Tensor network methods for quantum-inspired image processing and classical optics

This paper explores the application of quantum-inspired tensor network methods to enhance the efficiency of classical image processing and optical simulations, such as wave-front propagation and image formation.

The Gist

Imagine you are trying to describe a massive, incredibly detailed landscape—like the Grand Canyon—to a friend over a very slow, old-fashioned telephone line. You can’t send a high-definition photo; you don't have the "bandwidth." If you try to describe every single pebble and grain of sand, the call will drop before you finish.

This paper, written by Nicolas Allegra, explores a brilliant way to "cheat" this limitation. Instead of describing every pebble, he uses a mathematical trick borrowed from the world of Quantum Physics to describe the patterns of the landscape. This allows him to compress massive amounts of information into tiny, efficient packages without losing the "soul" of the image.

Here is the breakdown of how this works, using everyday analogies.

---

1. The Problem: The "Data Explosion"

In the digital world, a high-resolution image is like a giant spreadsheet where every single pixel has its own number. As images get bigger, the amount of memory needed doesn't just grow; it explodes. If you double the width and height of a photo, you don't just double the data—you quadruple it. For scientists studying stars or microscopic cells, this "data explosion" makes simulations painfully slow.

2. The Solution: "Tensor Networks" (The Art of Summarizing)

The author uses something called Tensor Networks. Think of this as the ultimate "Smart Summary" tool.

Imagine you are reading a 1,000-page novel.

• Standard Computing is like trying to memorize every single word in the book.
• Tensor Networks are like reading the book and realizing that most of the chapters follow the same themes. Instead of memorizing every word, you memorize the plot points and the character arcs.

By focusing on the "correlations" (how one part of the image relates to another), you can represent a massive image using only a tiny fraction of the original data.

3. The "Quantum Inspiration": The Ghost in the Machine

The most exciting part of the paper is that these methods are "Quantum-Inspired."

In Quantum Mechanics, particles aren't just in one spot; they exist in a web of relationships and probabilities. The author realized that natural images (like a photo of a forest) behave a lot like quantum systems. They have "scales"—you have big shapes (the trees) and tiny details (the leaves).

He uses two specific "quantum" ways to organize this data:

• The Hilbert Curve (The Efficient Path): Imagine a massive city. If you want to map it, you could just list addresses randomly. Or, you could use a "space-filling curve"—a single, continuous line that snakes through every single house in a way that neighbors stay close to each other on the map. This keeps the "local" information together, making it much easier to compress.
• The Tree Structure (The Family Tree): Instead of a long line of data, he organizes the image like a family tree. The "grandparents" are the big, blurry shapes of the image, and the "grandchildren" are the tiny, sharp details. This allows the computer to handle the big picture first and only worry about the tiny details when absolutely necessary.

4. Real-World Magic: Fast-Forwarding Light

The paper doesn't just stop at saving space; it uses these tricks to speed up physics simulations.

Specifically, he looks at Classical Optics (how light travels through lenses and air). Normally, calculating how light waves bounce around a complex lens is a mathematical nightmare that takes a long time.

The author shows that we can treat a beam of light like a "quantum wave." By using his "Smart Summary" (Tensor Networks), he can simulate how light moves through an optical system much faster than traditional methods. It’s like the difference between calculating the path of every single raindrop in a storm versus simply calculating the flow of the river.

Summary: Why does this matter?

If we can master these "Quantum-Inspired" summaries, we can:

• Astronomy: Process massive images from space telescopes much faster.
• Microscopy: See deeper into cells without needing supercomputers the size of a building.
• Imaging: Create faster, smarter cameras and sensors for Earth observation.

In short: He is teaching computers how to see the "big picture" so they don't get lost in the "tiny details."

Technical Summary

Technical Summary: Tensor Network Methods for Quantum-Inspired Image Processing and Classical Optics

Author: Nicolas Allegra (Thales Alenia Space)
Date: February 10, 2026

---

1. Problem Statement

The paper addresses the computational bottleneck in simulating complex classical systems, specifically in high-resolution image processing and wave-front propagation (classical optics). Traditional methods, such as the Fast Fourier Transform (FFT), scale linearly or super-linearly with the number of pixels ($L^2$), making very large-scale simulations (e.g., in astronomy or microscopy) memory-intensive and slow. The author seeks to leverage the "middle ground" of tensor networks (TN)—methods inspired by quantum mechanics that provide massive compression for systems with specific correlation structures—to achieve faster, more efficient classical computations.

2. Methodology

The research follows a three-step pipeline: Encoding $\rightarrow$ Compression $\rightarrow$ Operation.

A. Quantum-Inspired Encodings:
To apply tensor networks, 2D classical images must be mapped into higher-dimensional tensor structures. The paper evaluates two primary methods:

• FRQI (Flexible Representation of Quantum Images): A scheme that factorizes pixel intensity and position into a multi-qubit state using space-filling curves (e.g., Hilbert curves) to preserve spatial locality.
• Quantics Representation: A multi-scale encoding that factorizes spatial indices into binary components. This naturally encodes scale separation, where coarse features reside in high-order qubits and fine details in low-order qubits.

B. Tensor Network Topologies:
The author compares two main compression architectures:

• Matrix Product States (MPS) / Tensor Trains (TT): A linear chain of tensors.
• Tensor Tree Networks (TTN): A hierarchical, acyclic tree structure.

C. Optical Mapping:
The author draws a formal mathematical parallel between Fourier Optics and Quantum Mechanics. Specifically, the paraxial Helmholtz equation (governing light propagation) is shown to be formally equivalent to the Schrödinger equation for a free quantum particle. This allows classical optical propagators (like the Fresnel or Angular Spectrum transfer functions) to be represented as Matrix Product Operators (MPO).

3. Key Contributions

• Scale-Invariant Compression Analysis: The paper demonstrates that natural images follow a power-law distribution, allowing them to be compressed using renormalization group (RG) principles.
• Development of the "qTCI" Algorithm: To avoid the exponential cost of standard Singular Value Decomposition (SVD) on large images, the author utilizes Quantics Tensor Cross-Interpolation (qTCI), a sampling-based approach that constructs the tensor network with only a linear number of calls.
• MPO Construction for Wave Optics: The paper provides a systematic way to construct low-rank MPOs for:
  1. Fresnel Propagation: Using a Bessel function expansion (Jacobi-Anger expansion) to represent the unitary evolution of light.
  2. Relativistic (Non-paraxial) Propagation: Using a Dirac-inspired Hamiltonian to represent the Angular Spectrum Transfer Function.

4. Key Results

• Compression Efficiency:
  • MPS is effective for images following an "area law" of entanglement, but its performance is limited by its linear topology.
  • TTN significantly outperforms MPS in the "intermediate regime" (noisy or high-precision data) because its hierarchical structure allows it to push noise/entropy to the top of the tree, preserving local features more efficiently.
• Computational Speed-up:
  • For Point Spread Function (PSF) calculations and image formation (convolution), the tensor network formulation (MPS/MPO) exhibits constant run-time scaling (with respect to system size $n$) for a fixed approximation error, whereas the FFT scales linearly with the number of pixels.
  • The error in the Bessel-based MPO construction decays super-exponentially, allowing for machine-precision accuracy with very low bond dimensions.
• Complexity Scaling: The complexity of contracting an image with an optical operator scales as $O(2^n \chi^2_1 \chi^2_2)$, where $\chi$ is the bond dimension. For low-rank operators, this is highly efficient.

5. Significance

This work establishes a robust framework for "Quantum-Inspired Classical Computing." By treating images as quantum-like states and optical propagation as unitary evolution, the author provides a path to simulate massive optical systems that are currently computationally prohibitive.

Potential Applications:

• Astronomy & Earth Observation: Processing ultra-high-resolution satellite or telescope imagery.
• Microscopy: Efficiently modeling complex wave-front aberrations and light-matter interactions.
• Advanced Imaging: Faster implementation of deconvolution, denoising, and phase retrieval algorithms.