Improper Integrals

A perpetual bond pays a coupon forever — there's no maturity date to stop the clock. To value it you'd want to add up every discounted future coupon: today's, next year's, the one in year 500, the one in year 5,000, on and on with no end. An ordinary integral assumes you're summing area over a finite, well-behaved strip. Here the strip has no right edge. An improper integral is the tool that asks: does that endless sum settle down to a finite number, or does it blow up? Sometimes an infinitely long region holds a finite amount of "area," the way an infinitely long trough that narrows fast enough still holds a finite amount of water.

The analogy: a trough that narrows fast enough

Picture pouring water into a trough that stretches out to infinity but gets thinner and thinner the further out you go. If it thins out fast enough, the total water it can hold is finite — even though the trough itself has no end. If it thins too slowly, the trough can hold an infinite amount. Improper integrals are exactly this question applied to area under a curve: does the curve shrink toward zero fast enough, far enough out (or near enough to a spike), that the total area is finite?

The mechanics, one symbol at a time

An integral $\int_a^b f(x)\,dx$ is "improper" when either the interval is unbounded ($b = \infty$, or $a=-\infty$) or the function $f$ itself becomes unbounded somewhere in $[a,b]$. You handle both the same way: replace the problem spot with a variable, take the ordinary integral up to that variable, then let the variable slide toward the trouble point as a limit.

$$\int_a^{\infty} f(x)\,dx = \lim_{N \to \infty} \int_a^{N} f(x)\,dx.$$

In words: integrate normally up to some finite cutoff $N$, then ask what happens to that finite answer as you push $N$ further and further out. If the limit exists and is a finite number, the integral converges; if it grows without bound or oscillates forever, it diverges — and in that case, the "area" is simply not a well-defined finite quantity, no matter how the calculation looks on paper.

Worked example 1: valuing a perpetuity

A perpetual bond pays a continuous coupon at rate $50 per year, discounted at $r=5\%$. Its value is the improper integral

$$V = \int_0^{\infty} 50\,e^{-0.05t}\,dt = \lim_{N\to\infty}\left[-\frac{50}{0.05}e^{-0.05t}\right]_0^N = \lim_{N\to\infty}\left(1000 - 1000\,e^{-0.05N}\right).$$

As $N\to\infty$, $e^{-0.05N}\to 0$, so the limit is $1000$: the perpetuity is worth $1,000. The discounting shrinks each future coupon's present value fast enough that the infinite stream sums to a finite price — exactly the "trough narrows fast enough" case.

Worked example 2: a spike that's still finite, and one that isn't

Consider $\int_0^1 x^{-1/2}\,dx$. The integrand blows up as $x\to 0$, so this is improper at the left endpoint.

$$\int_0^1 x^{-1/2}\,dx = \lim_{\epsilon\to 0^+}\int_\epsilon^1 x^{-1/2}\,dx = \lim_{\epsilon\to 0^+}\left[2\sqrt{x}\right]_\epsilon^1 = \lim_{\epsilon\to 0^+}(2 - 2\sqrt\epsilon) = 2.$$

The spike near zero is tall but thin enough to have finite area. Compare $\int_0^1 x^{-1}\,dx = \lim_{\epsilon\to 0^+}[\ln x]_\epsilon^1 = \lim_{\epsilon\to 0^+}(0-\ln\epsilon)$, and $-\ln\epsilon \to \infty$ as $\epsilon\to 0^+$: this one diverges. The exponent $-1$ is the razor's edge between finite and infinite area — a fact that reappears constantly wherever power-law tails show up in finance.

Drag the exponent above and watch the curve near zero and near infinity: small changes in how fast a curve decays are the entire difference between a convergent and a divergent improper integral.

What this means in practice

Improper integrals show up whenever a quant integrates over an infinite horizon (perpetuities, infinite-maturity credit spreads) or across the full real line (the normal density in Black-Scholes pricing, characteristic-function methods, Laplace and Fourier transforms of payoff functions). Expected shortfall and other tail-risk measures are themselves improper integrals of a loss density over an unbounded region, and whether they're even finite depends on how fat the tail is — this is precisely why some heavy-tailed distributions have no finite mean or variance.