Professional Documents
Culture Documents
Fundamental Connections Between Utility Theories of Wealth and Information Theory
Fundamental Connections Between Utility Theories of Wealth and Information Theory
the advantage provided by side information in betting tasks for utility theories of wealth ratios.
Moreover, we show that the Ilić-Djordjević conditional entropy [1] satisfies a type of generalised
chain rule, which generalises that of Arimoto-Rényi. Finally, we address the implications of these
results on the quantum resource theories of informative measurements and non-constant channels.
Altogether, these results further help strengthening the bridge between the theory of expected utility
from the economic sciences and Shannon’s theory of information.
I. INTRODUCTION
describing behavioural tendencies of rational agents such
as the agent’s aversion to risk. Another important ob-
ject of study, and related to the utility function, is that
Information theory (IT) is a theoretical framework of the certainty equivalent. The certainty equivalent (for
which deals with the manipulation, storage, and process- a given lottery) represents the certain amount of wealth,
ing of information, and which was formalised by Shannon goods, or services, which the rational agent is willing to
in 1948 [2], with important earlier contributions made by accept, so in order to walk away from participating in the
Nyquist and Hartley during the 1920’s [3, 4]. One of the given lottery. In other words, it is the certain amount of
main objects of study here is that of entropic quantities, wealth that is equivalent (from the agent’s point of view)
which come in the form of: unconditional and conditional to the lottery in question. The characterisation of agents’
entropies, mutual information measures, capacities, di- behavioural tendencies has arguably been the bread and
vergences, conditional divergences, amongst others [5]. butter, so to speak, of the economic sciences since their
In addition to their role within IT itself, these quantities formal introduction during the 1950’s and throughout
have been found to play fundamental roles in the devel- the second half of the previous century, with many of
opment of scientific disciplines such as: thermodynamics, these efforts being recognised with the Sveriges Riks-
statistical mechanics, black hole thermodynamics, quan- bank Prize in Economic Sciences in Memory of Alfred
tum gravity, quantum information theory, biological and Nobel. Whilst initially a concept of interest mostly to
social sciences, amongst others [5]. It therefore becomes the economic sciences, the behavioural tendencies of ra-
of foundational and practical importance to keep explor- tional agents turned out to be a ubiquitous concept that
ing the mathematical properties as well as the physical has emerged and found usefulness in other scientific dis-
and operational significance of such information-theoretic ciplines such as behavioural ecology, neuroscience, and
quantities. information theory, with this latter case being the main
Expected utility theory (EUT) on the other hand, is a focus of this manuscript.
theoretical framework first formalised by von Neumann
and Morgenstern within the theory of games and eco- One of the earliest examples of the interplay between
nomic behaviour in 1944 [6]. In broad terms, EUT deals expected utility theory and information theory, to the
with the behaviour of rational agents when faced with de- best of our knowledge, was addressed in the work of
cision problems. One of the main objects of study within Kelly in 1956 [7]. In this work, Kelly formalised and
EUT is that of the utility function (of a rational agent), characterised the operational tasks of betting on horse
a real-valued function representing the agent’s level of races, or horse betting for short and, in particular, ad-
satisfaction when acquiring an amount of wealth, goods, dressed a notion of wealth which was found to be related
or services. Utility functions are particularly useful at to Shannon’s information-theoretic quantities such as the
entropy and the mutual information [5, 7]. Whilst the
concept of risk-aversion was not explicitly addressed as
an ingredient in Kelly’s original proposal, it is now under-
∗ andres.ducuara@yukawa.kyoto-u.ac.jp stood that the scenario introduced by Kelly corresponds
† paul.skrzypczyk@bristol.ac.uk to a risk-averse gambler with constant relative risk aver-
2
sion (CRRA) given by R = 1 (relative risk aversion being cations of these results, about the utility theory of wealth
one way to quantify the agent’s aversion to risk). Fast ratios, on quantum state betting games (QSB) within the
forward to the twenty-first century, in 2020 Bleuler, Lapi- quantum resource theory (QRT) of informative measure-
doth, and Pfister (BLP) took an important step in [8], ments, as well as noisy QSB games within the QRT of
by considering horse betting more generally for gamblers non-constant channels.
with risk-aversion values spanning R ∈ R≥0 ∪ {+∞}. In This work is organised as follows. In Sec. II we
addition to this, BLP introduced a new conditional Rényi start with preliminaries of the background theories. In
divergence and a new mutual information measure, which Sec. II A we address some information-theoretic quan-
helped characterising horse betting scenarios where the tities and introduce new conditional Rényi divergences,
gambler has access to side information [8]. Later on, whilst in Sec. II B we describe expected utility theory,
another step was taken in 2022 in [9], where the the au- the concept of risk-aversion, the certainty equivalent, and
thors introduced a framework for: negative wealth, risk- prospect theory. In Sec. III we start our results sec-
aversion values spanning the whole extended line of real tions. In Sec. III A we address operational tasks based
numbers R ∈ R ∪ {+∞, −∞}, as well as explicitly identi- on betting, risk-aversion, and double side information.
fying the role of the certainty equivalent in horse betting. In Sec. III B we characterise betting tasks with double
The authors in [9] also showed that Arimoto’s mutual side information. In Sec. III C we consider a generalisa-
information measure quantifies the usefulness that side tion to original prospect theory. In Sec. III D we intro-
information provides for these tasks, and extended horse duce an utility theory framework for wealth ratios, and
betting to the quantum domain, where families of ad- in Sec. III E a characterisation of betting tasks in such
ditional betting tasks naturally emerge: quantum state a scenario. In Sec. III F we address the implications of
betting, quantum channel/subchannel betting, amongst these results on the QRTs of informative measurements
other variants. The current manuscript aims to be a spir- and non-constant channels. We finish in Sec. IV with
itual successor of this line of research. conclusions.
In this work we further explore and expand the scope
of the connections between expected utility theory and
information theory by showing that the above-mentioned II. PRELIMINARIES
cases are not isolated examples, but that these two the-
ories are further intimately connected at a fundamen- In this section we address some generalised entropic
tal level. We do this by uncovering additional scenarios measures, the theory of expected utility, the concept of
where both expected utility theory and information the- risk-aversion, risk aversion measures, the certainty equiv-
ory jointly describe decision problems and betting tasks alent, and prospect theory.
for utility theories of wealth. Our main findings are the
following.
First, within the utility theory of wealth, we address A. Information-theoretic quantities
betting tasks where both gambler and bookmaker have
access to side information, or betting tasks with double One important generalisation of the entropic quanti-
side information for short. The exploration of these op- ties à la Shannon [2] is the extension due to Rényi [11].
erational tasks leads us to the identification of new Rényi In this direction, the Shannon entropy and the Kullback-
conditional divergences which precisely characterise the Leibler (KL) divergence (or KL relative entropy) are gen-
advantage provided by side information in such betting eralised as the Rényi entropy and the Rényi divergence,
tasks. In particular, we find that side information be- respectively [11, 12]. Interestingly however, there are in-
ing available to the bookmaker can actually help the stead various Rényi alternatives for the conditional en-
gambler, but that this advantage cannot be larger than tropy [13], as well as for the mutual information such as
when the gambler has direct access to such side infor- the proposals of: Arimoto [14], Sibson [15], Csiszár [16],
mation. Second, we further generalise these results to Lapidoth-Pfister [17], Tomamichel-Hayashi [18], amongst
the so-called original prospect theory (OPT), a gener- others [1]. Whilst all of these mutual information mea-
alised version of EUT, in which decision-making agents sures have found usefulness within information theory
are allowed to behave irrationally, albeit in a systematic [1], Arimoto’s measures stand out as strong candidates
manner [10]. In this regard, we prove that a specific due to them satisfying a considerable number of desirable
Rényi conditional entropy characterises such deviation properties [13].
from rationality. Third, we introduce an utility theory The previous addressed extension à la Rényi is however
of wealth ratios and prove that the recently introduced not the only way to generalise Shannon’s entropy. The
Ilić-Djordjević measure of mutual information [1] quan- Tsallis-Havrda-Charvát entropy, first introduced in the
tifies the advantage provided by side information in bet- context of information theory by Havrda and Charvát
ting tasks for such an utility theory. Moreover, we prove [19], and later on independently introduced in the realm
that the Ilić-Djordjević conditional entropy [1] satisfies a of statistical physics by Tsallis [20], is a non-additive gen-
type of generalised chain rule, which generalises that of eralisation of Shannon’s. More generally, the two param-
Arimoto-Rényi. Fourth and finally, we address the impli- eter (q, r) entropy introduced by Sharma and Mittal [21]
3
recovers both Rényi and Tsallis in the limits (q, 1) and 28] and, similarly, the so-called Landsberg-Vedral en-
(q, q), respectively. Similarly to the Rényi case however, tropy (q, 2 − q) [29–31]. We now move on to the informa-
whilst there is a generally accepted unique way of in- tion theoretic quantities, based upon the Sharma-Mittal
troducing a unconditional entropy à la Tsallis as well as entropy, recently introduced by Ilić and Djordjević [1].
à la Sharma-Mittal, here again one important challenge
is the generalisation of additional information-theoretic Definition 2. (Ilić-Djordjević information theoretic
quantities [1]) The Ilić-Djordjević conditional entropy
quantities. In this regard, an important step was recently
taken in the work of Ilić and Djordjević [1], where various and mutual information measures of order (q, r) ∈ R × R
ID
entropic measures à la Sharma-Mittal were introduced: of a joint PMF pXG are denoted by Hq,r (X|G) and
ID
a conditional entropy, a mutual information measure, Iq,r (X; G) respectively. The orders q, r ∈ (−∞, 0) ∪
and a channel capacity. In particular, the Ilić-Djordjević (0, 1) ∪ (1, ∞) are defined by:
measures reduce to the Arimoto-Rényi conditional en-
tropy, Arimoto’s mutual information, and Arimoto-Rényi
! 1 q(1−r)
(1−q)
q
ID sgn(q) X X
p(x, g)q
capacity, in the limit (q, 1). Furthermore, these mea- Hq,r (X|G) := − 1 , (3)
1−r
sures happen to satisfy a considerable number of desir- g x
with HαR (·) the Rényi entropy. The orders α ∈ Definition 5. (Second new conditional Rényi diver-
{1, 0, ∞, −∞} are defined define by their respective con- gence) The second new conditional Rényi divergence (n2-
tinuous extensions. CR-divergence) of order α ∈ R of PMFs pX|GY , qX|GY ,
pG , and pY |G is denoted as Dαn2 (pX|GY ||qX|GY |pG , pY |G ).
Further details about the relationship between the The orders α ∈ (−∞, 0) ∪ (0, 1) ∪ (1, ∞) are defined as:
Rényi divergence, conditional Rényi divergences, and
Rényi conditional entropies in Appendix A. We now in- n2 |α|
Dα (pX|GY ||qX|GY |pG , pY |G ) := (11)
troduce two new conditional Rényi divergences, address α−1
1
!α !
some of their properties, as well as the way they relate X X X 1−α
α
Dα (pX ||qX ) ≤ Dαn1 (pX|Y ||qX|Y |pY ), (9) Dαn2(pX|GY ||qX|GY |pGY )
P
with the PMFs pX , qX given by p(x) := y p(x|y)p(y), (X, Y ) ⊥
⊥G X ⊥⊥ Y
P
q(x) := y q(x|y)p(y).
Dαn1(pX|Y ||qX|Y |pY ) DαBLP(pX|G||qX|G|pG)
The proof of this lemma is in Appendix C. It will also
prove useful to consider the following inequality between X ⊥⊥ Y X ⊥⊥ G
the n1-CR-divergence and the BLP-CR-divergence. Dα(pX ||qX )
Lemma 2. (Inequality between the n1-CR divergence FIG. 1. Relation between the new conditional Rényi diver-
gences, the BLP-CR divergence, and the Rényi divergence.
and the BLP-CR divergence) Consider conditional PMFs C
pX|Y , qX|Y , PMF pY , and orders α ∈ (−∞, 0) ∪ (0, 1) ∪ A → B means that “quantity A becomes equal to quantity B
when condition C holds”. The notation X ⊥⊥ Y means that
(1, ∞), we have:
X is independent of Y .
Dαn1 (pX|Y ||qX|Y |pY ) ≤ DαBLP (pX|Y ||qX|Y |pY ), (10)
rational agents and their prospects of earning, losing, Qualitatively, the CE represents how attractive the lot-
hoarding, and interchanging goods like wealth, services, tery is to the rational agent. In other words, it represents
and similar assets. A central object of study here is the minimum certain amount of wealth that the rational
the concept of utility function, a real-valued function agent is willing to receive, so to get persuaded of not play-
u : A → R, with A a set of “alternatives” endowed with ing the game. We remark here that the CE is a quantity
a binary relation ≺. The general idea is for the utility that depends both on the lottery (pX , wX ) as well as on
function to represent the level of satisfaction of a ratio- the agent’s utility function u, and so we can explicitly
nal agent (u(a) ∈ R) with the prospect of acquiring a write this as:
specific alternative (a ∈ A). The binary relation a ≺ b
can be regarded as “alternative b represents more wealth wuCE (pX , wX ) = u−1 (EpX (u(wX ))) . (17)
or more services than alternative a”. Now, it is then nat-
ural for rational agents to be more pleased with b than The CE stands out as an important quantity because op-
with a, and so this should be reflected in the agent’s util- timising it is equivalent to optimising the expected utility,
ity function as u(a) < u(b), which therefore imposes the due to the utility function being strictly increasing and
utility function to be a monotone for the binary relation. because u(wCE ) = EpX [u(wX )]. Taking the latter into
This is an appealing way of formalising utility, since it account, and that the CE has units of wealth [w] ($, £,
is natural for a rational agent to regard more goods or ¥, ...), it is usually better to consider the CE, instead of
wealth, as something generally better than less goods or the expected utility, as the figure of merit for the setups
wealth. involving rational agents placing bets that we will address
In this work we deal with the set of alternatives being later on. In addition to this, the CE helps establishing
either wealth, or wealth ratios (to be introduced later on the characterisation of risk tendencies of rational agents
in the document), and so it is enough to consider an inter- [9]. Using the CE, the concept of risk-aversion in the
val in the real numbers A ⊆ R, with the binary relation context of utility theory emerges via the highly nontriv-
being the standard relation <. In this section we then ial realisation that the risk-attitude of a rational agent
address the quintessential example of an utility theory of is related to the concavity (risk-averse), convexity (risk-
wealth. The utility function being monotonic gets trans- seeking), or linearity (risk-neutral) of the agent’s utility
lated into being an strictly increasing function and, addi- function [33–36] (a detailed derivation for both positive
tionally, we ask for it to be twice-differentiable, for math- and negative wealth is in [9]). One common measure of
ematical convenience, and because it is natural to con- risk-aversion is the so-called relative risk aversion (RRA)
sider that small changes in wealth imply small changes measure given by [34–36]:
in the agent’s satisfaction. We note here that the utility 00
u (w)
function does not necessarily need to be positive or nega- RRAu (w) := −w . (18)
tive, since the idea is for it to compare alternatives only. u0 (w)
We also note that whilst one could exclusively work with
This quantifier is dimensionless, which is a characteris-
positive wealth, one can also construct scenarios where
tic not satisfied by all quantifiers of risk-aversion [34–36].
wealth is negative, and so the utility function would then
The RRA measure does not assign a global value for risk
naturally represent the level of (dis)satisfaction of the ra-
aversion, so we ask here for utility functions where the
tional agent when having to pay or give back the amount
agent’s RRA is constant. We can solve (18) assuming
of wealth in question [9].
