The Fokker-Planck Equation

Release a swarm of Brownian particles into a bowl-shaped potential — a colloid in an optical trap, an ion in a molecular cage — and two things happen at once. Each particle jitters unpredictably, its trajectory a different scribble every time you run the experiment. Yet the cloud of particles settles into a perfectly reproducible shape: it slides downhill and, at long times, freezes into the Boltzmann bell $e^{-U/k_BT}$ , no matter how you started it. The Fokker–Planck equation is the law that governs this reproducible cloud. It is the deterministic partial differential equation for the probability density $P(x,t)$ that a single noisy trajectory would visit, and it converts an intractable stochastic problem into the far better-developed machinery of linear PDEs. Its structure is that of a conservation law. Probability is neither created nor destroyed, so $\partial_t P = -\partial_x J$ with a current $J$ made of two competing pieces: a drift that carries probability bodily downhill along the force, and a diffusion that flattens gradients. The drift sharpens the density into the potential's minima; the diffusion smears it back out. Their balance sets the width of the equilibrium distribution — and when the drift derives from a potential and the diffusion coefficient obeys the Einstein relation $D=k_BT/\gamma$ , that balance produces exactly the Boltzmann distribution. This is the equation's most important statement: fluctuation and dissipation, tuned to each other, force stochastic dynamics to relax to thermal equilibrium. The Fokker–Planck equation sits at the center of nonequilibrium statistical mechanics. In the force-free limit it collapses to the pure diffusion equation; with a linear restoring drift it becomes the exactly solvable Ornstein–Uhlenbeck process; in the overdamped limit it is the Smoluchowski equation, and in full phase space the Klein–Kramers equation behind Kramers' escape-rate theory. It is the deterministic partner of the Langevin trajectory — the "hydrodynamics" of a single noisy degree of freedom — and its systematic derivation from a Markov process is the Kramers–Moyal expansion, which truncates exactly at second order for Gaussian white noise.

Key takeaways

The Fokker–Planck equation is a deterministic PDE for the probability density $P(x,t)$ of a system driven by Langevin dynamics with Gaussian white noise — the ensemble description behind the stochastic trajectory.
It is a continuity equation $\partial_t P=-\partial_x J$ with current $J=aP-\tfrac12\partial_x(b^2P)$ : a drift term $a$ carrying probability downhill and a diffusion term $\tfrac12 b^2$ spreading it (with $D=\tfrac12 b^2$ ).
Its stationary, zero-current state is the Boltzmann distribution $P_{\text{ss}}\propto e^{-U/k_BT}$ — but only when the drift derives from a potential and the Einstein relation $D=k_BT/\gamma$ holds.
Setting the drift to zero recovers the diffusion equation exactly; a linear drift gives the Ornstein–Uhlenbeck process (relaxation to a stationary Gaussian of variance $D/\gamma$ ).
For multiplicative (state-dependent) noise the drift depends on the Itô-versus-Stratonovich convention, differing by the spurious drift $\tfrac12 b b'$ ; you must fix the calculus before writing the equation.
Key idea: a Langevin trajectory and a Fokker–Planck density are two faces of the same Markov process; the drift and diffusion are the first two jump moments, and the Einstein relation is precisely the condition that makes the Boltzmann distribution the zero-current fixed point.

The core insight

A Langevin equation with Gaussian white noise and its Fokker–Planck density are two descriptions of one Markov process: the trajectory is random, the density is deterministic. The drift $a$ and diffusion $D=\tfrac12 b^2$ are the first two Kramers–Moyal jump moments; the stationary zero-current state is Boltzmann $\propto e^{-U/k_BT}$ exactly when the Einstein relation $D=k_BT/\gamma$ holds.

