Latent-Space Bayesian Filtering

• Bayesian filtering in latent space is a method that recasts high-dimensional, nonlinear state-space models into lower-dimensional latent representations for efficient filtering.
• It leverages learned or prescribed latent constructions, such as deep generative models and factor copulas, to yield analytical and robust filtering recursions similar to Kalman updates.
• Empirical evaluations demonstrate that latent-space filtering outperforms classical methods in tasks like Moving MNIST and Lorenz96 by reducing RMSE and enhancing calibration.

Bayesian filtering in latent space refers to the execution of sequential state estimation, prediction, or signal smoothing by recasting the original high-dimensional, nonlinear, or non-Gaussian state-space model into a lower-dimensional (or otherwise more tractable) latent representation. The goal is to exploit the statistical, spectral, or structural properties of the latent space—often learned or constructed via neural or probabilistic model families—so that the filtering recursion becomes more efficient, more robust to non-Gaussianity, and sometimes even analytically tractable. Below, the key foundational mechanisms, representative methodologies, and consequences are surveyed.

1. Fundamentals of Latent-Space Bayesian Filtering

Bayesian filtering is the process of sequentially updating the posterior distribution of hidden states $x_t$ in a dynamical model given observations $y_{1:t}$. Standard approaches (Kalman, particle, ensemble, or variational filters) work in the ambient state space. However, for high-dimensional, nonlinear, or non-Gaussian systems, this is computationally difficult or even infeasible.

Bayesian filtering in latent space reparametrizes the original model, or constructs an auxiliary latent variable model. This latent space (typically $\mathbb{R}^d$, $d\ll\text{dim}(x_t)$) is chosen so that the filtering recursion—prediction and correction—attains favorable analytic, algorithmic, or statistical properties. Notably, latent construction may be learned (deep generative models, flow-based mappings), prescribed (factor models, spectral separations), or induced by problem structure (manifolds, factor copulas).

The archetype is the Deep Bayesian Filter (DBF) (Tarumi et al., 2024), which introduces a latent $h_t$ with linear-Gaussian latent-space transitions and a neural "inverse observation operator" trained to ensure the latent-space filter is analytically Gaussian-recursive.

2. Mathematical Frameworks and Analytical Structures

Multiple recent approaches implement Bayesian filtering in latent space; their main distinctions are in latent construction, transition/observation structure, and the analytical form of filtering recursion.

Deep Bayesian Filter (DBF) (Tarumi et al., 2024):

• Latent state $h_t \in \mathbb{R}^d$ with linear-Gaussian dynamics,

$$p(h_t \mid h_{t-1}) = \mathcal{N}\big(h_t \,\vert\, Ah_{t-1}, Q\big)$$

• Neural inverse observation operator (IOO) $$r(h_t|o_t)\approx \mathcal{N}(h_t|f_\theta(o_t), G_\theta(o_t))$$
• Filtering recursion is closed-form Gaussian due to Gaussian prior and IOO. The correction step is

$$\tilde{p}(h_t|o_{1:t}) \propto q_\theta(h_t|o_t)\tilde{p}(h_t|o_{1:t-1})$$

yielding a fully closed-form Kalman-like update for mean and covariance.

Gaussian Process Low-Pass Filtering (Valenzuela et al., 2019):

• The time series is decomposed into low- and high-frequency latent GPs.
• Posterior over the low-frequency latent directly corresponds to an optimal Wiener filter but with computable uncertainty, adaptable to sampling irregularity and noise.

Score-Based and Flow-Based Nonparametric Filtering (Christensen et al., 27 Oct 2025, Wang et al., 22 Feb 2025):

• High-dimensional latent signals $X$ are recovered via manifold learning and score-based diffusions, such that the Bayesian posterior in the high-dimensional space contracts to the projection onto the low-dimensional manifold.
• Flow-based Bayesian filter (FBF) learns invertible maps ($f$, $g$) to latent space, defining Gaussian state-space models there. Efficient filtering and posterior sampling are performed using standard Kalman formulas in latent space, with densities pushed back via the learned flows (Wang et al., 22 Feb 2025).

Latent Ensemble Score Filtering (Latent-EnSF) (Si et al., 2024):

• State and (possibly sparse) observation are encoded into a common latent Gaussian manifold (via a coupled variational autoencoder).
• Score-based filtering (e.g., via SDEs in latent space) updates state posteriors even in unobserved coordinates, circumventing traditional degeneracies in sparse observation settings.

Hybrid Physics-Neural Latent Filtering (Imbiriba et al., 2022):

• The latent state evolves jointly by known physics-based dynamics and a learnable neural correction, all formulated and estimated in latent space using Gaussian assumed-density filtering and cubature integration.

3. Algorithmic Recursions and Training Paradigms

Across these methods, latent-space filtering consists of recursively alternating prediction (time update in latent dynamics) and correction (conditioning on new, possibly nonlinearly observed data mapped to latent space):

Framework	Latent Transition	Observation Update	Posterior Approximation
DBF (Tarumi et al., 2024)	Linear-Gaussian	Neural Gaussian IOO	Closed-form Gaussian
FBF (Wang et al., 22 Feb 2025)	Learned linear	Linear in latent obs	Kalman (latent), Flow (orig)
Latent-EnSF (Si et al., 2024)	VAE-encoded	Score-based SDE	Latent EnSF ensemble
Score-based (Christensen et al., 27 Oct 2025)	N/A (manifold)	Learned reverse SDE	Sampling via SDE
Bayesian Copula (Lavine et al., 2020)	DGLMs with latent	Copula for joint marginals	Gaussian copula

Training regimes typically combine maximizing a filtering-recursion variational lower bound (evidence lower bound, ELBO), conditional likelihood, and explicit learning of latent-to-observation or latent-to-state maps, often by neural parameterizations.

The DBF trains by maximizing an ELBO that includes expectations over the (analytically available) latent-space posteriors and KL divergence between filtering updates (Tarumi et al., 2024). Similarly, flow-based and VAE-based approaches optimize a joint likelihood (including change-of-variables determinants) over observed trajectories (Wang et al., 22 Feb 2025, Si et al., 2024).

4. Empirical Performance and Comparative Advantages

Empirical results across settings distinguish latent-space Bayesian filtering by accuracy and computational efficiency, especially in high-dimensional, nonlinear, or non-Gaussian scenarios.

• In the Moving MNIST and Double Pendulum tasks, DBF recovers unobserved states and uncertainties more accurately and robustly than EnKF, ETKF, PF, and neural SSMs, particularly under non-Gaussian posteriors.
• On the Lorenz96 high-dimensional PDE benchmark, under large noise and saturating nonlinear observations, DBF and Latent-EnSF outperform classical ensemble methods and standard score filters by an order of magnitude in RMSE, and retain better calibration (Tarumi et al., 2024, Si et al., 2024).
• Flow-based Bayesian filtering attains higher efficiency and accuracy than sequential Monte Carlo and variational SSMs; both density evaluation and sampling are tractable via latent-to-physical transformations (Wang et al., 22 Feb 2025).
• In nonparametric high-dimensional settings with only sample access to the signal manifold, score-based diffusion filtering provides theoretically Bayes-optimal estimation as dimensionality increases (Christensen et al., 27 Oct 2025).
• Latent-space approaches remain robust to sparse, irregular, or noisy observations, for instance, GP-based low-pass filtering exactly reproduces Wiener filtering but with principled uncertainty quantification and native support for unequally sampled series (Valenzuela et al., 2019).

5. Theoretical Properties and Guarantees

Several methods furnish rigorous analysis of the stability, convergence, and optimality of latent-space filters.

• DBF and related approaches guarantee no accumulation of Monte Carlo approximation error over time—the latent-space recursion is fully analytic and stable under Gaussian assumptions (Tarumi et al., 2024).
• In the context of in-context learning for LLMs, Bayesian latent filtering (via Kalman recursion in an adaptation subspace) guarantees stability, exponential contraction of the epistemic covariance, and mean-square error bounds under standard observability conditions (Kiruluta, 2 Jan 2026).
• For model-free, high-dimensional score-based filtering, posterior concentration on the manifold is proven at an exponential rate in dimensionality, guaranteeing Bayes-optimal filtering in the high-dimensional limit (Christensen et al., 27 Oct 2025).

6. Extensions, Practical Limitations, and Research Directions

Open questions center on the latent construction and tractability-accuracy tradeoff:

• The effectiveness of Bayesian filtering in latent space depends critically on the fidelity of latent encoders, quality of observation decoders, and capacity to approximate nonlinear, non-Gaussian statistics. If the latent mapping or decoder is insufficient, filter performance degrades (Si et al., 2024).
• Several methods, such as Latent-EnSF or flow-based filters, rely on amortized inference or neural parameterization—training stability and generalization remain ongoing challenges.
• For physics-augmented filters, there is a tradeoff between model interpretability (retained via explicit physics-based dynamics) and adaptability/expressiveness granted by neural augmentations, managed by constraint filtering (Imbiriba et al., 2022).
• Future extensions are anticipated to further integrate score-based and diffusion models with robust uncertainty quantification, directly optimize filtering objectives, and enable scalable filtering in extremely high-dimensional, multimodal latent spaces.

7. Connections with Related Methodologies

Bayesian filtering in latent space unifies ideas from state-space modeling, modern representation learning, score-based generative modeling, and structured probabilistic inference:

• Latent-space recasting enables incorporation of powerful learned representations (e.g., VAEs, flows) without sacrificing the basic algorithmic principles of probabilistic filtering.
• Integration with score/diffusion approaches bridges the gap to nonparametric, model-free filtering pipelines now prominent in generative modeling (Christensen et al., 27 Oct 2025, Si et al., 2024).
• These advances facilitate efficient, scalable, and theoretically sound Bayesian sequential estimation for problems at the intersection of physics, statistics, and machine learning.

---

References:

• Deep Bayesian Filter (Tarumi et al., 2024)
• Low-pass filtering as Bayesian inference (Valenzuela et al., 2019)
• Model-free filtering with score-based diffusions (Christensen et al., 27 Oct 2025)
• Latent-EnSF (Si et al., 2024)
• Hybrid neural-physics latent filtering (Imbiriba et al., 2022)
• Flow-based Bayesian filtering (Wang et al., 22 Feb 2025)
• Bayesian latent filtering in LLMs (Kiruluta, 2 Jan 2026)
• EnLLVM (Lin et al., 2017)
• Neural Bayesian Filtering (Solinas et al., 4 Oct 2025)
• Copula-based latent factor filtering (Lavine et al., 2020)