{"id":"6ab7cfe8-9de7-4afb-8425-c3c41df3a0fb","arxiv_id":"2502.03958","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A federated optimization algorithm decouples proximal steps from communication, uses local updates and drift correction, and converges sublinearly or linearly to a bounded residual for non-convex composite losses with heterogeneous data.","lead":"A new federated learning algorithm for non-convex problems with non-smooth penalties lets each client run several local updates and send one vector per round, while a correction term prevents client drift. The paper proves convergence to a small neighborhood of the optimum under general non-convexity and linear convergence under a stronger condition, without assuming clients' data are similar.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Assumption 3.1's bounded-subgradient requirement drives the residual error in Theorems 3.5 and 3.6, making the guarantee vacuous for constraint-type regularizers; Lemma A.1 also needs a stronger step-size condition than it states.","rationale":"I followed the proof of Theorems 3.5 and 3.6 through Lemmas A.1 and A.2. The algebraic structure is internally consistent: the drift bound, the auxiliary-function descent, and the telescoping produce the stated constants, and the theorem-level step-size condition (13) is strong enough to cover the stricter step-size actually used in Lemma A.1's proof. The central algorithmic idea — sending the pre-proximal local model so the server recovers the average gradient without prox distortion — is coherent and is backed by the compact-form derivation in Appendix A.1. However, Assumption 3.1 is genuinely load-bearing: every place the proof needs control of subgradients of g, it uses the uniform bound Bg, and this is what creates the residual error. For indicators of convex sets, which are a standard nonsmooth composite term in constrained federated optimization, the bound is infinite and the theorems say nothing. The paper acknowledges this in Remark 3.7, so it is an honest limitation rather than a hidden flaw, but it narrows the force of the 'non-smooth composite FL' claim. I also verified the reader's observation that Lemma A.1's proof uses 8(t+1)η²L²≤1/(8τ), i.e. η̃≤ηg/(8L), which is stronger than the lemma's stated √20 condition; the main theorems are unaffected because (13) implies the stronger bound. These are revision-level issues, not reasons to reject the work, so the existing CONDITIONAL verdict is appropriate and I do not propose changing it.","tokens_in":23462,"tokens_out":13812,"duration_ms":142833,"concrete_test":"Run the full-gradient synthetic experiment of Fig. 2 with the same smooth non-convex f but with g replaced by the indicator of a bounded convex set such as {x:∥x∥≤1}, and record whether ∥G(Pη̃(x^r))∥ reaches machine precision or plateaus. If it plateaus, the Bg residual is an intrinsic algorithmic limit for constraint regularizers; if it reaches machine precision, the residual is an artifact of the proof, and the main question becomes whether the theorem can be re-proved without Assumption 3.1.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing condition is Assumption 3.1. It is used in Lemma A.1 at (A.13) to bound ∥∇g(˜x_pgd)−∇g(˜x_{r+1})∥ by 2Bg; this bound propagates through the drift estimate (A.8) into the server-side recursion and becomes the O(L²η̃²Bg²/ηg²) residual in Theorem 3.5 and O(L²η̃²Bg²/(μηg²)) residual in Theorem 3.6. For g equal to the indicator of a convex constraint set, Assumption 3.1 fails since subgradients on the boundary are unbounded, so the main theorems give no guarantee for a canonical composite FL setting. Remark 3.7 and the conclusion explicitly say the Bg term cannot be removed in the non-convex case, so this is a real scope restriction rather than an artifact. A separate precision issue is that Lemma A.1 is stated under η̃≤ηg/√(20L), but the proof of (A.20) uses 8(t+1)η²L²≤1/(8τ), which is equivalent to η̃≤ηg/(8L); condition (13) in the theorems does imply the stricter bound, so the main claim survives, but the lemma as written is not fully proven.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the composite federated learning problem (1) with a non-convex smooth loss f and a convex, possibly non-smooth regularizer g. It proposes Algorithm 1, in which each client maintains both a pre-proximal and a post-proximal local model, performs τ local updates using mini-batch stochastic gradients plus a client-drift correction term, and transmits only the pre-proximal model to the server. The server then updates a pre-proximal global model so that the server-side recursion mimics centralized proximal gradient descent. Under Assumptions 3.1–3.4 and step-size condition (13), the paper proves an average bound on the gradient mapping norm for general non-convex f (Theorem 3.5) and a linear convergence bound for the function value under the proximal PL inequality (Theorem 3.6), both with explicit residual errors depending on the stochastic variance and on the subgradient bound Bg. Numerical experiments on sparse logistic regression and CNN training on MNIST support the algorithmic claims.","tokens_in":23740,"tokens_out":8618,"duration_ms":75365,"significance":"If the results hold, Algorithm 1 is a meaningful step for composite federated learning: it decouples proximal evaluation from communication, requires only one d-dimensional vector per client per communication round, does not assume bounded data heterogeneity, and handles non-smooth regularizers. The convergence analysis is non-trivial and the appendix contains a full Lyapunov-style proof with explicit lemmas. The paper is also candid about its limitations, notably the residual term depending on Bg. The main concerns are a step-size condition mismatch in the appendix and the restrictive bounded-subgradient assumption on g; both are local and fixable, and the central algorithmic idea appears sound.","major_comments":[{"comment":"Lemma A.1 is stated under the condition η̃ ≤ ηg/√(20L), but the proof of (A.20) uses the inequality 8(t+1)η²L² ≤ 1/(8τ), which is equivalent to η̃ ≤ ηg/(8L). For typical values of L this is a stricter condition than the one stated, so the lemma as written is not fully proven. Lemma A.2 has the same issue and also contains a typographical garble in its stated condition (\"η̃ ≤ ηg√(20L)\"). The step-size condition (13) in Theorems 3.5 and 3.6 does imply the stricter bound, so the main results are not invalidated, but the lemma statements must be corrected and Lemma A.2's hypothesis should match the proof requirements.","section":"Appendix A.4, Lemma A.1 and Lemma A.2"},{"comment":"Assumption 3.1 requires the subgradient of g to be uniformly bounded by Bg. For many canonical composite problems, g is the indicator function of a convex constraint set, for which subgradients on the boundary are unbounded, making Bg infinite and the residual terms 187L²η̃²Bg²/ηg² (Theorem 3.5) and 168L²η̃²Bg²/(μηg²) (Theorem 3.6) vacuous. The authors acknowledge this limitation in Remark 3.7 and the conclusion, but the abstract and introduction present the result as a general treatment of non-convex composite FL. I recommend adding an explicit qualification in the abstract and a short discussion of settings where Bg is finite (e.g., ℓ1, ℓ∞, ReLU) and of possible avenues for removing the Bg residual under additional assumptions.","section":"Assumption 3.1 and Theorems 3.5–3.6"}],"minor_comments":[{"comment":"In the compact form (2), the two sums in the correction term are typeset identically, although one is the averaged gradient across clients and the other is the per-client gradient; please add the overbar or a textual clarification to avoid confusion.","section":"Equation (2) and Appendix A.1"},{"comment":"The text reads \"The remaining 3,0000 samples\" and should be \"The remaining 30,000 samples\".","section":"Section 4.2"},{"comment":"The stated condition in Lemma A.2 appears as \"η̃ ≤ ηg√(20L)\" but should presumably be \"η̃ ≤ ηg/√(20L)\", and in any case must be updated to the stricter condition used in the proof.","section":"Appendix A.4, Lemma A.2"},{"comment":"The phrase \"sends only a single d-dimensional vector per communication round\" is not fully precise because the server also broadcasts a d-dimensional vector; the per-client communication cost is correct, but the wording in the abstract could be clarified.","section":"Notation and introduction"},{"comment":"The proof of Lemma A.1 cites \"[22, Corollary C.1]\" for an inequality on the variance of averaged stochastic gradients; please verify that the cited result indeed appears in reference [22] and that the reference is complete.","section":"Reference [22]"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the central contribution is sound. The main issue is the step-size condition mismatch in Lemma A.1 (and Lemma A.2), which is a proof gap in a load-bearing lemma, even though the theorem-level condition covers the stricter bound. The bounded-subgradient limitation is acknowledged by the authors and is acceptable as a scope restriction, but it should be made more visible. I recommend a major revision to correct the lemma statements and to sharpen the presentation of the assumption's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid theory paper with a real algorithmic idea, and the main convergence claims survive scrutiny. The authors decouple proximal evaluation from communication by having clients send the pre-proximal model to the server, so the server can recover average gradients without the nonlinear distortion of the proximal map. They add a correction term built from the broadcast global model to cancel client drift while exchanging only one d-dimensional vector per client per round. That design is genuinely useful for sparse and regularized federated training. Relative to FedDR, FedADMM, FedDA, and Fast-FedDA, the non-convex and proximal-PL analyses for this decoupled local-update scheme are new; the strongly convex case appeared in their earlier ICASSP paper, and they cite it. The appendix gives a full Lyapunov proof with explicit lemmas and constants. Credit where it is earned: this is a substantive, reproducible theoretical contribution.\n\nSoft spots, in proportion. Assumption 3.1 requires a uniform bound on subgradients of the regularizer. That covers l1 and ReLU but fails for indicator functions of constraint sets, so the main theorems give no guarantee in a canonical composite setting. The residual error in Theorems 3.5 and 3.6 scales with Bg^2, and the authors explicitly say they cannot remove it in the non-convex case. That is an honest scope restriction, not a fatal flaw. Second, Lemma A.1 states step-size tilde eta <= eta_g / sqrt(20 L), but the proof at (A.20) needs tilde eta <= eta_g / (8 L). The theorem's global condition (13) implies the stricter bound, so the main results are unaffected; the lemma as written is imprecise and should be fixed. Third, the numerical section is thin: no error bars, no FedDR/FedADMM baselines, no code. The experiments show the algorithm works and that the Bg residual does not show up in practice, but the word 'superiority' overstates the evidence.\n\nThe citation pattern looks fair: they position themselves against FedDA, Fast-FedDA, FedDR, FedADMM, and their own conference paper. The math is not machine-checked, and I did not verify every inequality, but the structure of the proof is coherent and the claimed rates match the method.\n\nWho is this for: researchers working on communication-efficient composite federated learning with non-smooth regularizers. The proof machinery and the decoupling trick are worth studying. My recommendation: yes, send this to peer review. It deserves a serious referee. Require fixing Lemma A.1's condition, tightening the empirical claims, and adding at least one constraint-set experiment to confront the Bg limitation.","headline":"A solid non-convex composite FL paper with one repairable lemma inconsistency and an honest Bg residual; worth full peer review.","tokens_in":24282,"tokens_out":1982,"would_cite":true,"duration_ms":19513,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","90C15","68W15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sending the pre-proximal vector to the server recovers the true average gradient and removes client drift in composite federated learning.","keywords":["non-convex composite optimization","federated learning","heterogeneous data","client drift","proximal operator","local updates","proximal Polyak-Lojasiewicz inequality","communication efficiency"],"falsifier":"A direct calculation settles the internal consistency: the proof of Lemma A.1 invokes $8(t+1)\\eta^2 L^2 \\leq 1/(8\\tau)$, which requires $\\tilde{\\eta} \\leq \\eta_g/(8L)$, stricter than the lemma's stated $\\tilde{\\eta}\\leq \\eta_g/\\sqrt{20L}$; the theorem's condition (13) with $\\eta_g\\ge 1.5$ and $\\tilde{\\eta}\\le 1/(10L)$ implies the stricter bound, so the main theorems survive but the lemma statement is imprecise. Separately, run Algorithm 1 with full local gradients on a non-convex composite problem where $g$ is the indicator function of a box, so Assumption 3.1 fails and $B_g=\\infty$, and check whether $\\|G(P_{\\tilde{\\eta}}(x^r))\\|^2$ goes to machine zero; exact convergence would show the $B_g$ residual is an artifact of the proof.","tokens_in":23254,"feed_emoji":"📡","tokens_out":16256,"duration_ms":137562,"temperature":0.7,"pith_summary":"The paper proposes a federated algorithm for objectives of the form $F(x)=f(x)+g(x)$, where $f$ is smooth and possibly non-convex, averaged across clients, and $g$ is convex and non-smooth, such as an $\\ell_1$ penalty. Each client runs $\\tau$ local proximal-gradient steps while maintaining two quantities, a pre-proximal model and a post-proximal model, and transmits only the final pre-proximal model to the server. The paper claims that this decouples the proximal operation from communication: the transmitted vectors carry the average of the local gradients, so the server can mimic centralized proximal gradient descent without distortion, and a locally reconstructed correction term cancels client drift without extra communication. Under a global bound on the subgradients of $g$, it proves sublinear convergence to a bounded residual for general non-convex problems and linear convergence to a bounded residual when the composite objective satisfies the proximal Polyak–Łojasiewicz inequality. If the analysis holds, composite federated learning no longer has to choose between sparse regularizers, local updates, and low communication cost.","feed_headline":"One vector per round ends client drift in federated learning","feed_subtitle":"Sending models before the proximal step lets servers recover true gradients; rates proven for non-convex losses.","key_machinery":"The argument is carried by three linked devices. The first is the two-model bookkeeping: only the pre-proximal local model is transmitted, so the server sees the sum of gradients before the regularizer's proximal operator acts on them. The second is the correction term $c^{r+1}_i = \\frac{1}{\\eta_g\\eta\\tau}(P_{\\tilde{\\eta}}(x^r)-x^{r+1}) - \\frac{1}{\\tau}\\sum_{t=0}^{\\tau-1}\\nabla f_i(z^r_{i,t};B^r_{i,t})$, which each client reconstructs from the broadcast global model; its average over clients is zero, which is exactly what keeps the transmitted average undistorted. The third is the auxiliary function $\\Omega^r = F(P_{\\tilde{\\eta}}(x^r))-F^\\star + \\frac{1}{n\\tilde{\\eta}}\\|\\Lambda^r - \\bar{\\Lambda}^r\\|^2$, whose first term tracks global suboptimality and whose second term bounds the local-drift error; Lemma A.1 and Lemma A.2 close the recursion on this drift term. The parameter $(t+1)\\eta$ inside the local proximal operator is chosen so that the post-proximal local iterates approximate a centralized proximal gradient step on the global model, which is what makes the drift bound small enough to close.","core_discovery":"The load-bearing claim is that each client keeps two local models at every step, the pre-proximal variable $\\hat{z}^r_{i,t}$ and the post-proximal variable $z^r_{i,t}=P_{(t+1)\\eta}(\\hat{z}^r_{i,t})$, evaluates stochastic gradients at the post-proximal model, and after $\\tau$ updates sends only $\\hat{z}^r_{i,\\tau}$ to the server. Because the drift-correction terms $c^r_i$ average to zero across clients, the server's update becomes $x^{r+1}=P_{\\tilde{\\eta}}(x^r)-\\tilde{\\eta}\\,\\frac{1}{n\\tau}\\sum_{t=0}^{\\tau-1}\\sum_{i=1}^n\\nabla f_i(z^r_{i,t};B^r_{i,t})$, which is a centralized proximal gradient step affected only by stochastic sampling, with no distortion from the nonlinear proximal map. This is the decoupling: the server can run a proximal step on the global model using the true average local gradient. The convergence analysis measures progress with the gradient mapping $G(P_{\\tilde{\\eta}}(x^r))=\\frac{1}{\\tilde{\\eta}}(P_{\\tilde{\\eta}}(x^r)-\\tilde{x}^{r+1})\\in \\nabla f(P_{\\tilde{\\eta}}(x^r))+\\partial g(\\tilde{x}^{r+1})$, and proves $\\frac{1}{R}\\sum_{r=1}^R \\mathbb{E}\\|G(P_{\\tilde{\\eta}}(x^r))\\|^2 \\leq \\frac{\\mathbb{E}[\\Omega^1]}{0.3\\tilde{\\eta}R} + \\frac{20\\sigma^2}{n\\tau b} + \\frac{187L^2\\tilde{\\eta}^2 B_g^2}{\\eta_g^2}$ for general non-convex problems, and $\\mathbb{E}[\\Omega^{R+1}] \\leq (1-\\frac{\\mu\\tilde{\\eta}}{3})^R \\mathbb{E}[\\Omega^1] + \\frac{18\\sigma^2}{\\mu n\\tau b} + \\frac{168L^2\\tilde{\\eta}^2 B_g^2}{\\mu\\eta_g^2}$ under the proximal PL inequality. These are stated as Theorem 3.5 and Theorem 3.6, with residuals set by stochastic variance and by the squared subgradient bound of the regularizer.","pith_inferences":["The pre-proximal transmission idea may extend to other nonlinear client-side maps, such as variance-reduced proximal operators or adaptive-regularizer steps, whenever the server needs the average of a pre-image rather than a post-image.","The $B_g^2$ residual likely comes from the two-sided subgradient bound used in Lemma A.1; a drift bound that exploits one-sided or structure-dependent properties of $g$ could remove or shrink it for specific regularizers.","The proof of Lemma A.1 requires the stricter condition $\\tilde{\\eta} \\leq \\eta_g/(8L)$ at the step where it uses $8(t+1)\\eta^2 L^2 \\leq 1/(8\\tau)$, while the lemma states only $\\tilde{\\eta} \\leq \\eta_g/\\sqrt{20L}$; the theorem's global condition (13) appears to imply the stricter bound when $\\eta_g\\ge 1.5$, so the main theorems survive, but the lemma as written is imprecise.","Because the correction term uses the previous round's global gradient information, the drift correction is inherently one round stale; a testable variant with momentum or variance reduction over older gradients could shrink the residual at no extra communication cost."],"forward_implications":["A single $d$-dimensional vector per client per communication round is enough to remove client drift for non-convex composite objectives; the correction term is reconstructed locally from the broadcast global model, so no control-variate communication is needed.","Communication frequency can be reduced through local updates while the method keeps a per-round convergence guarantee; the variance term $\\sigma^2/(n\\tau b)$ shrinks with larger mini-batches and vanishes when full local gradients are used.","Under the proximal PL inequality the composite loss converges linearly to a neighborhood whose radius is set by the stochastic variance and by $B_g^2$, the squared subgradient bound of the regularizer.","If the regularizer has unbounded subgradients, such as an indicator function of a constraint set, the stated residual bounds are vacuous; the paper explicitly notes in Remark 3.7 that it cannot remove the $B_g$ term in the non-convex case.","Full-gradient experiments on sparse logistic regression show exact convergence with no visible $B_g$ residual, suggesting that the theoretical residual is conservative rather than a fundamental obstruction."],"supporting_citations":[{"why":"Introduces FedAvg and the local-update paradigm of running multiple steps per client before communicating; the proposed algorithm inherits this communication-reduction design.","marker":"[19]"},{"why":"Introduces control-variate drift correction for smooth FL; the paper's correction term achieves the same goal without transmitting the control variates.","marker":"[13]"},{"why":"Defines composite FL baselines FedMid and FedDA and identifies the 'curse of primal averaging' that the pre-proximal transmission is designed to overcome.","marker":"[29]"},{"why":"Provides the proximal gradient-mapping measure and the descent inequality (Fact 1) used to prove both non-convex convergence theorems.","marker":"[10]"},{"why":"Defines the proximal Polyak–Łojasiewicz inequality that appears as Assumption 3.3(2) and yields the linear-convergence result.","marker":"[11]"},{"why":"Presents FedDR, the main prior method for non-convex composite FL, whose need for increasingly accurate local prox subproblems motivates the paper's decoupling strategy.","marker":"[25]"},{"why":"Shows a control-variate solution for smooth FL; the comparison clarifies that the proposed method gets the drift correction with one vector per round.","marker":"[12]"}],"fun_headline_variants":["Decoupling proximal and communication cuts client drift","Single-vector updates decouple proximal and server in FL","Proven rates for non-convex FL with local updates and decoupling","Heterogeneous data no longer a barrier: one vector per round"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The regularizer $g$ must have globally bounded subgradients (norm at most $B_g$), which fails for constraint-set indicator functions and makes the stated residual error infinite.","fun_headline_variants_meta":{"raw":{"variants":["Decoupling proximal and communication cuts client drift","Single-vector updates decouple proximal and server in FL","Proven rates for non-convex FL with local updates and decoupling","Heterogeneous data no longer a barrier: one vector per round"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000305,"raw_usage":{"total_tokens":1869,"prompt_tokens":1184,"completion_tokens":685,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":800,"completion_tokens_details":{"reasoning_tokens":616}},"tokens_in":800,"tokens_out":685,"duration_ms":6973,"temperature":1.0,"reasoning_tokens":616,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T00:07:02.595573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct calculation settles the internal consistency: the proof of Lemma A.1 invokes $8(t+1)\\eta^2 L^2 \\leq 1/(8\\tau)$, which requires $\\tilde{\\eta} \\leq \\eta_g/(8L)$, stricter than the lemma's stated $\\tilde{\\eta}\\leq \\eta_g/\\sqrt{20L}$; the theorem's condition (13) with $\\eta_g\\ge 1.5$ and $\\tilde{\\eta}\\le 1/(10L)$ implies the stricter bound, so the main theorems survive but the lemma statement is imprecise. Separately, run Algorithm 1 with full local gradients on a non-convex composite problem where $g$ is the indicator function of a box, so Assumption 3.1 fails and $B_g=\\infty$, and check whether $\\|G(P_{\\tilde{\\eta}}(x^r))\\|^2$ goes to machine zero; exact convergence would show the $B_g$ residual is an artifact of the proof.","supporting_citations":[{"cited_title":"Communication- efficient learning of deep networks from decentralized data","cited_arxiv_id":null,"evidence_quote":"Introduces FedAvg and the local-update paradigm of running multiple steps per client before communicating; the proposed algorithm inherits this communication-reduction design."},{"cited_title":"Scaffold: Stochastic controlled averaging for federated learning","cited_arxiv_id":null,"evidence_quote":"Introduces control-variate drift correction for smooth FL; the paper's correction term achieves the same goal without transmitting the control variates."},{"cited_title":"Federated composite optimization","cited_arxiv_id":null,"evidence_quote":"Defines composite FL baselines FedMid and FedDA and identifies the 'curse of primal averaging' that the pre-proximal transmission is designed to overcome."},{"cited_title":"Proximal stochastic methods for nonsmooth nonconvex finite-sum optimization","cited_arxiv_id":null,"evidence_quote":"Provides the proximal gradient-mapping measure and the descent inequality (Fact 1) used to prove both non-convex convergence theorems."},{"cited_title":"Linear convergence of gradient and proximal-gradient methods under the Polyak- Lojasiewicz condition","cited_arxiv_id":null,"evidence_quote":"Defines the proximal Polyak–Łojasiewicz inequality that appears as Assumption 3.3(2) and yields the linear-convergence result."},{"cited_title":"FedDR–randomized Douglas-Rachford splitting algorithms for nonconvex federated composite optimization","cited_arxiv_id":null,"evidence_quote":"Presents FedDR, the main prior method for non-convex composite FL, whose need for increasingly accurate local prox subproblems motivates the paper's decoupling strategy."}],"review_version":1}