Discover millions of ebooks, audiobooks, and so much more with a free trial

Only $11.99/month after trial. Cancel anytime.

Nonlinear Filtering and Smoothing: An  Introduction to Martingales, Stochastic Integrals and Estimation
Nonlinear Filtering and Smoothing: An  Introduction to Martingales, Stochastic Integrals and Estimation
Nonlinear Filtering and Smoothing: An  Introduction to Martingales, Stochastic Integrals and Estimation
Ebook515 pages3 hours

Nonlinear Filtering and Smoothing: An Introduction to Martingales, Stochastic Integrals and Estimation

Rating: 0 out of 5 stars

()

Read preview

About this ebook

Most useful for graduate students in engineering and finance who have a basic knowledge of probability theory, this volume is designed to give a concise understanding of martingales, stochastic integrals, and estimation. It emphasizes applications. Many theorems feature heuristic proofs; others include rigorous proofs to reinforce physical understanding. Numerous end-of-chapter problems enhance the book's practical value.
After introducing the basic measure-theoretic concepts of probability and stochastic processes, the text examines martingales, square integrable martingales, and stopping times. Considerations of white noise and white-noise integrals are followed by examinations of stochastic integrals and stochastic differential equations, as well as the associated Ito calculus and its extensions. After defining the Stratonovich integral, the text derives the correction terms needed for computational purposes to convert the Ito stochastic differential equation to the Stratonovich form. Additional chapters contain the derivation of the optimal nonlinear filtering representation, discuss how the Kalman filter stands as a special case of the general nonlinear filtering representation, apply the nonlinear filtering representations to a class of fault-detection problems, and discuss several optimal smoothing representations.
LanguageEnglish
Release dateOct 17, 2013
ISBN9780486781839
Nonlinear Filtering and Smoothing: An  Introduction to Martingales, Stochastic Integrals and Estimation

Related to Nonlinear Filtering and Smoothing

Related ebooks

Science & Mathematics For You

View More

Related articles

Reviews for Nonlinear Filtering and Smoothing

Rating: 0 out of 5 stars
0 ratings

0 ratings0 reviews

What did you think?

Tap to rate

Review must be at least 10 words

    Book preview

    Nonlinear Filtering and Smoothing - Venkatarama Krishnan

    Errata

    1.1 INTRODUCTION

    Probability theory is the mathematical study of phenomena occurring due to chance mechanism. If we toss a coin, we cannot say a priori whether we will get heads or tails. Outcomes of a random experiment can be analyzed or modeled only in an abstract manner. A random experiment or a mathematical experiment is one in which the possible outcomes may be finite or infinite. In the experiment of tossing a coin there are two outcomes, heads and tails. In the tossing of a die there are six outcomes. On the other hand, the weight of a full-term new-born baby may vary continuously from 4 to 10 pounds. Each of these outcomes is known as an elementary outcome. The collection of all elementary outcomes of a random experiment is called sample space and is denoted by Ω. In set terminology the sample space is termed the universal set. Thus, the sample space Ω is a set consisting of mutually exclusive, collectively exhaustive listing of all possible outcomes of a random experiment. That is, Ω = {ω¹, ω²,...,ωn} denotes the set of all finite outcomes, Ω = {ω¹, ω²,...} denotes the set of all countably infinite outcomes, and Ω = {0 ≤ t T} denotes the set of uncountably infinite outcomes.

    1.2 ALGEBRA OF SETS

    Let Ω represent the sample space which is a collection of ω-points as defined earlier. The various set operations are (1) complementation, (2) union, and (3) intersection. Let A and B be two subsets of the sample space Ω, denoted by A ⊂ Ω, B ⊂ Ω. The complement of A, denoted by Ac, represents the set of all ω-points not contained in A :

    Evidently the complement of Ω is the empty set Ø. Two sets A and B are equal if and only if A is contained in B and B is contained in A :

    The union of sets A and B, denoted by A B or A + B, represents the occurrence of ω-points in either A or B. Similarly, the intersection of sets A and B, denoted by A B or AB, represents the occurrence of ω-points in A and B. Clearly, if there is no commonality of ω-points in A and B, then A B is the empty set Ø.

    Example 1.2.1

    Let Ω be the ω-points on the real line R.

    Define

    Then the set operations yield

    The unions and intersections of an arbitrary collection of sets are defined by

    where N is an arbitrary index set which may be finite or countably infinite.

    The unions and intersections follow the reflexive, commutative, associative, and distributive laws.

    The complements (∪n N An)c and (∩n N An)c are given by de Morgan’s laws as follows:

    Sequences

    A sequence of sets An, n N, is increasing if An + 1 An and decreasing if An + 1 An for every n N.

    A sequence which is either increasing or decreasing is called a monotone sequence. We can write the limits (N countably infinite) of monotone sequences as

    The limit of monotone sequences {An} is written as An ↑ A when it is increasing and An ↓ A when it is decreasing.

    Example 1.2.2

    Let Ω be the real line R. If An = {ω: 0 < ω < a − 1/n}, then An ↑ A = {ω: 0 < ω < a}. On the other hand, if Bn = {ω: 0 < ω < a + 1/n}, then Bn B = {ω: 0 < ω a}.

    We can define a superior limit and an inferior limit for any sequence {An} not necessarily monotone. We first define sequences {Bn} and {Cn} derived from {An} as follows:

    Clearly the sequences {Bn} and {Cn} are monotone and decreasing and increasing, respectively. We can now define a limit from eq. 1.2.7 for these monotone sequences:

    Hence

    If lim supnAn = lim infnAn, then {An} is a convergent sequence and limnAn = A, say, exists, that is,

    Example 1.2.3

    Let Ak be the set of points (x, y) of the Cartesian plane R² in the region {0 ≤ x < k, 0 ≤ y < 1/k}, that is,

    Here {Ak} does not belong to the monotone class.

    But

    is a decreasing sequence, and hence

    Similarly,

    is an increasing sequence, and hence

    Since lim supnAn = lim infnBn = {x, y R²: 0 ≤ x < ∞, y = 0}, we have limnAn = B = C = {x, y R²: 0 ≤ x < ∞, y = 0}.

    Example 1.2.4

    Let Ω be the positive real line R+. Consider the sequence

    Here

    Similarly

    Since lim supnAn ≠ lim infnAn, the sequence {An} does not converge and has no limit.

    Even though the limit of a sequence may not exist, superior and inferior limits will always exist, as is evident from the definitions.

    1.3 FIELDS, σ-FIELDS, AND EVENTS

    is a field or an algebra of sets in Ω if it satisfies the following definition.

    Definition 1.3.1 Field or Algebra

    A class of a collection of subsets Aj is a field when the following conditions are satisfied:

    Given the above two conditions, de Morgan’s law ensures that finite intersections also belong to the field. Thus a class of subsets is a field if and only if it is closed under all finite set operations like unions, intersections, and complementations. Since every Boolean algebra of sets is isomorphic to an algebra of subsets of Ω, we can also call the field a Boolean field or Boolean algebra. Every field contains as elements the sample space Ω and the empty set Ø.

    Example 1.3.1

    Let Ω = R of all intervals of the form (a, b], that is, {x ∈ R: a < x ≤ b}:

    is closed under intersections. However,

    is not a field.

    Example 1.3.2

    The smallest field containing A ⊂ Ω is

    If a class of subsets is closed under finite set operations, it does not necessarily mean that it is also closed under countably infinite set operations. Very often we come across the sequence of sets {An} as n → ∞ and the convergence of such sequences (lim supnAn, lim infnAn). In is called a σ-field or σ-algebra.

    Definition 1.3.2 σ-Field or σ-Algebra

    A class of a countably infinite collection of subsets Aj is a σ-field when the following conditions are satisfied:

    In general a σ-field is a field, but a field may not be a σ-field.

    Example 1.3.3

    Let Ω = R be the class of all intervals of the form (− ∞, a], (b, c], and (d, ∞):

    From is a field. However, for infinite intersections of the form

    is not a σ-field.

    Proposition 1.3.1 Intersection of σ-Fields

    The intersection of any nonempty but arbitrary collection of σ-fields in Ω is a σ-field in Ω.

    In general the arbitrary union of a collection of σ-fields may not be a σ-field.

    , then

    Example 1.3.4

    Let the sample space Ω contain ω-points of the toss of a die. Ω is the set {1, 2, 3, 4, 5, 6}. We shall now define a class of sets

    by

    .

    So far we have not considered the nature of the sample space Ω, except that it is nonempty. A set A and a class of subsets of A called open sets of A, such that this class contains Ø and A, and closed under finite intersections and arbitrary unions, is called a topologicai space. If Ω is a topological space, an example being the real line R, the minimum σ-field over the collection of open sets of the topological space is called the Borel σ-field or Borel field, as given by the following definition.

    Definition 1.3.3 Borel σ-Field

    The minimum σ-field generated by the collection of open sets of a topological space Ω is called the Borel σ-field or Borel field. Members of this σ-field are called Borel sets.

    Clearly the Borel σ-field is a σ-field, and hence each closed set is also a Borel set.

    The important topological space with which we will be concerned is the real line R. . From the relationships

    we find that the intervals (a, b], [a, b) and [a, bcontains all subsets of the form given above and their complements, countable unions, and intersections, Each set {b} = [b, b] = (– ∞, b) ∩ (b, ∞) consisting of a single point b , and so are countable unions of single points.

    ) thus created is called a measurable space. Subsets of Ω which are elements in the σ-field are called events. Elements of Ω are points.

    If {Ai, i = 1, 2,...,n, then the {Ai} collectively exhaust Ω. The class {Ai} is called a partition of Ω.

    Definition 1.3.4 Atom

    be a σ-field of subsets of a sample space Ω. We call a set A ≠ Ø an atom if A and if

    ².

    Example 1.3.5

    In Example 1.3.4 where Ω = {1, 2, 3, 4, 5, 6} we define two σ-fields

    ¹.

    1.4 PROBABILITY SPACE

    A probability measure is a set function P of subsets of a sample space Ω such that it satisfies the following axioms of Kolmogorov for any A :

    Any set function μ ) satisfying axioms 1 and 3 is called a measure. A probability measure is a normed or scaled measure because of axiom 2. There are other measures, such as counting measure, Lebesgue-measure, and spectral measure, which are not necessarily normalized. However, any bounded measure with suitable normalization can be converted into a probability measure. If μ(A) is finite for each A , then μ is a finite measure. However, if μ(A) = ∞ but if there exists a sequence {Anand μ(An) is finite for each n, then μ is a σ-finite measure. An example of a measure that is σ-finite is length defined on the space (R, , μ) is a measure space. , P) is called a probability space, which serves to describe any random experiment where

    1.  Ω is a nonempty set called the sample space, whose elements are the elementary outcomes of a random experiment.

    is a σ-field of subsets of Ω.

    3.  P ).

    A signed measure (Halmos, 18, p. 118) is a σ-additive set function μ ) taking positive and negative values such that μ(Ø) = 0 and assuming at most one of two values, + ∞ or − ∞.

    As an example if we define a measure v(A) = μ¹(A) – μ²(A), then v is a signed measure. The Jordan–Hahn decomposition theorem (Halmos, 18, p. 123) guarantees that a signed measure v can be given as the difference between two measures, v = v+ − v− one of which is finite.

    We would now like to enquire whether a probability measure P . The answer is in the affirmative, as given by the following proposition (Halmos, 18, p. 54).

    Proposition 1.4.1 Extension of Probability Measure

    . Let P )) such that for any A (A) = P(A).

    Example 1.4.1

    Let Ω = R be the σ-field generated by the class of all intervals. For any finite interval (a, b] we shall define a measure μ given by μ((a, b]) = b − a, b > a. Then μ can be extended to a unique σ-finite measure on (R). Note that the measure of any countable number of points on the real line is zero.

    Completion (Doob, 14, pp. 48, 606)

    , P) and has a measure 0, that is, P(Λ⁰) = 0. Hence Λ⁰ is an event. Let Λ be a proper subset of Λ⁰, Λ ⊂ Λ⁰. If Λ and Λ⁰ are two subsets with the property that the occurrence of Λ⁰ implies the occurrence of Λ, and if P(Λ⁰) = 0, then Λ need neither be assigned a probability nor can it be called an event. However, if a probability is assigned to Λ, then necessarily it has to have a measure 0 for consistency. We denote subsets of this type with zero probability as null sets. and the null sets. By the Extension Proposition 1.4.1, the probability measure P . This is called completion of the probability measure P. ) is called the complete probability space, from P , P) is complete.

    In Example 1.4.1 if the σ-finite measure μ is completed, then we have the Lebesgue measure (Halmos, 18, p. 62).]

    Product Spaces

    We now define the concept of product spaces and product measures. If ω¹ ∈ Ω¹ and ω² ∈ Ω², then the set of all pairs (ω¹, ω²) is called the Cartesian product ²) be two measurable spaces and let sets A¹ and A², respectively. Then the set {(ω¹, ω²): ω¹ ∈ A¹, ω² ∈ A²} is a measurable rectangle A¹ × A². However, the class {A¹ × A²: A¹, A²} may not be a σ-field. We will call the minimum σ-field generated by this class the product ¹, μ², μ²) be two σ-finite measure spaces. The product measure ². It is a σ-finite measure such that for every measurable rectangle A¹ × A², (A¹, A²)

    Next we establish the concept of ω-sections²) be a product measure space. Let B be a set in Ω¹ × Ω². The ω¹-section of B, denoted by ¹, is defined as

    and the ω²-section of B, denoted by ², is defined as

    As an example, if B is the rectangle A¹ × A², then ¹ = A² and ² = A¹.

    We now state Fubini’s Theorem, which relates the integrals on a product space to integrals on component spaces.

    Theorem 1.4.1 Fubini’s Theorem (Halmos, 18, p. 148)

    ¹, μ², μ²) be two σ-finite measure spaces and let λ be a product measure μ¹ × μ². Let h be an integrable function on Ω¹ × Ω². Then:

    1.  Almost every ω-section of h is integrable.

    2.  If functions f¹ and f² are defined by

    and f¹ and f² are integrable, then

    In Section 1.2 we discussed the algebra of sets. Since events are sets, we can also have a sequence of events {An} where we can define supnAn, infnAn, lim supnAn and infnAn. We now discuss the convergence of the probability of a sequence of events.

    Lemma 1.4.1 Sequential Monotone Continuity

    Let {Ansuch that An + ¹ ⊂ An, and let limn → ∞An = Ø. Then

    The probability measure is said to satisfy the sequential monotone continuity at Ø.

    Proposition 1.4.2 Sequential Continuity

    Let {An, with limnAn = A. Then

    The probability measure is sequentially continuous.

    Proof

    1.  If A = Ø, then this is exactly Lemma 1.4.1.

    2.  If A is a nonempty set and {An} is a monotone sequence, An ↓ A, (Figure 1.4.1),

    FIGURE 1.4.1

    FIGURE 1.4.2

    since (An − A) and

    Enjoying the preview?
    Page 1 of 1