pith. sign in

arxiv: 2604.09513 · v1 · submitted 2026-04-10 · 🧮 math.ST · stat.ME· stat.TH

Harmonic Map Regression: Rate-Optimal Nonparametric Estimation on Manifolds with Topological Recovery

Xiaoyu Chen This is my paper

Pith reviewed 2026-05-10 16:27 UTC · model grok-4.3

classification 🧮 math.ST stat.MEstat.TH
keywords harmonic map regressionmanifold-valued regressiontopological recoveryDirichlet energynonparametric estimationharmonic mapsFréchet riskpiecewise-geodesic spline
0
0 comments X

The pith

Harmonic map regression recovers the correct topological class of manifold-valued maps while attaining the minimax nonparametric rate.

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

The paper introduces harmonic map regression, which estimates a map from a domain into a manifold by minimizing the empirical Fréchet risk plus the Dirichlet energy. This penalty imports the classical theory of harmonic maps, so the estimator satisfies an Euler-Lagrange equation that makes its solutions piecewise-geodesic splines, supplies an equivalent kernel for pointwise analysis, and reduces the infinite-dimensional problem to finite dimensions. The same penalty produces topological sensitivity: when the target manifold permits regression curves in distinct homotopy classes separated by geometric energy barriers, the estimator selects the correct class with probability tending to one. A curvature-dependent oracle inequality shows that the procedure attains the minimax rate n^{-2s/(2s+1)} for Sobolev smoothness s, matching the Euclidean rate on non-positively curved manifolds. Five geometric obstructions establish that this complete structural theory, including the topological phase transition, holds only for the Dirichlet energy.

Core claim

By penalizing the empirical Fréchet risk with the Dirichlet energy, the harmonic map regression estimator satisfies an Euler-Lagrange equation that characterizes it as a piecewise-geodesic spline. This connection to harmonic map theory yields an equivalent kernel that controls pointwise risk at the rate n^{-2/3} and reduces the infinite-dimensional variational problem to finite-dimensional optimization. On manifolds with regression curves in distinct homotopy classes, maps from different classes are separated by intrinsic energy barriers, and the Dirichlet penalty renders the estimator sensitive to these barriers, recovering the correct topological class with probability tending to one. A 4.

What carries the argument

The Dirichlet energy penalty applied to the empirical Fréchet risk, which connects the estimator to the Euler-Lagrange equation of harmonic maps and induces sensitivity to homotopy classes via intrinsic energy barriers.

If this is right

  • The estimator attains the minimax rate n^{-2s/(2s+1)} for Sobolev order s on suitable manifolds.
  • The estimator recovers the correct homotopy class with probability tending to one when the manifold admits separated energy barriers.
  • Solutions are piecewise-geodesic splines and the variational problem reduces exactly to a finite-dimensional optimization.
  • The full structural theory, including the topological phase transition, holds uniquely for the Dirichlet energy.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Regression estimators of this form could serve as data-driven probes for manifold topology without requiring the homotopy class to be specified in advance.
  • Other penalty functions may achieve the same rate but would lose the automatic topological selection property.
  • The curvature dependence in the oracle inequality suggests that topological recovery may behave differently on positively curved manifolds even when rates remain optimal.

Load-bearing premise

The target manifold must admit regression curves in distinct homotopy classes that are separated by positive energy barriers intrinsic to its geometry, and the Dirichlet energy must be the unique penalty that produces the full structural theory including piecewise-geodesic solutions and topological sensitivity.

What would settle it

An experiment on a manifold such as the circle or torus in which the estimator selects an incorrect homotopy class with probability bounded away from zero for large sample sizes, or empirical convergence rates that deviate from n^{-2s/(2s+1)} on a non-positively curved target.

Figures

Figures reproduced from arXiv: 2604.09513 by Xiaoyu Chen.

Figure 1
Figure 1. Figure 1: Winding ambiguity in circle-valued regression ( [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Phase transition for homotopy class recovery on [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: A demonstration of the fitted curves on S 2 , H2 , and T 2 (n = 400). replications, with reference slopes n −2/3 (oracle rate at s = 1) and n −1/3 shown as dotted lines [PITH_FULL_IMAGE:figures/full_fig_p024_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Log-log MISE versus sample size on five manifolds. Reference slope [PITH_FULL_IMAGE:figures/full_fig_p025_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Wind direction regression on S 1 (n = 649 hourly observations, June). (a) Fitted curves for all methods. (b) Zoom near wrap-around events. (c) Boxplot of test-set geodesic errors. folds in round-robin order so that each validation fold consists of 4 non-adjacent blocks spread throughout the month. The CV loss is the average held-out geodesic prediction error n −1 val Pd 2 S1 ( ˆf(tj ), Yj ) [PITH_FULL_IMA… view at source ↗
Figure 6
Figure 6. Figure 6: Normalized MISE versus sample size on S 2 (R) for varying curvature κ = 1/R2 . 67 [PITH_FULL_IMAGE:figures/full_fig_p067_6.png] view at source ↗
read the original abstract

We study harmonic map regression, a nonparametric estimator for manifold-valued responses, that penalizes the empirical Fr\'echet risk by the Dirichlet energy. By connecting penalized regression to the theory of harmonic maps, the estimator acquires a structural theory that parallels the classical Euclidean smoothing spline. The Euler-Lagrange equation characterizes the solution as a piecewise-geodesic spline, an equivalent kernel controls pointwise risk at the rate $n^{-2/3}$, and the infinite-dimensional variational problem reduces exactly to a finite-dimensional optimization. Such newly established connection reveals a topological phenomenon that has no analogue in Euclidean nonparametric regression and, to our knowledge, has not been studied in the manifold regression literature. On manifolds whose regression curves can wrap around in topologically distinct ways, maps in distinct homotopy classes are separated by energy barriers intrinsic to the geometry of the target, and the Dirichlet penalty makes the estimator sensitive to this structure, recovering the correct topological class with probability tending to one, a phase transition we call topological recovery. A curvature-dependent oracle inequality yields the minimax rate $n^{-2s/(2s+1)}$ for Sobolev order $s$, matching the Euclidean constant on non-positively curved targets, while five geometric obstructions show that the full structural theory is unique to the Dirichlet energy ($s=1$). Simulations on $S^2$, $\mathbb{H}^2$, $SO(3)$, $\mathrm{Sym}^+(2)$, and $T^2$ corroborate the theory, and an application to wind-direction data on $S^1$ illustrates practical advantages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 3 minor

Summary. The paper introduces harmonic map regression for nonparametric estimation of manifold-valued responses, penalizing the empirical Fréchet risk by the Dirichlet energy. Connecting to classical harmonic map theory yields a piecewise-geodesic spline characterization via the Euler-Lagrange equation, an equivalent kernel giving pointwise rate n^{-2/3}, and an exact reduction of the variational problem to finite-dimensional optimization. The work identifies a topological recovery phenomenon: on manifolds admitting regression curves in distinct homotopy classes, the Dirichlet penalty ensures the estimator recovers the correct class with probability tending to one due to intrinsic energy barriers. A curvature-dependent oracle inequality establishes the minimax rate n^{-2s/(2s+1)} for Sobolev order s (matching Euclidean constants on non-positively curved targets), with uniqueness to s=1 shown via five geometric obstructions. Theory is supported by simulations on S^2, H^2, SO(3), Sym^+(2), T^2 and an application to wind-direction data on S^1.

Significance. If the derivations hold, the manuscript offers a substantive advance by forging a direct link between harmonic maps and manifold regression, producing structural features (piecewise geodesics, topological sensitivity) with no Euclidean analogue. The exact finite-dimensional reduction, rate optimality, and multi-manifold simulations are clear strengths. Topological recovery, if rigorously quantified, would be a genuinely new contribution to geometric statistics.

major comments (3)
  1. [Topological recovery section] The topological recovery claim (abstract and the section developing the homotopy-class argument) asserts that positive energy barriers intrinsic to the target geometry dominate empirical Fréchet fluctuations so that the correct class is recovered with probability tending to one. No explicit quantitative threshold relating barrier height, noise variance, and the smoothing parameter is supplied; without it the global minimizer could select an incorrect class when a finite-sample fit advantage offsets the barrier, as the empirical risk term fluctuates at rate roughly n^{-1/2}.
  2. [Oracle inequality theorem] The curvature-dependent oracle inequality (the theorem establishing the rate n^{-2s/(2s+1)}) is stated to match the Euclidean constant on non-positively curved targets. The precise manner in which sectional curvature enters the leading constant, and whether the bound remains rate-optimal (rather than merely rate-consistent) on positively curved manifolds such as S^2, requires explicit display.
  3. [Finite-dimensional reduction argument] The exact reduction to finite-dimensional optimization (the argument following the Euler-Lagrange characterization) is presented as holding without approximation. The derivation should clarify how the piecewise-geodesic property follows directly from the variational equation and whether any discretization or truncation is hidden in the passage from the infinite-dimensional problem to the finite set of manifold-valued knots.
minor comments (3)
  1. [Introduction] The abstract refers to 'five geometric obstructions' showing uniqueness to the Dirichlet energy; a concise enumeration of these obstructions in the introduction would improve readability.
  2. [Notation and estimator definition] Notation for the smoothing parameter and the precise definition of the Dirichlet energy functional should be fixed at first use and used consistently in all subsequent statements of the estimator and the oracle inequality.
  3. [Simulation studies] In the simulation section, the practical procedure used to identify the recovered homotopy class (e.g., via winding numbers or energy comparison) should be stated explicitly so that the reported success rates can be reproduced.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments. We address each major point below. Where clarification or strengthening is needed, we will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: The topological recovery claim (abstract and the section developing the homotopy-class argument) asserts that positive energy barriers intrinsic to the target geometry dominate empirical Fréchet fluctuations so that the correct class is recovered with probability tending to one. No explicit quantitative threshold relating barrier height, noise variance, and the smoothing parameter is supplied; without it the global minimizer could select an incorrect class when a finite-sample fit advantage offsets the barrier, as the empirical risk term fluctuates at rate roughly n^{-1/2}.

    Authors: We agree that an explicit quantitative threshold would strengthen the argument. The current proof shows that the energy barrier exceeds the O_p(n^{-1/2}) fluctuation of the empirical Fréchet risk with probability tending to one, but does not display a concrete sufficient condition on λ. In the revision we will insert a lemma giving an explicit lower bound on the barrier height relative to the noise level and λ that guarantees topological recovery with probability at least 1 - C exp(-c n). revision: yes

  2. Referee: The curvature-dependent oracle inequality (the theorem establishing the rate n^{-2s/(2s+1)}) is stated to match the Euclidean constant on non-positively curved targets. The precise manner in which sectional curvature enters the leading constant, and whether the bound remains rate-optimal (rather than merely rate-consistent) on positively curved manifolds such as S^2, requires explicit display.

    Authors: Sectional curvature enters the leading constant of the oracle inequality through the comparison estimates used to control the remainder term in the proof of Theorem 3.4; the constant C is monotone in the upper bound on sectional curvature. On non-positively curved targets the constant reduces exactly to the Euclidean value. On positively curved manifolds the rate n^{-2s/(2s+1)} is preserved, but the prefactor is strictly larger. The lower bound in the minimax theorem holds independently of curvature, so the rate remains optimal (up to the curvature-dependent constant). We will display the explicit dependence of C on the curvature bound in the revised statement of the theorem. revision: yes

  3. Referee: The exact reduction to finite-dimensional optimization (the argument following the Euler-Lagrange characterization) is presented as holding without approximation. The derivation should clarify how the piecewise-geodesic property follows directly from the variational equation and whether any discretization or truncation is hidden in the passage from the infinite-dimensional problem to the finite set of manifold-valued knots.

    Authors: The piecewise-geodesic characterization follows directly from the Euler-Lagrange equation for the Dirichlet energy: away from the data points the map satisfies the harmonic-map equation, whose solutions on a Riemannian manifold are geodesics. Because the solution is completely determined by its values at the finitely many knots (the points at which the weak derivative jumps), the infinite-dimensional variational problem reduces exactly to a finite-dimensional optimization over the knot positions on the target manifold; no discretization or truncation is introduced. We will add a short remark after the Euler-Lagrange derivation to make this reduction explicit. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected; derivation rests on external classical theory.

full rationale

The paper connects the Dirichlet-penalized Fréchet estimator to the classical theory of harmonic maps (an external, pre-existing mathematical framework not defined within the paper). The Euler-Lagrange characterization as piecewise-geodesic splines, the exact reduction to finite-dimensional optimization, the energy-barrier separation of homotopy classes, and the curvature-dependent oracle inequality are all derived from this external connection rather than from quantities fitted to the same data or from self-referential definitions. The claim of uniqueness to the Dirichlet energy (s=1) is supported by five geometric obstructions, which are independent checks rather than tautological. No load-bearing step reduces by construction to its own inputs, and no self-citation chain is invoked for the central results. The derivation is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The claims rest on standard Riemannian geometry and Sobolev-space assumptions for the regression function; no new free parameters or invented entities are introduced in the abstract.

free parameters (1)
  • smoothing parameter
    The strength of the Dirichlet penalty; its selection rule is not specified in the abstract.
axioms (2)
  • domain assumption The response space is a compact Riemannian manifold
    Required for Fréchet means, geodesic splines, and homotopy classes to be well-defined.
  • domain assumption The regression map belongs to a Sobolev space of order s on the manifold
    Used to obtain the rate n^{-2s/(2s+1)}.

pith-pipeline@v0.9.0 · 5584 in / 1518 out tokens · 54630 ms · 2026-05-10T16:27:52.385574+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages

  1. [1]

    R. A. Adams and J. J. F. Fournier. Sobolev Spaces . Academic Press, 2nd edition, 2003

    work page 2003

  2. [2]

    B. Afsari. Riemannian L^p center of mass: existence, uniqueness, and convexity . Proc.\ Amer.\ Math.\ Soc., 139 0 (2): 0 655--673, 2011

    work page 2011

  3. [3]

    M. S. Birman and M. Z. Solomjak. Piecewise-polynomial approximations of functions of the classes W_p^ . Math.\ USSR-Sb., 2 0 (3): 0 295--317, 1967

    work page 1967

  4. [4]

    N. Boumal. An Introduction to Optimization on Smooth Manifolds . Cambridge University Press, 2023

    work page 2023

  5. [5]

    Cornea, H

    E. Cornea, H. Zhu, P. Kim, and J. G. Ibrahim. Regression models on R iemannian symmetric spaces . J.\ R.\ Stat.\ Soc.\ Ser.\ B, 79 0 (2): 0 463--482, 2017

    work page 2017

  6. [6]

    M. P. do Carmo. Riemannian Geometry . Birkh\"auser, 1992

    work page 1992

  7. [7]

    Eells and J

    J. Eells and J. H. Sampson. Harmonic mappings of R iemannian manifolds . Amer.\ J.\ Math., 86 0 (1): 0 109--160, 1964

    work page 1964

  8. [8]

    P. T. Fletcher. Geodesic regression and the theory of least squares on R iemannian manifolds . Int.\ J.\ Comput.\ Vision, 105 0 (2): 0 171--185, 2013

    work page 2013

  9. [9]

    M. Gromov. Filling R iemannian manifolds . J.\ Differential Geom., 18 0 (1): 0 1--147, 1983

    work page 1983

  10. [10]

    H\' e lein

    F. H\' e lein. Harmonic Maps, Conservation Laws and Moving Frames . Cambridge University Press, 2002

    work page 2002

  11. [11]

    Hinkle, P

    J. Hinkle, P. T. Fletcher, and S. Joshi. Intrinsic polynomials for regression on R iemannian manifolds . J.\ Math.\ Imaging Vision, 50 0 (1--2): 0 32--52, 2014

    work page 2014

  12. [12]

    G. Y. Jiang. 2-harmonic maps and their first and second variational formulas . Chinese Ann.\ Math.\ Ser.\ A, 7 0 (4): 0 389--402, 1986

    work page 1986

  13. [13]

    J. Jost. Riemannian Geometry and Geometric Analysis . Springer, 7th edition, 2017

    work page 2017

  14. [14]

    H. Karcher. Riemannian center of mass and mollifier smoothing . Comm.\ Pure Appl.\ Math., 30 0 (5): 0 509--541, 1977

    work page 1977

  15. [15]

    K.-R. Kim, I. L. Dryden, H. Le, and K. E. Severn. Smoothing splines on R iemannian manifolds, with applications to 3D shape space . J.\ R.\ Stat.\ Soc.\ Ser.\ B, 83 0 (1): 0 108--132, 2021. doi:10.1111/rssb.12402

    work page doi:10.1111/rssb.12402 2021

  16. [16]

    N. J. Korevaar and R. M. Schoen. Sobolev spaces and harmonic maps for metric space targets . Comm.\ Anal.\ Geom., 1 0 (4): 0 561--659, 1993

    work page 1993

  17. [17]

    Lamm and T

    T. Lamm and T. Rivi\`ere. Conservation laws for fourth order systems in four dimensions . Comm.\ Partial Differential Equations, 33 0 (2): 0 245--262, 2008

    work page 2008

  18. [18]

    John M. Lee. Introduction to R iemannian Manifolds , volume 176 of Graduate Texts in Mathematics. Springer, 2 edition, 2018

    work page 2018

  19. [19]

    Lin and H.-G

    Z. Lin and H.-G. M\" u ller. Total variation regularized F r\' e chet regression for metric-space valued data . Ann.\ Statist., 49 0 (6): 0 3510--3533, 2021

    work page 2021

  20. [20]

    Machado, F

    L. Machado, F. Silva Leite, and K. Krakowski. Higher-order smoothing splines versus least squares problems on R iemannian manifolds . J.\ Dyn.\ Control Syst., 16 0 (1): 0 121--148, 2010. doi:10.1007/s10883-010-9080-1

    work page doi:10.1007/s10883-010-9080-1 2010

  21. [21]

    Muralidharan and P

    P. Muralidharan and P. T. Fletcher. Sasaki metrics for analysis of longitudinal data on manifolds . In Proc.\ IEEE Conf.\ Computer Vision and Pattern Recognition (CVPR), pages 1027--1034, 2012

    work page 2012

  22. [22]

    J. Nash. The imbedding problem for R iemannian manifolds . Ann.\ of Math., 63 0 (1): 0 20--63, 1956

    work page 1956

  23. [23]

    Integrated Surface Database (ISD)

    National Oceanic and Atmospheric Administration . Integrated Surface Database (ISD) . https://www.ncei.noaa.gov/products/land-based-station/integrated-surface-database, 2023. Station USAF 725300, WBAN 94846 (Chicago O'Hare)

    work page 2023

  24. [24]

    Noakes, G

    L. Noakes, G. Heinzinger, and B. Paden. Cubic splines on curved spaces . IMA J.\ Math.\ Control Inform., 6 0 (4): 0 465--473, 1989

    work page 1989

  25. [25]

    Petersen and H.-G

    A. Petersen and H.-G. M\" u ller. Fr\' e chet regression for random objects with E uclidean predictors . Ann.\ Statist., 47 0 (2): 0 691--719, 2019

    work page 2019

  26. [26]

    Schoen and K

    R. Schoen and K. Uhlenbeck. A regularity theory for harmonic maps . J.\ Differential Geom., 17 0 (2): 0 307--335, 1982

    work page 1982

  27. [27]

    Sch\" o tz

    C. Sch\" o tz. Convergence rates for the generalized F r\' e chet mean via the quadruple inequality . Electron.\ J.\ Statist., 13 0 (2): 0 4280--4345, 2019. doi:10.1214/19-EJS1618

    work page doi:10.1214/19-ejs1618 2019

  28. [28]

    Sch\" o tz

    C. Sch\" o tz. Nonparametric regression in nonstandard spaces . Electron.\ J.\ Statist., 16 0 (2): 0 4679--4741, 2022. doi:10.1214/22-EJS2056

    work page doi:10.1214/22-ejs2056 2022

  29. [29]

    X. Shi, M. Styner, J. Lieberman, J. G. Ibrahim, W. Lin, and H. Zhu. Intrinsic regression models for manifold-valued data . In Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 192--199, 2012

    work page 2012

  30. [30]

    B. W. Silverman. Spline smoothing: the equivalent kernel method . Ann.\ Statist., 12 0 (3): 0 898--916, 1984

    work page 1984

  31. [31]

    Steinke and M

    F. Steinke and M. Hein. Non-parametric regression between manifolds . In Advances in Neural Information Processing Systems 21, pages 1561--1568, 2009

    work page 2009

  32. [32]

    C. J. Stone. Optimal global rates of convergence for nonparametric regression . Ann.\ Statist., 10 0 (4): 0 1040--1053, 1982

    work page 1982

  33. [33]

    K.-T. Sturm. Probability measures on metric spaces of nonpositive curvature . In Heat Kernels and Analysis on Manifolds, Graphs, and Metric Spaces, volume 338 of Contemp.\ Math., pages 357--390. Amer.\ Math.\ Soc., 2003

    work page 2003

  34. [34]

    J. Su, S. Kurtek, E. Klassen, and A. Srivastava. Statistical analysis of trajectories on R iemannian manifolds: bird migration, hurricane tracking and video surveillance . Ann.\ Appl.\ Statist., 8 0 (1): 0 530--552, 2014

    work page 2014

  35. [35]

    A. B. Tsybakov. Introduction to Nonparametric Estimation . Springer, 2009

    work page 2009

  36. [36]

    G. Wahba. Spline Models for Observational Data . SIAM, 1990

    work page 1990

  37. [37]

    Y. Yuan, H. Zhu, W. Lin, and J. S. Marron. Local polynomial regression for symmetric positive definite matrices . J.\ R.\ Stat.\ Soc.\ Ser.\ B, 74 0 (4): 0 697--719, 2012

    work page 2012

  38. [38]

    Zhang and S

    H. Zhang and S. Sra. First-order methods for geodesically convex optimization . In Conference on Learning Theory (COLT), pages 1617--1638, 2016

    work page 2016

  39. [39]

    Michele Benzi and Gene H. Golub. Bounds for the entries of matrix functions with applications to preconditioning. BIT, 39 0 (3): 0 417--438, 1999. doi:10.1023/A:1022362401426

    work page doi:10.1023/a:1022362401426 1999

  40. [40]

    R. Bhatia. Positive Definite Matrices . Princeton University Press, 2009

    work page 2009

  41. [41]

    Boucheron, G

    S. Boucheron, G. Lugosi, and P. Massart. Concentration Inequalities: A Nonasymptotic Theory of Independence . Oxford University Press, 2013

    work page 2013

  42. [42]

    2002 , issn =

    O. Bousquet. A B ennett concentration inequality and its application to suprema of empirical processes. In C.\ R.\ Math.\ Acad.\ Sci.\ Paris, volume 334, pages 495--500, 2002. doi:10.1016/S1631-073X(02)02292-6

    work page doi:10.1016/s1631-073x(02)02292-6 2002

  43. [43]

    Cheeger and D

    J. Cheeger and D. G. Ebin. Comparison Theorems in Riemannian Geometry . AMS Chelsea Publishing, revised edition, 2008

    work page 2008

  44. [44]

    Moss, and Philip W

    Stephen Demko, William F. Moss, and Philip W. Smith. Decay rates for inverses of band matrices. Math. Comp., 43 0 (168): 0 491--499, 1984. doi:10.1090/S0025-5718-1984-0758197-9

    work page doi:10.1090/s0025-5718-1984-0758197-9 1984

  45. [45]

    Koltchinskii

    V. Koltchinskii. Local R ademacher complexities and oracle inequalities in risk minimization. Ann.\ Statist., 34 0 (6): 0 2593--2656, 2006. doi:10.1214/009053606000001019

    work page doi:10.1214/009053606000001019 2006

  46. [46]

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

    work page 2000

  47. [47]

    A. W. van der Vaart. Asymptotic Statistics . Cambridge University Press, 1998

    work page 1998