Diffusion

The unifying principle

Start from a random walker. In time $\Delta t$ it steps $\pm\Delta x$ with equal probability. The probability density $p(x,t)$ obeys, in the continuum limit, the diffusion equation:

$\frac{\partial p}{\partial t} = D\,\frac{\partial^2 p}{\partial x^2}, \qquad D = \lim \frac{(\Delta x)^2}{2\,\Delta t}$

Two facts fall out and reappear everywhere:

• The spread grows as the square root of time: $\langle x^2\rangle = 2Dt$, so $\sigma \propto \sqrt{t}$. Diffusion is slow over distance.
• Fick's / Fourier's law: flux is proportional to the gradient, $J = -D\,\nabla p$. Combined with conservation $\partial_t p + \nabla\!\cdot J = 0$, you recover the same PDE.

The microscopic driver is the stochastic differential equation (Langevin form):

$dx = -\nabla U(x)\,dt + \sqrt{2D}\; dW_t$

whose density evolves by the Fokker–Planck equation. Every field below is a reading of these same two lines.

How it shows up in each field

Physics — heat and Brownian motion

Heat conduction is diffusion of thermal energy: Fourier's law $\mathbf{q} = -k\nabla T$ gives $\frac{\partial T}{\partial t} = \alpha\,\nabla^2 T, \qquad \alpha = \frac{k}{\rho c_p}.$ Same equation, $T$ replacing $p$, thermal diffusivity $\alpha$ replacing $D$. Einstein (1905) tied $D$ to the microscopic jitter: $D = \mu k_B T$ (fluctuation–dissipation). Example: a hot poker end; the temperature front reaches distance $L$ in time $t \sim L^2/\alpha$ — double the rod, quadruple the wait.

Chemistry — molecular mixing and reaction fronts

Concentration $c$ obeys Fick's second law $\partial_t c = D\nabla^2 c$. Couple diffusion to reaction and you get reaction–diffusion: $\frac{\partial c}{\partial t} = D\nabla^2 c + R(c).$ Example: a drop of ink in still water spreads with $\sigma\sim\sqrt{2Dt}$ (small-molecule $D\approx 10^{-9}\,\mathrm{m^2/s}$ — hours to cross a cup). With reaction terms, the same equation makes Turing patterns (spots, stripes).

Biology — neuron signaling and cable theory

Ions diffuse across membranes down electrochemical gradients (Nernst–Planck, a drift–diffusion law). Along a dendrite, membrane voltage $V$ obeys the cable equation: $\tau\frac{\partial V}{\partial t} = \lambda^2 \frac{\partial^2 V}{\partial x^2} - V,$ a diffusion equation ($\lambda^2$ plays the role of $D\tau$) with a leak term $-V$. Example: a passive voltage bump spreads and decays over the length constant $\lambda$; this is why axons need active regeneration (spikes) — pure diffusion of charge would die out in millimeters.

AI-ML — generative diffusion models

Add noise to data step by step until it's pure Gaussian — the forward process is exactly the Langevin SDE: $dx = f(x,t)\,dt + g(t)\,dW_t.$ The magic: this SDE has a reverse-time SDE (Anderson) that runs it backward using the score $\nabla_x \log p_t(x)$: $dx = \big[f - g^2\nabla_x\log p_t(x)\big]dt + g(t)\,d\bar W_t.$ A neural net learns the score; sampling = denoising noise back into images. Example: Stable Diffusion / DDPM start from $x_T \sim \mathcal N(0,I)$ and integrate the reverse process to produce a photo — literally reversing diffusion.

Why this bridge matters

• Scaling intuition transfers: the $\sqrt{t}$ law tells a biologist why signaling needs active amplification, a chemist how long mixing takes, an ML engineer why many small denoising steps beat few large ones.
• The gradient view unifies: "flux flows down gradients" is Fourier's law, Fick's law, Nernst–Planck, and score-based sampling (the score $\nabla\log p$ is the gradient that guides flow).
• Reversibility is the big payoff: physics taught that diffusion is entropy-increasing and "irreversible" — yet the SDE is exactly reversible if you know the score. That physics insight (Anderson's reverse SDE, from stochastic control) is precisely what made modern generative AI work.
• Fluctuation–dissipation ($D = \mu k_B T$) is the same bargain everywhere: the noise that mixes you is the noise that lets you explore/generate.

Connections

• 01 Random Walks & Brownian Motion
• 02 Fokker–Planck & Langevin Dynamics
• 03 Heat Equation & Thermal Diffusivity
• 04 Fick's Laws & Reaction–Diffusion
• 05 Cable Theory & Nernst–Planck
• 06 Score-Based Generative Models
• 07 Turing Patterns
• 08 Fluctuation–Dissipation Theorem

#bridge