{"id":"9e1e15f8-bda7-4d3c-a38e-e85558c14117","arxiv_id":"2511.19045","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Constant overparametrization makes every second-order critical point of the nonsmooth amplitude loss a statistically optimal phase-retrieval estimator from O(d) sub-Gaussian measurements.","lead":"This paper proves a benign-landscape guarantee for phase retrieval with the nonsmooth amplitude loss: every second-order critical point of the overparametrized factorization recovers the signal at the statistically optimal noise rate using only a constant amount of overparametrization. It is the first global landscape analysis of this loss, and it extends to infinite-dimensional Hilbert spaces.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's printed limsup definition of second-order critical point is too weak; Lemma 2's proof uses a liminf-along-every-ray inequality, so the main nonsmooth landscape theorems are not proven as stated.","rationale":"I read the central claim as the proof that, with constant overparametrization, every second-order critical point of the nonsmooth amplitude loss is a good estimator. The deterministic landscape machinery (Lemma 1 -> Theorem 2 -> Theorem 5) is the load-bearing structure, and the definition of the objects to which the theorems apply is the least secure part. The printed limsup definition does not imply the per-direction second-order ascent used in the proof of Lemma 2; the exhibited 2x2 example shows Lemma 2 is false under that definition. This is exactly the reader's weakest_assumption, so I agree with the conditional disposition. I do not count the reliance on Lemma 5 from [34] as a separate objection: importing a published concentration/identifiability statement is standard practice, and the quoted lemma appears well matched to the theorem's assumptions. The PhaseCut portion may be insulated from the flaw because it uses the separate Lemma 3, but the nonsmooth half of Theorem 5 is not. The defect appears easily repairable, which is why I would not move to reject; conditional acceptance with a required definitional correction is the right call.","tokens_in":30002,"tokens_out":13406,"duration_ms":130440,"concrete_test":"Independently verify Lemma 2 on the 2x2 instance A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1). Compute the printed limsup condition for the amplitude loss ℓ(b,y)=(√b-y)^2 (it is satisfied), then check the two conclusions of Lemma 2: finiteness of ℓ'(0,1) and inequality (16) along a perturbation with ⟨A1,Xdot Xdot*⟩>0. If either conclusion fails while the printed definition holds, Lemma 2 is false as stated. Then re-run the proof of Lemma 2 with the SOCP definition replaced by liminf_{t->0} [f(X+t Xdot)-f(X)]/t^2 >= 0 for all unit Xdot and confirm the Taylor-expansion step is valid under that definition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 defines a second-order critical point by limsup_{X'->X} [f(X') - f(X)] / ||X'-X||^2 >= 0. This does not rule out quadratic descent along any particular direction: if the function has both positive- and negative-curvature directions, the global limsup can be positive because of the positive direction, even when a strict saddle is present. The proof of Lemma 2 (Section 3.4) nevertheless replaces second-order criticality by the inequality L((X+t Xdot)(X+t Xdot)^*) - L(XX^*) >= o(t^2) for every fixed unit Xdot, i.e. by liminf_{t->0} of the second-order difference along every ray. That inference is not licensed by the printed definition. In fact Lemma 2 is false as stated: with d=p=2, A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1), the printed limsup condition is satisfied (perturbations in the positive-curvature direction give ratio ~1) while ℓ'(0,1)=-∞ and the first-order stationarity conclusion of Lemma 2 fails. Since Lemma 2 underlies Lemma 1 and Theorems 1-3, and Theorem 5's nonsmooth part is obtained from Theorem 2, the central claim about every second-order critical point of (8) is currently unsupported. The likely fix is to define SOCP by requiring the liminf to be nonnegative along every unit ray (equivalently, no quadratic descent in any direction); this appears to be a typo-level quantifier error rather than a flaw in the overall strategy, but as written the proof does not go through.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the global nonconvex landscape of phase retrieval formulated as overparametrized low-rank semidefinite matrix sensing, with two objectives: the nonsmooth amplitude least-squares loss (8) and a smooth PhaseCut-type reformulation (26). The main deterministic results are a general lemma (Lemma 1) on second-order critical points for losses of the form L_λ(XX^*), specialized to the quartic, amplitude, and Poisson losses (Theorems 1–3), plus a parallel landscape theorem for the PhaseCut formulation (Theorem 4). These are used to prove high-probability recovery guarantees with constant overparametrization: Theorem 5 for finite-dimensional phase retrieval with isotropic sub-Gaussian measurements, and Theorem 6 for Gaussian measurements in a Hilbert space. The claimed statistical error rates are optimal up to constants in the noise level, improving on prior quartic-loss results that required logarithmic overparametrization.","tokens_in":30341,"tokens_out":14520,"duration_ms":141472,"significance":"If the results are correct, the paper makes a substantial contribution: it provides the first global landscape guarantees for the nonsmooth amplitude loss with only constant overparametrization, and it shows that the same techniques cover a smooth PhaseCut formulation. The deterministic inequalities are explicit and parameter-free, and the statistical statements are sharp in their noise dependence. The paper is honest about its limitations, including the suboptimal constant in the non-adversarial noise comparison in Section 5.1. However, the central definition of second-order critical point is not consistent with the proofs, and Theorem 6 omits a necessary condition on the optimization rank. These are load-bearing issues that currently prevent the main theorems from being accepted as proved, although they appear to be fixable without changing the overall strategy.","major_comments":[{"comment":"The printed definition of second-order critical point is limsup_{X'→X} [f(X')-f(X)]/||X'-X||^2 ≥ 0. This condition is strictly weaker than the property used in the proof of Lemma 2. It can hold at a strict saddle with both positive- and negative-curvature directions: for f(x,y)=x^2-y^2 at 0, the limsup is +1 along y=0, while along the y direction the ratio is -1. The proof of Lemma 2 replaces the definition with the per-direction condition L((X+t Xdot)(X+t Xdot)^*) - L(XX^*) ≥ o(t^2) for every fixed unit Xdot, i.e. liminf_{t↓0} of the second-order difference along every ray is nonnegative. No argument in §3.4 justifies this replacement. Since Lemma 2 is used to prove Lemma 1 and hence Theorems 1–3 and the nonsmooth branch of Theorem 5, the central landscape results are not proved as stated. The fix is to define SOCP by requiring the liminf to be nonnegative along every unit ray (equivale","section":"§3 (definition of SOCP) and §3.4 (proof of Lemma 2)"},{"comment":"The statement of Theorem 6 imposes no condition on the optimization rank p. The proof, however, explicitly uses p≥c (after equation (47)) and, in the real case, Theorem 2 is only valid for p≥2 because the denominator c_F p − 1 vanishes at p=1. Thus Theorem 6 as written is not valid for p=1 in the real case, and the current proof does not cover arbitrary p. The theorem statement should add a hypothesis such as 'for all p≥c3' (with c3 independent of n,d), consistent with Theorem 5 and with the proof.","section":"§5.2, Theorem 6"}],"minor_comments":[{"comment":"The typeset formula contains a stray 'ww' and is unreadable; please correct it.","section":"§3.4, displayed equation after (18)"},{"comment":"The display 'min_{u∈F^n_1} 1/n ...' contains an extra '1' in the LaTeX source and should be cleaned up.","section":"§4, after the PhaseCut reformulation"},{"comment":"Typo: 'we gives' should be 'we give'.","section":"§5, first paragraph"},{"comment":"The lemma asserts a bound for 'any X that is a first-order critical point of (13)', but first-order criticality of the nonsmooth objective is never defined. The proof uses ∇L_λ(XX^*)X=0, which is a conclusion of Lemma 2 for second-order critical points. Please restate Lemma 8 for the class of points actually used (those satisfying the criticality conditions of Lemma 2) or define the intended first-order notion.","section":"Lemma 8"},{"comment":"Typo: 'in the regimen≲dlogd' should be 'in the regime n≲d log d'.","section":"§2"}],"recommendation":"major_revision","confidential_remarks":"The main gap appears to be a quantifier/definition error rather than a fundamental flaw in the landscape strategy; the PhaseCut branch and the statistical machinery are plausible once the definition is repaired. The missing p condition in Theorem 6 is a simple but necessary correction. I recommend major revision: the author should fix the definition of second-order critical point, revisit all statements that depend on it, and add the missing rank hypothesis to Theorem 6."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is the first global landscape analysis of the nonsmooth amplitude loss (7)/(8) for phase retrieval, and it does a fair amount of useful work: a deterministic landscape lemma, a PhaseCut/synchronization reformulation that gives essentially the same bounds, constant-overparametrization statistical guarantees with quasi-optimal sample complexity under sub-Gaussian measurements, and an infinite-dimensional Hilbert-space extension. The PhaseCut reformulation and the way the proof avoids the logarithmic factors of the quartic loss are genuinely new and worth knowing.