{"id":"d3cf6da7-3aa2-41f2-8a0c-3ba40be23737","arxiv_id":"2509.10834","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a simplified tensor-sensing surrogate of the BDPR problem, the population risk has only benign critical points and Riemannian gradient descent converges linearly under a restricted-isometry condition.","lead":"This paper analyzes the mathematics of an optimization problem that combines blind deconvolution and phase retrieval, a task that appears in phase imaging with partially coherent light. It proves convergence guarantees for a gradient method on a simplified surrogate version of the problem, while the real structured sensing model is left unproven.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Structured sensing TRIP is the unproven bridge; without it the convergence theorems do not apply to the original BDPR problem.","rationale":"The reader's weakest assumption identifies the correct critical point: the paper's relevance to BDPR depends on the structured sensing operator satisfying TRIP, and the formal proof is explicitly deferred. I agree with that assessment. The surrogate-theoretic results are self-contained: Theorem IV.1 appears to correctly characterize the population-risk critical points, and Theorems V.3 and V.5 are coherent conditional results under TRIP. There are smaller caveats, such as Theorem IV.1 not by itself implying global convergence of arbitrary first-order methods because the saddle class has flat directions, and the spectral-initialization-to-local-basin link requiring a much smaller delta than 3/15, but the TRIP bridge is the load-bearing issue. A failure of TRIP for (3), or a TRIP that only holds at m much larger than NI, would leave no theorem about the original BDPR problem, only the authors' own structured experiments, which they state are outside the theory. Because the limitation is acknowledged and the surrogate analysis can stand on its own, the appropriate verdict remains CONDITIONAL, requiring either a structured TRIP proof or an explicit scoping-away of the BDPR claim.","tokens_in":27898,"tokens_out":21641,"duration_ms":189326,"concrete_test":"Using the exact structured operator (3) with N=25, K=6, I=60 (m=1500) and resampled Gaussian masks, P_i, and B, estimate the rank-five restricted-isometry spread rho_min and rho_max of m^{-1}||A(T)||^2 over unit-Frobenius CP tensors T = sum_{i=1}^5 x_i o x_i o h_i, via a projected power method with many random restarts and over several independent draws of A. Check whether [rho_min, rho_max] lies within [1-3/15, 1+3/15]. If either ratio departs from this interval, the structured operator violates the TRIP level required by Theorem V.5, and the claimed transfer of the surrogate theorems to BDPR is not justified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the structured sensing operator A_{i,n} in (3), built from DFT matrices, Gaussian masks, and random P_i,B, satisfies the rank-r TRIP that Theorems V.3 and V.5 require. Theorem V.1 supplies TRIP only for i.i.d. subgaussian sensing tensors; immediately after it, the paper states that 'We leave the formal proof of TRIP for the structured sensing operator (3) used in the original BDPR problem as future work.' Section VI-B likewise admits that the structured experiments are outside the current theory. Theorems V.3 and V.5 are conditional on TRIP with r=5 and delta_r at most 3/15. The structured operator is not a subgaussian ensemble, so its restricted isometry constants are not implied by Theorem V.1. If A in (3) fails TRIP, or satisfies it only when m is much larger than NI, then the convergence guarantees do not transfer to BDPR. In that case the abstract's claim that the surrogate 'equivalently characterizes the original BDPR problem' is unsupported, and the paper's contribution reduces to a generic rank-one partial-symmetric tensor sensing analysis.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the simultaneous blind deconvolution and phase retrieval (BDPR) problem through a structured low-rank tensor recovery surrogate. The target tensor is T* = x* ∘ x* ∘ h*, and the authors analyze the population risk f(x,h) = 1/2 ||x∘x∘h − T*||_F^2 on the unit sphere. Theorem IV.1 classifies all critical points as either the global optima (±x*, h*) or non-strict saddles of the form x⊥x*, h=0. Theorem IV.2 proves local linear convergence of a hybrid Riemannian gradient descent (RGD) scheme for the factorization problem. The paper then extends the analysis to the tensor sensing problem (12) under a tensor restricted isometry property (TRIP): Theorems V.2–V.5 give spectral initialization accuracy, local linear convergence, and a noise-robust variant. Numerical experiments cover both Gaussian sensing tensors and the structured sensing tensors (3) arising from the original BDPR model.","tokens_in":28077,"tokens_out":11121,"duration_ms":97261,"significance":"The critical-point analysis in Section IV is clean and clearly derived, and the local convergence proofs give explicit contraction rates and noise floors, which are useful and falsifiable. If the TRIP-based theorems hold, the paper provides a rigorous landscape analysis for rank-one partial-symmetric tensor sensing, a model that is more general than the original BDPR problem. The main weakness is that the connection to BDPR is conditional: the TRIP for the structured sensing operator (3) is explicitly left to future work, and Section VI-B admits that the structured experiments are outside the current theory. The paper is therefore stronger as a standalone tensor-sensing analysis than as a landscape analysis of BDPR itself; the abstract and conclusion overstate the equivalence between the surrogate and the original problem.","major_comments":[{"comment":"The load-bearing bridge from the surrogate to the original BDPR problem is unproven. Theorem V.3 and Theorem V.5 require A to satisfy the TRIP with r=5, but Theorem V.1 supplies TRIP only for i.i.d. subgaussian sensing tensors. Immediately after Theorem V.1, the paper states that the formal proof of TRIP for the structured sensing operator (3) is left as future work, and Section VI-B states that the structured experiments are outside the current theory. Despite this, the abstract and conclusion claim that the surrogate \"equivalently characterizes the original BDPR problem\" and provides \"principled guidance\" for it. If the structured operator (3) fails TRIP, or satisfies it only with a much larger measurement complexity than the subgaussian bound (11), none of the convergence theorems transfers to BDPR. The authors should either prove an appropriate TRIP for (3), or explicitly reframe the contribution as an analysis of generic rank-one partial-symmetric tensor sensing and remove the BDPR-equivalence claims.","section":"V (after Theorem V.1), VI-B, Abstract"},{"comment":"Lemma B.3 is stated for TRIP with r=2 and for \"any CP format tensors X1, X2\". As stated, this is false when X1±X2 has CP rank larger than 2, because TRIP with r=2 gives no control over such pairs. The proof of Theorem V.2 invokes \"Lemma B.3 with r=3\", and the proof of Theorem V.3 invokes \"Lemma B.3 with r=5\", so the paper implicitly relies on a rank-aware polarization bound of the form: if A satisfies TRIP with r ≥ rank(X1) + rank(X2), then |(1/m)<A(X1),A(X2)> − <X1,X2>| ≤ δ_r ||X1||_F ||X2||_F. The lemma must be restated in that general form and the rank budget verified in each application. Without this correction, Theorem V.2 and the convergence proofs built on it are not fully supported as written.","section":"Appendix B, Lemma B.3 (used in Appendices D, E, F, G)"},{"comment":"The noisy theorems assume only that e has i.i.d. entries with mean zero and variance γ², but the proofs use the concentration bound (47) cited from [54, eq. (D.6)], which requires at least sub-Gaussian (or bounded) noise entries. Finite variance alone does not give the exponentially small failure probability needed to control the spectral initializer and the iterates. The authors should either strengthen the noise assumption to sub-Gaussian entries, or supply a proof under the stated second-moment assumption. As written, the noise-robustness guarantee is not fully justified.","section":"V-B (Theorems V.4 and V.5), Appendix G"}],"minor_comments":[{"comment":"In the displayed bound for ||c2||_2, the second inequality reads ||c2−b2||_2 + ||c2||_2; it should be ||c2−b2||_2 + ||b2||_2. The numerical bound is unchanged, but the typo obscures the argument.","section":"Appendix E, after equation (40)"},{"comment":"In the display after equation (54), \"3−15δ_c\" should be \"3−15δ_r\"; this appears to be a typographical error.","section":"Appendix G, proof of Theorem V.5"},{"comment":"The text says the authors \"replace ||T*||_F^2 with ||T_t||_F^2 in the RGD updates (8)\". This modification is not covered by Theorems IV.2, V.3, or V.5, which use the fixed quantity ||T*||_F^2. A remark explaining why this algorithmic change is benign would improve clarity.","section":"VI-B, structured experiments"},{"comment":"The noise floor is written as O((5(N+K)+5^3)/m γ²), while the proof tracks terms of the form O(sqrt((5(N+K)+5^3)/m) γ) before applying Young's inequality. Please clarify the exact constants and how the final expression is obtained.","section":"Theorem V.5 statement"},{"comment":"The sentence that any first-order method producing h≠0 \"will converge to the second class of critical points\" is informal: gradient descent can in principle converge to a non-strict saddle or move slowly along the flat directions unless additional conditions are imposed. This statement should be softened or given a precise qualification.","section":"Section IV-A, after Theorem IV.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest about the unproven TRIP for the structured operator, but the title, abstract, and conclusion claim more BDPR relevance than the results support. If the authors cannot prove TRIP for (3), the contribution should be reframed as a tensor-sensing landscape analysis. The Lemma B.3 gap is a correctness issue in the proof apparatus but appears fixable by stating the standard rank-aware polarization bound. The noise-assumption mismatch is similarly fixable. I would like the editor to verify that the revised framing matches the journal's scope in eess.SP."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core of this paper is a clean, elementary analysis of the population risk for the partial symmetric rank-one tensor x∘x∘h. Theorem IV.1 correctly identifies all critical points: the ground truth (up to sign) and non-strict saddles of the form x⊥x*, h=0. The Hessian computation is explicit and the classification is new, since existing tensor landscape results do not cover the shared-factor structure. That part is solid and citable.\n\nThe paper also proves local linear convergence of RGD for the unstructured subgaussian tensor sensing surrogate, with spectral initialization and noise robustness. These results are plausible and follow the established TRIP machinery. The experiments support the theory on the surrogate, and the paper is honest enough to state that the structured experiments fall outside the theory.\n\nThe soft spots are real but not fatal to the core contribution. First, the abstract and contributions claim the surrogate 'equivalently characterizes' the original BDPR problem. That is not supported. The structured sensing operator in (3) is not subgaussian, and the paper explicitly leaves the formal TRIP for that operator as future work. Without that bridge, the convergence theorems simply do not apply to BDPR; they apply to a generic rank-one tensor sensing model. The authors know this, but the abstract overstates it. Second, the claim that 'any first-order method can converge to the global optimum' is not what is proven. The theorems give local convergence near the ground truth, not a global guarantee. The population landscape contains zero-curvature saddles, so a first-order method could stall. Third, the proofs in Appendices C-G contain algebra I could not fully verify. The dist^4 bounds in equations (30)-(34) and the expansion around (27) are unusual; they may be fixable, but they need a careful rewrite before publication.\n\nWho is this for? Researchers working on landscape analysis of tensor factorization and nonconvex inverse problems. The critical point classification is a useful contribution; the surrogate results are a stepping stone, not a resolution of the BDPR problem. I would send it to serious referees, but with the expectation of major revision: soften the claims about BDPR, fix the appendix details, and either prove the structured TRIP or explicitly scope the paper as an analysis of the unstructured surrogate.","headline":"A correct and novel landscape classification for the x∘x∘h population risk, but the paper's claim to characterize the original BDPR problem rests on an unproven TRIP bridge.","tokens_in":28615,"tokens_out":2302,"would_cite":false,"duration_ms":22678,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["15A69","65K10","90C26","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"For the surrogate of simultaneous blind deconvolution and phase retrieval, every critical point is either the global optimum or a non-strict saddle.","keywords":["blind deconvolution","phase retrieval","tensor factorization","tensor sensing","geometric landscape","rank-one tensor recovery","Riemannian gradient descent"],"falsifier":"Compute the restricted isometry constant of the structured operator (3) numerically: draw many rank-one tensors x∘x∘h, evaluate deviations of (1/m)||A(x∘x∘h)||² from ||x∘x∘h||², and check whether they stay below the theory's δ when m=O(N+K); a violating tensor, or a large gap between the structured and Gaussian phase transitions of RGD, would break the surrogate bridge.","tokens_in":27648,"feed_emoji":"📡","tokens_out":8032,"duration_ms":64213,"temperature":0.7,"pith_summary":"The paper analyzes simultaneous blind deconvolution and phase retrieval (BDPR), where one must recover both an unknown signal x and source-shape coefficients h from intensity-only blurred measurements. Because the sensing tensor of the original problem is too complicated to analyze directly, the authors study a surrogate: recovering the rank-one partial-symmetric tensor T* = x*∘x*∘h* from linear measurements. For the population risk of this factorization, they prove that all critical points are either the global optimum x=±x*, h=h* or non-strict saddles x⊥x*, h=0, so there are no spurious local minima. They then prove that Riemannian gradient descent with a spectral initialization converges linearly under a tensor restricted isometry property, and degrades gracefully under noise. If the surrogate faithfully captures the original problem, these results provide the first convergence guarantees for a class of BDPR algorithms that previously had none.","feed_headline":"No spurious local minima in BDPR surrogate","feed_subtitle":"A rank-one tensor surrogate makes gradient descent provably recover x and h, even under noise.","key_machinery":"The load-bearing object is the partial-symmetric rank-one CP tensor x∘x∘h, in which two modes share the factor x. The machinery has four parts: the population risk f(x,h)=1/2||x∘x∘h−T*||$_F^{2}$ over the unit sphere, whose hybrid Riemannian Hessian at critical points takes the closed form in equation (7) and immediately shows positive semidefiniteness; the factor distance dist²(x,h)=min_{a=±1} 2||T*||_F²||x−ax*||²+||h−h*||² together with Lemma IV.1 relating it to the tensor reconstruction error; the tensor restricted isometry property (TRIP) for CP rank at most r, which preserves inner products under the linear sensing map; and a hybrid Riemannian gradient descent update with a spectral initialization obtained from the SVD of the mode-1 unfolding of the adjoint applied to the measurements. These ingredients convert a nonconvex tensor recovery problem into one where all critical points are classified and descent contracts the distance metric at a linear rate.","core_discovery":"On its own terms, the paper's central discovery is that the factorized objective f(x,h)=1/2||x∘x∘h−T*||$_F^{2}$ with ||x||_2=1 has a benign landscape despite the repeated factor x. Theorem IV.1 characterizes every critical point: the global minima are x=±x*, h=h*, and the only other critical points are x⊥x*, h=0, where the hybrid Riemannian Hessian is positive semidefinite. This means first-order methods that keep h away from zero cannot be trapped away from the truth. For the tensor sensing surrogate (12), the paper shows that under TRIP with rank up to 5, the spectral initializer lands within a controlled neighborhood, and RGD contracts the distance metric by a factor of 1−O(µ) per iteration; with noise, convergence stops at a floor proportional to γ√((N+K)/m). The authors present these results as principled guidance for the original BDPR problem because the underlying tensor x*∘x*∘h has the same structure, while explicitly deferring a direct landscape analysis of the structured sensing tensor (3) and a proof of TRIP for it to future work.","pith_inferences":["If the structured operator (3) is later shown to satisfy TRIP, Theorems V.2-V.5 would immediately supply recovery guarantees for the original BDPR problem; the missing link is purely a verification of the isometry constant.","The same Hessian calculation should extend to other factor-sharing low-rank templates, such as tensors of the form x∘y∘y or higher-order symmetric slices, which would give a unified proof tool for repeated-factor tensor factorization landscapes.","A quantitative test of surrogacy would compare the phase transition of RGD on the structured tensors (3) with the Gaussian phase transition: if the required number of measurements tracks N+K with similar constants, the surrogate captures the original geometry.","The benign saddle structure suggests that even without exact TRIP, relaxed isometry conditions or batch-stochastic variants of RGD may still work, so the practical guidance of the paper is likely stronger than its theorems."],"forward_implications":["Because Theorem IV.1 rules out spurious local minima, any first-order method that reaches a point with h≠0 will converge to the global optimum up to the sign ambiguity.","With subgaussian sensing tensors, m=Ω(N+K) measurements suffice for the TRIP prerequisite, so RGD with spectral initialization is sample-efficient in the problem's ambient dimension (Theorem V.1 and Theorem V.3).","In the noisy case, RGD converges linearly down to a noise floor of order γ√((N+K)/m), meaning measurement noise degrades the estimate gracefully rather than derailing convergence (Theorem V.5).","Lemma IV.1 transfers the linear contraction in factor distance to linear contraction in tensor Frobenius error, so the guarantee applies to the actual reconstruction error.","The structured sensing experiments show the same linear convergence pattern, which the paper offers as empirical evidence that the surrogate analysis guides the original BDPR problem even though the theory does not yet cover it."],"supporting_citations":[{"why":"formulates BDPR as structured rank-one tensor recovery x*∘x*∘h and supplies the measurement model (3) that the paper analyzes","marker":"[19]"},{"why":"supplies the low-CP-rank tensor restricted isometry property definition and iterative-hard-thresholding context adopted as Definition 1","marker":"[47]"},{"why":"provides the subgaussian tensor measurement-complexity theorem invoked as Theorem V.1 for TRIP with rank-r CP tensors","marker":"[48]"},{"why":"supplies the Riemannian gradient descent proof template, the inner-product preservation lemma (Lemma B.3), and the spectral initialization strategy","marker":"[29]"},{"why":"provides the singular-subspace perturbation bound used in Lemma B.2 to relate factor distance to tensor error","marker":"[51]"},{"why":"supplies the nonexpansive retraction property on the Stiefel manifold used in the RGD contraction proofs","marker":"[52]"},{"why":"supplies the concentration bound on noise-corrupted tensor measurements used in Theorem V.5's noise floor","marker":"[54]"}],"fun_headline_variants":["Benign landscape in tensor surrogate for BDPR","RGD provably converges for BDPR surrogate","Surrogate tensor model eliminates spurious minima","Landscape analysis yields linear convergence for RGD","Tensor surrogate enables provable recovery in BDPR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the structured sensing tensors of the original BDPR problem obey the same tensor restricted isometry condition as the random subgaussian surrogate, so the surrogate's benign landscape carries over; the paper leaves that proof to future work.","fun_headline_variants_meta":{"raw":{"variants":["Benign landscape in tensor surrogate for BDPR","RGD provably converges for BDPR surrogate","Surrogate tensor model eliminates spurious minima","Landscape analysis yields linear convergence for RGD","Tensor surrogate enables provable recovery in BDPR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000849,"raw_usage":{"total_tokens":3718,"prompt_tokens":996,"completion_tokens":2722,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2659}},"tokens_in":612,"tokens_out":2722,"duration_ms":391216,"temperature":1.0,"reasoning_tokens":2659,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:53:42.329422+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the restricted isometry constant of the structured operator (3) numerically: draw many rank-one tensors x∘x∘h, evaluate deviations of (1/m)||A(x∘x∘h)||² from ||x∘x∘h||², and check whether they stay below the theory's δ when m=O(N+K); a violating tensor, or a large gap between the structured and Gaussian phase transitions of RGD, would break the surrogate bridge.","supporting_citations":[{"cited_title":"Simultaneous blind deconvolution and phase retrieval with tensor iterative hard thresholding,","cited_arxiv_id":null,"evidence_quote":"formulates BDPR as structured rank-one tensor recovery x*∘x*∘h and supplies the measurement model (3) that the paper analyzes"},{"cited_title":"Iterative hard thresholding for low CP-rank tensor models,","cited_arxiv_id":null,"evidence_quote":"supplies the low-CP-rank tensor restricted isometry property definition and iterative-hard-thresholding context adopted as Definition 1"},{"cited_title":"Low rank tensor recovery via iterative hard thresholding,","cited_arxiv_id":null,"evidence_quote":"provides the subgaussian tensor measurement-complexity theorem invoked as Theorem V.1 for TRIP with rank-r CP tensors"},{"cited_title":"Guaranteed nonconvex factorization approach for tensor train recovery,","cited_arxiv_id":null,"evidence_quote":"supplies the Riemannian gradient descent proof template, the inner-product preservation lemma (Lemma B.3), and the spectral initialization strategy"},{"cited_title":"Provable tensor-train format tensor completion by Riemannian optimization,","cited_arxiv_id":null,"evidence_quote":"provides the singular-subspace perturbation bound used in Lemma B.2 to relate factor distance to tensor error"},{"cited_title":"Weakly convex optimization over Stiefel manifold using Riemannian subgradient-type methods,","cited_arxiv_id":null,"evidence_quote":"supplies the nonexpansive retraction property on the Stiefel manifold used in the RGD contraction proofs"},{"cited_title":"An optimal statistical and computational framework for generalized tensor estimation,","cited_arxiv_id":null,"evidence_quote":"supplies the concentration bound on noise-corrupted tensor measurements used in Theorem V.5's noise floor"}],"review_version":2}