{"id":"b2f80252-4cbe-4cf0-946b-2b07508ca6b6","arxiv_id":"2507.21347","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A MUSIC algorithm and Cramér-Rao bounds are developed for DOA estimation with continuous aperture arrays, with simulations showing near-CRLB accuracy.","lead":"The authors bring MUSIC, a standard high-resolution direction-finding algorithm, to continuous aperture arrays (CAPAs), which sense the electromagnetic field across a whole surface rather than at discrete antennas. They also derive Cramér-Rao lower bounds for CAPA-based DOA estimation, reporting near-bound performance and an accuracy gain over conventional discrete arrays.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Z->0 limit in Eqs. (16)-(29) is inconsistent: the discrete voltages vanish as Z->0, so the sample covariance and the likelihood used for the CRLB are not the limits claimed; the CAPA accuracy gain in Remark 5 is therefore not established by the derivation.","rationale":"The reader's CONDITIONAL verdict is appropriate. Of the two premises the reader flags, the finite-rank reduction of Section III-B is the less dangerous: it is explicitly approximate, and the reported convergence in Fig. 3 gives empirical support in the tested regime. The Z->0 limit is more load-bearing because it underwrites both the covariance used by Algorithm 1 and the CRLB that supports the CAPA-vs-SPDA gain. As written, Eq. (16) sets x_n = Z E and noise variance Z*sigma^2, so the N-dimensional signal vectors in Eqs. (17)-(20) have norm tending to zero; Eq. (21) then drops two powers of Z inconsistently, and Eq. (29) replaces the vanishing Riemann sum by the finite integral. The likelihood and FIM inherit the same problem through sigma_nu^2 -> 0. The final CRLB formulas are finite only if sigma_nu^2 is silently identified with sigma^2, which is not the model of Eq. (16). If the limit is taken carefully, e.g., x_n = sqrt(Z) E with noise variance Z*sigma^2, the continuous covariance and likelihood are recovered, and the CRLB formulas (89)-(90) may be correct with sigma_nu^2 replaced by sigma^2. But that correction is not in the paper. The numerical results do not resolve this, because the algorithm computes K by quadrature on E directly and never exercises the Z-scaling. Thus the paper's strongest theoretical assertion is not currently supported. I would keep the CONDITIONAL verdict: the algorithmic core appears to work in simulation and the flaw is a fixable normalization/limiting issue, but the authors must redo the limit and re-derive the CRLB comparison.","tokens_in":19379,"tokens_out":9336,"duration_ms":117153,"concrete_test":"Re-derive the CRLB directly from the continuous-field model (10)-(12) without the Z-discretization: J_{ij} = (2*T*R_s/sigma^2) * Re{ int_S (partial a/partial theta_i)^* (partial a/partial theta_j) dr }, and compare this FIM with Eqs. (47)-(58). If the corrected FIM agrees with the paper only under the substitution sigma_nu^2 = sigma^2, then the discrete-to-continuous limit is inconsistent. Additionally, for a simple function f(r)=1 evaluate the Riemann sum sum_{n=1}^N Z^2 f(r_n) as Z->0; it tends to 0, directly contradicting Eq. (29).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is the Z->0 discretization limit in Section II. In Eq. (16), each element output is x_n(t)=|S_n|(a_n^T s(t)+n(r_n,t))=Z E(r_n,t), with noise variance Z*sigma^2. Therefore x(t)=Z[A s+n] and the covariance in Eq. (21) should be Z^2 A R_s A^H + Z*sigma^2 I_N, not A R_s A^H + sigma^2 I_N. Consequently Eq. (29) is dimensionally wrong: (1/T) * sum_n Z^2 E^*(r_n,i)E(r_n,j) tends to 0 as Z->0, whereas the right-hand side is the finite integral (1/T) * int_S E^*(r,i)E(r,j) dr. The K matrix used in Algorithm 1 is the correct finite object only if the normalization x_n = sqrt(Z) E is adopted, which is never stated. The same inconsistency enters the likelihood: Eq. (44) writes L = -(1/sigma_nu^2) * int |E-a^H s|^2 dr + C, and Eq. (46) has FIM proportional to 1/sigma_nu^2 = 1/(Z*sigma^2). Under the stated model, the CRLB in Eqs. (89)-(90) would tend to zero as Z->0, i.e., infinite precision from a finite aperture. The formulas are finite only if one renormalizes sigma_nu^2 -> sigma^2, a substitution that contradicts Eq. (16). Because Remark 5's CAPA-vs-SPDA factor lambda^2/(8*l_r^2) is computed from these CRLBs, the central theoretical claim of CAPA superiority is not supported by the derivation as written. This is an internal inconsistency, not a matter of consensus.