The Diffusion Equation

Drop a bead of ink into still water and watch the boundary blur. No stirring, no current — yet the dark cloud swells, its edge softening from a sharp line into a smooth halo, until the color fades evenly through the glass. What you are seeing is the collective effect of countless molecular collisions, each nudging an ink molecule a tiny random step. The astonishing fact is that this microscopic chaos obeys a single, clean law at the macroscopic scale: the density $P(\mathbf{x},t)$ of the diffusing stuff changes at a rate set by its own curvature, $\partial_t P = D\,\nabla^2 P$ , with one number $D$ — the diffusion coefficient — fixing the pace. The equation is deceptively simple but structurally deep. Because it can be written as a continuity equation $\partial_t P + \nabla\cdot\mathbf{J}=0$ with the down-gradient current $\mathbf{J}=-D\nabla P$ (Fick's first law), it conserves total probability: diffusion rearranges without creating or destroying. Its fundamental solution — the response to a point release — is a Gaussian that broadens with variance $2Dt$ , so a spreading front advances only as $\sqrt{t}$ , never at constant speed. That $\sqrt{t}$ law is the fingerprint of uncorrelated random stepping, and it is exactly the mean-square displacement $\langle x^2\rangle=2Dt$ that a Brownian particle traces out one trajectory at a time. Diffusion is where reversible microscopic mechanics turns into an irreversible macroscopic arrow. Run the equation backward and it is mathematically ill-posed — high-wavenumber ripples explode — mirroring the monotonic growth of entropy. Within statistical mechanics this equation is a hub: it is the force-free heart of the Fokker–Planck equation, the long-wavelength limit of the master equation for a hopping particle, and the ensemble companion of the Langevin trajectory. Master it and drift–diffusion, Kramers' escape problem, and the whole machinery of stochastic transport open up.

Key takeaways

The diffusion equation $\partial_t P = D\nabla^2 P$ governs how a probability or concentration density spreads under random microscopic motion; the single constant $D$ has units of $(\text{length})^2/\text{time}$ .
It is a conservation law: written as continuity $\partial_t P + \nabla\cdot\mathbf{J}=0$ with the Fick current $\mathbf{J}=-D\nabla P$ , it conserves $\int P\,\mathrm{d}^dx$ .
Its fundamental solution is a Gaussian with variance $2Dt$ , so a point release spreads as $\sqrt{t}$ and the mean-square displacement is $\langle x^2\rangle=2Dt$ (and $2dDt$ in $d$ dimensions) — the diffusive scaling.
Each Fourier mode decays as $e^{-Dk^2t}$ : short wavelengths vanish fastest, which is why diffusion smooths sharp features and why running it backward is ill-posed.
It is the force-free special case of the Fokker–Planck equation; adding drift or a potential generalizes it to advection–diffusion and the Smoluchowski form.
Key idea: uncorrelated random microscopic steps of zero mean and finite variance produce, after coarse-graining, a conserved density whose entire evolution is fixed by one number $D$ — a spreading Gaussian of width $\sqrt{2Dt}$ , slower than ballistic, and irreversible despite reversible microscopics.

The core insight

Uncorrelated random microscopic steps of zero mean and finite variance produce, after coarse-graining in space and time, a conserved density $P(\mathbf{x},t)$ whose evolution is fixed entirely by one number $D$ : $\partial_t P = D\nabla^2 P$ . The response to a point release is a Gaussian of variance $2Dt$ , so the front spreads as $\sqrt{t}$ — the signature of diffusion.

Where this lives today

The diffusion equation is arguably the most reused partial differential equation in science, because the same operator governs any conserved quantity spread by unbiased random motion. It is the heat equation of thermal engineering (with $D$ the thermal diffusivity), the diffusion (zero-drift) part of the Black–Scholes equation in quantitative finance, the model for dopant profiles in semiconductor fabrication (junction depth $\sim\sqrt{Dt}$ ), and the backbone of image-processing "diffusion filters" that blur noise while preserving structure. In modern machine learning, diffusion models are inspired by this mathematics: they begin with a simple forward process that gradually adds Gaussian noise and train a model to approximate the corresponding reverse denoising dynamics, transforming noise into images. Within statistical physics the equation is the gateway to nonequilibrium dynamics: it is the force-free heart of the Fokker–Planck equation, the coarse-grained limit of the master equation, and the ensemble face of the Langevin trajectory. Single-molecule tracking in living cells routinely measures $\langle r^2\rangle$ versus $t$ and reads off $D$ — and departures from the linear law (subdiffusion $\langle r^2\rangle\sim t^\alpha$ with $\alpha<1$ ) now serve as a sensitive probe of crowding and binding inside cells.

The physical picture

Think of the density $P$ as a crowd of independent walkers, each kicked at random by the surrounding molecules. Nothing tells a walker which way to go, yet more walkers sit where the density is higher, so random motion alone produces a net flux from regions of high concentration to regions of low concentration. Count the walkers crossing any plane per unit time and it is proportional to the concentration difference across it — that is Fick's law, $\mathbf{J}=-D\nabla P$ . Add the bookkeeping that walkers are neither born nor lost, and the diffusion equation follows with no further input. The Laplacian $\nabla^2 P$ is a local averaging detector: it measures how much $P$ at a point falls below the average of its neighbors. Diffusion pushes every point toward that local average, so peaks erode and hollows fill. A jagged initial profile is combed smooth, high-frequency wrinkles first, until the density is uniform. This is the intuitive content of " $\partial_t P$ is proportional to curvature." Two features distinguish diffusion from ordinary transport. First, it is slow at scale: to spread a distance $L$ takes time $t\sim L^2/D$ , quadratic in distance, which is why molecules cross a cell by diffusion in milliseconds but would need years to diffuse across a room — life uses convection and pumps for anything larger. Second, it is irreversible: although each molecular collision is time-reversible, the coarse-grained density only ever spreads. The arrow enters through the monotonic growth of the spatial entropy of $P$ , and mathematically through the fact that the backward equation blows up.

Figure

A point release spreading as a Gaussian at three times — the peak falls as t^−1/2 while the width grows as √t, with area conserved

The second figure highlights the defining signature of diffusion: the characteristic width of the distribution grows as $\sqrt{t}$ , rather than linearly in time. A signal that moved at constant speed would trace a straight line on a plot of displacement versus time; a ballistic object would trace a parabola in $\langle x^2\rangle$ . Diffusion does neither — its mean-square displacement is a straight line through the origin, $\langle x^2\rangle=2Dt$ , and that single fact is what a hundred years of single-particle tracking has confirmed.

Figure

Mean-square displacement vs time: diffusive ⟨x²⟩ = 2Dt is linear, the outlier against the ballistic ⟨x²⟩ ∝ t² parabola

The core equations

The diffusion equation follows from combining particle conservation with Fick's constitutive law for the diffusive current. Start from the statement that walkers are neither created nor destroyed, so the density can change only by flux through a surface, and close it with the empirical observation that the current runs down the gradient. Combining the two gives a single second-order equation for $P$ .

\frac{\partial P(\mathbf{x},t)}{\partial t} = D\,\nabla^2 P(\mathbf{x},t) \qquad (1)

where: $P(\mathbf{x},t)$ — probability density (or concentration) at position $\mathbf{x}$ and time $t$ ; $D$ — diffusion coefficient, units $(\text{length})^2/\text{time}$ ; $\nabla^2=\sum_i\partial^2/\partial x_i^2$ — the Laplacian, a local averaging operator. Physically: the density rises where it is locally concave-up (below its neighbors' average) and falls where it is concave-down, so any profile is combed toward flatness. The continuity form makes the conservation explicit and names the current whose divergence drives the density. Fick's first law supplies that current as proportional to minus the gradient.

