cond-mat0201430/sf.tex
1: \documentclass[prb,twocolumn,groupedaddress,showpacs]{revtex4}
2: \usepackage{graphicx}
3: \bibliographystyle{prsty}
4: \begin{document}
5: \title{The network topology of a potential energy landscape: A static scale-free network}
6: \author{Jonathan P.~K.~Doye}
7: \email{jon@clust.ch.cam.ac.uk}
8: \affiliation{University Chemical Laboratory, Lensfield Road, Cambridge CB2 1EW, United Kingdom}
9: \date{\today}
10: \begin{abstract}
11: Here we analyze the topology of the network formed by the minima and transition states on the
12: potential energy landscape of small clusters. We find that this network has both a small-world
13: and scale-free character. 
14: In contrast to other scale-free networks, where the topology results from the 
15: dynamics of the network growth, the potential energy landscape is a static entity.
16: Therefore, a fundamentally different organizing principle underlies this behaviour:
17: The potential energy landscape is highly heterogeneous with the low-energy minima 
18: having large basins of attraction and acting as the highly-connected hubs in the network. 
19: \end{abstract}
20: \pacs{89.75.Hc,61.46.+w,31.50.-x}
21: \maketitle
22: 
23: Energy landscapes have been at the forefront of many of the recent theoretical
24: developments in our understanding of biomolecules,\cite{Brooks01} clusters\cite{WalesMW98,WalesDMMW00} 
25: and the glass transition.\cite{Debenedetti01}
26: For example, this research has provided important new insights into how 
27: proteins fold\cite{Bryngelson95} and the origin of the unusual properties of supercooled liquids,
28: such as the distinction between ``strong'' and ``fragile'' liquids.\cite{Sastry01,SaikaVoivod01}
29: There has also been a surge of interest in modelling complex systems as
30: networks,\cite{Strogatz01} inspired by Watts and Strogatz's discovery that many 
31: networks behave as ``small worlds''.\cite{Watts98}
32: Intriguingly, a diverse range of such networks, e.g.\ the world-wide web,\cite{Albert99}
33: the internet,\cite{Faloutsos99}  scientific collaboration\cite{Newman01a} and citation\cite{Redner98} networks, 
34: and biochemical networks,\cite{Jeong00,Jeong01} 
35: also have a ``scale-free'' topology, where 
36: the distribution of the number of connections to each node, the degree, follows a power law.
37: This topology results from the dynamics of the network growth in these systems.\cite{Barabasi99}
38: Here we draw these two strands of research together by applying the techniques of
39: network analysis to probe the global structure of potential energy landscapes of clusters.
40: 
41: 
42: The potential energy landscape is a multi-dimensional surface 
43: representing the dependence of the potential energy on the 
44: positions of all the atoms of the system. 
45: For a system with many atoms the landscape will have a complex topography with 
46: higher-dimensional analogues of mountain ranges, valleys and passes.
47: Although the potential energy landscape determines the system's structure, thermodynamics
48: and dynamics, the nature of this relationship is complex. 
49: A particularly successful means of elucidating this relationship
50: is the inherent structure approach of Stillinger and Weber,\cite{StillW84a} 
51: in which the landscape is divided into basins of attraction surrounding 
52: each minimum (See Fig.\ \ref{fig:PES}).
53: This partition provides a natural way to describe the dynamics of the system, because, except
54: at high temperature, the system spends most of the time vibrating in the well surrounding
55: a minimum and only occasionally hops to a different well by passing over a transition state.
56: The interbasin dynamics can then be represented as a walk on a network whose
57: nodes correspond to the minima and where edges link those minimum which are directly connected
58: by a transition state.
59: Figure \ref{fig:PES} provides an illustration of such an inherent structure network (ISN)
60: for a two-dimensional energy surface. 
61: 
62: \begin{figure}
63: \includegraphics[width=8.6cm]{PES_d.eps}
64: \caption{
65: (a) A model two-dimensional energy surface. 
