Higher order derivatives 1

Chapter 3 Higher order derivatives

You certainly realize from single-variable calculus how very important it is to use derivatives
of orders greater than one. The same is of course true for multivariable calculus. In particular,
we shall definitely want a “second derivative test” for critical points.

A. Partial derivatives
First we need to clarify just what sort of domains we wish to consider for our functions.
So we give a couple of definitions.

DEFINITION. Suppose x0 ∈ A ⊂ Rn . We say that x0 is an interior point of A if there

exists r > 0 such that B(x0 , r) ⊂ A.

DEFINITION. A set A ⊂ Rn is said to be open if every point of A is an interior point of

A.

EXAMPLE. An open ball is an open set. (So we are fortunate in our terminology!) To see
this suppose A = B(x, r). Then if y ∈ A, kx − yk < r and we can let
² = r − kx − yk.
Then B(y, ²) ⊂ A. For if z ∈ B(y, ²), then the triangle inequality implies
kz − xk ≤ kz − yk + ky − xk
< ² + ky − xk
= r,
so that z ∈ A. Thus B(y, ²) ⊂ B(x, r) (cf. Problem 1–34). This proves that y is an interior
point of A. As y is an arbitrary point of A, we conclude that A is open.

PROBLEM 3–1. Explain why the empty set is an open set and why this is actually a
problem about the logic of language rather than a problem about mathematics.

PROBLEM 3–2. Prove that if A and B are open subsets of Rn , then their union
A ∪ B and their intersection A ∩ B are also open.
2 Chapter 3

PROBLEM 3–3. Prove that the closed ball B(x, r) is not open.

PROBLEM 3–4. Prove that an interval in R is an open subset ⇐⇒ it is of the form

(a, b), where −∞ ≤ a < b ≤ ∞. Of course, (a, b) = {x ∈ R|a < x < b}.

PROBLEM 3–5. Prove that an interval in R is never an open subset of R2 . Specifically,

this means that the x-axis is not an open subset of the x − y plane. More generally, prove
that if we regard Rk as a subset of Rn , for 1 ≤ k ≤ n − 1, by identifying Rk with the
Cartesian product Rk × {0}, where 0 is the origin in Rn−k , then no point of Rk is an
interior point relative to the “universe” Rn .

f
NOTATION. Suppose A ⊂ Rn is an open set and A − → R is differentiable at every point
x ∈ A. Then we say that f is differentiable on A. Of course, we know then that in particular
the partial derivatives ∂f /∂xj all exist. It may happen that ∂f /∂xj itself is a differentiable
function on A. Then we know that its partial derivatives also exist. The notation we shall
use for the latter partial derivatives is
µ ¶
∂ 2f ∂ ∂f
= .
∂xi ∂xj ∂xi ∂xj

Notice the care we take to denote the order in which these differentiations are performed. In
case i = j we also write µ ¶
∂ 2f ∂2f ∂ ∂f
= = .
∂x2i ∂xi ∂xi ∂xi ∂xi
By the way, if we employ the subscript notation (p. 2–20) for derivatives, then we have logically

∂ 2f
fxj xi = (fxj )xi = .
∂xi ∂xj

Partial derivatives of higher orders are given a similar notation. For instance, we have

∂ 7f
= fxzzyxxx .
∂x3 ∂y∂z 2 ∂x
Higher order derivatives 3

f
PROBLEM 3–6. Let (0, ∞) × R −
→ R be defined as

f (x, y) = xy .

Compute fxx , fxy , fyx , fyy .

f
PROBLEM 3–7. Let R2 −
→ R be defined for x 6= 0 as f (x) = log kxk. Show that

∂ 2f ∂ 2f
+ = 0.
∂x21 ∂x22

This is called Laplace’s equation on R2 .

f
PROBLEM 3–8. Let Rn →
− R be defined for x 6= 0 as f (x) = kxk2−n . Show that
n
X ∂2f
= 0.
j=1
∂x2j

(Laplace’s equation on Rn .)

ϕ
PROBLEM 3–9. Let R → R be a differentiable function whose derivative ϕ0 is also
f
differentiable. Define R2 −
→ R as

f (x, y) = ϕ(x + y).

Show that
∂ 2f ∂ 2f
− 2 = 0.
∂x2 ∂y
Do the same for g(x, y) = ϕ(x − y).

f
Now suppose A ⊂ Rn is open and A − → R is differentiable at every point x ∈ A. Then we
say f is differentiable on A. The most important instance of this is the case in which all the
partial derivatives ∂f /∂xj exist and are continuous functions on A — recall the important
4 Chapter 3

corollary on p. 2–32. We say in this case that f is of class C 1 and write

f ∈ C 1 (A).

(In case f is a continuous function on A, we write

f ∈ C 0 (A).)

Likewise, suppose f ∈ C 1 (A) and each ∂f /∂xj ∈ C 1 (A); then we say

f ∈ C 2 (A).

In the same way we define recursively C k (A), and we note of course the string of inclusions

· · · ⊂ C 3 (A) ⊂ C 2 (A) ⊂ C 1 (A) ⊂ C 0 (A).

The fact is that most functions encountered in calculus have the property that they have
continuous partial derivatives of all orders. That is, they belong to the class C k (A) for all k.
We say that these functions are infinitely (often) differentiable, and we write the collection of
such functions as C ∞ (A). Thus we have
∞
\
∞
C (A) = C k (A)
k=0

and
C ∞ (A) ⊂ · · · ⊂ C 2 (A) ⊂ C 1 (A) ⊂ C 0 (A).

PROBLEM 3–10. Show that even in single-variable calculus all the above inclusions
are strict. That is, for every k show that there exists f ∈ C k (R) for which f ∈
/ C k+1 (R).

We now turn to an understanding of the basic properties of C 2 (A). If f ∈ C 2 (A), then we

have defined the “pure” partial derivatives
µ ¶
∂ 2f ∂ ∂f
=
∂xi ∂xi ∂xi ∂xi
as well as the “mixed ” partial derivatives
µ ¶
∂ 2f ∂ ∂f
= .
∂xi ∂xj ∂xi ∂xj
Higher order derivatives 5

for i 6= j. Our next task is the proof that if f ∈ C 2 (A), then

∂ 2f ∂ 2f
=
∂xi ∂xj ∂xj ∂xi

(“the mixed partial derivatives are equal”). This result will clearly render calculations involv-
ing higher order derivatives much easier; we’ll no longer have to keep track of the order of
computing partial derivatives. Not only that, there are fewer that must be computed:

PROBLEM 3–11. If f ∈ C 2 (R2 ), then only three second order partial derivatives of
f need to be computed in order to know all four of its second order partial derivatives.
Likewise, if f ∈ C 2 (R3 ), then only six need to be computed in order to know all nine.
What is the result for the general case f ∈ C 2 (Rn )?

What is essentially involved in proving a result like this is showing that the two limiting
processes involved can be done in either order. You will probably not be surprised, then, that
there is pathology to be overcome. The following example is found in just about every text
on vector calculus. It is sort of a canonical example.

f
PROBLEM 3–12. Define R2 −
→ R by
(
xy(x2 −y 2 )
x2 +y 2
if (x, y) 6= 0,
f (x, y) =
0 if (x, y) = 0.
Prove that

a. f ∈ C 1 (R2 ),
∂ 2f ∂ 2f ∂2f ∂ 2f
b. , , , exist on R2 ,
∂x2 ∂x∂y ∂y∂x ∂y 2
∂2f ∂ 2f
c. (0, 0) 6= (0, 0).
∂x∂y ∂y∂x

All pathology of this nature can be easily eliminated by a reasonable assumption on the
function, as we now demonstrate.
6 Chapter 3

THEOREM. Let A ⊂ Rn be an open set and let f ∈ C 2 (A). Then

∂ 2f ∂2f
= .
∂xi ∂xj ∂xj ∂xi

PROOF. Since we need only consider a fixed pair i, j in the proof, we may as well assume
i = 1, j = 2. And since x3 , . . . , xn remain fixed in all our deliberations, we may also assume
that n = 2, so that A ⊂ R2 .
Let x ∈ A be fixed, and let δ > 0 and ² > 0 be arbitrary but small enough that the points
considered below belong to A (remember, x is an interior point of A).
The first step in our proof is the writing of a combination of four values of f near the point
x. Then we shall perform a limiting procedure to obtain the desired result. Before defining
the crucial expression we present the following guideline to our reasoning. Thus we hope that
approximately

∂f f (x1 + δ, x2 ) − f (x1 , x2 )
(x1 , x2 ) ≈ ,
∂x1 δ
∂f f (x1 + δ, x2 + ²) − f (x1 , x2 + ²)
(x1 , x2 + ²) ≈ .
∂x1 δ

Now subtract and divide by ²:

∂f ∂f
∂ 2f ∂x1
(x1 , x2 + ²) − ∂x1
(x1 , x2 )
(x1 , x2 ) ≈
∂x2 ∂x1 ²
f (x1 +δ,x2 +²)−f (x1 ,x2 +²) f (x1 +δ,x2 )−f (x1 ,x2 )
δ
− δ
≈ .
²
Thus we hope that

∂ 2f
≈ δ −1 ²−1 [f (x1 + δ, x2 + ²) − f (x1 , x2 + ²) − f (x1 + δ, x2 ) + f (x1 , x2 )] .
∂x2 ∂x1

Now we come to the actual proof. Let

D(δ, ²) = f (x1 + δ, x2 + ²) − f (x1 , x2 + ²) − f (x1 + δ, x2 ) + f (x1 , x2 ).

Then define the auxiliary function (this is a trick!) of one real variable,

g(y) = f (x1 + δ, y) − f (x1 , y).

Higher order derivatives 7

By the mean value theorem of single-variable calculus,

g(x2 + ²) − g(x2 ) = ²g 0 (t)

for some x2 < t < x2 + ². That is,

· ¸
∂f ∂f
D(δ, ²) = ² (x1 + δ, t) − (x1 , t) .
∂x2 ∂x2

Now we use the mean value theorem once again, this time for the term in brackets:

∂f ∂f ∂ 2f
(x1 + δ, t) − (x1 , t) = δ (s, t)
∂x2 ∂x2 ∂x1 ∂x2

for some x1 < s < x1 + δ. The result is

∂ 2f
D(δ, ²) = δ² (s, t).
∂x1 ∂x2

Here is a diagram illustrating this situation:

Now carry out the very same procedure, except with the roles of x1 and x2 interchanged.
The result can be immediately written down as

∂ 2f
D(δ, ²) = δ² (s0 , t0 ).
∂x2 ∂x1
8 Chapter 3

Comparing the two equations we have thus obtained, we find after dividing by δ²,

∂ 2f ∂ 2f
(s, t) = (s0 , t0 ).
∂x1 ∂x2 ∂x2 ∂x1

Finally, we let ² → 0 and at this time use the continuity of the two mixed partial derivatives
to obtain
∂ 2f ∂ 2f
(x) = (x).
∂x1 ∂x2 ∂x2 ∂x1

QED

REMARK. The proof we have presented certainly does not require the full hypothesis that
f ∈ C 2 (D). It requires only that each of ∂ 2 f /∂xi ∂xj and ∂ 2 f /∂xj ∂xi be continuous at the
point in question. As I have never seen an application of this observation, I have chosen
to state the theorem in its weaker version. This will provide all we shall need in practice.
However, the weakened hypothesis does indeed give an interesting result. Actually, the same
proof outline provides even better results with very little additional effort:

PROBLEM 3–13∗ . Assume only that ∂f /∂xi and ∂f /∂xj exist in A and that
∂ f /∂xi ∂xj exists in A and is continuous at x. Prove that ∂ 2 f /∂xj ∂xi exists at x and
2

that
∂ 2f ∂ 2f
= at x.
∂xi ∂xj ∂xj ∂xi
(HINT: start with the equation derived above,

∂ 2f
δ −1 ²−1 D(δ, ²) = (s, t).
∂x1 ∂x2
The right side of this equation has a limit as δ and ² tend to zero. Therefore so does the
left side. Perform this limit in the iterated fashion
³ ´
lim lim (left hand side)
²→0 δ→0

to obtain the result.)

Higher order derivatives 9

PROBLEM 3–14∗ . Assume only that ∂f /∂xi and ∂f /∂xj exist in A and are themselves
differentiable at x. Prove that
∂ 2f ∂ 2f
= at x.
∂xi ∂xj ∂xj ∂xi

(Hint: start with the expression for D(δ, ²) on p. 3–6. Then use the hypothesis that
∂f /∂x2 is differentiable at x to write

∂f ∂f ∂ 2f ∂2f
(x1 + δ, t) = (x1 , x2 ) + (x1 , x2 )δ + 2 (x1 , x2 )(t − x2 )
∂x2 ∂x2 ∂x1 ∂x2 ∂x2
+ small multiple of (δ + (t − x2 ))
∂f ∂f ∂ 2f
(x1 , t) = (x1 , x2 ) + 2 (x1 , x2 )(t − x2 )
∂x2 ∂x2 ∂x2
+ small multiple of (t − x2 ).

Conclude that
∂ 2f
D(δ, ²) = ²δ + ² (small multiple of (δ + ²))
∂x1 ∂x2

and carefully analyze how small the remainder is. Conclude that the limit of ²−2 D(², ²)
as ² tends to zero is the value of ∂ 2 f /∂x1 ∂x2 at (x1 , x2 ).

PROBLEM 3–15. Prove that if f is of class C k on an open set A, then any mixed par-
tial derivative of f of order k can be computed without regard to the order of performing
the various differentiations.

B. Taylor’s theorem (Brook Taylor, 1712)

We are now preparing for our study of critical points of functions on Rn , and we specifically
want to devise a “second derivative test” for the type of critical point. The efficient theoretical
way to study this question is to approximate the given function by an expression (the Taylor
polynomial) which is a polynomial of degree 2. The device for doing this is called Taylor’s
theorem.
We begin with a quick review of the single-variable calculus version. Suppose that g is a
10 Chapter 3

function from R to R which is defined and of class C 2 near the origin. Then the fundamental
theorem of calculus implies Z t
g(t) = g(0) + g 0 (s)ds.
0
Integrate partially after a slight adjustment:
Z t
g(t) = g(0) − g 0 (s)d(t − s) (t is fixed)
0
¯s=t Z t
¯
= g(0) − g (s)(t − s) ¯¯
0
+ g 00 (s)(t − s)ds
0
Z t s=0
= g(0) + g 0 (0)t + g 00 (s)(t − s)ds.
0

Now we slightly adjust the remaining integral by writing

³ ´
g 00 (s) = g 00 (0) + g 00 (s) − g 00 (0)

and doing the explicit integration with the g 00 (0) term:

Z t Z t
00 00
g (s)(t − s)ds = g (0) (t − s)ds + R
0 0
¯s=t
00 (t − s)2 ¯¯
= g (0) +R
−2 ¯s=0
1
= g 00 (0)t2 + R.
2
Here R, the “remainder,” is of course
Z t³ ´
00 00
g (s) − g (0) (t − s)ds.
0

Suppose
|g 00 (s) − g 00 (0)| ≤ Q for s between 0 and t.
Then ¯Z t ¯
¯ ¯ 1
|R| ≤ ¯ Q(t − s)ds ¯ = Qt2 .
0 2
(The absolute value signs around the integral are necessary if t < 0.) So the version of Taylor’s
theorem we obtain is
1
g(t) = g(0) + g 0 (0)t + g 00 (0)t2 + R,
2
Higher order derivatives 11

where
1 2
|R| ≤ Qt .
2
(By the way, if g is of class C 3 , then |g 00 (s) − g 00 (0)| ≤ C|s| for some constant C (Lipschitz
condition), so that
¯Z t ¯ 1
¯ ¯
|R| ≤ ¯ Cs(t − s)ds ¯ = C|t|3 .
0 6
Thus the remainder term in the Taylor series tends to zero one degree faster than the quadratic
terms in the Taylor polynomial. This is the usual way of thinking of Taylor approximations.)
f
Now we turn to the case Rn − → R, and we work first with the Taylor series centered at the
origin. (We can easily extend to other points later.) We assume f is of class C 2 near 0. If
x ∈ Rn is close to 0, we use the familiar ploy of moving between 0 and x along a straight line:
we thus define the composite function x

g(t) = f (tx),
tx for 0 < t < 1
O

regarding x as fixed but small. We then have as above, with t = 1,

1
g(1) = g(0) + g 0 (0) + g 00 (0) + R,
2
where
1
|R| ≤ Q,
2
where
|g 00 (s) − g 00 (0)| ≤ Q for 0 ≤ s ≤ 1.
Now we apply the chain rule:
n
X
0
g (t) = (Di f )(tx)xi
i=1

(Di f = ∂f /∂xi ) and again

n X
X n
g 00 (t) = (Dj Di f )(tx)xj xi .
i=1 j=1
12 Chapter 3

In particular,
n
X
g 0 (0) = Di f (0)xi ,
i=1
n
X
g 00 (0) = Di Dj f (0)xi xj .
i,j=1

We now see what the statement of theorem should be:

f
THEOREM. Let Rn − → R be of class C 2 near 0. Let ² > 0. Then there exists δ > 0 such
that if kxk ≤ δ, then
n
X n
1X
f (x) = f (0) + Di f (0)xi + Di Dj f (0)xi xj + R,
i=1
2 i,j=1

where
|R| ≤ ²kxk2 .

PROOF. We continue with the above development. Because f is of class C 2 , its second
order partial derivatives are all continuous. Hence given ² > 0, we may choose δ > 0 such that
for all kyk ≤ δ,
|Di Dj f (y) − Di Dj f (0)| ≤ ².
Then if kxk ≤ δ, we have a pretty crude estimate valid for 0 ≤ s ≤ 1:
¯Xn ¯
00 00 ¯ ¯
|g (s) − g (0)| = ¯ (Di Dj f (sx) − Di Dj f (0)) xi xj ¯
i,j=1
n
X
≤ ²|xi | |xj |
i,j=1
à n
!2
X
= ² |xi |
i=1
n
X n
X
≤ ² 1· x2i (Schwarz inequality)
i=1 i=1
= ²nkxk2 .
Higher order derivatives 13

The remainder term above therefore satisfies

1
|R| ≤ ²nkxk2 .
2
To obtain the final result, first replace ² by 2²/n.
QED

B0 . Taylor’s theorem bis

There is actually a much different proof of Taylor’s theorem which enables us to weaken
the hypothesis significantly and still obtain the same conclusion. As far as applications to
calculus are concerned, the C 2 version we have proved above is certainly adequate, but the
improved version presented here is nevertheless quite interesting for its minimal hypothesis.
f
THEOREM. Let Rn − → R be differentiable in a neighborhood of 0, and assume each partial
derivative ∂f /∂xi is differentiable at 0. Let ² > 0. Then there exists δ > 0 such that if
kxk ≤ δ, then
Xn n
1X
f (x) = f (0) + Di f (0)xi + Di Dj f (0)xi xj + R,
i=1
2 i,j=1

where
|R| ≤ ²kxk2 .

PROOF. We still use the function g(t) = f (tx), and we still have the chain rule
n
X
0
g (t) = Di f (tx)xi .
i=1

However, now we use the hypothesis that Di f is differentiable at 0 as follows: for any ² > 0
there exists δ > 0 such that
n
X
Di f (y) = Di f (0) + Dj Di f (0)yj + Ri (y),
j=1

where
|Ri (y)| ≤ ² kyk for all kyk ≤ δ.
14 Chapter 3

Therefore, if kxk ≤ δ and 0 ≤ t ≤ 1, we obtain

n
X n X
X n
0
g (t) = Di f (0)xi + Dj Di f (0)txj xi + R(t, x),
i=1 i=1 j=1

where
¯Xn ¯
¯ ¯
|R(t, x)| = ¯ xi Ri (tx) ¯
i=1
n
X
≤ |xi |²t kxk
i=1
√
≤ ²t n kxk2 .

At this point we would like to apply a theorem from single-variable calculus known as the
Cauchy mean value theorem. This is used in most proofs of l’Hôpital’s rule. However, instead
of appealing to this theorem we insert its proof in this context. Define

ϕ(t) = g(t) − g(0) − g 0 (0)t − [g(1) − g(0) − g 0 (0)]t2 ,

noting that ϕ(0) = ϕ(1) = 0 and that ϕ is differentiable for 0 ≤ t ≤ 1. Thus the mean value
theorem implies ϕ0 (t) = 0 for some 0 < t < 1. That is,

[g(1) − g(0) − g 0 (0)]2t = g 0 (t) − g 0 (0)

X n
= Dj Di f (0)txj xi + R(t, x).
i,j=1

Dividing by 2t gives
n
X n
1X R(t, x)
f (x) − f (0) − Di f (0)xi = Dj Di f (0)xj xi + ,
i=1
2 i,j=1 2t

where ¯ R(t, x) ¯
¯ ¯ ²√
¯ ¯≤ n kxk2 .
2t 2
QED
Higher order derivatives 15

PROBLEM 3–16. Here’s a simplification of the proof of the theorem in case n = 1.

Calculate
f (x) − f (0) − f 0 (0)x
lim
x→0 x2
and then deduce the theorem.

The hypothesis of this version of Taylor’s theorem is rather minimal. It is strong enough
to guarantee the equality of the mixed second order partial derivatives of f at 0, as we realize
from Problem 3–14 (although we did not require this fact in the proof). The following problem
shows that there is little hope of further weakening the hypothesis.

PROBLEM 3–17. Let f be the “canonical” pathological function of Problem 3–12.

This function is of class C 1 on R2 and has the origin as a critical point. Show that there
do not exist constants c1 , c2 , c3 such that

f (x, y) = c1 x2 + c2 xy + c3 y 2 + R(x, y)

and
R(x, y)
lim = 0.
(x,y)→(0,0) x2 + y 2

C. The second derivative test for R2

To get started with this analysis, here is a problem which you very much need to investigate
carefully.
16 Chapter 3

PROBLEM 3–18. Let A, B, C be real numbers such that

A > 0, C > 0, AC − B 2 > 0.

a. Prove that a number λ > 0 exists such that for all (x, y) ∈ R2 ,

Ax2 + 2Bxy + Cy 2 ≥ λ(x2 + y 2 ).

b. Calculate the largest possible λ.

q
(Answer: 2 − ( A+C
A+C
2
)2 − (AC − B 2 ))

Now we are ready to give the second derivative test. We suppose that we are dealing
f
with a function R2 −
→ R which is of class C 2 in a neighborhood of the origin (we shall later
immediately generalize to any point). Using coordinates (x, y) for R2 , we obtain from Taylor’s
theorem in Section B
f (x, y) = f (0, 0) + D1 f (0, 0)x + D2 f (0, 0)y
| {z } | {z }
0 order term 1st order terms
1¡ ¢
+ D11 f (0, 0)x2 + 2D12 f (0, 0)xy + D22 f (0, 0)y 2 + R(x, y),
|2 {z } | {z }
2nd order terms remainder term

and the remainder term satisfies the condition that for any ² > 0 there exists δ > 0 such that
x2 + y 2 ≤ δ 2 ⇒ |R(x, y)| ≤ ²(x2 + y 2 ).
(Here we have continued to employ the notation D11 f = ∂ 2 f /∂x2 etc.)
Now we suppose that (0, 0) is a critical point for f . That is, by definition,
D1 f (0, 0) = D2 f (0, 0) = 0.
Let us now denote
A = D11 f (0, 0) = ∂ 2 f /∂x2 (0, 0),
B = D12 f (0, 0) = ∂ 2 f /∂x∂y(0, 0),
C = D22 f (0, 0) = ∂ 2 f /∂y 2 (0, 0).
Then the Taylor expression for f takes the form
1
f (x, y) = f (0, 0) + (Ax2 + 2Bxy + Cy 2 ) + R(x, y).
2
Higher order derivatives 17

The underlying rationale for the second derivative test is that in the Taylor expression for
f the first order terms are absent, thanks to the fact that we have a critical point; therefore
we expect that the second order terms will determine the behavior of the function near the
critical point. In fact, this usually happens, as we now show.
Let us first assume that the numbers A, B, C satisfy the inequalities
A > 0 and C > 0,
AC − B 2 > 0.

We know from Problem 3–18 that there exists a positive number λ (depending only on A, B,
C) such that
Ax2 + 2Bxy + Cy 2 ≥ λ(x2 + y 2 ) for all (x, y) ∈ R2 .
Therefore, x2 + y 2 ≤ δ 2 ⇒
1
f (x, y) ≥ f (0, 0) + λ(x2 + y 2 ) − |R(x, y)|
2
1
≥ f (0, 0) + λ(x2 + y 2 ) − ²(x2 + y 2 )
2
µ ¶
1
= f (0, 0) + λ − ² (x2 + y 2 ).
2

This inequality is exactly what we need! For we may choose 0 < ² < 21 λ and then conclude
with the corresponding δ that

0 < x2 + y 2 ≤ δ ⇒ f (x, y) > f (0, 0).

The conclusion: at the critical point 0, f has a strict local minimum. (“Strict” in this
context means that near (0, 0) the equality f (x, y) = f (0, 0) holds only if (x, y) = (0, 0).)
If we assume instead that
A < 0 and C < 0,
AC − B 2 > 0,

then we obtain the corresponding conclusion: at the critical point 0, f has a strict local
maximum. The proof is immediate: we could either repeat the above type of analysis, or
we could simply replace f by −f and use the previous result. For replacing f by −f also
replaces A, B, C by their negatives, so that the assumed condition for f becomes the previous
condition for −f .
18 Chapter 3

Finally, assume
AC − B 2 < 0.
The quadratic form in question, Ax2 + 2Bxy + Cy 2 , can now be positive at some points,
negative at others. We check this quickly. If A 6= 0, then the equation Ax2 + 2Bxy + Cy 2 = 0
has distinct roots √
x −B ± B 2 − AC
=
y A
and therefore changes sign. The reason is clear. Renaming x/y as t, the function At2 +2Bt+C
has a graph which is a parabola (since A 6= 0). It hits the t axis at two distinct points, so
that its position is that of a parabola that genuinely crosses the t-axis. Likewise if C 6= 0. If
both A and C are 0, then we are dealing with 2Bxy, which has opposite signs at (say) (1, 1)
and (1, −1). (Notice that our assumption implies B 6= 0 in this case.)
Now suppose (x0 , y0 ) satisfies

Ax20 + 2Bx0 y0 + Cy02 = α < 0.

Then for small |t| we have

1
f (tx0 , ty0 ) = f (0, 0) + αt2 + R(tx0 , ty0 )
2
1
≤ f (0, 0) + αt2 + ²t2 (x20 + y02 )
2µ ¶
2 1 2 2
= f (0, 0) + t α + ²(x0 + y0 ) .
2

Thus if ² is small enough that 12 α + ²(x20 + y02 ) < 0 and |t| is sufficiently small and not zero,

f (tx0 , ty0 ) < f (0, 0).

Thus in any neighborhood of (0, 0) there are points where f < f (0, 0).
In the same way we show that in any neighborhood of (0, 0) there are points where f >
f (0, 0). The conclusion: at the critical point 0, f has neither a local minimum nor a
local maximum value.
To recapitulate, all the above analysis is based on the same observation: at a critical point
(0, 0), the behavior of f is governed by the quadratic terms in its Taylor expansion, at least in
the three cases we have considered. These three cases are all subcategories of the assumption

AC − B 2 6= 0.
Higher order derivatives 19

For if AC − B 2 > 0, then AC > 0 and we see that A and C are either both positive or both
negative. If AC − B 2 < 0, that is the third case.
The conclusion is that, if AC − B 2 6= 0, then the quadratic terms in the Taylor expansion
of f are determinative. If AC − B 2 = 0, we have no criterion.
For example, the function
f (x, y) = x2 + y 4
has the origin as a strict local minimum; its negative has the origin as a strict local maximum.
And the function
f (x, y) = x2 − y 4
has a critical point at the origin, which is neither a local minimum nor a local maximum. All
these examples satisfy AC − B 2 = 0.

PROBLEM 3–19. Find three functions f (x, y) which have (0, 0) as a critical point
and have AC = B = 1 (so that AC − B 2 = 0) such that the critical point (0, 0) is

a. a strict local minimum for one of them,

b. a strict local maximum for another,

c. neither a local minimum nor a local maximum for the other.

D. The nature of critical points

f
This section contains mostly definitions. We suppose that Rn −
→ R and that f is of class
1
C near a critical point x0 . Thus ∇f (x0 ) = 0. There are just three types of behavior for f
near x0 :
f has a local minimum at x0 :
f (x) ≥ f (x0 ) for all x ∈ B(x0 , ²) for some ² > 0.
f has a local maximum at x0 :
f (x) ≤ f (x0 ) for all x ∈ B(x0 , ²) for some ² > 0.
f has a saddle point at x0 :

neither of the above two conditions holds — in other words, for all ² > 0 there exist x,
x0 ∈ B(x0 , ²) such that f (x) < f (x0 ) < f (x0 ).

There is a further refinement in the first two cases:

f has a strict local minimum at x0 :
20 Chapter 3

f (x) > f (x0 ) for all 0 < kx − x0 k < ² for some ² > 0.
f has a strict local maximum at x0 :
f (x) < f (x0 ) for all 0 < kx − x0 k < ² for some ² > 0.
In Section C we analyzed the R2 case when the critical point was the origin. That is really
no loss of generality in view of the correspondence between the function f and its translated
version g given by
g(x) = f (x0 + x), x ∈ Rn .
For f has a critical point at x0 ⇐⇒ g has a critical point at 0. Moreover, the nature of the
critical points is the same. Moreover,
Di Dj g(0) = Di Dj f (x0 ),
so the information coming from the second derivatives is the same for f and g.
SUMMARY. We take this opportunity to summarize the R2 case. Assume
f
R2 −
→ R is of class C 2 near x0 , and x0 is a critical point of f .
Define
A = D1 D1 f (x0 ),
B = D1 D2 f (x0 ),
C = D2 D2 f (x0 ).
Assume
AC − B 2 6= 0.
Then we have the results:
A > 0 and C > 0,
=⇒ f has a strict local minimum at x0 .
AC − B 2 > 0

A < 0 and C < 0,

=⇒ f has a strict local maximum at x0 .
2
AC − B > 0

AC − B 2 < 0 =⇒ f has a saddle point at x0 .

In case AC − B 2 = 0, it is impossible to tell from A, B, C what the nature of the critical
point is. Here are some easy examples for the critical point (0, 0) in the x − y plane:
Higher order derivatives 21

1. f (x, y) = x2 local minimum

2. f (x, y) = −x2 local maximum (these are even global extrema)

3. f (x, y) = x2 + y 4 strict local minimum

4. f (x, y) = −x2 − y 4 strict local maximum

5. f (x, y) = x2 − y 4 saddle point

In each of the next five problems determine the nature of each critical point of the indicated
function on R2 .
1 1
PROBLEM 3–20. f (x, y) = x3 + y 2 + 2xy + 5x + y.
3 2
(Answer: one local minimum, one saddle)

PROBLEM 3–21. f (x, y) = x2 y 2 − 5x2 − 8xy − 5y 2 .

(Answer: one local maximum, four saddles)

PROBLEM 3–22. f (x, y) = xy(12 − 3x − 4y).

PROBLEM 3–23. f (x, y) = x4 + y 4 − 2x2 + 4xy − 2y 2 .

2 −y 2
PROBLEM 3–24. f (x, y) = (ax2 + by 2 )e−x (a > b > 0 are constants).

PROBLEM 3–25. The function of Problem 2–56 has just the critical point (1, 1). Use
the above test to determine its local nature.
(Answer: local minimum)

PROBLEM 3–26. The function of Problem 2–64 has just the critical point (1, 0), and
it was shown to be a local maximum. Use the above test to determine its local nature.
22 Chapter 3

PROBLEM 3–27. Here’s a function given me by Richard Stong:

f (x, y) = 24x + 6x2 y 2 + 6x2 y − x2 + x3 y 3 .

Show that it has just one critical point, which is a strict local maximum but not a global
maximum. (Stong’s point: this provides the same moral as does Problem 2–64, but with
a polynomial function.)

There is a beautiful geometric way to think of these criteria, in terms of level sets. For
f
instance, suppose R2 −→ R has a critical point at (0, 0), and that A = 2, B = 0, C = 1, so we
know the critical point is a minimum. If f were just a quadratic polynomial, then

1
f (x, y) = f (0, 0) + x2 + y 2 .
2

The level set description of f then is concentric ellipses. In case f is not a quadratic polynomial,
then near the critical point the level sets will still resemble ellipses, but with some distortion.
The distortion decreases as we approach the critical point.

Likewise, the level sets of a quadratic polynomial which is a saddle point are concentric
hyperbolas, but in general they will resemble distorted hyperbolas.
Higher order derivatives 23

E. The Hessian matrix

Now we return to the general n-dimensional case, so we shall be assuming throughout this
section that
f
Rn →
− R
is a function of class C 2 near its critical point x0 . The Taylor expansion from Section B
therefore becomes
n
1X
f (x0 + y) = f (x0 ) + Di Dj f (x0 )yi yj + R(y),
2 i,j=1

where for any ² > 0 the remainder satisfies

|R(y)| ≤ ² kyk2 if kyk is sufficiently small.
We are going to try to analyze this situation just as we did for the case n = 2, hoping that the
quadratic terms in the Taylor expansion are determinative. In order to begin our deliberations,
we single out the essence of the quadratic terms by giving the following terminology.

DEFINITION. The Hessian matrix of f at its critical point x0 is the n × n matrix

µ 2 ¶
∂ f
H = (x0 ) .
∂xi ∂xj
In case n = 2 and in the notation of the preceding sections,
µ ¶
A B
H = .
B C
24 Chapter 3

In case n = 1 we have simply

H = f 00 (x0 ).

Notice that in all cases the fact that f is of class C 2 implies that Di Dj f (x0 ) = Dj Di f (x0 ), in
other words that H is a symmetric matrix.
Here is a crucial concept:

DEFINITION. In the above situation the critical point x0 is nondegenerate if det H 6= 0.

And it is degenerate if det H = 0.
Our goal is to establish a “second derivative test” to determine the local nature of a
critical point. This will prove to be possible only if the critical point is nondegenerate. For
n = 1 this means f 00 (x0 ) 6= 0, and standard single-variable calculus gives a strict local minimum
if f 00 (x0 ) > 0 and a strict local maximum if f 00 (x0 ) < 0 (and no conclusion if f 00 (x0 ) = 0). For
n = 2 the nondegeneracy means AC − B 2 6= 0, and the results are given on p. 3–20.
To repeat: if the critical point is degenerate, no test involving only the second order partial
derivatives of f can detect the nature of a critical point. We shall therefore restrict our analysis
to nondegenerate critical points.

REMARK. If n = 1 there is no such thing as a nondegenerate saddle point. Thus the

structure of critical point behavior is much richer in multivariable calculus than in single-
variable calculus. It’s a good idea to keep in mind easy examples of nondegenerate saddle
points for n = 2, such as the origin for functions such as:

f (x, y) = xy,

or
f (x, y) = x2 − y 2 ,

or
f (x, y) = x2 + 5xy + y 2 .

Here is a simple but very interesting example which illustrates the sort of thing that can
happen at degenerate critical points.
Higher order derivatives 25

PROBLEM 3–28. Define

f (x, y) = (x − y 2 )(x − 2y 2 ).

a. Show that (0, 0) is the only critical point.

b. Show that it is degenerate.

c. Show that when f is restricted to any straight line through (0, 0), the resulting
function has a strict local minimum at (0, 0).

d. Show that (0, 0) is a saddle point.

We are going to discover that quadratic polynomials are much more complicated alge-
braically for dimensions greater than two than for dimension two.

PROBLEM 3–29. The quadratic polynomial below corresponds to a Hessian matrix

 
1 −1 1
−1 3 0 .
1 0 2

Use whatever ad hoc method you can devise to show that

x2 + 3y 2 + 2z 2 − 2xy + 2xz > 0

except at the origin.

REMARK. Since the terms in the above expression are purely quadratic, the homogeneity
argument you may have used for Problem 3–18 shows that there exists a positive number λ
such that
x2 + 3y 2 + 2z 2 − 2xy + 2xz ≥ λ(x2 + y 2 + z 2 )
for all (x, y, z) ∈ R3 . However, the problem of finding the largest such λ is a difficult algebra
problem, equivalent to finding a root of a certain cubic equation. We are very soon going to
discuss this algebra. (The maximal λ is approximately 0.1206.)
We have just defined nondegenerate critical points by invoking the determinant of the
associated Hessian matrix. Before we proceed further, we need to pause to define determinants
and develop their properties.
26 Chapter 3

F. Determinants
Every n×n matrix (every square matrix) has associated with it a number which is called its
determinant. We are going to denote the determinant of A as det A. This section is concerned
with the definition and properties of this extremely important function.
If n = 1, a 1 × 1 matrix is just a real number, so we define det a = a.
If n = 2, then we rely on our experience to define
µ ¶
a11 a12
det A = det = a11 a22 − a12 a21 .
a21 a22

Some students at this stage also have experience with 3 × 3 matrices, and know that the
correct definition in this case is
 
a11 a12 a13
det A = det a21 a22 a23 
a31 a32 a33
= a11 a22 a33 − a11 a23 a32 − a12 a21 a33 + a12 a23 a31 + a13 a21 a32 − a13 a22 a31 .

Often this definition for n = 3 is presented in the following scheme for forming the six possible
products of the aij ’s, one from each row and column:

a11 a12 a13

a21 a22 a23 the three indicated receive a plus sign, the three formed with the opposite slant a minus sign

a31 a32 a33

No such procedure is available for the correct definition for n > 3. However, meditation on
the formulas n = 2 and n = 3 reveals certain basic properties, which we can turn into axioms
for the general case. These properties are as follows:
MULTILINEAR: det A is a linear function of each column of A;
ALTERNATING: det A changes sign if two adjacent columns are interchanged;
NORMALIZATION: det I = 1.
Higher order derivatives 27

(In the first two of these properties we could replace “column” by “row,” but we choose not
to do that now. More on this later.) Incidentally, a function satisfying the first condition is
said to be multilinear, not linear. A linear function would satisfy det(A + B) = det A + det B,
definitely not true. In order to work effectively with these axioms it is convenient to rewrite
a matrix A in a notation which displays its columns. Namely, we denote the columns of A as
A1 , A2 , . . . , An , respectively. In other words,
 
a1j
 a2j 
 
Aj =  ..  .
 . 
anj

Then we present A in the form

A = (A1 A2 . . . An ).
Here is the major theorem of this section.

THEOREM. Let n ≥ 2. Then there exists one and only one real-valued function det defined
on all n × n matrices which satisfies
MULTILINEAR: det(A1 A2 . . . An ) is a linear function of each Aj ;
ALTERNATING: det(A1 · · · Aj Aj+1 · · · An ) = − det(A1 · · · Aj+1 Aj · · · An )
for each 1 ≤ j ≤ n − 1;
NORMALIZATION: det I = 1.

Rather than present a proof in our usual style, we are going to present a discussion of the
ramifications of these axioms. Not only will we prove the theorem, we shall also in the process
present useful techniques for evaluation of the determinant. Here’s an example.

PROBLEM 3–30. Prove that if det satisfies the alternating condition of the theorem,
then interchanging any two columns (not just adjacent ones) of A produces a matrix
whose determinant equals − det A.

PROBLEM 3–31. Prove that if two columns of A are equal, then det A = 0.
28 Chapter 3

PROBLEM 3–32. Conversely, prove that if det is known to be multilinear and satisfies
det A = 0 whenever two columns of A are equal, then A is alternating.
(HINT: expand
det(A1 · · · (Aj + Aj+1 )(Aj + Aj+1 )Aj+2 · · · An )
as a sum of four terms.)

Now we can almost immediately prove the uniqueness. It just takes a little patience with
the notation. Rewrite the column Aj in terms of the unit coordinate (column) vectors as
n
X
Aj = aij êi .
i=1

For each column we require a distinct summation index, so change i to ij to get

n
X
Aj = aij j êij .
ij =1

Now the linearity of det as a function of each column produces a huge multi-indexed sum:

det A = det(A1 A2 . . . An )
X
= ai1 1 . . . ain n det(êi1 . . . êin ).
i1 ,...,in

In this huge sum each ij runs from 1 to n, giving us nn terms. Now look at each individual
term in this sum. If the indices i1 , . . . , in are not all distinct, then at least two columns of the
matrix
(êi1 . . . êin )
are equal, so that Problem 3–31 implies the determinant is 0. Thus we may restrict the sum
to distinct i1 , . . . , in . Thus, the indices i1 , . . . , in are just the integers 1, . . . , n written in some
other order. Thus the matrix
(êi1 . . . êin )
can be transformed to the matrix
(ê1 . . . ên ) = I
by the procedure of repeatedly interchanging columns. Thus Problem 3–30 shows that

det(êi1 . . . êin ) = ±1,

Higher order derivatives 29

and the sign does not depend on det at all, but only on the interchanges needed. Thus

X
(∗) det A = ±ai1 1 . . . ain n ,
i1 ,...,in

and the signs depend on how i1 , . . . , in arise from 1, . . . , n. This finishes the proof of uniqueness
of det!
Now we turn to the harder proof of existence. A very logical thing to do would be to
use the formula we have just obtained to define det A; then we would “just” have to verify
the three required properties. However, it turns out that there is an alternate procedure that
gives a different formula and one that is very useful for computations. We now describe this
formula.
Temporarily denote by A0j the column Aj with first entry replaced by 0:
   
a1j 0
 0   a2j 
   
Aj =  ..  +  ..  = a1j ê1 + A0j .
 .   . 
0 anj

The multilinearity of det, supposing it exists, gives

det A = det(a11 ê1 + A01 . . . a1n ê1 + A0n )

= a sum of 2n terms,

where each term is a determinant arising from det A by replacing various columns of A by
a1j ê1 ’s. But if as many as two columns are so replaced the resulting determinant is 0, thanks
to its multilinearity and Problem 3–31. Thus (if it exists)
n
X
det A = a1j det(A01 . . . ê1 . . . A0n ) + det(A01 . . . A0n ).
j=1
↑
column j

The last of these determinants is 0, as the first row of the matrix consists of 0’s and (∗) gives
0. By interchanging the required j − 1 columns in the other matrices, we find
n
X
det A = (−1)j−1 a1j det(ê1 A01 . . . A0j−1 A0j+1 . . . A0n ).
j=1
30 Chapter 3

The matrix that appears in the j th summand is

 
1 0 ... 0 0 ... 0
0 a21 a2,j−1 a2,j+1 ... a2n 
 
 .. .. .. .. ..  .
. . . . . 
0 an1 an,j−1 an,j+1 . . . ann

It’s not hard to see that its determinant must be the same as that of the (n − 1) × (n − 1)
matrix that lies in rows and columns 2 through n. However, we don’t need to prove that at
this stage, as we can just use the formula to give the definition of det by induction on n.
Before doing that it is convenient to introduce a bit of notation. Given an n × n matrix
A and two indices i and j, we obtain a new matrix from A by deleting its ith row and j th
column. We denote this matrix by Aij , and call it the minor for the ith row and j th column.
It is of course an (n − 1) × (n − 1) matrix. Now here is the crucial result.

THEOREM. The uniqueness proof of the preceding theorem guarantees that only one deter-
minant function for n × n matrices can exist. It does exist, and can be defined by induction
on n from the starting case n = 1 (det a = a) and the inductive formula for n ≥ 2,

n
X
det A = (−1)i+j aij det Aij .
j=1

In this formula the row index i is fixed, and can be any of the integers 1, . . . , n (the sum is
independent of i).

The important formula in the theorem has a name: it is called expansion of det A by
the minors along the ith row.

PROOF. What we must do is for any 1 ≤ i ≤ n prove that det A as defined above satisfies
the three properties listed in the preceding theorem. In doing this we are of course allowed to
use the same properties for the determinants of (n − 1) × (n − 1) matrices. We consign the
proofs of the multilinearity and normalization to the problems following this discussion, as
they are relatively straightforward, and content ourselves with the alternating property about
switching two adjacent columns. Let us then assume that B is the matrix which comes from A
by switching columns k and k +1. Then if j 6= k or k +1 the minors Bij and Aij also differ from
one another by a column switch, so that the inductive hypothesis implies det Bij = − det Aij .
Higher order derivatives 31

Thus only j = k and k + 1 matter:

det B + det A = (−1)i+k bik det Bik + (−1)i+k+1 bi,k+1 det Bi,k+1
+ (−1)i+k aik det Aik + (−1)i+k+1 ai,k+1 det Ai,k+1
= (−1)i+k ai,k+1 [det Bik − det Ai,k+1 ]
+ (−1)i+k aik [− det Bi,k+1 + det Aik ] .

Aha! We have Bik = Ai,k+1 and Bi,k+1 = Aik . Thus det B + det A = 0.
QED

PROBLEM 3–33. Prove by induction that det A as defined inductively is a linear

function of each column of A.

PROBLEM 3–34. Prove by induction that det A as defined inductively satisfies

det I = 1.

PROBLEM 3–35. Let A be “upper triangular” in the sense that aij = 0 for all i > j.
Prove that det A = a11 a22 . . . ann . That is,
 
a11 a12 a13 . . . a1n
 0 a22 a23 . . . a2n 
 
 0 a33 . . . a3n 
det  0  = a11 a22 . . . ann .
 .. .. .. .. 
 . . . . 
0 0 0 ann

Do the same for lower triangular.

One more property is needed in order to be able to compute determinants efficiently. This
property states that all the properties involving column vectors and expansion along a row
are valid also for the row vectors of the matrix and expansion along columns. Because of the
uniqueness of the determinant, it will suffice to prove the following result.

THEOREM. The determinant as defined above satisfies

ROW MULTILINEAR: det A is a linear function of each row of A.
ROW ALTERNATING: det A changes sign if two rows of A are interchanged.
32 Chapter 3

PROOF. The row multilinearity is an immediate consequence of the formula (∗) or of the
formula for expansion of det A by the minors along the ith row. The latter formula actually
displays det A as a linear function of the row vector (ai1 , ai2 , . . . , ain ).
The row alternating property follows easily from the result of Problem 3–32, applied to
rows instead of columns. Thus, assume that two rows of A are equal. If n = 2, det A = 0
because of our explicit formula in this case. If n > 2, then expand det A by minors along a
different row. The corresponding determinants of the minors Aij are all zero (by induction on
n) because each Aij has two rows equal.
QED
The proofs of the following two corollaries are now immediate:

COROLLARY. det At = det A.

Here At is the transpose of A, as defined in Problem 2–83.

COROLLARY. Expansion by minors along the j th column:

n
X
det A = (−1)i+j aij det Aij .
i=1

If we had to work directly from the definition of determinants, calculating them would be
utterly tedious for n ≥ 4. We would essentially need to add together n! terms, each a product
of n matrix entries. Fortunately, the properties lend themselves to very efficient calculations.
The goal is frequently to convert A to a matrix for which a row or column is mostly 0.
The major tool for doing this is the property that if a scalar multiple of a column (or
row) is added to a different column (or row), then the determinant is unchanged.
Such procedures are called elementary column (or row) operations.
Higher order derivatives 33

For instance,

 
3 1 0 6
1 −1 −1 5
det 
0 2

2 1
2 0 −1 3
 
0 4 3 −9
1 −1 −1 5 
= det 0 2
 (added multiples of row 2 to rows 1 and 4)
2 1
0 2 1 −7
 
4 3 −9
= − det 2 2 1  (expanded by minors along column 1)
2 1 −7
 
22 21 −9
= − det  0 0 1  (added multiples of column 3 to columns 1 and 2)
16 15 −7
µ ¶
22 21
= det (expanded by minors along row 2)
16 15
µ ¶
11 7
= 2 · 3 det (used linearity in columns 1 and 2)
8 5
= 6(55 − 56)
= −6.

PROBLEM 3–36. Find and prove an expression for the n × n determinant

 
0 0 ... 0 1
0 0 . . . 1 0
 
 .. 
det  . .
 
0 1 . . . 0 0
1 0 ... 0 0
34 Chapter 3

PROBLEM 3–37. Given fixed numbers a, b, c, let xn stand for the n × n “tridiagonal”
determinant  
a b
c a b 
 
xn = det  c a b 

.
 . . .
c a
Find a formula for xn in terms of xn−1 and xn−2 . (Blank spaces are all 0.)

PROBLEM 3–38. Express the n × n tridiagonal determinant

 
1 1
−1 1 1 
 
det 
 −1 1 1 

 . . .
−1 1

in Fibonacci terms. The Fibonacci numbers may be defined by F0 = 0, F1 = 1, Fn =

Fn−1 + Fn−2 .

PROBLEM 3–39. Find and prove an expression for the n × n determinant

 
0 1 1 ... 1
1 0 1 . . . 1
 
 
det 1 1 0 . . . 1 .
 .. .. .. .. 
. . . .
1 1 1 ... 0
Higher order derivatives 35

PROBLEM 3–40. Here is a famous determinant, called the Vandermonde determi-

nant: prove that  
1 a1 a21 . . . an−1
1
1 a2 a2 . . . an−1  Y
 2 2 
det  .. .. .. ..  = (aj − ai ).
. . . .  i<j
1 an a2n . . . an−1
n

PROBLEM 3–41. Calculate

 
λ + a1 b1 a1 b2 a1 b3 . . . a 1 bn
 a2 b1 λ + a2 b2 a2 b3 . . . a 2 bn 
 
det  .. .. .. .. .
 . . . . 
an b1 a n b2 an b3 . . . λ + an bn

(Answer: λn−1 (λ + a • b))

PROBLEM 3–42. When applied to n × n matrices, determinant can be regarded as

2
a function from Rn to R. What is
∂ det A
?
∂aij

G. Invertible matrices and Cramer’s rule

Before entering into the main topic of this section, we give one more property of determi-
nant.

THEOREM. det(AB) = det A det B.

PROOF. Here of course A and B are square matrices of the same size, say n × n. The i − j
entry of the product AB is
Xn
aik bkj .
k=1
36 Chapter 3

Thus the j th column of AB is the column vector

 
a1k bkj
X n   n
 a2k bkj  X
 ..  = bkj Ak ,
k=1
 .  k=1
ank bkj

where, as usual, Ak is our notation for the k th column of A. We can now simply follow the
procedure of p. 3–28. In the above formula rename the summation index k = ij , so we have
by the multilinearity of det,
X
det AB = bi1 1 . . . bin n det(Ai1 . . . Ain ).
i1 ,...,in

Here the ij ’s run from 1 to n. If they are not distinct, then the corresponding determinant is
0. Thus the ij ’s are distinct in all nonzero terms in the sum, and by column interchanges we
have
det(Ai1 . . . Ain ) = ± det(A1 . . . An )
= ± det A.

In fact, the sign is the same as

det(êi1 . . . êin ).
Therefore
X
det AB = det A bi1 1 . . . bin n det(êi1 . . . êin )
i1 ,...,in

= det A det B.

QED
REMARKS. This result is a rather astounding bit of multilinear algebra. To convince
yourself of that, just write down what it says in the case n = 2. Amazingly, the identical
technique leads to a vast generalization, the Cauchy-Binet determinant theorem, which we
shall present in Section 11D. It is also amazing that there is an entirely different approach to
proving this result, as we shall discover in our study of integration in Chapter 10.

DEFINITION. A square matrix A is invertible if there exists a matrix B such that AB =

BA = I. The matrix B is called the inverse of A, and is written B = A−1 .
Higher order derivatives 37

PROBLEM 3–43. For the above definition of inverse to make sense we need to know
that the matrix B is unique. In fact, prove more: if A has a right inverse B (AB = I)
and a left inverse C (CA = I), then B = C.
(HINT: CAB.)

PROBLEM 3–44. Prove that if A and B are invertible, then AB is invertible and

(AB)−1 = B −1 A−1 .

PROBLEM 3–45. Prove that in general

(AB)t = B t At .

PROBLEM 3–46. Prove that A is invertible ⇐⇒ At is invertible. Also prove that

(A−1 )t = (At )−1 .

Now we are almost ready to state the major result of this section. There’s one more concept
we need, and that is the linear independence of points in Rn .

DEFINITION. The vectors x(1) , x(2) , . . . , x(k) in Rn are linearly dependent if there exist
real numbers c1 , c2 , . . . , ck , not all 0, such that

c1 x(1) + c2 x(2) + · · · + ck x(k) = 0.

In other words, the given vectors satisfy some nontrivial homogeneous linear relationship. The
vectors are said to be linearly independent if they are not linearly dependent.

PROBLEM 3–47. Prove that x(1) , . . . , x(k) are linearly dependent ⇐⇒ one of them is
equal to a linear combination of the others.
38 Chapter 3

A good portion of elementary linear algebra is dedicated to proving that if x(1) , . . . , x(n)
are linearly independent vectors in Rn — notice that it’s the same n — then every vector in
Rn can be expressed as a unique linear combination of the given vectors. That is, for any
x ∈ Rn there exist unique scalars c1 , . . . , cn , such that

x = c1 x(1) + · · · + cn x(n) .

PROBLEM 3–48. Prove that the scalars c1 , . . . , cn in the above equation are indeed
unique.

PROBLEM 3–49. Prove that any n + 1 vectors in Rn are linearly dependent.

PROBLEM 3–50. Prove that the column vectors of an upper triangular matrix (aij )
are linearly independent ⇐⇒ the diagonal entries aii are all nonzero.

DEFINITION. If x(1) , . . . , x(n) are linearly independent vectors in Rn , they are said to be
a basis for Rn .
Now for the result.

THEOREM. Let A be an n × n matrix. Then the following conditions are equivalent:

(1) det A 6= 0.

(2) A is invertible.

(3) A has a right inverse.

(4) A has a left inverse.

(5) The columns of A are linearly independent.

(6) The rows of A are linearly independent.

PROOF. We first prove that (3)=⇒(1)=⇒(5)=⇒(3). First, (3)=⇒(1) because the matrix
equation AB = I implies det A det B = det I = 1, so det A 6= 0. To see that (1)=⇒(5),
suppose the columns of A linearly dependent. Then one of them is a linear combination of the
Higher order derivatives 39

others, as we know from Problem 3–47. Inserting the resulting equation into the corresponding
column of A and using the linearity of det A with respect to that column leads to det A = 0.
Now suppose (5) holds. Then every vector can be expressed as a linear combination of the
columns A1 , . . . , An of A. In particular, each coordinate vector can be so expressed:
n
X
êj = bkj Ak
k=1

for certain unique scalars bij , . . . , bnj . Now we define the matrix B = (bij ), and we realize the
equations for êj simply assert that
I = AB.
Thus (3) is valid.
So (1), (3), (5) are equivalent. Likewise, (1), (4), (6) are equivalent. We conclude that (1),
(3), (4), (5), (6) are equivalent. Finally, if they all hold, then (3) and (4) imply A is invertible,
thanks to Problem 3–43. The converse assertions that (2) =⇒ (3) and (4) are trivial.
QED

We also remark that in case A is invertible, the two matrices A and A−1 commute. Though
this is rather obvious, still it is quite a wonderful fact since we do not generally expect two
matrices to commute. Whenever it happens that AB = BA, that is really worth our attention.
This is certainly a terrific theorem! It displays in dramatic fashion the importance of the
determinant. However, it does not show how we might go about computing A−1 . There is a
theorem that does this, and it goes by the name of Cramer’s rule.
Assume that det A 6= 0 and that B = A−1 . Then we examine the equation AB = I. As
we have seen, it can be written as the set of vector equations
Xn
bkj Ak = êj , 1 ≤ j ≤ n.
k=1
th
Now replace the i column of A by êj and calculate the determinant of the resulting matrix.
First, expansion by minors along the ith column produces
det(A1 . . . Ai−1 êj Ai+1 . . . An ) = (−1)i+j det Aji .
On the other hand, the formula for êj yields
n
X
det(A1 . . . Ai−1 êj Ai+1 . . . An ) = bkj det(A1 . . . Ai−1 Ak Ai+1 . . . An )
k=1
= bij det(A1 . . . Ai−1 Ai Ai+1 . . . An )
= bij det A.
40 Chapter 3

We conclude that
det Aji
bij = (−1)i+j .
det A
Now that we have proved it, we record the statement:

CRAMER’S RULE. Assume det A 6= 0. Then A−1 is given by the formula for its i − j
entry,
det Aji
(−1)i+j .
det A

Elegant as it is, this formula is rarely used for calculating inverses, on account of the effort
required to evaluate determinants. However, the 2 × 2 case is especially appealing and easy
to remember:
µ ¶−1 µ ¶
a b 1 d −b
= .
c d ad − bc −c a

COROLLARY. Assume det A 6= 0 and b ∈ Rn is a column vector. Then the unique column
vector x ∈ Rn which satisfies Ax = b is given by x = A−1 b, and the ith entry of x equals

det(A1 . . . Ai−1 bAi+1 . . . An )

xi = .
det A

PROBLEM 3–51. Prove the corollary.

The equation Ax = b, with x as the “unknown,” is of course the general system of n linear
equations in n unknowns, for if we display the coordinates we have

a11 x1
 + ···+ a1n xn = b1 ,
..
 .

an1 x1 + · · · + ann xn = bn .

We can conclude from our results that four conditions concerning this system of equations are
equivalent:
Higher order derivatives 41

PROBLEM 3–52. Given an n × n matrix A, prove that the following four conditions
are equivalent:

(1) det A 6= 0.

(2) If x ∈ Rn is a column vector such that Ax = 0, then x = 0.

(3) For every b ∈ Rn , there exists x ∈ Rn such that Ax = b.

(4) For every b ∈ Rn , there exists one and only one x ∈ Rn such that Ax = b.

PROBLEM 3–53. The classical adjoint of a square matrix A is denoted adjA and is
defined by prescribing its i − j entry to be

(−1)i+j det Aji .

As we have shown if det A 6= 0, then A−1 = adjA/ det A. Prove that in general

AadjA = (det A)I.

(HINT: write down the definition of the i − j entry of the product on the left side and
use expansion by minors along the j th row.

PROBLEM 3–54. Prove that A and adjA commute.

PROBLEM 3–55. Prove that

adj(adjA) = (det A)n−2 A.

PROBLEM 3–56. Prove that the classical adjoint of an upper triangular matrix is
also upper triangular.
42 Chapter 3

PROBLEM 3–57. Write out explicitly

 −1
a11 a12 a13
 0 a22 a23  .
0 0 a33

H. Recapitulation
As we have become involved in a great deal of mathematics in our analysis of critical
points, it seems good to me that we pause for a look at the forest.
f
1. We are considering a function Rn −
→ R of class C 2 .

2. We assume x0 is a critical point for f : ∇f (x0 ) = 0.

3. We consider the Taylor expansion of f (x0 + y):

n
1X
f (x0 + y) = f (x0 ) + Di Dj f (x0 )yi yj + R.
2 i,j=1

4. We define the Hessian matrix

H = (Di Dj f (x0 )) .

5. We say the critical point is nondegenerate if det H 6= 0.

6. Our goal is now to show that in the nondegenerate case, the Hessian matrix reveals the
local nature of the critical point.

7. We have finished this program in dimension 2, where we learned that if the critical point
is degenerate (det H = 0), we cannot detect the local behavior through H.

What remains is the analysis of symmetric matrices like H and the associated behavior of the
quadratic functions they generate. As this in itself is a fascinating and important topic, we
devote the next chapter to this study.
Higher order derivatives 43

PROBLEM 3–58. Although this problem is concerned with Rn , you can analyze it
directly without requiring the result of Chapter 4. This is a generalization of Problem 3–
f
24 (see also Problem 2–60). Let a1 > a2 > · · · > an > 0 be constants, and define Rn →− R
by
2
f (x) = (a1 x21 + · · · + an x2n )e−kxk .

a. Find all the critical points of f and determine the local nature of each one.

b. Clearly, f attains a strict global minimum at x = 0. Show also that f attains a

global maximum; is it strict?

I. A little matrix calculus

It is quite interesting to investigate the interplay between the algebra of matrices and
calculus. Specifically we wish to analyze the operation of the inverse of a matrix.
2
First we observe that the space of all n × n real matrices may be viewed as Rn , a fact we
observed on p. 2–48. Since the determinant of A is a polynomial expression in the entries of
2
A we conclude that det A is a C ∞ function on Rn . The set of invertible matrices is the same
as the set of matrices with nonzero determinant and is therefore an open subset of the set of
all n × n matrices.

DEFINITION. The general linear group of n×n matrices is the set of all n×n real invertible
matrices. It is commonly denoted
GL(n).
This is to be regarded as an open subset of the space of all n × n real matrices.
There is of course a very interesting function defined on GL(n), namely the function which
maps a matrix A to its inverse A−1 . Let’s name this function inv. Thus

inv(A) = A−1 .

Cramer’s rule displays A−1 as a matrix whose entries are polynomial functions of A, divided
by det A. Thus A−1 is a C ∞ function of A. That is,
inv
GL(n) → GL(n)

is a C ∞ function.
44 Chapter 3

PROBLEM 3–59. If the n × n matrix B is sufficiently small, I + B ∈ GL(n). Prove

that
(I + B)−1 = I − (I + B)−1 B.
Then prove that
(I + B)−1 = I − B + (I + B)−1 B 2 .

PROBLEM 3–60. Use the preceding problem to show that the differential of inv at I
is the linear mapping which sends B to −B. In terms of directional derivatives,

Dinv (I; B) = −B.

PROBLEM 3–61. Let A ∈ GL(n). Use the relation

A + B = A(I + A−1 B)

to prove that
Dinv (A; B) = −A−1 BA−1 .
In other words, ¯
d ¯
−1 ¯
(A + tB) ¯ = −A−1 BA−1 .
dt t=0