IEEE, VOL. XX, NO.
XX, X 2021
Cyber-Physical-Systems and Secrecy Outage
Probability: Revisited
Makan Zamanipour, Member, IEEE
Abstract—This paper technically explores the secrecy rate Λ
and a maximisation problem over the concave version
of the
secrecy outage probability (SOP) as Max P𝓇 Λ ≥ 𝜆 . We do
Δ
this from a generic viewpoint even though we use a traditional
Wyner’s wiretap channel for our system model − something
arXiv:2201.01841v2 [cs.IT] 14 Jun 2022
that can be extended to every kind of secrecy modeling and
analysis. We consider a Riemannian mani-fold for  it and we
mathematically define a volume for it as Vℴ𝓁 Λ . Through
achieving a new bound for the Riemannian mani-fold and its
volume, we subsequently relate it to the number of eigen-
values existing in the relative probabilistic closure. We prove
in-between some novel lemmas with the aid of some useful
inequalities such as the Finsler’s lemma, the generalised Young’s
inequality, the generalised Brunn-Minkowski inequality, the Tala-
grand’s concentration inequality. We additionally propose a novel
Markov decision process based reinforcement learning algorithm
in order to find the optimal policy in relation to the eigenvalue
distributions − something that is extended to a possibilisitically
semi-Markov decision process for the case of periodic attacks.
Index Terms—𝑁 𝑃−hard, Alice, Bob, eigenvalue distribution,
eigenvector transition, Eve, generalised Brunn-Minkowski in-
equality, Hofer-Zehnder capacity, semi-Markov model, periodic
attack, possibility-theory, projection method, Talagrand’s concen-
tration inequality.
I. I NTRODUCTION
Hysical-layer security inevitably plays a vital role in 5/6
P G and beyond. This widely supported concept [1], [2], [3],
[4], [5], [6], [7], [8], [9], [10], [11], [12] is emerged in parallel
with traditional cryptography techniques while information-
theoretic perspectives are promising.
In order to simultaneously enhance the fairness and the
quality of service among all the users, the physical character-
istics of the wireless channel are of an absolutely inconsistent
nature, which originally comes from the channel’s broadcast
behaviour − something that should be essentially managed.
The concept of secrecy outage probability (SOP) in telecom-
munication still shows up an open research field in the
literature. This concept is useful e.g. for: free-space optical
communications [1], vehicular communications [2], reflecting
intelligent surfaces [3]-[4], cognitive networks [5], coopera-
tive communications [6], power-line communications [7], the
internet of Things [8], terrestrial networks [9], mobile edge
computing networks [10], molecular communications [11] and
under-water networks [12].
Manuscript received NOV, 2021; revised X XX, XXXX. Copyright (c) 2015
IEEE. Personal use of this material is permitted. However, permission to use
this material for any other purposes must be obtained from the IEEE by
sending a request to pubs-permissions@ieee.org. Makan Zamanipour is with
Lahijan University, Shaghayegh Street, Po. Box 1616, Lahijan, 44131, Iran,
makan.zamanipour.2015@ieee.org.
1
In [1]-[12] and in totally various types of system models,
some novel and closed-form mathematical expressions have
been newly derived and proposed − some of them are optimi-
sation based, some of them are statistical oriented and some
of them are even jointly theoretical-practical.
A. Motivations and contributions
In this paper, we are interested in responding to the fol-
lowing question: How can we guarantee highly adequate
relaxations over the principle of SOP? With regard to the
non-complete version of the literature, the expressed question
strongly motivate us to find an interesting solution, according
to which our contributions are fundamentally described as
follows.
• (i) A new bound in relation to the maximisation prob-
lem over the SOP’s concave version is derived. We, in
addition, theoretically discuss about a totally novel inter-
pretation over the aforementioned maximisation problem
from a duality point of view. We consider a Riemannian
mani-fold for the SOP’s concave version and we math-
ematically define a volume for it for which we derive a
new bound. We use some insightful principles such as
•
Keyhole contour.
(ii) We subsequently relate the Riemannian mani-fold and
its bounded volume expressed above to the number of
eigen-values. We use in-between some useful lemmas and
inequalities such as the Finsler’s lemma, the generalised
Young’s inequality, the generalised Brunn-Minkowski in-
equality, the Talagrand’s concentration inequality.
• (iii) We additionally go over further discussions in terms
of the projection method technically relating it to the for-
mer parts. We also take into account the case of relaxing
the non-contractibility and how to decrease the relative
non-contractibility radius from a topological point of
view. In this context, we propose a novel Markov decision
process based reinforcement learning algorithm in order
to find the optimal policy in relation to the eigenvalue
distributions − something that was subsequently extended
to a semi-Markov decision one for the case of periodic
attacks and with regard to the possibility-theory.
B. General notation
The notations widely used throughout the paper is given in
Table I.
IEEE, VOL. XX, NO. XX, X 2021 2

TABLE I: List of notations.

Notation Definition Notation Definition

M𝑖𝑛 Minimisation M𝑎𝑥 Maximisation
E Expected-value I (·) Mutual-information
𝑑𝑒 𝑓
= Is defined as ≈ Is approximated to
Vℴ𝓁 Volume P𝓇 Probability
𝑑𝑒𝑡 (·) Matrix determinant (·)𝑇 Transpose
𝑇𝑟 [·] Trace of matrix (·) −1
Inverse of matrix
Λ Secrecy rate P𝓇 Λ ≥ 𝜆 SOP’s Concave Version
𝓉 Time ℋ Entropy
M Riemannian-manifold 𝜕M Boundary
𝜁 Eigen-value 𝑇𝑟 [·] Trace

legitimate reciever named Bob and an un-authorised one as

Bob
an eavesdropper named Eve, as shown in Fig. 4. The infor-
mation capacity of the communication system is theoretically
expressed by the general formula from Shannon. The secrecy
Alice capacity is interpreted as a bound of the security performance
of the communication system. We now have the following in-
equality for the secrecy capacity from an information-theoretic
Eve point of view
𝑑𝑒 𝑓


𝐶𝑠 = Max I X, A − I X, B
Fig. 1: A traditional Wyner’s wiretap channel: System diagram 𝑓X ( 𝑥)
(1)
of a communication which suffers from insecurity while X, A



≥ Max I X, A − Max I X, B ,
and B are random states relating to respectively Alice, Bob 𝑓X ( 𝑥) 𝑓X ( 𝑥)

and Eve. while X, A and B are random states relating to respectively

Alice2 , Bob3 and Eve4 , where the maximisation takes place
over the encoding function 𝑓 X (𝑥) − and consequently input
C. Organisation distributions5 − , that is, 𝑓 X (𝑥) − the input distributions −
The rest of the paper is organised as follows. The system should be optimally found by Alice in the sense that the overall
set-up and our main results are given in Sections II and III. performance can be technically guaranteed with regard to the
Subsequently, the evaluation of the framework and conclusions metric of the secrecy capacity6 𝐶𝑠 . Hereinafter, we r-call 𝐶𝑠
are given in Sections IV and V. In Fig. 2, the flow of the main as Λ.
problem and the solution to that is deppicted.
B. Main problem
II. S YSTEM MODEL AND PROBLEM FORMULATION The main problem w.r.t. the secrecy rate Λ is about the max-


imisation problem over the SOP’s
concave version P𝓇 Λ ≥ 𝜆 ,
In this section, we describe the system model, subsequently, that is,  Max P𝓇 Λ ≥ 𝜆 where the parameters are
we formulate the basis of our problem. Δ= 𝜆1 , ··· ,𝜆𝑛
defined in the next section. Furthermore, we discuss how to
A. System description: A traditional Wyner’s wiretap channel reach out the relative eigen-values.
− without loss of generality A SSUMPTION 1. We consider that the dynamical system
A traditional Wyner’s wiretap channel1 based communi- X (𝑡+1) = 𝑎 1 X (𝑡) + 𝓌0(𝑡+1) ,
cation scenario includes a transmitter named Alice and a (2)
A (𝑡) , B (𝑡) = 𝑎 2 , 𝑎 3 , X (𝑡) + 𝓌1(𝑡) , 𝓌2(𝑡) ,




1 Although our novel analysis can be undoubtedly extended to other sce-
is satisfied while (𝑎 1 , 𝑎 2 , 𝑎 3 ) and (𝓌0 , 𝓌1 , 𝓌3 ) are respec-
narios as well. For example, for a reconfigurable intelligent surface based
scheme, the lower-bound of the secret-key-rate as the maximal key bits tively the control parameter tuple and the noise one. is
n
generated from an observation is expressed as [13]: Max I ℎ𝑎𝑏 ; ℎ𝑏𝑎 −

unstable, that is, its spectral radius is



o
I ℎ𝑎𝑏 ; ℎ𝑎𝑒 , ℎ𝑏𝑒 , I ℎ𝑏𝑎 ; ℎ𝑎𝑏 − I ℎ𝑏𝑎 ; ℎ𝑎𝑒 , ℎ𝑏𝑒 while ( ·) 𝑎𝑏 , 𝜑(𝑎 1 ) = Max |𝜁 (𝑎 1 )| > 1, (3)
( ·) 𝑏𝑎 , ( ·) 𝑎𝑒 , ( ·) 𝑏𝑒 respectively declare the links Alice-to-Bob, Bob-to- 𝑖
Alice, Alice-to-Eve and Bob-to-Eve and ℎ stands for the stacked versions 2 Encoded
of measurements at the relative receiver. Nevertheless, the physical logic by Alice.
3 Observed by Bob.
behind the aforementioned rate here is similar to our current consideration and
4 Observed by Eve.
scheme, and we have nothing to to with their detail since we are supposed
5 [14].
to find a relaxation over the outage probability relating to the our security
oriented rates. 6 For the precise definition of rate and capacity, please refer e.g. to [14].
IEEE, VOL. XX, NO. XX, X 2021 3

is satisfied. is controllable, the error covariance matrix at Eve converges

exponentionally to a unique fixed-point [22]. We have conse-
III. M AIN RESULTS quently to consider inverse.
D EFINITION 3 − TANGENT C ONE11 . The term tangent
In this section, our main results are theoretically provided 𝑑𝑒 𝑓


𝑑𝑖𝑠𝑡 𝑥+ℎ 𝜔,M
in details. cone is defined as 𝜔 M (𝑥) = limℎ→0 𝑖𝑛 𝑓 ℎ = 0
D EFINITION 1. Let us theoretically assign a random vari- while 𝑑𝑖𝑠𝑡 (·, ·) stands literally for the information theoretic
𝑑𝑒 𝑓 
able Λ(𝓉) = 𝜆1 , · · · , 𝜆 𝑛 for the secrecy rate according to
distance(s).
which the SOP’s concave version can be defined as P𝓇 Λ ≥ 𝜆 D EFINITION 4 − I NWARD - POINTING12 . If 𝑓 (·) is a
as well. vector-field relating to M and the boundary 𝜕M exists, 𝑓 (·)
is said to be oint  inward to M at a point 𝑥 ∈ M if
𝑑𝑒 𝑓
 Lemma 1: For the random variable Λ(𝓉) =
𝑓 (𝑥) ∈ 𝜔 M (𝑥) − 𝜔 𝜕M (𝑥) holds13 .
𝜆1 (𝓉), · · · , 𝜆 𝑛 (𝓉) over the time horizon − for which
C OROLLARY 1. Any compact and convex set is con-
the term (𝓉) is neglected hereinafter for the ease of notation
tractible, but not vise versa14 . Meamwhile, every contractible
−, the expression
 set, even non-convex, has a concave volume15 . Finally, every
 E 𝑒 𝓉Λ − 1 vector field that is inward-point to M has an equilibrium [27],
E Λ ≤ , (4)
𝓉 [28] as there exists at least one contractible interior within the
is satisfied. boundary 𝜕M, if 𝜕M exists.
P ROOF : The proof is convenient to follow according to the O BJECTIVE . Our aim is to increase the contractibility
Taylor-expansion theorem, however, see Appendix A for more radius16 as much as possible.
detailed justifications. 
In the following, roughly speaking, we consider the secrecy
rate as a Riemannian mani-fold for which a volume is techni- A. How to define a maximisation problem

over the SOP’s
cally defined. concave version as  Max P𝓇 Λ ≥ 𝜆 w.r.t. the secrecy
R EMARK 1: The reason why we consider a vloume can Δ= 𝜆1 , ··· ,𝜆𝑛
be theoretically justified as follows. Unhesitatingly, since we rate Λ
consider the relative parameters as Riemannian mani-folds, Lemma 2: Vitale’s random Brunn-Minkowski inequality17 −
one can justify a volume over them from a generic point of The expression
view. In particular and as obvious e.g. from [15], [16] and n  o n  o
in the context of the Asymptotic Equipartition Property, one Vℴ𝓁 E Λ ≥ E Vℴ𝓁 Λ , (5)
can interpret the volume of a random variable − only for the

holds.
secrecy rate Λ but not the SOP’s concave version P𝓇 Λ ≥ 𝜆
Lemma 3: The expression
− as the exponential of Shannon’s entropy of it7 . Indeed, this
property can be elaborated via the Cramer’s large deviation n E  𝑒 𝓉Λ − 1 o n  o
Vℴ𝓁 ≥ E Vℴ𝓁 Λ , (6)
theorem8 . Meanwhile, in relation to the volume of the random 𝓉
variable, please do not misunderstand it with the standard holds.
deviation as we are generalising it. P ROOF : The proof is easy to follow by an integration
D EFINITION 2. Let us, for more generalisations, consider ofn Lemma
 o1 and Lemma
n  2. This
o is due to the fact that
the secrecy rate  as the Riemannian mani-fold M for which the E Vℴ𝓁 Λ ≤ Vℴ𝓁 E Λ holds. 
volume Vℴ𝓁 Λ is valid. Undoubtedly, M can be contractible |{z}
if its Euler characteristics gets 𝒳(M) = 19 , that is, if M is 
E 𝑒𝓉Λ −1
homotopically10 equivalent to a single-point, that is, when and ≤ 𝓉
if M is continously shrunk and topologically deformed into Lemma 4: The expression
a sigle-point. In other words, a space is contractible iff the 

E 𝑒 𝓉Λ = 𝑒 𝓉𝜆 P𝓇 Λ ≥ 𝜆 , (7)
identity map from it to itself − which is always a homotopy
equivalence − is null-homotopic [17], [18], [19], [20]. strongly holds.
In order to further justify Definition 2, re-call Assumption P ROOF : See Appendix B. 
1. Indeed, when a system is unstable, the more uncertainty Lemma 5: The problem
is amplified originating from the noise. Thus, Eve cannot n o n
o
𝓉𝜆
always predict the system close to an equilibrium [21]. Or  Max Vℴ𝓁 𝑒 Vℴ𝓁 P𝓇 Λ ≥ 𝜆 , (8)
√ Δ= 𝜆1 , ··· ,𝜆𝑛
in other words, if: (i) (𝑎 1 , 𝑎 2 ) is observable; and (ii) (𝑎 1 , 𝑎 2 )
 11 See e.g. [23], [24], [25] to understand what it is.
7
 ℋ Λ 
 That
is, Vℴ𝓁 Λ(𝓉) → 𝑒 holds. In other words, Vℴ𝓁 Λ(𝓉) ≈ 12 See e.g. [17], [18], [19] to understand what it is.
E I Λ . 13 This is more-and-less similar to Pincare-Hopf-Theorem in system theory
8 It states that the probability of a large deviation from mean decays and stabilizations.
14 See [26], Page 4.
exponentially with the number of samples. See the large deviations theory.
9 See e.g. [17], [18], [19] to understand what it is. 15 See [26], Page 4.
10 In topology, two continuous functions from one topological space to 16 See e.g. [17] to understand what it is.

another one − whether isolated or not − are called homotopic. 17 Generalised Brunn-Minkowski inequality [29].
IEEE, VOL. XX, NO. XX, X 2021 4

How to bound :

Lemma 1, Lemma 2, Lemma 3

In addition to Lemma 4

How to find a dual for :

Lemma 5
………………………………………………………………………………………………... ………....
In addition to Lemma 6 & 7
Eigen
Values
Proposition 1: How to reach out the eigen-values
Fig. 2: Flow of problem-and-solution.

can be18 a dual one for the problem P ROOF : The proof is easy to follow with the aid of the
n  o generalised Yong’s ineqality19 which says that
 Max E Vℴ𝓁 Λ . (9)
Δ= 𝜆1 , ··· ,𝜆𝑛 𝑓 0 (𝑥)𝑔 0 (𝑥) ≤ 𝑓 (𝑥) + 𝑔(𝑥), (12)
P ROOF : See Appendix C.  holds for the arbitary functions 𝑓 (·) and 𝑔(·), while (·) 0 stands
R EMARK 2: for the drivative. 
n o n o
• (i) Vℴ𝓁 𝑒 𝓉𝜆 . Whether  Max Vℴ𝓁 𝑒 𝓉𝜆 is of a
Δ= 𝜆1 , ··· ,𝜆𝑛 B. How to reach out the eigen-values
partially useless nature here for our main problem or not, R EMARK 3: Regarding to the fact that mainly most of the
we use it as a trick which is of a purely useful nature in secrecy rate problems can be discussed in the context of semi-
the next lemma.
n
o definite algebra [4]-[12], that is the format B𝑇 AB, we jump
• (ii) Vℴ𝓁 P𝓇 Λ ≥ 𝜆 . First of all, we have nothing to do in terms of the following to the next steps.
with its maximum version, i.e., 1. Additionally, recalling Lemma 7: Finsler’s lemma20 − The problem
Definition
 2 as well as Remark 1 in connection with
Vℴ𝓁 · , we see that the aforementioned value is not ∃X, X𝑇 AX = 𝜉, X𝑇 BX ≤ 𝜉 =⇒ ∃𝑧 : B − 𝑧A < 𝜉, (13)
by-default equated with 1.
• (iii) R ECAPTULATION . So far, we
have indeed recasted holds for the arbitary matrices A and B while 𝜉 and 𝑇
the problem  Max P𝓇 Λ ≥ 𝜆 into two parts, i.e., stand respectively for an arbitary threshold and the transpose
Δ= 𝜆1 , ··· ,𝜆𝑛 operand.

n o n Proposition 1: Let us assume the descriptor system B, A ,
 Max Vℴ𝓁 𝑒 𝓉𝜆 and  Max Vℴ𝓁 P𝓇 Λ ≥ so, the characteristic polynomial is given as
Δ= 𝜆1 , ··· ,𝜆𝑛 Δ= 𝜆1 , ··· ,𝜆𝑛

o

𝜆 which are deterministic ones. 𝒫(𝑧) = 𝑑𝑒𝑡 B − 𝑧A , (14)

Lemma 6: The problem while 𝑑𝑒𝑡 (·) stands for the matrix determinant. The number
∫ n o ∫ n of eigenvalues in the region associated with the polynomial

o n
o
 Max Vℴ𝓁 𝑒 𝓉𝜆 + Vℴ𝓁 P𝓇 Λ ≥ 𝜆 , (10) 𝒫(𝑧) over the Riemannian Vℴ𝓁 P𝓇 Λ ≥ 𝜆 is related to
Δ= 𝜆1 , ··· ,𝜆𝑛
−1
and 𝑑𝑒𝑡 B −𝑧A while (·) −1 stands for the inverse


B −𝑧A
can be a dual for the problem matrix.
n o
𝓉𝜆
n
o P ROOF : See Appendix D. 
 Max Vℴ𝓁 𝑒 Vℴ𝓁 P𝓇 Λ ≥ 𝜆 , (11)
Δ= 𝜆1 , ··· ,𝜆𝑛 C. Futher discussion
as its bound. Proposition 2: Let the random 𝑘−dimensional subspace
P𝓇 Λ ≥ 𝜆 be valid and let 𝑣 𝑖0 be the projection of the point


18 Not definitely, but in terms of one of the highly probably efficient and 19 See e.g. [31] to understand what it is.
acceptable one: See [30] for more details. 20 See e.g. [32] to understand what it is.
IEEE, VOL. XX, NO. XX, X 2021 5

Algorithm 1 A greedy algorithm to Λ.

I NITIALISATION .
while TRUE do ( )
n  o n
o
Find the solution to Δ ∈  Max E Vℴ𝓁 Λ ,  Max Vℴ𝓁 P𝓇 Λ ≥ 𝜆 .
Δ= 𝜆1 , ··· ,𝜆𝑛 Δ= 𝜆1 , ··· ,𝜆𝑛
endwhile
end
O UTPUT: Λ

𝑣 𝑖 ∈ Λ into P𝓇 Λ ≥ 𝜆 . Now, calling 𝐿 = ||𝑣 𝑖0 − 𝑣 0𝑗 || 2 and


𝐷 𝐴𝐵 = √ 100 𝑚 and the distance between Alice and Eve is
𝜏1 = 𝑘 ||𝑣 𝑖 − 𝑣 𝑗 || 2 . The value of P𝓇 𝐿 ≤ (1 − 𝜏2 )𝜏1 is bounded 𝐷 𝐴𝐸 = 502 + 𝑑 2 𝑚. The channel matrices are modeled as


2
𝑘𝜏

from above by ≤ 𝑒𝑥 𝑝 − 4 2 , ∀ 0 < 𝑘 < ∞. 𝐻 = 𝐷 −𝜖 /2 ˆ −𝜖 /2 ˆ
𝐴𝐵 𝐻 and 𝐺 = 𝐷 𝐴𝐸 𝐺 for respectively Alice-Bob link
P ROOF : See Appendix E.  and Alice-Eve one while 𝐻ˆ and 𝐺ˆ are small-scale Rayleigh
Proposition 3: Consider Θ ∈ R𝑟 ×𝑚 and 𝜃 ∈ R𝑟 ×𝑚 . The fading modeled matrices with i.i.d. complex Gaussian entries
1
element-wise projection operator 𝑃𝑟𝑜 𝑗 (·) :∈ R×R → R which with the zero-mean and the variance of 𝑠𝑖𝑧𝑒 where 𝑠𝑖𝑧𝑒 is
is a convex and continuously differentiable function is defined the size of the relative matrices, and 𝜖 stands for the path-
as [41] loss exponent − something that is chosen 3. Consequently,
the signal-to-noise-ratio (SNR) values at Bob and Eve are
𝑃𝑟𝑜 𝑗 (𝜃 𝑖 𝑗 , Θ𝑖 𝑗 ) = respectively deriven as 𝑆𝑁 𝑅 𝐵 = 𝐷 𝜖𝑃𝑎𝜕2 and 𝑆𝑁 𝑅 𝐸 = 𝐷 𝜖𝑃𝑎𝜕2
( 𝑑 𝑓 (𝜃 ) 𝐴𝐵 𝑏 𝐴𝐸 𝑒
Θ𝑖 𝑗 − Θ𝑖 𝑗 𝑓 (𝜃 𝑖 𝑗 ), 𝑖 𝑓 𝑓 (𝜃 𝑖 𝑗 ) > 0 & Θ𝑖 𝑗 ( 𝑑 𝜃𝑖 𝑗𝑖 𝑗 ) > 0, where 𝑃 𝑎 stands literally for the transmition power of Alice,
Θ𝑖 𝑗 , 𝑜.𝑤., while 𝜕𝑏2 and 𝜕𝑒2 are the noise variances at Bob and Eve,
(15) respectively.

Table II shows the SOP’s convex version vs. I X, A while
where the index (·)𝑖 𝑗 refers to the element in the 𝑖 𝑡 ℎ row and changing 𝜌 − something that is perfect for the evaluation here.
−𝜌2
the 𝑗 𝑡 ℎ column and the convex and continuously differentiable We indeed use 𝑒
, something that can be in connection
function 𝑓 (𝜃 𝑖 𝑗 ) is defined as P𝓇 Λ
with another lower-bound
from a traditional point of view,
(𝜃 𝑖 𝑗 − 𝜃 𝑖𝑚𝑖𝑛
𝑗 − 𝜂𝑖 𝑗 ) (𝜃 𝑖 𝑗 − 𝜃 𝑖 𝑗
𝑚𝑎𝑥 + 𝜂 )
𝑖𝑗 that is, P𝓇 𝑆𝑆 𝑁
𝑁 𝑅𝐵
𝑅𝐸 < 2 𝑅𝑠 while 𝑅 is our arbitary secrecy rate
𝑠
𝑓 (𝜃 𝑖 𝑗 ) = , (16) threshold. A comparison is also made with the Algorithm 2
(𝜃 𝑚𝑎𝑥 𝜂𝑖 𝑗 − 𝜃 𝑖𝑚𝑖𝑛
𝑗 − 𝜂𝑖 𝑗 )𝜂𝑖 𝑗
given below.
where 𝜂𝑖 𝑗 ∈ R+ is the projection tolerance of 𝜃 𝑖 𝑗 while
𝜃 𝑚𝑎𝑥 −𝜃 𝑚𝑖𝑛 Algorithm 2 A Projection method based algorithm.
𝜂𝑖 𝑗 < 𝑖 𝑗 2 𝑖 𝑗 and 𝜂𝑖 𝑗 < 𝜃 𝑖𝑚𝑎𝑥 𝑗 & |𝜃 𝑖𝑚𝑖𝑛
𝑗 | hold, and while
𝜃 𝑖 𝑗 > 0 and 𝜃 𝑖 𝑗 < 0 are respectively the upper-bound and
𝑚𝑎𝑥 𝑚𝑖𝑛 I NITIALISATION .
the lower-bound of 𝜃 𝑖 𝑗 . Now 𝑃𝑟𝑜 𝑗 (𝜃 𝑖 𝑗 , Θ𝑖 𝑗 ) is a calculable while TRUE do
function ofour eigen-values discussed in the previous parts, Compute 𝑓 (𝜃 𝑖 𝑗 ),
Compute ∇ 𝑓 (𝜃 𝑖 𝑗 ),

that is, P𝓇 Λ ≥ 𝜆 .
P ROOF : See Appendix F.  if 𝑓 (𝜃 𝑖 𝑗 ) > 0 & Θ𝑖 𝑗 (∇ 𝑓 (𝜃 𝑖 𝑗 )) > 0,
Proposition 4: Even if the secrecy rate Λ is not convex and Update 𝑃𝑟𝑜 𝑗 (𝜃 𝑖 𝑗 , Θ𝑖 𝑗 ) ← Θ𝑖 𝑗 − 𝑓 (𝜃 𝑖 𝑗 )Θ𝑖 𝑗
compact, or for more generality, it is not contractible, that can according to the Eqn. 15.
still be relaxed and an equilibrium can be consequently found elseif
in relation to the eigenvalues discussed above. Update 𝑃𝑟𝑜 𝑗 (𝜃 𝑖 𝑗 , Θ𝑖 𝑗 ) ← Θ𝑖 𝑗 according to the
P ROOF : See Appendix G.  Eqn. 15.
Proposition 5: For the case of periodic attacks, an extension endif
over our Markov decision process and with regard to the endwhile
possibility-theory is satisfied in term of a possibilisitically O UTPUT: 𝑃𝑟𝑜 𝑗 (𝜃 𝑖 𝑗 , Θ𝑖 𝑗 )
semi-Markov decision process. end
P ROOF : See Appendix H. 
Table III and Table IV also show the complexity/accu-
IV. N UMERICAL RESULTS racy comparison
n for the opossible greedy algorithmn 1 to find
Initially opening, we have done our simulations w.r.t. the  Max E Vℴ𝓁 Λ and  Max Vℴ𝓁 P𝓇 Λ ≥
Δ= 𝜆1 , ··· ,𝜆𝑛 Δ= 𝜆1 , ··· ,𝜆𝑛
Bernoulli-distributed data-sets using GNU Octave of version
o
4.2.2 on Ubuntu 16.04. 𝜆 . As obvious, it is proven that our derived bound is more
Consider a three-dimensional coordinate network setup acceptable.
consisting
√ of Alice located at (−50, 0, 0), Bob located at Finally speaking, in Fig. 3, and
the subfigures a, b and
(0, 50 3, 0) and Eve located at (0, −𝑑, 0) while 𝑑 is given c the average reward R S𝑡 , A𝑡 , the error in relation to
the Q-function Q𝑡 (𝑠, 𝑎), i.e., Q𝑡Q−Q
★
unkown. Indeed, the distance between Alice and Bob is ★ and the average optimal
IEEE, VOL. XX, NO. XX, X 2021 6

TABLE II: Simulations: SOP’s convex version vs. I X, A while changing 𝜌.




I X, A SOP’s convex version I X, A SOP’s convex version I X, A SOP’s convex version
0 𝑂𝑢𝑟 = 0.005, 𝜌 = 0.1 0.5 𝑂𝑢𝑟 = 0.0026, 𝜌 = 0.1 1 𝑂𝑢𝑟 = 0.0001, 𝜌 = 0.1
0 𝑂𝑢𝑟 = 0.005, 𝜌 = 0.2 0.5 𝑂𝑢𝑟 = 0.0027, 𝜌 = 0.2 1 𝑂𝑢𝑟 = 0.0002, 𝜌 = 0.2
0 𝑃𝑟𝑜 𝑗 = 0.005, 𝜌 = 0.1 0.5 𝑃𝑟𝑜 𝑗 = 0.0027, 𝜌 = 0.1 1 𝑃𝑟𝑜 𝑗 = 0.00011, 𝜌 = 0.1
0 𝑃𝑟𝑜 𝑗 = 0.005, 𝜌 = 0.2 0.5 𝑃𝑟𝑜 𝑗 = 0.00274, 𝜌 = 0.2 1 𝑃𝑟𝑜 𝑗 = 0.0002, 𝜌 = 0.2

n
o
TABLE III: Simulations: Complexity vs. Iterations derived from Δ =  Max Vℴ𝓁 P𝓇 Λ ≥ 𝜆 divided by the derived
Δ= 𝜆1 , ··· ,𝜆𝑛
n  o
one from Δ =  Max E Vℴ𝓁 Λ .
Δ= 𝜆1 , ··· ,𝜆𝑛

Iterations Complexity Iterations Complexity Iterations Complexity

0 0.99 50 0.94 100 0.93

n
o
TABLE IV: Simulations: Accuracy vs. Iterations derived from Δ =  Max Vℴ𝓁 P𝓇 Λ ≥ 𝜆 divided by the derived one
Δ= 𝜆1 , ··· ,𝜆𝑛
n  o
from Δ =  Max E Vℴ𝓁 Λ .
Δ= 𝜆1 , ··· ,𝜆𝑛

Iterations Accuracy Iterations Accuracy Iterations Accuracy

0 1.01 50 1.1 100 1.12

policy 𝜋𝑡 (𝑎|𝑠) − while (·)★ stands for the optimum-value − The proof is now completed. 
are respectively depicted versus the iteration regime while
|S| = 𝒰0 + 𝒱0 = 5 + 7 = 12, |A| = 𝑚𝑎𝑥{𝒰0 , 𝒱0 } = 7, A PPENDIX B
V𝑒𝑛𝑠𝑒𝑚𝑏𝑒𝑙 (𝜁) = 𝜁 2 are basically selected. P ROOF OF L EMMA 4
The cumulative distribution function (CDF)
V. CONCLUSION

A new bound and the relating interpretations over the ℱΛ (𝜆) = P𝓇 Λ ≤ 𝜆


concave version of the SOP maximisation problem were fun- = 1 − P𝓇 Λ ≥ 𝜆 (19)
damentally explored in this paper. We technically considered =1−𝑒 −𝓉Λ
𝜇Λ (𝓉) ,
a Riemannian mani-fold for the SOP’s concave version and
a volume for it. Towards such end, some highly professional holds while 𝜇Λ (𝓉) is the moment-generating function (MGF),
and insightful principles such as Keyhole contour, Finsler’s so, we have
lemma, the generalised Brunn-Minkowski inequality etc were 

E 𝑒 𝓉Λ = 𝑒 𝓉𝜆 P𝓇 Λ ≥ 𝜆 , (20)
used. In order to find the optimal policy in relation to the
eigenvalue distributions, a novel Markov decision process holds.
based reinforcement learning algorithm was also essentially The proof is now completed. 
proposed − something that was subsequently extended to a
possibilisitically semi-Markov decision process for the case A PPENDIX C
of periodic attacks and with regard to the possibility-theory. P ROOF OF L EMMA 5

A PPENDIX A Let us assume

n that we have the optimisation problem of
 o
P ROOF OF L EMMA 1  Max E Vℴ𝓁 Λ , something that is equivalent to the
Δ= 𝜆1 , ··· ,𝜆𝑛
The proof is performed according to the Taylor expansion maximisation over its supremum as in
of 
n E 𝑒 𝓉Λ − 1 o
𝑒 𝓉Λ = 1 + 𝓉Λ + 𝒪{·}, (17)  Max Vℴ𝓁 , (21)
Δ= 𝜆1 , ··· ,𝜆𝑛𝓉
while 𝒪{·} is the Big-O notation. Now, by applying an or with the aid of Lemma 4 − the Eqn. 7 − , as in
expected-value operand, we consequently reach out
 n 𝑒 𝓉𝜆 P𝓇 Λ ≥ 𝜆
− 1 o
 E 𝑒 𝓉Λ − 1  Max Vℴ𝓁 , (22)
E Λ ≤ . (18) Δ= 𝜆1 , ··· ,𝜆𝑛 𝓉
𝓉
IEEE, VOL. XX, NO. XX, X 2021 7

or according to which one can say

n
o ∮
𝓉𝜆 1 h i
 Max Vℴ𝓁 𝑒 P𝓇 Λ ≥ 𝜆 , (23) 𝑁= 𝑇𝑟 (B − 𝑧A) −1 (−A) 𝑑𝑧. (31)
Δ= 𝜆1 , ··· ,𝜆𝑛 2𝜋 𝑗 ℒ
or finally The last integral, i.e., the equation appeared above can
n o n
o be efficiently solved by some digitised methods such as the
𝓉𝜆
 Max Vℴ𝓁 𝑒 Vℴ𝓁 P𝓇 Λ ≥ 𝜆 . (24) Rayleigh-Ritz method [36], [37], [38].
Δ= 𝜆1 , ··· ,𝜆𝑛
In order to conclude the proof, let us ultimately go over the
The proof is now completed.  essential
n relevance between the number of eigen-values and

o
Vℴ𝓁 P𝓇 Λ ≥ 𝜆 in the context of the following lemma.
A PPENDIX D Lemma 8: The number of eigen-values

discussed above
P ROOF OF P ROPOSITION 1 relies fundamentally upon P𝓇 Λ . n

o
The proof is provided here in terms of the following Proof. In relation to the term Vℴ𝓁 P𝓇 Λ ≥ 𝜆 , we get in
solution. hands
Where 𝐾 is a constant scaling factor, one can re-write the 
 𝑒 −𝜌
2

polynomial as21 P𝓇 Λ ≥ 𝜆 ≥ 𝜌 ≤ 1 −
(32)

P𝓇 Λ
𝑛
Ö
24 , while
𝒫(𝑧) = 𝐾 (𝑧 − 𝜁𝑖 ), (25) according to the Talagrand’s Concentration inequality
n
o
𝑖=1 𝜌 is an arbitary threshold. This means that Vℴ𝓁 P𝓇 Λ ≥ 𝜆 ,
while 𝜁𝑖 , 𝑖 ∈ {1, · · · , 𝑛} stands literally for the 𝑖−th eigen-


𝒞 and ℒ are functions of 𝜌; P𝓇 Λ − something that proves
value.
Lemma 8.
n
o
Now, recall the term Vℴ𝓁 P𝓇 Λ ≥ 𝜆 versus
∫ n
o R EMARK 4. The accuracy of evaluating the eigen-values
Vℴ𝓁 P𝓇 Λ ≥ 𝜆 . By differentiating 𝒫(𝑧) with respect expressed here can be fully able to be controlled by 𝜌.
𝒫0 (𝑧) The proof is now completed. 
to 𝑧 as 𝒫 0 (𝑧), 𝒫 (𝑧) is obtained as

𝒫 0 (𝑧) ∑︁ 1
𝑛
= . (26) A PPENDIX E
𝒫(𝑧) 𝑖=1
𝑧 − 𝜁𝑖 P ROOF OF P ROPOSITION 2
√
For the aboven equation, where 𝑗 = −1 is the imaginary The sketch of the proof is given here which is similar to
[40].

o
unit, ℒ ⊇ Vℴ𝓁 P𝓇 Λ ≥ 𝜆 is a closed anti-clockwise curve
∫ n
o We know [40]
on the complex plane, and 𝒞 ⊇ Vℴ𝓁 P𝓇 Λ ≥ 𝜆 is the

P𝓇 𝐿 ≤ (1 − 𝜏2 )𝜏1
region enclosed by ℒ, it is achieved as22 𝑘
(33)
𝑛
( ≤ 𝑒𝑥 𝑝 (1 − (1 − 𝜏2 ) + 𝑙𝑛(1 − 𝜏2 ) ,
2
∮ ∑︁
1 2𝜋 𝑗, if 𝜁𝑖 ∈ 𝒞,
𝑑𝑧 = (27) which can be re-casted to
ℒ 𝑖=1 𝑧 − 𝜁𝑖 0, if 𝜁𝑖 ∉ 𝒞,


P𝓇 𝐿 ≤ (1 − 𝜏2 )𝜏1
accoring to which one can say that the number of the eigen-
𝜏22
 (34)
  
values in the region 𝒞 is 𝑘
≤ 𝑒𝑥 𝑝 𝜏2 − 𝜏2 + ,
1
∮
𝒫 0 (𝑧) 2 2
𝑁= 𝑑𝑧 2
2𝜋 𝑗 ℒ 𝒫(𝑧) since 𝑙𝑛(1 − 𝑥) ≤ −𝑥 − 𝑥2 , ∀ 0 ≤ 𝑥 < 1 holds − something that
𝑛 ∮ (28) can conclude the proof. 
1 ∑︁ 1
= 𝑑𝑧.
2𝜋 𝑗 𝑖=1 ℒ 𝑧 − 𝜁𝑖
A PPENDIX F
On the other hand, 𝒫 0 (𝑧) is obtained as [35]23 P ROOF OF P ROPOSITION 3
h 𝜕 (B − 𝑧A) i
𝒫 0 (𝑧) = 𝑑𝑒𝑡 (B − 𝑧A)𝑇𝑟 (B − 𝑧A) −1 , The proof is given as the following.
𝜕𝑧
| {z } (29) We know [41]
−A


𝑡𝑟 (𝜃𝑇 − 𝜃 ∗𝑇 ) − Θ + 𝑃𝑟𝑜 𝑗 (𝜃, Θ) < 0, (35)
while 𝑇𝑟 [·] stands for the trace of the matrix, something that
is equivalent to holds for 𝜃 ∗ ∈ [𝜃 𝑖𝑚𝑖𝑛
𝑗 + 𝜂𝑖 𝑗 , 𝜃 𝑖 𝑗
𝑚𝑎𝑥 − 𝜂 ] while the trace operator
𝑖𝑗
h i 𝑡𝑟 (·) is a function of − sum of − the eigen-values related to
𝒫 0 (𝑧) = 𝒫(𝑧)𝑇𝑟 (B − 𝑧A) −1 (−A) , (30) 𝜃, that is, Λ in our scheme and analysis.
21 See e.g. [33]. 24 See e.g. [39] to understand what it is: It says that the complement of
22 See e.g. [34]. the given random variable in a bounded probability closure is emphatically
23 Page 8, eqn. 46. upperbounded.
IEEE, VOL. XX, NO. XX, X 2021 8

Algorithm 3 A two-time-scale natural actor-critic 𝜖−greedy algorithm.

I NPUT.
• S0 ∼ arbitary; // Initial state
• A 0 ∼ 𝜋 𝑜 (·|𝑠0 ); // Initial action
1
• 𝜋 𝑜 (𝑎|𝑠) = 𝜋ˆ 𝑜 (𝑎|𝑠) = | A | ; // Initial ploicy to be learnt, as the probability distribution of taking the relative action at
the given state
• 𝑇 > 0; // Iteration number
• 𝛼𝑡 > 0; // Step size
• 𝛽𝑡 > 0; // Step size
• 𝜖 𝑡 > 0; // Greedy factor
• 𝜓 > 0; // Discount factor
• Q0 (𝑠, 𝑎) ∈ R | S | | A | . // Initial Q-function corresponding to the initial policy 𝜋𝑜 (𝑎|𝑠)
for 𝑡 = 0, 1, · · · , 𝑇 do

Sample S𝑡+1 ∼ P · |S𝑡 , A
𝑡 ; // According to the set of the transition probability matrices P
Sample A𝑡+1 ∼ 𝜋ˆ 𝑡 · |S𝑡+1 ;
𝛼𝑡 (𝑠, 𝑎) ← 𝛼𝑡 1 S𝑡 = 𝑠, A𝑡 = 𝑎, ∀𝑠, 𝑎h ;




i
Q𝑡+1 (𝑠, 𝑎) ← Q𝑡 (𝑠, 𝑎) + 𝛼𝑡 (𝑠, 𝑎) R S𝑡 , A𝑡 + 𝜓Q𝑡 S𝑡+1 , A𝑡+1 − Q𝑡 S𝑡 , A𝑡 , ∀𝑠, 𝑎; // With regard to the
 
1 ( ·) ( ·)

Reward-function R (·, ·) ∝ 𝑓 (·) (·)

while P 𝜁 𝑚𝑎𝑥 , 𝜁 𝑚𝑖𝑛 to be chosen
P 𝜁𝑚𝑎𝑥 ,𝜁𝑚𝑖𝑛

𝑒𝑥 𝑝 𝛽𝑡 Q𝑡+1 (𝑠,𝑎)
𝜋𝑡+1 (𝑎|𝑠) ← 𝜋𝑡 (𝑎|𝑠) Í 0 0

, ∀𝑠, 𝑎;
𝑎0 𝜋𝑡 (𝑎 |𝑠)𝑒𝑥 𝑝 𝛽𝑡 Q𝑡+1 (𝑠,𝑎 )
𝜖𝑡
𝜋ˆ 𝑡+1 ← | A | + (1 − 𝜖 𝑡 )𝜋 𝑡+1 .
endfor
Sample 𝑇ˆ from {0, 1, · · · , 𝑇 } by distribution P 𝑇ˆ = 𝑖 =

