Matrix Product States, Finite Entanglement, and Continuum Limits

An MPS recovers one-dimensional QFT information by turning the many-body state into finite-dimensional transfer and tangent-space problems, then removing finite-entanglement, finite-volume, local-Hilbert, lattice-spacing, and optimization controls separately. The MPS correlation length is an infrared scale induced by finite bond dimension; it is not itself the continuum limit.

Required background. Entanglement structure and tensor-network ansätze supplies physical and virtual indices, Schmidt truncation, and tensor gauge. Variational principles and field-theory ansätze supplies energy stationarity and residual tests.

Helpful background. Lines of constant physics and continuum extrapolation supplies the tuning and spacing-limit contract.

MPS transfer channels and physical length scales

For a translation-invariant infinite MPS with tensor $A^s\in\mathbb C^{\chi\times\chi}$, define

$$\mathcal E(X)=\sum_s A^sX(A^s)^\dagger.$$

Choose a canonical normalization so the spectral radius is $\lambda_0=1$ and the corresponding left and right fixed points normalize expectation values. If the channel is injective, $\lambda_0$ is nondegenerate and connected correlators have the spectral expansion

$$\langle O_0P_r\rangle_c =\sum_{k\geq1}c_k(O,P)\lambda_k^r, \qquad \xi_k=-\frac{a}{\log|\lambda_k|}.$$

A negative or complex $\lambda_k$ adds oscillation but does not change the decay length obtained from its modulus. A degenerate unit-modulus eigenvalue instead signals non-injectivity, symmetry breaking, or a cat-state structure; applying the injective formula without resolving sectors is a diagnostic failure. Transfer-channel methods and canonical gauges are developed in Schollwöck 2011, §§ 4–5, and the broader transfer-operator and injectivity structure is reviewed by Cirac et al. 2021.

Regulator and convention box. Record whether the MPS is open, periodic, or uniform; the unit cell; canonical gauge; bond dimension $\chi$; local dimension $d_{\rm loc}$; lattice spacing $a$; physical size $L$; symmetry sector; optimizer residual; and which transfer eigenvalue couples to each operator. Quote $\xi/a$ before scale setting. Do not combine a $\chi$ scan with changes in $d_{\rm loc}$ or bare couplings unless the joint path is explicitly modeled.

Exactly checkable transfer channel

Take a physical dimension four and bond dimension two with

$$A^0=\sqrt{1-3b}\,I, \qquad A^j=\sqrt b\,\sigma_j\quad(j=x,y,z), \qquad 0<b<\frac14.$$

Because $\sum_sA^s(A^s)^\dagger=I$, the channel is unital. Using $\sum_j\sigma_j\sigma_i\sigma_j=-\sigma_i$ gives

$$\mathcal E(I)=I, \qquad \mathcal E(\sigma_i)=(1-4b)\sigma_i.$$

Thus the transfer spectrum is $\{1,1-4b,1-4b,1-4b\}$ and

$$\frac{\xi}{a}=-\frac{1}{\log|1-4b|}.$$

This fixture checks channel construction, normalization, degeneracy, and the absolute value in $\xi$. It does not identify a Hamiltonian mass gap without showing that an operator couples to the mode and controlling the Euclidean-time or tangent-space relation.

Finite entanglement at a critical point

An exact critical ground state has divergent correlation length and logarithmic interval entropy. A finite-$\chi$ MPS instead produces a finite $\xi_\chi$. In the asymptotic finite-entanglement regime,

$$S_{\rm half}(\chi) =\frac{c}{6}\log\!\left(\frac{\xi_\chi}{a}\right)+s_0+\cdots,$$

and often $\xi_\chi/a\sim C\chi^\kappa$. The exponent and correction terms require the critical universality class and an appropriate scaling window; a log–log straight line over three small bond dimensions is not sufficient. The original finite-entanglement analyses and the separation from finite-size scaling are given by Tagliacozzo et al. 2008, §§ II–IV, Pollmann et al. 2009, and Pirvu et al. 2012, §§ II–III.

For a finite ring, two infrared scales compete:

$$\xi_\chi \quad\text{and}\quad L=Na.$$

Finite-entanglement scaling requires $L/\xi_\chi$ large enough that the bond cutoff sets the infrared scale. Conventional finite-size scaling requires $\xi_\chi/L$ large enough that the MPS resolves the finite ring. Data straddling the crossover should be fit with a two-variable scaling form, not silently pooled.

The geometry map locates MPS on the Hamiltonian-state branch and shows why its transfer spectrum still precedes, rather than replaces, the physical regulator limits.

For MPS, $\chi$ controls the variational representation and $\xi_\chi$ is read from the transfer channel. The schematic map emphasizes that neither quantity removes $d_{\rm loc}$, $L$, $a$, optimizer, operator-matching, or real-time controls.

Spectra and excitations

Ground-state transfer eigenvalues determine spatial decay scales. Energies require the Hamiltonian. A uniform-MPS tangent excitation with momentum $p$ has the form

$$|\Phi_p(B)\rangle =\sum_{n\in\mathbb Z}e^{ipna} \cdots A^{s_{n-1}}B^{s_n}A^{s_{n+1}}\cdots|\boldsymbol s\rangle.$$

After removing tangent gauge null directions, stationarity gives a generalized eigenproblem

$$H_{\rm eff}(p)b=\omega(p)N_{\rm eff}(p)b.$$

The resulting $\omega(p)$ is variational within the chosen tangent space, not an exact spectrum. Check momentum conventions, norm conditioning, multi-particle thresholds, and $\chi$ dependence. A relativistic continuum claim additionally tests

$$\omega(p)^2=M^2+v^2p^2+O(a^2p^4),$$

with the emergent velocity $v$ determined rather than assumed.

Joint continuum protocol

The safe sequence is observable dependent, but a useful default is:

1. At fixed bare lattice parameters, converge the optimizer and $d_{\rm loc}$ for several $\chi$.
2. Resolve finite-$L$ versus finite-$\chi$ scaling with both $L/\xi_\chi\gg1$ and $\xi_\chi/L\gg1$ checks where feasible.
3. Tune a line of constant physics using dimensionless inputs that are not reused as held-out validation.
4. Match or renormalize the operator at each $a$.
5. Extrapolate in $a$ with correlated uncertainties and at least one justified correction model.
6. Repeat a selected point with an alternative unit cell, initialization, or algorithm.

The joint control structure is shown below. Inspect the dashed failure path: apparent convergence in $\chi$ is rejected when any environment, local-space, volume, spacing, or optimization scan moves the observable.

An MPS error statement is joint and observable specific. The schematic diagram does not prescribe additive errors: it requires fixed-axis and crossed scans, covariance-aware extrapolation, and an explicit weaker claim whenever one control remains open.

The declarations needed for this program are collected in the chapter’s tensor-network regulator and error record.

A reproducible critical-Ising MPS and Euclidean comparison must control size, bond dimension, contraction, and optimization jointly; its data do not substitute for the spacing extrapolation.

Adversarial failure: a false critical exponent

Suppose $\xi_\chi$ grows smoothly for $\chi=32,48,64$, and a power-law fit returns a stable $\kappa$. If $d_{\rm loc}$ is too small, the local cutoff can move the critical bare coupling while leaving each fixed-$d_{\rm loc}$ fit visually excellent. Retuning at $d_{\rm loc}+2$ may change both $\kappa$ and the inferred central charge. The repair is a crossed $(\chi,d_{\rm loc})$ scan and a held-out observable such as an excitation ratio or scaling dimension.

Observable-level validation checklist

• Confirm canonical normalization and left/right transfer fixed points.
• Identify the transfer eigenmode coupled to the stated operator.
• Test both finite-size and finite-entanglement regimes rather than mixing them.
• Report energy variance or projected residual and initialization dependence.
• Vary $d_{\rm loc}$ independently for bosonic or gauge-link local spaces.
• Check dispersion and at least one held-out ratio against exact diagonalization or another method.
• Renormalize the observable and state the order of $\chi$, $L$, $d_{\rm loc}$, and $a$ limits.

What you should be able to do

After this page, you should be able to:

1. construct an MPS transfer channel, normalize it, and extract the operator-relevant correlation length or a tangent-space excitation scale; and
2. design a continuum analysis that distinguishes finite entanglement from finite size, local Hilbert truncation, lattice spacing, and optimization error.

Exercises

1. Transfer correlation length. For the Pauli-channel fixture with $b=1/8$, compute $\xi/a$.

Solution

Here $1-4b=1/2$, so $\xi/a=-1/\log(1/2)=1/\log2\simeq1.4427$.

2. Scaling regime. Two calculations have $(L/a,\xi_\chi/a)=(512,40)$ and $(64,400)$. Which is suited to finite-entanglement scaling and which to finite-size scaling?

Solution

The first has $L/\xi_\chi=12.8$, so the bond-induced correlation length is the smaller infrared scale and finite-entanglement scaling is plausible. The second has $\xi_\chi/L=6.25$, so the finite ring is the smaller infrared scale and finite-size scaling is plausible. Neither ratio alone proves asymptotia; nearby sizes and bond dimensions must confirm stability.

References

• Cirac, J. Ignacio, David Pérez-García, Norbert Schuch, and Frank Verstraete. “Matrix Product States and Projected Entangled Pair States: Concepts, Symmetries, Theorems.” Reviews of Modern Physics 93 (2021): 045003. DOI.
• Pirvu, Bogdan, Guifré Vidal, Frank Verstraete, and Luca Tagliacozzo. “Matrix Product States for Critical Spin Chains: Finite-Size Scaling versus Finite-Entanglement Scaling.” Physical Review B 86 (2012): 075117. DOI.
• Pollmann, Frank, Subroto Mukerjee, Ari M. Turner, and Joel E. Moore. “Theory of Finite-Entanglement Scaling at One-Dimensional Quantum Critical Points.” Physical Review Letters 102 (2009): 255701. DOI.
• Schollwöck, Ulrich. “The Density-Matrix Renormalization Group in the Age of Matrix Product States.” Annals of Physics 326 (2011): 96–192. DOI.
• Tagliacozzo, Luca, Thiago R. de Oliveira, Salvatore Iblisdir, and José I. Latorre. “Scaling of Entanglement Support for Matrix Product States.” Physical Review B 78 (2008): 024410. DOI.