{"id":"654851e4-094a-428c-bdfb-2f4dc5bf994c","arxiv_id":"2507.13463","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A two-stage estimator uses 2D-DFT spread widths to coarsely infer range and transverse velocity, then refines angle, range, and velocities with 1D MUSIC, achieving -40 dB NMSE at high SNR.","lead":"The paper proposes a low-complexity method for estimating the location and full velocity of a near-field target using a large antenna array, by reading the width and center of angular and Doppler spreads in a 2D Fourier transform. The method is aimed at 6G sensing scenarios where a single base station must detect both radial and transverse motion.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Doppler-spread analysis in Sec. III-B is spatially, not temporally, defined: Eq. (13) is an angle-domain projection of one pulse, and the vr term cancels in the magnitude, so the claimed center/width relationships for the Doppler spread are unsupported.","rationale":"The reader's weakest assumption (1D MUSIC under a multi-rank spatial signature) is a legitimate secondary concern for the refinement stage, but the more load-bearing issue is upstream: the paper's derivation of the Doppler spread itself. In Section III-B, Eq. (13) projects the spatial Doppler steering vector b(m) onto a spatial DFT codebook f(θn); the variable being scanned is the angle θn, not a Doppler frequency, and the radial-velocity term vr appears as a constant phase independent of n, so it disappears from the magnitude. The claim that the Doppler gain pattern is shifted by vr and that its midpoint estimates vr does not follow from Eq. (13). This matters because the coarse estimation stage uses the median of the Doppler spread for vr (Eq. 15) and builds a lookup table Kv relating Doppler spread to transverse velocity. If the displayed derivation is wrong, the central claim is not established; if it is merely a typo, the correct temporal DFT derivation is absent and must be supplied. I keep the CONDITIONAL verdict because the underlying qualitative observation may be salvageable: a 2D-DFT of the exact snapshot likely does produce a Doppler spread whose center is set by vr and whose width is set by vθ, but the paper must demonstrate this explicitly. I also note the Eq. (7) dimension mismatch and the O((NBSM)^3) complexity line as additional correctness/consistency issues, but they are less central than the missing Doppler-spectrum derivation. No ad hominem is intended; the critique is on the argument's analytical support.","tokens_in":7914,"tokens_out":9696,"duration_ms":113561,"concrete_test":"Construct the noiseless space-time snapshot X_{n,m}=a_n(θ,r)e^{-jπ m ω^{(n)}} with ω^{(n)}=2(v_r + v_θ n d cosθ / r)/(λ f_r), take the 2D DFT over antenna index n and time index m, and measure the 3-dB Doppler-spread width and centroid as functions of (v_r, v_θ, r). If the centroid does not track v_r or the width does not scale linearly with v_θ, then Eq. (13) and the lookup-table design are wrong. If the relationships do hold, rewrite Section III-B to derive the temporal DFT response explicitly and reconcile it with the spatial-projection formula of Eq. (13).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-B does not actually analyze the Doppler spectrum. Gdopp(vr,vθ;θn) in Eq. (13) is defined as |b^H(vr,vθ)f(θn)|^2, where b(m) is the NBS×1 spatial Doppler steering vector for the m-th symbol and f(θn) is a spatial DFT codebook; the summation runs over antenna index n. This is a spatial (angle) projection of one temporal snapshot, not a temporal DFT over m, so it cannot produce a Doppler spectrum. Moreover, the term containing vr in Eq. (13) is independent of n: it factors out of the squared magnitude, so the pattern is not shifted by vr. The text immediately after Eq. (13) claims the entire gain pattern is shifted by vr and that the midpoint of the Doppler spread gives vr; that conclusion does not follow from the displayed expression. Since the coarse radial-velocity estimate in Eq. (15) and the lookup table Kv are built on this relationship, the central claim that the Doppler spread width encodes vθ and its center encodes vr is unsupported as written. The simulations may be correct, but the analytical foundation is missing or mis-stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage estimator for the joint location and velocity of a moving target in the radiative near field of a UM-MIMO array. The first stage applies a 2D-DFT to the space-time snapshot and exploits the angular spread and the Doppler spread of the response to obtain coarse estimates of angle, range, radial velocity, and transverse velocity. The second stage refines each parameter with a separate 1D MUSIC scan. Simulation results with a 256-antenna array at 28 GHz report NMSEs approaching -40 dB at high SNR, comparable to polar-codebook benchmarks and better than gradient-based ML, at lower nominal complexity.","tokens_in":8114,"tokens_out":7838,"duration_ms":82141,"significance":"The problem is timely and important for near-field ISAC, and the idea of decoupling the four parameters through spectral spread characteristics is appealing. If the analytical characterization were correct, the proposed two-stage method would be a valuable low-complexity alternative to 4D ML search. The paper also explicitly models array calibration errors and shows graceful degradation. However, the central Doppler-spread analysis is currently mis-stated, and the range/transverse-velocity estimates rely on lookup tables generated from the same forward model used in the test simulations, so the evidence for the claimed spectral relationships is partly circular. The manuscript does not provide code or machine-checked proofs; the numerical claims are not independently reproducible from the text alone.","major_comments":[{"comment":"Equation (13) is not a Doppler spectrum. G_dopp(vr,vtheta; theta_n) is defined as an inner product over the antenna index n between a single-symbol Doppler steering vector b(vr,vtheta) and a spatial DFT codebook f(theta_n); it does not involve a temporal DFT over the symbol index m. Consequently, it cannot characterize the Doppler spread of the space-time snapshot. Moreover, the vr-dependent term inside the exponent in Eq. (13) is independent of n and therefore disappears from the squared magnitude, so the statement that the gain pattern is shifted by vr and that the midpoint of the Doppler spread gives vr does not follow from the displayed expression. Since Eqs. (14)-(15) and the lookup table K_v are built on this relationship, the analytical foundation of the velocity estimation is unsupported as written. Please provide a correct derivation of the 2D-DFT response of the space-time snapshot in Eq. (7), or present the approach as purely empirical and remove the claimed analytical relationships.","section":"III-B"},{"comment":"The signal model in Eq. (7) is dimensionally inconsistent. Y is defined as [V^1 s(1), ..., V^M s(M)] with V in C^{NBS x M} and s(m) in C^{NBS x 1}; the product V^m s(m) is not defined (V^m is not specified, and the dimensions do not match a standard narrowband receive model). This is load-bearing because the entire derivation and the definition of the space-time snapshot depend on it. Please clarify whether s(m) is a scalar transmitted symbol, whether V^m denotes a column or a diagonal matrix, and how the transmit vector is combined with the steering matrix.","section":"II-B"},{"comment":"The lookup tables K_a and K_v are populated by simulating the same forward model that is later used to generate the test data. The coarse range and transverse-velocity estimates therefore measure agreement with a self-consistent dictionary, not with an independent closed-form mapping. The claim that angular-spread width is uniquely linked to range and Doppler-spread width to transverse velocity needs to be supported by an independent analytical derivation (e.g., a closed-form relation between the 3-dB spread and the parameters) or by validation on data generated from a different forward model. As it stands, the simulations demonstrate self-consistency rather than predictive accuracy of the proposed spectral relationships.","section":"IV-A"},{"comment":"The 1D MUSIC refinement scans a single parameter-specific steering vector a(x) while keeping other parameters fixed. In the NF regime with Doppler spread, the spatial signature across the array can be multi-rank, and a single steering vector may not lie in the signal subspace. The paper does not bound or analyze the bias this introduces, and it is not obvious that sequential 1D scans over theta, r, vr, vtheta converge to the true parameters. Please provide a theoretical justification or a numerical subspace-distance analysis supporting the single-vector MUSIC scan in the multi-rank regime.","section":"IV-B"}],"minor_comments":[{"comment":"The summation limits in Eq. (10) are written as sum_{n=-NBS/2}^{NBS/2}, which is off by one for even NBS; please correct to sum_{n=-(NBS-1)/2}^{(NBS-1)/2} or an equivalent integer range.","section":"III-A"},{"comment":"The Doppler phase in Eq. (3) uses -j pi m omega^(n); a standard narrowband Doppler phase is -j 2 pi m omega^(n) (or +j 2 pi). Please resolve the factor-of-2 discrepancy or justify the definition of omega^(n).","section":"III-B"},{"comment":"The approximation v_r^(n) approximately v_r is described as neglecting O(1/r_F^2) terms, but it also discards the NF coupling of the radial velocity to the antenna index; please state explicitly that this is an additional modeling assumption beyond the Taylor expansion.","section":"II-A2"},{"comment":"The complexity expression O(NBS M (log NBS + log M)) + O((NBS M)^3 + G NBS M) is ambiguous: if the SVD is performed on the covariance R = Y Y^H in C^{NBS x NBS}, the cost should be O(NBS^3), not O((NBS M)^3). Please clarify the implementation and update the complexity comparison accordingly.","section":"V"},{"comment":"The lookup table description does not specify the grid sizes, the resolution of the angle/range/velocity grids, or the interpolation method used when matching measured spreads to table entries; please add these details for reproducibility.","section":"IV-A"},{"comment":"The NMSE definition averages over 1000 iterations, but reporting only NMSE can conceal bias; please also report per-parameter RMSE or bias to help interpret the -40 dB values.","section":"V"},{"comment":"There are several typos, including 'the the system behavior' in Section III-A; please proofread the manuscript carefully.","section":"III-A"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own prior work [9] for the angular-spread analysis, and the Doppler-spread derivation appears to be a direct analog; if the central Doppler-spread claim cannot be fixed, the analytical novelty is limited. The lack of independent validation for the lookup-table approach and the absence of public code make it difficult to assess the generality of the numerical results."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one before you trust the abstract. The central analytical claim — that a near-field target's transverse velocity creates a Doppler spread whose center gives radial velocity and width gives transverse velocity — is not derived correctly in the paper. Equation (13), which is supposed to be the Doppler-domain gain, is actually a spatial inner product between the Doppler steering vector b and a DFT codebook f, summed over the antenna index n. There's no temporal DFT over the symbol index m, so it cannot produce a Doppler spectrum. Worse, the radial velocity vr appears inside the phase as a constant that factors out of the squared magnitude, so the pattern is not shifted by vr at all. The text immediately after Eq. (13) claims the opposite.\n\nThat's the load-bearing flaw. If the Doppler-spread-to-vθ relationship is real, it is not established by this derivation. The authors might be seeing an effect in simulation that they haven't correctly characterized analytically.\n\nCredit where due: the idea itself is new relative to the cited near-field estimation works [6]-[8], which don't exploit Doppler spread across the array for transverse velocity. The two-stage DFT-coarse/1D-MUSIC-fine workflow is sensible, and the simulation results are internally consistent, reaching -40 dB NMSE at high SNR. The calibration error model is reasonable.\n\nOther soft spots, in decreasing order: the signal model in Eq. (7) has a dimension mismatch that makes it hard to verify the data generation; the 1D MUSIC refinement is applied without analyzing the multi-rank spatial signature induced by Doppler spread; the lookup tables are populated by simulating the same forward model used for the test data, so the coarse estimates are matched to a self-consistent dictionary; and the complexity expression includes an O((NBSM)^3) SVD term that overcounts the actual NBS×NBS covariance — conservative, but sloppy.\n\nWho should read it: researchers in near-field ISAC or automotive radar who want a low-complexity estimator for transverse velocity. They'll get the idea and maybe the simulation approach, but they'll have to redo the derivation themselves.\n\nRecommendation: worth sending to peer review because the observation is potentially valuable and the simulations could be reproduced. But I would not accept it without a corrected Doppler-spread analysis — either a proper temporal DFT derivation or an explicit statement that the relationship is empirical. For your own work, I wouldn't cite it until the math is fixed.","headline":"The paper's headline Doppler-spread claim is unsupported by its own Eq. (13), which is a spatial projection, not a Doppler spectrum; the idea is plausible and simulations are clean, but the core derivation must be redone.","tokens_in":8703,"tokens_out":4279,"would_cite":false,"duration_ms":51840,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper shows that in the near field of an ultra-massive antenna array, a target's angle, range, radial velocity, and transverse velocity can be read jointly from the width and center of its 2D-DFT angle-Doppler response, then refined…","keywords":["near-field sensing","UM-MIMO","2D-DFT","angular spread","Doppler spread","transverse velocity","MUSIC refinement","array calibration imperfections"],"falsifier":"Simulate one target whose transverse velocity is large enough that its Doppler spread covers several Fourier bins, run the proposed estimator, and compare the 1D-MUSIC-refined estimates against a full four-dimensional maximum-likelihood search; if the refined estimates become biased as the spread widens, the single-shape refinement assumption is wrong.","tokens_in":7642,"feed_emoji":"📡","tokens_out":12740,"duration_ms":125709,"temperature":0.7,"pith_summary":"The paper claims that in the radiative near field of an ultra-massive multiple-input multiple-output (UM-MIMO) array, a single moving target's angle, range, radial velocity, and transverse (sideways) velocity can be estimated jointly by examining how its response spreads in the two-dimensional discrete Fourier transform (2D-DFT) angle-Doppler domain. The angular spread is centered at the true angle and its width is set by range, while the Doppler spread is centered at the true radial velocity and its width grows with transverse velocity. Exploiting these relations, the paper develops a two-stage estimator: coarse 2D-DFT measurements of the two spreads, followed by one-dimensional multiple signal classification (MUSIC) refinement of each parameter in sequence. This replaces a four-dimensional search by one-dimensional scans. Simulations report a normalized mean squared error of about -40 dB for all four parameters at high signal-to-noise ratio, with lower complexity than gradient-based maximum likelihood, and the method is valid only inside the effective beamfocusing Rayleigh distance where the spreads exist.","feed_headline":"Doppler spread reveals a target's sideways speed in near-field MIMO","feed_subtitle":"A 2D Fourier transform plus a refinement step reads angle, range, and both velocity components at -40 dB error.","key_machinery":"The load-bearing object is the 2D-DFT angle-Doppler response of the space-time steering matrix $\\mathbf{V}(\\theta,r,v_r,v_\\theta)=\\sqrt{\\xi_t}(\\mathbf{A}\\odot\\mathbf{B})$, where $\\mathbf{A}$ carries the near-field spherical-wave phases and $\\mathbf{B}$ carries element-wise Doppler shifts. The paper's central identity is the Fresnel-integral form of the angular gain, $G_{\\mathrm{ang}}\\approx\\left|\\left(C(\\gamma_1,\\gamma_2)+jS(\\gamma_1,\\gamma_2)\\right)/(2\\gamma_2)\\right|^2$, where $C$ and $S$ are Fresnel integrals, with $\\gamma_1=\\sqrt{r_F/(d\\cos^2\\theta_u)}(\\sin\\theta_n-\\sin\\theta_u)$ and $\\gamma_2=(N_{\\mathrm{BS}}/2)\\sqrt{d\\cos^2\\theta_u/r_F}$; it shows the response is centered on the true angle and its width shrinks with range. The analogous Doppler-domain expression shows a linear phase slope whose width is set by $v_\\theta\\cos\\theta_u/(r_F f_r)$ and whose center is set by $v_r$. Offline lookup tables map measured 3 dB spread widths to range and transverse velocity, and 1D MUSIC on the sample covariance matrix then refines each parameter separately.","core_discovery":"The central discovery is that the two-dimensional DFT of the near-field space-time snapshot decouples location from motion: the angular spread width maps to range, the Doppler spread width maps to transverse velocity, and the centers of the two spreads give angle and radial velocity. Because the spherical wavefront makes the per-antenna Doppler shift vary linearly across the array, transverse motion appears not as a single Doppler tone but as a spread whose phase slope is proportional to $v_\\theta\\cos\\theta_u / r_F$. The paper derives this from a Fresnel approximation of the array response and shows that a coarse 2D-DFT stage plus per-parameter 1D MUSIC refinement reaches an NMSE of $-40$ dB under calibration imperfections while avoiding the exponential four-dimensional search.","pith_inferences":["Beyond the paper, the same width-to-range and width-to-transverse-velocity mappings suggest that a target crossing boresight produces the widest Doppler support, so sensitivity to transverse velocity is highest exactly where conventional radial-Doppler sensing is blind; the paper does not analyze this angular dependence of the resolution limit.","A natural untested extension is multi-target operation: when two targets' spreads overlap in the angle-Doppler plane, the lookup-table correlation step would need to be replaced by a sparse-recovery or iterative cancellation routine, since the paper's single-target model would otherwise merge the two spreads.","The offline lookup tables couple spread width to range and transverse velocity on a discrete grid; in practice the achievable resolution is set by how finely the 3 dB spread changes with each parameter, so a Cramér-Rao-style bound on spread-width estimation would predict where the coarse stage forces the MUSIC stage to work hardest."],"forward_implications":["A monostatic UM-MIMO base station can estimate transverse velocity from a single coherent processing interval, something far-field planar-wavefront radars cannot do without multistatic geometry.","The two-stage estimator replaces a four-dimensional grid search by one-dimensional scans, with complexity about $O(N_{\\mathrm{BS}}M(\\log N_{\\mathrm{BS}}+\\log M)) + O((N_{\\mathrm{BS}}M)^3 + G N_{\\mathrm{BS}}M)$, avoiding the $O(TG^4N_{\\mathrm{BS}}M)$ cost of the gradient-based search.","The angular and Doppler spreads shrink as range grows, so the method's validity is bounded by the effective beamfocusing Rayleigh distance; beyond it the near-field spreads disappear and the approach becomes inapplicable.","The coarse DFT stage already gives usable angle, radial-velocity, range, and transverse-velocity estimates, so the 1D MUSIC refinement only needs to search a small neighborhood around them."],"supporting_citations":[{"why":"Supplies the prior result that a near-field target's DFT angular response is centered on the true angle with spread width set by range, including the effective beamfocusing Rayleigh distance boundary.","marker":"[9]"},{"why":"Provides the maximum-likelihood velocity estimator used as the accuracy benchmark and the starting point for the paper's ML problem formulation.","marker":"[6]"},{"why":"Provides the polar-codebook estimator used as the high-accuracy benchmark that the DFT+MUSIC results are compared against.","marker":"[10]"}],"fun_headline_variants":["2D-DFT decouples angle, range, and velocity in near-field MIMO","Doppler spread width measures sideways motion in near-field","Near-field MIMO: Fourier spreads reveal range and speed","Spectral spreads unlock joint angle, range, and velocity","Low-complexity near-field MIMO estimation via Fourier bins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The refinement stage assumes that after the coarse Fourier step, each parameter can be recovered by matching a single idealized response shape to the measured signal, even though the moving target's Doppler spread makes the measured signature a mix of many shapes; the paper does not say how large the error from that mix can be.","fun_headline_variants_meta":{"raw":{"variants":["2D-DFT decouples angle, range, and velocity in near-field MIMO","Doppler spread width measures sideways motion in near-field","Near-field MIMO: Fourier spreads reveal range and speed","Spectral spreads unlock joint angle, range, and velocity","Low-complexity near-field MIMO estimation via Fourier bins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3130,"prompt_tokens":938,"completion_tokens":2192,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":554,"tokens_out":2192,"duration_ms":18116,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:24:06.262214+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate one target whose transverse velocity is large enough that its Doppler spread covers several Fourier bins, run the proposed estimator, and compare the 1D-MUSIC-refined estimates against a full four-dimensional maximum-likelihood search; if the refined estimates become biased as the spread widens, the single-shape refinement assumption is wrong.","supporting_citations":[{"cited_title":"Near-Field Motion Parameter Estimation: A Variational Bayesian Approach","cited_arxiv_id":"2502.14193","evidence_quote":"Supplies the prior result that a near-field target's DFT angular response is centered on the true angle with spread width set by range, including the effective beamfocusing Rayleigh distance boundary."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the maximum-likelihood velocity estimator used as the accuracy benchmark and the starting point for the paper's ML problem formulation."},{"cited_title":"Near-Field Beam Prediction Using Far-Field Codebooks in Ultra-Massive MIMO Systems","cited_arxiv_id":"2503.14317","evidence_quote":"Provides the polar-codebook estimator used as the high-accuracy benchmark that the DFT+MUSIC results are compared against."}],"review_version":1}