CS3: Cross-Scale State Space

• CS3 is a unifying framework that formalizes inter-scale mappings and state-space construction to integrate dynamics across varying dimensions and resolutions.
• It employs state space compression methods that balance predictive fidelity with computational cost using information-theoretic and algorithmic approaches.
• Applications span neuroscience, control theory, and time-series forecasting, demonstrating CS3’s capability to model and control multiscale systems effectively.

The Cross-Scale State Space (CS3) framework encompasses a set of mathematical, algorithmic, and modeling principles designed to unify multilayered dynamical systems by explicitly capturing structure and interactions across multiple scales of resolution, dimension, and abstraction. The CS3 paradigm has emerged independently in contemporary research domains such as neuroscientific modeling via state-space neural networks, algorithmic state-space compression for complex physical and social systems, property-driven coarse-graining of Markov chains, time-series modeling with deep architectures, and control theory for dimension-varying dynamical systems. The unifying feature of CS3 is the formal construction of inter-scale mappings, dynamics, information conservation, and performance guarantees, thereby enabling both theoretical analysis and high-performance computational schemes for systems that span heterogeneous scales.

1. Mathematical Definition and Universal State Space Construction

CS3 formalizes state spaces that bridge models of differing dimension, granularity, or abstraction via explicit metric and equivalence constructions. A canonical instance is the cross-dimensional Euclidean space $Ω$ (Cheng, 21 Jan 2026):

• $Ω$ is defined as the quotient of the disjoint union $\bigcup_{n=1}^\infty \mathbb R^n$ by the equivalence $x\sim y$ iff $x$ and $y$ can be represented as identical vectors in a common embedding via Kronecker products. The universal metric is

$$d_V(\bar x,\bar y) = \|\ x\otimes 1_{t/m} - y\otimes 1_{t/n} \| / \sqrt t,$$

where $x\in\mathbb R^m$, $y\in\mathbb R^n$, $t = \mathrm{lcm}(m,n)$.

• Each $Ω$0 is isometrically embedded into $Ω$1, which retains a path-connected topology and supports vector addition, inner products, and projections.
• Cross-dimensional projections, $Ω$2, provide the closest $Ω$3-representation to a given $Ω$4.
• State trajectories with dimension jumps are Lipschitz-continuous in $Ω$5, enabling generalized stability, controllability, and observability results for dimension-varying control systems.

This universal space $Ω$6 serves as the mathematical substrate for reasoning about and constructing dynamical models and control laws that interpolate or aggregate across variable system dimensionalities.

2. Cross-Scale Modeling and State-Space Compression

CS3 subsumes and operationalizes the theory of State Space Compression (SSC) (Wolpert et al., 2014), which posits that any high-dimensional dynamical system can be reduced to a lower-dimensional representation by explicitly optimizing a trade-off between predictive accuracy and computational cost:

• The compression map $Ω$7 projects the fine-scale state $Ω$8 onto a macrostate $Ω$9, with associated reconstruction $\bigcup_{n=1}^\infty \mathbb R^n$0 and observable prediction $\bigcup_{n=1}^\infty \mathbb R^n$1.
• The dynamics in $\bigcup_{n=1}^\infty \mathbb R^n$2 are induced via

$\bigcup_{n=1}^\infty \mathbb R^n$3

• The optimization objective is

$\bigcup_{n=1}^\infty \mathbb R^n$4

where $\bigcup_{n=1}^\infty \mathbb R^n$5 (e.g., a distortion or mutual information term) quantifies loss of predictive fidelity, and $\bigcup_{n=1}^\infty \mathbb R^n$6 quantifies resource or complexity costs.

• Information-theoretic metrics, such as transfer entropy and conditional mutual information, are used to quantify information content and causation across scales.
• SSC underlies the construction of multi-layered CS3 models that align macrostate and microstate dynamics via variational, Markovian, or learning-based parameterizations.

A plausible implication is that SSC provides both the formal basis and the computational pipeline for constructing scale-spanning CS3 models in scientific and engineering contexts.

3. Algorithmic Realizations in Deep Learning and Neuroscience

Recent work has instantiated CS3 in state-space models (SSMs) designed for both neuroscience and time-series machine learning (Lu et al., 18 Jul 2025, 2502.11340):

• The S5 variant of state-space neural models uses a diagonalizable linear recurrence,

$\bigcup_{n=1}^\infty \mathbb R^n$7

where each hidden unit evolves as $\bigcup_{n=1}^\infty \mathbb R^n$8 with $\bigcup_{n=1}^\infty \mathbb R^n$9 encoding time constants and resonance frequencies.

