Fokker–Planck Equation

Seen as a stochastic process, the Fokker–Planck equation is the propagator of a continuous Markov process — the deterministic evolution law for the probability density carried by an entire ensemble of noisy trajectories. This article takes that process viewpoint. The Fokker–Planck equation itself — its drift–diffusion structure, its origin as the second-order truncation of the Kramers–Moyal expansion of the master equation, Pawula's theorem, and its Boltzmann stationary state — is derived and owned by `fokker-planck-equation`; we recall those results by name rather than re-deriving them. What we develop here is the pair of Kolmogorov equations and what each is for. There are two adjoint operators. The forward (Kolmogorov-forward) equation, $\partial_t P=\mathcal L_{\rm FP}P$ , propagates a density forward in the final variable: given where you start, where will the probability cloud be later? The backward (Kolmogorov-backward) equation, $\partial_\tau Q=\mathcal L^\dagger_{\rm FP}Q$ , propagates an observable backward in the initial variable: as a function of where you start, what is the expected value of some future quantity? The two operators are adjoints of each other on the space of densities and observables, and this adjoint relationship is the engine behind first-passage theory. The single most useful thing the backward equation buys you is the mean first-passage time. The average time to first reach a boundary, as a function of the starting point $x$ , solves the deceptively simple boundary-value problem $\mathcal L^\dagger T(x)=-1$ . From this one equation flow escape rates, reaction times, the Kramers barrier-crossing formula, neuron firing times, and the exit statistics of gene-expression switches. The forward picture tells you how a cloud spreads; the backward picture tells you how long you wait. Reading the Fokker–Planck equation as the density-level twin of an ensemble of Langevin trajectories — the link owned by `langevin-equation` — ties the whole stochastic-methods chapter together.

Key takeaways

Viewed as a Markov process, the Fokker–Planck equation is the propagator of a continuous stochastic dynamics; the equation, Kramers–Moyal expansion, Pawula truncation, and Boltzmann stationary state are owned by `fokker-planck-equation` and recalled by name here.
It comes in two adjoint forms: the forward (Kolmogorov-forward) equation propagates densities in the final variable; the backward (Kolmogorov-backward) equation propagates observables/expectations in the initial variable.
The forward and backward operators $\mathcal L_{\rm FP}$ and $\mathcal L^\dagger_{\rm FP}$ are adjoints: $\int f\,(\mathcal L g)\,\mathrm{d}x=\int(\mathcal L^\dagger f)\,g\,\mathrm{d}x$ for decaying test functions.
The mean first-passage time $T(x)$ out of a domain solves the backward boundary-value problem $\mathcal L^\dagger_{\rm FP}\,T(x)=-1$ with absorbing conditions.
For free diffusion out of an interval $[0,L]$ the mean first-passage time is $T(x)=x(L-x)/2D$ — a closed number for given $L,D,x$ .
Key idea: The forward equation answers "where will the probability be?"; the backward equation answers "how long until it gets there?". Their adjoint pairing turns first-passage and escape-time questions into a one-line boundary-value problem $\mathcal L^\dagger T=-1$ .

The core insight

The forward Fokker–Planck equation answers "where will the probability be?"; its adjoint backward equation answers "how long until it gets there?". The adjoint pairing turns first-passage and escape-time questions into the boundary-value problem $\mathcal L^\dagger T=-1$ .

Where this lives today

The backward equation is the quantitative core of every escape-time problem in modern science. In chemical kinetics it delivers Kramers' reaction-rate theory; in neuroscience it gives the mean firing time of integrate-and-fire neurons driven by synaptic noise; in cell biology it sets the timescales of gene-expression switching and the search times of proteins diffusing to their DNA targets; in finance it prices barrier options through the first-passage statistics of the underlying. Numerically, the forward equation is integrated directly for low-dimensional problems, while for high-dimensional systems one samples the equivalent ensemble of Langevin trajectories (owned by `langevin-equation`) and estimates first-passage statistics by Monte-Carlo — the density and trajectory pictures traded off against dimension. The adjoint pair — forward for "where", backward for "how long" — is the fixed conceptual scaffolding underneath all of it.

