International Journal of Approximate Reasoning 130 (2021) 226–245

Contents lists available at ScienceDirect

International Journal of Approximate Reasoning

www.elsevier.com/locate/ijar

On nonlinear expectations and Markov chains under model

uncertainty ✩
Max Nendel
Center for Mathematical Economics, Bielefeld University, 33615 Bielefeld, Germany

a r t i c l e i n f o a b s t r a c t

Article history: The aim of this work is to give an overview on nonlinear expectations and to relate
Received 28 November 2019 them to other concepts that describe model uncertainty or imprecision in a probabilistic
Received in revised form 21 September framework. We discuss imprecise versions of stochastic processes with a particular interest
2020
in imprecise Markov chains. First, we focus on basic properties and representations
Accepted 14 December 2020
Available online 24 December 2020
of nonlinear expectations with additional structural assumptions such as translation
invariance or convexity. In a second step, we discuss how stochastic processes under
Keywords: nonlinear expectations can be constructed via primal and dual representations. We
Nonlinear expectation illustrate the concepts by means of imprecise Markov chains with a countable state space,
Imprecise probability and show how families of Markov chains give rise to imprecise versions of Markov chains.
Choquet capacity We discuss dual representations and differential equations related to the latter.
Imprecise Markov chain © 2020 Elsevier Inc. All rights reserved.
Nonlinear transition probability

1. Introduction

Model uncertainty appears in many scientific disciplines, when, for example, due to statistical estimation methods, only
a confidence interval for certain parameters of a model is known, or when certain aspects of a model cannot be determined
exactly. In this context, one often speaks of imprecision or polymorphic uncertainty. In mathematical finance, model un-
certainty or ambiguity is a frequent phenomenon since financial markets usually do not allow for repetition, whereas, in
other disciplines, experiments can be repeated under similar conditions arbitrarily often. The most prominent example for
ambiguity in mathematical finance is uncertainty with respect to certain parameters (drift, volatility, etc.) of the stochastic
process describing the value of an underlying asset. This leads to the task of modeling stochastic processes under model
uncertainty.
In mathematical finance, model uncertainty is often being described via nonlinear expectations, introduced by Peng [32].
Some of the most prominent examples of nonlinear expectations include the g-expectation, see Coquet et al. [11], describing
a Brownian Motion with uncertainty in the drift parameter, and the G-expectation or G-Brownian Motion introduced by
Peng [33,34], describing a Brownian Motion with uncertain volatility. There is a close connection between g-expectations,
backward stochastic differential equations (BSDEs) and semilinear partial differential equations. We refer to Coquet et al. [11]
and Pardoux and Peng [30] for more details on this topic. We also refer to Cheridito et al. [6] and Soner et al. [35,36] for the
connection between G-expectations, 2BSDEs and fully nonlinear partial differential equations. Moreover, there is a one-to-
one relation between sublinear expectations and coherent monetary risk measures as introduced by Artzner et al. [1] and

✩
Financial support through the German Research Foundation via CRC 1283 is gratefully acknowledged.
E-mail address: [email protected].

https://doi.org/10.1016/j.ijar.2020.12.013
0888-613X/© 2020 Elsevier Inc. All rights reserved.
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Delbaen [14,15]. Another related concept is the concept of a (Choquet) capacity (see e.g. Dellacherie-Meyer [16]) leading to
Choquet integrals (see Choquet [7]).
On the other side, there is a large community working on similar questions related to model uncertainty in the field of
imprecise probability. Here, the central objects are upper and lower previsions introduced by Walley [40]. In the sublinear
case, there is a one-to-one relation between sublinear expectations and coherent upper previsions, which creates a huge
intersection between the communities working on nonlinear expectations and upper/lower previsions. Within the field
of imprecise probability, many work has been done in the direction of defining, axiomatizing, and computing transition
operators of, both, discrete-time and continuous-time imprecise Markov chains, see e.g. De Bock [12], De Cooman et al. [13],
Krak et al. [26], and Škulj [37,38]. Concepts that are related to imprecise Markov chains include Markov set-chains, see
Hartfiel [25], and, in the field of mathematical finance, BSDEs on Markov chains, see Cohen and Elliott [9,10], and Markov
chains under nonlinear expectations, see Nendel [27] and Peng [32].
The aim of this paper is to link and compare the concepts and results obtained in the fields of imprecise probabil-
ity and mathematical finance. Since Markov chains under model uncertainty form the largest intersection between both
communities, we put a special focus on the latter.
The paper is organized as follows: In Section 2, we start by introducing nonlinear expectations, and discussing basic
properties and relations to upper/lower previsions, monetary risk measures and Choquet integrals. In Section 3, we present
extension procedures for pre-expectations due to Denk et al. [17]. Here, we focus on two different extensions, one in terms
of finitely additive measures, and the other in terms of countably additive measures. In Section 4, we discuss Kolmogorov-
type extension theorems and the existence of stochastic processes under nonlinear expectations due to Denk et al. [17].
We conclude, in Section 5, by constructing imprecise versions of transition operators for families of time-homogeneous
continuous-time Markov chains with countable state space. Here, we use an approach due to Nisio [29], which has been
used in various contexts to construct imprecise versions of Markov processes, such as Lévy processes, Ornstein-Uhlenbeck
processes, geometric Brownian Motions, and finite-state Markov chains, see Denk et al. [18], Nendel [27] and Nendel and
Röckner [28]. Finally, we compare the Nisio approach to the methods used for continuous-time imprecise Markov chains
in the field of imprecise probability. In Section 6, we conclude by summarizing the main insights and connections that are
illustrated in the paper.

2. Nonlinear expectations and related concepts

In this section, we give an introduction into the theory of nonlinear expectations, and discuss related concepts. Through-
out this section, let  be a nonempty set, and F ⊂ 2 be an arbitrary σ -algebra on , where 2 denotes the power set
of . We emphasize that, throughout this section, F = 2 is a possible choice for F . We denote the space of all bounded
F -B(R)-measurable random variables X :  → R by L∞ (, F ), where B(R) denotes the Borel σ -algebra on R. The space
L∞ (, F ) and subspaces thereof are endowed with the standard norm  · ∞ , defined by

 X ∞ := sup | X (ω)| ( X ∈ L∞ (, F )).

ω∈
For α ∈ R, we make use of the notation α := α 1 , and, for M ⊂ L∞ (, F ), we write R ⊂ M instead of {α 1 : α ∈ R} ⊂ M.
Here, 1 A stands for the indicator function of A ⊂ . We point out that, throughout, we use the notation “⊂” also for
inclusions that are not strict.
We write ba(, F ) for the space of all real-valued and finitely additive measures on (, F ) with finite total variation,
and ca(, F ) for the subspace of all σ -additive signed measures on (, F ). The subsets ba1+ (, F ) and ca1+ (, F ) stand for
all positive elements μ ∈ ba(, F ) and μ ∈ ca(, F ) with μ() = 1, respectively. For a normed  vector space M, we denote
by M  its topological dual space. Using the identification ba(, F ) = (L∞ (, F )) via μ X :=  X dμ for μ ∈ ba(, F ) and
X ∈ L∞ (, F ) (cf. [20, p. 258]), every monotone linear functional E : L∞ (, F ) → R with E(α ) = α , for all α ∈ R, is the
expectation of a finitely additive probability measure μ ∈ ba1+ (, F ). This motivates the following definition, which is due
to Peng [32].

Definition 2.1. Let M ⊂ L∞ (, F ) with R ⊂ M. A (nonlinear) pre-expectation E on M is a functional E : M → R with the
following properties:

(i) Monotonicity: E ( X ) ≤ E (Y ) for all X , Y ∈ M with X ≤ Y .

(ii) Constant preserving: E (α ) = α for all α ∈ R.

A pre-expectation on L∞ (, F ) is called an expectation.

Definition 2.2. Let M ⊂ L∞ (, F ) with R ⊂ M, and E : M → R be a pre-expectation on M.

a) If M + R := { X + α : X ∈ M , α ∈ R} ⊂ M and E ( X + α ) = E ( X ) + α for all X ∈ M and α ∈ R, we say that E is translation

invariant.

227
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

b) If M + M := { X + Y : X , Y ∈ M } ⊂ M and E ( X + Y ) ≤ E ( X ) + E (Y ) for all X , Y ∈ M, we say that E is subadditive.

c) If M is convex, i.e., λ X + (1 − λ)Y ∈ M for all X , Y ∈ M and λ ∈ [0, 1], and E (λ X + (1 − λY )) ≤ λE ( X ) + (1 − λ)E (Y ) for
all X , Y ∈ M and λ ∈ [0, 1], we say that E is convex.
d) If [0, ∞) · M := {λ X : λ ≥ 0, X ∈ M } ⊂ M and E (λ X ) = λE ( X ) for all λ ≥ 0 and X ∈ M, we say that E is positive homoge-
neous (of degree 1).
e) If M is a convex cone, i.e., M is convex and [0, ∞) · M ⊂ M, and E is convex and positive homogeneous, we say that E
is sublinear.
f) If M is a linear subspace of L∞ (, F ), i.e., λ X + Y ∈ M for all λ ∈ R and X , Y ∈ M, and E (λ X + Y ) = λE ( X ) + E (Y ) for
all λ ∈ R and X , Y ∈ M, we say that E is linear.
 
Remark 2.3. Assume that M is a convex cone. Then, X + Y = 2 12 X + 12 Y ∈ M for all X , Y ∈ M, and therefore M + M ⊂ M.
Since 0 ∈ M, any two of the following three conditions imply the remaining third:

(i) E is convex,
(ii) E is positive homogeneous (of degree 1),
(iii) E is subadditive.

Remark 2.4. A concept that is very much related to nonlinear expectations is the concept of an upper prevision as intro-
duced by Walley [40]. The latter are very prominent in the context of imprecise probabilities. Walley [40] defines a upper
prevision as a real-valued functional M → R on an arbitrary set of gambles M ⊂ L∞ (, F ). A nonlinear pre-expectation
can thus be seen as an upper prevision that is monotone and preserves constants. The concept of a sublinear expectation is
equivalent to the concept of a coherent upper prevision. More precisely, if E is a sublinear expectation, which is defined on
a linear subspace M of L∞ (, F ) with R ⊂ M, then P ( X ) := E ( X ), for X ∈ M, defines a coherent upper prevision. More-
over, if E is a convex expectation, which is defined on a linear subspace M of L∞ (, F ) with R ⊂ M, then P ( X ) := E ( X )
is a convex upper prevision, cf. Pelessoni and Vicig [31, Theorem 3.1].

In this work, we mainly focus on expectations that are translation invariant. Prominent examples for such expectations
are Choquet integrals, see Choquet [7] and (monetary) risk measures, see e.g. Föllmer and Schied [23]. We start with some
observations that help to verify translation invariance.

Lemma 2.5. Let M ⊂ L∞ (, F ) with R ⊂ M, and let E : M → R be a pre-expectation on M.

a) If E is translation invariant, then E is 1-Lipschitz, i.e., |E ( X ) − E (Y )| ≤  X − Y ∞ for all X , Y ∈ M.

b) Let M + R ⊂ M. Then, the following three statements are equivalent:
(i) E is translation invariant,
(ii) E ( X + α ) ≤ E ( X ) + α for all X ∈ M and all α ∈ R,
(iii) E ( X + α ) ≥ E ( X ) + α for all X ∈ M and all α ∈ R.
c) If E is subadditive, then E is translation invariant.
d) If E is convex, then E is translation invariant.

Proof. a) Assume that E is translation invariant. Then,

E ( X ) − E (Y ) ≤ E (Y +  X − Y ∞ ) − E (Y ) =  X − Y ∞
for all X , Y ∈ M. By a symmetry argument, we obtain that |E ( X ) − E (Y )| ≤  X − Y ∞ for all X , Y ∈ M.
b) First assume that E ( X + α ) ≤ E ( X ) + α for all X ∈ M and all α ∈ R. Then,
 
E ( X ) = E ( X + α) − α ≤ E ( X + α) − α.
Now, assume that E ( X + α ) ≥ E ( X ) + α for all X ∈ M and all α ∈ R. Then,
 
E ( X + α ) − α ≤ E ( X + α ) − α = E ( X ).
We have therefore established the equation between (ii) and (iii), which, in turn, implies the remaining equivalence.
c) This follows directly from part b).
d) Assume that E is convex. Let X ∈ M and α ∈ R. Then, for all λ ∈ [0, 1),
 α  ≤ λE ( X ) + (1 − λ)E  X  + α 
E ( X + α ) ≤ λE ( X ) + (1 − λ)E X + 1−λ ∞ 1−λ

= λE ( X ) + (1 − λ) X ∞ + α .
Letting λ → 1, we obtain that E ( X + α ) ≤ E ( X ) + α . By part b), it follows that E is translation invariant. 

228
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Remark 2.6. Since every convex pre-expectation E is translation invariant, ρ ( X ) := E (− X ) defines a convex monetary risk
measure, cf. Föllmer and Schied [22] and Frittelli and Rosazza Gianin [24]. If E is sublinear, i.e., convex and subadditive,
then ρ is a coherent monetary risk measure as introduced by Artzner et al. [1], see also Delbaen [14,15].

As in the theory of risk measures, random variables with positive expectation, play a special role, and we refer to them
as acceptable positions.

Definition 2.7. Let M ⊂ L∞ (, F ) with R ⊂ M.

a) A set A ⊂ M is called an acceptance set of M if

(i) inf{α ∈ R : α ∈ A} = 0,
(ii) for all X ∈ A and all Y ∈ M with Y ≥ X we have that Y ∈ A,
(iii) A is a closed subset of M.
b) Let E : M → R be a pre-expectation. Then, the set

AE := { X ∈ M : E ( X ) ≥ 0}
is called the acceptance set of E .

In the field of imprecise probability, acceptable positions are called desirable gambles, and acceptance sets are called
sets of desirable gambles. Translation invariant expectations are uniquely characterized via their acceptable positions. This is
a well-known fact within the theory of risk measures and imprecise probabilities, and directly carries over to translation
invariant expectations. For the reader’s convenience, we state the result and provide a short proof.

Proposition 2.8. Let M ⊂ L∞ (, F ) with R ⊂ M and M + R ⊂ M. Then, the mapping E → AE is a bijection between the set of all
translation invariant pre-expectations on M and the set of all acceptance sets of M. More precisely, the following holds:

a) Let E : M → R be a translation invariant pre-expectation. Then, AE is an acceptance set in the sense of Definition 2.7 a), and

E ( X ) = sup{α ∈ R : X − α ∈ AE }
for all X ∈ M.
b) Let A ⊂ M be an acceptance set. Then,

E ( X ) := sup{α ∈ R : X − α ∈ A} ( X ∈ M )
defines a translation invariant pre-expectation E : M → R with AE = A.

Proof. a) Since E is translation invariant, for all X ∈ M,

E ( X ) = sup{α ∈ R : E ( X ) ≥ α } = sup{α ∈ R : E ( X − α ) ≥ 0} = sup{α ∈ R : X − α ∈ AE }.

Clearly, AE satisfies the properties (i) and (ii) in Definition
 2.7. Since E is translation invariant, it is a 1-Lipschitz
continuous map M → R, and therefore AE = E −1 [0, ∞) is a closed subset of M.
b) One readily verifies that E defines a translation invariant pre-expectation on M. By definition, A ⊂ AE . On the other
hand, for X ∈ AE , there exists a sequence (αn )n∈N ⊂ (−∞, 0] with X − αn ∈ A and αn → 0 as n → ∞. Since A is a
closed subset of M, it follows that X = limn→∞ X − αn ∈ A. 

We now specialize on the case, where E is a convex expectation on a linear subspace of L∞ (, F ). Let M ⊂ L∞ (, F )
be a linear subspace of L∞ (, F ) with R ⊂ M. For a convex function E : M → R, we write E ∗ for its conjugate function or
Fenchel-Legendre transform, i.e., we define

E ∗ (μ) := sup μ X − E ( X )
X ∈M

for all linear functionals μ : M → R. Note that the conjugate function E ∗ may take the value +∞. In the following proposi-
tion, we will see that, for a convex pre-expectation E on M, its conjugate function E ∗ is concentrated on the class of linear
pre-expectations on M. That is, E ∗ is finite only for linear pre-expectations on M. As every linear pre-expectation on M is
continuous, we therefore obtain the representation

E ( X ) = sup μ X − E ∗ (μ) ( X ∈ M )
μ∈ M 

229
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

for all convex pre-expectations E on M, where M  denotes the topological dual space of M. Again, this type of representation
is well-known for convex risk measures, and the proof relies on a collection of several well-known facts from convex analysis
and duality theory. In order to keep this manuscript self-contained, we nevertheless provide the proof below.

Proposition 2.9. Let M ⊂ L∞ (, F ) be a linear subspace of L∞ (, F ) with R ⊂ M.

a) Let E : M → R be a convex pre-expectation. Then, every linear functional μ : M → R with E ∗ (μ) < ∞ is a linear pre-expectation
on M, and therefore μ ∈ M  with μ M  = 1, where  ·  M  denotes the operator norm on M  .
b) Let E : M → R be a convex pre-expectation, and let PE denote the set of all linear functionals μ : M → R with E ∗ (μ) < ∞.
Then, for all X ∈ M,

E ( X ) = max μ X − E ∗ (μ).
μ∈PE

Moreover, the map E ∗ : PE → R is convex and weak∗ lower semicontinuous with

min E ∗ (μ) = 0.
μ∈PE

c) Let P ⊂ M  be a set of linear pre-expectations, and σ : P → R be a map with infμ∈P σ (μ) = 0. Then,

E ( X ) := sup μ X − σ (μ), for X ∈ M ,

μ∈P
defines a convex pre-expectation on M. It holds P ⊂ PE and E ∗ (μ) ≤ σ (μ) for all μ ∈ P .

Proof. μ : M → R be linear with E ∗ (μ) < ∞. Then, for all λ > 0,

a) Let
 
1 − λ−1 E ∗ (μ) = −λ−1 E (−λ) + E ∗ (μ) ≤ −λ−1 μ(−λ) = μ1 = λ−1 μ(λ)
 
≤ λ−1 E (λ) + E ∗ (μ) = 1 + λ−1 E ∗ (μ).
Letting λ → ∞, we obtain that μ1 = 1. Moreover, for λ > 0 and all X , Y ∈ M with X ≤ Y ,
     
μ( X − Y ) = λ−1 μ λ( X − Y ) ≤ λ−1 E λ( X − Y ) + E ∗ (μ) ≤ λ−1 E ∗ (μ) → 0, λ → ∞.
This shows that μ : M → R is a linear pre-expectation on M. Therefore, Lemma 2.5 a) and c) imply that μ ∈ M  with
μ M  ≤ 1. Since μ1 = 1, it follows that μ M  = 1.
b) Let X ∈ M and E0 (Y ) := E ( X + Y ) − E ( X ) for all Y ∈ M. Then, E0 : M → R is convex and E0 (0) = 0. By the extension
theorem of Hahn-Banach, there exists a linear functional μ : M → R with μY ≤ E0 (Y ) for all Y ∈ M. That is,

μY − E (Y ) ≤ μ X − E ( X )
for all Y ∈ M. Therefore, by definition of the conjugate function,

E ( X ) = μ X − E ∗ (μ).
One readily checks that PE ⊂ M  is convex. Hence, for fixed X ∈ M, the map

PE → R, μ → μX − E(X)
is convex and weak∗ continuous. Taking the pointwise supremum over all X ∈ M, we see that the mapping PE →
R, μ → sup X ∈ M μ X − E ( X ) is convex and weak∗ lower semicontinuous. As 0 ∈ M with E (0) = 0, it follows that E ∗ (μ) ≥
0 for all μ ∈ PE . Again, as 0 ∈ M, there exists some μ ∈ PE with E ∗ (μ) = μ0 − E (0) = 0.
c) The map E is monotone and convex, as it is a supremum over monotone affine linear maps. Moreover,

E (α ) = sup μα − σ (μ) = α − inf σ (μ) = α .

μ∈P μ∈P

Therefore, E defines a convex pre-expectation on M. Let ν ∈ P . Then,

ν X − E ( X ) = ν X − sup μ X − σ (μ) ≤ σ (μ)

μ∈P
for all X ∈ M. Hence, E ∗ (ν ) ≤ σ (ν ) and, in particular, P ⊂ PE . 

The set PE from the previous proposition can be used to characterize sublinear and linear pre-expectations. In the field
of imprecise probability or, more precisely, for coherent upper previsions the set PE is called the credal set.

230
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Lemma 2.10. Let M ⊂ L∞ (, F ) be a linear subspace of L∞ (, F ) with R ⊂ M, and let E : M → R be a convex pre-expectation.

a) E is sublinear if and only if PE = {μ ∈ M  : E ∗ (μ) = 0}.

b) E is linear if and only if #PE = 1. In this case, PE = {E }.

