When a sensor keeps only the coefficients it considers large enough and sets every value below a threshold $\delta$ to zero, the natural expectation is that a smaller $\delta$ gives a better reconstruction. Boche and Mönich proved that for bandlimited signals this can fail badly: on the sampling grid the error is at most $\delta$, yet between grid points the reconstruction can blow up as $\delta \to 0$. Separately, Adcock, Hansen and Shadrin proved that recovering a smooth non-periodic function from its first $m$ Fourier coefficients cannot converge faster than roughly $\exp(-c\sqrt{m})$ unless the method becomes exponentially sensitive to perturbations. The two results have never been put together, although the second setting is a finite, practical version of the first. The Fourier coefficients of a non-periodic analytic function decay only like $1/k$, the same slow, conditionally convergent regime that drives the thresholding paradox.
This note proposes a single, computable quantity that governs how every linear Fourier-sample reconstruction method responds to thresholding. We call it the thresholding Lebesgue constant: the worst-case uniform error produced by coefficient perturbations that are each bounded by one. We will prove that it inherits the root-exponential stability barrier, measure it for the methods that currently attain that barrier, and test whether choosing method parameters by this constant, rather than by the usual least-squares condition number, gives a better accuracy for the same number of retained coefficients. Arithmetic Fourier analysis, through Ramanujan sums, enters in a supporting role. It supplies exactly computable, integer-coefficient test signals that let us separate thresholding error from round-off error, and a cyclic-group model in which the constant can be evaluated in closed form.
Core Claim
Write a linear reconstruction method as $R c = \sum_{|k| \le m} c_k\, r_k(x)$, where $c_k$ are the observed Fourier coefficients of $f$ on $[-1,1]$ and $r_k$ are the method's reconstruction functions. Examples are polynomial least squares of degree $n$, or a Fourier extension solved with truncated singular value decomposition (TSVD) at a fixed cutoff. Thresholding replaces $c$ by $\Theta_\delta c$, and every entry of the difference is bounded by $\delta$. Hence
$$\|f - R(\Theta_\delta c)\|_{\infty} \;\le\; \|f - R c\|_{\infty} \;+\; \delta\,\Lambda_m, \qquad \Lambda_m = \sup_{x \in [-1,1]} \sum_{|k|\le m} |r_k(x)|.$$
$\Lambda_m$ is exactly the finite-$m$ counterpart of the shift-sum $\sum_k |h_T(t-k)|$. Boche and Mönich showed that this sum characterizes pointwise divergence under thresholding at $m = \infty$ (Characterization of the pointwise and the peak value behavior of system approximation under thresholding). $\Lambda_m$ also sits between two known quantities. Because $\|c\|_\infty \le \|c\|_2$ and $\|g\|_{L^2[-1,1]} \le \sqrt{2}\|g\|_\infty$, we get $\Lambda_m \ge \kappa_m/\sqrt{2}$, where $\kappa_m$ is the least-squares condition number used in A Stability Barrier for Reconstructions from Fourier Samples (up to that paper's normalization). The barrier therefore transfers immediately: any method converging faster than $o(m^{-\sigma})$ has $\Lambda_m \gtrsim m^{\sigma-1}$, and geometric convergence forces exponential $\Lambda_m$.
The open question, and the substance of the project, is the upper side. For methods that sit exactly at the root-exponential limit, does $\Lambda_m$ track $\kappa_m$? Or does it pick up an extra growth factor, anywhere up to $\sqrt{m}$ times a mixed norm, that the least-squares analysis cannot see? If the extra factor is real, then least-squares stability is the wrong criterion for thresholded or quantized front-ends. The optimal joint schedule of $m$, $\delta$ and method parameters would then differ measurably from the one the barrier suggests.
Research Program
The theoretical work has three parts. First, we will state and prove the inherited barrier for $\Lambda_m$ in the form above. Second, we will derive upper bounds on $\Lambda_m$ for polynomial least squares with $n \asymp \sqrt{m}$. Its reconstruction functions are projections of Legendre polynomials onto Fourier modes, so $\Lambda_m$ reduces to the uniform norm of an explicit kernel, which we expect behaves like a Lebesgue function with at most polynomial growth. We will also analyze Fourier extension with TSVD, where the cutoff plays the role of a regularizer and the oversampling ratio $m/n$ is a free parameter. Third, we will study attainability. The bound $\delta\Lambda_m$ is reached only when the coefficients that get zeroed carry the signs of $r_k(x^*)$ at the worst point $x^*$. Boche and Mönich's constructions show that this alignment is achievable for $PW_\pi^1$ as $\delta \to 0$ (Unboundedness of thresholding and quantization for bandlimited signals). Here $PW_\pi^1$ is the space of bandlimited signals whose spectrum is absolutely integrable. We will build analogous analytic functions on $[-1,1]$ whose sub-threshold Fourier tail aligns with a chosen kernel sign pattern. This bears on the open question of how fast thresholding errors diverge, since at finite $m$ the blow-up is capped by $\delta\Lambda_m$, and the cap's growth in $m$ becomes the speed.
The oversampling knob connects to a problem Boche and Mönich leave explicitly open: which systems remain stable under thresholding when oversampling is used (Characterization of the pointwise and the peak value behavior of system approximation under thresholding). Their infinite-dimensional negative results say oversampling does not rescue the worst case in $PW_\pi^1$ (Strong Divergence for System Approximations). The finite-$m$ question is different and answerable. At fixed $m$, the ratio $m/n$ is a finite-dimensional window into that problem, even though it cannot settle the infinite-dimensional case. We will ask how $\Lambda_m$ depends on $m/n$ and on the extension length, and whether that dependence differs from how $\kappa_m$ depends on them.
The arithmetic component is deliberately small. Ramanujan sums $c_q(k)$, the sums of $k$-th powers of primitive $q$-th roots of unity, are integers whose discrete Fourier transform on the cyclic group $\mathbb{Z}_q$ is $q$ times the indicator of residues coprime to $q$. This is the basis Bibak and coauthors use to count solutions of restricted congruences (On a restricted linear congruence). Weighted combinations of these sums, together with their extensions to $[-1,1]$, give test signals with exactly known rational coefficients, spread over many magnitudes, and placed on a divisor structure. We can threshold them and reconstruct in exact rational arithmetic, which isolates the $\delta\Lambda_m$ term from floating-point effects. That separation matters because double-precision Fourier extensions are known to behave better than their asymptotic conditioning predicts (A Local Fourier Extension Method for Function Approximation). On the cyclic group itself, periodic interpolation to off-grid points has a kernel whose $\ell^1$ norm can be written through Ramanujan expansions, giving a closed-form reference for the periodic limit. We do not claim that arithmetic structure suppresses the instability. The no-go theorem of No-Go theorem for sampling-based signal processing shows that re-arranging sample geometry cannot remove unbounded operator norms, and a divisor-structured index set is such a rearrangement.
Baselines
The do-nothing reference is the same reconstruction method applied to unthresholded coefficients. Its error curve $\|f - Rc\|_\infty$ versus $m$ shows how much of the observed error thresholding adds. All methods receive identical coefficient sets, identical $m$, and identical tuning grids.
The strongest competitor on stability is polynomial least squares with degree $n \asymp \sqrt{m}$, which Adcock, Hansen and Shadrin prove attains the optimal stable rate $\exp(-c\sqrt{m})$ (A Stability Barrier for Reconstructions from Fourier Samples). That paper reports no empirical benchmark numbers on Lacuna, so we compare against the analytic rate. The strongest competitor on practical accuracy is local Fourier extension with TSVD. Zhao and Wang report uniform errors of "$10^{-14}$ to $10^{-15}$, effectively the limit of double-precision arithmetic" on synthetic smooth and singular functions (A Local Fourier Extension Method for Function Approximation), and Li et al. report similar figures for piecewise smooth functions (Reconstruction of Piecewise Smooth Functions Based on Fourier Extension). No reported result was found on Lacuna for any of these methods under coefficient thresholding. Each will be run with its parameters (degree, extension length, TSVD cutoff, subinterval count) tuned over the same grid.
The model-based baseline treats thresholding as additive noise. It predicts an error of roughly $\kappa_m$ times the $\ell^2$ norm of a random perturbation with entries of size $\delta$, and it selects parameters by minimizing $\kappa_m$ plus approximation error. The key ablation holds each method fixed and changes only the selection rule, choosing parameters to minimize $E_m + \delta\Lambda_m$ instead of $E_m + \delta\kappa_m$. The proposal succeeds if three things hold. The $\Lambda_m$-guided rule must give lower worst-case and median uniform error at a matched number of retained coefficients. The adversarial constructions must reach a constant fraction of $\delta\Lambda_m$. And the noise model must underpredict the adversarial error by a factor that grows with $m$.
Why not just…?
Why not just use the least-squares condition number $\kappa_m$, which the barrier already controls? This is the real threat. Since $\Lambda_m \ge \kappa_m/\sqrt{2}$, it is possible that $\Lambda_m/\kappa_m$ stays bounded for every method of interest, in which case the new quantity adds nothing. The first experiment measures this ratio across methods, $m$, and oversampling ratios. If the ratio is flat within a constant, we report that as the result: least-squares stability then suffices for thresholded front-ends, which is itself a useful negative finding. We would then drop the parameter-selection claim.
Why not just model thresholding as additive white noise? Boche and Mönich stress that the thresholding error depends on the signal and can align its signs adversarially, which a noise model hides (Characterization of the pointwise and the peak value behavior of system approximation under thresholding). A random-sign perturbation typically yields error of order the $\ell^2$ norm of the kernel, while aligned signs yield its $\ell^1$ norm, which is $\Lambda_m$. The noise model is included as a baseline. The adversarial-construction experiment directly measures the gap between the two, and a small gap would favor the noise model.
Why not just oversample more? In infinite dimensions oversampling does not remove divergence (Strong Divergence for System Approximations), but at finite $m$ it may shrink $\Lambda_m$. We treat the ratio $m/n$ as a measured variable rather than a fix. If $\Lambda_m$ becomes uniformly bounded at moderate oversampling, that answers the finite-dimensional version of the open oversampling problem in the positive, and we would say so.
Why not just train a neural denoiser on thresholded signals? The ECA-Net/CycleGAN model reports a 32.74% relative improvement in $SNR_{imp}$ over CycleGAN at 5 dB mixed noise on MIT-BIH (An efficient ECG denoising method by fusing ECA-Net and CycleGAN). However, that model is evaluated only at the sample points and on a different corruption process. It offers no bound between samples, which is exactly where thresholding errors concentrate. It is not a competitor for the worst-case question. We do not run it, because its training corruption differs from ours and the comparison would not be fair.
Proposed Experiments
On synthetic analytic non-periodic test functions, the setting of all of the file's theoretical benchmarks, we will compute $\Lambda_m$ and $\kappa_m$ exactly for each method with $m$ ranging from $32$ to $4096$. We will then threshold coefficients with $\delta$ spanning $10^{-2}$ to $10^{-10}$ and record uniform error on a fine off-grid mesh. Three decisive checks follow. The first is the $\Lambda_m/\kappa_m$ growth curve. The second compares adversarially constructed functions against the bound $\delta\Lambda_m$. The third is error versus number of retained coefficients under the $\Lambda$-guided and $\kappa$-guided selection rules. The Ramanujan-sum test family will be run in exact rational arithmetic and in double precision. Any discrepancy between the two measures floating-point effects, not thresholding. A non-arithmetic control family with random rational coefficients tests whether the arithmetic structure matters at all, and if it does not, the family is reported only as a convenience. Finally, windowed segments from the MIT-BIH Arrhythmia Database, which are non-periodic on each window, serve as a realistic case of sensor thresholding. There we report uniform reconstruction error against the unthresholded reconstruction. No reported result was found on Lacuna for thresholded Fourier-sample reconstruction on MIT-BIH, so these numbers are new rather than comparisons.
References
- Characterization of the pointwise and the peak value behavior of system approximation under thresholding
- A Stability Barrier for Reconstructions from Fourier Samples
- Unboundedness of thresholding and quantization for bandlimited signals
- Strong divergence for system approximations
- On a restricted linear congruence
- A Local Fourier Extension Method for Function Approximation
- No-Go theorem for sampling-based signal processing
- Reconstruction of Piecewise Smooth Functions Based on Fourier Extension
- An efficient ECG denoising method by fusing ECA-Net and CycleGAN