{"id":"1a527948-eefa-4228-b23e-aea758e8be34","arxiv_id":"2501.13890","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A federated linear state-space framework learns cross-client Granger causality from shared low-dimensional states, with convergence and differential-privacy guarantees.","lead":"This paper proposes a federated learning framework where industrial clients share only low-dimensional state estimates with a central server, which learns cross-client Granger causality from those states without seeing raw sensor data. It matters because it combines bandwidth savings, differential privacy, and causal analysis for distributed industrial monitoring.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Oracle-convergence proof is invalid: Proposition 5.3's zero-gradient condition does not imply zero expected residual, and Theorem 6.3's proof also shifts time indices, so the advertised Granger-causality bounds are unsupported.","rationale":"The reader's verdict rejects the paper on the strength of multiple theoretical gaps. My pass isolates the single most load-bearing one: Proposition 5.3 condition 1, the bridge from client-loss stationarity to zero expected residual, is mathematically false. The gradient in Eq. 21 is a product of the residual with the regressor y^{t-1}; stationarity only makes the residual orthogonal to that regressor in expectation. A simple scalar example with nonzero-mean y_t and zero-mean or nonzero-mean y_{t-1} shows that zero expected gradient can hold while the expected residual is nonzero. Since Theorem 6.1 invokes this condition directly and Theorem 6.3 suffers from a parallel time-index shift, the central guarantees advertised in the abstract and contribution 3 do not follow from the supplied proof. This is an internal inconsistency, not a disagreement with external consensus, and no amount of empirical support can repair a broken derivation. The reader's weakest_assumption focused on Assumption 4.2's linear form; I see the proof failure as even more fundamental because it invalidates the theory even when that assumption is granted. Hence I agree with the REJECT verdict but only partially with the stated weakest assumption. No verdict change is needed.","tokens_in":28632,"tokens_out":13094,"duration_ms":114278,"concrete_test":"Re-derive Proposition 5.3 condition 1 from Eq. 21 in a scalar setting: take D=P=1, C_mm=A_mm=1, (h_hat^{t-1}_m)_c=0, y_{t-1} ~ Uniform[1,2], and y_t = 2.5 + epsilon_t with epsilon_t independent and mean-zero. The normal equation from Eq. 21 yields theta* = E[y_t y_{t-1}]/E[y_{t-1}^2], and E[y_t - theta* y_{t-1}] = 2.5 - theta* * 1.5 != 0, so condition 1 fails under the paper's stated gradient formula. Then independently re-derive Theorem 6.3 using only the actual updates with time index t-1 (Eqs. 16 and 18) and check whether the displayed inequality can be obtained without the time-index substitution; if it cannot, the Granger-causality bound is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires augmented client states to converge to the centralized oracle and stationary off-diagonal blocks to be close to true A_mn. The proof hinges on Proposition 5.3 condition 1: E[y_t_m - C_mm A_mm((h_hat^{t-1}_m)_c + theta*_m y^{t-1}_m)] = 0. This does not follow from Eq. 21. Setting E[gradient] = 0 gives E[(C_mm A_mm)^T (r_t_m)_a (y^{t-1}_m)^T] = 0, an orthogonality condition; because y^{t-1}_m is a random vector, it does not imply E[(r_t_m)_a] = 0. Full rank of C_mm A_mm and y != 0 do not remove the regressor from the expectation. Theorem 6.1's proof explicitly uses the invalid implication 'from Proposition 5.3, E[(r_t_m)_a] = 0' to conclude convergence to the oracle. Theorem 6.3's proof has an additional time-index problem: Eqs. (47)-(48) evaluate the residual and stationarity condition at time t with y^t_m and h_hat^t_c, whereas the actual server gradient in Eq. (16) and the augmentation identity Eq. (18) involve y^{t-1}_m and h_hat^{t-1}_c. Replacing t-1 by t changes the expectation and is not justified. Thus the matrix bound in Corollary 6.4 is not derived. The empirical demonstration may be sound, but the advertised guarantees for oracle convergence and bounded Granger-causality error are not established as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a federated framework for learning Granger causality in multi-client linear time-invariant state-space systems. Each client runs a Kalman filter using only its own diagonal state-transition block, and augments its estimated state with a linear ML function of its own measurements. A central server estimates the off-diagonal blocks of the state matrix from the clients' states, and communicates gradients back to clients. The authors claim co-dependence of client and server updates, a unified recurrent linear form, sublinear and linear convergence rates, convergence of augmented client states to a centralized oracle, matrix bounds on the Granger-causality estimation error, and differential privacy guarantees for both communication directions. The empirical sections report synthetic and real-world industrial control system experiments, including robustness, scalability, and data-volume savings.","tokens_in":29046,"tokens_out":5746,"duration_ms":54946,"significance":"If the theoretical claims were valid, this would be a useful contribution to decentralized causality discovery: it addresses a realistic multi-client state-space setting, provides a communication-efficient protocol, includes a differential privacy analysis, and the authors ship code with their experiments. The empirical study is substantial and the privacy analysis is standard and plausible. However, the central theoretical contributions are not established. The proof of Proposition 5.3 replaces a valid orthogonality stationarity condition with the much stronger claim that the expected residual is zero, and Theorem 6.3's proof shifts time indices without justification. Since Theorem 6.1 and Corollary 6.4 depend on these steps, the advertised oracle-convergence and bounded-Granger-error guarantees are unsupported. The experiments may be sound, but they cannot compensate for the invalid derivations in the main theoretical claims.","major_comments":[{"comment":"The derivation of condition (1) is invalid. Setting E[∇θ(Lm)a]=0 gives (CmmAmm)^T E[(r_t_m)_a (y_{t-1}^m)^T]=0, an orthogonality condition between the residual and the lagged regressor. Even when CmmAmm has full column rank and y_{t-1}^m is nonzero, this does not imply E[(r_t_m)_a]=0, because CmmAmm is tall and the regressor remains inside the expectation. The step labeled \"simplifies to\" in Eq. (26) therefore does not follow. This is load-bearing: Theorem 6.1 uses E[(r_t_m)_a]=0 to conclude convergence to the oracle, and Theorem 6.3 also relies on the same implication.","section":"§5, Proposition 5.3 and Appendix A.4.3, Eqs. (21) and (25)"},{"comment":"The proof of Theorem 6.3 changes the time index without justification. Eqs. (47) and (48) evaluate the residual and the stationarity condition at time t using y_t^m and (hhat^t_n)_c, whereas the actual server gradient in Eq. (16) and the augmentation identity in Eq. (18) involve y_{t-1}^m and (hhat^{t-1}_n)_c. Replacing t-1 by t alters the expectation and is not a notational variant. Consequently the intermediate expression in Eq. (49) and the final matrix bound in Corollary 6.4 are not derived from the stated framework.","section":"§6, Theorem 6.3 and Appendix A.4.10, Eqs. (47)–(48)"},{"comment":"The convergence analysis of the unified recurrence is not justified. The matrix H in Eq. (4) is built from random, time-varying quantities y_{t-1}^m and (hhat^{t-1}_n)_c, so Lemma 5.5's claim that convergence is equivalent to ρ(H)<1 is not a valid criterion for a random time-varying linear recurrence. In addition, Theorems 5.6 and 5.7 assume a joint loss L_f that is convex/smooth/strongly convex and claim Eq. (5) is its gradient descent, but no such L_f is ever constructed and the actual updates in Eqs. (21), (24), and (16) are least-squares updates with data-dependent regressors, not gradient descent on a single fixed loss. The sublinear and linear convergence-rate claims are therefore not supported.","section":"§5, Theorem 5.4, Lemma 5.5, Theorems 5.6–5.7"},{"comment":"The transition from Eq. (51) to Eq. (52) requires more than collinearity: it requires that the vectors [A*_mn - A_mn](hhat^{t-1}_n)_o all point in the same direction, and even then the lower bound using min_{n≠m} E[||(hhat^{t-1}_n)_o||] is not justified because the expectation of the norm of the sum is not generally bounded below by the norm of the sum of expectations times the minimum expected norm. Additionally, if any of the relevant oracle states can be zero with positive probability, σ^n_min may fail to be positive. The corollary's bound is thus not established.","section":"§6, Corollary 6.4 and Appendix A.4.11, Eqs. (51)–(52)"}],"minor_comments":[{"comment":"The notation for the two learning rates η1 and η2 is introduced but the text does not specify how their magnitudes affect the convergence conditions in Theorem 5.4; the recurrence H depends on them, but the spectral-radius condition is stated without relating it to these rates.","section":"§4.1, Eq. (1)–(2)"},{"comment":"The table shows very small server losses, but the baselines and the proposed method are compared without confidence intervals or repeated-seed variability; adding standard errors would make the empirical claims more robust.","section":"§8, Table 4"},{"comment":"The reported centralized ground-truth A for the real-world datasets is itself estimated from the same data; the paper should state this more prominently, since the comparison is then between two estimators rather than against a true ground truth.","section":"§9, Table 7"},{"comment":"The limitations section does not mention that the main theoretical results require the linear-in-y augmentation of Assumption 4.2; given how much of the analysis depends on this assumption, it deserves explicit discussion as a limitation.","section":"Appendix A.9, Limitations"},{"comment":"Several equations contain inconsistent time superscripts (for example, Table 1 uses h_t but Section 4 uses h^t interchangeably, and Theorem 6.3 mixes t and t-1); a careful pass to standardize notation would improve readability.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The reader's report aligns with my own reading of the manuscript. The central theoretical claims—oracle convergence and bounded Granger-causality error—rest on an invalid simplification of the first-order stationarity condition and on an unjustified time-index shift. These are not local presentation issues; they affect the main advertised contributions. The empirical study and privacy analysis are reasonable, but they do not repair the theoretical core. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a genuine contribution to federated causality, but the theory that anchors it does not hold up on close reading. The empirical framework is worth knowing about; the proofs need substantial repair before the advertised guarantees can be taken seriously.\n\nWhat is actually new and useful: the paper is the first to formulate federated Granger causality in state-space models, using a client Kalman filter, a linear augmentation theta_m y, and a server-side estimate of the off-diagonal A blocks. The synthetic experiments show reasonable recovery of the cross-client causal structure, robustness to perturbations and topology changes, and scalability in both measurement dimension and client count. The real-world demonstration on HAI and SWaT reports concrete data savings, and the code is public. The differential privacy analysis is standard but correctly derived.\n\nThe soft spots are load-bearing. Proposition 5.3 claims that the zero-gradient condition for the client loss implies E[(r_t^m)_a]=0. Starting from Eq. 21, the gradient condition is actually E[(C_mm A_mm)^T (r_t^m)_a (y^{t-1}_m)^T]=0, an orthogonality condition. Because y^{t-1}_m sits inside the expectation, full rank of C_mm A_mm does not let you drop it and conclude the residual has zero mean. That invalid conclusion is then used in Theorem 6.1, so the oracle-convergence result is unsupported. Theorem 6.3 has a separate time-index problem: Eqs. (47)-(48) evaluate objects at time t that are defined at t-1, so the advertised matrix bound on the Granger causality error is not derived. Lemma 5.5 treats H as constant in the recurrence Delta^{k+1}=H Delta^k+J, but H and J actually depend on the time index through y and h_hat; the spectral-radius condition does not apply to a time-varying linear recurrence. Theorems 5.6/5.7 are conditional statements about an undefined loss L_f and therefore do not add support.\n\nThese are not cosmetic issues. The central claims of the paper are convergence to a centralized oracle and bounded error in the learned causality, and both rest on the invalid steps above. The empirical method may still work as a heuristic, and the experiments suggest it does, but the theory as written does not establish the guarantees.\n\nWho is this for? Someone working on federated time-series or distributed causal discovery who wants a concrete architecture and experimental baseline. A reader looking for rigorous guarantees should wait for a revision. The paper deserves a serious referee because the problem and experiments are substantive, but it needs major revision.\n\nIf this crossed my desk, I would send it to peer review and ask for major revision: repair or remove Proposition 5.3, rebuild the oracle-convergence and error-bound arguments with correct time indices and stochastic expectations, and either define L_f or drop the convergence-rate theorems.","headline":"A promising federated Granger-causality framework with a solid empirical core, but the theoretical guarantees advertised in the title are not established because of invalid gradient-stationarity and time-index arguments.","tokens_in":29536,"tokens_out":3107,"would_cite":false,"duration_ms":29375,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a federated system of clients can learn Granger causality between distributed industrial processes by sharing only low-dimensional Kalman-filter states, and that the learned causal matrix is provably close to the…","keywords":["federated learning","Granger causality","state space model","Kalman filter","differential privacy","cross-client interdependency","industrial control systems","time series"],"falsifier":"Run the framework on a two-client linear state-space system with known ground-truth $A$, but make client 2's influence on client 1 nonlinear (for example, the effective $A_{12}$ depends on $y_2^2$) while keeping the augmentation linear. If the estimated $\\hat{A}_{12}$ fails to track the true causal effect, or if the Frobenius bound of Corollary 6.4 is violated, the central claim's load-bearing linearity assumption is falsified.","tokens_in":28367,"feed_emoji":"🔗","tokens_out":8259,"duration_ms":67652,"temperature":0.7,"pith_summary":"The paper tries to establish that Granger causality between geographically distributed industrial clients can be learned without pooling high-dimensional measurements. Clients run Kalman filters on their own data and send only low-dimensional state estimates to a server; the server estimates the off-diagonal blocks of a shared linear state-transition matrix, and each client's model is augmented with a linear ML term so that its local augmented state converges to what a centralized oracle would compute. The paper proves convergence of the coupled client-server updates, gives sublinear and linear convergence rates for the equivalent standalone algorithm, and bounds the error between estimated and true cross-client causality. If these results hold, decentralized monitoring systems can detect fault propagation and operational interdependencies while saving bandwidth and adding differential privacy.","feed_headline":"States, not raw data, reveal cross-client Granger causality","feed_subtitle":"A server estimates off-diagonal state-matrix blocks with provably bounded error versus a centralized oracle, saving bandwidth.","key_machinery":"The central object is the block state-transition matrix $A$ of a linear time-invariant state-space model, whose off-diagonal blocks $A_{mn}$ with $m\\neq n$ encode Granger causality from client $n$ to client $m$. The mechanism is an additive augmentation: the estimated augmented state is $(\\hat{h}^{t}_{m})_{a} = (\\hat{h}^{t}_{m})_{c} + \\theta_{m} y^{t}_{m}$, where $\\theta_{m}$ is learned by gradient descent using both the client's reconstruction loss and the server's loss. The server predicts $(h^{t}_{m})_{s} = A_{mm}(\\hat{h}^{t-1}_{m})_{c} + \\sum_{n\\neq m} \\hat{A}^{t}_{mn}(\\hat{h}^{t-1}_{n})_{c}$, and the whole coupled update is recast as the linear recurrence $\\Delta^{k+1} = H\\Delta^{k} + J$, which converges if and only if the spectral radius $\\rho(H) < 1$. Theorems 6.1 and 6.3 then argue that augmented client states converge in expectation to the centralized oracle and that the stationary estimate $\\hat{A}^{*}_{mn}$ has bounded Frobenius distance to the true $A_{mn}$.","core_discovery":"On its own terms, the paper's central discovery is that cross-client Granger causality can be learned from low-dimensional state estimates rather than from pooled high-dimensional measurements. Clients run Kalman filters with their own diagonal state blocks, and the server estimates the off-diagonal blocks of the shared state-transition matrix from the states the clients share. The client updates are augmented by a linear ML term, the server and client updates are shown to co-depend, and the whole loop is reduced to a linear recurrence. The paper proves that the augmented client states converge in expectation to a centralized oracle (Theorem 6.1), and that the stationary off-diagonal estimates satisfy the bounded-error inequality of Theorem 6.3, with a Frobenius-norm corollary under a collinearity and minimum-norm condition (Corollary 6.4). The experiments on synthetic and real industrial control system data illustrate communication savings and robustness to causal perturbations.","pith_inferences":["Beyond the paper, the estimated off-diagonal blocks could serve as a live fault-propagation diagnostic: a shift in $\\hat{A}_{mn}$ would flag a new or altered dependency between clients $m$ and $n$. The paper reports training losses and estimation errors, not this diagnostic use.","Beyond the paper, replacing the linear augmentation with a polynomial or neural function of $y^{t}_{m}$ is a natural stress test; the closed-form recurrence and oracle-convergence proof would no longer apply, and the paper does not examine this regime.","Beyond the paper, the privacy analysis is stated for one communication round; composing Theorems 7.1 and 7.2 over $T$ time steps would yield an overall privacy budget that the paper does not compute.","Beyond the paper, the real-world experiments validate against a centralized estimate of $A$ rather than an independent causal ground truth, so a deployment test would need a separate benchmark to confirm the estimated causality on actual system faults."],"forward_implications":["Clients can learn cross-client causality without ever transmitting their high-dimensional measurements; only low-dimensional estimated states and gradient signals cross the network.","The augmented client state converges, in expectation, to the centralized oracle's state, so decentralized inference can match centralized monitoring once training converges.","The server's estimated off-diagonal blocks converge to the true causal structure up to a matrix bound that can be evaluated without knowing the ground-truth $A$.","Under convexity and smoothness conditions, the equivalent standalone ML algorithm converges at rate $O(1/k)$, and under strong convexity at a linear rate.","Both client-to-server and server-to-client communications can be made $(\\varepsilon, \\delta)$-differentially private with Gaussian noise calibrated to the sensitivities derived in Theorems 7.1 and 7.2."],"supporting_citations":[{"why":"Defines Granger causality, the concept the framework aims to learn in a decentralized state-space setting.","marker":"Granger (1969)"},{"why":"Establishes the federated learning paradigm whose communication constraints motivate sharing states rather than raw measurements.","marker":"McMahan et al. (2017)"},{"why":"Supplies the differential privacy definitions and Gaussian mechanism used in Theorems 7.1 and 7.2.","marker":"Dwork & Roth (2014)"},{"why":"Relates Granger noncausality to the graph of state-space representations, grounding the off-diagonal-block interpretation.","marker":"Jozsa et al. (2019)"},{"why":"Provides the SWaT industrial control system dataset used for real-world validation.","marker":"Mathur & Tippenhauer (2016)"},{"why":"Provides the HAI industrial control system dataset used for real-world validation.","marker":"Shin et al. (2023)"}],"fun_headline_variants":["Cross-client Granger causality from shared states, not raw data","Federated state-space learning reveals cross-client Granger causality","Server learns off-diagonal state blocks to infer Granger causality","State-based federated causality converges to centralized oracle","Low-dim state estimates drive cross-client Granger causality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on Assumption 4.2 that each client's augmentation term is exactly linear in its own measurement $y^{t}_{m}$; if cross-client influence cannot be represented that way, the learned off-diagonal blocks need not equal the true Granger causality.","fun_headline_variants_meta":{"raw":{"variants":["Cross-client Granger causality from shared states, not raw data","Federated state-space learning reveals cross-client Granger causality","Server learns off-diagonal state blocks to infer Granger causality","State-based federated causality converges to centralized oracle","Low-dim state estimates drive cross-client Granger causality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000558,"raw_usage":{"total_tokens":2685,"prompt_tokens":1011,"completion_tokens":1674,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1593}},"tokens_in":627,"tokens_out":1674,"duration_ms":12094,"temperature":1.0,"reasoning_tokens":1593,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:30:29.656793+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the framework on a two-client linear state-space system with known ground-truth $A$, but make client 2's influence on client 1 nonlinear (for example, the effective $A_{12}$ depends on $y_2^2$) while keeping the augmentation linear. If the estimated $\\hat{A}_{12}$ fails to track the true causal effect, or if the Frobenius bound of Corollary 6.4 is violated, the central claim's load-bearing linearity assumption is falsified.","supporting_citations":[{"cited_title":"The algorithmic foundations of differential privacy","cited_arxiv_id":null,"evidence_quote":"Supplies the differential privacy definitions and Gaussian mechanism used in Theorems 7.1 and 7.2."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Establishes the federated learning paradigm whose communication constraints motivate sharing states rather than raw measurements."},{"cited_title":"Hai security datasets, 2023","cited_arxiv_id":null,"evidence_quote":"Provides the HAI industrial control system dataset used for real-world validation."}],"review_version":1}