Complex Vector Gain-Based Annealer for Minimizing XY Hamiltonians

James S. Cummins and Natalia G. Berloff∗

Department of Applied Mathematics and Theoretical Physics,
University of Cambridge, Wilberforce Road, Cambridge CB3 0WA, United Kingdom
(Dated: November 5, 2024)
This paper presents the Complex Vector Gain-Based Annealer (CoVeGA), an analog computing
platform designed to overcome energy barriers in XY Hamiltonians through a higher-dimensional
representation. Traditional gain-based solvers utilizing optical or photonic hardware typically rep-
resent each XY spin with a single complex field. These solvers often struggle with large energy
barriers in complex landscapes, leading to relaxation into excited states. CoVeGA addresses these
limitations by employing two complex fields to represent each XY spin and dynamically evolving
arXiv:2411.02010v1 [cond-mat.dis-nn] 4 Nov 2024

the energy landscape through time-dependent annealing. Operating in a higher-dimensional space,

CoVeGA bridges energy barriers in this expanded space during the continuous phase evolution, thus
avoiding entrapment in local minima. We introduce several graph structures that pose challenges
for XY minimization and use them to benchmark CoVeGA against single-dimension XY solvers,
highlighting the benefits of higher-dimensional operation.

I. INTRODUCTION portfolio optimization [24]. Most proposed physical ana-

log machines use one-dimensional (‘hard’ or ‘soft’) spins
The growing complexity and sheer scale of modern sci- as variables in the discrete binary Ising Hamiltonian HI =
entific and industrial computing tasks are pushing us to − ∑Ni,j Jij si sj with si = ±1, or the continuous XY Hamil-
look beyond traditional von Neumann architectures for tonian HXY = − ∑N i,j Jij si ⋅ sj , where si = (cos θi , sin θi ),
solutions to hard optimization problems. These architec- θi ∈ [0, 2π), and N is the number of spins.
tures, which dominate computing today, rely on a clear The XY model describes a system of spins constrained
separation between memory and processing and execute to rotate within a plane, each possessing a continu-
tasks in a step-by-step manner. While traditional com- ous degree of freedom characterized by a phase θi rel-
puting has been reliable for decades, it’s starting to fall ative to a fixed axis. Interactions between phases give
behind when it comes to the needs of today’s special- rise to a rich array of dynamics, including vortices and
ized applications—particularly those that demand high Berezinski-Kosterlitz-Thouless transitions [25]. Origi-
speed, energy efficiency, and scalability. As areas like nally developed for statistical mechanics and condensed
machine learning, big data analysis, and real-time pro- matter physics, the XY model has found extensive ap-
cessing continue to grow, the limitations of the von Neu- plications in various physical systems where rotational
mann approach have become more evident, creating a symmetry is essential. Notably, it applies to superfluid-
bottleneck that is increasingly difficult to overcome. This ity and superconductivity [26], cosmology [27], nematic
is where analog systems come in, offering a tailored ap- liquid crystals [28], magnetic nanoparticle ensembles [29],
proach to specific types of computing tasks, bypassing protein folding [30], and phase retrieval problems [31].
the constraints of conventional architectures. Gain-based minimizers utilize soft-spin bifurcation dy-
Physics-inspired analog machines have been proposed namics via Andronov-Hopf bifurcations to minimize spin
using various platforms such as superconducting qubits Hamiltonians [32]. The enhanced dimensionality offered
[1–3], optical parametric oscillators [4–7], memristors [8], by soft-spin models reduces energy barriers present in
lasers [9–11], photonic systems [12, 13], trapped ions classical hard-spin Hamiltonians by representing fixed
[14], polariton condensates [15, 16], photon condensates spin amplitudes as continuous variables [33]. We recently
[17], and surface acoustic waves [18]. These specialized proposed the Vector Ising Spin Annealer (VISA) as an
physical machines minimize programmable spin Hamil- Ising minimization model capable of overcoming obsta-
tonians, where the couplings between spins – given by cles in solving combinatorial optimization problems [34].
the interaction matrix J – are designed such that the By employing three soft modes to represent the vector
global minimum corresponds to the optimal solution of a components of an Ising spin, VISA bridges minima sep-
combinatorial optimization problem. Problems like num- arated by significant energy barriers in complex energy
ber partitioning, traveling salesman, graph coloring, spin landscapes. VISA uses real-valued soft spins, which can,
glass systems, knapsack problem, binary linear program- for example, represent the optical parametric oscillator
ming, graph partitioning, and Max-Cut can be mapped quadrature in coherent Ising machines.
to spin Hamiltonians [19–21]. Further applications in- In this paper, we propose to use multiple vector com-
clude machine learning [22], financial markets [23], and ponents to represent the spins in networks of complex-
valued fields ψi , which are known to minimize XY and
Ising Hamiltonians in gain-based systems [35]. The
Stuart-Landau equation that governs the dynamics of
∗ correspondence address: [email protected] one-dimensional complex oscillators ψi in gain-based net-
 2

works is given by spins are represented as complex-valued vectors in two-

dimensional vector space Ψi ∈ C2 . The increased dimen-
ψ̇i = (γi − ∣ψi ∣2 ) ψi + α ∑ Jij ψj , (1) sion over typical one-dimensional spins allows to effec-
j tively overcome the barriers between minima.
This paper proposes a new approach to minimizing XY
where we introduced a regulation parameter α > 0 and
Hamiltonians that utilizes the ultra-fast energy-efficient
the effective gain rate γi is individually dynamically ad-
architecture of photonics-based analog machines. In Sec-
justed through feedback mechanism
tion II, we formalize CoVeGA, and provide expressions
γ̇i = ε (1 − ∣ψi ∣2 ) , (2) for each term in its composite Hamiltonian. In Section
III, we use the Kuramoto model to investigate the dif-
which can be implemented via optical delay lines for laser ficulty of various XY minimization problems. This al-
systems, or by spatial light modulators for polariton con- lows us to identify suitably hard benchmark problems to
densates [16]; see also Appendix A. The dynamical equa- test CoVeGA and existing XY minimization algorithms,
tion (1) can be written as ψ̇i = −∂H/∂ψi∗ , which describes as recovering the global minimum of these problems is
the process of gradient descent to the minima of the loss nontrivial. Lastly, Section IV compares the dynamics
function of CoVeGA to the one-dimensional Stuart-Landau net-
work as well as other continuous-variable methods such
1 N 2 2 α N ∗ ∗
as spin-vector Langevin and Kuramoto models. We con-
H= ∑ (γi − ∣ψi ∣ ) − ∑ Jij (ψi ψj + ψi ψj ) , (3) trast these methods by finding ground and excited state
2 i=1 2 i,j
probabilities and illustrating the distribution of recovered
where γ̇i = 0 for all i at the threshold steady state. By states.
representing each oscillator in its polar form as ψi =
ri exp(iθi ), Eq. (1) can be decomposed into real and
II. COMPLEX VECTOR GAIN-BASED
imaginary parts to get equations of the time evolution ANNEALER
of the amplitude ri and the phase θi as

ṙi = γi ri − ri3 + α ∑ Jij rj cos (θi − θj ) , (4) The CoVeGA model operates through a system of
(1)
j N two-dimensional complex-field vectors Ψi = (ψi =
rj (1) (1) (2) (2) (2)
θ̇i = −α ∑ Jij sin (θi − θj ) . (5) ri eiθi , ψi = ri eiθi ). It utilizes annealing,
j ri symmetry-breaking bifurcation, gradient descent, and
mode selection to drive the system to the global min-
Starting from below the steady state threshold in the imum. The Hamiltonian is the sum of three terms
vacuum state ri = 0, all oscillators are pumped equally. H = H1 + αH2 + H3 , where
Then, depending on the structure of J, nonzero ampli-
tudes emerge at different rates for each oscillator as the 1 N 2 2
pumping intensity increases. The feedback mechanism H1 = ∑ (γi (t) − ∣∣Ψi ∣∣2 ) , (7)
2 i=1
of Eq. (2) adjusts each oscillator so that they all reach
1 N
equal amplitudes at the steady state threshold. H2 = − ∑ Jij (Ψi ⋅ Ψ∗j + Ψ∗i ⋅ Ψj ) , (8)
Only under the condition of equal amplitudes at the 2 i,j
steady state will Eq. (5) reach the minimum of the XY N
Hamiltonian. The sum of the steady states of Eq. (4) H3 = −P (t) ∑ (Ψj ⊙ Ψ∗i ) ⋅ [Q(Ψ∗j ⊙ Ψi )] . (9)
gives N = ∑N i=1 γi + α/2 ∑i,j Jij cos (θi − θj ), so the global
i,j=1
minimum of the XY model corresponds to the smallest
Here, ⊙ indicates element-wise multiplication, and Q is
effective injection ∑i γi . Close to the threshold, Eq. (5)
a 2 × 2 permutation matrix given by Q = ( 01 10 ). As the
becomes fully analogous to the Kuramoto model
effective gain γi (t) increases with time t from negative
N (effective losses) to positive values, H1 anneals between
θ̇i = −α ∑ Jij sin (θi − θj ) . (6) a convex function with minimum at ∣∣Ψi ∣∣22 = 0 for all i,
j=1 to nonzero amplitudes. Writing the complex vectors in
H2 using their polar coordinates gives
Equation (1) can be adapted to minimize Ising Hamilto-
nians by restricting the state space of the phase, which (1) (1) (1) (2) (2) (2)
H2 = − ∑ Jij (ri rj cos θij + ri rj cos θij ) , (10)
we detail in Appendix B. i,j
Gain-based systems described by Eq. (1) can still set- (k) (k) (k)
tle in local minima during amplitude bifurcation, which where we define θij ≡ θi −θj . Equation (10) is analo-
limits the probability of finding the global minimum. To gous to the second term on the right-hand side of Eq. (3),
combat this, we introduce the complex vector gain-based but now we have two terms corresponding to the two di-
annealer (CoVeGA) that exploits the advantages of ex- mensions in the complex vector space that CoVeGA op-
tended spatial dimensions. In this model, continuous XY erates in. The effective gain is subject to the feedback
 3

governed by γ̇i = ε(1 − ∣∣Ψi ∣∣22 ). This drives the effective

gain rates from below and brings the amplitudes to 1
at the steady state. Finally, H3 is a penalty term with
time-dependent magnitude P (t) that enforces agreement
of the phase differences in each dimension. This 4-local
penalty term, expressed in Eq. (9), can be expanded into
its constituent complex fields, which transforms H3 to
(1) (2)∗ (1)∗ (2)
H3 = −P (t) ∑ ψi ψi ψj ψj + c.c. (11)
i,j Figure 1. (a) Ground state solution, up to a global phase
change, for an N = 12 4-regular Möbius ladder graph. In this
When the complex oscillators have equal unit-valued am- ground state, every XY spin has a phase difference of ±2π/3
plitudes, ∣∣Ψi ∣∣22 = 1 for all i, H3 becomes with each of its nearest neighbors. (b) Excited state of an
N = 102 4-regular Möbius ladder graph with D = 4. Here,
(1) (2) the phases are not illustrated, and instead color represents
H3 = −2P (t) ∑ cos (θij − θij ) , (12)
i,j
the chilarity of each triangular base. For a base with indices
{i, j, k}, then Cijk = −1 is shown in yellow, Cijk = 1 as red,
(1) (2) and Cijk = 0 as green. (c) Frequency density histogram of
which is minimized when θij = θij for all pairs (i, j).
excitation parameter D for 1000 runs of the Kuramoto model
As P (t) increases from P (0) = 0 to sufficiently large on N = {12, 102, 204} 4-regular Möbius ladder graphs. In each
(1) (2)
P (T ) > 0 at t = T , the phase differences θij and θij run, initial phases θi (0) are chosen uniformly at random from
become equal. Therefore, up to a global phase off- range [−π, π), and Eq. (6) is solved using the Euler scheme
(1) (2) with fixed time step ∆t = 0.1.
set, θi = θi . At the end of the annealing proto-
col, the target XY Hamiltonian HXY is minimized. We
evolve the CoVeGA Hamiltonian H using gradient de-
scent while simultaneously annealing parameters γi and reverse its chirality, representing a significant energy bar-
P (t). Therefore, for each oscillator i, the governing equa- rier to overcome. Therefore, in this case, local perturba-
tions are not enough to bridge local and global minima.
tion Ψ̇i = −∇i H is given by
N
Ψ̇i = Ψi (γi (t) − ∣∣Ψi ∣∣22 ) + α ∑ Jij Ψj III. GRAPHS FOR XY MINIMIZATION
j=1
(13)
N
To rigorously test CoVeGA’s capabilities, we select
+ P (t) ∑ Ψj ⊙ [Q(Ψ∗j ⊙ Ψi )] .
j=1
graph types that, due to their complex topologies and
high energy barriers, present distinct challenges for
The operation of CoVeGA, therefore, relies on the gra- gradient-based methods. We consider network graph
dient descent of an annealed energy landscape. The structures given by the coupling matrix J, resulting in
Hamiltonian H is 4-local due to the H3 penalty term. nontrivial minimization instances for simple gradient-
While a 4-local Ising Hamiltonian can be relaxed to a based solvers. To do this, we represent the gradient dy-
quadratic function without loss of generality [36], sub- namics by the Kuramoto model (6), as it represents a
ject to an overhead corresponding to the introduction of gradient descent with respect to XY phases θi . Moreover,
N ⌈k/2⌉ auxiliary variables, this is not the case for XY by using the Kuramoto model, we can infer the sizes of
Hamiltonians. In general, a 4-local XY Hamiltonian can the basins of attraction for each locally stable state. This
not be directly mapped to a 2-local Hamiltonian. How- is because, under gradient descent, the system evolves to
ever, some optical hardware can directly encode such the closest minima.
high-order interactions for XY systems [37]. 4-Regular Möbius Ladder
In the next section, we seek suitable graph structures The 4-regular Möbius ladder network is an unweighted
for benchmarking CoVeGA and alternate algorithms for 4-regular circulant graph with antiferromagnetic cou-
XY Hamiltonian minimization. For benchmarking, we plings Jij ∈ {−1, 0} between its N vertices. It is inspired
choose graphs with analytically tractable yet nontrivial by its 3-regular version, often used as an Ising solver
ground states and energies and whose basins of attrac- benchmark [33, 34, 41]; see Appendices C and D. We
tion are small in volume. In this way, simple gradient- only consider cases for which N /2 is divisible by 3, in
based algorithms are expected to falter in recovering the which case the ground state has no frustrations, and any
global minimum. Moreover, we seek technologically feasi- connected spin pair has phase difference ±2π/3. Then,
ble graphs on analog hardware [38–40]. Complex graphs the graph, illustrated in Figs. (1)(a) and (b) consist of N
for optimization may contain topological structures re- triangular bases connected to each other, with the first
sistant to simple local perturbations. In the XY regime, and last bases joined, creating a periodic circular geom-
these often arise as domain boundaries or vortices. For etry. For any triangle with nodes i, j, k, we define the
domain boundaries, the transformation from the excited chirality Cijk ∈ {0, ±1} depending on whether there is a
state to the ground state requires an entire domain to phase winding around these nodes and in which direc-
 4

tion. The ground state corresponds to alternating chiral-

ity between adjacent triangles. Therefore, we define the
excitation parameter D for any 4-regular Möbius ladder
graph as

D= ∑ (1 − ∣Cijk ∣) , (14)
{ijk}∈T

to account for excitations that occur as domain bound-

aries, where T is the set of vertices for each triangle. This Figure 2. (a) A ground state solution for an N = 4 × 4 trian-
is where triangular bases with zero chirality separate do- gular lattice graph, where every XY spin has a phase differ-
ence of ±2π/3 with each of its neighbors. (b) Excited state
mains with triangular bases of alternating chirality ±1.
of an N = 10 × 10 triangular lattice graph. Here, the phases
The frequency density histogram of excitation param- are not illustrated, and instead color represents the chiral-
eter D for states recovered by the Kuramoto model (6) ity of each triangular base. For a base with indices {i, j, k},
is shown in Fig. (1)(c). Since the system descends to the then Cijk = −1 is show in yellow, Cijk = +1 as red, and
closest minimum to the initial state according to Eq. (6), Cijk = 0 as green. Domains are more likely to exist when
the histogram represents the volumes of the basins of at- part of their boundaries are the edges of the triangular lattice
traction for each minima. Figure (1)(c) shows that as the since, on these edges, there are no excited (green) triangu-
system size N increases, the volume of the basin of at- lar bases. This is a consequence of the graph’s non-periodic
traction of the ground state decreases as a proportion of boundary conditions. (c) Frequency density histogram of ex-
the total volume of all basins, and recovering the ground citation parameter D for 1000 runs of the Kuramoto model on
N = {16, 100, 196} triangular lattice graphs. In each run, ini-
state using standard gradient descent becomes hard.
tial phases θi (0) are chosen uniformly at random from range
Triangular Lattice [−π, π), and Eq. (6) is solved using the Euler scheme with
The triangular lattice is a two-dimensional graph con- fixed time step ∆t = 0.1.
sisting of triangles in a regular arrangement. Each of the
N vertices corresponds to an XY spin, while edges rep-
resent unweighted antiferromagnetic couplings Jij = −1.
Here, the parameter 0 ≤ p ≤ 1 in Eq. (15) controls the
The ground state is recognized by its arrangement of
rank of J and influences the hardness of the XY mini-
phases, with every XY spin having phase difference ±2π/3
mization problem, as shown in Figs. (3)(b) and (c). As p
with each of its nearest neighbors. Similar to 4-regular
increases, the number of nonzero entries in J decreases,
Möbius ladder graphs, we can use the same excitation
and consequently, the rank decreases. In spatial pho-
parameter D from Eq. (14) to categorize ground and ex-
tonic XY and Ising machines, the rank of the coupling
cited states. The frequency density histogram of excita-
matrix is an important feature of the spin network; in-
tion parameter D for states recovered by the Kuramoto
deed, some current implementations are only feasible for
model (6) is shown in Fig. (2)(c). As the system size
low-rank coupling matrices [24].
N increases, the number of trials in the histogram bin
corresponding to the ground state D = 0 decreases.
Basic Kuratowskian
Basic Kuratowskian graphs are non-planar graphs
that serve as fundamental structures in graph theory;
they are sub-graphs of every two-dimensional or three-
dimensional non-planar graph. For edges with ran-
dom weights −1, 0, +1, the Ising problem on basic Kura-
towskian graphs is known to be NP-complete [42]; how-
ever, the analysis of these graphs in the continuous XY
spin regime has not been extensively studied. Unlike 4- Figure 3. (a) A configuration of XY phases on an N = 4 × 4
regular Möbius ladder and triangular lattice graphs, the basic Kuratowskian graph with p = 0.2. Ferromagnetic, anti-
ground state configuration and energy are not known a ferromagnetic, and zero couplings are illustrated as red, blue,
priori due to the random coupling weights. and black lines, respectively. (b) Rank of coupling matrix J
To construct the coupling matrix J for a basic Ku- as a function of the probability parameter p. Each error bar
in (b) corresponding to a different value of p is constructed
ratowskian graph, we assign each edge in Fig. (3)(a) a
from 100 Kuratowskian graphs, each with random coupling
weight randomly chosen from W = {−1, 0, +1}. We sam- weights generated by Eq. (15). (c) Box plot distributions of
ple the weights w from the discrete probability mass func- sample variance values s2 from final state energies of Eq. (6),
tion pW ∶ W → [0, 1], where obtained from N = 8 × 8 basic Kuratowskian graphs. Each
box plot in (c) is obtained by sampling the final state ener-
⎧ 1
⎪
⎪ (1 − p) if w = −1, gies over 100 runs on each of the 100 Kuratowskian graphs
⎪2
⎪
generated for each value of p.
pW (w) = ⎨p if w = 0, (15)
⎪
⎪
⎪ 1
⎩ 2 (1 − p)
⎪ if w = +1.
 5

IV. RESULTS

In this section, we compare CoVeGA to the one-

dimensional spin version, whose Hamiltonian is given by
Eq. (3) and governed by the canonical Stuart-Landau
network Eq. (1), as well as to Kuramoto gradient de-
scent and other methods. The phase annealing protocol
of CoVeGA begins with P (0) = 0 and gradually increases
for t > 0, ensuring that the local phase difference in each
dimension of the complex-field vector Ψi becomes the
same. We increase P (t) to a sufficiently large value, so
(2)
that θij (1) = θij holds for every pair (i, j), after which
we choose one dimension from which to extract the XY
(1)
phases. Under this constraint, each θi is equivalent
(2)
to θi up to a global phase shift, and hence the choice
of dimension is arbitrary. For simplicity, we choose a
monotonically increasing function P (t) that has a linear
dependence on time, such that P (t) = βt, although other
options may be chosen.
Figure (4) compares CoVeGA with the one-
dimensional Stuart-Landau model of Eq. (3) under
equivalent starting conditions on a 4-regular Möbius
ladder graph with N = 36. The figure demonstrates that Figure 4. Panels (a) and (b) compare the trajectories of XY
while the scalar version fails to find the ground state spins in the CoVeGA model and the single-dimension Stuart-
(D = 0), CoVeGA leverages multidimensional phase Landau model, respectively, for an N = 36 4-regular Möbius
ladder graph. Red circles indicate initial states, while in (a),
dynamics to recover the global minimum by bridging (1)
minima unreachable by the one-dimensional approach. blue and cyan circles correspond to final states for ψi and
(2)
Additionally, the CoVeGA mechanism allows for phase ψi . Final states ψi in the scalar version (b) are highlighted
changes at lower energy costs than required by the with blue circles. Panels (c) and (d) illustrate the effective
gain γi as the systems evolve. CoVeGA successfully recovers
scalar version, due to its ability to navigate through the
the ground state without frustrations D = 0 as seen in in-
multidimensional space to find the most energy-efficient set (e), while the one-dimensional scalar version reaches the
path to the global minimum. excited state D = 2 shown in inset (f). Both systems start
In addition, we compare CoVeGA to the spin-vector from equivalent initial conditions, with the one-dimensional
Langevin (SVL) model that was proposed as a clas- scalar version beginning at ψi (0) = 0.01 exp(ia) and CoVeGA
sical analog of a quantum annealing description us- at Ψi (0) = 0.01(exp(ia), 0.1 exp(ib)), where a and b are uni-
ing stochastic Langevin time evolution governed by the formly chosen from range [−π, π). Panel (g) compares the
fluctuation-dissipation theorem [44]. SVL is based on the CoVeGA Hamiltonian H against the one-dimensional model
time-dependent Hamiltonian used in quantum anneal- Eq. (3), while panel (h) shows the values of the XY Hamilto-
ing H(t) = A(t)H0 + B(t)HP , where the initial Hamil- nians HXY .
tonian is H0 = − ∑i σix . We choose the problem Hamil-
tonian so that it maps to the XY Hamiltonian as HP = and ξi (t) is an iid Gaussian noise. For quadratic uncon-
− ∑i,j Jij (σix σjx + σiz σjz ), with Pauli operator σi acting on strained continuous optimization, such as minimizing the
the i-th variable. Real annealing functions satisfy bound- XY model, the gradient term in Eq. (16) is
ary conditions A(0) = B(T ) = 1 and A(T ) = B(0) = 0,
where T is the temporal length of the annealing sched- ∂H(θ)
ule. If the rate of change of the functions is slow enough, = A(t) sin θi + B(t) ∑ Jij sin(θi − θj ), (17)
∂θi j=1
the system stays in the ground state of the instantaneous
Hamiltonian so that at t = T the XY Hamiltonian is mini- which in conjunction with fluctuation-dissipation rela-
mized. The SVL model replaces Pauli operators with real tions ⟨ξi (t)⟩ = 0 and ⟨ξi (t)ξj (t′ )⟩ = δij δ(t − t′ ), give 2N
functions of continuous angle σiz → sin θi , σix → cos θi , stochastic differential equations: dθi = (pi /m)dt and
and is therefore a classical annealing Hamiltonian using
continuous-valued phases θi . SVL dynamics is described ∂H(θ) b
dpi = ( + pi ) dt + dWi , (18)
by a system of coupled stochastic equations ∂θi m
∂H(θ) where dWi represents a real-valued continuous-time
mθ̈i + bθ̇i + + ξi (t) = 0, (16) stochastic Wiener process [44].
∂θi
CoVeGA distinguishes itself from one-dimensional
where m is the effective mass, b is the damping constant, Stuart-Landau networks, SVL, and the Kuramoto model
 6

