Unified Multi-Scale Causal Structure

• Unified multi-scale causal structure is a framework that unifies high-dimensional observations with aggregated causal abstractions using principled mathematical partitions.
• It leverages advanced algorithms such as sparse autoencoders, stationary wavelet transforms, and graph neural networks to model hierarchical and nested causal relationships.
• The approach has demonstrated empirical success in neuroscience, economics, and biology, offering scalable, interpretable causal inference from complex data.

A unified multi-scale causal structure provides a principled mathematical and algorithmic foundation for discovering and representing causal relationships across distinct levels of granularity—ranging from micro-level features (e.g., high-dimensional observations, fine temporal or spatial scales) to macro-level variables (e.g., aggregated states, system-level summaries, higher-order constructs). Such a structure addresses the challenge of linking rich, high-dimensional data to interpretable causal abstractions, while rigorously preserving the relevant interventionist semantics associated with each level. The development of this framework draws from domains including theoretical causal inference, deep learning, time-series analysis, statistical signal processing, and mathematical physics, yielding methodologies that generalize over complex systems in neuroscience, biology, economics, and quantum theory.

1. Fundamental Definitions: Micro- and Macro-Variables

The central construct in any multi-level causality framework is the operational distinction between micro-variables and macro-variables. Micro-variables $I$ and $J$ are assumed to encode high-dimensional or fine-grained system states, often observed as data points $i\in I \subset \mathbb{R}^m$ and $j\in J \subset \mathbb{R}^n$ drawn from discrete but possibly extremely large sets. The causal system is modeled as:

$$P(J, I) = \sum_{h} P(J\,|\,I, h)\, P(I\,|\,h)\, P(h)$$

where $h$ is a latent confounder.

A key insight is to define macro-variables $C$ and $E$ as partitions of the micro-state spaces, such that each macro-variable value comprises micro-states that share identical interventional consequences:

• For inputs $i_1, i_2$: $$i_1 \sim_I i_2 \iff \forall j \in J: P(j\,|\,man(i_1)) = P(j\,|\,man(i_2))$$
• Similarly for outputs, $j_1, j_2$: $$j_1 \sim_J j_2 \iff \forall i \in I: P(j_1\,|\,man(i)) = P(j_2\,|\,man(i))$$ The coarsest possible such partitions, denoted $\Pi_c(I)$ and $\Pi_c(J)$, are termed the fundamental causal partitions. Macro-variables are thus defined as $C(i)$ and $E(j)$ indexing the corresponding partition class of $i$ and $j$ (Chalupka et al., 2015).

2. Minimal Sufficiency and Hierarchical/Nested Causal Variables

The fundamental causal partitions play a unique role: the vector of macro-variable counts in an i.i.d. sample from an interventional distribution is the minimal sufficient statistic for the family of interventional distributions. No strictly coarser macro-variable can preserve the full interventionist predictive power. Furthermore, coarse-grainings of $C$ and $E$ (termed subsidiary macro-variables) are permitted only when their induced interventions yield invariant macro-responses—a property leveraged to construct hierarchical and non-interacting multi-level causal structures. This explicit formalization enables identification of nested non-interacting causal subsystems, where the fundamental variables factor as $C = C^1 \otimes C^2$, $E = E^1 \otimes E^2$ if $$P\big((E^1, E^2)\,|\,do(C^1, C^2)\big) = P(E^1\,|\,do(C^1))\cdot P(E^2\,|\,do(C^2))$$ (Chalupka et al., 2015).

3. Causal Structure Learning across Scales: Methodological Advances

Recent frameworks have developed algorithms for learning unified multi-scale causal structures from complex data:

• Fundamental Macro-Variable Discovery: Algorithmic procedures build density estimators of $P(j\,|\,man(i))$, form effect/cause-vectors for clustering, and merge clusters based on similarity in causal effect tables. This produces partition functions $g,h$ extracting macro-variables $C = g(I), E = h(J)$ (Chalupka et al., 2015).
• Multi-Granularity Causal Structure Learning (MgCSL): Employs sparse autoencoders for coarse-graining micro-variables to macro-variables, combined with multilayer perceptrons and a Schur-decomposition based constraint for efficient DAG learning at multiple granularities, unifying micro and macro causal discovery (Liang et al., 2023).
• Wavelet-Based Multiscale DAGs (MS-CASTLE, MN-CASTLE): High-resolution time series are decomposed using stationary wavelet transforms, and scale-dependent instantaneous and lagged relationships are fitted via block-diagonal VAR models with acyclicity constraints. In nonstationary settings, Bayesian Gaussian process models capture smooth time- and scale-varying DAGs, while learning a shared node ordering across scales (D'Acunto et al., 2022, D'Acunto et al., 2022).
• Spectral Deep Learning for Multimodal Structure: In structural brain modeling, Laplacian-harmonic–based spectral embeddings enable cross-modal, multi-scale alignment; Graph Variational Autoencoders disentangle scale-specific (causal vs. noncausal) features, with mutual information regularization enforcing robust causal inference (Xia et al., 12 Dec 2025).
• Multiscale Causal Backbone (MCB): Population-level multi-scale DAGS are aggregated into a shared causal backbone by extracting $p$-persistent arcs consistent across subjects and scales, optimizing penalized likelihood over individual joint models for interpretability and generalizability (D'Acunto et al., 2023).

Framework	Scale Handling	Key Mechanism
(Chalupka et al., 2015)	Micro & macro, hierarchical	Partition equivalence, minimal sufficiency
(Liang et al., 2023)	Multi-granularity	SAE + MLP, Schur-based acyclicity penalty
(D'Acunto et al., 2022)	Temporal multi-scale	SWT, block-diagonal SVAR, continuous DAG constr.
(Xia et al., 12 Dec 2025)	Multi-modal, spectral	Laplacian harmonics, GraphVAE, MI regularizer
(D'Acunto et al., 2023)	Temporal population backbone	Persistent arcs, MDL-optimized aggregation

4. Hierarchical and Multimodal Multi-Scale Causal Systems

Causal structure at multiple hierarchical levels is addressed, e.g., in the hierarchical structural causal model (HSCM) (Hermes et al., 25 Nov 2025), which integrates group-level (macro) and unit-level (micro) DAGs, together with cross-level dependencies, under additive-noise nonlinear SEMs. Group-level effects and confounders are absorbed via group-specific intercepts, and cross-level edges are identified via generalized additive modeling with smoothness-based hypothesis testing.

In AI-driven biology, multiscale frameworks explicitly encode the known DAG of biological scales (DNA, RNA, protein, cell, tissue, phenotype), aligning latent representations across omic modalities and species using VAEs and graph neural networks, and allowing cross-scale intervention simulations (e.g., in silico gene knockouts) (Wu et al., 2024).

5. Time-Frequency and Quantum Multi-Scale Causality

Unified time-frequency causal inference frameworks (e.g., MB-VLGC (Sookkongwaree et al., 1 Aug 2025)) extend classical Granger causality by modeling variable-lag, frequency-specific relationships using per-band dynamic time warping and band-specific OLS regressions, with pipelines for aggregate inference and explicit theoretical variance bounds.

In quantum settings, the global causal structure is reconstructed from local pseudo-density operator (PDO) marginals across space-time regions using maximum-entropy principles and neural variational representations. This unifies spatial and temporal causality, allowing constraints such as positivity and separability (for quantum or classical limits) to be imposed directly (Jia et al., 2023).

6. Higher-Order Mathematics: Category- and Topology-Based Unification

The Universal Causality Layered Architecture (UCLA) (Mahadevan, 2022) formalizes causality across four categorical/topological layers:

1. Simplicial Layer: Combinatorial models of interventions as face/degeneracy maps.
2. Categorical Layer: Graph- and string-diagrammatic causal models, with categorical surgery induced by the simplicial structure.
3. Data Layer: Set-valued functors encoding datasets as instances over models.
4. Homotopy Layer: Topological realization (e.g., nerve, homotopy colimit) of data/model combinations, with causal effect defined as (non-)homotopy equivalence of resulting topological spaces. Lifting problems and universal arrows formalize queries and intervention semantics across all layers, integrating combinatorial, graph-theoretic, statistical, and topological perspectives.

7. Empirical Evidence and Applications

Unified multi-scale causal structure enables:

• Precise, interpretable macro-causal models in systems neuroscience, with invariance under neuron permutation and hierarchy (Chalupka et al., 2015, Liang et al., 2023, Xia et al., 12 Dec 2025).
• Improved accuracy and scalability in high-dimensional time series, outperforming NOTEARS and other baselines in fMRI and economic risk propagation (Liang et al., 2023, D'Acunto et al., 2022, Sookkongwaree et al., 1 Aug 2025).
• Causal backbone extraction for individual differences in brain connectivity, enabling causal fingerprinting and system-level insights (D'Acunto et al., 2023).
• Robust multi-omics–to–phenotype prediction, simulation of interventions, biomarker prioritization, and generalized cross-domain adaptation in biological data (Wu et al., 2024).

Taken together, a unified multi-scale causal structure systematically extends the reach of causal discovery, inference, and explanation from raw high-dimensional data to abstract macro-level phenomena, providing both rigorous theoretical guarantees and scalable, empirically validated algorithms across multiple domains.