Hybrid Quantum Tensor Networks

• Hybrid quantum tensor networks are frameworks that combine classically contractable tensors with quantum state representations or circuits, enabling efficient many-body simulations.
• They integrate diverse architectures, including tree tensor networks, MPS-VQC models, and MERA variants, to leverage both classical resources and quantum capabilities.
• The approach strategically separates contraction processes to mitigate noise and optimize performance in tasks like ground-state computation and machine learning applications.

Hybrid quantum tensor networks are tensor-network constructions in which some tensors are treated as classically contractable objects and others are represented by measurable quantum states, parameterized quantum circuits, or quantum-executable subcircuits. In the literature surveyed here, the term covers a family of related architectures rather than a single ansatz: hybrid tree tensor networks for many-body simulation, end-to-end tensor-network/variational-circuit learning systems, executable hybrid tensor networks for circuit knitting, and more formal frameworks that interpolate between classical tensor networks and quantum tensor networks through reduction operators and post-selection (Yuan et al., 2020, Jäger et al., 4 May 2026).

1. Formal definitions and conceptual scope

A foundational distinction is between a classical tensor and a quantum tensor. In the noise-aware HTN formalism, a quantum tensor is written as

$\psi^{i_1,i_2,\dots,i_b}_{j_1,j_2,\dots,j_n},$

where the upper indices are classical indices and the lower indices are quantum indices associated with physical qubits. Classical-index contractions are performed classically, while quantum-index contractions are implemented via quantum measurements. The same work rewrites a hybrid tree tensor network state as

$$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$

and then reformulates it in density-matrix language through local expansion maps, making noise propagation and physicality explicit (Harada et al., 2023).

A broader many-body definition appears in the original hybrid-TN simulation framework, where hybrid tensor networks are networks constructed by connecting both classical tensors and quantum tensors, and a representative two-layer clustered state is

$$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$

This formulation already isolates the central idea: local or clusterwise structure may be stored in quantum states, while inter-cluster structure is carried by contractable tensors or by a higher-level quantum tensor (Yuan et al., 2020).

A more abstract unification is given by the formal hybrid tensor network of “Entanglement is Half the Story: Post-Selection vs. Partial Traces,” where any TN is an HTN if it can be split into three parts: an isometric tensor network, its complex conjugate, and a set of real diagonal reduction operators $0 \preceq D_j \preceq I$ connecting the first TN with its conjugate. Aggregating these operators gives $D=\bigotimes_j D_j$, and the associated map is

$\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$

Within this formalism, quantum tensor networks are recovered when $D_j=I$, while classical tensor networks are recovered, up to a global scaling factor, when the $D_j$ are one-hot diagonal matrices. The same paper argues that the decisive structural distinction is not entanglement alone but whether Hilbert-space reduction is performed by partial trace or by post-selection (Jäger et al., 4 May 2026).

2. Major architectural families

One major family is the hybrid tree tensor network. In the earliest simulation-oriented formulation, the lower layer consists of subsystem quantum tensors and the upper layer encodes inter-subsystem correlations; this made it possible to represent systems of up to $8\times 8$ and $9\times 8$ qubits while using operations only on $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$0 and $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$1 qubits, respectively (Yuan et al., 2020). A later hTTN formulation specialized this idea to loop-free binary TTNs, replacing the upper classical part of the tree by one or more quantum tensors. That work emphasized the “virtual bond dimension” $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$2 of a quantum tensor, meaning the bond dimension required to represent that tensor exactly as a classical TN, and used this viewpoint to explain why replacing the top of a TTN is natural when bond dimensions grow upward through the tree (Schuhmacher et al., 2024).

A second family couples classical tensor networks directly to variational quantum circuits for machine learning. In one version, a classically implemented MPS compresses $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$3-dimensional image data into a low-dimensional feature vector, which is then encoded into a shallow VQC and trained jointly end-to-end; the classical–quantum boundary is explicitly adjustable because “an MPS can be realized precisely by a quantum circuit” (Chen et al., 2021). Closely related work uses a trainable MPS with output dimension four as a feature extractor for a 4-qubit VQC in binary MNIST classification, again optimizing both components together (Chen et al., 2020). An aeroelastic pipeline extends this design to time series: data are represented as an MPS, contracted with a trainable MPO-like network mapping $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$4, normalized into a quantum-state amplitude object, converted into a state-preparation circuit by a matrix product disentangler, and processed by a tensor-network-inspired VQC for classification or regression (Hickmann et al., 7 Aug 2025). In speech emotion recognition, HQTN-SER uses a trainable affine projection from a 32-dimensional PCA representation to $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$5 qubit angles and an MPS-inspired nearest-neighbor variational circuit whose outputs are fused with a learned classical latent embedding (Mohtashim et al., 14 May 2026).

A third family is built around isometric or hierarchical tensor networks. One example is the hybrid MERA

$$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$6

where a fixed non-parametric quantum state prepared by digital quantum annealing acts as a boundary tensor resource, classical shadows provide the interface, and all variational optimization is shifted to the classical MERA (Schuhmacher et al., 20 May 2026). Another is the 2D isoTNS algorithm, in which the variational state remains an isometric tensor network, but the classically hard construction of local effective Hamiltonians is delegated to a quantum computer through tomography-inspired measurements (Leontica et al., 17 Nov 2025).

A fourth family treats the hybrid tensor network as an operational representation of computation rather than only of a wavefunction. In qTPU, circuit cutting is rewritten as the contraction of a hybrid tensor network whose nodes are classical tensors of quasiprobability coefficients and quantum tensors whose entries are expectation values of executable subcircuits (Tornow et al., 2024). Tensor Quantum Programming proposes the explicit workflow

$$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$7

using MPS/TT vectors and MPOs as the native classical representation, a quantum device for matrix–vector multiplication when state ranks become too large, and MPS tomography to recover the output in tensor-network form (Termanova et al., 2024). A different operational hybridization is the stabilizer-MPO method, which treats Clifford layers through stabilizer propagation and the residual non-Clifford evolution as a product of low-bond-dimension MPO layers (Mello et al., 2024).

3. Contraction, encoding, and representation change

A recurrent theme is that hybridization becomes effective only when contraction rules are reorganized around what is classically cheap and what is quantum-accessible. The formal CTN-to-QTN bridge in (Jäger et al., 4 May 2026) begins by grouping open indices into inputs and outputs, performing an SVD, embedding the isometric factors into unitary or isometric circuit components, implementing singular values through ancilla-assisted controlled rotations after rescaling by the largest singular value, and then using post-selection to realize the target non-unitary map. In the $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$8 case, the smaller singular value is written as $$\ket{\psi_{\rm HT}} = \frac{1}{C} \sum_{\vec{i}_1,\dots,\vec{i}_N} \psi_{\vec{i}_1,\dots,\vec{i}_N} \ket{\psi_1^{\vec{i}_1}} \otimes \cdots \otimes \ket{\psi_N^{\vec{i}_N}},$$9, implemented with a controlled $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$0-rotation on an ancilla, and retained only when the ancilla is measured in $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$1. This construction is precisely why the paper identifies post-selection as the architectural hinge between the classical and quantum limits (Jäger et al., 4 May 2026).

For hybrid tree tensor networks, expectation values factor through effective observables. In the density-matrix HTTN framework, for a two-layer state and an observable $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$2, one has

$$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$3

with

$$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$4

This reduces evaluation of an $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$5-qubit effective state to measurements using only $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$6 qubits (Harada et al., 2023). The earlier hybrid-TN simulation work uses the same principle in amplitude form: lower-layer contractions produce effective matrices $$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$7, and the final scalar is obtained from the upper tensor (Yuan et al., 2020).

The hTTN ground-state algorithm generalizes contraction to a genuinely hybrid setting. If the open indices are classical, the relevant matrix element is

$$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$8

if the open index is quantum,

$$\ket{\tilde\psi} = \sum_{i_1,\dots,i_k} \alpha^{i_1,\dots,i_k}\ket{\psi_1^{i_1}}\otimes\cdots\otimes\ket{\psi_k^{i_k}}.$$9

That work stresses a concrete contraction-order rule: contract classical indices first, and delay quantum contractions until they are turned into measurable expectation values, because otherwise one effectively incurs tomography-like overhead (Schuhmacher et al., 2024).

In executable h-TNs for circuit knitting, the same logic appears at the level of postprocessing. The starting identity is

$0 \preceq D_j \preceq I$0

with $0 \preceq D_j \preceq I$1, but qTPU replaces brute-force summation over global quasiprobability instances with contraction of a hybrid tensor network whose quantum tensors are evaluated on QPUs and then classically contracted with optimized tensor-network paths (Tornow et al., 2024). Tensor Quantum Programming extends the same representational strategy to matrix–vector multiplication: a target MPO is approximated by a unitary MPO, mapped to sequential gates on system qubits plus ancillas, and realized as a non-unitary slice of a larger unitary via post-selection (Termanova et al., 2024).

4. Optimization, training, and time evolution

Hybrid quantum tensor networks support several distinct optimization regimes. In supervised learning, the most common pattern is end-to-end joint optimization of a classical tensor-network front end and a quantum back end. The MNIST/Fashion-MNIST hybrid classifier combines an MPS feature extractor with a VQC and optimizes both parts together by gradient descent, using the parameter-shift rule for quantum gradients (Chen et al., 2021). A related TN-VQE framework optimizes a parameterized unitary tensor network $0 \preceq D_j \preceq I$2 together with a PQC $0 \preceq D_j \preceq I$3 through

$0 \preceq D_j \preceq I$4

with the tensor network reshaping the Pauli coefficients of the Hamiltonian seen by the circuit (Cores et al., 2024). In the aeroelastic pipeline, the trainable MPO compression and the TN-inspired VQC are optimized jointly by autodiff with Adam (Hickmann et al., 7 Aug 2025). HQTN-SER likewise trains the classical encoder, the affine projection into qubit angles, the MPS-structured variational circuit, and the final classifier end-to-end, using parameter-shift for the quantum parameters (Mohtashim et al., 14 May 2026).

For many-body ground states, hybridization is usually coupled to local tensor updates. The hTTN optimization of (Schuhmacher et al., 2024) is explicitly sweep-based and environment-based, with effective Hamiltonians constructed by hybrid contractions and a VQE subroutine used only for the quantum tensor. The 2D isoTNS method is even closer to DMRG: writing

$0 \preceq D_j \preceq I$5

one optimizes the current center tensor by diagonalizing the local effective Hamiltonian, while the quantum processor supplies the classically hard entries of $0 \preceq D_j \preceq I$6 through tomography-inspired measurement (Leontica et al., 17 Nov 2025).

A different strategy is to freeze the quantum part entirely. In the hybrid MERA construction, the quantum resource is a fixed annealing-prepared state $0 \preceq D_j \preceq I$7, and only the classical MERA tensors are trained. Classical shadows provide unbiased estimators for the expectation values and gradients needed during optimization, while Riemannian ADAM enforces the isometric constraints on the classical network (Schuhmacher et al., 20 May 2026). This removes quantum-gradient estimation from the loop altogether.

Time evolution introduces another optimization regime. The MPS-based hTN time-evolution algorithm of (Bauer et al., 26 Jun 2026) splits the chain into classical boundary tensors and a central quantum tensor, evolves the classical part with the BUG integrator and the quantum part with any compatible quantum time-evolution subroutine, and couples them through boundary environments and bookkeeping tensors. The authors emphasize that the classical and quantum components can run in parallel during a single time step and are not constrained by synchronization barriers or mid-circuit measurements (Bauer et al., 26 Jun 2026).

5. Empirical performance across domains

In quantum simulation, the earliest hybrid-TN benchmarks already showed that a quantum processor significantly smaller than the target system can still support meaningful many-body calculations. For hybrid tree tensor networks, numerical benchmarks were reported for 1D and 2D spin systems of up to $0 \preceq D_j \preceq I$8 and $0 \preceq D_j \preceq I$9 qubits using operations only on $D=\bigotimes_j D_j$0 and $D=\bigotimes_j D_j$1 qubits, respectively (Yuan et al., 2020). Later hTTN work showed that, for the critical Ising model and the Toric code, hTTNs can improve upon classical equivalents with equal bond dimension in the classical part; in the $D=\bigotimes_j D_j$2 Toric code benchmark, only the hTTN recovered the exact ground-state energy $D=\bigotimes_j D_j$3 (Schuhmacher et al., 2024). In the hybrid MERA study, the optimized hMERA energy was consistently lower than the original quantum-annealing energy, with an improvement factor larger than 5 for $D=\bigotimes_j D_j$4 and still above 2 for $D=\bigotimes_j D_j$5, without increasing circuit depth (Schuhmacher et al., 20 May 2026). For 2D isoTNS optimization of the transverse-field Ising model, the tomography-based method reproduced the exact-contraction optimization profile with about $D=\bigotimes_j D_j$6 shots per tomography circuit, while the Lanczos variant reduced the total number of shots by about an order of magnitude and enabled ground-state optimization on up to 25 qubits (Leontica et al., 17 Nov 2025).

In hybrid quantum-classical processing of circuits, qTPU reported an average postprocessing overhead reduction of $D=\bigotimes_j D_j$7, up to $D=\bigotimes_j D_j$8, an average postprocessing speedup of $D=\bigotimes_j D_j$9, and an average end-to-end runtime speedup of $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$0 relative to Qiskit Addon Cutting (Tornow et al., 2024). In hybrid stabilizer-MPO simulation, the method reduced entanglement growth relative to direct MPS evolution in random Clifford $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$1-doped circuits and reproduced the analytical Floquet magnetization decay $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$2 for $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$3 (Mello et al., 2024).

In machine learning, the 4-qubit MPS-VQC classifier reported

$\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$4

for the end-to-end MPS-VQC model with $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$5, compared with about $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$6 test accuracy for the PCA-VQC baseline (Chen et al., 2020). The TN-VQE study found that, with a three-parameter, two-layer TN, the relative error could be reduced by almost three orders of magnitude compared to VQE without TN, and that three-layer PQCs with TN assistance often reduced relative error below $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$7 (Cores et al., 2024). In speech emotion recognition, HQTN-SER reported RAVDESS $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$8, SAVEE $\Lambda[\sigma_i]_A=\tr_B(U\sigma_i U^\dag (D_B \otimes I_A)).$9, and MDER $D_j=I$0 under a unified preprocessing and training protocol; its ablation also showed that the hybrid model outperformed both the classical-only and quantum-only variants on all three datasets (Mohtashim et al., 14 May 2026). In aeroelasticity, the reported contribution was an end-to-end trainable TN–quantum pipeline for binary classification and univariate or multivariate regression, with $D_j=I$1 used for classification and $D_j=I$2 for regression during hyperparameter selection (Hickmann et al., 7 Aug 2025).

6. Limitations, bottlenecks, and open directions

Several bottlenecks recur across the literature. Noise and measurement overhead are central. In density-matrix HTNs for noisy devices, the expectation value of a measured observable was shown to exponentially vanish with the number of contracted quantum tensors, making noise propagation an architectural issue rather than a mere implementation detail (Harada et al., 2023). In the hybrid MERA setting, unbiased optimization with classical shadows requires fresh and independent shadow samples for gradient estimation and energy evaluation, and the empirical shot requirement for the $D_j=I$3, $D_j=I$4 case was estimated as $D_j=I$5 (Schuhmacher et al., 20 May 2026). In the 2D isoTNS algorithm, the total shot count per local step scales as $D_j=I$6 for full tomography and as $D_j=I$7 for the Lanczos-style alternative, so quantum-assisted contraction does not eliminate bond-dimension dependence (Leontica et al., 17 Nov 2025).

A second bottleneck is the classical–quantum interface itself. The formal HTN framework makes this explicit by locating the transition between classical and quantum tensor networks in the reduction operators $D_j=I$8: partial trace preserves CPTP structure, whereas post-selection is non-trace-preserving and effectively nonlinear after renormalization (Jäger et al., 4 May 2026). This observation suggests that “hybridity” is not only a matter of where the model is executed, but also of how reduction, normalization, and admissibility constraints are imposed. A plausible implication is that bond dimension alone is insufficient to characterize expressivity across the classical–hybrid–quantum continuum.

A third bottleneck is trainability and partition management. The BUG-based time-evolution algorithm explicitly leaves automated repartitioning criteria, shot complexity, noise analysis of interface measurements, and adaptive rank growth of the quantum tensor to future work (Bauer et al., 26 Jun 2026). The aeroelastic study states that hyperparameter selection remains a challenge and requires careful optimisation (Hickmann et al., 7 Aug 2025). The joint TN-VQE study is limited to a 4-site transverse-field Ising model and remains inconclusive on whether the observed improvements amount to genuine barren-plateau mitigation (Cores et al., 2024). The hTTN ground-state study identifies the VQE optimization of the quantum tensor as the main practical roadblock and reports much stronger tolerance when noise affects only the hybrid contractions than when it also affects the variational quantum optimization itself (Schuhmacher et al., 2024).

Taken together, these results define hybrid quantum tensor networks less as a single algorithm than as a research program for allocating tensor degrees of freedom across classical and quantum resources. The literature already contains fixed-boundary constructions, end-to-end differentiable pipelines, DMRG-like local solvers, circuit-knitting h-TNs, stabilizer–tensor hybrids, and parallel time integrators. The common objective is consistent: retain tensor-network structure where contraction and gauge control are advantageous, and use quantum resources where classical storage, bond growth, or executable circuit structure becomes the dominant obstacle.