High-Resolution Extrapolation Overview

• High-resolution extrapolation is a set of computational and mathematical strategies that reconstruct data beyond classical sampling limits using analytic continuation, sparsity, and learned priors.
• It leverages iterative phase retrieval, deep generative models, and operator composition to overcome conventional constraints in imaging, simulation, and communications.
• Empirical results demonstrate resolution improvements up to 2× or more, underscoring its practical applications and the challenges for further advancements.

High-resolution extrapolation is the set of computational strategies and mathematical frameworks that enable inference, synthesis, or reconstruction of data at scales or frequencies beyond those present in the training set, physical sensor, or simulation bandwidth. Research on high-resolution extrapolation spans physics (notably coherent imaging and holography), machine learning for generative models, numerical PDE simulation, communications, and sensor signal processing—each with domain-specific methodologies but unified challenges: bypassing sampling constraints, preventing information repetition, and preserving structural and fine-detail fidelity.

1. Mathematical and Physical Foundations

Classical sampling theory and diffraction/abbe-limited imaging establish strict upper bounds on resolvable detail, dictated by grid spacing or instrument numerical aperture (NA). Extrapolation aims to push beyond these limits by leveraging physical or statistical continuity, sparsity, or model-based priors.

a) Coherent imaging and analytic continuation: In coherent microscopy, the recorded complex field $U(x,y)$ is analytic owing to wave superposition. As shown in (Latychevskaia et al., 2013), this analyticity ensures that the unmeasured field outside a finite detector patch is uniquely determined (up to physical symmetries) from the measured region, provided the object support is finite. This foundation supports post-experimental field extrapolation, effectively increasing the NA and permitting super-resolution beyond the traditional $R = \lambda/(2\,NA)$ bound.

b) PDEs and operator composition: For numerical simulation, grid-induced spectral cutoffs limit the finest structures that can be directly resolved. High-resolution extrapolation in this domain uses learned coarse-to-fine operators to interpolate or forecast detail, preserving macroscopic governing dynamics while reconstructing subgrid features (Bassi et al., 2 Sep 2025).

c) Signal processing and communication theory: In wireless communications, frequency extrapolation relies on high-resolution multipath parametrization; channel responses $H(f)$ can be extrapolated far beyond the pilot band if the underlying multipath structure is well characterized (Rottenberg et al., 2019, Zeng et al., 2024). The ultimate bounds are dictated by SNR, antenna count, bandwidth, and path separability.

2. Iterative and Learning-Based Methodologies

Methodologies for high-resolution extrapolation differ by domain but fall into several canonical forms:

a) Iterative Phase Retrieval and Padding

Coherent imaging and digital holography employ iterative algorithms to reconstruct unmeasured high frequencies. After padding incomplete Fourier-domain measurements, iterative algorithms (Error-Reduction, Hybrid Input-Output) enforce measured-modulus constraints on the known region, real-space support, and physical bounds on the extrapolated domain (Latychevskaia et al., 2015, Latychevskaia et al., 2015, Latychevskaia et al., 2013, Latychevskaia et al., 2013). The process iteratively fills in the missing data, converging to a plausible full-field consistent with all constraints.

Table: Extrapolation Steps in Iterative CDI (Latychevskaia et al., 2015)

Step	Fourier Plane	Real Space
Initialization	Pad with random amplitudes	Set support mask, amplitude cap
Constraint	Known → measured modulus	Enforce finite support
Update	Amplitude+phase update	Thresholding, possible shrinkwrap

b) Extrapolation in Generative Deep Models

Text-to-image and video diffusion transformers are trained at fixed or limited resolutions, with positional encoding schemes (RoPE) encoding spatial structure. When synthesizing at higher resolutions, naively scaling positions generates periodic artifacts (repetition) and spatial incoherence. Modern strategies, including UltraImage (Zhao et al., 4 Dec 2025), DyPE (Issachar et al., 23 Oct 2025), SEGA (Rajabi et al., 21 May 2026), ExtraVAR (Yan et al., 11 May 2026), and AccDiffusion v2 (Lin et al., 2024), deploy frequency-wise remapping, entropy-guided or band-adaptive attention concentration, and content-aware patch prompting to maintain fidelity and minimize artifact proliferation.

