1: \documentstyle[pra,aps,epsfig]{revtex}
2:
3:
4: \input epsf
5:
6: \begin{document}
7:
8: %%\twocolumn[\hsize\textwidth\columnwidth\hsize\csname
9: %%@twocolumnfalse\endcsname
10: %%]
11:
12: \title{Quest for Fast Partial Search Algorithm}
13:
14: \author{ Vladimir\ E.\ Korepin$^1$ and Jinfeng Liao$^2$}
15: \address{ $^1$C.N. Yang Institute for Theoretical Physics, State
16: University of New York at Stony Brook, Stony Brook, NY 11794-3840 \\
17: $^2$Department of Physics and Astronomy, State University of New
18: York at Stony Brook, Stony Brook, NY 11794-3800
19: }
20:
21: \maketitle
22:
23: \begin{abstract}
24: A quantum algorithm can find a target item in a database faster
25: than a classical algorithm. One can trade accuracy for speed and
26: find a part of the database (a block) containing the target item
27: even faster, this is partial search. We consider different
28: partial search algorithms and suggest the optimal one. Efficiency
29: of an algorithm is measured by number of queries to the oracle.
30: \end{abstract}\vspace{0.1in}
31:
32:
33: \section{Introduction}
34:
35: Database search has many applications and is used widely. Grover
36: discovered a quantum algorithm that searches faster than a
37: classical algorithm \cite{Grover}. It consists of repetition of
38: the Grover iteration $\hat{G}_1$, which operates on the
39: computational quantum states.
40: % Recently nuclear magnetic resonance (NMR) was used to implement it \cite{jx}
41: The number of repetitions is:
42: \begin{equation}
43: j_{\mbox{full}}= \frac{\pi}{4} \sqrt{N} \label{full}
44: \end{equation}
45: for a database with large number of entries $N$. After
46: $j_{\mbox{full}} $ the algorithm finds the target item. For more
47: details, see also \cite{Bennett,Boyer,Brass}. Below we shall call
48: $\hat{G}_1$ a global iteration.
49:
50: Sometimes it is sufficient to find an approximate location of
51: the target item. A partial search considers the following problem:
52: a database is
53: separated into
54: %\begin{equation}
55: $K$ { blocks},
56: %\end{equation}
57: of a size
58: %\begin{equation}
59: $b={N}/{K}$.
60: %\end{equation}
61: %\footnote{Here we only consider a case when both $N$ and $K$ are powers of $2$.}
62: We want to find a block with the target item, not the target item
63: itself. Such partial search was first introduced by Mark Heiligman in
64: \cite{hei}, as a part of algorithm for list matching.
65: We can think of partial search in following terms: an exact address
66: of the target item is given by a sequence of $n$ bites , but we
67: want to find only first $k$ bites ($k<n$).
68: %It was first introduced by Mark Heiligman in \cite{hei},
69: %as a part of algorithm for list matching.
70: Fast quantum algorithm for a partial search was found by Grover and
71: Radhakrishnan in \cite{jaik}. They showed that classical partial
72: search takes $\sim (N-b)$ queries, but
73: % but quantum is $\sim \sqrt{b} $ faster then full search.
74: quantum algorithm takes only $\sim ( \sqrt{N}-
75: \mbox{coeff}\sqrt{b}) $ queries.
76: % and simplified in \cite{kg}.
77: % and simplified in \cite{kg}.
78: It uses several global iterations $\hat{G}^{j_1}_1$ and
79: then several local
80: iteration $\hat{G}^{j_2}_2$, see (\ref{liter}). Local searches are Grover iterations [searches] in each individual block made in
81: each block separately in parallel. Grover-Radhakrishnan algorithm
82: was improved and simplified in \cite{kg}. The number of queries to the
83: oracle in this algorithm was minimized
84: by in \cite{kor}, the $ \mbox{coeff}$ was maximized.
85: Below we shall explain minimized version of Grover-Radhakrishnan
86: algorithm.
87: We shall call it GRK algorithm.
88: In this paper we consider three other versions of partial search
89: algorithm. They use different sequences of global and local
90: searches: local-global, global-local-global and
91: local-global-local. We prove that GRK version still uses minimal
92: number of queries to the oracle. We conjecture that{\it GRK
93: algorithm is optimal among all partial search algorithms, which consist of
94: arbitrary sequence of local and global searches}.
95:
96: The plan of the paper is as follows. In the next section we
97: remind the Grover algorithm. After this we
98: formulate minimized version of Grover-Radhakrishnan algorithm
99: [GRK algorithm]. In the rest of the paper we consider other
100: partial search algorithms. We arrive at the conclusion that GRK
101: uses minimal number of queries comparing to other algorithms.
102: %The algorithm consists of three steps, described below.
103: %For many large blocks the algorithm was improved in \cite{kg}.
104: %In this paper we optimize the algorithm.
105: % partial search algorithm and compare it with
106: % other approaches for different numbers of blocks, denoted $K$.
107: %We also improve a lower bound for number of queries on an oracle, comparing
108: %to \cite{jaik}.
109: %We modify the partial search algorithm in such a way that the final state of
110: %the target block can be represented as a result of several applications of
111: %Grover iterations (\ref{iter}) to the standard initial state (\ref{ave}); see
112: %eq. (\ref{usk}). If a somebody will decide to make another quantum search [to
113: %use the target block as a new database] he/she can do it faster than
114: %(\ref{full}), as will be shown in eq. (\ref{luch}).
115:
116:
117: \section{Partial search}
118: \subsection{Global Iterations}
119: First let us remind the full Grover search. We shall consider a
120: database with one target item. The aim of the Grover algorithm is
121: to identify a target state $|t\rangle$ among an unordered set of
122: $N$ states. This is achieved by repeating global iteration which
123: is defined in terms of two operators. The first changes the sign
124: of the target state $|t\rangle$ only:
125: \begin{equation}
126: \hat{I_t}=\hat{I}-2|t\rangle \langle t|, \qquad \langle
127: t|t\rangle=1, \label{target}
128: \end{equation}
129: where $\hat{I}$ is the identity operator. The second operator,
130: \begin{equation}
131: \hat{I}_{s_1}=\hat{I}-2|s_1\rangle \langle s_1|, \label{average}
132: \end{equation}
133: changes the sign of the uniform superposition of all basis states
134: $|s_1\rangle$,
135: \begin{equation}
136: |s_1\rangle = \frac{1}{\sqrt{N}}\sum_{x=0}^{N-1}|x\rangle , \qquad
137: \langle s_1|s_1\rangle =1 . \label{ave}
138: \end{equation}
139: %The operator $I_{s_1}$ reflects amplitudes $ a_x$ about the average:
140: %\begin{eqnarray}
141: %-I_{s_1}\sum a_x|x\rangle
142: %=\sum_{x=0}^{N-1}(2\bar{a}-a_x)|x\rangle,
143: %\mbox{\u{a}}_x =
144: % \quad \bar{a}=\sum_{x=0}^{N-1}
145: %\frac{a_x}{N} \label{reflect}
146: %\end{eqnarray}
147: The {\bf global iteration} is defined as a unitary operator
148: \begin{equation}
149: \hat{G_1}=-\hat{I}_{s_1}\hat{I}_t . \label{iter}
150: \end{equation}
151: %and will be called a global iteration.
152: We shall use eigenvectors of $\hat{G_1}$:
153: %\begin{equation}
154: %Q|\psi^{\pm}_1\rangle =\lambda^{\pm}_1 |\psi^{\pm}_1\rangle
155: %\end{equation}
156: \begin{eqnarray}
157: \hat{G_1}|\psi^{\pm}_1\rangle = \lambda^{\pm}_1
158: |\psi^{\pm}_1\rangle , \qquad \lambda^{\pm}_1 =\exp[{\pm 2i
159: \theta_1}] , \qquad
160: |\psi^{\pm}_1\rangle = \frac{1}{\sqrt{2}}|t\rangle \pm \frac{i}
161: {\sqrt{2}}\left(
162: \sum^{N-1}_{\stackrel{\mbox{\small{x=0}}}{\mbox{\small{x $\neq$
163: t}}}} \frac{|x\rangle}{\sqrt{(N-1)}} \right). \label{value}
164: % \lambda^{\pm}_1& =&\exp[{\pm 2i \theta_1}] . \label{value}
165: \end{eqnarray}
166: They were found in \cite{Brass}, where the angle $\theta_1$ is
167: defined by
168: \begin{equation}
169: \sin ^2\theta_1 =\frac{1}{N}. \label{ang1}
170: \end{equation}
171:
172: %At the Step 1 we apply $j_1$ successive Grover iterations
173: %to the initial state (\ref{ave})
174: %After $j$ iterations, the result is \cite{Brass}
175: %\begin{eqnarray}
176: %G_1^{j_1} |s_1\rangle = \sin \left( (2j_1+1)\theta_1 \right) & |t\rangle& +
177: %\nonumber \\
178: % \frac{\cos \left( (2j_1+1)\theta_1 \right)}{\sqrt{N-1}}
179: %\sum^{N-1}_{\stackrel{\mbox{\small{x=0}}}{\mbox{\small{x $\neq$ t}}}}&
180: %|x\rangle& ,
181: %\label{first}
182: %\end{eqnarray}
183:
184:
185:
186: \subsection{GRK Algorithm for Partial Search}
187:
188: The first version of partial search was found in \cite{jaik}. The
189: algorithm uses $j_1$ global iteration and $j_2$ local iterations.
190: {\bf LOCAL ITERATIONS} are Grover iterations for each block:
191: \begin{equation}
192: \hat{G}_2=-\hat{I}_{s_2}\hat{I}_t .\label{liter}
193: \end{equation}
194: $\hat{I}_t $ is given by (\ref{target}), but $\hat{I}_{s_2}$ is
195: different. In one block it acts as:
196: \begin{equation}\label{is2}
197: \hat{I}_{s_2}{\big |}_{block}=\hat{I}{\big |}_{block}-2|s_2\rangle
198: \langle s_2|, \qquad |s_2\rangle =
199: \frac{1}{\sqrt{b}}\sum_{\mbox{\scriptsize{one block}}}|x\rangle
200: \label{nblock}\end{equation}
201: %where we sum up only within one block.
202: In the whole database $\hat{I}_{s_2}$ is the direct sum of
203: (\ref{is2}) with respect to all blocks. Both relevant
204: eigenvectors of $\hat{G_2}$ were found by in \cite{Brass}:
205: \begin{eqnarray}
206: \hat{G}_2|\psi^{\pm}_2\rangle =\lambda^{\pm}_2
207: |\psi^{\pm}_2\rangle, \qquad
208: \lambda^{\pm}_2 = \exp[{\pm 2i \theta_2}] , \label{vector2} \qquad
209: |\psi^{\pm}_2\rangle =\frac{1}{\sqrt{2}}|t\rangle \pm
210: \frac{i}{\sqrt{2}} |\mbox{ntt}\rangle
211: % \lambda^{\pm}_2 & =& \exp[{\pm 2i \theta_2}] .
212: \end{eqnarray}
213:
214: Here the $|\mbox{ntt}\rangle $ is a normalized sum of all
215: non-target items in the target block:
216: \begin{equation}
217: |\mbox{ntt}\rangle = \frac{1}{\sqrt{b-1}} \sum_{\stackrel{x \neq
218: t}{\mbox{\tiny{target block}}}} |x\rangle,\qquad \langle
219: \mbox{ntt} |\mbox{ntt}\rangle =1 .\label{vntt}
220: \end{equation}
221: We shall need an angle $\theta_2$ given by
222: \begin{equation}
223: \sin^2 \theta_2 =\frac{K}{N}=\frac{1}{b}. \label{ang2}
224: \end{equation}
225:
226: The partial search algorithm of \cite{jaik} creates a vector
227: \begin{equation}
228: |d\rangle =\hat{G}_1 \hat{G}^{j_1}_2 \hat{G}^{j_0}_1|s_1\rangle
229: .\label{gl}
230: \end{equation}
231: see \footnote{ We use a modification of \cite{jaik}, suggested
232: in \cite{kor}}. In the state $|d\rangle $ the amplitudes of all
233: items in non-target blocks are zero. Notice that this algorithm
234: uses global-local sequence of searches. We consider large blocks
235: $b=N/K\rightarrow \infty$. The number of blocks $K$ is an
236: important parameter. We shall replace it with
237: \begin{equation}
238: \sin \gamma =\frac{1}{\sqrt{K}}, \qquad 0\le \gamma \le
239: \frac{\pi}{4} \label{gam}
240: \end{equation}
241: The optimal version of this algorithm was find in \cite{kor}. It
242: can be described by the following equations:
243: \begin{eqnarray}
244: \cos (2j_1 \theta_2)=\frac{\sin \gamma \cos 2\gamma }{\cos \gamma
245: \sin 2\gamma}, \qquad
246: \tan (2j_0 \theta_1)&=& \frac{\cos 2\gamma }{(\sin \gamma) \sqrt{3-4
247: (\sin \gamma)^2} } \label{korepin}
248: \end{eqnarray}
249: Partial search is faster then full search (\ref{full}) by $\sim
250: \sqrt{b}$, the coefficient in front of $\sqrt{b}$ is explicitly
251: calculated in \cite{kor} and studied as a function of number of
252: blocks.
253:
254:
255: \subsection{Notation and Setup for General Partial Search Algorithm }
256:
257: Let us introduce a unite vector:
258: \begin{eqnarray}
259: %|t\rangle &,& \qquad |ntt\rangle \\
260: |u\rangle = \frac{1}{\sqrt{b(K-1)}} \sum_{\stackrel{\mbox{all
261: items in all}}{\mbox{{non-target blocks}}}} |x\rangle , \qquad
262: \langle u|u\rangle =1 .\label{ntb}
263: \end{eqnarray}
264: We shall use a three dimensional space. The orthonormal basis is
265: formed by the target item $|t\rangle$, sum of all non-target items
266: in the target block $|ntt\rangle$, defined in (\ref{vntt}) and
267: $|u\rangle $. All the state vectors involved in present quantum
268: search problem can be written in this basis as
269: \begin{equation} |V>=(a,b,c)^T \end{equation}
270: meaning
271: \begin{equation}\label{vectorform}|V>=a|t>+b|ntt>+c|u>\end{equation} For
272: example, the global uniform state which is used as initial state
273: of searching is given by
274: \begin{equation}
275: |s_1\rangle =( \sin \gamma \sin \theta_2, \sin \gamma \cos
276: \theta_2 , \cos \gamma )^T \label{initial}
277: \end{equation}
278: and the local uniform state is
279: \begin{equation}
280: |s_2\rangle =( \sin \theta_2, \cos \theta_2 , 0 )^T \label{local
281: initial}
282: \end{equation}
283:
284: The algorithms which we consider in this paper can be represented
285: as matrices in this linear space. For example $j_2$ repetitions of
286: the local iteration (\ref{liter}) is:
287: \begin{equation}
288: \hat{G}^{j_2}_2=\left( \begin{array}{clcr}
289: \cos (2j_2\theta_2) & \sin (2j_2\theta_2) &0 \\
290: -\sin (2j_2\theta_2) & \cos (2j_2\theta_2) &0 \\
291: 0 & 0& 1
292: \end{array}
293: \right) \label{loc}
294: \end{equation}
295: The ordering of eigenvectors is $|t\rangle$, $|ntt\rangle $ and
296: $|u\rangle $. The matrix has three eigenvectors:
297: \begin{eqnarray}
298: \hat{G}^{j_2}_2|v_2^{\pm}\rangle =\exp (\pm 2i\theta_2
299: j_2)|v_2^{\pm}\rangle ,\quad \hat{G}^{j_2}_2|v_2^{0}\rangle
300: =|v_2^{0}\rangle
301: \end{eqnarray}
302: The eigenvectors can be represented as:
303: \begin{equation}
304: |v_2^{\pm}\rangle=\frac{1}{\sqrt{2}}\left(\begin{array}{c}
305: 1\\
306: \pm i \\
307: 0
308: \end{array} \right), \qquad |v_2^{0}\rangle= \left(\begin{array}{c}
309: 0\\
310: 0 \\
311: 1
312: \end{array} \right).
313: \end{equation}
314:
315: Now let us turn our attention to global iterations (\ref{iter}),
316: $j_1$ repetitions of the global iterations can be represented as
317: \begin{equation}
318: \hat{G}^{j_1}_1= \left( \begin{array}{clcr} \cos (2j_1\theta_1), &
319: \sin (2j_1\theta_1) \sin \gamma, &
320: \sin (2j_1\theta_1) \cos \gamma \\
321: -\sin (2j_1\theta_1)\sin \gamma,\quad & (-1)^{j_1}\cos^2 \gamma +\cos (2j_1\theta_1) \sin^2 \gamma,\quad &\sin \gamma \cos \gamma \left( (-1)^{j_1+1}+\cos (2j_1\theta_1)\right) \\
322: - \sin (2j_1\theta_1) \cos \gamma, \quad &\sin \gamma \cos \gamma
323: \left( (-1)^{j_1+1}+\cos (2j_1\theta_1)\right), &
324: (-1)^{j_1}\sin^2 \gamma +\cos (2j_1\theta_1) \cos^2 \gamma
325: \end{array}
326: \right)
327: \end{equation}
328: This is a simplified asymptotic expression valid in the limit of
329: large blocks $b\rightarrow \infty$. We used (\ref{gam}). The
330: matrix has three eigenvectors:
331: \begin{eqnarray}
332: \hat{G}^{j_1}_1|v_1^{\pm}\rangle =\exp (\pm 2i\theta_1
333: j_1)|v_1^{\pm}\rangle ,\quad \hat{G}^{j_1}_1|v_1^{0}\rangle
334: =(-1)^{j_1}|v_1^{0}\rangle
335: \end{eqnarray}
336: The eigenvectors can be represented as:
337: \begin{equation}
338: |v_1^{\pm}\rangle=\frac{1}{\sqrt{2}}\left(\begin{array}{c}
339: 1\\
340: \pm i\sin \gamma \\
341: \pm i\cos \gamma
342: \end{array} \right), \qquad |v_1^{0}\rangle= \left(\begin{array}{c}
343: 0\\
344: \cos \gamma \\
345: -\sin \gamma
346: \end{array} \right).
347: \end{equation}
348:
349:
350: \subsection{General Problem of Partial Search}
351: The ultimate goal of partial search is to start with the uniform
352: state $|s_1>$ and locate the target block, and obviously GRK is
353: not the only means to achieve the goal. Based on the two types of
354: queries, global and local iterations, we naturally generalize GRK
355: into a wide set of partial search algorithms by alternate use of
356: the two iterations:
357: \begin{equation}
358: \hat{G}(j_k,j_{k-1},\cdot\cdot\cdot,j_2,j_1,j_0)=\hat{G_1}^{j_k}
359: \hat{G_2}^{j_{k-1}} \hat{G_1}^{j_{k-2}} \hat{G_2}^{j_{k-3}} \cdot
360: \cdot \cdot \hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0}
361: \end{equation}
362: which fulfills the following condition
363: \begin{equation} \label{constraint}
364: <u|\, \hat{G}(j_k,j_{k-1},\cdot\cdot\cdot,j_2,j_1,j_0) \, |s_1>=0
365: \end{equation}
366: This means that amplitude of each item in non-target block is
367: zero. Here all $j_i$ are non-negative integers. The total number
368: of queries for these algorithms is given by
369: $S=\sum_{i=0}^{_k}j_i$, and we should try to minimize $S$ to find
370: the optimal one.
371:
372: Now let's consider various sequences. A few discussions can be
373: made here:\\
374: 1) The last queries in the sequence should be the global
375: iterations. Note that local iteration (\ref{loc}) doesn't do
376: anything on $|u>$ but only rotates state component inside the
377: target block, so if an algorithm makes computational state in the
378: target block after the last local queries, it can simply waive the
379: last local queries since the state must already rest in the target
380: block before those unnecessary last local iterations. That is, if
381: $<u|\, \hat{G_2}^{j_{k+1}} \hat{G_1}^{j_{k}} \hat{G_2}^{j_{k-1}}
382: \cdot \cdot \cdot \hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0}
383: \, |s_1>=0$ then we must also have $<u|\, \hat{G_1}^{j_{k}}
384: \hat{G_2}^{j_{k-1}} \cdot
385: \cdot \cdot \hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0} \, |s_1>=0$. \\
386: 2) The first queries could be either local or global iterations.
387: Sequences starting with the global ($j_0>0$) include:
388: $\hat{G_1}^{j_0}$ (Global), $\hat{G_1}^{j_2} \hat{G_2}^{j_1}
389: \hat{G_1}^{j_0}$ (Global-Local-Global), and so on. The simplest
390: one, with only global queries involved, gives nothing but the
391: original Grover's full search (\ref{full}), which saves no steps
392: but precisely locates the target item. The next simplest one,
393: Global-Local-Global, will be studied in later section. Note that
394: the up-to-now established optimal partial search GRK algorithm
395: also falls into this category, with $j_2=1$ and $j_{0},j_1$
396: specified by
397: (\ref{korepin}). \\
398: 3) For those starting with local queries ($j_0=0$ and $j_1>0$), we
399: have: $\hat{G_1}^{j_2} \hat{G_2^{j_1}}$ (Local-Global),
400: $\hat{G_1}^{j_4} \hat{G_2^{j_3}} \hat{G_1}^{j_2} \hat{G_2^{j_1}}$
401: (Local-Global-Local-Global), and so on. We will later fully
402: consider the Local-Global algorithm and also discuss one
403: particular case of Local-Global-Local-Global which has $j_4$ to be
404: one, namely with only one global query applied after the last
405: local iterations (we call this as Local-Global-Local sequence). \\
406: 4) The argument based on the optimality of Grover's full search
407: puts lower limit of the partial search steps $S \ge
408: \frac{\pi}{4} \sqrt{N} - \frac{\pi}{4} \sqrt{b}$. It was shown in
409: \cite{jaik} that
410: % only expect partial search to achieve at most $\sqrt{b}$ speedup (at
411: %large block limit),
412: $S=\frac{\pi}{4}\sqrt{N}- R \sqrt{b}$. Here $R$ is a constant
413: independent of $b$. The optimal scheme we are seeking should
414: maximize $R$. Also according to this thought, we expect the total
415: number of global queries should scale asymptotically as
416: $\frac{\pi}{4}\sqrt{N}-\eta \sqrt{b}$, since in partial search we
417: should be faster than full search $\frac{\pi}{4}$ but save steps
418: only of order $\sqrt{b}$. As for total number of local queries it
419: should scale like $\alpha \sqrt{b}$, since $\pi \sqrt{b}$ times
420: local iteration will have rotated the state vector a whole lap on
421: the target block $|t> -- |ntt> $ plane. It follows from these
422: consideration that $R=\eta-\alpha$. Though the numbers of local
423: and global iterations $j_i$ are integer numbers and in principle
424: not continuous variables, in the large block limit we can
425: reasonably treat them as quasi-continuous and use them as function
426: arguments.
427: \\
428: 5) $R$ is a function of block number $K$ or instead the parameter
429: $\gamma=\arcsin(\frac{1}{\sqrt{K}})$ with $0< \gamma \le \pi/4$.
430: Another limit we will consider is the large $K$ limit, $K>>1$ or
431: equivalently $\gamma \to 0$. In this limit the mathematical
432: formulae will be significantly simplified, which will be very
433: helpful to analytical efforts as will be seen in later sections.
434: The cases with not so large $K$ could be complemented by direct
435: numerical verification.
436:
437:
438:
439: \section{Local-Global sequence of searches }
440: The Local-Global sequence of searches is the simplest partial
441: search scheme. For such a sequence $\hat{G}(j_2,j_1)$ with first
442: $j_1$ local iterations applied and then $j_2$ global, we have the
443: final computational state to be:
444: \begin{equation}
445: \hat{G_1}^{j_2} \hat{G_2}^{j_1} |s_1> = (A,B,C)^T
446: \end{equation}
447: Here A,B,C are the coefficients of components $|t>,|ntt>,|u>$
448: respectively (see (\ref{vectorform})), which are given by
449: \begin{eqnarray}
450: A=&& \sin\gamma \cos(2j_2\theta_1)\sin(2j_1\theta_2) + \sin^2\gamma \sin(2j_2\theta_1)\cos(2j_1\theta_2) + \cos^2\gamma \sin(2j_2\theta_1) \nonumber \\
451: B=&& -\sin^2\gamma \sin(2j_2\theta_1)\sin(2j_1\theta_2) + (-)^{j_2} \cos^2\gamma \sin\gamma \cos(2j_1\theta_2) \nonumber \\
452: && + \sin^3 \gamma \cos(2j_2\theta_1)\cos(2j_1\theta_2) + \sin\gamma \cos^2\gamma(\cos(2j_2\theta_1-(-)^{j_2})) \nonumber \\
453: C=&& -\sin\gamma\cos\gamma\sin(2j_2\theta_1)\sin(2j_1\theta_2) +
454: \sin^2\gamma \cos\gamma (\cos(2j_2\theta_1)-(-)^{j_2})\cos(2j_1\theta_2) \nonumber \\
455: && + (-)^{j_2} \sin^2\gamma\cos\gamma + \cos^3 \cos(2j_2\theta_1)
456: \end{eqnarray}
457: Then to accomplish the partial search we should have the
458: constraint equation
459: \begin{equation} <u|\hat{G_1}^{j_2}
460: \hat{G_2}^{j_1} |s_1> = C = 0
461: \end{equation}
462: which means that the amplitude of every item in non-target block
463: vanish (\ref{constraint}). Since there are only one set of local
464: and one set of global iterations, we apply the scaling as
465: discussed before and introduce
466: \begin{equation}
467: j_1=\alpha \sqrt{b} \,\,\, , \,\,\, j_2=\frac{\pi}{4}\sqrt{N}-\eta
468: \sqrt{b}
469: \end{equation}
470: We have the limitation $0 \le \alpha < \pi$ and $\eta <
471: \frac{\pi}{4} \sqrt{K}$. Now we can rewritten the constraint
472: equation as following:
473: \begin{equation}
474: \cos\gamma \cdot {\bigg ( }
475: [\sin^2\gamma\cos(2\alpha)+\cos^2\gamma]
476: \sin(\frac{2\eta}{\sqrt{K}})
477: -\sin\gamma\sin(2\alpha)\cos(\frac{2\eta}{\sqrt{K}}) +(-)^{j_2}
478: \sin^2\gamma[1-\cos(2\alpha)] { \bigg) } \, = \, 0
479: \end{equation}
480: Note that $K\ge 2$ so $\cos\gamma>0$, hence it drops out in the
481: above equation and we have
482: \begin{equation}\label{lg_cons}
483: F(\alpha,\eta)=[\sin^2\gamma\cos(2\alpha)+\cos^2\gamma]
484: \sin(\frac{2\eta}{\sqrt{K}})
485: -\sin\gamma\sin(2\alpha)\cos(\frac{2\eta}{\sqrt{K}}) +(-)^{j_2}
486: \sin^2\gamma[1-\cos(2\alpha)]=0
487: \end{equation}
488:
489: Remember we are interested in the fastest algorithm (using the
490: fewest queries), especially those using less steps than Grover's
491: full search. Now we have the total number of queries to be
492: $S=\frac{\pi}{4}\sqrt{N}-(\eta-\alpha)\sqrt{b}$, thus to minimize
493: $S$ we should maximize $R=\eta-\alpha$ under the constraint
494: (\ref{lg_cons}). The point here is that from the constraint we can
495: consider $\eta$ as a function of $\alpha$, thus $R$ is also a
496: function of $\alpha$, to which it should be optimized. We then
497: have $\frac{d R}{d\alpha}=-\frac{d\eta}{d\alpha}+1=0$, which leads
498: to
499: \begin{equation}
500: \sin(2\gamma) [ \cos^2\gamma
501: (1-\cos(2\alpha)\cos(\frac{2\eta}{\sqrt{K}}))+(-)^{j_2} \sin\gamma
502: \sin(2\alpha)]=0
503: \end{equation}
504: Again since $0<2\gamma\le \pi/2$, we have
505: \begin{equation}
506: \sin(2\gamma)>0
507: \end{equation}
508: so we simplify the above equation as
509: \begin{equation}\label{lg_dif}
510: H(\alpha,\eta)=2\sin \alpha [\cos^2\gamma \sin \alpha
511: \cos(\frac{2\eta}{\sqrt{K}})+(-)^{j_2} \sin \gamma \cos\alpha]=0
512: \end{equation}
513:
514: There is an apparent solution for (\ref{lg_cons}) and
515: (\ref{lg_dif}), namely
516: $\sin\alpha=\sin(\frac{2\eta}{\sqrt{K}})=0$. This gives $\alpha=0$
517: and $\eta=l\cdot \pi \cdot \sqrt{K}, l=0,-1,-2,\cdot\cdot\cdot$,
518: but to maximize $R=\eta-\alpha$ we should have $\alpha=\eta=0$,
519: which again recovers the Grover's full search solution and is a
520: trivial one for partial search. In the following we only consider
521: nontrivial solutions with $\sin\alpha \ne 0$.
522:
523: Let's first look at the large $K$ or small $\gamma$ limit. In
524: leading order, the two equations (\ref{lg_cons}) and
525: (\ref{lg_dif}) are reduced to be
526: \begin{equation}\label{lg_limit1}
527: F(\alpha,\eta)=\sin(\frac{2\eta}{\sqrt{K}})-\gamma
528: \sin(2\alpha)\cos(\frac{2\eta}{\sqrt{K}})+\gamma^2 (-1)^{j_2} 2
529: \sin^2 \alpha=0
530: \end{equation}
531: \begin{equation}\label{lg_limit2}
532: H(\alpha,\eta)=\sin\alpha[\sin\alpha
533: \cos(\frac{2\eta}{\sqrt{K}})+(-1)^{j_2}\gamma \cos\alpha]=0
534: \end{equation}
535: To satisfy (\ref{lg_limit1}), the first term
536: $\sin(\frac{2\eta}{\sqrt{k}})$ must be at least as small as $\sim
537: \gamma$, so in leading order we must have
538: $\cos(\frac{2\eta}{\sqrt{k}}) \sim 1$. This, combined with
539: (\ref{lg_limit2}), requires also $\sin\alpha$ must be at least as
540: small as $\sim \gamma$ (we don't consider $\sin\alpha=0$ as
541: mentioned before), which means $\cos\alpha \sim 1$. So finally we
542: reduce (\ref{lg_limit2}) in leading order to be
543: \begin{equation}
544: \sin\alpha + (-1)^{j_2} \gamma =0
545: \end{equation}
546: Remember we have $0\le \alpha < \pi$ thus $\sin\alpha \ge 0$, so
547: the above equation has solution ONLY for ODD values of $j_2$. For
548: odd $j_2$ the solution is $\sin\alpha=\gamma$, this yields two
549: solutions $\alpha=\gamma$ and $\alpha=\pi-\gamma$. By substituting
550: $\sin\alpha=\gamma$ back into (\ref{lg_limit1}) we get in leading
551: order $\sin{\frac{2\eta}{\sqrt{K}}}=2\gamma^2$, which means
552: $\frac{2\eta}{\sqrt{K}}=2\gamma^2+2l\pi$ or
553: $\frac{2\eta}{\sqrt{K}}=-2\gamma^2+(2l-1)\pi$,
554: $l=0,-1,-2,\cdot\cdot\cdot$. But again remember our goal is to
555: maximize $R=\eta-\alpha$, so we adopt the optimal Local-Global
556: solution $\alpha=\gamma$ and $\eta=\gamma$ in large K limit. This,
557: however, seems giving no speedup compared with full search since
558: $R=\eta-\alpha \sim 0$. To clarify this, we need go to higher
559: order of equations (\ref{lg_limit1}) and (\ref{lg_limit2}). By
560: properly including corrections up to $\sim \gamma^3$ we can find
561: the solution to be
562: \begin{equation}
563: \alpha=\gamma+\frac{1}{2}\gamma^3+o(\gamma^4) \,\,\, , \,\,\,
564: \eta=\gamma+\frac{2}{3}\gamma^3+o(\gamma^4)
565: \end{equation}
566: Now we see in large $K$ limit, the Local-Global sequences of
567: search do achieve speedup
568: \begin{equation}
569: R=\frac{1}{6} \gamma^3 + o(\gamma^4)
570: \end{equation}
571: which is very small but still nonzero.
572:
573: Let's now turn back to finite values of $K$. From (\ref{lg_dif})
574: we have
575: \begin{equation} \label{lg_eta}
576: \cos(\frac{2\eta}{\sqrt{K}})=\frac{(-)^{j_2+1}\sin\gamma\cos\alpha}{\cos^2\gamma\sin\alpha}
577: \end{equation}
578: Combine (\ref{lg_eta}) and (\ref{lg_cons}) together we get
579: \begin{equation}\label{lg_eta2}
580: \cos(\frac{2\eta}{\sqrt{K}})=\frac{\sin\gamma\cos\alpha}{\cos^2\gamma\sin\alpha}
581: \,\,\, , \,\,\, \sin(\frac{2\eta}{\sqrt{K}})=\frac{2\sin^2\gamma
582: (\cos^2\gamma \sin^2\alpha+\cos^2\alpha)}{\cos^2\gamma
583: [\sin^2\gamma\cos(2\alpha)+\cos^2\gamma]}
584: \end{equation}
585: By requiring
586: $\cos^2(\frac{2\eta}{\sqrt{K}})+\sin^2(\frac{2\eta}{\sqrt{K}})=1$
587: we obtain the equation determining the value of $\alpha$ and from
588: $\alpha$ we can obtain $\eta$. Generally it is hard to analyze the
589: problem analytically, so we proceed to numerical solutions , see
590: the figures Fig.\ref{fig_LGOdd},\ref{fig_LGEven}. In these figures
591: we plot $R$ as a function of $\alpha$ with different $K$ for both
592: even and odd values of $j_2$, at each point $\eta$ is determined
593: from the constraint (\ref{lg_cons}). For the odd $j_2$ case there
594: is always a positive maximum which is faster than full search, and
595: as $K$ approaches very large values, the maximum of the $\epsilon$
596: v.s. $\alpha$ curve moves gradually toward the origin which stands
597: for the full search solution. But for even $j_2$, $R$ is always
598: negative, hence slower than full search.
599:
600: \begin{figure}
601: \hspace{+0cm}
602: \centerline{\epsfxsize=13cm\epsffile{fig_lg_odd.eps}}
603: \hspace{+0.5cm}
604: \caption{\label{fig_LGOdd}
605: Dependence of $R$ on $\alpha$ with odd $j_2$ in Local-Global
606: sequence, the curves from top to bottom are for
607: $K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\alpha$ the
608: value of $\eta$ is solved from (\ref{lg_cons}). Right panel is the
609: amplification of the area near origin in Left panel.}
610: \end{figure}
611:
612: \begin{figure}
613: \hspace{+0cm}
614: \centerline{\epsfxsize=8cm\epsffile{fig_lg_even.eps}}
615: \hspace{+0.5cm}
616: \caption{\label{fig_LGEven}
617: Dependence of $R$ on $\alpha$ with even $j_2$ in Local-Global
618: sequence, the curves from top to bottom are for
619: $K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\alpha$ the
620: value of $\eta$ is solved from (\ref{lg_cons}).}
621: \end{figure}
622:
623: An important comparison is to be made between this Local-Global
624: algorithm and the established GRK one. Our numeric results show
625: that the odd $j_2$ Local-Global searches can get $R \approx 0.3$
626: with $K=2,3$ and get $0< R <0.09 $ for $K \ge 4$. For GRK,
627: however, it can achieve $R > 0.32$ for all $K\ge 2$. So the GRK is
628: much faster than the present Local-Global searches.
629:
630: To sum up with the Local-Global sequence of searches, we find it
631: can be faster than full search with $R \sqrt{b}$ speedup for block
632: number $K$ up to several tens, but it is always slower than the
633: GRK optimized partial search.
634:
635:
636:
637: \section{Global-Local-Global sequence of searches}
638: In this section we shall consider another algorithm for partial
639: search, the Global-Local-Global sequence of searches, which starts
640: with (\ref{ave}), first applies $j_0$ global iterations
641: (\ref{iter}), then $j_1$ local iterations (\ref{liter}) and
642: finally $j_2$ global iterations:
643: \begin{equation}
644: \hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0} |s_1\rangle
645: .\label{glg}
646: \end{equation}
647: Again amplitudes of all items in non-target blocks should vanish
648: at the end of algorithm:
649: \begin{equation}
650: \langle u| \hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0}
651: |s_1\rangle =0.\label{cancelation}
652: \end{equation}
653:
654: Let us present this equation in an explicit form. After first set
655: of global iterations the state of the database will be:
656: \begin{equation}
657: \hat{G_1}^{j_0}|s_1\rangle=\frac{\sin
658: \gamma}{\sqrt{b}}\left(\begin{array}{c}
659: \cos (2 j_0 \theta_1)\\
660: -\sin (2 j_0 \theta_1)\sin \gamma \\
661: -\sin (2 j_0 \theta_1) \cos \gamma
662: \end{array} \right) + \left(\begin{array}{c}
663: \sin (2 j_0 \theta_1)\\
664: \sin \gamma \cos (2 j_0 \theta_1)\\
665: \cos \gamma \cos (2 j_0 \theta_1)
666: \end{array} \right).
667: \end{equation}
668: After local iterations the state of the database is:
669: \begin{equation}
670: \hat{G_2}^{j_1} \hat{G_1}^{j_0}|s_1\rangle= \left(\begin{array}{c}
671: \sin (2 j_0 \theta_1)\cos (2 j_1\theta_2) +\sin \gamma \cos (2 j_0
672: \theta_1)
673: \sin (2j_1\theta_2) \\
674: -\sin (2 j_0 \theta_1)\sin (2j_1\theta_2)+\sin \gamma \cos (2 j_0 \theta_1)\cos (2j_1\theta_2)\\
675: \cos \gamma \cos (2 j_0 \theta_1)
676: \end{array} \right).
677: \end{equation}
678: Here we neglected $\sim 1/\sqrt{b}$ terms (namely taking large
679: block limit). After next set of global searches we have to
680: calculate only third component(coefficient of $|u\rangle$) of the
681: vector:
682: \begin{eqnarray} \label{GLG_cons}
683: 0&=&\langle u|\hat{G_1}^{j_2} \hat{G_2}^{j_1} \hat{G_1}^{j_0} |s_1\rangle \label{glgmain}\\
684: &=&\cos (2j_1\theta_2) \left\{ -\cos \gamma \sin
685: (2j_2\theta_1)\sin (2j_0\theta_1)+ \sin^2\gamma \cos
686: (2j_0\theta_1)\cos \gamma
687: \right[(-1)^{j_2+1}+\cos (2j_2\theta_1) \left] \right\} + \nonumber \\
688: &&\sin (2j_1\theta_2)\left\{-\sin \gamma \cos \gamma \sin
689: (2j_2\theta_1)\cos (2j_0\theta_1)- \sin \gamma \cos \gamma \sin
690: (2j_0\theta_1)
691: \right[(-1)^{j_2+1}+\cos (2j_2\theta_1) \left] \right\} + \nonumber \\
692: && + \cos \gamma \cos (2j_0\theta_1) [(-1)^{j_2} \sin^2 \gamma +
693: \cos^2 \gamma \cos (2j_2\theta_1) ] \nonumber
694: \end{eqnarray}
695: This constraint equation guarantees the amplitudes of all items in
696: all non-target blocks vanish to successfully complete the partial
697: search.
698:
699: First let us check the case $j_1=0$, no local searches. In this
700: case the constraint equation can be reduced to $\cos [ 2\theta_1
701: (j_0+j_2) ] = 0$. This is just the full search: we use
702: $j_0+j_2=\frac{\pi}{4}\sqrt{N} $ global iterations to find the
703: target item.
704:
705: Now let us consider more general and complicated case. We expect
706: the following scaling, namely the global iterations
707: $j_0=\frac{\pi}{4}\sqrt{N}-\eta \sqrt{b}$, $j_2=\beta \sqrt{b}$
708: and total local iterations $j_1=\alpha \sqrt{b}$ with $\eta \le
709: \frac{\pi}{4} \sqrt{K} $, $0\le \beta < \frac{\pi}{4}\sqrt{K}$ and
710: $0\le \alpha < \pi$. It should be remembered that our purpose is
711: to minimize $j_0+j_1+j_2=\frac{\pi}{4} \sqrt{N} - R \sqrt{b} ,\,\,
712: R = \eta-\beta-\alpha$, or equivalently maximize $R$. Our
713: strategy is similar to that used for analyzing Local-Global
714: sequence: study the large $K$ limit analytically while deal with
715: finite $K$ case numerically.
716:
717: In the large $K$ or small $\gamma$ limit, we can take the leading
718: order of (\ref{GLG_cons}) and simplify our constraint to a much
719: simpler form :
720: \begin{equation} \label{GLG_limit}
721: \eta=\beta \cos(2\alpha)+\frac{1-(-)^{j_3}}{2} \sin(2\alpha)
722: \end{equation}
723: From this equation, there are two possibilities: \\
724: 1) $j_2$ is even, thus $\eta=\beta \cos(2\alpha)$. In this case
725: the problem simplifies into maximizing $R=\beta (\cos(2\alpha)-1)
726: -\alpha$ with $0\le \beta < \frac{\pi}{4}\sqrt{K}$ and $0\le
727: \alpha < \pi$. Note that $\cos(2\alpha)-1 \le 0$ so the maximum of
728: R must occur at $\eta=\beta=\alpha=0$, which is again the trivial
729: full search solution (\ref{full}).
730: \\
731: 2) $j_2$ is odd, thus $\eta=\beta \cos(2\alpha)+\sin(2\alpha)$. We
732: then have to maximize $R=\beta
733: (\cos(2\alpha)-1)+\sin(2\alpha)-\alpha$ with $0\le \beta <
734: \frac{\pi}{4}\sqrt{K}$ and $0\le \alpha < \pi$. The solution is
735: \begin{equation}
736: \eta=\frac{\sqrt{3}}{2} \,\, , \,\, \beta=0 \, \, , \, \,
737: \alpha=\pi/6 ,
738: \end{equation}
739: which has achieved $R \sqrt{b} \approx 0.3424 \sqrt{b}$ speedup
740: with respect to full search. This nontrivial optimal solution is
741: exactly the GRK algorithm, with vanishingly small odd $j_2$ namely
742: $j_2=1$.
743:
744: Now let us discuss finite values of $K$.
745: We calculated dependence of $R$ on $\beta$
746: % We plot $R$ as a function of $\beta$,
747: by determining $\alpha$ from constraint equations and
748: optimizing $\eta$ numerically at each value of $\beta$.
749: We done this numerically for up to 200 blocks.
750: The dependence of $R$ on $\beta$ is monotonous.
751: Corresponding figures took too much memory, so we withdraw them from the paper.
752: % It can be seen from the figures Fig.\ref{fig_GLGOdd}\ref{fig_GLGEven} that
753: For even $j_2$ searches we consider are slower than full Grover
754: search, while odd $j_2$ these searches are faster than full search. The
755: optimal searches always occur with odd and vanishing $j_2$ (the
756: GRK case), being about $0.34\sqrt{b}$ faster than full search. The
757: large $K$ case also confirm our analysis above. Our results here
758: also confirm that the optimum partial search is GRK \cite{kor},
759: it has the form $\hat{G_1} \hat{G_2}^{j_1}
760: \hat{G_1}^{j_0}$.
761:
762: To conclude, we have found that Global-Local-Global sequence of
763: searches can be much faster than full search (\ref{full}) with odd
764: times global iterations applied in the end, and the optimal search
765: is determined to be the GRK algorithm for all values of $K$.
766:
767: %\begin{figure}
768: %\hspace{+0cm}
769: %\centerline{\epsfxsize=13cm\epsffile{fig_glg_odd.eps}}
770: %\hspace{+0.5cm}
771: % \caption{\label{fig_GLGOdd}
772: %Dependence of $R$ on $\beta$ with odd $j_2$ in Global-Local-Global
773: %sequence, the curves from top to bottom are for
774: %$K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\beta$ the
775: %value of $\alpha$ is solved from (\ref{GLG_cons}) with $\eta$
776: %optimized numerically. Right panel is the amplification of the
777: %area near origin in Left panel.}
778: % \end{figure}
779:
780: %\begin{figure}
781: % \centerline{\epsfxsize=8cm\epsffile{fig_glg_even.eps}}
782: %\hspace{+0.5cm}
783: % \caption{\label{fig_GLGEven}
784: %Dependence of $R$ on $\beta$ with even $j_2$ in
785: %Global-Local-Global sequence, the curves from top to bottom are
786: %for $K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\beta$
787: %the value of $\alpha$ is solved from (\ref{GLG_cons}) with $\eta$
788: %optimized numerically.}
789: % \end{figure}
790:
791:
792: \section{Local-Global-Local sequence of searches}
793:
794: In this section we shall discuss another algorithm of partial
795: search. It belongs to the category of Local-Global-Local-Global
796: sequence but with only one global query applied in the end, which
797: we call as Local-Global-Local sequence. We start with (\ref{ave})
798: first apply $j_1$ local iterations, then $j_2$ global iterations
799: (\ref{liter}) and then $j_3$ local iterations, and eventually a
800: single global query:
801: \begin{equation}
802: \hat{G_1} \hat{G_2}^{j_3} \hat{G_1}^{j_2} \hat{G_2}^{j_1}
803: |s_1\rangle .\label{lgl}
804: \end{equation}
805: One particular advantage of this type of sequence is that when
806: $j_1$ tends to zero our Local-Global-Local will degenerate to the
807: GRK type algorithm. Amplitudes of all items in non-target blocks
808: should vanish at the end:
809: \begin{equation}
810: \langle u| \hat{G_1} \hat{G_2}^{j_3} \hat{G_1}^{j_2}
811: \hat{G_2}^{j_1} |s_1\rangle =0.\label{lglcons}
812: \end{equation}
813: As before, we introduce the scaling of iteration numbers to be
814: $j_1=\alpha \sqrt{b}$, $j_3=\delta \sqrt{b}$, and
815: $j_2=\frac{\pi}{4}\sqrt{N}-\eta\sqrt{b}$. The total number of
816: queries will be $S=\frac{\pi}{4}\sqrt{N}-R \sqrt{b}$ with
817: $R=\eta-\alpha-\delta$ which we want minimize under constraint
818: (\ref{lglcons}). The explicit form of the above constraint
819: equation is as follows:
820: \begin{eqnarray}\label{LGL_cons}
821: 0&=&X \cdot \sin(2\alpha) + Y \cdot {\big [ } \cos(2\alpha)-1 {\big ]} + Z \\
822: && X=\sin\gamma \sin(2\gamma) \sin(2\delta) \sin(\frac{2\eta}{\sqrt{K}}) + \sin^2\gamma \sin(2\gamma)
823: \cos(2\delta) \cos(\frac{2\eta}{\sqrt{K}}) + \sin\gamma \cos\gamma \cos(2\gamma) \cos(\frac{2\eta}{\sqrt{K}}) \nonumber \\
824: && Y= \sin^2\gamma \sin(2\gamma) \sin(2\delta) \cos(\frac{2\eta}{\sqrt{K}}) - \sin^3\gamma \sin(2\gamma)\cos(2\delta)\sin(\frac{2\eta}{\sqrt{K}})
825: -\sin^2\gamma \cos\gamma {\bigg [} \cos(2\gamma)\sin(\frac{2\eta}{\sqrt{K}})+(-)^{j_2} {\bigg ]} \nonumber \\
826: && Z= \sin(2\gamma) \sin(2\delta) \cos(\frac{2\eta}{\sqrt{K}}) - \sin\gamma \sin(2\gamma) \cos(2\delta) \sin(\frac{2\eta}{\sqrt{K}})
827: -\cos\gamma \cos(2\gamma) \sin(\frac{2\eta}{\sqrt{K}}) \nonumber
828: \end{eqnarray}
829: Though the above equation is complicated, it is very easy to solve
830: numerically. Also taking large $K$ limit can significantly
831: simplify it. So as we did before, we analytically study the large
832: $K$ limit, complemented by numerical results from very small $K$
833: to very large $K$.
834:
835: By taking large $K$ or small $\gamma$ limit, we obtain from
836: leading order of (\ref{LGL_cons}) the following:
837: \begin{equation} \label{LGL_conssim}
838: \sin(2\alpha)+2\sin(2\delta)-2\eta=0
839: \end{equation}
840: We then have $\eta=\frac{1}{2} \sin(2\alpha) + \sin(2\delta)$ and
841: hence the total number of queries to be
842: \begin{equation}
843: R=\frac{1}{2}\sin(2\alpha)-\alpha+\sin(2\delta)-\delta
844: \end{equation}
845: By requiring $\frac{\partial R}{\partial \alpha}=\frac{\partial
846: R}{\partial \delta}=0$ we get the solution maximizing $R$
847: \begin{equation}
848: \alpha=0 \,\,\, , \,\,\, \delta=\frac{\pi}{6} \,\,\, , \,\,\,
849: \eta=\frac{\sqrt{3}}{2}
850: \end{equation}
851: This, with zero $j_1$, again recovers the GRK optimized solution.
852: So in large $K$ limit we see Local-Global-Local sequence is no
853: faster than GRK algorithm. Different from Local-Global and
854: Global-Local-Global, here in Local-Global-Local we notice that the
855: odd $j_2$ and even $j_2$ converge to each other when approaching
856: the optimal solution and the oscillation terms in (\ref{LGL_cons})
857: with factor $(-)^{j_2}$ disappear.
858:
859: \begin{figure}
860: \hspace{+0cm}
861: \centerline{\epsfxsize=13cm\epsffile{fig_lgl_odd.eps}}
862: \hspace{+0.5cm}
863: \caption{\label{fig_LGLOdd}
864: Dependence of $R$ on $\alpha$ with odd $j_2$ in Local-Global-Local
865: sequence, the curves from top to bottom are for
866: $K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\alpha$ the
867: value of $\eta$ is solved from (\ref{LGL_cons}) with $\delta$
868: optimized numerically. Right panel is the amplification of the
869: area near origin in Left panel.}
870: \end{figure}
871:
872: \begin{figure}
873: \centerline{\epsfxsize=13cm\epsffile{fig_lgl_even.eps}}
874: \hspace{+0.5cm}
875: \caption{\label{fig_LGLEven}
876: Dependence of $R$ on $\alpha$ with even $j_2$ in
877: Local-Global-Local sequence, the curves from top to bottom are for
878: $K=5,6,7,8,9,18,36,72,144,200$ respectively. For each $\alpha$ the
879: value of $\eta$ is solved from (\ref{LGL_cons}) with $\delta$
880: optimized numerically.}
881: \end{figure}
882:
883: For general values of $K$, we conduct the numerical method and
884: show the results in figures Fig.\ref{fig_LGLOdd}\ref{fig_LGLEven},
885: which plot $R$ as a function of $\alpha$ with $\eta$ solved from
886: (\ref{LGL_cons}) and $\delta$ optimized numerically. As can be
887: seen, the optimal solutions always occur with $\alpha=0$ which
888: goes back to GRK case. Also we note that even and odd $j_2$ give
889: same results around optimal point.
890:
891: So in this section we have established that the Local-Global-Local
892: sequence of searches can be much faster than full search, but is
893: no faster than GRK algorithm. In the appendix an alternative
894: approach for Local-Global-Local sequence based on a conjecture
895: about cancellation of oscillation terms in (\ref{LGL_cons}) is
896: briefly described, which though is not directly relevant here but
897: arrives at similar result and may shed light for future
898: exploration of even more complicated sequences.
899:
900: \section{Summary}
901: We considered different partial search algorithms, which consists of a
902: sequence of local and global searches.
903: We introduced a general framework for studying partial
904: quantum search algorithms and classified various possible sequences.
905: Particularly, we studied the Local-Global,
906: Global-Local-Global as well as Local-Global-Local sequences of
907: searches by combining numerical study for wide range values of $K$
908: and analytical results for large $K$ limit. All these algorithms
909: achieve $\sqrt{b}$ speedup compared to the Grover's full quantum
910: search. GRK algorithm \cite{kor} is the fastest among
911: partial search algorithms, which we considered.
912:
913:
914: \section*{Acknowledgements}
915:
916: The paper was supported by NSF Grant DMS-0503712.
917:
918: \appendix
919: \section*{Appendix: remarks on
920: Local-Global-Local sequence}
921:
922: Let us make two remarks:
923:
924: 1) Let's start from the explicit form of constraint equation
925: (\ref{LGL_cons}). We minimizing $j_1+j_2+j_3$. It is interesting
926: to note that at the minimum the oscillation terms cancel. The
927: coefficient at $(-1)^{j_2}$ vanishes because of GRK equation:
928: \begin{equation}
929: \cos \gamma \sin 2 \gamma \cos (2\theta_2j_3)= \sin \gamma \cos
930: 2 \gamma \label{uda}
931: \end{equation}
932: This is exactly the first equation of (\ref{korepin}). We also can
933: write it in the form:
934: \begin{equation}
935: \cos (2j_3\theta_2)= \frac{1-\tan^2 \gamma}{2} , \qquad \sin
936: (2j_3\theta_2)=\frac{\sqrt{3-4\sin^2 \gamma}}{2\cos^2 \gamma}
937: \label{uda}
938: \end{equation}
939:
940:
941: 2) Minimum number of iterations for Local-Global-Local sequences
942: corresponds to $j_1=0$ case and the algorithm is reduced back to
943: GRK version of partial search, see small $j_1$ increase $j_2+j_1$
944: very little since $d(j_2+j_1)/dj_1=0$ and $d^2(j_2+j_1)/dj_1^2=0$
945: at $j_1=0$.
946:
947:
948: \begin{thebibliography}{99}
949:
950: \bibitem{Grover} L.K.~Grover, STOC, 212(1996).
951:
952: \bibitem{Bennett} C.H. Bennett, E. Bernstein, G. Brassard, and
953: U. Vazirani, SIAM J.~Comput. {\bf 26}, 1510(1997); e-print
954: quant-ph/9701001.
955:
956: \bibitem{Boyer} M. Boyer, G. Brassard, P. Hoyer, and A. Tapp, Fortsch. Phys.
957: {\bf 46}, 493(1998); e-print quant-ph/9605034.
958:
959: \bibitem{Brass} G. Brassard, P. Hoyer, M.Mosca and A. Tapp: Quantum Amplitude Amplification and Estimation, Contemporary mathematics, vol 305, p 53, 2002
960: %\bibitem{Brass} G. Brassard, P. Hoyer, M.Mosca, and A. Tapp, vol 305, p 53, 2002
961:
962: %\bibitem{jx} Li Xiao, Jonathan A. Jones, quant-ph/0504054
963: %\bibitem{NC} M.A.~Nielsen and I.L.~Chuang: Quantum Computation and Quantum
964: %Information. Cambridge University Press (2000).
965: %\bibitem{zalka} C.~Zalka: Grover's quantum searching algorithm is optimal.
966: %Phys. Rev. A60 (1999) 2746-2751.
967: %\bibitem{mary} L. K.\ Grover lecture at second FEYNMAN FESTIVAL, August 2004,
968: %University of Maryland.
969: \bibitem {hei} M. Heiligman, Finding Matches between Two Databases on a Quantum Computer, quant-ph/0006136
970:
971: \bibitem{jaik} L. K.\ Grover and J. Radhakrishnan, e-print quant-ph/0407122.
972: \bibitem{kg} V. E. Korepin and L. K.\ Grover, accepted to Quantum Information
973: Processing,
974: quant-ph/0504157.
975: \bibitem{kor} V. E. Korepin, Journal of Physics A: Math. Gen. vol 38,
976: pages L731-L738, 2005 and e-print quant-ph/0503238.
977:
978:
979: \end{thebibliography}
980:
981:
982: \end{document}
983: