1: \section{Algorithmic Approach}
2: When trying to develop a good search strategy,
3: we have to balance theoretical quality with
4: practical applicability. More precisely, we have to
5: keep a close eye on the trade-off between these objectives:
6: An increase in theoretical quality may come at the
7: expense of higher mathematical difficulty, possibly requiring
8: more complicated tools.
9: In an online context, the use of such tools may cause both
10: theoretical and practical difficulties:
11: Complicated solutions may cause computational overhead that
12: can change the solution itself by causing extra delay;
13: on the practical side, actually applying such a solution
14: may be difficult (due to limited accuracy of the robot's motion)
15: and without significant use.
16: To put relevant error bounds into perspective:
17: The largest room available to us is the great hall of Schloss
18: Birlinghoven; even there, the size of Kurt and the object
19: is still in the order of 2\% of the room diameter.
20:
21: On the mathematical side, it should be noted that even in
22: the theoretical paper \cite{hikk-pep-01},
23: semi-circles are considered instead of the solution to the
24: differential equation, in order to allow analysis of the resulting
25: trajectories.
26:
27: In the following, we will start by giving some basic mathematical
28: observations and properties (Section~\ref{subsec:basic});
29: this is followed by a discussion
30: of globally optimal strategies (Section~\ref{subsec:global}).
31: Section ~\ref{subsec:circle} describes a natural heuristic solution
32: that is both easy to describe and fast to evaluate; we give a number
33: of computational and empirical results that suggest our heuristic
34: is within 2\% of an optimal strategy.
35: Finally, Section ~\ref{subsec:asymp} provides a number of
36: mathematical results, showing that our fast and easy
37: heuristic is asymptotically optimal.
38:
39: \subsection{Basic Observations}
40: \label{subsec:basic}
41:
42: First, we introduce some notation that will be used throughout this section.
43:
44: Let us assume that the corner that hides the object
45: is at distance $d$ from the start.
46: Let $x_i$ denote the distance the robot travels in the
47: $i-$th step, i.e., on its way from position $i-1$ to position $i$,
48: from which the $i-$th scan will be taken.
49: If the object was hidden infinitesimally behind position~$i$,
50: the optimal solution would go perpendicularly to the line $L_i$ that runs
51: from the corner through position $i$, and then take one scan from there.
52: Let $d_i$ denote the length of this line segment (observe that it meets $L_i$
53: at a point that lies on the semi-circle spanned by the start
54: and the corner).
55: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
56: %ACHTUNG NOTATIONEN: d_i ist die Sehnenlänge, wie in Lemma1 benutzt, nicht s_i
57: %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
58: Then the optimum cost to detect the object would be $1+d_i$, whereas the robot
59: would only see the object at position $i+1$, having accumulated a cost of
60: \[
61: i+1 + \sum_{j=1}^{i+1} x_j.
62: \]
63: Now suppose that $c$ is the smallest competitive ratio that can be achieved
64: in this setting. By local optimality, for any scan position, the ratio of the
65: solution achieved and the optimal solution must be equal to $c$.
66: Therefore,
67: \begin{equation}
68: x_{i+1} = c(1+d_i) - (i+1) - \sum_{j=1}^i x_j
69: \label{eq:rec}
70: \end{equation}
71: must hold for $i=1,2,\ldots$ In particular, we have $x_1 = c-1$ for the
72: first step.
73:
74: \subsection{Globally Optimal Strategies}
75: \label{subsec:global}
76:
77: The above recursion can be used for proving a lower bound.
78: \begin{theorem}
79: \label{th:lower}
80: There is no global $c$-competitive strategy with $c<2$.
81: \end{theorem}
82:
83: \proof
84: Assume the claim was false, and there was a $c$-competitive
85: strategy for $c=2-\delta$.
86: We show that $x_i \leq (1-\delta)^i$ holds,
87: making it impossible for the robot to get further than a distance
88: of $1/\delta$ away from the start, a contradiction.
89: Clearly, we have $x_1=1-\delta$ for step~1.
