Abstract
We analyse a continuous-time quantum walk on a chimera graph, which is a graph of choice for designing quantum annealers, and we discover beautiful quantum walk features such as localization that starkly distinguishes classical from quantum behaviour. Motivated by technological thrusts, we study continuous-time quantum walk on enhanced variants of the chimera graph and on diminished chimera graph with a random removal of vertices. We explain the quantum walk by constructing a generating set for a suitable subgroup of graph isomorphisms and corresponding symmetry operators that commute with the quantum walk Hamiltonian; the Hamiltonian and these symmetry operators provide a complete set of labels for the spectrum and the stationary states. Our quantum walk characterization of the chimera graph and its variants yields valuable insights into graphs used for designing quantum-annealers.
Export citation and abstract BibTeX RIS
Original content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
1. Introduction
The chimera graph [1], which we henceforth call χ, is the hardware graph for D-Wave computers. This graph underpins the chip design for computational tasks such as optimization [2–4], graph partitioning [5], and machine learning [6–8], which classical computers and algorithms struggle to perform efficiently and accurately. D-Wave computers solve problems by converting the problem graph into its hardware graph, using the minor embedding technique of representing one logical qubit by several physical qubits [9–11]. Long-range connections in χ [12, 13] are absent due to implementation constraints [14], which necessitates an embedding of problem graphs into the incomplete χ graph using the minor embedding technique [13–15]. Therefore, modifying the current D-Wave computers graph structure is important, and studying the symmetries of χ may offer another perspective to give insights to reduce overheads or improve minor embedding technology.
The quantum walk (QW) [16], which quantizes the ubiquitous classical random walk [17], has become a rich area of theoretical [18–20] and experimental [21–23] study for quantum computation [24, 25], quantum search [26], photosynthetic energy transfer [27–29], topological phases [30, 31], quantum algorithms [32, 33], quantum transport [34, 35] and the foundational quantum–classical divide [36]. Two equivalent versions are the discrete QW [37, 38] and the continuous-time QW [24, 39]. For the discrete QW, the evolution operator is applied in discrete time steps, whereas the evolution operator is defined for all times for the continuous-time QW [40]. In both discrete and continuous models, the topology on which QW is performed and its properties studied are discrete graphs [40]. The topology of the graph has significant affect in the evolution.
Through simulating the continuous-time QW on χ and its variants with different boundary conditions and initial positions, we discover that QW localization behaviour is starkly different from the classical walk. We study symmetries of χ and the role they play in state evolution. The properties of the walker's evolution on χ reflect the structures of the graphs thereby indicating that our QW approach could be used for state engineering and quantum-gate design. Our method provides a way to investigate the structure of graphs and sheds light on how to understand the symmetries of χ, which is used on the D-Wave machine. We are studying a single-particle QW, and quantum annealing on a χ graph involves the ground-state of a many-body Hamiltonian so how much we can learn from the single-particle QW is inherently limited. The nature of the many-body ground-state problem for classical Ising models [41] and quantum model [42] provides a rigorous and valuable complementary analyses for many-body cases.
The structure of our article is as follows. We provide relevant background regarding the continuous-time QW and χ in section 2. In section 3 we present our numerical simulation of the continuous-time QW on χ and its variants. We discuss the graph symmetries and spectra of the QW Hamiltonian in section 4. In section 5, we discuss the QW behaviour on chimera graph and its variants. We present our conclusions in section 6.
2. Background
In this section, we give a brief introduction regarding the continuous-time QW and χ.
2.1. Continuous-time quantum walk
The QW was originally motivated by the widespread use of the classical random walk in the design of randomized algorithms [32, 33]. A quantum walker's evolution on a graph is described by the Schrödinger equation [40]. In the continuous-time QW, a single-particle evolves over a connected graph , for the set of vertices of the graph and the set of edges where each edge can be expressed by the pair of vertices it connects. The Hamiltonian determining the walker's evolution corresponds to the adjacency matrix representation of the edges, and the edges can have weights , which are the couplings between pair of vertices.
The bridge from graphs to quantum dynamics is achieved by constructing the orthogonal vertex basis . The continuous-time QW evolution is generated by the adjacency matrix, serving as the Hamiltonian
with the onsite frequency term, and the transition rates between vertices (edge weights), and and the intracell and intercell edges. Later we allow h, j and k to be time-dependent to study adiabatic evolution [43], but initially we consider a time-independent whose rows and columns add to zero (thereby constraining) in order to generate an evolution operator whose rows and columns sum to one analogous to the properties of a bistochastic matrix [44].
For a time-independent , the unitary evolution operator is starting at time t = 0 (ℏ ≡ 1), and the evolving walker wavefunction is . If the walker starts at the vertex then the walker's state at time t is described by
Unlike the classical random walk [17], the walker in the QW is in a superposition of vertices. The transition amplitude from the vertex to at time t is
and the corresponding transition probability is [34]
The long-time average of , which is known as the limiting probability [39]
embodies a natural notion of QW convergence and captures the amount of time the walker spends in each subset of the nodes [37].
2.2. Chimera topology
The chimera graph , is the underlying topology for the D-Wave Two, D-Wave 2X, and D-Wave 2000Q [1, 45]. D-Wave Systems'M × N grid of unit cells is realized as a rectangular L × 2 qubit array [1], denoted here as χMNL. The structure of χ is shown in figure 1, which has MN unit cells and qubits. The chimera vertices shown in figure 1 are noted as
for m, n the cartesian coordinates of the unit cells and μ the vertex label for the L × 2 cell with μ ≤ L for left side and μ > L for the right side.
The edge set is
Intracell edges are given by the complete bipartite graph KLL [46] connecting each left vertex to all right vertices and vice versa. Intercell edges are established by connecting each left vertex of a unit cell to the corresponding left vertex in the cells above and below and by connecting each right vertex to the corresponding right vertex in the neighbouring unit cells left and right. Edges cross in χ, so χ, as well as the variants we consider, are nonplanar graphs.
The connectivity of χ affects the efficiency of quantum algorithms implemented on D-Wave computers. Motivated by technological thrusts, we study the continuous-time QW on χ's variants. Now we describe χ and its variants, with each considered under two different boundary-conditions indicated by superscripts: periodic, labelled , which is equivalent to the graph being on the torus, and reflecting, labelled . Our enhanced chimera graph, denoted by χ+, adds vertical intracell connections: the top and bottom vertices are connected for L = 2 on the left and right side; the top and bottom vertices are each connected to the middle vertex for L = 3 on the left and right side; and, for L ≥ 4, the top left and right are connected to the second-from-top left and right, respectively, and the bottom left and right are connected to the second-from-bottom left and right, respectively. We show χ+ in figure 1.
We also introduce the diminished chimera graph χ− resulting from randomly deleting 2% of the vertices, which corresponds to the case of real-world quantum annealers with a certain fraction of its vertices not working properly, and 2% is the failure rate for the D-Wave 2000Q [45]. Open-system effects play an important role in how D-Wave quantum computers works and open-system dynamics can even be beneficial: quantum annealing could exploit a thermal environment to speed up compared with closed-system dynamics [47]. Despite potential advantages of open-system dynamics, our focus is on the closed-system QW on chimera graph as the closed system is advantageous for unitary quantum computing and the chief advantage of the QW analysis is in discovering graph symmetries and effects of localization.
3. Quantum walk on chimera graph and its variants
In this section, we study the continuous-time QW on χ and its variants with different boundary conditions and initial positions. We compare the continuous-time QW and the random walk on χ.
First we consider time-independent with in equation (1) and constrain by requiring that rows and columns add to zero. If the walker is initialized at a single vertex in the left side of a cell, we observe that the QW is localized vertically as shown in figure 2(a), whereas initialization in the right side of a cell yields a horizontal localization as shown in figure 2(b) for the probability distribution
and the limiting probability distribution
In contrast, the classical random walk is seen to spread in all directions as shown in figure 2(c). The classical random walk probability distribution is for the stochastic walker's likelihood to be at any vertex [34]. The QW exhibits strong localization whereas the classical walk does not; this localization is evident in other QWs as well [48–50]. The contrast between periodic and boundary conditions is evident when the walker has evolved to reach a boundary and shows different interference effects, shown as limiting probability distributions in figures 2(e), (f).
Download figure:
Standard image High-resolution image Figure data fileThe quantum walker on χ+ also shows the similar localization behaviour as on χr , that is, if the walker is initialized at a single vertex in the left side of a cell, we observe that the QW is localized vertically as shown in figure 2(d) for the limiting probability distribution, whereas initialization in the right side of a cell yields a horizontal localization. The walker's distribution inside unit cells of χ+ is divergent compared with on χr as shown in figures 2(d), (f), which reflects the difference between the unit cell structures.
Consideration of χ− is driven by the experimental problem where 2% vertices do not work [45], Thus, we consider the class of random χ with uniform deletion of 2% of vertices, i.e., χ− with 2% of vertices deleted. We observe that the effect of broken vertices can be divided into two cases. In the first case, we forbid any broken vertices sharing the same m and n as the walker, whose initial position is (m, n, μ). In this case, broken vertices have little effect on the walker's dynamics, as seen by comparing figures 3(a) to 2(f). The only noticeable difference is in the low-probability features appearing as dark lines radiating orthogonal to the walker's line of confinement and also diagonally in figure 2(f) but not in figure 3(a). Quantitatively, we calculate the 1-norm distance between these two limiting probability distributions as
In the second case, we consider at least one broken vertex located with the same m or n or both. In this second case, the broken vertices have a significant effect on QW behaviour, seen by comparing figure 3(b) with figure 2(f). the distance between those two limiting probability distribution is
and the qualitative structure of the curve has been dramatically changed.
Download figure:
Standard image High-resolution image Figure data file4. Symmetry and spectrum analysis
From our simulation, we observe that the quantum walker is strongly localized, which is related to the structure of the graph. To analyse and explain the QW localized behaviour, we now determine the graph symmetries. Those symmetries manifest as self-adjoint operators that commute with , i.e.,
These symmetry operators provide sufficient eigenvalues to lift the energy degeneracy for eigenstates of and thus label all eigenstates uniquely. In quantum mechanics, this set of operators is also known as a complete set of commuting observables. We construct by determining a sufficient set of symmetries of χ, arising from a sufficient generating set for a subgroup of χ's automorphism group, and we obtain this sufficient generating set.
A graph automorphism is constructed from a vertex permutation such that an edge exists if and only if is an edge in . For each automorphism identified by vertex permutation σ, the corresponding symmetry operator is
If matrix S is not Hermitian, we replace S by
for S† the adjoint so we discuss operators as being self-adjoint below.
4.1. Symmetry and spectrum analysis for
In this subsection we study the symmetries of the periodic chimera graph χp. Figure 1 shows the symmetries of . Specifically, σ1,3 is an intracell left-side (right-side) translation permutation symmetry, is an intracell left-side (right-side) mirror permutation symmetry, σ5,7 is a translation permutation along the M (N) axis, and is a mirror permutation through the M (N) axis. The symmetry operators are
with matrices Si obtained from equation (13) and satisfying equation (12). Each Si is (not) Hermitian for even (odd) i and replaced by for odd i. Table 1 shows the detail of σi for . The first column shows the vertex permutation; the second column shows the vertex coordinate defined in equation (6); the third column shows the vertex coordinate after permutation; the fourth column gives the eigenvalues of S corresponding to σ given by equation (13).
Table 1. The permutation operators resulted from graph automorphism for
Vertex permutation | Eigenvalues of S | ||
---|---|---|---|
To analyse and explain the QW localized behaviour, we construct the spectrum and stationary states. Stationary states are eigenvectors, with each uniquely labelled by energy E and spectrum
We parameterize by unit-lattice coordinates
expressed as a disjoint union of two three-dimensional unit lattices and one two-dimensional lattices. For
these two three-dimensional lattices are
and
is the two-dimensional unit lattice.
Spectral labels are expressed as and . For a valid ,
and the energy is
The cardinalities are
In this subsection, we have obtained the spectrum for the χp QW Hamiltonian. Also we divided the set of stationary states into three cases for interpretation in section 5.
4.2. Symmetry and spectrum analysis for
In this subsection we study the symmetries of the reflecting chimera graph χr. Both and have identical intracell symmetries; i.e., for , the vertex permutation symmetries σ1,2,3,4 still hold. However, due to reflecting boundary conditions, translation permutation symmetry along the M (N) axis no longer holds in the reflecting case. In order to lift the intercell degeneracy of the QW , we need to construct other operators. Notice that the QW of the one-dimensional finite line has no degeneracy, so we write the matrix representation of directly as
with AM the QW for the one-dimensional finite line with M vertices.
The elements of AM are
and the eigenvalues of are
Thus, for , the symmetry operators are
with matrices Si obtained from equation (13).
Stationary states are eigenvectors, with each uniquely labelled by energy E and spectrum
We parameterize by unit-lattice coordinates
expressed as a disjoint union of two three-dimensional unit lattices
and one two-dimensional unit lattice
Spectral labels are expressed as and . For a valid ,
and the energy is
The cardinalities are the same as equation (27). In this subsection, we have obtained the spectrum of the χr QW Hamiltonian. Furthermore we divided the set of stationary states into three cases for interpretation in section 5.
4.3. Symmetry and spectrum analysis for
In this subsection we study the symmetries of the enhanced chimera graph χ+. For , intracell translation permutation symmetry does not exist. The following demonstration is with L = 4 as example.
Figure 4 shows the intracell symmetries of . Π1,3 is an intracell left-side (right-side) permutation symmetry that permutes intracell vertices 1 and 2 (5 and 6) and simultaneously permutes intracell vertices 3 and 4 (7 and 8). Π2,4 is an intracell left-side (right-side) permutation symmetry that permutes intracell vertices 1 and 3 (5 and 7) and simultaneously permutes intracell vertices 2 and 4 (6 and 8). Intercell symmetries are the same as for , which are . The symmetry operators are
with matrices Si obtained from equation (13).
Download figure:
Standard image High-resolution imageStationary states are eigenvectors, with each uniquely labeled by energy E and spectrum
We parameterize by unit-lattice coordinates
expressed as a disjoint union of two four-dimensional unit lattices
and one two-dimensional unit lattice
Spectral labels are expressed as and . For a valid ,
and the energy is
The cardinalities are the same as equation (27). In this subsection, we obtained the spectrum of the χ+ QW Hamiltonian. Furthermore we divided the set of stationary states into three cases for interpretation in section 5. For other L values, one could consider symmetry operators such as
but we do not study these because the method to lift the degeneracy is similar.
4.4. Spectrum analysis for χ−
In this subsection we study the spectrum for the diminished chimera graph χ−. Symmetry analysis for χ− is difficult as most symmetry operators fail to commute with the Hamiltonian corresponding to a graph with broken vertices. But we can compare χ− with χr which we already study above.
First, we consider the case of just one broken vertex as the simplest example of broken vertices in the graph, and we compare the spectrum for for the unbroken reflecting graph versus the reflecting graph with one broken vertex, namely . The two spectra are compared in figure 5 with the entire spectrum shown as a zoom-in plot online and as a full-spectrum plot with partial spectrum plot as an inset. By zooming in, we see that the unbroken reflecting graph has a doubly degenerate eigenvalue E = 1.4635, which reduces to nondegeneracy arising from the broken vertex. This nondegeneracy is manifested as one eigenvalue shifting to a slightly lower value E ≈ 1.4620 and the other eigenvalue moving to the slightly higher E ≈ 1.4640. This broken vertex only noticeably changes the degeneracy of this eigenvalue and leaves most other eigenvalue degeneracies intact, thereby showing how only a few stationary states are affected by breaking one vertex.
Download figure:
Standard image High-resolution image Figure data fileThe reason why a few eigenvalues have shifted can be understood in the context of Anderson localization [51], which we can see with the help of figure 6. We consider one of eigenstates in the two-dimensional eigenspace corresponding to the eigenvalue E = 1.4635, and this eigenstate is depicted in figure 6(a). We observe approximately two cycles at one frequency along the m axis and a beat between two frequencies along the n axis. Then we isolate, or break, the (1, 1, 1) vertex and show in figure 6(b) the eigenstate for the lower eigenvalue. Figures 6(a), (b) look similar but differ markedly at the broken vertex. To elucidate this difference, we plot the arithmetic difference of the two eigenstates in figure 6(c), which shows clearly a spike at the (1, 1, 1) vertex. This spike shows that this peak is missing from the eigenstate for the broken-vertex case.
Download figure:
Standard image High-resolution image Figure data fileIn figure 6(d), we depict the localized state with eigenvalue E = 0. This depiction shows the the effect of breaking a vertex, which is to lead to a completely localized eigenstate at (1, 1, 1). This shifting of an eigenvalue and corresponding revision of eigenstates to transition from no support to total support at the defect (broken vertex) is a manifestation of Anderson localization. Thus, the broken-vertex χ− graph leads to quantum walks experiencing defects and thus Anderson localization.
5. Results and discussion
When we lift the degeneracy of the QW , we analyse salient properties of the eigenstates according to the spectra. First we consider eigenstates belonging to the unit lattice , denoted by , where and according to equation (19). Through equations (37) and (38), we have and s3 = s4 = 1. As s3 and s4 correspond to intracell right-side translation and mirror permutation symmetry, if we consider , we conclude that
for any and on the right side of the same unit cell. Notice that s1 corresponds to intracell left-side translation symmetry. If , we must have
for any and on the right side of the same unit cell. Through the above analysis, we see that has nonzero component only on the left side of the unit cells as shown in figure 7(a). Similarly, has nonzero component only on the right side of the unit cells as shown in figure 7(b) and has nonzero component on both side of the cells as shown in figure 7(c). Eigenstates are oscillatory in both horizontal and vertical directions and are effective lattice 'modes'.
Download figure:
Standard image High-resolution image Figure data fileWe have a comprehensive description of the quantum-walk 'modes' and thereby analyse the walker's evolution on the graph, shown in figure 2(e) for χp. The mode description helps explain features such as the walker's localization in figure 2. The walker's state
is supported by subspaces yielding orthogonal states , respectively. Initialization at in the left side of a cell yields
Similarly, the walker starting on the right side of the cell yields
Thus, a walker starting on the left or on the right has vertex distribution supported by at least the remaining distribution is spread over all vertices and thus is a small nonzero 'floor' for the distribution over all vertices.
In figure 2(e) we see that a walker starting at a vertex in the left side of a cell is not only confined to the left side but also stays in just one column with high probability. This transport is also clear from the two-dimensional Fourier transform in figure 8. Localization is due to the flat momentum spectrum along the axis, and the k spectrum is consistent with the momentum distribution for an ordinary QW on a circle [26]. The walker is initialized on the left of thus has broadband support over all states drawn from resulting in the observed localization to one column of . This localized walk exhibits some (indiscernible) leakage outside the column due to support over as seen in figure 2(e).
Download figure:
Standard image High-resolution imageThe case of the walker localized at a vertex on the right is similar to the case of the walker localized to a cell on the left in that all states parametrized in are localized to the right side of the initial column cell instead of to the left side of the initial column cell. A walker commencing at a vertex on the right side has support over with weight and support over with weight . One contrast between localization on the right versus left is evident by comparing figures 2(a), (b). If all the inter-cell coupling is increased equally, then the QW only changes by having the walker move more quickly and all the vertical and horizontal localization effects do not change. That only the speed, and not the features, changes is easily understood by recognizing that rescaling the coupling strengths does not change the graph symmetries.
The QW can be understood from spectral and stationary-state properties; now we use this knowledge to examine χ+. The walker's evolutions on χ+ and χr are similar at the cellular scale, which is evident by comparing the limiting probability distributions for in and in figures 2(d), (f), respectively. This similarity between reflecting and enhanced graphs is readily understood from equations (40) and (49) where we see that and share the same complete set of commuting observables except that energy differs. In fact, has many triply degenerate eigenvalues, and adding intracell edges to obtain the enhanced graph causes these triple degeneracies to split into single and double degeneracies that are still concentrated in the same cell. Thus, the effect of enhancing χ by adding intracell edges only spreads the walker within the cell but does not change the QW behaviour on the cellular scale.
Our approach could help with quantum annealer design by making graph symmetries, and consequences on localization and rate of spreading, quite clear and explicit. For example, according to equations (54) and (55), increasing L causes the walker to become more localized. Through symmetry analysis we can decide whether to increase or decrease the connections of χ in order to lift the degeneracy of the QW but without significantly changing the eigenstates of the QW .
Although the eigenstate so far are used to analyse properties of the QW, they can also be prepared. Preparation of these intricate eigenstates can be achieved by adiabatic evolution [43] beginning with an easy-to-prepare state and then varying h, j and k in equation (1) as time-dependent labels. For distinct the initial Hamiltonian is , and the final Hamiltonian at time T is
for y a transcendental number and z = 2 as an example. Each initial evolves to a different mode adiabatically so each mode can be prepared.
6. Conclusion
We have studied the QW on the chimera χ graph and its variants, discovered intricate features and showed how a model, or eigenstate, analysis explains the QW behaviour. The features of QW localization are explained by spectral analysis, which builds on completely characterizing the graph symmetries. We show how these symmetry operators can be incorporated into a target Hamiltonian for adiabatically creating modal states from initial states localized at vertices.
Our analysis of the χ graph via QWs could possibly aid graph design for quantum annealing as the χ graph is used for designing the D-Wave chip. We caution, though, that our analysis is based on a single-particle QW, and quantum annealing is based on exploiting a many-body ground state [52]. Our method thus helps to explore the symmetries of the quantum Hamiltonian with a graph structure that is founded on χ or one of its variants but not to ascertain the many-body ground state or the hardness therein.
Our enhanced χ shows that the walker spreads significantly only within the cell and not outside, and our diminished χ shows a high probability for a low-error graph to significantly diminish QW dynamics. Furthermore we have introduced powerful techniques that are useful for studying QWs on other graphs with intrinsic symmetries. Our approach provides a simple and general way to analyse the quantum walk on complicated graphs, which have applications to quantum transport and quantum algorithms, and to designing quantum annealers, which is one of the most significant directions in practical quantum computing. An experimental implementation of a two-dimensional quantum walk on the chimera graph could be envisaged for a Bose–Einstein condensate in an optical lattice [53] or with photon interference [23]. Vacuum noise and thermal effects clearly influence the walker's behaviour and need further investigation. Our focus has been on the closed system as the closed-system analysis is intricate and reveals much about the role of symmetries and localization in quantum-walk behaviour on the chimera and related graphs.
Acknowledgments
We thank S K Goyal for valuable discussions and acknowledge China's 1000 Talent Plan and NSFC (Grant No. 11675164) for support. NA acknowledges support from CAS PIFI program (Grant No. 2016PT003) and BCS appreciates support from Alberta Innovates.