Gaussian Continuous Tensor Network States

• Gaussian Continuous Tensor Network States (GCTNS) are a finite-parameter subclass of Gaussian states that serve as continuum limits of discrete tensor network states for bosonic quantum fields.
• They achieve analytical tractability by reducing observables to explicit kernel manipulations through a Gaussian auxiliary field theory with an affine coupling.
• GCTNS have been effectively employed to benchmark ground states of quadratic and weakly quartic Hamiltonians, accurately capturing short-distance and ultraviolet behaviors.

Gaussian Continuous Tensor Network States (GCTNS) are a finitely-parameterized subclass of Gaussian states admitting an interpretation as continuum limits of discrete tensor network states. Within the broader class of continuous tensor network states for bosonic quantum fields, they are the analytically tractable Gaussian corner: the auxiliary field theory is Gaussian, the coupling to the physical field is affine, the resulting physical state is Gaussian in the usual sense, and observables reduce to explicit kernel manipulations rather than generic interacting functional integrals (Karanikolaou et al., 2020, Rigobello et al., 17 Feb 2026).

1. Historical placement within the continuous tensor-network program

The conceptual origin of GCTNS lies in the program of replacing discrete tensor contraction by a continuum functional integral. In the one-dimensional setting of continuous matrix product states (cMPS), the path-integral representation was introduced as the natural continuum analogue of summing over bond indices, and it was used to show both that dynamically evolving quantum-field states admit a cMPS representation and that all states in Fock space admit a cMPS representation when the number of variational parameters tends to infinity (Brockt et al., 2012).

This viewpoint was then extended to higher-dimensional continuum tensor-network field states. A central idea is that a physical field state is generated by an auxiliary system of one lower spatial dimension, and that symmetries of the physical state are encoded in the auxiliary dynamics. In particular, the higher-dimensional construction obtained a well-behaved field limit of PEPS-like networks and linked Euclidean symmetries of the physical field to Lorentz-invariant auxiliary dynamics, while also inheriting area-law structure from the underlying PEPS construction (Jennings et al., 2012).

The formal continuous tensor network states (cTNS) introduced for bosonic quantum fields generalized cMPS to spatial dimensions $d \ge 2$. They were presented as genuine continuum limits of discrete tensor network states, Euclidean invariant by construction, and equipped with both a functional integral and an operator representation. A later viewpoint summarized the essential continuum move as replacing a sum over discrete indices with a functional integral, and identified Gaussian-representable states as the analytically accessible sector of the new framework (Tilloy et al., 2018, Pervishko et al., 2019).

2. Defining data and Gaussian specialization

In the general cTNS construction, one introduces an auxiliary $D$-component scalar field $\phi=(\phi_1,\dots,\phi_D)$ on the same $d$-dimensional domain $\Omega\subset\mathbb{R}^d$ as the physical bosonic field $\psi(x)$, together with local functions $V,\alpha$ of the auxiliary field and, in the most general form, a boundary functional $B$. The core functional-integral definition is

$$|V,B,\alpha\rangle = \int \mathcal{D}\phi \, B(\phi|_{\partial \Omega}) \exp\bigg\{-\int_\Omega d^d x \; \frac{1}{2}\sum_{k=1}^D [\nabla \phi_k(x)]^2 + V[\phi(x)] - \alpha[\phi(x)]\,\psi^\dagger(x)\bigg\}\,|0\rangle .$$

Here $D$ is the bond field dimension, the continuum analogue of bond dimension. In the thermodynamic-limit formulation used for translationally invariant benchmark problems, the same structure appears without an explicit boundary term (Tilloy et al., 2018, Karanikolaou et al., 2020).

The Gaussian subclass is obtained by truncating $D$0 to at most quadratic order and $D$1 to at most linear order: $D$2 In this case the cTNS is called Gaussian, and the physical bosonic state is “also a Gaussian state in the usual sense” (Tilloy et al., 2018).

For translationally invariant GCTNS, several redundancies can be removed. $D$3 only affects normalization, $D$4 can be absorbed into a shift of $D$5, and if $D$6 is diagonalizable one may assume a diagonal auxiliary mass matrix $D$7. The resulting family is parameterized by one complex scalar $D$8, one complex vector $D$9, and one complex vector $\phi=(\phi_1,\dots,\phi_D)$0, giving only $\phi=(\phi_1,\dots,\phi_D)$1 complex parameters. This is a highly structured finite-dimensional submanifold inside the full manifold of continuum Gaussian states (Karanikolaou et al., 2020).

3. Exact solvability, generating functionals, and Gaussianity criteria

The principal technical advantage of GCTNS is that the doubled auxiliary path integral remains Gaussian. For generic cTNS, the generating functional for normal-ordered correlators involves a doubled auxiliary-field functional integral with a bra-ket coupling term $\phi=(\phi_1,\dots,\phi_D)$2, and this makes expectation values analytically inaccessible in general. In the Gaussian subclass, by contrast, the generating functional becomes explicitly Gaussian in the sources, so all local observables reduce to finite-dimensional matrix algebra and momentum integrals (Karanikolaou et al., 2020).

In the benchmark formulation, the generating functional takes the form

$\phi=(\phi_1,\dots,\phi_D)$3

where $\phi=(\phi_1,\dots,\phi_D)$4 is the inverse kernel of a $\phi=(\phi_1,\dots,\phi_D)$5 matrix differential operator. Translation invariance gives $\phi=(\phi_1,\dots,\phi_D)$6, and in momentum space

$\phi=(\phi_1,\dots,\phi_D)$7

The momentum-space two-point function then has the rational form

$\phi=(\phi_1,\dots,\phi_D)$8

with $\phi=(\phi_1,\dots,\phi_D)$9 the eigenvalues of $d$0. Thus a GCTNS correlator is an even rational function of $d$1, and every correlator reduces to matrix inversion plus explicit momentum integrals (Karanikolaou et al., 2020).

Because the state is Gaussian, higher correlators are obtained by Wick’s theorem. This is the mechanism behind the exact evaluation of quartic expectation values in Gaussian variational calculations (Karanikolaou et al., 2020).

A structurally important caveat is that quadratic $d$2 alone does not suffice. In the cTNS formalism, even if $d$3 is quadratic, non-Gaussianity enters immediately if $d$4 is nonlinear, because $d$5 appears quadratically through the overlap of coherent states. A cTNS is Gaussian if and only if the full exponent in the doubled functional integral remains quadratic in the auxiliary fields $d$6, which requires $d$7 to be at most quadratic and $d$8 to be at most affine (Tilloy et al., 2018).

The same Gaussian tractability appears in explicit examples. For the simplest $d$9 translation-invariant case with $\Omega\subset\mathbb{R}^d$0, the physical two-point function has ultraviolet asymptotic behavior $\Omega\subset\mathbb{R}^d$1, so the equal-point density is ultraviolet finite for $\Omega\subset\mathbb{R}^d$2 (Tilloy et al., 2018).

4. Variational role and benchmark performance

GCTNS were introduced as a tractable subclass of CTNS precisely because generic CTNS local observables are not analytically computable. In simple bosonic models, the Gaussian subclass proved sufficiently explicit for serious benchmarking. The central results are that GCTNS provide arbitrarily accurate approximations to the ground states of quadratic Hamiltonians, give decent estimates for quartic ones at weak coupling, and capture the short-distance behavior of the benchmark theories exactly well enough to renormalize away simple divergences variationally (Karanikolaou et al., 2020).

For the quadratic bosonic Hamiltonian

$\Omega\subset\mathbb{R}^d$3

the exact ground state is Gaussian, so it is a natural target for GCTNS. A finite-$\Omega\subset\mathbb{R}^d$4 GCTNS has correlator $\Omega\subset\mathbb{R}^d$5, which cannot equal the exact square-root correlator at finite $\Omega\subset\mathbb{R}^d$6, but it can approximate it arbitrarily well by increasing $\Omega\subset\mathbb{R}^d$7. Numerical evidence in $\Omega\subset\mathbb{R}^d$8 and $\Omega\subset\mathbb{R}^d$9 showed rapid convergence of both energy density and two-point functions, slower only near the gap-closing point $\psi(x)$0 (Karanikolaou et al., 2020).

In $\psi(x)$1, the same analysis exhibited a model-specific renormalization mechanism. With hard cutoff $\psi(x)$2, the GCTNS energy density has the form

$\psi(x)$3

and the divergent coefficient is minimized exactly on the submanifold

$\psi(x)$4

On that submanifold, the coefficient of the logarithmic divergence matches the exact theory (Karanikolaou et al., 2020).

Beyond exactly Gaussian targets, the one-dimensional Lieb–Liniger Hamiltonian

$\psi(x)$5

provided a test of the Gaussian restriction itself. Since the true ground state is non-Gaussian, no Gaussian ansatz can approximate it arbitrarily well at all couplings. Nevertheless, GCTNS energies approach the exact Bethe-Ansatz ground-state energy well at weak coupling, and $\psi(x)$6 already captures essentially all the improvement available within the Gaussian class (Karanikolaou et al., 2020).

This Gaussian tractability was already identified, at the viewpoint stage, as the main analytically controlled sector of the continuum tensor-network program. The same commentary emphasized that exact derivation of correlation functions was then available only for states representable in a Gaussian basis and that extension to the non-Gaussian case remained open (Pervishko et al., 2019).

5. Rational kernels, short-distance structure, and Lifshitz behavior

A later analysis sharpened the structural characterization of GCTNS. In Schrödinger form, a continuous tensor-network wavefunctional is written as

$\psi(x)$7

with $\psi(x)$8 auxiliary fields. In the Gaussian case, $\psi(x)$9 is quadratic, and after integrating out the virtual fields one obtains

$V,\alpha$0

For first-derivative, translation- and rotation-invariant GCTNS with $V,\alpha$1 auxiliary fields, the physical kernel is

$V,\alpha$2

a rational function of $V,\alpha$3. More generally, $V,\alpha$4-th derivative GCTNS with $V,\alpha$5 auxiliary fields have numerator and denominator degrees $V,\alpha$6 and $V,\alpha$7, respectively (Rigobello et al., 17 Feb 2026).

This rational characterization implies the paper’s main ultraviolet result: at short distance, finite-parameter GCTNS correspond to free Lifshitz vacua. Since a rational function of $V,\alpha$8 approaches a monomial at large momentum, the ultraviolet dispersion behaves as $V,\alpha$9 with even dynamical exponent

$B$0

Finite-$B$1 GCTNS therefore do not reproduce relativistic $B$2 asymptotics in a local canonical field basis; instead they flow to a Lifshitz ultraviolet fixed point (Rigobello et al., 17 Feb 2026).

Two explicit approximation schemes were developed on this basis for free bosonic theories, especially Klein–Gordon theory. The first uses rational approximants to the exact dispersion $B$3, often in continued-fraction form. The second uses Trotterized imaginary-time evolution, which yields a tridiagonal GCTNS at each finite Trotter step. Both schemes systematically improve the low-energy approximation while introducing a soft ultraviolet breakdown scale. For the continued-fraction approximation of Klein–Gordon theory, the crossover scale is estimated as $B$4; for the imaginary-time scheme it is $B$5 (Rigobello et al., 17 Feb 2026).

The same analysis connects short-distance structure to entanglement. In $B$6 dimensions, free Lifshitz vacua with exponent $B$7 have logarithmic ultraviolet coefficient $B$8. GCTNS approximants to Klein–Gordon theory therefore display an intermediate relativistic window with effective coefficient $B$9, but their true ultraviolet scaling crosses over to the Lifshitz value dictated by the finite-$$|V,B,\alpha\rangle = \int \mathcal{D}\phi \, B(\phi|_{\partial \Omega}) \exp\bigg\{-\int_\Omega d^d x \; \frac{1}{2}\sum_{k=1}^D [\nabla \phi_k(x)]^2 + V[\phi(x)] - \alpha[\phi(x)]\,\psi^\dagger(x)\bigg\}\,|0\rangle .$$0 ansatz. This identifies both the strength and the limitation of GCTNS as variational ansätze for relativistic quantum fields (Rigobello et al., 17 Feb 2026).

6. Symmetries, operator formulations, and extensions

GCTNS inherit the broader cTNS structural features. The original cTNS construction was Euclidean invariant by construction, admitted both a functional integral and an operator representation, and shared with discrete tensor networks expressiveness, invariance under gauge transformations, simple rescaling flow, and compact expressions for the $$|V,B,\alpha\rangle = \int \mathcal{D}\phi \, B(\phi|_{\partial \Omega}) \exp\bigg\{-\int_\Omega d^d x \; \frac{1}{2}\sum_{k=1}^D [\nabla \phi_k(x)]^2 + V[\phi(x)] - \alpha[\phi(x)]\,\psi^\dagger(x)\bigg\}\,|0\rangle .$$1-point functions of local observables. In operator form, a $$|V,B,\alpha\rangle = \int \mathcal{D}\phi \, B(\phi|_{\partial \Omega}) \exp\bigg\{-\int_\Omega d^d x \; \frac{1}{2}\sum_{k=1}^D [\nabla \phi_k(x)]^2 + V[\phi(x)] - \alpha[\phi(x)]\,\psi^\dagger(x)\bigg\}\,|0\rangle .$$2-dimensional physical field is generated by a non-unitary evolution of an auxiliary theory in one lower dimension, with a transfer-operator description for correlators (Tilloy et al., 2018).

The symmetry logic of the continuum tensor-network program is especially important. In the higher-dimensional field-state construction, physical spatial symmetries are encoded in the lower-dimensional auxiliary dynamics, and Euclidean symmetries of the physical state require Lorentz-invariant auxiliary theories before analytic continuation. The resulting continuum PEPS-like states inherit area-law structure from the PEPS class and are fully described by the dissipative dynamics of a lower-dimensional virtual field system. This provides the symmetry architecture within which Gaussian subclasses are most naturally formulated (Jennings et al., 2012).

Gauge-invariant continuum generalizations have also appeared. The continuum limit of certain gauged tensor network states defines gauged continuous PEPS (gCPEPS) and, in one dimension, gauged cMPS-like states. That construction is not itself a Gaussian theory, but it explicitly identifies the subset of “gauged Gaussian continuous PEPS” as a promising tractable direction. In this sense, gauge symmetry supplies a scaffold for future gauge-invariant Gaussian continuum tensor networks rather than an already completed GCTNS formalism (Roose et al., 13 Nov 2025).

Two nearby literatures are conceptually related but distinct. Efficient tensor-network constructions of many-body Gaussian states based on Bogoliubov and Bloch–Messiah decompositions produce explicit discrete MPO representations of bosonic and fermionic Gaussian states, but they do not define a genuinely continuous tensor network in the field-theoretic sense (Nüßeler et al., 2020). Likewise, continuous-valued MPS Born machines for machine learning model continuous random variables at finitely many sites and are explicitly “not the quantum-field-theoretic cMPS with continuous spatial coordinate,” so they should not be conflated with GCTNS (Meiburg et al., 2023).

GCTNS therefore occupy a specific position in the broader landscape: they are the exactly solvable Gaussian sector of continuum tensor networks for bosonic quantum fields, simultaneously interpretable as continuum limits of PEPS-like constructions and as finitely parameterized rational-kernel Gaussian states. Their main importance lies in combining continuum tensor-network structure, exact Gaussian computability, controlled variational benchmarking, and a sharply characterized ultraviolet limitation.