Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2609.23050

Quantum Generative Modeling with a Coherent Memory, Exact Training and a Memory-Dimension Advantage

Jamal Slim. Saverio Monaco, Ran Xue, Dirk Kruecker, Kerstin Borras

quant-phSubmitted Sep 19, 2026• Announced Sep 22, 2026• 1 citation• Revised Nov 2, 2026 (v2)
v1 · Sep 19, 2026v2 · Nov 2, 2026
arXivHTMLPDF

Abstract

We introduce the coherent memory Born machine (CoMB), an algorithm for quantum generative modeling on a register whose size does not depend on the dimension of the data. CoMB generates a sample one component per round. In each round, a work register is entangled with a memory register and measured, and the outcome gives one component. The memory is never measured and carries correlations between rounds. We show that the characteristic function of the generated distribution is a product of small matrices, one per round. This gives exact training, with maximum mean discrepancy losses and gradients costing time linear in the data dimension, without sampling or likelihood. We also compile every round onto one work qubit and evaluate device noise exactly. To isolate coherence, each CoMB is compared with its dephased twin and with equally trained classical hidden Markov models. We benchmark it on continuous data, $368$-cell calorimeter showers from five qubits, and on sequential data, handwritten digits from three qubits. The coherent memory provides a memory-dimension advantage. Four coherent levels do the work of about six classical states on both benchmarks, and of at least five with $95\%$ confidence on showers at every training budget, including convergence. On showers, removing their coherence raises the loss by $23$ to $31\%$. With eight levels, where all models approach a common floor, coherent and classical memories become indistinguishable with longer training. On an IBM Heron processor, the memory carries two thirds of the inter-round correlation through twelve mid-circuit measurements.

AI Summary

The paper asks whether a small coherent quantum memory can learn high-dimensional generative distributions with fewer memory levels than a classical hidden Markov model (HMM). Its coherent memory Born machine emits one token per round by measuring and resetting a work register while retaining an unmeasured memory. Continuous outputs use learned bin readouts and a fixed linear decoder. A tilted-instrument identity expresses characteristic functions as products of maps on the memory. This enables sampling-free evaluation and differentiation of random-feature maximum mean discrepancy (MMD) objectives, costing $O(RdK\chi^3)$ per step for $R$ frequencies, $d$ rounds, $K$ tokens and memory dimension $\chi$. The paper also gives binary-tree compilation onto one work qubit and exact evaluation under specified noise models. Comparisons use dephased twins and equally trained classical HMMs. On 368-cell calorimeter showers, five-qubit circuits with $\chi=4$ achieve a converged held-out MMD loss 15% below the four-state HMM; their equivalent classical memory is approximately 5.8 states, with a bootstrap 95% interval of 5.0–7.2. Their loss is 23–31% below the dephased twin across training budgets. Digits yield approximately 6.2 equivalent classical states. At $\chi=8$, coherent and classical performance becomes indistinguishable with longer training. A processor experiment preserves about two thirds of inter-round correlation over twelve rounds. All models are classically simulable; no computational or practical advantage is claimed. Classical baselines are trained, not provably optimal. Showers remain distinguishable from reference simulations, digits use fixed training budgets, shared readouts impose a fidelity floor, and hardware execution covers only twelve of 369 rounds.

Little sign of AI-generated textUnclear noveltyPossible error

Central claim may not hold: The claimed comparison with the most general classical memory uses factorized transition and emission matrices rather than unrestricted outcome-dependent classical transitions, so the reported memory-dimension advantage is established only against a restricted classical family. The abstract says dephasing raises the four-level shower loss by 23–31%, but the reported losses imply increases of approximately 30–45%; 23–31% uses the dephased loss as the denominator instead.

Written by gpt-6.1-sol from the paper's full text. Automated: it can be wrong.

Related papers

An IQP Born Machine for Calorimeter Image Generation at 64 Qubits with Compiled-IQP Deployment

Jamal Slim, Saverio Monaco, Florian Rehm +2 more

The challenge to scaling quantum generative models on near-term hardware is training. Variational circuit Born machines require repeated quantum sampling and are prone to barren plateaus. Instantaneous Quantum Polynomial-time (IQP) Born machines sidestep both, since their loss is built from low-order Pauli-Z correlators that admit an unbiased classical estimator, while sampling from worst-case circuits in the class is conjectured to be classically hard. We take this train-on-classical, deploy-on-quantum workflow to a real high-energy-physics generative task, learning calorimeter shower profiles at 64 qubits and running the trained model on an IBM Heron r2 superconducting processor at 67 physical qubits. Three ingredients make it work. A uniform mixture of IQP circuits (MoIQP) widens the model class at single-circuit training cost. The Pearson-Stabilized Correlation Kernel (PSCK) biases descent toward the pairwise correlations that carry the shower-development physics, which the standard heat kernel systematically compresses. An exact deferred-measurement compilation collapses the mixture into a single IQP circuit, realized on hardware as a constant-depth dynamic circuit with zero SWAP insertions on the device's native heavy-hex graph. The trained model reconstructs the correlation structure to within 0.016 of the floor imposed by the encoding itself. Raw device samples reproduce the per-cell energy spectra and the full pairwise correlation structure at Pearson r = 0.989, up to a single global amplitude compression of depolarizing origin. A Gaussian copula fitted to the same training split matches the pairwise target more accurately than the quantum model at negligible cost. The contribution is therefore the classical trainability, exact compilation, and hardware deployability of a quantum generative model at this scale, not superiority over classical surrogates.

quant-pharXiv:2605.27735•May 28, 2026•5 citations

"Train classical, deploy quantum" requires rethinking generalization

Snehal Raj, Natansh Mathur, Alejandro Perdomo-Ortiz

Generative models have become central across science and industry, from image and text synthesis to the design of molecules and materials. Quantum generative models are considered one of the most promising applications for quantum computers, since a quantum circuit naturally produces samples from the distribution it encodes, and for suitable circuits that distribution is believed to be hard for any classical computer to reproduce. A leading strategy trains these models on a classical computer and reserves the quantum device for generating samples at deployment. This is possible when the training loss can be evaluated on a classical computer. A prime example is the maximum mean discrepancy (MMD$^2$), a moment-matching loss that compares the model and the data through their Pauli-$Z$ correlations. Research so far has asked whether such models can be trained and whether their sampling is hard; whether minimizing such an objective yields a model that \emph{generalizes}, rather than one that merely reproduces the training statistics, remains poorly understood. We benchmark thirteen quantum and classical generative models by direct sampling on two application-inspired datasets: first a cardinality-constrained dataset at up to $30$ qubits and second a dataset of genomic single-nucleotide variants, whose valid set is the observed data. Models that converge the loss to the same value differ widely in how much of the unseen valid set they cover. These results indicate that a converged moment-matching loss is not a reliable measure of generalization, and that a train-classical, deploy-quantum workflow has to measure generalization by sampling the trained model, a step that at the sizes of interest is believed to require the quantum device.

quant-pharXiv:2608.31117•Sep 1, 2026•3 citations

Generative modeling with Gaussian Boson Sampling: classically trainable Bosonic Born Machines

Zoltán Kolarovszki, Bence Bakó, Michał Oszmaniec +2 more

Quantum generative modeling has emerged as a promising application of quantum computers, aiming to model complex probability distributions beyond the reach of classical methods. In practice, however, training such models often requires costly gradient estimation performed directly on the quantum hardware. Crucially, for certain structured quantum circuits, expectation values of local observables can be efficiently evaluated on a classical computer, enabling classical training without calls to the quantum hardware in the optimization loop. In these models, sampling from the resulting circuits can still be classically hard, so inference must be performed on a quantum device, yielding a potential computational advantage. In this work, we introduce a photonic quantum generative model built on parametrized Gaussian Boson Sampling circuits. The training is based on the efficient classical evaluation of expectation values enabled by the Gaussian structure of the state, allowing scalable optimization of the model parameters through the maximum mean discrepancy loss function. We demonstrate the effectiveness of the approach through numerical experiments on photonic systems with up to 805 modes and over a million trainable parameters, highlighting its scalability and suitability for near-term photonic quantum devices.

quant-pharXiv:2603.11195•Mar 13, 2026•13 citations

Fermionic Born Machines: Classical training of quantum generative models based on Fermion Sampling

Bence Bakó, Zoltán Kolarovszki, Zoltán Zimborás

Quantum generative learning is a promising application of quantum computers, but faces several trainability challenges, including the difficulty in experimental gradient estimations. For certain structured quantum generative models, however, expectation values of local observables can be efficiently computed on a classical computer, enabling fully classical training without quantum gradient evaluations. Although training is classically efficient, sampling from these circuits is still believed to be classically hard, so inference must be carried out on a quantum device, potentially yielding a computational advantage. In this work, we introduce Fermionic Born Machines as an example of such classically trainable quantum generative models. The model employs parameterized magic states and fermionic linear optical (FLO) transformations with learnable parameters. The training exploits a decomposition of the magic states into Gaussian operators, which permits efficient estimation of expectation values. Furthermore, the specific structure of the ansatz induces a loss landscape that exhibits favorable characteristics for optimization. The FLO circuits can be implemented, via fermion-to-qubit mappings, on qubit architectures to sample from the learned distribution during inference. Numerical experiments on systems up to 160 qubits demonstrate the effectiveness of our model and training framework.

