An iterated integral performs two operations: it integrates one variable while holding the other fixed, then integrates the resulting function. Before comparing it with a product integral, we must know that the sections are measurable and that the first integral produces a measurable function. For signed functions, we must also prevent cancellation from disguising an undefined difference of infinities.
Tonelli's theorem answers the nonnegative question even when the answer is infinite. Fubini's theorem answers the signed or complex question under absolute integrability. Their proofs share a path from measurable rectangles to indicators, then to simple functions and monotone limits. The signed theorem is obtained only after that nonnegative foundation has been established.
Motivation
On a rectangle, multiplying the measures of the sides is straightforward. On a general measurable subset of the product, a vertical section can change shape with the first variable. A section integral can even be infinite at an exceptional point while the product integral is finite. These are measure-theoretic issues, not difficulties removed by drawing a smooth region.
A second danger is more subtle. A signed function can have finite integrals in both orders, with every individual section integrable, yet give two different answers. In such a case, canceling positive and negative contributions within each section has happened before checking their total masses. The joint Lebesgue integral may not exist at all. An absolute-value check tells us whether the proposed change of order is legitimate.
Definitions
Let and be nonnegative measure spaces. Both measures are assumed -finite: each space has a countable measurable cover by finite-measure sets. Finite unions turn such a cover into increasing sets , with finite measures.
Definition 1.4.1
(Product -algebra and product measure)
The product -algebra is generated by rectangles with , . We assume the product-measure construction: there is a unique -finite measure on that -algebra with , using .
We do not define the product measure by the iterated-integral identity that we are about to prove. Its construction and rectangle property are inputs. The main statements use the product -algebra itself; a completed-product formulation needs additional representative and exceptional-section arguments.
Definition 1.4.2
(Sections and iterated integrals)
For a product set , write and . For a function, write and . An iterated integral such as requires both the inner integrals and measurability of the outer function.
Nonnegative integrals take values in . For real functions, and . Integrability means , so the two finite nonnegative integrals may be subtracted. Complex integrals are defined through real and imaginary parts, again requiring absolute integrability.
Theorem/proposition
Theorem 1.4.3
(Tonelli's theorem for -finite products)
Let both marginal measure spaces be -finite as above. If is product-measurable, then every section is measurable, both section-integral functions are measurable with values in , and
All three values may be infinite. No prior finiteness assumption is needed.
Theorem 1.4.4
(Fubini's theorem for absolutely integrable functions)
On the -finite product spaces above, let be real- or complex-valued and product-measurable, with . For -almost every , is -integrable, and for -almost every , is -integrable. Define either section integral to be zero on its exceptional set. The resulting measurable functions are integrable, and the three integrals in Tonelli's displayed identity are equal and finite. In particular,
where the inner integral uses that exceptional-set convention.
Corollary 1.4.5
(Null sections and an absolute-integrability test)
For the same -finite marginal spaces, a product-measurable set is -null exactly when for -almost every , and equivalently when for -almost every . Also, a product-measurable is integrable exactly when the iterated integral of in either chosen order is finite. Both orders then give the same finite absolute integral.
The last test uses Tonelli on before using Fubini on . Computing an iterated signed integral and finding a finite number is a different test and does not suffice.
Proof sketch or proof idea
Sections are measurable before they are integrated
Proof
(Sections of product-measurable sets)
Fix . The sets for which form a -algebra: sections preserve complements and countable unions. It contains all measurable rectangles, hence contains . The corresponding statement holds for horizontal sections. Applying this to threshold sets proves measurability of nonnegative sections; real and imaginary parts give the other cases.
The measurability of is a separate assertion. Simply knowing each section is measurable does not prove it. We establish it together with the integral identity, first where both total measures are finite.
Indicators on finite blocks
Proof
(The indicator identity on a finite block)
We use the elementary - principle: a family closed under finite intersections, contained in a -system, generates a -algebra also contained in that system. A -system contains the whole space, is closed under complements, and under countable disjoint unions. For completeness, let be the smallest -system containing the intersection-closed family. Fix a member of the original family. The class of for which is a -system containing the original family, using closure under differences of nested members. It therefore contains . Repeating with a fixed proves that is itself intersection-closed. Complements and disjoint unions now make it a -algebra. This proves the principle.
Work on with finite marginal measures. The trace of the product -algebra on equals the product of the two trace -algebras. To check this, let . Every restricted rectangle belongs to the trace of , giving one inclusion. For the other, the global sets whose intersections with belong to form a -algebra containing every measurable rectangle, hence contain the whole product -algebra.
Let consist of the measurable subsets of for which on is measurable and
The whole block and the empty set belong to . For a relative complement, the section measure is , so it is measurable and has the required integral by subtraction. Every number being subtracted here is finite. For disjoint unions, section measures add; monotone convergence for their finite partial sums proves the identity, and countable additivity gives the product measure on the other side. Thus is a -system.
For a measurable rectangle the formula is exactly the assumed rectangle property. Rectangles form an intersection-closed generating family for the restricted product -algebra. The - principle gives the indicator identity for every measurable set in the block. This also proves measurability of its section measure.
Assemble the -finite space, then pass to functions
Proof
(Exhaustion and simple approximation)
For any product-measurable , apply the block result to . Define on and zero outside . These functions are measurable and increase pointwise to . Monotone convergence and continuity from below of give
The reversed argument gives the horizontal identity. The increasing finite pieces are where -finiteness enters; we have never subtracted two infinite total masses.
Finite nonnegative linear combinations of indicators give the result for nonnegative simple functions. For measurable , choose simple . For each fixed , monotone convergence in gives . These are measurable functions of . Monotone convergence again, now in , and also on the product, gives the asserted equality for . Reverse the variables for the other order. This completes Tonelli's proof, including section-integral measurability and infinite values.
Proof X-ray
Where $\sigma$-finiteness and nonnegativity enter
-finiteness enters at exactly two places. Inside a block , the finiteness of and is what keeps the good class closed under relative complements: that step subtracts from and from . If , the same step on the whole product would subtract infinite quantities, which is why the proof works in blocks. The second use is the exhaustion , where monotone convergence and continuity from below of replace any subtraction. The key move is the - principle: the identity is checked only on rectangles, where it is the assumed rectangle property, and the -system closure transfers it to every product-measurable set. After that, every limit is a monotone limit of nonnegative quantities, so no cancellation can hide in disjoint unions of sections, in the exhaustion or in the simple approximants. The counterexample below shows that without -finiteness the conclusion itself fails, not merely this proof.
Absolute integrability makes signed subtraction legitimate
Proof
(Fubini by subtracting finite section integrals)
Apply Tonelli to and set . It is measurable and has finite integral, hence is finite outside a measurable -null set. Indeed a nonnegative function of finite integral cannot be infinite on a set of positive measure. Thus almost every section is integrable.
For real , Tonelli applied to and gives measurable section-integral functions with finite outer integrals. On the good set , both inner values are finite. On its null complement, set both to zero before subtracting them. Their difference is measurable, off that null set, and
The two finite product integrals justify this subtraction. The bound on proves the stated norm inequality. Reverse the variables for the other iterated integral. For complex , apply the real result to its real and imaginary parts, each bounded in absolute value by ; the section inequality gives the same norm bound.
Proof
(The corollary from Tonelli on indicators and on )
Finally, Tonelli applied to proves the null-section assertion, and applied to proves the integrability test. In particular, two product-measurable representatives agreeing off a product-null set have equal section integrals for almost every outer variable. The exceptional set need not be empty. None of this asserts that every section of a function measurable only for a completed product is measurable; that stronger formulation needs an additional representative argument.
Worked examples
Worked example 1.4.6
(An unbounded integrand on a triangular region)
On , let . It is nonnegative and measurable, so Tonelli applies before we know whether the integral is finite. Integrating first in gives
For fixed , the other section integral is , whose integral on is also one. This proves integrability despite the unbounded values near the corner. The section geometry explains the different inner limits; the theorem explains why the two computations describe one product integral.
Worked example 1.4.7
(Check absolute mass before using cancellation)
On , take . The bound and Tonelli give an absolute integral at most two, using . Now Fubini is available. Integrating first in gives ; first in gives . Both outer integrals are zero. The symmetry of the expression suggests cancellation, but the absolute-mass estimate is what first makes that cancellation legitimate on the product.
Worked example 1.4.8
(Finite iterated integrals can disagree)
On , define
For each fixed , the denominator is at least , so the vertical section is bounded and absolutely integrable on ; fixed works similarly. Yet
whereas . The outer integrals are respectively and .
For , the numerator is at least and the denominator at most . Therefore on that region. Tonelli for gives
Since , on the negative part is at least . Tonelli over that reflected region gives by the same divergent integral. Thus the joint signed integral is undefined, not a third number with which the two answers should agree. Sectionwise cancellation has concealed infinite positive and negative masses.
