{"id":"8ce62e2c-fb85-4216-a0d4-eef31cc6d238","arxiv_id":"2606.22784","paper_version":2,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Semiparametric efficiency theory is differential calculus on distributions: scores are velocities, influence functions are gradients, and efficient influence functions are their orthogonal projections.","lead":"This tutorial reframes semiparametric efficiency theory as ordinary differential calculus on probability distributions, with paths as curves, scores as velocities, and efficient influence functions as projected gradients. It aims to make the geometry of influence-function methods intuitive for causal inference and missing-data work.","discovery_kind":"review","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The Reader correctly identifies the paper as a high-quality tutorial that reorganizes known geometry rather than proving new results, and correctly notes that regularity is assumed rather than derived. That assumption is load-bearing for the classical theory itself, but not for the paper's claim that the classical theory can be presented as differential calculus on distributions. The manuscript is careful about scope (Abstract, §1, §9), does not overclaim novelty, and the internal logic of the Calculus–Statistics Dictionary is sound. No internal inconsistency or hidden gap that would undermine the pedagogical thesis was found. Verdict remains ACCEPT; no adjustment is warranted.","tokens_in":22931,"tokens_out":445,"duration_ms":5243,"concrete_test":"Independently re-derive the efficient influence function for the ATE under randomization (§8.4) from the pathwise derivative and the orthogonal decomposition T = TY|A,X ⊕ TA ⊕ TX without consulting the paper's final formula; if the projected gradient matches the displayed φ_eff, the geometric pipeline is self-consistent under the stated regularity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper is an expository reorganization of classical semiparametric geometry (paths, scores, tangent spaces, influence functions as Riesz representers, efficient influence functions as projections). Its central claim is pedagogical, not a new theorem. The regularity conditions the Reader flags (regular paths with scores in L2_0(P0) and continuous linear pathwise derivatives so that Riesz applies) are the standard hypotheses of the literature the paper surveys; they are invoked explicitly and not hidden. Because the manuscript does not claim to derive those conditions from first principles or to enlarge the class of models for which they hold, their presence is not a soft spot in the argument it actually makes. The worked examples in §8 are consistent with textbook EIFs, and the geometric dictionary is used coherently throughout.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"This tutorial reorganizes classical semiparametric efficiency theory as differential calculus on a space of probability distributions. Paths of distributions play the role of curves, scores the role of velocity vectors, influence functions the role of gradients (via Riesz representation of the pathwise derivative on the tangent space), and efficient influence functions the role of projected gradients onto the informative component T ∩ T_η^⊥. The manuscript develops a Calculus–Statistics Dictionary section by section, distinguishes model-based from parameter-based irrelevance (T^⊥ versus T_η), connects the geometry to one-step and TMLE-style estimators, and recovers textbook efficient influence functions for the average treatment effect under four nested models in §8.","tokens_in":23078,"tokens_out":870,"duration_ms":7962,"significance":"The contribution is pedagogical rather than a new theorem, but it is a high-value one for a methods journal. The geometric narrative unifies constructions that are usually introduced piecemeal (scores, tangent spaces, nuisance tangent spaces, influence functions, efficiency bounds) and answers several recurring conceptual questions (why directions are functions, why T depends only on M while T_η depends on ψ, why efficiency is a projection). The four ATE examples recover standard EIFs and illustrate when model restrictions improve the bound and when they do not. The exposition is careful, cites the foundational monographs appropriately, and should lower the barrier for students and applied researchers who use influence-function methods without a geometric picture. Strengths include the explicit Hilbert-space arguments, the orthogonal decompositions, and the concrete worked examples that match textbook results.","major_comments":[],"minor_comments":[{"comment":"The regularity conditions for regular paths (scores in L2_0(P0), continuous linear pathwise derivatives so that Riesz applies) are invoked throughout §§3–6 but never collected in one place. A short remark or appendix pointer to the standard references (e.g., van der Vaart 1991, Bickel et al. 1993) would help readers who want the technical hypotheses without interrupting the geometric narrative.","section":null},{"comment":"Figure 1 (right panel) and Figure 2 would benefit from slightly more explicit axis labels or a one-sentence caption note that the horizontal axis is the sample space and the vertical axis is density/mass; the geometric idea is clear but the panels are a bit sparse.","section":null},{"comment":"In §8 the path constructions (exponential tilting for X and Y|A,X; logit tilting for A|X) are convenient and standard, but a brief sentence noting that any other regular paths generating the same scores would yield the same tangent spaces and EIFs would forestall the impression that the particular tilting is essential.","section":null},{"comment":"Table 2 is a useful summary; adding a column or footnote that T and T^⊥ depend only on (P0,M) while T_η, T_η^⊥ and φ_eff also depend on ψ would reinforce the central geometric distinction drawn in the text.","section":null},{"comment":"A few minor typographical items: “ϵ” versus “ε” appear interchangeably for the path index in early sections; “nonpara-metric” line break in §2; and the arXiv date line (June 23, 2026) is presumably a placeholder.","section":null}],"recommendation":"accept","confidential_remarks":"The manuscript is pure exposition of classical material; novelty is in the pedagogical framing, not new results. That is appropriate for a tutorial/review slot. I see no citation or priority issues. Fit for a methods journal that publishes high-quality expository pieces is good; if the journal is strictly original-theory only, the editor may wish to route it as a tutorial rather than a research article. I would not require major revision for that reason alone."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a pure exposition paper that does exactly what it claims: it reorganizes the standard objects of semiparametric efficiency theory (paths, scores, tangent spaces, nuisance spaces, influence functions as Riesz representers, EIFs as projections) into a differential-calculus dictionary. The author states up front that nothing is new theoretically. That is accurate and refreshing.\n\nWhat it does well is the geometry. The Calculus–Statistics Dictionary is used consistently, the Hilbert-space arguments are clean, and the four ATE examples recover the textbook EIFs while illustrating when model restrictions do or do not change the bound. The distinction between model-determined directions and parameter-determined relevance is made sharper than in most monographs. For someone who already knows the material, it is a good organizing narrative; for someone learning it, it should reduce the usual “why functions, why project?” confusion.\n\nSoft spots are minor and proportional. Regularity (regular paths, scores in L2_0, continuous linear pathwise derivatives) is invoked rather than derived, but that is standard for a tutorial of this scope and is not hidden. There are no new theorems, no code, no data, and no claim that practice outside the subfield will change. Citation pattern is appropriate: Bickel, van der Vaart, Tsiatis, Robins, etc., used as background, not as circular support.\n\nThis is for people who teach or derive influence functions in causal inference and missing data. It deserves a serious referee as a pedagogical contribution; desk rejection would be a mistake. I would bring it to reading group for the geometric framing, and I would cite the dictionary sections when I need a clean reference for students. Engage with it.","headline":"Clean, honest tutorial that reorganizes classical semiparametric geometry as differential calculus; no new theory, but unusually clear and useful for teaching and derivation work.","tokens_in":23639,"tokens_out":430,"would_cite":true,"duration_ms":5059,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62F12","62G20"],"pacs":[],"model":"grok-4.5","headline":"Semiparametric efficiency theory is just differential calculus on spaces of probability distributions: paths are curves, scores are velocities, influence functions are gradients, and efficient influence functions are projected gradients.","keywords":["semiparametric efficiency","influence functions","tangent spaces","pathwise differentiability","efficient influence function","canonical gradient","one-step estimators","causal inference"],"falsifier":"Exhibit a pathwise-differentiable statistical parameter whose derivative map cannot be represented by any square-integrable influence function, or a setting in which the projected (efficient) influence function fails to give the minimal asymptotic variance among regular estimators; either would break the claimed calculus–geometry equivalence.","tokens_in":23846,"feed_emoji":"📐","tokens_out":685,"duration_ms":7293,"temperature":0.7,"pith_summary":"This tutorial reorganizes the standard machinery of semiparametric efficiency theory as ordinary multivariable calculus performed on a space whose points are probability distributions. Paths of distributions play the role of curves, scores play the role of velocity vectors, the model tangent space collects allowable directions of movement, the nuisance tangent space collects directions that leave the target parameter unchanged to first order, influence functions represent the pathwise derivative map as gradients, and the efficient influence function is the orthogonal projection of any such gradient onto the informative component of the tangent space. The same geometry explains why directions are functions rather than finite-dimensional vectors, why the model alone determines the tangent space while the parameter determines the nuisance directions, and why modern one-step, TMLE, and debiased estimators all correct first-order error by using that projected gradient. A sympathetic reader gains a single coherent picture that unifies scores, influence functions, efficiency bounds, and the construction of estimators used throughout causal inference and missing-data analysis.","feed_headline":"Efficiency theory is calculus on probability distributions","feed_subtitle":"Paths are curves, scores are velocities, and efficient influence functions are projected gradients","key_machinery":"The Calculus–Statistics Dictionary and the orthogonal decomposition L2_0(P0)=T⊥⊕Tη⊕(T∩Tη⊥): the efficient influence function is the unique element of the influence-function class that lies in T∩Tη⊥, equivalently the orthogonal projection of any influence function onto the model tangent space.","core_discovery":"Semiparametric efficiency theory is differential calculus on a space of probability distributions: once paths, scores, and the Hilbert space L2_0(P0) are identified with curves, velocities, and the ambient geometry of directions, influence functions become representations of the pathwise derivative map and the efficient influence function is uniquely the projection of that gradient onto the informative directions allowed by the model and relevant to the parameter.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Efficiency theory as differential calculus on distributions","Scores as velocities, influence functions as gradients","Projected gradients define efficient influence functions","Semiparametric efficiency: geometry of distribution paths","Why tangent spaces and projections yield efficient influence"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Everything rests on the existence of sufficiently regular paths whose scores live in a Hilbert space of mean-zero square-integrable functions and for which the pathwise derivative is a continuous linear map that can be represented by an inner product.","fun_headline_variants_meta":{"raw":{"variants":["Efficiency theory as differential calculus on distributions","Scores as velocities, influence functions as gradients","Projected gradients define efficient influence functions","Semiparametric efficiency: geometry of distribution paths","Why tangent spaces and projections yield efficient influence"]},"model":"grok-4.5","effort":"low","cost_usd":0.00411,"raw_usage":{"total_tokens":1259,"prompt_tokens":765,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":41100000,"prompt_tokens_details":{"text_tokens":765,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":427,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":765,"tokens_out":67,"duration_ms":4132,"temperature":1.0,"reasoning_tokens":427,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T12:48:13.388126+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Exhibit a pathwise-differentiable statistical parameter whose derivative map cannot be represented by any square-integrable influence function, or a setting in which the projected (efficient) influence function fails to give the minimal asymptotic variance among regular estimators; either would break the claimed calculus–geometry equivalence.","supporting_citations":[],"review_version":2}