⚛️ quantum physics

Spectral Quantum Memory for Implicit Neural Representations

This paper introduces the Spectral Quantum Memory Model (SQMM), a novel Quantum Implicit Neural Representation architecture that utilizes a shared coherent quantum memory register to enable sequential hybrid data reuploading layers to build upon each other's spectral computations, thereby improving reconstruction performance on audio and image benchmarks while providing a theoretical framework for spectral capacity allocation.

Original author: Daniel A. Moroz-Broitman, Eliya Nachmani

Published 2026-11-04

📖 4 min read🧠 Deep dive

CC BY 4.0

Original author: Daniel A. Moroz-Broitman, Eliya Nachmani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the author. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to describe a complex sound or a detailed photograph using only a mathematical formula. For decades, scientists have used a technique called implicit neural representation to do exactly this, treating signals not as a grid of pixels or a list of samples, but as a continuous function that can be queried at any point. The challenge has always been that these mathematical models struggle to capture both the broad, smooth shapes of a scene and the tiny, sharp details simultaneously. They tend to learn the easy, low-frequency parts first and often miss the high-frequency details, like the crisp edge of a building or the subtle crackle in a voice. To fix this, researchers have turned to quantum computing, hoping that the unique properties of quantum bits could naturally handle these different frequencies. However, early attempts at quantum models treated each frequency band as a separate, isolated worker, with no way for one part of the system to talk to another about what it had already learned.

A team of researchers at Ben Gurion University has introduced a new approach called the Spectral Quantum Memory Model, which changes how these quantum systems communicate. Instead of letting each part of the model work in isolation, they built a shared quantum memory that acts as a bridge between the different frequency bands. Think of it like a relay race where the runners don't just pass a baton; they also share a notebook of notes about the track they just ran, allowing the next runner to adjust their strategy based on what came before. In this new system, the first part of the model analyzes the signal and writes a summary of its findings into a shared quantum memory. The next part reads that summary, adds its own analysis, and writes an updated version back. This process repeats, with each stage building upon the spectral information gathered by the previous ones, creating a coherent chain of understanding rather than a collection of disjointed guesses.

The researchers proved mathematically that this method does not create random, chaotic frequencies. Instead, the information flowing through the memory remains organized, with the final output being a precise combination of the frequencies introduced at each step. To ensure the memory was actually doing its job, they added a special training rule that encouraged the memory to accurately reflect the spectrum of the signal it had processed so far. When they tested this model on reconstructing audio, such as a cello recording, and images, ranging from simple cartoons to complex photographs, the results were striking. In simulations, the new model consistently produced clearer, more accurate reconstructions than both standard classical models and previous quantum attempts, all while using a similarly small number of parameters.

The work suggests that the key to making quantum models effective for these tasks is not just in the individual quantum circuits, but in how they share information. By keeping a coherent quantum state alive between the different stages of the model, the system can condition its later calculations on the spectral content of earlier ones. The researchers ran these experiments on powerful computer simulators that mimic quantum hardware, as they did not yet run the model on actual quantum devices. While the simulations showed significant improvements in accuracy for both sound and image reconstruction, the team acknowledges that real-world quantum hardware faces challenges like noise and the high cost of measurement. Nevertheless, the study provides a clear path forward, demonstrating that a shared quantum memory can organize complex spectral information in a way that isolated quantum blocks cannot, offering a promising new direction for compact and efficient quantum representations of the world around us.

Technical Summary: Spectral Quantum Memory for Implicit Neural Representations

Problem Statement
Implicit Neural Representations (INRs) model signals as continuous functions mapping coordinates to values, enabling compact representations for images, audio, and physical fields. However, their performance is bottlenecked by "spectral bias," where standard networks learn low-frequency structures before high-frequency details. While various classical and Quantum INR (QINR) architectures (e.g., Fourier features, periodic activations, data reuploading circuits) have been proposed to manage spectral capacity, existing QINR models typically treat spectral components as independent outputs of separate quantum blocks. In these architectures, information between frequency modules (bands) is passed only through classical decoders, preventing earlier bands from influencing the spectral computation of later bands in a coherent manner.

Methodology: The Spectral Quantum Memory Model (SQMM)
The authors introduce SQMM, a memory-coupled QINR architecture designed to propagate coherent spectral context across sequential data reuploading bands.

