Stochastic Differential Equations

⚡ Key Takeaways

Stochastic Differential Equations (SDEs) are differential equations driven by random noise, written $dX_t = b(t,X_t)\,dt + \sigma(t,X_t)\,dW_t$ , whose solutions are stochastic processes rather than deterministic functions. The key result is that under Lipschitz and growth conditions, an SDE driven by Brownian motion has a unique strong solution given meaning by the Itô integral. They apply across financial mathematics, statistical physics, filtering, and population dynamics. Foundational work by Itô (1942–1951), building on Wiener, Kolmogorov, and later Stratonovich.

💡 The Core Insight

The "differential" $dX_t$ in an SDE is shorthand for an integral identity. The deterministic chain rule fails because Brownian motion accumulates quadratic variation: $(dW_t)^2 = dt$ at leading order, not $0$ . This single non-classical relation — that the square of the noise increment is of the same order as $dt$ — is the source of every distinctive feature of stochastic calculus, including the Itô correction term, the diffusion equation for the density, and the difference between Itô and Stratonovich interpretations.

DefinitionItô Stochastic Differential Equation

Let $(\Omega,\mathcal F,(\mathcal F_t)_{t\ge0},\mathbb P)$ be a filtered probability space satisfying the usual conditions, and let $W_t$ be an $m$ -dimensional $(\mathcal F_t)$ -Brownian motion. An $n$ -dimensional Itô SDE is the integral relation

1

\quad X_t = X_0 + \int_0^t b(s,X_s)\,ds + \int_0^t \sigma(s,X_s)\,dW_s,

written in differential shorthand as

2

\quad dX_t = b(t,X_t)\,dt + \sigma(t,X_t)\,dW_t,\qquad X_0=\xi.

Here $X_t\in\mathbb R^n$ is the state; $b:[0,\infty)\times\mathbb R^n\to\mathbb R^n$ is the drift coefficient; $\sigma:[0,\infty)\times\mathbb R^n\to\mathbb R^{n\times m}$ is the diffusion (or dispersion) coefficient; $\xi$ is an $\mathcal F_0$ -measurable initial condition; and the second integral is the Itô integral with respect to $W$ . The matrix $a(t,x)=\sigma(t,x)\sigma(t,x)^{\top}\in\mathbb R^{n\times n}$ is the diffusion matrix.

DefinitionThe Itô Integral

For a progressively measurable integrand $H$ with $\mathbb E\int_0^T |H_s|^2\,ds<\infty$ , the Itô integral is the mean-square limit of left-endpoint sums over partitions $0=t_0<\cdots<t_N=t$ :

3

\quad \int_0^t H_s\,dW_s := \lim_{\|\Pi\|\to0}\sum_{k=0}^{N-1} H_{t_k}\,(W_{t_{k+1}}-W_{t_k}),

the limit taken in $L^2(\Omega)$ . The defining property is the Itô isometry:

4

\quad \mathbb E\!\left[\left(\int_0^t H_s\,dW_s\right)^2\right]=\mathbb E\!\int_0^t H_s^2\,ds.

TheoremExistence and Uniqueness — Itô 1951

Suppose $b$ and $\sigma$ satisfy the global Lipschitz and linear growth conditions: there is $K>0$ such that for all $t\le T$ and $x,y\in\mathbb R^n$ ,

5

\quad |b(t,x)-b(t,y)|+\|\sigma(t,x)-\sigma(t,y)\|\le K|x-y|,

6

\quad |b(t,x)|^2+\|\sigma(t,x)\|^2\le K^2(1+|x|^2).

If $\mathbb E|\xi|^2<\infty$ , then [2] has a unique (up to indistinguishability) continuous, $(\mathcal F_t)$ -adapted strong solution $X_t$ on $[0,T]$ with $\mathbb E\sup_{t\le T}|X_t|^2<\infty$ .

TheoremItô's Formula

If $X_t$ solves [2] and $f\in C^{1,2}([0,\infty)\times\mathbb R^n)$ , then $f(t,X_t)$ is again an Itô process and

7

\quad df(t,X_t)=\Big(\partial_t f + \mathcal L f\Big)\,dt + (\nabla_x f)^{\top}\sigma\,dW_t,

where the generator $\mathcal L$ is the second-order operator

8

\quad \mathcal L f = \sum_i b_i\,\partial_{x_i}f + \tfrac12\sum_{i,j} a_{ij}\,\partial_{x_i x_j}f,\qquad a=\sigma\sigma^{\top}.

ProofItô's Formula via Quadratic Variation

Step 1 — Partition and Taylor expand. Fix $t$ and a partition $\{t_k\}$ . Write the telescoping sum

f(t,X_t)-f(0,X_0)=\sum_k\big[f(t_{k+1},X_{t_{k+1}})-f(t_k,X_{t_k})\big].

Expand each increment to second order in $X$ and first order in $t$ :

\Delta f_k=\partial_t f\,\Delta t_k+(\nabla f)^\top\Delta X_k+\tfrac12\Delta X_k^\top(\mathrm{Hess}\,f)\Delta X_k+o(\cdot).

Step 2 — Insert the dynamics. With $\Delta X_k\approx b_k\Delta t_k+\sigma_k\Delta W_k$ , the linear term gives, in the limit, $\int(\nabla f)^\top b\,dt+\int(\nabla f)^\top\sigma\,dW$ — the ordinary terms.

Step 3 — The quadratic term is the crux. Expand

\Delta X_k\Delta X_k^\top=\sigma_k\Delta W_k\Delta W_k^\top\sigma_k^\top+(\text{terms with }\Delta t_k).

Terms containing $\Delta t_k\,\Delta W_k$ or $(\Delta t_k)^2$ vanish in $L^2$ as $\|\Pi\|\to0$ . The leading piece uses the quadratic variation identity

9

\quad \sum_k \Delta W^i_k\,\Delta W^j_k \xrightarrow{L^2} \delta_{ij}\,t,

so $\sum_k\Delta W_k\Delta W_k^\top\to I\,dt$ . Hence $\sum_k\tfrac12\Delta X_k^\top(\mathrm{Hess}\,f)\Delta X_k\to\int\tfrac12\sum_{ij}a_{ij}\partial^2_{ij}f\,dt$ .

Step 4 — Assemble. Collecting Steps 2–3 gives exactly [7]–[8]. $\blacksquare$

∎

KI

Kiyosi Itô

1915–2008

Itô, working in wartime and postwar Japan and partly isolated from Western mathematics, sought to give rigorous meaning to Kolmogorov's diffusion processes at the level of individual sample paths rather than only their transition densities. Between 1942 and 1951 he constructed the stochastic integral against Brownian motion and proved both the existence/uniqueness theorem and the change-of-variables formula now bearing his name. His insight — that one must integrate using non-anticipating (left-endpoint) evaluation to preserve the martingale property — is what makes the Itô integral a centred, computable object. The resulting Itô calculus became the mathematical engine of modern probability, control theory, and quantitative finance.

🧠 Intuition

Think of $X_t$ as the position of a particle being simultaneously pushed by a smooth velocity field $b$ and kicked by a relentless storm of independent random impulses $\sigma\,dW$ . Over a tiny interval $dt$ , the drift moves you a distance $\propto dt$ , but the noise moves you $\propto\sqrt{dt}$ — much larger when $dt$ is small. This asymmetry of scales is everything. Average displacement is set by drift; spread is set by diffusion. Because the random step is $\sqrt{dt}$ , its square is $dt$ , and squares of random steps do not average away — they accumulate deterministically. That accumulation is the quadratic variation, and it is why a curved function $f$ of a noisy variable picks up a systematic bias (the Itô term): random jitter on a convex curve lands you, on average, higher than the curve's value at the mean position — Jensen's inequality made dynamical.

Figure

Sample paths of an SDE: drift sets the average, diffusion sets the spread

ExampleGeometric Brownian Motion — the Black–Scholes asset

Given. $dX_t=\mu X_t\,dt+\sigma X_t\,dW_t$ , $X_0=S_0=100$ , $\mu=0.08\,\text{yr}^{-1}$ , $\sigma=0.20\,\text{yr}^{-1/2}$ , horizon $t=2$ yr.

Step 1 — Apply Itô to $f=\ln x$ . Here $b=\mu x,\ a=\sigma^2x^2$ . With $f'=1/x,\ f''=-1/x^2$ :

d\ln X_t=\Big(\mu-\tfrac12\sigma^2\Big)dt+\sigma\,dW_t.

The $-\tfrac12\sigma^2$ is the Itô correction.

Step 2 — Integrate (this equation has constant coefficients).

\ln X_t-\ln X_0=(\mu-\tfrac12\sigma^2)t+\sigma W_t\ \Rightarrow\ X_t=S_0\exp\!\big[(\mu-\tfrac12\sigma^2)t+\sigma W_t\big].

Step 3 — Plug numbers. $\mu-\tfrac12\sigma^2=0.08-0.02=0.06$ . Mean: $\mathbb E[X_t]=S_0e^{\mu t}=100\,e^{0.16}=117.35$ . Median: $S_0e^{(\mu-\sigma^2/2)t}=100\,e^{0.12}=112.75$ .

\boxed{\mathbb E[X_2]=\$117.35,\qquad \text{median}=\$112.75}

The mean exceeds the median — the log-normal skew, a direct fingerprint of the Itô term.

ExampleOrnstein–Uhlenbeck process — mean-reverting velocity

Given. $dX_t=-\theta X_t\,dt+\sigma\,dW_t$ , $X_0=x_0$ , $\theta=2\,\text{s}^{-1}$ , $\sigma=1\,\text{(units)}\,\text{s}^{-1/2}$ , $x_0=5$ , evaluate at $t=1\,$ s.

Step 1 — Variation of constants. Multiply by integrating factor $e^{\theta t}$ and apply Itô to $e^{\theta t}X_t$ :

d(e^{\theta t}X_t)=\theta e^{\theta t}X_t\,dt+e^{\theta t}dX_t=\sigma e^{\theta t}\,dW_t.

Step 2 — Integrate.

X_t=x_0e^{-\theta t}+\sigma\int_0^t e^{-\theta(t-s)}\,dW_s.

Step 3 — Moments via Itô isometry [4].

\mathbb E[X_t]=x_0e^{-\theta t},\qquad \mathrm{Var}(X_t)=\sigma^2\int_0^t e^{-2\theta(t-s)}ds=\frac{\sigma^2}{2\theta}\big(1-e^{-2\theta t}\big).

Step 4 — Numbers. $\mathbb E[X_1]=5e^{-2}=0.677$ . $\mathrm{Var}(X_1)=\tfrac{1}{4}(1-e^{-4})=0.2454$ , so SD $=0.495$ . Stationary variance $\sigma^2/2\theta=0.25$ .

\boxed{\mathbb E[X_1]=0.677\ \text{units},\quad \mathrm{SD}(X_1)=0.495\ \text{units}}

⚠️ Common mistake

The chain rule does NOT hold. Writing $d(X_t^2)=2X_t\,dX_t$ is wrong. The correct identity from [7] with $f=x^2$ and $dX=\sigma\,dW$ is $d(X_t^2)=2X_t\,dX_t+\sigma^2\,dt$ . Omitting the $\sigma^2\,dt$ term is the single most common error in stochastic calculus. Equivalently: $\int_0^t W_s\,dW_s=\tfrac12 W_t^2-\tfrac12 t$ , NOT $\tfrac12W_t^2$ . The extra $-t/2$ is the price of left-endpoint evaluation.