\frac{\partial P}{\partial t} + \nabla\cdot\mathbf{J} = 0,\qquad \mathbf{J} = -D\,\nabla P \qquad (2)

where: $\mathbf{J}$ — probability/particle current density; the minus sign encodes flow from high to low concentration. Reading it physically: the first equation is bookkeeping (change = net inflow), the second is the microphysics (random walkers drift down-gradient on average), and eliminating $\mathbf{J}$ recovers equation (1). The response to a sharply localized release — a delta-function initial condition — is the Green's function of the equation, the heat kernel. Everything else follows by convolution because the equation is linear.

P(x,t\,|\,x_0,0) = \frac{1}{\sqrt{4\pi D t}}\,\exp\!\left[-\frac{(x-x_0)^2}{4D t}\right] \qquad (3)

where: $x_0$ — the release point; the prefactor $(4\pi Dt)^{-1/2}$ keeps $\int P\,\mathrm{d}x=1$ while the Gaussian widens. Physically this is a bell curve whose peak sinks as $t^{-1/2}$ and whose width $\sigma=\sqrt{2Dt}$ grows as $\sqrt{t}$ , spreading the same unit of probability over an ever-broader range. The second moment of that kernel is the mean-square displacement — the quantity a microscope measures by tracking a single particle, and the bridge to Brownian motion (the phenomenon that owns $\langle x^2\rangle=2Dt$ at the trajectory level, cross-referenced below).

\langle (x-x_0)^2\rangle = 2D t \;\;(\text{1D}),\qquad \langle r^2\rangle = 2dD t \;\;(d\text{ dimensions}) \qquad (4)

where: $d$ — spatial dimension; each Cartesian axis contributes an independent $2Dt$ . This linear-in- $t$ growth is the defining diffusive scaling; its slope directly gives $D$ . Because the equation is linear with constant coefficients, Fourier modes are its eigenfunctions: each spatial mode of wavevector $\mathbf{k}$ evolves independently as $e^{-Dk^2t}$ . Consequently, short-wavelength (large- $k$ ) fluctuations decay more rapidly than long-wavelength ones.

\tilde P(\mathbf{k},t)=\tilde P(\mathbf{k},0)\,e^{-D k^2 t} \qquad (5)

where: $\tilde P$ — the Fourier transform of $P$ ; $k=|\mathbf{k}|$ — wavenumber. Physically the decay rate $Dk^2$ grows quadratically with $k$ , so sharp features (large $k$ ) blur first and the broad envelope survives longest — the mathematical statement that diffusion smooths.

Derivation AFrom conservation and Fick's law to the diffusion equation

The diffusion equation is not postulated but assembled from two independent physical inputs: the exact statement that the number of diffusing walkers is conserved, and the constitutive assumption (Fick's law) that the current is proportional to the negative concentration gradient. Keeping these separate makes clear what is rigorous (conservation) and what is constitutive (Fick's law), and it exposes exactly where the single constant $D$ enters. This same two-ingredient structure — a continuity equation plus a current — recurs throughout nonequilibrium physics, from heat conduction to charge transport.

Step 1. Take any fixed region $\Omega$ . The rate of change of the walkers inside it equals minus the net current out through its boundary $\partial\Omega$ : $\frac{\mathrm{d}}{\mathrm{d}t}\int_\Omega P\,\mathrm{d}^dx = -\oint_{\partial\Omega}\mathbf{J}\cdot\mathrm{d}\mathbf{S}$ . Why: nothing is created or destroyed inside, so every change is accounted for by flux across the surface.

Step 2. Apply the divergence theorem to the surface integral, $\oint\mathbf{J}\cdot\mathrm{d}\mathbf{S}=\int_\Omega\nabla\cdot\mathbf{J}\,\mathrm{d}^dx$ , and shrink $\Omega$ to a point. Since the region was arbitrary, the integrands match: $\partial_t P + \nabla\cdot\mathbf{J}=0$ . Why: an integral identity that holds for every region forces the local (differential) form.

Step 3. Close the system with Fick's first law $\mathbf{J}=-D\nabla P$ : walkers on average drift from crowded to sparse regions at a rate proportional to the local gradient. Why: Fick's law is the constitutive assumption that closes the conservation law. Microscopically, it emerges as the long-wavelength, long-time limit of an unbiased random walk (or master equation), where the current becomes proportional to the density gradient.

