Circuit Knitting in Quantum Computing

• Circuit knitting is a technique for dividing quantum circuits into manageable subcircuits to overcome qubit capacity and connectivity limitations.
• It employs methods like gate cuts and wire cuts, using quasiprobability decompositions that replace nonlocal operations with combinations of local actions and classical post-processing.
• Recent advances leverage entanglement structure and hardware-aware optimization to significantly reduce the sampling overhead, enabling practical NISQ and distributed quantum applications.

Searching arXiv for recent and foundational papers on circuit knitting to ground the article in the literature. arxiv_search(query="circuit knitting quantum", max_results=10, sort_by="submittedDate") Circuit knitting is a family of techniques for executing or simulating quantum circuits that exceed the qubit capacity or connectivity of a single device by partitioning a target circuit into smaller subcircuits, running those fragments separately, and reconstructing observables through classical post-processing. In the contemporary literature, it is studied both as a method for limited-size NISQ hardware and as a distributed-quantum-computing primitive when many QPUs must be coordinated without high-fidelity quantum interconnects. Its central promise is modularization of large workloads; its central difficulty is the sampling and reconstruction overhead introduced by cuts, which in straightforward schemes grows exponentially with the number of inter-partition operations (Johnson et al., 12 Mar 2026, Brenner et al., 2023).

1. Formal definition and primitive operations

In the standard formulation, circuit knitting is based on a quasiprobability decomposition (QPD) of a nonlocal operation into a signed linear combination of physically realizable local operations. For a unitary $\mathcal{U}$ acting across two partitions $A$ and $B$,

$\mathcal{U} = \sum_j c_j \mathcal{F}_j,$

with $c_j \in \mathbb{R}$ and $\mathcal{F}_j$ implementable on the separated partitions. The associated sampling coefficient is

$\gamma = \sum_j |c_j|,$

and observable estimation incurs an overhead scaling as $O(\gamma^2/\epsilon^2)$, or $O(\gamma^{2n}/\epsilon^2)$ when $n$ two-qubit gates are cut (Johnson et al., 12 Mar 2026).

Two primitive cut types recur throughout the literature. A gate cut removes a nonlocal gate, typically a two-qubit entangling gate, and replaces it by a QPD over local operations. A wire cut severs the time evolution of a qubit line, replacing the severed channel by measurement and state preparation at the boundary. One review-style result states that all existing rigorous circuit knitting techniques can be understood in the quasiprobability framework, including both gate cuts and wire cuts (Brenner et al., 2023).

Classical communication materially changes the cost of some wire-cutting and multi-gate protocols. For cutting $A$0 wires, the optimal overhead scales as $A$1 without classical communication and $A$2 with classical communication (Brenner et al., 2023). In a related result for circuits containing $A$3 nonlocal CNOT gates connecting two circuit parts, the simulation overhead can be reduced from $A$4 to $A$5 if one allows for classical information exchange (Piveteau et al., 2022).

Primitive or regime	Mechanism	Stated overhead
Gate cut	QPD of a nonlocal gate into local operations	$A$6 per reconstructed observable
$A$7 cut two-qubit gates	Repeated inter-partition QPD	$A$8
$A$9 wire cuts, no classical communication	Measurement and preparation without feedforward	$B$0
$B$1 wire cuts, with classical communication	Measurement and preparation with feedforward	$B$2
$B$3 nonlocal CNOTs with communication	Collective simulation across two parts	$B$4 instead of $B$5

The operational significance of these primitives is that cutting trades quantum width and connectivity for classical sampling, classical post-processing, and, in some variants, adaptive coordination between fragments.

2. Sampling overhead, entanglement, and lower bounds

A major development in the theory of circuit knitting is the identification of entanglement as the quantity controlling unavoidable overhead. Adaptive Circuit Knitting (ACK) makes this explicit for bipartitions of a state $B$6 through the robustness of entanglement: $B$7 For a pure state with Schmidt coefficients $B$8,

$B$9

the corresponding overhead is

$\mathcal{U} = \sum_j c_j \mathcal{F}_j,$0

This is related to the half-Rényi entropy by

$\mathcal{U} = \sum_j c_j \mathcal{F}_j,$1

which means that the overhead is exponential in the half-Rényi entropy across the cut (Johnson et al., 12 Mar 2026).

A more general information-theoretic treatment proves that the regularized sampling overhead of any bipartite quantum channel is lower bounded by the exponential of its exact entanglement cost under the corresponding free-operation class. For a bipartite channel $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$2 and $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$3,

$\mathcal{U} = \sum_j c_j \mathcal{F}_j,$4

The same work gives efficiently computable lower bounds in terms of $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$5-entanglement and max-Rains information, thereby providing practical infeasibility benchmarks for specific cut channels (Jing et al., 2024).

Complexity-theoretic negative results sharpen this picture for local circuit cutting schemes. In the regime where the input state and measurement are fixed, the existence of an efficient locally acting scheme that could partition even a single qubit from the rest of a circuit would imply $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$6. In the regime requiring a cutting to work for a complete basis of input states and measurements, inefficiency is unconditional. The same study also shows that any circuit cutting scheme cannot function by only applying unital channels (Marshall et al., 2023).

These results delimit a central misconception: minimizing the number of cuts is not, by itself, a guarantee of tractability. The entanglement cost of what is cut, and the operational class allowed for the reconstruction, determine the true asymptotic barrier.

3. Structure-exploiting routes to lower overhead

Recent work has concentrated on regimes in which the generic exponential barrier can be softened by exploiting structure in the state, circuit, or inter-partition coupling.

Adaptive Circuit Knitting replaces blind or load-balanced cuts by entanglement-aware partitioning. Starting from a tensor-network representation such as an MPS, ACK iteratively partitions the state, variationally optimizes a circuit unitary for each partition, computes entanglement entropy between neighboring qubits, and updates cut locations using an entanglement “heatmap.” On 1D and 2D disordered mixed-field Ising models up to 60 qubits, ACK reports up to four orders of magnitude reduction in sampling overhead compared to load-balanced cuts; for 40-qubit, 60-qubit, and 32-qubit 2D models, the reported median savings are $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$7--$\mathcal{U} = \sum_j c_j \mathcal{F}_j,$8, with hardest instances reaching hundreds to ten-thousands fold savings (Johnson et al., 12 Mar 2026).

A different controlled-approximation route is the weak-coupling approximation for distributed quantum computing. When a circuit is partitioned through a qubit weakly coupled to the rest, the full $\mathcal{U} = \sum_j c_j \mathcal{F}_j,$9 basis-string explosion of standard CutQC can be truncated to strings with at most $c_j \in \mathbb{R}$0 flips. The number of retained strings is

$c_j \in \mathbb{R}$1

and the reconstruction error obeys

$c_j \in \mathbb{R}$2

The method is demonstrated for layered circuits based on those used for QAOA (Stenger et al., 17 Jun 2026).

A third line of work shows that exponential scaling is not universal for tree-structured circuits. Using concatenated quantum tomography rather than conventional QPD, expectation estimation for a tree-depth-1 circuit with branching factor $c_j \in \mathbb{R}$3 and bond dimension at most $c_j \in \mathbb{R}$4 can be performed with

$c_j \in \mathbb{R}$5

total measurements. For complete $c_j \in \mathbb{R}$6-ary trees of larger depth, the total measurement cost scales as

$c_j \in \mathbb{R}$7

polynomial in the number of cuts $c_j \in \mathbb{R}$8. In the comparable tree-depth-1 setting, QPD-based wire-cutting methods are shown to require at least

$c_j \in \mathbb{R}$9

measurements (Harada et al., 22 Dec 2025).

Taken together, these results suggest that “exponential overhead” is best interpreted as the generic cost of structure-agnostic exact reconstruction, not as a universal property of every restricted geometry or approximation regime.

4. Algorithmic frameworks, architecture search, and systems integration

Several frameworks specialize circuit knitting to concrete algorithm classes or hardware stacks, and these papers emphasize that partition quality is determined jointly by sampling overhead, trainability, routing cost, and repeated structure.

CKVQA applies circuit knitting to variational quantum algorithms, specifically QAOA and VQE, and embeds the cutting cost directly into a quantum architecture search objective,

$\mathcal{F}_j$0

with an extended form

$\mathcal{F}_j$1

For the gate set $\mathcal{F}_j$2, the overhead is stated as $\mathcal{F}_j$3, where $\mathcal{F}_j$4 is the number of gates crossing partitions. CKVQA also introduces subcircuit-level optimization so that each parameter is updated using only the affected subcircuits. Its reported results state that sampling overheads remain orders of magnitude smaller than with naive ansatzes, and that the framework can find ansatzes with up to $\mathcal{F}_j$5--$\mathcal{F}_j$6 orders of magnitude less sampling overhead while preserving performance (Wu et al., 5 Aug 2025).

For variational quantum dynamics, an overhead-constrained PVQD method imposes an explicit constraint on cut-induced overhead during optimization: $\mathcal{F}_j$7 For a cut two-qubit rotation, the per-gate overhead is

$\mathcal{F}_j$8

The method is tested on weakly entangled spin blocks and is also used to reduce circuit depth by cutting long-ranged gates; one reported example saves 18 SWAP gates for 4 layers (Gentinetta et al., 2023).

Hardware-aware cutting shifts attention from cut count alone to the post-cut compilation problem. One framework co-optimizes the number of gate cuts and the graph edit distance between a subcircuit interaction graph and the chip topology, using

$\mathcal{F}_j$9

It reports reductions of total subcircuit depth by up to $\gamma = \sum_j |c_j|,$0 ($\gamma = \sum_j |c_j|,$1 on average) compared to the state-of-the-art approach, and relative-fidelity enhancement up to $\gamma = \sum_j |c_j|,$2 (Ren et al., 2024).

CiFold approaches the same problem from workload structure rather than routing. It identifies repeated qubit-local modules by longest consecutive common subsequence analysis, folds them into a meta-graph, and then performs meta-graph-guided partitioning. Evaluated on QFT, Adder, BV, and GHZ workloads, it reports up to $\gamma = \sum_j |c_j|,$3 reduction in quantum resource usage, with partitioning runtime usually under 10 seconds for large circuits (Kan et al., 2024).

Another concrete framework, shardQ, combines SparseCut, MPS compilation, and global knitting. In an ablation study on an IBM Marrakesh superconducting-type QPU, it reports error rates below $\gamma = \sum_j |c_j|,$4 for 6-qubit encoder circuits when using optimal cuts, and a reduction of the number of executed subcircuits and total shots by about $\gamma = \sum_j |c_j|,$5 relative to naive or arbitrary cutting (Guo et al., 7 Nov 2025).

Framework	Technical focus	Reported result
CKVQA	Knitting-aware ansatz search for QAOA and VQE	Up to $\gamma = \sum_j |c_j|,$6--$\gamma = \sum_j |c_j|,$7 orders of magnitude less sampling overhead
PVQD with overhead constraint	Variational dynamics with $\gamma = \sum_j |c_j|,$8	18 SWAP gates saved for 4 layers
Hardware-aware gate cutting	Joint optimization of cuts and routing cost	Up to $\gamma = \sum_j |c_j|,$9 depth reduction; up to $O(\gamma^2/\epsilon^2)$0 relative fidelity
CiFold	Module folding and meta-graph partitioning	Up to $O(\gamma^2/\epsilon^2)$1 reduction in quantum resource usage
shardQ	SparseCut + MPS compilation + global knitting	About $O(\gamma^2/\epsilon^2)$2 fewer executed subcircuits and shots

A plausible implication is that practical circuit knitting is becoming less a single reconstruction primitive than a workflow layer spanning partition discovery, ansatz design, compilation, routing, and execution management.

5. Entanglement-assisted, virtual, and continuous-variable generalizations

Although many circuit-knitting papers study resource-free decomposition into local operations and classical post-processing, several recent works blur the boundary between knitting and entanglement-assisted distributed quantum computing.

Virtual entanglement purification operates at the level of expectation values rather than state purification. It uses noisy Bell states as a resource, projects onto the Bell-state stabilizer space by post-processing, and is presented as a protocol that bridges entanglement-based and circuit-knitting DQC methods. According to the reported comparison, it requires fewer sampling shots than circuit knitting, can surpass the fidelity limit of conventional entanglement purification in the presence of noise in LOCC, and is robust against infidelity fluctuations in shared noisy Bell states (Yamamoto et al., 2024).

Entanglement-assisted circuit knitting formalizes this intermediate regime through resource-assisted QPD. For wire cutting assisted by a general resource state $O(\gamma^2/\epsilon^2)$3,

$O(\gamma^2/\epsilon^2)$4

For a bipartite unitary

$O(\gamma^2/\epsilon^2)$5

a constructive protocol with one shared Bell pair gives

$O(\gamma^2/\epsilon^2)$6

whereas standard resource-free cutting for KAK-like unitaries is stated as

$O(\gamma^2/\epsilon^2)$7

This protocol is notable because it reduces the overhead to the asymptotic limit of standard circuit knitting without classical communication (Hu et al., 30 Oct 2025).

The framework has also been extended beyond finite-dimensional qubit settings. One continuous-variable treatment develops a general QPD framework for simulating non-Gaussian states from available CV resources, with overhead measured by $O(\gamma^2/\epsilon^2)$8, but proves a no-go theorem for exact knitting of non-local Gaussian unitaries with separable operations: $O(\gamma^2/\epsilon^2)$9 The same paper explores approximate Fock-state simulation, GKP-state generation, and cat-state amplification (Hu et al., 9 Sep 2025).

In hybrid qubit-boson systems, unitary-transformed projective squeezing supplies a different notion of “knitting.” It uses smeared projectors and virtual quantum error detection to implement virtual entanglement and state projection with only local operations and classical post-processing. The same formalism is used both for larger-device simulation and for preparing higher-squeezing, higher-nonlinearity cubic phase states, with numerical verification that projection can suppress photon-loss errors (Anai et al., 2024).

These generalizations show that circuit knitting is no longer restricted to qubit-only gate cutting. It now includes resource-assisted, expectation-value-level, and infinite-dimensional variants, but the generalizations introduce their own impossibility theorems and resource trade-offs.

6. Scope, limitations, and historical usage

A persistent misconception is that circuit knitting is simply “divide and conquer for qubits.” The literature instead presents a narrower and more technical claim: partitioning is only useful when the cut structure aligns with weak entanglement, weak coupling, tree geometry, or hardware-friendly routing. ACK, weak-coupling truncation, and tomography-based tree knitting all exploit such restricted structure, and their performance improvements are explicitly conditioned on it (Johnson et al., 12 Mar 2026, Stenger et al., 17 Jun 2026, Harada et al., 22 Dec 2025).

The opposite misconception is that all exponential overhead statements are merely pessimistic artifacts of early methods. The entanglement-cost lower bounds, the local-cutting complexity results, and the continuous-variable no-go theorem show that strong negative statements remain in force for generic exact settings, and in some CV cases the exact overhead diverges (Jing et al., 2024, Marshall et al., 2023, Hu et al., 9 Sep 2025).

Historically, the word “knitting” also appeared in an earlier, distinct sense in “Knitting distributed cluster state ladders with spin chains,” where engineered perfect-state-transfer spin chains were used to knit distributed cluster-state structures by repeated injection and extraction of spin qubits at the ends of the chain. That protocol relies on natural spin-chain dynamics rather than QPD-based circuit cutting, but it anticipates later modular and distributed uses of the term (Ronke et al., 2011).

Across these strands, circuit knitting has evolved from a narrow circuit-cutting primitive into a broader distributed-computation methodology. Its present meaning centers on replacing unavailable nonlocal quantum operations by smaller fragments plus classical reconstruction, while its modern research frontier concerns identifying exactly which structural, entanglement, and resource assumptions make that replacement practical rather than merely formal.