Global optimization of MPS in quantum-inspired numerical analysis

Paula García-Molina,Luca Tagliacozzo,Juan José García-Ripoll

2024-05-17

Abstract:This work discusses the solution of partial differential equations (PDEs) using matrix product states (MPS). The study focuses on the search for the lowest eigenstates of a Hamiltonian equation, for which five algorithms are introduced: imaginary-time evolution, steepest gradient descent, an improved gradient descent, an implicitly restarted Arnoldi method, and density matrix renormalization group (DMRG) optimization. The first four methods are engineered using a framework of limited-precision linear algebra, where operations between MPS and matrix product operators (MPOs) are implemented with finite resources. All methods are benchmarked using the PDE for a quantum harmonic oscillator in up to two dimensions, over a regular grid with up to $2^{28}$ points. Our study reveals that all MPS-based techniques outperform exact diagonalization techniques based on vectors, with respect to memory usage. Imaginary-time algorithms are shown to underperform any type of gradient descent, both in terms of calibration needs and costs. Finally, Arnoldi like methods and DMRG asymptotically outperform all other methods, including exact diagonalization, as problem size increases, with an exponential advantage in memory and time usage.

Quantum Physics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to use Matrix Product States (MPS) to solve Partial Differential Equations (PDEs), especially to find the lowest eigenstate of the Hamiltonian equation. Five algorithms are introduced in the paper to achieve this goal:

1. **Imaginary - time evolution**
2. **Steepest gradient descent**
3. **Improved gradient descent**
4. **Implicitly restarted Arnoldi method**
5. **Density Matrix Renormalization Group (DMRG) optimization**

These algorithms are implemented within the finite - precision linear algebra framework, where the operations between MPS and Matrix Product Operators (MPOs) are carried out with limited resources. The paper benchmarks these methods on PDE problems of one - dimensional and two - dimensional quantum harmonic oscillators. The results show that all MPS - based techniques are superior to vector - based exact diagonalization techniques in terms of memory usage. Specifically:

- The imaginary - time evolution algorithm performs poorly in terms of calibration requirements and costs.
- Arnoldi - type methods and DMRG have an exponential advantage in both memory and time usage over other methods, including exact diagonalization methods, as the problem size increases.

In conclusion, this paper aims to explore and compare different MPS optimization algorithms for efficiently solving PDE problems, especially their advantages in handling large - scale problems.