Stable Behaviour of Infinitely Wide Deep Neural Networks

Total Page:16

File Type:pdf, Size:1020Kb

Stable Behaviour of Infinitely Wide Deep Neural Networks Stable behaviour of infinitely wide deep neural networks Stefano Favaro Sandra Fortini Stefano Peluchetti [email protected] [email protected] [email protected] Department ESOMAS Department of Decision Sciences Cogent Labs University of Torino Bocconi University and Collegio Carlo Alberto Abstract (Lee et al., 2018; Garriga-Alonso et al., 2019); ii) ker- nel regression for infinitely wide networks which are trained with continuous-time gradient descent via the We consider fully connected feed-forward neural tangent kernel (Jacot et al., 2018; Lee et al., deep neural networks (NNs) where weights 2019; Arora et al., 2019); iii) analysis of the properties and biases are independent and identically of infinitely wide networks as function of depth via distributed as symmetric centered stable dis- the information propagation framework (Poole et al., tributions. Then, we show that the infi- 2016; Schoenholz et al., 2017; Hayou et al., 2019). It nite wide limit of the NN, under suitable has been shown a substantial gap between finite NNs scaling on the weights, is a stochastic pro- and their corresponding infinite (wide) GPs counter- cess whose finite-dimensional distributions parts in terms of empirical performance, at least on are multivariate stable distributions. The some of the standard benchmarks applications. More- limiting process is referred to as the stable over, it has been shown to be a difficult task to avoid process, and it generalizes the class of Gaus- undesirable empirical properties arising in the context sian processes recently obtained as infinite of very deep networks. Given that, there exists an in- wide limits of NNs (Matthews et al., 2018b). creasing interest in expanding GPs arising in the limit Parameters of the stable process can be com- of infinitely wide NNs as a way forward to close, or puted via an explicit recursion over the layers even reverse, this empirical performance gap and to of the network. Our result contributes to the avoid, or slow down, pathological behaviors in very theory of fully connected feed-forward deep deep NN. NNs, and it paves the way to expand recent lines of research that rely on Gaussian infinite Let (µ, σ2) denote the Gaussian distribution with N wide limits. mean µ R and variance σ2 R+. Following the celebrated2 work of Neal (1995), we2 consider the shallow NN 1 Introduction I (1) (1) (1) fi (x)= wi,j xj + bi j=1 The connection between infinitely wide deep feed- X forward neural networks (NNs), whose parameters 1 n f (2)(x, n)= w(2)φ(f (1)(x)) + b(2), at initialization are independent and identically dis- i pn i,j j i j=1 tributed (iid) as scaled and centered Gaussian distri- X butions, and Gaussian processes (GPs) is well known where φ is a nonlinearity, i =1,...,n, w(1),w(2) iid (Neal, 1995; Der and Lee, 2006; Lee et al., 2018; i,j i,j ⇠ Matthews et al., 2018a,b; Yang, 2019). Recently, this 2 (1) (2) iid 2 I (0,σw), bi ,bi (0,σb ) and x R is the input. intriguing connection has been exploited in many ex- ItN follows that ⇠N 2 citing research directions, including: i) Bayesian in- (1) iid 2 ference for GPs arising from infinitely wide networks f (x) 0,σ (1) (x) i ⇠N f (2) (1) iid ⇣ 2 ⌘ rd f (x, n) f 0,σ (2) (x, n) Proceedings of the 23 International Conference on Artifi- i | ⇠N f cial Intelligence and Statistics (AISTATS) 2020, Palermo, ⇣ I ⌘ Italy. PMLR: Volume 108. Copyright 2020 by the au- 2 2 2 1 2 σ (1) (x)=σ + σ x thor(s). f b w I j j=1 X Stable behaviour of infinitely wide deep neural networks n 2 2 2 1 (1) 2 each layer; iv) the convergence in distribution is es- σ (2) (x, n)=σ + σ φ(f (x)) . f b w n j tablished jointly for multiple inputs, namely the con- j=1 X vergence concerns the class of finite dimensional dis- tributions of the NN viewed as a stochastic process If x0 is another input we obtain bivariate Gaussian distributions in function space. See Neal (1995) and Der and Lee (2006) for early works on NNs under stable initializa- (1) (1) iid tion. (f (x),f (x0)) 0, ⌃ (1) (x, x0) i i ⇠N2 f (2) (2) (1) iid Within our setting, we show that the infinite wide limit (f (x, n),f (x0,n)) f 0, ⌃ (2) (x, x0,n ) , i i | ⇠N2 f of the NN, under suitable scaling on the weights, is a where stochastic process whose finite-dimensional distribu- tions are multivariate stable distributions (Samorad- 2 nitsky, 2017). This process is referred to as the stable σf (1) (x) cf (1) (x, x0) ⌃f (1) (x, x0)= 2 process. Our result may be viewed as a generalization cf (1) (x, x0) σ (1) (x0) " f # of the main result of Matthews et al. (2018b)tothe 2 σf (2) (x, n) cf (2) (x, x0,n) context of stable distributions, as well as a generaliza- ⌃f (2) (x, x0,n)= 2 c (2) (x, x ,n) σ (x ,n) tion of results of Neal (1995) and Der and Lee (2006) " f 0 f (2) 0 # to the context of deep NN. Our result contributes to I 2 2 1 the theory of fully connected feed-forward deep NNs, c (1) (x, x0)=σ + σ x x0 f b w I j j and it paves the way to extend the research directions j=1 X i) ii) and iii) that rely on Gaussian infinite wide lim- n 2 2 1 (1) (1) its. The class of stable distributions is known to be c (2) (x, x0,n)=σ + σ φ(f (x))φ(f (x0)). f b w n j j especially relevant. Indeed while the contribution of j=1 X each Gaussian weight vanishes as the width grows un- a.s. bounded, some of the stable weights retains significant Let denote the almost sure convergence. By the size, thus allowing them to represent ”hidden features” strong! law of large numbers we know that, as n + , (Neal, 1995). one has ! 1 n The paper is structured as follows. Section 2 contains 1 (1) 2 a.s. (1) 2 some preliminaries on stable distributions, whereas in φ(f (x)) E[φ(f (x)) ] n j ! 1 j=1 Section 3 we define the class of feedforward NNs con- X n sidered in this work. Section 4 contains our main 1 (1) (1) a.s. (1) (1) φ(f (x))φ(f (x0)) E[φ(f (x))φ(f (x0))], result: as the width tends to infinity jointly on net- n j j ! 1 1 j=1 work’s layers, the finite dimensional distributions of X the NN converges to a multivariate stable distribution from which one can conjecture that in the limit of whose parameters are compute via a recursion over (2) infinite width the stochastic processes fi (x) are the layers. The convergence of the NN to the sta- distributed as iid (over i)centeredGPwithkernel ble process then follows by finite-dimensional projec- 2 2 (1) (1) tions. In Section 5 we detail how our result extends K(x, x0)=σb + σwE[φ(f1 (x))φ(f1 (x0))]. Pro- vided that the nonlinear function φ is chosen so that previously established large width convergence results (1) and comment on related work, whereas in Section 6 φ(f1 (x)) has finite second moment, Matthews et al. (2018b) made rigorous this argument and extended it we discuss how our result applys to the research lines to deep NNs. highlighted in i) ii) and iii) which relies on GP lim- its. In Section 7 we comment on future research di- A key assumption underlying the interplay between rections. The Supplementary Material (SM) contains infinite wide NNs and GPs is the finiteness of the vari- all the proofs (SM A,B,C), a preliminary numerical ance of the parameters’ distribution at initialization. experiment on the recursion evaluation (SM D), an In this paper we remove the assumption of finite vari- empirical investigation of the distribution of trained ance by considering iid initializations based on stable NN models’ parameters (SM E). Code is available at distributions, which includes Gaussian initializations https://github.com/stepelu/deep-stable. as a special case. We study the infinite wide limit of fully connected feed-forward NN in the following general setting: i) the NN is deep, namely the NN 2 Stable distributions is composed of multiple layers; ii) biases and scaled weights are iid according to centered symmetric stable Let St(↵,σ) denote the symmetric centered stable dis- distributions; iii) the width of network’s layers goes to tribution with stability parameter ↵ (0, 2] and scale 2 infinity jointly on the layers, and not sequentially on parameter σ>0, and let S↵,σ be a random vari- Running heading author breaks the line (l) able distributed as St(↵,σ). That is, the characteristic That is, the characteristic function of wi,j St(↵,σw) itS↵,σ ⇠ function of S↵,σ St(↵,σ)is'S↵,σ (t)=E[e ]= is ↵ ↵ (l) ↵ ↵ σ t ⇠ itwi,j σw t e− | | . For any σ>0, a S↵,σ random variable with ' (l) (t)=E[e ]=e− | | , (4) wi,j 0 <↵<2 has finite absolute moments [ S ↵ "] E ↵,σ − (l) ↵ | | for any i 1, j 1 and l 1. Let b denote the bi- for any ">0, while E[ S↵,σ ]=+ . Note that ≥ ≥ ≥ i when ↵ = 2, we have that| S| 1(0, 2σ2). The ases of the l-th hidden layer, and assume that they are 2,σ ⇠N random variable S2,σ has finite absolute moments of independent and identically distributed as St(↵,σb).
Recommended publications
  • Introduction to Lévy Processes
    Introduction to Lévy Processes Huang Lorick [email protected] Document type These are lecture notes. Typos, errors, and imprecisions are expected. Comments are welcome! This version is available at http://perso.math.univ-toulouse.fr/lhuang/enseignements/ Year of publication 2021 Terms of use This work is licensed under a Creative Commons Attribution 4.0 International license: https://creativecommons.org/licenses/by/4.0/ Contents Contents 1 1 Introduction and Examples 2 1.1 Infinitely divisible distributions . 2 1.2 Examples of infinitely divisible distributions . 2 1.3 The Lévy Khintchine formula . 4 1.4 Digression on Relativity . 6 2 Lévy processes 8 2.1 Definition of a Lévy process . 8 2.2 Examples of Lévy processes . 9 2.3 Exploring the Jumps of a Lévy Process . 11 3 Proof of the Levy Khintchine formula 19 3.1 The Lévy-Itô Decomposition . 19 3.2 Consequences of the Lévy-Itô Decomposition . 21 3.3 Exercises . 23 3.4 Discussion . 23 4 Lévy processes as Markov Processes 24 4.1 Properties of the Semi-group . 24 4.2 The Generator . 26 4.3 Recurrence and Transience . 28 4.4 Fractional Derivatives . 29 5 Elements of Stochastic Calculus with Jumps 31 5.1 Example of Use in Applications . 31 5.2 Stochastic Integration . 32 5.3 Construction of the Stochastic Integral . 33 5.4 Quadratic Variation and Itô Formula with jumps . 34 5.5 Stochastic Differential Equation . 35 Bibliography 38 1 Chapter 1 Introduction and Examples In this introductive chapter, we start by defining the notion of infinitely divisible distributions. We then give examples of such distributions and end this chapter by stating the celebrated Lévy-Khintchine formula.
    [Show full text]
  • A Mathematical Theory of Network Interference and Its Applications
    INVITED PAPER A Mathematical Theory of Network Interference and Its Applications A unifying framework is developed to characterize the aggregate interference in wireless networks, and several applications are presented. By Moe Z. Win, Fellow IEEE, Pedro C. Pinto, Student Member IEEE, and Lawrence A. Shepp ABSTRACT | In this paper, we introduce a mathematical I. INTRODUCTION framework for the characterization of network interference in In a wireless network composed of many spatially wireless systems. We consider a network in which the scattered nodes, communication is constrained by various interferers are scattered according to a spatial Poisson process impairments such as the wireless propagation effects, and are operating asynchronously in a wireless environment network interference, and thermal noise. The effects subject to path loss, shadowing, and multipath fading. We start introduced by propagation in the wireless channel include by determining the statistical distribution of the aggregate the attenuation of radiated signals with distance (path network interference. We then investigate four applications of loss), the blocking of signals caused by large obstacles the proposed model: 1) interference in cognitive radio net- (shadowing), and the reception of multiple copies of the works; 2) interference in wireless packet networks; 3) spectrum same transmitted signal (multipath fading). The network of the aggregate radio-frequency emission of wireless net- interference is due to accumulation of signals radiated by works; and 4) coexistence between ultrawideband and nar- other transmitters, which undesirably affect receiver nodes rowband systems. Our framework accounts for all the essential in the network. The thermal noise is introduced by the physical parameters that affect network interference, such as receiver electronics and is usually modeled as additive the wireless propagation effects, the transmission technology, white Gaussian noise (AWGN).
    [Show full text]
  • Introduction to Lévy Processes
    Introduction to L´evyprocesses Graduate lecture 22 January 2004 Matthias Winkel Departmental lecturer (Institute of Actuaries and Aon lecturer in Statistics) 1. Random walks and continuous-time limits 2. Examples 3. Classification and construction of L´evy processes 4. Examples 5. Poisson point processes and simulation 1 1. Random walks and continuous-time limits 4 Definition 1 Let Yk, k ≥ 1, be i.i.d. Then n X 0 Sn = Yk, n ∈ N, k=1 is called a random walk. -4 0 8 16 Random walks have stationary and independent increments Yk = Sk − Sk−1, k ≥ 1. Stationarity means the Yk have identical distribution. Definition 2 A right-continuous process Xt, t ∈ R+, with stationary independent increments is called L´evy process. 2 Page 1 What are Sn, n ≥ 0, and Xt, t ≥ 0? Stochastic processes; mathematical objects, well-defined, with many nice properties that can be studied. If you don’t like this, think of a model for a stock price evolving with time. There are also many other applications. If you worry about negative values, think of log’s of prices. What does Definition 2 mean? Increments , = 1 , are independent and Xtk − Xtk−1 k , . , n , = 1 for all 0 = . Xtk − Xtk−1 ∼ Xtk−tk−1 k , . , n t0 < . < tn Right-continuity refers to the sample paths (realisations). 3 Can we obtain L´evyprocesses from random walks? What happens e.g. if we let the time unit tend to zero, i.e. take a more and more remote look at our random walk? If we focus at a fixed time, 1 say, and speed up the process so as to make n steps per time unit, we know what happens, the answer is given by the Central Limit Theorem: 2 Theorem 1 (Lindeberg-L´evy) If σ = V ar(Y1) < ∞, then Sn − (Sn) √E → Z ∼ N(0, σ2) in distribution, as n → ∞.
    [Show full text]
  • Fclts for the Quadratic Variation of a CTRW and for Certain Stochastic Integrals
    FCLTs for the Quadratic Variation of a CTRW and for certain stochastic integrals No`eliaViles Cuadros (joint work with Enrico Scalas) Universitat de Barcelona Sevilla, 17 de Septiembre 2013 Damped harmonic oscillator subject to a random force The equation of motion is informally given by x¨(t) + γx_(t) + kx(t) = ξ(t); (1) where x(t) is the position of the oscillating particle with unit mass at time t, γ > 0 is the damping coefficient, k > 0 is the spring constant and ξ(t) represents white L´evynoise. I. M. Sokolov, Harmonic oscillator under L´evynoise: Unexpected properties in the phase space. Phys. Rev. E. Stat. Nonlin Soft Matter Phys 83, 041118 (2011). 2 of 27 The formal solution is Z t x(t) = F (t) + G(t − t0)ξ(t0)dt0; (2) −∞ where G(t) is the Green function for the homogeneous equation. The solution for the velocity component can be written as Z t 0 0 0 v(t) = Fv (t) + Gv (t − t )ξ(t )dt ; (3) −∞ d d where Fv (t) = dt F (t) and Gv (t) = dt G(t). 3 of 27 • Replace the white noise with a sequence of instantaneous shots of random amplitude at random times. • They can be expressed in terms of the formal derivative of compound renewal process, a random walk subordinated to a counting process called continuous-time random walk. A continuous time random walk (CTRW) is a pure jump process given by a sum of i.i.d. random jumps fYi gi2N separated by i.i.d. random waiting times (positive random variables) fJi gi2N.
    [Show full text]
  • Levy Processes
    LÉVY PROCESSES, STABLE PROCESSES, AND SUBORDINATORS STEVEN P.LALLEY 1. DEFINITIONSAND EXAMPLES d Definition 1.1. A continuous–time process Xt = X(t ) t 0 with values in R (or, more generally, in an abelian topological groupG ) isf called a Lévyg ≥ process if (1) its sample paths are right-continuous and have left limits at every time point t , and (2) it has stationary, independent increments, that is: (a) For all 0 = t0 < t1 < < tk , the increments X(ti ) X(ti 1) are independent. − (b) For all 0 s t the··· random variables X(t ) X−(s ) and X(t s ) X(0) have the same distribution.≤ ≤ − − − The default initial condition is X0 = 0. A subordinator is a real-valued Lévy process with nondecreasing sample paths. A stable process is a real-valued Lévy process Xt t 0 with ≥ initial value X0 = 0 that satisfies the self-similarity property f g 1/α (1.1) Xt =t =D X1 t > 0. 8 The parameter α is called the exponent of the process. Example 1.1. The most fundamental Lévy processes are the Wiener process and the Poisson process. The Poisson process is a subordinator, but is not stable; the Wiener process is stable, with exponent α = 2. Any linear combination of independent Lévy processes is again a Lévy process, so, for instance, if the Wiener process Wt and the Poisson process Nt are independent then Wt Nt is a Lévy process. More important, linear combinations of independent Poisson− processes are Lévy processes: these are special cases of what are called compound Poisson processes: see sec.
    [Show full text]
  • Financial Modeling with L´Evy Processes and Applying L
    FINANCIAL MODELING WITH LEVY´ PROCESSES AND APPLYING LEVY´ SUBORDINATOR TO CURRENT STOCK DATA by GONSALGE ALMEIDA Submitted in partial fullfillment of the requirements for the degree of Doctor of Philosophy Dissertation Advisor: Dr. Wojbor A. Woyczynski Department of Mathematics, Applied Mathematics and Statistics CASE WESTERN RESERVE UNIVERSITY January 2020 CASE WESTERN RESERVE UNIVERSITY SCHOOL OF GRADUATE STUDIES We hereby approve the dissertation of Gonsalge Almeida candidate for the Doctoral of Philosophy degree Committee Chair: Dr.Wojbor Woyczynski Professor, Department of the Mathematics, Applied Mathematics and Statis- tics Committee: Dr.Alethea Barbaro Associate Professor, Department of the Mathematics, Applied Mathematics and Statistics Committee: Dr.Jenny Brynjarsdottir Associate Professor, Department of the Mathematics, Applied Mathematics and Statistics Committee: Dr.Peter Ritchken Professor, Weatherhead School of Management Acceptance date: June 14, 2019 *We also certify that written approval has been obtained for any proprietary material contained therein. CONTENTS List of Figures iv List of Tables ix Introduction . .1 1 Financial Modeling with L´evyProcesses and Infinitely Divisible Distributions 5 1.1 Introduction . .5 1.2 Preliminaries on L´evyprocesses . .6 1.3 Characteristic Functions . .8 1.4 Cumulant Generating Function . .9 1.5 α−Stable Distributions . 10 1.6 Tempered Stable Distribution and Process . 19 1.6.1 Tempered Stable Diffusion and Super-Diffusion . 23 1.7 Numerical Approximation of Stable and Tempered Stable Sample Paths 28 1.8 Monte Carlo Simulation for Tempered α−Stable L´evyprocess . 34 2 Brownian Subordination (Tempered Stable Subordinator) 44 i 2.1 Introduction . 44 2.2 Tempered Anomalous Subdiffusion . 46 2.3 Subordinators . 49 2.4 Time-Changed Brownian Motion .
    [Show full text]
  • Models Beyond the Dirichlet Process
    Models beyond the Dirichlet process Antonio Lijoi Igor Prünster No. 129 December 2009 www.carloalberto.org/working_papers © 2009 by Antonio Lijoi and Igor Prünster. Any opinions expressed here are those of the authors and not those of the Collegio Carlo Alberto. Models beyond the Dirichlet process Antonio Lijoi1 and Igor Prunster¨ 2 1 Dipartimento Economia Politica e Metodi Quantitatavi, Universit`adegli Studi di Pavia, via San Felice 5, 27100 Pavia, Italy and CNR{IMATI, via Bassini 15, 20133 Milano. E-mail: [email protected] 2 Dipartimento di Statistica e Matematica Applicata, Collegio Carlo Alberto and ICER, Universit`a degli Studi di Torino, Corso Unione Sovietica 218/bis, 10134 Torino, Italy. E-mail: [email protected] September 2009 Abstract. Bayesian nonparametric inference is a relatively young area of research and it has recently undergone a strong development. Most of its success can be explained by the considerable degree of flexibility it ensures in statistical modelling, if compared to parametric alternatives, and by the emergence of new and efficient simulation techniques that make nonparametric models amenable to concrete use in a number of applied sta- tistical problems. Since its introduction in 1973 by T.S. Ferguson, the Dirichlet process has emerged as a cornerstone in Bayesian nonparametrics. Nonetheless, in some cases of interest for statistical applications the Dirichlet process is not an adequate prior choice and alternative nonparametric models need to be devised. In this paper we provide a review of Bayesian nonparametric models that go beyond the Dirichlet process. 1 Introduction Bayesian nonparametric inference is a relatively young area of research and it has recently under- gone a strong development.
    [Show full text]
  • Final Report (PDF)
    Foundation of Stochastic Analysis Krzysztof Burdzy (University of Washington) Zhen-Qing Chen (University of Washington) Takashi Kumagai (Kyoto University) September 18-23, 2011 1 Scientific agenda of the conference Over the years, the foundations of stochastic analysis included various specific topics, such as the general theory of Markov processes, the general theory of stochastic integration, the theory of martingales, Malli- avin calculus, the martingale-problem approach to Markov processes, and the Dirichlet form approach to Markov processes. To create some focus for the very broad topic of the conference, we chose a few areas of concentration, including • Dirichlet forms • Analysis on fractals and percolation clusters • Jump type processes • Stochastic partial differential equations and measure-valued processes Dirichlet form theory provides a powerful tool that connects the probabilistic potential theory and ana- lytic potential theory. Recently Dirichlet forms found its use in effective study of fine properties of Markov processes on spaces with minimal smoothness, such as reflecting Brownian motion on non-smooth domains, Brownian motion and jump type processes on Euclidean spaces and fractals, and Markov processes on trees and graphs. It has been shown that Dirichlet form theory is an important tool in study of various invariance principles, such as the invariance principle for reflected Brownian motion on domains with non necessarily smooth boundaries and the invariance principle for Metropolis algorithm. Dirichlet form theory can also be used to study a certain type of SPDEs. Fractals are used as an approximation of disordered media. The analysis on fractals is motivated by the desire to understand properties of natural phenomena such as polymers, and growth of molds and crystals.
    [Show full text]
  • Modern Bayesian Nonparametrics
    Modern Bayesian Nonparametrics Peter Orbanz Yee Whye Teh Cambridge University and Columbia University Gatsby Computational Neuroscience Unit, UCL NIPS 2011 Peter Orbanz & Yee Whye Teh 1 / 72 OVERVIEW 1. Nonparametric Bayesian models 2. Regression 3. Clustering 4. Applications Coffee refill break 5. Asymptotics 6. Exchangeability 7. Latent feature models 8. Dirichlet process 9. Completely random measures 10. Summary Peter Orbanz & Yee Whye Teh 2 / 72 C. E. Rasmussen & C. K. I. Williams, Gaussian Processes for Machine Learning, the MIT Press, 2006, ISBN 026218253X. c 2006 Massachusetts Institute of Technology. www.GaussianProcess.org/gpml 20 Regression 3 PARAMETERS AND PATTERNS2 1 0 Parameters output, y −1 −2 P(Xjθ) =−3 Probability[datajpattern] −5 0 5 input, x (a), =1 3 3 2 2 1 1 0 0 output, y −1 output, y −1 −2 −2 −3 −3 −5 0 5 −5 0 5 input, x input, x (b), =0.3 (c), =3 Figure 2.5: (a) Data is generated from a GP with hyperparameters (, σf , σn)= Inference idea(1, 1, 0.1), as shown by the + symbols. Using Gaussian process prediction with these hyperparameters we obtain a 95% confidence region for the underlying function f (shown in grey). Panels (b) and (c) again show the 95% confidence region, but this timedata for hyperparameter = underlying values pattern (0.3, 1.08, 0 +.00005) independent and (3.0, 1.16, noise0.89) respectively. The covariance is denoted ky as it is for the noisy targets y rather than for the 2 Peter Orbanz & Yee Whye Teh underlying function f. Observe that the length-scale , the signal variance σf 3 / 72 2 hyperparameters and the noise variance σn can be varied.
    [Show full text]
  • Multifractal Analysis of Superprocesses with Stable
    The Annals of Probability 2015, Vol. 43, No. 5, 2763–2809 DOI: 10.1214/14-AOP951 c Institute of Mathematical Statistics, 2015 MULTIFRACTAL ANALYSIS OF SUPERPROCESSES WITH STABLE BRANCHING IN DIMENSION ONE By Leonid Mytnik1 and Vitali Wachtel2 Technion—Israel Institute of Technology and University of Munich We show that density functions of a (α, 1, β)-superprocesses are almost sure multifractal for α > β + 1, β ∈ (0, 1) and calculate the corresponding spectrum of singularities. 1. Introduction, main results and discussion. For 0 < α ≤ 2 and 1+ β ∈ d (1, 2], the so-called (α, d, β)-superprocess X = {Xt : t ≥ 0} in R is a finite measure-valued process related to the log-Laplace equation d (1.1) u =∆ u + au − bu1+β, dt α where a ∈ R and b> 0 are any fixed constants. Its underlying motion is de- α/2 scribed by the fractional Laplacian ∆α := −(−∆) determining a symmet- ric α-stable motion in Rd of index α ∈ (0, 2] (Brownian motion corresponds to α = 2), whereas its continuous-state branching mechanism (1.2) v 7→ −av + bv1+β, v ≥ 0, belongs to the domain of attraction of a stable law of index 1 + β ∈ (1, 2] (the branching mechanism is critical if a = 0). α Let d< β . Then, for any fixed time t, Xt(dx) is a.s. absolutely continuous with respect to the Lebesgue measure (cf. Fleischmann [3] for a = 0). In the case of β = 1, there is a continuous version of the density of Xt(dx) for all α> 1; see Konno and Shiga [12].
    [Show full text]
  • BEJ 22 4.Pdf
    Official Journal of the Bernoulli Society for Mathematical Statistics and Probability Volume Twenty Two Number Four November 2016 ISSN: 1350-7265 CONTENTS Papers MALLER, R.A. 1963 Conditions for a Lévy process to stay positive near 0, in probability LACAUX, C. and SAMORODNITSKY, G. 1979 Time-changed extremal process as a random sup measure BISCIO, C.A.N. and LAVANCIER, F. 2001 Quantifying repulsiveness of determinantal point processes SHAO, Q.-M. and ZHOU, W.-X. 2029 Cramér type moderate deviation theorems for self-normalized processes WANG, M. and MARUYAMA, Y. 2080 Consistency of Bayes factor for nonnested model selection when the model dimension grows LANCONELLI, A. and STAN, A.I. 2101 A note on a local limit theorem for Wiener space valued random variables HUCKEMANN, S., KIM, K.-R., MUNK, A., REHFELDT, F., 2113 SOMMERFELD, M., WEICKERT, J. and WOLLNIK, C. The circular SiZer, inferred persistence of shape parameters and application to early stem cell differentiation DÖRING, H., FARAUD, G. and KÖNIG, W. 2143 Connection times in large ad-hoc mobile networks DELYON, B. and PORTIER, F. 2177 Integral approximation by kernel smoothing CORTINES, A. 2209 The genealogy of a solvable population model under selection with dynamics related to directed polymers ARNAUDON, M. and MICLO, L. 2237 A stochastic algorithm finding p-means on the circle HEAUKULANI, C. and ROY, D.M. 2301 The combinatorial structure of beta negative binomial processes (continued) The papers published in Bernoulli are indexed or abstracted in Current Index to Statistics, Mathematical Reviews, Statistical Theory and Method Abstracts-Zentralblatt (STMA-Z), and Zentralblatt für Mathematik (also avalaible on the MATH via STN database and Compact MATH CD-ROM).
    [Show full text]
  • Random Walks at Random Times: Convergence to Iterated Lévy
    The Annals of Probability 2013, Vol. 41, No. 4, 2682–2708 DOI: 10.1214/12-AOP770 c Institute of Mathematical Statistics, 2013 RANDOM WALKS AT RANDOM TIMES: CONVERGENCE TO ITERATED LEVY´ MOTION, FRACTIONAL STABLE MOTIONS, AND OTHER SELF-SIMILAR PROCESSES By Paul Jung1 and Greg Markowsky2 University of Alabama at Birmingham and Monash University For a random walk defined for a doubly infinite sequence of times, we let the time parameter itself be an integer-valued process, and call the orginal process a random walk at random time. We find the scaling limit which generalizes the so-called iterated Brownian motion. Khoshnevisan and Lewis [Ann. Appl. Probab. 9 (1999) 629–667] suggested “the existence of a form of measure-theoretic duality” be- tween iterated Brownian motion and a Brownian motion in random scenery. We show that a random walk at random time can be consid- ered a random walk in “alternating” scenery, thus hinting at a mech- anism behind this duality. Following Cohen and Samorodnitsky [Ann. Appl. Probab. 16 (2006) 1432–1461], we also consider alternating random reward schema asso- ciated to random walks at random times. Whereas random reward schema scale to local time fractional stable motions, we show that the alternating random reward schema scale to indicator fractional stable motions. Finally, we show that one may recursively “subordinate” random time processes to get new local time and indicator fractional stable motions and new stable processes in random scenery or at random times. When α = 2, the fractional stable motions given by the recur- sion are fractional Brownian motions with dyadic H ∈ (0, 1).
    [Show full text]