{"id":"a042d099-4b58-47ef-9af6-70ef2a0d639a","arxiv_id":"2506.18282","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"For rank d phase retrieval with Gaussian measurements, descending gradient algorithms are predicted to succeed above a sample complexity ratio near 2.79 for d=2, with lifted bounds lowering this estimate and simulations showing rough agreement.","lead":"This paper extends Random Duality Theory to phase retrieval with rank d measurements, predicting sample-complexity thresholds above which gradient-type algorithms succeed. Simulations for the rank 2 (complex) case roughly match the predicted threshold, but the theory is bound-based and the comparison uses a hand-set adjustment.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"α≈2.79 rests on a strict lower-bound landscape, and the step from that landscape to actual dPR convergence is imported from [118], not proved here.","rationale":"The reader's weakest assumption identifies the unproved link between the fd-pro landscape and actual descending-algorithm convergence; that is indeed a load-bearing gap and I agree it warrants a conditional verdict. My read adds a second, compounding issue: the predicted α≈2.79 is computed from a strict lower-bound curve, and the paper itself shows through lifted RDT that lower values of α become admissible, so the numerical location is method-dependent rather than a determined phase transition of dPR. I do not think this demands rejection: the RDT derivation is systematic, the lower-bound nature is acknowledged in Section 2.1, the lifted-RDT extension is a reasonable attempt to tighten it, and the small-scale simulation shows a transition in the expected region. However, the abstract's phrasing that phase transition locations are 'determined' for both plain and lifted RDT overstates what is proved: no lifted-RDT critical α is actually computed, and the plain-RDT value is an upper-bound-based estimate. The requested revision should therefore prove or clearly label the thresholds as estimates, justify the landscape-to-algorithm correspondence for the actual gradbar objective, and release code and experimental details so the transition can be checked at larger n. These are exactly the conditions under which the paper's central claim would be acceptable, so the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":25696,"tokens_out":10267,"duration_ms":114322,"concrete_test":"Run the authors' gradbar (or a faithful reimplementation) for d=2 at n=200 and n=400, with both the spectral initializer of Eq. (52) and a random initializer, at α ∈ {2.2, 2.3, 2.5, 2.6, 2.79, 3.0, 3.2}, using at least 100 random instances per point and recording success probability and final overlap. If the empirical transition shifts by more than roughly 0.1-0.2 in α with n, or if the transition depends strongly on the initializer, then the landscape-to-algorithm correspondence and the claimed α≈2.79 location are not supported. As an analytical complement, compute the lifted-RDT curve of Section 3 at α=2.5, 2.6, and 2.7; if the lifted lower bound is already positive for all x≠1 below 2.79, the plain-RDT value is only an upper-bound estimate, not the determined phase transition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central number α≈2.79 is obtained in Section 2.2 from the plain-RDT lower bound φ0 on the fd-pro objective ξ(c,x), not from ξ itself. Section 2.1 explicitly states that strong random duality is not in place and that the RDT results are strict lower bounds; Section 3 then shows that lifted RDT lowers the relevant α to about 2.5 for the same landscape criterion. Thus the α at which the lower-bound curve loses its x=0 minimum is only an upper estimate for the true landscape threshold, and calling it \"the dPR phase transitioning sample complexity ratio\" is not supported by the derivation. Separately, the inference from \"ξ(c,x) has a single minimum\" to \"descending algorithms converge\" is taken from the companion paper [118] without a derivation for the actual gradbar dynamics. The simulation does not close this gap because gradbar minimizes a different objective, the squared-magnitude log-barrier objective fbar in Eqs. (49)-(50), whereas the theory analyzes the amplitude-based fd-pro objective. The finite-n 'safer compression' adjustment is hand-set at 10-20%, so the agreement in Figure 6 is evidence that some transition occurs near the adjusted prediction, but it does not establish the specific location α≈2.79 or that the landscape condition alone governs algorithm success.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends the author's random duality theory (RDT) program for descending phase retrieval algorithms (dPR) to rank d positive semidefinite measurements, with d = 2 treated as an emulation of complex phase retrieval. The author derives a lower bound on the scaled fundamental dPR objective ξ(c, x) via plain RDT (Section 2.1), observes that the lower-bound curve loses its spurious x = 0 minimum at α ≈ 2.79 (Section 2.2), develops a partially lifted RDT version that lowers the operating estimate to about α = 2.5 (Section 3), and reports simulations of a log-barrier gradient algorithm (gradbar) at n = 100 showing a transition near a \"safer compression\" adjusted value (Section 4, Figure 6). The abstract claims that phase transition locations are determined for both plain and lifted RDT.","tokens_in":25904,"tokens_out":4612,"duration_ms":45955,"significance":"If the quantitative phase-transition predictions were established, the paper would provide a useful design rule for practitioners and a rare precise statistical characterization of descending algorithms for rank d phase retrieval. Strengths include the explicit closed-form evaluation of f_q for d = 2 in Eq. (37), the step-by-step adaptation of the RDT machinery to the rank d setting, the recognition that the results are lower bounds, and a concrete, reproducible simulation protocol (gradbar with spectral initialization). However, the central threshold α ≈ 2.79 is inferred from a strict lower-bound curve, not from the actual objective, and the landscape-to-algorithm link is imported from the author's companion paper [118] rather than derived here. The significance is therefore conditional on these gaps being closed or the claims being appropriately weakened.","major_comments":[{"comment":"The claimed phase transition at α ≈ 2.79 is read off from the disappearance of the x = 0 minimum of the plain-RDT lower-bound curve √φ0, not from the actual fd-pro objective ξ(c, x). Section 2.1, step 4 explicitly states that strong random duality is not in place and that these results are strict lower bounds. A lower-bound curve losing a spurious minimum is only a sufficient condition for the landscape condition on the true objective; ξ(c, x) may lose its bad minimum at a smaller α. Consequently, the sentence \"RDT predicts the dPR phase transitioning sample complexity ratio to be α ≈ 2.79\" is not supported by the derivation. The manuscript should either present 2.79 as an upper estimate/guarantee threshold or provide a concrete way to quantify the gap (for example, by comparing the lifted threshold with a simulation using the amplitude-based objective).","section":"Section 2.2, Figure 1, Eq. (30)"},{"comment":"The inference from \"ξ(c, x) has a single minimum\" to \"descending algorithms converge\" is taken from the companion paper [118] and is not derived for the gradbar dynamics used in the simulations. Moreover, gradbar minimizes fbar (Eq. (49)) built on the squared-magnitude loss fplain (Eq. (50)), whereas the theory analyzes the amplitude-based fd-pro objective ξ(c, x). The agreement in Figure 6 therefore tests a different objective and a different algorithm than the one analyzed, so it cannot directly validate the predicted landscape condition or the specific value α ≈ 2.79. The authors should either supply a derivation of the landscape-to-algorithm correspondence for the actual dynamics or explicitly frame the simulation as a heuristic consistency check and state the objective mismatch in Section 4.","section":"Sections 2.2 and 4, Eqs. (49)–(50)"},{"comment":"The abstract claims that \"for both plain and lifted RDT we determine phase transitions locations,\" but the lifted analysis does not actually determine a threshold. Equation (48) is again a lower bound (via Theorem 2 and Eq. (39)), and the text selects α = 2.5 as a convenient operating point, states that \"there is really not much point in doing so\" for the limiting transition, and only loosely associates a transition \"around 2.3\" with the spectral initializer's overlap. To support the abstract claim, the manuscript should give a precise characterization of the lifted threshold or explicitly restrict the claim to the plain-RDT lower-bound estimate.","section":"Section 3, Eq. (48), Figures 3–5"}],"minor_comments":[{"comment":"The phrase \"Wirt inger flows\" contains a spacing typo and should be \"Wirtinger flows.\"","section":"Abstract"},{"comment":"The dependence of the main claims on the unpublished companion paper [118] should be stated more prominently, since the landscape-to-algorithm step and parts of the random-dual machinery are not re-derived here.","section":"Section 1 and throughout"},{"comment":"The constraint \"xT ¯x = x\" in Eq. (13) is confusing because the same symbol x denotes both the vector and the scalar overlap; the later notation x1 = x in Eq. (15) should be introduced earlier.","section":"Section 2, Eq. (13)"},{"comment":"Figure 6 would be more informative if the empirical transition point were reported numerically and compared explicitly with the plain-RDT, lifted-RDT, and safer-compression values, since the current text describes the agreement only qualitatively.","section":"Section 4, Figure 6"},{"comment":"The strict-lower-bound nature of the results is stated in the body, but the abstract and conclusion present the phase-transition locations as determined; adding a caveat there would accurately represent the strength of the results.","section":"Section 2.1, step 4, and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is an incremental extension of the author's own RDT line and leans heavily on the companion manuscript [118], which is not yet peer-reviewed; the editor may wish to ensure that [118] is available or under review before relying on this work. The core difficulty is that the announced phase-transition locations are supported only as lower-bound-derived estimates, so the manuscript needs either a significant weakening of the claims or a substantive additional comparison (e.g., amplitude-loss simulations or a rigorous bound on the lower-bound gap) before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a genuine extension of the companion RDT program from rank-1 to rank-d phase retrieval, with new explicit formulas for d=2. But the advertised α≈2.79 should be read as a lower-bound-derived landscape estimate, not an established phase transition threshold. The simulation is supportive but not sharp enough to fix the exact location.\n\nWhat the paper does well: the rank-d generalization is not in the cited prior literature; the d=2 Laguerre/Bessel expressions are new and concrete; the derivation is systematic and follows a documented methodology; and the lifted RDT section is a real improvement, giving a lower α around 2.5. The author is also candid about the lack of strong random duality and about finite-n jitteriness. No fitted constants enter the core φ0 formulas, so the central derivation is not circular in the strong sense.\n\nThe main soft spot is exactly where the stress-test note lands. Section 2.1 explicitly says plain RDT gives strict lower bounds on ξ(c,x), and Section 3 shows lifted RDT lowering the relevant α to about 2.5. So 'the dPR phase transitioning sample complexity ratio is α ≈ 2.79' overstates what the derivation supports: 2.79 is where the lower-bound curve loses its x=0 minimum, not the established location of the true landscape transition. The lifted α≈2.5 is also a partially lifted lower bound, not an exact threshold. These should be labeled as estimates from below, with the fully lifted RDT as the natural route to a definitive value.\n\nThe second soft spot is the landscape-to-algorithm step. The claim that a single minimum of ξ(c,x) at (c=1,x=1) makes descending algorithms succeed is imported from [118], not derived for the gradbar dynamics used here. The simulation partially covers this gap, but it minimizes a squared-magnitude log-barrier objective while the theory is for amplitude loss, and the 'safer compression' adjustment is hand-set at 10–20%. Figure 6 is consistent with a transition near the adjusted prediction, but it is not a sharp test of α≈2.79. No code or detailed protocol is given, which matters for a paper whose main evidence beyond the derivation is one simulation curve.\n\nThe citation pattern is heavily self-referential, but that is not a flaw here: the methodology genuinely comes from [118], and the author credits that dependence clearly. The paper is honestly written about its own limitations, which makes the few overstatements easy to isolate.\n\nWho gets value from this: researchers working on nonconvex phase retrieval and on RDT-style lower bounds, especially anyone wanting explicit rank-d formulas or a template for extending RDT to other rank-d problems. It deserves a serious referee. The main revision requests should be: clarify that the thresholds are lower-bound estimates, reconcile or justify the amplitude-versus-squared-magnitude mismatch, and release code and full experimental details. I would engage with it as a conditional paper, not a reject.","headline":"A real rank-d generalization of the author's RDT phase-transition program with new explicit d=2 formulas, but the headline thresholds are lower-bound estimates and the landscape-to-algorithm link is imported rather than proved here.","tokens_in":26460,"tokens_out":2301,"would_cite":false,"duration_ms":25457,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","90C26"],"pacs":[],"model":"deepseek-v4-flash","headline":"For rank-2 phase retrieval, descending algorithms hit a sharp success threshold at 2.79 measurements per unknown.","keywords":["phase retrieval","rank d measurements","descending algorithms","random duality theory","phase transition","Wirtinger flow","sample complexity ratio","complex phase retrieval"],"falsifier":"At $n=1000$ with Gaussian rank-2 measurements, run a descending solver (for instance the paper's log-barrier gradient or a Wirtinger flow) with optimal diagonal spectral initialization over at least 100 random trials per value of $\\alpha$, sweeping $\\alpha$ from 2.2 to 3.0 in steps of 0.05. If the empirical success probability transitions far from the predicted plain-RDT value $\\alpha \\approx 2.79$ (or from the $\\approx 2.3$--$2.5$ lifted range), or if a second local minimum of $\\xi(1,x)$ survives at $\\alpha > 2.79$, the landscape-to-algorithm prediction fails.","tokens_in":25446,"feed_emoji":"📉","tokens_out":15670,"duration_ms":138996,"temperature":0.7,"pith_summary":"This paper generalizes the companion rank-1 random duality theory (RDT) program to phase retrieval with rank-$d$ positive-semidefinite measurements, and uses it to predict when descending algorithms--gradient-descent-type solvers such as Wirtinger flow--succeed or fail in the high-dimensional limit. The central object is the fundamental RdM PR optimization $\\xi(c,x)$; the paper computes a random-dual lower bound $\\varphi_0$ whose shape, as a function of the overlap $x$, decides whether the objective has one minimum or two. For the practically important rank-2 case, which emulates complex phase retrieval, plain RDT predicts the phase transition at sample complexity ratio $\\alpha \\approx 2.79$, while lifted RDT lowers the estimate to roughly $2.3$--$2.5$ depending on initializer quality. A log-barrier gradient descent experiment at dimension $n=100$ finds transition points close to the safer compression adjusted theory. The result matters because it condenses the question of whether a descending algorithm works into a single threshold $\\alpha$ for each rank $d$.","feed_headline":"Descending phase retrieval turns on at 2.79 measurements per unknown","feed_subtitle":"For rank-2 complex-like measurements, gradient-type solvers cross a sharp success/failure threshold.","key_machinery":"The load-bearing object is the fundamental RdM PR optimization (fd-pro), $\\xi(c,x)$, defined as the minimum over $x$ and auxiliary $z$ of the sum of squared differences between measured and candidate amplitude roots, subject to $Ax=z$, $x^T\\bar{x}=x$, and $\\|x\\|_2^2=c$. Random duality theory replaces the random primal with a random dual whose expected value $\\varphi_0$ the paper evaluates via non-central chi distributions; for $d=2$ the evaluation reduces to integrals of modified Bessel functions $I_0$ and $I_1$. The predictive instrument is the curve $\\sqrt{\\varphi_0}$ as a function of the overlap $x$ at $c=1$: a second minimum at $x=0$ indicates that a descending algorithm can be trapped, and its disappearance marks the phase transition. Lifted RDT refines the bound with a partially lifted dual and a large-deviation functional, lowering the predicted transition. The safer compression adjustment is a finite-dimensional heuristic that moves the working $\\alpha$ about 10--20% above the asymptotic threshold to avoid local jitteriness.","core_discovery":"The paper claims that for rank-$d$ measurements $B^{(i)} = \\sum_{j=0}^{d-1} A_{jm+i,:}^T A_{jm+i,:}$ with iid standard Gaussian $A \\in \\mathbb{R}^{dm \\times dn}$, the behavior of descending algorithms is controlled by the fd-pro objective $\\xi(c,x)$: when the random-dual limit $\\varphi_0$ is positive, $\\xi(c,x)/(dn)>0$ for $x \\neq 1$ with probability tending to one, and the phase retrieval problem is uniquely solvable up to global phase. It then identifies the dPR phase transition with the disappearance of the secondary minimum of $\\sqrt{\\varphi_0}$ at $x=0$ for $c=1$: curves for $\\alpha=2.4$ and $\\alpha=2.6$ have that minimum, while around $\\alpha \\approx 2.79$ the curve flattens and the minimum disappears. Because strong random duality is not in place, the plain RDT estimates are strictly lower bounds, and a lifted RDT version--using the partially lifted dual--produces decreasing curves already at $\\alpha=2.5$ and, with optimal diagonal spectral initializers, a transition near $\\alpha=2.3$. The paper reports simulations at $n=100$ using a log-barrier gradient descent with spectral initialization; the empirical transition falls close to the safer compression adjusted theoretical prediction.","pith_inferences":["A direct test would run unconstrained Wirtinger flow (not the constrained log-barrier variant) on the same rank-2 Gaussian model: the paper's landscape-to-algorithm link suggests the same thresholds near 2.3--2.79 should hold, but that is not derived for unconstrained dynamics in this paper.","If the phase transition depends monotonically on $d$, the required $\\alpha$ should shrink as $d$ grows because each measurement carries $d$ independent Gaussian rows; computing the rank-3 and rank-4 transitions from the same Bessel/Laguerre integrals would settle this and is not done in the paper.","The Gaussian rotational invariance is central to the argument, so non-Gaussian or orthogonal measurement ensembles should shift or smear the transition; quantifying the shift would delimit how universal the predicted thresholds are.","The safer compression adjustment is an asymptotic-to-finite heuristic; a sharper finite-$n$ analysis could replace it by estimating the probability that local jitteriness creates traps as a function of $n$ and $\\alpha$."],"forward_implications":["Above