{"id":"1ecb1692-12d3-4282-8a10-a3015d205b03","arxiv_id":"2506.18275","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"The paper derives RDT-based lower bounds and predicts a phase transition at oversampling ratio α≈1.4 where descending phase retrieval algorithms transition from failing to succeeding, but the key isomorphism with convergence is asserted heuristically.","lead":"This paper analyzes gradient-style 'descending' algorithms for phase retrieval and predicts a sample-complexity threshold at which they start succeeding. It introduces parametric manifolds and 'funneling points' as the conceptual bridge between the shape of a lower-bound objective and algorithm convergence.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central 'isomorphism' between single funneling points and global convergence is asserted without proof; since the analyzed manifold is a strict RDT lower bound, the phase-transition thresholds are not established.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: the paper asserts an isomorphism between single-funneling-point manifolds and global convergence of descending algorithms without supplying a definition or proof. My reading of the full text confirms this. Section 2.2 gives only a pictorial description, and the RDT theorems are lower-bound statements about solvability, not algorithmic convergence. Because the manifold studied is phi0, a lower bound for the true f-pro objective, even a rigorous single-funnel statement for phi0 would not transfer automatically to xi(c,x), let alone to the specific barrier/hybrid algorithms run in Section 4. The paper contains substantial RDT machinery and interesting numerics, but the central claim is stated as established while it is in fact unproved. This is a correctness risk, not merely a disagreement with prior work. I therefore agree with the REJECT verdict; my concern does not change the reader's conclusion.","tokens_in":31684,"tokens_out":5067,"duration_ms":58504,"concrete_test":"Check the transfer step directly: take the paper's own phi0(c,x) at alpha = 1.7932 and add a small positive bump h(c,x) concentrated in x in (0.3,0.7), c near 1, with h(1,1)=0, to form a candidate true objective xi = phi0 + h. Since Theorem 1 only guarantees xi >= phi0, this is an admissible true objective; if its parametric manifold has an extra funneling point while phi0 has one, the claimed isomorphism fails. Run this counterexample check symbolically or numerically, and if it fails, the paper must add missing assumptions before the phase-transition claim can stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 states that if the parametric manifold has a single 'funneling point' then 'any descending algorithm will converge to the global optimum,' and the abstract and contributions promote this to an 'isomorphism ... established.' No formal definition of a funneling point and no proof of this implication appear anywhere; the water-pouring analogy is not an argument. This is load-bearing because Theorems 1 and 2, and their lifted variants, only establish that phi0 > 0 implies uniqueness/solvability (frp > 0). The paper itself repeatedly notes that strong random duality is absent, so phi0 is a lower bound on the true objective xi(c,x). A monotone single-funnel lower-bound manifold does not imply the true manifold is single-funnel: a positive perturbation can create extra local minima. Moreover, the simulated 'hybrid' algorithm minimizes fbar with squared magnitudes and a log barrier, an objective not shown equivalent to f-pro; Section 4.3 concedes that the flat lifted curve makes alpha = 1.4 'difficult to make a definite conclusion.' Thus the central algorithmic claim rests on an unsupported step, and the lower-bound derivation cannot rescue it. The numerical agreement is suggestive but does not replace the missing proof.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the performance of descending (gradient-type) algorithms for real phase retrieval in the proportional high-dimensional regime. It introduces a constrained optimization formulation called f-pro, whose optimal value ξ(c,x) depends on the squared norm c and the overlap x with the true signal. Using Random Duality Theory (RDT) and a lifted variant, the paper derives lower bounds φ0 on the scaled objective and defines a 'parametric manifold' PM(α) in the (c,x) plane. It then claims an isomorphism between a single 'funneling point' of this manifold and global convergence of descending algorithms, and reports thresholds α≈1.7932 for plain RDT and α≈1.4 for lifted RDT. A hybrid barrier/plain gradient algorithm is implemented for n=300 squared-magnitude measurements, and its simulated success probability is compared with the theoretical thresholds.","tokens_in":31904,"tokens_out":3739,"duration_ms":39735,"significance":"If the central claim were established, the paper would provide a quantitative, parameter-free prediction of the oversampling ratio at which descending phase retrieval algorithms stop being trapped in spurious local minima. The RDT lower-bound derivations follow a known template and are explicit; the thresholds are evaluated from the derived integrals rather than fitted to simulations, and the numerical experiments in Figure 7 are suggestive. However, the paper's headline contribution, the funneling-point isomorphism, is only asserted through an informal water-pouring analogy in Section 2.2 and is never formalized or proved. Since the derived φ0 is explicitly a lower bound and strong random duality is absent, the manifold-shape analysis cannot by itself establish algorithmic convergence on the true objective. The gap between the non-squared constrained theory and the squared unconstrained/log-barrier simulated objective is also not rigorously closed, and Section 4.3 concedes that the α≈1.4 threshold is difficult to confirm numerically.","major_comments":[{"comment":"The claimed 'isomorphism' between a single funneling point of the parametric manifold and global convergence of descending algorithms is asserted, not proved. The formal statements in Theorems 1 and 2 (and their lifted analogues) only show that φ0 > 0 implies that the random primal value is positive with probability tending to one, i.e., uniqueness/solvability of the feasibility problem. Because Section 2.1 states that strong random duality is not in place, φ0 is a strict lower bound on the true objective; a single-funnel structure of a lower-bounding manifold does not transfer to the true objective, since a positive perturbation of a monotone funnel can create additional local minima. No formal definition of 'funneling point' or statement of the isomorphism is supplied. This missing step is load-bearing for the central thresholds α≈1.7932 and α≈1.4.","section":"Section 2.2"},{"comment":"The simulations minimize fbar(t0;x) with squared magnitudes and a log barrier, whereas the theoretical analysis in Sections 2–3 treats a constrained non-squared objective. Theorems 3 and 4 provide lower-bound analogues for squared magnitudes, but they do not establish that the phase-transition threshold of the constrained non-squared problem governs the unconstrained/log-barrier squared objective actually run in the experiments. Moreover, Section 4.3 explicitly concedes that the lifted squared-magnitude curve is flat and that 'it is a bit difficult to make a definite conclusion' about α≈1.4, which directly weakens the use of this value as the simulated transition point. The numerical agreement in Figure 7 is suggestive but does not replace the missing mathematical bridge.","section":"Sections 4.2 and 4.3"},{"comment":"The hybrid algorithm whose success probabilities are reported in Figure 7 includes sign-reshuffling steps and an increasing barrier schedule; it is not an instance of a pure descending algorithm on the manifold analyzed in Sections 2 and 3. The theoretical claims concern 'any descending algorithm,' but the simulated procedure can leave the descent path through the reshuffle operation. Thus the comparison in Figure 7 is not a direct test of the funneling-point isomorphism, and the numerical agreement cannot validate an unproved universal algorithmic statement.","section":"Section 4, Eq. (54)"},{"comment":"The paper concedes that the plain gradient has no generic phase transition because for any α one can find c>1 with multiple funneling points, and it explains the observed transition by the empirical fact that trajectories stay below c≈1.4 in practice. This explanation is trajectory-dependent and algorithm-dependent; it does not support the universal claim that above the threshold any norm-constrained descending algorithm reaches the global optimum. A trajectory-dependent empirical observation cannot substitute for a manifold-level convergence theorem.","section":"Section 4.1"}],"minor_comments":[{"comment":"The manuscript contains many typos and grammatical errors, including 'Paramatric manifold', 'agrement', 'matheamtical', 'go9ng', 'proeprties', and 'Figure 8 and 9'; these should be corrected.","section":"Throughout"},{"comment":"Several references have incomplete bibliographic data, e.g., [117]–[119] are listed as 'available online at arxiv' without identifiers; full citations should be provided.","section":"References"},{"comment":"The derivation from the integral definitions in (28) to the closed form for f_q in (29) is difficult to verify; a step-by-step derivation or an appendix would improve reproducibility.","section":"Section 2.1, Eqs. (28)–(29)"},{"comment":"The symbol x is used both for the overlap variable and for the optimization variable; although the paper notes this convention, the double use makes equations such as (7) and (11) unnecessarily confusing.","section":"Section 2, notation"},{"comment":"The vertical lines for 'Theoretical phase transition – RDT' and 'Lifted RDT' are not defined precisely; the caption should state whether they mark the threshold values themselves or a transition band.","section":"Figure 7"}],"recommendation":"reject","confidential_remarks":"The paper's headline contribution is a theorem-like statement that is asserted rather than proved, and the omitted step is not local: the funneling-point isomorphism is the bridge from the RDT lower bound to the algorithmic phase transition. The heavy reliance on the author's own RDT references also makes independent verification of the random-dual steps more difficult. A resubmission that supplies a formal funneling-point theorem, a precise class of algorithms covered, and a rigorous link between the squared and non-squared objectives would merit reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead the Stojnic paper. The novel piece is real: the parametric manifold PM(α) with (c,x) coordinates and the notion of funneling points give a clean way to visualize why descending phase-retrieval might fail, and the paper produces concrete numbers—α≈1.7932 from plain RDT, ≈1.4 after partial lifting—that are not in the earlier RDT literature. The closed-form evaluation of the random dual for the amplitude objective is competently done, and the numerical experiments at n=300 land surprisingly close to the advertised transitions. Credit where due: the paper also flags its own weak spots, noting no strong random duality, and Section 4.3 admits the squared-magnitude lifted curve is flat enough that α=1.4 is hard to pin down.\n\nThe soft spot is exactly where the reader puts it. The claim that a single funneling point in the parametric manifold is isomorphic to global convergence of descending algorithms appears in the abstract, Section 1.2, and Section 2.2, but no formal definition of funneling point and no proof of the implication are given. The water-pouring analogy is an intuition, not an argument. Worse, the manifold is computed from φ0, which the paper itself states is a strict lower bound on the true objective. A lower-bound landscape with one funnel says essentially nothing about the actual objective, which can have extra local minima under a positive perturbation. So the thresholds are best read as predictions from a heuristic, not as theorems about descending algorithms. The simulated agreement is encouraging but cannot certify the missing proof, especially since the simulation uses squared magnitudes with a log-barrier while the derivation uses amplitudes.\n\nThe citation pattern is heavy on the author's own RDT work, but in context that is the natural toolkit; the thresholds are evaluated from integrals, not fit to data, so I don't see a circularity problem of that kind.\n\nBottom line: this is a substantial, thought-provoking paper with real numerical suggestions, but the central claim is not established. It deserves a serious referee—the machinery is nontrivial and the predictions are falsifiable—so send it out, but expect the main contribution to be reframed as conjecture plus strong evidence, or to require a genuinely new proof. I would not cite it as a proved phase transition, but I'd read follow-ups.","headline":"Novel manifold/funneling-point picture and sharp numerical thresholds for descending phase retrieval, but the load-bearing 'isomorphism' is asserted without proof, so the thresholds remain predictions.","tokens_in":32425,"tokens_out":3156,"would_cite":false,"duration_ms":34101,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","65K10","94A12"],"pacs":[],"model":"deepseek-v4-flash","headline":"For Gaussian phase retrieval, the paper claims that a single funneling point on a random-duality manifold is equivalent to global convergence of every descending algorithm, placing the transition near an oversampling ratio of 1.79, or…","keywords":["phase retrieval","descending algorithms","random duality theory","parametric manifold","funneling points","phase transition","gradient descent","oversampling ratio"],"falsifier":"Run a norm-constrained descending algorithm at α = 1.4 with n = 10,000 Gaussian measurements and many random starts, including starts with overlap near 0; if a positive fraction converge to a point with overlap far below 1, the single-funneling-point claim is false. A second check: numerically trace the true objective's stationary points over (c, x) at α = 1.5 and look for any local minimum besides ±x̄.","tokens_in":31445,"feed_emoji":"📉","tokens_out":7268,"duration_ms":70833,"temperature":0.7,"pith_summary":"Phase retrieval asks for a signal from only the squared magnitudes of linear measurements. This paper tries to establish exactly when the most natural class of solvers—descending algorithms such as gradient descent and Wirtinger flow—are guaranteed to find the true signal, up to global phase, regardless of initialization. The answer it proposes is geometric: performance is controlled by a two-parameter parametric manifold of lower bounds, and success coincides with the manifold having a single funneling point at the true solution. Using Random duality theory, the paper locates the transition at about 1.79 measurements per unknown, and at about 1.4 after a lifted bound. If correct, this gives a quantitative phase-transition limit for nonconvex phase retrieval and explains why a simple hybrid gradient method succeeds in simulations near those values.","feed_headline":"One funneling point makes descent phase retrieval global","feed_subtitle":"Past 1.79 measurements per unknown, any descending method finds the true signal; lifting lowers that to 1.4.","key_machinery":"The central objects are the parametric manifold PM(α) and its funneling points. PM(α) plots the random-duality lower bound φ0(c, x) on the scaled objective ξ(c, x)/n against two parameters: c = ‖x‖², the squared norm of the algorithmic iterate, and x = xᵀx̄, its overlap with the true signal. A funneling point is a collector of all descending paths on the manifold, and the paper's claimed isomorphism is that a single funneling point at (1, 1) guarantees global convergence of every descending algorithm. The machinery that produces the manifold is the Random duality theory recipe: rewrite phase retrieval as a random optimization problem, form a random dual through a Gaussian comparison inequality, solve the scalarized dual in closed form for amplitude objectives, and lift the resulting lower bound using a partially lifted RDT variant.","core_discovery":"The paper's central claim is that the success of descending phase retrieval algorithms is governed by the shape of a two-parameter random-duality manifold: for each allowed pair (c, x), representing the squared norm of the iterate and its overlap with the true signal, a lower bound φ0(c, x) defines a surface over which descent flows. If that manifold has exactly one funneling point, located at the true solution (c, x) = (1, 1), then any norm-constrained descending algorithm converges to the global optimum from any initialization. If it has more than one funneling point, descent generically fails by being captured at an undesired collector such as (c, x) = (1, 0). The paper derives this manifold from a fundamental phase-retrieval optimization via Random duality theory, observes that increasing the oversampling ratio α = lim_{n→∞} m/n changes the manifold from multi-funnel to single-funnel, and locates the transition at α ≈ 1.7932 for plain RDT and α ≈ 1.4 for partially lifted RDT.","pith_inferences":["If the funneling-point criterion transfers to the true optimization landscape, the finite-n success probability should sharpen toward a step function at the threshold as n grows; this is a testable prediction the paper does not run at large n.","The same two-parameter manifold analysis could be applied to other nonconvex recovery problems whose objectives admit random-duality lower bounds, such as matrix completion, blind deconvolution, or phase retrieval with generative priors.","Because plain RDT gives strict lower bounds and the lifted threshold still sits above the information limit α = 1, a fully lifted treatment may push the guaranteed-success threshold lower; the paper identifies this as a next step but does not claim the lower value.","Spectral initialization probably makes the practical transition appear at slightly lower α than the worst-case guarantee, because it places the start inside the good basin; the paper observes the favorable overlap but does not quantify this gain."],"forward_implications":["Above α ≈ 1.7932, plain RDT predicts that any norm-constrained descending algorithm reaches the global optimum from any initialization, and below that ratio descending algorithms generically fail.","Partially lifted RDT lowers the guaranteed-success oversampling to α ≈ 1.4, so the true algorithmic phase transition should lie between 1.4 and 1.7932 rather than at the plain-RDT value.","The manifold parameter c, the squared norm of the iterate, is load-bearing: unconstrained plain gradient can fail even for larger α because it enters the c > 1 region where undesired funneling points reappear.","The same single-funneling-point conclusion holds for squared-magnitude objectives, the form used in practical implementations, with the lifted bound flattening the curve at α = 1.4.","A hybrid alternating barrier-gradient and plain-gradient method run at n = 300 shows a simulated transition fairly close to both theoretical predictions, indicating that finite-dimension jitteriness does not wash out the effect."],"supporting_citations":[{"why":"Introduces Wirtinger flow, the descending gradient algorithm whose performance limits the paper analyzes.","marker":"[28]"},{"why":"Supplies the Gaussian comparison theorem used to form the random dual and derive the lower bound φ0.","marker":"[54]"},{"why":"Sets out the Random duality theory principles used to obtain the algebraic representation and random dual.","marker":"[104,111]"},{"why":"Provides the partially lifted RDT variant that produces the improved manifold and the lower threshold α ≈ 1.4.","marker":"[109,110,115]"},{"why":"Characterizes spectral initialization and overlap behavior, motivating the initializer used in the practical simulations.","marker":"[77,90]"},{"why":"Fully lifted RDT, identified as the route to an exact rather than lower-bound characterization of the manifold.","marker":"[114]"}],"fun_headline_variants":["Phase retrieval descent succeeds past 1.79 samples per unknown","Descent phase retrieval hits global success at 1.79x oversampling","Phase transition at 1.79: when descent always finds the signal","One funnel point on the manifold guarantees descent success","Sample complexity shapes descent: 1.79 marks the success edge"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that when the random-duality lower-bound surface has exactly one collecting point, every descending algorithm converges globally; this transfer from a lower bound to the real optimization landscape is asserted rather than proved.","fun_headline_variants_meta":{"raw":{"variants":["Phase retrieval descent succeeds past 1.79 samples per unknown","Descent phase retrieval hits global success at 1.79x oversampling","Phase transition at 1.79: when descent always finds the signal","One funnel point on the manifold guarantees descent success","Sample complexity shapes descent: 1.79 marks the success edge"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000617,"raw_usage":{"total_tokens":2889,"prompt_tokens":991,"completion_tokens":1898,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":1809}},"tokens_in":607,"tokens_out":1898,"duration_ms":11908,"temperature":1.0,"reasoning_tokens":1809,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:52:59.101389+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a norm-constrained descending algorithm at α = 1.4 with n = 10,000 Gaussian measurements and many random starts, including starts with overlap near 0; if a positive fraction converge to a point with overlap far below 1, the single-funneling-point claim is false. A second check: numerically trace the true objective's stationary points over (c, x) at α = 1.5 and look for any local minimum besides ±x̄.","supporting_citations":[],"review_version":2}