Relational State-Space Modeling

• Relational state-space modeling is a framework that incorporates explicit relational structure into traditional state-space models by leveraging graphical models, GNNs, and nonparametric Bayes.
• It employs advanced techniques such as sequential message passing, nonparametric Bayesian inference, and encoder–decoder architectures to capture dynamic interactions in complex systems.
• The paradigm unifies temporal dynamics with relational dependencies, offering scalability, uncertainty quantification, and rigorous theoretical guarantees across diverse application domains.

Relational state-space modeling is a paradigm within temporal modeling and reasoning that explicitly incorporates relational structure among entities into the state-space formalism. This approach unifies the representation of temporal dynamics with the statistical dependencies or semantic relations present among system components, leveraging tools from graphical models, graph neural networks (GNNs), nonparametric Bayes, dynamic logic, and neural architectures. Relational state-space models have demonstrated efficacy in diverse application domains ranging from network time-series, multi-agent reinforcement learning, stochastic multi-object systems, spatio-temporal forecasting, to vision tasks demanding relational feature integration.

1. Mathematical Foundations of Relational State-Space Models

The mathematical core of relational state-space modeling augments the classical notion of a state-space model with an explicit relational bias. In graphical models, this is formalized by introducing latent variables that represent the evolving state of each entity, as well as the relations (edges) between entities, which may themselves be time-varying and learned from data.

A typical probabilistic relational state-space model is defined by a sequence of latent states $\{h_t\}_{t=0}^{T}$, each representing a graph or a set of entity-wise latent vectors, with state transitions and observation processes conditioned on both historical states and the graph structure. This structure can be given (based on domain knowledge) or learned (via neural architectures or probabilistic procedures). The joint distribution is often expressed as

$$p(h_{0:T}, y_{1:T} \mid x_{1:T}) = p(h_0) \prod_{t=1}^T p(h_t|h_{t-1},x_t,A_t)p(y_t|h_t)$$

where $A_t$ encodes the edge set or adjacency of the latent relational graph at time $t$ (Zambon et al., 2023). In hierarchical latent variable models, both entity-level $\mathbf{z}_t^{(i)}$ and global latent $\mathbf{z}_t^g$ are included, with message passing or structured flows modeling their joint evolution (Yang et al., 2020).

Relational dependencies can be implemented as:

• Predefined relational graphs (e.g., fixed object interaction graphs)
• Learned, time-varying graph structures (e.g., Bernoulli-sampled adjacency matrices with neural parameterization)
• Relational abstractions via first-order logic (e.g., in relational MDPs and dynamic probabilistic logic models) (Kokel et al., 2021)
• Relation-driven neural architectures (e.g., spatial GNNs for agent and object interactions) (Utke et al., 2024).

2. State Evolution and Inference Mechanisms

Relational state-space models specify state evolution using either continuous-time processes, sequential message-passing on graphs, or neural SSM variants. Examples include:

• Nonparametric Bayesian State Evolution: Latent coordinates for each entity evolve in continuous time as Gaussian processes, with nonparametric shrinkage inducing automatic dimensionality selection. The observation model is typically Bernoulli-logistic, mapping the latent similarities to observed network edges. Inference is performed by Pólya–Gamma augmentation and Gibbs sampling, allowing exact conjugate updates and efficient exploration of the (potentially infinite) latent space (Durante et al., 2013).
• Sequential Hierarchical Latent Models: High-dimensional latent states for each object are propagated via GNN-based message passing, possibly augmented by global latent vectors to capture system-wide context. The generative process specifies both local and global transitions, with emissions modeled per-entity and inference performed using variational sequential Monte Carlo complemented by contrastive objectives to stabilize long-range interaction learning (Yang et al., 2020).
• Encoder–Decoder Graph SSMs: For spatio-temporal data, latent state graphs are learned end-to-end, with encoder modules selecting, reducing, and transforming input graphs to latent graphs, inferring the necessary adjacency by neural mechanisms. Decoder modules read out predictions by pooling or upscaling, maintaining differentiability via reparameterization and score-function estimators (Zambon et al., 2023).
• Relational GNN Critics: In MARL, the state is mapped into a relation-typed, directed graph encoding spatial positional predicates (left, right, adjacent, etc.). A multi-layer relational GCN computes node embeddings, which are then pooled to a permutation- and translation-invariant state summary used for value estimation or policy evaluation (Utke et al., 2024).

3. Modeling Relational Structure: Graphs, Logic, and Abstraction

Relational structure is represented across several frameworks:

• Graph-Structured SSMs: Entities are vertices, and interactions are encoded as (possibly typed, directed, time-varying) edges. Message passing or MPNN architectures model information exchange, where the adjacency is either fixed or sampled at each step, capturing uncertainty and sparsity (Zambon et al., 2023, Yang et al., 2020).
• Relational Logic via DPLM: In symbolic domains, dynamic probabilistic logic models (DPLMs) describe transition structure with dynamic first-order conditional influence (D-FOCI) statements, specifying how the truth of each child predicate at $t+1$ depends only on a finite, context-determined parent set at $t$ and the action taken (Kokel et al., 2021). Backward-influence procedures recover the minimal set of relevant literals per subtask, yielding bisimulation-based abstractions with optimality guarantees.
• Relational Abstractions in RL: In multi-agent RL, relational state abstraction transforms raw spatial and attribute state into a graph over entities, using GNN encoders to enable more efficient and invariant value estimation, as in the Multi-Agent Relational Critic (MARC) architecture (Utke et al., 2024).

4. Neural and Algorithmic Architectures

Neural implementations of relational SSMs leverage diverse architectural strategies:

• Relation-Driven State-Space Models (RD-UIE): Modifies standard SSMs (e.g., Mamba) by dynamically reordering sequence scans in vision models based on spatial sampling frequency, prioritizing structurally salient tokens for state updates. The Visually Self-adaptive State Block combines this relation-driven scan with dynamic convolution conditioned on global context, and a Cross-Feature Bridge fuses multi-scale relational features. This architecture outperforms prior Mamba variants by 0.55 dB average PSNR across benchmarks (Jiang et al., 2 May 2025).
• Graph Normalizing Flows: To capture complex joint distributions over entity states, graph-based normalizing flows generalize coupling layers and invertible linear mixing to the node-indexed setting, allowing flexible, invertible joint latent modeling (Yang et al., 2020).
• Relational Graph Convolutional Networks: Employed for permutation and translation invariance in MARL critics, these layers compute node updates as a sum of relation-typed and self-connection weighted combinations, enabling sample-efficient and generalizable value function approximation (Utke et al., 2024).

5. Theoretical Guarantees and Empirical Results

Relational SSMs offer strong theoretical properties:

• Universal Approximation: Nonparametric Bayes frameworks with multiplicative gamma shrinkage and Gaussian process priors admit full large support in uniform topology on dynamic network processes, ensuring flexible approximation of any smooth relational time series (Durante et al., 2013).
• Value-Preserving Abstractions: For reinforcement learning, abstraction functions derived from DPLMs are proven to preserve the optimal value function under specified bisimulation conditions (Kokel et al., 2021).
• Uncertainty Quantification: End-to-end probabilistic architectures with latent random graphs and node features propagate and calibrate uncertainty at every level, from state inference to forecasted outputs (Zambon et al., 2023).

Empirically:

• Models such as eGSS and gRNN (graph SSMs) closely match analytic optimal error on synthetic spatio-temporal VAR processes (MAE near 0.32 vs. optimum 0.319), and residuals pass correlation tests, while non-relational RNNs perform much worse (Zambon et al., 2023).
• In hierarchical MARL, relational GNN-based abstraction (MARC) yields both 7-fold improvements in sample efficiency and large gains in asymptotic performance over non-relational baselines, retaining performance under entity/role variation and out-of-distribution changes (Utke et al., 2024).
• In vision (UIE), relation-driven SSMs outperform standard state-space architectures on PSNR and SSIM, with notably improved structural fidelity in enhanced images (Jiang et al., 2 May 2025).

6. Connections to Broader State-Space and Relational Modeling Paradigms

Relational SSMs generalize classical linear-Gaussian SSMs, hidden Markov models, and Kalman filters in several key ways:

• Observation Model Generalization: They replace Gaussian emissions with Bernoulli-logistic links, neural decoders, or domain-specific measurement models (Durante et al., 2013, Zambon et al., 2023).
• State Dynamics: Nonparametric, continuous-time evolutions (GPs), graph-based message passing, and probabilistically sampled adjacency structures supplant recurrence relations or fixed Markovian state updates. The latent dimensionality and relational topology are often inferred from data, rather than fixed a priori (Durante et al., 2013, Zambon et al., 2023).
• Integration with Logic and Planning: DPLMs bridge statistical relational learning and temporal SSMs, supporting abstraction and decomposability in RL (Kokel et al., 2021).

This suggests that relational state-space modeling functions as a unifying language across statistical time-series, symbolic RL, graph representation learning, and structured neural architectures, facilitating principled, scalable, and adaptive modeling of complex relational-temporal phenomena.