In measure theory, the factorization lemma allows us to express a function f with another function T if f is measurable with respect to T. An application of this is regression analysis.
Let
T
:
Ω
→
Ω
′
be a function of a set
Ω
in a measure space
(
Ω
′
,
A
′
)
and let
f
:
Ω
→
R
¯
be a scalar function on
Ω
. Then
f
is measurable with respect to the σ-algebra
σ
(
T
)
=
T
−
1
(
A
′
)
generated by
T
in
Ω
if and only if there exists a measurable function
g
:
(
Ω
′
,
A
′
)
→
(
R
¯
,
B
(
R
¯
)
)
such that
f
=
g
∘
T
, where
B
(
R
¯
)
denotes the Borel set of the real numbers. If
f
only takes finite values, then
g
also only takes finite values.
First, if
f
=
g
∘
T
, then f is
σ
(
T
)
−
B
(
R
¯
)
measurable because it is the composition of a
σ
(
T
)
−
A
′
and of a
A
′
−
B
(
R
¯
)
measurable function. The proof of the converse falls into four parts: (1)f is a step function, (2)f is a positive function, (3) f is any scalar function, (4) f only takes finite values.
Suppose
f
=
∑
i
=
1
n
α
i
1
A
i
is a step function, i.e.
n
∈
N
∗
,
∀
i
∈
[
[
1
,
n
]
]
,
A
i
∈
σ
(
T
)
and
α
i
∈
R
+
. As T is a measurable function, for all i, there exists
A
i
′
∈
A
′
such that
A
i
=
T
−
1
(
A
i
′
)
.
g
=
∑
i
=
1
n
α
i
1
A
i
′
fulfils the requirements.
If f takes only positive values, it is the limit, for pointwise convergence, of a increasing sequence
(
u
n
)
n
∈
N
of step functions. For each of these, by (1), there exists
g
n
such that
u
n
=
g
n
∘
T
. The function
lim
n
→
+
∞
g
n
, which exists on the image of T for pointwise convergence because
(
u
n
)
n
∈
N
is monotonic, fulfils the requirements.
We can decompose f in a positive part
f
+
and a negative part
f
−
. We can then find
g
0
+
and
g
0
−
such that
f
+
=
g
0
+
∘
T
and
f
−
=
g
0
−
∘
T
. The problem is that the difference
g
:=
g
+
−
g
−
is not defined on the set
U
=
{
x
:
g
0
+
(
x
)
=
+
∞
}
∩
{
x
:
g
0
−
(
x
)
=
+
∞
}
. Fortunately,
T
(
Ω
)
∩
U
=
∅
because
g
0
+
(
T
(
ω
)
)
=
f
+
(
ω
)
=
+
∞
always implies
g
0
−
(
T
(
ω
)
)
=
f
−
(
ω
)
=
0
We define
g
+
=
1
Ω
′
∖
U
g
0
+
and
g
−
=
1
Ω
′
∖
U
g
0
−
.
g
=
g
+
−
g
−
fulfils the requirements.
If f takes finite values only, we will show that g also only takes finite values. Let
U
′
=
{
ω
:
|
g
(
ω
)
|
=
+
∞
}
. Then
g
0
=
1
Ω
′
∖
U
′
g
fulfils the requirements because
U
′
∩
T
(
Ω
)
=
∅
.
If the function
f
is not scalar, but takes values in a different measurable space, such as
R
with its trivial σ-algebra (the empty set, and the whole real line) instead of
B
(
R
)
, then the lemma becomes false (as the restrictions on
f
are much weaker).