ECS 289 / MAE 298, Network Theory Spring 2014 Problem Set # 1, Solutions Problem 1: Power Law Degree Distributions (Continuum approach) Consider the power law distribution p(k) = Ak −γ , with support (i.e., defined from) k = 1 to k → ∞. a) Show that we must have γ > 1 for this to be a properly defined probability distribution function (pdf ). Recall a pdf must have two properties: 1) p(k) ≥ 0 for all k, and 2) it must be normalized. ∞ Z Ak 1 −γ ∞ i −A h −A (−γ+1) −γ+1 k ( lim k )−1 = dk = (γ − 1) (γ − 1) k→∞ k=1 1) If γ < 1, the limit (limk→∞ k −γ+1 ) → ∞. 2) If γ = 1, the prefactor −A (γ−1) → ∞. 3) If γ > 1, then the integral is finite: only converges if γ > 1. R∞ 1 Ak −γ dk = A . (γ−1) In other words, the integral b) Solve for the normalization constant A. R∞ A From (3) above, we find 1 Ak −γ = (γ−1) for γ > 1. Setting the integral equal to 1, implies: A = (γ − 1). c) Show that if 1 < γ ≤ 2, the average value hki diverges. Z hki = ∞ (γ − 1)k −γ+1 1 ∞ i (γ − 1) (−γ+2) (γ − 1) h −γ+2 dk = − k = − ( lim k ) − 1 (γ − 2) (γ − 2) k→∞ k=1 1) If γ < 2, the limit (limk→∞ k −γ+2 ) → ∞. 2) If γ = 2, the prefactor (γ−1) (γ−2) → ∞. 3) If γ > 2, then the integral is finite: integral only converges if γ > 2. R∞ 1 1 (γ − 1)k −γ+1 dk = (γ−1) . (γ−2) In other words, the d) Show that if 2 < γ ≤ 3, the average is finite, but the standard deviation diverges. 2 k = Z ∞ (γ − 1)k −γ+2 1 ∞ i (γ − 1) (−γ+3) (γ − 1) h −γ+3 dk = − k = − ( lim k ) − 1 (γ − 3) (γ − 3) k→∞ k=1 1) If γ < 3, the limit (limk→∞ k −γ+3 ) → ∞. 2) If γ = 3, the prefactor (γ−1) (γ−3) → ∞. 3) If γ > 3, then the integral is finite: integral only converges if γ > 3. R∞ 1 (γ − 1)k −γ+2 dk = (γ−1) . (γ−3) In other words, the Thus σ 2 = hk 2 i − hki2 → ∞ for γ ≤ 3. In general, for a power law of the form p(k) ∝ k −γ , the moments hk n i diverge for all n ≥ (γ − 1). e) Plot p(k) = Ak −γ , for k = 1 to k = 100, 000 for γ = 3 (and properly normalize). Use matlab, R, or pen and paper, etc (but make sure to label axes clearly with values). The simple approach suggests that p(k) = 2k −3 since γ = 3. But the solution p(k) = 2k −3 is clearly problematic since it gives probability p(k = 1) > 1. R∞ P To properly normalize consider that this requires ∞ k=1 p(k) = 1, not k=1 p(k)dk = 1 as used in part (b). We can solve for the normalization constant numerically using a simple program P such as this C program which implements S = 100000 k −3 : k=1 2 #i n c l u d e <s t d i o . h> #i n c l u d e <math . h> main ( ) { do ub le x , S = 0 . ; long i ; f o r ( i =1; i $\ l e $ 1 0 0 0 0 0 ; ++i ) S+=(pow ( i , − 3 . ) ) ; p r i n t f ( ‘ % 0. 8 l f ’ , S ) } The inverse normalization constant, A−1 = S = 1.20206. (Note the “discrete” treatment – in detail next – gives this value). The correctly normalized plot is: 1 f(x) 0.01 0.0001 1e-06 1e-08 1e-10 1e-12 1e-14 1e-16 1 10 100 1000 10000 100000 f ) In a finite network with N nodes, what is the largest possible value of degree, kmax , that can ever be observed? So can we ever have hki → ∞ in a finite network? In a finite network of N nodes, kmax = (N − 1) (or N if we allow self-loops). Regardless, we can never observe hki → ∞ in a finite network. 3 Problem 1: Power Law Degree Distributions (Discrete approach) In most situations in networks, the k’s take on only discrete values. We can do a similar treatment, where now we consider k discrete: ∞ X p(k) = k=1 ∞ X k=1 0 −γ Ak 1 1 1 = A 1 + γ + γ + γ + ··· . 2 3 4 0 The sum in the brackets is known as the Riemann Zeta Function, RZ(γ). The value of RZ(γ), for many values of γ can be found in standard references (e.g., Mathworld, Wikipedia, etc). Commonly tabulated values are: RZ(0) = −1/2 RZ(1/2) ≈ −1.460 RZ(1) = ∞ RZ(3/2) ≈ 2.612 RZ(2) = π 2 /6 ≈ 1.645 RZ(3) ≈ 1.202 RZ(4) = π 4 /90 ≈ 1.082 Clearly, for γ < 0, RZ(γ) → ∞. Though the proof is beyond the scope of this assignment, the relevant information is that RZ(γ) < 0 for 0 ≤ γ < 1, RZ(1) → ∞, and RZ(γ) converges to a finite positive number for all γ > 1. P∞ P 0 −γ = A0 · RZ(γ). This is negative for 0 ≤ γ < 1, and it diverges (a) ∞ k=1 A k k=1 p(k) = P∞ for γ = 1. Thus for k=1 p(k) to converge to a finite positive number requires γ > 1. (b) Normalization constant, A0 = 1/RZ(γ). P 2 3 3 0 (c) hki = ∞ 1 kp(k) = A 1 + 2γ + 3γ + 4γ + · · · 1 1 1 = A0 1 + 2γ−1 + 3γ−1 + 4γ−1 + · · · = A0 RZ(γ − 1) = RZ(γ − 1)/RZ(γ), which is negative or diverges for γ − 1 ≤ 1 or equivalently, γ ≤ 2. Thus hki only converges to a positive finite number for γ > 2. (d) Likewise, hk 2 i = A0 RZ(γ − 2) = RZ(γ − 2)/RZ(γ), which is negative for 2 ≤ γ < 3 and diverges for γ = 3. Thus σ 2 = hk 2 i − hki2 = RZ(γ − 2)/RZ(γ) − [RZ(γ − 1)/RZ(γ)]2 only converges to a positive finite number for γ > 3. (e) For plot, see page 3. (f) In a finite network of N nodes, kmax = (N − 1) (or N if we allow self-loops). Regardless, we can never observe hki → ∞ in a finite network. 4 Problem 2: Random walk on a network (a) Let Aij = 1 in the following matrix denote the presence of a link from node j to node i and Aij = 0 denote the absence of a link. The adjacency matrix is given by 0 1 0 0 1 0 0 1 0 0 (1) A= 1 0 0 1 1 0 1 1 0 0 0 0 0 0 0 (b) In order to obtain the steady state distribution we normalize the adjacency matrix given above to obtain the transition matrix T . Entry Tij denotes the probability of going from node j to node i in the next time step. 0 0.5 0 0 0.5 0 0 0.5 0 0 T = (2) 0 1 0.5 1 0 0 0.5 0.5 0 0 0 0 0 0 0 The steady state probability corresponds to the eigenvector of T with eigenvalue 1. Using a standard software package we find that the eigenvector that corresponds to the eigenvalue of 1 is 0.182574 0.365148 π= (3) 0.730296 0.547722 0 5 Note that the above eigenvector is normalized such that the sum of the square of the elements is equal to 1. Renormalizing such that the sum of the terms is 1 we obtain 0.1 0.2 π= (4) 0.4 0.3 0 (c) For any undirected graph, the steady state probability of finding a random walker is directly proportional to the degree of the node. Therefore, 3 3 1 4 (5) π= 14 2 2 6