Finite-Element Matrix Product States for Continuum Models in One Dimension

This paper introduces a finite-element matrix product state framework that utilizes non-orthogonal single-particle basis sets to efficiently simulate one-dimensional continuum quantum many-body systems, enabling the solution of generalized eigenvalue problems via a density matrix renormalization group algorithm for applications such as the inhomogeneous Lieb-Liniger gas.

The Gist

Imagine you are trying to solve a massive, complex puzzle representing a quantum system (like a cloud of ultra-cold atoms) that exists in a smooth, continuous world. For decades, scientists have used a powerful tool called DMRG (Density Matrix Renormalization Group) to solve these puzzles, but it was originally designed for "pixelated" worlds—systems made of distinct, separate blocks (like a grid of squares).

The problem is that the real world isn't pixelated; it's smooth. When scientists tried to force the smooth world into a pixelated grid to use their old tools, they ran into three major headaches:

1. The "Pixelation" Error: Just like a low-resolution photo looks blocky, the math didn't always guarantee the answer was the "best" possible one. Sometimes, making the grid finer actually made the answer worse before it got better.
2. The "Rigid Grid" Problem: Standard grids are rigid. If you have a tiny, sharp feature (like a narrow wall inside a trap), you need a super-fine grid everywhere to see it, which is computationally expensive.
3. The "Overlap" Issue: To make the math work better, scientists sometimes use "tent functions" (shapes that look like triangular tents) that overlap with their neighbors. While this is great for capturing smooth curves, the overlapping pieces break the rules of the old DMRG tool, which expects pieces to be perfectly separate.

The New Solution: A "Translation" Layer

The authors of this paper (Shankar, Van Acoleyen, and Haegeman) propose a clever new framework called Finite-Element Matrix Product States (FE-MPS).

Think of their solution as building a translation layer or a specialized adapter.

1. The Physical World (The Messy Reality): They start with the real, smooth world using those overlapping "tent" functions. This is great for accuracy and handling smooth curves, but the math gets messy because the tents overlap (non-orthogonal).
2. The Computational World (The Clean Grid): They create a separate, imaginary "computational space" where the rules are simple and clean (like a standard grid with no overlaps).
3. The Adapter (The MPO): The magic happens in the middle. They build a mathematical "adapter" (called a Matrix Product Operator, or MPO) that translates the messy, overlapping reality into the clean computational language. This adapter is smart enough to keep track of exactly how much the tents overlap, so no information is lost.

By doing this, they can use the powerful, fast DMRG engine (which loves clean grids) to solve the messy, smooth problem. The engine thinks it's working on a simple grid, but the adapter ensures it's actually solving the complex, continuous physics correctly.

Why is this better?

• It's a "Guaranteed" Solution: Unlike the old pixelated methods that could give you a wrong answer that looked "close," this new method is variational. Think of it like climbing a mountain: the old method might let you slide down a false peak, but this method guarantees you are always climbing toward the true highest peak (the true ground state energy). You never get a result that is "better" than the true answer; you only get closer to it.
• It Handles "Zooming" Naturally: The paper introduces a multigrid strategy. Imagine you are drawing a map. First, you sketch the rough outline on a large piece of paper. Then, you take that sketch and paste it onto a much larger, finer piece of paper to add details.
  • In this new method, the "tent" functions have a special property: you can perfectly map a coarse sketch onto a fine grid without losing any data.
  • This allows the computer to solve the "big picture" quickly first, and then use that solution as a starting point to solve the "fine details" much faster. It's like getting a head start on the puzzle rather than starting from scratch every time you zoom in.

What did they test?

They tested this on a famous model called the Lieb-Liniger gas (a line of bosons that bump into each other). They looked at two scenarios:

1. A simple box: They showed their method converges steadily to the correct answer, whereas the old pixelated method sometimes jumped around or gave slightly wrong answers.
2. A trap with a tiny barrier: They put a very narrow "wall" (a Gaussian barrier) inside a trap. This is hard to see on a standard grid unless the grid is incredibly fine. Their method handled this "competing length scale" beautifully, using the multigrid approach to first find the general shape of the gas and then zoom in to resolve the tiny wall efficiently.

The Bottom Line

The authors have built a bridge between the messy, continuous world of real physics and the clean, efficient world of current quantum computing algorithms. By using a "translation adapter" to handle overlapping shapes, they allow scientists to simulate smooth quantum systems with high accuracy, guaranteed correctness, and the ability to zoom in on details efficiently without crashing the computer.

Technical Summary

Technical Summary: Finite-element matrix product states for continuum models in one dimension

Problem Statement
The development of the density matrix renormalization group (DMRG) and matrix product states (MPS) has established a standard for studying strongly correlated one-dimensional lattice models. However, extending these methods to continuum systems (non-relativistic field theories) remains a significant challenge. Existing approaches face a trade-off between three competing properties: a strict variational principle, basis orthogonality, and spatial locality.

• Finite-difference (FD) discretization preserves orthogonality and locality but fails to represent a well-defined Hilbert space projection, often resulting in non-monotonic convergence and energies that do not strictly bound the continuum ground state.
• Continuous MPS (cMPS) restores variationality but struggles with spatial inhomogeneities, broken gauge redundancy, and boundary conditions, making robust ground-state searches difficult.
• Truncated orthogonal bases (e.g., plane waves) maintain variationality but destroy interaction locality, leading to unfavorable computational scaling.
• Finite-element (FE) methods offer a variational subspace with local interactions using piecewise-linear "tent" functions but inherently lose basis orthogonality due to the physical overlap of adjacent functions. Previous attempts to combine FE with DMRG either destroyed locality by orthogonalizing the basis or required specialized gluing conditions.

Methodology
The authors propose a framework that resolves basis non-orthogonality directly at the many-body level without transforming single-particle orbitals. The core methodology involves:

1. Non-Orthogonal Basis Expansion: The continuous field operators are expanded in a set of linearly independent, non-orthogonal single-particle orbitals $\{\phi_i(x)\}$. This leads to a non-canonical commutation relation $[\hat{a}_i, \hat{a}^\dagger_j]_\pm = N_{ij}$, where $N$ is the overlap matrix.
2. Mapping to an Auxiliary Computational Space: The physical problem is mapped to an auxiliary computational Fock space $\mathcal{H}_c$ constructed from canonical creation operators $\{c^\dagger_i\}$ satisfying standard commutation relations. A linear map $W$ connects the computational space to the physical subspace.
3. Generalized Eigenvalue Problem: The Schrödinger equation in the computational basis becomes a generalized eigenvalue problem: $\mathcal{H}|\Phi\rangle_c = E \mathcal{N}|\Phi\rangle_c$. Here, $\mathcal{H} = W^\dagger \hat{H} W$ is the effective Hamiltonian, and $\mathcal{N} = W^\dagger W$ is the many-body overlap operator acting as the metric.
4. MPO Construction of the Overlap: A key technical achievement is the exact formulation of the many-body overlap operator $\mathcal{N}$ as a Matrix Product Operator (MPO). By utilizing a "particle-flow" picture where virtual indices track the number of particles crossing links between sites, the authors show that for localized orbitals with a small bandwidth $R$, $\mathcal{N}$ admits a compact MPO representation with a bond dimension independent of the total number of basis functions $L$.
5. Generalized DMRG: The ground-state search is performed using a modified DMRG algorithm that solves the generalized eigenvalue problem locally. To ensure numerical stability, the authors employ the Locally Optimal Block Preconditioned Conjugate Gradient (LOBPCG) algorithm, avoiding the explicit inversion of the effective metric.
6. Multigrid Optimization: The framework leverages the specific properties of finite-element tent functions to enable exact refinement. A coarse grid basis can be exactly mapped to a refined grid basis via a refinement map $R$, which is also constructed as an MPO. This allows for multigrid strategies where solutions on sparse grids serve as initial states for refined grids.

Key Contributions

• Generalization of FE-DMRG: The work generalizes a previous construction (originally for chiral fermions) to encompass bosonic statistics and arbitrary local orbitals, resolving the non-orthogonality issue without sacrificing locality.
• Exact MPO Representation of Overlap: The paper provides an explicit construction showing that the many-body overlap operator for localized non-orthogonal bases can be encoded efficiently as an MPO, enabling the use of standard tensor network algorithms.
• Variational Continuum Simulation: The approach recasts the variational ground-state search into a generalized eigenvalue problem, ensuring that the computed energy is a strict upper bound to the continuum ground state.
• Multigrid Capability: The framework provides a natural setting for exactly refining the lattice, enabling multigrid optimization strategies that avoid local minima and accelerate convergence.

Results
The framework is applied to the Lieb-Liniger gas in the presence of inhomogeneous potentials using a first-order finite-element expansion.

• Benchmarks: Comparisons between FE-MPS and standard finite-difference MPS (FD-MPS) for particles in an infinite square well and a harmonic trap show that FE-MPS provides a strictly variational upper bound that converges monotonically to the continuum limit. While FD-MPS may achieve smaller absolute errors for specific grid sizes, it lacks variationality and can exhibit non-monotonic behavior.
• Convergence Rates: The FE-MPS energy converges at a rate of $O(\Delta x)$, limited by the cusp in the many-body wavefunction induced by $\delta$-interactions, which restricts the expressivity of the piecewise-linear basis.
• Observables: FE-MPS naturally captures continuum observables, such as Tan's contact, without requiring interpolation, unlike FD-MPS which yields discrete points.
• Multigrid Efficiency: In scenarios with competing length scales (e.g., a narrow Gaussian barrier in a harmonic trap), the multigrid optimization scheme demonstrates robustness. It provides converged intermediate results essentially for free and consistently locates lower ground-state energies compared to random initialization, avoiding local minima.

Significance
The paper claims to provide a robust and efficient method for simulating one-dimensional quantum many-body systems in the continuum. By directly accounting for basis non-orthogonality through an auxiliary computational space and an MPO-encoded overlap metric, the authors overcome the historical trade-off between variationality and locality. The approach recovers the benefits of lattice DMRG while maintaining a well-defined variational principle for continuum models. Furthermore, the inherent compatibility with multigrid optimization strategies offers a practical solution to the convergence difficulties often encountered in continuum simulations with competing length scales. The work establishes a foundation for studying inhomogeneous bosonic systems, such as the Lieb-Liniger model, with improved stability and convergence properties.