Search on a Hypercubic Lattice using a Quantum Random Walk: I.

d > 2
Apoorva Patel1, 2, ∗ and Md. Aminoor Rahaman2, †
1
Centre for High Energy Physics, Indian Institute of Science, Bangalore-560012, India
2
Supercomputer Education and Research Centre, Indian Institute of Science, Bangalore-560012, India
(Dated: October 5, 2010)
Random walks describe diffusion processes, where movement at every time step is restricted to only
the neighbouring locations. We construct a quantum random walk algorithm, based on discretisation
of the Dirac evolution operator inspired by staggered lattice fermions. We use it to investigate the
spatial search problem, i.e. find a marked vertex on a d-dimensional hypercubic lattice. The
restriction on movement hardly matters for d > 2, and scaling behaviour close to Grover’s optimal
algorithm (which has no restriction on movement) can be achieved. Using numerical simulations, we
optimise the proportionality constants of the scaling behaviour, and demonstrate the approach to
arXiv:1003.0065v2 [quant-ph] 4 Oct 2010

that for Grover’s algorithm (equivalent to the mean field theory or the d → ∞ limit). In particular,
the scaling behaviour for d = 3 is only about 25% higher than the optimal d → ∞ value.

PACS numbers: 03.67.Ac, 03.67.Lx

I. SPATIAL SEARCH USING THE the marked vertex [8]. The optimization criterion for the
DIRAC OPERATOR algorithm is to concentrate the amplitude distribution
toward the marked vertex as quickly as possible. Grover
The spatial search problem is to find a marked object constructed the optimal global diffusion operator, but it
from an unsorted database of size N spread over dis- requires movement from any vertex to any other vertex
tinct locations. Its characteristic is that one can proceed in just one step. That corresponds to a fully connected
from any location to only its neighbours, while inspect- graph, or equivalently the mean field theory limit. When
ing the objects. The problem is conventionally repre- movements are restricted to be local (i.e. one can go from
sented by a graph, with the vertices denoting the loca- a vertex to only its neighbours in one step), the search
tions of objects and the edges labeling the connectivity algorithm slows down. It is then worthwhile to redesign
of neighbours. Classical algorithms for this problem are the search algorithm by understanding the extent of slow
O(N ), since they can do no better than inspect one lo- down due to local movements, and how it depends on the
cation after another until reaching the marked object. spatial arrangement of the database. That is the question
On the other hand, quantum algorithms can do better we address quantitatively in this article. As discussed in
in search problems by working with a superposition of what follows, for the best spatial search algorithms, the
states, Grover’s algorithm being the prime example [1]. slow down is in the scaling prefactor for d > 2 and in the
The spatial search problem has been a focus of inves- scaling form for d ≤ 2.
tigation in recent years using different quantum algo- A diffusion process can be modeled either using con-
rithmic techniques, both analytical and numerical, and tinuous time and a first order time derivative, or using
in a variety of spatial geometries ranging from a single discrete time and a single time step evolution. In its
hypercube to regular lattices [2–7]. These studies have discrete version, it is generically described as a random
mainly looked at the asymptotic scaling forms of the al- walk, whereby a probability distribution evolves in a non-
gorithms, and have not varied the database size N and deterministic manner at every time step. Such random
its dimension d independently. In this work, we study walks have been used to tackle a wide range of graph
the specific case of searching for a marked vertex on a theory problems, usually with a local and translation-
d-dimensional hypercubic lattice with N = Ld vertices. ally symmetric evolution rule. We use a quantum ver-
We let N and d be independent, as well as determine the sion of this process, i.e. quantum random walks [9]. It
scaling prefactors, and thereby develop a broad picture of provides a unitary evolution of the quantum amplitude
how the dimension of the database (or the connectivity distribution, such that the amplitude at each vertex gets
of the graph) influences the spatial search problem. redistributed over itself and its neighbours at every time
step. It is worth noting that quantum random walks are
The quantum algorithmic strategy for spatial search is
deterministic, unlike classical random walks, with quan-
to construct a Hamiltonian evolution, where the kinetic
tum superposition allowing simultaneous exploration of
part of the Hamiltonian diffuses the amplitude distribu-
multiple possibilities. Several quantum algorithms have
tion all over the lattice and the potential part of the
used them as important ingredients, and an introductory
Hamiltonian attracts the amplitude distribution toward
overview can be found in Ref.[10].
On a periodic lattice with translational symmetry, spa-
tial propagation modes are characterized by their wave
∗ Electronic address: [email protected] vectors ~k. Quantum diffusion depends on the energy of
† Electronic address: [email protected] these modes according to U (~k, t) = exp(−iE(~k)t). The
 2

lowest energy mode, ~k = 0, corresponding to a uni- arbitrarily separated vertices. To make U also local, we
form distribution, is an eigenstate of the diffusion op- split H in to block-diagonal Hermitian parts and then
erator and does not propagate. The slowest propagating exponentiate each part separately [13]. For a bipartite
modes are the ones with the smallest nonzero |~k|. The lattice, partitioning of H in to two parts (which we label
commonly used diffusion operator is the Laplacian, with “odd” and “even”) is sufficient for this purpose [14]:
E(~k) ∝ |~k|2 . The alternative Dirac operator, available Hfree = Ho + He . (6)
in a quantum setting, provides a faster diffusion of the
slowest modes with E(~k) ∝ |~k|. Its quadratically faster This partition is illustrated in Fig.1 for d = 1 and d = 2.
spread makes it better suited to quantum spatial search Each part contains all the vertices, but only half of the
algorithms [3, 4]. links attached to each vertex. Each link is associated
An automatic consequence of the Dirac operator is the with a term in Hfree providing propagation along it, and
appearance of an internal degree of freedom correspond- appears in only one of the two parts. H thus gets divided
ing to spin, whereby the quantum state becomes a multi- in to a set of non-overlapping blocks of size 2d × 2d (each
component spinor. These spinor components can guide block corresponds to an elementary hypercube on the lat-
the diffusion process, and be interpreted as the states of tice) that can be exactly exponentiated. The amplitude
a coin [3, 4]. While this is the only possibility for the distribution then evolves according to
continuum theory, another option is available for a lat- ψ(~x; t) = W t ψ(~x; 0) , W = Ue Uo = e−iHe τ e−iHo τ . (7)
tice theory, i.e. the staggered fermion formalism [11]. In
this approach, the spinor degrees of freedom are spread in Each block of the unitary matrices Uo(e) mixes ampli-
coordinate space over an elementary hypercube, instead tudes of vertices belonging to a single elementary hy-
of being in an internal space. We follow it to construct percube, and the amplitude distribution spreads in time
a quantum search algorithm on a hypercubic lattice, re- because the two alternating matrices do not commute.
ducing the total Hilbert space dimension by 2d and elim- Note that W does not perform evolution according to
inating the coin toss instruction. Hfree exactly. Instead, W = exp(−iHfree τ ) + O(τ 2 ).
Still, the truncation is such that W is exactly unitary,
i.e. W = exp(−iH̃τ ) for some H̃.
A. Bipartite Decomposition We adopt the convention where the walk operator W
is real. For d = 1, we set
The free particle Dirac Hamiltonian in d dimensions is ih i
Ho |xi = − (−1)x |x + (−1)x i , (8)
2
Hfree = −i~ ~ + βm .
α·∇ (1) i h i
He |xi = (−1)x |x − (−1)x i . (9)
On a hypercubic lattice, with the simplest discretization 2
of the gradient operator, The resultant Hamiltonian blocks are HoB = −σ2 /2, and
HeB = σ2 /2 when operating on the block with coordi-
1 nates flipped in sign. With (HoB )2 = (HeB )2 = 41 I,
∇n f (~x) = [f (~x + n̂) − f (~x − n̂)] , (2)
2
Uo(e) = cI − 2isHo(e) , |c|2 + |s|2 = 1 , (10)
and a convenient choice of basis, the anticommuting ma-
trices α
~ , β are spin-diagonalized to location dependent
signs: −2 −1 0 1 2
ψ → T ψ, H → T HT , †
T =
xd−1
αxd d αd−1 · · · αx2 2 αx1 1 ,(3) (a) o e o e
n−1
Y d
Y
αn → (−1)xj , β→ (−1)xj β . (4)
j=1 j=1 e
e
With this choice, translational invariance holds in steps
11111
00000 01 010,1 011,1
11111
00000
of 2 (instead of 1), and the squared Hamiltonian,
00000
11111
o 00000
11111
00000
11111 00000
11111 o
2
Hfree = −∇ ~ ·∇
~ + m2 , (5)
(b) 00000
11111
−1,001 00000
11111
010,0 011,0
2
has eigenvalues n sin (kn )+m2 on a hypercubic lattice.
P e e
Inclusion of a mass term slows down the diffusion process, −1,−1
11111
00000 01 010,−1 01
11111
00000
but it also regulates the ~k = 0 mode. In the present 00000
11111 00000
11111
article, we set m = 0; in an accompanying article [12], 00000
11111
o
00000
11111 o
we investigate the advantage a nonzero mass offers when 00000
11111 00000
11111
the problem has an infrared divergence.
Even when the Hamiltonian H is local, the discrete FIG. 1: Bipartite decomposition of the hypercubic lattice in
time evolution operator U = exp(−iHτ ) may connect to odd and even parts: (a) for d = 1, (b) for d = 2.
 3

where s = sin(τ /2) is a parameter that can be tuned. B. Search Algorithm

These expressions can be rearranged to separate the
translationally invariant part of the walk from the loca- To search for a marked vertex, say the origin, we at-
tion dependent signs. Let the amplitude distribution in tract the quantum random walk toward it by adding a
a two spinor (or coin) component notation be potential to the free Hamiltonian,

βV0 δ~x,0 (gravitational) ,
 
ψ(2x, t) V = (17)
Ψ(X, t) ≡ . (11) V0 δ~x,0 (electromagnetic) .
ψ(2x − 1, t)

Then the walk operator becomes Exponentiation of this potential produces a phase change
for the amplitude at the marked vertex. For the steep-

c s
 est attraction of the quantum random walk toward the
W = Ũ C , C = UoB = , (12) marked vertex, it is optimal to choose V0 so as to make
−s c
X the corresponding evolution phase maximally different
Ũ |Xi = c|Xi − s (±σ± )|X ± 1i . (13) from 1, i.e. e−iV0 τ = −1. The sign provided by β does
± not matter in this case, and the evolution phase becomes
the binary oracle,
The operator C = UoB just mixes the components of Ψ
within the block. The first term of the operator Ũ is sta- R = I − 2|~0ih~0| . (18)
tionary, while the other moves the amplitude to neigh- The search algorithm alternates between diffusion and
bouring blocks. The movement is accompanied by chi- oracle operators—a discrete version of Trotter’s formula
rality flip, a characteristic feature of the Dirac opera- involving Ho , He and V —yielding the evolution
tor, because of the raising and lowering Pauli operators
σ± = (σ1 ± iσ2 )/2. ψ(~x; t1 , t2 ) = [W t1 R]t2 ψ(~x; 0, 0) . (19)
In d dimensions, the preceding structure generalises
to 2d × 2d Hamiltonian blocks that can be expressed in Here t2 is the number of search oracle queries, and t1 is
terms of tensor products of Pauli matrices [15]. When the number of walk steps between queries. Both should
operating on hypercubes with coordinate labels {0, 1}⊗d, be minimised to find the the quickest solution to the
search problem.
d Since this algorithm iterates a unitary operator on the
1 X ⊗(d−j) ⊗(j−1) initial state, it produces periodic results. As a function of
HoB = − I ⊗ σ2 ⊗ σ3 , (14)
2 j=1 the number of iterations, the probability of being at the
marked vertex first increases, reaches a peak value P after
and HeB = −HoB when operating on a hypercube with t2 search oracle queries, then decreases toward zero, and
all coordinates flipped in sign. Note that Eq.(14) is the thereafter keeps on repeating this cycle. Grover’s algo-
sum of d terms, which mutually anticommute and whose rithm is designed to evolve the quantum state in a two-
squares are proportional to identity. This is the charac- dimensional Hilbert space (spanned by the initial and
teristic signature of the Dirac Hamiltonian, Eq.(1). The marked states), and is able to reach P = 1. That does
block-diagonal matrices satisfy Ho2 = He2 = d4 I, and ex- not hold for spatial search, and the resultant values of P
ponentiate to are less than 1. The remedy is to augment the algorithm
by the amplitude amplification procedure [16], to find
2 the marked vertex with Θ(1) probability. The complex-
Uo(e) = cI − is √ Ho(e) , |c|2 + |s|2 = 1 . (15) ity of the algorithm is then characterised by the effective
d √
number of search oracle queries, t2 / P .
√
The parameter s = sin( dτ /2) can be optimised, to
achieve the fastest diffusion across the lattice. Note that
C. Complexity Bounds
the d-dimensional walk is not the tensor production of d
one-dimensional walks.
Again separating the degrees of freedom within each Spatial search obeys two simple lower bounds. One
~
local hypercube as ~x ≡ X⊗{0,-1} ⊗d arises from the fact that while the marked vertex could
, the location depen-
be anywhere on the lattice, one step of the local walk
dent signs can be rewritten as operators acting on the
can move from any vertex to only its nearest neighbours.
spinor (or coin) degrees of freedom. The walk operator
Since the marked vertex cannot be located without reach-
takes the form: W = Ũ C, C = UoB , and
ing it, the worst case scenario on a hypercubic lattice re-
~ = c|Xi
~ quires Ω(dL) steps. This bound is the strongest in one
Ũ |Xi (16) dimension, where quantum search is unable to improve
d
s X X ⊗(j−1) ⊗(d−j) ~
on the O(N ) classical search.
− √ (I ⊗ (±σ± ) ⊗ σ3 )|X ± ĵi . The other bound follows from the fact that spatial
d ± j=1
search cannot outperform Grover’s optimal algorithm,
 4
√
and so it must require Ω( N ) oracle queries. That is II. OPTIMISATION OF PARAMETERS
independent of d, and stronger than the first bound for
d > 2. As a matter of fact, the first bound weakens with The fastest search amounts to finding the shortest uni-
increasing d, as the number of neighbours of a vertex in- tary evolution path between the initial state |si and the
crease and steps required to go from one corner of the marked state |~0i. This path is a geodesic arc on the uni-
lattice to another decrease. In analogy with mean field
tary sphere from |si to |~0i, and Grover’s optimal algo-
theory behaviour of problems in statistical mechanics,
rithm strides along this path perfectly. In our algorithm,
we can thus expect that in the d → ∞ limit the local-
Eq.(19), we have replaced Grover’s optimal diffusion op-
ity restriction on the walk would become irrelevant and
erator, G = 2|sihs| − 1, with the walk operator W t1 . So
the complexity of spatial search would approach that of
to optimise our algorithm, we need to tune the param-
Grover’s algorithm. (Note that the maximum value of d
eters appearing in W t1 , i.e. s (or c) and t1 , such that
is log2 N for finite N .) Our results
√ in this work demon- W t1 R approximates a rotation in the two-dimensional
strate that not only is the Ω( N ) complexity scaling |si-|~0i subspace by the largest possible angle.
achievable for d > 2, but one can get pretty close to the The operator G has |si as an eigenstate with eigen-
asymptotic prefactor π/4 as well. value 1. All other states orthogonal to√|si, in particular
√
Combined together, the two √ bounds make the complex- the combination |s⊥ i = (|si− N |~0i)/ N − 1 in the |si-
ity of spatial search Ω(dN 1/d ,
√ N ). The two bounds are |~0i subspace, are its eigenstates with eigenvalue −1. The
of the same magnitude, Ω( N ), for the critical case of walk operator W = Ue Uo also has |si as an eigenstate
d = 2. It is a familiar occurrence in statistical mechan- with eigenvalue 1. However, its other eigenvalues are
ics of critical phenomena that an interplay of different spread around the unit circle, and evolution under W t1 R
dynamical features produces logarithmic correction fac- does not remain fully confined to the |si-|~0i subspace. If
tors in critical dimensions due to infrared divergences. the outward diffusion is not controlled, the probability of
The same seems to be true for spatial search in d = 2, remaining in the |si-|~0i subspace would decay exponen-
where algorithms are slowed down by extra logarithmic tially with the number of iterations. To reach the marked
factors [3–5]. It is an open question to design algorithms, state with a constant probability, therefore, parameters
perhaps using additional parameters, that suppress the must be tuned so that part of the outward diffusion sub-
logarithmic factors as much as possible. We look at the sequently returns to the |si-|~0i subspace.
special d = 2 case in an accompanying article.
The way the lower bounds arise in spatial search also
illustrates an interplay of two distinct physical princi- A. Analytical Criteria
ples, special relativity (or no faster than light signaling)
and unitarity. It is well-known that the two are compati- The optimisation condition that makes W t1 as close
ble, but just barely so, in the sense that physical theories an approximation to G as possible can be formulated in
that are unitary but not relativistic or relativistic but not several different ways. Though they all lead to the same
unitary exist. Furthermore, the best algorithms should result, describing them separately is instructive.
arise in a framework that respects both the principles, (1) Maximise the overlap of W t1 and G, i.e.
i.e. relativistic quantum mechanics. In spatial search, X
locality of the quantum walk and the square-root speed T r(W t1 G) = T r(2|sihs|−W t1 ) = 2− hi|W t1 |ii . (20)
up produced by the Dirac equation (through change in i
the dispersion relation) follow from special relativity. On
the other hand, optimality of square-root speed up pro- hi|W t1 |ii is the amplitude for the random walk to return
vided by Grover’s algorithm is a consequence of unitarity to the starting state |ii after t1 time steps. Since W is
[17, 18]. Why these two distinct physical principles lead translationally invariant along each coordinate direction
to the same scaling constraint is not understood, and is in steps of 2, we need to evaluate hi|W t1 |ii only for states
a really interesting question to explore. At present, how- |ii in an elementary hypercube. Now, any closed path
ever, we have nothing more to add. on a hypercubic lattice takes an even number of steps
along each coordinate direction, and all location depen-
We point out that preparation of the unbiased initial dent propagation signs (cf. Eq.(4)) get squared to 1 in
state for the search problem, i.e. the translationally
P √ in- the process. As a result, hi|W t1 |ii is independent of |ii,
variant uniform superposition state |si = x |~xi/ N , is and the optimisation condition becomes the minimisation
not at all difficult using local directed walk steps. For of A(t1 ) ≡ h~0|W t1 |~0i.
instance, one can start at the origin, step by step trans- (2) Make |s⊥ i approximate an eigenvector of W t1 , with
fer the amplitude to the next vertex along an axis, and an eigenvalue approaching −1, i.e. minimise
achieve an amplitude L−1/2 at all the vertices on the axis
after L steps. Repeating the procedure for each remain- 1
hs⊥ |W t1 |s⊥ i =


ing coordinate direction produces the state |si after dL − 1 + N A(t1 ) , (21)
N −1
steps in total. Clearly, this preparation does not add to
the complexity of the algorithm. which happens to be the same condition as given earlier.
 5
qP
In addition, unitarity of the evolution operator implies exp(±iτ 2
j sin kj ), with the eigenphases in the inter-
√ √ 2
2 val [−τ d, τ d]. The average value of Hfree is d/2,
p and
|hs⊥ |W t1 |s⊥ i| ≤ 1 =⇒ − 1 + ≤ A(t1 ) ≤ 1 . (22)
N the corresponding eigenvalues of W are exp(±iτ d/2).
We find that this average value plays an important role
(3) Maximise the rotation provided by the operator
in determination of the optimal parameters.
W t1 R along the |si-|~0i subspace. This rotation is given
by the projection

hs|W t1 R|si hs|W t1 R|s⊥ i

 
B. Numerical Optimisation
Projs,s⊥ [W t1 R] =
hs⊥ |W t1 R|si hs⊥ |W t1 R|s⊥ i
√ The considerations of the previous subsection focused
1 − N2 2
!
N N −1
= √2 (23) on what happens during a single iteration of the algo-
N −1
(A(t1 ) − N1 ) (1 − N2 ) 1−N A(t1 )
N −1 rithm. Over many iterations, there is outward diffusion
√

1 0

1 − N2 2
N −1
 of the amplitude distribution from the |si-|~0i subspace,
= √ N . and subsequent partial return. That plays an important
0 1−N A(t1 )
N −1 − N2 N − 1 1 − N2
role in the overall success of the algorithm, but we have
The factor on the right in Eq.(23) is precisely the ro- not been able to estimate it analytically. To study the
tation matrix produced by the operator GR √ in Grover’s performance of the whole algorithm, therefore, we nu-
algorithm, i.e. rotation by an angle 2 sin−1 (1/ N ) in the merically tune the parameters s and t1 to (a) maximise
|si-|~0i subspace. So the total projection can be viewed the probability of reaching the marked vertex, and (b)
as first applying an iteration of Grover’s algorithm to the minimise the number of search oracle queries required to
state, and then shrinking the |s⊥ i component by a fac- reach that stage. The ensuing results of our optimisation
tor (1 − N A(t1))/(N − 1). The shrinking disappears and of P , t2 and A(t1 ) are described in what follows.
the projection tends to an exact orthogonal transforma- We first looked at how the probability distribution
tion, as A(t1 ) approaches its lower bound (2 − N )/N . It evolves as a function of time during the search √ process.
also follows that when the shrinking is controlled, cutting To illustrate our results, we take s = c = 1/ 2 that
down outward diffusion from the ~ corresponds to maximal mixing between even and odd
√ |si-|0i subspace, the al- vertices, and analyse the evolution as a function of t1
gorithm will have the same O( N ) scaling as in case of
Grover’s algorithm. and t2 on a 323 lattice. As displayed in Fig.2, the proba-
The minimisation condition for A(t1 ) is obtained here bility at the marked vertex undergoes a cyclic evolution
only using the translational invariance of the walk op- pattern as a function of time t2 . We observe that the
erator W , together with W |si = |si and hs|W = hs|. evolution is essentially sinusoidal for t1 = 1, 2, 3, and
It is easy to perform this optimisation numerically, as persists for more than 30 cycles without any visible devi-
a function of the parameters s and t1 , and our results ation. This behaviour implies that a constant fraction of
are described later. Before that, an understanding of the the quantum state remains in the |si-|~0i subspace, and
eigenspectrum of W gives us some idea regarding what the operator W t1 R rotates it at a uniform rate within the
to expect. subspace, fully matching the exact solution of Grover’s
Since Ho2 = He2 = d4 I, eigenvalues of Ho and He algorithm. The difference among t1 = 1, 2, 3 is that the
√ cycle time and the peak probability vary, meaning that
are ± d/2. Moreover, Ho and √ He are traceless, so the fraction of the quantum state present in the |si-|~0i
that
√ half the eigenvalues are d/2 and the other half subspace depends on t1 .
− d/2. Consequently,
√ eigenvalues of Uo and Ue are For t1 > 3, we find that the evolution rapidly loses
c ± is = exp(±i dτ /2), equally divided among complex its sinusoidal nature and the peak probability plummets.
conjugate pairs. Furthermore, Ho(e) and Uo(e) have the The change in the pattern is so drastic that the algorithm
same eigenvectors. no longer performs a reasonable search. This behaviour
Since Uo and Ue do not commute, we do not have a implies that most of the quantum state has moved out
general solution for the eigenspectrum of W . Still it is of the |si-|~0i subspace and only a negligible component
straightforward to work it out in the continuum time is left behind. An appropriate choice of t1 is thus crucial
limit, τ → 0. Hfree is diagonalised by a Fourier transform, for constructing a successful search algorithm.
d We also show in Fig.3 what the probability distribu-
2 1 Xh i
tion looks like at the instance when the probability at
Hfree |~xi = − |~x + 2n̂i − 2|~xi + |~x − 2n̂i
4 n=1 the marked vertex attains its peak value. We note that
d
the distribution is strongly peaked around the marked
−→
X
sin2 (kn )|~ki , (24) vertex in an essentially uniform background. The peak
n=1
is sharper when the peak probability is larger, and it al-
most disappears for t1 > 3. This pattern suggests that
with a uniform distribution in the Brillouin zone. Up to the part of the quantum state that diffuses outside the
O(τ 2 ), therefore, the eigenvalues of W (≈ e−iHfree τ ) are |si-|~0i subspace is almost featureless.
 6

Combining our observations in this particular example,

we find that the best results—the shortest period, the
largest peak probability and the sharpest peak—clearly
correspond to t1 = 3. The results improve as t1 increases
from 1 to 3, and rapidly deteriorate thereafter. The im- 0.1

Probability
t =1
portant fact that all optimal features appear simultane- 1

ously at t1 = 3 simplifies tuning of the parameters. 0.05

Next we scanned through values of s, for different val-
ues of t1 and in different dimensions, to determine the 0
optimal values at which the peak probability P at the 30
marked vertex is the largest and the time t2 required to 20 30
reach it is the smallest. A typical situation, with t1 = 3 20
10 10
on a 323 lattice, is shown in Fig.4. We notice that con-
0 0
tinuous P has a smoother behaviour than discrete t2 ,
and maximisation of P and minimisation of t2 occur at
roughly the same value of s. So we simplify matters, and
select maximisation of P as our optimality condition.
0.1

Probability
t =2
1

0.05
Probability

t1 = 1
0.1 0
30
0.05 20 30
20
0 10 10
0 500 1000 1500 0 0
Time
Probability

t1 = 2
0.1 0.1
Probability

t =3
1

0.05 0.05
0
0 500 1000 1500 0
30
Time 30
20
20
10 10
Probability

t1 = 3
0.1 0 0

0.05
0
0.1
Probability

0 500 1000 1500 t =4

1

Time 0.05
Probability

t1 = 4 0
0.1 30
20 30
0.05 20
10 10
0 0 0
0 500 1000 1500
Time
FIG. 3: (Color online) Probability distribution for the spatial
search problem on a 323 lattice, at the instance when the
FIG. 2: (Color online) Time evolution of the probability at the probability at the marked vertex (16, 16, 16) attains its peak
√
marked√vertex for the spatial search problem on a 323 lattice. value. The displayed results are for the z = 16 slice. s = 1/ 2
s = 1/ 2 and t1 increases from 1 (top) to 4 (bottom). and t1 increases from 1 (top) to 4 (bottom).
 7

A comparison of our scans, for t1 = 3 and different

TABLE I: Results of optimal parameter determination for
number of dimensions, is presented in Fig.5. We find
different t1 and in different dimensions.
that at fixed t1 , the shape of the P vs. s curve is al-
most dimension independent, and the optimal value of s Spatial Search Walk
d L t1
is approximately the same for all dimensions. We also s t2 P θ s A(t1 )min θ
observe that for fixed s, P roughly scales as 2−d . In the 2 0.9507 59 0.0942 3.551 0.9258 −0.7143 3.346
staggered fermion implementation of the Dirac operator, 3 0.7015 55 0.1001 3.299 0.6737 −0.7618 3.136
different vertices of an elementary hypercube correspond 3 32
4 0.5363 55 0.1016 3.202 0.5194 −0.7748 3.089
to different degrees of freedom [11]. So our observation 8 0.2755 54 0.1027 3.158 0.2665 −0.7860 3.052
suggests that to a large extent only the degree of free- 20 0.1114 54 0.1030 3.157 0.1074 −0.7890 3.044
dom corresponding to the marked vertex evolves during
2 0.9541 54 0.0528 3.583 0.9428 −0.7778 3.482
the search, while other degrees of freedom remain spec-
tators. Moreover, the feature that the maximum value 3 0.6986 54 0.0548 3.281 0.6827 −0.8190 3.188
4 16
of P approaches 2−d with increasing d implies that the 4 0.5411 53 0.0553 3.234 0.5257 −0.8300 3.131
efficiency of our algorithm increases with d. 8 0.2771 52 0.0558 3.177 0.2694 −0.8395 3.086
All our quantitative results for optimal values of s are 20 0.1115 52 0.0559 3.160 0.1084 −0.8420 3.072
collected in Table I. We note that the best performance 2 0.9500 150 0.0276 3.545 0.9535 −0.8182 3.577
parameters, P and t2 , depend on d but hardly on t1 or 3 0.6920 148 0.0284 3.242 0.6880 −0.8541 3.219
s. (With increasing t1 , there is a slight increase in P and 5 16
4 0.5376 147 0.0286 3.211 0.5292 −0.8636 3.155
a small decrease in t2 , but that is a rather tiny improve-
8 0.2726 147 0.0288 3.124 0.2710 −0.8718 3.105
ment.) More interestingly, we find a relation between the
optimal values of s and t1 , in terms of the variable 20 0.1108 147 0.0288 3.140 0.1092 −0.8739 3.095
3 0.6891 51 0.0145 3.225 0.6913 −0.8778 3.238
√ p 6 8
θ = 2t1 sin−1 s = t1 τ d/2 . (25) 20 0.1106 51 0.0147 3.135 0.1094 −0.8951 3.101
3 0.6932 102 0.0073 3.250 0.6937 −0.8949 3.252
Since sin−1 s describes the unitary rotation performed 7 8
20 0.1097 102 0.0074 3.109 0.1098 −0.9102 3.112
by the single step operators Uo(e) , and t1 is the number
of walk steps, θ is a measure of the unitary rotation per-
formed by the operator W t1 . The corresponding operator
in Grover’s optimal algorithm, G, is a reflection providing
isation of θ implies that the optimal W t1 ≡ e−iH̃τ t1 is a
a phase change of π. We find that our optimal W t1 is also
reflection operator for the “average eigenvalue” of the as-
close to a reflection, in a sense, with θ ≈ π. Unlike the
sociated H̃ 2 . Appearance of the average eigenvalue in the
discrete ±1 eigenvalues of G though, eigenvalues of W t1
normalisation indicates that all spatial modes contribute
form a distribution. We do not know this distribution in
to the search process with roughly equal strength.
general, but we know it in the limit τ → 0, as described
at the end of Subsection II.A. This limit corresponds to Table I also lists our results for the variable A(t1 ) de-
small s and large t1 , and Table I shows that θ indeed fined in Subsection II.A, which characterises how close
gets close to π in this limit. To be precise, this normal-

1
110
d=3
100 d=4
1000 P 0.8
d=5
90 t
2 d=6
80 d=7
2d P

0.6
70
60
0.4
50
40
0.2
30 0.4 0.5 0.6 0.7 0.8
0.4 0.5 0.6 0.7 0.8 s
s
FIG. 5: (Color online) Optimisation of s for t1 = 3 in different
FIG. 4: (Color online) Optimisation of s for t1 = 3 on a 323 dimensions. Lattice sizes L = 32, 16, 16, 8, 8 were used for
lattice, using peak probability P as well as oracle queries t2 . d = 3, 4, 5, 6, 7, respectively.
 8
√
W t1 is to G. We computed A(t1 ), by starting with unit 1/ 2 for t1 = 3 and s = 0.5410 for t1 = 4.
amplitude at the marked vertex, and then measuring the To extract the scaling properties of the algorithm, we
amplitude at the marked vertex after t1 walk steps. At studied the behaviour of P and t2 , as a function of L and
the beginning of this subsection we mentioned that check- d. We found that the finite size effects in our data, as
ing A(t1 ) is not as elaborate an optimisation test as us- L → ∞ at fixed d, are well described by series in inverse
ing the full search algorithm. Nevertheless, the results powers of L. In Figs.6-7, we show some of our results
corroborate our optimisation of spatial search: The min- with the fitting functions:
imum value of A(t1 ) is reasonably close to −1, and occurs
at approximately the same value of s (and hence θ) that b1 t2 b2
P = a1 + , √ = a2 + . (26)
provides the optimal search (the mismatch decreases with L N L
increasing d). We also observe that A(t1 ) gets closer to
−1, suggesting that the algorithm becomes more efficient, All the fit parameters are listed in Table II, where error
with increasing d as well as t1 . refers to the r.m.s. deviation of the data from the fit. We
Given the strong correlation between the optimal val- do not have sufficient data for d = 8 and d = 9, to make
ues of s and t1 , the computational cost of the algorithm accurate fits or to see the asymptotic behaviour. Our
is minimised by choosing the smallest t1 . For t1 = 1, analysis, therefore, relies on the patterns seen for d = 3
Eq.(25) does not have a solution with θ ≈ π, the optimal to d = 7, while we use d = 8 and d = 9 values only for
value of s is very close to 1, and our algorithm does not consistency checks.
work well there. Then t1 = 2 is the choice that min- We observe the following features:
imises the number of walk steps in the algorithm. Even (1) There is not much difference among the a1 and a2
when the emphasis is on optimising P and t2 , most of values for different optimised s-t1 pairs. This confirms
the improvement can be obtained by going from t1 = 2 our earlier inference that implementing our algorithm for
to t1 = 3, and we need not go beyond that. t1 = 2 minimises the running cost.
(2) P approaches a constant as L → ∞, with a1 decreas-
ing approximately as 2−d with increasing d. As men-
III. SIMULATION RESULTS tioned before, this is a consequence of the fact that in
the staggered fermion implementation of the Dirac op-
erator, only the degree of freedom corresponding to the
We carried out numerical simulations of the quantum location of the marked vertex on the elementary hyper-
spatial search problem, with a single marked vertex and cube participates in the search process. Since the number
lattice dimension ranging from 3 to 9. The matrices Uo(e) of vertices in an elementary hypercube is 2d , P is upper
are real with our conventions, and that was convenient bounded by 2−d . Our algorithm achieves that bound up
for simulations. The optimal values of s-t1 pairs depend to a constant multiple close to 1.
a little on the lattice size and dimension. We are only (3) b1 decreases toward zero with increasing d, which
interested in the asymptotic behaviour of the algorithm, is also true for |b2 | to a large extent. Thus the small-
however, so we used the same parameter values for all est d = 3 has the largest finite size corrections to the
lattice sizes and dimensions: s = 0.9539 for t1 = 2, s =

0.35
0.12
0.3
0.1 d=3
d=3 d=4
0.25
d=4 d=5
0.08 d=5 d=6
t2 / N

d=6 0.2 d=7

√

d=8
0.06
P

0.15 d=9

0.04
0.1

0.02 0.05

0 0
5 10 15 20 25 5 10 15 20 25
log2 N log2 N

FIG. 6: (Color online) Peak probability at the marked vertex FIG. 7: (Color online) Number of search oracle queries as a
as a function of database size (ranging from 28 to 227 ) in differ- function of database size (ranging from 28 to 227 ) in different
ent dimensions. The points are the data from the simulations dimensions. The points are the data from √ the simulations
with t1 = 3, and the curves are the fits P = a1 + (b1 /L). with t1 = 3, and the curves are the fits (t2 / N ) = a2 +(b2 /L).
 9

√
TABLE II: Fit parameters for peak probability P and search oracle queries t2 / N vs. L.
√
s t1 d L a1 b1 Error a2 b2 Error a2 / a1
3 64,128,256,512 0.0911 0.0964 2.43×10−5 0.3237 0.1440 4.57×10−4 1.072
4 16,32,48,64 0.0522 0.0100 3.02×10−5 0.2167 −0.0832 1.05×10−3 0.948
5 12,16,24,32 0.0275 0.0009 1.70×10−6 0.1486 −0.0243 2.39×10−4 0.896
0.9539 2
6 12,14,16,18 0.0141 0.0001 2.31×10−7 0.1043 −0.0208 1.78×10−4 0.878
7 6,8,10 0.0072 −0.0004 1.40×10−6 0.0757 −0.0338 2.47×10−4 0.892
8 6,8 0.0036 — 3.50×10−6 0.0509 — 7.23×10−5 0.848
9 6 0.0018 — — 0.0356 — — 0.839
−6 −3
3 64,128,256,512 0.0968 0.0920 2.73×10 0.3141 −0.0306 1.23×10 1.010
4 16,32,48,64 0.0542 0.0076 1.46×10−5 0.2097 0.0151 6.71×10−4 0.901
√ 5 12,16,24,32 0.0283 0.0010 4.66×10−6 0.1470 −0.0300 2.12×10−4 0.874
1/ 2 3
6 12,14,16,18 0.0145 0.0001 5.79×10−7 0.1035 −0.0269 1.88×10−4 0.860
7 6,8,10 0.0074 −0.0003 1.46×10−6 0.0750 −0.0296 3.39×10−4 0.872
8 6,8 0.0037 — 3.00×10−6 0.0498 — 4.55×10−5 0.819
9 6 0.0019 — — 0.0353 — — 0.810
3 64,128,256,512 0.0984 0.0936 1.84×10−5 0.3123 −0.1239 8.46×10−4 0.996
4 16,32,48,64 0.0548 0.0087 2.24×10−5 0.2103 −0.0500 3.57×10−4 0.898
5 12,16,24,32 0.0285 0.0013 8.05×10−6 0.1455 −0.0142 2.25×10−4 0.862
0.5410 4
6 12,14,16,18 0.0146 0.0002 7.67×10−7 0.1015 0.0043 6.28×10−5 0.840
7 6,8,10 0.0074 −0.0003 2.86×10−6 0.0733 −0.0194 2.01×10−4 0.852
8 6,8 0.0037 — 4.50×10−6 0.0505 — 4.39×10−4 0.830
9 6 0.0019 — — 0.0353 — — 0.810

asymptotic L → ∞ behaviour. This suggests that as the This is in agreement with the fact that, with only one
number of neighbours of a vertex increase, the finite size degree of freedom per elementary hypercube participat-
of the lattice becomes less√and less relevant. ing in the search process, the effective number of vertices
(4) t2 is proportional to N as L → ∞, with a2 de- being searched is N/2d . √
creasing approximately as 2−d/2 with increasing pd. Thus (5) In the combination t2 / P , describing the effective
t2 scales with the database size as (L/2)d/2 = N/2d. number of oracle queries for our algorithm, the factors of
2d arising from the number of vertices √ in an elementary
√
hypercube fully cancel. That makes t2 / P scale as N ,
√
and a2 / a1 describes the complexity scaling of our al-
2 gorithm. As listed in the rightmost column of Table II,
√
1 a2 / a1 shows a gradual decrease with increasing d, and
0 log2 a1
our results are just above the corresponding value π/4 for
log2 a2 Grover’s optimal algorithm. More explicitly, our result
−2 a2 / a1
√ for d = 3 is about 25% above π/4, decreasing to about
10% above π/4 for d = 7. These values imply that our
−4 algorithm improves in efficiency with increasing d, and
is not too far from the optimal behaviour even for the
smallest dimension d = 3.
−6
To analyse the dimension dependence of our algorithm
in more detail, we fit the values in Table II to the func-
−8 tional forms,

−10 log2 a1 = c1 + d1 d , log2 a2 = c2 + d2 d , (27)

3 4 5 6 7 8 9
d
a2 d3
FIG. 8: (Color online) Behaviour of the asymptotic fit param- √ = c3 + . (28)
a1 d
eters as a function of dimension. The data points are from
Table II, with t1 = 3, and the curves are the fits described in The results for t1 = 3 are displayed in Fig.8, showing
Eqs.(27,28). that the fits work reasonably well.
 10

√
TABLE III: Fit parameters for log2 a1 , log2 a2 and a2 / a1 vs. d.
s t1 Fitted d c1 d1 Error c2 d2 Error c3 d3 Error
0.9539 2 −0.553 −0.938 4.77×10−2 −0.085 −0.526 2.53×10−2 0.721 0.995 1.71×10−2
√
1/ 2 3 3,4,5,6,7,8 −0.458 −0.947 4.48×10−2 −0.138 −0.521 2.57×10−2 0.729 0.791 1.63×10−2
0.5410 4 −0.421 −0.951 3.78×10−2 −0.151 −0.521 2.09×10−2 0.725 0.761 1.32×10−2

The fit parameters for the three different optimised s- d) available. For our staggered fermion implementation
t1 pairs are listed in Table III. They support our earlier of the Dirac operator, L has to be even, and the small-
scaling observations: d1 close to −1 means that P scales est L possible is 4. Because of small values of d that we
−d
roughly as 2p , and d2 close to −0.5 means that t2 scales have worked with, and discrete nature of t2 , our data for
roughly as N/2d . Interpreting Grover’s algorithm as L = 4 do not show smooth asymptotic behaviour. We
the d → ∞ limit of the spatial search algorithm, we ex- therefore omitted L = 4 from our analysis, and our best
pect c3 to be close to π/4. It turns out to be somewhat results are for L = 6.
smaller, but is not very accurately determined due to siz- We can fit the data reasonably well as
able correction from d3 for the values of d that we have.
t2 bl
We believe that simulations for larger lattice sizes and √ = al + . (29)
higher dimensions can remedy this problem. NP d
To look at our data in a different manner, we analyse The fit parameters for the three different optimised s-t1
√
the scaling behaviour of t2 / N P as a function of the pairs are listed in Table IV. We note that al is close to
database size. In Fig.9, we plot a subset of the data, il- π/4 when data for d > 6 are available, but it cannot be
√
lustrating how t2 / N P behaves for fixed lattice size L accurately determined using data for small d only due to
when the database size is changed by changing d. The ap- sizable corrections from bl .
proach to the asymptotic
√ behaviour is particularly clear
in this form, with t2 / N P decreasing toward π/4 as the
database size increases. It is also obvious that, for fixed IV. RESULTS FOR MULTIPLE
MARKED VERTICES
N , our algorithm improves in efficiency with decreasing
L (or equivalently with increasing d). This is a conse-
quence of the increase in the number of neighbours of We also carried out a few simulations of spatial search
a vertex with increasing d. An important implication is with more than one marked vertex. The change from
that, given a database of size N , it is best to implement the single marked vertex algorithm is that the search or-
the algorithm using the smallest L (and hence the largest acle now flips the sign of amplitudes at several distinct
vertices, say M . To see clear patterns in the results, we
needed to keep the marked vertices well separated and
avoid interference effects.
1.1 With the staggered fermion implementation of the
Dirac operator, it matters whether the marked vertices

1 √
L=6 TABLE IV: Fit parameters for t2 / N P vs. d at fixed L.
L=8
t2 / N P

L=16 s t1 L Fitted d al bl Error

0.9 0.766 0.551 1.06×10−2
√

L=32 6 5,6,7,8,9
0.9539 2 8 4,5,6,7,8 0.805 0.293 5.31×10−4
16 3,4,5,6 0.663 1.132 1.73×10−2
π/4 32 3,4,5 0.625 1.304 5.77×10−3
Grover
6 4,5,6,7,8,9 0.819 0.006 1.28×10−2
√ 0.803 0.234 3.28×10−3
1/ 2 3 8 4,5,6,7,8
0.7
5 10 15 20 25 16 3,4,5,6 0.773 0.471 6.46×10−3
log2 N 32 3,4,5 0.724 0.705 3.23×10−3
√ 6 4,5,6,7,8,9 0.837 −0.092 1.35×10−2
FIG. 9: (Color online) t2 / N P as a function of the database
size for different lattice sizes. The points 0.5410 4 8 4,5,6,7,8 0.791 0.272 3.79×10−3
√ are the data for
t1 = 3, and the curves are the fits (t2 / N P ) = al + (bl /d). 16 3,4,5,6 0.767 0.450 1.29×10−3
Also shown is the limiting value corresponding to Grover’s 32 3,4,5 0.711 0.726 1.13×10−3
optimal algorithm.
 11

TABLE V: Spatial search results for multiple marked vertices in d = 3.

s t1 M Marked vertices P t2
1 (32,32,32) 0.09829 161
(0,32,32), (32,32,32) 0.04919, 0.04919 112, 112
√ 2
1/ 2 3 (0,32,33), (32,32,32) 0.09868, 0.09790 161, 161
(0,0,0), (16,16,16), (32,32,32) 0.03530, 0.03264, 0.03082 94, 92, 92
3 (0,0,1), (16,16,16), (32,32,32) 0.09380, 0.05590, 0.04507 161, 117, 109
(0,0,1), (16,16,16), (33,32,32) 0.09298, 0.09838, 0.10347 157, 161, 162

p
belong to the same corner of the elementary hypercube i.e. proportional to N/M .
or to different corners. So we considered various hyper-
cube locations for the marked vertices. In Table V, we
compare our illustrative results for two and three marked
vertices with those for a single marked vertex. The val- V. SUMMARY
ues for peak probability √ P and corresponding time t2
were obtained with s = 1/ 2 and t1 = 3 on a 643 lattice.
When the marked vertices belong to the same corner of We have formulated the spatial search problem us-
the elementary hypercube (e.g. (0,32,32) and (32,32,32)), ing staggered fermion discretisation of the Dirac oper-
they involve the same degree of freedom. In that case, ator on a hypercubic lattice in arbitrary dimension. Our
we observe that P decreases by a factor of M and t2 de- search strategy, Eq.(19), mimics that of Grover’s algo-
√
creases by a factor of M . On the other hand, when rithm. The locality restriction on the walk operator be-
the marked vertices belong to different corners of the ele- comes less and less relevant as the dimensionality of the
mentary hypercube (e.g. (0,32,33) and (32,32,32)), they space increases, and we expect to reach the optimal scal-
involve different degrees of freedom. In that case, we ing behaviour of Grover’s algorithm as d → ∞.
find that P and t2 are more or less the same as for the We have verified our theoretical expectations by nu-
single marked vertex case. This pattern is also followed merical simulations of the algorithm for d = 3 to d = 9.
when some marked vertices are at the same corner of the We clearly demonstrate the approach to the d → ∞
elementary hypercube and some at different corners. limit, and find that even for d = 3 our algorithm is only
We can infer the following equipartition rule from these 25% less efficient. From a computational cost point of
observations. A fixed peak probability (upper bounded view, for a fixed database size N , our best parameters
by 2−d ) is available for each corner degree of freedom are t1 = 2 and the largest d available. With the stag-
of the elementary hypercube. Multiple marked vertices gered fermion implementation, our algorithm works in
corresponding to the same degree of freedom share this total Hilbert space dimension N . That is a big reduction
peak probability equally, while different degrees of free- from the total Hilbert dimension 2d N required by the
dom evolve independently of each other. Apart from this algorithms that use coin or internal degrees of freedom
caveat, the dependence of t2 on M for our spatial search (for the best d = Θ(ln N ), 2d = O(N ) is substantial),
algorithm is the same as in the case of Grover’s algorithm, and makes our exercise worthwhile.

[1] L.K. Grover, Proceedings of STOC’96 (ACM Press, New [8] L.K. Grover, Pramana 56 (2001) 333,
York, 1996), p. 212, arXiv:quant-ph/9605043. arXiv:quant-ph/0109116.
[2] N. Shenvi, J. Kempe and K. Birgitta Whaley, Phys. Rev. [9] Y. Aharonov, L. Davidovich and N. Zagury, Phys. Rev.
A67 (2003) 052307, arXiv:quant-ph/0210064. A48 (1993) 1687.
[3] A. Ambainis, J. Kempe and A. Rivosh, Proceedings of [10] J. Kempe, Contemporary Physics 44 (2003) 307,
ACM-SIAM SODA’05 (ACM Press, New York, 2005), arXiv:quant-ph/0303081.
p. 1099, arXiv:quant-ph/0402107. [11] L. Susskind, Phys. Rev. D16 (1977) 3031.
[4] A.M. Childs and J. Goldstone, Phys. Rev. A70 (2004) [12] A. Patel, K.S. Raghunathan and Md.A. Rahaman, Phys.
042312, arXiv:quant-ph/0405120. Rev. A82 (2010) 032331, arXiv:1003.5564.
[5] A. Tulsi, Phys. Rev. A78 (2008) 012310, arXiv:0801.0497. [13] A. Patel, K.S. Raghunathan and P. Rungta, Phys. Rev.
[6] B. Hein and G. Tanner, J. Phys. A42 (2009) 085303, A71 (2005) 032347, arXiv:quant-ph/0405128.
arXiv:0906.3094. [14] This is in spite of the fact that Dirac spinors require
[7] G. Abal, R. Donangelo, F.L. Marquezino and R. Por- 2[(d+1)/2] components in d dimensions. For example, in
tugal, Math. Struct. Comp. Sci. (2010), to appear, d = 3, a Rubik’s cube can have the central cube labeled
arXiv:1001.1139. “o” and the eight corner cubes labeled “e”.
 12

[15] A. Patel, K.S. Raghunathan and P. Rungta, Pro- and H.E. Brandt (AMS, Providence, 2002), p. 53,
ceedings of the Workshop on Quantum Information, arXiv:quant-ph/0005055.
Computation and Communication (QICC-2005), IIT [17] C. Zalka, Phys. Rev. A60 (1999) 2746,
Kharagpur, India, (Allied Publishers, 2006), p. 41, arXiv:quant-ph/9711070.
arXiv:quant-ph/0506221. [18] C. Bennett, E. Bernstein, G. Brassard and U.
[16] G. Brassard, P. Hoyer, M. Mosca and A. Tapp, in Quan- Vazirani, SIAM J. Comput. 26 (1997) 1510,
tum Computation and Information, AMS Contempo- arXiv:quant-ph/9701001.
rary Mathematics Series Vol. 305, eds. S.J. Lomonaco