Multiple Integrals

Motivation: why this matters in quant finance

The moment you move beyond a single asset, you need multiple integrals. The joint expected value

\mathbb{E}[f(X, Y)] = \iint f(x, y)\,p(x, y)\,dx\,dy

is a double integral over the joint density. A basket option on

n

assets has price:

V = e^{-rT}\int \cdots \int \left(\sum w_i S_i - K\right)^+ p(S_1, \ldots, S_n)\,dS_1\cdots dS_n

an $n$ -dimensional integral. Covariance, correlation, and multi-asset risk measures are all defined through multiple integrals. Any quant working with portfolios, correlation products, or multi-factor models must understand how to set up, evaluate, and (when necessary) numerically approximate integrals in multiple dimensions.

Definition

Double integrals

For a function $f(x, y)$ over a region $R \subseteq \mathbb{R}^2$ :

\iint_R f(x, y)\,dA = \lim_{\|\Delta\| \to 0} \sum_{i,j} f(x_i^*, y_j^*)\,\Delta A_{ij}

where the region is partitioned into small rectangles of area

\Delta A_{ij}

. The limit exists whenever

f

is continuous (or bounded with discontinuities on a set of measure zero).

Iterated integrals and Fubini's theorem

The key computational tool: a double integral can be evaluated as two nested single integrals.

Fubini's theorem: If

f

is continuous on a rectangle

R = [a, b] \times [c, d]

, then:

\iint_R f(x, y)\,dA = \int_a^b \left(\int_c^d f(x, y)\,dy\right)dx = \int_c^d \left(\int_a^b f(x, y)\,dx\right)dy

The order of integration can be swapped. For more general regions, Fubini's theorem holds whenever $\iint |f|\,dA < \infty$ (the integral converges absolutely).

Why this matters: Swapping the order of integration is not just a computational trick — it is often the difference between a tractable and an intractable calculation. In quant finance, the order in which you integrate over time, price, and measure can dramatically simplify pricing formulas.

$n$ -dimensional integrals

For $f: \mathbb{R}^n \to \mathbb{R}$ :

\int_{\mathbb{R}^n} f(\mathbf{x})\,d\mathbf{x} = \int_{-\infty}^{\infty} \cdots \int_{-\infty}^{\infty} f(x_1, \ldots, x_n)\,dx_1\cdots dx_n

By Fubini, this equals any iterated integral (in any order) when $f$ is absolutely integrable.

Change of variables in multiple integrals

The multivariable substitution formula generalises the single-variable change of variables. If

\mathbf{x} = \mathbf{g}(\mathbf{u})

is a smooth, invertible transformation:

\int f(\mathbf{x})\,d\mathbf{x} = \int f(\mathbf{g}(\mathbf{u}))\,|\det J_{\mathbf{g}}(\mathbf{u})|\,d\mathbf{u}

where

J_{\mathbf{g}}

is the Jacobian matrix of

\mathbf{g}

and

|\det J_{\mathbf{g}}|

is the absolute value of its determinant — the "volume scaling factor."

Polar coordinates (2D)

$x = r\cos\theta$ , $y = r\sin\theta$ , $|\det J| = r$ :

\iint f(x, y)\,dx\,dy = \int_0^{2\pi}\int_0^{\infty} f(r\cos\theta, r\sin\theta)\,r\,dr\,d\theta

This is how the Gaussian integral $\int e^{-x^2/2}\,dx = \sqrt{2\pi}$ is evaluated: square the integral, convert to polar coordinates, and the resulting radial integral is elementary.

Cholesky decomposition for correlated normals

In quant finance, the most important change of variables for multiple integrals is the Cholesky decomposition. If

\mathbf{Z} = (Z_1, \ldots, Z_n)

are independent standard normals and

L

is the lower-triangular Cholesky factor of the correlation matrix

\Sigma = LL^T

, then

\mathbf{X} = L\mathbf{Z}

is multivariate normal with correlation

\Sigma

This converts an

n

-dimensional integral over a correlated density into an

n

-dimensional integral over independent standard normals — a form suitable for Monte Carlo simulation. The Jacobian of the transformation is

|\det L| = \prod_i L_{ii}

, but in practice you generate

\mathbf{Z}

and transform rather than computing the integral analytically.

Key results

Covariance as a double integral

For continuous random variables

(X, Y)

with joint density

p(x, y)

\text{Cov}(X, Y) = \iint (x - \mu_X)(y - \mu_Y)\,p(x, y)\,dx\,dy

= \iint x\,y\,p(x, y)\,dx\,dy - \mu_X \mu_Y = \mathbb{E}[XY] - \mathbb{E}[X]\mathbb{E}[Y]

If $X$ and $Y$ are independent, $p(x,y) = p_X(x)p_Y(y)$ , and the double integral factors into a product of single integrals: $\mathbb{E}[XY] = \mathbb{E}[X]\mathbb{E}[Y]$ , so $\text{Cov} = 0$ .

Multivariate normal density

The density of $\mathbf{X} \sim \mathcal{N}(\boldsymbol{\mu}, \Sigma)$ in $\mathbb{R}^n$ is:

p(\mathbf{x}) = \frac{1}{(2\pi)^{n/2}|\Sigma|^{1/2}} \exp\left(-\frac{1}{2}(\mathbf{x} - \boldsymbol{\mu})^T \Sigma^{-1}(\mathbf{x} - \boldsymbol{\mu})\right)

Every probability computed from this density is an $n$ -dimensional integral. The normalisation $\int p\,d\mathbf{x} = 1$ is a standard multivariable Gaussian integral, evaluated by diagonalising $\Sigma$ (change of variables to principal components) and factoring into $n$ independent one-dimensional Gaussian integrals.

Independence ↔ factorisation

$X$ and $Y$ are independent if and only if $p(x, y) = p_X(x)\,p_Y(y)$ . When this holds, any double integral factors:

\iint f(x)\,g(y)\,p(x,y)\,dx\,dy = \left(\int f(x)\,p_X(x)\,dx\right)\left(\int g(y)\,p_Y(y)\,dy\right)

This factorisation is what makes independent models tractable: an $n$ -asset pricing problem reduces to $n$ one-dimensional problems. When assets are correlated, the double (or $n$ -fold) integral does not factor, and the problem is fundamentally harder — this is the "curse of dimensionality."

Examples and applications

Example 1: basket option pricing

A basket call on two assets with weights $w_1, w_2$ and joint log-normal distribution:

V = e^{-rT}\iint (w_1 s_1 + w_2 s_2 - K)^+ \, p(s_1, s_2)\,ds_1\,ds_2

The sum

w_1 S_1 + w_2 S_2

is not log-normal even if

S_1

and

S_2

are individually log-normal (sums of log-normals are not log-normal — see Log-Normal Distribution). This integral has no closed-form solution. Practical approaches include moment-matching approximations (approximate the basket as a single log-normal), Monte Carlo with Cholesky-generated correlated normals, and numerical quadrature.

Example 2: bivariate normal probability

The probability that both $X > a$ and $Y > b$ where $(X, Y)$ is bivariate standard normal with correlation $\rho$ :

\mathbb{P}(X > a, Y > b) = \int_a^{\infty}\int_b^{\infty} \frac{1}{2\pi\sqrt{1-\rho^2}}\exp\left(-\frac{x^2 - 2\rho xy + y^2}{2(1-\rho^2)}\right)dx\,dy

This is the bivariate normal CDF, often denoted $M(a, b; \rho)$ or $\Phi_2(a, b; \rho)$ . It appears in pricing compound options, options on the best/worst of two assets, and default correlation models. It must be computed numerically (Drezner-Wesolowsky algorithm or Gauss-Hermite quadrature).

Example 3: computing the variance of a portfolio

For a two-asset portfolio with returns $R_p = w_1 R_1 + w_2 R_2$ :

\text{Var}(R_p) = w_1^2\sigma_1^2 + w_2^2\sigma_2^2 + 2w_1 w_2\,\text{Cov}(R_1, R_2)

The covariance is the double integral $\text{Cov}(R_1, R_2) = \iint r_1 r_2\,p(r_1, r_2)\,dr_1\,dr_2 - \mu_1\mu_2$ . Under the multivariate normal assumption, this integral evaluates to $\rho\sigma_1\sigma_2$ , giving the familiar formula $\text{Var}(R_p) = \mathbf{w}^T\Sigma\mathbf{w}$ .

The curse of dimensionality

As the number of dimensions $n$ grows, numerical evaluation of multiple integrals becomes exponentially harder for grid-based methods (trapezoidal, Simpson's): if you use $m$ points per dimension, the total number of evaluations is $m^n$ . For $n = 50$ assets and $m = 100$ points, this is $100^{50}$ — astronomically impossible.

This is the curse of dimensionality, and it is the reason Monte Carlo integration dominates multi-asset pricing. Monte Carlo convergence rate is

O(1/\sqrt{N})

regardless of dimension — 10,000 samples give roughly 1% error whether you have 2 assets or 200. The price is slow convergence, but it is the only feasible approach for high-dimensional integrals.

Variance reduction techniques (antithetic variates, control variates, importance sampling) improve the constant factor without changing the $O(1/\sqrt{N})$ rate.

Common confusions and pitfalls

Forgetting the Jacobian when changing variables. The factor

|\det J|

is essential. Omitting it when converting from Cartesian to polar coordinates (missing the

r

factor) or when transforming correlated normals gives wrong results.

Swapping integration order without checking Fubini's conditions. Fubini's theorem requires absolute integrability. If the integral is only conditionally convergent, swapping the order can change the value. In quant finance, this is rarely an issue (pricing integrands are typically non-negative), but it matters for signed integrands like P&L distributions.

Assuming the basket distribution is log-normal. The sum of correlated log-normals is not log-normal. Treating it as such (e.g., using Black-Scholes with a "basket volatility") introduces approximation error that can be significant, especially for options near the money or with high correlation.

Where this goes next

Multiple integrals connect to:

Numerical Integration: Monte Carlo and quasi-Monte Carlo for high-dimensional integrals.
Change of Variables: The Jacobian-based substitution formula and Cholesky decomposition for correlated normals.
Improper Integrals: Multi-asset expected values are improper integrals over $\mathbb{R}^n$ ; convergence depends on joint tail behaviour.
Partial Derivatives: Multi-asset Greeks (cross-gammas, correlation sensitivities) are partial derivatives of the multiple integral pricing function.
Normal Distribution: The multivariate normal density and the bivariate normal CDF are the workhorses of multi-asset Gaussian models.

References

Stewart, J. (2008). Single Variable Calculus: Early Transcendentals (6th ed.). Thomson Brooks/Cole. The uploaded resource is single-variable; this note extends the same Riemann-sum construction to multivariable integrals and should be grounded in a multivariable calculus source when available.