Random Walks & Brownian Motion: From Pollen to Finance

1. The Simplest Random Walk

Imagine standing at position 0 on the number line. At each time step you flip a fair coin: heads → step right (+1), tails → step left (−1). After n steps your position is:

X_n = S₁ + S₂ + \dots + Sₙ where each Sᵢ = +1 or -1 with equal probability E[X_n] = 0 (expected position: right back at the origin) Var(X_n) = n (variance grows linearly) σ(X_n) = \sqrtn (typical displacement \propto \sqrtn, NOT n)

This √n scaling is the central fact: a random walker at step 1,000,000 is typically only ~1,000 steps from the origin. Contrast with a directed walker who would be at position 1,000,000.

The position after many steps follows a normal distribution by the Central Limit Theorem: X_n ~ N(0, n). This emerges beautifully even though each step is a discrete ±1.

2. Brownian Motion

Take the discrete random walk, shrink the step size and time interval simultaneously, and in the limit you get Brownian motion (Wiener process) — a continuous-time stochastic process W(t) with:

Properties of Brownian motion W(t): 1. W(0) = 0 2. Independent increments: W(t)-W(s) is independent of W(r) for r \leq s 3. W(t)-W(s) ~ N(0, t-s) (Gaussian with variance = elapsed time) 4. Continuous paths (but nowhere differentiable!) Scaling: W(t) has typical magnitude \sqrtt ⟨W(t)²⟩ = t

Einstein's 1905 paper showed that the mean-squared displacement of a Brownian particle is:

⟨r²⟩ = 2\cdotd\cdotD\cdott d = number of spatial dimensions D = diffusion coefficient (m²/s) t = time (s) Example: a 1 μm polystyrene sphere in water at 20°C D \approx 4.4 \times 10⁻¹³ m²/s After 1 second: \sqrt⟨r²⟩ \approx \sqrt(2\cdot3\cdot4.4\times10⁻¹³\cdot1) \approx 1.6 μm (about its own diameter)

Historical note: Einstein's Brownian motion paper (1905) provided the first concrete evidence for the existence of atoms — because the diffusion coefficient D is linked to Boltzmann's constant and Avogadro's number. Jean Perrin confirmed this experimentally in 1908, winning the Nobel Prize in 1926.

3. Diffusion & the Heat Equation

There's a deep connection between random walks and the diffusion equation (which is also the heat equation):

\partialP/\partialt = D \cdot \nabla²P P(x, t) = probability density of finding the random walker at position x at time t D = diffusion coefficient Solution for a point source at origin: P(x, t) = (1/\sqrt(4πDt)) \cdot exp(-x²/(4Dt)) This is a Gaussian that spreads as \sqrtt — the same \sqrtn scaling from discrete walks!

This means: (1) heat spreading through a solid, (2) ink diffusing in water, and (3) a random walker's position distribution are all described by the same mathematics. Each molecule of ink is undergoing its own random walk, buffeted by water molecules.

The connection runs deeper: you can solve the heat equation numerically by simulating random walkers (Monte Carlo method). Conversely, you can compute random walk statistics by solving differential equations.

4. Higher Dimensions & Return Probabilities

A stunning result in random walk theory due to Pólya (1921):

Will a random walker ever return to the starting point? 1D: P(return) = 1 (certain — always returns) 2D: P(return) = 1 (certain — "a drunk in a plane always finds home") 3D: P(return) \approx 0.3405 ("a drunk bird may be lost forever") 4D+: P(return) \to 0 (almost certainly never returns) Technically: the walk is "recurrent" in 1D and 2D, "transient" in 3D+.

This has profound implications in physics: defects diffusing on a 2D surface will always eventually meet (and potentially annihilate or react). In 3D, diffusing particles can escape each other permanently — which is why reactions in 3D are diffusion-limited differently than in 2D.

Self-Avoiding Walks

If the walker cannot revisit a site it has already visited, the walk is self-avoiding. This is the model for polymer chains (each monomer occupies a unique position). Self-avoiding walks have different scaling: the end-to-end distance grows as n^ν where ν ≈ 0.588 in 3D (Flory exponent), not 0.5 as in the ordinary walk.

5. Monte Carlo Methods

Random walks are the backbone of Monte Carlo simulation — using randomness to solve deterministic problems:

Monte Carlo integration: Estimate ∫f(x)dx by sampling random points and averaging f. Converges as 1/√N regardless of dimensionality — beats grid methods in >4 dimensions.
Markov Chain Monte Carlo (MCMC): Construct a random walk in the space of possible states such that the walk visits each state proportionally to its probability. The Metropolis-Hastings algorithm (1953/1970) does this for any target distribution. Used in Bayesian statistics, protein folding, and lattice QCD.
Path-integral Monte Carlo: Feynman showed that quantum mechanics can be formulated as a sum over all possible paths (each a random walk in spacetime). Monte Carlo sampling of these paths computes quantum properties of materials.

Random walks in games: The "drunken sailor" random walk underpins gambling theory. The Gambler's Ruin problem — a random walk with absorbing barriers — shows that a player with finite money against an infinite casino will always go broke, regardless of the fairness of the game.

6. Random Walks in Finance

Louis Bachelier's 1900 thesis modelled stock prices as Brownian motion — five years before Einstein's paper. The random walk hypothesis states that price changes are independent and identically distributed — past prices give no information about future prices.

Geometric Brownian Motion (GBM) — the basis of the Black-Scholes model: dS = μ\cdotS\cdotdt + σ\cdotS\cdotdW S = stock price μ = drift (expected return) σ = volatility dW = Wiener process increment Solution: S(t) = S(0) \cdot exp((μ - σ²/2)\cdott + σ\cdotW(t)) This gives log-normal price distributions. Black-Scholes option pricing (1973 Nobel): C = S\cdotN(d₁) - K\cdote⁻ʳᵗ\cdotN(d₂) where d₁,d₂ depend on σ, t, S, K, r

Real markets deviate from GBM: they exhibit fat tails (extreme events are far more common than Gaussian distributions predict), volatility clustering (calm and turbulent periods), and long-range correlations. Lévy flights and fractional Brownian motion are used to capture these features.

7. Broader Applications

Ecology (Lévy flights): Animals like albatrosses follow Lévy flight patterns — mostly short steps with occasional long jumps. This strategy is mathematically optimal for searching sparse, randomly distributed food sources.
Network science (PageRank): Google's original algorithm models a random walker clicking links on the web. A page's importance equals the probability the walker lands on it at steady state — a random walk on a graph.
Polymer physics: A polymer chain in solution behaves as a self-avoiding random walk. The radius of gyration, elasticity, and chain melting all follow from walk statistics.
Bacterial chemotaxis: E. coli navigates by alternating straight "runs" with random "tumbles." Runs last longer when moving up a chemical gradient — a biased random walk that achieves directed motion from random rotations.
Image segmentation: Random walks from labelled seed pixels to unlabelled pixels; each pixel is assigned the label of the seed it's most likely to reach — an elegant solution to the segmentation problem.
Quantum walks: The quantum analogue of classical random walks, where the walker exists in superposition. Quantum walks spread as t (not √t), enabling quadratic speedups in search algorithms (Grover's algorithm can be viewed as a quantum walk).