{"id":"417d6bd5-d719-433c-815f-ffcec2845cfb","arxiv_id":"2506.14950","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"A DML estimator for conditional moment restrictions is proposed, but its central N^{-1/2} rate theorem is broken because the selected score is degenerate at the truth.","lead":"This paper proposes DML-CMR, a double machine learning estimator for conditional moment restrictions including IV regression and proximal causal learning, and claims a root-N convergence rate. The proof of that rate is invalid because the squared score has zero Jacobian and zero covariance at the true parameter.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The score ψ=(s(c)-g(f,c))² is zero at the true parameter, so the Jacobian J0 and score covariance required by Condition 4(d)/(e) are degenerate; the N^{-1/2} claim of Theorem 6 is unsupported.","rationale":"I read the paper in good faith: the intended contribution is a debiased two-stage CMR estimator with a proved O(N^{-1/2}) rate, plus strong empirical results on IV and PCL benchmarks. The reader's weakest-assumption analysis correctly identifies a structural degeneracy rather than a missing minor regularity condition. The degeneracy follows directly from the definitions of s0 and g0: the score ψ=(s(c)-g(f,c))² vanishes identically at the true parameter, so the Jacobian and covariance needed by the DML theorem are both zero. This is not a matter of choosing constants poorly; the theorem's stated assumptions contradict the paper's own equation ψ(D;f0,s0,g0)=0, which the proof of Theorem 3 establishes. The Appendix C.1 verification of Condition (e) is also circular or vacuous in exactly the way the reader states: the trace argument cannot produce a positive eigenvalue for a matrix that is identically zero. Consequently, Theorem 6, Corollary 7, and Proposition 9, which all rely on the non-singularity of J0, do not follow. The empirical evaluations and released code are valuable but do not substitute for the central theoretical guarantee that the abstract and introduction advertise. Because the main claim is unsupported by the paper's own equations, the REJECT verdict should stand unchanged.","tokens_in":38448,"tokens_out":4591,"duration_ms":50138,"concrete_test":"Use the simplest linear IV example: X=γZ+ν, Y=θ0X+ε with Cov(ν,ε)≠0, so the CMR is E[Y-θX|Z]=0. Then s0(Z)=E[Y|Z]=θ0γZ and g0(fθ,Z)=E[θX|Z]=θγZ, giving ψ=(θ0-θ)²γ²Z². At θ=θ0, ψ=0 a.s., so computing J0=∂θ E[(θ0-θ)²γ²Z²]|θ=θ0 gives 0, and E[ψ²]|θ=θ0 gives 0. This directly falsifies Condition 4(d)-(e) and makes the variance in Theorem 6 undefined. A second check is to re-derive the Appendix C.1 proof of Condition (e) for this scalar score; the trace argument must establish E[ψ²]>0, but explicit computation gives E[ψ²]=0, so the claimed positive definiteness cannot hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is Theorem 6: DML-CMR is concentrated in an N^{-1/2} neighbourhood of θ0 and satisfies √N(θ̂-θ0) → N(0,σ²). To invoke the DML theorem, Condition 4(d) requires the Jacobian J0 of E[ψ] w.r.t. θ to have singular values bounded away from zero, and Condition 4(e) requires the covariance E[ψψ^T] at θ0 to be positive definite. Both fail for the proposed score. By the CMR definition and the paper's own equations, s0(C)=E[Y|C]=E[f0(X)|C]=g0(f0,C), so ψ(D;f0,s0,g0)=(s0(C)-g0(f0,C))²=0 almost surely. Therefore E[ψ(D;f0,s0,g0)]=0, J0=∂θ E[(s0(C)-g0(fθ,C))²]|θ=θ0=2E[(s0(C)-g0(fθ0,C))∂θ g0(fθ0,C)]=0, and E[ψψ^T]=E[ψ²]=0. Condition 4(d) reduces to a vacuous inequality with J0=0, and Condition 4(e) is plainly false. The trace-trick verification of Condition (e) in Appendix C.1 is invalid: for a scalar ψ that is identically zero on the support of D, the matrix E[ψψ^T] is the scalar 0, not a positive-definite matrix. The squared objective may be a reasonable loss to minimize, but it is not a valid DML moment function with nondegenerate first-order behaviour. Theorem 6's variance formula σ²=J0^{-1}E[ψψ^T](J0^{-1})^T is undefined, so the central theoretical claim is not proved.