{"id":"d1a7a662-c589-494a-a8b3-7bc8ec4f7619","arxiv_id":"2607.27716","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A sparse four-corner RIS plus coded scalar observations can reconstruct a virtual full-aperture surface field and perform 2D DoA estimation nearly as accurately as a fully programmable RIS.","lead":"This paper proposes V-RIS, a way to do high-resolution angle-of-arrival sensing with a reconfigurable surface that programs only four corner patches instead of the whole aperture. A single receiver plus a neural representation reconstructs the missing surface field, and an outdoor test with 25% programmed elements kept errors within about one degree.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Direct-LS baseline nearly matches V-RIS DoA accuracy, so the paper does not establish that recurrence-based virtual-aperture reconstruction is what enables sparse-aperture performance.","rationale":"The reader's weakest assumption concerns model mismatch (K unknown, far-field, LoS), which is a genuine correctness risk for the reconstruction mechanism. My concern is different: even under the paper's ideal assumptions, the empirical advantage of V-RIS over a much simpler direct estimator is negligible in Table I. The paper's central claim is not merely that the reconstruction is possible, but that it is what enables sparse-aperture DoA accuracy. Direct-LS nearly matches that accuracy without reconstruction, so the causal claim is unsupported. This is a load-bearing issue because it affects the significance of the contribution, not just its robustness. However, it does not falsify the paper; the reconstruction might still be valuable for other applications, and V-RIS does produce the cleanest spectrum. Thus a CONDITIONAL verdict remains appropriate, with the added condition that the paper must demonstrate either a clear DoA-accuracy advantage over direct estimation in harder regimes or a concrete use of the reconstructed field. The reader already noted the Direct-LS near-match in the rationale, so there is partial agreement, but the reader's formal 'weakest_assumption' was focused on the model-mismatch risk, which is a different (though related) concern.","tokens_in":19449,"tokens_out":10361,"duration_ms":114488,"concrete_test":"Re-run the Table I experiment at N=50 and η=12.5% with 100 Monte Carlo trials, and again with K=5 targets at N=200, comparing V-RIS and Direct-LS DoA RMSE. If Direct-LS stays within 0.1° of V-RIS in both settings, the claim that virtual-aperture reconstruction enables the sparse-aperture accuracy is not supported by the evidence; the paper should then demonstrate a concrete task (e.g., imaging or multi-target tracking) that requires the reconstructed field.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I shows Direct-LS—a direct least-squares estimator that uses the same 200 coded observations and the same four-corner sparse deployment—recovering all three targets with errors (0.02°,0.02°), (0.04°,0.02°), and (0.05°,0.10°), whereas V-RIS yields (0.01°,0.00°), (0.01°,0.00°), and (0.01°,0.01°). Both are far more accurate than SHGD-O and RIS-NEAR, so the observed DoA accuracy may simply be a property of the large sparse aperture and coded observations, not of the proposed virtual-aperture reconstruction. The central claim is that the finite-order spatial recurrence lets the missing field be reconstructed, thereby enabling sparse-RIS DoA accuracy comparable to a full aperture. But Direct-LS achieves comparable accuracy without reconstructing any virtual aperture, undermining the causal role assigned to the reconstruction. The paper states that Direct-LS 'does not recover a reusable virtual-aperture surface field,' yet no experiment demonstrates a downstream task that benefits from such a field. Without a head-to-head comparison in regimes where reconstruction is supposedly essential—lower snapshot counts, lower deployment ratios, or larger target numbers—the experiments do not isolate the contribution of the central mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes V-RIS, a framework for 2D DoA estimation with a sparsely programmed RIS and a single-antenna receiver. The RIS physically deploys only four corner subarrays, and the missing aperture is treated as a virtual aperture. The full virtual-aperture surface field is reconstructed from RIS-coded scalar observations using an implicit neural representation (INR) enforced by two constraints: a bias-invariant receiver-domain data loss and a propagation-consistency loss based on a finite-order spatial recurrence of far-field surface fields. DoA estimation is then performed with a 2D Bartlett spectrum on the reconstructed field. The paper includes a four-corner deployment geometry justified by CRB reasoning, extensive simulations comparing against sparse-completion and direct-localization baselines, ablation studies, scalability experiments, robustness tests, and an outdoor prototype with 25% programmable elements.","tokens_in":19753,"tokens_out":3900,"duration_ms":40963,"significance":"If the central claim holds, V-RIS would provide a practical way to decouple effective sensing aperture from the number of programmable RIS elements, which is a timely and relevant result for low-cost RIS-aided localization. The paper has several strengths: Proposition 1 gives a clean closed-form affine nuisance alignment; Proposition 2 is a standard linear-prediction property and is properly attributed to the NEAR method; the outdoor prototype is a valuable proof of concept, showing sub-degree errors with only 25% programmable elements. However, the experiments do not currently isolate the contribution of the recurrence-based virtual-aperture reconstruction, because the Direct-LS baseline—which performs no such reconstruction—achieves nearly the same DoA accuracy. In addition, an ablation (V-RIS w/o Bias) behaves in a way that is inconsistent with the stated simulation protocol. These issues bear directly on the paper's main claim and need to be resolved before the contribution can be assessed.","major_comments":[{"comment":"Direct-LS achieves DoA errors of (0.02°,0.02°), (0.04°,0.02°), and (0.05°,0.10°) for the three targets, while V-RIS yields (0.01°,0.00°), (0.01°,0.00°), and (0.01°,0.01°). The paper's central claim is that recurrence-based virtual-aperture reconstruction enables sparse-aperture performance, but Direct-LS obtains comparable accuracy without reconstructing any virtual aperture. This does not refute V-RIS, but it means the experiments do not demonstrate the causal role of the reconstruction. The authors should add head-to-head comparisons in regimes where reconstruction is argued to be essential: lower N, lower deployment ratios η, larger K, or scenarios with model mismatch (e.g., near-field or multipath). Without such results, the 'enabling' claim is not supported.","section":"§V-B, Table I and Fig. 5"},{"comment":"The ablation 'V-RIS w/o Bias' collapses completely, but the simulation protocol in §V-A synthesizes observations according to (1) with no global gain or configuration-invariant additive term. Under that protocol, the bias-invariant loss should reduce to the ordinary data loss, and the two variants should behave almost identically. The observed collapse indicates either an unstated bias in the simulation or an implementation artifact. This must be explained. Similarly, 'V-RIS w/ Full Dep.' fails despite having more deployed elements and the same N=200; the paper attributes this to underdetermination, but the sparse four-corner case also has far fewer observations than unknowns. The explanation in terms of phase-configuration diversity is not convincing without a controlled experiment or convergence diagnostics.","section":"§V-D, Table II"},{"comment":"The four-corner deployment is a load-bearing design choice, but the CRB argument in Eq. (16) is taken from direct-array direction-cosine CRBs and is applied heuristically to a single-antenna RIS-coded observation model with INR reconstruction. The layout-dependent CRB term does not account for the coding mechanism, the number of snapshots, or the reconstruction error of the field. The conclusion that support must be confined to the corners is therefore not justified by the stated equations. Either derive a CRB for the actual coded-inverse problem or provide a more direct experiment that rules out other layouts while controlling for the number of deployed elements and measurement budget.","section":"§IV-C, Prop. 3"},{"comment":"Proposition 2 holds only when the target–RIS channel is a superposition of exactly K far-field plane waves and K is known. All simulations set K equal to the true target count (Sec. V-A), and the prototype has a single target. The text suggests AIC/MDL or validation residuals for selecting K, but no experiment tests model-order mismatch (K too small, K too large, or an unknown K). Since the recurrence loss is a hard structural constraint, misspecification of K can violate the propagation-consistency term and distort the reconstructed field. A sensitivity experiment with K mismatch is needed, or the paper should explicitly scope the claims to known K.","section":"§IV-B and §V"}],"minor_comments":[{"comment":"The axis labels in Fig. 7 appear garbled (e.g., '5 01 502 503 50/s8722/s51/s48...'). The horizontal and vertical axes should be clearly labeled as 'Number of snapshots N' and 'NMSE (dB)'.","section":"§V-A / Fig. 7"},{"comment":"The simulation setting states that target gains are constant over N configurations and noise is AWGN, but it does not specify how many Monte Carlo trials are averaged for the main comparisons in Table I and Fig. 5. This should be stated for reproducibility and to indicate statistical significance.","section":"§V-A"},{"comment":"The label 'V-RIS w/ Full Dep.' is confusingly similar to the 'full-aperture benchmark' mentioned elsewhere (e.g., the oracle field in Fig. 5 and the full-aperture baseline in the prototype). Clarify the distinction between the oracle full-aperture field, the full-deployment reconstruction variant, and the prototype full-aperture programming.","section":"§V-D, Table II"},{"comment":"The prototype reports a single measured realization for each configuration. Since the hardware setup includes quasi-static coupling and switching effects, reporting repeated measurements or a small number of trials would strengthen the claim that the error increments (0.41°, 0.21°) are not due to a single unlucky realization.","section":"§VI, Table III"},{"comment":"There are minor typographical issues, e.g., 'involved' in the introduction and the phrase 'jointly seeks a full-aperture estimate bH and a deployed set' is slightly ungrammatical. These do not affect the technical content.","section":"Intro and §II"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely problem and the prototype is a useful contribution. However, the strongest challenge is the Direct-LS baseline: the current experiments do not show that the virtual-aperture reconstruction is what provides the DoA accuracy, and the ablation inconsistency (V-RIS w/o Bias) suggests that the simulations may contain an unstated implementation detail. The authors should be asked to either add experiments that isolate the reconstruction mechanism or substantially temper the central claim. The CRB-based deployment justification also needs tightening for the coded-observation setting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a read if you work on RIS-aided DoA. The core trick is to treat the far-field surface field as satisfying a finite-order spatial recurrence, so a sparse four-corner deployment plus INR can reconstruct the full aperture from a single-antenna receiver's coded scalar observations. Prop. 2 is correctly attributed to NEAR, Prop. 1 is a standard closed-form affine nuisance elimination, and the geometry argument for four-corner deployment via CRBs is reasonable. The bias-invariant loss is a nice practical touch, and the outdoor prototype, while single-shot, is real evidence that the pipeline works at 25% programmability with sub-degree error.\n\nThe main soft spot is the causal claim. In Table I, Direct-LS—which uses the same coded observations and the same sparse corner deployment—nearly matches V-RIS on all three targets. So the paper doesn't yet demonstrate that virtual-aperture reconstruction is what enables the sparse-aperture performance; the large sparse aperture and coded observations themselves might be doing most of the work. The authors say Direct-LS doesn't yield a reusable field, but they never show a downstream task that benefits from that field. A head-to-head at lower snapshot counts, lower deployment ratios, or larger target counts would help isolate the contribution.\n\nTwo smaller issues. The ablation in Table II is internally confusing: V-RIS w/o Bias collapses, but the stated simulation protocol (Sec V-A) doesn't inject a global gain/offset or configuration-invariant bias, so it's unclear what the bias-invariant loss is removing. And the full-deployment ablation performance drop looks counterintuitive and deserves an explanation (maybe the fixed N=200 with random 1-bit codes over 4096 elements is under-determined, but that needs saying). Also, everywhere in simulation K is set equal to the true target count; robustness to unknown K and multipath is only mentioned as future work.\n\nOverall, the math is standard and the engineering is careful. No code or data are available for independent verification, which is a limitation but not disqualifying for a prototype paper. The paper deserves a serious referee and could be accepted after the ablation inconsistency is fixed and after a clearer demonstration that the recurrence prior earns its keep against the direct baseline. I'd suggest a conditional positive recommendation, with the isolation experiments as the main requested revision.","headline":"A plausible and well-executed combination of NEAR-style recurrence with RIS-coded observations, but the experiments don't yet pin down whether virtual-aperture reconstruction is what buys the accuracy.","tokens_in":20268,"tokens_out":1762,"would_cite":true,"duration_ms":19162,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","94A13"],"pacs":[],"model":"deepseek-v4-flash","headline":"Sparse RIS with 25% of elements matches full-aperture DoA","keywords":["reconfigurable intelligent surface","direction-of-arrival estimation","virtual aperture","spatial recurrence","implicit neural representation","sparse deployment","surface-field reconstruction","single-antenna receiver"],"falsifier":"Deploy the same four-corner setup with a target in the near field (e.g., at 1 m for the 16x16 prototype) and run V-RIS with K equal to the declared target count; if the reconstructed field's Bartlett spectrum peaks are displaced by more than a few degrees, or the recurrence residual stays large, the far-field recurrence premise is falsified. A cleaner test: feed receiver observations synthesized from K=3 plane waves but run the algorithm with K=2; the recurrence residual should fail to vanish and DoA estimates should degrade.","tokens_in":19292,"feed_emoji":"📡","tokens_out":3027,"duration_ms":30740,"temperature":0.7,"pith_summary":"The paper claims that the angular resolution of a large reconfigurable intelligent surface can be achieved by programming only four small corner subarrays, because the field across the full aperture is determined by a finite-order spatial recurrence. A single-antenna receiver collects coded scalar observations under many phase configurations; a neural representation of the surface field is fit to those observations while enforcing the recurrence. The reconstructed virtual-aperture field then feeds a standard 2D DoA estimator. If correct, this decouples sensing aperture size from the number of programmable elements, cutting hardware and control cost. Simulations and an outdoor prototype with 25% programmable elements report sub-degree elevation and azimuth errors.","feed_headline":"Four corner subarrays reconstruct a full RIS aperture for DoA","feed_subtitle":"A physics-based recurrence recovers the missing surface field, keeping angular errors near 1 degree at a fraction of the hardware cost.","key_machinery":"The finite-order spatial recurrence of the far-field surface field (Proposition 2): for K plane-wave targets, each field entry equals a fixed linear combination of the K previous entries along the same row or column. This recurrence links deployed and virtual elements, turning an ill-posed inverse problem into a constrained reconstruction. The four-corner deployment supplies the K+1 consecutive samples needed to estimate the recurrence coefficients while preserving large aperture baselines. A closed-form affine alignment eliminates global complex gain and configuration-invariant additive bias from the receiver loss.","core_discovery":"Under far-field illumination, the discretized RIS surface field obeys a K-th order linear recurrence along both aperture axes, because each target contributes a rank-one planar phase progression. V-RIS uses this recurrence as a propagation-consistency constraint to extend the field from four corner subarrays (which provide contiguous local samples and long baselines) to the whole virtual aperture. The reconstructed field is then processed by Bartlett beamforming. The paper reports that with only 25% of elements programmable, DoA errors stay within about a degree of a full-aperture baseline, in simulation and in an outdoor 5.8 GHz prototype.","pith_inferences":["A testable extension is near-field generalization: replacing the plane-wave recurrence with a spherical-wave one would extend V-RIS to short-range settings, where the current model would fail.","The recurrence prior is essentially a structured low-rank/latent-subspace assumption; one could compare it against matrix-completion baselines on the same coded observations to see how much the physics prior adds over generic low-rankness.","Because the receiver only sees scalar projections, the number of independently recoverable targets may be bounded by the corner-block size; the paper's K-known assumption hides this limit — a model-order mismatch experiment would reveal it.","If the configuration-invariant additive term is actually configuration-dependent (e.g., moving scatterers), the affine alignment in the loss would break; a dynamic-multipath test would expose that boundary."],"forward_implications":["If correct, high-resolution DoA sensing becomes cheaper: the number of programmable elements no longer determines aperture size.","The reconstructed virtual-aperture field is reusable for other array-processing tasks, not just the specific estimator used here.","The finite-order recurrence provides a physical prior that could transfer to other coded-aperture or sparse-array sensing problems.","Measurement budget scales with the number of corner elements and target count, not with aperture size, making very large surfaces feasible.","Sub-degree accuracy with 1-bit phase control suggests that cheap, coarse RIS hardware is sufficient for angular sensing."],"fun_headline_variants":["Four corners, full-aperture DoA: V-RIS","Sparse RIS: virtual aperture from 4 subarrays","25% RIS programmability, 1° DoA error","Spatial recurrence enables RIS virtual aperture"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The target-to-RIS channel must be a superposition of exactly K far-field line-of-sight plane waves with the direct link blocked; if the target count K is wrong, the scene is near-field, or multipath varies with RIS configuration, the recurrence constraint does not describe the true field and reconstruction collapses.","fun_headline_variants_meta":{"raw":{"variants":["Four corners, full-aperture DoA: V-RIS","Sparse RIS: virtual aperture from 4 subarrays","25% RIS programmability, 1° DoA error","Spatial recurrence enables RIS virtual aperture"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000236,"raw_usage":{"total_tokens":1366,"prompt_tokens":793,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":505}},"tokens_in":537,"tokens_out":573,"duration_ms":5898,"temperature":1.0,"reasoning_tokens":505,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T02:44:14.581662+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the same four-corner setup with a target in the near field (e.g., at 1 m for the 16x16 prototype) and run V-RIS with K equal to the declared target count; if the reconstructed field's Bartlett spectrum peaks are displaced by more than a few degrees, or the recurrence residual stays large, the far-field recurrence premise is falsified. A cleaner test: feed receiver observations synthesized from K=3 plane waves but run the algorithm with K=2; the recurrence residual should fail to vanish and DoA estimates should degrade.","supporting_citations":[],"review_version":1}