Where Thresholding Error Goes in a Fourier Extension

Many sensor and compression pipelines save bandwidth by discarding every sample whose magnitude is below a small threshold $\delta$. It seems obvious that the reconstruction should improve as $\delta \to 0$. Boche and Mönich showed that it does not. For bandlimited signals on the whole real line, the error between sample points can blow up as $\delta$ shrinks, even though every individual sample is off by at most $\delta$ (Unboundedness of thresholding and quantization for bandlimited signals). Real signals, however, live on bounded intervals. On a bounded interval the preferred spectrally accurate tool is Fourier extension. This method embeds the interval $[a,b]$ in a larger periodic box $[-T,T]$ and fits a Fourier series whose values in the extra "buffer" region are left free. Its known stability rests on one fact: the badly conditioned directions of the fit live almost entirely in the buffer, so numerical noise gets dumped there (On the Numerical Stability of Fourier Extensions). Nobody has asked what happens to that mechanism when the samples have been thresholded.

This note proposes two things. The first is a quantitative answer: on a finite grid, thresholding instability does not vanish. Instead of diverging in $\delta$, it diverges with the resolution of the grid. The second is a fix. A transmitted zero is not a measurement of zero. It is a certificate that the true value lies in $[-\delta,\delta]$. Reconstructing with that interval constraint, instead of with the zero, should make the error independent of resolution. There is one sharp prediction here that a colleague can carry away. The two obvious linear treatments either fill discarded samples with zero or drop them as missing, and both fail on signals with quiet regions. Dropping them fails worst, because every run of dropped samples inside $[a,b]$ becomes an artificial buffer zone where Fourier extension deposits its exponentially ill-conditioned modes.

Core Claim

Let $\tau_\delta$ set samples with $|f(x_k)|<\delta$ to zero, and let $R$ be any linear reconstruction from $M$ samples on $[a,b]$ with cardinal functions $\ell_k$. Thresholding changes each sample by at most $\delta$, so

$$\|f - R\,\tau_\delta f\|_{L^\infty[a,b]} \le \|f - Rf\|_{L^\infty[a,b]} + \delta\,\Lambda_M, \qquad \Lambda_M = \sup_{t\in[a,b]} \sum_{k=1}^{M} |\ell_k(t)|,$$

where $\Lambda_M$ is the Lebesgue constant, the worst-case amplification from sample errors to pointwise error. This bound is trivial. The non-trivial parts are two.

The first concerns the lower bound. Thresholding errors are not arbitrary: they equal $f(x_k)$ on the discarded set, so they are correlated with the signal. We conjecture that, within natural smoothness classes, one can still construct signals whose small samples align in sign with $\ell_k(t)$, so that the $\delta\Lambda_M$ term is attained up to constants. For truncated-SVD Fourier extension, the Lebesgue constant $\Lambda_M$ grows with $M$. If both claims hold, refining the grid at fixed $\delta$ makes thresholded reconstructions worse. That would be a bounded-interval analogue of the Boche–Mönich divergence, with $\Lambda_M$ playing the role of their divergent shift-sum (Characterization of the pointwise and the peak value behavior of system approximation under thresholding).

The second part is the fix. We replace the linear map with a consistent reconstruction:

$$\min_{c}\ \|c\|_2 \quad \text{s.t.}\quad |(\Phi c)_k - y_k| \le \eta \ \text{(kept samples)},\qquad |(\Phi c)_k| \le \delta \ \text{(discarded samples)},$$

Here $\Phi$ is the Fourier-extension sampling matrix, $c$ the Fourier coefficients on $[-T,T]$, and $\eta$ the noise level on retained samples. The true signal is feasible for this problem. The error of the fix is therefore bounded by how large, in sup norm on $[a,b]$, a low-norm extension can become while its samples stay within about $2\delta$. That quantity is a discretization inequality for the extension space, not a Lebesgue constant. We conjecture it is bounded independently of $M$ in the interior of $[a,b]$ under linear oversampling $M=\gamma N$ (with $N$ the number of Fourier modes), and grows only near the endpoints.

The endpoint growth is where the root-exponential barrier enters. Endpoint derivative bounds of Markov type force effectively $N \sim \sqrt{M}$ near the edges. This is the same mechanism behind the $\exp(-c\sqrt{m})$ ceiling for stable recovery proved in A Stability Barrier for Reconstructions from Fourier Samples, and the $n=O(\sqrt{m})$ restriction for least squares in On the Stability and Accuracy of Least Squares Approximations. We expect thresholding to make that barrier visible at the scale of $\delta$ rather than of machine precision.

Research Program

The work has three strands.

Theory. We would compute and bound $\Lambda_M$ for continuous and equispaced truncated-SVD Fourier extension as a function of $M$, the oversampling $\gamma$, and the truncation level $\varepsilon$. Adcock et al. leave even the equispaced stability constants unproved, and only in $L^2$, so a sup-norm characterization is new and useful on its own. We would then attempt the adversarial lower bound for analytic and fractional-Sobolev classes. For the consistent reconstruction, we would prove an interior $O(\delta + \varepsilon\|c\|)$ bound and an endpoint bound whose $M$-dependence matches the root-exponential scaling.

Missing-data mechanism. We would analyze the "drop the zeros" variant separately. Our hypothesis is that a run of $L$ dropped samples inside $[a,b]$ acts as an interior buffer of width $L/M$. The singular vectors that On the Numerical Stability of Fourier Extensions shows are harmless in the outer buffer would then concentrate inside the physical domain, giving error exponential in the gap width.

Function-space scope. The lens from function-space generalization is to ask how the answer depends on the signal class. We would test analytic signals; fractional Sobolev $H^s$ signals with $s\in(0,1)$, the rough regime studied in Minimax Optimal Nonsmooth Nonparametric Regression via Fractional Laplacian Eigenmaps; and piecewise-smooth signals with shocks. In each class we would identify where $\delta\Lambda_M$ or the approximation error dominates. A final downstream study would ask whether a learned operator fed thresholded sensor data inherits the linear instability.

The scope is deliberately one-dimensional. This matches the file's benchmarks and avoids the unresolved problem of extending Fourier extension to complex geometries in higher dimensions.

Baselines

Naive reference. The do-nothing baseline is zero-filled linear Fourier extension: truncated-SVD Fourier extension with $T=2$ applied directly to $\tau_\delta f$, exactly as in On the Numerical Stability of Fourier Extensions. It is the method a practitioner would run today. Beating it means a lower sup-norm error on $[a,b]$ at every tested $(\delta,M)$ pair, and above all an error that stops growing with $M$.

Strongest existing method. This is the partitioned local Fourier extension of A Local Fourier Extension Method for Function Approximation. It splits $[a,b]$ into subintervals, reuses one truncated-SVD operator per piece, and on clean data reaches errors of $10^{-14}$ to $10^{-15}$ at linear cost. It is the strongest competitor because localization could plausibly confine thresholding damage to the few subintervals containing quiet regions. It also has a built-in detector, spikes in coefficient norm, which might flag trouble. No reported result was found on Lacuna for any Fourier extension method on thresholded samples.

Other linear baselines. We also run polynomial least squares with degree $n\sim\sqrt{m}$, the stable choice identified in A Stability Barrier for Reconstructions from Fourier Samples. We run cosine-transform reconstruction as well, the even-reflection basis used by Transform Once: Efficient Operator Learning in Frequency Domain.

Fairness. All methods get identical sample sets, the same mode or degree budget, and the same held-out grid of 20× resolution for measuring sup error. Truncation levels, subinterval counts, and degrees are each tuned by the same grid search on validation signals drawn from the same class. The consistent method must beat the tuned local Fourier extension by at least an order of magnitude in interior sup error at $\delta \ge 10^{-4}$ before we claim a win.

Ablations. Three ablations isolate the key ingredient. Consistency with a polynomial basis instead of Fourier extension separates the value of the interval constraint from the value of the extension frame. Fourier extension with discarded samples simply omitted tests the interior-buffer hypothesis. Consistency using the wrong interval, $[-\delta/10,\delta/10]$, checks that the gain comes from correct censoring information rather than from regularization alone.

Why not just…?

Why not just oversample? Boche and Mönich prove that oversampling does not cure thresholding instability for linear sampling-based reconstruction (No-Go theorem for sampling-based signal processing, Strong Divergence for System Approximations). If our first claim holds, oversampling at fixed $\delta$ is actively harmful, since error grows with $M$. We test this directly by sweeping $\gamma$ at fixed $\delta$ for the linear baselines. Their theorems do not cover consistent reconstruction, which is nonlinear. That is exactly where we expect oversampling to start helping.

Why not just treat discarded samples as missing? This is the most natural simple fix, and we take it seriously. Our hypothesis is that it performs worst on signals with quiet stretches, because the gaps become interior buffer zones. For signals whose discarded samples are isolated points near sign changes, dropping them may work about as well as consistency. The experiment separates these regimes by varying the fraction of $[a,b]$ where $|f|<\delta$. If dropping matches consistency in both regimes, the interval constraint adds nothing and we would say so.

Why not just let a neural operator learn to undo thresholding? This is a real threat downstream. A Fourier neural operator trained on zero-filled inputs, especially with a channel marking which samples were discarded, might learn to fill them in. On clean inviscid Burgers, Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities reports a median relative $L^1$ error of 1.57% for the Fourier neural operator, against 28.5% for DeepONet. That is a strong learner. We test it head to head. If the masked operator matches consistent Fourier extension at inference, our contribution narrows to the classical, training-free setting and to guarantees rather than accuracy. The sup-norm and adversarial evaluations are where we expect learned imputation to lose, since training minimizes average error.

Why not just use the cosine transform? Even reflection creates a derivative jump at the endpoints, which limits convergence to algebraic rates, as noted in the analysis of Transform Once: Efficient Operator Learning in Frequency Domain. It is therefore a floor baseline rather than a competitor for smooth non-periodic signals.

Proposed Experiments

Synthetic study. The core experiment uses three signal families on $[0,1]$. The first is analytic non-periodic functions, including Runge-type and oscillatory profiles of the kind in the local Fourier extension test suite. The second is random Fourier series with coefficient decay set to target $H^s$, $s\in\{0.3,0.6,0.9\}$. The third is compact bumps and pulses with long near-zero stretches. We sweep $\delta\in[10^{-8},10^{-1}]$ relative to the signal peak and $M\in[64,8192]$ at $\gamma\in\{1,2,4\}$. We report three things: sup error in the interior and within $10/M$ of each endpoint, the measured $\Lambda_M$, and worst-case error found by gradient ascent over signals in the unit ball of each class.

The decisive check has two parts. At fixed $\delta$, the linear methods' worst-case error should grow with $M$, tracking $\delta\Lambda_M$. The consistent method's interior error should stay flat. A second check is that the drop-the-zeros variant should show error growing exponentially with gap width on the bump family. If $\Lambda_M$ turns out bounded, or the adversarial search cannot approach $\delta\Lambda_M$, the first claim is wrong and the paper becomes a stability result for Fourier extension under thresholding. That outcome would also be worth reporting.

Downstream study. We use the one-dimensional inviscid Burgers benchmark from Fourier Neural Operator for Parametric Partial Differential Equations and Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities. Initial conditions are thresholded at $\delta\in\{10^{-3},10^{-2},10^{-1}\}$ times their maximum, then passed through one of three front ends before a Fourier neural operator trained with a fixed budget. The front ends are zero-filling with a mask channel, linear Fourier extension, and consistent Fourier extension. The reference number is the 1.57% clean median relative $L^1$ error. No reported result was found on Lacuna for thresholded inputs, so we establish that curve ourselves.

The decisive check is whether the consistent front end keeps the error within a small factor of the clean value as $\delta$ grows while the other two degrade. We also check whether any advantage survives when the evaluation resolution differs from the training resolution. That is the setting where a resolution-dependent instability should show up, and where a learned imputer has had no chance to adapt.