Spectral Coarse-Graining and Rescaling for Preserving Structural and Dynamical Properties in Graphs

F. Caccioli; M. Schmidt; T. Aste

REVIEW 3 major objections 4 minor 29 references

Spectral Coarse-Graining and Rescaling for Preserving Structural and Dynamical Properties in Graphs

T0 review · 3 major / 4 minor · reviewed 2026-08-12 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A coarse-grained Laplacian rescaled by the spectral gap generates smaller graphs that preserve diffusion dynamics and large-scale topology.

desk verdict A clearly specified spectral Laplacian coarse-graining scheme whose central preservation claim outruns its proof — worth refereeing, but only after the authors close the gap between the truncation bound and the full contraction/rescaling map. read the letter →

arxiv 2411.11991 v1 pith:4XIO7WQK submitted 2024-11-18 cond-mat.stat-mech cond-mat.dis-nnphysics.bio-phphysics.data-an

M. Schmidt , F. Caccioli , T. Aste This is my paper · ORCID

classification cond-mat.stat-mechcond-mat.dis-nnphysics.bio-phphysics.data-an

keywords graphrenormalizationcoarse-grainedLaplacianspectralgapdiffusiondynamicsheatkernelEEGbrainnetworksscaleinvariance

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces a graph-renormalization procedure that, for each characteristic scale identified by a spectral gap in the graph Laplacian, produces a smaller graph with fewer vertices. Its central claim is that this renormalized graph reproduces the original diffusion probabilities approximately, up to an error that decays at least as $e^{-t\lambda_{k+1}}$, while preserving the zoomed-out topology. The motivation is that many real networks, such as the brain, are not organized by geometric closeness, so a topological renormalization based on diffusion is needed. Applied to EEG-derived functional graphs of human brain activity, the method reports collective coordinated clusters, and it finds that attention states show more specialized occipital activity and stronger scale invariance than rest states.

What carries the argument

The central mechanism is the coarse-grained Laplacian, the projection $L^{(1)}=\sum_{\alpha:\lambda_\alpha\le\lambda_k}\lambda_\alpha u_\alpha u_\alpha^\top$ onto the slowest eigenmodes of the graph Laplacian. Its entries act as a weighted cosine similarity between vertices within the retained modes, and the negative entries of $A^{(1)}=\mathrm{diag}(L^{(1)})-L^{(1)}$ mark vertices that should be contracted. The rescaling $A_R=(1/\lambda_k)A^{(3)}$ restores the original resolution after contraction, analogous to restoring the lattice spacing in renormalization-group schemes. The error control is the spectral-truncation bound $\epsilon(t)\le e^{-t\lambda_{k+1}}\epsilon(0)$, which quantifies how quickly neglected fast modes disappear.

What would settle it

Take a graph with a known spectral gap, run the full renormalization to obtain $G_R$, and compare the true diffusion probabilities $e^{-tL_R}p(0)$ of the renormalized graph with $e^{-tL}p(0)$ of the original. The paper's preservation claim predicts the mismatch decays at rate at least $\lambda_{k+1}$; if a graph can be found where the mismatch does not decay at that rate, or where it stays large even for $t>1/\lambda_{k+1}$, the claim would be falsified.

Watch

Extended reading notes

Core claim

The core discovery is that a projection of the graph Laplacian onto its $k$ slowest eigenmodes can serve as the generator of a renormalized graph. Writing $L^{(1)}=\sum_{\alpha:\lambda_\alpha\le\lambda_k}\lambda_\alpha u_\alpha u_\alpha^\top$, the entries of $L^{(1)}$ are similarity values between vertices from a zoomed-out perspective. The paper defines $A^{(1)}=\mathrm{diag}(L^{(1)})-L^{(1)}$; negative entries identify pairs of similar vertices that are contracted into effective vertices, and the contracted adjacency matrix is rescaled by $1/\lambda_k$ to restore the original resolution. The resulting $G_R$ is claimed to keep both the diffusion dynamics and the skeleton of $G$: errors from neglecting fast modes are bounded by $e^{-t\lambda_{k+1}}\epsilon(0)$, so a large spectral gap and a large $\lambda_{k+1}$ make the approximation better. The paper tests this on two synthetic graphs and on EEG-based functional brain graphs, where renormalized graphs reveal mesoscale clusters and scale-dependent reorganization between rest and attention.

Load-bearing premise

The paper proves only that fast eigenmodes can be safely ignored in the original spectral expansion; it assumes, without proof, that the subsequent steps of reassigning weights, contracting negative-edge vertices, and rescaling by $1/\lambda_k$ keep the renormalized graph's diffusion close to that truncated dynamics.

Editorial extensions

If this is right

For each spectral gap, one obtains a smaller graph whose diffusion probabilities match the original to within an error that decays at rate at least $\lambda_{k+1}$.
Because the method is not based on geometric closeness, it applies to networks where interactions are not distance-driven, such as functional brain connectivity.
The renormalized graphs group vertices into effective vertices that absorb fast-diffusion regions, exposing mesoscale structures that single-scale methods miss.
In the EEG application, attention states maintain diffusion dynamics better than rest states and yield more compact occipital clusters, supporting a scale-invariance interpretation of task-focused brain activity.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A natural extension is to seek a rigorous bound on the full heat-kernel distance between the original and renormalized graphs in terms of $\lambda_{k+1}$ and the contraction choices; the paper provides only a truncation bound.
Iterating the procedure on its own outputs would define a renormalization-group flow whose fixed points, if they exist, would characterize self-similar graph families; the paper does not investigate this flow.
The EEG results suggest a quantitative index of scale invariance: the ratio of diffusion-probability mismatch before and after renormalization could serve as a data-driven measure of how well slow modes describe brain dynamics.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: this is a genuine variant on Laplacian renormalization, with a clearly specified algorithm, but the paper's central claim — that the renormalized graph preserves diffusion dynamics — is not established by the math it provides. Eq. 7 bounds only the error from dropping fast eigenmodes; it says nothing about the later steps (thresholding A(1), contracting negative-weight vertices, rescaling by 1/λ_k). Those are nonlinear operations on the graph, and no theorem connects their output L_R back to the truncated Laplacian L(1). The examples in Fig. 2 are two small hand-built graphs, not enough to validate a general renormalization scheme.

What's genuinely new: while Villegas et al. use the heat kernel, this paper projects the graph Laplacian directly and reads vertex similarity from L(1) entries. The contraction rule — merging vertices connected by negative coarse-grained weights — is a clear, deterministic prescription. The rescaling step is a natural RG analogue. The method is self-contained: the scale is read from the spectral gap, not fitted to conclusions. That part is fine. The Louvain comparison on graph (b) is a nice sanity check, though anecdotal.

Where it's soft: beyond the missing error bound, Eq. 5 discards all similarity entries between vertices not already connected in G. But L(1) is dense; those off-diagonal entries code long-range diffusive coupling. Throwing them away could matter exactly when the graph has no geometric embedding. The paper doesn't address that. The brain application is qualitative: no confidence intervals, no null model, and the claim that attention states are more scale-invariant rests on visual inspection of two states. That's fine as an illustration, but it doesn't independently support the preservation claim.

The paper is honest about the heuristic spectral-gap choice and doesn't oversell the synthetic tests. The citation pattern looks appropriate, including the TMFG self-citation. Overall: a promising method with a real gap between claim and proof. A serious referee could push the authors to either prove an error bound for the full renormalization map or validate on a broader class of graphs (e.g., random graphs with known community structure and a range of spectral gaps). I'd send it out — the method deserves scrutiny, and the gap is fixable, not fatal.

