Improper Integrals

Some integrals run to infinity or over blowups yet still settle on finite numbers; a limit decides whether the total converges or diverges.

Two tails from 1 to infinity side by side: 1 over x squared encloses a finite area of exactly 1 that the tail completes by 10, while 1 over x keeps accumulating without bound

Why Does This Exist?

Probability lives on infinite domains: Gaussians run to $\pm\infty$ , waiting times run to $+\infty$ , and every density must integrate to exactly 1 over its whole range. Those normalizing totals are improper integrals. Expected values add more: heavy-tailed losses and returns integrate payoffs against densities out to infinity, and whether that integral converges decides whether the "expected" value even exists. Cauchy-distributed noise has no mean for exactly this reason.

Definite and indefinite integrals covers finite-interval totals. This page extends the machinery past its two failure points: infinite limits and infinite integrand values.

The one idea you need first

An improper integral is a limit of proper ones: push the boundary out (or up to the blowup) and ask where the totals head. $\int_1^{\infty} f = \lim_{b \to \infty} \int_1^b f$ . Converges means the limit lands on a finite number; diverges means it runs away or oscillates forever.

Think of It Like This

Painting an infinitely long fence that needs finite paint

A fence runs forever but gets thinner fast: the first meter needs a liter, the next half a liter, the next a quarter. The fence never ends, yet the paint bill converges to 2 liters, because the coats shrink faster than the fence grows. That is $\int_1^{\infty} x^{-2}\,dx = 1$ : infinite extent, finite total.

Now a fence whose coats thin slowly, a liter, then half, then a third: the bill is the harmonic series and it diverges, growing without bound albeit slowly. Same infinite fence, opposite verdict. Convergence is a race between growing domain and shrinking height, and only the race result matters.

How It Actually Works

Infinite limits. Replace $\infty$ with $b$ , integrate, take $b \to \infty$ :

\int_1^{\infty} \frac{1}{x^2}\,dx = \lim_{b \to \infty} \left[ -\frac{1}{x} \right]_1^b = \lim_{b \to \infty} \left(1 - \frac{1}{b}\right) = 1

In plain words: the partial total $1 - 1/b$ creeps toward 1 and the leftover $1/b$ vanishes. Finite answer from an infinite domain.

Singularities. Split at the blowup and approach from each side. $\int_0^1 x^{-1/2}\,dx = \lim_{a \to 0^+} [2\sqrt{x}]_a^1 = 2$ : the spike at 0 is narrow enough to contain finite area.

The p-test. $\int_1^{\infty} x^{-p}\,dx$ converges exactly when $p > 1$ (giving $1/(p-1)$ ) and diverges for $p \leq 1$ . So $\int_1^{\infty} 1/x\,dx = \lim_{b\to\infty}\ln b = \infty$ : diverges, slowly but surely. Memorize the boundary: barely-faster-than- $1/x$ decay still diverges.

Watch Out For

Treating infinity as a number to plug in

The symptom is " $F(\infty) - F(1)$ " written with $\infty$ evaluated like a constant, which skips the limit that decides everything and mishandles oscillation like $\int_0^{\infty}\sin x\,dx$ (no limit exists, so it diverges by oscillation). Always write the limit explicitly, integrate finitely, then let the boundary travel. If the limit does not exist, the integral diverges, however bounded the integrand looks.

Assuming every density has a mean

The Cauchy density $1/(\pi(1+x^2))$ is a valid density integrating to 1, but $\int_{-\infty}^{\infty} |x|/(\pi(1+x^2))\,dx$ diverges: both tails behave like $1/|x|$ , the $p = 1$ boundary case. The symptom is a sample average over Cauchy noise that never settles no matter how much data arrives. Check tail decay against the p-test before quoting expectations, and prefer medians for heavy-tailed data where means do not exist.

The Quick Version

Improper integrals are limits of proper ones: push the boundary, watch the total.
$\int_1^{\infty} x^{-2}\,dx = 1$ converges; $\int_1^{\infty} x^{-1}\,dx$ diverges like $\ln b$ .
The p-test: $\int_1^{\infty} x^{-p}\,dx$ converges exactly when $p > 1$ .
Split at singularities and approach each side separately.
Never plug in infinity; write the limit and check it exists.
Heavy tails can break expectations: Cauchy noise has no mean.