See discussions, stats, and author profiles for this publication at: [Link]

net/publication/392096134

Multi-SuperHyperGraph Neural Networks: A Generalization of Multi-

HyperGraph Neural Networks

Preprint · October 2025

DOI: 10.13140/RG.2.2.28787.03362

CITATIONS READS

0 62

1 author:

Takaaki Fujita

470 PUBLICATIONS 3,198 CITATIONS

SEE PROFILE

All content following this page was uploaded by Takaaki Fujita on 26 May 2025.

The user has requested enhancement of the downloaded file.

 Multi-SuperHyperGraph Neural Networks: A Generalization of
Multi-HyperGraph Neural Networks
Takaaki Fujita 1 ∗
1 Independent Researcher, Shinjuku, Shinjuku-ku, Tokyo, Japan.

Abstract
Graph theory provides a mathematical framework for modeling relationships among entities via vertices
(nodes) and edges [1–3]. A hypergraph extends this framework by allowing hyperedges to connect any
number of vertices, thereby capturing complex multi-way interactions [4, 5]. The SuperHyperGraph concept
generalizes hypergraphs further through iterated power-set constructions and has recently drawn significant
research interest [6, 7].

Graph Neural Networks (GNNs) propagate and aggregate node features across graph topologies via learnable
message-passing to capture structural context [8–10]. Extensions such as Hypergraph Neural Networks,
SuperHyperGraph Neural Networks, Multigraph Neural Networks, and MultiHyperGraph Neural Networks
have likewise been explored [11, 12].

In this paper, we introduce and analyze the Multi 𝑛-SuperHyperGraph Neural Network, a theoretical extension
of SuperHyperGraph Neural Networks built upon Multi-SuperHyperGraph structures. We expect that this
framework will stimulate further advances in the study and application of GNNs.

Keywords: Graph Neural Networks (GNNs), HyperGraph, SuperHyperGraph, Multigraph Neural Networks,
MultiHyperGraph Neural Networks, Hypergraph Neural Networks, SuperHyperGraph Neural Networks

1 Preliminaries and Definitions

This section introduces the foundational concepts and definitions required for the discussions in this paper.
Throughout this work, all sets and structures are assumed to be finite. Unless otherwise specified, the parameter
𝑛 is taken to be a non-negative integer.

1.1 SuperHyperGraph

A hypergraph generalizes a classical graph by introducing hyperedges, which can connect any number of
vertices—not just two—making it suitable for modeling complex, multi-way relationships [4, 11, 13–19]. A
SuperHyperGraph takes this concept even further. Recently introduced and actively studied in a growing body
of literature [6, 7, 20–29], the SuperHyperGraph incorporates recursive structures into hypergraphs through
iterated applications of the power set operation. Conceptually, a SuperHyperGraph can be viewed as a
hierarchical generalization of a hypergraph, in which both vertices and hyperedges are drawn from higher-order
powersets of a base vertex set. The formal definition is presented below.
Definition 1.1 (Powerset). (cf. [30]) Let 𝑆 be any set. The powerset of 𝑆, denoted P (𝑆), is the collection of all
subsets of 𝑆:
P (𝑆) = { 𝐴 | 𝐴 ⊆ 𝑆}.
In particular, ∅ ∈ P (𝑆) and 𝑆 ∈ P (𝑆).
Definition 1.2 (𝑛-th Powerset). (cf. [31, 32])

The 𝑛-th powerset of a set 𝐻, denoted 𝑃𝑛 (𝐻), is constructed iteratively. Beginning with the standard powerset,
the process is defined as:

𝑃1 (𝐻) = 𝑃(𝐻), 𝑃𝑛+1 (𝐻) = 𝑃(𝑃𝑛 (𝐻)), for 𝑛 ≥ 1.

In a similar manner, the 𝑛-th non-empty powerset, represented as 𝑃𝑛∗ (𝐻), is recursively defined as:

𝑃1∗ (𝐻) = 𝑃∗ (𝐻), ∗

𝑃𝑛+1 (𝐻) = 𝑃∗ (𝑃𝑛∗ (𝐻)).

Here, 𝑃∗ (𝐻) refers to the powerset of 𝐻 excluding the empty set.

1
Definition 1.3 (Hypergraph [4, 33]). A hypergraph 𝐻 = (𝑉 (𝐻), 𝐸 (𝐻)) is a pair where:

• 𝑉 (𝐻): A non-empty set of vertices.

• 𝐸 (𝐻): A set of hyperedges, each of which is a subset of 𝑉 (𝐻).

This paper focuses exclusively on finite hypergraphs.

Definition 1.4 (n-SuperHyperGraph). [7, 24] Let 𝑉0 be a finite base set of vertices. For each 𝑘 ≥ 0, define the
iterative powerset P 𝑘 (𝑉0 ) by
P 0 (𝑉0 ) = 𝑉0 , P 𝑘+1 (𝑉0 ) = P P 𝑘 (𝑉0 ) ,

where P (·) denotes the power set. An n-SuperHyperGraph is a pair

SHT (𝑛) = (𝑉, 𝐸),

with
𝑉 ⊆ P 𝑛 (𝑉0 ) and 𝐸 ⊆ P 𝑛 (𝑉0 ).
Each element of 𝑉 is an n-supervertex, and each element of 𝐸 is an n-superedge.
Example 1.5 (2-SuperHyperGraph). Let the base set be

𝑉0 = {𝑎, 𝑏}.

Then
𝑃1 (𝑉0 ) = ∅, {𝑎}, {𝑏}, {𝑎, 𝑏} , 𝑃2 (𝑉0 ) = P 𝑃1 (𝑉0 ) .


We choose the set of 2-supervertices

n o
𝑉 = 𝑣 1 = {{𝑎}}, 𝑣 2 = {{𝑏}, {𝑎, 𝑏}} ⊆ 𝑃2 (𝑉0 ),

and the set of 2-superedges

n o
𝐸 = 𝑒 1 = {{𝑎}, {𝑏}}, 𝑒 2 = {∅, {𝑎, 𝑏}} ⊆ 𝑃2 (𝑉0 ).

Thus
SHT (2) = 𝑉, 𝐸

is a 2-SuperHyperGraph in which:

• 𝑣 1 and 𝑣 2 are two distinct 2-supervertices drawn from the second iterated powerset of 𝑉0 .
• 𝑒 1 and 𝑒 2 are two distinct 2-superedges, each a 2-element subset of 𝑃1 (𝑉0 ).
• Every supervertex and superedge lies within 𝑃2 (𝑉0 ), illustrating the hierarchical structure.

1.2 Multi 𝑛-SuperHyperGraph

A multigraph is a graph in which multiple edges connecting the same pair of vertices are allowed, enabling edge
multiplicities [34–39]. A multihypergraph is a hypergraph variant where hyperedges, each potentially connect-
ing any number of vertices, can appear repeatedly with multiplicities [40–46]. A Multi n-SuperHyperGraph
generalizes hypergraphs by iteratively lifting vertices and edges into n-th powerset hierarchies, enabling super-
vertex and superedge multiplicities [23].
Definition 1.6 (MultiHypergraph). (cf. [40–43]) A multihypergraph is a triple

H = (𝑉, E, 𝜇),

where

2
 • 𝑉 is a finite set of vertices,
• E is a (multi)set of nonempty subsets of 𝑉, called hyperedges,
• 𝜇 : E → N>0 is a multiplicity function, assigning to each hyperedge 𝑒 ∈ E the number of times it appears.

Equivalently, one may regard E itself as a multiset, in which each hyperedge 𝑒 occurs with multiplicity 𝜇(𝑒).
Example 1.7 (MultiHypergraph). Let the vertex set be

𝑉 = {𝑣 1 , 𝑣 2 , 𝑣 3 }.

Define the multiset of hyperedges 

E = 𝑒1 , 𝑒1 , 𝑒2 , 𝑒3 , 𝑒3 , 𝑒3 ,
where
𝑒 1 = {𝑣 1 , 𝑣 2 }, 𝑒 2 = {𝑣 2 , 𝑣 3 }, 𝑒 3 = {𝑣 1 }.
The multiplicity function 𝜇 is given by

𝜇(𝑒 1 ) = 2, 𝜇(𝑒 2 ) = 1, 𝜇(𝑒 3 ) = 3.

