Deterministic dynamical bounds on moments of nonstationary stochastic processes
P. Carrette
Department of Electrical Engineering Linkping University, S-581 83 Linkping, Sweden
WWW:
http://www.control.isy.liu.seEmail:
carrette@isy.liu.seApril 8, 1998
REGLERTEKNIK
AUTOMATIC CONTROL LINKÖPING
Report no.: LiTH-ISY-R-2022
Submitted to Systems and Control Letters
Technical reports from the Automatic Control group in Linkping are available by anony-
mous ftp at the address
ftp.control.isy.liu.se. This report is contained in the com-
pressed postscript le
2022.ps.Z.
Deterministic dynamical bounds on moments of nonstationary stochastic processes
P. Carrette - carrette@isy.liu.se
Department of Electrical Engineering, Linkoping University S-58183 Linkoping, Sweden
Abstract
In this contribution, we deal with the deterministic dominance of the proba- bility moments of stochastic processes. More precisely, given a positive stochastic process, we propose to dominate its probability moment sequence by the trajectory of appropriate lower and upper dominating deterministic processes. The analysis of the behavior of the original stochastic process is then transferred to the stability analysis of the deterministic dominating processes.
The result is applied to a nonstationary auto-regressive process that appears in the system identication literature.
Keywords
: nonstationary stochastic process, probability theory, nonlinear dynamic system, stability analysis, trajectory bounding.
1 Introduction
In general, the evolution of nonstationary stochastic processes is hard to obtain from the statement of the underlying stochastic equation (see 2, 7, 9, 3] and 6, chap. 13]). As a motivation example, let us consider the positive scalar stochastic process
rkdened by
r
k
= (1
;)
rk;1+
'2k k >0 (1) where
r00,
20
1) and
'kdenotes a random variable whose distribution is subject to the following \excitation" condition
' 2
k
r
k;1
(2)
with
20
1]. Note that this process arises in time-varying system identication by use of a constrained forgetting factor recursive least square algorithm (as introduced in 4], see also 8]).
Obviously, the process
rkcannot be considered anymore as an auto-regressive (AR) pro- cess 1, chap. 5]. Roughly speaking, it can be viewed as a nonstationary AR process.
1
As valuable characteristics of stochastic processes lie in their probability moments, it is natural to ask for the trajectory of the probability moments of
rkalong
k. But this is a hard problem to solve. The reason for this is that the excitation condition (2) is such that its probability distribution of
'kdepends on all its past samples (due to
rk;1). Thus, the distribution of the sample
rkdepends in a very intricate way on its past values, so does the evaluation of its moments.
Then, instead of asking for the exact value of these moments, one may be less ambitious and desire to only characterize the evolution of their trajectory (along
k). This is the purpose of the paper.
Here, we propose to lower and upper dominate the trajectory of the probability moments of nonstationary stochastic processes by the solutions of deterministic dynamical equa- tions. Our contribution is as follows.
Given a positive scalar stochastic process
xk, we show that under functional assump- tions on its conditional (onto past samples) probability moments, it is possible to trace the evolution of its probability moments on that of the output of appropriately dened lower and upper bounding deterministic dynamic systems, i.e.
wk E(
xk)
zkwith
w
k
=
g(
wk;1) and
zk=
f(
zk;1). Hence, valuable properties of these moments can be obtained from the stability analysis of these bounding dynamic systems, e.g. equilibrium points, convergence rates.
For an illustration purpose, our results will be applied to the stochastic process
rkin order to derive bounds upon its probability moments, i.e.
E(
rnk), and on those of its inverse process
pk= 1
=rk, i.e.
E(
pnk).
The structure of the paper is as follows. Our main result is stated in Section 2. It deals with the convex and concave functional boundedness of the trajectory of the conditional expectation of a positive scalar stochastic process. Consequences of this functional prop- erty on the evolution of the process expectation are provided. In Section 3, we develop a simple algorithm for practically evaluating convex (lower) and concave (upper) functional bounds on a given function representing conditional expectation dynamics. In Section 4, we apply our results to the stochastic equation (1) under the excitation condition (2).
More precisely, we derive deterministic dynamics dominating that of the moments of the stochastic processes
rkand
pk. Finally, simulations are provided for a particular distribu- tion of the sequence
'k.
2 Deterministic dominance of stochastic processes
In this section, we are interested in evaluating convergence bounds on the expectation of a positive scalar stochastic process
xk. Therefore, we propose to dominate this expectation by the trajectories of appropriate lower and upper bounding deterministic processes, i.e
w
k
E
(
xk)
zkfor
k0.
The convergence analysis of the original expectation is then transferred to that of the deterministic dominating dynamics.
2
The following theorem presents our main result.
Theorem 1 Let
xk(with
k >0) be a positive stochastic process such that
g
(
xk;1)
E(
xkjFk;1)
f(
xk;1) a.e. (3) where
Fk;1=
fxj0
j <kgis the
-algebra generated by the past events of the process and
Fk;1 Fk, and where the functions
g(
x) and
f(
x) are continuous nonnegative convex and concave functions in
R+, respectively. Then,
w
i
E
(
xkjFk;i)
zi(4)
where
wizi >0 are the samples of particular trajectories of the following deterministic scalar processes:
wi=
g(
wi;1) and
zi=
f(
zi;1) with
w0=
z0=
xk;i.
Before going into the proof, let us note that the stochastic inequality (3) holds uniformly in
k. For example, in the case of a stochastic process
xk=
h(
ek) with a random sequence
e
k
possibly dependent on the past
xk(i.e.
xk;1x0), we can write
E
(
xkjFk;1) =
hk(
xk;1)
where
hk(
x) is possibly non-uniform in
k. Then, by dening
h;(
x) := min
khk(
x) and
h
+
(
x) = max
khk(
x) over
x >0, we obtain
g(
x)
h;(
x) as well as
h+(
x)
f(
x) with the desired properties for
g(
x) and
f(
x), if possible. If not, the associated deterministic process bound does not hold.
Proof: By use of Jensen's inequality 5, page 47], the concavity (resp. convexity) property of
f(
x) (resp.
g(
x)) leads to
E
(
f(
x))
f(
E(
x)) (resp.
g(
E(
x))
E(
g(
x)))
for any positive random variable
x. Now, the quantity
E(
xkjFk;i) is recursively dened by
E
(
xkjFk;i) =
E(
E(
xkjF(k;i)+1)
jFk;i) for
i<k. So that
E
(
xkjFk;i) =
E(
E(
E(
xkjFk;1)
jFk;2)
jFk;i)
E
(
E(
f(
xk;1)
jFk;2)
jFk;i)
E
(
f(
f(
xk;2)
jFk;i)
f
(
f(
xk;i)
)
with
icompositions of the concave function
f(
x). For the lower bound (i.e. in term of the convex function
g(
x)), we similarly have
E
(
xkjFk;i)
g(
g(
xk;i)
)
Finally, the denition of the
wiand
ziprocesses leads to
wi=
g(
g(
z0)
) and
zi=
f
(
f(
y0)
) with
icompositions of
g(
x) and
f(
x), respectively. Hence, the proof is completed by taking
w0=
z0=
xk;i.
3
It follows from this result that the convergence properties of the expectation of the stochas- tic process
xkcan be estimated by the analysis of particular deterministic positive pro- cesses. The two following lemmas exhibit properties of their underlying dynamics, i.e.
f
(
x) and
g(
x) respectively.
Lemma 2 Let
zi(with
i>0) be the following positive scalar process
z
i
=
f(
zi;1) with
z0 >0
where
f(
z) is a nonnegative nondecreasing concave function in
R+. If there exists
z >0 such that
f(
z) =
zand
f(
z)
< zfor
z < z, then
zis an attractive equilibrium point for
z <z, i.e.
z
i
z
+
z;1(
zi;1;z) (5)
with
z >1. Globally, we have that lim
izi zfor
i!1.
Proof: First, we derive some properties of the function
f(
z).
It is nonnegative:
f(
z)
0 for
z0. It is nondecreasing and concave: 0
f+0(
z2)
f 0
+
(
z1) for 0
z1 z2, with
f+0(
z), the right derivative of
f(
z). By assumption, 0
f 0
+
(
z)
<1 and
f(
z)
>0, so that
f(
z)
>0 for all
z >0 and either
z < f(
z)
zor
f
(
z) =
zfor 0
< z < z. We also have
z f(
z)
< zfor
z > z, by the nondecreasing property.
Now, we show that
zzM] (with
zM < 1) is a positively invariant compact set and we derive the result in (5). From above, if
zi;1 2 zzM] then
z zi=
f(
zi;1)
zi;1, so that
zi 2zzM]. Moreover, as
f(
z)
<zfor
z <z, we have that
zi <zi;1for
z <zi;1. This means that the equilibrium point
z(i.e.
f(
z) =
z) is attractive from above. And simple calculations give :
f
(
z)
;zz;z
f
(
z) +
f+0(
z)(
z;z)
;zz;z
=
f+0(
z) for
z <z. Hence,
z;1=
f+0(
z)
<1 in (5).
Finally, the positive invariance of 0
z] (i.e. if
zi;1 20
z] then 0
zi=
f(
zi;1)
z:
z
i
2
0
z]) completes the proof of the lemma.
Similarly, we have for the process
wiin Theorem 1.
Lemma 3 Let
wi(with
i>0) be the following positive scalar process
w
i
=
g(
wi;1) with
w0 >0
where
g(
w) is a positive nondecreasing convex function in
R+. If there exists
w >0 such that
g(
w) =
wand
g(
w)
>wfor
w<w, then
wis an attractive equilibrium point for
w<w
, i.e.
w
i
w
;
;1
w
(
w;wi;1) with
w >1. Globally, we have that lim
iwi wfor
i!1.
4
Proof: it is similar to the one of Lemma 2. In this case,
wcan be linked with the left derivative of
g(
w) evaluated at
w=
w:
w;1=
g;0(
w)
<1 with
g0;(
w), the left derivative of
g(
w).
Hence, provided that the stochastic process
xksatises the condition (3) in Theorem 1 and that the corresponding functions
f(
x) and
g(
x) exhibit the characteristics described in Lemma 2 and Lemma 3, respectively, we have that
w
;
;k
w
(
w;x0)
Ix0<w E(
xk)
z+
;kz(
x0;z)
Ix0>z(6) with
E(
xk) =
E(
xkjF0) and
IXdenotes the indicator function on to the condition
X. Of course, if either
g(
x) or
f(
x) cannot be found then the corresponding convergence property does not hold anymore. Actually, either
w= 0 or
z=
1.
3 Algorithm for practical dominating functions
For a constructive use of the result presented in Section 2, we now propose to evaluate, at least numerically, dominating convex and concave functions (i.e.
g(
x) and
f(
x), respec- tively) of a particular function
h(
x) for
x0.
The two corresponding procedures read as:
For the convex functional lower-bound on
h(
x) with a positive and nondecreasing function
g(
x), we rst make
g(0) = min
x0h(
x). Then, we estimate
g(
x) for
x>0 by successive Euler integration steps (with an increment
xfor
x), i.e.
g
(
xi) =
g(
xi;1) +
g;0(
xi)
xfor
xi=
xi;1+
xwhere
g;0(
xi) = max(
g0;(
xi;1)
i) and
iis such that the line
g(
xi;1)+
i(
x;xi;1) is tangent to
h(
x), from below, for
x >xi;1while
g0;(0) = 0.
For the concave functional upper-bound on
h(
x) with a nonnegative and nonde- creasing function
f(
x), we rst state
f(0) =
h(0). Then, we successively perform
f
(
xi) =
f(
xi;1) +
f;0(
xi)
xfor
xi=
xi;1+
xwhere
f;0(
xi) = min(
f;0(
xi;1)
i) and
iis such that the line
f(
xi;1) +
i(
x;xi;1) is tangent to
h(
x), from above, for
x > xi;1while
f;0(0) is as large as possible in order to have
f;0(
x1) =
1.
In Figure 1, we present the result obtained by these two procedures for a given function
h
(
x). We have also displayed the attractive equilibrium point
z(resp.
w) associated to the
z(resp.
w) process in Lemma 2 (resp. 3).
It appears that the growth rate of these two dominating function is asymptotically iden- tical, i.e.
h0(
x) for
x1. Note also that the equilibrium point interval, i.e.
wz], is rather extended over the abscissas where the original function
h(
x) varies.
5
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
x
Figure 1: Practical concave (
;;) and convex (
;) domi- nating functions of the original
h(
x) (
|). Corresponding equilibrium points:
z('
') and
w('
o').
4 Application to the stochastic process example
In this section, we analyze in some details the stochastic process
rkintroduced in (1).
More precisely, we provide lower and upper dynamical bounds on the trajectory of its probability moments, i.e.
E(
rnk), and those of its inverse, i.e.
E(
pnk).
In view of Theorem 1, such dynamical bounding is achieved in deriving lower convex and upper concave functional bounds on the conditional (onto the past events) expectation on the corresponding power of these processes, i.e.
E(
xkjFk;1) with
xk=
rknfor the
n-th probability moments of
rk. Let us then evaluate bounding functionals for the conditional expectation of the
n-th power of the two processes successively.
The
n-th power of the stochastic equation governing the process
rkis written as
r n
k
= (1
;)
rk;1+
'2k]
n=
rk;1n(1
;)
n+
Pn(
kjk;1)]
with
kjk;1=
'2k=rk;1and
Pn(
x) = ((1
;) +
x)
n;(1
;)
n >0 for
x>0. Note that
P
n
(
x) is monotonically increasing in
xwith
Pn(0) = 0.
The evaluation of the conditional expectation of
rknwith respect to the past events gives
E
(
rknjFk;1) =
rk;1n(1
;)
n+
E(
Pn(
kjk;1)
jFk;1)]
Then, we derive the following functional bounds that are uniform in
kr n
k;1
(1
;)
n+
Q;n(
rnk;1)]
E(
rknjFk;1)
rk;1n(1
;)
n+
Q+n(
rnk;1)] (7)
6
for appropriate functions
Q;n(
x) (resp.
Q+n(
x)) dened similarly to
g(
x) (resp.
f(
x)) in Theorem 1 for
Pn(
kjk;1).
Moreover, when appropriate (lower convex and upper concave, respectively) dominating functions are estimated for
E(
rknjFk;1) (as presented in Section 3), we can dene ~
Q;n(
x) and ~
Q+n(
x) as the functions that make the bounding expressions in (7) identical to the corresponding dominating function estimates. The attractive equilibrium points of these dominating processes, i.e.
wrand
zrare found by solving
~
Q
;
n
(
wr) = ~
Q+n(
zr) = 1
;(1
;)
nwhile the lower bounds for the convergence rate to these solutions (see expression (6)) are given by
;1
wr
= 1 +
wr( ~
Q;n(
wr))
0;and
zr;1= 1 +
zr( ~
Q+n(
zr))
0+where (
h(
x))
0;(resp. (
h(
x))
0+) stands for the rst left (resp. right) derivative of
h(
x) evaluated at
x=
x.
The
n-th power of the process
pk, i.e.
pnk, is treated similarly. We rst write
p n
k
=
pnk;1(1
;)
n+
Pn(
kjk;1)]
;1where
Pn(
x) is the same polynomial as before and
kjk;1can be written as
kjk;1=
'2k pk;1. Then, the uniform (in
k) functional bounds on
E(
pnkjFk;1) takes the following form
p n
(1
;)
n+
k;1Tn;(
pnk;1)
E(
pnkjFk;1)
(1
;)
np+
nk;1Tn+(
pnk;1) (8) for appropriate functions
Tn;(
x) and
Tn+(
x). In fact, by Jensen's inequality, it can be shown that
Tn+(
x)
Q+n(1
=x)
Tn;(
x) with
Q+n(
x) from above.
Finally, when convex and concave dominating functions are estimated for the lower and upper bounds of
E(
pnkjFk;1), we similarly obtain the functions ~
Tn;(
x) and ~
Tn+(
x). The attractive equilibrium points of these dominating processes, i.e.
wpand
zpare found by solving
~
T
;
n
(
wp) = ~
Tn+(
zp) = 1
;(1
;)
nwhile the lower bounds for the convergence rate to these solutions are given by
;1
wp
= 1
;wp( ~
Tn;(
wp))
0;and
zp;1= 1
;zp( ~
Tn+(
zp))
0+In the next section, we give simulations of the dynamical (lower convex and upper concave) bounds we have derived for the probability moments of the processes
rkand
pkin the case of a particular distribution of their \independent" random variable
'k. The role of the corresponding equilibrium points, i.e.
wand
z, will also be demonstrated.
7
5 Simulation results
Here, we illustrate the theoretical results presented in the preceding sections. More pre- cisely, we evaluate asymptotic bounds on the trajectories of the second probability mo- ments of the stochastic processes
rkand
pk= 1
=rk.
As seen above, these bounds are made of the equilibrium points of the dominating deter- ministic dynamics associated to the corresponding conditional moments trajectories, i.e.
E
(
rk2jFk;1) and
E(
p2kjFk;1). By use of Jensen's inequality, we further have that
1
=zp E(
p2k)
E(
rk2)
zrfor large
k(9) where
zr(resp.
zp) is related to the estimated function ~
Q+2(
r2) (resp. ~
T2+(
p2)) introduced in Section 4.
First, let us introduce the density function of the \input" random variable. Each sample
'
k
is taken independently of the others. Its energy density function is based upon a reference density function, i.e.
d(
'2), whose distribution function is denoted
D(
'2). We then consider a modication of this reference density function in order to generate energy sequences that satisfy the excitation condition
'2k rk;1along
k0 for a chosen value of
20
1]. By dening the conditional (on
2) reference density function as
d
(
'2j2) = 1
1
;D(
2)
d(
'2)
I'22the sample
'2kcan be seen as a random variable following a density function identical to
d
(
:jrk;1).
For the simulations, we consider an energy sequence
'2that has a small probability, i.e.
say, of being large with a density function centered at
02and a complementary probability, i.e. 1
;, of being small. In Figure 2, we present the density function of
j'jcorresponding to a particular example of such a reference density function
d(
'2) for
= 0
:1 and
02= 1.
We also show the constitutive density functions.
In Figure 3, we have presented two realizations of the process
rkfor this reference density function
d(
'2) with
identical either to zero or 0
:3 for
= 0
:05. Obviously, the two real- izations behave very dierently: for zero
, it exhibits small
rk(leading to large
pk) values due to irrelevant samples
'2kwhile, for
= 0
:3, its focuses on the signicant samples out of that density function. Note that memory of the process (or of the initial condition
r0) is similar to the inverse of the forgetting factor
, i.e.
1
=.
Now, let us turn to the trajectory of the second probability moment of the processes
r
k
and
pk. Although these two processes are not auto-regressive (AR) per se, they can roughly speaking be seen (from their realizations in Figure 3) as almost-stationary AR processes.
The independence of the samples
'kover
kand the fact that their density function is at most
rk;1-dependent imply that the conditional (onto the past) probability moments of these processes are uniform in
k, i.e.
E(
xkjFk;1) =
h(
xk;1). Therefore, the results
8
0 0.5 1 1.5 2 2.5 3 0
0.5 1 1.5 2 2.5 3 3.5
Figure 2: Example of the density function of
j'j j'jwith
= 0
:1 and
02= 1.
0 100 200 300 400 500 600 700
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
=0:05
=0:0
=0:3
Figure 3: Two di erent realizations of the
knonstationary process
rk. derived in Section 2 are easily applied.
Let us then consider the deterministic dynamics that upper dominate the second prob- ability moment of these processes, i.e.
E(
rk2) and
E(
p2k). These dynamics are related to the bounding functions ~
Q+2(
r2) and ~
T2+(
p2) that are represented in Figure 4 (normalized to 1
;(1
;)
2]). As mentionned in Section 4, the equilibrium points of the dominating trajectories are found by making these bounding functions identical to 1
;(1
;)
2],
0.7 0.8 0.9 1 1.1 1.2 1.3 1.4 1.5 1.6 1.7
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6
~
Q +
2 (r
2
)
1;(1;) 2
~
T +
2 (p
2
)
1;(1;) 2
ror1=p
=0:4
=0:4
Figure 4: Bounding functions for the
rk2and 1
=p2kprocesses. Equilibrium points:
1
=(
zp)
1=2and (
zr)
1=2('
'). Estimated values of
krkk2('
') and 1
=kpkk2('
o').
0 0.1 0.2 0.3 0.4 0.5
0 0.5 1 1.5
0.01 0.05 0.10 0.18
Figure 5: Asymptotic bounds on
krkk2and
kp
k k
;1
2
as a function of the dynamic relative threshold
.
i.e. leading to
zrand
zp, respectively. Furthermore, from the expression (9), the interval
9
made of these two equilibria, i.e. 1
=zpzr], will asymptotically (in
k) contain the expec- tation of the square processes. This can be seen in the gure where we have displayed the estimated values of these process 2-norms, i.e.
krkk2and 1
=kpkk2with
kxk2=
E(
x2)]
1=2. These estimates has been obtained by averaging a particular realization of the associated processes (for
=
= 0
:4). In fact, the equilibria interval appears to provide bounds that tightly surround the estimated moments.
Finally, in Figure 5, we present the estimated (lower and upper) bounds on the asymptotic value of these process 2-norms, i.e. 1
=(
zp)
1=2and (
zr)
1=2, as functions of the value
for several values of
.
It can be emphasized that these bounding intervals are not linear in the
value. Indeed, for small
(i.e.
<0
:02), the
rkprocess exhibits small sample values that characterize the global distribution of
'2. For increasing
's (i.e. 0
:02
< <0
:2), the process
rktends to exhibit the distribution of more energetic regressors samples
'k. For larger
values (i.e.
>