{"id":"c1a1a517-5b8a-4de9-87c2-ff9e21756bcd","arxiv_id":"2502.00811","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A low-complexity variational Bayesian algorithm jointly performs 3D scatterer localization, multiuser channel estimation, and data recovery in ISAC systems by exploiting shared sparsity and refining the position grid.","lead":"This paper presents an algorithm that lets a 6G base station jointly locate scatterers in the environment, estimate channels, and decode data from multiple users using the same wireless signals. It exploits the fact that different users share the same physical scatterers, and uses variational Bayesian inference to keep the computation cheap.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (6)'s virtual-scatterer equivalence is the load-bearing physical assumption: for multi-bounce paths the virtual position depends on user location, so the joint-sparsity prior that drives the reported gains has no guarantee of holding in real environments.","rationale":"The reader's verdict already identifies Eq. (6) as the weakest assumption, and my stress-test converges on the same point after considering alternatives. The more algorithmic concerns I examined—the fixed-Bα gradient updates in the proof of Theorem 3, the heuristic M-step, and the complexity accounting—are either fixable proof gaps or practical engineering issues that do not strike at the core of what the paper promises: joint physical scatterer localization and channel/data recovery. The physical equivalence in Eq. (6), by contrast, determines whether the estimated 'scatterers' correspond to real objects in a multi-bounce or extended-reflector environment. The paper's simulations, which draw on a Quadriga UMA scenario, are not described with enough detail to establish that they contain the multi-bounce/diffuse effects that would break the virtual-scatterer alignment. The algorithmic contributions—subspace-constrained VBI, turbo message passing, and coarse estimation—are coherent and the complexity reduction relative to the full inverse is real. The Theorem 3 proof is terse and would benefit from a more rigorous treatment of the inexact block updates, but I do not see an immediate counterexample that would make the theorem false as an algorithmic guarantee for the constrained AVBI problem. Therefore the most load-bearing concern is the physical modeling assumption, and it is exactly the one the reader highlighted. My recommendation agrees with the reader's CONDITIONAL verdict: the paper should be accepted only if the authors add a robustness study or clearly scope the claims to single-bounce point-scatterer environments.","tokens_in":25013,"tokens_out":17898,"duration_ms":200585,"concrete_test":"Generate two-bounce and extended-reflector channels (e.g., with a ray tracer) using the same BS/user geometry and true physical scatterer positions. Run the full EM-Turbo-BiSVBI pipeline on these channels and on the single-bounce point-scatterer channels used in Section VI. Compare scatterer localization RMSE against the true physical positions and compare the gain of EM-Turbo-BiSVBI over EM-Turbo-BiSVBI-IID. If the multi-bounce RMSE is large or the joint-sparsity gain disappears, Eq. (6) does not deliver physical scatterer localization, and the central sensing claim is not supported beyond the exact-model case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central argument's load-bearing premise is the single-bounce virtual-scatterer equivalence in Eq. (6), used to justify the location-domain sparse representation (10) and the BGG joint-sparsity prior. The mathematical equivalence is valid for a single path: for any angle (θ,φ) and delay τ there is an r solving the equation. But the virtual position p=(θ,φ,r) depends on the user position p_u,k through Eq. (6). For a genuine multi-bounce path, different users whose signals bounce off the same physical object will generally have different virtual positions; the same holds for diffuse or extended reflectors, which do not map to a small number of grid points. The joint support s in Section III-B and the coarse-estimation clustering in Section V-C both rely on the union over users of nonzero α entries aligning on a common grid. That alignment is a structural property of single-bounce point-scatterer geometries, not of the channels the paper claims to sense. Section VI generates the channel 'according to the UMA scenario in the Quadriga model,' but the paper does not report how the Quadriga channel relates to the ground-truth scatterer positions used for the localization RMSE, nor does it include a mismatch experiment with multi-bounce or diffuse scattering. Without such robustness evidence, the near-genie results in Figs. 7-8 cannot be read as evidence for real-environment ISAC sensing; they only verify the algorithm when the model is exactly true.