Tensor Network States for Vibrational Spectroscopy

Nina Glaser, Alberto Baiardi, Markus ReiherWORLD 2022 · DOI · arXiv · OpenAlex

Abstract

This review elaborates on the foundation, the advantages, and the prospects of tensor network representations for quantum states in vibrational spectroscopy. The focus is on the recently introduced matrix product state decomposition of nuclear quantum states and its optimization by the density matrix renormalization group algorithm.

Introduction: Overcoming the Exponential Wall

In quantum chemistry, the primary challenge is solving the many-body Schrödinger equation exactly. While the Full Configuration Interaction (FCI) method provides an exact solution, it is hindered by an "exponential wall"—the computational cost grows exponentially with the number of vibrational modes. For molecules beyond a few atoms, calculating anharmonic vibrational energy levels becomes computationally prohibitive. Traditional approximations like truncated CI or perturbation theory often fail in systems with "strong correlation," where vibrational modes are intricately coupled. Tensor network states offer a solution by "compressing" the many-body wave function, capturing essential physics while discarding redundant information.

Tensor Network Foundations

Tensor networks provide a graphical language for representing complex multi-dimensional arrays. In this notation, a tensor is represented by a node, and its rank is determined by the number of "legs" or indices attached to it.

Basic tensor representations: (a) Scalar (rank 0), (b) Vector (rank 1), (c) Matrix (rank 2), and (d) Tensor of rank 3.

By connecting these legs, one can represent contractions and products, effectively building complex quantum states from simpler building blocks.

Matrix Product States and vDMRG

The core of recent advancements in vibrational spectroscopy is the Matrix Product State (MPS). An MPS represents a high-rank tensor—such as the CI coefficient tensor C—as a chain of lower-rank tensors. This decomposition transforms the exponentially large global tensor into a product of local matrices, leading to linear scaling ($O(L)$) with respect to the number of modes.

Comparison of (a) the full CI tensor C with (b) its Matrix Product State (MPS) representation. The MPS reduces the complex global structure into a linear chain of tensors.

The vibrational Density Matrix Renormalization Group (vDMRG) algorithm is used to optimize these MPS representations. By "sweeping" across the chain and updating tensors locally, vDMRG converts a massive global optimization problem into a series of manageable local ones. The accuracy is controlled by the "bond dimension" ($m$), which determines the size of the matrices in the chain.

Advanced Quantization and Excited States

To handle highly anharmonic systems, researchers utilize different quantization schemes. While Canonical Quantization maps each vibrational mode to a single site, $n$-mode Quantization maps individual basis functions to sites, offering superior performance for complex potentials.

Spectroscopy requires the calculation of excited states. Several specialized algorithms have been developed:

  • vDMRG[ortho]: Iteratively finds excited states by enforcing orthogonality to previously determined states.
  • vDMRG[FEAST]: Integrates the FEAST algorithm to find all states within a specific energy window in parallel using contour integration.
  • Time-Dependent DMRG (TD-DMRG): Directly simulates absorption spectra by solving the time-dependent Schrödinger equation on the MPS manifold, allowing for the direct simulation of absorption spectra.

Key Insights and Results

Benchmarks on molecules like ethylene ($C_2H_4$) and pyrazine ($C_4H_4N_2$) demonstrate the power of this approach. For ethylene, zero-point vibrational energies (ZPVE) converge to sub-cm⁻¹ accuracy with bond dimensions as low as $m=20$. In the case of pyrazine, TD-DMRG successfully simulated the complex 24-mode vibronic absorption spectrum with $m=100$, a feat typically impossible for exact methods.

A critical insight is the use of entanglement metrics (e.g., von Neumann entropy) to order sites within the tensor network. Optimal ordering significantly accelerates convergence by grouping strongly coupled modes together.

Conclusion and Future Outlook

Tensor network states, particularly MPS, provide a robust framework for simulating large, strongly correlated molecular systems with linear scaling. While 1D chains are highly efficient, future research points toward higher-dimensional networks like Projected Entangled Pair States (PEPS) to better match 3D molecular structures. Additionally, "Pre-Born-Oppenheimer" DMRG aims to treat electrons and nuclei simultaneously, potentially capturing non-adiabatic effects with unprecedented accuracy.

Summary updated

Abstract

This review elaborates on the foundation, the advantages, and the prospects of tensor network representations for quantum states in vibrational spectroscopy. The focus is on the recently introduced matrix product state decomposition of nuclear quantum states and its optimization by the density matrix renormalization group algorithm.
BibTeX
@inproceedings{glaser2022tensor,
  author = {Nina Glaser and Alberto Baiardi and Markus Reiher},
  title = {{Tensor Network States for Vibrational Spectroscopy}},
  booktitle = {WORLD},
  year = {2022},
  doi = {10.1142/9789811237911\_0003},
}

Loading related content...

Figure