1: \documentstyle [12pt]{article}
2: \textheight 22.5cm
3: \textwidth 16.5cm
4: \topmargin=-2cm
5: \oddsidemargin -.5cm
6: \begin{document}
7: \baselineskip .3in
8: \begin{titlepage}
9: \begin{center}{\large {\bf The travelling salesman problem on randomly diluted
10: lattices: Results for small-size systems}}
11: \vskip .2in
12: {\bf Anirban Chakraborti}$~^{(1)}$ and
13: {\bf Bikas K. Chakrabarti}$~^{(2)}$\\
14: {\it Saha Institute of Nuclear Physics},\\
15: {\it 1/AF Bidhan Nagar, Calcutta 700 064, India.}\\
16: \end{center}
17: \vskip .3in
18: {\bf Abstract}\\
19: \noindent
20: If one places $N$ cities randomly on a lattice of size $L$,
21: we find that $\bar l_E\sqrt p$ and $\bar l_M\sqrt p$ vary with
22: the city concentration $p=N/L^2$, where $\bar l_E$ is the average optimal travel
23: distance per city in the Euclidean metric
24: and $\bar l_M$ is the same in the Manhattan metric. We have studied such optimum
25: tours for visiting all the cities using a branch and bound algorithm, giving the
26: exact optimized tours for small system sizes ($N\leq 100$) and near-optimal
27: tours for bigger system sizes ($100<N\leq 256$).
28: Extrapolating the results for $N\rightarrow \infty$, we find that $\bar l_E\sqrt
29: p = \bar l_M\sqrt p = 1$ for $p=1$, and $\bar l_E\sqrt p=0.73\pm 0.01$ and
30: $\bar l_M\sqrt p=0.93\pm 0.02$ with $\bar l_M/\bar l_E \simeq 4/\pi$ as
31: $p\rightarrow 0$. Although the problem is
32: trivial for $p=1$, for $p\rightarrow 0$ it certainly reduces to the standard
33: travelling salesman problem on continuum which is NP- hard.
34: We did not observe any irregular behaviour at any intermediate point.
35: The crossover from the triviality to the NP- hard
36: problem presumably occurs at $p=1$.
37: \vskip 2.5in
38: \noindent
39: {\bf PACS No. :} 05.50+q
40: \end{titlepage}
41: \newpage
42: \noindent
43: {\bf 1 Introduction}\\
44: \noindent
45: The travelling salesman problem (TSP) is a simple example of a multivariable combinatorial
46: optmization problem and perhaps the most famous one. Given a certain set of
47: cities and the intercity distance metric, a travelling salesman must find the
48: shortest tour in which he visits all the cities and comes back to his starting
49: point. It is a non-deterministic polynomial complete (NP- complete) problem
50: [1-3].
51: In the standard formulation of TSP, we have $N$ number of cities distributed
52: randomly on a continuum plane and we determine the average optimal
53: travel distance per city $\bar l_E$ in the Euclidean metric (with
54: $\Delta r_E= \sqrt {\Delta x^2+\Delta y^2}$), or $\bar l_M$ in the Manhattan
55: metric (with $\Delta r_M= |\Delta x|+|\Delta y|$).
56: Since the average distance per city scales (for fixed area) with the number of
57: cities $N$ as $1/ \sqrt N$, we find that the normalized travel distance per
58: city $\Omega_E=\bar l_E \sqrt N$ or
59: $\Omega_M=\bar l_M \sqrt N$ become the optimized constants and their values
60: depend on the method used to optimize the travel distance. Extending the
61: analytic estimates of the average nearest neighbour distances, in particular
62: within a strip and varying the width of the strip to extremize (single parameter
63: optimization approximation), one gets $\frac {5}{8}<\Omega_E<0.92$ [4] and
64: $\frac {5}{2\pi}<\Omega_M<1.17$ [5]. Careful (scaling, etc.) analysis of the
65: numerical results obtained so far indicates that $\Omega_E\simeq 0.72$ [6].
66:
67:
68: Similar to many of the
69: statistical physics problems redefined on the lattices, e.g., the statistics of
70: self-avoiding walks on lattices (for investigating the linear polymer
71: conformational statistics), the TSP can also be
72: defined on randomly dilute lattices. The (percolation) cluster statistics of
73: such dilute lattices is now extensively studied [7].
74: The salesman's optimized path on a dilute lattice is necessarily a
75: self-avoiding one; for optimized tour the salesman cannot afford to visit any
76: city more than once and obviously it is one where the path is non-intersecting.
77: The statistics of self-avoiding walks on dilute lattices has also been
78: studied quite a bit (see e.g., [8]). However, this knowledge
79: is not sufficient to understand the TSP on similar lattices. The TSP on dilute
80: lattices is a very intriguing one, but has not been studied intensively so far.
81:
82:
83: The lattice version of the TSP was first studied by Chakrabarti [9].
84: In the lattice version of the TSP, the $N$ cities are represented by randomly
85: occupied lattice sites of a two- dimensional square lattice ($L \times L$),
86: the fraction of sites occupied being $p$ ($=N/L^2$, the lattice occupation
87: concentration). One must then find the shortest tour in which the salesman
88: visits each city only once and comes back to its starting point. The average
89: optimal travel distance in the Euclidean metric
90: $\bar l_E$, and in the Manhattan metric $\bar l_M$, are functions of the lattice
91: occupation concentration $p$ [10]. We intend to study here the variation of the
92: normalised travel distance per city,
93: $~\Omega_E=\bar l_E \sqrt p$ and $\Omega_M=\bar l_M \sqrt p$,
94: with the lattice concentration $p$ for different system sizes.
95: It is obvious that at $p=1$, all the self-avoiding walks passing through all the
96: occupied sites will satisfy the requirements of TSP and $\Omega_E=1=\Omega_M$
97: (the distance between the neighbouring cities is equal to the unit lattice
98: constant and the path between neighbouring sites makes discrete angles of
99: of $\pi /2 $ or its multiples with the Cartesian axes). The problem becomes
100: nontrivial as $p$ decreases from unity: isolated occupied cities and branching
101: configurations of occupied cities are found here with finite probabilities and
102: self-avoiding walks through all the occupied cities, and only through the
103: occupied cities, become impossible.
104: As $p$ decreases from unity, the discreteness of the distance of the path
105: connecting the two cities and of the angle which the path makes with the Cartesian
106: axes, tend to disappear. The problem reduces to the standard TSP on the
107: continuum in the $p\rightarrow 0$ limit when all the continuous sets of
108: distances and angles become possible.
109: We study here the TSP on dilute lattice employing a
110: computer algorithm which gives the exact optimized tours for small system sizes
111: ($N\leq 100$) and near-optimal tours for bigger system sizes ($100<N\leq 256$).
112: Our study indeed indicates that $\Omega_E$ and $\Omega_M$ vary
113: with $p$ and $\Omega_E\simeq 0.73$ and $\Omega_M\simeq 0.93$ as
114: $p\rightarrow 0$.
115:
116: \noindent
117: {\bf 2 Computer Simulation and Results}\\
118: \noindent
119: We generate the randomly diluted lattice configurations following the standard
120: Monte Carlo procedure for different system sizes. For each system size $N$, we
121: vary the lattice size $L$ so that the lattice concentration $p$ varies.
122: For each such lattice configuration, the optimum tour with open boundary
123: conditions, is obtained with the help of the {\it GNU tsp\_ solve} [11]
124: developed using a branch and bound algorithm (see Fig. 1).
125: It claims to give exact results for $N\leq 100$ and near-optimal solutions for
126: $100<N\leq 256$. It may be noted that the program works
127: essentially with the Euclidean distance. However there exists a geometric
128: relationship between the Euclidean distance and the Manhattan distance. We may
129: write $l_E=\sum_{i=1}^{N} r_i$, and $l_M=\sum_{i=1}^N r_i \alpha_i$, where $r_i$
130: is the magnitude of the Euclidean path vector between two neighbouring cities
131: and $r_i\alpha_i=r_i(|\sin \theta_i |+|\cos \theta_i |)$ is the sum of the components
132: of the Euclidean path projected along the Cartesian axes.
133: Naturally, $1\leq \alpha_i \leq \sqrt 2$. If $l_E$ corresponds to the shortest
134: Euclidean path, then $\sum_{i=1}^N r_i^{\prime} > \sum_{i=1}^N r_i$ ,
135: for any other path denoted by the primed set. If the optimized Euclidean
136: path does not correspond to the optimized Manhattan path, then one will have
137: $\sum_{i=1}^N r_i^{\prime } \alpha_i^{\prime } < \sum_{i=1}^N r_i\alpha_i$,
138: where all the $\alpha_i$ and $\alpha_i^{\prime}$ satisfy the previous bounds.
139: Additionally, for random orientation of the Euclidean distance with respect to
140: the Cartesian axes, $\langle \alpha_i \rangle =\langle \alpha_i^{\prime}\rangle=
141: (2/\pi) \int_0^{\pi /2} (\sin \theta + \cos \theta ) d\theta =4/\pi $. It seems,
142: with all these constraints on $\alpha$'s and $\alpha^{\prime}$'s, it would be
143: impossible to satisfy the above inequalities on
144: $\sum r_i$, and $\sum r_i \alpha_i$. In fact, we checked for a set of $50$
145: random optimized Euclidean tours for small $N$ ($<10$), obtained using the algorithm, whether the
146: optimized Manhattan tours correspond to different sequence (of visiting the
147: cities), and did not find any. We believe that the optimized Euclidean tour
148: necessarily corresponds to the optimized Manhattan tour.
149: We then calculate $l_E$ and $l_M$ for each such optimized tour.
150:
151:
152: At each lattice concentration $p$, we take about $100$
153: lattice configurations (about 150 configurations at some special points near
154: $p\rightarrow 0$) and then obtain the averages $\bar l_E$ and
155: $\bar l_M$. We then determine $~\Omega_E=\bar l_E ~\sqrt p~$ and
156: $~\Omega_M=\bar l_M ~\sqrt p~$ and study the variations of $\Omega_E$ and
157: $\Omega_M$, and of the ratio $\Omega_M /\Omega_E$ with $p$.
158: We find that $\Omega_E$ and $\Omega_M$ both have variations starting
159: from the exact result of unity for $p=1$ to the respective constants
160: in the $p\rightarrow 0$ limit. In fact we noted that although $\Omega_M$
161: continuously decreases as $p\rightarrow 0$, it
162: remains close to unity for all values of $p$.
163: We studied the numerical results for $N=~64,~81,~100,~121,~144,~169,
164: ~196,~225~\rm {and}~256$.
165: The results for $N=64~\rm {and}~100$ have been shown in Figs. 2 and 3
166: respectively.
167: We have studied the variations in the values of $\Omega_E$ and $\Omega_M$
168: against $1/N$ for $p\rightarrow 0$, to extrapolate its value in the
169: $N\rightarrow \infty$ limit. It appears that for the large $N$ limit (see
170: Fig. 4), $\Omega_E(p\rightarrow 0)$ and $\Omega_M(p\rightarrow 0)$ eventually
171: extrapolate to $0.73\pm 0.01$ (as in
172: continuum TSP) and to $0.93\pm 0.02$, respectively.
173: This result for $\Omega_E$ (at $p\rightarrow 0$) compares very well with the
174: previous estimates [6]. As $p$ changes from $1$ to $0$, the ratio $\Omega_M /
175: \Omega_E $ changes continuously from $1$ to about $1.27~(\simeq 4/\pi )$ (see
176: Fig. 4), which is the average ratio of the Manhattan distance between two random
177: points in a plane and the Euclidean distance between them [10, 5].
178:
179:
180: \noindent
181: {\bf 3 Conclusions}\\
182: \noindent
183: We note that the TSP
184: on randomly diluted lattice is certainly a trivial problem when $p=1$ (lattice
185: limit) as it reduces to the one-dimensional TSP (the connections in the optimal
186: tour are between the nearest neighbours along the lattice).
187: Here $\Omega_E(p)=\Omega_M(p)=1$.
188: However, it is certainly NP- hard at the $p\rightarrow 0$ (continuum) limit,
189: where $\Omega_E \simeq 0.73$ and $\Omega_M \simeq 0.93$ (extrapolated for large
190: system sizes $N$). We note that
191: $\Omega_M$ remains practically close to unity for all values of
192: $p<1$. Our numerical results also suggest that $\Omega_M/\Omega_E\simeq 4/\pi$
193: as $p\rightarrow 0$.
194: It is clear that the problem crosses from triviality (for $p=1$) to the NP- hard
195: problem (for $p\rightarrow 0$) at a certain value of $p$.
196: We did not find any irregularity in the variation of $\Omega$ at any $p$.
197: A naive expectation might be that around
198: the percolation point, beyond which the marginally connected lattice spanning
199: path is snapped off [7], the $\Omega_E$ or $\Omega_M$ suffers some irregularity.
200: The absence of any such irregularity can also be justified easily: the
201: travelling salesman
202: has to visit all the occupied lattice sites (cities), not necessarily those on
203: the spanning cluster. Also, the TSP on dilute lattices has got to accomodate
204: the same kind of frustration as the (compact) self-avoiding chains on dilute
205: (percolating) lattices, although there the (collapsed) polymer is confined only
206: to the spanning cluster. This indicates that the
207: transition occurs either at $p=1_-$ or at $p=0_+$. From the consideration of
208: frustration for the TSP even at $p=1_-$, it is almost certain that the
209: transition occurs at $p=1$.
210: However, this point requires further investigations.
211:
212: \vskip 0.3in
213: \noindent
214: {\bf Acknowledgement} : We are grateful to O. C. Martin and A. Percus
215: for very useful comments and suggestions.
216:
217: \newpage
218: \noindent
219: {\bf References}\\
220: \vskip .1in
221: \noindent
222: {\it e-mail addresses} :
223:
224: \noindent
225: $^{(1)}$anirban@cmp.saha.ernet.in
226:
227: \noindent
228: $^{(2)}$bikas@cmp.saha.ernet.in
229: \vskip .2 in
230:
231: \noindent
232: 1. M. R. Garey and D. S. Johnson, {\it Computers and Intractability: A Guide
233: to the Theory of NP- Completeness} (Freeman; San Franscisco) (1979).\\
234: \noindent
235: 2. S. Kirkpatrick, C. D. Gelatt, Jr., and M. P. Vecchi, {\it Science},
236: {\bf 220}, 671 (1983).\\
237: \noindent
238: 3. M. Mezard, G. Parisi and M. A. Virasoro, {\it Spin Glass Theory and Beyond}
239: (World Scientific; Singapore) (1987).\\
240: \noindent
241: 4. J. Beardwood, J. H. Halton and J. M. Hammersley, {\it Proc. Camb. Phil.
242: Soc.} {\bf 55}, 299 (1959); R. S. Armour and J. A. Wheeler, {\it Am. J. Phys.}
243: {\bf 51}, 405 (1983).\\
244: \noindent
245: 5. A. Chakraborti and B. K. Chakrabarti, {\it cond-mat/0001069} (2000).\\
246: \noindent
247: 6. A. Percus and O. C. Martin, {\it Phys. Rev. Lett.}, {\bf 76}, 1188 (1996).\\
248: \noindent
249: 7. D. Stauffer and A. Aharony, {\it Introduction to Percolation Theory} (Taylor
250: and Francis; London) (1985).\\
251: \noindent
252: 8. K. Barat and B. K. Chakrabarti, {\it Phys. Rep.}, {\bf 258}, 377 (1995).\\
253: \noindent
254: 9. B. K. Chakrabarti, {\it J. Phys. A: Math. Gen.}, {\bf 19}, 1273 (1986).\\
255: \noindent
256: 10. D. Dhar, M. Barma, B. K. Chakrabarti and A. Tarapder,
257: {\it J. Phys. A: Math. Gen.}, {\bf 20}, 5289 (1987);
258: M. Ghosh, S. S. Manna and B. K. Chakrabarti,
259: {\it J. Phys. A: Math. Gen.}, {\bf 21}, 1483 (1988);
260: P. Sen and B. K. Chakrabarti,
261: {\it J. Phys. (Paris)}, {\bf 50}, 255, 1581 (1989).\\
262: \noindent
263: 11. C. Hurtwitz, {\it GNU tsp\_ solve}, available at: http://www.cs.sunysb.edu/\~\\
264: \noindent
265: algorith/implement/tsp/implement.shtml\\
266:
267: \newpage
268: \noindent
269: {\bf Figure captions}\\
270: \vskip .1 in
271: \noindent
272: {\bf Fig. 1} : A typical TSP for ($N=)~64$ cities on a
273: dilute lattice of size $L=30$. The cities are represented by black dots which
274: are randomly occupied sites of the lattice with concentration $p=N/L^2\simeq
275: 0.07$. The optimized Euclidean path is indicated.\\
276: \noindent
277: {\bf Fig. 2} : Plot of $\Omega_E$, $\Omega_M$ and $\Omega_M/\Omega_E$ against
278: $p$ for $N=64$ cities, obtained using the optimization programs (exact).
279: The error bars are due to configurational fluctuations.
280: The extrapolated values of $\Omega_E$, $\Omega_M$ and $\Omega_M/\Omega_E$ are
281: indicated by horizontal arrows on the y-axis.\\
282: \noindent
283: {\bf Fig. 3} : Plot of $\Omega_E$, $\Omega_M$ and $\Omega_M/\Omega_E$ against
284: $p$ for $N=100$ cities, obtained using the optimization programs (exact).
285: The error bars are due to configurational fluctuations.
286: The extrapolated values of $\Omega_E$, $\Omega_M$ and $\Omega_M/\Omega_E$ are
287: indicated by horizontal arrows on the y-axis.\\
288: \noindent
289: {\bf Fig. 4 } : Plots of $\Omega_E$ , $\Omega_M $ and of $\Omega_M/\Omega_E$ in
290: the $p\rightarrow 0$ limit, against $1/N$.
291: The error bars are due to configurational fluctuations.
292: The extrapolated value of $\Omega_E$ , $\Omega_M$ and $\Omega_M/\Omega_E$ in
293: this $p\rightarrow 0$ limit for $N\rightarrow \infty$ are indicated by
294: horizontal arrows on the y-axis.\\
295:
296: \end{document}
297: