pith. sign in

arxiv: 2604.22320 · v1 · submitted 2026-04-24 · 📊 stat.ME · stat.ML

Nonparametric Estimation of Isotropic Covariance Function

Yiming Wang , Sujit K. Ghosh This is my paper

Pith reviewed 2026-05-08 10:59 UTC · model grok-4.3

classification 📊 stat.ME stat.ML
keywords nonparametric covariance estimationisotropic covarianceBernstein polynomialssieve maximum likelihoodspatial statisticsincreasing domain asymptoticspositive definite functions
0
0 comments X

The pith

A sequence of Bernstein polynomials approximates arbitrary isotropic covariance functions valid in infinite-dimensional space, supporting consistent nonparametric estimation via sieve maximum likelihood.

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

The paper develops a nonparametric approach to estimate isotropic covariance functions that remain valid no matter how many dimensions are involved. It builds an approximating sequence from Bernstein polynomials that preserves the required positive-definiteness properties. A computationally efficient sieve maximum likelihood procedure then estimates the unknown function from spatial observations. Consistency of this estimator is proved when the sampling domain grows with the number of data points. Numerical comparisons on simulated data and a precipitation example show lower bias and smaller L-infinity and L2 errors than standard parametric or other nonparametric alternatives.

Core claim

A nonparametric model is constructed using a sequence of Bernstein polynomials to approximate any isotropic covariance function that is valid in R^∞. Approximation properties are established in the L∞ and L2 norms. A sieve maximum likelihood estimator is developed for the unknown function, and its consistency is proved under an increasing-domain asymptotic regime. The method is shown numerically to reduce bias from model misspecification relative to parametric models and to achieve smaller expected L∞ and L2 norms than existing nonparametric estimators.

What carries the argument

A sequence of Bernstein polynomials that approximates the covariance function while maintaining positive definiteness in infinite dimensions, paired with a sieve maximum likelihood estimator.

If this is right

  • The estimator remains consistent when the observation region expands with the number of samples.
  • It can be applied directly to real spatial datasets such as precipitation measurements.
  • Bias from assuming an incorrect parametric form is reduced compared with standard parametric covariance models.
  • Expected L∞ and L2 approximation errors are lower than those obtained from other nonparametric covariance estimators.

Where Pith is reading between the lines

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

  • The approach could improve spatial prediction accuracy in environmental applications by avoiding errors from choosing the wrong parametric family.
  • Extensions might allow estimation of non-isotropic or non-stationary covariance structures using similar polynomial sieves.
  • In high-dimensional settings the method offers a way to let the data determine the decay rate of dependence without fixing a parametric form in advance.

Load-bearing premise

The unknown covariance function must be continuous, isotropic, and positive definite in every finite dimension, and the data must be collected under an increasing-domain regime in which the spatial region expands with sample size.

What would settle it

The central claim would be falsified if the sieve maximum likelihood estimator fails to converge in probability to the true covariance as the domain expands, or if repeated simulations show no consistent reduction in expected L∞ and L2 errors relative to parametric or competing nonparametric methods.

Figures

Figures reproduced from arXiv: 2604.22320 by Sujit K. Ghosh, Yiming Wang.

Figure 1
Figure 1. Figure 1: presents curves of basis {Ak,m} m k=1 for m = 5 and m = 25. Recall that the approximation result in (9) is based on the assumption that g(s) is bounded on (0, 1) in which case one can define g(0) = lim inf s→0 g(s) and g(1) = lim sup s→1 g(s). Otherwise, there exists no such gm(s) of the form (8) satisfying (9) when g(s) explodes to infinity at the endpoints 0 or 1. For example, when true covariance functi… view at source ↗
Figure 2
Figure 2. Figure 2: Relative approximation error vs. m in (a)–(c), and true vs. ap￾proximation covariance in (d)–(f). Each row includes results for Matern, Cauchy, and Gaussian. 3 Sieve Maximum Likelihood Estimation Method Consider scenarios where observations are generated from a Gaussian pro￾cess with an unknown isotropic covariance function valid in R∞. We are interested in the statistical estimation of the model (e.g., th… view at source ↗
Figure 3
Figure 3. Figure 3: (a) √ semivariance (inch) vs. distance (100 km), and (b) 189 sta￾tions (dots) in the US together with the mean precipitation surface over forty years. In this section, we applied the three aforementioned nonparametric and twelve parametric methods to an annual total precipitation dataset obtained from NOAA. We chose the study region to be between latitudes 30 and 46 22 view at source ↗
Figure 4
Figure 4. Figure 4: (a) Estimated covariance and (b) correlation for candidate view at source ↗
Figure 5
Figure 5. Figure 5: (a) Estimated covariance and (b) semivariance of nonparametric view at source ↗
Figure 6
Figure 6. Figure 6: (a) Estimated covariance and (b) semivariance of parametric meth view at source ↗
Figure 7
Figure 7. Figure 7: Boxplots of estimation results based on 100 MCs. Each row view at source ↗
read the original abstract

A nonparametric model using a sequence of Bernstein polynomials is constructed to approximate arbitrary isotropic covariance functions valid in $\mathbb{R}^\infty$ and related approximation properties are investigated using the popular $L_{\infty}$ norm and $L_2$ norms. A computationally efficient sieve maximum likelihood (sML) estimation is then developed to nonparametrically estimate the unknown isotropic covaraince function valid in $\mathbb{R}^\infty$. Consistency of the proposed sieve ML estimator is established under increasing domain regime. The proposed methodology is compared numerically with couple of existing nonparametric as well as with commonly used parametric methods. Numerical results based on simulated data show that our approach outperforms the parametric methods in reducing bias due to model misspecification and also the nonparametric methods in terms of having significantly lower values of expected $L_{\infty}$ and $L_2$ norms. Application to precipitation data is illustrated to showcase a real case study. Additional technical details and numerical illustrations are also made available.

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 / 2 minor

Summary. The paper constructs a sieve of Bernstein polynomials to nonparametrically approximate isotropic covariance functions that are valid in R^∞, studies the resulting L∞ and L2 approximation properties, develops a computationally efficient sieve maximum likelihood estimator, establishes its consistency under increasing-domain asymptotics, and reports numerical superiority over parametric and existing nonparametric methods in bias and expected error norms, with an application to precipitation data.

Significance. If the central claims hold, the work supplies a flexible nonparametric tool for covariance estimation that avoids parametric misspecification while targeting the strong validity condition required for infinite-dimensional isotropy. The reported numerical gains in L∞/L2 error and the consistency result under increasing domains would be useful additions to the spatial statistics literature, particularly for applications where dimension-independent positive-definiteness matters.

major comments (3)
  1. [§3] §3 (sieve construction): The Bernstein-polynomial sieve is defined directly on the covariance function without coefficient restrictions that enforce complete monotonicity of C(√t) on [0,∞). Schoenberg’s theorem requires this property for validity in every dimension; the paper reports only L∞/L2 approximation error and does not verify or constrain the alternating-derivative sign pattern, so the estimator can converge in the stated norms while producing functions that cease to be positive definite for large spatial dimensions.
  2. [§4] §4 (consistency theorem): The consistency statement for the sieve ML estimator is given without an explicit convergence rate, without conditions on the growth of the polynomial degree with sample size, and without details on how the degree is selected in practice or in the asymptotics. This leaves the result non-quantitative and makes it impossible to judge whether the sieve dimension choice is compatible with the increasing-domain regime.
  3. [Numerical results] Numerical results section (Tables/Figures on L∞ and L2 errors): The computation of the reported expected L∞ and L2 norms is not described (e.g., how the supremum is discretized, whether the estimated functions are projected onto the valid cone, or the precise data-exclusion rules used to avoid circularity). Without these details the claimed superiority over other nonparametric methods cannot be reproduced or assessed.
minor comments (2)
  1. [Abstract] Abstract: “covaraince” is misspelled; “couple of existing nonparametric” should read “a couple of existing nonparametric.”
  2. [§2–3] Notation: The transition from the population covariance C(h) to the sieve approximant is not clearly distinguished from the estimated version; a single symbol is used in several places, which obscures whether statements refer to approximation error or estimation error.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive report. We address each major comment below and will revise the manuscript to incorporate the suggested improvements.

read point-by-point responses

  1. Referee: [§3] §3 (sieve construction): The Bernstein-polynomial sieve is defined directly on the covariance function without coefficient restrictions that enforce complete monotonicity of C(√t) on [0,∞). Schoenberg’s theorem requires this property for validity in every dimension; the paper reports only L∞/L2 approximation error and does not verify or constrain the alternating-derivative sign pattern, so the estimator can converge in the stated norms while producing functions that cease to be positive definite for large spatial dimensions.

    Authors: We agree that Schoenberg’s theorem requires complete monotonicity of C(√t) for the covariance to be valid in all dimensions, and that the current sieve construction does not explicitly impose the alternating-sign coefficient restrictions needed to guarantee this property. While Bernstein polynomials can approximate completely monotone functions, the lack of explicit constraints means the estimator could in principle produce non-valid functions for large dimensions. We will revise §3 to incorporate the necessary coefficient restrictions (or a post-estimation projection step) that enforce complete monotonicity, and we will verify that the resulting sieve satisfies the required sign pattern. revision: yes

  2. Referee: [§4] §4 (consistency theorem): The consistency statement for the sieve ML estimator is given without an explicit convergence rate, without conditions on the growth of the polynomial degree with sample size, and without details on how the degree is selected in practice or in the asymptotics. This leaves the result non-quantitative and makes it impossible to judge whether the sieve dimension choice is compatible with the increasing-domain regime.

    Authors: The consistency theorem in §4 is established under increasing-domain asymptotics, with the sieve dimension permitted to grow with sample size. However, we acknowledge that the statement lacks an explicit rate, precise growth conditions on the polynomial degree, and practical selection guidance, rendering it non-quantitative. We will revise the theorem and surrounding discussion to include these details, specifying admissible growth rates for the sieve dimension that remain compatible with the increasing-domain regime and describing the data-driven selection procedure used in the numerical studies. revision: yes

  3. Referee: Numerical results section (Tables/Figures on L∞ and L2 errors): The computation of the reported expected L∞ and L2 norms is not described (e.g., how the supremum is discretized, whether the estimated functions are projected onto the valid cone, or the precise data-exclusion rules used to avoid circularity). Without these details the claimed superiority over other nonparametric methods cannot be reproduced or assessed.

    Authors: We agree that the numerical section lacks sufficient detail on the computation of the expected L∞ and L2 norms. We will expand this section to describe the discretization grid used for the supremum, whether any projection onto the valid cone was performed, and the exact cross-validation or data-exclusion rules employed to prevent circularity, thereby enabling full reproducibility and assessment of the reported superiority. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper defines a Bernstein polynomial sieve directly on the covariance function, separately establishes its uniform and L2 approximation properties for continuous functions, then defines the sieve ML estimator via likelihood maximization over that sieve and proves consistency under increasing-domain asymptotics using standard arguments. None of these steps reduce the estimator, the consistency theorem, or the validity claim to a quantity fitted from the same data or to a self-citation whose content is the target result itself. The numerical comparisons and real-data illustration are post-derivation evaluations and do not feed back into the theoretical claims.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed on abstract only; therefore the ledger records only the minimal assumptions stated or implied by the abstract. No explicit free parameters or invented entities are named.

axioms (2)
  • domain assumption The target covariance function is continuous and isotropic on R^∞
    Required for the Bernstein approximation to be uniformly valid and for the sieve to be well-defined.
  • domain assumption Observations follow an increasing-domain asymptotic regime
    Invoked to obtain consistency of the sieve maximum-likelihood estimator.

pith-pipeline@v0.9.0 · 5460 in / 1496 out tokens · 46056 ms · 2026-05-08T10:59:55.493179+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

37 extracted references · 37 canonical work pages

  1. [1]

    (1965), Handbook of Mathematical Functions, 9th ed., New York

    Abramowitz, M., and Stegun, I. (1965), Handbook of Mathematical Functions, 9th ed., New York

    work page 1965

  2. [2]

    (1985), Advanced Econometrics, Harvard University Press, pp

    Amemiya, T. (1985), Advanced Econometrics, Harvard University Press, pp. 105--158. Asymptotic properties of extremum estimators

    work page 1985

  3. [3]

    (1970), Estimation of covariance matrices which are linear combinations or whose inverse are linear combinations of given matrices , Essays in Probability and Statistics, pp

    Anderson, T.W. (1970), Estimation of covariance matrices which are linear combinations or whose inverse are linear combinations of given matrices , Essays in Probability and Statistics, pp. 1--24

    work page 1970

  4. [4]

    (2014), Asymptotic analysis for the role of spatial sampling for covariance parameter estimation of Gaussian processes , Journal of Multivariate Analysis, pp

    Bachoc, F. (2014), Asymptotic analysis for the role of spatial sampling for covariance parameter estimation of Gaussian processes , Journal of Multivariate Analysis, pp. 1--25

    work page 2014

  5. [5]

    (2021), Oxford Handbook of Innovation, Springer, Cham, chap

    Bachoc, F. (2021), Oxford Handbook of Innovation, Springer, Cham, chap. Asymptotic analysis of Maximum Likelihood Estimation of Covariance Parameters for Gaussian Processes: An Introduction with Proofs, pp. 283--303

    work page 2021

  6. [6]

    (2016), On the smallest eigenvalues of covariance matrices of multivariate spatial processes , Stat, 1, 102--107

    Bachoc, F., and Furrer, R. (2016), On the smallest eigenvalues of covariance matrices of multivariate spatial processes , Stat, 1, 102--107

    work page 2016

  7. [7]

    (1912), Demonstration of a theorem of Weierstrass based on the calculus of probabilities , Communications of the Kharkov Mathematical Society

    Bernstein, S. (1912), Demonstration of a theorem of Weierstrass based on the calculus of probabilities , Communications of the Kharkov Mathematical Society

    work page 1912

  8. [8]

    (2007), Large Sample Sieve Estimation of Semi-Nonparametric Models , in Handbook of Econometrics, Vol

    Chen, X. (2007), Large Sample Sieve Estimation of Semi-Nonparametric Models , in Handbook of Econometrics, Vol. 6B, 1st ed., eds. J. Heckman and E. Leamer, chap. 76

    work page 2007

  9. [9]

    (2013), Nonparametric Estimation of Spatial and Space-Time Covariance Function , Journal of Agricultural, Biological, and Environmental Statistics, 18, 611--630

    Choi, I., Li, B., and Wang, X. (2013), Nonparametric Estimation of Spatial and Space-Time Covariance Function , Journal of Agricultural, Biological, and Environmental Statistics, 18, 611--630

    work page 2013

  10. [10]

    (2004), Bayesian estimation of the spectral density of a time series , Journal of the American Statistical Association, 99, 1050--1059

    Choudhuri, N., Ghosal, S., and Roy, A. (2004), Bayesian estimation of the spectral density of a time series , Journal of the American Statistical Association, 99, 1050--1059

    work page 2004

  11. [11]

    (1993), Statistics for Spatial Data, John Wiley & Sons, New York

    Cressie, N. (1993), Statistics for Spatial Data, John Wiley & Sons, New York

    work page 1993

  12. [12]

    (1995), A Strong Law for Weighted Sums of i.i.d Random Variables , Statistics & Probability Letters, 76, 1482--1487

    Cuzik, J. (1995), A Strong Law for Weighted Sums of i.i.d Random Variables , Statistics & Probability Letters, 76, 1482--1487

    work page 1995

  13. [13]

    (2004), Least angle regression , Annals of Statistics, 32, 407--451

    Efron, B., Hastie, T., Johnstone, I., and Tibshirani, R. (2004), Least angle regression , Annals of Statistics, 32, 407--451

    work page 2004

  14. [14]

    (2012), The Bernstein polynomial basis: a centennial retrospective , Computer Aided Geometric Design, 29, 379--419

    Farouki, R.T. (2012), The Bernstein polynomial basis: a centennial retrospective , Computer Aided Geometric Design, 29, 379--419

    work page 2012

  15. [15]

    (1981), Abstract Inference, Wiley, New York

    Grenander, U. (1981), Abstract Inference, Wiley, New York

    work page 1981

  16. [16]

    (1991), Topics in Matrix Analysis, Cambridge University Press

    Horn, R., and Johnson, C. (1991), Topics in Matrix Analysis, Cambridge University Press

    work page 1991

  17. [17]

    (2011), Nonparametric estimation of the variogram and its spectrum , Biometrika, 98, 775--789

    Huang, C., Hsing, T., and Cressie, N. (2011), Nonparametric estimation of the variogram and its spectrum , Biometrika, 98, 775--789

    work page 2011

  18. [18]

    (1997), Sieve Estimation for the Proportional-Odds Failure-Time Regression Model With Interval Censoring , Journal of the American Statistical Association, 92, 960--967

    Huang, J., and Rossini, A.J. (1997), Sieve Estimation for the Proportional-Odds Failure-Time Regression Model With Interval Censoring , Journal of the American Statistical Association, 92, 960--967

    work page 1997

  19. [19]

    (2007), Semiparametric Estimation of Spectral Density With Irregular Observations , Journal of the American Statistical Association, 102, 726--735

    Im, H., Stein, M., and Zhu, Z. (2007), Semiparametric Estimation of Spectral Density With Irregular Observations , Journal of the American Statistical Association, 102, 726--735

    work page 2007

  20. [20]

    (2008), Tapering for Likelihood-Based Estimation in Large Spatial Data Sets , Journal of the American Statistical Association, pp

    Kaufman, C.G., Schervish, M.J., and Nychka, D.W. (2008), Tapering for Likelihood-Based Estimation in Large Spatial Data Sets , Journal of the American Statistical Association, pp. 1545--1555

    work page 2008

  21. [21]

    (2016), On the consistency of inversion-free parameter estimation for Gaussian random fields , Journal of Multivariate Analysis, 150, 245--266

    Keshavarz, H., Scott, C., and Nguyen, X. (2016), On the consistency of inversion-free parameter estimation for Gaussian random fields , Journal of Multivariate Analysis, 150, 245--266

    work page 2016

  22. [22]

    (1991), Uniform Convergence in Probability and Stochastic Equicontinuity , Econometrica, 59, 1161--1167

    Newey, W.K. (1991), Uniform Convergence in Probability and Stochastic Equicontinuity , Econometrica, 59, 1161--1167

    work page 1991

  23. [23]

    (1994), Handbook of Econometrics, Elsevier Science, chap

    Newey, W.K., and McFadden, D. (1994), Handbook of Econometrics, Elsevier Science, chap. 4, pp. 2111--2245. Large sample estimation and hypothesis testing

    work page 1994

  24. [24]

    Global Summary of The Month (GSOY) data retrieved at https://www.ncei.noaa.gov/data/gsoy/

    NOAA (2020). Global Summary of The Month (GSOY) data retrieved at https://www.ncei.noaa.gov/data/gsoy/

    work page 2020

  25. [25]

    (2017), Maximum likelihood estimation for linear Gaussian covariance models , Journal of Royal Statistical Society, 79, 1269--1292

    Piotr, Z., Caroline, U., and Donald, R. (2017), Maximum likelihood estimation for linear Gaussian covariance models , Journal of Royal Statistical Society, 79, 1269--1292

    work page 2017

  26. [26]

    (2012), Nonparametric Bayesian models for a spatial covariance , Statistical Methodology, 9, 265--274

    Reich, B., and Fuentes, M. (2012), Nonparametric Bayesian models for a spatial covariance , Statistical Methodology, 9, 265--274

    work page 2012

  27. [27]

    (1938), Metric spaces and completely monotone functions , Annals of Mathematics, 39, 811--841

    Schoenberg, I. (1938), Metric spaces and completely monotone functions , Annals of Mathematics, 39, 811--841

    work page 1938

  28. [28]

    (2012), Tapered Covariance: Bayesian Estimation and Asymptotics , Journal of Computational and Graphical Statistics, 21, 433--452

    Shaby, B., and Ruppert, D. (2012), Tapered Covariance: Bayesian Estimation and Asymptotics , Journal of Computational and Graphical Statistics, 21, 433--452

    work page 2012

  29. [29]

    (1994), Convergence rate of sieve estimates , Annals of Statistics, 22

    Shen, X., and Wong, W. (1994), Convergence rate of sieve estimates , Annals of Statistics, 22

    work page 1994

  30. [30]

    (1999), Interpolation of Spatial Data: Some Theory for Kriging, Springer, New York

    Stein, M. (1999), Interpolation of Spatial Data: Some Theory for Kriging, Springer, New York

    work page 1999

  31. [31]

    (1996), Weak Convergence and Empirical Processes, Springer, New York

    van der Vaart, A.W., and Wellner, J.A. (1996), Weak Convergence and Empirical Processes, Springer, New York

    work page 1996

  32. [32]

    Xue, H., Miao, H., and Wu, H. (2010), Sieve estimation of constant and time-varying coefficients in nonlinear ordinary differential equation models by considering both numerical error and measurement error , Annals of Statistics, 38, 2351--2387

    work page 2010

  33. [33]

    (2022), Covariance Function Estimation and Causal Inference Method, North Carolina State University

    Wang, Y. (2022), Covariance Function Estimation and Causal Inference Method, North Carolina State University

    work page 2022

  34. [34]

    (2004), Inconsistent Estimation and Asymptotically Equivalent Interpolations in Model-based Geostatistics , Journal of the American Statistical Association, 99, 250--261

    Zhang, H. (2004), Inconsistent Estimation and Asymptotically Equivalent Interpolations in Model-based Geostatistics , Journal of the American Statistical Association, 99, 250--261

    work page 2004

  35. [35]

    (2010), Nonparametric Bayesian Inference for the Spectral Density Function of a Random Field , Biometrika, 97, 238--245

    Zheng, Y., Zhu, J., and Roy, A. (2010), Nonparametric Bayesian Inference for the Spectral Density Function of a Random Field , Biometrika, 97, 238--245

    work page 2010

  36. [36]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence...

    work page

  37. [37]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in " " * FUNCTION format....

    work page