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DML-CMR, a two-stage estimator for conditional moment restrictions built on the double/debiased machine learning framework. The estimator uses the score function ψ(D; f, (s, g)) = (s(c) − g(f, c))², where s(c) = E[Y|C=c] and g(f, c) = E[f(X)|C=c] are treated as nuisance parameters, and it employs K-fold cross-fitting so that nuisance estimators are trained on the complement of the fold used for the second-stage loss. Under a parametric assumption on the function of interest, the paper claims in Theorem 6 that the resulting estimator is concentrated in an N^{−1/2} neighbourhood of the true parameter and satisfies asymptotic normality, and it also claims this rate is minimax optimal. The experiments compare DML-CMR against existing IV regression and proximal causal learning methods on synthetic, semi-synthetic, and high-dimensional benchmarks.","tokens_in":38791,"tokens_out":10013,"duration_ms":119736,"significance":"If the central convergence claim were valid, the paper would make a useful contribution: it provides a concrete, non-adversarial DML-style algorithm for general conditional moment problems, it supports both neural and tree-based nuisance estimation, and the experimental evaluation is broad, includes external baselines, and is backed by released code. The connection drawn between DML's identification condition and bounded ill-posedness is also potentially interesting. However, the load-bearing theoretical result is not established: the proposed squared score is degenerate at the true parameter, so the Jacobian and score-covariance conditions required by the DML theorem fail. Because the claimed O(N^{−1/2}) rate is the paper's primary theoretical contribution, the soundness of the central claim is in question.","major_comments":[{"comment":"The score function is identically zero at the true parameter. Since s0(C) = E[Y|C] = E[f0(X)|C] = g0(f0, C), we have ψ(D; f0, (s0, g0)) = (s0(C) − g0(f0, C))² = 0 almost surely. Consequently the Jacobian J0 = ∂θ′ E[(s0(C) − g0(fθ′, C))²]|θ′=θ0 = 2E[(s0(C) − g0(fθ0, C))(−∂θ g0(fθ0, C))] = 0, and the covariance E[ψ(D; f0, (s0, g0))ψ(D; f0, (s0, g0))T] = E[ψ²] = 0. This directly contradicts Condition 4(d), which requires the singular values of J0 to be bounded away from zero, and Condition 4(e), which requires the score covariance to have all eigenvalues strictly positive. Moreover, the variance formula σ² = J0^{-1}E[ψψT](J0^{-1})T in Theorem 6 is undefined because J0 is not invertible. Thus the central N^{-1/2} convergence claim is not proved.","section":"Section 4.1, Eq. (5) and Appendix C.1, proof of Theorem 6"},{"comment":"The trace-trick argument for positive definiteness is invalid. The proof notes that each outer product ψψT has eigenvalue ∥ψ∥² ≥ 0 and concludes that E[ψψT] is positive definite, but non-negativity of eigenvalues is not positivity. For the proposed scalar score, ψ = 0 on the support at the true parameter, so the only eigenvalue is 0 and the covariance matrix is the zero matrix. This is not a technical gap but a direct failure of Condition 4(e).","section":"Appendix C.1, verification of Condition (e)"},{"comment":"The identification condition is also internally inconsistent with the squared score. A first-order Taylor expansion gives E[ψ(D; fθ, (s0, g0))] = J0(θ − θ0) + O(∥θ − θ0∥²), and since J0 = 0, the leading term vanishes and ∥E[ψ]∥ = O(∥θ − θ0∥²). No constant c0 > 0 can therefore satisfy 2∥E[ψ]∥ ≥ c0∥θ − θ0∥ in a neighbourhood of θ0. Proposition 9, which derives bounded ill-posedness from Condition 4(d), rests on the same impossible assumption.","section":"Section 4.3, Condition 4(d) and Section 4.4, Proposition 9"},{"comment":"Even setting aside the degeneracy of the score, the proof that Condition 4(f) holds is not correct as written. Lemma 5 bounds the squared norm ∥bs − s0∥₂² by c1(δN(S*)² + √(log(1/ζ)/N)), but the proof of Theorem 6 converts this into ∥bs − s0∥₂ ≤ O(δN + N^{-1/2}). Taking square roots of the lemma's bound yields O(N^{-1/4} log^{1/4}) at best when δN = o(N^{-1/4}), which is not o(N^{-1/4}) as Condition 4(f) requires. The nuisance-rate condition therefore does not follow from the stated lemma.","section":"Appendix C.1, Lemma 5 and Condition 4(f)"}],"minor_comments":[{"comment":"Equation (6) has a typo: the second-stage loss should read bEk[(bsk(c) − bgk(fbθ, c))²], with the closing parenthesis and square placed correctly.","section":"Section 4.2, Eq. (6)"},{"comment":"The caption reads \"Real-World Datsets\"; this should be \"Real-World Datasets\".","section":"Figure 3 caption"},{"comment":"The text refers to \"Lemma 3.3\" and \"Theorem 3.4 and 3.5\", but these numbered results do not exist in the paper; the intended references are likely Lemma 5 and Theorem 6 (or Corollary 7).","section":"Appendix G.2"},{"comment":"The notation bEck_i in line 11 of Algorithm 1 is not defined; the paper should define it as the empirical expectation over the mini-batch sampled from fold Ik.","section":"Algorithm 1"}],"recommendation":"reject","confidential_remarks":"The main reason for rejection is the degeneracy of the proposed score: the moment function vanishes at the true parameter, so the DML rate theorem cannot be applied. This is a load-bearing structural issue that cannot be fixed by a small modification to the proof; the squared score would need to be replaced by a different first-order identifying moment, which would change the method. I would not recommend inviting a revision unless the authors can provide a fundamentally different theoretical argument for their estimator."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper's main theoretical claim, the root-N convergence rate for DML-CMR, is not supported by the paper's own equations. The proposed score ψ=(s(c)−g(f,c))² is exactly zero at the true functions, because s0(c)=g0(f0,c). That makes both the Jacobian J0 and the covariance E[ψψ^T] zero at θ0, so Conditions 4(d) and 4(e) of the DML theorem fail. The trace-trick argument in Appendix C.1 does not fix this; for a scalar ψ that is identically zero on the support, the matrix E[ψψ^T] is the scalar 0, not positive-definite. The variance formula σ²=J0^{-1}E[ψψ^T](J0^{-1})^T is undefined. The proof of Theorem 6 thus does not go through.\n\nWhat's new and good: the paper generalises the authors' earlier DML-IV to a wider class of conditional moment restrictions, adds proximal causal learning experiments, and clearly explains the connection between DML's identifiability condition and bounded ill-posedness (Proposition 9, conditional on J0 being non-singular). The experiments are thorough—multiple benchmarks, high-dimensional inputs, weak IVs, tree-based estimators—and the code is released. As an empirical method, the squared-difference loss is a plausible heuristic and the results are decent.\n\nSoft spots: besides the broken theorem, the 'first DML for CMR with neural nets' claim is a stretch; it is an extension of their own DML-IV. The theoretical analysis only covers the parametric case, which is standard for DML but leaves the nonparametric setting—the main motivation—outside the guarantees.\n\nWho gets value: someone building debiased two-stage neural estimators might pick up the algorithmic recipe and the experimental findings. But as a theory paper, it's not there. The headline claim in the abstract is exactly what fails.\n\nI would not send it to peer review in this form. The authors need a Neyman-orthogonal score that has non-degenerate first-order behaviour at θ0 before the rate result can be taken seriously. If they can fix that, it might be worth revisiting.","headline":"The root-N rate theorem is unsupported because the proposed score is identically zero at the true parameter, so J0 and E[ψψ^T] vanish and the DML conditions fail.","tokens_in":39405,"tokens_out":3897,"would_cite":false,"duration_ms":34240,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62D20","62F12","62G05","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage estimator with a new Neyman-orthogonal score solves conditional moment restrictions at the minimax-optimal root-N rate.","keywords":["conditional moment restrictions","double machine learning","Neyman orthogonality","instrumental variable regression","proximal causal learning","neural network estimators","cross-fitting","ill-posed inverse problems"],"falsifier":"Compute $J_0 = \\partial_\\theta E[(s_0(C) - g_0(f_\\theta, C))^2]$ at $\\theta = \\theta_0$ for any CMR with $s_0(C) = E[Y \\mid C] = E[f_0(X) \\mid C]$; the derivative is zero, so the bounded-singular-values condition of Theorem 6 fails. The same holds for the covariance $E[\\psi \\psi^T]$, whose entries vanish at $\\theta_0$.","tokens_in":38158,"feed_emoji":"🎯","tokens_out":9347,"duration_ms":78811,"temperature":0.7,"pith_summary":"Conditional moment restrictions ask for a function $f_0$ satisfying $E[Y - f_0(X) \\mid C] = 0$; instrumental-variable regression and proximal causal learning are prominent examples. Standard two-stage solvers fit a nuisance function in the first stage and plug it into the second, which injects regularisation and overfitting bias. This paper proposes DML-CMR, a two-stage estimator built around a new Neyman-orthogonal score $\\psi(D; f, (s, g)) = (s(c) - g(f, c))^2$, where $s$ estimates $E[Y \\mid C]$ and $g$ estimates $E[f(X) \\mid C]$. With K-fold cross-fitting, the estimator is claimed to be asymptotically unbiased and to converge at the minimax-optimal $O(N^{-1/2})$ rate under parameterisation and mild regularity assumptions. On IV and proximal causal learning benchmarks, including high-dimensional and semi-synthetic real-world data, the reported mean squared errors improve on existing methods.","feed_headline":"Debiased two-stage estimator hits the optimal root-N rate","feed_subtitle":"A Neyman-orthogonal score removes first-stage bias in nonlinear IV and proximal causal learning.","key_machinery":"The load-bearing object is the Neyman-orthogonal score $\\psi(D; f, (s, g)) = (s(c) - g(f, c))^2$. Here $s(c)$ estimates $E[Y \\mid C]$, and $g(f, c)$ estimates $E[f(X) \\mid C]$, so the squared residual compares two conditional expectations rather than $Y$ and a prediction; at the truth the two coincide, making the expected score zero and its first derivative with respect to $(s, g)$ zero. Cross-fitting then trains these nuisance estimates on the complement of each fold before the score is minimised over $f_\\theta$, which is the mechanism that converts first-stage estimation error into a second-order term and yields the $O(N^{-1/2})$ rate.","core_discovery":"The paper's central claim is that the plug-in bias of two-stage CMR estimation can be removed without influence-function calculations by replacing the standard squared loss $(Y - g(f, c))^2$ with $(s(c) - g(f, c))^2$, where $s_0(c) = E[Y \\mid C]$ and $g_0(f, c) = E[f(X) \\mid C]$ are nuisance functions. Since $s_0(c) = g_0(f_0, c)$ at the true parameter, the score vanishes at the truth and its Gateaux derivative with respect to the nuisance parameters is zero, which is Neyman orthogonality. Under a finite-dimensional parameterisation of $f_0$ and regularity conditions, Theorem 6 asserts that the cross-fitted estimator $\\hat{\\theta}$ is concentrated in an $N^{-1/2}$ neighbourhood of $\\theta_0$ and satisfies $\\sqrt{N}(\\hat{\\theta} - \\theta_0) \\to N(0, \\sigma^2)$, matching the minimax parametric rate; the variance has the usual sandwich form. The paper also shows that the Jacobian non-singularity required by the double machine learning identification condition implies bounded ill-posedness of the underlying CMR problem.","pith_inferences":["For nonparametric $f_0$, the same Neyman-orthogonal score could be analysed through orthogonal statistical learning bounds; the paper names this as future work, and a concrete outcome would be rate guarantees that depend on function-class complexity rather than parameter dimension.","For proximal causal learning, the score's orthogonality could be ported to bridge-function estimators that currently require explicit influence-function derivations, potentially yielding doubly robust estimators for continuous treatments in a unified way.","A direct transfer to other conditional moment problems, such as demand estimation with endogeneity or treatment choice, is plausible because the score only requires estimating $E[Y \\mid C]$ and $E[f(X) \\mid C]$ as nuisance functions."],"forward_implications":["If the rate claim holds, DML-CMR offers a single debiasing recipe for any conditional moment restriction problem, not just IV and proximal causal learning, with neural-network nuisance estimators.","First-stage nuisance estimators only need to achieve the $o(N^{-1/4})$ convergence rate, a condition satisfied by many standard learners, so the method inherits the debiasing promise of double machine learning for nonlinear problems.","Because Jacobian non-singularity implies bounded ill-posedness, the paper connects the usual identifiability assumption in conditional moment problems to the double machine learning rate theorem.","Experimental comparisons on ticket demand, MNIST-augmented, IHDP, PM-CMR, and dSprites benchmarks show lower mean squared error than existing IV and proximal causal learning methods, with especially strong gains under weak instruments.","The computationally efficient version without cross-fitting, CE-DML-CMR, performs comparably on low-dimensional datasets, suggesting the score itself provides most of the debiasing benefit there."],"supporting_citations":[{"why":"Supplies the double machine learning framework, cross-fitting definition, and the $N^{-1/2}$ convergence theorem used as the template for the rate guarantee.","marker":"Chernozhukov et al. (2018)"},{"why":"DeepIV two-stage network and the ticket-demand benchmark; the paper follows its approach to estimate $g$ via conditional distributions.","marker":"Hartford et al. (2017)"},{"why":"Establishes nonparametric instrumental-variable regression as a conditional moment problem and provides the baseline sieve estimator and ill-posedness discussion.","marker":"Newey and Powell (2003)"},{"why":"Introduces proximal causal learning and the bridge-function identification that makes PCL a conditional moment restriction.","marker":"Miao et al. (2018)"},{"why":"Precursor DML-IV algorithm whose IV results are reused as experimental comparisons for the new method.","marker":"Shao et al. (2024)"},{"why":"Orthogonal statistical learning analysis of Neyman-orthogonal scores, cited for the debiasing property and for future nonparametric extensions.","marker":"Foster and Syrgkanis (2019)"},{"why":"Defines the ill-posedness measure used in Definition 8 and provides penalised sieve CMR estimation.","marker":"Chen and Pouzo (2012)"},{"why":"Classic sieve-minimum-distance estimator for CMRs with unknown functions, the theoretical standard the paper extends.","marker":"Ai and Chen (2003)"},{"why":"Effective dimension and critical radius analysis used to verify the $o(N^{-1/4})$ nuisance convergence condition.","marker":"Chernozhukov et al. (2021)"},{"why":"Local Rademacher complexity and critical radius tools used in Lemma 5 to bound nuisance estimation error.","marker":"Bartlett et al. (2005)"}],"fun_headline_variants":["Neyman-orthogonal score removes plug-in bias in conditional moment models","Root-N optimal rate without plug-in bias","Double ML removes first-stage bias in IV and proximal causal learning","Neyman-orthogonal score debiases conditional moment estimation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The rate guarantee depends on the Jacobian of the expected score and the score covariance being non-degenerate at the true parameter, but for the proposed squared score both are identically zero at the truth because $s_0(c) = g_0(f_0, c)$.","fun_headline_variants_meta":{"raw":{"variants":["Neyman-orthogonal score removes plug-in bias in conditional moment models","Root-N optimal rate without plug-in bias","Double ML removes first-stage bias in IV and proximal causal learning","Neyman-orthogonal score debiases conditional moment estimation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000744,"raw_usage":{"total_tokens":3387,"prompt_tokens":1081,"completion_tokens":2306,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":697,"completion_tokens_details":{"reasoning_tokens":2235}},"tokens_in":697,"tokens_out":2306,"duration_ms":14431,"temperature":1.0,"reasoning_tokens":2235,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:10:01.168055+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute $J_0 = \\partial_\\theta E[(s_0(C) - g_0(f_\\theta, C))^2]$ at $\\theta = \\theta_0$ for any CMR with $s_0(C) = E[Y \\mid C] = E[f_0(X) \\mid C]$; the derivative is zero, so the bounded-singular-values condition of Theorem 6 fails. The same holds for the covariance $E[\\psi \\psi^T]$, whose entries vanish at $\\theta_0$.","supporting_citations":[],"review_version":1}