Hybrid Stabilizer-Tensor Simulation

• Hybrid stabilizer–tensor network simulation is a collection of methods that merges efficient stabilizer (Clifford) techniques with tensor network compression to manage non-Clifford operations.
• It leverages diverse architectures—such as Clifford tensor networks, stabilizer tensor networks, and MPO/MPS variants—to decisively offload Clifford dynamics and focus computational resources on residual non-stabilizerness.
• Empirical scaling studies and noise-model analyses demonstrate its capability to simulate large qubit systems and support quantum error correction by optimizing resource allocation and contraction schemes.

Hybrid stabilizer–tensor network simulation denotes a family of classical simulation methods that combine the polynomial-time tractability of stabilizer and Clifford structure with tensor-network compression of the remaining non-stabilizer content. Across the literature, the hybrid boundary is realized in several closely related forms: stabilizer-channel tableaux contracted as binary tensor networks over $GF(2)$; Clifford tensor networks of the form $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$; stabilizer tensor networks in which a tableau-defined basis carries Clifford entanglement while an MPS stores the expansion coefficients; and operator-space variants that move Clifford dynamics to observables or vectorized operators (Yashin, 18 Apr 2025, Masot-Llima et al., 17 Feb 2026, Masot-Llima et al., 2024, Mello et al., 2024). The unifying principle is to offload the part of the computation governed by the Gottesman–Knill theorem into a stabilizer representation, while reserving tensor-network resources for non-Clifford gates, magic, coherent noise, decoding likelihoods, or residual correlations.

1. Conceptual scope and representational forms

The topic encompasses several representational choices rather than a single simulator design. In the stabilizer-channel viewpoint, circuits are rewritten as compositions of Clifford channels represented by modified stabilizer tableaux, and contraction reduces to linear-algebraic matching of tableau row spaces (Yashin, 18 Apr 2025). In Clifford tensor networks, the state is written as $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$, where the Clifford frame $\mathcal{C}$ stores a large portion of the entanglement and the tensor network $\mathcal{T}$ stores the residual non-Clifford structure (Masot-Llima et al., 17 Feb 2026). In stabilizer tensor networks, a state is expanded in an orthonormal basis $\{d^{\hat{\imath}}\ket{\psi_S}\}$ generated from a stabilizer state and its destabilizers, while the coefficient vector is encoded as a tensor network, typically an MPS (Masot-Llima et al., 2024). In Clifford-augmented MPO and MPS approaches, Clifford layers are pushed into a frame or into observables, leaving only Clifford-conjugated local rotations or imaginary-time updates to be handled by the tensor network (Mello et al., 2024, Qian et al., 2024).

These variants differ in what is treated as the “hard” object. Some methods treat non-Clifford gates as tensor-network perturbations on top of a stabilizer backbone; others treat measurements, projections, or likelihood factors as the tensor-network layer; still others encode the entire circuit as a graph whose Clifford parts become binary constraints and whose non-Clifford parts remain complex tensors (Yashin, 18 Apr 2025, Harper et al., 28 May 2026). A plausible implication is that “hybrid stabilizer–tensor network simulation” is best understood as an umbrella methodology defined by an interface, not by a fixed data structure.

A central operational distinction is whether the stabilizer component carries entanglement, operator algebra, or decoding structure. In CTN and GCAMPS-style simulators, the stabilizer frame absorbs entanglement generated by Clifford dynamics, leaving the tensor network to encode non-stabilizerness or coherent error wavefunctions (Masot-Llima et al., 17 Feb 2026, Harper et al., 28 May 2026). In tensor-network stabilizer codes, the tensor network itself is an indicator representation of stabilizer cosets and is contracted to perform maximum-likelihood decoding or code analysis (Farrelly et al., 2020, Farrelly et al., 2021). In the operator-space formulations, the vectorized operator $\mathcal V[O]$ or a Pauli observable is the hybrid object, and the relevant complexity measure may be time entanglement rather than only state entanglement (Shang et al., 14 May 2025).

2. Tableau, channel, and Boolean-linear-algebra foundations

A particularly explicit formulation is the stabilizer-channel tableau formalism. An $n$-qubit Pauli observable is encoded by a $(2n+1)$-bit string $(u|c)$, where $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$0 and $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$1 encodes the sign. Commutation is controlled by the symplectic form $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$2, and a stabilizer tableau stores generators in a Boolean matrix $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$3 with sign-free block obeying $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$4, where

$\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$5

(Yashin, 18 Apr 2025).

The key generalization is from stabilizer states to arbitrary Clifford channels. A channel $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$6 is encoded by rows of a modified tableau $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$7 representing Pauli superoperators

$\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$8

Tensor product corresponds to a direct sum of tableaux, while channel composition reduces to matching the shared-side labels $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$9. Algorithmically, composition picks the intersection of the two $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$0-side row spaces and forms the composed rows with sign $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$1 (Yashin, 18 Apr 2025). This is the algebraic core of the hybrid picture: wire contraction in the circuit diagram is literally row-space intersection plus sign bookkeeping.

For non-adaptive strong simulation, the fully contracted output tableau has the form $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$2 with $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$3-columns zero, and the output distribution is determined by a linear system over $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$4:

$\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$5

This reproduces the known reduction of strong stabilizer simulation to solving linear systems over $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$6, and the contraction procedure becomes a sequence of Gaussian eliminations (Yashin, 18 Apr 2025).

The diagrammatic interpretation makes the connection to tensor networks explicit. Each channel-tableau can be viewed as a binary tensor or factor enforcing affine constraints over $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$7), and an internal edge imposes equality of Pauli labels on the shared subsystem. Contracting the diagram therefore corresponds to eliminating internal variables by Gaussian elimination, with graph structure, sparsity, and treewidth governing practical cost (Yashin, 18 Apr 2025). This binary viewpoint is also the natural interface used when Clifford subgraphs are collapsed into boundary indicator tensors $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$8 before being merged with neighboring complex-valued tensors in near-Clifford simulations.

3. Hybrid architectures for near-Clifford circuits and dynamics

The most direct state-space architecture is the Clifford tensor network. Here the state is represented as

$\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$9

with $\mathcal{C}$0 a Clifford unitary and $\mathcal{C}$1 a tensor-network ansatz such as MPS, PEPS, or MERA (Masot-Llima et al., 17 Feb 2026). Clifford gates are absorbed into $\mathcal{C}$2, while non-Clifford resources are stored in $\mathcal{C}$3. The central optimization strategy is “entanglement cooling”: exact disentangling when $\mathcal{C}$4 contains a separable stabilizer site aligned with a global Pauli rotation, and heuristic $\mathcal{C}$5-local greedy sweeps over Clifford gates that minimize entanglement across local bonds. The exact protocol decomposes a global Pauli rotation into a local rotation on one stabilizer site together with controlled-Pauli Clifford gates absorbed into the frame $\mathcal{C}$6 (Masot-Llima et al., 17 Feb 2026).

A second architecture is the stabilizer tensor network. A state is expanded as

$\mathcal{C}$7

where $\mathcal{C}$8 is a stabilizer state, $\mathcal{C}$9 is a product of destabilizers, and the coefficient vector $\mathcal{T}$0 is stored as a tensor network, typically an MPS (Masot-Llima et al., 2024). Clifford gates update only the tableau and do not alter the coefficient TN. Non-Clifford gates are rewritten as structured multi-qubit rotations acting on the coefficient network. Measurements are handled by expectation-value evaluation on the coefficient TN followed by a non-unitary projection and a tableau collapse. This framework therefore shifts Clifford-generated entanglement into the choice of basis rather than into the tensor-network amplitudes.

A third architecture is the hybrid stabilizer MPO. Predominantly Clifford circuits with sparse single-qubit non-Clifford rotations are rewritten as a product of Clifford-conjugated local rotations followed by a global Clifford unitary. Each Clifford-conjugated rotation becomes a Pauli-controlled MPO of bond dimension $\mathcal{T}$1, while the global Clifford is pushed to the observable side via stabilizer conjugation (Mello et al., 2024). This construction is tailored to expectation values of Pauli observables and to regimes in which Clifford dynamics would otherwise dominate entanglement growth in a direct state evolution.

A fourth architecture uses magic state injection to alter where non-Clifford complexity appears. In the magic-state-injected STN framework, each $\mathcal{T}$2 gate is replaced by a magic ancilla $\mathcal{T}$3 and a Clifford-only injection gadget, so that the full intermediate evolution is Clifford and the expensive step becomes a deferred final projection sweep (Nakhl et al., 2024). The same philosophy appears in coherent-noise simulators such as GCAMPS, which represent the evolving state as

$\mathcal{T}$4

with $\mathcal{T}$5 a global Clifford operator and the MPS encoding the non-Clifford “error wavefunction” generated by coherent $\mathcal{T}$6 crosstalk or other non-Clifford updates (Harper et al., 28 May 2026).

4. Algorithmic regimes and empirical scaling

The CTN entanglement-cooling literature identifies three regimes for random 1D Clifford+$\mathcal{T}$7 circuits as a function of the number $\mathcal{T}$8 of injected $\mathcal{T}$9 gates relative to system size $\{d^{\hat{\imath}}\ket{\psi_S}\}$0. For $\{d^{\hat{\imath}}\ket{\psi_S}\}$1, cooling is highly effective and often eliminates entanglement. For $\{d^{\hat{\imath}}\ket{\psi_S}\}$2, entanglement grows comparably to uncooled dynamics. For $\{d^{\hat{\imath}}\ket{\psi_S}\}$3, entropy saturates below the Page bound, and $\{d^{\hat{\imath}}\ket{\psi_S}\}$4-local cooling fails to produce significant improvements (Masot-Llima et al., 17 Feb 2026). The same work reports that increasing the local search radius to $\{d^{\hat{\imath}}\ket{\psi_S}\}$5 yields no improvement over $\{d^{\hat{\imath}}\ket{\psi_S}\}$6 in the tested regimes, and that increasing sweep depth beyond approximately $\{d^{\hat{\imath}}\ket{\psi_S}\}$7–$\{d^{\hat{\imath}}\ket{\psi_S}\}$8 shows no systematic gain (Masot-Llima et al., 17 Feb 2026). These results delimit a stabilizer-dominant regime in which frame updates are genuinely beneficial.

The magic-state-injected STN protocol yields a different scaling regime. For random $\{d^{\hat{\imath}}\ket{\psi_S}\}$9-doped $\mathcal V[O]$0-qubit Clifford circuits, the computational cost scales as $\mathcal V[O]$1 when the circuit has $\mathcal V[O]$2 $\mathcal V[O]$3-gates, while plain STN exhibits exponential scaling. In systems of up to $\mathcal V[O]$4 qubits, the average MPS bond dimension remains bounded by $\mathcal V[O]$5 in this regime; a crossover appears for $\mathcal V[O]$6; and for $\mathcal V[O]$7 both MAST and STN saturate at $\mathcal V[O]$8 (Nakhl et al., 2024). On the Hidden Bit Shift circuit, the same framework efficiently simulates $\mathcal V[O]$9 qubits and $n$0 $n$1-gates, and for $n$2 it reaches $n$3 $n$4 gates (Nakhl et al., 2024).

In the hybrid stabilizer MPO setting, the MPO bond dimension contributed by each Clifford-conjugated local rotation is fixed at $n$5, and the method was demonstrated on random Clifford $n$6-doped circuits with $n$7 and $n$8 layers, as well as on random $n$9-symmetric Clifford Floquet dynamics with analytical magnetization decay $(2n+1)$0 (Mello et al., 2024). In the reported Floquet simulations, bond dimensions $(2n+1)$1 and $(2n+1)$2 reproduce the analytical magnetization for small $(2n+1)$3 (Mello et al., 2024). In finite-temperature simulations, Clifford augmentation within purification-based TDVP achieves the accuracy of plain TDVP at roughly one-third the bond dimension in the $(2n+1)$4 $(2n+1)$5–$(2n+1)$6 Heisenberg tests, with CA-TDVP with $(2n+1)$7 matching TDVP with $(2n+1)$8 at the $(2n+1)$9 yielding the largest error (Qian et al., 2024).

These results do not point to a single universal scaling law. Instead they identify multiple favorable regimes: sparse magic on top of large Clifford structure; low-to-intermediate non-Clifford density; operator-space or observable-space tasks where Clifford conjugation can be moved away from the state; and settings where projection cost is lower than inline non-Clifford evolution. This suggests that performance hinges less on raw circuit size than on how the non-Clifford content is exposed to the tensor network.

5. Error correction, decoding, and noisy-circuit simulation

Hybrid stabilizer–tensor network methods are also central in quantum error-correction simulation. Tensor-network stabilizer codes encode stabilizer cosets by indicator tensors

$(u|c)$0

and use tensor contraction to evaluate logical-coset likelihoods $(u|c)$1 for maximum-likelihood decoding (Farrelly et al., 2020, Farrelly et al., 2021). For the holographic code family considered in the exact decoder work, the depolarizing threshold is reported as $(u|c)$2 with finite-size scaling exponent $(u|c)$3, and the exact tensor-network decoder is polynomial in the number of physical qubits even with locally correlated noise represented by a boundary MPS of bond dimension $(u|c)$4 (Farrelly et al., 2020). Local tensor-network codes further show that code distance and full logical-coset weight histograms can be obtained by contracting a tensor network, and that injecting a non-CSS five-qubit code tensor into the rotated surface-code TN reduces the number of minimum-weight logical operators from $(u|c)$5 to $(u|c)$6 at $(u|c)$7, with up to $(u|c)$8 higher success probability under depolarizing noise in the perfect-measurement setting (Farrelly et al., 2021).

In circuit-level surface-code simulation with coherent crosstalk, the hybrid STN/CAMPS simulator GCAMPS represents the evolving state as $(u|c)$9, where $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$00 tracks the ideal circuit and all Clifford noise, and the MPS carries the non-Clifford coherent error wavefunction (Harper et al., 28 May 2026). The noise model includes coherent nearest-neighbor $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$01 rotations $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$02 inserted after each entangling gate during syndrome extraction, with realistic parameters $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$03 and $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$04 giving $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$05 (Harper et al., 28 May 2026). Using $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$06 and $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$07 shots per data point, the study reports a baseline depolarizing threshold of approximately $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$08, a threshold of approximately $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$09 with crosstalk PTA, and coherent crosstalk that increases the sub-threshold logical error metric

$\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$10

relative to the Pauli-twirled approximation (Harper et al., 28 May 2026). Fixed-sign coherent crosstalk produces constructive interference and higher $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$11, whereas random-sign coherent crosstalk yields $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$12 values close to PTA below threshold, despite sharing the same twirled channel (Harper et al., 28 May 2026). This directly illustrates why a hybrid representation is needed: pure stabilizer simulation cannot retain the coherent phase information, while pure TN evolution is burdened by the large Clifford circuit.

A related but distinct line uses hybrid TTNs with quantum tensors for variational simulation of models with stabilizer structure. In the toric-code Hamiltonian, where all local terms are Pauli stabilizers, hTTN achieved the exact $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$13 ground-state energy $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$14 and outperformed both a $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$15 classical TTN and a $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$16-qubit full VQE, while on the $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$17 toric code the hTTN with $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$18 outperformed both $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$19 and $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$20 TTN baselines (Schuhmacher et al., 2024). Although this setting is variational rather than purely classical, it exemplifies the same organizing principle: stabilizer-compatible structure simplifies the quantum-classical interface, while the tensor component repairs the limitations of a low-bond classical ansatz.

6. Structural limits, complexity measures, and open directions

A recurrent misconception is that Clifford-based augmentation can universally remove hard entanglement. The CTN no-go theorem rules this out: beyond stabilizer settings, there is no Clifford operation that can universally disentangle even a single qubit from an arbitrary non-Clifford rotation (Masot-Llima et al., 17 Feb 2026). The constructive exact disentangler exists only when the tensor network retains a separable stabilizer site on which the relevant Pauli acts nontrivially, and heuristic cooling loses effectiveness as separable stabilizer sites disappear and entropy approaches Haar-like values (Masot-Llima et al., 17 Feb 2026). Likewise, the binary stabilizer-channel formalism captures stabilizer operations but does not directly encode amplitudes of non-Clifford gates; stabilizer-rank and quasiprobability methods require additional machinery beyond binary tableaux (Yashin, 18 Apr 2025).

A second structural theme is that graph complexity and contraction order are as decisive as gate count. The Boolean-linear-algebraic tableau formalism already notes that sparsity, locality, pivoting, and treewidth determine practical performance (Yashin, 18 Apr 2025). This perspective is developed into an exact strong-simulation framework unifying stabilizer decompositions and tensor-network contraction, with runtimes $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$21 and $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$22 for $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$23, together with refined notions of focused tree-width and focused rank-width that are never worse than their standard counterparts (Codsi et al., 6 Mar 2026). The emphasis here is not on tensor truncation but on separator structure, ZX simplification, and linear-memory branching over non-Clifford vertices.

A third direction is resource-theoretic unification. In the “cutting in time” framework, the central quantity is the time entanglement of a vectorized operator $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$24, which determines whether tensor-network complexity is governed primarily by coherence or by magic (Shang et al., 14 May 2025). Low time entanglement favors TN propagation because space entanglement growth is coherence-limited, whereas high time entanglement aligns tensor-network hardness with the magic dependence already present in operator stabilizer simulation (Shang et al., 14 May 2025). This suggests a more adaptive hybridization strategy in which one switches between tensor-network propagation and stabilizer-based operator methods according to a diagnostic on $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$25, rather than according only to the gate set of $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$26.

The open problems therefore cluster around interfaces. The literature repeatedly identifies optimal contraction order, projection order, and disentangler search as unresolved algorithmic questions (Yashin, 18 Apr 2025, Nakhl et al., 2024, Masot-Llima et al., 17 Feb 2026). Extending these methods beyond qubits, beyond Pauli noise, or into higher-dimensional PEPS-like settings remains active territory (Yashin, 18 Apr 2025, Qian et al., 2024). Another plausible implication is that future progress may depend less on inventing an entirely new hybrid formalism than on improving interface selection: when to collapse a Clifford subgraph into $\ket{\psi}=\mathcal{C}\ket{\psi_{\mathcal T}}$27 constraints, when to use a Clifford frame, when to inject magic states, when to switch from state-space TNs to operator-space stabilizer methods, and how to make those choices geometry-aware.