{"id":"d1fa42ec-b785-4ec1-9f17-2a0973b3cff4","arxiv_id":"2502.01583","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"For multi-index models, the eigenvalues, eigenvector overlaps, and optimal preprocessing of spectral estimators are characterized exactly in the proportional asymptotics.","lead":"This paper derives exact formulas for how well spectral methods recover the hidden low-dimensional structure in multi-index models, as data size and dimension grow together. It also finds the data preprocessing that needs the fewest samples for weak recovery, with numerical checks.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4.3 silently assumes y has a density; for discrete/mixed outputs the optimal threshold formula and T* are undefined.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing gap: Theorem 4.3 uses conditional densities and Lebesgue integrals without assuming y is absolutely continuous. This is the most serious issue because it affects the central optimal-threshold claim and the construction of the optimal preprocessing function, and it cannot be repaired by a purely technical remark: for discrete outputs the formula is not merely unproven but undefined. Other potential concerns are secondary. The eigenspace-invariance condition in Theorem 4.2 is explicit, is described as a proof artifact, and does not feed into the threshold argument of Theorem 4.3, which relies only on the lim-inf overlap positivity in (4.5) and the vanishing in (4.7). The use of lim rather than lim inf in the definition of δc(T) is a definitional imprecision that could be fixed by replacing the limit with lim inf to match Definition 3.1; it does not invalidate the formula for the threshold itself. Both of these are less fundamental than the missing density assumption. Since the reader already flagged the same concern and assigned a conditional verdict, the appropriate recommendation is to keep the verdict unchanged: the paper should state an absolute-continuity assumption for Theorem 4.3 (or provide a discrete-output version), after which the main claims are likely correct.","tokens_in":40336,"tokens_out":11303,"duration_ms":101728,"concrete_test":"Instantiate a single-index model with binary labels: y = 1_{s_1 + ε > 0}, ε ~ N(0,1), u = e_1. This satisfies (A1)-(A4) and any bounded T satisfies (A5), but y is Bernoulli and has no Lebesgue density. Compute the optimal threshold δc by directly optimizing over the two values T(0), T(1) using the eigenvalue outlier condition (E.1), which is well-defined from the distribution of z = T(y). Then attempt to evaluate the right-hand side of (4.9): the integrand Es[p(y|s)(<s,u>^2-1)]^2 / Es[p(y|s)] requires a density p(y|s), so the expression is undefined. If one formally replaces the Lebesgue integral by a sum over y ∈ {0,1} using conditional probabilities, the resulting number will generally differ from the directly optimized threshold, showing that Theorem 4.3 as stated does not cover discrete outputs and needs an explicit density assumption or a separate discrete formulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central optimality claim rests on Theorem 4.3, whose statement defines p(y|s) as the conditional density of y given s and uses integrals over Lebesgue measure in (4.9) and (4.10). Assumptions (A1)-(A5) do not require y to be absolutely continuous: the link q and noise ε are unrestricted apart from moment conditions, so y may be discrete (e.g., classification labels) or mixed. In such cases p(y|s) does not exist as a density, the ratio Es[p(y|s)(<s,u>^2-1)]/Es[p(y|s)] in the proof of Theorem 4.3 (Appendix E, lines around (E.4)-(E.5)) is not well-defined, and the proposed optimal preprocessing T* in (4.10) cannot be formed. This is not a harmless technicality: the paper's stated scope is general multi-index models, and the optimal weak recovery threshold is a headline contribution. Without an explicit absolute-continuity assumption (or a discrete analog with sums and probability mass functions), the theorem overclaims. The proof of the upper bound and the achievability construction both depend on the density representation, and no alternative is supplied for non-atomic outputs.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes spectral estimators for multi-index models in the proportional regime n/d -> δ with fixed signal dimension p. For a preprocessing function T, the spectral matrix is D_n = (1/n) A^T Z A with z_i = T(y_i). The main results are: Theorem 4.1 locates the top-p eigenvalues of D_n and exhibits a BBP-type phase transition, with outliers converging to ζδ(α_i) where α_i solve det(ζδ(α)I - R∞(α)) = 0; Theorem 4.2 characterizes the overlaps of the corresponding eigenvectors with the signal subspace; Theorem 4.3 gives the optimal weak-recovery threshold δc over all bounded preprocessing functions and constructs an optimal T* in (4.10). The proofs are based on an equivalent spectral characterization of D_n through the functions L_i, L_{i,j} and on results from Bai-Yao random matrix theory adapted from the single-index case. The paper also proves an equivalence with the threshold of Troiani et al. under simultaneous diagonalizability of the conditional second-moment matrices.","tokens_in":40644,"tokens_out":5260,"duration_ms":49864,"significance":"If the results are correct, this is a substantial advance: it provides the first precise asymptotic characterization of spectral estimators for general multi-index models with correlated signals, it identifies the optimal preprocessing function for weak recovery, and it resolves the conjecture of the parallel work DDM+25 in a wide class of cases. The derivation is self-contained: the threshold is obtained by optimizing a variational bound over preprocessing functions, the optimal T* is constructed by saturating Hölder's inequality, and no fitted parameters enter the asymptotic formulas. The appendices contain detailed proofs, and the numerical experiments in Figures 1 and 2 show good agreement with the predicted overlaps. The main caveat is a hidden regularity assumption in Theorem 4.3: the statement and proof use the conditional density p(y|s) and Lebesgue integrals over y, while the model assumptions allow y to be discrete or mixed. This is a load-bearing gap in the optimality claim, but it is local and can likely be fixed by adding an explicit absolute-continuity condition or by reformulating the result for general output distributions.","major_comments":[{"comment":"The statement and proof of the optimality result assume that y admits a conditional density p(y|s) with respect to Lebesgue measure and that the integrals over dy are Lebesgue integrals. Assumptions (A1)-(A5) do not imply this: the link function q and the noise ε are only required to satisfy moment conditions, so y may be discrete or mixed (for example, a classification label or a quantized response). In such cases the conditional density p(y|s) is not defined, the ratio appearing in (E.5) is not meaningful, and the proposed optimal preprocessing T* in (4.10) cannot be formed. Because Theorem 4.3 is the central optimality claim, the manuscript overclaims for the stated model class. Please add an explicit assumption that the conditional law of y given s is absolutely continuous with respect to Lebesgue measure, or reformulate Theorem 4.3 using conditional expectations/Radon-Nikodym derivatives so that discrete and mixed outputs are covered. In addition, the integrand in (4.9) implicitly requires E_s[p(y|s)] > 0 almost everywhere; this should also be stated explicitly.","section":"Theorem 4.3 and Appendix E (Eqs. (4.9), (4.10), (E.4), (E.5))"},{"comment":"The theorem says 'Let α1 ≥ ... ≥ α_j > τ (for some j ∈ [p]) be all the solutions' and then discusses the 'remaining p − j eigenvalues'. If (4.2) has no solutions in ]τ, ∞[, the statement as written does not cover the case j = 0, which is exactly the no-outlier regime discussed in the proof. Please either allow j = 0 explicitly or add a separate sentence stating that, when (4.2) has no solutions, all top-p eigenvalues converge to ζδ(λ̄δ). This is a presentation gap in a main theorem, but it is easy to repair.","section":"Theorem 4.1 (statement, page 6)"}],"minor_comments":[{"comment":"The eigenspace-invariance condition preceding (4.6) is an explicit additional assumption, and the authors note that it may be a proof artifact. Since the condition is not derived from the model assumptions, it would be helpful to state in the theorem or in a remark which natural classes of q are known to satisfy it (for example, permutation-invariant links via Proposition D.1) and whether there is a known example where it fails. This would clarify the scope of the overlap formula.","section":"Theorem 4.2 (page 7)"},{"comment":"The equivalence with the threshold of [TDD+24] is proved only under the conditions that the supremum is achieved by a rank-1 matrix and that the matrices E(y) are simultaneously diagonalizable. The main text is careful about this, but the appendix title 'Equivalence to [TDD+24]' could be read as unconditional. Consider a more guarded title or an explicit sentence in the appendix stating the exact hypotheses under which the two thresholds coincide.","section":"Remark 4.1 and Appendix G"},{"comment":"The phrase 'the remaining p − j eigenvalues' is slightly ambiguous when j = p; in that case there are no remaining eigenvalues and the statement is vacuous. Clarifying the notational convention for j = 0 and j = p would improve readability.","section":"Section 4, paragraph after Theorem 4.1"}],"recommendation":"major_revision","confidential_remarks":"This is a strong paper with a self-contained derivation and substantial technical content. The main issue is the unstated absolute-continuity assumption in Theorem 4.3, which affects the headline optimality claim. The fix is likely straightforward: either add an explicit assumption on the conditional law of y given s or provide a measure-theoretic version of the optimality result. I do not see a reason to reject on this basis, and I recommend major revision rather than minor revision because the density assumption is load-bearing for the central theorem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper delivers what it promises: precise asymptotics for spectral estimators in multi-index models with correlated signals, plus an optimal preprocessing rule. The proof is genuinely random-matrix-theoretic, extends [MM19] and [ZMV22], and resolves a conjecture from [DDM+25] for permutation-invariant links. The fixed-point characterization and the overlap formula are new. The appendices are detailed, and I found no circularity. This is serious work.\n\nThe main soft spot is Theorem 4.3. The statement defines p(y|s) as a conditional density and optimizes over integrals with respect to Lebesgue measure, but assumptions (A1)-(A5) don't require y to be absolutely continuous. If y is discrete, as in classification, the ratio in (4.9)-(4.10) is undefined. The authors don't flag this; as stated, the theorem overclaims. The fix is either to add an explicit absolute-continuity assumption or to prove a discrete analogue. Until that's done, the headline optimality result is conditional on a regularity condition that is not stated.\n\nThe second caveat, the invariance condition in (4.6), is milder. The authors acknowledge it, explain it's a proof artifact, and show it holds for permutation-invariant links. I wouldn't block on that.\n\nSo: the math is plausible, the proofs are detailed, the results are significant. But the main theorem needs a corrected statement. I'd send it to a serious referee and ask for a revision that addresses the density assumption. It's worth a reading group slot.","headline":"Sharp spectral asymptotics for correlated multi-index models, but the optimality theorem silently assumes y has a density and needs a corrected statement.","tokens_in":41087,"tokens_out":1495,"would_cite":true,"duration_ms":14392,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["60B20","62H25"],"pacs":[],"model":"deepseek-v4-flash","headline":"Under Gaussian design, spectral estimators for multi-index models have exactly characterized top eigenvalues and eigenvector overlaps, and the minimal sample ratio for weak recovery is attained by an explicit preprocessing function.","keywords":["multi-index models","spectral estimators","weak recovery","phase transition","random matrix theory","Gaussian design","optimal preprocessing","high-dimensional asymptotics"],"falsifier":"Run the model $y=\\mathbf{1}\\{s_1s_2+\\varepsilon>0\\}$ with $p=2$, a bounded $T$, and $n/d=\\delta$. The expressions in (4.9)-(4.10) require $p(y|s)$ as a Lebesgue density, which does not exist for this atomic output, so the claimed optimal threshold cannot be evaluated directly; any discretization that makes it computable will produce a number that can be compared with the simulated outlier onset, settling whether the hidden regularity assumption is essential to the