⚛️ quantum physics

Classical training of conditional instantaneous quantum polynomial-time Born machines

This paper introduces a classically trainable conditional IQP Born machine that encodes conditions in the input state using a shared unitary, demonstrating superior efficiency and accuracy in modeling conditional distributions and phase transitions compared to previous diagonally conditioned models and classical generative models.

Original author: Oriol Balló-Gimbernat, Emiliano Tolotti, Laia Domingo, Fernando Vilariño, Adan Garriga

Published 2026-11-05

📖 6 min read🧠 Deep dive

CC BY 4.0

Original author: Oriol Balló-Gimbernat, Emiliano Tolotti, Laia Domingo, Fernando Vilariño, Adan Garriga

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

In the quiet corners of physics and computer science, researchers are exploring a new kind of artificial intelligence that lives inside quantum computers. Unlike the classical computers in our pockets, which process information as a simple sequence of zeros and ones, quantum machines use particles that can exist in many states at once. This unique ability allows them to tackle problems that would take ordinary computers thousands of years to solve. One promising application is generative modeling, where a machine learns the hidden patterns of a complex system—like the way atoms arrange themselves in a material or how pixels form a face—and then creates new, realistic examples of that system. The challenge has always been training these machines. Because quantum computers are still fragile and expensive to run, scientists have been searching for ways to teach them using classical computers first, saving the quantum hardware only for the final act of generating data.

A team of researchers has now developed a method to do exactly this for a specific type of quantum model called a Born machine. These models are designed to mimic the statistical behavior of quantum systems. The researchers discovered a way to teach a single quantum circuit to handle a whole family of different scenarios, rather than just one fixed situation. Imagine a single artist who can paint a landscape in summer, winter, spring, and autumn, simply by being told which season to depict, without needing to learn a completely new painting technique for each one. The team found that by encoding the "season" directly into the starting state of the quantum computer, they could train the machine entirely on a classical computer. Once trained, the machine could be deployed on a real quantum device to generate samples for any condition, even those it had never seen before.

The core of their discovery lies in how they prepared the quantum computer to receive instructions. Previous methods tried to adjust the internal knobs of the quantum circuit based on the condition, a bit like trying to change the engine of a car while it is driving. The new approach instead changes the fuel before the engine starts. They feed the quantum circuit a specific starting state that already contains the information about the condition, such as a specific temperature or a particular setting. The circuit then applies a single, shared set of operations to this starting state. Because the starting state carries the condition, the circuit does not need to be retrained or reconfigured for every new scenario. This allows the researchers to train the model using classical simulations, which are fast and reliable, and then run the final version on a quantum processor.

To test this idea, the team focused on two distinct challenges. First, they looked at a mathematical family of distributions where the average and the spread could be tuned independently. They trained a model to handle these variations and found that it could successfully predict outcomes for combinations of settings it had never encountered during training. In this test, their method was significantly more accurate than older techniques that tried to adjust the circuit's internal parameters for each condition. The new approach captured the underlying structure of the problem so well that it generalized to new situations with much less error.

The second, more demanding test involved the Curie-Weiss model, a famous system used to describe how magnets behave as they heat up or cool down. As the temperature changes, the magnetic spins in the material undergo a dramatic shift, moving from a disordered state to an ordered one. This transition is notoriously difficult to model because the shape of the probability distribution changes fundamentally. The researchers trained a single sixteen-qubit quantum model to handle this entire range of temperatures. The results were striking: the model accurately captured the transition from a single peak of probability to a double peak, a behavior that represents the physical shift in the material. When they compared their results to classical generative models, the quantum approach performed just as well as classical models that had seven times more adjustable parameters. Even more impressively, the model learned the physics so well that it could predict the behavior of the system at temperatures it had never been shown during training.

The team did not stop at small simulations. They pushed the method to its limits, training models with up to two thousand and forty-eight qubits. At this scale, it is impossible to simulate the system on a classical computer to check the answers directly. However, because the physical laws governing this specific system are well understood, the researchers could calculate the expected behavior of the magnetism and compare it to their model's output. They found that the model successfully learned the long-range order of the cold phase and could extrapolate this behavior to sizes far beyond what was used in training. This demonstrated that the method scales effectively, maintaining its ability to learn complex patterns even as the system grows to a size where classical verification becomes impossible.

The study also highlighted the importance of how the condition is encoded. They tested different ways of preparing the starting state and found that a more sophisticated preparation method, which involved twisting the quantum state in a specific way, yielded much better results than a simpler approach. This suggests that the way information is fed into the quantum computer is just as critical as the learning algorithm itself. By carefully designing the input, they could impose physical symmetries on the model, ensuring it respected the fundamental laws of the system it was trying to mimic.

Ultimately, this work extends the reach of quantum machine learning. It shows that we do not need to wait for perfect, error-free quantum computers to begin training complex models. By using classical computers to learn the patterns and quantum computers only to generate the final data, researchers can now tackle problems involving conditional generation—where the output depends on a specific input variable. This opens the door to simulating physical systems under various conditions, from different temperatures to different external fields, using a single, shared quantum circuit. The ability to train these models classically and then deploy them on quantum hardware represents a significant step forward in making quantum generative models practical for real-world scientific discovery.

Technical Summary: Classical Training of Conditional Instantaneous Quantum Polynomial-Time Born Machines

Problem Statement
Conditional generative models aim to approximate a family of target distributions $p(\cdot | c)$ using a single model, capturing dependencies on conditioning variables such as physical temperature or class labels. While Quantum Circuit Born Machines (QCBMs) are promising architectures for this task, training them at scale typically requires quantum hardware due to the exponential cost of classical simulation. Furthermore, training on noisy intermediate-scale quantum (NISQ) hardware faces challenges like barren plateaus and noise-induced optimization issues. Existing conditional IQP (Instantaneous Quantum Polynomial-time) Born machines have attempted to address this by making parameters of the diagonal layer condition-dependent, but these approaches often fail to learn the condition dependence jointly across the entire family or require separate training blocks for each condition.

Methodology
The authors propose a framework for Conditional IQP Born Machines where the condition $c$ is encoded directly into the input state $|\phi(c)\rangle$ , while a single, shared IQP unitary $U_\theta$ is applied across all conditions. The model represents the distribution as:
$q_\theta(x | c) = |\langle x| U_\theta |\phi(c)\rangle|^2$
where $U_\theta = H^{\otimes n} D(\theta) H^{\otimes n}$ , with $D(\theta)$ being a diagonal layer of phase gates.

Key methodological components include:

Sampling-and-Ratio Interface: To enable classical training, the input state (after the initial Hadamard layer) must support efficient sampling of computational-basis strings and evaluation of amplitude ratios. The authors utilize encoders based on product states, matrix product states, or permutation-symmetric states that satisfy this requirement.
Classical Estimation of Gradients: The training objective is the squared Maximum Mean Discrepancy (MMD) with a Gaussian kernel. The authors show that Pauli-Z moments and their gradients can be estimated classically to additive accuracy using only the sampling and amplitude ratios of the reference state. This allows the optimization of the shared parameters $\theta$ without quantum hardware.
Encoder Design: The paper introduces specific encoders, such as the parity encoder and the twisted parity encoder (incorporating one-axis twisting), which impose symmetries (e.g., global spin-flip invariance) and encode condition-dependent magnitudes into the reference state, not just phases.

Key Contributions

Theoretical Expressivity: The authors prove (Proposition 1) that input conditioning allows the model to represent conditional families that no diagonally conditioned IQP model (with a fixed input and the same number of qubits) can reproduce. This is because the IQP unitary cannot alter the output's distribution in the X-basis, which is fixed by the input state's magnitudes. Input conditioning breaks this limitation by varying the X-basis distribution across conditions.
Classical Training Framework: The work extends the classical training framework of Recio-Armengol et al. to condition-dependent inputs, demonstrating that efficient classical estimation of the MMD objective and its gradients is possible for these models.
Encoder Impact: The study highlights that the choice of encoder significantly impacts performance. Encoders that encode condition-dependent magnitudes (non-diagonal-equivalent) outperform those that only encode phases or rely on diagonal conditioning.

Experimental Results
The framework was evaluated on two target families using 16 qubits, with large-scale simulations up to 2048 qubits:

Hamming-Weight Family: A model trained on conditions where coordinates were varied separately successfully generalized to unseen combinations of coordinates. The input-conditioned model achieved a Total Variation Distance (TVD) approximately three times lower than a diagonal-conditioning baseline with the same circuit structure.
Curie–Weiss Spin Model:
- A single conditional model successfully captured the transition from unimodal to bimodal magnetization profiles across different temperatures, including held-out temperatures.
- The twisted parity encoder achieved a mean TVD six times lower than the diagonal-conditioning baseline (0.046 vs. 0.28) under the same protocol.
- The shared conditional model was comparable in accuracy to classical generative models (Transformers, RBMs, VAEs) with up to seven times more parameters.
- Large-Scale Training: The authors trained models classically with up to 2048 qubits. These models successfully reproduced the mean squared magnetization away from the crossover region, demonstrating scalability beyond exact simulation capabilities.
- Hardware Deployment: Models with up to 30 qubits were sampled on real quantum hardware (IQM's Emerald), qualitatively reproducing the phase transition, though with significant errors in the ordered regime due to device noise.

Significance and Claims
The paper claims that this work extends classically trained IQP Born machines from modeling single distributions to conditional families. The primary significance lies in demonstrating that:

Input conditioning is a more powerful mechanism for conditional generation than diagonal conditioning, offering greater expressivity without increasing the number of qubits.
Classical training is feasible and effective for conditional IQP models, provided the encoder supports the sampling-and-ratio interface. This bypasses the need for quantum hardware during the training phase, mitigating noise and optimization issues associated with NISQ devices.
The approach allows for the verification of quantum generative models at scales (e.g., 2048 qubits) where exact simulation is impossible, by leveraging classically tractable observables (like mean squared magnetization) for validation.

The authors remain modest regarding quantum advantage, noting that the specific targets used (permutation-invariant) are classically tractable for sampling, meaning the experiments test representation and generalization capabilities rather than a computational speedup over classical sampling. They identify the construction of efficient encoders and the characterization of which encoders preserve sampling hardness as open questions for future work.

Technical Summary: Classical Training of Conditional Instantaneous Quantum Polynomial-Time Born Machines

More like this