Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing

arXiv:2509.15486 · quant-ph · Submitted 2025-09-18 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Quantum Radio. Generated commentary on the latest quantum physics and condensed matter papers.

Kai: We're looking at the paper: "Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing".

Mira: Here's the gist of the paper. The gist: GFlowNets introduce a Generative Flow Networks-based model that colors graph representations of molecular qubit Hamiltonians to sample non-overlapping commuting groupings, which can reduce measurement requirements and two-qubit gate counts in quantum chemical simulations.

Kai: Measurement Optimization Challenge Quantum chemical simulations on quantum computers are limited by a measurement bottleneck: estimating molecular Hamiltonians to chemical accuracy requires a large number of measurements, which remain costly even on fault-tolerant devices The paper addresses this by formulating Hamiltonian grouping as a reward-driven generative search problem and introducing GFlowNets to sample non-overlapping commuting groupings GFlowNets Framework GFlowNets are a class of generative models over a directed acyclic graph (DAG) G = (S, A) of states with initial state s0 and terminal set Sf The model learns the probability of sampling a compositional object sf ∈ Sf, where the probability of reaching it is proportional to a positive reward function R(sf) This is achieved by parameterizing PF and PB so that the conservation law holds at all internal states and the boundary matches the reward at terminals The training objective uses the trajectory balance (TB) objective, which enforces flow-consistency over the entire trajectory in a single step, minimizing L(τ; θ) = log Zθ + Xn t=1 log PF,θ(stst−1) − log R(sf) − Xn t=1 log PB,θ(st−1st) Reward Function Design The total reward R(x) is defined as a weighted sum of three components to enable multi-objective optimization without differentiable cost functions or model pretraining R(x) = λ0 RM(x) + λ1 RG(x) + λ2 RN2q(x) The measurement term RM(x) is given by 1/ε squared M(x), where M(x) is the total number of measurements needed to achieve a target accuracy ε, and the regularization term RG(x) penalizes excessive groupings, defined as NP − NG(x), where NG = max color(x) The third term RN2q(x) takes into account the total number of two-qubit gates (N2q), given by 1/N2q(x) Performance and Initialization Across molecular Hamiltonian benchmarks, the GFlowNets sampler finds fully commuting non-overlapping groupings with average measurement requirements 18% lower than SI, and produces Pareto sets that expose trade-offs among shots, circuits, and two-qubit resources When used to initialize iterative coefficient splitting (ICS), GFlowNet-generated groupings reduce post-ICS measurement estimates by up to 40% for Jordan-Wigner-mapped fully commuting Hamiltonians relative to SI initialization The results show that GFlowNets provide a flexible workflow for resource-aware measurement and quantum-resource optimization in quantum chemistry, replacing single-heuristic outputs with diverse candidate groupings that can be selected according to hardware-specific priorities Sorted GFlowNets Variant A variant called sorted GFlowNets (S-GFN) was designed by ordering the Pauli operators in terms of the absolute value of the coefficients ωk, in a similar fashion to SI, and this variant performed remarkably well owing to the covariance distributions of most of the systems studied here Table 2 shows that for FC grouping and JW mapping, S-GFN consistently falls below the SI baseline, even for molecules and settings where our initial formulation struggled The proposed variant for JW-mapped Hamiltonians in the FC setting yields an average 18.5% reduction in measurement requirements for non-overlapping groupings compared to SI, with maximum reductions of 33% Composite Reward Analysis By considering composite rewards that account for both measurement and the number of groups, the study illustrates how the reward weights modify the sampled trade-off between measurement count and the number of groups For H4, LiH, and BeH2, solutions containing 8 groups were consistently found in a grouping-heavy regime when λ0 was small Furthermore, for composite rewards including two-qubit gates (RN2q(x)), the GFN-ICS approach showed a reduction of 85 two-qubit gates for the implemented circuits at λ0 = 100 The central conceptual contribution is a principled mechanism for sampling multiple candidate non-overlapping groupings under a specified reward instead of committing to a single heuristic output The diversity of samples, inherent to GFlowNets and possibly other generative-based models, is shown to be useful for candidate ICS initialization selection and trade-off analysis in the tested systems The results indicate that the sampled candidates have structurally distinct measurement/two-qubit trade-offs and can improve downstream ICS initialization beyond selecting the single best non-overlapping grouping Future Directions Two main limitations to scaling this class of algorithms are the trajectory length τ and the cost of evaluating the reward function Future developments may include introducing a pre-training phase where the model is “warmed up” using circuits generated by heuristic approaches such as RLF Leveraging precomputed covariance dictionaries from classical, efficient methods may further enhance learning efficiency and reduce the computational overhead of reward evaluation The potential to generate transferable models through our sampling strategy is also noted as a direction for future work Acknowledgments This research was enabled by support from the Digital Research Alliance of Canada and NSERC Discovery Grant No.

Mira: RGPIN-2024-06594 The code for our GFlowNets implementation is available at https://github.com/ChemAI-Lab/GFlowNetsMOpt Data Availability The code for our GFlowNets implementation is available at https://github.com/ChemAI-Lab/GFlowNetsMOpt The sampled graphs and outputs for the FC groupings obtained with the GINEw model in 21, as well as the sampled graphs, outputs, and models for the measurement plus two-qubit gate count composite-reward results in Table 4, are available through Zenodo at doi.org/10.5281/zenodo.21050287 References

1: Peruzzo, A.; McClean, J.; Shadbolt, P.; Yung, M.-H.; Zhou, X.-Q.; Love, P.

Kai: J.; Aspuru-Guzik, A.; O’Brien, J.

Title and authors: Kai: So to get into this, the title is "Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing". It sounds a bit like a lot of math, but it’s pointing to the method they used.

Mira: They are using discrete flow-based generative models, which means they’re not just guessing; they have a structured way to generate potential groupings based on how the states connect in some graph structure.

Lev: The authors are Huidobro-Meezs, Dai, and Vargas-Hernández. They come from chemistry and operations research at McMaster and the University of Montréal. That mix is important because it grounds the quantum problem in a physical context while using operations research tools for the search.

Kai: So what does this title actually mean for us? It means they are moving beyond just finding one good grouping, which is what methods like SI do, and instead generating a whole family of candidates that represent different resource trade-offs.

Mira: It’s about structured generation to solve a search problem that has many competing objectives simultaneously.

Lev: They are trying to find groupings—like fully commuting or qubit-wise commuting ones—that reduce measurement overhead while keeping the resulting circuit depth manageable, which is a real headache for us on current hardware.

The paper's summary: Kai: So the summary explains that they frame Hamiltonian grouping as a generative search problem, and GFlowNets are the tools they introduce to learn policies that sample high-quality groupings with probability proportional to some task-specific reward.

Mira: They build this GFlowNet over a directed acyclic graph of states, and the probability of sampling a certain grouping is linked to a reward function R(sf). The training objective uses something called trajectory balance, which enforces flow consistency across the entire path in one step.

Lev: That trajectory balance thing sounds like they are ensuring that as you move from one state to the next in this graph, the flow respects some physical conservation law while simultaneously trying to hit that reward target at the end.

Kai: And this reward function R(x) is defined as a weighted sum of three parts: a measurement term, a grouping penalty term, and a two-qubit gate count term.

Mira: That's the key part for me; they explicitly put in terms for circuit costs, not just measurement counts. They have R(x) = lambda0 * M(x) plus lambda1 * RG(x) plus lambda2 * RN2q(x).

Lev: So if we tune those lambdas, we can tell the model whether it should prioritize reducing measurements, reducing the number of groups, or minimizing two-qubit gates. That’s a lot of control.

The paper's improvements: Kai: The paper shows that this GFlowNets approach finds fully commuting non-overlapping groupings with average measurement requirements eighteen percent lower than SI methods on molecular Hamiltonian benchmarks <ref:2509.15486#pg2>.

Mira: And it doesn't stop there; they produce Pareto sets that are really useful because they expose the trade-offs between shots, circuits, and two-qubit resources when you look at those different candidates.

Lev: But the real impact is when you use these GFlowNets to initialize iterative coefficient splitting or ICS methods. They found that GFlowNet-generated groupings reduce post-ICS measurement estimates by up to forty percent for Jordan-Wigner-mapped fully commuting Hamiltonians compared to SI initialization.

Kai: That forty percent reduction in measurement estimate is substantial, especially when you compare it back to the initial cost of setting up those measurements.

Mira: They also have a specific variant called sorted GFlowNets, S-GFN, which they ordered Pauli operators by their coefficient absolute values and that actually performed really well across most systems studied here.

Lev: And for the S-GFN variant in the fully commuting grouping setting with Jordan-Wigner mapping, they saw an average eighteen point five percent reduction in measurement requirements compared to SI, with maximum reductions hitting thirty-three percent.

Conclusion: Kai: So looking at the conclusion of "Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing", the main thing is that this framework provides a principled mechanism for sampling multiple candidate non-overlapping groupings instead of committing to just one heuristic output.

Mira: They argue that the diversity inherent in GFlowNets is useful because it lets you select initialization candidates based on what you need—whether it’s minimizing measurement shots or minimizing two-qubit gate counts.

Lev: That diversity means that even if the single best grouping isn't the one with the absolute minimum measurements, sampling from this set helps improve post-ICS measurement estimates by up to fifteen percent for certain molecular systems <ref:2509.15486#pg1>.

Kai: It really gives us a flexible workflow for resource-aware optimization in quantum chemistry, letting us choose groupings based on hardware priorities instead of just following a single heuristic output.

Mira: It’s about moving from finding one answer to having a set of diverse answers that map out the trade-offs.

Lev: So, this paper shows how generative models can be used to discover structural differences in measurement and two-qubit resource usage that standard greedy algorithms might miss.

Isaac L. Huidobro-Meezs, Jun Dai, Rodrigo A. Vargas-Hernáez

Department of Chemistry and Chemical Biology, McMaster University · Mila - Quebec AI Institute, Université de Montréal · Département d’informatique et de recherche opérationnelle, Université de Montréal · Brockhouse Institute for Materials Research, McMaster University

quant-ph

Submitted: 2025-09-18

Updated: 2026-08-25

Comments: 26 pages, 9 figures and 4 tables. Contact authors for SM

DOI: 10.1039/d6dd00430j

Code: https://github.com/ChemAI-Lab/GFlowNetsMOpt

License: http://creativecommons.org/licenses/by/4.0/

The gist: The gist: GFlowNets introduce a Generative Flow Networks-based model that colors graph representations of molecular qubit Hamiltonians to sample non-overlapping commuting groupings, which can reduce

Key concepts

GFlowNets Framework: GFlowNets is a generative model built over a directed acyclic graph (DAG) of states. It learns the probability of sampling a specific grouping by maximizing a reward function that enforces flow consistency across the entire trajectory, ensuring the sampled groupings are physically valid and reachable.
Reward Function Design: The reward function R(x) combines three objectives: minimizing measurement cost (RM), penalizing excessive groupings (RG), and accounting for two-qubit gate count (RN2q). By adjusting the weights ($\lambda_0, \lambda_1, \lambda_2$), users can tune the search to prioritize different trade-offs between these competing resources.
Non-overlapping Commuting Groupings: The model identifies sets of Pauli operators that commute with each other and do not overlap. Finding these groupings is crucial because they allow for fewer measurements and fewer two-qubit gates during quantum chemical simulations, directly reducing the computational bottleneck.
Sorted GFlowNets Variant (S-GFN): This variant orders Pauli operators based on the absolute value of their coefficients. It performs well because it leverages the covariance distributions of many molecular systems, consistently outperforming standard models like SI in finding better groupings for Jordan-Wigner mappings.

Terminology

Accuracy — 1× in this paper · explained in 1 episode
Adaptation — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Generative Model — 1× in this paper · explained in 2 episodes
Generative Models — 1× in this paper · explained in 3 episodes
Multi-objective Optimization — 1× in this paper · explained in 5 episodes
Reward Function — 1× in this paper · explained in 4 episodes

Summary

The gist: GFlowNets introduce a Generative Flow Networks-based model that colors graph representations of molecular qubit Hamiltonians to sample non-overlapping commuting groupings, which can reduce measurement requirements and two-qubit gate counts in quantum chemical simulations.

Measurement Optimization Challenge

Quantum chemical simulations on quantum computers are limited by a measurement bottleneck: estimating molecular Hamiltonians to chemical accuracy requires a large number of measurements, which remain costly even on fault-tolerant devices The paper addresses this by formulating Hamiltonian grouping as a reward-driven generative search problem and introducing GFlowNets to sample non-overlapping commuting groupings

GFlowNets Framework

GFlowNets are a class of generative models over a directed acyclic graph (DAG) G = (S, A) of states with initial state s0 and terminal set Sf The model learns the probability of sampling a compositional object sf ∈ Sf, where the probability of reaching it is proportional to a positive reward function R(sf) This is achieved by parameterizing PF and PB so that the conservation law holds at all internal states and the boundary matches the reward at terminals The training objective uses the trajectory balance (TB) objective, which enforces flow-consistency over the entire trajectory in a single step, minimizing L(τ; θ) = log Zθ + Xn t=1 log PF,θ(stst−1) − log R(sf) − Xn t=1 log PB,θ(st−1st)

Reward Function Design

The total reward R(x) is defined as a weighted sum of three components to enable multi-objective optimization without differentiable cost functions or model pretraining R(x) = λ0 RM(x) + λ1 RG(x) + λ2 RN2q(x) The measurement term RM(x) is given by 1/ε squared M(x), where M(x) is the total number of measurements needed to achieve a target accuracy ε, and the regularization term RG(x) penalizes excessive groupings, defined as NP − NG(x), where NG = max color(x) The third term RN2q(x) takes into account the total number of two-qubit gates (N2q), given by 1/N2q(x)

Performance and Initialization

Across molecular Hamiltonian benchmarks, the GFlowNets sampler finds fully commuting non-overlapping groupings with average measurement requirements 18% lower than SI, and produces Pareto sets that expose trade-offs among shots, circuits, and two-qubit resources When used to initialize iterative coefficient splitting (ICS), GFlowNet-generated groupings reduce post-ICS measurement estimates by up to 40% for Jordan-Wigner-mapped fully commuting Hamiltonians relative to SI initialization The results show that GFlowNets provide a flexible workflow for resource-aware measurement and quantum-resource optimization in quantum chemistry, replacing single-heuristic outputs with diverse candidate groupings that can be selected according to hardware-specific priorities

Sorted GFlowNets Variant

A variant called sorted GFlowNets (S-GFN) was designed by ordering the Pauli operators in terms of the absolute value of the coefficients ωk, in a similar fashion to SI, and this variant performed remarkably well owing to the covariance distributions of most of the systems studied here Table 2 shows that for FC grouping and JW mapping, S-GFN consistently falls below the SI baseline, even for molecules and settings where our initial formulation struggled The proposed variant for JW-mapped Hamiltonians in the FC setting yields an average 18.5% reduction in measurement requirements for non-overlapping groupings compared to SI, with maximum reductions of 33%

Composite Reward Analysis

By considering composite rewards that account for both measurement and the number of groups, the study illustrates how the reward weights modify the sampled trade-off between measurement count and the number of groups For H4, LiH, and BeH2, solutions containing 8 groups were consistently found in a grouping-heavy regime when λ0 was small Furthermore, for composite rewards including two-qubit gates (RN2q(x)), the GFN-ICS approach showed a reduction of 85 two-qubit gates for the implemented circuits at λ0 = 100

Conclusion on Diversity

The central conceptual contribution is a principled mechanism for sampling multiple candidate non-overlapping groupings under a specified reward instead of committing to a single heuristic output The diversity of samples, inherent to GFlowNets and possibly other generative-based models, is shown to be useful for candidate ICS initialization selection and trade-off analysis in the tested systems The results indicate that the sampled candidates have structurally distinct measurement/two-qubit trade-offs and can improve downstream ICS initialization beyond selecting the single best non-overlapping grouping

Future Directions

Two main limitations to scaling this class of algorithms are the trajectory length τ and the cost of evaluating the reward function Future developments may include introducing a pre-training phase where the model is “warmed up” using circuits generated by heuristic approaches such as RLF Leveraging precomputed covariance dictionaries from classical, efficient methods may further enhance learning efficiency and reduce the computational overhead of reward evaluation The potential to generate transferable models through our sampling strategy is also noted as a direction for future work

Acknowledgments

This research was enabled by support from the Digital Research Alliance of Canada and NSERC Discovery Grant No. RGPIN-2024-06594 The code for our GFlowNets implementation is available at https://github.com/ChemAI-Lab/GFlowNetsMOpt

Data Availability

The code for our GFlowNets implementation is available at https://github.com/ChemAI-Lab/GFlowNetsMOpt The sampled graphs and outputs for the FC groupings obtained with the GINEw model in 21, as well as the sampled graphs, outputs, and models for the measurement plus two-qubit gate count composite-reward results in Table 4, are available through Zenodo at doi.org/10.5281/zenodo.21050287

References

[1] Peruzzo, A.; McClean, J.; Shadbolt, P.; Yung, M.-H.; Zhou, X.-Q.; Love, P. J.; Aspuru-Guzik, A.; O’Brien, J. L. Nature Communications 2014

[2] Tilly, J.; Chen, H.; Cao, S.; Picozzi, D.; Setia, K.; Li, Y.; Grant, E.; Wossnig, L.; Rungger, I.; Booth, G. H.; Tennyson, J. Physics Reports 2022

[3] Cao, Y.; Romero, J.; Olson, J. P.; Degroote, M.; Johnson, P. D.; Kieferová, M.; Kivlichan, I. D.; Menke, T.; Peropadre, B.; Sawaya, N. P. D.; Sim, S.; Veis, L.

[4] Bauer, B.; Bravyi, S.; Motta, M.; Chan, G. K.-L. Chemical Reviews 2020

[5] McArdle, S.; Endo, S.; Aspuru-Guzik, A.; Benjamin, S. C.; Yuan, X. Rev. Mod. Phys. 2020

[6] Bharti, K.; Cervera-Lierta, A.; Kyaw, T. H.; Haug, T.; Alperin-Lea, S.; Anand, A.; Degroote, M.; Heimonen, H.; Kottmann, J. S.

[7] Patel, S.; Jayakumar, P.; Yen, T.-C.; Izmaylov, A. F. Chemical Reviews 2025

[8] Huggins, W. J.; McClean, J. R.; Rubin, N. C.; Jiang, Z.; Wiebe, N.; Whaley, K. B.; Babbush, R.

[9] Oumarou, O.; Scheurer, M.; Parrish, R. M.; Hohenstein, E. G.; Gogolin, C. Quantum 2024

[10] Izmaylov, A. F.; Yen, T.-C.; Lang, R. A.; Verteletskyi, V. Journal of Chemical Theory and Computation 2020

[11] Yen, T.-C.; Ganeshram, A.; Izmaylov, A. F. npj Quantum Information 2023

[12] Yen, T.

Improvements for AI systems

Bold Header: GFlowNets for Non-Overlapping Groupings

The system can sample non-overlapping commuting groupings of molecular qubit Hamiltonians by learning a generative policy proportional to a reward function that combines measurement cost, circuit count, and compiled two-qubit gate count.

Bold Header: Improved Initialization for ICS

GFlowNets-generated groupings can be used to initialize iterative coefficient splitting (ICS), reducing post-ICS measurement estimates by up to 40% for Jordan-Wigner-mapped fully commuting Hamiltonians relative to SI initialization.

Bold Header: Multi-Objective Resource Optimization

The reward function allows the system to optimize multiple objectives simultaneously by adjusting hyperparameters, such as in Equation (9): R(x) = λ0RM(x) + λ1RG(x) + λ2RN2q(x), enabling selection of groupings based on trade-offs between measurement requirements, circuit count, and two-qubit gate counts.

Bold Header: Hardware-Aware Adaptation

The reward function provides flexibility for hardware-aware adaptations via its reward function, allowing the model to prioritize metrics relevant to specific hardware constraints like circuit fidelity overheads or restrictions on hardware-efficient measurement circuits.

Bold Header: Robustness Against Heuristic Failures

The system can outperform conventional greedy algorithms by exploring diverse solution spaces, as demonstrated by the finding that the GFlowNets generate groupings with up to 37% fewer two-qubit gate requirements, even when standard heuristics struggle, such as in the QWC setting.

Bold Header: Scalable Solution Selection via Pareto Fronts

The system can select optimal initialization candidates by analyzing the Pareto front, showing that the solution with the lowest measurement estimate is not necessarily the best ICS initialization and that retaining samples from this set can improve post-ICS measurement estimates by up to 15% for specific molecular systems.

Abstract

Estimating molecular Hamiltonians to chemical accuracy requires a large number of measurements. Hamiltonian overlapping grouping methods focus on reducing measurement counts, employing greedy initializations, such as sorted insertion (SI), leaving useful measurement/circuit trade-offs unexplored. Here, we formulate Hamiltonian grouping as a reward-driven generative search problem and introduce a Generative Flow Networks (GFlowNets)-based model that colors graph representations of molecular qubit Hamiltonians to sample non-overlapping commuting groupings. The reward function can combine measurement cost, circuit count, and compiled two-qubit-gate count, enabling multi-objective optimization without differentiable cost functions or model pretraining. Across molecular Hamiltonian benchmarks, the GFlowNets sampler finds fully commuting non-overlapping groupings with average measurement requirements 18% lower than SI, and produces Pareto sets that expose trade-offs among shots, circuits, and two-qubit resources. When used to initialize iterative coefficient splitting (ICS), GFlowNet-generated groupings reduce post-ICS measurement estimates by up to 40% for Jordan-Wigner-mapped fully commuting Hamiltonians relative to SI initialization. Composite rewards further identify lower two-qubit gate groupings, including cases with more than 100 fewer compiled two-qubit gates, while retaining comparable post-ICS measurement benefits. GFlowNets provide a flexible workflow for resource-aware measurement and quantum-resource optimization in quantum chemistry, replacing single-heuristic outputs with diverse candidate groupings that can be selected according to hardware-specific priorities. Our results show that our GFlowNets' generative policy framework not only reduces measurement and two-qubit gate costs but also provides flexibility for hardware-aware adaptations via its reward function.