Fast Fourier methods are accurate and cheap, but they assume the function repeats periodically. Fourier extension removes that assumption. It fits the data on the physical interval with Fourier modes that are periodic on a larger interval, so the artificial jump at the boundary disappears. The fit is badly ill-conditioned, so in practice it is solved with a truncated singular value decomposition (TSVD), which keeps only the well-determined directions of the matrix. Our proposal is to reuse that same decomposition as a smoothness sensor. We project each block of data onto the singular directions and ask whether the projected coefficients fall off at least as fast as the singular values. When they do, the block is smooth. When they flatten out, the block contains a jump. The Gibbs phenomenon (the roughly 9% overshoot and ringing that any global smooth basis produces at a jump) then shows up as a measurable, amplitude-independent signature, and it is available at no extra cost inside a Fourier-continuation solver for shock-forming equations.
This matters because current Fourier-continuation shock solvers either train a neural network to find shocks or rely on mollifier edge detectors that are sensitive to noise and steep smooth gradients. The one purely linear-algebraic detector in this area, a spike in the TSVD coefficient norm, has only been used on static one-dimensional functions with hand-set thresholds. We want to turn that observation into a principled, training-free, scale-invariant sensor and put it inside a time-dependent solver with moving shocks.
Core Claim
Following A Local Fourier Extension Method for Function Approximation, we split the domain into $K$ uniform blocks. Every block then has the same frame matrix $A = U\Sigma V^*$, computed once. For block data $b_k$, the discrete Picard coefficients $\beta_{k,i} = |u_i^* b_k|$ form an orthonormal expansion of that block's samples in the physical domain, ordered by singular value $\sigma_i$. We define the block's Picard slope $s_k$ as the least-squares slope of $\log \beta_{k,i}$ against $\log \sigma_i$ over the retained modes $\sigma_i \ge \tau$, restricted to coefficients above a small multiple of the block's noise or rounding floor. Our hypothesis is that $s_k \gtrsim 1$ on smooth blocks and $s_k \approx 0$ on blocks with an interior jump. The reasoning: the singular values decay geometrically, while a jump's projections onto oscillatory modes decay only algebraically.
The mechanism comes from On the Numerical Stability of Fourier Extensions. That paper shows that singular vectors with tiny $\sigma_i$ live mostly in the buffer zone, so analytic data barely excite them. A jump inside the block has no smooth extension. It is forced to excite those modes, and this is our explanation for the norm spike reported by Zhao and Wang. Because $s_k$ is a slope and not a norm, it does not change when the data are rescaled. That removes the need for per-problem thresholds.
The sensor drives a localized artificial viscosity (numerical smoothing added only near shocks),
$$\nu_k = \nu_{\max}\,\mathrm{clip}\big((s^\ast - s_k)/s^\ast,\,0,\,1\big),$$
where $\nu_{\max}$ is a mesh-scaled viscosity ceiling and $s^\ast$ is a slope threshold fixed once on the smooth Adcock–Ruan functions. This replaces the trained shock-detecting network of FC-based shock-dynamics solver with neural-network localized artificial-viscosity assignment.
The claim is deliberately narrow. TSVD alone does not remove Gibbs ringing. It is a linear reconstruction, and Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities proves that any linear reconstruction of a moving discontinuity is limited to $O(1/p)$ error in the number of basis functions $p$. That result constrains our design: the SVD is used to decide where to act, and the viscosity or subdivision does the nonlinear work.
Research Program
The work has three parts.
Analysis. For a model jump inside a block, we will bound the decay of $\beta_{k,i}$ using the prolate-type structure of the frame, which is the same spectral object whose asymptotics Parameter selection and numerical approximation properties of Fourier extensions from fixed data leaves open. We will also derive how $s_k$ behaves for a steep but smooth tanh front of width $\delta$. We expect a "delayed" regime: slope near zero for the first modes, recovering to at least one once the front is resolved. That suggests fitting the slope over the tail rather than over all modes, which is how the sensor can separate a true jump from a steep smooth gradient. Reconstruction of Piecewise Smooth Functions Based on Fourier Extension identifies exactly this confusion as the weakness of mollifier detectors.
Static validation. We will test on the function suites of Adcock–Ruan, Li–Zhao–Jia and Zhao–Wang.
Dynamic solver. We will integrate the sensor into a Fourier-continuation solver with SSPRK-4 time stepping (fourth-order strong-stability-preserving Runge–Kutta), matching the FC-SDNN setup. It will be run on Burgers and compressible Euler equations in one dimension and on tensor-product two-dimensional cases. This is the moving-discontinuity setting that both 2025 local Fourier extension papers leave open.
Baselines
Naive reference. Fourier continuation with no shock treatment, plus a version with a global uniform viscosity tuned to just suppress blow-up. The first shows the raw Gibbs ringing. The second shows the cost of smearing smooth regions, the over-dissipation problem noted by Leibovici et al.
Strongest existing method. FC-SDNN FC-based shock-dynamics solver with neural-network localized artificial-viscosity assignment. It is the only method in the file that combines Fourier continuation with localized, shock-targeted viscosity on non-periodic shock problems, so it answers exactly our question. We will re-implement it on identical grids, time steps and viscosity kernels, differing only in the detector. The network's training data come from its paper's protocol, and our threshold $s^\ast$ gets a matched tuning budget on disjoint problems. The file contains no reported error numbers for FC-SDNN, so we will report both detectors from our own runs. Beating FC-SDNN means matching its $L^1$ error and shock width on Burgers and shock-tube problems within run-to-run variation, with zero training cost and lower per-step detector cost. It also means strictly lower false-positive viscosity on smooth regions. For context on problem difficulty only, Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities reports 28.5% error for DeepONet and 1.57% for FNO on inviscid Burgers. Those are learned-operator numbers, not solver numbers, and are not a target.
Two detector-level baselines. The first is Zhao–Wang's coefficient-norm spike with its heuristic threshold, applied to the same blocks. The second is the mollifier edge detector of Li–Zhao–Jia.
Key ablation. Replacing the Picard slope with the raw coefficient norm, holding everything else fixed, isolates whether scale invariance is the active ingredient.
Why not just…?
Why not just use the TSVD coefficient-norm spike that Zhao and Wang already found? The norm grows with signal amplitude, so a fixed threshold misfires when shock strength or units change. In a time-dependent flow, shock strength varies both along the domain and over time. The ablation above tests this directly. We will sweep jump heights over four orders of magnitude and measure whether one threshold serves all of them. If the norm works equally well, our contribution shrinks to the analysis and the dynamic integration, and we will say so.
Why not just use a classical modal-decay sensor on local Legendre coefficients? This is the real threat. The Picard slope is, in effect, a modal-decay sensor. The difference is that ours is expressed in the SVD basis of the frame the solver already inverts, so it costs one extra small matrix product per block and adds no second basis. We will implement a Legendre modal-decay sensor on the same blocks as a fourth detector. If it matches ours on accuracy, we will present the method honestly as a zero-overhead instance of that idea. The remaining claimed advantage would then be the buffer-mode explanation and the tail-slope fit that separates steep smooth fronts from jumps, and we test that separately.
Why not just keep the neural shock detector? It works, but it needs training data, carries inference overhead, and has no guarantee outside its training distribution. The head-to-head comparison described under Baselines, plus out-of-distribution tests on Euler states unseen in training, will show whether a training-free linear-algebra sensor is competitive.
Proposed Experiments
Static discrimination. On the nine Adcock–Ruan functions, which are smooth, including Runge-type and flat $C^\infty$ cases, we measure the fraction of blocks falsely flagged. On the Li–Zhao–Jia and Zhao–Wang piecewise-smooth suites, we measure jump-location recall and precision. Both groups of authors report $10^{-14}$ to $10^{-15}$ reconstruction error once breakpoints are known. Our detector plus their local extension should reach that level with no manual partition.
Decisive steep-front test. We use a family of tanh fronts whose width $\delta$ shrinks toward the grid spacing, alongside true jumps. The sensor passes if flags turn on only once $\delta$ falls below what the block can resolve, and it clearly beats both the mollifier and the norm detectors there.
Dynamic tests. We run inviscid Burgers with shock formation and the Sod-type shock tube from the Lanthaler et al. benchmarks, plus FC-SDNN's one- and two-dimensional Euler cases. We report $L^1$ error against high-resolution references, shock width in cells, total variation increase, error in smooth regions away from shocks, and wall-clock time per step.
Kill criterion. The idea is abandoned if, on the shock tube, the Picard sensor either smears smooth regions more than FC-SDNN or misses shocks that the Legendre modal sensor catches.
References
- A Local Fourier Extension Method for Function Approximation
- On the Numerical Stability of Fourier Extensions
- FC-based shock-dynamics solver with neural-network localized artificial-viscosity assignment
- Nonlinear Reconstruction for Operator Learning of PDEs with Discontinuities
- Parameter selection and numerical approximation properties of Fourier extensions from fixed data
- Reconstruction of Piecewise Smooth Functions Based on Fourier Extension