Resource Bounds for Quantum Circuit Mapping via Quantum Circuit Complexity

Matthew Steinberg1,2 *, Medina Bandić1,2 *, Sacha Szkudlarek1 , Carmen G. Almudever3 , Aritra Sarkar1,2 , and Sebastian Feld1,2
1
Quantum Computing Division, QuTech, Delft University of Technology, Delft, the Netherlands
2
Department of Quantum & Computer Engineering, Delft University of Technology, Delft, the Netherlands and
3
Computer Engineering Department, Technical University of Valencia, Valencia, Spain
(Dated: February 2, 2024)
Efficiently mapping quantum circuits onto hardware is an integral part of the quantum compilation process,
wherein a quantum circuit is modified in accordance with the typically stringent architectural demands of a
quantum processor. Many techniques exist for solving the quantum circuit mapping problem, in addition to
several theoretical perspectives that relate quantum circuit mapping to problems in classical computer science.
This work considers a novel perspective on quantum circuit mapping, in which the routing process of a simplified
circuit is viewed as a composition of quantum operations acting on density matrices representing the quantum
circuit and processor. Drawing on insight from recent advances in quantum resource theory, circuit complexity,
arXiv:2402.00478v1 [quant-ph] 1 Feb 2024

and information geometry, we show that a minimal SWAP gate count for executing a quantum circuit on a
device emerges via the minimization of the distance between quantum states using the quantum Jensen-Shannon
divergence. Additionally, we develop a novel initial placement algorithm based on a graph similarity search that
selects the partition nearest to a graph isomorphism between interaction and coupling graphs. From these two
ingredients, we then construct a polynomial-time algorithm for calculating the SWAP gate lower bound, which
is directly compared alongside the IBM Qiskit compiler for over 600 realistic benchmark experiments, as well
as against a brute-force method for smaller benchmarks. In our simulations, we unambiguously find that neither
the brute-force method nor the Qiskit compiler surpass our bound, implying utility as a precise estimation of
minimal overhead when realizing quantum algorithms on constrained quantum hardware. This work constitutes
the first use of quantum circuit uncomplexity to practically-relevant quantum computing. We anticipate that this
method may have diverse applicability outside of the scope of quantum information science, and we discuss
several of these possibilities.

I. INTRODUCTION circuit-mapping problem is NP-complete [14, 15], and has

been likened to the traveling salesman problem (TSP) on a
The promise of quantum technology extends to many areas torus [16]. The QCMP is also related to token swapping [17].
of modern theoretical physics, computer science and cryptog- Many competing approaches have been proposed for solving
raphy, among others [1]. In spite of much success over the past the QCMP, with all of the state-of-the-art strategies trading ac-
30 years, current-generation quantum technology is character- curacy for speed, among other considerations [18–34]. How-
ized by noisy, intermediate-scale devices that are severely lim- ever, to our knowledge, no work has attempted to formulate
ited not only by the depth and size of the quantum circuits that the QCMP from a standpoint grounded in theoretical physics,
can be executed, but also by the qubit connectivity of such de- specifically in quantum information theory. The motivation
vices [2]. Such processors, have allowed for the first genera- for such an endeavor is twofold. Firstly, since the QCMP is a
tion of quantum-technology demonstrations, ranging from ex- physical process, such a description can provide new insights
perimental realizations of hybrid quantum-classical optimiza- and perspectives on how best to solve it. Secondly, by pro-
tion techniques [3, 4] to resource-intensive algorithms such as viding a fundamental description of the QCMP, we lay the
fault-tolerant quantum error-correction codes (QECCs) [5–7]. groundwork for uniting the various contemporary approaches
With such promise as is forecasted for quantum technol- towards a solution, and show how they relate to each other in
ogy, full-stack design approaches have emerged, in order to a self-consistent framework. In short, a physics-motivated de-
delegate resources efficiently and to ensure high success rates scription of the QCMP offers consensus for current and future
for a given quantum circuit, realized on a quantum proces- solution strategies, and how best to compare them.
sor [8–10]. As such, one of the cardinal issues to emerge Underpinning the advances in quantum technology, quan-
for practical quantum computing is that of quantum compila- tum information theory seeks to quantify the achievable lim-
tion, which can be broadly defined as the various engineering- its of information processing on a fundamental mathemat-
level steps required to translate and prepare a quantum circuit ical basis using quantum physics [1, 35, 36]. While such
for execution on a quantum processor [11]. Central to the progress is already notable, many fields outside of the im-
quantum compilation is the quantum circuit-mapping prob- mediate scope of quantum information theory have benefited
lem (QCMP), which concerns the assignment and rearrange- from incorporating quantum-information-theoretic interpreta-
ment of qubits from an algorithm to a processor as a quan- tions to outstanding research problems, including theoretical
tum circuit is executed, in order to guarantee high fidelity physics [37, 38], network science (which studies the behav-
of the resulting state [12, 13]. It is known that the quantum ior of complex networks from the standpoint of statistical me-
chanics and graph theory) [39–41], among many others. Bear-
ing in mind such potential, we apply notions of quantum infor-
mation theory, in particular, quantum resource theory [42, 43],
* These authors contributed equally to this work. spectral graph entropy [39, 44–46], and the quantum opera-
 2

tions formalism [35, 47–49] to the QCMP, in order to describe compiler, finding that in all cases, the SWAP gate count as
the problem of preparing certain quantum states on a quan- calculated by the SWAP uncomplexity algorithm is never sur-
tum processor whose qubit connectivity is restricted. As a passed, serving as a useful proof of concept. This minimal
quantum circuit itself describes a sequence of unitary trans- bound is of great importance, as such resource constraints can
formations under which a quantum state transforms, address- help with the prediction of compilation performance, as well
ing such quantum operations under the guise of a processor’s as for making design choices relevant in application-specific
connectivity is not only reasonable using quantum informa- mapping strategies and quantum devices [8].
tion theory, but also embodies a natural extension of quantum This article is organized as follows. We review the QCMP
information theory to the setting of practical quantum com- in Section II. Following this section, we immediately discuss
puting. in Section III how to mathematically map objects such as the
Several recent proposals have sought to establish links be- graph Laplacian from graph theory into the density matrix
tween graph theory and the QCMP [28, 29, 50–52]. Since formalism, highlighting in the process the importance of en-
then, it has become commonplace in the literature to con- tropic divergence measures (Section III A). Next, we show in
sider an interaction graph (IG), in which edges represent the Section IV how to translate the QCMP in density-matrix form
necessary two-qubit interactions for implementing a quantum to a statement about quantum circuit complexity, and we de-
circuit, and a coupling graph (CG), whose edges determine scribe how to optimize for a quantity known as uncomplexity
allowed two-qubit interactions between neighboring subsys- [42, 43]. In Section V we describe an algorithmic implemen-
tems on the processor [53]. In many of these proposals, the tation for calculating the SWAP uncomplexity; here we also
SWAP gate count required to realize a quantum algorithm on treat several subtleties of the problem (Sections V A and V B).
a given quantum processor is considered to be a typical metric We present the numerical results from over 600 benchmarks
for the objective function of a mapping strategy [22]. tested on realistic quantum circuits and hardware layouts in
In this vein, we strengthen this connection by initiating a Section VI. Next, a physical interpretation of the SWAP un-
study of the QCMP from the standpoint of the quantum op- complexity is given, using a Penrose diagram as a guide, in
erations formalism, using notions from graph theory and net- Section VII. Finally, we offer final conclusions and ideas for
work science as a foundation. More concretely, we propose a future work in Section VIII.
special case of the QCMP in which all two-qubit interactions
of a given IG can be compressed into a single time slice of
the quantum circuit. Starting from this point, we reformulate II. QUANTUM CIRCUIT MAPPING
the QCMP as a superoperator mapping in the space of density
matrices; as such, we employ doubly-stochastic quantum op- Generally, quantum algorithms and their associated circuit-
erations [49, 54], wherein both the initial density matrix rep- level descriptions are developed without considering the
resents the graph Laplacian of the associated IG, subject to the architecture-specific limitations of particular devices, i.e.,
constraints of the CG. Using recent results in resource theory, they are developed in an architecture-free manner. For ex-
quantum circuit complexity [42, 43], and methods from quan- ample, the elementary gate set (or primitives) for a particular
tum information geometry [55], we methodically show that device may differ significantly from what has been indicated
entropic divergence measures can be used in order to mini- at the level of a generic circuit description; as such, several
mize the distance between density matrices describing the IG actions must be performed in order to translate the quantum
and CG, and that a minimal SWAP gate count can be ascer- algorithm into a circuit that a quantum device can actually
tained for any IG / CG pair using this method. This minimal execute. Another example can be seen in the physical connec-
SWAP gate count is shown to exactly coincide with the quan- tivity properties of a quantum processor, which must be con-
tum circuit uncomplexity [42, 43]; as such, we name this lower sidered to ensure that the necessary qubit-qubit interactions of
bound the SWAP uncomplexity. the circuit can be performed on the device. Although certain
In addition, we develop a novel algorithm for the qubit exceptions may exist (in which several of the aforementioned
assignment (or initial placement) of qubits from the IG to steps may not necessarily be carried out), these procedures are
the CG, based on a subgraph isomorphism and graph sim- collectively known as quantum circuit mapping [11].
ilarity search [56–58], which has applications for multi- The task of quantum circuit mapping itself is usually di-
programming on a quantum device [59] and may be of in- vided into several steps which typify the process: A) elemen-
dependent interest. In this case, however, it serves as a neces- tary gate-set decomposition, which involves the translation of
sary step in our formulation and further enables our approach a circuit to a native gate set utilized by a quantum processor;
by constraining the coupling graph to match the size of an IG, B) scheduling, which concerns the formation of a logical time
which is one of the method’s crucial requirements. This al- ordering for algorithm execution, and includes considerations
gorithm also facilitates a calculation that we use to compute a for parallelism of operations and for the shortening of circuit
maximal SWAP gate count. depth; C) qubit assignment, which relates to the initial assign-
Together with the formalism introduced, a combined ap- ment of qubits from an algorithm to the physical qubits on a
proach is constructed that searches for the best qubit assign- quantum architecture; and D) qubit routing, which examines
ment in terms of the graph-edit distance (GED), and then the increase in gate overhead as extra operations are inserted
calculates the SWAP uncomplexity, given an IG/CG pair as into the algorithm as a function of physically moving qubits
inputs. We test the resultant algorithm against IBM’s Qiskit around the processor, such that the required two-qubit opera-
 3

(a) (b) (c) (d)

FIG. 1. An example of the QCMP as a sequence of steps needed to assign qubits from an algorithm to a quantum device. The two-qubit gates
in the two circuit diagrams are used to represent general two-qubit unitary operations (with the exception of SWAP gates); here, we do not
consider single-qubit gates, and the two-qubit interactions shown in (a) and (d) are taken to be general two-qubit unitary operations. (a) The
quantum circuit is transformed into an interaction graph, as shown in (b). Next, it is compared with the connectivity properties of the coupling
graph (c). As no graph isomorphism (i.e., no exact matching between the vertices of the IG and CG which upholds all of the edge relations of
both) exists between the IG and CG, one can compensate for the lack of connectivity by introducing SWAP operations to the circuit in order
to realize the circuit. (d) These operations degrade the fidelity of the final output state.

tions are realizable [11]. Typically, operations such as SWAP III. GRAPHS AS DENSITY MATRICES
gates are utilized in order to adapt the circuit to hardware;
these amount to classical permutation operations on a product In quantum physics, the most general manner of describing
state, but other approaches exist as well [32, 60, 61]. a quantum state involves the use of density matrices [1, 64]. A
density matrix ρ is a Hermitian, positive semidefinite matrix,
As a simple example, consider the quantum circuit-   A system |ψ⟩ is termed pure if
whose trace is equal to unity.
mapping procedure in Figure 1. The circuit on the left is de- and only if the bound Tr ρ2 ≤ 1 is saturated. The density
composed into IG form (a)-(b), wherein we do not consider matrix admits a spectral decomposition as
single-qubit gates for simplicity, and we assume that the two-
qubit gates shown in the circuit diagrams are taken as gen- X
eral two-qubit operations. Upon comparing the IG with the ρ= pj |ψj ⟩ ⟨ψj | , (1)
available qubit-qubit interactions afforded by a quantum de- j
vice, it is apparent that the connectivity available on the device
and the connectivity required by the algorithm differ (the two- for an orthonormal basis {|ψj ⟩}, where pj are non-negative
qubit interaction represented by the edge ed0 d1 is not possible, eigenvalues summing to 1.
as shown in (b)-(c)). As such, this discrepancy can be com- The density matrix formalism allows one to define the Von
pensated for by adding detrimental SWAP operations to the Neumann entropy (VNE), which takes the form
circuit’s initial assignment (d). This example provides an il-
lustration of quantum circuit mapping based on an exact graph  
matching between IG and CG. Implicit to this example was S(ρ) = − Tr ρ log ρ . (2)
the assumption that exactly such a mapping between the ver-
tices of the IG and CG exists which preserves all of the edge Here we take the convention 0 log2 0 := 0.
relationships of the IG; this is known as a graph isomorphism, In this work, as in [39, 40], we make use of the concept of
and will be discussed in more detail in Appendix A. a density matrix to describe a complex network (i.e. a graph
with many edges and vertices, and assumed topological struc-
ture [65]), by defining a matrix from a network which fulfills
In this paper, we formulate and solve a special case of the the mathematical properties of a density matrix. One such
QCMP. In contrast to Figure 1, our formalism exhibits three candidate was previously shown in [39] to be promising for
main simplifications. Firstly, as is typical in the QCMP, we adhering to the property of subadditivity for the VNE; this
do not consider single-gate interactions in the formalism that Gibbs state formulation is defined as
we present starting in Section III. Secondly, we assume the
existence of a noiseless quantum device, as such an ideal sce-
nario precisely allows for the emergence of a lower bound. e−βL
Finally, in real quantum hardware, typically we only consider ρL = , (3)
Z
that one multiqubit gate can be performed on a given hardware
qubit during a given moment; this necessitates the division of where ρL , e(·) , β, and Z represent: the density matrix of
the quantum circuit into time slices. Instead, we consider the graph Laplacian L; the matrix exponential; the inverse tem-
scenario in which all two-qubit gate interactions of the circuit perature (or diffusion time [44]);
 and  the partition function,
can be performed within a single unit time slice, i.e. the causal which is defined as Z = Tr e−βL , respectively. We de-
structure of our circuits are taken to be indefinite [62, 63]. We fine the graph Laplacian as L := D − A, following [39, 66].
devote more detail to these concepts in Section VII. Throughout the text, we will refer to the graph Laplacians
 4

for the IG and CG using the notation LIG , LCG , respectively, orthogonal states; in this case, the divergence is unbounded
and ρIG , σCG to refer to their derived IG and CG density- and gives DQRE (ρ||σ) 7→ ∞ [1]. In the practical setting of
matrix forms, respectively. Additionally, we refer to edges in the QCMP, such a measure is therefore not useful and does
a graph-theoretic context as a line connecting two vertices; in not convey the necessary distance information.
the density-matrix formalism, we will make reference to this Secondly, the qJSD is symmetric. This property is concomi-
instead with the term subsystem interactions. tant to the previous property related to metric spaces, but we
Using these objects to describe complex networks is ad- address it here separately. Symmetricity means that the qJSD
vantageous for several reasons. Firstly, although it is known obeys the relation
that the graph Laplacian is uniquely determined up to vertex-
numbering assignments [67], the eigenvalue spectrum of the
graph Laplacian does not allow for unique identification of DqJSD (ρ||σ) = DqJSD (σ||ρ). (6)
a graph. For example, two graphs can be cospectral, i.e.
possessing the same eigenvalue spectrum, but with different For the QCMP, we observe that this relation is desirable, as
connectivity [67]. As such, the approach we detail in this we wish for the notion of distance between two density matri-
work is motivated by the fact that entropic divergence mea- ces to stay the same, regardless of whether one is derived from
sures allow for a unique differentiation between two quantum LIG or LCG . As we will see in Section IV, it is in fact this
states ρ and σ. Secondly, the VNE is permutation-invariant, distance quantity that we relate to the quantum circuit uncom-
i.e., the VNE is invariant under a reordering of subsystems. plexity [42, 43]. Additionally, the concept of symmetricity is
For example, suppose we have a state vector of five subsys- paramount, as it permits us to directly relate the qJSD to the
tems a, b, c, d, and e. If twoPsuch subsystem orderings give Fisher information metric; it was in fact shown through ap-
rise toPdensity matrices η = abcde∈Z2 |abcde⟩ ⟨abcde| and plication of the Fubini-Study metric that the qJSD is directly
ξ = abcde∈Z2 |baced⟩ ⟨baced|, it can be shown that the
related to the thermodynamic distance between two equilib-
equality S(η) = S(ξ) holds [49]. Lastly, as discussed in rium quantum states, and lower bounds their divergence on a
[39, 40], previous attempts to calculate the classical entropy Riemannian manifold [71].
of a complex network fail, as these measures are dependent Moreover, the qJSD is non-increasing under the action of a
on a probability distribution resultant from a specific network CP map [68], which can be formally stated as
descriptor. In contrast, the quantum approach we utilize does
not depend on a specific network descriptor, but rather the en-
tire network, rescaled and normalized as a Gibbs state. DqJSD (ρ||σ) ≤ DqJSD (Λ(ρ)||Λ(σ)), (7)

where Λ(·) represents the superoperator of a quantum op-

A. Entropic Divergence Measures eration. The most general form of a quantum operation can
be written in several representations; in this work we will
The task of actually distinguishing two quantum states ρ concentrate on the Kraus representation (also known as the
and σ can be accomplished through the use of the entropic operator-sum representation), stated as
divergence measures [1, 36, 45, 46, 49, 68, 69]. In particular,
we employ the quantum Jensen-Shannon divergence (qJSD),
Ei · Ei† ,
X
Λ(·) := (8)
which is defined as
i

where Ei is the ith term in the sum of operators, and Λ(·) is

 
ρ+σ 1

DqJSD (ρ||σ) = S − S(ρ) + S(σ) . (4) taken to be a general quantum operation superoperator, con-
2 2
strained to the completely-positive (CP) condition [1, 35, 36].
One may immediately ask why we choose to utilize the We also introduce here the class of doubly-stochastic (DS)
qJSD, and not other quantum entropic measures, such as the quantum channels, with the term quantum channel distin-
mutual information or the quantum relative entropy (QRE) guishing from quantum operation in that, in addition to the
[1]. The reason for this choice is severalfold, and is intimately CP constraint, we additionally impose trace preservation (TP)
related to the QCMP. Firstly, the qJSD defines a bounded met- [1]. In this work, we will refer to CPTP maps using Φ(·).
ric space (see Appendix B). More specifically, we note that Moreover, DS quantum channels are unital, meaning that the
fixed point of the channel upholds the equality Φ(In ) = In
q [47, 49]. In defining the class of DS quantum channels, we
0≤ DqJSD (ρ||σ) ≤ 1, (5) use the fact that any Kraus operator can be factorized, as all
systems of Kraus operators implementing a quantum opera-
with a value of 0 signifying that ρ = σ, and a value of 1 tion are related by a unitary transformation. A particular de-
used for the case of ρ ⊥ σ [68, 70]. As we are comparing the composition can be defined as
density matrices related to the IG and CG of a quantum circuit
and processor, it is imperative to understand the closeness of Xp
one to the other, using some bounded distance measure. As Ei = θ j Pj . (9)
a contrasting incentive, consider measuring the QRE of two j
 5

Here Pj ∈ Pn refers to permutation matrices from the set to the ith iteration of edge removal performed on the original
of n × n permutation matrices, and θj refers to a probability ρIG .
distribution [49]. The existence of this class of convex de- Once ρiIG is derived from its modified graph Laplacian, the
composition comes from the Birkhoff-Von-Neumann Theorem erasure process can commence. In order to erase a subsys-
[36, 49, 54] for which it is known that such a decomposition tem interaction from ρiIG , we must first permute a pair of sub-
can be found in polynomial time [72]. systems from σCG such that the corresponding interaction be-
Lastly, we introduce the notion of projective measurements comes possible on the device.
for density matrices constructed from graph Laplacians. Fol- Let us imagine first, as a simple scenario, that ρiIG contains
lowing the treatment ofP [41], we define a set of orthogonal only one subsystem interaction remaining. In order to permute
projectors Πk such that k Πk = In . The post-measurement the subsystems of σCG , we employ the DS quantum channels
Πi ρΠi
state of a general density matrix ρ is Tr[Π i ρ]
, where Tr[Πi ρ] introduced in the previous section. The resulting calculation
th
represents the probability of the i measurement outcome. proceeds as
Note that projective measurements are known as a specific
example of a CP map [1, 35, 36]. In Section IV, we shall
use projective measurements to erase subsystem interactions DqJS (ρiIG ||Φ(σCG )) ≥ 0, (11)
from the density-matrix form of the IG, ρIG , as well as for se- P
lecting appropriate subsystem permutations of the CG density where Φ(σCG ) = j θj Pj σCG Pj . There is one subtlety to
matrix σCG . this point. Not every possible permutation matrix in Pn can be
defined for the DS quantum channel, as we are limited by the
connectivity of the quantum device. In view of this, we restrict
IV. QUANTUM CIRCUIT MAPPING AS A RESOURCE Pj ∈ PCG CG
n , where Pn represents the permutation matrices
THEORY allowed by the available two-qubit interactions in ECG .
As the superoperator Φ(σCG ) yields a superposition of pos-
In the QCMP, if graph Laplacians LIG and LCG exactly sible permutation choices, one may ask then how to decide
match, i.e. if LIG = LCG , then all of the necessary two- on which permutation to apply to σCG . In order to resolve
qubit interactions from the IG can be performed on the CG this complication, we perform a projective measurement cor-
without further circuit modification, effectively requiring zero responding to the permutation matrix appearing in the Kraus
SWAP gate insertions. Similarly, as discussed in Section III A, operator which obtains the largest value of θj . Conversely, if
if DqJS (ρIG ||σCG ) = 0, then we can analogously conclude that we obtain several values of θj which are equal (this can hap-
ρIG = σCG , and all subsystem interactions from the IG are re- pen in larger complex networks, where several shortest paths
alizable on the CG without the addition of SWAP gates. We of permuting subsystems exist), we make no distinction as to
will consider this case to be trivial from the standpoint of the which component is chosen (in Section V, we treat this is-
QCMP. sue in the practical sense by employing an optimization al-
Under these conditions, the qJSD provides an efficient gorithm based on gradient descent from the Python package
method to check whether a quantum circuit is implementable SciPy [73]). The physical motivation for this procedure can
on a given quantum device. However, one might wonder how be justified from tools in quantum information geometry [55].
to address the case of DqJS (ρIG ||σCG ) ̸= 0, which is the aim It is well known in quantum information geometry that the
of this work; in order to confront this, we return to physical distance between two equilibrium quantum states can be pa-
considerations of the problem. rameterized by the Fubini-Study metric [42, 55, 74]. This met-
We first address how to ascertain exactly which two-qubit ric in turn can be shown to be a special case of the Fisher
interactions are not possible on the quantum device. This aim information metric applied to projective Hilbert spaces. The
can be accomplished by means of projective measurements; Fisher information metric is also directly related to the qJSD
the idea here is to simply project out all two-qubit interactions [71]. In light of this knowledge, minimizing the qJSD is then
from ρIG which are already possible on σCG . From [41] it immediately associated with a minimization of the thermody-
was shown that the graph operation known as edge removal namic path length between two equilibrium quantum states,
was possible by considering projective measurements, which and thus lower bounds their distance on a Riemannian state
are known to be CP operations. As such, our overarching goal space manifold. Additionally, it is known that VNE monoton-
will be to erase all of the subsystem interactions from ρIG as ically increases under the action of a CP map [75]. Therefore,
quickly as possible, as this corresponds to all of the circuit we argue that by performing a selection of the maximum-
two-qubit gates being accounted for on the device. Firstly, we probability permutation matrix on σCG , we in effect choose
define a new density matrix ρiIG as the quantum operation which minimally increases the VNE,
ergo minimizing the qJSD between σCG and ρiIG .
As such, we modify Equation 11 in order to accomodate
ρiIG := ΠĒ ρIG ΠĒ , (10) this choice as a projective measurement:

where the projectors ΠĒ leave only those subsystem inter-  

actions which are in the IG’s edge set EIG but not in the CG’s i
min DqJS (ρIG ||Πθmax (Φ(σCG ))) ≥ 0. (12)
edge set ECG , i.e. Ē = EIG \ECG . The superscript i refers θ̄
 6

Here, Πθmax signifies a projective superoperator map onto input for the algorithm; USWAP is set to zero. Next, the algo-
the maximum-probability θmax after minimizing. If we require rithm described in Section V B is used to determine the op-
several subsystem permutations before reaching the minimal timal initial assignment of qubits from ρIG to the qubits of
value of qJSD, then we simply extend the map by defin- σCG . The object ρi=0IG is generated using the projective mea-
ing a composition of such quantum operations as ξ(·) = surement approach detailed in Section IV. Once an initial as-
Πθmax (Φ(·)), and the action of these operations can be written signment is completed, an orientation of the two density ma-
concisely as ξ ◦li (·) = Πlθimax (Φli (· · · Π1θmax (Φ1 (·)))), where trices to each other is set, and they can be compared using
the superscript li refers to the lith superoperator composition the qJSD. If the value of the qJSD is initially zero, then both
for the ith remaining subsystem interaction in ρiIG . This can of the density matrices are the same, and no SWAP gates are
be amended to Equation 12 as necessary. However, if the qJSD is non-zero, we then perform
a minimization of the divergence, allowing for permutations
   of subsystems in σCG via the DS quantum channel; the mini-
mization takes place over the vector of squared probabilities θ̄.
min min DqJS (ρiIG ||ξ ◦li (σCG )) ≥ 0. (13)
li θ̄ As a result of the optimization, the permutation matrix whose
associated θj value is highest is selected and performed on the
Equation 13 still only takes into account one possible sub- device, in agreement with previous observations on the mini-
system interaction remaining on ρiIG , however. We may gen- mization of thermodynamic length via the Fisher information
eralize this logic further for |EIG | subsystem interactions re- metric [71, 77]. The superscript l is used to count the num-
maining in ρiIG , such that many possible remaining subsystem ber of iterations of the DS quantum channel as the algorithm
interactions can be treated: proceeds. Continuing, the divergence of the newly-modified
CG density matrix, ξ li +1 (σCG ), should be lower than before.
 
  If instead it is found that the divergence is not lower (or is
Πi i ◦li
min min DqJS (ρIG ||ξ (σCG )) ≥ 0. (14) the same), then the previous iteration, ξ l (σCG ), is taken to the
li θ̄ i∈ĒIG next step of the algorithm. The optimization scaling for this
  calculation is expected to be O(n2 + c), as is dictated by the
Here, Πi · represents the projective superoperator map convexity of the problem [72]. Finally, we check explicitly to
which erases the ith subsystem interaction from ρiIG . In the see if ρiIG = In ; if it is, then we first sum the li in the vari-
end, we successfully project out all of the subsystem inter- able USWAP and immediately stop the algorithm, as our goal
actions from ρiIG , as they become realizable on the quantum of erasing subsystem interactions in ρIG is achieved. If not,
device denoted as σCG . After the final step, ρiIG = In . we proceed to remove additional subsystem interactions from
However, we can further sum all of the gate operations ρiIG , generating a further iterated object ρi+1
IG .
which were required in order to transport σCG as close as pos- At this point, the erasure of a remaining subsystem inter-
sible to all of the ρiIG . Indeed, summing over the collected action implies that the qJSD will again increase, as we know
li permutations per subsystem interaction i allows us to find that the VNE under a CP map always increases [75]. We must
the minimal SWAP gate count for performing a circuit on a then perform the commensurate DS quantum channel opera-
device. Continuing, we are left with the equation tion(s) again and select the appropriate θj -valued permutation
such that the divergence decreases to its minimal value once
more. The algorithm terminates upon the successful erasure of
Ē
X all subsystem interactions in ρiIG , leaving a maximally mixed
li = C(In ) − C(σCG ) = USWAP , (15) state. Once this is completed, we sum all of the collected li for
i
each edge i; this sum is then taken as the SWAP uncomplexity
where USWAP represents the SWAP uncomplexity between USWAP and is returned.
ρiIGmax = In and σCG [42, 43], and C(·) signifies the quantum
circuit complexity. It is known that quantum circuit uncom-
plexity defines a resource theory and can be optimized for A. The β Parameter
[76]. As it is known that most states of maximal circuit com-
plexity are maximally mixed [42], the formalism presented One of the subtleties of our algorithmic implementation lies
here portrays a precise lower bound distance measure in terms in the estimation of an appropriate value for the inverse tem-
of SWAP gates required for executing a quantum circuit on a perature β, and can be tricky. As shown in Figure 2, we have
quantum device. graphed the value of β versus the VNE in the resulting den-
sity matrix. The figure depicts a dependency of the VNE on β.
As shown, a phase transition is present at and about the point
V. ALGORITHMIC IMPLEMENTATION of β ∼ 100 , while the high-temperature regime generally be-
gins at β ∼ 10−1 and extends leftwards. In this work, we
Pseudocode for calculating the SWAP uncomplexity is search over a range of β values with varying increments; the
shown in Algorithm 1. The algorithm proceeds in a very sim- preferred value yields the minimal SWAP gate count for an IG
ilar form to the mathematical breakdown. Firstly, the density- and CG pair. We were able to ascertain this minimal SWAP
matrix objects ρIG and σCG are initialized and provided as gate count using a search over the range of 10−5 ≤ β ≤ 105 .
 7

Algorithm 1: Pseudocode for SWAP Uncomplexity

USWAP .
Input: ρIG , σCG
Initial Qubit Assignment ← Qubit Assignment(ρIG , σCG )
USWAP ← 0
Output: USWAP
Assert: DqJS (ρIG ||σCG ) == 0

if DqJS (ρIG ||σCG ) ! = 0 then

ρi=0
CG ← Remove Trivial Edges(ρCG )
for i ∈ [0, |ĒIG |] do
for li ∈ [0, |ECG |] do 
qjsd1 ← minθ̄ DqJS (ρiIG || j θj Pj σCG Pj )
P

ξ li +1 (σCG ) ← Πmax θj Pj σCG Pj Πmax

 
 
qjsd2 ← minθ̄ DqJS (ρiIG ||ξ li +1 (σCG ))
if qjsd2 < qjsd1 then
continue
else
USWAP ← USWAP + li
break
end if
end for FIG. 2. The fluctuation of SG (ρi ) as a function of β for 4-node
if ρiIG == In then graphs. We have utilized different marker types in order to distin-
USWAP ← USWAP + li guish curves with very similar VNE.
break
else
ρi+1 i
IG ← Remove Trivial Edges(ρIG ) spite its potential. Instead, most existing approaches focus
end if
on alternative factors such as sequential gate flow in the cir-
end for
cuit or the number of interactions between qubits like in [20,
end if
79, 80]. Nevertheless, some work has explored the subgraph-
return USWAP
isomorphism concept for the QCMP [28, 52, 81, 82].
Building upon the foundation of the well-known VF2 algo-
rithm [28, 83], our approach to qubit assignment searches for
Keeping this matter in mind, the definition of SWAP uncom- an exact location on the quantum device where our circuit can
plexity states that there exists a minimal bound for the num- run without requiring additional gates. If a solution is feasi-
ber of SWAP gates at some value of β, but does not say di- ble, we are left with an optimal assignment. In cases where a
rectly how one may find the most appropriate β, consistent solution is not possible, we conduct a graph similarity search.
with USWAP . As we show in Section VI, it so happens that This process involves the GED calculation and comparison
almost all of our results (∼ 97%) achieve the minimal SWAP- of the IG to all distinct subgraphs of the same size within
gate count in the high-temperature regime, largely following the CG, which opens up alternative assignment possibilities.
the empirically-derived results of [78]. We will discuss possi- In this fashion, we condense the search space for alternative
ble improvements to this technique in Sections VII and VIII. solutions, while also highlighting the potential utility of our
approach for multi-programming applications (i.e. executing
multiple circuit in parallel on a quantum device) [59].
B. Qubit Assignment Let |VIG | be the number of qubits in the IG, and |VCG | be the
number of physical qubits on the CG , as per the definitions
As mentioned previously in Section II, the initial stage of from previous sections. Our qubit assignment process consists
the QCMP is known as qubit assignment (also known as initial of the following steps:
placement, qubit allocation, or initial mapping) [51, 52]. This
procedure plays a pivotal role in quantum circuit execution 1. Preprocessing:
[12]. In our proposal for calculating the SWAP uncomplexity (a) Select a quantum algorithm described as a quan-
in the QCMP, we also must assign qubits from the IG to the tum circuit and extract its interaction graph
CG initially in an optimal way, as this influences how many GIG (EIG , VIG ), where |EIG | represents the number
SWAP gates will be utilized. of edges in the IG.
In [51], the concept of qubit assignment was introduced as (b) Choose a quantum device to execute the cir-
a search for a subgraph isomorphism for an IG/CG pair. To cuit on and extract its CG, represented as graph
our knowledge, this technique has not yet seen widespread GCG (ECG , VCG ), where |ECG | stands for the num-
implementation in practical qubit-assignment techniques, de- ber of edges in the CG.
 8

(c) In order to increase the efficiency of steps later on, (a) 0 1 0 1 2

and reduce the search space, we find all distinct
subgraphs of size |VIG | within graph GCG .
2. Subgraph isomorphism using VF2 and subgraph simi-
3
larity search:
(a) Use the VF2 algorithm to check if a subgraph iso- 3 2
morphism exists between graphs GIG and GCG .
4 5 6
i. If a subgraph isomorphism is found, we im-
mediately determine the location within the (b)
CG for qubit assignment.
ii. When a subgraph isomorphism does not ex-
ist, we utilize the graph-edit distance (GED)
(Appendix A) to identify structurally most
similar subgraph of the CG when compared
to the IG. During this process, we compare
IGs only to distinct subgraphs of a CG de-
rived from Item 1c.
(c)
(0,0) (1,1)
(b) Assign the IG to the CG in accordance with the
result from the previous step.
3. Calculating the maximal SWAP gate count as it is qubit
(2,3)
assignment-dependent. We describe the computation of
this bound in more detail in Appendix C.

The steps in the algorithm are exemplified in Figure 3. (3,5)

Here, we take as a simple example the case of a 4-qubit IG
assigned to a 7-qubit architecture, as shown in (a). In (b), we
FIG. 3. Steps taken for the qubit-assignment algorithm described in
display the two distinct 4-qubit subgraphs that are identified in Section V B: (a) The task at hand is to find the best-fit initial place-
Item 1c. Finally, in (c) we show the final assignment of qubits ment for the qubits in the 4-qubit IG (shown with blue numbering)
as ordered pairs; in this case, an graph isomorphism was not for the 7-qubit CG (shown with red numbering); in our case, the the
found. Therefore, we select the subgraph with the lowest GED CG corresponds to the connectivity of the IBM Casablanca quan-
calculated, as per Item 2(a)ii. The final initial placement and tum device. (b) Shows the distinct subgraphs found from Item 1c
related information obtained during this process serve as in- in Section V B. After verifying that no direct subgraph isomorphism
puts for the SWAP uncomplexity algorithm, as described in between IG and one of these graphs exists, a similarity search is em-
Sections IV and V. ployed. (c) The subgraph of the CG with the lowest GED relative to
In addition to implementing the technique described above, an IG is retrieved. The resulting initial placement is shown in blue,
a specific approach for complete graphs (i.e., all-to-all IGs) with the final GED calculated to be 1. The actual qubit assignment is
shown in the form of several colored ordered pairs (blue numbering
was utilized. For such cases, we automatically locate the
represents IG qubits, while red numbering represents CG qubits).
most-connected subgraph of that size within the CG. This
method can also be applied to circuits larger than 20 qubits.
However, for the purpose of this paper, we focus on smaller alizing such a comparison provides a practical first test. Sec-
circuits as a demonstration of the concept. ondly, to the best of our knowledge, there is scant literature on
bounding the required SWAP gates, and such work addresses
only up to quantum circuits of 6 qubits via a brute-force op-
VI. BENCHMARK EVALUATION & RESULTS
timization algorithm [85, 86]. Our simulation results demon-
strate scalability that greatly exceeds the current state of the
Here we describe numerical results taken from comparing art [85], as we achieved results for circuits of up to 16 qubits.
the SWAP uncomplexity against IBM’s Qiskit compiler [84]*, 57 benchmark circuits were selected from the qbench suite
as well as against a brute-force approach [85]. These exper- [53]. These benchmarks cover a range of 3 to 20 qubits and
iments were carried out for several reasons. Firstly, as the represent a wide spectrum of possible IG connectivities (47
SWAP uncomplexity algorithm does in fact solve for the min- different connectivities) encountered in quantum algorithms.
imal SWAP gate count, then we should be able to confirm More details about the selected benchmarks can be found in
this fact by comparing with any compiler. By such logic, a Table II. As for the CGs, we chose connectivity graphs from a
compiler should be able to approach but not find fewer SWAP set of 16 in-use quantum devices, ranging from 5 to 72 qubits.
gates for an arbitrary IG / CG pair. In order to perform this
empirical check, we choose to run our algorithm (Section V)
against the Qiskit compiler. As the Qiskit compiler is con-
* Used version of Qiskit is 0.24.1.
sidered to be the state-of-the-art approach at the moment, re-
 9

1.0

0.8
Normalized number of SWAPs

0.6
Maximal Bound
Compilation Results
SWAP Uncomplexity
0.4

0.2

0.0
0 100 200 300 400 500 600 700
Benchmark circuits
FIG. 4. Simulation results for various IG / CG benchmark pairs. The X-axis enumerates each benchmark circuit tested, and the Y-axis describes
the normalized number of SWAPs, due to very high maximal bounds (the SWAP-bound values of each benchmark are divided by their sum).
The results are color-coded as follows: the SWAP uncomplexity of Section IV (green); the Qiskit compiler with default options Sabre router
and circuit optimization level 1 [84] (orange); and the maximal maximal bound calculated as in Appendix C (blue). In every IG/CG pair,
the bound calculated captures the SWAP uncomplexity that is either approachable or unattainable by the Qiskit compiler, thus empirically
demonstrating our formulation.

The specific details of these devices are provided in Table I. positive correlation (34% and 73%, respectively) with the ac-
As some of the benchmarks are too large to be run on some tual results obtained from the compiler. The correlations here
of the smaller processors from our list, in total we devised exemplify the non-triviality of the bounds; in other words, the
675 simulation experiments with the Qiskit compiler [84]. In SWAP uncomplexity and maximal bounds grow proportion-
these simulations, we utilized Qiskit’s transpiler with the de- ally with the actual compilation results. The results at best co-
fault circuit-optimization setting. incide with each other, meaning that the lower bound equals
The results of our simulations are shown in Figures 4–6. In the actual SWAP gate count from the presence of a graph iso-
Figure 4, we display the normalized number of SWAP gates morphism; in this case, the SWAP uncomplexity, Qiskit result,
found by: the SWAP uncomplexity from Section IV (shown and the maximal bound all obtain the same amount (which is
in green); the Qiskit compiler (orange); and the maximum zero if a graph isomorphism is present). These checks provide
SWAP gate count (which we name MSWAP in Appendix C) not only hard evidence for the usability of our methods, but
(blue). We observe that the SWAP uncomplexity can be additionally serve as pragmatic sanity test for our algorithmic
reached, but never surpassed by the Qiskit compiler for se- implementations.
lect benchmark trials. As expected, the Qiskit compiler sig- Although not shown in Figure 5, it is also worth observ-
nificantly outperforms the maximum SWAP gate count calcu- ing the considerable impact of the initial placement on re-
lated. trieving bounds, which particularly depends on the GED and
The non-triviality of the maximal and minimal SWAP gate the number of missing edges in the chosen CG partition com-
counts becomes evident in Figure 5, where we present a ma- pared to the IG. We therefore calculated the correlation coeffi-
trix with correlation coefficients whose values range from −1 cients of these two parameters as well when compared to our
to 1, with 0 indicating no correlation [87]. This matrix com- retrieved bounds, resulting in correlations of 79% and 61%
pares the results that we obtained throughout the simulation; for the SWAP uncomplexity and maximal bounds, respec-
in particular, we compare the effective correlation between the tively). Furthermore, using the same initial placement for the
SWAP uncomplexity; the Qiskit compiler SWAP calculation Qiskit compiler resulted in a 45% correlation with the initial
results; and the maximal bound as calculated in Appendix C. placement-related parameters.
Notably, both of our bounds (i.e., those obtained from our Additionally, our exact qubit assignment method was ini-
minimal bound with the SWAP uncomplexity algorithm, as tially tested with 729 benchmarks, showing a success rate of
well as the maximum SWAP gate count) exhibit a substantial 92.6%. The remaining 7.4% of the benchmarks could not
 10

1.00
lex AP
200
ity
0.75
mp SW
1 0.34 0.46

# of benchmarks
0.50 150
mp Unco

0.25
su on

100
lts
Re ilati

0.34 1 0.73 0.00

50
0.25
Co

0.50 0
un al

0.46 0.73 1
Bo xim

-1.0
1e-05
2e-05
3e-05
5e-05
8e-05
9e-05
0.0001
0.0002
0.0003
0.0004
0.0005
0.0006
0.0007
0.0008
0.0009
0.001
0.002
0.003
0.005
1.0
10.0
250.0
100000.0
d

0.75
Ma

SWAP Compilation Maximal 1.00 Values

Uncomplexity Results Bound
FIG. 6. The distribution of β values for minimal bounds found by
FIG. 5. The Pearson correlation matrix [87] between the critical the SWAP uncomplexity algorithm. Results in blue (∼ 97%) were
parameters measured for our benchmark investigation. The values found in the high-temperature regime described in [78]. We searched
range between −1 and 1 for negative and positive correlation, respec- values in the range from 10− 5 - 105 based on the following formula
tively. When one of the parameters changes, the other one changes A ∗ 10a where A ∈ 1, 2, 3...9 and a ∈ −5, −4, ..., 4, 5.
in the same direction. In this figure, we observe a high positive cor-
relation between all the selected parameters where the Pearson cor- 0.9
relation coefficient ranges between 0.34 and 1.0. Brute-force approach
0.8 SWAP uncomplexity
0.7
be finished due to insufficient computing resources (mem-
ory, long runtime, etc.). Recognizing the limited scalabil- 0.6
ity of the approach (up to 16 circuit qubits), we developed
0.5
a more relaxed method for complete graphs, mentioned in
Section V B. Indeed, the scalability of our exact algorithm al- 0.4
ready exceeded that of the exact state-of-the-art algorithms,
0.3
which struggled beyond 6 qubits [51]. Furthermore, our ini-
tial placement encountered no difficulties in exploring a vast 0.2
search space. It successfully executed circuits on all tested
0.1
devices, extending up to a size of 72 physical qubits in our

hwb4_52
3_17_13

4gt11_8
rd32-v0_67

decod24-v3_46
case.
Figure 6 displays the β values obtained during the course
of the simulation. As we must search over a range in order
to find the most-appropriate β, it is helpful and interesting to
catalog roughly how many benchmarks exhibited the minimal
bound obtained and at which β values. In particular, we find FIG. 7. Comparison of our minimal bounds with results of [85]. We
that the vast majority of the benchmark pairs (∼ 97%) led to notice that, just like before, our results (shown in orange) are always
minimal SWAP counts within the range of ∼ 10−5 − 10−3 , smaller or equal to the actual number of SWAPs. The benchmarks
consistent with the high-temperature regime studied in [78]. employed in this study are identical to those utilized in [85]. The
We will comment on this more in Section VII. results were normalized in the same way as in Fig. 4 .
Lastly, Figure 7 shows the results of a comparison lever-
aged between our SWAP uncomplexity algorithm (Section V)
and the brute-force optimization results from [85]. In this ap- tions, as it is known that such gates do not heavily affect calcu-
proach, the authors utilize an optimizer which essentially tries lated success rates [13]. Secondly, we have removed all two-
every permutation of SWAP placements which is possible, re- qubit interaction noise from the coupling graph; this should
specting the gate-dependency graph and weighted IG of the come as no surprise, as we are mainly interested in finding
original quantum circuit. In all cases, we see clearly that the a lower bound for the number of SWAP gates required, and
brute-force algorithm only achieves the minimal bound, but such a lower bound mandates the existence of a hypotheti-
never surpasses it. cally noiseless quantum processor. Thirdly, we consider the
limit in which gate dependencies for the IG are not consid-
ered; implicitly, we assume the existence of not only a noise-
VII. DISCUSSION less quantum processor, but additionally one that can perform
all two-qubit gate interactions required in one unit time slice.
It is known that the QCMP is NP-complete [14]. We have At the outset, one may consider a graph-theoretic interpre-
made three simplifications in order to derive the SWAP un- tation of the SWAP uncomplexity as considering only con-
complexity. Firstly, we do not consider single-qubit interac- nectivity dissimilarities between the interaction and coupling
 11

+
(a) 𝒥𝑅
𝜌𝐼𝐺 𝑅

(b)

Time

𝜌𝐼𝐺 𝑅 ∝ 𝕀𝑛
(c) 𝜌𝐼𝐺 𝑅

FIG. 8. Distinctions between the standard IG and the minimal ver-

sion of the same IG. (a) illustrates an archetypal quantum circuit with
arbitrary CU interactions, although making an exception with SWAP
FIG. 9. A Penrose diagram representing the quantum circuit com-
gates. (b) displays the standard weighted IG, together with its cor-
plexity for the evolution of a quantum state, which is related to a
responding gate-dependency graph; we follow the gate-dependency
quantum circuit ρIG (for the sake of simplicity, we choose the state
notation used in [88]. (c) shows the unweighted form of our IG; here
ρIG to be derived from the K4 graph, but one can choose other ex-
we do not take into account the gate ordering nor the number of gate
amples). It is possible to generate different quantum complexities
calls realized per qubit, opting instead towards an unweighted IG.
by adding different amounts and orderings of SWAP gates to the cir-
cuit as we approach the event horizon of a black hole (shown as a
shaded triangle). Here, the connectivity limitations of the CG cap-
graphs. Indeed, not only are gate dependencies explicitly not ture the role of the background spacetime geometry [89], as both de-
termine the ease by which certain operations can be performed. The
taken into account, but we assume that qubits can interact with
SWAP uncomplexity, as it ignores the effects of time ordering and the
each other potentially an infinite number of times within a unit amount of qubit-qubit interaction present, can be associated with the
time subdivision of a compiler’s scheduling scheme. These null lightlike geodesic + J [R], shown as a green dotted line. Below
differences are shown in Figure 8. In (a), we depict a generic this arrow, the bottom-right part of the diagram represents possible
quantum circuit, again omitting the single-qubit gates and tak- states that are inaccessible to us, given the restrictions of the coupling
ing every two-qubit gate shown to be a general CU gate for graph, as well as the operations available to us (SWAP gates, in our
simplicity (the obvious exception to this rule would be the case).
use of SWAP gates). In (b) the standard weighted IG and its
accompanying gate-dependency graph are shown, following
the notation of [88]. Finally, (c) makes manifest the differ- a given set represents a unique spectral signature which can
ences between the standard IG for the QCMP and for our un- be shared by multiple four-node subgraphs. The vertical axis
weighted version. Here, we take any number of two-qubit depicts the evolution of time, and is known as the null time
interactions to 1 on the IG, and we omit the gate-dependency geodesic, under which the set of trivial time-ordered gate op-
graph, preferring an indefinite causal structure. erations (i.e. idling, which in our simplified picture, is noise-
Such a graph-theoretic picture as ours can be related using less) evolve ρIG from point R to the same later state. Here,
Penrose diagrams [89, 90], as shown in Figure 9. A Penrose the trivial minimal SWAP gate count for the QCMP is repre-
diagram typically shows the causal structure of events unfold- sented, under which no SWAP gates are ever applied in order
ing in a spacetime geometry [89]. In Figure 9, the horizon- to adapt the quantum circuit to the device and its connectivity
tal axis refers to purely spatial evolutions, which in our case restrictions; in effect, the state is left to freely evolve for an
are shown by potential SWAP erasures. Each of the sets out- infinite amount of time, with no regards to gate operations.
lined in orange represent elements of the same total number Conversely, the red dotted line represents possible space-
of edges, but different spectral properties. Each graph within time evolutions arising from the application of distinct in-
 12

stances of time-ordered quantum operations Λ1 (·) (which in sumptions that we utilize in our formulation above is that pos-
this case are insertions of SWAP gates). In this way, every tra- sible interactions can occur only if a subsystem interaction be-
jectory on the diagram can be associated with a given SWAP tween qubits exists in the density matrix picture; intrinsically,
uncomplexity from a sequence of quantum operations. The moving outside the region embellished by the future light-
possible endpoints of the quantum circuit are shown along the like geodesic corresponds to new operations which must be
shaded triangle in dark gray, which represents ρIG = In , i.e. taken into account. One simple example lies in teleportation-
the state of maximal circuit complexity, a maximally mixed based quantum-circuit mapping, which can be used between
state. CG qubits which do not share a subsystem interaction, and
Furthermore, erasure transformations on the original quan- can allow for a smaller quantum circuit complexity [60]. One
tum circuit proceed according to restrictions dictated by the may suspect that architectures in the future may benefit from
background geometry (which in our present case is analogous such on-chip teleportation procedures, as work has shown that
to the CG connectivity). Evolution commences at a spacetime speedups exist over classical SWAP methods for exchanging
point R. The green dotted line traces out the lightlike null distant qubits [32, 60, 61], albeit with larger circuit and entan-
geodesic (i.e. future lightcone) + J [R]. This geodesic signi- glement overhead.
fies the SWAP uncomplexity, which is the path that the state Finally, we close this section by discussing the β parameter
takes under the minimal set of causally-indefinite operations in more detail. For the vast majority of our results (approxi-
such that we approach the maximally-mixed state in minimal mately ∼ 97%, as shown in Figure 6), the β values yielding
time. As we solve for the SWAP uncomplexity, without con- the SWAP uncomplexity for an IG/CG pair coincide with the
sideration of time ordering and as dictated by the thermody- high-temperature regime noted in [78]. Indeed, it was shown
namic path length calculable via the Fisher information met- there that the high-temperature regime allows for optimiza-
ric [71, 91], we can interpret our bound as the lightcone evo- tion of the quantum relative entropy for density matrices con-
lution of our initial state ρIG towards the event horizon of a trived from complex networks. The main interpretation pro-
black hole (shown as the shaded triangle). Previous work has vided in [78] is that the parameter β controls the diffusion of
already alluded to the concept of optimization over thermody- information about the graph neighborhood to other vertices.
namic distance [77, 92]; as such, our results point to a natural Conversely, if we take the β 7→ 0 limit, diffusion over the ver-
and reasonable extension of this trend for the quantum circuit tices of the graph is limited, and therefore, only information
mapping problem. about the degrees of links is conferred. This trait is known to
Taking stock, we would then expect that any realistic quan- exhibit a first-order linear dependency in the adjacency ma-
tum compiler which takes into account time ordering and fi- trix, and as such, the tendency for information to diffuse over
nite qubit-qubit interactions per unit of time slicing to be lim- the network becomes uniform. As we can see from Figure 6,
ited by the lightlike null geodesic. Indeed, surpassing the less than 3% of our benchmarks do not fall within the high-
lightlike null geodesic would introduce operations which are temperature β range described in [78] (shown in red). Those
not inside of the lightcone, giving access to the uncomputable outliers, however, can be explained with the triviality of the
region to the bottom-right. Such trajectories could be made correct β search process, which was particularly based on
possible using a larger set of routing resources, such as tele- sweeping over a range of values between 10−5 − 105 . One
portation [60, 61]. On the other hand, the SWAP uncomplex- future direction, therefore, is to devise more efficient schemes
ity bound implies an indefinite causal order, which would give for finding the optimal β value, guaranteeing a minimal bound
rise to superpositions of quantum gates that are executed in in less computation time. One could imagine that such a goal
tandem [62, 63]. could be completed by use of a gradient optimization method,
As we stated earlier, we can alternatively consider the but we leave such exploration to future work.
shaded triangle in Figure 9 as the event horizon into a black
hole. Indeed, Equation 15 gives rise to a new perspective for
the QCMP, due to its apparent relation with quantum uncom- VIII. CONCLUSIONS & OUTLOOK
plexity [42, 43]. Consider a benign black hole scenario in
which the black hole itself can only erase information from The QCMP itself has been described using several ap-
a density matrix in accordance with only certain SWAP gates proaches from computer science, many of which have allowed
from some constrained architecture (i.e., the black hole itself for the development of entirely new strategies for solving the
is described with respect to a background geometry, which problem. Our contribution here serves a different purpose. At
constrains which operations can be performed). As black the level of theoretical physics, quantum information theory,
holes are known to be the fastest information scramblers in and information geometry, we solve a simplified subproblem
nature [93], the QCMP can be viewed through the lens of of the QCMP. We have shown that solving this subproblem
a scrambling process, yielding the most-efficient method to quantifies a minimal bound, underlying the full QCMP. At
maximally mix the information of the IG density matrix, at a the limit where this bound is applicable, we assume implic-
given β value. This process exemplifies the traits of a quantum itly that: the quantum device is noiseless; the two-qubit gate
resource theory and we have shown that the quantity USWAP interactions given by the device can be performed with an in-
can in fact be optimized for using our technique. definite causal order; and that any number of two-qubit in-
As a counterexample, consider a hypothetical quantum teractions can be performed in tandem within one unit time
compiler which could surpass USWAP . One of the main as- slice. Our conceptualization of the QCMP provides, to our
 13

knowledge, the first instance of a solvable lower bound for lation of USWAP can be completed in O(n2 + c) steps,
the SWAP gate count when adapting a quantum circuit to a as per the Birkhoff-von-Neumann algorithm [72]. A fu-
connectivity-restricted processor. The SWAP uncomplexity ture research goal could involve making further scala-
was derived using tools from graph theory and quantum in- bility improvements to the qubit-assignment algorithm.
formation theory, using insights from quantum resource the- 3. Searching for the optimal β value. In our work, we
ory and quantum circuit complexity. Potential applications of have taken a somewhat naive approach to optimizing
uncomplexity for quantum machine learning are discussed in for β; however, because of the similarity of Figure 2 to
[94]. To the best of our knowledge, this work represents the a phase diagram, one may be able to use concepts from
first demonstration of an application of the concept of quan- condensed matter theory [95–97] in order to devise a
tum circuit uncomplexity to the realm of practical quantum suitable gradient-based optimization method.
information processing. 4. Analytical expression for tightness of the SWAP uncom-
Our purpose for deriving the bound has therefore been to plexity to the brute-force solution. We have given em-
inform and to create a meaningful method of comparison be- pirical evidence for tightness, but it still remains to de-
tween quantum compilers and strategies for solving the full fine an analytical expression for how similar in general
QCMP. This goal has been accomplished in several key ways. our solution is, compared to the brute-force solution
Firstly, the SWAP uncomplexity sheds light onto the physi- proposed in [85], and how tightness scales as the size
cal nature of the QCMP using information geometry, as well of the quantum circuit to be mapped increases in both
as theoretical physics. We have systematically shown that the register and depth.
concept of SWAP uncomplexity provides a minimal SWAP 5. Extension to incorporate bridge gates and
gate count for the full QCMP via considerations of the ther- teleportation-based quantum circuit mapping. There
modynamic length and Fisher information metric, providing are other methods commonly in use, in addition to
strong analytical arguments to supplant the empirical results the SWAP gate, for conforming a quantum circuit to
obtained in Section VI. Therefore, the use of SWAP uncom- hardware. Our approach is extendable for the Bridge
plexity as a metric for routing efficiency is pragmatic and jus- gates mentioned in [98–102], as well as the quantum
tified, and can be used to quantify and compare routing strate- teleportation-based protocols of [32, 60, 61, 103].
gies, as well as helping to inform architectural decisions by 6. Extension for quantum error correction codes, in par-
quantum architects and processor designers. ticular syndrome extraction circuits. It is well-known
Of independent interest may be the qubit-assignment al- that various fault-tolerance protocols are required in or-
gorithm which was designed to aid in the calculation of the der to ensure that quantum error correction codes func-
SWAP uncomplexity. This algorithm, grounded as a graph tion up to their full code distance [104–115]. As our
similarity search, inspects distinct n-qubit partitions of a given bound constitutes a non-trivial resource requirement, it
coupling graph, and returns the most-similar resultant to the may be useful to adapt fault-tolerance protocols further
interaction graph provided. Employing this method has en- to the setting of quantum compilation, in which an error
abled us to map circuits with up to 16 qubits onto devices with correction code is adapted to a device not specifically
up to 72 physical qubits. For larger circuits, we devised an designed for a particular code family [116, 117].
alternative approach. The initial placement precedes the min- 7. Extension for entanglement/qubit routing in quantum
imal SWAP gate count solution, which is further utilized for communications networks and modular architectures.
routing and minimal bound calculation. Additionally, we cal- Several other extensions may be possible as well, in-
culated a maximal bound by leveraging known classical graph cluding those allowing for bounds on the QCMP for
metrics. Both of these new structures provide additional tools modular scenarios [118] as well as for entanglement
of interest outside of the scope of this work. distribution in noisy quantum networks [119].
We would like to draw attention to several open problems Finally, we remark that the problem of assessing similar-
regarding our work, as well as several future possible direc- ities between two complex networks is a problem spanning
tions for research: many disciplines. Indeed, our work follows recent trends of
1. Formally associating the qJSD to circuit uncomplexity. utilizing quantum information theory and statistical mechan-
As we stated previously, it has been shown that both ics to study complex networks [39, 40, 44, 78]. As the task of
the quantum circuit uncomplexity as well as the qJSD comparing the distance between graphs appears in many dif-
are closely related to the Fisher information and Fubini- ferent areas of science [65, 120], we expect the implications of
Study metrics [42, 55, 71]. However, an exact relation- our work to stretch beyond the realm of quantum information
ship has not yet been established, as far as we are aware. science.
Establishing this relationship on more rigorous mathe-
matical footing could be very helpful.
2. Improvements to the subgraph similarity search algo-
rithm. In this work, although our qubit-assignment al- IX. SOFTWARE AVAILABILITY
gorithm outperforms the current state of the art solver
[86], we were still limited by the scalability of the qubit-
assignment algorithm constructed in Section V B. How- The software developed for this project is available at
ever, once a suitable qubit assignment is set, the calcu- https://github.com/QML-Group/QCMP-complexity-bound.
 14

X. ACKNOWLEDGEMENTS project “QuTech Part III Application-based research” (project

no. 601.QT.001 Part III-C—NISQ).
We thank Kenneth Goodenough, Hans van Someren, Pablo
le Henaff, Luise Prielinger, David Elkouss, and Tariq Bon-
tekoe for insightful discussions and useful manuscript feed-
XI. AUTHOR CONTRIBUTIONS
back. MS, MB, and SF are grateful for financial support
from Intel. CGA acknowledges support from Spanish Min-
isterio de Ciencia e Innovación and European ERDF un- MS, AS, and SS developed the theoretical framework and
der grant PID2021-123627OB-C51. CGA also acknowledges formalism. MB, SS, and AS developed the numerical algo-
support from the QuantERA grant EQUIP with the grant num- rithm based on the framework and implemented all numerical
ber PCI2022-133004, funded by Agencia Estatal de Inves- simulations. MB developed the subgraph isomorphism qubit-
tigación, Ministerio de Ciencia e Innovación, Gobierno de assignment algorithm. Part of this work was conducted during
España, MCIN/AEI/10.13039/501100011033, and by the Eu- the Master thesis for SS, and was supervised by MS, MB, and
ropean Union “NextGenerationEU/PRTR. AS acknowledges SF. CA and SF supervised and coordinated the project, and
funding from the Dutch Research Council (NWO) through the provided guidance during the writing process.

[1] M. Nielsen and I. Chuang, Quantum Computation and Quan- [11] M. Bandic, H. Zarein, E. Alarcon, and C. G. Almudever, On
tum Information, ISBN 978-1-107-00217-3 (Cambridge Uni- structured design space exploration for mapping of quantum
versity Press, Cambridge, UK, 2010). algorithms, in 2020 XXXV conference on design of circuits and
[2] J. Preskill, Quantum computing in the nisq era and beyond, integrated systems (DCIS) (IEEE, 2020) pp. 1–6.
Quantum 2, 79 (2018). [12] G. Li, Y. Ding, and Y. Xie, Tackling the qubit mapping
[3] C. Hempel, C. Maier, J. Romero, J. McClean, T. Monz, problem for nisq-era quantum devices, in Proceedings of the
H. Shen, P. Jurcevic, B. P. Lanyon, P. Love, R. Babbush, Twenty-Fourth International Conference on Architectural Sup-
A. Aspuru-Guzik, R. Blatt, and C. F. Roos, Quantum chem- port for Programming Languages and Operating Systems, AS-
istry calculations on a trapped-ion quantum simulator, Phys. PLOS ’19 (Association for Computing Machinery, New York,
Rev. X 8, 031022 (2018). NY, USA, 2019) p. 1001–1014.
[4] P. J. J. O’Malley, R. Babbush, I. D. Kivlichan, J. Romero, [13] M. Y. Siraichi, V. F. d. Santos, C. Collange, and F. M. Q.
J. R. McClean, R. Barends, J. Kelly, P. Roushan, A. Tran- Pereira, Qubit allocation, in Proceedings of the 2018 In-
ter, N. Ding, B. Campbell, Y. Chen, Z. Chen, B. Chiaro, ternational Symposium on Code Generation and Optimiza-
A. Dunsworth, A. G. Fowler, E. Jeffrey, E. Lucero, tion, CGO 2018 (Association for Computing Machinery, New
A. Megrant, J. Y. Mutus, M. Neeley, C. Neill, C. Quin- York, NY, USA, 2018) p. 113–125.
tana, D. Sank, A. Vainsencher, J. Wenner, T. C. White, P. V. [14] A. Cowtan, S. Dilkes, R. Duncan, A. Krajenbrink, W. Sim-
Coveney, P. J. Love, H. Neven, A. Aspuru-Guzik, and J. M. mons, and S. Sivarajah, On the qubit routing problem, in 14th
Martinis, Scalable quantum simulation of molecular energies, Conference on the Theory of Quantum Computation, Commu-
Phys. Rev. X 6, 031007 (2016). nication and Cryptography (TQC 2019) (Schloss Dagstuhl-
[5] D. G. Cory, M. D. Price, W. Maas, E. Knill, R. Laflamme, Leibniz-Zentrum fuer Informatik, 2019).
W. H. Zurek, T. F. Havel, and S. S. Somaroo, Experimental [15] N. Alon, F. R. Chung, and R. L. Graham, Routing permuta-
quantum error correction, Phys. Rev. Lett. 81, 2152 (1998). tions on graphs via matchings, in Proceedings of the twenty-
[6] J. Chiaverini, D. Leibfried, T. Schaetz, M. D. Barrett, fifth annual ACM symposium on Theory of Computing (1993)
R. Blakestad, J. Britton, W. M. Itano, J. D. Jost, E. Knill, pp. 583–591.
C. Langer, et al., Realization of quantum error correction, Na- [16] A. Paler, A. Zulehner, and R. Wille, Nisq circuit compilation is
ture 432, 602 (2004). the travelling salesman problem on a torus, Quantum Science
[7] B. M. Terhal, Quantum error correction for quantum memo- and Technology 6, 025016 (2021).
ries, Reviews of Modern Physics 87, 307 (2015). [17] T. Miltzow, L. Narins, Y. Okamoto, G. Rote, A. Thomas,
[8] P. Murali, N. M. Linke, M. Martonosi, A. J. Abhari, N. H. and T. Uno, Approximation and hardness for token swapping,
Nguyen, and C. H. Alderete, Full-stack, real-system quantum arXiv preprint arXiv:1602.05150 (2016).
computer studies: Architectural comparisons and design in- [18] F. Wagner, A. Bärmann, F. Liers, and M. Weissenbäck, Im-
sights, in Proceedings of the 46th International Symposium on proving quantum computation by optimized qubit routing,
Computer Architecture (2019) pp. 527–540. Journal of Optimization Theory and Applications , 1 (2023).
[9] M. Bandic, S. Feld, and C. G. Almudever, Full-stack quan- [19] P. Murali, J. M. Baker, A. Javadi-Abhari, F. T. Chong, and
tum computing systems in the nisq era: algorithm-driven and M. Martonosi, Noise-adaptive compiler mappings for noisy
hardware-aware compilation techniques, in 2022 Design, Au- intermediate-scale quantum computers, in International Con-
tomation & Test in Europe Conference & Exhibition (DATE) ference on Architectural Support for Programming Languages
(IEEE, 2022) pp. 1–6. and Operating Systems (2019) pp. 1015–1029.
[10] K. Bertels, A. Sarkar, T. Hubregtsen, M. Serrao, A. Moue- [20] S. S. Tannu and M. K. Qureshi, Not all qubits are created
denne, A. Yadav, A. Krol, I. Ashraf, and C. G. Almudever, equal: A case for variability-aware policies for NISQ-era
Quantum computer architecture toward full-stack quantum ac- quantum computers, in International Conference on Archi-
celerators, IEEE Transactions on Quantum Engineering 1, 1 tectural Support for Programming Languages and Operating
(2020). Systems (2019) pp. 987–999.
 15

[21] G. LI, Y. Ding, and Y. Xie, Towards efficient superconduct- [38] D. Harlow, Jerusalem lectures on black holes and quantum in-
ing quantum processor architecture design, in Proceedings formation, Rev. Mod. Phys. 88, 015002 (2016).
of the Twenty-Fifth International Conference on Architectural [39] M. De Domenico and J. Biamonte, Spectral entropies as
Support for Programming Languages and Operating Systems information-theoretic tools for complex network comparison,
(2020) pp. 1031–1045. Physical Review X 6, 10.1103/physrevx.6.041062 (2016).
[22] A. Zulehner, A. Paler, and R. Wille, An efficient methodology [40] M. Faccin and J. Biamonte, Complex networks from
for mapping quantum circuits to the IBM QX architectures, classical to quantum, Nature Communications Physics 2,
IEEE Transactions on Computer-Aided Design of Integrated 10.1038/s42005-019-0152-6 (2019).
Circuits and Systems (2018). [41] S. L. Braunstein, S. Ghosh, and S. Severini, The laplacian of
[23] D. Venturelli, M. Do, B. O’Gorman, J. Frank, E. Rieffel, a graph as a density matrix: A basic combinatorial approach
K. E. C. Booth, T. Nguyen, P. Narayan, and S. Nanda, Quan- to separability of mixed states, Annals of Combinatorics 10,
tum circuit compilation: An emerging application for auto- 291–317 (2006).
mated reasoning, in Scheduling and Planning Applications [42] A. R. Brown and L. Susskind, Second law of quantum com-
Workshop (2019). plexity, Phys. Rev. D 97, 086015 (2018).
[24] L. Lao, B. van Wee, I. Ashraf, J. van Someren, N. Khammassi, [43] N. Yunger Halpern, N. B. T. Kothakonda, J. Haferkamp,
K. Bertels, and C. Almudever, Mapping of lattice surgery- A. Munson, J. Eisert, and P. Faist, Resource theory of quan-
based quantum circuits on surface code architectures, Quan- tum uncomplexity, Phys. Rev. A 106, 062417 (2022).
tum Science and Technology 4, 015005 (2019). [44] A. Glos, A. Krawiec, and Ł. Pawela, Asymptotic entropy of
[25] L. Lao, D. M. Manzano, H. van Someren, I. Ashraf, and C. G. the gibbs state of complex networks, Scientific Reports 11, 1
Almudever, Mapping of quantum circuits onto nisq supercon- (2021).
ducting processors, arXiv preprint arXiv:1908.04226 (2019). [45] J. Briët and P. Harremoës, Properties of classical and quan-
[26] S. Herbert and A. Sengupta, Using reinforcement learning to tum jensen-shannon divergence, Physical Review A 79,
find efficient qubit routing policies for deployment in near- 10.1103/physreva.79.052311 (2009).
term quantum computers, arXiv:1812.11619 (2018). [46] P. W. Lamberti, A. P. Majtey, A. Borras, M. Casas,
[27] A. Lye, R. Wille, and R. Drechsler, Determining the minimal and A. Plastino, Metric character of the quantum jensen-
number of swap gates for multi-dimensional nearest neighbor shannon divergence, Physical Review A 77, 10.1103/phys-
quantum circuits, in Asia and South Pacific Design Automa- reva.77.052311 (2008).
tion Conference (2015) pp. 178–183. [47] M. Wolf, Quantum channels and operations: A guided tour
[28] S. Li, X. Zhou, and Y. Feng, Qubit mapping based on subgraph (2012).
isomorphism and filtered depth-limited search, IEEE Transac- [48] M. Christandl, Quantum information theory lecture notes
tions on Computers 70, 1777 (2020). (2018).
[29] A. Biuki, N. Mohammadzadeh, R. Wille, and S. Sargaran, Ex- [49] M. Nielsen, An introduction to majorization and its applica-
act mapping of quantum circuit partitions to building blocks of tions to quantum mechanics: Lecture notes (2002).
the saqip architecture, in 2022 IEEE Computer Society Annual [50] M. A. Steinberg, S. Feld, C. G. Almudever, M. Marthaler,
Symposium on VLSI (ISVLSI) (IEEE, 2022) pp. 402–405. and J.-M. Reiner, Topological-graph dependencies and scal-
[30] A. Molavi, A. Xu, M. Diges, L. Pick, S. Tannu, and A. Al- ing properties of a heuristic qubit-assignment algorithm, IEEE
barghouthi, Qubit mapping and routing via maxsat, in 2022 Transactions on Quantum Engineering 3, 1 (2022).
55th IEEE/ACM International Symposium on Microarchitec- [51] M. Y. Siraichi, V. F. d. Santos, C. Collange, and F. M. Q.
ture (MICRO) (IEEE, 2022) pp. 1078–1091. Pereira, Qubit allocation, in Proceedings of the 2018 Inter-
[31] L. Moro, M. G. Paris, M. Restelli, and E. Prati, Quantum national Symposium on Code Generation and Optimization
compiling by deep reinforcement learning, Communications (2018) pp. 113–125.
Physics 4, 178 (2021). [52] M. Y. Siraichi, V. F. d. Santos, C. Collange, and F. M. Q.
[32] D. Devulapalli, E. Schoute, A. Bapat, A. M. Childs, and Pereira, Qubit allocation as a combination of subgraph iso-
A. V. Gorshkov, Quantum routing with teleportation (2022), morphism and token swapping, Proceedings of the ACM on
arXiv:2204.04185 [quant-ph]. Programming Languages 3, 1 (2019).
[33] S. Upadhyay, A. A. Saki, R. O. Topaloglu, and S. Ghosh, A [53] M. Bandic, C. G. Almudever, and S. Feld, Interaction graph-
shuttle-efficient qubit mapper for trapped-ion quantum com- based characterization of quantum benchmarks for improving
puters, in Proceedings of the Great Lakes Symposium on VLSI quantum circuit mapping techniques, Quantum Machine Intel-
2022 (2022) pp. 305–308. ligence 5, 40 (2023).
[34] N. Nottingham, M. A. Perlin, R. White, H. Bernien, F. T. [54] T. Sagawa, Entropy, divergence, and majorization in classi-
Chong, and J. M. Baker, Decomposing and routing quantum cal and quantum thermodynamics (2020), arXiv:2007.09974
circuits under constraints for neutral atom architectures, arXiv [quant-ph].
preprint arXiv:2307.14996 (2023). [55] I. Bengtsson and K. Życzkowski, Geometry of quantum states:
[35] M. M. Wilde, Quantum Information Theory, 2nd ed. (Cam- an introduction to quantum entanglement (Cambridge univer-
bridge University Press, 2017). sity press, 2017).
[36] J. Watrous, The theory of quantum information (Cambridge [56] L. A. Zager and G. C. Verghese, Graph similarity scoring and
university press, 2018). matching, Applied mathematics letters 21, 86 (2008).
[37] A. Di Meglio, K. Jansen, I. Tavernelli, C. Alexandrou, [57] D. Koutra, A. Parikh, A. Ramdas, and J. Xiang, Algorithms
S. Arunachalam, C. W. Bauer, K. Borras, S. Carrazza, for graph similarity and subgraph matching, in Proc. Ecol. in-
A. Crippa, V. Croft, et al., Quantum computing for high- ference conf, Vol. 17 (Citeseer, 2011).
energy physics: State of the art and challenges. summary of [58] K. Samanvi and N. Sivadasan, Subgraph similarity search in
the qc4hep working group, arXiv preprint arXiv:2307.03236 large graphs, arXiv preprint arXiv:1512.05256 (2015).
(2023). [59] S. Niu and A. Todri-Sanial, Multi-programming cross plat-
form benchmarking for quantum computing hardware, arXiv
 16

preprint arXiv:2206.03144 (2022). Computing 12, 29 (2015).

[60] A. Bapat, A. M. Childs, A. V. Gorshkov, and E. Schoute, Ad- [80] G. Li et al., Tackling the qubit mapping problem for NISQ-
vantages and limitations of quantum routing, PRX Quantum era quantum devices, in International Conference on Archi-
4, 010313 (2023). tectural Support for Programming Languages and Operating
[61] S. Hillmich, A. Zulehner, and R. Wille, Exploiting quantum Systems (2019) pp. 1001–1014.
teleportation in quantum circuit mapping, in Proceedings of [81] H. Jiang, Y. Deng, and M. Xu, Quantum circuit transforma-
the 26th Asia and South Pacific Design Automation Confer- tion based on tabu search, arXiv preprint arXiv:2104.05214
ence, ASPDAC ’21 (Association for Computing Machinery, (2021).
New York, NY, USA, 2021) p. 792–797. [82] T. Peham, L. Burgholzer, and R. Wille, On optimal subarchi-
[62] Č. Brukner, Quantum causality, Nature Physics 10, 259 tectures for quantum circuit mapping, ACM Transactions on
(2014). Quantum Computing 4, 1 (2023).
[63] K. Goswami and J. Romero, Experiments on quantum causal- [83] L. P. Cordella, P. Foggia, C. Sansone, and M. Vento, A (sub)
ity, AVS Quantum Science 2 (2020). graph isomorphism algorithm for matching large graphs, IEEE
[64] J. J. Sakurai and E. D. Commins, Modern quantum mechanics, transactions on pattern analysis and machine intelligence 26,
revised edition (1995). 1367 (2004).
[65] S. H. Strogatz, Exploring complex networks, nature 410, 268 [84] Qiskit contributors, Qiskit: An open-source framework for
(2001). quantum computing (2023).
[66] F. Green, Review of handbook of graph theory, combinatorial [85] A. Lye, R. Wille, and R. Drechsler, Determining the minimal
optimization, and algorithms, SIGACT News 50, 6–11 (2019). number of swap gates for multi-dimensional nearest neighbor
[67] C. Godsil and G. F. Royle, Algebraic graph theory, Vol. 207 quantum circuits, in The 20th Asia and South Pacific Design
(Springer Science & Business Media, 2001). Automation Conference (2015) pp. 178–183.
[68] A. P. Majtey, P. W. Lamberti, and D. P. Prato, Jensen-shannon [86] A. Shafaei, M. Saeedi, and M. Pedram, Qubit placement to
divergence as a measure of distinguishability between mixed minimize communication overhead in 2d quantum architec-
quantum states, Phys. Rev. A 72, 052310 (2005). tures, in 2014 19th Asia and South Pacific Design Automation
[69] M. De Domenico, V. Nicosia, A. Arenas, and V. Latora, Struc- Conference (ASP-DAC) (2014) pp. 495–500.
tural reducibility of multilayer networks, Nature communica- [87] D. Freedman, R. Pisani, and R. Purves, Statistics (international
tions 6, 6864 (2015). student edition), Pisani, R. Purves, 4th edn. WW Norton &
[70] L. Rossi, A. Torsello, and E. R. Hancock, Attributed graph Company, New York (2007).
similarity from the quantum jensen-shannon divergence, in [88] L. Lao, H. van Someren, I. Ashraf, and C. G. Almudever,
Similarity-Based Pattern Recognition: Second International Timing and resource-aware mapping of quantum circuits to
Workshop, SIMBAD 2013, York, UK, July 3-5, 2013. Proceed- superconducting processors, IEEE Transactions on Computer-
ings 2 (Springer, 2013) pp. 204–218. Aided Design of Integrated Circuits and Systems (2021).
[71] G. E. Crooks, Measuring thermodynamic length, Phys. Rev. [89] S. M. Carroll, Spacetime and geometry (Cambridge University
Lett. 99, 100602 (2007). Press, 2019).
[72] D. M. Johnson, A. Dulmage, and N. Mendelsohn, On an al- [90] R. Penrose, Asymptotic properties of fields and space-times,
gorithm of g. birkhoff concerning doubly stochastic matrices, Phys. Rev. Lett. 10, 66 (1963).
Canadian Mathematical Bulletin 3, 237 (1960). [91] M. Scandi and M. Perarnau-Llobet, Thermodynamic length in
[73] P. Virtanen, R. Gommers, T. E. Oliphant, M. Haber- open quantum systems, Quantum 3, 197 (2019).
land, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, [92] S. Ito, Stochastic thermodynamic interpretation of information
W. Weckesser, J. Bright, S. J. van der Walt, M. Brett, J. Wil- geometry, Phys. Rev. Lett. 121, 030605 (2018).
son, K. J. Millman, N. Mayorov, A. R. J. Nelson, E. Jones, [93] Y. Sekino and L. Susskind, Fast scramblers, Journal of High
R. Kern, E. Larson, C. J. Carey, İ. Polat, Y. Feng, E. W. Moore, Energy Physics 2008, 065 (2008).
J. VanderPlas, D. Laxalde, J. Perktold, R. Cimrman, I. Hen- [94] A. Sarkar, Applications of Quantum Computation and Algo-
riksen, E. A. Quintero, C. R. Harris, A. M. Archibald, A. H. rithmic Information: for Causal Modeling in Genomics and
Ribeiro, F. Pedregosa, P. van Mulbregt, and SciPy 1.0 Con- Reinforcement Learning, Phd thesis, Delft University of Tech-
tributors, SciPy 1.0: Fundamental Algorithms for Scientific nology (2022).
Computing in Python, Nature Methods 17, 261 (2020). [95] N. Goldenfeld, Lectures on phase transitions and the renor-
[74] R. Cheng, Quantum geometric tensor (fubini-study metric) in malization group (CRC Press, 2018).
simple quantum system: A pedagogical introduction, arXiv [96] G. R. Brightwell and P. Winkler, Graph homomorphisms and
preprint arXiv:1012.1337 (2010). phase transitions, Journal of combinatorial theory, series B 77,
[75] F. Benatti and H. Narnhofer, Entropy behaviour under com- 221 (1999).
pletely positive maps, letters in mathematical physics 15, 325 [97] S. N. Dorogovtsev, A. V. Goltsev, and J. F. F. Mendes, Critical
(1988). phenomena in complex networks, Rev. Mod. Phys. 80, 1275
[76] E. Chitambar and G. Gour, Quantum resource theories, Rev. (2008).
Mod. Phys. 91, 025001 (2019). [98] S. Niu, A. Suau, G. Staffelbach, and A. Todri-Sanial, A
[77] P. Abiuso, H. J. Miller, M. Perarnau-Llobet, and M. Scandi, hardware-aware heuristic for the qubit mapping problem in
Geometric optimisation of quantum thermodynamic pro- the nisq era, IEEE Transactions on Quantum Engineering 1,
cesses, Entropy 22, 1076 (2020). 1 (2020).
[78] C. Nicolini, V. Vlasov, and A. Bifone, Thermodynamics of [99] T. Itoko, R. Raymond, T. Imamichi, and A. Matsuo, Optimiza-
network model fitting with spectral entropies, Phys. Rev. E 98, tion of quantum circuit mapping using gate transformation and
022322 (2018). commutation, Integration 70, 43 (2020).
[79] T. Bahreini and N. Mohammadzadeh, An MINLP model for [100] S. Park, D. Kim, M. Kweon, J.-Y. Sim, and S. Kang, A fast and
scheduling and placement of quantum circuits with a heuris- scalable qubit-mapping method for noisy intermediate-scale
tic solution approach, Journal on Emerhing Technologies in quantum computers, in Proceedings of the 59th ACM/IEEE
 17

Design Automation Conference, DAC ’22 (Association for [120] R. Albert and A.-L. Barabási, Statistical mechanics of com-
Computing Machinery, New York, NY, USA, 2022) p. 13–18. plex networks, Rev. Mod. Phys. 74, 47 (2002).
[101] J. Liu, E. Younis, M. Weiden, P. Hovland, J. Ku- [121] P. F. C. Sansone, M. Vento, L. Cordella, P. Foggia, C. San-
biatowicz, and C. Iancu, Tackling the qubit mapping sone, and M. Vento, An improved algorithm for matching
problem with permutation-aware synthesis, arXiv preprint large graphs, in Proc. of the 3rd IAPR-TC-15 International
arXiv:2305.02939 (2023). Workshop on Graph-based Representations, Vol. 57 (2001).
[102] S. Park, D. Kim, J.-Y. Sim, and S. Kang, Mcqa: Multi- [122] V. Bonnici, R. Giugno, A. Pulvirenti, D. Shasha, and A. Ferro,
constraint qubit allocation for near-ftqc device, in Proceedings A subgraph isomorphism algorithm and its application to bio-
of the 41st IEEE/ACM International Conference on Computer- chemical data, BMC bioinformatics 14, 1 (2013).
Aided Design, ICCAD ’22 (Association for Computing Ma- [123] C. Solnon, Alldifferent-based filtering for subgraph isomor-
chinery, New York, NY, USA, 2022). phism, Artificial Intelligence 174, 850 (2010).
[103] S. Hillmich, A. Zulehner, and R. Wille, Exploiting quantum [124] J. R. Ullmann, Bit-vector algorithms for binary constraint sat-
teleportation in quantum circuit mapping, in 2021 26th Asia isfaction and subgraph isomorphism, Journal of Experimental
and South Pacific Design Automation Conference (ASP-DAC) Algorithmics (JEA) 15, 1 (2011).
(IEEE, 2021) pp. 792–797. [125] R. M. Roth, Introduction to coding theory, IET Communica-
[104] J. Preskill, Reliable quantum computers, Proceedings of the tions 47, 4 (2006).
Royal Society of London. Series A: Mathematical, Physical [126] Z. Abu-Aisheh, R. Raveaux, J.-Y. Ramel, and P. Martineau,
and Engineering Sciences 454, 385 (1998). An exact graph edit distance algorithm for solving pattern
[105] D. Bhatnagar, M. Steinberg, D. Elkouss, C. G. Almudever, and recognition problems, in 4th International Conference on Pat-
S. Feld, Low-depth flag-style syndrome extraction for small tern Recognition Applications and Methods 2015 (2015).
quantum error-correction codes, in 2023 IEEE International [127] P. H. Sellers, On the theory and computation of evolution-
Conference on Quantum Computing and Engineering (QCE), ary distances, SIAM Journal on Applied Mathematics 26, 787
Vol. 01 (2023) pp. 63–69. (1974).
[106] D. Gottesman, An introduction to quantum error correction [128] J. Nocedal and S. J. Wright, Numerical optimization (Springer,
and fault-tolerant quantum computation, in Quantum informa- 1999).
tion science and its contributions to mathematics, Proceedings [129] S. J. Russell and P. Norvig, Artificial intelligence a modern
of Symposia in Applied Mathematics, Vol. 68 (2010) pp. 13– approach (London, 2010).
58. [130] J. M. Hernández and P. Van Mieghem, Classification of graph
[107] D. Gottesman, Theory of fault-tolerant quantum computation, metrics, Delft University of Technology: Mekelweg, The
Phys. Rev. A 57, 127 (1998). Netherlands , 1 (2011).
[108] C. Chamberland and M. E. Beverland, Flag fault-tolerant er- [131] A collection of quantum circuits for benchmarking quantum
ror correction with arbitrary distance codes, Quantum 2, 53 system software, https://github.com/QML-Group/
(2018). qbench (2023 (accessed Jan 31, 2024)).
[109] A. M. Steane, Overhead and noise threshold of fault-tolerant
quantum error correction, Phys. Rev. A 68, 042322 (2003).
[110] A. W. Cross, D. P. DiVincenzo, and B. M. Terhal, A compar-
ative code study for quantum fault-tolerance, arXiv preprint
arXiv:0711.1556 (2007).
[111] E. T. Campbell, B. M. Terhal, and C. Vuillot, Roads towards
fault-tolerant universal quantum computation, Nature 549, 172
(2017).
[112] P. Aliferis, D. Gottesman, and J. Preskill, Quantum accuracy
threshold for concatenated distance-3 codes, arXiv preprint
quant-ph/0504218 (2005).
[113] D. Aharonov and M. Ben-Or, Fault-tolerant quantum compu-
tation with constant error, in Proceedings of the twenty-ninth
annual ACM symposium on Theory of computing (1997) pp.
176–188.
[114] E. Knill, Quantum computing with realistically noisy devices,
Nature 434, 39 (2005).
[115] R. Chao and B. W. Reichardt, Quantum error correction with
only two extra qubits, Phys. Rev. Lett. 121, 050502 (2018).
[116] D. A. Lidar and T. A. Brun, Quantum error correction (Cam-
bridge university press, 2013).
[117] L. Lao and C. G. Almudever, Fault-tolerant quantum error cor-
rection on near-term quantum processors using flag and bridge
qubits, Phys. Rev. A 101, 032333 (2020).
[118] M. Bandic, L. Prielinger, J. Nüßlein, A. Ovide, S. Rodrigo,
S. Abadal, H. van Someren, G. Vardoyan, E. Alarcon, C. G.
Almudever, et al., Mapping quantum circuits to modular archi-
tectures with qubo, arXiv preprint arXiv:2305.06687 (2023).
[119] M. Pant, H. Krovi, D. Towsley, L. Tassiulas, L. Jiang, P. Basu,
D. Englund, and S. Guha, Routing entanglement in the quan-
tum internet, npj Quantum Information 5, 25 (2019).
 18

Appendix A: Graph Theory target graphs. It starts with an empty matching and gradually
extends the matching by adding pairs of nodes, one from the
In Figure 1, the IG and CG are examples of simple, undi- pattern graph and one from the target graph, until a complete
rected graphs [66], with simple referring to a restriction of matching is found or it is determined that no matching exists.
only one edge between any two vertices, and undirected It has a worst-case time complexity of O(|V |2 ∗ |E|), where
meaning without directionality indicated by arrows. For this |V | is the number of nodes in the graph and |E| is the number
work, we restrict ourselves to this regime only. An example of of edges.
a simple, undirected graph can be seen in Figure 1B. Here, we If an exact isomorphism cannot be found, one must choose
define a graph as an ordered-pair object G(EG , VG ) with edge a subgraph known to be close to a graph isomorphism, i.e., as
set EG = {eij |i, j ∈ VG , i ̸= j} and vertex set VG contain- close to a graph homomorphism as possible. Just as for the
ing all nodes of G [67]. Additionally, we define the adjacency SIP, many approaches exist for extending into the regime of
matrix A, degree matrix D, and combinatorial graph Lapla- subgraphs [56–58] exhibiting graph homomorphisms. Typ-
cian L from literature [39, 40]. As usual, the graph Laplacian ically, one employs some graph-theoretic distance metric in
takes on the form L = D − A. order to locate the most similar subgraph [125–127].
From here onwards, we will refer to the combinatorial Finally, as we will require the use of the graph-edit distance
graph Laplacian as simply the graph Laplacian. Several (GED) in Section V B, we give a brief definition of this met-
forms of the graph Laplacian exist in the literature [39– ric. The GED itself is known as a classical similarity measure
41, 44, 66, 67]; as such, we have opted to use the definition between two graphs G and H [66, 67]. Given a set of graph-
of the graph Laplacian which upholds the triangle inequality edit operations (such as edge addition or edge removal in our
[39, 40]. The graph Laplacian in the current context is known case), one may define the GED as:
to be symmetric and positive semidefinite. X
Let us also define the notions of graph isomorphism and GED(G, H) = min c(ei ), (A1)
{e1 ...ei }∈Sops(G,H)
i
homomorphisms, as we will use these later. For two graphs
G(EG , VG ) and H(EH , VH ) with |VG | = |VH |, we say that where {ei } ∈ Sops(G,H) represents the set of graph-edit op-
G and H are cospectral if they share the same eigenvalue erations along all possible graph-edit paths between graphs G
spectrum [67]. Note that the graph Laplacian’s eigenvalue and H. c(ei ) is the cost of the graph-edit operation. An exact
spectrum is not sufficient in order to determine whether or algorithmic implementation of the GED usually can be found
not two graphs are the same; this reason motivates us towards in the A∗ search algorithm, wherein the problem of finding the
the treatment described in Section III and originally treated in minimal graph-edit cost is transformed into a shortest-path al-
[39, 40] with Gibbs states. gorithm [128, 129]. However, in this work, we introduce an
Additionally, we say that G and H are graph homomor- algorithm based on the depth-first GED (DF-GED) algorithm
phic if there exists a map G 7→ H such that the vertex and [126], which we detail in Section V B.
edge connectivities of the two graphs are preserved; we rep-
resent this relation in the text as G ≜ H. The task of de-
termining whether or not two graphs exhibit a graph isomor- Appendix B: Quantum Information Theory
phism or homomorphism is in general NP-complete [66, 67],
although certain exceptions exist. If we are looking for an In this section, we review some of the basics of quantum
embedding of graph G onto graph H such that, for some sub- information theory which may be useful later. We refer the
graph H ′ ⊆ H, we have G ≜ H ′ . The generalization of reader to [1, 35, 36] for a more nuanced treatise of quantum
this problem is known as the subgraph-isomorphism problem information theory.
(SIP). As we allow |VG | ⊆ |VH |, the task adds an additional In quantum computation and quantum information theory,
layer of complexity, since we must identify a suitable sub- the fundamental information unit is known as a qubit. Any
graph for comparison; this subgraph is known as an induced pure qubit state can represented as:
subgraph, and is defined as a subgraph G′ ⊆ G for which
VG′ ⊆ VG and EG′ ⊆ EG . |ψ⟩ = α |0⟩ + β |1⟩ , (B1)
We further remark that solving the SIP efficiently is still an
active area of research, and various approaches exist for ad- where |α|2 + |β|2 = 1. Qubits are defined in a Hilbert space
dressing it. These methods range from brute-force approaches H; these are known as inner-product spaces on the field of
to more sophisticated algorithms that exploit specific proper- complex numbers.
ties of the graphs being analyzed. Some of the most common Representing systems of many qubits becomes a cumber-
algorithms used for solving the SIP include the Vento-Foggia some task when such systems are entangled with other sys-
algorithm (VF2) [83, 121], RI [122], The LAD (Labeled tems. For pure quantum states, it suffices to utilize Dirac nota-
Anatomy Directed) [123] and Ullman [124] algorithms. Each tion, but, when mixed states are involved, one conventionally
algorithm has its strengths and weaknesses, and the choice uses the language of density matrices ρ, whose explicit form
depends on the specific requirements of the application. VF2 is given by:
algorithm is a most commonly used algorithm for solving the X
SIP known for its speed and effectiveness. It works by main- ρ= pj |ψ⟩ ⟨ψ| , (B2)
taining a partial matching between the nodes of the pattern and j
 19

where pj represents the probability of each pure state in the allow for the construction of a method to compare density ma-
ensemble. trices.
Density matrices exhibit several important properties for We have, until now, mainly discussed static properties of
the present work. Density matrices in general are: density matrices in the formalism of quantum physics. For
• Defined formally as objects ρ such that ρ ∈ Md (C) ∼= considering evolutions, let us introduce the notion of a quan-
B(Hd ), equipped with a Hilbert-Schmidt scalar prod- tum operation (or quantum evolution) [1, 35, 36]. Succinctly,
uct as usual (where d is the dimension of the Hilbert a quantum operation is a map from the space of density ma-
space). Here, Md (C) represents the set of complex- trices to the space of density matrices. More formally, we
valued d × d square matrices, and B(Hd ) is the set of define a quantum operation as a map Λ : Md (C) 7→ Md (C)
bounded linear operators on a Hilbert space [47]; that is both linear and completely positive (CP). We may ad-
• Normalized, such that Tr(|ψ⟩ ⟨ψ|) = 1; ditionally equip Λ(·) with the property of trace preservation
• Positive, such that ρ ≥ 0; (TP), in which case we refer to this operation as a quan-
• Hermitian, meaning ρ = ρ† ; and tum channel (or completely-positive, trace-preserving (CPTP)
• Projectors, where ρ = ρ2 . map) [1, 35, 47, 48]. Finally, one may also impose unitality
In quantum information theory, we customarily distinguish as an extra condition.
the amount of entanglement present in multipartite quantum Although there are several representations of quantum
systems using the Von Neumann entropy (VNE) [1, 49], which channels [47, 48], in this work we will utilize the Kraus-
has the following definition: operator representation of a quantum channel, described as:

Ei ρEi† = σ,
X
S(ρ) = − Tr(ρ log ρ), (B3) Λ(ρ) = (B5)
i
where all logarithms are of natural base. The VNE exhibits
th
several interesting properties: where Ei represents the i term in the sum of operators and
• Permutation-invariance with respect to subsystem σ represents the output density matrix.
ordering: that is, given a multipartite quantum state, Measurement operations can also readily be described us-
the VNE is invariant under the specific ordering of sub- ing the quantum-operation formalism; in the case of our work,
systems that we choose. For example, given a state we will utilize the case of projective measurements, operators
vector of five subsystems
P a, b, c, d, and e, we have Πi that fulfill the relation Π2i = Πi . For representing a pro-
S(ρ)P = S(σ), if ρ = abcde∈Z2 |abcde⟩ ⟨abcde|, and jective measurement in√the Kraus formalism, we simply set
σ = abcde∈Z2 |baced⟩ ⟨baced|. the convention Ei = Πi . More information about projec-
• Unitary-transformation invariance: a density matrix tive measurements in quantum information theory and graph
ρ is invariant under S(ρ) = S(U ρU † ), where U is a theory can be found in [1, 41].
unitary transformation.
• Additivity: the VNE is additive for independent sys-
tems. For example, for independent subsystems A and Appendix C: Calculating the Maximal Bound
B, we have S(ρA ⊗ ρB ) = S(ρA ) + S(ρB ).
• Subadditivity: crucially, we see that a conjoined sys- We briefly introduce here a metric related to qubit assign-
tem AB has S(ρAB ) ≤ S(ρA ) + S(ρB ). ment in order to approximate a maximal bound. This bound
The task of distinguishing two quantum states can be ad- was employed in Section VI. The metric MSWAP involves as-
dressed using entropic divergences [36, 45, 46, 49, 68, 69]. sessing the diameter D(·) [130] of the CG subgraph G′CG that
For our purposes, we define the quantum Jensen-Shannon di- is selected for qubit assignment. We then simply multiply the
vergence (qJSD) as: diameter with the amount of two-qubit gates, equivalent to the
number of edges, |EIG | in the circuit. This forms our approxi-
ρ + σ
1

mate maximal bound for the number of SWAP gates, and takes
DqJSD (ρ||σ) = S − S(ρ) + S(σ) . (B4)
2 2 on the following algebraic form:
The quantum Jensen-Shannon divergence has several useful MSWAP (EIG , G′CG ) = |EIG | · D(G′CG ) − 1 .


(C1)
properties, but arguably the most important one originates
from the square root of the qJSD lies in a metric space D(x, y) By considering the maximal distance between any two
for two objects x, y that we wish to distinguish. A metric points in the chosen CG subgraph and imagining the worst-
space is endowed with the properties of: case scenario in which all the two-qubit gates are on the path
• Distance: Let x, y, z be the elements inside a set X, between those two points, we get this maximal bound. This
then the function D : X × X 7→ R upholds D(x, y) ≥ approach, although approximate, offers a broader perspective
0, with the case of D = 0 if x = y. on the possible SWAP overhead associated with the qubit as-
• Symmetricity: The function D(x, y) also obeys signment, enhancing the depth and utility of our analysis. It is
D(x, y) = D(y, x). worth noting that, unlike the minimal bound derived in Sec-
• Adherence to the Triangle inequality: lastly, tion IV, our approach for the maximal bound involves using
D(x, y) + D(x, z) ≥ D(y, z). weighted IGs, taking into account the two-qubit gate depth of
If these conditions are all upheld, we say that D(·, ·) is the underlying circuit; this makes our maximal bound specific
known as a metric space [46]. In this way, the VNE and qJSD to each benchmark.
 20

Appendix D: Benchmarks and quantum devices used for

experiments

TABLE I. Quantum devices used for simulations: We chose 16 of the

most renowned device layouts of superconducting technology. We
opted for superconducting qubits as the limit in connectivity charac-
terizes them. The coupling graphs range in size from 5 to 72 qubits.
Quantum device Number of qubits
IBM Athens 5
QuTech Starmon-5 5
IBM Yorktown 5
IBM Ourense 5
QuTech Surface-7 7
IBM Casablanca 7
Rigetti Agave 8
IBM Melbourne 15
Rigetti Aspen-1 16
QuTech Surface-17 17
IBM Singapore 20
IBM Johannesburg 20
IBM Tokyo 20
IBM Paris 27
IBM Rochester 53
Google Bristlecone 72
 21

TABLE II. Benchmarks used for experiments taken from [131]. The benchmarks are characterized by different IG connectivities and range
from 3 to 20 qubits.
Benchmark Number of qubits Number of gates 2-qubit gate %
basis change n3 3 79 0,126582278
fredkin n3 3 51 0,156862745
grover n3 3 102 0,117647059
teleportation n3 3 20 0,1
adder n4 4 63 0,158730159
bell n4 4 66 0,106060606
cuccaroAdder 1b 4 83 0,204819277
q=4 s=19996 2qbf=02 1 4 20000 0,20365
q=4 s=2996 2qbf=08 1 4 3000 0,814
variational n4 4 94 0,170212766
vbeAdder 1b 4 74 0,189189189
4gt10-v1 81 5 424 0,155660377
4gt13 92 5 190 0,157894737
4gt5 75 5 239 0,158995816
alu-v1 28 5 105 0,171428571
alu-v2 31 5 1295 0,152895753
decod24-v1 41 5 241 0,157676349
error correctiond3 n5 5 278 0,176258993
q=5 s=2995 2qbf=09 1 5 3000 0,899
qec en n5 5 61 0,163934426
qec sm n5 5 61 0,163934426
quantum volume n5 5 411 0,124087591
simon n6 5 92 0,152173913
4gt12-v0 87 6 711 0,157524613
4gt4-v0 72 6 740 0,152702703
alu-v2 30 6 1446 0,154218534
ex3 229 6 1153 0,151777971
graycode6 47 6 15 0,333333333
mod5adder 127 6 1577 0,151553583
q=6 s=2994 2qbf=08 1 6 3000 0,802
q=6 s=54 2qbf=022 1 6 60 0,233333333
qaoa n6 6 528 0,102272727
sf 274 6 2221 0,151283206
xor5 254 6 17 0,294117647
4mod5-bdd 287 7 196 0,158163265
alu-bdd 288 7 240 0,158333333
C17 204 7 1341 0,152870992
ham7 104 7 922 0,161605206
majority 239 7 1754 0,152223489
q=7 s=2993 2qbf=08 1 7 3000 0,795333333
q=7 s=29993 2qbf=08 1 7 30000 0,799866667
dnn n8 8 1904 0,100840336
f2 232 8 3456 0,151909722
hwb7 59 8 70093 0,152383262
q=8 s=2992 2qbf=01 1 8 3000 0,091666667
vqe uccsd n8 8 24136 0,22737819
q=9 s=19991 2qbf=08 1 9 20000 0,79645
q=9 s=2991 2qbf=01 1 9 3000 0,101
q=9 s=51 2qbf=012 1 9 60 0,116666667
adder n10 10 328 0,198170732
q=10 s=990 2qbf=091 1 10 1000 0,899
sqn 258 10 29333 0,152013091
sym9 148 10 61824 0,152173913
shor 15 11 13588 0,131586694
16QBT 100CYC QSE 1 16 1776 0,18018018
20QBT 45CYC 0D1 2D2 0 20 270 0,333333333