theorem.","tokens_in":40167,"feed_emoji":"📈","tokens_out":11174,"duration_ms":86923,"temperature":0.7,"pith_summary":"This paper analyzes spectral estimators for multi-index models $y_i = q(\\langle a_i,w_1^*\\rangle,\\ldots,\\langle a_i,w_p^*\\rangle,\\varepsilon_i)$ with Gaussian design, where the estimator forms $D_n = \\frac{1}{n}A^\\top Z A$ with $z_i = T(y_i)$ and returns the top-$p$ eigenvectors. It establishes that as $n,d\\to\\infty$ with $n/d\\to\\delta$, the top-$p$ eigenvalues converge almost surely to the explicit values $\\zeta_\\delta(\\alpha_i)$, where $\\alpha_i$ solve the $p$-dimensional equation $\\det(\\zeta_\\delta(\\alpha)I - R_\\infty(\\alpha))=0$. Eigenvalues that emerge from the bulk correspond to eigenvectors with non-vanishing overlap with the signal subspace, so weak recovery happens exactly when such an outlier appears. The paper then optimizes over the preprocessing function $T$ and proves that the smallest sample ratio at which any spectral estimator can weakly recover the subspace is $\\delta_c$, given in closed form in (4.9), with a matching optimal preprocessing $T^*_\\delta$ in (4.10). This matters because spectral methods are widely used warm-starts, and the result turns their design from a heuristic choice into a provably optimal one.","feed_headline":"Exact formula fixes optimal spectral weak-recovery threshold","feed_subtitle":"Exact eigenvalues and overlaps for any preprocessing yield the minimal sample ratio and the provably optimal T.","key_machinery":"The load-bearing object is the $p\\times p$ matrix $R_\\infty(\\alpha)=\\mathbb{E}[\\alpha s s^\\top z/(\\alpha-z)]$ together with the scalar function $\\zeta_\\delta(\\lambda)=\\psi_\\delta(\\max\\{\\bar\\lambda_\\delta,\\lambda\\})$, where $\\psi_\\delta(\\lambda)=\\lambda(1/\\delta+\\mathbb{E}[z/(\\lambda-z)])$ and $\\bar\\lambda_\\delta$ is the minimizer of $\\psi_\\delta$. Eigenvalue locations are the solutions of $\\det(\\zeta_\\delta(\\alpha)I-R_\\infty(\\alpha))=0$; the proof compresses the $(d-p)$-dimensional spectral problem to this $p$-dimensional equation through a rank-$p$ perturbation formula, and the eigenvector overlaps are expressed through $\\zeta_\\delta'(\\alpha)$ and $R_\\infty'(\\alpha)$. The fixed-point functions $\\tilde L_i(\\mu)$ of (5.2)-(5.3), whose limits are obtained from low-rank-perturbation asymptotics, carry the convergence argument, and the optimality proof uses H\\\"older's inequality to show no $T$ can beat the threshold in (4.9) while $T^*_\\delta$ attains it.","core_discovery":"The paper claims that for any bounded preprocessing function $T$ with $P(z=0)<1$, the spectral matrix $D_n$ has a precisely describable spectrum: under the stated assumptions, the top $p$ eigenvalues converge almost surely to $\\zeta_\\delta(\\alpha_1)\\ge\\cdots\\ge\\zeta_\\delta(\\alpha_j)>\\zeta_\\delta(\\bar\\lambda_\\delta)$, where $\\alpha_i$ are the solutions of the secular equation, and the remaining $p-j$ eigenvalues collapse to the bulk edge $\\zeta_\\delta(\\bar\\lambda_\\delta)$. Whenever an eigenvalue is an outlier, i.e. $\\alpha_i>\\bar\\lambda_\\delta$, the corresponding eigenvector subspace has asymptotically non-vanishing squared overlap with the signal subspace, and if the eigenvalue stays in the bulk the overlap vanishes. For the weak-recovery threshold $\\delta_c$, defined as the infimum over all admissible $T$ of the smallest $\\delta$ at which the top-$p$ eigenvectors achieve non-vanishing overlap, the paper proves $\\delta_c$ equals the closed-form expression (4.9), and the preprocessing $T^*_\\delta$ in (4.10) achieves weak recovery for every $\\delta>\\delta_c$. This simultaneously provides the first exact asymptotic characterization of spectral methods in general multi-index models and an optimality certificate for one particular choice of $T$.","pith_inferences":["If (4.9) is taken as the fundamental threshold, its integrand resembles a Fisher-information-type quantity for the direction $u$; one could test whether $\\delta_c$ also lower-bounds every estimator, spectral or not, in models where computational and statistical thresholds coincide.","The hidden absolute-continuity assumption on $y$ suggests a concrete extension: for discrete or mixed outputs, the variational problem should be reformulated with probability mass functions, and the resulting threshold may differ from (4.9).","The optimal $T^*_\\delta$ depends on the conditional density $p(y|s)$; in practice this density must be estimated, and a finite-sample guarantee for the plug-in spectral estimator (how many samples suffice to stay above threshold) is left implicit by the paper."],"forward_implications":["For any admissible $T$, the top-$p$ eigenvalues and the phase transition are determined by a $p\\times p$ equation, so the outlier onset of a spectral estimator is directly computable without simulation.","Weak recovery by the top-$p$ eigenvectors occurs exactly when some $\\alpha_i>\\bar\\lambda_\\delta$; inside the bulk all overlaps vanish, identifying $\\delta_c(T)$ with the appearance of the first outlier.","The optimal preprocessing $T^*_\\delta$ provably attains the minimal sample ratio over all bounded preprocessing functions, which implies that existing heuristic choices for $T$ are suboptimal in general multi-index models.","The threshold formula generalizes the single-index and independent-mixture results to arbitrarily correlated signals and requires only the link function $q$ and the signal covariance $\\Sigma$, which can be estimated on a separate sample."],"supporting_citations":[{"why":"It supplies the single-index weak-recovery threshold and optimal preprocessing that this paper extends to p correlated signals.","marker":"[MM19]"},{"why":"It provides the spectral initialization eigenvalue equation and the limiting lemmas used to take high-dimensional limits.","marker":"[LL20]"},{"why":"It gives the generalized spiked-model asymptotics used to evaluate the low-rank perturbation functions.","marker":"[BY12]"},{"why":"It analyzes spectral methods for mixtures of single-index models, whose independence-of-signals assumption is removed here.","marker":"[ZMV22]"},{"why":"It computes the computational weak-learnability threshold to which the paper's $\\delta_c$ is shown equivalent for permutation-invariant models.","marker":"[TDD+24]"},{"why":"It describes the BBP phase transition that gives the outlier-from-bulk phenomenon characterized in Theorem 4.1.","marker":"[BBAP05]"},{"why":"It derives the single-index optimal spectral preprocessing for phase retrieval, used as a baseline in the numerical comparison.","marker":"[LAL19]"},{"why":"It supplies the perturbation theory used to prove convergence of eigenprojectors and the overlap formula in Theorem 4.2.","marker":"[Kat95]"}],"fun_headline_variants":["Exact asymptotics pin down optimal spectral weak-recovery threshold","Optimal spectral estimator found via exact spectrum","Spectral weak-recovery threshold proven optimal via exact spectrum","Phase transition sets optimal spectral weak-recovery limit","Minimal sample size for spectral weak recovery now exact"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The main optimality result assumes the response $y$ admits a conditional density $p(y|s)$ with respect to Lebesgue measure, even though the model assumptions do not guarantee this for discrete or mixed outputs.","fun_headline_variants_meta":{"raw":{"variants":["Exact asymptotics pin down optimal spectral weak-recovery threshold","Optimal spectral estimator found via exact spectrum","Spectral weak-recovery threshold proven optimal via exact spectrum","Phase transition sets optimal spectral weak-recovery limit","Minimal sample size for spectral weak recovery now exact"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00119,"raw_usage":{"total_tokens":4947,"prompt_tokens":1015,"completion_tokens":3932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":631,"completion_tokens_details":{"reasoning_tokens":3856}},"tokens_in":631,"tokens_out":3932,"duration_ms":22033,"temperature":1.0,"reasoning_tokens":3856,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:53:36.896315+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the model $y=\\mathbf{1}\\{s_1s_2+\\varepsilon>0\\}$ with $p=2$, a bounded $T$, and $n/d=\\delta$. The expressions in (4.9)-(4.10) require $p(y|s)$ as a Lebesgue density, which does not exist for this atomic output, so the claimed optimal threshold cannot be evaluated directly; any discretization that makes it computable will produce a number that can be compared with the simulated outlier onset, settling whether the hidden regularity assumption is essential to the theorem.","supporting_citations":[],"review_version":1}