Bridge Sampling for Bayesian and Stochastic Processes

• Bridge sampling is a computational design that replaces complex target distributions with auxiliary densities or processes to improve overlap and enhance estimability.
• It is widely applied in Bayesian model selection and marginal likelihood estimation as well as reverse-time sampling within diffusion and Schrödinger bridge frameworks.
• Modern implementations utilize neural flows and adaptive proposals to efficiently tackle high-dimensional problems and rare-event sampling challenges.

Bridge sampling denotes a family of Monte Carlo and stochastic-process constructions in which an auxiliary bridge is introduced between otherwise difficult objects. In Bayesian computation, it estimates a normalizing constant or marginal likelihood by bridging a target density and a proposal density (Gronau et al., 2017, Jia et al., 2019). In stochastic-process settings, it means simulating a diffusion process conditioned to hit a prescribed endpoint at a fixed terminal time (Grong et al., 2024). In recent machine-learning work, diffusion bridges and Schrödinger bridges are used to transport samples between source and target distributions, to sample from unnormalized densities, or to construct efficient reverse-time solvers (Pan et al., 23 May 2025, Liu et al., 27 Jun 2025). This suggests that bridge sampling is best understood not as a single algorithm, but as a common design principle: replace a hard target by an auxiliary density, path law, or controlled process chosen to improve overlap, conditioning, or transport.

1. Core meanings and formal settings

The literature uses the term in several technically distinct senses, each centered on a bridge object with better computational properties than the original target.

Setting	Bridge object	Representative formulation
Normalizing-constant estimation	Density bridge between $p$ and $q$	$$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$ (Jia et al., 2019)
Conditioned stochastic processes	Endpoint-constrained diffusion path	Bridge from $x_0$ to $x_T$ over $[0,T]$ is the law of $(X_t)$ conditioned on $X_T=x_T$ (Grong et al., 2024)
Diffusion-bridge transport	Controlled stochastic path between source and target distributions	Unified Diffusion Bridge and Schrödinger Bridge formulations as stochastic optimal control problems (Pan et al., 23 May 2025, Liu et al., 27 Jun 2025)
Rare-event path sampling	Bridge ensemble reweighted to recover target path statistics	Sample endpoint-constrained trajectories and assign exact statistical weights (Aguilar et al., 2021)

In Bayesian statistics, the key normalizing constant is the marginal likelihood

$$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$

which is central for Bayes factors, model selection, and Bayesian model averaging (Micaletto et al., 20 Aug 2025, Gronau et al., 2017). In conditioned-diffusion work, the target is instead a path law, and the bridge is a stochastic process pinned at both endpoints (Grong et al., 2024). In diffusion-model and Schrödinger-bridge work, the target may be an unnormalized density $\nu(x)\propto e^{-E(x)}$, and the bridge is a controlled diffusion whose terminal marginal approximates $q$0 (Liu et al., 27 Jun 2025).

A plausible implication is that the common computational role of a bridge is to convert a rare, poorly overlapping, or intractable target into a sequence or family of better matched objects. The precise form of the bridge—density ratio, endpoint-conditioned path, or controlled transport—depends on the problem class.

2. Classical bridge sampling for normalizing constants and marginal likelihoods

In its classical statistical form, bridge sampling estimates a ratio of normalizing constants by introducing a second density $q$1 with known normalizing constant $q$2 and substantial overlap with an unnormalized target $q$3. For any bridge function $q$4,

$q$5

which yields

$q$6

(Jia et al., 2019). The tutorial treatment aimed at marginal likelihoods presents the same idea in posterior–proposal form and emphasizes that bridge sampling is more robust than importance sampling or the generalized harmonic mean estimator because it is less sensitive to tail behavior (Gronau et al., 2017).

The optimal bridge function in the Meng–Wong sense appears in several of the papers. For unnormalized densities $q$7 with sample fractions $q$8, the asymptotically optimal bridge is

$q$9

and the corresponding asymptotic variance is controlled by the overlap between the normalized densities (Wang et al., 2016). In practical Bayesian implementations, the bridge function depends on the unknown marginal likelihood, so the estimator is computed iteratively. One standard form is

$$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$0

with $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$1 and $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$2 given by posterior-to-proposal ratios (Micaletto et al., 20 Aug 2025, Gronau et al., 2017).

A further generalization appears for energy-based models. When the parameter $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$3 is fixed at the true value and only the partition function $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$4 is unknown, the paper on contrastive learning, importance sampling, and bridge sampling shows that noise-contrastive estimation in $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$5-space, reverse logistic regression, and optimal bridge sampling are equivalent under the paper’s assumptions (Martino, 9 Apr 2026). In that formulation, the optimal bridge estimator becomes a fixed-point recursion for $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$6, and bridge sampling occupies the central position in a broader family of classification- and importance-sampling-based partition-function estimators.

3. Overlap engineering, adaptive proposals, and reliability assessment

The dominant determinant of bridge-sampling performance is overlap. The Warp bridge literature makes this explicit: Warp-I, Warp-II, and Warp-III are location, location-scale, and random symmetrization transforms effective for uni-modal targets, while Warp-U is designed for multi-modal densities (Wang et al., 2016). Warp-U begins from a Normal mixture

$$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$7

then applies the stochastic componentwise inverse map that sends $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$8 back to $$\mathcal Z_p=\frac{\langle \alpha(\mathbf x)p(\mathbf x)\rangle_q}{\langle \alpha(\mathbf x)q(\mathbf x)\rangle_p}\mathcal Z_q$$9. If $x_0$0, its Warp-U transform is $x_0$1, and the transformed density has the same normalizing constant as the original target. The key theorem is

$x_0$2

so overlap cannot get worse in any $x_0$3-divergence sense (Wang et al., 2016). The later Warp-U sampler paper extends this idea using location-scale-skew mixtures and neural ODE-based maps, and proposes a stochastic Warp-U bridge estimator that estimates component normalizers separately, reducing evaluation cost while preserving the variance-reduction effect of the transformation (Ding et al., 2024).

A different route to proposal design is Gaussianized Bridge Sampling. There the proposal is constructed by fitting a normalizing flow to posterior samples and then using Optimal Bridge Sampling on the learned analytic density (Jia et al., 2019). The target is $x_0$4, the learned proposal is

$x_0$5

and $x_0$6. The proposal is built with an Iterative Neural Transform, the posterior draws are split into two batches to avoid evidence underestimation, and the optimal ratio $x_0$7 is obtained as the unique root of a monotone score equation (Jia et al., 2019). On a 16D Funnel, 32D Banana, 48D Cauchy mixture, and 64D Ring, the paper reports that GBS is substantially more accurate than Nested Sampling and AIS at comparable cost, that AIS / RAIS needs about $x_0$8–$x_0$9 more evaluations to reach comparable accuracy, and that a reduced-cost GBSL uses about $x_T$0 fewer evaluations than full GBS while remaining unbiased (Jia et al., 2019).

Bridge construction can also be adaptive along a sequence of intermediate targets. The Shortened Bridge Sampler starts from a deterministic approximation $x_T$1 and bridges to the exact posterior through

$x_T$2

with $x_T$3 chosen by a conditional ESS criterion (Donnet et al., 2017). This paper emphasizes robustness when the approximation underestimates variance or is poorly centered, and it reports that inference on network datasets can lead to different statistical conclusions from the standard variational Bayes approximation (Donnet et al., 2017).

Reliability assessment has become a topic in its own right. The diagnostics paper derives a delta-method Monte Carlo standard error for bridge sampling, applies Pareto-$x_T$4 to the numerator and denominator summands, and proposes block reshuffling to assess instability induced by posterior splitting and iterative proposal fitting without rerunning full inference (Micaletto et al., 20 Aug 2025). On 41 posteriors from posteriordb plus a Gaussian-process example, 36 out of 42 have MCSE below $x_T$5, which the paper treats as excellent; for harder posteriors the required number of posterior draws can increase dramatically (Micaletto et al., 20 Aug 2025). This suggests that practical bridge sampling is now understood not only as an estimator, but also as a diagnostic workflow involving overlap design, uncertainty quantification, and failure detection.

4. Conditioned diffusion processes and geometric bridge simulation

In stochastic-process theory, bridge sampling means simulating a diffusion process conditioned on a terminal endpoint. If $x_T$6 is an unconditioned diffusion with generator $x_T$7, the bridge from $x_T$8 to $x_T$9 over $[0,T]$0 is the law of $[0,T]$1 conditioned on $[0,T]$2 (Grong et al., 2024). The bridge generator is

$[0,T]$3

and time reversal yields a computationally useful form in which the reversed bridge uses the score from the starting point,

$[0,T]$4

(Grong et al., 2024). The paper’s main difficulty is sub-Riemannian geometry: diffusion is constrained to the horizontal bundle $[0,T]$5, the metric is degenerate on $[0,T]$6, the generator is hypoelliptic rather than elliptic, and the score depends on a horizontal gradient.

