IsoTNS: Isometric Tensor Network States

• IsoTNS are tensor-network ansätze defined by local isometry conditions that extend the 1D MPS canonical form to higher dimensions.
• They enable efficient contraction by reducing complex 2D problems to effective 1D computations along an orthogonality hypersurface.
• IsoTNS can represent intricate quantum phenomena including topological order, critical behavior, fermionic systems, and thermal states.

Isometric tensor network states (isoTNS) are tensor-network ansätze in which local tensors satisfy directional isometry constraints, extending the canonical form of one-dimensional matrix product states (MPS) to higher-dimensional networks. In the standard two-dimensional construction, one chooses an orthogonality row and column whose intersection is the orthogonality center; tensors away from this orthogonality hypersurface contract to identities, so many local computations reduce from a two-dimensional contraction to an effective one-dimensional problem (Zaletel et al., 2019). The modern theory of isoTNS was shaped by two complementary developments: the original formulation of 2D isoTNS as a constrained, efficiently contractible subclass of PEPS, and the later proof that exact finite-bond-dimension isoTNS can represent Levin–Wen string-net fixed points and their finite-depth local-unitary deformations, establishing that long-range entanglement is not by itself an obstruction to an isometric representation (Soejima et al., 2019).

1. Canonical definition and geometric structure

In the two-dimensional PEPS language, a state is written as

$$\ket{\psi}=\sum_{\vec i}\mathcal{C}\!\left(\{T^{i_v}_{\alpha\beta\gamma\varepsilon}\}\right)\ket{\vec i},$$

with $\mathcal C$ denoting contraction of virtual indices. A local tensor becomes isometric once its legs are partitioned into incoming and outgoing sets. In one standard convention, if $i_v,\alpha,\beta$ are incoming and $\gamma,\varepsilon$ are outgoing, the local condition is

$$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$

This generalizes the MPS left- or right-canonical condition to higher dimension (Soejima et al., 2019).

The defining global structure is the orthogonality hypersurface. In 2D, one selects a row and a column and requires every tensor off that hypersurface to be isometric with arrows pointing toward it; their intersection is the orthogonality center. The hypersurface itself behaves like an MPS and admits its own movable center, so contraction of bra and ket cancels the exterior exactly and leaves only the effective one-dimensional object on the hypersurface. In this sense, isoTNS inherit from canonical MPS the identities $W^\dagger W=I$ locally and norm reduction to a distinguished central object globally (Soejima et al., 2019).

This geometry can also be formulated in a corner-centered orientation. In the qisoTNS construction, the orthogonality center is fixed at $(u,v)=(0,0)$, so all arrows point toward a corner and each local tensor satisfies a uniform directional isometry,

$\sum_{k,l,p}A_{x_{(m,n)}^{ijklp}A_{x_{(m,n)}^{\dag ijklp}=\mathbf{1}.$

The same paper emphasizes that each row or column of an isoTNS behaves like an MPS in canonical form, making the 2D causal structure explicit (Slattery et al., 2021).

2. Manipulating the orthogonality hypersurface

The central algorithmic problem in isoTNS is not defining the canonical structure but moving it while preserving locality and manageable bond dimension. The original 2D framework introduced the Moses Move, whose purpose is to factor a column wavefunction as

$\Lambda^\ell \approx A^\ell \Lambda,$

with $A^\ell$ an isometric column and $\mathcal C$0 the remaining orthogonality hypersurface. Locally, the core step rewrites a tripartite center tensor $\mathcal C$1 by splitting $\mathcal C$2 into $\mathcal C$3 with an isometry $\mathcal C$4, then minimizing the entanglement across $\mathcal C$5. This allows iterative conversion of a 1D MPS representation of a 2D state into a genuine 2D isoTNS and also underlies the nested time-evolution algorithm $\mathcal C$6 (Zaletel et al., 2019).

Real- and imaginary-time evolution use the same nested logic. In $\mathcal C$7, one Trotterizes the Hamiltonian into row and column terms, performs a 1D TEBD update on the active orthogonality hypersurface, and then shifts the hypersurface using the Moses Move. A later implementation study revisited this structure in detail, comparing disentanglers used inside the Moses Move and introducing $\mathcal C$8, where local optimization on the orthogonality center reduces to an ordinary eigenvalue problem because the local norm tensor is the identity. That work also stressed a limitation that remains characteristic of isoTNS algorithms: $\mathcal C$9 is not strictly variational globally, because each Moses Move is approximate (Lin et al., 2021).

Finite-temperature simulations preserve the same architecture. In purification-based thermal isoTNS, each physical tensor acquires an ancilla leg, the $i_v,\alpha,\beta$0 state is a product of local Bell pairs, and imaginary-time evolution

$i_v,\alpha,\beta$1

is implemented by a double-column $i_v,\alpha,\beta$2 scheme. The paper gives explicit costs for this purified double-column variant, including $i_v,\alpha,\beta$3 for the Moses Move and $i_v,\alpha,\beta$4 for the TEBD step, and uses the construction to represent thermal states of the 2D transverse-field Ising model (Kadow et al., 2023).

3. Expressive power, topological order, and criticality

A longstanding question was whether the isometric restriction excludes intrinsically two-dimensional long-range-entangled phases. The decisive result is that every Levin–Wen string-net fixed-point wavefunction admits an exact finite-bond-dimension isoTNS representation, with the orthogonality hypersurface placeable anywhere, and that any state related to such a fixed point by a finite-depth local unitary circuit also has an exact isoTNS representation with finite bond dimension. Since string-net models realize a broad class of nonchiral topological phases, including discrete gauge theories and all bosonic Abelian topological orders with gappable edges, this shows that topological order and long-range entanglement do not by themselves prevent an isoTNS form (Soejima et al., 2019).

Subsequent constructions showed that isoTNS can also interpolate between distinct topological phases and host nontrivial critical points. A bond-dimension-$i_v,\alpha,\beta$5 plumbed isoTNS path was constructed between distinct $i_v,\alpha,\beta$6-enriched $i_v,\alpha,\beta$7 topological phases. At the transition point $i_v,\alpha,\beta$8, the corresponding classical weight matrix reduces to a gapless point of the six-vertex model, and the critical wavefunction displays power-law correlations $i_v,\alpha,\beta$9 along one spatial direction while remaining long-range ordered in another (Liu et al., 2023).

A different route to criticality arises from stochastic automata. Mapping the Domany–Kinzel process to a two-dimensional quantum wavefunction yields an isoTNS with bond dimension $\gamma,\varepsilon$0 whose local isometry is exactly probability conservation,

$\gamma,\varepsilon$1

At the directed-percolation critical point, the resulting 2D quantum state has algebraic correlations in all spatial directions, with exponents $\gamma,\varepsilon$2 along the time-like direction and $\gamma,\varepsilon$3 in all other directions. The same construction produces a frustration-free local parent Hamiltonian with a degenerate ground-state manifold consisting of a product state and a second state that undergoes a transition from a $\gamma,\varepsilon$4-like long-range pairwise entanglement pattern to a short-range-entangled trivial state (Boesl et al., 25 May 2026).

The notion of an isoTNS “skeleton” extends these ideas away from fixed points for Abelian string-net models. In that framework, continuous deformations preserve both local isometry and the virtual symmetries that encode topological order, producing stable finite-correlation-length branches between string-net fixed points. The paper shows that these branches can meet at shared critical points and that generalized Pauli strings of arbitrary weight can be evaluated efficiently by mapping the 2D isoTNS to a 1D stochastic automaton with local update rules (Boesl et al., 17 Nov 2025).

4. Generalizations across statistics, dimensions, and target states

The isoTNS framework has been generalized in several orthogonal directions. For fermions, the main modification is a sign-dressed isometry in one quadrant of the network: in the upper-right quadrant, $\gamma,\varepsilon$5 rather than $\gamma,\varepsilon$6 itself is isometric. With this change, the orthogonality-hypersurface contraction logic survives, enabling a fermionic $\gamma,\varepsilon$7 algorithm for real- and imaginary-time evolution. Benchmarks cover gapped systems, a Dirac semimetal, and systems with gapless edge modes, while real-time simulations capture two-fermion scattering and chiral edge motion in a Chern insulator (Dai et al., 2022).

In three dimensions, the canonical structure becomes an embedded tree tensor network inside a cubic lattice. The distinctive new ingredient is a tetrahedral site-splitting with a tripartite disentangler, needed to move the embedded canonical tree from one 2D slice to the next. This leads to $\gamma,\varepsilon$8, a 3D imaginary-time algorithm benchmarked on the 3D transverse-field Ising model, with exact contraction in canonical form and overall linear scaling in system size for fixed $\gamma,\varepsilon$9, but steep local costs such as $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$0 in the dominant regime (Tepaske et al., 2020).

The ansatz class itself has also been refined. Alternating isoTNS reverse the vertical arrow orientation from column to column, and the paper argues that entanglement is mediated along isometric arrows. This improves representability for isotropic and strongly entangled states, including free-fermion Fermi surfaces, band insulators, $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$1 superconductors, and the interacting transverse-field Ising model. The same work introduces isoGfTNS, a Gaussian fermionic framework where isometry reduces to a covariance-matrix zero-block condition, making global Riemannian optimization possible (Wu et al., 15 Feb 2025).

For multistate calculations, block-isoPEPS generalize block-MPS to two dimensions by placing a block index on the orthogonality-center tensor while keeping all surrounding tensors shared and isometric. This yields exact block orthogonalization at the center and supports an inexact subspace-iteration algorithm for low-lying excitations in the 2D transverse-field Ising and Heisenberg models (Dektor et al., 22 Oct 2025).

A further extension goes in the opposite direction: holographic isoTNS for one-dimensional states. There the network is $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$2-dimensional, with physical degrees of freedom only on the bottom row and an auxiliary “holographic” direction supplying additional expressivity. Random instances exhibit volume-law entanglement, and the paper gives analytic or variational constructions for arbitrary fermionic Gaussian states, Clifford states, and certain short-time-evolved states (Kobayashi et al., 12 Dec 2025).

5. Circuit, sampling, and holographic viewpoints

Because local tensors are isometries, isoTNS admit direct circuit interpretations. The qisoTNS construction compiles each rank-5 site tensor into a unitary block $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$3 with one ancilla input fixed to $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$4. On an $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$5 lattice, the resulting holographic circuit needs only $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$6 qubits rather than $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$7, because the second spatial direction is represented by circuit time and physical qubits are reused after measurement. This produces a VQE ansatz that is qubit efficient and naturally layerwise in its trainability structure (Slattery et al., 2021).

A closely related line constructs exact linear-depth circuits for families of 2D topological states. In the plumbed isoTNS subclass, the local tensor

$$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$8

satisfies the isometry condition iff $$(T^{i_v}_{\alpha \beta \gamma' \varepsilon'})^* T^{i_v}_{\alpha \beta \gamma \varepsilon} = \delta_{\gamma,\gamma'}\delta_{\varepsilon,\varepsilon'}.$$9. This allows exact sequential preparation, and the same causal structure supports a holographic algorithm using only a 1D array of qubits: one prepares a row, measures local observables on it, resets the measured qubits, and reuses them for the next row (Liu et al., 2023).

Sampling algorithms make the same structure algorithmic on classical hardware. For a normalized $W^\dagger W=I$0 isoTNS, one can sample a full computational-basis configuration row by row, using exact local conditionals inside each row and approximate MPO–MPS contraction between rows. The resulting independent-sampling algorithm has total cost

$W^\dagger W=I$1

while a greedy search for the top-$W^\dagger W=I$2 high-probability configurations scales as

$W^\dagger W=I$3

Benchmarks on GHZ, $W^\dagger W=I$4, and random states show exactness in the no-truncation regime and controlled degradation when row-contraction truncation is introduced (Dektor et al., 2 Feb 2026).

