Spectrally-Filtered Direct-Insertion Downscaling

• The paper introduces a spectrally-filtered direct-insertion method that achieves exponential synchronization by preconditioning data with spectral smoothing before insertion.
• It integrates Fourier and de la Vallée-Poussin polynomial filters with direct-insertion schemes to reduce aliasing in PDE models and image processing tasks.
• Applications span data assimilation in Navier–Stokes simulations and high-quality image downscaling, demonstrating improved accuracy and robustness over traditional methods.

A spectrally-filtered direct-insertion downscaling method combines spectral filtering with direct-insertion techniques to reconstruct or downscale fields or signals—especially in data assimilation, numerical simulation of partial differential equations (PDEs), and image processing. This approach enables exponential convergence and robust synchronization properties even when the original observational data reside in non-orthogonal, physical-space representations or are prone to aliasing. Distinct variants of the method appear in both mathematical PDE filtering (Celik et al., 2018, Celik et al., 11 Jan 2026) and spectral polynomial interpolation for signal and image processing (Occorsio et al., 2021), unified by the core principle of spectral smoothing preconditioning before insertion or resampling.

1. Mathematical Foundations and Spectral Filtering

Spectral filtering, central to these downscaling methods, entails projecting or smoothing data onto a space of low-wavenumber (smooth) components prior to usage in reconstruction or model correction. For PDEs such as the 2D Navier–Stokes equations, this is realized through Fourier projection:

• The projection operator $P_K$ restricts the field to Fourier modes $|k|^2 < K$.
• For interpolant observables $I_h$—physical-space interpolants such as local averaging or nodal sampling—a composite insertion operator $J = P_K P_\sigma I_h$ is constructed, where $P_\sigma$ (or $P_H$) is the Leray projection onto divergence-free subspaces.
• The operator $J$ nearly preserves $L^2$ and $H^1$ norms of smooth fields with proper choice of cutoff $K$ and spatial scale $h$, ensuring that insertion does not introduce high-frequency artifacts (Celik et al., 2018).

Spectral interpolation for image processing realizes spectral filtering via polynomial bases. The de la Vallée-Poussin (VP) filter is applied to damp high-frequency modes within the Chebyshev polynomial expansion (Occorsio et al., 2021).

2. Direct-Insertion Schemes and Algorithmic Structure

Direct-insertion, as opposed to nudging or continuous correction, modifies the simulated or reconstructed field at discrete times by partially or fully replacing components with observed values. The procedure in data assimilation is as follows:

• At observation times $t_n$, the model state is advanced by the Navier–Stokes semi-group, $\hat{u}(t_{n+1}) = S(t_{n+1}, t_n; u_n)$.
• The updated state is constructed as $u_{n+1} = E \hat{u}(t_{n+1}) + J U(t_{n+1})$, with $E = I - J$.
• This process repeats, resulting in a piecewise-strong trajectory $u(t)$ which restarts at each $t_n$ from a spectrally-filtered, observationally-informed state.

In multidimensional polynomial interpolation (for downscaling images), the direct-insertion method evaluates a global, spectrally-filtered polynomial interpolant at the sample points of the desired lower-resolution grid, omitting convolutional or localized filtering traditional in standard kernels (Occorsio et al., 2021).

3. Interpolant Classes, Parameterization, and Filter Design

Interpolant observables $I_h$ are central and fall into two broad classes:

• Type-I: $I_h: V \to L^2$—physical observables or averages satisfying $\|u - I_h u\|_{L^2}^2 \leq c_1 h^2 \|u\|_{H^1}^2$.
• Type-II: More general, incorporating higher regularity, $$\|u - I_h u\|_{L^2}^2 \leq c_1 h^2(\|u\|_{H^1}^2 + h^2 \|A u\|_{L^2}^2)$$.

The parameters $\delta$ (insertion interval), $h$ (mesh scale), and $K$ (spectral cutoff) are interrelated:

• $\delta$ must be large enough to prevent over-insertion but not so large as to allow high-frequency error to grow.
• $h$ and $K$ are linked via $c_1 h^2 K \lesssim \epsilon \ll 1$.
• In image downscaling, the VP filter parameter $m$ and fractional cutoff $\theta = m/n$ dictate the spectral attenuation and tradeoff between smoothness and accuracy.

For image processing, the Chebyshev zero grid, combined with the VP-filtered basis, minimizes interpolation error and aliasing by controlling the spectral content passed through each stage.

4. Theoretical Results and Convergence

For the 2D Navier–Stokes setting, significant convergence theorems are established (Celik et al., 2018, Celik et al., 11 Jan 2026):

• Exponential synchronization: The difference $\|u(t) - U(t)\|_{L^2}$ decays exponentially, uniformly over time, provided $\delta, h, K$ satisfy the theoretical bounds.
• For Type-II observables, convergence is additionally proved in the vorticity norm.
• Introduction of a relaxation parameter $\alpha \in (0,1]$ in the discrete-in-time insertion formula allows robust convergence even as the observation frequency increases, overcoming the instability observed with $\alpha = 1$ and small $\delta$ (Celik et al., 11 Jan 2026).

In the continuous-time limit as $\delta \to 0$ and $\alpha = \mu \delta$, the method converges to the continuous-in-time nudging algorithm, rigorously connecting the discrete-in-time insertion scheme to standard nudging frameworks.

5. Applications and Algorithmic Implementation

Data Assimilation and Dynamical Systems

The spectrally-filtered direct-insertion method rigorously enables discrete-time synchronization for general classes of interpolants—encompassing sensor networks, pointwise measurements, and volume averages. It extends the applicability of discrete Fourier-mode replacement, previously limited to orthogonal projections, to physically-motivated non-orthogonal data (Celik et al., 2018).

The relaxation mechanism (Celik et al., 11 Jan 2026) allows for varying observational cadence without degrading convergence, facilitating robust real-time assimilation systems.

Image and Signal Processing

The direct-insertion, spectrally-filtered strategy, notably the de la Vallée-Poussin filtered interpolation (d-VPI), is applied to image scaling, yielding superior quantitative results compared to standard kernel-based methods. Specifically:

• The method operates via precomputing Chebyshev-based VP-basis matrices followed by matrix multiplications, resulting in computational efficiency and BLAS/GPU suitability.
• On image downscaling benchmarks, d-VPI provides 1–2 dB higher PSNR than bicubic interpolation and better SSIM scores. For specific integer scales, exact reconstruction is achieved with zero MSE (Occorsio et al., 2021).

6. Spectral Attenuation: Control, Benefits, and Trade-offs

The VP filter and its analogs enforce flat passbands up to controllable cutoff frequencies and a linear roll-off for higher frequencies. This guarantees faithful reproduction of low-frequency content while automatically suppressing high-frequency noise or artifacts that would induce aliasing or instability upon resampling or model insertion.

• In data assimilation, this creates an almost-orthogonal projection, crucial for exponential contractivity and stability even when the physical-space interpolant $I_h$ has no spectral structure.
• In image processing, it yields downscaled images with minimal artifacts, sharp edges, and accurate texture reproduction, outperforming both Lagrange interpolants and traditional smoothing kernels (Occorsio et al., 2021).

A plausible implication is that parametrizable spectral filters enable domain-transfer and adaptive reconstruction strategies across disparate contexts—PDE modeling and signal processing—while maintaining theoretical guarantees.

7. Computational and Practical Considerations

The per-step computational cost in both PDE and image contexts is dominated by basis evaluation and matrix multiplications. The method is efficient for moderate-to-large-scale applications, requiring only the storage of dense basis matrices and input data. Instability for too-small $\delta$ in the original unrelaxed scheme is resolved by the introduction of relaxation (Celik et al., 11 Jan 2026). In image processing, real-time downscaling is achievable for several thousand-pixel grids due to the algorithm’s amenability to vectorized and parallel computation (Occorsio et al., 2021).

Summary Table: Spectrally-Filtered Direct-Insertion Downscaling Variants

Context	Spectral Filter Mechanism	Principal Benefit
PDE/Assimilation	Fourier cutoff, $J=P_K I_h$	Exponential synchronization
Image Processing	de la Vallée-Poussin polynomial	Alias suppression, sharp edges
Relaxed Updating	Direct-insertion with $\alpha<1$	Robustness to observation freq

The spectrally-filtered direct-insertion downscaling method unifies advances from spectral theory, operator approximation, and numerical analysis to achieve optimal downscaling and data insertion. The method is notable for its rigorous convergence properties, flexibility regarding observational modalities, and practical efficiency across physical and computational domains (Celik et al., 2018, Celik et al., 11 Jan 2026, Occorsio et al., 2021).