{"id":"03499901-2d2f-4884-9323-00096dc5a260","arxiv_id":"2608.09812","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"With bounded potential outcomes, independent random assignment plus a midpoint-centered no-intercept regression is minimax optimal for average treatment effects, while balanced and paired designs can be strictly worse.","lead":"This paper proves that when outcomes are known to lie in a fixed range, randomizing each unit independently and using an unconventional recentered regression can be minimax optimal for estimating average treatment effects. The result overturns the usual preference for balanced designs and provides explicit finite-sample minimax estimators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant objection identified: the minimax theorem and proofs are internally consistent; the acknowledged worst-case modeling choice limits practical interpretation but not the central claim.","rationale":"The reader's weakest_assumption correctly identifies the pure worst-case criterion as the most fragile premise for the paper's practical recommendation. However, that is a modeling choice, not a flaw in the theorem. The paper itself acknowledges the limitation in Section 4 and after Proposition 3, so the mathematical claims are properly conditioned. I independently verified the key proof steps: the Cauchy-Schwarz lower bounds in Propositions 1-3, the endpoint averaging, the symmetrization over swaps and permutations in Theorem 1, the separate-convexity argument that reduces the continuous cube to vertices, and the affine minimax calculation in Corollary 1. I found no internal inconsistency, no unjustified step, and no sign that the stated minimax values are incorrect. The numerical figure lacks code but the convex program is precisely specified and would be straightforward to reproduce. Because the theorem and corollaries are sound, the reader's ACCEPT verdict stands without change. My partial agreement reflects that the identified limitation is real but not load-bearing for the paper's stated claims, since the claims are explicitly conditional on the worst-case objective.","tokens_in":12261,"tokens_out":21580,"duration_ms":174872,"concrete_test":"For N=4 and N=8, exhaustively enumerate the 2^(2N) endpoint configurations (A_i,C_i in {0,1}) and solve the finite zero-sum game between the researcher's choice of a decision rule on binary observations and the adversary's choice of endpoint configuration; compare the resulting value with (U-L)^2 kappa_N from the convex program (4). If the game value differs from kappa_N, the lower-bound or attainment argument in Theorem 1 is incomplete.","verdict_should_be":"UNCHANGED","load_bearing_attack":"I read the paper's central claim as a conditional mathematical statement: under the criterion of worst-case MSE over all configurations in [L,U]^(2N), independent randomization attains the minimax value (U-L)^2 kappa_N (or (U-L)^2/N for equivariant affine estimators). The proofs are internally consistent. I checked the adversarial averaging lower bounds (Propositions 1-3), the symmetrization and Jensen steps in Theorem 1, the separate-convexity argument for the Bernstein extension, and the affine minimax calculation in Corollary 1; each step is valid. The one caveat that could limit the practical force of the headline is the pure worst-case criterion: the paper itself notes (Section 4) that stratified designs based on informative observables are likely to dominate in practice. This is a modeling choice, not a technical flaw, and it is explicitly acknowledged. Conditional on the stated objective, I find no load-bearing concern for the mathematical result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies minimax estimation of the finite-population average treatment effect when all potential outcomes are known to lie in a bounded interval [L,U]. The main results are: (i) Proposition 1 shows that, among all assignment mechanisms and affine estimators satisfying the equivariance condition (2), independent Bernoulli(1/2) randomization paired with the no-intercept regression of the support-midpoint-centered outcome on the recentered treatment attains the minimax worst-case MSE (U-L)^2/N; (ii) Proposition 2 shows balanced complete randomization has worst-case MSE at least (U-L)^2/(N-1), attained by the difference-in-means estimator; (iii) Proposition 3 shows paired randomization has worst-case MSE at least (U-L)^2/J = 2(U-L)^2/N; (iv) Theorem 1 extends the design optimality to all measurable estimators, giving the minimax value (U-L)^2 κ_N from the convex program (4) and characterizing a generally nonlinear optimal estimator; (v) Corollary 1 solves the unrestricted affine problem, yielding value (U-L)^2/(√N+1)^2 via a shrunk version of the Proposition 1 estimator. The paper explicitly notes in Section 4 that the pure worst-case criterion may be pessimistic for stratified designs when strata are prognostically informative, and that such designs may dominate in practice.","tokens_in":12437,"tokens_out":17796,"duration_ms":149304,"significance":"If correct, the paper provides a sharp and surprising decision-theoretic reversal: with bounded potential outcomes, the usual prescription of balanced complete randomization and difference-in-means is not