Skip to content

Understanding Solutions of Stochastic Differential Equations

Stochastic differential equations (SDEs) describe systems that evolve under both deterministic forces and random fluctuations. They appear throughout physics, finance, biology, and engineering.

Because most SDEs cannot be solved in closed form, the goal of this chapter is to understand what it means to solve an SDE, when analytical solutions exist, and how to recognize the structures that make them tractable.

Learning Goals

After completing this chapter you should be able to:

  • explain what it means to solve an SDE
  • distinguish between explicit pathwise solutions and distributional characterizations
  • recognize structural classes of SDEs
  • understand why transformations are central to solvability
  • identify when analytical methods are unlikely to succeed

The discussion proceeds in four stages:

  1. What a solution means
  2. Structural classification of SDEs
  3. Transformation-based thinking
  4. Limits of closed-form solvability

1. General Form of an SDE

An Itô stochastic differential equation has the form

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

where

Symbol Meaning
\(X_t\) stochastic process
\(\mu(X_t, t)\) drift term
\(\sigma(X_t, t)\) diffusion coefficient
\(W_t\) Brownian motion

The drift describes the deterministic trend, while the diffusion term represents random fluctuations.

2. What Does "Solving" an SDE Mean?

In deterministic calculus we solve for a function \(x(t)\).

For stochastic differential equations the solution is itself a random process.

Types of Solution

There are several distinct senses in which an SDE can be "solved":

Explicit pathwise solution. This is the strongest case. One writes \(X_t\) explicitly in terms of time, the initial condition, and Brownian motion — for example, \(X_t = X_0 + \mu t + \sigma W_t\). Such formulas allow direct simulation and transparent moment computation.

Distributional characterization. Sometimes no simple pathwise formula exists, but the law of \(X_t\) is fully known. Examples include the Gaussian law of the OU process and the log-normal law of GBM. Knowing the distribution suffices for computing option prices, risk measures, and other statistical quantities.

PDE / generator characterization. A diffusion may be analyzed indirectly through the partial differential equations satisfied by conditional expectations. This approach connects SDEs to the Kolmogorov equations and the Feynman–Kac formula.

Numerical solvability. When closed forms are unavailable, the SDE may still be studied effectively through simulation, moment approximations, or PDE solvers.

A solution is often expressed in integral form as

\[ X_t = X_0 + \int_0^t \mu(X_s, s)\,ds + \int_0^t \sigma(X_s, s)\,dW_s \]

Analytical solutions allow us to compute distributions, derive moments, analyze long-term behavior, and benchmark numerical algorithms.

Important

Most SDEs do not admit elementary closed-form pathwise solutions. In many cases one can still characterize the process through its law, transition density, or associated PDE.

3. Types of SDE Structures

Different analytical techniques apply depending on the structure of the equation. Recognizing the structure is the most important step in solving an SDE.

flowchart TD

A["General SDE
dX = μ(X,t) dt + σ(X,t) dW"]

A --> B[Additive noise]
A --> C[Multiplicative noise]
A --> D[Linear SDE]
A --> E[State-dependent diffusion]

B --> B1[Brownian motion with drift]
C --> C1[Geometric Brownian Motion]
D --> D1[Vasicek / OU]
E --> E1[CIR Process]

B1 --> M1[Direct integration]
C1 --> M2[Itô transformation]
D1 --> M3[Integrating factor]
E1 --> M4[Lamperti transform]

4. The Transformation Viewpoint

Many solvable SDEs become manageable only after a suitable change of variables. The central theme is

\[ \text{complicated SDE} \;\rightarrow\; \text{simpler transformed SDE} \;\rightarrow\; \text{solution or characterization} \]

We do not usually solve difficult SDEs by brute force, but by transforming them into forms we already understand. Typical transformations include:

  • log transforms for multiplicative noise — removes state dependence from the diffusion
  • integrating factors for linear drift — cancels the deterministic decay term
  • Lamperti transforms for state-dependent diffusion — normalizes the diffusion coefficient to a constant

Although the details differ from model to model, the strategy is the same: identify the equation's structure, apply the appropriate transformation, solve or simplify, then interpret the result in the original variables.

5. Diffusion Model Cheat Sheet

Model SDE Key Property Typical Use
Brownian Motion \(dB_t = dW_t\) pure randomness building block
BM with Drift \(dX_t = \mu\,dt + \sigma\,dW_t\) additive noise physical diffusion
GBM \(dS_t = \mu S_t\,dt + \sigma S_t\,dW_t\) log-normal stock prices
Vasicek / OU \(dr_t = a(\theta - r_t)\,dt + \sigma\,dW_t\) Gaussian, mean-reverting interest rates
CIR \(dr_t = a(\theta - r_t)\,dt + \sigma\sqrt{r_t}\,dW_t\) non-negative short-rate models
Heston \(dv_t = \kappa(\bar{v} - v_t)\,dt + \xi\sqrt{v_t}\,dW_t\) stochastic volatility option pricing

Heston Notation

The Heston variance SDE uses the conventional parameters \(\kappa\) (mean-reversion speed), \(\bar{v}\) (long-run variance), and \(\xi\) (vol of vol) to distinguish it from the CIR/Vasicek notation \(a\), \(\theta\), \(\sigma\) used elsewhere in this chapter.

Key Structural Differences

Structure Example Model Key Feature
Additive noise Brownian motion with drift volatility independent of state
Multiplicative noise GBM volatility proportional to state
Mean reversion Vasicek process pulled toward long-run mean
Square-root diffusion CIR helps enforce non-negativity
Stochastic volatility Heston volatility itself follows an SDE

Quick Method Reference

Model Main Technique
Brownian motion with drift direct integration
GBM Itô transformation (log)
Vasicek / OU integrating factor
CIR Lamperti transform; Bessel-type connection

6. Analytical Tools for Studying SDEs

Some techniques do not directly produce an explicit pathwise solution, but still provide powerful analytical information.

Martingale Methods

If one can identify a function \(g(X_t,t)\) such that \(M_t = g(X_t,t)\) is a martingale, then conditional expectations can be analyzed through the generator of the diffusion. This viewpoint leads naturally to the Kolmogorov backward equation.

Feynman–Kac Formula

For an SDE \(dX_t = b(X_t)\,dt + \sigma(X_t)\,dW_t\), quantities of the form

\[ u(t, x) = \mathbb{E}[\phi(X_T) \mid X_t = x] \]

satisfy a backward PDE

\[ u_t + b(x)\,u_x + \frac{1}{2}\sigma^2(x)\,u_{xx} = 0 \]

This creates a deep connection between SDEs and PDEs, and is the foundation of the Black–Scholes equation in mathematical finance.

Girsanov's Theorem

Girsanov's theorem allows one to change the drift of a stochastic process by changing the probability measure. This is a fundamental tool in risk-neutral pricing: under the physical measure \(\mathbb{P}\) the drift involves the true return \(\mu\), while under the risk-neutral measure \(\mathbb{Q}\) it involves the risk-free rate \(r\).

7. When Closed-Form Solutions Do Not Exist

Most SDEs cannot be solved analytically in an elementary pathwise sense. Examples include nonlinear stochastic volatility models, SABR-type models, multi-factor interest rate models, jump-diffusions, and coupled nonlinear systems.

Even when a simple explicit pathwise formula is unavailable, one may still have access to transition densities, characteristic functions, moment equations, PDE representations, and simulation methods. In all these cases the structural classification of §3 remains the starting point: knowing the equation's form determines which analytical or numerical tool to reach for.

8. Decision Framework

A practical workflow when encountering a new SDE:

flowchart TD

A[Identify SDE structure]

A --> B{Additive noise?}
B -->|Yes| C[Direct integration]

B -->|No| D{Multiplicative noise?}
D -->|Yes| E[Log / Itô transform]

D -->|No| F{Linear equation?}
F -->|Yes| G[Integrating factor]

F -->|No| H{State-dependent diffusion?}
H -->|Yes| I[Lamperti transform]

H -->|No| J[Use analytical approximations or numerical methods]

Key Takeaways

  • Only a small class of SDEs admit elementary closed-form pathwise solutions
  • Many solvable models rely on transformations
  • Structural recognition is the first and most important step
  • When common transformations fail, one usually turns to PDE methods, transform methods, or numerical simulation

9. Bridge to Solution Techniques

In this chapter we focused on conceptual foundations: what it means to solve an SDE, the four senses of solvability, and the structural patterns that determine which techniques apply.

In the next chapter we turn to the main solution techniques themselves: direct integration, Itô transformations, integrating factors, Lamperti transforms, and the practical workflow for applying them to classical solvable models.

Exercises

Exercise 1. For each of the following, state which type of solution is being described: explicit pathwise solution, distributional characterization, PDE/generator characterization, or numerical solvability.

(a) \(S_t = S_0 \exp[(\mu - \sigma^2/2)t + \sigma W_t]\)

(b) The transition density of \(r_t\) satisfies the Fokker–Planck equation.

(c) \(r_t\) follows a noncentral chi-squared distribution with known parameters.

(d) Euler–Maruyama simulation with \(10^6\) sample paths gives \(\widehat{\mathbb{E}}[X_T] = 3.14 \pm 0.02\).

Solution to Exercise 1

(a) \(S_t = S_0 \exp[(\mu - \sigma^2/2)t + \sigma W_t]\) — this is an explicit pathwise solution. The process is written as a deterministic function of time, the initial condition, and the Brownian motion path.

(b) The transition density of \(r_t\) satisfies the Fokker-Planck equation — this is a PDE/generator characterization. The process is described indirectly through the partial differential equation governing its probability density.

(c) \(r_t\) follows a noncentral chi-squared distribution with known parameters — this is a distributional characterization. The full law of \(r_t\) is specified, even though no simple pathwise formula may exist.

(d) Euler-Maruyama simulation with \(10^6\) sample paths gives \(\widehat{\mathbb{E}}[X_T] = 3.14 \pm 0.02\) — this is numerical solvability. The SDE is studied through Monte Carlo simulation rather than analytical formulas.

Exercise 2. For the SDE \(dX_t = \alpha X_t(1 - X_t)\,dt + \sigma X_t(1 - X_t)\,dW_t\):

(a) Classify the noise structure (additive, multiplicative, or state-dependent diffusion).

(b) Suggest a transformation that might simplify the diffusion coefficient.

(c) Is a closed-form pathwise solution likely? Justify your reasoning.

Solution to Exercise 2

(a) The diffusion coefficient is \(\sigma(X_t) = \sigma X_t(1 - X_t)\), which depends on the state \(X_t\) in a nonlinear way. This is state-dependent diffusion (neither purely additive nor simply multiplicative in the standard sense).

(b) The Lamperti transform \(h(x) = \int^x \frac{1}{\sigma s(1-s)}\,ds\) would normalize the diffusion to a constant. Using partial fractions:

\[ \frac{1}{s(1-s)} = \frac{1}{s} + \frac{1}{1-s} \]

so \(h(x) = \frac{1}{\sigma}\ln\frac{x}{1-x}\), i.e., the logit transform \(Y_t = \frac{1}{\sigma}\ln\frac{X_t}{1-X_t}\) would make the diffusion coefficient constant.

(c) A closed-form pathwise solution is unlikely. The drift \(\alpha X_t(1 - X_t)\) is nonlinear (logistic growth), and after the Lamperti transform the resulting drift in \(Y_t\) will generally be a complicated nonlinear function of \(Y_t\). Standard transformation techniques are unlikely to reduce this to a directly integrable form. One would typically resort to numerical simulation, PDE methods, or moment analysis.

Exercise 3. Consider the three canonical transformations:

Transformation Target structure
\(Y_t = \log X_t\) ?
\(Y_t = e^{at}(X_t - \theta)\) ?
\(Y_t = \int^{X_t} \frac{dx}{\sigma(x)}\) ?

Fill in the "Target structure" column by describing what type of SDE each transformation is designed to simplify.

Solution to Exercise 3
Transformation Target structure
\(Y_t = \log X_t\) Removes multiplicative noise: if \(dX_t = \mu X_t\,dt + \sigma X_t\,dW_t\), then \(Y_t = \log X_t\) satisfies an SDE with additive (constant) diffusion coefficient, namely \(dY_t = (\mu - \sigma^2/2)\,dt + \sigma\,dW_t\)
\(Y_t = e^{at}(X_t - \theta)\) Eliminates linear mean-reverting drift: if \(dX_t = a(\theta - X_t)\,dt + \sigma\,dW_t\), then \(Y_t\) satisfies a driftless SDE \(dY_t = \sigma e^{at}\,dW_t\) that can be integrated directly
\(Y_t = \int^{X_t} \frac{dx}{\sigma(x)}\) Normalizes state-dependent diffusion to unit (constant) diffusion coefficient — this is the Lamperti transform, converting the SDE to one where only the drift remains state-dependent

Exercise 4. The SABR model for forward rates is

\[ dF_t = \sigma_t F_t^\beta\,dW_t^{(1)}, \qquad d\sigma_t = \alpha\,\sigma_t\,dW_t^{(2)} \]

with \(\langle dW_t^{(1)}, dW_t^{(2)} \rangle = \rho\,dt\).

(a) Does this model admit an elementary closed-form pathwise solution? Why or why not?

(b) Which of the four senses of solvability (pathwise, distributional, PDE, numerical) are available for this model?

Solution to Exercise 4

(a) The SABR model does not admit an elementary closed-form pathwise solution. Several factors prevent this:

  • The forward rate \(F_t\) has state-dependent diffusion \(\sigma_t F_t^\beta\), where the local volatility itself is stochastic.
  • The volatility \(\sigma_t\) follows its own GBM-type SDE.
  • The two Brownian motions are correlated (\(\rho \neq 0\)), creating a coupled two-dimensional system.
  • No single transformation can simultaneously simplify both equations into integrable form.

(b) Available senses of solvability:

  • Pathwise: Not available in elementary form.
  • Distributional: Asymptotic expansions for the implied volatility are available (the Hagan et al. formula), which provide approximate distributional information. The exact transition density is not known in closed form.
  • PDE: The Kolmogorov backward equation for the two-dimensional diffusion \((F_t, \sigma_t)\) can be written down and solved numerically or asymptotically.
  • Numerical: Monte Carlo simulation is straightforward and is the standard approach for pricing under SABR. Euler-Maruyama or more specialized schemes can be applied to the coupled system.

Exercise 5. Explain why the Feynman–Kac formula creates a connection between SDEs and PDEs. Specifically, if \(u(t,x) = \mathbb{E}[\phi(X_T) \mid X_t = x]\) for an SDE with drift \(b(x)\) and diffusion \(\sigma(x)\), write down the PDE that \(u\) satisfies and explain why this is useful when direct simulation is expensive.

Solution to Exercise 5

For an SDE \(dX_t = b(X_t)\,dt + \sigma(X_t)\,dW_t\), consider the function \(u(t, x) = \mathbb{E}[\phi(X_T) \mid X_t = x]\). By the Feynman-Kac formula, \(u\) satisfies the backward PDE

\[ \frac{\partial u}{\partial t} + b(x)\frac{\partial u}{\partial x} + \frac{1}{2}\sigma^2(x)\frac{\partial^2 u}{\partial x^2} = 0 \]

with terminal condition \(u(T, x) = \phi(x)\).

This connection is useful because it provides two complementary approaches to the same problem. When Monte Carlo simulation is expensive (high-dimensional expectations, path-dependent payoffs, need for full solution surface), one can instead solve the PDE numerically using finite difference or finite element methods. Conversely, when PDE methods are expensive (high-dimensional state spaces), Monte Carlo simulation of the SDE provides an efficient alternative. The Feynman-Kac formula guarantees that both approaches compute the same quantity, and it forms the mathematical foundation of the Black-Scholes equation, where the option price satisfies a PDE derived from the underlying GBM dynamics.

Exercise 6. A colleague proposes modeling a stock price with the SDE \(dS_t = \mu\,dt + \sigma\,dW_t\) (Brownian motion with drift) instead of geometric Brownian motion. Identify at least two problems with this choice from a modeling perspective, referring to the structural classification discussed in this chapter.

Solution to Exercise 6

Problem 1: Non-positivity. Brownian motion with drift \(S_t = S_0 + \mu t + \sigma W_t\) is a Gaussian process. Since a Gaussian random variable takes all real values with positive probability, \(S_t < 0\) occurs with positive probability for any \(t > 0\). Stock prices cannot be negative, so this model is structurally inconsistent with the fundamental constraint of limited liability.

Problem 2: Constant absolute volatility vs proportional volatility. In Brownian motion with drift, the diffusion coefficient is the constant \(\sigma\), meaning that a $1000 stock and a $1 stock experience the same absolute random fluctuations. In reality, price fluctuations are roughly proportional to the price level. GBM captures this with multiplicative noise (\(\sigma S_t\,dW_t\)), producing percentage returns that are stationary, which is consistent with empirical observations.

In terms of the structural classification: the additive-noise model uses a constant diffusion coefficient independent of the state, while GBM uses multiplicative noise where the diffusion is proportional to the state. The multiplicative structure is what ensures positivity (through the log-normal distribution) and produces realistic return dynamics.