Generalized Doob's h-Transform

• Generalized Doob's h-transform is a measure-change construction that reweights a reference Markov process using positive harmonic or space-time harmonic functions, endpoint weights, and Feynman–Kac factors.
• It adjusts dynamics through modifications in drift, jump intensity, or endpoint laws while preserving the Markov property, offering a unified view of conditioning, duality, and bridge constructions.
• The framework bridges finite- and infinite-dimensional systems and connects variational principles and control formulations to practical applications in diffusions, jump processes, and generative models.

Generalized Doob’s $h$-transform is a family of measure-change constructions that turns a reference Markov process into another Markov process by reweighting trajectories with a positive harmonic or space-time harmonic function $h$, or, in broader formulations, with endpoint weights and Feynman–Kac factors. In its classical semigroup form, the transform replaces $P_t$ by $P_t^h f(x)=P_t(hf)(x)/h(x)$; in generalized form it acts on path space through densities such as $f_0(X_0)\exp(-\int V)\,g_1(X_1)$, conditional success probabilities, ground states of Schrödinger operators, or terminal bridge densities. Across diffusions, jump processes, self-similar Markov processes, measure-valued systems, and modern generative models, the common structure is that a reference law is tilted so that the transformed dynamics preserve Markovianity while acquiring a corrected drift, jump kernel, or endpoint law [(Léonard, 2011); (Pieper-Sethmacher et al., 6 Feb 2026); (Denker et al., 2024)].

1. General form of the transform

A broad path-space formulation starts from a continuous-time Markov reference process $R$ on $[0,1]$ and defines a transformed law

$$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$

with $f_0,g_1\ge 0$ measurable and $V$ measurable and bounded below. This is called a generalized $h$0-transform, or generalized $h$1-process. The transformation preserves the Markov property and induces forward and backward space-time functions

$h$2

and

$h$3

When $h$4 is stationary with invariant law $h$5, the one-time marginals satisfy

$h$6

This formulation extends the classical case in which $h$7 is purely harmonic and the transform is interpreted as conditioning or steering by $h$8 (Léonard, 2011).

A second standard formulation appears in exit conditioning. For a diffusion absorbed at $h$9, a measurable exit set $P_t$0, and

$P_t$1

the conditioned semigroup is

$P_t$2

The corresponding conditional law satisfies the identity

$P_t$3

for $P_t$4-measurable $P_t$5. In this sense, generalized Doob transforms are exact conditioning mechanisms, not merely heuristic biases (Bakhtin et al., 2013).

These two formulations already display the main extension beyond the classical textbook picture. The transform need not be limited to a positive harmonic function on state space; it can involve endpoint weights, rough potentials, boundary events, and space-time harmonic functions defined by conditional expectations or Feynman–Kac formulas (Léonard, 2011).

2. Generator calculus and transformed dynamics

At generator level, the generalized transform modifies the infinitesimal dynamics in a structured way. For the generalized $P_t$6-process above, a central formula is

$P_t$7

where $P_t$8 is the carré du champ operator

$P_t$9

The identification relies on the paper’s main conceptual result that the extended generator equals the stochastic derivative,

$P_t^h f(x)=P_t(hf)(x)/h(x)$0

Under finite entropy $P_t^h f(x)=P_t(hf)(x)/h(x)$1, this yields a probabilistic generator theory without continuity or differentiability assumptions on $P_t^h f(x)=P_t(hf)(x)/h(x)$2, $P_t^h f(x)=P_t(hf)(x)/h(x)$3, or $P_t^h f(x)=P_t(hf)(x)/h(x)$4 (Léonard, 2011).

For diffusions

$P_t^h f(x)=P_t(hf)(x)/h(x)$5

the transformed drift becomes

$P_t^h f(x)=P_t(hf)(x)/h(x)$6

with

$P_t^h f(x)=P_t(hf)(x)/h(x)$7

For jump processes on a discrete space, the same mechanism acts multiplicatively on intensities: $P_t^h f(x)=P_t(hf)(x)/h(x)$8 Thus the diffusion case appears as a drift correction, while the jump case appears as a ratio-of-$P_t^h f(x)=P_t(hf)(x)/h(x)$9 reweighting of transitions (Léonard, 2011).

A ground-state version arises for non-local Schrödinger operators $f_0(X_0)\exp(-\int V)\,g_1(X_1)$0. If $f_0(X_0)\exp(-\int V)\,g_1(X_1)$1 satisfies

$f_0(X_0)\exp(-\int V)\,g_1(X_1)$2

the intrinsic semigroup is

$f_0(X_0)\exp(-\int V)\,g_1(X_1)$3

and the transformed Lévy-type generator contains both a drift term involving $f_0(X_0)\exp(-\int V)\,g_1(X_1)$4 and a jump modulation factor $f_0(X_0)\exp(-\int V)\,g_1(X_1)$5. The resulting ground state-transformed process is càdlàg, stationary under $f_0(X_0)\exp(-\int V)\,g_1(X_1)$6, and solves a jump SDE in weak form (Lorinczi et al., 2017).

A related operator-theoretic extension is the Darboux transformation of diffusion processes. There the probabilistic construction is explicitly

$f_0(X_0)\exp(-\int V)\,g_1(X_1)$7

and the transformed killing rate is

$f_0(X_0)\exp(-\int V)\,g_1(X_1)$8

This realizes Darboux factorization directly at the level of Markov semigroups (Kuznetsov et al., 2024).

Setting	Transform object	Effect on dynamics
Continuous diffusion	$f_0(X_0)\exp(-\int V)\,g_1(X_1)$9	Added drift correction (Léonard, 2011)
Jump process	$R$0	Reweighted jump intensities (Léonard, 2011)
Ground state transform	$R$1	Drift and jump-kernel modulation (Lorinczi et al., 2017)

3. Conditioning, duality, inversion, and pathwise realizations

In self-similar Markov process theory, inversion, duality, and Doob $R$2-transforms are presented as one structure. Every $R$3-valued self-similar Markov process can be represented via an extended Lamperti transformation from a Markov additive process $R$4, and the inverted process

$R$5

is in weak duality with $R$6 with respect to

$R$7

if and only if the underlying MAP is reversible with respect to $R$8. For isotropic stable Lévy processes, this yields the explicit $R$9-function

$[0,1]$0

recovering the Bogdan–Żak theorem in the paper’s formulation (Alili et al., 2016).

For one-dimensional regular diffusions on $[0,1]$1, the transformed law is

$[0,1]$2

with generator

$[0,1]$3

The dual diffusion $[0,1]$4 can be realized pathwise by a deterministic involution $[0,1]$5 followed by a random clock,

$[0,1]$6

In scale coordinates the involution is Möbius, so the theory generalizes Euclidean inversion and unifies Brownian motion, drifted Brownian motion, Bessel processes, and hyperbolic Bessel processes (Alili et al., 2012).

A discrete example is the two-dimensional simple random walk conditioned on never hitting the origin. Here the $[0,1]$7-function is the potential kernel

$[0,1]$8

harmonic on $[0,1]$9. The transformed kernel is

$$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$0

with reversible measure $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$1 and explicit Green’s function

$$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$2

This is a canonical example in which the transform defines a transient process corresponding to a recurrent process conditioned on an event of probability zero in the naive sense (Popov, 2019).

Measure-valued analogues admit pathwise genealogical realizations. In the lookdown particle system, a product-type transform for a Generalized Fleming–Viot process uses

$$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$3

and is realized by changing the initial configuration and forgetting reproduction events that involve more than one of the first $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$4 levels. The additive transform uses

$$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$5

and yields a spine-type picture in which the total mass is size-biased and the first-level particle follows the $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$6-transform of the underlying motion. The paper identifies this with an immortal particle representation for Dawson–Watanabe processes and a pathwise version of Overbeck’s suggestion for Fleming–Viot systems (Hénard, 2012).

A geometric hypoelliptic version appears on the Heisenberg group. There, the Cayley transform maps Brownian motion on $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$7 to a time-changed Brownian motion on the CR sphere conditioned to be at the south pole at a random time, while the Kelvin transform maps Brownian motion started away from $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$8 to a time-changed Brownian bridge conditioned to be at the origin. The conformal identities are expressed through explicit $$P := f_0(X_0)\exp\!\Big(-\int_0^1 V_t(X_t)\,dt\Big)\, g_1(X_1)\,R,$$9-functions and carré-du-champ formulas (Wang, 2016).

4. Infinite-dimensional and bridge-type extensions

A rigorous infinite-dimensional generalization is developed for Hilbert-space-valued diffusions and SPDEs. The reference process is

$f_0,g_1\ge 0$0

and, assuming a Gaussian reference measure $f_0,g_1\ge 0$1 dominating the target $f_0,g_1\ge 0$2 in the required sense, the space-time harmonic function is defined by

$f_0,g_1\ge 0$3

The path law is tilted by

$f_0,g_1\ge 0$4

and the transformed process satisfies

$f_0,g_1\ge 0$5

with corrected initial law

$f_0,g_1\ge 0$6

Under the transformed measure, the terminal marginal is exact: $f_0,g_1\ge 0$7 The construction requires transition densities with respect to $f_0,g_1\ge 0$8, Fréchet differentiability of $f_0,g_1\ge 0$9 with at most polynomial growth, a martingale property for $V$0, and then Girsanov’s theorem to identify the drift correction. The paper further states a Wasserstein-2 bound in which sampling error decomposes into initialisation error, score-estimation error, and numerical discretisation error, and shows that minimizing a score-matching-type loss is equivalent to minimizing the KL divergence between the true forced process and the approximating one (Pieper-Sethmacher et al., 6 Feb 2026).

The same bridge principle appears in image restoration through the Generalized Ornstein–Uhlenbeck Bridge. Starting from

$V$1

the generalized Doob transform adds

$V$2

to the drift, producing a point-to-point bridge toward a fixed $V$3. The paper presents closed-form conditional transitions, a reverse SDE, and a probability-flow or Mean-ODE formulation; it also identifies Brownian bridge, BBDM, DDBM (VE), and $V$4SB as special cases or equivalent cases within this broader bridge viewpoint (Yue et al., 2023).

A variational version appears in transition path sampling. For a reference diffusion

$V$5

conditioning on $V$6 or $V$7 yields the classical drift correction

$V$8

The paper "Doob’s Lagrangian" reformulates this exact conditioned process as a path-space KL minimization and a constrained minimization over marginals $V$9 and controls $h$00, with hard endpoint constraints. At optimum,

$h$01

so the variational solution is not merely analogous to the Doob transform; it is an equivalent characterization of it (Du et al., 2024).

5. Computational approximations and diffusion-model guidance

Modern computational work often uses the generalized Doob transform as an exact target and then approximates $h$02 or $h$03. In diffusion models, the common pattern is that a pretrained reverse process is left largely unchanged while an additional drift term, interpreted as Doob guidance, transports the base generative distribution toward a conditional or reward-tilted target.

Framework	$h$04-object	Main use
DEFT (Denker et al., 2024)	$h$05	Small-network correction on a frozen unconditional diffusion model
DOIT (Zhu et al., 18 Feb 2026)	$h$06	Training-free reward adaptation with Monte Carlo Doob correction
Doob’s matching (Chang et al., 10 Jan 2026)	$h$07	Regression-based guidance estimation with non-asymptotic rates
h-Edit (Nguyen et al., 4 Mar 2025)	$h$08 on the backward chain	Reverse-time bridge editing and compositional editing terms
CDT filtering (Chopin et al., 2022)	$h$09	Approximate fully adapted auxiliary particle filters

In DEFT, the generalized posterior transform is

$h$10

and only a small network $h$11 is trained while the large unconditional model remains fixed. The paper reports speedups of up to $h$12 on image reconstruction tasks and relates DPS, $h$13GDM, reconstruction guidance, and conditional score models to exact or approximate forms of the same generalized $h$14-transform (Denker et al., 2024).

DOIT defines

$h$15

and corrects the reverse drift by $h$16. The key estimator uses backward rollouts and differentiates only Gaussian reverse kernels, not the reward itself, so the method applies to non-differentiable rewards. The theory provides a high-probability convergence guarantee to the target high-reward distribution and a dimension-independent local Monte Carlo rate that deteriorates when $h$17 is small, i.e. near the end of the reverse trajectory (Zhu et al., 18 Feb 2026).

Doob’s matching starts from a tilted target

$h$18

and defines

$h$19

A gradient-penalized regression loss estimates both $h$20 and $h$21, producing a consistent guidance estimator and non-asymptotic convergence guarantees for the generated distribution in $h$22 (Chang et al., 10 Jan 2026).

In h-Edit, the generalized transform is applied directly to the backward chain: $h$23 The resulting Langevin update decomposes into a reconstruction term and an editing term, and products of $h$24-functions yield additive editing scores for multi-objective editing (Nguyen et al., 4 Mar 2025).

Outside generative modeling, online filtering of discretely observed diffusions uses the exact conditioned drift

$h$25

which defines the fully adapted auxiliary particle filter. Because the exact $h$26-transform is usually intractable, the paper approximates the backward Kolmogorov solution with neural networks through nonlinear Feynman–Kac and BSDE representations, producing a locally optimal particle filter that is reported to be orders of magnitude more efficient than state-of-the-art particle filters in difficult regimes (Chopin et al., 2022).

A parameter-efficient conditional adaptation variant appears in virtual try-on. DEFT-VTON freezes the pre-trained backbone and trains an $h$27-transform network, using only $h$28 percent of the frozen parameters compared to $h$29 percent in the baseline PEFT setup, and reports operation with as few as $h$30 denoising steps (Xu et al., 16 Sep 2025).

6. Conceptual boundaries, equivalences, and recurrent misconceptions

A frequent misconception is that Doob’s $h$31-transform is only a finite-dimensional conditioning device based on a strictly harmonic state function. The generalized theory is broader. It includes endpoint weights, integrated potentials, and rough measurable data under finite entropy conditions; it also includes additive and product-type transforms on measure-valued processes, ground-state transforms of non-local Schrödinger semigroups, and space-time harmonic functions built from conditional expectations rather than explicit PDE solutions [(Léonard, 2011); (Hénard, 2012); (Lorinczi et al., 2017)].

A second misconception is that the transform is synonymous with time reversal. In modern diffusion modeling, the distinction is explicit. Infinite-dimensional generative diffusion via Doob’s transform imposes the target at time $h$32 by an exponential change of measure and an initial-law correction, rather than by reversing a noising process; the paper emphasizes that the transformed process reaches $h$33 exactly at time $h$34 regardless of whether $h$35 is large or small (Pieper-Sethmacher et al., 6 Feb 2026). This differs conceptually from approaches whose correctness depends on the forward noising law approaching a tractable reference.

A third misconception is that practical guidance methods are automatically exact instances of the transform. The literature distinguishes carefully between exact transformed dynamics and approximations. In inverse problems, DPS-style reconstruction guidance and Gaussian moment-matching are presented as approximations to the exact conditional term $h$36 (Denker et al., 2024). In filtering, the fully adapted proposal is the exact Doob-transformed diffusion, while the learned BSDE-based method is a computational approximation to that proposal (Chopin et al., 2022). In reward-guided sampling, DOIT and Doob’s matching both provide theory for approximation error rather than claiming closed-form access to $h$37 (Zhu et al., 18 Feb 2026, Chang et al., 10 Jan 2026).

A fourth point concerns equivalence with control, bridges, and variational principles. Several papers show that the same transformed process can be viewed as a bridge, a control problem, or a path-space KL minimizer. Doob’s Lagrangian proves that the conditioned diffusion is the exact minimizer of a constrained variational problem over path laws, with optimal control $h$38 (Du et al., 2024). The exit-conditioning literature rewrites $h$39 through a Hopf–Cole transform

$h$40

so that the conditioned drift becomes $h$41, linking $h$42-transforms to viscous Hamilton–Jacobi–Bellman asymptotics and Gaussian scaling limits for rare exits (Bakhtin et al., 2013). In self-similar process theory, MAP reversibility, inversion, weak duality, and $h$43-transforms are shown to be different manifestations of the same reversible structure (Alili et al., 2016).

Taken together, these results define the generalized Doob’s $h$44-transform not as a single formula but as a unifying probabilistic mechanism: a Markov law is reweighted by a positive space-time factor, the transformed law remains Markov, and the new dynamics can be read as exact conditioning, duality, bridge construction, control, or inference-time steering, depending on the ambient problem class.