Hierarchical Reference Structures

• Hierarchical reference structures are multi-level architectures that organize information recursively, balancing specificity with abstraction.
• They are pivotal in applications such as communication protocols, data compression like HRLZ, and video/image coding, enhancing efficiency and retrieval.
• Their formal models utilize structured dependencies and causal embeddings to secure high accuracy and significant gains in coding performance.

Hierarchical reference structures are multi-level architectures for organizing, retrieving, or encoding information in which reference relationships are defined recursively across levels of abstraction, generalization, or temporal/spatial scale. These structures arise in diverse domains, including natural language reference, data compression, video/image coding, and geometric representations of knowledge hierarchies. At their core, hierarchical reference systems facilitate efficient access, communication, or coding by leveraging structured dependencies, often balancing specificity and generality or exploiting correlations across the hierarchy.

1. Formal Models of Hierarchical Reference

Across application domains, hierarchical reference systems are formalized by imposing structured relations between entities (objects, views, strings, concepts, or events), such that each level refines or abstracts the level beneath it. In natural language reference games, each object is described by an attribute vector $o \in \{1, \ldots, k\}^n$, and the relevant attributes for reference are encoded in a Boolean vector $r \in \{0,1\}^n$. A concept $C(o, r)$ is defined as the set of objects $o'$ such that $o'_i = o_i$ on all dimensions for which $r_i = 1$:

$$C(o, r) = \{ o' \in \{1, \ldots, k\}^n \mid \forall i, (r_i = 1 \implies o'_i = o_i) \}.$$

Hierarchical levels correspond to varying the count of relevant ($r_i=1$) attributes: leaves are fully specific, while higher levels abstract over attribute sets, creating a tree or DAG of nested concepts (Ohmer et al., 2022).

In data compression, hierarchical reference structures are often modeled as rooted arborescences (directed trees) among data instances. For example, in hierarchical relative Lempel-Ziv (HRLZ) compression, a set of strings $\mathcal{S} = \{S_1,\ldots,S_m\}$ is organized as a tree, where each string is parsed relative to its unique parent. The optimal hierarchy minimizes the total number of dictionary phrases across all nodes and is found as a minimum-weight spanning arborescence on a complete cost graph defined by pairwise parse costs (Bille et al., 2022).

Geometrically, hierarchical data can be embedded as a set of events in $D=3$ Minkowski spacetime such that parent-child pairs are realized as proper-time-separated lightcones. The injective map $r \in \{0,1\}^n$0 ensures that each $r \in \{0,1\}^n$1 is in the causal future of its parents $r \in \{0,1\}^n$2, preserving all hierarchical (transitive) parent relations (Anabalon et al., 7 May 2025).

2. Architectural and Algorithmic Realizations

The implementation of hierarchical reference structures varies by domain:

• Multi-agent communication: Agent architectures use encoders (MLPs) to process attributes and relevance, initializing RNNs (GRUs) to generate symbol sequences that reflect the hierarchical conceptual structure. The emergent protocol composes reference at different specificity levels and supports both implicit omission and explicit symbol marking for irrelevance. Training is performed with standard cross-entropy loss for referential success in a multi-level reference game (Ohmer et al., 2022).
• Video and image coding: Hierarchical reference is central to video coding standards. For instance, in pseudo-sequence based 2D hierarchical reference for light-field image compression, views are organized spatially into quadrants, each encoded hierarchically to maximize inter-view prediction under buffer size constraints. Reference set selection is distance-based in microlens coordinates, and motion vector scaling leverages 2D relative positions, not just temporal intervals as in 1D sequences (Li et al., 2016). In neural video codecs, explicit multi-level reference structures are aligned with group-of-picture (GOP) quality layers, and fusion networks adaptively weight spatial/temporal reference features (Liao et al., 4 Sep 2025).
• Data compression: HRLZ employs an explicit parent-selection process: construct the complete reference graph, weight edges by RLZ cost (number of phrases), and extract the optimal hierarchy via Tarjan's minimum arborescence. Locality-sensitive hashing can sparsify candidate references while preserving near-optimality. Compression proceeds by RLZ parsing with respect to parents, decoding by BFS traversal of the tree (Bille et al., 2022).
• Geometric embedding: The causally-constrained, iterative adjustment of event times in Minkowski space enforces all parent-child conditions, producing perfect hierarchical embeddings. Retrieval consists of causal cone queries: for any event $r \in \{0,1\}^n$3, parents are those $r \in \{0,1\}^n$4 in the past light cone achieving minimal proper time separation (Anabalon et al., 7 May 2025).

