1. Introduction
In (Delbaen, 2009) , the problem of defining a risk measure on a solid, rearrangement invariant subspace of
-space of random variables with respect to some atomless probability space
. We recall that a vector space E, being a vector subspace of
is called rearrangement invariant if for random rariables
, which have the same distribution,
implies
. Also, the space E is solid if for andom viariables
,
, implies
. In (Delbaen, 2009) , there is an extensive treatment of this problem, related to the role of the spaces
and
, compared to E, especially in (Delbaen, 2009) . On the other hand, the whole paper (Delbaen, 2002) is devoted to the difficulties of defining coherent risk measures on subspaces of
, while it is proved that if the probability space is atomless, no coherent risk measure is defined all over
(Delbaen, 2002) . Of course these attempts of moving from
to appropriately defined subspaces of
, are related to the tail propertes of the random variables in actuarial science and finance and more specifically to heavy-tailed distributed random variables. The actual problem behind these seminal article by F. Delbaen is since we cannot define a coherent risk measure on the entire
, whether subspaces of
which are both alike
and preserve nice distributional properties (from the aspect of heavy-tails). Especially, we treat the rearrangement invariance in the sense of remaining in the same class of distributions and not by requiring distributional invariance. This is the topic of our paper.
2. Ideals of L0 and Heavy-Tailed Distributions
It is well-known that since
is a Riesz space, being ordered by the pointwise-
-a.e. partial ordering ³, it would be taken as a Riesz subspace of
. Hence, it may be considered to be an order-complete Riesz space. Let us take an element y of
, which corresponds to a heavy-tailed random variable. This indicates that either for
for
,
![]()
for any real number
, where
or
. Heavy-tailed random variables may not have even a finite moment
. On the other hand, according to (Aliprantis & Border, 1999) , the principal ideal
generated by y in E, endowed by the norm
![]()
is an AM-space with order unit
. We also have to mention the following relevant.
Lemma 2.1 If
and y is a heavy-tailed random variable, then every
is a heavy-tailed random variable.
Proof. Since
, we get that for the sets
, the inclusion
holds, which implies
for the corresponding cumulative distri- bution functions. Since for the integral
![]()
holds for any
, this implies
![]()
for any
.
We recall the class
of dominated variation distributions:
![]()
This class is a sub-class of heavy -tailed distributions, see (Cai & Tang, 2004) .
Theorem 2.2 If
, where
denotes the class of dominated variation distributions respectively, then for every
,
.
Proof. According to what is proved in (Cai & Tang, 2004) , the class
is convolution-closed, namely if
, then
. First, we have to prove that if
, then
, for any
. Since
, there exists some
such that
. But
. This is easy to prove, since if
![]()
for any
, then in order to prove that
![]()
for any
, then we get that the above limsup is equal to
![]()
for any
. Hence,
. Moreover, we have to prove that if
, then
. From the previous Lemma,
![]()
From the properties of the tail function of z we also have that since
for
, then
![]()
Hence,
![]()
Since
,
![]()
which is the desired conclusion.
Hence we obtain subspaces E of
, which are actually the ideals
which satisfy the rearrangement invariance property, while they contain non-integrable distributions, in the sense that for any
there is a maximum p for which the moment
exists in
. Let us discuss more this question. A notion which is very important is the one of the moment index. We recall that the moment index for a non-negative random variable x is equal to
![]()
We also recall that if
, then
, see in (Seneta, 1976) , (Tang & Tsitsiashvili, 2003) . The use of the moment index in the specific case is that despite the validity of the (Delbaen, 2002) , due to the fact that the elements of
distributions lie in the class
, we assure that at least in the ideal
, we assure a general level of non-integrability of
, given by a finite
. About the question whether the class
is the greatest in which the specific Theorem holds, we have to mention that if we move up to the class of the subexponential distributions, it is not convolution-closed, see for example in (Leslie, 1989) . As it is also well- known from (Aliprantis & Border, 1999) , the dual space
of
is an AL-space, since the ideal
is an AM-space with unit
, as mentioned above. Hence, we keep the dual pair
![]()
for any of the y described above.
3. Expected-Shortfall on Ideals of L0
Taking any
whose
, and defining the corresponding dual pair
![]()
we may define an Expected Shortfall-form risk measure on
. We have to notice that
satisfies both the order and the distributional rearrangement property, as a subspace of
. This is due to the properties of the class
of the dominated variation distributions. Hence we use (Kaina & Rüschendorf, 2009) of the dual (robust) representation of the usual Expected Shortfall in order to prove the following.
Theorem 3.1 The functional
, where
![]()
is an
-coherent risk measure, where
is such that
.
Proof.
1)
![]()
for any
, due to the order completeness of the ideal
(y-Translation Invariance).
2)
![]()
for any
(Subadditivity).
3)
![]()
for any
and
(Positive Homogeneity).
4) If
then for any
we get
. Hence by taking suprema all over
, we ger
(
-Monotonicity).
Finally, if we suppose that the dual pair
is a symmentric Riesz pair, or else that
has order-
continuous norm (see also (Aliprantis & Border, 1999) ), then the values of R are finite since they represent the supremum value of a weak-star continuous linear functional on a weak-star compact set, which is the box of
functionals
. Otherwise, the infinity of the values of R may be excused by the presence of heavy-tailed distributions.