An integration formula has three pieces: the region, the integrand and the area element. Changing their order or coordinates can simplify a calculation, but it does not automatically preserve all three. A successful setup explains which points are counted, how many times they are counted and why the integral exists. This chapter develops those checks for proper integrals and then shows exactly what can go wrong at a singularity.
Slicing a region before integrating
The model result is the rectangle case.
Theorem 2.4.1
(Iterated integrals on a rectangle)
Let be continuous on a compact rectangle . Then
On a bounded Jordan region with suitable piecewise descriptions, the same conclusion can be obtained using zero extension and the appropriate integrability hypotheses. This is a theorem about a fixed proper integral; it is not permission to interchange arbitrary limits of improper integrals.
A vertical slice holds fixed and lets vary. A horizontal slice holds fixed and lets vary. The endpoints of a slice must describe the same set as the original inequalities. Sketch first, but verify the bounds algebraically: drawings do not decide which branch of a square root is valid or whether a boundary is included.
Worked example 2.4.2
(Reversing the order over a curved region)
Let . Reverse the order of integration and evaluate both ways.
For the displayed description, gives . The other inequality, since , gives , or . The two curves meet when , so . Therefore
The sign condition matters: squaring an inequality involving a negative would not be equivalent, but here the original bounds force . Now the reversed integral is
In the original order,
Agreement checks the setup, although two equal numerical results alone would not prove the regions were equal. The algebra above establishes the equivalence of the descriptions.
Quick check
(Reverse the order of integration in and evaluate it.)
Sketch the region , and read off a horizontal slice.
Solution · Answer
The region is , . Since , the curve is , so a horizontal slice at height runs from to . Hence , which agrees with the original value .
A determinant measures local scale, not coverage
Definition 2.4.3
(Jacobian determinant and area scale)
For a map , let be the matrix whose columns are and . Its determinant is the Jacobian determinant of . The number is the local area scale, and the sign of records whether preserves or reverses orientation.
The columns of approximate the image of a small coordinate rectangle by a parallelogram. Its ordinary area is multiplied by . A negative determinant reverses orientation; it does not make scalar area negative. The proper change-of-variables theorem turns this into a statement about integrals.
Theorem 2.4.4
(Proper change of variables)
Let be a compact Jordan region, let be a diffeomorphism on suitable open neighborhoods of and , and let be continuous on . Then
These hypotheses are sufficient, not necessary. A nonzero Jacobian at every interior point establishes local behavior but does not by itself establish a globally one-to-one parametrization. A coordinate map that covers a positive-area region twice will generally produce twice the desired integral unless multiplicity is accounted for.
Worked example 2.4.5
(A parallelogram as the image of a square)
Evaluate on the parallelogram
Use , . The inverse is , , so corresponds exactly to . The map is a global invertible linear transformation, with
The determinant is negative because the map reverses orientation, while the area scale is . The continuous integrand and compact square satisfy the stated theorem. Consequently
The inner and outer limits are both because the inverse inequalities, rather than the shape of alone, produced the parameter square. A quick area check is ; the negative determinant cannot mean negative area.
Common mistake
(A negative Jacobian is not a negative area)
Writing with the signed determinant would give in the example above, a negative value for the integral of a positive function. The theorem uses ; the sign belongs to orientation, not to area.
Quick check
(The map is applied to the unit square . What is the area of the image, and does preserve orientation?)
Compute from the columns and .
Solution · Answer
. The map is linear with nonzero determinant, hence injective, and the image has area . The negative sign means that orientation is reversed.
Polar seams and the origin
Polar coordinates have determinant . For a bounded continuous integrand on a disk, restrict initially to and a once-only angular interval; cover the disk using suitable pieces or truncate the origin and seam. The seam has area zero, and if the omitted disk of radius contributes at most . This justifies the limit for the bounded proper problem under the stated piecewise hypotheses. It does not show that an arbitrary unbounded integrand has a negligible singularity.
Common mistake
(Reading the Jacobian as a count of preimages)
The range represents each nonzero point once, whereas traces it twice. For the unit disk, is twice the area, despite the Jacobian being locally correct. The Jacobian answers “how much local area?”, not “how many preimages?” Before integrating, state a parameter domain on which the map is one-to-one outside a set of zero area, such as .
Quick check
(Evaluate and compare it with the area of the disk .)
Do the inner integral first, then count how often each point of the disk is traced.
Solution · Answer
, so the value is . The disk has area : since runs over , every point other than the origin is traced twice.
Two improper iterated integrals disagree
Worked example 2.4.6
(Iterated integrals of a singular function)
On the punctured closed square consider
Compute both iterated integrals by first taking the proper inner integral at each positive value of the outer variable, then taking the outer limit as that variable approaches zero. Explain why their disagreement does not contradict the proper Fubini theorem.
Fix . Since
integrating in first gives the proper integral . The outer limit is . Conversely, for fixed ,
so the proper inner integral is and . The resulting inner-integral functions extend continuously to the outer endpoint, but the excluded slices themselves do not have finite integrals: diverges to , and diverges to . The field is unbounded at the origin, so it is not a continuous function on the closed rectangle; the proper theorem quoted above does not apply. Values assigned on the axes cannot remove the singular behavior near their intersection.
The failure of absolute integrability is visible in polar coordinates on the first-quadrant sector , . There and . The absolute integral on a truncated sector is at least
Thus no absolutely integrable version of Fubini can be invoked either. The polar calculation is made on truncated sectors, where the map is regular, before taking the displayed limit. It does not presume an unproved change-of-variables theorem at the singular point.
Concept lensAnalytic
The integrand as a mixed derivative of the polar angle
For , the integrand is , where is the polar angle. An iterated integral over therefore evaluates at the four corners: it equals . As , has limit along the -axis and along the -axis, so the corner values depend on how the corner approaches the origin. That is why the answers , and differ.
A repeatable audit
Before reversing an order, identify every slice, including any split into pieces. Before a coordinate change, state the parameter domain, inverse or multiplicity, Jacobian sign and absolute scale, and treatment of seams. For an improper integral, state which truncations define it and prove convergence independently; agreement of selected iterated integrals does not by itself prove a joint improper integral exists. For a nonnegative integrand, zero or positive contributions cannot be hidden by cancellation, which often supplies a useful divergence test.
Quick check
(If is nonzero everywhere on an open parameter region, must be injective there?)
Test polar coordinates on a long angular interval.
Solution · Answer
No. Polar coordinates on have nonzero Jacobian but cover most points twice. Local invertibility is not global coverage.
A nonlinear map with an explicit inverse
Worked example 2.4.7
(A rectangle mapped onto a trapezoid)
Consider on . Its image is , because and . Conversely, these inequalities put in and in . The inverse is on the open half-plane , and is a diffeomorphism on the corresponding open half-plane . Its determinant is , positive throughout .
As a check, for the coordinate theorem gives
Direct vertical slicing gives . The inverse identifies the global coverage, while the determinant supplies local scaling. The restriction keeps us away from the singular line ; neither step should be silently carried over to a domain crossing that line.
An exhaustion is itself a choice
Counterexample mode
Different truncations, different limits
Suppose one claims that every way of cutting out the singularity of must give the same limit. It does not. Here . For every , the integral over the symmetric square is zero: swapping the two coordinates leaves the square unchanged and reverses the integrand's sign. Thus this particular simultaneous square exhaustion tends to zero, even though the two iterated procedures give and .
There is no contradiction: the singular integrand is not absolutely integrable near zero, and these limiting procedures are not covered by a theorem asserting independence of truncation. One chosen exhaustion producing a finite number is a well-defined statement about that exhaustion; it is not a proof that every natural way of approaching the singularity gives the same number. Writing the truncation explicitly exposes the extra choice.
Why zero extension may still be Riemann integrable
On the unit square define when and otherwise. This is the zero extension of the constant function on a triangle. It is discontinuous along the diagonal, so it would be wrong to call it continuous on the entire rectangle. Nevertheless it is Riemann integrable.
Proposition 2.4.8
(A zero extension across a diagonal)
The function is Riemann integrable on , and .
Proof
(Counting the cells that meet the diagonal)
Divide the square into equal squares. Only cells meeting the diagonal can contain both zero and one; at most cells meet that diagonal, including cells touching it at a corner. Their total area is at most . On all other cells the oscillation is zero, while on exceptional cells it is at most one. The upper–lower sum gap is therefore at most . The Darboux criterion proves integrability, and horizontal slices give
Proof X-ray
Why only linearly many cells matter
The proof uses two facts and nothing more. First, a cell that does not meet the diagonal lies wholly in or wholly in , where is constant, so its oscillation is zero. Second, the diagonal meets only cells with , at most cells of area each. With this bounds the gap between upper and lower sums by . Without the hypotheses the argument fails: if were unbounded near the diagonal, the oscillation on an exceptional cell would no longer be at most one; if the discontinuity set were not thin, as for the indicator of the points with both coordinates rational, every cell would be exceptional and the gap would stay equal to .
This argument makes the boundary-null claim concrete for this region. For a different domain one must establish the corresponding boundary control rather than infer it merely from a picture. It also does not cover an unbounded singular integrand: the common oscillation bound by one was used essentially.
One slice can have two pieces
For the annulus , a horizontal slice at fixed behaves differently inside and outside the inner circle. If , it is the union
If , the inner exclusion is absent and the slice is the single interval . The boundary values can be assigned consistently to either adjacent description without changing a proper area integral. A single interval from the left outer boundary to the right outer boundary for every would fill the central hole and integrate over the disk instead. When a slice splits, write a sum of integrals or split the range; changing the order may be correct only after this geometric bookkeeping.
Even when a change of variables makes the algebra shorter, the transformed integrand must include both and the area factor. Checking a constant integrand is a fast diagnostic: if the transformed integral of does not equal the known area, there is an error in the Jacobian, bounds or multiplicity. Such a check can expose a mistake, although it cannot replace the hypotheses or a proof of coverage.