{"id":"9d90caed-2d43-40f0-9536-c0d8c44ac121","arxiv_id":"1908.04462","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For smooth strictly increasing signals, the bias of isotonic regression is at most C(log n/n)^{β/3}, with matching lower bounds up to log factors; flat signals make bias as large as the error.","lead":"This paper proves sharp bounds on the bias of isotonic regression, showing that for smooth strictly increasing signals the bias decays as n to the -2/3, faster than previously known. The result clarifies when the bias is negligible compared with the estimation error.","discovery_kind":"first_principles","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's parameter condition (25) is unsatisfiable for the stated construction, so the matching lower bound is not established as written.","rationale":"The upper-bound proof (Theorem 1) checks out: Step 2's use of the L0 gap is internally consistent, Step 3's linearization has zero bias only at the midpoint of an odd-length subvector (the symmetry argument is valid even though E[iso(Z)] is nonzero at boundaries), and the breakpoint lemma's variance condition is satisfied at the chosen scale m ≈ n^{2/3}/(log n)^{1/3} because m/n ≈ n^{-1/3}(log n)^{-1/3}. The strict-monotonicity boundary highlighted by the reader is real but is explicit in the theorem statements rather than a hidden assumption. The main defect I found is in Theorem 2's proof: condition (25) contains a lower bound that is impossible for the stated choice an = (L1+L0)/2 and incompatible with L1-Lipschitz. Since (25) is invoked before the breakpoint-lemma application, the lower-bound proof is not valid as written. However, the impossible inequality appears unused; only the upper bound an ≤ C4√(log n) is needed to ensure m is large enough, suggesting a one-line correction. This supports the existing CONDITIONAL verdict rather than changing it, and the proposed concrete test would confirm whether the fix is indeed that simple.","tokens_in":24303,"tokens_out":48912,"duration_ms":480941,"concrete_test":"Recompute the inequalities in Section 3.4 with an = (L1+L0)/2 and cn = n^{1/3}(log n)^{5/3}. Verify: (a) C3√(n log n) ≤ an fails for any fixed C3 > 0 as n → ∞; (b) the Step 4 bound P{k_i(Y_lin) < i + C6 n/cn} ≤ 0.5 follows using only b_n c_n ≤ a_n, C1 ≤ c_n ≤ C2 n, and a_n ≤ C4√(log n), with no lower bound on a_n. If (b) holds, delete the false lower-bound condition from (25) and confirm the rest of Theorem 2's proof goes through unchanged; if not, Theorem 2 requires a new construction and the paper's sharpness claim lacks support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Section 3.4, the proof of the matching lower bound (Theorem 2) assumes the parameter regime (25), which includes C3√(n log n) ≤ an. The proof then sets an = (L1+L0)/2, bn = min{M/2, (L1-L0)/2}·(n(log n)^5)^{-β/3}, and cn = n^{1/3}(log n)^{5/3}. For any fixed C3 > 0, the inequality C3√(n log n) ≤ (L1+L0)/2 fails for all sufficiently large n, and the paper even states that C3 should be chosen sufficiently large. Moreover, the L1-Lipschitz assumption (5) forces an/n ≤ L1/n, hence an ≤ L1, so an cannot grow with n to satisfy this condition. Since (25) is explicitly invoked before applying the breakpoint lemma to certify m ≥ log n/(C2(C4)^{2/3}), the proof of Theorem 2 as written is internally inconsistent. The theorem may still be true: the later arguments appear to need only the upper bound an ≤ C4√(log n) (with the stated cn), not the impossible lower bound, so a corrected (25) could restore the proof. But the sharpness claim is not currently established by the text.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the bias of the least-squares isotonic regression estimator for a monotone mean vector. Under a Lipschitz assumption (5), strict monotonicity (6), Hölder smoothness (7), subexponential noise (9), and a lower-bounded Lipschitz variance (10), Theorem 1 gives an entrywise bias bound of order (log n / n)^{β/3} for β ∈ [1,2) and ((log n)^5 / n)^{2/3} for β = 2, away from the endpoints. Theorem 2 constructs a signal satisfying the same conditions whose bias is at least a constant times n^{-β/3}(log n)^{-5β/3} on a constant fraction of indices, and Theorem 3 shows that if strict monotonicity is relaxed to monotonicity with a flat point, the worst-case bias can be as slow as n^{-α/(2α+1)}, approaching n^{-1/2}. The proofs combine Gaussian coupling via Sakhanenko, a breakpoint lemma for Gaussian isotonic regression, subvector restriction, and a linearization/symmetry step.","tokens_in":24556,"tokens_out":19088,"duration_ms":165365,"significance":"If correct, the paper provides a sharp characterization of the bias of one-dimensional isotonic regression: for smooth strictly increasing signals the bias vanishes relative to the n^{-1/3} estimation error, while for merely Lipschitz signals the bias can be of the same order as the error. The proof structure is attractive and largely modular, and the lower-bound constructions are natural. The paper also includes reproducible simulation code and a clear demonstration that strict monotonicity is essential. The main caveat is that the matching lower bound (Theorem 2) is not established by the proof as written because a stated parameter condition is unsatisfiable.","major_comments":[{"comment":"The parameter regime (25) includes the lower bound C3√(n log n) ≤ a_n. With the stated choice a_n = (L1+L0)/2, a constant independent of n, this inequality fails for all sufficiently large n for any fixed C3 > 0. The text states in Section 3.4 that the chosen parameters satisfy (25) and asserts this is trivial to verify; Section 3.4.4 then explicitly invokes (25) to certify m ≥ log n/(C2(C4)^{2/3}) and to control constants. Because the lower bound in (25) is impossible while the subsequent arguments appear to need only the upper bound a_n ≤ C4(c_n)^{3/2}/((log n)^2√n) — which is compatible with a_n constant, since c_n^{3/2}/((log n)^2√n) = (log n)^{1/2} — the proof of Theorem 2 is internally inconsistent as written. The theorem may be true after replacing (25) with the upper-bound condition, but the matching lower bound is not currently established by the text.","section":"Section 3.4, condition (25)"},{"comment":"The final sentence of Step 4 states that the required conditions hold 'recalling condition (25), with C3 chosen to be sufficiently large and C4 sufficiently small.' This is not meaningful for C3: the impossible lower bound C3√(n log n) ≤ a_n cannot hold for any choice of C3 when a_n is constant, and the displayed inequalities in Step 4 do not otherwise involve C3. This reinforces that the parameter condition must be revised before the lower-bound claim can be accepted.","section":"Section 3.4.4, Step 4"}],"minor_comments":[{"comment":"The abstract states that the result 'only requires a strictly monotone mean and that the noise distribution has subexponential tails,' but Theorem 1 also assumes the Lipschitz bound (5), the Hölder smoothness condition (7), the variance lower bound, and the Lipschitz variance condition (10). Please restate the assumptions accurately.","section":"Abstract"},{"comment":"The displayed implication in the proof of Lemma 3 has the wrong direction and an invalid index: the subvector Ỹ has length m, so the expression 'iso(Ỹ)_m < iso(Ỹ)_{m+1}' is undefined. The proof should use the contrapositive of property (15) and map a break at position i in the subvector to a break at position m in the original vector; for example, with Ỹ_j = Y_{m-i+j}, the correct implication is iso(Y)_m ≠ iso(Y)_{m+1} ⇒ iso(Ỹ)_i ≠ iso(Ỹ)_{i+1}.","section":"Lemma 3 proof"},{"comment":"In the paragraph introducing ˇY(i), the expression 'E[|iso(Ỹ(i))_{m+1} − iso(ˇY(i))_i|]' should read 'iso(ˇY(i))_{m+1}' rather than 'iso(ˇY(i))_i', since the subvector has length 2m+1 and its midpoint is index m+1; the subsequent equations consistently use the latter notation.","section":"Section 3.3.3"},{"comment":"The proof of Theorem 2 begins 'Fix any L1 > L0 ≥ 0' while the theorem statement requires L1 > L0 > 0. Please align the proof with the theorem statement.","section":"Section 3.4"},{"comment":"The notation '3√' for cube roots, for example in Step 2 of Section 3.3, is ambiguous in the text; using the explicit ∛{\\cdot} or ∛[3]{\\cdot} notation would improve readability.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The upper-bound theorem appears sound and is a substantive contribution. The main obstacle is the parameter inconsistency in the proof of Theorem 2, which is likely repairable by replacing condition (25) with the upper-bound condition actually needed. I would encourage the editor to request a revision rather than reject, provided the authors fix the stated condition and the Lemma 3 direction typo."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline result is real: the paper pins down the entrywise bias of isotonic regression on strictly increasing Hölder signals as n^{-β/3} up to logs, and shows strict monotonicity is what makes the fast rate possible. That is a genuine advance over the earlier n^{-7/15} (Banerjee–Durot–Sen) and n^{-1/2} (Durot) bounds. The three-step upper-bound proof — Gaussian coupling, subvector truncation, linearization with the zero-bias symmetry argument — is coherent, and the breakpoint lemma is a genuinely useful technique. The simulations back the scaling and the code is posted. The citation pattern is clean: the Yang–Barber bound is an external tool, not a circular dependency.\n\nThe soft spot is the matching lower bound. Theorem 2's construction chooses a_n = (L1+L0)/2, a constant, but condition (25) requires C3√(n log n) ≤ a_n, and the proof later says to take C3 'sufficiently large.' That inequality cannot hold for large n, and L1-Lipschitz actually forces a_n ≤ L1, so a_n cannot grow to satisfy it. Since (25) is explicitly invoked in Step 4 to certify m ≥ log n/(C2(C4)^{2/3}), the proof of Theorem 2 as written is internally inconsistent. The fix looks easy — the later steps appear to need only an upper bound on a_n, and the chosen constant a_n satisfies that — so the theorem is probably true, but the sharpness claim is not established in this version.\n\nThe two minor flags the reader raised are real but cosmetic: the abstract's 'only requires' phrasing omits the variance conditions in (10), and Lemma 3's displayed implication has the wrong subvector indices (should be i, i+1 rather than m, m+1). Neither affects the upper-bound argument.\n\nNet: the upper bound is the main contribution and it holds; the lower bound needs a corrected parameter condition. This is for statisticians in shape-constrained and order-restricted inference — anyone estimating monotone functions will want the sharp bias rate. A serious journal should send it to referees; the problem is classical, the advance is real, and the gap is identifiable and repairable. I'd ask the authors to fix (25) and re-check Step 4, after which it should be fine.","headline":"The n^{-β/3} bias rate for strictly increasing signals is a real advance and the upper-bound proof holds; the matching lower bound has a repairable but real gap in its parameter condition.","tokens_in":25109,"tokens_out":13401,"would_cite":true,"duration_ms":111030,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G08","62G20","62G05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proves sharp bounds on the bias of isotonic regression: for strictly increasing Hölder-smooth signals, interior bias is at most $(\\log n/n)^{\\beta/3}$, with matching lower bounds, so for smooth signals bias vanishes relative to…","keywords":["isotonic regression","bias","shape constraints","Hölder smoothness","strict monotonicity","pool adjacent violators","breakpoint lemma","nonparametric regression"],"falsifier":"Take a strictly increasing smooth mean such as $\\mu_i=i/n$, add i.i.d. Gaussian noise, and estimate $\\mathbb{E}[\\mathrm{iso}(Y)_{n/2}]$ by Monte Carlo for $n=1000,10000,100000$. The theorem predicts the absolute bias at this interior point decays like $n^{-2/3}$ up to logs, so a log-log regression of $|\\text{bias}|$ on $n$ should have slope near $-2/3$; observing a slope near $-1/2$ or no decay would contradict the claimed sharp rate.","tokens_in":24128,"feed_emoji":"📉","tokens_out":8061,"duration_ms":78886,"temperature":0.7,"pith_summary":"This paper characterizes the bias of isotonic regression, the difference between what the monotone least-squares fit averages to and the true signal. For a strictly increasing signal with Hölder smoothness exponent $\\beta$ between 1 and 2, it proves the bias at interior points is at most $(\\log n/n)^{\\beta/3}$ up to constants, and supplies matching lower bounds. For smooth signals ($\\beta=2$) the bias decays like $n^{-2/3}$ up to logs, so it is negligible next to the usual $n^{-1/3}$ estimation error; for merely Lipschitz signals ($\\beta=1$) the bias can be as large as the error itself. The result requires only strict monotonicity and subexponential noise, not constant variance or symmetric noise.","feed_headline":"Bias of isotonic regression shrinks as n^-2/3 on smooth signals","feed_subtitle":"Strictly increasing smooth means make the estimator's bias vanish relative to its error; merely Lipschitz means do not.","key_machinery":"The load-bearing tool is a breakpoint lemma: if the means and variances are nearly constant on a window of length $2m$ around index $i$, then the probability that isotonic regression creates a block boundary at $i$ is at most $C(\\log m)/m$. This is proved using the min-max formula for the isotonic projection, the pool-adjacent-violators truncation properties, and a classical result on the number of constant blocks in isotonic regression of standard Gaussian noise. The upper-bound proof then works in three steps: couple the noise to a Gaussian with the same means and variances; truncate to a subvector of length roughly $n^{2/3}$ centered at $i$; and compare with a linearly increasing, constant-variance subproblem whose midpoint bias is zero by symmetry. Smoothness controls the error of the linear approximation, while the breakpoint lemma controls how often truncation changes the estimate.","core_discovery":"The paper establishes that when observations satisfy independent subexponential noise and the mean is Lipschitz, strictly increasing, and $(\\beta,M)$-Hölder smooth, the isotonic projection $\\hat{\\mu}=\\mathrm{iso}(Y)$ obeys $|\\mathbb{E}[\\hat{\\mu}_i]-\\mu_i| \\le C(\\log n/n)^{\\beta/3}$ for $1\\le \\beta<2$ and $\\le C((\\log n)^5/n)^{2/3}$ for $\\beta=2$, for all interior indices bounded away from the endpoints. A matching lower bound constructs means satisfying the same assumptions for which at least a constant fraction of indices have bias at least $C' n^{-\\beta/3}(\\log n)^{-5\\beta/3}$, showing the exponent $\\beta/3$ is tight up to logarithmic factors. A further result shows that if the mean is merely nondecreasing and is flat at one point, the bias can be as large as $n^{-\\alpha/(2\\alpha+1)}$ up to logs, which approaches $n^{-1/2}$ as $\\alpha\\to\\infty$; strict monotonicity is therefore essential to the fast rate.","pith_inferences":["If the bias for smooth strictly increasing signals is really $O(n^{-2/3}\\mathrm{polylog})$, then debiasing isotonic regression may be unnecessary for such signals: the dominant error is stochastic, and standard error-based confidence intervals could be correctly centered.","The same combination of a breakpoint lemma and a local linear approximation may transfer to related shape-constrained estimators such as the Grenander estimator of a monotone density, where the analogous bias-variance split is still open.","The lower-bound construction with a linear mean plus a high-frequency sine oscillation suggests a concrete diagnostic: fit isotonic regression on a smooth increasing trend with a small superimposed oscillation and measure bias at the oscillation peaks; it should scale as the amplitude $n^{-\\beta/3}$.","For non-Gaussian noise with $\\beta=2$, the paper's bound carries an extra $\\log^5 n$ factor that it calls likely an artifact; a sharper proof removing it may be testable by simulation with subexponential non-Gaussian noise."],"forward_implications":["For smooth strictly increasing signals, bias decays as $n^{-2/3}$ up to logs while the stochastic error is of order $n^{-1/3}$, so the bias is asymptotically negligible relative to the noise.","For merely Lipschitz signals, the bias can be of order $n^{-1/3}$, the same size as the error, so the estimator can be systematically miscentered unless smoothness is assumed.","The $n^{-\\beta/3}$ rate is tight for every $\\beta\\in[1,2]$ up to logs, meaning no strictly increasing smoothness class with exponent $\\beta$ can guarantee a faster interior bias rate in the worst case.","The older $n^{-1/2}$ bias bound under constant variance is not the general picture: with nonconstant subexponential noise the worst-case bias is slower, and the present bounds do not require constant variance.","Strict monotonicity is necessary for the fast rate: a flat point in an otherwise nondecreasing signal can push the bias up to $n^{-\\alpha/(2\\alpha+1)}$, which for smooth flat signals is $n^{-2/5}$, much larger than the strict-case $n^{-2/3}$."],"supporting_citations":[{"why":"Supplies the min-max formula (12) and the PAVA truncation properties (15) and (16) used throughout the proofs.","marker":"Barlow et al. [1972]"},{"why":"Provides the high-probability entrywise error bound (2) and the contraction lemma used to compare isotonic projections of coupled vectors.","marker":"Yang and Barber [2018]"},{"why":"Establishes the minimax $n^{-1/3}$ error rate that frames the slowest possible bias rate at $\\beta=1$.","marker":"Chatterjee et al. [2015]"},{"why":"Provides the Gaussian coupling with exponential moments used in Lemma 4 to reduce general subexponential noise to Gaussian noise.","marker":"Sakhanenko [1985, Theorem 1]"},{"why":"Gives the distribution of the number of blocks in isotonic regression of standard Gaussian noise, which is the core of the breakpoint lemma.","marker":"Andersen [1954]"},{"why":"Supplies the flat-point condition and the $n^{-\\alpha/(2\\alpha+1)}$ error scaling that Theorem 3 matches as a bias lower bound.","marker":"Wright [1981]"},{"why":"Is the previous best general bias upper bound ($n^{-7/15}$) that this paper improves.","marker":"Banerjee et al. [2019]"},{"why":"Is the earlier $n^{-1/2}$ bias bound under constant variance, which the present results place in context.","marker":"Durot [2002]"},{"why":"Supplies the chi-square tail bound used in the appendix to compare a non-constant-variance Gaussian to a constant-variance Gaussian.","marker":"Laurent and Massart [2000, Lemma 1]"}],"fun_headline_variants":["Isotonic regression bias: sharp n^{-β/3} bound","Bias of isotonic regression shrinks at rate n^{-β/3}","Tight bias rate for isotonic regression: n^{-β/3}","Isotonic bias: n^{-β/3} up to logs, no symmetry needed","Strict monotonicity gives fast bias decay in isotonic regression"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's rate relies on the mean being strictly increasing everywhere, with a positive minimum step between adjacent values; if the mean is merely nondecreasing and flattens at one point, the bias can be much larger.","fun_headline_variants_meta":{"raw":{"variants":["Isotonic regression bias: sharp n^{-β/3} bound","Bias of isotonic regression shrinks at rate n^{-β/3}","Tight bias rate for isotonic regression: n^{-β/3}","Isotonic bias: n^{-β/3} up to logs, no symmetry needed","Strict monotonicity gives fast bias decay in isotonic regression"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000836,"raw_usage":{"total_tokens":3616,"prompt_tokens":886,"completion_tokens":2730,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":502,"completion_tokens_details":{"reasoning_tokens":2628}},"tokens_in":502,"tokens_out":2730,"duration_ms":20399,"temperature":1.0,"reasoning_tokens":2628,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:43:56.069166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a strictly increasing smooth mean such as $\\mu_i=i/n$, add i.i.d. Gaussian noise, and estimate $\\mathbb{E}[\\mathrm{iso}(Y)_{n/2}]$ by Monte Carlo for $n=1000,10000,100000$. The theorem predicts the absolute bias at this interior point decays like $n^{-2/3}$ up to logs, so a log-log regression of $|\\text{bias}|$ on $n$ should have slope near $-2/3$; observing a slope near $-1/2$ or no decay would contradict the claimed sharp rate.","supporting_citations":[],"review_version":1}