arxiv: 2604.06523 · v1 · submitted 2026-04-07 · 🪐 quant-ph · cs.AI· cs.LG

Recognition: no theorem link

Soft-Quantum Algorithms

Basil Kyriacou , Mo Kordzanganeh , Maniraman Periyasamy , Alexey Melnikov

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:29 UTC · model grok-4.3

classification 🪐 quant-ph cs.AIcs.LG

keywords variational quantum circuitsquantum machine learningunitary regularizationcircuit alignmenthybrid quantum-classical networkssupervised classificationreinforcement learning

0 comments

The pith

Direct matrix optimization with unitarity regularization trains variational quantum circuits faster than gate decomposition.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper establishes that for small numbers of qubits and large datasets, quantum operations can be optimized by treating them as matrices rather than as sequences of gates, with a single added term in the loss function keeping the matrices close to unitary. A follow-up alignment step then converts the resulting soft-unitary matrix into an equivalent gate circuit. This two-step route is shown to reach lower loss values in minutes instead of hours on a five-qubit classification problem and to produce a competitive hybrid agent on a cartpole reinforcement-learning task. A sympathetic reader would care because the method removes the main computational bottleneck that currently forces most quantum machine learning to stay on simulators with tiny data sets.

Core claim

Quantum operations on pure states can be trained directly as matrices by minimizing a loss that includes one regularization term to enforce near-unitarity; a subsequent circuit-alignment procedure recovers a gate-based variational circuit from the trained soft-unitary matrix. On a five-qubit supervised classification task with 1000 data points the procedure yields lower binary cross-entropy loss in under four minutes, versus more than two hours for conventional circuit training. The same soft-unitary matrices can be embedded inside a hybrid quantum-classical network that outperforms a purely classical baseline on a cartpole reinforcement-learning task.

What carries the argument

The soft-unitary training procedure: direct optimization of matrix elements under a single unitarity-regularization term, followed by circuit alignment to recover an equivalent gate decomposition.

If this is right

For few-qubit problems with large data sets, the matrix-first route bypasses gate decomposition during the expensive optimization phase.
The resulting trained circuits can be deployed on actual quantum hardware after the alignment step.
Soft-unitary matrices can be inserted into hybrid quantum-classical architectures, as demonstrated by the cartpole reinforcement-learning agent that exceeds a classical baseline of similar size.
Training time scales with matrix size rather than with the number of gates, offering a practical speed-up when datasets are large.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Because the method works on classical simulators, it could let researchers test variational quantum models on data sets an order of magnitude larger than those currently feasible with gate-level training.
The approach implicitly decouples the search for good unitary operators from the choice of gate set, which may allow systematic comparison of different ansatz families once the matrices are obtained.
If the regularization strength can be tuned automatically, the same pipeline might be applied to deeper circuits where direct gate training becomes prohibitive.

Load-bearing premise

A single regularization term added to the loss is sufficient to keep the learned matrices close enough to unitary that the subsequent alignment step produces a usable circuit without large fidelity loss.

What would settle it

On the same five-qubit, 1000-point classification task, run the two-step procedure and measure whether the final aligned circuit achieves binary cross-entropy loss comparable to or lower than direct gate training; if the aligned circuit instead shows markedly higher loss or requires comparable wall-clock time, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2604.06523 by Alexey Melnikov, Basil Kyriacou, Maniraman Periyasamy, Mo Kordzanganeh.

**Figure 2.** Figure 2: FIG. 2. Training loss versus wall-clock time for the soft-unitary [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: FIG. 3. Circuit alignment loss, measured as the matrix norm [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

**Figure 4.** Figure 4: FIG. 4. Difference in output values between the soft-unitary [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

**Figure 6.** Figure 6: FIG. 6. Comparison of the cartpole task between a purely classical and a hybrid quantum-classical neural network. Both [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗

read the original abstract

Quantum operations on pure states can be fully represented by unitary matrices. Variational quantum circuits, also known as quantum neural networks, embed data and trainable parameters into gate-based operations and optimize the parameters via gradient descent. The high cost of training and low fidelity of current quantum devices, however, restricts much of quantum machine learning to classical simulation. For few-qubit problems with large datasets, training the matrix elements directly, as is done with weight matrices in classical neural networks, can be faster than decomposing data and parameters into gates. We propose a method that trains matrices directly while maintaining unitarity through a single regularization term added to the loss function. A second training step, circuit alignment, then recovers a gate-based architecture from the resulting soft-unitary. On a five-qubit supervised classification task with 1000 datapoints, this two-step process produces a trained variational circuit in under four minutes, compared to over two hours for direct circuit training, while achieving lower binary cross-entropy loss. In a second experiment, soft-unitaries are embedded in a hybrid quantum-classical network for a reinforcement learning cartpole task, where the hybrid agent outperforms a purely classical baseline of comparable size.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper gives a workable shortcut for training small variational circuits by optimizing soft-unitary matrices first then aligning to gates, but leaves the key unitarity preservation claim without numbers.

read the letter

The core contribution is a two-step procedure: train the matrix elements of the quantum operation directly while adding one regularization term to push toward unitarity, then run a circuit-alignment step to recover an explicit gate decomposition. On the five-qubit supervised classification task with 1000 points this finishes in under four minutes versus over two hours for standard circuit training and reaches lower binary cross-entropy. The cartpole reinforcement-learning hybrid also beats a same-size classical baseline. That empirical time saving on modest qubit counts is the clearest practical takeaway and matches what one would expect when the problem is small enough to simulate classically anyway. Direct matrix training sidesteps the overhead of gate decomposition during optimization, and the alignment step is a logical way to hand the result back to a hardware-compatible form. The soft spot is the missing quantitative check on how well the regularization actually works. The abstract and stress-test note both flag the absence of any reported ||U†U − I||_F, spectral deviation, or post-alignment fidelity numbers. Without those, it is possible that the optimizer trades classification accuracy for the penalty and leaves the aligned circuit with degraded effective unitarity, especially if the deviation grows with qubit number or data size. Error bars on the timing and loss comparisons are also not mentioned. The method stays within the regime of classical simulation of few-qubit circuits and does not claim broader quantum advantage. This is useful reading for groups that already run small variational experiments and want faster classical prototypes. I would send it to peer review once the authors add the unitarity diagnostics and statistical details; the idea is simple enough that referees can evaluate it without a long turnaround.

Referee Report

3 major / 2 minor

Summary. The paper proposes training variational quantum circuits by directly optimizing matrix elements as 'soft-unitaries' via gradient descent on a composite loss that includes a single regularization term to enforce approximate unitarity, followed by a circuit-alignment step that recovers an equivalent gate-based variational circuit. On a 5-qubit supervised classification task with 1000 datapoints, the method reportedly trains in under 4 minutes (vs. >2 hours for direct circuit training) while achieving lower binary cross-entropy loss; a second experiment embeds soft-unitaries in a hybrid quantum-classical agent that outperforms a classical baseline on cartpole reinforcement learning.

Significance. If the regularization reliably keeps the learned matrices sufficiently close to the unitary manifold for the alignment step to recover high-fidelity circuits, the approach could provide a practical route to faster classical simulation and training of small-qubit variational quantum models on large datasets by borrowing techniques from classical neural-network training. The reported empirical speed-up and loss improvement constitute a concrete, falsifiable demonstration that merits attention, though its broader significance depends on quantitative validation of unitarity preservation and generalization beyond the 5-qubit regime.

major comments (3)

[Method and regularization description] The central claim that a single additive regularization term suffices to produce soft-unitaries close enough to the unitary group for circuit alignment to preserve performance is load-bearing, yet the manuscript supplies no post-training measurement of ||U†U − I||_F, spectral deviation, or average gate fidelity, nor any scaling analysis with qubit number or dataset size. Gradient descent on the composite loss can therefore trade classification accuracy against unitarity without the reader being able to assess the resulting fidelity loss in the aligned circuit.
[Experimental results on the five-qubit classification task] The timing and loss comparisons (under 4 min vs. >2 h, lower BCE) are presented without error bars, details on the exact form and coefficient of the regularization term, the circuit-alignment algorithm, or the precise gate decomposition used in the baseline. This makes it impossible to determine whether the reported advantage is robust or sensitive to hyper-parameter choices and implementation details.
[Reinforcement-learning experiment] In the hybrid RL experiment, the manner in which soft-unitary matrices are embedded and simulated within the quantum component is not specified quantitatively (e.g., whether they are projected back onto unitaries or how non-unitary deviations propagate through the policy network), leaving open whether the performance gain truly stems from the quantum part or from the hybrid architecture itself.

minor comments (2)

The term 'soft-unitary' is introduced without a precise mathematical definition (e.g., a bound on the deviation from U†U = I) or comparison to related concepts such as approximate unitaries or Stiefel-manifold optimization already present in the quantum-machine-learning literature.
Notation for the loss function, regularization coefficient, and circuit-alignment procedure should be introduced with explicit equations rather than descriptive prose to improve reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and constructive comments on our manuscript. We address each major comment below and have revised the manuscript to incorporate the suggested clarifications and additions.

read point-by-point responses

Referee: [Method and regularization description] The central claim that a single additive regularization term suffices to produce soft-unitaries close enough to the unitary group for circuit alignment to preserve performance is load-bearing, yet the manuscript supplies no post-training measurement of ||U†U − I||_F, spectral deviation, or average gate fidelity, nor any scaling analysis with qubit number or dataset size. Gradient descent on the composite loss can therefore trade classification accuracy against unitarity without the reader being able to assess the resulting fidelity loss in the aligned circuit.

Authors: We agree that explicit post-training unitarity metrics were not reported. In the revised manuscript we will include measurements of ||U†U − I||_F, the spectral norm deviation, and average gate fidelity for the trained soft-unitaries on the 5-qubit task. We will also add a short scaling study with qubit number (up to 6–7 qubits) and dataset size to quantify how the regularization coefficient affects the trade-off. These additions will allow readers to directly evaluate the fidelity preserved by the alignment step. revision: yes
Referee: [Experimental results on the five-qubit classification task] The timing and loss comparisons (under 4 min vs. >2 h, lower BCE) are presented without error bars, details on the exact form and coefficient of the regularization term, the circuit-alignment algorithm, or the precise gate decomposition used in the baseline. This makes it impossible to determine whether the reported advantage is robust or sensitive to hyper-parameter choices and implementation details.

Authors: We acknowledge the lack of these implementation details and statistical measures. The revised version will report error bars computed over 10 independent runs, specify the regularization term as λ‖U†U − I‖_F² with the exact λ value used, provide pseudocode and a description of the circuit-alignment procedure, and detail the gate set and decomposition method employed for the direct circuit baseline. These changes will enable reproducibility and assessment of robustness. revision: yes
Referee: [Reinforcement-learning experiment] In the hybrid RL experiment, the manner in which soft-unitary matrices are embedded and simulated within the quantum component is not specified quantitatively (e.g., whether they are projected back onto unitaries or how non-unitary deviations propagate through the policy network), leaving open whether the performance gain truly stems from the quantum part or from the hybrid architecture itself.

Authors: We thank the referee for highlighting this ambiguity. In the revision we will add a quantitative description of the embedding: the soft-unitary matrices are used directly in the quantum simulation layer without projection, with any non-unitary component propagating through the density-matrix evolution; we will also report the observed deviation norm during training and include an ablation comparing the hybrid agent against a version that forces unitarity via projection. This will clarify the contribution of the soft-unitary quantum component. revision: yes

Circularity Check

0 steps flagged

No circularity: proposal is self-contained empirical method

full rationale

The paper introduces a two-step procedure—direct matrix training with an added regularization term to enforce approximate unitarity, followed by a separate circuit-alignment step to recover a gate decomposition. No equation or claim reduces the final trained circuit or reported speed/accuracy gains to a fitted parameter renamed as a prediction, nor to a self-citation chain, uniqueness theorem, or ansatz imported from prior work by the same authors. The regularization term is presented as a direct proposal rather than derived from the target result, and performance numbers are obtained from explicit experiments on concrete tasks. The derivation chain therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim relies on the effectiveness of the regularization term and the circuit alignment procedure, whose details and any additional parameters are not specified in the abstract.

free parameters (1)

regularization coefficient
The strength of the unitarity penalty term, which must be chosen or tuned to balance task loss and unitarity.

axioms (1)

standard math Quantum operations on pure states can be represented by unitary matrices
This is a fundamental principle of quantum mechanics for closed systems.

invented entities (1)

soft-unitary matrix no independent evidence
purpose: A matrix that is trained directly and is close to but not exactly unitary
The paper introduces this concept to enable direct optimization while approximating quantum operations.

pith-pipeline@v0.9.0 · 5515 in / 1405 out tokens · 63606 ms · 2026-05-10T18:29:48.368655+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

34 extracted references · 5 canonical work pages · 1 internal anchor

[1]

Peruzzo, J

A. Peruzzo, J. McClean, P. Shadbolt, M.-H. Yung, X.-Q. Zhou, P. J. Love, A. Aspuru-Guzik, and J. L. O’brien, A variational eigenvalue solver on a photonic quantum processor, Nature Communications5, 4213 (2014). 6

2014
[2]

Mitarai, M

K. Mitarai, M. Negoro, M. Kitagawa, and K. Fujii, Quan- tum circuit learning, Physical Review A98, 032309 (2018)

2018
[3]

Benedetti, E

M. Benedetti, E. Lloyd, S. Sack, and M. Fiorentini, Pa- rameterized quantum circuits as machine learning models, Quantum Science and Technology4, 043001 (2019)

2019
[4]

Sagingalieva, M

A. Sagingalieva, M. Kordzanganeh, N. Kenbayev, D. Kosichkina, T. Tomashuk, and A. Melnikov, Hybrid quantum neural network for drug response prediction, Cancers15, 2705 (2023)

2023
[5]

Cerezo, A

M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio, and P. J. Coles, Variational quantum algo- rithms, Nature Reviews Physics3, 625 (2021)

2021
[6]

Melnikov, M

A. Melnikov, M. Kordzanganeh, A. Alodjants, and R.-K. Lee, Quantum machine learning: from physics to software engineering, Advances in Physics: X8, 2165452 (2023)

2023
[7]

Preskill, Quantum computing in the NISQ era and beyond, Quantum2, 79 (2018)

J. Preskill, Quantum computing in the NISQ era and beyond, Quantum2, 79 (2018)

2018
[8]

Bharti, A

K. Bharti, A. Cervera-Lierta, T. H. Kyaw, T. Haug, S. Alperin-Lea, A. Anand, M. Degroote, H. Heimonen, J. S. Kottmann, T. Menke,et al., Noisy intermediate- scale quantum algorithms, Reviews of Modern Physics 94, 015004 (2022)

2022
[9]

Kordzanganeh, M

M. Kordzanganeh, M. Buchberger, B. Kyriacou, M. Po- volotskii, W. Fischer, A. Kurkin, W. Somogyi, A. Sagin- galieva, M. Pflitsch, and A. Melnikov, Benchmarking sim- ulated and physical quantum processing units using quan- tum and hybrid algorithms, Advanced Quantum Tech- nologies6, 2300043 (2023)

2023
[10]

Kuzmin, B

V. Kuzmin, B. Kyriacou, T. Protasevich, M. Papierz, M. Kordzanganeh, and A. Melnikov, TQml simulator: optimized simulation of quantum machine learning, arXiv preprint arXiv:2506.04891 (2025)

work page arXiv 2025
[11]

Bowles, S

J. Bowles, A. Shahnawaz, and M. Schuld, Better than classical? the subtle art of benchmarking quantum ma- chine learning models, arXiv preprint arXiv:2403.07059 (2024)

work page arXiv 2024
[12]

Kandala, A

A. Kandala, A. Mezzacapo, K. Temme, M. Takita, M. Brink, J. M. Chow, and J. M. Gambetta, Hardware- efficient variational quantum eigensolver for small molecules and quantum magnets, Nature549, 242 (2017)

2017
[13]

M´ at´ e, B

B. M´ at´ e, B. L. Saux, and M. Henderson, Beyond Ans¨ atze: Learning quantum circuits as unitary operators, arXiv preprint arXiv:2203.00601 (2022)

work page arXiv 2022
[14]

Arjovsky, A

M. Arjovsky, A. Shah, and Y. Bengio, Unitary evolution recurrent neural networks, inInternational Conference on Machine Learning(PMLR, 2016) pp. 1120–1128

2016
[15]

Lezcano-Casado and D

M. Lezcano-Casado and D. Mart´ ınez-Rubio, Cheap or- thogonal constraints in neural networks: A simple parametrization of the orthogonal and unitary group, in International Conference on Machine Learning(PMLR,
[16]

Kuzmin, W

V. Kuzmin, W. Somogyi, E. Pankovets, and A. Melnikov, Method for noise-induced regularization in quantum neu- ral networks, Advanced Quantum Technologies8, e00603 (2025)

2025
[17]

J. R. McClean, S. Boixo, V. N. Smelyanskiy, R. Bab- bush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nature Communications9, 4812 (2018)

2018
[18]

Cerezo, A

M. Cerezo, A. Sone, T. Volkoff, L. Cincio, and P. J. Coles, Cost function dependent barren plateaus in shallow parametrized quantum circuits, Nature Communications 12, 1791 (2021)

2021
[19]

Ragone, B

M. Ragone, B. N. Bakalov, F. Sauvage, A. F. Kemper, C. O. Marrero, M. Larocca, and M. Cerezo, A Lie algebraic theory of barren plateaus for deep parameterized quantum circuits, Nature Communications15, 7172 (2024)

2024
[20]

Holmes, K

Z. Holmes, K. Sharma, M. Cerezo, and P. J. Coles, Con- necting ansatz expressibility to gradient magnitudes and barren plateaus, PRX Quantum3, 010313 (2022)

2022
[21]

S. Sim, P. D. Johnson, and A. Aspuru-Guzik, Express- ibility and entangling capability of parameterized quan- tum circuits for hybrid quantum-classical algorithms, Ad- vanced Quantum Technologies2, 1900070 (2019)

2019
[22]

Patapovich, M

V. Patapovich, M. Periyasamy, M. Kordzanganeh, and A. Melnikov, Superposed parameterised quantum circuits, arXiv preprint arXiv:2506.08749 (2025)

work page arXiv 2025
[23]

M. A. Nielsen and I. L. Chuang,Quantum computation and quantum information(Cambridge University Press, 2010)

2010
[24]

V. V. Shende, S. S. Bullock, and I. L. Markov, Synthesis of quantum logic circuits, inProceedings of the 2005 Asia and South Pacific Design Automation Conference(2005) pp. 272–275

2005
[25]

Khatri, R

S. Khatri, R. LaRose, A. Poremba, L. Cincio, A. T. Sorn- borger, and P. J. Coles, Quantum-assisted quantum com- piling, Quantum3, 140 (2019)

2019
[26]

Schuld, R

M. Schuld, R. Sweke, and J. J. Meyer, Effect of data encoding on the expressive power of variational quantum- machine-learning models, Physical Review A103, 032430 (2021)

2021
[27]

Kordzanganeh, P

M. Kordzanganeh, P. Sekatski, L. Fedichkin, and A. Mel- nikov, An exponentially-growing family of universal quan- tum circuits, Machine Learning: Science and Technology 4, 035036 (2023)

2023
[28]

V. Mnih, K. Kavukcuoglu, D. Silver, A. Graves, I. Antonoglou, D. Wierstra, and M. Riedmiller, Playing Atari with deep reinforcement learning, arXiv preprint arXiv:1312.5602 (2013)

work page internal anchor Pith review arXiv 2013
[29]

Jerbi, C

S. Jerbi, C. Gyurik, S. C. Marshall, H. J. Briegel, and V. Dunjko, Parametrized quantum policies for reinforce- ment learning, inAdvances in Neural Information Pro- cessing Systems, Vol. 34 (2021) pp. 28362–28375

2021
[30]

Skolik, S

A. Skolik, S. Jerbi, and V. Dunjko, Quantum agents in the Gym: a variational quantum algorithm for deep Q- learning, Quantum6, 720 (2022)

2022
[31]

Lockwood and M

O. Lockwood and M. Si, Playing Atari with hybrid quantum-classical reinforcement learning, inNeurIPS 2020 Workshop on Pre-registration in Machine Learn- ing(PMLR, 2021) pp. 285–301

2020
[32]

Kordzanganeh, D

M. Kordzanganeh, D. Kosichkina, and A. Melnikov, Par- allel hybrid networks: an interplay between quantum and classical neural networks, Intelligent Computing2, 0028 (2023)

2023
[33]

Kurkin, J

A. Kurkin, J. Hegemann, M. Kordzanganeh, and A. Mel- nikov, Forecasting steam mass flow in power plants using the parallel hybrid network, Engineering Applications of Artificial Intelligence160, 111912 (2025)

2025
[34]

Z. Li, N. Kovachki, K. Azizzadenesheli, B. Liu, K. Bhat- tacharya, A. Stuart, and A. Anandkumar, Fourier neural operator for parametric partial differential equations, in International Conference on Learning Representations (2021)

2021