Probability and statistics

Brownian motion and stochastic calculus

A continuous random process modeling erratic motion, with a calculus of its own used in finance and physics.

IntuitionWhy does a speck of pollen dancing in water need its own kind of calculus?

In 1827 the botanist Robert Brown watched pollen grains suspended in water under a microscope and saw them jitter endlessly, changing direction at every instant with no pattern he could discern. The explanation, only understood decades later, is that each grain is being bombarded from all sides by billions of invisible water molecules per second, and the net random kick at each moment nudges the grain a little further in a completely unpredictable direction. Brownian motion is the mathematical idealization of this endless, directionless jitter: a random path that is continuous — it never jumps — yet so jagged that it has no well-defined velocity at any instant, ever. That single fact, that the path is continuous but nowhere smooth, is what forces mathematicians to build an entirely new calculus, since ordinary calculus assumes you can zoom in on a curve and see it look like a straight line.

At any time

t > 0

, the position

W_t

of a standard Brownian motion has a Gaussian bell-curve density

\mathcal{N}(0, t)

, with

\approx 68\%

probability inside

[-\sqrt{t}, +\sqrt{t}]

The distribution curve above describes where a particle may be at one fixed time. A two-dimensional Brownian motion uses two independent copies of this process; with time as a third coordinate, one sample becomes a jagged space-time path.

One deterministic sample of two independent Brownian coordinates against time. Coordinates x and y show position, and z is time; change the step count to compare reproducible discrete samples at different resolutions.

UndergraduateThe Wiener process: a rigorous definition

Definition: Standard Wiener process (Brownian motion)

A standard Wiener process (or Brownian motion) is a continuous-time stochastic process $(W_t)_{t\ge0}$ such that: (1) $W_0=0$ almost surely; (2) it has independent increments — for any $0\le t_1<t_2<t_3<\cdots$ , the increments $W_{t_2}-W_{t_1},\ W_{t_3}-W_{t_2},\dots$ are mutually independent; (3) each increment is normally distributed, $W_t-W_s\sim N(0,t-s)$ for $s<t$ ; and (4) almost every sample path $t\mapsto W_t$ is continuous. Property (3) alone already forces the process to look identical in distribution at every time scale — this self-similarity is why Brownian motion appears as the universal limit of so many different discrete random walks.

W_t-W_s\sim N(0,t-s)

A defining, almost paradoxical feature of Brownian paths is their quadratic variation: summing the squared increments of $W$ over a fine partition of $[0,t]$ converges (in probability) to $t$ itself, $\sum_{i}(W_{t_{i+1}}-W_{t_i})^2 \to t$ , no matter how fine the partition — the squared wiggles never vanish, they accumulate to exactly $t$ . Compare this to any ordinary differentiable function $f$ , whose quadratic variation over any interval is always $0$ , since its increments shrink like $\Delta t$ rather than $\sqrt{\Delta t}$ . This single difference in scaling — Brownian increments are of order $\sqrt{\Delta t}$ , not $\Delta t$ — is the root cause of everything unusual about stochastic calculus.

\sum_{i}(W_{t_{i+1}}-W_{t_i})^2 \to t

Ordinary calculus vs. stochastic (Itô) calculus
Aspect	Smooth function $f(t)$	Brownian path $W_t$
Differentiability	Differentiable, tangent line exists	Nowhere differentiable, no tangent anywhere
Increment size over $\Delta t$	Order $\Delta t$	Order $\sqrt{\Delta t}$
Quadratic variation on $[0,t]$	$0$	$t$ (never zero)
Chain rule for $F(f(t))$	$dF=F'(f)f'(t)\,dt$	Needs an extra second-derivative term (Itô's lemma)

Theorem: Brownian motion has positive quadratic variation and is nowhere differentiable

For a standard Wiener process $W$ and any $t>0$ , $\sum_{i}(W_{t_{i+1}}-W_{t_i})^2 \to t$ where the sum is over a partition of $[0,t]$ whose mesh tends to $0$ , the limit holding in probability (in fact almost surely along dyadic partitions). Consequently, almost every sample path $s\mapsto W_s$ is differentiable at no point $s\in[0,t]$ .

Why is it true?

This is the theorem that makes stochastic calculus necessary in the first place: it rules out treating $dW_t$ as an ordinary infinitesimal the way $dt$ is treated in classical calculus, and it justifies the heuristic rule $(dW_t)^2=dt$ that drives Itô's lemma below.

Proof

Quadratic variation. Partition $[0,t]$ into $n$ equal pieces of length $t/n$ and let $Q_n=\sum_{i=1}^n(W_{it/n}-W_{(i-1)t/n})^2$ . Each term $(W_{it/n}-W_{(i-1)t/n})^2$ has mean $t/n$ (since the increment is $N(0,t/n)$ ) and, using the fourth moment of a normal variable, variance $2(t/n)^2$ . Summing $n$ independent such terms, $E[Q_n]=t$ exactly and $\mathrm{Var}(Q_n)=2t^2/n\to0$ as $n\to\infty$ , so $Q_n\to t$ in probability (and mean square) — this proves the quadratic variation identity.

Nowhere differentiability. Suppose, for contradiction, that $W$ were differentiable at some point $s$ with derivative $W'(s)=L$ finite. Then for small $h$ , $W_{s+h}-W_s\approx Lh$ , so the squared increment over a tiny interval of length $h$ would be of order $h^2$ — but the quadratic variation computation above shows the typical squared increment over an interval of length $h$ is of order $h$ (much larger than $h^2$ for small $h$ ), and summing these order- $h$ terms over $t/h$ intervals is exactly what produces a total that converges to the finite, positive number $t$ rather than to $0$ .

This mismatch — differentiability would force quadratic variation to be $0$ , but it is provably $t>0$ — is irreconcilable, so no point of differentiability can exist; a full measure-theoretic argument (due to Paley, Wiener and Zygmund in 1933) makes this rigorous by directly bounding the probability that any difference quotient stays bounded near any point, showing this probability is exactly zero simultaneously for every point on the path.

Theorem: Itô's lemma

Let $W_t$ be a standard Wiener process and let $F(t,x)$ be twice continuously differentiable in $x$ and once in $t$ . Then the process $F(t,W_t)$ satisfies $dF=\Big(\partial_tF+\tfrac12\partial_x^2F\Big)dt+\partial_xF\,dW_t$ — a stochastic chain rule with an extra second-derivative ("Itô correction") term compared to the ordinary chain rule.

Why is it true?

This is the single most-used tool in stochastic calculus: it tells you exactly how to differentiate a function of a random path, which is the starting point for deriving the dynamics of any quantity (an option price, a physical observable) that depends on Brownian motion.

Proof

Taylor expand. For a smooth $F$ , the ordinary second-order Taylor expansion in both variables over a small step $\Delta t$ with $\Delta W=W_{t+\Delta t}-W_t$ reads $\Delta F\approx \partial_tF\,\Delta t+\partial_xF\,\Delta W+\tfrac12\partial_x^2F\,(\Delta W)^2+\tfrac12\partial_t^2F\,(\Delta t)^2+\partial_t\partial_xF\,\Delta t\,\Delta W+\cdots$ — this much is pure calculus, valid for any smooth path.

Order the terms by size. Because $\Delta W$ is of order $\sqrt{\Delta t}$ (not $\Delta t$ , as shown in the previous theorem), the terms scale as: $\partial_tF\,\Delta t$ is order $\Delta t$ ; $\partial_xF\,\Delta W$ is order $\sqrt{\Delta t}$ (the dominant, leading-order random term); $(\Delta t)^2$ and $\Delta t\,\Delta W$ are order $(\Delta t)^2$ and $(\Delta t)^{3/2}$ respectively, negligible compared to $\Delta t$ ; but $(\Delta W)^2$ is order $\Delta t$ — the same order as $\partial_tF\,\Delta t$ , not negligible at all, unlike in ordinary calculus where $(\Delta x)^2$ is always negligible next to $\Delta x$ .

Replace $(\Delta W)^2$ by its mean. By the quadratic variation theorem, summed over many small steps $(\Delta W)^2$ behaves like $\Delta t$ (its mean, with fluctuations around that mean vanishing as the steps shrink and are summed), which is the informal justification for the heuristic substitution rule $(dW_t)^2=dt$ in the limit of infinitesimal steps.

Take the limit. Dropping the negligible higher-order terms and substituting $(\Delta W)^2\to dt$ in the limit $\Delta t\to0$ turns the Taylor expansion into the differential form $dF=\Big(\partial_tF+\tfrac12\partial_x^2F\Big)dt+\partial_xF\,dW_t$ exactly as claimed — the $\tfrac12\partial_x^2F\,dt$ term is precisely the contribution that ordinary calculus discards but stochastic calculus must keep.

UndergraduateReal-World Applications and Worked Examples

Beyond describing physical diffusion (pollen in water, heat spreading through a solid, gas molecules mixing), Brownian motion and Itô calculus became the mathematical foundation of modern quantitative finance. A stock price $S_t$ is commonly modeled as geometric Brownian motion, $dS_t=\mu S_t\,dt+\sigma S_t\,dW_t$ , where $\mu$ is the average growth rate and $\sigma$ the volatility. Applying Itô's lemma to the value $V(t,S_t)$ of a financial derivative (like an option) leads to the celebrated Black–Scholes partial differential equation, $\partial_tV+\tfrac12\sigma^2S^2\partial_S^2V+rS\partial_SV-rV=0$ , whose solution gives the fair price of options traded on every major exchange. The same mathematics — Wiener processes and Itô's lemma — also underlies models of neuron membrane potentials in neuroscience, interest-rate models in economics, and turbulent dispersion in fluid dynamics.

Example: Expected value and variance of a Brownian displacement

A particle's position follows a standard Wiener process $W_t$ in one dimension, starting at $W_0=0$ . Find $E[W_5]$ and $\mathrm{Var}(W_5-W_2)$ .

Solution

For the first quantity, use $W_t-W_s\sim N(0,t-s)$ with $s=0$ : $W_5-W_0=W_5$ is $N(0,5)$ , so $E[W_5]=0$ directly — Brownian motion has no drift, so its expected position never moves away from the start.

For the second quantity, apply the same defining property with $s=2$ , $t=5$ : $W_5-W_2\sim N(0,5-2)$ , i.e. $N(0,3)$ .

The variance of a $N(0,\sigma^2)$ random variable is $\sigma^2$ by definition, so reading off the parameter directly, $\mathrm{Var}(W_5-W_2)=3$ .

Notice the variance only depends on the elapsed time $t-s=3$ , not on the starting time $s=2$ itself — this reflects the time-homogeneity built into the definition of the Wiener process.

Example: Applying Itô's lemma to $F(t,x)=x^2$

Use Itô's lemma to find $d(W_t^2)$ , the stochastic differential of the squared Wiener process, and use the result to confirm $E[W_t^2]=t$ .

Solution

Take $F(t,x)=x^2$ , so $\partial_tF=0$ , $\partial_xF=2x$ , $\partial_x^2F=2$ .

Substituting into Itô's lemma $dF=\Big(\partial_tF+\tfrac12\partial_x^2F\Big)dt+\partial_xF\,dW_t$ with these derivatives (evaluated at $x=W_t$ ): $d(W_t^2)=\big(0+\tfrac12\cdot2\big)dt+2W_t\,dW_t$ , which simplifies to $d(W_t^2)=dt+2W_t\,dW_t$ .

Integrating both sides from $0$ to $t$ (using $W_0^2=0$ ): $W_t^2=t+2\int_0^tW_s\,dW_s$ .

Taking expectations, and using the key fact that an Itô integral $\int_0^tW_s\,dW_s$ always has mean zero (it is built from increments independent of the past, so there is no systematic drift to accumulate): $E[W_t^2]=t+2E\big[\int_0^tW_s\,dW_s\big]=t+0=t$ , confirming directly what we already knew from $W_t\sim N(0,t)$ (whose variance is $t$ ), but this time derived purely from the stochastic calculus machinery rather than from the definition.

Common mistake. The single most common error is writing

dW_t

and manipulating it with ordinary calculus rules, in particular treating

(dW_t)^2

as negligible the way

(dt)^2

is negligible — this silently throws away the entire second-derivative term in Itô's lemma and gives wrong answers. A second, subtler error is assuming Brownian motion has bounded variation (finite total "distance traveled") because it is continuous; in fact its total variation is infinite on every interval, even though its quadratic variation is finite — continuity alone says nothing about how wiggly a path can be. Finally, remember Itô's lemma is not unique: a different convention (the Stratonovich integral) gives a chain rule that does look like ordinary calculus, at the cost of losing the martingale property that makes Itô integrals so useful; the two are related by a precise conversion formula, but mixing conventions inside one calculation is a frequent source of sign errors.

Historical note

Robert Brown's 1827 microscope observations went unexplained for nearly 80 years, until Albert Einstein's 1905 paper derived the diffusion equation governing the particle's probability density from molecular kinetic theory, giving physicists the first quantitative, testable theory of the phenomenon (soon confirmed experimentally by Jean Perrin, helping settle the debate over whether atoms were real). The rigorous probabilistic construction of the process itself is due to Norbert Wiener, whose 1923 paper proved that a process with the properties described above actually exists as a well-defined mathematical object with continuous paths — hence the name "Wiener process." The calculus needed to integrate against this wildly irregular path was built by Kiyoshi Itô, whose 1944 paper "Stochastic Integral" defined the stochastic integral $\int H_s\,dW_s$ rigorously (via the isometry $E\Big[\big(\int_0^t H_s\,dW_s\big)^2\Big]=E\Big[\int_0^t H_s^2\,ds\Big]$ ) and derived the change-of-variables formula that bears his name, providing the toolkit that, decades later, Fischer Black, Myron Scholes and Robert Merton would use to derive the option-pricing formula that transformed financial markets.

Kiyoshi Itô

Research frontier as of 2026

Brownian motion sits at the crossroads of several live research fronts. Rough path theory and regularity structures, developed by Martin Hairer (Fields Medal 2014), extend Itô-style calculus to a much broader class of irregular driving signals (rougher than Brownian motion itself), making rigorous sense of stochastic PDEs like the KPZ equation that were previously ill-defined, with ongoing work extending these techniques to increasingly singular equations arising in quantum field theory. In probability theory proper, the Schramm–Loewner evolution (SLE), which describes the scaling limit of 2D random curves (self-avoiding walks, percolation boundaries), continues to reveal deep and only partially understood connections between 2D Brownian-motion-adjacent processes and conformal field theory. On the applied side, rough volatility models — which replace standard Brownian motion in option-pricing models with a rougher fractional variant, motivated by high-frequency market data showing volatility is far less smooth than classical models assume — are an active area of quantitative finance research, and machine-learning-based numerical schemes for high-dimensional stochastic differential equations (deep BSDE solvers) are being developed to price derivatives in dimensions where classical PDE methods become computationally intractable.

For a standard Wiener process, the increment $W_t-W_s$ (for $s<t$ ) is distributed as:

The quadratic variation of Brownian motion on $[0,t]$ is:

Compared to the ordinary chain rule, Itô's lemma $dF=\Big(\partial_tF+\tfrac12\partial_x^2F\Big)dt+\partial_xF\,dW_t$ has an extra term because:

The Black–Scholes PDE for option pricing is derived by applying:

References

Ioannis Karatzas, Steven E. Shreve (1991). Brownian Motion and Stochastic Calculus
Kiyosi Itô (1944). Stochastic Integral
Fischer Black, Myron Scholes (1973). The Pricing of Options and Corporate Liabilities