Classical dynamical systems are usually written as deterministic differential equations, dxdt=f(x,t)\frac{dx}{dt}=f(x,t); once the initial condition is given, the future trajectory is, in principle, uniquely determined. But real systems are often subject to continuous, unpredictable random perturbations, and putting such a perturbation directly into a continuous-time dynamical system gives a stochastic differential equation (SDE).

1. From ODE to SDE

1.1 From deterministic ODEs to stochastic dynamical systems

The simplest idea is: the change over the next small interval = a deterministic trend + a random perturbation. In continuous time this is usually written dXt=μ(Xt,t) dt+σ(Xt,t) dBtdX_t=\mu(X_t,t)\,dt+\sigma(X_t,t)\,dB_t, where μ\mu is called the drift, σ\sigma controls the strength of the diffusion, and BtB_t is Brownian motion.

1.2 Drift and diffusion

μ(Xt,t) dt\mu(X_t,t)\,dt represents the system’s average direction of evolution in the absence of random noise: in dXt=μ dt+σ dBtdX_t=\mu\,dt+\sigma\,dB_t, μ>0\mu>0 means the process drifts upward overall, and μ<0\mu<0 means it drifts downward. σ(Xt,t)dBt\sigma(X_t,t)dB_t represents the random perturbation; because the scale of a Brownian increment is dt\sqrt{dt}, an entirely different order of magnitude from an ordinary differential dtdt, dBtdB_t cannot simply be treated as a “small change” in the ordinary derivative sense.

2. Itô Calculus

2.1 Why ordinary calculus isn’t enough

Brownian motion is almost nowhere differentiable, so dBt/dtdB_t/dt cannot be treated as an ordinary function, which is why a new integration theory had to be developed — the most classic being the Itô integral and the Stratonovich integral — giving rise to stochastic calculus.

2.2 Itô’s formula

The chain rule of ordinary calculus is df(Xt)=f′(Xt)dXtdf(X_t)=f'(X_t)dX_t, but in Itô calculus, because (dBt)2(dB_t)^2 leaves behind a contribution of order dtdt, an extra second-order term appears: the typical one-dimensional form of Itô’s formula is df(Xt)=f′(Xt)dXt+12f′′(Xt)σ2(Xt,t)dtdf(X_t)=f'(X_t)dX_t+\frac12f''(X_t)\sigma^2(X_t,t)dt. This extra term may look like a mere correction, but it is the most central difference between stochastic calculus and ordinary calculus — Section 3.2 uses it directly to derive the analytic solution of an important model.

3. Two Classic Models, Worked Out

3.1 The Ornstein–Uhlenbeck process: mean reversion in concrete numbers

The classic SDE dXt=θ(μ−Xt)dt+σdBtdX_t=\theta(\mu-X_t)dt+\sigma dB_t (with θ>0\theta>0) describes mean reversion: when Xt>μX_t>\mu, the drift pulls it downward, and when Xt<μX_t<\mu, it pushes it upward. This equation has an analytic solution: E[Xt]=μ+(X0−μ)e−θtE[X_t]=\mu+(X_0-\mu)e^{-\theta t} and Var⁡(Xt)=σ22θ(1−e−2θt)\operatorname{Var}(X_t)=\frac{\sigma^2}{2\theta}(1-e^{-2\theta t}), and as t→∞t\to\infty the variance approaches the stationary value σ2/(2θ)\sigma^2/(2\theta). Taking θ=2\theta=2, μ=5\mu=5, σ=1\sigma=1, X0=10X_0=10, the expected value after one year (t=1t=1) is E[X1]=5+5e−2≈5.68E[X_1]=5+5e^{-2}\approx5.68 — already pulled noticeably from the initial 10 toward the mean of 5; the variance is Var⁡(X1)=14(1−e−4)≈0.245\operatorname{Var}(X_1)=\frac14(1-e^{-4})\approx0.245, very close to the long-run stationary variance σ2/(2θ)=0.25\sigma^2/(2\theta)=0.25. The OU process is widely used for velocity models in physics, financial interest rates, and various noise processes that need to be “pulled back after deviating from the mean.”

3.2 Geometric Brownian motion: deriving the analytic solution with Itô’s formula

The classic financial-mathematics model dSt=μSt dt+σSt dBtdS_t=\mu S_t\,dt+\sigma S_t\,dB_t is called geometric Brownian motion (GBM): the size of the random perturbation is proportional to the current value StS_t, making it suitable for describing positive, multiplicative random growth, and it is the foundation of the Black–Scholes model. Applying Itô’s formula to f(x)=ln⁡xf(x)=\ln x: with f′(x)=1/xf'(x)=1/x and f′′(x)=−1/x2f''(x)=-1/x^2, substituting gives d(ln⁡St)=f′(St)dSt+12f′′(St)σ2St2dt=(μ−σ22)dt+σdBtd(\ln S_t)=f'(S_t)dS_t+\frac12f''(S_t)\sigma^2S_t^2dt=\big(\mu-\frac{\sigma^2}2\big)dt+\sigma dB_t — this extra −σ2/2-\sigma^2/2 term is exactly the direct consequence of the second-order correction term from Section 2.2; applying the ordinary chain rule directly would miss it. Integrating gives the analytic solution St=S0exp⁡((μ−σ22)t+σBt)S_t=S_0\exp\big((\mu-\frac{\sigma^2}2)t+\sigma B_t\big), so ln⁡(St/S0)∼N((μ−σ2/2)t, σ2t)\ln(S_t/S_0)\sim N\big((\mu-\sigma^2/2)t,\ \sigma^2t\big). Taking S0=100S_0=100, μ=0.1\mu=0.1, σ=0.2\sigma=0.2, t=1t=1: the median of S1S_1 is 100 e(0.1−0.02)≈108.33100\,e^{(0.1-0.02)}\approx108.33, but the expected value is E[S1]=S0eμt≈110.52E[S_1]=S_0e^{\mu t}\approx110.52 — the median and the expected value are not equal, a direct consequence of the lognormal distribution’s right skew. In practice the two are often conflated: mistaking the “typical scenario” (the median) for the “average return” (the expectation) systematically underestimates the long-run average growth rate.

4. Distributional Evolution and Conceptual Connections

4.1 The Fokker–Planck equation

An SDE describes how a sample path moves randomly; the same random system can also be described from the perspective of “how does the probability density evolve,” which gives the Fokker–Planck equation. A stochastic dynamical system therefore has two complementary views: the SDE looks at a single random trajectory, while the Fokker–Planck equation looks at how the entire probability distribution evolves over time.

4.2 How SDEs relate to earlier concepts

At this point, everything from earlier in the series converges: Brownian motion supplies continuous random noise, the Markov property describes state evolution, martingales supply the conditional-expectation toolkit, Gaussian processes supply an important class of stochastic processes, and stationarity and ergodicity describe long-run statistical properties — an SDE essentially puts all of these concepts into one continuous-time dynamical system, making it the confluence point of the entire series.

5. Summary

The core form of a stochastic differential equation is dXt=μ(Xt,t)dt+σ(Xt,t)dBtdX_t=\mu(X_t,t)dt+\sigma(X_t,t)dB_t, which can be understood as deterministic dynamics plus a continuous random perturbation. The Ornstein–Uhlenbeck process showed how mean reversion can be worked out concretely with an analytic solution, while geometric Brownian motion showed how the second-order correction term in Itô’s formula directly determines that the “median” and the “expected value” — two quantities that look similar but are in fact systematically different — diverge. From here, the natural next steps are the fuller machinery of stochastic calculus, the Fokker–Planck equation, stochastic control, and more general diffusion processes.