Step 4. Substitute Fick's law into the continuity equation: $\partial_t P = -\nabla\cdot(-D\nabla P) = D\nabla^2 P$ for constant $D$ . Why: for constant $D$ , $\nabla\cdot(D\nabla P)=D\nabla^2P$ , giving the boxed equation (1).

Go deeper · Derivation B · 1 minThe heat kernel by Fourier transform, and why the front spreads as √t

Derivation BThe heat kernel by Fourier transform, and why the front spreads as √t

The single most useful solution of the diffusion equation is its response to a point release, because linearity makes the point-source solution (the Green's function) the building block for arbitrary initial conditions through superposition. Deriving it by Fourier transform turns the partial differential equation into a one-line ordinary differential equation in each mode, and the inverse transform of the resulting Gaussian is again a Gaussian — the heat kernel of equation (3). The width of that Gaussian encodes the entire diffusive scaling law and fixes the mean-square displacement.

Step 1. Transform equation (1) in space, $\tilde P(\mathbf{k},t)=\int P\,e^{-i\mathbf{k}\cdot\mathbf{x}}\,\mathrm{d}^dx$ . The Fourier transform diagonalizes the Laplacian, replacing $\nabla^2$ by multiplication with $-k^2$ , so $\partial_t\tilde P = -Dk^2\tilde P$ . Why: derivatives in real space are algebraic factors in Fourier space, decoupling the modes.

Step 2. Solve the mode ODE: $\tilde P(\mathbf{k},t)=\tilde P(\mathbf{k},0)\,e^{-Dk^2t}$ . For a point source at $x_0$ , $\tilde P(\mathbf{k},0)=e^{-i\mathbf{k}\cdot\mathbf{x}_0}$ , giving equation (5). Why: each Fourier mode is an independent exponentially decaying amplitude — no mode-mixing in a linear constant-coefficient equation.

Step 3. Invert the transform. The inverse Fourier transform of the Gaussian $e^{-Dk^2t}$ in $k$ is a Gaussian in $x$ of variance $2Dt$ , centered at $x_0$ : $P=(4\pi Dt)^{-d/2}\exp[-(\mathbf{x}-\mathbf{x}_0)^2/4Dt]$ . Why: a Gaussian is its own Fourier transform up to a width inversion — a narrow spectrum in $k$ is a broad bell in $x$ .

Step 4. Read off the scaling. The variance $\sigma^2=2Dt$ means the width grows as $\sigma=\sqrt{2Dt}\propto\sqrt{t}$ and the peak, fixed by normalization, falls as $t^{-d/2}$ . The second moment is $\langle(x-x_0)^2\rangle=2Dt$ in 1D and $2dDt$ in $d$ dimensions. Why: the $\sqrt{t}$ width, not a constant-speed front, is the fingerprint of adding independent random increments — the central limit theorem in action. *One can confirm these results by direct substitution: the heat kernel satisfies $\partial_tP=D\partial_x^2P$ identically, its normalization is $\int P\,\mathrm{d}x=1$ for all $t$ , and its second moment gives $\int x^2 P\,\mathrm{d}x=2Dt$ in 1D and $6Dt$ in 3D.*

Limiting cases

The diffusion equation is the simplest member of a broader family of transport equations, obtained by adding drift, external forces, spatially varying diffusivity, or microscopic lattice dynamics. Two of these — the drift-carrying and potential forms — are exactly the territory of the Fokker–Planck equation, which owns them.

Case	Result	Meaning
Add a drift velocity $v$	$\partial_t P = -v\,\partial_x P + D\,\partial_x^2 P$	Advection–diffusion: the Gaussian packet drifts at speed $v$ while spreading. With $v=\mu F$ this is force-driven transport (owned by the Fokker–Planck equation).
Add an external potential $U$	$\partial_t P = \partial_x\!\big[\tfrac1\gamma U'(x)P + D\,\partial_x P\big]$	The Smoluchowski (overdamped Fokker–Planck) equation; the stationary state is Boltzmann $\propto e^{-U/k_BT}$ when the Einstein relation $D=k_BT/\gamma$ holds.
Position-dependent $D(\mathbf{x})$	$\partial_t P = \nabla\!\cdot\!\big(D(\mathbf{x})\nabla P\big)$	Inhomogeneous medium; the divergence form (not $D\nabla^2$ ) is needed to conserve probability, an Itô-versus-Stratonovich subtlety.
Discrete lattice, $a,\tau\to0$	$\dot P_n = w(P_{n+1}+P_{n-1}-2P_n)$	Nearest-neighbor hopping (a master equation); Taylor-expanding recovers $D\partial_x^2 P$ with $D=wa^2$ .
Long-time, finite box	$P\to 1/V$ (uniform)	All nonzero Fourier modes die as $e^{-Dk^2t}$ ; the slowest mode $D(\pi/L)^2$ sets $\tau_{\text{relax}}\sim L^2/D$ .

Common trapThe width grows as √t, not as t — and the factor of 2 matters

In one dimension the mean-square displacement is $\langle x^2\rangle=2Dt$ , while in $d$ dimensions $\langle r^2\rangle=2dDt$ . Accordingly, the root-mean-square displacement grows as $\sqrt{2Dt}$ in one dimension and $\sqrt{2dDt}$ in $d$ dimensions. Do not confuse this with a variance of $Dt$ (the factor of two is fixed by the standard definition of the diffusion equation $\partial_tP=D\nabla^2P$ ) or with a front that advances at constant speed: a diffusing edge slows down, covering distance $\propto\sqrt{t}$ , so it never has a well-defined velocity. Equally, $D$ here is the macroscopic diffusion coefficient in $(\text{length})^2/\text{time}$ , not a hopping rate; and the equation is irreversible even though every underlying molecular collision is time-reversible — the arrow of time is generated by coarse-graining, not put in by hand.

Worked examples

Example 1 — Extracting the diffusion coefficient from a point release, and its microscopic value. A pollutant is released as a point source and observed to have spread to a root-mean-square radius of $1\ \text{m}$ after $100\ \text{s}$ in three dimensions. What is $D$ , and does it match the molecular Stokes–Einstein prediction for a micron-scale particle in water?

Step 1. Use the 3D mean-square displacement $\langle r^2\rangle = 2dDt = 6Dt$ . With $\langle r^2\rangle=(1\ \text{m})^2$ and $t=100\ \text{s}$ , solve $D=\langle r^2\rangle/(6t)=1/600\approx \mathbf{1.67\times10^{-3}\ \text{m}^2/\text{s}}$ .

Step 2. Compare to the molecular scale. For a bead of radius $a=1\ \mu\text{m}$ in water ( $\eta=10^{-3}\ \text{Pa}\,\cdot\,\text{s}$ ) at $T=300\ \text{K}$ , the Stokes–Einstein value is $D=k_BT/(6\pi\eta a)$ . Numerically $D=\dfrac{(1.381\times10^{-23})(300)}{6\pi(10^{-3})(10^{-6})}\approx \mathbf{2.20\times10^{-13}\ \text{m}^2/\text{s}}$ — ten orders of magnitude smaller. The pollutant dispersion is dominated by turbulent (eddy) diffusion rather than molecular diffusion. Using the molecular diffusivity above gives a characteristic diffusion time $t\sim L^2/(2D)\approx(1\,\mathrm{m})^2/(2\times2.2\times10^{-13}\,\mathrm{m^2\,s^{-1}})\approx2.3\times10^{12}\,\mathrm{s}$ , about $7\times10^4$ years, illustrating why molecular diffusion alone is ineffective over metre-scale distances. This illustrates the quadratic scaling $t\propto L^2$ : molecular diffusion is efficient over microscopic distances but becomes extremely slow on macroscopic scales. (The Stokes–Einstein $D$ here is the same number the Brownian-motion and Einstein-relation articles compute — a cross-article closure.)

Example 2 — How far, and how fast, does a molecule diffuse in a cell? A small protein has $D\approx1\times10^{-10}\ \text{m}^2/\text{s}$ in cytoplasm. Estimate the time to diffuse across a $10\ \mu\text{m}$ cell and across a $1\ \text{mm}$ tissue.

Step 1. From $\langle x^2\rangle\sim L^2\sim 2dDt$ with $d=3$ (use $L^2\approx 6Dt$ ), the cell-crossing time is $t\approx L^2/(6D)=(10^{-5})^2/(6\times10^{-10})\approx \mathbf{1.7\times10^{-2}\ \text{s}}$ — about $17\,\mathrm{ms}$ , showing that diffusion is an efficient transport mechanism across the dimensions of a typical cell.

Step 2. For $L=1\ \text{mm}=10^{-3}\ \text{m}$ the same estimate gives $t\approx(10^{-3})^2/(6\times10^{-10})\approx \mathbf{1.7\times10^{3}\ \text{s}}$ — about 28 minutes. The quadratic scaling means a hundredfold larger distance costs a ten-thousandfold longer time, which is precisely why tissues need vasculature and cells above a critical size need active transport. These estimates illustrate the quadratic dependence of diffusion time on distance.

Origins · 1 minAdolf Fick and how this was discovered

Adolf Fick

1829–1901 · German physiologist and physician who, in 1855, formulated the phenomenological laws of diffusion by direct analogy with Fourier's law of heat conduction — the first quantitative statement that matter flows down a concentration gradient. Trained in medicine, Fick was drawn to the physics of the body: he applied his diffusion laws to gas exchange in the lungs and blood, and later devised the Fick principle for measuring cardiac output from oxygen uptake. His diffusion laws, written for salt spreading in water, gave the constitutive current J = −D∇P that closes the continuity equation into the diffusion equation, and they remain the working definition of the diffusion coefficient across chemistry, biology, and materials science.

History

1822

— Fourier publishes the heat equation, mathematically identical to the diffusion equation, from Fourier's law plus energy conservation

1827

— Robert Brown observes the incessant jiggling of pollen grains in water under the microscope

1855

— Adolf Fick formulates his laws of diffusion, defining the current

J = -D\nabla P

and the diffusion coefficient

1905

— Einstein derives the diffusion equation from molecular random motion and identifies

D = k_B T/(6\pi\eta a)

, fixing the atomic scale

1906

— Smoluchowski gives an independent kinetic derivation of the same spreading law

1908

— Perrin begins tracking Brownian particles, measuring

\langle r^2\rangle = 2dDt

to determine Avogadro's number

1917

— Smoluchowski's and later Fokker–Planck extensions add drift and potentials to the bare diffusion equation

Going further

References. Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 15 (Fluctuations and Nonequilibrium Statistical Mechanics) · Kardar, Statistical Physics of Particles, ch. 9 · Reif, Fundamentals of Statistical and Thermal Physics, ch. 15 · Sethna, Statistical Mechanics: Entropy, Order Parameters, and Complexity, ch. 2 · Landau & Lifshitz, Physical Kinetics (vol. 10) and Fluid Mechanics (vol. 6, heat equation) · Einstein, Ann. Phys. 17, 549 (1905) · Risken, The Fokker–Planck Equation. Key results to keep in mind: the heat kernel follows from the Fourier transform; $\langle x^2\rangle=2Dt$ in 1D and $2dDt$ in $d$ dimensions; the random-walk continuum limit gives $D=a^2/2\tau$ ; and the Einstein relation $D=k_BT/\gamma$ with Stokes drag $\gamma=6\pi\eta a$ fixes the molecular value of $D$ .

Builds on: Brownian Motion · The Langevin Equation · The Einstein Relation. Leads to: The Fokker–Planck Equation · The Master Equation · The Ornstein–Uhlenbeck Process · Kramers' Escape Rate Theory.

Is the diffusion equation the same as the heat equation?

Mathematically, yes: both are $\partial_t u = D\nabla^2 u$ , a linear parabolic PDE. Fourier derived it for temperature in 1822 (with $D$ the thermal diffusivity), Fick wrote it for concentration in 1855. The physics differs only in what $u$ and $D$ represent — temperature and thermal diffusivity, versus probability/concentration and a mass diffusion coefficient. The identical structure is why every result about heat kernels, Green's functions, and $\sqrt{t}$ penetration depths transfers directly between heat conduction and molecular diffusion.

Where does the diffusion equation come from microscopically?

From a random walk. Put a walker on a lattice of spacing $a$ that hops left or right with equal probability every time $\tau$ . Its probability obeys $P(x,t+\tau)=\tfrac12 P(x-a,t)+\tfrac12 P(x+a,t)$ . Taylor-expand in $a$ and $\tau$ : the first-order time term matches the second-order space term, and taking $a,\tau\to0$ with $D=a^2/2\tau$ fixed leaves $\partial_t P=D\partial_x^2P$ . The single number $D$ packages the microscopic step size and rate; the central limit theorem guarantees the Gaussian kernel regardless of the exact step distribution, as long as it has finite variance.

Why does a spreading front go as √t rather than at constant speed?

Because independent random steps add in variance, not in displacement. After $N$ uncorrelated steps of typical size $a$ , the mean displacement is zero but the mean-square displacement is $Na^2$ , so the typical distance is $\sqrt{N}\,a$ . With $N\propto t$ , the spread grows as $\sqrt{t}$ . A constant-speed front would require the steps to be correlated (all pointing the same way), which is ballistic motion, not diffusion. The crossover from ballistic (short times, before the first collision) to diffusive (long times) is exactly what the velocity autocorrelation of the Langevin equation captures.

How does adding an external force change the equation?

It adds a drift current. In a potential $U(x)$ the current gains an advective piece $-(1/\gamma)U'(x)P$ alongside the Fickian $-D\partial_x P$ , giving the Smoluchowski equation $\partial_t P=\partial_x[(1/\gamma)U'P+D\partial_x P]$ . Its stationary, zero-current state is the Boltzmann distribution $P_{\text{eq}}\propto e^{-U/k_BT}$ — but only if $D=k_BT/\gamma$ , the Einstein relation. This drift-plus-diffusion structure is exactly the Fokker–Planck equation, which owns the force-driven case; the bare diffusion equation is its force-free heart.

Does the diffusion equation conserve energy or increase entropy?

It conserves total probability (or total mass/heat), because it is a continuity equation with no source. It does not conserve the "energy" $\int P^2\,\mathrm{d}x$ , which strictly decreases as gradients smooth out. And it monotonically increases the spatial (Gibbs–Shannon) entropy $S=-k_B\int P\ln P\,\mathrm{d}x$ : differentiating along the flow gives $\dot S = k_BD\int(\partial_x P)^2/P\,\mathrm{d}x\ge0$ , with equality only for a uniform distribution. This entropy increase is the microscopic arrow of time expressed at the level of the density.

When does normal diffusion break down?

Whenever the underlying steps are not independent with finite variance. If the walker gets trapped for heavy-tailed waiting times, or moves in a crowded, obstacle-filled medium, the mean-square displacement grows as $\langle x^2\rangle\sim t^\alpha$ with $\alpha\ne1$ — subdiffusion ( $\alpha<1$ ) in cytoplasm and porous media, superdiffusion ( $\alpha>1$ ) for active or Lévy-flight motion. These anomalous regimes require fractional or continuous-time-random-walk generalizations; the ordinary diffusion equation is recovered only in the $\alpha=1$ limit where the central limit theorem applies.

Why must position-dependent diffusion be written in divergence form?

Because only $\partial_t P=\nabla\cdot(D(\mathbf{x})\nabla P)$ conserves probability. The naive form $D(\mathbf{x})\nabla^2 P$ would create or destroy probability wherever $D$ varies in space, since the current would no longer be a pure divergence. The divergence form also encodes a real physical ambiguity: for a Langevin equation with state-dependent noise, whether the extra drift term appears depends on the Itô-versus-Stratonovich interpretation. Getting this right matters for particles diffusing near walls, in temperature gradients, or through membranes where the local mobility changes.