With the explicit introduction of uncertainties and risks the overall performance of a de-cision
x
becomes a tradeo between dierent socioeconomic and environmental indicators (costs, benets, incomes, damages) and indicators of risks. The classical example is the mean-variance ecient strategies providing a tradeo between expected returns and the variance. Unfortunately, the concept of the mean-variance ecient strategies may be mis-leading and even wrong for nonnormal probability distributions (especially for catastrophic risks) which require more sophisticated risk indicators and corresponding concepts of ro-bust strategies. More precisely, in practice a given decisionx
results in dierent outcomesf
(x;!
) = (f
1(x;!
);:::;f
m(x;!
)) aected by some uncertain (random) variables!
. For-mally, the overall performance ofx
can be often summarized in the form of an expectation functionU
(x
) =E u
(f
1(x;!
);:::;f
m(x;!
));
where
u
() is a \utility" function dened onf
2R
m. The mean-variance ecient solutions maximizingE f
(x;!
);N E
[f
(x;!
);E f
(x;!
)]2,N >
0, can also be obtained from the maximization of the following type of function:maxx;y
E
hf
(x;!
);N
(f
(x;!
);y
)2i:
This representation convexies the problem for
f
(x;!
) = ;jf
(x;!
)j, where jf
(;!
)jis a convex (cost) function.Traditionally the utility function is assumed to be continuous and dierentiable. It is easy to see that all risk functions discussed in this section can be represented in the same form but with nonsmooth and even discontinuous utility functions. For example, if
u
() is the indicator function for the eventff
2R
mjf
c
g, thenU
(x
) =P
ff
(x;!
)c
g:
(17) Ifu
(f
1;f
2) =f
1I
ff
20g=( 0
; f
2<
0; f
1; f
20;
then we obtain function (2)U
(x
) =Zf2
(x;!)0
f
1(x;!
)P
(d!
):
(18) In the particular casef
1(x;!
)f
2(x;!
) =f
(x;!
)U
(x
) =E
maxf0;f
(x;!
)P
(d!
):
Functions
U
(x
) with nonsmooth and discontinuous integrandu
() can be used as a unied concept to analyze quite dierent risk management problems. In short, we can call suchU
(x
) the risk functions andu
(f
) the sample risk function or (extended) utility function. We can callU
(x
) also extended expected utility function. Note that although in-dicators (13), (14), (15) are dened through stopping time(x
), they can also be expressed in the formE u
(R
0;R
1(x
);:::;R
T(x
)) with some discontinuous functionu
().3 Risk functions
Consider the following risk function given in the form of extended expected utility
U
(x
) =E u
(f
(x;!
));
(19) wheref
:R
n ;!R
m is a continuous inx
and measurable in!
vector function,u
:R
m ;!R
1 is a Borel (extended utility) function,E
(orE
!) denotes mathematical expectation over measureP
(orP
!) on . In general, as we discussed in Section 2, functionu
() may be discontinuous on a setD
R
m.Proposition 3.1
(Continuity of risk function). Assume that (i)f
(x;!
) is a.s. continuous at pointx
,(ii)
P
ff
(x;!
)2D
g= 0,(iii)
u
(f
(y;!
))M
(!
) for ally
from a vicinity ofx
with integrable functionM
(!
).Then function
U
(x
) is continuous.The proposition follows from Lebesgue's dominance convergence theorem.
Denote
D
=fy
2R
mjdist
(y;D
)g; dist
(y;D
) = infz2Dk
y
;z
k: Proposition 3.2
(Lipschitz continuity). Assume that(i)
u
() is uniformly Lipschitzian in any ball outside the discontinuity setD
; (ii)f
(x;!
) are a.s. Lipschitzian inx
2X
uniformly in!
;(iii)
P
ff
(x;!
)2D
gC
for allx
2X
,y
2R
m,>
0 and some constantC
; (iv)u
(f
(x;!
) +y
)M
for allx
2X
,y
2R
m and some constantM
.Then function
U
(x;y
) =E u
(f
(x;!
) +y
) is Lipschitz continuous in (x;y
)2X
R
m, and hence risk functionU
(x
) =U
(x;
0) is Lipschitzian inx
2X
.Proof.
LetL
uandL
f be Lipschitz constants foru
andf
, respectively. For givenx
1,x
2,y
1,y
2denex
=x
1+(x
2;x
1),y
=y
1+(y
2;y
1) with2[0;
1],=L
fkx
2;x
1k+ky
2;y
1k, 2 = f!
2 j (f
(x
1;!
) +y
1) 2D
2g, Obviously, kf
(x
;!
) +y
;f
(x
1;!
);y
1kL
fkx
2;x
1k+ky
2;y
1k. Note that if (f
(x
1;!
)+y
1)2D
2, then (f
(x
2;!
)+y
2)2D
3, and if (f
(x
1;!
)+y
1)2D
2, then (f
(x
;!
)+y
)2D
for any 2[0;
1]. We haveU
(x
2;y
2);U
(x
1;y
1) = R2+Rn2[u
(f
(x
2;!
) +y
2);
u
(f
(x
1;!
) +y
1)]P
(d!
)
M P
f(f
(x
2;!
) +y
2)2D
3g+
M P
f(f
(x
1;!
) +y
1)2D
2g+Rn2
L
ukf
(x
2;!
) +y
2;f
(x
1;!
);y
1kP
(d!
)(5
MC
+L
u)(L
fkx
2;x
1k+ky
2;y
1k):
2If function
u
() is discontinuous then it can be approximated in dierent ways by continuous functionsu
() for some parameter in such a way thatu
(y
) ;!u
(y
) as ;!0 for ally
2D
. Then functionU
(x
) is approximated by functionsU
(x
) =E u
(f
(x;!
)):
(20)Proposition 3.3
(Convergence of approximations). Assume that (i)lim!0P
ff
(x;!
)2D
g= 0, pointwise (uniformly) inx
2X
; (ii)lim!0u
(z
) =u
(z
), uniformly inz
2D
for any>
0;(iii)
u
(f
(x;!
)) andu
(f
(x;!
)) are bounded by an integrable in square functionM
(!
) uniformly inx
2X
and>
0.Thenlim!0
U
(x
) =U
(x
) pointwise (uniformly) inx
2X
.Proof.
Dene 1 =f!
2jf
(x;!
)2D
g and 2= n1. Then The rst term on the right-hand side of (21) can be made arbitrarily small by choosing small enough due to (i), (iii). For a given the second term on the right-hand side of (21) can be made arbitrary small by choosing small enough due to (ii).2One way to construct approximations
U
(x
) is to consider stochastically disturbed performance indicatorsf
(x;!;
) =f
(x;!
) +;
where
is a small positive parameter, 2R
m is a random vector independent of!
with densityK
(). The corresponding disturbed risk function takes the formU
(x
) =E
E
!u
(f
(x;!;
))=
E
!E
u
(f
(x;!
)+)=
E
!u
(f
(x;!
));
where
u
(f
) is the so-called smoothed (or mollied) utility functionu
(y
) =E
u
(y
+) = 1mZ
u
(z
)K
z
;y
dz
used in kernel density estimation (see, for example, [7]), in probability function optimiza-tion (see [22], [27]) and in nonsmooth optimizaoptimiza-tion (see [25], [17] and references therein).
Proposition 3.4
(Convergence of mollied utilities at continuity points). Letu
(x
) be a real-valued Borel measurable function onR
m,K
(x
) be a bounded, integrable, real valued density function onR
m and one of the following holds(i)
u
() is bounded onR
m; (ii)K
() has a compact support;(iii)k
y
kK
(y
);!0 as
ky
k;!1, where kk denotes the Euclidean norm onR
m. Thenu
(y
);!u
(y
) as ;!0 at any continuity point ofu
().The statement of the proposition under assumption (i) can be found in [3], and under (ii), (iii) it is available in [6].
Proposition 3.5
(Uniform convergence outside discontinuity points ). Assume that (i)u
() is a Borel function with closed setD
of discontinuity points;(ii) density
K
() has a compact support.where
S
(K
) = fz
jK
(Z
)>
0g denotes support of densityK
(). SinceD
is closed andExample 3.1
(Partial smoothing). If in (18) we disturb only functionf
2 thenU
(x
) =E
E
!f
1(x;!
)I
f2(x;!)+0=
E
!f
1(x;!
)E
I
f2(x;!)+0=
E
!f
1(x;!
)(1;F(;f
2(x;!
)=
));
where F is a cumulative distribution function of random variable.Proposition 3.6
(Uniform convergence under partial smoothing). Assume that con-ditions of Proposition 3.5 are fullled and(i) function
E
jf
1(x;!
)jis bounded onX
;(ii)
P
fjf
2(x;!
)jg;!0 as ;!0 uniformly inx
2X
. ThenU
(x
) converges toU
(x
) uniformly inx
2X
.Proof.
For arbitrary numbersC
, estimate the dierencej
U
(x
);U
(x
)jE
jf
1(x;!
)jj1;F(;f
2(x;!
)=
);I
f2(x;!)0j The rst term on the right-hand side of (22) is made arbitrarily small by takingC
su-ciently large by (i). The second term for givenC
is made small by taking suciently small by (ii). GivenC
and the third term can be made small by taking small by Proposition 3.5. 2Example 3.2
(Smoothing probability function ). Consider probability functionU
(x
) =P
!ff
1(x;!
)0;:::;f
m(x;!
)0gWe can also approximate by using
i =,i
= 1;:::;m
, where random variable has the cumulative distribution function F. ThenU
(x
) =P
!P
f;f
1(x;!
)=;:::;
;f
m(x;!
)=
g=
P
!P
f;1 max1imf
i(x;!
)g=
E
!F
;
1max1im
f
i(x;!
):
If functions
u
andf
(x;!
) in (20) are continuously (or generalized) dierentiable, then compound functionu
(f
(x;!
)) is also continuously (generalized) dierentiable with (sub)dierential@
xu
(f
(x;!
)), which can be calculated by a chain rule (see [16], [26] for the nondierentiable case).If (sub)dierential
@
xu
(f
(x;!
)) is majorized by an integrable (Lipschitz) constant, (x;!
) is a measurable selection of@
xu
(f
(x;!
)), then functionF
(x
) is also (generalized) dierentiable with (sub)dierential@U
(x
) =E @
xu
(f
(x;!
))3E
(x;!
):
(23) For optimization ofF
(x
) one can apply specic stochastic gradient methods (see Section 6) based on samples of (x;!
) with ;! 0. For a given it is also possible to use the sample mean optimization methods.4 Stochastic smoothing of risk processes
To optimize risk functions we can apply molliers [17]) over decision variables
x
. Similarly, we can mollify risk process over some parameters, for example, initial state. In addition to smoothing eects, which are usually weaker than in the rst case, the signicant advantage of the parametric smoothing is the possibility to obtain fast statistical estimators of the risk functions and their derivatives [12].Beside standard risk process (11) consider a process with random initial capital
R
0+ [12]:Q
t(x;
) =R
0++ t(x
);C
t(x
) =R
t(x
) +;
0t
T;
(24) where is an independent of all claimsC
t(x
) one-dimensional random variable with a continuously dierentiable distribution functionF(
y
) =P
f< y
g;
is a small (smoothing) parameter (!0).We can think of (24) as risk process (11) with disturbed initial values
R
0 orR
1(x
).Through dynamic equation (24) the disturbance
is transferred to further valuesR
t(x
),t
1, of the process. Similarly we can independently disturb allR
t(x
), 0t
T
, and interpret these disturbances as the presence of insignicant lines of business of the insurance company.In subsection 2.3 we introduced important performance functions of process (11): prob-ability of insolvency T(
x
), partial expected protF
T(x
), expected shortfallH
(x
). Under assumption P(i) they are continuous, and under P(ii) they are Lipschitz continuous. Here we consider the same performance functions also for the disturbed process (24). Under assumption P(ii) by the results of section 3 (Propositions 3.3, 3.5) these approximates con-verge uniformly inx
to the original undisturbed performance functions as the disturbance goes to zero. The smoothing eects enable us to derive their subdierentials.4.1 The probability of ruin
Dene measure
P
as the product ofP
! andP
,P
=P
P
!. Then the probability of ruin till momentT
of the disturbed risk processfQ
t(x;
) =R
t(x
)+; t
= 0;
1;:::;T
gisT(
x;
) = 1;P
fQ
t(x;
)0;
0t
T
g= 1;
P
f;R
t(x
)=;
0t
T
g= 1;
P
fmax0tT;R
t(x
)=
g=
P
f<
;min0tTR
t(x
)=
g=