Thus the multihypergraph

H = (𝑉, E, 𝜇)
has:

• three vertices 𝑣 1 , 𝑣 2 , 𝑣 3 ,
• one hyperedge {𝑣 1 , 𝑣 2 } appearing twice,
• one hyperedge {𝑣 2 , 𝑣 3 } appearing once,
• one hyperedge {𝑣 1 } appearing three times.
Definition 1.8 (Multi 𝑛-SuperHyperGraph). [23] Let 𝑉0 be a finite base set of vertices. For each integer 𝑘 ≥ 0,
define the iterated powerset
𝑃0 (𝑉0 ) = 𝑉0 , 𝑃 𝑘+1 (𝑉0 ) = P 𝑃 𝑘 (𝑉0 ) ,

where P (·) denotes the standard power-set operator. A Multi 𝑛-SuperHyperGraph is a triple

MSHT (𝑛) = 𝑉, 𝐸, 𝜇 ,

where
𝑉 ⊆ 𝑃 𝑛 (𝑉0 ) , 𝐸 ⊆ 𝑃 𝑛 (𝑉0 )
and
𝜇 : 𝐸 −→ N
is a multiplicity function assigning to each superedge 𝑒 ∈ 𝐸 a positive integer 𝜇(𝑒), indicating that 𝑒 appears
𝜇(𝑒) times.

• Each element of 𝑉 is called an 𝑛-supervertex.

• Each element 𝑒 ∈ 𝐸 is called an 𝑛-superedge, and its multiplicity is 𝜇(𝑒).
Remark 1.9. When 𝜇(𝑒) = 1 for all 𝑒 ∈ 𝐸, MSHT (𝑛) reduces exactly to an ordinary 𝑛-SuperHyperGraph.
Example 1.10 (Multi 2-SuperHyperGraph). Let the base set be

𝑉0 = {𝑎, 𝑏}.

For 𝑛 = 2, we have
𝑃1 (𝑉0 ) = ∅, {𝑎}, {𝑏}, {𝑎, 𝑏} , 𝑃2 (𝑉0 ) = P 𝑃1 (𝑉0 ) ,


which is the collection of all subsets of 𝑃1 (𝑉0 ). We now select:

3
 • The set of 2-supervertices
n o
𝑉 = {{𝑎}}, {{𝑏}, {𝑎, 𝑏}} ⊆ 𝑃2 (𝑉0 ).

• The set of 2-superedges

n o
𝐸 = {{𝑎}, {𝑏}}, {∅, {𝑎, 𝑏}} ⊆ 𝑃2 (𝑉0 ).

• The multiplicity function

𝜇 : 𝐸 −→ N, 𝜇 {{𝑎}, {𝑏}} = 2, 𝜇 {∅, {𝑎, 𝑏}} = 1.

Thus we obtain the Multi 2-SuperHyperGraph

MSHT (2) = 𝑉, 𝐸, 𝜇 ,

in which:

• The supervertex {{𝑎}} represents the singleton subset {𝑎} of 𝑉0 .

• The supervertex {{𝑏}, {𝑎, 𝑏}} encodes the two-element collection {{𝑏}, {𝑎, 𝑏}} ⊆ 𝑃1 (𝑉0 ).
• The superedge {{𝑎}, {𝑏}} appears with multiplicity 2, indicating two parallel occurrences.
• The superedge {∅, {𝑎, 𝑏}} appears once.
Example 1.11 (Multi 3-SuperHyperGraph). Let the base set be

𝑉0 = {𝑎}.

Then
𝑃1 (𝑉0 ) = {∅, {𝑎}}, 𝑃2 (𝑉0 ) = ∅, {∅}, {{𝑎}}, {∅, {𝑎}} ,


and
𝑃3 (𝑉0 ) = P 𝑃2 (𝑉0 ) .

Now select two 3-supervertices:

𝑈1 = {∅, {∅}}, 𝑈2 = {{{𝑎}}, {∅, {𝑎}}},

so that
𝑉 = {𝑈1 , 𝑈2 } ⊆ 𝑃3 (𝑉0 ).
Next, choose two 3-superedges:

𝐸 1 = {{∅}, {{𝑎}}}, 𝐸 2 = {∅, {∅, {𝑎}}},

and set
𝐸 = {𝐸 1 , 𝐸 2 } ⊆ 𝑃3 (𝑉0 ).
Define the multiplicity function
𝜇(𝐸 1 ) = 2, 𝜇(𝐸 2 ) = 4.
Then
MSHT (3) = (𝑉, 𝐸, 𝜇)
is a Multi 3-SuperHyperGraph in which:

• 𝑈1 and 𝑈2 are two distinct 3-supervertices,

• 𝐸 1 occurs twice, 𝐸 2 occurs four times,
• all vertices and edges lie in the third iterated powerset of 𝑉0 .

4
1.3 MultiHypergraph Neural Network

Graph Neural Networks and Hypergraph Neural Networks have been the subject of extensive research across
a multitude of publications [11, 47–56]. A Multigraph Neural Network processes multiple graph instances
via parallel graph convolutional layers, then aggregates their vertex embeddings into a unified representation
[57–62]. A MultiHypergraph Neural Network generalizes this by applying hypergraph convolution to several
hypergraph structures in parallel, integrating both hyperedge and vertex features into a combined embedding
[43, 63–67].
 𝑀
Definition 1.12 (MultiHypergraph Neural Network). (cf. [67]) Let 𝐻𝑚 = (𝑉, E 𝑚 ) 𝑚=1 be a collection of 𝑀
hypergraphs over the same vertex set 𝑉. Denote by

𝐻𝑚 ∈ {0, 1} |𝑉 | × | E𝑚 | the incidence matrix of 𝐻𝑚 ,

and let 𝑋 ∈ R |𝑉 | ×𝐹 be the matrix of input vertex features. Define for each 𝑚 the hypergraph Laplacian
1 1
−2 −1 ⊤ −2
e𝑚 = 𝐷 𝑣,𝑚
𝐻 𝐻𝑚 𝐷 𝑒,𝑚 𝐻𝑚 𝐷 𝑣,𝑚 , (1)

where 𝐷 𝑣,𝑚 and 𝐷 𝑒,𝑚 are the diagonal degree matrices of vertices and hyperedges respectively. A MultiHy-
pergraph Neural Network with 𝐿 layers is defined by the layerwise propagation
(ℓ+1) (ℓ )
𝑊𝑚(ℓ ) , (0)


𝑍𝑚 =𝜎 𝐻
e𝑚 𝑍 𝑚 𝑍𝑚 = 𝑋, (2)

for ℓ = 0, 1, . . . , 𝐿 − 1, where each 𝑊𝑚(ℓ ) is a learnable weight matrix and 𝜎 an activation function. Finally, the
outputs from all hypergraphs are fused by

𝑍 = AGG 𝑍1(𝐿) , 𝑍2(𝐿) , . . . , 𝑍 𝑀

(𝐿)

, (3)

where AGG is an aggregation operator (e.g. average or concatenation). This architecture processes multiple
hypergraph structures in parallel and integrates their learned representations.
Example 1.13 (MultiHypergraph Neural Network). Let the base vertex set be

𝑉 = {𝑣 1 , 𝑣 2 , 𝑣 3 }, 𝑀 = 2.

We define two hypergraphs on 𝑉:

𝐻1 : E1 = {{𝑣 1 , 𝑣 2 }, {𝑣 2 , 𝑣 3 }}, 𝐻2 : E2 = {{𝑣 1 , 𝑣 3 }, {𝑣 2 }}.

Their incidence matrices (rows 𝑣 1 , 𝑣 2 , 𝑣 3 ; columns ordered as the hyperedges above) are

1 0 1 0
𝐻 1 = ­1 1® , 𝐻2 = ­0 1® .
© ª © ª

«0 1¬ «1 0¬

Compute the degree matrices:

𝐷 𝑣,1 = diag(1, 2, 1), 𝐷 𝑒,1 = diag(2, 2), 𝐷 𝑣,2 = diag(2, 1, 1), 𝐷 𝑒,2 = diag(2, 1).

The normalized Laplacians (cf. (1)) are

1 1
0.5000 0.3536 0
e1 = 𝐷 − 2 𝐻1 𝐷 −1 𝐻 ⊤ 𝐷 − 2 ≈ ©­0.3536 0.5000 0.3536ª® ,
𝐻 𝑣,1 𝑒,1 1 𝑣,1
« 0 0.3536 0.5000¬

1 1
0.5000 0 0.5000
e2 = 𝐷 − 2 𝐻2 𝐷 −1 𝐻 ⊤ 𝐷 − 2 = ©­ 0
𝐻 1.0000 0 ®.
ª
𝑣,2 𝑒,2 2 𝑣,2
«0.5000 0 0.5000¬

5
Choose input features
1 0
𝑋 = ­0 1® ,
© ª

«1 1¬
use weight matrices 𝑊𝑚(0) = 𝐼, and the identity activation 𝜎(𝑥) = 𝑥. Then one propagation layer (2) yields
0.5000 0.3536 1.0000 0.5000
𝑍1(1) = 𝐻
e1 𝑋 ≈ ©­0.7071 0.8536ª® , 𝑍2(1) = 𝐻
e2 𝑋 = ©­0.0000 1.0000ª® .
«0.5000 0.8536¬ «1.0000 0.5000¬

Finally, fuse the two outputs by averaging (3):

0.7500 0.4268
1
𝑍1(1) + 𝑍2(1)


𝑍= ≈ ­0.3536 0.9268® .
© ª
2
«0.7500 0.6768¬
This example illustrates a concrete forward pass of a MultiHypergraph Neural Network with two hypergraph
structures.

1.4 Undirected n-SuperHyperGraph Neural Network (n-SHGNN)

The definition of the Undirected n-SuperHyperGraph Neural Network (n-SHGNN) is presented as follows [12].
Definition 1.14 (n-SuperHyperGraph Neural Network (n-SHGNN)). [12] Let 𝐻 (𝑛) = 𝑉 (𝑛) , 𝐸 (𝑛) be an


𝑛-SuperHyperGraph over a base vertex set 𝑉0 , and let

𝐻 ′ = 𝑉0 , 𝐸 ′

be its Expanded Hypergraph, where

Ø
𝐸 ′ = 𝑒 ′ ⊆ 𝑉0 𝑒 ′ = 𝑒 ∈ 𝐸 (𝑛) .

𝑣,
𝑣 ∈𝑒

Let
𝑋 ∈ R |𝑉0 | ×𝑑
be the input feature matrix whose 𝑖-th row 𝑥 𝑖 ∈ R𝑑 is the feature vector of base vertex 𝑣 𝑖 ∈ 𝑉0 . Define:

′
• The incidence matrix 𝐻 ′ ∈ {0, 1} |𝑉0 | × | 𝐸 | with entries
(
′ 1, 𝑣 𝑖 ∈ 𝑒 ′𝑗 ,
𝐻𝑖 𝑗 =
0, otherwise.

′ | × |𝐸 ′ |
• The diagonal vertex-degree matrix 𝐷 𝑉 ∈ R |𝑉0 | × |𝑉0 | and hyperedge-degree matrix 𝐷 𝐸 ∈ R | 𝐸
defined by
|𝐸 ′ |
∑︁ |𝑉0 |
∑︁
(𝐷 𝑉 )𝑖𝑖 = 𝐻𝑖′ 𝑗 𝑤(𝑒 ′𝑗 ), (𝐷 𝐸 ) 𝑗 𝑗 = 𝐻𝑖′ 𝑗 ,
𝑗=1 𝑖=1

where 𝑤(𝑒 ′𝑗 ) > 0 is a learnable weight for hyperedge 𝑒 ′𝑗 ∈ 𝐸 ′.

• A learnable hyperedge-weight matrix
′ | × | 𝐸′ |
𝑊 ∈ R| 𝐸 , Θ ∈ R𝑑×𝑐 ,

and a non-linear activation 𝜎(·) (e.g. ReLU).

Then one layer of the 𝑛-SHGNN is given by the convolution

𝑌 = 𝜎 𝐷 𝑉−1/2 𝐻 ′ 𝑊 𝐷 𝐸
−1 ′⊤ −1/2

𝐻 𝐷𝑉 𝑋 Θ ,

where 𝑌 ∈ R |𝑉0 | ×𝑐 is the updated feature matrix.

6
Example 1.15 (Concrete Undirected 2-SuperHyperGraph Neural Network). Let the base vertex set be

𝑉0 = {1, 2, 3}, 𝑛 = 2.

Define an 𝑛-SuperHyperGraph
𝐻 (2) = 𝑉 (2) , 𝐸 (2)

by choosing
𝑉 (2) = {1, 2}, {2, 3} , 𝐸 (2) = 𝑒 1 = {{1, 2}}, 𝑒 2 = {{2, 3}} .
 

Its expanded hypergraph is

𝐻 ′ = 𝑉0 , 𝐸 ′ , 𝐸 ′ = { {1, 2}, {2, 3}}.

The incidence matrix 𝐻 ′ ∈ {0, 1}3×2 (rows 1, 2, 3; cols 𝑒 1′ , 𝑒 2′ ) is

1 0
𝐻 ′ = ­1 1® .
© ª

«0 1¬

Assign learnable hyperedge weights

𝑤(𝑒 1′ ) = 1, 𝑤(𝑒 2′ ) = 2,
so that
𝐷 𝑉 = diag 𝐻 ′ 𝑤(𝐸 ′ ) = diag(1, 3, 2), 𝐷 𝐸 = diag 𝐻 ′⊤ 1 = diag(2, 2),

and form
𝑊 = diag(1, 2).
Let the input feature vector be
1
𝑋 = ­2® ,
© ª

«3¬
choose a single output channel (Θ = 1) and identity activation 𝜎(𝑥) = 𝑥. Then one layer of the 2-SHGNN
computes
1.0774
−1 −1 ′⊤ − 2
1
𝑌 = 𝐷𝑉 2 𝐻′ 𝑊 𝐷 𝐸 𝐻 𝐷 𝑉 𝑋 ≈ ­2.5133® .
© ª

«2.3165¬
Thus each base-vertex’s new feature is a weighted, normalized aggregation of its neighbors according to the
2-SuperHyperGraph structure.

2 Results: Multi-SuperHyperGraph Neural Networks

The definition of the Multi-SuperHyperGraph Neural Network is presented as follows. This concept extends
the 𝑛-SuperHyperGraph Neural Network by employing the Multi-SuperHyperGraph framework. As it is purely
a theoretical extension, we anticipate future work to include empirical experiments on real datasets to validate
its effectiveness.
(𝑛)
Definition 2.1 (Multi 𝑛-SuperHyperGraph Neural Network). Let {MSHT𝑚 = (𝑉𝑚(𝑛) , 𝐸 𝑚
(𝑛)
, 𝜇 𝑚 )} 𝑚=1
𝑀 be a
collection of 𝑀 Multi 𝑛-SuperHyperGraphs over the same base set 𝑉0 . For each 𝑚, define its expanded
hypergraph nØ o
′ ′
′ (𝑛)
𝐻𝑚 = 𝑉0 , 𝐸 𝑚 , 𝐸𝑚 = 𝑣 𝑒 ∈ 𝐸𝑚 .
𝑣 ∈𝑒
Let
′ ′
𝐻𝑚 ∈ {0, 1} |𝑉0 | × | 𝐸𝑚 | be its incidence matrix,
and set
′
′⊤


𝐷 𝑉 ,𝑚 = diag 𝐻𝑚 𝜇𝑚 1 , 𝐷 𝐸,𝑚 = diag 𝐻𝑚 1, 𝑊𝑚 = diag 𝜇 𝑚 (𝑒) ′ .
𝑒∈𝐸𝑚

Given input features 𝑋 ∈ R |𝑉0 | ×𝐹 , a Multi 𝑛-SuperHyperGraph Neural Network with 𝐿 layers computes for
each 𝑚 and ℓ = 0, . . . , 𝐿 − 1:
 −1 1
2 𝐻 ′ 𝑊 𝐷 −1 𝐻 ′ ⊤ 𝐷 − 2 𝑍 (ℓ ) Θ (ℓ ) ,

(ℓ+1) (0)
𝑍𝑚 = 𝜎 𝐷 𝑉 ,𝑚 𝑚 𝑚 𝐸,𝑚 𝑚 𝑉 ,𝑚 𝑚 𝑚 𝑍𝑚 = 𝑋,

7
 (ℓ )
where each Θ𝑚 is a learnable weight matrix and 𝜎 an activation. Finally, the per-graph outputs are fused:

𝑍 = AGG 𝑍1(𝐿) , 𝑍2(𝐿) , . . . , 𝑍 𝑀

(𝐿)

,

with AGG an aggregation operator (e.g. mean or concat).

