{"id":"c758eddd-cd0f-42bd-a5a6-cf3aa790b239","arxiv_id":"2509.11381","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Greedy causal trees cannot uniformly estimate heterogeneous treatment effects faster than any polynomial rate, even in a constant-effect randomized benchmark, and honesty only removes a log-log factor.","lead":"This paper proves that causal decision trees grown by greedy splitting can have very slow worst-case accuracy, even when the true treatment effect is flat and treatment is randomized. It explains why common 'honesty' fixes do not save them, and why standard theoretical assumptions for causal forests do not match how these trees are actually built.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 1–3 include the SSE estimator under Assumption 1, but every SSE proof in the supplement requires V[ε_i(0)] = V[ε_i(1)], an assumption absent from the main text; Remark SA-2 explicitly leaves the unequal-variance case as a conjecture.","rationale":"The paper's central claim is the impossibility of polynomial uniform rates for causal trees, stated uniformly for DIM, IPW, and SSE, and for NSS, HON, and X-adaptive designs. The missing equal-variance condition is load-bearing because it is precisely an assumption in the proof of the SSE versions of the main theorems, yet it is absent from Assumption 1 and from the theorem statements. The authors are transparent in Remark SA-2 that the unequal-variance case is conjectural, which makes the issue unambiguous: as submitted, the SSE half of the main results is not proven for part of the stated parameter space. This warrants a CONDITIONAL verdict—the paper should be revised either to add V[ε_i(0)] = V[ε_i(1)] to Assumption 1 (or the relevant theorems) or to supply a proof for the unequal-variance case. The reader's identified weakest assumption, covariate independence, is less directly damaging because independence is explicitly assumed; my concern is an internal mismatch rather than a boundary of the stated assumptions. I also note a secondary issue in Lemma SA-28 whose stated limsup condition (ρ_n log log n → 0) conflicts with the parameter choice ρ_n ≫ 1/log log n in the proof of Theorem SA-19, but the equal-variance gap is more central to the paper's headline claims.","tokens_in":75066,"tokens_out":13541,"duration_ms":160236,"concrete_test":"For p=1 with constant effects, take ε(0) ~ N(0,1), ε(1) ~ N(0,4), ξ=1/2, τ=0, n=10^4–10^5, and simulate the SSE split index at the root over many replications. Estimate P(î ∈ [n^a, n^b]) for, say, a=0.1, b=0.4. If the probability is materially below the equal-variance lower bound (b−a)/(2e) ≈ 0.055, Theorem 1's SSE claim fails without the equal-variance condition. If the probability matches, the gap is proof-only, but the theorem still needs restating or a proof extension. Analytically, re-derive Step 5 of SA-4.30 without assuming equal variances: if the two Ornstein–Uhlenbeck components cannot be normalized to be identically distributed, Horváth's Lemma 2.1 does not apply and Remark SA-2's conjecture remains open.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The main text states Theorem 1 for l ∈ {DIM, IPW, SSE} under Assumption 1, which only requires E[ε_i(d)^2] > 0 for each d, with no equality of variances. The supplement's SSE results, however, all add `with V[ε_i(0)] = V[ε_i(1)]`: Theorem SA-29 (imbalanced splits), Corollaries SA-30/31/33/35 (stump, deep trees, honest, X-adaptive), and the L2 results SA-32/34/36. Remark SA-2 says this condition is needed to apply the two-dimensional Darling-Erdős theorem and merely conjectures that the conclusion holds without it. Since Assumption 1 does not impose equal error variances, Theorems 1–3 as stated assert results for SSE in regimes the supplied proof does not cover. This is an internal mismatch between the theorem statements and the proof, not a matter of consensus or of an unstated but harmless regularity condition: the authors themselves flag the missing case as open. The DIM/IPW arguments are unaffected, but the SSE component—one of the three splitting rules in the paper's central claim—is only established under an extra assumption the main text never states.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the uniform and integrated accuracy of causal tree estimators based on the Athey--Imbens recursive partitioning framework, covering DIM, IPW, and SSE splitting, with no sample splitting (NSS), honesty (HON), and X-adaptive versions. Under a constant-treatment-effect DGP with independent continuous covariates and sub-exponential errors, the authors prove lower bounds showing that these estimators cannot attain polynomial-in-n uniform convergence rates, with explicit non-vanishing probability bounds involving constants such as b/e. They also prove that X-adaptive trees of depth at least proportional to log log n are uniformly inconsistent, while integrated squared error can be small. The proofs, given in a long supplement, use empirical process approximations, high-dimensional CLTs, Gaussian comparison inequalities, and Ornstein-Uhlenbeck process calculations, and they correct an error in Eicker (1979). The paper includes a simulation study.","tokens_in":75452,"tokens_out":6347,"duration_ms":73123,"significance":"If the results hold as stated, this is an important negative result for a widely used methodology: it shows that canonical causal tree implementations need not satisfy the balanced-cell conditions behind existing consistency and inference guarantees, and that honesty alone does not cure the problem. The explicit lower-bound probabilities and the contrast between uniform and L2 behavior are valuable, as are the L2 upper bounds and the clarification of the role of α-regularity. The supplement is unusually detailed, with many named technical tools and explicit proof steps, and the correction to Eicker (1979) is a useful by-product. The main caveat is that the SSE component of the central theorems is only proved in the supplement under an equal-variance condition that is absent from the main theorem statements and is explicitly left as a conjecture for the general case.","major_comments":[{"comment":"The main text states Theorem 1 and Theorem 2 for l ∈ {DIM, IPW, SSE} under Assumption 1, which only requires E[ε_i(d)^2] > 0 for d = 0,1 and does not impose equality of variances. However, every SSE result in the supplement adds the condition V[ε_i(0)] = V[ε_i(1)]: Theorem SA-29, Corollaries SA-30/31/33/35, and the L2 bounds SA-32/34/36. Remark SA-2 explicitly states that this condition is needed to apply a two-dimensional Darling–Erdős theorem and leaves the unequal-variance case as a conjecture. Thus Theorems 1 and 2 assert SSE results in a regime for which no proof is supplied. This is load-bearing because SSE is one of the three splitting rules in the paper's central claim. The authors should either state the equal-variance condition in the main theorems for SSE (and adjust the abstract and discussion accordingly) or prove the missing case; DIM and IPW are not affected by this issue.","section":"§SA-3.3, Theorem SA-29; main Theorems 1–2"},{"comment":"Lemma SA-18 is stated under the condition lim sup_n ρ_n log log n = 0, with s_n = exp((log n)^{ρ_n}). But the proof of Theorem SA-19 chooses ρ_n satisfying log log log log n / log log n ≪ ρ_n ≪ 1, which implies ρ_n log log n → ∞, not 0. The same proof relies on the resulting truncation bound to restrict the argmax to [s_n, n-s_n], so as written Lemma SA-18 does not apply. Lemma SA-28, the SSE analogue, has a different condition (lim sup ρ_n log log n = ∞), which is also not the 'ρ_n log log n → ∞' condition used in its proof. This is likely a typographical/condition error, but it must be corrected because Theorem SA-19 feeds directly into the proof of Theorem 1 for the DIM estimator.","section":"§SA-4.19–SA-4.20, Lemma SA-18"}],"minor_comments":[{"comment":"The proof of Lemma SA-17 contains the sentence 'The assumption that lim inf_n ρ_n log log n = ∞ implies lim inf r_n = ∞', but Lemma SA-17 has no ρ_n and only assumes r_n → ∞. This looks like a leftover from Lemma SA-18 and should be removed or rewritten.","section":"§SA-4.18"},{"comment":"Theorem SA-25 states the condition lim inf K_n / log log(K_n) > 0, while the main-text Theorem 3 uses lim inf K_n / log log n = κ > 0. The two statements should be aligned, or the difference should be explained.","section":"§SA-3.2.3, Theorem SA-25"},{"comment":"The text says 'The second is consistent with the X-results of Theorem 1', but X-adaptive results are presented in Theorem 3. Please correct the cross-reference.","section":"§7"},{"comment":"Some references contain broken markup: 'Chernozhuokov' (should be 'Chernozhukov') in the main text and references, and 'Lata/suppress la' appears in place of 'Latała'. These should be fixed in the final version.","section":"References"},{"comment":"The independence of covariate components in Assumption 1(iv) is used to obtain the factor 1/p in equation (6) and the asymptotic independence of split coordinates. It would be helpful to state explicitly that the quantitative lower bounds are not proven for dependent covariates, since the multivariate coupling argument in the supplement relies on this assumption.","section":"Assumption 1(iv)"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the mismatch between the main theorems and the supplement for the SSE estimator. The paper's most broadly stated claims are stronger than what the supplied proofs establish. If the authors restrict the SSE parts of Theorems 1 and 2 to the equal-variance case and correct the truncation-lemma condition, the remaining contributions appear substantial and publishable. The DIM/IPW results and the X-adaptive inconsistency results are the core strengths."