Curves in R2: Graphs vs Level Sets

Graphs (y = f (x)): The graph of f : R → R is

{(x, y) ∈ R2 | y = f (x)}.
Example: When we say “the curve y = x2 ,” we really mean: “The graph of the function
f (x) = x2 .” That is, we mean the set {(x, y) ∈ R2 | y = x2 }.

Level Sets (F (x, y) = c): The level set of F : R2 → R at height c is

{(x, y) ∈ R2 | F (x, y) = c}.

Example: When we say “the curve x2 + y 2 = 1,” we really mean: “The level set of the
function F (x, y) = x2 + y 2 at height 1.” That is, we mean the set {(x, y) ∈ R2 | x2 + y 2 = 1}.

Note: Every graph is a level set (why?). But not every level set is a graph.
Graphs must pass the vertical line test. (Level sets may or may not.)

Surfaces in R3: Graphs vs Level Sets

Graphs (z = f (x, y)): The graph of f : R2 → R is

{(x, y, z) ∈ R3 | z = f (x, y)}.

Example: When we say “the surface z = x2 + y 2 ,” we really mean: “The graph of the func-
tion f (x, y) = x2 + y 2 .” That is, we mean the set {(x, y, z) ∈ R3 | z = x2 + y 2 }.

Level Sets (F (x, y, z) = c): The level set of F : R3 → R at height c is

{(x, y, z) ∈ R3 | F (x, y, z) = c}.

Example: When we say “the surface x2 + y 2 + z 2 = 1,” we really mean: “The level set of the
function F (x, y, z) = x2 + y 2 + z 2 at height 1.” That is, {(x, y, z) ∈ R3 | x2 + y 2 + z 2 = 1}.

Again: Every graph is a level set (why?). But not every level set is a graph.
Graphs must pass the vertical line test. (Level sets may or may not.)

Q: Do you see the patterns here? For example, suppose G : R7 → R.

◦ What does the graph of G look like? (A: 7-dimensional object in R8 .
That is, {(x1 , . . . , x7 , x8 ) ∈ R8 | x8 = G(x1 , . . . , x7 )}.)
◦ What do the level sets of G look like? (A: They are (generically) 6-dim
objects in R7 . That is, {(x1 , . . . , x7 ) ∈ R7 | G(x1 , . . . , x7 ) = c}.)
Curves in R2: Examples of Graphs
2

{(x, y) ∈ R  y = ax + b}: Line (not vertical)
{(x, y) ∈ R2  y = ax2 + bx + c}: Parabola

Curves in R2: Examples of Level Sets


2
{(x, y) ∈ R ax + by = c}:Line
  x2 y 2
(x, y) ∈ R2  2 + 2 = 1 : Ellipse

a b
  x2 y 2 
(x, y) ∈ R2  2 − 2 = 1 : Hyperbola

a b
  x2 y 2 
(x, y) ∈ R2  − 2 + 2 = 1 : Hyperbola

a b

Surfaces in R3: Examples of Graphs

n  o
3
(x, y, z) ∈ R  z = ax + by : Plane (not containing a vertical line)
x2 y 2
  
3
(x, y, z) ∈ R  z = 2 + 2 : Elliptic Paraboloid
a b 
2
y2

3
 x
(x, y, z) ∈ R  z = − 2 − 2 : Elliptic Paraboloid
a b
2 2
  x y
(x, y, z) ∈ R3  z = 2 − 2 : Hyperbolic Paraboloid (“saddle”)

a b 
2
y2

3
 x
(x, y, z) ∈ R  z = − 2 + 2 : Hyperbolic Paraboloid (“saddle”)
a b

Surfaces in R3: Examples of Level Sets

n  o
3
(x, y, z) ∈ R  ax + by + cz = d : Plane
  x2 y 2 z 2 
(x, y, z) ∈ R3  2 + 2 + 2 = 1 : Ellipsoid

a b c
  x2 y 2 z 2 
(x, y, z) ∈ R3  2 + 2 − 2 = 1 : Hyperboloid of 1 Sheet

a b c
  x2 y 2 z 2 
(x, y, z) ∈ R3  2 − 2 − 2 = 1 : Hyperboloid of 2 Sheets

a b c
Two Model Examples
Example 1A (Elliptic Paraboloid): Consider f : R2 → R given by

f (x, y) = x2 + y 2 .

The level sets of f are curves in R2 . Level sets are {(x, y) ∈ R2  x2 + y 2 = c}.
The graph of f is a surface in R3 . Graph is {(x, y, z) ∈ R3  z = x2 + y 2 }.

Notice that (0, 0, 0) is a local minimum of f .

∂2f ∂2f
Note that ∂f ∂f
∂x (0, 0) = ∂y (0, 0) = 0. Also, ∂x2 (0, 0) > 0 and ∂y 2 (0, 0) > 0.
Sketch the level sets of f and the graph of f :

Example 1B (Elliptic Paraboloid): Consider f : R2 → R given by

f (x, y) = −x2 − y 2 .

The level sets of f are curves in R . Level sets are {(x, y) ∈ R −x2 −y 2 = c}.
2 2

The graph of f is a surface in R3 . Graph is {(x, y, z) ∈ R3  z = −x2 − y 2 }.

Notice that (0, 0, 0) is a local maximum of f .

∂2f ∂2f
Note that ∂f
∂x (0, 0) = ∂f
∂y (0, 0) = 0. Also, ∂x2 (0, 0) < 0 and ∂y 2 (0, 0) < 0.
Sketch the level sets of f and the graph of f :

Example 2 (Hyperbolic Paraboloid): Consider f : R2 → R given by

f (x, y) = x2 − y 2 .

The level sets of f are curves in R2 . Level sets are {(x, y) ∈ R2  x2 − y 2 = c}.
The graph of f is a surface in R3 . Graph is {(x, y, z) ∈ R3  z = x2 − y 2 }.

Notice that (0, 0, 0) is a saddle point of the graph of f .

∂2f ∂2f
Note that ∂f
∂x (0, 0) = ∂f
∂y (0, 0) = 0. Also, ∂x2 (0, 0) > 0 while ∂y 2 (0, 0) < 0.
Sketch the level sets of f and the graph of f :
Introduction: Derivatives of Functions Rn → Rm
Def: Let F : Rn → Rm be a function, say

F (x1 , . . . , xn ) = (F1 (x1 , . . . , xn ), . . . , Fm (x1 , . . . , xn )).

Its derivative at the point (x1 , . . . , xn ) is the linear transformation

DF (x1 , . . . , xn ) : Rn → Rm whose (m × n) matrix is
 ∂F ∂F1

∂x
1
· · · ∂xn
 .. 1 .. .
DF (x1 , . . . , xn ) =  . . 
∂Fm ∂Fm
∂x1 · · · ∂xn

Note: The columns are the partial derivatives with respect to x1 , then to
x2 , etc. The rows are the gradients of the component functions F 1 , F 2 , etc.
∇F1
   
∂F
DF (x1 , . . . , xn ) =  ∂x ∂F 
· · · ∂x = .. .
1 n
.
∇Fm

Example: Let F : R2 → R3 be the function

F (x, y) = (x + 2y, sin(x), ey ) = (F1 (x, y), F2 (x, y), F3 (x, y)).

Its derivative at (x, y) is a linear transformation DF (x, y) : R2 → R3 . The

matrix of the linear transformation DF (x, y) is:
 ∂F ∂F  
1 1

∂x ∂y 1 2
DF (x, y) =  ∂F
 2 ∂F2  
∂y  = cos(x) 0 .

∂x
∂F3 ∂F3
∂x ∂y
0 ey

Notice that (for instance) DF (1, 1) is a linear transformation, as is DF (2, 3),

etc. That is, each DF (x, y) is a linear transformation R2 → R3 .

Goals: We will:
◦ Interpret the derivative of F as the “best linear approximation to F .”
◦ State a Chain Rule for multivariable derivatives.
Introduction: Gradient of Functions Rn → R
Def: Let f : Rn → R be a function.
Recall: The derivative of f : Rn → R at the point x = (x1 , . . . , xn ) is the
1 × n matrix h i
∂f ∂f
Df (x) = ∂x1 · · · ∂xn .
The gradient of f : Rn → R at the point x = (x1 , . . . , xn ) is the vector
h i
∂f ∂f
∇f (x) = ∂x1 , . . . , ∂xn .

The directional derivative of f : Rn → R at the point x = (x1 , . . . , xn )

in the direction v ∈ Rn is the dot product of the vectors ∇f (x) and v:

Dv f (x) = ∇f (x) · v.

We will give geometric interpretations of these concepts later in the course.

Introduction: Hessian of Functions Rn → R

Theorem: Let f : Rn → R be a function whose second partial derivatives
are all continuous. Then:
∂ 2f ∂ 2f
= .
∂xi ∂xj ∂xj ∂xi
In brief: “Second partials commute.”

Def: Let f : Rn → R be a function.

The Hessian of f : Rn → R at the point x = (x1 , . . . , xn ) is the n × n
matrix  ∂2f ∂2f

∂x2 · · · ∂x1 ∂xn
 .. 1 .. .
Hf (x) =  . . 
∂2f ∂2f
∂xn ∂x1 ··· ∂x2n
The directional second derivative of f : Rn → R at the point x =
(x1 , . . . , xn ) in the direction v ∈ Rn is

vT Hf (x) v.

Again, we will give geometric interpretations of these concepts later on.

Linear Approximation: Single-Variable Calculus
Review: In single-variable calc, we look at functions f : R → R. We write
y = f (x), and at a point (a, f (a)) write:

∆y ≈ dy.

Here, ∆y = f (x) − f (a), while dy = f 0 (a)∆x = f 0 (a)(x − a). So:

f (x) − f (a) ≈ f 0 (a)(x − a).

Therefore:
f (x) ≈ f (a) + f 0 (a)(x − a).

The right-hand side f (a) + f 0 (a)(x − a) can be interpreted as follows:

◦ It is the best linear approximation to f (x) at x = a.
◦ It is the 1st Taylor polynomial to f (x) at x = a.
◦ The line y = f (a) + f 0 (a)(x − a) is the tangent line at (a, f (a)).

Linear Approximation: Multivariable Calculus

Now consider functions f : Rn → Rm . At a point (a, f (a)), we have exactly
the same thing:
f (x) − f (a) ≈ Df (a)(x − a).
That is:
f (x) ≈ f (a) + Df (a)(x − a). (∗)
Note: The object Df (a) is a matrix, while (x − a) is a vector. That is,
Df (a)(x − a) is matrix-vector multiplication.

Example: Let f : R2 → R. Let’s write x = (x1 , x2 ) and a = (a1 , a2 ). Then

(∗) reads:
h i x − a 
∂f ∂f 1 1
f (x1 , x2 ) ≈ f (a1 , a2 ) + ∂x1 (a1 , a2 ) ∂x2 (a1 , a2 )
x2 − a2
∂f ∂f
= f (a1 , a2 ) + (a1 , a2 )(x1 − a1 ) + (a1 , a2 )(x2 − a2 ).
∂x1 ∂x2
Review: Taylor Polynomials in Single-Variable Calculus
Review: In single-variable calculus, we look at functions f : R → R.
At a point a ∈ R, the linear approximation (1st-deg Taylor polynomial)
to f is:
f (x) ≈ f (a) + f 0 (a)(x − a).
More accurate is the quadratic approximation (2nd-deg Taylor polynomial)
1
f (x) ≈ f (a) + f 0 (a)(x − a) + f 00 (a)(x − a)2 .
2!
We would like to have similar ideas for multivariable functions.

Linear Approximation: 1st-Deg Taylor Polynomials

Let f : Rn → Rm . The linear approximation of f at the point a ∈ Rn
is:
f (x) ≈ f (a) + Df (a)(x − a). (1)
Note that Df (a) is a matrix, while (x − a) is a vector. That is, Df (a)(x − a)
is matrix-vector multiplication.
Note that (1) is the best linear approximation to f (x) for points x near a.
It is the 1st-degree Taylor polynomial to f (x) at a.

Quadratic Approximation: 2nd-Deg Taylor Polynomials

Let f : Rn → R. The quadratic approximation of f at the point a ∈ Rn
is:
1
f (x) ≈ f (a) + Df (a)(x − a) + (x − a)T Hf (a)(x − a). (2)
2!
Note that (x − a)T is a row vector, while (x − a) is a column vector, while
Hf (a) is a matrix. So, 2!1 (x − a)T Hf (a)(x − a) is of the form vT Av.
Note that (2) is the best quadratic approximation to f (x) for points x
near a. It is the 2nd-degree Taylor polynomial to f (x) at a.

Example: Let f : R2 → R be f (x, y) = x3 sin(y). For (x, y) near a = (2, π2 ):

   
π π x−2 1 π
 π x−2
f (x, y) ≈ f (2, 2 ) + Df (2, 2 ) + x − 2 y − 2 Hf (2, 2 )
y − π2 2! y − π2
    
  x−2 1 π
 12 0 x−2
= 8 + 12 0 + x−2 y− 2
y − π2 2 0 −8 y − π2
= 8 + 12(x − 2) + 6(x − 2)2 − 4(y − π2 )2 .
Tangent Lines/Planes to Graphs
Fact: Suppose a curve in R2 is given as a graph y = f (x). The equation of
the tangent line at (a, f (a)) is:
y = f (a) + f 0 (a)(x − a).
Okay, you knew this from single-variable calculus. How does the multivari-
able case work? Well:

Fact: Suppose a surface in R3 is given as a graph z = f (x, y). The equation

of the tangent plane at (a, b, f (a, b)) is:
∂f ∂f
z = f (a, b) + (a, b)(x − a) + (a, b)(y − b).
∂x ∂y
Notice the similarity between this and the linear approximation to f at (a, b).

Tangent Lines/Planes to Level Sets

Def: For a function F : Rn → R, its gradient is the vector in Rn given by:
 
∂F ∂F ∂F
∇F = , , ..., .
∂x1 ∂x2 ∂xn

Theorem: Consider a level set F (x1 , . . . , xn ) = c of a function F : Rn → R.

If (a1 , . . . , an ) is a point on the level set, then ∇F (a1 , . . . , an ) is normal to
the level set.

Corollary 1: Suppose a curve in R2 is given as a level curve F (x, y) = c.

The equation of the tangent line at a point (x0 , y0 ) on the level curve is:
∂F ∂F
(x0 , y0 )(x − x0 ) + (x0 , y0 )(y − y0 ) = 0.
∂x ∂y
Corollary 2: Suppose a surface in R3 is given as a level surface F (x, y, z) = c.
The equation of the tangent plane at a point (x0 , y0 , z0 ) on the level surface
is:
∂F ∂F ∂F
(x0 , y0 , z0 )(x − x0 ) + (x0 , y0 , z0 )(y − y0 ) + (x0 , y0 , z0 )(z − z0 ) = 0.
∂x ∂y ∂z
Q: Do you see why Cor 1 and Cor 2 follow from the Theorem?
Composition and Matrix Multiplication
Recall: Let f : X → Y and g : Y → Z be functions. Their composition is
the function g ◦ f : X → Z defined by

(g ◦ f )(x) = g(f (x)).

