Adaptive estimation of Sobolev-type energy functionals on the sphere

TL;DR

We address the problem of estimating the quadratic Sobolev-type functional $T_r(f)=\left\|(-\Delta_{\mathbb S^d})^{r/2} f\right\|_{L^2(\mathbb S^d)}^2$ for densities on the unit sphere. The authors develop a spherical needlet framework that yields a natural multiscale decomposition of Sobolev energy and construct unbiased estimators for truncated versions of $T_r(f)$ using a sample-splitting strategy; they quantify the bias and variance and identify the minimax rate $n^{-4(s-r)/(2s+d+4r)}$ when $f\in H^s(\mathbb S^d)$ with $s>r$. To achieve adaptivity to the unknown smoothness, a Lepski-type procedure selects a global resolution level $\widehat{J}$, yielding an estimator that attains the same minimax rates uniformly over $s\in[r+\varepsilon, s_{\max}]$ without relying on nonlinear or sparsity-based methods. The work combines harmonic analysis on the sphere with multiscale needlet tools and rigorous bias--variance analysis to deliver sharp adaptive risk bounds, supported by numerical illustrations of the bias--variance tradeoff and resolution selection. These results have implications for global energy and smoothness assessment of spherical densities in directional data settings, with practical guarantees for adaptive estimation on manifolds.

Abstract

We study the estimation of quadratic Sobolev-type integral functionals of an unknown density on the unit sphere. The functional is defined through fractional powers of the Laplace--Beltrami operator and provides a global measure of smoothness and spectral energy. Our approach relies on spherical needlet frames, which yield a localized multiscale decomposition while preserving tight frame properties in the natural square-integrable function space on the sphere. We construct unbiased estimators of suitably truncated versions of the functional and derive sharp oracle risk bounds through an explicit bias--variance analysis. When the smoothness of the density is unknown, we propose a Lepski-type data-driven selection of the resolution level. The resulting adaptive estimator achieves minimax-optimal rates over Sobolev classes, without resorting to nonlinear or sparsity-based methods.

Figures (5)

• Figure 1: Spherical needlet $\psi_{j,k}$ at resolution level $j=3$ together with its first and second Sobolev derivatives $\psi_{j,k}^{(1)} = (-\Delta_{\mathbb S^2})^{1/2}\psi_{j,k}$ and $\psi_{j,k}^{(2)} = (-\Delta_{\mathbb S^2})\psi_{j,k}$, plotted as functions of the angular distance $\Theta$ from the center $\xi_{j,k}$. Differentiation increases oscillatory behavior while preserving spatial localization, illustrating the scale-dependent amplification induced by Sobolev derivatives.
• Figure 3: Numerical illustration of Theorem \ref{['thm:adaptive']}. Left: bias--variance tradeoff and oracle resolution level across Sobolev regimes (see Figure \ref{['fig:oracle_regimes']}). Right: oracle versus adaptive risk as a function of the sample size, for different Sobolev smoothness levels.
• Figure : Classical Sobolev regime ($d=2$, $r=1$, $s=2.5$, $n=4000$).
• Figure : Classical Sobolev regime ($d=2$, $r=1$, $s=2.5$, $n=4000$).
• Figure : Low-regularity boundary regime ($d=2$, $r=0$, $s=0.6$, $n=2\times10^5$).

Theorems & Definitions (18)

• Remark 2.1: Harmonic versus multiscale representations
• Lemma 2.2: Derivatives and needlet coefficients
• Remark 2.3: Multiscale structure and adaptive resolution selection
• Remark 2.4: Sobolev geometry versus frame tightness
• Remark 2.5: Sobolev and Besov representations via needlets, and on $r$-dependent constructions
• Remark 3.1: Model-independent nature of the truncation scheme
• Remark 3.2: Sample splitting and $U$-statistic constructions
• Proposition 3.3: Bias--variance decomposition for truncated Sobolev estimators
• Lemma 4.1: Stochastic fluctuation bound across resolutions
• Lemma 4.2: Control of over-smoothing events