minimax, and independent randomization with a random treated share is optimal. The proof strategy is elegant, connecting the experiment-design problem to Hodges-Lehmann-style least-favorable-prior arguments, and the convex program (4) makes the optimal nonlinear estimator computable. The appendix proofs are self-contained and the main derivations balance; the reader's stress-test concern about the worst-case modeling choice is explicitly acknowledged in the paper itself and is a modeling limitation rather than a technical flaw. The paper also credits and builds on the recent recentering literature, and the claimed novelty relative to Aronow and Lopatto (2026) and Harshaw et al. (2024) is clearly delineated.","major_comments":[],"minor_comments":[{"comment":"The abstract's phrase \"Among all assignment mechanisms and a broad class of affine estimators\" could be misread as covering all affine estimators; Proposition 1 applies to affine estimators satisfying the equivariance condition (2), while Corollary 1 covers all affine estimators with a different minimax value and a shrunk estimator. Please clarify this distinction in the abstract or at the end of the introduction.","section":"Abstract and Section 2"},{"comment":"The symbol \\hat\\beta^*_{NL} is used for the shrunk affine estimator in Corollary 1 as well as for the nonlinear estimator in Theorem 1. Since the two estimators are different objects, a distinct notation for the affine shrunk estimator would avoid confusion.","section":"Corollary 1"},{"comment":"Proposition 2 assumes N is even, which is natural for balanced complete randomization, but the paper does not state what happens for odd N; a brief remark on the odd-N case would make the scope of the strict suboptimality claim precise.","section":"Proposition 2"},{"comment":"The notation p_{k,m}(x) in (4) and p_x(z) immediately after Theorem 1 are visually close; renaming one of them, for example q_{k,m}(x) or r_x(z), would improve readability.","section":"Theorem 1"}],"recommendation":"minor_revision","confidential_remarks":"This is a strong paper with complete, internally consistent proofs and a clear contribution. The only issues I found are local presentation points; no load-bearing technical concern emerged. The paper's heavy reliance on the author's own recentering framework is visible but the proofs do not depend on those companion papers, so I do not see a circularity problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result is real. Hull proves that for bounded potential outcomes, independent randomization and a recentered, intercept-free regression estimator achieve minimax MSE among all assignment mechanisms and a broad class of affine estimators. Complete randomization is strictly worse by an N/(N-1) factor, and paired randomization is twice as bad. Extending to arbitrary estimators, independent randomization remains optimal and the optimal estimator is characterized by a convex program. This is genuinely new relative to Aronow and Lopatto (2026), who fix marginal inclusion probabilities and restrict to design-unbiased estimators, and to Harshaw et al. (2024), who fix the Horvitz-Thompson estimator. The paper also connects cleanly to the Hodges and Lehmann (1982) sampling result. The proof structure is solid: adversarial averaging gives the lower bounds, symmetrization reduces to independent assignment, and the separate-convexity argument pins down the unrestricted minimax value. I checked the main steps and they balance; the appendix is self-contained and does not lean on the companion papers. The main soft spot is not technical but interpretive. The minimax criterion lets the adversary choose any configuration in [L,U]^2N, with no structure linking Y(0) and Y(1). That is the natural worst case, but the author himself notes in Section 4 that stratification based on informative observables will likely dominate in practice, and the gain over complete randomization is asymptotically negligible. So the practical prescription is less sharp than the abstract suggests. The numerical figure has no code, but the program is precisely specified and easily reproducible, so this is minor. Self-citations are present but not circular; they provide context, not premises. The paper deserves a serious referee. It is a well-argued, technically clean contribution to the minimax design literature. I would send it out and expect acceptance after minor revision, with the main discussion focused on how much weight to put on the pure worst-case criterion.","headline":"A clean finite-sample minimax result showing independent randomization beats complete randomization for bounded outcomes; the proofs check out, and the main caveat is the author's own worst-case framing.","tokens_in":12930,"tokens_out":1122,"would_cite":true,"duration_ms":12089,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62K05","62C20","62D05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that when potential outcomes are known to lie in $[L,U]$, the minimax way to estimate an average treatment effect randomizes each unit independently and uses a midpoint-centered recentered regression; balanced designs…","keywords":["bounded potential outcomes","average treatment effect","minimax experimental design","independent random assignment","complete randomization","paired randomization","finite population","recentered regression"],"falsifier":"For a small fixed $N$ (say $N=4$), enumerate all assignment distributions on $\\{0,1\\}^N$ and all affine estimators satisfying equivariance condition (2), and evaluate each at all $2^N$ no-effect endpoint configurations; if any pair's worst-case squared error is strictly below $(U-L)^2/N$, Proposition 1 is false. For the unrestricted claim, compute $\\kappa_N$ from (4) and search over measurable estimators via the least-favorable-prior representation; if any design-estimator pair beats $(U-L)^2\\kappa_N$, Theorem 1 is false. A separate unresolved check is the paper's conjecture that complete randomization is strictly suboptimal for general estimators: finding an even $N$ where complete randomization attains the independent-randomization value would weaken that subclaim but not the main theorem.","tokens_in":12040,"feed_emoji":"🎲","tokens_out":16623,"duration_ms":128772,"temperature":0.7,"pith_summary":"The paper studies how to run and analyze a randomized experiment when each unit's two potential outcomes are known to lie in a fixed interval $[L,U]$, as with binary outcomes. It claims that under a worst-case mean-squared-error criterion, the optimal design is independent random assignment with a random treatment share, not balanced complete randomization, because randomness in the treated share becomes informative once the outcome level is anchored. The optimal estimator in a broad affine class is an unconventional regression of the support-midpoint-centered outcome on the recentered treatment, with no intercept, attaining worst-case MSE $(U-L)^2/N$. Complete randomization costs at least $(U-L)^2/(N-1)$, and paired randomization with pair-fixed effects has twice the optimum. If these results are right, the usual advice to balance treatment shares needs a bounded-outcome caveat: balance is minimax only when outcome levels are unrestricted.","feed_headline":"Random share beats complete randomization when outcomes are bounded","feed_subtitle":"A midpoint-centered no-intercept regression attains worst-case MSE $(U-L)^2/N$; complete and paired designs do worse.","key_machinery":"The engine is the no-effect endpoint adversary: for each unit, the worst case can be taken to have both potential outcomes equal to either $L$ or $U$, so the outcome vector is $M + r s$ with $s\\in\\{-1,1\\}^N$ and $r=(U-L)/2$. Under such a schedule, the risk of an equivariant affine estimator becomes $r^2\\|b(D)\\|^2$, and the equivariance condition plus Cauchy-Schwarz forces $\\|b(D)\\|^2\\ge 4/N$. Independent randomization makes the assignment pairs uncorrelated, so no pattern of endpoint signs can align with assignment dependence to inflate risk; any negative pairwise dependence (as in complete or paired randomization) gives the adversary extra leverage. For the unrestricted measurable-estimator problem, the same endpoint reduction is carried by a symmetrization argument: averaging any design over coordinate swaps and permutations yields independent randomization and a symmetric estimator, and the minimax value $\\kappa_N$ is the optimum of the convex program (4) over the endpoint states $(k,m)$.","core_discovery":"Formally, fix $N$ units with $Y_i(0),Y_i(1)\\in[L,U]$ and target the finite-population average treatment effect $\\beta = N^{-1}\\sum_i(Y_i(1)-Y_i(0))$. Over all assignment mechanisms and affine estimators satisfying the equivariance condition $b(D)'(D-\\tfrac12\\mathbf{1})=1$ (which holds for difference-in-means, OLS with controls or fixed effects, and matching estimators), the minimax worst-case MSE is $(U-L)^2/N$. It is attained by independent Bernoulli$(1/2)$ assignment and $\\hat\\beta^* = \\frac{2}{N}\\sum_i(2D_i-1)(Y_i-M)$, where $M=(U+L)/2$: the no-intercept regression of $Y_i-M$ on $D_i-1/2$. Complete randomization with any equivariant affine estimator has worst-case MSE at least $(U-L)^2/(N-1)$, and difference-in-means attains that bound; paired randomization with pair-fixed effects has worst-case MSE $(U-L)^2/J$, exactly twice the optimum. The paper then removes the affine restriction: over all measurable estimators the minimax value is $(U-L)^2\\kappa_N$, where $\\kappa_N$ is the value of a finite convex program with quadratic constraints, and independent randomization remains optimal, with a generally nonlinear estimator that has a least-favorable-prior representation. Restricting to affine estimators without the equivariance condition, the best procedure shrinks $\\hat\\beta^*$ by $\\sqrt{N}/(\\sqrt{N}+1)$, giving minimax MSE $(U-L)^2/(\\sqrt{N}+1)^2$.","pith_inferences":["If bounds were unit-specific, the same midpoint-centering derivation would replace the common midpoint $M$ by each unit's own $(L_i+U_i)/2$ and likely preserve the optimality of independent assignment; the paper does not develop this case.","Adding a mild within-stratum homogeneity or Lipschitz bound to the minimax problem would create a threshold where paired randomization overtakes independent randomization; locating that threshold would quantify the cost of balance that the paper leaves open.","The numerical decline of the nonlinear estimator's gain with $N$ suggests the correction matters most for experiments with a few hundred units; for large samples the simple no-intercept recentered regression is nearly minimax. This extrapolation follows from the paper's Figure 1, not from a stated theorem."],"forward_implications":["Within the equivariant affine class, any completely randomized experiment has worst-case MSE at least $N/(N-1)$ times the minimax value, and paired randomization with pair-fixed effects has exactly twice the minimax value.","The minimax affine estimator is the no-intercept regression of the midpoint-centered outcome on the recentered treatment; difference-in-means is optimal only under complete randomization, where it attains the larger $(U-L)^2/(N-1)$ bound.","When the estimator is unrestricted, the minimax value is $(U-L)^2\\kappa_N$ with independent assignment, and the optimal estimator is nonlinear; numerically it reduces worst-case MSE by 15-30% for moderate $N$ relative to the best affine estimator.","Without the equivariance restriction, the best affine estimator shrinks $\\hat\\beta^*$ by $\\sqrt{N}/(\\sqrt{N}+1)$, with minimax MSE $(U-L)^2/(\\sqrt{N}+1)^2$, and a further nonlinear correction helps for every $N>2$.","The paper does not claim balance is always bad: when strata are chosen using informative covariates, it concludes that stratified or paired designs 'are likely to dominate' independent randomization, so the reversal applies to worst-case comparisons with uninformative strata."],"supporting_citations":[{"why":"Establishes the bounded-support midpoint-adjustment logic for finite-population totals and shows pairwise-independent inclusion indicators attain a sharp MSE bound, the closest precedent for this paper's treatment of ATEs.","marker":"Aronow and Lopatto (2026)"},{"why":"Provides the least-favorable-prior minimax argument for estimating a bounded-population mean under simple random sampling, which Theorem 1 adapts to the ATE problem.","marker":"Hodges and Lehmann (1982)"},{"why":"Proves complete randomization is minimax when potential outcomes are unrestricted, the baseline prescription that bounded support overturns.","marker":"Kallus (2021)"},{"why":"Gives minimax optimality of complete randomization and Neyman allocation under permutation invariance, the contrast for the reversal in Proposition 1.","marker":"Bai (2023)"},{"why":"Supplies the recentering principle for design-based identification that justifies the no-intercept recentered treatment regression in the optimal estimator.","marker":"Borusyak and Hull (2023)"},{"why":"Provides the close antecedent that, with a fixed Horvitz-Thompson estimator and treatment probability one-half, independent randomization minimizes worst-case MSE over an $\\ell^2$ bound.","marker":"Harshaw et al. (2024, Lemma 4.1 and Proposition 4.2)"}],"fun_headline_variants":["Random share beats complete randomization for bounded outcomes","Independent randomization optimal for bounded potential outcomes","Bounded outcomes? Randomize independently","Minimax design: random share beats complete randomization","For bounded outcomes, no-intercept regression is minimax"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result depends on evaluating designs against the worst possible configuration of bounded potential outcomes, with no restriction that units in the same pair or stratum have similar outcomes; the paper's own conclusion (Section 4) notes that this is 'likely overly pessimistic' when stratification uses informative covariates.","fun_headline_variants_meta":{"raw":{"variants":["Random share beats complete randomization for bounded outcomes","Independent randomization optimal for bounded potential outcomes","Bounded outcomes? Randomize independently","Minimax design: random share beats complete randomization","For bounded outcomes, no-intercept regression is minimax"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001512,"raw_usage":{"total_tokens":6113,"prompt_tokens":1053,"completion_tokens":5060,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":669,"completion_tokens_details":{"reasoning_tokens":5002}},"tokens_in":669,"tokens_out":5060,"duration_ms":32965,"temperature":1.0,"reasoning_tokens":5002,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T10:15:30.702000+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For a small fixed $N$ (say $N=4$), enumerate all assignment distributions on $\\{0,1\\}^N$ and all affine estimators satisfying equivariance condition (2), and evaluate each at all $2^N$ no-effect endpoint configurations; if any pair's worst-case squared error is strictly below $(U-L)^2/N$, Proposition 1 is false. For the unrestricted claim, compute $\\kappa_N$ from (4) and search over measurable estimators via the least-favorable-prior representation; if any design-estimator pair beats $(U-L)^2\\kappa_N$, Theorem 1 is false. A separate unresolved check is the paper's conjecture that complete randomization is strictly suboptimal for general estimators: finding an even $N$ where complete randomization attains the independent-randomization value would weaken that subclaim but not the main theorem.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the least-favorable-prior minimax argument for estimating a bounded-population mean under simple random sampling, which Theorem 1 adapts to the ATE problem."}],"review_version":1}