You are on page 1of 4

We begin with the classical one-dimensional change of variables

Z g(b) Z b
f (x) dx = f (g(u)) g ′ (u) du.
g(a) a

This holds if g ′ is strictly positive on [a, b] and and also holds when g ′ is strictly
negative on [a, b]. In the latter case it can written as
Z g(a) Z b
f (x) dx = f (g(u)) (−g ′ (u)) du.
g(b) a

We can express both of these inequalities in a single form by writing


Z Z b
f (x) dx = f (g(u)) |g ′ (u)| du.
g[a,b] a

This covers the cases of both g being strictly increasing and strictly decreasing on
[a, b] and also provides a stepping point to extend matters to higher dimensions.
We would like to prove something analogous on Rn . The formula is the following:
Z Z
(1) f (x) dx = f (g(u)) |g ′ (u)| du
g[Ω] Ω

where g is 1−1 continuously differentiable mapping from an open rectangle U contain-


ing a compact subset Ω such that det g ′ never vanishes on U . We are also assuming
here that f is integrable over g[Ω] and (f ◦ g) | det g ′ | is integrable over Ω.

We first verify (1) when g is a linear mapping, i.e., it has the form g(x) = Bx for
some n × n matrix B, f = 1, Ω is a rectangle and n = 2.
Then if the rectangle R is determined by the vectors (t1 , 0) and (0, t 2 ), 
where
t1
t1 , t2 > 0, then the parallelogram B[R] is determined by the vectors B and
0
 
0
B . These vectors are
t2
             
t1 b11 b12 t1 b11 0 b11 b12 0 b
B = = t1 , B = = t2 12
0 b21 b22 0 b21 t2 b21 b22 t2 b22
and the area of the parallelogram formed by these vectors is
 
i j k
det t1 b11 t1 b21 0 = t1 t2 | det B| = | det B| Area(R).
t2 b12 t2 b22 0
This argument gives that
(2) V ol(B[R]) = | det B|V ol(R)
for any rectangle R with sides parallel to the axes and any matrix B.
We have the following lemma:
1
2

Lemma 1. Suppose that Φ is a 1 − 1 and onto continuously differentiable mapping


from an open subset U of Rn to another open subset of Rn such that det Φ′ (x) ̸= 0
for all x ∈ U . Then for every ε > 0 (smaller than | det Φ′ (c)|/2const) there is δ0 > 0
such that for any closed cube Q contained in U with side length δ < δ0 centered at c
and with sides parallel to the axes we have
n n
1 − const ε V ol(Q)| det Φ′ (c)| ≤ V ol(Φ[Q]) ≤ V ol(Q)| det Φ′ (c)| 1 + const ε .
Proof. By the definition of differentiability, given ε > 0, there is a δ > 0 such that if
Q has center c and side length δ we have
x ∈ Q =⇒ Φ(x) − Φ(c) − Φ′ (c)(x − c) < ∥x − c∥ε.
This fact implies
√ 
′ nε 
Φ[Q] − Φ(c) ⫅ Range[Φ (c)[Q − c]] + δB 0,
2
and √

 n ε
Range[Φ (c)[Q − c]] ⫅ Φ[Q] − Φ(c) + δB 0, .
2
It follows from this and (2) that
V ol(Φ[Q]) ≤ | det Φ′ (c)|V ol(Q) + (1 + constε)n V ol(Q)
and
| det Φ′ (c)|V ol(Q) ≤ V ol(Φ[Q])(1 + constε)n V ol(Q)
Combining these facts
n n
1 − constε V ol(Q)| det Φ′ (c)| ≤ V ol(Φ[Q]) ≤ V ol(Q)| det Φ′ (c)| 1 + constε .

PROOF OF (1)
Let ε > 0 be given. Then there is δ1 > 0 such that for any grid R (we by refinement
we may assume that it contains cubes Qi of equal side length) such that ∪i Qi covers
Ω and if the diameter of the cubes Qi are less than δ1 , then
Z Z

(U ) |f ◦ g(u)| | det g (u)|du ≤ |f ◦ g(u)| | det g ′ (u)| du + ε
Ω Ω

and Z Z

(L) |f ◦ g(u)| | det g (u)|du ≥ |f ◦ g(u)| | det g ′ (u)| du − ε
Ω Ω
By uniform continuity there is a δ2 > 0 such that if

x, y ∈ Qi =⇒ | det g ′ (x)| − | det g ′ (y)| < ε.


3

Choose a grid of cubes {Qi }i such that the diameter of each cube is at most δ.
Also, restricting to the cubes Qi such that Qi ∩ Ω is not empty, we have that that
∪i g[Qi ] covers g[Ω]. Additionally,
Z X Z  [  

(3) f (x) dx− f (x) dx ≤ sup |f | V ol g[Qi ] \g[Ω]
g[Ω] g[Qi ] g[Ω]
i:Qi ∩Ω̸=∅ i:Qi ∩Ω̸=∅

and choose the grid {Qi } such that if each cube in it has diameter smaller that δ3 ,
then
 [  

(4) sup |f | V ol g[Qi ] \ g[Ω] < ε.
g[Ω]
i:Qi ∩Ω̸=∅

By Lemma 1 there is also a δ0 associated with this ε > 0. Now set


δ = min(δ0 , δ1 , δ2 , δ3 )
and pick a grid of cubes covering Ω with diameter at most δ.
By (3) and (4) we have
Z XZ
f (x) dx − f (x) dx < ε
g[Ω] i g[Qi ]

and we also trivially have


X XZ X
( inf f )V ol(g[Qi ]) < f (x) dx < (sup f )V ol(g[Qi ]).
g[Qi ] g[Qi ] g[Qi ]
i i i

We then obtain
X Z X
( inf f )V ol(g[Qi ]) − ε < f (x) dx < (sup f )V ol(g[Qi ]) + ε
g[Qi ] g[Ω] g[Qi ]
i i

But Lemma 1 gives


X X
(sup f )V ol(g[Qi ]) + ε ≤ (sup f )| det g ′ (ci )|V ol(Qi )(1 + cε)n + ε
i g[Qi ] i g[Qi ]

which is bounded by
X Z
sup |f ◦ g|| det g |V ol(Qi )(1 + cε) + ε ≤ (U ) |f ◦ g(u)|| det g ′ (u)|du(1 + cε)n + ε.
′ n
Qi Ω
i
Z
≤ |f ◦ g(u)|| det g ′ (u)| du (1 + cε)n + 2ε.

Now for the lower inequality we write


X X
( inf f )V ol(g[Qi ]) − ε ≥ ( inf f )| det g ′ (ci )| V ol(Qi )(1 − cε)n − ε
g[Qi ] g[Qi ]
i i
X X
≥ ( inf f )| det g ′ (xi )|V ol(Qi )(1−cεn )− ( inf f ) | det g ′ (xi )|−| det g ′ (ci )| V ol(Qi )(1−cε)n −ε
g[Qi ] g[Qi ]
i i
4

where we picked xi such that | det g ′ (xi )| = inf Qi |g ′ | and we set M = sup |f |. We
obtain X
( inf f ) | det g ′ (ci )| V ol(g[Qi ])(1 − cε)n − ε
g[Qi ]
i
X
≥ ( inf f ) | det g ′ (xi )| V ol(Qi )(1 − cε)n − M εV ol(R)(1 − cε)n − ε
g[Qi ]
i
Z
≥ (L) |f ◦ g(u)| | det g ′ (u)|du(1 − cε)n − M εV ol(R)(1 − cε)n − ε
Z Ω
≥ |f ◦ g(u)| | det g ′ (u)|du(1 − cε)n − M εV ol(R)(1 − cε)n − 2ε.

We have therefore proved the upper inequality
Z Z
f (x) dx ≤ |f ◦ g(u)| | det g ′ (u)|du(1 + cε)n + 2ε
g[Ω] Ω

and the lower inequality


Z Z
f (x) dx ≥ |f ◦ g(u)| | det g ′ (u)|du(1 − cε)n − M εV ol(R)(1 − cε)n − 2ε.
g[Ω] Ω

Letting ε → 0 we obtain the validity of (1).


It follows from the proof that we can eliminate the hypothesis that f is integrable
over g[Ω] as this would follow from the inequalities proved and the assumption that
(f ◦ g) | det g ′ | is integrable over Ω.

You might also like