Lorenzo Grimaldi, Pablo Villegas, Alessandro Vezzani, Raffaella Burioni, Davide Cassi, Andrea Gabrielli (2025), Physical Review E 114, L022301 (2026) (Letter); first posted as arXiv preprint — ARXIV-2510.20520
Key takeaways
- In composite “bundled” networks (a copy of a fiber graph attached to every node of a base graph) the Fiedler (spectral-gap) dimension generically decouples from the spectral dimension, so networks with identical density-of-states exponents can have parametrically different equilibration times.
Seeded from the abstract and a skim, before the reading in NOTE-005. What follows is what the work says about itself.
Many real networks are assembled by reusing and recombining existing modules (“tinkering”), and it is an open problem how the properties of the parts determine the behaviour of the whole. The authors study networks of networks built by modular composition. They show that such compositions generically separate two notions of dimension that coincide in homogeneous structures: the spectral dimension (scaling of the Laplacian eigenvalue density) and the Fiedler dimension (scaling of the smallest nonzero Laplacian eigenvalue with size). Consequently, two networks can share the same thermodynamic exponents yet relax to equilibrium on very different timescales.
Standing in the record
Filed by the reading-time triage of 2026-09-25: 6,080 seconds of active reading over 7 sessions in the papers-feed tracker. It was filed Deferred, unread. NOTE-005 is the close reading of 2026-09-25, and it placed the work: Active — the base-controlled gap rule and its derivation are correct and easy to check, and they are worth reading as a clean, minimal example of the density-of-states exponent and the gap exponent coming apart. Read the SM finite-size scaling forms (Eqs. 25, 29) with care: their N prefactors are wrong except when d_s,b = d_g.
When seeded, before the reading in NOTE-005, the entry gave a priority for a deeper reading: medium — the owner has already spent the most time here (t=6080 s) and the 5-page letter is largely captured by this skim; the remaining value is verifying the SM derivation, which only matters if spectral network dimension is an active interest.
What the seeding asked a deeper reading to check (NOTE-005 is that reading):
- Strongest claim is the closed-form composition rule (Eq. 6); a deeper read should check the perturbative derivation and its stated validity range (SM S2-S3) and whether it holds beyond the base/fiber sizes scaling together (L_base = L_fiber is assumed in the examples).
- Generality claim (“independent of microscopic details” given finite Hausdorff dimension) is backed by a handful of synthetic families; no real-world network is analysed in the main text.
- ML link is indirect: graph Laplacian spectra, spectral gaps and diffusion timescales matter for spectral GNNs, graph diffusion and mixing times, but the paper itself makes no ML claim.
Access when seeded: arXiv abs page (v1 and v2) and full v2 PDF (13 pp. incl. Supplemental Material) read via arxiv.org; DOI and journal reference confirmed on the arXiv page and via Crossref (PRE, published online 2026-08-18). The list title “Unveiling the Dimensionality of Networks of Networks” is the v1 title (23 Oct 2025); v2 (31 Aug 2026) and the published PRE Letter use the title above, and the abstract was reworded between versions. export.arxiv.org API returned HTTP 406, so metadata came from the abs page. Supplemental sections listed but not read in detail.