Stochastic Processes — Brownian Motion and Martingales

Brownian motion is the canonical continuous-time stochastic process with independent, stationary, Gaussian increments and continuous paths; a martingale is a process whose conditional expected future value equals its present value. Key result: Brownian motion is the prototypical martingale, and the two notions are linked by Lévy's characterization (a continuous martingale with quadratic variation

t

is Brownian motion) and by the Itô stochastic calculus. Applies to: mathematical finance, statistical physics, PDE theory, filtering, population genetics. Originated with Brown (1827), Einstein (1905), Bachelier (1900), formalized by Wiener (1923), Lévy, Doob, and Itô (1940s–50s).

A stochastic process is a family ${X_{t}}_{t \geq 0}$ of random variables on a common probability space, modeling a quantity evolving randomly in time. Brownian motion (the Wiener process) is the unique—up to scaling—process with continuous sample paths, independent stationary increments, and $W_{t} - W_{s} \sim N (0, t - s)$ . A martingale $M_{t}$ is a process adapted to a filtration with $E [M_{t} ∣ F_{s}] = M_{s}$ , encoding the idea of a "fair game." These objects matter because they form the foundation of continuous-time probability, the Itô integral, and modern finance. The surprising fact: Brownian paths are continuous everywhere yet differentiable nowhere, have infinite total variation, yet possess finite, deterministic quadratic variation equal to elapsed time.

The Core Insight

Randomness accumulated continuously over time, when made to have independent increments of variance proportional to elapsed time, forces a Gaussian limit (via the CLT) and a path so rough it has no derivative but a perfectly predictable "energy" $⟨ W ⟩_{t} = t$ . The martingale property—conditional fairness—is what allows this roughness to be tamed: stochastic integrals against martingales are again martingales, and quadratic variation replaces the chain rule, giving the Itô formula. Brownian motion is simultaneously a Markov process, a Gaussian process, and a continuous martingale; these three identities, when forced to coexist, pin it down uniquely.

Formal Statement

Filtered Probability Space

A filtered probability space $(Ω, F, {F_{t}}_{t \geq 0}, P)$ consists of a probability space together with a filtration: an increasing family of $σ$ -algebras $F_{s} \subseteq F_{t} \subseteq F$ for $s \leq t$ . It satisfies the usual conditions if it is right-continuous ( $F_{t} = ⋂_{u > t} F_{u}$ ) and $F_{0}$ contains all $P$ -null sets. A process $X$ is adapted if $X_{t}$ is $F_{t}$ -measurable for all $t$ .

Brownian Motion

A real-valued process ${W_{t}}_{t \geq 0}$ is a standard Brownian motion (Wiener process) if:

$W_{0} = 0$ almost surely;
independent increments: for $0 \leq t_{0} < t_{1} < \dots < t_{n}$ , the increments $W_{t_{i}} - W_{t_{i - 1}}$ are independent;
stationary Gaussian increments: $W_{t} - W_{s} \sim N (0, t - s)$ for $s < t$ ;
continuous paths: $t \mapsto W_{t}$ is continuous almost surely.

P (W_{t} - W_{s} \in d x) = \frac{1}{2 π ( t - s )} exp (- \frac{x ^{2}}{2 ( t - s )}) d x

⟳this is the transition density; it solves the heat equation

\partial_{t} p = \frac{1}{2} \partial_{xx} p

with

p (0, \cdot) = δ_{0}

, encoding the Gaussian increment law.

Martingale, Submartingale, Supermartingale

An adapted, integrable ( $E ∣ M_{t} ∣ < \infty$ ) process is a martingale if

E [M_{t} ∣ F_{s}] = M_{s} for all s \leq t .

⟳replace "

=

" with "

\geq

" for a submartingale and "

\leq

" for a supermartingale; the equality is the tower property applied to a fair-game increment with zero conditional mean.

The defining relation says the best forecast of $M_{t}$ given information up to $s$ is the current value $M_{s}$ .

Brownian Motion is a Martingale; Associated Martingales

Let $W$ be a standard Brownian motion with its natural filtration. Then each of the following is a martingale:

W_{t}, W_{t}^{2} - t, exp (λ W_{t} - \frac{1}{2} λ^{2} t) (λ \in R) .

⟳for

W_{t}

, use independent increments and zero mean; for

W_{t}^{2} - t

, expand

W_{t}^{2} = (W_{t} - W_{s})^{2} + 2 W_{s} (W_{t} - W_{s}) + W_{s}^{2}

and take conditional expectation; for the exponential, use the Gaussian MGF

E [e^{λ (W_{t} - W_{s})} ∣ F_{s}] = e^{λ^{2} (t - s) /2}

Lévy's Characterization

Let $M$ be a continuous local martingale with $M_{0} = 0$ and quadratic variation $⟨ M ⟩_{t} = t$ . Then $M$ is a standard Brownian motion.