Referee Report

3 major / 4 minor

Summary. The manuscript proposes a graph renormalization procedure based on truncating the graph Laplacian to its slowest eigenmodes (Eq. 3), using the resulting coarse-grained Laplacian to define similarity weights, thresholding those weights to existing edges (Eq. 5), contracting vertices connected by negative effective weights, and rescaling by 1/λ_k (Eq. 6). The authors claim that this procedure approximately preserves the diffusion dynamics of the original graph while reducing the number of vertices and retaining large-scale topological structure. They illustrate the method on two synthetic graphs and apply it to TMFG graphs derived from EEG recordings, reporting that attention states show more specialized and scale-invariant brain activity than resting states.

Significance. The proposed method is self-contained and does not fit free parameters to the conclusions: the spectral gap is read off from the graph spectrum, and the central quantity λ_k is a property of the input graph, not a tuned target. The idea of using the coarse-grained Laplacian rather than the heat kernel to derive effective interactions is a reasonable direction and could be useful for reducing large graphs while keeping some dynamical information. However, the central claim that the renormalized graph G_R approximately preserves the original diffusion dynamics is not supported by a proof: the only quantitative error bound (Eq. 7) concerns spectral truncation, not the subsequent contraction and rescaling steps. The validation on two small synthetic graphs is too limited to establish the method's general correctness, and the brain application is qualitative. The paper therefore presents a plausible and potentially useful heuristic, but the load-bearing preservation claim needs substantially stronger support.

major comments (3)

[Renormalization, Eq. (7) and surrounding text] The error bound in Eq. (7) applies only to the difference between the full heat kernel e^{-tL} and the truncated spectral projection e^{-tL(1)}, where L(1) is the coarse-grained Laplacian of Eq. (3). The actual renormalized graph G_R is produced by additional operations that are not covered by this bound: thresholding the dense matrix A(1) via Eq. (5), contracting vertices connected by negative entries, and rescaling all weights by 1/λ_k via Eq. (6). These operations are nonlinear functions of the graph, and no theorem, identity, or error estimate connects the heat kernel of G_R (or its Laplacian) to either L(1) or L. The statement immediately after Eq. (6), that 'the diffusion dynamics of the original system are approximately preserved in the renormalized system G_R', is therefore unsupported by the derivations in the manuscript.
[Renormalization, Eq. (5) and contraction step] Even if one accepted that L(1) encodes the slow-mode dynamics, the step in Eq. (5) discards all similarity information between vertices that are not already connected in the original graph A. Since L(1) is generically dense, these discarded entries can represent long-range diffusive couplings that affect the heat kernel at finite times. The manuscript gives no argument that this thresholding preserves the truncated dynamics, and the subsequent contraction rule is not fully specified: when a vertex is involved in multiple negative-edge pairs, or when two contracted vertices share neighbors, the resulting effective weights and possible parallel edges or self-loops are not defined. A precise algorithmic specification and a bound on the error introduced by these steps are needed before the preservation claim can be evaluated.
[Fig. 4 and the brain application] The claim that 'diffusion dynamics are better preserved for attention states compared to rest states' (p. 6) is supported only by a qualitative visual comparison in Fig. 4 and by the observation that λ_{k+1} is smaller for rest states. No quantitative error metric, such as the actual ϵ(t) of Eq. (7) computed for G_R versus G, is reported, and there is no confidence interval or null model to assess whether the difference between rest and attention is meaningful. As a result, the brain application does not independently validate the central preservation claim and should be presented as an illustrative observation until the method's accuracy is established.

minor comments (4)

[Introduction, second paragraph] The phrase 'the spectral gap δ = |λ_{k+1} − λ_k|, defined as the largest difference between the consecutive smallest eigenvalues' is ambiguous: if δ is defined as the largest gap, then the index k should be chosen accordingly, but the text later treats λ_k as a freely chosen threshold. Clarify whether k is selected by the largest gap or by another criterion.
[Fig. 1 caption] The caption contains an apparent typo: it says 'positive weights (black line) indicate dissimilar vertices connected in the original graph' but likely should read 'negative weights' for consistency with the main text. Please correct.
[References] Reference [18] duplicates reference [16]; the two entries for von Luxburg 2007 should be merged, and the in-text citation for the spectral clustering perspective should point to a single entry.
[Conclusion] The conclusion states that 'a simpler, and more universal representation of neural dynamics can be found via renormalization,' but the manuscript does not test universality or criticality directly; this sentence overstates the evidence. Consider softening it to reflect that the results are consistent with, but do not establish, scale invariance.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the renormalization construction is self-contained, and the unproved contraction/rescaling step is a correctness concern rather than a circular reduction.

full rationale

The central construction is not fitted to its conclusions. The coarse-grained Laplacian L(1) in Eq. (3) is the spectral projection of L onto the k slowest modes, and Eq. (7) is a standard bound on the truncation error of the heat-kernel expansion; this part is a genuine derivation. The spectral gap and λ_k are read off from the graph rather than chosen to match target diffusion curves. The subsequent thresholding (Eq. 5), contraction of negative-edge vertices, and rescaling by 1/λ_k (Eq. 6) are heuristic operations that are not covered by Eq. (7), so the claim that G_R preserves diffusion dynamics is not fully proven; however, this is a correctness gap, not circularity, because nothing in those steps is defined in terms of the quantity being predicted. The TMFG citation [27] is a self-citation (Aste is a coauthor), but it is used only to construct the EEG graphs for the application and does not bear the weight of the renormalization claim. Brain-state interpretations are post hoc and do not feed back into the algorithm. No step reduces by definition or by fitted input to its own conclusion.

Assumptions & free parameters 1 free parameters · 4 assumptions · 1 invented entities

The method's central claim rests on standard spectral theory plus two domain assumptions about ergodicity and the meaning of Laplacian eigenvectors. The most fragile element is the ad hoc contraction criterion, which is introduced without validation. One free parameter, the spectral threshold λ_k, controls how many modes are kept and thus the coarse-graining scale.

free parameters (1)

Spectral threshold λ_k (number of slow modes kept) = chosen per graph; e.g., λ_k=0.07 for graph (a), 0.119 for graph (b); two smallest λ_k for EEG graphs
The method requires selecting which eigenmodes to keep. The paper uses the largest gap between consecutive small eigenvalues as a heuristic, and for the brain data it renormalizes at the two smallest λ_k. This selection is a free parameter affecting the output.

assumptions (4)

standard math The graph Laplacian L is symmetric and admits an orthonormal eigenbasis, with non-negative eigenvalues.
Invoked in the spectral decomposition section; standard spectral theory for undirected weighted graphs.
domain assumption The graph must be ergodic for meaningful time-scale separation.
The authors state 'the graph must be ergodic; otherwise, it may have oscillating modes that do not decay with time.' This restricts applicability to graphs with a non-degenerate zero eigenvalue.
domain assumption Eigenvectors of the Laplacian represent graph partitions at different scales, with smaller eigenvalues giving coarser partitions.
This is the basis for interpreting u_1 as the Fiedler vector and for treating slow modes as large-scale structure. It is a standard but non-trivial assumption in spectral clustering.
ad hoc to paper Entries of the coarse-grained Laplacian L(1) can be interpreted as similarities; negative entries in A(1) indicate vertices that can be contracted.
The paper states 'A positive entry in L(1) or a negative entry in A(1) indicates that vertices are similar in G at a zoomed-out perspective and can be contracted.' This is a modeling assumption specific to this method and is not proven.

invented entities (1)

Effective vertices (supernodes)
purpose: Aggregate clusters of original vertices that have negative coarse-grained weights, reducing graph size.
Effective vertices are constructs of the method, not independently observable entities. They have no falsifiable handle outside the algorithm.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Spectral Coarse-Graining and Rescaling for Preserving Structural and Dynamical Properties in Graphs." pith.science (2026). https://pith.science/paper/4XIO7WQK

@misc{pith2026241111991,
  author       = {Pith},
  title        = {Pith review of: Spectral Coarse-Graining and Rescaling for Preserving Structural and Dynamical Properties in Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4XIO7WQK}},
  note         = {Machine review of arXiv:2411.11991}
}

read the original abstract

We introduce a graph renormalization procedure based on the coarse-grained Laplacian, which generates reduced-complexity representations for characteristic scales identified through the spectral gap. This method retains both diffusion probabilities and large-scale topological structures, while reducing redundant information, facilitating the analysis of large graphs by decreasing the number of vertices. Applied to graphs derived from EEG recordings of human brain activity, our approach reveals macroscopic properties emerging from neuronal interactions, such as collective behavior in the form of coordinated neuronal activity. Additionally, it shows dynamic reorganization of brain activity across scales, with more generalized patterns during rest and more specialized and scale-invariant activity in the occipital lobe during attention-focused tasks.

Figures

Figures reproduced from arXiv: 2411.11991 by the authors.

**Figure 2.** FIG. 2: Diffusion dynamics. The probabilities [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** FIG. 3: Spectral gap and renormalization of graph (b). A [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 5.** Figure 5: FIG. 5: Spectral gap and renormalization of the brain. For each [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 23 canonical work pages

[1]

K. G. Wilson and J. Kogut, Physics Reports 12, 75 (1974)

1974
[2]

Villegas, T

P. Villegas, T. Gili, G. Caldarelli, and A. Gabrielli, Na- ture Physics 19, 445 (2023)

work page 2023
[3]

Chung, Spectral Graph Theory, CBMS Regional Con- ference Series No

F. Chung, Spectral Graph Theory, CBMS Regional Con- ference Series No. 92 (Conference Board of the Mathe- matical Sciences)

work page
[4]

Lambiotte and M

R. Lambiotte and M. T. Schaub, Modularity and Dynam- ics on Complex Networks, Elements in the Structure and Dynamics of Complex Networks (Cambridge University Press, 2022)

work page 2022
[5]

Strogatz, Nonlinear Dynamics and Chaos: With Appli- cations to Physics, Biology, Chemistry and Engineering, Studies in nonlinearity (Westview, 2000)

S. Strogatz, Nonlinear Dynamics and Chaos: With Appli- cations to Physics, Biology, Chemistry and Engineering, Studies in nonlinearity (Westview, 2000)

work page 2000
[6]

L. P. Kadanoff, Physics Physique Fizika 2, 263 (1966)

1966
[7]

K. G. Wilson, Rev. Mod. Phys. 47, 773 (1975)

1975
[8]

K. G. Wilson, Scientific American 241, 158 (1979)

work page 1979

Show all 29 references

[9]

Ambjørn, J

J. Ambjørn, J. Jurkiewicz, and R. Loll, Phys. Rev. Lett. 95, 171301 (2005)

2005
[10]

Garc ´ ıa-P´ erez, M

G. Garc ´ ıa-P´ erez, M. Bogu˜ n´ a, and M.´A. Serrano, Nature Physics 14, 583 (2018)

2018
[11]

Calcagni, D

G. Calcagni, D. Oriti, and J. Th¨ urigen, Classical and Quantum Gravity 31, 135014 (2014)

2014
[12]

Villegas, A

P. Villegas, A. Gabrielli, F. Santucci, G. Caldarelli, and T. Gili, Phys. Rev. Res. 4, 033196 (2022)

2022
[13]

Masuda, M

N. Masuda, M. A. Porter, and R. Lambiotte, Physics Reports 716-717, 1 (2017)

2017
[14]

Cimini, T

G. Cimini, T. Squartini, F. Saracco, D. Garlaschelli, A. Gabrielli, and G. Caldarelli, Nature Reviews Physics 1, 58 (2019)

2019
[15]

Stewart and J

G. Stewart and J. Sun, Matrix Perturbation Theory, Computer Science and Scientific Computing (Elsevier Science, 1990)

1990
[17]

A. Ng, M. Jordan, and Y. Weiss, in Advances in Neu- ral Information Processing Systems, Vol. 14 (MIT Press, 2001)

2001
[18]

von Luxburg, Statistics and Computing 17, 395 (2007)

U. von Luxburg, Statistics and Computing 17, 395 (2007)

2007
[19]

De Domenico and J

M. De Domenico and J. Biamonte, Phys. Rev. X 6, 041062 (2016)

2016
[20]

Chung, Proceedings of the National Academy of Sci- ences 104, 19735 (2007)

F. Chung, Proceedings of the National Academy of Sci- ences 104, 19735 (2007)

2007
[21]

Fortunato and M

S. Fortunato and M. Barth´ elemy, Proceedings of the Na- tional Academy of Sciences 104, 36 (2007)

2007
[22]

Mora and W

T. Mora and W. Bialek, Journal of Statistical Physics 144, 268 (2011)

2011
[23]

P. Bak, C. Tang, and K. Wiesenfeld, Phys. Rev. Lett. 59, 381 (1987)

1987
[24]

Friedman, S

N. Friedman, S. Ito, B. A. W. Brinkman, M. Shimono, R. E. L. DeVille, K. A. Dahmen, J. M. Beggs, and T. C. Butler, Phys. Rev. Lett. 108, 208102 (2012)

2012
[25]

J. M. Beggs and D. Plenz, Journal of Neuroscience 23, 11167 (2003)

2003
[26]

Ku´ smierz, S

L. Ku´ smierz, S. Ogawa, and T. Toyoizumi, Phys. Rev. Lett. 125, 028101 (2020)

2020
[27]

G. P. Massara, T. Di Matteo, and T. Aste, Journal of Complex Networks 5, 161 (2016)

2016
[28]

A. T. Gifford, K. Dwivedi, G. Roig, and R. M. Cichy, NeuroImage 264, 119754 (2022)

2022
[29]

E. T. Bullmore and O. Sporns, Nature Reviews Neuro- science 10, 186 (2009)

2009
[30]

W. L. Shew, H. Yang, S. Yu, R. Roy, and D. Plenz, The Journal of Neuroscience 31, 55 (2010)

2010

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.

[1] [1]

K. G. Wilson and J. Kogut, Physics Reports 12, 75 (1974)

1974

[2] [2]

Villegas, T

P. Villegas, T. Gili, G. Caldarelli, and A. Gabrielli, Na- ture Physics 19, 445 (2023)

work page 2023

[3] [3]

Chung, Spectral Graph Theory, CBMS Regional Con- ference Series No

F. Chung, Spectral Graph Theory, CBMS Regional Con- ference Series No. 92 (Conference Board of the Mathe- matical Sciences)

work page

[4] [4]

Lambiotte and M

R. Lambiotte and M. T. Schaub, Modularity and Dynam- ics on Complex Networks, Elements in the Structure and Dynamics of Complex Networks (Cambridge University Press, 2022)

work page 2022

[5] [5]

Strogatz, Nonlinear Dynamics and Chaos: With Appli- cations to Physics, Biology, Chemistry and Engineering, Studies in nonlinearity (Westview, 2000)

S. Strogatz, Nonlinear Dynamics and Chaos: With Appli- cations to Physics, Biology, Chemistry and Engineering, Studies in nonlinearity (Westview, 2000)

work page 2000

[6] [6]

L. P. Kadanoff, Physics Physique Fizika 2, 263 (1966)

1966

[7] [7]

K. G. Wilson, Rev. Mod. Phys. 47, 773 (1975)

1975

[8] [8]

K. G. Wilson, Scientific American 241, 158 (1979)

work page 1979