1 Introduction and notation
Although our results are valid in more general filtrations, we start with a two-period model. In this setting, we work with a probability space equipped with three sigma-algebras,
\((\Omega ,{\mathcal{F}}_{0}\subseteq {\mathcal{F}}_{1}\subseteq { \mathcal{F}}_{2},{\mathbb{P}})\). The sigma-algebra
\({\mathcal{F}}_{0}\) is supposed to be trivial, i.e., every
\(A\in {\mathcal{F}}_{0}\) satisfies
\({\mathbb{P}}[A]=0\text{ or } 1\), whereas
\({\mathcal{F}}_{2}\) is supposed to express innovations with respect to
\({\mathcal{F}}_{1}\). Since we do not put topological properties on the set
\(\Omega \), we make precise definitions later that do not use conditional probability kernels. But essentially, we could say that we suppose that conditionally on
\({\mathcal{F}}_{1}\), the probability ℙ is atomless on
\({\mathcal{F}}_{2}\). We shall show that such a hypothesis implies that there is an atomless sigma-algebra
\({\mathcal{B}}\subseteq {\mathcal{F}}_{2}\) which is independent of
\({\mathcal{F}}_{1}\). The space
\(L^{\infty }({\mathcal{F}}_{i})\) is the space of bounded
\({\mathcal{F}}_{i}\)-measurable random variables modulo equality almost surely (a.s.). We say that two random variables
\(\xi ,\eta \) are
commonotonic1 if there are two nondecreasing functions
\(f,g\colon {\mathbb{R}}\rightarrow {\mathbb{R}}\) and a random variable
\(\zeta \) such that
\(\xi =f(\zeta ), \eta =g(\zeta )\). Commonotonicity can be seen as the opposite of diversification. If
\(\zeta \) increases, then both
\(\xi \) and
\(\eta \) increase (or, better, do not decrease). By the way, if
\(\xi \) and
\(\eta \) are commonotonic, then one can choose
\(\zeta =\xi +\eta \); see Delbaen [
7, Chap. 2.4]. It can be shown that in this case one can choose representatives — still denoted by
\((\xi ,\eta )\) — such that
\((\xi (\omega )-\xi (\omega '))(\eta (\omega )-\eta (\omega '))\ge 0\) for all
\(\omega ,\omega '\). Since we do not need this result, we do not include a proof. We say that a set
\(E\subseteq {\mathbb{R}}^{2}\) is commonotonic if
\((x,y),(x',y')\in E\) implies
\((x-x')(y-y')\ge 0\). In convex function theory, such sets are also called monotone or monotonic sets. Random variables
\(\xi ,\eta \) are commonotonic if and only if the support of the image measure of
\((\xi ,\eta )\) is a commonotonic set.
The present paper deals with time-consistent utility functions. This means that for
\(0\le i< j\le 2\), there are functions
\(u_{i,j}\colon L^{\infty }({\mathcal{F}}_{j})\rightarrow L^{\infty }({ \mathcal{F}}_{i})\) such that we have
\(u_{0,2}=u_{0,1}\circ u_{1,2}\). These utility functions satisfy the following properties; see [
7, Chap. 11] for more information on the relation between these properties:
1) \(u_{i,j}\colon L^{\infty }({\mathcal{F}}_{j})\rightarrow L^{\infty }({ \mathcal{F}}_{i})\), and if \(\xi \ge 0\), then also \(u_{i,j}(\xi )\ge 0\), and \(u_{i,j}(0)=0\).