Origins · 1 minstochastic differential equation and how this was discovered

stochastic differential equation

formalism, 20th c. · The mathematical object

\mathrm{d}X=a\,\mathrm{d}t+b\,\mathrm{d}W

driving a continuous Markov process, whose ensemble of solutions is described at the density level by the Fokker–Planck equation. Developed rigorously by Kolmogorov (forward and backward equations, 1931) and Itô (the stochastic integral, 1944), it links each noisy trajectory to the deterministic evolution of its probability density — the trajectory/density duality at the heart of this article.

History

1900

— Bachelier models market fluctuations with a diffusion equation

1905

— Einstein derives the diffusion equation for Brownian motion

1914

— Fokker writes the drift–diffusion equation for the density (Planck gives the general form in 1917)

1931

— Kolmogorov formulates the forward and backward analytic equations

1933

— Pontryagin, Andronov and Vitt solve first-passage problems via the backward equation

1940

— Kramers computes the barrier-escape rate from the diffusion equation

1944

— Itô constructs the stochastic integral, making the SDE picture rigorous

The physical picture

Hold two questions apart. First: I release a probability cloud at $x_0$ ; how does it drift and spread as time runs forward? That is the forward equation, and it acts on the final position — the cloud advects with the drift and blurs with the diffusion, exactly the dye-in-flowing-water image owned by `fokker-planck-equation`. Second: I care about some future event — hitting a wall, crossing a barrier — and I want its expected value as a function of where I started. That is the backward equation, and it acts on the initial position. The same dynamics, read from the two ends.

Figure

Forward spreads the cloud, backward maps the starting point to a waiting time

The backward equation is the natural home of first-passage problems because the question "how long until I hit the boundary?" is fundamentally a question about the starting point: two trajectories launched from different places have different expected escape times, and $T(x)$ is exactly that map. Its structure is remarkable — an ordinary source-term boundary-value problem, $\mathcal L^\dagger T=-1$ , whose right-hand side " $-1$ " simply counts one unit of elapsed time per unit time. The punchline figure shows the archetype: for diffusion out of an interval the waiting time is a downward parabola, longest from the middle, zero at the absorbing walls.

:::svg[Punchline — mean first-passage time out of [0,L] is a parabola peaking at the centre]

::: Underneath, this is the ensemble-of-Langevin-paths reading. Each realisation of the associated stochastic differential equation (owned by `langevin-equation`) is one wiggly trajectory; the forward equation is the exact deterministic law for the density of that ensemble, and expectations over the ensemble are precisely what the backward equation computes. The two Kolmogorov equations are the density-level face of trajectory-level noise.

The core equations

The forward operator generates density evolution. We display the standard drift–diffusion form only to fix notation and to name its adjoint; the derivation of this operator from the master equation via the Kramers–Moyal expansion is owned by `fokker-planck-equation`.

\frac{\partial P(x,t)}{\partial t}=\mathcal L_{\rm FP}\,P,\qquad \mathcal L_{\rm FP}=-\frac{\partial}{\partial x}A(x)+\frac12\frac{\partial^2}{\partial x^2}B(x) \qquad (1)

where: $P(x,t)$ — probability density; $A(x)$ — drift coefficient; $B(x)\ge0$ — diffusion coefficient; $\mathcal L_{\rm FP}$ — forward (Kolmogorov-forward) operator acting on densities. The backward operator is its formal adjoint. It acts on observables — functions of the initial state — and propagates their conditional expectations.

\frac{\partial Q(x,\tau)}{\partial\tau}=\mathcal L^\dagger_{\rm FP}\,Q,\qquad \mathcal L^\dagger_{\rm FP}=A(x)\frac{\partial}{\partial x}+\frac12 B(x)\frac{\partial^2}{\partial x^2} \qquad (2)

where: $Q(x,\tau)$ — expectation of an observable given start at $x$ , propagated over remaining time $\tau$ ; $\mathcal L^\dagger_{\rm FP}$ — backward (Kolmogorov-backward) operator; note the drift multiplies $\partial_x$ (not $\partial_x A$ ) — that is the adjoint. The two operators are adjoints under the flat $L^2$ inner product for decaying densities and observables: this is the identity that makes first-passage theory work.