RRA(w) = R, leading to the constant relative risk aver-
Another important aspect of the theory of expected sion (CRRA) function, or isoelastic utility function, for
utility is the behaviour of a rational agent in the presence both positive and negative wealth as:
of uncertainty. Consider an random event distributed ac-
cording to the PMF pX on a finite alphabet X , and a uIR (w) = sgn(w) lnR (|w|), (19)
distribution of wealth given by wX , with w(x) ∈ R≥0 ,
∀x ∈ X . The pair (pX , wX ) is often refereed to as a lot- with the auxiliary “sign” function (1) and the R-
tery. Fix a lottery (pX , wX ) and consider a game where deformed natural logarithm as:
a rational agent, represented by a utility function u, is (
asked to predict outcome x ∈ X , which occurs with prob- ln x, R = 1,
ability p(x), and for which is rewarded (after correctly lnR (x) := x1−R −1 (20)
1−R , R 6= 1,
guessing) an amount of wealth w(x) ∈ R≥0 . The ex-
pected utility of the rationalPagent playing such a game with the deformation parameter R ∈ R. The parame-
is given by EpX (u(wX )) = x u(w(x))p(x). Using this ter R varies from minus to plus infinity, describing all
expected utility, a central object of interest is that of possible risk tendencies of the rational agent, for either
the certainty equivalent (CE). The CE is defined as the positive or negative wealth [9]. The certainty equivalent
amount of wealth wCE which has an utility that exactly of the isoelastic function, or isoelastic certainty equivalent
matches the expected utility of the lottery as: (ICE), the figure of merit in this work, is given by:
event distributed according to the PMF pX . Similar to a values R ∈ (−∞, 1) ∪ (1, ∞) can be written as:
decision problem, the referee is going to ask the gambler ICE
to decide between a lottery or a fixed (certain) amount wR (bX|G , oX|Y , pXGY )
= (uIR )−1 EpXGY [uIR (wXGY )]
of wealth. The lottery in question is given by a pair (27)
(pX , wX ), with wX a distribution of wealth. The main " 1
# 1−R
difference with a standard decision problem is that now X 1−R
the gambler can play a more active role when it comes = b(x|g) o(x|y) p(x, g, y) . (28)
g,x,y
to the amount of wealth that he is going to receive. This
is because the wealth to be rewarded is now considered a The cases R ∈ {1, ∞, −∞} are defined by continu-
function with two components as wX = bX oX , with oX a ous extension of (28). A betting task is specified by
function refereed to as the odds (not necessarily a PMF), (oX|Y , pGXY ), and the gambler plays this game with a
which is proposed by the referee, and a PMF bX , which betting strategy bX|G .
represents the bet placed by the gambler. The idea is
that after correctly guessing the outcome of the random We now address some particular cases of interest.
event, the referee rewards the gambler with an amount Remark 2. (Betting tasks with individual side informa-
of wealth given by w(x) = b(x)o(x). It is then explicit tion) It is useful to specify here two natural scenarios.
that the gambler can influence the amount of wealth to First, a scenario where only the gambler has access to
be rewarded. In this decision problem involving betting, side information (or betting task with gambler’s side in-
or betting task for short, the quantity that describes the formation):
gambler’s assessment of the lottery (pX , wX ) is the cer- " # 1
1−R
tainty equivalent of wealth given by: ICE
wR (bX|G , oX , pXG ) =
X
b(x|g) o(x)
1−R
p(x, g) . (29)
g,x
wuCE (bX , oX , pX ) = u−1 (EpX [u(wX := bX oX )]) , (26)
Second, a scenario where only the bookmaker has access
which represents the minimum amount of wealth that to side information (or betting task with bookmaker’s side
the gambler is willing to accept so to walk away from information):
the lottery. We can similarly introduce scenarios dealing " # 1
1−R
with negative wealth, where the decision problem and ICE
wR (bX , oX|Y , pXY ) =
X
b(x) o(x|y)
1−R
p(x, y) . (30)
consequently the betting task is now dealing with losses y,x
instead of gains. This scenario can naturally be included
by allowing the odds function to be negative (full details One useful concept to introduce here is that of the fair-
about this construction in [9]). ness of the odds [5, 8]. Given a conditional odds function
One common characteristic amongst works in the lit- oX|Y , we consider the quantity:
erature addressing betting tasks so far, is the assump- !−1
tion of side information being available exclusively to (o)
X
−1
c (y) := |o(x|y)| , (31)
the gambler [7–9]. In this work we further extend this to
x
include a more elaborate scenario in which both gambler
and bookmaker can now have access to side information. and classify the fairness of the odds as follows:
We coin this scenario as betting tasks with double side
information, and formalise it as follows. c(o) (y) = 1, ∀y, Fair odds. (32)
(o)
c (y) > 1, ∀y, Superfair odds, (33)
Definition 6. (Betting tasks with double side informa-
(o)
tion) Consider a random event described by the PMF pX . c (y) < 1, ∀y, Subfair odds. (34)
Consider also random variables G (gambler’s side infor-
Finally, we allow the gambler to optimise over all possible
mation) and Y (bookmaker’s side information) correlated
betting strategies, for a given game setup (oX|Y , pXGY ):
with the random variable X and described by the joint
PMF pXGY . Consider now a game (betting task) with a ICE
max wR (bX|G , oX|Y , pXGY ), (35)
referee and a gambler (rational agent), with the latter be- bX|G
ing represented by an isoelastic utility function uIR (·) with
with the maximisation over all conditional PMFs. It is
a constant relative risk aversion (CRRA) factor R ∈ R. also going to be useful to introduce the following auxiliary
The game consists on the gambler placing “a bet” with function, which we here simply address as the logarithm
the help of side information G as bX|G (a conditional of the ICE:
PMF) on the output of the random event. The referee ICE
then rewards the gambler (when guessing correctly) with UR (bX|G , oX|Y , pXGY ) (36)
h i
an amount of wealth given by w(x, g, y) = b(x|g) o(x|y), := sgn(o) ln wR ICE
(bX|G , |oX|Y |, pXGY ) .
with oX|Y the conditional “odds”, a pre-established func-
tion proposed by the referee (not necessarily a PMF), and We are now ready to address our main results concerning
known to the gambler. The figure of merit of interest here the characterisation of these operational tasks in terms
is the isoelastic certainty equivalent (ICE), which for risk of information-theoretic quantities.
8
B. Information-theoretic characterisation of with sgn(o) = sgn(R). From the previous corollary, and
betting games with double side information using the fact that the n1-CR-divergence reduces to the
Rényi divergence when X ⊥⊥ Y , we have:
In this subsection we address the characterisation of (o)
various types of betting tasks in terms of information- max URICE (bX , oX , pX ) = D1/R (pX ||rX ). (40)
bX
theoretic quantities. We start with betting tasks with
bookmaker’s side information (only the bookmaker has With these two corollaries in place, we can now anal-
access to side information), followed by general betting yse the performance of a rational agent when playing a
tasks with double side information (both gambler and betting task with bookmaker’s side information in com-
bookmaker have access to side information). We com- parison to a scenario with no side information.
pare the performance of a rational agent when playing Corollary 3. (Ratio of betting games between book-
betting tasks where there is bookmaker’s side informa- maker’s side-information and no side-information) Con-
tion, against a scenario with no side information at all. sider R ∈ [1, +∞], with the
Similarly, we then compare the performance of a rational P odds function for each betting
game related as o(x) := y o(x|y)p(y), then:
agent when playing betting tasks where there is book-
maker’s side information against a scenario with gam-
max wRICE
(bX , oX|Y , pXY )
bler’s side information. ln X
b
(41)
ICE
max wR (bX , oX , pX )
Result 1. (Characterisation of betting games with book- bX
maker’s side information) Consider a betting game de- = n1
D1/R
(o) (o)
||rX|Y |pY ) − D1/R (pX ||rX ) ≥ 0,
(pX|Y
fined by the pair (oX|Y , pXY ), with fair odds as c(o) (y) =
−1
P
x |o(x|y)| = 1, ∀y. Consider now a gambler play-
P
with the PMFs pX , rX given by p(x) := y p(x|y)p(y),
ing this game with a betting strategy given by bX . Then, P
r(x) := y r(x|y)p(y).
