This Quantum Network Can Represent More Than Its Classical Twin. It Doesn't Beat It Yet.

Quantum computing makes two kinds of claim, and they are not the same size. The loud one is that a quantum machine will someday finish a calculation a classical one never could. The quiet one is about room: that at a given size, a quantum circuit can hold more. On Sept. 3 a group at Shanghai Jiao Tong University published a proof of the quiet claim for one particular machine-learning model, and spent the rest of the paper making sure it would not be read as the loud one.

The model belongs to a family called neural operators. Train an ordinary network on a differential equation and it learns one solution; a neural operator learns the map itself, so a new starting condition gets a new answer without retraining. Classical versions, DeepONet and the Fourier neural operator among them, have become working tools for problems like fluid flow and diffusion. Quantum versions have lagged behind, mostly because the obvious designs want either more qubits than anyone has or circuits too deep to survive the noise in the machines that exist.

Ruocheng Wang and Xiaoqiu Zhong share first authorship on the paper; Junchi Yan, of the university's School of Artificial Intelligence, is the senior author. Their argument is that the circuit gets its capacity from somewhere other than the usual place. The standard case for quantum machine learning points at the exponentially large space a register of qubits can in principle occupy. These authors prove something narrower and more concrete: that the way their circuit represents data implicitly builds a quadratic feature frame, a structured set of building blocks that is quadratic rather than linear in the circuit's own state.

What they prove from it is a bound. For operators, the model's capacity grows as p², where p is its latent dimension, against O(p) for a matched classical model. p is not a parameter count. It is set by the qubit register: the public code defines it as p = 2ⁿ for n qubits, so five qubits give p = 32 and each further qubit doubles it. The bound is about what the model can represent, not how fast it runs or how accurately it lands, and the authors do not claim otherwise.

Whether that room turns into performance is a separate question, and the paper's own answer is modest. The trainable-frequency version of the model, TF-QuanONet, "notably outperforms quantum baselines" and reaches "competitive accuracy against classical frameworks under strictly matched-parameter conditions," in the abstract's wording. Matched parameters is a different kind of fairness from matched latent dimension: that comparison holds the number of trainable weights equal on both sides. Against the classical models it is a tie, not a win.

The one place the paper claims more than parity is stability. Pushed to a latent dimension of 256, the model converges to what the authors call the intrinsic error floor, while classical baselines at that setting swing widely from run to run. That is a claim about being easier to train at size, not about being more accurate at its best.

The architecture itself is not new. QuanONet, the trainable-frequency variant and the first runs on real hardware were published by the same group at ICML 2025. What is new is the theory underneath (the implicit frame and the capacity bound) along with the head-to-head against classical frameworks at matched parameters and the scaling study.

The model was deployed on an IBM processor called ibm_fez and kept working under the machine's noise; the authors' own description of the exercise is "a qualitative proof of concept." The acknowledgements reveal the often informal ways quantum experiments come together: the device runs were executed by S. Wei, who is not an author, working in a personal capacity during a visit to Paris.

The result is peer reviewed; Nature names Natansh Mathur as one of the reviewers on the published record. The classical baselines it ties with were implemented and run by the same team, in their own framework, which is the usual arrangement and the usual reason to want the code open.

The code is public under an MIT license, on GitHub and archived on Zenodo, and it is unusually complete: five simulation backends, pretrained weights, and a script named for each experiment in the paper. None of it needs a quantum computer. A pretrained model runs an inference in about a minute on a desktop CPU, while training the quantum model on the simplest benchmark takes roughly 80 minutes by the repository's own estimate, against about five minutes for the classical DeepONet baseline it is compared with. Simulating a quantum circuit is expensive, and none of this is a story about speed.

Expressivity and trainability pull against each other in quantum models: a circuit with more room in it can be harder to optimize, and hardware noise takes back part of what it gains. That says how much room the circuit has, and anyone with the repository and an ordinary CPU can start checking what fits in it.