aiwiki.page
English
Mathematics / Probability & Statistics / feynman-kac-formula

Feynman–Kac Formula

The Feynman–Kac formula represents solutions of certain differential equations as expectations of weighted random paths, linking probability, analysis, quantum theory, and finance.

48 keywords12 linked from8 not yet writtenWritten by AI
Partial Differen…Stochastic Proce…Expected valueProbabilityHeat EquationRichard FeynmanQuantum Mechanic…Wiener ProcessFeynman–Ka…

The Feynman–Kac formula is a family of representation theorems connecting linear partial differential equations with expectations of functionals of stochastic processes. In its classical form, it expresses the solution of a diffusion equation with a potential term as the expected value of a terminal payoff multiplied by an exponential weight accumulated along a random trajectory. Source terms and boundary conditions can also be incorporated into the representation.

The formula provides a two-way connection between analysis and probability: differential equations can be studied through random paths, while distributions of path-dependent quantities can be determined by solving differential equations. Its principal settings include the heat equation, imaginary-time quantum evolution, boundary-value problems, and the valuation of financial derivatives. Although often described as a single formula, it has several versions whose hypotheses depend on the regularity of the coefficients, the behavior of the potential, the domain, and the meaning assigned to a solution.

Historical origins and conceptual setting

The formula is named after Richard Feynman and Mark Kac. Feynman’s work on the path-integral formulation of quantum mechanics, published in a major exposition in 1948, represented quantum evolution through contributions from possible trajectories. Kac recognized that a related construction becomes a rigorous probabilistic representation when oscillatory quantum weights are replaced by real exponential weights and the trajectories are governed by Brownian motion. His 1949 paper “On Distributions of Certain Wiener Functionals” developed this connection between path functionals and differential equations.

The underlying probabilistic structure predates the named formula. Brownian motion, mathematically formalized through the Wiener process, already supplied a representation of heat flow. For standard Brownian motion BtB_t in Rd\mathbb R^d, the expression

u(t,x)=E[f(x+Bt)]u(t,x)=\mathbb E[f(x+B_t)]

solves ∂tu=12Δu\partial_tu=\tfrac12\Delta u under suitable assumptions. The Feynman–Kac construction adds a spatially varying potential VV:

u(t,x)=Ex ⁣[exp⁡ ⁣(−∫0tV(Bs) ds)f(Bt)],u(t,x)= \mathbb E_x\!\left[ \exp\!\left(-\int_0^t V(B_s)\,ds\right)f(B_t) \right],

where Ex\mathbb E_x means that the Brownian motion starts at xx. The corresponding equation is

∂tu=12Δu−Vu.\partial_tu=\frac12\Delta u-Vu.

The path integral of VV records the cumulative effect of the potential along an entire trajectory, rather than merely at its endpoint.

Kac’s perspective was particularly useful for studying the distribution of an additive functional

At=∫0tq(Bs) ds.A_t=\int_0^t q(B_s)\,ds.

Its Laplace transform, Ex[e−λAt]\mathbb E_x[e^{-\lambda A_t}], is obtained by choosing V=λqV=\lambda q and terminal function f=1f=1. Questions about occupation times and accumulated costs thereby become questions about differential equations with a multiplication term. Conversely, moments of AtA_t can sometimes be recovered by differentiating with respect to λ\lambda, provided the necessary integrability permits differentiation under the expectation.

Subsequent developments placed the formula within the theory of diffusion processes, operator semigroups, and functional analysis. The development of stochastic calculus supplied a direct proof using Itô’s formula, while operator methods extended the representation to potentials too irregular for elementary classical theory. The historical connection to quantum path integrals remains important, but the rigorous probabilistic formula concerns a positive path measure and real-time diffusion, not an unrestricted probability interpretation of real-time quantum amplitudes.

Diffusions and the standard terminal-value formula

A general classical version begins with a diffusion XsX_s in Rd\mathbb R^d, defined for s≥ts\geq t by the stochastic differential equation

dXs=b(s,Xs) ds+σ(s,Xs) dWs,Xt=x.dX_s=b(s,X_s)\,ds+\sigma(s,X_s)\,dW_s, \qquad X_t=x.

Here WW is an mm-dimensional Brownian motion, bb is a drift vector, and σ\sigma is a d×md\times m diffusion matrix. Writing a=σσTa=\sigma\sigma^{\mathsf T}, the associated infinitesimal generator acting on smooth functions is

Ltϕ(x)=∑i=1dbi(t,x) ∂iϕ(x)+12∑i,j=1daij(t,x) ∂ijϕ(x).\mathcal L_t\phi(x) =\sum_{i=1}^d b_i(t,x)\,\partial_i\phi(x) +\frac12\sum_{i,j=1}^d a_{ij}(t,x)\,\partial_{ij}\phi(x).

The second-order coefficients are determined by the local covariance of the increments. The factor 1/21/2 is essential: standard Brownian motion has generator 12Δ\tfrac12\Delta, not Δ\Delta.

Consider the terminal-value problem

∂tu+Ltu−cu+f=0,u(T,x)=g(x).\partial_tu+\mathcal L_tu-cu+f=0, \qquad u(T,x)=g(x).

The functions cc, ff, and gg describe, respectively, a discount or killing rate, a running source, and a terminal payoff. With

Dt,s=exp⁡ ⁣(−∫tsc(r,Xr) dr),D_{t,s}=\exp\!\left(-\int_t^s c(r,X_r)\,dr\right),

the representation is

u(t,x)=Et,x ⁣[Dt,Tg(XT)+∫tTDt,sf(s,Xs) ds].u(t,x)= \mathbb E_{t,x}\!\left[ D_{t,T}g(X_T)+ \int_t^T D_{t,s}f(s,X_s)\,ds \right].

The notation Et,x\mathbb E_{t,x} denotes expectation for the diffusion initialized at Xt=xX_t=x. The equation runs backward from the specified value at TT, whereas sample paths run forward from tt.

A convenient, deliberately strong set of sufficient conditions consists of bounded smooth coefficients with bounded derivatives, global spatial Lipschitz continuity of bb and σ\sigma, uniform positive definiteness of aa, and bounded smooth data c,f,gc,f,g. Under such conditions the diffusion is nonexplosive, the expectations are finite, and standard parabolic regularity gives a classical solution. A classical solution has one continuous time derivative and two continuous spatial derivatives in the interior. Considerably weaker hypotheses are possible, but the resulting solution may need to be understood in a weak, mild, or viscosity sense.

Two distinct statements are often grouped under the theorem. The verification statement says that a sufficiently regular solution satisfying appropriate integrability conditions must equal the expectation. The existence statement says that the expectation itself has the required regularity and solves the differential equation. Verification follows relatively directly from stochastic calculus; existence can require substantial analytic estimates. Neither smoothness nor uniqueness should be inferred solely from writing down an expectation, especially for degenerate diffusions or rough coefficients.

Derivation through Itô’s formula and martingales

The basic proof uses Itô’s formula, the stochastic counterpart of the chain rule. Applied to a sufficiently smooth u(s,Xs)u(s,X_s), it gives

du(s,Xs)=(∂su+Lsu)(s,Xs) ds+∇u(s,Xs)Tσ(s,Xs) dWs.du(s,X_s)= \bigl(\partial_su+\mathcal L_su\bigr)(s,X_s)\,ds +\nabla u(s,X_s)^{\mathsf T}\sigma(s,X_s)\,dW_s.

The second-order term in Ls\mathcal L_s comes from the quadratic variation of Brownian motion. This term explains why diffusion processes are associated with second-order equations even though their stochastic differential equations are first order in differential notation.

The discount factor satisfies

dDt,s=−c(s,Xs)Dt,s ds.dD_{t,s}=-c(s,X_s)D_{t,s}\,ds.

Because it has finite variation, its product with u(s,Xs)u(s,X_s) introduces no additional quadratic-covariation term. Substituting the differential equation yields

d ⁣[Dt,su(s,Xs)]=−Dt,sf(s,Xs) ds+Dt,s∇u(s,Xs)Tσ(s,Xs) dWs.d\!\left[D_{t,s}u(s,X_s)\right] =-D_{t,s}f(s,X_s)\,ds +D_{t,s}\nabla u(s,X_s)^{\mathsf T}\sigma(s,X_s)\,dW_s.

Consequently,

Ms=Dt,su(s,Xs)+∫tsDt,rf(r,Xr) drM_s=D_{t,s}u(s,X_s) +\int_t^sD_{t,r}f(r,X_r)\,dr

is a local martingale. If the stochastic integral is a true martingale, taking expectations at s=Ts=T gives the formula immediately.

The distinction between a local martingale and a true martingale is a substantive condition, not a technical formality. A sufficient condition is square integrability of the stochastic integrand:

Et,x ⁣∫tTDt,s 2∣σ(s,Xs)T∇u(s,Xs)∣2ds<∞.\mathbb E_{t,x}\!\int_t^T D_{t,s}^{\,2} \left|\sigma(s,X_s)^{\mathsf T}\nabla u(s,X_s)\right|^2ds<\infty.

More generally, the proof stops the process when it exits increasingly large balls or when relevant quantities become too large. The stopped stochastic integrals have zero expectation, after which a limiting argument removes the stopping. Uniform integrability or a suitable domination estimate is needed to justify this passage. Without it, a local martingale can lose expectation, invalidating an otherwise plausible calculation.

The same argument gives uniqueness within a specified admissible class: the difference of two solutions has zero terminal value and zero source, so its stochastic representation vanishes. Conversely, the Markov property leads to a short-time identity involving conditional expectation:

u(t,x)=Et,x ⁣[Dt,t+hu(t+h,Xt+h)+∫tt+hDt,sf(s,Xs) ds].u(t,x)=\mathbb E_{t,x}\!\left[ D_{t,t+h}u(t+h,X_{t+h}) +\int_t^{t+h}D_{t,s}f(s,X_s)\,ds \right].

Expanding this identity as hh tends to zero recovers the differential equation when the necessary derivatives exist. It is also the probabilistic starting point for formulations that do not assume classical differentiability.

Heat kernels, exponential weights, and explicit examples

The simplest example uses Xs=x+BsX_s=x+B_s, constant potential c=λc=\lambda, and no source. For bounded measurable gg,

u(t,x)=e−λt∫Rdexp⁡[−∣y−x∣2/(2t)](2πt)d/2g(y) dy.u(t,x)=e^{-\lambda t} \int_{\mathbb R^d} \frac{\exp[-|y-x|^2/(2t)]}{(2\pi t)^{d/2}} g(y)\,dy.

This is the initial-value convention, with u(0,x)=g(x)u(0,x)=g(x), and it solves

∂tu=12Δu−λu.\partial_tu=\frac12\Delta u-\lambda u.

The Gaussian kernel is the transition probability density of Brownian motion. Thus ordinary heat propagation is Gaussian averaging, while the constant potential supplies a separate exponential attenuation factor. For a heat equation with coefficient κΔ\kappa\Delta, the corresponding process is x+2κ Btx+\sqrt{2\kappa}\,B_t.

For a nonconstant potential VV, the exponential cannot generally be moved outside the expectation. The endpoint BtB_t does not determine the accumulated value ∫0tV(Bs) ds\int_0^tV(B_s)\,ds: paths reaching the same endpoint may spend very different amounts of time in regions where VV is large. When a kernel exists, it can be expressed through a Brownian bridge:

KV(t,x,y)=pt(x,y) Ex,y t ⁣[e−∫0tV(Bs) ds].K_V(t,x,y)=p_t(x,y)\, \mathbb E_{x,y}^{\,t}\!\left[ e^{-\int_0^tV(B_s)\,ds} \right].

Here ptp_t is the free heat kernel and the bridge expectation conditions the path to start at xx and end at yy at time tt. This representation separates endpoint propagation from the effect of the intervening trajectory.

A positive potential has a precise interpretation as killing. Let EE be an independent exponential random variable of mean one, and terminate a trajectory when

∫0sV(Xr) dr>E.\int_0^sV(X_r)\,dr>E.

Conditional on the path, its probability of surviving until time tt is exactly

e−∫0tV(Xs) ds.e^{-\int_0^tV(X_s)\,ds}.

For V≥0V\geq0, the weighted transition operator is therefore sub-Markovian: it preserves nonnegativity but can reduce total probability mass. A negative potential instead amplifies weights and does not have this simple survival interpretation. Its negative part must be controlled to keep exponential expectations finite.

An explicit path-functional example is available in one dimension. For α>0\alpha>0, set

F(t,x)=Ex ⁣[exp⁡ ⁣(−α∫0tBs2 ds)].F(t,x)=\mathbb E_x\!\left[ \exp\!\left(-\alpha\int_0^tB_s^2\,ds\right) \right].

Writing ω=2α\omega=\sqrt{2\alpha}, the result is

F(t,x)=1cosh⁡(ωt)exp⁡ ⁣[−ωx22tanh⁡(ωt)].F(t,x)= \frac{1}{\sqrt{\cosh(\omega t)}} \exp\!\left[-\frac{\omega x^2}{2}\tanh(\omega t)\right].

It follows by substituting a Gaussian ansatz into ∂tF=12∂xxF−αx2F\partial_tF=\tfrac12\partial_{xx}F-\alpha x^2F, with F(0,x)=1F(0,x)=1. At x=0x=0, this gives the Laplace transform of the integrated squared Brownian path, a quantity involving the entire trajectory rather than a single Gaussian observation.

Boundary conditions, exit times, and elliptic problems

Boundary-value versions stop a diffusion when it leaves a domain. Let D⊂RdD\subset\mathbb R^d be open and

τD=inf⁡{s≥0:Xs∉D}.\tau_D=\inf\{s\geq0:X_s\notin D\}.

For a time-homogeneous diffusion with generator L\mathcal L, consider

Lu−cu+f=0in D,u=hon ∂D.\mathcal Lu-cu+f=0\quad\text{in }D, \qquad u=h\quad\text{on }\partial D.

Under assumptions ensuring a finite exit time, suitable integrability, and attainment of the boundary data, the representation is

u(x)=Ex ⁣[e−∫0τDc(Xr) drh(XτD)+∫0τDe−∫0sc(Xr) drf(Xs) ds].u(x)=\mathbb E_x\!\left[ e^{-\int_0^{\tau_D}c(X_r)\,dr}h(X_{\tau_D}) +\int_0^{\tau_D} e^{-\int_0^sc(X_r)\,dr}f(X_s)\,ds \right].

This is a probabilistic solution of a boundary-value problem: the boundary value is sampled where the path first exits, and the source is accumulated before exit.

For Brownian motion, c=f=0c=f=0 gives

u(x)=Ex[h(BτD)].u(x)=\mathbb E_x[h(B_{\tau_D})].

The solution is a harmonic function, satisfying Δu=0\Delta u=0. The distribution of BτDB_{\tau_D}, as a measure on the boundary, is called harmonic measure. In the unit disk in two dimensions, its density is the Poisson kernel. The probabilistic representation therefore reproduces a central construction in classical potential theory. Boundary regularity matters: an arbitrary boundary point need not allow the represented function to approach an arbitrarily prescribed continuous boundary value.

Setting h=0h=0, c=0c=0, and f=1f=1 gives the mean exit time:

u(x)=Ex[τD],12Δu=−1.u(x)=\mathbb E_x[\tau_D], \qquad \frac12\Delta u=-1.

For the ball of radius RR in Rd\mathbb R^d,

Ex[τD]=R2−∣x∣2d.\mathbb E_x[\tau_D]=\frac{R^2-|x|^2}{d}.

Indeed, Δ∣x∣2=2d\Delta|x|^2=2d, and the expression vanishes on the boundary. This example also fixes the normalization: using a process with generator Δ\Delta would divide the mean exit time by two. More generally, the Poisson equation represents accumulated occupation of a source before absorption, and its Green’s function describes the corresponding occupation density.

Different boundary conditions require different processes. Absorbing a path corresponds to a Dirichlet condition, whereas reflecting diffusion corresponds to a Neumann condition. Robin conditions involve additional exponential weights built from boundary local time, which measures boundary contact rather than ordinary elapsed time. Parabolic problems combine boundary stopping with a fixed terminal time by stopping at whichever occurs first. On unbounded domains, exit may fail to occur, so a representation can require an additional contribution from trajectories that remain inside indefinitely, or conditions that make that contribution vanish.

Semigroups and imaginary-time quantum mechanics

For time-independent coefficients and potential, define

(Ttf)(x)=Ex ⁣[e−∫0tV(Xs) dsf(Xt)].(T_tf)(x)= \mathbb E_x\!\left[ e^{-\int_0^tV(X_s)\,ds}f(X_t) \right].

The Markov property and additivity of the path integral imply

Tt+s=TtTs.T_{t+s}=T_tT_s.

Thus the weighted operators form a semigroup, on appropriate function spaces and under suitable hypotheses. Its generator is formally L−V\mathcal L-V. The formula does not merely solve one initial-value problem: it represents an entire evolution family. If V≥0V\geq0, this evolution is positivity preserving and contractive on bounded functions in the supremum norm.

For Brownian motion, the corresponding operator is

H=−12Δ+V,Tt=e−tH.H=-\frac12\Delta+V, \qquad T_t=e^{-tH}.

Here Δ\Delta is the Laplace operator. In a Hilbert space setting such as L2(Rd)L^2(\mathbb R^d), suitable assumptions make HH a semibounded self-adjoint operator. The exponential is then defined using the spectral theorem. This framework accommodates certain singular potentials by defining the operator through a closed quadratic form rather than requiring pointwise second derivatives of every function in its domain.

The connection to the Schrödinger equation depends on a distinction between real and imaginary time. Real-time quantum evolution is e−itH/ℏe^{-itH/\hbar}, an oscillatory unitary operator. Imaginary-time evolution is e−τH/ℏe^{-\tau H/\hbar}, a heat-like operator represented by a Feynman–Kac expectation. For the physical Hamiltonian

H=−ℏ22mΔ+U,H=-\frac{\hbar^2}{2m}\Delta+U,

the associated imaginary-time diffusion has covariance rate ℏ/m\hbar/m, while its weight is

exp⁡ ⁣[−1ℏ∫0τU(Xs) ds].\exp\!\left[-\frac1\hbar\int_0^\tau U(X_s)\,ds\right].

A formal substitution of complex time helps explain the historical connection, but it does not by itself establish a positive probabilistic representation of real-time quantum evolution.

Operator splitting provides another route to the formula. Under appropriate conditions, the Trotter product formula gives

e−t(H0+V)=lim⁡n→∞(e−tH0/ne−tV/n)n,H0=−12Δ.e^{-t(H_0+V)} =\lim_{n\to\infty} \left(e^{-tH_0/n}e^{-tV/n}\right)^n, \qquad H_0=-\frac12\Delta.

Each free factor propagates by a Gaussian step, and each potential factor multiplies by an exponential weight. Their product approximates a random path weighted by a Riemann sum for ∫V\int V. Spectral information is encoded in the resulting long-time evolution: when an isolated lowest eigenvalue and suitable overlap conditions exist, the slowest-decaying contribution identifies the ground-state energy. This is the mathematical basis for several imaginary-time methods in quantum computation and statistical physics.

Financial valuation and the Black–Scholes equation

In mathematical finance, the formula links discounted expectations to pricing equations. In the Black–Scholes model, a non-dividend-paying asset under a risk-neutral probability measure satisfies

dSt=rSt dt+σSt dWt,dS_t=rS_t\,dt+\sigma S_t\,dW_t,

where rr is a constant interest rate and σ>0\sigma>0 is volatility. For a European payoff g(ST)g(S_T), the valuation formula is

v(t,s)=e−r(T−t)E[g(ST)∣St=s].v(t,s)=e^{-r(T-t)} \mathbb E[g(S_T)\mid S_t=s].

The associated equation is

∂tv+12σ2s2∂ssv+rs∂sv−rv=0,v(T,s)=g(s).\partial_tv+\frac12\sigma^2s^2\partial_{ss}v +rs\partial_sv-rv=0, \qquad v(T,s)=g(s).

The first-order drift coefficient and discount rate are both rr, but they play different roles: one governs the risk-neutral asset dynamics, while the other discounts the payoff.

The use of risk-neutral probabilities is a separate financial argument, not a consequence of the Feynman–Kac theorem alone. Replication and absence of arbitrage lead to the pricing equation in the idealized complete-market model; an appropriate change of probability measure makes the discounted traded asset a martingale. Feynman–Kac then identifies the equation’s solution with a discounted expectation. Historical asset-return probabilities generally cannot be inserted into the pricing expectation without changing its meaning. In more general models, the existence or uniqueness of an admissible pricing measure requires additional assumptions.

For the call payoff g(s)=(s−K)+g(s)=(s-K)^+, the terminal value is continuous but not differentiable at the strike KK. This does not prevent the representation. The expectation remains well defined, and diffusion smooths the solution for t<Tt<T. Since log⁡ST\log S_T is normally distributed conditional on StS_t, evaluating the expectation gives

v(t,s)=sΦ(d1)−Ke−rτΦ(d2),v(t,s)=s\Phi(d_1)-Ke^{-r\tau}\Phi(d_2),

where

d1=log⁡(s/K)+(r+σ2/2)τστ,d2=d1−στ,τ=T−t.d_1=\frac{\log(s/K)+(r+\sigma^2/2)\tau}{\sigma\sqrt{\tau}}, \qquad d_2=d_1-\sigma\sqrt{\tau}, \qquad \tau=T-t.

The function Φ\Phi is the cumulative distribution function of the standard normal distribution.

The same structure accommodates stochastic interest rates:

v(t,x)=Et,x Q ⁣[e−∫tTr(Xs) dsg(XT)].v(t,x)= \mathbb E_{t,x}^{\,Q}\!\left[ e^{-\int_t^Tr(X_s)\,ds}g(X_T) \right].

Running payments add a source integral, while barrier contracts introduce stopping at a boundary. Path-dependent contracts may require additional state variables: an Asian payoff, for example, can be made Markovian by adjoining the running integral of the asset price. American-style exercise is different because the holder chooses a stopping time. Its value involves an optimization over stopping rules and leads to a variational inequality, rather than merely the linear terminal-value equation of the basic formula.

Monte Carlo computation and numerical limitations

The representation naturally produces a Monte Carlo method. If YY denotes the discounted terminal payoff plus any accumulated source contribution, independent simulated trajectories produce the estimator

u^N=1N∑k=1NYk.\widehat u_N=\frac1N\sum_{k=1}^NY_k.

For exact independent samples and finite expectation, the law of large numbers gives convergence to uu. If Var⁡(Y)<∞\operatorname{Var}(Y)<\infty, the central limit theorem gives an asymptotic standard error of

Var⁡(Y)N.\sqrt{\frac{\operatorname{Var}(Y)}{N}}.

The nominal N−1/2N^{-1/2} sampling rate does not directly depend on the state-space dimension, which makes probabilistic representations attractive when deterministic grids would become prohibitively large.

Most diffusions cannot be sampled exactly. A basic time-discrete approximation is Euler–Maruyama:

Xn+1=Xn+b(tn,Xn)Δt+σ(tn,Xn)Δt Zn,X_{n+1}=X_n+b(t_n,X_n)\Delta t +\sigma(t_n,X_n)\sqrt{\Delta t}\,Z_n,

where the ZnZ_n are independent standard Gaussian vectors. The path integral in the exponential must also be approximated. This introduces discretization bias in addition to sampling error. Under suitable smoothness and growth conditions, Euler–Maruyama commonly has strong order 1/21/2 and weak order 11; discontinuous payoffs, boundary crossings, or irregular coefficients can change those rates. An uncertainty interval based only on sample variability does not measure this bias.

Boundary problems present a distinctive numerical difficulty. A simulated path may leave and reenter a domain between two sampled times, so checking only the grid points can miss the first exit. For some processes and simple boundaries, Brownian-bridge crossing probabilities correct this effect. Otherwise, refined time steps or specialized exit-time methods are required. Likewise, replacing a continuously monitored barrier by discrete monitoring changes the mathematical payoff, not merely its computational implementation. In high dimensions, the cost of each trajectory, the number of necessary time steps, and the variance of the weight can all grow substantially despite the dimension-independent nominal sampling exponent.

Variance reduction can be decisive. Control variates use correlated quantities with known expectations; importance sampling changes the simulation law and compensates with a likelihood ratio; multilevel Monte Carlo combines coupled paths at several time resolutions. Large positive potentials can leave almost all weights near zero, while negative potentials can create rare but enormous weights. Both situations can make a straightforward estimator inefficient or invalidate finite-variance assumptions. Moreover, direct Monte Carlo usually estimates the solution at selected starting points rather than constructing a global approximation. Methods based on finite differences or finite elements may therefore be preferable when a full solution field is required in a low-dimensional domain.

Extensions beyond the classical diffusion formula

The basic mechanism is not restricted to continuous paths. For a continuous-time Markov chain with generator matrix QQ, let VV be a diagonal matrix of state-dependent rates. Then, under the convention that QQ acts on column vectors of function values,

u(t)=et(Q−V)gu(t)=e^{t(Q-V)}g

has the representation

ui(t)=Ei ⁣[e−∫0tV(Xs) dsg(Xt)].u_i(t)= \mathbb E_i\!\left[ e^{-\int_0^tV(X_s)\,ds}g(X_t) \right].

The matrix exponential therefore has the same weighted-path interpretation as a diffusion semigroup. More general jump processes replace the local second-order differential operator by a nonlocal generator, often producing integro-differential equations. The essential ingredients are a Markov evolution and a multiplicative functional accumulated along its paths.

Rough potentials and coefficients require careful choices of analytical framework. An unbounded nonnegative potential can often be treated by truncation and monotone convergence. A potential with a large negative part is more delicate because the exponential weight grows; conditions controlling short-time path integrals are used to establish finiteness and operator bounds. Depending on the problem, the represented function may be a weak solution, a mild semigroup solution, or a viscosity solution. These notions are not interchangeable, and each carries its own hypotheses for existence, uniqueness, and boundary behavior. There is no single unrestricted version covering all measurable coefficients and all potentials.

Semilinear equations have related, but different, stochastic representations. For

∂tu+Lu+F ⁣(t,x,u,σT∇u)=0,u(T,x)=g(x),\partial_tu+\mathcal Lu+ F\!\left(t,x,u,\sigma^{\mathsf T}\nabla u\right)=0, \qquad u(T,x)=g(x),

a backward stochastic differential equation takes the form

Ys=g(XT)+∫sTF(r,Xr,Yr,Zr) dr−∫sTZr dWr.Y_s=g(X_T)+ \int_s^T F(r,X_r,Y_r,Z_r)\,dr -\int_s^T Z_r\,dW_r.

In a sufficiently regular Markovian setting,

Ys=u(s,Xs),Zs=σ(s,Xs)T∇u(s,Xs).Y_s=u(s,X_s), \qquad Z_s=\sigma(s,X_s)^{\mathsf T}\nabla u(s,X_s).

The linear Feynman–Kac formula is recovered when FF is affine in YY and independent of ZZ after the forward dynamics have been fixed. For nonlinear FF, the representation is no longer simply an expectation of a prescribed path payoff: the unknown backward process appears inside its own driver.

Controlled diffusions introduce optimization over trajectories and lead to Hamilton–Jacobi–Bellman equations, connecting the subject to the Bellman equation and mathematical optimization. Other extensions use branching particles to represent particular nonlinear reaction terms, or attach matrix-valued transport to paths when the unknown is vector-valued. In the latter case, coefficients at different times may not commute, so an ordinary scalar exponential must be replaced by a time-ordered evolution. These extensions retain the principle of translating analytic evolution into stochastic path information, but their integrability, ordering, and uniqueness requirements differ from those of the scalar linear theorem.