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should read this one. The DIM/IPW part is real: for a constant-effect benchmark with randomized assignment, honest and non-honest causal trees with DIM or IPW splitting have uniform error slower than any polynomial in n, with explicit constants and a clean proof of the end-cut mechanism. That settles the authors' old conjecture, shows honesty buys almost nothing in the sup-norm, and explains why alpha-regularity guarantees for causal forests do not apply to CART implementations. The IMSE upper bounds in the same setting are a useful contrast, and the correction to Eicker (1979) is a nice byproduct. The proof strategy is coherent, and the supplement is detailed.\n\nThe soft spot is the SSE part. Theorems 1 and 2 in the main text are stated for l ∈ {DIM, IPW, SSE} under Assumption 1, which only requires positive error variances. But every SSE result in the supplement, from Theorem SA-29 through the relevant corollaries, adds V[ε_i(0)] = V[ε_i(1)], and Remark SA-2 openly says the unequal-variance case is left as a conjecture. That is not a stylized assumption hidden in a lemma; it is an internal mismatch between what the main text claims and what the supplied proofs deliver. The DIM/IPW results are untouched, so the paper's central message can survive, but the authors either need to prove the vanishing-variance case or restrict the main theorems to the condition they can actually prove.\n\nA smaller technical issue: Lemma SA-18 states a condition on ρ_n (limsup ρ_n log log n = 0) that is violated by the choice of ρ_n made in the proof of Theorem SA-19, where ρ_n log log n actually diverges. That looks like a typo or a stale condition; it should be fixed, but it is not in the same league as the SSE gap. Also, no code is shipped for the simulations, though the simulation design is simple enough that this is not a serious concern.\n\nBottom line: this deserves a serious referee. The DIM/IPW negative result is important for people working on tree-based HTE estimation, and the proof structure is worth engaging with even if the SSE gap requires revision. I would cite the DIM/IPW result and bring it to a reading group, with a warning to read the supplement carefully before trusting the SSE claims.","headline":"A serious, technically substantial negative result for DIM/IPW causal trees, but the main theorems overclaim for SSE splitting: the supplement only proves the SSE case under an extra equal-variance condition that the main text never states.","tokens_in":75846,"tokens_out":6483,"would_cite":true,"duration_ms":74118,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G20","62G08"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper establishes lower bounds showing that standard recursive-partitioning causal tree estimators—difference-in-means, inverse-probability-weighted, or squared-error splitting, with or without honesty—cannot achieve polynomial-in-samp","keywords":["causal trees","heterogeneous treatment effects","recursive partitioning","CART","honesty","uniform convergence","end-cut preference","lower bounds"],"falsifier":"Simulate the constant-effect randomized benchmark with p=2 independent uniform covariates and n=10^5, recording the root split fraction i-hat/n over many replications. Theorem 1 implies liminf of P(i-hat <= n^b or i-hat >= n-n^b) >= b/e for every b in (0,1); if these boundary-tail probabilities decay to zero as n grows, the paper's central lower bound is wrong.","tokens_in":75035,"feed_emoji":"📉","tokens_out":7992,"duration_ms":89510,"temperature":0.7,"pith_summary":"Greedy decision trees are among the most widely used tools for estimating heterogeneous treatment effects, so this paper asks how accurately they can really estimate the conditional average treatment effect function. In the simplest possible benchmark—randomized treatment, constant true effect, independent continuous covariates—the paper proves that the uniform (worst-case over covariate values) error of nine standard causal-tree estimators decays more slowly than any power of the sample size. The mechanism is the 'end-cut preference': with nonvanishing probability, the first greedy split falls very close to the boundary of a node, so one child contains very few observations and the estimator has huge variance there. Sample splitting ('honesty') removes only a sqrt(log log n) factor from the lower-bound rate and does not restore polynomial convergence, and X-adaptive trees become uniformly inconsistent once depth grows like log log n. These findings matter because they show that average accuracy—which can be near-parametric—is not informative about how well the tree estimates the effect in any particular subpopulation, and they undercut the balanced-split assumptions behind existing asymptotic guarantees for causal forests.","feed_headline":"Uniform error of causal trees falls slower than any power of n","feed_subtitle":"Even with a constant treatment effect, greedy splits starve some leaves of data; honesty barely helps.","key_machinery":"The load-bearing object is the first-split index i-hat, the number of observations in the left child at the root split. The proof's engine is an approximation of the split-criterion process (a partial-sum empirical process indexed by split location and coordinate) by the square of a time-transformed Ornstein-Uhlenbeck (stationary Gaussian) process, achieved through high-dimensional Gaussian coupling, Gaussian comparison inequalities, Gaussian correlation inequalities, and a corrected Darling-Erdos limit. This yields the imbalanced-split probability bound P(n^a <= i-hat <= n^b or n-n^b <= i-hat <= n-n^a) >= (b-a)/(2pe) for each coordinate, which is then applied recursively to deeper nodes. Th","core_discovery":"On the paper's own terms, the central claim is Theorem 1: in a randomized benchmark with constant treatment effect, independent continuous covariates, and sub-exponential errors, the DIM, IPW, and SSE causal tree estimators—with or without honesty—each have uniform error at least C n^{-b/2} sqrt(log log n) (NSS) or C n^{-b/2} (HON) with probability bounded away from zero, for every b in (0,1); no polynomial-in-n sup-norm rate is possible. The root cause is an imbalanced-split bound: at the first split, the probability that a child contains between n^a and n^b observations is at least (b-a)/(2pe). Greedy recursion propagates such small cells through the tree, and local averaging inside them p","pith_inferences":["Editorial inference: if the same end-cut mechanism operates in forests built from unregularized CART base learners, uniform confidence intervals for heterogeneous effects from those ensembles could inherit the slow convergence; the paper does not analyze forests, so this is an extrapolation.","Editorial inference: a direct diagnostic for practitioners is the empirical distribution of the first-split fraction in a null-effect simulation; the theorem predicts probability mass accumulating at 0 and 1 with magnitude governed by b/e, and a flat histogram would contradict the mechanism.","Editorial inference: the lower bound suggests that simple regularizations such as minimum leaf size or balance constraints can restore polynomial uniform rates, but only by introducing smoothing bias in high-curvature regions—a trade-off the paper notes but does not quantify."],"forward_implications":["In the constant-effect benchmark, no DIM, IPW, or SSE causal tree, honest or not, can achieve uniform error o_p(n^{-b/2}) for all b>0; at least a b/e fraction of samples exhibit superslow local errors.","Honesty only removes the sqrt(log log n) factor from the lower bound; the uniform rate remains slower than polynomial, so sample splitting is not a cure for this inaccuracy.","X-adaptive trees—those sharing covariates across splits but refreshing outcomes—are uniformly inconsistent whenever depth K_n grows like log log n, meaning even very shallow trees can break worst-case accuracy at realistic sample sizes.","Integrated squared error bounds remain near-parametric, so evaluations based on average error will not detect the local inaccuracy that Theorem 1 establishes.","Balanced-cell assumptions such as alpha-regularity are incompatible with standard CART-type splitting in the constant-effect benchmark; existing polynomial-rate results for causal forests therefore do not apply to these implementations."],"fun_headline_variants":["Causal trees flunk worst-case accuracy even with constant effects","Greedy splits doom causal trees to snail's-pace convergence","Honesty can't save causal trees from slow uniform error","Causal trees' average accuracy masks severe local blind spots","Even constant effects can't beat causal trees' slowest rates"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The proof's multivariate lower bound relies on Assumption 1(iv), that the covariate components are independent and continuously distributed: this independence is what lets the split-coordinate criteria be treated as asymptotically independent, producing the 1/p factor in the bound; with dependent covariates the quantitative bound is unproven, though the univariate end-cut mechanism may persist.","fun_headline_variants_meta":{"raw":{"variants":["Causal trees flunk worst-case accuracy even with constant effects","Greedy splits doom causal trees to snail's-pace convergence","Honesty can't save causal trees from slow uniform error","Causal trees' average accuracy masks severe local blind spots","Even constant effects can't beat causal trees' slowest rates"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000145,"raw_usage":{"total_tokens":1031,"prompt_tokens":772,"completion_tokens":259,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":174}},"tokens_in":516,"tokens_out":259,"duration_ms":3732,"temperature":1.0,"reasoning_tokens":174,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:47:47.155506+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate the constant-effect randomized benchmark with p=2 independent uniform covariates and n=10^5, recording the root split fraction i-hat/n over many replications. Theorem 1 implies liminf of P(i-hat <= n^b or i-hat >= n-n^b) >= b/e for every b in (0,1); if these boundary-tail probabilities decay to zero as n grows, the paper's central lower bound is wrong.","supporting_citations":[],"review_version":1}