Many functions that scientists want to learn are smooth almost everywhere but jump across a surface. Examples are the solution of a fluid equation with a shock, a material property across a phase boundary, or a classifier's decision boundary. In many input dimensions, two things go wrong at once. Smooth approximators such as Fourier modes, polynomials, or smooth networks ring around the jump; this is the Gibbs phenomenon. Piecewise-constant bases such as Haar wavelets leave blocky artifacts and need a number of cells that grows exponentially with dimension. Our idea is to move the jump out of $d$ dimensions and into one. We write the function as a sum of one-dimensional "outer" functions applied to smooth learned "inner" maps, following the Kolmogorov superposition theorem, which says any continuous multivariate function can be built from sums and compositions of univariate functions. A jump across a curved surface then becomes an ordinary jump in a scalar function of one variable. Placing a discontinuity in one dimension, cheaply and without ringing, is a problem classical wavelet theory already knows how to solve.
The proposal therefore treats the three directions unequally. The curse of dimensionality is the regime. Basis discontinuity is the phenomenon we must represent. The wavelet support trade-off decides which one-dimensional basis to use for the outer functions. In one dimension, the rule that a scalar wavelet can be orthogonal, symmetric, and compactly supported only if it is the discontinuous Haar wavelet stops being a dead end. Multiwavelets, which use a small vector of scaling functions instead of one, escape it, and shifted redundant frames remove Haar's grid blind spots. We will build and test a superposition network whose inner maps are smooth. Its outer functions are expanded in a sparse, shift-redundant frame of balanced, orthogonal, symmetric, compactly supported multiwavelets. The number and location of the jumps are found by coefficient sparsity rather than specified in advance.
Core Claim
Consider a target of the form $f(x) = s(x) + \sum_{i} h_i \, \mathbf{1}[g_i(x) > c_i]$ on $[0,1]^d$. Here $s$ is smooth, each $g_i$ is a smooth scalar map whose level set is the discontinuity surface, $h_i$ is the jump height, and the number of jumps is unknown. We claim three things.
First, a two-level model $f(x) \approx \sum_{q} \Phi_q(u_q(x))$ can represent this target. The inner maps $u_q$ are smooth learned scalars. Each outer function $\Phi_q$ is written as $\Phi_q(t) = \sum_{j,k,\tau} a^{q}_{j,k,\tau}\, \psi_{j,k}(t-\tau)$, a combination of compactly supported multiwavelets $\psi_{j,k}$ at scale $j$, position $k$, and grid shift $\tau \in \{0, 1/3, 2/3\}$. In such a model the jumps cost only a few fine-scale coefficients in the $\Phi_q$. That count depends on the number of jumps and the resolution, not on $d$.
Second, choosing orthogonal, symmetric, balanced multiwavelets with shifted copies has measurable consequences. It gives lower overshoot at the jump and no spurious oscillation in the flat regions between jumps. It also locates the jump with less dependence on the dyadic grid than B-splines, Haar, or unshifted multiwavelets under the same parameter and FLOP budget.
Third, because jumps show up as isolated large fine-scale coefficients, a sparsity penalty discovers how many jumps there are. The closest existing method needs that count as an input.
Where This Sits in Prior Work
Several papers motivate the gap. Hybrid Surrogate Models splits a solution into Chebyshev polynomials plus explicit Heaviside steps and trains with distributional derivatives. It reports an $l_1$ error of $7.33\times10^{-5}$ against $4.92\times10^{-3}$ for the best neural network on a non-linearly propagating shock. The method is strictly one-dimensional and needs the number of shocks fixed in advance. We keep its insight that jumps should be represented explicitly, but we make the jump surface a learned level set in $d$ dimensions and let sparsity choose the count. Multiwavelet-based Operator Learning shows that compactly supported Legendre and Chebyshev multiwavelets compress PDE operators well. It reports relative $L_2$ error of about 0.0018 to 0.0033 on 1D Burgers and KdV, a 2x to 10x improvement over the Fourier Neural Operator. That work uses the multiwavelets on the spatial grid, never addresses balancing, and has no explicit shock representation. We use the same family on the one-dimensional outer axis, where the jump lives.
Two papers supply the dimensional structure and also constrain it. KAN: Kolmogorov–Arnold Networks places learnable B-splines on network edges. The Kolmogorov Superposition Theorem can Break the Curse of Dimensionality proves an $\mathcal{O}(1/n)$ rate that does not depend on $d$ for a restricted Lipschitz class, using smoothed splines. Neither targets discontinuous functions: B-spline edges smooth over jumps, and the KST bound excludes them. KAN or MLP: A Fairer Comparison is the constraining result we take most seriously. It finds that KAN's advantage on symbolic tasks comes mostly from the B-spline activation, and that an MLP with B-spline activations matches it. This means our contribution must be the outer basis, not the edge-based architecture. The experiments are designed so an MLP inner map with our outer layer is a first-class variant.
The one-dimensional basis design draws on four papers. Compactly Supported Orthogonal Symmetric Scaling Functions establishes the scalar trilemma: with dilation 2, only Haar satisfies all three properties. It also constructs symmetric orthogonal scaling functions for dilations of 3 or more, which we use as an alternative outer basis in ablations. Haar-type multiwavelet bases shows that matrix-valued filters achieve orthogonality, symmetry, and short support together. The Design of Matched Balanced Orthogonal Multiwavelets warns that unbalanced multiwavelets inject high-frequency oscillation into constant signals, which is exactly the flat region between two jumps. It also warns that balancing to second order uses up most of the design freedom. We therefore fix a balanced family rather than learning filter shapes. Regularity properties of Haar Frames shows that three shifted Haar bases remove the dyadic blind spots and recover Hölder regularity at every point. We borrow the shift construction so that a jump at a non-dyadic location is still captured by a few coefficients. Wavelet-Based Density Estimation for Persistent Homology shows that hard-thresholded Haar coefficients can be very sparse (98.5% compression) but produce blocky, piecewise-constant output. This is the failure we expect an unbalanced or low-order outer basis to reproduce.
Research Program
We will build the model as follows. The inner maps $u_q$ come in two versions: a small KAN-style additive layer $u_q(x)=\sum_p \phi_{q,p}(x_p)$ with cubic splines, or a small MLP. Each outer $\Phi_q$ is a finite expansion in a balanced, symmetric, orthogonal, compactly supported multiwavelet family with the three grid shifts. Coarse-scale coefficients carry the smooth part and fine-scale coefficients carry jumps. Training minimizes squared error plus an $\ell_1$ penalty on fine-scale coefficients weighted by scale, followed by hard thresholding. The inner maps are normalized to $[0,1]$ so the outer grid stays meaningful.
The analysis has one theoretical target. For targets whose jump surfaces are level sets of functions in the inner-map class, we will bound the number of nonzero fine-scale outer coefficients needed for accuracy $\epsilon$ and for overshoot below $\delta$ near the surface. We expect this count to be proportional to the number of jumps times $\log(1/\epsilon)$ and free of $d$. Compact support lets each jump touch only a bounded number of coefficients per scale, and the shifted frame removes the dependence on where the jump falls relative to the grid.
We also state honestly what the model cannot do. Jump surfaces outside the inner class, such as intersecting fronts that no single level set describes, need several outer channels. We will measure how many.
Baselines
The naive reference point is a plain ReLU MLP, trained with Adam and matched in FLOPs to our model. A ReLU network is piecewise linear and can build steep ramps, so it is a genuine competitor on jumps. It anchors every comparison. The HSM paper's reported neural-network $l_1$ of $4.92\times10^{-3}$ on its shock test tells us roughly where it should land.
The strongest existing method on the one-dimensional shock testbed is the Hybrid Surrogate Model from Hybrid Surrogate Models, with a reported $l_1$ error of $7.33\times10^{-5}$ and a runtime of about 19 to 29 seconds on the transport equation. On that testbed we reimplement it and give it the true shock count. Beating it there means matching its $l_1$ within a factor of two without being told the count. In higher dimensions we extend HSM ourselves into a Heaviside of a learned smooth map plus a smooth network, again given the true count. This is the fairest possible version of the "explicit jump" approach. Beating it means lower RMSE and lower overshoot once the count is wrong or unknown, and no worse accuracy when the count is right.
For smooth-basis competition in operator-style settings, the strongest reported number we have is the Multiwavelet Neural Operator's relative $L_2$ error of about 0.0018 to 0.0033 on 1D Burgers from Multiwavelet-based Operator Learning. Those figures come from operator learning, not pointwise approximation, so we cite them for scale rather than as a direct target. The PDEBench paper PDEBench gives no subset error numbers that we can use; no reported result was found on Lacuna for Burgers' RMSE or fRMSE per dataset.
The architecture baselines are two. The first is KAN with B-spline edges and grid extension from KAN: Kolmogorov–Arnold Networks. The second, following KAN or MLP: A Fairer Comparison, is an MLP whose activations are B-splines, which that paper shows matches or beats KAN on symbolic tasks.
The ablations isolate the key ingredient by holding the inner maps, parameter count, FLOPs, optimizer, and a shared tuning budget of 50 random hyperparameter trials fixed. Only the outer basis changes. The outer basis options are cubic B-splines, order-zero splines (Haar-like), Haar with shifts, unbalanced orthogonal multiwavelets, balanced multiwavelets without shifts, and smooth Daubechies wavelets. Our claim holds only if the full balanced, shifted multiwavelet variant beats each of these on overshoot and jump localization without losing smooth-region RMSE. Every method sees identical training samples and five random seeds.
Why not just…?
Why not just use a deep ReLU MLP, which already adapts to low-dimensional structure? A ReLU network can in principle build a jump across a curved surface. When Do Neural Networks Outperform Kernel Methods? shows that trained networks learn relevant low-dimensional directions where fixed kernels fail. The evidence against it on jumps is only one-dimensional: the two-orders-of-magnitude gap reported by the HSM paper. We do not assume that gap carries over to higher dimensions. The dimension-sweep experiment below tests it directly. If a FLOP-matched MLP reaches the same overshoot and localization, the claim fails.
Why not just refine the B-spline grid in a KAN, or use an MLP with B-spline activations? This is the most serious threat. The fairer-comparison paper shows the spline activation is doing the work, and a fine enough B-spline grid does localize a jump. B-splines are symmetric and compactly supported but not orthogonal. That makes sparse thresholding of their coefficients poorly conditioned, and they keep overshooting near a jump that falls between knots. The basis ablation is designed to settle this. If B-splines with adaptive knots match our overshoot and localization at equal FLOPs, the correct conclusion is that the trilemma is irrelevant here, and we would report it.
Why not just extend the Hybrid Surrogate Model with learned level sets? When the shock count is known and the fronts do not merge, this may well win, because it spends parameters only on what is needed. Our advantage has to come from unknown counts, fronts that appear or merge, and tolerance to misspecifying the count. We give HSM the true count, and also a count off by one, so the comparison makes clear where each approach is preferable.
Why not just use Haar on the outer axis, since it is already orthogonal, symmetric, and compact? Haar puts jumps exactly where they belong, but it represents the smooth part in blocks and has one vanishing moment. As a result it cannot capture smoothness beyond Hölder order 1, as Regularity properties of Haar Frames shows. The ablation with shifted Haar measures whether higher-order balanced multiwavelets buy smooth-region accuracy at equal cost.
Proposed Experiments
The first testbed is synthetic piecewise-smooth targets in dimensions 2, 5, 10, and 20. Each has one to four jump surfaces, both additive and non-additive, including a merging pair. The smooth parts are drawn from the ten KST test functions of The Kolmogorov Superposition Theorem can Break the Curse of Dimensionality, with added jump terms. The second testbed is the one-dimensional shock and transport problems from the HSM paper. The third uses parametric Burgers solution fields from PDEBench, treated as functions of space, time, and viscosity. For this third testbed we report RMSE and PDEBench's Fourier-spectrum deviation (fRMSE). Across all testbeds the metrics are RMSE away from the jumps, maximum overshoot within a band around the jumps (the Gibbs measure), jump-location error, recovered versus true jump count, FLOPs, and wall-clock training time.
The decisive check is the dimension sweep. We fix the number of jumps and the smooth-part difficulty, then plot the parameters needed to reach a target overshoot and RMSE against $d$. If the claim is right, our model's curve stays nearly flat for jump surfaces within the inner class. The MLP and B-spline curves should rise, or plateau at higher overshoot.
A second check places jumps at dyadic versus irrational positions on the outer axis. Unshifted bases should show a clear accuracy gap between the two cases, and the shifted frame should not.
A third check hides the jump count. Our method should recover the correct count on at least most seeds, while count-specified HSM degrades when given the wrong count.
If B-spline outer functions match ours on all three checks at equal FLOPs, we will conclude that one-dimensional placement of the jump is what matters and the choice of wavelet basis is not. That finding would still be useful, but it would be a smaller claim than the one we set out to test.
References
- Hybrid Surrogate Models: Circumventing Gibbs Phenomenon for Partial Differential Equations with Finite Shock-Type Discontinuities
- Multiwavelet-based Operator Learning for Differential Equations
- KAN: Kolmogorov–Arnold Networks
- The Kolmogorov Superposition Theorem can Break the Curse of Dimensionality When Approximating High Dimensional Functions
- KAN or MLP: A Fairer Comparison
- Compactly Supported Orthogonal Symmetric Scaling Functions
- Haar-type multiwavelet bases and self-affine multi-tiles
- The Design of Matched Balanced Orthogonal Multiwavelets
- Regularity properties of Haar Frames
- Wavelet-Based Density Estimation for Persistent Homology
- PDEBench: An Extensive Benchmark for Scientific Machine Learning
- When Do Neural Networks Outperform Kernel Methods?