2) For
\(\xi ,\eta \in L^{\infty }({\mathcal{F}}_{j})\) and
\(0\le \lambda \le 1\) and
\({\mathcal{F}}_{i}\)-measurable, we have
$$ u_{i,j}\big(\lambda \xi +(1-\lambda )\eta \big)\ge \lambda u_{i,j}( \xi )+(1-\lambda ) u_{i,j}(\eta ). $$
3) Since commonotonicity implies (as easily seen) positive homogeneity, we use a stronger property and suppose
coherence. For
\(\xi \in L^{\infty }({\mathcal{F}}_{j})\) and
\(\lambda \geq 0\) and
\({\mathcal{F}}_{i}\)-measurable, we have
$$ u_{i,j}(\lambda \xi )=\lambda u_{i,j}(\xi ). $$
4) For
\(\xi \in L^{\infty }({\mathcal{F}}_{j})\) and
\(a\in L^{\infty }({\mathcal{F}}_{i})\), we have
$$ u_{i,j}(\xi +a)=u_{i,j}(\xi ) + a. $$
5) We need Lebesgue-continuity which means that if \((\xi _{n}) \subseteq L^{\infty }({\mathcal{F}}_{j})\) is a uniformly bounded sequence such that \(\xi _{n}\rightarrow \eta \) in probability, then \(u_{i,j}(\xi _{n})\) tends to \(u_{i,j}(\eta )\) in probability.
6) The Lebesgue property is stronger than the Fatou property which says that for a sequence \((\xi _{n}) \subseteq L^{\infty }\) such that a.s. \(\xi _{n}\downarrow \eta \in L^{\infty }\), we have \(u_{ij}(\xi _{n})\rightarrow u_{ij}(\eta )\) a.s.
The utility functions we need are coherent and hence we can use their dual representation; see Delbaen [
6, end of the proof of Theorem 6]. This means that there is a uniquely defined convex closed set
\({\mathcal{S}}\subseteq L^{1}\) of probability measures, absolutely continuous with respect to ℙ, such that
$$ u_{0,2}(\xi )=\inf _{{\mathbb{Q}}\in {\mathcal{S}}}{\mathbb{E}}_{ \mathbb{Q}}[\xi ]. $$
The set
\({\mathcal{S}}\) is viewed as a subset of
\(L^{1}\) via the Radon–Nikodým theorem. The Lebesgue-continuity is equivalent to the weak compactness of
\({\mathcal{S}}\). We suppose that our utility functions are
relevant, i.e., for each
\(A\) with
\({\mathbb{P}}[A]>0\), we have
\(u(-\mathbf {1}_{A})<0\); see [
7, Chap. 4.14]. By the Halmos–Savage theorem, this means that
\({\mathcal{S}}\) contains an equivalent probability measure. We need this property in order to avoid some problems with negligible sets appearing in the definition and with comparisons of conditional expectations.
Without further notice, we always assume that our utility functions are relevant and Lebesgue-continuous. These assumptions are not always needed; sometimes Fatou-continuity is sufficient. Since we want to put more emphasis on the methods of proof, we do not aim for the most general results.
One may ask in which way the utility functions
\(u_{i,j}\) can be constructed from the utility function
\(u_{0,2}\). The construction is easier when
\(u_{0,2}\) is relevant. The Fatou or Lebesgue property is less important for this development. As shown in [
7, Chap. 11], there is a way to check whether the utility function
\(u_{0,2}\) can be embedded in a time-consistent family of utility functions. To do this, we introduce the acceptability cones
$$\begin{aligned} {\mathcal{A}}_{0,2} &=\{\xi \in L^{\infty }({\mathcal{F}}_{2}) \colon u_{0,2}(\xi )\ge 0\}, \\ {\mathcal{A}}_{0,1}&=\{\xi \in L^{\infty }({\mathcal{F}}_{1})\colon u_{0,2}( \xi )\ge 0\}, \\ {\mathcal{A}}_{1,2}&=\{\xi \in L^{\infty }({\mathcal{F}}_{2})\colon \text{for all } A\in {\mathcal{F}}_{1} , u_{0,2}(\xi \mathbf {1}_{A})\ge 0 \}. \end{aligned}$$
The necessary and sufficient condition for the existence of a time-consistent extension is
\({\mathcal{A}}_{0,2}={\mathcal{A}}_{0,1}+{\mathcal{A}}_{1,2}\). If this is fulfilled, we put
$$ u_{1,2}(\xi )=\mathop{\mathrm{ess\,inf}}\{\eta \in L^{\infty }({ \mathcal{F}}_{1})\colon \xi -\eta \in {\mathcal{A}}_{1,2}\}, $$
and
\(u_{0,1}\) is simply the restriction of
\(u_{0,2}\) to
\(L^{\infty }({\mathcal{F}}_{1})\). This gives sense to expressions such as “
\(u_{0,2}\) is time-consistent”.