\n\nThe catch is in the definition of second-order critical point. The paper defines it via a limsup over all nearby points. That condition does not rule out quadratic descent along a specific direction. The proof of Lemma 2 silently replaces it with a liminf along every unit ray, which is a strictly stronger condition. As printed, Lemma 2 is false: with d=p=2, A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1), the limsup condition holds but the first-order stationarity conclusion fails. Since Lemma 2 feeds into Lemma 1 and Theorems 1–3, and Theorem 5's nonsmooth part rests on those, the main landscape theorem is not proven as stated. The likely fix is straightforward—define a second-order critical point by the per-direction liminf—and I'd expect the proofs to go through with that change, but as it stands the central claim is unsupported.\n\nThe paper also openly notes that no algorithm is proven to converge to these critical points for the nonsmooth objective; the smooth PhaseCut formulation is the only algorithmic bridge. That's a real limitation, and I'd want the revision to address it or at least frame the claims accordingly.\n\nThe statistical results are stated cleanly and the deterministic inequalities are parameter-free; the reliance on [34] for the phase-retrieval lower bound is standard and clearly cited. The noise-to-error tradeoff discussion is honest.\n\nBottom line: this is a serious piece of work with a fixable but load-bearing flaw. I'd send it to peer review, and in the revision I'd ask for a corrected SOCP definition, a repaired Lemma 2, and a careful statement of what is proved for the nonsmooth problem versus the smooth reformulation.","headline":"First landscape analysis of the nonsmooth amplitude loss with constant overparametrization, but a quantifier error in the SOCP definition undercuts the main theorems as printed.","tokens_in":30870,"tokens_out":2184,"would_cite":true,"duration_ms":20629,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C26","49J52","90C46"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes that the nonsmooth amplitude least-squares loss for phase retrieval has a benign landscape under constant overparametrization: every second-order critical point is a statistically optimal estimate, achieving noise-flo","keywords":["phase retrieval","nonconvex optimization","overparametrization","nonsmooth amplitude loss","second-order critical points","low-rank matrix sensing","sub-Gaussian measurements","PhaseCut"],"falsifier":"Run the two-line check on Lemma 2 using the amplitude loss ℓ(b,y)=(√b−y)² with d=p=2, A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1). This point satisfies the printed limsup definition of a second-order critical point, yet ⟨A1,XX*⟩=0 gives ℓ′(0,1)=−∞, which contradicts the first claim of Lemma 2 and violates the quadratic-ascent inequality used throughout. Reproducing this computation would settle whether the main results hold with the definition as printed.","tokens_in":29821,"feed_emoji":"📡","tokens_out":9724,"duration_ms":84094,"temperature":0.7,"pith_summary":"The paper studies a nonconvex, nonsmooth phase-retrieval loss — minimize the average squared difference between measured amplitudes |⟨fi,x⟩| and observations — but over a rank-p variable X with p larger than the true rank, an overparametrization. It proves that, for isotropic sub-Gaussian measurements and n at least a constant multiple of d, every second-order critical point of this amplitude loss lies close to the true signal: the recovery error is bounded by the statistical noise floor, ∥ε∥/√n, with only constant overparametrization. The same guarantee holds for a nonconvex PhaseCut reformulation as a smooth synchronization problem over spheres, and a parallel result covers Gaussian measurements in any Hilbert space with ridge regularization. A sympathetic reader would care because the amplitude loss is the one practitioners prefer, and this shows its global nonconvex landscape is benign with minimal overparametrization, matching guarantees that previously required semidefinite programming or logarithmic rank inflation.","feed_headline":"Constant overparametrization: noise-optimal nonconvex phase retrieval","feed_subtitle":"Second-order critical points of the amplitude loss recover x* up to the noise floor with n≈d measurements.","key_machinery":"The engine is Lemma 1, a deterministic second-order-criticality inequality for factored losses Lλ(XX*) = (1/n)∑ ℓ(⟨Ai,XX*⟩,y_i) + λ tr(XX*) with PSD measurement matrices Ai and convex ℓ. It shows that any second-order critical point X must satisfy ∇Lλ(XX*)X = 0 and, for every PSD Z′, 0 ≤ ⟨∇Lλ(XX*), Z′⟩ + (2/(c_F p))∑_{⟨Ai,XX*⟩>0} ℓ″(⟨Ai,XX*⟩,y_i)⟨Ai,XX*⟩⟨Ai,Z′⟩, where c_R=1 and c_C=2. The rank p sits in the denominator of the curvature penalty, which is exactly why overparametrization weakens the nonconvex obstruction and brings the condition close to convex SDP optimality. For the amplitude loss ℓ(b,y)=(√b−y)², the curvature term collapses to a controlled form, yielding Theorem 2; the Phase","core_discovery":"The central claim is Theorem 5: under the measurement model y_i = |⟨f_i,x*⟩| + ε_i with isotropic sub-Gaussian measurement vectors satisfying mild moment and low-coherence conditions, if n ≥ c1 d then with probability at least 1 − c2 n^{−2}, for every p ≥ c3, every second-order critical point X of the nonsmooth amplitude problem (1/2n)∑(√⟨Ai,XX*⟩ − y_i)² satisfies ∥XX* − x*x*∥_* ≤ c4(∥x*∥∥ε∥/√n + ∥ε∥²/n), and the nearest rank-1 factor x̂ satisfies min_{|s|=1}∥x̂ − sx*∥ ≤ c5∥ε∥/√n. The identical conclusion holds when X = F† diag(y) U is the solution map of the PhaseCut formulation evaluated at any second-order critical point U. The theorem says there are no spurious second-order critical poin","pith_inferences":["If the stronger second-order-criticality interpretation is retained, the amplitude loss's constant-overparametrization success suggests smoothing the loss (as several nearby works do) may be unnecessary for global landscape guarantees; the unsmoothed objective is already benign.","Because the deterministic landscape lemma is separated from the statistical concentration, the same proof template should transfer to other measurement models — coded diffraction patterns, Poisson noise, or non-isotropic covariances — once a stable phaseless injectivity bound is available.","The infinite-dimensional result hints that phase retrieval in an RKHS could be solved by the same regularized factored formulation with feature-space inner products, requiring only O(n²) kernel-matrix operations rather than explicit high-dimensional vectors.","In practice, constant p means the factorized algorithm stores O(np) measurement-phase variables instead of O(d log d), which could make the amplitude loss the method of choice for large-scale imaging if the constant c3 is small; the paper does not compute c3 numerically."],"forward_implications":["Constant-overparametrization nonconvex optimization of the amplitude loss achieves order-optimal sample complexity n=O(d) for sub-Gaussian phase retrieval, matching semidefinite programming guarantees while optimizing over roughly d×p instead of d×d variables.","The nonconvex PhaseCut formulation inherits the same landscape guarantee, so a smooth sphere-synchronization problem has no bad second-order critical points either.","Ridge-regularized amplitude loss is provably statistically optimal in infinite-dimensional Gaussian phase retrieval, with the usual linear-regression bias term; this is a step toward nonparametric phase retrieval in reproducing kernel Hilbert spaces.","The deterministic inequality generalizes beyond amplitude loss, giving landscape control for quartic and Poisson losses (Theorems 1 and 3) in the same factored-matrix framework."],"fun_headline_variants":["Constant overparametrization: optimal noise floor in phase retrieval","No spurious critical points with constant overparametrization","Overparametrized amplitude loss: recovery up to noise floor","n≈d suffices: overparametrized nonconvex phase retrieval"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that every second-order critical point of the nonsmooth loss satisfies the quadratic-ascent inequality liminf_{t↓0}[f(X+tẊ)−f(X)]/t² ≥ 0 in every direction Ẋ, as used in Lemma 2; the paper's printed definition of second-order critical point uses a weaker limsup, and under the printed definition Lemma 2 is false, so the main theorems depend on the stronger interpretation being the intended one.","fun_headline_variants_meta":{"raw":{"variants":["Constant overparametrization: optimal noise floor in phase retrieval","No spurious critical points with constant overparametrization","Overparametrized amplitude loss: recovery up to noise floor","n≈d suffices: overparametrized nonconvex phase retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2529,"prompt_tokens":844,"completion_tokens":1685,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":588,"completion_tokens_details":{"reasoning_tokens":1610}},"tokens_in":588,"tokens_out":1685,"duration_ms":13125,"temperature":1.0,"reasoning_tokens":1610,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:36:42.432319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the two-line check on Lemma 2 using the amplitude loss ℓ(b,y)=(√b−y)² with d=p=2, A1=diag(1,0), A2=diag(0,1), y=(1,1), X=(0,1). This point satisfies the printed limsup definition of a second-order critical point, yet ⟨A1,XX*⟩=0 gives ℓ′(0,1)=−∞, which contradicts the first claim of Lemma 2 and violates the quadratic-ascent inequality used throughout. Reproducing this computation would settle whether the main results hold with the definition as printed.","supporting_citations":[],"review_version":1}