Observations:
(1) For this to make sense, we must have: co-domain(f ) = domain(g).
(2) Composition is not generally commutative: that is, f ◦ g and g ◦ f are
usually different.
(3) Composition is always associative: (h ◦ g) ◦ f = h ◦ (g ◦ f ).

Fact: If T : Rk → Rn and S : Rn → Rm are both linear transformations, then

S ◦ T is also a linear transformation.

Question: How can we describe the matrix of the linear transformation S ◦T

in terms of the matrices of S and T ?

Fact: Let T : Rn → Rn and S : Rn → Rm be linear transformations with

matrices B and A, respectively. Then the matrix of S ◦ T is the product AB.

We can multiply an m × n matrix A by an n × k matrix B. The result,

AB, will be an m × k matrix:

(m × n)(n × k) → (m × k).

Notice that n appears twice here to “cancel out.” That is, we need the number
of rows of A to equal the number of columns of B – otherwise, the product
AB makes no sense.

Example 1: Let A be a (3 × 2)-matrix, and let B be a (2 × 4)-matrix. The

product AB is then a (3 × 4)-matrix.

Example 2: Let A be a (2 × 3)-matrix, and let B be a (4 × 2)-matrix. Then

AB is not defined. (But the product BA is defined: it is a (4 × 3)-matrix.)
Chain Rule
Chain Rule (Matrix Form): Let f : Rn → Rm and g : Rm → Rp be any
differentiable functions. Then

D(g ◦ f )(x) = Dg(f (x)) · Df (x).

Here, the product on the right-hand side is a product of matrices.

Many texts describe the chain rule in the following more classical form.
While there is a “classical” form in the general case of functions g : Rm → Rp ,
we will keep things simple and only state the case of functions g : Rm → R
with codomain R.

Chain Rule (Classical Form): Let g = g(x1 , . . . , xm ) and suppose each

x1 , . . . , xm is a function of the variables t1 , . . . , tn . Then:
∂g ∂g ∂x1 ∂g ∂x2 ∂g ∂xm
= + + ··· + ,
∂t1 ∂x1 ∂t1 ∂x2 ∂t1 ∂xm ∂t1
..
.
∂g ∂g ∂x1 ∂g ∂x2 ∂g ∂xm
= + + ··· + .
∂tn ∂x1 ∂tn ∂x2 ∂t1 ∂xm ∂tn

Example 1: Let z = g(u, v), where u = h(x, y) and v = k(x, y). Then the
chain rule reads:
∂z ∂z ∂u ∂z ∂v
= +
∂x ∂u ∂x ∂v ∂x
and
∂z ∂z ∂u ∂z ∂v
= + .
∂y ∂u ∂y ∂v ∂y

Example 2: Let z = g(u, v, w), where u = h(t), v = k(t), w = `(t). Then

the chain rule reads:
∂z ∂z ∂u ∂z ∂v ∂z ∂w
= + + .
∂t ∂u ∂t ∂v ∂t ∂w ∂t
Since u, v, w are functions of just a single variable t, we can also write this
formula as:
∂z ∂z du ∂z dv ∂z dw
= + + .
∂t ∂u dt ∂v dt ∂w dt
For Clarification: The Two Forms of the Chain Rule
Q: How exactly are the two forms of the chain rule the same?
A: If we completely expand the matrix form, writing out everything in compo-
nents, we end up with the classical form. The following examples may clarify.

Example 1: Suppose z = g(u, v), where u = h(x, y) and v = k(x, y).

This setup means we essentially have two functions:
g : R2 → R and f : R2 → R2
g(u, v) = z f (x, y) = (h(x, y), k(x, y)) = (u, v).
The matrix form of the Chain Rule reads:
D(g ◦ f )(x, y) = Dg(f (x, y)) · Df (x, y)
" #
h i  ∂h ∂h
∂(g◦f ) ∂(g◦f ) ∂g ∂g
 ∂x ∂y
∂x , ∂y
= ∂u , ∂v ∂k ∂k
∂x ∂y
h i h i
∂z ∂z ∂g ∂h ∂g ∂k ∂g ∂h ∂g ∂k
∂x , ∂y = ∂u ∂x + ∂v ∂x , ∂u ∂y + ∂v ∂y

Setting components equal to each other, we conclude that

∂z ∂g ∂h ∂g ∂k
= +
∂x ∂u ∂x ∂v ∂x
and
∂z ∂g ∂h ∂g ∂k
= + .
∂y ∂u ∂y ∂v ∂y
This is exactly the classical form of the Chain Rule. 

Example 2: Suppose z = g(u, v, w), where u = h(t), v = k(t), w = `(t).

This setup means we essentially have two functions:
g : R3 → R and f : R → R3
g(u, v, w) = z f (t) = (h(t), k(t), `(t)) = (u, v, w).
The matrix form of the Chain Rule reads:
D(g ◦ f )(t) = Dg(f (t)) · Df (t)
 ∂h 
h
∂(g◦f )
i 
∂g ∂g
 ∂t
∂g  ∂k 
∂t
= ∂u , ∂v , ∂w ∂t
∂`
∂t
∂z ∂g ∂h ∂g ∂k ∂g ∂`
= + + .
∂t ∂u ∂t ∂v ∂t ∂w ∂t
Again, we recovered the classical form of the Chain Rule. 
Inverses: Abstract Theory
Def: A function f : X → Y is invertible if there is a function f −1 : Y → X
satisfying:
f −1 (f (x)) = x, for all x ∈ X, and
f (f −1 (y)) = y, for all y ∈ Y.
In such a case, f −1 is called an inverse function for f .

In other words, the function f −1 “undoes” the function f . For example,

√
an inverse function of f : R → R, f (x) = x3 is f −1 : R → R, f −1 (x) = 3 x.
An inverse of g : R → (0, ∞), g(x) = 2x is g −1 : (0, ∞) → R, g −1 (x) = log2 (x).

Whenever a new concept is defined, a mathematician asks two questions:

(1) Uniqueness: Are inverses unique? That is, must a function f have at
most one inverse f −1 , or is it possible for f to have several different inverses?
Answer: Yes.

Prop 16.1: If f : X → Y is invertible (that is, f has an inverse), then the

inverse function f −1 is unique (that is, there is only one inverse function).

(2) Existence: Do inverses always exist? That is, does every function f
have an inverse function f −1 ?
Answer: No. Some functions have inverses, but others don’t.

New question: Which functions have inverses?

Prop 16.3: A function f : X → Y is invertible if and only if f is both “one-

to-one” and “onto.”
Despite their fundamental importance, there’s no time to talk about “one-
to-one” and “onto,” so you don’t have to learn these terms. This is sad :-(

Question: If inverse functions “undo” our original functions, can they help
us solve equations? Yes! That’s the entire point:

Prop 16.2: A function f : X → Y is invertible if and only if for every b ∈ Y ,

the equation f (x) = b has exactly one solution x ∈ X.
In this case, the solution to the equation f (x) = b is given by x = f −1 (b).
Inverses of Linear Transformations
Question: Which linear transformations T : Rn → Rm are invertible? (Equiv:
Which m × n matrices A are invertible?)

Fact: If T : Rn → Rm is invertible, then m = n.

So: If an m × n matrix A is invertible, then m = n.

In other words, non-square matrices are never invertible. But square ma-
trices may or may not be invertible. Which ones are invertible? Well:

Theorem: Let A be an n × n matrix. The following are equivalent:

(i) A is invertible
(ii) N (A) = {0}
(iii) C(A) = Rn
(iv) rref(A) = In
(v) det(A) 6= 0.

To Repeat: An n × n matrix A is invertible if and only if for every b ∈ Rn ,

the equation Ax = b has exactly one solution x ∈ Rn .
In this case, the solution to the equation Ax = b is given by x = A−1 b.