• These dynamics, without internal nonlinearity, recapitulate features of single-neuron biophysics (membrane potential, gating variables, decay, resonance), and yield emergent population behaviors such as time cells, ramping activity, and traveling waves.
• Pure rotational dynamics in the complex plane induce sequences and spatial patterns analogous to experimentally observed neurocognitive phenomena.
• SSMs in time series forecasting, exemplified by S2TX, incorporate Mamba-type global SSM modules for capturing long-range dependencies and Transformers for high-resolution local dynamics, with cross-attention mechanisms fusing global and local scales (2502.11340).
• Cross-variate and cross-scale interactions are realized by combining module outputs via cross-attention, yielding robust benchmarks for multivariate time-series forecasting.

These findings demonstrate that state-space models, when constrained by CS3 principles, enable interpretable, scalable, and domain-general architectures capable of matching the multiscale complexity of biological and artificial systems.

4. Property-Driven Coarsening and Behavioral Preservation

For Markovian and stochastic systems, a CS3 approach emphasizes property preservation across scales (Michaelides et al., 2016):

• Given a continuous-time Markov chain $x\sim y$0 and a set of temporal logic specifications $x\sim y$1, the state space $x\sim y$2 is partitioned into macro-states $x\sim y$3 to minimize the maximum difference in property-satisfaction probabilities,

$x\sim y$4

• Gaussian process (GP) emulation, multi-dimensional scaling (MDS), and clustering algorithms are employed to construct low-dimensional embeddings and macro-state aggregations optimized for property coherence rather than mere transition-rate similarity.
• Empirical analysis in epidemic and performance models confirms that the resulting macro-dynamics preserve targeted behavioral properties while substantially reducing computational cost (e.g., 82% reduction in transitions for coarse-grained simulations).
• This property-driven, specification-centric reduction is a key distinguishing feature of behaviorally meaningful CS3 modeling pipelines.

The significance of this approach is its capacity to deliver semantic compression, directly aligned with the modeler's observational or control objectives.

5. Hierarchical Organization, Aggregation, and Large-Scale Networks

CS3 extends naturally to systems exhibiting hierarchical or networked organization and dimension-varying subsystems (Cheng, 21 Jan 2026):

• Aggregation across network hierarchies is accomplished via projection-based approximations and least-squares optimal mappings between high-dimensional subsystem states and lower-dimensional aggregate representations.
• Linear dynamics are approximated in aggregate subspaces using operators,

$x\sim y$5

minimizing the discrepancy over relevant time scales.

• The universal CS3 space $x\sim y$6 supports seamless embedding, continuity, and operation of subsystems with dynamic or heterogeneous dimensions, facilitating stability analysis, robust switching, and controllability via the structure of subspace lattices.
• Empirical results indicate that error bounds for such aggregations scale favorably, and relative error remains small under moderate dimension reduction.

This framework enables analysis and design for large-scale, multi-resolution, and multi-dimensional networks, providing a tractable mathematical foundation for modern complex systems.

6. Cross-Scale Information Flow, Causality, and Complexity Quantification

CS3 models support explicit quantification of information flow and complexity between scales:

• Cross-scale transfer entropy quantifies bottom-up causation,

$x\sim y$7

while $x\sim y$8 measures top-down influences.

• The compression complexity of a system is defined as the minimum achievable value of the CS3/SSC trade-off objective, normalized by the uncompressed baseline.
• Systems that tolerate large-scale reductions with minimal predictive loss are "highly compressible"; those that resist meaningful compression are deemed complex in this operational sense (Wolpert et al., 2014).
• The alignment of observations, predictions, and decisions across scale is formalized via loss functionals summing per-level errors and regularization penalties.

These mechanisms lend CS3 the capacity for rigorous, information-theoretic analysis of structure, efficiency, and causality in multiscale systems.

7. Applications, Advantages, and Limitations

CS3 has found application in a diverse array of fields:

Domain	CS3 Realization	Salient Features
Neuroscience	S5 SSM, time cells, traveling waves	Biophysical interpretability, emergent cognitive function
Control Theory	$x\sim y$9 universal state space	Cross-dimensional switching/stability/controllability
Statistical Physics	SSC for macrostates (volume, pressure)	Thermodynamic law recovery
Time Series Forecasting	S2TX, Mamba + Transformer, cross-attention	Multiscale predictive performance, efficiency
Systems Biology	Property-driven Markov chain coarsening	Specification-preserving reductions

Chief advantages of the CS3 paradigm include scale-consistency, interpretability, domain-general applicability, and support for efficient learning and simulation. Limitations observed in current implementations include a focus on binary or fixed scale splits, lack of explicit intermediate-scale modeling, and restricted capacity for local cross-variate interactions in some architectures (2502.11340). Addressing these limitations remains an open direction for future research.

CS3 unifies multiscale mathematical modeling, efficient algorithmic compression, and principled information flow quantification, providing a robust platform for the study and engineering of complex systems across scientific and technological domains.