Finite-Entanglement Scaling in Tensor Networks

• Finite-entanglement scaling is a framework where finite bond dimensions in tensor network states induce an effective correlation length, acting as an infrared cutoff.
• It enables extraction of key universal properties such as central charge, scaling dimensions, and critical exponents through controlled numerical procedures.
• The approach bridges finite-size scaling with tensor network methods and extends to various models, including 1D critical chains and 2D systems via iPEPS.

Searching arXiv for papers on finite-entanglement scaling and closely related developments. Finite-entanglement scaling is the scaling framework in which the finite bond dimension $D$ or $\chi$ of a matrix product state (MPS), continuous MPS, or related tensor-network ansatz is treated as the control parameter that induces an effective infrared cutoff. In a critical system, finite bond dimension bounds the bipartite entanglement entropy, generates an emergent correlation length $\xi_D \sim D^\kappa$, and perturbs the state away from the conformal fixed point while preserving universal scaling relations in terms of the induced length scale. First established for infinite translationally invariant critical spin chains, the framework has since been used to extract central charges, scaling dimensions, and critical exponents, to distinguish finite-size and finite-entanglement regimes, and to formulate finite-correlation-length scaling in higher-dimensional tensor networks (0712.1976, Pirvu et al., 2012, Stojevic et al., 2014, Xu et al., 2024, Mortier et al., 2023).

1. Foundational definition and scaling laws

The starting point is the incompatibility between critical entanglement growth and the finite Schmidt rank of an MPS. For an infinite, translationally invariant critical chain, a block of $L$ contiguous sites has entanglement entropy

$S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$

where $c$ is the central charge. By contrast, an MPS of bond dimension $\chi$ can never exceed

$S_{\max}=\log \chi.$

Calabrese–Cardy scaling away from criticality gives

$S(\xi)\simeq (c/6)\,\log \xi,$

with $\xi$ the correlation length. At the true critical point, a finite-$\chi$0 MPS exhibits saturation of the half-chain entropy according to

$\chi$1

where $\chi$2 is the emergent MPS correlation length. Numerical evidence then supports

$\chi$3

and hence

$\chi$4

For the Ising universality class, $\chi$5 and $\chi$6 give $\chi$7; for the Heisenberg chain, the reported value is $\chi$8 (0712.1976).

A complementary formulation uses $\chi$9 rather than $\xi_D \sim D^\kappa$0 and emphasizes conformal invariance. At a conformally invariant critical point, one observes the universal scaling laws

$\xi_D \sim D^\kappa$1

together with the analytic relation

$\xi_D \sim D^\kappa$2

This relation yields $\xi_D \sim D^\kappa$3 for the critical quantum Ising chain and $\xi_D \sim D^\kappa$4 for the Heisenberg chain (Nagy, 2011, Pirvu et al., 2012, Huang et al., 2023).

Model or universality class	Reported $\xi_D \sim D^\kappa$5	Reported $\xi_D \sim D^\kappa$6
Quantum Ising	$\xi_D \sim D^\kappa$7	$\xi_D \sim D^\kappa$8, $\xi_D \sim D^\kappa$9, or $L$0
Spin-$L$1 Heisenberg	$L$2	$L$3 or $L$4
ANNNI floating phase	$L$5	$L$6

These laws motivate the term “finite-entanglement scaling”: the role played by physical size in finite-size scaling is taken over by the entanglement support permitted by the variational class (0712.1976, Nagy, 2011).

2. Effective correlation length and the transfer-matrix perspective

The induced length scale is defined from the transfer matrix of the tensor-network state. In an infinite MPS, the leading transfer-matrix eigenvalues determine a finite correlation length,

$L$7

or, in equivalent conventions,

$L$8

For iPEPS, the analogous expression is

$L$9

In all of these settings, finite bond dimension produces a nonzero transfer-matrix gap and therefore a finite effective correlation length even at criticality (0712.1976, Xu et al., 2024).

A major refinement is to use the induced correlation length itself as the scaling variable. In the notation of translation-invariant (c)MPS, one defines

$S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$0

with $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$1 the transfer-matrix eigenvalue with largest nonzero real part. The finite-entanglement-scaling hypothesis is then written directly as

$S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$2

and, for a two-point function of an operator of scaling dimension $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$3,

$S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$4

This choice is motivated by several theoretical arguments and by the numerical observation that scaling with $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$5 produces significantly smaller error bars than scaling with $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$6 (Stojevic et al., 2014).

The same transfer-matrix viewpoint underlies higher-dimensional generalizations. In iPEPS, one builds the row-to-row transfer matrix $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$7, approximates its leading eigenvector by an iMPS of bond dimension $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$8 using CTMRG, and extracts $S(L)\simeq (c/3)\,\log L \qquad (L\gg 1),$9 to obtain $c$0. In this formulation, one tunes $c$1 so that the effective correlation length is dominated by $c$2 (Xu et al., 2024).

3. Empirical diagnostics and extraction of conformal data

The original finite-entanglement-scaling analysis assembled five independent tests of the scaling hypothesis. First, direct measurement of $c$3 gives $c$4 with a single exponent $c$5. Second, at criticality the order parameter obeys

$c$6

For the Ising model, with $c$7 and $c$8, fitting $c$9 against $\chi$0 yields $\chi$1. Third, the pseudo-critical point $\chi$2 shifts according to

$\chi$3

giving $\chi$4 in the Ising case. Fourth, the entropy of a finite block satisfies $\chi$5 for $\chi$6 and saturates for $\chi$7, with the saturation length obeying $\chi$8. Fifth, critical exponents can be extracted by data collapse, for example by plotting $\chi$9 versus $S_{\max}=\log \chi.$0 and tuning $S_{\max}=\log \chi.$1 until the curves collapse onto a universal curve (0712.1976).

The algorithmic setting in which these tests were first carried out was the infinite time evolving block decimation algorithm. In the two-site unit-cell ansatz,

$S_{\max}=\log \chi.$2

the diagonal matrix $S_{\max}=\log \chi.$3 contains the Schmidt coefficients, and the half-chain entropy is

$S_{\max}=\log \chi.$4

The iTEBD workflow consists of starting from a random translationally invariant MPS of bond dimension $S_{\max}=\log \chi.$5, applying a Suzuki–Trotter decomposition of $S_{\max}=\log \chi.$6 into two-site gates with small time step $S_{\max}=\log \chi.$7, truncating after each gate to the $S_{\max}=\log \chi.$8 largest singular values, and iterating $S_{\max}=\log \chi.$9 until energy, entropy, and local observables converge. The results were cross-checked against standard infinite- and finite-DMRG runs, with agreement to $S(\xi)\simeq (c/6)\,\log \xi,$0 in energy and $S(\xi)\simeq (c/6)\,\log \xi,$1 in entropy away from the thin critical window (0712.1976).

Subsequent work turned these observations into a systematic procedure for extracting conformal data. In translationally invariant (c)MPS and MPS simulations, one performs linear regression of $S(\xi)\simeq (c/6)\,\log \xi,$2 versus $S(\xi)\simeq (c/6)\,\log \xi,$3 to obtain $S(\xi)\simeq (c/6)\,\log \xi,$4, and of $S(\xi)\simeq (c/6)\,\log \xi,$5 versus $S(\xi)\simeq (c/6)\,\log \xi,$6 to obtain $S(\xi)\simeq (c/6)\,\log \xi,$7. Applied to the Lieb–Liniger model, the massless relativistic boson, and the critical quantum Ising chain, this yielded central charges within a few percent and scaling dimensions accurate to $S(\xi)\simeq (c/6)\,\log \xi,$8 for primary fields (Stojevic et al., 2014).

4. Finite-entanglement versus finite-size scaling, and unified crossover theory

In a critical one-dimensional system with periodic boundary conditions, finite size and finite entanglement generate two distinct artificial gaps. The finite-size gap is

$S(\xi)\simeq (c/6)\,\log \xi,$9

where $\xi$0 is the smallest scaling dimension of the CFT. The finite-entanglement gap is

$\xi$1

This produces two regimes in the $\xi$2 plane. Finite-size scaling (FSS) corresponds to

$\xi$3

so that the MPS is accurate enough to resolve the periodic-chain gap and standard CFT formulas such as

$\xi$4

apply. Finite-entanglement scaling (FES) is the complementary regime

$\xi$5

in which the MPS is effectively blind to the chain’s finite size or boundary conditions. A sharp crossover is observed when $\xi$6, and only the FSS regime maintains a finite overlap with the exact periodic-boundary-condition ground state in the large-$\xi$7 limit (Pirvu et al., 2012).

This distinction resolves a common misconception: finite-$\xi$8 simulations can yield correct universal local quantities while failing to reproduce global finite-$\xi$9 data. In the FES regime, local observables and local scaling laws remain meaningful, whereas boundary-condition-sensitive quantities require remaining in the FSS regime and extrapolating $\chi$00 afterwards (Pirvu et al., 2012).

A broader unification includes equilibrium criticality, finite bond dimension, finite system size, and linear driving. With $\chi$01, the four competing length scales are the equilibrium correlation length $\chi$02, the finite-bond-dimension length $\chi$03, the system size $\chi$04, and the finite-driving length $\chi$05 with $\chi$06. The scaling ansatz for the half-chain entanglement entropy is

$\chi$07

Choosing $\chi$08 so that one scaled variable is of order unity yields the leading forms for equilibrium scaling, FES, FSS, and finite-time scaling. In this unified picture, whichever of $\chi$09, $\chi$10, $\chi$11, or $\chi$12 is shortest controls the leading logarithmic growth of $\chi$13, while the longer scales enter as analytic subleading corrections (Cao et al., 2018).

5. Generalizations beyond the original one-dimensional setting

Finite-entanglement scaling was quickly applied to models with more intricate phase structure. In the one-dimensional ANNNI model, iMPS finite-entanglement scaling was used to study second-order transitions and the supposedly infinite-order Kosterlitz–Thouless transition. The analysis used the correlation length, entanglement entropy, and the second derivative of the energy with respect to the external field, together with an off-critical ansatz

$\chi$14

At $\chi$15, the FM–PM transition gave $\chi$16, $\chi$17, and $\chi$18. At $\chi$19 and $\chi$20, the floating phase was found to be critical with $\chi$21 and measured correlation-length exponent $\chi$22. At the floating–paramagnetic transition, $\chi$23 held up to $\chi$24, beyond which $\chi$25 bent downward, and fitting the crossover ansatz yielded $\chi$26 together with an essential singularity in $\chi$27 (Nagy, 2011).

In two-dimensional iPEPS, the framework reappears as finite-correlation-length scaling. For the disorder parameter

$\chi$28

one writes, in the critical region,

$\chi$29

Dimensional analysis gives

$\chi$30

For the $\chi$31-dimensional transverse-field Ising model, the extrapolated values were $\chi$32, $\chi$33, $\chi$34, and $\chi$35, implying a pure perimeter law at criticality. Data collapse used the scaling variable

$\chi$36

and the rescaled observables $\chi$37 and $\chi$38 (Xu et al., 2024).

A further extension addresses two-dimensional metals with a Fermi surface. For the optimal tensor-network state approximation to a metallic state, the entanglement entropy of a contractible region of linear size $\chi$39 is proposed to scale as

$\chi$40

with

$\chi$41

The scaling function satisfies $\chi$42 for $\chi$43 and $\chi$44 for $\chi$45, reproducing respectively the Widom-type $\chi$46 growth and area-law saturation. In the Gaussian-fermion PEPS study of a spinless circular Fermi surface at half filling, the reported scaling exponent was $\chi$47, and moderate bond dimensions $\chi$48 already produced correlation lengths of order $\chi$49 lattice sites. The same work also emphasized generalized Lieb–Schultz–Mattis–Oshikawa–Hastings constraints at fractional filling and their implications for charge-conserving PEPS (Mortier et al., 2023). This suggests that finite-entanglement scaling is not confined to one-dimensional conformal critical points, although the universal objects being scaled differ substantially.

6. Effective-field-theory and boundary-CFT interpretations

A more recent line of work interprets finite bond dimension as an actual relevant deformation of the critical theory rather than merely a numerical cutoff. In one formulation,

$\chi$50

where the most relevant operator allowed by the residual MPS symmetry controls the deformation. In another,

$\chi$51

with $\chi$52 at the lattice scale. Under RG,

$\chi$53

and the induced gap forms at a scale $\chi$54. This picture is explicitly compatible with the standard law $\chi$55 and with the central-charge relation

$\chi$56

The finite-entanglement perturbation therefore has a field-theoretic interpretation without invalidating the original scaling hypothesis (Huang et al., 2023, Schneider et al., 2024).

An important consequence is a boundary-CFT interpretation of the entanglement Hamiltonian. The reduced density matrix can be written as

$\chi$57

and the spectrum of $\chi$58 is that of a BCFT on a strip of width $\chi$59, with two boundaries: an entanglement boundary fixed by the bipartition and a physical boundary determined by the relevant deformation selected by the MPS symmetry. In the critical Ising chain, a generic MPS without $\chi$60 symmetry activates the $\chi$61 field and yields the $\chi$62 tower, whereas a $\chi$63-symmetric MPS forbids $\chi$64, activates $\chi$65, and yields the identity and $\chi$66 towers. Analogous constructions were demonstrated for the three-state Potts model and the free compact boson, including symmetry-engineered conformal boundary conditions and the recovery of the boson radius from the entanglement spectrum (Huang et al., 2023).

A related development is the proposal of a renormalization-group self-congruent point. The claim is that the running coupling at the induced infrared scale,

$\chi$67

approaches a $\chi$68-independent constant. The transfer-matrix spectrum then becomes scale-invariant but differs from the unperturbed BCFT values. For the Ising model, the reported universal transfer-matrix gap ratios are

$\chi$69

rather than the free–free BCFT values

$\chi$70

The paper argues that this does not alter the validity of finite-entanglement scaling because the self-congruent point lies within the scaling regime: bulk observables continue to scale with $\chi$71, while boundary-spectrum data are shifted in a universal, $\chi$72-independent way (Schneider et al., 2024).

Several caveats recur across the literature. Corrections of order $\chi$73 can bias estimates of $\chi$74 and $\chi$75 if the accessible bond dimensions are too small. The off-critical ansatz used in the ANNNI analysis is phenomenological and could in principle be refined. Weakly gapped regions, especially near Kosterlitz–Thouless transitions, may require sufficiently large $\chi$76 before deviations from critical scaling become visible. For finite systems with periodic boundary conditions, faithful extraction of global finite-$\chi$77 quantities still requires remaining in the finite-size-scaling regime rather than the finite-entanglement regime. In cMPS-based conformal-data extraction, the extraction of OPE coefficients remains to be implemented (Nagy, 2011, Pirvu et al., 2012, Stojevic et al., 2014).

Finite-entanglement scaling is therefore best understood as a family of scaling theories centered on an induced correlation length. In its original MPS form, it provides a thermodynamic-limit analogue of finite-size scaling. In its transfer-matrix formulation, it turns the emergent length scale into the primary observable. In its more recent effective-field-theory and BCFT formulations, it interprets finite bond dimension as a symmetry-constrained relevant perturbation with a nontrivial boundary spectrum. Across these formulations, the core statement remains the same: finite entanglement imposes a controlled infrared scale that can be used, rather than merely tolerated, to extract universal critical information (0712.1976, Stojevic et al., 2014, Huang et al., 2023, Schneider et al., 2024).