1. Scale Invariance in Complex Systems Geoff Rodgers School of Information Systems, Computing and Mathematics

2. Plan 1.Physics – equilibrium statistical mechanics. 2.Self-organised criticality. 3.Networks. 4.Financial systems. 5.Conclusions.

3. Physics Consider a classical system, like an Ising model of a magnet. This has a Hamiltonian H = - J  S i S j Consider it in equilibrium at temperature T in dimension d >2.

4. Consider the pair correlation function F(r) = ) (S j - ) >, where r = | i - j |. Simple matter to show that F(r) = r -(d-2+  ) g(r/  ) where the correlation length  behaves like  = | T - T c | - 

5. So at the critical temperature T=T c, the correlation length diverges to infinity, and the pair correlation function is power-law F(r) ~ r -(d-2+  ) i.e. scale invariant.

6. In equilibrium statistical mechanics – power laws only occur at the critical point. In non-equilibrium statistical mechanics power-laws are also quite rare.

7. Other physical quantities for this system have a scaling form, and are power-law at criticality. Also, because F(r) = r -(d-2+  ) g(r | T - T c |  ), plots of F(r)r d-2+  against r | T - T c | ,for different temperatures, lie on a universal curve g.

8. Other systems have the same scaling forms, with different scaling functions g(x) and different exponents. Systems with the same exponents have identical physics and are said to be in the same universality class. Scale free, scale invariant, power-law behaviour only occurs at criticality.

9. Self-organised criticality Simple models of avalanches Have a square lattice with grains of sand on each site. At each time step, add a grain to the central site. Then allow toppling and hence avalanches to occur.

10. Site (i,j) has z i j grains on it For all z i j  4 toppling occurs z i j  z i j – 4 z i  1 j  z i  1 j +1 z i j  1  z i j  1 +1 Lattice is finite, so grains falls off – this is the avalanche.

11. 113 123 222 113 133 222 113 143 222 123 204 232 124 210 233 130 211 233 Three time steps; Avalanche sizes 0, 0 and 3. Time step 1Time step 2Time step 3 Time step 4 130 221 233

12. Measure the avalanches over a long period of time on a large lattice – the distribution of avalanches sizes is power-law N(x) ~ x – 3/2 Such models are called self-organised critical models – there is no tunable parameter such as temperature to make them critical.

13. At first models such as this were believed to be robust – with powerful toppling kinetics that made the details of the lattice, toppling rule etc.. irrelevant, and drove the model to a large attractor state with power-law distributions. Later became apparent that introducing physical characteristics to the kinetics – for instance, friction, randomness in the toppling, anisotropy, momentum, inhomogeneous grains – destroyed the power-laws.

15. However, a lot of experimental evidence suggests that spatial and temporal correlations in avalanche systems are power-law. But these real experimental systems obviously have anisotropies, inertia, friction etc…

16. Networks A lot of networks in economic, technological and social systems have been found to have a power-law degree distribution. That is, the number of vertices N(m) with m edges is given by N(m) ~ m - 

17. Examples include Web graph Telephone call graph E-mail graph Citation graph Co-authorship graph and many others

18. Web-graph Vertices are web pages Edges are html links Degree distribution is power law

20. These graphs are all grown, i.e. vertices and edges added over time. Imagine a model system in which we add a new vertex at each time step. Connect the new vertex to an existing vertex of degree k with rate proportional to k.

21. Total degree of network = 18, 10 vertices Connect new vertex number 11 to vertex 1 with probability 5/18 vertex 2 with probability 3/18 vertex 7 with probability 3/18 all others, probability 1/18 each

22. This network is completely solvable analytically – the number of vertices of degree k at time t, n k (t), obeys the differential equation where M(t) =  kn k (t) is the total degree of the network.

23. Simple to show that as t   n k (t) ~ k -3 t and the behavior is power-law. However, adding other ingredients – for Instance aging, heterogeneities and probabilistic attachment - destroys the power-law.

24. Financial Systems Intra-day Returns Empirical analyses of financial price data show that the price of different assets deviate from Gaussian, which would be expected if the agents were acting independently.

25. Anomalously large oscillations are observed, which are distributed according to an exponentially truncated power law x -  exp { -x/x 0 } where  takes a value between 1.4 and 1.6 for a number of market indices and stocks.

26. If the price at time t, is p(t), then normally measure the return at time t defined by Z(t) = p(t+  t)-p(t) for small  t.

27. S&P 500 price changes, with Gaussian and Levy stable distribution for comparison.

28. Physically motivated models assume that this behaviour is caused by co-operative behaviour of the agents, and in particular by a phenomenon called herding or crowding. This is the traditional explanation provided by economists.

29. Models used to explain this phenomena are based on either (i)Kinetics of group aggregation and dissolution (ii)Processes on social networks (iii)The kinetics of the order book (iv)Multiplicative random exchange processes. All these models give rise to power laws with varying degrees of generality.

30. e.g. Kinetics of the order book A limit order to sell (or buy) is an instruction to sell (or buy) a share if its price rises above (falls below) the execution price of the limit order. A market order is an instruction to immediately sell or buy at the currently available price.

31. Simple model At each time step, with equal probability, place a limit order to buy a limit order to sell a market order to buy a market order to sell Each new order is placed randomly within ±  of the current price.

32. Kinetics of the Order Book

33. The probability of a price fluctuation of size x, is given by x -  with   2.0.

34. How to make sense of it all? Traditionally physicists search for the necessary and sufficient ingredients for a phenomena to occur. Then seek to define universality classes – classes of behaviour as a function of the ingredients.

35. Of course, systems considered in this talk may be too diverse to make this possible. Have a go anyway.

36. Power laws seem to be ubiquitous in a wide range of physical, technological, social and financial systems. Traditional models, be they either equilibrium or non equilibrium, only give rise to power laws in rather special circumstances, such as at a critical point.

37. The models considered in this talk were designed to understand empirical measurements of power-law distributions. The models themselves were kinetic. They can be characterised or modeled by a kinetic process. growing or driven; at each time step something is added. random – not deterministic. Thus a growing or driven random kinetic process might be regarded as necessary for power-laws distribution.

38. Although these models exhibit power laws for a wide range of some parameters, the power- laws are destroyed by the addition of some ingredients. Unfortunately, these ingredients, such as (i)Inertia in sand piles (ii)Randomness in the www or collaboration graph (iii)Strategy scoring in the stock market are all present in the real systems.

39. Hence the question of why scale free behaviour is observed is such a wide range of systems still remains open.

41. Seminar next week Networks: Structure and Transport. Bosa Tadic, Institute Josef Stefan, Ljubljana. Wednesday 4pm, Maths 128.