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

Isaac L. Huidobro-Meezs, Jun Dai, Rodrigo A. Vargas-Hernández · 2025

Read Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing by Isaac L. Huidobro-Meezs, Jun Dai, Rodrigo A. Vargas-Hernández (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-15486.txt.

Introduction

Isaac Huidobro-Meezs, Jun Dai, and Rodrigo Vargas-Hernández cut the cost of measuring quantum states of molecules. To know a molecule's energy to chemical accuracy takes a huge count of measurements. A common trick groups terms that can be measured together. The usual greedy way to form groups leaves better trade offs unseen.

The trio recast grouping as a search game with rewards. A flow based generative model colors a graph of the molecule's terms to propose groups that do not overlap. The reward can weigh shot count, circuit count, and gate count at once. On test molecules, their groups need about a fifth fewer measurements than the greedy start. Fed into a later step, they cut estimates by up to two fifths.

Catalog record

CREATORIsaac L. Huidobro-Meezs, Jun Dai, Rodrigo A. Vargas-Hernández
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE77 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2509-15486

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Discrete Flow-Based Generative Models for Measurement Optimization in Quantum Computing Isaac L. Huidobro-Meezs, Jun Dai, Rodrigo A. Vargas-Hernández Abstract Achieving chemical accuracy in quantum simulations is often constrained by the measurement bottleneck: estimating operators requires a large number of shots, which remains costly even on fault-tolerant devices and is further exacerbated on today’s noisy hardware by finite circuit fidelity and error-mitigation overhead. Addressing this challenge involves a multiobjective optimization problem that balances total shot count, the number of distinct measurement circuits, and hardware-specific compilation constraints. Existing methods typically rely on heuristic graph-coloring strategies to group commuting or qubit-wise commuting Hamiltonian terms, or on greedy allocation schemes for distributing measurements. Such approaches explore only a limited portion of the combinatorial solution space, potentially missing superior solutions. We introduce an algorithm that adapts Generative Flow Networks (GFlowNets) for coloring graph representations of Hamiltonians, enabling the generation of diverse, high-quality groupings. Our approach samples colored graphs in proportion to a user-defined reward, naturally capturing multiobjective trade-offs and discovering multiple competitive solutions. On benchmark molecular Hamiltonians, our method outperforms sorted-insertion baselines by reducing measurement costs. We further analyze the role of composite rewards, incorporating both the number of circuits and measurement costs, leading to additional improvements. This generative policy framework not only reduces measurement costs but also offers flexibility for potential hardware-aware adaptations through its reward function. keywords Quantum computing, Measurement optimization, GFlowNets, Measurement problem, Generative models 1 Introduction In the context of quantum computing for the electronic structure problem, reducing the cost of measurements remains a critical challenge for scaling up quantum algorithms. A prominent example is the variational quantum eigensolver (VQE) [ 1 , 2 , 3 , 4 , 5 , 6 ] , where estimating the expectation values of molecular Hamiltonians often requires a prohibitively large number of measurements to achieve chemical accuracy, constituting the measurement problem in VQE. Consequently, various strategies have been developed to reduce measurement requirements [ 7 ] . On the algorithmic side, clever grouping of commuting operators has been shown to reduce measurement overhead significantly. Examples include splitting the Hamiltonian into simultaneously measurable fragments using (anti)commutation structure, unitary factorization, fermionic-space transformations, and symmetry considerations [ 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 , 16 , 17 ] . These approaches typically rely on greedy sequential optimization strategies [ 17 , 11 ] to identify optimal groupings with precomputed covariance dictionaries and tackle fully commuting (FC) or qubit-wise commuting (QWC) groupings using coloring algorithms [ 13 ] . Taken together, these measurement-reduction strategies underscore the importance of combinatorial optimization in quantum computing. Generative models and reinforcement learning (RL) are increasingly being developed to tackle problems in quantum science and related fields, including those with a strong sequential and combinatorial nature [ 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 ] . These AI-driven methods have proposed solutions that were neither previously proposed nor anticipated by humans. For example, in Refs. [ 25 , 31 ] , the authors developed an autonomous algorithm for quantum optics experiments that identified a configuration capable of producing significantly greater entanglement than expected. The insight gained from this unexpected result led the researchers to formulate a new principle of entanglement generation [ 19 , 32 ] . Similarly, machine learning (ML) methods are emerging as powerful tools for solving challenging combinatorial optimization problems [ 33 , 34 , 35 ] . A recent study on the elementary shortest path problem employed an unsupervised graph neural network to identify near-optimal routes, outperforming classical heuristic methods on graphs, and was capable of generalizing to much larger unseen graphs (with more than 100 nodes) [ 36 ] . These examples illustrate how AI/ML models can explore vast solution spaces, often revealing high-quality solutions or hidden patterns that conventional greedy optimization methods might overlook. For the measurement problem, ML-based approaches have been developed in recent years. For example, Liang et al. [ 37 ] developed an RL agent that dynamically allocates measurement shots during a VQE optimization, decreasing the total measurements while still converging to the ground-state energy. Each of these developments, from analytic grouping to RL-based policies, contributes to reducing quantum measurement overhead within the current noisy intermediate-scale quantum (NISQ) era. Additionally, early evidence of efficiency gains from generative models in quantum computation is already emerging. Zou et al. [ 38 ] demonstrated that embedding a flow-based generative model into VQE can further accelerate its convergence [ 39 ] . Additionally, Nakaji et al. optimized a classical generative model to produce quantum circuits with desired properties. Despite these advances, most current methods focus on optimizing a single objective or tasks (e.g., minimizing variance and identifying one locally optimal shot distribution). Greedy optimization and RL-based algorithms often converge to a single local optimal solution, neglecting alternative candidates that may be as effective or more optimal. This highlights a limitation of both approaches, particularly in RL, where it is well known that agents can be trapped in a single mode of the solution space [ 40 , 41 ] . Such a lack of diversity can be problematic in scientific applications, where multiple distinct measurement schemes or experimental designs could yield satisfactory results. Restricting exploration to a single solution risks overlooking other viable and potentially more practical alternatives for quantum technologies. Generative models offer a complementary perspective; rather than producing a single optimal solution, they aim to learn a distribution over many possible solutions. Traditionally, this class of models relies on the existence of a dataset for the task problem and can be adapted to different tasks through conditional sampling schemes, making them powerful but inherently data-dependent [ 42 ] . The recently proposed Generative Flow Networks (GFlowNets) overcome this limitation by providing a framework for sampling complex objects in proportion to a reward function [ 43 , 41 , 20 ] , thereby bypassing the need for extensive training datasets. Rather than optimizing for a single objective, GFlowNets learn to generate a diverse set of high-reward solutions, effectively sampling from an unnormalized target distribution of interest, allowing for the inclusion of different figures of merit in the reward. Each sampled solution is constructed step-by-step, similar to an RL policy, but the learning objective ensures that the probability of generating any given solution is proportional to its reward. This fundamentally different probabilistic approach enables a GFlowNet to output a rich ensemble of distinct, high-quality solutions independently by learning a direct generative policy [ 43 ] . Sampling methods such as Markov chain Monte Carlo can be considered to address this problem, but they have the disadvantage of a lengthy stochastic search of the solution space and the challenge of mode mixing [ 44 ] . The potential of GFlowNets is especially promising for the measurement problem, where multiple disparate configurations (of required circuits, experimental parameters, commutativity conditions, etc.) might yield near-optimal outcomes. By covering many modes of the solution landscape, a GFlowNet-based approach could avoid the "dominant mode" trap of conventional RL. There is growing evidence that this diversity-driven approach pays off. In domains ranging from graph combinatorial problems [ 20

Read the complete text