Score matching is used to learn this intractable score. In a local orthonormal frame $[0,T]$7, the network outputs a horizontal vector field

$[0,T]$8

and the intractable score-matching objective

$[0,T]$9

is converted by integration by parts into the divergence loss

$(X_t)$0

(Grong et al., 2024). The denoising version must be modified because short-time increments in sub-Riemannian geometry are not approximately Gaussian in the full tangent space: first-order motion is horizontal, while missing directions arise only through Lie brackets and Lévy area terms. The paper therefore uses a stochastic Taylor expansion in non-holonomic frames and adapted coordinates, motivating the short-time approximation

$(X_t)$1

(Grong et al., 2024).

The Heisenberg group provides the explicit test case. Its horizontal bundle is generated by

$(X_t)$2

with bracket relation $(X_t)$3, making the vertical direction bracket-generated (Grong et al., 2024). Numerical experiments show that a bridge from $(X_t)$4 to $(X_t)$5 over $(X_t)$6 follows a path close to the minimizing geodesic, and that bridges concentrate more strongly around the minimizing geodesic as $(X_t)$7. The paper also reports that both denoising and divergence losses produce qualitatively correct score fields near the origin, but the divergence loss is less stable in practice, especially in low-density regions, so the denoising loss is more robust for these sub-Riemannian examples (Grong et al., 2024).

Bridge proposals can also be embedded inside inference procedures. The differentiable diffusion bridge importance-sampling paper uses score matching to approximate bridge drifts, then uses bridge-path samples as proposals in an importance sampler for otherwise intractable transition densities (Boserup et al., 2024). The path-space identity

$(X_t)$8

is turned into a differentiable log-likelihood estimator via Euler–Maruyama Gaussian approximations, stable log-weights, and a log-sum-exp aggregation. The paper then performs gradient ascent for variance-parameter inference and diffusion-mean estimation in biological morphometry (Boserup et al., 2024).

5. Diffusion bridges and Schrödinger bridges in modern generative sampling

Recent machine-learning literature has repurposed bridge ideas for high-dimensional transport and sampling from unnormalized distributions. In UniDB, diffusion bridges are formulated as a stochastic optimal control problem in which a controlled diffusion minimizes

$(X_t)$9

subject to a linear SDE; classic diffusion-bridge constructions such as Doob’s $X_T=x_T$0-transform are recovered only as a limiting case when $X_T=x_T$1 (Pan et al., 23 May 2025). UniDB++ addresses the slow, error-prone Euler sampling of the original reverse-time SDE by deriving exact closed-form reverse-time updates, replacing noise prediction with a more stable data-prediction model, and adding an SDE-Corrector for the $X_T=x_T$2–$X_T=x_T$3-step regime. The paper reports about $X_T=x_T$4 speedup in standard comparisons and up to $X_T=x_T$5 acceleration relative to UniDB, and gives concrete image-restoration numbers such as 28.40 dB PSNR and 0.8045 SSIM at 5 NFE on DIV2K super-resolution, compared with UniDB’s 25.46 dB and 0.6856 at 100 NFE (Pan et al., 23 May 2025).

The Schrödinger-bridge line of work reframes sampling as kinetic-optimal transport. The Adjoint Schrödinger Bridge Sampler considers target distributions $X_T=x_T$6 known only through an unnormalized energy and learns the Schrödinger bridge between a source $X_T=x_T$7 and $X_T=x_T$8 without target samples (Liu et al., 27 Jun 2025). The controlled process minimizes

$X_T=x_T$9

and the paper converts this into alternating Adjoint Matching and Corrector Matching steps, interpreted as an Iterative Proportional Fitting procedure on half-bridges (Liu et al., 27 Jun 2025). A central point is that ASBS relaxes the memoryless Dirac-source restriction of earlier Adjoint Sampling, allowing arbitrary priors and non-memoryless base processes. The convergence guarantee is conditional: ASBS converges to the Schrödinger bridge solution provided all matching stages achieve their critical points (Liu et al., 27 Jun 2025).

Bridge Matching Sampler pushes this line further by recasting bridge learning as a fixed-point iteration rooted in Nelson’s relation (Blessing et al., 28 Feb 2026). For a reference Brownian bridge, the path-space factorization $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$0 and the Markovian projection

$$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$1

lead to the least-squares update

$$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$2

The damped version

$$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$3

adds a proximal-style regularization term to stabilize training and mitigate mode collapse (Blessing et al., 28 Feb 2026). The paper reports tests up to dimension $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$4 and emphasizes that BMS uses a single scalable least-squares objective, arbitrary priors, and an unnormalized target-density oracle rather than target samples (Blessing et al., 28 Feb 2026).

Training objectives themselves have become contentious. The loss-analysis paper argues that for diffusion bridge samplers the Log Variance loss is not equivalent to reverse KL once forward dynamics, shared parameters, or diffusion coefficients are learned, and that LV lacks the usual data-processing-inequality motivation available for true $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$5-divergences (Sanokowski et al., 12 Jun 2025). The proposed alternative is reverse KL with the log-derivative trick, rKL-LD, which the paper reports to be more stable, to require less hyperparameter tuning, and to outperform LV on both Bayesian and synthetic benchmarks, especially when diffusion coefficients are learned (Sanokowski et al., 12 Jun 2025).

High-dimensional data-driven Schrödinger bridges can also be localized. The Localized Schrödinger Bridge Sampler replaces one global bridge in $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$6 by $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$7 low-dimensional bridge problems based on conditional independence, with local kernels

$$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$8

local denoisers $$p(\mathbf y\mid \mathcal M)=\int_\Theta p(\mathbf y\mid \theta,\mathcal M)\,p(\theta\mid\mathcal M)\,d\theta,$$9, and a split-step sampler for each coordinate (Gottwald et al., 2024). The paper states that the localized sampler is stable and geometrically ergodic, extends naturally to conditional sampling and Bayesian inference, and has a direct connection to multi-head self attention transformer architectures (Gottwald et al., 2024).

6. Rare events, discrete-state bridges, and domain-specific applications

Bridge constructions are also used to sample rare trajectories and low-probability events. One exact path-space method constructs an associated backward bridge process with transition probabilities

$\nu(x)\propto e^{-E(x)}$0

and reweights bridge trajectories via

$\nu(x)\propto e^{-E(x)}$1

With $\nu(x)\propto e^{-E(x)}$2, one obtains true endpoint-constrained bridges and exact reweighting by $\nu(x)\propto e^{-E(x)}$3 (Aguilar et al., 2021). This supports importance sampling for rare path ensembles, including first-passage-time distributions, and the paper reports that bridge trajectories collapse onto the WKB optimal path as the noise level is reduced, giving a practical way to judge the accuracy of WKB approximations at finite noise (Aguilar et al., 2021).

For discrete-state birth–death processes, the integer grid bridge sampler identifies a one-to-one correspondence between bridge paths from $\nu(x)\propto e^{-E(x)}$4 to $\nu(x)\propto e^{-E(x)}$5 over time $\nu(x)\propto e^{-E(x)}$6 and the product of a temporal simplex and a finite integer-grid bridge set (Sun et al., 2022). Partitioning by the number of upward jumps $\nu(x)\propto e^{-E(x)}$7 yields a uniform sampler over restricted bridge path space and the estimator

$\nu(x)\propto e^{-E(x)}$8

The paper emphasizes that even the near zero probability of rare event can now be evaluated with controlled relative error, and that the “haunting filtering failure” of a bootstrap particle filter finds no position in the new scheme (Sun et al., 2022).

Bridge sampling also appears inside broader rare-event pipelines. Neural Bridge Sampling for safety-critical autonomous systems introduces intermediate distributions

$\nu(x)\propto e^{-E(x)}$9

decomposes the failure probability into a product of normalizing-constant ratios, and estimates each ratio with a geometric bridge

$q$00

(Sinha et al., 2020). The method then learns invertible neural warps so that adjacent bridge distributions are closer to Gaussian, improving the Bhattacharyya coefficient and reducing bridge-sampling variance. The paper states that the efficiency gain over naive Monte Carlo scales like

$q$01

for very small failure probabilities, and demonstrates the approach on MountainCar, rocket-design problems, and CarRacing (Sinha et al., 2020).

A more specialized conditional object is the time-integrated stochastic bridge

$q$02

The finance-oriented paper uses Stochastic Collocation Monte Carlo and the Seven-League scheme to approximate the conditional distribution through a polynomial interpolant built from collocation points, while an ANN learns the collocation tensor from model parameters (Perotti et al., 2021). The expensive work is off-line; the on-line stage outputs collocation points and then samples

$q$03

very rapidly. The paper reports thousands or even millions of bridge samples in milliseconds to seconds depending on the model, with applications to Heston and SABR simulation and validations on ABM, GBM, and CIR (Perotti et al., 2021).

Across these domains, the recurring limitations are also consistent. Classical density-ratio bridge sampling depends on overlap and can become unstable with poor proposals or difficult posterior geometry (Micaletto et al., 20 Aug 2025, Wang et al., 2016). Endpoint-conditioned diffusion bridges require scores or transition densities that are often unavailable and may need learned approximations (Grong et al., 2024, Boserup et al., 2024). Learned diffusion-bridge samplers introduce further issues of solver stability, objective choice, damping, and low-step artifacts (Pan et al., 23 May 2025, Sanokowski et al., 12 Jun 2025, Blessing et al., 28 Feb 2026). The unifying pattern is that bridge sampling is effective precisely when the bridge object is chosen so that overlap, conditioning, or transport becomes substantially easier than in the original formulation.