Stochastic Processes and the Fokker-Planck Equation

Stochastic processes describe systems evolving randomly in time; the Fokker-Planck equation (FPE) is the deterministic partial differential equation governing the time evolution of the probability density of a continuous Markov (diffusion) process. In one line: if a particle obeys the Itô SDE

d X_{t} = a d t + b d W_{t}

, then its density

p (x, t)

obeys

\partial_{t} p = - \partial_{x} (a p) + \frac{1}{2} \partial_{x}^{2} (b^{2} p)

. It applies wherever drift competes with diffusion — Brownian motion, chemical kinetics, neuron firing, option pricing, population genetics. Introduced by Fokker (1914) and Planck (1917); rigorous foundations from Kolmogorov (1931).

A stochastic process is a family of random variables ${X_{t}}_{t \in T}$ indexed by time on a common probability space. The most tractable and physically central class consists of Markov diffusion processes, whose transition probabilities are completely characterized by an infinitesimal drift and diffusion coefficient. The Fokker-Planck equation (equivalently the Kolmogorov forward equation) converts the stochastic dynamics of individual trajectories into a deterministic PDE for the ensemble probability density. This is the great conceptual pivot: random microscopic motion yields smooth, predictable macroscopic evolution of probabilities. A surprising fact is that the same operator, read in two adjoint directions, gives both the forward evolution of densities and the backward evolution of expectations — a duality that underlies the Feynman-Kac formula and modern mathematical finance.

The Core Insight

A continuous Markov process is locally Gaussian: over an infinitesimal time $d t$ , its increment has mean $a (x, t) d t$ and variance $b^{2} (x, t) d t$ , with all higher cumulants $o (d t)$ . These two functions — drift $a$ and diffusion $D = \frac{1}{2} b^{2}$ — encode everything. The Fokker-Planck equation is simply the statement of probability conservation (a continuity equation) with a probability current built from advection by the drift plus Fickian diffusion. Forward in time it propagates densities; backward in time its adjoint propagates expectations.

Formal Statement

Stochastic Process and Filtration

A stochastic process on a probability space $(Ω, F, P)$ is a map $X : T \times Ω \to R^{d}$ such that $X_{t} = X (t, \cdot)$ is measurable for each $t$ . A filtration ${F_{t}}$ is an increasing family of sub- $σ$ -algebras, $F_{s} \subseteq F_{t}$ for $s \leq t$ , representing information available up to time $t$ . The process is adapted if $X_{t}$ is $F_{t}$ -measurable.

A process is Markov if for $s < t$ and any Borel $B$ ,

P (X_{t} \in B ∣ F_{s}) = P (X_{t} \in B ∣ X_{s}) =: P (s, X_{s}; t, B) .

⟳This is the defining "memorylessness" — condition on the full past, recover dependence on the present alone. Define the transition kernel as the regular conditional probability.

Diffusion Process and the SDE

A diffusion is a continuous-path Markov process whose transition kernel satisfies, for any $ε > 0$ ,

h ↓ 0 lim \frac{1}{h} \int_{∣ y - x ∣ > ε} P (t, x; t + h, d y) = 0,

⟳This is the "no jumps" condition — the probability of a finite displacement in time

h

vanishes faster than

h

, forcing sample paths to be continuous (Lindeberg-type condition).

with drift and diffusion coefficients

a_{i} (x, t) = h ↓ 0 lim \frac{1}{h} E [(X_{t + h}^{i} - x) X_{t} = x], b_{ij} (x, t) = h ↓ 0 lim \frac{1}{h} E [(X_{t + h}^{i} - x^{i}) (X_{t + h}^{j} - x^{j}) X_{t} = x] .

⟳These are the first and second infinitesimal conditional moments; Taylor-expand the conditional moment generating function and keep

O (h)

terms.

Equivalently, $X_{t}$ solves the Itô stochastic differential equation

d X_{t} = a (X_{t}, t) d t + σ (X_{t}, t) d W_{t}, b_{ij} = k \sum σ_{ik} σ_{j k} = (σ σ^{⊤})_{ij},

⟳

W_{t}

is a

d

-dimensional Brownian motion. Match infinitesimal moments:

E [d X] = a d t

because

E [d W] = 0

, and

E [d X d X^{⊤}] = σ σ^{⊤} d t

using the Itô rule

d W_{i} d W_{j} = δ_{ij} d t

where $W_{t}$ is standard Brownian motion ( $E [W_{t}] = 0$ , $E [W_{s} W_{t}] = min (s, t)$ ).

Fokker-Planck (Kolmogorov Forward) Equation

Let $X_{t}$ solve the Itô SDE [4] with smooth, suitably bounded coefficients, and let $p (x, t)$ be the probability density of $X_{t}$ (assumed to exist). Then $p$ satisfies

\frac{\partial p}{\partial t} = - i = 1 \sum d \frac{\partial}{\partial x _{i}} [a_{i} (x, t) p] + \frac{1}{2} i, j = 1 \sum d \frac{\partial ^{2}}{\partial x _{i} \partial x _{j}} [b_{ij} (x, t) p] .

⟳Apply the adjoint of the generator to a test function and integrate by parts twice; the boundary terms vanish. See proof below.

In one dimension with $D (x, t) = \frac{1}{2} b (x, t)$ this reads

\frac{\partial p}{\partial t} = - \frac{\partial}{\partial x} [a (x, t) p] + \frac{\partial ^{2}}{\partial x ^{2}} [D (x, t) p] = - \frac{\partial J}{\partial x}, J = a p - \frac{\partial}{\partial x} (D p) .

⟳Rewrite [5] as a continuity equation

\partial_{t} p + \partial_{x} J = 0

by grouping all spatial derivatives into the probability current

J

The adjoint (Kolmogorov backward) equation for the transition density $p (y, t ∣ x, s)$ as a function of the initial data $(x, s)$ is

- \frac{\partial p}{\partial s} = i \sum a_{i} (x, s) \frac{\partial p}{\partial x _{i}} + \frac{1}{2} i, j \sum b_{ij} (x, s) \frac{\partial ^{2} p}{\partial x _{i} \partial x _{j}} =: L p .

⟳

L

is the infinitesimal generator; the forward operator in [5] is its formal

L^{2}

-adjoint

L^{†}

. Backward evolution acts on the starting point.

Why It Holds / Derivation

Forward Equation via the Generator and Itô's Lemma

Step 1 — The generator. For a

C_{b}^{2}

test function

f

, define the infinitesimal generator

(L f) (x, t) = h ↓ 0 lim \frac{E [ f ( X _{t + h} ) ∣ X _{t} = x ] - f ( x )}{h} .

⟳This is the time derivative of the conditional expectation; it measures how the average of

f

changes per unit time along trajectories.

Step 2 — Itô's lemma. For

f (X_{t})

with

X_{t}

solving [4], Itô's formula gives

df (X_{t}) = (i \sum a_{i} \partial_{i} f + \frac{1}{2} ij \sum b_{ij} \partial_{i} \partial_{j} f) d t + ik \sum \partial_{i} f σ_{ik} d W_{t}^{k} .

⟳Taylor-expand

f (X_{t + d t})

to second order; replace

d X_{i} d X_{j} \to b_{ij} d t

via the Itô multiplication table and discard

d t d W

and higher terms.

Taking conditional expectation kills the martingale term, so $(L f) (x, t) = \sum_{i} a_{i} \partial_{i} f + \frac{1}{2} \sum_{ij} b_{ij} \partial_{i} \partial_{j} f$ .

Step 3 — Duality. Evolution of the expectation

u (t) = E [f (X_{t})] = \int f (x) p (x, t) d x

satisfies

\frac{d}{d t} \int f p d x = \int (L f) p d x .

⟳Differentiate under the integral and use that

E [f (X_{t})]

advances by the generator; this is the weak form of the FPE.

Step 4 — Integrate by parts. Substituting

L f

and integrating by parts (boundary terms vanish for compactly supported

f

or fast-decaying

p

\int (L f) p d x = \int f [- \partial_{i} (a_{i} p) + \frac{1}{2} \partial_{i} \partial_{j} (b_{ij} p)] d x = \int f (L^{†} p) d x .

⟳Move one derivative off

f

in the drift term (one integration by parts) and two derivatives off

f

in the diffusion term (two integrations by parts), each producing a sign and a derivative landing on the coefficient times

p

Step 5 — Conclude. Since [10] equals

\int f \partial_{t} p d x

and must hold for all test

f

\partial_{t} p = L^{†} p = - \partial_{i} (a_{i} p) + \frac{1}{2} \partial_{i} \partial_{j} (b_{ij} p),

⟳Equate the integrands by the fundamental lemma of calculus of variations (arbitrariness of

f

■

Adriaan Fokker & Max Planck | 1914–1917

Adriaan Fokker (1887–1972), a Dutch physicist and musician (he built the 31-tone "Fokker organ"), derived in his 1914 doctoral work an equation for the probability distribution of a Brownian particle's velocity under radiation and collisions. Max Planck (1858–1947), founder of quantum theory, generalized and gave the equation rigorous standing in 1917 while studying fluctuation phenomena. Neither had the modern stochastic-calculus machinery; the rigorous probabilistic underpinning came only with Andrei Kolmogorov's 1931 paper, which is why physicists say "Fokker-Planck" while mathematicians often say "Kolmogorov forward equation." The equation crowned a decade of fluctuation physics opened by Einstein and Smoluchowski.

Building Intuition

Imagine an enormous swarm of identical particles, each independently obeying the same noisy law of motion. The Fokker-Planck equation does not track any single particle — it tracks the shape of the swarm's cloud. The drift term $- \partial_{x} (a p)$ is a conveyor belt: it bodily transports the cloud along the deterministic flow $\overset{x}{˙} = a$ , conserving area but shifting and shearing it. The diffusion term $\partial_{x}^{2} (D p)$ is a blurring lens: it spreads probability from regions of high density to low, exactly like heat smoothing out. Equilibrium is the standoff where the deterministic conveyor pushing probability "uphill" toward a potential minimum exactly balances the diffusive blur pushing it back out — yielding the Boltzmann distribution. The probability current $J$ in [6] is the bookkeeping that makes "probability is never created or destroyed, only moved" precise.

Drift-diffusion: a Gaussian packet advects and spreads while conserving area

Worked Examples

Free Brownian Motion — the Heat Kernel

Given:

d X_{t} = σ d W_{t}

(zero drift, constant

σ

), initial condition

X_{0} = x_{0}

, i.e.

p (x, 0) = δ (x - x_{0})

. Here

a = 0

D = \frac{1}{2} σ^{2}

. Step 1 — Write the FPE. From [6]:

\partial_{t} p = \frac{1}{2} σ^{2} \partial_{x}^{2} p

. This is the heat equation with diffusivity

\frac{1}{2} σ^{2}

. Step 2 — Fourier transform. With

\overset{p}{^} (k, t) = \int e^{- ik x} p d x

\partial_{t} \overset{p}{^} = - \frac{1}{2} σ^{2} k^{2} \overset{p}{^}

, giving

\overset{p}{^} (k, t) = e^{- ik x_{0}} e^{- σ^{2} k^{2} t /2}

. Step 3 — Invert. This is a Gaussian; transform back:

p (x, t) = \frac{1}{2 π σ ^{2} t} exp (- \frac{( x - x _{0} ) ^{2}}{2 σ ^{2} t}) .

⟳The inverse Fourier transform of a Gaussian in

k

is a Gaussian in

x

with variance

σ^{2} t

Step 4 — Numbers. Take

σ^{2} = 4 c m^{2} /s

x_{0} = 0

t = 2 s

. Variance

= σ^{2} t = 8 c m^{2}

, so spread

8 \approx 2.83 cm

p (x, 2) = \frac{1}{16 π} e^{- x^{2} /16} c m^{- 1}, std = 2.83 cm

Ornstein-Uhlenbeck Process — Relaxation to Equilibrium

Given:

d X_{t} = - γ X_{t} d t + σ d W_{t}

(linear restoring drift

a = - γ x

, constant

D = \frac{1}{2} σ^{2}

), with

γ > 0

. Model a particle in a harmonic trap or velocity of a Brownian particle (Langevin). Step 1 — FPE.

\partial_{t} p = γ \partial_{x} (x p) + \frac{1}{2} σ^{2} \partial_{x}^{2} p

. Step 2 — Gaussian ansatz. Means and variances close:

\overset{m}{˙} = - γ m

\dot{V} = - 2 γ V + σ^{2}

m (t) = m_{0} e^{- γ t}, V (t) = \frac{σ ^{2}}{2 γ} (1 - e^{- 2 γ t}) + V_{0} e^{- 2 γ t} .

⟳Multiply the FPE by

x

and

x^{2}

, integrate, and integrate by parts to get closed ODEs for the moments.

Step 3 — Stationary distribution. As

t \to \infty

m \to 0

V \to V_{\infty} = σ^{2} / (2 γ)

p_{st} (x) = \frac{γ}{π σ ^{2}} exp (- \frac{γ x ^{2}}{σ ^{2}}) .

⟳Set

\partial_{t} p = 0

and require zero probability current

J = 0

; solve the resulting first-order ODE, recovering a Boltzmann form

\propto e^{- U / D}

with

U = \frac{1}{2} γ x^{2}

Step 4 — Numbers. Take

γ = 2 s^{- 1}

σ^{2} = 8 (units)^{2} /s

m_{0} = 5

V_{0} = 0

. Relaxation time

τ = 1/ γ = 0.5 s

. Equilibrium variance

V_{\infty} = 8/4 = 2

m (0.5) = 5 e^{- 1} \approx 1.84, V (0.5) = 2 (1 - e^{- 2}) \approx 1.73, V_{\infty} = 2

Common Misconceptions

Itô vs. Stratonovich — the FPE depends on the calculus convention

When the diffusion coefficient is state-dependent ( $b = b (x)$ ), the SDE is ambiguous until you specify how the stochastic integral is interpreted. The Itô convention gives FPE [5] directly. The Stratonovich convention (denoted $\circ d W$ ) produces an extra "spurious drift": the Stratonovich SDE $d X = a d t + b \circ d W$ is equivalent to the Itô SDE with drift $a + \frac{1}{2} b b^{'}$ , so its FPE is

\partial_{t} p = - \partial_{x} [(a + \frac{1}{2} b b^{'}) p] + \frac{1}{2} \partial_{x}^{2} (b^{2} p) .

⟳Convert Stratonovich to Itô via the drift correction

\frac{1}{2} b \partial_{x} b

, then apply the standard FPE [6].

Failing to state the convention makes the equation ill-defined. Physicists modeling multiplicative noise from a smooth physical process should usually use Stratonovich (Wong-Zakai theorem); modelers wanting the martingale/no-anticipation property use Itô.

Not every Markov process has a Fokker-Planck equation

The FPE requires continuous paths and the vanishing-large-jump condition [2]. A Poisson process or any jump (Lévy) process is Markov but violates [2]: it makes finite jumps in arbitrarily short times. Its evolution is governed instead by a master equation / integro-differential Kolmogorov-Feller equation containing a jump kernel, not by a second-order PDE. Attempting to fit a pure-jump process with an FPE — the so-called Kramers-Moyal truncation at second order — is generally inconsistent: by Pawula's theorem, the Kramers-Moyal expansion must either stop at the second term (yielding the FPE) or contain infinitely many terms; truncating at any finite order $\geq 3$ can produce negative probabilities.

Connections

The FPE sits at a crossroads of analysis, probability, and physics. It is the forward dual of the backward Kolmogorov equation, which via the Feynman-Kac Formula connects to expectations and to the Schrödinger equation under Wick rotation. The associated SDE is the realm of Itô Calculus and Brownian Motion. In overdamped form (the Smoluchowski equation) it reduces from the full phase-space Kramers equation, the prototype of Kinetic Theory transport equations. Its stationary solutions reproduce equilibrium Statistical Mechanics (the Boltzmann distribution), while detailed balance vs. its violation is the gateway to stochastic thermodynamics and the Fluctuation-Dissipation Theorem. Numerically, the FPE is a parabolic PDE solvable by the same finite-difference and spectral tools as the Heat Equation, to which free diffusion literally reduces.

Real-World Applications

Mathematical finance (Black-Scholes-Merton). Asset prices modeled as geometric Brownian motion $d S = μ S d t + σ S d W$ ; the option-pricing PDE is the backward Kolmogorov equation, and risk-neutral densities solve the forward FPE. Used by every quant desk for derivative pricing.
Neuroscience (integrate-and-fire neurons). Membrane potential under synaptic noise obeys an OU-type FPE; population firing rates are computed from the stationary flux across the spiking threshold (Brunel's mean-field theory).
Chemical kinetics & systems biology. The chemical Fokker-Planck / Langevin equation approximates the chemical master equation for moderate molecule numbers, modeling gene-expression noise and reaction networks.
Plasma & accelerator physics. Velocity-space relaxation of charged particles via the Landau-Fokker-Planck collision operator; essential for tokamak transport and beam dynamics.
Population genetics. Wright-Fisher allele-frequency diffusion: $a$ from selection/mutation, $D = x (1 - x) /2 N$ ; the FPE predicts fixation probabilities and times (Kimura's diffusion theory).
Laser physics & quantum optics. The phase-space (Glauber-Sudarshan P or Wigner) representation maps quantum master equations onto FPEs for the field amplitude, explaining laser linewidth and photon statistics.
Climate & finance risk (rare events). First-passage and Kramers escape-rate theory compute mean exit times over barriers — molecular reaction rates, default times, climate tipping.

Historical Development

1827 — Robert Brown observes erratic motion of pollen grains in water; phenomenon unexplained for decades.
1900 — Louis Bachelier, in Théorie de la spéculation, models stock prices with Brownian motion and writes a diffusion equation — the first FPE, in finance, five years before Einstein.
1905 — Albert Einstein derives the diffusion equation for Brownian particles, linking diffusivity to temperature and molecular size ( $D = k_{B} T /6 π η a$ ).
1906 — Marian Smoluchowski independently develops the kinetic theory of Brownian motion and the overdamped (Smoluchowski) equation.
1908 — Paul Langevin introduces the stochastic differential equation viewpoint via a fluctuating force.
1914–1917 — Adriaan Fokker (thesis) and Max Planck establish the equation bearing their names for velocity/probability distributions.
1923 — Norbert Wiener gives the first rigorous mathematical construction of Brownian motion (the Wiener process/measure).
1931 — Andrei Kolmogorov publishes Über die analytischen Methoden in der Wahrscheinlichkeitsrechnung, deriving the forward and backward equations rigorously and founding the analytic theory of Markov processes.
1940 — Hendrik Kramers analyzes barrier escape (the Kramers problem) and the full phase-space equation; 1944–1951 Kiyosi Itô builds stochastic calculus, completing the SDE↔FPE correspondence.

Subtleties & Limits

Several deep issues persist. (1) Existence and smoothness of densities: Hörmander's hypoellipticity theorem guarantees a smooth density even for degenerate diffusion (rank-deficient $σ$ , as in the Kramers equation) provided a Lie-bracket spanning condition holds — but sharp conditions for general non-smooth or unbounded coefficients remain active research (the Stroock-Varadhan martingale problem, well-posedness for low-regularity drift à la DiPerna-Lions/Figalli). (2) Boundary conditions: absorbing, reflecting, and sticky boundaries each modify the current $J$ ; classifying boundary behavior (Feller's classification: regular, exit, entrance, natural) is delicate. (3) Long-time behavior and rates of convergence to equilibrium for non-reversible (current-carrying) FPEs are governed by hypocoercivity (Villani), where the operator is non-self-adjoint and spectral-gap arguments fail. (4) Non-Markovian generalizations: systems with memory require generalized Langevin equations and fractional Fokker-Planck operators whose probabilistic foundations are still incomplete. (5) High-dimensional numerics: solving the FPE directly suffers the curse of dimensionality; tensor methods and neural-network surrogates are open engineering frontiers.