Structural Entropy-based Evolution (SEE)

• Structural Entropy-based Evolution (SEE) is a framework that quantifies system complexity using entropy measures defined on networks, molecular descriptors, or event graphs.
• SEE integrates efficient, incremental algorithms to monitor and adjust entropy levels, ensuring scalable real-time evolution control across dynamic systems.
• Applied in diverse domains, SEE enables identification of key structural patterns and extraction of driving mechanisms, enhancing automated mechanism inference and topic tracking.

Structural Entropy-based Evolution (SEE) encompasses a set of theoretical concepts, algorithmic methodologies, and applied frameworks in which structural entropy—variously defined over degree sequences, partitions, or molecular descriptors—guides or interprets the evolution of complex systems. Originating in the context of network science and adopted in dynamic graph partitioning, molecular trajectory analysis, and social event tracking, SEE combines order-theoretic entropy metrics with data-driven or algorithmic approaches that optimize, monitor, or explain system change over time. The central object is a structural entropy measure (e.g., Shannon entropy of the normalized degree sequence, multi-level encoding tree entropy, or conditional information entropy), which acts as an interpretable, often scalable, and theoretically principled signal for network evolution, mechanism extraction, or temporal event alignment.

1. Structural Entropy Foundations and Key Formulas

Structural entropy, in SEE, is a graph or system complexity measure derived from the uncertainty present in specific structural representations. Various formalizations exist:

• Graph-based Structural Entropy (degree sequence): For a graph $G=(V,E)$ with degree sequence $d_i$, the structural information (SI) is defined as the Shannon entropy of the normalized degree distribution, $p_i = d_i/2m$:

$SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$

• Multi-level Partition Structural Entropy: For a weighted undirected graph and a two-level encoding tree $\mathcal{T}$ with partition $\pi = \{c_1, ..., c_k\}$:

$$H(G; \pi) = -\sum_{c \in \pi} \frac{g(c)}{\mathrm{vol}(G)} \log_2 \left( \frac{\mathrm{vol}(c)}{\mathrm{vol}(G)} \right)$$

where $g(c)$ denotes cut size and $\mathrm{vol}(c)$ the cluster volume.

• Conditional Structural Entropy in Molecular Dynamics: For features $X$ and label $d_i$0, typically reflecting structure and cluster assignment,

$d_i$1

This quantifies how much uncertainty in $d_i$2 remains after knowing $d_i$3.

These entropy measures are central for quantifying complexity, identifying key structural differences, and driving evolutionary processes in dynamic systems.

2. Structural Entropy as an Evolutionary Driver in Networks

SEE provides an operational framework for controlling or interpreting the evolution of networked systems by manipulating or monitoring structural entropy:

• Entropy-driven Network Design: The MaxEntropy problem is to add up to $d_i$4 edges to an input graph to maximize its von Neumann entropy. The SI proxy offers a scalable algorithm—greedily selecting edge additions that most increase $d_i$5—which provably approximates the spectral entropy within an absolute error of $d_i$6, with runtime $d_i$7 for $d_i$8 additions (Liu et al., 2021).
• Evolution Control: For evolving graph streams, one monitors $d_i$9. At each time step, local edge additions/deletions can be triggered to maintain a desired entropy level $p_i = d_i/2m$0:
  • Compute $p_i = d_i/2m$1 incrementally in $p_i = d_i/2m$2.
  • Implement local policies to steer $p_i = d_i/2m$3 towards $p_i = d_i/2m$4.

Because SI reflects degree heterogeneity and can be computed in linear time, SEE enables real-time network evolution at scale while preserving interpretability and theoretical guarantees (Liu et al., 2021).

3. Incremental and Dynamic Structural Entropy Measurement

In dynamic graph applications, measuring or optimizing structural entropy across time is crucial for understanding and controlling evolving community structure:

• Incremental Frameworks: The Incre-2dSE framework supports efficient computation of two-dimensional (height-2 encoding tree) structural entropy $p_i = d_i/2m$5 for dynamic graphs. Two main update strategies are used:
  • Naive Adjustment: Update only edge/node-level quantities without allowing node migration between communities. Computational cost is $p_i = d_i/2m$6 where $p_i = d_i/2m$7 is the number of edge/node updates.
  • Node-Shifting: Iteratively allow nodes to migrate to their optimal-preference community to minimize $p_i = d_i/2m$8, with convergence guarantees and improved entropy reduction. Cost can reach $p_i = d_i/2m$9, with $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$0 the iteration cap and $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$1 number of communities.

The incremental formula allows $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$2 updates in $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$3 time, contrasted with static recomputation $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$4 plus full community detection (Yang et al., 2022).

• Empirical Results: Incre-2dSE showed entropy reductions of up to 10–12% (for node-shifting), time savings of 4x–140x over offline Infomap re-partition, and stable monotonic curves for $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$5 over time.

A plausible implication is that incremental SEE methods are best suited for scenarios with small, frequent update batches (e.g., streaming social or biological networks), with offline recalculation reserved for large, global structural changes.

4. Structural Entropy-guided Structural Evolution in Molecular Dynamics

SEE has been generalized beyond graphs to high-dimensional molecular data, particularly in non-adiabatic molecular dynamics (NAMD):

• Workflow: Starting from ensembles of trajectory surface hopping data, SEE identifies distinct reaction channels and their active coordinates nonparametrically:

1. Represent structures using geometric/spectral high-dimensional descriptors (e.g., Cartesian, RIC, MBTR, SOAP).
2. Reduce dimensionality using a deep autoencoder to learn nonlinear 2D latent coordinates.
3. Cluster the low-dimensional embeddings with DBSCAN, identifying distinct decay channels.
4. Compute conditional entropy $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$6, identifying original descriptors with minimal entropy (i.e., maximal discriminative power for each channel).
5. Extract “active coordinates”—minimal sets of features driving each channel (Liu et al., 17 Nov 2025).

• Application Results: For keto isocytosine and methaniminium cation, SEE autonomously retrieved known and previously-unreported reaction channels, with channel purity ≥95% and dimensionality reductions from $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$7 descriptors to $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$8 key coordinates.

SEE in molecular dynamics thus demonstrates the domain-agnostic nature of structural entropy: leveraged for efficient, automated mechanistic inference in high-dimensional, temporally-evolving datasets.

5. Structural Entropy-based Evolution in Social Event Tracking

SEE principles have been applied in the context of social media event detection and topic evolution:

• Formalism: Each day, events detected via retrieval-augmented generation (RAG) form nodes in a weighted graph, with edges determined by shared keywords and embedding similarity.
• Algorithm:

1. Build a daily event-graph $SI(G) = -\sum_{i=1}^n p_i \log_2 p_i$9.
2. Apply two-level (height-2) structural entropy minimization—repeatedly merging clusters to minimize

   $\mathcal{T}$0

3. Inherited alignment, enforcing temporal continuity by preventing merges of events unless new evidence arises.
4. Forgetting mechanism: drop events with no new support.

• Performance: On Events2012, RagSEDE (incorporating SEE) achieved high topic diversity (TD = 0.88) and coherence ($\mathcal{T}$1), outperforming static and dynamic topic models; ablation confirmed the essential role of SEE’s inheritance and forgetting modules (Liu et al., 17 Jan 2026).

This suggests that SEE enables scalable, unsupervised, dynamically-aligned topic evolution—identifying and tracking temporally coherent, semantically distinct social events with interpretable evolution keywords.

6. Theoretical Guarantees, Efficiency, and Limitations

SEE methodologies are underpinned by rigorous theoretical analysis:

• Provable accuracy of SI as a von Neumann entropy proxy: For any undirected, unweighted graph, the entropy gap $\mathcal{T}$2, ensuring SI delivers both scalability ($\mathcal{T}$3 time) and interpretability (Liu et al., 2021).
• Correctness of incremental entropy measurement: For node addition, communities of sufficient size guarantee optimal assignment; error terms in incremental updates decay as $\mathcal{T}$4 with graph size (Yang et al., 2022).
• Algorithmic complexity: Incremental SEE updates scale with batch size rather than full graph size, enabling real-time evolution monitoring for systems with small, frequent modifications.

Limitations include the two-level (height-2) restriction of encoding trees—generalization to higher $\mathcal{T}$5 remains open—, sensitivity to large global structural changes, and, for node-shifting strategies, potential non-convergence in pathological cases (cycle of shifts), necessitating practical iteration caps.

7. Applications and Extensions

SEE’s structurally-entropic core generalizes across domains:

Domain	Entropy Definition	SEE Role/Output
Network design/graph streams	Degree sequence SI or partition entropy	Real-time topology control, community tracking
Non-adiabatic molecular dynamics	Conditional entropy over descriptors	Automated mechanism extraction, key coordinate identification
Social event detection	Structural entropy on event graphs	Aligned, evolving topic clusters with evolution keywords

Ongoing and proposed extensions include hierarchical adjustment for deeper trees, adaptation to directed/weighted graphs, and hybrid approaches mixing offline and incremental recalibration for real-time monitoring and anomaly detection.

A plausible implication is that SEE will continue to expand into new areas—anywhere interpretable, incremental, and theoretically principled evolution control or monitoring is required across high-dimensional, dynamically structured data.