PaperlayerView exact revision on arXiv

Source-linked AI summary

MiNCE: Nonparametric, Strongly Consistent Confidence Envelopes for Band-Limited Functions and their Smoothed Spectra

Balázs Csanád Csáji, Bálint Horváth

arXiv:2609.09436v1math.STcs.LGeess.SPstat.MEstat.ML

TL;DR

Existing MiNCE envelopes provide nonasymptotic coverage, but their consistency had not been established. This paper proves strong uniform consistency for noise-free and mildly noisy settings, extends the framework to smoothed spectra, and validates contraction numerically.

  • Problem

    The paper addresses the missing consistency analysis of MiNCE confidence envelopes that already had finite-sample coverage guarantees.

  • Method

    The paper analyzes MiNCE bands in spatial and frequency domains using RKHS minimum-norm constructions, including noisy-output confidence ellipsoids and smoothed spectral tubes.

  • Results

    MiNCE bands converge almost surely and uniformly to the true regression function in noise-free and mildly noisy settings, while spectral tubes inherit coverage and strong uniform consistency.

  • Takeaways & Limitations

    MiNCE provides a mathematically supported basis for simultaneous nonparametric inference on band-limited functions and smoothed spectra.

  • Takeaways & Limitations

    The analysis assumes that the input sampling distribution is known and does not provide convergence rates.

Abstract

from arXiv · show

Minimum-norm confidence envelope strategies offer a nonparametric approach to constructing nonasymptotic, simultaneous confidence regions for band-limited functions, exploiting the theory of Reproducing Kernel Hilbert Spaces (RKHS). While the finite-sample coverage guarantees of these envelopes have been established, their consistency has not been analyzed so far. In this paper, we study this construction, here termed the Minimum-Norm Confidence Envelope (MiNCE) framework, and establish the strong uniform consistency of the resulting bands, both for noise-free and noisy observation models, under mild assumptions on the measurement noises. We further extend this formulation to the frequency domain, deriving nonasymptotic, simultaneous, strongly uniformly consistent confidence bands for the smoothed spectra. Numerical experiments in nonparametric regression and spectral estimation empirically confirm our theoretical results, illustrating the contraction of the confidence envelopes toward the target function as the sample size increases.

I. INTRODUCTION

The paper develops MiNCE, a nonparametric framework for simultaneous, nonasymptotic confidence bands for band-limited functions, addressing the missing consistency analysis of previously covered envelopes. It also connects RKHS minimum-norm interpolation with confidence-band construction and extends the framework toward spectral applications.

  • Motivation: Regression estimates conditional-mean functions from finite input-output samples, whereas standard methods typically select a point estimate from a hypothesis class.The paper motivates region estimation as a complement to point estimation.
  • MiNCE framework: MiNCE constructs simultaneous, nonparametric confidence bands with nonasymptotic coverage for band-limited functions.The target class is foundational in signal reconstruction, digital communications, remote sensing, and frequency-division multiplexing.
  • MiNCE framework: The framework addresses a gap: prior MiNCE envelopes had finite-sample coverage guarantees, but their consistency had not been established.The paper studies strong uniform consistency for both abstract regions and induced bands.
  • Results: The paper reports strong uniform consistency for MiNCE regions and bands in noise-free and mildly noisy settings, and numerical contraction toward the target as sample size increases.The reported experiments cover nonparametric regression and spectral estimation.
  • RKHS construction: RKHSs provide the function-space setting, with kernels reproducing point evaluations and minimum-norm interpolants supplying the central construction object.For interpolation data, the minimum-norm solution is represented through the kernel matrix; its norm measures smoothness and, in Paley–Wiener spaces, signal energy.

C. Main Assumptions

The framework assumes i.i.d. data, known absolutely continuous input sampling with positive density, and a Paley–Wiener target satisfying an informativeness bound. MiNCE combines finite-sample confidence information, kernel-norm bounds, and minimum-norm interpolation to form query-wise confidence intervals.

  • Main assumptions: The data are i.i.d. input-output pairs, with finite-variance outputs and measurement noise in the regression model.The regression function is f∗(t) = E[Y | X = t].
  • Main assumptions: The input distribution is known, absolutely continuous, and has density h∗(x) > 0 everywhere, enabling information about the whole regression function.The known-distribution condition simplifies the theoretical analysis.
  • Main assumptions: The target f∗ belongs to a Paley–Wiener space and satisfies f∗(x)^2 ≤ ϱ h∗(x), linking function magnitude to sampling density.The bound supports generalization from random observations and constrains compatible input-distribution tails.
  • MiNCE construction: MiNCE first obtains a confidence set for noiseless outputs at observed inputs, then bounds the target’s kernel norm before constructing query-point intervals through minimum-norm interpolation.In the noise-free case, the preliminary confidence-set step is unnecessary because the noiseless outputs are directly available.
  • MiNCE construction: The norm bound can be obtained using concentration inequalities, while the abstract confidence region is converted into a practical band by computing envelopes of potential outputs.The construction seeks user-chosen finite-sample reliability and uniform shrinkage with increasing sample size.
  • MiNCE construction: The ideal abstract region has coverage at least 1 − α, but constructing it directly is hard in practice, motivating an efficient confidence-band algorithm.This practical difficulty is explicitly identified in the construction discussion.

B. Confidence Bands and Interval Endpoints

MiNCE constructs pointwise confidence intervals by finding feasible outputs under a minimum RKHS-norm bound, then envelopes these intervals into a confidence band. Under the stated assumptions, the resulting bands are strongly uniformly consistent as the sample size grows.

  • Confidence-band construction: The confidence band Bn is obtained by computing an envelope for the potential outputs allowed by the abstract confidence region Cn.The construction satisfies Cn ⊆ Bn, transferring coverage from the function-space region to the band.
  • Interval endpoints: At a query input x0, the interval endpoints are the minimum and maximum y0 values interpolable by an RKHS function with squared norm at most κn.These extrema arise from two convex optimization problems with quadratic constraints.
  • Interval endpoints: The resulting intervals are centered around the minimum-norm interpolant fn̂ and collapse to the observed output when x0 equals a sampled input.The endpoints can be computed analytically using the kernel matrix and, efficiently, Schur complements.
  • Strong uniform consistency: Under A1–A3 and noise-free observations, the norm bounds are consistent, with the excess bound converging to zero as n →∞.The proof uses the minimum-norm interpolant and its projection properties in the RKHS.
  • Strong uniform consistency: Theorem 3 and its corollaries establish strong uniform consistency of the MiNCE bands and a consistent norm-ball outer approximation.The consistency argument first controls abstract confidence regions in RKHS norm and then transfers that control to confidence bands uniformly over inputs.

V. OUTPUTS WITH MEASUREMENT NOISE

The noisy observation model assumes outputs y_k=f*(x_k)+ε_k while retaining a known input distribution, and the paper aims to show MiNCE remains strongly uniformly consistent.

  • The noisy model observes y_k=f*(x_k)+ε_k instead of the regression function values directly.
  • The input distribution is assumed to be known a priori under measurement noise.
  • The stated goal is to establish strong uniform consistency of MiNCE under these noisy observations.

A. Norm Estimation from Noisy Outputs

Because noisy outputs hide the true sampled function values, the method constructs confidence ellipsoids for those values and uses them to extend the norm-bound construction under explicit noise assumptions.

  • A. Norm Estimation from Noisy Outputs: Noisy observations prevent direct application of the norm-bound construction because the values f*(x_k) are unobserved.
  • A. Norm Estimation from Noisy Outputs: The method constructs a confidence ellipsoid for the true outputs at a subset of sample inputs.
  • A. Norm Estimation from Noisy Outputs: The ellipsoid is represented by a center vector and shape matrix, with its radius normalized to 1 by rescaling the shape matrix.
  • A. Norm Estimation from Noisy Outputs: One supported assumption set requires zero-mean noises independent of inputs and distributional invariance under a known compact matrix group.
  • A. Norm Estimation from Noisy Outputs: Exchangeable noises permit permutation transformations, while independent noises symmetric about zero permit sign-change transformations.
  • A. Norm Estimation from Noisy Outputs: The resulting bound optimizes over all output vectors in the ellipsoid, but the capped problem is nonconvex and only an upper bound on its optimum is needed.
  • A. Norm Estimation from Noisy Outputs: Lemma 5 combines the ellipsoid uncertainty and norm-estimation uncertainty through risk probabilities β and α, respectively.

B. Confidence Bands and Interval Endpoints

The noisy confidence band is formed by optimizing interval endpoints over ellipsoid-compatible outputs, while preserving simultaneous coverage of the target function under the stated assumptions.

  • B. Confidence Bands and Interval Endpoints: The abstract confidence set is converted into interval endpoints by solving minimization and maximization problems whose optimal values define the lower and upper bounds.
  • B. Confidence Bands and Interval Endpoints: Under the assumptions, the construction satisfies P(f*∈D_n) ≥ 1−α−β because the true sampled outputs lie in the ellipsoid with that probability.
  • B. Confidence Bands and Interval Endpoints: Although the direct confidence-set construction is computationally challenging, an algorithm is needed to generate the set efficiently for arbitrary inputs.
  • B. Confidence Bands and Interval Endpoints: For a query input, the noisy construction optimizes over all possible sampled-output vectors contained in the confidence ellipsoid.
  • B. Confidence Bands and Interval Endpoints: The extended Gram matrix includes the query input and has dimension (n0+1) × (n0+1).
  • B. Confidence Bands and Interval Endpoints: If the optimization problem is infeasible, such as for an empty KGP ellipsoid, the construction returns an empty confidence band.
  • B. Confidence Bands and Interval Endpoints: The resulting band guarantees z_min(x0) ≤ f*(x0) ≤ z_max(x0) simultaneously for all input points.

C. Strong Uniform Consistency

The MiNCE construction is shown to be strongly uniformly consistent under measurement noise, provided the stated assumptions hold. The proof proceeds through shrinking confidence ellipsoids, consistent norm bounds, and convergence of auxiliary confidence regions to the true regression function.

  • Proof strategy: The noisy-case analysis uses a larger confidence-band construction whose consistency implies consistency of MiNCE.This strategy simplifies the proof while preserving the target consistency result.
  • Assumptions: The confidence ellipsoids are assumed to shrink as the sample size increases, with their centers strongly consistent in the RKHS norm.Euclidean convergence alone is insufficient because the smallest eigenvalue of the kernel matrix may vanish.
  • Norm control: The norm bound τn is strongly consistent under assumptions A1–A5, and admissible upper bounds may widen the bands when chosen larger.The capped bound τ = min{τ°0, τn} is available as an implementable upper bound.
  • Consistency proof: Technical lemmas establish convergence of the auxiliary confidence sets, including convergence of minimum-norm interpolant norms to the regression-function norm.These results provide the ingredients for the abstract MiNCE consistency theorem.
  • Main result: Under A1–A5, the abstract MiNCE regions and the induced confidence bands are strongly uniformly consistent almost surely as n increases.The result applies to the noisy observation setting and follows by combining the abstract-region theorem with the band construction.

VI. ROBUST SPECTRAL ESTIMATION

The paper extends MiNCE to frequency-domain estimation by applying a smoothing operator to the minimum-norm interpolant. This yields nonasymptotic confidence regions and strongly uniformly consistent bands for smoothed spectra, including magnitude and phase.

  • Motivation: Direct frequency-domain MiNCE is unavailable because the spectrum is not directly observed and L2([−η, η]d) is not an RKHS.The paper instead constructs confidence regions through smoothed Fourier transforms of spatial-domain functions.
  • Construction: The smoothed spectrum is obtained by convolving the Fourier transform with a user-chosen nonnegative L1-normalized window function.Rectangular and triangular smoothers are discussed as practical choices.
  • Consistency: The smoothed spectrum of the minimum-norm interpolant is strongly uniformly consistent for the target smoothed spectrum under the stated assumptions.The result follows from convergence of the interpolant and Plancherel’s theorem.
  • Confidence regions: MiNCE confidence regions for the smoothed spectrum are formed as images of spatial-domain regions under the smoothing operator, with radius bounds inherited from the original construction.The L2 and pointwise deviations are bounded using Young’s and Cauchy–Schwarz inequalities.
  • Magnitude and phase: The resulting bands provide nonasymptotic coverage and strong uniform consistency for smoothed magnitude and phase spectra.Magnitude and phase intervals are determined from extrema over the frequency-domain confidence ball.

VII. NUMERICAL EXPERIMENTS

Numerical experiments in spatial-domain regression and spectral estimation support the theoretical contraction of MiNCE confidence envelopes as sample size increases. The experiments cover noise-free and nontrivial noisy observations, as well as smoothed magnitude and phase spectra.

  • Experimental setup: The experiments use one-dimensional Paley–Wiener RKHS models with Student t inputs and both noise-free and Laplacian-mixture measurement-noise settings.The noise-free and noisy demonstrations use n = 40 and n = 400 observations, respectively.
  • Noisy regression: The noisy experiments use KGP confidence ellipsoids built from n0 = 20 points and a capped norm bound.The permutation group is used because the noise mixture is exchangeable but not symmetric about zero.
  • Norm bounds: 4.79, 3.20 and 2.92 are the median capped norm bounds for n = 100, 250 and 500, respectively.The corresponding uncapped medians are 13.02, 19.75 and 51.02, illustrating the practical need for capping.
  • Spatial-domain regression: The mean maximum errors and their standard deviations shrink by more than an order of magnitude as the sample size increases.This experiment directly illustrates strong consistency of the spatial-domain MiNCE bands.
  • Spectral estimation: Smoothed magnitude-spectrum bands are easier to estimate than phase-spectrum bands, while both are harder than spatial-domain signal estimation.The spectral experiment uses n = 10 000 noise-free observations and a triangular smoother with h = 1.
  • Spectral consistency: Average maximum errors and variances of the smoothed-spectrum intervals decrease with sample size, although simultaneous estimation over the full spectrum requires a large sample.The frequency-domain evaluation reports relative maximum errors over randomly selected frequencies.

VIII. CONCLUSIONS

The paper establishes strong uniform consistency for MiNCE confidence bands in spatial and frequency domains, with experiments showing contraction toward the target as sample size grows. The framework still assumes known input sampling distributions, leaving relaxation of this assumption and convergence rates for future work.

  • VIII. CONCLUSIONS: MiNCE bands converge almost surely and uniformly over the domain to the true regression function as the sample size grows.This holds for both noise-free observations and noisy observations under a mild confidence-ellipsoid condition.
  • VIII. CONCLUSIONS: The framework assumes a priori knowledge of the sampling distribution of the inputs.Relaxing this assumption and adding convergence rates are identified as future directions.
  • VIII. CONCLUSIONS: The frequency-domain extension provides simultaneous confidence tubes for smoothed spectra, inducing bands for magnitude and phase spectra.These regions inherit nonasymptotic coverage and strong uniform consistency from the spatial-domain construction.
  • VIII. CONCLUSIONS: Experiments in nonparametric regression and spectral estimation show systematic contraction of the confidence regions as n grows.The supplied table and figure captions identify spatial-domain errors and frequency-domain magnitude-spectrum errors as evaluated quantities.

APPENDIX A PROOF OF LEMMA 1

The proof establishes convergence properties by controlling kernel-based quantities with concentration inequalities, reproducing-kernel bounds, and strong-law arguments. It concludes that the relevant asymptotic error term vanishes almost surely.

  • APPENDIX A PROOF OF LEMMA 1: Hoeffding’s inequality is used to select a threshold t satisfying the required probability bound.The proof transforms the confidence requirement into exp(−2nt^2/ϱ^2) ≤ α and solves it logarithmically.
  • APPENDIX A PROOF OF LEMMA 1: The RKHS reproducing property and Cauchy–Schwarz inequality provide pointwise control of functions and their confidence-band deviations.The argument applies these properties to the bounded kernel and the constructed quantity L_n(x).
  • APPENDIX A PROOF OF LEMMA 1: The minimum-norm interpolant is uniquely identified through the Gram-matrix system Kα = y.For an almost surely invertible Gram matrix, α = K^−1y gives the unique interpolating function in the finite-dimensional kernel span.
  • APPENDIX A PROOF OF LEMMA 1: The proof shows that the relevant norm-bound terms converge almost surely by combining bounded differences, truncation, and the strong law of large numbers.Splitting observations according to h∗(xk) ≥ t or h∗(xk) < t makes the residual contribution vanish as t decreases.

APPENDIX E PROOF OF LEMMA 8

The proof analyzes the minimum-norm quadratic form over confidence ellipsoids by separating the noiseless component from the ellipsoid perturbation. Both components converge almost surely to the target kernel norm, forcing the infimum and supremum quantities to converge as well.

  • APPENDIX E PROOF OF LEMMA 8: The finite-dimensional kernel span is closed, allowing minimum-norm interpolants and orthogonal decompositions to be used within H_n0.The proof compares interpolants associated with z and z′ through this subspace structure.
  • APPENDIX E PROOF OF LEMMA 8: The noiseless quadratic-form term converges almost surely to κ∗, the kernel norm of the minimum-norm interpolant of the noiseless outputs.This convergence is invoked from the corresponding minimum-norm interpolation result.
  • APPENDIX E PROOF OF LEMMA 8: The confidence-ellipsoid contribution is controlled by expanding z = z∗ + (z − z∗), then applying the triangle and Cauchy–Schwarz inequalities.The resulting cross-term and perturbation terms are bounded through the ellipsoid’s quadratic form.
  • APPENDIX E PROOF OF LEMMA 8: Both arguments inside the maximum converge almost surely to κ∗, so the overall supremum also converges almost surely to κ∗.The proof then sandwiches the infimum between two terms with the same limit.
Loading 2609.09436v1…