Split-Fuse Construction in Multidomain Systems

• Split-fuse construction is a two-stage process where an object is decomposed into subcomponents (split) and then recombined (fuse) to achieve global, meaningful structures.
• In integrable models and ribbon-knot theory, the split stage isolates rank-deficient or quotient components while the fuse stage reconstructs the system via operator factorization and fusion gates.
• The design principle spans various domains—from computer vision and quantum-information processing to deterministic filtering—enhancing performance by local decomposition followed by controlled recombination.

Searching arXiv for the cited papers on split-fuse constructions across domains. Across the cited arXiv literature, the expression split-fuse construction does not denote a single universal formalism. It denotes a family of two-stage constructions in which a composite object is first decomposed into subspaces, components, quotient pieces, or specialized branches, and is then recombined by an operator factorization, a handle attachment, a circuit conversion, a fusion gate, an optimal-transport alignment, or an auxiliary data structure. In integrable models, the split-fuse step is an algebraic factorization of a rank-deficient $R$-matrix; in ribbon-knot theory it is the passage between births-and-saddles and complement handle decompositions; in graph-state preparation it is a QASST-guided decomposition followed by Type-II fusions; in unsupervised saliency it is entropy-guided dual clustering followed by OT alignment; and in ZOR filters it is deterministic key abandonment followed by auxiliary recovery of false-positive-only semantics (Beisert et al., 2015, Hom et al., 2020, Connolly et al., 25 Mar 2026, Ramzan et al., 20 Oct 2025, Limasset, 3 Feb 2026).

1. Conceptual schema

A common structural pattern is visible across otherwise unrelated domains. The “split” stage isolates a latent structure that is difficult to handle directly: a bound-state subspace inside $V^F\otimes V^F$, the quotient pieces of a graph under strong splits, high-entropy boundary pixels in saliency detection, sub-target domains in open compound domain adaptation, or a residual core in peeling-based filter construction. The “fuse” stage then reconstructs an object with the intended global semantics: a fused $R$-matrix, a ribbon disk or its complement, a target graph state, a globally consistent pseudo-mask, a fused segmentation prediction, or a membership filter with false-positive-only behavior (Beisert et al., 2015, Connolly et al., 25 Mar 2026, Ramzan et al., 20 Oct 2025, Gong et al., 2020, Limasset, 3 Feb 2026).

A recurrent misconception is that split-fuse is merely a synonym for ordinary fusion. The surveyed literature shows otherwise. In the generalized fusion formalism for integrable models, the construction is formulated directly in terms of a rank-deficient $R$-matrix, not merely a projector onto an obvious irreducible component (Beisert et al., 2015). In ribbon-knot theory, the split and fuse sides are not even equivalent complexity measures: the fusion number $F(K)$ and the strong homotopy fusion number $\mathcal F_{sh}(K)$ can diverge dramatically under cabling (Hom et al., 2020). In graph-state preparation, the split-fuse language is algorithmic rather than representation-theoretic: the target graph is recovered from quotient graph states by Type-II fusions prescribed by the quotient-augmented strong split tree (Connolly et al., 25 Mar 2026).

2. Algebraic split-fuse in integrable models

The most explicit algebraic formulation appears in the construction of bound-state $R$-matrices for AdS/CFT worldsheet scattering and the one-dimensional Hubbard model (Beisert et al., 2015). One starts from a fundamental $R$-matrix

$R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$

satisfying the Yang–Baxter equation and the involution property. At special points $u_{12}\in M^B$, its rank drops,

$V^F\otimes V^F$0

and this signals a composite or bound-state sector $V^F\otimes V^F$1. The split-fuse factorization is

$V^F\otimes V^F$2

with

$V^F\otimes V^F$3

subject to

$V^F\otimes V^F$4

In the paper’s language, $V^F\otimes V^F$5 extracts the bound-state subspace, $V^F\otimes V^F$6 reinserts it, and $V^F\otimes V^F$7 is the effective operator on that fused space.

Two equivalent constructions are given. In the image-based version, $V^F\otimes V^F$8 is the image of $V^F\otimes V^F$9, with $R$0 the inclusion, $R$1, and $R$2, provided $R$3 is invertible. In the eigenvector version, for trivial Jordan form and nonzero eigenvalues $R$4 with right and left eigenvectors $R$5,

$R$6

so that

$R$7

The identities

$R$8

drive the fusion proof. Mixed and doubly fused $R$9-matrices are then constructed by inserting fundamental $R$0-matrices between $R$1 and $R$2, for example

$R$3

and

$R$4

which makes the order-independence of fusion manifest.

The fused operators inherit the integrable structure: they satisfy involution relations such as

$R$5

and the corresponding mixed Yang–Baxter equations. If the original $R$6-matrix intertwines a Hopf-algebra coproduct, the fused representation is

$R$7

and the fused $R$8-matrix intertwines this induced representation. The paper’s main application is the non-standard centrally extended $R$9 $F(K)$0-matrix of AdS/CFT, or equivalently Shastry’s Hubbard $F(K)$1-matrix. At the special points

$F(K)$2

the $F(K)$3-matrix drops to rank $F(K)$4, producing the symmetric short multiplet bound-state representation. The formalism also accommodates complementary fusion, opposite fusion ordering, similarity transforms, and singular singlet cases with nontrivial Jordan form, including the singularity at

$F(K)$5

3. Geometric and topological variants

In ribbon-knot theory, split-fuse language is realized literally in the ribbon-disk movie (Hom et al., 2020). A ribbon knot is built from disjoint disks by attaching only $F(K)$6-handles, and the paper formalizes two complexity measures: the fusion number $F(K)$7, the minimal number of $F(K)$8-handles needed to construct a ribbon disk for $F(K)$9, and the strong homotopy fusion number $\mathcal F_{sh}(K)$0, the minimal number of $\mathcal F_{sh}(K)$1-handles in a handle decomposition of the ribbon disk complement $\mathcal F_{sh}(K)$2. These satisfy

$\mathcal F_{sh}(K)$3

The movie has births and saddles on the knot side, while in Kirby-diagram language each birth becomes a dotted circle and each saddle becomes a $\mathcal F_{sh}(K)$4-framed $\mathcal F_{sh}(K)$5-handle in the complement. The split-fuse terminology therefore links the primal ribbon construction to the dual handle decomposition.

The paper’s main theorem shows that these invariants behave completely differently under $\mathcal F_{sh}(K)$6-cabling. If $\mathcal F_{sh}(K)$7 is ribbon and $\mathcal F_{sh}(K)$8, then

$\mathcal F_{sh}(K)$9

and, more generally, for an iterated $R$0-cable,

$R$1

The geometric reason is that the cable disk $R$2 is built from $R$3 parallel copies of $R$4 and $R$5 half-twisted bands, while the complement can be simplified because the $R$6-cable of the unknot is again the unknot. The lower bound on $R$7 comes from knot Floer homology via the torsion order

$R$8

together with the inequality $R$9. A unit-box summand in $R$0 is the seed that cabling amplifies, and Hanselman–Watson’s immersed-curve cabling formula yields

$R$1

A different geometric use of splitting appears in the conformally covariant split system for the vacuum Einstein constraint equations (Mach et al., 2018). Here the unknowns are a conformal factor $R$2 and a $R$3-form or vector field $R$4, and the system is written so as to be conformally covariant rather than in the standard Lichnerowicz–York form. If $R$5 solves the split system, the physical initial data are reconstructed by

$R$6

Using the implicit function theorem, the paper proves existence of solutions near a constant-mean-curvature seed and thereby constructs non-CMC vacuum initial data on compact $R$7-manifolds, with or without boundary. In the boundary case, $R$8 is treated as an apparent horizon through boundary conditions enforcing $R$9 and $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$0. The framework is extended to nonzero cosmological constant and illustrated numerically on $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$1 and $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$2. This suggests a geometric split-fuse pattern in which one solves for split conformal variables and then fuses them back into physical vacuum data.

4. Quantum-information constructions

One quantum-information use of split-fuse is the conversion between two qubits and one four-dimensional qudit (Moussa, 2015). The fusion map is

$R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$3

and the inverse-style fission map is

$R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$4

Treating a $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$5-level qudit as a pair of qubits produces the nesting

$R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$6

where $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$7 denotes the $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$8-th level of the $R(u_1,u_2):V^F\otimes V^F\to V^F\otimes V^F$9-dimensional Clifford hierarchy. In this representation, qudit Pauli operators are realized as qubit Clifford operators, whereas qudit Clifford generators become qubit non-Clifford circuits involving gates such as $u_{12}\in M^B$0, controlled-$u_{12}\in M^B$1, and Toffoli. Complete fusion and fission are implemented by stabilizer circuits that consume a resource state $u_{12}\in M^B$2, itself a fused qubit stabilizer state with a fault-tolerant preparation using stabilizer circuits. The construction is therefore simultaneously a code-conversion scheme and a route from qudit Clifford structure to qubit non-Clifford power.

A second quantum-information instance is the preparation of graph states by split decomposition and the quotient-augmented strong split tree (QASST) (Connolly et al., 25 Mar 2026). A split of a graph $u_{12}\in M^B$3 is a bipartition

$u_{12}\in M^B$4

such that the crossing edges form a complete bipartite subgraph. Strong splits yield a tree of quotient graphs $u_{12}\in M^B$5, with split-nodes $u_{12}\in M^B$6 and $u_{12}\in M^B$7 on adjacent quotients. The core reconstruction statement is that if $u_{12}\in M^B$8 has split decomposition $u_{12}\in M^B$9, then $V^F\otimes V^F$00 can be recovered from $V^F\otimes V^F$01 using Type-II fusions on all connected split-node pairs described by the QASST. For distance-hereditary graphs, every quotient graph is star- or complete-type, hence LC-equivalent to a star; the preparation protocol therefore builds star graph states for the quotients, applies local Clifford transformations if needed, and performs one Type-II fusion per QASST edge.

The resulting resource bounds are explicit. With $V^F\otimes V^F$02, $V^F\otimes V^F$03 quotient graphs, $V^F\otimes V^F$04 QASST edges, and $V^F\otimes V^F$05 auxiliary split-node qubits, the protocol uses

$V^F\otimes V^F$06

$V^F\otimes V^F$07

and

$V^F\otimes V^F$08

Because $V^F\otimes V^F$09, all three resources scale linearly for arbitrary distance-hereditary graph states. The paper’s worked example $V^F\otimes V^F$10 uses $V^F\otimes V^F$11 total qubits, $V^F\otimes V^F$12 auxiliary qubits, $V^F\otimes V^F$13 CZ gates, and $V^F\otimes V^F$14 time steps. For complete multipartite and clique-star families, the method eventually outperforms direct preparation of an optimal LC representative, with first observed depth improvement around $V^F\otimes V^F$15 and CZ-gate improvement around $V^F\otimes V^F$16.

5. Split-fuse in machine learning and computer vision

In unsupervised salient-object detection, the Split-Fuse-Transport design of POTNet is introduced because one clustering strategy is not sufficient for heterogeneous pixel geometry (Ramzan et al., 20 Oct 2025). The split criterion is pixel-wise CAM entropy,

$V^F\otimes V^F$17

where high entropy indicates global ambiguity and low entropy indicates local consistency. High-entropy pixels are routed to spectral clustering, low-entropy pixels to k-means, and a soft gate

$V^F\otimes V^F$18

blends the two assignments. The fused memberships define class-wise prototype sets

$V^F\otimes V^F$19

which are intended to capture intra-class diversity and semantic parts. These prototypes are then aligned to pixel features by optimal transport with cosine-distance cost and prototype-aware column marginals. The transport plan is used for a hard global reassignment, after which prototype CAMs are reweighted by OT mass and Otsu thresholding produces the pseudo-mask. Those pseudo-masks supervise AutoSOD, a MaskFormer-style encoder-decoder trained without pixel-level labels.

The ablation evidence in the paper is tied directly to the split-fuse components. On DUTS-TE, full AutoSOD reports

$V^F\otimes V^F$20

whereas “Only K-means Clustering” reports

$V^F\otimes V^F$21

“Only Spectral Clustering” reports

$V^F\otimes V^F$22

“w/o OT Consistency Loss” reports

$V^F\otimes V^F$23

and “w/o Reweighting” reports

$V^F\otimes V^F$24

The authors also state that AutoSOD outperforms unsupervised methods by up to $V^F\otimes V^F$25 and weakly supervised methods by up to $V^F\otimes V^F$26 in $V^F\otimes V^F$27-measure on five benchmarks.

A related but distinct split-fuse architecture appears in open compound domain adaptive semantic segmentation (Gong et al., 2020). The target domain is first clustered into $V^F\otimes V^F$28 sub-target domains by style codes extracted with MUNIT and clustered by K-means. The segmentation network is then split into $V^F\otimes V^F$29 independent branches with compound-domain-specific batch normalization, one branch per sub-target domain. A hypernetwork $V^F\otimes V^F$30 takes the style code $V^F\otimes V^F$31 and outputs fusion weights

$V^F\otimes V^F$32

so that the final prediction is

$V^F\otimes V^F$33

This replaces the discrete partition by a continuous mixture over domain-specialized experts. The model is then updated by MAML, with inner-loop self-entropy minimization and an outer-loop loss combining the fuse objective and entropy. The paper’s “cluster, split, fuse, and update” sequence makes the split-fuse step a bridge between domain specialization and open-domain generalization.

6. Deterministic filtering and operator-algebraic fusion

In probabilistic membership filters, ZOR filters implement a split-fuse construction at build time rather than query time (Limasset, 3 Feb 2026). XOR and fuse filters are compact and fast, but their peeling-based construction can fail when the residual hypergraph has no degree-$V^F\otimes V^F$34 cell, forcing a restart with new hash seeds. ZOR replaces restart-on-failure with deterministic peeling. When the peel queue is empty but keys remain, it chooses a cell $V^F\otimes V^F$35 of minimum current degree

$V^F\otimes V^F$36

keeps one incident key $V^F\otimes V^F$37, abandons the other $V^F\otimes V^F$38 keys, removes those abandoned keys from all their incident cells, and continues ordinary peeling. Termination is guaranteed because each step removes at least one still-active key. The abandoned set $V^F\otimes V^F$39 has fraction $V^F\otimes V^F$40; to recover false-positive-only semantics, one stores the subset $V^F\otimes V^F$41 that would otherwise return “absent” in a compact auxiliary structure, such as a fuse filter or an MPHF-plus-fingerprint table. The combined false-positive rate is

$V^F\otimes V^F$42

and the total bits per original key are

$V^F\otimes V^F$43

For the auxiliary sizing rule,

$V^F\otimes V^F$44

The reported abandonment rates for large tested sets are $V^F\otimes V^F$45 for $V^F\otimes V^F$46-wise, $V^F\otimes V^F$47 for $V^F\otimes V^F$48-wise, $V^F\otimes V^F$49 for $V^F\otimes V^F$50-wise, $V^F\otimes V^F$51 for $V^F\otimes V^F$52-wise, $V^F\otimes V^F$53 for $V^F\otimes V^F$54-wise, and $V^F\otimes V^F$55 for $V^F\otimes V^F$56-wise filters. The paper states that, in its experiments, ZOR achieves overhead within $V^F\otimes V^F$57 of the information-theoretic lower bound $V^F\otimes V^F$58, while the main runtime penalty is concentrated on negative queries because they may need both main and auxiliary checks.

An operator-algebraic notion of fusion appears in the theory of implementers for odd spinors on the circle (Kristel et al., 2019). The Hilbert space

$V^F\otimes V^F$59

is decomposed as

$V^F\otimes V^F$60

where $V^F\otimes V^F$61 are the Hilbert completions of spinors supported on the lower and upper semicircles. This decomposition induces a graded Clifford-algebra splitting

$V^F\otimes V^F$62

On the implementer side, there is a central extension

$V^F\otimes V^F$63

and the paper constructs a lift of fusion to this central extension using Tomita–Takesaki theory. The key device is a fusion factorization, a smooth group homomorphism

$V^F\otimes V^F$64

satisfying $V^F\otimes V^F$65 on doubled paths. The resulting fusion product is

$V^F\otimes V^F$66

and it satisfies associativity, smoothness, and multiplicativity. The modular conjugation is computed explicitly as

$V^F\otimes V^F$67

which identifies the reflection operator needed for fusion with the modular-theoretic structure of the Clifford-von Neumann algebra. Specializing to $V^F\otimes V^F$68, the construction yields a fusion product on the basic central extension that is compatible with the canonical connection and with the transgression model from the basic gerbe.

7. Comparative features and recurrent issues

The surveyed papers suggest that split-fuse constructions are most useful when the target object contains a hidden heterogeneity that a one-piece construction handles poorly. In integrable models that heterogeneity appears as a rank-deficient $V^F\otimes V^F$69-matrix with a nontrivial bound-state image (Beisert et al., 2015). In ribbon knots it appears as the divergence between band complexity and complement-handle complexity under cabling (Hom et al., 2020). In graph states it is the decomposition into quotient graphs connected by strong splits (Connolly et al., 25 Mar 2026). In saliency detection it is the difference between low-entropy interior pixels and high-entropy boundary pixels (Ramzan et al., 20 Oct 2025). In domain adaptation it is the presence of multiple unknown homogeneous sub-domains inside the target distribution (Gong et al., 2020). In ZOR filters it is the stubborn residual core left by failed peeling (Limasset, 3 Feb 2026).

A second recurrent feature is that the fusion stage rarely means a naive inverse to the split stage. In the generalized $V^F\otimes V^F$70 construction, the fused object is defined only up to similarity transformations and may coexist with a meaningful complementary fused space (Beisert et al., 2015). In ribbon-knot cabling, the complement can remain strongly simple while the ribbon-disk fusion number grows linearly with $V^F\otimes V^F$71 (Hom et al., 2020). In graph-state preparation, fusion consumes auxiliary split-node qubits and relies on Type-II fusion operations, so the reconstruction is operational rather than purely formal (Connolly et al., 25 Mar 2026). In ZOR filters, auxiliary recovery restores semantics but introduces an extra negative-query cost (Limasset, 3 Feb 2026). In operator algebra, fusion is mediated by a central extension and a connection-preserving factorization rather than by direct concatenation of operators (Kristel et al., 2019).

A final common theme is that split-fuse methods typically trade global monolithic optimization for structured local processing followed by a controlled recombination rule. This suggests why they recur in such distant settings. The split stage isolates a tractable subproblem—an image, a quotient graph, a sub-target branch, a complement handlebody, a bound-state image, or a peeled key set—while the fuse stage enforces the global semantics that would be difficult to maintain if those local pieces were treated independently. The cited literature therefore presents split-fuse construction not as a single theorem, but as a recurrent design principle linking decomposition, specialization, and reconstruction across algebra, geometry, quantum information, machine learning, data structures, and operator theory (Beisert et al., 2015, Mach et al., 2018, Moussa, 2015, Ramzan et al., 20 Oct 2025).