quant-pharXiv:2511.13844•Nov 19, 2025•19 citations

The Born supremacy: quantum advantage and training of an Ising Born machine

Brian Coyle, Daniel Mills, Vincent Danos +1 more

Abstract The search for an application of near-term quantum devices is widespread. Quantum machine learning is touted as a potential utilisation of such devices, particularly those out of reach of the simulation capabilities of classical computers. In this work, we study such an application in generative modelling, focussing on a class of quantum circuits known as Born machines. Specifically, we define a subset of this class based on Ising Hamiltonians and show that the circuits encountered during gradient-based training cannot be efficiently sampled from classically up to multiplicative error in the worst case. Our gradient-based training methods use cost functions known as the Sinkhorn divergence and the Stein discrepancy, which have not previously been used in the gradient-based training of quantum circuits, and we also introduce quantum kernels to generative modelling. We show that these methods outperform the previous standard method, which used maximum mean discrepancy (MMD) as a cost function, and achieve this with minimal overhead. Finally, we discuss the ability of the model to learn hard distributions and provide formal definitions for ‘quantum learning supremacy’. We also exemplify the work of this paper by using generative modelling to perform quantum circuit compilation.

arXiv:1904.02214•npj Quantum Information•Jul 8, 2020•185 citations

Trainable Quantum-Native Language Models: Dissipative Memory, Capacity Limits, and Hardware Validation

Deniz Yunus Göğüş

We study byte-level autoregressive language models whose hidden state is a small quantum register, whose transitions are learned token-conditioned quantum channels, and whose next-byte probabilities come directly from Born-rule measurements without a learned classical vocabulary projection or softmax. In the fixed-width recurrent setting, unitary dynamics cannot contract or reset stored distinctions; experimentally, unitary recurrence loses to its retrained one-token control in every tested variant. Introducing learned dissipation, then token-conditioned dissipative gating, yields useful memory beyond one token. The gated model improves test negative log-likelihood by 0.20 +/- 0.03 nats/byte over retrained parameter-matched one-token controls at context length 12, and by 0.29-0.33 at longer contexts, but never beats retrained two-token controls. A key-value recall benchmark shows a width-dependent threshold: one stored pair meets the pre-registered accuracy bar at 4-5 memory qubits, two pairs at 6-7, and three pairs remain unsolved. This is consistent with, but does not establish, a capacity explanation for the short effective memory horizon. As a secondary result, one-qubit models reach lower stationary memory entropy than a heuristically optimized classical predictive frontier at matched error in 9/9 seeds. On IBM Heron hardware, computational-basis memory-register distributions show signal against a uniform null in 12/12 circuits across prefixes of up to 12 tokens, next-bit probabilities pass pre-registered accuracy bars at four-token prefixes, and bytes are generated autoregressively by sampling physically measured bit probabilities. The tested models are classically simulable and substantially worse than classical n-gram baselines. We claim no computational quantum advantage.

quant-pharXiv:2610.12806•Oct 12, 2026•0 citations
We introduce the coherent memory Born machine (CoMB), an algorithm for quantum generative modeling on a register whose size does not depend on the dimension of the data. CoMB generates a sample one component per round. In each round, a work register is entangled with a memory register and measured, and the outcome gives one component. The memory is never measured and carries correlations between rounds. We show that the characteristic function of the generated distribution is a product of small matrices, one per round. This gives exact training, with maximum mean discrepancy losses and gradients costing time linear in the data dimension, without sampling or likelihood. We also compile every round onto one work qubit and evaluate device noise exactly. To isolate coherence, each CoMB is compared with its dephased twin and with equally trained classical hidden Markov models. We benchmark it on continuous data, 368368368-cell calorimeter showers from five qubits, and on sequential data, handwritten digits from three qubits. The coherent memory provides a memory-dimension advantage. Four coherent levels do the work of about six classical states on both benchmarks, and of at least five with 95%95\%95% confidence on showers at every training budget, including convergence. On showers, removing their coherence raises the loss by 232323 to 31%31\%31%. With eight levels, where all models approach a common floor, coherent and classical memories become indistinguishable with longer training. On an IBM Heron processor, the memory carries two thirds of the inter-round correlation through twelve mid-circuit measurements.
The paper asks whether a small coherent quantum memory can learn high-dimensional generative distributions with fewer memory levels than a classical hidden Markov model (HMM). Its coherent memory Born machine emits one token per round by measuring and resetting a work register while retaining an unmeasured memory. Continuous outputs use learned bin readouts and a fixed linear decoder.
A tilted-instrument identity expresses characteristic functions as products of maps on the memory. This enables sampling-free evaluation and differentiation of random-feature maximum mean discrepancy (MMD) objectives, costing O(RdKχ3)O(RdK\chi^3)O(RdKχ3) per step for RRR frequencies, ddd rounds, KKK tokens and memory dimension χ\chiχ. The paper also gives binary-tree compilation onto one work qubit and exact evaluation under specified noise models.
Comparisons use dephased twins and equally trained classical HMMs. On 368-cell calorimeter showers, five-qubit circuits with χ=4\chi=4χ=4 achieve a converged held-out MMD loss 15% below the four-state HMM; their equivalent classical memory is approximately 5.8 states, with a bootstrap 95% interval of 5.0–7.2. Their loss is 23–31% below the dephased twin across training budgets. Digits yield approximately 6.2 equivalent classical states. At χ=8\chi=8χ=8, coherent and classical performance becomes indistinguishable with longer training. A processor experiment preserves about two thirds of inter-round correlation over twelve rounds.
All models are classically simulable; no computational or practical advantage is claimed. Classical baselines are trained, not provably optimal. Showers remain distinguishable from reference simulations, digits use fixed training budgets, shared readouts impose a fidelity floor, and hardware execution covers only twelve of 369 rounds.
The challenge to scaling quantum generative models on near-term hardware is training. Variational circuit Born machines require repeated quantum sampling and are prone to barren plateaus. Instantaneous Quantum Polynomial-time (IQP) Born machines sidestep both, since their loss is built from low-order Pauli-Z correlators that admit an unbiased classical estimator, while sampling from worst-case circuits in the class is conjectured to be classically hard. We take this train-on-classical, deploy-on-quantum workflow to a real high-energy-physics generative task, learning calorimeter shower profiles at 64 qubits and running the trained model on an IBM Heron r2 superconducting processor at 67 physical qubits. Three ingredients make it work. A uniform mixture of IQP circuits (MoIQP) widens the model class at single-circuit training cost. The Pearson-Stabilized Correlation Kernel (PSCK) biases descent toward the pairwise correlations that carry the shower-development physics, which the standard heat kernel systematically compresses. An exact deferred-measurement compilation collapses the mixture into a single IQP circuit, realized on hardware as a constant-depth dynamic circuit with zero SWAP insertions on the device's native heavy-hex graph. The trained model reconstructs the correlation structure to within 0.016 of the floor imposed by the encoding itself. Raw device samples reproduce the per-cell energy spectra and the full pairwise correlation structure at Pearson r = 0.989, up to a single global amplitude compression of depolarizing origin. A Gaussian copula fitted to the same training split matches the pairwise target more accurately than the quantum model at negligible cost. The contribution is therefore the classical trainability, exact compilation, and hardware deployability of a quantum generative model at this scale, not superiority over classical surrogates.
Generative models have become central across science and industry, from image and text synthesis to the design of molecules and materials. Quantum generative models are considered one of the most promising applications for quantum computers, since a quantum circuit naturally produces samples from the distribution it encodes, and for suitable circuits that distribution is believed to be hard for any classical computer to reproduce. A leading strategy trains these models on a classical computer and reserves the quantum device for generating samples at deployment. This is possible when the training loss can be evaluated on a classical computer. A prime example is the maximum mean discrepancy (MMD2^22), a moment-matching loss that compares the model and the data through their Pauli-ZZZ correlations. Research so far has asked whether such models can be trained and whether their sampling is hard; whether minimizing such an objective yields a model that \emph{generalizes}, rather than one that merely reproduces the training statistics, remains poorly understood. We benchmark thirteen quantum and classical generative models by direct sampling on two application-inspired datasets: first a cardinality-constrained dataset at up to 303030 qubits and second a dataset of genomic single-nucleotide variants, whose valid set is the observed data. Models that converge the loss to the same value differ widely in how much of the unseen valid set they cover. These results indicate that a converged moment-matching loss is not a reliable measure of generalization, and that a train-classical, deploy-quantum workflow has to measure generalization by sampling the trained model, a step that at the sizes of interest is believed to require the quantum device.
Quantum generative modeling has emerged as a promising application of quantum computers, aiming to model complex probability distributions beyond the reach of classical methods. In practice, however, training such models often requires costly gradient estimation performed directly on the quantum hardware. Crucially, for certain structured quantum circuits, expectation values of local observables can be efficiently evaluated on a classical computer, enabling classical training without calls to the quantum hardware in the optimization loop. In these models, sampling from the resulting circuits can still be classically hard, so inference must be performed on a quantum device, yielding a potential computational advantage. In this work, we introduce a photonic quantum generative model built on parametrized Gaussian Boson Sampling circuits. The training is based on the efficient classical evaluation of expectation values enabled by the Gaussian structure of the state, allowing scalable optimization of the model parameters through the maximum mean discrepancy loss function. We demonstrate the effectiveness of the approach through numerical experiments on photonic systems with up to 805 modes and over a million trainable parameters, highlighting its scalability and suitability for near-term photonic quantum devices.
Quantum generative learning is a promising application of quantum computers, but faces several trainability challenges, including the difficulty in experimental gradient estimations. For certain structured quantum generative models, however, expectation values of local observables can be efficiently computed on a classical computer, enabling fully classical training without quantum gradient evaluations. Although training is classically efficient, sampling from these circuits is still believed to be classically hard, so inference must be carried out on a quantum device, potentially yielding a computational advantage. In this work, we introduce Fermionic Born Machines as an example of such classically trainable quantum generative models. The model employs parameterized magic states and fermionic linear optical (FLO) transformations with learnable parameters. The training exploits a decomposition of the magic states into Gaussian operators, which permits efficient estimation of expectation values. Furthermore, the specific structure of the ansatz induces a loss landscape that exhibits favorable characteristics for optimization. The FLO circuits can be implemented, via fermion-to-qubit mappings, on qubit architectures to sample from the learned distribution during inference. Numerical experiments on systems up to 160 qubits demonstrate the effectiveness of our model and training framework.
Abstract The search for an application of near-term quantum devices is widespread. Quantum machine learning is touted as a potential utilisation of such devices, particularly those out of reach of the simulation capabilities of classical computers. In this work, we study such an application in generative modelling, focussing on a class of quantum circuits known as Born machines. Specifically, we define a subset of this class based on Ising Hamiltonians and show that the circuits encountered during gradient-based training cannot be efficiently sampled from classically up to multiplicative error in the worst case. Our gradient-based training methods use cost functions known as the Sinkhorn divergence and the Stein discrepancy, which have not previously been used in the gradient-based training of quantum circuits, and we also introduce quantum kernels to generative modelling. We show that these methods outperform the previous standard method, which used maximum mean discrepancy (MMD) as a cost function, and achieve this with minimal overhead. Finally, we discuss the ability of the model to learn hard distributions and provide formal definitions for ‘quantum learning supremacy’. We also exemplify the work of this paper by using generative modelling to perform quantum circuit compilation.
We study byte-level autoregressive language models whose hidden state is a small quantum register, whose transitions are learned token-conditioned quantum channels, and whose next-byte probabilities come directly from Born-rule measurements without a learned classical vocabulary projection or softmax. In the fixed-width recurrent setting, unitary dynamics cannot contract or reset stored distinctions; experimentally, unitary recurrence loses to its retrained one-token control in every tested variant. Introducing learned dissipation, then token-conditioned dissipative gating, yields useful memory beyond one token. The gated model improves test negative log-likelihood by 0.20 +/- 0.03 nats/byte over retrained parameter-matched one-token controls at context length 12, and by 0.29-0.33 at longer contexts, but never beats retrained two-token controls. A key-value recall benchmark shows a width-dependent threshold: one stored pair meets the pre-registered accuracy bar at 4-5 memory qubits, two pairs at 6-7, and three pairs remain unsolved. This is consistent with, but does not establish, a capacity explanation for the short effective memory horizon. As a secondary result, one-qubit models reach lower stationary memory entropy than a heuristically optimized classical predictive frontier at matched error in 9/9 seeds. On IBM Heron hardware, computational-basis memory-register distributions show signal against a uniform null in 12/12 circuits across prefixes of up to 12 tokens, next-bit probabilities pass pre-registered accuracy bars at four-token prefixes, and bytes are generated autoregressively by sampling physically measured bit probabilities. The tested models are classically simulable and substantially worse than classical n-gram baselines. We claim no computational quantum advantage.