Infinite-Dimensional Diffusion Bridges

• Infinite-dimensional diffusion bridges are the laws of stochastic processes in separable Hilbert spaces conditioned to meet terminal constraints, generalizing classical finite-dimensional bridges.
• They employ techniques like Doob’s h-transform, stochastic optimal control, and neural operator learning to address intractable infinite-dimensional transition densities.
• These bridges enable high-fidelity conditional simulation for applications such as SPDEs, inverse problems, and generative models for function-valued data.

An infinite-dimensional diffusion bridge is the law of a stochastic process governed by a stochastic differential equation (SDE) in a separable Hilbert space, conditioned to attain a given terminal value or to satisfy a linear constraint at a specified time. This construction generalizes classical finite-dimensional diffusion bridges and is fundamental for modeling conditioned dynamics in function spaces, such as for stochastic partial differential equations (SPDEs), conditional simulation, inverse problems, and modern score-based generative frameworks for function-valued data. Rigorous frameworks for infinite-dimensional bridges address the technical challenges arising from intractable transition densities and the lack of Lebesgue measure or closed-form conditioning in infinite-dimensional settings. The theory leverages analytic tools including Doob's $h$-transform, stochastic optimal control, operator-learning, and stochastic interpolant methods adapted to function spaces (Pieper-Sethmacher et al., 17 Mar 2025, Yang et al., 2024, Yu et al., 2 Feb 2026, Park et al., 2024).

1. Mathematical Foundations of Infinite-Dimensional Bridges

Consider the stochastic evolution equation (in mild form)

$$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$

where $H$ is a real, separable Hilbert space, $A$ generates a $C_0$-semigroup $(S_t)_{t\ge 0}$, $F$ is nonlinear, and $W_t$ is a $Q$-Wiener process with $Q$ trace-class and positive-definite. The mild solution is

$$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$0

guaranteed to exist and be unique under standard Lipschitz and trace-class assumptions. The diffusion bridge is defined as the law of $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$1 conditioned on $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$2 for a bounded linear map $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$3 (Pieper-Sethmacher et al., 17 Mar 2025).

The conditioning is rigorously formalized via the $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$4-function,

$$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$5

leading to the Radon–Nikodym derivative for the bridge law on $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$6:

$$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$7

This “Doob $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$8-transform” modifies the drift to guide the process toward the conditioning. In practice, $$dX_t = [A X_t + F(t, X_t)]\,dt + \sqrt{Q}\,dW_t, \quad X_0 = x_0 \in H,$$9 is generally unavailable in nonlinear/infinite-dimensional settings, motivating auxiliary constructions.

Absolute continuity and the existence of a bridge law on $H$0 require: (i) nondegeneracy of $H$1 in observed directions, (ii) invertibility of $H$2 for $H$3, and (iii) sufficient regularity for exponential martingale properties. Under these, the shift induced by the $H$4-transform lies in the Cameron–Martin space (i.e., the reproducing kernel Hilbert space of the Gaussian reference measure), ensuring absolute continuity (Pieper-Sethmacher et al., 17 Mar 2025, Park et al., 2024).

2. Doob’s $H$5-Transform and Stochastic Optimal Control

Doob’s $H$6-transform realizes the bridge as a solution to a controlled SDE:

$H$7

where the additional drift term $H$8 pins the process to the desired terminal value. The change-of-measure formula for the path-space density is

$H$9

When $A$0 is unavailable, practical methods substitute computable and tractable “guided” processes (e.g., based on the Ornstein–Uhlenbeck bridge), sampled under an auxiliary law $A$1, with appropriate correction via importance sampling or MCMC (Pieper-Sethmacher et al., 17 Mar 2025).

An alternative, rigorous construction frames the bridge as a stochastic optimal control (SOC) problem in the Hilbert space:

• Control $A$2 with values in the Cameron–Martin space $A$3; the controlled SDE is

$A$4

• The optimal control solution yields a bridge where

$A$5

and the controlled process solves the same SDE as the $A$6-transformed bridge (Park et al., 2024). Notably, this framework covers both linear and nonlinear dynamics with mild solutions and is tightly linked to infinite-dimensional Hamilton–Jacobi–Bellman equations.

3. Operator-Based and Machine-Learned Bridge Simulation

Recent advances deploy operator-learning—specifically, mesh-free neural operator architectures—for learning the intractable correction drift $A$7 directly in function space. The key steps include:

• Parameterizing the true correction operator as $A$8.
• Training $A$9 to minimize a variational (score-matching/KL) loss implicitly matching the law of the time-reversed bridge:

$C_0$0

where $C_0$1 approximates the score via local transitions (Yang et al., 2024).

The G_\theta operator is realized via a continuous-time U-shaped Fourier neural operator (CT-UNO), which is inherently discretization-invariant: once trained, the operator can be deployed at any spatial or temporal resolution without retraining. The algorithm proceeds by training on finite-dimensional projections and drawing bridge paths via a backward Euler–Maruyama scheme, applying the learned correction (Yang et al., 2024). This procedure enables high-fidelity, mesh-free sampling of infinite-dimensional bridges when analytic drift corrections are unavailable.

4. Stochastic Interpolants and Bridge SDEs in Hilbert Spaces

A complementary framework—stochastic interpolants in Hilbert spaces—defines a process that smoothly interpolates between arbitrary functional distributions $C_0$2 on $C_0$3. The interpolant

$C_0$4

can be lifted to a conditional bridge SDE (CB-SDE):

$C_0$5

where the drift $C_0$6 splits into deterministic “velocity” $C_0$7 and “denoiser” $C_0$8 terms, formally:

$C_0$9

CB-SDEs admit rigorous existence, uniqueness, and error bounds under realistic conditions, including well-posedness in the Cameron–Martin norm and explicit Wasserstein error estimates for neural approximations of drift terms (Yu et al., 2 Feb 2026).

5. Numerical Methods and Sampling Algorithms

Bridge simulation in infinite dimensions centers on discretization (e.g., spectral-Galerkin via $(S_t)_{t\ge 0}$0 eigenfunctions) and auxiliary “guided” SPDE solvers. The leading computational frameworks and their key steps include:

Method	Principle	Correction Downstream
Guided MCMC (Metropolis–Hastings) (Pieper-Sethmacher et al., 17 Mar 2025)	Simulate guided process with tractable auxiliary drift, reweight paths using importance weight $(S_t)_{t\ge 0}$1	Accept/reject proposals based on ratio of path weights
Score-matching operator learning (Yang et al., 2024)	Learn drift correction with neural operators on function space, train using variational loss	Backward Euler–Maruyama with learned correction
Stochastic Interpolants (SI) (Yu et al., 2 Feb 2026)	Use interpolation schemes, fit $(S_t)_{t\ge 0}$2 with neural operators, simulate bridge SDE	Error bounds in Wasserstein distance for learned drift

Concrete examples include stochastic reaction–diffusion SPDEs (Michaelis–Menten kinetics, stochastic Allen–Cahn), functional Brownian bridges, conditioned biological shape trajectories, and PDE inverse problems (e.g., Darcy flow, Navier–Stokes) (Pieper-Sethmacher et al., 17 Mar 2025, Yang et al., 2024, Yu et al., 2 Feb 2026).

6. Theoretical Guarantees and Empirical Results

Theoretical contributions across the frameworks include:

• Path-space absolute continuity of bridge measures under sufficient nondegeneracy and regularity (Pieper-Sethmacher et al., 17 Mar 2025, Park et al., 2024).
• Existence, uniqueness, and well-posedness of strong solutions to the bridge SDE/SPDE in $(S_t)_{t\ge 0}$3 or the Cameron–Martin space (Yu et al., 2 Feb 2026, Park et al., 2024).
• Explicit error bounds: CB-SDE discretization error in Wasserstein-2 is controlled by the mean-square errors for the learned drift components, with convergence as the approximation improves (Yu et al., 2 Feb 2026).
• Score-matching loss for operator-learned drift is provably equivalent (up to constants) to KL divergence of true and learned reversed bridge laws (Yang et al., 2024).

Empirically, infinite-dimensional bridge frameworks achieve state-of-the-art performance for function-valued conditional generation, including:

• Low mean-squared error and stable sampling for functional bridges across resolutions (Yang et al., 2024).
• Relative $(S_t)_{t\ge 0}$4-error for stochastic interpolants: e.g., 2D Navier–Stokes forward/inverse tasks reaching 1.0%/4.6% error, outperforming several diffusion PDE baselines (Yu et al., 2 Feb 2026).
• Effective mesh-free generative models for images, time series, and probabilistic function-valued data, capable of zero-shot upscaling or interpolation (Park et al., 2024).

7. Applications and Broader Significance

Infinite-dimensional diffusion bridges underpin a wide spectrum of modern statistical and computational sciences:

• Conditional path sampling for SPDEs, enabling data assimilation, rare event simulation, and uncertainty quantification (Pieper-Sethmacher et al., 17 Mar 2025).
• Operator-based generative modeling, supporting mesh-invariant image synthesis, functional regression, or imputation, bridging initial and terminal distributions without retraining (Yang et al., 2024, Park et al., 2024).
• Conditional generation and statistical inference for PDE-based scientific benchmarks, such as conditioned solutions to Navier–Stokes and Darcy flow equations (Yu et al., 2 Feb 2026).
• The rigorous grounding in stochastic control, functional analysis, and operator learning connects classical probabilistic methods with modern machine-learned simulation, extending the scope and tractability of bridge-based approaches in infinite-dimensional function spaces.

Collectively, these methodologies establish both a theoretical and practical foundation for simulating and analyzing conditioned stochastic dynamics in infinite-dimensional settings, with broad utility across inverse problems, generative modeling, computational physics, and uncertainty-aware scientific discovery.