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper studies uplink massive MIMO-OFDM ISAC with multiple single-antenna users, where the base station jointly localizes scatterers, estimates multiuser channels, and recovers data. The authors introduce a 3D location-domain sparse channel model over a dynamic position grid, a three-layer Bernoulli-Gamma-Gaussian prior for joint support, and formulate the problem as bilinear structured sparse recovery with uncertain parameters such as user positions and timing offsets. They propose EM-Turbo-BiSVBI: a turbo E-step combining a bilinear subspace-constrained variational Bayesian inference module (diagonal covariance, subspace-constrained inverse for the posterior mean, gradient refinement) with message passing on the support graph, and an EM M-step that refines the dynamic grid and imperfect factors by gradient ascent. They also propose a SMUSIC-SCVBI coarse estimation stage to build compact position grids. Theorem 3 claims monotone KLD decrease and convergence to stationary points of the constrained VBI problem. Simulations against VB-CESD, ST-MUSIC, a two-stage scheme, an iid-prior variant, and a full-inverse variant show NMSE and RMSE gains and near-genie performance.","tokens_in":25381,"tokens_out":6386,"duration_ms":71325,"significance":"If the central claims hold, the paper contributes a scalable algorithmic framework for an important ISAC problem: it reduces the dominant matrix inverse from O(K(Q+1))^3 to O(|S_hat|^3) with a support-subspace update, integrates data-aided sensing with dynamic-grid refinement, and demonstrates the value of location-domain joint sparsity by ablating it with an iid-prior version. The variational update equations follow from a clearly stated constrained KL objective, and the comparison against EM-Turbo-BiVBI is a useful check on the subspace approximation. However, the physical premise that all NLoS paths are equivalent to single-bounce scatterers is user-dependent for multi-bounce paths, and the joint-support prior that drives the reported gains depends on this premise. The convergence proof for Theorem 3 also has a gap. Both issues are discussed below; the algorithmic and simulation contributions remain a solid basis for revision.","major_comments":[{"comment":"The virtual-scatterer equivalence in Eq. (6) is valid per path, but the virtual position p=(theta,phi,r) is a function of the user position p_u,k through Eq. (6). For a genuine multi-bounce path via the same physical reflector, two users at different positions will in general be assigned different virtual positions. Consequently, the common-support prior p(s_k|s) and the joint support s defined in Sec. III-B need not hold for the channels the paper claims to sense. Section VI generates channels with the Quadriga UMA model, but the paper does not state whether the generated paths are single-bounce point-reflector paths, nor how the ground-truth scatterer positions used in the localization RMSE are extracted from the Quadriga realization. As written, Figs. 7 and 8 verify the algorithm only when the model is exactly true. I request either (a) restrict the system model and claims to single-bounce point scatterers and justify the Quadriga channel as an approximation of that model, or (b) add a mismatch experiment with multi-bounce or diffuse scattering and report the resulting localization and channel NMSE. This is load-bearing because the joint-sparsity prior is the source of the largest performance gain over the iid version in the simulations.","section":"Sec. II-B, Eq. (6); Sec. III-B"},{"comment":"The proof of Theorem 3 has two gaps. First, it asserts 'strong convexity of KL(...) w.r.t. xi_j' for all xi_j not equal to mu_alpha, but no strong-convexity argument is supplied for the Gamma parameters (a_kq, b_kq), the support probabilities lambda_kq, or the noise parameters c and d; the KL is not generally strongly convex in these variables. Second, the theorem's statement covers the BiSVBI algorithm, but the implemented algorithm updates the support set S_hat by thresholding mu_alpha in Eq. (32), and S_hat enters the posterior-mean update through the subspace-constrained inverse in Eq. (29) and the robust initialization in Eq. (61). The parameterization xi in Eq. (58) and the alternating update in Eq. (59) do not include S_hat, so the proof does not cover the algorithm as implemented. Please either provide a corrected proof that handles the support update and does not rely on unproved strong convexity, or restate the theorem to match what is actually proven.","section":"Appendix B, Theorem 3; Eq. (32)"},{"comment":"The abstract and Sec. IV claim that the subspace-constrained inverse reduces complexity 'with almost no sacrifice' on performance and convergence speed. The evidence for this is the convergence comparison in Figs. 5 and 6 for one scenario, with no wall-clock time, per-iteration runtime, or complexity-matched benchmark. The complexity orders in Eqs. (42) and (53) are asymptotic and do not by themselves justify 'almost no sacrifice', especially because the thresholding in Eq. (32) makes the achieved support S_hat a random quantity. Please report measured runtime or a complexity-matched comparison, or soften the claim to 'negligible performance loss in the tested scenarios'.","section":"Sec. IV-E, Sec. V-D, Figs. 5-7"}],"minor_comments":[{"comment":"The lemma statement says 'for given q(v_k), v_k not equal to alpha', but the intended meaning, based on the proof and the sentence before Eq. (27), is that all densities q(v_l) with l not equal to k are given. Please correct the wording.","section":"Lemma 2, Sec. IV-C2"},{"comment":"The caption of Fig. 8(a) reads 'Scatterer localization RMSE' while the vertical axis in the figure is labeled 'Channel estimation NMSE'. This mismatch should be fixed, since Fig. 8(c) already reports RMSE.","section":"Fig. 8"},{"comment":"The text says 'we very N p from 2 to 32'; this should read 'we vary N p'. In addition, the footnote in Sec. VI refers to 'VR-CESD' and 'TS-MUSIC', while the baselines are elsewhere called VB-CESD and ST-MUSIC.","section":"Sec. VI-C"},{"comment":"The definition of C_kq contains a typographical artifact: the term labeled 'bn' should be the appropriate Gamma rate parameter, and the overbar notation for the inactive-state parameters should be applied consistently. Please re-check the display.","section":"Eq. (36)"},{"comment":"Algorithm 1 line 5 says the gradient update is performed 'for Bx times', but the surrounding text and Theorem 3 use B_alpha. Also, line 11 updates the support S_hat before the outer support-to-modules message is sent; it would be clearer to specify whether S_hat is updated once per outer EM iteration or once per turbo iteration.","section":"Algorithm 1"},{"comment":"The simulation section does not report several parameters that the algorithm depends on: the inner iteration count I, the outer iteration count T, the gradient iteration counts B_alpha and B_x, and the support threshold epsilon (Eq. (32)). These choices are needed for reproducibility and should be stated.","section":"Sec. VI"},{"comment":"There are minor typos: 'estimating the the location-domain' in the abstract, and 'blinear' in the index terms.","section":"Abstract and Index Terms"}],"recommendation":"major_revision","confidential_remarks":"This is a strong systems paper with a clear algorithmic contribution. I do not see a circularity problem: the authors' earlier SCVBI and Turbo-VBI results are used as components, not as evidence for the main performance claims. The main risks are the physical validity of the shared-scatterer model under multi-bounce propagation and the rigor of Theorem 3's proof. Both are addressable in a major revision. The paper is well within the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a serious, well-executed extension of the authors' prior VBI line, and the main caveat is not the math but the physical model. The 3D location-domain sparse representation and the bilinear subspace VBI updates are genuinely new, and the complexity reduction from O(K(Q+1))^3 to O(|S_hat|^3 + ...) is real within the model. I'd send it to review.\n\nWhat the paper does well: the derivation of Lemma 2 is careful, the update equations follow, and the complexity analysis is concrete. The EM-M step is honestly labeled heuristic. The paper also credits prior work properly, including its own; the circularity concern is low because the bilinear derivation is new.\n\nSoft spots, in order of importance. First, the virtual-scatterer equivalence in Eq. (6) is load-bearing. For a multi-bounce path, the virtual position depends on the user location, so the joint-sparsity prior (common grid support across users) can be misspecified in real environments. The paper generates channels from Quadriga UMA but never explains how the Quadriga multipath is mapped to ground-truth scatterer positions for the localization RMSE, and there is no mismatch experiment with multi-bounce or diffuse scattering. Without that, the near-genie results are verification of the algorithm when the model is exactly true, not evidence for real-environment ISAC. This is addressable: add a mismatch study or reframe the claim as virtual-scatterer sensing.\n\nSecond, Theorem 3's proof asserts strong convexity of the KL in each block and does not explicitly handle the support set update in (32). The support update only affects the initial point, so I think the proof can be repaired, but as written it is a gap. Minor.\n\nThird, no code or data and no error bars on the figures. For a 50-page algorithm paper that's common, but it makes the performance claims hard to verify.\n\nWho is this for: people working on ISAC, structured sparse recovery, or VBI for bilinear models. It is not a broad scientific breakthrough but it is a solid step within that niche. The paper deserves a serious referee; I would send it out with a request to fix the model-mismatch issue and the proof gap.","headline":"Solid algorithmic extension with a real but addressable model-mismatch gap; the convergence proof has a small unproven step.","tokens_in":25888,"tokens_out":3324,"would_cite":true,"duration_ms":38320,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that one Bayesian loop can localize 3D scatterers, estimate multiuser channels, and recover uplink data in an ISAC system, while its subspace-constrained inverse keeps complexity practical and performance near genie-aided.","keywords":["integrated sensing and communication","ISAC","3D scatterer localization","bilinear sparse recovery","variational Bayesian inference","massive MIMO-OFDM","joint sparsity","dynamic position grid"],"falsifier":"Run the same EM-Turbo-BiSVBI pipeline in a scene whose scattering includes an extended reflector such as a wall or large flat surface instead of only isolated point scatterers, and compare scatterer localization RMSE and channel NMSE against the point-scatterer case at identical SNR and pilot count. If the estimates degrade sharply or the recovered virtual scatterers drift off the true surface, the single-bounce equivalence in Eq. (6) is falsified as a general model of the propagation environment.","tokens_in":24784,"feed_emoji":"📡","tokens_out":7137,"duration_ms":61577,"temperature":0.7,"pith_summary":"This paper claims that in an uplink integrated sensing and communication (ISAC) system, the base station can sense the three-dimensional scattering environment, estimate the multiuser channels, and recover the users' data in a single inference procedure rather than treating sensing and communication as separate tasks. The key modeling move is a 3D location-domain sparse channel representation in which every non-line-of-sight path is assigned to a grid point corresponding to a virtual scatterer, so that different users' channels share a common sparse support. On top of that model, the paper builds an EM-Turbo-BiSVBI algorithm that alternates a variational Bayesian estimation step for channels and data with a gradient-based learning step for the grid, time offsets, and user positions. The payoff, if the claims hold, is that sensing accuracy is largely insensitive to pilot overhead, with localization staying accurate even at two pilot subcarriers, and the whole procedure approaches the performance of an ideal genie-aided receiver in simulation.","feed_headline":"One Bayesian loop maps scatterers and decodes uplink data","feed_subtitle":"Data signals double as sensing probes, so scatterer localization stays sharp with as few as two pilots.","key_machinery":"The load-bearing object is the virtual-scatterer equivalence in Eq. (6), which turns any NLoS path, regardless of how many bounces it physically took, into a single-bounce reflection from a virtual point at position $(\\theta,\\phi,r)$, so the whole environment can be described by a sparse grid of locations. Combined with the 3D location-domain sparse representation of Eq. (10), this yields a linear-in-channel, bilinear-in-data observation model. The inference machinery is EM-Turbo-BiSVBI: Module A applies BiSVBI with an independence-constrained variational posterior, a subspace-constrained inverse on the estimated support $\\hat{\\mathcal S}$, and a gradient update for the posterior mean; Module B passes extrinsic messages for the joint support variable; the M-step performs gradient ascent on an EM surrogate to update the dynamic grid, time offsets, and user positions. A SMUSIC-SCVBI coarse estimator first identifies candidate angles and builds a compact position grid, which is what makes the $O(|\\hat{\\mathcal S}|^3)$ complexity claim meaningful. Theorem 3 is the convergence certificate that connects these approximations to stationary solutions of the original variational problem.","core_discovery":"On its own terms, the paper's contribution is a tractable solution to the joint sensing and data recovery problem in a massive MIMO-OFDM uplink, formulated as bilinear structured sparse recovery with a dynamic position grid and imperfect parameters. The paper introduces the 3D location-domain sparse channel model of Eq. (10), driven by the virtual-scatterer equivalence of Eq. (6), and a three-layer Bernoulli-Gamma-Gaussian prior that encodes joint sparsity across users. The algorithmic core is BiSVBI: a subspace-constrained matrix inverse replaces the high-dimensional inversion $W_\\alpha^{-1}$ by working only on the estimated support $\\hat{\\mathcal S}$, and a gradient update with Armijo step size refines the posterior mean. Theorem 3 states that BiSVBI monotonically decreases the AVBI KLD objective and that every limiting point is a stationary solution of Problem AVBI. The EM outer loop then refines the position grid and learns time offsets and user positions, and a SMUSIC-SCVBI coarse stage shrinks the search region. Simulations indicate that this combination achieves near-genie channel, data, and localization performance while cutting the dominant per-iteration complexity from $O((K(Q+1))^3)$ to $O(|\\hat{\\mathcal S}|^3+\\dots)$.","pith_inferences":["Editor's inference: the scatterer map produced by this algorithm is an environmental description rather than a per-user channel, so it could plausibly be reused for downlink beamforming or predictive handover; the paper does not test that reuse.","Editor's inference: the reported performance is demonstrated for isolated point scatterers in a 100 m by 100 m by 10 m area; extended reflectors or diffuse scattering would likely violate the single-bounce equivalence in Eq. (6), so a robustness study against such environments is a natural next experiment.","Editor's inference: because the algorithm outputs soft posterior distributions for data symbols, it could be nested with a channel decoder in a turbo equalization loop; the paper stops at a Gaussian data prior and does not explore coded iterative decoding."],"forward_implications":["High-accuracy scatterer localization persists even with $N_p=2$ pilot subcarriers, because data subcarriers act as extra sensing observations; the paper's simulations show localization RMSE stays small when pilots are scarce.","The three-layer BGG prior's joint sparsity gives a substantial gain over an i.i.d. prior and over angular-domain baselines, so sharing scatterers across users is what drives the mutual enhancement of sensing and communication.","The subspace-constrained inverse cuts the per-iteration complexity of the posterior mean update from $O((K(Q+1))^3)$ to $O(|\\hat{\\mathcal S}|^3+\\dots)$, and simulations show convergence speed and steady-state performance nearly identical to the full-inverse version.","Learning time offsets and user positions inside the EM loop brings performance close to the genie-aided case in data recovery NMSE and localization RMSE.","Theorem 3 gives a monotone-decreasing KLD guarantee with stationary limiting points, so unlike heuristic bilinear message-passing baselines, the algorithm carries a convergence certificate under the model assumptions."],"supporting_citations":[{"why":"Supplies the subspace-constrained variational Bayesian inference and dynamic-grid machinery that BiSVBI generalizes from linear to bilinear observation models.","marker":"[16]"},{"why":"Provides the SMUSIC pseudo-spectrum used in the coarse estimation stage to narrow down the angle search region.","marker":"[15]"},{"why":"Is the variational Bayes joint channel estimation and symbol decoding baseline that the proposed scheme is compared against.","marker":"[13]"},{"why":"Companion baseline work that supplies sparse channel estimation and Cramer-Rao bound context for the massive MIMO-OFDM comparison.","marker":"[14]"},{"why":"Introduces the bilinear observation model and BiGAMP algorithm that motivate the bilinear formulation and serve as a comparison point for robustness and guarantees.","marker":"[10]"},{"why":"Supplies the three-layer Bernoulli-Gamma-Gaussian structured-sparsity prior and turbo approach used for the location-domain channel model.","marker":"[17]"},{"why":"Provides the Turbo-VBI framework and the EM-surrogate gradient justification used in the M-step for grid and parameter learning.","marker":"[19]"},{"why":"Introduces the two-dimensional location-domain sparse channel model that this paper extends to three dimensions.","marker":"[8]"},{"why":"Generates the 3GPP UMa channel scenarios used in the simulations that produce the reported near-genie results.","marker":"[27]"}],"fun_headline_variants":["One Bayesian loop maps scatterers and decodes data in ISAC","Subspace VBI cracks joint sensing and data recovery in ISAC","Bayesian ISAC: sense and decode with few pilots","One algorithm maps scatterers and recovers data in ISAC","Near-genie ISAC sensing and data recovery with low complexity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model assumes every non-line-of-sight propagation path can be represented as reflection from a single virtual point scatterer lying exactly on the dynamic position grid; if real environments contain diffuse scattering or extended reflectors such as walls, this equivalence in Eq. (6) can be misspecified and the paper provides no robustness analysis for that case.","fun_headline_variants_meta":{"raw":{"variants":["One Bayesian loop maps scatterers and decodes data in ISAC","Subspace VBI cracks joint sensing and data recovery in ISAC","Bayesian ISAC: sense and decode with few pilots","One algorithm maps scatterers and recovers data in ISAC","Near-genie ISAC sensing and data recovery with low complexity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3275,"prompt_tokens":1060,"completion_tokens":2215,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":2128}},"tokens_in":676,"tokens_out":2215,"duration_ms":31185,"temperature":1.0,"reasoning_tokens":2128,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T17:38:02.323747+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same EM-Turbo-BiSVBI pipeline in a scene whose scattering includes an extended reflector such as a wall or large flat surface instead of only isolated point scatterers, and compare scatterer localization RMSE and channel NMSE against the point-scatterer case at identical SNR and pilot count. If the estimates degrade sharply or the recovered virtual scatterers drift off the true surface, the single-bounce equivalence in Eq. (6) is falsified as a general model of the propagation environment.","supporting_citations":[{"cited_title":"Subspace constrained variational Bayesian inference for structured compressive sensing with a dynamic grid,","cited_arxiv_id":null,"evidence_quote":"Supplies the subspace-constrained variational Bayesian inference and dynamic-grid machinery that BiSVBI generalizes from linear to bilinear observation models."},{"cited_title":"TST-MUSIC for joint DOA- delay estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the SMUSIC pseudo-spectrum used in the coarse estimation stage to narrow down the angle search region."},{"cited_title":"Massive MIMO-OFDM systems with low resolution ADCs: Cramér-rao bound, sparse channel estimation, and soft symbol decod- ing,","cited_arxiv_id":null,"evidence_quote":"Companion baseline work that supplies sparse channel estimation and Cramer-Rao bound context for the massive MIMO-OFDM comparison."},{"cited_title":"Bilinear generalized approxi- mate message passing-part I: Derivation,","cited_arxiv_id":null,"evidence_quote":"Introduces the bilinear observation model and BiGAMP algorithm that motivate the bilinear formulation and serve as a comparison point for robustness and guarantees."},{"cited_title":"Cloud-assisted cooperative localization for vehicle platoons: A turbo approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the three-layer Bernoulli-Gamma-Gaussian structured-sparsity prior and turbo approach used for the location-domain channel model."},{"cited_title":"Robust recovery of structured sparse signals with uncertain sensing matrix: A Turbo-VBI approach,","cited_arxiv_id":null,"evidence_quote":"Provides the Turbo-VBI framework and the EM-surrogate gradient justification used in the M-step for grid and parameter learning."},{"cited_title":"Joint scattering environment sensing and channel estimation based on non-stationary Markov random field,","cited_arxiv_id":null,"evidence_quote":"Introduces the two-dimensional location-domain sparse channel model that this paper extends to three dimensions."},{"cited_title":"Wiley 5G Ref, 2021","cited_arxiv_id":null,"evidence_quote":"Generates the 3GPP UMa channel scenarios used in the simulations that produce the reported near-genie results."}],"review_version":1}