DOCSLIB.ORG

Rates of Convergence for Distributed Average Consensus with Probabilistic Quantization

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Rates of Convergence for Distributed Average Consensus with Probabilistic Quantization Rates of Convergence for Distributed Average Consensus with Probabilistic Quantization Tuncer C. Aysal, Mark J. Coates, Michael G. Rabbat Department of Electrical and Computer Engineering McGill University, Montr´eal, Qu´ebec, Canada {tuncer.aysal, mark.coates, michael.rabbat}@mcgill.ca Abstract— Probabilistically quantized distributed averaging the weighted linear combination, (PQDA) is a fully decentralized algorithm for performing aver- age consensus in a network with finite-rate links. At each itera- xi(t +1)= Wi,ixi(t)+ Wi,j xj (t). tion, nodes exchange quantized messages with their immediate j:(i,j)∈E neighbors. Then each node locally computes a weighted average X of the messages it received, quantizes this new value using a For a given initial state, x(0), and reasonable choices of randomized quantization scheme, and then the whole process weights Wi,j , it is easy to show that limt→∞ xi(t) = is repeated in the next iteration. In our previous work we 1 N , N i=1 xi(0) x. The DA algorithm was introduced by introduced PQDA and demonstrated that the algorithm almost Tsitsiklis in [17], and has since been pursued in various surely converges to a consensus (i.e., every node converges to P the same value). The present article builds upon this work forms by many other researchers (e.g., [3,6,8,11,14,20]). by characterizing the rate of convergence to a consensus. Of course, in any practical implementation of this algo- We illustrate that the rate of PQDA is essentially the same rithm, communication rates between neighboring nodes will as unquantized distributed averaging when the discrepancy be finite, and thus quantization must be applied to xi(t) among node values is large. When the network has nearly before it can be transmitted. In applications where heavy converged and all nodes’ values are at one of two neighboring quantization points, then the rate of convergence slows down. quantization must be applied (e.g., when executing multiple We bound the rate of convergence during this final phase by consensus computations in parallel, so that each packet trans- applying lumpability to compress the state space, and then using mission carries many values), quantization can actually affect stochastic comparison methods. convergence properties of the algorithm. Figure 1 shows the trajectories, xi(t), for all nodes superimposed on one set of axes. In Fig. 1(a), nodes apply deterministic uniform I. INTRODUCTION quantization with ∆ =0.1 spacing between quantization points. Although the algorithm converges in this example, A fundamental problem in decentralized networked sys- clearly the limit is not a consensus; not all nodes arrive at tems is that of having nodes reach a state of agreement. the same value. For example, the nodes in a wireless sensor network must In [1,2] we introduced probabilistically quantized dis- be synchronized in order to communicate using a TDMA tributed averaging (PQDA). Rather than applying deter- scheme or to use time-difference-of-arrival measurements for ministic uniform quantization, nodes independently apply a localization and tracking. Similarly, one would like a host of simple randomized quantization scheme. In this scheme, the unmanned aerial vehicles to make coordinated decisions on random quantized value is equal to the original unquantized a surveillance strategy. This paper focuses on a prototypical value in expectation. Through this use of randomization, we example of agreement in networked systems, namely, the guarantee that PQDA converges almost surely to a consensus. average consensus problem: each node initially has a scalar However, since x¯ is not divisible by ∆ in general, the 1 N value, yi, and the goal is to compute the average, N i=1 yi value we converge to is not precisely the average of the at every node in the network. initial values. We presented characteristics of the limiting P Distributed averaging (DA) is a simple iterative distributed consensus value, in particular, showing that it is equal to x¯ algorithm for solving the average consensus problem with in expectation. many attractive properties. The network state is maintained The main contribution of the present paper is to character- N in a vector x(t) R , where xi(t) is the value at node i after ize the rates of convergence for PQDA. We show that, when ∈ t iterations, and there are N nodes in the network. Network the values xi(t) lie in an interval which is large relative to the connectivity is represented by a graph G = (V, E), with quantizer precision ∆, then PQDA moves at the same rate vertex set V = 1,...,N and edge set E V 2 such that as regular unquantized consensus. On the other hand, the (i, j) E implies{ that nodes} i and j communicate⊆ directly transition from when node values are all within ∆ of each with each∈ other. (We assume communication is symmetric.) other (i.e., each node is either at k∆ or (k+1)∆, for some k), In the t+1st DA iteration, node i receives values xj (t) from is slower than unquantized consensus. We present a scheme all the nodes j in its neighborhood and updates its value by for characterizing the time from when PQDA iterates enter 1.8 1.8 1.6 1.6 1.4 1.4 1.2 1.2 1 1 0.8 0.8 0.6 0.6 0.4 0.4 Individual Node Trajectories Individual Node Trajectories 0.2 0.2 Probabilistically Quantized Consensus Deterministic Uniformly Quantized Consensus 0 0 −0.2 −0.2 10 20 30 40 50 60 70 80 10 20 30 40 50 60 70 80 Iteration Number Iteration Number (a) (b) Fig. 1. Individual node trajectories (i.e., xi(t), ∀i) taken by the distributed average consensus using (a:) deterministic uniform quantization and (b:) probabilistic quantization. The number of nodes is N = 50, the nodes’ initial average is x = 0.85, and the quantization resolution is set to ∆ = 0.1. The consensus value, in this case, is 0.8. this “final bin” until the algorithm is absorbed at a consensus the quantization noise diminishes and nodes converge to a state. consensus. The remainder of the paper is organized as follows. Kashyap et al. examine the effects of quantization in Section II describes the probabilistic quantization scheme consensus algorithms from a different point of view [7]. employed by PQDA. Section III formally defines the PQDA N They require that the network average, x¯ =1/N i=1 xi(t), algorithm and lists fundamental convergence properties. Sec- be preserved at every iteration. To do this using quantized tion IV explores rates of convergence when the node values transmissions, nodes must carefully account forP round-off are far from a consensus, relative to the quantization preci- errors. Suppose we have a network of N nodes and let ∆ sion ∆. Section V presents our technique for characterizing denote the “quantization resolution” or distance between to convergence rates when all nodes reach the “final bin” and quantization lattice points. If x¯ is not a multiple of N∆, are within ∆ of each other. We conclude in Section VI. then it is not possible for the network to reach a strict Before proceeding, we briefly review related work. consensus (i.e., limt→∞ maxi,j xi(t) xj (t) = 0) while also preserving the network average,| −x¯, since| nodes only A. Related Work ever exchange units of ∆. Instead, Kashyap et. al define the While there exists a substantial body of work on average notion of a “quantized consensus” to be such that all xi(t) consensus protocols with infinite precision and noise–free take on one of two neighboring quantization values while peer–to–peer communications, little research has been done preserving the network average; i.e., xi(t) k∆, (k+1)∆ ∈{ } introducing distortions in the message exchange. In [12], for all i and some k, and i xi(T )= Nx¯. They show that, Rabani, Sinclair, and Wanka examine distributed averaging under reasonable conditions, their algorithm will converge P as a mechanism for load balancing in distributed computing to a quantized consensus. However, the quantized consensus systems. They provide a bound on the divergence between is clearly not a strict consensus, i.e., all nodes do not have quantized consensus trajectories, q(t), and the trajectories the same value. Even when the algorithm has converged, as which would be taken by an unquantized averaging algo- many as half the nodes in the network may have different rithms. The bound reduces to looking at properties of the values. If nodes are strategizing and/or performing actions averaging matrix W. based on these values (e.g., flight formation), then differing Recently, Yildiz and Scaglione, in [22], explored the values may lead to inconsistent behavior. impact of quantization noise through modification of the Of note is that the related works discussed above all consensus algorithm proposed by Xiao and Boyd [21]. They utilize standard deterministic uniform quantization schemes note that the noise component in [21] can be considered to quantize the data. In contrast to [22], where quantization as the quantization noise and they develop an algorithm noise terms are modeled as independent zero-mean random for predicting neighbors’ unquantized values in order to variables, we explicitly introduce randomization in our quan- correct errors introduced by quantization [22]. Simulation tization procedure. Careful analysis of this randomization studies for small N indicate that if the increasing correlation allows us to provide concrete theoretical rates of convergence among the node states is taken into account, the variance of in addition to empirical results. Moreover, the algorithm proposed in this paper converges to a strict consensus, as for some quantization point τ ∗ τ . opposed to the approximate “quantized consensus” achieved Proofs of all the theorems stated∈ in this section can be found in [7] which is clearly not a strict consensus and does not in [2].
Load more

Recommended publications
  • Numerical Solution of Ordinary Differential Equations
    NUMERICAL SOLUTION OF ORDINARY DIFFERENTIAL EQUATIONS Kendall Atkinson, Weimin Han, David Stewart University of Iowa Iowa City, Iowa A JOHN WILEY & SONS, INC., PUBLICATION Copyright c 2009 by John Wiley & Sons, Inc. All rights reserved. Published by John Wiley & Sons, Inc., Hoboken, New Jersey. Published simultaneously in Canada. No part of this publication may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, recording, scanning, or otherwise, except as permitted under Section 107 or 108 of the 1976 United States Copyright Act, without either the prior written permission of the Publisher, or authorization through payment of the appropriate per-copy fee to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, (978) 750-8400, fax (978) 646-8600, or on the web at www.copyright.com. Requests to the Publisher for permission should be addressed to the Permissions Department, John Wiley & Sons, Inc., 111 River Street, Hoboken, NJ 07030, (201) 748-6011, fax (201) 748-6008. Limit of Liability/Disclaimer of Warranty: While the publisher and author have used their best efforts in preparing this book, they make no representations or warranties with respect to the accuracy or completeness of the contents of this book and specifically disclaim any implied warranties of merchantability or fitness for a particular purpose. No warranty may be created ore extended by sales representatives or written sales materials. The advice and strategies contained herin may not be suitable for your situation. You should consult with a professional where appropriate. Neither the publisher nor author shall be liable for any loss of profit or any other commercial damages, including but not limited to special, incidental, consequential, or other damages.
    [Show full text]
  • Polarization Fields and Phase Space Densities in Storage Rings: Stroboscopic Averaging and the Ergodic Theorem
    Physica D 234 (2007) 131–149 www.elsevier.com/locate/physd Polarization fields and phase space densities in storage rings: Stroboscopic averaging and the ergodic theorem✩ James A. Ellison∗, Klaus Heinemann Department of Mathematics and Statistics, University of New Mexico, Albuquerque, NM 87131, United States Received 1 May 2007; received in revised form 6 July 2007; accepted 9 July 2007 Available online 14 July 2007 Communicated by C.K.R.T. Jones Abstract A class of orbital motions with volume preserving flows and with vector fields periodic in the “time” parameter θ is defined. Spin motion coupled to the orbital dynamics is then defined, resulting in a class of spin–orbit motions which are important for storage rings. Phase space densities and polarization fields are introduced. It is important, in the context of storage rings, to understand the behavior of periodic polarization fields and phase space densities. Due to the 2π time periodicity of the spin–orbit equations of motion the polarization field, taken at a sequence of increasing time values θ,θ 2π,θ 4π,... , gives a sequence of polarization fields, called the stroboscopic sequence. We show, by using the + + Birkhoff ergodic theorem, that under very general conditions the Cesaro` averages of that sequence converge almost everywhere on phase space to a polarization field which is 2π-periodic in time. This fulfills the main aim of this paper in that it demonstrates that the tracking algorithm for stroboscopic averaging, encoded in the program SPRINT and used in the study of spin motion in storage rings, is mathematically well-founded.
    [Show full text]
  • Approximations in Numerical Analysis, Cont'd
    Jim Lambers MAT 460/560 Fall Semester 2009-10 Lecture 7 Notes These notes correspond to Section 1.3 in the text. Approximations in Numerical Analysis, cont'd Convergence Many algorithms in numerical analysis are iterative methods that produce a sequence fαng of ap- proximate solutions which, ideally, converges to a limit α that is the exact solution as n approaches 1. Because we can only perform a finite number of iterations, we cannot obtain the exact solution, and we have introduced computational error. If our iterative method is properly designed, then this computational error will approach zero as n approaches 1. However, it is important that we obtain a sufficiently accurate approximate solution using as few computations as possible. Therefore, it is not practical to simply perform enough iterations so that the computational error is determined to be sufficiently small, because it is possible that another method may yield comparable accuracy with less computational effort. The total computational effort of an iterative method depends on both the effort per iteration and the number of iterations performed. Therefore, in order to determine the amount of compu- tation that is needed to attain a given accuracy, we must be able to measure the error in αn as a function of n. The more rapidly this function approaches zero as n approaches 1, the more rapidly the sequence of approximations fαng converges to the exact solution α, and as a result, fewer iterations are needed to achieve a desired accuracy. We now introduce some terminology that will aid in the discussion of the convergence behavior of iterative methods.
    [Show full text]
  • On the Three Methods for Bounding the Rate of Convergence for Some Continuous-Time Markov Chains
    On the Three Methods for Bounding the Rate of Convergence for some Continuous-time Markov Chains Alexander Zeifman∗ Yacov Satiny Anastasia Kryukovaz Rostislav Razumchik§ Ksenia Kiseleva{ Galina Shilovak Abstract. Consideration is given to the three different analytical methods for the computation of upper bounds for the rate of convergence to the limiting regime of one specific class of (in)homogeneous continuous- time Markov chains. This class is particularly suited to describe evolutions of the total number of customers in (in)homogeneous M=M=S queueing systems with possibly state-dependent arrival and service intensities, batch arrivals and services. One of the methods is based on the logarithmic norm of a linear operator function; the other two rely on Lyapunov functions and differential inequalities, respectively. Less restrictive conditions (compared to those known from the literature) under which the methods are applicable, are being formulated. Two numerical examples are given. It is also shown that for homogeneous birth-death Markov processes defined on a finite state space with all transition rates being positive, all methods yield the same sharp upper bound. ∗Vologda State University, Vologda, Russia; Institute of Informatics Problems, Federal Research Center “Computer Science and Control”, Russian Academy of Sciences, Moscow, arXiv:1911.04086v2 [math.PR] 24 Feb 2020 Russia; Vologda Research Center of the Russian Academy of Sciences, Vologda, Russia. E-mail: [email protected] yVologda State University, Vologda, Russia zVologda State
    [Show full text]
  • CONVERGENCE RATES of MARKOV CHAINS 1. Orientation 1
    CONVERGENCE RATES OF MARKOV CHAINS STEVEN P. LALLEY CONTENTS 1. Orientation 1 1.1. Example 1: Random-to-Top Card Shuffling 1 1.2. Example 2: The Ehrenfest Urn Model of Diffusion 2 1.3. Example 3: Metropolis-Hastings Algorithm 4 1.4. Reversible Markov Chains 5 1.5. Exercises: Reversiblity, Symmetries and Stationary Distributions 8 2. Coupling 9 2.1. Coupling and Total Variation Distance 9 2.2. Coupling Constructions and Convergence of Markov Chains 10 2.3. Couplings for the Ehrenfest Urn and Random-to-Top Shuffling 12 2.4. The Coupon Collector’s Problem 13 2.5. Exercises 15 2.6. Convergence Rates for the Ehrenfest Urn and Random-to-Top 16 2.7. Exercises 17 3. Spectral Analysis 18 3.1. Transition Kernel of a Reversible Markov Chain 18 3.2. Spectrum of the Ehrenfest random walk 21 3.3. Rate of convergence of the Ehrenfest random walk 23 1. ORIENTATION Finite-state Markov chains have stationary distributions, and irreducible, aperiodic, finite- state Markov chains have unique stationary distributions. Furthermore, for any such chain the n−step transition probabilities converge to the stationary distribution. In various ap- plications – especially in Markov chain Monte Carlo, where one runs a Markov chain on a computer to simulate a random draw from the stationary distribution – it is desirable to know how many steps it takes for the n−step transition probabilities to become close to the stationary probabilities. These notes will introduce several of the most basic and important techniques for studying this problem, coupling and spectral analysis. We will focus on two Markov chains for which the convergence rate is of particular interest: (1) the random-to-top shuffling model and (2) the Ehrenfest urn model.
    [Show full text]
  • Convergence Rates for the Quantum Central Limit Theorem
    Commun. Math. Phys. 383, 223–279 (2021) Communications in Digital Object Identifier (DOI) https://doi.org/10.1007/s00220-021-03988-1 Mathematical Physics Convergence Rates for the Quantum Central Limit Theorem Simon Becker1 , Nilanjana Datta1, Ludovico Lami2,3, Cambyse Rouzé4 1 Department of Applied Mathematics and Theoretical Physics, Centre for Mathematical Sciences University of Cambridge, Cambridge CB3 0WA,UK. E-mail: [email protected]; [email protected] 2 School of Mathematical Sciences and Centre for the Mathematics and Theoretical Physics of Quantum Non-Equilibrium Systems, University of Nottingham, University Park, Nottingham NG7 2RD, UK. 3 Institute of Theoretical Physics and IQST, Universität Ulm, Albert-Einstein-Allee, 89069 Ulm, Germany. E-mail: [email protected] 4 Zentrum Mathematik, Technische Universität München, 85748 Garching, Germany. E-mail: [email protected] Received: 20 February 2020 / Accepted: 15 January 2021 Published online: 15 February 2021 – © The Author(s) 2021 Contents 1. Introduction ................................. 224 2. Notation and Definitions ........................... 228 2.1 Mathematical notation ......................... 228 2.2 Definitions ................................ 229 2.2.1 Quantum information with continuous variables .......... 229 2.2.2 Phase space formalism ....................... 231 2.3 Moments ................................ 233 2.4 Quantum convolution .......................... 235 3. Cushen and Hudson’s Quantum Central Limit Theorem .......... 236 4. Main Results ................................. 237 4.1 Quantitative bounds in the QCLT .................... 237 4.2 Optimality of convergence rates and necessity of finite second moments in the QCLT ............................... 239 4.3 Applications to capacity of cascades of beam splitters with non-Gaussian environment ............................... 240 4.4 New results on quantum characteristic functions ............ 242 5. New Results on Quantum Characteristic Functions: Proofs ........
    [Show full text]
  • High Order Regularization of Dirac-Delta Sources in Two Space Dimensions
    INTERNSHIP REPORT High order regularization of Dirac-delta sources in two space dimensions. Author: Supervisors: Wouter DE VRIES Prof. Guustaaf JACOBS s1010239 Jean-Piero SUAREZ HOST INSTITUTION: San Diego State University, San Diego (CA), USA Department of Aerospace Engineering and Engineering Mechanics Computational Fluid Dynamics laboratory Prof. G.B. JACOBS HOME INSTITUTION University of Twente, Enschede, Netherlands Faculty of Engineering Technology Group of Engineering Fluid Dynamics Prof. H.W.M. HOEIJMAKERS March 4, 2015 Abstract In this report the regularization of singular sources appearing in hyperbolic conservation laws is studied. Dirac-delta sources represent small particles which appear as source-terms in for example Particle-laden flow. The Dirac-delta sources are regularized with a high order accurate polynomial based regularization technique. Instead of regularizing a single isolated particle, the aim is to employ the regularization technique for a cloud of particles distributed on a non- uniform grid. By assuming that the resulting force exerted by a cloud of particles can be represented by a continuous function, the influence of the source term can be approximated by the convolution of the continuous function and the regularized Dirac-delta function. It is shown that in one dimension the distribution of the singular sources, represented by the convolution of the contin- uous function and Dirac-delta function, can be approximated with high accuracy using the regularization technique. The method is extended to two dimensions and high order approximations of the source-term are obtained as well. The resulting approximation of the singular sources are interpolated using a polynomial interpolation in order to regain a continuous distribution representing the force exerted by the cloud of particles.
    [Show full text]
  • Strongly Mixed Random Errors in Mann's Iteration Algorithm for a Contractive Real Function
    Pr´e-Publica¸c~oesdo Departamento de Matem´atica Universidade de Coimbra Preprint Number 17{29 STRONGLY MIXED RANDOM ERRORS IN MANN'S ITERATION ALGORITHM FOR A CONTRACTIVE REAL FUNCTION HASSINA ARROUDJ, IDIR ARAB AND ABDELNASSER DAHMANI Abstract: This work deals with the Mann's stochastic iteration algorithm under α−mixing random errors. We establish the Fuk-Nagaev's inequalities that enable us to prove the almost complete convergence with its corresponding rate of con- vergence. Moreover, these inequalities give us the possibility of constructing a confidence interval for the unique fixed point. Finally, to check the feasibility and validity of our theoretical results, we consider some numerical examples, namely a classical example from astronomy. Keywords: Fixed point; Fuk-Nagaev's inequalities; Mann's iteration algorithm; α-mixing; Rate of convergence; confidence domain. 1. Introduction In many mathematical problems arising from various domains, the exis- tence of a solution is the same as the existence of a fixed point by some appropriate transformation of the problem. The most known problem in that framework is the root existence which can be tackled easily as the ex- istence of a fixed point and vice versa. Therefore, the fixed point theory is of paramount importance in engineering sciences and many areas of mathe- matics. Fixed point theory provides conditions under which maps have the existence and uniqueness of solutions. Over the last decades, that theory has been revealed as one of the most significant tool in the study of nonlinear problems. In particular, in many fields, equilibria or stability are fundamen- tal concepts that can be described in terms of fixed points.
    [Show full text]
  • Convergence of Markov Processes
    Convergence of Markov Processes May 26, 2021 Martin Hairer Mathematics Department, Imperial College London Email: [email protected] Abstract The aim of this minicourse is to provide a number of tools that allow one to de- termine at which speed (if at all) the law of a diffusion process, or indeed a rather general Markov process, approaches its stationary distribution. Of particular in- terest will be cases where this speed is subexponential. After an introduction to the general ergodic theory of Markov processes, the first part of the course is devoted to Lyapunov function techniques. The second part is then devoted to an elementary introduction to Malliavin calculus and to a proof of Hormander’s¨ famous ”sums of squares” regularity theorem. 1 General (ergodic) theory of Markov processes In this note, we are interested in the long-time behaviour of Markov processes, both in discrete and continuous time. Recall that a discrete-time Markov process x on a state space X is described by a transition kernel P, which we define as a measurable map from X into the space of probability measures on X . In all that follows, X will always be assumed to be a Polish space, that is a complete, separable metric space. When viewed as a measurable space, we will always endow it with its natural Borel σ-algebra, that is the smallest σ-algebra containing all open sets. This ensures that X endowed with any probability measure is a Lebesgue space and that basic intuitive results of probability and measure theory (Fubini’s theorem, regular conditional probabilities, etc) are readily available.
    [Show full text]
  • On Spectral Accuracy of Quadrature Formulae Based on Piecewise Polynomial Interpolation
    ON SPECTRAL ACCURACY OF QUADRATURE FORMULAE BASED ON PIECEWISE POLYNOMIAL INTERPOLATION A. KURGANOV∗ AND S. TSYNKOV† Abstract. It is well-known that the trapezoidal rule, while being only second-order accurate in general, improves to spectral accuracy if applied to the integration of a smooth periodic function over an entire period on a uniform grid. More precisely, for the function that has a square integrable derivative of order r the convergence rate is o N −(r−1/2), where N is a number of grid nodes. Accordingly, for a C∞-function the trapezoidal quadrature converges with the rate faster than any polynomial. In this paper, we prove that the same property holds for all quadrature formulae obtained by integrating fixed degree piecewise polynomial interpolations of a smooth integrand, such as the midpoint rule, Simpson’s rule, etc. Key words. Numerical quadrature, piecewise polynomial, convergence rate, trapezoidal rule, midpoint rule, Simpson’s rule, spectral accuracy. Let f = f(x) be a function defined in the interval a ≤ x ≤ b, |b − a| = L. To approximate the value Z b of its integral f(x) dx, let us introduce a partition of the interval [a, b] into N equal subintervals of size a h = L/N. The partition is rendered by the nodes of the grid: xj = a + jh, j = 0, 1,...,N, so that x0 = a and xN = b. On each subinterval [xj, xj+1], j = 0, 1,...,N − 1, we approximate the function f by the interpolating polynomial of degree 1: x − x x − x f(x) ≈ Q (x, f, x , x ) = f j+1 + f j , (1) 1 j j+1 j h j+1 h where fj = f(xj), j = 0, 1,...,N.
    [Show full text]
  • On Almost Sure Rates of Convergence for Sample Average Approximations
    On almost sure rates of convergence for sample average approximations Dirk Banholzer1, J¨orgFliege1, and Ralf Werner2 1 Department of Mathematical Sciences, University of Southampton, Southampton, SO17 1BJ, UK [email protected] 2 Institut f¨urMathematik, Universit¨atAugsburg 86159 Augsburg, Germany Abstract. In this paper, we study the rates at which strongly con- sistent estimators in the sample average approximation approach (SAA) converge to their deterministic counterparts. To be able to quantify these rates at which convergence occurs in the almost sure sense, we consider the law of the iterated logarithm in a Banach space setting and first establish under relatively mild assumptions convergence rates for the approximating objective functions. These rates can then be transferred to the estimators for optimal values and solutions of the approximated problem. Based on these results, we further investigate the asymptotic bias of the SAA approach. Especially, we show that under the same as- sumptions as before the SAA estimator for the optimal value converges in mean to its deterministic counterpart, at a rate which essentially co- incides with the one in the almost sure sense. Further, optimal solutions also convergence in mean; for convex problems with the same rates, but with an as of now unknown rate for general problems. Finally, we address the notion of convergence in probability and conclude that in this case (weak) rates for the estimators can also be derived, and that without im- posing strong conditions on exponential moments, as it is typically done for obtaining exponential rates. Results on convergence in probability then allow to build universal confidence sets for the optimal value and optimal solutions under mild assumptions.
    [Show full text]
  • A Statistical Application in Astronomy: Streaming Motion in Leo I
    Streaming motion in Leo I Separating signal from background A statistical application in astronomy: Streaming motion in Leo I Bodhisattva Sen DPMMS, University of Cambridge, UK Columbia University, New York, USA [email protected] ETH Zurich, Switzerland 31 May, 2012 Streaming motion in Leo I Separating signal from background Outline 1 Streaming motion in Leo I Modeling Threshold models 2 Separating signal from background Method Theory Streaming motion in Leo I Modeling Separating signal from background Threshold models What is streaming motion? Local Group dwarf spheroidal (dSph) galaxies: small, dim Is Leo I in equilibrium or tidally disrupted by Milky Way? Such a disruption can give rise to streaming motion: the leading and trailing stars move away from the center of the main body of the perturbed system Streaming motion in Leo I Modeling Separating signal from background Threshold models Data on 328 stars (Y ; Σ): Line of sight velocity; std. dev. of meas. error (R; Θ): Projected position, orthogonal to line of sight Contaminated with foreground stars (in the line of sight) Magnitude of streaming motion Likely to increase beyond a threshold radius Likely to be aligned with the major axis of the system Streaming motion in Leo I Modeling Separating signal from background Threshold models Statistical questions Is streaming motion evident in Leo I? If so, how can it be described and estimated? To what extent can it be described by a threshold model? Findings [Sen et al. (AoAS, 2009)]: The magnitude of streaming motion appears to
    [Show full text]