Where this lives today

The Fokker–Planck equation is the workhorse of stochastic modeling far beyond its statistical-mechanics origins. It governs the phase-space distribution of electrons in plasmas and accelerators, the intensity fluctuations of laser light, the allele frequencies of population genetics (the Wright–Fisher diffusion), and the evolution of option prices in mathematical finance (where its adjoint is the backward Kolmogorov equation behind Black–Scholes). Its overdamped Smoluchowski form is the standard description of molecular motors, protein folding on rugged energy landscapes, and ion transport through channels; its Klein–Kramers form underlies modern reaction-rate theory. In statistical physics specifically, the equation is the bridge between the Langevin trajectory picture and the ensemble picture, and its stationary Boltzmann state is the anchor that ties nonequilibrium stochastic dynamics back to equilibrium thermodynamics. It is the parent of the Ornstein–Uhlenbeck process and Kramers' escape problem (both developed downstream), and its systematic derivation from a Markov process — the Kramers–Moyal expansion — connects it to the master equation for discrete-state systems. Numerically, its adjoint delivers mean first-passage times for everything from neuron firing to genetic switches, making it one of the most practically deployed equations in computational biophysics.

The physical picture

Picture $P(x,t)$ as a fluid of probability. The Fokker–Planck equation is the continuity equation for this fluid, and its current has two ingredients. The drift current $aP$ is advection: it carries the whole distribution bodily in the direction of the force, like a river sweeping a dye downstream. The diffusive current $-\tfrac12\partial_x(b^2P)$ is the same down-gradient flattening as ordinary diffusion. Everything the equation does is the competition between these two — advection concentrating probability where the force points, diffusion spreading it back. Read this way, the deep lesson is that randomness at the trajectory level produces determinism at the distribution level. A single Brownian particle is unpredictable, but the density of infinitely many copies obeys a unique, deterministic evolution — exactly as the chaotic motion of $10^{23}$ molecules produces smooth, deterministic hydrodynamics. The Fokker–Planck equation is the hydrodynamics of one noisy coordinate.

Figure

The probability current has two competing pieces: a drift current aP carrying the density downhill and a diffusive current −∂(DP) flattening it; their balance is equilibrium

The balance point is equilibrium. In a confining potential, drift pulls probability into the minimum while diffusion pushes it out; the width where they balance is set by temperature. If — and only if — the diffusion coefficient and the friction obey the Einstein relation $D=k_BT/\gamma$ , the stationary distribution is the Boltzmann factor $e^{-U/k_BT}$ , with the current vanishing everywhere. This is the precise sense in which the fluctuation–dissipation tie forces a stochastic system to thermalize. (The general fluctuation–dissipation theorem that this instance anticipates is developed downstream in the fluctuation-dissipation-theorem article.)

Figure

Free diffusion spreads without bound while a confining drift drives the same particle to a stationary Boltzmann bell — the zero-current equilibrium

The core equations

The Fokker–Planck equation combines a drift and a diffusion into one second-order PDE for the density. Written with a general drift $a(x)$ and a general noise amplitude $b(x)$ (so that $D=\tfrac12 b^2$ is the diffusion coefficient), it reads:

\frac{\partial P(x,t)}{\partial t} = -\frac{\partial}{\partial x}\big[a(x)\,P\big] + \frac{1}{2}\frac{\partial^2}{\partial x^2}\big[b^2(x)\,P\big] \qquad (1)

where: $P(x,t)$ — probability density at position $x$ , time $t$ ; $a(x)$ — drift coefficient, the first Kramers–Moyal jump moment $\lim_{\tau\to0}\langle\Delta x\rangle/\tau$ ; $b^2(x)=2D(x)$ — noise variance rate, twice the diffusion coefficient (the second jump moment). Physically the first term advects probability along the drift, the second spreads it by diffusion. Because probability is conserved, equation (1) is exactly a continuity equation, and reading off the current makes the two ingredients explicit.

\frac{\partial P}{\partial t} = -\frac{\partial J}{\partial x},\qquad J(x,t) = a(x)P - \frac{1}{2}\frac{\partial}{\partial x}\big[b^2(x)P\big] \qquad (2)

where: $J(x,t)$ — probability current; it vanishes at the boundaries for any normalizable solution. The drift piece $aP$ is advective; the term $-\tfrac12\partial_x(b^2P)$ is the Fickian diffusive flow. Stationarity means $\partial_t P=0$ , i.e. a spatially constant current; in a one-dimensional potential landscape that constant is zero — detailed balance. The trajectory-level partner of equation (1) is the Langevin equation, whose drift and noise amplitude are exactly the same $a$ and $b$ (the equivalence is owned by the Langevin-equation article; here it is the input).

\dot{x} = a(x) + b(x)\,\xi(t),\qquad \langle\xi(t)\rangle=0,\quad \langle\xi(t)\xi(t')\rangle = \delta(t-t') \qquad (3)

where: $\xi(t)$ — Gaussian white noise: zero mean, delta-correlated, unit strength (defined in the Wiener-process/white-noise article); $b(x)$ — noise amplitude, $b^2=2D$ . Physically each realization of $\xi$ generates one random trajectory; equation (1) is the density obtained by averaging over all realizations. The systematic route from the trajectory to the density is the Kramers–Moyal expansion, an infinite series in the jump moments that truncates exactly at second order for Gaussian white noise. (Its full derivation and Pawula's theorem are developed downstream in the Kramers–Moyal-expansion article; here it is the origin of the drift and diffusion.)

\frac{\partial P}{\partial t} = \sum_{n=1}^{\infty}\Big(-\frac{\partial}{\partial x}\Big)^{n}\big[M_n(x)P\big],\qquad M_n=\frac{1}{n!}\lim_{\tau\to0}\frac{\langle(\Delta x)^n\rangle}{\tau} \qquad (4)

where: $M_n$ — the $n$ -th jump moment per unit time; $M_1=a$ (drift), $M_2=\tfrac12 b^2=D$ (diffusion). Physically, for Gaussian white noise the short-time increment is Gaussian, so all moments beyond the second vanish and only drift and diffusion survive. The equation's payoff is its stationary state. When the drift comes from a potential and the diffusion obeys the Einstein relation, the zero-current condition integrates to the Boltzmann distribution.

a(x) = -\frac{U'(x)}{\gamma},\quad D=\frac{k_BT}{\gamma}\;\Longrightarrow\; P_{\text{ss}}(x)\propto e^{-U(x)/k_BT} \qquad (5)

where: $U(x)$ — external potential; $\gamma$ — friction coefficient; $k_BT$ — thermal energy. The friction $\gamma$ cancels between drift and diffusion, so only the ratio $U/k_BT$ survives — equilibrium depends on temperature, not on the dissipation rate.

Derivation AThe continuity form and the zero-current Boltzmann state

The single most important property of the Fokker–Planck equation is that it enforces thermal equilibrium: given a potential drift and an Einstein-tuned diffusion, its unique stationary state is the Boltzmann distribution. Establishing this cleanly requires first recognizing the equation as a conservation law with a specific probability current, then demanding that current vanish. The calculation shows that the friction coefficient — the very quantity measuring dissipation — cancels, leaving only the ratio of energy to temperature, which is precisely why the fluctuation and dissipation must be tied by the Einstein relation for the Boltzmann state to emerge.

Step 1. Rewrite equation (1) as a continuity equation. Total probability $\int P\,\mathrm{d}x=1$ is conserved, so $\partial_t P=-\partial_x J$ for some current; matching to (1) gives $J=aP-\tfrac12\partial_x(b^2P)$ , an advective drift current plus a diffusive current. Why: a locally conserved density can always be written as minus the divergence of a current that vanishes at infinity.

Step 2. Impose stationarity. Setting $\partial_t P=0$ forces $\partial_x J=0$ , so $J$ is constant in $x$ ; for a normalizable distribution in a potential landscape the current must vanish at the boundaries, hence $J\equiv0$ everywhere. Why: a nonzero constant current would carry probability out to infinity, contradicting normalization — the equilibrium of a closed system carries no net flow.

Step 3. Solve $J=0$ for constant diffusion $D=\tfrac12 b^2$ : $aP=D\,P'$ , so $P'/P=a/D$ and $P_{\text{ss}}\propto\exp[\int^x a(x')/D\,\mathrm{d}x']$ . Why: the zero-current condition is a first-order ODE separable in $P$ .

Step 4. Insert the potential drift $a=-U'/\gamma$ and the Einstein value $D=k_BT/\gamma$ : the exponent becomes $\int^x \frac{-U'/\gamma}{k_BT/\gamma}\mathrm{d}x'=-U(x)/k_BT$ , so $P_{\text{ss}}\propto e^{-U/k_BT}$ . Why: the friction $\gamma$ appears in both drift and diffusion and cancels — only the Einstein-tied combination gives the temperature-controlled Boltzmann weight; any other relation between $D$ and $\gamma$ would give the wrong stationary distribution. *Verified: the Boltzmann state $P_{\text{ss}}\propto e^{-\beta U}$ is shown symbolically to give zero current $J=aP-\partial_x(DP)=0$ for $a=-\beta D U'$ , that $a=-U'/\gamma$ equals $-\beta DU'$ with $D=k_BT/\gamma$ , and that the equation reduces to the diffusion equation when $a=0$ — all in verify_aq-sm-fokker-planck-equation.py.*

Go deeper · Derivation B · 1 minDrift and diffusion from the short-time moments of the Langevin equation

Derivation BDrift and diffusion from the short-time moments of the Langevin equation

To see that the Fokker–Planck density and the Langevin trajectory describe the same physics, one computes how a sharply localized density evolves over an infinitesimal time step directly from the stochastic equation. The first two moments of the displacement identify the drift and diffusion coefficients, and the Gaussianity of white noise ensures that all higher moments are subleading — the reason the expansion (4) truncates exactly at second order rather than being an uncontrolled approximation. This calculation is the concrete content of the claim that a Langevin equation "is" a Fokker–Planck equation.

Step 1. Integrate the Langevin equation (3) over a short interval $\tau$ , holding $a,b$ at $x(t)$ : $\Delta x = a(x)\tau + \int_t^{t+\tau} b(x)\xi\,\mathrm{d}t'$ . Why: over a short enough step the coefficients are approximately constant, so the increment splits into a deterministic and a random part.

Step 2. Take the first moment. The noise averages to zero, so $\langle\Delta x\rangle = a(x)\tau + O(\tau^2)$ , giving $M_1=a$ . Why: $\langle\xi\rangle=0$ kills the random part at leading order, leaving only the drift.

Step 3. Take the second moment. Using $\langle\xi(t')\xi(t'')\rangle=\delta(t'-t'')$ , the double integral of the noise gives $\langle\Delta x^2\rangle = b^2(x)\tau + O(\tau^2) = 2D\tau$ , so $M_2=\tfrac12 b^2=D$ . Why: the delta-correlation makes the variance grow linearly in $\tau$ — the diffusive $\sqrt{\tau}$ scaling of the increment — while the drift contributes only at $O(\tau^2)$ .

Step 4. Show higher moments are subleading. Odd moments vanish by symmetry; even moments factor into products of deltas scaling as $\tau^{n/2}$ , which is $O(\tau^2)$ or smaller for $n\ge3$ , so $M_{n\ge3}=0$ . Assembling $M_1=a$ , $M_2=D$ with all higher terms absent gives equation (1). Why: Gaussian white noise has no cumulants beyond the second, so the Kramers–Moyal series stops at drift and diffusion — Pawula's theorem forbids stopping at any other finite order. *Verified: the leading jump moments $M_1=a$ and $M_2=2D$ (so the diffusion coefficient is $D=\tfrac12 b^2$ ) and the continuity identity $-\partial_x[aP-\partial_x(DP)]=-\partial_x(aP)+\partial_x^2(DP)$ are checked symbolically in verify_aq-sm-fokker-planck-equation.py.*

Go deeper · Derivation C · 1 minRelaxation to equilibrium: relative entropy as a Lyapunov function

Derivation CRelaxation to equilibrium: relative entropy as a Lyapunov function

Beyond identifying the stationary state, one wants to prove the density actually reaches it. The tool is the relative entropy (Kullback–Leibler divergence) of $P$ from the Boltzmann state, which for the potential Fokker–Planck equation is a free-energy functional. Differentiating it along the flow yields a manifestly non-positive rate, so it decreases monotonically to zero — a rigorous H-theorem for the Fokker–Planck equation, the stochastic-dynamics counterpart of the second law and the mechanism by which the confining drift and diffusion drive any initial cloud to thermal equilibrium.

Step 1. Define $H(t)=\int P\ln(P/P_{\text{ss}})\,\mathrm{d}x\ge0$ , the relative entropy, which vanishes if and only if $P=P_{\text{ss}}$ . For the constant- $D$ potential case this equals $(F[P]-F_{\text{eq}})/k_BT$ , a free energy above its minimum. Why: relative entropy is non-negative by Jensen's inequality and is a natural distance to equilibrium.

Step 2. Write the current in the form $J=-D\,P_{\text{ss}}\,\partial_x(P/P_{\text{ss}})$ , valid for the potential Fokker–Planck equation with $P_{\text{ss}}\propto e^{-U/k_BT}$ . Why: factoring out the stationary state exposes the gradient of the ratio $P/P_{\text{ss}}$ , which is the true driving force toward equilibrium.

Step 3. Differentiate $H$ along the flow, using $\partial_t P=-\partial_x J$ and integrating by parts (boundary terms vanish for normalizable $P$ ): $\dot H = -\int J\,\partial_x\ln(P/P_{\text{ss}})\,\mathrm{d}x$ . Why: the time derivative of the entropy is driven entirely by the probability current times the gradient of the log-ratio.

Step 4. Substitute the current from Step 2 to get $\dot H = -\int \dfrac{D\,P_{\text{ss}}}{(P/P_{\text{ss}})}\big[\partial_x(P/P_{\text{ss}})\big]^2\,\mathrm{d}x \le 0$ , non-positive because $D,P,P_{\text{ss}}>0$ , with equality only at $P=P_{\text{ss}}$ . Why: the integrand is a manifestly non-negative squared gradient with a positive prefactor, so $H$ can only decrease — the system relaxes irreversibly to Boltzmann. *Verified: the H-theorem integrand prefactor $D\,P_{\text{ss}}/(P/P_{\text{ss}})=D\,P_{\text{ss}}^2/P>0$ is checked positive symbolically, and the Ornstein–Uhlenbeck moment ODEs with stationary variance $D/\gamma$ are verified in verify_aq-sm-fokker-planck-equation.py.*

Limiting cases

The Fokker–Planck equation is a template that specializes to many familiar equations depending on the drift and diffusion. Several of these limits are owned by their own articles; the equation is the parent.

Case	Result	Meaning
No drift, constant $D$	$\partial_t P = D\,\partial_x^2 P$	The pure diffusion equation: a Gaussian packet spreading with variance $2Dt$ (free Brownian motion).
Linear restoring drift $a=-\gamma x$	$\partial_t P = \gamma\,\partial_x(xP) + D\,\partial_x^2 P$	The Ornstein–Uhlenbeck process: relaxation to a stationary Gaussian of variance $D/\gamma$ (developed downstream).
Overdamped potential	$\partial_t P = \partial_x\!\big[\tfrac1\gamma U'(x)P + \tfrac{k_BT}\gamma\partial_x P\big]$	The Smoluchowski equation; stationary state $\propto e^{-U/k_BT}$ by the Einstein relation.
Full phase space $(x,v)$	Klein–Kramers equation	Position and velocity together — the starting point of Kramers' escape-rate theory (developed downstream).
Multiplicative noise $b(x)$	drift shifts by $\tfrac12 bb'$	Itô and Stratonovich forms differ; the convention must be fixed (owned by the Itô–Stratonovich article).

Worked examples

Example 1 — Verifying that the Boltzmann state is the equilibrium, and that fluctuation must match dissipation. Show that $P_{\text{ss}}\propto e^{-\beta U}$ (with $\beta=1/k_BT$ ) is a zero-current stationary state of the overdamped Fokker–Planck equation, and find the condition on $D$ .

Step 1. Write the zero-current condition for constant $D$ : $J=aP-D\,\partial_x P=0$ . With $P=e^{-\beta U}$ , $\partial_x P=-\beta U'\,e^{-\beta U}$ , so $J=(a+\beta D U')e^{-\beta U}$ . For this to vanish for all $x$ requires $a=-\beta D\,U'$ .

Step 2. Compare to the Langevin drift $a=-U'/\gamma$ . Matching $-U'/\gamma=-\beta D U'$ gives $D=1/(\beta\gamma)=\mathbf{k_BT/\gamma}$ — the Einstein relation. The Boltzmann distribution is the stationary state if and only if the diffusion coefficient is Einstein-tied to the friction; any other $D$ gives a different, non-thermal steady state. (This is the fluctuation–dissipation balance the Einstein-relation article owns; the Fokker–Planck equation is where it becomes the equilibrium condition.)

Example 2 — Ornstein–Uhlenbeck relaxation: how fast and how wide. A colloidal particle in a harmonic optical trap has a linear restoring drift $a=-\gamma x$ and diffusion $D$ . Starting from a point at $x_0$ , find the mean and variance at time $t$ and the stationary width.

