Discrete-Time Stationary Stochastic Processes Lecture Notes
Discrete-Time Stationary Stochastic Processes Lecture Notes
Lecture Notes
Marco Lippi
Dipartimento di Scienze Economiche
Università di Roma “La Sapienza”
Introduction
These Lecture Notes assume that the reader is already acquainted with the basic
notions on stochastic processes and stationarity. The present text does not pretend
to be self-contained. Rather, I will try to provide a lively presentation, just like in
a lecture course. More precisely, the Notes contain:
2. A guide to the proof of the main results. Again, examples and counterexam-
ples will be very often used as substitutes of rigorous proofs, the latter being
left to the readers to work out themselves or to look up in reference texts.
4. A guide to the mathematical results that are necessary for a rigorous under-
standing of the content of the lectures.
The subject index of the Lecture Notes is, for the moment,
b. Linear filtering.
Definitions
3
4 CHAPTER 1. DEFINITIONS
Example 1.1 At time t a coin is tossed. Outcomes at different times are assumed
to be independent, while the outcomes at time t are equiprobable. The stochastic
variable x t is defined as being 1 if the outcome at time t is “Head”, 0 if “Tail” (H
and T henceforth). Independence implies that
However, this infinite moving average does not make sense unless we (1) define
its sample space and (2) specify the kind of convergence in (1.5) (in probability, in
mean square, almost surely).
Let us discuss this point a little further. Assume that in (1.4) x t is the process
defined in Example 1.1. The variable x t has sample space
˝ t D fH t ; T t g;
i.e. the set whose elements are “H at time t ”, and “T at time t ”, or, if you prefer
that different coins C t be used at different times, “H with coin C t ’ and “T with
coin C t ”, the probability P t being 1=2 for both sample points. The infinite sum in
1.1. STOCHASTIC PROCESSES. KOLMOGOROV’S THEOREM 5
˝t ˝t 1 ˝t k;
1 ˛ 2.k h/
var.Sk t Sht / D var.x t /˛ hC1 ;
1 ˛2
which converges to zero as h ! 1 (note that var.x t / is independent of t ). How-
ever, the mean-square Cauchy condition is necessary for convergence in mean
square, not sufficient in general. Moreover, the limit, if existing, should be a
stochastic variable defined on the infinite product
˝t ˝t 1 ˝t k ;
and the latter should therefore be endowed with a -field and a probability measure.
The situation is fairly uncomfortable, the solution of a simple problem like equation
(1.4) requiring the introduction of a new, infinite-dimensional probability space.
This difficulty can be overcome by an elegant re-definition of the process x
in such a way that the stochastic variables are defined all on the same probability
space. This construction is known as the Kolmogorov’s Existence Theorem.
We do not give a proof of the theorem. The following is an illustration based
on the tossing-coin example. Consider the infinite product space
˝ Z D ˝t 1 ˝ t ˝ tC1 ;
K.; A / D fr 2 ˝ Z; r 2 A g;
i.e. the set of all bilateral sequences of H and T such that at t D the outcome
belongs to A . For example, if D 1 and A1 D fH1 g, then K.1; A1 / is the set
of all sequences that have an H at t D 1. Then consider the -field G generated
by all the sets K. Firstly, observe that all finite-dimensional events have their
“copies” in G. For example, H at times t1 ; : : : ; tm , corresponds to the set of all
6 CHAPTER 1. DEFINITIONS
P .K.t; A t // D P t .A t /:
This law is then extended to G. It is quite obvious that such probability is consistent
with the probability (1.2) on the finite dimensional sets ˝ t1 ˝ tn . It is also
consistent with fairly elementary limit results; for example, the probability of the
set “T for all t ” is zero (for that matter, the probability of any elementary event
in ˝ Z is zero). In conclusion, we may say that .˝ Z; G; P / is an extension of
the set of all probability spaces ˝ t1 ˝ tn , each endowed with its product
probability.
Lastly, we define copies of the stochastic variables x t W ˝ t ! R. Precisely, let
Z t W ˝ Z ! R be defined by Z t .r / D x t .r t /. Given the sequence r , Z t takes on
the value 1 if r t is H, 0 otherwise. Thus Z t does exactly what x t does. Z t is called
the t -th coordinate variable. The distribution of Z t is the same as that of x t , 1 with
probability 1=2 and zero with probability 1=2. Moreover, the probability measure
of .Z t1 ; : : : ; Z tm / is the same as that of .x t1 ; : : : ; x tm /. The big difference, the
result of the construction above, is that the variables Z t are all defined on the same
probability space. Equation (1.4) can be rewritten as
y t D ˛y t 1 C Zt ;
Exercise 1.2 Discuss the distribution of the stochastic variable Y t . Show firstly
that it is independent of t . The Matlab program
1.1. STOCHASTIC PROCESSES. KOLMOGOROV’S THEOREM 7
function y=autor(ALPHA,OBS,REPL)
a=floor(2*rand(OBS,REPL));
% 2*rand generates a uniform
% stochastic variable in (0 2).
% Taking floor we obtain 0 if the uniform
% lies in (0 1), 1 if it lies in (1 2).
% The matrix a has OBS rows and REPL columns
y=(ALPHA.ˆ(0:(OBS-1)))*a;
% ALPHA.ˆ(0:(OBS-1)) has 1 row and OBS columns,
% so that y has 1 row and REPL columns.
computes, for a given ALPHA, a number REPL of replications of Y t by generating
sequences
Z t ; Z t 1 ; : : : ; Z t OBSC1
(the series Y t is truncated at ALPHA OBS 1 Z t OBSC1 ). The output y contains the
replications. Draw the histogram by hist(y,100), and comment on the result.
Of course there exist stochastic processes, whose variables are defined directly
on the same probability space, so that the construction sketched above is not nec-
essary.
Example 1.3 Let be a given real number, and a and b stochastic variables on
the probability space .˝; F ; P /, and let
What is stochastic in x t does not depend on t . In this case values for a and b are
drawn, so to speak, before the beginning of time. When time “starts going” no
further stochastic events occur. As particular cases we have x t D a, the constant
process, for D 0, and x t D . 1/t a, for D .
Example 1.4 Same as in Example 1.3, but now also is a stochastic variable.
Same considerations apply.
K.t1 ; t2 ; : : : ; tn ; A1 ; A2 ; : : : ; An / D fr 2 RZ ; r ts 2 As ; s D 1; 2; : : : ; ng;
8 CHAPTER 1. DEFINITIONS
where t is any integer and As a Borel subset of R; then define the probability
measure on ˝ extending t1 t2 tn on the whole F . Lastly, define y t as the t -th
coordinate variable Z t , which is defined by Z t .r / D r t , for all r 2 RZ. (Here the
infinite dimensional space is constructed on RZ , while in the tossing-coin example
we have used the product of the spaces ˝ t , for t 2 Z; of course the results are
equivalent.)
Thus, with no loss of generality, the definition of a stochastic process can be
restated as follows: