{"id":"ee1cdb1f-06fd-4666-b7e5-a9d335a3e34f","arxiv_id":"2506.21036","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"An information-theoretic upper bound on a Markov-chain-weighted generalization error for federated learning under concept drift, together with a regularized ERM algorithm and a Pareto cost-performance analysis.","lead":"This paper studies federated learning when the data distribution shifts over time, and derives a mathematical bound on how well a model trained now will perform on future data. It also proposes a training method that adds KL-divergence and noise regularizers to help models adapt, and tests it on a Raspberry Pi testbed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1 relies on an unstated Bregman-divergence loss assumption, so the main bound is not established for the losses used in the experiments.","rationale":"The reader's verdict correctly flags the large-N simplification in Section IV-B as a gap, but the more fundamental issue is that the main theorem is proved only under a hidden Bregman-divergence loss assumption. This is not a minor formality: the proof's crucial step for samples from the previous distribution (Eq. 33) requires this assumption to control the cross-distribution excess risk. The theorem statement, Assumption 1, and the experimental section give no indication that the losses used (e.g., cross-entropy) satisfy this condition, so the paper's central theoretical claim is not actually established for its own experiments. The Bregman issue is therefore the single most load-bearing concern: if it fails, the entire upper bound and all subsequent corollaries collapse, whereas the large-N issue only weakens the finite-sample tightness of an already-proven bound. The concern is addressable by stating the assumption explicitly, proving or citing that the experimental losses satisfy it, and re-deriving the bound without the hidden condition if possible. Given that the contribution is primarily theoretical and the gap is concrete, a conditional verdict is appropriate: the paper should be accepted only after the authors clarify and justify the Bregman assumption, or prove the theorem without it. This does not change the reader's overall verdict category, but it shifts the emphasis from an asymptotic technicality to a correctness condition on the main result.","tokens_in":25605,"tokens_out":15710,"duration_ms":167437,"concrete_test":"Independently re-derive Eq. (33) of Appendix A without invoking the Bregman-divergence condition, and test the resulting inequality on a concrete counterexample: squared loss, \\pi_{pre} = N(0,1), \\pi_{cur} = N(\\mu,1), with w drawn independently of the sample. If E_{w,Z~\\pi_{pre}}[L_{\\pi_{cur}}(w) - \\ell(w,Z)] is not bounded by \\psi^{*-1}(I(w;Z) + D_{KL}(\\pi_{pre}\\|\\pi_{cur})) for some \\mu, then Theorem 1 as stated is false. Additionally, check whether the cross-entropy loss used in the experiments satisfies the Bregman condition of [29, Theorem 4].","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The proof of Theorem 1 (Appendix A) explicitly invokes [29, Theorem 4] and states 'we assume that the loss function is in the form of Bregman divergence.' This assumption appears only inside the proof, not in Theorem 1, Assumption 1, or the system model. It is essential for bounding the terms in Eq. (33): the excess risk L_{\\pi_{cur}}(w) - \\ell(w, Z) for a sample Z ~ \\pi_{pre} can be large when the target and source distributions differ, and only for Bregman-divergence losses does the change-of-measure argument yield a bound of the form \\psi^{*-1}(I(w;Z) + D_{KL}(\\pi_{pre}\\|\\pi_{cur})). Without this condition, the inequality can fail: for squared loss with \\pi_{pre} = N(0,1), \\pi_{cur} = N(\\mu,1), and w independent of Z, E[L_{\\pi_{cur}}(w) - \\ell(w,Z)] = (\\mu - w)^2 + 1, which can exceed \\psi^{*-1}(D_{KL}) for large \\mu, even with I=0. The experiments use cross-entropy loss (CIRCLE, MNIST, CIFAR-10, Fashion-MNIST), and the paper never verifies that this loss satisfies the Bregman condition required by the proof. Thus Theorem 1 as stated is unproven for the central examples of the paper, and the corollaries (Eqs. 13-17) inherit this gap. This is more load-bearing than the large-N simplification in Section IV-B: even if the mutual information terms are retained, the bound itself may not be valid for non-Bregman losses.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an information-theoretic analysis of federated learning under concept drift. Concept drift is modeled as a Markov chain over data distributions, and a Stationary Generalization Error G is defined as the steady-state expected weighted generalization error of the current global model on both the current and the next distribution. Theorem 1 states an upper bound on G in terms of mutual information I(w_cur, Z) and KL divergences between successive data distributions. The bound is then simplified for large sample sizes and sub-Gaussian losses, leading to closed-form expressions for periodic, gradual, and random drift patterns (Corollary 1) and an ordering result (Proposition 1). The paper also proposes a regularized ERM algorithm (KLMI-FedAvg), analyzes a performance-cost tradeoff via a Pareto front, and reports experiments on a Raspberry Pi testbed and on CIFAR-10, MNIST, Fashion-MNIST, and CIRCLE.","tokens_in":25942,"tokens_out":8388,"duration_ms":91567,"significance":"If the main theorem were fully established, the paper would provide a useful information-theoretic vocabulary for comparing drift patterns in FL and would connect the bound to an algorithmic regularizer and to a cost-performance tradeoff. The paper is largely non-parametric: the bound is a direct application of cited information-theoretic lemmas, with no hidden constants beyond the algorithm's hyperparameter gamma, and the testbed experiments are a concrete attempt at validation. However, the central theoretical claim is currently not proven for the losses used in the experiments, and the pattern-ordering results rest on an asymptotic simplification that is presented as a finite-sample bound. These gaps are load-bearing for the paper's main conclusions, so the contribution is significant only conditionally on repairing them.","major_comments":[{"comment":"The proof of Theorem 1 invokes [29, Theorem 4] and, immediately after Eq. (33), states that 'we assume that the loss function is in the form of Bregman divergence.' This assumption is not stated in Theorem 1, Assumption 1, or the system model. It is load-bearing: the step from per-sample bounds (32) and (33) to the summed client bound (34) is exactly where the federated averaging structure is handled via [29, Theorem 4], and without the Bregman condition the change-of-measure argument does not produce the stated bound. The experiments in Sections VI-A and VI-B use cross-entropy losses, and the paper gives no argument that cross-entropy satisfies the Bregman condition. The theorem and the corollaries derived from it (Eqs. (15)–(17), Proposition 1) are therefore not established for the losses used in the paper. Please state the Bregman condition in Theorem 1 and either restrict the theoretical claims to Bregman losses or prove that the cross-entropy losses used in the experiments satisfy it.","section":"§IV-A and Appendix A, Eq. (12)"},{"comment":"Equation (13) is obtained by taking N to infinity and setting I(w_cur, Z) to zero, but it is introduced as an upper bound ('we can rewrite the bound as follows: G ≤ ...'). Because the omitted mutual-information terms are nonnegative, Eq. (13) is an asymptotic approximation of the bound in Theorem 1, not a finite-sample upper bound. Since Corollary 1 and Proposition 1 are derived from Eq. (13), the claimed ordering of periodic, gradual, and random patterns is only an asymptotic statement unless a finite-sample control of the mutual-information terms is provided. The experiments in Section VI-A use N=600 samples (Table I), so the verification in Figs. 7–10 does not directly test the proven finite-sample bound. Please restate Eq. (13) as an asymptotic result or add a finite-sample bound that accounts for the mutual-information terms.","section":"§IV-B, Eq. (13)"},{"comment":"Observation 3 states that, in contrast to the analytical result, no threshold p_th with B_p(p) > B_g(p) for p < p_th is observed in the CIFAR-10 experiments (Figs. 9 and 10). The paper attributes this to asymmetry of the KL divergence, but Proposition 1 is stated as a general result and Corollary 1 is derived under the symmetric Gaussian KL of Definition 3. The experiment therefore directly contradicts a stated theoretical claim. Please either make explicit the conditions (e.g., symmetric KL) under which Proposition 1 is claimed to hold, or revise the abstract and conclusion claims that the experiments corroborate the theoretical pattern ordering.","section":"§VI-A, Observation 3"}],"minor_comments":[{"comment":"The phrase 'Desipite the noncovexity' should read 'Despite the nonconvexity'.","section":"§I, Contributions"},{"comment":"The statement 'the function ψ*^{-1}(·) is concave' is asserted without proof and is not true for every Legendre dual arising in Theorem 1; the concavity claim should be stated under the sub-Gaussian or other specific ψ assumption used later.","section":"§IV-A, after Eq. (12)"},{"comment":"If Lemma 1 and the sub-Gaussian assumption give ψ*^{-1}(x) = sqrt(2 r^2 x), then Eq. (13) appears to be missing a factor of sqrt(2); please check the consistency of the factor across Eqs. (13), (15)–(17), and Figure 5.","section":"§IV-B, Eq. (13)"},{"comment":"The notation for the second sum in Eq. (2) and later in Theorem 1 is confusing: the upper limit is written as N while the normalization uses N_pre; please define N = N_cur + N_pre explicitly before Eq. (2) and keep the summation limits consistent.","section":"§III-A, Eq. (2)"},{"comment":"The tables report a single accuracy value per method without error bars or multiple seeds; the claim that the proposed method 'significantly' outperforms baselines would be strengthened by reporting variance across runs.","section":"§VI-B, Tables II–IV"}],"recommendation":"major_revision","confidential_remarks":"The main blocker is the Bregman-divergence assumption in the proof of Theorem 1; this is a correctness issue for the paper's central theoretical claim and for the experiments, which use cross-entropy losses. The large-N simplification in Eq. (13) is a second correctness issue for the corollaries. If the authors can qualify the theorem appropriately and either adapt the experiments to a Bregman loss or rigorously justify the cross-entropy case, the paper could become publishable. The asymptotic pattern-ordering claim and the observed contradiction in Observation 3 also need reconciliation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know up front. First, this paper gives the federated-learning community a real object to argue about: the Stationary Generalization Error, defined over a Markov chain of data distributions, plus closed-form bound comparisons across periodic, gradual, and random drift patterns. That is new and worth engaging. Second, the main theorem as stated is not actually proven for the losses used in the experiments. The proof in Appendix A invokes [29, Theorem 4] and quietly assumes the loss is a Bregman divergence. Theorem 1 and Assumption 1 never state that condition, and cross-entropy does not satisfy it. The change-of-measure step in Eq. (33) can fail badly for non-Bregman losses, so the bound is unproven for the paper's own central examples. That is more load-bearing than the N-to-infinity simplification in Section IV-B, though that issue is real too: Eq. (13) drops nonnegative mutual information terms and then calls the result an upper bound. It is an asymptotic approximation, and the corollaries inherit that status.\n\nWhat the paper does well: the Markov-chain formulation is a clean way to capture long-term behavior under drift, and the SGE metric sensibly combines current and future generalization. The bound itself is a natural extension of Wu et al., but the pattern-wise corollaries give concrete, checkable predictions about ordering. The KLMI-FedAvg algorithm follows from the bound's structure, which is good theory-to-algorithm transfer. The Raspberry Pi testbed is real hardware, not just simulation, and the parameter sweep over transition probability is the right kind of verification attempt.\n\nSoft spots, in proportion: the Bregman gap is serious because it undermines Theorem 1 for the paper's own experiments. The fix is to state the assumption openly and either prove the bound for a broader class of losses or restrict the claims accordingly. The large-N simplification needs an explicit finite-N error term before Eq. (13) can be called a bound. The experiments show mean curves only, with no error bars and no released code; for a theory-verification paper, that is a reproducibility gap. Minor point: the Pareto analysis in Section V is self-contained and looks correct, but it depends on the linear-in-tau2 bound, so it inherits the same caveats.\n\nWho this is for: researchers working on federated learning under concept drift, and anyone interested in information-theoretic generalization bounds. It deserves a serious referee, but not acceptance in current form. The gaps are addressable: state the Bregman assumption, provide a finite-N version of Eq. (13), and release code. If the authors do that, it becomes a solid contribution. I would send it out with a clear request for major revision.","headline":"Genuinely new SGE metric and pattern bounds, but Theorem 1 is unproven for the paper's own losses because the proof silently assumes Bregman divergence.","tokens_in":26462,"tokens_out":2016,"would_cite":false,"duration_ms":24906,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","94A17"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that the long-term performance of federated learning under concept drift is governed by an information-theoretic bound on a Stationary Generalization Error, and that this bound correctly orders the damage done by random…","keywords":["federated learning","concept drift","Stationary Generalization Error","Markov chain","mutual information","KL divergence","performance-cost tradeoff","information-theoretic generalization bound"],"falsifier":"Recompute the bounds in Eqs. (15)-(17) without setting $I(w_{\\mathrm{cur}},Z)$ to zero, using the $N=600$ sample size and the actual trained models from the testbed; if the mutual information terms are comparable to the KL terms, the predicted pattern ordering $B_r>B_p>B_g$ for small $p$ need not hold.","tokens_in":25383,"feed_emoji":"📉","tokens_out":8182,"duration_ms":84151,"temperature":0.7,"pith_summary":"This paper tries to establish an information-theoretic account of how concept drift degrades federated learning, and to turn that account into a usable training rule. It models the evolving data distribution as a Markov chain and defines the Stationary Generalization Error, the steady-state expected gap between population and empirical risk across a current and a future unseen distribution. Theorem 1 bounds this error by mutual information between the learned model and individual samples plus KL divergences between consecutive distributions; dropping the mutual information under a large-sample assumption yields closed-form bounds for periodic, gradual, and random drift (Corollary 1) and a threshold ordering in which random drift is worst, periodic intermediate, and gradual mildest when drift is drastic (Proposition 1). The paper then proposes an algorithm that regularizes empirical risk minimization with KL divergence on representation distributions and Gaussian noise, and reports that it outperforms four baselines on Fashion-MNIST and CIFAR-10 across all three patterns. The same bound is used to derive a closed-form, convex Pareto frontier between the Stationary Generalization Error and system cost.","feed_headline":"Where drift is drastic, random drift hurts FL the most","feed_subtitle":"A bound on FL's steady-state error orders drift patterns, and a KL-regularized retraining method beats baselines.","key_machinery":"The load-bearing object is the Stationary Generalization Error $G$, defined as the steady-state average, over a Markov chain of data distributions, of a weighted generalization error that mixes the model's fit to the current distribution (stage I) and its fit to the next unseen distribution (stage II). The argument runs through the upper bound of Theorem 1: using the Legendre dual $\\psi^{*-1}$ of the cumulant-generating function of the loss, each term of $G$ is bounded by $\\psi^{*-1}$ applied to a mutual information term plus a KL divergence between distribution states. With sub-Gaussian losses, $\\psi^{*-1}(x)=\\sqrt{2r^2x}$, and with the large-sample assumption $I(w_{\\mathrm{cur}},Z)\\to 0$, the bound collapses to a weighted sum of square-root KL divergences driven only by the Markov chain's transition probabilities. The periodic, gradual, and random drift patterns are encoded as three state machines, and the closed-form bounds of Corollary 1 come from substituting their stationary distributions into that simplified bound. The algorithm then acts on the same two knobs: KL regularization on representation distributions lowers the divergence terms, and injected Gaussian noise lowers the mutual information term.","core_discovery":"On its own terms, the paper's central claim is that the population risk of a federated model deployed under drifting data is not an unpredictable quantity but is controlled by two information-theoretic quantities: how much the retrained model $w_{\\mathrm{cur}}$ depends on each individual training sample (mutual information $I(w_{\\mathrm{cur}}, Z)$), and how far the previous, current, and next data distributions are from each other (KL divergences). Theorem 1 packages this in the Stationary Generalization Error $G$, averaged over the steady state of a Markov chain over distribution states. In the large-sample limit, the mutual information terms vanish and the bound becomes a sum of square-root KL distances weighted by transition probabilities; evaluating that sum for the three canonical drift state machines yields the explicit bounds in Corollary 1 and the threshold ordering in Proposition 1. The paper further claims that this analysis supports a practical training rule, Algorithm 1, that regularizes empirical risk minimization with symmetric KL matching of learned representations plus Gaussian noise, and that the same bound produces a closed-form convex Pareto frontier for the performance-cost tradeoff (Theorem 2).","pith_inferences":["Beyond the paper: the large-sample simplification is the load-bearing step; at the experimental size of $N=600$ samples per client, directly estimating $I(w_{\\mathrm{cur}},Z)$ on the trained models would show whether the closed-form ordering still holds or is only asymptotic.","If the ordering is correct, an FL operator can allocate adaptation effort by pattern: under random drift, increasing the number of clients gives the largest payoff, while under gradual drift it gives the least.","The representation-space KL penalty in Algorithm 1 is a concrete, testable way to lower the KL terms in Eq. (13); varying the symmetry of the penalty (one-sided versus two-sided KL) would map which direction of distributional shift matters more.","The convex Pareto frontier in Theorem 2 suggests that online controllers could tune client processing capacities in real time without nonconvex solvers, an application the paper gestures at but does not develop."],"forward_implications":["If Theorem 1 is correct, retraining algorithms can be compared by two quantities alone: how much the model depends on a single training sample ($I(w_{\\mathrm{cur}},Z)$) and how far consecutive data distributions are in KL divergence.","The simplified bound predicts that for small transition probability $p$ (drastic drift), random drift produces the largest stationary generalization error, periodic drift is intermediate, and gradual drift is the mildest; for larger $p$ the periodic ordering can reverse.","Proposition 2 directs system design: placing more weight on fresh data (larger $\\alpha$), using more clients (larger $K$), and shortening the adaptation window (smaller $\\tau_2$) all shrink the upper bound.","Theorem 2 gives a closed-form, convex Pareto frontier, so an operator can read off the minimum achievable Stationary Generalization Error for a given system-cost budget without solving a nonconvex problem directly.","Algorithm 1, by matching representation distributions with symmetric KL penalties and adding Gaussian noise, empirically reduces the accuracy drop at drift points on Fashion-MNIST and CIFAR-10 across all three patterns."],"supporting_citations":[{"why":"Supplies the mutual-information-based generalization bound and the Legendre-dual lemma that Theorem 1 builds on.","marker":"[20]"},{"why":"Extends the information-theoretic bound to transfer learning; the weighted empirical risk and KL terms in Eqs. (2)-(13) follow its setup.","marker":"[21]"},{"why":"Provides the distributed and federated version of the bound with Bregman loss that yields the $K^{-2}$ form and the assumptions used in the proof.","marker":"[29]"},{"why":"Justifies adding Gaussian noise to the loss to reduce mutual information between model and samples in Algorithm 1.","marker":"[32]"},{"why":"Motivates matching representation distributions with KL divergence, which Algorithm 1 uses for its regularization terms.","marker":"[35]"},{"why":"Supplies the FedAvg aggregation procedure used as the base training algorithm in the testbed and in Algorithm 1.","marker":"[36]"},{"why":"Defines the three drift patterns (periodic, gradual, random) whose state machines structure Corollary 1 and Proposition 1.","marker":"[6]"},{"why":"Provides the empirical risk minimization baseline and the objective that Algorithm 1 regularizes.","marker":"[22]"}],"fun_headline_variants":["Random drift hits federated learning hardest, new analysis finds","Information-theoretic bound predicts FL error under concept drift","KL regularization helps federated models survive concept drift","Drift patterns dictate FL performance, Pareto front emerges","Federated learning under drift: random is worst, theory says"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The ordering of drift patterns in the simplified bounds assumes the sample size is large enough that each single training sample's influence on the learned model, the mutual information term $I(w_{\\mathrm{cur}},Z)$, is zero; the experiments use 600 samples per client, so Eqs. (13)-(17) may not hold exactly at that size.","fun_headline_variants_meta":{"raw":{"variants":["Random drift hits federated learning hardest, new analysis finds","Information-theoretic bound predicts FL error under concept drift","KL regularization helps federated models survive concept drift","Drift patterns dictate FL performance, Pareto front emerges","Federated learning under drift: random is worst, theory says"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1550,"prompt_tokens":971,"completion_tokens":579,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":587,"completion_tokens_details":{"reasoning_tokens":500}},"tokens_in":587,"tokens_out":579,"duration_ms":7352,"temperature":1.0,"reasoning_tokens":500,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:35:49.742928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the bounds in Eqs. (15)-(17) without setting $I(w_{\\mathrm{cur}},Z)$ to zero, using the $N=600$ sample size and the actual trained models from the testbed; if the mutual information terms are comparable to the KL terms, the predicted pattern ordering $B_r>B_p>B_g$ for small $p$ need not hold.","supporting_citations":[{"cited_title":"Tightening mutual information- based bounds on generalization error,","cited_arxiv_id":null,"evidence_quote":"Supplies the mutual-information-based generalization bound and the Legendre-dual lemma that Theorem 1 builds on."},{"cited_title":"Information-theoretic analysis for transfer learning,","cited_arxiv_id":null,"evidence_quote":"Extends the information-theoretic bound to transfer learning; the weighted empirical risk and KL terms in Eqs. (2)-(13) follow its setup."},{"cited_title":"Improved information-theoretic generalization bounds for distributed, federated, and iterative learning,","cited_arxiv_id":null,"evidence_quote":"Provides the distributed and federated version of the bound with Bregman loss that yields the $K^{-2}$ form and the assumptions used in the proof."},{"cited_title":"Information-theoretic analysis of generaliza- tion capability of learning algorithms,","cited_arxiv_id":null,"evidence_quote":"Justifies adding Gaussian noise to the loss to reduce mutual information between model and samples in Algorithm 1."},{"cited_title":"KL guided domain adaptation,","cited_arxiv_id":null,"evidence_quote":"Motivates matching representation distributions with KL divergence, which Algorithm 1 uses for its regularization terms."},{"cited_title":"Communication-efficient learning of deep networks from decentralized data,","cited_arxiv_id":null,"evidence_quote":"Supplies the FedAvg aggregation procedure used as the base training algorithm in the testbed and in Algorithm 1."},{"cited_title":"Learning under concept drift: A review,","cited_arxiv_id":null,"evidence_quote":"Defines the three drift patterns (periodic, gradual, random) whose state machines structure Corollary 1 and Proposition 1."},{"cited_title":"Principles of risk minimization for learning theory,","cited_arxiv_id":null,"evidence_quote":"Provides the empirical risk minimization baseline and the objective that Algorithm 1 regularizes."}],"review_version":1}