Chinese Academy of Sciences Proves MPS Landscape Advantage

Decades of successful density matrix renormalization group calculations demonstrate the practical trainability of matrix product states (MPS), a surprising outcome given the known difficulties in training typical quantum circuits. Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) have now resolved this paradox, proving that the energy landscapes of MPS are remarkably free from poor local minima, unlike those found in brickwork circuits. The team reports this is due to the “gauge freedom of MPS” creating effective local overparametrization, concentrating local minima near the global minimum, an effect analogous to overparameterized classical neural networks. Numerical experiments further confirm that the optimization of sequential circuits converges to near-optimal solutions even for random Hamiltonians, in stark contrast to brickwork circuits, offering a valuable guide for designing better quantum algorithms in the future.

Trainability of Matrix Product States versus Quantum Circuits

Researchers, including Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) are elucidating this unexpected success with a new theoretical framework. The researchers prove that “random canonical MPS ensembles with different orthogonality centers are statistically identical,” meaning the resulting physical states remain consistent despite variations in the MPS representation. This inherent flexibility appears to smooth the optimization process, allowing algorithms to reliably converge, suggesting a clear pathway for designing future quantum algorithms with improved trainability and performance.

Barren Plateaus and Poor Local Minima in Variational Quantum Algorithms

This connection to classical machine learning is significant, as it suggests shared principles governing the trainability of complex optimization landscapes. Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) suggest this understanding provides a valuable guide for designing variational quantum circuits and algorithms with better trainability in the future, potentially unlocking more powerful and reliable quantum computation.

Gauge Freedom in MPS and Effective Overparametrization

The pursuit of reliable quantum algorithms has taken an unexpected turn, with Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) now understanding why matrix product states (MPS) consistently outperform conventional quantum circuits despite theoretical predictions of trainability issues. Their work reveals that the “gauge freedom of MPS” isn’t a hindrance, but a key advantage. This freedom, stemming from the ability to represent the same quantum state with different MPS configurations, creates what the researchers describe as effective local overparametrization, smoothing the optimization process and facilitating the discovery of optimal solutions.

Local Minimum Distribution Invariance via Orthogonality Center Moves

Practical success observed across numerous simulations of one-dimensional quantum systems predates a comprehensive theoretical explanation of why matrix product states (MPS) are so trainable. Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) are now demonstrating that the key lies in the inherent flexibility of MPS, stemming from what they term the “gauge freedom of MPS”. This means the distribution of possible states remains consistent regardless of how the MPS is configured, effectively creating a form of local overparametrization. Numerical experiments further confirm that the optimization of sequential circuits converges to near-optimal solutions even for random Hamiltonians, in stark contrast to brickwork circuits.

Sequential Circuits Converge to Near-Optimal Solutions

Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) have now provided a theoretical basis for this long-observed phenomenon, revealing a surprising connection to classical machine learning. As explained by the authors, this concentration arises because of the ability to manipulate the orthogonality center, effectively creating a more robust optimization pathway. The authors formalized this through two theorems, establishing that the energy landscapes exhibit favorable properties for gradient-based optimization, providing a valuable guide for designing future quantum algorithms. They rigorously prove that random canonical MPS ensembles with different orthogonality centers are statistically identical, implying that the induced distribution over physical states is independent of the orthogonality center. Numerical experiments further confirm that the optimization of sequential circuits converges to near-optimal solutions even for random Hamiltonians, in stark contrast to brickwork circuits.

Energy Landscapes: Absence of Poor Local Minima in MPS

Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) have now addressed this discrepancy, revealing a fundamental property of MPS that explains its robust optimization. In sharp contrast, brickwork circuits, a common alternative, become trapped by these undesirable local minima. Numerical experiments further confirm that the optimization of sequential circuits converges to near-optimal solutions even for random Hamiltonians, in stark contrast to brickwork circuits.

Density Matrix Renormalization Group and Time-Dependent Variational Principle

The persistent success of density matrix renormalization group (DMRG) calculations presents a curious challenge to current understanding of quantum algorithm trainability. Their work centers on the “gauge freedom of MPS,” a property allowing for effective local overparametrization. Hao-Kai Zhang (Institute of Physics, Chinese Academy of Sciences), Chenghong Zhu (The Hong Kong University of Science and Technology (Guangzhou)), Shuo Liu (Princeton University), Shi-Xin Zhang (Institute of Physics, Chinese Academy of Sciences), and Tao Xiang (Institute of Physics, Chinese Academy of Sciences) rigorously prove that random canonical MPS ensembles with different orthogonality centers are statistically identical. Numerical experiments further confirm that the optimization of sequential circuits converges to near-optimal solutions even for random Hamiltonians, in stark contrast to brickwork circuits.