\int f(x)\,\big[\mathcal L_{\rm FP}\,g\big](x)\,\mathrm{d}x=\int \big[\mathcal L^\dagger_{\rm FP}\,f\big](x)\,g(x)\,\mathrm{d}x \qquad (3)

where: $f,g$ — smooth, rapidly decaying test functions so that all boundary terms from integration by parts vanish. The mean first-passage time out of a domain solves the backward equation with a constant source. It is the single most useful application of the whole formalism.

\mathcal L^\dagger_{\rm FP}\,T(x)=-1,\qquad T=0 \text{ on absorbing boundaries} \qquad (4)

where: $T(x)$ — mean first-passage (exit) time starting from $x$ ; the source $-1$ counts one unit of elapsed time per unit time; absorbing (Dirichlet) conditions set $T=0$ where the particle is removed on contact. For free diffusion ( $A=0$ , $B=2D$ ) out of the interval $[0,L]$ with both ends absorbing, equation (4) integrates to a closed form — the parabola of the punchline figure.

T(x)=\frac{x(L-x)}{2D},\qquad T\Big(\frac L2\Big)=\frac{L^2}{8D} \qquad (5)

where: $x$ — starting position in $[0,L]$ ; $D$ — diffusion constant; $T$ is largest from the centre, zero at each wall.

The core derivations

Derivation AThe forward and backward operators are adjoints

The goal is to establish the pairing $\int f\,\mathcal L_{\rm FP}g=\int(\mathcal L^\dagger_{\rm FP}f)\,g$ by integration by parts, and to read off why the drift enters the two operators differently — inside the derivative for the forward (conservative, current) form, outside it for the backward (transport) form. This adjointness is not a formality: it is exactly what lets us convert a statement about densities (forward) into a statement about starting-point expectations (backward), which is the whole basis of first-passage and escape-time calculations.

Step 1. Write the forward action and pair it with a test function $f$ :

\int f\,\mathcal L_{\rm FP}g\,\mathrm{d}x=\int f\Big[-\partial_x(Ag)+\tfrac12\partial_x^2(Bg)\Big]\mathrm{d}x.

Why: insert the definition of $\mathcal L_{\rm FP}$ from equation (1); $g$ plays the role of a density, $f$ of an observable.

Step 2. Integrate the drift term by parts once and the diffusion term twice, discarding boundary terms:

\int f\big[-\partial_x(Ag)\big]\mathrm{d}x=\int (A\,\partial_x f)\,g\,\mathrm{d}x,\qquad \int f\,\tfrac12\partial_x^2(Bg)\,\mathrm{d}x=\int \tfrac12 B\,(\partial_x^2 f)\,g\,\mathrm{d}x.

Why: for rapidly decaying $f,g$ the surface terms vanish; each derivative moves off the $(Ag)$ or $(Bg)$ factor and onto $f$ , and the diffusion term's two integrations by parts leave $B$ un-differentiated multiplying $\partial_x^2 f$ .

Step 3. Collect the two results:

\int f\,\mathcal L_{\rm FP}g\,\mathrm{d}x=\int\Big[A\,\partial_x f+\tfrac12 B\,\partial_x^2 f\Big]g\,\mathrm{d}x=\int(\mathcal L^\dagger_{\rm FP}f)\,g\,\mathrm{d}x.

Why: the bracket is exactly $\mathcal L^\dagger_{\rm FP}f$ from equation (2) — drift outside the derivative, diffusion coefficient outside the second derivative. The adjoint is proven.

Step 4. Note the asymmetry: forward has $-\partial_x(A\,\cdot)$ (probability-current form, so it conserves $\int P$ ); backward has $A\,\partial_x$ (a transport/advection form acting on observables). Why: the forward operator must conserve total probability, which forces the drift inside the derivative; the backward operator propagates expectations, which do not need to be conserved, so the drift sits outside. Verified: the adjoint identity $\int f\mathcal L g=\int(\mathcal L^\dagger f)g$ is checked symbolically on an Ornstein–Uhlenbeck test case in verify_aq-sm-fokker-planck-equation-sm11.py.

Go deeper · Derivation B · 1 minMean first-passage time from the backward equation

Derivation BMean first-passage time from the backward equation

The goal is to show that the average exit time $T(x)$ solves $\mathcal L^\dagger T=-1$ with absorbing boundaries, and to integrate it explicitly for free diffusion to the closed parabola $T(x)=x(L-x)/2D$ . This is the payoff of the adjoint structure: a hard-looking question ("how long until escape?") becomes an elementary second-order boundary-value problem, and its solution is a real number for given parameters.

Step 1. Define $T(x)=\langle\text{time to first exit}\mid X(0)=x\rangle$ . Condition on the first infinitesimal step $x\to x+\mathrm{d}X$ over $\mathrm{d}t$ :

T(x)=\mathrm{d}t+\langle T(x+\mathrm{d}X)\rangle.

Why: the total exit time is the elapsed $\mathrm{d}t$ plus the remaining expected exit time from the new position, averaged over the stochastic step.

Step 2. Expand $\langle T(x+\mathrm{d}X)\rangle$ to second order using $\langle\mathrm{d}X\rangle=A\,\mathrm{d}t$ and $\langle\mathrm{d}X^2\rangle=B\,\mathrm{d}t$ :

\langle T(x+\mathrm{d}X)\rangle=T(x)+A\,T'(x)\,\mathrm{d}t+\tfrac12 B\,T''(x)\,\mathrm{d}t.

Why: these are the drift and diffusion jump moments of the process; higher moments are $O(\mathrm{d}t^2)$ and drop out.

Step 3. Substitute into Step 1 and cancel $T(x)$ ; divide by $\mathrm{d}t$ :

0=1+A\,T'(x)+\tfrac12 B\,T''(x)\ \Longrightarrow\ \mathcal L^\dagger_{\rm FP}\,T(x)=-1.

Why: the bracket $A\partial_x+\tfrac12 B\partial_x^2$ acting on $T$ is exactly $\mathcal L^\dagger_{\rm FP}T$ ; moving the $1$ across gives the source term.

Step 4. Solve for free diffusion ( $A=0,\ B=2D$ ) on $[0,L]$ with $T(0)=T(L)=0$ :

D\,T''(x)=-1\ \Rightarrow\ T(x)=\frac{x(L-x)}{2D},\qquad T\Big(\tfrac L2\Big)=\frac{L^2}{8D}.

Why: integrate twice; the two absorbing conditions fix the constants, yielding the symmetric parabola peaking at the midpoint. Verified: $\mathcal L^\dagger T=-1$ , the boundary values $T(0)=T(L)=0$ , and the midpoint value $T(L/2)=L^2/8D=1.163\,\mathrm{s}$ (L=2 µm, D=0.43 µm²/s) are all checked in verify_aq-sm-fokker-planck-equation-sm11.py.

Limiting cases

The two Kolmogorov equations specialise to the standard named equations depending on drift and diffusion; the underlying operator theory is owned by `fokker-planck-equation`.

Case	Result	Meaning
Free diffusion (forward)	$\partial_t P=D\,\partial_x^2 P$	Drift zero: the heat equation. Density spreads with variance $2Dt$ .
Free diffusion (backward, MFPT)	$\mathcal L^\dagger T=D\,T''=-1$	Exit-time parabola $T(x)=x(L-x)/2D$ on an absorbing interval.
Ornstein–Uhlenbeck (backward)	$\mathcal L^\dagger=-kx\,\partial_x+D\,\partial_x^2$	Restoring drift; expectations relax on time $1/k$ ; stationary variance $D/k$ .
High barrier (Kramers)	$\mathcal L^\dagger T=-1$ over a barrier $\Rightarrow T\sim e^{\Delta U/k_BT}$	Escape time is exponentially long in the barrier height — Arrhenius law from the backward equation.

Worked examples

Example 1 — Mean first-passage time of a colloid across a channel (a real number). A tracer with $D=0.43\ \mu\mathrm{m}^2/\mathrm{s}$ (a $1\ \mu$ m sphere in water at 300 K) diffuses freely in a channel of length $L=2\ \mu$ m with absorbing ends. How long, on average, does it take to escape when released from the centre?

Step 1. Free diffusion has $A=0,\ B=2D$ , so the backward equation is $D\,T''(x)=-1$ with $T(0)=T(L)=0$ ; the solution is $T(x)=x(L-x)/2D$ from equation (5).

Step 2. At the centre $x=L/2$ , $T(L/2)=\dfrac{(L/2)(L/2)}{2D}=\dfrac{L^2}{8D}$ .

Step 3. Plug numbers: $T=\dfrac{(2\ \mu\mathrm{m})^2}{8\,(0.43\ \mu\mathrm{m}^2/\mathrm{s})}=\dfrac{4}{3.44}\ \mathrm{s}=\mathbf{1.163\ s}$ . A particle released at the middle escapes in a little over one second on average; released nearer a wall it escapes faster, as the parabola shows.

Example 2 — Adjointness check on an Ornstein–Uhlenbeck process. For linear drift $A=-kx$ and constant diffusion $D$ , confirm that $\int f\,\mathcal L_{\rm FP}g\,\mathrm{d}x=\int(\mathcal L^\dagger_{\rm FP}f)\,g\,\mathrm{d}x$ using decaying Gaussians $f=e^{-x^2}$ , $g=x\,e^{-x^2}$ .

Step 1. Forward: $\mathcal L_{\rm FP}g=-\partial_x(-kx\,g)+D\,\partial_x^2 g=k\,\partial_x(xg)+D\,g''$ .

Step 2. Backward: $\mathcal L^\dagger_{\rm FP}f=-kx\,f'+D\,f''$ .

Step 3. Both integrals evaluate to the same value (the symbolic difference is zero), confirming the adjoint pairing; the stationary density of $\mathcal L_{\rm FP}$ is the Gaussian of variance $\mathbf{D/k}$ , whose zero-current condition $-kx\,P-D\,P'=0$ is satisfied identically. The number to remember: an OU process relaxes on timescale $1/k$ to a Gaussian of variance $D/k$ .

Going further

Texts. Pathria & Beale, Statistical Mechanics, 3rd ed., ch. 15 (Fluctuations and Brownian motion) · Risken, The Fokker–Planck Equation, 2nd ed., ch. 5–8 · Gardiner, Handbook of Stochastic Methods, ch. 5 (first-passage times) · van Kampen, Stochastic Processes in Physics and Chemistry, ch. VIII, XII · Redner, A Guide to First-Passage Processes · Kardar, Statistical Physics of Particles, ch. 9.

Provenance. The Fokker–Planck equation itself, the Kramers–Moyal expansion, Pawula's theorem, the drift/diffusion structure, and the Boltzmann stationary state are derived and owned by `fokker-planck-equation`; this article recalls them by name and does not reproduce those derivations. The forward/backward adjoint structure and the mean-first-passage-time boundary-value problem $\mathcal L^\dagger T=-1$ follow Gardiner ch. 5, Risken ch. 8, and Redner. A verifier should: (i) confirm the operator adjointness by integration by parts on a test case; (ii) solve $\mathcal L^\dagger T=-1$ on $[0,L]$ and recover $T(x)=x(L-x)/2D$ ; (iii) check the OU stationary variance $D/k$ . The Brownian number $D=0.43\ \mu\mathrm{m}^2/\mathrm{s}$ is the Stokes–Einstein value for a 1 μm sphere in water at 300 K.

Builds on: Markov Processes and the Markov Property · Chapman–Kolmogorov Equation · The Master Equation · Langevin Equation and Stochastic Differential Equations. Leads to: Fluctuation–Dissipation Theorem · Detailed Balance and Microscopic Reversibility · Monte Carlo Method · Kinetic Theory and the Boltzmann Equation.

If the Fokker–Planck equation is owned by another article, what does this one add?

The equation, its Kramers–Moyal derivation, Pawula's theorem, and the Boltzmann stationary state are owned by `fokker-planck-equation`. This article adds the process reading: the pair of adjoint Kolmogorov equations (forward for densities, backward for expectations), the adjointness identity itself, and — the practical payoff — mean-first-passage-time theory via $\mathcal L^\dagger T=-1$ . In short, the other article tells you what the equation is and where it comes from; this one tells you how to use it to answer "how long until escape?" and how it sits as the density twin of an ensemble of Langevin trajectories.

Why does the drift appear inside the derivative forward but outside it backward?

Probability conservation. The forward equation must keep $\int P\,\mathrm{d}x=1$ , and the only way a first-order term can conserve the integral is if it is a pure divergence, $-\partial_x(A P)$ — a probability current. The backward equation propagates expectations, which are not conserved quantities, so its drift term is a plain advection $A\,\partial_x$ . The difference is exactly the difference between an operator and its adjoint: integration by parts moves the derivative from acting on $(A\,\cdot)$ to acting on the test function, flipping the drift from inside to outside.

What does the source term "−1" in

\mathcal L^\dagger T=-1

mean physically?

It counts time. The mean first-passage time accumulates one unit of elapsed time for every unit of real time the particle survives inside the domain, so the "clock" contributes a constant rate of $+1$ per unit time; moving it to the other side of the backward operator gives the $-1$ source. Concretely, conditioning on one infinitesimal step, $T(x)=\mathrm{d}t+\langle T(x+\mathrm{d}X)\rangle$ , and the $\mathrm{d}t$ becomes the source. If instead you asked for the mean of some other accumulated observable $r(x)$ , the source would be $-r(x)$ .

How is this the "ensemble of Langevin paths"?

A single realisation of the stochastic differential equation $\mathrm{d}X=A\,\mathrm{d}t+\sqrt{B}\,\mathrm{d}W$ (the Langevin picture, owned by `langevin-equation`) is one jagged trajectory. Launch a whole ensemble of them from the same start and their instantaneous density obeys the forward Fokker–Planck equation exactly — that is the trajectory/density duality. Any expectation over the ensemble (an escape time, a hitting probability, a conditional average) is what the backward equation computes as a function of the launch point. So the two Kolmogorov equations are simply the deterministic, ensemble-level bookkeeping of the same noisy dynamics that the SDE runs one path at a time.

Does the backward equation give the Kramers escape rate?

Yes, in the high-barrier limit. Set up $\mathcal L^\dagger T=-1$ for a particle in a double well with absorbing conditions past the barrier. Because escape requires a rare thermal fluctuation over the barrier, the mean first-passage time is dominated by an Arrhenius factor $T\sim e^{\Delta U/k_BT}$ , and the escape rate is its reciprocal — Kramers' formula, with a friction- and curvature-dependent prefactor. The full Kramers prefactor and the turnover between friction regimes are developed in `fokker-planck-equation`; here the point is that the backward equation is the clean route to the exponential barrier-height dependence.

Can the first-passage problem be solved in more than one dimension?

The boundary-value problem $\mathcal L^\dagger T(\mathbf x)=-1$ with $T=0$ on the absorbing part of the boundary holds in any dimension — it is just an elliptic partial differential equation for $T(\mathbf x)$ . What changes is solvability: in one dimension you integrate twice and read off a closed form, but in higher dimensions closed forms exist only for special geometries (spheres, rectangles), and otherwise one solves the PDE numerically or samples Langevin trajectories. The conceptual structure — backward operator, constant source, absorbing boundary — is dimension-independent; only the ease of solution degrades.

What is the stationary variance of an Ornstein–Uhlenbeck process and why does it matter here?

For linear restoring drift $A=-kx$ and constant diffusion $D$ , the stationary density is a Gaussian of variance $D/k$ , obtained by setting the probability current to zero, $-kxP-DP'=0$ . It matters because the OU process is the linearised model near any potential minimum, so $D/k$ is the local equilibrium spread of essentially every confined stochastic system, and $1/k$ is the local relaxation time that sets how fast expectations (backward equation) approach their equilibrium values. It is the simplest non-trivial testbed for both Kolmogorov equations.