⚛️ quantum physics

Virtual quantum neural networks

This paper introduces "virtual quantum neural networks," an extended model that utilizes random sampling and classical processing to incorporate linear combinations of completely positive maps, thereby enhancing expressivity and noise robustness for tasks like error mitigation, classification, and energy estimation on near-term quantum hardware.

Original authors: Benchi Zhao, Xuanqiang Zhao, Yinan Li, Yingzhou Li, Giulio Chiribella

Published 2026-09-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Benchi Zhao, Xuanqiang Zhao, Yinan Li, Yingzhou Li, Giulio Chiribella

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 current era of quantum computing, machines are powerful enough to perform calculations that would stump even the most advanced classical supercomputers, yet they remain fragile. These devices, known as noisy intermediate-scale quantum processors, operate with a few hundred tiny units called qubits. While they can probe new realms of physics and chemistry, they suffer from a fundamental flaw: environmental interference causes errors that corrupt the data almost as soon as it is generated. To make these machines useful, scientists rely on a method called quantum neural networks. Think of this as a learning loop where a quantum circuit, adjusted by a set of knobs, processes information and produces an output. A classical computer then checks how close that output is to the desired result and tweaks the knobs to improve the next attempt. This cycle allows researchers to train the machine to solve problems like identifying patterns in data or finding the most stable energy state of a molecule, despite the hardware's imperfections. However, this traditional approach is limited because it can only explore a narrow path of possible calculations, restricted by what the physical machine can actually do in a single run.

A team of researchers has now proposed a way to break through this barrier by changing how the learning process works. Instead of relying on a single quantum circuit to do the heavy lifting, they introduced a new model called a virtual quantum neural network. In this approach, the computer does not just run one circuit; it randomly samples many different circuits, each with its own settings, and runs them on the quantum hardware. The results from these multiple runs are not simply averaged out. Instead, a classical computer takes the data from all these separate experiments and combines them using a specific mathematical recipe. This recipe involves adding and subtracting the results in a way that a single physical circuit could never achieve on its own. By mixing these outcomes, the system creates a "virtual" operation that is far more flexible and powerful than any single circuit could be. It is as if the researchers found a way to simulate a much more complex machine by cleverly stitching together the outputs of several simpler, noisier ones.

The researchers tested this idea on three distinct challenges to see if the virtual approach offered real advantages. First, they tackled the problem of correcting errors caused by unknown noise. In standard quantum computing, fixing errors usually requires knowing exactly what kind of noise is affecting the machine. The new method, however, worked even when the noise was a mystery. By training the virtual network to reverse the effects of the noise, the team found that their system could recover the correct information with near-perfect accuracy, even when the underlying hardware was flawed. The simulations showed that the method could find the best possible way to cancel out errors, performing almost as well as if the noise had been perfectly understood from the start.

Next, the team applied the model to a binary classification task, which is essentially teaching the machine to sort data into two categories, such as distinguishing between two types of shapes. In this test, the virtual network consistently outperformed traditional quantum classifiers. When the data was fed into the system, the virtual network learned to make more accurate predictions, even when the quantum gates used to process the data were noisy. The results demonstrated that by allowing the system to explore a wider range of possibilities through random sampling, the machine became more robust against the errors that typically plague quantum devices.

Finally, the researchers used their method to estimate the ground-state energy of a molecule, a calculation that determines the most stable configuration of atoms and is crucial for understanding chemical reactions. This is a notoriously difficult task for noisy quantum computers because small errors can lead to wildly incorrect energy values. The team introduced a new algorithm based on their virtual network to solve this. Their simulations showed that by combining the outputs of multiple noisy circuits, the system could estimate the molecule's energy with significantly higher precision than standard methods. The virtual network effectively filtered out the noise, allowing the researchers to get much closer to the true physical value than was previously possible with the same hardware.

The core finding of this work is that the limitations of current quantum hardware are not necessarily a dead end. By investing more time in sampling multiple circuits and using classical computers to process the results, scientists can unlock a much broader space of calculations. This approach does not require waiting for perfect, error-free machines; instead, it uses the existing noisy devices more intelligently. The virtual quantum neural network acts as a bridge, turning the weakness of noise into a manageable variable and expanding what these early-stage quantum computers can actually achieve. The study suggests that this flexible learning paradigm could become a standard tool for making near-term quantum hardware more useful for real-world applications in science and engineering.

Technical Summary: Virtual Quantum Neural Networks

Problem Statement
Quantum Neural Networks (QNNs) are a leading model for quantum machine learning (QML) in the Noisy Intermediate-Scale Quantum (NISQ) era. Standard QNNs are trained by minimizing a loss function over a parametrized family of quantum circuits, mathematically described by unitary operators or, more generally, completely positive trace-preserving (CPTP) maps. However, this framework faces a fundamental limitation: the optimization space is constrained to the subset of CPTP maps physically realizable on the available quantum hardware. This restriction limits the model's expressivity, often confining it to shallow circuits, and renders the model sensitive to hardware noise, which degrades performance as circuit depth increases.

Methodology: The Virtual QNN (VQNN) Framework
The authors propose an extended model called the Virtual Quantum Neural Network (VQNN). This framework leverages random sampling and classical post-processing to enlarge the optimization space beyond the set of physically implementable CPTP maps.

Core Mechanism: Instead of optimizing a single quantum circuit, a VQNN randomly samples $N$ parametrized quantum circuits (branches). Let these circuits be represented by CPTP maps $C_i(\theta^{(i)})$ . The VQNN computes a linear combination of the expectation values obtained from these branches, weighted by real coefficients $c_i$ .
Mathematical Formulation: The effective operation performed by the VQNN is a linear combination of the individual maps:
$M(\theta, c) = \sum_{i=1}^N c_i C_i(\theta^{(i)})$
Unlike standard QNNs, $M(\theta, c)$ is not restricted to being a CPTP map. It belongs to the broader class of Hermitian-preserving (HP) maps (also known as virtual quantum operations).
Training Loop:
1. Sampling: The input state $\rho_0$ is processed by one of the $N$ circuits chosen with probability $p_i = |c_i| / \|c\|_1$ .
2. Measurement: The output state is measured to estimate the expectation value of an observable $O$ .
3. Post-processing: The measurement outcomes are classically aggregated to estimate the argument of the loss function, $x(\theta, c) = \sum c_i \text{Tr}[O C_i(\theta^{(i)})(\rho_0)]$ .
4. Optimization: A classical optimizer (e.g., Adam) updates the circuit parameters $\theta$ and the coefficients $c$ to minimize the loss function $L = f(x(\theta, c))$ .
Sampling Overhead: The cost of this extension is a "sampling overhead" proportional to $\|c\|_1 = \sum |c_i|$ . The number of sampling rounds required to estimate the loss function within a certain error bound scales with $\|c\|_1^2$ .

Key Contributions and Applications
The paper demonstrates the efficacy of VQNNs through three representative tasks, showing that the enlarged optimization space leads to enhanced expressivity and noise robustness.

Variational Quantum Error Mitigation (VQEM):
- Goal: Mitigate unknown, uncharacterized quantum noise without prior knowledge of the noise channel.
- Approach: The VQNN is trained to find a linear combination of parametrized channels that approximates the inverse of the noise channel ( $N^{-1}$ ). The loss function minimizes the Hilbert-Schmidt distance between the virtual output state and the ideal target state (e.g., a Bell state).
- Results: Numerical simulations show that VQEM achieves near-optimal mitigation for depolarizing and amplitude damping noise. Even when the variational circuits themselves are noisy, the VQNN can nearly recover the performance of an ideal noiseless circuit by increasing the sampling budget.
Virtual Quantum Classifier (VQC):
- Goal: Binary classification of classical data encoded into quantum states.
- Approach: A VQC is trained on noisy quantum devices where the encoding and processing gates are subject to depolarizing noise.
- Results: The VQC consistently achieves higher prediction accuracy on both training and testing datasets compared to conventional QNN-based classifiers (which correspond to a VQNN with a sampling budget of 1.0). The accuracy improves as the sampling budget increases.
Virtual Variational Quantum Eigensolver (VVQE):
- Goal: Estimate the ground-state energy of a Hamiltonian (an unsupervised learning task).
- Approach: The authors address the issue that linear combinations of quantum states may not be valid density matrices (they may not be positive semi-definite). They introduce a loss function that includes a penalty term for the "negativity" of the averaged state, approximated via a truncated Chebyshev expansion.
- Results: In the estimation of the ground-state energy of the $HeH^+$ molecule, the VVQE yields significantly more accurate results than standard VQE when using noisy circuits, provided a sufficient sampling budget is allocated.

Significance and Claims
The paper claims that VQNNs offer a flexible learning paradigm that expands the space of achievable computations for near-term quantum hardware. The significance lies in two main advantages:

Enhanced Expressivity: By optimizing over HP maps rather than just CPTP maps, VQNNs can access a larger optimization space. This allows the model to simulate complex operations, such as long-range entanglement, using linear combinations of shallow circuits, potentially bypassing the "barren plateau" problem associated with deep circuits.
Improved Noise Robustness: The framework naturally functions as a variational protocol for probabilistic error cancellation. By learning optimal coefficients in the presence of noise, the VQNN can partially cancel out noise effects, effectively acting as a buffer against hardware imperfections.

The authors conclude that VQNNs represent a strategy to overcome hardware limitations by investing additional resources in sampling and classical computation, thereby strengthening the practical applications of NISQ devices.

Technical Summary: Virtual Quantum Neural Networks

More like this