PHYSICAL REVIEW A 67, 052307 共2003兲

Quantum random-walk search algorithm

Neil Shenvi,1 Julia Kempe,1,2,3 and K. Birgitta Whaley1
1
Department of Chemistry, University of California, Berkeley, California 94720
2
Computer Science Division, EECS, University of California, Berkeley, California 94720
3
CNRS-LRI, UMR 8623, Université de Paris-Sud, 91405 Orsay, France
共Received 9 October 2002; published 23 May 2003兲
Quantum random walks on graphs have been shown to display many interesting properties, including
exponentially fast hitting times when compared with their classical counterparts. However, it is still unclear
how to use these novel properties to gain an algorithmic speedup over classical algorithms. In this paper, we
present a quantum search algorithm based on the quantum random-walk architecture that provides such a
speedup. It will be shown that this algorithm performs an oracle search on a database of N items with O( 冑N)
calls to the oracle, yielding a speedup similar to other quantum search algorithms. It appears that the quantum
random-walk formulation has considerable flexibility, presenting interesting opportunities for development of
other, possibly novel quantum algorithms.

DOI: 10.1103/PhysRevA.67.052307 PACS number共s兲: 03.67.Lx, 89.70.⫹c

I. INTRODUCTION the continuous-time quantum random walk, introduced in

Recent studies of quantum random walks have suggested
that they may display different behavior than their classical
counterparts 关1–5兴. One of the promising features of these
quantum random walks is that they provide an intuitive
framework on which to build novel quantum algorithms.
Since many classical algorithms can be formulated in terms
of random walks, it is hoped that some of these may be
translated into quantum algorithms which run faster than
their classical counterparts. However, previous to a very re-
cent paper by Childs et al. 关6兴, there had been no quantum
algorithms based on the random-walk model. In this paper
we show that a quantum search algorithm can be derived
from a certain kind of quantum random walk. Optimal quan-
tum search algorithms are already well known 关7–9兴. The
search algorithm from a quantum random walk we present
here shows some differences from the established search al-
gorithms and may possess useful properties with respect to
robustness to noise and ease of physical implementation. It
also provides a new direction for design of quantum algo-
rithms from random walks, which may eventually lead to
entirely new algorithms.
Current research uses two distinct models for quantum
random walks, based on either discrete-time steps or on
continuous-time evolution. Discrete time quantum random
walks were introduced as a possible new tool for quantum
algorithms generalizing discrete classical Markov chains 关2兴.
The discrete-time walk can be thought of as a succession of
unitary operations, each of which has a nonzero transition
amplitude only between neighboring nodes of the graph. The
relation of these to classical Markov chains provides consid-
erable motivation for exploration of discrete random walks.
Within the field of classical algorithms, the application of
classical Markov chains in classical algorithms has been
quite revolutionary, providing new approximation and opti-
mization algorithms. By analogy, it might reasonably be
hoped that similar algorithmic advances could be obtained
for quantum algorithms from development of the quantum
random walks. The second quantum random-walk model is
Refs. 关4 – 6兴. In the continuous-time walk, the adjacency ma-
trix of the graph is used to construct a Hamiltonian which
gives rise to a continuous-time evolution. This model differs
from the discrete-time walk in that even for small times there
is an 共exponentially small兲 probability of transition to non-
adjacent nodes. In this paper, we will consider the discrete-
time model only.
The paper is organized as follows. Section II provides a
brief introduction to discrete-time quantum random walks.
Section III describes the random-walk search algorithm and
provides a proof of its correctness. Section IV summarizes
the similarities and differences between the random-walk
search algorithm and Grover’s search algorithm. Conclusions
are presented in Sec. V.
Notation. Following standard computer science notation
we will use the following to characterize the growth of cer-
tain functions: We will say f (n)⫽O„g(n)… if there are posi-
tive constants c and k such that 0⭐ f (n)⭐cg(n) for n⭓k.
Similarly f (n)⫽⍀„g(n)… if 0⭐cg(n)⭐ f (n) for constants
c,k⭓0 and n⭓k.
II. BACKGROUND
The discrete-time random walk can be described by the
repeated application of a unitary evolution operator U. This
operator acts on a Hilbert space H C 丢 H S , where H C is the
Hilbert space associated with a quantum coin 共coin space兲
and H S is the Hilbert space associated with the nodes of the
graph. The operator U can be written as 关2兴
U⫽SC, 共1兲
where S is a permutation matrix which performs a controlled
shift based on the state of the coin space, and C is a unitary
matrix which corresponds to ‘‘flipping’’ the quantum coin.
We will call C the quantum coin. This operation can be vi-
sualized by analogy to a classical random walk. In each it-
eration of a discrete-time classical random walk on a graph,
the coin is flipped. The walker then moves to an adjacent

1050-2947/2003/67共5兲/052307共11兲/$20.00 67 052307-1 ©2003 The American Physical Society

SHENVI, KEMPE, AND WHALEY
node specified by the outcome of the coin flip. An equivalent
process occurs in the quantum random walk, with the modi-
fication that the coin is a quantum coin, and can therefore
exist in a superposition of states. This modification can lead
to dramatic differences in behavior between the classical and
quantum random walks. However, it should be noted that if
the state of the coin is measured after each flip, then the
quantum random walk reverts to a classical random walk
共and similarly if the state of the nodes is measured after
every step兲.
An important feature of the discrete-time quantum ran-
dom walk that has significance for its use in the development
of quantum algorithms is that by virtue of its definition on a
quantum computer this walk will be efficiently implement-
able whenever its classical counterpart is efficiently imple-
mentable on a classical computer. 关By efficient we mean that
the walk can be simulated by a circuit with a number of gates
that is polynomial in the number of bits 共qubits兲.兴 This is due
to the very similar structure of both these walks. To illustrate
this, assume that we have an efficient way to implement the
classical random walk on the underlying graph, i.e., to per-
form the coin flip and subsequent shift. The shift is condi-
tional on the outcome of the coin flip 共which determines the
direction of the next step兲, i.e., we have a classical efficient
circuit that performs a controlled shift on the basis states. It
is straightforward 关10兴 to translate this circuit into a quantum
circuit that performs the unitary controlled shift of Eq. 共1兲.
Similarly, if there is an efficient procedure to flip the classi-
cal coin of the random walk, there will be an efficient way to
implement a quantum coin. Hence implementation of the
discrete-time random walk is automatically efficient if the
underlying classical walk is efficiently implementable.
Note that if no measurement is made, the quantum walk is
controlled by a unitary operator rather than a stochastic one.
This implies that there is no limiting stationary distribution
关2,11兴. Nevertheless, several recent works have shown that
consistent notions of mixing time can be formulated, and
have shown polynomial speedup in these quantum mixing
times relative to the classical analog 关2,11兴. Another quantity
for which quantum walks have shown speedup relative to
their classical analogs is the hitting time 关12,13兴. Under cer-
tain conditions this speedup can be exponential compared to
the classical analog. We refer the reader to the recent papers
关2,3,11兴, and 关12兴 for some results obtained from discrete-
time quantum random walks.
Our random-walk search algorithm will be based on a
random walk on the n cube, i.e., the hypercube of dimension
n 关11,12兴. The hypercube is a graph with N⫽2 n nodes, each
of which can be labeled by an n-bit binary string. Two nodes
on the hypercube described by bitstrings xជ and yជ are con-
nected by an edge if 兩 xជ ⫺yជ 兩 ⫽1, where 兩 xជ 兩 is the hamming
weight of xជ . In other words, if xជ and yជ differ by only a
single-bit flip, then the two corresponding nodes on the
graph are connected. Thus, each of the 2 n nodes on the n
cube has degree n 共i.e., it is connected to n other nodes兲, so
n
the Hilbert space of the algorithm is H⫽H n 丢 H 2 . Each
state in H can be described by a bit string xជ , which specifies
the position on the hypercube, and a direction d, which
052307-2
PHYSICAL REVIEW A 67, 052307 共2003兲
specifies the state of the coin. The shift operator S maps a
state 兩 d,xជ 典 onto the state 兩 d,xជ 丣 ជ
e d 典 , where ជ
e d is the dth basis
vector on the hypercube. S can be written explicitly as
n⫺1
S⫽ 兺 兺xជ 兩 d,xជ 丣 ជe d 典具 d,xជ 兩 .
d⫽0
共2兲
To completely specify the unitary evolution operator U,
the coin operator C must also be chosen. Normally, the coin
operator is chosen such that the same coin action is applied
to each node on the graph. This is the case in previous stud-
ies of discrete quantum walks on the line 关2,3,14兴 and on the
hypercube 关11,12兴. In other words, the coin operator C can
be written in a separable way as
C⫽C 0 丢 I, 共3兲
where C 0 is a n⫻n unitary operator acting on the coin space
H C . In this case the action on the coin space H C does not
depend on the state of the node space H S . If C is separable
according to Eq. 共3兲 then the eigenstates of U are simply the
tensor product of the eigenstates of an operator C kជ on the
coin space and of the Fourier modes of the hypercube 共la-
beled by n-bit strings kជ ) 关11兴. One frequently chosen sepa-
rable coin is Grover’s ‘‘diffusion’’ operator on the coin
space, given by
C 0 ⫽G⫽⫺I⫹2 兩 s C 典具 s C 兩 , 共4兲
where 兩 s C 典 is the equal superposition over all n directions,
i.e., 兩 s C 典 ⫽1/冑n 兺 d⫽1
n
兩 d 典 关11兴. This coin operator is invariant
to all permutations of the n directions, so it preserves the
permutation symmetry of the hypercube. The use of the
Grover diffusion operator as a coin for the hypercube was
proposed in Ref. 关11兴, where it was pointed out that this
operator is the permutation invariant operator farthest away
from the identity operator 关11兴. So, heuristically, it should
provide the most efficient mixing over states, from any given
initial state. The nontrivial eigenvalues and eigenvectors of
U are given by 关11兴
2k 2i
e ⫾i ␻ k ⫽1⫺ ⫾ 冑k 共 n⫺k 兲 , 共5兲
n n
2 ⫺n/2
兩 v kជ 典 , 兩 v kជ 典 * ⫽ 兺
xជ ,d
ជ ជ
共 ⫺1 兲 k •x
冑2
兩 d,xជ 典
再
1/冑k if k d ⫽1
⫻ 共6兲
⫿i/ 冑n⫺k if k d ⫽0.
Note that the equal superposition over all states, 兩 ␺ 0 典 ⫽ 兩 s C 典
丢 兩 s S 典 , where 兩 s S 典 is the equal superposition over the 2 n
nodes, is an eigenvector of U with eigenvalue 1. So repeated
application of U leaves the state 兩 ␺ 0 典 unchanged.
In order to create a search algorithm using the quantum
random-walk architecture, we now consider a small pertur-
bation of the unitary operator U. In the standard setting of a
search algorithm we have an oracle O f , which ‘‘marks’’ a
single bitstring xជ target . More specifically the oracle com-
QUANTUM RANDOM-WALK SEARCH ALGORITHM
FIG. 1. Modified standard oracle that simulates the coin oracle.
The standard oracle acts on the node space as U f : 兩 xជ 典 丢 兩 y 典 → 兩 xជ 典
丢 兩 y 丣 f (x) 典 . The controlled coin operation, denoted by C 0 /C 1 , ap-
plies C 0 on the coin space if the control qubit is in the state 兩 0 典 , and
C 1 if it is in the state 兩 1 典 .
putes a function f such that f (xជ target )⫽1 and f (xជ )⫽0 if xជ
⫽xជ target . The query complexity of a search algorithm is
defined to be the number of queries to O f that need to be
made to find the marked string xជ target with high probability.
In the quantum case the oracle U f is implemented via a
reversible unitary operation; in the standard setting the oracle
shifts the phase of the marked item. For our implementation
the coin operator will take the function of the oracle. Spe-
cifically, we consider ‘‘marking’’ a single arbitrary node by
applying a special coin action to that node. The oracle will
act by applying a ‘‘marking coin’’ C 1 to the marked node and
the original coin C 0 to the unmarked nodes, i.e., the coin
action will be conditioned on the node. 共Note that this modi-
fied coin is still unitary.兲
This ‘‘coin oracle’’ can be easily obtained from the stan-
dard oracle of quantum search. To simulate the coin oracle
we setup the standard oracle on the node space, and then add
a conditional C 0 or C 1 operation, respectively, at the output.
This is illustrated in Fig. 1.
Without loss of generality, we can assume that the marked
node corresponds to the all-zero string xជ target ⫽0ជ . Then our
coin operator becomes
C ⬘ ⫽C 0 丢 I⫹ 共 C 1 ⫺C 0 兲 丢 兩 0ជ 典具 0ជ 兩 . 共7兲
The marking coin C 1 can be any n⫻n unitary matrix. For
simplicity, we will consider here the case where C 1 ⫽⫺I.
Numerical studies 关15兴 have shown that other choices for the
coin C 1 yield similar results. As seen from Eq. 共7兲 the coin
operator is now a composite unitary and its action is condi-
tioned on the node register. Our perturbed unitary evolution
operator U ⬘ is given by
U ⬘ ⫽SC ⬘
⫽S 共 G 丢 I⫺ 共 G⫹I兲 丢 兩 0ជ 典具 0ជ 兩 兲
⫽U⫺2S 共 兩 s C 典具 s C 兩 丢 兩 0ជ 典具 0ជ 兩 兲 . 共8兲
Analysis of the effects of this perturbation leads directly to
the definition of the random-walk search algorithm, as is
described in the following section.
052307-3
PHYSICAL REVIEW A 67, 052307 共2003兲
III. RANDOM WALK SEARCH ALGORITHM
A. Overview of the algorithm
We define the search space of the algorithm to be the set
of all n-bit binary strings, xជ ⫽ 兵 0,1其 n . We consider the func-
tion f (xជ )⫽ 兵 0,1其 , such that f (xជ )⫽1 for exactly one input
xជ target . Our goal is to find xជ target . Using the mapping of
n-bit binary string to nodes on the hypercube, this search
problem is then equivalent to searching for a single marked
node amongst the N⫽2 n nodes on the n cube. For purposes
of the proof, we have set the marked node to be xជ target ⫽0ជ ,
but the location of the marked node has no significance.
The random-walk search algorithm is implemented as fol-
lows.
共1兲 Initialize the quantum computer to the equal superpo-
sition over all states, 兩 ␺ 0 典 ⫽ 兩 s C 典 丢 兩 s S 典 . This can be accom-
plished efficiently on the node space by applying n single-bit
Hadamard operations to the 兩 0ជ 典 state. A similar procedure
works for the direction space.
共2兲 Given a coin oracle C ⬘ which applies the coin C 0
⫽G to the unmarked states and the coin C 1 ⫽⫺I to the
marked state, apply the perturbed evolution operator, U ⬘
⫽SC ⬘ , t f ⫽ ␲ /2冑2 n times.
共3兲 Measure the state of the computer in the 兩 d,xជ 典 basis.
It is our claim that with probability 21 ⫺O(1/n), the out-
come of the measurement will be the marked state. By re-
peating the algorithm a constant number of times, we can
determine the marked state with an arbitrarily small degree
of error. In the remainder of this section we provide a proof
of this algorithm.
The general outline of the proof that we will present is the
following. We need to determine the result of the operation
(U ⬘ ) t on the initial state 兩 ␺ 0 典 . To do this, we will first sim-
plify the problem by showing that the perturbed walk on the
hypercube can be collapsed to a walk on the line 共Theorem
1兲. Next, by constructing two approximate eigenvectors of
U ⬘ , 兩 ␺ 0 典 and 兩 ␺ 1 典 , we will show that there are exactly two
eigenvalues of U ⬘ that are relevant 关i.e., the initial state 兩 ␺ 0 典
has high overlap with the space spanned by the correspond-
ing eigenvectors 共see Theorem 2 and Theorem 3兲兴. We de-
note these eigenvalues by e i ␻ 0⬘ and e ⫺i ␻ 0⬘ . We will then show
that the corresponding eigenvectors 兩 ␻ 0⬘ 典 and 兩 ⫺ ␻ 0⬘ 典 can be
well approximated by linear combinations of the initial state
兩 ␺ 0 典 and the second state 兩 ␺ 1 典 共Theorem 4兲. As a result, our
random walk search algorithm can be approximated by a
two-dimensional rotation in the 兩 ␻ 0⬘ 典 , 兩 ⫺ ␻ ⬘0 典 plane away
from the initial state 兩 ␺ 0 典 ⬇1/冑2( 兩 ␻ 0⬘ 典 ⫹ 兩 ⫺ ␻ 0⬘ 典 ) and towards
兩 ␺ 1 典 ⬇i/ 冑2(⫺ 兩 ␻ 0⬘ 典 ⫹ 兩 ⫺ ␻ 0⬘ 典 ), which constitutes a very close
approximation to the target state 兩 xជ target 典 . Finally, we show
that each application of the evolution operator U ⬘ corre-
sponds to a rotation angle of approximately 1/冑2 n⫺1 Theo-
rem 共5兲. Hence, the search is completed after approximately
( ␲ /2)冑2 n⫺1 steps, i.e., after O( 冑N) calls to the oracle,
where N⫽2 n is the number of nodes.
SHENVI, KEMPE, AND WHALEY PHYSICAL REVIEW A 67, 052307 共2003兲

B. Proof of correctness
In general, analytic determination of the eigenspectrum of
a large matrix is a daunting task, so we will take advantage
of the symmetries inherent in U ⬘ to simplify the problem.
Let us first show that the perturbed random walk on the
hypercube can be collapsed onto a random walk on the line.
Let P i j be the permutation operator which swaps the bits i
and j, in both the node space and the coin space. In other
words, given a state 兩 d,xជ 典 , under the permutation operator,
P i j , the ith and jth bits of xជ are swapped and the directions
d⫽i and d⫽ j are swapped. Clearly, the unperturbed evolu-
tion operator U commutes with P i j since every direction in
the unperturbed walk is equivalent.
Theorem 1. U ⬘ commutes with P i j , i.e., the perturbed
walk on the hypercube can be effectively regarded as a walk
on the line.
Proof.

P †i j U ⬘ P i j ⫽ P †i j U P i j ⫺2 P †i j S• 共 兩 s C 典具 s C 兩 丢 兩 0ជ 典具 0ជ 兩 兲 P i j
n⫺1
2 FIG. 2. Collapsing a random walk on the hypercube to a random
⫽U⫺ 兺 P †i j 兩 d,eជd 典具 d,0兩 P i j
冑n d⫽0
walk on the line. The states on the hypercube are mapped to the
state on the line based on their Hamming weight and the direction
⫽U ⬘ . 共9兲 in which they point 共see text兲.

冉 cos ␻ x sin ␻ x
冊
n
So, 关 U ⬘ , P i j 兴 ⫽0. 䊏
Because the initial state 兩 ␺ 0 典 is an eigenvector of P i j
C 0⫽ 兺
x⫽0 sin ␻ x ⫺cos ␻ x
丢 兩 x 典具 x 兩 , 共13兲
with eigenvalue 1 for all i and j, and 关 U ⬘ , P i j 兴 ⫽0, any
intermediate state 兩 ␺ t 典 ⫽(U ⬘ ) t 兩 ␺ 0 典 must also be an eigen- where cos ␻x⫽1⫺2x/n and sin ␻x⫽(2/n) 冑x(n⫺x) and
vector of eigenvalue 1 with respect to P i j . Thus, (U ⬘ ) t where the first part acts on the space spanned by 兵 兩 R 典 , 兩 L 典 其
preserves the symmetry of 兩 ␺ 0 典 with respect to bit and the second part acts on the positions 兵 兩 0 典 , . . . , 兩 n 典 其 on
swaps. It is therefore useful to define 2n basis states, the line. Note that the coin of the collapsed walk is not ho-
兩 R,0典 , 兩 L,1典 , 兩 R,1典 , . . . , 兩 R,n⫺1 典 , 兩 L•n 典 , where mogeneous in space any more. The unitary operator U on the
restricted space acts as

兩 R,x 典 ⫽
冑 冉冊 共 n⫺x 兲
1
n
x
兺 兺 兩 d,xជ 典 ,
兩 xជ 兩 ⫽x x ⫽0
d
共10兲
U⫽
n⫺1

兺
x⫽0
兩 R,x 典 (⫺cos ␻ x⫹1 具 L,x⫹1 兩 ⫹sin ␻ x⫹1 具 R,x⫹1 兩 兲

冑冉 冊兺 兺 1
⫹ 兺 兩 L,x 典 共 sin ␻ x⫺1 具 L,x⫺1 兩 ⫹cos ␻ x⫺1 具 R,x⫺1 兩 兲 .
兩 d,xជ 典 ,
x⫽1
兩 L,x 典 ⫽ 共11兲
n 兩 xជ 兩 ⫽x x d ⫽1 共14兲
x
x
Similarly,
which are also invariant to bit swaps P i j . These states span
the eigenspace of eigenvalue 1 of P i j . Using these basis U ⬘ ⫽U⫹⌬U⫽U⫺2 兩 L,1典具 R,0兩 . 共15兲
states, we can project out all but one spatial degree of free-
dom and effectively reduce the random walk on the hyper- Note that the only difference between U and U ⬘ is in the sign
cube to a random walk on the line. This is illustrated in Fig. of the matrix element in position ( 兩 L,1典 , 兩 R,0典 ). Finally,

冉 冑冉 冊
2. The marked node corresponds now to 兩 R,0典 . We can re-
write U, U ⬘ , and 兩 ␺ 0 典 in this collapsed basis. First note that n⫺1
n⫺1
the shift operator S in this basis acts as 1 1 x⫺1
兩 ␺ 0典 ⫽
冑2 n
兩 R,0典 ⫹
冑2 n
兩 L,n 典 ⫹ 兺
x⫽1 2n
兩 L,x 典

冑冉 冊 冊
n⫺1

S⫽ 兺
x⫽0
兩 R,x 典具 L,x⫹1 兩 ⫹ 兩 L,x⫹1 典具 R,x 兩 共12兲 n⫺1
x
⫹ 兩 R,x 典 . 共16兲
and the unperturbed coin acts as 2n

052307-4
QUANTUM RANDOM-WALK SEARCH ALGORITHM PHYSICAL REVIEW A 67, 052307 共2003兲

Re 冉兺i
冊 冉兺
具 ␻ ⬘i 兩 U ⬘ 兩 ␻ ⬘i 典 ⫽Re
i
冊
e i ␻ i⬘ 具 ␻ ⬘i 兩 ␻ ⬘i 典 ⬎3⫺2/n.
共19兲

Let us define ⍀ to be the subspace spanned by 兩 ␻ ⬘0 典 , 兩 ␻ ⬘1 典 ,

兩 ␻ 2⬘ 典 . Then we can write Eq. 共19兲 as the partial trace of U ⬘
over ⍀,

Re Tr⍀ U ⬘ ⬎3⫺2/n. 共20兲

Let us now define 兩 ␺ ⫺ 典 ⫽1/冑2( 兩 0,R 典 ⫺ 兩 1,L 典 ). We can ex-

pand the 兩 ␻ 0⬘ 典 , 兩 ␻ 1⬘ 典 , and 兩 ␻ 2⬘ 典 in terms of 兩 ␺ 0 典 , 兩 ␺ ⫺ 典 , and a
residual vector,

兩 ␻ ⬘0 典 ⫽c 00
⬘ 兩 ␺ 0 典 ⫹c 01
⬘ 兩 ␺ ⫺ 典 ⫹c 02
⬘ 兩 r ⬘0 典 ,
FIG. 3. The results of numerical spectral analysis of U and U ⬘
for n⫽8. The circles indicate eigenvalues of U. The crosses indi- 兩 ␻ 1⬘ 典 ⫽c 10
⬘ 兩 ␺ 0 典 ⫹c 11
⬘ 兩 ␺ ⫺ 典 ⫹c 11
⬘ 兩 r 1⬘ 典 , 共21兲
cate eigenvalues of U ⬘ .
兩 ␻ 2⬘ 典 ⫽c 20
⬘ 兩 ␺ 0 典 ⫹c 21
⬘ 兩 ␺ ⫺ 典 ⫹c 22
⬘ 兩 r 2⬘ 典 ,
Since U and P i j are mutually diagonalizable, the eigenvec-
tors of U in the reduced space are also bit-flip invariant. where 兩 r ⬘i 典 is a normalized vector orthogonal to 兩 ␺ 0 典 and
Examining Eq. 共5兲, it is clear that if we take the equal super- 兩 ␺ ⫺ 典 . We now observe that, due to the basis invariance of
positions of all eigenvectors of same eigenvalue 兩 v kជ 典 such the trace, Eq. 共20兲 holds for any linear combination of 兩 ␻ 0⬘ 典 ,
that 兩 kជ 兩 ⫽k, the resulting eigenvector will be bit-swap invari- 兩 ␻ 1⬘ 典 , and 兩 ␻ 2⬘ 典 . Thus, we can construct three new orthonor-
ant. Thus we define mal vectors, 兩 ␣ 0 典 , 兩 ␣ 1 典 , and 兩 ␣ 2 典 by taking linear combina-
tions of 兩 ␻ ⬘0 典 , 兩 ␻ ⬘1 典 , and 兩 ␻ ⬘2 典 , such that
1
兩 ␻ k典 ⫽ 兺 兩 v kជ 典 , 共17兲
冑冉 冊 n
k
兩 kជ 兩 ⫽k
具 ␣ 2 兩 ␺ 0 典 ⫽ 具 ␣ 2 兩 ␺ ⫺ 典 ⫽0.
In other words, we can expand 兩 ␣ 0 典 , 兩 ␣ 1 典 , and 兩 ␣ 2 典 as
共22兲

which are the eigenvectors of U with eigenvalues e i ␻ k in the 兩 ␣ 0 典 ⫽c 00兩 ␺ 0 典 ⫹c 01兩 ␺ ⫺ 典 ⫹c 02兩 r 0 典 ,
collapsed 共symmetric兲 space.
Note that both U and U ⬘ are represented by real matrices; 兩 ␣ 1 典 ⫽c 10兩 ␺ 0 典 ⫹c 11兩 ␺ ⫺ 典 ⫹c 12兩 r 1 典 , 共23兲
therefore, their eigenvalues and eigenvectors will come in
complex-conjugate pairs. 兩 ␣ 2典 ⫽ 兩 r 2典 .
Having determined these general properties of the per-
turbed matrix U ⬘ , we now turn to the problem of analyzing Since 兩 ␣ 0 典 , 兩 ␣ 1 典 , and 兩 ␣ 2 典 still form a basis for ⍀, from Eq.
the eigenvalue spectrum of U ⬘ . Let A be the arc on the unit 共20兲 it follows that
circle containing all complex numbers of unit norm with real
part greater than 1⫺2/(3n). In other words, 3⫺2/n⬍Re 兺i 具 ␣ i兩 U ⬘兩 ␣ i 典 . 共24兲

再
A⫽ z:Re z⬎1⫺
2
3n
, 兩 z 兩 ⫽1 . 冎 共18兲 Since U ⬘ is a unitary operator, we know that Re具 ␣ i 兩 U ⬘ 兩 ␣ i 典
⭐1 for all 兩 ␣ i 典 . Thus, applying this inequality to the first
two terms in the sum, we obtain
Figure 3 shows the geometrical representation of A together
with the eigenvalue spectra of the unperturbed and perturbed
matrices for n⫽8. We will prove that A contains exactly two Re 兺i 具 ␣ i兩 U ⬘兩 ␣ i 典 ⭐2⫹Re具 ␣ 2兩 U ⬘兩 ␣ 2 典 . 共25兲
eigenvalues e i ␻ ⬘0 and e ⫺i ␻ 0⬘ of U ⬘ . First, we will prove that
there are at most two eigenvalues with real part greater than Since, U ⬘ ⫽U⫹⌬U, we can write
1⫺2/(3n). Then we will show that there are at least two
eigenvalues on A. From these facts, it follows that there are Re具 ␣ 2 兩 U ⬘ 兩 ␣ 2 典 ⫽Re具 ␣ 2 兩 U 兩 ␣ 2 典 ⫹Re具 ␣ 2 兩 ⌬U 兩 ␣ 2 典 . 共26兲
exactly two eigenvalues of U ⬘ on A.
Theorem 2. There are at most two eigenvalues of U ⬘ with Let us first consider 具 ␣ 2 兩 U 兩 ␣ 2 典 . We can expand 兩 ␣ 2 典 in
real part greater than 1⫺2/(3n). terms of the unperturbed eigenstates, 兩 ␣ 2 典 ⫽ 兺 j b j 兩 ␻ j 典 . So,
Proof. We will prove by contradiction. Let us assume that Re具 ␣ 2 兩 U 兩 ␣ 2 典 ⫽ 兺 j 兩 b j 兩 2 cos ␻ j . However, since 具 ␣ 2 兩 ␺ 0 典 ⫽0,
there are three eigenvalues, e i ␻ ⬘0 , e i ␻ ⬘1 , and e i ␻ 2⬘ , with real there is no contribution from the eigenvalue with value 1.
part greater than 1⫺2/(3n). Let 兩 ␻ 0⬘ 典 , 兩 ␻ 1⬘ 典 , and 兩 ␻ ⬘2 典 be the The eigenvalue with the next-largest real part is e i ␻ 1 ⫽1
corresponding eigenvectors. Then, ⫺2/n⫹i(2/n) 冑n⫺1. Thus,

052307-5
SHENVI, KEMPE, AND WHALEY PHYSICAL REVIEW A 67, 052307 共2003兲

Re具 ␣ 2 兩 U 兩 ␣ 2 典 ⭐1⫺2/n. 共27兲 Hence,

Next, we consider 具 ␣ 2 兩 ⌬U 兩 ␣ 2 典 . Let 兩 ␺ ⫹ 典 ⫽1/冑2( 兩 0,R 典 1

冉 冊
⫹ 兩 1,L 典 ). Using Eq. 共15兲 we can express ⌬U in terms of 具 ␺ 1 兩 U ⬘ 兩 ␺ 1 典 ⫽1⫺ . 共36兲
兩 ␺ ⫺ 典 and 兩 ␺ ⫹ 典 , n⫺1
2
2c
n/2
⌬U⫽ 兩 ␺ ⫺ 典具 ␺ ⫺ 兩 ⫹ 兩 ␺ ⫺ 典具 ␺ ⫹ 兩 ⫺ 兩 ␺ ⫹ 典具 ␺ ⫺ 兩 ⫺ 兩 ␺ ⫹ 典具 ␺ ⫹ 兩 .
共28兲
Expanding Eq. 共34兲 we find 1⬍c 2 ⬍1⫹2/n for sufficiently
But since 具 ␣ 2 兩 ␺ ⫺ 典 ⫽0 关see Eq. 共22兲兴, large n. Thus, except for a small residual, 兩 ␺ 1 典 is ‘‘almost’’
an eigenvector of U ⬘ with eigenvalue 1.
具 ␣ 2 兩 ⌬U 兩 ␣ 2 典 ⫽ 共 ⫺ 兩 具 ␺ ⫹ 兩 ␣ 2 典 兩 2 兲 ⭐0. 共29兲 Now let us verify that there is at least one eigenvalue of
U ⬘ on A. Let us assume that there are no eigenvalues of U ⬘
Then, Re具 ␣ 2 兩 U ⬘ 兩 ␣ 2 典 ⭐1⫺2/n. Combining Eqs. 共25兲, 共27兲, on A. Then cos ␻⬘j ⬍1⫺2/(3n) for all j. Then using Eq. 共32兲,
and 共29兲, we obtain
1⫺1/2n⫺1 ⫽Re具 ␺ 0 兩 U ⬘ 兩 ␺ 0 典
Re 兺i 具 ␣ i兩 U ⬘兩 ␣ i 典 ⭐3⫺2/n. 共30兲
⫽ 兺j 兩 具 ␺ 0兩 ␻ ⬘j 典 兩 2 cos ␻ ⬘j
Since this contradicts Eq. 共24兲, our assumption must be
false. 䊏
Theorem 3. There are at least two eigenvalues of U ⬘ on ⬍„1⫺2/共3n 兲 … 兺j 兩 具 ␺ 0兩 ␻ ⬘j 典 兩 2
A.
Proof. We will construct two approximate eigenvectors of ⫽1⫺2/共3n 兲 , 共37兲
U ⬘ , 兩 ␺ 0 典 and 兩 ␺ 1 典 . 兩 ␺ 0 典 is given by Eq. 共16兲. Using Eq. 共15兲,
which is wrong for n⬎3. Hence our assumption is false and
U ⬘ 兩 ␺ 0 典 ⫽ 兩 ␺ 0 典 ⫺2/冑2 n 兩 L,1典 , 共31兲
there must be at least one eigenvalue of U ⬘ on A.
and Now let us assume that there is exactly one eigenvalue of
U ⬘ , e i ␻ 0⬘ , on A. Then,
具 ␺ 0 兩 U ⬘ 兩 ␺ 0 典 ⫽ 具 ␺ 0 兩 ␺ 0 典 ⫺ 具 ␺ 0 兩 L,1典具 R,0兩 ␺ 0 典
⫽1⫺1/2n⫺1 . 共32兲 1
1⫺ ⫽Re具 ␺ 0 兩 U ⬘ 兩 ␺ 0 典
2 n⫺1
So, apart from a small residual, 兩 ␺ 0 典 is also ‘‘almost’’ an
eigenvector of U ⬘ with eigenvalue 1. Now, we need to find a
second approximate eigenvector 兩 ␺ 1 典 . Let ⫽ 兺j 兩 具 ␺ 0兩 ␻ ⬘j 典 兩 2 cos ␻ ⬘j

冉
n/2⫺1
1
兩 ␺ 1典 ⫽ 兺 兩 R,x 典 ⫽ 兩 具 ␺ 0 兩 ␻ 0⬘ 典 兩 2 cos ␻ 0⬘ ⫹ 兺 兩 具 ␺ 0兩 ␻ ⬘j 典 兩 2 cos ␻ ⬘j
x⫽0
冑冉 冊 2
n⫺1
x
j⫽0

⭐ 兩 具 ␺ 0 兩 ␻ 0⬘ 典 兩 2 ⫹ 共 1⫺ 兩 具 ␺ 0 兩 ␻ 0⬘ 典 兩 2 兲 „1⫺2/共3n 兲 ….

⫺
冑冉 冊 2
1
n⫺1
x
兩 L,x⫹1 典
冊 冒 c, 共33兲
Rearranging terms,
共38兲

3n
where c is a normalization constant, 兩 具 ␺ 0 兩 ␻ ⬘0 典 兩 2 ⭓1⫺ . 共39兲

冑
2n
n/2⫺1
1
兺
冉 冊
c⫽ . 共34兲 If we use 兩 ␺ 1 典 as a trial vector and follow the same argu-
x⫽0 n⫺1
ments, we obtain the inequality
x

Using this definition and Eq. 共14兲, we see that 3n

冉 冊
兩 具 ␺ 1 兩 ␻ 0⬘ 典 兩 2 ⭓1⫺ . 共40兲
n⫺1
1 2
4c
U ⬘兩 ␺ 1典 ⫽ 兩 ␺ 1典 ⫺ 共 兩 R,n/2⫺1 典 ⫹ 兩 L,n/2⫹1 典 ).
冑冉 冊
n/2
n⫺1
c 2
n/2 But since 兩 ␺ 0 典 and 兩 ␺ 1 典 are orthonormal, this leads to a
共35兲 contradiction, since,

052307-6
QUANTUM RANDOM-WALK SEARCH ALGORITHM PHYSICAL REVIEW A 67, 052307 共2003兲

1⫽ 具 ␻ ⬘0 兩 ␻ 0⬘ 典 ⭓ 兩 具 ␺ 0 兩 ␻ 0⬘ 典 兩 2 ⫹ 兩 具 ␺ 1 兩 ␻ ⬘0 典 兩 2 3n
p 0 ⭓1/2⫺ . 共44兲
3n 3n 2 n⫹1

冉 冊
⭓2⫺ ⫺ , 共41兲
2n n⫺1 Using Eq. 共36兲 and the same arguments as above we obtain
2
4c
n/2
3n

冉 冊
which is not true for large n. Hence, there must be at least p 1 ⭓1/2⫺ . 共45兲
n⫺1
two eigenvalues on the arc A. 䊏 8c 2
As noted above, the eigenvalues and eigenvectors of U ⬘ n/2
come in complex-conjugate pairs. In particular, the two ei-
genvalues on A must be a complex-conjugate pair; let e ⫾i ␻ ⬘0 Up to a global phase 兩 ␻ 0⬘ 典 can be written as
be the two eigenvalues on A. The corresponding eigenvec-
兩 ␻ 0⬘ 典 ⫽ 兩 具 ␻ 0⬘ 兩 ␺ 0 典 兩兩 ␺ 0 典 ⫹ 兩 具 ␻ 0⬘ 兩 ␺ 1 典 兩 e i ␩ 兩 ␺ 1 典 ⫹ 冑1⫺ p 0 ⫺p 1 兩 r 0 典 ,
tors obey 兩 ⫺ ␻ 0⬘ 典 ⫽ 兩 ␻ ⬘0 典 * 共if e i ␻ 0⬘ ⫽e ⫺i ␻ ⬘0 ⫽1, then we can
共46兲
construct linear combinations of 兩 ␻ 0⬘ 典 and 兩 ⫺ ␻ 0⬘ 典 for which
this statement is true兲. We will now show that 兩 ⫾ ␻ 0⬘ 典 can be which yields Eq. 共42兲 for 兩 ␻ 0⬘ 典 and 兩 ⫺ ␻ ⬘0 典 .
well approximated by linear combinations of 兩 ␺ 0 典 and 兩 ␺ 1 典 . To estimate e i ␩ note that since 兩 ␻ 0⬘ 典 and 兩 ⫺ ␻ 0⬘ 典 are eigen-
Theorem 4. The two eigenvectors with eigenvalues close vectors of a unitary matrix, they must be orthogonal. Conse-
to 1 can be well approximated by linear combinations of the quently,
initial state 兩 ␺ 0 典 and the state 兩 ␺ 1 典 , as 兩 ⫾ ␻ ⬘0 典 ⬇1/冑2( 兩 ␺ 0 典
⫾i 兩 ␺ 1 典 ). More precisely, 0⫽ 具 ⫺ ␻ 0⬘ 兩 ␻ 0⬘ 典 ⫽ p 0 ⫹ p 1 共 e i ␩ 兲 2 ⫹ 共 1⫺ p 0 ⫺ p 1 兲 具 r 0* 兩 r 0 典 .
共47兲
兩 ␻ 0⬘ 典 ⫽ 冑p 0 兩 ␺ 0 典 ⫹ 冑p 1 e i ␩ 兩 ␺ 1 典 ⫹ 冑1⫺ p 0 ⫺p 1 兩 r 0 典 ,
Solving for e i ␩ , we obtain
兩 ⫺ ␻ ⬘0 典 ⫽ 冑p 0 兩 ␺ 0 典 ⫹ 冑p 1 e ⫺i ␩
兩 ␺ 1 典 ⫹ 冑1⫺ p 0 ⫺p 1 兩 r 0 典 * ,
共42兲 ⫺ p 0 ⫺ 共 1⫺ p 0 ⫺ p 1 兲 Re具 r *
0 兩 r 0典
Re共 e i ␩ 兲 2 ⫽ . 共48兲
p1
where p 0 ⫽ 兩 具 ␻ 0⬘ 兩 ␺ 0 典 兩 ⫽ 兩 具 ⫺ ␻ 0⬘ 兩 ␺ 0 典 兩 , p 1 ⫽ 兩 具 ␻
2 ˜ ⬘0兩 ␺ 1典 兩 2
2

⫽兩具⫺␻ ˜ ⬘0 兩 ␺ 1 典 兩 , and 兩 r 0 典 is a normalized vector orthogonal

2 Assume Re具 r *
0 兩 r 0 典 ⭓0. Then, using 1/(1⫺x)⭐1⫹2x for
to 兩 ␺ 0 典 and 兩 ␺ 1 典 . Furthermore, 1/2⭓ p 0 ⭓1/2⫺3n/2n⫹1 and small x, we get

3n 3n p0

冉 冊
1/2⭓p 1 ⭓1/2⫺ , ⫺1⫹ ⭓⫺
n⫺1 2 n p1
8c 2
n/2 ⭓Re共 e i ␩ 兲 2
with e i ␩ ⫽i⫹⌬, where

兩 ⌬ 兩 ⫽O
冉冉 冊冊 n
n⫺1
.
⭓⫺
p 0⫹
冉 2
3n
n⫹1
⫹
8c 2 冉 冊冊
3n
n⫺1
n/2
p1

冉 冉 冊冊
n/2
3n 3n
Proof. Since 兩 ␺ 0 典 and 兩 ␺ 1 典 are real vectors, 兩 具 ␻ 0⬘ 兩 ␺ 0 典 兩 2 ⭓⫺1⫺2 ⫹
⫽ 兩 具 ⫺ ␻ ⬘0 兩 ␺ 0 典 兩 2 ⭐1/2 and 兩 具 ␻ ⬘0 兩 ␺ 1 典 兩 2 ⫽ 兩 具 ⫺ ␻ ⬘0 兩 ␺ 1 典 兩 2 ⭐1/2. n⫺1
n⫹1
2
8c 2
Using Eq. 共32兲, n/2

1⫺
2
1
n⫺1
⫽Re具 ␺ 0 兩 U ⬘ 兩 ␺ 0 典 ⫺4
冉 8c 2
3n

冉n⫺1
n/2
冊冊
, 共49兲

⫽ 兺j cos ␻ ⬘j 兩 具 ␻ ⬘j 兩 ␺ 0 典 2 兩
which in turn implies that e i ␩ ⫽i⫹⌬ with 兩 ⌬ 兩 ⫽O( n/2 n⫺1
). A
similar reasoning holds if Re具 r * 0 0典
兩 r ⭐0. 䊏
⫽2 p 0 cos ␻ 0⬘ ⫹ 兺
j⫽0
cos ␻ ⬘j 兩 具 ␻ ⬘j 兩 ␺ 0 典 兩 2 This means that the initial state can be approximately
written as 兩 ␺ 0 典 ⬇1/冑2( 兩 ␻ ⬘0 典 ⫹ 兩 ⫺ ␻ 0⬘ 典 ) and evolves as
⬍2 p 0 ⫹„1⫺2/共3n 兲 …共 1⫺2p 0 兲 . 共43兲 (U ⬘ ) t 兩 ␺ 0 典 ⬇1/冑2(e it ␻ 0⬘ 兩 ␻ 0⬘ 典 ⫹e ⫺i ␻ ⬘0 兩 ⫺ ␻ 0⬘ 典 ).
As a last ingredient we need to bound the angle ␻ ⬘0 .
Rearranging terms, we obtain These bounds are provided in the following final theorem.

052307-7
SHENVI, KEMPE, AND WHALEY PHYSICAL REVIEW A 67, 052307 共2003兲

Theorem 5. Each application of the evolution operator U ⬘

corresponds to a rotation of angle approx. 1/冑2 n⫺1 in the
basis of the two eigenvectors 兩 ⫾ ␻ 0⬘ 典 . More precisely
⫺1/(c 冑2 n⫺1 )⫺ ␤ ⭐ ␻ ⬘0 ⭐⫺1/(c 冑2 n⫺1 )⫹ ␤ , where ␤
⫽O(n 3/2/2n )
Proof. We will approximate e i ␻ 0⬘ ⫽ 具 ␻ 0⬘ 兩 U ⬘ 兩 ␻ 0⬘ 典 by
具 ␣ 兩 U ⬘ 兩 ␣ 典 , where 兩 ␣ 典 ⫽1/冑2( 兩 ␺ 0 典 ⫹e i ␩ 兩 ␺ 1 典 ). Let us first
evaluate 兩 e i ␻ ⬘0 ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 . We can expand U ⬘ in terms of
its eigenvectors to obtain

冏
兩 e i ␻ ⬘0 ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 ⫽ e i ␻ ⬘0 ⫺ 兺j 冏 具 ␻ ⬘j 兩 ␣ 典 兩 2 e i ␻ ⬘j 兩 . 共50兲

We then note that from Eq. 共42兲,

兩 具 ␻ ⬘0 兩 ␣ 典 兩 2 ⫽ 冑p 0 /2⫹ 冑p 1 /2. 共51兲 FIG. 4. Geometric representation of Theorem 5, which proves

that the eigenvalue, e i ␻ 0⬘ , must be located on a disc of radius
So, n⫺1
n/8c 2 ( n/2 ) centered at 具 ␣ 兩 U ⬘ 兩 ␣ 典 . The position of the eigenvalue is

冏
denoted by a cross.
兩 e i ␻ 0⬘ ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 ⫽ e i ␻ 0⬘ ⫺ 共 冑p 0 /2⫹ 冑p 1 /2兲 e i ␻ 0⬘
Next, we evaluate Im具 ␣ 兩 U ⬘ 兩 ␣ 典 using Eqs. 共31兲 and 共35兲,

冏 Im具 ␣ 兩 U ⬘ 兩 ␣ 典 ⫽Im共 e i ␩ 具 ␺ 0 兩 U ⬘ 兩 ␺ 1 典 ⫺e i ␩ 具 ␺ 1 兩 U ⬘ 兩 ␺ 0 典 兲

冉
⫹ 兺 兩 具 ␻ ⬘j 兩 ␣ 典 兩 2 e i ␻ ⬘j
兩 ␻ ⬘j 典 ⫽ 兩 ␻ 0⬘ 典 1 e i␩
⫽Im ⫺ 共 具 ␺ 0 兩 R,n/2⫺1 典
⭐ 兩 e i ␻ 0⬘ 共 1⫺ 冑p 0 /2⫺ 冑p 1 /2兲 兩 2
c 冑冉 冊2
n⫺1
n/2
兺
冉 冊冊
⫹ 兩 具 ␻ ⬘j 兩 ␣ 典 兩 2
兩 ␻ ⬘j 典 ⫽ 兩 ␻ 0⬘ 典 2
⫹ 具 ␺ 0 兩 L,n/2⫹1 典 兲 ⫺e i ␩ ⫺ 具 ␺ 1 兩 L,1典
⭐2 共 1⫺ 冑p 0 /2⫺ 冑p 1 /2兲
冑2 n

⭐2
冉 2
3n
n⫹1
⫹
8c 2
冉
3n
n⫺1
n/2
冊冊
共52兲
⫽⫺Im
e i␩
c 冑2 n⫺1

with 冑1⫺x⭓1⫺x for 0⭐x⭐1. Using the fact that the bi-
nomial coefficients approach the Gaussian distribution for
large n, such that
⫽⫺
1
c 冑2 n⫺1
⫺O
冉 冑2 n
冉
n
n⫺1
n/2
冊冊
. 共56兲

冉冊 冑n
⫽
2 ⫺(x⫺n/2)/n/2 n
e 2 , 共53兲 Then, using Theorem 4, Eqs. 共54兲 and 共56兲, we can write

冉 冊 冊冏
␲n

冏
x

we can rewrite Eq. 共52兲 taking the leading order terms in n.

Recalling that c⬎1, we obtain
sin ␻ ⬘0 ⫹
1
c 冑2 n⫺1
⫹O
冑2 n
冉
n
n⫺1
⫽O 冉 冊
n 3/2
2n
.

冉 冊
n/2
3/2
i ␻ ⬘0
n 共57兲
兩e ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 ⫽O n
. 共54兲
2 Using sin x⫽x⫹O(x3) and keeping only leading order terms
Equation 共54兲 is an explicit formula which bounds the dis- solving for ␻ ⬘0 gives us

冉 冊 冉 冊
tance in the complex plane between the eigenvalue of inter-
1 n 3/2 1 n 3/2
est, e i ␻ ⬘0 , and the matrix element 具 ␣ 兩 U ⬘ 兩 ␣ 典 . Figure 4 shows ⫺ ⫺O ⭐ ␻ ⬘0 ⭐⫺ ⫹O . 䊏
the geometric representation of Eq. 共54兲. Note that c 冑2 n⫺1
2n c 冑2 n⫺1
2n
共58兲
兩 sin ␻ 0⬘ ⫺Im具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 ⫽ 兩 Im共 e i ␻ 0⬘ ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兲 兩
We can now quantitatively describe the overall operation
⭐ 兩 e i ␻ ⬘0 ⫺ 具 ␣ 兩 U ⬘ 兩 ␣ 典 兩 . 共55兲 of the algorithm. Starting with initial state 兩 ␺ 0 典 , we consider

052307-8
QUANTUM RANDOM-WALK SEARCH ALGORITHM PHYSICAL REVIEW A 67, 052307 共2003兲

the state of the computer after t applications of U ⬘ . Then
using Theorem 4 we can expand 兩 ␺ 0 典 as
兩 ␺ 0 典 ⫽ 冑p 0 共 兩 ␻ 0⬘ 典 ⫹ 兩 ⫺ ␻ 0⬘ 典 )⫹ ␦ 兩 r 典 ,
where ␦ ⫽ 冑1⫺2p 0 ⫽O( 冑n/2n ) and 兩 r 典 is a residual normal-
ized vector orthogonal to 兩 ␻ 0⬘ 典 and 兩 ⫺ ␻ 0⬘ 典 . Now
共 U ⬘ 兲 t 兩 ␺ 0 典 ⫽ 冑p 0 共 e i ␻ 0⬘ t 兩 ␻ ⬘0 典 ⫹e ⫺i ␻ 0⬘ t 兩 ⫺ ␻ 0⬘ 典 )⫹ ␦ 兩 r ⬘ 典
So the probability to measure 兩 xជ target 典 after t f
⫽( ␲ /2)冑2 n⫺1 steps is still p success ⫽1/2⫺O(1/n). Hence,
by repeating the algorithm a constant number of times, the
共59兲 probability of error can be made arbitrarily small. Note the
periodic nature of the evolution under U ⬘ 关Eq. 共60兲兴; this
means that if we measure at t⬎t f the probability of success
will decrease and later increase again.
In summary, we arrived at the final result that the marked
state is identified after O( 冑N) calls to the oracle.

⫽2p 0 cos ␻ 0⬘ t 兩 ␺ 0 典 ⫺2 冑p 0 p 1 共 sin ␻ 0⬘ t IV. CONNECTION TO GROVER’S ALGORITHM

⫹Re e i ␻ ⬘0 t ⌬ 兲 兩 ␺ 1 典 ⫹ 冑1⫺ p 0 ⫺p 1 共 e i ␻ 0⬘ t 兩 r 0 典 The main point of this paper is to give a first algorithm in
the discrete-time random-walk setting. We have shown how
⫹e ⫺i ␻ 0⬘ t 兩 r *
0 典 )⫹ ␦ 兩 r 典 to realize quantum search in this scenario, without losing any

冉 冊
of the quantum speedup obtained in Grover’s search algo-
n 3/4 rithm. Although the layout of our algorithm is very different
⫽cos ␻ 0⬘ t 兩 ␺ 0 典 ⫺sin ␻ ⬘0 t 兩 ␺ 1 典 ⫹O 兩 r̃ 典 ,
冑2 n from Grover’s search, there are several similarities to Grov-
er’s algorithm.
共60兲 Both algorithms begin in the equal superposition state
over all bit strings. Both algorithms make use of the Grover
where 兩 r̃ 典 is some residual normalized vector 共not necessar- diffusion operator G 共sometimes known as the Grover iter-
ily orthogonal to 兩 ␺ 0 典 and 兩 ␺ 1 典 ). ate兲. Both algorithms can be viewed as a rotation in a two-
Starting with 兩 ␺ 0 典 and applying U ⬘ for t f ⫽ ␲ /2兩 ␻ ⬘0 兩 steps, dimensional subspace. Both algorithms use an oracle which
we approximately rotate from 兩 ␺ 0 典 to 兩 ␺ 1 典 . From 兩 ␺ 1 典 we marks the target state with a phase of ⫺1. Both algorithms
can obtain 兩 xជ target 典 ⫽ 兩 0ជ 典 with high probability p using Eq. have a running time of O( 冑N). In both algorithms we have
共33兲 and 1⫹1/(2n)⭐c 2 ⭐1⫹2/n for large n 关following from to measure at a specific time to obtain maximum probability
Eq. 共34兲兴, of success. However, there are several important differences
between the two search algorithms. In this section, we call
attention to the ways in which the random-walk search algo-
p⫽ 兺d 兩 具 d,0兩 ␺ 1 典 兩 2 rithm is distinct from Grover’s algorithm, and consider how
these differences affect performance and implementation.
⫽ 兩 具 R,0兩 ␺ 1 典 兩 2 It is well known that Grover’s algorithm can be mapped
exactly onto a rotation in the two-dimensional subspace
1/2 1/2 spanned by the equal-superposition state 兩 ␺ 0 典 and the
⫽ ⭓ marked state 兩 0 典 关7兴. Each iteration in Grover’s algorithm
c 2 1⫹2/n
corresponds to a rotation in this subspace. In this paper, we
1 have shown that the random-walk search algorithm can also
⫽ ⫺O 共 1/n 兲 . 共61兲 be viewed as a rotation in a two-dimensional subspace. How-
2
ever, there are two important distinctions. First, the random
Finally, to obtain t f in terms of n, we make use of the bounds walk search algorithm can only be approximately mapped
on ␻ 0⬘ provided by Theorem 5, onto a two-dimensional subspace. Unlike Grover’s algo-
rithm, this mapping is not exact. Second, the two-

冋 冉 冊册
dimensional subspace in which the random-walk search al-
␲ c n⫺1 n 3/2
gorithm is approximately contained is spanned by 兩 ␺ 0 典 and
tf⫽ 冑2 1⫾O . 共62兲
2 冑2 n 兩 ␺ 1 典 , not by 兩 ␺ 0 典 and 兩 0 典 . Hence, the final state of the algo-
rithm is not exactly the pure marked state 兩 0 典 as it is in
Using the inequality 1⫹1/(2n)⭐c 2 ⭐1⫹2/n to get 1 Grover’s algorithm. It is a linear combination of states which
⫹1/(4n)⭐c⭐1⫹1/n, we obtain is composed primarily of the marked state, but also possesses
small contributions from its nearest neighbors, second-
tf⫽
␲ n⫺1
2
冑2 1⫹O
1
n 冋 冉 冊册 . 共63兲
nearest neighbors, etc. Thus, the random-walk search algo-
rithm contains traces of the underlying topology of the hy-
percube on which it is based.
If we set the number of time steps to be t f ⫽( ␲ /2)冑2 n⫺1 共or Another crucial difference is the locality of the unitary
the closest integer兲 then transformations used during the algorithm. In the random-
walk search algorithm the shift operator is local in the topol-
␲
⫺sin ␻ ⬘0 t f ⫽sin 1⫺O
2
1
n 冋 冉 冊册 ⫽1⫺O 冉冊1
n2
. 共64兲
ogy of the hypercube, i.e. it shifts amplitude only between
the n nearest neighbors. The coin operator shifts amplitudes
only on the n-dimensional coin space. So we can say that all

052307-9
SHENVI, KEMPE, AND WHALEY
our operations in an iteration are n local. Compared to this
the reflection operator used in Grover’s algorithm is highly
nonlocal.
Another difference between the two algorithms is their
use of the Grover diffusion operator G. In Grover’s algo-
rithm, this operator is applied to the entire 2 n -dimensional
search space 共corresponding to the node space in the
random-walk search algorithm兲. On the other hand, Grover’s
diffusion operator G in the random-walk algorithm is used as
the quantum coin, and acts only on the n-dimensional coin
space. This fact may be of practical use for certain physical
implementations since many physical implementations of
quantum computers contain multiple types of qubits, which
have different natural gate sets. We could exploit this variety
using the random-walk search algorithm by choosing the
coin space to be represented by qubits on which it is conve-
nient to implement the Grover diffusion operator. Similarly,
it might be natural and easy for some physical systems to
implement the shift operator rather than the gates required in
Grover’s search algorithm. It is ultimately the physical sys-
tem that will determine which of the search algorithms is
more advantageous.
Another similarity between the two algorithms is the
implementation of the oracle. In Grover’s algorithm, the
oracle marks the target state with a phase of ⫺1. To arrive at
this random walk search algorithm, we chose the marking
coin C 1 to be the ⫺I coin. This choice was actually moti-
vated because it yielded a result that was amenable to analy-
sis, and while the emergence of Grover’s algorithm appears
natural in hindsight, it was not obvious at the outset. How-
ever, more generally, it is not clear whether this choice of
marked coin is either optimal or unique. In fact, numerical
simulations have shown us that many different types of
marking coins will yield search algorithms 关15兴. Unfortu-
nately, analytic treatment of the quantum random walk for
more complicated coins has proven substantially more diffi-
cult than the instance analyzed here for C 1 ⫽⫺I. It is an
open question what 共constant factor兲 gains might be made by
using different marking coins to implement the search.
V. CONCLUSIONS
In this paper, we have shown that the random-walk search
algorithm can search a list of 2 n items in time proportional to
冑2 n . The lower bound on a quantum search of an N-item list
is known to be ⍀( 冑N) 关9兴. Thus, up to a constant factor, the
random-walk search algorithm is optimal. However, al-
关1兴 Y. Aharonov, L. Davidovich, and N. Zagury, Phys. Rev. A 48,
1687 共1993兲.
关2兴 D. Aharonov, A. Ambainis, J. Kempe, and U. Vazirani, in Pro-
ceedings of the 33rd Annual ACM Symposium on the Theory of
Computation (STOC) 共ACM, New York, 2001兲, pp. 50–59;
e-print quant-ph/0012090.
关3兴 A. Ambainis, E. Back, A. Nayak, A. Vishwanath, and J. Wa-
trous, in Proceedings of the 33rd Annual ACM Symposium on
the Theory of Computation (STOC) 共ACM, New York, 2001兲,
pp. 60– 69.
052307-10
PHYSICAL REVIEW A 67, 052307 共2003兲
though after repetition of the algorithm a constant number of
times the result is arbitrarily close to the result of Grover’s
search, the random-walk search algorithm is not identically
equivalent to Grover’s algorithm. In particular, the final so-
lution obtained by the random-walk search still retains some
of the underlying character of the hypercube on which it was
based, with a small admixture of states other than the solu-
tion at the marked node.
The random walk search analyzed here was based on a
discrete walk on the hypercube. In general, a similar meth-
odology can be applied to any regular graph, e.g., a two-
dimensional hexagonal lattice with periodic boundary condi-
tions, a three-dimensional rectangular lattice with periodic
boundary conditions, etc. We have numerical evidence indi-
cating that this methodology will yield quantum search algo-
rithms when applied to other regular n-dimensional lattices.
Future studies will investigate the extent of optimality of
such search algorithms.
The intriguing possibility of finding novel algorithms
based on the random walk also remains an open question.
The results described here indicate that the random-walk
search algorithm provides a suggestive framework for new
algorithms. Though the optimality of Grover’s algorithm pre-
cludes the construction of an improved oracle-based search
algorithm based on a quantum walk, nevertheless, many
other oracle problems still exist for which a quantum walk
may be advantageous. For instance, the lower bound on
quantum search holds only for oracles that provide ‘‘yes/no’’
information 关9兴. Our choice of marking coin here has a clear
relation to an identifiable component of Grover’s algorithm.
In general, the marking coin can be an arbitrary n⫻n unitary
matrix. The marking coin provides a intuitive means by
which to introduce a large amount of information to an
oracle problem. Thus, it is possible that unique coins with
interesting properties may give rise to an entirely new algo-
rithm. Overall we conclude that the quantum random walk
provides a means for insight into existing quantum algo-
rithms and offers a potentially vast source for development
of new algorithms.
ACKNOWLEDGMENTS
N.S. thanks the University of California, for a Berkeleyan
Fellowship. This effort is sponsored by the Defense Ad-
vanced Research Projects Agency 共DARPA兲 and the Air
Force Laboratory, Air Force Material Command, USAF, un-
der Contract No. F30602-01-2-0524. We also thank NSF
ITR/SY Grant No. 0121555.
关4兴 E. Farhi and S. Gutmann, Phys. Rev. A 58, 915 共1998兲.
关5兴 A. Childs, E. Farhi, and S. Gutmann, Quantum Inform. Pro-
cess. 1, 35 共2002兲.
关6兴 A. Childs, R. Cleve, E. Deotto, E. Farhi, S. Gutmann, and D.
Spielman, e-print quant-ph/0209131.
关7兴 L. Grover, in Proceedings of the 28th Annual ACM Symposium
on the Theory of Computation 共ACM Press, New York, 1996兲,
pp. 212–219.
关8兴 L. Grover, Phys. Rev. Lett. 79, 325 共1997兲.
QUANTUM RANDOM-WALK SEARCH ALGORITHM PHYSICAL REVIEW A 67, 052307 共2003兲

关9兴 C. Bennett, E. Bernstein, G. Brassard, and U. Vazirani, SIAM
J. Comput. 26, 1510 共1997兲.
关10兴 M. Nielsen and I. Chuang, Quantum Computation and Quan-
tum Information 共Cambridge University Press, Cambridge,
2000兲.
关11兴 C. Moore and A. Russell, in Proceedings of RANDOM, 2002,
edited by J. D. P. Rolim and P. Vadham 共Springer, Cambridge,
MA, 2002兲, pp. 164 –178.
关12兴 J. Kempe, e-print quant-ph/0205083.
关13兴 T. Yamasaki, H. Kobayashi, and H. Imai, in Proceedings of the
3rd UMC, edited by C. Calude, M. J. Dinneen, and F. Peper
共Springer, Kobe, 2002兲, pp. 315–330.
关14兴 J. Watrous, J. Comput. Syst. Sci. 62, 376 共2001兲.
关15兴 N. Shenvi 共unpublished兲.

052307-11