The fundamental challenge of quantum many-body physics is an exponential one. If you have a system of $N$ interacting particles, each capable of occupying $d$ possible states, the total wavefunction is described by a massive tensor containing $d^N$ complex coefficients. For a modest system of 100 spin-1/2 particles ($d=2$), storing this tensor would require more parameters than there are atoms in the observable universe.
Yet, nature routinely finds stable ground states for these systems. Why? Because the vast majority of that exponentially large Hilbert space consists of highly non-local, unphysical states that real systems never visit. Physical states—especially ground states of local Hamiltonians—occupy a tiny "physical corner" of Hilbert space governed by the area law of entanglement. This law dictates that the entanglement between a subregion and the rest of the system scales with the area of the boundary between them, rather than the subregion's volume.
Matrix Product States (MPS) exploit this property to bypass the exponential wall, compressing the high-dimensional tensor of quantum coefficients into a linear chain of low-rank matrices.
The Core Mechanism: Tensor Factorization
To understand why an MPS works, imagine a one-dimensional chain of particles. Instead of representing the global wavefunction coefficient $C_{i_1 i_2 \dots i_N}$ as a single gargantuan tensor with $N$ indices, we decompose it into a sequence of matrix products:
$$C_{i_1 i_2 \dots i_N} = \sum_{\alpha_1, \alpha_2, \dots, \alpha_{N-1}} A^{(1)}_{i_1, \alpha_1} A^{(2)}_{i_2, \alpha_1 \alpha_2} \dots A^{(N)}_{i_N, \alpha_{N-1}}$$
Here, each physical site $k$ is assigned a local tensor $A^{(k)}$ with one physical index ($i_k$) and two virtual indices ($\alpha_{k-1}, \alpha_k$). The size of these virtual indices is known as the bond dimension ($\chi$ or $D$).
The bond dimension acts as a direct throttle on the amount of entanglement the state can represent. If $\chi = 1$, the matrices collapse into scalars, and the MPS represents a completely unentangled product state. As $\chi$ is allowed to grow, the ansatz can capture increasingly complex quantum correlations. As Eriksen et al. describe, this decomposition reduces the complexity of many-body wavefunctions by breaking down high-dimensional tensors into a sequence of lower-rank matrices, optimizing them through a renormalization group process to efficiently capture correlation effects.
Because the area law guarantees that entanglement remains bounded in 1D ground states, we can cap $\chi$ at a relatively small constant value without sacrificing accuracy. This shifts the computational scaling from exponential ($d^N$) to polynomial ($N \cdot d \cdot \chi^3$), making classical simulation tractable.
Bypassing the Numerical Sign Problem
The true power of the MPS representation becomes apparent when contrasted with traditional stochastic simulation methods, such as Quantum Monte Carlo (QMC). QMC maps quantum systems to classical ones by sampling path integrals using statistical weights. However, for fermionic systems, frustrated magnets, or theories with topological terms, these weights can become negative or complex. This causes catastrophic cancellations during sampling—a hurdle known as the numerical sign problem.
MPS and broader tensor network formalisms circumvent the sign problem entirely by abandoning stochastic sampling. Instead, they operate as a deterministic variational ansatz directly on the Hamiltonian formulation of the system.
As demonstrated by Bañuls et al. and Jansen et al., this allows for the direct computation of physical observables and ground states in parameter regimes where the action is complex. By using deterministic tensor contractions rather than random sampling, researchers can precisely map out phase diagrams and topological vacuum structures that are completely blind to traditional Monte Carlo methods.
Applications Across Domains
Because the MPS framework provides an efficient, sign-problem-free representation of low-entanglement quantum states, it has become a foundational tool across several distinct areas of research.
Lattice Gauge Theories and High-Energy Physics
Simulating the fundamental forces of particle physics on a lattice is notoriously difficult. The 1+1 dimensional Schwinger model (which simulates quantum electrodynamics on a line) serves as a primary proving ground for these techniques. Cirac et al. demonstrated that MPS can efficiently capture the ground states and real-time dynamics of gauge theories without requiring infinite-dimensional systems.
This approach has allowed researchers to calculate complex high-energy phenomena classically. For example, Kühn explores how targeting the wavefunction directly via MPS circumvents the sign problem in lattice gauge theories, enabling the precise calculation of additive mass renormalization (Funcke et al.; Funcke et al.) and the exploration of CP-violating phase transitions (Kühn et al.; Jansen et al.).
Quantum Chemistry and Molecular Structure
In quantum chemistry, the Density Matrix Renormalization Group (DMRG) algorithm uses the MPS architecture to solve electronic structure problems. Keller shows that encoding the electronic wavefunction as an MPS offers a vastly more efficient representation than traditional combinatorial methods, allowing for polynomial rather than factorial scaling.
This methodology has expanded beyond electrons. Baiardi et al. adapted the framework to molecular vibrations by mapping vibrational modes to a chain of tensors. Furthermore, Feldmann et al. developed a framework that treats protons and electrons on equal footing, using MPS to capture complex inter-particle correlations without relying on the standard Born-Oppenheimer approximation.
Quantum Circuits and Classical Machine Learning
The utility of MPS extends into quantum information and generative modeling. Because many near-term quantum circuits generate limited entanglement, they can be efficiently simulated on classical hardware using tensor networks. Shang et al. demonstrated this by simulating large-scale variational quantum circuits using MPS, while Ballarin et al. utilized MPS to track entanglement entropy production and benchmark the "entangling speed" of Quantum Neural Networks.
Even in purely classical domains, the algebraic properties of MPS are valuable. Causer et al. showed that discrete generative diffusion models can be parametrized using MPS, allowing for exact algebraic manipulation and sampling of probability distributions without relying on neural network approximations.
Beyond One Dimension
While the standard MPS chain is inherently optimized for 1D geometries, its underlying logic—factorizing a global tensor into local, interconnected pieces—extends to higher dimensions.
For two-dimensional systems, the natural extension is the Projected Entangled Pair State (PEPS) or alternative tensor network states like string bond states. In 2D, the area law dictates that entanglement scales with the perimeter of a region, making exact contractions computationally heavier. However, even low-bond-dimension networks can be incredibly potent when paired with other numerical methods.
For instance, Lin et al. demonstrated that a low-bond-dimension PEPS can effectively capture the complex nodal structure of a wave function, allowing fixed-node Monte Carlo simulations to bypass the sign problem in frustrated 2D Heisenberg models. Similarly, Wang et al. and Dong et al. utilized tensor network states to simulate frustrated magnetic systems and exotic supersolid phases on 2D lattices, mapping out quantum fluctuations where traditional stochastic approaches break down.
Key Foundational Work
Quantum simulation of the Schwinger model: A study of feasibility (2014) — contributed 1 concept elements to this direction.
Determining Factors for the Accuracy of DMRG in Chemistry (2014) — contributed 1 concept elements to this direction.
Complete characterization of the ground-space structure of two-body frustration-free Hamiltonians for qubits (2011) — contributed 1 concept elements to this direction.