Already in the case where the utility functions are expected value and conditional expectations, the main theorem leads to the following result. (The notion “conditionally atomless” will be explained and analysed in the next section.)
Both concepts, time-consistency and commonotonicity, are important in the theory of risk evaluation. The concept of time-consistency (and -inconsistency) was introduced and investigated by Koopmans [
12]. The role of commonotonicity found its way into insurance and is present in several papers. The use of Choquet integration as premium principle was emphasised by Denneberg [
9] who was inspired by the pioneering work of Yaari [
21]. Schmeidler proved the relation between commonotonic principles, convex games and Choquet integration [
14]. Modern uses can be found for instance in Wang et al. [
17] and Wang [
18]. For more references and different proofs of these results, we refer to [
7, Chap. 7]. Although commonotonicity seems to be a desirable property, there might be some difficulties when insurance contracts are priced in this way; see Castagnoli et al. [
5] for some unexpected consequences.
The concept of risk measures (up to sign changes monetary utility functions) was introduced in Artzner et al. [
1,
2].
Using the general version of Theorem
1.1, we shall show that except in very restrictive cases, a utility function
\(u_{0,2}\) cannot be time-consistent and commonotonic at the same time. It seems that time-consistency is a strong property that excludes some other
desirable properties. For instance in Kupper and Schachermayer [
11], it is shown that in a filtration with innovations (comparable to the requirement of being conditionally atomless), utility functions that are time-consistent and law-determined are necessarily of entropic type. We refer to [
11] for the details and the precise form of the innovations. The present paper studies time-consistent utility functions that might depend on past history and are not necessarily law-determined. The methods we use are different from the approaches used for law-determined or law-invariant utility functions. Among the many papers on these utility functions, we could refer the reader to the cited papers and to e.g. Bellini et al. [
3], Bellini et al. [
4], Wang and Ziegel [
19], Weber [
20] and Ziegel [
22].
2 Atomless extension of sigma-algebras
In this section, we work with a probability space \((\Omega ,{\mathcal{F}}_{2},{\mathbb{P}})\) equipped with the filtration \({\mathcal{F}}_{0}\subseteq {\mathcal{F}}_{1}\subseteq { \mathcal{F}}_{2}\).
If the conditional expectation can be calculated with a – under extra topological conditions – regular probability kernel, say
\(K(\omega , A)\), then the above definition is a measure-theoretic way of saying that the probability measure
\(K(\omega , \cdot )\) is atomless for almost every
\(\omega \in \Omega \). The precise relation between these two notions is not the topic of this paper. See Delbaen [
8] for the details.
The main result of this section is the following.
The “if” part is easy, but requires some continuity argument. Because ℬ is atomless, there is a ℬ-measurable random variable
\(U\) uniformly distributed on
\([0,1]\). The sets
\(B_{t}=\{U\le t\}, 0\le t \le 1\), form an increasing family of sets with
\({\mathbb{P}}[B_{t}]=t\). Fix
\(A\in {\mathcal{F}}_{2}\) and let
\(F=\{ 0 < {\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}]\}\). We may suppose that
\({\mathbb{P}}[F]>0\) since otherwise there is nothing to prove. We now show that there is
\(t\in (0,1)\) with
\({\mathbb{P}}[ 0 < {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, { \mathcal{F}}_{1}]< {\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}] ] > 0\). According to Theorem
2.2,
\({\mathcal{F}}_{2}\) is atomless conditionally to
\({\mathcal{F}}_{1}\). Obviously for
\(0\le s\le t \le 1\), we have by independence of ℬ and
\({\mathcal{F}}_{1}\) that
$$ \Vert {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}] - { \mathbb{E}}[\mathbf {1}_{A\cap B_{s}} \, | \, {\mathcal{F}}_{1}]\Vert _{\infty }\le \Vert {\mathbb{E}}[\mathbf {1}_{B_{t}\setminus B_{s}} \, | \, { \mathcal{F}}_{1}]\Vert _{\infty }= t-s. $$
It follows that there is a set of measure 1, say
\(\Omega '\), such that for all
\(s\le t\),
\(s,t\) rational, and all
\(\omega \in \Omega '\),
\({\mathbb{E}}[\mathbf {1}_{A\cap B_{t}}\, | \, {\mathcal{F}}_{1}](\omega )\) can be taken to satisfy
$$ | {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}](\omega ) - {\mathbb{E}}[\mathbf {1}_{A\cap B_{s}} \, | \, {\mathcal{F}}_{1}](\omega ) | \le t-s. $$
For each
\(\omega \in \Omega '\), we can extend the function
$$ [0,1] \cap {\mathbb{Q}}\ni q \mapsto {\mathbb{E}}[\mathbf {1}_{A\cap B_{q}}\, | \, {\mathcal{F}}_{1}](\omega ) $$
to a continuous function on
\([0,1]\). The resulting continuous extension then represents the equivalence classes of random variables
\(({\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}])_{t \in [0,1]}\). For
\(t=0\), we have zero, and for
\(t=1\), we find
\({\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}]\). Because the trajectories are continuous for
\(\omega \in \Omega '\), a simple application of Fubini’s theorem shows that the real valued function
$$ t\mapsto {\mathbb{P}}\left [ 0 < {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}]< {\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}] \right ] $$
becomes strictly positive for some
\(t\). With some extra work – done later –, one can even show that there is
\(G\subseteq A\) such that
\({\mathbb{E}}[\mathbf {1}_{G}\, | \, {\mathcal{F}}_{1}]= (1/2){\mathbb{E}}[ \mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}]\).
For completeness, let us now give the details of the application of Fubini’s theorem. Suppose to the contrary that for all
\(t\in [0,1]\), we have
$$ {\mathbb{P}}\left [ 0 < {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, { \mathcal{F}}_{1}]< {\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}] \right ] =0. $$
Then on the product space
\([0,1]\times \Omega '\), we find that the (clearly measurable) set
$$ \{(t,\omega )\colon 0 < {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, { \mathcal{F}}_{1}](\omega )< {\mathbb{E}}[\mathbf {1}_{A}\, | \, { \mathcal{F}}_{1}] (\omega )\} $$
has
\((m\times {\mathbb{P}})\)-measure zero (
\(m\) denotes Lebesgue measure). By Fubini’s theorem, we have that for almost all
\(\omega \in \Omega '\), the set
$$ \{t \colon 0 < {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}]( \omega )< {\mathbb{E}}[\mathbf {1}_{A}\, | \, {\mathcal{F}}_{1}] (\omega ) \} $$
must have Lebesgue measure zero. However, for
\(\omega \in \Omega '\), this contradicts the continuity of the mapping
$$ t\mapsto {\mathbb{E}}[\mathbf {1}_{A\cap B_{t}} \, | \, {\mathcal{F}}_{1}]( \omega ). $$
The proof of the “only if” part is broken down into several steps stated in the lemmas that follow. Without further notice, we always suppose that \({\mathcal{F}}_{2}\) is atomless conditionally to \({\mathcal{F}}_{1}\).
The following proposition is Lemma
2.5 where we take
\(C=\Omega \). For didactic reasons, we give another proof that directly uses the existence of an independent sigma-algebra. We use the same assumptions and notations as in Theorem
2.3.
They then prove the following result.
There are several differences with our approach. There is the technical difference that [
16] suppose the existence of a continuously distributed random variable
\(X\). In doing so, they avoid the technical points between the more conceptual definition using conditional expectations and the construction of a suitable sigma-algebra with a uniformly distributed random variable. A further difference is that they use a dominating measure that later can be taken as the mean of
\(({\mathbb{Q}}_{1},\ldots ,{\mathbb{Q}}_{n})\). Of course, their result together with the results here show that the definition of
\(({\mathbb{Q}}_{1},\ldots ,{\mathbb{Q}}_{n})\) being conditionally atomless is equivalent to the statement that for the measure
\({\mathbb{Q}}_{0}=\frac{1}{n}({\mathbb{Q}}_{1}+\cdots +{\mathbb{Q}}_{n})\), the sigma-algebra
\({\mathcal{A}}\) is atomless conditionally to the sigma-algebra generated by the Radon–Nikodým derivatives
\((\frac{d{\mathbb{Q}}_{k}}{d{\mathbb{Q}}_{0}})_{k=1,\dots ,n}\). In [
16], it is also shown that one can take any strictly positive convex combination of the measures
\(({\mathbb{Q}}_{1},\ldots ,{\mathbb{Q}}_{n})\). Below we show that the sigma-algebra
\({\mathcal{A}}\) in some sense has a minimality property, a result that clarifies the relation between the two approaches. Before doing so, let us recall two easy results from introductory probability theory.
From Proposition
2.16, it follows that the sigma-algebra augmented with the class
\({\mathcal{N}}\) is the same for all strictly positive convex combinations. This shows that in the definition of atomless conditionally to ℱ, we can also add the nullsets
\({\mathcal{N}}\) to ℱ. To check that
\({\mathcal{A}}\) is atomless conditionally to a sigma-algebra ℱ, it is clear that the smaller ℱ, the easier it is to satisfy the condition. In our opinion, the above clarifies the relation between this paper and [
16].
4 Some special commonotonic set
In this section, we define a special norm on
\({\mathbb{R}}^{2}\). Part of its unit sphere will then be used as a commonotonic set. The reader could make some drawings to help visualise the constructions. The construction is done in several steps. The first step consists in taking the curve obtained as the concatenation of the convex intervals that join the points
$$ (-4,-4)\rightarrow (-4,-2)\rightarrow (0,0)\rightarrow (4,2) \rightarrow (4,4). $$
The convex hull of this set is a parallelogram
\(P_{0}\), with parallel vertical sides given by the line segments
$$ (-4,-4)\rightarrow (-4,-2)\qquad \text{and}\qquad (4,2) \rightarrow (4,4). $$
The set
\(P_{0}\) will be used as the unit ball of a norm on
\({\mathbb{R}}^{2}\). More precisely, we use the Minkowski functional
$$ \Vert (x,y)\Vert := \inf \{ \alpha >0: (x,y)\in \alpha P_{0}\}. $$
Note that every point of
\(P_{0}\) is the convex combination of points taken on the vertical sides. An easy and continuous way to obtain such convex combination goes as follows. Through a point in
\(P_{0}\), take a line parallel to the “skew” sides of
\(P_{0}\) and see where it intersects the vertical sides. Elementary calculations give us that for
\((x,y)\in P_{0}\), we may write
\((x,y)=(1-\lambda _{0})(u^{0}_{1},u^{0}_{2})+\lambda _{0}(v^{0}_{1},v^{0}_{2})\) with
\(u^{0}, v^{0} \in P\) and
\(0 \leq \lambda _{0} \leq 1\), or more explicitly
$$ (x,y)=\frac{4-x}{8}\left (-4,y-3-\frac{3x}{4}\right )+\frac{4+x}{8} \left (4,y+3-\frac{3x}{4}\right ). $$
For each
\(n\in {\mathbb{Z}}\), we now define
\(P_{n}=2^{n}P_{0}\) and similarly as for
\(n=0\), we define
\(\lambda _{n}\),
\((u^{n}_{1},u^{n}_{2})\),
\((v^{n}_{1},v^{n}_{2})\). These functions are obviously continuous. The set
\(E\) consists of all the vertical segments with the origin added. It forms a commonotonic set. This follows from the equality
$$ E=\{(0,0)\}\cup \bigcup _{n\in {\mathbb{Z}}}\Big( 2^{n}\big( [(-4,-4),(-4,-2)] \cup [(2,4),(4,4)] \big) \Big). $$
We now construct functions
\(\Lambda , U, V\) on
\({\mathbb{R}}^{2}\) as follows. For
\((x,y)\in P_{n}\setminus P_{n-1}\), we define
\(\Lambda (x,y)= \lambda _{n}(x,y)\),
\(U(x,y)=u^{n}(x,y)\),
\(V(x,y)=v^{n}(x,y)\). At
\((0,0)\), we put
\(\Lambda (0,0)=1\),
\(U(0,0)=(0,0)=V(0,0)\). These functions are no longer continuous, but are certainly Borel-measurable. They satisfy the following properties:
1) \(\Lambda \colon {\mathbb{R}}^{2}\rightarrow [0,1]\).
2) \(U \colon {\mathbb{R}}^{2}\rightarrow E\), \(V\colon {\mathbb{R}}^{2}\rightarrow E\).
3) We have \(\Vert U(x,y)\Vert \le 2 \Vert (x,y)\Vert \) and \(\Vert V(x,y)\Vert \le 2 \Vert (x,y)\Vert \). Indeed, for \((x,y)\in P_{n}\setminus P_{n-1}\), we have \(2^{n}=\Vert U(x,y)\Vert \ge \Vert (x,y)\Vert \ge 2^{n-1}\), and the same holds for \(V\).
4) For all \((x,y)\in {\mathbb{R}}^{2}\), \((x,y)=(1-\Lambda (x,y))U(x,y)+\Lambda (x,y)V(x,y)\).
5) The coordinates \(V_{1}(x,y)-U_{1}(x,y)\) and \(V_{2}(x,y)-U_{2}(x,y)\) of \(V-U\) are nonnegative.
5 The main result
We start by giving an extension of the usual definition of conditional expectation.
The reader can check that the existence and definition of an extended conditional expectation are independent of the choice of the \({\mathcal{F}}_{1}\)-measurable partition. We sometimes drop the word “extended”.
Again we suppose that \({\mathcal{F}}_{2}\) is atomless conditionally to \({\mathcal{F}}_{1}\). The utility function \(u_{1,2}\) is Lebesgue-continuous.
Before giving the main result of the paper, we first prove a special case.
The next theorem is an improvement of the preceding result in the sense that we replace the conditional expectation by a more general utility function. The proof follows the same lines.
6 Commonotonicity and time-consistency
In this section, we use the same hypothesis on the filtration \(({\mathcal{F}}_{0},{\mathcal{F}}_{1},{\mathcal{F}}_{2})\). In particular, we suppose that \({\mathcal{F}}_{2}\) is atomless conditionally to \({\mathcal{F}}_{1}\). We start with a monetary coherent utility function \(u_{0,2}\colon L^{\infty }({\mathcal{F}}_{2})\rightarrow {\mathbb{R}}\). We suppose – as in the rest of the paper – that \(u_{0,2}\) is relevant.
7 A continuous-time result
In this section, we use a filtration indexed by the time interval \([0,T]\). This filtration \(\left ({\mathcal{F}}_{t}\right )_{0\le t\le T}\) does not necessarily fulfil the usual assumptions. The only assumption is that \({\mathcal{F}}_{T}\) is generated by \(\bigcup _{0\le t< T}{\mathcal{F}}_{t}\). We also suppose that we are given a family \(u_{t,s},0\le t\le s\le T\), \(u_{t,s}\colon L^{\infty }({\mathcal{F}}_{s})\rightarrow L^{\infty }({ \mathcal{F}}_{t})\), of coherent utility functions. We assume the following time-consistency: for \(t\le s\le v\), we have \(u_{t,v}=u_{t,s}\circ u_{s,v}\).
Publisher’s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.