Length versus cost The Poissonian city Scale-invariant random spatial networks Length versus cost The Poissonian city Scale-invariant random spatial networks 1: Length versus cost Stochastic geometry and the art of getting from A to B Wilfrid Kendall w.s.kendall@warwick.ac.uk Probability Approximations: A conference in Honour of Louis Chen on his 70th birthday Short-length routes versus low-cost networks Examples of frustrated optimization in networks Finessing trade-off using Poisson line process cell Mean perimeter length as double integral Asymptotically efficient networks A theoretical lower bound Simulations 25 June 2010 See Aldous and WSK (2008). Length versus cost The Poissonian city Scale-invariant random spatial networks Examples of frustrated optimization in networks Length versus cost The Poissonian city Scale-invariant random spatial networks Finessing trade-off using Poisson line process cell Consider a network formed from a Poisson line process; Connect x to y using enclosing cell Cx,y ; Compute mean length of ∂Cx,y . . . . . . using independent unit-intensity invariant Poisson line process Π2 ; and determine mean number of hits. Length versus cost The Poissonian city Scale-invariant random spatial networks Length versus cost Mean perimeter length as double integral Theorem h i E length ∂Cx,y − 2|x − y| = ZZ 1 (α − sin α) exp − 12 (η − n) d z 2 R2 Note that α = α(z) and η = η(z) both depend on z. Fixed α: locus of z is circle. Fixed η: locus of z is ellipse. Length versus cost The Poissonian city Scale-invariant random spatial networks A theoretical lower bound The Poissonian city Scale-invariant random spatial networks Asymptotically efficient networks h i Careful work shows E length ∂Cx,y − 2|x − y| is asymptotically logarithmic in |x − y|. Hence sparse line processes deliver efficient modifications of networks. Theorem √ For any N-point configuration x (N) in square side N and any sequence wN → ∞ there are connecting networks GN with: length(GN ) = average(GN ) = length(ST(x (N) )) + o(N) XX 1 kxi − xj k + o(wN log N) N(N − 1) i≠j The sequence {wN } can tend to infinity arbitrarily slowly. Length versus cost The Poissonian city Scale-invariant random spatial networks Simulations The above is provably nearly as good as we can hope: Theorem √ Given apconfiguration of N cities in [0, N]2 which is LN = o( log N)-equidistributed: random choice XN of city can be coupled to uniformly random point YN so that |XN − YN | E min 1, -→ 0 ; LN consider any connecting network GN with length bounded above by a multiple of N. This connects the cities with average connection length exceeding p average Euclidean connection length by at least Ω( log N) . One thousand semi-perimeters, with vertical exaggeration and statistical analysis. Length versus cost The Poissonian city Scale-invariant random spatial networks Length versus cost 2: The Poissonian city The Poissonian city Scale-invariant random spatial networks Describing near-geodesics The Poissonian city Describing near-geodesics Comparison with true geodesics Random variation Flow in the network Asymptotic distribution of maximum lateral deviation: uniformly distributed along the line segment, See WSK (2009). Length versus cost The Poissonian city Scale-invariant random spatial networks with conditional distribution of deviation given by length of Gaussian 4-vector, quadratic variance vanishing at the endpoints. Length versus cost Comparison with true geodesics The Poissonian city Scale-invariant random spatial networks Random variation Asymptotic variance of excess length: (one end only). 20 27 log dist(x, y) Ingredients: Paths formed by semi-cells have asymptotic excess 4 length (over Euclidean length) of 3 log dist(x, y); True geodesicshave asymptotic length excess no less than log 4 − 54 log dist(x, y) at each end. representation by random growth process; Lévy processes; Lamperti transform and self-similar processes; thin-tailed perpetuities. Length versus cost The Poissonian city Scale-invariant random spatial networks Flow in the network Suppose traffic is generated uniformly between all point-pairs in disk radius n. Condition on line through centre of disk. Length versus cost The Poissonian city Scale-invariant random spatial networks 3: Scale-invariant random spatial networks Scale-invariant random spatial networks Scale-invariance Model using marked Poisson line process Foundational results Let Tn be flow through centre of disk; √ E [Tn ] = 2n3 + O n2 n ; (Aldous and WSK; work in progress) Tn /n3 has a non-degenerate limiting distribution . . . . . . derived from flow for improper Poisson line process; convergence of moments of order up to 2 − . Length versus cost The Poissonian city Scale-invariant random spatial networks Scale-invariance Input: set of nodes x1 , . . . , xn ; Output: random network N(x1 , . . . , xn ) connecting nodes. Scale-invariance: L (N(λx1 , . . . , λxn )) = L (λN(x1 , . . . , xn )) for each Euclidean similarity λ. Consider example based on Poisson line process marked with speeds; build N(x1 , . . . , xn ) from fastest paths connecting the nodes x1 , . . . , xn . Length versus cost The Poissonian city Scale-invariant random spatial networks Model using marked Poisson line process Poisson line process: coordinatize lines by signed distance from origin o and angle in range [0, π ). Each line marked by a speed v > 0. Intensity measure 1 −γ 2v d v d r d θ. Lines at speed v0 or higher form Poisson line process of −(γ−1) v0 intensity γ−1 ; the full set of lines is everywhere dense. Regularity: γ > 1 or can travel anywhere in zero time; γ ≥ 2 or can reach infinity in finite time; γ > 2 or cannot leave o in finite time. Length versus cost The Poissonian city Scale-invariant random spatial networks Length versus cost Foundational results (I) The Poissonian city Scale-invariant random spatial networks Foundational results (II) Build path leaving o in finite time by working in reverse: choose fastest line from circle radius n to circle radius 2n , then fastest line from circle radius 2n to circle radius 4n also hitting previous line, repeat . . . . Mean time from o to circle of radius n is finite, and indeed (inevitably from scaling!) is proportional to Consider intensity for s = |v|, u = 1/v (“slowness”): we obtain π uγ−2 d u d s. Introduce notion of extended path: moves infinitely fast on circles centered at o; speed boosted so that radial speed always equals v. Study the fastest extended path from o: this is the lexicographic “frontier” of the Poisson point pattern. Hence immediately, if γ > 2, then paths cannot go to infinity in finite time, and extended paths can leave o in finite time. Length versus cost The Poissonian city Scale-invariant random spatial networks Bibliography Aldous, D. J. and J. Shun (2010). Connected Spatial Networks over Random Points and a Route-Length Statistic. Aldous, D. J. and WSK (2008, March). Short-length routes in low-cost networks via Poisson line patterns. Advances in Applied Probability 40(1), 1–21. Böröczky, K. J. and R. Schneider (2008). The mean width of circumscribed random polytopes. Canadian Mathematical Bulletin accepted. Rényi, A. and R. Sulanke (1968). Zufällige konvexe Polygone in einem Ringgebiet. Zeitschrift {f{ü}r} Wahrscheinlichkeitstheorie und Verwe Gebiete 9, 146–157. 1 1− γ−1 n . Path length is then almost-surely finite (soft argument) and actually of finite mean, hence proportional to n. Finally, we can show that almost surely there are fastest+finite-time / finite-distance paths from all points to all other points. Length versus cost The Poissonian city Scale-invariant random spatial networks WSK (2009, October). Geodesics and flows in a Poissonian city. pp. 36.