PaperlayerView exact revision on arXiv

Source-linked AI summary

Tensor network surrogate models for variational quantum computation

Ryo Watanabe, Dries Sels, Joseph Tindall

arXiv:2604.20180v1quant-ph

TL;DR

Device noise has limited hardware QAOA experiments to shallow circuits, motivating classical surrogates for probing deeper schedules and larger scales. This paper uses tensor networks to simulate and train QAOA on two-dimensional lattices, finding that small-instance parameter transfer saturates while representative-size training yields lower-energy samples.

  • Problem

    Because device noise has limited QAOA experiments to shallow circuits, accurate classical surrogates are sought to probe deep schedules and scales beyond current hardware.

    Source: §I · p. 1

  • Method

    The paper uses a connectivity-aware tensor-network framework to simulate QAOA on planar architectures and extends parameter optimization to larger systems using tensor-network simulations.

    Source: §I · p. 1, §V · p. 9

  • Results

    Transfer from much smaller systems saturates with depth, while representative-size training escapes local minima and yields lower-energy samples; parameter concentration also persists on square lattices.

    Source: §I · p. 2, §V · p. 9

  • Takeaways & Limitations

    Tensor-network methods with sampling provide a practical tool for benchmarking deep variational schedules on two-dimensional systems with hundreds of qubits and for classical-surrogate parameter optimization.

    Source: §V · p. 9

    Accurate sampling on square lattices requires significantly greater computational resources than on heavy-hexagonal lattices.

    Source: §I · p. 2

Abstract

from arXiv · show

We adopt a two-dimensional tensor-network (TN) ansatz to simulate variational quantum algorithms on two-dimensional qubit architectures, demonstrating its capability to accurately simulate deep circuits through the Quantum Approximate Optimization Algorithm (QAOA) applied to Ising spin-glass problems on heavy-hexagonal and square lattices. For heavy-hexagonal problems with up to three-body interactions, parameters trained on small instances and transferred to systems an order of magnitude larger improve the sampled energy distribution only up to intermediate depths, indicating a fundamental limit of parameter concentration as a transfer strategy. By extending the training itself with TN simulations on larger system sizes, we avoid local minima and obtain lower-energy samples. Analyses of entanglement growth and importance sampling show that the simulation remains classically feasible with moderate bond dimension. We find that parameter concentration also persists on square lattices, albeit at substantially higher computational cost to perform reliable sampling. Overall, our TN framework not only provides an efficient and controlled framework for benchmarking variational quantum algorithms on two-dimensional lattices, but also serves as an effective surrogate model for training variational algorithms.

I. INTRODUCTION

The paper uses connectivity-aware tensor networks to simulate and train QAOA on two-dimensional Ising spin-glass lattices beyond state-vector reach. Larger, more representative training instances improve parameter transfer, while entanglement and sampling analyses support tractability at moderate bond dimension.

Motivation

  • QAOA is a variational algorithm for combinatorial optimization that extracts candidate solutions by sampling output bitstrings, motivating classical surrogates for probing deeper schedules than noisy hardware has reached.The introduction frames practical quantum utility in terms of resolving scientifically meaningful questions and notes that device noise has limited QAOA demonstrations to shallow circuits.Source: §I · p. 1

Approach

  • The framework simulates QAOA on two-dimensional heavy-hexagonal and square lattices using a connectivity-aware TN ansatz, BP-conditioned truncations, and boundary-MPS bitstring sampling.Parameters are trained on small instances and transferred to larger ones, with low-dimensional interpolation used to extend them to deeper schedules.Source: §I · p. 1

Results

  • For heavy-hexagonal instances, parameter-transfer quality improves with schedule depth only up to a saturation point, beyond which added layers provide no further improvement.Training on systems closer in size to target instances is most effective for transfer.Source: §I · p. 2
  • Entanglement remains bounded during circuit evolution, while stable sampled-energy histograms indicate accurate sampling at moderate boundary-MPS bond dimension.Importance weights improve systematically with boundary-MPS rank, providing a measure of sampling accuracy.Source: §I · p. 2
  • Training on larger, more representative systems helps optimization escape local minima and produces energy distributions that converge to lower values.The optimization uses TN simulations on systems already beyond state-vector simulation reach, showing TNs can support parameter optimization as well as forward simulation.Source: §I · p. 2
  • On square lattices, parameter concentration remains effective, but accurate sampling requires significantly more computational resources than for heavy-hexagonal instances.The BP approximation is less reliable on square lattices because of their higher density of short loops; appropriate bond-dimension choices keep the TN framework applicable.Source: §I · p. 2

II. BACKGROUND · A. Quantum Approximate Optimization Algorithm · B. Problem definition

This section defines QAOA as a parameterized method for minimizing a cost Hamiltonian and specifies the heavy-hexagonal and square-lattice Ising spin-glass problems used in the paper. It also establishes how optimal configurations and measured samples are represented.

A. Quantum Approximate Optimization Algorithm

  • QAOA maps the cost function to a cost Hamiltonian, applies parameterized cost and mixing gates, optimizes their parameters, then samples bitstrings from the output state.The circuit has p layers with parameters γ and β; optimization minimizes the cost-Hamiltonian expectation value. [15]Source: A. Quantum Approximate Optimization… · p. 2

B. Problem definition

  • The study considers heavy-hexagonal lattices, IBM’s native superconducting-processor architecture, and square grids, which naturally implement two-dimensional surface codes.The heavy-hexagonal problem was introduced to benchmark QAOA on IBM heavy-hex hardware. [38]Source: B. Problem definition · p. 2
  • Heavy-hexagonal spin-glass instances have linear, quadratic, and cubic terms with coefficients sampled uniformly from {±1}.The lattice is bipartite, and its degree-2 bulk vertices in W have two degree-3 neighbors.Source: B. Problem definition · p. 2
  • The illustrated heavy-hexagonal instances range from 16 to 127 qubits and include three existing IBM processor architectures.The figure identifies ibmq_guadalupe (16), ibm_geneva (27), a 2 × 2 grid system (35), and ibm_washington (127).Source: Figure 1 · p. 3
  • Square-lattice instances use random nearest-neighbor two-body interactions, unlike the heavy-hexagonal model, which includes interactions of up to three bodies.The square-grid cost coefficients d_ij are drawn from {±1}.Source: B. Problem definition · p. 3
  • The optimal-bitstring set contains every configuration attaining the minimum cost, allowing for ground-state degeneracy; the cost Hamiltonian replaces each spin z_i with Pauli Z_i.Source: B. Problem definition · p. 3

III. METHOD

The method targets classical simulation of QAOA on large systems with deep circuits, a regime previously constrained by quantum hardware and classical simulation limits.

III. METHOD

  • The simulation strategy aims to tackle QAOA instances with large system sizes n and deep circuits (large p).The paper identifies this regime as largely unexplored.Source: §III · p. 3
  • Prior work benchmarked IBM hardware on n = 127 QAOA instances only up to p = 2, using globally optimized parameters from grid search.The broader regime remained unexplored mainly because of limits in quantum hardware and classical simulation techniques [38].Source: §III · p. 3

A. Parameter optimization

The section describes QAOA transfer learning through median parameters trained across small instances, alongside a low-dimensional schedule for optimizing deep circuits.

A. Parameter optimization

  • Ref. showed parameter concentration across Ising spin-glass sizes 16 to 127, permitting transfer from smaller to larger instances.Parameter concentration has also been reported across problem classes [41] [42] and is believed to aid practical QAOA use for large-scale optimization.Source: A. Parameter optimization · p. 3
  • For each layer, parameters optimized on small instances are aggregated by their median and transferred to target instances with n qubits.The training system has size n′ < n, and the median is computed across k problem instances.Source: A. Parameter optimization · p. 3
  • For deep circuits, parameters use an interpolation-based schedule with C ≪ p basis functions, whose coefficients are optimized using Chebyshev polynomials and BOBYQA.The basis functions are orthonormal on [0, 1], and Chebyshev polynomials are rescaled to that interval.Source: A. Parameter optimization · p. 3

B. Parameter optimization based on TN simulations

TN-based optimization on larger training instances improves the generalization of QAOA parameters to larger heavy-hexagonal systems, while training on n′ = 27 provides little refinement over n′ = 16. The larger-instance training also lowers bipartite entanglement entropy, keeping the resulting circuits suitable for TN simulation.

B. Parameter optimization based on TN simulations

  • For optimization on systems with n′ ≳ 30, the study uses TN methods with bond dimension χ = 8 and contracts cost-Hamiltonian expectations using BP message tensors.The small bond dimension accommodates repeated state evolutions during BOBYQA optimization.Source: B. Parameter optimization based on TN… · p. 6
  • n′ = 27 training yields nearly the same energy distributions as n′ = 16, suggesting its optimization landscape provides insufficient gradient to refine parameters for larger systems.The n′ = 16 parameters already yield exact solutions for n = 27 instances.Source: B. Parameter optimization based on TN… · p. 6
  • Approximate TN optimization generally yields schedules different from state-vector simulations; these discrepancies are surmised to become more pronounced as circuit depth p increases.The energy distributions for schedules with p = 10, 50, and 100 are each based on 1,000 samples; median parameters are computed from ten problem instances.Source: B. Parameter optimization based on TN… · pp. 6–7
  • Parameters trained on n′ = 35 escape the local minima of n′ = 16 and 27 training and improve sampled energies on a 127-qubit instance at p = 10, 50, and 100.The improvement is already apparent at p = 10 and is accompanied by reduced bipartite entanglement entropy.Source: B. Parameter optimization based on TN… · p. 7, Figure 5 · p. 7

C. Benchmarks on square lattice instances

On square lattices, transferred parameters continue to yield lower sampled energies at greater QAOA depths, while sampling validation is less conclusive at 8×8 than at 6×6.

C. Benchmarks on square lattice instances

  • Because short loops make BP messages less reliable on square lattices, importance weights are normalized by their sample mean, and Var(˜ω) diagnoses accuracy only after mean convergence.Without convergence, a biased sampler can produce a narrow weight distribution despite returning incorrect weights.Source: C. Benchmarks on square lattice… · pp. 7–8
  • For representative 6×6 and 8×8 square-lattice instances, increasing QAOA depth consistently yields lower sampled energies, showing parameter concentration remains effective.The instances have ground-state energies −46 and −92, respectively.Source: Figure 6 · p. 8, Figure 7 · p. 8, C. Benchmarks on square lattice… · p. 9
  • At 6×6, Mean(ω) converges across both norm-MPS ranks and Var(˜ω) falls as rank increases; at 8×8, convergence is unconfirmed within affordable ranks, though the mode approaches unity and variance falls.The authors use stability of Mean(ω) across ranks as a practical convergence indicator because directly computing the TN norm is expensive.Source: Figure 6 · p. 8, Figure 7 · p. 8, C. Benchmarks on square lattice… · p. 9

V. CONCLUSION

The paper presents a connectivity-aware tensor-network framework for simulating and training QAOA on heavy-hexagonal and square-lattice Ising spin glasses. Its results show limits to parameter transfer from small systems, while larger-system training and tensor-network simulations support deep-circuit analysis and benchmarking.

V. CONCLUSION

  • The framework simulates QAOA on heavy-hexagonal and square architectures using a connectivity-aware tensor-network ansatz, BP truncations, and boundary-MPS sampling.The study also uses tensor-network simulations to train on system sizes beyond state-vector reach.Source: §V · p. 9
  • Parameter transfer from systems an order of magnitude smaller improves with schedule depth but saturates once the smaller training systems reach optimality, delineating a practical limit.Training on representative system sizes mitigates this limitation, even when coarse tensor-network optimization introduces residual inaccuracies.Source: §V · p. 9
  • Parameter concentration also holds for square-lattice instances, which require significantly greater computational resources; GPUs help meet this demand.Tensor-network parameter optimization extends the analysis to square lattices not considered in Ref. [50].Source: §V · p. 9
  • Well-optimized schedules remain numerically tractable at p = 100 layers, beyond prior work, and the analysis examines entanglement growth and sampling quality [50].The observations agree with Ref. [50] on parameter concentration for deep heavy-hexagonal schedules and the suitability of tensor networks for locally interacting QAOA circuits.Source: §V · p. 9
  • Tensor networks with two-dimensional sampling provide a practical tool for analyzing variational algorithms and benchmarking deep schedules on systems with hundreds of qubits, while enabling classical-surrogate optimization [21] [51].The framework targets deep-circuit schedules on modern hardware.Source: §V · p. 9
  • The authors argue that quantum utility requires identifying problem families where QAOA is favorable, given efficient tensor-network methods for local spin glasses and strong classical combinatorial-optimization algorithms [52, 53].They also call for exploring quantum-computation models beyond the settings studied here.Source: §V · p. 9

Appendix A: Benchmark for ibm_geneva instances

On the ibm_geneva instance, transferred QAOA parameters recover the optimal solution by p = 50, but transfer performance has saturated at that depth. Entanglement results indicate that truncation and parameter schedules shape the simulated state, with suppressed growth supporting efficient simulation along optimal parameter paths.

Appendix A: Benchmark for ibm_geneva instances

  • At p = 50, the optimal cost C(z∗) = −42 is fully recovered and Var(ω) = 3.56 × 10−8; at p = 10, the optimum occurs in most cases.The sampling weights are tightly concentrated around ω = 1, indicating nearly exact sampling.Source: Appendix A · p. 11
  • Transfer from n′ = 16 training systems saturates by p = 50, with no further improvement; this may explain lost generalization on n = 127 instances.Source: Appendix A · p. 11
  • At p = 50, Scut converges nearly to zero after slight intermediate peaks, whereas p = 10 retains a nonzero plateau, potentially because BP truncation selects one configuration despite degenerate ground states [52].Scut is defined as the sum of Eq. (10) across the bonds corresponding to the dashed edges.Source: Appendix A · pp. 11–12, Figure 9 · p. 12
  • The (γ∗, β∗) schedule increases Scut monotonically to values significantly above other schedules’ peaks, while suppressed entanglement growth along optimal parameter paths supports stable, efficient TN simulation.Source: Appendix A · p. 12
Loading 2604.20180v1…