Q: How can we find inverse matrices? This is accomplished via:

Prop 16.7: If A is an invertible matrix, then rref[A | In ] = [In | A−1 ].

 
a b
Useful Formula: Let A=
c d
be a 2×2 matrix. If A is invertible (det(A) =
ad − bc 6= 0), then:  
−1 1 d −b
A = .
ad − bc −c a

Prop 16.8: Let f : X → Y and g : Y → Z be invertible functions. Then:

(a) f −1 is invertible and (f −1 )−1 = f .
(b) g ◦ f is invertible and (g ◦ f )−1 = f −1 ◦ g −1 .

Corollary: Let A, B be invertible n × n matrices. Then:

(a) A−1 is invertible and (A−1 )−1 = A.
(b) AB is invertible and (AB)−1 = B −1 A−1 .
The Gradient: Two Interpretations
Recall: For a function F : Rn → R, its gradient is the vector in Rn given
by:  
∂F ∂F ∂F
∇F = , , ..., .
∂x1 ∂x2 ∂xn
There are two ways to think about the gradient. They are interrelated.

Gradient: Normal to Level Sets

Theorem: Consider a level set F (x1 , . . . , xn ) = c of a function F : Rn → R.
If (a1 , . . . , an ) is a point on the level set, then ∇F (a1 , . . . , an ) is normal to
the level set.

Example: If we have a level curve F (x, y) = c in R2 , the gradient vector

∇F (x0 , y0 ) is a normal vector to the level curve at the point (x0 , y0 ).

Example: If we have a level surface F (x, y, z) = c in R3 , the gradient vector

∇F (x0 , y0 , z0 ) is a normal vector to the level surface at the point (x0 , y0 , z0 ).

Normal vectors help us find tangent planes to level sets (see the handout
“Tangent Lines/Planes...”) But there’s another reason we like normal vectors.

Gradient: Direction of Steepest Ascent for Graphs

Observation: A normal vector to a level set F (x1 , . . . , xn ) = c in Rn is the
direction of steepest ascent for the graph z = F (x1 , . . . , xn ) in Rn+1 .

Example (Elliptic Paraboloid): Let f : R2 → R be f (x, y) = 2x2 + 3y 2 .

The level sets of f are the ellipses 2x2 + 3y 2 = c in R2 .
The graph of f is the elliptic paraboloid z = 2x2 + 3y 2 in R3 . 
4
At the point (1, 1) ∈ R2 , the gradient vector ∇f (1, 1) = is normal to
6
the level curve 2x2 +3y 2 = 5. So, if we were hiking on the surface z = 2x2 +3y 2
in R3 and were at the point (1, 1, f (1, 1)) = (1,1, 5), to ascend the surface
4
the fastest, we would hike in the direction of . 
6
Warning: Note that ∇f is normal to the level sets of f . It is not a normal
vector to the graph of f .
Directional Derivatives
Def: For a function f : Rn → R, its directional derivative at the point
x ∈ Rn in the direction v ∈ Rn is:
Dv f (x) = ∇f (x) · v.
Here, · is the dot product of vectors. Therefore,
Dv f (x) = k∇f (x)kkvk cos θ, where θ = ](∇f (x), v).
Usually, we assume that v is a unit vector, meaning kvk = 1.
 
a
Example: Let f : R2 → R. Let v = . Then:
b
  " ∂f #  
Dv f (x, y) = ∇f (x, y) ·
a
= ∂f∂x · a = a ∂f + b ∂f .
b ∂y b ∂x ∂y

In particular, we have two important special cases:

 
1 ∂f
De1 f (x, y) = ∇f (x, y) · =
0 ∂x
 
0 ∂f
De2 f (x, y) = ∇f (x, y) · = .
1 ∂y
Point: Partial derivatives are themselves examples of directional derivatives!

Namely, ∂f
∂x is the directional derivative of f in the e1 -direction, while
∂f
∂y
is the directional derivative in the e2 -direction.

Question: At a point a, in which direction v will the function f grow the

most? i.e.: At a given point a, for which unit vector v is Dv f (a) maximized?

Theorem 6.3: Fix a point a ∈ Rn .

(a) The directional derivative Dv f (a) is maximized when v points in the
same direction as ∇f (a).
(b) The directional derivative Dv f (a) is minimized when v points in the
opposite direction as ∇f (a).

In fact: The maximum and minimum values of Dv f (a) at the point a ∈ Rn

are k∇f (a)k and −k∇f (a)k. (Assuming we only care about unit vectors v.)
Determinants
There are two reasons why determinants are important:
(1) Algebra: Determinants tell us whether a matrix is invertible or not.
(2) Geometry: Determinants are related to area and volume.

Determinants: Algebra
Prop 17.3: An n × n matrix A is invertible ⇐⇒ det(A) 6= 0.
Moreover: if A is invertible, then
1
det(A−1 ) = .
det(A)
Properties of Determinants (17.2, 17.4):
(1) (Multiplicativity) det(AB) = det(A) det(B).
(2) (Alternation) Exchanging two rows of a matrix reverses the sign of the
determinant.
(3) (Multilinearity): First:
     
a1 a2 · · · an b1 b2 · · · bn a1 + b1 a2 + b2 · · · an + bn
 c21 c22 · · · c2n   c21 c22 · · · c2n   c21 c22 ··· c2n 
det . + det = det
     
.. .. .. .. .. . . .. .. .. .. .. 
 ..
   
. . .   . . . .   . . . . 
cn1 cn2 · · · cnn cn1 cn2 · · · cnn cn1 cn2 ··· cnn
and similarly for the other rows; Second:
   
ka11 ka12 · · · ka1n a11 a12 · · · a1n
 a21 a22 · · · a2n   a21 a22 · · · a2n 
det . = k det
   
. .
. . . .
.   .. .. . . .. 
 . . . .   . . . . 
an1 an2 · · · ann an1 an2 · · · ann
and similarly for the other rows. Here, k ∈ R is any scalar.

Warning! Multilinearity does not say that det(A + B) = det(A) + det(B).

It also does not say det(kA) = k det(A). But: det(kA) = k n det(A) is true.

Determinants: Geometry
Prop 17.5: Let A be any 2 × 2 matrix. Then the area of the parallelogram
generated by the columns of A is |det(A)|.

Prop 17.6: Let T : R2 → R2 be a linear transformation with matrix A. Let

R be a region in R2 . Then:
Area(T (R)) = |det(A)| · Area(R).
Coordinate Systems
Def: Let V be a k-dim subspace of Rn . Each basis B = {v1 , . . . , vk } deter-
mines a coordinate system on V .
That is: Every vector v ∈ V can be written uniquely as a linear combina-
tion of the basis vectors:

v = c1 v1 + · · · + ck vk .

We then call c1 , . . . , ck the coordinates of v with respect to the basis B. We

then write  
c1
c 
 2
[v]B =  .. .
.
ck
Note that [v]B has k components, even though v ∈ Rn .

Note: Levandosky (L21: p 145-149) explains all this very clearly, in much
more depth than this review sheet provides. The examples are also quite
good: make sure you understand all of them.

Def: Let B = {v1 , . . . , vk } be a basis for a k-dim subspace V of Rn . The

change-of-basis matrix for the basis B is:
 

C = v1 v2 · · · vk .

Every vector v ∈ V in the subspace V can be written

v = c1 v1 + · · · + ck vk .

In other words:
v = C[v]B .
This formula tells us how to go between the standard coordinates for v and
the B-coordinates of v.

Special Case: If V = Rn and B is a basis of Rn , then the matrix C will be

invertible, and therefore:
[v]B = C −1 v.