On estimation and prediction in spatial functional linear regression model

Guy Martial Nkiet; Sophie Dabo-Niang; St\'ephane Bouka

REVIEW 3 major objections 6 minor 23 references

On estimation and prediction in spatial functional linear regression model

T0 review · 3 major / 6 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Under polynomial mixing spatial dependence, the smoothing-spline slope estimator in a functional linear regression attains the same convergence rate as for independent data, up to a logarithmic factor, and prediction at a non-visited site…

desk verdict A real but incomplete extension of functional linear regression to spatial mixing: Theorem 1 has substance, but the main rates are unproved and Theorem 2's proof has an L^1 gap. read the letter →

arxiv 1908.02143 v1 pith:3WPXAJNU submitted 2019-08-03 math.ST stat.TH

Stéphane Bouka , Sophie Dabo-Niang , Guy Martial Nkiet This is my paper · ORCID

classification math.STstat.TH MSC 60G6062F12

keywords functionallinearregressionspatialdatasmoothingsplinesmixingdependenceslopefunctionestimationpredictionerrorrandomfields

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper extends the smoothing-spline estimator for functional linear regression from independent functional data to data collected on a spatial lattice with polynomial mixing dependence. It aims to show that the unknown slope function $\beta$ can still be estimated at essentially the same rate as in the non-spatial setting, and that prediction at a non-visited location is consistent. The estimation rate, in the $\Gamma$-seminorm, is $\|\hat{\beta}-\beta\|_{\Gamma}^{2}=O_p(\rho +(n^d \rho^{1/(2m+2q+1)})^{-1}\ln n + n^{-d(2q+1)/2})$ as $\rho\to 0$, and the prediction error at a non-visited site is $O_p(n^{-d/(2m+2q+2)})$. If correct, this makes functional linear regression usable for geostatistical and gridded spatial data under mild conditions on how dependence decays.

What carries the argument

The central object is the smoothing spline estimator $\hat{\beta}(t)=D(t)^\top(D^\top D)^{-1}D^\top \tilde{\beta}$, where $\tilde{\beta}=(n^{-d}p^{-1}X^\top X+\rho A_m)^{-1} n^{-d}p^{-1}X^\top Y$, with $A_m$ a B-spline roughness penalty matrix, $\rho$ a smoothing parameter, and $D$ a basis for splines of order $m$. The load-bearing mechanism is the trace inequality $\mathrm{tr}(M^2)\le \mathrm{tr}(M)$ for the design matrix $M=(n^{-d}p^{-1}X^\top X+\rho A_m)^{-1}(n^{-d}p^{-1}X^\top X)$; this controls the diagonal variance contribution. The mixing assumption then converts all off-diagonal spatial covariance sums into bounded multiples of $\ln n$ and of $\sum_t t^{d-1}\alpha_{1,\infty}(t)$, giving the rate by balancing penalty bias $\rho$, variance, and discretization error.

What would settle it

Write out the variance-bias decomposition of $\|\hat{\beta}-\beta\|_{\Gamma}^{2}$ for a concrete spatially correlated Gaussian noise with exponential covariance on a $d$-dimensional grid and check whether the off-diagonal spatial terms are absorbed by the $\ln n$ factor in Corollary 1; if a term of larger order (such as $n^d$ times a covariance sum) survives, the rate is false. A simulation with fixed $\rho$ and growing $n$ that does not show the predicted decay would also be evidence against the corollary.

Watch

Extended reading notes

Core claim

The central claim is that stationary polynomial mixing spatial dependence does not change the rate of the penalized least-squares smoothing spline estimator. Concretely, the paper proves a finite-sample variance bound for the estimator conditional on the design functions (Theorem 1), and from it derives the estimation rate in Corollary 1 and the prediction rate $E((\hat{Y}_{i_0}-Y^*_{i_0})^2 \mid \hat{\beta}_0,\hat{\beta}) = O_p(n^{-d/(2m+2q+2)})$ at a non-visited site (Theorem 2). The variance bound contains the mixing effect as a $\ln n$ factor, while the bias and discretization terms are taken over from the independent-data analysis. The prediction result covers sites both inside and beyond the observation grid, which the paper contrasts with earlier spatial functional regression with derivatives that only handled sites beyond the grid.

Load-bearing premise

The load-bearing premise is that the proof of the independent-data smoothing spline estimator in reference [7] transfers unchanged to spatial mixing data; Corollary 1 is asserted by analogy with [7], and Theorem 2 inherits it, so if that transfer fails the rates are unsupported.

Editorial extensions

If this is right

The rates for slope estimation and prediction from the independent functional linear regression model carry over to spatially dependent data with only a logarithmic penalty coming from the mixing assumption.
The method provides prediction at an unvisited site both inside and beyond the observation grid, a feature the paper highlights as an extension over derivative-based spatial functional regression.
The smoothing parameter choice $\rho\sim n^{-d(2m+2q+1)/(2m+2q+2)}$ gives an explicit guideline for balancing bias and variance in spatial applications.
Because the prediction rate improves as the dimension $d$ of the spatial lattice grows, larger spatial datasets yield faster convergence in this model, other things equal.

Reading between the lines

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

Corollary 1, the paper's main rate, is asserted by arguing as in [7] rather than proved; a reader should treat the transfer of the independent-data bias decomposition to the spatial setting as an assumption until the details are written out.
The separable covariance structure in Assumption 8 is stronger than polynomial mixing alone; if it fails while mixing holds, the prediction-error proof may need an additional term controlling spatial cross-covariances of $X$.
A natural testable extension is data-driven selection of $\rho$ by spatial cross-validation; the paper's asymptotic choice of $\rho$ is not compared against such a procedure in the simulations.
The condition that the unvisited site sits at distance at least $n^{2d/\theta}$ means the prediction consistency is only for sites sufficiently separated from the sample; if $\theta$ is not large, prediction very close to the observed grid is not covered by Theorem 2.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Dear colleague,

Here's the short version: the paper extends Crambes–Kneip–Sarda's smoothing spline estimator for functional linear regression to spatially mixing data. That is a real gap, and the proof of Theorem 1 (the conditional variance bound) is a genuine technical contribution: it uses mixing inequalities and handles the correlation of the noise terms. If the rates are right, it's a useful paper for spatial functional data analysis.

But the paper's central quantitative claims are not actually proved. Corollary 1, the estimation rate, is stated with no proof; the text says 'arguing as in [7]' and moves on. That would be acceptable if the adaptation to spatial mixing were trivial, but it isn't—the proof of Theorem 1 needed a page of extra work just for the variance bound. The same adaptation has to be checked for the bias and approximation terms.

There's a second gap in Theorem 2. The proof bounds the expected prediction error by K5/(n^d ρ) + K6 E[||β̂−β||²_Γ], then applies Corollary 1. But Corollary 1 is O_p, not an expectation bound. The a.s. bound from Lemma 2, ||β̂−β||² = O(1/ρ), is too crude: it gives a term of order n^d times the claimed rate. So the prediction rate does not follow as written. You'd need an L^1 version of Corollary 1, or a separate moment argument.

Other issues: the simulation table has a likely typo (0.073 at n²=10², snr=5%, Case A), there are no error bars or code, and the conclusion's claim about predicting inside vs outside the grid is not backed by theory. None of this is fatal to the underlying approach, and I don't see circularity or data fabrication. The citations look appropriate; the estimator is indeed Crambes et al.'s, and the paper says so.

Who is this for? Researchers working on spatial functional data who want a starting point for rates under mixing. Right now the paper is more of a promising sketch than a finished theory. I'd send it to a serious referee with a strong request to obtain the missing proofs—especially the L^1 rate—before publication. It's worth engaging with, but not at face value.

Referee Report

3 major / 6 minor

Summary. This paper considers the spatial functional linear regression model Y_i = β0 + ∫ β(t)X_i(t)dt + ε_i for i ∈ Z^d, where the covariates are square-integrable spatial functional processes and the errors may be spatially dependent. The authors propose a smoothing-spline estimator of the slope function β along the lines of Crambes–Kneip–Sarda (2009). The main results are Theorem 1, a finite-sample conditional variance bound for the spline estimator under α-mixing; Corollary 1, an estimation rate for ||β̂ − β||²_Γ; and Theorem 2, a prediction-error bound at a non-visited site under an additional distance assumption. A simulation study on a two-dimensional grid illustrates the method.

Significance. If the stated rates are correct, this is a useful extension of functional linear regression to spatially dependent functional data: it provides explicit rates for estimation and for prediction at unsampled locations, and the technical framework (mixing random fields plus smoothing splines) is well motivated. The paper gives a relatively detailed proof of the conditional variance bound in Theorem 1 and a clear parametric rate for the prediction error. However, the central estimation rate is not proved in the manuscript and is instead deferred to an external reference, and the proof of Theorem 2 contains a separate gap where an O_p rate is used as an L^1 bound. The simulation is illustrative but its 'inside the grid' case does not satisfy the theorem's assumptions. The contribution is potentially valuable, but the main quantitative claims are unsupported as written.

major comments (3)

[Section 3.2, Corollary 1] Corollary 1 is the paper's main estimation-rate result, but it is not proved. The text says only 'Using Theorem 1 and Arguing as in [7], we obtain the Corollary below.' Theorem 1 in Crambes et al. (2009) is for independent functional data; carrying its proof over to a spatially α-mixing field requires controlling the bias term E(β̂) − β under spatial dependence and reworking the discretization and eigenvalue-sum arguments under Assumptions 2–4 and 8. None of these steps is supplied. Since Theorem 2 explicitly invokes Corollary 1, the paper's rates for both estimation and prediction rest on an unverified transfer of the proof in [7].
[Section 5.3.2, proof of Theorem 2] The proof of Theorem 2 derives B ≤ K5/(n^d ρ) + K6 E[||β̂ − β||²_Γ] and then concludes by 'Applying Corollary 1'. Corollary 1 provides only ||β̂ − β||²_Γ = O_p(ρ + (n^d ρ^{1/(2m+2q+1)})^{-1} ln n + n^{-d(2q+1)/2}); an O_p rate does not imply that the expectation has the same rate. Lemma 2 gives only the much cruder a.s. bound O(1/ρ). With ρ chosen as n^{-d(2m+2q+1)/(2m+2q+2)}, the bound O(1/ρ) is n^{d(2m+2q+1)/(2m+2q+2)}, which is far larger than the claimed n^{-d/(2m+2q+2)}. A separate L^1 (or higher-moment) version of Corollary 1, or a direct argument for the conditional prediction error, is required; as written, Theorem 2 is not established even if Corollary 1 is true.
[Assumption 9 and Section 4] Assumption 9 requires the non-visited site to be at distance at least n^{2d/θ} from all sampled sites. For large θ this threshold tends to 1, and for any θ it is at least 1 when n > 1, so the assumption excludes all sites within distance 1 of the sampling grid; in particular it cannot hold for a site inside the grid. Section 4 nevertheless reports prediction at i0 = (13.5, 5) 'inside the grid' for n = 15, where the distance to the nearest sampled site is 0.5, so the displayed simulation lies outside the theorem's assumptions. The statement in Section 3.2 that 'it is sufficient to choose θ large for doing the prediction at any non-visited site' is therefore inaccurate: for θ large the distance threshold tends to 1, and sites closer than 1 remain excluded.

minor comments (6)

[Section 4, Table 1] The entry 0.073 for case A, snr = 5%, n^2 = 10^2 is an order of magnitude larger than the adjacent entries and appears to be a typo; please check and correct it.
[Section 4] The text refers to the semi-norm ‖.‖_Γ 'defined in (15)', but that semi-norm is defined in equation (5); the cross-reference should be corrected.
[Section 5.2, equation (15)] Inequality (15) has c1 ln n / n^{2d} in the second factor, while Theorem 1 states c ln n / n^d. The theorem's form is weaker and therefore a valid upper bound, but the notation should be aligned to avoid the appearance of an error.
[Assumption 3 and Theorem 2] Assumption 3 states q ∈ (0,1), but Theorem 2 requires 2q > 1; this additional restriction should be stated explicitly in the assumptions rather than appearing only inside the theorem.
[Conclusion and Section 3.2] There are several language issues, e.g., 'The main diﬃcult is technical' should be 'The main difficulty is technical', and 'it is suﬃcient to choice θ large' should be 'it is sufficient to choose θ large'.
[References] Reference [21] displays a duplicated author field ('M.D., Ruiz-Medina MD'); the entry should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the estimator and rates are inherited from an external source, and the paper's gaps are omissions, not self-referential reductions.

full rationale

The paper's central quantities are not defined in terms of their own outputs. The spline estimator in (2)-(3) follows the construction of Crambes, Kneip and Sarda [7], an external published source, and the paper's contribution is carrying that construction over to polynomial mixing spatial fields. Corollary 1 is asserted by 'Using Theorem 1 and Arguing as in [7]' rather than proved; this is an omitted verification, and Theorem 2's final step invokes an O_p rate where an L^1 bound would be needed. These are gaps in mathematical support, not circular reductions. The only self-citation, [5], is used for comparison ('compared to [5]') and as prior context, not as the load-bearing justification for a rate or uniqueness claim. No equation in the paper is shown to be equivalent to its own input by construction, no fitted parameter is renamed as a prediction, and no uniqueness theorem is imported from the authors' prior work. Hence no circular step is established.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The estimation procedure is borrowed from Crambes et al. (2009); the paper contributes novel proofs under spatial mixing, but the central corollary is not derived here.

free parameters (1)

smoothing parameter ρ = not specified (tuning parameter)
The estimator in (2)-(3) depends on ρ; the rates assume ρ is chosen in certain asymptotic ranges, but the paper does not give a data-driven selection for ρ.

assumptions (5)

domain assumption Assumptions 5-6: strict stationarity and polynomial strong mixing of the error field and the (X,Y) field
The proofs of Theorems 1 and 2 rely on mixing inequalities from Tran (1990).
domain assumption Assumption 7: the functional predictor is bounded almost surely
Used in Lemma 2 and the proof of Theorem 2 to bound moments.
domain assumption Assumption 8: isotropic separable covariance structure for X
Used to bound the second term B2 in the proof of Theorem 2.
ad hoc to paper Assumption 9: the non-visited site is at distance at least n^{2d/θ} from sampled sites
This ensures the mixing coefficient between the new site and the sample is small enough to yield the prediction rate; it is a condition made specifically for the prediction result.
ad hoc to paper The adapted argument from Crambes et al. (2009) for Corollary 1 is valid under spatial dependence
Corollary 1 is not proved; it is deferred to 'Arguing as in [7]'.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On estimation and prediction in spatial functional linear regression model." pith.science (2026). https://pith.science/paper/3WPXAJNU

@misc{pith2026190802143,
  author       = {Pith},
  title        = {Pith review of: On estimation and prediction in spatial functional linear regression model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3WPXAJNU}},
  note         = {Machine review of arXiv:1908.02143}
}

read the original abstract

We consider a spatial functional linear regression, where a scalar response is related to a square integrable spatial functional process. We use a smoothing spline estimator for the functional slope parameter and establish a finite sample bound for variance of this estimator under mixing spatial dependence. Then, we give a bound of the prediction error. Finally, we illustrate our results by simulations

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

[7]

Smoothing splines estimators for fu nc- tional linear regression

Crambes C, Kneip A, Sarda P. Smoothing splines estimators for fu nc- tional linear regression. Ann. Statist. 2009; 37: 35–72

work page 2009
[1]

On the Prediction of Stationary Fu nc- tional Time Series

Aue A, Norinho DD, Hormann S. On the Prediction of Stationary Fu nc- tional Time Series. J. Amer. Statist. Assoc. 2015; 110:378–392

work page 2015
[2]

Nonparamatric Spatial Prediction

Biau G, Cadre B. Nonparamatric Spatial Prediction. Stat. Infer ence Stoch. Process. 2004; 7:327–349. 16

work page 2004
[3]

Optimal sampling for spatial pr e- diction of functional data

Bohorquez M, Giraldo R, Matheu J. Optimal sampling for spatial pr e- diction of functional data. Stat. Methods Appl. 2016; 25: 39–54

work page 2016
[4]

Multivariate functional rando m ﬁelds: prediction and optimal sampling

Bohorquez M, Giraldo R, Matheu J. Multivariate functional rando m ﬁelds: prediction and optimal sampling. Stoch. Environ. Res. Risk As - sess. 2017; 31:53–70

work page 2017
[5]

Bouka S, Dabo-Niang S, Nkiet GM. (2018). On estimation in spatial functional regression with derivatives. C. R. Acad. Sci. Paris Ser. I 2018; 356:558–562

work page 2018
[6]

Adaptive functional linear regression

Comte F, Johannes J. Adaptive functional linear regression. An n. Statist. 2012; 40:2765–2797

work page 2012
[8]

A partial overview of the theory of statistics with fun ctional data

Cuevas A. A partial overview of the theory of statistics with fun ctional data. J. Statist. Plan. Inf. 2014; 147:1-23

work page 2014

Show all 23 references

[9]

Smoothing parameter se lection methods for nonparametric regression with spatially correlated er rors

Francisco-Fernandez M, Opsomer JD. Smoothing parameter se lection methods for nonparametric regression with spatially correlated er rors. Canad. J. Statist. 2005; 33:279-295

2005
[10]

Asymptotic normality of the Parzen-Rosenbla tt den- sity estimator for strongly mixing random ﬁeldsStat

El Machkouri M. Asymptotic normality of the Parzen-Rosenbla tt den- sity estimator for strongly mixing random ﬁeldsStat. Inference St och. Process. 2011;, 14, 73–84

2011
[11]

geofd: An R Package for Function- Valued Geostatistical Prediction

Giraldo R, Mateu J, Delicado. geofd: An R Package for Function- Valued Geostatistical Prediction. Rev. Colombiana Estadst. 2012; 35:385 -407

2012
[12]

Cokriging based on curves, prediction and estimation o f the prediction variance, InterStat 2014; 2:1–30

Giraldo R. Cokriging based on curves, prediction and estimation o f the prediction variance, InterStat 2014; 2:1–30

2014
[13]

Statistical modelin g of spatial data big data: An approach from a functional data analysis perspective

Giraldo R, Dabo-Niang S, Mart ´ ınez S (2018). Statistical modelin g of spatial data big data: An approach from a functional data analysis perspective. Statist. Probab. Lett. 2018; 136:126–129

2018
[14]

Ordinary kriging for function-v alued spatial data

R., Giraldo R, Delicado P, Mateu J. Ordinary kriging for function-v alued spatial data. Environ. Ecol. Stat. 2011; 18:411–426. 17

2011
[15]

Weakly dependent functional data

H¨ ormann S, Kokoszka P. Weakly dependent functional data. Ann. Statist. 2010; 38:1845–1884

2010
[16]

Functional linear regression with points o f impact

Kneip A, Po D, Sarda P. Functional linear regression with points o f impact. Ann. Statist. 2016; 44:1–30

2016
[17]

Functional principal component analysis of spatially correlated data

Liu C, Ray S, Hooker G. Functional principal component analysis of spatially correlated data. Stat. Comput. 2017; 27:1639–1654

2017
[18]

Functional linear regression with derivat ives

Mas A, Pumo B (2009). Functional linear regression with derivat ives. J. Nonparametr. Stat. 2009; 21:19–40

2009
[19]

Cokriging for spatial f unctional data

Nerini D, Monestiez P, Mant´ e C (2010). Cokriging for spatial f unctional data. J. Multivariate Anal. 2010; 101:409–418

2010
[20]

Spatial autoregressive and moving average hilb ertian processes

Ruiz-Medina MD. Spatial autoregressive and moving average hilb ertian processes. J. Multivariate Anal. 2010; 102:292–305

2010
[21]

Spatial functional prediction from spatial au- toregressive hilbertian processes

M.D., Ruiz-Medina MD. Spatial functional prediction from spatial au- toregressive hilbertian processes. Environmetrics 2012; 23:119– 128

2012
[22]

Kernel density estimation on random ﬁelds

Tran LT. Kernel density estimation on random ﬁelds. J. Multivar iate Anal. 1990; 34:37–53

1990
[23]

Polynomial spline estimation for partial f unc- tional linear regression models

Zhou J, Chen Z, Peng Q. Polynomial spline estimation for partial f unc- tional linear regression models. Comput. Statist. 2016; 31:1107–1 129. 18

2016

Pith tools

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

[1] [7]

Smoothing splines estimators for fu nc- tional linear regression

Crambes C, Kneip A, Sarda P. Smoothing splines estimators for fu nc- tional linear regression. Ann. Statist. 2009; 37: 35–72

work page 2009

[2] [1]

On the Prediction of Stationary Fu nc- tional Time Series

Aue A, Norinho DD, Hormann S. On the Prediction of Stationary Fu nc- tional Time Series. J. Amer. Statist. Assoc. 2015; 110:378–392

work page 2015

[3] [2]

Nonparamatric Spatial Prediction

Biau G, Cadre B. Nonparamatric Spatial Prediction. Stat. Infer ence Stoch. Process. 2004; 7:327–349. 16

work page 2004

[4] [3]

Optimal sampling for spatial pr e- diction of functional data

Bohorquez M, Giraldo R, Matheu J. Optimal sampling for spatial pr e- diction of functional data. Stat. Methods Appl. 2016; 25: 39–54

work page 2016

[5] [4]

Multivariate functional rando m ﬁelds: prediction and optimal sampling

Bohorquez M, Giraldo R, Matheu J. Multivariate functional rando m ﬁelds: prediction and optimal sampling. Stoch. Environ. Res. Risk As - sess. 2017; 31:53–70

work page 2017

[6] [5]

Bouka S, Dabo-Niang S, Nkiet GM. (2018). On estimation in spatial functional regression with derivatives. C. R. Acad. Sci. Paris Ser. I 2018; 356:558–562

work page 2018

[7] [6]

Adaptive functional linear regression

Comte F, Johannes J. Adaptive functional linear regression. An n. Statist. 2012; 40:2765–2797

work page 2012

[8] [8]

A partial overview of the theory of statistics with fun ctional data

Cuevas A. A partial overview of the theory of statistics with fun ctional data. J. Statist. Plan. Inf. 2014; 147:1-23

work page 2014