Research article

Published: 2025-03-24
https://doi.org/10.20935/AcadQuant7587

Accelerating spectral clustering on quantum and analog

platforms
Xingzi Xu1,2 , Tuhin Sahai2,∗
Academic Editors: Misha Erementchouk, Yuhua Duan

Abstract
We introduce a novel hybrid quantum–analog algorithm to perform a graph clustering that exploits connections between the evolution
of dynamical systems on graphs and the underlying graph spectra. This approach constitutes a new class of algorithms that combine
emerging quantum and analog platforms to accelerate computations. Our hybrid algorithm is equivalent to spectral clustering and
significantly reduces the computational complexity from O(N3 ) to O(N), where N is the number of nodes in the graph. We achieve
this speedup by circumventing the need for explicit eigendecomposition of the normalized graph Laplacian matrix, which dominates the
classical complexity, and instead leveraging quantum evolution of the Schrödinger equation followed by efficient analog computation
for the dynamic mode decomposition (DMD) step. Specifically, while classical spectral clustering requires O(N3 ) operations to perform
eigendecomposition, our method exploits the natural quantum evolution of states according to the graph Laplacian Hamiltonian in
linear time, combined with the linear scaling for DMD that leverages efficient matrix–vector multiplications on analog hardware. We
prove and demonstrate that this hybrid approach can extract the eigenvalues and scaled eigenvectors of the normalized graph Laplacian
by evolving Schrödinger dynamics on quantum computers followed by DMD computations on analog devices, providing a significant
computational advantage for large-scale graph clustering problems. Our demonstrations can be reproduced using our code that has been
released at on github.

Keywords: quantum computing, analog matrix computing, spectral clustering

Citation: Xu X, Sahai T. Accelerating spectral clustering on quantum and analog platforms. Academia Quantum 2025;2. https://doi.
org/10.20935/AcadQuant7587

1. Introduction retrieve spectral clustering information. We evolve the underly-

ing Schrödinger dynamics on quantum devices [11] and perform
Graph clustering is a popular technique to identify and group
the subsequent spectrum computations on analog machines. The
densely connected subgraphs within a larger graph. It is a pow-
approach has a scaling of O(N), where N is the number of nodes
erful decomposition approach that enables the analysis of inter-
in the graph. Since the existing state-of-the-art classical methods
connected systems for a variety of applications in a wide range
scale as O(N3 ), our approach provides a polynomial speedup over
of fields, such as social networks [1–3], fraud detection [4, 5],
traditional computing platforms. To the best of our knowledge,
bioinformatics [6], uncertainty analysis in networked systems [7],
this is the first-of-a-kind algorithm that exploits the combination
decomposition for scientific computation [8], and transport net-
of quantum and analog computing platforms to extract a scaling
works [9]. Consequently, the versatility of graph clustering makes
superior to algorithms that run on a single computing platform.
it an invaluable tool in areas ranging from scientific discovery
to industrial applications. Spectral clustering, a prevalent graph In quantum computing, qubits, the fundamental elements of
decomposition approach, is particularly advantageous in high- quantum information, evolve according to the time-dependent
dimensional spaces where the geometric characteristics of the Schrödinger equation. Qubits differ from classical bits in that
data are not readily apparent [10]. Unlike traditional clustering they can exist in a superposition of states, meaning that a qubit
methods, spectral clustering uses the properties of the eigenvalues can be in states representing the traditional bits (0, 1) or a su-
and eigenvectors of the Laplacian of the underlying graph (or perposition of both. Quantum algorithms provide substantial re-
data) to find the optimal partitioning. This aspect enables spectral ductions in computational complexity for specific problems such
clustering to cluster points based on their interconnectedness as integer factorization [12], database search [13], solving lin-
rather than their raw distances from each other, typically resulting ear systems [14], simulating dynamical systems [15], and matrix
in better performance when dealing with data where the clusters multiplication [16] to name a few. On the other hand, analog
are irregular, intertwined, or lie on a complex surface. computing with resistive memory has a high response speed and
embeds parallelism, significantly accelerating matrix computa-
In this work, we propose a highly scalable quantum–analog hybrid
tion [17], a core step in a wide variety of numerical techniques.
algorithm based on the evolution of Schrödinger dynamics and
Analog computing has garnered attention due to its potential to
dynamic mode decomposition (DMD) to rapidly and accurately

1
Department of Electrical and Computer Engineering, Duke University, Durham, NC, 27708, USA.
2
Applied Sciences, SRI International, Menlo Park, CA, 94025, USA.
∗
email: [email protected]

ACADEMIA QUANTUM 2025, 2 1 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

reduce the data movement bottleneck in traditional computing a limit where it is increasingly difficult to continue with appre-
architectures. The essence of analog computing is performing ciable reductions in the size of transistors [21]. Additionally, due
computational operations directly at the data’s location, rather to the high computational complexity of MVM required for the
than moving data between memory and the central processing training and inference of artificial intelligence models on digital
unit (CPU). Integrating crosspoint arrays with negative feedback platforms, there is an urgent need to develop alternative com-
amplifiers allows for addressing linear algebra challenges like puting devices with more attractive scaling properties for these
solving linear systems and computing eigenvectors in a single computations. Due to the inherent architectural parallelism for
step. Analog computing improves the exponential solving of linear matrix operations, analog matrix computing (AMC) circuits based
systems and has a computational complexity of O(1) for matrix– on crosspoint resistive memory arrays or photonics provide a de-
vector multiplication (MVM) [17, 18]. sirable alternative [17]. We note that historically, analog comput-
ers have been used extensively during the Apollo space program
Despite the advantages of spectral clustering, it typically scales as
to simulate the trajectory of spacecraft. They fell out of vogue
O(N3 ) (where N is the number of nodes in the graph), limiting its
during the digital revolution, but have recently been becoming
applications in large graphs [19]. We show that spectral clustering
increasingly popular for a wide variety of applications.
has an O(N) complexity on a hybrid quantum–analog platform.
For example, the multiplication between an M × N matrix and
We organize this article as follows: Section 3.2 provides a brief
an N × 1 vector has the computational complexity of O(MN)
overview of classical spectral clustering. Section 3.3 delves into
on digital computers and O(1) on AMC circuits [17] or photonic
the classical decentralized algorithm for spectral clustering that
devices [22]. Figure 1 shows the schematic of a prototypical AMC
leverages the wave equation and DMD. In Sections 4.1 and 4.2,
circuit for computing the MVM between matrix A and vector x.
we discuss our novel approach that uses quantum evolution of
The conductance in the circuits captures the entries of A, and
wave dynamics and singular-value decomposition and eigende-
the applied voltages encode the values of x. The output currents
composition on analog computers. We also provide details on
provide the result of y = Ax, due to Ohm’s law [23]. Since
efficiently solving the linear system of equations for computing
conductance can only take positive values, matrices with negative
eigenvectors on quantum or analog computers in Section 4.3.
entries can be split as A = A+ − A− , where A+ and A− are both
We present experimental results in Section 5 and wrap up with
positive. The results are then calculated as Ax = A+ x − A− x.
conclusions in Section 6.
Unlike digital computers, analog computers compute the MVM
results for all entries in parallel after the setting of conductances
and application of voltages. To quantify the error and computa-
2. Analog matrix computing circuits tional time of MVM on analog computers, we prove the following
Conventional computers have significantly improved in speed and theorem using circuit dynamics.
efficiency over the past decades, as captured by Moore’s law, stat- Theorem 1. Let the minimum time for convergence up to an
ing that the number of transistors in an integrated circuit doubles error ϵ (where ϵ is the difference between the computed and
about every two years [20]. However, the industry is approaching ground truth solutions) of MVM between an N × N matrix A and

x1 x2 xn
...

+
y1
-
A-

+
A+ y2
-
...
...

+
ym
-

Figure 1 • Schematic of analog matrix computing circuits for matrix–vector multiplication between matrix A and vector x, where
A = A+ − A− [17]. A ∈ RM×N , x ∈ RN×1 , y ∈ RM×1 .

ACADEMIA QUANTUM 2025, 2 2 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

an N × 1 vector x on circuit-based analog computers be T. Then, We note that the above derivation not only provides performance
T is given by bounds on the time to convergence, but also enables the design
of efficient analog circuits. Moreover, the above derivation will
1 x0 β − αj
T = max ln( ), (1) provide the wall clock time for executing various steps within our
j βj ϵ
∑
algorithm (outlined in Section 4) on analog platforms.
A y
where x0 is the initial condition, αj = ϑω 1+i∑ jiAi , β = ϑω
∑
1+ i Aji
,
i ji
and j is the index that takes values {1, 2, . . . , N}.
3. Preliminaries
Proof. Given the evolution dynamics of analog circuits is given In this section, we will summarize our previously constructed clas-
by [24] sical spectral clustering algorithm that exploits wave dynamics.
This algorithm will subsequently be adopted for quantum and
dx(t)
= ϑωU(Ay − x(t)), (2) analog platforms. Before we get into the details of spectral clus-
dt
tering and our approach, we will provide a summary of existing
where ϑ is the DC loop gain, ω is the 3-dB bandwidth, y is an input methods for clustering.
vector, A is the conductance matrix, and
3.1. Overview of various methods for graph clustering
 1

∑ 0 ··· 0
1+ i A1i
∑1
0 ··· 0
 
1+ i A2i The overall goal of partitioning graphs is to separate nodes into
 
U = . . . . (3)
 
.. .. .. .. groups or clusters such that there are “weak” interactions between
 . 
clusters and strong interactions within the groups. Numerous
 
0 0 · · · 1+∑1 A
i ni
methods for partitioning graphs have been developed over the
By expanding the above equations, it is easy to show that the j-th years. A summary of some of the approaches and their drawbacks
equation is given by is provided in Table 1. Several of the methods require additional
information about the graph. We note that algorithms for mod-
dxj (t)
= αj − βj xj (t), (4) ularity maximization are popular for clustering large graphs and
dt
have been used extensively for community detection. As shown
where αj and β are constants given by in the table, the Louvain method has a scaling of O(N log N),
∑
Aji yi which is significantly better than the O(N3 ) scaling of spectral
i∑
αj = ϑω , clustering. Note that modularity maximization has a spectral for-
1 + i Aji
mulation [28]. Consequently, our hybrid analog–quantum algo-
ϑω
βj = . (5) rithm can easily adapt to the community detection (modularity
1 + i Aji
∑
maximization) scenario. We can achieve this by replacing the
Note that Eq. 4 are all decoupled and their solution at time t is graph Laplacian in our approach with the matrix that arises in the
given by modularity maximization formulation.
(x(0)βj − αj )e−βj t + αj
x(t) = . (6) 3.2. Short introduction for spectral clustering
βj

αj
Consider a graph denoted by G = (V, E) that consists of a node set
Now, let the solution as t → ∞ be xf = βj
. If we consider a time Tj V = 1, . . . , N and an edge set E ⊆ V × V. Each edge (i, j) ∈ E in
such that the error ||x(t) − xf || is less than equal to our threshold this graph is associated with a weight Wij > 0, and W represents
ϵ, we get the N × N weighted adjacency matrix of G. In this context, Wij = 0
only if (i, j) ̸∈ E. The normalized graph Laplacian, represented by
(x(0)βj − αj )e−βj t
≤ϵ (7) L, is defined as follows:
βj

or if i = j,
1,


∑N
1 x0 β − α j Lij = −Wij / l=1 Wil , if (i, j) ∈ E, (10)
Tj ≥ ln( ). (8) 
βj ϵ 
0, otherwise.
It now follows that the minimum time for convergence of the
An equivalent representation of the normalized graph Laplacian
system (as indicated by the error reducing below a threshold of
is L = I − D−1 W, where D is a diagonal matrix that consists of
ϵ) is given by the maximum overall j ∈ {1, 2, . . . , N},
the sums of the rows of W. We focus solely on undirected graphs,
1 x0 β − αj for which the Laplacian is symmetric and the eigenvalues are real
T = max ln( ). (9)
j βj ϵ numbers. The eigenvalues of L can be arranged in ascending order

Table 1 • Various graph clustering methods

Method Approach Complexity Drawbacks
Spectral clustering Eigendecomposition O(N3 ) Poor scalability
Louvain method Modularity maximization O(N log N) Requires resolution parameter
Label propagation Label diffusion O(E) Needs label data
Markov clustering Random walks O(N3 ) Unbalanced cuts
Hierarchical clustering Agglomerative or divisive O(N2 ) Assumes hierarchical structure

ACADEMIA QUANTUM 2025, 2 3 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

as 0 = λ1 ≤ λ2 ≤ · · · ≤ λN . Each eigenvalue corresponds to where x(t) := [u(t), u(t + 1), . . . , u(t + (M − 1))]T . Let H > 1 also
an eigenvector, denoted as v(1) , v(2) , . . . , v(N) , where v(1) = 1 = define
[1, 1, . . . , 1]T [10]. For our study, we assume λ1 < λ2 , which means  
1 1 ... 1
that the graph does not contain disconnected clusters. Spectral  eiζ1 eiζ2 ... eiζN 
clustering partitions the graph G into two clusters by employing  
(18)
ΦH :=  .
. .
. . . .. ,
the signs of the entries of the second eigenvector v(2) . k-Means . . . .
 
 
clustering on the signs of entries of higher eigenvectors enables ei(H−1)ζ1 ei(H−1)ζ2 . . . ei(H−1)ζN
one to partition graphs into more than two clusters [10]. The
= [ϕ1 , ϕ2 , . . . , ϕN ], (19)
spectral gap in the eigenvalues of the graph Laplacians typically
determines the number of clusters. where ϕj = [1, eiζj , . . . , ei(H−1)ζj ]T . An earlier study [30] proved
the following two lemmas to connect the above matrices and
3.3. Wave equation–based classical clustering method the associated DMD computations to the spectra of the graph
Laplacian.
According to earlier studies [29, 30], we consider the wave equa-
tion given by Lemma 1. For one-dimensional signal u(t) defined by Eq. (14),
if K ≥ J and M ≥ J of the matrices X and Y defined by (15)
∂2u and (17), respectively, the eigenvalues of A = YX+ are {eiζj }Jj=1
= c2 ∆u. (11)
∂t2 and the columns of ΦK defined by (18) are the corresponding
The discrete form of the wave equation on a graph is given by eigenvectors [30].

ui (t) = 2ui (t − 1) − ui (t − 2) − c2
∑
Lij uj (t − 1), (12) Lemma 2. At any node l, the DMD computations on matrices
j∈N (i)
X(ul ) and Y(ul ) using local snapshots ul (0), ul (1), . . . , ul (4N − 1)
defined by Eq. 13 with K = M = 2N yield exact eigenvalues of the
where N (i) is the set of neighbors of node i including the node i Laplacian and the corresponding eigenvectors (scaled) [30].
itself [31]. To update ui , one only needs the previous value of uj at
the neighboring nodes and the connecting edge weights. Based on Lemma 2, an earlier study [30] developed a DMD-based
√ algorithm for distributed spectral clustering, given in Algorithms 1
Given the initial condition du |
dt t=0
= 0 and 0 < c < 2, the
solution of the wave equation 12 can be written as and 2.

N
u(0)T v(j) (pj eitζj + qj e−itζj )v(j) , 4. Quantum–analog hybrid computation
∑
u(t) = (13)
j=1
of spectral clustering
where pj = (1 + i tan(ζj /2))/2, qj = (1 − i tan(ζj /2))/2 [30]. We now adapt our classical algorithm such that various steps are
Consequently, computing the eigenvectors of L is transformed executed on quantum or analog platforms. We now describe the
into the computation of the coefficients of the frequencies of the details of executing the various steps on the two platforms along
wave dynamics in Eq. 13. with associated benefits.
To extract the above coefficients, we use DMD, a powerful tool for
analyzing the dynamics of nonlinear systems. We interpret DMD 4.1. Algorithm step 1: wave evolution on graphs
as an approximation of Koopman modes [32]. We apply DMD
Conventional computers suffer from various computational chal-
to one-dimensional time series data with time-delay embedding
lenges and inherent truncation errors in the simulation process
by constructing the matrix X and Y for the exact DMD matrix
when performing dynamics simulations. In particular, matrix
A = YX+ (where + denotes the pseudoinverse) [33]. When DMD
multiplication, central to these computations, has a computa-
with time-delay embedding is implemented on one-dimensional
tional complexity of O(N2.371866 ) on digital computers [34]. The
signals of the form
algorithm with this complexity also has reduced numerical stabil-
J ity, compared to the naïve O(N3 ) algorithm [34]. On the contrary,
aj eiζj t ,
∑
u(t) = (14) quantum and analog computers demonstrate inherent prowess in
j=1
this domain, providing an exponential advantage for simulating
where ζj ∈ (−π, π), j = 1, 2, . . . , J are unique frequencies, it can dynamical systems compared to their digital counterparts [15,
successfully extract the coefficients aj [30, 33]. 35, 36]. For the specific case of wave dynamics, we can choose
to simulate the equations in their continuous form on a quan-
We form the following matrices with the one-dimensional signal tum computer or after discretization on an analog device. Note
u(t): that continuous simulations on quantum computers avoid the
truncation errors met in discrete simulations on analog or digital
u(0) u(1) . . . u(M − 1)
 
 u(1) computers. We now provide details on both computing options.
u(2) . . . u(M) 
X :=  (15)
 
 .. .. .. .. ,
. . . . 4.1.1. Quantum acceleration of wave dynamics

 
u(K − 1) u(K) . . . u(K + M − 2) evolution
The wave equation on a graph [37] is equivalent to Eq. 11, given as
[ ]
= xl (0) xl (1) . . . xl (M − 1) , (16)
[ ] d2 u
Y := xl (1) . . . xl (M − 1) xl (M) , (17) = −c2 Lu, (20)
dt2

ACADEMIA QUANTUM 2025, 2 4 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

Algorithm 1 DMD(X, Y): For computing eigenvalues and eigenvector components at node i [30].
1: Compute reduced SVD of X, i.e., X = UΣV∗ .
2: Define the matrix à ≡ U∗ YVΣ−1 .
3: Compute eigenvalues/vectors µ and ξ of Ã, i.e., Ãξ = µξ. Nonzero eigenvalues µ are DMD eigenvalues.
4: The DMD mode corresponding to µ is then given by ϕ̂ ≡ µ1 YVΣ−1 ξ.
5: Compute â by solving the linear system Φ̂â = x(0), where the columns of Φ̂ are the eigenvectors sorted in decreasing order based
on the real part of the eigenvalues.
(j)
6: ai = ϕ̂j,1 âj , j = 1, 2, . . . , k.

Algorithm 2 Wave equation–based graph clustering [30].

1: ui ← Random([0, 1])
2: ui (−1) ← ui (0)
3: t←1
4: while do t < Tmax
5: ui (t) ← 2ui (t − 1) − ui (t − 2) − c2 j∈N (i) Lij uj (t − 1)
∑
6: t←t+1
7: end while
8: Create the matrices Xi , Yi ∈ RK×M defined by Eqs. 15 and 17 at node i,
9: using ui (0), ui (1), . . . , ui (Tmax − 1), where K + M = Tmax .
10: vi ← ai from DMD(Xi , Yi ) by algorithm 1
11: for j ∈ {1, . . . , k} do
(j)
12: if vi > 0 then
13: Γj ← 1
14: else
15: Γj ← 0
16: end if
17: end for
18: Cluster number ← kj=1 Γj 2j−1
∑

where L is the graph Laplacian [31]. In its discrete form, the wave for a predetermined period before measurements collapse the
equation on a graph is as follows: wavefunction. In the above equation, H is the Hermitian static
∑ Hamiltonian [35]. We note that the original classical algorithm
ui (t) = 2ui (t − 1) − ui (t − 2) − c2 Lij uj (t − 1), (21)
used the random-walk version of the graph Laplacian [29]. How-
j∈N (i)
ever, for quantum computations, the symmetric form of the graph
Laplacian [10] is more amenable for Hamiltonian embeddings
where ui is the i-th element of u. In an earlier study [30],
(see Remark 1 for more details). Here, Lsym is the symmetric
the authors simulated wave dynamics with dt = 1 on a digital
normalized graph Laplacian [10] defined as
computer. As this simulation only involves MVM, we can replace
this O(N2 ) operation on digital computers with O(N) on quantum
Lsym = I − D−1/2 WD1/2 , (23)
platforms [17, 38] (or O(1) operation on analog computers).
∑n ∑n
Using the above approach, simulating the wave equation on a where the degree matrix D = diag( j=1 W1j , j=1 W2j , . . . ,
∑n
graph has a time complexity of O(D5/2 l/∆l + TD2 /∆l), where T is j=1 Wnj ).
the terminal time of simulation, D is the dimension of the wave, l
Now one can define the symmetric Laplacian using the normalized
is the diameter of the region, and ∆l is the step size [39]. Conse-
signed incidence matrix of the graph (denoted by B) as follows:
quently, for the setting of graphs, the dimension D is 2 (time and
Lsym = BBT . To define B, consider a graph with N nodes and M
space), the diameter of the region l is the number of nodes N, and
edges, and then the N × M incidence matrix ι has rows indexed
the step size a is 1, and so the complexity is O(N + T). In practice,
by nodes v and columns indexed by edges e. Specifically, the
we simulate the wave dynamics for around 2N steps, leading to
incidence matrix ι is defined as
an O(N) computational complexity. Since T = 2N leads to long
simulation times for large graphs, the approximation (round-off)

−1 if edge ej leaves node vi ,


errors accumulate for the discrete system of equations. Therefore,
ιij = 1 if edge ej enters node vi ,
an exact continuous simulation is favorable over their discretized 
otherwise.

0
counterparts.
To implement continuous simulations of the wave dynamics that Now, the normalized incidence matrix B is defined as
avoid truncation errors, the study [39] suggests using quantum
1
computers by exploiting native Hamiltonian simulations or the B = D− 2 ι. (24)
quantum linear system algorithm (QLSA). In our work, we evolve
the following Hamiltonian system on quantum platforms:
Since we restrain to undirected graphs, for any given edge ej , we
du randomly pick one node as the source node and the other as the
= −iHu(t), (22)
dt sink node.

ACADEMIA QUANTUM 2025, 2 5 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

Using the formulation given in an earlier study [39], if we now Proof. For the Lsym case, the dynamics of z(t) = [u(t); u(t − 1)]
define a Hermitian Hamiltonian in the following block form: evolves as
[ ] [ ] [ ][ ]
0 B u(t) 2I − c2 Lsym −I u(t − 1)
H=c , (25) z(t) = =
B 0 u(t − 1) I 0 u(t − 2)

the Schrödinger Hamiltonian system takes the form = Mz(t − 1) = Mt z(0).

[ ] [ ][ ]
d u1 0 B u1 Let matrix M have eigenvector mj = (mj , mj )T , where j
(1) (2)
= −ic , (26)
dt u2 B 0 u2 denotes the node number and the superscript splits the compound
vector into two equal parts. Now, let the corresponding eigenvalue
which implies that
[ ] [ ]2 [ ] [ ][ ] be ρj , and it then follows that
T
d2 u1 2 0 B u1 2 BB 0 u1
= −c = −c
[ (1) ] [ ] [ (1) ]
dt2 u2 B 0 u2 0 BT B u2 mj 2I − c2 Lsym −I mj
M (2) =
[ ][ ] mj I 0 (2)
mj
Lsym 0 u1
= −c2 . (27)
[ (1) (1) (2)
]
0 BT B u2 2mj − c2 Lsym mj − mj
= (1) ,
mj
When simulating dynamics of the Schrödinger system using
Eq. 27, the first N terms correspond to the desired wave dynamics. [ ] [ (1) ] [ ]
(1) (1)
When u has a nonzero initial derivative condition, the resulting mj mj ρj mj
M (2) = ρj (2) = (2) .
solution tends to increase over time, leading to instability [40]. mj mj ρj mj
Given that
[ ] [ ][ ] [ ]
d u1 0 B u1 Bu2 If we let mj
(2)
= vj (where vj is the eigenvector of Lsym ), the
= −ic = −ic , (28)
dt u2 B 0 u2 Bu1 (1)
above equation implies that mj = ρj mj
(2)
= ρj vj . Therefore, the
the stability criteria du 1
| = 0 translate to the initial condition eigenvector mj of M is mj = (ρj vj , vj )T .
dt t=0
u2 (0) = 0. Therefore, all nonzero initial values for u2 lead to Using the definition of eigenvalues and pulling everything to the
unstable dynamics. left-hand side of the equation,
Remark 1. We again note that since the graph Laplacian form [ ] [ ] [ ] [ ]
ρj vj ρj vj (ρj (2 − ρj )I − ρj c2 Lsym − I)vj 0
used in previous works [29, 30] is the random-walk variant of the M = ρj ⇒ = .
vj vj 0 0
graph Laplacian Lrw , it does not have a Hermitian form. Con-
sequently, the wave dynamics based on Lrw cannot be evolved
using quantum platforms. Consequently, on quantum devices, If we now denote the eigenvalue of Lsym as λj , we get
one is restricted to the symmetric graph Laplacian Lsym form
defined in Eq. 23. Moreover, since the simulation is Hamiltonian, ρj (2 − ρj )vj − c2 ρj Lsym vj − vj = 0 ⇒ ρ2j + (c2 λj − 2)ρj + 1 = 0,
it is guaranteed to remain stable on quantum platforms.
and so we have
Remark 2. In Eq. 26, the initial state can be generated by start-
ing from a fixed reference quantum state ψ0 and transforming −(c2 λj − 2) ± 2 − c2 λj c√ 2 2
√
(c2 λj − 2)2 − 4
it to |ψU ⟩ = U|ψ0 ⟩, where U is generated from the uniform ρj = = ± c λj − 4λj .
2 2 2
Haar measure [41, 42]. This result is a natural consequence of
Therefore, to generate stable dynamics with M, we need Lsym ’s
the results derived on the initial condition of the wave equation
eigenvalues λj ’s to satisfy c2 λ2j −4λj < 0. If we now use the fact that
evolution derived in an earlier work [29]. Therefore, state prepa- √
0 ≤ λj ≤ 2 (see [10] for more details), we get 0 ≤ c < 2. Thus, as
ration will not be a significant bottleneck in the implementation
long as the above conditions are true, the update equations remain
of our approach.
stable on analog devices.

4.1.2. Analog computer–based acceleration of wave 4.2. Algorithm step 2: eigenvalue/vector computations
dynamics
4.2.1. Analog computer–based acceleration of
When working with the discrete form of the equation, one can eigenvector computations
exploit the favorable O(1) scaling of analog platforms for MVM
Once the wave equation is evolved on either analog or quantum
operation. Unlike the quantum setting, the Lrw update equation
platforms, one has to compute the eigenvalues and eigenvector
that is used in the classical approach [29, 30, 43] can be used
components for node l using the DMD approach outlined pre-
on analog platforms. However, in the following lemma, we prove
viously. We begin with a reduced SVD of matrix X. Despite the
that the wave dynamics with Lsym remain stable under the same
superiority of existing quantum SVD methods (polylog complex-
conditions that were derived in an earlier work [29] for Lrw ,
ity), they only work on low-rank matrices, which is not the case
thereby significantly extending the original results for digital and
here [44]. We instead focus on efficient methods for SVD that
analog settings.
exploit MVM and are, therefore, tailored for analog computers.
Lemma 3. The wave equation iteration given by Eq. 28 is stable It’s important to note that the eigenvalues of XT X, represented as
√
on any graph as long as the numerical value of the wave speed γi , correspond to the squared singular values of X [45]. Since
√
c satisfies 0 < c < 2 and the iterations have initial conditions XT X is symmetric, it has distinct real eigenvalues, which can be
that satisfy u1 (−1) = u1 (0). determined using the power method.

