Statistics Theory · arXiv:2611.17082 v1 · submitted 12 Nov 2026 · 13 references
Isotonic least squares is robust to heteroscedasticity, unimodal least squares is not
The paper says
“We compare the least squares estimator (LSE) with the ℓ1-estimator recently introduced by Baraud and Maillard i̧teyearparBM-2026 for estimating a regression function under a fixed design, when the errors are independent but possibly heteroscedastic and heavy-tailed.”
more of the abstract
We show that the two estimators coincide on the class of nondecreasing (or nonincreasing) functions on an interval. As a consequence, the isotonic LSE inherits the non-asymptotic ℓ1-risk bounds of the ℓ1-estimator, which only require a moment of order p ∈ [1,2] on the errors and depend on the average of their p-th absolute moments rather than on their maximum. This coincidence no longer holds on the nonconvex class of unimodal functions: we exhibit a heteroscedastic Gaussian example in which the LSE is inconsistent while the ℓ1-estimator converges at the parametric rate. Finally, we show on a one-dimensional linear model that neither estimator dominates the other: depending on the way the noise is distributed among the design points, the LSE may converge at rate 1/n while the ℓ1-estimator is inconsistent.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 13 Nov 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. The coincidence proof is clean and the counterexamples are sharp, but the central risk-inheritance claim for isotonic LSE depends entirely on an unpublished theorem from a self-cited paper. The paper's own contribution is the coincidence, not the risk bounds. This line is wrong
This review says the paper's mathematics is wrong. Pith did not check that against the paper's source.
3 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 13 references say.
The paper says, at §3, Theorem 2 and surrounding text
The review did not quote this passage.
Pith says: Objection 1 of 3 · major
The central claim that the isotonic LSE inherits the ℓ1-risk bounds (4) and (5) is not proved in this paper. Theorem 2 shows that the LSE minimises T(Z,f) and satisfies T(Z,ef)=0, i.e., that it belongs to E_0(Z,F). The risk bounds (4) and (5) are then stated to follow from Theorem 1 of Baraud and Maillard (2026), which is cited as 'to appear'.
more of objection 1
The paper provides no derivation, sketch, or verification that the hypotheses of Theorem 1 are satisfied beyond the coincidence result itself. Since the entire quantitative content of the headline robustness claim rests on this external unpublished result, the paper's own contribution is incomplete. The authors should either provide a self-contained proof of the risk bounds for the isotonic LSE (perhaps exploiting the specific structure of the isotonic class) or, at a minimum, include a detailed verification that the conditions of Theorem 1 of Baraud and Maillard (2026) are met, and discuss the extent to which the bounds depend on the unpublished source. Without this, the paper's main advertised conclusion is contingent on an unverifiable result.
The paper says, at §4, Theorem 3
The review did not quote this passage.
Pith says: Objection 2 of 3 · major
The parametric rate for the ℓ1-estimator in (7) relies on Proposition 4.1 of Baraud and Maillard (2026) to bound the extremal degree of the piecewise constant function f⋆ in the unimodal class. This proposition is from the same unpublished source.
more of objection 2
While the argument that f⋆ is extremal with degree at most 6 is plausible, the paper does not provide a proof or even a sketch of how Proposition 4.1 applies. Given that the inconsistency of the LSE is the main result of this section, the rate for the ℓ1-estimator is an important benchmark, and its dependence on an unpublished result weakens the self-containedness of the paper.
The paper says, at §2.3, Theorem 1
The review did not quote this passage.
Pith says: Objection 3 of 3 · major
Theorem 1 itself is stated without proof and attributed to Baraud and Maillard (2026). While it is acceptable to quote results from a cited source, the fact that the source is unpublished ('to appear') means that the paper's entire risk analysis framework is built on unverifiable ground.
more of objection 3
The authors should either provide a proof of Theorem 1 in the present paper (or in a publicly available preprint) or restructure the paper to avoid relying on it for the main claims. At the very least, the authors should clarify the status of the reference and, if possible, include a preprint version or a detailed sketch.
6 minor objections
The paper says, at §1, Introduction
Pith says: minor objection 1 of 6
The sentence 'All these results assume that the errors are at least square integrable and that their variances are uniformly controlled (if not constant)' is a bit sweeping. While it is true for the cited works, it might be worth noting that some results allow for heteroscedasticity under bounded variance assumptions.
The paper says, at §3, Proof of Theorem 2
Pith says: minor objection 2 of 6
The proof uses Assumption 1, which is verified for the isotonic class. It would be helpful to explicitly note that this assumption is not satisfied for the unimodal class, which is why the coincidence fails there. This is implicit but could be stated more clearly.
show 4 more
The paper says, at §4, Theorem 3
Pith says: minor objection 3 of 6
The construction of the function f in the proof is somewhat intricate. A figure illustrating the configuration might help readability, though this is not essential.
The paper says, at §5, Proposition 1
Pith says: minor objection 4 of 6
The notation ∥f1∥_{1,n} and ∥f1∥_{2,n} is used but not explicitly defined until after the statement. It would be clearer to define them before or in the statement.
The paper says, at General
Pith says: minor objection 5 of 6
The paper would benefit from a brief discussion of how the results relate to the existing literature on isotonic regression under heteroscedasticity, e.g., whether any previous work has noted robustness properties of the isotonic LSE.
The paper says, at References
Pith says: minor objection 6 of 6
The reference Baraud and Maillard (2026) is cited as 'to appear in Math. Stat. Learn.' It would be helpful to provide an arXiv identifier or a preprint URL if available.
What would settle it
Independently verify that the LSE ef satisfies the conditions of Theorem 1 of Baraud and Maillard (2026) for the class F↑, and then derive the risk bounds (4) and (5) from first principles, without relying on the unpublished manuscript. If the bounds cannot be derived without additional assumptions, the central claim is unsupported.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 13 Nov 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 3 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
Yannick,
This paper has one genuinely new result and two well-constructed counterexamples. The new result is Theorem 2: on monotone classes, the isotonic LSE coincides with the ℓ1-estimator and even satisfies T(Z, ef) = 0. The proof is short, self-contained, and uses a clever convexity argument (Assumption 1). That is a nice piece of work.
The counterexamples are also solid. The unimodal Gaussian example (Theorem 3) is explicit and shows inconsistency of the LSE while the ℓ1-estimator converges at parametric rate. The linear model example (Proposition 1) is a clean demonstration that neither estimator dominates. These are useful contributions.
Now the soft spot, and it is a load-bearing one. The paper's headline claim—that the isotonic LSE inherits the ℓ1-risk bounds (4) and (5)—is not proved here. Theorem 2 only shows coincidence and T(Z, ef)=0. The risk bounds are then asserted to follow from Theorem 1 of Baraud and Maillard (2026), which is cited as "to appear." The paper does not derive them, does not sketch how the coincidence plus that external theorem yields the bounds, and provides no independent verification. If that unpublished theorem is false or has narrower hypotheses, the central robustness conclusion collapses. The paper's own contribution—the coincidence proof—is necessary but not sufficient for the risk bounds.
This is not a minor gap. The abstract and introduction advertise the risk-inheritance as the main consequence. A reader cannot evaluate whether the claim holds without access to the cited paper. The authors acknowledge the reliance (they cite BM26 for Theorem 1 and Proposition 4.1), but that does not reduce the burden: the paper's central advertised result is contingent on an external, unavailable source.
Who is this paper for? Researchers in shape-restricted regression and ℓ1-estimation. The coincidence result is interesting and may be useful. The counterexamples are pedagogically valuable. But the paper as a whole cannot be fully assessed until BM26 is available.
Recommendation: I would not desk-reject this, but I would send it to a referee with a clear note that the risk-inheritance claim depends on an unpublished source. The referee should be asked to evaluate the coincidence proof and the counterexamples on their own, and to note that the risk bounds are conditional. If BM26 appears and is verified, the paper becomes a solid contribution. As it stands, it is a conditional result.
Best, [Your name]
The strongest objection
The paper's main advertised result is that the isotonic LSE inherits the non-asymptotic ℓ1-risk bounds of the ℓ1-estimator (Theorem 2 and surrounding text). However, the proof of Theorem 2 only shows that the LSE minimises T(Z,f) and satisfies T(Z,ef)=0. The risk bounds (4) and (5) are then claimed to follow from Theorem 1 of Baraud and Maillard (2026), which is cited as 'to appear'. The paper does not provide any derivation of these bounds from the coincidence result alone; it simply states 'In particular, the LSE satisfies (4) and (5) with c=0.' This is a direct appeal to an external, unpublished source for the entire quantitative content of the central claim. The paper's own contribution—the coincidence proof—is necessary but not sufficient for the risk bounds, because Theorem 1 of Baraud and Maillard (2026) requires that the estimator be an element of Ec(Z,F) with c≥0. The paper shows that the LSE is in E0(Z,F), but the risk bounds themselves are not re-derived or even sketched. The reader's verdict correctly identifies this as the weakest assumption. The concern is not that the cited result is false, but that the paper's central claim is entirely contingent on a result that is not yet publicly available, and the paper does not provide independent verification or even a sketch of how the bounds follow. This is a load-bearing concern because without Theorem 1 of Baraud and Maillard (2026), the paper's headline conclusion—that isotonic LSE is robust to heteroscedasticity—is unsubstantiated. The paper's own proof only covers the coincidence, which is interesting but does not by itself imply the risk bounds.
The reader's verdict identifies the same weakest assumption: the reliance on an unpublished theorem. The paper's own proof only establishes that the LSE minimises T(Z,f) and satisfies T(Z,ef)=0, which places it in E0(Z,F). However, the risk bounds (4) and (5) are not proved in this paper; they are stated as consequences of Theorem 1 of Baraud and Maillard (2026), which is cited as 'to appear'. The paper does not provide any derivation or even a sketch of how the bounds follow from the coincidence result. This is a genuine load-bearing concern because the entire quantitative claim of robustness (the non-asymptotic ℓ1-risk bounds depending on average moments) rests on that external theorem. Without it, the paper's contribution is limited to the coincidence result, which is interesting but does not by itself imply the advertised risk guarantees. The paper does not contain any internal verification or alternative proof. The concern is not that the cited result is likely false, but that the paper's central claim is not self-contained and cannot be evaluated without access to the unpublished source. The concrete test proposed is to independently derive the risk bounds from the coincidence result, which would either validate the claim or reveal hidden assumptions. The verdict remains CONDITIONAL because the paper's own proof is sound and the coincidence result is valuable; the condition is that the risk-inheritance claim must be substantiated by the appearance of the cited work or by an independent derivation. The agreement with the reader is 'agree' because the reader's weakest_assumption directly matches this concern.
The referee's report
If the central claim holds, the paper makes a conceptually interesting observation: the isotonic LSE, a classical and widely used estimator, automatically enjoys the robustness properties of the more recently proposed ℓ1-estimator under heteroscedasticity and heavy tails. This would be practically relevant because isotonic regression is a standard tool and the ℓ1-estimator's risk bounds depend on the average of p-th absolute moments rather than their maximum, offering a genuine advantage in heteroscedastic settings. The unimodal counterexample (Theorem 3) is also valuable as it delineates the limits of this robustness. The paper is clearly written and the proof of the coincidence (Theorem 2) is self-contained and elegant. The linear model comparison (Proposition 1) provides a balanced perspective.
Pith's reading of the paper (a machine's summary, not the authors')
In a fixed-design regression problem, you observe noisy values at known input points and want to recover the underlying function. Usually the noise is assumed to have the same variance at every point. Here the errors may have different variances and may be heavy-tailed. A recently introduced ℓ1 estimator works with signs of differences instead of squares, and it has risk bounds that depend on the average of the p-th absolute moments of the noise, not on the largest one. The paper asks whether the classic least squares estimator has this same robustness on shape-restricted classes. It shows that for monotone (nondecreasing or nonincreasing) functions, the two estimators coincide, so isotonic least squares automatically gets the ℓ1 risk bounds. The proof is self-contained and uses the projection characterization of least squares plus a geometric property of monotone vectors. For unimodal functions, which can have one peak, the coincidence breaks down. The paper constructs a Gaussian example with one very noisy design point. On an event with probability roughly 30 percent, the least squares fit is far from the true indicator function, while the ℓ1 estimator converges at the parametric rate. Then, in a one-dimensional linear model, the opposite happens: least squares can converge at rate 1/n while the ℓ1 estimator is inconsistent. So neither estimator dominates in general.
Core claim. Theorem 2 states: "For F=F↑, the LSE ef minimises f↦T(Z,f) over F and even satisfies T(Z,ef)=0. In particular, the LSE satisfies (4) and (5) with c=0." If correct, the isotonic least squares estimator has the same non-asymptotic ℓ1-risk guarantees as the ℓ1-estimator, requiring only finite p-absolute moments for some p in [1,2] and depending on the average rather than the maximum of those moments.
The premise it rests on. The advertised inheritance of ℓ1 risk bounds by the isotonic LSE is not proved in this paper; it is a consequence of Theorem 1 of Baraud and Maillard (2026), which is stated without proof and cited as "to appear." The central robustness conclusion would collapse if that external theorem is false or has narrower hypotheses. Similarly, the parametric rate for the ℓ1 estimator in Theorem 3 relies on Proposition 4.1 of that same unpublished source.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
We compare the least squares estimator (LSE) with the ℓ1-estimator recently introduced by Baraud and Maillard i̧teyearparBM-2026 for estimating a regression function under a fixed design, when the errors are independent but possibly heteroscedastic and heavy-tailed. We show that the two estimators coincide on the class of nondecreasing (or nonincreasing) functions on an interval. As a consequence, the isotonic LSE inherits the non-asymptotic ℓ1-risk bounds of the ℓ1-estimator, which only require a moment of order p ∈ [1,2] on the errors and depend on the average of their p-th absolute moments rather than on their maximum. This coincidence no longer holds on the nonconvex class of unimodal functions: we exhibit a heteroscedastic Gaussian example in which the LSE is inconsistent while the ℓ1-estimator converges at the parametric rate. Finally, we show on a one-dimensional linear model that neither estimator dominates the other: depending on the way the noise is distributed among the design points, the LSE may converge at rate 1/n while the ℓ1-estimator is inconsistent.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-11-14
The authors' main claim, as the classifier picked it out 0.83
“This coincidence no longer holds on the nonconvex class of unimodal functions: we exhibit a heteroscedastic Gaussian example in which the LSE is inconsistent while the $\ell_{1}$-estimator converges at the parametric rate.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.03 · states limitation 0.76 · human participants 0.02 · controlled comparison 0.18 · randomized assignment 0.01 · states quantitative result 0.74 · compares with prior results 0.95 · mentions code or data availability 0.02
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| We compare the least squares estimator (LSE) with the $\ell_{1}$-estimator recently introduced by Baraud and Maillard~\citeyearpar{BM-2026} for estimating a regression function under a fixed design, when the errors are independent but possibly heteroscedastic and heavy-tailed. | contribution 0.57 | 0.57 | 0.04 | unclear 0.62 |
| We show that the two estimators coincide on the class of nondecreasing (or nonincreasing) functions on an interval. | result 1.00 | 0.91 | 0.10 | asserted 1.00 |
| As a consequence, the isotonic LSE inherits the non-asymptotic $\ell_{1}$-risk bounds of the $\ell_{1}$-estimator, which only require a moment of order $p\in[1,2]$ on the errors and depend on the average of their $p$-th absolute moments rather than on their maximum. | result 0.64 | 0.86 | 0.27 | asserted 1.00 |
| This coincidence no longer holds on the nonconvex class of unimodal functions: we exhibit a heteroscedastic Gaussian example in which the LSE is inconsistent while the $\ell_{1}$-estimator converges at the parametric rate. | result 0.99 | 0.94 | 0.50 | asserted 1.00 |
| Finally, we show on a one-dimensional linear model that neither estimator dominates the other: depending on the way the noise is distributed among the design points, the LSE may converge at rate $1/n$ while the $\ell_{1}$-estimator is inconsistent. | result 1.00 | 0.95 | 0.10 | asserted 0.78 |
What kind of paper the classifier reports this is: theoretical result 1.00 · evidence from proof or derivation 1.00
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-11-14: 1 request, 4,640 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
References (13)
- [1] Baraud, Y. and Maillard, G. (2026). Estimating a regression function under possible heteroscedastic and heavy-tailed errors. Application to shape-restricted regression. to appear in Math. Stat. Learn on Pith
- [2] Bellec, P. C. (2018). Sharp oracle inequalities for Least Squares estimators in shape restricted regression . The Annals of Statistics , 46(2):745 -- 780 on Pith
- [3] Chatterjee, S. (2014). A new perspective on least squares under convex constraint. Ann. Statist. , 42(6):2340--2381 on Pith
- [4] Chatterjee, S. (2016). An improved global risk bound in concave regression . Electronic Journal of Statistics , 10(1):1608 -- 1629 on Pith
- [5] Chatterjee, S., Guntuboyina, A., and Sen, B. (2015). On risk bounds in isotonic and other shape restricted regression problems. Ann. Statist. , 43(4):1774--1800 on Pith
- [6] Chatterjee, S. and Lafferty, J. (2019). Adaptive risk bounds in unimodal regression . Bernoulli , 25(1):1 -- 25 on Pith
- [7] Devroye, L., Lerasle, M., Lugosi, G., and Oliveira, R. I. (2016). Sub- G aussian mean estimators. Ann. Statist. , 44(6):2695--2725 on Pith
- [8] Efromovich, S. and Pinsker, M. (1996). Sharp-optimal and adaptive estimation for heteroscedastic nonparametric regression. Statistica Sinica , 6:925--942 on Pith
- [9] Feng, O. Y., Chen, Y., Han, Q., Carroll, R. J., and Samworth, R. J. (2022). Nonparametric, tuning-free estimation of S -shaped functions. Journal of the Royal Statistical Society Series B: Statistical Methodology , 84(4):1324--1352 on Pith
- [10] Gao, C., Han, F., and Zhang, C.-H. (2020). On estimation of isotonic piecewise constant signals. Ann. Statist. , 48(2):629--654 on Pith
- [11] Guntuboyina, A. and Sen, B. (2018). Nonparametric Shape-Restricted Regression . Statistical Science , 33(4):568 -- 594 on Pith
- [12] Minami, K. (2020). Estimating piecewise monotone signals . Electronic Journal of Statistics , 14(1):1508 -- 1576 on Pith
- [13] Zhang, C.-H. (2002). Risk bounds in isotonic regression . The Annals of Statistics , 30(2):528 -- 555 on Pith
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment