2250 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO.
9, SEPTEMBER 2012
Moderate Deviations of a Random Riccati Equation
Soummya Kar, Member, IEEE, and José M. F. Moura, Fellow, IEEE
Abstract—The paper characterizes the invariant filtering mea-
sures resulting from Kalman filtering with intermittent observa-
tions in which the observation arrival is modeled as a Bernoulli
process with packet arrival probability . Our prior work showed
that, for 0, the sequence of random conditional error co-
variance matrices converges weakly to a unique invariant distri-
bution . This paper shows that, as approaches one, the family
0 satisfies a moderate deviations principle with good rate
function ( ): 1) as 1, the family converges weakly to
the Dirac measure concentrated on the fixed point of the asso-
ciated discrete time Riccati operator; 2) the probability of a rare
event (an event bounded away from ) under decays to zero
as a power law of (1 ) as 1; and, 3) the best power law
decay exponent is obtained by solving a deterministic variational
problem involving the rate function ( ). For specific scenarios, the
paper develops computationally tractable methods that lead to ef-
ficient estimates of rare event probabilities under . stability results in robust filtering (e.g., [15]–[17]).
Index Terms—Intermittent observations, Kalman filtering, mod-
erate deviation principle (MDP), random Riccati equation (RRE).
I. INTRODUCTION
Background and Motivation:
OTIVATED by problems in networked sensing and
M control, Kalman filtering with non-classical informa-
tion patterns has received renewed attention. Networked control
systems suffer from unpredictable communication, leading to
inherent complications in estimation and control. Reference
[1] introduced a non-trivial networked control system, Kalman
filtering with intermittent observations, in which the sensor
observation packets transmitted through an imperfect commu-
nication medium are received at the estimator as a Bernoulli
process with packet arrival probability . In [2], we
established a dichotomy for the resulting filtering error process
by showing, in particular, that stochastic boundedness of the
sequence of conditional error covariance matrices [generated by
an associated random Riccati equation (RRE)] is necessary and
sufficient for its ergodicity. In other words, [2] shows that there
exists a critical value, , such that, if the observation packet
arrival probability , the sequence of random conditional
error covariance matrices converges weakly (in distribution) to
a unique invariant distribution ; otherwise the error process
diverges pathwise. We note here that stochastic boundedness is
Manuscript received June 01, 2010; revised April 24, 2011; accepted May
29, 2012. Date of publication June 05, 2012; date of current version August
24, 2012. This work was supported by the National Science Foundation (NSF)
grants CCF-1018509 and CCF-1011903 and by AFOSR grant FA9550101291.
Recommended by Associate Editor T. Zhou.
The authors are with the Department of Electrical and Computer Engi-
neering, Carnegie Mellon University, Pittsburgh, PA 15213 USA (e-mail:
soummyak@andrew.cmu.edu, moura@ece.cmu.edu).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TAC.2012.2203453
0018-9286/$31.00 © 2012 IEEE
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
a much weaker condition than moment stability, and, as shown
in [2], convergence to a unique attracting invariant distribution
is possible under a packet arrival probability for which sta-
bility in mean does not hold. Moreover, it was shown that for
observable and controllable systems , and, hence,
exists and is unique for every for such systems.1 Further,
[2] provides a sample-path analysis of the RRE, in contrast
to moment stability analysis, as is done conventionally in the
literature (e.g., [3]–[11], see also [2] for a detailed review of
the literature.) Finally, apart from its relevance to the area of
networked control systems, our study is intimately connected
to the design and analysis of robust and risk-sensitive linear
filtering approaches (e.g., [12], [13], and [14], in the context
of missing measurement data,) and more generally to pathwise
This paper characterizes the invariant measures in the non-
classical information case, the counterpart of the fixed point
of the Riccati equation in the classical case. For observable and
controllable systems, we establish as :
1) Moderate deviation principle (MDP). For , let
be the unique attracting measure of the ergodic conditional
error covariance process (Theorem 9 in [2]). We show that,
as , satisfies an MDP with good rate func-
tion and converges weakly to the Dirac measure
where is the unique fixed point of the associated de-
,
terministic Riccati equation.
2) Probability of rare events. The probability of a rare event
(bounded away from ) decays as a power law of .
3) Best decay: Variational problem. We show that the best
MDP decay rate of the probability of a rare event satisfies
4) Estimating probability of rare events. This reduces
to solving deterministic variational problems that char-
acterize their decay rate and give insight into how they
occur. As shown by examples, these variational problems
may be solved much more efficiently than numerically
estimating the invariant distributions when , since
the rare events become increasingly difficult to observe.
Our techniques go beyond Kalman filtering with intermittent
observations. There is a key difference in our MDP arguments
and conventional methods on the moderate (or large) deviations
for stationary measures of Markov processes. Typically these
methods assume that the underlying Markov process is posi-
tive recurrent and moderate deviations of stationary measures
then follow from that of finite dimensional distributions (e.g.,
[18], [19]). However, the Markov processes governing the RRE
are not, in general, positive recurrent, as is the case for a large
class of iterated function systems ([20]). Our analysis studies
1That =0 was obtained for invertible observation matrices in [2]. A
proof for general observable systems is provided in Appendix B of the current
paper.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2251

the probability measures induced on the space of random func- be an extended valued lower semicontinuous function and
tion compositions (strings) and develops their topological prop- be non-decreasing on (0, 1) with
erties. Several intermediate results are of independent interest
and are applicable to the analysis of more complex networked
control systems (e.g., [21]) and hybrid or switched systems.
We summarize the paper. The next paragraph presents nota- Definition 1.1: The family is said to satisfy a moderate
tion and preliminaries. Section II sets up the problem and re- deviations principle (MDP) with MDP rate function at scale
views briefly prior work. Key approximation results are in Sec- as if the following hold for each open set
tion III, whereas the main results are stated and discussed in and closed set , respectively:
Section IV. The MDP for finite dimensional distributions of
(2)
the RRE sequence is in Section V, whereas Sections VI and
VII carry out the steps to obtain the main results on MDP for
stationary distributions. Section VIII presents efficient ways to (3)
solve the variational problems involving the rate function. Nu-
merical studies justifying theoretical results for a scalar system Lower semicontinuity implies that the level sets of ,
are presented in Section IX. Section X concludes the paper. , are closed . If the level sets are also
Notation and Preliminaries: Denote by: , the reals; , compact , is said to be a good rate function. It can be
the non-negative reals; , the -dimensional Euclidean shown that, for every measurable set , the MDP, as stated in
space; , the identity matrix of appropriate dimension; , (2) and (3), is equivalent to
the integers; , the non-negative integers; , the natural
numbers; and , a generic space. For a subset ,
is the indicator function, which is 1 when
the argument is in and zero otherwise; and is the identity (4)
function on . A metric space with metric is denoted
by the pair . The corresponding Borel algebra is de- The above MDP notion is similar to that of the large deviations
noted by . For a set , we denote by and its principle (LDP). In fact, if the scale function is polynomial
interior and closure respectively. For , the open ball of in , is said to satisfy an LDP (e.g., [18], [23]).
radius centered at is denoted by . For any set Definition 1.2 (Rare Event): A set is said to be a
, the open and closed -neighborhoods of are given by rare event with respect to the family , if .
and , respectively. The To see the implications of an MDP on the probabilities of a
space of symmetric matrices is denoted by , whereas rare event, assume that is an -continuity set, i.e.,
and denote the subsets of positive semidefinite and . Then the limit in (4) exists and we have
positive definite matrices, respectively. For , we
write to denote ; to (5)
denote and ; to denote
. where denotes a functional of that decays to zero as .
Limit Notations: Let be measurable. The We will use the following as a short form of (5):
notation implies that for (6)
every sequence in with and
, we have . If is a rare event, by the lower bound in (5), , en-
Probability Measures on Metric Spaces: Let be a suring that the probabilities decay to zero. In turn, by the upper
complete separable metric space and the set of probability bound in (5), as long as an MDP holds in the sense of Defini-
measures on . For tion 1.1, the probability of a rare event not only goes to zero, but
does so exponentially at scale , with being the
best exponent (or rate) of decay.

is the support of , a closed set. The sequence in II. PROBLEM FORMULATION

converges weakly (in distribution and denoted by )
The state and the observation are
to if as for every bounded
continuous function on . The weak topology induced on
by weak convergence may be metrized by the Prohorov
metric [22]; is complete and separable. The Pro- where and are zero-mean (z.m.) Gaussian
horov distance between two probability measures and in with covariances and . The sequences and are
may be obtained as uncorrelated and mutually independent. The initial state is
z.m. Gaussian with covariance . Unless otherwise stated, we
(1) use the following standing assumption throughout the paper:
Assumption (E): The pair is observable and are
Moderate Deviations: Let be a family of probability positive definite. The assumption implies the control-
measures on the complete separable metric space in- lability of the pair . Some intermediate results hold
dexed by the real-valued parameter . Let under less stringent assumptions as will be noted.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2252 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

The m.m.s.e. predictor of given is given where and correspond to the Lyapunov and Riccati up-
by the Kalman filter with the conditional prediction error co- dates in (7) and (8). The length of a string is denoted by
variances, , given by the algebraic Riccati equation. . The set of all possible strings is denoted by .
Under controllability of and observability of , Remark 3.1: A string can be of length 0, then it is repre-
the deterministic converges to the unique fixed point sented as a 1-tuple, consisting only of the initial condition.
of the algebraic Riccati equation for all . We introduce notation. Let
With intermittent observations, packets are dropped and the and for , , such that, for all
estimator receives observations at random times. In [1], the , , . Let be a string of length
channel randomness is modeled by a sequence of i.i.d.
Bernoulli random variables with mean (the arrival proba-
bility), where corresponds to arrival of the packet at
time to the estimator and to a packet drop. Let the pair
. Under the TCP packet acknowledgement For brevity, we will write as
protocol in [1] (the estimator knows whether packets arrived or
not), the m.m.s.e. predictor of may be implemented through
a modified Kalman filter [1] in which , updated by RRE
e.g, the string is .
Definition 3.2 (Numerical Value of a String): To every string
we associate its numerical value, , the numerical eval-
uation of the function composition2 on the initial state , i.e.,
for
Let the Lyapunov and Riccati operators and be
(7)
(8) Thus, the numerical value can be viewed as a function
from the space of strings to . Denote the set of numerical
These operators are defined on . We then have for all values attainable by
Note the difference between a string and its numerical value.
Two strings are equal iff they are of the same order of function
composition applied to the same initial state.
Unlike the classical case, for each , is now random
Definition 3.3 (Concatenated Strings): Let ,
in . We let be its distribution, and and the
. If , choose .
probability and expectation, respectively, with the superscripts Then, a concatenated string of block length with initial 3
, emphasizing their dependence on the packet arrival proba-
and its numerical value
bility and the initial condition. The next result, [2], characterizes
the RRE convergence.
Theorem 2.1 (Theorems 9,10 in [2]): Under (E.1), for each
, there exists a unique invariant distribution , such that
the sequence (or the sequence of measures) con-
verges weakly to from any initial . Let

with its topological closure. Then, the support of is For fixed , , , denote the set of con-
catenated strings by and the set of numerical values
by . For , the set
consists of all strings with numerical value , i.e.
III. SOME APPROXIMATION RESULTS

A. Lyapunov and Riccati Operators: Preliminary Results

We develop elsewhere the algebra of such strings. Next, we
The RRE sequence is an iterated function system (e.g., [20])
present properties of strings (proof in Appendix A):
comprising random compositions of Lyapunov and Riccati op-
Proposition 3.1:
erators. To understand such random compositions, both their nu-
• For , and , we have, .
merical value and their pattern are relevant. We start with three
definitions: In particular, if for some , and
Definition 3.1 (String): A string with initial state 2For function composition, we adopt the notation with strings, e.g., f  f 

and length is a -tuple of the form f  f  f  f (P ) is written f  f  f (P ). By convention, for a function

f on , its zero-th iterate f is the identity function on .
3We again adopt the convention that the first block is simply P if t = 0.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2253

, the string Lemma 3.3: Let controllable and observ-

belongs to , we have able. Then, for every and initial covariance ,
, i.e., the sequence is stochastically bounded

• Fix , . Then, for

The proof is in Appendix B. The result establishes the impor-
tance of stochastic boundedness as a metric for system stability
and design as compared to various notions of moment stability
• Define to be the function (see [2] for detailed discussion). For example, as shown in [1],
the critical probability for mean stability may be quite large, de-
pending on the instability of . However, Lemma 3.3 and the
Let and be a string dichotomy relating stochastic boundedness to ergodicity estab-
and count the number of ’s in lished in [2] shows that, even in the sub-mean stability regime,
the system is stochastically bounded (for ) and converges
if (9) to a unique invariant distribution. As such, our analysis offers
otherwise insight into system design in the sub-mean stability regime.
The following result on limits of real number sequences (see,
Then, , depending on only such that e.g., Lemma 1.2.15 of [23]) will be useful.
Proposition 3.2: For and , let
be functions with

B. Riccati Equation and MDP: Approximation Results

The first result is on uniform convergence of the classical Ric- (we adopt the convention ) and is non-negative
cati operator; it is used to obtain tightness estimates required for with as a possible value. Then
establishing the MDP. Proof in the Appendix A.
Lemma 3.1: For every , there exists , such that,
for every , with

can be chosen independently of the initial state .

The following result is a corollary on the Lipschitz continuity IV. MAIN RESULTS AND DISCUSSIONS
of finite compositions of the Riccati operator.
We state and discuss the main results of the paper, the proofs
Lemma 3.2: For fixed and , define
being provided in Section VII.
the function by
The following result is a first step to understanding the be-
havior of the family of invariant distributions.
Theorem 4.1: The family of invariant distributions
Then is Lipschitz continuous with some constant . converges weakly to the dirac probability measure , i.e.,
Also, for every , there exists such that the function as . In particular, for each , we
is Lipschitz continuous with constant . have the following convergence rate asymptotics:
Proof: From (49), the function is Lipschitz contin-
uous with constant , where are positive (10)
constants defined in Lemma 3.1. Also, the affine function
is Lipschitz continuous with constant , where is the We discuss the consequences of Theorem 4.1. The first part
largest singular value of the matrix . Then, the function states that the family converges weakly to the Dirac mea-
defined above is Lipschitz continuous, being a finite composi- sure concentrated at , , as . This is quite intuitive, as
tion of Lipschitz continuous functions. with , the filtering problem reduces to the classical Kalman
For the second part, choose , such that, filtering setup with deterministic packet arrival (no dropouts), in
, where are defined in Lemma 3.1 (see (49)). Then, which case the deterministic sequence of conditional error co-
from (49), with the above choice of , the function is variances converges to the fixed point . Thus, as , we
Lipschitz continuous with constant . expect the RRE sequence to behave more and more similar to the
The following result concerns stochastic boundedness of the deterministic case leading to the convergence of the mea-
random sequence generated by the RRE, and, in par- sures to as . However, from a technical viewpoint,
ticular, completes the proof of Proposition 8 in [2] for general such a convergence is particularly interesting and non-trivial in
observable and controllable systems. the sense that the infinite supports exhibited by the measures
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2254 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

(see Theorem 2.1) for shrinks to the point at mating numerically the invariant distributions. A naive numer-
. An immediate consequence of Theorem 4.1 is ical approach of simulating the distributions as be-
comes infeasible since the rare events become increasingly dif-
ficult to simulate as . Even techniques such as importance
sampling require characterization of the distributions in ques-
Thus, with respect to , every event , such that , is a tion, the issue we addressed in this paper.
rare event (see Definition 1.2). This is intuitively clear, because
as , the measures become concentrated on arbitrarily V. MODERATE DEVIATIONS (MD) FOR FINITE-
small neighborhoods of , making such an event very diffi- DIMENSIONAL DISTRIBUTIONS
cult to observe. We establish MD for finite dimensional distributions of the
The second assertion estimates the convergence rate of the process as . We start with notation.
measures to and relates to the loss in performance Fix and and recall the sets
when . Essentially, Theorem 4.1 states that, for every and for .
, the probability of staying away from the -neighborhood By Proposition 3.1, the random object takes
of decreases at least as when , i.e.,4 only a finite number of values and, for all

Since this holds , the probability of any rare event van- We generalize (see (9)) to count the number of
ishes at least as . A complete characterization of the in the string by
family requires the exact decay rate of rare events as ,
and this is achieved next by establishing an MDP for the family if
as . otherwise.
Theorem 4.2: The family of invariant distributions sat-
isfies an MDP at scale as with a good rate Also, for , define
function , i.e., for every open and closed sets (11)

where we adopt the convention that the minimum of an

empty set is . The following result shows that the function
is a good rate function on .
Proposition 5.1: The function in (11)
is a good rate function on .
The rate function is given by Proof: The function is clearly non-negative. Its level
sets are compact, because its effective domain is finite

Theorem 4.2 characterizes the invariant measures as .

First, it establishes the important qualitative behavior of
that rare events decay exactly as power-laws of as .
The exact exponent of such a power law decay, which depends We then have the following result giving the MD for
on the particular rare event, is obtained as the solution of an the family (as ) of finite dimensional distributions
associated variational problem involving the minimization of . The proof is provided in Appendix C.
the rate function . The result shows how the system de- Theorem 5.1: Fix and and let
signer can trade-off estimation accuracy with communication be the family of finite dimensional distributions
overhead. For example, given a tolerance , the value of indexed by , starting from the same initial state . Then, for
such that the probability of lying outside the -neighborhood every , we have
of is less than some , is given by

By computing , the designer obtains an es- where

timate of the communication rate required to maintain a prob- if
ability of error less than . Section VIII shows techniques to otherwise.
solve these variational problems efficiently. Finally, we note
that reducing the estimation of probabilities to solving varia-
tional problems efficiently is much more practical than esti- VI. MDP FOR INVARIANT DISTRIBUTIONS
4For functions h(1), g (1), the notation h( ) = O (g ( )) implies the existence This Section establishes results of independent interest, key to
of a constant c > 0, such that h( )  cg ( ) for all 2 (0; 1). proving the MDP results in Section IV. It characterizes the sta-
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2255

tionary measures arising from general stabilizable iterated func- Proof: Let be an arbitrary initial state and
tion systems. The MDP for finite dimensional distributions pre- be the sequence generated by the RRE for
sented in the last Section suggests a rate function for the MDP
. From the weak convergence in Theorem 2.1
for the family of invariant distributions, as a suitable “limit”
as of the finite time marginal rate functions . How-
ever, even if such a limit exists, it is not a priori obvious whether
such an is lower semicontinuous and hence qualifies as a
rate function. In the following, we introduce such an and
its lower semicontinuous regularization as a candidate rate
function (Section VI-A).
(12)

A. A Rate Function Now consider the measurable set . Let

, where is the effective domain of , i.e., the set on
Let be given by which is finite. Then, there exists , sufficiently small,
such that, the closed ball . From (12) it follows:

where, as usual, we adopt the convention that the infimum of an (13)

empty set is . Note that, for We now set to estimate the R.H.S. of (13). To this end, re-
call the nonempty set of all strings of finite (but ar-
bitrary) length with initial state and numerical value .
For some and , let the string
. Define the function
and may be viewed as a natural generalization of the marginal by . Note that is
rate functions for all . It is not clear if is lower continuous, and, hence, there exists , such that
semicontinuous; hence, we consider its lower semicontinuous
regularization as a candidate rate function, where (14)

Also, by Lemma 3.1, there exists , such that

(15)
The following proposition states some readily verifiable prop-
erties of ; the proof is in Appendix D.
It then follows from (14) and (15) that for any such that
Proposition 6.1:
and any string of the form:
• is a good rate function on .
• For every , we have

where , we have

• For non empty, we have (16)

. If is open, the reverse inequality holds,
and . Indeed
• Let be a non-empty compact set. We have

where the last step follows from the fact that, by (15)

B. The MDP Lower Bound

Now, for as defined, , consider the set of strings
The following result establishes the MDP lower bound for the
sequence of invariant distributions as .
Lemma 6.1: Let . We have the lower bound

(i.e., the indices can be arbitrary). It then fol-

lows from (16) that , . From the
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2256 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

iterative construction of the sequence , it follows Lemma 6.2: Let be a closed set. Define the set of
that, for : strings and quantity as
(20)
(21)

where the infimum of an empty set is . Then, if ,

there exists sufficiently large, such that, for all
with , we have .
Remark 6.1: It follows from (20) and (21) that , :

(17)

From (13) and (17), we have

Also, if , i.e., the set is non-empty, the in-
(18) fimum in (21) is attained, i.e., there exists , such
that, . The above follows from the fact that the
and hence function takes only a countable number of values. Simi-
larly, in the case , by defining , we
have .
We now prove the MDP upper bound for the family as
Since the above holds for all , we have over compact sets.
Lemma 6.3: Let be a compact set. Then the
following upper bound holds:

Finally, from the fact, that, for , Proof: For , is the open -neighborhood of
and its closure. Since is open, we have by the weak
convergence of the sequence to
(19)

The Lemma follows from Proposition 6.1 (third assertion), since which in turn implies
is an open set, by showing the R.H.S. of (19) satisfies
(22)

We now estimate the probabilities on the L.H.S. of (22). Since

We extract (18) as an independent result for later use. is closed, Lemma 6.2 applies. Recall and in (20)
Corollary 6.1: Let be an open set and and (21). Also, for every and closed set , define
be a string, such that, . Then, there exists a positive
integer (depending on and ), such that

To establish the Lemma we consider first (i.e,

is non-empty) and then not. Fix . From Proposition 3.1
(second assertion)
C. MDP Upper Bound
We establish the MDP upper bound for the family of invariant
measures as . The proof is accomplished in three steps. (23)
First, Lemma 6.3 establishes the upper bound for compact sets. Since and , we have . Since
Then, Lemma 6.4 proves a tightness result on the family of in- is closed and Lemma 6.2 imply there exists , such
variant distributions. Finally, Lemma 6.5 establishes the MDP that, for every string with , we have
upper bound for closed sets. . In other words, for all
Definition 6.1 (Truncated String): Let the string be given
by , where ,
and ., for , the truncated string of length is
given by . Now consider and define the set of strings by
The next result on the topological properties of strings will be
used. It is proved in Appendix D.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2257

In other words, the set consists of all strings of length PROOFS OF THEOREMS
, such that there are at least occurrences of in the Theorem 4.2: The MDP lower and upper bounds obtained in
truncated string . For , the following iholds: Lemma 6.1 and Lemma 6.5 respectively show that the family
(24) satisfies an MDP at scale as with a
good rate function . To complete the proof of Theorem 4.2
By the Markovian dynamics of the RRE, clearly, for it suffices to show that
(29)

i.e., the function is lower semicontinuous. Clearly, if

(25) , the claim in (29) follows from (63). We thus
consider the case . Since
We then have from (23) and (24)

and the integer-valued quantity is non-de-

creasing w.r.t. , there exists , such that

Taking the log-limits on both sides and noting that is inde-

pendent of , we have The infimum above is achieved for every and we conclude
that there exists a sequence , such that

Again, taking the limit on both sides as we conclude Recall the set of strings
(26)

We then have
From Lemma 6.1 (third assertion), we have for all

Finally, taking the limit and from Lemma 6.1 (fourth assertion) Since is closed, by Lemma 6.2, there exists ,
such that, for with , we have
(27)
(30)

The Lemma then follows from (26) and (27). By the existence of , there exists a sequence of
The following tightness result extends the upper bound from strings in , such that
compact sets to arbitrary closed sets (proof in Appendix D).
(31)
Lemma 6.4: The family of invariant distributions sat-
isfies the following tightness property: For every , there Note that, without loss of generality, we can assume that
exists a compact set , such that for all . Indeed, if , we can
(28) modify by appending the requisite number of at
the right end, still satisfying (31). On the other hand, if
, we note that must be of the form
We now complete the proof of the MDP upper bound for
arbitrary closed sets by the upper bound for compact sets and
the tightness estimate obtained in Lemma 6.4. This is achieved
in Lemma 6.5 and has parallels in the theory of large devia- where the truncated string
tions. The tightness analogue of Lemma 6.4 is called exponen- (32)
tial tightness in the context of LDP. It can be shown that the
LDP upper bound for closed sets follows from that of compact satisfies
sets and exponential tightness, see, for example, [18]. Since the
(33)
proof approaches are closely related (with obvious modifica-
tions due to a different MDP scale function), we omit the proof The second inequality in (33) follows from (30), whereas (32)
of Lemma 6.5 in the following. follows from the fact that implying
Lemma 6.5: Let be a closed set. Then the fol- . Thus the right end of does not contain any , ex-
lowing upper bound holds: plaining the form in (32). The key conclusion of the above dis-
cussion is that, if , we may instead consider the
truncated string , which also satisfies (31).
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2258 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

We thus assume that the sequence with the properties claim in (10) then follows from the MDP upper bound for closed
in (31) further satisfy for all . The number of sets (Lemma 6.5.)
distinct strings in the sequence is at most (in fact,
lesser than that, because of the constraint ) VII. COMPUTATIONS WITH THE RATE FUNCTION
and hence at least one pattern is repeated infinitely often in the A complete characterization of probabilities under the in-
sequence , i.e., there exists a string , such that variant distributions are obtained in Theorem 4.2, which shows
that the probability of rare events decay as powers of
as . The best exponent of this power law decay is charac-
and a subsequence of , such that, terized by the MDP rate function . As Theorem 4.2 shows,
for all . The corresponding subsequence of nu- the best decay exponent of such a rare event can be computed as
merical values then satisfy the infimum of the over that set. This reduces the problem
of estimating probabilities under the invariant distributions to
solutions of a related variational problem, namely, minimizing
over Borel sets in . From a numerical analysis point of
and, hence view, one may simulate the function and use a look-up table
to numerically solve the associated variational problems. In this
section we show that for a class of events of interest, the vari-
ational problem of computing the best decay exponent can be
Thus the string and we obtain simplified to a great extent. In particular, we are interested in es-
(34) timating probabilities of the form for every .
Theorem 4.1 shows that such probabilities decay to zero at least
as , i.e., for every , we have
The other inequality is obvious and hence we
conclude from (34) that , which completes the (36)
proof of Theorem 4.2.
Theorem 4.1: Define the class of sets However, this upper bound becomes loose as increases and the
best exponent of decay, i.e., the best power of in (36)
can be computed precisely by solving the variational problem
in Theorem 4.2. The main result of this section shows that for a
Then the following equivalence may be readily verified: particular class of systems, this computation can be extremely
simplified and for general systems we present bounds on the
(35)
decay exponent which are simple to obtain in contrast to solving
where the Prohorov metric is defined in (1). Consider the full-fledged variational problem.
, small enough. Then , such that for every To this end, define the function by
, we have . (Note that the constant can be
chosen independently of , but depends on .) The string
belongs to and hence by Corollary 6.1, there exists Also, define by
an integer , such that

We note that is a non-decreasing right continuous function

Thus for all and we have for all

Then for we have for all Also recall

It then follows from (35) that for :

Definition 8.1 (Class Systems): Let be a
system satisfying the assumption (E.1). Following Bucy ([24]),
Hence we define the set

Then the system is called a class system if

Since is arbitrary, by passing to the limit as , we (37)
conclude that as .
For the second part we note that for , the closed set We then have the following MDP asymptotics for class
does not contain . Hence . The systems (see Appendix E for a proof):
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2259

Lemma 8.1: Let be a class system. Then we . The following proposition outlines a simple algorithm for
have for all solving the variational problems of interest leading to the decay
exponents of rare events for general systems:
Proposition 8.2: Let and define
(41)

Define the set

Lemma 8.1 shows that for class systems the variational

problem of computing the best decay exponent can be simpli- Then, .
fied to a great exponent. In particular, rather than generating the Proof: The proof is straight-forward and follows from the
set of all possible strings (which grows exponentially with fact, that, and so it suffices to look at strings
the length of the strings) one can systematically look into strings of length at most.
of the form , and obtain the decay exponent. Proposition 8.2 concludes that the computation of the decay
The next natural question is whether there exists a suitable char- exponent may be reduced to obtaining in (41) and subse-
acterization of class systems. Determining whether a given quently looking for the minimizer in the set as constructed
system is class is numerically simple, as one only needs to above.
check the condition in (37). From a theoretical point of view, it
would be relevant to offer a characterization of class systems
VIII. A SCALAR EXAMPLE
through properties of the system matrices. A detailed study to
that end will be a digression from the main theme of the paper For a scalar system: , . Solving the
and we intend to pursue it elsewhere. In the following we show algebraic Riccati equation, . By Proposition 8.1,
that scalar systems are included in class , thereby confirming the system is of class . By Lemma 8.1, for
that the latter is not empty.
Proposition 8.1: Let be a scalar system satis-
fying (E.1) (i.e., , and .) Then,
belongs to class . (42)
Proof: Define the function by
, where is the scalar Riccati operator. Under the Let . We estimate the decay rate of the rare event
assumptions, has only one fixed point in , , the steady as . Since
state solution. Thus, in the domain of interest
(43)
(38)
By (43), is an -continuity set [see text following (6)],
We note that and by [24], there exists hence the limit in (42) exists and we have
sufficiently large, such that
(44)
(39)
Our theory then predicts that for close to 1
We now claim that
(45)
(40)
We estimate the empirical decay rate of the event by simu-
Indeed, if this was not true, then by (39) this would imply the lating samples from for with a step size
existence of an interval in not containing , such that the of.05. We iterated the RRE 100 times to make sure the random
sign of changes over this interval. This in turn would imply covariance converged in distribution to . Thus, a total of
from the continuity of , the existence of another solution computations were involved. The resulting empirical
to the equation on other than . Clearly, this cumulative distributions functions (cdf) are plotted in Fig. 1 (on
contradicts with the hypothesis (see (38)) and hence, the claim the top). As , the empirical invariant measures converge to
in (40) holds. , thus verifying Theorem 4.1.
Since , then from (40) that, for To obtain the empirical decay rate of , for each , we
estimated from the empirical cdfs.
This is plotted in Fig. 1 (bottom) as a function of (solid line).
The result agrees with the theoretical findings: (1) qualitatively,
thus showing that scalar systems belong to class . the rare event decays as a power law of ; (2) the best
As shown in Lemma 8.1, for class systems the computation decay exponent approaches 4 (theoretically established in (44))
of the decay exponents of rare events can be greatly simplified. as gets closer to 1; (3) even for , the empirical decay
For general systems such a simplification may not be possible, rate is close to 4, justifying (45).
nonetheless, the solution of the variational problem in Theorem Even estimation of the probabilities of rare events for this
4.2 may be made more efficient than an exhaustive search over modest scalar example required order computations,
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2260 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

Now choose , such that, . Then,

, and it follows from the order-preserving property of the
operators and , that
(46)

where . Note that the claim is trivial

for , so assume . Also, for , we have

We now consider the nontrivial case . Let

be the indices corresponding to in from
left to right. Consider the last segment of , i.e., the string
. We then have
(47)

This follows from the fact, that, , which im-

plies the sequence is decreasing and we have
. The last inequality in (47) uses the
following ordering between the operators , and :
(48)

Noting that the iterates of are non-decreasing by construc-

tion, we conclude that , and
. Now considering the segment of from to

Fig. 1. Upper: Weak convergence of (empirical) measures  to  as " 1.

Lower: Decay exponent of probability of the rare event B (P ).

whereas we find the best decay exponent by solving a much

simpler variational problem.
where we use the fact that and (48). Since
IX. CONCLUSION , we can repeat the above argument inductively
The paper studied the RRE arising from the problem of and obtain . The claim then follows
Kalman filtering with intermittent observations. We showed from (46).
that for every the conditional mean-squared error process Proof of Lemma 3.1: The proof of Lemma 3.7 may be ex-
is ergodic and the resulting family of invariant distributions tracted from established results on Riccati operators, see, for ex-
(as they converge weakly to as ) satisfies an ample, [25]. However, to present the result in the form most suit-
MDP with good rate function . The rate function is able for our presentation, we provide a brief self-contained proof
completely characterized and the asymptotic decay rate of rare here. We use the following result on uniform convergence over
events are formulated as solutions of deterministic variational compact sets of the Riccati iterates to . Let , .
problems. The intermediate results obtained are of independent Then it can be shown (see Theorem 7.5 in [26])5 that there exist
interest and our methods are fairly general to be applicable to constants , , such that
the analysis of more complex networked control systems (see, (49)
for example, [21]) and hybrid or switched systems.
Taking , and noting that we
APPENDIX A have from the above
PROOFS OF PROPOSITION 3.1, LEMMA 3.1
Proof of Proposition 3.1: The first assertion follows from
the fact, that, , whereas the second is an imme- Thus for every compact subset , there exists , de-
diate consequence of the iterated construction of the sequence pending on , such that
. We prove the third assertion in the following.
Recall the set . Under the (50)
assumptions of observability and controllability, it can be shown 5The result in [26] applies to time-variant system matrices under the assump-
(see [24]), there exists , such that tions of uniform complete observability and uniform complete controllability,
which reduces to observability and controllability in the current time-invariant
system setting.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2261

To prove the Lemma, we need to transform the above uniform We then have
convergence over compact sets to uniform convergence over the
entire space . To this end, we note that for a observable and
controllable system, the following uniform boundedness of Ric-
cati iterates holds from arbitrary initial state (see, for
example, Lemma 7.1 in [26]):

(51)
where we have used the fact, that, for all
. It then follows from the above and (55), that
where is a sufficiently large number. Now consider
the compact set . From (51) (57)
(52)
We now estimate the probability . To this end,
Now from (50) choose , such that first consider the case . On the event , it is not hard
to see that the following events are equal:
(53)

Then defining we have from (52) and (53),

for all , and the result follows.

It then follows by elementary manipulations and the indepen-

dence of packet arrivals:
APPENDIX B
STOCHASTIC BOUNDEDNESS OF ERROR COVARIANCES
Proof of Lemma 3.3: For stable , the proposition is trivial.
We consider unstable . We bound -step Riccati iterates for
controllable and observable systems and use arguments in [2]
with more generality. It can be shown (see Lemma 7.1 in [26])
that for controllable and observable systems

(54)

where is sufficiently large. In other words, the above

states that an application of the Riccati operator more than
times in succession leads to a covariance bounded above by a
specific constant, irrespective of the initial state. Now, for
and sufficiently large, define On the event , using similar arguments, we can show

(55)
We thus have the upper bound
where and . Since is unstable
, it follows that as . To estimate
the probability for , define by

Rearranging and summing the geometric series above we have

where the maximum of an empty set is taken to be zero. Thus,
if ,6 it denotes the time closest to , such that, there were
successive packet arrivals in the time interval .
Then, using the above arguments, we have

(56) (58)

Indeed, if , then and (56) holds by the definition

of . On the contrary, if , we have by (54) From (57) and (58) we have for all and sufficiently large

6Note that, if not zero, t  N.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2262 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

Since and as , it follows: It then follows from Proposition 3.2 that:

Thus is s.b. (for all initial conditions ) for every

and hence, the Lemma follows.
which establishes the Theorem.

APPENDIX C APPENDIX D
PROOF OF THEOREM 5.1 PROOFS OF RESULTS IN SECTION VI
Proof: For , we have Proof of Proposition 6.1: That is lower semicontin-
uous follows from its definition (e.g., [27].) For , con-
sider the level set . By lower
semicontinuity, is closed. We show that is bounded and
(59) hence compact. Note that for all
(62)
which follows from the Markov property. From the above
for a constant , which can be chosen independent of
. Indeed, implies that is non-empty and

Also, if , from (59), simple manipulations, Since is integer valued, this infimum is attained and there
and the independence of the packet dropouts exists with . Then, by Proposition 3.1,
there exists (depending on only)

(60)
This verifies the claim in (62).
Let . For , the sequence is
non-decreasing w.r.t. ; hence implies
By manipulating each term on the R.H.S. above, we have

Since is integer valued, the infimum is attained and there

exists , such that, . From (62) then

(61) Since we have

It then follows from Proposition 3.2 and (60) and (61) that:
The goodness of follows from the level sets of being
closed and bounded, , as is bounded:

(the case is absorbed above by using the For the second assertion, we note that, for arbitrary
convention, that, the minimum of an empty set is .) Now con-
sider . If , then the
claim is obvious. Hence, assume (this
intersection is necessarily finite) and note that
The assertion follows by passing to the limit as .
For the third assertion, in general, for arbitrary , we
have ; passing to the limit

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION 2263

This immediately gives for any set and, hence, as . It then follows from the
lower semicontinuity of and (65) that:
(63)

For the reverse inequality, when is open, consider .

Then, there exists (depending on ) such that, for every
, the open ball . Then The above contradicts that is the minimizer of over .
We conclude that (64), and, hence, the result, holds.
Proof of Lemma 6.2: The case is trivial as the
assertion follows by choosing an arbitrary positive .
Taking the limit on both sides we obtain Consider . By induction it suffices to prove:
, for ,
(66)

Thus, , . Taking the in- Start with . Assume on the contrary, that there is no such
fimum over on the left hand side gives the desired . Since is non-empty, by Proposition 3.1 (first
. The result follows. assertion), there exists , such that
We prove the fourth assertion. For limits involving continuous (67)
arguments, it suffices to show that with
Thus the non-existence of , implies, that, for every ,
there exists a string , with , such that,
.7 Such a string is then necessarily of the form
we have

(64)
where . Thus denoting

Consider now such a sequence . Assume on the contrary,

that (64) is not satisfied. Since , we clearly have
we note that
(68)

Now consider the sequence of such strings as

Thus the hypothesis that (64) is not satisfied implies
. Let be arbitrary. By Proposition 3.1, the uniform
convergence of Riccati iterates implies there exists ,
such that, for every , for all
(the constant chosen independently of .) Then, defining
The sets are compact. Since a lower semicontinuous , it follows from (68) that for :
function attains its minimum over a compact set, ,
. Similarly,
. Note is a non de- Since above is arbitrary, it follows that the sequence
creasing sequence (hence, the limit exists) and of numerical values converges to as .
However, by construction, the sequence belongs
(65) to the set and we conclude that is a limit point of the set
. Since is closed, we have . Then
Also, given , it follows from the continuity of the metric,
that the function , , is
continuous and hence attains its minimum over a compact set.
Hence, in particular, the string . The fact, that
Thus , there exists , such that
then contradicts the hypothesis .
Thus by contradiction we establish that, if , there
exists satisfying the properties in (66) for , completing
the proof of the Lemma if . We consider now the
where the last inequality follows since . Note the general case .
sequence belongs to the compact set ; hence, there exists Assume now . Assume on the contrary that (66)
a subsequence in , which converges to some does not hold for all . By the previous argu-
. Now consider the sequence . We then have ments, (66) clearly holds for . Then, let
be the largest integer such that (66) holds for all .
The hypothesis implies there exists no
7In the above, R corresponds to the truncated string, see Definition 6.1.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 2264 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 57, NO. 9, SEPTEMBER 2012

satisfying (66) for . Since (66) holds for , there Thus in converges to as .
exists , such that for all with , Hence, as is closed and is a limit point
we have . The non-existence of and (67) im- of . This implies . Since , this contra-
plies that, for every , there exists a string , dicts the hypothesis and the claim in (66) holds for
with , such that all , establishing the Lemma.
Proof of Lemma 6.4: Let be arbitrary and be
(69) such that . From Proposition 3.1 (third assertion), there
exists , such that, for all
We now study the structure of the strings for sufficiently
. Define such that and
large . To this end, define . Then by the ex-
istence of and (69) it follows that for . consider the compact set . Also
Hence for , is necessarily of the form define the closed set, , by . As
per Lemma 6.2, define the set by

where , such that, We then have the following inclusion:

and . Now consider
and define the set . Also, define
. Consider the function
and, hence
by

Since the cardinality of is finite and is countably infinite, Since is closed, by Lemma 6.2
there exists , such that, the set is
countably infinite. This implies, that we can extract a subse-
quence from , such that To estimate , follow a similar set of arguments as in
Lemma 6.3. By weak convergence

In other words, if for fixed (70)

, , the string is of the form
For define the sets . For
we have [see also (25)]

where are arbitrary. Denoting

(71)

we have A familiar set of arguments (Lemma 6.3) yields from (70)–(71)

Since as , by Proposition 3.1 in similar way from which we obtain

Noting that is continuous (com- Thus for every there exists a compact set such that
position of continuous functions), conclude (28) is satisfied and the Lemma follows.

APPENDIX E
PROOF OF LEMMA 8.1
Proof: By Theorem 4.2 it suffices to show

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
 KAR AND MOURA: MODERATE DEVIATIONS OF A RANDOM RICCATI EQUATION
We prove the left assertion; the second is similar. Since
, the sequence is
non-decreasing. Then, for class systems, the following holds:
for any positive integer . Then an inductive argument that re-
peatedly uses the definition of class systems and the mono-
tonicity of the Riccati operator yields the following:
(72)
By definition it follows that and hence
(73)
Now assume on the contrary strict inequality in (73). Then there
exists with , such that,
. Then, from (72)
which implies
This contradicts the definition of that it is the smallest
non-negative integer , such that, . We
thus conclude that the left assertion and the claim holds.
ACKNOWLEDGMENT
The authors wish to thank K. Ramanan for general discus-
sions on large deviations of Markov processes, the associate ed-
itor and the anonymous reviewers for insightful comments and
suggestions that significantly improved the clarity of the paper.
REFERENCES
[1] B. Sinopoli, L. Schenato, M. Franceschetti, K. Poolla, M. Jordan, and S.
Sastry, “Kalman filtering with intermittent observations,” IEEE Trans.
Autom. Control, vol. 49, no. 9, pp. 1453–1464, Sep. 2004.
[2] S. Kar, B. Sinopoli, and J. M. F. Moura, “Kalman filtering with inter-
mittent observations: Weak convergence to a stationary distribution,”
IEEE Trans. Autom. Control, vol. 57, no. 2, pp. 405–420, Feb. 2012.
[3] H. Chen, P. Kumar, and J. van Schuppen, “On Kalman filtering for
conditionally Gaussian systems with random matrices,” Syst. Control
Lett., vol. 13, pp. 397–404, 1989.
[4] Y. Wang and L. Guo, “On stability of random Riccati equations,” Sci.
China Series E: Technol. Sci., vol. 42, no. 2, pp. 136–148, Apr. 1999.
[5] X. Liu and A. Goldsmith, “Kalman filtering with partial observation
losses,” in Proc. IEEE Conf. Decision Control, Paradise, The Bahamas,
Dec. 2004, vol. 4, pp. 4180–4186.
[6] V. Gupta, T. Chung, B. Hassibi, and R. M. Murray, “On a stochastic
sensor selection algorithm with applications in sensor scheduling and
sensor coverage,” Automatica, vol. 42, no. 2, pp. 251–260, 2006.
[7] Y. Xu and J. Hespanha, “Estimation under controlled and uncontrolled
communications in networked control systems,” in Proc. IEEE Conf.
Decision Control, Sevilla, Spain, Dec. 2005, pp. 842–847.
[8] M. Huang and S. Dey, “Stability of Kalman filtering with Markovian
packet losses,” Automatica, vol. 43, no. 4, pp. 598–607, 2007.
[9] K. Plarre and F. Bullo, “On Kalman filtering for detectable systems
with intermittent observations,” IEEE Trans. Autom. Control, vol. 54,
no. 3, pp. 386–390, Feb. 2009.
[10] C. Robinson and P. R. Kumar, “Sending the most recent observation is
not optimal in networked control: Linear temporal coding and towards
the design of a control specific transport protocol,” in Proc. IEEE CDC,
Dec. 2007, pp. 334–339.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KHARAGPUR. Downloaded on June 05,2023 at 21:30:18 UTC from IEEE Xplore. Restrictions apply.
2265
[11] L. Xie and L. Xie, “Stability of a random Riccati equation with
Markovian binary switching,” in Proc. IEEE CDC, Dec. 2007, pp.
1565–1570.
[12] A. H. Sayed, “A framework for state-space estimation with uncertain
models,” IEEE Trans. Autom. Control, vol. 46, no. 7, pp. 998–1013,
Jul. 2001.
[13] R. K. Boel, M. R. James, and I. R. Petersen, “Robustness and risk-sensi-
tive filtering,” IEEE Trans. Autom. Control, vol. 47, no. 3, pp. 451–461,
Mar. 2002.
[14] H. Y. Liang and T. Zhou, “Robust state estimation for uncertain
discrete-time stochastic systems with missing measurements,” Auto-
matica, vol. 47, no. 7, pp. 1520–1524, 2011.
[15] J. M. C. Clark and D. Crisan, “On a robust version of the integral repre-
sentation formula of nonlinear filtering,” Prob. Theory Related Fields,
vol. 133, no. 1, pp. 43–56, Sep. 1996.
[16] N. J. Newton, “Observations sampling and quantisation for continuous-
time estimators,” Stoch. Processes Appl., vol. 87, pp. 311–337, 2000.
[17] S. Kar, J. M. F. Moura, and K. Ramanan, “Sample path large devia-
tions for the randomly sampled continuous-discrete Kalman filter,” in
Proc. 49th IEEE Conf. Decision Control, Atlanta, GA, Dec. 2010, pp.
7375–7382.
[18] J. Deuschel and D. Stroock, Large Deviations. New York: Academic,
1984.
[19] M. I. Freidlin and A. D. Wentzell, Random Perturbations of Dynamical
Systems, 2nd ed. New York: Springer-Verlag, 1998.
[20] P. Diaconis and D. Freedman, “Iterated random functions,” SIAM Rev.,
vol. 41, no. 1, pp. 45–76, 1999.
[21] S. Kar, B. Sinopoli, and J. M. F. Moura, “A random dynamical systems
approach to filtering in large-scale networks,” in Proc. Amer. Control
Conf., Jun. 2010, pp. 1027–1034.
[22] J. Jacod and A. Shiryaev, Limit Theorems for Stochastic Processes.
Berlin, Germany: Springer-Verlag, 1987.
[23] A. Dembo and O. Zeitouni, Large Deviations Techniques and Applica-
tions. Boston, MA: Jones and Barlett, 1993.
[24] R. S. Bucy, “Apriori bounds for the Riccati equation,” in Proc. 6th
Berkeley Symp. Math. Stat. Prob., 1972, vol. 3, pp. 645–656.
[25] B. D. O. Anderson and J. Moore, Optimal Filtering. Englewood
Cliffs, NJ: Prentice-Hall, 1979.
[26] A. Jazwinski, Stochastic Processes and Filtering Theory. New York:
Dover, 1970.
[27] J. Feng and T. G. Kurtz, Large Deviations for Stochastic Processes.
Providence, RI: American Mathematical Society, 2006.
Soummya Kar (S’05–M’10) received the B.Tech.
degree in electronics and electrical communication
engineering from the Indian Institute of Technology,
Kharagpur, in May 2005 and the Ph.D. degree in
electrical and computer engineering from Carnegie
Mellon University, Pittsburgh, PA, in 2010.
From June 2010 to May 2011 he was with the
Electrical Engineering Department, Princeton Uni-
versity, Princeton, NJ, as a Postdoctoral Research
Associate. He is currently an Assistant Research
Professor of electrical and computer engineering
at Carnegie Mellon University, Pittsburgh, PA. His research interests include
performance analysis and inference in large-scale networked systems, adaptive
stochastic systems, stochastic approximation, and large deviations.
José M. F. Moura (S’71–M’75–SM’90–F’94)
received the E.E. degree from the Instituto Superior
Técnico (IST), Portugal, and the Ph.D. degree from
the Massachusetts Institute of Technology (MIT),
Cambridge, MA.
He is a University Professor at Carnegie Mellon
University (CMU), Pittsburgh, PA, was on the faculty
at IST and has been Visiting Professor at MIT, Cam-
bridge, MA. In 2010, he was elected University Pro-
fessor at CMU. His interests include statistical and
algebraic signal and image processing.
Dr. Moura is Division IX Director of The IEEE, was President of the
IEEE Signal Processing Society (SPS), and Editor-in-Chief of the IEEE
TRANSACTIONS ON SIGNAL PROCESSING. He is a Fellow of the AAAS and
a Member of the Academy of Sciences of Portugal. He received the SPS
Technical Achievement Award.