Tensor networks are the working answer to a brutal arithmetic problem: a quantum system of N interacting particles lives in a Hilbert space whose dimension grows exponentially with N, making exact simulation hopeless beyond a few dozen sites. By decomposing one enormous tensor into a network of small, interconnected ones, tensor network methods make many of these systems tractable. Within this family, the practical choice usually comes down to dimensionality: Matrix Product States (MPS) for one-dimensional chains, and Projected Entangled Pair States (PEPS) for two-dimensional and higher lattices. This article explains why, and gives a decision framework for picking the right architecture.

Introduction to Tensor Networks

A tensor is simply a generalization of a matrix: an array indexed by any number of labels. A quantum many-body state is naturally such an object, with one index per particle or site. Written out in full, the coefficient array of even a modest chain of spins is astronomically large. Tensor network methods, as Jacob Biamonte and Ville Bergholm put it in their 2017 tutorial "Tensor Networks in a Nutshell," "can provide an efficient approximation to certain classes of quantum states, and the associated graphical language makes it easy to describe and pictorially reason about quantum circuits, channels, protocols, open systems and more."

The key insight is that not every quantum state needs its full exponential description. Physically relevant states—especially ground states of local Hamiltonians—carry structured entanglement, and tensor networks exploit that structure to compress the representation. A comprehensive text on quantum physics with tensors notes that this decomposition of large tensors into interconnected networks "enables computational efficiency, especially for simulating large quantum systems," and that specific architectures such as MPS and Tree Tensor Networks (TTN) demonstrate wide applicability across quantum physics and quantum computation. The efficiency is achieved by exploiting the inherent structure of quantum states, particularly their entanglement properties, to reduce computational complexity.

Graphical Notation and Tensor Decomposition

Tensor networks are almost always drawn, not written. In the standard diagrammatic language, each tensor is a shape—a circle, square, or triangle—and each index is a line, often called a leg, sticking out of it. A matrix has two legs; a higher-rank tensor has more. When two tensors share an index, their legs are connected, and summing over that shared index is called a contraction. The whole network, once all internal legs are contracted, evaluates to a single number or a reduced tensor with only the open legs remaining.

This graphical notation simplifies the understanding and application of complex tensor operations: complicated index gymnastics become topological statements about which shapes are connected to which. The deeper purpose, though, is computational. Manipulating large tensors comes at a high computational cost, as a review of tensor networks in machine learning in the European Mathematical Society Magazine observes—whether those tensors represent large datasets in data science or states of quantum systems in condensed matter physics and quantum computer science. Decomposing a big tensor into a network of smaller ones replaces an intractable monolithic computation with a sequence of cheap contractions, which is particularly beneficial for high-dimensional problems that would otherwise be intractable.

Matrix Product States (MPS): The 1D Solution

The Matrix Product State, also called the tensor train format, is the archetypal tensor network. Structurally it is a chain of tensors—rank three in the interior—with the physical or data indices left open. Each tensor carries one physical index connecting it to a site of the system, and two virtual indices bonding it to its neighbors. The state of the entire chain is thus encoded as a product of matrices, one per site, with the bond dimensions controlling how much entanglement the ansatz can capture.

MPS are a special case of tensor network states representing variational wave functions, and their typical uses are one-dimensional quantum states and the compression of high-order arrays. The reason MPS work so well in one dimension is entanglement geometry: in a 1D chain, a cut separates the system into two pieces whose boundary is a single point, so entanglement across the cut stays bounded. An MPS with modest bond dimension can represent such states faithfully, which is why the format dominates simulations of 1D quantum chains and has also been adopted as a compression tool far outside physics.

Projected Entangled Pair States (PEPS): Extending to Higher Dimensions

PEPS generalize the chain to a lattice. Tensors sit on a two- or higher-dimensional lattice or general graph, with one physical index per site, and virtual indices connect each tensor to its neighbors in every spatial direction. Typical uses are higher-dimensional quantum states and classical partition functions—precisely the objects MPS cannot efficiently represent, because a cut through a 2D system has a boundary that grows with its width.

This is where the area law of entanglement enters. In physically relevant higher-dimensional systems, entanglement tends to scale with the boundary of a region rather than its volume—a property that is key in critical quantum systems and that PEPS are built to capture. A tensor network perspective on quantum advantage identifies the effective dimensionality of the system—quasi-1D, higher-dimensional lattice, or irregular—as the first decision axis guiding the choice of tensor network dimensionality, alongside correlation geometry and the presence or absence of causal structure such as entanglement growth and area laws.

PEPS pay for their expressiveness in cost. Contracting a 2D network exactly is itself computationally hard, so practical PEPS algorithms rely on approximate contraction schemes. There is also a boundary caveat from application work: in a PEPS-based machine learning study of spin configurations, researchers found that while the two-dimensional structure of PEPS outperformed one-dimensional networks like MPS and TTN in learning the configurations, "close to critical point it fails to learn the configurations accurately due to increase in the long range correlation in the vicinity of the phase transition," motivating interest in MERA-style architectures for long-range ordering.

Tensor Network Architecture Decision Framework

The framework below maps the two decisive system properties—dimensionality and entanglement structure—onto an architecture recommendation, with the computational trade-off made explicit.

Choosing between MPS and PEPS by system property
System propertyRecommended architectureWhyTrade-off to watch
One-dimensional chain (or quasi-1D geometry)MPSChain structure matches the network; entanglement across a single cut stays bounded, so modest bond dimension sufficesCannot efficiently represent states whose entanglement grows faster than a 1D cut allows
Two-dimensional or higher latticePEPSTensors on a lattice capture area-law entanglement, which scales with a region's boundary rather than its volumeExact contraction is computationally hard; practical use requires approximate contraction methods
Higher-dimensional system near a critical pointPEPS with caution; consider MERA-family networksLong-range correlations grow near a phase transition, and PEPS-based learning has been reported to fail thereOpen research question; no single architecture dominates
High-order data arrays regardless of physicsMPS / tensor trainCompression of high-order arrays is a typical tensor-train useCompression quality depends on the data's correlation structure

Read the table as a two-step check. First, fix the effective dimensionality of your system: a chain or quasi-1D geometry points to MPS, a genuine 2D or 3D lattice points to PEPS. Second, examine the entanglement and correlation structure: area-law behavior fits both architectures' strengths, while growing long-range correlations near criticality strain PEPS and may call for layered networks such as MERA, which are designed for critical systems and real-space renormalization.

Beyond Quantum Physics: Diverse Applications of Tensor Networks

Tensor networks originated in condensed matter physics, but their utility extends to various scientific and computational domains because of their ability to handle high-dimensional data and complex correlations. In data science, tensors represent large datasets; in condensed matter physics and quantum computer science, they represent states of quantum systems—and the EMS Magazine review notes significant overlap in how tensors are treated across these areas. In machine learning, MPS-style decompositions have been used to compress neural network weight layers, reducing parameter counts while preserving function.

Other application areas include quantum chemistry, quantum error correction—connected to holographic tensor network models in a topical review cited within the EMS piece—and quantum computation more broadly, where tensor networks model quantum dynamics from optimization problems to open-system evolution. Biamonte and Bergholm's tutorial likewise emphasizes pictorial reasoning about quantum circuits, channels, protocols, and open systems. The common thread is that wherever high-dimensional structure with limited correlations appears, a tensor network can compress it.

Choosing the Right Tensor Network for Your Quantum System

Computational scientists and quantum physicists should select their tensor network architecture based on the dimensionality of the quantum system they are modeling, opting for MPS in 1D scenarios and PEPS for 2D and higher-dimensional systems. The decision is not merely a matter of taste: matching the network geometry to the system's entanglement structure is what makes the compression both faithful and affordable, while a mismatch—such as forcing a 2D lattice into an MPS—squanders the efficiency that motivated the method.

The measurable indicator of a good choice is the efficiency and accuracy of quantum simulations for many-body systems, particularly in terms of computational resources and the faithful representation of entanglement properties. Concretely, watch two numbers as you scale your simulation: the bond dimension needed to converge observables, and the contraction cost per sweep. If either grows sharply as system size increases, that is the signal that your architecture's dimensional assumption no longer matches your system—and, per the PEPS criticality finding, the point to consider a different network family.

Sources