Figure 5. The probability distribution of recovering a state Figure 6. The probability distribution of recovering a state
with excitation parameter D for CoVeGA, one-dimensional with excitation parameter D for CoVeGA, one-dimensional
Stuart-Landau, SVL, Kuramoto, and BFGS methods on N = Stuart-Landau, SVL, Kuramoto, and BFGS methods on N =
12 (blue), 102 (orange), and 204 (green) 4-regular Möbius 16 (blue), 100 (orange), and 194 (green) triangular lattice
ladder graphs. One thousand runs are used to calculate the graphs. One thousand runs starting from random initial con-
probability of recovering excitation parameters D for each ditions are used to calculate the probability for each method
value of N . Parameter values for CoVeGA are taken from and for each value of N .
Table (I) in Appendix E [43], and so are values for the Stuart-
Landau network solver where applicable. For SVL, m = 1.0,
ξ ∼ N (0, 0.1), and γ = 0.9. In all cases, a fixed time step of imity gap metric to evaluate performance. The proximity
∆t = 0.1 is used with annealing time length T = 1000. gap provides a measure of the distance between any solu-
tion state and the best-known state by comparing their
objective values. More precisely, it is defined as the ra-
through its gain-based annealing strategy in higher- tio of the objective value obtained by a given method to
dimensional spaces. In Figs. (5) and (6), we compare the best objective value found among all methods. Here,
CoVeGA to these models and the Broyden-Fletcher- the objective value is given by the XY Hamiltonian HXY .
Goldfarb-Shanno (BFGS) algorithm on 4-regular Möbius When comparing optimization methods on basic Kura-
ladder graphs and triangular lattice graphs, respectively. towskian graphs, we choose p = 0.1, the hardest value of
BFGS is a classical quasi-Newton method for uncon- p, which is the probability that a coupling is set to zero.
strained nonlinear optimization that approximates the From Fig. (3)(f), we see that this occurs when p = 0.1,
Hessian matrix to guide the search for a local minimum. where the average sample variance s2 of Kuramoto runs
By leveraging a multidimensional approach, the CoV- is maximized.
eGA model recovers the ground state (D = 0) of 4-regular Figure (7)(a) compares the objective values achieved
Möbius ladder and triangular lattice graphs more often by CoVeGA with those obtained by one-dimensional
than any of the compared methods. As a result, the dis- Stuart-Landau, SVL, Kuramoto, and BFGS methods
tribution of excitations D for CoVeGA is skewed closer for N = 64 and N = 144 basic Kuratowskian graphs.
to D = 0 for all tested values of N , compared to the We define the quality improvement of CoVeGA over
single-dimension Stuart-Landau network, and this effect another method X in terms of objective values O as
becomes more pronounced as N increases. While the an- (OX − OCoVeGA )/OCoVeGA , showcasing these metrics in
nealing schedules used in these approaches and in SVL Fig. (7)(b), where X represents the best-performing com-
demonstrate efficient convergence to low-energy solutions peting method for each instance.
compared to Kuramoto and BFGS, CoVeGA’s higher di- These results demonstrate that CoVeGA consis-
mensionality enables XY spins to traverse paths connect- tently achieves ground states more reliably than one-
ing minima and converge to the ground state. This en- dimensional models, particularly in larger and more com-
hances system robustness, making the final system less plex systems.
sensitive to initial conditions.
Next, we consider basic Kuratowskian graphs, which
exhibit significant difficulty for standard gradient descent V. CONCLUSIONS
methods, as illustrated in Fig. (3) for small values of the
parameter p. However, the randomness of the couplings This paper presents the Complex Vector Gain-Based
can introduce statistical convergence issues in small sam- Annealer (CoVeGA). This approach combines multidi-
ple sizes. Despite this, graphs with random couplings are mensional continuous spin systems, gain-based oper-
popular for benchmarking physical simulators [45–47]. ations, and soft-amplitude annealing to optimize XY
Because basic Kuratowskian graph instances do not Hamiltonians on challenging graph structures. By ex-
have analytically known ground states, we use the prox- ploiting higher-dimensional spaces, CoVeGA improves
 7

with reservoir dynamics ṅR = −(b0 +b1 ∣ψi ∣2 )nR +Pi , where
g, b0 , and b1 are dimensionless parameters, and Pi is
the pumping intensity. Networks of oscillators are con-
structed by coupling multiple lasers or condensates. In
this case, using the tight-binding approximation from
the mean-field complex Ginzburg-Landau equation, or
the mean-field Maxwell-Bloch equations for laser cavi-
ties, Eq. (A1) becomes
⎡ ⎤
⎢ ⎥
ψ̇i = −i∣ψi ∣2 ψi − ψi + (1 − ig) ⎢Ri ψi + α ∑ Jij ψj ⎥ , (A2)
⎢ j≠i
⎥
⎣ ⎦

where Ri is the i-th reservoir density that follows dynam-

ics Ṙi = b0 (γi − Ri − ξRi ∣ψi ∣2 ), where ξ = b1 /b0 , γi is the
Figure 7. (a) Proximity gap box plots for each method effective pumping rate, α quantifies the overall strength
{CoVeGA, Stuart-Landau (SL), SVL, Kuramoto, BFGS} on of coupling, and Jij is the individual coupling strength
the same 200 instances of basic Kuratowskian graphs with between the i-th and j-th oscillators. In the fast reservoir
random bimodel weights, with instances divided equally be- relaxation limit, b0 ≫ 1 and the reservoir dynamics can
tween graph sizes N = 64 and N = 144. In all cases, the be replaced with its steady-state Ri = γi /(1 + ξ∣ψi ∣2 ) to
matrix weight elements are drawn from the discrete probabil- lead to a Stuart-Landau network (1) of coupled oscilla-
ity mass function given by Eq. (15) with p = 0.1. (b) Violin tors. For lasers, the nonlinear self-interactions given by
plots demonstrate the distribution of CoVeGA’s quality im- the first term on the right-hand side of Eq. (A2) are zero.
provement performance compared to the best solution found
by any competing methods across the N = 64 and N = 144
basic Kuratowskian graph instances.
Appendix B: From XY to Ising in Gain-Based
Solvers
phase mobility, enabling it to navigate complex energy
landscapes more effectively, overcome energy barriers, To obtain the minimum of the Ising Hamiltonian us-
and avoid local minima. This approach allows CoVeGA ing Eq. (1), a mechanism is required to restrict phases
to reach optimal solutions with greater efficiency and ac- to θi ∈ {0, π}. This may be achieved in many ways, for
curacy compared to one-dimensional methods. instance, by (i) coupling the real parts of the field in
CoVeGA paves the way for the development of con- the coupling term of Eq. (1) as Jij (ψj + ψj∗ ). Then the
tinuous analog optimization machines. Applying dimen- phases θi will automatically be projected onto 0 or π
sionality annealing techniques in optical-based spin ma- [32]. Or, by (ii) externally forcing the system at a fre-
chines suggests a promising future for ultra-fast compu- quency resonant with the frequency of the coherent state
tation and accurate ground state recovery, representing such that h(t)ψ ∗(q−1) is added to the right-hand side of
a significant advancement in optimization technologies. Eq. (1), where h(t) is the real time-dependent magni-
tude of the resonant pumping term. The presence of this
external forcing produces q solutions whose phases differ
by 2π/q [35]. For q > 2, the network emulates the q-state
ACKNOWLEDGEMENTS
Potts model with phases restricted to discrete values that
are multiples of 2π/q. Minimizing the Ising Hamiltonian,
J.S.C. acknowledges the PhD support from the therefore, requires taking q = 2. For sufficiently large
EPSRC. N.G.B. acknowledges the support from the h(t), a feasible Ising solution is obtained by penalizing
HORIZON EIC-2022-PATHFINDERCHALLENGES-01 phases that are not equal to 0 or π.
HEISINGBERG Project 101114978 and Weizmann-UK
Make Connection Grant 142568.
Appendix C: Möbius Ladder Graphs for XY
Networks
Appendix A: Complex Ginzburg-Landau Equation
Möbius ladder graphs are cyclic graphs with an even
Gain-based models that describe the dynamics of com- number of vertices. They can be visualized as nodes
plex oscillators ψi are based on physical implementations arranged in a ring with antiferromagnetic circle cou-
of laser or non-equilibrium condensate networks coupled plings Jij = −1 between nearest neighbors, and variable
to a reservoir nR . The systems are described by the gen- cross-circle couplings Jij = −J between diametrically op-
eralized complex Ginzburg-Landau equation posite vertices. At some critical value J = Jcrit , the
ground state configuration changes. We denote these
iψ̇i = −∇2 ψi + ∣ψi ∣2 ψi + gnR ψi + i(nR − 1)ψi , (A1) states as S0 and S1 , respectively. Notably, below and
 8

Appendix D: Circulant Graphs

The weighted adjacency matrix J of a 4-regular

Möbius ladder graph is circulant because it is con-
structed through cyclical permutations of any N -vector.
The graph inherently has vertex permutation symmetry,
signifying boundary periodicity and uniform neighbor-
hoods. Any circulant matrix can be expressed as a poly-
nomial of a shift matrix P. For 4-regular Möbius ladder
graphs
J = −P − P2 − PN −2 − PN −1 , (D1)
where P is the N × N canonical shift matrix P = ( 01 0I ),
and here I is an (N − 1) × (N − 1) identity matrix. The
structure of a circulant matrix is contained in any row,
Figure 8. (a) Eigenvalues of an N = 8 Möbius ladder graph and its eigenvalues and eigenvectors can be analytically
as a function of J. Je , illustrated by the vertical red dashed, derived using the N roots of unity of a polynomial ω =
is the point at which the two largest eigenvalues cross. XY exp(2πi/N ), where the row components c of the matrix
(0) (1)
ground states correspond to SXY for J < Jcrit and SXY for J act as coefficients. The eigenvectors of J are the same
J > Jcrit , where Jcrit = Je for the XY model. XY spin config- as the shift matrix P while the N eigenvalues λn are the
(0) (1)
urations for (b) SXY , and (c) SXY . components of the product Fc, where F is the N × N
Fourier matrix. This gives the eigenvalues as
λn = −ω n − ω 2n − ω (N −2)n − ω (N −1)n
(D2)
= −4 cos (2πn/N ) cos (3πn/N ) ,
for n = 0, 1, . . . , N − 1. The two indices n corresponding
to the largest positive eigenvalues are given by
above J = Je = 1 − cos(2π/N ), the leading eigenvalue of
the coupling matrix J changes from S0 to S1 , where for 3N ± N ± 6 N − 12
n= ±⌊ ⌋, (D3)
Ising spins Jcrit ≠ Je . Ising Hamiltonian minimization on 6 24
Möbius ladder graphs is known to be computationally dif- where N /2 is divisible by 3 and hence N is a factor of
ficult for certain parameter regimes [33]. Specifically, for 6. The eigenvectors of J form an orthogonal basis, and
cross-circle couplings Je < J < Jcrit , many physically in- since λn ∈ R and J ∈ RN ×N , then we can always choose
spired optimization algorithms fail to recover the ground the eigenvectors to be real. In this case, the eigenvectors
state with high probability [34]. This originates from the corresponding to the eigenvalues with index n are
behaviour of J, where for Je < J < Jcrit , the eigenvector
of the principal eigenvalue does not correspond to the ⎛ 2 ⎞ ⎛ 1 ⎞
n −n cos( 2πn
Ising ground state. Figure (8)(a) illustrates how these ⎜ ω + ω ⎟ ⎜ N
) ⎟
1⎜ ⎟ ⎜ ⎟
eigenvalues vary as a function of J. qn = ⎜ ω 2n + ω −2n ⎟=⎜ cos( 4πn
N
) ⎟ ⎟ , (D4)
2⎜ ⎟ ⎜
⎜ ⎟
⎜ ⋮ ⎟ ⎜ ⋮ ⎟
⎝ω (N −1)n −(N −1)n ⎠ ⎝cos( 2(N −1)πn
+ω N
)⎠
Similar to the Ising model, in the continuous XY where the eigenvectors corresponding to the largest eigen-
regime, there are two distinct ground states as J is var- values are 4-cyclic vectors (1, 0, −1, 0, 1, 0, . . . , −1, 0). In-
(0) (1)
ied; namely SXY for J < Jcrit and SXY for J > Jcrit . deed, this is the ground state solution of the 4-regular
(0) (1)
These have Hamiltonians HXY = (J − 2)N /2 and HXY = Möbius ladder graph when you associate each compo-
[2 cos((N − 2)π/N ) − J]N /2 respectively, and are shown nent of the eigenvector with the cosine of the phase.
in Figs. (8)(b)-(c). The critical value Jcrit at which these Graphs with circulant coupling matrices are realizable in
energies intercept is Jcrit = 1−cos(2π/N ). This is exactly current experimental platforms [38–40]. This, combined
equal to Je – the point at which the leading eigenval- with the accessibility to the analytical energy spectrum
ues intercept one another. Therefore in the XY regime, they allow, makes them natural platforms for analyzing
Je = Jcrit , and the eigenvectors of the principal eigen- and contrasting properties of different platforms.
values correspond to the ground state solutions overall
J. We conclude that Möbius ladder graphs are compu-
tationally simple in the XY model. This motivates the Appendix E: Materials and Methods
use of other graphs for benchmarking XY minimizers.
Specifically, in Section III we utilize frustration and the In all comparisons between methods presented in this
promotion of domain boundaries to devise suitably hard paper, numerical integration is performed by the fourth-
graphs. order Runge-Kutta scheme with discrete time step ∆t =
 9

0.1. For CoVeGA, a linear annealing schedule is cho- N (0, 0.1). The Kuramoto model is implemented with-
sen such that P (t) = βt. Each time evolution is to a out hyperparameters, given by Eq. (6) with α = 1. The
maximum time T = 1000. The total parameter space BFGS algorithm is ran with the SciPy library using the
for CoVeGA is {ε, α, β, γ(0)}, with optimal choices de- scipy.optimize.minimize function.
duced by the Python machine-learner online optimiza-
tion package (M-LOOP) [43]. The hyperparameter α is Graph Type N ε α′ β γ(0)
scaled by the inverse of the largest positive eigenvalue of 12 0.032 2.882 0.020 −0.139
coupling matrix J such that α = α′ /λmax , and the initial 4-Regular Möbius 102 0.022 2.402 0.003 −0.453
effective gain is γi (0) = γ(0) for all i. For a selection
of graph sizes for 4-regular Möbius ladder graphs, tri- 204 0.002 2.529 0.008 −1.235
angular lattice graphs, and basic Kuratowskian graphs, 16 0.045 1.129 0.018 −0.900
the optimal parameter choices are detailed in Table (I). Triangular 100 0.034 2.280 0.004 −0.526
Since the value of λmax varies for random instances of 196 0.010 1.301 0.005 −0.986
basic Kuratowskian graphs, the values of α′ in Table (I) 16 0.047 1.148 0.016 −0.295
for this case are chosen as the best over a sample of 100 Kuratowskian 64 0.023 2.763 0.007 −1.275
p = 0.1 graph instances.
144 0.012 2.745 0.014 −1.285
For results obtained using the one-dimensional Stuart-
Landau network given by Eq. (1), parameters ε, α, and Table I. Optimal sets of hyperparameters for different graphs
γ(0) are taken from Table (I). For SVL we fix m = 1.0, and sizes.
b = 0.9, and the Gaussian noise ξ is sampled from

[1] M. W. Johnson, M. H. Amin, S. Gildert, T. Lanting, son, V. Čeperić, et al., Nature Communications 11, 249
F. Hamze, N. Dickson, R. Harris, A. J. Berkley, J. Jo- (2020).
hansson, P. Bunyk, et al., Nature 473, 194 (2011). [14] K. Kim, M.-S. Chang, S. Korenblit, R. Islam, E. E. Ed-
[2] V. S. Denchev, S. Boixo, S. V. Isakov, N. Ding, R. Bab- wards, J. K. Freericks, G.-D. Lin, L.-M. Duan, and
bush, V. Smelyanskiy, J. Martinis, and H. Neven, Phys- C. Monroe, Nature 465, 590 (2010).
ical Review X 6, 031015 (2016). [15] N. G. Berloff, M. Silva, K. Kalinin, A. Askitopoulos,
[3] F. Arute, K. Arya, R. Babbush, D. Bacon, J. C. Bardin, J. D. Töpfer, P. Cilibrizzi, W. Langbein, and P. G.
R. Barends, R. Biswas, S. Boixo, F. G. Brandao, D. A. Lagoudakis, Nature Materials (2017).
Buell, et al., Nature 574, 505 (2019). [16] K. P. Kalinin, A. Amo, J. Bloch, and N. G. Berloff,
[4] P. L. McMahon, A. Marandi, Y. Haribara, R. Hamerly, Nanophotonics 9, 4127 (2020).
C. Langrock, S. Tamate, T. Inagaki, H. Takesue, S. Ut- [17] M. Vretenar, B. Kassenberg, S. Bissesar, C. Toebes, and
sunomiya, K. Aihara, et al., Science 354, 614 (2016). J. Klaers, Physical Review Research 3, 023167 (2021).
[5] T. Inagaki, Y. Haribara, K. Igarashi, T. Sonobe, S. Ta- [18] A. Litvinenko, R. Khymyn, R. Ovcharov, and
mate, T. Honjo, A. Marandi, P. L. McMahon, T. Umeki, J. Åkerman, arXiv preprint arXiv:2311.06830 (2023).
K. Enbutsu, et al., Science 354, 603 (2016). [19] A. Lucas, Frontiers in Physics 2, 5 (2014).
[6] Y. Yamamoto, K. Aihara, T. Leleu, K.-i. Kawarabayashi, [20] M. Yamaoka, C. Yoshimura, M. Hayashi, T. Okuyama,
S. Kako, M. Fejer, K. Inoue, and H. Takesue, npj Quan- H. Aoki, and H. Mizuno, IEEE Journal of Solid-State
tum Information 3, 1 (2017). Circuits 51, 303 (2015).
[7] T. Honjo, T. Sonobe, K. Inaba, T. Inagaki, T. Ikuta, [21] K. Tanahashi, S. Takayanagi, T. Motohashi, and
Y. Yamada, T. Kazama, K. Enbutsu, T. Umeki, R. Kasa- S. Tanaka, Journal of the Physical Society of Japan 88,
hara, et al., Science Advances 7, eabh0952 (2021). 061010 (2019).
[8] F. Cai, S. Kumar, T. Van Vaerenbergh, X. Sheng, R. Liu, [22] A. Momeni, B. Rahmani, B. Scellier, L. G. Wright, P. L.
C. Li, Z. Liu, M. Foltin, S. Yu, Q. Xia, et al., Nature McMahon, C. C. Wanjura, Y. Li, A. Skalli, N. G. Berloff,
Electronics 3, 409 (2020). T. Onodera, et al., arXiv preprint arXiv:2406.03372
[9] M. Babaeian, D. T. Nguyen, V. Demir, M. Akbulut, (2024).
P.-A. Blanche, Y. Kaneda, S. Guha, M. A. Neifeld, [23] M. Gilli, D. Maringer, and E. Schumann, Numerical
and N. Peyghambarian, Nature Communications 10, 1 Methods and Optimization in Finance (Academic Press,
(2019). 2019).
[10] V. Pal, S. Mahler, C. Tradonsky, A. A. Friesem, and [24] R. Z. Wang, J. S. Cummins, M. Syed, N. Stroev,
N. Davidson, Physical Review Research 2, 033008 (2020). G. Pastras, J. Sakellariou, S. Tsintzos, A. Askitopou-
[11] M. Parto, W. Hayenga, A. Marandi, D. N. los, D. Veraldi, M. C. Strinati, et al., arXiv preprint
Christodoulides, and M. Khajavikhan, Nature Ma- arXiv:2406.01400 (2024).
terials 19, 725 (2020). [25] J. M. Kosterlitz and D. J. Thouless, in Basic Notions Of
[12] D. Pierangeli, G. Marcucci, and C. Conti, Physical Re- Condensed Matter Physics (CRC Press, 2018) pp. 493–
view Letters 122, 213902 (2019). 515.
[13] C. Roques-Carmes, Y. Shen, C. Zanoci, M. Prabhu, [26] P. Minnhagen, Reviews of Modern Physics 59, 1001
F. Atieh, L. Jing, T. Dubček, C. Mao, M. R. John- (1987).
 10

[27] W. H. Zurek, in Formation and Interactions of Topo- vokin, N. Berloff, and P. Lagoudakis, Physical Review
logical Defects: Proceedings of a NATO Advanced Study X 6, 031032 (2016).
Institute on Formation and Interactions of Topological [39] M. C. Strinati, D. Pierangeli, and C. Conti, Physical
Defects (Springer, 1995) pp. 349–378. Review Applied 16, 054022 (2021).
[28] P.-G. De Gennes and J. Prost, The Physics of Liquid [40] A. B. Ayoub and D. Psaltis, Scientific Reports 11, 1
Crystals, 83 (Oxford University Press, 1993). (2021).
[29] D. Gallina and G. Pastor, Physical Review X 10, 021068 [41] K. P. Kalinin and N. G. Berloff, arXiv preprint
(2020). arXiv:2008.00466 (2020).
[30] K. A. Dill, S. B. Ozkan, M. S. Shell, and T. R. Weikl, [42] S. Istrail, in Proceedings of the thirty-second annual ACM
Annu. Rev. Biophys. 37, 289 (2008). symposium on Theory of computing (2000) pp. 87–96.
[31] E. J. Candes, Y. C. Eldar, T. Strohmer, and V. Voronin- [43] P. B. Wigley, P. J. Everitt, A. van den Hengel, J. W.
ski, SIAM Review 57, 225 (2015). Bastian, M. A. Sooriyabandara, G. D. McDonald, K. S.
[32] M. Syed and N. G. Berloff, IEEE Journal of Selected Hardman, C. D. Quinlivan, P. Manju, C. C. Kuhn, et al.,
Topics in Quantum Electronics 29, 1 (2023). Scientific Reports 6, 25890 (2016).
[33] J. S. Cummins, H. Salman, and N. G. Berloff, arXiv [44] D. Subires, F. J. Gómez-Ruiz, A. Ruiz-Garcı́a, D. Alonso,
preprint arXiv:2311.17359 (2023). and A. Del Campo, Physical Review Research 4, 023104
[34] J. S. Cummins and N. G. Berloff, arXiv preprint (2022).
arXiv:2403.16608 (2024). [45] R. Hamerly, T. Inagaki, P. L. McMahon, D. Venturelli,
[35] K. P. Kalinin and N. G. Berloff, Scientific Reports 8, A. Marandi, T. Onodera, E. Ng, C. Langrock, K. Inaba,
17791 (2018). T. Honjo, et al., Science Advances 5, eaau0823 (2019).
[36] E. Boros and P. L. Hammer, Discrete Applied Mathe- [46] M. P. Harrigan, K. J. Sung, M. Neeley, K. J. Satzinger,
matics 123, 155 (2002). F. Arute, K. Arya, J. Atalaya, J. C. Bardin, R. Barends,
[37] N. Stroev and N. G. Berloff, Physical Review Letters S. Boixo, et al., Nature Physics 17, 332 (2021).
126, 050504 (2021). [47] F. Böhm, G. Verschaffelt, and G. Van der Sande, Nature
[38] H. Ohadi, R. Gregory, T. Freegarde, Y. Rubo, A. Ka- Communications 10, 1 (2019).