The integral is over a region, not a box
A double integral of f over a region R is, at heart, the same limit of Riemann sums you met in Volume I: chop R into tiny patches, multiply each patch's area by the value of f on it, and add. The previous guide turned that sum into an iterated integral — two ordinary single integrals nested one inside the other — and over a rectangle the bookkeeping was easy because every horizontal slice runs between the same two x-values and every vertical slice between the same two y-values. The constant limits are exactly what made it painless.
Real regions are almost never boxes. They are triangles, disks, the sliver trapped between a line and a parabola, the wedge under a curve. The whole skill of this guide is translating such a shape into the limits of an iterated integral — and the key realization is that for a non-rectangular region at least one pair of limits must be variable, a function of the other variable rather than a constant. Picturing how the slices stretch as you move across the region is the entire game; the antiderivatives at the end are usually the easy part.
Two ways to slice: Type I and Type II
There are exactly two natural ways to cut a planar region of integration into slices, and naming them keeps the setup honest. A Type I region is one you sweep with vertical slices: x runs over a fixed interval a to b, and for each frozen x the slice rises from a bottom curve y = g(x) to a top curve y = h(x). A Type II region is the mirror image — horizontal slices, with y running over c to d and x sweeping from a left curve x = p(y) to a right curve x = q(y). Picking Type I or Type II is exactly picking which variable to integrate first.
Make it concrete with the triangle whose corners are (0, 0), (1, 0) and (1, 1) — the region under the line y = x and above the x-axis, out to x = 1. Read it as Type I: x runs from 0 to 1, and for each such x the vertical slice climbs from the floor y = 0 up to the diagonal y = x. So the inner y-limits are 0 and x — variable, because the triangle gets taller as x grows. Read the SAME triangle as Type II: y runs from 0 to 1, and for each y the horizontal slice runs from the diagonal x = y across to the right wall x = 1. Now the inner x-limits are y and 1. One shape, two faithful descriptions.
Why both orders give the same number
It is not obvious that slicing a region vertically and slicing it horizontally should produce the same total. The guarantee is Fubini's theorem: as long as the integrand is reasonable over the region — for a bounded, continuous f on a bounded region you are safe — the iterated integral computed in either order equals the single genuine double integral, and therefore the two orders equal each other. Geometrically it is just the obvious fact that adding up the area-by-height contributions does not care whether you bundle them into vertical columns first or horizontal rows first.
Be honest about the fine print, because it is not mere pedantry. Fubini's theorem can fail when f is not absolutely integrable — when the positive and negative parts each pile up to infinity. The standard cautionary tale is f(x, y) = (x^2 - y^2) / (x^2 + y^2)^2 on the unit square: integrate x-first and you get one number, y-first and you get its negative. Nothing is wrong with the arithmetic; the hypothesis simply was not met, because the integral of |f| diverges. For the bounded, continuous integrands of this rung you will never hit this, but it is why the theorem carries conditions rather than being a free law.
Reversing the order, step by step
Because both orders give the same number, you are free to choose whichever is easier — and reversing a given order of integration is a concrete, mechanical skill. The trap is to merely swap the dx and dy and shuffle the existing limits around; that almost always produces nonsense. The only reliable method is to throw away the limits entirely, reconstruct the region they describe, and then re-read that region the other way. The limits are a coded description of a shape; you must decode it before you can re-encode it.
- Read the given limits as inequalities. For example the integral from 0 to 1 (outer, dx) of the integral from x to 1 (inner, dy) decodes to 0 <= x <= 1 and x <= y <= 1.
- Sketch the region those inequalities carve out, and identify each bounding curve by name (here: the line y = x, the top y = 1, the left wall x = 0).
- Now slice the OTHER way. To put y outside, find the full range of y over the whole region (0 to 1), then for a fixed y read off where x enters and exits (here x from 0 to y).
- Write the new iterated integral with constant OUTER limits and the variable INNER limits you just read, then sanity-check a corner or two against the picture.
A frequent stumble: when the region is shaped like a disk or a curved cap, swapping the order can split it into two pieces, each needing its own iterated integral, because along the new slicing direction the entry or exit curve changes partway through. That is not an error — it is the region honestly telling you it is not Type I and Type II in one clean piece. Sketching is what reveals the split before it bites you. (When a disk forces this, that is precisely the cue to reach for polar coordinates, the subject of a later guide in this rung.)
When the order is the whole point
Sometimes reversing the order is not a convenience but the only way through. The famous example is the integral, over the triangle 0 <= x <= y, 0 <= y <= 1, of e^{-y^2}. Set up y-inner and you are stuck immediately: the inner antiderivative needs the integral of e^{-y^2} dy, which has no elementary closed form — this is a genuine non-elementary integral, the very kernel of the Gaussian. "Non-elementary" does not mean impossible or uncomputable; it means no formula in powers, exponentials, logs and trig functions exists. The single integral is a dead end for an exact answer this way.
Now flip it. Put y outside, x inside. For fixed y the inner integral is the integral from x = 0 to x = y of e^{-y^2} dx — and here e^{-y^2} is a constant as far as x is concerned, so the inner integral is just e^{-y^2} times the length y. The outer integral becomes the integral from 0 to 1 of y e^{-y^2} dy, which the substitution u = y^2 cracks instantly: it equals (1/2)(1 - e^{-1}), about 0.316. The integrand never changed; reversing the order moved the bad variable to the outside, where a friendly extra factor of y appeared and rescued the whole computation.
Region (a triangle): 0 <= x <= y , 0 <= y <= 1
y
1 | *--------* slice the BAD way (y inner): each vertical
| | \ | slice needs integral of e^{-y^2} dy -- no
| | \ | elementary antiderivative. STUCK.
| | \ |
0 +--*-------*---- x slice the GOOD way (x inner, y outer):
0 1
inner: integral_{x=0}^{y} e^{-y^2} dx = e^{-y^2} * (y - 0) = y e^{-y^2}
outer: integral_{y=0}^{1} y e^{-y^2} dy -- let u = y^2, du = 2y dy
= (1/2) integral_{0}^{1} e^{-u} du = (1/2)(1 - e^{-1}) ~ 0.316
Same region, same integrand -- only the ORDER changed.From areas up to volumes
The same logic, one dimension richer, sets up a triple integral over a solid. Now you have three orders of nesting to choose among (six, if you count all permutations), and the inner limits can depend on the two outer variables while the middle limits depend only on the outermost. The discipline is unchanged: project the solid onto a coordinate plane to fix the outer two limits, then for a frozen point in that shadow read where the innermost variable enters and exits the solid — a bottom surface z = g(x, y) and a top surface z = h(x, y). Sketching a representative spear through the solid is again the move that prevents errors.
Reordering helps in three dimensions for the same two reasons it helps in two: a cleaner region description, or a tractable innermost antiderivative. And the payoff connects straight back to Volume I. Recall that a single definite integral of a positive function is the area under a curve, justified by the fundamental theorem of calculus; a double integral of a positive function is the volume under a surface; a triple integral of the constant 1 is the plain volume of the solid. Choosing the order well is, in every dimension, the difference between an integral you can finish and one you only stare at.