Spectral Analysis of Fake News Propagation

Reza Zafarani; Weibin Cai

arxiv: 2605.13861 · v1 · pith:2U6U4M6Pnew · submitted 2026-04-18 · 💻 cs.SI · cs.AI

Spectral Analysis of Fake News Propagation

Weibin Cai , Reza Zafarani This is my paper

Pith reviewed 2026-05-15 06:34 UTC · model grok-4.3

classification 💻 cs.SI cs.AI

keywords fake news detectioninformation propagationgraph spectraspectral boundscascade analysisstructural optimizationsocial networks

0 comments

The pith

Spectral bounds on graph propagation patterns distinguish fake news cascades from real ones and enable classification plus structural interpretation.

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

The paper establishes a spectral framework that links graph spectra to the structural properties of how information spreads through networks. It derives several new mathematical bounds on these properties and combines them with prior bounds into one unified representation of propagation. This representation is then applied directly to classify fake versus real news and to optimize cascade structures for clearer interpretation of the patterns. A sympathetic reader would care because the method replaces ad hoc topological features with rigorous spectral connections that capture distinguishing traits. If the bounds hold, they supply both a detection tool and a way to trace how fake propagation evolves differently.

Core claim

We introduce several new bounds and integrate them with existing ones into a unified spectral representation of information propagation. We then use these spectral bounds for downstream classification and design a discrete structural optimization framework to interpret learned propagation patterns. For efficient optimization we rely on a first-order perturbation approximation and consider both score-guided and bound-guided objectives. Experiments on real-world data reveal meaningful spectral differences between fake and real news, competitive classification performance from spectral bounds, and interpretable evolution trajectories from structural optimization.

What carries the argument

unified spectral representation of information propagation formed by integrating new and existing graph-spectrum bounds on cascade structure

If this is right

Spectral bounds produce competitive accuracy in distinguishing fake from real news on real data.
Discrete structural optimization yields interpretable trajectories showing how propagation patterns evolve.
Real-world cascades exhibit consistent spectral differences between fake and real news.
First-order perturbation approximations enable tractable optimization over discrete graph structures.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same spectral bounds could be tested on other diffusion processes such as rumor spreading or product adoption.
Optimization trajectories might predict future structural changes in new cascades if the approximation remains stable.
The unified representation offers a route to compare propagation models from different domains under one spectral lens.

Load-bearing premise

The derived spectral bounds together with the first-order perturbation approximation faithfully capture the structural differences between real and fake news cascades without substantial loss of distinguishing information.

What would settle it

Running the spectral bounds as features in a classifier on held-out real-world cascades yields accuracy no better than standard topological baselines, or the optimized structures diverge from observed fake-news trajectories.

Figures

Figures reproduced from arXiv: 2605.13861 by Reza Zafarani, Weibin Cai.

**Figure 1.1.** Figure 1.1: The benefits of spectral analysis of news. Illustration of the designed structural optimization from the same initial news propagation tree (nodes are spreaders and edges are spreading patterns). Optimizing (after some steps) toward a more “fake-like" pattern (higher fake probability) produces a deeper and more imbalanced structure, while optimizing toward a more “real-like" pattern yields a shallower … view at source ↗

**Figure 5.1.** Figure 5.1: Mean z-score differences across node-count bins for three eigenvalues, λ1, µn−1, and νn−1, on the Weibo22 dataset. Red indicates that eigenvalues of fake news are larger than real news, while green indicates the opposite. Algorithm 5.3 Bound-guided Optimization Input: Initial tree G0, Max iterations T, Direction d ∈ {1, −1}, Convergence threshold τ Output: Evolution sequence H = {G0, G1, . . . , Gk} 1: D… view at source ↗

**Figure 5.** Figure 5: shows the distributions of [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗

**Figure 6.1.** Figure 6.1: Propagation graph evolution toward a more fake news like structure under score-guided optimization. Yellow, blue, and gray nodes denote the root, internal, and leaf nodes, respectively. The title of each graph shows the current iteration t and the classifier prediction score. Initial P(fake)=0.5245 Iter 1 P(fake)=0.5701 max-breadth=15.3653 Iter 2 P(fake)=0.5665 max-breadth=16.5932 Iter 3 P(fake)=0.5675 m… view at source ↗

**Figure 6.2.** Figure 6.2: Propagation graph evolution toward higher max-breadth under bounds-guided optimization. Yellow, blue, and gray nodes denote the root, internal, and leaf nodes, respectively. The title of each graph shows the current iteration t, the classifier prediction score, and bound value. close to their exact properties. This suggests that spectral bounds still capture highly informative structural signals for dis… view at source ↗

**Figure 6.** Figure 6: shows that the [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗

read the original abstract

The propagation structure of fake news has been shown to be an important cue for detecting it; yet, existing propagation-based fake news detection methods have mainly relied on ad hoc topological features, and a unified view of cascade patterns is still lacking. To address this, we study news propagation from a spectral view by connecting graph spectra to propagation-related structural properties through rigorous spectral bounds. In particular, we introduce several new bounds and integrate them with existing ones into a unified spectral representation of information propagation. We then use these spectral bounds for downstream classification and design a discrete structural optimization framework to interpret learned propagation patterns. For efficient optimization, we rely on a first-order perturbation approximation and consider both score-guided and bound-guided objectives. Experiments on real-world data reveal meaningful spectral differences between fake and real news, competitive classification performance from spectral bounds, and interpretable evolution trajectories from structural optimization. The findings demonstrate the value of spectral analysis for understanding and modeling news propagation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

New spectral bounds tied to propagation structure are the real addition here, but the first-order perturbation for structural optimization is unlikely to hold up on actual cascade edits.

read the letter

The core contribution is a set of new spectral bounds that link graph eigenvalues to properties like spread and connectivity in news cascades, then combine them with prior bounds into one representation. This moves past the usual collection of hand-picked topological features and gives a more principled handle on what distinguishes fake from real propagation. The classification experiments use these bounds directly and report competitive results on real data, which is a straightforward win if the bounds are tight. The optimization framework for interpreting patterns by suggesting edge changes is also new in this setting, and the authors show example trajectories that look plausible at first glance. That part earns credit for trying to make the spectral view actionable rather than just descriptive. The main soft spot is the reliance on first-order perturbation to approximate the effect of discrete structural edits. First-order spectral perturbation works when the change is small relative to the eigenvalue gap, but cascade graphs routinely see larger edits involving multiple edges or high-degree nodes. The abstract gives no error analysis, no comparison to exact recomputation on held-out graphs, and no regime where the approximation stays accurate, so the interpretability claims rest on an assumption that probably does not hold for typical data. This is not fatal to the bounds or the classification results, but it does limit what the optimization section can reliably deliver. The paper is aimed at people working on graph-based misinformation detection who want something more unified than feature engineering. A reader who cares about spectral methods applied to social graphs will find the bounds and the empirical differences worth looking at. It deserves a serious referee because the central idea is grounded in graph theory, the experiments are competitive, and the approximation issue is fixable with additional checks rather than a load-bearing flaw in the whole argument. I would send it out for review.

Referee Report

1 major / 1 minor

Summary. The paper claims to advance fake news detection by shifting from ad hoc topological features to a unified spectral representation of propagation cascades. It derives several new rigorous spectral bounds on graph properties, integrates them with existing bounds, applies the resulting spectral features to downstream classification, and introduces a discrete structural optimization framework (using first-order perturbation approximations and both score-guided and bound-guided objectives) to recover interpretable evolution trajectories of cascades. Experiments on real-world data are reported to show meaningful spectral differences between fake and real news, competitive classification performance, and interpretable structural changes.

Significance. If the new bounds are rigorously derived without hidden parameters and the perturbation-based optimization is shown to be accurate in the relevant regime, the work would provide a principled, graph-theoretic foundation for analyzing cascade structure that could improve both detection accuracy and interpretability over existing ad-hoc methods.

major comments (1)

[discrete structural optimization framework (as described in abstract)] The discrete structural optimization framework relies on a first-order perturbation approximation to model discrete edge additions/deletions. This approximation is accurate only when the perturbation norm is small relative to the eigenvalue gap; however, the abstract provides no explicit error bounds, regime validation, or comparison against exact recomputation for the (potentially large) edits that occur in real cascades. This assumption is load-bearing for the interpretability claims.

minor comments (1)

[Abstract] The abstract asserts 'rigorous new bounds' and 'competitive classification performance' but does not name the datasets, report quantitative metrics (e.g., F1, AUC), or list the exact existing bounds being unified; adding these details would strengthen the summary.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address the major comment on the discrete structural optimization framework below and describe the planned revisions.

read point-by-point responses

Referee: [discrete structural optimization framework (as described in abstract)] The discrete structural optimization framework relies on a first-order perturbation approximation to model discrete edge additions/deletions. This approximation is accurate only when the perturbation norm is small relative to the eigenvalue gap; however, the abstract provides no explicit error bounds, regime validation, or comparison against exact recomputation for the (potentially large) edits that occur in real cascades. This assumption is load-bearing for the interpretability claims.

Authors: We agree that the first-order perturbation approximation is accurate primarily when the perturbation norm is small relative to the eigenvalue gap, and that explicit error bounds, regime validation, and comparisons to exact recomputation are needed to support the interpretability claims for real cascades. The manuscript discusses the approximation in the methods section for modeling small structural changes and reports experimental results on real-world data, but does not include formal error bounds or systematic comparisons across perturbation sizes. In the revised version, we will add a dedicated subsection deriving explicit error bounds using standard perturbation theory results (such as Weyl's inequalities and Davis-Kahan sin-theta bounds) and include new experiments comparing the approximate trajectories to exact recomputation for varying numbers of edge edits. These additions will clarify the regime of applicability and strengthen the claims regarding interpretable evolution trajectories. revision: yes

Circularity Check

0 steps flagged

No significant circularity; spectral bounds derived independently and applied downstream

full rationale

The paper derives new spectral bounds from graph spectra and propagation properties, integrates them with existing bounds from the literature into a unified representation, then applies the result to classification and a perturbation-based optimization framework. No step reduces by construction to its own inputs: the bounds are presented as rigorous derivations rather than fitted parameters or self-definitions, the first-order perturbation is an explicit efficiency assumption (not a renaming or load-bearing self-citation), and no uniqueness theorem or ansatz is smuggled via self-reference. The derivation chain remains self-contained against external graph-theoretic benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only access provides no explicit free parameters, axioms, or invented entities; full manuscript required for complete ledger.

pith-pipeline@v0.9.0 · 5449 in / 1161 out tokens · 43050 ms · 2026-05-15T06:34:15.919427+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages

[1]

maximum of out degree

work page
[2]

depth of the maximum of out degree node

work page
[3]

width entropy: P i p(wi) log(p(wi)),p(w i) = wiP i wi , wherew i is the width of depthi

work page
[4]

Category Meaning / Property Bound Equation C1

average of leaf depth; Copyright©2026 by SIAM Unauthorized reproduction of this article is prohibited Table C.2: Spectral bounds categorized by propagation-related structural properties. Category Meaning / Property Bound Equation C1. Branching Capacity Max degree / Mean degree ¯d≤λ 1 ≤∆[4] Maximum sum of degrees of adjacent nodesµ 1 ≤max x∼y(dx +d y)[4] S...

work page 2026
[5]

mean branching: P i∈I(T) di |I(T)| , where|I(T)|is the number of internal nodes,di is the degree of nodei

work page
[6]

variance of branching

work page
[7]

sackin index:S(T) = P v∈L(T) d(v), i.e., the sum of the depths of all leaf nodes, whereT is a tree,L(T)is a leaf node set ofT,d(v)is the depth of nodev

work page
[8]

colless index:C(T) = P v∈I(T) |Lv −R v|, whereL v andR v represent number of leaf nodes inv’s left subtree and right subtree

work page
[9]

number of internal node

work page
[10]

A larger value indicates a more un- even concentration of node degrees

degree gini:G d = Pn i=1 Pn j=1 |di−dj | 2n Pn i=1 di , which quantifies the inequality of the degree distri- bution. A larger value indicates a more un- even concentration of node degrees

work page
[11]

chromatic number: it is a constant number2 for trees

work page
[12]

approximate independent number

work page
[13]

maximum of degrees of a pair of connected nodes

work page
[14]

sum of top 30% degree

work page
[15]

sum of top 60% degree; 29.e+ t+1 2 , wheret= 0.3∗ ⌊n⌋, which is an upper bound for sum of part laplacian eigenvalues (check bounds of number of edges in Table C.2); 30.e+ t+1 2 , wheret= 0.6∗ ⌊n⌋, which is an upper bound for sum of part laplacian eigenvalues (check bounds of number of edges in Table C.2)

work page
[16]

bandwidth:bw(L) = max Lij ̸=0 |i−j|+ 1, whereLis the laplacian matrix of graph, which measures how far the nonzero entries of its matrix representation spread away from the main diagonal under a given node order- ing

work page
[17]

wiener index: the sum of the shortest- path distances between all unordered pairs of nodes

work page
[18]

number of leaf nodes

work page
[19]

number of spanning tree: its a constant num- ber1for trees; D More Structural Optimization Results. In addition to the two cases presented in the main text, we provide more examples of the proposed discrete structural optimization algorithms to further illustrate thedifferencesbetweenfakeandrealnewsdissemination patterns. Copyright©2026 by SIAM Unauthoriz...

work page 2026

[1] [1]

maximum of out degree

work page

[2] [2]

depth of the maximum of out degree node

work page

[3] [3]

width entropy: P i p(wi) log(p(wi)),p(w i) = wiP i wi , wherew i is the width of depthi

work page

[4] [4]

Category Meaning / Property Bound Equation C1

average of leaf depth; Copyright©2026 by SIAM Unauthorized reproduction of this article is prohibited Table C.2: Spectral bounds categorized by propagation-related structural properties. Category Meaning / Property Bound Equation C1. Branching Capacity Max degree / Mean degree ¯d≤λ 1 ≤∆[4] Maximum sum of degrees of adjacent nodesµ 1 ≤max x∼y(dx +d y)[4] S...

work page 2026

[5] [5]

mean branching: P i∈I(T) di |I(T)| , where|I(T)|is the number of internal nodes,di is the degree of nodei

work page

[6] [6]

variance of branching

work page

[7] [7]

sackin index:S(T) = P v∈L(T) d(v), i.e., the sum of the depths of all leaf nodes, whereT is a tree,L(T)is a leaf node set ofT,d(v)is the depth of nodev

work page

[8] [8]

colless index:C(T) = P v∈I(T) |Lv −R v|, whereL v andR v represent number of leaf nodes inv’s left subtree and right subtree

work page

[9] [9]

number of internal node

work page

[10] [10]

A larger value indicates a more un- even concentration of node degrees

degree gini:G d = Pn i=1 Pn j=1 |di−dj | 2n Pn i=1 di , which quantifies the inequality of the degree distri- bution. A larger value indicates a more un- even concentration of node degrees

work page

[11] [11]

chromatic number: it is a constant number2 for trees

work page

[12] [12]

approximate independent number

work page

[13] [13]

maximum of degrees of a pair of connected nodes

work page

[14] [14]

sum of top 30% degree

work page

[15] [15]

sum of top 60% degree; 29.e+ t+1 2 , wheret= 0.3∗ ⌊n⌋, which is an upper bound for sum of part laplacian eigenvalues (check bounds of number of edges in Table C.2); 30.e+ t+1 2 , wheret= 0.6∗ ⌊n⌋, which is an upper bound for sum of part laplacian eigenvalues (check bounds of number of edges in Table C.2)

work page

[16] [16]

bandwidth:bw(L) = max Lij ̸=0 |i−j|+ 1, whereLis the laplacian matrix of graph, which measures how far the nonzero entries of its matrix representation spread away from the main diagonal under a given node order- ing

work page

[17] [17]

wiener index: the sum of the shortest- path distances between all unordered pairs of nodes

work page

[18] [18]

number of leaf nodes

work page

[19] [19]

number of spanning tree: its a constant num- ber1for trees; D More Structural Optimization Results. In addition to the two cases presented in the main text, we provide more examples of the proposed discrete structural optimization algorithms to further illustrate thedifferencesbetweenfakeandrealnewsdissemination patterns. Copyright©2026 by SIAM Unauthoriz...

work page 2026