Photonic Fusion Networks

Updated 9 July 2026

Photonic fusion networks are architectures that create scalable computational resources by joining small entangled photonic states using probabilistic fusion gates.
Boosted fusion gates and temporal multiplexing techniques reduce photon overhead and synchronize resource state generation, enhancing network reliability.
Hybrid designs integrating quantum emitters with photonic circuits achieve near-deterministic entangled state generation with robust fault-tolerance thresholds.

Photonic fusion networks are architectures in which small entangled photonic resource states are generated and then joined by probabilistic entangling measurements, or fusion gates, to create larger states for measurement-based quantum computing, quantum communication, and distributed quantum information processing. In fusion-based quantum computation, the central primitives are near-deterministic generation of constant-size entangled states and probabilistic entangling measurements between them; the resulting network is naturally described as resource states connected by fusion operations, with photonic delays, switches, and feedforward providing synchronization and control (Chan et al., 2024, Meng et al., 2023, Bombin et al., 2021).

1. Foundational model and network representation

Fusion-based photonic architectures replace the direct preparation of a monolithic cluster state with a compositional procedure: small photonic graph states are prepared locally and are then fused into a larger computational resource. In the FFCC-based construction, the cluster state is decomposed through edge and node splits into a sequence of small entangled resource states, and all fusions on a given “layer” can be performed in parallel. Resource states are local, interact with only a few neighbors, and may be linear cluster states, possibly with repetition code encoding (Chan et al., 2024).

At the level of network description, a fusion graph treats resource states as vertices, pairwise fusions as edges, and single-qubit measurements as half-edges. In a 6-ring architecture, this yields a 3D cubic lattice in which each resource state is fused with six others along the axes. This formulation is central to modular surface-code constructions and lattice-surgery protocols, because fusion outcomes supply the syndrome information used for error correction (Bombin et al., 2021).

A more formal description models a fusion network as an open graph decorated by a multiset of fusion operations, each specified by the nodes it fuses and the measurement plane. In this setting, X- and Y-type fusions are the principal stabilizer cases, and the all-success branch defines a target open graph. This description is designed to support deterministic correction of measurement randomness through graph-theoretic flow conditions and feedforward (Felice et al., 2024).

Operationally, photonic fusion gates are linear-optical Bell-state analyzers acting on dual-rail qubits. In one standard formulation, successful patterns correspond to projections onto

$\ket{\psi^\pm}=\frac{1}{\sqrt{2}}(\ket{01}\pm\ket{10}),$

whereas outcomes associated with

$\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$

are failures. This 50% success limit for passive linear optics is a recurring design constraint throughout the literature (Meng et al., 2023).

2. Fusion operations, success probabilities, and gate generalizations

For dual-rail photonic qubits, standard type-I and type-II fusion gates are limited to success probability $1/2$ when no ancillary resources are used. A recent development is a boosted type-I fusion architecture that reaches the linear-optical upper bound of $3/4$ using only four ancillary single photons and passive linear optics, with no entangled ancillary states. The type-I map is

$|0\rangle\langle 00|+|1\rangle\langle 11|.$

In that scheme, direct fusion succeeds with probability $5/8$, a single distillation round raises the total to $11/16$, and full recursive distillation raises it to $3/4$. The practical motivation is explicit: earlier $3/4$-efficient type-I schemes relied on ancillary Bell-pair states whose preparation is itself probabilistic and resource-intensive (Melkozerov et al., 28 Mar 2026).

The same work quantifies the resource impact of higher-probability fusion. For representative large entangled states, the proposed gate reduces the required photonic overhead relative to standard type-I fusion. The data block reports, for example, 4-qubit GHZ generation requiring 320–448 photons with standard type-I versus 169–304 photons with Bell-boosted or proposed $3/4$-efficient type-I fusion, and a 6-ring graph requiring 2560–3840 versus 613–1273 photons, respectively (Melkozerov et al., 28 Mar 2026).

Fusion has also been generalized beyond qubits. For two $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 0-rail single-photon qudits, passive pairwise fusion succeeds with probability $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 1, with failures confined to the diagonal logical subspace. Identical single-mode squeezers applied to the $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 2 interferometer outputs before photon-number-resolving detection recover part of this structured failure sector. The exact logical-space POVM admits a simple acceptance rule: a diagonal pattern is accepted if and only if its photon-number-imbalance vector has exactly two nonzero components of equal magnitude. The resulting success probability increases from $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 3 to $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 4 for $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 5, and from $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 6 to $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 7 for $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 8; with detector saturation threshold $\ket{\phi^\pm}=\frac{1}{\sqrt{2}}(\ket{00}\pm\ket{11})$ 9, the certified values remain $1/2$0 and $1/2$1 (Laha et al., 28 Jun 2026).

Beyond graph-state fusion, photonic W-state networks admit a distinct fusion primitive. A nonunitary partial-swap gate fuses arbitrary small photonic W-states into larger W-states without ancillary photons. Its success probability for inputs $1/2$2 and $1/2$3 is

$1/2$4

and the nonsuccessful branch yields $1/2$5, so all “garbage” states are recyclable and there is no complete failure output in principle (Wei et al., 2020).

3. Resource-state generation, temporal multiplexing, and photonic routing

Fusion networks depend not only on the fusion measurement itself but also on the way small resource states are produced and synchronized. A key experimental direction uses solid-state spin-photon interfaces to deterministically generate entangled resource states. In one demonstration, repetitive operation of a single quantum emitter produces two entangled resource states sequentially, and a delay line causes photons emitted at different times to interfere in a common fusion gate. The resulting protocol creates entanglement between the quantum states of the same spin at two different instances in time. This temporal multiplexing provides a resource-efficient route to scaling many-body photonic entanglement with a single emitter rather than an array of independent sources (Meng et al., 2023).

Temporal multiplexing directly shapes the architecture of photonic fusion networks. Fiber delays and active switching align photons emitted at different times into simultaneous fusion inputs; more generally, switching hardware is required both for feedforward at the logical level and for multiplexing stages that allocate heralded photons to resource-state generation and fusion circuits. Because each switch stage adds loss and error, low switch depth is a primary design objective (Bartolucci et al., 2021).

A substantial body of work therefore studies switch-network topology as part of the fusion-network problem. The reported designs include spatial log-tree and chain muxes, generalized Mach-Zehnder interferometers, binary delay networks, storage loops, de Bruijn networks, rastering schemes, and shared multiplexers for parallel resource-state generation. The same source reports that sharing strategies for multiple resource-generation circuits achieve 2.5–5× yield improvements over independent muxes with the same resources, while fully ballistic approaches require 3× more sources to reach the same probability as optimal muxing (Bartolucci et al., 2021).

Long fixed-time delays can also be promoted from a synchronization tool to an architectural principle. Interleaving modules combine one resource-state generator, associated fusion devices, and a few fiber delays, and exploit the multiplicative power of delays so that each module can add thousands of physical qubits to the computational Hilbert space. In a network of modules containing 1-km-long fiber delays, each resource-state generator can generate four logical distance-35 surface-code qubits while tolerating photon loss rates above 2% in addition to the fiber-delay loss (Bombin et al., 2021).

4. Loss, erasure, percolation, and fault-tolerance thresholds

The dominant noise processes in photonic fusion networks are photon loss, imperfect indistinguishability, and memory errors in the physical system that generates the resource states. Architectures tailored to quantum emitters analyze these errors at the level of fusion-based fault tolerance and report explicit thresholds for synchronous FFCC constructions (Chan et al., 2024).

Architecture	Logical fusion	Thresholds
sFFCC REP ($1/2$6)	REP	Loss $1/2$7; distinguishability $1/2$8; spin error $1/2$9; spin-Z $3/4$0
sFFCC RUS ($3/4$1)	RUS	Loss $3/4$2; distinguishability $3/4$3; spin error $3/4$4; spin-Z $3/4$5

These values reflect several architectural choices that recur across the literature: synchronous layered decompositions, local resource states that interact with only a constant number of neighbors, repetition encoding, and repeat-until-success logical fusion. The same analysis emphasizes that present-day quantum emitters have demonstrated indistinguishabilities and spin coherence compatible with the fault-tolerance requirements highlighted (Chan et al., 2024).

Percolation theory supplies a complementary way to quantify robustness. In lossy photonic fusion networks, the percolation threshold $3/4$6 is the minimal photon survival probability above which a spanning cluster is likely to form. Several non-standard percolation models are relevant: loss on a pre-existing graph state, lossy ballistic fusion networks, hybrid emitter-photonic networks, and adaptive repeat-until-success fusion networks. An important result is that higher vertex degree can make a network less robust to loss, because a lost photon can force deletion of its neighbors through $3/4$7-basis measurements. Hybrid networks in which central qubits are quantum emitters are more robust than all-photonic networks, and simulated thresholds can vary from $3/4$8 to $3/4$9 depending on architecture and loss model. Modified Newman-Ziff methods enable simulations on lattices of $|0\rangle\langle 00|+|1\rangle\langle 11|.$ 0 nodes and are reported to be orders of magnitude faster than naïve approaches (Löbl et al., 2023).

A related distinction, emphasized in compiler-level work, is between fusion failure and fusion erasure. Fusion failure measures out affected qubits while keeping the graph state well defined; fusion erasure, caused by photon loss, leaves the graph structure unknown and is more damaging. To suppress erasure, tree-encoded fusion uses a logical qubit encoded in a tree graph with multiple branches, indirect $|0\rangle\langle 00|+|1\rangle\langle 11|.$ 1 measurement via the stabilizer rule

$|0\rangle\langle 00|+|1\rangle\langle 11|.$ 2

5. Formal methods, graph decompositions, and automated optimization

Because fusion-based photonic computation combines linear optics, probabilistic measurements, and classical feedforward, several works formalize photonic fusion networks as programmable and verifiable objects. One framework brings together linear optics, ZX calculus, and dataflow programming, characterizes fusion measurements that induce Pauli errors, and shows that these errors are correctable using a novel flow structure for fusion networks. It proves the correctness of repeat-until-success protocols for arbitrary fusions and gives a graph-theoretic proof of universality for linear optics with entangled photon sources (Felice et al., 2024).

A related extension places these ideas in a synchronous dataflow model with discrete-time dynamics and explicit interfaces for classical control and feedforward. Within this setting, entangling photonic fusion measurements are classified, Pauli-error correction is organized via a novel flow structure, and repeat-until-success protocols are incorporated into architectures built layer by layer from beam splitters, switches, and photon sources. The stated goal is verifiable compilation and automated optimization for distributed and networked photonic quantum computing (Felice et al., 13 Jan 2026).

$|0\rangle\langle 00|+|1\rangle\langle 11|.$ 9

This connects photonic compilation to trail decompositions, path covers, and trail covers under resource-state length bounds. The unbounded X-only problem, MinTrailDecomposition, is in P; MinPathCover and MinTrailCover are NP-hard; and bounded versions are NP-hard in all cases. The same study introduces heuristic algorithms, including a reduction to the travelling salesman problem, and graph rewrites that lower the average required number of fusions and photons by about 10%. On a Shor-encoded 6-cycle benchmark, the reported fusion counts are 352 for Y-fusions only, 87 for X-fusions only, and 78 for mixed XY-fusions (Cashman et al., 25 Aug 2025).

Automated circuit discovery offers another optimization route. A two-pass optimization framework based on polynomial simulation and gradient-based search first finds a unitary that prepares a target graph state with perfect fidelity and maximal success probability, then sparsifies it into a compact circuit with minimal beamsplitter count while preserving performance. For 4-qubit states, the discovered circuits achieve success probabilities from $5/8$0 to $5/8$1, outperforming the fusion baseline by up to $5/8$2; for 5-qubit states, the values are $5/8$3 to $5/8$4, with improvements up to $5/8$5. The same work reports the first known state-preparation circuits for certain 5-qubit graph states (Hartnett et al., 22 Aug 2025).

6. Experimental realizations and expanding physical scope

Experimental photonic fusion predates the current fusion-based architecture literature and already exposed many of the practical constraints that remain central. An all-fiber experiment fused photons from two independent photonic crystal fiber sources into polarization-entangled states using a fiber-based polarizing beam splitter. For input $5/8$6, the ideal fusion transformation is

$5/8$7

with ideal success probability $5/8$8. The reported fidelity with respect to $5/8$9 reached $11/16$0 at 5.3 mW, with 3.2 four-fold detections per second, while higher pump power increased the success rate to 111.6 four-folds per second with entanglement still present. Full process reconstruction gave $11/16$1, $11/16$2, $11/16$3, and process fidelity $11/16$4 (Bell et al., 2011).

A later solid-state experiment demonstrated the two core primitives of fusion-based photonic quantum computing in one platform: deterministic generation of entangled resource states from a quantum dot spin-photon interface and their fusion using a photonic linear-optical circuit. The experiment verified post-fusion quantum correlations by measuring entanglement between the spin at two times, thereby realizing temporal multiplexing with a single emitter (Meng et al., 2023).

Deterministic graph-state fusion has also been demonstrated with matter-assisted photonic generation. An optical resonator containing two individually addressable $11/16$5Rb atoms was used to generate and fuse ring and tree graph states with up to eight qubits. The fusion itself was implemented by a cavity-assisted gate between the two atoms. The reported Bell-state fidelity after fusion was up to $11/16$6 unfiltered; box-state fidelity ranged from $11/16$7 and tree-state fidelity from $11/16$8. Source-to-detector efficiency was close to 0.5 per photon, with coincidence rates up to 2/min (Thomas et al., 2024).

These experiments also delimit a persistent misconception: post-selected fusion is not automatically scalable. In the fiber-source experiment, the implemented protocol was explicitly “not scalable” because success relied on destructive post-selection. More scalable directions therefore emphasize heralded fusion, deterministic or near-deterministic resource generation, feedforward switching, quantum-emitter interfaces, and boosted or recycled failure sectors in the fusion measurement itself (Bell et al., 2011, Melkozerov et al., 28 Mar 2026, Laha et al., 28 Jun 2026).

Photonic fusion networks now encompass all-photonic and hybrid emitter-photonic architectures, qubit and qudit encodings, graph-state and W-state assembly, and both probabilistic and boosted fusion strategies. The research trajectory recorded in the cited work suggests a field organized around a common set of problems—fusion success probability, loss tolerance, resource overhead, feedforward control, and formal compilability—rather than a single device family or gate construction (Chan et al., 2024, Bartolucci et al., 2021).