Counterexample mode
$\sigma$-finiteness is a real hypothesis
Give Lebesgue measure and counting measure on all subsets, using the -algebra . Counting measure here is not -finite: a finite-measure set is finite, and countably many such sets cannot cover an uncountable interval. The diagonal is product-measurable: partition at each dyadic scale into half-open cells with the last cell closed at one, take the union of matching-cell rectangles, then intersect over the scales. Each vertical section is a singleton of counting measure one, whereas each horizontal section is a Lebesgue-null singleton. Its indicator therefore has iterated integrals one and zero. Outside the stated -finite setting, the conclusion equating both orders is false; product-measure construction and uniqueness also require separate care there.
Common mistakes
Common mistake
(Tonelli on the parts does not license subtracting infinities)
Tonelli does not require an integrable dominating function. Nonnegativity permits infinite values and monotone limits. For a signed function, applying Tonelli separately to its positive and negative parts does not authorize subtracting two infinities. The singular example shows that even finite answers in both signed orders do not repair that problem.
Common mistake
(Sections are integrable only almost everywhere)
Fubini guarantees integrability of sections almost everywhere, not at every point. A product-null vertical line can carry a section with infinite absolute integral. Similarly, a null change of a product-measurable representative preserves section integrals only outside an outer null set. The direction and location of that exceptional set matter.
Common mistake
(Joint measurability is a hypothesis)
Joint measurability is another explicit assumption. Separate statements about each section do not by themselves provide the measurable outer integral used in the proof. The measurable-rectangle argument supplies this for product-measurable functions. Finally, an iterated integral of is a valid integrability check even before finiteness is known, because Tonelli already applies to that nonnegative function.
Yes, under the continuum hypothesis: a set whose sections are all measurable can have iterated integrals zero and one.
A natural next question
Can measurable sections hide a set that is not product-measurable?
Yes, under the continuum hypothesis: a set whose sections are all measurable can have iterated integrals zero and one.
Assume the continuum hypothesis, and well-order so that every element has only countably many predecessors. Let . Each horizontal section is countable, so , while each vertical section has countable complement, so . Every section is measurable, yet is not product-measurable, since otherwise Tonelli would make the two values equal.
Summary
Product rectangles give the indicator identity on finite blocks; -finiteness and monotone convergence assemble those blocks. Simple approximation then proves Tonelli for all nonnegative measurable functions. Applying it to isolates good integrable sections and makes the signed subtraction in Fubini valid. Absolute mass, measurability and almost-everywhere qualifications are separate parts of the conclusion.
Exercises
Quick check
(On , let , with real . Determine its product integral and the exact range of integrability. Explain why Tonelli applies for every .)
Solution · Answer
The function is nonnegative and measurable for every real . Tonelli gives . This is when and infinite when . The infinite answer is permitted by Tonelli; absolute integrability holds exactly in the first range.
Quick check
(On with counting measures, let if , if , and zero otherwise. Compute the two iterated sums and the sum of absolute values. Identify the failed Fubini hypothesis.)
Solution · Answer
Every row has one and one , so summing rows first gives zero. Column one has sum one, while each later column has a and a , so summing columns first gives one. The sum of absolute values is infinite; in fact both positive and negative masses are infinite. Both factors are -finite, but absolute integrability fails. A joint signed sum is therefore not defined by Lebesgue integration.
Quick check
(On with Lebesgue measures, take . Find its product integral and identify the section that is not integrable. Why does this agree with Fubini?)
Solution · Answer
Every section with is zero, while the section at zero is and has infinite integral. Tonelli gives product integral zero, since the section-integral function is zero except for an infinite value on a null singleton. Thus is absolutely integrable on the product. Fubini permits that exceptional outer point and defines the outer integral after assigning its section value zero there.
Quick check
(On a -finite measure space , for measurable , prove the layer-cake formula . Include joint measurability of the indicator used in the proof.)
Solution · Answer
The set is measurable because it equals . Its vertical section has length , including infinite height, and its horizontal section is for . Apply Tonelli to with Lebesgue measure on . No integrability assumption on is needed.
Quick check
(On the same -finite product spaces, let a finite-valued nonnegative product-measurable kernel satisfy for -almost every . For measurable integrable on , prove that is defined almost everywhere and belongs to , with . Set it to zero on exceptional outer points.)
Solution · Answer
Choose a finite-valued measurable representative of , changing it on a null set if necessary. Define . The pullback is product-measurable, so both and are product-measurable. Tonelli gives . The exceptional column set has zero contribution to this nonnegative integral. Thus is absolutely integrable, and Fubini supplies almost-everywhere existence and measurability of . The section triangle inequality followed by Tonelli gives . This also shows that changing the representative of does not change as an element.