Example 2.2 (Concrete Multi 1-SuperHyperGraph Neural Network). Let the base set be

𝑉0 = {1, 2, 3}, 𝑛 = 1, 𝑀 = 2.

We define two Multi 1-SuperHyperGraphs over 𝑉0 :

MSHT1(1) : 𝑉1(1) = {1}, {2, 3} , 𝐸 1(1) = {1, 2}, {2, 3} ,

 
𝜇1 ({1, 2}) = 1, 𝜇1 ({2, 3}) = 2.

MSHT2(1) : 𝑉2(1) = {2}, {3} , 𝐸 2(1) = {1, 2, 3} ,

 
𝜇2 ({1, 2, 3}) = 1.

Their expanded hypergraphs are

𝐻1′ = 𝑉0 , 𝐸 1′ , 𝐸 1′ = 𝐸 1(1) , 𝐻2′ = 𝑉0 , 𝐸 2′ , 𝐸 2′ = 𝐸 2(1) .

Thus the incidence matrices (rows indexed by 1, 2, 3, columns by hyperedges) are

1 0 1
𝐻1′ = ­1 1® , 𝐻2′ = ­1® .
© ª © ª

«0 1¬ «1¬
The corresponding multiplicity diagonal matrices are

𝑊1 = diag(1, 2), 𝑊2 = diag(1).

Compute the degree matrices:

𝐷 𝑉 ,1 = diag(1, 3, 2), 𝐷 𝐸,1 = diag(2, 2), 𝐷 𝑉 ,2 = diag(1, 1, 1), 𝐷 𝐸,2 = diag(3).

Now choose input features and weights:

1 0
𝑋 = ­0 1® , Θ1(0) = Θ2(0) = 𝐼2 , 𝜎 = identity.
© ª

«1 1¬

One graph-convolution layer yields

0.5000 0.2887
− 12 − 21
𝑍1(1) = 𝐷 𝑉 ,1 −1
𝐻1′ 𝑊1 𝐷 𝐸,1 𝐻1′⊤ 𝐷 𝑉 ,1 𝑋 ≈ ­0.6969 0.9082® ,
© ª

«0.5000 0.9082¬
2 2
1 1 1
− 21 − 21 1© © 23 3ª
𝑍2(1) = 𝐻2′ 𝑊2 −1
𝐻2′⊤ 𝑋 = ­1 1 1® 𝑋 = ­ 3 2
ª
𝐷 𝑉 ,2 𝐷 𝐸,2 𝐷 𝑉 ,2 3® .
3 1 1 1¬ 2 2
« «3 3¬
Finally, fuse by averaging:
0.5833 0.4777
𝑍1(1) + 𝑍2(1)
𝑍= ≈ ­0.6818 0.7875® .
© ª
2
«0.5833 0.7875¬
This illustrates a concrete forward pass of a Multi 1-SuperHyperGraph Neural Network.

8
Example 2.3 (Concrete Multi 1-SuperHyperGraph Neural Network with 𝑀 = 3). Let the base set be

𝑉0 = {1, 2, 3}, 𝑛 = 1, 𝑀 = 3.

Define three Multi 1-SuperHyperGraphs over 𝑉0 :

MSHT1(1) : 𝑉1(1) = {1}, {2}, {3} , 𝐸 1(1) = {1, 2}, {2, 3} ,

 
𝜇1 ({1, 2}) = 2, 𝜇1 ({2, 3}) = 1.

MSHT2(1) : 𝑉2(1) = {1}, {2}, {3} , 𝐸 2(1) = {1, 2}, {1, 3}, {2, 3} , 𝜇2 (𝑒) = 1 ∀ 𝑒 ∈ 𝐸 2(1) .
 

MSHT3(1) : 𝑉3(1) = {1}, {2}, {3} , 𝐸 3(1) = {1}, {2, 3} ,

 
𝜇3 ({1}) = 1, 𝜇3 ({2, 3}) = 2.

(1)
Their expanded hyperedges coincide with 𝐸 𝑚 , so the incidence matrices (rows indexed by 1, 2, 3) are

1 0 1 1 0 1 0
𝐻1′ = ­1 1® , 𝐻2′ = ­1 0 1® , 𝐻3′ = ­0 1® .
© ª © ª © ª

«0 1¬ «0 1 1¬ «0 1¬

The multiplicity diagonal matrices are

𝑊1 = diag(2, 1), 𝑊2 = 𝐼3 , 𝑊3 = diag(1, 2).

Compute
′
′⊤

𝐷 𝑉 ,𝑚 = diag 𝐻𝑚 𝑊𝑚 1 , 𝐷 𝐸,𝑚 = diag 𝐻𝑚 1,
and let the input feature matrix and weights be

1 0
(0)
𝑋 = ­0 1® , Θ𝑚 = 𝐼2 , 𝜎(𝑥) = 𝑥.
© ª

«1 1¬

One layer of graph convolution yields

0.5000 0.4082 0.7500 0.5000

𝑍1(1) ≈ ­0.6969 0.7887® , 𝑍2(1) ≈ ­0.5000 0.7500® ,
© ª © ª

«0.5000 0.7887¬ «0.7500 0.7500¬

1.0000 0.0000
𝑍3(1) = ­0.5000 1.0000® .
© ª

«0.5000 1.0000¬
Finally, fuse by element-wise averaging:

𝑍1(1) + 𝑍2(1) + 𝑍3(1) 0.7500 0.3027

𝑍= ≈ ­0.5656 0.8462® .
© ª
3
«0.5833 0.8462¬
This demonstrates a detailed forward pass of a Multi 1-SuperHyperGraph Neural Network with three distinct
superhypergraph structures.
Example 2.4 (Concrete Multi 2-SuperHyperGraph Neural Network). Let the base set be

𝑉0 = {1, 2}, 𝑛 = 2, 𝑀 = 2.

We define two Multi 2-SuperHyperGraphs:

𝑉 (2) = {{1}}, {{2}, {1, 2}} ,

 
 1(2) 


MSHT1(2)

: 𝐸 1 = {{1}, {2}}, {∅, {1, 2}} ,

 𝜇1 ({{1}, {2}}) = 1, 𝜇1 ({∅, {1, 2}}) = 2,


9
 𝑉 (2) = {{1}}, {{2}} ,
 
 2(2) 


MSHT2(2)

: 𝐸 2 = {{1}}, {{2}} ,

 𝜇2 ({{1}}) = 1, 𝜇2 ({{2}}) = 1.



Their expanded hypergraphs are

𝐸 1′ = {1, 2}, {1, 2} , 𝐸 2′ = {1}, {2} .

 

Thus the incidence matrices (rows indexed by 1,2; columns by hyperedges) are
   
1 1 1 0
𝐻1′ = , 𝐻2′ = .
1 1 0 1

The multiplicity diagonal matrices are

𝑊1 = diag(1, 2), 𝑊2 = diag(1, 1).

Compute degree matrices:

𝐷 𝐸,1 = diag(2, 2), 𝐷 𝑉 ,1 = diag(3, 3), 𝐷 𝐸,2 = diag(1, 1), 𝐷 𝑉 ,2 = diag(1, 1).

Choose input features and weights:

 
1
𝑋= , Θ1(0) = Θ2(0) = (1), 𝜎(𝑥) = 𝑥.
2

One-layer propagation gives

 
− 21 − 12 1.5
𝑍1(1) = 𝐷 𝑉 ,1 𝐻1′ 𝑊1 −1
𝐷 𝐸,1 𝐻1′⊤ 𝐷 𝑉 ,1 𝑋= ,
1.5
 
− 21 − 12 1
𝑍2(1) = 𝐷 𝑉 ,2 𝐻2′ 𝑊2 −1
𝐷 𝐸,2 𝐻2′⊤ 𝐷 𝑉 ,2 𝑋= .
2
Finally, fuse by averaging:
𝑍1(1) + 𝑍2(1)

1.25
𝑍= = .
2 1.75
This concrete example illustrates feature propagation and fusion in a Multi 2-SuperHyperGraph Neural Network.
Example 2.5 (Concrete Multi 3-SuperHyperGraph Neural Network). Let the base set be

𝑉0 = {1, 2}, 𝑛 = 3, 𝑀 = 2.

We define two Multi 3-SuperHyperGraphs over 𝑉0 :

𝑉1(3) = 𝑝 1 = {{1}}, 𝑝 2 = {{2}}, 𝑝 3 = {∅} ,

 


MSHT1(3) : 𝐸 1(3) = 𝑒 11 = {𝑝 1 , 𝑝 2 }, 𝑒 12 = {𝑝 1 , 𝑝 3 } ,

 

 𝜇1 (𝑒 11 ) = 1, 𝜇1 (𝑒 12 ) = 2,



𝑉 (3) = 𝑞 = {{1, 2}}, 𝑞 2 = {{2}}, 𝑞 3 = {∅} ,

 
 2(3)  1


MSHT2(3)

: 𝐸 2 = 𝑒 21 = {𝑞 1 , 𝑞 3 }, 𝑒 22 = {𝑞 2 , 𝑞 3 } ,

 𝜇2 (𝑒 21 ) = 3, 𝜇2 (𝑒 22 ) = 1.



Their expanded hyperedges (by one-level union) are

𝐸 1′ = {1, 2}, {1} , 𝐸 2′ = {1, 2}, {2} .

 

10
Hence the incidence matrices (rows indexed by 1, 2; columns by each hyperedge) are
   
′ 1 1 ′ 1 0
𝐻1 = , 𝐻2 = .
1 0 1 1

The multiplicity diagonal matrices are

𝑊1 = diag(1, 2), 𝑊2 = diag(3, 1).

Compute the degree matrices:

𝐷 𝐸,1 = diag(2, 1), 𝐷 𝑉 ,1 = diag(3, 1), 𝐷 𝐸,2 = diag(2, 1), 𝐷 𝑉 ,2 = diag(3, 4).

Choose the input feature matrix and weights:

 
1 0 (0)
𝑋= , Θ𝑚 = 𝐼2 , 𝜎(𝑥) = 𝑥.
0 1

Then one graph-convolution layer yields

1 1  
−2 ′ −2 0.8333 0.2887
𝑍1(1) = 𝐷 𝑉 ,1 𝐻1 𝑊1 −1
𝐷 𝐸,1 𝐻1′⊤ 𝐷 𝑉 ,1 𝑋 ≈ ,
0.2887 0.5000
1 1  
−2 ′ −2 0.5000 0.4330
𝑍2(1) = 𝐷 𝑉 ,2 𝐻 2 𝑊2 −1
𝐷 𝐸,2 𝐻2′⊤ 𝐷 𝑉 ,2 𝑋 ≈ .
0.4330 0.6250
Finally, fuse by averaging:
𝑍1(1) + 𝑍2(1)
 
0.6667 0.3608
𝑍= ≈ .
2 0.3608 0.5625
This detailed example demonstrates the forward pass of a Multi 3-SuperHyperGraph Neural Network.
Theorem 2.6. The Multi 𝑛-SuperHyperGraph Neural Network generalizes both

1. the MultiHypergraph Neural Network (when 𝑛 = 0), and

2. the 𝑛-SuperHyperGraph Neural Network (when 𝑀 = 1).

(0)
Proof. 1. Case 𝑛 = 0. Then 𝑃0 (𝑉0 ) = 𝑉0 , so each MSHT𝑚 = (𝑉𝑚(0) , 𝐸 𝑚
(0)
, 𝜇 𝑚 ) satisfies 𝑉𝑚(0) ⊆ 𝑉0 , 𝐸 𝑚
(0)
⊆ 𝑉0 .
Ð (0) (0)
The expanded hyperedge set { 𝑣 ∈𝑒 𝑣 | 𝑒 ∈ 𝐸 𝑚 } coincides with 𝐸 𝑚 . Hence the incidence, degree and weight
matrices match those of a MultiHypergraph Neural Network, and the layer propagation and fusion reduce
exactly to that model.

2. Case 𝑀 = 1. With a single graph, the final aggregation AGG(𝑍1(𝐿) ) is the identity, and the update rule
coincides with the definition of the 𝑛-SuperHyperGraph Neural Network. Thus the architecture reduces to the
𝑛-SHGNN. □
Theorem 2.7. Each layer of the Multi 𝑛-SuperHyperGraph Neural Network faithfully encodes the Multi
𝑛-SuperHyperGraph structure, including supervertex connectivity and superedge multiplicity.

Proof. Fix 𝑚. By construction, the incidence matrix 𝐻𝑚 ′ encodes which base vertices belong to each expanded
(𝑛)
hyperedge derived from 𝐸 𝑚 . The diagonal weight matrix 𝑊𝑚 = diag(𝜇 𝑚 (𝑒)) injects the superedge multiplicity
into the convolution. Moreover, 𝐷 𝐸,𝑚 = diag(𝐻𝑚 ′ ⊤ 1) and 𝐷 ′
𝑉 ,𝑚 = diag(𝐻 𝑚 𝑊𝑚 1) correctly normalize by
hyperedge sizes and weighted vertex degrees. Therefore each graph convolution layer
−1
2 ′ −1 ′⊤ 2 −1
𝐷 𝑉 ,𝑚 𝐻𝑚 𝑊𝑚 𝐷 𝐸,𝑚 𝐻𝑚 𝐷 𝑉 ,𝑚

operates precisely on the Multi 𝑛-SuperHyperGraph (𝑉𝑚(𝑛) , 𝐸 𝑚(𝑛)

, 𝜇 𝑚 ), preserving both its higher-order connec-
tivity and multiplicity structure in the feature propagation. □

11
 (𝑛) 𝑀
Theorem 2.8 (Permutation Equivariance). Let {MSHT𝑚 } 𝑚=1 be Multi 𝑛-SuperHyperGraphs over the same
|𝑉 | × |𝑉 | (𝑛) Π
base set 𝑉0 , and let Π ∈ {0, 1} 0 0 be any permutation matrix on 𝑉0 . Denote by MSHT𝑚 the graph
obtained by permuting the labels of 𝑉0 by Π. Then for any layer index ℓ,
(ℓ ) (𝑛)
(ℓ ) (𝑛) Π

𝑍𝑚 = 𝑓ℓ MSHT𝑚 ,𝑋 =⇒ Π 𝑍 𝑚 = 𝑓ℓ MSHT𝑚 , Π𝑋 ,
i.e. the layer outputs commute with any permutation of the base vertices.

Proof. Fix 𝑚 and ℓ. Let 𝐻𝑚 ′ be the incidence matrix and 𝐷

𝑉 ,𝑚 , 𝐷 𝐸,𝑚 , 𝑊𝑚 the diagonal degree and weight
matrices defined in the network. Under a permutation Π, the incidence transforms as
′ ′
𝐻𝑚 ↦−→ Π 𝐻𝑚 ,
while 𝐷 𝐸,𝑚 and 𝑊𝑚 remain unchanged (since they depend only on hyperedges), and
′ ′
𝐷 𝑉 ,𝑚 = diag(𝐻𝑚 𝑊𝑚 1) ↦−→ diag(Π 𝐻𝑚 𝑊𝑚 1) = Π 𝐷 𝑉 ,𝑚 Π ⊤ .
Hence the normalized convolution operator
−1
2 ′ −1 ′⊤ 2 −1
L 𝑚 = 𝐷 𝑉 ,𝑚 𝐻𝑚 𝑊𝑚 𝐷 𝐸,𝑚 𝐻𝑚 𝐷 𝑉 ,𝑚
satisfies
−1 ′
⊤ − 21
Π⊤ ′
−1
Π⊤ = Π L 𝑚 Π ⊤ .
2



L 𝑚 ↦−→ Π 𝐷 𝑉 ,𝑚 Π 𝐻𝑚 𝑊𝑚 𝐷 𝐸,𝑚 Π 𝐻𝑚 Π 𝐷 𝑉 ,𝑚
(ℓ+1) (ℓ )
Therefore, if 𝑍 𝑚 = 𝜎(L 𝑚 𝑍 𝑚 Θ), then
(ℓ+1) (ℓ )
(ℓ )
Θ = 𝜎 (Π L 𝑚 Π ⊤ ) (Π 𝑍 𝑚


Π 𝑍𝑚 = 𝜎 Π L𝑚 𝑍𝑚 )Θ ,
showing equivariance at layer ℓ + 1. By induction from 𝑍 (0) = 𝑋, the claim follows. □
Theorem 2.9 (Lipschitz Stability). Assume the activation 𝜎 is 1-Lipschitz (e.g. ReLU) and that for each 𝑚 the
(ℓ ) (ℓ )
weight matrices Θ𝑚 satisfy ∥Θ𝑚 ∥ 2 ≤ 𝐵. Then each layer mapping
𝑍 (ℓ ) ↦→ 𝑍 (ℓ+1)
of the Multi 𝑛-SuperHyperGraph Neural Network is Lipschitz continuous with constant 𝐿 = 𝐵 max L 𝑚 2 ,
𝑚
− 21 1
′ 𝑊 𝐷 −1 𝐻 ′ ⊤ 𝐷 − 2 .
where L 𝑚 = 𝐷 𝑉 ,𝑚 𝐻𝑚 𝑚 𝐸,𝑚 𝑚 𝑉 ,𝑚

Proof. Write the layer update for all 𝑚 concatenated as

𝑍 ↦→ 𝜎 L 𝑍 Θ ,
(ℓ )
where L is block-diagonal with blocks L 𝑚 , and Θ is block-diagonal with Θ𝑚 . Then for any two inputs 𝑍, 𝑍 ′ ,
𝑍 (ℓ+1) − 𝑍 ′(ℓ+1) = 𝜎(L𝑍Θ) − 𝜎(L𝑍 ′ Θ) L (𝑍 − 𝑍 ′ )Θ
≤
′
≤ max ∥L 𝑚 ∥ 2 ∥𝑍 − 𝑍 ′ ∥ 𝐵.


≤ ∥L ∥ 2 ∥𝑍 − 𝑍 ∥ ∥Θ∥ 2
𝑚

Thus the mapping is Lipschitz with constant 𝐿 = 𝐵 max𝑚 ∥L 𝑚 ∥ 2 . □

Theorem 2.10 (Universality on Fixed-Size Graphs). Let F be any continuous, permutation-equivariant func-
tion on the space of feature-labelled Multi 𝑛-SuperHyperGraphs with base vertex set 𝑉0 of size 𝑁. Then for
any 𝜀 > 0, there exists a Multi 𝑛-SuperHyperGraph Neural Network with sufficiently many layers and hidden
units that approximates F uniformly within 𝜀.

Proof. 1. Reduction to MLP: By Theorem 1, the network is permutation-equivariant. Using the standard
“sum-aggregation” readout, one can reduce the graph-structured input to a multiset of per-vertex embeddings.

2. Universal approximation on multisets: It is known that any continuous permutation-invariant function of a

multiset can be approximated arbitrarily well by a sum-decomposition of MLPs.

3. Combining both: Since each layer of our network implements a learnable permutation-equivariant map
followed by a permutation-invariant fusion across the 𝑀 graphs, stacking sufficiently many layers and choosing
wide enough MLPs in the final fusion yields an approximation of F within 𝜀.

Hence the architecture is universal on the fixed finite domain of size 𝑁. □

12
 (𝑛) 𝑀 (𝑛) 𝑀
Theorem 2.11 (Robustness to Hyperedge Perturbations). Let {MSHT𝑚 } 𝑚=1 and {MSHT
›𝑚 } 𝑚=1 be two
collections of Multi 𝑛-SuperHyperGraphs over the same base set 𝑉0 , differing only in at most 𝐾 superedges
(additions or deletions) per graph. Denote by L 𝑚 and L b𝑚 the corresponding normalized convolution operators,
and let 𝑍 (ℓ ) and 𝑍
b(ℓ ) be the layer-ℓ feature matrices (with identical initial features 𝑍 (0) = 𝑍
b(0) ). If all weight
(ℓ )
matrices satisfy ∥Θ𝑚 ∥ 2 ≤ 𝐵 and 𝜎 is 1-Lipschitz, then after one layer:
𝑀
(1) b(1)
∑︁ ∥ 𝑋 ∥𝐹
𝑍 −𝑍 𝐹
≤ 𝐵 L𝑚 − L
b𝑚
2
∥ 𝑋 ∥ 𝐹 ≤ 2𝐵 𝑀 𝐾 ,
𝑚=1
𝑑min

where 𝑑min is the minimum (weighted) vertex degree across both collections.

Proof. Since 𝜎 is 1-Lipschitz,

∥𝑍 (1) − 𝑍
b(1) ∥ 𝐹 =


𝜎(L 𝑋Θ) − 𝜎( L
b𝑋Θ)
𝐹
≤ ∥ L−L
b 𝑋 Θ∥ 𝐹 .

(0)
Here L and L
b are block-diagonal with blocks L 𝑚 , L
b𝑚 , and Θ is block-diagonal with Θ𝑚 . Hence
𝑀
∑︁


∥ L − L 𝑋 Θ∥ 𝐹 ≤ 𝐵
b ∥L 𝑚 − L
b𝑚 ∥ 2 ∥ 𝑋 ∥ 𝐹 .
𝑚=1

2𝐾
Each perturbed superedge changes one row and one column of the incidence matrix, hence ∥L 𝑚 − L
b𝑚 ∥ 2 ≤
𝑑min .
Summing over 𝑚 yields the result. □

Theorem 2.12 (Spectral Filter Approximation). Let 𝑔 : [0, 1] → R be any continuous function on the spectrum
of each L 𝑚 . Then for any 𝜀 > 0 and each 𝑚, there exists a polynomial 𝑝 𝑚 of degree 𝑇 such that

max 𝑝 𝑚 (𝜆) − 𝑔(𝜆) < 𝜀.

𝜆∈Spec( L 𝑚 )

(ℓ )
Consequently, a Multi 𝑛-SuperHyperGraph Neural Network with 𝑇 layers and shared weights Θ𝑚 can ap-
proximate the spectral filter 𝑔(L 𝑚 ) arbitrarily well on all graphs.

Proof. By the Stone–Weierstrass theorem, on the compact interval containing Spec(L 𝑚 ) ⊂ [0, 1], there exists
a polynomial 𝑝 𝑚 of sufficiently large degree 𝑇 such that sup𝜆 | 𝑝 𝑚 (𝜆) − 𝑔(𝜆)| < 𝜀. In a GNN framework,
stacking 𝑇 linear layers with appropriate weight matrices allows implementing any polynomial filter in L 𝑚 up
to order 𝑇; non-linearities can be absorbed or interleaved while preserving polynomial expressivity. Hence the
network can approximate 𝑔(L 𝑚 ) within 𝜀. □
(ℓ )
Theorem 2.13 (Injectivity of Layer Maps). Assume each weight matrix Θ𝑚 ∈ R𝐹 ×𝐹 is invertible and the
activation 𝜎 : R → R is strictly monotonic and applied elementwise. Then each layer map

Φℓ : 𝑍 (ℓ ) ↦−→ 𝑍 (ℓ+1) = 𝜎 L 𝑍 (ℓ ) Θ (ℓ )

is injective on the space of feature matrices.

Proof. Write Φℓ (𝑍) = 𝜎( 𝐴𝑍Θ) with 𝐴 = L. If Φℓ (𝑍) = Φℓ (𝑍 ′ ), then 𝜎( 𝐴𝑍Θ) = 𝜎( 𝐴𝑍 ′ Θ). Since 𝜎 is
strictly monotonic and applied entrywise, we have

𝐴𝑍Θ = 𝐴𝑍 ′ Θ.

Post-multiplying by Θ−1 and using invertibility of 𝐴 (since it is positive-definite when all weights are positive)
gives 𝑍 = 𝑍 ′ . Thus Φℓ is injective. □

Funding

This study did not receive any financial or external support from organizations or individuals.

13
Acknowledgments

We extend our sincere gratitude to everyone who provided insights, inspiration, and assistance throughout this
research. We particularly thank our readers for their interest and acknowledge the authors of the cited works
for laying the foundation that made our study possible. We also appreciate the support from individuals and
institutions that provided the resources and infrastructure needed to produce and share this paper. Finally, we
are grateful to all those who supported us in various ways during this project.

Author Contributions

The paper has been solely authored by the corresponding author at this stage.

Data Availability

This research is purely theoretical, involving no data collection or analysis. We encourage future researchers
to pursue empirical investigations to further develop and validate the concepts introduced here.

Ethical Considerations

This work does not involve any experiments or studies involving human participants or animals, and therefore
no ethical approvals were required.

Conflicts of Interest

The authors confirm that there are no conflicts of interest related to the research or its publication.

Research Integrity

The authors hereby confirm that, to the best of their knowledge, this manuscript is their original work, has not
been published in any other journal, and is not currently under consideration for publication elsewhere at this
stage.

Disclaimer (Note on Computational Tools)

No computer-assisted proof, symbolic computation, or automated theorem proving tools (e.g., Mathematica,
SageMath, Coq, etc.) were used in the development or verification of the results presented in this paper. All
proofs and derivations were carried out manually and analytically by the authors.

Disclaimer (Limitations and Claims)

The theoretical concepts presented in this paper have not yet been subject to practical implementation or
empirical validation. Future researchers are invited to explore these ideas in applied or experimental settings.
Although every effort has been made to ensure the accuracy of the content and the proper citation of sources,
unintentional errors or omissions may persist. Readers should independently verify any referenced materials.

To the best of the authors’ knowledge, all mathematical statements and proofs contained herein are correct and
have been thoroughly vetted. Should you identify any potential errors or ambiguities, please feel free to contact
the authors for clarification.

The results presented are valid only under the specific assumptions and conditions detailed in the manuscript.
Extending these findings to broader mathematical structures may require additional research. The opinions
and conclusions expressed in this work are those of the authors alone and do not necessarily reflect the official
positions of their affiliated institutions.

14
References
[1] Reinhard Diestel. Graph theory 3rd ed. Graduate texts in mathematics, 173(33):12, 2005.
[2] Reinhard Diestel. Graduate texts in mathematics: Graph theory.
[3] Jonathan L Gross, Jay Yellen, and Mark Anderson. Graph theory and its applications. Chapman and Hall/CRC, 2018.
[4] Claude Berge. Hypergraphs: combinatorics of finite sets, volume 45. Elsevier, 1984.
[5] Muhammad Akram, A Nagoor Gani, and A Borumand Saeid. Vague hypergraphs. Journal of Intelligent & Fuzzy Systems,
26(2):647–653, 2014.
[6] Takaaki Fujita and Florentin Smarandache. Fundamental computational problems and algorithms for superhypergraphs. HyperSoft
Set Methods in Engineering, 3:32–61, 2025.
[7] Florentin Smarandache. Extension of HyperGraph to n-SuperHyperGraph and to Plithogenic n-SuperHyperGraph, and Extension
of HyperAlgebra to n-ary (Classical-/Neutro-/Anti-) HyperAlgebra. Infinite Study, 2020.
[8] Boyu Du, Jingya Zhou, Ling Liu, and Xiaolong She. Fl-gnn: Efficient fusion of fuzzy neural network and graph neural network. In
ECAI 2024, pages 1768–1775. IOS Press, 2024.
[9] Dalibor Krleža and Krešimir Fertalj. Graph matching using hierarchical fuzzy graph neural networks. Ieee transactions on fuzzy
systems, 25(4):892–904, 2016.
[10] Haotian Chen and Jialiang Xie. Eeg-based tsk fuzzy graph neural network for driver drowsiness estimation. Information Sciences,
679:121101, 2024.
[11] Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural networks. In Proceedings of the AAAI
conference on artificial intelligence, volume 33, pages 3558–3565, 2019.
[12] Takaaki Fujita and Florentin Smarandache. Superhypergraph neural networks and plithogenic graph neural networks: Theoretical
foundations. Infinite Study, 2025.
[13] Georg Gottlob, Nicola Leone, and Francesco Scarcello. Hypertree decompositions and tractable queries. In Proceedings of the
eighteenth ACM SIGMOD-SIGACT-SIGART symposium on Principles of database systems, pages 21–32, 1999.
[14] Chi Chen, Yongcheng Wang, Yuxi Zhang, Ning Zhang, Hao Feng, and Dongdong Xu. Hypergraph neural network for remote sensing
hyperspectral image super-resolution. Knowledge-Based Systems, page 113755, 2025.
[15] Bibin K Jose and Zsolt Tuza. Hypergraph domination and strong independence. Applicable Analysis and Discrete Mathematics,
3(2):347–358, 2009.
[16] Song Feng, Emily Heath, Brett Jefferson, Cliff Joslyn, Henry Kvinge, Hugh D Mitchell, Brenda Praggastis, Amie J Eisfeld, Amy C
Sims, Larissa B Thackray, et al. Hypergraph models of biological networks to identify genes critical to pathogenic viral response.
BMC bioinformatics, 22(1):287, 2021.
[17] Xiaowei Liao, Yong Xu, and Haibin Ling. Hypergraph neural networks for hypergraph matching. In Proceedings of the IEEE/CVF
International Conference on Computer Vision, pages 1266–1275, 2021.
[18] Yue Gao, Zizhao Zhang, Haojie Lin, Xibin Zhao, Shaoyi Du, and Changqing Zou. Hypergraph learning: Methods and practices.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2548–2566, 2020.
[19] Yifan Feng, Jiashu Han, Shihui Ying, and Yue Gao. Hypergraph isomorphism computation. IEEE Transactions on Pattern Analysis
and Machine Intelligence, 2024.
[20] Takaaki Fujita. An introduction and reexamination of molecular hypergraph and molecular n-superhypergraph. Asian Journal of
Physical and Chemical Sciences, 13(3):1–38, 2025.
[21] Takaaki Fujita. A theoretical investigation of quantum 𝑛-superhypergraph states. Neutrosophic Optimization and Intelligent Systems,
6:15–25, 2025.
[22] N. B. Nalawade, M. S. Bapat, S. G. Jakkewad, G. A. Dhanorkar, and D. J. Bhosale. Structural properties of zero-divisor hypergraph
and superhypergraph over Z𝑛 : Girth and helly property. Panamerican Mathematical Journal, 35(4S):485–495, 2025.
[23] Takaaki Fujita and Florentin Smarandache. A concise study of some superhypergraph classes. Neutrosophic Sets and Systems,
77:548–593, 2024.
[24] Florentin Smarandache. n-superhypergraph and plithogenic n-superhypergraph. Nidus Idearum, 7:107–113, 2019.
[25] Takaaki Fujita. Review of some superhypergraph classes: Directed, bidirected, soft, and rough. Advancing Uncertain Combinatorics
through Graphization, Hyperization, and Uncertainization: Fuzzy, Neutrosophic, Soft, Rough, and Beyond (Second Volume), 2024.
[26] Mohammad Hamidi, Florentin Smarandache, and Elham Davneshvar. Spectrum of superhypergraphs via flows. Journal of Mathe-
matics, 2022(1):9158912, 2022.
[27] Takaaki Fujita. Hypergraph and superhypergraph approaches in electronics: A hierarchical framework for modeling power-grid
hypernetworks and superhypernetworks. Journal of Energy Research and Reviews, 17(6):102–136, 2025.
[28] Yan Cao. Integrating treesoft and hypersoft paradigms into urban elderly care evaluation: A comprehensive n-superhypergraph
approach. Neutrosophic Sets and Systems, 85:852–873, 2025.
[29] Takaaki Fujita. Unifying grain boundary networks and crystal graphs: A hypergraph and superhypergraph perspective in material
sciences. Asian Journal of Advanced Research and Reports, 19(5):344–379, 2025.
[30] Thomas Jech. Set theory: The third millennium edition, revised and expanded. Springer, 2003.
[31] Florentin Smarandache. Foundation of superhyperstructure & neutrosophic superhyperstructure. Neutrosophic Sets and Systems,
63(1):21, 2024.
[32] Florentin Smarandache. The cardinal of the m-powerset of a set of n elements used in the superhyperstructures and neutrosophic
superhyperstructures. Systems Assessment and Engineering Management, 2:19–22, 2024.

15
[33] Alain Bretto. Hypergraph theory. An introduction. Mathematical Engineering. Cham: Springer, 1, 2013.
[34] WB Vasantha Kandasamy, K Ilanthenral, and Florentin Smarandache. Subset Vertex Multigraphs and Neutrosophic Multigraphs for
Social Multi Networks. Infinite Study, 2019.
[35] Hiroshi Nagamochi, Takashi Shiraki, and Toshihide Ibaraki. Augmenting a submodular and posi-modular set function by a multigraph.
Journal of Combinatorial Optimization, 5:175–212, 2001.
[36] Ryan C. Bunge, M. K. Chwee, Andrew Michael Wokingham Cooper, Saad I. El-Zanati, Katlyn Kennedy, Dan P. Roberts, and C. C.
Wilson. The spectrum for a multigraph on 4 vertices and 7 edges. 2018.
[37] Ryan C. Bunge, Joel Jeffries, Julie Kirkpatrick, Dan P. Roberts, and A. L. Sickman. Spectrum for multigraph designs on four vertices
and six edges. 2017.
[38] Armen S. Asratian and Raffi R. Kamalian. Interval colorings of edges of a multigraph. ArXiv, abs/1401.8079, 2014.
[39] Federico Romaniello. On 2-bisections and monochromatic edges in claw-free cubic multigraphs. 2023.
[40] Kelly J Pearson and Tan Zhang. The laplacian tensor of a multi-hypergraph. Discrete Mathematics, 338(6):972–982, 2015.
[41] Zhe Yang, Liangkui Xu, and Lei Zhao. Efbh: Collaborative filtering model based on multi-hypergraph encoder. IEEE Transactions
on Consumer Electronics, 70(1):2939–2948, 2023.
[42] Le An, Xiaojing Chen, Songfan Yang, and Xuelong Li. Person re-identification by multi-hypergraph fusion. IEEE transactions on
neural networks and learning systems, 28(11):2763–2774, 2016.
[43] Liping Nong, Jie Peng, Wenhui Zhang, Jiming Lin, Hongbing Qiu, and Junyi Wang. Adaptive multi-hypergraph convolutional
networks for 3d object classification. IEEE Transactions on Multimedia, 25:4842–4855, 2022.
[44] Maxime Bollengier, Abel Abel Dı́az Berenguer, and Hichem Sahli. Dynamic multi-hypergraph structure learning for disease
diagnosis on multimodal data. In 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology
Society (EMBC), pages 1–5. IEEE, 2024.
[45] Tongjie Pan, Yalan Ye, Yangwuyong Zhang, Kunshu Xiao, and Hecheng Cai. Online multi-hypergraph fusion learning for cross-
subject emotion recognition. Information Fusion, 108:102338, 2024.
[46] Cheng Zheng, Haojie Xu, and Xiao Sun. Mhg-erc: Multi-hypergraph feature aggregation network for emotion recognition in
conversations. ACM Transactions on Asian and Low-Resource Language Information Processing, 22(10):1–22, 2023.
[47] Yue Gao, Yifan Feng, Shuyi Ji, and Rongrong Ji. Hgnn+: General hypergraph neural networks. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 45(3):3181–3199, 2022.
[48] Xinyu Guo, Bingjie Tian, and Xuedong Tian. Hfgnn-proto: Hesitant fuzzy graph neural network-based prototypical network for
few-shot text classification. Electronics, 11(15):2423, 2022.
[49] Yixuan He, Quan Gan, David Wipf, Gesine D Reinert, Junchi Yan, and Mihai Cucuringu. Gnnrank: Learning global rankings
from pairwise comparisons via directed graph neural networks. In international conference on machine learning, pages 8581–8612.
PMLR, 2022.
[50] Mengqi Lei, Yihong Wu, Siqi Li, Xinhu Zheng, Juan Wang, Yue Gao, and Shaoyi Du. Softhgnn: Soft hypergraph neural networks
for general visual recognition. arXiv preprint arXiv:2505.15325, 2025.
[51] Tien Dang and Truong-Son Hy. Equihgnn: Scalable rotationally equivariant hypergraph neural networks. arXiv preprint
arXiv:2505.05650, 2025.
[52] Yuxin Wang, Quan Gan, Xipeng Qiu, Xuanjing Huang, and David Wipf. From hypergraph energy functions to hypergraph neural
networks. In International Conference on Machine Learning, pages 35605–35623. PMLR, 2023.
[53] Junwu Chen and Philippe Schwaller. Molecular hypergraph neural networks. The Journal of Chemical Physics, 160(14), 2024.
[54] Xiaojun Kang, Xinchuan Li, Hong Yao, Dan Li, Bo Jiang, Xiaoyue Peng, Tiejun Wu, Shihua Qi, and Lijun Dong. Dynamic
hypergraph neural networks based on key hyperedges. Information Sciences, 616:37–51, 2022.
[55] Mengfan Li, Xuanhua Shi, Chenqi Qiao, Teng Zhang, and Hai Jin. Hyperbolic hypergraph neural networks for multi-relational
knowledge hypergraph representation. arXiv preprint arXiv:2412.12158, 2024.
[56] Peng Zhou, Zongqian Wu, Xiangxiang Zeng, Guoqiu Wen, Junbo Ma, and Xiaofeng Zhu. Totally dynamic hypergraph neural
networks. In International Joint Conference on Artificial Intelligence, 2023.
[57] Federico Monti, Michael Bronstein, and Xavier Bresson. Geometric matrix completion with recurrent multi-graph neural networks.
Advances in neural information processing systems, 30, 2017.
[58] Ding Yao, Zhang Zhi-li, Zhao Xiao-feng, Cai Wei, He Fang, Cai Yao-ming, and Wei-Wei Cai. Deep hybrid: multi-graph neural
network collaboration for hyperspectral image classification. Defence Technology, 23:164–176, 2023.
[59] Du Yin, Renhe Jiang, Jiewen Deng, Yongkang Li, Yi Xie, Zhongyi Wang, Yifan Zhou, Xuan Song, and Jedi S Shang. Mtmgnn:
Multi-time multi-graph neural network for metro passenger flow prediction. GeoInformatica, 27(1):77–105, 2023.
[60] Yuzhi Song, Hailiang Ye, Ming Li, and Feilong Cao. Deep multi-graph neural networks with attention fusion for recommendation.
Expert Systems with Applications, 191:116240, 2022.
[61] Yi Ouyang, Bin Guo, Xing Tang, Xiuqiang He, Jian Xiong, and Zhiwen Yu. Learning cross-domain representation with multi-graph
neural network. arXiv preprint arXiv:1905.10095, 2019.
[62] Yaqin Ye, Yue Xiao, Yuxuan Zhou, Shengwen Li, Yuanfei Zang, and Yixuan Zhang. Dynamic multi-graph neural network for traffic
flow prediction incorporating traffic accidents. Expert Systems with Applications, 234:121101, 2023.
[63] Junjie Zhu, Xibin Zhao, Han Hu, and Yue Gao. Emotion recognition from physiological signals using multi-hypergraph neural
networks. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 610–615. IEEE, 2019.
[64] Ziang Li, Jie Wu, Guojing Han, Chi Ma, and Yuenai Chen. Multi-hypergraph neural network with fusion of location information for
session-based recommendation. IAENG International Journal of Applied Mathematics, 53(4), 2023.

16
 [65] Haojie Xu, Cheng Zheng, Zhuoer Zhao, and Xiao Sun. Multi-hypergraph neural networks for emotion recognition in multi-party
conversations. Applied Sciences, 13(3):1660, 2023.
[66] Cheng Zheng, Haojie Xu, and Xiao Sun. Multi-hypergraph neural networks for emotion recognition in multi-party conversations. In
National Conference on Man-Machine Speech Communication, pages 44–58. Springer, 2022.
[67] Jing Huang, Xiaolin Huang, and Jie Yang. Residual enhanced multi-hypergraph neural network. In 2021 IEEE international
conference on image processing (ICIP), pages 3657–3661. IEEE, 2021.

17

View publication stats