FREQUENCY RESPONSE AND MEAN-SQUARED STABILITY
4.8 Appendices
Thus, we can write the following due to the binary nature of the index selections:
E h
DTπ X DTπ
π, π
i
=ππ, π P[π β Tπ, π β Tπ]. (4.83) Regarding the probabilities in (4.83), we have the following for the non-diagonal (πβ π) entries:
P[π β Tπ, π β Tπ] =P[π β Tπ] P[π β Tπ] = ππ ππ, (4.84) which follows from the fact that indices get updated independently from each other.
On the other hand, we have the following for the diagonal (π = π) entries:
P[π β Tπ, π β Tπ] =P[π β Tπ] = ππ, (4.85) which follows simply from the fact thatπ β Tπ if and only if π β Tπ whenπ = π. Thus, we can write the following:
E
DTπ X DTπ
π, π
=


ο£²

ο£³
ππ ππ ππ, π π β π , ππ ππ,π, π = π ,
(4.86)
which is equivalent to the identity (4.81).
4.8.2 Proof of Theorem 4.1
Due to asynchronous updates described in (4.13), state vectorxπ can be written as follows:
xπ =(IβDTπ)xπ-1+DTπ A xπ-1+
π
Γ
π=1
Bπ πΌπ-1
π +wπ-1
= I+DTπ (AβI)
xπ-1+DTπ
π
Γ
π=1
Bπ πΌπ-1
π +wπ-1
!
. (4.87)
Taking expectation of (4.87) and using the facts that updated indices are selected independently, the input noise has zero mean, and the noise is uncorrelated with the index selections, we have the following:
E[xπ] =AΒ― E[xπ-1] +
π
Γ
π=1
BΒ―ππΌπ-1
π =AΒ―πE[x0] +
πβ1
Γ
π=0
AΒ―π
π
Γ
π=1
BΒ―ππΌπ-1-π
π (4.88)
=
π
Γ
π=1
(πΌπIβA)Β― -1BΒ―ππΌπ
π +AΒ―π
E[x0] β
π
Γ
π=1
(πΌπIβA)Β― -1BΒ―π
,
=xssπ +AΒ―π
E[x0] βxss0
=xssπ +xtrπ (4.89)
where Β―Aand Β―Bare as in (4.19).
4.8.3 Proof of Theorem 4.2
Using the definition of the error term in (4.29) and substituting xπ =qπ+xssπ into (4.87), we have the following:
qπ+1+xssπ
+1 = I+DTπ+1 (AβI)
(qπ+xssπ) +DTπ+1 wπ+DTπ+1P-1
π
Γ
π=1
BΒ―ππΌπ
π, (4.90) which can be written as follows by rearranging the terms:
qπ+1= I+DTπ+1 (AβI)
qπ+DTπ+1 wπ + DTπ+1P-1βI
Ξ΄π, (4.91) where thedeterministicvectorΞ΄π is defined as in (4.36).
Using (4.91), we can express the outer product qπ
+1qHπ
+1 recursively in terms of qπqHπ as follows:
qπ
+1qHπ
+1 = I+DTπ+1 (AβI)
qπqHπ (AHβI)DTπ+1 +I +DTπ+1wπ wHπ DTπ+1
+ DTπ+1P-1βI
Ξ΄π Ξ΄Hπ P-1DTπ+1 βI + I+DTπ+1 (AβI)
qπ Ξ΄Hπ P-1DTπ+1 βI + DTπ+1P-1βI
Ξ΄π qHπ I+ (AHβI)DTπ+1
, (4.92)
where the cross terms includingwπ are left-out intentionally because these terms will disappear when we take the expectation since wπ has a zero mean and it is uncorrelated with the index selections.
We now take the expectation of both sides of (4.92) and use the identities given by Lemma 4.6, and the independence assumption regarding the index selections, input noise and the initial condition. Then, we obtain the following:
Qπ+1 =π(Qπ) +PπͺP+πͺ PβP2
+ Ξ΄π Ξ΄Hπ
P-1βI +
(AΒ― βI) xtrπ Ξ΄Hπ
P-1βI +
Ξ΄π (xtrπ)H(AΒ―HβI)
P-1βI
, (4.93) where the function π(Β·)is defined in (4.33). We also note thatE[qπ] =xtrπ as given in (4.35).
AlthoughX+XH β 2<{X} in general, we note that the following equality holds true for anyX βCπΓπ:
X+XH
P-1βI
=2<
X P-1βI
, (4.94)
where<{Β·} denotes the real part of its argument. So, using the identity (4.94) in (4.93) gives the result in (4.35).
4.8.4 Proof of Corollary 4.2 We first define the following:
Zπ =Qπ βQrπβQn, (4.95)
where Qrπ and Qn are given as the solutions of (4.44) and (4.43), respectively.
Substituting (4.95) into the recursion (4.35), we get
Zπ+1+Qrπ+1+Qn =π(Zπ) +π(Qrπ) +π(Qn) (4.96) +PπͺP+πͺ PβP2
+ <n Ξ΄πΞ΄πH
o
Pβ1βI + <n
2(AΒ― βI)xtrπΞ΄Hπ
o
Pβ1βI ,
which can be simplified as follows due to the defining equations in (4.44) and (4.43):
Zπ+1 =π(Zπ) + <n
2(AΒ― βI)xtrπ Ξ΄Hπ
o
Pβ1βI
. (4.97)
Due to the stability assumption (4.41) and Lemma 4.3 we have π(A)Β― <1, so limπββxtrπ =0. As a result,
lim
πββZπ =0, (4.98)
which gives the desired result.
Necessity of the condition (4.41) follows from (4.40). That is, when π(π½) β₯1 there exists a nonzero positive semi-definite matrixXthat cannot be reduced by the functionπ(Β·).
4.8.5 Proof of Lemma 4.1
Assume that the stability condition (4.41) is met, and solution to (4.43) exists. Leteπ
denote theππ‘ β standard basis vector. It is readily verified that the following identity holds true for anyXβCπΓπ and any index 1β€ π β€ π:
eπH π(X) eπ =(1β ππ) eHπ X eπ+ππeHπ A X AHeπ. (4.99) Furthermore, we have the following:
eHπ
PπͺP+πͺ PβP2
eπ = ππ eHπ πͺeπ. (4.100) So, by left and right multiplying (4.43) with eHπ and eπ respectively, we get the following:
eπH Qnβπͺ
eπ ππ = ππ eHπ A QnAHeπ β₯ 0, (4.101) where the inequality follows fromQn 0, and the desired result follows from the fact that ππ > 0 for allπ.
4.8.6 Proof of Lemma 4.2
Enumerate all possible index update sets. Namely, letTπ denote the ππ‘ β update set, and let πΎπ denote the probability of selecting the set Tπ for 1β€ π β€ 2π. So, the system is equivalent to the following state transition matrix:
Aπ =I+DTπ (AβI) (4.102)
which now depends on the index π. This has probabilityπΎπ in model (4.50), where the probability is given as follows (due to the stochastic model in (4.15)):
πΎπ = Γ
πβTπ
ππ
! Γ
πβTπ
(1β ππ)
!
. (4.103)
Then, for a givenXβCπΓπ we have the following:
2π
Γ
π=1
πΎπ Aπ X AHπ =
2π
Γ
π=1
πΎπ
I+DTπ (AβI)
X
I+ (AHβI)DTπ
=E h
I+DTπ (AβI) X
I+ (AHβI)DTπ
i
=X+P(AβI)X+X(AHβI)P+E h
DTπ (AβI)X (AHβI)DTπ
i
(4.104)
=X+P(AβI)X+X(AHβI)P+P(AβI) X (AHβI)P +
(AβI) X (AHβI)
(PβP2)
=
I+P(AβI) X
I+ (AHβI)P +
(AβI) X (AHβI)
(PβP2)
=π(X),
where we use the identity (4.81) in (4.104). Thus, the statement of the corollary follows directly from [42, Corollary 3.26].
4.8.7 Proof of Corollary 4.3
Using the recursive definition of the error correlation matrix in (4.35), we have the following regarding the trace of the error correlation matrix:
tr(Qπ+1)=tr(π(Qπ)) +tr(Pπͺ) +tr
Ξ΄πΞ΄Hπ
Pβ1βI +tr
<
n
2(AΒ― βI)xtrπΞ΄Hπ
o
Pβ1βI
, (4.105)
where we use the fact that tr(XD) =tr(XD)holds true for any diagonal matrixD. Next, we will provide bounds for individual elements on the right-hand-side of (4.105). We first get the following trace equality by summing (4.99) over all indices:
tr π(X)
=tr
X I+AHP AβP
. (4.106)
Now defineΨas follows:
Ξ¨ =π
max I+AHP AβP
. (4.107)
Then we have the following:
Ξ¨I I+AHP AβP=AΒ―HAΒ― + (AHβI) (PβP2) (AβI) AΒ―HAΒ―, (4.108) which implies the following:
kAΒ―k2 β€Ξ¨1/2 and tr π(X)
β€ Ξ¨ tr(X). (4.109) Using the bound in (4.53), we can write the following:
tr
Ξ΄πΞ΄Hπ
Pβ1βI
=Ξ΄Hπ Pβ1βI
Ξ΄π β€ π«2
Pβ1βI
2. (4.110) For the last term on the right-hand-side of (4.105), we can write the following inequalities:
tr
<n
2(AΒ― βI)xtrπΞ΄Hπ
o
Pβ1βI
=<n
2Ξ΄Hπ (P-1βI) (AΒ― βI)xtrπ o
β€
2Ξ΄Hπ (IβP) (AβI)AΒ―π (E[x0] βxss
0)
β€ πΞ¨π/2, (4.111) where (4.111) follows from the definition of xtrπ in (4.18), and the constant π in (4.111) is given as follows:
π =2π«
(IβP) (AβI) 2
E[x0] βxss
0
2, (4.112)
where we use the bounds from (4.53) and (4.109).
Using the bounds in (4.109), (4.110), and (4.111) in the equality (4.105), we get the following:
tr(Qπ+1) β€Ξ¨tr(Qπ) +tr(Pπͺ) +π«2kPβ1βIk2+πΞ¨π/2. (4.113) Using the inequality (4.113) recursively, we get the following:
tr(Qπ) β€ Ξ¨πtr(Q0) +π Ξ¨π/2βΞ¨π
/ Ξ¨1/2βΞ¨ + 1βΞ¨π
1βΞ¨
tr(Pπͺ) +π«2kPβ1βIk2
. (4.114)
In the final step we use the assumptionAHP A⺠Pin (4.51), and conclude from (4.107) thatΨ < 1. So, the upper bound in (4.114) gives the following result:
lim sup
πββ tr(Qπ) β€ tr(Pπͺ) +π«2kPβ1βIk2
1βΞ¨ , (4.115)
which is equivalent to (4.52) since 1βΞ¨ =π
min(PβAHP A).
4.8.8 Proof of Lemma 4.5
Assume thatAβCπΓπ is a triangular matrix with π΄π,π being theππ‘ β diagonal entry.
Then, we first note that both Β―Aandπ½are triangular matrices, which follows simply from the fact that P and J are diagonal matrices, and the Kronecker product of triangular matrices is a triangular matrix. In particular, we note that theππ‘ βdiagonal entry of Β―Ais as follows:
Β― π΄π,π =
1+ ππ(π΄π,πβ1)
. (4.116)
Then, it is straightforward to verify that theππ‘ βdiagonal entry ofπ½is as follows for π =π+ (π β1)πwith 1 β€ π, π β€ π:
ππ , π = π΄Β―β
π , π π΄Β―π,π+πΏπ, π (ππβ π2
π) |π΄π,πβ1|2, (4.117) whereπΏπ, π =1 if and only ifπ= π.
Sinceπ½is triangular, the diagonal entries ofπ½correspond to the eigenvalues ofπ½. Thus,
π(π½)= max
1β€πβ€π2
|ππ , π|= max
1β€π, πβ€π
π΄Β―β
π , π π΄Β―π,π+πΏπ, π (ππβ π2
π) |π΄π,πβ1|2
(4.118)
= max
1β€πβ€π|π΄Β―π,π|2+ (ππβ π2
π) |π΄π,πβ1|2 = max
1β€πβ€π1+ ππ |π΄π,π|2β1
. (4.119) We now prove the equivalence in (4.64). Assume thatAis a stable matrix. SinceA is a triangular matrix, its diagonal entries are the eigenvalues. Thus,
π(A) < 1 =β |π΄π,π| <1 βπ =β π(π½) < 1, (4.120) where the last implication follows from (4.119).
For the converse direction,
π(A) β₯ 1 =β βπ s.t. |π΄π,π| β₯ 1 =β π(π½) β₯1, (4.121) where the last implication follows from (4.119).
C h a p t e r 5