we have that the logarithm of the ICE is characterised by
the n1-CR divergence Dαn1 (·|| · |·) and the R-divergence Details about this corollary in Appendix G. The in-
Dα (·||·) as: equality in the corollary follows from the data processing
inequality derived in Lemma 1. This corollary is telling
URICE bX , oX|Y , pXY
(37) us that it is more useful for the gambler when the book-
n1 (o) maker has access to side information, as opposed to no
= sgn(o) sgn(R) D1/R (pX|Y ||rX|Y |pY ) side-information at all. This is a rather counterintuitive
− sgn(o) sgn(R) DR (hX
(R,o,p)
||bX ), statement, because one can naively expect that it should
actually be worse for the gambler, when the bookmaker is
with the PMFs r(o) (x|y) := |o(x|y)|−1 , and: the one having access to side information. This is how-
ever, as the maths is telling us, not the case. The ex-
P R1 planation of this seemingly paradoxical situation is that,
y p(xy) |o(x|y)|1−R even though the bookmaker is the agent explicitly having
h(R,o,p) (x) := R1 . (38) access to the side information, we should not forget that
P P
x0 y p(x0 y) |o(x0 |y)|1−R she is still in charge of proposing the conditional odds
function oX|Y , which itself depends on Y , and that this
The proof of this result is in Appendix E. This result odds in turn are going to be known to the gambler (as
characterises betting tasks with bookmaker’s side infor- per the rules of the betting game), so that ultimately the
mation. In particular, a consequence of this result is gambler does have access to the side information Y , al-
the explicit form of the optimal betting strategy for the though via this implicit manner. From a technical point
gambler to play these games, as well as an operational of view, it is also interesting that the inequality holds for
interpretation of the n1-CR divergence. R ≥ 1, and so it could prove insightful to understand its
behaviour for other values, but we leave this however for
Corollary 1. (Optimal betting strategy for betting tasks
future research. We now can also compare rational agents
with bookmaker’s side information) Consider a betting
scenario with sgn(o) = sgn(R). We can readily check, playing betting tasks with bookmaker’s side information
from the previous decomposition, that the optimal betting and gambler’s side information.
(R,o,p)
strategy is given by b∗X = hX , which achieves the Corollary 4. (Ratio of betting games between
quantity: bookmaker’s side-information and gambler’s side-
ICE n1 (o) information) Consider R ∈ R then:
max UR (bX , oX|Y , pXY ) = D1/R (pX|Y ||rX|Y |pY ). (39)
bX
ICE
max wR (bX|G , oX , pXG )
We can also recover the known case where there is no bX|G
sgn(R) ln (42)
ICE
side information as a corollary [8].
max wR (bX , oX|G , pXG )
bX
Corollary 2. (Optimal betting strategy for betting tasks BLP (o) n1 (o)
= D1/R (pX|G ||rX|G |pG ) − D1/R (pX|G ||rX|G |pG ) ≥ 0,
with no side information) Consider a betting scenario
9
with thePodds function for each betting game related as C. Generalisation to original prospect theory
o(x) := g o(x|g)p(g).
Details about this corollary in Appendix H. This corol- In this section we address betting tasks from the point
lary is telling us that side-information is more useful of view of prospect theory. In particular, we analyse here
when given to the gambler directly, as opposed to when betting tasks with no side information, as well as betting
it is given to the bookmaker. This confirms what one tasks with gambler’s side information. Additionally, we
can intuitively expect, with the added benefit that this consider here the certainty equivalent for prospect the-
corollary allows us to quantify the advantage provided ory with the isoelastic utility function together with the
by giving the side information directly to the gambler. power probability weighting function.
We now move on to the more general scenario of betting
tasks with double side information.
Result 3. (Characterisation of betting tasks with no side
Result 2. (Characterisation of betting games with dou- information within prospect theory) Consider a betting
ble side information) Consider a betting game defined by game specified by the pair (oX , pX ), and a gambler repre-
the pair (oX|Y , pXGY ), and fair odds so that c(o) (y) = sented by the pair (R, S) (R is the standard CRRA whilst
P −1 S is the coefficient related to the agent’s sensitivity when
x |o(x|y)| = 1, ∀y. Consider also a gambler play-
ing this game with a betting strategy bX|G . Then, the assessing probabilities) playing this game with a betting
logarithm of the ICE is characterised by the the n2-CR- strategy bX . The logarithm of the ICE within prospect
divergence Dαn2 (·|| · |·) and the R-divergence Dα (·||·) as: theory is characterised by the R-divergence Dα (·||·) and
the Rényi entropy Hα (·) as:
UR (bX|G , oX|Y , pXGY ) (43)
h i
n2 (o) CE−PT
= sgn(o) sgn(R)D1/R pX|GY ||rX|Y |pG , pY |G sgn(o) ln wR,S (bX , |oX |, pX ) (46)
1−S
(R,o,p) (R,o,p) (R,o,p)
− sgn(o) sgn(R)DR hX|G hG bX|G hG , = sgn(o) sgn(S) HS (pX ) + sgn(o) log |c(o) |+
1−R
(S,p) (o)
with the conditional PMF r(o) (x|y) := |o(x|y)|−1 , and the + sgn(o) sgn(R) D1/R (qX ||rX )
conditional PMF and PMF: (R,S,o,p)
− sgn(o) sgn(R) DR (hX ||bX ),
P 1
1−R R
y p(xy|g) |o(x|y)|
h(R,o,p) (x|g) := 1 , (44) with the parameter and the PMFs:
0 y|g) |o(x0 |y)|1−R R
P P
x0 y p(x
1 R !−1
P P 1−R R c(o)
p(g) y p(xy|g) |o(x|y)|
X
x c(o) := o(x)−1 , r(o) (x) := , (47)
h(R,o,p) (g) := 1 R . x
o(x)
P P P R
g0 p(g 0 ) p(xy|g 0 ) |o(x|y)|1−R 1 1−R
x y q(x) R o(x) R
h(R,S,o,p) (x) := P 1 1−R , (48)
0) R o(x0 )
x0 q(x
R
The proof of this result is in Appendix F. Similar to
S
the previous case, we can also identify here the optimal p(x)
q (S,p) (x) := P . (49)
betting strategy to be implemented by a gambler playing 0 S
x0 p(x )
these games, and provide an operational interpretation
of the n2-CR divergence. (o) (R,S,o,p)
The functions rX and hX define valid PMFs even
Corollary 5. (Optimal betting strategy for betting tasks for negative odds (o(x) < 0, ∀x).
with double side information) Consider a betting scenario
with sgn(o) = sgn(R). We can readily check, from the
previous decomposition, that the optimal betting strategy The proof of this result is in Appendix I. We now con-
(R,o,p)
is given by b∗X|G = hX|G , which achieves the quantity: sider a CE of wealth with the isoelastic utility function
and the power probability weighting function in a sce-
nario with gambler’s side information. We find the fol-
max URICE (bX|G , oX|Y , pXGY ) (45)
bX|G lowing characterisation.
n2 (o)
= D1/R (pX|GY ||rX|Y |pG , pY |G ).
Result 4. (Characterisation of betting tasks with gam-
From this general result, and using the relationship bler’s side information within prospect theory) Consider
between conditional divergences, we can recover the pre- a betting game specified by the pair (oX , pXG ), and a
vious corollaries as particular cases. We now move on gambler represented by the pair (R, S) playing this game
to further generalise these results, from the theory of ex- with a betting strategy bX|G . The logarithm of the ICE is
pected utility to prospect theory. characterised by the R-divergence Dα (·||·), the BLP-CR
10
divergence DαBLP (·||·), and the Rényi entropy Hα (·) as: We note that reducing to standard expected utility the-
h i ory (by specifying S = 1), the second terms vanishes,
CE−PT
sgn(o) ln wR,S (bX|G , |oX |, pXG ) (50) and we recover the Result 5 in [9]. This provides an op-
erational interpretation for the second Rényi conditional
1−S entropy; it quantifies the deviation achieved when con-
= sgn(o) sgn(S) HS (pXG ) + sgn(o) log |c(o) |+
1−R sidering prospect theory against standard expected util-
BLP (S,p) (o) (S,p) ity theory. Finally, we note that we can similarly define
+ sgn(o) sgn(R) D1/R (qX|G ||rX |qG )+
variants for betting tasks with bookmaker’s side infor-
(R,S,o,p) (R,S,o,p) (R,S,o,p)
− sgn(o) sgn(R) DR (hX|G hG ||bX|G hG ), mation, and double side information and derive similar
decompositions. This finishes our main results concern-
with the parameter and the PMF: ing the utility theory of wealth. We now consider a utility
!−1 theory of wealth ratios.
X c(o)
c(o) := o(x)−1 , r(o) (x) := , (51)
o(x)
x D. Utility theory of wealth ratios
new utility theory based on such wealth ratios, which can Ilić-Djordjević conditional entropy of orders (q, r) ∈ R ×
be interpreted as the level of satisfaction that a rational R satisfies the condition:
agent would experience with the prospect of achieving ID
h
SM
i
such an advantage. The quantities which naturally arise sgn(q)Hq,r (X|G) ≥ sgn(q) Hq,r (XG) r lnr (K) , (63)
in such utility theory based on wealth ratios are then
given by: with r the pseudo-substraction.
The proof of this result is in Appendix L. In particular,
v A1,2 b1X , b2X , oX , pX ,
u (60) we recover a known case in the literature as a limit case.
SI
v Au (oX , pXG ) , (61)
Corollary 7. We recover the case for the Arimoto-Rényi
conditional entropy. For r = 1, q ≥ 0, HqA (X|G) ≥
with v : [1, ∞) → R a utility function for wealth ratios,
and u(·) a standard utility function for wealth. We note HqR (X) − log(K) [13].
that these two utility functions can be different, as ra- Considering this property, together with the opera-
tional agents can have a specific attitude with respect to tional interpretation in terms of betting tasks, the Ilić-
wealth, and a different when considering the growth of Djordjević conditional entropy stands out as a very ap-
such wealth. Whilst this formalism can be introduced for pealing conditional entropy à la Sharma-Mittal. In ad-
general utility functions (u, v), for simplicity, let us as- dition to this, it is also worth highlighting that its as-
sume rational agents described by isoelastic utility func- sociated mutual information satisfies various desirable
tions as u(·) = uIR (·), v(·) = uIS (·), with different attitude properties, which are not simultaneously valid for other
to risk as S 6= R. We are now ready to describe betting alternatives à la Sharma-Mittal [1]. Altogether, these
tasks from the point of view of utility theories of wealth results therefore make the quantities introduced by Ilić-
ratios. Djordjević stand as strong generalisations of entropic
quantities à la Sharma-Mittal. We now address the im-
plications of these latter results on the field of quantum
E. Sharma-Mittal information-theoretic quantities resource theories.
and the EUT of wealth ratios
In this section we characterise the advantage provided F. Implications on the QRTs of informative
by side information within utility theories of wealth ra- measurements and non-constant channels
tios. We then derive a generalised chain rule for the ID
conditional entropy. We now consider the implications of the previous
results on quantum resource theories, specifically, the
Result 5. (Operational interpretation of the Ilić- QRTs of measurement informativeness and non-constant
Djordjević mutual information measure) The Ilić- channels. A quantum state is a trace one (Tr(ρ) = 1)
ID
Djordjević mutual information measure Iq,r (X; G) of or- positive semidefinite operator (ρ ≥ 0) acting on a finite-
ders (q, r) ∈ R × R of a joint PMF pXG is given by: dimensional Hilbert space. A positive operator-valued
ID
measure (POVM) or quantum measurement is a set of
Iq,r (X; G) (62) positive semidefinite operators M = {Ma ≥ 0} such that
P
sgn(q)c
a Ma = 1. An ensemble of quantum states is a set
ICE
max w1/q (bX|G , oX , pXG )
bX|G E = {ρx , p(x)}, x ∈ {1, ..., K}, with pX a PMF. A quan-
= sgn(q) ur .
ICE sgn(q)c
max w1/q (bX , oX , pX ) tum channel is a completely positive trace-preserving
bX (CPTP) map [46].
In the framework of quantum resource theories
The Ilić-Djordjević entropy quantifies the utility of a r- (QRTs), it is common to first specify the mathematical
agent on the wealth ratio (advantage) provided by side objects of the theory, followed by a property of such ob-
information G achieved by a 1/q-agent when the latter jects to be considered as a resource [47]. We invoke here a
plays betting tasks where odds are constant. QRT of measurements with the resource of informative-
The proof of this result is in Appendix K. This re- ness [48]. A POVM M = {Mg } is called uninformative
sult holds true for any joint PMF pXG as well as for any when there exists a PMF qG such that the POVM effects
combination of orders (q, r) ∈ R × R. Finally, due to the can be written as Mg = q(g)1, ∀g [48, 49]. The set of
generality of this result, we can derive operational inter- uninformative measurements is denoted as UI. We also
pretations for the generalised mutual information mea- invoke here a QRT of channels with the resourceful chan-
sures of Rényi, Tsallis, Gauss, when taking the respec- nels being non-constant channels. A quantum channel N
tive limits. We now derive a generalised chain rule for is called constant when there exists a state σN such that
the Ilić-Djordjević conditional entropy. N (ρ) = σN , ∀ρ. The set of constant channels is denoted
as C. There are two natural betting tasks to be con-
Result 6. (Generalised chain rule for the Sharma-Mittal sidered in the quantum domain, quantum state betting
entropy and the Ilić-Djordjević conditional entropy) The (QSB) and noisy quantum state betting (NQSB) [9].
12
One pragmatical way of introducing quantum state sgn(q) C, C > 0, ∀x, an ensemble of states E =
betting (QSB) games is to consider a standard betting {ρx , p(x)}, and a quantum channel N . Consider a first
game where the conditional PMF pG|X is now given by rational agent (gambler) represented by the isoelastic util-
p(g|x) = Tr[Mg ρx ] for a given measurement M = {Mg } ity function u1/q . We now compare the performance of
and a given ensemble of states E = {ρx , p(x)}. Opera- this gambler having access to the quantum channel N , in
tionally speaking, here the referee prepares the ensem- comparison to being allowed to implement any possible
constant channel N 0 ∈ N . In each case the gambler is
ble of states E, sends one of these states to the gam-
allowed to play the betting task with the optimal betting
bler (say ρx ), and then the gambler’s goal is to try to strategy, and to optimise over all possible measurements.
identify the received state, by performing the measure- We also remember here that the gamblers are interested
ment M = {Mg }, and then using the outcome of this in maximising the isoelastic certainty equivalent (ICE).
measurement to propose a betting strategy b(x|g) [9]. Finally, consider a rational agent represented by a utility
In a similar manner, in a noisy quantum state betting function ur , who assesses the performance of the advan-
(NQSB) game, the conditional PMF pG|X is given by tage achieved by the 1/q-agent. This latter assessment is
p(g|x) = Tr[Mg N (ρx )], with the operational interpreta- given by the ID mutual information measure as:
tion being same as for a QSB game, with the exception ID
Iq,r (X; G)E,N = (66)
that the state the gambler receives is now affected by the
NQSB sgn(q)c
(potentially noisy) channel N [9]. With these considera-
max max w1/q
M bX|G
bX|G , M, oX , E, N
tions in place, Result 5 gets translated as follows. sgn(q) ur
.
NQSB sgn(q)c 0
max
0
max max w1/q bX|G , N, oX , E, N
N ∈C bX|G
Result 7. Consider a QSB game defined by the pair N
sgn(q)c
(oX , E) with constant odds as osgn(q)c (x) := sgn(q) C, This result follows from Result 5 and from the ob-
C > 0, ∀x, and an ensemble of states E = {ρx , p(x)}. servation that constant channels are not able to ex-
Consider a first rational agent (gambler) represented by tract information contained in the received state since
the isoelastic utility function u1/q . We now compare the p(g|x) = Tr[Mg N (ρx )] = Tr[Mg σN ] = q(g), and then we
performance of this gambler when the gambler uses a can check that:
fixed measurement M = {Mg }, in comparison to being al-
NQSB sgn(q)c 0
lowed to implement any possible uninformative measure- max
0
max max w1/q b X|G , N, oX , E, N (67)
N ∈C N bX|G
ment N ∈ UI. We remember here that the gamblers are
ICE sgn(q)c
interested in maximising the isoelastic certainty equiva- = max w1/q (bX , oX , pX ).
bX
lent (ICE), and so in each case the gambler is allowed
to play the betting task with the optimal betting strategy.
Finally, consider a second rational agent represented by
IV. CONCLUSIONS
an isoelastic utility function ur , who assesses the advan-
tage (wealth ratio) achieved by the first 1/q-agent. This
latter assessment is given by the ID mutual information In this work we establish fundamental connections be-
measure as: tween utilities of wealth and information theory. Specif-
ically, we derive results for general utility theories of
ID
Iq,r (X; G)E,M (64) wealth as well as for utility theories of wealth ratios.
First, regarding utility theories of wealth, we introduce
sgn(q)c
ICE
max w1/q bX|G , M, oX ,E new operational tasks in the form of betting tasks in
bX|G
= sgn(q) ur
.
which both gambler and bookmaker have access to side in-
ICE sgn(q)c
max max w1/q bX|G , N, oX ,E formation, or betting tasks with double side information
N∈UI bX|G
for short. This, as an extension of works in the literature
where it is usually assumed that only the gambler has
This result follows from Result 5 and from the obser-
access to side information. In particular, we show the
vation that uninformative measurements do not allow for
seemingly counterintuitive fact that, betting tasks with
the extraction of valuable information contained in the
bookmaker’s side information can be more advantageous
received state since p(g|x) = Tr[Mg ρx ] = q(g) Tr[1ρx ] =
(for the gambler) than betting tasks without any side in-
q(g), and therefore we can check that:
formation at all. This apparently paradoxical situation
ICE
sgn(q)c
can be explained by remembering that whilst in this sce-
max max w1/q bX|G , N, oX ,E (65) nario the bookmaker is indeed the only one having access
N∈UI bX|G
sgn(q)c
to side information, the bookmaker still needs to make
ICE
= max w1/q (bX , oX , pX ). the odds function public (as per the rules of the betting
bX
task) and therefore, the gambler can still implicitly have
We now address the noisy quantum state betting (NQSB) access to the side information in question. Notwithstand-
games and the QRT of non-constant channels. ing this however, we moreover prove that this scenario
cannot be more advantageous (for the gambler) than re-
Result 8. Consider a NQSB game defined by the tu- ceiving the side information directly. We prove this by
sgn(q)c
ple (oX , E, N ) with constant odds as osgn(q)c (x) := linking the scenarios of bookmaker’s side information and
13
The orders α ∈ {0, 1, ∞, −∞} are defined by their respec- for x ≥ 0 and with the deformation parameter r ∈ R.
tive continuous extensions. We have the pseudo-additivity identity:
Remark 3. ([8, 32]) Relationship between CR- lnr (xy) = lnr (x) + lnr (y) + (1 − r) lnr (x) lnr (y). (B2)
divergences and the Rényi divergence. For any condi-
tional PMFs pG|X , qG|X , and any PMF pX we have: It is also convenient to have:
S
Dα (pG|X ||qG|X |pX ) = Dα (pG|X pX ||qG|X pX ), (A7) x
lnr = [lnr (x) − lnr (y)] y 1−r . (B3)
y
C
X
Dα (pG|X ||qG|X |pX ) = p(x)Dα (pG|X=x ||qG|X=x ), (A8)
x
BLP
In particular we have:
Dα (pG|X ||qG|X |pX )
" # 1
=
α
log
X α−1
p(x)2 α Dα (p G|X=x ||qG|X=x )
. (A9) lnr = − lnr (y) y 1−r . (B4)
α−1 y
x
For the q-exponential we have:
Remark 4. (Relationship between CR-divergences and
conditional entropies) exq eyq = eqx+y+(1−q)xy , (B5)
DαS (pX|G ||uX |pG ) Hα4 (X|G),
1
= sgn(α) ln K − (A10) exq := (1 + (1 − q)x) 1−q
(B6)
DαC (pX|G ||uX |pG ) = sgn(α) ln K − Hα1 (X|G), (A11)
These functions are important because many of the quan-
DαBLP (pX|G ||uX |pG ) = sgn(α) ln K − HαA (X|G). (A12) tities considered in this work can be written in terms of
r-deformed logarithms. For instance, the isoelastic util-
These identities can be seen as the Rényi conditional ity function can be written as:
counterpart of the Rényi unconditional relationship:
uR (w) = sgn(w) lnR (|w|). (B7)
Dα (pX ||uX ) = sgn(α) ln K − HαR (X). (A13)
It is also going to be useful to introduce the function:
We address these two remarks in Fig. 2.
(
x, r = 1,
Dα(·||·) ηr (x) := e(1−r)x −1 (B8)
1−r , r 6= 1.
with the quantities: The first line is the definition of the Rényi divergence. In
the second line we replace the PMFs pX and qX . In the
1
third line we reorganise the expression so to use Hölder’s
! (q−1)
X
pq (X) := p(x)q , (B17) inequality. In the fourth line we use Hölder’s inequality
x with p = 1/α and q = 1/(1 − α). In the fifth line we
q
! q1 (q−1) reorganise so to use Jensen’s inequality. In the sixth line
X X we use Jensen’s inequality with the function f (·) = (·)α
pq (X|G) := p(x, g)q . (B18)
which is convex on R≥0 . In the seventh line we iden-
g x
tify the n1-CR-divergence of the conditional PMFs pX|Y ,
Finally, we will also need some useful inequalities. qX|Y , and PMF pY . The factor sgn(α)(α − 1) is negative
Jensen’s inequality: for α ∈ (0, 1), so this reverses the inequality and get:
f (EpX (X)) ≤ EpX (f (X)), f convex. (B19) Dα (pX ||qX ) ≤ Dαn1 (pX|Y ||qX|Y |pY ). (C10)
1 1
E(|XY |) ≤ E(|X|p ) p E(|Y |q ) q , (B20)
Appendix D: Proof of Lemma 2
1 1
for p, q ∈ [1, ∞] , p + q = 1. Minkowski inequality:
Proof. We start with the first new conditional Rényi di-
!p ! 1 !1
X X p X X p vergence. Considering α > 1 we have:
a(x, y) ≤ a(x, y)p , p > 1, (B21)
x y y x sgn(α)(α − 1)Dαn1 (pX|Y ||qX|Y |pY ) (D1)
!p ! 1 !1 !α
X X p X X p
1−α
a(x, y)p , 1
X X
a(x, y) ≥ p < 1, (B22) = ln p(y) p(x|y) q(x|y) α , (D2)
x y y x
x y
!α ! α1
for non-negative values {a(x, y)}. 1−α
2
X X
= α ln p(y) p(x) q(x|y) α , (D3)
x y
Appendix C: Proof of Lemma 1 ! α1
(M) X X 1−α
α
≤ α ln p(y) p(x|y) q(x|y) α , (D4)
Proof. Considering the PMFs: y x
X X ! α1
p(x) := p(x|y)p(y), q(x) := q(x|y)p(y). (C1) 4
X X
= α ln p(y) p(x|y)α q(x|y)1−α , (D5)
y y
y x
!α !1−α p(y) p(x|y) q(x|y) α , and we also take into account that
3
X X 1
α α
X 1 log(·) is increasing. In the fourth line we reorganise. In
= p(y)(p(x|y) ) p(y)(q(x|y)1−α ) 1−α , the fifth line we identify the BLP-CR-divergence. This
x y y
then proves the case for α > 1. We now analyse the
(C5)
additional cases.
(H)
There are three factors to consider when analysing the
XX
≥ p(y) p(x|y)α q(x|y)1−α
, (C6)
x y direction of the inequality. First, whether α > 1 or α < 1
5
XX 1−α α
so to use Minkowski inequality. Second, whether α > 0
= p(y) p(x|y)q(x|y) α , (C7)
or α < 0 because this factor multiplies the whole expres-
x y
!α sion. Third, the sign of the factor sgn(α)(α − 1). Con-
(J)
≥
X X
p(y) p(x|y)q(x|y)
1−α
α , (C8)
sidering now the case α < 1 we have that the Minkowski
x y
inequality gets reversed, and depending on sgn(α) we ei-
7 n1
(pX|Y ||qX|Y |pY )
ther keep the inequality or further reverse it again, and
= esgn(α)(α−1)Dα . (C9) then a similar analysis with the factor sgn(α)(α − 1). In
16
summary, a careful inspection of the two remaining cases (α < 0, and 0 < α < 1) yield the inequality:
p(xy)|o(x|y)|1−R
P
R y
h(x) := . (E1)
P P R1 R
x0 y p(x0 y) |o(x0 |y)|1−R
and so:
! R1 R
X X X
p(xy) |o(x|y)|1−R = p(x0 y) |o(x0 |y)|1−R h(x)R . (E2)
y x0 y
In the first line we use the definition of the ICE. In the second line we group terms and sum over y. In the third line
we replace the previous equality. In the fourth line we reorganise. In the fifth line we identify the Rényi divergence
of order R. In the sixth and final line we replace the PMF rX|Y and identify the n1-CR divergence. We can then
multiply both sides of the equality by sgn(o) and this finishes the proof.
p(xy|g)o(x|y)1−R
P
R y
h(x|g) := . (F1)
P P R1 R
x0 y p(x0 y|g) o(x0 |y)1−R
17
and so:
! R1 R
X X X
p(xy|g) o(x|y)1−R = p(x0 y|g) o(x0 |y)1−R h(x|g)R . (F2)
y x0 y
! ! R1 R
X X X
p(g) p(xy|g) o(x|y)1−R = p(g) p(x0 y|g) o(x0 |y)1−R h(x|g)R . (F3)
y x0 y
and so:
! R1 R ! R1 R
X X X X X
h(g) p(g 0 ) p(xy|g 0 ) o(x|y)1−R = p(g) p(xy|g) o(x|y)1−R . (F5)
g0 x y x y
In the first line we use the definition of the ICE. In the second line we group terms and sum over y. In the third line
we replace the previous equality. In the fourth line we reorganise. In the fifth line we identify the Rényi divergence
of order R. In the sixth and final line we replace the PMF rX|Y and identify the n1-CR divergence. We can then
multiply both sides of the equality by sgn(o) and this finishes the proof.
18
ICE
maxbX wR (bX , oX|Y , pXY )
sgn(R) ln ICE
max wR (bX , oX , pX )
bX
ICE
1 sgn(R) maxbX sgn(R) wR (bX , |oX|Y |, pXY )
= sgn(R) ln ICE
, (G1)
sgn(R) max sgn(R) wR (bX , |oX |, pX )
bX
2 ICE ICE
= sgn(R) ln sgn(R) max sgn(R)wR (bX , |oX|Y |, pXY ) − sgn(R) ln sgn(R) max sgn(R)wR (bX , |oX |, pX ) , (G2)
bX bX
3 ICE ICE
= max sgn(R) ln wR (bX , |oX|Y |, pXY ) − max sgn(R) ln wR (bX , |oX |, pX ) , (G3)
bX bX
4
= max URICE (bX , oX|Y , pXY ) − max URICE (bX , oX , pX ), (G4)
bX bX
5 n1 (o) (o)
= D1/R (pX|Y ||rX|Y |pY ) − D1/R (pX ||rX ). (G5)
In the first line we multiply and divide by sgn(R) inside the natural logarithm, and take out sgn(o) = sgn(R) from
the ICE, leaving the ICE with positive odds. In the second line we split the natural logarithm. In the third line we
take out the maxbX , which jumps through two sgn(R), and therefore keeps being a maximisation. In the third line we
remember that sgn(o) = sgn(R) and identify the functions URICE . In the fourth line we use Corollary 1 and Corollaries
6 and 7 from [9]. This finishes the statement.
ICE
maxbX|G wR (bX|G , oX , pXG )
sgn(R) ln ICE
max wR (bX , oX|G , pXG )
bX
ICE
1 sgn(R) maxbX sgn(R) wR (bX|G , |oX |, pXG )
= sgn(R) ln ICE
, (H1)
sgn(R) max sgn(R) wR (bX , |oX|G |, pXG )
bX
2 ICE ICE
= sgn(R) ln sgn(R) max sgn(R)wR (bX|G , |oX |, pXG ) − sgn(R) ln sgn(R) max sgn(R)wR (bX , |oX|G |, pXG ) ,
bX|G bX|G
(H2)
3 ICE
ICE
= max sgn(R) ln wR (bX|G , |oX |, pXG ) − max sgn(R) ln wR (bX , |oX|G |, pXG ) , (H3)
bX|G bX
4
= max URICE (bX|G , oX , pXG ) − max URICE (bX , oX|G , pXG ), (H4)
bX|G bX
5 BLP (o) n1 (o)
= D1/R (pX|G ||rX|G |pG ) − D1/R (pX|G ||rX|G |pG ). (H5)
In the first line we multiply and divide by sgn(R) inside the natural logarithm, and take out sgn(o) = sgn(R) from
the ICE, leaving the ICE with positive odds. In the second line we split the natural logarithm. In the third line we
take out the maxbX , which jumps through two sgn(R), and therefore keeps being a maximisation. In the third line we
remember that sgn(o) = sgn(R) and identify the functions URICE . In the fourth line we use Corollary 1 and Corollaries
6 and 7 from [9]. This finishes the statement.
19
Proof. Consider the natural logarithm of the certainty equivalent within prospect theory:
" #
h
CE−PT
i
1 1 X
S 1−R 1−R
ln VR,S (bX , |oX |, pX ) = ln p(x) |o(x)| b(x) , (I1)
1−R x
" ! #
2 1 X X p(x)S
0 S 1−R 1−R
= ln p(x ) P 0 S
|o(x)| b(x) , (I2)
1−R x0 p(x )
x0 xg
" # " #
3 1 X
0 S 1 X
1−R 1−R
= ln p(x ) + ln q(x) |o(x)| b(x) , (I3)
1−R 1−R
x0 x
4 1−S h
(S,p)
i
= sgn(S) HS (pX ) + ln VRCE−EU bX , |oX |, qX . (I4)
1−R
The first equality is the certainty equivalent within prospect theory. In the second inequality we multiply and divide
(S,p)
by the coefficient x0 p(x0 )S . In the third equality we reorganise the expression and identify the escort PMF qX .
P
In the fourth equality we identify the Rényi entropy of order S and the certainty equivalent within expected utility
theory. The final step is to input the expression for the CE-EU in terms of the Rényi divergence, and this finishes
the proof.
Appendix J: Proof of Result 4 (PT2)
Proof. Consider the natural logarithm of the certainty equivalent within prospect theory:
" #
h
CE−PT
i
1 1 X
S 1−R 1−R
ln VR,S (bX|G , |oX |, pXG ) = ln p(x, g) |o(x)| b(x|g) , (J1)
1−R xg
S
2 1 X X p(x, g)
= ln p(x0 , g 0 )S P 0 0 S
|o(x)|1−R b(x|g)1−R , (J2)
1−R 0 0 xg x0 g 0 p(x , g )
xg
" #
3 1 X 1 X
= ln p(x0 , g 0 )S + ln q(x, g) |o(x)|1−R b(x|g)1−R , (J3)
1−R 0 0
1−R xg
xg
4 1−S h
(S,p)
i
= sgn(S) HS (pXG ) + ln VRCE−EU bX|G , |oX |, qXG . (J4)
1−R
The first equality is
Pthe certainty equivalent within prospect theory. In the second inequality we multiply and divide
by the coefficient x0 g0 p(x0 , g 0 )S . In the third equality we reorganise the expression and identify the escort PMF
(S)
qXG . In the fourth equality we identify the Rényi entropy of order S and the certainty equivalent within expected
utility theory. The final step is input the known expression for this latter CE in terms of the Rényi divergence and
the BLP conditional Rényi divergence, and this finishes the proof.
Appendix K: Proof of Result 5 tropy of order q ∈ R for a joint PMF pXG reads:
ICE sgn(q)c
sgn(q) C pq (X) = max w1/q (bX , oX , pX ), (K1)
bX with the maximisation over all possible betting strategies
bX|G , and the odds osgn(q)c (x) := sgn(q) C, C > 0, ∀x.
with the maximisation over all possible betting strategies
bX , and the odds osgn(q)c (x) := sgn(q) C, C > 0, ∀x.
Lemma 4. ([9]) The Arimoto-Rényi conditional en- Proof. (of Result 5) Consider the Ilić-Djordjević mutual
20
−1 4
h i h i
−1
ID
Iq,r (X; G) lnr pq (X|G) ≥ ηr ln pq (XG) − ln [K] ,
SM
= Hq,r (X) r
ID
Hq,r (X|G), (K3) (L4)
5
h i
= sgn(q)[ηr sgn(q)HqR (X) r ηr sgn(q)HqA (X|G) ], −1
≥ ηr ln pq (XG) K −1 , (L5)
(K4) 6 h i
−1
= sgn(q)[ηr ln p−1
q (X)
r ηr ln p−1
q (X|G) ]. (K5) ≥ lnr pq (XG) K −1 . (L6)
[1] V. M. Ilić and I. B. Djordjević, Entropy 23, 702 (2021). [25] P. A. Kluza, Entropy 23, 1688 (2021).
[2] C. E. Shannon, The Bell System Technical Journal 27, [26] T. Frank and A. Daffertshofer, Physica A: Statistical Me-
379 (1948). chanics and its Applications 285, 351 (2000).
[3] H. Nyquist, The Bell System Technical Journal 3, 324 [27] T. Frank and A. Plastino, The European Physical Jour-
(1924). nal B - Condensed Matter 30, 543 (2002).
[4] R. V. L. Hartley, Bell System [28] T. Frank, Physica A: Statistical Mechanics and its Ap-
Technical Journal 7, 535 (1928), plications 340, 251 (2004).
https://onlinelibrary.wiley.com/doi/pdf/10.1002/j.1538- [29] P. T. Landsberg and V. Vedral, Physics Letters A 247,
7305.1928.tb01236.x. 211 (1998).
[5] T. M. Cover and J. A. Thomas, Elements of Information [30] A. K. Rajagopal and S. Abe, Phys. Rev. Lett. 83, 1711
Theory (Wiley, 2005). (1999).
[6] J. von Neumann and O. Morgenstern, Theory of Games [31] T. Yamano, Entropy 3, 280 (2001).
and Economic Behavior (60th Anniversary Commemo- [32] C. Pfister, On Renyi Information Measures and Their
rative Edition) (Princeton University Press, 2007). Applications, Doctoral thesis, ETH Zurich (2019).
[7] J. L. Kelly, The Bell System Technical Journal 35, 917 [33] D. Bernoulli, Econometrica 22, 23 (1954).
(1956). [34] K. Arrow, Aspects of the theory of risk-bearing (Yrjö
[8] C. Bleuler, A. Lapidoth, and C. Pfister, Entropy 22, 316 Jahnssonin Säätiö, 1965).
(2020). [35] J. W. Pratt, Econometrica 32, 122 (1964).
[9] A. F. Ducuara and P. Skrzypczyk, PRX Quantum 3, [36] B. de Finetti, Giornale degli Economisti e Annali di
020366 (2022). Economia 11, 685 (1952).
[10] D. Kahneman and A. Tversky, Econometrica 47, 263 [37] N. C. Barberis, Journal of Economic Perspectives 27, 173
(1979). (2013).
[11] A. Rényi, in Proceedings of the Fourth Berkeley Sympo- [38] A. Chiu and G. Wu, “Prospect theory,” (2011).
sium on Mathematical Statistics and Probability, Volume [39] G. Wu, J. Zhang, and R. Gonzalez, in Blackwell Hand-
1: Contributions to the Theory of Statistics (University book of Judgment and Decision Making (Blackwell Pub-
of California Press, Berkeley, Calif., 1961) pp. 547–561. lishing Ltd) pp. 399–423.
[12] T. van Erven and P. Harremos, IEEE Transactions on [40] C. Starmer, Journal of Economic Literature 38, 332
Information Theory 60, 3797 (2014). (2000).
[13] S. Fehr and S. Berens, IEEE Transactions on Information [41] A. Tversky and D. Kahneman, Journal of Risk and Un-
Theory 60, 6801 (2014). certainty 5, 297 (1992).
[14] S. Arimoto, Topics in Information Theory (1977). [42] R. Gonzalez and G. Wu, Theory and Decision 92, 647
[15] R. Sibson, Zeitschrift für Wahrscheinlichkeitstheorie und (2022).
Verwandte Gebiete 14, 149 (1969). [43] J. Pan, C. S. Webb, and H. Zank, Theory and Decision
[16] I. Csiszar, IEEE Transactions on Information Theory 41, 87, 201 (2019).
26 (1995). [44] M. Abdellaoui, O. L’Haridon, and H. Zank, Journal of
[17] A. Lapidoth and C. Pfister, Entropy 21, 778 (2019). Risk and Uncertainty 41, 39 (2010).
[18] M. Tomamichel and M. Hayashi, IEEE Transactions on [45] E. U. Weber, Psychological Bulletin 115, 228 (1994).
Information Theory 64, 1064 (2018). [46] M. A. Nielsen and I. L. Chuang, Quantum Computation
[19] J. Havrda and F. Charvát, Kybernetika 3, 30 (1967). and Quantum Information (Cambridge University Press,
[20] C. Tsallis, Journal of Statistical Physics 52, 479 (1988). 2000).
[21] B. D. Sharma and D. P. Mittal, J. Math. Sci. 10, 28 [47] E. Chitambar and G. Gour, Rev. Mod. Phys. 91, 025001
(1975). (2019).
[22] M. Masi, Physics Letters A 338, 217 (2005). [48] P. Skrzypczyk and N. Linden, Phys. Rev. Lett. 122,
[23] M. Masi, Physica A: Statistical Mechanics and its Appli- 140403 (2019).
cations 377, 67 (2007). [49] A. F. Ducuara and P. Skrzypczyk, Phys. Rev. Lett. 125,
[24] T. Wada and H. Suyari, Physics Letters A 368, 199 110401 (2020).
(2007).