iterferômetros baseados na geometria fractal de hilbert - UFPE
Minimax games for stochastic systems subject to relative entropy uncertainty: applications to SDEs...
Transcript of Minimax games for stochastic systems subject to relative entropy uncertainty: applications to SDEs...
Minimax Games for Stochastic Systems Subject
to Relative Entropy Uncertainty:
Applications to SDE’ s on Hilbert Spaces
N.U. AHMED1 and C.D. CHARALAMBOUS2
Abstract-In this paper we consider minimax games for stochastic uncertain systems
with the pay-off being a nonlinear functional of the uncertain measure where the
uncertainty is measured in terms of relative entropy between the uncertain and the
nominal measure. The maximizing player is the uncertain measure, while the mini-
mizer is the control which induces a nominal measure. Existence and uniqueness of
minimax solutions are derived on suitable spaces of measures. Several examples are
presented illustrating the results. Subsequently, the results are also applied to con-
trolled Stochastic Differential Equations (SDE’s) on Hilbert Spaces. Based on infinite
dimensional extension of Girsanov’s measure transformation, martingale solutions are
used in establishing existence and uniqueness of minimax strategies. Moreover, some
basic properties of the relative entropy of measures on infinite dimensional spaces are
presented and then applied to uncertain systems described by a Stochastic Differen-
tial inclusion on Hilbert space. The worst case measure representing the maximizing
player (adversary) is found and it is shown that it is described by an evolution equa-
tion on the space of measures on the original state space.
1N.U. Ahmed is with the School of Information Technology and Engineering and De-partment of Mathematics University of Ottawa, Ottawa, Ontario, CANADA.E-mail: [email protected]
2C.D. Charalambous is with the Department of Electrical and Computer Engineering,University of Cyprus, Nicosia, CYPRUS. Also with the School of Information Technologyand Engineering, University of Ottawa, Ottawa, Ontario, CANADA.E-mail: [email protected] work is supported by the European Commission under the project ICCCSYSTEMSand by the University of Cyprus under an internal medium size research project.
1
1 Motivation
Minimax games are important when dealing with an uncertain system which may
consist of a whole family of (mathematically) well defined systems. In the signal
processing and control community, the designer is the estimator or the control player
which wishes to minimize the pay-off challenged (presented) by the nature. This is
modeled as a functional of measures induced by the uncertain system. Specifically,
robustness problems in control theory can be formulated using this framework. In the
past several papers have appeared dealing with minimax games [11, 12, 14]. In [15],
the stochastic analog is addressed, in which the maximizing players are described by
measures. Moreover, in [16, 17], discrete-time minimax problems are also introduced
in which the maximizing measure is parameterized by random processes describing the
uncertainty of the system. In [18, 19], an alternative approach is put forward in which
the uncertain measure for general nonlinear partially observable uncertain systems is
computed. These models as well as the results presented are directly or indirectly
related to H∞ problems [6], dissipation inequalities [7], and risk-sensitive problems
[8-19]. An explanation of the various connections can be found in [20]. Uncertain
systems and Variational inequalities can be described by differential inclusions in
a unified framework covering both deterministic and stochastic systems. Optimal
control and nonlinear filtering of such systems are considered in several papers [21-
25].
This paper is fundamentally different from the above described references in both the
theory and applications considered. The theoretical development is concerned with
minimax games in which the pay-off is a nonlinear functional of the uncertain measure.
The maximizing player belongs to a set described by relative entropy between the
uncertain and the nominal measures, while the control affects the nominal measure.
Existence and uniqueness of the minimax strategies are presented on the space of
2
measures defined on Polish spaces. The results are then applied to controlled Stochas-
tic Differential equations (SDE’s) and inclusions defined on Hilbert Spaces. Infinite
dimensional extension of Girsanov’s measure transformation is used in establishing
existence and uniqueness of minimax strategies. The worst case measure is found and
it is shown that it is governed by an evolution equation on the space of countably ad-
ditive probability measures on Hilbert Space. To the best of the authors knowledge,
the treatment found here appears to be the first one dealing with minimax games
subject to relative entropy constraints for SDE’s and inclusions on Hilbert spaces.
2 Introduction
Let X be a Polish space, in particular, a complete separable metric space and let
BX the sigma algebra of Borel sets generated by the metric topology. Let M1(X)
denote the space of countably additive regular probability measures defined on BX .We assume throughout the paper that M1(X) is furnished with the weak topology.
It is well known that the weak topology is metrizable with respect to the Prohorov
metric which makes it a complete metric space. We shall denote this metric by dP .
Thus, a net µαw−→ µo if and only if dP (µα, µo) −→ o. Since the weak topology and
the metric topology are equivalent we note the following facts. A set Γ ⊂ M1(X) is
compact with respect to the weak topology, if and only if, it is compact with respect
to the metric topology dP . A functional g : M1(X) −→ R is weakly continuous, if
and only if, it is continuous with respect to the Prohorov metric dP .
Let ν, µ ∈ M1(X) and suppose ν is absolutely continuous with respect to the measure
µ, which is denoted by ν ≺ µ. Then, the Radon-Nikodym derivative of ν with respect
to µ exists and this is given by an h ∈ L1(X,µ) such that
dν = hdµ.
We note that, given that ν ≺ µ, the RND is unique, that is, there is only one
3
h ∈ L1(X,µ) that satisfies the above identity. Indeed, if both h and f satisfy the
above identity, then ∫Xϕ(x)h(x) − f(x) dµ = 0
for all ϕ ∈ L∞(X,µ). This is impossible unless h = f µ a.e. Identifying h with f that
differ from h only on a set of µ measure zero, we have the uniqueness. If further,
h log h ∈ L1(X,µ), then the entropy of ν relative to the measure µ, known as relative
entropy, is defined and it is given by
H(ν|µ) ≡∫Xh log hdµ =
∫X(dν/dµ) log(dν/dµ)dµ
=∫Xlog(dν/dµ)dν. (2.1)
The basic problem we wish to study can be stated as follows.
Basic Problem: Let η : M1(X) −→ [−∞,∞), be an extended real valued continu-
ous function. Let M0 ⊂ M1(X) denote the set of measures induced by a controlled
stochastic system under the assumptions that the system is perfectly known (system
operators and other parameters and coefficients etc., are all perfectly known). In
other words, M0 denotes the set of attainable measures on X associated with an
unambiguously defined (perfectly known) controlled stochastic system. However, in
the real world situation, the system parameters are not entirely known to the con-
troller. This introduces some degree of uncertainty in the system dynamics around
the nominal controlled system and hence the set of measures induced by the true
system may differ from those represented by the set M0. For each fixed r > 0 and
µ ∈ M0 ⊂ M1(X), define the set
Ar(µ) ≡ ν ∈ M1(X) : H(ν|µ)) ≤ r. (2.2)
This set represents the set of true measures induced by the actual physical system
given that the nominal controlled system induces the measure µ. This is the set of
uncertainty about the system measured in terms of relative entropy, relative to the
4
measure µ. Thus, the set of measures induced by the uncertain controlled system is
given by
Mr ≡⋃
(Ar(µ)), µ ∈ M0
.
Clearly M0 ⊂ Mr for all r ≥ 0 and limr↓0 Mr = M0. Since the relative entropy is
nonnegative, the parameter r determines the degree of uncertainty. Thus the larger
the r is, the greater is the uncertainty.
The problem is to find a µo ∈ M0 that minimizes the functional
Fr(µ) ≡ supη(ν); ν ∈ Ar(µ) (2.3)
over the set M0. In other words, we wish to find a solution to the min-max problem
(Pis) : infµ∈M0
supη(ν) : ν ∈ Ar(µ) (2.4)
In [15] a specific linear functional η(ν) is considered and results related to Large De-
viations Theory are derived. Here, η(ν) is a general nonlinear functional on M1(X).
3 Existence of the Min-Max Solution
In the following section we investigate the question of existence of solution to the
basic problem (Pis).
Lemma 3.1 Let η : M1(X) −→ [−∞,+∞) be an upper semi continuous function
and strictly concave. Then, for every µ ∈ M0 ⊂ M1(X), there exists a unique
νo ∈ Ar(µ) at which η attains its supremum.
Proof. Without loss of generality, we may assume that η(ν) ≡ −∞. If so there is
nothing to prove. It is well known that, for each µ ∈ M1(X), Ar(µ) is a closed convex
5
subset of M1(X) and compact with respect to the metric topology dP [1,Lemma 1.4.3,
pp. 37]. Let
supη(ν), ν ∈ Ar(µ) ≡ mo.
We show that there exists a νo ∈ Ar(µ) at which the sup is attained. Let να ∈ Ar(µ)
be a maximizing net so that
limαη(να) = supη(ν) : ν ∈ Ar(µ) = mo. (3.5)
Since Ar(µ) is compact, there exists a subnet of the net να, relabeled as the original
net, and a νo ∈ Ar(µ) so that να −→ νo in the Prohorov topology. It follows from
upper semi continuity of η and (3.5) that
mo = limαη(να) ≤ lim sup
αη(να) ≤ η(νo). (3.6)
Since νo ∈ Ar(µ), it is apparent that η(νo) ≤ mo. It follows from this and the above
inequality that η(νo) = mo. Thus η attains its supremum on Ar(µ). Since η is strictly
concave and Ar(µ) is convex, it is clear that the maximizer is unique. Indeed, suppose
there is another element µo ∈ Ar(µ) at which η attains its maximum. By convexity of
Ar(µ), we have (1/2)νo+(1/2)µo ∈ Ar(µ), and since η is a strictly concave functional,
we have
η((1/2)νo + (1/2)µo) > (1/2)η(νo) + (1/2)η(µo) = mo. (3.7)
Since mo is the maximum on Ar(µ), this is impossible. This contradiction completes
the proof. •
Remark 3.2 It is clear from the above result that under the given assumptions, η
attains its supremum at a unique point in the set Ar(µ).
The following Lemma is used to established existence of the minimizing player.
6
Lemma 3.3 Suppose the assumptions of Lemma 2.1 hold. Then the graph G(Ar) of
the multifunction Ar is sequentially closed and the function Fr as defined by (2.3) is
lower semi continuous.
Proof. By definition the function Fr is given by
Fr(µ) ≡ supη(ν); ν ∈ Ar(µ).
By virtue of the previous lemma, for every µ ∈ M0, there exists a unique ν ∈ Ar(µ)
so that
Fr(µ) = η(ν) ≡ η(ν(µ)),
that is, the maximizer ν is uniquely determined by µ. Thus Fr is a well defined single
valued functional. First we prove that the graph of the multifunction µ −→ Ar(µ)
given by
G(Ar) ≡ (ν, µ) ∈ M1(X) ×M1(X) : ν ∈ Ar(µ)
is closed. Suppose (θn, βn) ∈ G(Ar) and θndP−→ θo, βn
dP−→ βo. We must show that
(θo, βo) ∈ G(Ar). Since θn ∈ Ar(βn), it follows from the definition of the multifunction
Ar that we have
H(θn|βn) ≤ r, for all n ∈ N.
Thus, it suffices to show that H(θo|βo) ≤ r. It is known that, by virtue of the Donsker-
Varadhan variational formula, the map (ν, µ) −→ H(ν|µ) is convex and lower semi-
continuous on its domain of definition. Thus
H(θo|βo) ≤ lim infH(θn|βn) ≤ r,
and hence the graph of Ar is sequentially closed, that is, θo ∈ Ar(βo). Now we show
that Fr is lower semi continuous. Let µo ∈ M1(X) and suppose µndP−→ µo. We show
that, along a subsequence if necessary, Fr(µo) ≤ lim inf Fr(µ
n). For the fixed positive
number r and given µo ∈ M1(X), consider the set
Ar(µo) ≡ ν ∈ M1(X) : H(ν|µo) ≤ r. (3.8)
7
Take any ε > 0 and consider the family of sets
Ar+ε(µn) ≡ ν ∈ M1(X) : H(ν|µn) ≤ r + ε, ε > 0. (3.9)
Since µndp−→ µo, for every ε > 0 there exists an nε ∈ N such that
Ar(µo) ⊂ Ar+ε(µ
n) ∀ n ≥ nε. (3.10)
By our assumption η is upper semi continuous and strictly concave. Since the above
sets are convex and compact (in the Prohorov topology), it follows from Lemma 3.1
that η attains unique supremum on each of the above sets and they are determined
by µo and µn, ε alone. Thus, by our definition of Fr given by the expression (2.3),
we have
Fr(µo) = supη(ν), ν ∈ Ar(µ
o) (3.11)
Fr+ε(µn) ≡ supη(ν), ν ∈ Ar+ε(µ
n). (3.12)
Clearly, it follows from the inclusion (3.10) and the identities (3.11), (3.12) that
Fr(µo) ≤ Fr+ε(µ
n), ∀ n ≥ nε. (3.13)
Hence
Fr(µo) ≤ lim inf
n→∞ Fr+ε(µn), ε > 0. (3.14)
Since this holds for any ε > 0, we have
Fr(µo) ≤ lim inf
n→∞ Fr(µn). (3.15)
Thus µ −→ Fr(µ) is lower semi continuous at µo. Since µo is arbitrary, this implies
that Fr is lower semi continuous on M1(X). •
Combining these results we prove the existence of solution of the problem (Pis).
Theorem 3.4 Consider the problem (Pis) and suppose that η : M1(X) −→ [−∞,+∞)
is upper semi continuous and strictly concave and the set M0 is compact with respect
to the Prohorov topology. Then the problem (Pis) has a solution.
8
Proof. The proof follows from Lemma 3.3. Indeed, by virtue of Lemma 3.3, the map
µ −→ Fr(µ) is lower semi continuous and by assumption M0 is compact. Hence Fr
attains its minimum on M0. •.
So far we have considered inf-sup problem. In certain situations the problem may be
reversed and we are required to solve the sup-inf problem
(Psi) : supµ∈M0
infη(ν) : ν ∈ Ar(µ).
This is solved in the following theorem.
Theorem 3.5 Consider the problem (Psi) and suppose η : M1(X) −→ (−∞,∞] is
lower semi continuous and the set M0 is compact in the Prohorov topology. Then the
problem (Psi) has a solution.
Proof. Define the function
Gr(µ) ≡ infη(ν) : ν ∈ Ar(µ).
Since η is lower semi continuous and bounded away from −∞ and Ar(µ) is compact,
this is a well defined real valued function on M1(X). As in Lemma 3.3, one can verify
that for any (dP ∼= weakly) convergent sequence µndP−→ µo in M1(X), we have
Gr(µo) ≥ lim sup
nGr(µ
n).
Thus µ −→ Gr(µ) is upper semi continuous. Since M0 is compact, Gr attains its
supremum on M0. This completes the proof. •
A Special Case. Let g : Rn −→ R be a real valued function and ϕi, i = 1, 2, · · ·n ∈BM(X), a family of bounded Borel measurable functions defined on X. Define the
9
functional
ηg(ν) ≡ g(ν(ϕ1), ν(ϕ2), · · · , ν(ϕn)) (3.16)
where ν(ϕ) ≡ ∫X ϕ(x)ν(dx), ϕ ∈ BM(X). The cost functional is given by
Fg(µ) ≡ supηg(ν), ν ∈ Ar(µ). (3.17)
The problem is to find a µ ∈ M0 that imparts a minimum to the functional Fg.
As a corollary of Theorem 3.4 we have the following result.
Corollary 3.6 Let g : Rn −→ [−∞,+∞) be upper semi continuous and strictly
concave, ϕi1≤i≤n an arbitrary family from BM(X), and M0 a compact subset of
M1(X). Then there exists a µ∗ ∈ M0 at which the functional Fg, given by (3.17),
attains its minimum.
Remark 3.7 It is important to point out that the concavity assumption imposed on
the functional η guarantees uniqueness of the maximizer. But this is not necessary if
we are only interested in the question of existence of a solution to the inf-sup problem
(Pis) as described by the expression (2.4). In this case all the preceding results remain
valid without the uniqueness. To clarify this further, take any element µ ∈ M1(X)
and define
Cr(µ) ≡ν ∈ Ar(µ) : η(ν) = supη(ϑ), ϑ ∈ Ar(µ)
.
This is the set of maximizers of η on the set Ar(µ). It follows from the upper semi
continuity of η that this set is closed and hence, being a closed subset of a compact set,
is compact. Clearly, the functional Fr(µ), as given by the expression (2.3), satisfies
the obvious identity,
Fr(µ) = η(Cr(µ)).
Thus the lack of concavity does not affect the functional Fr and hence the results on
existence.
10
3.1 Examples
We present some examples illustrating the preceding results.
(E1)(Target seeking Problem): Let µd ∈ M1(X) be a desired measure, and
suppose we wish to find a µ ∈ M0 (the attainable set) that approximates µd as closely
as possible. Clearly the obvious choice of the objective functional is the Prohorov
metric giving
η(ν) ≡ dP (µd, ν).
Since dP is a metric, the functional η as defined is continuous. If one disregards
the uncertainty, one would be willing to minimize this functional on the attainable
set of the perfect system given by M0. By our assumption this set is compact and
since η is continuous it attains its minimum on it. Let µo ∈ M0 be a minimizer.
Then η(µo) ≤ η(µ), µ ∈ M0. But since the system is uncertain, the actual law (true
measure) in force may be any element from Ar(µo) and one may face the worst
situation,
η(ν∗) ≡ supη(ν), ν ∈ Ar(µo) = Fr(µ
o).
Clearly η(ν∗) ≥ η(µo) with H(ν∗|µo) ≤ r. So instead of choosing µo, we may choose
one that minimizes the maximum distance. That is, we choose an element µ∗ ∈ M0
so that
Fr(µ∗) = infFr(µ), µ ∈ M0
where
Fr(µ) ≡ supη(ν) ≡ dP (µd, ν) : ν ∈ Ar(µ).
Clearly
Fr(µ∗) ≤ Fr(µ
o) = η(ν∗).
Thus, we conclude that the inf-sup strategy is closer to the desired goal than the
strategy based on the deterministic dynamics with attainable set M0.
11
(E2)(Evasion Problem): Let Γk, 1 ≤ k ≤ n ⊂ BX be a family of disjoint sets
and suppose we wish to find a measure from the attainable set M0 that has minimum
concentration of mass on these sets. This represents an evasion problem. But since
the system is uncertain, we have a sort of games problem, game against natural
uncertainty. Thus we may formulate the problem as follows. Let g : Rn+ −→ R+ ≡
[0,∞) be any bounded continuous function, increasing in all its arguments. Define
the functional η as follows,
η(ν) ≡ g(ν(Γ1), · · · , ν(Γn))≡ g(ν(χΓ1), · · · , ν(χΓn)),
where χΓ denotes the indicator function of the set Γ ∈ BX . Disregarding uncertainty,
and noting that η is continuous, and M0 compact, there exists a µo ∈ M0 at which
η attains its minimum, η(µo) ≤ η(µ), µ ∈ M0. Considering the uncertainty at µo we
have Fr(µo) ≥ η(µo) where the function Fr is as defined before. Since Fr is lower semi
continuous and M0 is compact, letting µ∗ ∈ M0 denote the point at which Fr attains
its infimum, we conclude that Fr(µ∗) ≤ Fr(µ
o). In other words the inf-sup strategy
exists and is better than pure deterministic strategy which ignores the presence of
uncertainty.
In example (E2), if we assume that the sets Γk, 1 ≤ k ≤ n are closed subsets of
X, then for the existence of inf-sup strategy, it is sufficient if g is merely upper semi
continuous.
(E3)(Hitting Problem) Let O1,O2, · · · ,On be a family of open subsets of X
and g : Rn −→ (−∞,∞] an increasing function of all its arguments. Consider the
functional ηg given by
ηg(ν) ≡ g(ν(O1), · · · , ν(On)).
The objective is to find a ν from the attainable set of measures that maximizes the
12
concentration of mass on the given sets. This is equivalent to hitting the target with
maximum probability. Again due to the presence of uncertainty, we must maximize
the minimum concentration giving rise to the sup-inf problem
supµ∈M0
infηg(ν), ν ∈ Ar(µ).
By virtue of Theorem 3.5, it suffices to verify that the functional ηg, as defined above,
is lower semi continuous on M1(X) in the Prohorov topology. Suppose νkw−→ νo,
then, since Oi are open sets, we have
lim infk→∞
νk(Oi) ≥ νo(Oi), i = 1, 2, · · · , n.
Thus, g being an increasing function of it’s arguments, we have
lim infk→∞
g(νk(O1), · · · , νk(On)) ≥ g(lim infkνk(O1), · · · , lim inf
kνk(On))
≥ g(νo(O1), · · ·νo(On)).
Hence
lim infk→∞
ηg(νk) ≥ ηg(ν
o)
and so ηg is lower semi continuous on M1(X). Since for each µ ∈ M1(X), Ar(µ) is
compact, ηg attains its infimum on it determined by µ alone. Thus the function
Gr(µ) ≡ infηg(ν), ν ∈ Ar(µ)
is well defined. Now using similar arguments as in Lemma 3.3, we can verify that for
any convergent sequence µn with µndP−→ µo, we have
Gr(µo) ≥ lim sup
n→∞Gr(µ
n).
Thus Gr is upper semi continuous in the Prohorov topology and so attains it maximum
on M0. Hence the (sup-inf) problem as stated above has a solution.
13
4 Uncertain Stochastic Differential Equations on
Hilbert Space
In this section we wish to apply the results of the previous section to a class of
stochastic control problems in infinite dimensional spaces. This is a considerable gen-
eralization of the results found in [15-19]. Consider the controlled stochastic system
dx = Axdt+ fo(x, u)dt+B(x)dW, x(0) = x0, t ∈ I ≡ [0, T ], (4.18)
on a separable Hilbert space H, with A being the infinitesimal generator of a C0
semigroup S(t), t ≥ 0, on H , f0 : H × U −→ H and B : H −→ L(H) suitable
maps and W a cylindrical Brownian motion adapted to the filtered probability space
(Ω,F ,Ft ↑, P ). It is assumed that all the parameters A, f0, B, x0 defining the system
are perfectly known. Very often the system parameters, possibly based on noisy
estimates, are not exactly known. In this case the true physical system may be
described by a perturbed version of (4.18) as follows,
dx = Axdt+ fo(x, u)dt+ f(x)dt+B(x)dW, t ∈ I ≡ [0, T ],
x(0) = x0, f(x) ∈ F (x), x ∈ H, (4.19)
where f is any measurable selection of a Borel measurable multifunction F represent-
ing the uncertainty in the nominal drift f0. Note that F may absorb any bounded
perturbation of the unbounded operator A and those associated with the nominal
drift. We identify the system (4.18) as the nominal system and system (4.19) as the
true physical system with parametric uncertainty as described above.
In this section we consider a control problem for the system (3.19) which is described
by parametric uncertainty. Such problems arize in robust control problems on infinite
dimensional spaces.
Admissible Controls: Let U be a closed subset of another Polish space and Uad
14
the class of admissible controls taking values from U. More detailed information on
the structure of the class of admissible controls will be presented later.
Existence of Solutions, Properties of Attainable Measures: Standard assump-
tions imposed on f0, B guaranteeing the existence of mild solutions are as stated
below [3] . There exist constants K,L > 0, such that for all v ∈ U , f0 and B satisfy
the following Lipschitz and growth properties:
‖ f0(x, v) ‖2 + ‖ B(x) ‖2H.S≤ K(1+ ‖ x ‖2), (4.20)
‖ f0(x, v) − f0(y, v) ‖2 + ‖ B(x) − B(y) ‖2H.S≤ L ‖ x− y ‖2, (4.21)
for all x, y ∈ H where ‖ B(x) ‖H.S denotes the Hilbert-Schmidt norm of the operator
B(x). Under the above assumptions, it is well known [3] that for each control u ∈Uad the (nominal) system (4.18) has a unique Ft-adapted mild solution x with the
properties
(p1): supt∈I E ‖ x(t) ‖2H<∞ and (p2): x ∈ C(I,H) P − a.s.
In case controls are based on state feedback, the Lipschitz property imposed on the
drift f0 is rather restrictive and may be lost. Thus it is necessary to generalize this.
This can be done by accepting martingale solution (weak solution).
Lemma 4.1 Consider the system
dx = Axdt+B(x)dW, x(0) = x0, (4.22)
and suppose A generates a C0-semigroup S(t), t ≥ 0, on H, B satisfy the assumptions
(4.20) and (4.21) and that E ‖ x0 ‖2< ∞. Then (4.22) has a unique mild solution
with sample paths x ∈ C(I,H) P -a.s satisfying Esupt∈I ‖ x(t) ‖2H <∞.
Proof See Da Prato- Zabczyk [ 3, Chapter 7].
15
Lemma 4.2 Consider the control system
dx = Axdt+ f0(x, u)dt+B(x)dW, x(0) = x0. (4.23)
Suppose A and B satisfy the assumptions of Lemma 4.1 and further there exists a
finite positive number γ such that
sup‖ B−1(x)f0(x, v) ‖H , (x, v) ∈ H × U ≤ γ (4.24)
and f0 satisfy the growth assumption (4.20), and that E ‖ x0 ‖2< ∞. Then for any
given control u ∈ Uad, the system (4.23) has a martingale solution uniquely determined
by the solution of (4.22) and the control u ∈ Uad.
Proof Under the given assumption (4.24) it follows from Theorem 10.14 (see also
Proposition 10.17) [3] that the process
W (t) ≡W (t) −∫ t
0B−1(x(s))f0(x(s), us) ds (4.25)
evaluated along the mild solution of (4.22) is a cylindrical Brownian motion on an
extended (Skrokhod extension) probability space (Ω, F , Ft ↑, P ). Thus, on this prob-
ability space the mild solution of (4.22) is in fact the martingale solution of (4.23)
driven by the Browniam motion W . By Girsanov theorem [3, Da Prato-Zabczyk,
Theorems 10.14, 10.18, pp.290], the probability measure induced by the controlled
process governed by (4.23) is absolutely continuous with respect to that induced by
the process governed by (4.22). More precisely, let X ≡ C(I,H) denote the space of
continuous functions on I with values in H. Furnished with the standard sup norm
topology this is a Banach space. Let BX denote the Borel algebra of subsets of the
set X making (X,BX) a measurable space. This is the canonical sample space, a
Borel measurable space (X,BX), on which we may define different measures. Let µ0
denote the measure induced by the system (4.22) on BX and µu the one induced by
the system (4.23). Then by virtue of Girsanov theorem as mentioned above, we have
µu ≺ µ0 and the RND is given by
(dµu/dµ0) = exp∫ T
0(B−1f0, dW ) − (1/2)
∫ T
0‖ B−1f0 ‖2
H ds≡ Λ(u). (4.26)
16
This means that the martingale (or weak) solution of (4.23), described in terms of
the probability measure it induces on the path space X, is uniquely determined by
the measure induced by (4.22) and the control u ∈ Uad through the relation
dµu = Λ(u)dµ0. (4.27)
This completes the proof. •
Remark 4.3 The assumption (4.24) is satisfied if f0 satisfies the growth condition
(4.20) and the operator B(x) is surjective satisfying
‖ B(x)z ‖H≥ c(1+ ‖ x ‖q) ‖ z ‖H , q ≥ 1, (4.28)
for a constant c > 0.
Theorem 4.4 Consider the control system (4.23) and suppose the assumptions of
Lemma 4.2 hold, and the set Drnd ≡ Λ(u), u ∈ Uad is a closed bounded subset of
L1(X,µ0) satisfying
limµ0(Γ)→0
∫Γ
Λ(u)dµ0 = 0 uniformly with respect u ∈ Uad.
Then the set of attainable measures M0, induced by the control system (4.23), is
compact in the Prohorov topology.
Proof. The proof is a direct consequence of the celebrated Dunford-Pettis theorem
[4, pp.93]. By virtue of our assumptions, the set Drnd is a closed, bounded, uniformly
integrable subset of L1(X,µ0) and so by Dunford-Pettis theorem it is a weakly se-
quentially compact subset of L1(X,µ0). Hence, the attainable measures described
by
M0 ≡ µ ∈M1(X) : µ(K) =∫Kgdµ0, g ∈ Drnd, K ∈ BX
is a weakly compact subset of M1(X). Since H is a separable Hilbert space, C(I,H)
is a separable metric space with respect to the supnorm topology and hence a Polish
17
space. Thus M1(X) is a Polish space and the weak topology on M1(X) is equivalent
to the Prohorov metric topology. Hence M0 is compact in the Prohorov topology.
This completes our proof. •
Remark 4.5 Sufficient conditions guaranteeing the assumed properties of the set
Drnd are (i): u −→ f0(x, u) is continuous from U to H, (ii): Uad is a compact
metric space, and (iii): sup(x,u)∈H×U ‖ B−1(x)f0(x, u) ‖< ∞. This later condition
follows easily from the properties (4.20) and (4.28).
Perturbed System. The uncertain system described by (4.19) is essentially a
stochastic differential inclusion
dx ∈ Axdt+ f0(x, u)dt+ F (x)dt+B(x)dW, x(0) = x0, t ∈ I. (4.29)
We assume throughout that F : H −→ 2H \ ∅ is a measurable multifunction in the
sense that for every Γ ∈ BH the set x ∈ H : F (x) ∩ Γ = ∅ ∈ BH .
Let c(H) ⊂ 2H \ ∅ denote the class of nonempty closed subsets of H . The following
result relates the uncertain system (4.29) to the nominal system (4.23) in terms of
relative entropy (information distance).
Theorem 4.6 Consider the system (4.29) and suppose A, f0, B satisfy the assump-
tions of Lemma 4.2. Further, suppose that F is a measurable multifunction mapping
H to c(H) and there exists a finite positive number β such that
sup‖ B−1(x)h ‖H , h ∈ F (x)
≤ β, ∀ x ∈ H. (4.30)
Then for each fixed u ∈ Uad, the system has a nonempty set of martingale solutions
denoted by MF ,u ⊂ M1(X) and that each member ν of MF ,u is absolutely continuous
with respect to the corresponding martingale solution µu of the nominal (or perfect)
system (4.23). Further, there exists a finite positive number r such that
H(ν|µu) ≤ r, ∀ ν ∈ MF ,u u ∈ Uad.
18
Proof. Since H is a separable Hilbert space, it is also a Polish space with respect to
its standard topology. Clearly (H,BH) is a measurable space and by our assumption
F : H −→ c(H) is measurable. Thus it follows from standard selection theorem [ see
2, Theorem 2.1, pp.154] that F has a nonempty set of measurable selections which
we may denote by SF . Let f ∈ SF and consider the system
dx = Axdt+ f0(x, u)dt+ f(x)dt+B(x)dW, x(0) = x0, t ∈ I. (4.31)
For any u ∈ Uad, let µu ∈ M1(X) be the measure corresponding to the martingale
solution of (4.23). Then following identical arguments as in the proof of Lemma 4.2,
one can verify that ρ given by
ρ ≡ Exp∫
I(B−1f , dW ) − (1/2)
∫I‖ B−1f ‖2
H dt
(4.32)
is the RND of the measure νu ∈ M1(X), associated with the martingale solution of
(4.31), with respect to the measure µu, associated with the martingale solution of
(4.23). Computing the entropy of νu relative to µu, we find that
H(νu|µu) = (1/2)∫Xϕ(ξ)dνu(ξ) (4.33)
where
ϕ(x) ≡∫I‖ B−1(x(t))f (x(t)) ‖2
H dt , x ∈ X ≡ C(I,H).
By our assumption (4.30), it follows from (4.33) that
H(νu|µu) ≤ (1/2)β2T ≡ r. (4.34)
Since (4.30) holds for all selections f ∈ SF , and every admissible control u, we con-
clude that H(νu|µu) ≤ r for all u ∈ Uad. •
Pay-Off Functional. Let (U, d) be a compact metric space and Bρ(H,U) denote the
class of Borel measurable functions from H to U furnished with the metric topology
ρ(f, g) ≡ supx∈H
d(f(x), g(x)), f, g ∈ Bρ(H,U).
19
Let Uo denote the class of measurable functions on I with values in the metric space
Bρ(H,U). We may introduce a metric on it as follows.
γ(u, v) ≡ λt ∈ I, ρ(ut, vt) = 0,
where λ denotes the Lebesgue measure on I. With respect to this topology, (U0, γ) is a
complete metric space. Let Uad ⊂ U0 denote the class of admissible controls assumed
to be compact with respect to the γ topology. Thus, this class of controls represent
a class of compact Markovian feedback controls.
Now consider the classical control problem (Pc) : Find a control uo ∈ Uad at which
the following inf-sup is attained
Jo ≡ infu∈Uad
supν∈MF ,u
Eν
(∫I (t, x, u)dt+ Ψ(x(T ))
), (4.35)
where Eν(·) denotes integration with respect to the measure ν, a martingale solu-
tion of (4.29) corresponding to a given Markovian control u and a given measurable
selection f ∈ SF . Defining
ψu(x) ≡∫I (t, x(t), u(t, x(t)))dt+ Ψ(x(T )),
our problem can be compactly presented as follows: Find u ∈ Uad that minimizes the
functional Jo(u) given by
Jo(u) ≡ supν∈MF ,u
∫Xψu(x)dν(x). (4.36)
In other words,
Jo ≡ infu∈Uad
Jo(u). (4.37)
Clearly this is a particular case of the general problem studied in Section 3. In view of
the discussions of Section 3, this problem can be formulated as the original problem
studied in Section 2, expression (2.4). For this problem one has to choose η,M0, and
the multifunction Ar as given below:
η(ν) ≡∫Xψu(x)dν(x), ν ∈ MF ,u
M0 ≡ µ ∈ M1(X) : dµ = gdµo, g ∈ Drnd (4.38)
Ar(µ) ≡ ν ∈ M1(X) : H(ν|µ) ≤ (1/2)β2T ≡ r, µ ∈ M0.
20
The min-max problem (Pc) has a solution as stated in the following corollary.
Corollary 4.7 Consider the control problem (Pc) for the uncertain system described
by the Stochastic Differential Inclusion) (SDI) (4.29) and suppose the assmptions
of Theorem 4.4 and 4.6 hold and that and Ψ are bounded away from −∞. Let
η,M0,Ar be as described by (4.38). Then the problem (Pc) has a min-max strategy
uo ∈ Uad.
Proof. The proof is similar to that of Theorem 3.4. The important difference here is
to note that the functional η must be maximized on MF ,u ⊂ Ar(µu) instead of Ar(µ
u).
Since F is closed valued, the set of measurable selections denoted by SF is closed.
That is, the limit of any (uniformly) convergent sequence of measurable selections
is a measurable selection of F . This implies that the set MF ,u is a closed subset of
the compact set Ar(µu), and hence itself compact (in the Prohorov topology). If the
functional Jo given by
Jo(u) ≡ supη(ν), ν ∈ MF ,u
is identically +∞ for every u ∈ Uad, then there is nothing to prove. So we may assume
that there is a nonempty set Us ⊂ Uad on which the supremum is finite. Without loss
of generality we may assume that the whole admissible set has this property. Then
for any u ∈ Uad, η is a bounded linear and hence continuous functional and it attains
its supremum on MF ,u. Thus we may conclude that the functional Jo(u) is a well
defined real valued functional. Hence the functional Fr given by
Fr(µu) = Jo(u)
is well defined. Since M0 is compact and Fr is lower semi continuous, there exists a
µo ∈ M0 at which Fr attains its minimum and hence there exists a control uo ∈ Uadat which Jo(·) attains its minimum. This completes our proof. •
21
5 Construction of the Maximizing Measure
In general a non parametric model may cover a wider class of uncertain stochastic
control systems provided the uncertainty is measured in terms of Prohorov metric
topology. But for a fairly large class of physical systems, relative entropy (though it is
not a metric) provides a good and mathematically convenient measure of uncertainty.
This is the measure we have use throughout the paper. Let X denote a Polish space,
BX the Borel algebra of subsets of the set X, and M1(X) the space of probability
measures defined on BX . This is the state space. Let Uad denote the class of admissible
controls. The perfect system can be described by the single valued map,
Φ : Uad −→ M1(X),
that assigns a unique element µ ∈ M1(X) corresponding to each choice of control
policy u giving Φ(u) = µ. Clearly, the attainable measures, denoted by the set
M0 ≡ Φ(u), u ∈ Uad,
describes the power of the control system in the sense of its capacity to produce
a wider class of probability measures (possibly with different support properties)
guaranteeing broader freedom of choice and control. The model for the uncertain
system may be chosen as follows: for any choice of µ ∈ M0 and any real number
r > 0, define the set valued map as before by
Ar(µ) ≡ ν ∈ M1(X) : H(ν|µ) ≤ r.
Clearly this set is always non empty, since it contains µ itself. Then the uncertain
system is governed by the multi valued map
Φr(u) = Ar(Φ(u)) = Ar(µu),
with range given by
Mr ≡⋃Φr(u), u ∈ Uad =
⋃Ar(µ);µ ∈ M0.
22
The min-max problems studied in Section 3 directly applies to this later class of
uncertain systems.
The Worst Case Measure. This section is concerned with reformulating con-
strained optimization problems as unconstrained ones. For the uncertain system
described in Section 4, with the optimization problem (4.35) or equivalently (4.36),
(4.37), this is not easy. This is because the set MF ,u ⊂ Ar(µu) may not have simple
geometric and topological properties (such as convexity, compactness etc.) as does
the set Ar(µu). However, for nonparametric uncertainty as discussed in the introduc-
tion of this section, this is relatively easy. In other words the optimization problem
(4.35) or its equivalent (4.36), (4.37) is relaxed to the following simpler problem:
Find u ∈ Uad that minimizes the following functional.
Jo(u) ≡ sup∫Xψu(x)dν(x), ν ∈ Ar(µ
u)
≡ inf∫X−ψu(x)dν(x), ν ∈ Ar(µ
u). (5.39)
First we consider the problem (5.39). Using duality theory (Lagrange functionals),
we show that this constrained problem is equivalent to an unconstrained problem.
Throughout this section, for convenience of notation, we write µu for Φ(u) and con-
sider the number r > 0, describing Ar(µu), fixed. For an arbitrary but fixed u ∈ Uad
or equivalently the associated martingale solution µu, define the Lagrangian
L(u; s, ν) ≡ Eν(ψu) − s(H(ν|µu) − r), s ∈ R0 ≡ [0,∞), u ∈ Uad, ν ∈ M1(X), (5.40)
and suppose there exist a ν∗s ∈ M1(X) and s∗ ≥ 0 such that
L(u; s, ν∗s ) = supL(u; s, ν), ν ∈ M1(X), (5.41)
and
ϕ(u, s∗)) ≡ infs≥0
L(u; s, ν∗s ). (5.42)
23
Lemma 5.1 Suppose that for each u ∈ Uad, ψu ∈ L1(X, ν) for all ν ∈ Ar(µu). Then
the problem (5.39) is equivalent to the unconstrained problem
Jo(u) = infs≥0
supν∈M1(X)
L(u, s, ν). (5.43)
Proof The proof follows directly from [5, Theorem 1, pp.224] once we identify the
relevant functions f,G introduced there, as
f(ν) ≡ Eν(−ψu), and G(ν) ≡(H(ν|µu) − r
)
and the relevant function spaces denoted by Ξ,Z as follows. Let M(X) denote the
vector space of bounded countably additive signed measures on the Borel field of
subsets of the Polish space X. Furnished with the total variation norm, this is a
Banach space. Choose Ξ = M(X) and for the ordered vector space Z take Z ≡ R
the real line with the standard topology and natural ordering so that it is a ordered
vector space. For the constraint set we take Ω ≡ Ar(µu). Clearly f, as defined above,
is a convex functional and, since the set Ar(µu) is a convex subset of M1(X), it is also
a convex subset of M(X) and so the constraint set Ω is a convex subset of M(X).
Since
ν −→ H(ν|µu)
is convex, the function G(ν), as defined above is a convex functional. Further note
thatG(µu) = −r ≤ 0. Thus all the assumptions of [5, Theorem 1, pp.224] are satisfied.
Hence the conclusion follows. •
Theorem 5.2 Suppose the assumptions of Lemma 4.1 hold. Then
Jo(u) = ϕ(u, s∗) ≡ infs≥0
L(u, s, ν∗s ).
Proof. Follows directly from the above Lemma. •
24
Next, we present the expression of the worst case measure and some of the properties
of the dual functional.
Theorem 5.3 Let u ∈ Uad and µu ∈ M1(X) the measure induced by the unperturbed
system. Suppose that for all s > 0, ψu/s and exp(ψu/s) ∈ L1(X,µu). Then the
following statements hold:
(S1): The dual functional L(u, s, ν∗s ) given by (5.41) is related to the cumulant gen-
erating function of ψu with respect to µu through the following relation
L(u, s, ν∗s ) = s sup(1/s)
∫ψu(x)dν −H(ν|µu), ν ∈ M1(X)
+ sr (5.44)
= s log(∫
Xexpψu(x)/s)dµu
)+ sr = sCµu(1/s) + sr, (5.45)
where
Cµu(θ) ≡ log∫Xeθψu(x)dµu, θ ∈ R.
(S2): The measure ν∗s that maximizes the expression within the parenthesis of equa-
tion (5.44) is given by
dν∗s =exp(ψu(x)/s)dµ
u
∫X exp(ψu(x)/s)dµ
u, (5.46)
and further we have
∫Xψu(x)dν
∗s (x) = s log
∫Xexpψu(x)/sdµu + sH(ν∗s |µu), s ∈ (0,∞). (5.47)
(S3): There exists a unique s∗ ∈ (0,∞) at which ϕ(u, s) ≡ L(u, s, ν∗s ) attains its
infimum.
(S4): The measure ν∗s∗ that maximizes the functional (5.39) belongs to the boundary
of the constraint set, that is, ν∗s∗ ∈ ∂Ar(µu).
25
Proof. (S1): By virtue of Legendre-Fenchel duality,
supν∈M1(X)
∫X
(ψu/s)dν −H(ν|µu)
= log(∫
Xexp(ψu/s)dµ
u). (5.48)
Hence (5.45) follows from (5.44) and (5.48). The second identity is merely the def-
inition of cumulant generating function. (S2): This follows from direct verification.
Substituting ν∗s given by (5.46) in the expression within the parenthesis of equation
(5.44) we have
∫X
(ψu/s)dν∗s −H(ν∗s |µu) = log
∫X
exp(ψu/s)dµu. (5.49)
Thus it follows from (5.48) of (S1) that ν∗s , given by (5.46), is the maximizer of
L(u, s, ν). Identity (5.47) is an obvious consequence of this result. (S3): By definition
ϕ(u, s) = L(u, s, ν∗s ). Hence using ν∗s given by (5.46) in this expression we have
ϕ(u, s) = L(u, s, ν∗s ) = s log(∫
Xexp(ψu/s)dµ
u)
+ sr. (5.50)
For fixed u, this is a C1 function in the variable s ∈ (0,∞]. By direct computation
we find that
dϕ(u, s)/ds = log∫X
exp(ψu/s)dµu −
∫X(ψu/s) exp(ψu/s)dµ
u
∫X exp(ψu/s)dµu
+ r. (5.51)
By equating this to zero we determine s∗ that minimizes (or maximizes) ϕ(u, s).
Define the function h by
h(θ) ≡∫X(θψu) exp(θψu)dµ
u
∫X exp(θψu)dµu
− log∫X
exp(θψu)dµu, θ ∈ [0,∞). (5.52)
By equating (5.51) to zero, it follows from (5.52) that finding s that minimizes ϕ(u, s)
is equivalent to finding a θ∗ ∈ (0,∞) at which h satisfies the following equation
h(θ) = r. (5.53)
It is clear that h(0) = 0; we show that h is a monotonically increasing function over
the set R0 ≡ [0,∞). By taking the derivative of h with respect to θ followed by some
elementary manipulations we find
h′(θ) = θ
∫X
(ψu)2dλθ −
(∫Xψudλθ
)2(5.54)
26
where λθ ≡ ν∗1/θ. Note that λ0 = µu and that the expression within the parenthesis
of equation (5.54) is strictly positive for all θ ∈ [0,∞). It is zero only when ψu is
a constant ( a trivial case) . Hence h is monotonically increasing and so, for any
finite positive number r, there exists a unique θ∗ such that h(θ∗) = r implying that
equation (4.53) has a unique solution. Hence (S3) follows upon taking s∗ = (1/θ∗).
(S4): Using the measure ν∗s∗ or equivalently ν∗1/θ∗ in the expression for its entropy
relative to the measure µu, we have
H(ν∗1/θ∗|µ) =∫X
log(dν∗1/θ∗dµu
)dν∗1/θ∗ (5.55)
=∫X
(θ∗ψu) exp(θ∗ψu)dµu∫X exp(θ∗ψu)dµu
− log(∫X
exp(θ∗ψu)dµu). (5.56)
Clearly this last expression coincides with the expression (5.52) evaluated at θ∗. Hence
H(ν∗1/θ∗|µu) = h(θ∗) = r (5.57)
proving that ν∗1/θ∗ ∈ ∂Ar(µu). This completes the proof. •
Evolution of the Worst Case Measure. The worst case measure (5.46) is defined
on the path space C(I,H) ≡ X. It will be desirable to introduce another equivalent
problem to (5.39), in which the nominal and uncertain measures are defined on H .
This is done as follows. Substituting ν∗s into the unconstraint Lagrangian, we found
as seen in Theorem 5.3, (S1) that
L(u; s, ν∗s ) = sr + s log ∫
Xe
1sψu(x)dµu
. (5.58)
Next, define the measure valued process πut ; 0 ≤ t ≤ T on H by
πut (φ) = Eµuφ(x(t)) exp
(1
s
∫ t
0 (τ, x(τ), u(τ))dτ
)=
∫Hφ(z)dπut (z), (5.59)
for t ∈ I ≡ [0, T ]. Also, introduce the operator L(u) associated with the nominal
system (4.18). This is given by
(L(u)φ)(t, x) = (A∗Dφ(x), x)H + (Dφ(x), fu(t, x))H + (1/2)Tr(D2φBB∗)(x),
27
for t > 0, x ∈ D(A) ⊂ H. Then one can verify that for any φ ∈ D(L(u)), πut (φ), 0 ≤t ≤ T satisfies the following identity
πut (φ) = π0(φ) +∫ t
0πuτ (L(u(τ))φ)(τ)dτ +
1
s
∫ t
0πuτ ( (u(τ))φ)(τ)dτ, (5.60)
where π0 ∈ M1(H) is the law of the initial state. Since this identity holds for all
φ ∈ D(L(u)), πu is the weak solution of the Forward Kolmogorov equation
dπut = L∗(u)πut dt+B∗uπ
ut dt, π0 = p0, (5.61)
where B∗u is the adjoint of the operator Bu given by Buφ ≡ (1/s) uφ which is a
multiplication operator with (1/s) u acting as a potential.
Using the definition of the measure-valued process πut (·); 0 ≤ t ≤ T, given by (5.59),
we deduce
∫Xe
1sψu(x)dµu =
∫Xe
1sΨ(x(T ))+ 1
s
∫ T
0(t,x(t),u(t))dtdµu = πuT (e
1sΨ). (5.62)
Substituting (5.62) into (5.58) we have
L(u; s, ν∗s ) = sr + s log ∫
Xe
1sψu(x)dµu
= sr + s log
∫He
1sΨ(z)dπuT (z)
. (5.63)
Notice that the resulting Langrangian is expressed in terms of the measure πuT (·) ∈M1(H), instead of the measure µu ∈ M1(X). Now using Legender-Fenchel duality,
the last term of (5.63) can be written as
s log ∫
He
1sΨ(z)dπuT (z)
= supχ∈M1(H)
∫H
Ψ(z)dχ(z) − sH(χ|πuT ). (5.64)
Moreover, as seen before, the measure χu,∗s that maximizes the above expression is
given by
dχu,∗s (z) =exp1
sΨ(z)dπuT (z)∫
H exp1sΨ(z)dπuT (z).
(5.65)
28
Thus, the original problem is transformed to an equivalent one on H rather than X,
and we have the following identity.
L(u; s, ν∗s ) = supν∈M1(X)
∫Xψu(x)dν − s
(H(ν|µu) − r
)
= sr + s log ∫
Xe
1sψu(x)dµu
= sr + s log ∫
He
1sΨ(z)dπuT (z)
= supχ∈M1(H)
∫H
Ψ(z)dχ(z) − s(H(χ|πuT ) − r
). (5.66)
It is important to note that the radius of uncertainty r used here may be different
from the one used in the original path space X.
Hence, we have proved the following theorem.
Theorem 5.4 Define the sphere of uncertainty by
Br(πuT ) =χ ∈ M1(H) : H(χ|πuT ) ≤ r
.
Then Jo(u) defined by (5.39) is equivalent to the transformed optimization problem
Jotran(u) = sup ∫
HΨ(z)dχ(z), χ ∈ Br(πuT )
(5.67)
where the nominal measure is πuT ∈ M1(H) obtained from the weak solution of equa-
tion (5.61). Moreover, the worst case measure, that is, the measure at which (5.67)
is realized, is given by (5.65).
Furthermore, the statements of Theorem 5.3 (S1), (S2) hold with the obvious changes
ψu → Ψ, µu → πu, νu → χu, X → H. In addition, when = 0 statements of Theo-
rem 5.3 (S3), (S4) also hold, with πu given by (5.59) with = 0.
Proof. The equivalence of the two problems follows from the fact that L(u; s, ν∗s )
given by the first equality in (5.66) is the Lagrangian associated with Jo(u) defined
by (5.39), while the last equality in (5.66) is the Lagrangian associated with (5.67).
The worst case measure is given by (5.65). The last statement holds following the
derivation of Theorem 5.3.
29
Remark 5.5 For complete solution of the problem we convert (5.67) into an uncon-
straint problem using the last two identities in (5.66) to obtain the cost functional
Jotrans(u) and then minimize this functional subject to measure dynamics (5.61). This
can be formulated as a terminal control problem as follows. Note that
Jotrans(u) = infs>0
sr + s log
∫He
1sΨ(z)dπuT (z)
≡ inf
s>0Φ(πuT ).
Thus, the problem is
Φ(πuT ) −→ infu∈Uad
infs>0
subject to the dynamic constraint
dπut = L∗(u)πut dt+B∗uπ
ut dt, π0 = p0.
Moreover, when = 0, then the problem is
Φ(πuT ) −→ infu∈Uad
infs>0
subject to the dynamic constraint
dπut = L∗(u)πut dt, π0 = p0,
and since the solution πut (·); 0 ≤ t ≤ T is independent of s, then s∗ corresponds to
the boundary constraint H(χu,∗|πuT )|s=s∗ = r, while the infimum over u ∈ Uad is found
independently.
We leave this problem for future research.
6 Conclusion
This paper derives several results for minimax games of stochastic uncertain sys-
tems, when the pay-off is a nonlinear functional of the uncertain measure, and the
uncertainty is measured in terms of relative entropy between the uncertain and the
nominal measure. Existence and uniqueness of minimax solutions are derived on
suitable spaces of measures. Subsequently, the results are also applied to controlled
Stochastic Differential Equations (SDE’s) on Hilbert Spaces. Based on infinite dimen-
sional extension of Girsanov’s measure transformation, martingale solutions are used
30
in establishing existence and uniqueness of minimax strategies. The worst case mea-
sure representing the maximizing player is found and it is shown that it is described
by an evolution equation on the space of measures on the original state space.
7 References
[1]. P.Dupuis and R.S.Ellis, (1997), A Weak Convergence Approach to the Theory of
Large Deviations, John Wiley.
[2]. S.Hu and N.S.Papageorgiou,(1997), Handbook of Multivalued Analysis, Vol 1:
Theory, Kluwer Academic Publishers, DOrdrecht/Boston/London.
[3]. G.Da Prato and J. Zabczyk, Stochastic Equations in Infinite Dimensions,(1992)
Encyclopedia of Mathematics and its Applications, Cambridge University Press, (1992).
[4].J.Diestel,(1984), Sequences and Series in Banach Spaces, Springer-Verlag New
York,Inc
[5]. D. G. Luenberger, (1969), Optimization by Vector Space Methods, John Wiley
[6] G. Zames, “Feedback and optimal sensitivity: Model reference transformations,
multiplicative seminorms, and approximate inverses,” IEEE Transactions on Auto-
matic Control, vol. 26, no. 2, pp. 301–320, 1981.
[7] J. Willems, “Dissipative dynamical systems part I: General theory,” Arch. Rational
Mech. Anal., vol. 45, pp. 321–351, 1972.
[8] D. Jacobson, “Optimal stochastic linear systems with exponential performance
criteria and their relation to deterministic differential games,” IEEE Transactions on
Automatic Control, vol. 18, no. 2, pp. 124–131, 1973.
31
[9] A. Bensoussan and J. H. van Schuppen, “Optimal control of partially observable
stochastic systems with an exponential-of-integral performance index,” SIAM Journal
on Control and Optimization, vol. 23, no. 4, pp. 599–613, 1985.
[10] P. Whittle, “A risk-sensitive maximum principle,” Systems and Control Letters,
vol. 15, pp. 183–192, 1990.
[11] W. H. Fleming and W. M. McEneaney, “Risk-sensitive control and differential
games, pp.185-197,” in Stochastic Theory and Adaptive Control, eds., T.E. Duncan
and B. Pasik-Duncan, (New York), Springer-Verlag 1992.
[12] M. James, J. Baras, and R. Elliott, “Risk-sensitive control and dynamic games for
partially observed discrete-time nonlinear systems,” IEEE Transaction on Automatic
Control, vol. 39, no. 4, pp. 780–792, 1994.
[13] C. Charalambous and J. Hibey, “Minimum principle for partially observable non-
linear risk-sensitive control problems using measure-valued decompositions,” Stochas-
tics and Stochastics Reports, vol. 57, 1996.
[14] C. Charalambous, “The role of informations state and adjoint in relating nonlin-
ear output feedback risk-sensitive control and dynamic games,” IEEE Transactions
on Automatic Control, vol. 42, no. 8, pp. 1163–1170, 1997.
[15] P. D. Pra, L. Meneghini, and W. Runggaldier, “Some connections between
stochastic control and dynamic games.”Mathematics of Control, Signals, and Sys-
tems , vol. 9, pp. 303–326, 1996.
[16] V. A. Ugrinovskii and I. R. Petersen, “Finite Horizon Minimax Optimal Control
of Stochastic Partially Observed Time Varying Systems,” Mathematics of Control,
Signals and Systems, vol. 12, pp. 1–23, 1999.
32
[17] I. R. Petersen and M. R. James and P. Dupuis, “Minimax Optimal Control
of Stochastic Uncertain Systems with Relative Constraints,” IEEE Transactions on
Automatic Control, vol. 45, no. 3, pp. 398–412, 2000.
[18] C.D. Charalambous and F. Rezaei and S. Djouadi, ”Optimization of Stochastic
Uncertain Systems: Large Deviations and Robustness,” in Proceedings of the 42nd
IEEE Conference on Decision and Control, Hawaii, U.S.A., December 9-12, Vol.4,
pages 4249-4253, 2003.
[19] C.D. Charalambous and F. Rezaei, ”Characterization of the Optimal Disturbance
Attenuation for Nonlinear Stochastic Uncertain Systems,” in Proceedings of the 42nd
IEEE Conference on Decision and Control, Hawaii, U.S.A., December 9-12, Vol.4,
pages 4260-4264, 2003.
[20] C.D. Charalambous, F. Rezaei and A. Kyprianou, ”Relations Between Infor-
mation Theory, Robustness and Statistical Mechanics of Stochastic Systems,” in
Proceedings of the 43nd IEEE Conference on Decision and Control, pp.3479-3484,
Bahamas, December 14-17, 2004.
[21] N.U.Ahmed, Existence of Solutions of Nonlinear Stochastic Differential Inclu-
sions, Proc. world Congress of Nonlinear Analysis’ 92, (Ed. V Lakshmikantham),
(1992),1699-1712.
[22] N.U.Ahmed, Optimal Relaxed Controls for Nonlinear Infinite Dimensional Stochas-
tic Differential Inclusions, in Optimal Control of Differential Equations, Lec. Notes
in Pure and applied Mathematics, Marcel Dekker Inc, Vol.160,(1994), 1-19.
[23] N.U.Ahmed, Optimal Control of Impulsive Stochastic Evolution Inclusions, Dis-
cussiones Mathematicae, Differential Inclusions, Control and Optimization, 22(2002),155-
184.
33
[24] N.U.Ahmed, Optimal Relaxed Controls for Infinite Dimensional Stochastic Sys-
tems of Zakai Type, SIAM J. Control and Optim. 34 (5), (1996), 1592-1615.
[25] N.U.Ahmed and X.Xiang, Nonlinear Uncertain Systems and Necessary Condi-
tions of Optimality, SIAM J. Control and Optim. 35(5), (1997), 1755-1772.
34