Photonic Qubit Fusion in Quantum Architectures

• Photonic qubit fusion is a measurement-induced process that joins photonic qubits into larger resource states using linear optical techniques.
• It encompasses protocols like cluster-state fusion, path-qubit fusion, and quantum state fusion, each with distinct experimental methods and success probabilities.
• Advancements such as boosted and generalized fusion improve success rates and loss tolerance, with significant implications for scalable and fault-tolerant quantum architectures.

Searching arXiv for recent and foundational papers on photonic qubit fusion, boosted fusion, and FBQC. Photonic qubit fusion denotes a family of operations that join photonic quantum resources by measurement-induced entanglement, usually within linear optics, and it occupies a central place in measurement-based and fusion-based photonic quantum computing. In the fusion-based usage, computation is built from small pre-entangled resource states and fusions, with gates implemented by choosing which photons to fuse and in which basis, while the overall computation is determined by the pattern of fusion outcomes (Hauser et al., 2024). The term is not uniform across the literature: it can refer to Browne–Rudolph cluster-state fusion, path-qubit fusion in graph-state growth, generalized Bell-type fusion measurements, or the distinct process of combining two qubits from two photons into a single higher-dimensional photon (Weinstein, 2011).

1. Terminological scope and conceptual distinctions

A persistent source of confusion is that “photonic qubit fusion” does not name a single protocol. In cluster- and graph-state construction, it usually means a probabilistic entangling measurement that joins smaller photonic resource states into a larger graph. In a different usage, “quantum state fusion” means relocating two qubits carried by two photons into the four-dimensional internal state space of one photon, without changing the abstract quantum information content (Vitelli et al., 2012).

Usage	Representative operation	Characteristic feature
Cluster-state fusion	Type I or Type II fusion	Joins separate cluster fragments
Path-qubit fusion	$$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}$$	Merges two path qubits into one shared path qubit
Quantum state fusion	$\Psi_{12}\rightarrow \Psi_3$	Encodes two qubits into one 4D photon

In the cluster-state setting analyzed in “Fusing Imperfect Photonic Cluster States,” the fusion primitive is Browne and Rudolph’s Type I fusion, which succeeds with probability $1/2$. If it succeeds on edge qubits of two chains of lengths $m$ and $n$, the result is a longer chain of length $q=m+n-1$; if it fails, each cluster loses one edge qubit, leaving shorter fragments that may be recycled (Weinstein, 2011). In the graph-state scaling experiment based on path encoding, fusion is instead a path-qubit analog of the type-I fusion gate, implemented by superposing modes at a $50{:}50$ NPBS and postselecting the cases in which the photons exit in separate detectors, thereby realizing the effective projection

$$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$

This operation merges two input path qubits into a single fused path qubit (Lee et al., 2011).

The non-graph-state usage is more radical. “Quantum state fusion in photons” defines a process in which two input qubits,

$$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$

are mapped to a single-photon four-dimensional state,

$$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$

Here fusion changes the physical carrier rather than the entangling structure of a graph state (Vitelli et al., 2012). This distinction is essential: graph-state fusion is a measurement-induced network operation, whereas quantum state fusion is an encoding transformation.

2. Standard linear-optical fusion primitives and their limits

In FBQC, the standard physical realization of a fusion is a Bell-state measurement on two photonic qubits. The incoming photons are projected onto the Bell basis

$\Psi_{12}\rightarrow \Psi_3$0

$\Psi_{12}\rightarrow \Psi_3$1

The attraction of this construction is that it requires only beam splitters, polarizing beam splitters, and detectors; its limitation is equally fundamental. With passive linear optics and no ancillary entanglement, only two of the four Bell states can be unambiguously distinguished, so

$\Psi_{12}\rightarrow \Psi_3$2

In the usual scheme, $\Psi_{12}\rightarrow \Psi_3$3 and $\Psi_{12}\rightarrow \Psi_3$4 give distinct coincidence patterns, while $\Psi_{12}\rightarrow \Psi_3$5 and $\Psi_{12}\rightarrow \Psi_3$6 remain indistinguishable (Hauser et al., 2024).

This 50% ceiling appears in several strands of the literature. Standard type-II fusion in graph-state generation is correspondingly bounded by $\Psi_{12}\rightarrow \Psi_3$7 success probability, which is insufficient for some scalable percolative architectures (Guo et al., 2024). In the dual-rail type-II fusion reviewed in the redundantly encoded resource-state literature, bunching into the same detector corresponds to failure, whereas one photon in each detector pair projects the input into an entangled output; in the ideal case this again succeeds with probability $\Psi_{12}\rightarrow \Psi_3$8 (Sheldon et al., 2 Dec 2025).

The consequences are architectural rather than merely local. In FBQC, fusion failures increase the effective erasure rate in the network, raise the logical error rate, and reduce the tolerable photon-loss level. This is why the quality of the fusion primitive strongly impacts scalability and loss tolerance (Hauser et al., 2024). A common misconception is that fusion is “just” a Bell analyzer embedded inside a larger system. In practice, the success statistics of the analyzer determine whether the graph percolates, whether encoded redundancy is consumed too quickly, and whether arbitrarily low logical error rates remain available under scaling.

3. Resource-state growth and experimental realizations

One major line of work treats fusion as the mechanism for enlarging graph states. In the seven-qubit graph-state experiment, two separate two-photon four-qubit graph states are fused by merging one path qubit from each into a common path qubit, producing a four-photon seven-qubit graph state composed of $\Psi_{12}\rightarrow \Psi_3$9 polarization qubits $1/2$0 and $1/2$1 path qubits $1/2$2 (Lee et al., 2011). Genuine seven-qubit entanglement is certified by the witness

$1/2$3

for which the measured value is $1/2$4, with a lower bound on fidelity $1/2$5. Six qubits from the fused graph state were then used to execute the general two-qubit Deutsch–Jozsa algorithm with success probability $1/2$6 (Lee et al., 2011).

A second experimental line focuses on the minimal seed states needed for FBQC. “Heralded three-photon entanglement from a single-photon source on a photonic chip” identifies the heralded dual-rail encoded $1/2$7-GHZ state as the minimal initial entangled state for fusion-based architectures. The target state,

$1/2$8

is obtained on a programmable $1/2$9-mode silicon nitride chip using six indistinguishable telecom-wavelength photons. The reported population, coherence, and fidelity are $m$0, $m$1, and $m$2, respectively, exceeding the $m$3 entanglement threshold (Chen et al., 2023). This suggests that fusion-based architectures are not only a matter of fusion gates, but also of generating the correct seed states with heralded signatures.

A third line replaces probabilistic photon-pair sources with deterministic emitters. In “Fusion of deterministically generated photonic graph states,” two individually addressable $m$4Rb atoms in a high-finesse optical cavity emit graph-state “arms” that are fused by a cavity-assisted gate between the atoms. The protocol yields ring and tree graph states with up to eight qubits, including box, pentagon, hexagon, and depth-two tree states (Thomas et al., 2024). The Bell-pair fusion fidelity is reported as $m$5, and the tree-state fidelity interval is

$m$6

Here fusion is still heralded, but state generation is deterministic rather than SPDC-based.

Not all fusion targets are graph states. A distinct non-linear-optical protocol uses weak cross-Kerr nonlinearities to fuse two polarization-entangled W states,

$m$7

with overall success probability

$m$8

The protocol is qubit-loss-free in the sense that there is no complete failure output and the garbage states are recyclable (Wang et al., 2017). This broadens the meaning of photonic fusion beyond cluster-state assembly.

4. Boosted and generalized fusion measurements

Because standard passive linear optics saturates at $m$9, a major research direction is ancilla-assisted boosting. The Grice-style boosted Bell-state measurement adds an ancillary entangled photon pair,

$n$0

and a $n$1 multiport interferometer implementing

$n$2

In the ideal scheme, $n$3 remain identifiable, $n$4 become distinguishable in $n$5 of cases, and the total Bell-state discrimination success probability rises to $n$6 (Hauser et al., 2024). An experiment using two Sagnac photon-pair sources, a fibre-based $n$7 multiport beam splitter, and $n$8 SNSPDs achieved

$n$9

with correct-identification probabilities $q=m+n-1$0, $q=m+n-1$1, $q=m+n-1$2, and $q=m+n-1$3 for $q=m+n-1$4, $q=m+n-1$5, $q=m+n-1$6, and $q=m+n-1$7, respectively (Hauser et al., 2024).

A closely related boosted type-II fusion experiment addresses percolation directly. Using two auxiliary two-photon states $q=m+n-1$8 and $q=m+n-1$9, with theoretical success probability $50{:}50$0, the measured Bell-state discrimination success probability was $50{:}50$1, above the graph-growth percolation threshold of $50{:}50$2 for $50{:}50$3-photon GHZ resource states (Guo et al., 2024). The same work fused two Bell states with fidelity

$50{:}50$4

This is an experimental statement about graph-state generation rather than merely Bell-state analysis.

Generalized fusion broadens the target beyond Bell projections. Numerical optimization over static linear optics shows that with $50{:}50$5 unentangled single-photon ancillas, generalized-fusion efficiencies follow the hierarchy

$50{:}50$6

for $50{:}50$7, respectively (Schmidt et al., 2024). The same study conjectures an asymptotic ceiling of $50{:}50$8 for unentangled ancillas and argues, on numerical evidence, that unit efficiency appears impossible without entangled ancilla states. At the code level, a global generalized fusion on QPC$50{:}50$9 converges numerically to $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$0, exceeding the best known Bell-measurement efficiency for that code (Schmidt et al., 2024). This directly refutes the notion that “fusion” and “Bell measurement” are synonymous in all useful settings.

The qudit generalization makes the resource constraint explicit. In generalized type-II fusion of qudit graph states, passive linear optics plus number-resolving detection yields a reduced-density-matrix rank bounded by the number $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$1 of measured systems,

$$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$2

For two-cluster fusion without ancillae,

$$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$3

so a correct $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$4-dimensional fusion requires at least $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$5 ancilla qudits (Rimock et al., 18 Oct 2025). For qubits, $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$6, and the bound reduces to the familiar ancilla-free Bell-fusion setting.

5. Noise, loss, encoding, and fault-tolerant thresholds

Fusion is constrained not only by success probability but also by storage noise, photon loss, and encoding overhead. In the dephasing-based cluster-state analysis, the primitive two-qubit cluster

$$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$7

undergoes storage-induced dephasing modeled by Kraus operators $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$8 and $$|0\rangle\langle 00|_{12} + |1\rangle\langle 11|_{12}.$$9, with $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$0 as an example time dependence (Weinstein, 2011). Recycling failed fusion remnants improves resource efficiency but lowers fidelity because stored fragments undergo additional dephasing. The paper’s qualitative conclusion is that recycling preserves resource efficiency but hurts fidelity, and that a late failure is more damaging than an early one (Weinstein, 2011).

In FBQC proper, photon-loss thresholds make the role of fusion performance quantitative. For a six-ring fusion network built from $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$1-Shor encoded six-ring resource states, the erasure probability is defined by

$$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$2

and using measured boosted-BSM data yields a photon-loss threshold of $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$3, compared with $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$4 for the non-boosted case (Hauser et al., 2024). The improvement factor is about $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$5.

Broader FBQC comparisons show that success-probability boosting and loss tolerance are distinct objectives. The survey of highly loss-tolerant photonic FBQC defines the loss per photon threshold as the maximum tolerable i.i.d. per-photon loss probability and reports that unencoded boosted fusion gives only $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$6, while a $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$7-encoded six-ring with randomized fusion failure reaches $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$8 (Bartolucci et al., 13 Jun 2025). Adaptive schemes perform substantially better: exposure-based adaptivity gives $$\Psi_{12} = \alpha\gamma H_1H_2+\alpha\delta H_1V_2+\beta\gamma V_1H_2+\beta\delta V_1V_2,$$9 for the $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$0-encoded six-ring and $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$1 for the $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$2-encoded loopy diamond. The same comparison states a non-adaptive limit including scraps of $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$3, and an adaptive asymptotic ceiling of $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$4 with single-photon measurements (Bartolucci et al., 13 Jun 2025). It also states that boosted fusion can improve fusion success probabilities, but does not provide intrinsic loss tolerance and often adds extra photons for modest gain.

Resource-state generation overhead interacts strongly with these thresholds. For the $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$5-photon Shor-encoded $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$6 $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$7-ring resource state, three source architectures were benchmarked under the fault-tolerance condition

$$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$8

The RUS architecture requires $$\Psi_3=\alpha\gamma\,0_3+\alpha\delta\,1_3+\beta\gamma\,2_3+\beta\delta\,3_3.$$9 deterministic single-photon sources with a single matter qubit degree of freedom, has optical depth $\Psi_{12}\rightarrow \Psi_3$00, allows per-component loss $\Psi_{12}\rightarrow \Psi_3$01, and yields a resource-efficiency upper bound $\Psi_{12}\rightarrow \Psi_3$02 (Wein et al., 2024). This suggests that the practical value of a fusion protocol cannot be separated from the source model that feeds it.

6. Formal protocols, switching, and compiler-level fusion

Large-scale photonic fusion also depends on control flow, switching, and compilation. In switch-network analysis for FBQC, fusion is embedded in a routing problem: photons must be synchronized, multiplexed, and conditionally directed into entangling circuits with low switch depth, because each active layer adds optical loss and error (Bartolucci et al., 2021). The paper reports that sharing photons across multiple generators can raise overall yield from about $\Psi_{12}\rightarrow \Psi_3$03 to about $\Psi_{12}\rightarrow \Psi_3$04 for two generators and about $\Psi_{12}\rightarrow \Psi_3$05 for three generators, while rastering and de Bruijn timing schemes trade source count against time-bin structure (Bartolucci et al., 2021). This makes clear that fusion efficiency is partly a network-level quantity.

A complementary formal line uses ZX calculus and dataflow semantics to classify fusion measurements and prove protocol correctness. “Fusion and flow” characterizes fusions with green failure and Pauli error, identifies Type II fusion as an X-fusion and the linear-optical measurement used for CZ-type entangling as a Y-fusion, and proves that any fusion with green failure can be boosted with a repeat-until-success protocol (Felice et al., 2024). The success probability after $\Psi_{12}\rightarrow \Psi_3$06 steps is

$\Psi_{12}\rightarrow \Psi_3$07

The same work defines an XY-flow condition for fusion networks and proves a universality result for linear optics with entangled photon sources (Felice et al., 2024).

Recent compiler work distinguishes fusion failure from fusion erasure. In the spin-memory-based MBQC architecture using caterpillar states, type-II fusion can succeed, fail in a heralded way, or undergo erasure when a photon is lost before detection (Ren et al., 23 Apr 2026). Tree-encoded fusion protects logical qubits by encoding each into a root plus $\Psi_{12}\rightarrow \Psi_3$08 three-qubit branches, enabling indirect $\Psi_{12}\rightarrow \Psi_3$09 measurement of a lost leaf through the stabilizer relation $\Psi_{12}\rightarrow \Psi_3$10. The logical success rate is written as

$\Psi_{12}\rightarrow \Psi_3$11

With $\Psi_{12}\rightarrow \Psi_3$12 and $\Psi_{12}\rightarrow \Psi_3$13, the resulting MemTree compiler reduces execution time relative to OneAdapt by $\Psi_{12}\rightarrow \Psi_3$14, photon resource use by $\Psi_{12}\rightarrow \Psi_3$15, compilation runtime by $\Psi_{12}\rightarrow \Psi_3$16, and improves fidelity by $\Psi_{12}\rightarrow \Psi_3$17 on average (Ren et al., 23 Apr 2026). This indicates that fusion has become a compilation problem as much as an optical one.

Automated circuit synthesis further complicates the historical picture in which graph states are assumed to be built only by sequential fusion. A differentiable optimization framework for heralded ballistic graph-state generators discovers direct linear-optical circuits for $\Psi_{12}\rightarrow \Psi_3$18-, $\Psi_{12}\rightarrow \Psi_3$19-, and $\Psi_{12}\rightarrow \Psi_3$20-qubit graph states with success probabilities up to $\Psi_{12}\rightarrow \Psi_3$21 better than the fusion baseline for $\Psi_{12}\rightarrow \Psi_3$22-qubit states and up to $\Psi_{12}\rightarrow \Psi_3$23 better for $\Psi_{12}\rightarrow \Psi_3$24-qubit states (Hartnett et al., 22 Aug 2025). A plausible implication is that future photonic architectures may mix direct state synthesis with fusion-based assembly rather than treating them as mutually exclusive strategies.

Taken together, these results define photonic qubit fusion as a layered concept: at the physical level it is a measurement-induced entangling operation; at the encoding level it is a redundancy-management problem; at the architectural level it determines percolation, loss thresholds, and source overhead; and at the compiler level it is governed by flow, switching, and repeat-until-success control. The field’s central technical tension remains stable across these formulations: linear optics makes fusion natural, but probabilistic success, loss, and mode management determine whether that natural primitive becomes a scalable one.