Pith. sign in

REVIEW 2 major objections 5 minor 22 references

The bias of isotonic regression

T0 review · 2 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 1908.04462 v2 pith:XEA5DKOK submitted 2019-08-13 math.ST stat.TH

classification math.STstat.TH MSC 62G0862G2062G05
keywords isotonicregressionbiasshapeconstraintsHöldersmoothnessstrictmonotonicitypooladjacentviolatorsbreakpointlemmanonparametric
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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 $n^{-2/3}$.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

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.

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 (2)
  1. [Section 3.4, condition (25)] 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.
  2. [Section 3.4.4, Step 4] 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.
minor comments (5)
  1. [Abstract] 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.
  2. [Lemma 3 proof] 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}.
  3. [Section 3.3.3] 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.
  4. [Section 3.4] 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.
  5. [Throughout] 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.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity: the bias rates are derived from external coupling and breakpoint theorems, with self-citations used only as independent tools.

full rationale

The paper's derivation chain is self-contained. Theorem 1 first reduces to Gaussian noise via Sakhanenko's coupling theorem (external), then the breakpoint lemma is proved from Andersen's classical distribution of the number of blocks (Lemma 2) plus a likelihood-ratio comparison (Lemma 5). Step 3 approximates the mean by a linear vector and uses symmetry to identify zero bias in the approximating problem; the target rate emerges from the choice of block length m and the Holder exponent beta, rather than being inserted by construction. No fitted parameter is renamed as a prediction, and no known result is merely relabeled. The self-citations to Yang and Barber [2018] appear only as tools: equation (2) supplies a high-probability uniform error bound under Lipschitz and subgaussian assumptions, and their Lemma 1 is the standard 1-Lipschitz property of the isotonic projection. These results do not contain the bias rate, so they constitute independent support rather than a circular input. Theorem 2's lower bound explicitly constructs an oscillating mean and proves the bias bound through the breakpoint lemma, with bn chosen to give the desired rate; no part of the lower bound is defined in terms of the conclusion. The reviewer's concern about condition (25) in Section 3.4, where C3 sqrt(n log n) <= an fails for the chosen constant an, is a proof-gap or internal-consistency issue affecting the written justification of Theorem 2, not a circularity: it does not amount to assuming the target rate or to fitting a parameter to the predicted quantity. Overall, the central claims are not forced by definition or by a self-citation chain.

Assumptions & free parameters 0 free parameters · 8 assumptions · 0 invented entities

The paper introduces no fitted parameters and no new entities. The theorem bounds depend on assumption parameters (L0,L1,M,λ,τ,Lσ,σ_min) that are inputs, not fitted. Both upper and lower bounds are proven from structural properties and external theorems.

assumptions (8)
  • standard math Isotonic projection min-max formula and PAVA block properties (Barlow et al. (1.9), (15), (16))
    Used throughout Section 3 to represent iso(Y)_i and to justify truncating to subvectors.
  • standard math Andersen's distribution of the number of blocks of isotonic regression of iid standard Gaussian noise
    Core of Lemma 3 and hence the breakpoint lemma.
  • standard math Sakhanenko's Gaussian coupling theorem for sums with exponential moments
    Step 1 of Theorem 1 reduces subexponential noise to Gaussian noise.
  • standard math Laurent-Massart chi-square concentration inequality
    Used in Lemma 5's likelihood-ratio comparison of Gaussians.
  • standard math Yang and Barber's uniform high-probability error bound for isotonic regression
    Used in Step 2 to bound the probability that a subvector truncation changes iso(Y)_i. One author overlaps, but the cited result is a published theorem independent of the present claim.
  • domain assumption Strict monotonicity of means, assumption (6)
    Needed to control the subvector truncation in Step 2; Theorem 3 shows it cannot be dropped.
  • domain assumption Hölder smoothness of means, assumption (7)
    Controls the linearization error in Step 3 and the lower-bound construction.
  • domain assumption Subexponential tails (9) and variance conditions (10)
    Used for Gaussian coupling and for the breakpoint lemma's approximate-constancy conditions.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The bias of isotonic regression." pith.science (2026). https://pith.science/paper/XEA5DKOK

@misc{pith2026190804462,
  author       = {Pith},
  title        = {Pith review of: The bias of isotonic regression},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XEA5DKOK}},
  note         = {Machine review of arXiv:1908.04462}
}
abstract

We study the bias of the isotonic regression estimator. While there is extensive work characterizing the mean squared error of the isotonic regression estimator, relatively little is known about the bias. In this paper, we provide a sharp characterization, proving that the bias scales as $O(n^{-\beta/3})$ up to log factors, where $1 \leq \beta \leq 2$ is the exponent corresponding to H{\"o}lder smoothness of the underlying mean. Importantly, this result only requires a strictly monotone mean and that the noise distribution has subexponential tails, without relying on symmetric noise or other restrictive assumptions.

Figures

Figures reproduced from arXiv: 1908.04462 by the authors.

Figure 1
Figure 1. Top: illustration of the mean vectors for the simulation, for the case n = 10000. The indices for which we measure the bias are indicated in each plot. Bottom: simulation results, plotting bias against n (each on the log scale), for the smooth and non-smooth case. The line and slope in each plot are the least-squares regression line (for log(|bias|) regressed on log(n)). For both the smooth and non-smooth mean, writ… view at source ↗
Figure 2
Figure 2. Illustration of the proof idea for Lemma 3. A break between indices i and i + 1 in the isotonic regression of Y˜ , corresponds to a break between indices m and m + 1 for Y . Finally, the proof of Lemma 1 for the general case is established by comparing the distribution of Y to the distribution of a standard Gaussian random vector, using our assumptions on the low variability in the µj ’s and σj ’s near the index i. … view at source ↗
Figure 3
Figure 3. Illustration of the linear approximation (Step 3 of the proof of Theorem 1, with the distribution of Y˜ (i) shown in black, and Yˇ (i) in red. The means µj and ˇµj are drawn as solid black and red lines, respectively, while the standard deviations of Z˜ j and Zˇ j are represented by the black and red shaded regions. Combining (21) with (20) we have P n iso(Y˜ )i 6= iso(Y˜ (i) )m+1o ≤ 1/n2 . Returning to our work abo… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Illustration of the construction for the lower bound result for smoothness (The￾orem 2). The figure on the left illustrates the linear mean µ lin, while the figure on the right illustrates the oscillating mean µ = µ lin + bn∆. • Step 3: We will show that, for all indic…
Figure 5
Figure 5. Figure 5: Illustration of the construction for the lower bound result for strict monotonicity (Theorem 3). The figure illustrates the two different signal vectors µ (0) and µ (1) used in the construction for the proof. • Step 2: We will show that If ki0 (Y (1)) ≥ i0 + 10nn then…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages

  1. [1]

    E. S. Andersen. On the fluctuations of sums of random variables ii. Mathematica Scandinavica, 2: 0 194--222, Dec. 1954. doi:10.7146/math.scand.a-10407. URL https://www.mscand.dk/article/view/10407

  2. [2]

    Banerjee, C

    M. Banerjee, C. Durot, and B. Sen. Divide and conquer in nonstandard problems and the super-efficiency phenomenon. Ann. Statist., 47 0 (2): 0 720--757, 2019. ISSN 0090-5364. doi:10.1214/17-AOS1633. URL https://doi.org/10.1214/17-AOS1633

  3. [3]

    R. E. Barlow, D. J. Bartholomew, J. Bremner, and H. Brunk. Statistical inference under order restrictions: The theory and application of isotonic regression. Wiley New York, 1972

  4. [4]

    D. J. Bartholomew. A test for homogeneity for ordered alternatives. Biometrika, 46: 0 36--48, 1959a

  5. [5]

    D. J. Bartholomew. A test for homogeneity for ordered alternatives ii. Biometrika, 46: 0 328--335, 1959b

  6. [6]

    H. D. Brunk. Maximum likelihood estimates of monotone parameters. Annals of Mathematical Statistics, 26: 0 607--616, 1955

  7. [7]

    H. D. Brunk. Estimation of isotonic regression. In Nonparametric T echniques in S tatistical I nference ( P roc. S ympos., I ndiana U niv., B loomington, I nd., 1969) , pages 177--197. Cambridge Univ. Press, London, 1970

  8. [8]

    E. Cator. Adaptivity and optimality of the monotone least-squares estimator. Bernoulli, 17 0 (2): 0 714--735, 05 2011. doi:10.3150/10-BEJ289. URL https://doi.org/10.3150/10-BEJ289

Show all 22 references
  1. [9]

    Chatterjee, A

    S. Chatterjee, A. Guntuboyina, and B. Sen. On risk bounds in isotonic and other shape restricted regression problems. The Annals of Statistics, 43: 0 1774--1800, 08 2015. doi:10.1214/15-AOS1324

  2. [10]

    de Leeuw, K

    J. de Leeuw, K. Hornik, and P. Mair. Isotone optimization in r: Pool-adjacent-violators (pava) and active set methods. Journal of Statistical Software, 32: 0 1--24, 2009

  3. [11]

    C. Durot. Sharp asymptotics for isotonic regression. Probability Theory and Related Fields, 122 0 (2): 0 222--240, Feb 2002. ISSN 1432-2064. doi:10.1007/s004400100171. URL https://doi.org/10.1007/s004400100171

  4. [12]

    Durot, V

    C. Durot, V. N. Kulikov, and H. P. Lopuha \"a . The limit distribution of the l_ -error of grenander-type estimators. Ann. Statist., 40 0 (3): 0 1578--1608, 06 2012. doi:10.1214/12-AOS1015. URL https://doi.org/10.1214/12-AOS1015

  5. [13]

    C. Gao, F. Han, and C.-H. Zhang. On estimation of isotonic piecewise constant signals. arXiv preprint arXiv:1705.06386, 2017

  6. [14]

    Grenander

    U. Grenander. On the theory of mortality measurement: part ii. Scandinavian Actuarial Journal, 1956 0 (2): 0 125--153, 1956

  7. [15]

    Guntuboyina and B

    A. Guntuboyina and B. Sen. Nonparametric shape-restricted regression. Statistical Science, 33: 0 568--594, 2018

  8. [16]

    Q. Han, T. Wang, S. Chatterjee, and R. J. Samworth. Isotonic regression in general dimensions. ArXiv e-prints, 2017

  9. [17]

    Laurent and P

    B. Laurent and P. Massart. Adaptive estimation of a quadratic functional by model selection. Annals of Statistics, pages 1302--1338, 2000

  10. [18]

    R. E. Miles. The complete amalgamation into blocks, by weighted means, of a finite set of real numbers. Biometrika, 46: 0 317--327, 1959

  11. [19]

    Robertson, F

    T. Robertson, F. Wright, and R. Dykstra. Order Restricted Statistical Inference. Probability and Statistics Series. Wiley, 1988. ISBN 9780471917878. URL https://books.google.com/books?id=sqZfQgAACAAJ

  12. [20]

    A. I. Sakhanenko. Convergence rate in the invariance principle for non-identically distributed variables with exponential moments. Advances in Probability Theory: Limit Theorems for Sums of Random Variables, pages 2--73, 1985

  13. [21]

    F. T. Wright. The asymptotic behavior of monotone regression estimates. Ann. Statist., 9 0 (2): 0 443--448, 1981. ISSN 0090-5364

  14. [22]

    Yang and R

    F. Yang and R. F. Barber. Contraction and uniform convergence of isotonic regression. ArXiv e-prints, 2018

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.