Tensor Networks: Computational Guide

Tensor networks can represent many-body states using far fewer parameters than a full state vector. Whether they make a calculation practical depends on three separate tasks: representing the state accurately, contracting the network to evaluate observables, and finding useful tensors by optimization. Each can limit the result.

Required background. Tensor-network concepts explains indices, graph cuts and network families; computational method choice introduces numerical and physical convergence.

Representing, evaluating, and optimizing a state

Consider a fixed finite Hamiltonian and a family of tensor-network states. The following questions identify different sources of error.

State approximation

Does the selected family contain a sufficiently accurate state at accessible bond dimension?

For a family $\mathcal M_D$, the best variational energy is

$$E_D^* := \inf_{\lvert\phi\rangle\in\mathcal M_D} \frac{ \langle\phi|H|\phi\rangle }{ \langle\phi|\phi\rangle }.$$

The existence of $E_D^*$ says nothing yet about whether it can be evaluated or found.

Contraction

Can norms, energies, gradients, reduced states, and target observables be contracted with controlled error?

Chains permit exact sequential contraction at polynomial cost in the MPS bond dimension. Looped networks such as PEPS generally require an approximate environment or a cost that grows rapidly with width. The environment cutoff is therefore a convergence variable distinct from the state bond dimension.

Optimization

Can the optimization reach a good point in the chosen family?

Let $E_D^{\mathrm{found}}$ be the exactly evaluated Rayleigh quotient of the returned state. Then

$$\delta E_{\mathrm{opt}} := E_D^{\mathrm{found}} - E_D^* \ge 0.$$

An approximate environment can give a reported energy below even the exact ground energy, so the inequality does not apply to an uncontrolled contracted estimate. Small iteration-to-iteration changes alone establish neither stationarity nor a small $\delta E_{\mathrm{opt}}$; check gradients or local residuals and compare initializations. Restarts, symmetry sectors, unit cells, noise schedules, update type, and initialization can expose optimization dependence.

Physical limits

Does the finite calculation support the claimed physical limit and observable?

Even an exact optimum inside $\mathcal M_D$ can be a poor approximation to the intended thermodynamic, continuum, long-time, or critical result. Bond dimension must be varied alongside size, width, time step, boundary condition, and every other active cutoff.

Family selection

Network names describe state families and contraction geometries. Optimization algorithms are separate choices.

Family or workflow	Natural first target	Main strength	Decisive controls
finite MPS with DMRG	low-energy states of finite 1D chains	exact contractions and strong Schmidt compression	$L$, $\chi$, sweeps, restarts, sector, boundary
uniform or infinite MPS	translation-invariant 1D bulk states	direct thermodynamic representation	$\chi$, unit cell, transfer fixed point, initialization
MPS time evolution	short- and intermediate-time 1D dynamics	local gates or projected evolution	time step, truncation, $\chi$, conservation, reachable time
MPS on cylinders	quasi-1D views of 2D systems	mature 1D machinery and controlled long direction	width, length, ordering, $\chi$, edge and sector effects
PEPS	2D ground-state structure	lattice-matched area-law capacity	state $D$, environment cutoff, update, contraction scheme
tree tensor network	hierarchical or cluster geometry	loop-free contraction and multiscale partitions	tree choice, bond profile, translation artifacts
MERA	scale-resolved and critical states	bounded causal cones and explicit scale layers	bond dimension, layers, architecture, scaling-operator convergence
MPO or purification	operators and thermal states	structured representation of mixed-state tasks	operator bond, ancilla choice, positivity, temperature step

No row is a guarantee. A method is a candidate only after the target, geometry, and observable are declared.

From area law to algorithm

For a single-bond cut of an open-boundary MPS with bond dimension $\chi$,

$$\operatorname{rank}\rho_A \le \chi, \qquad S_A \le \ln\chi.$$

This is a capacity statement. It does not say that:

• every area-law state has a small useful $\chi$ at the desired precision;
• the Schmidt tail decays rapidly enough for the observable of interest;
• an optimizer will find the best finite-$\chi$ state;
• a higher-dimensional network is easy to contract;
• a finite-$\chi$ result has reached the physical correlation length.

On a cylinder of width $W$, an area law of the form

$$S_{\mathrm{cut}} \sim \alpha W$$

implies the capacity requirement

$$\chi \gtrsim e^{\alpha W}$$

for a chain-ordered MPS that reproduces this entropy across the cut. For approximation at a specified fidelity, use the retained Schmidt probability rather than treating the exact-state entropy as an accuracy bound. The MPS cost can therefore grow exponentially with cylinder width even when the two-dimensional state obeys an area law.

PEPS changes the virtual geometry: a two-dimensional boundary cuts many bonds, so fixed local $D$ can supply boundary-law capacity. That representational advantage comes with a harder contraction problem. MERA distributes bonds by scale, which favors critical structure but introduces a more constrained architecture and optimization.

The correct implication is:

$$\begin{gathered} \text{favorable entanglement structure} \\ \text{may enable a compact ansatz}, \\ \text{compactness} \not\Longrightarrow \text{easy evaluation}, \\ \text{compactness} \not\Longrightarrow \text{easy optimization}. \end{gathered}$$

Bond dimension is not an error bar

The symbols $\chi$ and $D$ are resource controls, not observables.

Their meaning depends on:

• network topology;
• physical site grouping and ordering;
• symmetry-resolved block structure;
• open, periodic, finite, or infinite geometry;
• state versus operator representation;
• local physical dimension;
• the contraction and optimization method.

Two calculations at “bond dimension 256” need not have comparable state capacity or cost.

Use observable convergence

For each target observable, compare successive controls:

$$\Delta_O(D_2,D_1) = \left| O(D_2) - O(D_1) \right|.$$

A small $\Delta_O$ is useful only if:

• both calculations are independently optimized;
• environment errors are smaller than the change;
• size and boundary effects are separately controlled;
• the sequence is not trapped on one metastable branch;
• the observable is not artificially pinned by a symmetry or unit-cell choice.

Use state diagnostics

For a normalized candidate eigenstate,

$$\sigma_H^2 = \langle H^2\rangle - \langle H\rangle^2.$$

An exact eigenstate has zero variance. A decreasing exactly evaluated variance is a stronger eigenstate check than an unchanged energy, though fidelity conclusions still require spectral information.

Schmidt values, truncation weights, canonical residuals, transfer spectra, and symmetry charges diagnose other aspects of the state. These diagnostics answer different questions and should be interpreted together.

Convergence designs by regime

Finite gapped chain

A defensible finite-chain study varies:

$$\begin{gathered} L, \qquad \chi, \\ \text{sweeps or iterations}, \\ \text{initialization}. \end{gathered}$$

It checks energy variance, central versus edge observables, symmetry sector, representative Schmidt tails, and agreement with exact diagonalization at smaller $L$.

Open boundaries often reduce MPS cost, but they introduce edge profiles. Bulk claims should use a central window or a controlled boundary extrapolation.

Critical chain

For the usual optimized clustering MPS approximations to a critical state, finite size and finite bond dimension produce competing infrared scales:

$$\xi_{\mathrm{eff}} \sim \min \left( L, \xi_\chi \right).$$

This estimate is not a theorem for every finite-$\chi$ state. A noninjective MPS such as a GHZ state can have long-range order at $\chi=2$. Inspect the transfer eigenvalues and sector structure before interpreting $\xi_\chi$ as a finite correlation length. In the clustering case, increasing $L$ far beyond $\xi_\chi$ reveals the finite-entanglement plateau. A critical study therefore needs a two-dimensional grid in $(L,\chi)$ or a justified finite-entanglement scaling protocol.

Cylinder or ladder

For a cylinder, vary width and length separately:

$$W \longrightarrow \infty, \qquad L \longrightarrow \infty.$$

Also vary the one-dimensional path through the sites, boundary pinning, flux or topological sector, and bond dimension. A stable result on one width is a quasi-one-dimensional result, not automatically a two-dimensional thermodynamic conclusion.

PEPS calculation

At minimum, vary both:

$$D \qquad \text{and} \qquad \chi_{\mathrm{env}},$$

where $\chi_{\mathrm{env}}$ denotes the boundary or environment control used by the contraction scheme.

Convergence in $D$ at one fixed $\chi_{\mathrm{env}}$ can merely converge the bias of an under-resolved environment. Compare contraction schemes or environment constructions when the claimed accuracy is close to their differences.

Real-time evolution

Real-time tensor-network evolution has at least three horizons:

$$\delta t, \qquad \chi, \qquad t_{\max}.$$

Decrease the time step, increase the retained bond dimension or decrease the truncation threshold, and monitor conservation laws. Entanglement growth can make the reachable time increase only slowly with computational effort.

Beyond the converged time interval, a smooth curve can still have a large truncation error.

Thermal and mixed-state calculations

Specify whether the density operator is represented as an MPO, a purification, a minimally entangled ensemble, or another construction. Control:

• inverse-temperature step;
• operator or purification bond dimension;
• ancilla gauge and normalization;
• positivity of the density operator; complete positivity is a condition on a represented evolution map, not on the state itself;
• spatial size and thermal correlation length;
• observable-specific contraction error.

Sources of numerical error

Vary the controls associated with each approximation. They can interact: an under-resolved environment may hide the change obtained by increasing the state bond dimension.

Error source	Typical control	Typical false conclusion
onsite basis	enlarge local cutoff or orbital set	apparent convergence in a truncated model
finite size	increase length and compare boundaries	crossover or edge effect called a phase
finite width	compare cylinders or ladders	quasi-1D behavior called 2D bulk physics
finite bond	increase $\chi$ or $D$	artificial entropy or correlation-length saturation
environment	increase $\chi_{\mathrm{env}}$ and compare contractions	biased PEPS energy called variationally converged
optimization	restarts, sectors, unit cells, residuals	metastable tensor called the ground state
time or temperature step	reduce step and change integrator	discretization drift called dynamics
observable extraction	independent contractions and sum rules	normalization or insertion error called physics

Choosing stopping criteria

Stopping because the code no longer changes visibly is not a scientific criterion.

A defensible stopping rule ties numerical changes to the target:

$$\left| O_{k+1}-O_k \right| < \epsilon_O$$

over a declared sequence of controls, with independent checks that other errors are below $\epsilon_O$ or included in the uncertainty.

For a phase classification, stopping criteria may require more than one local observable:

• gap or correlation-length behavior;
• symmetry quantum numbers and order parameters;
• entanglement spectrum or topological-sector diagnostics;
• stability across boundary, initialization, and unit-cell choices;
• finite-size or finite-entanglement scaling.

Use the observables relevant to the physical question.

Exercises

Exercise 2: Cylinder-width cost

Suppose the entropy across a cylinder cut behaves as

$$S = \alpha W, \qquad \alpha = \frac{1}{2}.$$

Use $S\le\ln\chi$ to estimate the minimum capacity at widths $W=8$ and $W=12$, and find their ratio.

Solution

The capacity condition gives

$$\chi \ge e^{\alpha W}.$$

At $W=8$,

$$\chi_8 \gtrsim e^4 \approx 55.$$

At $W=12$,

$$\chi_{12} \gtrsim e^6 \approx 403.43,$$

Rounding upward gives entropy-capacity bounds of $55$ and $404$. The ratio of the unrounded bounds is

$$\frac{e^{\alpha\,12}}{e^{\alpha\,8}} = e^2 \approx 7.4.$$

This is a ratio of lower bounds, not a lower bound on the ratio of actual required bond dimensions: two separate lower bounds do not constrain that ratio. These entropy-capacity estimates are not accuracy predictions. Schmidt-tail structure and the desired observables can require larger bonds.

Exercise 3: Diagnose false PEPS convergence

A study reports energies at $D=2,3,4,5$ using one fixed environment cutoff. The final two energies agree to eight digits, so the authors claim convergence in $D$. Which convergence test is missing?

Solution

The contraction error is not controlled. A fixed, insufficient environment can make every state energy approach the same biased value.

At each relevant $D$, the study should increase the environment cutoff until the energy and target observables stabilize. It should also compare environment constructions when their systematic differences matter, verify normalization, and check whether the energy evaluation preserves the claimed variational property.

Only after contraction error is below the observed $D$ dependence does the sequence test finite-bond convergence.

Exercise 4: Design a critical-chain grid

A finite-MPS calculation targets a critical exponent. Explain why a sequence with increasing $L$ at fixed $\chi$ is insufficient and propose a minimal convergence design.

Solution

For an optimized clustering MPS at fixed $\chi$, the transfer operator generally gives a finite-entanglement correlation length $\xi_\chi$. Once $L\gg\xi_\chi$, increasing $L$ probes the finite-$\chi$ state rather than the scale-free target.

A minimal design uses several $L$ values at each of several increasing $\chi$ values. It records $\xi_\chi$, energy variance, fitted observables, and the scaling window. The analysis should show either:

• a finite-size regime with $L\ll\xi_\chi$ that is stable as $\chi$ increases; or
• a justified finite-entanglement scaling collapse with stable exponents.

Unit cell, boundaries, symmetry sector, and initialization should also be tested when they can affect the critical state.

Exercise 5: Evaluate an efficiency claim

A paper states: “Our two-dimensional state has an exact PEPS with $D=4$, so all local observables can be computed efficiently and exactly.” Identify the valid part and the unsupported conclusion.

Solution

The valid statement is representational: an exact $D=4$ PEPS is a compact description of the state in the declared geometry.

The unsupported step is from compact representation to efficient exact contraction. A PEPS norm or local expectation value forms a looped two-dimensional contraction, and exact contraction can remain computationally hard at fixed $D$. The paper must specify a contraction algorithm, its environment controls, and an error analysis.

Special isometric, free, stabilizer, or otherwise structured PEPS may be exactly tractable, but that requires an additional theorem beyond finite bond dimension.

References

• M. C. Bañuls, “Tensor Network Algorithms: A Route Map,” Annual Review of Condensed Matter Physics 14, 173–191 (2023). doi:10.1146/annurev-conmatphys-040721-022705
• F. Verstraete and J. I. Cirac, “Matrix Product States Represent Ground States Faithfully,” Physical Review B 73, 094423 (2006). doi:10.1103/PhysRevB.73.094423
• J. Haegeman, C. Lubich, I. Oseledets, B. Vandereycken, and F. Verstraete, “Unifying Time Evolution and Optimization with Matrix Product States,” Physical Review B 94, 165116 (2016). doi:10.1103/PhysRevB.94.165116
• M. B. Hastings, “An Area Law for One-Dimensional Quantum Systems,” Journal of Statistical Mechanics: Theory and Experiment 2007, P08024 (2007). doi:10.1088/1742-5468/2007/08/P08024
• U. Schollwöck, “The Density-Matrix Renormalization Group in the Age of Matrix Product States,” Annals of Physics 326, 96–192 (2011). doi:10.1016/j.aop.2010.09.012
• N. Schuch, M. M. Wolf, F. Verstraete, and J. I. Cirac, “Computational Complexity of Projected Entangled Pair States,” Physical Review Letters 98, 140506 (2007). doi:10.1103/PhysRevLett.98.140506
• F. Verstraete, V. Murg, and J. I. Cirac, “Matrix Product States, Projected Entangled Pair States, and Variational Renormalization Group Methods for Quantum Spin Systems,” Advances in Physics 57, 143–224 (2008). doi:10.1080/14789940801912366
• G. Vidal, “Efficient Classical Simulation of Slightly Entangled Quantum Computations,” Physical Review Letters 91, 147902 (2003). doi:10.1103/PhysRevLett.91.147902
• G. Vidal, “Entanglement Renormalization,” Physical Review Letters 99, 220405 (2007). doi:10.1103/PhysRevLett.99.220405
• S. R. White, “Density Matrix Formulation for Quantum Renormalization Groups,” Physical Review Letters 69, 2863–2866 (1992). doi:10.1103/PhysRevLett.69.2863