Pith. sign in

REVIEW 3 major objections 8 minor 59 references

Synergy can hide behind redundancy in complex systems

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 15:06 UTC pith:OT535M5M

load-bearing objection Clean framework for separating dependency-driven synergy from genuine non-additive mechanisms, with one live confound in the empirical applications the 3 major comments →

arxiv 2607.07286 v1 pith:OT535M5M submitted 2026-07-08 stat.ME

From Statistical to Structural Synergy: A Predictability Framework to Quantify the Effects due to High-Order Mechanisms

Yuri Antonacci , Chiara Bar\`a , Laura Sparacino , Daniele Marinazzo , Luca Faes , Sebastiano Stramaglia This is my paper
classification stat.ME
keywords hobsstructuralsynergyhigh-orderdynamicsframeworkhomsmechanism
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper introduces a quantity called structural synergy, defined as the gap between the best possible prediction of a target variable from two source variables and the best prediction achievable when those sources are constrained to contribute additively. The authors argue that this gap separates two distinct phenomena that are routinely conflated: high-order behaviours (HOBs), which are statistical patterns of synergy or redundancy visible in observed data, and high-order mechanisms (HOMs), which are genuine non-additive interaction rules in the data-generating process. The central claim is that these two levels can dissociate. A system can display redundancy-dominated HOBs — where the two sources appear to provide overlapping, redundant information about the target — while simultaneously containing a significant non-additive mechanism that is only detectable through structural synergy. The authors call this the masked structural synergy regime. They prove it analytically for a multiplicative model Y = cX₁X₂ + U, where structural synergy is strictly positive for all source correlations below one, even when the observed synergy-redundancy balance becomes negative (redundancy-dominated) as source correlation increases. They validate the framework in stochastic autoregressive simulations and demonstrate it in climate indices (ENSO) and source-reconstructed EEG motor data, both of which show redundancy-dominated HOBs coexisting with statistically significant structural synergy.

Core claim

The paper's core result is the decomposition of observed high-order predictability into a dependency-driven component and a mechanism-driven component, operationalized through the comparison of two prediction errors: R* (the best achievable mean squared prediction error using both sources jointly and without restrictions) and R*_A (the best achievable error when the prediction is constrained to be additive in the two sources). Their difference, structural synergy Ss = R*_A − R*, is zero when the sources contribute additively to the target and positive when a non-additive combination of the sources is required. The analytical proof that Ss > 0 for all |r₁₂| < 1 in the multiplicative model Y =

What carries the argument

The framework operates on triplets of variables (two sources, one target). Four regression models are fit: single-source polynomial regressions for each source, a full bivariate polynomial model with interaction terms, and an additive bivariate model without interaction terms. From the residual variances of these models, three quantities are computed: Δ (interaction predictability, the observed synergy-redundancy balance), Δ_A (additive interaction predictability, the same balance under the additive constraint), and Ss = Δ − Δ_A (structural synergy). The classification of interaction regimes into five types — no high-order effect, dependency-driven synergy, dependency-driven redundancy,

Load-bearing premise

The framework assumes that the additive model class — functions of the form h(X₁) + h(X₂) — adequately captures all effects arising from source dependencies, so that any excess predictability beyond the additive benchmark can be attributed to non-additive mechanism structure. In practice the additive benchmark is implemented via polynomial regression, where the choice of polynomial order and basis functions determines what counts as additive versus non-additive. If the basis

What would settle it

A system with a genuinely non-additive mechanism that is poorly captured by polynomial basis functions could yield Ss ≈ 0 (false negative), while a system with no non-additive mechanism but with a complex dependency structure not well captured by the additive polynomial model could yield Ss > 0 (false positive). The framework's ability to distinguish mechanism-driven from dependency-driven synergy therefore depends on the polynomial basis being a faithful representation of both the additive and non-additive components.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If structural synergy is present even when observed redundancy dominates, then intervention strategies based solely on observed synergy-redundancy patterns may miss genuine non-additive mechanisms, leading to incorrect predictions about how a system responds to perturbations.
  • The framework could be extended beyond two sources using partial information decomposition, which would allow explicit separation of synergistic and redundant components rather than relying on the whole-minus-sum principle that conflates them.
  • In neuroscience, the finding that motor-network EEG dynamics show masked structural synergy suggests that redundancy-dominated information-theoretic analyses may have systematically underestimated the prevalence of non-additive neural interactions.
  • In climate science, the selective detection of structural synergy in ENSO-related index pairs (NINO34 with PDO or NTA) provides a data-driven criterion for identifying which variable combinations require mechanism-based models rather than additive descriptions.

Where Pith is reading between the lines

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

  • The masked structural synergy regime suggests a general principle: whenever sources are correlated and the target depends on their interaction, increasing source correlation will eventually push the observed synergy-redundancy balance negative while leaving structural synergy intact. This means that in any real system with correlated inputs — which is most real systems — observed redundancy is not
  • The framework could be applied as a diagnostic for when linear or additive models are insufficient: if Ss is significant, one knows that additive models will fail to capture the system's predictive structure, which has direct consequences for model selection in fields ranging from econometrics to systems biology.
  • The polynomial regression implementation means structural synergy is sensitive to the choice of basis functions and polynomial order. A system with a non-additive mechanism that is poorly approximated by polynomials (e.g., discontinuous or highly oscillatory interactions) might show low Ss not because the mechanism is absent but because the model class cannot represent it.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 8 minor

Summary. The manuscript introduces a predictability-based framework to distinguish high-order behaviours (HOBs) — observed synergistic or redundant statistical dependencies among variables — from high-order mechanisms (HOMs) — structural or dynamical rules containing genuine non-additive interaction terms. The core quantity, structural synergy Ss, is defined as the excess predictive power of the full joint model over the best additive model (Eq. 19). The framework is validated through analytical derivations for linear Gaussian and multiplicative non-linear systems, stochastic autoregressive simulations, and two empirical applications (climate indices and source-reconstructed EEG). The central conceptual contribution is the identification of a 'masked structural synergy' regime, where redundancy-dominated HOBs (Δ < 0) coexist with a genuine non-additive mechanism (Ss > 0).

Significance. The dissociation between HOBs and HOMs is a conceptually important contribution to the high-order interactions literature. The analytical derivations are clean and verifiable: the linear Gaussian case correctly yields Ss = 0 (Eq. 23), and the multiplicative case derivation in the Appendix is rigorous, using Isserlis' theorem for Gaussian moments to obtain the closed-form Ss = c²(1−r₁₂²)²/(1+r₁₂²) (Eq. 26). The masked structural synergy regime is a concrete, falsifiable theoretical prediction. The authors provide publicly available MATLAB code, which supports reproducibility. The two empirical applications are illustrative and connect to existing information-theoretic literature on the same datasets.

major comments (3)
  1. The practical implementation relies on polynomial regression (Eqs. 20–22), where the additive model (Eq. 22) and the full model (Eq. 21) share the same polynomial basis for individual source terms. When sources are correlated (r₁₂ ≠ 0), interaction monomials X₁ᵏX₂ʲ are correlated with individual source variables. If the true generative model is additive but non-polynomial (Y = h₁(X₁) + h₂(X₂) + U with smooth non-polynomial h₁, h₂), the polynomial additive model approximates h₁ and h₂ imperfectly. The full model can then use interaction terms to capture residual non-polynomial additive structure — because X₁X₂ carries information about X₁ when r₁₂ ≠ 0 — yielding Ss > 0 and a significant surrogate test even though the mechanism is genuinely additive. This is a false positive for HOMs. The analytical examples avoid this confound because they use polynomial functions (linear or multiplicable
  2. The surrogate test (Methods, 'Statistical assessment of structural synergy') permutes the interaction terms while leaving additive terms unchanged. This tests whether interaction terms provide additional predictive contribution beyond the additive model, but it does not test whether that contribution reflects a genuine non-additive mechanism versus an artifact of basis misspecification. A more informative null would involve fitting the additive model with a richer or alternative basis (e.g., splines, kernels) and comparing against the polynomial interaction model. The authors should either (a) add a simulation where the true model is additive but non-polynomial (e.g., Y = sin(X₁) + cos(X₂) + U with correlated Gaussian sources) and show whether Ss exhibits false positives, or (b) explicitly state in the Methods and Applications that Ss is defined relative to the polynomial basis and that,
  3. In the EEG application (Fig. 4), the results are framed as evidence of a 'masked structural synergy regime' compatible with non-additive mechanisms. Given the confound above, and given that EEG source-reconstructed signals have unknown functional forms and weak but positive source correlations (~0.1), the interpretation should be more cautious. The authors' caveat in the Conclusions ('structural synergy should be interpreted as evidence of non-additive predictive structure rather than as a direct identification of the underlying generative mechanism') partially addresses this, but the application sections still use language like 'compatible with a non-additive predictive component' and 'masked structural synergy regime' that may overstate what the polynomial-regression-based Ss can establish. A brief additional caveat in the application sections, referencing the basis-dependence of Ss,
minor comments (8)
  1. Eq. (21): the double summation notation 'p−1 Σ_{k=1} p−k Σ_{j=1}' is correct but dense; explicitly stating that the constraint is k+j ≤ p in the equation line would prevent confusion.
  2. Table I: the 'Dependency-driven synergy' and 'Dependency-driven redundancy' rows list ΔA = Δ, which is correct, but the column header 'HOM' lists 'Absent' for these rows. This is a consequence of Ss = 0, but a brief footnote clarifying that 'Absent' means 'no non-additive mechanism detectable by this framework' rather than 'no mechanism of any kind' would improve precision.
  3. Fig. 2: the y-axis scales differ substantially across panels (e.g., panel d ranges from −3 to 0, panel a from 0 to ~0.04). A note in the caption or shared scaling for comparable panels would aid interpretation.
  4. Fig. 3: the asterisks denoting significant increases in Ss when increasing polynomial order are described in the caption but their placement relative to specific bars is ambiguous; clarifying which polynomial orders are being compared would help.
  5. Climate application: the Spearman correlation values in Fig. 3 (top row) appear to be plotted on a scale from −0.15 to 0.2, but the text does not state whether the indices were standardized or detrended before correlation. A brief preprocessing note would be appropriate.
  6. EEG application: the Bonferroni correction across 10 trials is mentioned, but it is unclear whether correction was also applied across the 3 ROI triplets or the 3 polynomial orders. Clarification of the full multiple-comparison structure would be appropriate.
  7. Ref [22] (Caprioglio et al., 2026) and Ref [1] (Peixoto et al., 2026) have 2026 dates; verify these are correct and not placeholder dates.
  8. The phrase 'high-order mechanisms (HOMs) as intended in the present work' (Introduction, final paragraph) is slightly awkward; rephrasing to 'as defined in the present work' would improve readability.

Circularity Check

0 steps flagged

No circularity: the framework is derived from first principles with parameter-free analytical results and controlled simulations

full rationale

The paper's central derivation chain is self-contained and non-circular. The core quantities are defined from standard probability and regression theory: the law of total variance (Eq. 10), optimal predictors as conditional expectations (Eqs. 6–9), and the additive predictive class H_add (Eq. 14). Structural synergy Ss = R*_A − R* (Eq. 19) follows by definition from comparing the best additive predictor against the best unrestricted predictor — this is a mathematical identity, not a fitted-then-predicted result. The two analytical examples are parameter-free derivations: for the linear Gaussian model Y = aX1 + bX2 + U, the authors show Ss = 0 because the optimal predictor is already additive (E[Y|X1,X2] = aX1 + bX2), so R*_A = R* by construction — this is a genuine mathematical result, not circular. For the multiplicative model Y = cX1X2 + U, the authors derive Ss = c²(1−r₁₂²)²/(1+r₁₂²) (Eq. 26) from first principles using Isserlis' theorem for Gaussian moments (Appendix Eqs. A.21–A.33). No parameter is fitted to data and then presented as a prediction. The stochastic autoregressive simulations (Eqs. 27a–c) use known ground-truth generative mechanisms with controlled parameters c, c_A, r₁₂, and the polynomial order p=2 matches the quadratic interaction term — so the recovery of expected regimes is not forced by construction. The real-world applications (climate indices, EEG) do not fit framework parameters to the data being explained; they apply the estimation procedure and report Ss values with surrogate-based significance testing. Self-citations exist (refs 8–15 share authors), but these cite prior methodological work on predictability measures and information-theoretic decompositions — they provide context and motivation, not the load-bearing mathematical content. The mathematical derivations in the Appendix are self-contained, using only standard results (Isserlis' theorem, properties of conditional expectation for Gaussian variables, L2 projection theory from ref [47] which is an external textbook). No step in the derivation chain reduces to its own inputs by definition, fit, or self-citation.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 1 invented entities

See above.

free parameters (2)
  • Polynomial order p = Selected via greedy surrogate-based search (p ∈ {2,3,4} in applications)
    The maximum polynomial order controls the dimensionality of the regression and determines which interaction terms are included. Selected data-adaptively via surrogate testing rather than fixed a priori.
  • eLORETA regularization λ = 0.05
    Regularization parameter for EEG source reconstruction, set based on prior literature (ref 43). Not a parameter of the structural synergy framework itself.
axioms (5)
  • standard math The optimal predictor minimizing mean squared prediction error is the conditional expectation E[Y|X1,X2] (Eq. 6).
    Standard result from prediction theory; invoked in Methods section to define R* and the predictive powers λ, λ1, λ2.
  • standard math The law of total variance: σ²_Y = λ + R* (Eq. 10), decomposing target variance into explained and unexplained components.
    Standard probability result; used to derive the interaction predictability Δ in Eq. 13.
  • domain assumption The additive model class H_add = {h(X1,X2) = h1(X1) + h2(X2)} captures all dependency-driven predictive effects, so that Ss = R*_A − R* isolates non-additive mechanism contributions.
    This is the core domain assumption of the framework. It is exact in the two analytical examples but is an approximation when polynomial regression is used in practice. Invoked implicitly throughout Eqs. 14–19 and Table I.
  • standard math Isserlis' theorem for computing moments of jointly Gaussian variables (used in Appendix, Eq. A.7).
    Standard result in probability theory; used to derive E[X1²X2²] = 1 + 2r₁₂² for the multiplicative system.
  • domain assumption The WMS (whole-minus-sum) principle adequately characterizes the synergy-redundancy balance (Eq. 13).
    The WMS approach is known to conflate synergy and redundancy (acknowledged in Conclusions, ref 46). The framework inherits this limitation. Invoked in the definition of Δ.
invented entities (1)
  • Structural synergy (Ss) independent evidence
    purpose: Quantify the component of target predictability that cannot be reproduced by additive source effects, interpreted as evidence of non-additive interaction mechanisms.
    Ss is defined as a difference of mean squared prediction errors (Eq. 19), which is directly computable from data. It has closed-form analytical expressions in two model systems (Eqs. 23, 26) and is validated in simulations with known ground truth. The surrogate testing provides a falsifiable null hypothesis.

pith-pipeline@v1.1.0-glm · 24635 in / 4152 out tokens · 234473 ms · 2026-07-09T15:06:07.737091+00:00 · methodology

0 comments
read the original abstract

High-order interactions are increasingly recognized as a hallmark of collective dynamics in complex systems. The relationship between high-order behaviours (HOBs), observed as synergistic or redundant statistical dependencies, and high-order mechanisms (HOMs), related to the structural or dynamical rules of the data-generating process, remains difficult to establish from data. We introduce a predictability-based framework to disentangle these two levels of description in complex network systems. Structural synergy is defined as the excess predictive power gained when two sources are considered jointly beyond the best additive description and is estimated through polynomial regression by comparing a model with interaction terms against an additive model. Simulations show that dependencies among sources reflecting HOBs can arise even in the absence of HOMs, while synergy due to a non-additive mechanism may remain hidden when the observed synergy-redundancy balance is dominated by redundancy and become detectable only through structural synergy. Applications to climate and source-reconstructed cortical EEG dynamics reveal significant non-additive predictive components despite predominantly redundancy-dominated HOBs. These findings emphasize that HOBs and HOMs can dissociate: a system may display redundancy-dominated HOBs while still containing a significant synergistic mechanism. The proposed framework supports a mechanistically informed interpretation of complex-system dynamics and may help to identify when mechanism-based models are needed to predict the response of a system to perturbations or interventions, while also recognizing that structural synergy should be interpreted as evidence of non-additive predictive structure rather than as a direct identification of the underlying generative mechanism.

Figures

Figures reproduced from arXiv: 2607.07286 by Chiara Bar\`a, Daniele Marinazzo, Laura Sparacino, Luca Faes, Sebastiano Stramaglia, Yuri Antonacci.

Figure 1
Figure 1. Figure 1: Theoretical trends of interaction predictability ( [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Estimated trends of interaction predictability ( [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bar plots reporting the Spearman correlation co [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Boxplots and individual subject values, aver [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages · 3 internal anchors

  1. [1]

    Analogously, the MSPE associated with prediction from X2 isR ∗ 2 =a 2(1−r 2

    +σ 2 U. Analogously, the MSPE associated with prediction from X2 isR ∗ 2 =a 2(1−r 2

    work page

  2. [2]

    Given thatE[Y] = 0, the relevant variance is defined asσ 2 Y =E[Y 2] = E[(aX1 +bX 2 +U) 2] =a 2 +b 2 + 2abr12 +σ 2 U by just exploiting the fact thatE[X1U] =E[X 2U] = 0

    +σ 2 U. Given thatE[Y] = 0, the relevant variance is defined asσ 2 Y =E[Y 2] = E[(aX1 +bX 2 +U) 2] =a 2 +b 2 + 2abr12 +σ 2 U by just exploiting the fact thatE[X1U] =E[X 2U] = 0. Substi- tuting these quantities into Eq. (13) yields: ∆ =−r 2 12(a2 +b 2)−2abr 12.(23) The interaction predictability∆is a quadratic function of the source correlationr12, and the...

    work page

  3. [3]

    +σ 2 U andσ 2 Y =c 2(1 +r 2

    work page

  4. [4]

    Substituting these quantities into Eq

    +σ 2 U (derivations in Appendix). Substituting these quantities into Eq. (13) yields ∆ =c 2(1−3r 2 12).(24) 6 (a) (b) (c) 0 0.2 -0.2 -0.4 -0.6 -0.8 -1 0.4 0.6 0.8 1 a 0 0.2 -0.2 -0.4 -0.6 -0.8 -1 0.4 0.6 0.8 1 a 0 0.2 -0.2 -0.4 -0.6 -0.8 -1 0.4 0.6 0.8 1 c r12 0 0.2-0.2-0.4-0.6-0.8-1 0.40.6 0.81 r12 0 0.2-0.2-0.4-0.6-0.8-1 0.40.6 0.81 r12 0 0.2-0.2-0.4-0....

    work page

  5. [5]

    Specifically,∆>0for|r 12|<1/ √ 3, ∆ = 0for|r 12|= 1/ √ 3(or trivially forc= 0), and ∆<0for|r 12|>1/ √

    work page

  6. [6]

    1(c), where the positive values of∆occur in the central region of ther12 domain, while negative values emerge as the source correlation increases in magnitude

    This behavior is illustrated in Fig. 1(c), where the positive values of∆occur in the central region of ther12 domain, while negative values emerge as the source correlation increases in magnitude. To assess whether such interaction predictability can be explained without invoking genuine interaction mechanisms, we compute the best additive predictor h(X1,...

    work page 1950

  7. [7]

    T. P. Peixoto, L. Peel, T. Gross, and M. De Domenico, arXiv preprint arXiv:2602.16937 (2026)

    work page arXiv 2026

  8. [8]

    Boccaletti, P

    S. Boccaletti, P. De Lellis, C. I. del Genio, K. Alfaro- Bittner, R. Criado, S. Jalan, and M. Romance, Physics Reports1018, 1 (2023)

    work page 2023

  9. [9]

    F. Fang, J. Ma, Y.-J. Ma, and S. Boccaletti, Chaos, Solitons & Fractals186, 115149 (2024)

    work page 2024

  10. [10]

    F. E. Rosas, P. A. Mediano, M. Gastpar, and H. J. Jensen, Physical Review E100, 032305 (2019)

    work page 2019

  11. [11]

    Stramaglia, T

    S. Stramaglia, T. Scagliarini, B. C. Daniels, and D. Marinazzo, Frontiers in Physiology11, 595736 (2021)

    work page 2021

  12. [12]

    P. L. Williams and R. D. Beer, arXiv preprint arXiv:1004.2515 (2010)

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  13. [13]

    Stramaglia, L

    S. Stramaglia, L. Faes, J. M. Cortes, and D. Marinazzo, Physical Review Research6, L032007 (2024)

    work page 2024

  14. [14]

    L. Faes, G. Mijatovic, L. Sparacino, and A. Porta, IEEE Transactions on Biomedical Engineering (2025)

    work page 2025

  15. [15]

    Sparacino, G

    L. Sparacino, G. Mijatovic, Y. Antonacci, L. Ricci, D. Marinazzo, S. Stramaglia, and L. Faes, Physical Review E112, 044313 (2025)

    work page 2025

  16. [16]

    C. Barà, Y. Antonacci, H. Pinto, L. Sparacino, M. Ja- vorka, S. Stramaglia, and L. Faes, IEEE Transactions on Biomedical Engineering (2026)

    work page 2026

  17. [17]

    Mijatovic, L

    G. Mijatovic, L. Sparacino, Y. Antonacci, M. Javorka, D. Marinazzo, S. Stramaglia, and L. Faes, IEEE Open Journal of Engineering in Medicine and Biology5, 846 (2024)

    work page 2024

  18. [18]

    Marinazzo, S

    Y.Antonacci, L.Minati, D.Nuzzi, G.Mijatovic, R.Per- nice, D. Marinazzo, S. Stramaglia, and L. Faes, IEEE Access9, 149486 (2021)

    work page 2021

  19. [19]

    Antonacci, C

    Y. Antonacci, C. Barà, L. Sparacino, I. Pirovano, A. Mastropietro, G. Rizzo, and L. Faes, IEEE Trans- actions on Biomedical Engineering (2024)

    work page 2024

  20. [20]

    T. F. Varley, M. Pope, J. Faskowitz, and O. Sporns, Communications biology6, 451 (2023)

    work page 2023

  21. [21]

    L. Faes, L. Sparacino, G. Mijatovic, Y. Antonacci, L. Ricci, D. Marinazzo, and S. Stramaglia, Physical Review Letters135, 187401 (2025)

    work page 2025

  22. [22]

    F. E. Rosas, P. A. Mediano, A. I. Luppi, T. F. Varley, J. T. Lizier, S. Stramaglia, H. J. Jensen, and D. Mari- nazzo, Nature Physics18, 476 (2022)

    work page 2022

  23. [23]

    Malizia, A

    F. Malizia, A. Corso, L. V. Gambuzza, G. Russo, V. La- tora, and M. Frasca, Nature Communications15, 5184 (2024)

    work page 2024

  24. [24]

    L. Faes, G. Mijatovic, Y. Antonacci, R. Pernice, C. Barà, L. Sparacino, M. Sammartino, A. Porta, D. Marinazzo, and S. Stramaglia, IEEE Transactions on Signal Processing70, 5766 (2022)

    work page 2022

  25. [25]

    Marinazzo, Physics18, 71 (2025)

    D. Marinazzo, Physics18, 71 (2025)

    work page 2025

  26. [26]

    A. B. Barrett, Physical Review E91, 052802 (2015)

    work page 2015

  27. [27]

    Robiglio, M

    T. Robiglio, M. Neri, D. Coppes, C. Agostinelli, F. Bat- tiston, M. Lucas, and G. Petri, Physical review letters 134, 137401 (2025)

    work page 2025

  28. [28]

    Caprioglio, P

    E. Caprioglio, P. A. Mediano, and L. Berthouze, Phys- ical Review Letters136, 127401 (2026)

    work page 2026

  29. [29]

    L. Faes, D. Marinazzo, S. Stramaglia, F. Jurysta, A. Porta, and N. Giandomenico, Philosophical Trans- actions of the Royal Society A: Mathematical, Physical and Engineering Sciences374(2016)

    work page 2016

  30. [30]

    Porta, V

    A. Porta, V. Bari, B. De Maria, A. C. Takahashi, S. Guzzetti, R. Colombo, A. M. Catai, F. Raimondi, 13 and L. Faes, IEEE Transactions on Biomedical Engi- neering64, 2628 (2017)

    work page 2017

  31. [31]

    Disentangling Interactions and Dependencies in Feature Attribution

    G. König, E. Günther, and U. von Luxburg, arXiv preprint arXiv:2410.23772 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    McGill, Transactions of the IRE Professional Group on Information Theory4, 93 (1954)

    W. McGill, Transactions of the IRE Professional Group on Information Theory4, 93 (1954)

    work page 1954

  33. [33]

    Hastie, R

    T. Hastie, R. Tibshirani, J. H. Friedman, and J. H. Friedman,The elements of statistical learning: data mining, inference, and prediction, Vol. 2 (Springer, 2009)

    work page 2009

  34. [34]

    J. K. Blitzstein and J. Hwang,Introduction to probabil- ity(Chapman and Hall/CRC, 2019)

    work page 2019

  35. [35]

    Timme, W

    N. Timme, W. Alford, B. Flecker, and J. M. Beggs, Journal of computational neuroscience36, 119 (2014)

    work page 2014

  36. [36]

    D. R. Brillinger, inSelected Works of David Brillinger (Springer, 2011) pp. 589–606

    work page 2011

  37. [37]

    Lütkepohl,Introduction to multiple time series anal- ysis(Springer Science & Business Media, 2013)

    H. Lütkepohl,Introduction to multiple time series anal- ysis(Springer Science & Business Media, 2013)

    work page 2013

  38. [38]

    Theiler, S

    J. Theiler, S. Eubank, A. Longtin, B. Galdrikian, and J. D. Farmer, Physica D: Nonlinear Phenomena58, 77 (1992)

    work page 1992

  39. [39]

    Palus, Physics Letters A235, 341 (1997)

    M. Palus, Physics Letters A235, 341 (1997)

    work page 1997

  40. [40]

    Kilian and H

    L. Kilian and H. Lütkepohl,Structural vector autore- gressive analysis(Cambridge University Press, 2017)

    work page 2017

  41. [41]

    Teräsvirta, D

    T. Teräsvirta, D. Tjøstheim, and C. W. Granger,Mod- elling nonlinear economic time series(Oxford Univer- sity Press, 2010)

    work page 2010

  42. [42]

    M. J. McPhaden, S. E. Zebiak, and M. H. Glantz, sci- ence314, 1740 (2006)

    work page 2006

  43. [43]

    Silini, G

    R. Silini, G. Tirabassi, M. Barreiro, L. Ferranti, and C. Masoller, Climate Dynamics61, 79 (2023)

    work page 2023

  44. [44]

    Schalk, D

    G. Schalk, D. J. McFarland, T. Hinterberger, N. Bir- baumer, and J. R. Wolpaw, IEEE Transactions on biomedical engineering51, 1034 (2004)

    work page 2004

  45. [45]

    A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley, Circulation101, e215 (2000 (June 13)), circulation Electronic Pages: http://circ.ahajournals.org/content/101/23/e215.full PMID:1085218; doi: 10.1161/01.CIR.101.23.e215

    work page doi:10.1161/01.cir.101.23.e215 2000

  46. [46]

    Delorme and S

    A. Delorme and S. Makeig, Journal of neuroscience methods134, 9 (2004)

    work page 2004

  47. [47]

    R. D. Pascual-Marqui, arXiv preprint arXiv:0710.3341 (2007)

    work page internal anchor Pith review Pith/arXiv arXiv 2007

  48. [48]

    Mazziottaet al., Philosophical Transactions of the Royal Society of London

    J. Mazziottaet al., Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences 356, 1293 (2001)

    work page 2001

  49. [49]

    Pirovanoet al., IEEE Transactions on Neural Systems and Rehabilitation Engineering31, 4549 (2023)

    I. Pirovanoet al., IEEE Transactions on Neural Systems and Rehabilitation Engineering31, 4549 (2023)

    work page 2023

  50. [50]

    Grefkes, S

    C. Grefkes, S. B. Eickhoff, D. A. Nowak, M. Dafotakis, and G. R. Fink, Neuroimage41, 1382 (2008)

    work page 2008

  51. [51]

    Van de Steen, L

    F. Van de Steen, L. Faes, E. Karahan, J. Songsiri, P. A. Valdes-Sosa, and D. Marinazzo, Brain topography32, 643 (2019)

    work page 2019

  52. [52]

    P. A. Mediano, F. E. Rosas, A. I. Luppi, R. L. Carhart- Harris, D. Bor, A. K. Seth, and A. B. Barrett, Pro- ceedings of the National Academy of Sciences122, e2423297122 (2025)

    work page 2025

  53. [53]

    A. Buja, T. Hastie, and R. Tibshirani, The Annals of Statistics , 453 (1989). Appendix Derivation of the predictive terms for the non-additive multiplicative system Individual and joint predictive terms When one predictor is considered, it is necessary to compute the termE[Y|X 1], which can be written as E[Y|X 1] =E[cX 1X2 |X 1]+E[U|X 1] =cX 1E[X2 |X 1], ...

    work page 1989

  54. [54]

    Therefore, E (cX1W+U) 2 =E[c 2X 2 1 W 2] +E[2cX 1W U] +E[U 2] =c 2E[X 2 1]E[W 2] +σ 2 U =c 2σ2 X1(1−r 2

    work page

  55. [55]

    +σ 2 U , (A.5) so thatR ∗ 1 =c 2(1−r 2

    work page

  56. [56]

    +σ 2 U (and symmetricallyR ∗ 2 = R∗ 1). The variance ofYcan be retrieved from the computa- tion of all statistical moments forE[(cX1X2+U−cr 12)2] as follows: =E[c 2X 2 1 X 2 2 +U 2 +c 2r2 12] +E[2cX 1X2U−2c 2X1X2r12 −2U cr 12] =c 2E[X 2 1 X 2 2] +σ 2 U +c 2r2 12 −2c 2r2 12 =c 2(1 +r 2

    work page

  57. [57]

    (A.7) Thus, σ2 Y =c 2(1 +r 2

    +σ 2 U , (A.6) which follows from Isserlis’ theorem [25], since E[X 2 1 X 2 2] =E[X 2 1]E[X 2 2] + 2(E[X1X2])2 = 1 + 2r2 12. (A.7) Thus, σ2 Y =c 2(1 +r 2

    work page

  58. [58]

    2 2r 2 12 2r2 12 2 #

    +σ 2 U .(A.8) 14 Best additive predictor and additive MSPE We herein derive the computation of∆A by finding the best additive predictorh(X 1, X2) =d+h 1(X1) + h2(X2)(whered∈Ris an intercept andE[h 1(X1)] = E[h2(X2)] = 0) in the sense of minimizing R∗ A = min d,h1,h2∈L2 E[(Y−d−h 1(X1)−h 2(X2))2].(A.9) According with [47], differentiating with respect tod y...

    work page

  59. [59]

    (A.29) The minimum value of (A.19) is equal to the variance of Zminus the amount of variance explained by its opti- mal additive approximation

    =r 12 ⇒α ∗ =β ∗ = r12 1 +r 2 12 .(A.27) Therefore, the optimal additive predictors are h∗ 1(X1) =c r12 1 +r 2 12 (X 2 1 −1),(A.28a) h∗ 2(X2) =c r12 1 +r 2 12 (X 2 2 −1).(A.28b) Thus, the best additive predictor can be written as h∗(X1, X2) =cr 12 +c r12 1 +r 2 12 h (X 2 1 −1) + (X 2 2 −1) i . (A.29) The minimum value of (A.19) is equal to the variance of ...

    work page