Saved in:
| Main Author: | |
|---|---|
| Format: | Recurso digital |
| Language: | |
| Published: |
Zenodo
2026
|
| Subjects: | |
| Online Access: | https://doi.org/10.5281/zenodo.20363911 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Table of Contents:
- <p>A vector database index built from millions of embeddings can be searched using only its edge structure-no coordinates required. Sub-Laplacian Spectral Hashing (SLSH) assigns compact binary identifiers to graph nodes by recursively decomposing local sub-Laplacians via their Fiedler eigenvectors, operating exclusively on topology after raw embedding vectors have been discarded. The scheme admits an O(1) candidate-pruning oracle at query time via a single XOR instruction over prefix bits. Three experimental regimes expose a sharp topological dichotomy: substantial accuracy gains as a positional encoding in community-structured graphs; complete signal collapse in scale-free networks, explained by the metric homogenization of preferential attachment; and sustained neighbor enrichment in simulated proximity manifolds. We further derive and empirically validate a Topological Stopping Criterion-the maximum hash depth beyond which spectral computation degrades to noise-expressed entirely in terms of graph degree statistics, and refine it into a deployment rule confirmed across four distinct graph configurations. </p>