arXiv paper2026-09-03· machine learning (causal discovery / federated learning)· method paper

Federated Causal Discovery via Regression-Directed Cumulants

Pablo Torrijos, Fabio Stella, José A. Gámez, José M. Puerta

cs.LG

Quick verdict

FedRCD federates LiNGAM via 4th-order cumulants; sound and honest, but it shows the whole cumulant family ranks by variance ladder, not causal asymmetry.

Summary

The paper studies LiNGAM causal discovery in federated settings where privacy regulation forbids pooling raw data. It first formalises why FedISHC, the existing federated cumulant method, fails under symmetric noise: its identification score and deflation coefficient both divide by the candidate source's third self-cumulant, which vanishes for symmetric distributions (Proposition 1), and swapping in a fourth-order score alone cannot help because the third-order deflation never updates fourth-order tensors. The authors therefore introduce FedRCD, three variants (NI, Def, X) pairing the fourth-order score with an OLS covariance-based deflation coefficient, trading communication rounds (one vs p−1) against algebraic noise; NI and X are exact federated counterparts of centralised HC-LiNGAM and HC, and single-round variants support exact unlearning at any granularity. On Erdős–Rényi and eight bnlearn DAGs across eleven noise families, FedRCD-X comes closest to centralised DirectLiNGAM (all-noise mean Ancestral Accuracy 0.981 at p=20, K=10, versus 0.802 for FedISHC). The key diagnostic (Section 4.5) shows the whole cumulant family — federated and centralised — actually ranks variables by a DAG-induced variance ladder (the cumulant analogue of varsortability): marginal standardisation drops all cumulant methods to near-random or below-chance levels (FedRCD-NI 0.940→0.303) while scale-invariant DirectLiNGAM is unaffected. Limitations: synthetic-only experiments, no plain variance-sorting baseline, and privacy of transmitted moments deferred to future work.

Review complete

Assessment at a glance

machine learning (causal discovery / federated learning)

risk flags identified

● high 0

● medium 3

● low 2

Reproducibility readinesshigh

Code & data availableyes

Review confidencehigh

AI text signal0/20 passages flagged

Deep research viableCode-reproducible

5 strengths · 5 weaknesses · 5 author questions

AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.

Levels reflect verification complexity — not a quality score. Evolve never rates your paper.

Key claims under review

FedISHC breaks under near-symmetric noise because its identification score and deflation coefficient both divide by the candidate source's third self-cumulant, giving Θ(κ3⁻²/N) variance (Proposition 1), and replacing only the score cannot repair it because the third-order deflation leaves fourth-order tensors invariant.
FedRCD-NI and FedRCD-X are exact federated counterparts of centralised HC-LiNGAM and HC respectively, working over horizontal, vertical and hybrid partitions, with the single-round variants supporting exact unlearning at any granularity.
At deployment-scale sample sizes the entire cumulant-based family — federated and centralised — ranks variables by a DAG-induced variance ladder (the cumulant counterpart of varsortability), collapsing to near-random or below-chance ordering under marginal standardisation while scale-invariant DirectLiNGAM is unaffected.
FedRCD-X reaches the ceiling of exact federation under cumulant aggregation, outperforming FedISHC under both symmetric and skewed noise (all-noise mean AncAcc 0.981 vs 0.802 at p=20, K=10).

Strengths

Clean diagnosis of the prior method's failure mode
Proposition 1 quantifies the divergence of FedISHC's deflation coefficient (Var = Θ(κ3⁻²/N) as κ3→0), and the FedISHC(τ4) ablation (Fig. 2a) is exemplary design: replacing only the score lifts AncAcc from 0.802 to 0.959 at p=20, K=10, and the paper correctly explains why this matches deflation-free FedRCD-NI (0.960) — the third-order deflation is inert on fourth-order tensors.
Exactness of the federated protocol is verifiable, not just claimed
Cumulant aggregation is lossless by additivity over independent groups (Speed 1983); Fig. 2c shows flat AncAcc curves over K∈{1,...,50}, and the text states FedRCD-NI/X produce the same orderings as HC-LiNGAM/HC given identical aggregates — a checkable internal consistency property.
The Section 4.5 stratification diagnostic is a standalone contribution
Practical GDPR unlearning at any granularity
Broad, fully specified experimental grid

Weaknesses

The obvious control baseline is missing
Synthetic-only evidence for a deployment-motivated method
FedRCD-Def is dominated and rests on population-only approximations
Statistical reporting is thin for the ordering claims
Privacy of the transmitted statistics is unanalysed

Full Review Report

$1.99 · one-time

Unlock the complete preliminary analysis Evolve produced for this paper.

Report preview

Scope & Claims

This PGM 2026 paper studies LiNGAM causal discovery when data are split across clients that may not pool raw observations (GDPR motivation, Section 1). Four claims: (1) FedISHC (Chen et al., 2026) fails under near-symmetric noise because both its identification score and deflation coefficient divide by κ3 of the candidate source (Proposition 1: Θ(κ3⁻²/N) variance); (2) swapping only the score for a fourth-order criterion cannot repair it, since the third-order deflation never touches fourth-order tensors; (3) the FedRCD-NI/Def/X family fixes both, with NI and X being exact federated counterparts of HC-LiNGAM and HC, and single-round variants supporting exact unlearning; (4) the candid headline: at deployment-scale N, the entire cumulant family — federated and centralised — ranks variables by a DAG-induced variance ladder, the cumulant analogue of varsortability (Reisach et al., 2021).

Methodology Assessment

The diagnosis is the paper's best technical work. Proposition 1 applies the delta method to κ̂2,1/κ̂3 informally but with the right scaling; Proposition 2 contrasts the OLS coefficient β̂js = Σ̂js/Σ̂ss, whose variance is O(N⁻¹) uniformly over noise laws. The FedISHC(τ4) ablation (Fig. 2a) is exemplary experimental design: replacing only the score lifts Ancestral Accuracy from 0.802 to 0.959 at p=20, K=10, and the paper correctly explains why this matches deflation-free FedRCD-NI (0.960) — the old deflation is inert on fourth-order statistics. Two soft spots: identifi

The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.

All 5 strengths & 5 weaknesses, analyzed in detail
5 risk flags with severity, section & reviewer questions
Methods & statistical evidence assessment
5 questions a peer reviewer would ask
Full six-dimension evaluation report & recommendations

One-time purchase · this paper only

One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.

Frequently asked questions

What is FedRCD and how does it differ from FedISHC?

FedRCD is a family of federated LiNGAM estimators in which clients send raw-moment sums once (or p−1 times for the X variant) and the server reconstructs cumulant tensors exactly. Unlike FedISHC — whose identification score and deflation coefficient both divide by the candidate source's third-order cumulant, a quantity that vanishes for symmetric noise — FedRCD pairs a fourth-order identification score with an ordinary-least-squares deflation coefficient whose denominator (the source variance) is always positive.

Does FedRCD support GDPR right-to-erasure (federated unlearning)?

Yes, for the single-round variants (FedRCD-NI and FedRCD-Def). Because raw moments are additive over samples, the server can subtract the contribution of any forgotten subset — a single individual, a cohort, or an entire client — and rescale, recovering exactly the statistics that would have existed had those samples never participated. FedRCD-X, which needs p−1 communication rounds, does not retain this property.

What is the paper's variance-stratification (varsortability) finding?

The authors show that at sample sizes in the thousands, the entire cumulant-based method family — including the centralised baselines HC and HC-LiNGAM, not just the new federated methods — ranks variables mainly by the variance ladder a DAG induces along its directed paths, not by the non-Gaussian asymmetry the theory targets. After standardising each variable to unit variance, these methods drop to near-random or below-chance ordering, while DirectLiNGAM, which is scale-invariant, is unaffected.

Is the code for FedRCD available?

Yes. Source code and experiment scripts are at github.com/ptorrijos99/FedLiNGAM, implemented in Python 3.13 with the Flower federated framework and the lingam library. All experiments are synthetic and were run on a laptop-class AMD Ryzen machine, so replication is feasible without special hardware.

About this review

This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.

Quantitative results live mainly in figures (Figs. 1, 3–5); beyond the printed labels of Fig. 2a and Table 2, exact per-regime values cannot be read from the extracted text.
The metadata abstract is cut mid-sentence, but the complete abstract appears at the start of the embedded paper text; the extraction ends cleanly at page 17 with references and appendices present, so no content truncation was detected.

Go deeper on this paper

Wrote a paper? Stress-test it before reviewers do.

Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.

Scan your paper