⟨ M ⟩_{t} = t and M cont. local mart. ⟹ M_{t} = d W_{t}

⟳apply Itô's formula to

exp (iλ M_{t} + \frac{1}{2} λ^{2} ⟨ M ⟩_{t})

; the stochastic integral term is a martingale, forcing the conditional characteristic function to be that of

N (0, t - s)

, the law of an increment.

Why It Holds / Derivation

Quadratic Variation of Brownian Motion

We show that over $[0, t]$ with partition $Π_{n} = {0 = t_{0} < \dots < t_{n} = t}$ of mesh $∥ Π_{n} ∥ \to 0$ , the sums of squared increments converge to $t$ .

Step 1. Define

Q_{n} = \sum_{i = 1}^{n} (W_{t_{i}} - W_{t_{i - 1}})^{2}

. By stationary increments,

E [(W_{t_{i}} - W_{t_{i - 1}})^{2}] = t_{i} - t_{i - 1}

, so

E [Q_{n}] = i = 1 \sum n (t_{i} - t_{i - 1}) = t .

⟳variance of a centered Gaussian increment equals its time span; sum telescopes.

Step 2. Increments are independent, so the variance of

Q_{n}

adds:

Var (Q_{n}) = i = 1 \sum n Var ((W_{t_{i}} - W_{t_{i - 1}})^{2}) = 2 i = 1 \sum n (t_{i} - t_{i - 1})^{2} .

⟳for

Z \sim N (0, σ^{2})

Var (Z^{2}) = 2 σ^{4}

; here

σ^{2} = t_{i} - t_{i - 1}

Step 3. Bound the sum by the mesh:

Var (Q_{n}) = 2 i = 1 \sum n (t_{i} - t_{i - 1})^{2} \leq 2 ∥ Π_{n} ∥ i = 1 \sum n (t_{i} - t_{i - 1}) = 2 ∥ Π_{n} ∥ t \to 0.

⟳factor out the largest interval; the remaining sum telescopes to

t

Step 4. Since

E [Q_{n}] = t

and

Var (Q_{n}) \to 0

, Chebyshev gives

Q_{n} \to t

L^{2}

(and in probability). For nested partitions or summable meshes, the convergence is almost sure. Hence

⟨ W ⟩_{t} = t

■

$W_t^2 - t$ is a Martingale

Step 1. For

s < t

write

W_{t} = W_{s} + (W_{t} - W_{s})

, so

W_{t}^{2} = W_{s}^{2} + 2 W_{s} (W_{t} - W_{s}) + (W_{t} - W_{s})^{2} .

⟳binomial expansion of

(W_{s} + Δ)^{2}

Step 2. Take

E [\cdot ∣ F_{s}]

. Since

W_{s}

F_{s}

-measurable and the increment is independent of

F_{s}

with mean

0

and variance

t - s

E [W_{t}^{2} ∣ F_{s}] = W_{s}^{2} + 2 W_{s} \cdot 0 + (t - s) = W_{s}^{2} + (t - s) .

⟳pull out

F_{s}

-measurable factors; use

E [Δ ∣ F_{s}] = 0

and

E [Δ^{2} ∣ F_{s}] = t - s

Step 3. Subtract

t

from both sides:

E [W_{t}^{2} - t ∣ F_{s}] = W_{s}^{2} + (t - s) - t = W_{s}^{2} - s .

⟳rearrange; the result is the value at

s

, confirming the martingale property.

■

Itô's Formula

Let $f \in C^{2} (R)$ and $W$ a Brownian motion. Then a.s.,

f (W_{t}) = f (W_{0}) + \int_{0}^{t} f^{'} (W_{s}) d W_{s} + \frac{1}{2} \int_{0}^{t} f^{''} (W_{s}) d s .

⟳Taylor-expand

f (W_{t_{i}}) - f (W_{t_{i - 1}})

to second order; the first-order terms build the Itô integral, while

\sum f^{''} (\cdot) (W_{t_{i}} - W_{t_{i - 1}})^{2} \to \int f^{''} d ⟨ W ⟩ = \int f^{''} d s

because

(d W)^{2} = d t

Brownian sample path with its variance envelope $\pm\sqrt{t}$

Norbert Wiener | 1894–1964

Wiener gave the first rigorous mathematical construction of Brownian motion (1923), proving that a probability measure—the Wiener measure—exists on the space $C [0, \infty)$ of continuous functions realizing Einstein's physical predictions. He established the continuity and nowhere-differentiability of almost every path, and developed the abstract machinery (the Wiener integral) that anticipated Itô's calculus. Beyond probability, Wiener founded cybernetics, contributed to harmonic analysis (the Paley–Wiener theorems), and shaped wartime work on prediction and filtering. The Wiener process bears his name precisely because he transformed a physical phenomenon into a well-defined object of pure mathematics.

Building Intuition

Picture a drunkard taking $n$ steps of size $\pm 1/ n$ per unit time. After time $t$ , position is a sum of $\sim n t$ tiny independent steps; the CLT makes this Gaussian with variance $t$ . Send $n \to \infty$ : the discrete walk becomes Brownian motion. The variance growing linearly in time, not quadratically, is the signature of diffusion—random cancellation means displacement scales like $t$ , not $t$ .

The martingale property is "no information advantage": knowing the entire past gives no edge in predicting the next increment's direction, because $E [W_{t} - W_{s} ∣ F_{s}] = 0$ . The roughness ( $(d W)^{2} = d t$ ) is what distinguishes stochastic from ordinary calculus: a Brownian path wiggles so much that second-order effects survive in the limit, producing the extra $\frac{1}{2} f^{''}$ term in Itô's formula. Think of it as a deterministic "drift" generated purely by the path's roughness.

Albert Einstein | 1879–1955

In his 1905 annus mirabilis paper on Brownian motion, Einstein derived the diffusion equation for suspended particles and predicted that the mean-squared displacement grows linearly in time: $⟨ x^{2} ⟩ = 2 D t$ , with $D = k_{B} T / (6 π η r)$ relating the diffusion constant to temperature, viscosity, and particle radius. This gave a route to measuring Avogadro's number, providing decisive evidence for the atomic hypothesis (confirmed experimentally by Perrin, Nobel 1926). Einstein's physical reasoning—random molecular bombardment producing a Gaussian spreading—is exactly mirrored by the mathematical transition density $[1]$ .

Worked Examples

Hitting Time and Optional Stopping

Given: Standard Brownian motion

W

starting at

0

. Let

a > 0

b > 0

, and

τ = in f {t : W_{t} = a or W_{t} = - b}

be the first exit time from

(- b, a)

. Find

P (W_{τ} = a)

. Steps:

$W_{t}$ is a martingale; $τ$ is a stopping time with $E [τ] < \infty$ (since the gap is bounded). By the Optional Stopping Theorem, $E [W_{τ}] = W_{0} = 0$ .
$W_{τ}$ takes value $a$ with probability $p$ and $- b$ with probability $1 - p$ . So

0 = E [W_{τ}] = a p + (- b) (1 - p) = p (a + b) - b .

⟳equate the martingale's expected stopped value to its start, then solve the two-point distribution.

Solve: $p = \frac{b}{a + b}$ .

Boxed answer:

P (W_{τ} = a) = \frac{b}{a + b}

. With

a = 2, b = 3

: probability

= 3/5 = 0.6

(dimensionless).

Expected Exit Time via $W_t^2 - t$

Given: Same setup; compute

E [τ]

, the expected time to exit

(- b, a)

. Steps:

The process $W_{t}^{2} - t$ is a martingale (Eq. $[10]$ ). Optional stopping gives $E [W_{τ}^{2} - τ] = 0$ , hence $E [τ] = E [W_{τ}^{2}]$ .
Using $p = b / (a + b)$ from the previous example:

E [W_{τ}^{2}] = a^{2} \cdot \frac{b}{a + b} + b^{2} \cdot \frac{a}{a + b} = \frac{ab ( a + b )}{a + b} = ab .

⟳weight the squared exit values by their probabilities; factor the common

ab

Therefore $E [τ] = ab$ .

Boxed answer:

E [τ] = ab

. With

a = 2, b = 3

E [τ] = 6

(time units).

Common Misconceptions

Brownian motion is not differentiable. It is a frequent error to write "

d W_{t} / d t

" as if it were an ordinary velocity. Almost every path is nowhere differentiable, has infinite total variation on every interval, and "white noise"

\dot{W}

exists only as a distribution (generalized process). Consequently the Itô integral

\int H d W

is not a Riemann–Stieltjes integral; it must be defined as an

L^{2}

limit of left-endpoint Riemann sums, and the choice of evaluation point matters—the Stratonovich integral (midpoint) obeys the ordinary chain rule but is not a martingale. Optional stopping requires hypotheses. The identity

E [M_{τ}] = E [M_{0}]

can fail without integrability/uniform-integrability conditions or a bounded stopping time. A classic trap:

τ_{a} = in f {t : W_{t} = a}

for

a > 0

has

P (τ_{a} < \infty) = 1

yet

E [τ_{a}] = \infty

, and

E [W_{τ_{a}}] = a \neq = 0 = W_{0}

—optional stopping fails because

{W_{t \land τ_{a}}}

is not uniformly integrable. A martingale need not converge. One might guess every martingale settles down. But

W_{t}

itself is a martingale that oscillates forever: by the law of the iterated logarithm,

lim sup_{t \to \infty} W_{t} / 2 t lo g lo g t = 1

and

lim inf = - 1

a.s., so

W_{t}

does not converge—it visits

+ \infty

and

- \infty

in the limsup sense. Martingale convergence requires an

L^{1}

-bound:

sup_{t} E ∣ M_{t} ∣ < \infty

. Brownian motion has

E ∣ W_{t} ∣ = 2 t / π \to \infty

, so the hypothesis fails.

Connections

Brownian motion and martingales sit at the crossroads of several theories. Via the Feynman–Kac formula, expectations of functionals of $W$ solve parabolic PDEs, linking Partial Differential Equations to probability: $u (t, x) = E [ϕ (W_{t}) ∣ W_{0} = x]$ solves the heat equation. The Itô integral and Itô's formula generate the theory of Stochastic Differential Equations, whose solutions are diffusions. The Martingale Representation Theorem states every $L^{2}$ Brownian martingale is a stochastic integral $\int H d W$ , underpinning Mathematical Finance and market completeness. Lévy's characterization $[4]$ shows Brownian motion is the only continuous martingale with quadratic variation $t$ , and the Dambis–Dubins–Schwarz theorem reveals every continuous martingale is a time-changed Brownian motion. The increment structure makes $W$ a Markov process and a Gaussian process simultaneously, connecting to Markov Processes and Gaussian field theory.

Real-World Applications

Black–Scholes option pricing (finance): asset prices modeled as geometric Brownian motion $d S = μ S d t + σ S d W$ ; risk-neutral pricing exploits the martingale property under an equivalent measure (Girsanov).
Statistical mechanics (physics): diffusion of particles, the Langevin equation, and the fluctuation–dissipation theorem all rest on Brownian dynamics (Einstein–Smoluchowski).
Kalman–Bucy filtering (engineering): continuous-time signal estimation in noise uses martingale innovations and stochastic differential equations.
Population genetics (biology): the Wright–Fisher diffusion and Kimura's models describe allele-frequency drift as a bounded diffusion process.
Neuroscience: integrate-and-fire neuron models use Brownian motion with drift to model membrane potential crossing a firing threshold (first-passage problems).
Queueing and heavy-traffic limits (operations research): reflected Brownian motion approximates congested queues under diffusion scaling.
Climate and physical oceanography: stochastic energy-balance models add Brownian forcing to capture internal variability.

Historical Development

1827 — Robert Brown observes erratic motion of pollen grains suspended in water under a microscope, describing the phenomenon now bearing his name.
1900 — Louis Bachelier, in Théorie de la spéculation, models stock prices with Brownian motion—five years before Einstein—introducing the heat-equation connection.
1905 — Albert Einstein derives the diffusion law $⟨ x^{2} ⟩ = 2 D t$ , linking microscopic molecular motion to macroscopic diffusion and Avogadro's number.
1906 — Marian Smoluchowski independently develops a kinetic theory of Brownian motion.
1908 — Paul Langevin introduces the stochastic differential equation $m \overset{x}{¨} = - γ \overset{x}{˙} + ξ (t)$ , the first SDE in physics.
1923 — Norbert Wiener constructs Brownian motion rigorously as a measure on continuous-function space, proving path continuity and nowhere-differentiability.
1933–1939 — Paul Lévy develops fine path properties (law of iterated logarithm, local time, characterization theorem) and the modulus of continuity.
1940s — Joseph Doob systematizes martingale theory: the convergence theorem, optional stopping, and maximal inequalities, providing the abstract backbone.
1944–1951 — Kiyosi Itô defines the stochastic integral and proves Itô's formula, founding stochastic calculus and the theory of diffusion processes.

Subtleties & Limits

While Brownian motion itself is fully understood, surrounding questions remain deep and active. The construction and uniqueness of solutions to SDEs with irregular (merely measurable or distributional) coefficients—regularization by noise—is an ongoing program (Davie, Krylov–Röckner, and the rough-paths/regularity-structures approaches of Lyons and Hairer). The multidimensional and infinite-dimensional analogues—stochastic PDEs driven by space-time white noise—require renormalization, as in Hairer's Fields-medal-winning theory of regularity structures resolving the KPZ and $Φ_{3}^{4}$ equations. Fine geometric properties of the Brownian zero set, the exact Hausdorff measure of paths, and intersection local times in critical dimensions continue to be studied. In finance, rough volatility models (Gatheral–Jaisson–Rosenbaum) replace Brownian volatility with fractional Brownian motion of Hurst index $H \approx 0.1$ , raising new questions about non-Markovian, non-semimartingale calculus where Itô's formula no longer directly applies.