{"id":"54dd3292-8d01-4580-bc47-2bdec4307f8f","arxiv_id":"2602.14914","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"β*-IPS, the optimal additive control-variate estimator, asymptotically dominates SNIPS in MSE; the exact variance gap is (V(π)σ²_w − σ_{w,wr})²/(nσ²_w).","lead":"An off-policy estimator with an optimally tuned additive baseline (β*-IPS) is proved to be asymptotically at least as accurate, and usually strictly more accurate, than the standard self-normalized estimator SNIPS. The paper supplies an exact variance-gap formula and extends the dominance to ranked-list evaluation under the Item-Position Model.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 3.1's proof never bounds the remainder on the dangerous tail {0<W̄<1/2}; it only dismisses the exponentially rare W̄=0, so E[R_n^2]=O(n^{-2}) is not established as written.","rationale":"The reader's weakest assumption correctly identifies the tail handling in Theorem 3.1's Step 2: the proof bounds the remainder on the good event and then discards the bad event by invoking the exponentially small probability of W̄=0, but small-positive W̄ is the event that can blow up 1/W̄ and the text never controls R_n there. This is a genuine gap in the proof as written, and it is load-bearing because the O(n^{-2}) remainder is what converts the exact variance gap into asymptotic dominance. However, the gap is not fatal: bounded rewards give |X̄/W̄|≤R, so R_n is bounded on {W̄>0} and the missing tail term is exponentially small. The fix uses only assumptions already in the theorem. I therefore agree with the CONDITIONAL verdict: the central claim is true but the proof needs tightening. The plug-in-β bias issue in §3.4 is secondary because the plug-in estimator has the same leading variance and only O(n^{-1}) bias, matching SNIPS; but the unresolved tail argument is the more fundamental obstruction to accepting the theorem as stated.","tokens_in":10289,"tokens_out":13502,"duration_ms":127077,"concrete_test":"Repair Step 2 formally: insert the missing bound |R_n|≤C on {W̄>0} using |X̄/W̄|≤R (from |r_i|≤R, w_i≥0), then recompute E[R_n^2] = E[1_E R_n^2] + E[1_{E^c} R_n^2] and verify both terms are O(n^{-2}) using only Assumptions (1)-(3). If the repair requires an extra condition such as E[w^{-1}]<∞ or a lower bound on W̄, the theorem statement must be weakened; if it goes through unchanged, the central claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Step 2 of the Theorem 3.1 proof bounds R_n only on E={W̄≥1/2}. For E^c it says only that W̄=0 has exponentially small probability. The threatening event for the O(n^{-2}) remainder is 0<W̄<1/2, where the 1/W̄ factor in R_n=L_n(1-W̄)/W̄ is large and is not controlled by Hoeffding. The manuscript never bounds E[R_n^2 1_{E^c}] or the cross-term on that event. This is load-bearing because the central inequality MSE(β*-IPS)≤MSE(SNIPS)+O(n^{-2}) and the strict-dominance conclusion depend directly on the remainder being O(n^{-2}). The gap is repairable under the stated assumptions: since |r_i|≤R and w_i≥0, |X̄/W̄|≤R, hence |R_n|=|(X̄/W̄ - V)(1-W̄)|≤(R+|V|)(1+W) wherever W̄>0, giving E[R_n^2 1_{E^c}]≤C^2 P(E^c)=O(exp(-n/(2W^2))). But that argument is absent from the text; the sentence about W̄=0 addresses the wrong event. As published, the proof of the main theorem is incomplete exactly where self-normalisation's denominator is dangerous.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies off-policy evaluation (OPE) and compares two variance-reduction strategies: self-normalised importance sampling (SNIPS) and additive baseline corrections (β-IPS). The main theorem (Theorem 3.1) claims that the estimator using the population-optimal additive baseline, β*-IPS, asymptotically dominates SNIPS in mean squared error: MSE(β*-IPS) ≤ MSE(SNIPS) + O(n^{-2}), with strict dominance for sufficiently large n when β* ≠ V(π) and Var(w) > 0. The proof decomposes SNIPS as β-IPS(V(π)) plus a remainder, bounds the remainder, and then compares the optimal baseline with V(π). Proposition 3.2 gives an exact asymptotic variance gap between the two estimators. Section 3.3 extends the argument per-position to ranking under the Item-Position Model, defining β⊥⊥-IPM and claiming position-wise dominance over SNIPM. Section 3.4 discusses the practical use of an empirically estimated baseline and its finite-sample bias.","tokens_in":10501,"tokens_out":8740,"duration_ms":89074,"significance":"If correct, the result provides a clean theoretical justification for preferring additive control variates over self-normalisation in OPE, filling a gap left by prior empirical work (Gupta et al., RecSys 2024). The paper's main strengths are the exact decomposition in Eq. (7), the closed-form variance gap in Eq. (13), and the extension to ranking via a per-position argument. The dominance inequality is a genuine in-paper derivation rather than a restatement of known results. The result is also falsifiable: the variance gap has an explicit algebraic form and the strict-dominance condition is testable from data. The main caveats are that the proof of the O(n^{-2}) remainder is incomplete as written and the practical plug-in claim is not formally established.","major_comments":[{"comment":"The proof of the O(n^{-2}) remainder is incomplete. The event E^c = {W̄ < 1/2} is dismissed via Hoeffding's inequality, but the text only says 'W̄ = 0 occurs with exponentially vanishing probability'. The dangerous event for the factor 1/W̄ in R_n = L_n(1-W̄)/W̄ is {0 < W̄ < 1/2}, where the factor is large and is not controlled by the fourth-moment bound on L_n. Thus E[R_n^2 1_{E^c}] is not bounded as written. This is load-bearing: the central inequality MSE(β*-IPS) ≤ MSE(SNIPS) + O(n^{-2}) and the strict-dominance conclusion both depend on this remainder being O(n^{-2}). The gap is repairable under the stated assumptions: on {W̄>0}, |X̄/W̄| ≤ max|r_i| ≤ R, hence |R_n| ≤ (R+|V|)(1+W), so E[R_n^2 1_{E^c}] ≤ C^2 P(E^c) = O(exp(-n/(2W^2))). The manuscript should state this argument or an equivalent; it currently addresses the wrong event. The cross-term analysis in Step 3 also uses 'R_n ≈ L","section":"§3.1, Step 2 (Eqs. (8)–(11))"},{"comment":"The practical claim that the empirically estimated β̂*-IPS 'maintains the same asymptotic MSE convergence rate as SNIPS whilst benefiting from the superior variance constant' is asserted without proof. Theorem 3.1 concerns the population-optimal β*, not the data-dependent β̂*. The extra term introduced by plug-in estimation is (β̂*-β*)(1-W̄), which is O_p(n^{-1}) if β̂* is n^{1/2}-consistent, so the asymptotic variance constant is indeed unchanged; but the mean-squared-error claim, including the O(n^{-1}) bias, should be stated as a lemma or given a rigorous derivation. Since the paper's practical recommendation is to replace SNIPS with β̂*-IPS, this step should not be left as an informal note.","section":"§3.4 (plug-in β̂*)"}],"minor_comments":[{"comment":"The optimal baseline β* = σ_{w,wr}/σ_w^2 is undefined when Var(w)=0. The theorem and Proposition 3.2 should explicitly assume Var(w_i) > 0, or treat the degenerate case separately.","section":"§3.1, Assumptions and Theorem 3.1"},{"comment":"The one-sided Hoeffding bound is fine, but the sentence 'W̄ = 0 occurs with exponentially vanishing probability' is misleading because P(E^c) is not the same as P(W̄=0), and the latter is not the event that needs controlling. Rephrase once the E^c bound is fixed.","section":"§3.1, Eq. (10)"},{"comment":"The notation for the per-position optimal baseline is inconsistent: the theorem and text use β*_{⊥⊥,j}, while Eq. (17) writes β_{⊥⊥,j} without the star. Please align the notation.","section":"§3.3, Eq. (17)"},{"comment":"The reference to cross-fitting [2] is appropriate but the 'negligible reduction in effective sample size' claim is informal; a precise statement of the variance cost would be useful.","section":"§3.4"},{"comment":"The abstract and introduction state the result as a 'definitive answer' and 'theoretically justify shifting' from SNIPS. Given that Theorem 3.1 is asymptotic and relies on bounded importance weights and the oracle β*, the wording could be softened to reflect the scope of the assumptions.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The central claim is likely correct under the stated boundedness assumptions, and the proof gap in Step 2 of Theorem 3.1 is localized and repairable with a short boundedness argument. The paper should be sent back for a revision rather than rejected. I would also ask the authors to provide a formal statement for the plug-in claim in §3.4, since the practical recommendation depends on it. No concerns about novelty or scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline is that the paper gets the core comparison right: β*-IPS is asymptotically no worse than SNIPS in MSE, and generically strictly better. The variance algebra checks out, and the exact gap formula in Eq. (13) is a clean, usable result. But the novelty is mostly repackaging classical ratio-estimator vs control-variate theory (Owen, Joachims et al.) in OPE notation. The genuinely new pieces are the O(n^{-2}) remainder statement and the position-wise ranking extension.\n\nWhat it does well: Proposition 3.2's variance gap is exact, Theorem 3.3 is a correct per-position application, and §3.4 is honest about the plug-in bias. The heavy self-citation to [6] is not circular because β*'s optimality is derived in-paper.\n\nWhere it's soft:\n\n1. Theorem 3.1's proof has a real hole. Step 2 bounds the remainder on {W̄ ≥ 1/2} and then dismisses W̄=0 as exponentially unlikely. The dangerous event for the 1/W̄ factor is 0 < W̄ < 1/2, and the text never bounds E[R_n^2 1_{E^c}] there. The stress-test note is right, and also right that it's repairable: with bounded rewards and nonnegative weights, |X̄/W̄| ≤ R, which bounds R_n pointwise on W̄>0. But that argument is absent. As written, the main theorem is incomplete at exactly the spot self-normalisation is delicate.\n\n2. The practical recommendation outruns the theory. §3.4 acknowledges the estimated β̂* introduces O(n^{-1}) bias, and mentions cross-fitting, but gives no analysis of whether the dominance survives plug-in estimation. The paper's own §3.5 claims finite-sample implications, but those are not proven.\n\n3. The ranking theorem is over-sold. Theorem 3.3 proves position-wise dominance, but the IPM total is a sum; its variance includes across-position covariances, which are never compared. So \"dominance for ranking and recommendation\" in the abstract is stronger than what's shown.\n\nNone of these sink the central claim, which I believe is true. But the proof needs tightening and the claims need scaling back.\n\nWho it's for: OPE practitioners and recsys researchers who want a formal reason to switch from SNIPS to additive baselines. It deserves serious peer review; a good reviewer can fix the tail bound and push for more modest claims. I'd engage with it.\n\nRecommendation: send to review, expect a revision.","headline":"The paper's main theorem is true but classical; the proof has a repairable hole in the tail bound, and the ranking claims overreach.","tokens_in":11188,"tokens_out":2524,"would_cite":true,"duration_ms":25528,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves that the estimator with the variance-minimising additive control variate (β★-IPS) asymptotically dominates self-normalised IPS (SNIPS) in mean squared error, and that the same holds at every position for ranking evaluation","keywords":["off-policy evaluation","self-normalised importance sampling","control variates","variance reduction","ranking policies","mean squared error","importance weights","baseline corrections"],"falsifier":"Simulate off-policy evaluation with heavy-tailed importance weights (e.g., log-normal) and compare finite-sample MSE of β★-IPS and SNIPS at large n; if SNIPS ever has strictly lower MSE, the bounded-weight remainder bound in the proof is the point of failure. Alternatively, compute the exact asymptotic variance gap from the data and check whether empirical MSE differences converge to it at the claimed rate.","tokens_in":9995,"feed_emoji":"📊","tokens_out":3904,"duration_ms":38409,"temperature":0.7,"pith_summary":"Off-policy evaluation estimates how a new policy would perform using logged data, without running online tests. SNIPS is a standard variance-reduction trick that divides by the average importance weight, trading bias for lower variance. The paper proves that an additive baseline correction—subtracting an optimally chosen constant from each reward—is asymptotically at least as good as SNIPS, and strictly better unless the optimal baseline happens to equal the true policy value. By deriving the exact variance gap, it shows SNIPS is asymptotically equivalent to using the generally sub-optimal baseline equal to the policy value. This gives practitioners a theoretical reason to prefer baseline corrections over self-normalisation, including for ranked outputs.","feed_headline":"Additive control variates beat self-normalized estimators","feed_subtitle":"New proof shows optimal baseline corrections dominate SNIPS in mean squared error for large samples.","key_machinery":"The key mechanism is the exact decomposition SNIPS = β-IPS(V) + R_n, meaning self-normalisation is asymptotically nothing but an additive-baseline estimator whose baseline is the true policy value. The variance-gap identity Δ = (V(π)σ²_w − σ_{w,wr})² / (nσ²_w) ≥ 0 quantifies the loss from this sub-optimal baseline, and the O(n⁻²) bound on R_n, obtained via Hoeffding and Rosenthal inequalities under bounded weights and rewards, ensures the gap is the leading term in the MSE comparison.","core_discovery":"The central claim is Theorem 3.1: for i.i.d. logged data with bounded rewards and bounded importance weights, MSE(β★-IPS) ≤ MSE(SNIPS) + O(n⁻²), and the inequality is strict for all sufficiently large n whenever β★ ≠ V(π) and Var(w) > 0. The proof constructs the exact identity SNIPS = β-IPS(V) + R_n, where β-IPS(V) is the additive-baseline estimator with baseline fixed to the true policy value V(π), and shows that the remainder R_n has mean-squared size O(n⁻²). Proposition 3.2 then gives the exact asymptotic variance gap: Δ = (V(π)σ²_w − σ_{w,wr})² / (nσ²_w) ≥ 0, so SNIPS's variance equals that of the additive estimator with baseline V(π), which is generally not the variance-minimising β★ =","pith_inferences":["A natural testable extension is to check whether the dominance survives when importance weights are heavy-tailed; the O(n⁻²) remainder bound relies on bounded weights, and simulation could reveal whether a modified theorem holds in the unbounded case.","The same decomposition might apply to other ratio estimators beyond SNIPS, such as weighted importance sampling with different normalisers, suggesting additive control variates are a general recipe for variance reduction.","The paper leaves cross-position dependencies in rankings open; a globally optimal vector of baselines accounting for these dependencies could dominate the per-position approximation β★⊥⊥ even further."],"forward_implications":["Practitioners can replace SNIPS with β★-IPS and provably reduce asymptotic MSE whenever the optimal baseline differs from the policy value.","For ranking evaluation under the Item-Position Model, per-position additive baselines dominate the self-normalised SNIPM estimator at every rank.","Cross-fitting the empirically estimated β★ restores exact finite-sample unbiasedness while retaining the variance advantage.","Because the remainder decays as O(n⁻²), the variance gap dominates except in very small samples, so gains appear precisely in the policy-mismatch regimes where variance reduction matters most.","The exact gap formula gives a diagnostic: estimating β★ and V(π) from data reveals how much variance SNIPS is leaving on the table."],"fun_headline_variants":["Additive control variates dominate self-normalised OPE","Proof: additive baselines beat SNIPS in MSE","Optimal additive baseline dominates SNIPS asymptotically","Self-normalisation lag: additive control variates lead","Additive baselines surpass self-normalized estimators"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The theorem requires bounded importance weights (0 ≤ w ≤ W < ∞) and bounded rewards to control the O(n⁻²) remainder; if weights are heavy-tailed, the declared asymptotic dominance is not proven and may fail.","fun_headline_variants_meta":{"raw":{"variants":["Additive control variates dominate self-normalised OPE","Proof: additive baselines beat SNIPS in MSE","Optimal additive baseline dominates SNIPS asymptotically","Self-normalisation lag: additive control variates lead","Additive baselines surpass self-normalized estimators"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":1924,"prompt_tokens":720,"completion_tokens":1204,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":464,"completion_tokens_details":{"reasoning_tokens":1129}},"tokens_in":464,"tokens_out":1204,"duration_ms":10231,"temperature":1.0,"reasoning_tokens":1129,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T23:03:23.522378+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate off-policy evaluation with heavy-tailed importance weights (e.g., log-normal) and compare finite-sample MSE of β★-IPS and SNIPS at large n; if SNIPS ever has strictly lower MSE, the bounded-weight remainder bound in the proof is the point of failure. Alternatively, compute the exact asymptotic variance gap from the data and check whether empirical MSE differences converge to it at the claimed rate.","supporting_citations":[],"review_version":2}