FUTON: Fourier Tensor Networks for Implicit Neural Representations

arXiv:2602.13414 · eess.IV, cs.CV, cs.LG, eess.SP · Submitted 2026-02-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: We're looking at the paper: "FUTON: Fourier Tensor Networks for Implicit Neural Representations".

Dev: Here's the gist of the paper. Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, overfitting to noise, and poor extrapolation FUTON implicitly expresses signals as weighted combinations of orthonormal, separable basis functions, combining complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure Novel Architecture FUTON introduces a novel architecture that models signals as generalized Fourier series whose coefficients are parameterized via a low-rank tensor decomposition This design embeds complementary inductive biases: Fourier bases capture smoothness and periodicity, while low-rank factorization captures the intrinsic low-dimensional spectral structure of natural signals The architecture leverages the insight that structured signals exhibit frequency-domain correlations amenable to low-rank approximation Theoretical Foundations and Efficiency The paper provides theoretical guarantees through a universal approximation theorem establishing that FUTON can represent any L2 function to arbitrary precision Furthermore, it establishes FUTON’s linear complexity through efficient tensor contraction, enabling scalability to high-dimensional signals The forward pass proceeds in three stages with a total complexity of O(CKR + CR + DR), which scales linearly with the input and output dimensions C and D Superior Performance Across Tasks Extensive experiments demonstrate that FUTON achieves higher reconstruction quality, faster convergence, and better generalization on inverse problems compared to state-of-the-art methods For signal representation, FUTON substantially outperforms state-of-the-art MLP-based INRs such as WIRE while converging significantly faster In inverse problems, it generalizes better and converges faster on image super-resolution and CT reconstruction Ablation Studies Ablation studies investigate the effect of key design choices, including the CP rank R, number of basis functions K, and the one-dimensional basis φ(·) Increasing R improves PSNR with diminishing returns beyond R = 512 for image representation Increasing K to K = 512 monotonically improves reconstruction quality across all ranks The cosine basis is favored because it achieves higher PSNR and faster convergence compared to polynomial bases Training Efficiency Analysis FUTON demonstrates superior wall-clock performance across all tasks, reaching 30 dB in approximately 10s for super-resolution compared to over 25s for WIRE In image denoising, FUTON achieves 27 dB in under 20s while WIRE requires approximately 60s—a 3× speedup The CP decomposition decouples learning of global spectral structure from local refinements, enabling efficient gradient-based optimization FUTON presents a structured INR that unifies generalized Fourier series with a low-rank CP decomposition of the coefficient tensor This design addresses common limitations of MLP-based INRs (e.g., slow convergence, overfitting, and poor extrapolation), while offering a linear-time forward pass and universal approximation in L2 Extensive experiments across diverse image and signal tasks validate the approach Promising directions include exploring alternative orthonormal bases (e.g.

Paper summary: Rosa: So we're looking at this paper called FUTON: Fourier Tensor Networks for Implicit Neural Representations, and the authors are from Ghent University and Vrije Universiteit Brussel. It seems like they’re tackling some big problems with how we use neural networks to represent signals.

Dev: Yeah, the core idea here is that typical neural network designs, especially those based on MLPs, have these issues: slow convergence, they tend to overfit noise in the data, and they struggle when you try to predict things that are outside what they've seen before.

Rosa: Exactly. So FUTON tries to fix that by modeling signals as generalized Fourier series and parameterizing the coefficients with a low-rank tensor decomposition. They claim this combines two different ways of thinking about the signal, which is pretty interesting because it adds complementary inductive biases.

Dev: It’s combining Fourier bases, which are great for capturing smoothness and periodicity in a signal, with this low-rank structure that enforces a lower dimensional spectral shape on the data. That suggests they think structured signals naturally have correlations in their frequency domain that can be captured efficiently.

Taro: From my side, I'm curious how this structural representation holds up when the real world throws weird stuff at it, like noise or complex deformations. Does this low-rank structure allow it to generalize better than just a standard MLP?

Rosa: Well, the paper claims they provide a universal approximation theorem which basically says FUTON can represent any function in L two with arbitrary precision <ref:2602.13414#pg1>. That’s a strong theoretical guarantee for how good the representation itself can be.

Dev: But theory is one thing, and running it fast in practice is another. They establish linear complexity through efficient tensor contraction, which means the computational cost scales linearly with both the input dimension and the output dimension. That’s a big deal for scalability when you deal with high-dimensional signals.

Taro: Linear complexity sounds promising for real-time applications or very large datasets, but what about latency? If we're talking about real-time loops, how does this tensor decomposition factor into the actual inference speed on hardware?

Rosa: The forward pass is broken down into three stages, and the total complexity ends up being O of C K R plus C R plus D R. That means it scales linearly with those input and output dimensions C and D.

Dev: So, if we look at performance across tasks, they say FUTON achieves higher reconstruction quality, faster convergence compared to state-of-the-art MLP-based INRs like WIRE. They also show better generalization on challenging inverse problems like image super-resolution and CT reconstruction.

Taro: Faster convergence is definitely something I care about when we're testing systems, but how does that translate when the world misbehaves? If the input signal is noisy or corrupted, does this structure help it recover gracefully?

Rosa: The ablation studies show that increasing the rank R improves PSNR with diminishing returns beyond a certain point of five hundred twelve for image representation. And increasing the number of basis functions K to five hundred twelve actually improves reconstruction quality across all ranks.

Paper summary: Dev: They also mention favoring the cosine basis because it achieves higher PSNR and faster convergence than using polynomial bases. That’s a concrete choice they made for their setup.

Taro: It sounds like they tuned these parameters quite a bit to find the sweet spot for performance on those specific tasks, but I wonder what happens when you move away from those specific image or CT reconstruction scenarios into something totally different, say, simulating fluid dynamics?

Rosa: The paper focuses heavily on image and signal representation experiments, so the immediate implication is that for those structured signals with smoothness or periodicity, this approach is significantly better than a standard MLP. It’s a solid method for those domains.

Dev: And the training efficiency analysis shows they hit really fast wall-clock performance across all tasks. For instance, in image denoising, FUTON reaches twenty seven decibels in under twenty seconds, while WIRE takes around sixty seconds. That’s a three times speedup for that specific task.

Taro: A three times speedup is significant when you're running large simulations or real-time analysis on embedded systems. It suggests the decomposition helps decouple learning the broad spectral structure from fine local details, which makes gradient optimization more efficient.

Rosa: That decoupling is key; they suggest this CP decomposition separates learning the global spectral structure from local refinements, which helps with how fast it can be optimized via gradients. It’s a clever way to manage complexity during training.

Dev: So, the main thing here is that FUTON offers a structured INR that unifies generalized Fourier series with a low-rank CP decomposition of the coefficient tensor. This design seems specifically aimed at overcoming the common limitations we see with MLP-based INRs, like slow convergence and poor extrapolation.

Taro: The theoretical guarantees about L two functions are solid, but I'm still looking at those practical limits <ref:2602.13414#pg1>. What are they saying about where this method might break down if the signal doesn't conform to that generalized Fourier structure?

Rosa: They did state a limitation, and it’s tied to how well the input signal fits that specific Fourier decomposition; they say the method relies on signals exhibiting some degree of smoothness or periodicity. If the signal is totally chaotic without any underlying structure, performance might drop significantly.

Dev: So it’s not a magic fix for every single possible signal representation, but it’s very effective for the structured data we usually deal with in imaging and reconstruction problems. The paper FUTON: Fourier Tensor Networks for Implicit Neural Representations really offers a concrete alternative to standard MLPs in that space.

Taro: It means if we need high fidelity reconstruction on images or CT scans, this architecture is definitely worth looking at because the speed and generalization claims are backed by those experiments.

Rosa: Yeah, it seems like a promising direction for structured signal processing using these tensor network ideas. That’s what we’ve got from this paper.

Conclusion: Rosa: So FUTON is this new way of building neural networks for signals, using Fourier series and low-rank tensor math instead of just standard multi-layer perceptrons.

Dev: Yeah, that's right. The paper by Rosa and Dev says they’re tackling the slow convergence and overfitting problems that plague these standard implicit representations.

Taro: They use this structure to capture both smoothness from the Fourier part and the low-dimensional nature of the data with the tensor decomposition part.

Rosa: Basically, they're trying to build a representation that respects how natural signals are structured in the frequency domain.

Dev: It’s about using those complementary biases—smoothness and low rank—to make the network learn much faster and generalize better to new data.

Taro: And theoretically, they prove it can represent any function in L two with arbitrary precision, which is a big theoretical statement.

Rosa: But what matters for me is that it runs linearly with respect to the input and output dimensions, which means it scales well for really high-dimensional signals.

Dev: That linear scaling through tensor contraction is key to making it practical, because we need efficient loops for real-time applications.

Taro: So they’re promising a way to get high fidelity reconstruction on things like images and CT scans without the massive computational cost of some other methods.

Rosa: It seems like this structured approach could really open up new avenues for how we use neural networks in signal processing, especially where structure matters a lot.

Ghent University · Vrije Universiteit Brussel

eess.IV, cs.CV, cs.LG, eess.SP

Submitted: 2026-02-13

Updated: 2026-11-10

Comments: Accepted at NeurIPS 2026. 41 pages, 18 figures, 14 tables. v2: the NeurIPS 2026 camera-ready version, with the learning-rate schedule of the radiance fields corrected (Sec. 5, App. G). Project page: https://pashtari.github.io/inro/

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, overfitting to noise…

Key concepts

Implicit Neural Representations (INRs)
These are neural networks used to encode signals or functions implicitly rather than explicitly through a fixed output layer. FUTON is an advanced type of INR that uses structured representations, specifically Fourier series, to model these signals effectively.
Low-Rank Tensor Decomposition
This technique breaks down the large set of coefficients representing the signal into a smaller set of core components. This enforces the idea that natural signals have a low-dimensional spectral structure, allowing the network to learn efficiently and generalize better.
Fourier Bases
Fourier bases are mathematical functions (like sines and cosines) used to represent any periodic or smooth signal as a sum of these simple waves. They allow the model to capture essential properties like smoothness and periodicity inherent in many real-world signals.
Linear Complexity
This refers to how fast the network can process input and generate output. FUTON is designed with linear complexity, meaning its computational cost scales linearly with the size of the input and output dimensions, making it highly scalable for high-dimensional data.

Terminology

Summary

Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, overfitting to noise, and poor extrapolation FUTON implicitly expresses signals as weighted combinations of orthonormal, separable basis functions, combining complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure

Novel Architecture

FUTON introduces a novel architecture that models signals as generalized Fourier series whose coefficients are parameterized via a low-rank tensor decomposition This design embeds complementary inductive biases: Fourier bases capture smoothness and periodicity, while low-rank factorization captures the intrinsic low-dimensional spectral structure of natural signals The architecture leverages the insight that structured signals exhibit frequency-domain correlations amenable to low-rank approximation

Theoretical Foundations and Efficiency

The paper provides theoretical guarantees through a universal approximation theorem establishing that FUTON can represent any L2 function to arbitrary precision Furthermore, it establishes FUTON’s linear complexity through efficient tensor contraction, enabling scalability to high-dimensional signals The forward pass proceeds in three stages with a total complexity of O(CKR + CR + DR), which scales linearly with the input and output dimensions C and D

Superior Performance Across Tasks

Extensive experiments demonstrate that FUTON achieves higher reconstruction quality, faster convergence, and better generalization on inverse problems compared to state-of-the-art methods For signal representation, FUTON substantially outperforms state-of-the-art MLP-based INRs such as WIRE while converging significantly faster In inverse problems, it generalizes better and converges faster on image super-resolution and CT reconstruction

Ablation Studies

Ablation studies investigate the effect of key design choices, including the CP rank R, number of basis functions K, and the one-dimensional basis φ(·) Increasing R improves PSNR with diminishing returns beyond R = 512 for image representation Increasing K to K = 512 monotonically improves reconstruction quality across all ranks The cosine basis is favored because it achieves higher PSNR and faster convergence compared to polynomial bases

Training Efficiency Analysis

FUTON demonstrates superior wall-clock performance across all tasks, reaching 30 dB in approximately 10s for super-resolution compared to over 25s for WIRE In image denoising, FUTON achieves 27 dB in under 20s while WIRE requires approximately 60s—a 3× speedup The CP decomposition decouples learning of global spectral structure from local refinements, enabling efficient gradient-based optimization

Conclusion

FUTON presents a structured INR that unifies generalized Fourier series with a low-rank CP decomposition of the coefficient tensor This design addresses common limitations of MLP-based INRs (e.g., slow convergence, overfitting, and poor extrapolation), while offering a linear-time forward pass and universal approximation in L2 Extensive experiments across diverse image and signal tasks validate the approach Promising directions include exploring alternative orthonormal bases (e. p. 7

Where the paper says it

This design addresses common limitations of MLP-based INRs (e. , slow convergence, overfitting, and poor extrapolation), while offering a linear-time forward pass and universal approximation in L 2 .

g.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Fast training for implicit neural representations by replacing dominant MLP-based designs with Fourier Tensor Network (FUTON). FUTON can train 2 to 5 times faster than state-of-the-art MLP-based INRs on image and volume representation tasks.

  2. Improved generalization in inverse problems by using the intrinsic prior of low-rank spectral parameterization. FUTON generalizes better and converges faster on inverse problems such as image denoising and super-resolution compared to state-of-the-art methods.

  3. High efficiency for high dimensional signals through linear time complexity. FUTON achieves linear time complexity O(CKR + CR + DR) with respect to input and output dimensions C and D, enabling scalability to high dimensional signals without exponential scaling of computation.

  4. Enhanced reconstruction quality in image representation by leveraging the cosine basis function. The cosine basis achieves 37.12 dB PSNR and 96.53% SSIM, significantly outperforming other orthonormal bases like Legendre or Chebyshev polynomials on the Kodak dataset.

  5. Better convergence behavior during training via favorable optimization landscape properties of the CP decomposition structure. The CP decomposition decouples learning of global spectral structure from local refinements, enabling efficient gradient-based optimization compared to MLP-based INRs which require coordinated layer-wise updates.

Abstract

Implicit neural representations (INRs) have emerged as powerful tools for encoding signals, yet dominant MLP-based designs often suffer from slow convergence, and how their accuracy depends on their size is known only empirically. We introduce FUTON (Fourier Tensor Network), which models signals as generalized Fourier series whose coefficients are parameterized by a low-rank tensor decomposition, read out by a linear or MLP head. FUTON combines complementary inductive biases: Fourier bases capture smoothness and periodicity, while the low-rank parameterization enforces low-dimensional spectral structure. Beyond universal approximation, we show that with a linear head, the best FUTON's error splits exactly into the truncation of the signal's series and the best low-rank error of its coefficients, which we bracket from the samples before training, with a FUTON that attains the upper bound. On image fitting, occupancy volumes, and novel view synthesis, against twelve INRs of matched size, FUTON achieves the best average quality, reaches the final quality of each task's strongest baseline 4-34x sooner than that baseline, infers fastest, and trains in the least memory.

Sources

Related papers