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies direction-of-arrival (DOA) estimation with continuous aperture arrays (CAPAs). It proposes CAPA-MUSIC, which avoids eigendecomposition of an infinite-dimensional covariance operator by forming a T-by-T matrix K from field snapshots, reconstructing noise-subspace eigenvectors through Gauss-Legendre quadrature and a pseudo-inverse, and then searching a MUSIC spectrum. It derives Cramér-Rao lower bounds (CRLBs) for DOA estimation in both known-snapshot and unknown-snapshot cases, and claims a theoretical accuracy advantage of CAPA over conventional spatially discrete arrays (SPDAs), quantified in Remark 5 as a reduction by roughly lambda^2/(8 l_r^2). Numerical experiments show the proposed algorithm's MSE closely tracking the derived CRLB for single- and multi-target scenarios.","tokens_in":19735,"tokens_out":14121,"duration_ms":159195,"significance":"If the central claims are correct, the paper would provide a practical, low-complexity MUSIC algorithm for continuous apertures and a useful CRLB benchmark, together with a theoretical argument that CAPAs offer a substantial accuracy gain over SPDAs. The algorithm is clearly specified with a complexity analysis, and the numerical validation is extensive. However, the theoretical comparison and the CRLB formulas rest on a limiting discretization that is internally inconsistent. The claimed CAPA-over-SPDA gain is not established by the derivation as written. With a repaired normalization of the continuous-limit model, the contribution could be significant for the CAPA sensing literature.","major_comments":[{"comment":"The limiting discretization is internally inconsistent. From Eq. (16), x_n(t)=Z(a^T(r_n)s(t)+n(r_n,t)) with noise variance sigma_nu^2=Z sigma^2, so Eq. (20) is a vector whose entries scale as Z; taken literally the limit is the zero vector, and the covariance in Eq. (21) should be Z^2 A R_s A^H + Z sigma^2 I_N, not A R_s A^H + sigma^2 I_N. Consequently, in Eq. (29), Z^2 sum_n E^*(r_n,i)E(r_n,j) tends to 0 as Z->0, not to the surface integral; the displayed equality is dimensionally wrong. The same scaling issue affects the eigenvector reconstruction in Eqs. (34)-(38), where the components of u_i acquire a Z-dependent normalization. The paper needs to define the limiting continuous observation model explicitly (for example, by normalizing voltages per unit area or by working directly with the integral covariance operator) and redo Eqs. (21)-(38) under that model.","section":"§II-B and §III-B, Eqs. (16), (20), (21), (29)"},{"comment":"The likelihood limit in Eq. (44) has the same normalization error. With x(r_n,t)=Z E(r_n,t) and sigma_nu^2=Z sigma^2, the exponent in Eq. (43) equals -(1/(Z sigma^2)) sum_t sum_n Z^2 |E-a^H s|^2 = -(Z/sigma^2) sum_t sum_n |E-a^H s|^2, whose limit is -(1/sigma^2) sum_t int_S |E-a^H s|^2 dr, not -(1/sigma_nu^2) sum_t int_S |E-a^H s|^2 dr. Thus the FIM in Eq. (46) should be proportional to 1/sigma^2 rather than 1/sigma_nu^2 = 1/(Z sigma^2). As written, sigma_nu^2->0 makes the CRLBs in Eqs. (89)-(90) tend to zero, i.e., infinite precision from a finite aperture, which is unphysical. The CRLB derivation must be redone with a consistent noise normalization.","section":"§IV-A, Eqs. (43)-(46)"},{"comment":"The claimed CAPA-over-SPDA gain factor lambda^2/(8 l_r^2) is not obtained from the displayed formulas under any consistent assignment of noise variances. If one evaluates the elevation CRLB ratio from Eqs. (90) and (98) with P=2 L_x/lambda, Q=2 L_y/lambda, and a common noise variance, the ratio is approximately 2 l_r^2/lambda^2 (i.e., CAPA improves by lambda^2/(2 l_r^2)), not lambda^2/(8 l_r^2). If instead the infinitesimal-element variance sigma_nu^2=Z sigma^2 is retained, the CAPA CRLB vanishes with Z and the comparison is meaningless. A corrected comparison after the normalization fix is required before the theoretical superiority claim can be sustained.","section":"§IV-C, Remark 5"},{"comment":"The reduction of the infinite-dimensional eigendecomposition to the T-by-T matrix K is not fully justified. The exact algebraic relation between eigenvectors of X X^H and X^H X is u_i = (1/lambda_i) X e_i; the paper instead reconstructs u_i from e_i via the pseudo-inverse of the quadrature matrix E^H Omega in Eq. (38). It is not shown that this pseudo-inverse recovers the noise-subspace eigenvectors, nor that the Gauss-Legendre quadrature error remains controlled after the pseudo-inverse. Remark 2 concedes information loss but does not quantify it. Please provide a consistency argument or error bound for this step, or replace it by the exact algebraic reconstruction.","section":"§III-B, Eqs. (34)-(38) and Remark 2"}],"minor_comments":[{"comment":"The dimension of A(alpha, phi) is stated as C^{M x N}, but the construction [a(r_1), ..., a(r_N)]^T with a(r_n) in C^M gives an N x M matrix; the product A s(t) in Eq. (20) also requires A to be N x M.","section":"Eq. (18)"},{"comment":"There are notation slips: Eq. (34) integrates E^*(r, k) where the index should be t, and Eq. (35) states omega = diag{omega_1,...,omega_K} in R^{K^2 x K^2}; the intended object is likely a K x K diagonal matrix before the Kronecker product.","section":"Eqs. (34)-(35)"},{"comment":"In the M=3 target list, the third target location z3 = [200, 50, 15] is identical to the second target's location; this appears to be a typo and should be corrected for reproducibility.","section":"§V-C"},{"comment":"Proposition 1 proves that CRLB_k equals CRLB_u for a symmetric single-target aperture, but Remark 4 then states that uncertainty in the signal still affects estimation performance; the remark appears to contradict the proposition and should be clarified or removed.","section":"Proposition 1 and Remark 4"},{"comment":"The delta-correlated noise in Eq. (12) is not pointwise well-defined; the derivation should integrate the noise over each cell before taking the limit, rather than writing n(r_n, t) as a pointwise random variable.","section":"Eq. (12)"}],"recommendation":"major_revision","confidential_remarks":"The normalization issue in the Z->0 limit is the central technical problem. It appears to be a genuine internal inconsistency rather than a disagreement with existing literature, and it undermines the advertised CAPA-over-SPDA gain. The paper's algorithmic idea and numerical methodology are nonetheless promising; a corrected limiting model and rederived CRLBs could make the paper suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper has a genuine new piece: a MUSIC variant for continuous-aperture arrays, built on the standard X^H X versus XX^H eigendecomposition trick and Gauss-Legendre quadrature, with numerical experiments showing MSE tracking the CRLB. That part is worth reading. The complexity analysis is sensible, and the CAPA-MUSIC spectrum is clearly specified.