{"id":"b1d5a64e-060c-421b-a65e-c869ed0ef95f","arxiv_id":"2606.30958","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Introduces nonconvex dual-TV regularizers based on transformed L1 for exponential-family tensor completion and proves error bounds of order O(n3 rt (max sk^2) log / n) that approach minimax rates up to O(max sk^2 / max(n1,n2)) gap.","lead":"This paper proposes a family of dual total-variation regularizers for tensor completion under exponential-family noise and derives recovery error upper bounds that nearly match minimax lower bounds in some regimes. A smart generalist might read it for better methods to impute missing entries in image, video, or count-based multi-way data with non-Gaussian noise.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Error bounds rest on gradient tensor satisfying simultaneous sparsity+low-rank structure exactly matched by the transformed-L1 DTV family; this is invoked but not shown to be necessary or sufficient for the stated rates.","rationale":"The reader's weakest_assumption directly identifies the modeling choice that the claimed rates depend on. Because the full text was not supplied to the first reader, the current UNVERDICTED status remains appropriate until the derivation steps that invoke this assumption are inspected.","tokens_in":1846,"tokens_out":372,"duration_ms":24191,"concrete_test":"Extract the precise statement of the main recovery theorem (likely Theorem 3.x or 4.x) and the definition of the DTV regularizer; re-derive the error bound while replacing the transformed-L1 penalty with a standard (convex) anisotropic TV penalty on the same gradient tensor; if the O(n3 rt (maxk sk^2) log / n) rate is lost, the specific nonconvex dual construction is load-bearing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract states that upper bounds of order O(n3 rt (maxk sk^2) log((n1+n2)n3)/n) are attained 'in certain cases' and that the gap to the minimax lower bound is O(maxk sk^2 / max(n1,n2)) up to logs. Both results are derived by assuming the gradient tensor admits a structure that the specific dual-TV regularizer family is designed to capture. If this modeling assumption fails to hold (or if the regularizer does not enforce it under the exponential-family likelihood), the error bounds do not follow. The paper provides no independent verification that the transformed-L1 DTV construction is the minimal or canonical choice that produces the claimed rates.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a family of nonconvex dual total-variation (DTV) regularizers based on the transformed L1 function for tensor completion under exponential-family noise. It claims to derive upper bounds on the recovery error of the proposed estimator that attain the rate O(n_3 r_t (max_k s_k^2) log((n1+n2)n3)/n) in certain cases, and presents minimax lower-bound analysis showing that the upper bounds approach optimality with a gap of order O(max_k s_k^2 / max(n1,n2)) up to logarithmic factors. Experiments on synthetic, image, and video tensors are included to support the claims.","tokens_in":2014,"tokens_out":579,"duration_ms":30283,"significance":"If the stated bounds hold under the paper's modeling assumptions, the work would provide a useful theoretical contribution to tensor completion by extending TV regularization to general exponential-family noise while jointly capturing sparsity and low-rank structure in the gradient tensor. The near-matching of upper and lower bounds would be a positive feature, and the experimental results on real data would add practical support.","major_comments":[{"comment":"Abstract and theoretical sections: the manuscript asserts rigorous derivation of the upper bound O(n_3 r_t (max_k s_k^2) log((n1+n2)n3)/n) and the minimax lower-bound gap, yet supplies no derivation steps, no explicit assumption list, and no verification that the transformed-L1 DTV construction produces the claimed rates under the exponential-family likelihood; these steps are load-bearing for the central theoretical contribution.","section":"Abstract"},{"comment":"Modeling section on DTV regularizers: the claim that the gradient tensor admits simultaneous sparsity and low-rank structure exactly captured by the transformed-L1 family is invoked to obtain the error bounds, but the manuscript does not demonstrate necessity or sufficiency of this structure nor verify that the regularizer enforces it under the given noise model.","section":"Modeling section on DTV regularizers"},{"comment":"Theoretical analysis: the upper and lower bounds are presented as following from the estimator definition, but without shown intermediate steps it is impossible to confirm whether the rates reduce to quantities defined solely by the fitted parameters or require additional unstated conditions on the gradient tensor.","section":"Theoretical analysis"}],"minor_comments":[{"comment":"Notation for dimensions (n, n1, n2, n3) and parameters (r_t, s_k) should be defined once at first use and used consistently.","section":"Notation"},{"comment":"The phrase 'in certain cases' for the upper-bound rate should be replaced by an explicit statement of the conditions under which the rate holds.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the careful reading and constructive comments. We address each major comment below and will revise the manuscript accordingly to improve clarity of the theoretical contributions.","responses":[{"response":"We agree that the main text presents the bounds without full intermediate derivation steps. The complete proofs appear in the appendix, but to address this we will add an explicit assumption list and a high-level proof sketch in the main theoretical section of the revision, showing how the transformed-L1 DTV yields the stated rates under the exponential-family likelihood.","revision_made":"yes","referee_comment":"[Abstract] Abstract and theoretical sections: the manuscript asserts rigorous derivation of the upper bound O(n_3 r_t (max_k s_k^2) log((n1+n2)n3)/n) and the minimax lower-bound gap, yet supplies no derivation steps, no explicit assumption list, and no verification that the transformed-L1 DTV construction produces the claimed rates under the exponential-family likelihood; these steps are load-bearing for the central theoretical contribution."},{"response":"The modeling choice is motivated by the observation that gradient tensors of image and video data commonly exhibit joint sparsity and low-rank structure. The transformed-L1 family is selected as a nonconvex surrogate that promotes both. We will expand the modeling section in revision to include a discussion of sufficiency for the error bounds and interaction with the exponential-family noise model.","revision_made":"yes","referee_comment":"[Modeling section on DTV regularizers] Modeling section on DTV regularizers: the claim that the gradient tensor admits simultaneous sparsity and low-rank structure exactly captured by the transformed-L1 family is invoked to obtain the error bounds, but the manuscript does not demonstrate necessity or sufficiency of this structure nor verify that the regularizer enforces it under the given noise model."},{"response":"The bounds rely on the structural assumptions (sparsity and low-rankness) of the gradient tensor that are captured by the regularizer. We will include the key intermediate steps and clarify in the revised theoretical analysis which quantities depend only on fitted parameters versus the gradient-tensor assumptions.","revision_made":"yes","referee_comment":"[Theoretical analysis] Theoretical analysis: the upper and lower bounds are presented as following from the estimator definition, but without shown intermediate steps it is impossible to confirm whether the rates reduce to quantities defined solely by the fitted parameters or require additional unstated conditions on the gradient tensor."}],"tokens_in":1559,"tokens_out":535,"duration_ms":37452,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper introduces a family of dual total-variation regularizers built from transformed L1 penalties. These are meant to jointly enforce sparsity and low-rank structure on the gradient tensor while handling exponential-family noise in tensor completion. That construction and its application to the broader noise model look new compared with earlier TV tensor work.\n\nIt does a clean job stating the target rates: upper bounds of order O(n3 rt (maxk sk^2) log((n1+n2)n3)/n) in certain cases, plus a minimax lower bound whose gap is only O(maxk sk^2 / max(n1,n2)) up to logs. The abstract also flags experiments on synthetic, image, and video data.\n\nThe central modeling choice is that the gradient tensor must admit exactly the simultaneous sparsity-plus-low-rank pattern that this particular regularizer family is designed to capture. The bounds are derived under that assumption. If the data do not satisfy it, or if the regularizer does not enforce it under the exponential-family likelihood, the guarantees do not follow. The abstract gives no derivation steps or explicit assumption list, so it is impossible to check whether the transformed-L1 construction is minimal or whether hidden gaps exist in the argument.\n\nThis is for researchers who work on statistical tensor estimation with non-Gaussian observations and want theoretical rates rather than purely algorithmic papers. A reader already familiar with TV regularization in tensors would see a targeted extension, but would still need to verify the proofs before relying on the bounds.\n\nI would send it to peer review. The claims are specific enough to be checked, the setting is relevant, and the modeling assumption is stated plainly even if it turns out to be restrictive.","headline":"New dual-TV regularizer via transformed L1 for exponential-family tensor completion, with claimed near-minimax rates that rest on a strong modeling assumption about the gradient tensor.","tokens_in":2520,"tokens_out":423,"would_cite":false,"duration_ms":20968,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Dual total-variation regularizers based on transformed L1 yield recovery error bounds for exponential-family tensor completion that approach the minimax lower bound.","keywords":["tensor completion","dual total variation","exponential family","nonconvex regularization","recovery error bounds","minimax analysis","gradient sparsity","low-rank structure"],"falsifier":"A concrete counter-example would be an instance where the gradient tensor satisfies the sparsity and low-rank conditions yet the observed recovery error exceeds the stated upper bound by more than the predicted gap of order O(max sk^2 / max(n1,n2)).","tokens_in":2746,"feed_emoji":"","tokens_out":752,"duration_ms":34917,"temperature":0.7,"pith_summary":"The paper develops regularization for recovering tensors from partial observations under exponential-family noise, which includes Gaussian and Poisson cases. It introduces a family of dual-TV regularizers that enforce both sparsity and low-rank structure on the gradient tensor. Theoretical analysis derives upper bounds on the estimator's recovery error that reach order O(n3 rt (max sk^2) log((n1+n2)n3)/n) in some regimes. Minimax lower-bound analysis shows these upper bounds are within a gap of order O(max sk^2 / max(n1,n2)) up to a logarithmic factor. Experiments on synthetic, image, and video data illustrate the approach.","feed_headline":"Dual-TV regularizers bound tensor error near minimax rate","feed_subtitle":"Upper bounds reach O(n3 rt sk^2 log / n) and close the gap to the lower bound by O(sk^2 / n) for exponential-family data.","key_machinery":"The family of dual-TV (DTV) regularizers based on the transformed L1 function, which simultaneously enforce sparsity and low-rank structure on the gradient tensor of the target.","core_discovery":"We propose a family of dual-TV regularizers based on the transformed L1 function for exponential-family tensor completion. These regularizers simultaneously capture sparsity and low-rank structures in the gradient tensor. We establish upper bounds on the recovery error that can attain the order O(n3 rt (maxk sk^2) log((n1+n2)n3)/n), and show via minimax analysis that these bounds approach the lower bound with a gap of O(maxk sk^2 / max(n1,n2)) up to a logarithmic factor.","pith_inferences":["The same regularizer structure might be tested on other inverse problems where data gradients are expected to be sparse and low-rank.","If the simultaneous sparsity-low-rank assumption is weakened, the error rates could still hold under milder conditions on the gradient.","The derived rates suggest that computational implementations could be scaled to larger tensors while preserving the theoretical guarantees."],"forward_implications":["Recovery error is bounded by O(n3 rt (max sk^2) log((n1+n2)n3)/n) under the modeling assumptions.","The estimator approaches minimax optimality because the upper bound is within O(max sk^2 / max(n1,n2)) of the lower bound up to logs.","The framework covers general exponential-family noise, recovering Gaussian and Poisson tensor completion as special cases.","The regularizer is nonconvex and handles both synthetic and real tensor data such as images and videos."],"fun_headline_variants":["Dual-TV nears minimax bounds for exponential tensor completion","Nonconvex DTV bounds tensor error close to lower bound","Transformed L1 DTV achieves near-minimax tensor error bounds","Recovery bounds with dual TV approach minimax rate"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The gradient tensor of the target admits simultaneous sparsity and low-rank structure that the specific transformed-L1 dual-TV regularizer can capture.","fun_headline_variants_meta":{"raw":{"variants":["Dual-TV nears minimax bounds for exponential tensor completion","Nonconvex DTV bounds tensor error close to lower bound","Transformed L1 DTV achieves near-minimax tensor error bounds","Recovery bounds with dual TV approach minimax rate"]},"model":"grok-4.3","cost_usd":0.006415,"raw_usage":{"total_tokens":3054,"prompt_tokens":761,"num_sources_used":0,"completion_tokens":65,"cost_in_usd_ticks":64149500,"prompt_tokens_details":{"text_tokens":761,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2228,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":761,"tokens_out":65,"duration_ms":24591,"temperature":1.0,"reasoning_tokens":2228,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-01T01:11:41.165598+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A concrete counter-example would be an instance where the gradient tensor satisfies the sparsity and low-rank conditions yet the observed recovery error exceeds the stated upper bound by more than the predicted gap of order O(max sk^2 / max(n1,n2)).","supporting_citations":[],"review_version":1}