The integral of a nonnegative function is defined by approximation from below. For a signed function, a finite integral requires both positive- and negative-part integrals to be finite. One may define an extended signed integral when exactly one is infinite, but not when both are infinite. We will prove the algebra that makes this construction linear, rather than assuming that cancellation behaves as it does for finite sums.
Work on an arbitrary measure space , without assuming finite measure, -finiteness or completeness. We use nonnegative simple-function integration and the monotone convergence theorem. All functions named below are measurable; and .
Adding nonnegative functions
Lemma 1.6.1
(Additivity for nonnegative functions)
For measurable and ,
with the conventions on both sides. Any integral may be infinite.
Proof
(Simple approximations and monotone convergence)
Choose nonnegative simple functions and pointwise. Refining their finite measurable level partitions into intersections proves ; this is valid even when a level set has infinite measure, using . Since , monotone convergence on each side yields (1). For a zero coefficient the corresponding sequence and limit are identically zero.
In particular, if and , then is finite and . This follows directly from the simple-function supremum definition and requires no signed subtraction.
Adding signed functions
Definition 1.6.2
(Integrable function and its integral)
A real measurable function is integrable, written , when . Then both and are finite, and
Theorem 1.6.3
(Linearity of the integral on )
If and , then and
Here may be read as integrable measurable representatives; the same statements hold for their almost-everywhere equivalence classes. An extended-real integrable representative is finite outside a measurable null set and can be set to zero on that set before forming sums (see the null-set argument below).
Proof
(Move the negative parts to the other side)
The pointwise inequality , monotonicity and (1) give . The crucial identity involves only nonnegative functions:
Indeed (4) is the rearrangement of the finite pointwise identity . Every term in (4) is integrable. Repeated application of (1) gives finite-sum additivity; integrate (4) and rearrange finite real numbers to obtain .
For , ; for , and . Equation (1) gives in either case, and gives integrability. Combine the two results for . Finally, (2) and give , the last inequality in (3).
Proof X-ray
Why add the negative parts first
The tempting equation is exactly what must be proved, and it cannot be obtained by splitting into minus , because is in general not . The key move is identity (4): moving every negative part to the side where it is added turns a signed equation into an equality of nonnegative functions, and for those the lemma already gives additivity with no hypothesis beyond measurability. Integrability is used only at the end, to know that all six integrals are finite real numbers, so that the integrated form of (4) may be rearranged by ordinary subtraction. The scalar case is a case split on the sign of , with no cancellation at all.
Common mistake
(Splitting an integral into two infinite pieces)
Linearity requires . On take . Then and , but the expression is and has no value. A rule applied without checking integrability can produce any answer.
Worked example 1.6.4
(A negative scalar swaps the two parts)
On with Lebesgue measure let , so , and . For the parts exchange roles: and . Hence , as linearity predicts. The edge case gives and , with no use of because both parts of are finite.
Proposition 1.6.5
(Integrals do not see null sets)
If two measurable integrable representatives agree almost everywhere, their integrals agree. This holds even if the measure space is not complete.
Proof
(Restrict to the complement of the difference set)
The set on which the two functions differ is measurable, because both are measurable, and it is null. A nonnegative measurable function supported on has integral zero by the simple-function definition, so by (1) each positive- and negative-part integral is unchanged by restricting to . Thus (3) is independent of the chosen measurable representative.
Completeness would be needed to conclude that an arbitrary modification on an arbitrary subset of a null set remains measurable.
Integrals over sets form a signed measure
Theorem 1.6.6
(The indefinite integral is a signed measure)
For define for . Then is a finite signed measure with , and for disjoint with union the series converges absolutely to .
Proof
(Control the tail by a convergent nonnegative series)
Clearly , and (3) applied to gives . For disjoint , finite additivity follows from (3). If , nonnegative finite additivity and monotone convergence give . Consequently
The tail integral equals , the remainder of a convergent nonnegative series. Hence ; in fact by the same nonnegative additivity and monotone convergence.
Worked example 1.6.7
(A signed measure with alternating pieces)
On with Lebesgue measure let and . Since ,
so and . Using for with ,
The series of converges absolutely, with , so any rearrangement gives . Splitting into parts, (odd ) and (even ): their difference is and their sum is .
Cancellation is not a definition
Counterexample mode
Symmetric cancellation does not define an integral
Claim: if the symmetric integrals converge as , their limit is the Lebesgue integral of . On , let . Each symmetric integral is , but and . Formula (2) is undefined, not zero: a symmetric limiting convention cannot turn into a Lebesgue integral. The repair is the integrability hypothesis , under which dominated convergence does give .
By contrast, cancellation is harmless when both parts are finite.
Worked example 1.6.8
(Cancellation between two finite parts)
On let . Monotone convergence on gives , so linearity yields . Yet is positive on and negative on . The positive part has integral , and the negative part has , so . Here the cancellation is between two finite quantities.
Practice with solutions
Quick check
(If , why does have finite integral?)
Solution · Answer
and (1) give the finite upper bound .
Quick check
(Let be disjoint and . Can depend on the order of the sets?)
Solution · Answer
No: the series is absolutely convergent as shown above, so every rearrangement has the same sum.
Quick check
(On , is in although it is finite everywhere?)
Solution · Answer
No: by monotone convergence on .
Quick check
(If a measurable function is changed on a subset of a null set in a noncomplete measure space, is the new function automatically measurable?)
Solution · Answer
Not necessarily; the subset itself need not be measurable. If the new function is measurable, however, it is integrable and its integral agrees with the original whenever the original is integrable.