Proof. a) First, assume that E is sublinear. Let μ ∈ PE and X ∈ M be arbitrary. Then,

 
λ μ X − E ( X ) = μ(λ X ) − E (λ X ) ≤ E ∗ (μ) < ∞
for all λ > 0, and therefore μ X − E ( X ) ≤ 0. Taking the supremum over all X ∈ M, we obtain that E ∗ (μ) ≤ 0. By
Proposition 2.9 b), this implies that E ∗ (μ) = 0.
Now assume that PE = {μ ∈ M  : E ∗ (μ) = 0}. Then, for all X ∈ M and all λ ≥ 0 we have that

E (λ X ) = max μ(λ X ) = λ max μ X = λE ( X ).

μ∈P μ∈P
b) Assume that E is linear and let μ ∈ PE . Then, by part a), E ∗ (μ) = 0, and therefore μ X ≤ E ( X ) for all X ∈ M. As E is
linear, we thus obtain that μ = E . This shows that P = {E }.
Now let #P = 1 and let μ ∈ P . Then, by Proposition 2.9 b), E ∗ (μ) = 0, and therefore E = μ is linear. 

Proposition 2.9 together with Lemma 2.10 yields the following characterization of sublinear pre-expectations.

Proposition 2.11. Let M ⊂ L∞ (, F ) be a linear subspace of L∞ (, F ) with R ⊂ M. Then, the map E → PE is a bijection between
the set of all sublinear expectations and the set of all nonempty, convex and weak∗ compact sets of linear pre-expectations. More
precisely, the following holds:

a) Let E : M → R be a sublinear pre-expectation. Then,

E ( X ) = max μ X .
μ∈PE
Moreover, the set PE is nonempty, convex, and weak∗ compact.
b) Let P be a nonempty set of linear pre-expectations. Then,

E ( X ) := max μ X ( X ∈ M)
μ∈P
gives rise to a sublinear pre-expectation on M. If P is convex and weak∗ compact, then P = PE .

Proof. a) By Lemma 2.10, PE = {μ ∈ M  : E ∗ (μ) = 0}. Therefore, by Proposition 2.9 b),

E ( X ) = max μ X .
μ∈PE
Clearly, PE is convex and, by Proposition 2.9, nonempty. Since

PE = {μ ∈ M  : μ X ≤ E ( X )}
X ∈M

is closed, and PE ⊂ {μ ∈ M  : μ M  ≤ 1}, we obtain that PE is weak∗ compact by the Banach-Alaoglu Theorem.
b) One readily verifies that E defines a sublinear pre-expectation on M. Assume that P is convex and weak∗ compact.
Then, by Proposition 2.9 c), P ⊂ PE . In order to prove the other inclusion, let ν ∈ M  \ P . Then, by the separation
theorem of Hahn-Banach, there exists some X ∈ M with

E ( X ) = sup μ X < ν X ,
μ∈PE

where we used the fact that the topological dual of M  (endowed with the weak∗ topology) is M. Hence, E ∗ (ν ) > 0 and
/ PE . This shows that P = PE . 
therefore, by Lemma 2.10 a), ν ∈

Example 2.12. Let c : F → [0, 1] be a capacity, i.e., function with c (∅) = 0, c () = 1, and c ( A ) ≤ c ( B ) for all A , B ∈ F with
A ⊂ B. Then, c gives rise to a translation invariant expectation E : L∞ (, F ) → R via
 ∞
E ( X ) := X dc := c ( X ≥ x) dx, for X ∈ L∞ (, F ) with X ≥ 0,
0

231
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

  
and E ( X ) := X dc := E X +  X ∞ −  X ∞ for X ∈ L∞ (, F ). The nonlinear expectation E is called the Choquet integral
of X w.r.t. c, see Choquet [7]. By definition, E is positive homogeneous, translation invariant, and satisfies E (1 A ) = c ( A ) for
all A ∈ F . Assume that c is 2-alternating, i.e.

c( A ∪ B ) + c( A ∩ B ) ≤ c( A) + c( B ) for all A , B ∈ F .

Then, it is well-known that E is subadditive and thus sublinear, see, e.g., Denneberg [19, Theorem 6.3]. By definition,
μ X ≤ E ( X ) for all X ∈ L∞ (, F ) and every linear expectation μ ∈ ba1+ (, F ) with μ( A ) ≤ c ( A ) for all A ∈ F . Recall that
linear expectations can be identified by finitely additive measures via μ( A ) := μ1 A , for A ∈ F . This implies that

PE = μ ∈ ba1+ (, F ) : ∀ A ∈ F : μ( A ) ≤ c ( A ) ,

and thus Ẽ ( X ) ≤ E ( X ) for all X ∈ L∞ (, F ) and every sublinear expectation Ẽ : L∞ (, F ) → R with Ẽ (1 A ) ≤ c ( A ) for all
A ∈ F . That is, E is the largest sublinear expectation with E (1 A ) ≤ c ( A ) for all A ∈ F . The following corollary shows that a
sublinear expectation that coincides with a finitely additive measure on all F -measurable sets is already linear.

Corollary 2.13. Let E : L∞ (, F ) → R be a sublinear expectation and μ be a finitely additive measure with E (1 A ) ≤ μ( A ) for all
A ∈ F . Then, E ( X ) = μ X for all X ∈ L∞ (, F ), and therefore E is a linear expectation.

Proof. Let ν ∈ PE . Then, ν is a linear expectation, i.e., a finitely additive measure, with ν ( A ) ≤ E (1 A ) ≤ μ( A ) for all A ∈ F .
Hence, μ = ν , i.e., PE = {μ}, which, by Lemma 2.10 b), implies that E = μ. 

For sublinear expectations, the following version of Jensen’s inequality holds.

Lemma 2.14. Let E : L∞ (, F ) → R be a sublinear expectation and h : R → R be a convex function. Then, for all X ∈ L∞ (, F ),
   
h E ( X ) ≤ E h( X ) .

Proof. Since h : R → R is convex, h is continuous, and therefore h( X ) ∈ L∞ (, F ) for all X ∈ L∞ (, F ). Moreover,

h(x) = max λx − h∗ (λ), for all x ∈ R,

λ∈U

where h∗ : R → R is the conjugate function or Fenchel-Legendre transform of h and U := {λ ∈ R : h∗ (λ) < ∞}. Now, let
X ∈ M and λ ∈ U . If λ ≥ 0, then
   
λE ( X ) − h∗ (λ) = E λ X − h∗ (λ) ≤ E h( X ) .
If λ < 0, then
   
λE ( X ) − h∗ (λ) ≤ E (λ X ) − h∗ (λ) = E λ X − h∗ (λ) ≤ E h( X ) ,
where, in the first inequality, we used the sublinearity of E . Taking the supremum over all λ ∈ U yields the assertion. 

As in the linear case, one can define the concept of a distribution for nonlinear expectations.

Remark 2.15. Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a pre-expectation. Moreover, let S = ∅ be a set (state
space) and T :  → S be an arbitrary map. Then, N := { f ∈ L∞ ( S , 2 S ) : f ( T ) := f ◦ T ∈ M } contains all constant functions
S → R, and one readily verifies that
 
E ◦ T −1 : N → R, Y → E f (T )

defines a pre-expectation on N. We call E ◦ T −1 the distribution of T under E . Note that, if M = L∞ (, F ), then N =
L∞ ( S , S ), where

S := B ∈ 2 S : { T ∈ B } := T −1 ( B ) ∈ F
is the σ -algebra generated by T . In particular, N ⊂ L∞ ( S , S ) for all M ⊂ L∞ (, F ). We would like to point out that, for a
linear expectation E, the distribution ν = E ◦ T −1 is given by

   
f dν = E ◦ T −1 ( f ) = E f ( T ) ( f ∈ L∞ ( S , S )).
S

232
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

The concept of distribution for nonlinear expectations is therefore consistent with the distribution for expectations w.r.t.
probability measures. If M is a linear subspace of L∞ (, F ) and E is sublinear, then N is a linear subspace of L∞ ( S , S )
and E ◦ T −1 is sublinear expectation. In this case,

PE ◦ T −1 = ν ∈ N  : (E ◦ T −1 )∗ (ν ) < ∞ = {μ ◦ T −1 : μ ∈ PE } =: PE ◦ T −1 .
That is, the dual representation of the sublinear distribution E ◦ T −1 is given in terms of the distributions of the dual
representation of E . In particular,
 
E ◦ T −1 ( f ) = sup μ f ( T ) ( f ∈ N ).
μ∈P

In fact, as the map M  → N  , μ → μ ◦ T −1 is linear and weak∗ continuous, the set PE ◦ T −1 is convex and weak∗ compact.
Moreover, for all Y ∈ N,
   
(E ◦ T −1 )( f ) = E f ( T ) = max μ ( f ( T ) = max (μ ◦ T −1 ) f = max νf.
μ∈PE μ∈PE ν ∈PE ◦ T −1

By Lemma 2.11 b), it follows that PE ◦ T −1 = PE ◦ T −1 .

3. Extension of pre-expectations

Let M ⊂ L∞ (, F ) with R ⊂ M. Given a pre-expectation E : M → R, we are looking for extensions of E to an ex-
pectation on L∞ (, F ). Here, the main challenge is to preserve monotonicity. We start with the extension of linear
pre-expectations.

Remark 3.1. Let M ⊂ L∞ (, F ) be a linear subspace of L∞ (, F ) with 1 ∈ M. We denote by ba1+ ( M ) the space of all linear
pre-expectations on M. A natural question is if the mapping

ba1+ (, F ) → ba1+ ( M ), ν → ν |M (3.1)

is bijective. The following theorem by Kantorovich shows that this mapping is surjective, i.e., any linear pre-expectation on
M can be extended to a linear expectation on L∞ (, F ). For the reader’s convenience, we state this theorem and provide
a short sketch of the proof. For more details we refer to [39, p. 277]. However, in Example 3.15, we will see that, in general,
the mapping in (3.1) is not injective, not even if F = σ ( M ), i.e., even if F = σ ( M ), a linear pre-expectation on M usually
admits various extensions to an expectation on L∞ (, F ).

Theorem 3.2 (Kantorovich). Let M ⊂ L∞ (, F ) be a linear subspace of L∞ (, F ) with 1 ∈ M and μ : M → R be a linear pre-
expectation on M. Then, there exists a linear expectation ν : L∞ (, F ) → R with ν | M = μ.

Proof. Let

μ( X ) := inf{μ X 0 : X 0 ∈ M , X 0 ≥ X }
for all X ∈ L∞ (, F ). Then, μ : L∞ (, F ) → R is a sublinear expectation with μ| M = μ. By the extension theorem of
Hahn-Banach, there exists a linear functional ν : L∞ (, F ) → R with ν | M = μ and ν X ≤ μ( X ) for all X ∈ L∞ (, F ). Thus,

ν X − ν Y = ν ( X − Y ) ≤ μ( X − Y ) ≤ μ(0) = 0
for all X , Y ∈ L∞ (, F ) with X ≤ Y . 

In the proof of the previous theorem, before applying the Hahn-Banach Theorem, the linear pre-expectation μ: M → R
is extended to a sublinear expectation μ : L∞ (, F ) → R via

μ( X ) := inf{μ X 0 : X 0 ∈ M , X 0 ≥ X } ( X ∈ L∞ (, F )).

The idea for the first extension procedure therefore is, to extend a pre-expectation E : M → R on M ⊂ L∞ (, F ) with
R ⊂ M via

E ( X ) := inf{E ( X 0 ) : X 0 ∈ M , X 0 ≥ X } ( X ∈ L∞ (, F )).

The following proposition is due to Denk et al. [17], and shows that E is an expectation with E | M = E . Moreover, if M is
assumed to be convex or a convex cone, then convexity or sublinearity of E carry over to the extension E , respectively. For
related extension results on niveloids, we refer to Maccheroni et al. [5]. In the context of lower previsions, for F = 2 , the
extension E is usually referred to as the natural extension, cf. Walley [40, Section 3.1].

233
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Proposition 3.3. Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a pre-expectation on M. Further, let

E ( X ) := inf{E ( X 0 ) : X 0 ∈ M , X 0 ≥ X }
for all X ∈ L∞ (, F ). Then, the following assertions hold:

a) E : L∞ (, F ) → R is an expectation with E | M = E .

b) If E is translation invariant, then E is translation invariant.
c) If E is convex, then E is convex.
d) If E is sublinear, then E is sublinear.

Proof. a) Let X ∈ L∞ (, F ). As R ⊂ M, we have that  X ∞ ∈ M with  X ∞ ≥ X . Thus, the set {E ( X 0 ) : X 0 ∈ M , X 0 ≥ X }

is nonempty. Since X 0 ≥ − X ∞ for all X 0 ∈ M with X 0 ≥ X , we obtain that

E ( X 0 ) ≥ E (− X ∞ ) = − X ∞
for all X 0 ∈ M with X 0 ≥ X . Hence, E : L∞ (, F ) → R is well-defined. Further, if X ∈ M we have that E ( X ) ≤ E ( X 0 )
for all X 0 ∈ M with X 0 ≥ X and, therefore, E ( X ) = E ( X ). Since R ⊂ M, we thus obtain that E (α ) = α for all α ∈ R.
Now, let X , Y ∈ L∞ (, F ) with X ≤ Y . Then, Y 0 ≥ X for all Y 0 ∈ M with Y 0 ≥ Y and, therefore, E ( X ) ≤ E (Y ).
b) Assume that M + R ⊂ M and that E is translation invariant. Let X ∈ L∞ (, F ) and α ∈ R. Then, for all X 0 ∈ M with
X 0 ≥ X + α we have that X 0 − α ≥ X , and therefore

E ( X0) = E ( X0 − α) + α ≥ E ( X ) + α.
Taking the infimum over all X 0 ∈ M with X 0 ≥ X + α yields that E ( X + α ) ≥ E ( X ) + α , which, by Lemma 2.5 b) implies
that E is translation invariant.
c) Assume that M is convex and that E is convex. Let X , Y ∈ L∞ (, F ) and λ ∈ [0, 1]. Moreover, let X 0 , Y 0 ∈ M with
X 0 ≥ X and Y 0 ≥ Y . Since M is convex, λ X 0 + (1 − λ)Y 0 ∈ M with

λ X 0 + (1 − λ)Y 0 ≥ λ X + (1 − λ)Y .
Due to convexity of E , we thus obtain that

E (λ X + (1 − λ)Y ) ≤ E (λ X 0 + (1 − λ)Y 0 ) ≤ λE ( X 0 ) + (1 − λ)E (Y 0 ).

Taking the infimum over all X 0 , Y 0 ∈ M with X 0 ≥ X and Y 0 ≥ Y , we get that

E (λ X + (1 − λ)Y ) ≤ λE ( X ) + (1 − λ)E (Y ).
d) Now assume that M is a convex cone and that E is sublinear. Then, E is convex and part b) yields that E is convex as
well. Moreover, as λ X 0 ∈ M for all X 0 ∈ M and λ > 0 we have that

E (λ X ) = inf{E (λ X 0 ) : X 0 ∈ M , X 0 ≥ X } = inf{λE ( X 0 ) : X 0 ∈ M , X 0 ≥ X } = λE ( X )
for all X ∈ L∞ (, F ) and all λ > 0. Hence, E is convex and positive homogeneous, and therefore sublinear. 

Remark 3.4.

a) Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a pre-expectation on M. Let E : L∞ (, F ) → R be an expectation

with E | M = E and X ∈ L∞ (, F ). Then,

E ( X ) ≤ E ( X0) = E ( X0)
for all X 0 ∈ M with X 0 ≥ X . Taking the infimum over all X 0 ∈ M with X 0 ≥ X , we see that E ( X ) ≤ E ( X ). That is, E is
the largest expectation, which extends E .
b) Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a pre-expectation on M. For X ∈ M, let

Ě ( X ) := sup{E ( X 0 ) : X 0 ∈ M , X 0 ≤ X }.
Then, one readily verifies that Ě : L∞ (, F ) → R is the smallest expectation, which extends E . However, convexity of
E usually does not carry over to Eˇ .

The following lemma provides an alternative extension procedure for translation invariant pre-expectations, which is an
extension via acceptable positions and is intimately related to the idea of superhedging. As we will see, however, it leads to
the same expectation.

234
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Lemma 3.5. Let M ⊂ L∞ (, F ) with R ⊂ M and let E : M → R be a translation invariant pre-expectation on M. Let

AE := { X ∈ L∞ (, F ) : X 0 ∈ AE for all X 0 ∈ M with X 0 ≥ X }.

Then, AE = AE , i.e., AE is the acceptance set of E , and therefore

E ( X ) = sup α ∈ R : X − α ∈ AE (3.2)
for all X ∈ L∞ (, F ). Thus, (3.2) provides a second extension procedure for E , which is extending E via its acceptable positions AE .

Proof. Let X ∈ AE , i.e., E ( X ) ≥ 0. By definition of E , it follows that E ( X 0 ) ≥ 0 for all X 0 ∈ M with X 0 ≥ X , i.e., X ∈ AE . Now
assume that X ∈ AE , i.e., E ( X 0 ) ≥ 0 for all X 0 ∈ M with X 0 ≥ X . Then,

E ( X ) = inf{E ( X 0 ) : X 0 ∈ M , X 0 ≥ X } ≥ 0,
i.e., X ∈ AE . Now, (3.2) follows directly from Proposition 2.8. 

Although Proposition 3.3 implies the existence of an extension E for every pre-expectation E : M → R, this extension is
not necessarily unique. However, as translation invariant pre-expectations are 1-Lipschitz by Lemma 2.5 a), the extension is
uniquely determined on the closure M of M for translation invariant pre-expectations.

Proposition 3.6. Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a translation invariant pre-expectation. Then, there exists
exactly one translation invariant pre-expectation E : M → R with E | M = E . Here, M denotes the closure of M as a subset of L∞ (, F )
and E is given as in Proposition 3.3. If E is convex or sublinear, then E is convex or sublinear, respectively.

Proof. As M + R ⊂ M it follows that M + R ⊂ M. Hence, by Proposition 3.3, there exists a translation invariant pre-
expectation E : M → R with E | M = E . Since, by Lemma 2.5 a), every translation invariant pre-expectation on M is 1-
Lipschitz, it is uniquely determined by its values on M. 

Lemma 3.7. Let M and N be two linear subspaces of L∞ (, F ) with R ⊂ M ⊂ N, and let E : N → R a convex pre-expectation. Then,

μ ∈ M  : (E |M )∗ (μ) < ∞ = ν |M : ν ∈ PE
and, for all μ ∈ M  with (E | M )∗ (μ) < ∞,

(E | M )∗ (μ) = min E ∗ (ν ).
ν ∈PE ,ν |M =μ

Proof. By Proposition 2.9 c),

μ ∈ M  : (E |M )∗ (μ) < ∞ ⊃ ν |M : ν ∈ N  , E ∗ (ν ) < ∞

and (E | M )∗ (ν | M ) ≤ E ∗ (ν ) for all ν ∈ PE . Therefore, let μ ∈ M  with (E |M )∗ (μ) < ∞. Then, we have that

μ X ≤ E ( X ) − E ∗ (μ)
for all X ∈ M. Hence, by the extension theorem of Hahn-Banach, there exists a linear functional ν : N → R with ν | M = μ
and

ν X ≤ E ( X ) + E ∗ (μ)
for all X ∈ N. Therefore, E ∗ (ν ) ≤ (E | M )∗ (μ), i.e., ν ∈ PE with ν |M = μ and E ∗ (ν ) = (E |M )∗ (μ). 

We apply Lemma 3.7 to the case N = L∞ (, F ) and obtain the following corollary.

Corollary 3.8. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and E : M → R be a convex pre-expectation. If E : L∞ (, F ) →
R be a convex expectation, which extends E , then

PE = ν | M : ν ∈ PE
and, for all μ ∈ PE ,

E ∗ (μ) = min E ∗ (ν ).
ν ∈PE ,ν |M =μ

235
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

In view of Proposition 2.9 and Corollary 3.8, another natural approach to extend a convex pre-expectation E on M would
be to consider PE := {ν ∈ ba1+ (, F ) : ν | M ∈ PE }, and to define E : L∞ (, F ) → R by

E ( X ) := sup ν X − E ∗ (ν | M ) (3.3)
ν ∈PE

for all X ∈ L∞ (, F ). By Proposition 2.9 c) and Theorem 3.2, E : L∞ (, F ) → R is an expectation with E | M = E . In the
following proposition, we will prove that E = E and that the supremum in (3.3) is attained.

Proposition 3.9. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and E : M → R be a convex pre-expectation. Further, let

PE := {ν ∈ ba1+ (, F ) : ν | M ∈ PE }.
Then, PE = PE and E ∗ (ν ) = E ∗ (ν | M ) for all ν ∈ PE . In particular,

E ( X ) = max ν X − E ∗ (ν | M ) ( X ∈ L∞ (, F )).

ν ∈PE

Proof. By Corollary 3.8, PE ⊂ PE and E ∗ (ν ) ≥ E ∗ (ν | M ) for all ν ∈ PE . Proposition 2.9 c), thus implies that PE = PE with
E ∗ (ν ) = E ∗ (ν | M ) for all ν ∈ PE . 

Corollary 3.10. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and E : M → R be a sublinear pre-expectation. Then, PE = PE ,
and therefore

E ( X ) = max ν X for all X ∈ L∞ (, F ).

ν ∈PE

Corollary 3.11. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and E : M → R be a sublinear pre-expectation. Then, there
exists a convex weak∗ compact set P ⊂ ba1+ (, F ) such that

E ( X ) = max μ X for all X ∈ M .

μ∈P

We return to the setting of Theorem 3.2. For a given linear pre-expectation μ : M → R on M, we consider the sublinear
expectation μ : L∞ (, F ) → R which extends μ. Then, for every ν ∈ M  we have that μ∗ (ν ) = 0 if and only if ν = μ. With
the previous results, we therefore obtain the following corollary, which states that μ is the pointwise maximum of all linear
expectations ν ∈ ba1+ (, F ) that extend μ.

Corollary 3.12. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and μ : M → R be a linear pre-expectation. Further, let

P := {ν ∈ ba1+ (, F ) : ν | M = μ}.

Then, P = {ν ∈ ba(, F ) : μ∗ (ν ) = 0}, and therefore

μ( X ) = max ν X for all X ∈ L∞ (, F ).

ν ∈P

Let E : M → R be a convex pre-expectation. Considering the sublinear expectation μ, which extends μ ∈ PE , one could
also think of

E ( X ) := sup μ( X ) − E ∗ (μ) ( X ∈ L∞ (, F )) (3.4)

μ∈PE

as another possible extension of E . Clearly, we have that E | M = E , and therefore E (α ) = α for all α ∈ R. Moreover, as μ is
monotone for all μ ∈ PE , we also have that E is monotone. Hence, E is an expectation which extends E . In the following
proposition, we will use Corollary 3.12 to show that the expectation E coincides with E and that the supremum in (3.4) is
attained.

Proposition 3.13. Let M be a linear subspace of L∞ (, F ) with R ⊂ M and E : M → R be a convex pre-expectation. Then,

E ( X ) = max μ( X ) − E ∗ (μ) for all X ∈ L∞ (, F ).

μ∈PE

236
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Proof. Let X ∈ L∞ (, F ) and E : L∞ (, F ) → R be given by (3.4). We have already seen that E : L∞ (, F ) → R is an
expectation that extends E . Remark 3.4 thus yields that E ( X ) ≤ E ( X ). By Proposition 3.9, there exists a linear expectation
ν ∈ ba1+ (, F ) which satisfies μ := ν |M ∈ PE and E ( X ) = ν X − E ∗ (μ). Hence, by Corollary 3.12, we get that

E ( X ) = ν X − E ∗ (μ) ≤ μ( X ) − E ∗ (μ) ≤ E ( X ) ≤ E ( X ).
This shows that E ( X ) = ν X − E ∗ (μ) = μ( X ) − E ∗ (μ). 

Remark 3.14. Let E : M → R be a convex pre-expectation on M. In Lemma 3.5, Proposition 3.3, Proposition 3.9 and Proposi-
tion 3.13, we have seen that the following four extension procedures all lead to the same (maximal) expectation extending
E.

(i) E → E ,  
(ii) E → AE → AE → X → sup α∈R : X − α ∈ AE ,
   
(iii) E → (P , E ∗ ) → P , [ν → E ∗ (ν | M )] → X → maxν ∈P ν X − E ∗ (ν | M ) ,
   
(iv) E → (P , E ∗ ) → {μ : μ ∈ P }, E ∗ → X → maxμ∈P μ X − E ∗ (μ) .

Example 3.15. Let  := R, F be the Borel σ -algebra on R, and M := C b () be the space of all bounded continuous
functions  → R. Let δ0 be the Dirac measure with center 0, E := δ0 | M be the restriction of the Dirac measure to M, and
X := 1(0,∞) . Then,

E ( X ) = inf{Y ∈ M : Y ≥ X } = 1,
while δ0 ( X ) = 0. This shows that PE = δ0 , i.e., E admits several extensions in terms of finitely additive measures.

The previous example shows that the extension E , already in the linear case, usually does not result in uniqueness of any
kind. In order to obtain uniqueness, one therefore needs additional continuity properties of the expectation as an analogue
of σ -additivity in the linear case.

Definition 3.16.

a) Let M ⊂ L∞ (, F ) with R ⊂ M and E : M → R be a pre-expectation. Then, we say that E is continuous from above
or below if E ( X n ) → X as n → ∞ for all sequences ( X n )n∈N ⊂ M with X n ≥ X n+1 or X n ≤ X n+1 for all n ∈ N and
X := limn→∞ X n ∈ M, respectively.
b) Let E : L∞ (, F ) → R be a convex expectation. Assume that there exists a set P of probability measures, i.e., linear
expectations, which are continuous from above, and a function ρ : P → R with infμ∈P ρ (μ) = 0 such that

E ( X ) = sup μ X − ρ (μ), for all X ∈ L∞ (, F ),

μ∈P

then we say that (, F , E ) is a convex expectation space. If E is sublinear or linear, we say that (, F , E ) is a sublinear
or linear expectation space, respectively.

Notice that every linear expectation space is a probability space and vice versa.
Let E : M → R be a convex pre-expectation on a linear subspace M of L∞ (, F ) with R ⊂ M. Then, one can show that
the following statements are equivalent:

(i) E is continuous from above.

(ii) Every μ ∈ PE is continuous from above.

Remark 3.17. Assume that M is a Riesz subspace of L∞ (, F ), i.e., a linear subspace of L∞ (, F ) with X ∨ Y ∈ M and
X ∧ Y ∈ M for all X , Y ∈ M, with R ⊂ M. Then, by the Daniell-Stone Theorem, every linear pre-expectation
 on M, which
is continuous from above, can uniquely be extended to a linear expectation μ̄ on L∞ , σ ( M ) , which is continuous from
above. Now, assume that E is a convex expectation, and assume that there exists a set P ⊂ M  of linear pre-expectations,
which are continuous from above, and a map ρ : P → R with infμ∈P ρ (μ) = 0. Then,
  
Ē ( X ) := sup μ̄ X − ρ (μ) X ∈ L∞ , σ ( M )
μ∈P
 
defines a convex expectation, which extends E to L∞ , σ ( M ) (see Bartl et al. [3]). If E : M → R is continuous from
above, using a modification of Choquet’s theorem for capacities [8], one can show that

237
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245


N
Ē ( X ) := sup inf E ( X n ) : ( X n )n∈N ∈ M , X n ≥ X n+1 (n ∈ N), X ≥ inf X n
n∈N n∈N

for P = PE and ρ := E ∗ , and that uniqueness within a certain class of expectations can be achieved. We refer to Bartl [2]
and Denk et al. [17] for more details on the uniqueness of this extension.

4. Stochastic processes under nonlinear expectation

In this section, we apply the extension results of the previous chapter to a Kolmogorov-type setting. That is, given a
consistent family of finite-dimensional marginal expectations, we are looking for an expectation on a suitable path space
with these marginals. As in the linear case, the formulation of Kolmogorov’s consistency condition requires the distribution
for finite-dimensional projections, cf. Remark 2.15. Our formulation of the consistency condition is very much in the spirit
of the formulation in the linear case as, for example, in the textbook by Bauer [4].
Throughout this section, let I = ∅ be an index set, H := J ⊂ I : | J | ∈ N the set of all finite, nonempty subsets of I and
( S , S ) be a measurable (state) space. For each J ∈ H let M J ⊂ L∞ ( S J , B J ) be a linear subspace with 1 ∈ M J , where B J
is the product σ -algebra on S J . For all K ⊂ J ⊂ I let

pr J K : S J → S K , ( xi )i ∈ J → ( xi )i ∈ K
be the projection (restriction) from J to K . Moreover, we define pr J := pr I J as the projection from I to J ∈ H . Throughout
this section, we assume that

M K ◦ pr J K := f ◦ pr J K = f (pr J K ) : f ∈ M K ⊂ M J
for all J , K ∈ H with K ⊂ J . Typical examples for the family ( M J ) J ∈H are:

(i) the space L∞ ( S J ) := L∞ ( S J , B J ) of all bounded B J -B (R)-measurable functions, where B J denotes the product σ -
algebra on S J ,
(ii) the space C b ( S J ) of all bounded continuous functions S J → R, where S J is endowed with the product topology,
(iii) the space BUC( S J ) of all bounded uniformly continuous functions S J → R w.r.t. a fixed metric, which generates the
topology on S.

Let J , K ∈ H with K ⊂ J . For a pre-expectation E J : M J → R we then denote by E J ◦ pr−J K1 the restriction of the
distribution (see Remark 2.15) of E J under pr−J K1 to M K , i.e.

E J ◦ pr−J K1 : M K → R, f → E J ( f ◦ pr J K ).
In [32], Peng defines a consistency condition for nonlinear expectations and proves an extension to the subspace

M := f ◦ pr J = f (pr J ) : J ∈ H , f ∈ L∞ ( S J , B J )

of L∞ ( S I , B I ). Here B I denotes the product σ -algebra of B , i.e., the σ -algebra generated by the sets of the form pr−J 1 ( B J )
with J ∈ H and B J ∈ B . In the sequel, we use the same notion of consistency as Peng and apply the extension results
J

from the previous section in order to obtain an extension to L∞ ( S I , B I ).

Definition 4.1. For all J ∈ H let E J : M J → R be a pre-expectation. Then, the family (E J ) J ∈H is called consistent if, for all
J , K ∈ H with K ⊂ J ,

E K ( f ) = E J ( f ◦ pr J K ) for all f ∈ M K ,

i.e., if E K = E J ◦ pr−J K1 .

Remark 4.2. For all J ∈ H let E J : M J → R be a pre-expectation. Then, the family (E J ) J ∈H is consistent if and only if

E K = E J ◦ pr−J K1

for all J , K ∈ H with K ⊂ J and | J | = | K | + 1. In fact, assume that E K = E J ◦ pr−J K1 for all J , K ∈ H with K ⊂ J and
| J | = | K | + 1. We prove that

E K = E J ◦ pr−J K1
for all J ∈ H with K ⊂ J by induction on n = | J | − | K | ∈ N0 . For n = 0 the statement is trivial. Now, assume that there
exists some n ∈ N0 such that

238
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

E K = E J ◦ pr−J K1

for all J ∈ H with K ⊂ J and | J | = | K | + n. Let J ∈ H with | J | = | K | + n + 1, J  := J \ {i } for some i ∈ J \ K and f ∈ M K .

Then, we have that g := f ◦ pr J  K ∈ M J  with g ◦ pr J J  = f ◦ pr J K . Therefore, by the induction hypothesis, we have that

E J ( f ◦ pr J K ) = E J ( g ◦ pr J J  ) = E J  ( g ) = E J  ( f ◦ pr J  K ) = E K ( f ).

The following theorem due to Denk et al. [17] is a finitely additive and nonlinear version of Kolmogorov’s extension
theorem.

Theorem 4.3. Let (E J ) J ∈H be a consistent family of pre-expectations E J : M J → R for J ∈ H . Then, there exists an expectation
E : L∞ ( S I , B I ) → R such that

E ( f ◦ pr J ) = E J ( f ) for all J ∈ H and all f ∈ M J .

If the pre-expectations E J are translation invariant, convex or sublinear for all J ∈ H , then E is translation invariant, convex or
sublinear, respectively.

Proof. Let M := f ◦ pr J : f ∈ M J , J ∈ H . Then M is a linear subspace of L∞ ( S I , B I ) with 1 ∈ M. For every J ∈ H and

f ∈ M J let E ( f ◦ pr J ) := E J ( f ). Since the family (E J ) J ∈H is consistent, the functional E : M → R is well-defined. Moreover,
E : M → R is a pre-expectation on M. The assertion now follows from Proposition 3.3. 

In Theorem 4.3, we proved the existence of an extension without any continuity properties and any structural assump-
tions. The following theorem due to Denk et al. [17] can be viewed as a continuous and convex version of Theorem 4.3.

Theorem 4.4. Let S be a Polish space and S be the Borel σ -algebra on S. For all J ∈ H let M J be a Riesz subspace of L∞ ( S J , B J )
with σ ( M J ) = B J and E J : M J → R be a convex
 pre-expectation,
 which is continuous from above. If the family (E J ) J ∈H is consistent,
then there exists a convex expectation space S I , B I , Ē with

E J ( f ) = Ē ( f ◦ pr J ) for all J ∈ H and f ∈ M J .

If the pre-expectations E J are sublinear or linear for all J ∈ H , then Ē can be chosen sublinear or linear, respectively.

Proof. Let M := f ◦ pr J : f ∈ M J , J ∈ H , and define E ( f ◦ pr J ) := E J ( f ) for all f ∈ M J and all J ∈ H . Since the family
(E J ) J ∈H is consistent, E : M → R defines a convex pre-expectation on M. Let μ ∈ M  with E ∗ (μ) < ∞. We will first show
that μ : M → R is continuous from above. Let μ J := μ ◦ pr−J 1 for all J ∈ H . Then, E ∗J (μ J ) ≤ E ∗ (μ) < ∞, and therefore
μ J : M J → R is continuous from above. By the theorem of Daniell-Stone, there exists a unique ν J ∈ ca1+ ( S J , B J ) with
ν J |M J = μ J for all J ∈ H . Let J , K ∈ H with K ⊂ J and f ∈ M K . Then,

μ K f = μ J ( f ◦ pr J K ) = ν J ( f ◦ pr J K ) for all f ∈ M K ,
and therefore

ν K f = ν J ( f ◦ pr J K ) for all f ∈ L∞ ( S K , B K )
as the extension of μ K to a probability measure is unique. By Kolmogorov’s extension theorem, there exists a unique
ν ∈ ca1+ ( S I , B I ) with ν ( f ◦ pr J ) = ν J f for all J ∈ H and f ∈ L∞ ( S J , B J ). Hence, we get that ν |M = μ, which implies
that μ : M → R is continuous from above as well. Using Remark 3.17, we thus obtain that there exists an expectation
Ē : L∞ (, σ ( M )), which extends E and results in a convex expectation space S I , σ ( M ), Ē .
It remains to show that B I = σ ( M ). It is clear that B I ⊃ σ ( M ). In order to show the inverse inclusion, let J ∈ H and
B J ∈ B J . Then, by assumption, B J ∈ σ ( M J ), and therefore we obtain that pr−J 1 ( B J ) ∈ σ ( M J ◦ pr J ) ⊂ σ ( M ). 

 
In the situation of Theorem 4.4, considering the canonical process (pr{i } )i ∈ I on the convex expectation space S I , B I , E ,
we obtain the following two corollaries on the existence of stochastic processes under nonlinear expectations.

Corollary 4.5. Let S be a Polish space and S be the Borel σ -algebra on S. For all J ∈ H let M J be a Riesz subspace of L∞ ( S J , B J )
with σ ( M J ) = B J and E J : M J → R be a convex pre-expectation which is continuous from above. Then, the following two statements
are equivalent:

(i) The family (E J ) J ∈H is consistent.

239
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

(ii) There exists a convex expectation space (, F , E ) and a stochastic process ( X i )i ∈ I with

E ( f ( X J )) = E J ( f )

for all J ∈ H and f ∈ M J , where X J := ( X i )i ∈ J .

Corollary 4.6. Let S be a Polish space and S be the Borel σ -algebra on S. For all J ∈ H let M J be a Riesz subspace of L∞ ( S J , B J )
with σ ( M J ) = B J and E J : M J → R be a sublinear pre-expectation which is continuous from above. Then, the following two state-
ments are equivalent:

(i) The family (E J ) J ∈H is consistent.

(ii) There exists a sublinear expectation space (, F , E ) and a stochastic process ( X i )i ∈ I with

E ( f ( X J )) = E J ( f )

for all J ∈ H and f ∈ M J , where X J := ( X i )i ∈ J .

We conclude this section with the following example on discrete-time Markov chains, which is taken from Denk et
al. [17].

Example 4.7. Let S be a finite state space and S := 2 S , so that L∞ ( S , B ) = R S . Let

P : L∞ ( S , B) → L∞ ( S , B) and μ0 : L∞ ( S , B) → R

be convex, constant preserving, i.e., P (α ) = α and μ0 (α ) = α for all α ∈ R, and monotone, i.e., P ( f ) ≤ P ( g ) and μ0 ( f ) ≤
μ0 ( g ) for all f , g ∈ L∞ ( S , B) with f ≤ g. For every k, l ∈ N0 with k < l, we define

Ek,l ( · , f ) := P l−k ( f ) for all f ∈ L∞ ( S , B). (4.1)

Let H := { J ⊂ N0 : # J ∈ N} be the set of all finite, nonempty subsets of N0 . For k ∈ N0 we define

E{k} ( f ) := μ0 (P k ( f )) for all f ∈ L∞ ( S , B),

where P 0 is the identity. For n ∈ N , k1 , . . . , kn+1 ∈ N0 with k1 < . . . < kn+1 and f ∈ L∞ ( S n+1 , Bn+1 ) we now define recur-
sively

E{k1 ,...,kn+1 } ( f ) := E{k1 ,...,kn } ( g ), (4.2)

where g (x1 , . . . , xn ) := Ekn ,kn+1 (xn , f (x1 , . . . , xn , · )) for all x1 , . . . , xn ∈ S. Then, E J : L∞ ( S J , B J ) → R is a convex expectation
and, therefore, continuous from above since S J is finite for all J ∈ H . By Remark 4.2, the family (E J ) J ∈H is consistent.
Hence, Theorem 4.4 or, more precisely, Corollary 4.5 implies that there exists a convex expectation space (, F , E ) and a
stochastic process ( X i )i ∈N0 with

E ( f ( X J )) = E J ( f )

for all J ∈ H and f ∈ M J , where X J := ( X i )i ∈ J . The process ( X i )i ∈N0 can be viewed as a convex version of a discrete-time
Markov chain. In fact, assume that P and μ0 are linear, i.e., P = P ∈ R S × S and μ0 ∈ R S . Then, E = E is a linear expectation
(w.r.t. a probability measure), and equations (4.1) and (4.2) translate to

   
E f ( Xl )|Fk = E f ( Xl )| Xk = μ0T P l−k f for k, l ∈ N0 with k ≤ l,

where Fk := σ ( X 0 , . . . , X k ). If P is sublinear, the set

μ ∈ R S × S : μ f ≤ P ( f ) for all f ∈ R S
induces a Markov-set chain, see Hartfiel [25], and the operator P coincides with the concept of a conditional coherent upper
prevision or upper transition operator, cf. De Cooman et al. [13] and Škulj [37].

240
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

5. Continuous-time Markov chains under nonlinear expectation

In this section, we consider time-homogeneous continuous-time Markov chains with a countable state space S (endowed
 2 ). We identify (measurable) functions S → R via sequences of the form (u i )i ∈ S ∈ R , and use
S S
with the discrete topology
the notation ∞ := L∞ S , 2 S . We call a (possibly nonlinear) map P : ∞ → ∞ a kernel if

(i) P u ≤ P v for all u , v ∈ ∞ with u ≤ v,

(ii) P α = α for all α ∈ R.

Here and throughout, we use the notation P u = P (u ) for a kernel P and u ∈ ∞ . In this section, we consider an arbitrary
family of linear Markov chains and construct a sublinear Markov chain, which can be interpreted as an imprecise version
of the family of Markov chains. The proofs and statements in this section rely on an approach proposed by Nisio [29],
which has been applied in various settings in order to construct Markov processes under model uncertainty, see e.g. Denk
et al. [18], Nendel [27], and Nendel and Röckner [28].  
Throughout, we consider the following setup: Let be a nonempty index set. For each λ ∈ , let S λ = S λ (t ) t ≥0 be the
family of transition probabilities of a time-homogeneous Markov chain, i.e.,

(i) S λ (t )i j ≥ 0 for all t ≥ 0 and i , j ∈ S,


(ii) j ∈ S S λ (t )i j = 1 for all t ≥ 0 and i ∈ S,
(iii) S λ (0) = I , where I is the identity,
(iv) S λ (s) S λ (t ) = S λ (s + t ) for all s, t ≥ 0.
 
For the family ( S λ )λ∈ , we construct a semigroup of sublinear kernels, i.e., a family S (t ) t ≥0 of sublinear kernels ∞ →
∞ with

(i) S (0) = I ,
(ii) S (s) S (t )u = S (s + t )u for all s, t ≥ 0 and u ∈ ∞ .

To this end, we consider the set of finite partitions

P := π ⊂ [0, ∞) : 0 ∈ π , |π | < ∞ .
For a partition π ∈ P , π = {t 0 , t 1 , . . . , tm } with 0 = t 0 < t 1 < . . . < tm we set

|π |∞ := max (t j − t j −1 ).
j =1,...,m

Moreover, we define |{0}|∞ := 0. The set of partitions with end-point t will be denoted by P t , i.e., P t := {π ∈ P : max π = t }.
Note that

P= Pt .
t ≥0

For all h ≥ 0 and u ∈ ∞ we define

Eh u := sup S λ (h)u ,
λ∈

where the supremum is taken componentwise. Then, Eh is well-defined since

 S λ (h)u ∞ ≤ u ∞
for all λ ∈ . Moreover, Eh is a sublinear kernel as it is monotone and Eh α = α for all α ∈ R. For a partition π =
{t 0 , t 1 , . . . , tm } ∈ P with m ∈ N and 0 = t 0 < t 1 < . . . < tm , we set

Eπ := Et1 −t0 . . . Etm −tm−1 .

Moreover, we set E{0} := E0 . Then, Eπ is a sublinear kernel for all π ∈ P as it is a concatenation of sublinear kernels. We
define

S (t )u := sup Eπ u
π ∈Pt
 
for all u ∈ ∞ and t ≥ 0, and call S (t ) t ≥0 the Nisio semigroup or (upper) semigroup envelope of ( S λ )λ∈ . Note that
S (t ) : ∞ → ∞ is well-defined and a sublinear kernel for all t ≥ 0 since Eπ is a sublinear kernel for all π ∈ P .

241
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

For h1 , h2 ≥ 0,

Eh1 +h2 u = sup S λ (h1 + h2 )u = sup S λ (h1 ) S λ (h2 )u ≤ sup S λ (h1 )Eh2 u = Eh1 Eh2 u ,
λ∈ λ∈ λ∈

which implies the inequality

Eπ1 u ≤ Eπ2 u (5.1)

for π1 , π2 ∈ P with π1 ⊂ π2 . The following lemma shows that S (t ) can be obtained by a pointwise monotone approxima-
tion with finite partitions letting the mesh size tend to zero.

Lemma 5.1. Let u ∈ ∞ and t ≥ 0. Then, there exists a sequence (πn )n∈N ⊂ P t with Eπn u ≤ Eπn+1 u for all n ∈ N and

Eπn u → S (t )u as n → ∞.

Proof. For t = 0 the statement is trivial. We therefore assume that t > 0. Then, for every i ∈ S, there exists a sequence
(πni )n∈N ⊂ P t with πni ⊂ πni +1 for all n ∈ N and
   
Eπ i u (i ) → S (t )u (i ) as n → ∞.
n

Since S is countable, there exists a sequence ( S n )n∈N with S n ⊂ S n+1 ⊂ S for all n ∈ N and S = n∈N S n . Let

πn := πni
i∈ Sn

for all n ∈ N . Then, πni ⊂ πn ⊂ πn+1 for all n ∈ N and i ∈ S n , and therefore

Eπ i ≤ Eπn ≤ Eπn+1
n

for all n ∈ N and i ∈ S n , which implies that S (t )u = supn∈N Eπn u. 

Proposition 5.2. S (s + t ) = S (s)S (t ) for all s, t ≥ 0.

Proof. Let u ∈ ∞ . If s = 0 or t = 0 the statement is trivial. Therefore, let s, t > 0, π0 ∈ P s+t and π := π0 ∪ {s}. Then, π ∈ P s+t
with π0 ⊂ π . Hence, by (5.1),

Eπ0 u ≤ Eπ u .
Let m ∈ N and 0 = t 0 < t 1 < . . . tm = s + t with π = {t 0 , . . . , tm } and k ∈ {1, . . . , m} with tk = s. Then, π1 := {t 0 , . . . , tk } ∈ P s
and π2 := {tk − s, . . . , tn − s} ∈ P t with

Eπ1 = Et1 −t0 · · · Et i −t i−1

and

Eπ2 = Et i+1 −t i · · · Etm −tm−1 .

Therefore,
  
Eπ0 u ≤ Eπ u = Et1 −t0 · · · Etm −tm−1 u = Et1 −t0 · · · Et i −t i−1 Et i+1 −t i · · · Etm −tm−1 u
= Eπ1 Eπ2 u ≤ Eπ1 S (t )u ≤ S (s)S (t )u .
Taking the supremum over all π0 ∈ P s+t , we obtain that S (s + t )u ≤ S (s)S (t )u.
Now, let (πn )n∈N ⊂ P t with Eπn u ≤ Eπn+1 u for all n ∈ N and Eπn u → S (t )u as n → ∞ (see Lemma 5.1), and fix π0 ∈ P s .
Then, for all n ∈ N ,

πn := π0 ∪ {s + τ : τ ∈ πn } ∈ P s+t
with Eπn = Eπ0 Eπn . As Eπ0 is continuous from below, we obtain that
 
Eπ0 S (t )u = lim Eπ0 Eπn u = lim Eπn u ≤ S (s + t )u .
n→∞ n→∞

Taking the supremum over all π0 ∈ P s , we see that S (s)S (t )u ≤ S (s + t )u. 

242
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

Remark 5.3. The semigroup S is the smallest semigroup that dominates the family ( S λ )λ∈ . In fact, let T be an upper
bound of the family ( S λ )λ∈ , i.e.

S λ (t )u ≤ T (t )u

for all λ ∈ , u ∈ ∞ and t ≥ 0. Then,

S λ (h)u ≤ Eh u ≤ T (h)u

for all λ ∈ , u ∈ ∞ and h ≥ 0. Since S λ and T are semigroups, it follows that

S λ (t )u ≤ Eπ u ≤ T (t )u

for all λ ∈ , u ∈ ∞ , t ≥ 0 and π ∈ P t . Taking the supremum over all π ∈ P t , we obtain that

S λ (t )u ≤ S (t )u ≤ T (t )u
for all λ ∈ , u ∈ ∞ and t ≥ 0.
 
The Nisio semigroup S (t ) t ≥0 has been used by various authors in the field of imprecise probability in order to describe
imprecise Markov chains via conditional upper previsions, cf. Škulj [38] and Krak et al. [26]. In fact, the operator S (t )
coincides with the conditional upper prevision over a time interval of the length t ≥ 0. However, one can actually go one
step further and extend the family of transition operators to a Markov chain on a canonical path space. In [27], this has
been done in the finite-state-case using the Theorem 4.4, which is due to Denk et al. [17]. However, in the case of infinitely
many states, the continuity from above is not trivial and leads to restrictions for the family of Markov chains indexed by
the set . Nevertheless, using an explicit dual representation of the Nisio semigroup one can extend the family of transition
operators to a Markov chain on the canonical path space without requiring the continuity from above.
In the following remark, we start by deriving a dual representation of the Nisio semigroup by viewing it as the cost
functional of an optimal control problem, where, roughly speaking, “nature” tries to control the system into the worst
possible scenario (using controls within the set ).

Remark 5.4. For λ = (λi )i ∈ S ∈ S

and t ≥ 0, let S λ (t ) : ∞ → ∞ be given by
   
S λ (t )u i
:= S λi (t )u i (5.2)

for all u 0 ∈ ∞ and i ∈ S. That is, S λ (t ) is the matrix whose i-th row is the i-th row of S λi (t ) for all i ∈ S. Here, the
interpretation is that, in every state i ∈ S, “nature” is allowed to choose a different Markov chain, which is indexed by
λ ∈ . We now add a dynamic component, and define


m 
 S
m
Q t := (λk , hk )k=1,...,m ∈ × [0, t ] : m ∈ N, hk = t .
k =1

Roughly speaking, the set Q t corresponds to the set of all (space-time discrete) admissible controls for the control set .
For θ = (λk , hk )k=1,...,m ∈ Q t with m ∈ N and u ∈ ∞ , we define

S θ u := S λ1 (h1 ) · · · S λm (hm )u ,

where S λk (hk ) is defined as in (5.2) for k = 1, . . . , m. Then, for all u ∈ ∞ ,

S (t )u = sup Eπ u = sup S θ u . (5.3)

π ∈Pt θ∈Q t

In fact, by definition of Q t , it follows that supθ∈ Q t S θ u ≤ S (t )u for all t ≥ 0 and u ∈ ∞ . On the other hand, let ε > 0
and π = {t 0 , . . . , tm } ∈ P t with 0 = t 0 < . . . < tm = t, and define hk := tk − tk−1 for k = 1, . . . , m. Choosing, by a backward
recursion, an m ε -optimizer in each coordinate of E · · · E u for k = m, . . . , 1, there exist λ , . . . , λ ∈ S with
hk hm 1 m

Eπ u = Eh1 · · · Ehm u ≤ S λ1 (h1 ) · · · S λm (hm )u + ε = S θ u + ε ,

where θ := (λk , hk )k=1,...,m ∈ Q t . Letting ε → 0 and taking the supremum over all π ∈ P t and θ ∈ Q t , yields S (t )u ≤
supθ∈ Q t S θ u for all t ≥ 0 and u ∈ ∞ .

The explicit dual representation from the previous remark can be used as in [18, Proposition 5.12] in order to obtain a
sublinear Markov chain in the following sense:

243
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

 
Definition 5.5. A (time-homogeneous) sublinear Markov chain is a quadruple , F , E , ( X t )t ≥0 , where

(i ) (, F ) is a measurable space.

(ii ) X t :  → S is F -measurable for all t ≥ 0.  
(iiii ) E = (Ei )i ∈ S , where (, F , Ei ) is a sublinear expectation space for all i ∈ S with Ei u ( X 0 ) = u (i ) for all u ∈ ∞.
 n+1 
(i v ) For all s, t ≥ 0, n ∈ N , 0 ≤ t 1 < . . . < tn ≤ s and v ∈ L∞ S n+1 , 2 S ,
    
E v 0 ( X t1 , . . . , X tn , X s+t ) = E S (t ) v 0 ( X t1 , . . . , X tn , · ) ( X s )
     
with S (t )u (i ) = Ei u ( X t ) for all u ∈ ∞ and i ∈ S. The family of kernels S (t ) t ≥0 is called the transition semigroup
 
of the Markov chain , F , E , ( X t )t ≥0 .

Remark 5.6. Assume that for all λ ∈ , there exists an infinite matrix qλ = (qλi j )i , j ∈ S ∈ R S × S such that, for all u ∈ ∞
 
 S λ (h)u − u 
 λ 
− q u → 0 as h  0.
 h ∞

Then, for each λ ∈ , qλ satisfies the following:

(i) qλii ≤ 0 for all i ∈ S,

(ii) qλi j ≥ 0 for all i , j ∈ S with i = j,

(iii) j∈ S qλi j = 0 for all i ∈ S.

Then, the above conditions imply that, for all λ ∈ ,

 
|qλi j | = −qλii + qλi j = −2qλii = 2|qλii | < ∞. (5.4)
j∈ S j ∈N\{i }

In particular, qλ : ∞ → ∞ is a bounded linear operator if and only if qλ  := 2 supi ∈ S |qλii | < ∞. If

1
sup sup |qλii | = sup qλ  < ∞, (5.5)
λ∈ i∈ S 2 λ∈
then, by [28, Example 6.7], it follows that
 
 S (h)u − u 
 − Q u  → 0 as h  0
 h 
∞

with Qu := supλ∈ qλ u for all u ∈ ∞ . Moreover, the function v : [0, ∞) → ∞ , t → S (t )u is the unique classical solution
to the differential equation

v  (t ) = Q v (t ), for all t ≥ 0, v (0) = u .

Therefore, S (t )u can be computed by solving the above differential equation. In the finite-state case, this approach has
been used by Škulj [38] and by Nendel [27] for Markov chains under convex expectations in order to define and compute
the transition operator as a solution to an ordinary differential equation. Krak et al. [26] use the explicit Euler method in
order to discretize the time interval, and come up with the transition operator as the limit of the Euler approximation. Both
approaches result in the same transition operator as shown in [26] and [27]. Examples for a family (qλ )λ∈ that satisfies
(5.5) are given by qλ = λq for λ ∈ , ⊂ [0, ∞) compact, and a fixed bounded operator q satisfying the above conditions
(i) - (iii). The
 operator
 q could, for example, be the generator of a Poisson process (with intensity 1). In this case, the Nisio
semigroup S (t ) t ≥0 can be viewed as the family of transition operators for an imprecise Poisson process with imprecision
in the intensity of the process, see also Erreygers and De Bock [21].

6. Conclusion

In the present paper, we have discussed and compared the concept of nonlinear expectations to other concepts that
are being used to describe model uncertainty or imprecision in a mathematical framework. In Section 2, we investigated
basic properties of nonlinear expectations, their acceptance sets, and their convex conjugates. Moreover, we indicated how
these concepts relate to concepts in the field of imprecise probability such as upper previsions, Choquet integrals, sets
of desirable gambles, and credal sets. We further showed how stochastic processes under nonlinear expectations can be
constructed using a Kolmogorov-type extension theorem for convex expectations (see Section 4). The latter is based on
general extension results for nonlinear expectations, which were discussed in Section 3. Here, similar to the construction

244
M. Nendel International Journal of Approximate Reasoning 130 (2021) 226–245

of the Lebesgue measure or the Daniell-Stone Theorem, the basic question is how to extend a pre-expectation, which
might a priori be defined on a relatively small set of bounded measurable functions, to the set of all bounded measurable
functions for a suitable σ -algebra. In Section 5, we addressed the existence of imprecise versions of Markov chains with
countable state space. The key ingredient in this context is the proper definition of a family of kernels satisfying the so-
called Chapman-Kolmogorov equations (semigroup property). This is done using an approach proposed by Nisio, where the
idea is to consider a worst-case scenario among various models (indexed by a set ) of linear Markov chains on a finer
and finer partition of a time interval. The resulting kernels are intimately related to stochastic control problems as we have
seen in Remark 5.4. Moreover, we indicated how these kernels and thus the finite-dimensional marginals of the resulting
imprecise Markov chain can be computed numerically by solving an ordinary differential equation in infinite dimensions;
an approach that has been used by various authors in the field of imprecise probability.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have
appeared to influence the work reported in this paper.

References

[1] P. Artzner, F. Delbaen, J.-M. Eber, D. Heath, Coherent measures of risk, Math. Finance 9 (3) (1999) 203–228.
[2] D. Bartl, Conditional nonlinear expectations, Stoch. Process. Appl. 130 (2) (2020) 785–805.
[3] D. Bartl, P. Cheridito, M. Kupper, L. Tangpi, Duality for increasing convex functionals with countably many marginal constraints, Banach J. Math. Anal.
11 (1) (2017) 72–89.
[4] H. Bauer, Probability Theory, De Gruyter Studies in Mathematics, vol. 23, Walter de Gruyter & Co., Berlin, 1996. Translated from the fourth (1991)
German edition by Robert B. Burckel and revised by the author.
[5] S. Cerreia-Vioglio, F. Maccheroni, M. Marinacci, A. Rustichini, Niveloids and their extensions: risk measures on small domains, J. Math. Anal. Appl.
413 (1) (2014) 343–360.
[6] P. Cheridito, H.M. Soner, N. Touzi, N. Victoir, Second-order backward stochastic differential equations and fully nonlinear parabolic PDEs, Commun. Pure
Appl. Math. 60 (7) (2007) 1081–1110.
[7] G. Choquet, Theory of capacities, Ann. Inst. Fourier 5 (131–295) (1955) 1953–1954.
[8] G. Choquet, Forme abstraite du théorème de capacitabilité, Ann. Inst. Fourier 9 (1959) 83–89.
[9] S.N. Cohen, R.J. Elliott, Solutions of backward stochastic differential equations on Markov chains, Commun. Stoch. Anal. 2 (2) (2008) 251–262.
[10] S.N. Cohen, R.J. Elliott, Comparisons for backward stochastic differential equations on Markov chains and related no-arbitrage conditions, Ann. Appl.
Probab. 20 (1) (2010) 267–311.
[11] F. Coquet, Y. Hu, J. Mémin, S. Peng, Filtration-consistent nonlinear expectations and related g-expectations, Probab. Theory Relat. Fields 123 (1) (2002)
1–27.
[12] J. De Bock, The limit behaviour of imprecise continuous-time Markov chains, J. Nonlinear Sci. 27 (1) (2017) 159–196.
[13] G. de Cooman, F. Hermans, E. Quaeghebeur, Imprecise Markov chains and their limit behavior, Probab. Eng. Inf. Sci. 23 (4) (2009) 597–635.
[14] F. Delbaen, Coherent risk measures, Cattedra Galileiana. [Galileo Chair]. Scuola Normale Superiore, Classe di Scienze, Pisa, 2000.
[15] F. Delbaen, Coherent risk measures on general probability spaces, in: Advances in Finance and Stochastics, Springer, Berlin, 2002, pp. 1–37.
[16] C. Dellacherie, P.-A. Meyer, Probabilities and Potential, North-Holland Mathematics Studies., vol. 29, North-Holland Publishing Co., Amsterdam, New
York, 1978.
[17] R. Denk, M. Kupper, M. Nendel, Kolmogorov-type and general extension results for nonlinear expectations, Banach J. Math. Anal. 12 (3) (2018) 515–540.
[18] R. Denk, M. Kupper, M. Nendel, A semigroup approach to nonlinear Lévy processes, Stoch. Process. Appl. 130 (3) (2020) 1616–1642.
[19] D. Denneberg, Non-additive Measure and Integral, Theory and Decision Library. Series B: Mathematical and Statistical Methods, vol. 27, Kluwer Aca-
demic Publishers Group, Dordrecht, 1994.
[20] N. Dunford, J.T. Schwartz, Linear Operators. Part I General Theory, Wiley Classics Library, John Wiley & Sons, Inc., New York, 1988. With the assistance
of William G. Bade and Robert G. Bartle, Reprint of the 1958 original, a Wiley-Interscience Publication.
[21] A. Erreygers, J. De Bock, First steps towards an imprecise Poisson process, in: Proceedings of Machine Learning Research, vol. 103, Thagaste, Ghent,
Belgium, 03–06 Jul 2019, PMLR, pp. 175–184.
[22] H. Föllmer, A. Schied, Convex measures of risk and trading constraints, Finance Stoch. 6 (4) (2002) 429–447.
[23] H. Föllmer, A. Schied, Stochastic Finance. An Introduction in Discrete Time, extended edition, Walter de Gruyter & Co., Berlin, 2011.
[24] M. Frittelli, E.R. Gianin, Putting order in risk measures, J. Bank. Finance 26 (7) (2002) 1473–1486.
[25] D.J. Hartfiel, Markov Set-Chains, Lecture Notes in Mathematics, vol. 1695, Springer-Verlag, Berlin, 1998.
[26] T. Krak, J. De Bock, A. Siebes, Imprecise continuous-time Markov chains, Int. J. Approx. Reason. 88 (2017) 452–528.
[27] M. Nendel, Markov chains under nonlinear expectation, Math. Finance (2020), https://doi.org/10.1111/mafi.12289, forthcoming.
[28] M. Nendel, M. Röckner, Upper envelopes of families of Feller semigroups and viscosity solutions to a class of nonlinear Cauchy problems, Preprint,
arXiv:1906.04430, 2019.
[29] M. Nisio, On a non-linear semi-group attached to stochastic optimal control, Publ. Res. Inst. Math. Sci. 12 (2) (1976/1977) 513–537.
[30] E. Pardoux, S. Peng, Backward stochastic differential equations and quasilinear parabolic partial differential equations, in: Stochastic Partial Differential
Equations and Their Applications, Charlotte, NC, 1991, in: Lect. Notes Control Inf. Sci., vol. 176, Springer, Berlin, 1992, pp. 200–217.
[31] R. Pelessoni, P. Vicig, Convex imprecise previsions, Reliab. Comput. 9 (6) (2003) 465–485. Special issue on dependable reasoning about uncertainty.
[32] S. Peng, Nonlinear expectations and nonlinear Markov chains, Chin. Ann. Math., Ser. B 26 (2) (2005) 159–184.
[33] S. Peng, G-expectation, G-Brownian motion and related stochastic calculus of Itô type, in: Stochastic Analysis and Applications, in: Abel Symp., vol. 2,
Springer, Berlin, 2007, pp. 541–567.
[34] S. Peng, Multi-dimensional G-Brownian motion and related stochastic calculus under G-expectation, Stoch. Process. Appl. 118 (12) (2008) 2223–2253.
[35] H.M. Soner, N. Touzi, J. Zhang, Martingale representation theorem for the G-expectation, Stoch. Process. Appl. 121 (2) (2011) 265–287.
[36] H.M. Soner, N. Touzi, J. Zhang, Quasi-sure stochastic analysis through aggregation, Electron. J. Probab. 16 (67) (2011) 1844–1879.
[37] D. Škulj, Discrete time Markov chains with interval probabilities, Int. J. Approx. Reason. 50 (8) (2009) 1314–1329.
[38] D. Škulj, Efficient computation of the bounds of continuous time imprecise Markov chains, Appl. Math. Comput. 250 (2015) 165–180.
[39] B.Z. Vulikh, Introduction to the Theory of Partially Ordered Spaces, Wolters-Noordhoff Scientific Publications, Ltd., Groningen, 1967. Translated from
Russian by Leo F. Boron, with the editorial collaboration of Adriaan C. Zaanen and Kiyoshi Iséki.
[40] P. Walley, Statistical Reasoning with Imprecise Probabilities, Monographs on Statistics and Applied Probability, vol. 42, Chapman and Hall, Ltd., London,
1991.

245