90: Moreover,
91: \[
92: d_i \leq \sum_{j=1}^i x_j
93: \]
94: holds, because $d_i$ is the shortest path from the start to line $L_i$,
95: whereas the sum denotes the length of the robot's path.
96: Plugging this into our recursion yields
97: \[
98: x_{i+1} \leq (1 - \delta) (1+\sum_{j=1}^i x_j) -i.
99: \]
100: By induction, we have $x_j \leq (1-\delta)^j$, hence
101: \begin{equation*}
102: x_{i+1} \leq (1 - \delta)\frac{1-(1-\delta)^{i+1}}{\delta} -i \leq
103: (1-\delta)^{i+1},
104: \end{equation*}
105: using the Bernoulli inequality $1-(1+i)\delta \leq (1-\delta)^{i+1}$.
106: \qed
107:
108: \medskip
109: Instead of increasing the distance $d$ we could as well
110: consider a situation where start and corner are a distance~1 apart, but the
111: scan cost is only $1/d$. Now Theorem~\ref{th:lower}
112: shows a remarkable discontinuity: Even for a scan cost arbitrarily small,
113: a lower bound of~2 cannot be beaten, whereas for zero scan cost, a factor
114: of $1.212\dots$ can be obtained~\cite{ikm-hlac-93}.
115:
116: On the positive side, for $n$ intermediate scan points, Equation~(\ref{eq:rec})
117: provides $n$ optimality conditions. As there are $2n$ degrees
118: of freedom (the coordinates of intermediate scan points),
119: we get an underdetermined nonlinear optimization problem for any given
120: distance $d$, provided that we know the number of scan points.
121: For $d=1$, this can be used to derive an optimal competitive factor of
122: 1.808201..., achieved with one intermediate scan point.
123: For larger $d$ (and hence, larger $n$) one could derive additional
124: geometric optimality conditions and use them in combination with more
125: complex numerical methods.
126: However, this approach appears impractical for real applications,
127: for reasons stated above. As we will see in the following, there is
128: a better approach.
129:
130: \subsection{A Simple Heuristic Strategy}
131: \label{subsec:circle}
132:
133: Now we describe a simple strategy for the searching
134: problem that uses trajectories inscribed into a circle. This
135: reduces the degrees of freedom to the point where evaluation
136: is fast and easy. What is more, it works very
137: well in realistic settings, and it is asymptotically optimal for
138: decreasing cost of scanning, or growing size of the environment.
139:
140: The robot simply follows a polygonal path inscribed into
141: the semi-circle of diameter $d$, spanned by start
142: and corner. It remains to determine those points where it
143: stops for scanning its environment. This is done by applying the
144: optimality condition
145: derived in Section~\ref{subsec:basic}.
146: In step $j$, the robot moves along a chord of
147: length $x_j$. From the corner, this chord is visible
148: under an angle of $\varphi_j= \arcsin(x_j/d)$.
149: The chord connecting the start to position $i$ is of length
150: \[
151: d_i = d\ \sin( \sum_{j=1}^i \varphi_j),
152: \]
153: so that the recursion (\ref{eq:rec})
154: obtained in Section~\ref{subsec:basic} turns into
155: \begin{eqnarray*}
156: x_{i+1}&=&c\left(1+d\ \sin\left(\sum_{j=1}^{i}\arcsin\left(\frac{x_j}{d}\right)\right)\right)
157: -(i+1)-\sum_{j=1}^{i}x_j.
158: \end{eqnarray*}
159: Given any $c>1$, we can tentatively compute steps of length $x_i$
160: by this formula, starting with $x_1=c-1$.
161: If the resulting sequence reaches the corner,
162: the ratio of $c$ can indeed be achieved.
163: If it collapses prematurely (by returning negative values) $c$ was too small.
164: (For example, $c=2.001525...$ is optimal for $d=40$; see Figure~\ref{fig:40}
165: for an illustration of upper and lower bounds on this value.)
166:
167: \begin{figure}[htbp]
168: \centerline{\epsfig{file=40.20016.eps,width=0.6\textwidth}} %\textwidth
169: \centerline{\epsfig{file=40.20015.eps,width=0.6\textwidth}} %\textwidth
170: \label{fig:40}
171: \caption{An example for $d=40$: (top) For $c=2.0016$, the circle sequence
172: reaches the corner, showing that the chosen $c$ can be achieved.
173: (bottom) For $c=2.0015$, the sequence collapses before reaching
174: the corner, showing that the chosen $c$ cannot be achieved.
175: The actual optimum is about 2.001525...}
176: \end{figure}
177:
178: By performing a binary search, the optimal ratio and the necessary
179: step lengths can be computed extremely fast. Moreover, an analysis of
180: the optimal ratio as a function of $d$ shows that a maximum is reached
181: for $d=4.400875...$
182: which is precisely at the threshold between three and four necessary scans,
183: with a competitive ratio of 2.168544.
184: (See Table~\ref{tab:thresh} for an overview of the critical values
185: for which the number of scans increases, and Figure~\ref{fig:func}
186: for the achievable ratios as a function of the distance.)
187: This is still within about 2\% of the global optimum, which appears to be
188: at about 2.12 (see Figure~\ref{fig:212}.)
189: Moreover, numerical evidence shows
190: that the ratio approaches 2 quite rapidly
191: as $d$ tends to infinity. This is all the more surprising, as the resulting
192: initial step length converges to 1, while a constant step length
193: of 1 yields a competitive ratio of $\pi$.
194: In the following Section~\ref{subsec:asymp} we give a mathematical
195: proof of this observation.
196:
197: \begin{table}[t!]
198: \begin{center}
199: \noindent
200: \begin{tabular}{|c|c|c|}
201: \hline
202: \ \ Number \ \ & \ \ Maximal \ \ & \ $c$ at \ \\
203: \ \ of scans \ \ & \ \ $d$ \ \ & \ upper bound \ \\
204: \hline
205: \hline
206: 0 & 0.618034& 1.618034\\
207: 1 & 1.530414& 2.040287\\
208: 2 & 2.799395& 2.155363\\
209: 3 & 4.400876& 2.168544\\
210: 4 & 6.316892& 2.147994\\
211: 5 & 8.514200& 2.118498\\
212: %0 & 0.618033995635122 & 1.618033995635122 \\
213: %1 & 1.530414115619580 & 2.040286988611578 \\
214: %2 & 2.799395458454474 & 2.155362902918181 \\
215: %3 & 4.400875777167743 & 2.168544137757960 \\
216: %4 & 6.316892402768754 & 2.147994262968274 \\
217: %5 & 8.514199850252105 & 2.118498141207955 \\
218: \hline
219: \end{tabular}
220: \caption{Threshold values for small numbers of scans, rounded to six digits.}
221: \label{tab:thresh}
222: \end{center}
223: \end{table}
224:
225: \begin{figure}[htbp]
226: \centerline{\epsfig{file=212.eps,width=0.8\textwidth}} %\textwidth
227: \caption{A solution for $d=4.4$ that achieves competitive ratio 2.12:
228: The starting position is at $A$, the corner at $B$.}
229: \label{fig:212}
230: \end{figure}
231:
232: \begin{figure}[htbp]
233: \centerline{\epsfig{file=curve2.eps,width=0.8\textwidth}} %\textwidth
234: \vspace*{5mm}
235: \centerline{\epsfig{file=curve.eps,width=0.8\textwidth}} %\textwidth
236: \caption{The competitive ratio as a function of $d$:
237: (top) for small values of $d$;
238: (bottom) for larger values of $d$;
239: note the cusps at threshold values, the sharp peak at about (4.4,2.17), and
240: the clear asymptotic behavior. The first step length, $x_1$ is given by
241: $c-1$.}
242: \label{fig:func}
243: \end{figure}
244:
245: %%\begin{table}[t!]
246: %\begin{center}
247: %\noindent
248: %\begin{tabular}{|c|c|c|}
249: %\hline
250: %\ \ $d$ \ \ & \ \ $c$ \ \ \\
251: %\hline
252: %\hline
253: %0.1 & 1.100000000000000\\
254: %0.2 & 1.200000000000000\\
255: %0.3 & 1.300000000000000\\
256: %0.4 & 1.400000000000000\\
257: %0.5 & 1.500000000000000\\
258: %0.6 & 1.600000000000000 \\
259: %0.7 & 1.683560375413682\\
260: %0.8 & 1.743303547736261\\
261: %0.9 & 1.793846198523264\\
262: %1.0 & 1.839286755214161\\
263: %1.1 & 1.881366659010949\\
264: %1.2 & 1.921004390426161\\
265: %1.3 & 1.958749927759779\\
266: %1.4 & 1.994962024018910\\
267: %1.5 & 2.029889774672129\\
268: %1.6 & 2.041541885858511\\
269: %1.7 & 2.047085370601057\\
270: %1.8 & 2.054799981848935\\
271: %1.9 & 2.063663342108746\\
272: %2.0 & 2.073191047336831\\
273: %2.1 & 2.083117497383667\\
274: %2.2 & 2.093283433353382\\
275: %2.3 & 2.103587435265142\\
276: %2.4 & 2.113962101519117\\
277: %2.5 & 2.124361227114723\\
278: %2.6 & 2.134752412775745\\
279: %2.7 & 2.145112571450110\\
280: %2.8 & 2.155306352907740\\
281: %2.9 & 2.148487573089050\\
282: %3.0 & 2.144877197329530\\
283: %3.1 & 2.143106360512665\\
284: %3.2 & 2.142541771356012\\
285: %3.3 & 2.142829356158246\\
286: %3.4 & 2.143748285358150\\
287: %3.5 & 2.145150584746137\\
288: %3.6 & 2.146931925327941\\
289: %3.7 & 2.149015902786048\\
290: %3.8 & 2.151344890601265\\
291: %3.9 & 2.153874390950044\\
292: %4.0 & 2.156569377418749\\
293: %4.1 & 2.159401834991393\\
294: %4.2 & 2.162349052201169\\
295: %4.3 & 2.165392403501791\\
296: %4.4 & 2.168516461220506\\
297: %4.5 & 2.161677539592321\\
298: %4.6 & 2.156779978243191\\
299: %4.7 & 2.153166666113382\\
300: %4.8 & 2.150448040478810\\
301: %4.9 & 2.148392646922383\\
302: %5.0 & 2.146849901386174\\
303: %5.1 & 2.145715522433242\\
304: %5.2 & 2.144913918842684\\
305: %5.3 & 2.144388350353874\\
306: %5.4 & 2.144095037308599\\
307: %5.5 & 2.143999438703545\\
308: %5.6 & 2.144073796113323\\
309: %5.7 & 2.144295454116049\\
310: %5.8 & 2.144645676941181\\
311: %5.9 & 2.145108793338261\\
312: %6.0 & 2.145671565018517\\
313: %6.1 & 2.146322711300967\\
314: %6.2 & 2.147052545337888\\
315: %6.3 & 2.147852691611595\\
316: %6.4 & 2.143258308154723\\
317: %6.5 & 2.138786892308755\\
318: %6.6 & 2.135232840819421\\
319: %6.7 & 2.132336704808192\\
320: %6.8 & 2.129937590066217\\
321: %6.9 & 2.127928233454718\\
322: %7.0 & 2.126233215274163\\
323: %7.1 & 2.124797299166372\\
324: %7.2 & 2.123578707638580\\
325: %7.3 & 2.122545009483175\\
326: %7.4 & 2.121670482613897\\
327: %7.5 & 2.120934355726916\\
328: %7.6 & 2.120319596953788\\
329: %7.7 & 2.119812055818866\\
330: %7.8 & 2.119399840737092\\
331: %7.9 & 2.119072857898155\\
332: %8.0 & 2.118822463402640\\
333: %8.1 & 2.118641196562492\\
334: %8.2 & 2.118522572466230\\
335: %8.3 & 2.118460918546419\\
336: %8.4 & 2.118451244311209\\
337: %8.5 & 2.118489136412374\\
338: %8.6 & 2.114760318625351\\
339: %8.7 & 2.111222560556033\\
340: %8.8 & 2.108317261908449\\
341: %8.9 & 2.105878603657950\\
342: %9.0 & 2.103797969795149\\
343: %10.0 & 2.092874692523618\\
344: %11.0 & 2.087508078072620\\
345: %12.0 & 2.071665617012323\\
346: %13.0 & 2.066282781812281\\
347: %14.0 & 2.056125991435708\\
348: %15.0 & 2.048886293937643\\
349: %16.0 & 2.045645969380369\\
350: %17.0 & 2.036362979005524\\
351: %18.0 & 2.032489129552330\\
352: %19.0 & 2.028667571821866\\
353: %20.0 & 2.023217625959561\\
354: %21.0 & 2.021094140382032\\
355: %22.0 & 2.016945185614920\\
356: %23.0 & 2.014673201350458\\
357: %24.0 & 2.013511048835425\\
358: %25.0 & 2.010345528644003\\
359: %26.0 & 2.009301004935581\\
360: %27.0 & 2.007503285715742\\
361: %28.0 & 2.006530266540061\\
362: %29.0 & 2.006004938386025\\
363: %30.0 & 2.004716551512916\\
364: %31.0 & 2.004293620179882\\
365: %32.0 & 2.003530653625454\\
366: %33.0 & 2.00318504086358\\
367: %34.0 & 2.002762230232777\\
368: %35.0 & 2.002458099231800\\
369: %36.0 & 2.002283399777191\\
370: %37.0 & 2.001971602742876\\
371: %38.0 & 2.001833684640574\\
372: %39.0 & 2.001637142556429\\
373: %40.0 & 2.001525539548887\\
374: %45.0 & 2.001086079638081\\
375: %50.0 & 2.000839017941836\\
376: %55.0 & 2.000679327582184\\
377: %60.0 & 2.000566388986747\\
378: %65.0 & 2.000480652275790\\
379: %70.0 & 2.000413390542211\\
380: %75.0 & 2.000359458222390\\
381: %80.0 & 2.000315488139118\\
382: %85.0 & 2.000279149773209\\
383: %90.0 & 2.000248763585558\\
384: %95.0 & 2.000223092914125\\
385: %100.0 & 2.000201207935604\\
386: %\hline
387: %\end{tabular}
388: %\caption{Best achievable competitive ratios as a function of the distance.}
389: %\label{tab:dist}
390: %\end{center}
391: %\end{table}
392:
393: \subsection{Asymptotics}
394: \label{subsec:asymp}
395:
396: As we have seen in Theorem~\ref{th:lower}, there is a
397: lower bound of 2 on the competitive ratio for all strategies and large $d$.
398: In the following we will show that for large $d$, there is a matching
399: upper bound on our circle strategy presented in Section~\ref{subsec:circle},
400: proving it to be asymptotically optimal. For limited physical distances,
401: it shows that even for arbitrarily small scan times, there is a
402: relatively simple strategy that achieves the optimal ratio of 2.
403:
404: Our proof of the upper bound proceeds as follows. Let us assume that we are given
405: some fixed $\varepsilon>0$.
406: We then proceed to show that
407: for $c=2+\varepsilon$, the recursion presented in Section~\ref{subsec:circle}
408: does not collapse before the corner is reached, if the diameter $d$ of the
409: semi-circle is large enough.
410:
411: In proving the lower bound
412: stated in Theorem~\ref{th:lower}, we have used the obvious fact that
413: the length $d_n$ of the optimal path cannot exceed the length of the robot's
414: path. Now we are turning this argument around: The robot's path to position $n$
415: does not exceed the length of the circular arc leading from the start to
416: position $n$. As this arc is not much longer than
417: $d_n$, the length of the chord
418: from the start to $n$, if the diameter $d$ of the circle is large enough.
419: More precisely, we use the following.
420:
421: \begin{lemma}
422: \label{le:arcchord}
423: (i) There is an upper bound on the total length of the first $n$ steps
424: of the circle strategy that does only depend on $n$ and $\varepsilon$,
425: but not on $d$.\\
426: (ii) Given any $A>0$, we can find $d_0$ such that each
427: arc of length $\leq A$ in a circle of diameter $\geq d_0$ exceeds the
428: length of its chord by at most $\varepsilon^2$.
429: \end{lemma}
430:
431: \proof
432: Claim (i) can be shown by the same technique as in the proof of
433: Theorem~\ref{th:lower}. In order to prove claim~(ii), let $a$ and $c$
434: denote the maximum lengths of an arc and its chord in a circle of
435: diameter $d$ satisfying $a \leq b + \varepsilon^2$. Let $2\beta$ denote
436: the angle of the arc, as seen
437: from the center, so that $a=d \beta$ and $c=d \sin \beta$ hold. The maximum arc
438: satisfying the condition is of length $a=d \beta_d$ where $\beta_d$ is the
439: solution of the equation $\beta_d - \sin \beta_d = \varepsilon^2 /d$. In the
440: equivalent expression
441: \[
442: \ d \beta_d\ \left(1 - \frac{\sin \beta_d}{\beta_d} \right) = \varepsilon^2
443: \]
444: the fraction tends to one, so $a= d \beta_d$ must be unbounded.
445: \qed
446:
447: These facts will now be used in providing a lower bound for the first steps
448: along the semi-circle, aiming for a competitive
449: ratio of $c=2+\varepsilon$.
450:
451: \begin{lemma}
452: \label{le:liftoff}
453: Let $\varepsilon>0$ and $N$ be given. Then there is a
454: number $d_0$ such that for each diameter $d\geq d_0$ we have
455: $x_n\geq 1+\left(2^{n}-1\right)\varepsilon$, for
456: $n\leq N$.
457: \end{lemma}
458:
459: \proof
460: Using Lemma~\ref{le:arcchord} we can choose $d_0$ large enough that
461:
462: \[
463: \sum_{i=1}^{n} x_i \leq d_n + \varepsilon^2
464: \]
465: holds for all $n \leq N$ if $d \geq d_0$.
466: Now we proceed by induction. For $x_1:=(1+\varepsilon)$
467: the claim is fulfilled. For $n=2$ we observe that
468: $d_1 = x_1$ holds, so the recursive formula~(\ref{eq:rec})
469: yields
470: \begin{eqnarray*}
471: x_2 &=& (2+\varepsilon) (1+d_1) - 2 - x_1 \\
472: &=& (2+\varepsilon)^2 -3 -\varepsilon \geq 1+3\varepsilon.
473: \end{eqnarray*}
474: Now assume the claim was true for $x_1,\ldots,x_{n-1}$, where $n\geq 3$,
475: and let $d_{n-1}$ be the $(n-1)$st chord, arising by connecting
476: the start point with the $(n-1)$st scan point.
477: The induction hypothesis implies
478: $$\sum_{i=1}^j x_i
479: \geq \sum_{i=1}^j \left( 1+\left(2^{i}-1\right)\varepsilon\right)
480: = j + (2^{j+1}-j-2)\varepsilon.$$
481: From the recursion we obtain
482:
483: $$x_n=(2+\varepsilon)(1+d_{n-1})-n-\left(\sum_{i=1}^{n-1} x_{i}\right).$$
484:
485: By choice of $d$
486: we have $d_{n-1}\geq \left(\sum_{i=1}^{n-1} x_{i}\right)-\varepsilon^2$
487: for $n\leq N$.
488: Thus, we get
489: \begin{eqnarray*}
490: x_n&\geq& (2+\varepsilon)\left(1+\left(\sum_{i=1}^{n-1} x_{i}\right)-\varepsilon^2\right)
491: - n- \left(\sum_{i=1}^{n-1} x_{i}\right)\\
492: &=& (1+\varepsilon)\left(\sum_{i=1}^{n-1} x_i \right) +
493: (2+\varepsilon)(1-\varepsilon^2)-n\\
494: &=& 1+(2^n-1)\varepsilon + (2^n -n-3-\varepsilon)\ \varepsilon^2 \\
495: &\geq& 1+(2^n-1)\varepsilon,
496: \end{eqnarray*}
497: as $n\geq 3$.
498: \qed
499:
500: \medskip
501: Under the assumptions of Lemma~\ref{le:liftoff} we
502: can now prove the following.
503: \begin{lemma}
504: \label{le:rise}
505: For the first $N$ steps of the robot,
506: $\frac{1}{N}\sum_{i=0}^{N-1} x_i\geq 5$ holds.
507: \end{lemma}
508:
509: \proof
510: We may assume that
511: $$x_n\geq 1+\left(2^{n}-1\right)\varepsilon$$
512: holds for $n \leq N$. If $N$ is large enough and $n\geq N/2$, we get
513: $$ x_{n}\geq 1+ \left(\frac{10}{\varepsilon}-1\right)\varepsilon \geq 10.$$
514: Thus,
515: $$\sum_{i=1}^{N-1} x_i\geq \sum_{i=N/2}^{N-1} x_i\geq 5N,$$
516: as claimed.
517: \qed
518:
519: \medskip
520: To conclude the proof, we consider a diameter $d$ large
521: enough for Lemma~\ref{le:rise} to hold, so we have a lower
522: bound of 5 on the average size for the first $N$ steps.
523: This suffices to show that all following steps are at
524: least of length 5.
525:
526: \begin{lemma}
527: \label{le:glide}
528:
529: Assume that for some $N\geq 12$, we have
530: $\sum_{i=1}^{N-1} x_i\geq 5N$.
531: Then $x_n\geq 5$ for all $n\geq N$.
532: \end{lemma}
533:
534: \proof
535: Again we proceed by induction and consider
536:
537: $$x_n=(2+\varepsilon)(1+d_{n-1})-n-\left(\sum_{i=1}^{n-1} x_{i}\right).$$
538: As all $x_i$ are lengths of chords of the semi-circle
539: with diameter $d$, we have
540:
541: $$d_{n-1}\geq \frac{2}{\pi}\sum_{i=1}^{n-1} x_i.$$
542: By a similar argument as before, we get
543:
544: \begin{eqnarray*}
545: x_n&\geq& (2+\varepsilon)\left(1+\frac{2}{\pi} \left(\sum_{i=1}^{n-1} x_{i}\right)\right)- n
546: - \left(\sum_{i=1}^{n-1} x_{i}\right)\\
547: &\geq& \left(\frac{4}{\pi}-1\right)\left(\sum_{i=1}^{n-1} x_{i}\right) - n+2\\
548: &\geq& \left(\frac{4}{\pi}-1\right)\ 5n - n+2 \geq 5,
549: \end{eqnarray*}
550: since $n\geq 12$, as claimed.
551: \qed
552:
553: \medskip
554: With the help of these lemmas, we get
555:
556: \begin{theorem}
557: \label{th:2asymp}
558: The circle strategy is asymptotically optimal:
559: For any $\varepsilon>0$, there is a $d_\varepsilon$, such that
560: for all $d\geq d_\varepsilon$, the strategy is $(2+\varepsilon)$-competitive.
561: \end{theorem}
562:
563: \proof
564: The preceding Lemmas~\ref{le:liftoff}, \ref{le:rise}, \ref{le:glide}
565: show that for any large enough $d$, the sequence will consist of step lengths
566: that are all at least 5. This implies that the sequence will reach the
567: corner in a finite number of steps, showing that a competitive factor
568: of $(2+\varepsilon)$ can be reached.
569: \qed
570: