Skip to content

Transition Density as Green's Function

The transition density of a diffusion and the Green's function of its generator are the same mathematical object viewed through two lenses. This page is the probability lens: it establishes the equivalence, derives the forward/backward equations, and develops the financial interpretation (state-price density, Arrow-Debreu price, Breeden-Litzenberger). The PDE lens -- operator definition, superposition principle, and smoothing -- is in Green's Function for Parabolic PDEs.


The Two Perspectives

Given the SDE \(dX_t = \mu(X_t)\,dt + \sigma(X_t)\,dW_t\) with generator \(\mathcal{L}\), define:

  • Probabilistic. The transition density \(p(t, x \mid s, y)\) by \(\mathbb{P}(X_t \in A \mid X_s = y) = \int_A p(t, x \mid s, y)\,dx\).
  • Analytical. The Green's function \(G(t, x; s, y)\) as the fundamental solution of \(\partial_t - \mathcal{L}\) (see Green's Function for Parabolic PDEs).

The central claim:

\[ \boxed{\;p(t, x \mid s, y) = G(t, x; s, y)\;} \]

The transition density of a diffusion is the Green's function of its generator.


Proof of the Equivalence

Let \(g\) be a bounded continuous test function and set

\[ u(s, y) = \mathbb{E}[g(X_t) \mid X_s = y] = \int g(x)\,p(t, x \mid s, y)\,dx \]

The Kolmogorov backward equation (derived from Itô's formula applied to \(u(s, X_s)\)) gives

\[ -\partial_s u = \mathcal{L}_y u, \qquad u(t, y) = g(y) \]

The Green's function representation of the same backward problem gives

\[ u(s, y) = \int g(x)\,G(t, x; s, y)\,dx \]

Both representations hold for every bounded continuous \(g\). By the density of such \(g\) in the space of test functions, the kernels must agree:

\[ p(t, x \mid s, y) = G(t, x; s, y) \qquad \square \]

The delta-function initial condition of \(G\) reflects the point-mass initial law \(X_s = y\) a.s.: both encode "the system starts with certainty at \(y\)".


Forward and Backward Equations

Recall (see § Kolmogorov Forward Equation and § Kolmogorov Backward Equation): \(p(t,x\mid s,y)\) satisfies \(\partial_t p = \mathcal{L}_x^* p\) in the destination \((t,x)\) and \(-\partial_s p = \mathcal{L}_y p\) in the origin \((s,y)\), each with the delta initial/terminal condition. The forward operator \(\mathcal{L}^* p = -\partial_x[\mu p] + \tfrac12\partial_{xx}[\sigma^2 p]\) is the formal adjoint of the generator; this adjoint structure is what enforces conservation of probability.

Equation Varies Fixed Operator Role
Forward (Fokker-Planck) \((t, x)\) destination \((s, y)\) origin \(\mathcal{L}_x^*\) Density evolution
Backward (Kolmogorov) \((s, y)\) origin \((t, x)\) destination \(\mathcal{L}_y\) Expectation / pricing

Here the Green's-function lens adds one observation: the same kernel \(G = p\) solves both PDEs simultaneously -- forward in \((t,x)\), backward in \((s,y)\) -- because the delta-source condition is symmetric in the two variable pairs. The two Kolmogorov equations are thus two views of one object, not two unrelated facts.


Chapman-Kolmogorov Equation

For \(s < r < t\):

\[ \boxed{\;p(t, x \mid s, y) = \int p(t, x \mid r, z)\,p(r, z \mid s, y)\,dz\;} \]

Probabilistic derivation. By the Markov property and total probability,

\[ \mathbb{P}(X_t \in dx \mid X_s = y) = \int \mathbb{P}(X_t \in dx \mid X_r = z)\,\mathbb{P}(X_r \in dz \mid X_s = y) \]

Dividing by \(dx\) gives the displayed equation.

PDE interpretation. Solve the point-source problem in two stages: first evolve from \(s\) to \(r\), obtaining \(p(r, z \mid s, y)\); then use this as initial data for the segment \(r \to t\). By uniqueness, the two-stage result equals the direct transition \(p(t, x \mid s, y)\). This is the semigroup property of \(G\) from Green's Function for Parabolic PDEs.

The pricing analogue is the law of iterated expectations: \(\mathbb{E}[\,\cdot \mid \mathcal{F}_s] = \mathbb{E}[\mathbb{E}[\,\cdot \mid \mathcal{F}_r] \mid \mathcal{F}_s]\).


Brownian Motion with Drift

Recall (see § Fundamental Solution and § Transition Densities for Standard SDEs): for \(dX_t = \mu\,dt + \sigma\,dW_t\) with \(X_t - X_s \sim N(\mu(t-s), \sigma^2(t-s))\),

\[ p(t, x \mid s, y) = \frac{1}{\sigma\sqrt{2\pi(t-s)}}\exp\!\left(-\frac{(x - y - \mu(t-s))^2}{2\sigma^2(t-s)}\right) \]

is the heat kernel translated by the drift -- and identifying it as \(G\) makes the Black-Scholes operator's fundamental solution an Arrow-Debreu kernel.


Ornstein-Uhlenbeck

Recall (see § Transition Densities for Standard SDEs): for \(dX_t = -\kappa X_t\,dt + \sigma\,dW_t\), \(p(t,x\mid s,y)\) is Gaussian with mean \(ye^{-\kappa\tau}\) and variance \(v(\tau) = \sigma^2(1-e^{-2\kappa\tau})/(2\kappa)\), \(\tau = t-s\).

As \(\tau \to \infty\), \(p \to N(0, \sigma^2/(2\kappa))\) independent of \(y\): the process forgets its initial condition and relaxes to the stationary density. The same phenomenon appears in Spectral Decomposition as the decay of all non-zero eigenmodes -- and in Free vs Bounded Domains as the contrast between free-space dissipation and bounded-domain equilibration.


Financial Interpretation

State-Price Density

Under the risk-neutral measure \(\mathbb{Q}\), the transition density of the asset price is the state-price density (Arrow-Debreu kernel):

\[ V(t, S) = e^{-r(T-t)}\int g(S_T)\,p^{\mathbb{Q}}(T, S_T \mid t, S)\,dS_T \]

For each terminal state \(S_T\), the quantity \(e^{-r(T-t)}p^{\mathbb{Q}}(T, S_T \mid t, S)\,dS_T\) is the price today of an Arrow-Debreu security paying $1 in that state and nothing otherwise. Knowing \(p^{\mathbb{Q}}\) is equivalent to knowing all European option prices simultaneously. The integral-against-\(G\) formulation is developed in Green's Function for Parabolic PDEs.

Breeden-Litzenberger

The second derivative of call prices with respect to strike recovers the risk-neutral density:

\[ \boxed{\;p^{\mathbb{Q}}(T, K \mid t, S) = e^{r(T-t)}\,\frac{\partial^2 C}{\partial K^2}(t, S; T, K)\;} \]

This is the basis of implied-density estimation and non-parametric calibration from option quotes.

Black-Scholes

Recall (see § Geometric Brownian Motion and § Black-Scholes via Heat Equation): the risk-neutral GBM transition density is the lognormal \(p^{\mathbb{Q}}(T,S_T\mid t,S)\), and integrating any payoff against \(e^{-r(T-t)}p^{\mathbb{Q}}\) reproduces the Black-Scholes price. The Green's-function lens identifies this kernel as the Arrow-Debreu state-price density for the Black-Scholes operator.


Dictionary

Object Probability PDE Finance
\(p(t, x \mid s, y) = G\) Transition density Green's function State-price density
Action \(\int G\cdot f\) Expected value Solution of IVP Option price
Semigroup / CK Markov property Uniqueness / superposition Iterated expectations
\(\delta(x - y)\) initial \(X_s = y\) a.s. Point source Unit Arrow-Debreu

Summary

\[ \boxed{\;p(t, x \mid s, y) = G(t, x; s, y) \quad \Longleftrightarrow \quad \text{transition density} = \text{Green's function}\;} \]

The identification collapses two parallel theories into one. Probabilistic questions become PDE problems; PDE problems become stochastic computations. The forward equation tracks density evolution; the backward equation prices claims. The Chapman-Kolmogorov / semigroup property expresses the Markov property in analytic form. And in finance, the Green's function is the Arrow-Debreu pricing kernel.


See Also


Exercises

Exercise 1. The direct verification of the forward equation \(\partial_t p = -\mu\partial_x p + \tfrac12\sigma^2\partial_{xx}p\) for the Gaussian density of \(dX_t = \mu\,dt + \sigma\,dW_t\) is carried out in § Kolmogorov Forward Equation. Explain why this same Gaussian must also solve the backward equation \(-\partial_s p = \mu\partial_y p + \tfrac12\sigma^2\partial_{yy}p\), without redoing the differentiation.

Solution to Exercise 1

The Gaussian depends on \((t,x,s,y)\) only through \(\tau = t-s\) and \(w = x - y - \mu\tau\). Under the symmetry \((t,x) \leftrightarrow (s,y)\), \(\tau \to -\tau\) and \(w \to -w\), but the density depends only on \(\tau\) and \(w^2\), so the forward operator in \((t,x)\) and the backward operator in \((s,y)\) produce identical actions on \(p\). This is the Green's-function identity at work: one kernel, two PDEs.


Exercise 2. Verify the Chapman-Kolmogorov equation for standard Brownian motion by explicit Gaussian convolution.

Solution to Exercise 2

With \(\tau_1 = r - s\), \(\tau_2 = t - r\), combine exponents:

\[ -\frac{(x - z)^2}{2\tau_2} - \frac{(z - y)^2}{2\tau_1} = -\frac{(\tau_1 + \tau_2)(z - z^*)^2 + \tau_1\tau_2(x - y)^2/(\tau_1+\tau_2)}{2\tau_1\tau_2} \]

with \(z^* = (\tau_1 x + \tau_2 y)/(\tau_1 + \tau_2)\). The \(z\)-integral gives \(\sqrt{2\pi\tau_1\tau_2/(\tau_1+\tau_2)}\). Combining prefactors:

\[ \frac{1}{\sqrt{2\pi(\tau_1 + \tau_2)}}\exp\!\left(-\frac{(x - y)^2}{2(\tau_1 + \tau_2)}\right) = p(t, x \mid s, y) \quad\checkmark \]

The result says: adding independent Gaussian increments adds variances -- the defining property of Brownian motion.


Exercise 3. State the forward and backward equations for the Ornstein-Uhlenbeck process \(dX_t = -\kappa X_t\,dt + \sigma\,dW_t\) in the two pairs of variables, and identify which PDE governs the option price \(V(s, y) = \mathbb{E}[g(X_t) \mid X_s = y]\).

Solution to Exercise 3

With \(\mathcal{L} = -\kappa x\,\partial_x + \tfrac{1}{2}\sigma^2\partial_{xx}\), the adjoint is \(\mathcal{L}^* p = \kappa\,\partial_x(x p) + \tfrac{1}{2}\sigma^2\partial_{xx}p\).

Forward (in \((t, x)\)): \(\partial_t p = \kappa\,\partial_x(x p) + \tfrac{1}{2}\sigma^2\partial_{xx}p\).

Backward (in \((s, y)\)): \(-\partial_s p = -\kappa y\,\partial_y p + \tfrac{1}{2}\sigma^2\partial_{yy}p\).

The option price \(V(s, y) = \mathbb{E}[g(X_t) \mid X_s = y]\) depends on \((s, y)\) and satisfies the backward equation \(-\partial_s V = \mathcal{L}_y V\) with \(V(t, y) = g(y)\). The backward equation is the pricing equation; the forward equation governs implied-distribution and risk analysis.


Exercise 4. The full derivation of \(V = S\mathcal{N}(d_1) - Ke^{-r(T-t)}\mathcal{N}(d_2)\) from the lognormal transition density of geometric Brownian motion is carried out in § Black-Scholes PDE: Analytic Solutions. In the Green's-function language of this page, state which object \(G\) is being integrated against and identify the financial role of the prefactor \(e^{-r(T-t)}\).

Solution to Exercise 4

\(G\) is the risk-neutral lognormal transition density \(p^{\mathbb{Q}}(T,S_T \mid t,S)\) for GBM under \(\mathbb{Q}\); the pricing integral \(V(t,S) = e^{-r(T-t)}\int (S_T - K)^+ G\,dS_T\) integrates the payoff against \(G\) over the in-the-money region.

The prefactor \(e^{-r(T-t)}\) converts the kernel into the Arrow-Debreu state-price density: \(e^{-r(T-t)}G(T,S_T;t,S)\,dS_T\) is the price today of $1 delivered if \(S_T \in [S_T, S_T+dS_T]\). The Black-Scholes formula is then the integral of the call payoff against this state-price density.


Exercise 5. Explain Breeden-Litzenberger: derive \(p^{\mathbb{Q}}(T, K \mid t, S) = e^{r(T-t)}\partial_K^2 C(t, S; T, K)\) by differentiating the pricing integral twice.

Solution to Exercise 5

\(C(t, S; T, K) = e^{-r(T-t)}\int_K^\infty (S_T - K)\,p^{\mathbb{Q}}(T, S_T \mid t, S)\,dS_T\).

Differentiate once in \(K\): the boundary term \((S_T - K)p^{\mathbb{Q}}\) at \(S_T = K\) vanishes, leaving

\[ \partial_K C = -e^{-r(T-t)}\int_K^\infty p^{\mathbb{Q}}(T, S_T \mid t, S)\,dS_T \]

Differentiate again, applying Leibniz to the moving lower limit:

\[ \partial_K^2 C = e^{-r(T-t)}\,p^{\mathbb{Q}}(T, K \mid t, S) \]

Rearranging gives the Breeden-Litzenberger formula. Model-free content: given a dense set of call quotes in \(K\), the risk-neutral marginal density at maturity \(T\) is recovered by numerical second differencing.


Exercise 6. For a diffusion killed at an absorbing barrier \(B\), the killed density \(p_B(t, x \mid s, y) \le p(t, x \mid s, y)\) integrates to less than \(1\). Relate the deficit to the first-passage probability \(\mathbb{P}(\tau_B \le t \mid X_s = y)\) and derive the flux formula for the rate at which mass leaks out.

Solution to Exercise 6

The killed density is

\[ p_B(t, x \mid s, y)\,dx = \mathbb{P}(X_t \in dx,\ \tau_B > t \mid X_s = y) \]

Integrating over \(x\):

\[ \int p_B(t, x \mid s, y)\,dx = \mathbb{P}(\tau_B > t \mid X_s = y) \]

so the deficit is \(1 - \int p_B\,dx = \mathbb{P}(\tau_B \le t \mid X_s = y)\).

The leakage rate follows from the Fokker-Planck equation with absorbing condition \(p_B = 0\) at \(x = B\):

\[ -\frac{d}{dt}\int p_B\,dx = \pm\tfrac{1}{2}\sigma^2(B)\,\partial_x p_B(t, B \mid s, y) \]

(sign determined by whether \(B\) is upper or lower barrier). In barrier-option pricing, this deficit is the knock-out probability; see Free vs Bounded Domains for the full boundary-value-problem treatment.