Matrix Product State (MPS) Representation

Updated 28 January 2026

Matrix Product State (MPS) is a tensor network ansatz that represents high-dimensional quantum states by decomposing them into local tensors with finite bond dimensions.
MPS enables scalable simulation in quantum many-body physics, quantum chemistry, and machine learning by reducing the exponential storage complexity of full state vectors.
Leveraging SVD-based canonical forms, MPS efficiently controls truncation errors and captures system entanglement, underlining its role in methods like DMRG and TEBD.

A matrix product state (MPS) is a tensor network ansatz for representing vectors (typically wavefunctions) of high-dimensional Hilbert spaces, most naturally adapted to one-dimensional systems. It replaces the exponential complexity of storing an arbitrary state vector with a product of local tensors coupled along a chain via finite-dimensional “bond” indices. This representation underlies much of modern computational many-body physics, quantum chemistry, tensor network theory, and efficient (compressed) data representation in quantum information and machine learning.

1. Mathematical Formalism and Algebraic Structure

Let $|\psi\rangle$ be a pure quantum state on $n$ sites, each with physical basis $\{|s_i\rangle\}$ ( $d$ -dimensional), for $i=1,\dots,n$ . The most general state,

$|\psi\rangle = \sum_{s_1,\ldots,s_n} T^{s_1 s_2\ldots s_n} |s_1\rangle\otimes\cdots\otimes|s_n\rangle,$

requires $d^n$ amplitudes $T^{s_1\ldots s_n}$ . The MPS ansatz decomposes $T$ as a product of local three-index tensors: $T^{s_1\ldots s_n} = \sum_{\alpha_1,\ldots,\alpha_{n-1}} A^{[1]s_1}_{\alpha_1} A^{[2]s_2}_{\alpha_1,\alpha_2} \cdots A^{[n]s_n}_{\alpha_{n-1}}.$ Here $A^{[k]s_k}$ is a $D_{k-1}\times D_k$ matrix for each physical index $s_k$ , with $D_0=D_n=1$ for open boundaries. The $D_k$ are bond dimensions.

Canonical forms: Through successive SVDs, MPS tensors can be brought into left, right, or mixed canonical forms necessary for measuring entanglement and for efficient variational algorithms (Dolfi et al., 2014). Schmidt values at each bipartition correspond to singular values of reshaped tensors and control the local entanglement rank; their logarithms yield the local von Neumann entropy (Green et al., 23 Feb 2025).

Translation-invariant and algebraic geometry: For infinite systems or periodic boundary conditions, a uniform MPS is specified by fixed $A^{s}$ matrices for all sites. The algebraic-geometric study connects the set of all MPS of given bond dimension to varieties defined by polynomial relations in the amplitudes, with explicit relations known for small system sizes and specific choices of $D$ and $d$ (Critch et al., 2012, Seynnaeve, 2022).

2. Construction, Parameterization, and Variants

Exact MPS Decomposition: Any vector in $(\mathbb{C}^d)^{\otimes n}$ can be brought into MPS form with maximal bond dimension $D\le d^{\lfloor n/2\rfloor}$ , using recursive SVDs (“canonical form”). For practical purposes, the bond dimension is truncated at some finite value, incurring a controlled error quantified by the sum of discarded Schmidt coefficients (Green et al., 23 Feb 2025, Nakhl et al., 14 Jan 2026).

Truncation and error control: For a truncation to $D$ bonds, the Frobenius-norm error is the sum of the squares of discarded singular values $\epsilon_i$ across all bipartitions; the total MPS error is $\sum_{i=1}^{n-1} \epsilon_i$ (Jeon et al., 2024). This is the theoretical foundation for DMRG and all SVD-based tensor-network compression schemes.

Fermionic and bosonic MPS: Fermionic MPS can be constructed via explicit bond-space representations implementing Jordan-Wigner strings (Li et al., 2020, Silvi et al., 2012). For bosonic systems, variants handle infinite-dimensional local spaces by “folding” approaches or special soft-cutoff tensor forms (Frenzel et al., 2012, Reslen, 2013).

Stochastic MPS: In classical settings, an sMPS provides a decomposition of probability distributions as products of nonnegative matrices with a canonical “Schmidt-like” structure, encoding correlations quantified by the entropy cost $S_C$ (Temme et al., 2010).

Infinite MPS (iMPS): The algebraic conditions for a finite MPS to define a consistent infinite-volume state involve contractive (completely positive, trace-preserving) transfer operators (Souissi et al., 2024). This enables unique shift-invariant infinite-volume states with well-defined correlation decay, classified by the spectral data of the transfer map.

3. Applications in Quantum Many-Body Physics

Low-dimensional quantum systems and DMRG: MPS form the theoretical backbone of the density-matrix renormalization group (DMRG), capturing area-law entangled states efficiently with small $D$ (Dolfi et al., 2014, Liu et al., 2014). This includes generic quantum spin systems, fermionic chains, and the simulation of real-time or thermodynamic evolution via TEBD and related time-evolution algorithms (Li et al., 2020).

Fractional quantum Hall (FQH) model states: Many FQH trial states (Laughlin, Moore–Read, Read–Rezayi, Halperin, Gaffnian, Haldane–Rezayi) can be written as conformal field theory (CFT) correlators, which map naturally onto MPS representations on the cylinder or torus geometries. The auxiliary (bond) spaces correspond to chiral CFT Hilbert spaces truncated at fixed conformal level, with finite bond-dimension controlling topological entanglement entropy and correlation length (Crepel et al., 2018, Wu et al., 2015, Crépel et al., 2019, Kjäll et al., 2017). Exact MPS forms permit efficient calculation of observables, overlap integrals, and Berry phases associated with braiding.

Electronic structure and Slater determinants: Any Slater determinant of $N$ fermions in $L$ orbitals has an explicit exact MPS representation of bond dimension $2^N$ , which is optimal. Extensions to configuration interaction (CI) wavefunctions involve MPO stacking (superposition of determinants) (Silvi et al., 2012).

4. Quantum Information Processing and Machine Learning

State encoding and data compression: Amplitude encoding of classical data (images, functions, neural network features) into quantum states via MPS allows for drastic circuit-depth and resource reductions, with rigorous control over encoding infidelity as a function of bond dimension and data structure (Green et al., 23 Feb 2025, Jeon et al., 2024). Optimal qubit orderings minimize truncation errors for fixed $D$ ; search algorithms (e.g., uniform-cost permutation search with symmetry pruning) find such mappings for data compression and fidelity maximization in quantum machine learning (Jeon et al., 2024).

Quantum machine learning and robustness: MPS-based state preparation enables low-depth circuits that retain high classification performance for variational quantum classifiers, even with approximate/truncated encodings. Empirically, decreased circuit depth via aggressive MPS truncation can, in some cases, enhance adversarial robustness relative to exact encoding (Nakhl et al., 14 Jan 2026).

Boolean function representation: MPS (tensor-train) techniques underlie the Binary Matrix Product (BMP) representation of Boolean functions, which is closely related to reduced ordered binary decision diagrams (BDDs). All Boolean functions can be written exactly as an MPS, with bond dimension controlled by variable ordering; variable-order optimization is crucial for tractable circuit synthesis and classical–quantum algorithm design (Usturali et al., 3 May 2025).

5. Topological, Geometric, and Symmetry Properties

Topological order and entanglement: The topological entanglement entropy $\gamma$ can be extracted from MPS by real-space or orbital bipartition on infinite cylinders, matching predictions of Chern-Simons and CFT-based plasma analogies. For FQH states, gaplessness and ground state degeneracy are probed via MPS transfer matrices; non-Abelian statistics and monodromy appear as explicit braiding matrix entries in the auxiliary space (Crepel et al., 2018, Crépel et al., 2019, Wu et al., 2015).

Symmetry and invariant theory: The parametrization and classification of MPS spaces draws on invariant theory, specifically trace algebras and polynomial trace relations, which encode the hierarchy of entanglement ranks and spectral properties (Critch et al., 2012, Seynnaeve, 2022). The SU(2) spin-singlet symmetry can be exactly preserved in MPS form by tensor network truncations respecting zero-mode Ward identities (Crepel et al., 2018).

Algebraic identifiability: For small systems ( $D=d=2$ ), all translation-invariant MPS correspond to solutions of explicit quartic or sextic polynomial equations in the amplitudes, and the algebraic geometry of parameter identifiability (number of equivalent parameterizations for the same state) is nontrivial and grows with system size (Critch et al., 2012).

6. Numerical Algorithms and Computational Aspects

Variational optimization: Ground- and excited-state search in the MPS manifold is done via local tensor updates under fixed effective environments (single-site or two-site DMRG, with optional density-matrix correction) (Dolfi et al., 2014, Liu et al., 2014). SVD-based truncation controls the bond-dimension and trade-off between computational complexity and representation fidelity. Time evolution is efficiently performed via TEBD or equivalent MPO-based methods.

Error control and scaling: The representation error from bond-dimension truncation is strictly governed by the tail of the Schmidt (singular value) spectra. In both quantum and stochastic MPS, error bounds on the $L_2$ (quantum) and $L_1$ (classical) norms are available directly from the canonical form (Green et al., 23 Feb 2025, Temme et al., 2010). The scaling of computational cost is $\mathcal{O}(N D^3 d)$ for most core routines, with low-order polynomial dependence on system size and bond dimension.

Symmetry exploitation: Conservation of Abelian charges (particle number, spin, $\mathbb{Z}_2$ parity) is encoded in block-diagonal site tensors, dramatically reducing memory and computational cost and allowing efficient handling of large-scale systems (Dolfi et al., 2014, Liu et al., 2014).

7. Extensions and Open Directions

Higher-dimensional and non-MPS tensor networks: MPS generalizes to projected entangled pair states (PEPS) in 2D, tree tensor networks (TTN), and multi-scale entanglement renormalization ansatz (MERA), which capture spatially extended correlations at higher entropic cost (Jeon et al., 2024). Variable ordering and entanglement-minimization ideas can be transferred to these settings.

Nontrivial infinite-volume states: The extension of finite MPS to infinite systems is governed by algebraic compatibility conditions on local tensors, with full C*-algebraic foundations. This mathematical perspective unifies tensor network states with quantum Markov chains and operator algebras, providing rigorous criteria for injectivity, ergodicity, and finite-correlation properties (Souissi et al., 2024).

Open algebraic questions: The precise dimension of the linear space (variety) of MPS of fixed bond and local dimension, and the identifiability of their algebraic parameters, remain challenging in invariant theory for $D>2$ (Seynnaeve, 2022).