Step 1. Multiply the Fokker–Planck equation by $x$ and by $x^2$ and integrate by parts to get closed moment equations: $\dot\mu=-\gamma\mu$ and $\dot{\sigma^2}=-2\gamma\sigma^2+2D$ . Solving, $\mu(t)=x_0 e^{-\gamma t}$ (the mean relaxes exponentially to the trap center) and $\sigma^2(t)=\dfrac{D}{\gamma}\big(1-e^{-2\gamma t}\big)$ .

Step 2. Take the stationary limit $t\to\infty$ : $\mu\to0$ and $\sigma^2\to \mathbf{D/\gamma}$ . For a trap with $D=2\times10^{-13}\ \text{m}^2/\text{s}$ and $\gamma=10\ \text{s}^{-1}$ , the equilibrium variance is $\sigma^2=D/\gamma=\mathbf{2\times10^{-14}\ \text{m}^2}$ , i.e. an RMS excursion of $\sqrt{\sigma^2}\approx1.4\times10^{-7}\ \text{m}\approx140\ \text{nm}$ . This Gaussian is exactly the Boltzmann distribution for the harmonic potential $U=\tfrac12\kappa x^2$ with $\kappa=\gamma^2 m$ -type stiffness, and it is what optical-tweezer experiments measure. The closed-form OU solution is the exactly solvable window into Fokker–Planck relaxation.

Origins · 1 minAdriaan Fokker and how this was discovered

Adriaan Fokker

1887–1972 · Dutch physicist and musician, a doctoral student of Lorentz and a cousin of the aircraft pioneer Anthony Fokker. In 1914 he derived the equation for the probability distribution of a Brownian particle's velocity under the influence of random collisions, which Max Planck cast in general form in 1917 — together giving the Fokker–Planck equation. Fokker worked on relativity and radiation theory, but he is equally remembered for his passion for microtonal music: he built the 31-tone "Fokker organ" and championed equal temperaments beyond the standard twelve. The equation bearing his name became the central tool of nonequilibrium statistical mechanics, governing everything from Brownian motion to laser physics and stochastic finance.

History

1827

— Robert Brown reports the ceaseless motion of pollen grains suspended in water

1905

— Einstein derives the diffusion equation from molecular motion and the relation D = k_BT/γ

1914

— Adriaan Fokker obtains the equation for the velocity distribution of a Brownian particle

1917

— Max Planck casts the equation in general form; it becomes the Fokker–Planck equation

1930

— Ornstein and Uhlenbeck solve the harmonic-drift case in closed form (the OU process)

1940

— Kramers applies the phase-space (Klein–Kramers) equation to thermally activated escape

1951

— Kramers and Moyal formulate the systematic jump-moment expansion; Pawula's theorem (1967) proves the second-order truncation is exact for Gaussian noise

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 & 6 · Risken, The Fokker–Planck Equation: Methods of Solution and Applications, 2nd ed. · Gardiner, Handbook of Stochastic Methods, Ch. 5 · van Kampen, Stochastic Processes in Physics and Chemistry, Ch. VIII.

Provenance. All equations are standard textbook results. A verifier should reproduce the drift/diffusion identification $M_1=a$ , $M_2=2D$ from the Langevin short-time moments (Risken Ch. 4; Gardiner §5.3); confirm the second-order truncation and Pawula's theorem (Risken §4.3); check the OU moment equations and stationary variance $D/\gamma$ (Risken Ch. 5); verify the Boltzmann stationary state under the Einstein relation $D=k_BT/\gamma$ (Pathria & Beale Ch. 15); confirm the H-theorem sign-definiteness (Risken §6.1; Sethna Ch. 6); and check the Itô–Stratonovich drift shift $\tfrac12 bb'$ (Gardiner §4.3). These are reproduced in verify_aq-sm-fokker-planck-equation.py.

Builds on: The Langevin Equation · Wiener Process and White Noise · Brownian Motion · The Diffusion Equation. Leads to: The Ornstein–Uhlenbeck Process · Kramers' Escape Rate Theory · The Kramers–Moyal Expansion · Itô versus Stratonovich Calculus · The Einstein Relation.

