A variational mollification approach to circular deconvolution

1 Introduction

• Because measurements may not support reliable reconstruction across all frequencies, a stable approximation at data-compatible resolution can be more natural than exact recovery.The paper notes that only a limited frequency range can generally be reconstructed reliably.Source: §1 · p. 1
• Earlier work connects resolution-limited reconstruction to stabilized Fourier synthesis, approximate inverses, and variational mollification for truncated Fourier operators and probability-density deconvolution.The cited density-deconvolution work established deterministic rates under Besov-Nikolskii source conditions.Source: §1 · p. 2
• The paper derives a variational reconstruction formula for a prescribed mollified target and provides a deterministic convergence analysis.The target retains a direct resolution-based interpretation.Source: §1 · pp. 2–3
• Ordinary smooth kernels yield order-optimal algebraic rates, while supersmooth kernels yield classical logarithmic rates under the stated smoothness assumptions.The approach retains a direct interpretation of the reconstructed object.Source: §1 · p. 3
• The paper advocates choosing the interpretable target before constructing the inversion algorithm, then analyzing its consistency, stability, and convergence rates.It presents this as a bridge between finite physical resolution and regularization theory.Source: §1 · p. 3

2 Convolution on the unit circle

• The paper represents the unit circle as T = R/2πZ and uses normalized Haar measure for integration on T.It introduces Lp(T) for measurable functions satisfying the stated integrability condition.Source: §2 · p. 3
• The section establishes the L2 Fourier framework through the orthonormal basis and Parseval identity.It also states that Fourier coefficients of an L2 function belong to the corresponding square-summable sequence space.Source: §2 · p. 4
• For p = 2, convolution by γ maps L2(T) into L2(T), and its Fourier coefficients are products of the kernel and function coefficients.The convolution operator is diagonalized by the Fourier transform.Source: §2 · pp. 4–5

3 Deconvolution on the unit circle

• The reduced minimum modulus of T is introduced as a way to characterize the ill-posedness of deconvolution.The operator is diagonal in the Fourier basis, with Fourier multiplier given by the kernel coefficients.Source: §3 · pp. 5–6
• If the kernel’s Fourier coefficients have infinite support and tend to zero along that support, the inverse problem is ill-posed.The range of T is then not closed and the inverse on ran T is not continuous.Source: §3 · p. 6
• For an L1(T) kernel with infinitely supported Fourier coefficients, the Riemann-Lebesgue lemma supplies the coefficient decay that meets this ill-posedness condition.The paper identifies this property as motivation for variational mollification.Source: §3 · p. 6

3.1 Mollification

• An approximate identity is a family of functions whose convolution with f approaches f in Lp as β decreases to zero.Fejér kernels are given as a familiar example.Source: §3.1 · p. 6
• The prescribed target at resolution β is convolution by φβ, represented by the operator Cβ.The completion condition is stated for T and Qβ := I − Cβ.Source: §3.1 · p. 6
• When the Morozov Completion Condition holds, the associated operator is boundedly invertible and the variational functional is defined.The text identifies the functional’s first term as approximation error and its second as data-perturbation propagation.Source: §3.1 · p. 6
• The solution depends continuously on the data and has an explicit representation in Fourier coefficients.The section presents this formula for stable reconstruction of the prescribed mollified target.Source: §3.1 · p. 7
• If bγ(0) is nonzero and φβ is nonnegative and even, the Morozov Completion Condition is satisfied.This is the stated condition for Proposition 2.Source: §3.1 · p. 7

• The section’s main result is consistency of the proposed reconstruction.Source: §3.2 · pp. 8–9
• The mollifier’s Fourier coefficient at each frequency is assumed to tend to 1 as β tends to 0.Source: §3.2 · p. 8
• The proof uses dominated convergence, with uniform L1 boundedness of the mollifiers providing a summable dominating sequence.Source: §3.2 · p. 9
• Only frequencies where the convolution operator has nonzero Fourier coefficient contribute to the error, because the exact solution lies in (ker T)⊥.Source: §3.2 · p. 9

3.3.1 Bias-noise decomposition

• The reconstruction error splits into approximation bias and propagated data perturbation; Fourier diagonalization reduces analysis to the bias and a kernel-dependent multiplier norm.Source: §3.3.1 · pp. 10–11

3.3.2 Besov-Nikolskii smoothness

• The bias analysis assumes the exact solution belongs to a periodic Besov-Nikolskii space, characterized through its Fourier-tail behavior.Source: §3.3 · p. 10, §3.3.2 · p. 11

3.3.3 Assumptions on the kernel

• The analysis distinguishes ordinary-smooth kernels with polynomial Fourier decay from supersmooth kernels with exponential decay; the latter yield severely ill-posed problems.Source: §3.3.3 · p. 11

3.3.4 Ordinary smooth case

• For ordinary-smooth kernels, the rate is minimax order-optimal under polynomial Fourier decay, but smoothness beyond u=b+d cannot improve it unless mollifier order rises.Source: §3.3.4 · p. 15

3.3.5 Supersmooth case

• For supersmooth kernels, the logarithmic rate is minimax order-optimal and has no saturation; exponential Fourier decay dictates it.Source: §3.3.5 · p. 18

4.1 Problem Formulation and Data Simulation

• The simulation uses 512 observations of a two-mode Wrapped Gaussian wind profile, blurred by a periodic kernel and corrupted by approximately 20% additive white Gaussian noise.The profile represents Westerly and North-Easterly wind components; at this noise level, the secondary peak is indistinguishable in the raw data.Source: §4.1 · pp. 18–19

4.2 Parameter Selection via the L-curve Method

• Because the noise level is unknown, the L-curve selects β at the trade-off curve’s maximum-curvature corner, balancing data fidelity and regularization.The experiment evaluates β over [10^-5, 10^-1]; the selected value lies near the error minimum without requiring ground truth.Source: §4.2 · p. 19, Figure 1 · p. 19, §4.3 · p. 20

4.3 Stability Analysis

• The selected β lies close to the relative-error minimum and reduces the condition number from approximately 1022 to 1020, though it remains large.The condition number remains large, as expected for this severely ill-posed problem.Source: Figure 2 · p. 20, §4.3 · p. 20

4.4 Reconstruction Results

• The selected β is approximately 3.4 × 10^-3; non-negativity projection and renormalization make the final estimate a valid discrete probability distribution.The convergence results apply to unconstrained fβ; projection is applied only to ensure interpretability of the final estimate.Source: §4.4 · p. 20
• At δ ≈20%, reconstruction recovers the bimodal wind profile and the obscured secondary North-Easterly mode despite noise and blur.The observed sensor data merges the two modes, while the reconstruction using the optimal β restores the bimodal structure.Source: §4.4 · pp. 20–21, Figure 3 · p. 21