Improper Integrals and Convergence

Motivation: why this matters in quant finance

Almost every integral in quant finance is improper. The expected value of a continuous random variable is

\mathbb{E}[X] = \int_{-\infty}^{\infty} x\,f(x)\,dx

— an integral over an infinite domain. The normal distribution CDF is

\Phi(z) = \int_{-\infty}^{z} \phi(t)\,dt

. The present value of a perpetuity is

\int_0^{\infty} c\,e^{-rt}\,dt

. Every option pricing formula involves integrating a payoff against a density over an unbounded range.

The question that improper integrals force you to ask is: does this integral converge? In finance, convergence is not a technicality — it determines whether a price, an expected value, or a risk measure is well-defined. If

\mathbb{E}[X]

diverges, the "expected payoff" is infinite and no finite price makes sense. If a moment integral diverges, the corresponding risk measure (variance, VaR) does not exist. Understanding convergence conditions tells you when your model's outputs are meaningful.

Definition

Type I: infinite limits of integration

When one or both limits are infinite, define the integral as a limit:

\int_a^{\infty} f(x)\,dx = \lim_{b \to \infty} \int_a^b f(x)\,dx

\int_{-\infty}^{b} f(x)\,dx = \lim_{a \to -\infty} \int_a^b f(x)\,dx

\int_{-\infty}^{\infty} f(x)\,dx = \int_{-\infty}^{c} f(x)\,dx + \int_c^{\infty} f(x)\,dx

The integral converges if the limit exists and is finite; it diverges otherwise. For the doubly-infinite case, both halves must converge independently.

Type II: unbounded integrands

When $f$ is unbounded at an interior or endpoint $c \in [a, b]$ :

\int_a^{b} f(x)\,dx = \lim_{\varepsilon \to 0^+} \int_a^{c-\varepsilon} f(x)\,dx + \lim_{\varepsilon \to 0^+} \int_{c+\varepsilon}^{b} f(x)\,dx

Both limits must exist independently.

Convergence tests

Comparison test

If $0 \leq f(x) \leq g(x)$ for all $x \geq a$ :

$\int_a^{\infty} g\,dx$ converges $\implies$ $\int_a^{\infty} f\,dx$ converges.
$\int_a^{\infty} f\,dx$ diverges $\implies$ $\int_a^{\infty} g\,dx$ diverges.

Benchmark:

\int_1^{\infty} x^{-p}\,dx

converges if and only if

p > 1

Limit comparison test

If $f, g > 0$ and $\lim_{x \to \infty} f(x)/g(x) = L$ with $0 < L < \infty$ , then $\int f$ and $\int g$ either both converge or both diverge.

Absolute convergence

\int_a^{\infty} f\,dx

converges absolutely if

\int_a^{\infty} |f|\,dx

converges. Absolute convergence implies convergence (but not vice versa).

Key integrals in quant finance

The Gaussian integral

\int_{-\infty}^{\infty} e^{-x^2/2}\,dx = \sqrt{2\pi}

This is the normalisation constant for the normal distribution. It converges because

e^{-x^2/2}

decays faster than any power

x^{-p}

. This integral has no elementary antiderivative — it must be evaluated by the polar-coordinates trick (squaring the integral and converting to

\int_0^{\infty} r\,e^{-r^2/2}\,dr\,d\theta

The exponential discount integral

\int_0^{\infty} e^{-rt}\,dt = \frac{1}{r}, \quad r > 0

This gives the value of a perpetual unit cash flow stream discounted at rate

r

. Convergence requires

r > 0

: if

r \leq 0

, the integral diverges (a perpetuity has infinite value when the discount rate is zero or negative). See Discounting.

Moments of the log-normal distribution

If $X \sim \text{LogNormal}(\mu, \sigma^2)$ , the $n$ th moment is:

\mathbb{E}[X^n] = \int_0^{\infty} x^n \cdot \frac{1}{x\sigma\sqrt{2\pi}}e^{-(\ln x - \mu)^2/(2\sigma^2)}\,dx = e^{n\mu + n^2\sigma^2/2}

This converges for all

n

— the log-normal has finite moments of all orders. In contrast, distributions with power-law tails (Pareto, Student's

t

) may have divergent higher moments.

The option pricing integral

C = e^{-rT}\int_K^{\infty} (s - K)\,f_{S_T}(s)\,ds

Under GBM (

f_{S_T}

is log-normal), this integral converges because the log-normal density decays exponentially in

\ln s

. But under heavy-tailed models (e.g., stable distributions), the integral may diverge, meaning the option has no finite Black-Scholes-style price. This is a genuine model constraint: your distributional assumption must produce convergent pricing integrals.

Moment existence and tail behaviour

The connection between tail decay and moment existence is fundamental:

\mathbb{E}[|X|^n] = \int_{-\infty}^{\infty} |x|^n f(x)\,dx < \infty \iff \text{the } n\text{th moment exists}

For the integral to converge, $|x|^n f(x)$ must decay fast enough as $|x| \to \infty$ .

Distribution	Tail decay	Moments that exist
Normal	$\sim e^{-x^2/2}$	All moments
Log-normal	$\sim e^{-(\ln x)^2}$	All moments
Student's $t$ ( $\nu$ d.f.)	$\sim	x
Cauchy ( $\nu = 1$ )	$\sim	x
Pareto ( $\alpha$ )	$\sim x^{-\alpha - 1}$	$n < \alpha$ only

In finance, if your return model has a Student's $t$ distribution with $\nu = 3$ degrees of freedom, the variance exists ( $n = 2 < 3$ ) but the kurtosis does not ( $n = 4 > 3$ ). This is not a computational problem — it is a mathematical fact that the kurtosis integral diverges. Risk measures that depend on the fourth moment (like kurtosis-adjusted VaR) are undefined under this model.

Examples and applications

Example 1: present value of a growing perpetuity

A cash flow growing at rate $g$ and discounted at rate $r$ :

\text{PV} = \int_0^{\infty} c\,e^{(g - r)t}\,dt = \frac{c}{r - g}, \quad r > g

If $g \geq r$ (growth matches or exceeds the discount rate), the integral diverges — the present value is infinite. The condition $r > g$ is the convergence condition and is the continuous-time analogue of the Gordon growth model constraint.

Example 2: expected shortfall

The Expected Shortfall (CVaR) at level

\alpha

is:

\text{ES}_\alpha = -\frac{1}{\alpha}\int_0^{\alpha} F^{-1}(u)\,du

where $F^{-1}$ is the quantile function. For this to be finite, we need $\mathbb{E}[|X|] < \infty$ (the first moment must exist). Under a Cauchy distribution, $\mathbb{E}[|X|] = \infty$ , so ES is undefined — you literally cannot compute the average loss beyond VaR. This is a convergence failure of a financial risk measure, not a computational error.

Example 3: normalisation of the normal density

To verify that $\phi(z) = \frac{1}{\sqrt{2\pi}}e^{-z^2/2}$ is a valid density, we must check:

\int_{-\infty}^{\infty} \phi(z)\,dz = \frac{1}{\sqrt{2\pi}}\int_{-\infty}^{\infty} e^{-z^2/2}\,dz = \frac{1}{\sqrt{2\pi}} \cdot \sqrt{2\pi} = 1

Both improper integrals $\int_{-\infty}^{0}$ and $\int_0^{\infty}$ converge separately (by comparison: $e^{-z^2/2} \leq e^{-z/2}$ for $z \geq 1$ , and $\int_1^{\infty} e^{-z/2}\,dz = 2e^{-1/2} < \infty$ ).

Common confusions and pitfalls

"The integral $\int_{-\infty}^{\infty} x\,dx$ is zero by symmetry." No. The two halves

\int_{-\infty}^{0} x\,dx = -\infty

and

\int_0^{\infty} x\,dx = +\infty

each diverge. The doubly-infinite integral is undefined (not zero), because

\infty - \infty

is an indeterminate form. The Cauchy distribution illustrates this: its density is symmetric, but

\mathbb{E}[X]

does not exist.

Confusing convergence with having a formula. An integral can converge without having a closed-form antiderivative (

\int e^{-x^2}\,dx

), and an antiderivative can exist without the integral converging (

\int_0^{\infty} \cos x\,dx

oscillates and does not converge, even though

\sin x

is the antiderivative). Convergence and closed-form solvability are independent properties.

Ignoring convergence when choosing a model. If you assume Student's

t_3

returns, your model has no finite kurtosis. If you assume Cauchy returns, it has no finite mean. These are not bugs in your code — they are features of the model that you chose. Always check that the integrals your analysis requires (moments, pricing expectations, risk measures) actually converge under your distributional assumption.

Where this goes next

Improper integrals connect to:

Introduction to Integration: Proper Riemann integrals are the building block; improper integrals extend them to unbounded domains and integrands.
Numerical Integration: In practice, improper integrals are truncated and computed numerically. The convergence rate determines how much of the tail you can safely ignore.
Normal Distribution and Log-Normal Distribution: The Gaussian integral and the log-normal moment integrals are the most important convergent improper integrals in quant finance.
Multiple Integrals: Multi-asset expected values are improper integrals in multiple dimensions, with convergence conditions that depend on the joint tail behaviour.

References

Stewart, J. (2008). Single Variable Calculus: Early Transcendentals (6th ed.). Thomson Brooks/Cole. Ch. 7 Section 7.8 (Improper Integrals) for infinite intervals, unbounded integrands, and convergence tests.