{"id":"0fb3b489-baf7-4128-bc05-9b1048129f87","arxiv_id":"1908.09730","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"The DPLMS algorithm, a diffusion network of probabilistic LMS filters, shows lower steady-state error and faster convergence than DSE-LMS, DRVSSLMS, and DLLAD in impulsive-noise simulations.","lead":"This paper combines two existing adaptive filtering ideas, probabilistic LMS and diffusion networks, into a new algorithm for distributed estimation under impulsive noise. The authors claim their DPLMS algorithm recovers an unknown signal more accurately and robustly than three existing algorithms, and they support this with simulations.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The variance recursion in Table 1 uses sigma^2(i-1)-sigma_rho^2, whose fixed point is negative, so the published DPLMS algorithm becomes undefined during the 4000-iteration simulations.","rationale":"The paper's central claim is that the DPLMS algorithm, as defined in Table 1 and Eq. (16), is more robust than DSE-LMS, DRVSSLMS, and DLLAD. For that claim to be meaningful, the published recursion must define a stable step size over the 4000 iterations. The minus sign in the variance update is not a parameter choice; it is an algebraic inconsistency with the Gaussian state model in Eq. (2), and it drives the variance negative. A negative variance has no probabilistic interpretation and makes alpha invalid. The absence of sigma_rho^2 and sigma^2(0) values, and of any clipping, means the simulations cannot be reproduced or audited. I therefore regard this as the single most load-bearing concern. The reader's weakest assumption was the independence E{S(i)R(i)W(i-1)} = E{S(i)R(i)}E{W(i-1)} in Eq. (21). That concern is real but less decisive: under Assumption 1, alpha_n(i) depends on current and past regressors but not on W(i-1), so S(i)R(i) is independent of W(i-1). The more serious issue is the variance recursion itself. The reader's CONDITIONAL verdict remains appropriate: a sign correction, complete parameter disclosure, and reproducible code could resolve the concern. If the sign error is confirmed and no code is provided, the empirical robustness claim would be unsupported.","tokens_in":9489,"tokens_out":16714,"duration_ms":165906,"concrete_test":"Run the exact Table 1 recursion for one node with M=16, zero-mean unit-variance Gaussian regressors, sigma_epsilon^2=0.01, sigma_rho^2=0.001, and initial sigma^2(0)=1, with no clipping, for 4000 iterations; record the first iteration at which sigma_n^2(i)<=0. If it occurs before 4000, the published DPLMS algorithm is undefined in the reported simulation regime. As an analytical check, solve sigma^2 = c(sigma^2 - sigma_rho^2) for c in (0,1) and confirm the fixed point is negative.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central defect is the variance recursion that defines the step size. Under the state model in Eq. (2), the predictive variance in Eq. (6) should be sigma^2(i-1)+sigma_rho^2, but Eq. (6), Eq. (9), and Table 1 all use sigma^2(i-1)-sigma_rho^2. With this minus sign, Table 1 iterates sigma_n^2(i) = c_n(i) [sigma_n^2(i-1)-sigma_rho^2], where c_n(i)=1-alpha_n(i)||X_n(i)||^2/L. For the intended operating regime c_n(i) is between 0 and 1; the only fixed point of this recursion is negative, sigma^2* = -c sigma_rho^2/(1-c) < 0, so sigma_n^2 crosses zero in finite time from any positive initial value. If c_n(i)<0, sigma_n^2 is negative immediately. Once sigma_n^2 is negative, alpha_n(i) is no longer a valid step size and the update in Eq. (16) is undefined or diverging. The paper never states sigma_rho^2, the initial sigma_n^2(0), or any clipping floor, so the published algorithm cannot be the one that produced the MSD curves in Section 4. This is more load-bearing than the independence step in Eq. (21), because it concerns the algorithm itself rather than an approximation in its mean analysis.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed adaptive filtering algorithm, diffusion probabilistic LMS (DPLMS), which combines the adapt-then-combine diffusion strategy with the probabilistic least-mean-squares (PLMS) algorithm. The authors derive the local update from a Gaussian posterior approximation, introduce a constant parameter tau, and present a network-level two-step adaptation-combination scheme. The paper claims that DPLMS is more robust to input signal characteristics and impulsive interference than DSE-LMS, DRVSSLMS, and DLLAD, and it provides a mean-stability analysis and a computational-complexity comparison. The empirical section reports network MSD curves for several impulsive-noise scenarios with 60 Monte Carlo runs.","tokens_in":9798,"tokens_out":6374,"duration_ms":63573,"significance":"If the proposed algorithm worked as claimed, it would be a useful addition to distributed estimation under impulsive noise, with a step-size mechanism derived from a probabilistic model and with complexity equal to that of DLLAD and lower than that of DRVSSLMS. The paper's combination of a probabilistic step-size with diffusion is an interesting idea, and the complexity table is a useful summary. However, the current manuscript contains a fundamental error in the variance recursion that defines the algorithm, and the mean-stability analysis relies on an unjustified independence assumption. As a result, the algorithm as published is not well-defined for the reported simulations, and the main theoretical guarantee is not established. The empirical robustness claim is also not supported by the reported evidence because essential parameter values and statistical error information are missing. The paper would require substantial correction and re-analysis before its central claims can be accepted.","major_comments":[{"comment":"The variance recursion contains a sign error that makes the published algorithm undefined during the reported simulations. Under the state model in Eq. (2), the predictive variance of W_o(i) given Z^{i-1} is sigma^2(i-1) + sigma_rho^2, not sigma^2(i-1) - sigma_rho^2. This minus sign is then used in Eq. (9) and in Table 1's alpha_n(i) and sigma_n^2(i) update. The iteration sigma_n^2(i) = c_n(i) [sigma_n^2(i-1) - sigma_rho,n^2] with 0 < c_n(i) < 1 has only the negative fixed point -c_n(i) sigma_rho,n^2 / (1 - c_n(i)), so sigma_n^2(i) becomes negative in finite time from any positive initial value. Once sigma_n^2(i) is negative, alpha_n(i) is no longer a valid step size and the update in Eq. (16) is undefined or diverging. The paper never states sigma_rho,n^2, sigma_epsilon,n^2, the initial sigma_n^2(0), or any clipping floor, so the algorithm that produced the MSD curves in Section 4 cannot be reproduced. This error must be corrected and the simulations re-run.","section":"Section 2.1, Eq. (6), Eq. (9), and Table 1"},{"comment":"The mean-stability derivation is not valid as written. The step E{A^T [I - S(i)R(i)] W~(i-1)} = A^T [I - E{S(i)R(i)}] E{W~(i-1)} in Eq. (21) requires S(i)R(i) to be independent of W~(i-1). But alpha_n(i) depends on sigma_n^2(i-1), which is a deterministic function of past regressors through the variance recursion in Table 1, and W~(i-1) is correlated with those past regressors. Therefore the independence assumption is unjustified, and the stability condition in Eq. (22) does not follow. Additionally, Eq. (22) still contains the stochastic alpha_l(i) inside the bound, so it is not a checkable step-size condition without further expectation or concentration arguments. The paper's only theoretical guarantee for DPLMS is therefore unsupported.","section":"Section 3.1, Eqs. (21) and (22)"},{"comment":"The central robustness claim is not supported by the reported evidence. The simulations do not list the values of sigma_rho,n^2, sigma_epsilon,n^2, the initial sigma_n^2(0), or the constant tau, all of which are needed to run DPLMS. The MSD curves are averaged over 60 runs, but no error bars, confidence intervals, or run-to-run spread are shown, so the claim that DPLMS is more robust than DSE-LMS, DRVSSLMS, and DLLAD rests on visual inspection of single curves. The text in Experiments 1 and 2 also refers to 'DNLMS' where DPLMS is presumably intended, further complicating interpretation. These omissions are load-bearing because the empirical comparison is the main evidence for the paper's headline claim.","section":"Section 4"}],"minor_comments":[{"comment":"The 'global cost function' in Eq. (15) is not well defined: it places argmax_{W_o(i)} p(W_o(i)|Z_i) inside a sum of local costs and is never used to derive the adaptation and combination updates in Eqs. (16) and (17). Please rewrite or remove this part to avoid implying that DPLMS minimizes a well-defined global objective.","section":"Section 2.2, Eq. (15)"},{"comment":"Equation (7) omits the Bayes normalization and confuses the predictive and posterior distributions. A cleaner derivation would state the exact Gaussian posterior before the isotropic approximation and then apply the trace-based variance update.","section":"Section 2.1, Eq. (7)"},{"comment":"The notation is inconsistent: the noise variance is called sigma_p^2 in Eq. (2) and sigma_rho^2 afterward, and the matrix definitions in Section 3.1 contain garbled expressions (e.g., 'M C C I' and the definition of O(i)). These should be corrected for readability.","section":"Section 2.1 and Section 3.1"},{"comment":"The complexity comparison would be clearer if the entries for DRVSSLMS were explained: the '>' markers and the roles of the sgn and absolute-value columns are not defined, and the total counts should be consistent with the equations cited.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The variance-sign error is likely a typo, but it propagates into every simulation and makes the current algorithm undefined; this alone justifies a major revision. The independence assumption in Eq. (21) is also a serious gap in the theoretical contribution. I would be willing to consider a revised version that corrects the variance recursion, reports all simulation parameters (including a variance floor if one is used), and either proves the mean stability under a stated assumption or removes the theoretical claim. If the authors cannot provide the missing parameters or the re-run simulations, the paper should not be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on 1908.09730. The core idea is a plausible mashup: take the probabilistic LMS (PLMS) variable step-size rule and run it through ATC diffusion. That combination is new in the literature, and the computational complexity table is straightforward and honest. But the algorithm as written is not defined, and the supporting theory does not do the work the authors claim.\n\nThe serious problem is in the variance recursion that sets the step size. Under the state model in Eq. (2), the predictive variance should be sigma^2(i-1)+sigma_rho^2. Eq. (6), Eq. (9), and Table 1 all use a minus sign instead. With the minus sign, sigma^2(i) = c(i)[sigma^2(i-1)-sigma_rho^2], which has a negative fixed point. The variance crosses zero in finite time, alpha becomes negative, and the update in Eq. (16) diverges. The paper never specifies sigma_rho^2, the initial variance, or a floor, so the published algorithm cannot have generated the MSD curves in Section 4. This is not a nitpick; it is load-bearing. Either the sign is a typo and the simulations used a different recursion, or the simulations are not reproducible. Either way, the paper needs a major revision.\n\nSeparately, the mean-stability analysis in Eq. (21) assumes S(i)R(i) is independent of W-hat(i-1), which is not true because alpha depends on the regressor norm and past errors. That assumption is unstated and unverified. The stability bound in Eq. (22) is therefore not established. The simulations also lack error bars and key parameter values (mu for DPLMS, sigma_rho^2, sigma^2(0)), so I can't take the robustness claim at face value.\n\nWhat's worth keeping: the recipe of combining PLMS with diffusion is reasonable, and the complexity analysis is clean. The idea that a probabilistic step-size rule could help in impulsive noise is worth exploring. But as submitted, the paper is under-specified and internally inconsistent.\n\nWho is this for? Someone working on distributed adaptive filtering in impulsive noise might find the basic idea useful, but they'd have to reverse-engineer the algorithm to test it. I would not bring it to a reading group, and I would not cite it in its current form. It deserves a serious referee only because the combination is new and the flaw is potentially fixable; a referee could ask for a corrected recursion and a proper convergence proof. My recommendation: send it to review if you think the authors can repair the sign error and provide reproducible simulations; otherwise desk reject. I'd lean toward rejecting in current form but allowing a resubmission.","headline":"A plausible PLMS-plus-diffusion recipe, but the published algorithm has a sign error in its variance recursion that makes it undefined, and the simulations are not reproducible.","tokens_in":10313,"tokens_out":2838,"would_cite":false,"duration_ms":25822,"reading_group":"no","serious_thinker":"no","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 distributed estimation algorithm whose step size is derived from a Gaussian posterior approximation is more robust to input-signal statistics and impulsive interference than three established diffusion LMS variants.","keywords":["distributed estimation","diffusion strategy","probabilistic LMS","impulsive interference","variable step size","adaptive filtering","system identification","mean stability"],"falsifier":"Simulate a two-node network with strongly correlated inputs or a node whose step size is driven by large past errors, compute the Monte Carlo average of $E[\\mathbf{S}(i)\\mathbf{R}(i)\\hat{\\mathbf{W}}(i-1)]$ and compare it with $E[\\mathbf{S}(i)\\mathbf{R}(i)]E[\\hat{\\mathbf{W}}(i-1)]$; if the difference is large and the mean weight error grows even when $\\mu$ satisfies Eq. (22), the independence assumption is what failed.","tokens_in":9268,"feed_emoji":"📡","tokens_out":5895,"duration_ms":53025,"temperature":0.7,"pith_summary":"The paper proposes diffusion probabilistic LMS (DPLMS), a distributed adaptive filtering algorithm in which each node estimates a common unknown vector by combining an adapt-then-combine diffusion strategy with a probabilistic variable step-size LMS update. The step size at every node is not fixed but is recalculated from an isotropic Gaussian approximation of the posterior distribution of the unknown coefficients, so it shrinks when the input power is large and adapts to the measurement-noise variance. The paper's central claim is that this makes DPLMS more robust to input signal statistics and impulsive interference than three established robust diffusion algorithms: DSE-LMS, DRVSSLMS, and DLLAD. Simulation results on a 20-node system identification task are presented for several input-signal and Bernoulli-Gaussian impulse-noise settings. A mean-stability condition and a computational-complexity comparison are also derived.","feed_headline":"Bayesian step size beats three rivals under impulsive noise","feed_subtitle":"DPLMS adapts each node's learning rate to input power and noise, cutting steady-state error.","key_machinery":"The load-bearing device is the time-varying step size $\\alpha_n(i) = \\frac{\\sigma_n^2(i-1) + \\sigma_\\rho^2}{[\\sigma_n^2(i-1) - \\sigma_\\rho^2]\\|\\mathbf{X}_n(i)\\|^2 + \\sigma_{\\varepsilon,n}^2}$ together with the variance recursion $\\sigma_n^2(i) = \\left[1 - \\frac{\\alpha_n(i)\\|\\mathbf{X}_n(i)\\|^2}{L}\\right][\\sigma_n^2(i-1) - \\sigma_\\rho^2]$. This $\\alpha_n(i)$ is the PLMS step size obtained by approximating the posterior $p(\\mathbf{W}_o(i)|\\mathcal{Z}_i)$ as an isotropic Gaussian and taking the MAP estimate; it automatically down-weights large-input-power updates and nodes with high noise variance. Combining this update through the adapt-then-combine diffusion step is what produces the claimed robustness.","core_discovery":"The central claim is that replacing the fixed step size that appears in existing diffusion LMS variants with the PLMS variable step size—defined from the variance of a Gaussian posterior—yields a distributed estimator that converges faster and reaches lower steady-state misalignment under both correlated input signals and impulsive interference. The paper argues this by constructing the DPLMS recursion, proving a sufficient mean-stability condition, and showing in Monte Carlo simulations that DPLMS outperforms DSE-LMS, DRVSSLMS, and DLLAD across several combinations of impulse probability and intensity, with equal or lower computational cost.","pith_inferences":["The same probabilistic construction likely extends to other likelihood models: replacing the Gaussian likelihood with a heavier-tailed one (e.g., Student-t or Laplace) could give even stronger resistance to extreme impulses, since the Gaussian posterior is the current source of the quadratic error term.","A direct consequence not tested here is that DPLMS should track time-varying systems better than fixed-step-size algorithms when the unknown vector drifts, because the step size already incorporates the process-noise variance $\\sigma_\\rho^2$ as an explicit parameter.","The load-bearing independence assumption in the mean analysis suggests a testable refinement: measure the correlation between $\\mathbf{S}(i)\\mathbf{R}(i)$ and $\\hat{\\mathbf{W}}(i-1)$ in simulation; if non-negligible, the stability bound may need to be corrected for correlated inputs."],"forward_implications":["If DPLMS is correct, distributed estimators can be made resistant to impulsive noise without sign operations or nonlinear clipping; the adaptation is achieved entirely by the Bayesian step-size schedule.","The mean-stability condition in Eq. (22) gives users an explicit step-size bound $\\mu < 2/\\rho_{\\max}(\\sum_l \\alpha_l(i) c_{l,l} \\mathbf{R}_{xx,l}(i))$, so deployment can start from a principled range rather than trial-and-error.","Because its per-iteration complexity matches DLLAD and is lower than DSE-LMS and DRVSSLMS, the claimed robustness comes without added computational burden.","The construction suggests that posterior-approximation techniques from Bayesian filtering can be transplanted into diffusion networks one agent at a time."],"supporting_citations":[{"why":"Supplies the adapt-then-combine diffusion LMS formulation that DPLMS adopts for cooperation among nodes.","marker":"[2]"},{"why":"Introduces the probabilistic least-mean-squares filter from which the DPLMS step size is derived.","marker":"[20]"},{"why":"Provides the PLMS mean-square-deviation analysis and the variable step-size form that DPLMS modifies with the constant $\\tau$.","marker":"[21]"},{"why":"Baseline DSE-LMS algorithm used in the comparison; its sign-error recursion is what DPLMS is claimed to outperform.","marker":"[19]"},{"why":"Baseline DRVSSLMS algorithm; the comparison measures robustness and complexity against this robust variable step-size method.","marker":"[16]"},{"why":"Baseline DLLAD algorithm; DPLMS matches its complexity while improving robustness.","marker":"[17]"},{"why":"Supplies the Bernoulli-Gaussian impulse noise model and the input assumptions used in the mean stability analysis.","marker":"[15]"}],"fun_headline_variants":["Diffusion probabilistic LMS beats rivals under impulsive noise","Bayesian diffusion LMS stays robust when inputs are impulsive","Probabilistic step size makes diffusion LMS impulsive-noise proof","DPLMS: Gaussian posterior step size outperforms three adaptive filters","Diffusion with Bayesian step size tames impulsive interference"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The mean-stability proof assumes the matrix product $\\mathbf{S}(i)\\mathbf{R}(i)$ is independent of the current weight-error vector $\\hat{\\mathbf{W}}(i-1)$, even though the step-size factors in $\\mathbf{S}(i)$ depend on input power and past estimation errors.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion probabilistic LMS beats rivals under impulsive noise","Bayesian diffusion LMS stays robust when inputs are impulsive","Probabilistic step size makes diffusion LMS impulsive-noise proof","DPLMS: Gaussian posterior step size outperforms three adaptive filters","Diffusion with Bayesian step size tames impulsive interference"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000125,"raw_usage":{"total_tokens":1029,"prompt_tokens":791,"completion_tokens":238,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":407,"completion_tokens_details":{"reasoning_tokens":159}},"tokens_in":407,"tokens_out":238,"duration_ms":2808,"temperature":1.0,"reasoning_tokens":159,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:47:24.428731+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a two-node network with strongly correlated inputs or a node whose step size is driven by large past errors, compute the Monte Carlo average of $E[\\mathbf{S}(i)\\mathbf{R}(i)\\hat{\\mathbf{W}}(i-1)]$ and compare it with $E[\\mathbf{S}(i)\\mathbf{R}(i)]E[\\hat{\\mathbf{W}}(i-1)]$; if the difference is large and the mean weight error grows even when $\\mu$ satisfies Eq. (22), the independence assumption is what failed.","supporting_citations":[{"cited_title":"Cattivelli, A.H","cited_arxiv_id":null,"evidence_quote":"Supplies the adapt-then-combine diffusion LMS formulation that DPLMS adopts for cooperation among nodes."},{"cited_title":"Fernandez -Bes, V","cited_arxiv_id":null,"evidence_quote":"Introduces the probabilistic least-mean-squares filter from which the DPLMS step size is derived."},{"cited_title":"Mean-square-deviation analysis of probabilistic LMS algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Provides the PLMS mean-square-deviation analysis and the variable step-size form that DPLMS modifies with the constant $\\tau$."},{"cited_title":"Steady -state and stability analyses of diffusion sign-error LMS algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Baseline DSE-LMS algorithm used in the comparison; its sign-error recursion is what DPLMS is claimed to outperform."},{"cited_title":"Diffusion robust variable step -size LMS algorithm over distributed networks [J]","cited_arxiv_id":null,"evidence_quote":"Baseline DRVSSLMS algorithm; the comparison measures robustness and complexity against this robust variable step-size method."},{"cited_title":"Diffusion least logarithmic absolute difference algorithm for distributed estimation [J]","cited_arxiv_id":null,"evidence_quote":"Baseline DLLAD algorithm; DPLMS matches its complexity while improving robustness."},{"cited_title":"Diffusion normalized Huber adaptive filtering algorithm [J]","cited_arxiv_id":null,"evidence_quote":"Supplies the Bernoulli-Gaussian impulse noise model and the input assumptions used in the mean stability analysis."}],"review_version":1}