1 / 37

An Introduction to Point Processes

Definitions & examples Conditional intensity & Papangelou intensity Models a) Renewal processes b) Poisson processes c) Cluster models d) Inhibition models. An Introduction to Point Processes. Point pattern : a collection of points in some space.

moses-nixon
Download Presentation

An Introduction to Point Processes

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Definitions & examples • Conditional intensity & Papangelou intensity • Models • a) Renewal processes • b) Poisson processes • c) Cluster models • d) Inhibition models An Introduction to Point Processes

  2. Point pattern: a collection of points in some space. Point process: a random point pattern. Centroids of Los Angeles County wildfires, 1960-2000

  3. Aftershocks from global large earthquakes

  4. Epicenters & times of microearthquakes in Parkfield, CA

  5. Marked point process: a random variable (mark) with each point. Hollister, CA earthquakes: locations, times, & magnitudes

  6. Los Angeles Wildfires: dates and sizes

  7. Time series:

  8. Time series: Palermo football rank vs. time Marked point process: Hollister earthquake times & magnitudes

  9. Antiquated definition: a point process N(t) is a right-continuous, Z+-valued stochastic process: --x-------x--------------x-----------------------x---x-x--------------- 0 t T N(t) = Number of points with times < t. Problem: does not extend readily to higher dimensions. Modern definition: A point process N is a Z+-valued random measure N(a,b) = Number of points with times between a & b. N(A) = Number of points in the set A.

  10. More Definitions: • s-finite: finite number of pts in any bounded set. • Simple: N({x}) = 0 or 1 for all x, almost surely. (No overlapping pts.) • Orderly: N(t, t+ D)/D ---->p 0, for each t. • Stationary: The joint distribution of {N(A1+u), …, N(Ak+u)} does not • depend on u. • Notation & Calculus: • ∫A f(x) dN = ∑f(xi ), for xi in A. • ∫A dN = N(A) = # of points in A.

  11. Intensities (rates) and Compensators • -------------x-x-----------x----------- ----------x---x--------------x------ • 0 t- t t+ T • Consider the case where the points are observed in time only. • N[t,u] = # of pts between times t and u. • Overall rate: m(t) = limDt -> 0 E{N[t, t+Dt)} / Dt. • Conditional intensity: l(t) = limDt -> 0 E{N[t, t+Dt) | Ht} / Dt, where Ht = history of N for all times before t. • If N is orderly, then l(t) = limDt -> 0 P{N[t, t+Dt) > 0 | Ht} / Dt. • Compensator: predictable process C(t) such that N-C is a martingale. If l(x) exists, then ∫otl(u) du = C(t). • Papangelou intensity: lp(t) = limDt -> 0 E{N[t, t+Dt) | Pt} / Dt, • where Pt = information on N for all times before and after t.

  12. Intensities (rates) and Compensators -------------x-x-----------x----------- ----------x---x--------------x------ 0 t- t t+ T These definitions extend to space and space-time: Conditional intensity: l(t,x) = limDt,Dx -> 0 E{N[t, t+Dt) x Bx,Dx | Ht} / DtDx, where Ht = history of N for all times before t, and Bx,Dx is a ball around x of size Dx. Compensator: ∫Al(t,x) dt dx = C(A). Papangelou intensity: lp(t,x) = limDt,Dx -> 0 E{N[t, t+Dt) x Bx,Dx | Pt,x} / DtDx, where Pt,x = information on N for all times and locations except (t,x).

  13. Some Basic Properties of Intensities: • Fact 1 (Uniqueness). If l exists, then it determines the distribution of N. (Daley and Vere-Jones, 1988). • Fact 2 (Existence). For any simple point process N, the compensator C exists and is unique. (Jacod, 1975) Typically we assume that l exists, and use it to model N. • Fact 3 (Kurtz Theorem). The avoidance probabilities, P{N(A)=0} for all measurable sets A, also uniquely determine the distribution of N. • Fact 4 (Martingale Theorem). For any predictable process f(t), E ∫ f(t) dN = E ∫ f(t) l(t) dt. • Fact 5 (Georgii-Zessin-Nguyen Theorem). For any ex-visible process f(x), • E ∫ f(x) dN = E ∫ f(x) lp(x) dx.

  14. Some Important Point Process Models: Renewal process. The inter-event times: t2 - t1, t3 - t2, t4 - t3, etc. are independent and identically distributed random variables. (Classical density estimation.) Ex.: Normal, exponential, power-law, Weibull, gamma, log-normal.

  15. 2) Poisson process. Fact 6: If N is orderly and l does not depend on the history of the process, then N is a Poisson process: N(A1), N(A2), … , N(Ak) are independent, and each has the Poisson dist.: P{N(A) = j} = [C(A)]j exp{-C(A)} / j!. Recall: C(A) = ∫Al(x) dx. Stationary (homogeneous) Poisson process: l(x) = m. Fact 7: Equivalent to a renewal process with exponential inter-event times. Inhomogeneous Poisson process: l(x) = f(x), where f(x) is some fixed, deterministic function.

  16. The Poisson process is the limiting distribution in many important results: Fact 8 (thinning; Westcott 1976): Suppose N is simple, stationary, & ergodic.

  17. Fact 9 (superposition; Palm 1943): Suppose N is simple & stationary. Then Mk --> stationary Poisson.

  18. Fact 10 (translation; Vere-Jones 1968; Stone 1968): Suppose N is stationary. For each point xi in N, move it to xi + yi, where {yi} are iid. Let Mk be the result of k such translations. Then Mk --> stationary Poisson.

  19. Fact 11 (rescaling; Meyer 1971): Suppose N is simple and has at most one point on any vertical line. Rescale the y-coordinates: move each point (xi, yi) to (xi , ∫oyil(xi,y) dy). Then the resulting process is stationary Poisson.

  20. 3) Some cluster models. • Neyman-Scott process: clusters of points whose centers are formed from a stationary Poisson process. Typically each cluster consists of a fixed integer k of points which are placed uniformly and independently within a ball of radius r around each cluster’s center. • Cox-Matern process: cluster sizes are random: independent and identically distributed Poisson random variables. • Thomas process: cluster sizes are Poisson, and the points in each cluster are distributed independently and isotropically according to a Gaussian distribution. • Hawkes (self-exciting) process: “mothers” are formed from a stationary Poisson process, and each produces a cluster of “daughter” points, and each of them produces a cluster of further “daughter” points, etc. l(t, x) = m + ∑ g(t-ti, ||x-xi||). ti < t

  21. 4) Some inhibition models. • Matern (I) process: first generate points from a stationary Poisson process, and then if there are any pairs of points within distance d of each other, delete both of them. • Matern (II) process: generate a stationary Poisson process, then index the points j = 1,2,…,n at random, and then successively delete any point j if it is within distance d from any retained point with smaller index. c) Simple Sequential Inhibition (SSI): Keep simulating points from a stationary Poisson process, deleting any if it is within distance d from any retained point, until exactly k points are kept. • Self-correcting process: Hawkes process where g can be negative: l(t, x) = m + ∑ g(t-ti, ||x-xi||). ti < t

  22. Poisson (100) Poisson (50+50x+50y) Neyman-Scott(10,5,0.05) Cox-Matern(10,5,0.05) Thomas (10,5,0.05) Matern I (200, 0.05) Matern II (200, 0.05) SSI (200, 0.05)

  23. In modeling a space-time marked point process, usually directly model l(t,x,a). • For example, for Los Angeles County wildfires: • Windspeed. Relative Humidity, Temperature, Precipitation, • Tapered Pareto size distribution f, smooth spatial background m. l(t,x,a) = b1exp{b2R(t) + b3W(t) + b4P(t)+ b5A(t;60) + b6T(t) + b7[b8 - D(t)]2} m(x) g(a). Could also include fuel age, wind direction, interactions…

  24. r = 0.16 (sq m)

  25. (sq m) (F)

  26. In modeling a space-time marked point process, usually directly model l(t,x,a). • For example, for Los Angeles County wildfires: • Windspeed. Relative Humidity, Temperature, Precipitation, • Tapered Pareto size distribution f, smooth spatial background m. l(t,x,a) = b1exp{b2R(t) + b3W(t) + b4P(t)+ b5A(t;60) + b6T(t) + b7[b8 - D(t)]2} m(x) g(a). Could also include fuel age, wind direction, interactions…

  27. In modeling a space-time marked point process, usually directly model l(t,x,a). • For example, for Los Angeles County wildfires: • Relative Humidity, Windspeed, Precipitation, Aggregated rainfall over previous 60 days, Temperature, Date • Tapered Pareto size distribution f, smooth spatial background m. l(t,x,a) = b1exp{b2R(t) + b3W(t) + b4P(t)+ b5A(t;60) + b6T(t) + b7[b8 - D(t)]2} m(x) g(a). Could also include fuel age, wind direction, interactions…

  28. (Ogata 1998)

  29. Simulation. • Sequential. a) Renewal processes are easy to simulate: generate iid random variables z1, z2, … from the renewal distribution, and let t1=z1, t2= z1+ z2, t3= z1+z2+z3, etc. b) Reverse Rescaling. In general, can simulate a Poisson process with rate 1, and move each point (ti, xi) to (ti , yi), where xi = ∫oyil(ti,x) dx. • Thinning. If m = sup l(t, x), first generate a Poisson process with rate m, and then keep each point (ti, xi) with probability l(ti, xi)/m.

  30. Summary: • Point processes are random measures: N(A) = # of points in A. • l(t,x) = Expected rate around x, given history < time t. • Classical models are renewal & Poisson processes. • For Poisson processes, l(t,x) is deterministic. • Poisson processes are limits in thinning, superposition, translation, and rescaling theorems. • Non-Poisson processes may have clustering (Neyman-Scott, Cox-Matern, Thomas, Hawkes) or inhibition (MaternI, MaternII, SSI, self-correcting). • Next time: How to estimate the parameters in these models, and how to tell how well a model fits….

More Related