9.4 Stochastic differential equations

Definition of Stochastic Differential Equations

A stochastic differential equation (SDE) describes a dynamical system where both a deterministic trend and random noise shape the evolution over time. The general form of an SDE is:

$dX_t = \mu(X_t, t)\,dt + \sigma(X_t, t)\,dW_t$

Here, $\mu(X_t, t)$ is the drift coefficient (the deterministic part), $\sigma(X_t, t)$ is the diffusion coefficient (scaling the noise), and $W_t$ is a Brownian motion. The solution $X_t$ is itself a stochastic process, not a single function.

SDEs show up across finance (stock prices), physics (particle diffusion), biology (population dynamics), and engineering. They extend ordinary differential equations by adding a term driven by Brownian motion that captures inherent randomness in the system.

Brownian Motion in Stochastic Calculus

Brownian motion (also called the Wiener process) is the stochastic process that drives most SDEs. A process $W_t$ is a standard Brownian motion if it satisfies:

• $W_0 = 0$
• Continuous sample paths: the trajectory $t \mapsto W_t$ is continuous (no jumps), but it's nowhere differentiable
• Independent increments: for any $0 \le s < t < u < v$, the increments $W_t - W_s$ and $W_v - W_u$ are independent
• Normally distributed increments: $W_t - W_s \sim \mathcal{N}(0, t - s)$

The nowhere-differentiable nature of Brownian paths is exactly why you can't just use ordinary calculus to handle SDEs. The paths are too rough for classical integration, which motivates the Itô integral.

Itô Integral for Stochastic Integrals

Definition of the Itô Integral

The Itô integral extends integration to stochastic processes, letting you integrate a process against Brownian motion:

$\int_0^T X_t \, dW_t$

You build it as a limit of sums, similar in spirit to Riemann sums but with a critical constraint: the integrand $X_t$ must be non-anticipating (also called adapted). This means $X_t$ can only depend on information available up to time $t$, never on future values of $W_t$. This left-endpoint evaluation is what distinguishes the Itô integral from the Stratonovich integral.

Properties of the Itô Integral

• Linearity: $\int_0^T (\alpha X_t + \beta Y_t)\,dW_t = \alpha \int_0^T X_t\,dW_t + \beta \int_0^T Y_t\,dW_t$
• Zero mean: If $X_t$ is adapted and square-integrable, then $\mathbb{E}\left[\int_0^T X_t\,dW_t\right] = 0$. This reflects the fact that Brownian increments have zero mean.
• Itô isometry: This is the workhorse property for computing second moments.

Itô Isometry

For an adapted, square-integrable process $X_t$:

$\mathbb{E}\left[\left(\int_0^T X_t \, dW_t\right)^2\right] = \mathbb{E}\left[\int_0^T X_t^2 \, dt\right]$

This converts a difficult stochastic computation (the squared Itô integral) into an ordinary integral of $X_t^2$. You'll use this constantly when computing variances of SDE solutions.

Itô's Lemma for Stochastic Calculus

Chain Rule in Stochastic Calculus

Itô's lemma is the stochastic version of the chain rule. If $X_t$ satisfies $dX_t = \mu\,dt + \sigma\,dW_t$ and $f(t, x)$ is twice continuously differentiable, then:

$df(t, X_t) = \frac{\partial f}{\partial t}\,dt + \frac{\partial f}{\partial x}\,dX_t + \frac{1}{2}\frac{\partial^2 f}{\partial x^2}\,(dX_t)^2$

Expanding using the rules $dt \cdot dt = 0$, $dt \cdot dW_t = 0$, and $(dW_t)^2 = dt$, you get:

$df = \left(\frac{\partial f}{\partial t} + \mu\frac{\partial f}{\partial x} + \frac{1}{2}\sigma^2\frac{\partial^2 f}{\partial x^2}\right)dt + \sigma\frac{\partial f}{\partial x}\,dW_t$

That extra $\frac{1}{2}\sigma^2 f_{xx}$ term is the key difference from ordinary calculus. It arises because Brownian motion has nonzero quadratic variation: $(dW_t)^2 = dt$ rather than zero.

Applications of Itô's Lemma

A classic application: deriving the solution to geometric Brownian motion (GBM), $dS_t = \mu S_t\,dt + \sigma S_t\,dW_t$. Apply Itô's lemma to $f(S_t) = \ln S_t$:

1. Compute $f'(x) = 1/x$ and $f''(x) = -1/x^2$

2. Apply the formula: $d(\ln S_t) = \left(\mu - \frac{\sigma^2}{2}\right)dt + \sigma\,dW_t$

3. Integrate both sides to get: $S_t = S_0 \exp\left[\left(\mu - \frac{\sigma^2}{2}\right)t + \sigma W_t\right]$

Notice the drift of $\ln S_t$ is $\mu - \sigma^2/2$, not $\mu$. This correction factor is a direct consequence of the extra second-order term in Itô's lemma and trips up many students.

Itô's lemma is also the foundation for deriving the Black-Scholes PDE in financial mathematics.

Existence and Uniqueness of Solutions

Lipschitz Conditions for Existence

Not every SDE has a well-defined solution. The standard existence and uniqueness theorem requires two conditions on the coefficients $\mu(x,t)$ and $\sigma(x,t)$:

• Lipschitz continuity in $x$: There exists a constant $K$ such that for all $x, y$ and $t$: $|\mu(x,t) - \mu(y,t)| + |\sigma(x,t) - \sigma(y,t)| \le K|x - y|$

• Linear growth condition: There exists a constant $C$ such that: $|\mu(x,t)| + |\sigma(x,t)| \le C(1 + |x|)$

The Lipschitz condition prevents the coefficients from changing too abruptly, while the linear growth condition prevents solutions from exploding to infinity in finite time.

Uniqueness of Solutions

Under the conditions above, the SDE has a unique strong solution that is adapted to the filtration generated by $W_t$. Uniqueness here means both pathwise uniqueness (any two solutions with the same Brownian motion and initial condition are indistinguishable) and uniqueness in law (the probability distribution of the solution is unique).

In some cases, weaker conditions suffice. Local Lipschitz continuity (Lipschitz on bounded sets) combined with a growth bound can still guarantee uniqueness, though the solution may only exist up to a stopping time.

Numerical Methods for Stochastic Differential Equations

Since most SDEs don't have closed-form solutions, numerical methods are essential.

Euler-Maruyama Method

This is the simplest and most widely used scheme. For the SDE $dX_t = \mu(X_t)\,dt + \sigma(X_t)\,dW_t$, discretize with time step $\Delta t$:

$X_{n+1} = X_n + \mu(X_n)\,\Delta t + \sigma(X_n)\,\Delta W_n$

where $\Delta W_n \sim \mathcal{N}(0, \Delta t)$. This is the stochastic analogue of the forward Euler method for ODEs.

The Euler-Maruyama method achieves strong convergence of order 0.5 and weak convergence of order 1.0. Strong convergence measures pathwise accuracy; weak convergence measures accuracy of expectations.

Milstein Method

The Milstein method adds a correction term that accounts for the quadratic variation of the diffusion:

$X_{n+1} = X_n + \mu(X_n)\,\Delta t + \sigma(X_n)\,\Delta W_n + \frac{1}{2}\sigma(X_n)\sigma'(X_n)\left[(\Delta W_n)^2 - \Delta t\right]$

where $\sigma'$ denotes the derivative of $\sigma$ with respect to $x$. This achieves strong convergence of order 1.0, a significant improvement over Euler-Maruyama. The trade-off is that you need to compute $\sigma'(x)$, which can be expensive or analytically difficult for complex diffusion coefficients.

Convergence and Stability

Two types of convergence matter for SDE numerics:

• Strong convergence: $\mathbb{E}[|X_N - X(T)|] \to 0$ as $\Delta t \to 0$. Relevant when you care about individual sample paths.
• Weak convergence: $|\mathbb{E}[g(X_N)] - \mathbb{E}[g(X(T))]| \to 0$ for smooth test functions $g$. Relevant when you only need expectations (e.g., option prices).

Stability requires that the numerical scheme doesn't amplify errors or produce unbounded oscillations. For stiff SDEs (where the drift has a large negative eigenvalue), implicit methods like the implicit Euler-Maruyama scheme may be necessary.

Applications of Stochastic Differential Equations

Financial Mathematics and Option Pricing

The Black-Scholes model uses geometric Brownian motion for the asset price:

$dS_t = \mu S_t\,dt + \sigma S_t\,dW_t$

Applying Itô's lemma to the option price as a function of $S_t$ and $t$ leads to the Black-Scholes PDE, whose solution gives the famous option pricing formula. Extensions include stochastic volatility models (e.g., the Heston model, where volatility itself follows an SDE) and interest rate models (e.g., the Vasicek and Cox-Ingersoll-Ross models).

Physics and Statistical Mechanics

The Langevin equation describes a particle subject to friction and random thermal forces:

$m\,dv = -\gamma v\,dt + \sigma\,dW_t$

This connects to the Ornstein-Uhlenbeck process and models Brownian particles in a fluid. In quantum mechanics, stochastic Schrödinger equations incorporate noise and dissipation into quantum evolution.

Biology and Population Dynamics

The stochastic logistic equation adds noise to the classical logistic growth model:

$dN_t = rN_t\left(1 - \frac{N_t}{K}\right)dt + \sigma N_t\,dW_t$

This captures demographic stochasticity (randomness from individual birth/death events) and environmental stochasticity (random fluctuations in growth conditions). SDEs also model gene expression noise, epidemic spread, and neural dynamics.

Relationship to Partial Differential Equations

SDEs and PDEs are deeply connected. This link lets you convert probabilistic problems into analytical ones and vice versa.

Kolmogorov Backward Equation

For the SDE $dX_t = \mu(X_t)\,dt + \sigma(X_t)\,dW_t$, define $u(x,t) = \mathbb{E}[f(X_T) \mid X_t = x]$. Then $u$ satisfies:

$\frac{\partial u}{\partial t} + \mu(x)\frac{\partial u}{\partial x} + \frac{1}{2}\sigma^2(x)\frac{\partial^2 u}{\partial x^2} = 0$

with terminal condition $u(x,T) = f(x)$. The differential operator $\mathcal{L} = \mu \partial_x + \frac{1}{2}\sigma^2 \partial_{xx}$ is the infinitesimal generator of the diffusion process.

Feynman-Kac Formula

The Feynman-Kac formula generalizes this connection. If $u(x,t)$ solves:

$\frac{\partial u}{\partial t} + \mu(x)\frac{\partial u}{\partial x} + \frac{1}{2}\sigma^2(x)\frac{\partial^2 u}{\partial x^2} - r(x)u = 0$

with terminal condition $u(x,T) = f(x)$, then $u$ has the probabilistic representation:

$u(x,t) = \mathbb{E}\left[f(X_T)\exp\left(-\int_t^T r(X_s)\,ds\right) \,\bigg|\, X_t = x\right]$

This is powerful because it lets you solve PDEs using Monte Carlo simulation of the SDE, or conversely, compute expectations by solving a PDE. In finance, the Feynman-Kac formula underpins risk-neutral pricing: the discounted option price satisfies a PDE whose solution equals an expected payoff under the risk-neutral measure.

Advanced Topics in Stochastic Differential Equations

Stochastic Control Theory

Stochastic control optimizes a system governed by an SDE when you can influence the drift or diffusion through a control process. The value function $V(x,t)$ satisfies the Hamilton-Jacobi-Bellman (HJB) equation:

$\frac{\partial V}{\partial t} + \sup_u \left[\mu(x,u)\frac{\partial V}{\partial x} + \frac{1}{2}\sigma^2(x,u)\frac{\partial^2 V}{\partial x^2} + \ell(x,u)\right] = 0$

where $u$ is the control and $\ell$ is the running cost. Applications include Merton's portfolio optimization problem and optimal execution strategies in trading.

Malliavin Calculus

Malliavin calculus provides a differential calculus on Wiener space, allowing you to take derivatives of random variables with respect to the underlying Brownian motion. The Malliavin derivative $D_t F$ measures the sensitivity of a functional $F$ of the Brownian path to perturbations at time $t$.

Key applications include computing Greeks (sensitivities of option prices to parameters) without requiring smoothness of the payoff function, and establishing regularity of probability densities for SDE solutions.

Rough Path Theory

Classical Itô calculus requires the driving process to be a semimartingale. Rough path theory (developed by Terry Lyons) extends this to processes with lower regularity, such as fractional Brownian motion with Hurst parameter $H < 1/2$.

The central idea is to "enhance" the path with additional iterated integral data, which provides enough information to define integrals and solve differential equations pathwise. This framework is useful for modeling systems with long-range dependence and has connections to stochastic PDEs through Martin Hairer's regularity structures.