the predicted threshold, descending phase retrieval algorithms succeed with probability tending to one on Gaussian rank-$d$ measurements; below it, the objective's second minimum at $x=0$ means badly initialized descent can be trapped.","For the rank-2 case that emulates complex phase retrieval, plain RDT gives $\\alpha \\approx 2.79$ as the dPR phase transition, and lifted RDT lowers the usable transition to about $\\alpha=2.5$ for the decreasing-curve regime and $\\alpha \\approx 2.3$ with optimal diagonal spectral initializers.","Because the same integrals extend to any rank $d$, the framework yields explicit phase-transition predictions for higher-rank measurements, with $d=1$ and $d=2$ recovering the real and complex phase retrieval scenarios.","In finite dimensions the safer compression rule recommends operating 10--20% above the asymptotic threshold; the $n=100$ simulations show the empirical transition near this adjusted value rather than at the raw asymptotic point.","The lack of strong random duality means the plain RDT numbers are strict lower bounds, so the lifted RDT estimates, not the plain ones, are the ones to use for predicting actual algorithm performance."],"supporting_citations":[{"why":"Supplies the companion rank-1 RDT program: the fd-pro landscape-to-algorithm link, the random-dual derivation, and the phase-transition methodology that this paper extends to rank-d measurements.","marker":"[118]"},{"why":"Introduces Wirtinger flow as a successful gradient-type phase retrieval solver, the prototypical descending algorithm whose high-dimensional behavior the paper characterizes.","marker":"[24]"},{"why":"Analyzes optimal spectral initializers and their impact on phase retrieval phase transitions; the paper uses it to place initializers on the right downside of the lifted curves.","marker":"[117]"},{"why":"Establishes the optimal diagonal spectral initializer for any oversampling ratio; this is the initialization scheme used in the rank-2 simulations.","marker":"[76]"},{"why":"Defines the weak threshold for diagonal spectral initializers and supplies the spectral initialization theory that the numerical experiments rely on.","marker":"[89]"},{"why":"Lays out the four RDT principles (algebraic representation, random dual, handling, strong-duality check) that the paper adapts from rank 1 to rank d.","marker":"[111]"},{"why":"Provides the comparison theorems for Gaussian processes used to prove the random-dual lower bounds in Theorems 1 and 2.","marker":"[112, 113]"},{"why":"Develops partially lifted RDT, the mechanism the paper uses to improve the plain RDT phase-transition estimate.","marker":"[109, 110, 115]"}],"fun_headline_variants":["Phase transition at 2.79 measurements for descending phase retrieval","Rank-2 measurements give sharp threshold at 2.79 per unknown","Descending algorithms flip at 2.79 for complex-like measurements","2.79: critical measurement ratio for descending phase retrieval"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the landscape of the fd-pro objective--specifically whether $\\xi(c,x)$ or its RDT lower-bound curve has a single minimum at $(c=1,x=1)$--determines whether descending algorithms converge on random instances; this landscape-to-algorithm link is imported from the companion paper and is never derived for the actual gradient dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Phase transition at 2.79 measurements for descending phase retrieval","Rank-2 measurements give sharp threshold at 2.79 per unknown","Descending algorithms flip at 2.79 for complex-like measurements","2.79: critical measurement ratio for descending phase retrieval"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001101,"raw_usage":{"total_tokens":4639,"prompt_tokens":1037,"completion_tokens":3602,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":653,"completion_tokens_details":{"reasoning_tokens":3529}},"tokens_in":653,"tokens_out":3602,"duration_ms":22955,"temperature":1.0,"reasoning_tokens":3529,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:52:29.694165+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"At $n=1000$ with Gaussian rank-2 measurements, run a descending solver (for instance the paper's log-barrier gradient or a Wirtinger flow) with optimal diagonal spectral initialization over at least 100 random trials per value of $\\alpha$, sweeping $\\alpha$ from 2.2 to 3.0 in steps of 0.05. If the empirical success probability transitions far from the predicted plain-RDT value $\\alpha \\approx 2.79$ (or from the $\\approx 2.3$--$2.5$ lifted range), or if a second local minimum of $\\xi(1,x)$ survives at $\\alpha > 2.79$, the landscape-to-algorithm prediction fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Analyzes optimal spectral initializers and their impact on phase retrieval phase transitions; the paper uses it to place initializers on the right downside of the lifted curves."}],"review_version":2}