FREQUENCY RESPONSE AND MEAN-SQUARED STABILITY
4.2 Asynchronous Linear Systems with Exponential Inputs
Consider a discrete time-invariant system with ๐ inputs, ๐ outputs, and ๐ state variables, whose state-space description is given as follows:
x๐+1 =A x๐ +B u๐ +w๐, (4.1)
y๐ =C x๐ +D u๐, (4.2)
wherex0โC๐is the initial state vector (initial condition), andw๐ โC๐ is the noise term with the following statistics:
E[w๐] =0, E w๐ wH๐
=๐ฟ(๐ โ๐ ) ๐ช, (4.3) where ๐ฟ(ยท) denotes the discrete Dirac delta function, and๐ช is allowed to be non- diagonal.
The matrices in the state-space model in (4.1) have the following dimensions:
AโC๐ร๐, B โC๐ร๐ , CโC๐ร๐, D โC๐ร๐ , (4.4) whereAis referred to as thestate-transition matrix, and the columns of the matrices BandDwill be denoted as follows:
B=[B1 ยท ยท ยท B๐ ], D=[D1 ยท ยท ยท D๐ ]. (4.5) We further assume that the input signalu๐ consists of ๐ exponential signals in the following form:
u๐ = ๐ผ๐
1 ยท ยท ยท ๐ผ๐
๐
T
, (4.6)
where๐ผ๐โs are assumed to be distinct without loss of generality. Furthermore, we always assume that
|๐ผ๐| โค 1, โ1โค ๐ โค ๐ , (4.7)
so that u๐ stays bounded throughout the iterations. While exponential inputs may seem restrictive, they form the basis for more general practical signals, making this study useful.
In the noise free case, i.e.,๐ช=0, it is well-known from linear system theory that the output vectory๐ โC๐in (4.2) can be written as follows:
y๐ =yss๐ +ytr๐, (4.8)
whereyss๐ denotes the steady-state component, andytr๐ denotes the transient compo- nent that are given as follows:
yss๐ =
๐
ร
๐=1
H๐(๐ผ๐) ๐ผ๐
๐, ytr๐ =C A๐ x
0โxss
0
. (4.9)
whereH๐(๐ง) โC๐denotes the transfer function that relates the๐๐ก โinput to the output, which is given as follows:
H๐(๐ง) =D๐+C ๐งIโA-1
B๐. (4.10)
We also note that the termxss
0 in (4.9) is given as follows:
xss0 =
๐
ร
๐=1
๐ผ๐IโA-1
B๐. (4.11)
It is clear from (4.9) that when the state-transition matrixAis a stable matrix, i.e., the following holds true:
๐(A) <1, (4.12)
then the transient componentytr๐ converges to zero as the iterations progress leaving only the steady-state componentyss๐ in the output signal. In fact, stability ofAis also necessary for the transient part to converge to zero, which is a well-known result from linear system theory [95, 200].
In this chapter we consider the following randomized model:
(x๐+1)๐ =
๏ฃฑ๏ฃด
๏ฃด
๏ฃฒ
๏ฃด๏ฃด
๏ฃณ
A x๐ +B u๐+w๐
๐, ๐ โ T๐+1, (x๐)๐, ๐ โT๐+1,
(4.13)
y๐ =C x๐ +D u๐, (4.14)
whereT๐ denotes the set of indices updated at the ๐๐ก โ iteration. The model (4.13) is very similar to the standard synchronous recursions in (4.2) except the fact that only a random subset of indices (denoted byT๐) are updated in every iteration, and the remaining indices stay the same. The specific stochastic model regarding the selection of the indices will be elaborated next.
4.2.1 Random Selection of the Update Sets
In the asynchronous model considered in (4.13), we assume that the๐๐ก โ index (state variable) is updated independently with probability๐๐in every iteration. That is,
P[๐ โ T๐] = ๐๐ โ๐ , (4.15)
where๐๐will be referred to asthe update probability of the๐๐ก โ index. We will useP to denote the diagonal matrix consisting of the index selection probabilities. More precisely,
P=diag
[๐
1 ๐
2 ยท ยท ยท ๐๐]
. (4.16)
It is assumed thatPsatisfies0โบ P I, where the positive definiteness follows from the fact that no index should be left out permanently during the updates of (4.13).
See Section 4.4 for further details. Additionally, tr(P) =E[|T๐|]denotes the number of indices updated per iteration on average.
4.2.2 Frequency Response in the Mean
Due to the random updates of the state variables it is clear from (4.13) that the state vectorx๐ is a random vector. So, the state vector will not have an exponential behavior exactly even when the input is a simple exponential (๐ = 1 in (4.6)).
Nevertheless, we will show thatx๐ still behaves like a sum of exponential signalsin a statistical sense:
Theorem 4.1. Assume that the randomized asynchronous state recursions in(4.13) are initialized independently and randomly. Then, the expectation of the state vector in(4.13)is as follows:
E[x๐] =xss๐ +xtr๐, (4.17) where
xss๐ =
๐
ร
๐=1
(๐ผ๐IโA)ยฏ โ1Bยฏ๐๐ผ๐
๐, xtr๐ =Aยฏ๐ E[x
0] โxss
0
, (4.18)
and
Aยฏ =I+P (AโI), Bยฏ๐ =P B๐. (4.19)
Proof. See Section 4.8.2.
Here, ยฏAwill be referred to asthe average state-transition matrix, and ยฏBthe average input matrix. We can also representxss๐ in (4.18) as follows:
xss๐ =
๐
ร
๐=1
xss
0,๐ ๐ผ๐
๐ where xss
0,๐ =(๐ผ๐IโA)ยฏ โ1Bยฏ๐, (4.20) which will be useful later in the chapter.
As an immediate corollary to Theorem 4.1, we present the following result regarding the expected behavior of the output:
Corollary 4.1. Assume that the randomized asynchronous state recursions in(4.13) are initialized independently and randomly. Then, the expectation of the output in (4.14)is as follows:
E[y๐] =yss๐ +ytr๐, (4.21) where
yss๐ =
๐
ร
๐=1
Hยฏ๐(๐ผ๐) ๐ผ๐
๐, ytr๐ =CAยฏ๐ E[x0] โxss0
, (4.22)
and
Hยฏ๐(๐ง) =D๐+C ๐งIโAยฏ-1Bยฏ๐. (4.23) This can, therefore, be regarded as the โtransfer function in the meanโ from the๐๐ก โ input node to the output.
Proof. Due to (4.6), (4.14), and Theorem 4.1, the expectation ofy๐ can be written as follows:
E[y๐] =CE[x๐] +
๐
ร
๐=1
D๐๐ผ๐
๐ =yss๐ +ytr๐, (4.24)
whereyss๐ andytr๐ are as in (4.22).
Regarding the form in (4.21) we first note that the termsyss๐ andytr๐ are deterministic quantities, and the expectation is with respect to the random selection of the indices, the input noise, and the random selection of the initial condition.
Corollary 4.1 shows that the random output vector y๐ behaves the same as its deterministic counterpart (4.8) in expectation. That is,E[y๐] can be decomposed into steady-state and transient parts similar to (4.8). Therefore, the quantity ยฏH๐(๐ง) given in (4.23) can be regarded as the โtransfer functionโ from the๐๐ก โ input to the output in theexpectation sense.
It is clear from (4.22) that as long as the average state transition matrix ยฏAis stable, i.e., the following holds true:
๐(A)ยฏ <1, (4.25)
the component ytr๐ converges to zero irrespective of the observation matrixC and the statistics of the initial condition. Thus, the condition (4.25) is both necessary and sufficient forE[y๐]to behave like a sum of exponentials, that is,
๐limโโE
y๐ โyss๐
=0. (4.26)
This shows that when (4.25) is satisfied an exponential input results in an exponential outputin expectationeven with the randomized asynchronous state recursions.
4.2.3 A Running Numerical Example
In order to demonstrate the behavior of the random vector y๐, we consider the following state-space model with ๐ =4 state variables, ๐ =1 input, and ๐=1 output:
A= 1 10
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ
-4 -1 2 -6 4 -6 -5 3
2 -2 7 2
5 9 -3 1
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
, B=
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ 1 4 2 3
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
, C=
๏ฃฎ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฏ
๏ฃฐ 1 1 1 1
๏ฃน
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃบ
๏ฃป
T
, D=0, (4.27)
where we point out that the matrix A is not stable since ๐(A) โ1.0441. As we shall discuss later in Section 4.3, stability of the randomized asynchronous state recursions does not require stability of the state-transition matrix in general.
In the following numerical example we assume thatP=๐I, i.e., all nodes have the update probability๐and assume that๐ช=0. Furthermore, we assume that the input signal has the following complex exponential form:
๐ผ=๐๐2๐/100 =โ u๐ =๐๐2๐ ๐/100. (4.28) In Figure 4.1 we visualize a realization of the output signal y๐ together with the steady-state componentyss๐ as well as the input signal๐ข๐ for three different update probabilities, namely, ๐ โ {0.1, 0.3, 0.6}. The figure shows only the real part of the signals for convenience.
From Figure 4.1 it is clear that the random vectory๐is not a complex exponential in a strict sense, yet it โbehaves likeโ one. We also note thaty๐has the same โfrequencyโ
as the input signal irrespective of the update probabilities.
Figure 4.1 shows also that the response of the random asynchronous system depends on the update probabilities, which is also apparent from the expression in (4.23). In fact, the response of the random asynchronous updates running on a system denoted with (A,B,C,D) can be represented in an average sense as the response of a syn- chronous system(Aยฏ,Bยฏ,C,D). As a result, even when a single state variable updates its value with a different probability, the response of the overall system changes. In short, different update probabilities result in different โfrequency responsesโ while leaving the output frequency unchanged.