ACADEMIA QUANTUM 2025, 2 6 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

The power method approximates dominant eigenvalues (and their O(M log M) [17, 49, 50]. Although quantum algorithms provide
corresponding eigenvectors) of matrices. Especially powerful for the same scaling, existing quantum algorithms are restricted to
large-scale matrices, this method iteratively refines estimates of matrices with specific properties, such as a Hermitian structure
the dominant eigenvector, capitalizing on the property that re- and diagonalizability [51–53]. Consequently, in our approach,
peated multiplication accentuates the contribution of the domi- we exploit analog platforms for these computations due to their
nant eigenvalue. The following recurrence relation describes the flexibility. As quantum algorithms develop further, it is possible
power method: that future variants of this algorithm will rely on quantum devices.

XT Xbk
bk+1 = . (29) 4.3. Algorithm step 3: solving linear systems
∥XT Xbk ∥
4.3.1. Quantum-accelerated solutions of linear systems
Using the eigenvector bk , the estimation for the dominant eigen-
As mentioned previously, the final step within the step of DMD
value is the Rayleigh quotient [45]
computations is to solve a system of linear equations of the form
bTk XT Xbk Φ̂â = x(0). Here, â provides the coefficients whose signs are used
γk = . (30)
bTk bk to assign the nodes to corresponding clusters. Quantum comput-
ers excel at this task, with a time complexity of O(poly(log N, κ)),
To calculate other eigenvalues and eigenvectors using the power where N and κ are the dimension and condition number of Φ̂,
method, we need to remove the dominant eigenvalue while pre- respectively. By contrast, classical algorithms have a scaling of
√
serving the spectrum. Deflation methods such as Wielandt’s and O(N κ) (we refer the reader to [14] for further details on solving
Hotelling’s deflation provide such tools [46,47]. The statement for linear systems on quantum platforms).
the Hotelling deflation is captured by Lemma 4.
4.4. Summary of steps
Lemma 4. If ψ1 ≥ ψ2 ≥ · · · ≥ ψN are the eigenvalues of matrix
B, and ν (1) , ν (2) , . . . , ν (N) are the corresponding eigenvectors, In summary, the algorithm steps and their associated scaling on
define B̂ = B − ν (j) ν (j)T Bν (j) ν (j)T for some j ∈ {1, 2, . . . , N}. the various computing platforms are shown in Table 2. This table
Then, B̂ has eigenvectors ν (1) , ν (2) , . . . , ν (N) with corresponding represents the best-known scaling of the current generation of
eigenvalues ψ1 , . . . , ψj−1 , 0, ψj+1 , . . . , ψN [47]. algorithms. As new algorithms for quantum and analog platforms
are developed, the benefits and drawbacks associated with exe-
Hotelling’s deflation maintains the full spectrum of a matrix, cuting various steps of spectral clustering on these platforms are
nullifying a chosen eigenvalue while keeping the rest intact. expected to evolve.
Implementation of Hotelling’s deflation is straightforward with
just MVMs. Due to the approximation error inherent in the Table 2 • A summary of the computational complexities of vari-
ous components of the clustering algorithm. H is the Hermitian
power method, Hotelling’s deflation does not exactly eliminate
static Hamiltonian, N is the number of nodes, Nt is the number
the eigenvalues, causing numerical instabilities for large matrices.
of time steps, K is the number of iterations taken for the power
Schur’s deflation method mitigates these challenges and remains method, M is the number of components left in reduced singular
stable despite the numerical errors in the eliminated eigenval- value decomposition (SVD), and T is the terminal time [14, 17,
ues. When the elimination is exact, Schur’s deflation reduces to 25]. The complexity of quantum matrix–vector multiplication is
Hotelling’s deflation [47]. The power method struggles to de- adapted from quantum matrix multiplication algorithms, which
termine smaller eigenvalues due to numerical issues. Since we have a complexity of O(N2 ) [26]. Mizuno and Komatsuzaki [27]
exploit the reduced SVD (where we disregard exceedingly small proposed a quantum version of the dynamic mode decomposition
singular values), these numerical issues do not affect the efficiency algorithm to achieve an O(poly log N) complexity for eigendecom-
of our algorithm [45]. position
Digital Analog Quantum
Let us define the SVD of X as X = UΣVT . The eigendecompo- Simulate wave O(N2 Nt ) O(Nt ) O(N)
sition with the power method and Hotelling’s deflation provides dynamics
us the singular values of X (given by the diagonal of Σ) as well Reduced SVD O(M2 N + M3 ) O(KM) O(poly log N)
as the eigenvectors of XT X (given by the columns of V). When Eigendecomposition O(M3 ) O(M log M) O(poly log N)
performing a reduced SVD, we keep only M ≪ N singular values Matrix–vector O(N2 ) O(1) O(N)
and columns of V. We can calculate the reduced U using the multiplication
pseudoinverse of VTred , given by Solving linear system O(N3 ) O(log N) O(log N)

Ured ≈ XVT+ −1
red Σred , (31)

where VT+ T
5. Results
red is the pseudoinverse of Vred . Both quantum and analog
computers provide efficient methods for computing the pseudoin- We implemented our proposed algorithm using the IBM Qiskit
verse [17, 48], and in our approach, either can be used. package and performed numerical experiments to verify that its
results match those of classical algorithms. Our computations are
The subsequent step within DMD involves calculating the eigen-
performed on quantum emulators since the current generation
values of à = U∗ YVΣ−1 . The power method is inapplica-
of quantum computers does not have enough qubits to handle
ble in the setting of general unsymmetric matrices such as Ã.
meaningful examples.
However, given that à is of dimensions RM×M and M ≪ N,
eigendecomposition techniques are not impacted by large N val-
ues. Eigendecomposition on analog computers has a scaling of

ACADEMIA QUANTUM 2025, 2 7 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

First, as shown in Figure 2, we experiment with a benchmark Second, we test the performance of our method on the Twitter
called Zachary’s karate club graph [54]. Zachary’s karate club interaction network for the US Congress, which represents the
graph is a widely studied social network representing the relation- Twitter interaction network of both the House of Representatives
ships among a karate club’s 34 members at a US university in the and Senate that formed the 117th US Congress [56, 57]. As shown
1970s. In the graph, each node represents a member of the karate in Figure 3, our method correctly classifies 467 out of 475 nodes,
club, and each edge indicates a tie between two members outside thereby achieving an accuracy of 98.32%. This discrepancy arises
of the club activities. The network is undirected and unweighted, due to the numerical error of simulating quantum dynamics using
reflecting the mutual nature of social relationships. We simulated the Qiskit dynamics package. We also test the proposed method
the wave dynamics using Qiskit for the time interval t = [0, 99], on the social circle graph on Facebook [58]. In particular, we take
measuring the dynamics at integer times [55]. As expected, the 200 nodes from the dataset. As shown in Figure 4, our method
eigenvector estimated by the proposed method is a constant factor correctly clusters 196 out of the 200 nodes, achieving an accuracy
of the ground truth (see Figure 2). Since the ground truth and of 98%. We expect 100% accuracy of cluster assignments on an
computed eigenvectors share consistent signs for assigning nodes error-corrected (fault-tolerant) quantum platform. We note that
to clusters, the estimated clusters exactly match. the error of computing the eigenvector components (from the
Schrödinger dynamics data) was below the threshold required to
perform accurate cluster assignments and did not contribute to
the error rates.

To numerically verify the stability of the wave equation simulated

with the Lsym (proved in Lemma 3), we plot the wave dynamics
used for clustering on the Twitter dataset in Figure 5. Here, it is
easy to see that the dynamics at the second node remains bounded
for all time, as should be expected in the evolution of Hamiltonian
systems. Similar evolution plots can be generated at all nodes in
the graph. Note that the Qiskit dynamics are not expected to match
the discrete equation evolution since the governing equations for
both the cases are different.

Figure 2 • The estimated and the ground truth eigenvector of the

adjacency matrix of the karate club network

Figure 3 • (a) The clusters calculated according to the eigenvector estimated with the proposed method. (b) Ground truth cluster
calculated with spectral clustering. Clusters of the Twitter interaction network for the US Congress, comparing the ground truth and
the proposed method’s results.

Lastly, we test the proposed method on a synthetic network with have weak interaction. In this case, as can be seen in Figure 6,
40 nodes and four clusters. The network is generated such that our method perfectly estimates the clusters with 100% accuracy.
it contains four clusters of nodes, where the nodes within each All results can be reproduced with our codebase available online
cluster are closely connected, and the four clusters themselves [59]. I confirm

ACADEMIA QUANTUM 2025, 2 8 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

Figure 4 • Comparison of the ground truth spectral clustering results and the proposed quantum–analog framework on 200 nodes from
the Facebook social network graph.

Figure 5 • The wave dynamics of the second node on the Twitter dataset simulated with Qiskit dynamics verify the stability of
quantum-simulated dynamics.

(a) (b)

Figure 6 • (a) The clusters calculated according to the eigenvector estimated with the proposed method. (b) The estimated and ground
truth eigenvectors of the adjacency matrix of the network. Estimated clusters of the synthetic network and the comparison of the
estimated and ground truth eigenvectors. DMD, dynamic mode decomposition.

ACADEMIA QUANTUM 2025, 2 9 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

6. Conclusions digital, analog, and quantum platforms will provide a unique

opportunity for developing novel and efficient algorithms that
We propose a first-of-a-kind quantum–analog hybrid algorithm
exploit the inherent strength of each platform. As these computing
to cluster large-scale graphs. The algorithm contains two com-
platforms mature, we expect an increase in the development of
ponents that are sequentially applied. In the first component,
such algorithms that exploit the dynamic nature of these problems
we simulate wave dynamics on a graph using analog (discretized
and devices. These embeddings also provide deep insights into
dynamics) or quantum (continuous dynamics) computers. Both
the fundamental limitations of algorithm construction and related
analog and quantum platforms accelerate the computations by
complexity classes [61]. The use of these embeddings for com-
a polynomial factor. However, we expect fault-tolerant quantum
putational complexity analysis of problem classes on emerging
computers to provide a significant advantage over analog com-
computing platforms remains an underexplored area of research.
puters for the task of simulating wave dynamics due to their
inherent lack of discretization error, which provides a key advan-
tage of robustness for large graphs. In the second component,
we analyze the data from the wave dynamics step using DMD on
Acknowledgments
analog platforms. In this step, we achieve polynomial speedups The authors thank SRI International for their financial support for
over digital computers, arising from the O(1) complexity of MVM the work under Internal Research & Development (IRAD) project
on analog computers. Overall, we accelerate the existing O(N3 ) number IRHOME.A.2023.SNAR.0000.
complexity on digital computers to O(N) by using the combi-
nation of quantum and analog platforms. We demonstrate our
proposed algorithm on diverse benchmark datasets, including Funding
the Zachary karate club example, Twitter interaction networks,
This research was sponsored by SRI International under in-
Facebook social circles, and random graphs by reproducing the
ternal research and development project (project number:
clusters in these graphs with over 98% accuracy. The error rate
IRHOME.A.2023.SNAR.0000). This research received no exter-
is due to the non-native simulation of the Schrödinger dynamics
nal funding.
and expected to be rectified when running these computations
on quantum devices. However, in practice, the devices will have
an error rate that will influence the accuracy of the approach.
Performing sensitivity analysis of the approach concerning error
Author contributions
rates in the analog and quantum portions of the computation is Conceptualization, T.S.; methodology, T.S. and X.X.; software,
subject to future work. X.X.; validation, T.S. and X.X.; formal analysis, T.S. and X.X.; in-
As both of the computing paradigms mature, we expect our vestigation, T.S. and X.X.; resources, T.S. and X.X.; data curation,
method to become increasingly attractive for practical applica- X.X.; writing—original draft preparation, X.X.; writing—review
tions. Currently, quantum computers are limited in the number and editing, T.S. and X.X.; visualization, X.X.; supervision, T.S.;
and fidelity of qubits. Similarly, analog devices are still under project administration, T.S.; funding acquisition, T.S. All authors
development and exist only as experimental prototypes. Conse- have read and agreed to the published version of the manuscript.
quently, we have used classical emulators for both platforms to
perform the above experiments.
Conflict of interest
We expect that our method will be particularly useful in the analy-
sis of very large graphs that require dynamic updates to the parti- The authors declare no conflict of interest.
tioning. For example, the approach will be useful for partitioning
of proliferated satellite networks (for task allocation), formation
of teams within heterogeneous autonomous systems, analysis of Data availability statement
cellular networks, and biological graphs such as metabolic and Data supporting these findings are available within the article, at
genetic networks. The primary drawback of the approach is that https://doi.org/10.20935/AcadQuant7587, or upon request.
it requires access to large-scale quantum and analog devices that
can be coupled to one another. Moreover, we assume that the com-
putations are fault-tolerant and error-corrected. Such platforms Institutional review board statement
are not expected to be portable in the near future, restricting their
deployment on mobile agents. Not applicable.

Given that the development of algorithms that leverage both

analog and quantum computing devices is at a nascent stage, we
Informed consent statement
anticipate that several existing and new algorithms will leverage
a combination of these paradigms. However, a comprehensive Not applicable.
understanding of the properties of problems and algorithms that
make them attractive for these settings remains elusive.

More broadly, this work demonstrates that embeddings of dis-

Additional information
crete problems in continuous spaces enable the construction of Received: 2025-01-02
efficient algorithms on emerging computing devices. Given deep
Accepted: 2025-03-02
connections between dynamical systems and combinatorial op-
timization problems [60], we expect that novel combinations of Published: 2025-03-24

ACADEMIA QUANTUM 2025, 2 10 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

Academia Quantum papers should be cited as Academia Quan- 10. Von Luxburg U. A tutorial on spectral clustering. Stat Com-
tum 2025, ISSN 3064-979X, https://doi.org/10.20935/Acad put. 2007;17:395–416. doi: 10.1007/s11222-007-9033-z
Quant7587. The journal’s official abbreviation is Acad. Quant.
11. Sahai T, Krovi HK. Spectral clustering of graphs on fault
tolerant and noisy quantum devices. Google Patents; 2023.
Publisher's note US Patent App. 17/843,801.
Academia.edu Journals stays neutral with regard to jurisdictional 12. Shor PW. Polynomial-time algorithms for prime factorization
claims in published maps and institutional affiliations. All claims and discrete logarithms on a quantum computer. SIAM Rev.
expressed in this article are solely those of the authors and do 1999;41(2):303–32. doi: 10.1137/S0097539795293172
not necessarily represent those of their affiliated organizations, or
those of the publisher, the editors, and the reviewers. Any product 13. Grover LK. A fast quantum mechanical algorithm for
that may be evaluated in this article, or claim that may be made by database search. In: Proceedings of the Twenty-Eighth An-
its manufacturer, is not guaranteed or endorsed by the publisher. nual ACM Symposium on Theory of Computing; 1996 May
22–24; Philadelphia , PA, USA. New York (NY): ACM; 1996.
p. 212–9.
Copyright
© 2025 copyright by the authors. This article is an open ac- 14. Harrow AW, Hassidim A, Lloyd S. Quantum algorithm for
cess article distributed under the terms and conditions of the linear systems of equations. Phys Rev Lett. 2009;103:150502.
Creative Commons Attribution (CC BY) license (https://creative doi: 10.1103/PhysRevLett.103.150502
commons.org/licenses/by/4.0/).
15. Surana A, Gnanasekaran A, Sahai T. An efficient quantum al-
gorithm for simulating polynomial dynamical systems. Quan-
References tum Inf Process. 2024;23(3):105. doi: 10.48550/arXiv.2212.
10775
1. Mishra N, Schreiber R, Stanton I, Tarjan RE. Clustering social
networks. In: Bonato A, Chung FRK, editors. International 16. Li H, Jiang N, Wang Z, Wang J, Zhou R. Quantum ma-
Workshop on Algorithms and Models for the Web-Graph. trix multiplier. Int J Theor Phys. 2021;60:2037–48. doi:
New York (NY): Springer; 2007. p. 56–67. 10.1007/s10773-021-04816-x
2. Liu R, Feng S, Shi R, Guo W. Weighted graph clustering for 17. Sun Z, Ielmini D. Invited tutorial: analog matrix computing
community detection of large social networks. Procedia Com- with crosspoint resistive memory arrays. IEEE Trans
put Sci. 2014;31:85–94. doi: 10.1016/j.procs.2014.05.248 Circuits Syst II Exp Briefs. 2022;69(7):3024–9. doi:
3. Bu Z, Cao J, Li HJ, Gao G, Tao H. GLEAM: a graph clustering 10.48550/arXiv.2205.05853
framework based on potential game optimization for large-
18. Sun Z, Pedretti G, Mannocci P, Ambrosi E, Bricalli A, Ielmini
scale social networks. Knowl Inf Syst. 2018;55:741–70. doi:
D. Time complexity of in-memory solution of linear sys-
10.1007/s10115-017-1105-6
tems. IEEE Trans Electron Devices. 2020;67(7):2945–51.
4. Cai F, Le-Khac N-A, Kechadi T. Clustering approaches for fi- doi: 10.48550/arXiv.2005.04530
nancial data analysis: a survey. arXiv:1609.08520 [Preprint].
2016. doi: 10.48550/arXiv.1609.08520 19. White S, Smyth P. A spectral clustering approach to finding
communities in graphs. In: Kargupta H, Srivastava J, Kamath
5. Pourhabibi T, Ong KL, Kam BH, Boo YL. Fraud detection: a C, Goodman A, editors. Proceedings of the 2005 SIAM In-
systematic literature review of graph-based anomaly detec- ternational Conference on Data Mining (SDM). Philadelphia
tion approaches. Decis Support Syst. 2020;133:113303. doi: (PA): Society for Industrial and Applied Mathematics; 2005.
10.1016/j.dss.2020.113303 p. 274–85.
6. Higham DJ, Kalna G, Kibble M. Spectral clustering and its use
20. Schaller RR. Moore’s law: past, present and future. IEEE
in bioinformatics. J Comput Appl Mathe. 2007;204(1):25–37.
Spectr. 1997;34(6):52–9. doi: 10.1109/6.591665
doi: 10.1016/j.cam.2006.04.026

7. Surana A, Sahai T, Banaszuk A. Iterative methods for scal- 21. Powell JR. The quantum limit to Moore’s law. Proc IEEE.
able uncertainty quantification in complex networks. Int J 2008;96(8):1247–8. doi: 10.1109/JPROC.2008.925411
Uncertain Quantif. 2012;2(4):413–39. doi: 10.48550/arXiv.
22. Xu S, Wang J, Yi S, Zou W. High-order tensor flow pro-
1110.2435
cessing using integrated photonic circuits. Nat Commun.
8. Klus S, Sahai T, Liu C, Dellnitz M. An efficient algorithm 2022;13(1):7970. doi: 10.1038/s41467-022-35723-2
for the parallel solution of high-dimensional differential
equations. J Comput Appl Math. 2011;235(9):3053–62. doi: 23. Gilbert-Kawai ET, Wittenberg MD. Ohm’s law, voltage, cur-
10.48550/arXiv.1003.5238 rent and resistance. New York (NY): Cambridge University
Press; 2014.
9. Ahmed S, Kanhere SS. Cluster-based forwarding in delay tol-
erant public transport networks. In: 32nd IEEE Conference 24. Sun Z, Huang R. Time complexity of in-memory matrix-
on Local Computer Networks (LCN 2007); 2007 Oct 15–18; vector multiplication. IEEE Trans Circuits Syst II Exp Briefs.
Dublin, Ireland. Piscataway (NJ): IEEE; 2007. p. 625–34. 2021;68(8):2785–9. doi: 10.1109/TCSII.2021.3068764

ACADEMIA QUANTUM 2025, 2 11 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

25. Li X, Wang S, Cai Y. Tutorial: complexity analysis of singu- 40. Evans LC. Partial differential equations. Vol. 19. Ann Arbor
lar value decomposition and its variants. arXiv:1906.12085 (MI): American Mathematical Society; 2022.
[Preprint]. 2019. doi: 10.48550/arXiv.1906.12085
41. Reck M, Zeilinger A, Bernstein HJ, Bertani P. Experimental
26. Shao C. Quantum algorithms to matrix multiplication; 2018. realization of any discrete unitary operator. Phys Rev Lett.
arXiv:1803.01601 [Preprint]. 2018. doi: 10.48550/arXiv.18 1994;73(1):58. doi: 10.1103/PhysRevLett.73.58
03.01601
42. Tang H, Banchi L, Wang TY, Shang XW, Tan X, Zhou WH,
27. Mizuno Y, Komatsuzaki T. Quantum algorithm for dynamic et al. Generating Haar-uniform randomness using stochastic
mode decomposition integrated with a quantum differential quantum walks on a photonic chip. Phys Rev Lett. 2022;128
equation solver. Phys Rev Res. 2024;6(4). doi: 10.1103/Phys- (5):050503. doi: 10.1103/PhysRevLett.128.050503
RevResearch.6.043031
43. Sahai T, Speranzon A, Banaszuk A. Wave equation based
28. Newman ME. Spectral methods for community detection and algorithm for distributed eigenvector computation. In: 49th
graph partitioning. Phys Rev E–Stat Nonlin Soft Matter Phys. IEEE Conference on Decision and Control (CDC); 2010 Dec
2013;88(4):042822. doi: 10.1103/PhysRevE.88.042822 15–17; Atalanta, GA, USA. Atalanta (GA): IEEE; 2010. p.
7308–15.
29. Sahai T, Speranzon A, Banaszuk A. Hearing the clusters of a
graph: a distributed algorithm. Automatica. 2012;48(1):15– 44. Rebentrost P, Steffens A, Marvian I, Lloyd S. Quantum
24. doi: 10.1016/j.automatica.2011.09.019 singular-value decomposition of nonsparse low-rank matri-
ces. Phys Rev A. 2018;97:012327. doi: 10.1103/PhysRevA.97.
30. Zhu H, Klus S, Sahai T. A dynamic mode decomposition 012327
approach for decentralized spectral clustering of graphs. In:
2022 IEEE Conference on Control Technology and Appli- 45. Anthony M, Harvey M. Advanced linear algebra. Hanover
cations (CCTA); 2022 Aug 22–25; Trieste, Italy. Piscataway (PA): Springer; 2006.
(NJ): IEEE; 2022. p. 1202–7. 46. Saad Y. Numerical methods for large Eigenvalue problems:
31. Friedman J, Tillich JP. Wave equations for graphs and the revised edition. Philadelphia (PA): SIAM; 2011.
edge-based Laplacian. Pac J Math. 2004;216:229–66. doi: 47. Mackey L. Deflation methods for sparse PCA. In: Koller D,
10.2140/pjm.2004.216.229 Schuurmans D, Bengio Y, Bottou L, editors. Advances in
32. Tu J, Rowley C, Luchtenburg D, Brunton S, Kutz N. On neural information processing systems. Vol. 21. Vancouver
dynamic mode decomposition: theory and applications. ACM (BC): Curran Associates, Inc.; 2008.
J Comput Doc. 2013;1:391–421. doi: 10.48550/arXiv.1312. 48. Wiebe N, Braun D, Lloyd S. Quantum algorithm for data
0041 fitting. Phys Rev Lett. 2012;109:050505. doi: 10.1103/Phys-
33. Dylewsky D, Kaiser E, Brunton SL, Kutz JN. Principal compo- RevLett.109.050505
nent trajectories for modeling spectrally-continuous dynam- 49. Sun Z, Pedretti G, Ielmini D. Fast solution of linear sys-
ics as forced linear systems. College Park (MD): American tems with analog resistive switching memory (RRAM). In:
Physical Society; 2022. 2019 IEEE International Conference on Rebooting Comput-
34. Williams VV, Xu Y, Xu Z, Zhou R. New bounds for ma- ing (ICRC). San Mateo (CA): IEEE; 2019. p. 1–5.
trix multiplication: from alpha to omega. arXiv:2307.07970 50. Pan L, Zuo P, Luo Y, Sun Z, Huang R. BlockAMC: scalable in-
[Preprint]. 2023. doi: 10.48550/arXiv.2307.07970 memory analog matrix computing for solving linear systems.
arXiv:2401.10042 [Preprint]. 2024. doi: 10.48550/arXiv.24
35. Tagliacozzo L. Optimal simulation of quantum dynamics. Nat
01.10042
Phys. 2022;18(9):970–1. doi: 10.1038/s41567-022-01696-8
51. Nghiem NA, Wei TC. Quantum algorithm for estimating
36. Gnanasekaran A, Surana A, Sahai T. Efficient quantum algo-
largest eigenvalues. Phys Lett A. 2023;488:129138. doi:
rithms for nonlinear stochastic dynamical systems. In: 2023
10.1016/j.physleta.2023.129138
IEEE International Conference on Quantum Computing and
Engineering (QCE); 2023 Sep 17–22; Bellevue, WA, USA. 52. Shao C. Computing Eigenvalues of diagonalizable matrices on
Vol. 2. Piscataway (NJ): IEEE; 2023. p. 66–75. a quantum computer. ACM Trans Quantum Comput. 2022;3
(4):1–20. doi: 10.1145/3527845
37. Belkin M, Niyogi P. Laplacian Eigenmaps and spectral tech-
niques for embedding and clustering. In: Dietterich T, Becker 53. Lloyd S, Mohseni M, Rebentrost P. Quantum principal
S, Ghahramani Z, editors. Advances in neural information component analysis. Nat Phys. 2014;10(9):631–3. doi:
processing systems. Vol. 14. Vancouver (BC): MIT Press; 10.1038/nphys3029
2001.
54. Girvan M, Newman MEJ. Community structure in social and
38. Zhang XD, Zhang XM, Xue ZY. Quantum hyperparallel algo- biological networks. Proc Nat Acad Sci. 2002;99(12):7821–6.
rithm for matrix multiplication. Sci Rep. 2016;6:24910. doi: doi: 10.1073/pnas.122653799
10.1038/srep24910
55. Puzzuoli D, Wood CJ, Egger DJ, Rosand B, Ueda K. Qiskit
39. Costa PCS, Jordan S, Ostrander A. Quantum algorithm for dynamics: a python package for simulating the time dynamics
simulating the wave equation. Phys Rev A. 2019;99:012323. of quantum systems. J Open Source Softw. 2023;8(90):5853.
doi: 10.1103/PhysRevA.99.012323 doi: 10.21105/joss.05853

ACADEMIA QUANTUM 2025, 2 12 of 13

https://www.academia.edu/journals/academia-quantum/about https://doi.org/10.20935/AcadQuant7587

56. Fink CG, Omodt N, Zinnecker S, Sprint G. A congressional 59. GitHub. XingziXu/quantum-analog-clustering; [cited 2025
Twitter network dataset quantifying pairwise probability of Feb 11]. Available from: https://github.com/XingziXu/quan
influence. Data Brief. 2023;50:109521. doi: 10.1016/j.dib.20 tum-analog-clustering
23.109521
60. Sahai T. Dynamical systems theory and algorithms for NP-
57. Fink CG, Fullin K, Gutierrez G, Omodt N, Zinnecker S, Sprint hard problems. In: Junge O, Schütze O, Froyland G, Ober-
G, et al. A centrality measure for quantifying spread on Blöbaum S, Padberg-Gehle K, editors. Advances in dynam-
weighted, directed networks. Phys A. 2023;626:129083. doi: ics, optimization and computation: a volume dedicated to
10.48550/arXiv.2303.09684 Michael Dellnitz on the occasion of his 60th birthday. Cham:
Springer; 2020. p. 183–206.
58. Leskovec J, Mcauley J. Learning to discover social circles in
ego networks. In: Pereira F, Burges CJ, Bottou L, Weinberger 61. Sahai T, Gnanasekaran A. On the emergence of ergodic dy-
KQ, editors. Advances in neural information processing sys- namics in unique games. arXiv:2404.16024 [Preprint]. 2014.
tems. Vol. 25. Lake Tahoe (NV): Curran Associates, Inc.; 2012. doi: 10.48550/arXiv.2404.16024

ACADEMIA QUANTUM 2025, 2 13 of 13