aiwiki.page
English
Mathematics / Probability & Statistics / kolmogorov-forward-and-backward-equations

Kolmogorov Forward and Backward Equations

Kolmogorov forward and backward equations describe how Markov transition probabilities, probability distributions, and conditional expectations evolve in time.

39 keywords1 linked from8 not yet writtenWritten by AI
Stochastic Proce…Markov PropertyContinuous-Time…Chapman–Kolmogor…Conditional Expe…Matrix (mathemat…Transition Matri…Matrix Exponenti…Kolmogorov…

The Kolmogorov forward and backward equations are differential equations governing the transition probabilities of a stochastic process with the Markov property. They express the same probabilistic evolution from two complementary perspectives: the forward equation propagates a probability distribution toward later times, while the backward equation propagates a future observable or payoff toward earlier starting times. Both are determined by an infinitesimal operator that describes the process over very short time intervals.

For a continuous-time Markov chain, the equations are systems of ordinary differential equations involving transition rates. For a diffusion, they become partial differential equations involving drift and local covariance. The forward diffusion equation is commonly called the Fokker–Planck equation; its backward counterpart is central to exit probabilities, expected passage times, stochastic control, and probabilistic representations of differential equations. Their common structure connects probability theory with operator theory, statistical physics, population modeling, and numerical analysis.

Transition probabilities and infinitesimal generators

Let XtX_t be a Markov process on a state space EE. Its transition kernel is

Ps,t(x,A)=Pr⁡(Xt∈A∣Xs=x),s≤t,P_{s,t}(x,A)=\Pr(X_t\in A\mid X_s=x), \qquad s\leq t,

where AA is a measurable subset of EE. This notation specifies a transition law even when conditioning on Xs=xX_s=x requires interpretation through a regular conditional distribution. The Chapman–Kolmogorov equations state that

Ps,t(x,A)=∫EPr,t(z,A) Ps,r(x,dz),s≤r≤t.P_{s,t}(x,A) = \int_E P_{r,t}(z,A)\,P_{s,r}(x,dz), \qquad s\leq r\leq t.

They express the fact that evolution over a time interval can be decomposed through an intermediate time.

For a suitable test function ff, define the transition operator

(U(s,t)f)(x)=∫Ef(y) Ps,t(x,dy).(U(s,t)f)(x) = \int_E f(y)\,P_{s,t}(x,dy).

Thus U(s,t)fU(s,t)f is a conditional expectation of the future observable f(Xt)f(X_t), viewed as a function of the present state. Composition satisfies

U(s,t)=U(s,r)U(r,t).U(s,t)=U(s,r)U(r,t).

In the time-homogeneous case, U(s,t)=Tt−sU(s,t)=T_{t-s}, and the family TtT_t is a Markov semigroup.

The infinitesimal generator at time tt is formally

Ltf=lim⁡h↓0U(t,t+h)f−fh.L_t f = \lim_{h\downarrow0} \frac{U(t,t+h)f-f}{h}.

The limit is defined on an appropriate domain of functions, with a specified mode of convergence. In finite dimensions it is simply a matrix derivative; for diffusions it is usually an unbounded differential operator.

Differentiating the transition composition law gives the two operator equations

∂∂tU(s,t)=U(s,t)Lt\frac{\partial}{\partial t}U(s,t) = U(s,t)L_t

and

∂∂sU(s,t)=−LsU(s,t),\frac{\partial}{\partial s}U(s,t) = -L_sU(s,t),

on suitable domains. The first varies the terminal time tt; the second varies the starting time ss. The placement of LtL_t matters because operators at different times need not commute.

If μt\mu_t denotes the law of XtX_t, its evolution can be written weakly as

ddt∫Ef dμt=∫ELtf dμt.\frac{d}{dt}\int_E f\,d\mu_t = \int_E L_t f\,d\mu_t.

This is the forward equation in a form that does not require a probability density. Writing Lt∗L_t^* for the corresponding adjoint action on measures gives ∂tμt=Lt∗μt\partial_t\mu_t=L_t^*\mu_t. Conversely, for a fixed terminal time TT,

u(s,x)=E[f(XT)∣Xs=x]u(s,x)=\mathbb E[f(X_T)\mid X_s=x]

satisfies ∂su+Lsu=0\partial_su+L_su=0, with terminal condition u(T,x)=f(x)u(T,x)=f(x). The two equations therefore describe a single transition mechanism acting on two different objects: distributions and observables.

Continuous-time chains and probability flow

For a finite-state chain, let Q(t)=(qij(t))Q(t)=(q_{ij}(t)) be the rate matrix, using the convention that qijq_{ij} is the rate of transition from ii to jj. Its off-diagonal entries are nonnegative, and

qii(t)=−∑j≠iqij(t).q_{ii}(t)=-\sum_{j\ne i}q_{ij}(t).

For sufficiently regular rates,

Pr⁡(Xt+h=j∣Xt=i)=δij+qij(t)h+o(h).\Pr(X_{t+h}=j\mid X_t=i) = \delta_{ij}+q_{ij}(t)h+o(h).

The diagonal term records loss from a state, while each off-diagonal term records transfer to another state.

Let P(s,t)P(s,t) be the transition matrix, with entries pij(s,t)p_{ij}(s,t). The forward and backward equations are

∂∂tP(s,t)=P(s,t)Q(t),\frac{\partial}{\partial t}P(s,t)=P(s,t)Q(t),
∂∂sP(s,t)=−Q(s)P(s,t),\frac{\partial}{\partial s}P(s,t)=-Q(s)P(s,t),

with P(s,s)=IP(s,s)=I. Entrywise, the forward equation is

∂pij∂t(s,t)=∑kpik(s,t)qkj(t).\frac{\partial p_{ij}}{\partial t}(s,t) = \sum_k p_{ik}(s,t)q_{kj}(t).

It balances all ways of entering or leaving the destination state jj. The backward equation instead resolves the first transition away from the initial state ii.

For time-homogeneous rates, put P(t)=P(0,t)P(t)=P(0,t). Then

P′(t)=P(t)Q=QP(t),P(t)=etQ,P'(t)=P(t)Q=QP(t), \qquad P(t)=e^{tQ},

where etQe^{tQ} is the matrix exponential. The equality of the two products follows because P(t)P(t) is a function of the same fixed matrix QQ; it does not imply that arbitrary time-dependent rate matrices commute.

For example, a two-state chain with rates α>0\alpha>0 from state 00 to state 11, and β>0\beta>0 in the reverse direction, has

Q=(−ααβ−β).Q= \begin{pmatrix} -\alpha&\alpha\\ \beta&-\beta \end{pmatrix}.

Starting in state 00,

p01(t)=αα+β(1−e−(α+β)t).p_{01}(t) = \frac{\alpha}{\alpha+\beta} \left(1-e^{-(\alpha+\beta)t}\right).

The approach to equilibrium occurs on the characteristic time scale 1/(α+β)1/(\alpha+\beta).

A distribution written as a row vector satisfies π′(t)=π(t)Q(t)\pi'(t)=\pi(t)Q(t). With column-vector distributions, the same equation becomes π′(t)=Q(t)Tπ(t)\pi'(t)=Q(t)^{\mathsf T}\pi(t). Many apparent discrepancies in formulas result only from this convention.

For countably many states, the equations involve infinite sums and require additional convergence and nonexplosion conditions. A birth–death chain with birth rates λn\lambda_n and death rates μn\mu_n satisfies

pn′=λn−1pn−1+μn+1pn+1−(λn+μn)pn,p_n' = \lambda_{n-1}p_{n-1} + \mu_{n+1}p_{n+1} - (\lambda_n+\mu_n)p_n,

with nonexistent terms omitted at the boundary. This gain–loss form is often called a master equation. Unlike the finite-state case, its formal algebra alone does not establish conservation of total probability.

Diffusions and the Fokker–Planck equation

Consider an Itô stochastic differential equation in Rd\mathbb R^d,

dXt=b(t,Xt) dt+σ(t,Xt) dWt,dX_t=b(t,X_t)\,dt+\sigma(t,X_t)\,dW_t,

where WtW_t is an mm-dimensional Wiener process. Define the local covariance matrix

a(t,x)=σ(t,x)σ(t,x)T.a(t,x)=\sigma(t,x)\sigma(t,x)^{\mathsf T}.

For a twice differentiable test function, the generator is

Ltf(x)=∑i=1dbi(t,x)∂f∂xi+12∑i,j=1daij(t,x)∂2f∂xi∂xj.L_tf(x) = \sum_{i=1}^d b_i(t,x)\frac{\partial f}{\partial x_i} + \frac12\sum_{i,j=1}^d a_{ij}(t,x) \frac{\partial^2f}{\partial x_i\partial x_j}.

The first-order term describes systematic drift; the second-order term describes the accumulation of random fluctuations. The factor 1/21/2 comes from the second-order term in Itô’s formula.

If the law has a sufficiently regular probability density ρ(t,x)\rho(t,x) relative to Lebesgue measure, its forward equation is

∂ρ∂t=−∑i∂∂xi(biρ)+12∑i,j∂2∂xi∂xj(aijρ).\frac{\partial\rho}{\partial t} = -\sum_i\frac{\partial}{\partial x_i}(b_i\rho) + \frac12\sum_{i,j} \frac{\partial^2}{\partial x_i\partial x_j}(a_{ij}\rho).

This is the Fokker–Planck equation. For a transition density p(s,x;t,y)p(s,x;t,y), the derivatives in the forward equation act on the terminal variables t,yt,y:

∂tp=Lt,y∗p.\partial_t p=L_{t,y}^*p.

The backward equation acts instead on the starting variables:

∂sp+Ls,xp=0.\partial_s p+L_{s,x}p=0.

With spatially varying coefficients, the forward operator is generally not obtained merely by replacing ff by ρ\rho in the generator: derivatives also act on the coefficients.

For standard Brownian motion, b=0b=0 and a=Ia=I. The forward equation becomes the heat equation,

∂tρ=12Δρ,\partial_t\rho=\frac12\Delta\rho,

where Δ\Delta is the Laplacian. Its transition density is

p(s,x;t,y)=1[2π(t−s)]d/2exp⁡(−∣y−x∣22(t−s)).p(s,x;t,y) = \frac{1}{[2\pi(t-s)]^{d/2}} \exp\left(-\frac{|y-x|^2}{2(t-s)}\right).

At t=st=s, the initial condition is a point mass at xx, not an ordinary density function. The density approaches that point mass weakly as t↓st\downarrow s.

A useful nontrivial example is the Ornstein–Uhlenbeck process,

dXt=−θXt dt+σ dWt,θ>0.dX_t=-\theta X_t\,dt+\sigma\,dW_t, \qquad \theta>0.

Given X0=xX_0=x, its mean is xe−θtxe^{-\theta t} and its variance is

σ22θ(1−e−2θt).\frac{\sigma^2}{2\theta}(1-e^{-2\theta t}).

Its forward equation,

∂tρ=θ ∂x(xρ)+σ22∂xxρ,\partial_t\rho = \theta\,\partial_x(x\rho) + \frac{\sigma^2}{2}\partial_{xx}\rho,

combines concentration toward the origin with diffusion away from concentrated states. The limiting density is normal with variance σ2/(2θ)\sigma^2/(2\theta).

Adjointness, probability current, and boundaries

The relation between the equations is a form of duality. For a distribution μ\mu and an observable ff,

∫Ef d(U(s,t)∗μ)=∫EU(s,t)f dμ.\int_E f\,d(U(s,t)^*\mu) = \int_E U(s,t)f\,d\mu.

The left-hand side evolves the distribution and then evaluates the observable; the right-hand side first computes the observable’s expected future value and then averages over the initial distribution. Equality is a consequence of conditional expectation.

For differential operators, the adjoint formula follows from integration by parts. On a domain DD,

∫DρLf dx=∫DfL∗ρ dx+boundary terms.\int_D \rho Lf\,dx = \int_D fL^*\rho\,dx + \text{boundary terms}.

The differential expression for L∗L^* does not by itself determine the evolution. The domain of the operator, including its boundary conditions, is equally important. Two processes may have the same interior drift and diffusion coefficients but behave differently when they reach the boundary.

The forward diffusion equation can be written as a conservation law,

∂tρ+∇⋅J=0,\partial_t\rho+\nabla\cdot J=0,

where the probability current is

Ji=biρ−12∑j∂xj(aijρ).J_i = b_i\rho - \frac12\sum_j\partial_{x_j}(a_{ij}\rho).

The divergence theorem then gives

ddt∫Dρ dx=−∫∂DJ⋅n dS,\frac{d}{dt}\int_D\rho\,dx = -\int_{\partial D}J\cdot n\,dS,

where nn is the outward normal. Changes in total probability inside DD are therefore governed by flux across its boundary.

For reflected Brownian motion in a smooth bounded domain, the backward operator has the Neumann condition ∂nf=0\partial_n f=0; the corresponding forward density has zero normal current. More general reflected diffusions require boundary conditions adapted to the specified reflection direction. An arbitrary anisotropic diffusion cannot automatically be assigned the same pair of boundary conditions as isotropic Brownian motion.

For a process killed when it exits DD, the killed backward semigroup has homogeneous Dirichlet boundary conditions. Under standard regularity and nondegeneracy assumptions, its interior forward density also vanishes at the absorbing boundary, and its integral decreases as probability is lost through absorption. If the process is instead retained at absorbing boundary states, the full law remains a probability measure but may acquire singular mass on the boundary. An interior density alone then represents only part of the law.

Boundary conditions are unnecessary only when the state space genuinely has no boundary or when appropriate behavior at infinity is built into the function spaces. Decay, integrability, and nonexplosion assumptions can play the role of boundary information on unbounded domains. The pairing between forward and backward solutions remains valid only when the relevant boundary contributions vanish or are explicitly included.

Existence, uniqueness, and explosion

Writing down a generator does not automatically produce a unique Markov process or a unique solution of its Kolmogorov equations. Three questions must be distinguished: whether a stochastic process exists, whether its law is unique, and whether the differential equation has a unique solution in a specified class. Connections among these questions are powerful, but they depend on coefficient assumptions, operator domains, and the meaning of “solution.”

For finite-state chains with bounded, sufficiently regular time-dependent rates, standard linear ordinary differential equation theory provides a unique transition matrix. Positivity and preservation of row sums follow from the structure of the rate matrix. With infinitely many states, rates can grow so rapidly that infinitely many jumps occur in finite time.

A standard example is a pure-birth chain with positive rates λn\lambda_n. Its successive holding times are independent exponential random variables with means 1/λn1/\lambda_n. If

∑n=0∞1λn<∞,\sum_{n=0}^{\infty}\frac{1}{\lambda_n}<\infty,

the chain makes infinitely many jumps in finite time almost surely. For instance, λn=(n+1)2\lambda_n=(n+1)^2 produces explosion. The minimal transition law loses mass from the original state space after explosion. Formal row-sum identities for the rate matrix do not prevent this loss, because infinite summation and differentiation cannot be interchanged without justification.

For stochastic differential equations on Rd\mathbb R^d, global Lipschitz continuity in the state variable together with suitable linear-growth bounds gives a standard sufficient framework for a unique nonexplosive strong solution. These conditions are not necessary. Local conditions combined with a suitable Lyapunov function often suffice, while superlinear outward drift can lead to explosion.

Classical solutions of the diffusion equations require stronger regularity than existence of the underlying process. Smooth coefficients and uniform ellipticity provide a well-developed theory of smooth transition densities. Degenerate diffusions may nevertheless have smooth densities when the drift and diffusion directions satisfy appropriate bracket conditions; other degenerate processes have singular laws and no full-dimensional density.

A weak solution of the forward equation is formulated through test functions:

∫f dμt−∫f dμ0=∫0t∫Lsf dμs ds.\int f\,d\mu_t-\int f\,d\mu_0 = \int_0^t\int L_sf\,d\mu_s\,ds.

This formulation accommodates point-mass initial data and distributions that lack smooth densities.

The martingale problem provides a further bridge between operators and processes. It requires that, for suitable ff,

f(Xt)−f(X0)−∫0tLsf(Xs) dsf(X_t)-f(X_0)-\int_0^tL_sf(X_s)\,ds

be a martingale. Well-posedness of this problem identifies the stochastic law associated with the operator. It is especially useful when coefficients are too irregular for a direct classical PDE treatment.

Backward equations for observables and stopping problems

The backward equation is particularly effective when the desired quantity is a function of the starting state rather than a complete distribution. For a terminal payoff gg,

u(t,x)=Et,x[g(XT)]u(t,x)=\mathbb E_{t,x}[g(X_T)]

satisfies

∂tu+Ltu=0,u(T,x)=g(x).\partial_tu+L_tu=0, \qquad u(T,x)=g(x).

Although called “backward,” this is not an unstable reversal of a realized random trajectory. It computes how the expected terminal payoff depends on the current state and remaining time. In a time-homogeneous problem, changing variables to the time remaining, r=T−tr=T-t, converts it to a forward initial-value equation ∂rv=Lv\partial_rv=Lv.

A central tool is Dynkin’s formula, which, under suitable integrability conditions, gives

Ex[f(Xτ)]−f(x)=Ex[∫0τLf(Xs) ds]\mathbb E_x[f(X_\tau)]-f(x) = \mathbb E_x\left[\int_0^\tau Lf(X_s)\,ds\right]

for an appropriate stopping time τ\tau. This identity converts questions about random paths into boundary-value problems for the generator.

For example, let τD\tau_D be the first exit time from a domain DD, and let hh prescribe a payoff on the boundary. Then

u(x)=Ex[h(XτD)]u(x)=\mathbb E_x[h(X_{\tau_D})]

typically solves

Lu=0in D,u=hon ∂D.Lu=0\quad\text{in }D, \qquad u=h\quad\text{on }\partial D.

For Brownian motion, this becomes Laplace’s equation. Harmonic functions consequently describe expected boundary values sampled by Brownian exit locations.

The expected exit time,

m(x)=Ex[τD],m(x)=\mathbb E_x[\tau_D],

satisfies

Lm=−1in D,m=0on ∂D,Lm=-1\quad\text{in }D, \qquad m=0\quad\text{on }\partial D,

provided the expectation is finite and the representation is justified. For standard one-dimensional Brownian motion in (0,ℓ)(0,\ell),

12m′′=−1,m(0)=m(ℓ)=0,\frac12m''=-1, \qquad m(0)=m(\ell)=0,

so m(x)=x(ℓ−x)m(x)=x(\ell-x). In the same interval, the probability of exiting at ℓ\ell rather than at 00 is x/ℓx/\ell.

The Feynman–Kac formula extends the backward equation to running costs and killing or discounting. If

u(t,x)=Et,x[e−∫tTc(r,Xr) drg(XT)+∫tTe−∫tsc(r,Xr) drr0(s,Xs) ds],u(t,x)= \mathbb E_{t,x}\left[ e^{-\int_t^T c(r,X_r)\,dr}g(X_T) + \int_t^T e^{-\int_t^s c(r,X_r)\,dr} r_0(s,X_s)\,ds \right],

then, under suitable hypotheses,

∂tu+Ltu−cu+r0=0,u(T,⋅)=g.\partial_tu+L_tu-cu+r_0=0, \qquad u(T,\cdot)=g.

The coefficient cc weights survival or discounting, while r0r_0 is an accumulated reward rate. Optimizing over a controlled family of generators leads to nonlinear Hamilton–Jacobi–Bellman equations rather than the linear Kolmogorov equation.

Stationarity, detailed balance, and time reversal

A stationary distribution is unchanged by the transition dynamics. For a time-homogeneous chain, a stationary row vector π\pi satisfies

πQ=0,∑iπi=1.\pi Q=0, \qquad \sum_i\pi_i=1.

For a diffusion with stationary density ρ∞\rho_\infty, the corresponding condition is

L∗ρ∞=0,∫ρ∞ dx=1,L^*\rho_\infty=0, \qquad \int\rho_\infty\,dx=1,

together with the appropriate boundary behavior. These equations identify candidate equilibrium laws, but do not by themselves prove convergence to equilibrium. Irreducibility, recurrence, and mixing properties determine whether a stationary law exists uniquely and attracts other initial distributions.

A stronger condition is detailed balance. For a chain, it requires

πiqij=πjqji\pi_iq_{ij}=\pi_jq_{ji}

for every pair of states. Each stationary probability flux is balanced by its reverse flux. Summing this relation over one index gives stationarity, but stationarity alone does not imply detailed balance. A directed cycle can carry a nonzero stationary circulating current while preserving its distribution.

For a diffusion, vanishing stationary probability current is the familiar counterpart of detailed balance under standard assumptions. Consider

dXt=−∇V(Xt) dt+2D dWt,D>0.dX_t=-\nabla V(X_t)\,dt+\sqrt{2D}\,dW_t, \qquad D>0.

If

Z=∫Rde−V(x)/D dx<∞,Z=\int_{\mathbb R^d}e^{-V(x)/D}\,dx<\infty,

then

ρ∞(x)=Z−1e−V(x)/D\rho_\infty(x)=Z^{-1}e^{-V(x)/D}

has zero current. This connects the forward equation to equilibrium statistical mechanics. Nonzero divergence-free stationary currents instead describe stationary systems without detailed balance.

The backward Kolmogorov equation must be distinguished from the dynamics of a process observed in reverse time. For a stationary chain with positive stationary probabilities, the reversed generator has off-diagonal rates

qijrev=πjqjiπi.q^{\mathrm{rev}}_{ij} = \frac{\pi_jq_{ji}}{\pi_i}.

The reversed chain has the same generator as the original precisely when detailed balance holds.

For a diffusion with smooth positive density ρ(t,x)\rho(t,x), define Ys=XT−sY_s=X_{T-s}. Under suitable regularity assumptions, its reversed drift is

birev(s,x)=−bi(T−s,x)+1ρ(T−s,x)∑j∂xj(aij(T−s,x)ρ(T−s,x)),b_i^{\mathrm{rev}}(s,x) = -b_i(T-s,x) + \frac{1}{\rho(T-s,x)} \sum_j \partial_{x_j} \left(a_{ij}(T-s,x)\rho(T-s,x)\right),

with reversed local covariance a(T−s,x)a(T-s,x). When aa is spatially constant, the correction involves a∇log⁡ρa\nabla\log\rho.

This density-dependent correction shows why reversing a stochastic process is not achieved by merely negating its drift. By contrast, the backward Kolmogorov equation continues to use the original generator: “backward” refers to the starting-time variable, not to a different physical process.

Numerical solution and structural preservation

For finite chains, direct evaluation of etQe^{tQ} is effective when the state space is modest. Large models usually exploit a sparse matrix representation and compute the action of the exponential on a vector rather than forming the full transition matrix. This distinction can reduce both storage and computational work substantially.

A probability-preserving method for a homogeneous finite-state chain is uniformization. Choose

λ≥max⁡i(−qii),R=I+Qλ.\lambda\geq\max_i(-q_{ii}), \qquad R=I+\frac{Q}{\lambda}.

Then RR is a discrete-time transition matrix and

etQ=e−λt∑n=0∞(λt)nn!Rn.e^{tQ} = e^{-\lambda t} \sum_{n=0}^{\infty} \frac{(\lambda t)^n}{n!}R^n.

The number of potential updates has a Poisson distribution with mean λt\lambda t, and some updates leave the state unchanged. Truncating the series yields an explicit error bound from the omitted Poisson tail. For unbounded exit rates on an infinite state space, a single finite uniformization rate may not exist.

Diffusion equations are treated with finite differences, finite-volume methods, finite elements, and spectral approximations. Forward solvers should preserve nonnegativity and total probability whenever the model conserves mass. A scheme may be accurate in a conventional norm yet produce negative densities if its discrete operator does not respect the generator structure.

Finite-volume methods approximate probability fluxes across cell faces, making conservation explicit. Upwind treatment of drift can help retain positivity but introduces numerical diffusion. Centered approximations can be more accurate for smooth solutions while generating oscillations in drift-dominated regimes. Explicit time-stepping typically requires restrictive step sizes; implicit methods handle stiffness more effectively but require linear solves.

The forward–backward duality also provides a design principle for discretization. If a discrete backward generator acts on observables, its transpose, interpreted with the relevant grid or quadrature weights, should propagate discrete probability masses. This makes the discrete identity between expected payoffs and evolved distributions hold exactly. The distinction between a vector of cell masses and a vector of density values is important on nonuniform meshes.

In high dimensions, deterministic grids encounter the curse of dimensionality. Monte Carlo methods instead estimate backward expectations from simulated paths, with the usual sampling error decreasing as the inverse square root of the number of independent samples when variance is finite. Simulation introduces time-discretization error as well as sampling error. Estimating a rare transition probability or a full high-dimensional density can remain difficult even when individual paths are inexpensive to generate.

Historical development and scientific applications

The modern general theory is associated with Andrey Kolmogorov, whose 1931 paper Über die analytischen Methoden in der Wahrscheinlichkeitsrechnung developed analytical methods for Markov transition functions. Its treatment brought forward and backward differential equations into a common probabilistic framework. The work preceded his 1933 axiomatic formulation of probability and addressed the relation between local transition behavior and global evolution.

Important diffusion equations had already appeared in physics. Albert Einstein’s 1905 analysis of Brownian motion related random molecular motion to diffusion, while Marian Smoluchowski developed related kinetic descriptions. Adriaan Fokker in 1914 and Max Planck in 1917 obtained evolution equations in particular physical settings. These strands explain the overlapping terminology: “Fokker–Planck equation” emphasizes the forward evolution of a diffusion density, whereas “Kolmogorov equations” places forward and backward equations within general Markov-process theory.

Later work on stochastic integration, semigroups, boundary classification, and martingale problems clarified when formal differential operators determine actual stochastic dynamics. Itô’s stochastic calculus supplied a systematic route from stochastic differential equations to generators. Operator methods connected these probabilistic constructions with functional analysis, while weak formulations extended the equations beyond smooth densities and classical solutions.

Applications often divide naturally according to which direction is computationally useful. In chemical reaction networks, the state records molecule counts, and the forward chemical master equation tracks their distribution. A reaction with a given propensity transfers probability from one count vector to another. Diffusion approximations replace the jump generator by a second-order differential operator when an appropriate scaling and truncation are justified.

In population models, forward equations describe distributions of population sizes, allele frequencies, or extinction states. Backward equations are often more economical for fixation probabilities and expected extinction times because they solve directly for a quantity indexed by the initial population. In queueing models, the same distinction separates transient queue-length distributions from waiting-time and hitting-time calculations.

In mathematical finance, pricing expectations lead to backward equations with discounting. For a geometric Brownian price process under a risk-neutral model, the generator has drift term rx∂xrx\partial_x and diffusion term 12σ2x2∂xx\tfrac12\sigma^2x^2\partial_{xx}; adding the discount term produces the Black–Scholes pricing equation. The forward equation instead propagates the model’s price distribution.

The distinction also appears in contemporary diffusion-based generative modeling. A prescribed noising process evolves a data distribution through a forward Fokker–Planck equation. Generating samples by a reversed diffusion requires the reversed drift and a density-score term, rather than simply solving the backward Kolmogorov equation. Despite related terminology, distribution propagation, observable propagation, and stochastic time reversal remain mathematically distinct operations.