Architecture: SQMM consists of $n$ $n$ sequential hybrid quantum-classical bands. Unlike previous models where bands operate independently, SQMM utilizes a shared, unmeasured quantum memory register ( $M$ $M$ ) with $m$ $m$ qubits that is carried from band to band.
- Data Flow: For a coordinate $x$ , the memory is initially prepared via a one-time upload $\gamma_M(x)$ . Each band $i$ reads the current memory state $\sigma^{(i-1)}_M(x)$ , processes its own coordinate upload $\gamma_i(x)$ through a variational quantum circuit, and writes a portion of its local state back into the memory register before discarding its local data qubits.
- Readout: Each band emits Pauli expectation values as classical features, which are concatenated and passed to a classical decoder. Crucially, the memory state itself is never measured during inference; it serves solely as a coherent carrier of spectral information between bands.
Spectral Theory: The paper provides a theoretical analysis proving that SQMM preserves the finite Fourier structure inherent to data reuploading circuits.
- Operator-Valued Fourier Series: The memory state is shown to form an operator-valued Fourier series. The support of the frequencies in the memory state after band $i$ is contained within the Minkowski sum of the initial memory spectrum and the local spectra of all preceding bands ( $\Omega^{(i)}_{M,x} \subseteq \Omega^{(0)}_{M,x} + \sum_{j=1}^i \Omega^{loc}_{j,x}$ ).
- No New Primitive Frequencies: The memory coupling does not introduce arbitrary new primitive frequencies. Instead, it allows downstream bands to condition their Fourier coefficients on the spectral content (operator-valued summary) stored by earlier bands.
Training Objective: To ensure the memory effectively summarizes the computation, the authors propose a Spectral Memory Alignment Loss. This auxiliary loss compares the estimated spectrum of the memory state against the accumulated spectrum of the features emitted by the preceding bands (the "prefix"). A stop-gradient is applied to the prefix features, training the memory to align with the spectral occupancy of the deployed model prefix.

Key Contributions

Architecture: Introduction of SQMM, the first memory-coupled QINR that propagates coherent spectral context through a shared quantum register across sequential data reuploading bands.
Theoretical Proof: Demonstration that input-independent memory read/write operations preserve the finite Fourier structure of data reuploading circuits. The memory state remains an operator-valued Fourier series whose support is strictly bounded by the Minkowski sum of band spectra.
Training Mechanism: Development of a spectral memory alignment loss that trains the memory to act as a readable spectral summary of the model's prefix, facilitating better coefficient allocation in downstream bands.
Empirical Performance: Demonstration that SQMM achieves superior reconstruction results compared to compact classical INR baselines (e.g., SIREN, FINER, MSA-INR) and existing QINR baselines (QIREN, QVF) on audio and image reconstruction tasks under strict parameter budgets.

Experimental Results
The authors evaluated SQMM on audio and image reconstruction benchmarks using a mixed-state simulator (PennyLane/PyTorch).

Audio: On single-signal reconstruction (Bach cello excerpt) and dataset-level reconstruction (GTZAN, UrbanSound8K, VCTK), SQMM achieved the lowest Mean Squared Error (MSE) and Log Spectral Distance (LSD) among all compared models, including strong adaptive classical baselines like FINER.
Images: On grayscale CIFAR10 and larger images (Kodak, DIV2K), SQMM outperformed classical and quantum baselines in MSE, PSNR, SSIM, and LPIPS. Notably, it achieved the best results with fewer parameters than competitors like NeRF (which used 1.54x more parameters).
Ablation Studies: Mechanism ablations confirmed that the shared memory significantly improves reconstruction over a "no memory" control (where bands have local qubits but no shared state). Furthermore, the spectral alignment loss ( $L_{align}$ ) yielded better results than simply supervising the final output spectrum, validating the utility of aligning the internal memory state with the prefix spectrum.

Significance and Claims
The paper claims that SQMM establishes spectral memory as a practical design principle for compact Quantum INRs. By treating the memory as a trainable spectral interface rather than an opaque hidden state, the model allows later bands to reuse and reweight the spectral content of earlier bands without relearning them independently.

The authors position this work as a method to organize spectral computation in quantum implicit representations, potentially extending to higher-dimensional signals, time-varying representations, and physical fields. They explicitly state that the work does not claim a quantum computational or hardware advantage; rather, it demonstrates that coherent spectral memory is a useful mechanism for improving the representational capacity of compact quantum models, with simulation cost and hardware measurement overhead identified as the primary current limitations.

Technical Summary: Spectral Quantum Memory for Implicit Neural Representations

More like this