arxiv: 2604.11578 · v1 · submitted 2026-04-13 · 🪐 quant-ph · cs.AI· cs.LG· stat.ML

Recognition: unknown

Minimizing classical resources in variational measurement-based quantum computation for generative modeling

Arunava Majumder , Hendrik Poulsen Nautrup , Hans J. Briegel

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:49 UTC · model grok-4.3

classification 🪐 quant-ph cs.AIcs.LGstat.ML

keywords variational MBQCgenerative modelingquantum channelsmeasurement-based quantum computationunitary circuitsprobability distributionsclassical resourcesquantum machine learning

0 comments

The pith

Restricting variational measurement-based quantum computation to one extra classical parameter per layer generates probability distributions unreachable by the corresponding unitary model.

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

The paper introduces a restricted form of variational MBQC that converts a unitary circuit into a quantum channel model by adding only a single trainable parameter instead of doubling the total count. This minimal change incorporates the randomness from uncorrected qubit measurements to expand the set of representable probability distributions. The authors prove algebraically and confirm numerically that the restricted model can learn certain distributions that the unitary version fundamentally cannot. By cutting the classical parameter overhead, the approach makes optimization more practical while retaining the generative advantage that comes from treating measurement outcomes as part of the variational family.

Core claim

A restricted variational measurement-based quantum computation model extends the unitary setting to a channel-based one using only a single additional trainable parameter. This minimal extension is sufficient to generate probability distributions that cannot be learned by the corresponding unitary model, as established through both numerical training experiments on example distributions and algebraic arguments showing the expressivity separation.

What carries the argument

The restricted VMBQC channel model, which incorporates measurement-induced randomness by adding exactly one trainable parameter to the unitary variational circuit of width N and depth D.

If this is right

Generative modeling tasks can exploit measurement randomness in MBQC with far lower classical resource cost than a full channel parameterization.
Optimization landscapes become more tractable because the parameter count scales only as N times D plus one instead of twice that.
The separation in representable distributions arises specifically from the probabilistic channel behavior introduced by uncorrected measurements.
The algebraic proof technique used to show the expressivity gap can be applied to other restricted quantum circuit families.

Where Pith is reading between the lines

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

Similar single-parameter restrictions could be tested in other variational quantum algorithms that currently suffer from over-parameterization when moving from unitary to channel descriptions.
The result suggests that the generative advantage may not require the full freedom of arbitrary quantum channels, only a targeted form of randomness injection.
Practical implementations on near-term hardware might prioritize this minimal extension to balance expressivity against trainability.

Load-bearing premise

The chosen restriction on how the single extra parameter controls measurement randomness must preserve enough expressivity to reach distributions beyond the unitary model's reach without creating new training problems.

What would settle it

An explicit probability distribution (or family of distributions) that the unitary model cannot represent but where the restricted single-parameter model also fails to match the target statistics even after full optimization.

Figures

Figures reproduced from arXiv: 2604.11578 by Arunava Majumder, Hans J. Briegel, Hendrik Poulsen Nautrup.

**Figure 1.** Figure 1: Quantum circuit Born machines with VMBQC: (a) Target distribution Y (x) (usually unknown), from which samples are given, x ∼ Y . In this paper, the target distribution is generated from the VMBQC model itself with a randomly chosen set of parameters {θ, p} on the cluster state with some given width (N) and depth (D). (b) The learning model, with width N and depth D, is parametrized by measurement angles θ … view at source ↗

**Figure 3.** Figure 3: Propagation of byproducts in E˜c(θ, p) model: Top: Consider a byproduct Z s 1 3 3 on qubit 3 at position (3, 1). Propagating this through the cluster state via repeated stabilizer relations generates X s 1 3 i -byproducts on qubits i within the forward light-cone (green), which flips (if s 1 3 = 1) the measurement angles of the qubits. For instance, the angle θ 2 3 at position (3, 2) is transformed to (−… view at source ↗

**Figure 4.** Figure 4: VMBQC models with one correction probability: We consider four variants of the VMBQC model in a restricted setting, where the number of trainable correction probabilities is limited to a single parameter, i.e., p := {p j i = p}i,j . (a) Channel model E (A) c corresponding to Eq. (8) where measurement outcomes of all qubits (blue) are partially adapted and all are sharing the same correction probability p <… view at source ↗

**Figure 5.** Figure 5: Learning performances of VMBQC models: Four families of channel models, defined in Eqs. (8)–(11), together with the unitary model in Eq. (1), are trained to approximate a target distribution generated by the channel Ec(θt, pt) in Eq. (5). Top row: The target model has width N = 7 and depth D = 6, and generates target samples x ∼ Y(θt,pt)(x) according to the born rule (in blue box). The parameters θt = {θ j… view at source ↗

**Figure 6.** Figure 6: Effect of a single byproduct: Here we examine how well the unitary model Uc can learn the channel model E˜ (D) c (θ, p) when only the measurement outcome of a single qubit is probabilistically adapted (see Eq. (15)). The x-axis labels the different target models, while the y-axis reports the distribution of minimum MMD losses achieved by 10 independently initialized unitary learning models Uc over 200 epoc… view at source ↗

**Figure 7.** Figure 7: Effect of two byproducts: Here we examine how well the unitary model Uc can learn the channel model E˜c(θ, p) with two partially uncorrected qubits whose measurement outcomes are probabilistically adapted. The x-axis labels the different target models, while the y-axis reports the distribution of minimum MMD losses achieved by 10 independent randomly initialized unitary models Uc over 200 epochs. (a) Botto… view at source ↗

read the original abstract

Measurement-based quantum computation (MBQC) is a framework for quantum information processing in which a computational task is carried out through one-qubit measurements on a highly entangled resource state. Due to the indeterminacy of the outcomes of a quantum measurement, the random outcomes of these operations, if not corrected, yield a variational quantum channel family. Traditionally, this randomness is corrected through classical processing in order to ensure deterministic unitary computations. Recently, variational measurement-based quantum computation (VMBQC) has been introduced to exploit this measurement-induced randomness to gain an advantage in generative modeling. A limitation of this approach is that the corresponding channel model has twice as many parameters compared to the unitary model, scaling as $N \times D$, where $N$ is the number of logical qubits (width) and $D$ is the depth of the VMBQC model. This can often make optimization more difficult and may lead to poorly trainable models. In this paper, we present a restricted VMBQC model that extends the unitary setting to a channel-based one using only a single additional trainable parameter. We show, both numerically and algebraically, that this minimal extension is sufficient to generate probability distributions that cannot be learned by the corresponding unitary model.

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

This paper shows that one extra trainable parameter on top of unitary VMBQC is enough to reach some generative distributions the unitary model cannot, with algebra and numerics to support it.

read the letter

The punchline is that the authors define a restricted channel version of variational measurement-based quantum computation that adds only a single parameter while still producing probability distributions outside the reach of the corresponding unitary model. They give both an algebraic argument and numerical evidence on generative tasks to back the claim. This is the concrete new piece: a minimal extension that keeps classical overhead low instead of jumping to the full N by D channel parameters used in earlier VMBQC work.

Referee Report

2 major / 2 minor

Summary. The paper claims to introduce a restricted variational measurement-based quantum computation (VMBQC) model for generative modeling. By adding only a single additional trainable parameter to the unitary VMBQC model, it constructs a variational channel family and asserts, via algebraic proof and numerical experiments, that this minimal extension suffices to generate probability distributions unreachable by the corresponding unitary model, thereby reducing classical parameter overhead.

Significance. If verified for the single-parameter restriction, this provides a concrete method to achieve non-unitary generative advantages in MBQC with minimal added resources, potentially easing optimization difficulties noted for full channel models. The algebraic demonstration of expressivity gain from one parameter could serve as a template for other variational quantum algorithms balancing expressivity and trainability.

major comments (2)

Abstract: The central claim that one additional parameter suffices requires the algebraic argument to derive non-unitary distributions using solely this single parameter (rather than the full N×D channel parameters). The abstract does not specify the restriction form, so the proof must be shown to apply exactly to the restricted model as per the stress-test concern; otherwise the advantage may not hold for the minimal extension.
Numerical experiments section: The numerics must optimize and evaluate the restricted single-parameter model on the tested distributions while confirming the unitary model cannot reach them. If the experiments instead use the unrestricted channel, they do not support the claim that the minimal one-parameter extension is sufficient.

minor comments (2)

Introduction: Clarify early the exact parameter scaling (unitary vs. restricted channel) with a side-by-side comparison to make the resource minimization explicit.
Throughout: Ensure all equations defining the restricted channel are labeled and cross-referenced so readers can trace how the single parameter enters the measurement outcomes.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful review and constructive suggestions. We address each major comment below and will revise the manuscript for improved clarity on the single-parameter restriction.

read point-by-point responses

Referee: Abstract: The central claim that one additional parameter suffices requires the algebraic argument to derive non-unitary distributions using solely this single parameter (rather than the full N×D channel parameters). The abstract does not specify the restriction form, so the proof must be shown to apply exactly to the restricted model as per the stress-test concern; otherwise the advantage may not hold for the minimal extension.

Authors: We agree that the abstract would benefit from explicitly stating the single-parameter restriction. The algebraic proof (Section 3) is derived specifically for the restricted model with one additional trainable parameter, showing via direct construction that this minimal extension generates distributions unreachable by the unitary VMBQC. We will revise the abstract to specify the restriction form and reference the relevant section. revision: yes
Referee: Numerical experiments section: The numerics must optimize and evaluate the restricted single-parameter model on the tested distributions while confirming the unitary model cannot reach them. If the experiments instead use the unrestricted channel, they do not support the claim that the minimal one-parameter extension is sufficient.

Authors: The numerical experiments (Section 4) optimize the restricted single-parameter model and explicitly demonstrate that it reaches distributions the unitary model cannot. We will add clarifying text and pseudocode in the revised version to confirm that only the single additional parameter is trained in the channel experiments, with direct comparisons to the unitary baseline. revision: yes

Circularity Check

0 steps flagged

No significant circularity; algebraic and numerical claims appear independent of inputs

full rationale

The paper defines a one-parameter restriction on the VMBQC channel model and asserts that algebraic derivation plus numerical experiments establish generative distributions unreachable by the unitary model. No load-bearing step reduces by the paper's own equations to a fitted parameter renamed as prediction, nor to a self-citation chain. The abstract and description present the algebraic argument as deriving the claimed separation directly from the restricted model without presupposing the target distributions, and the numerical validation is described as testing the restricted model against the unitary baseline. This satisfies the criteria for a self-contained derivation with independent content.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The model rests on standard quantum mechanics and MBQC resource-state assumptions plus the variational optimization framework; the single extra parameter is the only new trainable element introduced.

free parameters (1)

single additional trainable parameter
The minimal channel extension is controlled by one extra trainable scalar whose value is determined by optimization.

axioms (1)

standard math Standard assumptions of quantum mechanics and the MBQC framework hold, including the existence of suitable resource states and the validity of one-qubit measurements.
The entire construction is built on established MBQC theory.

pith-pipeline@v0.9.0 · 5529 in / 1181 out tokens · 34165 ms · 2026-05-10T15:49:42.240802+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

36 extracted references · 6 canonical work pages

[1]

III A 2), with randomly initialized parameters, and then train different learning models (described in Sec

Training process First, we generate a dataset with samples drawn from the output distribution of the target model (described in Sec. III A 2), with randomly initialized parameters, and then train different learning models (described in Sec. III A 3) to learn this target distribution via samples using the MMD loss in Eq. (12). All learning models are train...
[2]

(5) withθ=θ t andp=p t, as the target

Target model We randomly select an instantiation of the channel model Ec(θt,p t), defined in Eq. (5) withθ=θ t andp=p t, as the target. In particular, for Fig. 5, we chooseN= 7and D= 6for the target model. The target is initialized with 8 measurement anglesθ t :={θ j i ∈[0,2π)} i,j and correction probabilitiesp t :={p j i ∈[0.9,1]} i,j, both drawn uniform...
[3]

(1)) and four families of channel models defined in Eqs

Learning models The VMBQC learning models considered in this work in- clude a purely unitary model (Eq. (1)) and four families of channel models defined in Eqs. (8)–(11) (Fig. 4 (a-d)), which differ only in the placement of the partially uncorrected qubits on the cluster. We numerically compare the learning perfor- mances of these channel models with that...
[4]

5 (middle row, panels (a)-(d)), we present the av- eraged learning curves comparing four families of channel models with the purely unitary model

Analysis In Fig. 5 (middle row, panels (a)-(d)), we present the av- eraged learning curves comparing four families of channel models with the purely unitary model. In all cases, the mod- els are trained to learn the same target distribution gener- ated by another channel model defined in Eq. (5). Across all panels, a significant separation is observed bet...
[5]

#",# Sampling targetparameters{𝑥∼𝑌%!,&!𝑥}Learning target distribution10.50 10.50 10.50 10.50 Target Model 𝑝!=𝑝

Model with one byproduct We now demonstrate the effect of a single byproduct, aris- ing from one partially adapted qubit, on both the output dis- tribution of the VMBQC channel model and the learning per- formance of a purely unitary model. Specifically, we use the channel model ˜Ec(θ,p)defined in Eq. (7) as the target. This model distills the non-Cliffor...
[6]

This publication was also made possible through financial support of WOST (WithOutSpaceTime) grant from the John Templeton Foundation

Model with two byproducts In this section, we extend the previous analysis with byproducts at two different positions but with same correction probability and their impact on the output distribution of the ˜Ec(θ,p)model. Similar to Sec. III B 1, we conduct a series of simulations with four different target models (Fig. 7 (a-d)), and the results follow a t...

work page doi:10.55776/f71
[7]

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

2010
[8]

Arute, K

F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A. Buell, et al., Quantum supremacy using a programmable supercon- ducting processor, nature574, 505 (2019)

2019
[9]

A. J. Daley, I. Bloch, C. Kokail, S. Flannigan, N. Pearson, M. Troyer, and P. Zoller, Practical quantum advantage in quan- tum simulation, Nature607, 667 (2022)

2022
[10]

Observation of constructive interference at the edge of quantum ergodicity, Nature646, 825 (2025)

2025
[11]

P. W. Shor, Polynomial-time algorithms for prime factorization and discrete logarithms on a quantum computer, SIAM review 41, 303 (1999)

1999
[12]

H. J. Briegel, D. E. Browne, W. D ¨ur, R. Raussendorf, and M. Van den Nest, Measurement-based quantum computation, Nature Physics5, 19 (2009)

2009
[13]

Raussendorf and H

R. Raussendorf and H. J. Briegel, A one-way quantum com- puter, Phys. Rev. Lett.86, 5188 (2001)

2001
[14]

Raussendorf, D.E

R. Raussendorf, D. E. Browne, and H. J. Briegel, Measurement- based quantum computation on cluster states, Physical Review A68, 10.1103/physreva.68.022312 (2003)

work page doi:10.1103/physreva.68.022312 2003
[15]

H. J. Briegel and R. Raussendorf, Persistent entanglement in arrays of interacting particles, Physical Review Letters86, 910 (2001)

2001
[16]

Huang, M

H.-Y . Huang, M. Broughton, N. Eassa, H. Neven, R. Babbush, and J. R. McClean, Generative quantum advantage for classi- cal and quantum problems, arXiv preprint arXiv:2509.09033 (2025)

work page arXiv 2025
[17]

Majumder, M

A. Majumder, M. Krumm, T. Radkohl, L. J. Fiderer, H. P. Nautrup, S. Jerbi, and H. J. Briegel, Variational measurement- based quantum computation for generative modeling, Physical Review A110, 062616 (2024)

2024
[18]

Coyle, S

B. Coyle, S. Raj, N. Mathur, E. A. Cherrat, N. Jain, S. Kazdaghli, and I. Kerenidis, Training-efficient density quan- tum machine learning, npj Quantum Information11, 172 (2025)

2025
[19]

Kurkin, K

A. Kurkin, K. Shen, S. Pielawa, H. Wang, and V . Dunjko, Note on the universality of parameterized iqp circuits with hid- den units for generating probability distributions, arXiv preprint arXiv:2504.05997 (2025)

work page arXiv 2025
[20]

Liu and L

J.-G. Liu and L. Wang, Differentiable learning of quantum cir- cuit born machines, Physical Review A98, 062324 (2018)

2018
[21]

Benedetti, D

M. Benedetti, D. Garcia-Pintos, O. Perdomo, V . Leyton-Ortega, Y . Nam, and A. Perdomo-Ortiz, A generative modeling ap- proach for benchmarking and training shallow quantum cir- cuits, npj Quantum Information5, 45 (2019)

2019
[22]

J. M. Tomczak,Deep Generative Modeling(Springer Interna- tional Publishing, 2022)

2022
[23]

J. R. McClean, S. Boixo, V . N. Smelyanskiy, R. Babbush, and H. Neven, Barren plateaus in quantum neural network training landscapes, Nature communications9, 4812 (2018)

2018
[24]

S. Wang, E. Fontana, M. Cerezo, K. Sharma, A. Sone, L. Cin- cio, and P. J. Coles, Noise-induced barren plateaus in variational quantum algorithms, Nature communications12, 6961 (2021)

2021
[25]

Learning in Implicit Generative Models

S. Mohamed and B. Lakshminarayanan, Learning in im- plicit generative models, arXiv preprint arXiv:1610.03483 https://doi.org/10.48550/arXiv.1610.03483 (2016)

work page Pith review doi:10.48550/arxiv.1610.03483 2016
[26]

Measurement-based quantum com- putation from clifford quantum cellular au- tomata

H. Poulsen Nautrup and H. J. Briegel, Measurement-based quantum computation from clifford quantum cellular automata, arXiv preprint arXiv:2312.13185 (2023)

work page arXiv 2023
[27]

Mantri, T

A. Mantri, T. F. Demarie, and J. F. Fitzsimons, Universality of quantum computation with cluster states and (x, y)-plane mea- surements, Scientific Reports7, 42861 (2017)

2017
[28]

D. T. Stephen, H. P. Nautrup, J. Bermejo-Vega, J. Eisert, and R. Raussendorf, Subsystem symmetries, quantum cellular au- tomata, and computational phases of quantum matter, Quantum 3, 142 (2019)

2019
[29]

Raussendorf, C

R. Raussendorf, C. Okay, D.-S. Wang, D. T. Stephen, and H. Poulsen Nautrup, Computationally universal phase of quan- tum matter, Phys. Rev. Lett.122, 090501 (2019)

2019
[30]

Raussendorf, Quantum computation via translation-invariant operations on a chain of qubits, Phys

R. Raussendorf, Quantum computation via translation-invariant operations on a chain of qubits, Phys. Rev. A72, 052301 (2005)

2005
[31]

Cerezo, A

M. Cerezo, A. Arrasmith, R. Babbush, S. C. Benjamin, S. Endo, K. Fujii, J. R. McClean, K. Mitarai, X. Yuan, L. Cincio,et al., Variational quantum algorithms, Nature Reviews Physics3, 625 (2021)

2021
[32]

Gretton, K

A. Gretton, K. M. Borgwardt, M. J. Rasch, B. Sch ¨olkopf, and A. Smola, A kernel two-sample test, The Journal of Machine Learning Research13, 723 (2012)

2012
[33]

Majumder, VMBQC 1p (2026)

A. Majumder, VMBQC 1p (2026). 13

2026
[34]

Duchi, E

J. Duchi, E. Hazan, and Y . Singer, Adaptive subgradient meth- ods for online learning and stochastic optimization., Journal of machine learning research12(2011). A. Gradients of loss function For training purposes, we require calculating the gradient of the loss function in Eq. 12. The resulting gradient with respect to a variational measurement angleθ j...

2011
[35]

Learning quantum channel distribution In this appendix, we provide the training details corre- sponding to the results presented in Sec. III A. We employ gradient-based optimization to train the learning models, us- ing the gradients derived in Eqs. (A1) and (A2). Specifically, both the variational measurement anglesθand the correction probabilitiespare u...
[36]

III B 1 and III B 2

Effect of byproducts on the learning performance This appendix provides the training details corresponding to the results presented in Secs. III B 1 and III B 2. In contrast to Sec. III A, here we fix the learning model to be the unitary 14 VMBQCU c in Eq. (1) and evaluate its ability to learn a variety of target models. For each target model, the paramet...