⚛️ quantum physics

Few-sample regression with an adaptively grown variational quantum Kolmogorov--Arnold network

This study provides a rigorous, reproducible evaluation of an adaptively grown variational quantum Kolmogorov-Arnold network, demonstrating that while it offers implicit regularization benefits over classical and quantum baselines in extreme few-sample regimes, it lacks a general expressivity advantage and is outperformed by classical methods on larger datasets.

Original authors: Hikaru Wakaura, Rahmat Mulyawan, Andriyan B. Suksmono

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hikaru Wakaura, Rahmat Mulyawan, Andriyan B. Suksmono

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 authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the quiet corners of modern physics and computer science, researchers are constantly searching for ways to make sense of the world when data is scarce. Imagine a scientist trying to predict the weather or the behavior of a new material, but they only have a handful of expensive measurements to work with. In this "few-sample" regime, the sheer power of a computer model matters less than its built-in intuition, or what experts call its inductive bias. This is the specific problem that Kolmogorov–Arnold networks were designed to solve. Unlike standard neural networks that learn by adjusting fixed switches at their nodes, these networks learn by shaping flexible, one-dimensional curves along their connecting edges. This structure makes the model's logic easier to interpret and, in theory, better suited for learning from very little data. Recently, scientists have tried to build these networks using the strange rules of quantum mechanics, hoping that the unique properties of quantum particles would give them an edge over classical computers. The big question remains: do these quantum versions actually offer a practical benefit, or are they just complex ways of doing what classical computers already do well?

A team of researchers set out to answer this question with a rigorous, no-nonsense approach, evaluating a new type of quantum model called an adaptively grown variational quantum Kolmogorov–Arnold network. Instead of guessing which quantum settings might work best, they built a system that grows its own structure, adding one quantum operator at a time only if it improves the model's performance. To ensure their results were trustworthy, they designed a study that avoided common pitfalls: they compared their model against others using the exact same random starting points, they never let the model peek at the test data during training, and they locked in their analysis plan before running a single experiment. They tested this quantum model on a series of mathematical challenges, ranging from simple four-variable problems to more complex scenarios with up to eighteen dimensions, using only ten training points for each task.

The results painted a clear and somewhat humbling picture. When the researchers tested the model on a small, four-qubit system, it performed no better than a standard quantum neural network of the same size and was significantly outperformed by simple classical computer models. However, the story changed when they moved to a more difficult, high-dimensional challenge where the model had to learn a complex pattern from just ten data points. In this specific "few-sample" regime, the quantum model did indeed beat the best unregularized classical models and a tuned quantum neural network. It managed to generalize well, making accurate predictions on new data where the classical competitors failed. Yet, this victory was not due to some mysterious quantum power. When the researchers compared the quantum model to a classical method that uses a specific type of smoothing technique called kernel ridge regression, the two performed almost identically. The quantum model's success came not from being more expressive or powerful, but from being naturally constrained; its small size and specific structure acted as a built-in filter that prevented it from overfitting the tiny dataset.

As the researchers increased the amount of data available, the quantum model's advantage vanished. When they doubled the training points from ten to twenty, the classical models caught up and surpassed the quantum one. Similarly, when they increased the complexity of the problem to eighteen dimensions, the quantum model's performance dropped to the level of a simple guess, while a well-tuned classical model continued to improve. This confirmed that the quantum model's benefit was limited to a very narrow window where data is extremely scarce and the model's capacity is deliberately kept low. The study also tested the model's resilience against real-world imperfections. They simulated the noise found in actual quantum hardware and ran the trained circuits on a real 156-qubit quantum processor from IBM. The model held up remarkably well, with its performance on the physical machine differing from the ideal simulation by less than a fraction of a percent. This proved that the model is robust enough to run on current hardware, even with the noise and measurement errors inherent in today's quantum devices.

Ultimately, this research provides a reproducible reference point for what these quantum networks can and cannot do. It shows that the adaptive quantum Kolmogorov–Arnold network is not a magic bullet that solves all learning problems, nor does it possess a fundamental quantum advantage in expressiveness. Instead, it functions as a highly effective, low-capacity tool that offers a form of implicit regularization, making it useful only when data is extremely limited. The study concludes that for these specific tasks, a well-chosen classical method can achieve the same results. The value of this work lies in its clarity: by stripping away the hype and using a strict, pre-registered protocol, the authors have shown that the path forward for quantum machine learning is not about finding bigger models, but about understanding exactly where and why these specific quantum structures might offer a unique, albeit limited, advantage.

Technical Summary: Few-sample regression with an adaptively grown variational quantum Kolmogorov–Arnold network

Problem Statement
Machine learning is increasingly applied to interpolate physical quantities from a limited number of expensive measurements or simulations. In this "few-sample" regime, a model's inductive bias is often more critical than its raw capacity. Kolmogorov–Arnold Networks (KANs) offer a specific bias by placing learnable one-dimensional functions on network edges rather than fixed activations on nodes. While several quantum realizations of KANs have been proposed, their practical utility remains unclear. Previous comparisons suffer from methodological flaws: they rely on single training runs, use untuned baselines (both quantum and classical), and often utilize test sets for model selection, introducing data leakage. This study aims to rigorously evaluate an Adaptive Variational Quantum KAN (Adaptive VQKAN) under a protocol designed to eliminate these confounds.

Methodology
The authors propose and evaluate an Adaptive VQKAN, a variational quantum algorithm where the ansatz is grown operator-by-operator.

Architecture: The model maps an input $x \in [0, 1]^d$ to a scalar prediction via the expectation value of a Hamiltonian $\hat{H}$ on $N_q = \max(4, d)$ qubits. Input encoding uses rotation gates $R_y(\arccos x_j + \pi/2)$ . A KAN layer consists of a list of Pauli strings, where each string carries a learnable edge function realized by a rotation angle dependent on measured single-qubit expectations. The edge functions are quadratic B-splines with 8 learnable control values.
Adaptive Growth: Training begins with a single Pauli operator ( $X_1$ ). Every $T_g$ optimization steps, operators from a pool (all one-body and two-body Pauli strings) are tentatively appended with zeroed control values. The operator that most reduces the training cost is permanently adopted. This mimics the ADAPT-VQE selection rule but uses cost minimization instead of energy minimization.
Experimental Protocol: To ensure reproducibility and fairness, the study employs:
- Seed-paired comparisons: All methods are tested on identical random seeds for training and test sets.
- Strict model selection: Model selection (including hyperparameter tuning for baselines) relies solely on training cost; the test set is never used for selection.
- Pre-registration: A confirmatory study on 12-, 16-, and 18-dimensional targets was pre-registered with fixed hypotheses, seeds, and analysis scripts before execution.
- Baselines: Comparisons include a hardware-efficient Quantum Neural Network (QNN) (both untuned and tuned), classical KANs, multilayer perceptrons (MLPs), linear regression, and regularized classical regressors (Ridge, Kernel Ridge with Leave-One-Out Cross-Validation, Gaussian Processes).

Key Results
The evaluation was conducted on four-qubit benchmarks and a difficulty-controlled family of targets in 12, 16, and 18 dimensions, using only 10 training points.

Four-Qubit Benchmarks: On synthetic 4-qubit targets, the Adaptive VQKAN (24 parameters) is statistically indistinguishable from a QNN of the same parameter count. Both are significantly outperformed by classical regressors (linear regression and MLPs). No advantage for the KAN structure was observed at this scale.
Scaling Study (12 and 16 Dimensions): With 10 training points, the 32-parameter Adaptive VQKAN significantly outperforms unregularized classical regressors and a tuned QNN (Holm-corrected $p \le 0.017$ $p \leq 0.017$ ).
- At $d=12$ , the VQKAN error is 11.1 vs. 13.0 for the best unregularized classical baseline.
- At $d=16$ , the VQKAN error is 11.7 vs. 15.0.
- However, a Kernel Ridge Regressor with hyperparameters chosen via Leave-One-Out Cross-Validation (LOO-CV) on the same 10 points matches the VQKAN's performance (e.g., 11.2 vs. 11.1 at $d=12$ ).
Scaling Limits (18 Dimensions): At $d=18$ , the quantum model's advantage vanishes; its performance becomes indistinguishable from the classical oracle and the constant predictor.
Robustness: The model's performance is robust to finite measurement shots (256 shots yield results comparable to exact simulation) and gate noise. The trained circuits were successfully executed on the 156-qubit IBM Heron processor ( $ibm\_marrakesh$ ) with test errors within 0.3 of exact values.
Sample Size Sensitivity: As the number of training points increases (from 10 to 100), the error of the quantum model remains nearly flat (saturating around 9.5), while classical models (especially the cross-validated kernel ridge) continue to improve significantly.

Mechanism and Interpretation
The authors conclude that the observed benefit of the Adaptive VQKAN in the few-sample regime is not a quantum expressivity advantage. Instead, it arises from the implicit regularization of a deliberately low-capacity model.

The adaptive growth rule selects a small number of operators (typically 4 two-body strings), resulting in a fixed parameter count (32) that cannot overfit the 10 training points in high dimensions.
Unregularized classical models overfit these few points, leading to poor generalization.
The quantum model's performance is effectively replicated by a classical kernel method with appropriate regularization (LOO-CV), suggesting the quantum feature space is not richer than accessible classical kernels but is simply more constrained.

Significance and Claims
The paper claims to provide a reproducible reference point for the resources and limits of variational quantum KANs.

It demonstrates that previous claims of quantum advantage in KANs may have been artifacts of poor experimental design (untuned baselines, single runs, data leakage).
It establishes that for the specific targets studied, the quantum model offers no advantage over a well-regularized classical kernel method, nor does it possess superior expressivity (as evidenced by the error saturation at higher sample sizes).
The study highlights that the "few-sample" benefit is a capacity effect, present only when the model is too small to overfit, and disappears as dimensionality increases or sample size grows.
The authors emphasize that the protocol and code are released to allow future testing of whether different encodings or target families (specifically those theoretically suited to KAN structures) might yield genuine quantum advantages.

Technical Summary: Few-sample regression with an adaptively grown variational quantum Kolmogorov–Arnold network

More like this