Synergy Between Quantum Circuits and Tensor Networks: Short-cutting the Race to Practical Quantum Advantage

I. INTRODUCTION

• PQC proposals target problems in quantum chemistry, materials science, and quantum machine learning [13] that are difficult for classical algorithms.The paper frames practical value in terms of solving problems of high interest with limited quantum resources.Source: §I · p. 1
• Practical quantum advantage remains challenging because PQCs face barren plateaus and unfavorable local-minimum guarantees, while task-specific initialization is less used in QML than quantum chemistry.The cited challenges typically concern generic ansätze and initialization schemes.Source: §I · p. 1
• Classical resources are expected to remain more abundant than quantum resources, while tensor-network simulations can efficiently describe PQCs with limited-entanglement intermediate states.Access to GPUs and TPUs further strengthens classical simulation capabilities.Source: §I · p. 1
• The framework trains an MPS on a classical simulation, transfers its state into PQC layers, then adds gates beyond classical simulation to improve on the MPS solution.The MPS is bounded by the entanglement available at its bond dimension χ; the added gates are initialized near identity.Source: §I · p. 1, Figure 1 · p. 2
• The authors report improved performance and trainability across generative modeling and Hamiltonian ground-state search, with circuits converging to high-quality solutions.For circuits up to 100 qubits, reported gradient variances and magnitudes remain stable as qubit count and depth increase.Source: §I · p. 2

II. RELATED WORK

• The framework provides a concrete realization and robust experimental verification of the benefits anticipated by Ref. across circuit architectures and learning tasks.The authors characterize their method as broadly similar to that proposal.Source: §II · p. 2
• The earlier MPS pretraining method decomposed χ = 2 states into staircase two-qubit gates, improving PQC performance and trainability but limiting classical resources.The proposed framework instead supports arbitrarily large classical and quantum resources, with continued practical returns reported.Source: §II · p. 2
• A suitable MPS-to-PQC decomposition must accept arbitrary bond dimensions, output one- and two-qubit circuits of desired depth, and converge toward the MPS as depth increases.These requirements respectively support increasing classical resources, increasing quantum resources, and avoiding fidelity plateaus; Refs. [39] also satisfy them.Source: §II · pp. 2–3

A. Parametrized Quantum Circuits

• PQC parameters θ are optimized, or trained when learning from data, according to a loss function.The prepared state is denoted ψθ.Source: A. Parametrized Quantum Circuits · p. 3
• Because the prepared state is not explicitly accessible on classical hardware, the loss L(θ) must be estimated using quantum circuit measurements.The loss depends on the wavefunction prepared by the circuit.Source: A. Parametrized Quantum Circuits · p. 3
• Training approaches include gradient descent using finite-distance gradients or the parameter-shift rule, as well as gradient-free CMA-ES.Source: A. Parametrized Quantum Circuits · p. 3

1. Quantum Circuit Born Machines

• A QCBM assigns probabilities to binary strings through measurement probabilities of its parametrized wavefunction, and can generate samples by measuring that state.QCBMs can represent complicated distributions, though their performance on near- to mid-term devices and complex real-world datasets remains uncertain.Source: §1 · p. 3
• The training loss is the KL divergence between the QCBM distribution qθ and the evenly weighted empirical distribution pD associated with dataset D.Source: §1 · p. 3
• For non-uniform dataset weighting, the final line of Eq. 3 must instead use the expectation Ex∼pD(x).Source: §1 · p. 3

2. Variational Quantum Eigensolver

• VQE finds a Hamiltonian ground state or ground-state energy by minimizing the variational energy of a parametrized trial wavefunction.The trial wavefunction is sampled on a quantum computer.Source: §2 · p. 3
• Sampling the trial wavefunction in multiple bases estimates each Hamiltonian operator’s expectation value with finite precision.Source: §2 · p. 3
• VQE can calculate useful Hamiltonian properties, including molecular energies in quantum chemistry, where the qubit Hamiltonian can be obtained using the Jordan-Wigner transformation.The passage also notes that practical quantum-chemistry problems have motivated specialized circuit ansätze and parameter initializations.Source: §2 · p. 3

B. Tensor Networks

• Tensor networks factorize large tensors into lower-order tensors that recover the original wavefunction when contracted.The factorization can reduce the complexity of storing and handling tensors with exponentially many entries.Source: B. Tensor Networks · pp. 3–4
• A tensor network’s graph specifies which tensor cores are contracted, while its topology determines the forms of entanglement it can reproduce.Source: B. Tensor Networks · p. 4
• Matrix product states (MPS) connect tensors along a line graph, with physical indices for qubits and virtual links between neighboring tensor cores.MPS tensors have order 3 in the bulk and order 2 at the boundary.Source: B. Tensor Networks · p. 4
• An MPS bond dimension χ sets an upper bound on the simulated state’s entanglement; when that is insufficient, SVD can produce a low-rank approximation with near-optimal fidelity.Source: B. Tensor Networks · p. 4
• The MPS results can be extended to more complex tensor-network architectures, which offer different tradeoffs between expressivity and classical computational complexity.Source: B. Tensor Networks · p. 4

Tensor Network Born Machines

• Tensor network Born machines (TNBMs) are generative models that represent probability distributions with simulated quantum wavefunctions parameterized by tensor networks.They provide classical solutions for quantum circuit Born machines (QCBMs).Source: Tensor Network Born Machines · p. 4
• QCBMs can leverage quantum hardware to reproduce complex-data correlations but are limited by current NISQ devices, whereas TNBMs use classical accelerators but are limited by efficiently simulable entanglement.Source: Tensor Network Born Machines · p. 4
• TNBMs allow exact calculation of probabilities qθ(x) and loss-function gradients, while conversion between these models complicates practical hybrid approaches.The work specifically considers TNBMs implemented by one-dimensional MPS rather than general tensor networks.Source: Tensor Network Born Machines · p. 4

C. MPS to PQC mapping

• Direct MPS-to-PQC conversion can yield multi-qubit gates that are difficult to compile, whereas two-qubit gates are preferred for hardware compatibility and more economical parametrization.For bond dimension χ, a direct mapping can require gates acting on ⌈log2(χ)⌉+1 qubits.Source: C. MPS to PQC mapping · p. 4
• The Ref. [31] protocol decomposes an MPS into k next-neighbor SU(4) circuit layers using classical constrained optimization, with decomposition quality limited by the MPS entanglement for fixed k.The circuit can be grown sequentially to a desired fidelity.Source: C. MPS to PQC mapping · p. 4
• The decomposition is classically tractable because its linear circuit layers tend to lower MPS bond dimension, so an initially feasible MPS should remain feasible to decompose.The passage contrasts this with brute-force optimization, whose intermediate states are not guaranteed to stay within the original maximum bond dimension.Source: C. MPS to PQC mapping · p. 5
• The mapped circuit is extended by changing its final SU(4) layer to all-to-all connectivity and initializing added gates near zero, then training all gates.The extension is intended to increase bond dimension beyond what classical simulation could handle.Source: C. MPS to PQC mapping · p. 5

• The evaluation covers two 12-bit generative datasets and a nine-qubit 2D Heisenberg problem, using CMA-ES to optimize PQCs initialized from MPS, random parameters, or near-identity gates.Cardinality has moderately low neighboring-bit correlations, while BAS has strong two-dimensional row-or-column correlations.Source: §IV · p. 5
• MPS-initialized circuits achieve better final losses than random or near-identity initializations across the studied tasks, despite identical circuit layouts and parameter counts within each comparison.The benchmarks include cardinality and bars-and-stripes QCBMs and a 3 × 3 Heisenberg VQE.Source: §IV · pp. 5–6, Figure 2 · p. 6
• MPS initialization changes convergence behavior: circuits can reach high-quality solutions even when random circuits plateau, and larger bond dimensions can improve final performance.For BAS, MPS-initialized circuits start at relatively high KL divergence but converge further; random circuits plateau at higher KL values.Source: §IV · p. 6
• In the VQE simulation, χ = 2 MPS initialization is insufficient for the PQC to outperform what classical-hardware MPS may achieve; larger χ, enabled by layer-efficient decomposition, enables significant enhancements.Source: §IV · p. 6
• Randomly initialized QCBMs show exponentially decaying gradient variance with qubit count and circuit depth, whereas MPS initialization avoids that decay and can retain or increase variance with depth.A single all-to-all layer can saturate barrenness; MPS-initialized gradients remain stable in the studied setting.Source: §IV · p. 7, Figure 3 · p. 7
• The gradient-scaling study is constrained by statevector simulation to systems and datasets of up to 20 qubits, and its all-to-all topology is unlikely to be practical at scale on near- to mid-term hardware.A separate BAS study uses an MPS-based circuit simulator to examine gradient magnitude scaling up to 100 qubits.Source: §IV · pp. 7–8
• On the BAS task, pretrained gradient magnitudes remain more stable than random-circuit gradients up to 100 qubits, with χ = 4 eventually outperforming χ = 2.In the reported scaling study, χ = 2 gradients begin to decay after 81 qubits.Source: §IV · p. 8, Figure 4 · p. 8

V. CONCLUSIONS

• The framework uses abundant classical resources to find promising solutions before transferring optimization to quantum hardware, reserving scarce quantum resources for further improvement.The authors identify very-large-χ performance as uncharacterized and anticipate that more topology-adapted tensor networks may improve results.Source: §V · pp. 8–9
• Pretrained PQCs achieved better training losses with identical quantum resources and turned deep circuits from practically untrainable into circuits that reliably converge to high-quality solutions.Gradient variances and magnitudes remained essentially constant with system size; at up to 100 qubits, larger bond dimensions favored maintaining this scaling.Source: §V · p. 9
• The study has not characterized performance at problem sizes requiring very large χ, and the line-graph MPS ansatz may be mismatched to some circuit architectures.The authors identify tree tensor networks as a possible next direction.Source: §V · p. 9

Appendix A: Distribution of gradient magnitudes

• Pretrained QCBMs break the exponential decay in gradient magnitudes seen with random parameters, although gradient distributions vary substantially across individual pretrained models.The reported individual-model distributions should be treated as representative examples.Source: Appendix A · p. 12

Appendix B: Performance enhancement when extending linear layers

• On the 10-qubit Cardinality task, extending the final linear layer to all-to-all connectivity lets classically initialized QCBMs reach better solutions, while random initialization suffers reduced trainability.Without added gates, classically initialized circuits can only aim to recover the MPS performance; the all-to-all topology also benefits from larger-χ MPS initializations.Source: Appendix B · p. 12, Figure 6 · p. 13

Appendix C: Median Optimization Performance from Sec. IV

• Across six-run optimization statistics, MPS initialization benefits all tested cases; increasing χ improves Cardinality and Heisenberg results, but BAS’s 2D correlations can mismatch the MPS bias.Even for BAS, the synergistic approach with any χ outperforms random and near-identity initializations.Source: Appendix C · p. 13

1. MPS optimization

• Generative-model MPS are trained by gradient descent using DMRG-computed KL-divergence gradients, adaptive bond dimensions, and a 5·10^-5 SVD truncation threshold; ground-state MPS use analogous DMRG energy gradients.Generative runs use learning rate 0.01 and 50 sweeps for Fig. 2 or 30 for Fig. 3; ground-state calculations truncate an exact χ=16 MPS to χ=2, 4, or 8.Source: §1 · p. 13

2. SU(4) gates

• Throughout the work, PQCs used SU(4) gates between qubit pairs with 15 parameters per gate; in MPS-initialized PQCs, decomposed tensors mapped to U(4) rotations then trainable SU(4) rotations with a fixed global phase.Additional gates extending the MPS-initialized PQC layer to an all-to-all topology were initialized from a zero-mean normal distribution with σ = 0.01, except in the VQE simulation in Fig. 2 and large-scale gradient simulations in Fig. 4, where σ = 0 was chosen.Source: §2 · pp. 13–14

3. PQC simulation

• All PQCs were simulated with Qulacs through ORQUESTRA using exact statevector simulation of the states and corresponding probabilities.Source: §3 · p. 14

4. PQC optimization

• CMA-ES optimizes PQC parameters with a population of 20 and task-specific step sizes; runs also use iteration caps and a loss-tolerance stopping condition.QCBM initial step size is 10^-2; VQE step sizes vary by initialization, and the stated caps are 10,000 or 15,000 steps.Source: §4 · p. 14

5. Gradients

• Gradients for Fig. 3 are estimated with a finite-distance method using ε=10^-8 for parameter index 8.Source: §5 · p. 14

Appendix E: MPS decomposition

• The MPS decomposition sequentially creates linear two-qubit circuit layers and optimizes existing layers to maximize fidelity, stopping at a depth limit or target fidelity.The protocol combines analytical truncation/disentangling with tensor-network optimization and is described as avoiding cumulative approximation-error buildup when using few layers.Source: Appendix E · p. 15