3. Emergent Properties and Empirical Outcomes

Hierarchical reference structures induce a range of desirable properties:

• Efficiency and generalization: In referential games, emergent protocols achieve $r \in \{0,1\}^n$598% accuracy and high zero-shot generalization to novel objects (near 100% for $r \in \{0,1\}^n$6) and abstractions. Implicit abstraction reduces message length; explicit symbols signal ignored attributes. Emergent messages display compositional structure, evidenced by topographic and bag-of-symbols disentanglement metrics (Ohmer et al., 2022).
• Compression gains: In HRLZ compression of genomic data, the optimal hierarchy typically halves the number of LZ phrases compared to the best single-reference scheme (e.g., E. coli data: $r \in \{0,1\}^n$7 phrases for single-ref, $r \in \{0,1\}^n$8 for HRLZ), with only a modest penalty in decode speed and shallow hierarchy depth ($r \in \{0,1\}^n$9200) (Bille et al., 2022).
• Coding efficiency: For light-field compression, the pseudo-sequence 2D hierarchical reference structure achieves up to 14.2% bitrate reduction over prior methods. Efficient quadrant-based buffer management ensures reference memory is reduced by over an order of magnitude (peak=12 vs. 164) with minimal encoding complexity increase (Li et al., 2016). Neural video codecs with synchronized hierarchical reference-quality structures yield $C(o, r)$017% average bitrate savings over traditional codecs across several datasets (Liao et al., 4 Sep 2025).
• Perfect semantic embedding: Hierarchical data from WordNet are embedded in three-dimensional Minkowski space with zero distortion: mean rank 1, mean average precision 1. The geometry enforces hierarchical inheritance via light-cone relations, generalizes to both unique and multiple parent structures, and converges efficiently for large vocabularies (Anabalon et al., 7 May 2025).

4. Reference Selection and Disentanglement Strategies

Reference structure selection is governed by context, proximity, and task:

• Reference context: In multi-level reference games, the number of relevant attributes determines specificity; unnecessary details are dropped for efficiency. In communication, both implicit omission (shorter messages) and explicit marking (operator symbols) emerge. For image compression, reference sets are chosen to maximize prediction given spatial layout and buffer requirements (Ohmer et al., 2022, Li et al., 2016).
• Distance and cost-based selection: In HRLZ, minimum-cost parent references are chosen by evaluating per-string RLZ costs. In 2D coding, reference lists are sorted by Euclidean distance; in motion estimation, scaling formulas account for geometric displacements rather than just temporal order (Li et al., 2016, Bille et al., 2022).
• Fusion and synchronization: In neural coding, per-level reference sets are fused by learned weights and aligned with quality/quantization structures. Randomized training on quality scales inhibits overfitting to rigid hierarchies and supports robustness at inference (Liao et al., 4 Sep 2025).
• Geometric separability: The Minkowski embedding approach ensures disentanglement of inheritance (causality) from general proximity, allowing efficient cone queries and theoretically perfect recovery of DAG or tree relations (Anabalon et al., 7 May 2025).

5. Practical Applications and Broader Implications

Hierarchical reference structures support diverse applications:

Domain	Reference Structure	Empirical Benefit
Multi-agent communication	Concept tree; symbol protocol	High context-sensitive accuracy, compositionality (Ohmer et al., 2022)
Light-field coding	2D quadrant hierarchy	6.5–14.2% average bitrate reduction, efficient memory use (Li et al., 2016)
Genomic/data compression	Hierarchical RLZ tree	2× fewer phrases, near-single-ref decode time (Bille et al., 2022)
Video coding	Temporal hierarchy (GOP)	$C(o, r)$117% BD-rate reduction in learned codecs (Liao et al., 4 Sep 2025)
Knowledge representation	Minkowski spacetime DAG	Perfect parametric embeddings, new causal queries (Anabalon et al., 7 May 2025)

In symbolic domains, hierarchical reference is closely tied to the emergence of compositionality and abstraction. In coding and compression, it enables efficient buffer management, scalable parsing, and exploitation of inter-object or inter-frame correlations. Geometric formalisms establish a physical analogy between information inheritance and causality, connecting hierarchical representation with conformal and relativistic symmetries.

A plausible implication is that hierarchical reference architectures, when properly aligned with data or communicative task structure, systematically yield improvements in efficiency, interpretability, or downstream task performance, often beyond what flat or naïve reference schemes can achieve.

6. Limitations, Challenges, and Future Directions

Key limitations are domain-specific. In HRLZ compression, the cost of constructing the pairwise cost graph is quadratic in the number of objects, necessitating sparsification via locality-sensitive hashing for scalability, with no formal guarantees but robust empirical performance (Bille et al., 2022). In deep hierarchical video coding, balancing multi-reference structures with quality layers and managing cross-level buffer constraints remain challenging, requiring adaptive learning strategies (Liao et al., 4 Sep 2025). Geometric approaches are currently verified only for acyclic or weakly cyclic hierarchies with low out-degree, and their extension to arbitrary knowledge graphs or highly entangled structures requires further investigation (Anabalon et al., 7 May 2025).

Hierarchical reference structures will continue to play a central architectural role in multi-modal communication, efficient storage and transmission, and interpretable knowledge systems, with ongoing research into optimization algorithms, learnable reference selection, and geometry-inspired index structures.