{"id":"b641fb77-7be2-48b9-9260-40f4e2c53c7e","arxiv_id":"2502.05583","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper casts sensor placement for graph-filtered regularized estimation as four cost functions, provides greedy and projected-gradient solvers, and reports up to 50% MSE reduction over A-, E-, and LR-design in tested scenarios.","lead":"This paper proposes a regularized graph-signal estimator and four sensor-placement criteria that choose which nodes to measure when only some can be observed. In simulations on power-grid and road-network data, the proposed placements reduce recovery error by up to 50% relative to standard A-, E-, and LR-design sampling.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The general-R claims are unsupported: Eq. (21) sets the covariance of Dy to R rather than D R D, so bCRB, WC-MSE, BMSE, and WC-BMSE are misspecified for correlated noise; since all experiments use diagonal R, the empirical MSE reductions stand but must be scoped to white/diagonal noise.","rationale":"Good-faith reading: the paper's contribution is a sampling-allocation framework for regularized graph-filtered estimation, with two practical algorithms and a coherent set of experiments. The condition that must hold for the central claim as stated is that the optimized cost functions are the bCRB/WC-MSE/BMSE/WC-BMSE of the model. That condition is not met for general R because Eq. (21) misstates the distribution of the sampled observation vector; the true covariance is D R D. This is an internal consistency issue rather than a disagreement with an external consensus. The error does not touch the reported simulations, all of which use scalar/white noise, so the empirical comparison is still meaningful. I would therefore not reject the paper; I would keep the reader's CONDITIONAL verdict, requiring either a corrected derivation for non-diagonal R or an explicit statement restricting the framework to diagonal/white noise. The reader's convexity concern is real but secondary for the headline numerical claim, since the reported MSEs are evaluated at the solution Algorithm 2 actually returns; hence my partial rather than full agreement with the reader's weakest_assumption.","tokens_in":31779,"tokens_out":11994,"duration_ms":115605,"concrete_test":"Use a three-node graph with hM(L)=I, R = [[1,0.5,0.4],[0.5,1,0.5],[0.4,0.5,1]], and S={1,2}. Compare the true Fisher information of Dy, hM^T D(D R D)^+ D hM, with the expression hM^T D R^{-1} D hM used in Eq. (25), and compute the trace term of Eq. (20) against the exact covariance of the estimator in Eq. (14) under this correlated e. If the two evaluations differ, the bCRB and the four derived cost functions are misspecified for non-diagonal R, and the paper's general-R scope must be narrowed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Most load-bearing is the correlated-noise misspecification. In the model e~N(0,R); after applying the sampling mask D, Dy has covariance D R D, not R. Eq. (21) states Dy~N(DhM(L)x,R), which is correct only when D and R commute, e.g. diagonal R. This propagates: the noise term in Eq. (20) should be K^{-1}hM D R^{-1} D R D R^{-1} D hM K^{-1}, and the FIM used in Eq. (25) should involve hM D(D R D)^+ D hM rather than hM D R^{-1} D hM. For non-diagonal R the printed bCRB and the cost functions built on it do not match the stated biased-CRB/WC-MSE/BMSE/WC-BMSE of the actual partial-observation model. All Section VI experiments use R = sigma^2 I or 0.01I, so the up-to-50% MSE reductions are not invalidated; but the abstract and Section III/IV claims for arbitrary noise covariance are unsupported without correction or scoping.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper considers the recovery of graph signals from partial noisy observations under a linear graph-filter measurement model y = hM(L)x + e, with a regularization term ||x - x0||^2_{h_R^+(L)}. It introduces the GFR-ML estimator in Eqs. (14)-(15) and proposes four sampling-allocation cost functions: the biased Cramér-Rao bound (bCRB), the worst-case MSE (WC-MSE), the Bayesian MSE (BMSE), and the worst-case Bayesian MSE (WC-BMSE). Two solvers are presented: a greedy algorithm with efficient rank-one updates, and an alternating projected gradient descent (PGD) algorithm based on a convex relaxation. The paper reports Monte Carlo simulations on synthetic graphs, the IEEE 118-bus power system, and the Minnesota road network, showing up to 50% MSE reduction relative to A-design, E-design, and LR-design sampling.","tokens_in":31976,"tokens_out":10739,"duration_ms":97436,"significance":"The paper offers a potentially unifying task-based framework for sampling in graph-signal recovery, with bandlimited and smoothness-based designs arising as special cases. The authors supply explicit gradient expressions and complexity analyses for both solvers, and the empirical evaluation spans synthetic data and two realistic network datasets. If the theoretical claims are restricted to diagonal/white noise (as used in all experiments) or corrected for general R, the framework is a useful contribution. The rank-one update formulas in Eqs. (41)-(44) and the gradient expressions in Claim 3 are practical assets.","major_comments":[{"comment":"The paper states in Eq. (21) that Dy ~ N(DhM(L)x, R) for the partial-observation model in Eq. (9). This is incorrect for a general noise covariance R: the noise term De has covariance D R D, so the correct distribution is Dy ~ N(DhM(L)x, D R D). Consequently, the FIM in Eq. (22), the bCRB in Eq. (25), and the MSE expression in Eq. (20) are valid only when D and R commute (in particular, for diagonal R), because the simplification D R^{-1} D R D R^{-1} D = D R^{-1} D is used implicitly. All experiments in Section VI use R = sigma^2 I or R = 0.01I, so the numerical MSE comparisons are not invalidated; however, the general claims about arbitrary noise covariance in the abstract and in Sections III and IV need to be corrected or explicitly restricted to diagonal or block-diagonal R.","section":"Section IV-A, Eqs. (20)-(25)"},{"comment":"The convexity proof in Appendix H establishes convexity of tr(K_affine(w)^{-1}) and lambda_max(K_affine(w)^{-1}) in the reparameterized variables w_i = d_i^2, with K_affine(w) defined in Eq. (124). However, Algorithm 2 performs projected gradient descent directly on d, using the gradient with respect to d from Claim 3 and the projections in Eqs. (52)-(53) on d. Convexity in w does not imply convexity in d, so the stated global-convergence guarantee does not apply to the implemented algorithm. The sentence in Section V-B2 that PGD 'is expected to converge to the global optimum' should either be softened or the algorithm should be reparameterized and implemented in the w-variables.","section":"Section V-B2 and Appendix H"}],"minor_comments":[{"comment":"The derivation of Eq. (20) mentions that the covariance of De is D R D, but the displayed trace term contains D R^{-1} D; please add the intermediate expression D R^{-1} D R D R^{-1} D and state explicitly where diagonal R is assumed.","section":"Eq. (20)"},{"comment":"The supplement refers to 'Theorem 1 in [67]', where [67] is the arXiv version of this same paper; this should be replaced by an internal cross-reference to the paper's own Theorem 1.","section":"Supplement, Appendix B"},{"comment":"The input list says 'number of nodes, q', but q denotes the number of selected sensors; please correct the label.","section":"Algorithm 1"},{"comment":"The caption item '(3) hM(L) = hDiff' should read '(c) hM(L) = hDiff' to match the panels.","section":"Fig. 1 caption"},{"comment":"The backtracking line search in Eq. (48) uses the projection P defined in Eq. (49), which is a projection onto the non-convex binary set, but the actual update in Eq. (50) and the projection steps in Eqs. (52)-(53) use the relaxed convex set. This inconsistency should be fixed so that the line search condition matches the projection actually applied in the update.","section":"Section V-B, Eqs. (48)-(53)"}],"recommendation":"major_revision","confidential_remarks":"The main technical issue is the mismatch between the general-R formulation and the diagonal-R validation. If the authors are unwilling to re-derive the bCRB for general correlated noise using (D R D)^+, they should explicitly scope the title, abstract, and theoretical sections to diagonal or block-diagonal noise covariance. The convexity issue is secondary but should be addressed by either reparameterizing the PGD algorithm or softening the convergence claim. The supplementary material's self-citation of the paper itself as [67] should also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate contribution. The GFR-ML estimator plus the four cost functions (bCRB, WC-MSE, BMSE, WC-BMSE) give practitioners a unified way to do task-based sensor placement for graph signals under general filters and regularization. The gradient formulas in Claim 3 are genuinely useful, and the greedy/PGD solvers are sensible. The simulations on the IEEE 118-bus and Minnesota road data are clean and show up to 50% MSE reduction over A-, E-, and LR-designs in the tested regimes. I believe those empirical claims.\n\nThe main problem is the correlated-noise claim. Eq. (21) states Dy ~ N(DhM(L)x, R), but the covariance of Dy is D R D, not R. That error propagates into the bCRB derivation in Eqs. (22)-(25) and into the WC-MSE/BMSE/WC-BMSE cost functions built on it. For diagonal R the formulas work, because D R^{-1} D is the pseudo-inverse of D R D, but for general R they are misspecified. All experiments use R = sigma^2 I, so the reported gains are not invalidated; the authors just need to scope the claims to diagonal noise or redo the FIM with (D R D)^+.\n\nSecond, the convexity section. Appendix H proves convexity in w_i = d_i^2, while Algorithm 2 optimizes d directly with gradients in Claim 3 taken w.r.t. d. The global-convergence sentence in Section V-B2 is therefore not backed by the proof. This is an overstatement, not a fatal error; the PGD still works empirically.\n\nMinor: no code or error bars, which weakens reproducibility but doesn't change the qualitative conclusions. The citation pattern looks fine; the self-citations are to prior work this framework extends.\n\nBottom line: the paper deserves a serious referee. It should be accepted after revision, with the R = diagonal assumption stated upfront and the convexity claim either matched to the algorithm or softened. I'd point the authors to the pseudo-inverse fix and ask them to either extend experiments to one correlated-noise case or explicitly scope the theory.","headline":"A generally solid extension of optimal-design sampling to graph-filtered models, with a real scoping flaw in the correlated-noise claims and a convexity guarantee that does not match the implemented algorithm.","tokens_in":32597,"tokens_out":2557,"would_cite":true,"duration_ms":25631,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Choosing sensors by recovery-error cost functions rather than bandlimitedness or smoothness assumptions cuts graph-signal recovery MSE by up to 50% in underdetermined systems.","keywords":["graph signal processing","sampling allocation","sensor placement","graph filters","regularized maximum likelihood","biased Cramer-Rao bound","Bayesian mean squared error","underdetermined systems"],"falsifier":"Run the paper's synthetic-data Monte Carlo comparison with a correlated noise covariance such as $R_{ij}=0.5^{|i-j|}$ at $q/N=70\\%$; if sensors chosen by the bCRB or BMSE criteria no longer achieve lower empirical MSE than A- and E-design, the misspecified Fisher-information model is what carries the reported gain.","tokens_in":31480,"feed_emoji":"📡","tokens_out":13719,"duration_ms":121191,"temperature":0.7,"pith_summary":"The paper is trying to establish that, in underdetermined graph-signal recovery, the sampling set should be chosen to minimize the recovery error of the estimator itself, and that this can be done with four signal-independent cost functions. It introduces the GFR-ML estimator, which combines a known measurement graph filter with a positive-semidefinite regularization graph filter, and shows that its MSE splits into a noise term and a signal-dependent bias term. Because that bias term blocks direct MSE optimization, the paper proposes four surrogate costs — the bCRB, WC-MSE, BMSE, and WC-BMSE — all functions of the regularized matrix $K(d)$, and gives a greedy solver and an alternating projected-gradient solver. If this is right, sensor placement no longer has to assume the signal is bandlimited or smooth; the paper's simulations show up to 50% MSE reduction against A-, E-, and LR-design on synthetic graphs, the IEEE 118-bus power system, and the Minnesota road network.","feed_headline":"Recovery-driven sensor placement cuts graph-signal MSE by up to 50%","feed_subtitle":"Four error-aware cost functions pick sensor locations without assuming the signal is bandlimited or smooth.","key_machinery":"The object doing the work is the regularized sampled-information matrix $K(d)=\\mathbf{h}_M(L)DR^{-1}D\\mathbf{h}_M(L)+\\mu\\mathbf{h}_R^+(L)$, where $\\mathbf{h}_M(L)$ is the measurement graph filter and $\\mathbf{h}_R^+(L)$ is a positive-semidefinite graph filter encoding prior knowledge. The GFR-ML estimator inverts this matrix, and each proposed cost function is a tractable function of it: $\\mathrm{bCRB}=\\mathrm{tr}(K^{-1}\\mathbf{h}_M(L)DR^{-1}D\\mathbf{h}_M(L)K^{-1})$, $\\mathrm{BMSE}=\\mathrm{tr}(K^{-1})$, WC-MSE adds $\\mu^2$ times the squared largest singular value of $K^{-1}\\mathbf{h}_R^+(L)$, and WC-BMSE is $\\lambda_{\\min}^{-1}(K)$. The shared gradient formula $\\nabla C(d)=-2\\mathrm{diag}(R^{-1}D\\mathbf{h}_M(L)K^{-1}QK^{-1}\\mathbf{h}_M(L))$, with a different matrix $Q$ for each criterion, is what allows the alternating projected-gradient algorithm to optimize all four cost functions with the same code path, and the rank-one structure of adding one sensor is what makes the greedy updates fast.","core_discovery":"The paper's central claim is that for the underdetermined measurement model $y=\\mathbf{h}_M(L)x+e$, the right sensor set is the one that minimizes the recovery error of the graph-filtered regularized maximum likelihood (GFR-ML) estimator, not a proxy tied to bandlimitedness or smoothness. Concretely, with $D=\\mathrm{diag}(d)$ marking the sampled nodes, the estimator is $\\hat{x}=K^{-1}(d)(\\mathbf{h}_M(L)DR^{-1}Dy+\\mu\\mathbf{h}_R^+(L)x_0)$ with $K(d)=\\mathbf{h}_M(L)DR^{-1}D\\mathbf{h}_M(L)+\\mu\\mathbf{h}_R^+(L)$, and its MSE decomposes into a noise term plus a bias term that depends on the unknown signal $x$. Since direct MSE minimization is therefore impossible, the paper proposes four signal-independent cost functions — bCRB (a lower bound on the MSE of estimators with a prescribed bias), worst-case MSE, Bayesian MSE, and worst-case Bayesian MSE — and shows in Monte Carlo experiments that optimizing them reduces recovery MSE by up to 50% relative to A-design, E-design, and LR-design sampling.","pith_inferences":["Because the proposed costs collapse to A- and E-design in the exact-bandlimited limit, the relative advantage should be largest under model mismatch; a direct test is whether the gap shrinks as signals approach perfect bandlimitedness.","The bCRB and the derived costs treat the Fisher information as $\\mathbf{h}_M(L)DR^{-1}D\\mathbf{h}_M(L)$, which is exact for diagonal or block-diagonal noise; re-deriving them with the exact observed covariance would extend the same machinery to correlated noise, a regime the reported experiments (all with $R=\\sigma^2 I$) do not test.","With polynomial or Chebyshev-approximated graph filters, the matrix-vector products in $K(d)$ are local, so the same sampling criteria could plausibly run in a distributed way on much larger networks."],"forward_implications":["A fixed sensor budget buys lower recovery error: in the tested scenarios the proposed designs match or beat A-, E-, and LR-design, so the same accuracy can be obtained with fewer measurements.","Bandlimited and smooth signal models become special cases: under strict bandlimitedness with $\\mu\\to\\infty$, the bCRB and BMSE reduce to A-design and the WC-BMSE reduces to E-design.","For diagonal noise, adding a sensor is a rank-one update of $K(d)$, reducing the greedy algorithm's total cost to $O(qN^3)$ or $O(qN^2)$ instead of $O(qN^4)$.","The alternating projected-gradient solver scales the criteria to large graphs, demonstrated on a 2,642-node road network.","In the power-system experiments, bCRB-based sampling is the most robust to noise and topology mismatch, while BMSE and WC-BMSE give the best average performance in Bayesian settings."],"supporting_citations":[{"why":"Defines the A-design sampling baseline that the bCRB and BMSE reduce to in the strict bandlimited limit.","marker":"[11]"},{"why":"Defines the E-design sampling baseline that the WC-BMSE criterion reduces to under strict bandlimitedness.","marker":"[13]"},{"why":"Supplies the Laplacian-regularized sampling baseline (LR-design) that the WC-BMSE criterion generalizes.","marker":"[25]"},{"why":"Provides the biased Cramér–Rao bound used to define the bCRB cost function.","marker":"[49]"},{"why":"Introduces the convex-relaxation sensor-selection approach on which the alternating projected-gradient solver builds.","marker":"[32]"},{"why":"Supplies the IEEE 118-bus test-case network used in the power-system experiments.","marker":"[35]"},{"why":"Supplies the Minnesota road-network graph used in the large-scale experiments.","marker":"[36]"}],"fun_headline_variants":["Recovery-aware sensor placement reduces graph-signal MSE by up to 50%","Four cost functions for sensor selection trim graph-signal recovery MSE by 50%","Target recovery error, not smoothness: sensor placement cuts graph-signal MSE 50%","Sensor selection that minimizes recovery error halves graph-signal MSE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that sampling a set of nodes only reweights noise through a diagonal mask, which is exact only when the noise at different nodes is uncorrelated or correlated only within known groups; with general correlated noise, the bCRB and the four cost functions built on it are misspecified.","fun_headline_variants_meta":{"raw":{"variants":["Recovery-aware sensor placement reduces graph-signal MSE by up to 50%","Four cost functions for sensor selection trim graph-signal recovery MSE by 50%","Target recovery error, not smoothness: sensor placement cuts graph-signal MSE 50%","Sensor selection that minimizes recovery error halves graph-signal MSE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000768,"raw_usage":{"total_tokens":3470,"prompt_tokens":1080,"completion_tokens":2390,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":696,"completion_tokens_details":{"reasoning_tokens":2305}},"tokens_in":696,"tokens_out":2390,"duration_ms":15386,"temperature":1.0,"reasoning_tokens":2305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T18:46:49.432331+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the paper's synthetic-data Monte Carlo comparison with a correlated noise covariance such as $R_{ij}=0.5^{|i-j|}$ at $q/N=70\\%$; if sensors chosen by the bCRB or BMSE criteria no longer achieve lower empirical MSE than A- and E-design, the misspecified Fisher-information model is what carries the reported gain.","supporting_citations":[{"cited_title":"Efficient sampling set selection for bandlimited graph signals using graph spectral proxies,","cited_arxiv_id":null,"evidence_quote":"Defines the A-design sampling baseline that the bCRB and BMSE reduce to in the strict bandlimited limit."},{"cited_title":"Discrete signal processing on graphs: Sampling theory,","cited_arxiv_id":null,"evidence_quote":"Defines the E-design sampling baseline that the WC-BMSE criterion reduces to under strict bandlimitedness."},{"cited_title":"F ast graph sampling set selection using gershgorin disc alignment,","cited_arxiv_id":null,"evidence_quote":"Supplies the Laplacian-regularized sampling baseline (LR-design) that the WC-BMSE criterion generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the biased Cramér–Rao bound used to define the bCRB cost function."},{"cited_title":"Sensor selection via convex optim ization,","cited_arxiv_id":null,"evidence_quote":"Introduces the convex-relaxation sensor-selection approach on which the alternating projected-gradient solver builds."},{"cited_title":"Power systems test case archive","cited_arxiv_id":null,"evidence_quote":"Supplies the IEEE 118-bus test-case network used in the power-system experiments."},{"cited_title":"GSPBOX: A toolbox for signal processing on graphs,","cited_arxiv_id":null,"evidence_quote":"Supplies the Minnesota road-network graph used in the large-scale experiments."}],"review_version":1}