CounterexampleItô ≠ Stratonovich

Consider the same-looking SDE $dX=X\,dW$ interpreted two ways. The Itô solution is $X_t=X_0\exp(W_t-\tfrac12 t)$ , a martingale with $\mathbb E[X_t]=X_0$ . The Stratonovich solution (denoted $dX=X\circ dW$ ) obeys the ordinary chain rule and gives $X_t=X_0\exp(W_t)$ , with $\mathbb E[X_t]=X_0e^{t/2}\ne X_0$ . The conversion is $b_{\text{Itô}}=b_{\text{Strat}}+\tfrac12\sigma\,\partial_x\sigma$ . So identical symbolic equations describe processes with different drifts, different means, and different physics. One must always declare the calculus.

connection

SDEs are the probabilistic dual of parabolic PDEs. The transition density $p(t,x)$ of [2] solves the Fokker–Planck (Kolmogorov forward) equation $\partial_t p=\mathcal L^* p$ , while expectations $u(t,x)=\mathbb E_x[g(X_t)]$ solve the backward equation $\partial_t u=\mathcal L u$ — this is the Feynman–Kac Formula, linking Partial Differential Equations to path integrals. The Itô integral is the prototypical Martingale, placing SDEs inside Stochastic Processes and Brownian Motion. Numerically, SDEs are solved by the Euler–Maruyama Method and Milstein schemes, the stochastic analogues of Numerical Methods for ODEs. The Girsanov theorem relates SDEs under different drifts via Measure Theory changes of probability.

🔬 Where this appears

⏳ History

- 1827 — Robert Brown observes the erratic motion of pollen grains, the empirical phenomenon later named Brownian motion.

- 1900 — Louis Bachelier, in Théorie de la spéculation, models stock prices with Brownian motion — the first SDE-like financial model, predating Einstein.

- 1905 — Albert Einstein derives the diffusion equation for Brownian particles, linking it to molecular collisions and Avogadro's number.

- 1908 — Paul Langevin writes the first stochastic differential equation of physics for a particle in a fluid, introducing a random force.

- 1923 — Norbert Wiener constructs Brownian motion rigorously as a measure on path space (the Wiener measure).

- 1931 — Andrei Kolmogorov derives the forward and backward equations for Markov diffusion processes (analytic theory).

- 1942–1951 — Kiyosi Itô develops the stochastic integral, the existence/uniqueness theorem, and Itô's formula, founding pathwise stochastic calculus.

- 1960s — Ruslan Stratonovich and D. Fisk introduce the symmetric (midpoint) integral obeying the classical chain rule, clarifying the physics/calculus distinction.

- 1973 — Black, Scholes, and Merton apply Itô calculus to option pricing, triggering the explosion of quantitative finance and the 1997 Nobel Prize.

open-problem

Beyond the Lipschitz regime, the theory becomes delicate and partly open. (1) Irregular coefficients: when $b$ is merely bounded measurable and $\sigma$ uniformly elliptic, strong uniqueness still holds (Veretennikov, Krylov–Röckner), but for $\sigma$ degenerate or non-Lipschitz the picture is incomplete; pathwise uniqueness vs. uniqueness-in-law (Yamada–Watanabe) remains subtle. (2) Rough-path and singular SPDEs: equations driven by noise rougher than Brownian motion (fractional, white-in-space) required entirely new frameworks — Lyons's rough path theory and Hairer's regularity structures (Fields Medal 2014) — and many singular SPDEs (e.g. the KPZ equation, $\Phi^4_3$ ) were given meaning only in the 2010s; their fine structure is active research. (3) Numerical complexity: sharp strong convergence rates for SDEs with non-globally-Lipschitz coefficients can be arbitrarily slow, and the construction of stable schemes (tamed/implicit methods) is ongoing. (4) High-dimensional and McKean–Vlasov / mean-field SDEs, where the coefficients depend on the law of the solution, raise open well-posedness and propagation-of-chaos questions.

Formal Statement

Why It Holds / Derivation

Building Intuition

Worked Examples

Common Misconceptions

Real-World Applications

Historical Development

Subtleties & Limits

See Also