{"id":"31470e69-ca35-4db2-9988-fa48e4ba7bcf","arxiv_id":"2501.08776","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A dual-purpose DFT/polar codebook design coupled with near-field STAP is proposed to reduce ISAC beam-training overhead and cut STAP complexity by about three orders of magnitude, with simulation-only support.","lead":"This paper proposes a two-stage beam-training pipeline for near-field integrated sensing and communication: a coarse DFT codebook sweep, polar codebook refinement, and a low-complexity space-time adaptive processing step for detecting slow-moving targets. The authors report roughly 1000x lower STAP complexity from narrowing the search space, but the algorithm details needed to verify that reduction are not included.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Complexity reduction depends on an unspecified reduced-dimension NF-STAP; the O(lc(M nc)^3) claim is an assumed model, not a derived result, and its training-sample support is unverified.","rationale":"The reader's weakest-assumption analysis correctly identifies the central gap. I re-read the relevant sections and confirmed that 'reduced-dimension NF-STAP' is named but not defined: Section 'Low-Complexity NF-STAP' only gives the before/after big-O expressions, and the case study's 'Reduced-dimension NF-STAP' paragraph provides no algorithm. The complexity reduction is therefore not a tested result but a placeholder. This is load-bearing because the abstract and conclusion advertise 'three orders of magnitude' complexity reduction as the main validation, and the simulation section uses that reduced-dimension method to produce the 15 dB SINR and rate curves. The lack of an algorithm also prevents any independent check of the 4M nc training-sample statement; 4M nc=4096 samples for a 1024x1024 covariance may be reasonable, but the paper does not say where those samples come from if only lc range cells are processed or if L is hundreds. These are correctable omissions, not evidence of a false result. The framework is plausible and builds on cited prior work on polar codebooks and near-field propagation, so the appropriate verdict remains CONDITIONAL. Since the reader already chose CONDITIONAL, my stress test does not change the verdict.","tokens_in":8709,"tokens_out":7490,"duration_ms":81621,"concrete_test":"Specify the missing algorithm in the form of a beamspace reduced-dimension STAP: project each of the M pulse snapshots onto an nc-column beamspace matrix (e.g., the nc strongest DFT/polar beams), estimate the (M nc)x(M nc) covariance from K training range cells, and invert it once per CPI. Then compute the exact flop count for the Fig. 5 case study with the paper's L, lc, nc, M, N values and verify two conditions: (i) the actual complexity ratio to O(L(MN)^3) is >= 1000; (ii) K >= 2M nc (RMB rule) or an explicit reduced-rank sample-support argument is satisfied with the available number of range cells. If either condition fails, the claimed complexity reduction or the 15 dB SINR gain cannot be sustained.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central quantitative claim, a reduction from O(L(MN)^3) to O(lc(M nc)^3) with ratio LN^3/(lc nc^3) >= 1000, depends entirely on a 'reduced-dimension NF-STAP' that is never specified. In the section 'Low-Complexity NF-STAP', the reduction is asserted as the complexity expression, and the case study invokes 'Reduced-dimension NF-STAP' without defining the preprocessor, the reduced covariance estimator, the training-cell selection, or the inversion procedure. Without those details, O(lc(M nc)^3) is not a derived complexity; it is an assumed model. The assumption is also in tension with the stated training-sample count. For the case-study parameters M=128 and nc=8, the reduced covariance matrix is 1024x1024, and the text says 4M nc = 4096 training samples approach optimal STAP performance. If L is on the order of hundreds, as the complexity argument assumes, the radar data cube cannot supply 4096 independent range-bin snapshots. If L is large enough to supply them, the cost of forming and exploiting those snapshots must be included in the complexity budget, but no such accounting appears. The 1000x reduction and the accompanying SINR/rate results therefore rest on an unverified algorithmic and sample-support assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a unified near-field ISAC framework in which a DFT codebook provides coarse range/angle estimates from the angular spread of NF users, a polar codebook refines those estimates within a candidate polar region, and a customized NF-STAP detector uses the refined range/angle candidates to reduce computational complexity. The central quantitative claim is that the STAP complexity falls from O(L(MN)^3) to O(lc(M nc)^3), yielding a reduction factor LN^3/(lc nc^3) of at least 1000, while preserving detection and communication performance. Simulation results report a 15 dB SINR gain over conventional Doppler filtering, near-optimal STAP performance with 4M nc training samples, and communication rates close to the perfect-CSI benchmark.","tokens_in":8915,"tokens_out":2933,"duration_ms":32310,"significance":"If the claimed complexity reduction and performance are substantiated, the paper would make a useful contribution to NF-ISAC by connecting beam-training codebooks with adaptive radar processing. The strength of the paper is its system-level vision: reusing DFT beam-training signals for coarse sensing and polar codebook candidates to narrow the STAP search is a sensible architectural idea, and the SINR/rate results are suggestive. The paper also explicitly identifies practical issues such as heterogeneous training cells and hardware impairments. However, the central complexity reduction rests on a reduced-dimension NF-STAP that is never specified, the sample-support argument is not reconciled with the stated complexity model, and the simulation evidence is presented without error bars or statistical detail. These gaps must be addressed before the quantitative claims can be accepted.","major_comments":[{"comment":"The reduction from O(L(MN)^3) to O(lc(M nc)^3) is asserted rather than derived. The 'reduced-dimension NF-STAP' is never defined: the text does not specify the preprocessor that maps the full MN-dimensional snapshot to dimension M nc, the reduced covariance estimator, the training-cell selection, or the inversion procedure. Without these details, O(lc(M nc)^3) is an assumed complexity model, not a result. Please provide the explicit reduced-dimension algorithm and a step-by-step complexity count, including the cost of forming and inverting the reduced covariance matrix.","section":"Section IV, Step 3 (Low-Complexity NF-STAP) and Section V (NF-STAP Case Study)"},{"comment":"The claimed complexity reduction is inconsistent with the stated training-sample requirement. For M=128 and nc=8, the reduced covariance matrix is 1024x1024, and the text says that 4M nc = 4096 training samples approach optimal performance. If L is on the order of hundreds, as the complexity argument assumes, the radar data cube cannot supply 4096 independent range-bin snapshots. If L is large enough to supply them, then the cost of forming and exploiting those L snapshots must be included in the complexity budget. This tension undermines the 1000x reduction claim and the simulation setup; please clarify how the training samples are obtained and how their cost is accounted for.","section":"Section V, NF-STAP Case Study and Section IV, Step 3"},{"comment":"The statement that 'the reduction factor is surely LN^3/(lc nc^3) ≥ 1000, given that N and L are on the order of hundreds, while lc and nc are on the order of tens' is not justified by those order-of-magnitude assumptions. For example, with L=100, N=100, lc=90, and nc=90, the ratio is approximately 1.5, far below 1000. Please replace the 'surely' claim with explicit ranges for L, N, lc, and nc under which the 1000x reduction holds, and state the actual parameters used in the simulation.","section":"Section IV, Step 3, complexity-reduction factor statement"},{"comment":"The quantitative performance claims, including the 15 dB SINR gain over Doppler filtering and the statement that 4M nc training samples 'closely approaches the optimal STAP performance,' are based on simulation curves without error bars or confidence intervals. Since the central evidence for the framework is these curves, please provide multiple independent trials or an analytical variance estimate, and report the number of Monte Carlo runs used to generate each panel.","section":"Section V, Figure 5 and accompanying text"}],"minor_comments":[{"comment":"The DFT angular-spread mapping and the lookup-table construction rely on reference [13], which is the authors' own arXiv preprint and is not summarized or validated in the paper. A brief derivation or a reproduction of the key mapping would help the reader assess the coarse-estimation step.","section":"Section IV, Step 1"},{"comment":"The EBRD concept and the polar codebook sampling are taken from reference [11], which is listed as 'Submitted to IEEE Trans. Wireless Commun.' and is not yet peer-reviewed. Please either provide the relevant definitions in the paper or cite a peer-reviewed source.","section":"Section II and reference [11]"},{"comment":"The statement 'beamwidth is constant in both NF and FF' should be qualified; the beamwidth is constant for a given array and steering direction in the FF, but the NF beam characteristics vary with focus distance and angle. Please clarify to avoid ambiguity.","section":"Section II, 'Lateral vs. Axial Resolution'"},{"comment":"The simulation setup omits several specific parameters such as the noise and clutter statistics, the number of Monte Carlo runs, the exact polar codebook sizes, and the construction of the reduced-dimension STAP. These should be included for reproducibility.","section":"Section V, Case study parameters"},{"comment":"The phrase 'ookup tables' on page 7 should be 'lookup tables'. There is also a missing space in 'Mutli-modal' in the futures section heading; these typos should be corrected.","section":"Conclusion and Future Directions"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own prior work for three load-bearing components: the DFT angular-spread mapping [13], the EBRD/polar codebook design [11], and the low-complexity STAP concept that is only named, not specified, in this manuscript. These are disclosed self-citations rather than hidden duplication, but the editor may wish to consider whether the contribution of this paper is sufficiently self-contained for the journal's readership. The main technical risk is the unspecified reduced-dimension STAP: the complexity reduction and the sample-support argument cannot be evaluated until that algorithm is fully described."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read on 2501.08776. The genuinely new thing is the system integration: DFT angular-spread mapping for coarse range-angle estimates, polar codebook refinement, and a NF-STAP whose search space is narrowed by the codebook outputs. That combination, with the 28 GHz slow-target case study, is not in the prior work. The paper also does a decent job explaining why NF beam focusing changes the STAP problem, and the self-citations to the authors' own DFT mapping and polar boundary work are disclosed and relevant rather than hidden.\n\nThe soft spot is exactly where the reader puts it. The central complexity reduction, from O(L(MN)^3) to O(lc(M nc)^3) with a claimed 1000x factor, is asserted rather than derived. The paper never defines the 'reduced-dimension NF-STAP' it uses in the case study—no preprocessor, no reduced covariance estimator, no training-cell selection, no inversion procedure. Without that, O(lc(M nc)^3) is an assumed model, not a result. The stress-test note adds a real inconsistency: the case study says 4M nc = 4096 training samples approach optimal STAP, but if L is on the order of hundreds—as the complexity argument assumes—the data cube cannot supply 4096 independent training snapshots. If L is large enough to supply them, the cost of forming those snapshots belongs in the complexity budget. Neither accounting appears. The paper actually lists 'reduced-dimension STAP' as a future direction in the conclusion while using it in the simulations, which is a tell.\n\nThe simulation curves are suggestive but lack error bars, and no code or detailed parameter settings are provided, so the performance claims are hard to check. These are all addressable, not fatal. The framework is plausible and worth a serious referee, but it needs major revision: specify the reduced-dimension STAP algorithm, derive the complexity with sample-support constraints, add sensitivity analysis, and release enough detail to reproduce the 15 dB SINR gain and rate curves.\n\nI'd take it to review, not desk reject, but I would not cite it in its current form.","headline":"Plausible framework for NF-ISAC beam training plus STAP, but the headline 1000x complexity reduction rests on an unspecified reduced-dimension STAP and a training-sample inconsistency.","tokens_in":9507,"tokens_out":3028,"would_cite":false,"duration_ms":29861,"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":"Near-field ISAC can cut space-time adaptive processing cost by three orders of magnitude by feeding it codebook-based position estimates.","keywords":["near-field integrated sensing and communication","space-time adaptive processing","polar codebook","DFT codebook","beam training","clutter suppression","ultra-massive MIMO"],"falsifier":"Place a target just outside the polar-codebook candidate window that the DFT stage selects, run the proposed NF-STAP, and compare detection with a full-dimension