66: (b) A contour plot of this surface
67: illustrating the inherent structure partition of the configuration space 
68: into basins of attraction surrounding minima.  
69: The basin boundaries are represented by red lines, 
70: and the minima and transition states by blue and green points, respectively.
71: (c) The resulting representation of the landscape as a network.
72: }
73: \label{fig:PES}
74: \end{figure}
75: 
76: The ISN should provide the starting point for an energy landscape view
77: of the global dynamics of a system. Indeed, as is increasingly being done,
78: it is relatively easy to calculate the dynamics from this network using 
79: a master equation approach.\cite{BerryK95,Cieplak98a,WalesDMMW00} 
80: However, fundamental questions about the topology of the ISN have received little attention.
81: By contrast the global topography of energy landscapes has been the 
82: focus of much research.\cite{Brooks01,WalesMW98,Bryngelson95} 
83: To give one example, this emphasis has revealed that when
84: a landscape is like a `funnel'\cite{Bryngelson95} the system is guided 
85: towards the global minimum, be it the native state of a protein,\cite{Bryngelson95} 
86: an ordered nanoparticle\cite{Ball96} or a bulk crystal.
87: However, the topological aspects of this idea remain open despite their importance:
88: if the average number of steps to reach the global minimum from an
89: arbitrary starting minimum scales unfavourably with size, 
90: the location of this structure would become significantly hindered at large size.
91: 
92: \begin{figure}
93: \includegraphics[width=8.6cm]{smallworld.eps}
94: \caption{
95: (a) The dependence of the average separation between nodes (in steps) 
96: on the size of the network, $N_{\rm min}$. 
97: (b) The size dependence of the clustering coefficient, $c$, compared to that for a random graph,
98: where $c$ is the fraction of the pairs of nodes with a common neighbour that are also 
99: connected.\cite{Watts98}
100: The data points are for Lennard-Jones clusters with from 7 to 14 atoms.
101: }
102: \label{fig:SW}
103: \end{figure}
104: 
105: To characterize the topology of the ISN 
106: we study small clusters for which we are able to locate nearly all 
107: the minima and transition states on the potential energy landscape.\cite{WalesDMMW00,Tsai93a}
108: The atoms of the cluster interact with a Lennard-Jones potential, which provides
109: a reasonable description for rare gas clusters. 
110: The numbers of minima and transition states are expected to increase roughly as 
111: $N_{\rm min}\approx e^{\alpha N}$ and 
112: $N_{\rm ts}\approx N e^{\alpha N}$, respectively,\cite{Still99,Doye02a} 
113: where $N$ is the number of atoms in the cluster.
114: Therefore, the largest network that we are able to consider is for a 14-atom cluster 
115: for which we have located 4196 minima and $87\,219$ transition states
116: 
117: Small-world networks have characteristics typical of both random graphs and lattices.
118: The average separation between nodes scales logarithmically with network size,
119: while the network is highly clustered, i.e.\ any two neighbours of a node 
120: are also likely to be connected. 
121: From Fig.\ \ref{fig:SW} it is clear that the ISNs for the clusters
122: show both these features and so are small worlds. 
123: The clustering is unsurprising given that the connections between basins on a 
124: potential energy landscape are based on adjacency in configuration space,\cite{cluster} 
125: but to interpret Fig.\ \ref{fig:SW}(a) properly we must take into account the increase in both the 
126: dimension of configuration space and the average degree, $\langle k\rangle$, 
127: as the size of the network increases. 
128: 
129: For example, for a hypercubic lattice 
130: with a constant number of lattice points, $L$, along each edge
131: and dimension $3N$, 
132: the number of lattice points, $N_{\rm latt}$, 
133: scales exponentially with $N$ and 
134: the average number of steps between any two lattice points is 
135: $3N(L+1)/3=(L+1)\log N_{\rm latt}/ 3 \log L$.
136: By contrast, if the network behaves as a random graph, 
137: the average separation should scale as $\log N_{\rm min}/\log \langle k\rangle\propto 
138: \log N_{\rm latt}/\log(\log N_{\rm latt})$ 
139: because $\langle k\rangle\propto N_{\rm ts}/N_{\rm min}\propto N$. 
140: The sublogarithmic scaling suggested by 
141: Fig.\ \ref{fig:SW}(a) points to the latter scenario.
142: This result is somewhat surprising.
143: In Watts and Strogatz's small-world model the random-graph character results from the 
144: introduction of random links into the lattice, which can potentially connect up distant nodes, 
145: but there is no obvious equivalent of the random links on the potential energy landscapes.
146: 
147: \begin{figure}
148: \includegraphics[width=8.6cm]{degree.eps}
149: \caption{
150: The cumulative distribution for the number of nodes that have more
151: than $k$ connections. The curves correspond to clusters of different sizes, as labelled.
152: An additional straight line with slope $-(\gamma-1)$, where $\gamma$=2.78, has
153: been plotted to emphasise the power law tail.
154: In the inset the cumulative probability distribution for the 12-, 13- and 14-atom clusters
155: is plotted against $k$ normalized by its average value, $\langle k \rangle$, to 
156: bring out the universal form of the distribution.
157: }
158: \label{fig:degree}
159: \end{figure}
160: 
161: If we now examine the distributions for the numbers of connections for each node
162: we find that as the size of the cluster increases a clear power-law tail develops, which
163: has a universal form independent of the cluster size (Fig.\ \ref{fig:degree}). 
164: The exponent of the power law, $\gamma$=2.78, is similar to other scale-free networks.\cite{Albert01}
165: The cause of the random-graph like scaling of the average separation is thus
166: the scale-free topology of the ISN.
167: The network is extremely heterogeneous with a few hubs that have
168: a very large number of connections, but with the majority of nodes only connected to 
169: a relatively small number of other minima.
170: 
171: This is a particularly surprising result 
172: because all other scale-free networks are dynamic in origin. 
173: They grow and change over time, be it on an almost continuous basis as in the WWW 
174: or on evolutionary time scales in the case of biochemical networks. 
175: Even the recently introduced deterministic scale-free networks are based on an
176: iterative growth procedure.\cite{Barabasi01,Dorogovtsev02}
177: Furthermore, models of network growth\cite{Barabasi99,Albert01} 
178: and studies on the time evolution of real networks\cite{Newman01d,Jeong01b} strongly suggest that
179: the heterogeneity at the heart of the scale-free topology develops as 
180: a result of new nodes preferentially linking to those nodes which 
181: have many connections, be they much-cited papers or popular web-sites.
182: However, the network associated with a potential energy landscape is static. It is simply
183: determined by the form of the interatomic interactions and the number of atoms in the cluster.
184: 
185: The source of the heterogeneity in the ISNs is apparent from Fig.\ \ref{fig:LJ14},
186: where we see that the degree of a node increases somewhat faster than exponentially 
187: as the energy of the minimum decreases.\cite{mindist} 
188: The low-energy minima act as the hubs in the network. 
189: Thus, for the 14-atom cluster, 76\% of the nodes in the network are connected to the global minimum.
190: To measure the extent of the catchment basin around a minimum,
191: we can calculate the distance in configuration space to all the 
192: transition states connected to a minimum.
193: For LJ$_{14}$ we find that this distance is 2.7 times larger for the global minimum than
194: for the surrounding minima. 
195: When the multi-dimensionality of configuration space is taken into account this result
196: suggests that the hyperarea of this catchment basin is many orders of magnitude 
197: larger than the average.
198: Similarly, it has been previously found that on average the basin area falls off 
199: approximately exponentially with the energy of a minimum.\cite{Doye98e}
200: 
201: \begin{figure}
202: \includegraphics[width=8.6cm]{LJ14_b.eps}
203: \caption{
204: The dependence of the degree of a node on the potential energy 
205: of the corresponding minimum for the 14-atom cluster. The data points are 
206: for each individual minimum and the solid line is a binned average.
207: }
208: \label{fig:LJ14}
209: \end{figure}
210: 
211: These results show that the global topology and topography of the potential energy landscape 
212: are intimately connected because the deep minima have very large catchment basins which are 
213: connected to lots of smaller basins around their edges. 
214: By contrast, if a potential energy landscape were flat and all basins of 
215: attraction had the same area the scale-free topology would be lost.
216: For example, an investigation of the network topology of the configuration space 
217: of a {\em non-interacting} lattice polymer\cite{Scala01b} found 
218: the connectivity distribution to be a Gaussian.\cite{Amaral00} 
219: 
220: This contrasting example naturally leads one to ask how general is the topology 
221: that we have found for the Lennard-Jones clusters.
222: Although such a question can only be definitively answered by similar analyses for a 
223: variety of systems, there is nothing ``special'' about the Lennard-Jones potential
224: and so there is no reason why similar behaviour should not be seen for other 
225: materials, as long as there are no constraints present that would
226: prevent the formation of the high degrees associated with the hubs.  
227: A polymer provides an example of the latter, because the connectivity of the chain 
228: limits the number of transition states that can surround a minimum. 
229: For example, a similar analysis for Lennard-Jones polymers\cite{Calvo02a} 
230: did not find a power-law tail to the degree distribution 
231: (although it was still longer than exponential).\cite{JDunpub}
232: There is no equivalent of many of the transition states for the equivalent Lennard-Jones 
233: cluster because they involve the breaking of the polymer chain.
234: 
235: The scale-free topology of the ISN is 
236: potentially good news for global optimization, the task of locating the global minimum.
237: Even though the number of minima increases exponentially with the size of the system,\cite{Still99}
238: the average number of steps in the shortest path to the global minimum grows 
239: sublinearly with system size. Of course, finding this path is not necessarily easy.
240: Our calculations of the shortest paths required information on the global structure of the 
241: potential energy landscape, whereas a global optimization algorithm usually takes a
242: step based on only local information.
243: 
244: Some path finding strategies to efficiently navigate scale-free networks have been suggested
245: that make use of the fact that most of the shortest paths pass through the highly-connected
246: hubs.\cite{Adamic01,Kim02} In our case the clear link between the topology and 
247: topography of the potential energy landscape provides an additional advantageous strategy. 
248: A downhill step to a lower-energy minimum is likely to take 
249: one to a minimum that is more connected and closer to the global minimum.
250: How well-obeyed the latter correlation is, depends upon the global topography
251: of the potential energy landscape and is a good indicator of the difficulty
252: of global optimization. Thus, when the landscape is like a single funnel, global 
253: optimization algorithms can achieve near to the ideal scaling.
254: For example, the basin-hopping algorithm can locate the global minimum 
255: of the 55-atom Lennard-Jones cluster after on average less than 150 minimizations 
256: when started from a random configuration, 
257: even though there are an estimated $10^{21}$ minima.\cite{Doye98e} 
258: The increasing number of links as the energy decreases evident in Fig.\ \ref{fig:LJ14}
259: further adds to the efficacy of a funnel in guiding the system towards its bottom,
260: and provides clear evidence of the convergence of pathways into the hub at 
261: the funnel bottom that is often postulated.
262: By contrast, when an energy landscape has multiple funnels 
263: and there is a tendency to enter a funnel that leads the system to a low-energy minimum
264: that is far from the global minimum, global optimization can be very difficult.
265: 
266: The topology of the ISN will of course significantly affect the dynamics. 
267: This connection can be probed for very small systems where the network can be completely
268: characterized and the inherent structure dynamics obtained by a master equation approach.
269: However, this approach is not practical for the system sizes that are of most interest. 
270: Therefore, models of protein folding and the glass transition usually have to assume a 
271: simplified topology for the interstate dynamics,\cite{Shak89,Kohen00} 
272: or relate the dynamics to static quantities through phenomenological equations, 
273: such as the Adam-Gibbs equation which relates the relaxation time in supercooled liquids 
274: to the configurational entropy.\cite{AdamG65}
275: To fully unlock the potential insights from the inherent structure view of the dynamics, a means
276: of statistically modelling the network topology from a partial characterization of 
277: the potential energy landscape is therefore needed.
278: Our results could significantly advance this goal.
279: 
280: The author is grateful to Emmanuel College, Cambridge 
281: and the Royal Society for financial support.
282: 
283: \begin{thebibliography}{10}
284: 
285: \bibitem{Brooks01}
286: C.~L. Brooks~III, J.~N. Onuchic, and D.~J. Wales, Science {\bf 293},  612
287:   (2001).
288: 
289: \bibitem{WalesMW98}
290: D.~J. Wales, M.~A. Miller, and T.~R. Walsh, Nature {\bf 394},  758  (1998).
291: 
292: \bibitem{WalesDMMW00}
293: D.~J. Wales {\it et~al.}, Adv. Chem. Phys. {\bf 115},  1  (2000).
294: 
295: \bibitem{Debenedetti01}
296: P.~G. Debenedetti and F.~H. Stillinger, Nature {\bf 410},  259  (2001).
297: 
298: \bibitem{Bryngelson95}
299: J.~D. Bryngelson, J.~N. Onuchic, N.~D. Socci, and P.~G. Wolynes, Proteins:
300:   Structure, Function and Genetics {\bf 21},  167  (1995).
301: 
302: \bibitem{Sastry01}
303: S. Sastry, Nature {\bf 409},  164  (2001).
304: 
305: \bibitem{SaikaVoivod01}
306: I. Saika-Voivod, P.~H. Poole, and F. Sciortino, Nature {\bf 412},  514  (2001).
307: 
308: \bibitem{Strogatz01}
309: S.~H. Strogatz, Nature {\bf 410},  268  (2001).
310: 
311: \bibitem{Watts98}
312: D.~J. Watts and S.~H. Strogatz, Nature {\bf 393},  440  (1998).
313: 
314: \bibitem{Albert99}
315: R. Albert, H. Jeong, and A.~L. Barab\'{a}si, Nature {\bf 401},  130  (1999).
316: 
317: \bibitem{Faloutsos99}
318: M. Faloutsos, P. Faloutsos, and C. Faloutsos, Comput. Commun. Rev. {\bf 29},
319:   251  (1999).
320: 
321: \bibitem{Newman01a}
322: M.~E.~J. Newman, Proc. Natl. Acad. Sci. {\bf 98},  404  (2001).
323: 
324: \bibitem{Redner98}
325: S. Redner, Eur. Phys. J. B {\bf 4},  131  (1998).
326: 
327: \bibitem{Jeong00}
328: H. Jeong {\it et~al.}, Nature {\bf 407},  651  (2000).
329: 
330: \bibitem{Jeong01}
331: H. Jeong, S. Mason, A.~L. Barab\'{a}si, and Z.~N. Oltvai, Nature {\bf 411},  41
332:    (2001).
333: 
334: \bibitem{Barabasi99}
335: A.~L. Barab\'{a}si and R. Albert, Science {\bf 286},  509  (1999).
336: 
337: \bibitem{StillW84a}
338: F.~H. Stillinger and T.~A. Weber, Science {\bf 225},  983  (1984).
339: 
340: \bibitem{BerryK95}
341: R.~S. Berry and R.~E. Breitengraser-Kunz, Phys. Rev. Lett. {\bf 74},  3951
342:   (1995).
343: 
344: \bibitem{Cieplak98a}
345: M. Cieplak, M. Henkel, J. Karbowski, and J.~R. Banavar, Phys. Rev. Lett. {\bf
346:   80},  3654  (1998).
347: 
348: \bibitem{Ball96}
349: K.~D. Ball {\it et~al.}, Science {\bf 271},  963  (1996).
350: 
351: \bibitem{Tsai93a}
352: C.~J. Tsai and K.~D. Jordan, J. Phys. Chem. {\bf 97},  11227  (1993).
