{"id":"e53b0061-b55d-44ae-90a2-1d729f6778b0","arxiv_id":"1908.04433","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For Gaussian one-bit measurements, the correlation of any convex-loss estimator is sharply predicted by a system of three equations, yielding new per-estimator comparisons and an optimality bound.","lead":"This paper proves sharp asymptotic formulas for how well convex estimators recover a vector from corrupted one-bit measurements in high dimensions. It covers least-squares, LAD, hinge, and logistic losses, and derives a fundamental upper bound on accuracy for any smooth convex loss.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem III.1's key convergence step is deferred: random minimizers of (39) are asserted to converge to deterministic minimizers of (40) without a supplied uniform-convergence or argmin-consistency proof, so the general claim is not yet established.","rationale":"The reader's weakest assumption names exactly the step I also regard as load-bearing: convergence of the random optimizers of (39) to the deterministic optimizers of (40). The paper explicitly defers this as 'it can be shown' and marks uniform convergence as an omitted technical detail, so this is not a manufactured concern. Because the theorem's final correlation formula is fully determined by the limiting values (µ,α), a failure at this step would invalidate the central claim. At the same time, I do not see an independent reason to reject the result: the least-squares special case is verified in closed form, the CGMT framework is standard, and the boundedness and uniqueness assumptions are stated explicitly. The typo in Remark 2 and the conjectural solvability of the equations for hinge/logistic losses are real but secondary issues; they strengthen the need for a conditional verdict rather than a rejection. The appropriate disposition is therefore to keep the reader's CONDITIONAL verdict unchanged until the missing convergence argument is supplied.","tokens_in":16199,"tokens_out":4843,"duration_ms":53561,"concrete_test":"Extract the deferred uniform-convergence lemma used at the (39)-to-(40) step and verify it against the theorem's assumptions: it must cover non-smooth losses, permit a valid compactification of (µ,α,τ,γ), and imply argmin convergence. Independently, evaluate the auxiliary objective (39) by Monte Carlo for the hinge-loss at δ = 4, ε = 0.1, for m = 1000, 5000, and 10000, and compare the computed minimizers with the solution of (40); a systematic non-vanishing discrepancy would refute the step, while agreement would support but not replace the missing proof.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof sketch in Appendix B reduces the estimator to the random scalar min-max problem (39), but the passage from convergence of the objective values to convergence of the minimizers is not proved. The text says 'it can be shown', and Appendix B itself lists 'uniform convergence in going from (39) to (40)' among omitted technical details. This is a substantive gap, not a cosmetic one: pointwise convergence in probability of the objective does not imply convergence of its argmin without uniform/equicontinuity bounds and compactness control. Since Theorem III.1 claims to cover general convex losses, including non-smooth LAD and hinge loss, the missing compactification and uniform-law arguments must hold in that generality. The theorem's stated assumptions (bounded minimizers, unique solution of (9)) do not by themselves supply those arguments. If the convergence of the random optimizers fails, the displayed formula (12) does not follow even when the CGMT setup is valid. Separately, Remark 2's inequality 'unbounded if δ > δε' contradicts Remark 3 and is presumably a typo, and solvability of (9) for hinge and logistic losses remains conjectural; these are secondary but add to the conditional status.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies high-dimensional estimation of a signal from one-bit Gaussian measurements with bit flips. For estimators obtained by minimizing a general convex loss function over linear predictors, it claims an exact asymptotic characterization of the correlation with the true signal in the limit m,n -> infinity with m/n -> delta > 1. The characterization is given by a system of three scalar equations (Eq. (9)) involving the Moreau envelope of the loss. The authors specialize the general result to least squares, least absolute deviations, and hinge loss, provide a closed-form result for least squares that matches the earlier result of Thrampoulidis et al. [29], and derive a universal upper bound on correlation over all continuously differentiable convex losses. Numerical simulations for n=128 are presented as corroboration.","tokens_in":16483,"tokens_out":4970,"duration_ms":52428,"significance":"If the main theorem is fully established, the paper would be a valuable contribution: it extends sharp CGMT-based predictions from least-squares nonlinear measurements to general convex losses and gives a principled way to compare estimators such as LAD and hinge loss, where only order-wise guarantees existed. The explicit closed-form least-squares result and the universal lower bound on sigma are concrete, testable achievements. The paper is honest about several deferred technical steps, but the central claim currently rests on one convergence assertion whose proof is not supplied, and the threshold condition for bounded minimizers contains contradictory statements. The overall framework is sound and the result is likely correct, but the manuscript as written does not yet meet the standard for a full journal proof.","major_comments":[{"comment":"The proof of Theorem III.1 hinges on the assertion that the random optimizers alpha_n and mu_n of the auxiliary problem (39) converge to the deterministic optimizers alpha and mu of the limiting problem (40). The text states \"Based on that, it can be shown\" and defers details to the long version, while Appendix B explicitly lists \"uniform convergence in going from (39) to (40)\" among omitted technical details. This is a load-bearing step: pointwise convergence in probability of the objective functions does not imply convergence of their argmins without uniform/equicontinuity estimates and compactness control. The stated assumptions of Theorem III.1 (bounded set of minimizers of (4), unique solution of (9)) do not, on their own, provide those estimates, and the step must hold for non-smooth losses such as LAD and hinge loss. The gap must be closed by a proof or the theorem must be restated with this convergence as an explicit hypothesis.","section":"Appendix B-C, Eqs. (39)-(40)"},{"comment":"The threshold condition for boundedness of the minimizer set is stated inconsistently. Remark 2 says the minimizer set is \"unbounded if delta > delta_epsilon\", while Remark 3 concludes that for logistic and hinge loss the set is bounded iff delta > delta_epsilon, i.e. unbounded for delta < delta_epsilon. The quotation of [7] around Eq. (15) also has a sign issue: the condition \"1/delta <= delta_epsilon\" is not equivalent to the stated conclusion \"delta < delta_epsilon\". Since boundedness of the minimizer set is an explicit hypothesis of Theorem III.1, the correct threshold and its direction must be stated precisely and, if necessary, proved for the present signed-measurement model rather than transferred from the logistic model in [7]. As written, the applicability of Theorem III.1 to hinge and logistic losses is not established.","section":"Remark 2 / Remark 3, Eq. (15)"},{"comment":"The theorem assumes that the system of equations (9) has a unique solution, but for hinge and logistic losses solvability is left as a conjecture. Consequently, the abstract and introduction claim that the results include hinge-loss and logistic-loss as special cases is stronger than what is proved: for these losses neither the bounded-minimizer condition nor the solvability of (9) is verified. The numerical simulations are suggestive but do not replace a proof. Please either prove the needed conditions for these losses or clearly restate the claims as conditional on them.","section":"Theorem III.1 and Remark 2"}],"minor_comments":[{"comment":"In Remark 5, the text refers to \"the system of non-linear equations in (4)\" and \"(4) is equivalent to v = F(v)\"; these should refer to system (9), not to the optimization problem (4).","section":"Remark 5"},{"comment":"The sentence \"We obtain the hinge-loss estimator in by setting\" is missing a word; it should read \"in this section\" or \"in (4)\".","section":"Section IV-C"},{"comment":"\"In other works\" should be \"In other words\".","section":"Remark 7"},{"comment":"The caption contains the typo \"Numeical\" instead of \"Numerical\".","section":"Figure 5 caption"},{"comment":"\"Cauchy-Schwartz\" should be \"Cauchy-Schwarz\".","section":"Theorem III.2 proof"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads as an extended abstract for a longer paper: the central proof of Theorem III.1 is presented as a sketch, with boundedness, interchange of min and max, and uniform convergence all deferred. For a full math.ST paper, the argmin-consistency argument connecting (39) and (40) should appear in the paper itself rather than in a promised long version. The reliance on the authors' own CGMT exposition [30] is methodologically legitimate and not circular, but the self-containedness of the present manuscript is below what the claims require. The threshold inconsistency in Remarks 2 and 3 should be fixed before the paper can be evaluated for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Henk,\n\nQuick take on 1908.04433. The paper does something genuinely useful: it takes the CGMT machinery that previously handled least-squares one-bit recovery and applies it to arbitrary convex losses. For LAD, hinge, and logistic losses, the paper gives explicit asymptotic correlation predictions and matches them with simulations. The Fisher-information bound in Theorem III.2, which gives a loss-independent ceiling on correlation, is a nice addition and the derivation is straightforward once you accept the main theorem. The LS special case reproduces known results, which is good independent support. The writing is clear about what is a corollary and what is new.\n\nThe soft spot is exactly where the stress-test note lands. The general theorem depends on the assertion that the random minimizers of the auxiliary program (39) converge to the deterministic minimizers of (40). The text says 'it can be shown' and defers details. That is a genuine gap: pointwise convergence of the objective function does not imply convergence of argmins without uniform-convergence and compactness arguments. Since the theorem claims to cover non-smooth losses like LAD and hinge, those arguments have to work in that generality, and the stated assumptions don't automatically provide them. As it stands, formula (12) is conditional on that step.\n\nThere are also smaller issues. Remark 2 has the inequality backwards — it says the minimizer set is unbounded when δ > δε, but Remark 3 and Figure 1 say the opposite. That's a typo, but it should be fixed. And solvability of the equation system for hinge/logistic losses is explicitly conjectural, which is honest but means those sections are not fully theorem-backed.\n\nNone of this makes me think the main result is wrong. The LS case checks out, the equations are plausible, and the simulations are clean. But the manuscript is incomplete in its present form. A referee should ask for the deferred convergence proof, or for the theorem to be stated with that step as an assumption. That is a major revision, not a desk reject.\n\nI'd bring it to the reading group and I'd cite the LS corollary and the overall framework if I worked in this area. Who it's for: people doing sharp asymptotics for signed measurements, one-bit compressed sensing, and CGMT applications. It deserves serious refereeing; the missing proof step is exactly what the referee process is for.","headline":"General convex losses in one-bit recovery: the paper extends the CGMT sharp-analysis program beyond least squares, and the new predictions are probably right, but the main theorem is not fully proved in this manuscript because a key convergence step is deferred.","tokens_in":16953,"tokens_out":2426,"would_cite":true,"duration_ms":23328,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62F12","62J07"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under Gaussian one-bit measurements, every convex-loss estimator's asymptotic correlation is exactly $\\sqrt{1/(1+(\\alpha/\\mu)^2)}$, where $(\\mu,\\alpha,\\lambda)$ solve the three equations (9); the paper also proves a universal upper bound…","keywords":["one-bit measurements","signed measurements","convex loss functions","Gaussian measurement vectors","Moreau envelope","high-dimensional asymptotics","convex Gaussian min-max theorem","universal performance bound"],"falsifier":"Pick a loss in the paper's class, say logistic loss, fix $\\varepsilon=0.1$ and $\\delta=4$, solve (9) numerically to get the predicted correlation, then simulate the estimator at $n=1024$ over many trials; a systematic mismatch between the simulated average correlation and the prediction would refute the formula. The same test can be sharpened by computing the minimizers of the auxiliary problem (39) and checking whether they converge to the solution of (40), the exact deferred step the proof relies on.","tokens_in":16031,"feed_emoji":"🎯","tokens_out":10856,"duration_ms":103532,"temperature":0.7,"pith_summary":"This paper seeks exact, not merely order-of-magnitude, predictions for a basic high-dimensional estimation problem: recovering the direction of a signal from Gaussian inner products that have been reduced to signs and then flipped with probability $\\varepsilon$. The authors prove that for any convex loss function used in the estimator in (4), the limiting correlation between the estimate and the true signal is a deterministic function of the loss, the oversampling ratio $\\delta=m/n$, and the noise level $\\varepsilon$. That function is $\\sqrt{1/(1+(\\alpha/\\mu)^2)}$, where $(\\mu,\\alpha,\\lambda)$ is the unique solution of the three nonlinear equations (9), which involve the loss only through its Moreau envelope. They further prove a universal upper bound on correlation across all continuously differentiable convex losses, so the result also delimits what the whole convex class can achieve. Because the prediction is exact in the limit, it lets practitioners compare estimators such as least squares, least absolute deviations, hinge loss, and logistic loss by solving a small system of equations rather than running large simulations.","feed_headline":"Three equations fix one-bit recovery accuracy for convex losses","feed_subtitle":"For Gaussian one-bit data, the correlation of LS, LAD, hinge, and logistic estimators comes from one small system.","key_machinery":"The load-bearing object is the Moreau envelope of the loss, $M_\\ell(x;\\lambda)=\\min_v \\frac1{2\\lambda}(x-v)^2+\\ell(v)$, which acts as a smoothed surrogate that encodes both the loss and the proximal scale $\\lambda$. The proof reduces the original high-dimensional program to a scalar min-max problem by a convex Gaussian min-max comparison argument; the random optimizers of that auxiliary problem are then asserted to converge to the deterministic optimizers of (40). Taking first-order conditions in that limiting problem produces exactly the system (9), whose unknowns $\\mu,\\alpha,\\lambda$ are the limiting signal bias, fluctuation norm, and proximal scale of the estimator. The Moreau envelope is what makes the reduction work: the inner minimization over coordinates in the auxiliary problem evaluates in closed form as a Moreau envelope, and its derivatives supply the expectation identities in (9).","core_discovery":"On the paper's own terms, the central discovery is that the asymptotic behavior of $\\hat x_\\ell=\\arg\\min_x \\frac1m\\sum_{i=1}^m \\ell(y_i a_i^T x)$ is governed exactly by the scalar system (9): under Gaussian measurements and $m/n\\to\\delta>1$, the absolute correlation converges almost surely to $\\sqrt{1/(1+(\\alpha/\\mu)^2)}$, and $\\|\\hat x_\\ell-\\mu x_0/\\|x_0\\|_2\\|_2^2\\to\\alpha^2$, where $(\\mu,\\alpha,\\lambda)$ is the unique solution of (9). The loss enters the equations only through the Moreau envelope $M_\\ell(x;\\lambda)=\\min_v \\frac1{2\\lambda}(x-v)^2+\\ell(v)$ and its derivatives. A consequence of the same analysis is the universal bound $\\sigma_\\ell^2 I(\\sigma_\\ell G+SY)\\ge 1/\\delta$ on the ratio $\\sigma_\\ell=\\alpha/\\mu$, with $G,S$ standard normal and $Y$ the corrupted sign of $S$; inverting it gives a numerical upper bound on correlation that holds for every continuously differentiable convex loss.","pith_inferences":["If correct, the paper reduces the optimal-loss design problem to a scalar program: choose the convex loss whose Moreau envelope minimizes $\\alpha/\\mu$ in (9). The authors flag this as open; pursuing it could turn the universal bound into an achievability result.","The universal bound is proved from a Fisher-information inequality, suggesting the optimal estimator may be one that saturates an information-theoretic limit for the binary channel; a natural test is whether a piecewise-linear loss designed from the channel statistics attains the bound.","The paper's LS result shows one-bit measurements behave like linear measurements with a known scaling and noise variance; if the same equivalence holds for other losses, one could extend these correlation predictions to full distributional statements about the estimator error."],"forward_implications":["For any convex loss satisfying the theorem's conditions, the limiting correlation is now a computable function of $\\delta$ and $\\varepsilon$, so LS, LAD, hinge, and logistic estimators can be compared exactly without simulation.","For least squares the system closes in closed form: $\\mu=(1-2\\varepsilon)\\sqrt{2/\\pi}$ and $\\alpha^2=(1-(1-2\\varepsilon)^2 2/\\pi)/(\\delta-1)$, recovering the interpretation of one-bit least squares as a noisy linear inverse problem.","For hinge and logistic losses, recovery is only well-posed above a threshold $\\delta^*_\\varepsilon$ that runs from $+\\infty$ at $\\varepsilon=0$ down to $2$ at $\\varepsilon=1/2$; simulations show hinge can outperform LS and LAD at larger $\\delta$.","The universal bound of Theorem III.2 means no continuously differentiable convex loss can exceed the correlation upper bound derived from $\\sigma_\\ell^2 I(\\sigma_\\ell G+SY)\\ge 1/\\delta$."],"supporting_citations":[{"why":"Supplies the least-squares special case and the reduction pattern this paper extends to general convex losses.","marker":"[29]"},{"why":"Provides the formal convex Gaussian min-max theorem statement and the deferred technical arguments for the auxiliary-problem convergence.","marker":"[30]"},{"why":"Introduces the convex Gaussian min-max comparison inequality on which the proof's reduction rests.","marker":"[31]"},{"why":"Gives the Gaussian min-max comparison inequality that underlies the convex Gaussian min-max theorem.","marker":"[17]"},{"why":"Supplies the threshold (15) used to characterize when hinge and logistic losses have bounded minimizers.","marker":"[7]"},{"why":"Provides the Moreau-envelope derivative identities used to derive the optimality equations.","marker":"[24]"},{"why":"Supplies the Fisher-information scaling identities used in the universal lower bound.","marker":"[1]"},{"why":"Provides the lemma whose steps the universal lower-bound proof follows.","marker":"[10]"}],"fun_headline_variants":["One-bit recovery exactly solved for convex losses","Scalar system pins down one-bit convex estimator accuracy","Exact asymptotics for one-bit inversion with convex losses","Universal correlation bound for one-bit convex estimators","Moreau envelope only input for exact one-bit asymptotics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing step is the deferred claim that the data-dependent minimizers of the simplified auxiliary problem converge to the deterministic solution of the limiting problem; if that convergence fails, the correlation formula does not follow even when the rest of the setup holds.","fun_headline_variants_meta":{"raw":{"variants":["One-bit recovery exactly solved for convex losses","Scalar system pins down one-bit convex estimator accuracy","Exact asymptotics for one-bit inversion with convex losses","Universal correlation bound for one-bit convex estimators","Moreau envelope only input for exact one-bit asymptotics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000606,"raw_usage":{"total_tokens":2806,"prompt_tokens":909,"completion_tokens":1897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":1823}},"tokens_in":525,"tokens_out":1897,"duration_ms":13434,"temperature":1.0,"reasoning_tokens":1823,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:42:31.578876+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pick a loss in the paper's class, say logistic loss, fix $\\varepsilon=0.1$ and $\\delta=4$, solve (9) numerically to get the predicted correlation, then simulate the estimator at $n=1024$ over many trials; a systematic mismatch between the simulated average correlation and the prediction would refute the formula. The same test can be sharpened by computing the minimizers of the auxiliary problem (39) and checking whether they converge to the solution of (40), the exact deferred step the proof relies on.","supporting_citations":[{"cited_title":"Lasso with non-linear measurements is equivalent to one with linear mea- surements","cited_arxiv_id":null,"evidence_quote":"Supplies the least-squares special case and the reduction pattern this paper extends to general convex losses."},{"cited_title":"Regu- larized linear regression: A precise analysis of the estimation error","cited_arxiv_id":null,"evidence_quote":"Introduces the convex Gaussian min-max comparison inequality on which the proof's reduction rests."},{"cited_title":"On Milman’s inequality and random subspaces which escape through a mesh in Rn","cited_arxiv_id":null,"evidence_quote":"Gives the Gaussian min-max comparison inequality that underlies the convex Gaussian min-max theorem."},{"cited_title":"Monotonic central limit theorem for densities","cited_arxiv_id":null,"evidence_quote":"Supplies the Fisher-information scaling identities used in the universal lower bound."},{"cited_title":"High dimensional robust m- estimation: Asymptotic variance via approximate message passing","cited_arxiv_id":null,"evidence_quote":"Provides the lemma whose steps the universal lower-bound proof follows."}],"review_version":1}