Pith. sign in

REVIEW 2 major objections 4 minor 50 references

Near-optimal Delta-convex Estimation of Lipschitz Functions

T0 review · 2 major / 4 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read A tractable estimator matches the optimal error rate for Lipschitz regression, adapting to the intrinsic dimension of the data.

desk verdict A genuinely new convex-optimization estimator for Lipschitz regression with a plausible near-minimax proof; the main vulnerability is a load-bearing external covering lemma whose local justification in the paper is suspect. read the letter →

arxiv 2511.15615 v2 pith:RITXYIKZ submitted 2025-11-19 stat.ML cs.LG

classification stat.MLcs.LG MSC 62G0862G0562C2041A30
keywords nonparametricregressionLipschitzfunctiondelta-convexminimaxrateintrinsicdimensionadaptivepartitioningmax-affineapproximationempiricalriskminimization
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

The paper claims that estimating a Lipschitz regression function from noisy data can be done at the statistically optimal rate — the minimax rate n^{-2/(2+d_*)} up to logarithmic factors — where d_* is the intrinsic dimension of the covariates rather than the ambient dimension. The new estimator, called delta-convex fitting (DCF), builds a finite covering of the covariate data, fits a max-concave (delta-convex) function on that cover via a tractable convex program, and refines the result locally. The main theorem gives a PAC bound valid for every sample size, without requiring the true Lipschitz constant as an input. The method is also adapted to convex regression, where it provides the first tractable PAC guarantee in the random-design setting without knowledge of the Lipschitz constant.

What carries the argument

The load-bearing object is the feature map φ▷(x, x̂) = (x−x̂, ‖x−x̂‖▷), which turns each affine piece into a 'cone' term; the function class F▷(X̂) is the max over such terms centered at the cover points. The key identity is that the classical Lipschitz extension f̂(x)=max_{x̂∈X_ε}(f(x̂)−λ‖x−x̂‖) approximates any λ-Lipschitz f to within O(λε) on an ε-cover and is itself Lipschitz with constant O(λ). DCF then solves a convex second-order-cone program over these parameters, with constraints enforced only at the cover centers, and a refinement step that prunes unused terms and re-centers the intercept.

What would settle it

Generate n i.i.d. samples from a subgaussian distribution whose support is a 1-dimensional manifold embedded in R^10, run the adaptive farthest-point clustering as defined in Algorithm 1, and check whether the returned number of centers K is O(n^{1/3}) and the covering radius is O(√(K/n)). A single example where K grows faster than n^{d_*/(2+d_*)} or the covering radius exceeds the stated bound would refute the cited lemma and collapse Theorem 1; alternatively, a constructed distribution with d_*≪d where the estimator's squared error decays slower than n^{-2/(2+d_*)} times a polylog factor wou

Watch

Extended reading notes

Core claim

The central claim is that functions of the form max over centers of (affine term plus norm term) — a subclass of delta-convex functions — are universal approximators of Lipschitz functions at uniform error O(λε) while preserving the Lipschitz constant up to a constant factor. Combined with an adaptive farthest-point clustering that produces K ≈ n^{d_*/(2+d_*)} centers covering the data at resolution ε ≈ n^{-1/(2+d_*)}, this yields an empirical-risk-minimization estimator whose squared error is within logarithmic factors of the minimax lower bound. The proof decomposes the error into approximation and estimation terms, controls the former with the covering argument and the latter with metric-

Load-bearing premise

The rate proof relies on a previously published covering lemma, cited without proof here, that the adaptive farthest-point clustering always returns K = O(n^{d_*/(2+d_*)}) centers whose covering radius is O(R_X√(K/n)); if this data-dependent covering guarantee fails on some subgaussian distribution, the n^{-2/(2+d_*)} rate does not follow.

Editorial extensions

If this is right

  • If Theorem 1 holds, Lipschitz regression can be solved at the minimax rate by a polynomial-time algorithm that adapts to low intrinsic dimension without any external model-selection procedure.
  • The near-minimax guarantee also covers max-min-affine variants and the symmetric difference representation F^∆, so practitioners can choose richer function classes without losing the rate.
  • The same machinery gives a convex regression estimator with a PAC guarantee for all n and no knowledge of the Lipschitz constant, improving on earlier results that hold only for large n or high dimension.
  • The O((1+d I{▷≠2})) factor identifies Euclidean norm features (▷=2) as the most parameter-efficient choice, with other norms costing a dimension factor.

Reading between the lines

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

  • The analysis suggests that the covering-resolution tradeoff is the real obstacle to adaptivity: any clustering method that returns O(n^{d_*/(2+d_*)}) centers with covering radius O(n^{-1/(2+d_*)}) would slot into the same rate proof, so other data-dependent quantizations could plausibly match DCF.
  • The proof's exclusion of quadratic features (‖x−x̂‖²) points to an open avenue: a locally Lipschitz regularizer that keeps the quadratic coefficient independent of ε would likely extend near-minimax rates to smooth regression functions.
  • The documented sensitivity of empirical performance to the slope penalty θ₂ suggests that data-dependent selection of this regularization parameter is a natural next target; the theory only requires θ₂ to lie in a range, not a specific value.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a polynomial-time estimator (delta-convex fitting, DCF) for nonparametric regression of an unknown Lipschitz function in the random-design setting, under squared loss and subgaussian covariate/noise assumptions. The estimator combines adaptive farthest-point clustering (AFPC) to select center points, a convex optimization problem over a class of piecewise affine/"delta-convex" functions, a regularization that avoids knowing the true Lipschitz constant, and an optional local refinement. The main theoretical claim (Theorem 1) is a PAC bound of order O(d(1+d I{▷≠2}) n^{-2/(2+d_*)} β), with β polylogarithmic, matching Stone's minimax lower bound up to logarithmic factors in terms of the intrinsic dimension d_*. The paper also gives approximation results for max-min-affine and other delta-convex classes and extends the framework to convex regression. Experiments compare DCF with k-NN, Nadaraya-Watson, random forests, and XGBoost on three public datasets.

Significance. If the main theorem holds, this is a genuinely valuable contribution: it is, to my knowledge, the first tractable estimator for general Lipschitz regression that is proven to achieve the adaptive near-minimax rate in the random-design setting without knowledge of the Lipschitz constant, while also admitting a convex initialization. The approximation construction based on McShane's extension and the use of AFPC to adapt to intrinsic dimension are elegant, and the paper explicitly provides a PAC bound for all n rather than only asymptotic convergence. The experiments are a useful complement, and the code is made available. However, two load-bearing points in the proof currently prevent me from accepting the paper as written: the adaptive covering lemma is cited without proof and with an incorrect local justification, and an algebraic simplification in the concentration step appears invalid and would, as written, break the n^{-2/(2+d_*)} rate. Both appear repairable, but they need to be fixed before the central claim is established.

major comments (2)
  1. [§2.1 and §4.3 (Lemma 8)] The paper states that the AFPC stopping rule implies 'K−1 < k̂(X_K) ≤ K' and uses this to justify ε² ≈ R² K/n. This inequality is not guaranteed: adding the final farthest point can sharply reduce ε_n(X_K) (e.g., two tight clusters), making k̂(X_K) much smaller than K−1. The useful upper bound ε = O(R√(K/n)) does follow directly from K ≥ k̂(X_K), and the complexity bound K ≤ k* requires K−1 < k̂(X_{K−1}) together with a covering/doubling argument. Neither is supplied. Since Theorem 1's rate enters exactly through Lemma 8, the manuscript must either prove Lemma 8 (especially the d* part) or provide a correct derivation; the current local justification is incorrect.
  2. [§4.4.2, Lemma 16, Eq. (20)] With δ_n = r_σ√(dK/n), the term r_σ δ_n in the concentration bound equals r_σ²√(dK/n) up to constants. The displayed simplification in Eq. (20) claims this is O(dKσ² ln(β2/γ)/n), which is false when dK/n < 1 — exactly the regime relevant for the near-minimax rate, where dK/n → 0. If this term is retained, the bound on E_approx in Eq. (23) would contain a √(K/n) term rather than K/n, destroying the n^{-2/(2+d_*)} rate. The proof can likely be repaired by choosing δ_n = r_σ dK/n (or a similar smaller scale), since δ appears only logarithmically in Lemma 15; but as written the central rate is not established.
minor comments (4)
  1. [Throughout §4] The prose consistently refers to lemmas as 'Theorem' (e.g., 'Theorem 4', 'Theorem 8', 'Theorem 16'). The cross-references should be corrected to Lemma numbers.
  2. [§2.1] The sentence 'The stopping condition of AFPC ensures that K−1 < k̂(X_K) ≤ K' is false as noted in Major 1. It should be rephrased, for example as 'K ≥ k̂(X_K) and K−1 < k̂(X_{K−1})'.
  3. [Proof of Theorem 1, after Eq. (33)] The simplification '(r_ρ β_0)/(ηδ) = O(n²√d n)' is ambiguous and appears to be a typo; it should be O(n²√d ln n) (or, if using ln n ≤ n, written with the explicit substitution).
  4. [Various] Typos and stylistic issues: 'Cauchy-Schwartz' should be 'Cauchy-Schwarz'; 'strenghten' → 'strengthen'; 'straightforwadly' → 'straightforwardly'; 'implmentation' → 'implementation'; 'samples sizes' → 'sample sizes'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the near-minimax rate follows from a genuine upper-bound proof with independent geometric and concentration lemmas.

full rationale

Theorem 1's rate is derived by bounding approximation error via Theorem 2 (McShane extension over an epsilon-cover) and estimation error via ERM/concentration inequalities, then plugging in the AFPC covering guarantee (Lemma 8). Lemma 8 is cited from the author's prior work, but it is a parameter-free clustering statement about AFPC's output (K and epsilon), not a restatement of the regression rate, and it is externally checkable. The statistical analysis—error decomposition (11), the 'basic inequality' (16), and concentration lemmas (11, 17)—is carried out in the paper. The regularization (7) is chosen as a function of K and n rather than fitted to the target loss, and the proof does not tune parameters to match the minimax rate. The only notable burden is reliance on self-cited AFPC and prior concentration results; these are independent published results and do not make the derivation circular. A separate correctness concern: Section 2.1's sentence 'The stopping condition of AFPC (Algorithm 1) ensures that K−1 < khat(X_K) ≤ K' is not an obvious consequence of the loop condition, since the loop condition is evaluated before the final insertion and the covering radius can drop sharply; this affects the informal justification of Lemma 8 but is a proof-rigor issue, not circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The DCF proof rests on four main inputs: the McShane extension (standard), the subgaussian model and bounded doubling dimension (domain assumptions), and the self-cited AFPC covering guarantee. The four regularization hyperparameters θ_0–θ_3 are chosen by formula or grid, with θ_2 flagged as performance-sensitive. No new entities with external handles are introduced.

free parameters (4)
  • θ_0 = O((R_Yn/max{1,R_Xn}) ln n); experiments use (R_Yn/R_Xn) ln n
    Regularization parameter in (3), mitigates over-conservative Lipschitz bound. Data-dependent but not fitted to the regression function.
  • θ_1 = Θ(max{1,R_Xn^2}(dK/n))
    Scales the z^2 penalty in (3); chosen by formula (7).
  • θ_2 = 0≤θ_2≤θ_1/K; experiments use R_Xn^2/n or (R_Xn/n)^2
    Penalizes slope norms; the paper notes performance is sensitive to this choice (Fig. 4).
  • θ_3 = 1≤θ_3=O(ln n); experiments use ln n
    Allows refinement Lipschitz slack; appears in the rate via θ_3^2.
assumptions (4)
  • standard math McShane extension theorem: every λ-Lipschitz f on X extends to R^d via sup_{\hat x∈X} f(\hat x)-λ\|x-\hat x\|.
    Used in Theorem 2 to construct the max-over-cover approximation with controlled Lipschitz constant.
  • domain assumption Doubling-dimension covering bound: N_{\|·\|}(X_n, ε) ≤ max{1,(4R_Xn/ε)^{d_*}} where d_* is the intrinsic dimension.
    Assumes X_* has bounded doubling dimension d_*; cited to Kpotufe-Dasgupta and Wainwright. Needed for Lemma 8 and the final rate.
  • domain assumption Subgaussian model: E[e^{\|X-E[X]\|^2/ρ^2}] ≤ 2 and E[e^{(f^*(X)-Y)^2/σ^2}|X] ≤ 2 a.s.
    Assumed in Theorem 1 for concentration; not verified empirically.
  • domain assumption AFPC guarantee (Balázs 2022, Lemma 4.2): AFPC output \hat X has K=O(n^{d_*/(2+d_*)}) a.s. and ε_n(\hat X)=O(R_Xn√(K/n)).
    Cited from self-authored prior work; not reproved in this paper, but load-bearing for the rate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Near-optimal Delta-convex Estimation of Lipschitz Functions." pith.science (2026). https://pith.science/paper/RITXYIKZ