353: 
354: \bibitem{Still99}
355: F.~H. Stillinger, Phys. Rev. E {\bf 59},  48  (1999).
356: 
357: \bibitem{Doye02a}
358: J.~P.~K. Doye and D.~J. Wales, J. Chem. Phys. {\bf 116},  in press  (2002).
359: 
360: \bibitem{cluster}
361: For the smaller clusters the clustering coefficients are not that much larger
362:   than random simply because these small networks are so highly connected.
363: 
364: \bibitem{Albert01}
365: R. Albert and A.~L. Barab\'{a}si, Rev. Mod. Phys.  submitted
366:   (cond-mat/0106096).
367: 
368: \bibitem{Barabasi01}
369: A.~L. Barab\'{a}si and E. Ravasz, Physica A {\bf 299},  559  (2001).
370: 
371: \bibitem{Dorogovtsev02}
372: S.~N. Dorogovtsev, A.~V. Goltsev, and J.~F.~F. Mendes,   (cond-mat/0112143).
373: 
374: \bibitem{Newman01d}
375: M.~E.~J. Newman, Phys. Rev. E {\bf 64},  025102  (2001).
376: 
377: \bibitem{Jeong01b}
378: H. Jeong, Z. Neda, and A.~L. Barab\'{a}si,   (cond-mat/0104131).
379: 
380: \bibitem{mindist}
381: This dependence of the degree on the energy of a minimum combines with an
382:   energetic distribution of minima, which is roughly Gaussian at the most
383:   common energies\cite{Sciortino99a,Buchner99a} but with a low-energy tail that
384:   falls off less rapidly, to give the power law tail of the degree
385:   distribution.
386: 
387: \bibitem{Doye98e}
388: J.~P.~K. Doye, D.~J. Wales, and M.~A. Miller, J. Chem. Phys. {\bf 109},  8143
389:   (1998).
390: 
391: \bibitem{Scala01b}
392: A. Scala, L.~A. Nunes~Amaral, and M. Barth\'{e}l\'{e}my, Europhys. Lett. {\bf
393:   55},  594  (2001).
394: 
395: \bibitem{Amaral00}
396: L.~A.~N. Amaral, A. Scala, M. Barth\'{e}l\'{e}my, and H.~E. Stanley, Proc. Nat.
397:   Acad. Sci. USA {\bf 97},  11149  (2000).
398: 
399: \bibitem{Calvo02a}
400: F. Calvo, J.~P.~K. Doye, and D.~J. Wales, J. Chem. Phys. {\bf 116},  in press
401:   (2002).
402: 
403: \bibitem{JDunpub}
404: J.~P.~K.~Doye, unpublished.
405: 
406: \bibitem{Adamic01}
407: L.~A. Adamic, R.~M. Lukose, A.~R. Puniyani, and B.~A. Huberman, Phys. Rev. E
408:   {\bf 64},  046135  (2001).
409: 
410: \bibitem{Kim02}
411: B.~J. Kim, C.~N. Yoon, S.~K. Han, and H. Jeong,   (cond-mat/011232).
412: 
413: \bibitem{Shak89}
414: E.~I. Shakhnovich and A.~M. Gutin, Europhys. Lett. {\bf 9},  569  (1989).
415: 
416: \bibitem{Kohen00}
417: D. Kohen and F.~H. Stillinger, Phys. Rev. E {\bf 61},  1176  (2000).
418: 
419: \bibitem{AdamG65}
420: G. Adam and J.~H. Gibbs, J. Chem. Phys. {\bf 43},  139  (1965).
421: 
422: \bibitem{Sciortino99a}
423: F. Sciortino, W. Kob, and P. Tartaglia, Phys. Rev. Lett. {\bf 83},  3214
424:   (1999).
425: 
426: \bibitem{Buchner99a}
427: S. B\"{u}chner and A. Heuer, Phys. Rev. E {\bf 60},  6507  (1999).
428: 
429: \end{thebibliography}
430: 
431: \end{document}
432: 
433: