Description
Preparing complex quantum states on noisy intermediate-scale quantum (NISQ) devices remains a fundamental challenge. Variational quantum algorithms (VQAs) provide a flexible framework for this task. However, their optimization often suffers from the barren plateau phenomenon, where gradients vanish exponentially as the system size increases. This has motivated two common classes of variational ansätze: problem-specific ansätze and hardware-efficient ansätze (HEAs). Problem-specific ansätze improve trainability by exploiting prior knowledge of the target state, while HEAs are designed to satisfy hardware constraints and remain broadly applicable. As a result, current approaches face a trade-off between trainability and expressivity, which remains a central challenge for NISQ.
Existing VQAs generally treat quantum states as generic vectors in Hilbert space without exploiting their underlying structure. In this work, we construct a tensor-network-inspired variational quantum circuit by translating matrix product state (MPS) representations into an ancilla-assisted parameterized quantum circuit architecture. We compare this architecture with conventional HEAs for quantum state preparation. We expect the resulting architecture to exploit the compact, low-rank structure of MPS. For target states admitting efficient MPS representations, this yields a compact variational parameterization with $\mathcal{O}(4nr^2)$ parameters within a general variational optimization framework. The proposed construction is also inherently probabilistic, introducing a design feature whose influence on optimization remains largely unexplored.
Preliminary numerical results indicate that the proposed MPS-inspired architecture exhibits improved trainability over conventional HEAs for several state-preparation tasks under comparable resource constraints. We observe more favorable optimization behavior and reduced susceptibility to barren plateau effects while achieving comparable state-preparation performance. These findings suggest that tensor-network-inspired circuit architectures provide a more stable optimization landscape for suitable problem classes. More generally, our comparison suggests that such architectures offer a promising middle ground between highly specialized problem-driven ansätze and generic HEAs, providing a new perspective for designing trainable, resource-efficient quantum algorithms for near-term quantum hardware.
| I am the presenting author | Yes |
|---|