Statisticians and physicists routinely add a quantity up across an endless range — the total probability under a bell curve that never quite touches the axis, the whole dose delivered by a slowly decaying isotope, the total energy in a fading signal. Each is an integral running out to infinity, and the make-or-break question is always the same: does the never-ending sum settle on a finite number, or does it run away without bound?

Here is a question that sounds like a riddle: can a region that never ends have a finite area? Take the curve y = 1/x^2 and shade everything under it from x = 1 out to the right — forever. The shaded strip is infinitely long. It never stops. And yet, as we are about to prove, its total area is exactly 1. Not "roughly one", not "one-ish in the limit of hand-waving" — exactly one square unit, full stop.

Now shade under y = 1/x instead. The two curves look like siblings: both slide down toward the axis, both get as close to zero as you like. But this second strip has infinite area — no amount of paint covers it. Two curves you could barely tell apart at a glance sit on opposite sides of a razor-thin boundary, and the whole subject of improper integrals is about locating that boundary precisely.

The definite integral \int_a^b f(x)\,dx was built for a finite interval and a bounded integrand, so questions like these are technically outside its remit. An improper integral relaxes one of the two assumptions: the interval may run off to \infty (Type 1), or the integrand may blow up at a point (Type 2). Either way we give it meaning the only honest way — as a limit.

The definition: stop short, then let the stop run away

The recipe is the same every time, and it has three moves: replace the bad endpoint by a letter, integrate as usual, then take the limit. Over an infinite interval, integrate up to a finite cut-off t — a perfectly ordinary definite integral — and then let the cut-off run away:

Where the integrand has a vertical asymptote at an endpoint — say f is unbounded near a — creep up to the bad point from the safe side:

If the limit is a finite number, the integral converges to that number. If the limit is infinite or fails to exist, the integral diverges. A two-sided improper integral — infinite at both ends, or with a bad point in the middle — is split into pieces, and every piece must converge on its own; one divergent piece sinks the whole thing.

Notice what the definition refuses to do: it never treats \infty as a number you can substitute. Every improper integral is an ordinary integral wearing a limit, and the limit is where all the action is.

The star contrast: 1/x^2 converges, 1/x does not

Let us settle the riddle from the top of the page, both halves, side by side. First the well-behaved sibling. Replace \infty by t, integrate, take the limit:

The running total 1 - 1/t is the whole story in one expression: out to t = 10 you have collected area 0.9; out to t = 1000, area 0.999. The infinite tail beyond any cut-off t holds only 1/t of area — the region thins out fast enough that what remains is always small. Converges, to exactly 1.

Now the near-identical twin, with one power of x less:

The running total is \ln t, and the logarithm — famously the laziest unbounded function in mathematics — still gets everywhere eventually. Want the shaded area to pass 20? You must integrate out to t = e^{20} \approx 485{,}000{,}000. Past 100? Out to e^{100}, a number with 44 digits. It crawls, but it never stops climbing. Slow divergence is still divergence.

So the boundary between "finite area" and "infinite area" sits somewhere between the exponents 2 and 1. The next card pins it down exactly — and the answer is razor-thin.

The benchmark: \displaystyle\int_1^{\infty} x^{-p}\,dx

One family settles almost every example you will meet by comparison, so we evaluate it once and for all. The behaviour turns entirely on whether the exponent p is bigger than 1.

Step 1 — replace \infty by a finite cut-off t. By definition the improper integral is the limit of an ordinary one:

Step 2 — antidifferentiate, assuming p \neq 1. The power rule for integration adds one to the exponent and divides by the new exponent (1 - p):

Step 3 — take the limit, and watch the sign of the exponent 1 - p. Everything hinges on what t^{1-p} does as t \to \infty.

Case p > 1. Then 1 - p < 0, so t^{1-p} = 1/t^{\,p-1} \to 0. The lone survivor is the -1:

Case p < 1. Then 1 - p > 0, so t^{1-p} \to \infty and the whole expression blows up:

Case p = 1. The power rule breaks (it would divide by 1 - p = 0), and the antiderivative is the logarithm instead — the computation we did in the previous card:

So the integral converges precisely when p > 1 — and not a hair sooner. p = 1.01 converges (to 100); p = 1 diverges; p = 0.99 diverges. The boundary really is razor-thin, and it sits exactly on the borderline exponent.

Watch convergence happen

Below is y = x^{-p} on [1, \infty) with the tail area shaded out to a large cut-off. Slide p and watch the readout: for p > 1 the area accumulated out to b = 1000 has already settled almost exactly on the limiting value 1/(p-1); for p \le 1 it just keeps climbing as the cut-off marches right.

Three experiments worth doing. (1) Park the slider at p = 2 and then at p = 3 — the curves are nearly indistinguishable to the eye, yet the areas are 1 and \tfrac12: your eye is a terrible convergence test. (2) Creep from p = 1.5 down toward p = 1 and watch the limit 1/(p-1) explode — 2, then 5, then 10 — as you approach the razor-thin boundary. (3) Sit exactly on p = 1: the area out to 1000 is only \ln 1000 \approx 6.9. It looks tame — but it is on its way to infinity, at a logarithm's unhurried pace.

Type 2: when the function blows up

The second kind of impropriety hides in plain sight: the interval is finite, but somewhere on it the integrand shoots off to infinity. Consider \displaystyle\int_0^1 \frac{dx}{\sqrt{x}}. The interval [0,1] looks harmless, but as x \to 0^{+} the integrand 1/\sqrt{x} climbs without bound — the region is an infinitely tall spike rather than an infinitely long tail. Same medicine: retreat to a safe starting point t > 0, integrate, then send t creeping down to the singularity:

An infinitely tall region with area exactly 2. The spike is tall but thin — near x = 0 the region narrows fast enough that the area it adds stays under control. Converges.

Now try the same trick on \displaystyle\int_0^1 \frac{dx}{x}, and catch the divergence properly rather than waving at it:

As t \to 0^{+}, \ln t \to -\infty, so -\ln t \to +\infty. Diverges — and note the verdict came from an honest limit computation, not from "it looks infinite". In general \int_0^1 x^{-p}\,dx converges exactly when p < 1: the inequality from the tail test, flipped. Small powers tame the spike at zero but fail at infinity; large powers tame the tail but fail at zero. And poor 1/x, sitting exactly on the boundary, fails at both ends — too fat at infinity, too tall at zero.

See it explained