See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220133694

Grover's Quantum Algorithm Applied to Global Optimization

Article  in  SIAM Journal on Optimization · January 2005

DOI: 10.1137/040605072 · Source: DBLP

CITATIONS READS
40 699

3 authors, including:

Graham Wood
The University of Warwick
130 PUBLICATIONS   1,173 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

C D Lai View project

Cotranslational protein folding View project

All content following this page was uploaded by Graham Wood on 05 June 2014.

The user has requested enhancement of the downloaded file.

SIAM J. OPTIM. 
c 2005 Society for Industrial and Applied Mathematics
Vol. 15, No. 4, pp. 1170–1184

GROVER’S QUANTUM ALGORITHM APPLIED TO GLOBAL

OPTIMIZATION∗
W. P. BARITOMPA† , D. W. BULGER‡ , AND G. R. WOOD‡

Abstract. Grover’s quantum computational search procedure can provide the basis for imple-
menting adaptive global optimization algorithms. A brief overview of the procedure is given and a
framework called Grover adaptive search is set up. A method of Dürr and Høyer and one introduced
by the authors fit into this framework and are compared.

Key words. discrete optimization, global optimization, Grover iterations, Markov chains, quan-
tum computers, random search

AMS subject classifications. 90C30, 68Q99, 68Q25

DOI. 10.1137/040605072

1. Introduction. This paper aims to provide the global optimization commu-

nity with some background knowledge of quantum computation and to explore the
importance of this topic for the future of global optimization.
Quantum computing [7] holds great potential to increase the efficiency of stochas-
tic global optimization methods. Current estimates are that quantum computers are
likely to be in commercial production within two or three decades. These devices
will be in many respects similar to the computers of today but will utilize circuitry
capable of quantum coherence [7], enabling data to be manipulated in entirely new
ways.
Grover introduced in [9] a quantum algorithm (that is, an algorithm to be ex-
ecuted on a quantum computer) for locating a “marked” item in a database. This
was extended in [2] to a quantum algorithm for locating one of an unknown number
of marked items. The latter method was incorporated into a minimization algorithm
by Dürr and Høyer in [8] (unpublished, but available electronically—see the reference
list).
Dürr and Høyer’s algorithm can be viewed as an example of Grover adaptive
search (GAS), an algorithmic framework we introduced in [4]. GAS in turn is a
quantum-computational implementation of hesitant adaptive search [6], a parameter-
ized pseudoalgorithm whose performance is well understood. Here we analyze Dürr
and Høyer’s method, present another version of GAS, and explore the relative merits
of the two methods via numerical simulation.
Outline. Section 2 presents the general optimization problem and introduces
some notation and terminology. Section 3 gives a brief overview of quantum compu-
tation and Grover’s algorithm. Section 4 describes the GAS framework. Section 5
discusses the considerations involved in choosing the “rotation count sequence,” the
parameter distinguishing one GAS algorithm from another. Section 6 presents Dürr

∗ Received by the editors March 11, 2004; accepted for publication (in revised form) November 22,

2004; published electronically August 3, 2005. This research was supported by the Marsden Fund of
the Royal Society of New Zealand.
http://www.siam.org/journals/siopt/15-4/60507.html
† Department of Mathematics and Statistics, University of Canterbury, Christchurch, New Zealand

(b.baritompa@math.canterbury.ac.nz).
‡ Department of Statistics, Macquarie University, NSW 2109, Australia (dbulger@efs.mq.edu.au,

gwood@efs.mq.edu.au).
1170
 APPLYING GROVER ADAPTIVE SEARCH 1171

and Høyer’s algorithm, extending and correcting the theoretical analysis in [8]. In
section 7, we present a refined version of GAS, and in section 8 this version is com-
pared to that of Dürr and Høyer by numerical simulation. Section 9 concludes the
paper.
2. Optimization problem. We consider the following finite global optimization
problem:

minimize f (x)
subject to x ∈ S,

where f is a real-valued function on a finite set S.

Throughout this paper we associate with the objective function f the following
definitions. Let N = |S|, the cardinality of the finite set S. We will usually assume N
to be a power of two. Let 1 < · · · < K be the distinct objective function values in
the range of f . Notice that there may be more than K points in S. Given the uniform
probability measure μ on S, we let π be the range measure given by the stochastic
vector (π1 , . . . , πK ) induced by f . That is, πj = |f −1 (j )|/N for j = 1, 2, . . . , K. Let
j
pj denote i=1 πi , the probability that a random point has value of j or less. In
particular, pK = 1. Corresponding to each function value is an improving region, that
part of the domain having a strictly better value, and we call its measure under μ the
improving fraction p. (Usually the specified function value will be the best yet seen,
and thus the improving region will be the set of points with objective function values
better than any yet seen.)
3. Quantum computing. This paper concerns optimization algorithms that
require the use of a quantum computer. The characteristic feature of a quantum
computer is that, in place of conventional computer bits, quantum bits, or qubits,
are used. A qubit can be in a simultaneous superposition of “off” and “on” and
thus allows quantum parallelism, where a single quantum circuit can simultaneously
perform a calculation on a superposed input, corresponding to very many conventional
inputs.
The Grover mechanism. The quantum procedure germane to our purposes is
Grover search [9]. This is one of the major advances to date in the fledgling field
of quantum computation. More details are given in [4], but we reiterate the salient
features here and give an intuitive discussion.
Consider the following general search problem. Let n be a positive integer, and
let S = {0, 1}n , so that the domain size N = 2n . Let h : S → {0, 1}. We wish to find
a point u ∈ S such that h(u) = 1. We further assume that h is a black-box, that is,
that knowledge of h can only be gained by sampling (evaluation), but no structural
information is available.
With conventional computing, the Boolean function h could be implemented as a
subroutine, i.e., a conventional logic circuit constructed to take an input string of bits,
representing a point of S, and output the associated bit value of h. The subroutine
could then be applied to all points of S, in succession, to find a required point. Such
a conventional program would require on average N/2 evaluations to find a marked
point.
In quantum computing, the circuit implementing h (using gates that work with
qubits) inputs and outputs superpositions. Thus it “sees” many possible answers at
once. On a quantum computer, observing the output will collapse it into a conven-
1172 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

tional bit string, according to a probability distribution determined by the superpo-

sition; thus quantum computing has a stochastic side. Rather than loop through the
N points in S, a quantum computer can operate on superposed states in such a way
that the probability distribution governing the collapse
√ can be changed. Grover in [9]
showed if exactly one point is marked, then only π4 N such operations are required
to find the marked point.
Denote the set of marked points by M = {u ∈ S|h(u) = 1} and denote the number
of these marked target points by t. We may or may not be aware of the value of t.
Let p be the proportion of marked points, t/N .
Grover introduced the Grover rotation operator, which incorporates the oracle for
h and provides a means of implementing a certain phase-space rotation of the states
of a quantum system encoding points in the domain S. Repeated applications of
this rotation can be used to move from the equal amplitude state, which is simple to
prepare within a quantum computer, toward the states encoding the unknown marked
points. For details see [4, 2, 9].
A Grover search of r rotations applies the above rotation operator r times, start-
ing from the equal amplitude superposition of states, and then observes (and hence
collapses to a point) the output state. The mathematical details in [2] show that exe-
cuting such a search of r rotations generates domain points according to the following
probability distribution γ on S:
⎧
⎪ g (p)
⎨ r , x ∈ M,
(3.1) γ({x}) = t
⎩ 1 − gr (p) , x ∈ S\M,
⎪
N −t
where
√
(3.2) gr (p) = sin2 [(2r + 1) arcsin p] .

Note that in the special case of r = 0, Grover search observes only the prepared equal
amplitude superposition of states and so reduces to choosing a point uniformly from
the domain.
Most of the work in implementing the Grover rotation operator is in the oracle
query, so the cost of a Grover search of r rotations is taken as the cost of r oracle
queries. The output is a point in S, and as one would usually want to know if it is in
M or not, a further oracle query (acting on the point) would give the function value
under h.
Grover search is sometimes portrayed as a method for the database table lookup
problem. This is only one elementary application, however. Other interesting appli-
cations concern “marking functions” h which are more than simple tests of indexed
data. Examples relating to data encryption and the satisfiability problem are given
in [2, 9].
From searching to optimizing. Grover search solves a special global opti-
mization problem: it finds a global maximum of h. For the more general problem
introduced in section 2, our intention is to use Grover search repeatedly within a
global optimization method of the adaptive search variety. Adaptive search methods
produce, or attempt to produce, an improving sequence of samples, each uniformly
distributed in the improving region of the previous sample (see [16, 15, 5]).
Given an objective function f : S → R and a point X ∈ S with f (X) = Y , we
use Grover’s algorithm to seek a point in the improving region {w ∈ S : f (w) < Y }.
 APPLYING GROVER ADAPTIVE SEARCH 1173

As described above, Grover’s algorithm requires an oracle, a quantum circuit able to

classify a point w ∈ S as inside or outside the target set (see [10]). This will be the
oracle for the Boolean function h(w) = (f (w) < Y ).
We denote by a “boxed” name the oracle for a given function. Symbolically h
is found as shown:

w −→ f

< −→ 0 or 1.

y

The additional comparison logic circuitry < to construct h is minimal, and we will
take the cost of h and f to be the same.
As far as Grover’s algorithm is concerned, h is simply a black box quantum
circuit, inputting a point w in S (or a superposition of such points) and outputting

1, f (w) < y,
0, f (w) ≥ y

(or the appropriate superposition of such bits).

Grover search of r rotations, using the compound oracle depicted above, will
require r uses of the objective function suboracle f and will output a random domain
point. An additional oracle query is required to determine whether the output is an
improvement or not. Therefore, for practical purposes, we can consider the cost of
running Grover’s algorithm to be r + 1 objective function evaluations (plus additional
costs, such as the cost of the comparisons, which we will ignore).
As a point of departure for the mathematics to follow, we can condense this
subsection into the following axiom, and henceforth dispense with any direct consid-
eration of quantum engineering. (Note that the content of this axiom is taken for
granted in [2] and many other recent publications on quantum searching.)
Axiom 1. Given f : S → R and Y ∈ R, there is a search procedure on a quantum
computer, which we shall call a “Grover search of r rotations on f with threshold Y ,”
outputting a random point x ∈ S distributed uniformly in

{w ∈ S : f (w) < Y } with probability gr (p), or uniformly in

{w ∈ S : f (w) ≥ Y } otherwise,

where p = |{w ∈ S : f (w) < Y }|/|S|. The procedure also outputs y = f (x). The
cost of the procedure is r + 1 objective function evaluations.

4. Grover adaptive search. This section presents the GAS algorithm intro-
duced in [4]. The algorithm requires as a parameter a sequence (rn : n = 1, 2, . . . ) of
rotation counts. Initially, the algorithm chooses a sample uniformly from the domain
and evaluates the objective function at that point. At each subsequent iteration, the
algorithm samples the objective function at a point determined by a Grover search.
The Grover search uses the best function value yet seen as a threshold. Here is the
algorithm in pseudocode form:
1174 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

Grover Adaptive Search (GAS).

1. Generate X1 uniformly in S, and set Y1 = f (X1 ).
2. For n = 1, 2, . . . , until a termination condition is met, do:
(a) Perform a Grover search of rn rotations on f with threshold Yn ,
and denote the outputs by x and y.
(b) If y < Yn , set Xn+1 = x and Yn+1 = y; otherwise, set Xn+1 = Xn
and Yn+1 = Yn .
GAS fits into the adaptive search framework developed in [5, 6, 15, 16, 17] which
has proved useful for theoretical studies of convergence of stochastic global optimiza-
tion methods. All adaptive algorithms assume “improving” points can be found (at
some cost). If Grover’s algorithm were only applicable to database lookup, one might
get the impression that GAS would require all function values to be first computed
and tabled before they could then be marked. However, Grover’s algorithm can find
points in an unknown target set, specified by an oracle. GAS exploits this ability
by constructing, at each iteration, an oracle targeting the current improving region.
In this way, it builds a sequence of domain points, each uniformly distributed in the
improving region of the previous point. Such a sequence is known to converge to the
global optimum very quickly; for instance, a unique optimum in a domain of size N
will be found after 1 + ln N such improvements, in expectation (see [17]).
Unfortunately this does not mean that GAS can find the global optimum for a
cost in proportion to ln N . The reason is that as the improving fraction p decreases,
larger rotation counts become necessary to make improvements probable; thus the
cost of GAS varies superlinearly in the number of improvements required. Note also
that not every iteration finds a point in the improving region. The probability of
finding an improvement is given by (3.2), and for a known p  1, a rotation count r
can be found making this probability very nearly 1. But since in general we can only
guess at p, lower probabilities result.
Readers may wonder why we use the best value yet seen as the threshold in the
Grover search. In a sense, all of the work of the algorithm is done in the last step,
when a Grover search is performed using a threshold only a little larger than the global
minimum. This final Grover search is not made any easier by the information gained
in earlier steps. In the general case, however, where we have no prior knowledge of
the objective function’s range, these earlier steps are an efficient way of finding a good
value to use as a threshold in the final step. The earlier steps are not great in number.
Moreover, the cost of each step is roughly inversely proportional to the square root
of the improving fraction; thus, if the sequence of rotation counts is chosen suitably,
most of the earlier steps will be much quicker than the final one.

5. Choosing the rotation count sequence. This section provides a general

discussion of the selection of the rotation count sequence used in the GAS algorithm
as a precursor to sections 6 and 7, each of which presents a specific selection method.

Why the rotation count should vary. In [4] we considered the possibility
of using the same rotation count at each iteration. Although it is easy to construct
objective functions for which this method works well, they are exceptional, and in
general it is preferable to vary the rotation count as the algorithm progresses.
To see why, suppose that at a certain point in the execution of the GAS algorithm,
the best value seen so far is Y , and the improving fraction is p = |{w : f (w) < Y }|/N .
For any given rotation count r, the probability of success of each single iteration of the
algorithm is given by gr (p). Although the rationale for using Grover’s algorithm is to
 APPLYING GROVER ADAPTIVE SEARCH 1175

Fig. 1. The probability of a step of three Grover rotations finding an improvement, as a function
of the improving fraction p.

increase the probability of finding improving points, there are combinations of values
of r and p where the opposite effect occurs. For instance, Figure 1 plots g3 (p) versus
p. If p = 0.2, then the step is almost guaranteed not to find an improvement. If the
rotation count varies from each iteration to the next, then this is only an occasional
nuisance. But if it is fixed at r, and if the algorithm should happen to sample a point
x such that the improving fraction p for Y = f (x) has gr (p) zero or very small, then
the algorithm will become trapped.
How the rotation count should vary. In fact, at each iteration during the
execution of the algorithm, some optimal rotation count r is associated with the
improving fraction p of the domain (assuming p > 0). If it is used for the next Grover
search, then an improving point will almost certainly be found. This r is the first
positive solution to the equation gr (p) = 1. (Actually of course we must round this to
the nearest integer, and therefore success is not absolutely guaranteed, but this would
contribute little to the expected cost of the algorithm.)
Unfortunately, in the general case the improving fraction p is unknown, so we
are somewhat in the dark as to the choice of rotation counts. In order to make
the most use of all the information available to us at each iteration, we could take
a Bayesian approach and keep track of a sequence of posterior distributions of the
improving fraction at each iteration and choose each rotation count to optimize the
change in some statistic of this posterior distribution. As might be expected, this
kind of approach appears to be very complex and unwieldy. The methods outlined in
the following two sections, however, strike a happy balance between implementability
and optimality of rotation count selection.
6. Dürr and Høyer’s random method. In this section we outline a method
due to Dürr and Høyer for randomly choosing rotation counts and correct two key
arguments in its originators’ analysis.
Grover’s search algorithm provides a method of finding a point within a subset
1176 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

of a domain. If the size of the target subset is known, the algorithm’s rotation count
parameter can easily be tuned to give a negligible failure probability. The case of a
target subset of unknown size is considered in [2], where the following algorithm is
presented:
Boyer et al. search algorithm.
1. Initialize m = 1.
2. Choose a value for the parameter λ (8/7 is suggested in [2]).
3. Repeat:
(a) Choose an integer j uniformly at random such that 0 ≤ j < m.
(b) Apply Grover’s algorithm with j rotations, giving outcome i.
(c) If i is a target point, terminate.
(d) Set m = λm.
√
Actually, in [2],√ the final step updates m to min{λm, N }. It is pointless to
allow m to exceed N , because for a target set √ of any size, it is known [2] that the
optimal rotation count will be no more than
π N /4. In the global optimization
context, however, this point will usually be immaterial, since the target region, though
comprising a small proportion of the domain, will normally be large in absolute terms.
For instance, suppose the domain contains 1020 elements and suppose finding one
of the smallest 10,000 points is required. The optimal√rotation count to find a target
set of this size is 108 π/4, substantially less than
π N /4. The actual target size
will be unknown, and therefore the actual optimal rotation count will be unknown.
But when m reaches this magnitude, if not before, each step will have a substantial
probability (on the order of 1/2) of finding a target point.√Therefore, unless λ is very
large, there will be negligible probability of m reaching N = 1010 before √ a target
point is produced. For simplicity, therefore, in this article we ignore the N ceiling
on the growth of m.
In the quant-ph internet archive, Dürr and Høyer [8] propose using the Boyer
et al. algorithm as the nucleus of a minimization algorithm. Their paper gives the
impression that the algorithm is just for the database problem. They begin with “an
unsorted table of N items each holding a value from an ordered set. The minimum
searching problem is to find the index y such that T [y] is minimum.” Again we
stress their algorithm fits in the GAS framework and is thus applicable to the general
optimization problem.
In their paper, they indicate that every item that is improving is explicitly marked.
However, this is a mistake, as it is incompatible with their complexity analysis later
in the paper. We describe a corrected version of their method using the terminology
of this paper.
Dürr and Høyer’s algorithm.
1. Generate X1 uniformly in S, and set Y1 = f (X1 ).
2. Set m = 1.
3. Choose a value for the parameter λ (as in the previous algorithm).
4. For n = 1, 2, . . . until a termination condition is met, do:
(a) Choose a random rotation count rn uniformly distributed
on {0, . . . ,
m − 1}.
(b) Perform a Grover search of rn rotations on f with threshold Yn ,
and denote the outputs by x and y.
(c) If y < Yn , set Xn+1 = x, Yn+1 = y, and m = 1; otherwise, set
Xn+1 = Xn , Yn+1 = Yn , and m = λm.
This is the special case of GAS arising when the rotation count rn is chosen
 APPLYING GROVER ADAPTIVE SEARCH 1177

randomly from an integer interval which is initialized√to {0} at each improvement but
which grows exponentially to a maximum of {0, . . . ,
N −1} between improvements.
The analysis of the algorithm reported in the archive [8] uses incorrect constants
from a preprint of [2]. In the following analysis, we correct this by using the published
version of [2]. Because the Boyer et al. algorithm underpins that of Dürr and Høyer,
we begin with an analysis of the former algorithm. Theorem 3 in [2] is an order of
magnitude result, but inspection of the proof implies that the expected time required
by the Boyer et al. algorithm to find one of t marked items among a total of N items
is bounded by 8 N/t. This constant can be improved upon, though, as we shall see
after the following theorem.
Theorem 6.1. The expected number of oracle queries required by the Boyer et
al. algorithm with parameter λ to find and verify a point from a target subset of size
t from a domain of size N is

∞
j−1  

λj 
1 sin(4θ
λi )
(6.1) + ,
j=0
2 i=0
2 4
λi  sin(2θ)


where θ = arcsin( t/N ).
Proof. Conditioned on reaching iteration j, the expected number of oracle queries
required at that iteration is
λj /2 (including the test of the output of Grover’s
algorithm for target subset membership). The probability of reaching iteration j is
a product of failure rates; the probability of the algorithm failing to terminate at
iteration j, having reached this iteration, is

1 sin(4θ
λi )
+
2 4
λi  sin(2θ)

(this is [2, Lemma 2]). Thus the expected number of oracle queries required at itera-
tion j, not conditioned on whether the iteration is reached, is
j−1  

λj 
1 sin(4θ
λi )
+ ,
2 i=0 2 4
λi  sin(2θ)

and summing over all possible iterations j = 0 · · · ∞ gives the result.

It is straightforward to evaluate the
geometrically convergent series (6.1) numer-
ically. By graphing the ratio of (6.1) to N/t versus t for a range of λ, empirically λ
that gave the lowest maximum is 1.34. The plot of Figure 2 uses this value of λ, and
it justifies the following observation.
Observation 1. The expected number of oracle queries required by the Boyer et
al. algorithm with parameter λ = 1.34 to find and verify a point from a target subset
of size t from a domain of size N is at most 1.32 N/t.
Now we can derive a performance bound for Dürr and Høyer’s algorithm. This
is similar to and extends the result in [8]; the main difference is in our treatment of
the coefficient of the order bound. Also we correct another technical error in their
argument, which is pointed out in our proof below.
Theorem 6.2. Assume the validity of the above observation. Let 1 ≤ s ≤ N
and assume that there are s points in the domain with strictly better objective function
values than the remaining N −s points. The expected number of oracle queries required
by Dürr and Høyer’s algorithm with λ = 1.34 to find one of these s points is bounded
1178 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

 Fig. 2. The ratio between the partial sums of the geometrically convergent series (6.1) and
N/t when λ = 1.34, plotted against t/N . Note that 1.32 appears to be an upper bound.

above by
√
N
1
1.32 N √ .
r=s+1
r r−1

Note that if s is small compared to N , then the above bound approximately equals

2.46 N/s.
Proof. Assign the domain points ranks from 1 to N , giving the best point rank
1 and so forth. Where several points have equal objective function value, break ties
arbitrarily, but let l(r) be the least rank and h(r) the greatest rank among the points
with the same value as the rank r point. (In the distinct values case we will have
l(r) = h(r) = r for each r ∈ {1, . . . , N }.)
Since Dürr and Høyer’s algorithm will move through a succession of threshold
values with rank above s before finding the desired target point, the bound on the
expectation in question is given by

N
(6.2) p(N, r)B(N, l(r) − 1),
r=s+1
where p(N, r) is the probability of the rank r point ever being chosen and B(N, l(r)−1)
is the expected number of iterations required by the Boyer et al. algorithm to find
and verify a point from a target subset of size l(r) − 1.
 APPLYING GROVER ADAPTIVE SEARCH 1179

The probability p(N, r) = 1/h(r). This is demonstrated in the proof of The-

orem1 in [17] and in Lemma 1 of [8]. Also, by Observation 1, B(N, l(r) − 1) ≤
1.32 N/(l(r) − 1).
In the distinct values case, substitution of the above value for p(N, r) and bound
for B(N, l(r) − 1) = B(N, r − 1) into (6.2) gives the theorem immediately. In [8] it
is claimed for the case of repeated objective function values that since the equation
p(N, r) = 1/r becomes the inequality p(N, r) ≤ 1/r, the bound still holds. This
argument ignores that the value of B(N, l(r) − 1) increases (for a given r) when
repeated values are allowed. Nevertheless, the theorem holds as follows. Consider
r̂ ∈ {1, . . . , N } with l(r̂) < h(r̂). We examine just that part of the summation in (6.2)
with the index going from l(r̂) to h(r̂),

h(r̂)

√ h(r̂) 1
p(N, r)B(N, l(r) − 1) ≤ 1.32 N 
r=l(r̂) r=l(r̂)
h(r) l(r) − 1


h(r̂)
1
= 1.32 N (l(r̂) − 1)
h(r̂)(l(r̂) − 1)
r=l(r̂)


h(r̂)
1
= 1.32 N (l(r̂) − 1)
r(r − 1)
r=l(r̂)

√ h(r̂) 1
≤ 1.32 N √ .
r=l(r̂)
r r−1

Remark 1. Dürr and Høyer’s method can be viewed as an implementation of pure

adaptive search [17], requiring no more than 1.32(N/t)1/2 iterations in expectation to
find an improvement, when t is the cardinality of the improving region.

7. A new method. In this section we propose an explicit sequence of integers

to be used as the GAS rotation count sequence. This gives a special case of GAS that
can be identified with an inhomogeneous Markov chain having states 1 , . . . , K .
For this paper we have sought an efficient choice for the rotation count sequence
used in GAS. This has led us to the special case of GAS arising when the sequence
(rn ) is fixed in advanced and determined by the following pseudocode. Note that the
sequence of rotation counts it produces is independent of the particular optimization
task; its first 33 entries are

0, 0, 0, 1, 1, 0, 1, 1, 2, 1, 2, 3, 1, 4, 5, 1, 6, 2, 7, 9,
(7.1)
11, 13, 16, 5, 20, 24, 28, 34, 2, 41, 49, 4, 60, . . . .

Here is the pseudocode:

1180 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

Rotation schedule construction algorithm.

1. Initialize u to be the polynomial u(y) = y.
2. For i = 1, 2, . . . , do:
1
(a) Set Eu = 1 − 0 u dy.
(b) Set b = 0.
(c) For r = 0, 1, . . . , until Eu /(r + 1) ≤ 2b , do:
1
i. Set v = u + y y (gr (t)/t)du(t).
1
ii. Set Ev = 1 − 0 v dy.
iii. Set b = (Eu − Ev )/(r + 1).
iv. If b > b , then,
A. Set r = r.
B. Set b = b.
C. Set v  = v.
(d) Set u = v  .
(e) Output ith rotation count r .
The resulting sequence (7.1) is heuristically chosen to maximize a benefit-to-cost
ratio, denoted b in the pseudocode, at each GAS iteration. The reader can verify that
u and Eu are the cumulative distribution function and expectation, respectively, of the
improving fraction of the domain, after the first i − 1 iterations of the GAS algorithm.
The symbols v and Ev denote the corresponding cumulative distribution function
and expectation after a further GAS step of r rotations. The benefit is (somewhat
arbitrarily) taken to be the expected decrease in the improving fraction of the domain,
Eu − Ev . The cost is r + 1, where r is the number of rotations chosen, as per Axiom 1.
The inner loop at (2c) terminates since even if gr were identically one, the expected
improving region measure would halve. Thus, higher rotation counts need not be
considered once we pass the point where half the expected improving region measure,
divided by the cost, exceeds the current best found benefit-to-cost ratio.
8. Computational results. In section 6 we presented a corrected version of
Dürr and Høyer’s demonstration
 of a performance bound for their algorithm. This
readily establishes the O( N/s) complexity, inherited from Grover’s algorithm. How-
ever, even the improved coefficient of 2.46 suggested by Theorem 6.2 is based on an
upper bound and may be a poor indicator of the algorithm’s actual performance. In
this section we study the methods described in sections 6 and 7 using numerical sim-
ulation. Our aim is twofold: to tune the parameter λ appearing in Dürr and Høyer’s
algorithm and then to compare their tuned method against the method of section 7.
Our simulations will determine the length of time each algorithm requires to
sample a point in a target region, constituting a certain proportion of the domain.
Intuitively, the algorithm terminates upon finding a value equal to or lower than the
quantile determined by a proportion α.
Recall that the proportion of the domain with value lower than or equal to j
is pj . More precisely, we specify an intended quantile proportion αnominal and set
k = min{j : pj ≥ αnominal }. We require the algorithm to find a point with value less
than or equal to k . The target set is f −1 ({1 , 2 , . . . , k }). Let s be its cardinality.
So α = pk = s/N and gives the quantile the algorithm will find. Note that it is the
measure under π of {1 , 2 , . . . , k }. It may be inevitable that α and αnominal differ
since it is possible that pk−1 < αnominal < α = pk .
Thus the quantity α is often unknown in practice and is a “global” piece of in-
formation. The dependence of performance on global information is unavoidable [14],
but we will see that for certain methods, the dependence is primarily on α. For the
 APPLYING GROVER ADAPTIVE SEARCH 1181

Fig. 3. Performance graphs for Dürr and Høyer’s algorithm for various values of the parameter
λ and two domain sizes. The third graph repeats the second with a finer mesh of λ values.

rest of this paper we assume α is close to αnominal .

Methodology. For the performance of either algorithm under consideration, the
distribution of objective function values influences performance only via the range
measure π. Our primary focus here will be the case where π is uniformly distributed
over a finite set of distinct function values. Without loss of generality we can take this
finite set to be {1, . . . , K}. For example, to explore seeking the best 1% of the domain
under a uniform range distribution (i.e., αnominal = 0.01), using K = 100 will be fairly
representative. At the end of this section we look briefly at other distributions.
To compare the algorithms, we plot their performance graphs [11] which relate
practical computational effort to the probability of finding a point in the target set.
The performance graph is simply the cumulative distribution function of the effort to
success, defined as the number of objective function evaluations before a point in the
target set is sampled. We compute these with MATLAB, using standard techniques
for Markov chains and stochastic processes.
Tuning λ. Observation 1 suggests the parameter choice λ = 1.34 for Dürr and
Høyer’s algorithm. Numerical experimentation agrees with this choice. Figure 3 shows
the performance graphs, seeking 1% (K = 100) or 0.2% (K = 500) of the domain, of
Dürr and Høyer’s algorithm using a selection of values of λ ranging from 1.05 to 30,
1182 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

Fig. 4. Performance graphs comparing Dürr and Høyer’s method to the method of section 7,
for a uniform range distribution.

and including the values 8/7 and 1.34 suggested by [2] and Figure 2. Performance
deteriorates slowly outside of the range from 1.34 to 1.44, but within that range
there is no visible performance gradient. The value of λ may become more important
for smaller values of α, but for the remainder of this section we shall use the value
λ = 1.34.
Comparing the new method to Dürr and Høyer. Having settled on the
parameter value λ = 1.34 for Dürr and Høyer’s method, we can compare it to the
method of section 7. Figure 4 shows that, in the two cases studied, the new method
dominates that of Dürr and Høyer. For instance, to sample a target comprising 0.2%
of the domain with probability 90% or more, Dürr and Høyer’s method requires more
than 100 units of effort, whereas the new method requires only 79 (and in fact it then
samples the target with probability 96%).
Note also, in the two situations depicted in Figure 4, the estimated bound of

2.46 N/s on the expected time required by Dürr and Høyer’s algorithm, mentioned
following Theorem 6.2, amounts to 24.6 and 55.0. While the true expectations cannot
be computed from any finite portion of the performance graphs, these figures do
appear visually to be in approximate agreement with the numerical results.
Nonuniform range distributions. Until now in this section we have assumed
a uniform range distribution. This corresponds to the assumption of injectivity of the
objective function, that is, that different points in the domain map to different values
in the range. In many cases, however, for instance in combinatorial optimization,
there may be a unique optimum, or a small number of optimal domain points, but
large sets of the domain sharing values in the middle of the range; this results in a
nonuniform range distribution.
Experimentation indicates that nonuniformity of the range distribution improves
the performance of both methods under study. To produce Figure 5, we randomly
created five stochastic vectors of length 20 with first element 0.002 (the remainder
of each vector was a point uniformly distributed in [0, 1]19 and then scaled to sum
to 0.998) and simulated the performance of both methods. Compare this with the
last plot of Figure 4. Nonuniformity has improved the performance of the method of
section 7 somewhat. However, a greater improvement in Dürr and Høyer’s method
has allowed it to overtake the method of section 7. Here, for most of the five sample
 APPLYING GROVER ADAPTIVE SEARCH 1183

Fig. 5. Performance graphs comparing Dürr and Høyer’s method to the method of section 7,
for a nonuniform range distribution.

range distributions, Dürr and Høyer’s method reaches the target with probability 90%
or more after 61 or fewer units of effort, whereas the new method now requires 67.
9. Conclusion. This paper outlines the significance ofGrover’s quantum search
algorithm (with its performance characteristics implying O( N/t) performance taken
as an axiom) for global optimization. Grover search can provide the basis of imple-
menting adaptive global optimization algorithms. One example is an algorithm of
Dürr and Høyer’s introduced as a method for finding minimum values in a database.
An improved analysis of Dürr and Høyer’s algorithm suggests increasing its parameter
λ from 8/7 to 1.34. Also, that algorithm fits the Grover adaptive search framework,
and thus is applicable to the more general global optimization problem. A new algo-
rithm within the same framework is proposed in section 7. Our numerical experiments
in section 8 show that the algorithms have similar performance. The method proposed
in section 7 had its parameters tuned for the distinct objective function value case
and shows superior performance to that of Dürr and Høyer’s in that case. On the
other hand, Dürr and Høyer’s method (with λ = 1.34) overtakes the new method if
there is a great deal of repetition in objective function values.
A final comment concerning implementation on a quantum computer. This is
work mainly for computer engineers of the future, but some indications are known
at the present time. A fully functional quantum computer would be able to evaluate
an objective function in just the same way as a conventional computer, by executing
compiled code. A technical requirement to control quantum coherence, which we have
not mentioned previously, is that the gates must implement reversible operations. The
code implementing the objective function must be run in the forward direction and
then in the reverse direction. This obviously at most doubles the computational effort
for a function evaluation compared to a conventional computer.
 1184 W. P. BARITOMPA, D. W. BULGER, AND G. R. WOOD

REFERENCES

[1] W. P. Baritompa, Handling Different Cost Algorithms, Department of Mathematics and

Statistics, Technical Report, University of Canterbury Christchurch, New Zealand,
preprint, 2001.
[2] M. Boyer, G. Brassard, P. Høyer, and A. Tapp, Tight bounds on quantum searching,
Fortschr. Phys., 46 (1998), pp. 493–506.
[3] S. H. Brooks, A discussion of random methods for seeking maxima, Oper. Res., 6 (1958),
pp. 244–251.
[4] D. W. Bulger, W. P. Baritompa, and G. R. Wood, Implementing pure adaptive search with
Grover’s quantum algorithm, J. Optim. Theory Appl., 116 (2003), pp. 517–529.
[5] D. W. Bulger, D. L. J. Alexander, W. P. Baritompa, G. R. Wood, and Z. B. Zabinsky,
Expected hitting time for backtracking adaptive search, Optimization, 53 (2004), pp. 189–
202.
[6] D. W. Bulger and G. R. Wood, Hesitant adaptive search for global optimisation, Math.
Programming, 81 (1998), pp. 89–102.
[7] Centre for Quantum Computation, Oxford, http://www.qubit.org.
[8] C. Dürr and P. Høyer, A quantum algorithm for finding the minimum, http://lanl.arxiv.
org/abs/quant-ph/9607014, version 2 (7 Jan. 1999).
[9] L. K. Grover, A fast quantum mechanical algorithm for database search, in Proceedings of the
28th Annual ACM Symposium on the Theory of Computing, Philadelphia, 1996, pp. 212–
219.
[10] L. K. Grover, A framework for fast quantum mechanical algorithms, in Proceedings of the
30th Annual ACM Symposium on the Theory of Computing, Dallas, 1998, pp. 53–62.
[11] E. M. T. Hendrix and O. Klepper, On uniform covering, adaptive random search and rasp-
berries, J. Global Optim., 18 (2000), pp. 143–163.
[12] R. Laflamme, Los Alamos scientists make seven bit quantum leap, http://www.lanl.gov/
worldview/news/releases/archive/00-041.html (2000).
[13] P. W. Shor, Polynomial-time algorithms for prime factorization and discrete logarithms on a
quantum computer, SIAM J. Comput., 26 (1997), pp. 1484–1509.
[14] C. P. Stephens and W. P. Baritompa, Global optimization requires global information, J. Op-
tim. Theory Appl., 96 (1998), pp. 575–588.
[15] G. R. Wood, Z. B. Zabinsky, and B. P. Kristinsdottir, Hesitant adaptive search: The
distribution of the number of iterations to convergence, Math. Programming, 89 (2001),
pp. 479–486.
[16] Z. B. Zabinsky and R. L. Smith, Pure adaptive search in global optimization, Math. Pro-
gramming, 53 (1992), pp. 323–338.
[17] Z. B. Zabinsky, G. R. Wood, M. A. Steel, and W. P. Baritompa, Pure adaptive search for
finite global optimization, Math. Programming, 69 (1995), pp. 443–448.
[18] C. Zalka, Grover’s quantum searching algorithm is optimal, Phys. Rev. A, 60 (1999), pp. 2746–
2751.

View publication stats