c) Learning-Based Operator Composition in PDEs

The SRaFTE framework (Bassi et al., 2 Sep 2025) for numerical simulation proceeds by learning a super-resolution "lifting" operator $S_\theta$ from coarse to fine grid, then embedding $S_\theta$ within a predictor–corrector time loop. The process alternates between physics-respecting coarse dynamics and data-driven fine structure extrapolation, stabilizing error growth over long temporal horizons and leveraging pronounced scale separation when present.

3. Frequency and Band-Limited Analysis

A central theme is the precise analysis and manipulation of frequency content:

a) RoPE Dominant Frequency Analysis: In image diffusion transformers, the period $T_k = 2\pi/\theta_k$ of the dominant RoPE frequency matches the training resolution. When the extrapolation scale $H$ exceeds $T_k$, spatial positions alias, causing periodic repetition. Correction involves resetting $\theta_k'$ so $T_k' \geq H$, recursively updating all competing frequency peaks until residual repetition disappears (Zhao et al., 4 Dec 2025).

b) Shannon/Entropy-based Attention Scaling: Attention maps at super-resolution tend to flatten (increase entropy), diluting local structure. Adaptive scaling—either globally via attention "temperature" or per-head via entropy calibration—concentrates attention where detail is under-resolved, preserving both local texture and global structure (Zhao et al., 4 Dec 2025, Rajabi et al., 21 May 2026, Yan et al., 11 May 2026).

c) Theoretical Extrapolation Bounds: In coherent diffraction imaging, the maximum extrapolation factor $R = \lambda/(2\,NA)$0 (i.e., the size ratio between extended and original domains) is tightly governed by the oversampling ratio $R = \lambda/(2\,NA)$1; for amplitude-only constraints, $R = \lambda/(2\,NA)$2; for full complex-amplitude recovery, $R = \lambda/(2\,NA)$3 (Latychevskaia et al., 2015).

4. Domain-Specific Algorithmic Innovations

a) Patch-wise and Adaptive Conditioning

Patch-wise extrapolation in high-resolution generative models suffers from repeated object instantiation and detail collapse. Approaches such as AccDiffusion v2 split prompts into patch-specific token sets derived via attention map thresholding, while ControlNet-based local structural information injects edge priors to prevent structural distortion. Dilated sampling with window interaction further induces global coherence across patches (Lin et al., 2024).

b) Multi-Band and Stage-Aware Positional Remapping

In visual autoregressive models, naive position interpolation methods fail due to mismatch between dominant RoPE bands in different generation stages. ExtraVAR assigns each frequency band a stage-specific remapping: low-frequencies receive strict PI scaling to preserve global layout, mid-frequencies are interpolated between PI and YaRN to maintain correct object scale, and high-frequencies are preserved for local detail sharpness. Entropy-driven attention calibration is then used to match attention dispersion to the training-resolution baseline (Yan et al., 11 May 2026).

c) Adaptive and Physically-Motivated Representations

Extrapolation in physical measurement or simulation frequently benefits from sparsity or analytical properties: JSDE recovers fine-grid image content by leveraging aliasing noise-like spread in non-regular sensor layouts and enforces sparse Fourier representations to both deconvolve and extrapolate unmeasured values (Seiler et al., 2022); in channel estimation, Slepian-function or low-rank eigenbasis projections optimize extrapolation filters for MIMO systems (Zeng et al., 2024).

5. Empirical Results and Practical Limits

Empirical validation across domains demonstrates significant, though domain-limited, resolution enhancements:

• Imaging and Holography: Iterative extrapolation methods have achieved resolution improvements up to 2× or more over the measured region, with phase-retrieval transfer function (PRTF) improvements and demonstration of previously unresolved features (Latychevskaia et al., 2015, Latychevskaia et al., 2013, Latychevskaia et al., 2013, Latychevskaia et al., 2015).
• Diffusion Transformers: UltraImage reduces FID from 206 to 83 on Flux at $R = \lambda/(2\,NA)$4, and from 86.9 to 78.2 on Qwen at $R = \lambda/(2\,NA)$5 (Zhao et al., 4 Dec 2025). SEGA yields further improvement (e.g., FID from 183 to 150 at $R = \lambda/(2\,NA)$6, CLIP score increase from 25.3 to 29.2) (Rajabi et al., 21 May 2026).
• PDEs: SRaFTE outpaces bicubic and fully autoregressive baselines in both immediate resolution and error stability (e.g., relative L2 error $R = \lambda/(2\,NA)$7 vs. baseline $R = \lambda/(2\,NA)$8 on Navier-Stokes tasks) (Bassi et al., 2 Sep 2025).
• Communications: Extrapolation in SIMO/MIMO settings with high-resolution paths and $R = \lambda/(2\,NA)$9 antennas achieves reliable channel prediction over multi-bandwidth frequency spans impractical for SISO (Rottenberg et al., 2019, Zeng et al., 2024).
• Automated Segmentation: HRSAM’s length extrapolation, enabled by flexible local attention and cycle-scan SSM modules, allows scaling standard distilled models to $H(f)$0 resolutions with matched or improved quality at under half the latency of previous state-of-the-art (Huang et al., 2024).

Practical limits are dictated by memory (quadratic scaling in tokens for transformers), domain-specific constraints (analyticity, support, bandwidth), noise, and the stability of the extrapolation or phase retrieval process. Empirically, extrapolation is typically effective up to $H(f)$1–$H(f)$2 the native range for iterative phase retrieval, and up to $H(f)$3 for UltraImage before requiring further algorithmic advances (Zhao et al., 4 Dec 2025, Latychevskaia et al., 2015).

6. Open Challenges and Future Directions

Research directions motivated by the current limits and findings include:

• Learnable and multi-band periodicity correction: Existing frequency correction assumes a single dominant period; heterogeneous or variable-resolution training requires more sophisticated multi-band approaches or even learnable correction schedules (Zhao et al., 4 Dec 2025, Rajabi et al., 21 May 2026).
• Extension to video and spatio-temporal domains: Video diffusion models require spatio-temporal RoPE and adaptation for high-resolution temporal extrapolation, as addressed in preliminary form by ViBe (Wu et al., 24 Mar 2026).
• Dynamic adaptive attention block sizing: For large panoramas or ultra-wide inputs, resizing blockwise attention dynamically could offset memory limitations.
• Unified latent-manifold approaches: Findings in VAE student–teacher distillation suggest opportunities for manifold-aligned compression and high-resolution encoding transferable across tasks (Chu et al., 15 Mar 2026).
• Entropy and spectral structure–guided adaptation: Both per-head entropy and Fourier spectral profiles are leveraged to adapt attention or positional encoding; further exploration into content-adaptive, learnable adaptation remains open (Rajabi et al., 21 May 2026, Yan et al., 11 May 2026).
• Task-specific and uncertainty-aware extrapolation: Quantification of uncertainty, error bounds, and task-aligned fidelity across modalities could inform more robust extrapolation methods, especially for decision-critical applications.

7. Theoretical and Practical Synthesis

High-resolution extrapolation is anchored by the interplay of physical and algorithmic constraints (analyticity, sampling, support) and modern statistical, deep learning, and adaptive attention mechanisms. Core principles include the preservation or controlled modulation of dominant frequencies, the enforcement of physical or data-driven constraints via iterative or learning-based operators, the explicit characterization and mitigation of artifacts (periodic repeats, attention flattening), and adaptive focus on both global structure and local detail. Contemporary advances bridge physically-motivated analytic continuation with data-driven frequency- and entropy-guided adaptation, enabling state-of-the-art extrapolation in applications ranging from imaging and sensing to generative modeling and automated segmentation (Zhao et al., 4 Dec 2025, Latychevskaia et al., 2015, Rajabi et al., 21 May 2026, Yan et al., 11 May 2026, Zeng et al., 2024, Bassi et al., 2 Sep 2025, Huang et al., 2024).