Abstract
We prove logarithmic upper bounds for the diameters of the random-surfer Webgraph model and the PageRank-based selection Webgraph model, confirming the small world phenomenon holds for them. In the special case when the generated graph is a tree, we provide close lower and upper bounds for the diameters of both models.
Similar content being viewed by others
References
Barabási, A.-L., Albert, R.: Emergence of scaling in random networks. Science 286(5439), 509–512 (1999)
Bhamidi, S.: Universal techniques to analyze preferential attachment trees: global and local analysis. (2007). (preprint) http://www.unc.edu/~bhamidi/
Bianconi, G., Barabási, A.-L.: Competition and multiscaling in evolving networks. Europhys. Lett. 54(4), 436–442 (2001)
Blum, A., Chan, T.-H.H., Rwebangira, M.R.: A random-surfer web-graph model. In: Proceedings of the 8th Workshop on Algorithm Engineering and Experiments and the 3rd Workshop on Analytic Algorithmics and Combinatorics, ALENEX/ANALCO ’06, pp. 238–246 (2006)
Bollobás, B., Riordan, O.: The diameter of a scale-free random graph. Combinatorica 24(1), 5–34 (2004)
Bonato, A., Chung, F.: Complex networks. In: Gross, J.L., Yellen, J., Zhang, P. (eds.) Handbook of Graph Theory, Chapter 12.1, 2nd edn, pp. 1456–1476. Chapman & Hall, London (2013)
Broutin, N., Devroye, L.: Large deviations for the weighted height of an extended class of trees. Algorithmica 46(3–4), 271–297 (2006)
Chakrabarti, D., Faloutsos, C.: Graph mining: laws, generators, and algorithms. ACM Comput. Surv. 38(1), (2006). Article 2
Chebolu, P., Melsted, P.: Pagerank and the random surfer model. In: Proceedings of the 19th Annual ACM-SIAM Symposium on Discrete Algorithms. SODA ’08, pp. 1010–1018. Philadelphia, PA, USA (2008)
Dembo, A., Zeitouni, O.: Large Deviations Techniques and Applications, Volume 38 of Stochastic Modelling and Applied Probability. Springer, Berlin (2010). Corrected reprint of the second (1998) edition
Devroye, L., Fawzi, O., Fraiman, N.: Depth properties of scaled attachment random recursive trees. Random Struct. Algorithms 41(1), 66–98 (2012)
Dommers, S., van der Hofstad, R., Hooghiemstra, G.: Diameters in preferential attachment models. J. Stat. Phys. 139(1), 72–107 (2010)
Drinea, E., Frieze, A., Mitzenmacher, M.: Balls and bins models with feedback. In: Proceedings of the 13th Annual ACM-SIAM Symposium on Discrete Algorithms. SODA ’02, pp. 308–315. Philadelphia, PA, USA (2002)
Ergün, G., Rodgers, G.J.: Growing random networks with fitness. Phys. A Stat. Mech. Appl. 303(1–2), 261–272 (2002)
Krapivsky, P.L., Redner, S.: Organization of growing random networks. Phys. Rev. E 63, 066123 (2001)
Leskovec, J., Kleinberg, J., Faloutsos, C.: Graph evolution: densification and shrinking diameters. ACM Trans. Knowl. Discov. Data (TKDD) 1(1), (2007). Article 2
Pandurangan, G., Raghavan, P., Upfal, E.: Using pagerank to characterize web structure. Internet Math. 3(1), 1–20 (2006). Conference version in COCOON 2002
Pittel, B.: Note on the heights of random recursive trees and random \(m\)-ary search trees. Random Struct. Algorithms 5(2), 337–347 (1994)
Acknowledgments
The authors thank the referees for their careful readings of the manuscript and their many useful comments.
Author information
Authors and Affiliations
Corresponding author
Additional information
Abbas Mehrabian: Supported by the Vanier Canada Graduate Scholarships program.
Nick Wormald: Supported by Australian Laureate Fellowships Grant FL120100125.
A preliminary version of this paper has appeared in Proceedings of the 18th International Workshop on Randomization and Computation (RANDOM’14), pages 857–871.
Appendix: Omitted Proofs
Appendix: Omitted Proofs
Proof of Lemma 6
We first prove the upper bound. If \(x>1\) then \(\exp (-\Upsilon (x) m)=1\), so we may assume that \(0 < x \le 1\). We use Chernoff’s technique. Let \(\theta = 1 - 1/x\). Then we have
We now prove the lower bound. If \(x>1\), then the result follows from Markov’s inequality, so we may assume that \(0<x\le 1\). Let \(\Lambda ^*(x) = \sup \{ \lambda x - \log (\mathbb {E}\left[ {e^{\lambda E_1}}\right] ) : \lambda \le 0\} \). Since \(\mathbb {E}\left[ {e^{\lambda E_1}}\right] =1/(1-\lambda )\) for all \(\lambda <1\), the supremum here occurs at \(\lambda = 1 - 1/x\), which implies \(\Lambda ^*(x)=\Upsilon (x)\). Then by Cramér’s Theorem (see, e.g., [10, Theorem 2.2.3, p. 27]) we have
as required. \(\square \)
Proof of Lemma 7
We use Chernoff’s technique. Let \(\theta \) satisfy
We have
Thus we have
\(\square \)
Proof of Lemma 8
We consider two cases.
Case 1: \(c\ge 1\). In this case we prove
Notice that we have \(1-c\Upsilon (1/c) = c + c \log (1/c)\), so, as \(\eta (1-p)\le \eta -c\), the conclusion is implied by
Dividing by c and since \(c>0\), this statement is equivalent to
where \(r:=\eta /c\). Exponentiating and plugging the definition of f, this statement is equivalent to
Since \((r/(r-1))^{r-1} < e\), and \(1-p < (1-p^3)e^{-p}\), for this inequality to hold it suffices to have
which follows from the fact that \(x^2 e^{-x} \le 4e^{-2}\) for all \(x\ge 1\).
Case 2: \(c< 1\). In this case we prove
Since \(\Upsilon (1/c)=0\), this is equivalent to
Note that
so we have
where we have used \(\log (1-p)\le -p\) in the last inequality. Hence to prove (36), since \(c>0\), it suffices to show that
Since \( p\eta \ge 4e^p\ge 4>4c\), we have
which imply (37). \(\square \)
Proof of Lemma 14
(a) The conclusion is clear for \(a\in \{0,1\}\), so we may assume that \(a\in (0,1)\). Since \(\Phi (a,a)<0\) and \(\Phi (a,1)>0\), there exists at least one \(s\in (a,1)\) with \(\Phi (a,s)=0\). We now show that there is a unique such s. Fixing a, since \(\Phi \) is differentiable with respect to s, it is enough to show that
Let \(\sigma = p(1-p) \). We have
At a point (a, s) with \(\Phi (a,s)=0\), we have
so at this point,
which is strictly positive because
and this proves (38).
(b) Plugging the definition of f from (4) and using \(\Phi (a,s)=0\) gives this equation.
(c) We first show that \(\phi \) is differentiable and increasing on (0, 1). Let \(a\in (0,1)\) and let \(s=\phi (a)\). We have
and \(\partial \Phi / \partial s\) is positive as proved in part (a). Hence by the implicit function theorem \({\mathrm {d}s}/{\mathrm {d}a}\) exists and is positive, so \(\phi \) is differentiable and increasing on (0, 1). Since \(\phi (0)=0\) and \(\phi (1)=1\), \(\phi \) is increasing on [0, 1].
(d) Let \(s\in [0,1]\). Then \(\Phi (0,s) \Phi (s,s) \le 0\) and so there exists at least one \(a_0\in [0,s]\) with \(\Phi (a_0,s)=0\). The function \(\Phi (a,s)\) is linear in a and the coefficient of a is non-zero, hence this root \(a_0\) is unique. The function \(\phi ^{-1}\) is increasing since \(\phi \) is increasing. The last two statements follow from similar statements proved for \(\phi \) in (a). \(\square \)
Proof of Lemma 15
We have
where k denotes the number of \(\hat{Y}_i\)’s whose value was determined to be equal to \(Y_i\).
If \(p> 1/2\) and \(0 < a < 1 - \frac{1}{2p}\), then letting \(k = \lceil m/2 \rceil \) gives
by Stirling’s approximation, and
by Lemma 13(c). This gives
as required.
Otherwise, let \(s=\phi (a)\). Then letting \(k = \lceil am/s \rceil \) gives
by Stirling’s approximation, and
by Lemma 13(b) and since f is continuous. This implies, in particular, that \(f(s)\in [0,1]\). Thus we have
where we have used Lemma 14(b) for the last equality. Using (39) we get
by the definition of \(g_L\) in (19). \(\square \)
Proof of Lemma 16
(a) The function \(r(p)=\log \left( \frac{1-p}{p} \right) - \frac{1-p}{1-2p}\) approaches \(+\infty \) when \(p\rightarrow 0^+\) and approaches \(-\infty \) when \(p\rightarrow { \frac{1}{2} }^{-}\). Moreover,
for \(p\in (0,1/2)\). Hence r(p) has a unique root \(p_0\), and \(r(p) \ge 0\) if and only if \(p\le p_0\).
(b) The function
approaches \(-\infty \) as \(s \rightarrow 0^+\), and approaches \(+\infty \) as \(s \rightarrow 1^{-}\), and its derivative is positive in (0, 1), hence it has a unique root \(s_0\) in (0, 1). Also we have \(\mu (2-p^{-1}) = p/(1-2p)\), which means that if \(p>1/2\) then \(s_0>2-p^{-1}\). Moreover, if \(p_0 < p \le 1/2\), then by part (a),
which means \(s_0 > \frac{1-2p}{1-p}\). \(\square \)
Proof of Lemma 19
The conclusion is obvious if \(p\ge \frac{1}{2}\) and \(a \le 2 - \frac{1}{p}\), or if \(a=0\), since in these cases \(h(a)=1\). Also, \(\mathbb {P}\left[ {X_1+\dots +X_m > m}\right] =0\) so the conclusion is true if \(a=1\), so we may assume that \(\max \{ 0, 2 - \frac{1}{p} \} < a < 1\).
Observe that if \(X_1+\dots +X_m > a m\), there is a subsequence of the form \(Y_{m-k+1},\dots ,Y_m\) whose sum is at least am, and this subsequence contains at least am elements since \(Y_i \le 1\) for all i. Hence we have
as the \(Y_i\)’s are i.i.d.
For any integer \(k \in [am, m]\), by Lemma 13(a) we have
for an absolute constant C, since \(am/k \ge a > 2 - \frac{1}{p}\). Let \(r = k/m \in [a,1]\). So we find that
Let us define
So to complete the proof we just need to show that
The function \(\xi (r)\) is positive and differentiable for each \(a\in (0,1)\), hence the supremum here occurs either at a boundary point or at a point with zero derivative. The derivative of \(\log (\xi (r))\) equals
Thus \(\xi '(r)\) has the same sign as \(\overline{\xi }(r) = p(1-p)(2r-a)^2 - r(r-a)\) in \(r\in [a,1]\). Notice that \(\overline{\xi }(r)\) has two roots
We may consider several cases.
Case 0: \(p=1/2\). The function \(\overline{\xi }\) is positive, so \(\xi \) is increasing in [a, 1], hence the supremum in (40) happens at \(r=1\) and its value is f(a).
Case 1: \(p>1/2\). Since \(a > 2 - \frac{1}{p}\), we find that \(r_1>1\) and \(r_2<0\). Moreover, \(\overline{\xi }(a)\ge 0\). Thus \(\overline{\xi }\) is non-negative in [a, 1], which implies \(\xi \) is increasing in [a, 1]. Thus the supremum in (40) happens at \(r=1\) and its value is f(a).
Case 2: \(p<1/2\) and \(a \le \frac{1-2p}{1-p}\). In this case \(r_1<0\) and \(a \le r_2 \le 1\). Since \(\overline{\xi }(a) \ge 0\) and \(\overline{\xi }(r_1)=\overline{\xi }(r_2)=0\) and \(\overline{\xi }\) is quadratic, the function \(\overline{\xi }\) goes from positive to negative at \(r_2\). Therefore, the function \(\xi \) attains its supremum at \(r_2\) and the supremum value in (40) equals
Case 3: \(p<1/2\) and \(a > \frac{1-2p}{1-p}\). We find that \(r_1<0\) and \(r_2>1\), and \(\overline{\xi }(a) \ge 0\), so \(\overline{\xi }\) is non-negative in [a, 1], hence \(\xi \) is increasing in [a, 1]. Thus the supremum in (40) happens at \(r=1\) and its value is f(a). This completes the proof of (40) and the lemma. \(\square \)
Proof of Lemma 20
(a) First, the case \(a=0\) is obvious since \(g_U(0) = 1/2\), and the case \(a=1\) is easy since \(\mathbb {P}\left[ {\hat{X}_1+\dots +\hat{X}_m > m}\right] = 0\). So we may assume that \(a\in (0,1)\).
Letting k of the \(\hat{X}_i\)’s being equal to \(X_i\) and the rest equal to zero, we get
For a given \(r \in [a,1]\), Lemma 19 gives
Moreover, by Stirling’s approximation
So, we find that
Thus to complete the proof of part (a) we just need to show
where we have used the change of variable \(\zeta = a/r\). For analysing this infimum we define the two variable function
with domain \(\{(a,\zeta ) : 0 < a < 1, a \le \zeta \le 1 \}\), and consider two cases depending on the value of p.
Case 1: \(p \ge 1/2\). By the definition of h in (22) we have
where f is defined in (4). Since \(f(2-p^{-1}) = 1\), \(\psi \) is continuous here. Let us define \(\psi _1(\zeta ) = \frac{\zeta -a}{\zeta } \left( \frac{a}{\zeta -a}\right) ^{a/\zeta }\) and \(\psi _2(\zeta )= \frac{\zeta -a}{\zeta } \left( \frac{a}{(\zeta -a)f(\zeta )}\right) ^{a/\zeta } \).
The derivative of \(\log \psi _1(\zeta )\) is
which is negative for \(\zeta <2a\) and positive for \(\zeta >2a\). This implies \(\psi _1(\zeta )\) is decreasing when \(\zeta \le 2a\) and increasing when \(\zeta \ge 2a\). So \(\psi _1\) achieves its minimum at \(\zeta = 2a\), and its minimum value is 1 / 2.
The derivative of \(\log \psi _2(\zeta )\) is
Comparing with (17) we find that this derivative has the same sign as \(\Phi (a,\zeta )\). So by Lemma 14(a) it vanishes at a unique point \(\zeta = \phi (a)\). Also at \(\zeta =\phi (a)\) we have \(\partial \Phi / \partial \zeta > 0\) (see (38)), which implies \(\Phi (a,\zeta )\) is non-positive when \(\zeta \le \phi (a)\) and non-negative when \(\zeta \ge \phi (a)\). Thus \(\psi _2\) achieves its minimum at \(\phi (a)\), and its minimum value is
by Lemma 14(b).
We conclude that:
(i) If \(2a \le 2 - 1/p\), then the infimum of \(\psi \) occurs at \(\zeta =2a\) and its value is \(\psi (a,2a) = \psi _1(2a)=1/2\). The reason is that on \([a,2-1/p]\), \(\psi = \psi _1\) achieves its minimum at 2a, and on \([2-1/p,1]\), \(\psi = \psi _2\) is increasing since \(\Phi (a,2-1/p) \ge 0\).
(ii) If \(a \le 2 - 1/p\) and \(2a > 2 - 1/p\), then the infimum occurs at \(\zeta =\phi (a)\) and its value is \(\frac{\phi (a)-a}{\phi (a)} \left( (1-p) (2-\phi (a)) / (1-\phi (a)) \right) ^a\). The reason is that on \([a,2-1/p]\), \(\psi = \psi _1\) is decreasing, and on \([2-1/p,1]\), \(\psi = \psi _2\) achieves its minimum at \(\phi (a)\) since \(\Phi (a,2-1/p)\le 0\) and \(\Phi (a,1)\ge 0\).
(iii) If \(a > 2 - 1/p\), then the infimum occurs at \(\zeta =\phi (a)\) and its value is equal to \(\frac{\phi (a)-a}{\phi (a)} \left( (1-p) (2-\phi (a)) / (1-\phi (a)) \right) ^a\). The reason is that on [a, 1], \(\psi =\psi _2\) achieves its minimum at \(\phi (a)\) since \(\Phi (a,a)\le 0\) and \(\Phi (a,1)\ge 0\).
Case 2: \(p < 1/2\). By the definition of h in (22) we have
The function \(\psi \) is continuous here since
Let us define \(\psi _3(\zeta ) = \left( \frac{1-p}{p}\right) ^{a} \frac{\zeta -a}{\zeta } \left( \frac{a}{\zeta -a}\right) ^{a/\zeta } \). Since \(\psi _3(\zeta ) = \left( \frac{1-p}{p}\right) ^{a} \psi _1(\zeta )\), the function \(\psi _3(\zeta )\) is decreasing when \(\zeta \le 2a\) and increasing when \(\zeta \ge 2a\). So \(\psi _3\) achieves its minimum at \(\zeta = 2a\) and its minimum value is \(\left( \frac{1-p}{p}\right) ^{a}/2\). We conclude that
(iv) If \(a \le 1 - p/(1-p)\) and \(2a \le 1 - p/(1-p)\), then the infimum in (41) occurs at \(\zeta =2a\) and at this point we have \(\psi (a,\zeta ) = \left( \frac{1-p}{p}\right) ^a/2\). The reason is that on \([a, 1 - p/(1-p)]\), \(\psi = \psi _3\) achieves its minimum at 2a, and on \([1-p/(1-p),1]\), \(\psi = \psi _2\) is increasing since \(\Phi (a,1 - p/(1-p))\ge 0 \).
(v) If \(a \le 1 - p/(1-p)\) and \(2a > 1 - p/(1-p)\), then the infimum in (41) occurs at \(\zeta =\phi (a)\) and its value is equal to \(\frac{\phi (a)-a}{\phi (a)} \left( (1-p) (2-\phi (a)) / (1-\phi (a)) \right) ^a\). The reason is that on \([a, 1 - p/(1-p)]\), \(\psi = \psi _3\) is decreasing, and on \([1-p/(1-p),1]\), \(\psi = \psi _2\) achieves its minimum at \(\phi (a)\) since \(\Phi (a,1 - p/(1-p))\le 0\) and \(\Phi (a,1)\ge 0\).
(vi) If \(a > 1 - p/(1-p)\), then the infimum in (41) occurs at \(\zeta =\phi (a)\) and its value is equal to \(\frac{\phi (a)-a}{\phi (a)} \left( (1-p) (2-\phi (a)) / (1-\phi (a)) \right) ^a\). The reason is that on [a, 1], \(\psi = \psi _2\) achieves its minimum at \(\phi (a)\) since \(\Phi (a,a)\le 0 \) and \(\Phi (a,1) \ge 0\).
In all cases we proved that \(g_U(a)\) actually gives the value of the infimum in (41), and this concludes the proof of (41) and of part (a).
(b) Consider the definition of \(g_U\) in (23). The formulae in (25) for the cases ‘\(p\ge 1/2\) and \(0 < a \le 1 - 1/2p\)’ and ‘\(p < 1/2\) and \(0 < a \le \frac{1-2p}{2-2p}\,\)’ are clearly true, so we assume that a is in the ‘otherwise’ case. We use the equality (41). Note that as proved in part (a), the infimum in (41) occurs at the point \(\zeta =\phi (a)\) that has \(\left. \frac{\partial \psi }{\partial \zeta }\right| _{(a,\phi (a))} = 0\). This implies for every \(a_0\),
and (25) follows from computing this partial derivative and putting \(\zeta = \phi (a_0)\).
We next prove the continuity of \(g_U\) and its derivative. Note that by Lemma 14(a), if \(a\in (0,1)\) then \(\phi (a)\in (0,1)\). First, \(g_U\) is continuous at \(a=1\) since
For \(p\ge 1/2\), the only discontinuity for \(g_U\) can possibly occur at \(b = 1 - 1/2p\). However at this point we have \(\phi (b) = 2b = 2 - p^{-1}\) so that \((1-p)(2-\phi (b)) = 1-\phi (b)\). Hence the left and right limits of \(g_U\) equal 1 / 2, and the left and right limits of \(g'_U\) equal 0. Therefore, both \(g_U\) and \(g'_U\) are continuous at b.
For \(p < 1/2\), the only discontinuity for \(g_U\) can possibly occur at \(c = (1-2p)/(2-2p)\). However at this point \(\phi (c) = 2c = (1-2p)/(1-p)\) so that \(\frac{(1-p)(2-\phi (c))}{1-\phi (c)} = \frac{1-p}{p}\). Hence the left and right limits of \(g_U\) equal \(\left( p^{-1}-1\right) ^{c}/2\), and the left and right limits of \(g'_U\) equal \(\displaystyle \left( \log \left( p^{-1}-1\right) \right) \left( p^{-1}-1\right) ^{c}/2\). Therefore, both \(g_U\) and \(g'_U\) are continuous at c.
(c) Note that \(g_U\) is positive everywhere, so \(\log (g_U)\) is (strictly) increasing if and only if \(g_U\) is (strictly) increasing. By the formulae for \(g'_U\) in part (b), it is easy to see that \(g'_U\) is always non-negative, and is positive when \(g_U(a) > 1/2\). To show \(\log (g_U)\) is convex, we need to show its derivative, i.e., \(g'_U/g_U\) is increasing. This also follows from part (b), noting that \(\phi \) is increasing by Lemma 14(c). \(\square \)
Rights and permissions
About this article
Cite this article
Mehrabian, A., Wormald, N. It’s a Small World for Random Surfers. Algorithmica 76, 344–380 (2016). https://doi.org/10.1007/s00453-015-0034-6
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00453-015-0034-6