---
title: Doob Conditioning in Markov Processes
url: https://www.emergentmind.com/topics/doob-conditioning
type: topic
---

# Doob Conditioning in Markov Processes

Doob conditioning refers to a broad set of constructions for Markov processes, stochastic differential equations, and statistical decision theory in which a process is transformed or "conditioned" to satisfy specified constraints—typically by modifying its transition dynamics via a space-time harmonic (or excessive) function. Central to these methodologies is the Doob $h$-transform, which yields a new process whose law matches that of the original process under a conditioning event, such as achieving a rare terminal state, remaining in a cone, non-extinction, or attaining a particular occupation measure. The framework is closely related to the general theory of conditional expectations (Doob–Dynkin lemma), large deviations, optimal control, the analysis of rare events, and the structure of Markov process boundaries.

## 1. The Doob $h$-Transform and Conditioning Principle

Let $X_t$ be a Markov process on state space $E$ with generator $L$. Suppose $h:E\to(0,\infty)$ is harmonic for $L$ ($L h=0$ on $D\subset E$) or, more generally, excessive. The Doob $h$-transform constructs a new Markov process with semigroup
\[
P^h_t f(x)=\frac{1}{h(x)}\,\mathbb{E}_x\left[h(X_t)f(X_t)\mathbf{1}_{t<\tau_{D^c}}\right],
\]
and generator
\[
L^h f(x) = \frac{1}{h(x)}\, L(h f)(x),
\]
where $\tau_{D^c}$ is the exit time from $D$. For continuous-time diffusions,
\[
dX_t = b(X_t)\,dt + \sqrt{\varepsilon}\,\sigma\,dW_t, \quad X_0=x_0,
\]
if one conditions on $X_T=a$, the appropriate space-time harmonic function $h(x,t) = \mathbb{P}(X_T = a \mid X_t = x)$ solves the backward Kolmogorov equation
\[
\left(\partial_t + \mathcal{L}^\varepsilon\right) h = 0, \quad h(x,T) = \delta(x-a),
\]
with $\mathcal{L}^\varepsilon f = b\cdot\nabla f + \frac{\varepsilon}{2} \nabla\cdot (D\nabla f),\ D=\sigma\sigma^T$. The $h$-transformed process has drift
\[
b^*(x,t) = b(x) + \varepsilon D \nabla_x \ln h(x,t).
\]
This construction ensures the new process is distributed as the original process conditioned (e.g., on $X_T = a$) [2605.24537], [2103.12179].

## 2. Large Deviations, Optimal Control, and Pathwise Representations

In the small-noise regime ($\varepsilon \ll 1$), Doob conditioning facilitates the efficient simulation and analysis of rare events by reweighting trajectories so that the prescribed rare event becomes typical. For an additive observable $Q_T$, the moment generating function under Doob conditioning exhibits a large deviation scaling:
\[
G(\lambda) \asymp \exp\left(-\frac{I_\lambda}{\varepsilon}\right),
\]
where the rate function $I_\lambda$ arises as the value of a variational (optimal control) problem:
\[
I_\lambda = \inf_{x(\cdot): x(0)=x_0, x(T)=a} \left\{ \int_0^T L(x, \dot{x}) \, dt + \Phi(x(T)) \right\}.
\]
Here $L(x, v)$ is a Lagrangian encoding the deviation from typical dynamics and the observable tilt, and $\Phi$ enforces endpoint constraints [2605.24537].

The Doob transform can also be interpreted as a post-selection procedure: the conditioned law is realized by reweighting the original path measure with the time-dependent harmonic function $h$, reflecting the likelihood of achieving the constraint.

## 3. Examples: Bridges, Lévy Processes, and Conditioned Trees

- **Brownian Bridge:** Conditioning Brownian motion to hit $a$ at time $T$ yields the classical Brownian bridge. The Doob $h$-transform drift is $(a-x)/(T-t)$ [2605.24537].
- **Ornstein-Uhlenbeck Bridge:** Conditioning on an endpoint yields a nontrivial, time-dependent drift deduced from the explicit $h(x,t)$ (a Gaussian kernel) [2605.24537].
- **Stable Lévy Processes:** For the one-dimensional symmetric $\alpha$-stable process, the $h$-transform with $h(x)=x$ on $(0,\infty)$ yields a process equivalent in law to the radial part of a three-dimensional isotropic $\alpha$-stable process—the Doob–McKean identity generalizes to jump processes [2103.12179].
- **Rémy's Tree Growth Chain:** The Doob–Martin compactification for Rémy’s Markov chain on binary trees explicitly identifies all ways the process can “go to infinity.” The Doob $h$-transform provides the canonical way to condition on convergence to a boundary point, realized as an extremal (minimal) harmonic function in terms of sampled subtrees and limiting $\mathbb{R}$-tree structures [1411.2526].

## 4. Doob Conditioning in Self-Interacting and Constrained Processes

Markov processes with path-constrained observables (e.g., occupation time, empirical measures) can be optimally sampled by constructing a self-interacting dynamics, where the transition probabilities at each step depend on the constrained empirical history. The Doob conditioning framework systematically constructs such processes:
\[
L_t^h(x,y) = \frac{h(t, y)}{h(t, x)} L(x, y),
\]
with $h(t,x)$ solving the appropriate backward equation. This approach realizes optimal simulation of rare events (e.g., bridges, excursions, forced excursions) and aligns with dynamic programming in optimal control [2503.01574].

## 5. Conditioning Branching and Population Processes

Conditioning continuous-state branching processes, such as the logistic CB process, on non-extinction or infinite progeny is realized by an explicit Doob $h$-transform. The excessive function $h$ is constructed via an integral involving the process' scale functions, and the resulting process incorporates an additional, density-dependent immigration term. Under the transformed law, the process exhibits either explosion (finite lifetime) or enters from zero, with generator
\[
\mathscr{L}^\uparrow f(z) = \frac{1}{h(z)}\,\mathscr{L}(h f)(z),
\]
where $\mathscr{L}$ is the original generator. Convergence criteria, duality relations, and explicit martingale characterizations are established [2408.14993].

## 6. Applications in Conditional Generative Modeling and Machine Learning

Doob’s $h$-transform is directly exploited for conditioning generative models, especially diffusion models. In adaptation problems, the $h$-transform enables measure transportation to a target distribution by dynamically modifying the drift with $\nabla \log h$, where $h(x,t)$ matches the desired event or high-reward property. This approach supports efficient, training-free, simulation-based adaptation, including for non-differentiable objectives, with finite-sample guarantees and established convergence rates. Practical approximations circumvent explicit calculation of $h$, using Monte Carlo evaluation or one-step surrogates. Empirical results validate these theoretical properties in offline RL and text-to-image generation [2602.16198].

## 7. The Doob–Dynkin Lemma and Conditional Expectation

In a foundational sense, Doob conditioning generalizes to the factorization properties given by the Doob–Dynkin lemma: any random variable measurable with respect to another function can be written as a measurable function thereof, underpinning $E[X|Y=y]$ as a concrete function of $y$. This principle is crucial in statistical learning, Bayesian estimation, and conditional Monte Carlo, providing the technical justification for representing conditional expectations as explicit functions and supporting optimal prediction and inference [1801.00974].

---

**References**  
- "Rare events of small-noise Doob conditioned processes" [2605.24537]  
- "The Doob-McKean identity for stable Lévy processes" [2103.12179]  
- "Doob--Martin boundary of Rémy's tree growth chain" [1411.2526]  
- "Self-interacting processes via Doob conditioning" [2503.01574]  
- "Conditioning the logistic continuous-state branching process on non-extinction via its total progeny" [2408.14993]  
- "Training-Free Adaptation of Diffusion Models via Doob's $h$-Transform" [2602.16198]  
- "Optimal Learning from the Doob-Dynkin lemma" [1801.00974]

Source: https://www.emergentmind.com/topics/doob-conditioning