IsoTNS also provide a natural language for holographic channel-state correspondences. A repeated $W^\dagger W=I$5-dimensional quantum channel can be represented as the transfer process of a $W^\dagger W=I$6-dimensional isoTNS, with the steady state appearing as the boundary reduced density matrix on virtual bonds. In that setting, strong symmetry of the channel becomes a virtual symmetry of the isoTNS, and strong-to-weak spontaneous symmetry breaking in the boundary mixed state is interpreted as boundary anyon condensation of bulk topological order (Lu et al., 24 Nov 2025).

6. Computational complexity, optimization, and open problems

Despite their efficient local contraction properties, isoTNS are not generically classically easy. A complexity-theoretic analysis maps 2D isoTNS contraction to $W^\dagger W=I$7D ancilla dynamics and proves that computing local expectation values in isoTNS is $W^\dagger W=I$8-complete. The same work introduces injective isoTNS, characterized by an injectivity parameter $W^\dagger W=I$9, and shows that injectivity induces depolarizing noise at rate

$(u,v)=(0,0)$0

Weakly injective isoTNS remain $(u,v)=(0,0)$1-complete, but for strongly injective isoTNS with $(u,v)=(0,0)$2 there is an efficient classical algorithm for local expectation values based on subcritical percolation of the effective noisy circuit (Malz et al., 2024).

Optimization theory for isometric tensor networks is correspondingly geometric. A general framework based on Riemannian optimization treats local isometric tensors as points on Stiefel or Grassmann manifolds, with explicit formulas for projected gradients, retractions, vector transport, and preconditioning by physically induced norm tensors. Although developed concretely for canonical MPS and MERA, this machinery transfers naturally to isoTNS because their variational parameters are likewise local isometries (Hauru et al., 2020).

Trainability is another area where isoTNS benefit from isometric structure. For MPS, TTNS, and MERA, energy minimization of local Hamiltonians is provably free of barren plateaus under Haar-random initialization of the isometric tensors; the analysis relies on bounded causal cones and gapped doubled transfer channels rather than on full Hilbert-space concentration. The same paper explicitly states that the techniques should also work for isometric PEPS, but it does not provide a theorem for general 2D isoTNS (Barthel et al., 2023). This suggests a distinction between rigorous results for lower-dimensional or tree-like isometric networks and plausible extensions to full 2D isoTNS.

A complementary open problem is approximate canonicalization. In one dimension, QR or SVD sweeps propagate the orthogonality defect exactly because the residual factor lives on a single bond. In higher dimensions, the residual generally has support on several outgoing directions. A recent gauge-propagation approach addresses this by approximating a local tensor or local cluster as

$(u,v)=(0,0)$3

where each $(u,v)=(0,0)$4 is locally isometric and the factors $(u,v)=(0,0)$5 can be absorbed independently along propagation directions. Benchmarks show that increasing the number of retained structured terms or enlarging the local cluster reduces both local residuals and accumulated gauge-propagation errors, making the method a plausible initializer or preconditioner for variational isoTNS algorithms (Jiang et al., 22 Jun 2026).

Several substantive questions remain unresolved. Exact representability is now known for large classes of nonchiral topological phases, but chiral and more generally ungappable-edge phases remain outside the proven scope (Soejima et al., 2019). Approximation theory is also incomplete: even when exact fixed-point or finite-depth-circuit representations exist, the bond-dimension scaling needed to approximate generic states in a phase is largely open. More specialized open questions include whether alternating isoGfTNS can represent gapless chiral free-fermion states (Wu et al., 15 Feb 2025), how best to restore fixed sitewise geometry after cluster-based isometrization (Jiang et al., 22 Jun 2026), and whether the stronger manifold-optimization theory already available for other isometric tensor networks can be extended to full 2D isoTNS with equally sharp guarantees (Hauru et al., 2020).

Taken together, these developments place isoTNS at an unusual intersection of tensor-network canonical structure, topological-state representability, circuit synthesis, stochastic-process duality, and computational complexity. The isometric restriction is severe enough to make contraction, measurement, and local optimization structurally controlled, but not so severe as to exclude broad classes of topological order, finite-temperature states, excitations, stochastic-critical wavefunctions, or even $(u,v)=(0,0)$6-hard instances.