{"id":"02cf3377-f643-4052-bbaf-3f849069fc52","arxiv_id":"2501.03163","paper_version":3,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For a broad class of single-index models with Gaussian design, the existence of the unregularized M-estimator and the existence of a unique solution to the state-evolution system are both governed by a single explicit threshold delta_infinity.","lead":"This paper proves that in high-dimensional single-index models with Gaussian covariates, the unregularized M-estimator exists exactly when the sample-size-to-dimension ratio exceeds a computable threshold, and that the standard asymptotic equations have a unique solution precisely on the same side of that threshold.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the proof chain for Theorem 2.7 is internally consistent; the C1 loss assumption is explicit and self-acknowledged, not a hidden flaw.","rationale":"The reader's weakest assumption was the C1 strict convexity hypothesis in Assumption 2.5(1), used in Lemma 3.2 to rule out the degenerate minimizer v* = 0. I agree that this is the only delicate point in the proof of Theorem 2.7: the entire if-and-only-if statement passes through the non-degeneracy lemma, and the authors themselves note that a different threshold appears for non-differentiable losses. However, this is not a hidden assumption or a gap. Theorem 2.7 is stated under Assumption 2.5, the proof uses the assumption exactly where needed, and the limitation is acknowledged in the text. Consequently I do not see a load-bearing objection that would change the verdict. I reviewed the algebra in Lemma C.2/C.3 connecting KKT conditions to the nonlinear system, the proximal identity v* = prox[gamma* l_Y](a* U + sigma* G) - a* U, the sign conventions in Lemma E.1, and the coercivity argument in Lemma E.2; all are consistent. The numerical experiments in Figures 2 and 3 are consistent with Theorem 2.6 but are not a substitute for proof; they do not bear on the correctness of Theorem 2.7. The recommendation ACCEPT with moderate confidence remains appropriate.","tokens_in":25687,"tokens_out":25442,"duration_ms":242203,"concrete_test":"Run a numerical continuation for system (2) in the Poisson model of Section 4 with n = 4000 and p = n/delta, for delta in [delta_infinity - 0.02, delta_infinity + 0.02] at several kappa values; compute delta_infinity from (4) and check that a unique solution to (2) is found exactly when delta > delta_infinity. This directly tests the if-and-only-if phase transition of Theorem 2.7, including the boundary case delta = delta_infinity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I checked the central claim (Theorem 2.7) and its proof route: equivalence Theorem 3.1 between system (2) and the infinite-dimensional program (5); non-degeneracy Lemma 3.2; uniqueness Lemma 3.3; and the two-sided phase transition in Theorem 3.4 via Lemmas E.1 and E.2. I could not find a missing condition or a circular step. The one place where the argument is genuinely delicate is Lemma 3.2, where v* != 0 is proved by a contradiction using the Moreau envelope and the C1 assumption on the loss. This is exactly the assumption flagged by the authors: for non-differentiable losses a different threshold delta_perfect emerges (Bellec and Koriyama, 2023). Because Assumption 2.5(1) is stated up front and the limitation is explicitly acknowledged in the text, this is a scope restriction rather than an internal inconsistency. Minor presentation issues (equation-numbering inconsistencies, numerics without code or error bars) do not touch the theorem. The strongest claim appears sound under its stated assumptions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies unregularized M-estimators in single-index models with Gaussian covariates under proportional asymptotics n/p → δ ∈ (1, ∞). It defines a threshold δ∞ as the infimum of an explicit convex functional φ(t) of the data distribution and proves two main results. Theorem 2.6 establishes a phase transition for the existence of the M-estimator at δ∞, generalizing Candès & Sur (2020) beyond binary logistic regression. Theorem 2.7 proves that the asymptotic nonlinear system (2) of Sur & Candès has a unique solution if and only if δ > δ∞, closing a gap noted in the literature. The proof proceeds through an equivalence (Theorem 3.1) between system (2) and an infinite-dimensional convex program (5), a non-degeneracy lemma (Lemma 3.2), a uniqueness lemma (Lemma 3.3), and a phase-transition theorem for the convex program (Theorem 3.4).","tokens_in":25729,"tokens_out":19875,"duration_ms":160486,"significance":"If correct, the results provide a comprehensive theoretical foundation for proportional-asymptotic analyses of M-estimators that assume existence of a solution to the asymptotic system. The threshold δ∞ is given by an explicit, parameter-free convex program, and both theorems are proved with complete arguments in the appendices. The main results are numerically verified for Poisson and binomial losses. The paper is careful about its scope: the smoothness assumption on the loss (Assumption 2.5(1)) is stated up front, and the authors explicitly note that the non-degeneracy conclusion changes for non-differentiable losses (Bellec & Koriyama, 2023). The central claim is internally consistent under the stated assumptions.","major_comments":[],"minor_comments":[{"comment":"The sentence referring to 'the nonlinear system (4)' should refer to 'the nonlinear system (2)'.","section":"Section 3, paragraph after Theorem 2.6"},{"comment":"The displayed bound '∥v − ṽ∥ ≤ C^(2)(ξ)(1 + ∥v∥²)' should use a linear term '1 + ∥v∥' rather than '1 + ∥v∥²'. The subsequent inequalities (28)–(29) require the linear version, and the proof's earlier bound on |a| is indeed linear in ∥v∥.","section":"Appendix E, Lemma E.2"},{"comment":"The computation of E[p_*²] contains a typographical clutter with repeated terms 'E[p_*²] + 2t_* E[p_*U] E[p_*U] = 0'; this should be simplified to a clean derivation.","section":"Appendix A, Lemma A.1 proof"},{"comment":"The phrase 'using linear programming' for the Poisson loss is potentially confusing because the loss is exponential; presumably the authors solve the linear feasibility problem (12) to detect existence, and this should be stated explicitly.","section":"Section 4, Numerical Simulation"},{"comment":"The figures would be more informative if they included error bars or standard errors over the 20 repetitions.","section":"Section 4, Figures 2 and 3"},{"comment":"The theorem does not address the boundary case δ = δ∞; a sentence noting that this case is not covered (and why) would improve precision.","section":"Theorem 2.6"}],"recommendation":"minor_revision","confidential_remarks":"This is a solid paper. The central theorems are correct under the stated assumptions, and the proof strategy is well executed. The only issues I found are typographical and precision issues in the appendix and figures. The typo in Lemma E.2's displayed bound should be corrected because it appears in a proof, even though the intended linear bound is clear from context."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The two theorems are the real thing. Theorem 2.6 extends the Candès–Sur phase transition for existence of the MLE to Poisson and binomial single-index models, with an explicit threshold delta_infinity. More importantly, Theorem 2.7 proves that the asymptotic nonlinear system (2) has a unique solution iff delta > delta_infinity, which was known only under the global null for logistic regression. That closes a gap that Sur and Candès left open in their supplement. The proof strategy, via the infinite-dimensional convex program (5) and KKT equivalence, is a genuine adaptation of the Bellec–Koriyama/Montanari et al. machinery, and the structure is transparent: equivalence, non-degeneracy, uniqueness, then the two-sided phase transition.\n\nThe paper is also honest about its boundary. Assumption 2.5(1) requires C^1 strictly convex losses, and Lemma 3.2 uses this to rule out v*=0. The authors explicitly note that for non-differentiable losses a different threshold delta_perfect appears (Bellec–Koriyama 2023). So the iff is not generic, but it is exactly what the assumption states, not a hidden flaw. The other soft spots are minor: equation numbering inconsistencies in the appendix, and the simulations are illustrative (no code, no data, no error bars). The numerics agree with the threshold, which is reassuring but not the point.\n\nThe math looks sound under its stated assumptions. I did not find a circular step: delta_infinity is defined from the data distribution, independent of the system (2), and the paper proves both sides of the equivalence from that same variational object. The proofs are long but complete in the appendix; they are not machine-checked, so there is room for a referee to comb through Lemma 3.2 and Lemma E.2, which are the delicate spots. But the burden is on those lemmas, and they hold up on inspection.\n\nWho is this for? Anyone working on CGMT-based asymptotics for GLMs, high-dimensional M-estimators, or phase transitions in convex programs. It deserves a serious referee. I would send it out.\n\nRecommendation: engage.","headline":"A solid, important paper that closes a real proof gap in high-dimensional M-estimator theory; the C^1 caveat is explicit and does not sink the central theorem.","tokens_in":26396,"tokens_out":1717,"would_cite":true,"duration_ms":14829,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62J12"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that one scalar threshold, $\\delta_\\infty$, simultaneously controls whether an unregularized M-estimator exists and whether the asymptotic system characterizing it has a unique solution, across smooth single-index models.","keywords":["M-estimators","phase transition","single-index models","Gaussian design","logistic regression","Poisson regression","binomial regression","proportional asymptotics"],"falsifier":"Choose the Poisson model with a nonzero signal, compute $\\delta_\\infty$ by minimizing $\\varphi$ from (4), then numerically solve the system (2) over a grid of $\\delta$ values; any solution found for $\\delta < \\delta_\\infty$, or no solution found for $\\delta > \\delta_\\infty$, would refute Theorem 2.7.","tokens_in":25343,"feed_emoji":"📊","tokens_out":7994,"duration_ms":143979,"temperature":0.7,"pith_summary":"The paper establishes that in high-dimensional single-index models with Gaussian covariates and smooth, strictly convex losses, a single scalar threshold $\\delta_\\infty$ controls two separate facts: whether the unregularized M-estimator exists, and whether the nonlinear system that describes its asymptotic behavior has a unique solution. The threshold is given explicitly as the reciprocal of the minimum of a one-dimensional convex function $\\varphi(t)$ that depends only on the law of the response and the link. This unifies and generalizes the known logistic-regression phase transition to models such as Poisson and binomial regression. The proof closes a gap that had previously only been verified numerically outside the global null.","feed_headline":"One formula decides when unregularized M-estimators exist","feed_subtitle":"The threshold $\\delta_\\infty$ separates existence from nonexistence across logistic, binomial, and Poisson single-index models.","key_machinery":"The central object is the infinite-dimensional convex program $\\min_{(a,v)\\in\\mathbb{R}\\times H} \\mathbb{E}[\\ell_Y(aU + v)]$ subject to $\\|v\\| - \\mathbb{E}[vG]/\\sqrt{1-\\delta^{-1}} \\le 0$, where $H$ is the Hilbert space of square-integrable functions of $(G,U,Y)$ and $G$ is an independent standard normal. Theorem 3.1 shows that minimizers of this program correspond one-to-one with solutions of the nonlinear system whenever the Lagrange multiplier is positive, so the phase transition for the system reduces to the existence of this minimizer. The threshold $\\delta_\\infty$ enters through the cone $C$ of directions along which the loss can only be constant or decreasing; the variational representation $\\delta_\\infty^{-1} = \\inf_{(t,p)\\in C} \\mathbb{E}[(G-p)^2]$ yields the direction $p^*$ used to show that no minimizer exists for $\\delta \\le \\delta_\\infty$, while the same $p^*$ gives coercivity for $\\delta > \\delta_\\infty$. The smoothness of the loss is used in Lemma 3.2 to rule out the degenerate minimizer $v^* = 0$.","core_discovery":"For $n/p \\to \\delta$, define $\\delta_\\infty$ by $1/\\delta_\\infty = \\inf_{t\\in\\mathbb{R}} \\varphi(t)$, where $\\varphi$ is a convex function built from the events where the loss is coercive, increasing, or decreasing. Theorem 2.6 states that the unregularized M-estimator exists with probability tending to one when $\\delta > \\delta_\\infty$ and with probability tending to zero when $\\delta < \\delta_\\infty$. Theorem 2.7 states that the nonlinear system of equations characterizing the asymptotic behavior of the estimator has no solution when $\\delta \\le \\delta_\\infty$ and a unique solution when $\\delta > \\delta_\\infty$. The two theorems are proved through an infinite-dimensional convex optimization problem whose minimizer is equivalent to a solution of the system, with the technical core being a non-degeneracy lemma showing that the minimizer does not collapse to $v^* = 0$.","pith_inferences":["The authors do not pursue non-Gaussian designs, but the cone-projection form of $\\delta_\\infty$ suggests an analogous variational formula for elliptical covariate distributions.","A consequence the authors leave implicit is that for smooth losses, existence of the estimator and solvability of its asymptotic system are the same event, which may be a general structural fact rather than a special property of the examples treated.","The remark on $\\delta_{\\text{perfect}}$ invites a concrete follow-up: for nonsmooth losses, test whether the phase transitions for the estimator and for the nonlinear system occur at different thresholds."],"forward_implications":["For Poisson and binomial regression with Gaussian covariates and smooth losses, the threshold $\\delta_\\infty$ from (4) rigorously separates existence from nonexistence of the unregularized M-estimator.","Any asymptotic analysis of these M-estimators based on the Convex Gaussian Minmax Theorem can now invoke Theorem 2.7 to justify the required unique solution of the system whenever $\\delta > \\delta_\\infty$.","The global-null logistic result $\\delta_\\infty = 2$ is contained as a special case, and the threshold is available for non-null single-index models.","In the non-existence regime $\\delta \\le \\delta_\\infty$, the proof exhibits an explicit ray along which the objective decreases, giving a constructive certificate that no minimizer exists."],"supporting_citations":[{"why":"Establishes the M-estimator phase transition at $\\delta_\\infty$ for binary logistic regression, the result this paper generalizes to other single-index models","marker":"Candès & Sur (2020)"},{"why":"Introduces the nonlinear system (2) and reports numeric evidence that solutions exist exactly when $\\delta > \\delta_\\infty$, the gap this paper closes","marker":"Sur & Candès (2019)"},{"why":"Proves the global-null logistic case with threshold $\\delta_\\infty = 2$, the only previous proof of the if-and-only-if system claim","marker":"Sur et al. (2019)"},{"why":"Supplies the Gaussian kinematic formula used to prove the existence phase transition in Theorem 2.6","marker":"Amelunxen et al. (2014)"},{"why":"Provides the Convex Gaussian Minmax Theorem machinery whose prerequisites include existence of a unique solution to system (2)","marker":"Thrampoulidis et al. (2018)"},{"why":"Pioneers the infinite-dimensional optimization technique and the non-degeneracy argument adapted here, and supplies the non-smooth threshold $\\delta_{\\text{perfect}}$","marker":"Bellec & Koriyama (2023)"},{"why":"Provides the KKT and convex-analysis results used to link minimizers of (5) to Lagrange multipliers and system solutions","marker":"Bauschke & Combettes (2017)"},{"why":"Introduces the infinite-dimensional convex program approach to existence of solutions for max-margin systems, adapted here to M-estimators","marker":"Montanari et al. (2023)"}],"fun_headline_variants":["Existence of M-estimators set by one threshold","A single δ∞ decides M-estimator solvability","Sharp phase transition for M-estimator existence","When do M-estimators exist? δ∞ gives the answer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The if-and-only-if result depends on every loss being $C^1$ and strictly convex; the paper itself notes that non-differentiable losses can introduce a different threshold, so if smoothness fails the central claim is not generic.","fun_headline_variants_meta":{"raw":{"variants":["Existence of M-estimators set by one threshold","A single δ∞ decides M-estimator solvability","Sharp phase transition for M-estimator existence","When do M-estimators exist? δ∞ gives the answer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1312,"prompt_tokens":985,"completion_tokens":327,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":259}},"tokens_in":601,"tokens_out":327,"duration_ms":3773,"temperature":1.0,"reasoning_tokens":259,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:52:49.949364+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Choose the Poisson model with a nonzero signal, compute $\\delta_\\infty$ by minimizing $\\varphi$ from (4), then numerically solve the system (2) over a grid of $\\delta$ values; any solution found for $\\delta < \\delta_\\infty$, or no solution found for $\\delta > \\delta_\\infty$, would refute Theorem 2.7.","supporting_citations":[{"cited_title":"B., and Tropp, J","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian kinematic formula used to prove the existence phase transition in Theorem 2.6"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the KKT and convex-analysis results used to link minimizers of (5) to Lagrange multipliers and system solutions"}],"review_version":1}