Multi-Star Topologies in Networks

• Multi-star topologies are network structures based on multiple interconnected star hubs that combine short communication paths with distributed load and redundancy.
• They are applied across domains—including distributed consensus, quantum networking, and high-performance interconnects—where trade-offs between efficiency and robustness are critical.
• Analytical frameworks such as eigenvalue convergence, graph state transformations, and star-product constructions are used to optimize connectivity, fault tolerance, and scalability.

Multi-star topologies are families of structures organized around more than one star-like center or around star-based composition rules. Across the cited literature, the term does not denote a single canonical graph class: it includes $k$-star networks with interconnected centers, fused stars sharing a center, complete-cored and $k$-cored star variants for consensus, graph-state multi-stars in quantum networks, star-product interconnects such as PolarStar, decentralized multi-hub overlays, and even multi-arm star polymers in active matter. What unifies these uses is a recurring attempt to combine short paths and aggregation around hubs with some degree of load distribution, redundancy, or modularity (Dhamal et al., 2017, Jafarizadeh et al., 2010, Blanco et al., 10 Sep 2025, Isham et al., 2024).

1. Canonical constructions and terminology

A recurrent source of ambiguity is terminological. In strategic network formation, a $k$-star network consists of $k$ center nodes forming a clique among themselves, each center having an equal or nearly equal number of leaf nodes attached only to it, with the difference in the number of leaves per center at most one (Dhamal et al., 2017). In distributed consensus, a symmetric star has one central node and $n$ path branches of length $m$; a Complete-Cored Symmetric (CCS) star replaces the center with a complete core of $n$ nodes at the branch bases; and a K-Cored Symmetric (KCS) star replaces the center with $k$ parallel central nodes that are not connected to each other, while each tail connects to all $k$ central nodes (Jafarizadeh et al., 2010). In the “Two Fused Star” construction, two different symmetric star networks share the same central node (Jafarizadeh, 2010). In graph-state quantum networking, a multi-star is a chain of $m$ interconnected switches, each attached to $k$0 or $k$1 client leaves (Blanco et al., 10 Sep 2025). In high-performance interconnect design, star-product topologies arise from composing a structure graph and a supernode graph, producing modular low-diameter networks that include Slim Fly, Bundlefly, and PolarStar (Isham et al., 2024, Lakhotia et al., 2023).

Variant	Defining structure	Representative context
$k$2-star	Clique of $k$3 centers with balanced leaves	Strategic network formation (Dhamal et al., 2017)
CCS / KCS star	Complete core or $k$4 parallel central nodes	Fastest distributed consensus (Jafarizadeh et al., 2010)
Two Fused Star	Two symmetric stars sharing one center	Consensus on heterogeneous branches (Jafarizadeh, 2010)
Multi-star graph state	Chain of switches, each with leaf clients	Logical quantum connectivity (Blanco et al., 10 Sep 2025)
Star product	$k$5 composition of factor graphs	Diameter-3 interconnects and EDSTs (Isham et al., 2024)

This plurality matters because performance claims are definition-dependent. A statement about convergence on CCS stars, for example, is not automatically a statement about graph-state multi-stars or about star polymers.

2. Analytical frameworks and observables

The literature evaluates multi-star topologies through domain-specific observables rather than a single universal criterion. In distributed average consensus, node states evolve according to

$k$6

and the convergence rate is governed by the Second Largest Eigenvalue in Magnitude,

$k$7

The fastest distributed consensus problem is then the optimization of edge weights subject to sparsity and stochasticity constraints (Jafarizadeh et al., 2010, Jafarizadeh, 2010).

In graph-state quantum networking, the central formal tools are graph states and LOCC. A node measured in the $k$8 basis is deleted; a $k$9 measurement performs local complementation and deletion; and an $k$0 measurement is implemented by a sequence of local complementations and deletion. These operations permit the transformation of a physical multi-star into logical cliques, bi-stars, or tri-stars (Blanco et al., 10 Sep 2025). In multipartite entanglement routing on real topologies, performance is measured by average waiting time $k$1 and rate

$k$2

with repeater trimming analyzed through the retained-rate ratio $k$3 (Ooi et al., 26 Mar 2026).

In optically interconnected quantum data centers, communication cost is measured by the number of required inter-QPU entanglement links and quality by RCNOT and GHZ-state fidelity under transduction and fiber noise (Elyasi et al., 21 May 2026). In multi-agent LLM communication, the diagnostic object is the successor representation

$k$4

derived from the row-stochastic communication operator $k$5; the relevant summaries are the spectral radius $k$6, spectral gap $k$7, and condition number $k$8 (Park et al., 12 May 2026). In active-matter studies of star polymers, the principal observables are the radius of gyration, mean squared displacement, diffusivity, VACF, persistence time, inter-arm separation distributions, and arm-angle distributions (Yadav et al., 13 Oct 2025).

A plausible implication is that “multi-star topology” is better understood as a structural motif than as a single model: the repeated object is a hub-distributed architecture, but the operative notion of optimality depends on whether the task is consensus, routing, entanglement distribution, robustness, or transport under nonequilibrium drive.

3. Consensus, aggregation, and communication efficiency

For distributed consensus, the basic symmetric star suffers from a central bottleneck, and this motivates multi-center refinements. In the symmetric star with $k$9 branches of length $k$0, the optimal central-edge weight is

$k$1

while other tail edges have weight $k$2, and the convergence factor is $k$3 where $k$4 is the smallest root of

$k$5

In the CCS star, core edges have weight $k$6, tail edges again have weight $k$7, and

$k$8

which is independent of $k$9. In the KCS star, the center-connection weight becomes

$n$0

and increasing $n$1 reduces SLEM up to an optimal value. The empirical ordering reported is

$n$2

so CCS converges fastest at the cost of a denser core (Jafarizadeh et al., 2010).

The fused-star case shows how branch heterogeneity degrades performance. In a Two Fused Star network with $n$3 branches of length $n$4 and $n$5 branches of length $n$6, most inner-path edges again receive optimal weight $n$7, while the central incident weights depend on $n$8 through

$n$9

with $m$0. The corresponding transcendental equation determines SLEM numerically. For fixed total size, the paper reports that a symmetric star converges faster than the fused asymmetric alternative (Jafarizadeh, 2010).

Aggregation has a different but related role in multi-agent LLM systems. For a star with one hub and $m$1 leaves at $m$2, the successor-representation spectrum gives

$m$3

The study reports that the condition number is a perfect rank-order predictor of empirical perturbation robustness, that the spectral gap partially predicts consensus dynamics, and that the spectral radius is perfectly inverted with respect to cumulative error. The inversion is attributed to a regime in which linear spectra are blind to non-contracting bias drift, and the paper therefore introduces a drift-corrected diagnostic. Star topologies reduce cumulative error relative to chains through aggregation, but they are also more fragile to adversarial manipulation unless individual leaf influence is capped (Park et al., 12 May 2026).

These results establish a consistent theme: multi-star refinements are often introduced to weaken a single-hub bottleneck, but the same concentration of influence that accelerates averaging or consensus can also create sensitivity to malformed or malicious inputs.

4. Quantum-network and quantum-data-center realizations

In optically interconnected quantum data centers, Elyasi et al. emulate ring, star, and line topologies on quantum hardware through a three-layer framework: a circuit layer for GHZ-state generation, a configuration layer that maps the circuit onto a physical topology and injects transduction and fiber noise, and an experimental layer that maps the noisy topology-aware circuit onto superconducting QPU layouts. The paper uses the cat-comm protocol for RCNOT gates, models transduction noise and fiber noise by a collisional Markovian process with environment resets, and represents the environmental interaction as

$m$4

Multi-hop entanglement is realized via SWAP operations, which introduce extra delay and decoherence. Under the paper’s communication-cost metric for all-to-all GHZ connectivity among $m$5 QPUs,

$m$6

For four QPUs, the cost ordering is star $m$7 ring $m$8 line. Monolithic RCNOT fidelity is reported as $m$9, while distributed RCNOT drops to $n$0 after several hops with 10 meters of fiber and transduction coupling constant $n$1. GHZ benchmarking shows that the star topology provides the best fidelity because direct connections to the hub minimize hops, routing, and noise. The experiments were executed on the IBM Quantum Torino backend, a Heron 133-qubit processor supporting mid-circuit measurements. Multi-star arrangements are not explicitly named in that study, but the methodology is stated to apply to scalable star-centric or hub-and-spoke ensembles (Elyasi et al., 21 May 2026).

A complementary quantum-network perspective treats multi-stars as graph-state resources for logical connectivity. Extending earlier bi-star configurations associated with Chen et al. to $n$2 switches arranged in a linear chain, each with $n$3 clients, the graph-state protocol deletes all clients on every other switch via $n$4 measurements and then performs $n$5 measurements on those pruned switches. For odd $n$6, the maximal clique size distilled from this architecture is

$n$7

using $n$8 measurements. In the asymmetric case with $n$9 clients per switch,

$k$0

where the choice of the deleted-switch set $k$1 determines the optimum. The paper also emphasizes two-colorability constraints for the resulting graph states (Blanco et al., 10 Sep 2025).

Routing studies on 81 real topologies qualify the limits of pure star-based quantum distribution. Four performance regimes are identified: globally adverse topologies, tree-dominant topologies, multi-path-dominant topologies, and globally favourable topologies. The evaluated protocols are SPS, SPT, MPS, and MPT. Star-based protocols are constrained by the need for disjoint paths to a central node; consequently, they are only optimal when the network closely resembles a pure star. In well-performing topologies, when only 80% of nodes operate as repeaters, over 90% of the distribution rate can be retained, whereas sparse weakly connected graphs retain less than half. In favourable topologies, over 38% of active repeaters can be trimmed while still preserving 50% of the maximum rate, especially for MPT. The study describes large, well-connected, high-redundancy topologies as “multi-star-like,” and reports that these are substantially more robust than sparse star-like cases under repeater trimming (Ooi et al., 26 Mar 2026).

The quantum literature therefore distinguishes between two roles for multi-star structure: as a physical topology that reduces hops and noise, and as a logical resource that can be distilled into high-connectivity graph states.

5. Star products, large-scale interconnects, and resilience

Star-product constructions generalize multi-star thinking to large modular interconnects. For graphs $k$2 and $k$3, the star product $k$4 has vertex set

$k$5

with intra-supernode edges inherited from $k$6 and inter-supernode edges defined by bijections along edges of $k$7. The resulting network satisfies

$k$8

This framework encompasses modern low-diameter topologies such as Slim Fly, Bundlefly, PolarStar, HyperX, mesh, and torus (Isham et al., 2024).

PolarStar is a diameter-3 family derived from the star product of an Erdős-Rényi polarity graph $k$9 and either Inductive-Quad or a Paley graph. Its order is

$k$0

with approximate maximal order

$k$1

For almost all feasible radixes $k$2, PolarStar is reported as the largest known diameter-3 network, with a geometric-mean scale increase of $k$3 over Bundlefly, $k$4 over Dragonfly, and $k$5 over 3-D HyperX. Its average bisection is approximately $k$6 of all links, exceeding Bundlefly, Dragonfly, and HyperX, and random link failure of up to 60% can be tolerated before disconnection, with graceful diameter growth under initial failures (Lakhotia et al., 2023).

The same star-product formalism supports edge-disjoint spanning tree constructions. If $k$7 has $k$8 EDSTs and $k$9 residual non-tree edges, while $m$0 has $m$1 EDSTs and $m$2 residual edges, then the star product admits

$m$3

The theoretical graph upper bound is

$m$4

and one construction yields tree depth of order the depths of the factor graphs, with the generic bound

$m$5

These EDSTs are motivated by collective communication, fault tolerance, network recovery, and congestion management (Isham et al., 2024).

At the overlay level, the Elevator protocol generates emergent multi-hub peer-to-peer networks by combining preferential attachment with random attachment. Each node maintains a cache of size $m$6, allocates exactly $m$7 hub-oriented connections, and fills the remaining $m$8 slots through random or backward sampling. The number of hubs is thus tunable by the parameter $m$9. The reported overlay has average path length approximately 2 and diameter approximately 2, and remains connected under random failure of up to 80% of nodes; after targeted hub attacks, hubs reemerge within a few cycles. The protocol is decentralized, asynchronous, and uses only local information (Legheraba et al., 2024).

Taken together, these studies portray multi-star structure as a scalable design principle: modular enough for large order, but still sufficiently centralized to preserve low diameter.

6. Formation mechanisms, non-network analogues, and conceptual boundaries

Not all multi-star topologies are engineered explicitly; some arise from incentives or dynamics. In strategic network formation with recursive entry and myopic best responses, a $k$00-star uniquely emerges only under stringent conditions. The network must start from a base graph consisting of a clique among the $k$01 centers, each with one attached leaf; for $k$02, formation from a single node is not possible under the basic rules. The sufficient conditions reported are

$k$03

Under these settings, the resulting topology is pairwise stable and unique. The paper also reports that the efficiency of a $k$04-star relative to the star, for large $k$05, is approximately

$k$06

Small deviations in $k$07 or $k$08 destroy $k$09-star uniqueness or stability (Dhamal et al., 2017).

A different non-engineering manifestation appears in soft matter. In a dilute bath of active Brownian particles, three-armed star polymers develop arm pairing at high Péclet number: two arms pair while the third remains extended, due to asymmetric accumulation of active particles in regions of higher curvature. The inter-arm separation distribution becomes bimodal, and the arm-angle distribution sharpens at low values. Semiflexible stars undergo the pairing transition at lower $k$10 than fully flexible stars, although rigidity effects diminish at high $k$11. Dynamically, the star polymer’s center of mass exhibits enhanced mean squared displacement,

$k$12

with $k$13 at high activity, and its long-time diffusivity

$k$14

grows faster with activity than that of a single colloidal bead or a linear chain. The VACF remains positive over extended times, and the persistence time

$k$15

rises markedly when arm pairing becomes prominent. When a three-arm star with 61 beads is compared with a 41-bead linear chain chosen to match the paired-arm conformation, structural and dynamical measures converge at high activity; yet for equal total bead count, the star remains more compact and exhibits faster, more persistent dynamics than the linear polymer (Yadav et al., 13 Oct 2025).

A recurrent misconception is therefore that “multi-star” is exclusively a graph-theoretic or communication-network term. The polymer study shows that the same language can describe a molecular topology whose salient phenomena are conformational asymmetry and transport in active baths rather than routing or consensus. This suggests that the concept’s true common denominator is not a particular algorithmic property, but a structural decomposition into multiple arms or hubs coupled through a more central core or set of cores.

Across these domains, the central trade-off remains stable even as the metrics change. Multi-star organization can shorten paths, accelerate aggregation, and expose parallel routes, but it also introduces bottlenecks, hub sensitivity, or balancing constraints. The detailed literature thus treats multi-star topologies not as universally optimal objects, but as a family of architectures whose value depends on how centralization, redundancy, and symmetry interact with the governing dynamics (Jafarizadeh et al., 2010, Elyasi et al., 21 May 2026, Lakhotia et al., 2023, Park et al., 12 May 2026).