\n\nThe problem is the continuous-limit theory. Equation (16) defines x_n(t) = Z E(r_n,t) with noise variance Z sigma^2, so the covariance of x is Z^2 A R_s A^H + Z sigma^2 I. Equation (21) instead writes A R_s A^H + sigma^2 I, dropping the Z factors. Equation (29) then claims Z^2 sum_n E*E tends to the integral, but for a Riemann sum with N proportional to 1/Z, Z^2 sum_n is about Z times the integral, which goes to 0. So the K matrix used in Algorithm 1 is not the limit stated; it is the right finite object only under a different normalization, which is not given. The same inconsistency enters the likelihood in (44) and the FIM in (46): they use sigma_nu^2 = Z sigma^2 without the corresponding Z^2 in the signal, so the CRLBs in (89)–(90) are not the limits of the stated model. The CAPA-vs-SPDA factor lambda^2/(8 l_r^2) in Remark 5 is computed from those CRLBs, so the central theoretical claim is unsupported. This is an internal contradiction, not a matter of interpretation.\n\nThere is also a smaller contradiction: Proposition 1 says known and unknown snapshot CRLBs are identical for a single symmetric target, which follows from the FIM block-diagonal structure. Remark 4 then says uncertainty in the signal still affects performance “via other statistical dependencies,” which directly contradicts the proposition. As written, that remark is wrong.\n\nWhat is good: the algorithm itself may work; the simulations show near-CRLB behavior, and the Gauss-Legendre approximation is sensible. But the paper does not ship code or data, so the empirical claim is not independently reproducible from the text.\n\nWho is this for? People working in CAPA sensing / holographic MIMO. They will want to cite it if it survives revision. My recommendation: send it to review, but it needs major revision. The authors must fix the Z->0 normalization, redo the CRLB derivation, and qualify or remove the CAPA superiority factor. If they do, it could be a solid contribution. As is, I would not accept.","headline":"The CAPA-MUSIC algorithm is a plausible practical method, but the paper's continuous-limit derivation is internally inconsistent, so the CRLB and the CAPA-vs-SPDA gain claim are not established as written.","tokens_in":20360,"tokens_out":5695,"would_cite":false,"duration_ms":66165,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","65D32"],"pacs":[],"model":"deepseek-v4-flash","headline":"CAPA-MUSIC estimates DOAs near the Cramér-Rao bound and proves continuous apertures beat discrete arrays.","keywords":["continuous aperture array","CAPA-MUSIC","DOA estimation","Cramér-Rao lower bound","MUSIC algorithm","Gauss-Legendre quadrature","eigendecomposition","spatial degrees of freedom"],"falsifier":"Generate a simulated CAPA with a large but finite number of very small antenna elements, run CAPA-MUSIC on a single target over many independent trials at a fixed SNR, and compare the empirical azimuth and elevation MSE with the claimed $CRLB_u$ from Eqs. (75)-(76); if the MSE remains clearly above the bound after increasing the quadrature order $K$ and the snapshot count $T$, the finite-rank noise-subspace reconstruction is not faithful and the central claim fails.","tokens_in":19093,"feed_emoji":"📡","tokens_out":7138,"duration_ms":72490,"temperature":0.7,"pith_summary":"This paper attempts to establish that direction-of-arrival (DOA) estimation can be performed directly on a continuous aperture array (CAPA), a receiving surface that measures the field at every point rather than at discrete antenna elements, and that such arrays are fundamentally more accurate than conventional discrete arrays. To make this practical, it proposes CAPA-MUSIC, a MUSIC-type algorithm that works with infinite-dimensional continuous measurements by reducing the covariance operator to a finite matrix and approximating integrals with Gauss-Legendre quadrature. It derives Cramér-Rao lower bounds for DOA estimation with CAPAs, both when the target snapshot signals are known and when they are unknown, and proves that CAPAs lower the bound by a factor on the order of $\\lambda^2/(8 l_r^2)$ relative to a spatially discrete array of short dipoles. For a typical dipole length this is roughly two orders of magnitude, so the paper's message is that continuous apertures offer a concrete accuracy advantage along with a workable estimator. If the central claim holds, it supplies both a theoretical limit and a practical algorithm for the sensing side of 6G continuous-aperture systems.","feed_headline":"MUSIC on continuous apertures approaches the Cramér-Rao DOA bound","feed_subtitle":"CAPA-MUSIC gives near-optimal azimuth and elevation estimates and proves a large accuracy gain over discrete arrays.","key_machinery":"The load-bearing device is the continuous-to-discrete transformation in Section III-B: the eigendecomposition of the infinite-dimensional covariance operator is reduced to that of the finite $T \\times T$ matrix $K = X^H X/T$, and the eigenvectors of the noise subspace are reconstructed by the Moore-Penrose pseudo-inverse in Eq. (38). The continuous inner products in $K$ and in the MUSIC spectrum are then evaluated by Gauss-Legendre quadrature (Eqs. (30)-(33)), giving a finite, tunable approximation and a complexity of $O(T^3 + K^4 T + N_S K^4)$. The CRLB analysis rests on a second device, the limit model $Z \\to 0$ in Eqs. (16)-(21), in which the aperture is split into infinitesimal elements; that limit produces a finite non-vanishing covariance from which the Fisher information matrix is computed.","core_discovery":"The central claim is that CAPA-MUSIC (Eq. (40), Algorithm 1) estimates azimuth and elevation of multiple targets with mean squared error close to the CAPA Cramér-Rao lower bound, and that a CAPA reduces that bound by a factor on the order of $\\lambda^2/(8 l_r^2)$ compared to a conventional spatially discrete array (Remark 5). The paper derives closed-form CRLBs for both known and unknown snapshot signals, proves that unknown snapshots never improve the bound, and shows that for a symmetric aperture with a single target the two bounds coincide (Proposition 1). It attributes the accuracy gain to the continuous aperture's large spatial degrees of freedom and demonstrates numerically that the algorithm's MSE approaches the bound as the quadrature order $K$ grows.","pith_inferences":["For real hardware, the continuous limit is a mathematical idealization; at finite element densities the gain over discrete arrays will be smaller, and a quantitative model of discretization loss would sharpen the practical prediction.","The same continuous-discrete transformation could carry other subspace methods (e.g., ESPRIT or tensor-based estimators) over to CAPAs, and the CRLB framework could extend to near-field DOA or joint position-and-attitude estimation.","The paper's own Remark 2 suggests a testable design rule: the number of snapshots $T$ needed for a faithful noise subspace may scale with the number of targets $M$, and a systematic study of that scaling would tell practitioners how large $T$ must be.","Because the CRLB expressions depend explicitly on aperture geometry, optimizing the aperture shape (not just its size) could yield further accuracy gains beyond the square CAPAs tested here."],"forward_implications":["CAPA-MUSIC achieves estimation performance close to the Cramér-Rao bound for multiple targets, with computational complexity $O(T^3 + K^4 T + N_S K^4)$.","CAPAs lower the CRLB by a factor on the order of $\\lambda^2/(8 l_r^2)$ versus half-wavelength-spaced short-dipole arrays, implying an accuracy gain of roughly $8\\pi^2$ for $l_r = \\lambda/(8\\pi)$.","The CRLB with known snapshots is always no larger than the CRLB with unknown snapshots, and the two are equal for a symmetric aperture with a single target.","Enlarging the CAPA dimensions $L_x$ and $L_y$ sharpens the MUSIC peaks and reduces the CRLB through terms like $L_x^3 L_y$ and $L_x L_y^3$.","With multiple targets, the algorithm's MSE approaches the CAPA CRLB and worsens gracefully as the number of closely spaced targets grows."],"supporting_citations":[{"why":"Introduces the CAPA model and its spatial degrees-of-freedom advantage that motivates the paper.","marker":"[14]"},{"why":"Provides the Fisher-information and CRLB methodology for holographic positioning that the paper extends to DOA estimation.","marker":"[22]"},{"why":"Extends CAPA CRLB analysis to near-field sensing, the framework the authors build on for known and unknown snapshot cases.","marker":"[23]"},{"why":"Sets out the MUSIC subspace-orthogonality principle that CAPA-MUSIC adapts to continuous measurements.","marker":"[26]"},{"why":"Supplies the Gauss-Legendre quadrature rule used to approximate all continuous integrals in the algorithm and CRLB.","marker":"[30]"},{"why":"Provides the partitioned-matrix inverse formula used to derive the unknown-snapshot CRLB.","marker":"[31]"},{"why":"Gives the dominated convergence theorem used to justify exchanging differentiation and integration in Lemma 1.","marker":"[32]"}],"fun_headline_variants":["CAPA-MUSIC nears Cramér-Rao bound for DOA","Continuous-aperture MUSIC achieves near-optimal DOA","MUSIC on continuous apertures nears optimal DOA","Continuous aperture MUSIC outperforms discrete arrays","MUSIC for continuous apertures approaches Cramér-Rao"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the finite snapshot matrix $K = X^H X/T$ captures the continuous noise subspace faithfully enough that the pseudo-inverse reconstruction of eigenvectors yields a valid MUSIC spectrum; the paper acknowledges information loss in Remark 2 and supports the assumption only numerically.","fun_headline_variants_meta":{"raw":{"variants":["CAPA-MUSIC nears Cramér-Rao bound for DOA","Continuous-aperture MUSIC achieves near-optimal DOA","MUSIC on continuous apertures nears optimal DOA","Continuous aperture MUSIC outperforms discrete arrays","MUSIC for continuous apertures approaches Cramér-Rao"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000834,"raw_usage":{"total_tokens":3624,"prompt_tokens":911,"completion_tokens":2713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":2626}},"tokens_in":527,"tokens_out":2713,"duration_ms":23144,"temperature":1.0,"reasoning_tokens":2626,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:54:46.920588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate a simulated CAPA with a large but finite number of very small antenna elements, run CAPA-MUSIC on a single target over many independent trials at a fixed SNR, and compare the empirical azimuth and elevation MSE with the claimed $CRLB_u$ from Eqs. (75)-(76); if the MSE remains clearly above the bound after increasing the quadrature order $K$ and the snapshot count $T$, the finite-rank noise-subspace reconstruction is not faithful and the central claim fails.","supporting_citations":[{"cited_title":"CAPA: Continuous-Aperture Arrays for Revolutionizing 6G Wireless Communications","cited_arxiv_id":"2412.00894","evidence_quote":"Introduces the CAPA model and its spatial degrees-of-freedom advantage that motivates the paper."},{"cited_title":"Cram´er-Rao bounds for holographic positioning,","cited_arxiv_id":null,"evidence_quote":"Provides the Fisher-information and CRLB methodology for holographic positioning that the paper extends to DOA estimation."},{"cited_title":"Cram ´er-Rao bound optimization for near-field sensing with continuous-aperture arrays,","cited_arxiv_id":null,"evidence_quote":"Extends CAPA CRLB analysis to near-field sensing, the framework the authors build on for known and unknown snapshot cases."},{"cited_title":"Gridless DOA estimation and root- MUSIC for non-uniform linear arrays,","cited_arxiv_id":null,"evidence_quote":"Sets out the MUSIC subspace-orthogonality principle that CAPA-MUSIC adapts to continuous measurements."},{"cited_title":"Ralston and P","cited_arxiv_id":null,"evidence_quote":"Supplies the Gauss-Legendre quadrature rule used to approximate all continuous integrals in the algorithm and CRLB."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the partitioned-matrix inverse formula used to derive the unknown-snapshot CRLB."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the dominated convergence theorem used to justify exchanging differentiation and integration in Lemma 1."}],"review_version":1}