@misc{pith2026251115615,
  author       = {Pith},
  title        = {Pith review of: Near-optimal Delta-convex Estimation of Lipschitz Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RITXYIKZ}},
  note         = {Machine review of arXiv:2511.15615}
}
read the original abstract

This paper presents a tractable algorithm for estimating an unknown Lipschitz function from noisy observations and establishes an upper bound on its convergence rate. The approach extends max-affine methods from convex shape-restricted regression to the more general Lipschitz setting. A key component is a nonlinear feature expansion that maps max-affine functions into a subclass of delta-convex functions, which act as universal approximators of Lipschitz functions while preserving their Lipschitz constants. Leveraging this property, the estimator attains the minimax convergence rate (up to logarithmic factors) with respect to the intrinsic dimension of the data under squared loss and subgaussian distributions in the random design setting. The algorithm integrates adaptive partitioning to capture intrinsic dimension, a penalty-based regularization mechanism that removes the need to know the true Lipschitz constant, and a two-stage optimization procedure combining a convex initialization with local refinement. The framework is also straightforward to adapt to convex shape-restricted regression. Experiments demonstrate competitive performance relative to other theoretically justified methods, including nearest-neighbor and kernel-based regressors.

Figures

Figures reproduced from arXiv: 2511.15615 by the authors.

Figure 1
Figure 1. AFPC partition size (K) for sample sizes n ∈ {1024, 2048, 4096}, and average cell size distribution for n = 4096. The upper bound of K is n d/(2+d) . The black vertical lines on the average cell size axes mark the value of d. The plots for pumadyn-8nh are similar to those of pumadyn-8nm and are omitted for brevity. analogously in (4), as has been done in convex regression practice (Aybat and Wang, 2016; Chen and Maz… view at source ↗
Figure 2
Figure 2. Test MSEs of the estimators trained on samples sizes n ∈ {1024, 2048, 4096}. The performance is very similar across all estimators for both MM and STD scalings of the pumadyn datasets; therefore, the plots for the latter are omitted for brevity. at least as well as the other theoretically justified methods (k-NN and NW) on all problems, except for the cpusmall dataset with STD scaling. In that case, [PITH_FULL_IMAG… view at source ↗
Figure 3
Figure 3. Training and prediction times (in seconds and milliseconds, respectively) are shown for the pumadyn-8nm dataset with MM scaling in the left and center panels. Prediction times are measured on the entire test set (whose size varies with n) and normalized to 1000 samples. The right panel shows the number of parameters used by the initial and final DCF estimators, fn and f + n , respectively. DCF on the pumadyn dataset… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Test MSEs, using the same notations as above, where all DCF models are trained with the weaker regularizer θ2 = (RXn /n) 2 . weaker regularizer θ2 = (RXn /n) 2 , the left panel shows that all DCF methods perform well on the cpusmall dataset with STD scaling (unlike usi…
Figure 5
Figure 5. Figure 5: Approximation of a function f ∈ Fλ,X by the max-concave ˆf and the min-convex ˇf as defined above. The left two plots use f(x) = x sin(x), while the right two plots use f(x) = max{1 − |x − 1|, 2 − |x − 3|, 1 − |x − 5|/2}, both over X = [0, 6]. The shaded regions repres…
Figure 6
Figure 6. Figure 6: Approximation of a smooth function (left two plots) by ˜f1 and ˘f1 of Theorem 23, and of a non-smooth Lipschitz function (right two plots) by ˜f0 and ˘f0 of Theo￾rem 24. The setting and notation are the same as in [PITH_FULL_IMAGE:figures/full_fig_p032_6.png]
Figure 7
Figure 7. Figure 7: DCF approximations, each column showing the result for F▷(·), F − ▷ (·), F ∆ ▷ (·), respectively. We used the norms ▷ ∈ {1, 2,∞} which are equivalent for d = 1. The settings and the notations are the same as on [PITH_FULL_IMAGE:figures/full_fig_p034_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 4 linked inside Pith

  1. [1]

    A parallelizable dual smoothing method for large scale convex regression problems

    Necdet Serhat Aybat and Zi Wang. A parallelizable dual smoothing method for large scale convex regression problems. arXiv preprint arXiv:1608.02227, 2016

  2. [2]

    An algorithm for the estimation of a regression function by continuous piecewise linear functions

    Adil Bagirov, Conny Clausen, and Michael Kohler. An algorithm for the estimation of a regression function by continuous piecewise linear functions. Computational Optimization and Applications, 45 0 (1): 0 159--179, 2010

  3. [3]

    Bagirov, Sona Taheri, Napsu Karmitsa, Nargiz Sultanova, and Soodabeh Asadi

    Adil M. Bagirov, Sona Taheri, Napsu Karmitsa, Nargiz Sultanova, and Soodabeh Asadi. Robust piecewise linear L1 -regression via nonsmooth DC optimization. Optimization Methods and Software, 37 0 (4): 0 1289--1309, 2022

  4. [4]

    Convex Regression: Theory, Practice, and Applications

    G\'abor Bal\'azs. Convex Regression: Theory, Practice, and Applications. PhD thesis, University of Alberta, 2016

  5. [5]

    Adaptively partitioning max-affine estimators for convex regression

    G\'abor Bal\'azs. Adaptively partitioning max-affine estimators for convex regression. In Gustau Camps-Valls, Francisco J. R. Ruiz, and Isabel Valera, editors, Proceedings of The 25th International Conference on Artificial Intelligence and Statistics, volume 151 of Proceedings of Machine Learning Research, pages 860--874. PMLR, 2022

  6. [6]

    Near-optimal max-affine estimators for convex regression

    G\'abor Bal\'azs, Andr\'as György, and Cs aba Szepesv\'ari. Near-optimal max-affine estimators for convex regression . In Guy Lebanon and S. V. N. Vishwanathan, editors, Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, volume 38 of Proceedings of Machine Learning Research, pages 56--64, San Diego, Californi...

  7. [7]

    Chaining bounds for empirical risk minimization

    G\'abor Bal\'azs, Andr\'as Gy\"orgy, and Cs aba Szepesv\'ari. Chaining bounds for empirical risk minimization. arXiv preprint arXiv:1609.01872v1, 2016

  8. [8]

    Bartlett and Shahar Mendelson

    Peter L. Bartlett and Shahar Mendelson. Empirical minimization. Probability Theory and Related Fields, 135 0 (3): 0 311--334, 2006

Show all 50 references
  1. [9]

    Bartlett, St\'ephane Boucheron, and G\'abor Lugosi

    Peter L. Bartlett, St\'ephane Boucheron, and G\'abor Lugosi. Model selection and error estimation. Machine Learning, 48 0 (1): 0 85--113, 2002

  2. [10]

    Multivariate distributionally robust convex regression under absolute error loss

    Jose Blanchet, Peter W Glynn, Jun Yan, and Zhengqing Zhou. Multivariate distributionally robust convex regression under absolute error loss. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Proces...

  3. [11]

    Concentration Inequalities: A nonasymptotic theory of independence

    St\'ephane Boucheron, G\'abor Lugosi, and Pascal Massart. Concentration Inequalities: A nonasymptotic theory of independence. Oxford University Press, 2013

  4. [12]

    Convex Optimization

    Stephen Boyd and Lieven Vandenberghe. Convex Optimization. Cambridge University Press, 2004

  5. [13]

    Random forests

    Leo Breiman. Random forests. Machine Learning, 45 0 (1): 0 5--32, 2001

  6. [14]

    XGB oost: A S calable T ree B oosting S ystem

    Tianqi Chen and Carlos Guestrin. XGB oost: A S calable T ree B oosting S ystem. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, KDD '16, page 785–794. Association for Computing Machinery, 2016. doi:10.1145/2939672.2939785

  7. [15]

    Subgradient regularized multivariate convex regression at scale

    Wenyu Chen and Rahul Mazumder. Subgradient regularized multivariate convex regression at scale. SIAM Journal on Optimization, 34 0 (3): 0 2350--2377, 2024

  8. [16]

    Random projection trees and low dimensional manifolds

    Sanjoy Dasgupta and Yoav Freund. Random projection trees and low dimensional manifolds. In Proceedings of the Fortieth Annual ACM Symposium on Theory of Computing, STOC '08, page 537–546, New York, NY, USA, 2008. Association for Computing Machinery

  9. [17]

    DeVore, Ralph Howard, and Charles Micchelli

    Ronald A. DeVore, Ralph Howard, and Charles Micchelli. Optimal nonlinear approximation. Manuscripta Mathematica, 63 0 (4): 0 469--478, 1989

  10. [18]

    Gonzalez

    Teofilo F. Gonzalez. Clustering to minimize the maximum intercluster distance. Theoretical Computer Science, 38: 0 293--306, 1985

  11. [19]

    M axout N etworks

    Ian Goodfellow, David Warde-Farley, Mehdi Mirza, Aaron Courville, and Yoshua Bengio. M axout N etworks. In Sanjoy Dasgupta and David McAllester, editors, Proceedings of the 30th International Conference on Machine Learning, volume 28 of Proceedings of Machine Learning Research...

  12. [20]

    Goulart and Yuwen Chen

    Paul J. Goulart and Yuwen Chen. Clarabel: An interior-point solver for conic programs with quadratic objectives. arXiv preprint arXiv:2405.12762, 2024

  13. [21]

    Gupta, R

    A. Gupta, R. Krauthgamer, and J.R. Lee. Bounded geometries, fractals, and low-distortion embeddings. In 44th Annual IEEE Symposium on Foundations of Computer Science, 2003. Proceedings., pages 534--543, 2003

  14. [22]

    A distribution-free theory of nonparametric regression

    L\'aszl\'o Gy\"orfi, Michael Kohler, Adam Krzyzak, and Harro Walk. A distribution-free theory of nonparametric regression. Springer, 2002

  15. [23]

    Qiyang Han and Jon A. Wellner. Multivariate convex regression: Global risk bounds and adaptation. arXiv preprint arXiv:1601.06844v1, 2016

  16. [24]

    On functions representable as a difference of convex functions

    Philip Hartman. On functions representable as a difference of convex functions. Pacific Journal of Mathematics, 9 0 (3): 0 707--713, 1959

  17. [25]

    Extension of L ipschitz functions

    J.-B Hiriart-Urruty. Extension of L ipschitz functions. Journal of Mathematical Analysis and Applications, 77 0 (2): 0 539--554, 1980

  18. [26]

    Generalized differentiability, duality and optimization for problems dealing with differences of convex functions

    Jean-Baptiste Hiriart-Urruty. Generalized differentiability, duality and optimization for problems dealing with differences of convex functions. Convexity and Duality in Optimization, pages 37--70, 1985

  19. [27]

    Hochbaum and David B

    Dorit S. Hochbaum and David B. Shmoys. A best possible heuristic for the k -center problem. Mathematics of Operations Research, 10 0 (2): 0 180--184, 1985

  20. [28]

    The curse of dimension in nonparametric regression

    Samory Kpotufe. The curse of dimension in nonparametric regression. PhD thesis, University of California, 2010

  21. [29]

    A tree-based regressor that adapts to intrinsic dimension

    Samory Kpotufe and Sanjoy Dasgupta. A tree-based regressor that adapts to intrinsic dimension. Journal of Computer and System Sciences, 78 0 (5): 0 1496--1515, 2011

  22. [30]

    Kulkarni and S.E

    S.R. Kulkarni and S.E. Posner. Rates of convergence of nearest neighbor estimation under arbitrary sampling. IEEE Transactions on Information Theory, 41 0 (4): 0 1028--1039, 1995

  23. [31]

    Convex regression in multidimensions: Suboptimality of least squares estimators

    Gil Kur, Fuchang Gao, Adityanand Guntuboyina, and Bodhisattva Sen. Convex regression in multidimensions: Suboptimality of least squares estimators . The Annals of Statistics, 52 0 (6): 0 2791--2815, 2024

  24. [32]

    On convergence rates of convex regression in multiple dimensions

    Eunji Lim. On convergence rates of convex regression in multiple dimensions. INFORMS Journal on Computing, 26 0 (3): 0 616--628, 2014

  25. [33]

    Convex regression with a penalty

    Eunji Lim. Convex regression with a penalty. arXiv preprint arXiv:2509.19788v1, 2025

  26. [34]

    E. J. McShane . Extension of range of functions. Bulletin of the American Mathematical Society, 40 0 (12): 0 837--842, 1934

  27. [35]

    E. A. Nadaraya. On estimating regression. Theory of Probability & Its Applications, 9 0 (1): 0 141--142, 1964

  28. [36]

    Nesterov

    Yu. Nesterov. Smooth minimization of non-smooth functions. Mathematical Programming, 103 0 (1): 0 127--152, 2005

  29. [37]

    Interior-Point Polynomial Algorithms in Convex Programming

    Yurii Nesterov and Arkadii Nemirovskii. Interior-Point Polynomial Algorithms in Convex Programming. Society for Industrial and Applied Mathematics, 1994

  30. [38]

    Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, 2006

  31. [39]

    Max-min representation of piecewise linear functions

    Sergei Ovchinnikov. Max-min representation of piecewise linear functions. Contributions to Algebra and Geometry, 43 0 (1): 0 297--302, 2002

  32. [40]

    Scikit-learn: Machine learning in python

    Fabian Pedregosa, Ga \"e l Varoquaux, Alexandre Gramfort, Vincent Michel, Bertrand Thirion, Olivier Grisel, Mathieu Blondel, Peter Prettenhofer, Ron Weiss, Vincent Dubourg, Jake Vanderplas, Alexandre Passos, David Cournapeau, Matthieu Brucher, Matthieu Perrot, and \'E douard D...

  33. [41]

    Torrecilla, Miguel Carbajo Berrocal, Pablo Marcos Manchón, and Alberto Suárez

    Carlos Ramos-Carreño, José L. Torrecilla, Miguel Carbajo Berrocal, Pablo Marcos Manchón, and Alberto Suárez. scikit-fda: A P ython P ackage for F unctional D ata A nalysis . Journal of Statistical Software, 109 0 (2): 0 1--37, 2024

  34. [42]

    Piecewise linear regression via a difference of convex functions

    Ali Siahkamari, Aditya Gangrade, Brian Kulis, and Venkatesh Saligrama. Piecewise linear regression via a difference of convex functions. In Hal Daumé III and Aarti Singh, editors, Proceedings of the 37th International Conference on Machine Learning, volume 119 of Proceedings o...

  35. [43]

    Charles J. Stone. Optimal global rates of convergence for nonparametric regression. The Annals of Statistics, 10 0 (4): 0 1040--1053, 1982

  36. [44]

    Least squares estimation of weakly convex functions

    Sun Sun and Yaoliang Yu. Least squares estimation of weakly convex functions. In Kamalika Chaudhuri and Masashi Sugiyama, editors, Proceedings of the Twenty-Second International Conference on Artificial Intelligence and Statistics, volume 89 of Proceedings of Machine Learning ...

  37. [45]

    Fitting piecewise linear continuous functions

    Alejandro Toriello and Juan Pablo Vielma. Fitting piecewise linear continuous functions. European Journal of Operational Research, 219 0 (1): 0 86--95, 2012

  38. [46]

    Empirical Processes in M-Estimation

    Sara van de Geer. Empirical Processes in M-Estimation. Cambridge University Press, 2000

  39. [47]

    Strong and weak convexity of sets and functions

    Jean-Philippe Vial. Strong and weak convexity of sets and functions. Mathematics of Operations Research, 8 0 (2): 0 231--259, 1983

  40. [48]

    Wainwright

    Martin J. Wainwright. High-Dimensional Statistics: A Non-Asymptotic Viewpoint. Cambridge University Press, 2019

  41. [49]

    Geoffrey S. Watson. Smooth regression analysis. Sankhyā: The Indian Journal of Statistics, Series A (1961-2002), 26 0 (4): 0 359--372, 1964

  42. [50]

    C. T. Zahn. Black box maximization of circular coverage. Journal of Research of the National Bureau of Standards -- B. Mathematics and Mathematical Physics, 66B 0 (4): 0 181--216, 1962

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.