Diffusion Bridge Variational Inference

• DBVI is a variational inference technique that leverages diffusion processes with bridge structures to approximate complex posterior distributions in deep Gaussian processes.
• It utilizes forward and reverse-time stochastic differential equations with Doob’s h-transform to conditionally correct drifts, enabling data-dependent inference.
• Empirical results show that DBVI improves performance on tasks such as regression, classification, and image reconstruction by optimizing an ELBO over path-space approximations.

Diffusion Bridge Variational Inference (DBVI) denotes variational inference constructions in which the variational approximation is defined through a diffusion process with bridge structure, or through a bridge-like path measure whose reverse-time dynamics are learned. In the strictest named usage, DBVI was proposed for inducing-variable inference in deep Gaussian processes (DGPs) as an extension of denoising diffusion variational inference (DDVI): instead of starting the reverse diffusion from a fixed unconditional Gaussian, it starts from a learnable data-dependent Gaussian and reinterprets the resulting reference process through a Doob-bridged diffusion, while retaining a Girsanov-based ELBO and reverse-time SDE machinery (Xu et al., 23 Sep 2025). More broadly, adjacent literature studies augmented path-space ELBOs, conditioned diffusion bridges, and variational Schrödinger bridge relaxations that supply many of the same structural ingredients—forward reference dynamics, reverse-time corrections, tractable bridge marginals, and learned score or drift terms (Geffner et al., 2022).

1. Conceptual scope and defining features

A central feature of DBVI is that the variational family is not specified only by a static density $q(z)$, but by a stochastic process or an induced terminal law obtained from such a process. In the DGP formulation, the posterior over inducing variables is approximated by the terminal marginal of a reverse-time diffusion whose start distribution is data-dependent and whose reference forward dynamics are Doob-bridged (Xu et al., 23 Sep 2025). This contrasts with simpler Gaussian variational families, and also with MCMC approaches such as SGHMC, because the approximation is learned as a transport mechanism over latent variables rather than as a fixed parametric density.

The literature surrounding DBVI uses “bridge” with different degrees of strictness. “Langevin Diffusion Variational Inference” constructs a stochastic interpolation from $q$ to $p$ by simulating underdamped Langevin dynamics and an approximate reverse-time diffusion, but explicitly notes that it “is not a diffusion bridge in the strict sense of conditioning a diffusion on fixed endpoints” (Geffner et al., 2022). By contrast, “Black-box Variational Inference for Stochastic Differential Equations” builds a variational path family for latent SDEs whose Gaussian transitions “bridge between observations as the conditioned diffusion process does,” making it much closer to classical endpoint- or observation-conditioned bridge inference (Ryder et al., 2018).

This terminological spread matters. In strict bridge formulations, the reference object is a conditioned diffusion law or a Doob $h$-transform. In broader DBVI-style usage, the essential point is that inference is performed over diffusion-induced paths or terminal laws, with the bridge structure providing either exact endpoint conditioning or a bridge-like transport from a tractable distribution toward a target posterior.

2. Path-space variational structure

One recurring foundation for DBVI-style methods is the replacement of a static variational density by an augmented path-space approximation. In “Langevin Diffusion Variational Inference,” forward kernels $F_k$ and backward kernels $B_k$ define augmented laws over discretized trajectories, and the resulting objective is the augmented ELBO

$$\mathrm{ELBO} = \mathbb E_q\!\left[ \log \frac{p(z_K,\rho_K)}{q(z_1,\rho_1)} + \sum_{k=1}^{K-1} \log \frac{B_k(z_k,\rho_k\mid z_{k+1},\rho_{k+1})}{F_k(z_{k+1},\rho_{k+1}\mid z_k,\rho_k)} \right].$$

This makes the variational problem one of choosing a forward stochastic dynamics and a compatible reverse or backward path law with tractable density ratios (Geffner et al., 2022).

In latent SDE inference, the same shift appears in a different guise. “Black-box Variational Inference for Stochastic Differential Equations” writes

$$q(\theta,x;\phi)=q(\theta;\phi_\theta)\,q(x\mid\theta;\phi_x),$$

where $q(x\mid\theta;\phi_x)$ is an autoregressive Gaussian path distribution produced by a recurrent network, and optimizes the joint ELBO

$$\mathcal L(\phi) = \mathbb E_{q(\theta)q(x\mid\theta)} \!\left[ \log p(\theta)+\log p(x\mid\theta)+\log p(y\mid x,\theta)-\log q(\theta)-\log q(x\mid\theta) \right].$$

Here the variational object is explicitly a latent diffusion path conditional on parameters and observations, not merely a smoothing marginal at a single time (Ryder et al., 2018).

In the DGP-specific DBVI formulation, the surrounding variational target remains the standard DGP ELBO,

$q$0

but $q$1 is no longer a simple Gaussian family. Instead it is induced by a reverse-time diffusion over inducing variables, with the bridge construction entering through the reference process and the learned conditional score (Xu et al., 22 May 2026). This is a defining DBVI move: preserve the variational objective at the model level, but replace the static posterior family by a diffusion-bridge sampler whose terminal law acts as the variational approximation.

3. Bridge dynamics, time reversal, and learned scores

Classical diffusion bridge theory enters through Doob’s $q$2-transform. For a diffusion

$q$3

the bridge conditioned on $q$4 and $q$5 satisfies

$q$6

with $q$7 and $q$8. This formula identifies bridge inference as drift correction by a transition-density score (Heng et al., 2021).

The explicit DBVI construction for DGPs adopts this bridge viewpoint at the reference-process level. Its learnable initial law is

$q$9

and the forward bridge dynamics are written as

$p$0

where the Doob correction $p$1 encodes conditioning on the learned initialization (Xu et al., 23 Sep 2025). The practical reverse-time bridge SDE then uses a conditional score

$p$2

so the learned reverse dynamics are corrected both by the neural score model and by the analytic bridge term (Xu et al., 23 Sep 2025).

A later DGP study makes this characterization more explicit by describing DBVI as a score-based ELBO method whose forward Doob-bridge has closed-form Gaussian marginals

$p$3

and whose conditional score network is trained against those bridge marginals by conditional denoising score matching (Xu et al., 22 May 2026). In this reading, DBVI occupies a middle ground: it uses exact bridge structure where available, but still relies on learned scores to close the reverse-time variational approximation.

4. Relation to Schrödinger bridges and adjacent variational bridge methods

DBVI belongs to a broader family of path-space variational bridge methods tied to Schrödinger bridges, controlled diffusions, and time-reversal identities. In “Variational Schrödinger Diffusion Models,” the dynamic Schrödinger bridge problem is written as

$p$4

and the intractable forward bridge score is replaced by a linear variational surrogate $p$5, restricting the forward path law to a multivariate Ornstein–Uhlenbeck family with closed-form conditionals (Deng et al., 2024). “Variational Schrödinger Momentum Diffusion” makes the analogous move in underdamped phase space, using a linearized momentum-bridge family and a critical-damping transform to reduce the number of score terms that must be estimated (Rojas et al., 28 Jan 2025).

A complementary conceptual formulation appears in “Schödinger Bridge Type Diffusion Models as an Extension of Variational Autoencoders,” which casts SB-type diffusion training as minimization of a path-space KL,

$p$6

and decomposes that objective into a prior loss $p$7 plus a drift-matching term between the reverse encoder drift and the decoder drift (Kaba et al., 2024). This gives a VAE-style interpretation of bridge learning on path space and clarifies why bridge-based diffusion methods naturally resemble variational inference.

For Bayesian computation, “Diffusion Schrödinger Bridges for Bayesian Computation” frames posterior or target sampling as a path-space KL projection against a reference diffusion and solves the resulting bridge by iterative proportional fitting. The output is primarily a bridge-based sampler rather than an ELBO-trained posterior density, but the formal object is again a controlled path measure constrained at the endpoints (Heng et al., 2023). This suggests that DBVI is one point in a wider continuum: some methods emphasize posterior approximation and ELBOs, others emphasize exact bridge constraints or transport efficiency, and many trade off simulation-free training against fidelity to the exact bridge.

5. Terminological boundaries and neighboring methods

Not every diffusion bridge method is a DBVI method, and not every diffusion-based variational method uses a strict bridge. The neighboring literature is best read by distinguishing the variational object from the bridge mechanics.

Work	Variational object	Relation to DBVI
“Langevin Diffusion Variational Inference” (Geffner et al., 2022)	Augmented path-space ELBO over forward and backward kernels	DBVI-adjacent; reverse-time diffusion VI, but not a strict endpoint-conditioned bridge
“VIPaint” (Agarwal et al., 2024)	Hierarchical variational posterior over selected diffusion states	DBVI-like observation-conditioned diffusion inference; coarse and hybrid rather than full-time bridge VI
“Denoising Diffusion Bridge Models” (Zhou et al., 2023)	Conditional bridge score field learned from paired endpoints	Bridge modeling, not variational inference
“Neural Guided Diffusion Bridges” (Yang et al., 17 Feb 2025)	Reverse-KL variational family over guided bridge SDE path measures	Very close to DBVI in path-space form, but focused on conditioned diffusion simulation rather than ELBO posterior inference
“UniDB++” (Pan et al., 23 May 2025)	None; training-free fast sampler for SOC bridge models	Bridge sampler, not variational inference

A common misconception is that endpoint conditioning alone is sufficient to make a method DBVI. The bridge literature shows otherwise. “Denoising Diffusion Bridge Models” learns endpoint-conditioned bridge scores from paired data and samples a reverse bridge SDE, but its objective is denoising score matching on bridge marginals rather than variational inference (Zhou et al., 2023). Conversely, “VIPaint” is explicitly variational and approximates an observation-conditioned posterior over diffusion states, but it parameterizes only a subset of times and then reverts below $p$8 to DPS-like refinement, so it is best viewed as a coarse hierarchical posterior approximation rather than a full diffusion-bridge VI method (Agarwal et al., 2024). “Neural Guided Diffusion Bridges” is perhaps the closest non-DGP relative: it introduces a neural residual drift on top of a guided proposal and minimizes a reverse KL on path space via Girsanov, which is structurally very close to DBVI, but the target is a conditioned diffusion bridge law rather than an ELBO-defined Bayesian posterior over inducing variables (Yang et al., 17 Feb 2025).

6. DGP instantiation, empirical profile, and later reinterpretation

In its explicit DGP form, DBVI targets inference over inducing variables $p$9 in sparse deep Gaussian processes, where the joint model factorizes layerwise and the inducing posterior is “high-dimensional, non-Gaussian, and entangled across layers through the deep composition” (Xu et al., 23 Sep 2025). The central change from DDVI is the start distribution

$h$0

with $h$1 produced by an amortized network. To make this scalable, the amortizer acts on the inducing inputs $h$2, which the paper treats as structured, low-dimensional summaries aligned with the shape of the inducing variables (Xu et al., 23 Sep 2025).

Training remains an SVI-style Monte Carlo procedure. The bridge marginal coefficients are precomputed, $h$3 is sampled from the learned initial Gaussian, the reverse bridge SDE is discretized by Euler–Maruyama, the score-matching integral is accumulated numerically, and the sampled endpoint $h$4 is propagated through the sparse DGP to evaluate the minibatch likelihood. The learned components are the score network, the amortized initializer, and the DGP hyperparameters; the paper emphasizes that ELBO gradients adapt the initial distribution through the boundary terms and through its effect on the simulated reverse path (Xu et al., 23 Sep 2025).

The reported empirical profile is consistent across task families.

Task family	Reported DBVI outcome	Comparison reported
UCI regression	Lower RMSE and NLL than DSVI, IPVI, SGHMC, and DDVI; largest gains on YearMSD and Airline	Faster early RMSE reduction than DDVI on Energy
Image classification	MNIST $h$5, Fashion-MNIST $h$6, CIFAR-10 $h$7	Modest but consistent gains over DDVI, with small runtime increase
Large-scale classification	SUSY AUC $h$8 and HIGGS AUC $h$9 across depths $F_k$0–$F_k$1	Beats DDVI across all reported depths
Frey Faces reconstruction	RMSE $F_k$2, NLL $F_k$3	Improves on DDVI’s RMSE $F_k$4, NLL $F_k$5

The same paper also states the main caveats: DBVI is more complex than standard DSVI because it requires reverse SDE simulation, bridge correction, score networks, and precomputation of bridge-marginal ODEs; the initial covariance remains fixed and isotropic; tractable bridge formulas rely on a linear/isotropic reference setup; and the method inherits diffusion-model sensitivities to discretization, schedule choice, and score-network optimization (Xu et al., 23 Sep 2025).

A later DGP paper, “Onsager-Machlup Posterior Transport for Deep Gaussian Processes,” treats DBVI as its immediate stochastic predecessor and thereby sharpens its empirical profile. Under matched-seed paired Wilcoxon tests on seven UCI benchmarks, OM-Path records statistically significant wins over DBVI on the two largest datasets, power and protein; statistical ties on yacht and qsar; and losses to DBVI on boston, energy, and concrete. That paper interprets the pattern as a deterministic-path versus stochastic-posterior tradeoff: DBVI remains stronger on small-$F_k$6, noisy data where SDE noise regularization helps, while deterministic path objectives can dominate when objective variance becomes the main bottleneck (Xu et al., 22 May 2026).

Taken together, these results situate DBVI as a bridge-based, score-parameterized variational method that is materially more expressive than Gaussian DGP variational inference and more posterior-aware than unconditional diffusion starts, while remaining subject to the computational and numerical burdens typical of reverse-time diffusion inference.