From a line of strips to a field of boxes
Recall how the definite integral of Volume I was built. You sliced the interval [a, b] into tiny pieces of width dx, multiplied each width by the height f(x) there, and added up the thin rectangles. That sum, a Riemann sum, crept toward an exact answer as the slices shrank — the signed area under a curve. Everything about a double integral is this same instinct, lifted up one dimension: instead of summing strips along a line, you sum little columns standing over a patch of the plane.
Picture a function z = f(x, y) as a landscape — a curved surface floating above a flat region R in the xy-plane. To find the volume trapped between that surface and the floor, chop R into a grid of tiny rectangles, each with area dA = dx dy. Over one such rectangle the surface is nearly flat at some height f(x*, y*), so the column standing on it has volume roughly f(x*, y*) dA. Add up every column and you get a number that improves as the grid gets finer. That limit of sums is the [[double-integral|double integral]], written as the double integral over R of f(x, y) dA.
Formally, you take the limit of these Riemann sums as the largest cell diameter goes to 0: the double integral is the limit of the sum over i of f(x_i*, y_i*) times dA_i. When that limit exists and is the same no matter how you cut the grid or where inside each cell you sample, f is called integrable on R. Continuous functions on a sensible bounded region always qualify, so for the working problems ahead you rarely have to worry — but the value is defined by this sum-and-limit, not by any formula. Volume is just the first interpretation; the very same machine will later add up mass, charge, or probability.
Fubini's theorem: a hard sum, one slice at a time
The definition is beautiful but useless for computing — nobody sums infinitely many shrinking columns by hand. The escape is [[fubinis-theorem-computational|Fubini's theorem]], which says a double integral can be evaluated as an iterated integral: two ordinary single integrals nested inside each other, each of which you already know how to do. The picture is slicing the solid with parallel cuts. Hold y fixed; the cross-section there is a thin slab whose area is the single integral over x of f(x, y) dx. That area is a function of y alone, A(y). Then sweep y across its range and integrate A(y) — you have summed the slabs into the full volume.
On a plain rectangle a <= x <= b, c <= y <= d, Fubini's theorem reads: the double integral equals the integral from c to d of [ the integral from a to b of f(x, y) dx ] dy. The inner integral treats y as a frozen constant — exactly the spirit of a partial derivative run in reverse — and produces a function of y; the outer integral then finishes the job. Concretely, the double integral of x*y over the unit square: inner integral over x of x*y dx from 0 to 1 gives y/2; outer integral over y of y/2 dy from 0 to 1 gives 1/4. One genuinely two-dimensional sum, dismantled into two one-dimensional ones.
When the region is not a box
Real problems rarely hand you a tidy rectangle. The region might be a triangle, a disk, or the area pinched between two curves. The fix is to let the inner limits depend on the outer variable. Suppose you integrate first in y over a [[region-of-integration|region of integration]] that runs from a bottom curve y = g(x) up to a top curve y = h(x), while x ranges over [a, b]. Then the iterated integral becomes the integral from a to b of [ the integral from g(x) to h(x) of f(x, y) dy ] dx. The outer limits are always plain constants; the inner limits may be functions that describe the moving walls of the region.
Try the triangle with corners (0, 0), (1, 0), (1, 1) — the region where 0 <= y <= x and x runs from 0 to 1. Integrate f = 1 there and you should recover the triangle's area, 1/2. Inner integral over y from 0 to x of 1 dy is x; outer integral over x from 0 to 1 of x dx is 1/2. The inner limit x is not a typo — it is the slanted edge y = x of the triangle written into the bookkeeping. This is also the cleanest way to compute an area between two curves you met in Volume I: it is simply the double integral of 1 over the enclosed region.
Because the same region can be described two ways — sweeping in y first, or in x first — you can often swap the [[order-of-integration|order of integration]], and one order is frequently far easier than the other. Swapping is not a mechanical relabel of the limits: you must redraw the region, then read its bounds off the other axis. A famous payoff is the integral over x from 0 to 1 of [ the integral over y from x to 1 of e^{y^2} dy ] dx. The inner integral has no elementary antiderivative and stalls you cold. Reverse the order — same triangular region, now 0 <= x <= y, 0 <= y <= 1 — and the inner integral over x of e^{y^2} dx becomes y times e^{y^2}, which integrates instantly to (e - 1)/2. Same region, same answer, wildly different difficulty.
Same triangle {0<=x<=y, 0<=y<=1} = {x<=y<=1, 0<=x<=1}, two readings:
inner dy, outer dx inner dx, outer dy
-------------------- --------------------
int_{x=0}^{1} int_{y=x}^{1} ... int_{y=0}^{1} int_{x=0}^{y} ...
outer limits: constants 0,1 outer limits: constants 0,1
inner limits: y from x to 1 inner limits: x from 0 to y
RULE: outer limits are ALWAYS constants; only inner limits may carry
the other variable. To swap: redraw R, re-read the bounds.Stacking up to triple integrals
Nothing new in spirit happens when you climb to three dimensions. A [[triple-integral|triple integral]] chops a solid region E in space into tiny boxes of volume dV = dx dy dz, weights each box by the value of f(x, y, z) inside it, and sums. It is again a limit of Riemann sums, now over a three-dimensional grid. The honest mental image: you are no longer measuring the volume under a surface — for that, integrate the constant 1, which adds up dV into the plain volume of the solid. With a nonconstant f you are summing a density spread through the solid, and the natural reading is total amount, not volume.
Fubini's theorem extends cleanly: a triple integral unwinds into three nested single integrals. The discipline is the same, just one layer deeper — the innermost limits may depend on both outer variables, the middle limits on the single outermost variable, and the outermost limits must be bare constants. For a box-shaped region every limit is a constant and you simply integrate z, then y, then x in any order you like. For a curved solid you describe it as 'z runs from a bottom surface to a top surface, over a shadow region R in the xy-plane', do the z-integral first, and what remains is exactly a double integral over the shadow R — collapsing the new problem back onto the one you just learned.
- Sketch the solid E and pick an order, say dz then dy then dx. The sketch is not optional — almost every wrong answer is a wrongly-drawn region, not a botched antiderivative.
- Find the innermost limits: for fixed (x, y), where does z enter and leave the solid? These give the bottom and top surfaces z = z_low(x, y) and z = z_high(x, y).
- Project the solid down onto the xy-plane to get its shadow R, then set the middle and outer limits exactly as you would for a double integral over R.
- Integrate from the inside out — z first, then y, then x — treating all not-yet-integrated variables as constants at each stage, just like reversing a partial derivative.
Why this is the engine for the whole rung
Step back and notice what you actually own now. A multiple integral is the only honest way to add a quantity up over an area or a volume, and Fubini's theorem makes that addition computable by reducing it to the single integrals from Volume I. Swap in different integrands and the same machinery reads off different physical truths: integrate density rho(x, y, z) to get total mass, integrate x weighted by density to find a centroid coordinate, integrate the squared distance to an axis to get a moment of inertia. The integral does not change; only the story you tell about the integrand does.
There is one large gap left, and the rest of this rung is built to close it. So far every region has been described with straight x-y-z slabs, which is miserable for anything round: a disk, a cylinder, a sphere all fight a rectangular grid. The next guides introduce polar, cylindrical, and spherical coordinates, where a circle is just 'r = constant'. Changing coordinates inside an integral is not free, though — areas and volumes stretch under the new ruler, and a correction factor called the Jacobian determinant must ride along to keep dA and dV honest. Master the plain x-y-z setup here, and that change of variables becomes a clean upgrade rather than a leap into fog.