STAP; a miss in the reduced-dimension processor would show the search-space confinement premise is false. Alternatively, measure the wall-clock time of both processors on the same 256-element, 128-pulse CPI; if the realized speedup is far below 1000x, the complexity claim fails.","tokens_in":8445,"feed_emoji":"📡","tokens_out":6851,"duration_ms":67275,"temperature":0.7,"pith_summary":"This paper argues that near-field integrated sensing and communication can be made practical by letting the communication codebook feed the radar processor: a DFT codebook finds a coarse position, a polar codebook refines it, and the refined candidates shrink a space-time adaptive processor's search so much that its cost drops by roughly three orders of magnitude. This matters because conventional STAP is computationally prohibitive for ultra-massive MIMO arrays, and because slow-moving targets are otherwise buried in clutter. On the paper's terms, the complexity falls from $O(L(MN)^3)$ to $O(l_c(M n_c)^3)$, with a reduction factor of at least 1000. The paper also demonstrates a 15 dB SINR gain over Doppler filtering and rate performance close to perfect-CSI bounds, so the claimed gain is not only computational.","feed_headline":"Near-field ISAC cuts sensing cost 1000x via codebook training","feed_subtitle":"DFT beams give coarse position, polar beams refine it, and STAP only searches that tiny window.","key_machinery":"The load-bearing mechanism is the dual-purpose codebook chain. A DFT codebook, already used for initial access, measures angular spread and matches it to a lookup table to get coarse range and angle. A polar codebook, sampled in range by beam-depth inside the effective beam-focused Rayleigh distance (EBRD), refines this to $l_c$ range and $n_c$ angle candidates. These candidates define the search space of a reduced-dimension NF-STAP whose steering vector is the Hadamard product of a Doppler vector and a spherical-wavefront spatial steering vector. The reduction in search space, not any new radar hardware, is what converts the cubic cost in $MN$ into a cubic cost in the much smaller $M n_c$.","core_discovery":"The central claim is that near-field STAP, normally too expensive for ultra-massive MIMO because its covariance matrix is $MN \\times MN$, can be made practical by preceding it with codebook-based beam training. The DFT codebook's angular spread gives a coarse range-angle estimate; the polar codebook then restricts the target to $l_c$ candidate range bins and $n_c$ candidate angle samples; and the STAP processor needs to evaluate only those candidates. On the paper's terms this lowers the complexity from $O(L(MN)^3)$ to $O(l_c(M n_c)^3)$, a factor $LN^3/(l_c n_c^3) \\ge 1000$, while a case study shows a 15 dB SINR gain over Doppler filtering and a rate close to perfect-CSI performance.","pith_inferences":["A testable extension is to make the polar-codebook window size adaptive: since the reduction factor is $LN^3/(l_c n_c^3)$, shrinking $l_c$ and $n_c$ as the DFT estimate becomes more reliable would buy even more speedup, at the cost of a higher miss probability when the target sits outside the window.","The DFT lookup table encodes angular spread as a function of range, angle, and carrier frequency; the same construction could be regenerated for other bands or array geometries, and its range resolution could be improved by using multiple frequencies jointly.","The architecture suggests a general recipe for ISAC: any coarse localizer, not only a DFT codebook, can feed a constrained STAP search; camera or LiDAR cues would plug into the same slot and inherit the complexity reduction, a direction the paper lists as future work."],"forward_implications":["With $N=256$, $M=128$, and typical $l_c$, $n_c$ on the order of tens, NF-STAP becomes feasible on a per-CPI basis rather than a batch offline computation.","Initial access sweeps can double as sensing: the DFT stage needs no extra reference signals, so the communication overhead stays at 5G levels.","Slow-moving targets whose Doppler is buried in mainlobe clutter are recoverable, with about 15 dB SINR gain over plain Doppler filtering.","After refinement with only a handful of extra beams, achievable rate approaches the perfect-CSI upper bound, so sensing does not come at a communication cost.","The same angular-spread lookup table can be reused for beam tracking, so the framework covers both initial access and mobility scenarios."],"supporting_citations":[{"why":"Supplies the angular-spread-to-position mapping that the DFT stage uses for coarse range and angle estimation.","marker":"[13]"},{"why":"Defines the EBRD boundary and beam-depth-based sampling used to size the polar codebook candidates.","marker":"[11]"},{"why":"Provides the two-stage DFT-angle then polar-range beam training approach that the paper adapts.","marker":"[3]"},{"why":"Serves as the near-field hierarchical beam training baseline against which the proposed framework compares rate and overhead.","marker":"[4]"},{"why":"Provides a subarray-based two-stage beam training alternative that the paper contrasts with its own two-stage codebook design.","marker":"[5]"},{"why":"Supplies the STAP covariance estimation, training-cell, and reduced-dimension processing concepts used in the NF-STAP formulation.","marker":"[12]"},{"why":"Establishes the finite beam-depth concept that underlies range-domain resolution in near-field beam focusing.","marker":"[7]"},{"why":"Motivates dual-purpose use of periodic DFT beam-sweeping signals for sensing in 5G-Advanced and 6G.","marker":"[6]"}],"fun_headline_variants":["Near-field ISAC slashes STAP cost 1000x via codebooks","Codebook duo makes near-field STAP 1000x cheaper","DFT and polar codebooks cut ISAC sensing complexity 1000x","Near-field ISAC: codebooks shrink STAP burden 1000x","1000x STAP complexity cut with dual-purpose codebooks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole thousandfold speedup rests on the premise that the codebook estimates shrink the target to a handful of candidate range-angle bins, and that the cheaper adaptive filter used on those bins is actually the same kind of STAP as the full one; the paper assumes this without specifying the reduced-dimension filter.","fun_headline_variants_meta":{"raw":{"variants":["Near-field ISAC slashes STAP cost 1000x via codebooks","Codebook duo makes near-field STAP 1000x cheaper","DFT and polar codebooks cut ISAC sensing complexity 1000x","Near-field ISAC: codebooks shrink STAP burden 1000x","1000x STAP complexity cut with dual-purpose codebooks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000605,"raw_usage":{"total_tokens":2823,"prompt_tokens":948,"completion_tokens":1875,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1778}},"tokens_in":564,"tokens_out":1875,"duration_ms":11743,"temperature":1.0,"reasoning_tokens":1778,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:17:56.529474+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place a target just outside the polar-codebook candidate window that the DFT stage selects, run the proposed NF-STAP, and compare detection with a full-dimension STAP; a miss in the reduced-dimension processor would show the search-space confinement premise is false. Alternatively, measure the wall-clock time of both processors on the same 256-element, 128-pulse CPI; if the realized speedup is far below 1000x, the complexity claim fails.","supporting_citations":[{"cited_title":"Redefining polar boundaries for near-field channel estimation for ultra-massive MIMO antenna array,","cited_arxiv_id":null,"evidence_quote":"Defines the EBRD boundary and beam-depth-based sampling used to size the polar codebook candidates."},{"cited_title":"Fast near-field beam training for extremely large-scale array,","cited_arxiv_id":null,"evidence_quote":"Provides the two-stage DFT-angle then polar-range beam training approach that the paper adapts."},{"cited_title":"Hierarchical beam training for extremely large-scale MIMO: From far-field to near-field,","cited_arxiv_id":null,"evidence_quote":"Serves as the near-field hierarchical beam training baseline against which the proposed framework compares rate and overhead."},{"cited_title":"Two-stage hierarchical beam training for near-field communications,","cited_arxiv_id":null,"evidence_quote":"Provides a subarray-based two-stage beam training alternative that the paper contrasts with its own two-stage codebook design."},{"cited_title":"Adaptive Doppler compensation for mitigating range dependence in forward-looking airborne radar,","cited_arxiv_id":null,"evidence_quote":"Supplies the STAP covariance estimation, training-cell, and reduced-dimension processing concepts used in the NF-STAP formulation."},{"cited_title":"Finite beam depth analysis for large arrays,","cited_arxiv_id":null,"evidence_quote":"Establishes the finite beam-depth concept that underlies range-domain resolution in near-field beam focusing."},{"cited_title":"Downlink sensing in 5G-advanced and 6G: SIB1-assisted SSB approach,","cited_arxiv_id":null,"evidence_quote":"Motivates dual-purpose use of periodic DFT beam-sweeping signals for sensing in 5G-Advanced and 6G."}],"review_version":1}