Hierarchical State Space (HiSS) Framework

• Hierarchical State Space (HiSS) is a modeling framework that organizes state representations and inference across multiple, explicitly tiered levels.
• It employs layered state-space models with mechanisms like temporal hierarchy and spatial abstraction to achieve robust, sample-efficient learning.
• HiSS architectures have demonstrated superior performance in areas such as robotics, fMRI analysis, and power systems, offering benefits in interpretability and safety.

A Hierarchical State Space (HiSS) is a compositional modeling framework in which the state representation, inference, and/or policy structure of a dynamical system or learning agent are organized across multiple, explicitly tiered levels. Each level operates at a distinct spatiotemporal or semantic scale, either compressing or abstracting information from finer-grained, subordinate levels. Originating independently in control theory, reinforcement learning, robotics, neuroscience, biomedicine, and power systems, HiSS architectures exploit this stratification to achieve tractable inference, sample-efficient learning, interpretability, and robustness across a broad spectrum of domains.

1. Core Formulation and Mathematical Foundation

The essential feature of HiSS frameworks is the recursive or stacked application of state-space models (SSMs) or their extensions, enabling each layer to process information at its characteristic timescale or abstraction level.

A two-level continuous HiSS for sequence modeling (Bhirangi et al., 2024) is defined as: $$\begin{aligned} \text{Low-level SSM:} \qquad & x^L_t = A^L x^L_{t-1} + B^L u_t \ & o_t = C^L x^L_t + D^L u_t \ \text{Chunking:} \qquad & c_i = o_{i \cdot k} \qquad (i = 1, 2, ..., \lfloor T/k \rfloor) \ \text{High-level SSM:} \qquad & x^H_i = A^H x^H_{i-1} + B^H c_i \ & y_i = C^H x^H_i + D^H c_i \end{aligned}$$ where the low-level operates at the fast input rate $u_t$ and the high-level at a coarser, chunked stride $k$.

In hierarchical representation learning for Markov decision processes (MDPs), the core is a state partitioning function $f_\psi: S \to \Delta(Z)$, mapping raw states to distributions over discrete abstract states. This induces abstract subtasks (options) and defines transitions and policies at each abstraction $z \in Z$ (Steccanella et al., 2021). The general principle extends to control-affine systems: $$\dot{x}_i = A_i(p_i) x_i + B_i(p_i) u_i + b_{x,i}(p_i)$$ where parameters depend locally on operating point $p_i$, and aggregation yields global system behavior (Liu et al., 25 Aug 2025).

Hierarchical null-space projection in robotics defines a sequence of subspaces, each constructed in the null-space of higher-priority task Jacobians, restricting exploration and policy adaptation to safe manifolds (Lundell et al., 2018).

2. Hierarchical Modeling Mechanisms

HiSS architectures implement hierarchical decomposition and modeling via several mechanisms:

• Temporal hierarchy: Lower levels capture rapid, local, or noisy dynamics; higher levels summarize global, long-range, or slowly varying dependencies (Bhirangi et al., 2024, Wu et al., 26 Jun 2025).
• Spatial/semantic abstraction: State compression or clustering into macro-states/modules, with subtasks linking these abstract regions (e.g., navigation “rooms” or anatomical phases) (Steccanella et al., 2021).
• Parallel or multi-branch encoding: Independent state-space modeling of complementary factors (e.g., spatial vs temporal in FST-Mamba for fMRI (Wei et al., 2024), or device-level vs grid-level in power systems (Liu et al., 25 Aug 2025)).
• Hierarchical vector quantization: Multi-stage codebooks discretize both state (macrostate) and state transitions, capturing metastability and transition kinetics as discrete hierarchies (Yang et al., 28 Jun 2025).

These mechanisms enable both bottom-up feature aggregation and (optionally) top-down feedback.

3. Optimization and Inference in HiSS

The typical inference pipeline for HiSS is strictly feedforward through the hierarchy, though recurrent or gating feedback is possible. Training objectives combine local reconstruction errors, cross-level consistency losses, and, when applicable, task-specific supervision signals.

• Sequence prediction and filtering: Mean squared error between predicted and target sequences, with all state-space parameters trained end-to-end via gradient descent; Adam or AdamW optimizers are standard (Bhirangi et al., 2024, Wei et al., 2024, Wu et al., 26 Jun 2025).
• Representation learning: In discrete abstraction HiSS, the partitioning function $f_\psi$ is optimized by a weighted combination of temporal-consistency, marginal entropy, and determinism losses, all amenable to stochastic gradient descent (Steccanella et al., 2021).
• Reinforcement learning and options: HiSS frameworks define hierarchical options parameterized by abstract state transitions. Option initiation, termination, and policy learning are modular; high-level master policies operate over the induced SMDP and are optimized via tabular or function-approximation methods (Steccanella et al., 2021).
• Constraint-enforced optimization: Hierarchical null-space methods recursively solve quadratic programs or least-squares problems, projecting lower-priority policies/passive tracking commands into the null spaces of higher-priority, safety-enforcing constraints (Lundell et al., 2018).

4. Empirical Results and Application Domains

HiSS models have been validated across a range of technical domains:

Domain	Empirical Setting	HiSS Gain Over Baselines
Sequence Prediction	Tactile, IMU, MoCap (CSP-Bench) (Bhirangi et al., 2024)	Median MSE ↓23% vs. flat S4/Mamba/LSTM/Transformers
Brain Dynamics/fMRI	ADHD-200, SchizoConnect (Yang et al., 28 Jun 2025, Wei et al., 2024)	Classification accuracy gains, robust metastability quantification, biomarker interpretability
Surgical Video Analysis	Cholec80, MICCAI2016, Heichole (Wu et al., 26 Jun 2025)	+2.8% / +4.3% / +12.9% absolute SOTA improvement
Power Systems	Two-machine, 500-bus grid (Liu et al., 25 Aug 2025)	Error ≲0.03%; superior eigenmode recovery; ×5 faster convergence vs NODE baselines
Robotics RL	Real/Sim manipulation (Lundell et al., 2018)	30–50% fewer roll-outs to convergence, strict safety
RL in MDP Navigation	Multi-room, KeyDoor (Steccanella et al., 2021)	2–8× faster exploration, effective transfer

These results consistently show HiSS approaches outperforming flat (single-layer) SSMs, standard RNNs, attention-based models, and conventional filter pipelines under both full- and low-data regimes.

5. Interpretability, Safety, and Transfer

A defining advantage of HiSS is the inherent interpretability of its decomposed structure:

• Interpretable dynamics: Explicit $A,B,C,D$ parameterization at each level, and, when present, clear mapping to semantic regions (e.g., rooms, surgical phases, grid components) (Liu et al., 25 Aug 2025, Steccanella et al., 2021).
• Safety and constraint enforcement: By construction, hierarchical null-space projections restrict learning and inference to low-dimensional safe manifolds, preventing collisions and constraint violations (Lundell et al., 2018).
• Metastability and brain state transitions: Analyzing dwell times and empirical transition matrices extracted from HiSS representations allows quantitative study of metastability in neural systems, with significant group-level differences demonstrated in clinical data (Yang et al., 28 Jun 2025).
• Task transfer: Once an abstract task graph or set of option policies is learned, new downstream tasks sharing the same structure benefit from accelerated learning through re-use of high-level skills (Steccanella et al., 2021).

6. Computational Scaling and Model Complexity

A result of the recursive structuring, HiSS models inherit the linear time and space complexity of their SSM building blocks (e.g., S4/Mamba), even as modeling expressive long-range correlations and complex dependency structures:

• Linear inference: $O(T)$ in sequence length $u_t$0 for both base and hierarchical SSM compositions (Bhirangi et al., 2024, Wu et al., 26 Jun 2025).
• Hierarchical parameterization: Only a single additional hyperparameter (chunk/stride size or hierarchy depth) typically required over flat SSMs (Bhirangi et al., 2024, Yang et al., 28 Jun 2025).
• Parallelizability: Modular SSM blocks or null-space projections can often be solved/batched independently at a given level (Lundell et al., 2018, Liu et al., 25 Aug 2025).
• Transformer-HiSS hybrids: When combined with chunked attention, computational complexity reduces to $u_t$1, substantially below full-attention's $u_t$2 for practical block sizes $u_t$3 (Bhirangi et al., 2024).

7. Theoretical and Practical Generalization

HiSS frameworks generalize to any domain where:

• The system can be meaningfully decomposed into interacting subsystems, time-scales, or semantic categories.
• Only partial observations are available at different scales, motivating multilevel observers or clustering.
• Safety, interpretability, or modularity are operational imperatives.

Concrete examples beyond those already surveyed include bio-signal modeling, industrial process control, traffic and resource flow networks, and large-scale sensor arrays (Liu et al., 25 Aug 2025). By modularizing dynamics and inference, HiSS architectures define a unifying principle for scalable, interpretable modeling across scientific and engineering disciplines.