{"id":"8c58d494-76f0-480a-a81d-18a209a85544","arxiv_id":"2411.09495","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"The paper proves that lifted atomic norm minimization can recover target parameters and communication data in a MIMO ISAC receiver from measurements whose count is sublinear in the degrees of freedom.","lead":"This paper proposes a receiver for integrated sensing and communication (ISAC) systems that simultaneously locates radar targets and decodes communication symbols without known pilot signals, using lifted atomic norm minimization. A mathematician or engineer might read it because it extends off-the-grid super-resolution theory to the joint radar-communication problem, with a sample complexity guarantee and robustness to noise and jamming.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1's proof relies on L^4 independent random coding vectors d_n indexed by 4-tuples, but the stated model supplies only \\bar L independent columns d_l; the Bernstein concentration in Lemma 3 therefore does not apply to the stated problem.","rationale":"The central claim is exact recovery via the lifted atomic norm program, and the only proof of that claim is the dual-certificate argument in Appendix X. That argument depends wholly on concentration of the random matrix Gamma. The system model defines D as an \\bar L x T coding matrix, so the only randomness supplied by Assumption 1 is \\bar L independent T-dimensional columns. The proof, however, builds Gamma from a sum over a four-dimensional index set J with a vector d_n at every 4-tuple n, and then applies matrix Bernstein as though all of those d_n were independent. If those d_n are actually d_{n1} reused across n2,n3,n4, the summands are dependent and the variance calculation in Appendix B is wrong; if D is silently redefined to have |J| columns, then D is no longer the matrix used in the observation model (5)-(9), so Theorem 1 does not prove recovery for the stated system. Either way, the sampled complexity guarantee does not follow. This is more specific than the reader's concern about the practical realism of random D: it says that even accepting all random-model assumptions, the proof's concentration step is internally inconsistent with the model. At the same time, the reader's other observations remain relevant: Lemma 2 is unproved, Appendix F contains an empty citation, and the complexity derivation has an algebraic slip. These reinforce a conditional verdict rather than a rejection, because the core idea may be repairable by rewriting the proof with the correct independence structure or by reformulating the model so that D genuinely has L^4 independent columns; but the paper as written does not establish Theorem 1.","tokens_in":21257,"tokens_out":13177,"duration_ms":135290,"concrete_test":"Re-derive Lemma 3 from the actual definitions, replacing the copied 4D kernel by the true tilde D_j in Eq. (5): write Gamma as a sum over the available independent columns d_l, with any repeated indices made explicit. Count the number of independent random columns actually used; if it is \\bar L = 2N+1 rather than L^4, compute Cov(S_n, S_{n'}) for two indices sharing n_1. If the cross-covariance is nonzero, the Bernstein bound in Appendix B undercounts sigma^2. Then run a numerical simulation for \\bar L = 64, T = 8, K = 3, N_r = 1 and compare ||Gamma - Phi \\otimes I_T|| with the Lemma 3 prediction. If the spectral norm does not decay according to the claimed L^4 concentration, the sample complexity in Theorem 1 is not supported by the proof.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The proof of Theorem 1 breaks at the point where the random coding matrix enters the concentration argument. Assumption 1 gives D in C^{\\bar L x T}, so the independent random vectors are d_l for l=-N,...,N — only \\bar L of them. Yet in Eq. (34) and Eq. (44), the construction introduces K(tau) and Gamma as averages over n in J with weights s_{n1}s_{n2}s_{n3}s_{n4}, treating d_n for each four-dimensional index n as an independent draw from F and forming (v_n \\otimes d_n)(v_n \\otimes d_n)^H. There are not L^4 independent d_n in the model: the same d_l is reused across the other three indices. The summands S_n in Eq. (44) are therefore dependent, the matrix Bernstein step in Appendix B (Eqs. 69-72) does not apply, and the bound L^4 >= C mu K T log(10KT/delta) in Lemma 3 and Theorem 1 does not follow. A related symptom is the dimension mismatch: K(tau) is claimed to lie in C^{L x L} although d_n d_n^H lies in C^{T x T}, and U is typed as C^{K x L^2} in Eq. (7) while U = sum |alpha| h a^H has row dimension T. The proof appears to be a four-dimensional copy of [31] without adapting the random coding operator to the actual ISAC observation model.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a bistatic MIMO ISAC receiver in which the transmit probing signals are unknown but lie in a known low-dimensional subspace, x_k = D h_k, and the goal is to recover both the communication vectors h_k and the radar target parameters (delay, Doppler, AoA, AoD) from the echoes. The received signal is recast as y = X(U) with U = sum_k |alpha_k| h_k a(tau_k)^H, and a lifted atomic norm minimization (LANM) problem (12) is proposed. The main theoretical claim, Theorem 1, is that under random-coding isotropy/incoherence, random h_k, and a minimum separation condition, exact recovery holds with high probability once L^4 >= C mu K T log(10KT/delta). The paper also gives a semidefinite relaxation of the dual, a robust formulation for AWGN plus jamming, a complexity analysis, and simulations against pilot-aided ANM and ell_1 methods.","tokens_in":21574,"tokens_out":7513,"duration_ms":69100,"significance":"If Theorem 1 were valid, the paper would make a substantial contribution: it would provide the first off-the-grid recovery guarantee for a four-parameter ISAC scenario without a pilot link, with sample complexity proportional to the degrees of freedom, and with a claimed computational cost comparable to pilot-aided ANM. The problem formulation and the proposed SDR are original and potentially useful, and the numerical experiments show encouraging behavior. However, the paper does not provide machine-checked proofs or reproducible code, and the central proof has serious gaps that prevent the main result from being accepted as established.","major_comments":[{"comment":"Assumption 1 supplies only \\bar L = 2N+1 independent coding vectors d_l (l = -N, ..., N). In Eq. (34) and Eq. (44), however, the proof averages over four-dimensional multi-indices n in J and treats d_n d_n^H as independent summands, e.g. in Gamma = (1/L^4) sum_{n in J} s_{n1}s_{n2}s_{n3}s_{n4} (v_n v_n^H) tensor (d_n d_n^H). The same d_l is reused whenever a coordinate of n equals l, so the summands S_n in Eq. (44) are dependent. The matrix Bernstein step in Eqs. (69)-(72) therefore does not apply, and the sample-complexity condition L^4 >= C mu K T log(10KT/delta) in Lemma 3 and Theorem 1 is not established for the stated model.","section":"Section X-A, Eqs. (34), (44), and Appendix B"},{"comment":"There are dimension mismatches that accompany the independence problem. K(tau) in Eq. (34) is declared to lie in C^{L x L}, while d_n d_n^H is T x T; U is typed as C^{K x L^2} in Eq. (7), although U = sum |alpha_k| h_k a(tau_k)^H has row dimension T when h_k is in C^T. The definition of a(tau) in Eq. (3) also produces a vector whose length is not reconciled with L = \\bar L N_r. These mismatches make it unclear what object the operator X in Eq. (7) acts on and prevent the proof from being checked.","section":"Eqs. (3), (7), (34)-(35)"},{"comment":"A load-bearing bound is left unsupported: Appendix F states 'we know ||\\bar E^{(i)}(tau)||_F <= C1 from []' with an empty citation, and this bound is needed to control I_2^{(i)}(tau) in Lemma 8. In addition, the proof of Lemma 2 is omitted (see the paragraph before Eq. (43)), and Lemma 9 is dismissed as 'based on [31, Lemma 8]' without translating that argument to the present four-dimensional ISAC operator. These gaps mean that the construction of the dual polynomial, which is the core of the proof of Theorem 1, is not fully proven.","section":"Appendix F and Lemma 9"},{"comment":"The complexity calculation is algebraically incorrect. With E approximately ((2N+1)2N_t N_r)^2 ~ 16 N^2 N_t^2 N_r^2, and F+1 of the same order, the interior-point bound O((E+F)^{1.5} E^2) scales as O(N^7 N_t^7 N_r^7) up to constants, not as O((4N^2 N_t N_r)^5). The claimed equivalence of the proposed estimator to pilot-aided ANM is therefore not demonstrated by Eq. (29).","section":"Section VI, Eq. (29)"}],"minor_comments":[{"comment":"The definition of the index set J and its relation to L is ambiguous: J is said to consider all elements of s, r, l, k with j = 1, ..., L, but the ranges of l and k and the mapping to j are not specified clearly.","section":"Section II, after Eq. (3)"},{"comment":"'identical matrix' should read 'identity matrix', and the coherence condition max |d^{(i)}|^2 <= mu should specify which entry index i is being ranged over.","section":"Assumption 1, Eq. (14)"},{"comment":"The caption says 'Figs. 4(a), 4(a), and 4(b)' but should refer to 4(a), 4(b), and 4(c).","section":"Section VII, Fig. 4 caption"},{"comment":"The caption mentions the (phi, rho) domain while the text and the panel labels use (tau, v); these should be reconciled.","section":"Fig. 2 caption and Section IV"},{"comment":"There is an extra closing parenthesis after '[\\hat z]_j)', which makes the expression hard to parse.","section":"Eq. (27)"}],"recommendation":"reject","confidential_remarks":"The proof of Theorem 1 is not merely incomplete; it appears to apply a four-dimensional independence structure that the stated model does not possess, since only \\bar L random coding vectors d_l exist in Assumption 1. The dimension mismatches in Eqs. (3), (7), and (34) suggest that the proof was adapted from [31] without adjusting the random coding operator to the actual ISAC observation model. Combined with the exponent error in the complexity analysis, these are load-bearing issues that cannot be fixed by local edits within the manuscript's scope. The problem formulation and simulations are interesting, but the central theoretical guarantee is not currently supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing you should know: the algorithmic idea is sensible and the simulations look okay, but the proof of Theorem 1 does not work as written. The concentration argument in Appendix B (and the construction of Gamma in Eq. 44) treats the random vectors d_n as independent draws indexed by a four-dimensional set J. Your model only gives you \\bar L independent columns d_l (l=-N,...,N), one per time sample. The same d_l is reused across all four indices, so the summands in Eq. (44) are dependent. Matrix Bernstein does not apply, and the sample-complexity bound L^4 >= C mu K T log(...) doesn't follow. There are also dimension mismatches: Eq. (34) claims K(tau) is in C^{L x L} while d_n d_n^H is T x T, and U is written as C^{K x L^2} in Eq. (7) but is actually T x L. The proof reads like a direct 4D copy of Chi's 2016 paper without adapting the random coding operator to the actual ISAC observation model.\n\nWhat is genuinely new: the paper extends lifted atomic norm minimization to joint 4D radar parameter estimation and communication decoding in MIMO ISAC, which I don't think appears elsewhere directly. The jamming formulation is a reasonable addition. The numerical results show performance comparable to pilot-aided ANM, so the approach is at least empirically promising.\n\nOther soft spots worth mentioning: Lemma 2's proof is omitted ('eager readers can follow'), Appendix F cites '[]' for a key bound, and the computational complexity derivation in Eq. (29) has an algebraic error: with E roughly A^2, the complexity should be O(A^7), not O(A^5). These are fixable with care.\n\nBottom line: the core theoretical guarantee is not established. If the authors can fix the concentration argument by using the actual dependency structure of the coding matrix, the paper could be a reasonable contribution. As it stands, I would not trust Theorem 1. The topic is timely, so I'd still send it to a referee, but with the expectation that the proof needs substantial rework.","headline":"Plausible LANM extension to 4D ISAC, but the main theorem's proof has a load-bearing independence flaw that the authors need to fix.","tokens_in":22120,"tokens_out":5324,"would_cite":false,"duration_ms":46974,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["94A12","90C22","94A20"],"pacs":[],"model":"deepseek-v4-flash","headline":"A single lifted convex program recovers radar target parameters and communication symbols together from received echoes, with no pilot link.","keywords":["ISAC","lifted atomic norm minimization","off-the-grid super-resolution","semidefinite relaxation","MIMO radar","joint radar and communication estimation","dual certificate","blind deconvolution"],"falsifier":"Run the proposed SDR (20) with two targets satisfying the separation condition but with $D$ chosen as a deterministic low-coherence matrix (for example, a partial Fourier or DCT matrix) at a sample size around the theorem's bound; if recovery fails consistently while a random Gaussian $D$ of the same size succeeds, the randomness assumption is load-bearing and the guarantee does not extend to deterministic coding matrices. Directly, compute $\\|\\Gamma - \\mathbb{E}\\Gamma\\|$ for such a $D$: Lemma 3 predicts concentration below $\\epsilon$ for $L^4 \\geq C\\mu K T \\log(10KT/\\delta)$, so a persistent large deviation would falsify the key concentration step.","tokens_in":21072,"feed_emoji":"📡","tokens_out":10268,"duration_ms":84213,"temperature":0.7,"pith_summary":"The paper tries to establish that an integrated sensing and communication (ISAC) receiver can simultaneously locate radar targets and decode the data-carrying transmit signal, even though neither the channel nor the transmit waveform is known. The proposed estimator, lifted atomic norm minimization (LANM), is a convex program whose exact-recovery guarantee is Theorem 1: with probability at least $1-\\delta$, the lifted matrix $U = \\sum_k |\\alpha_k| h_k a(\\tau_k)^H$ is recovered exactly from the linear measurements $y = \\mathcal{X}(U)$ as soon as $L^4 \\geq C\\mu K T \\log(10KT/\\delta)$, under a minimum-separation condition on the target parameters. If this is right, a single off-the-grid estimator can replace pilot-aided channel estimation and radar-only super-resolution at the same order of computational cost. Simulations support the theorem and show performance comparable to pilot-aided atomic norm minimization, which knows the transmit signal.","feed_headline":"One convex program recovers radar and data at once","feed_subtitle":"Lifted atomic norm minimization needs samples proportional to degrees of freedom, matching pilot-aided ANM's cost.","key_machinery":"The load-bearing object is the lifted atomic norm and its dual. The atom set is $\\mathcal{A} = \\{v a(\\tau)^H : \\tau \\in [0,1)^4, \\|v\\|_2 = 1\\}$, and the lifted atomic norm of $U$ is the gauge of $\\mathrm{conv}(\\mathcal{A})$; minimizing it under the measurement constraint $y = \\mathcal{X}(U)$ turns the unknown-waveform problem into a low-rank matrix recovery problem. The proof machinery is a randomized vector-valued dual polynomial $Q(\\tau) = \\mathcal{X}^*(q) a(\\tau)$, built from the squared Fejér kernel and its derivatives, that interpolates the sign patterns at the true target parameters and is bounded away from 1 elsewhere; concentration of the random Gram matrix $\\Gamma$ around its expectation $\\Phi \\otimes I_L$ supplies the sample-complexity bound. Computationally, the infinite-dimensional dual constraint $\\|Q(\\tau)\\|_2 \\leq 1$ is implemented by a semidefinite relaxation using Toeplitz matrix constraints from trigonometric polynomial theory.","core_discovery":"On the paper's own terms, the central discovery is that the bilinear ISAC reception problem—unknown target parameters multiplied by unknown communication vectors—can be lifted into a linear inverse problem over a low-rank matrix $U = \\sum_{k=1}^K |\\alpha_k| h_k a(\\tau_k)^H$, where $a(\\tau_k)$ encodes the four continuous parameters (delay, Doppler, angle of departure, angle of arrival) and $h_k$ carries the communication data. The lifted atomic norm $\\|U\\|_{\\mathcal{A}}$ over the atom set $\\{v a(\\tau)^H : \\tau \\in [0,1)^4, \\|v\\|_2=1\\}$ promotes exactly this structured low-rankness, and its dual certificate $Q(\\tau) = \\mathcal{X}^*(q) a(\\tau)$ is constructed from a randomized Fejér kernel so that it peaks at the true target locations and stays below 1 elsewhere. Theorem 1 states that this dual certificate exists with high probability once the sample count satisfies $L^4 \\geq C\\mu K T \\log(10KT/\\delta)$, yielding exact recovery of the target parameters and the data-carrying vectors $h_k$.","pith_inferences":["The theorem's randomness assumptions are used only for concentration, and the paper suggests $D$ can be a known coding matrix with $h_k$ drawn from QAM constellations; a natural follow-up is to test whether deterministic low-coherence coding matrices satisfy the same guarantee, since the proof as written needs the isotropy and incoherence of random columns.","The bound's logarithmic dependence on the subspace dimension $T$ suggests communication rate can be increased almost for free once the radar parameters are resolvable; if the scaling is tight, the main cost of data rate is in the constant and coherence, not the number of symbols per target.","The separation condition takes a max over the four parameters, so targets close in one dimension but separated in another are not covered by the current analysis; a refined condition using a joint metric could enlarge the recoverable regime.","For the jamming scenario the paper only conjectures the logarithmic phase-transition scaling; verifying that conjecture with a rigorous sample-complexity bound for the two-atomic-norm problem would complete the robust-recovery story."],"forward_implications":["An ISAC receiver can drop pilot signaling and the direct transmitter-receiver link; the same received echoes yield target angles, delay, Doppler, and the QAM symbols in the probing signal.","Recovery is off the grid: target parameters are continuous, so there is no basis-mismatch error of the kind that degrades grid-based $\\ell_1$ methods.","The sample complexity scales as $L^4 \\gtrsim \\mu K T \\log(KT/\\delta)$, proportionate to the degrees of freedom of the ISAC scene rather than the ambient dimension.","With AWGN and jamming, the same framework separates the radar signal, the communication data, and the jammer's angle, using one extra atomic norm for jammers.","The overall complexity is the same order as pilot-aided ANM, so data decoding comes at negligible extra computational cost; the paper also points to ADMM as a faster implementation route."],"supporting_citations":[{"why":"Introduces the lifting-and-convex-optimization proof machinery for blind sparse spike deconvolution that the paper adapts to four-parameter MIMO ISAC.","marker":"[31]"},{"why":"Supplies the Fejér-kernel bounds and the invertibility lemma for the matrix $\\Phi$ that control the dual polynomial and its derivatives.","marker":"[38]"},{"why":"Establishes atomic norm minimization off the grid and the dual-certificate strategy, including the perturbation bound used in the paper's Lemma 4.","marker":"[23]"},{"why":"Provides the pilot-aided ANM baseline whose computational complexity the proposed estimator matches.","marker":"[26]"},{"why":"Provides the positive trigonometric polynomial theory used to convert the infinite-dimensional dual search into a semidefinite relaxation.","marker":"[34]"},{"why":"Shows super-resolution radar recovery of continuous delay-Doppler pairs, the setting this paper generalizes to unknown waveforms and four parameters.","marker":"[29]"},{"why":"Extends atomic norm recovery to continuous angle-delay-Doppler triplets in MIMO radar, a predecessor of the four-parameter problem.","marker":"[30]"},{"why":"Applies lifted atomic norm minimization to separate radar and communication signals in spectrum sharing, motivating the dual-blind ISAC formulation.","marker":"[4]"}],"fun_headline_variants":["One lifted atomic norm solves ISAC's blind recovery","Atomic norm lifting yields radar targets and data together","Off-grid ISAC: a single convex lift finds everything","Lifted atomic norm cracks bilinear ISAC in one convex step","Blind ISAC: single lift decodes data and locates targets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each transmit probing signal lies on a known random low-dimensional subspace, $x_k = D h_k$, with the coding matrix $D$ having statistically isotropic, incoherent columns; if $D$ is deterministic or the signal is not confined to such a subspace, the concentration arguments and the exact-recovery guarantee of Theorem 1 no longer follow.","fun_headline_variants_meta":{"raw":{"variants":["One lifted atomic norm solves ISAC's blind recovery","Atomic norm lifting yields radar targets and data together","Off-grid ISAC: a single convex lift finds everything","Lifted atomic norm cracks bilinear ISAC in one convex step","Blind ISAC: single lift decodes data and locates targets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000288,"raw_usage":{"total_tokens":1728,"prompt_tokens":1021,"completion_tokens":707,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":637,"completion_tokens_details":{"reasoning_tokens":624}},"tokens_in":637,"tokens_out":707,"duration_ms":7601,"temperature":1.0,"reasoning_tokens":624,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T20:35:06.822675+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the proposed SDR (20) with two targets satisfying the separation condition but with $D$ chosen as a deterministic low-coherence matrix (for example, a partial Fourier or DCT matrix) at a sample size around the theorem's bound; if recovery fails consistently while a random Gaussian $D$ of the same size succeeds, the randomness assumption is load-bearing and the guarantee does not extend to deterministic coding matrices. Directly, compute $\\|\\Gamma - \\mathbb{E}\\Gamma\\|$ for such a $D$: Lemma 3 predicts concentration below $\\epsilon$ for $L^4 \\geq C\\mu K T \\log(10KT/\\delta)$, so a persistent large deviation would falsify the key concentration step.","supporting_citations":[{"cited_title":"Towards a mathematical theory of super-resolution,","cited_arxiv_id":null,"evidence_quote":"Supplies the Fejér-kernel bounds and the invertibility lemma for the matrix $\\Phi$ that control the dual polynomial and its derivatives."}],"review_version":1}