (t 3)
2
+ 5, t)
for t [1, 5] (the minimum and maximum values of y are 3 2).
24.3. Review: antiderivatives of functions of more than one variable.
Problem 24.3. What are all functions f(x, y) on the right half plane x > 0 satisfying
f
x
= x
2
+y
3
+
y
x
?
Solution: By integration with respect to x, treating y as a constant, compute one such
antiderivative:
x
3
3
+y
3
x +y ln x
(we could write ln x instead of ln [x[ since the function is on the right half plane x > 0 only).
Then the general antiderivative is this plus a constant but the constant can depend on y:
x
3
3
+y
3
x +y ln x +g(y)
for any function g(y).
Its easy to check the answer: just take the derivative with respect to x.
24.4. Review: computing a potential function using antiderivatives.
Problem 24.4. Explain why the vector eld
F(x, y) = 3 cos(x + 2y) + 2x +y, 6 cos(x + 2y) +x + 3y
2
)
on R
2
is conservative, and nd a potential function.
Solution: Since R
2
is simply connected, to say that F is conservative is the same as
saying that curl F = 0. As usual, let P and Q be the component functions of F, so P :=
3 cos(x + 2y) + 2x +y and Q := 6 cos(x + 2y) +x + 3y
2
. Then
Q
x
= 6 sin(x + 2y) + 1
P
y
= 6 sin(x + 2y) + 1
curl F = Q
x
P
y
= 0,
84
so F is conservative. (Note: another way to show that F is conservative would be to wait
until after we nd the potential function; at that point, we would know that F is a gradient
vector eld and hence conservative.)
Now lets nd a potential function. One way would be to compute
_
(a,b)
(0,0)
F dr, but instead
lets use the method of antiderivatives. We seek a function f(x, y) such that
f
x
= 3 cos(x + 2y) + 2x +y
f
y
= 6 cos(x + 2y) +x + 3y
2
.
Integrating the rst equation with respect to x gives
f = 3 sin(x + 2y) +x
2
+xy +g(y)
for some function g(y). Taking the derivative with respect to y shows that
f
y
= 6 cos(x + 2y) +x +g

(y),
which must match the earlier condition on f
y
, so g

(y) = 3y
2
. Integrating with respect to y
gives g(y) = y
3
+ c for some constant c. (Note that g is a function of the variable y only,
so c cannot depend on x.) Plugging this back into the formula for f gives the most general
potential function
f = 3 sin(x + 2y) + x
2
+xy +y
3
+c.
In other words, there are innitely many potential functions, one for each number c. In
particular, if we want just one potential function, we can choose a specic value for c; for
instance,
f = 3 sin(x + 2y) +x
2
+xy +y
3
+ 17
is one potential function.
The answer can be checked by computing f.
24.5. Review: computing the centroid of a curve.
Problem 24.5. What is the centroid of the upper half C of the circle x
2
+y
2
= 4? (Assume
constant density.)
Note: the problem is about the half-circle (a 1-dimensional object), not the half-disk (a
2-dimensional object).
Solution: The curve C is symmetric under reection in the y-axis, so the centroid ( x, y)
must lie on the y-axis. In other words, x = 0. On the other hand, y is the average value of
the y-coordinate:
y :=
_
C
y dm
_
C
dm
=
_
C
y ds
_
C
ds
=

_
C
y ds

_
C
ds
=
_
C
y ds
length(C)
.
85
To compute the numerator, we choose a parametrization of C, namely
r(t) := 2 cos t, 2 sin t)
for t [0, ] (we want only the upper half of the circle). Then
dr = 2 sin t, 2 cos t) dt
ds = [dr[ =
_
(2 sin t)
2
+ (2 cos t)
2
dt = 2 dt
_
C
y ds =
_
t=
t=0
2 sin t (2 dt)
= 4
_

0
sin t dt
= 4 cos t[

0
= 4((1) 1)
= 8.
On the other hand, the length of C is
1
2
(22) = 2.
Dividing gives
y =
8
2
=
4

,
so the centroid is (0, 4/). (Plotting this point suggests that this is a reasonable answer,
always a good sign!)
25. April 23
25.1. Triple integrals. Triple integral:

T
f(x, y, z) dV
where f(x, y, z) is continuous on a 3-dimensional region T. (Think of dividing T into tiny
blocks, multiply the value of f at a sample point in each block by the volume of the block,
and add the results to get an approximation.)
Problem 25.1. Let T be the region in R
3
where x
2
+ y
2
+ z
2
1 and z 0. So T is the
upper half of the unit ball in R
3
. Find the centroid of T. (Assume constant density.)
Solution: By symmetry, it must be of the form (0, 0, z), where z is the average value of z
on the half-ball:
z :=

T
z dV
1
2
_
4
3
1
3
_ .
86
At a given height z, the cross-section is the disk D
z
dened by x
2
+ y
2
1 z
2
(with z
viewed as constant), a disk of radius

1 z
2
. So

T
z dV =
_
z=1
z=0
_

Dz
z dx dy
_
dz.
At this point we could write the inner double integral as an iterated integral and obtain the
following:
_
z=1
z=0
_
y=

1z
2
y=

1z
2
_
x=

1y
2
z
2
x=

1y
2
z
2
z dx dy dz.
But it is easier instead to factor the z out of the inner double integral (z is a constant for
the inner integral, in which x and y are the variables), and to recognize what remains as an
area:

T
z dV =
_
z=1
z=0
z Area(D
z
) dz
=
_
z=1
z=0
z
_

1 z
2
2
_
dz
=
_
1
0
(z z
3
) dz
=

4
,
so
z =
/4
1
2
_
4
3
1
3
_
=
3
8
.
25.2. Cylindrical coordinates. In R
3
, replace x, y by polar coordinates, but keep z: this
leads to cylindrical coordinates.
To convert from cylindrical coordinates (r, , z) to rectangular coordinates (x, y, z):
x = r cos
y = r sin
z = z.
87
The other way:
r =
_
x
2
+y
2
tan =
y
x
z = z.
What happens to volume?
dV = dx dy dz = dz r dr d.
This can also be visualized geometrically.
Lets redo the half-ball integral: In cylindrical coordinates, the half-ball is given by r
2
+
z
2
1 and z 0. For each xed point (r, ) in the plane, look at the vertical segment
above/below it contained in T to get the range of integration for z:

T
z dV =
_
=2
=0
_
r=1
r=0
_
z=

1r
2
z=0
z dz r dr d
=
_
2
0
_
1
0
1 r
2
2
r dr d
=
_
2
0
1
8
d
=

4
,
which agrees with what we got before.
25.3. Spherical coordinates. Spherical coordinates are (, , ) (pronounced rho, phi,
theta), where
:= distance to (0, 0, 0)
:= angle down from z-axis (angle between position vector and positive z-axis)
:= same as in cylindrical coordinates, longitude, depends only on x, y.
Warning: Some books use dierent Greek letters here, so when reading another book,
check the denitions of the variables.
It may help to imagine an r z swinging door hinged along the z-axis, with one corner
at (0, 0, 0) and one corner at (x, y, z), so is length of the diagonal of the door, and is
the angle the diagonal forms with the z-axis; by trigonometry in the upper half of the door
(above the diagonal), r = sin is the width, and z = cos is the height (at least if z > 0).
Finally, controls how far counterclockwise the door has swung.
88
Range of possible values:
[0, )
[0, ]
[0, 2] (or [, ] or . . . )
Conversion: Given (, , ), we have r = sin , so
x = r cos = sin cos
y = r sin = sin sin
z = cos .
To convert backwards: Given (x, y, z), compute
:=
_
x
2
+y
2
+z
2
,
then convert (x, y) to polar coordinates (r, ) to get , and nally get from
cos =
z

.
(That is better than using
sin =
r

since the latter cannot distinguish between and .)

26. April 25
26.1. Spherical coordinates, continued.
26.1.1. Relationship to latitude and longitude. A great circle on a sphere S is the intersection
of S with a plane through the center. If P, Q are in S, the shortest path along the sphere
connecting P and Q is an arc of a great circle. (To visualize this, rotate the sphere so that
P and Q are on the equator.)
Set up an xyz-coordinate system with the origin at the center of the earth, with the
positive z-axis passing through the North Pole, with the xy-plane containing the Equator,
and with Greenwich (near London) contained in the xz-plane. Then

(i.e., measured in
degrees) is longitude (east if > 0). And 90

is latitude (number of degrees up from the

Equator).
89
26.1.2. Volume in spherical coordinates. The region with spherical coordinates in the in-
tervals [, + d], [, + d], [, + d] denes a tiny box of volume approximately d
times the surface area of a little rectangle on the sphere of sides d and r d (these
sides are actually tiny arcs of circles, and the formula for the length of such an arc is
(radius)(arc measure)). The volume of the box is approximately
(d)( d)(r d) =
2
sin d dd,
so
dV = dx dy dz =
2
sin d dd.
Another way to nd this formula: use the change of variable formula
dx dy dz =

(x, y, z)
(, , )

d dd.
(The volume scaling factor is the absolute value of the 3 3 Jacobian determinant.)
26.2. Gravitation. Let F be the gravitational force of a point of mass M with position
vector r acting on a point of mass m at (0, 0, 0). Let = [r[. Then Newton says
F =
GmM

2
r

.
If instead of a point of mass M, there is a solid body T whose density is given by = (x, y, z),
so dM = dV , then the z-component of the gravitational force is

T
GmdM

2
r

k =

T
GmdM

2
z

= Gm

T
cos

2
dM
= Gm

T
cos

2
dV
= Gm

T
cos sin d dd.
(We took the z-component so that we would be integrating a scalar function, since we havent
talked about triple integrals of vector-valued functions.)
Example 26.1. Suppose that T is a solid sphere of radius a centered at (0, 0, a), of constant
density = 1, and suppose that m = 1. Then by symmetry, the gravitational force F is
90
pointing straight up, so [F[ equals the z-component of F, which is
G

T
cos sin d dd = G
_
2
0
_
/2
0
_
2a cos
0
cos sin d dd
= G
_
2
0
_
/2
0
2a cos
2
sin dd
= 2Ga
_
2
0

1
3
cos
3

/2
0
d
= 2Ga
_
2
0
1
3
d
=
4Ga
3
.
On the other hand, since = 1, the mass of T is M = Volume(T) =
4
3
a
3
, so if all the
mass of T were concentrated at its center, then Newtons law of gravitation for point masses
would give
[F[ =
GM
a
2
=
4Ga
3
.
Same answer!
Remark 26.2. Newton proved more generally that on any external mass, a solid sphere of
constant density exerts the same force as would a point mass M placed at its center.
27. April 26
27.1. Parametrized surfaces. A region R in R
2
can be mapped to a curved surface S in
R
3
. Example: A disk can be mapped to a Salvador Dal watch.
In general, a parametrized surface S is the image of a region in the uv-plane under a function
r(u, v) = x(u, v), y(u, v), z(u, v)).
It is smooth at a point corresponding to (u, v) R if the vectors
r
u
:= x
u
, y
u
, z
u
)
r
v
:= x
v
, y
v
, z
v
)
are independent (i.e., nonzero and nonparallel). This guarantees that tiny rectangles in
R get mapped to tiny parallelograms in S (instead of being compressed into a curve, for
example). The quotation marks are there because the parallelograms in S could be slightly
warped and hence not actual parallelograms.
Problem 27.1. Let S be the lateral surface of the cone x
2
+y
2
= z
2
for 0 z 5. (Lateral
means on the side. In cylindrical coordinates, the equation would be r = z; thats why
its a cone.) What is a parametrization of S? Is S smooth?
91
Solution: We need to describe a point on the cone using two numbers u, v. Let u be the
r of cylindrical coordinates, measuring the distance of a point is from the z-axis. Let v be
the of cylindrical coordinates. Then for a point (x, y, z) on the cone,
x = r cos = ucos v
y = r sin = usin v
z = r = u.
Also, the range for u is [0, 5] (since we want the part with 0 z 5) and the range for v is
[0, 2]. In other words, the parametrization is
r(u, v) := ucos v, usin v, u),
dened on the rectangle [0, 5] [0, 2] in the uv-plane. The function r maps this rectangle
onto the cone.
Now lets check smoothness. We have
r
u
= cos v, sin v, 1)
r
v
= usin v, u cos v, 0).
At a point where u = 0, we get r
v
= 0. But at a point where u ,= 0, we have that r
v
is
a nonzero vector in the xy-plane while r
u
has a nonzero vertical component so r
u
and r
v
are nonzero and nonparallel. Thus S is smooth except at the points where u = 0, which
corresponds to r = 0. This makes sense: the cone is smooth except at its vertex.
27.2. Surface area. Suppose that S is a surface parametrized by r: R S. To approxi-
mate the surface area of S, we cut R into tiny rectangles, look at their images under r, and
add up the areas of these tiny parallelograms covering S.
Imagine a tiny rectangle [u, u + du] [v, v + dv] in the uv-plane. It will be mapped to a
parallelogram whose sides are the vectors r
u
du and r
v
dv. Dene
dS := (r
u
du) (r
v
dv)
= r
u
r
v
du dv.
What is the geometric meaning of dS?
The length of dS is the area of the parallelogram, which should be thought of as a
piece of surface area, denoted
dS := [r
u
r
v
[ du dv.
The direction of dS is the same as the unit normal vector n to the surface.
92
So
dS = ndS.
Dene the surface area of S to be
Area(S) :=

R
dS =

R
[r
u
r
v
[ du dv.
27.2.1. Surface area of a hemisphere.
Problem 27.2. Let S be the upper half of the unit sphere centered at the origin. What is
the surface area?
(There are many ways to do this; our purpose here is to use the general method for
calculating surface area via a parametrization this will also serve as a review of spherical
coordinates.)
Solution: First we choose a parametrization of S. To describe a point on S, use the
spherical coordinates and . (We dont need , since it equals 1 at every point of S.) The
region in the -plane is the rectangle R where 0 /2 and 0 2. The map that
sends R onto S is
r(, ) := sin cos , sin sin , cos ).
Then
r

= cos cos , cos sin , sin )

r

= sin sin , sin cos , 0)

r

i j k
cos cos cos sin sin
sin sin sin cos 0

= (sin
2
cos )i + (sin
2
sin )j + (sin cos cos
2
+ sin cos sin
2
)k
= (sin
2
cos )i + (sin
2
sin )j + (sin cos )k
= (sin )r
[r

[ = sin
dS = sin dd
Area(S) =
_
2
0
_
/2
0
sin dd
=
_
2
0
d
= 2.
93
Note: The geometric interpretation of the cross product shows that the vector r

is
perpendicular to the sphere at each point; this explains why it turned out to be a scalar
multiple of r.
27.3. 3D vector elds. A 3-dimensional vector eld has the form
F(x, y, z) = P(x, y, z) i +Q(x, y, z) j +R(x, y, z) k.
Denition 27.3. The divergence of F is
div F := F
=

x
,

y
,

z
) P, Q, R)
= P
x
+Q
y
+R
z
.
Vector or scalar? Its a scalar, just as in the 2-dimensional case.
27.4. Surface integrals. Recall: For a curve C,
_
C
f ds =
_
b
a
f(r(t)) [r

(t)[ dt.
(We compute the left side by choosing a parametrization r(t) of C, for t [a, b].
Similarly: If S is a surface, and f = f(x, y, z), then

S
f dS =

R
f(r(u, v)) [r
u
r
v
[ du dv,
where r(u, v) is a parametrization of S, mapping the points (u, v) R to the points of S.
Important reminder: To compute a 1-dimensional integral on a curve C, you must choose
a parametrization r(t) to convert it to an integral on a straight interval [a, b] in the real line.
Generally, to compute a 2-dimensional integral on a curved surface S, you must choose a
parametrization r(u, v) to convert it to an integral over a at region R in the uv-plane. The
only exceptions to this are:
If the integrand is a constant c, then
_
C
c ds = c Length(C)
_
S
c dS = c Area(S),
and you might know Length(C) or Area(S) already by geometry.
Sometimes you can use Greens theorem or some other theorem to convert the integral
to some other kind of integral.
94
27.4.1. Applications of surface integrals.
Denition 27.4. The average value of f on a surface S is

S
f dS
Area(S)
The denitions of mass, centroid, and moment of inertia extend to surfaces as well.
27.5. 3D ux. Recall: In 2D,
ux of F across C :=
_
C
F nds.
In 3D, the ux of a continuous vector eld F across a surface S is a special kind of surface
integral:
ux of F across S :=

S
F ndS =

S
F dS,
where n is a unit normal vector to the surface, and
dS := ndS.
There are actually two choices for n at each point. If S bounds a 3D region, then one
usually chooses n to be the outward unit normal.
27.5.1. Physical meaning of 3D ux. Intuitive explanation: Imagine a 3D uid with constant
velocity eld F. The amount of uid that ows across a tiny parallelogram of area dS in
unit time is the uid in a parallelepiped of base dS and height F n. Summing over all tiny
parallelepipeds comprising a surface S gives

S
F ndS.
Thats ux!
This intuition works more generally for any continuous F and smooth S, since F is almost
constant on small regions, and S is well approximated by tiny parallelograms. Flux measures
the rate of ow (volume per unit time, measured in m
3
/s, say).
27.5.2. Computing 3D ux. To compute a ux integral

S
F dS :
(1) Choose a parametrization of S, say r(u, v) for (u, v) in the at 2D region R. (The
whole point is to convert the surface integral over the curved surface S into a double
integral over the at region R.)
(2) Compute the partial derivatives r
u
, r
v
, and their cross product r
u
r
v
.
(3) Make sure that r
u
r
v
is in the same direction as the desired outward unit normal
n. (If it is in the opposite direction, negate your answer at the end.)
95
(4) Substitute
dS = r
u
r
v
du dv.
to get

S
F dS =

R
F(r(u, v)) (r
u
r
v
) du dv.
(5) Evaluate the double integral over the at 2-dimensional region R; usually this will
be done by converting it to an iterated integral.
Next time: an example of computing 3D ux.
28. April 30
Example 28.1. Let S in R
3
be dened by x
2
+ y
2
= 4 and 0 z 3. (So S is the lateral
surface of a cylinder.) Let F(x, y, z) = yz, x
2
+ 2, 5). What is the outward ux of F across
S?
Solution: Choose the parameterization where u is the of cylindrical coordinates, and
v = z. So r(u, v) = 2 cos u, 2 sin u, v), for 0 u 2 and 0 v 3. Then
r
u
= 2 sin u, 2 cos u, 0)
r
v
= 0, 0, 1)
r
u
r
v
=

i j k
2 sin u 2 cos u 0
0 0 1

= 2 cos u, 2 sin u, 0).

Notice that r
u
r
v
is in the same direction as the outward unit normal n. Next
dS = r
u
r
v
du dv
= 2 cos u, 2 sin u, 0) du dv.
So the ux is

S
F dS =
_
3
0
_
2
0
yz, x
2
+ 2, 5) 2 cos u, 2 sin u, 0) du dv
=
_
3
0
_
2
0
_
yz(2 cos u) + (x
2
+ 2)(2 sin u)
_
du dv
=
_
3
0
_
2
0
_
(2 sin u)v(2 cos u) + ((2 cos u)
2
+ 2)(2 sin u)
_
du dv
=
_
3
0
_
2
0
_
2v sin 2u + 8 cos
2
usin u + 4 sin u
_
du dv
= 0,
96
because
_
2
0
sin 2u du is the integral over two complete cycles, which is 0, and the last two
terms in the integrand are negated by u u + so their integrals over [0, ] cancel with
the integrals over [, 2].
28.1. Divergence theorem. Recall Greens theorem for ux:
Let C be a positively-oriented simple closed (piecewise smooth) curve bounding a region
R in R
2
. Let F be a vector eld that is continuously dierentiable not only on C but also
on R. Then _
C
F nds =

R
div FdA
The divergence theorem is the 3D analogue.
Denition 28.2. A region T in R
3
is called bounded if there is a sphere that contains it.
Denition 28.3. A closed surface S is a (piecewise smooth) surface that is the entire bound-
ary of a bounded region T in R
3
. To say that it is positively-oriented means that at each
(smooth) point of S we choose the outward unit normal n. An integral over a closed surface
S is sometimes written using the notation

S
, although

S
is also correct.
Examples: A sphere, a cube, a torus, a nite cylinder (including the top and bottom
disks).
Theorem 28.4 (Divergence theorem, also known as Gausss theorem or Ostrogradskys theorem
actually discovered earlier by Lagrange). Let S be a positively-oriented closed surface
bounding a region T in R
3
. Let F be a vector eld that is continuously dierentiable not only
on S but also on T. Then

S
F dS =

T
div FdV.
Well prove this later. There is also an extended version in which the boundary of T
consists of several surfaces. For instance, if T is a spherical shell, its boundary S consists of
an outer sphere and an inner sphere, oriented appropriately.
28.1.1. Physical interpretation. The physical interpretation of the divergence theorem is the
same as that of Greens theorem for ux. Suppose that F is the velocity eld of an incom-
pressible 3D uid. Then div F is the source rate (the rate at which uid is being created
per unit volume) so the right side is total rate of uid creation inside T. The left side is
the outward ux across S, the rate at which uid is crossing S. The divergence theorem
is saying that the rate at which uid is created inside T must equal the rate at which it is
overowing and leaking across S.
Why is it true that for F = P, Q, R) the formula div F = P
x
+ Q
y
+ R
z
gives the source
rate?
Imagine a tiny rectangular box with coordinates in [x, x + x], [y, y + y], [z, z + z].
97
Parametrize the top face by r(u, v) = x+u, y +v, z +z), so dS = du dv. The ux across
the top face is
_
top face
F ndS =
_
y
v=0
_
x
u=0
F(x +u, y +v, z + z) kdu dv
=
_
y
v=0
_
x
u=0
R(x +u, y +v, z + z) du dv.
Similarly, the ux across the bottom face is
_
y
v=0
_
x
u=0
F(x +u, y +v, z) (k) du dv =
_
y
v=0
_
x
u=0
R(x +u, y +v, z) du dv.
Adding gives the net ux across the top and bottom faces:
_
y
v=0
_
x
u=0
(R(x +u, y +v, z + z) R(x +u, y +v, z)) du dv
_
y
v=0
_
x
u=0
R
z
(x +u, y +v, z) z du dv
R
z
x y z
by the approximation formula, because R
z
is approximately constant on the tiny box.
A similar computation shows that the other pairs of opposite sides contribute a net ux
of approximately
P
x
x y z and Q
y
x y z
The total source rate for the box is the net ux across all six sides, which is
(P
x
+Q
y
+R
z
) x y z = (div F) x y z.
So the source rate per unit volume is
(div F) x y z
x y z
= div F.
29. May 2
29.1. Proof of the divergence theorem. Last time we introduced the divergence theorem:

S
F dS =

T
div FdV.
Why is this true?
Suppose that F = P i +Qj +Rk. We need to show

S
(P i +Qj +Rk) dS =

T
(P
x
+Q
y
+R
z
) dV.
Its enough to show
(5)

S
Rk dS =

T
R
z
dV
since this together with the similar equalities for P and Q will sum to give the whole theorem.
98
Well prove (5) in the case of a vertically simple region T whose bottom and top surfaces
are graphs, say z = f
1
(x, y) and z = f
2
(x, y) above a region D in the plane.
Right side of (5):

T
R
z
dV =

D
_
z=f
2
(x,y)
z=f
1
(x,y)
R
z
dz dA
=

D
R(x, y, z)[
z=f
2
(x,y)
z=f
1
(x,y)
dA
=

D
(R(x, y, f
2
(x, y)) R(x, y, f
1
(x, y)) dx dy.
Flux through top? Use the parametrization r(x, y) = x, y, f
2
(x, y)) for (x, y) D, so
dS = r
x
r
y
dx dy
=
_
1, 0,
f
2
x
_

_
0, 1,
f
2
y
_
dx dy
=

i j k
1 0
f
2
x
0 1
f
2
y

dx dy
=
_

f
2
x
,
f
2
y
, 1
_
dx dy,
k dS = dx dy.
so

top
Rk dS =

D
R(x, y, f
2
(x, y)) dx dy.
Flux through bottom? Similar, except r
u
r
v
is in the opposite direction of the outward
n, so

bottom
Rk dS =

D
R(x, y, f
1
(x, y)) dx dy.
Flux through sides: Rk is vertical, but dS = ndS is horizontal, so the ux through the
sides is 0.
Total ux: Adding the ux through the top, bottom, and sides gives

S
Rk dS =

D
(R(x, y, f
2
(x, y)) R(x, y, f
1
(x, y)) dx dy.
This matches the right side calculation!
To prove the divergence theorem for a more general region, one could cut it into simple
pieces.
99
29.2. Divergence and gravitation.
Lemma 29.1. Let F be a 3D vector eld pointing radially outward and whose magnitude is
1/
2
(dened everywhere except the origin). Then div F = 0.
Proof. Since the radially outward unit vector is r/[r[ = r/, this means that an explicit
formula for F is
F =
1

2
r

=
3
x, y, z) =
3
x,
3
y,
3
z).
To calculate div F, we need to calculate partial derivatives of the three coordinate functions,
while remembering that is really a function of x, y, z. We could just substitute =
_
x
2
+y
2
+z
2
and calculate the partial derivatives explicitly. Alternatively, taking

x
of

2
= x
2
+y
2
+z
2
gives
2

x
= 2x
so

x
=
x

.
Now

x
(
3
x) =
3
+ (3
4
)

x
x
=
3
3
5
x
2
.
Summing this with the corresponding equations for y and z gives
div F = 3
3
3
5
(x
2
+y
2
+z
2
)
= 3
3
3
5

2
= 0.
Let F = F(x, y, z) be the gravitational eld of a point mass M at (0, 0, 0) (i.e., the force
that it would exert on a unit mass at (x, y, z)). By Newtons inverse square law,
F =
GM

2
r

.
This is just a constant times the vector eld in the lemma, so
div F = 0
at every point in R
3
except (0, 0, 0).
Question 29.2. Let F be the gravitational eld of a point of mass M at the origin. Let S
a
be the sphere of radius a centered at the origin. Let S
2a
be the sphere of twice the radius.
Lets compare the ux across S
a
with the ux across S
2a
. Which of the following is correct?
100
(1) The ux across S
2a
is 4 times as much, because the integral is over a surface area
that is 4 times bigger.
(2) The ux across S
2a
is 1/4 as much, because the gravitational eld is 1/4 as strong.
(3) The uxes are equal and nonzero.
(4) The uxes are both 0, because the divergence theorem says

Sa
F dS =

T
div FdV = 0,
since div F = 0 everywhere.
Answer: The uxes are equal and nonzero. (The two eects in the rst two answers cancel
each other out. The application of the divergence theorem in the last answer is wrong: F
is not dened at (0, 0, 0), so the right side of the divergence theorem does not even make
sense.)
Question 29.3. What is the ux across the sphere S
a
of radius a centered at (0, 0, 0)?
Answer: At every point, F and n are in opposite directions, so
ux =

Sa
F ndS
=

Sa
[F[ dS
=

Sa

GM
a
2
dS
=
GM
a
2
Area(S
a
)
=
GM
a
2
(4a
2
)
= 4GM.
It is independent of the radius a!
Even better, if S is any closed surface enclosing the point mass, and T is the 3D region
between S and a small sphere S
a
centered at the point mass (so T has a bubble inside), then
the extended divergence theorem shows that

S
F dS

Sa
F dS =

T
div FdV
= 0,
101
so

S
F dS =

Sa
F dS
= 4GM.
The same claim is true if there are many point masses inside S, or even some planets
inside S, because the force elds add. This proves Gausss law:
gravitational ux across S = 4GM,
where M is the mass enclosed by S.
30. May 3
30.1. Gravitation, continued.
Example 30.1. What is the gravitational eld inside a hollow spherical planet? Inside a
centered sphere S of radius r inside the hollow part, symmetry implies that F = cn for some
c depending only on r. Then the ux across S is 4r
2
c, but Gausss law says that it equals
0, so c = 0. Thus the gravitational eld is 0 everywhere inside the hollow part.
Question 30.2. For a donut-shaped planet, if you are standing on the inner circle of the
planet, is gravity pulling you towards the center of mass or is it pulling you towards the
planet under your feet?
Hint: Imagine lling in the donut hole with a cylinder of very small height. The total ux
through the cylinder is 0 by Gausss law. On the other hand, is the outward ux through the
top and bottom disks positive or negative? If you gure that out, that can help answer the
question, because it must be cancelled by the ux through the lateral surface of the cylinder.
30.2. Application of the divergence theorem to an electric eld. Let E = E(x, y, z)
be the electric eld of a point charge Q at (0, 0, 0). Coulombs law:
E =
Q/4
0

2
r

where
0
is a constant.
The physics is dierent, but the math is the same as for gravitation, with the constant
Q/4
0
in place of GM. The gravitational ux was 4GM, which is 4 times the constant
appearing in the inverse square law for gravitation. Similarly, the electric ux is
4(Q/4
0
) =
Q

0
.
Summary:
102
eld ux
gravitational F =
GM

2
r

4GM
electric E =
Q/4
0

2
r

Q/
0
So we get Gausss law for an electric eld (also called the Gauss-Coulomb law):
The electric ux across a closed surface S equals the charge enclosed divided
by
0
,

S
E dS =
Q

0
.
30.3. 3D Curl. Recall: if F = P, Q) is a 2D vector eld, then
curl F =

y
P Q

= Q
x
P
y
.
Denition 30.3. If F = P, Q, R) is a 3D vector eld, then
curl F := F
=

i j k

z
P Q R

= (R
y
Q
z
)i + (P
z
R
x
)j + (Q
x
P
y
)k.
Warning:
If F is a 2D vector eld, then curl F is a scalar-valued function of (x, y).
If F is a 3D vector eld, then curl F is a vector-valued function of (x, y, z).
30.3.1. Physical interpretation. If F is a velocity eld, curl F measures the rotational com-
ponent of F:
direction of curl F is along the axis of rotation (have the ngers of your right hand
curl in the direction of rotation; then your thumb gives the direction)
length of curl F is 2(angular velocity).
Example 30.4. If F = y, x, 0) (rotation around z-axis with angular velocity ), then
curl F = 2k.
103
30.4. Line integrals in 3D. Let C be a piecewise smooth curve in R
3
.
Three kinds of line integrals:
Line integral with respect to arc length:
_
C
f(x, y, z) ds. (Remember: s = distance
traveled = arc length.)
Line integral with respect to coordinate variables:
_
C
P(x, y, z) dx + Q(x, y, z) dy +
R(x, y, z) dz
Line integral of a vector eld:
_
C
F dr.
Each of these is approximated by a sum of products; e.g.,
_
C
f(x, y, z) ds
n

i=1
f(P
i
) s
i
To compute it: convert to an ordinary integral over a straight line segment by substituting
a parametrization of C, say r(t) for t [a, b].
30.5. Fundamental theorem of calculus for line integrals. FTC for line integrals: If C
starts at A and ends at B, and f(x, y, z) is a function that is continuously dierentiable at
each point of C, then
_
C
f dr = f(B) f(A).
Warning: This can be used to calculate
_
C
F dr only when F is a gradient!
Denition 30.5. A dierential P dx +Qdy +Rdz is exact if it equals df for some f.
Since df := f
x
dx+f
y
dy +f
z
dz, exact means that there is an f such that P = f
x
, Q = f
y
,
R = f
z
.
30.6. Test for gradient. Let F = P, Q, R) be a continuously dierentiable vector eld in
a 3D region T.
Six conditions:
(1)
_
C
F dr = 0 for every closed curve C inside T
(2)
_
B
A
F dr is path independent (for paths inside T)
(3) F = f (i.e., P = f
x
, Q = f
y
, and R = f
z
) for some f = f(x, y, z) on T
(4) P dx +Qdy +Rdz is exact (equal to df for some f)
(5) R
y
= Q
z
, P
z
= R
x
, and Q
x
= P
y
at every point of T
(6) curl F = 0 at every point of T.
Conditions (1)-(4) are equivalent, and if satised, F is called conservative.
Conditions (5)-(6) are equivalent.
Conditions (1)-(4) imply (5)-(6). E.g., (3) says P = f
x
, Q = f
y
, R = f
z
, which imply
R
y
= (f
z
)
y
= f
zy
= f
yz
= (f
y
)
z
= Q
z
, and so on.
If T is simply connected, then all six conditions are equivalent.
104
30.7. Computing a potential function. Consider the vector eld
F = (3x
2
y + 5yz) i + (x
3
+ 7z + 5xz) j + (7y + 5xy +e
z
) k
on R
3
.
Problem 1: Is there a function f such that f = F?
Solution: It turns out that R
3
is simply connected, so we can use any of the six conditions.
Lets test (6):
curl F := F
=

i j k

z
3x
2
y + 5yz x
3
+ 7z + 5xz 7y + 5xy +e
z

= i((7 + 5x) (7 + 5x)) j(5y 5y) +k((3x

2
+ 5z) (3x
2
+ 5z))
= 0.
So the answer is yes.
Problem 2: Can you nd such an f?
First solution (antiderivative method): We know
f
x
= 3x
2
y + 5yz
f
y
= x
3
+ 7z + 5xz
f
z
= 7y + 5xy +e
z
.
The f
x
equation implies
f = x
3
y + 5xyz +g(y, z)
for some g(y, z). Taking

y
gives
f
y
= x
3
+ 5xz +g
y
and comparing with the given f
y
equation shows that
g
y
= 7z
g = 7yz +h(z) for some h(z).
f = x
3
y + 5xyz + 7yz +h(z).
Taking

z
gives
f
z
= 5xy + 7y +h
z
105
and comparing with the given f
z
equation shows that
h
z
= e
z
+c for some constant c.
f = x
3
y + 5xyz + 7yz +e
z
+c.
We check that this really has the right gradient.
Lecture actually ended here, due to lack of time.
Second solution (FTC in reverse): If f = F, then
f(a, b, c) f(0, 0, 0) =
_
C
f dr
=
_
C
F dr
=
_
(a,b,c)
(0,0,0)
(3x
2
y + 5yz) dx + (x
3
+ 7z + 5xz) dy + (7y + 5xy +e
z
) dz.
Choose the path C
1
+C
2
+C
3
where C
1
goes from (0, 0, 0) to (a, 0, 0), C
2
goes from (a, 0, 0)
to (a, b, 0), and C
3
goes from (a, b, 0) to (a, b, c). Add up the three integrals. For C
1
, use the
parametrization r(t) := t, 0, 0) for t [0, a]; then x = t, y = 0, z = 0, dx = dt, dy = 0,
dz = 0, so
_
C
1
F dr =
_
a
0
0 dt = 0.
Similarly, for C
2
use a, t, 0) for t [0, b]; then x = a, y = t, z = 0, dx = 0, dy = dt, dz = 0,
so
_
C
2
F dr =
_
b
0
a
3
dt = a
3
b.
Finally, for C
3
, use the parametrization a, b, t) for t [0, c], so x = a, y = b, z = t, dx = 0,
dy = 0, dz = dt, so
_
C
3
F dr =
_
c
0
(7b + 5ab +e
t
) dt
= 7bc + 5abc +e
c
1.
Summing yields
_
C
F dr = a
3
b + 7bc + 5abc +e
c
1.
Thus
f(x, y, z) = x
3
y + 7yz + 5xyz +e
z
1
is one possibility. (The others are obtained by adding any constant.)
106
Third solution: Just guess one possible potential function f, and check that it has the
right gradient; then the complete set of solutions is the set of functions of the form f + c,
where c is any number.
MIDTERM #4 COVERS EVERYTHING UP TO HERE
31. May 7
31.1. Stokes theorem.
_
C
F dr ux
2D Greens theorem Greens theorem for ux
3D Stokes theorem Divergence theorem
Remember Greens theorem?
_
C
F dr =

R
curl FdA.
Stokes theorem is the 3D analogue.
Setup: Let C be a closed curve, and let S be any surface bounded by C. Roughly speaking,
S lls in C. (Here S is usually not a closed surface; it does not bound a 3-dimensional
region.) For example, C could be the Equator on the Earth, in which case S could be
the Northern Hemisphere. Let F be a 3D vector eld that is continuously dierentiable
everywhere on S.
Stokes theorem
1
:
_
C
F dr =

S
curl F dS.
Here curl F means F, and dS means ndS.
One subtlety: both sides depend on a choice of orientation, and these must be chosen
compatibly in order for the theorem to hold. Namely, the left side depends on the choice of
direction along C, and the right side depends on the choice of unit normal vector n that is
implicit in dS = ndS.
What does it mean for these to be compatible? Two ways to say it:
If you walk along C in the chosen direction, with S to your left, then n is pointing
up.
If your right hand thumb is pointing in the direction of n, then your ngers point in
the chosen direction along C.
1
Actually rst proved by Lord Kelvin, who mailed it to Stokes, who included it as a question in a physics
competition for students.
107
31.1.1. Stokes theorem for a at surface. Special case: Suppose that S is a piece of the
xy-plane bounded by a closed curve C oriented counterclockwise, and suppose that
F(x, y, z) = P(x, y) i +Q(x, y) j + 0 k.
Which way should n be pointing? Up.
In this special case, what are the two sides of Stokes theorem?
Left side:
_
C
F dr =
_
C
P, Q, 0) dx, dy, dz)
=
_
C
P dx +Qdy.
Right side:
F =

i j k

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

= (Q
x
P
y
)k
dS = ndS
= kdS,

S
F dS =

S
(Q
x
P
y
) dS.
So Stokes theorem for a at surface says
_
C
P dx +Qdy =

S
(Q
x
P
y
) dS.
This is just Greens theorem!
31.1.2. Intuition for why Stokes theorem is true.
If C and S are in the xy-plane, then Stokes follows from Green.
If C and S are in an arbitrary plane, the same holds, because of the geometric
invariance of work and curl under rotations of space. (They have a meaning that is
independent of the coordinate system you are working in.)
If S consists of many at polygons hinged together (like a disco ball), then adding
up Stokes theorem for each polygon gives Stokes theorem for S (the line integrals
on shared edges cancel).
Any piecewise smooth surface S can be approximated by a collection of tiny polygons.
108
31.1.3. Example.
Problem 31.1. Let S be the boundary of the cylinder x
2
+y
2
9, 0 z 2 excluding the
base in the xy-plane. Let F = y, x, y
3
). Compute the outward ux of curl F across S in
as many ways as you can.
Solution 1: Use Stokes theorem to convert it to a line integral on the circle C given by
x
2
+y
2
= 9 in the xy-plane.
ux of curl F across S =

S
curl F dS
Stokes
=
_
C
F dr.
To compute the latter, choose a parametrization of C (counterclockwise, so as to agree with
the orientation of S). Lets use
r(t) = 3 cos t, 3 sin t, 0) for 0 t 2
so
dr = 3 sin t, 3 cos t, 0) dt
_
C
F dr =
_
2
0
3 sin t, 3 cos t, (3 sin t)
3
) 3 sin t, 3 cos t, 0) dt
=
_
2
0
_
9 sin
2
t 9 cos
2
t
_
dt
=
_
2
0
9 dt
= 18.
Solution 2: Compute the ux directly from the denition. The surface S consists of the
top disk S
1
and the lateral surface S
2
of the cylinder. We have
curl F =

i j k

z
y x y
3

= 3y
2
i 2k.
On S
1
, the unit normal is n = k, so
curl F n = 2

S
1
curl F ndS = 2 Area(S
1
)
= 18.
109
For S
2
, we use the parametrization in which u = z is the height, and v is the of cylindrical
coordinates:
r(u, v) := 3 cos v, 3 sin v, u)
for u [0, 2] and v [0, 2]. Then
r
u
= 0, 0, 1)
r
v
= 3 sin v, 3 cos v, 0)
r
u
r
v
=

i j k
0 0 1
3 sin v 3 cos v 0

= (3 cos v) i + (3 sin v) j.
This gives the reverse of the desired orientation for n, so

S
1
curl F ndS =
_
2
0
_
2
0
curl F (r
u
r
v
) du dv
=
_
2
0
_
2
0
3(3 sin v)
2
, 0, 2) 3 cos v, 3 sin v, 0) du dv
=
_
2
0
_
2
0
81 sin
2
v cos v dudv
=
_
2
0
162 sin
2
v cos v dv
= 54 sin
3
v

2
0
= 0.
So the total ux of curl F across S is
18 + 0 = 18.
Lecture actually ended here.
Solution 3: Use the divergence theorem to compute the ux of curl F across the entire
boundary of the cylinder, and then subtract the ux across the bottom disk. We already
calculated that
curl F = 3y
2
i 2 k
so
div curl F =

x
3y
2
+

y
0 +

z
(2) = 0 + 0 + 0 = 0.
110
By the divergence theorem, the outward ux of curl F across the entire boundary of the
cylinder is

entire boundary
(curl F) dS =

solid cylinder
div curl FdV
=

solid cylinder
0 dV
= 0.
On the other hand, the outward ux across the bottom disk of the cylinder is

bottom disk
(curl F) dS =

bottom disk
(curl F) (k) dS (since the outward unit normal is downward)
=

bottom disk
2 dS
= 2 Area(bottom disk)
= 2( 3
2
)
= 18.
Subtracting gives the ux of curl F through S:

S
(curl F) dS = 0 18 = 18.
32. May 9
32.1. Review: extended divergence theorem.
Problem 32.1. Let S
1
be the unit sphere centered at the origin. Let S
2
be the sphere of
radius 2 centered at the origin. Let F be an unknown continuously dierentiable 3D vector
eld such that F(x, y, z) = x, y, 3) for all (x, y, z) S
1
, and F(x, y, z) = x, y, z) for all
(x, y, z) S
2
. What is the average value of div F on the region T between the two spheres?
Solution: The average of any function f on T is dened as

T
f dV
Volume(T)
,
so the average value of div F is

T
div FdV
Volume(T)
.
The denominator is just the dierence of volumes of two balls:
Volume(T) =
4
3
2
3

4
3
1
3
=
28
3
.
111
What about the numerator? We dont have a formula for F on points between the two
spheres, so we cannot compute

T
div FdV directly. But the extended divergence theorem
says that
outward ux of F across the boundary of T
. .
T
=

T
div FdV,
so we can try computing the left side instead. The boundary T consists of two pieces, S
2
with the outward unit normal, and S
1
with the unit normal pointing towards the origin (lets
call that the inward unit normal, even though it is outward from the point of view of T). So
the extended divergence theorem in reverse gives

T
div FdV = outward ux of F across T
= (outward ux of F across S
2
) + (inward ux of F across S
1
)
= (outward ux of F across S
2
) (outward ux of F across S
1
).
Lets now compute each of the latter two uxes separately. At each point of S
2
, the value
F(x, y, z) = x, y, z) is a radially outward vector of length 2, and n is a unit vector in the
same direction, so we can use a shortcut:
F n = [F[[n[ cos 0 = (2)(1)(1) = 2,
so
outward ux of F across S
2
=

S
2
F ndS
=

S
2
2 dS
= 2

S
2
dS
= 2 Area(S
2
)
= 2(42
2
)
= 32.
For S
1
, it is not true that F n is constant, so the shortcut doesnt work. This means we
have to parametrize S
1
. We can use two of the spherical coordinates, namely and (since
= 1 on S
1
). Then the parametrization is
r(, ) := sin cos , sin sin , cos ),
112
for [0, ] and [0, 2]. To compute ux, we compute
r

= cos cos , cos sin , sin )

r

= sin sin , sin cos , 0)

r

i j k
cos cos cos sin sin
sin sin sin cos 0

= (sin
2
cos )i + (sin
2
sin )j + (sin cos cos
2
+ sin cos sin
2
)k
= sin
2
cos , sin
2
sin , sin cos )
(this is outward, so no need to change the sign for computing outward ux)
dS = r

dd
= sin
2
cos , sin
2
sin , sin cos ) dd,
so
outward ux of F across S
1
=

S
1
F dS
=

S
1
x, y, 3) sin
2
cos , sin
2
sin , sin cos ) dd
=

S
1
sin cos , sin sin , 3) sin
2
cos , sin
2
sin , sin cos ) dd
=

S
1
(sin
3
(cos
2
+ sin
2
) + 3 sin cos ) dd
=
_
2
0
_

0
(sin
3
+ 3 sin cos ) dd.
Now
_

0
3 sin cos d = 0
113
by symmetry, and
_

0
sin
3
d =
_

0
(1 cos
2
) sin d
=
_
1
1
(1 u
2
)(du) (where u = cos )
=
_
1
1
(1 u
2
)du
= u
u
3
3

1
1
= (2/3) (2/3)
= 4/3,
so we get
outward ux of F across S
1
=
_
2
0
4
3
d =
8
3
.
Thus

T
div FdV = (outward ux of F across S
2
) (outward ux of F across S
1
)
= 32
8
3
=
88
3
,
and nally, the average value of div F on T is

T
div FdV
Volume(T)
=
88/3
28/3
= 22/7.
Remark 32.2. Actually, there is a sneaky trick that could have been used to simplify the
computation of the ux through S
1
. Namely, if G is the vector eld dened by x, y, 3) on
all of R
3
, then G has the same values as F on S
1
, so
outward ux of F across S
1
= outward ux of G across S
1
=

ball of radius 1
div GdV (by the divergence theorem)
=

ball of radius 1
(1 + 1 + 0) dV
= 2
_
4
3
1
3
_
=
8
3
.
114
32.2. Review: a parade of dierentials. What is the dierence between ds, dS, dS, etc.?
dr can be thought of as a tiny vector measuring change of position along a curve.
Use: Line integrals of vector elds (e.g., work) have the shape
_
C
F dr.
dr = dx, dy, dz) = Tds.
ds can be thought of as the length of a little piece of a curve. Uses: Length of a
curve, line integrals of scalar functions.
ds = [dr[ =
_
dx
2
+dy
2
+dz
2
=
_
x

(t)
2
+y

(t)
2
+z

(t)
2
dt.
dA represents a piece of area of a region in R
2
:
dA = dx dy (rectangular)
= r dr d (polar)
=

(x, y)
(u, v)

du dv (change of variable).
(The last formula is for a change of variable x = x(u, v) and y = y(u, v).)
nds in 2D is dy, dx), obtained by rotating dr = Tds = dx, dy) clockwise 90

.
Use: Flux of a 2D vector eld F across a curve C in R
2
is
_
C
F nds.
dS can be thought of as the area of a little piece of a surface. Uses: Surface area,
surface integrals of scalar functions. If the surface is parametrized by r(u, v), then
dS = [r
u
r
v
[ du dv
(Thats the area of a tiny parallelogram of sides r
u
du and r
v
dv.) For a sphere of
radius parametrized by , ,
dS =
2
sin dd.
dS means ndS. Use: Flux of a 3D vector eld F across a surface S is

S
F dS. In
terms of r(u, v):
dS = r
u
r
v
du dv.
The length of dS is dS. The direction of dS is the unit normal
n =
r
u
r
v
[r
u
r
v
[
.
dV represents a piece of volume of a region in R
3
:
dV = dx dy dz (rectangular)
= dz r dr d (cylindrical)
=
2
sin d dd (spherical)
=

(x, y, z)
(u, v, w)

du dv dw (change of variable).
115
dm can be thought of as a tiny bit of mass. It is ds, or dS, or dV , depending on
whether the object is of dimension 1, 2, or 3. (E.g., a wire, a warped metal plate, or
a planet.) The density is mass per unit length, or mass per unit area, or mass per
unit volume.
33. May 14
116