Generalized Doob Transform in Stochastic Dynamics

• The generalized Doob transform is a measure-changing construction that uses positive eigenfunctions to modify transition probabilities and drifts, making rare events typical.
• It systematically recasts conditioning problems into altered dynamics by reweighting stochastic processes, preserving key structural properties.
• Its applications span discrete Markov chains, continuous diffusions, infinite-dimensional SPDEs, and quantum systems, providing a unified framework for rare-event sampling and optimal control.

Searching arXiv for papers on generalized Doob transforms and closely related Doob $h$-transform formulations across Markov chains, diffusions, SPDEs, and quantum transport. The generalized Doob transform is a change-of-measure construction that converts a reference Markov dynamics into a new dynamics in which a prescribed conditioning, bias, or rare-event ensemble becomes typical. In the discrete-time setting it is defined from a positive eigenfunction $h$ of a stochastic matrix; in diffusion settings it appears as a drift modification by $\nabla \log h$; in infinite-dimensional SPDEs it is an exponential change of measure that forces the terminal law to match a target distribution exactly; and in open quantum systems it is a gauge transformation of a tilted GKSL generator that produces a completely positive, trace-preserving dynamics whose stationary state realizes the biased ensemble (Chen et al., 2019, Pieper-Sethmacher et al., 6 Feb 2026, Esteve et al., 6 Aug 2025).

1. Definition and core mechanism

In its classical discrete-time form, let $P=(P_{ij})_{i,j\in\Omega}$ be an irreducible stochastic matrix on a countable state space $\Omega$, and let $h:\Omega\to(0,\infty)$ satisfy

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$

The generalized Doob transform of $P$ by $(h,\lambda)$ is the stochastic matrix

$$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$

Its stochasticity follows immediately from the eigenfunction relation (Chen et al., 2019).

The same mechanism can be expressed as a martingale change of measure. For a Markov chain $h$0, if $h$1, then

$h$2

is a positive martingale, and the transformed path measure satisfies

$h$3

Under the new measure, the one-step transition kernel becomes exactly $h$4 (Chen et al., 2019).

For one-dimensional diffusions with generator

$h$5

a positive harmonic function $h$6 yields the transformed generator

$h$7

so the diffusion coefficient is unchanged and the drift is shifted by $h$8 (Alili et al., 2012). In finite-dimensional fixed-time diffusion conditioning, the same structure appears as

$h$9

where $\nabla \log h$0 solves the backward Kolmogorov equation (Mamede et al., 23 May 2026).

In infinite dimensions, the transformed drift has the analogous form

$\nabla \log h$1

with $\nabla \log h$2 defined by a forward-backward density formula and the resulting process satisfying $\nabla \log h$3 exactly (Pieper-Sethmacher et al., 6 Feb 2026). In quantum Markov dynamics, the transform is implemented by a positive left eigenmatrix $\nabla \log h$4 of a tilted Lindbladian, giving

$\nabla \log h$5

which produces a bona fide GKSL generator (Esteve et al., 6 Aug 2025).

2. Conditioning, tilting, and rare-event realization

A central role of the generalized Doob transform is to replace conditioning or exponential tilting at the path level by a Markovian dynamics on a suitable state space. In endpoint conditioning for a discrete-time Markov chain with transition matrix $\nabla \log h$6, the value functions

$\nabla \log h$7

solve the backward equation

$\nabla \log h$8

and the conditioned dynamics is generated by

$\nabla \log h$9

This is the standard finite-horizon Doob construction (Coghi et al., 3 Mar 2025).

The same paper extends the construction to conditioning on trajectory functionals depending on the empirical occupation measure. If

$P=(P_{ij})_{i,j\in\Omega}$0

then the value functions must depend on both $P=(P_{ij})_{i,j\in\Omega}$1 and $P=(P_{ij})_{i,j\in\Omega}$2, and the generalized Doob process becomes

$P=(P_{ij})_{i,j\in\Omega}$3

The chain is therefore non-Markovian in $P=(P_{ij})_{i,j\in\Omega}$4 alone but Markovian in the enlarged state space $P=(P_{ij})_{i,j\in\Omega}$5 (Coghi et al., 3 Mar 2025).

In large-deviation settings, the generalized Doob transform makes rare events typical. For open quantum transport, the tilted generator $P=(P_{ij})_{i,j\in\Omega}$6 is built by inserting counting factors $P=(P_{ij})_{i,j\in\Omega}$7 into jump terms, and its dominant eigenvalue $P=(P_{ij})_{i,j\in\Omega}$8 is the scaled cumulant generating function. The typical current in the tilted ensemble is

$P=(P_{ij})_{i,j\in\Omega}$9

and the Doob-transformed steady-state current satisfies

$\Omega$0

so choosing $\Omega$1 systematically increases the stationary current (Esteve et al., 6 Aug 2025).

For fixed-time conditioning of small-noise diffusions, the conditioned ensemble can also be treated through a variational principle. The moment-generating function

$\Omega$2

has the weak-noise scaling

$\Omega$3

with $\Omega$4 represented by a Freidlin-Wentzell action or, equivalently, by Hamilton-Jacobi-Bellman dynamics. The controlled drift

$\Omega$5

has the same form as the Doob transform when $\Omega$6 to leading order (Mamede et al., 23 May 2026). This suggests a close operational relationship between Doob conditioning and optimal control in the weak-noise regime.

3. Diffusions, SPDEs, and transformed generators

For regular linear diffusions on an interval $\Omega$7, the generalized Doob transform is intertwined with scale and speed structures. If $\Omega$8 is $\Omega$9-harmonic, then the transformed semigroup satisfies

$h:\Omega\to(0,\infty)$0

and the transition density changes by

$h:\Omega\to(0,\infty)$1

with respect to the speed measure (Alili et al., 2012). In the Darboux-transformation formulation for second-order diffusion operators,

$h:\Omega\to(0,\infty)$2

and the transformed density becomes

$h:\Omega\to(0,\infty)$3

when $h:\Omega\to(0,\infty)$4 (Kuznetsov et al., 2024).

The one-dimensional theory also admits a pathwise representation by inversion and time change. With

$h:\Omega\to(0,\infty)$5

the process

$h:\Omega\to(0,\infty)$6

is precisely the Doob-$h:\Omega\to(0,\infty)$7 transform $h:\Omega\to(0,\infty)$8 (Alili et al., 2012). In self-similar Markov processes on $h:\Omega\to(0,\infty)$9, an analogous inversion

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$0

coincides with a Doob $$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$1-process for

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$2

under a reversibility assumption on the underlying Markov additive process (Alili et al., 2016).

In infinite-dimensional generative diffusions, the transform is defined from the strictly positive space-time harmonic function

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$3

where $$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$4 are transition densities with respect to a Gaussian reference measure $$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$5. Under the changed measure,

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$6

with steering field

$$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$7

and the transformed SPDE has terminal law $$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$8 exactly (Pieper-Sethmacher et al., 6 Feb 2026).

A practical consequence is that approximate steering fields $$(Ph)(i)=\sum_{j\in\Omega}P_{ij}h(j)=\lambda h(i), \qquad \lambda>0.$$9 can be trained by minimizing a conditional score-matching loss. The path-space KL divergence satisfies

$P$0

which leads to a variational principle for approximating the exact transformed process (Pieper-Sethmacher et al., 6 Feb 2026).

4. Quantum and non-commutative formulations

In Markovian open quantum systems, the generalized Doob transform is formulated at the level of GKSL generators. Starting from

$P$1

one defines a tilted generator $P$2 by inserting counting factors $P$3 into jump terms associated with the time-extensive observable $P$4 (Esteve et al., 6 Aug 2025).

If $P$5 and $P$6 are the right and left eigenmatrices corresponding to the eigenvalue $P$7 of largest real part,

$P$8

then the quantum Doob generator is

$P$9

It is completely positive and trace-preserving and has unique steady state

$(h,\lambda)$0

(Esteve et al., 6 Aug 2025).

In GKSL form, the transformed dynamics reads

$(h,\lambda)$1

with

$(h,\lambda)$2

and

$(h,\lambda)$3

Both the coherent part and the dissipative part are therefore modified in general (Esteve et al., 6 Aug 2025).

A distinct non-commutative direction appears in operator-algebraic work on Markov chains. There, generalized Doob transforms characterize when two irreducible stochastic matrices are Doob equivalent, meaning that their multi-step conditional probabilities coincide up to relabeling. The main characterization states that $(h,\lambda)$4 is Doob equivalent to $(h,\lambda)$5 if and only if $(h,\lambda)$6 is conjugate to a transform $(h,\lambda)$7 for some positive eigenfunction $(h,\lambda)$8 and eigenvalue $(h,\lambda)$9 (Chen et al., 2019). This feeds into the classification of the associated tensor algebras $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$0, which are completely isometrically isomorphic precisely when the underlying chains are Doob equivalent (Chen et al., 2019).

5. Structural relations: duality, inversion, symmetry, and equivalence

Several works identify the generalized Doob transform with deeper structural relations rather than only with conditioning. For self-similar Markov processes, weak duality with respect to

$$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$1

is equivalent to reversibility of the underlying MAP, and the inverted process $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$2 is precisely the Doob $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$3-transform with

$$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$4

In the isotropic case this reduces to $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$5 (Alili et al., 2016).

For one-dimensional diffusions, inversion is again central. The involution $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$6 satisfies $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$7, and the transformed process is obtained from the original one by the deterministic inversion $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$8 and a random clock $$P^{(h,\lambda)}_{ij}=\frac{1}{\lambda}\,\frac{h(j)}{h(i)}\,P_{ij}.$$9 (Alili et al., 2012). In the Brownian-motion-with-drift and Bessel-process examples, this yields explicit transformed drifts, inversions, and clocks (Alili et al., 2012).

The Darboux-transform framework combines Doob’s $h$00-transform with Siegmund duality. Starting from a killed diffusion $h$01, one applies a first Doob transform to obtain a conservative diffusion $h$02, then passes to the Siegmund dual $h$03, and then applies a second Doob transform to obtain $h$04. The resulting transition kernels satisfy the explicit relation

$h$05

for $h$06 (Kuznetsov et al., 2024). This suggests that generalized Doob transforms can serve as one component of larger intertwining constructions.

In the quantum-transport setting, the transform is also linked to symmetry. Centrosymmetry of an $h$07-site Hamiltonian is measured by

$h$08

and numerical results show that the Doob-transformed Hamiltonian $h$09 almost always has a smaller $h$10 for the same $h$11 that boosts the current $h$12 (Esteve et al., 6 Aug 2025). A plausible implication is that the transform not only reweights trajectories but can reveal latent geometric structure associated with efficient transport.

6. Applications and domain-specific realizations

The generalized Doob transform now appears across several distinct research programs.

In quantum transport, a single diagonalization of the tilted Liouvillian yields an optimized GKSL generator that tailors both Hamiltonian and dissipative contributions to improve currents and activities. Robustness can be probed by constraining the transformed dynamics, for example by replacing the transformed jump operators with the original set or by resetting a specific matrix element of the transformed Hamiltonian (Esteve et al., 6 Aug 2025).

In generative diffusions on infinite-dimensional spaces, the transform replaces time reversal by an exponential change of measure relative to a reference diffusion. Under the stated assumptions on $h$13, the transition densities, and the regularity of $h$14, the transformed process exists uniquely and samples the target terminal law exactly (Pieper-Sethmacher et al., 6 Feb 2026). Approximation by learned steering fields admits a Wasserstein error bound,

$h$15

when the score field is Lipschitz, the initial law is approximated by a Langevin sampler, and discretization uses semi-implicit Euler (Pieper-Sethmacher et al., 6 Feb 2026).

In self-interacting processes, the generalized Doob transform shows that Markov processes with constrained occupation measures are realized optimally by self-interacting dynamics (Coghi et al., 3 Mar 2025). For random-walk bridges, excursions, and forced excursions, the transformed rates are explicit. For example, in the bridge case,

$h$16

while the excursion case uses value functions from the Catalan triangle and produces modified rates with both positivity and finite-horizon corrections (Coghi et al., 3 Mar 2025).

In diffusion LLMs, Doob guidance is used as a token-ordering rule rather than as a drift for a physical diffusion. Given a base reveal law $h$17 and terminal reward $h$18, one defines

$h$19

and the exact transformed policy is

$h$20

This reward-tilted Gibbs reveal law is approximated stagewise by Soft-BoN, with terminal-KL bound

$h$21

under $h$22, finite action sets, and $h$23 (Bu et al., 27 Apr 2026).

Across these settings, the common pattern is the replacement of rejection-based or post-selected conditioning by an explicit transformed dynamics. This suggests that the generalized Doob transform functions as a unifying device for exact conditioning, rare-event sampling, and structure-preserving model design across classical, quantum, and data-driven stochastic systems.