{"id":"c7c43ff4-6603-4ab6-8b9b-170c10810231","arxiv_id":"2502.04654","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A sliced-Wasserstein and k-nearest-neighbor minimum-distance estimator for the distribution of random coefficients β is consistent with polynomial-in-dimension computation, while its diffusion and causal extensions remain heuristic.","lead":"This paper proposes a new way to estimate the distribution of random coefficients in linear models Y = <β, X> from observed (X, Y) pairs, combining sliced Wasserstein distance with nearest-neighbor weights. The method is computationally polynomial in dimension and comes with convergence-rate guarantees for an idealized estimator, plus exploratory diffusion-based and causal-inference extensions.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Rate theorem covers only global minimizers of F_k, while Algorithms 1–2 provably reach only local minima of the Monte Carlo surrogate sF_k; the computed estimator's advertised rate is therefore unestablished.","rationale":"Re-reading the manuscript confirms that the reader's weakest assumption is the correct primary concern. The core theoretical result—a rate for any global minimizer of the sliced-Wasserstein-plus-kNN objective—is internally coherent: the proof via the minimum-distance inequality and Fournier-Guillin bounds is standard, and the compact-support and positive-density assumptions are explicit. The decisive gap is the algorithm-to-estimator link. The method actually implemented minimizes sF_k by block coordinate descent; Theorem 4.3 certifies only local minima. The proof of Theorem 4.2 crucially uses F_k(μ̂β) ≤ F_k(μ_N), and no analogue is shown for local minima of sF_k. Since the sorting-permutation parametrization makes sF_k non-convex, local minima need not be global. Section 5 explicitly disclaims theory for the diffusion variant, and the Algorithm inputs incorrectly include the unobservable βi; these are secondary. I also noticed a technical gap in Lemma A.2, Step 3: the inclusion S(eV, eγ) ⊂ S(V, sγ) relies on a lower bound ⟨V,U⟩ ≥ 1 - ∥V−eV∥^2 - eγ^2, which can fail; a correct bound requires an O(∥V−eV∥) term, but this affects only the Assumption 4.4 rates and appears repairable with adjusted constants. The proposed MIQP test can settle whether the local-minimum gap is practically real on small instances, directly probing whether Algorithm 1's output deviates from the global minimizer. Since the reader's CONDITIONAL verdict already captures this unresolved gap, no verdict change is needed.","tokens_in":26161,"tokens_out":21794,"duration_ms":206041,"concrete_test":"Fix d=2, n=100, k=5, m=20 and generate data from the Section 7.1 simulation model. Formulate the global minimization of sF_k as a mixed-integer quadratic program (MIQP) by introducing binary variables for the sorting permutation of each sampled direction, and solve it to global optimality with a solver such as Gurobi. Run Algorithm 1 from 1000 uniform random initializations on the same data and record the best objective value. Repeat over 50 independent datasets. If the best Algorithm 1 objective exceeds the MIQP global minimum by more than a 1e-6 relative tolerance in any dataset, the algorithm provably stops at a non-global local minimum, confirming that Theorem 4.2's rates are not inherited by the computed estimator. If the two objectives coincide in all trials, the practical gap is weakened, though an analytic proof would still be needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 4.2 guarantees the n^{-1/(d+5)} / n^{-1/(2d-1)} SW2 rate for any global minimizer of the population objective F_k in (2.3). Algorithms 1 and 2, however, optimize the Monte Carlo surrogate sF_k in (3.2). Theorem 4.3 shows only that Algorithm 1 terminates at a local minimum of sF_k and that Algorithm 2 tracks Algorithm 1 for one step as m grows. Nothing in the paper shows that such a local minimum satisfies the minimum-distance inequality F_k(μ̂) ≤ F_k(μ_N) that drives the proof of Theorem 4.2 via equation (A.1), where μ_N is the empirical measure of N iid β-copies. Without this inequality or a substitute uniform suboptimality bound, the computed output cannot be claimed to attain the advertised rates. This missing link is the step from the oracle estimator to the practical method, so the abstract-level claim that the proposed method is consistent is not supported by the theorems as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a minimum-distance estimator for the distribution of random coefficients β in the linear model Y = ⟨β, X⟩, with β and X independent. The estimator minimizes an integrated squared Wasserstein objective F_k(μ) = ∫_{S^{d-1}} W_2^2(sμ^V_k, μ^V)dσ(V), where sμ^V_k is a k-nearest-neighbor empirical measure of projected observations, and the minimization is over N-point discrete measures supported on a ball. Theorem 4.2 claims polynomial rates, in expectation, for the sliced-Wasserstein distance between any global minimizer and the true distribution under Assumptions 4.3 or 4.4. The computation section introduces a Monte Carlo surrogate sF_k and two block coordinate descent algorithms; Theorem 4.3 states that Algorithm 1 reaches a local minimum of sF_k and that Algorithm 2 tracks it as m→∞. Sections 5–6 extend the approach to a diffusion-based generative algorithm and to treatment-effect estimation, with numerical experiments on simulated data and the ACTG 175 data.","tokens_in":26373,"tokens_out":16314,"duration_ms":152448,"significance":"The oracle consistency result (Theorem 4.2) is a solid contribution: it provides polynomial n^{-1/(d+5)} or n^{-1/(2d-1)} SW2 rates for a minimum-distance estimator without smoothness conditions on the density of X, and the proof uses a clean decomposition with Fournier–Guillin empirical-measure bounds and a new k-NN radius lemma. The proposed algorithms have polynomial complexity in dimension, and reproducible code is provided. The connections to sliced-Wasserstein flows and causal random coefficient models are interesting and will likely stimulate further work. However, the advertised practical consistency of the proposed method is not established: the algorithms are only shown to converge to local minima of the Monte Carlo surrogate, and the bridge to the oracle estimator is missing. This gap must be addressed before the central claim can be accepted.","major_comments":[{"comment":"The consistency rates in Theorem 4.2 apply to any global minimizer of the population objective F_k in (2.3), whereas Algorithms 1 and 2 minimize the Monte Carlo surrogate sF_k in (3.2). Theorem 4.3 establishes only that Algorithm 1 outputs a local minimum of sF_k and that Algorithm 2 tracks Algorithm 1 as m→∞; no result shows that a local minimum of sF_k approximately minimizes F_k, nor is there a bound on the suboptimality gap F_k(μ̂_alg) − min_{P_N} F_k. Uniform convergence of sF_k to F_k (Proposition 4.3) does not transfer local minima to global minima, even in the limit. Consequently, the computed estimator is not shown to achieve the rates of Theorem 4.2, and the abstract's claim that 'the proposed method is consistent' is not supported by the theorems as written. The authors should either add a bridging result (e.g., an approximate global optimality guarantee for Algorithm 1 or 2) or explicitly restrict the consistency claim to the oracle estimator and describe the algorithms as heuristics with empirical support.","section":"Theorem 4.2 vs. Section 4.3 and Theorem 4.3"},{"comment":"The stated rates in Theorem 4.2(ii) contain a log n factor, but the proof derives an upper bound of order n^{-α/(dα+5α+2)} log n for E[SW2^2] and then applies Jensen's inequality to obtain E[SW2] ≤ C n^{-α/(2dα+10α+4)} (log n)^{1/2} (and similarly in the d ≥ 6 case). The theorem as printed omits the square root on the logarithmic factor. This is a minor correction to the statement, but it should be fixed to accurately reflect the proof.","section":"Theorem 4.2(ii) and proof, Step 3 Case II"}],"minor_comments":[{"comment":"The input lists for Algorithms 1–3 include βi, but β is unobserved and the algorithms' computations do not use βi; this is confusing and should be corrected by removing βi from the inputs.","section":"Algorithm 1 (and 2, 3) inputs"},{"comment":"The statement of Proposition 4.2 refers to 'the random functions F_k ◦ η' but the proposition concerns the surrogate objective sF_k ◦ η; both the proof and the surrounding text make clear that the functions are sF_k ◦ η, so the statement should be amended.","section":"Proposition 4.2"},{"comment":"The claim that when k = n, F_k reduces to the SW2 distance between the empirical measure and the target is inaccurate: sμV_n is built from the transformed observations eY_i eX_i^T V, which coincide with the true projections β_i^T V only approximately, so F_n is not the SW2 distance between the empirical measure of β and the target.","section":"Remark 2.1"},{"comment":"Typographical errors: 'casual inference' in the paper-organization paragraph should be 'causal inference', and 'folllowing' in Section 7.1 should be 'following'.","section":"Introduction and Section 7.1"},{"comment":"The notation paragraph defines the closed ball 'in(R, ∥·∥2)', which should presumably be R^d; the typo is minor but may confuse readers.","section":"Notation paragraph"},{"comment":"In Table 2, the empirical distance for µdis decreases more slowly than for µsph and µdeg as n increases, but the paper does not comment on this; a brief remark would improve the interpretation of the experiments.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The oracle consistency theorem is a genuine contribution, but the manuscript's framing overstates what is proved for the implemented algorithms. I recommend requesting a revision that either proves a global-optimality or suboptimality-gap result connecting the algorithms' outputs to the oracle estimator, or carefully reframes the consistency claim to apply only to the oracle minimum-distance estimator with the algorithms treated as empirically validated heuristics. The numerical study and the reproducible code are valuable, but the theoretical core currently supports an oracle result rather than the proposed method as implemented."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: the oracle estimator is solid, and Theorem 4.2's rates are a real contribution. But there's a gap between the estimator defined in (2.3) and the algorithms: the theorems only show Algorithms 1 and 2 reach local minima of the Monte Carlo surrogate sF_k, not that those local minima achieve the n^{-1/(d+5)} or n^{-1/(2d-1)} rates. That means the abstract-level claim that the proposed method is consistent is stronger than the results support.\n\nWhat's actually new: a minimum-distance estimator for random coefficient distributions that combines sliced Wasserstein distance with k-NN-induced projected measures. Theorem 4.2 gives polynomial convergence for any global minimizer of F_k, avoiding smoothness assumptions on the density of X. The proof is detailed and largely self-contained, using Fournier-Guillin empirical measure bounds and a new NN-radius lemma. The claimed polynomial-in-d complexity for Algorithm 2 is plausible.\n\nThe main soft spot is the local-minimum gap described above. Proposition 4.3's uniform convergence of sF_k to F_k doesn't close it: a local minimizer of the surrogate need not have near-optimal F_k, and Theorem 4.3 only guarantees one-step tracking as m grows. This is fixable by proving a uniform suboptimality bound or by restating the rates as applying to the oracle estimator only; either way it needs to be addressed. The diffusion section explicitly disclaims theory, which is honest but makes Section 5 a heuristic extension. The simulations compare only the three proposed algorithms, so they don't show competitiveness against existing estimators. Minor issues: Algorithm 1's input lists beta_i, which are unobserved, and the relationship to Fan and Park (2024) deserves a clearer discussion.\n\nThe core idea is sound, the novelty is real, and the authors are candid about limitations. I'd accept this for peer review with a request to close or clearly scope the algorithm-versus-oracle gap, and to add at least one baseline to the experiments.","headline":"Solid oracle estimator theory and rates, but the algorithms only provably reach local minima of a Monte Carlo surrogate, so the advertised consistency of the computed output is not yet backed by the theorems.","tokens_in":26908,"tokens_out":3544,"would_cite":true,"duration_ms":34463,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes a k-nearest-neighbor sliced Wasserstein estimator for the distribution of random coefficients in the linear model $Y=\\langle\\beta,X\\rangle$ and proves polynomial convergence rates without smoothness conditions on the…","keywords":["sliced Wasserstein distance","minimum-distance estimation","random coefficient models","k-nearest neighbors","diffusion process","treatment effect distribution","nonparametric estimation","consistency rates"],"falsifier":"For a low-dimensional case such as $d=2$, $n=2000$, compare the output of Algorithm 2 with an exact or near-global minimizer of the Monte Carlo objective $\\hat F_k$ computed by fine-grained search over candidate $N$-point measures. If the algorithm's error in $SW_2$ decays slower than the theorem's $n^{-1/7}$ while the global minimizer attains it, the practical consistency claim would fail.","tokens_in":25940,"feed_emoji":"📊","tokens_out":13615,"duration_ms":123121,"temperature":0.7,"pith_summary":"The paper studies the linear random coefficient model $Y=\\langle\\beta,X\\rangle$, where each unit has its own coefficient vector $\\beta$ and the goal is to estimate the whole distribution of $\\beta$, not just its mean and variance, from observations of $(X,Y)$ only. It proposes an estimator that minimizes, over discrete candidate distributions, the integrated squared two-Wasserstein distance between the candidate's projections and $k$-nearest-neighbor empirical measures built from normalized data. The central theoretical claim is that any global minimizer of this objective is consistent at a polynomial rate: $n^{-1/(d+5)}$ for dimensions $2\\le d\\le5$ and $n^{-1/(2d-1)}$ for $d\\ge6$, with an analogous rate under a weaker tail condition that covers intercept models. This matters because the result is obtained without any smoothness assumption on the density of $X$, and the proposed algorithms have complexity that scales polynomially in the dimension rather than exponentially as in grid-based approaches.","feed_headline":"Sliced Wasserstein estimator recovers random-coefficient laws","feed_subtitle":"Nearest-neighbor projections recover full coefficient laws at polynomial rates, even with rough covariate densities","key_machinery":"The central object is the k-NN sliced objective $F_k(\\mu)=\\int_{\\mathbb{S}^{d-1}}W_2^2(\\hat\\mu_k^V,\\mu^V)\\,d\\sigma(V)$. For each direction $V$ on the unit sphere, the data are normalized to $\\tilde X_i=X_i/\\|X_i\\|_2$ and $\\tilde Y_i=Y_i/\\|X_i\\|_2$; the $k$ nearest neighbors of $V$ among the $\\tilde X_i$ define a $k$-point empirical measure $\\hat\\mu_k^V$ of projected responses. Minimizing $F_k$ over $N$-point discrete measures in a bounded ball matches each candidate distribution's one-dimensional projections against these neighborhood empirical projections. The convergence proof separates the error into an empirical-measure term controlled by standard Wasserstein bounds and a localization term controlled by the expected squared angular radius $\\gamma(V,k)^2$ of the $k$-nearest-neighbor ball, which Lemma A.2 bounds via the density lower bound or tail condition. For computation, the paper replaces the integral over directions with a Monte Carlo average over $m$ directions and minimizes it by block coordinate descent, yielding a projected-gradient variant whose per-iteration cost is $O(mdn\\log n + tmk\\log k)$.","core_discovery":"The paper claims that the distribution of the random coefficient vector $\\beta$ in the linear model $Y=\\langle\\beta,X\\rangle$, with $X$ independent of $\\beta$ and $\\beta$ supported on a bounded ball, is consistently recoverable by the k-NN sliced Wasserstein minimum-distance estimator $\\hat\\mu_\\beta$ defined in (2.3). Theorem 4.2 states that every global minimizer of the population objective satisfies $\\mathbb{E}[SW_2(\\mu_\\beta,\\hat\\mu_\\beta)]\\le C n^{-1/(d+5)}$ for $2\\le d\\le5$ and $\\le C n^{-1/(2d-1)}$ for $d\\ge6$ under a lower-bound density condition on the direction of $X$, with a further $\\alpha$-dependent bound under Assumption 4.4 that permits intercept regressors and polynomial-tailed densities. The paper emphasizes that these rates do not require smoothness of the density of $X$ or its normalized direction, in contrast to earlier estimators in this problem class.","pith_inferences":["The paper proves its rate for global minimizers of the population objective but only shows its algorithms reach local minima of the Monte Carlo objective; closing that gap would require a landscape analysis, for example an incoherence condition under which every local minimum lies near a global one.","The diffusion-based Algorithm 3 is presented without rate guarantees; a natural extension is to prove that the Euler-Maruyama particle system converges to the regularized minimum-distance estimator as $m,L\\to\\infty$ and $h\\to0$, putting the generative variant on the same footing as Algorithms 1 and 2.","Because $F_k$ is a distance between data-induced and candidate measures, the same construction could be turned into a distributional goodness-of-fit test for random coefficient models, although the paper does not pursue hypothesis testing.","The causal application assumes the unit-level treatment effect $R$ is independent of covariates $Z$ but may be correlated with the noise $U$; if that independence fails, the estimator would recover a mixture involving the joint law of $(R,U)$, pointing toward an extension to correlated coefficient models."],"forward_implications":["Any global minimizer of $F_k$ is consistent at a polynomial rate for every dimension $d\\ge2$, so the full distribution of $\\beta$ is estimable from $(X,Y)$ observations alone under the stated assumptions.","The rate theorem requires no smooth density for $X$: a lower bound on the spherical density, or a polynomial tail condition, suffices, so intercept models and regressors with limited continuous variation are covered.","The proof makes the tuning explicit: choose $N=k=n^{6/(d+5)}$ for $2\\le d\\le5$ and $N=k=n^{d/(2d-1)}$ for $d\\ge6$, with analogous choices (up to $\\log n$) under the tail condition.","Algorithm 2 has per-iteration cost $O(mdn\\log n + tmk\\log k)$, giving polynomial-in-$d$ computation instead of the exponentially many grid points used by existing nonparametric estimators.","Under the regularized causal working model, the same objective estimates the distribution of unit-level treatment effects in a randomized trial with discrete treatment, and the simulation on the ACTG 175 data illustrates the estimated treatment-effect distributions."],"supporting_citations":[{"why":"Establishes the minimum-distance estimation approach for random coefficient regression that this paper modifies with k-NN projections and the sliced Wasserstein distance.","marker":"(Beran and Millar, 1994)"},{"why":"Supplies the Wasserstein empirical-measure rate bounds invoked twice in the proof of Theorem 4.2, once for a fresh empirical measure of beta and once for the oracle neighborhood measure.","marker":"Fournier and Guillin (2015, Theorem 1)"},{"why":"Provides the block coordinate descent template for discrete sliced Wasserstein losses that Algorithms 1 and 2 adapt, and its Monte Carlo approximation analysis that Proposition 4.3 generalizes.","marker":"Tanguy et al. (2024, Algorithm 1)"},{"why":"Its sliced-Wasserstein flow theorem yields the continuity equation and particle approximation that the diffusion-based Algorithm 3 implements in Section 5.","marker":"Liutkus et al. (2019, Theorem 2)"},{"why":"Defines the causal random coefficient model whose unit-level treatment effect distribution is the target of the Section 6 application.","marker":"Heckman et al. (1997, Section 7)"},{"why":"Provides the prior rate-optimal estimator in this problem class and the assumption that Assumption 4.4 refines to allow polynomial-tailed regressors.","marker":"Holzmann and Meister (2020)"}],"fun_headline_variants":["Sliced Wasserstein estimator recovers coefficient laws at polynomial rates","k-NN sliced Wasserstein recovers random coefficient distributions","Random coefficient laws recovered without smooth covariate densities","Sliced Wasserstein yields fast recovery of random coefficient models","Sliced Wasserstein recovers coefficient laws without density smoothness"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The convergence-rate theorem is proven for the best possible (global) answer to the estimation problem, while the implemented algorithms are only shown to reach a locally best answer to a randomly approximated version of that problem; the paper assumes without proof that the local answer inherits the global rate.","fun_headline_variants_meta":{"raw":{"variants":["Sliced Wasserstein estimator recovers coefficient laws at polynomial rates","k-NN sliced Wasserstein recovers random coefficient distributions","Random coefficient laws recovered without smooth covariate densities","Sliced Wasserstein yields fast recovery of random coefficient models","Sliced Wasserstein recovers coefficient laws without density smoothness"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001887,"raw_usage":{"total_tokens":7336,"prompt_tokens":815,"completion_tokens":6521,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":431,"completion_tokens_details":{"reasoning_tokens":6437}},"tokens_in":431,"tokens_out":6521,"duration_ms":45418,"temperature":1.0,"reasoning_tokens":6437,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:58:22.931762+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a low-dimensional case such as $d=2$, $n=2000$, compare the output of Algorithm 2 with an exact or near-global minimizer of the Monte Carlo objective $\\hat F_k$ computed by fine-grained search over candidate $N$-point measures. If the algorithm's error in $SW_2$ decays slower than the theorem's $n^{-1/7}$ while the global minimizer attains it, the practical consistency claim would fail.","supporting_citations":[{"cited_title":"and Millar, P","cited_arxiv_id":null,"evidence_quote":"Establishes the minimum-distance estimation approach for random coefficient regression that this paper modifies with k-NN projections and the sliced Wasserstein distance."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the block coordinate descent template for discrete sliced Wasserstein losses that Algorithms 1 and 2 adapt, and its Monte Carlo approximation analysis that Proposition 4.3 generalizes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Its sliced-Wasserstein flow theorem yields the continuity equation and particle approximation that the diffusion-based Algorithm 3 implements in Section 5."},{"cited_title":"J., Smith, J., and Clements, N","cited_arxiv_id":null,"evidence_quote":"Defines the causal random coefficient model whose unit-level treatment effect distribution is the target of the Section 6 application."},{"cited_title":"and Meister, A","cited_arxiv_id":null,"evidence_quote":"Provides the prior rate-optimal estimator in this problem class and the assumption that Assumption 4.4 refines to allow polynomial-tailed regressors."}],"review_version":1}