What exactly is the difference between the Fokker–Planck and the diffusion equation?

The diffusion equation is the drift-free special case. The Fokker–Planck equation adds an advective drift current $aP$ alongside the diffusive current, so it describes probability that is both carried (by a force) and spread (by noise). Set the drift to zero and $\partial_t P=-\partial_x J$ collapses to $\partial_t P=D\partial_x^2P$ , the pure diffusion equation. The crucial new feature the drift brings is a nontrivial stationary state: free diffusion spreads forever, but a confining drift lets probability settle into the Boltzmann distribution, which is why the Fokker–Planck equation, not the diffusion equation, is the bridge to equilibrium.

Why does the Einstein relation appear as the equilibrium condition?

Because equilibrium requires the systematic pull of the force and the random spread of the noise to be tuned to each other. In the zero-current stationary state, the drift $a=-U'/\gamma$ pushes probability into the potential minimum at a rate set by friction $\gamma$ , while diffusion $D$ pushes it back out. The Boltzmann weight $e^{-U/k_BT}$ emerges only when the exponent $\int a/D$ reduces to $-U/k_BT$ , which forces $D=k_BT/\gamma$ . Physically the same collisions that damp the particle (dissipation, $\gamma$ ) also kick it (fluctuation, $D$ ); the Einstein relation is the statement that these two effects of one microscopic mechanism are locked together — the seed of the general fluctuation–dissipation theorem.

Is the Fokker–Planck equation the same as the Langevin equation?

They are two descriptions of the identical Markov process. The Langevin equation follows one random trajectory $x(t)$ ; the Fokker–Planck equation follows the probability density $P(x,t)$ obtained by averaging over infinitely many trajectories. The drift $a$ and diffusion $D=\tfrac12 b^2$ in the Fokker–Planck equation are exactly the deterministic drift and the noise-variance rate of the Langevin equation — they are the first two Kramers–Moyal jump moments. Which description to use is a matter of convenience: the Langevin form is natural for simulation and intuition, the Fokker–Planck form for exact solutions, spectra, and first-passage times.

What is the stationary state if the drift is not from a potential?

Then there can be a nonzero steady current, and the stationary state is generally not Boltzmann. In one dimension with reflecting boundaries any drift can be written as a gradient, so a potential always exists and the state is Boltzmann-like. But in higher dimensions or with periodic boundaries, a non-conservative (rotational) drift sustains a persistent probability current — a nonequilibrium steady state that breaks detailed balance. Such states carry entropy production and are the subject of stochastic thermodynamics; the Boltzmann state is the special, detailed-balance case where the current vanishes everywhere.

Why is the density deterministic if every trajectory is random?

Because it is an ensemble average. A single realization of the noise gives one unpredictable path, but the probability density is defined over all realizations, and averaging over infinitely many of them yields a unique, smooth, deterministic function of time. This is the same logic by which the random motion of $10^{23}$ molecules produces the deterministic Navier–Stokes equations: individual microscopic randomness averages into macroscopic determinism. The Fokker–Planck equation is that averaged, deterministic law for a single noisy coordinate.

What does the adjoint (backward) equation compute?

The adjoint operator $\mathcal{L}^\dagger=a\,\partial_x+D\,\partial_x^2$ acts on observables rather than densities and governs quantities like the mean first-passage time $T(x)$ to reach a boundary, which solves $a\,T'+D\,T''=-1$ . Where the forward Fokker–Planck equation evolves "where is the probability now," the backward equation answers "starting here, what is the expected time (or probability) to reach a target." This is how the same operator delivers reaction rates, neuron firing times, and the timing of genetic switches — practical quantities that the forward density does not directly give.

Does the second-order truncation ever fail?

Yes — whenever the noise is not Gaussian and white. Shot noise, jump processes, and heavy-tailed noise have nonzero higher jump moments, so the Kramers–Moyal expansion does not truncate at second order and the dynamics is governed by a full master equation with a jump kernel, not a Fokker–Planck equation. Pawula's theorem is the sharp statement: for a positive transition probability the expansion must either stop at order two or continue to infinity — there is no consistent third-order theory. The Fokker–Planck equation is exact precisely in the Gaussian-white-noise (diffusive) limit and is otherwise a diffusive approximation to the master equation.