𝛽
Í𝑇 𝑖
𝛽𝑗
.
𝑗=0
O UTPUT: 𝜋ˆ𝑇ˆ
end

We furthermore know that Sizes of random projections of
sets, i.e., Thereom 7.7.1 in [42] may help us to prove that if
we have a bounded set 𝜃 ∈ R𝑟1 , 𝑟 1 < 𝑟 while 𝑟 was defined
in Proposition 2, with a projected set Θ ∈ R𝑟2 , 𝑟 2 < 𝑟, with a
probability of at least 1 − 2𝑒 −𝑟2 we have
Now, we initially see that the contractibility radius as well
as the equilibrium we are supposed to go over rely deeply
upon the principal curvatures, i.e., the eigen-vectors.
28 says that the Systol of
Additionally, Pu-1952 inquality
a manifold M as SYS M as the least lenght29 of a non-

 √︂

 contractible loop of the homeomorphic manifold M − to the


𝑟2
𝑑𝑖𝑎𝑚 Θ𝜃 ≤ 𝐶0 𝑤 𝑠 𝜃 + 𝑑𝑖𝑎𝑚 𝜃 , (36) real projective plan − , i.e., the lowerbound of the lenghts of
𝑟1 non-contractible closed curves over M 30 satisfies
while 𝐶0 is a constant, 𝑑𝑖𝑎𝑚(·) stands for the diameter, 2
S𝑎𝑟 𝑒𝑎 ≥ SYS 2 M ,


and 𝑤 𝑠 (𝜃) denotes the Gaussian width as E Sup h𝑥, ℊi, ℊ ∼ (38)
𝑥∈𝜃
𝜋
𝒩(0, 𝐼𝑟1 ). while the equlity holds31 for the constant Gaussian curvatures,
The proof is now completed.  i.e., when M is locally isometric. Or correspondingly32 ,

 𝑛1
(39)


A PPENDIX G C𝑛 Vℴ𝓁 M ≥ SYS M , ∃C𝑛 ∈ R𝑛 .
P ROOF OF P ROPOSITION 4
Thus, it has so far been proven that, in order to work on
Let us start the proof with the Gauss-Bonnet-Theorem25 . the contractibility radius as well as the equilibrium discussed
It says that for a manifold M
with the boundary 𝜕M, with above, it is necessary and sufficient for us to only focus on
the Euler characterisitcs X M and the Gaussian Curvature26 the eigen-vectors.
K and the Geodesic Curvature27 K𝑔 relating to 𝜕M, the Now, in every kind of manifold and space, there may exist
following is satidfied multiple maximum-eigenvalues or/and minimum-eigenvalues,
for example, a hemisphere has 3 maximum-eigenvalues and
∫ ∫


K𝑑S𝑎𝑟 𝑒𝑎 + K𝑔 𝑑𝑠 = 2𝜋X M , (37) only 1 minimum-eigenvalue. However, the distribution of
M 𝜕M
while S𝑎𝑟 𝑒𝑎 stands theoretically for the area of M and 𝑠 ⊂ 28 [45].
29 [46].
S𝑎𝑟 𝑒𝑎 .
30 That 𝑑𝑒 𝑓 
is, = 𝑖𝑛 𝑓 𝓁 (𝑐) | 𝑐 : non-contractible closed curves from a
25 See e.g. [43], [44] to understand what it is. mathematical point of view, while 𝓁 (𝑐) denotes the lenght of 𝑐.
26 See e.g. [17], [18], [19], [20] to understand what it is. 31 Minding’s theorem.
27 See e.g. [17], [18], [19], [20] to understand what it is. 32 See e.g. [47], [48].
IEEE, VOL. XX, NO. XX, X 2021 9

0.8
Averaage Optimal Policy

0.6

0.4

0.2

Fig.
 (1) 4: The (𝓋) Markov model in relation
 (1) to our scheme and
0
(𝓊)
0 10 20 30 40 50 𝜁 𝑚𝑎𝑥 , · · · , 𝜁 𝑚𝑎𝑥 as well as the 𝜁 𝑚𝑖𝑛 , · · · , 𝜁 𝑚𝑖𝑛 .
Iterations

(a) Average optimal policy 𝜋𝑡 (𝑎 |𝑠)

1
1
In relation to Algorithm 334 , the reward function R (·, ·)
plays a vital role. Regarding the fact that we aim at
finding an equilibrium as discussed above, and due to
0.8
the fact that in equilibria, the maximum-eigenvalues and
the minimum ones tend to get closer to each other35
Averaage Reward

0.6 as much as possible,

 one may  select the reward func-
1 ( ·) ( ·)

tion R (·, ·) as 𝑓 (·) (·)

while P 𝜁 𝑚𝑎𝑥 , 𝜁 𝑚𝑖𝑛 stands
P 𝜁𝑚𝑎𝑥 ,𝜁𝑚𝑖𝑛
0.4 fundamentally for the joint probability distribution for the
maximum-eigenvalues and the minimum ones − something
( ·) ( ·)
𝑑𝑒 𝑓 (𝓋)
that is36 writen as P 𝜁 𝑚𝑎𝑥
Î Î
0.2 , 𝜁 𝑚𝑖𝑛 = 𝓊 𝓋 |𝜁 𝑚𝑎𝑥 −
𝒰0 +𝒱0
(𝓊) 𝛽 Î −𝛽 2 V
𝜁 𝑚𝑖𝑛 | 𝓊+𝓋 𝑒 𝑒𝑛𝑠𝑒𝑚𝑏𝑒𝑙 (𝜁), 𝓊 ∈ {1, · · · , 𝒰0 }, 𝓋 ∈
0
{1, · · · , 𝒱0 } while 𝛽 is the Dyson index and V𝑒𝑛𝑠𝑒𝑚𝑏𝑒𝑙 (𝜁) is
0 10 20 30 40 50
Iterations theoretically the variant ensemble which is e.g. for Gaussian

case 𝜁 2 .
(b) Average reward R S𝑡 , A𝑡

C OROLLARY 2 − E XAMPLE 1. In ∫ case of X M∫ = 1, the
1
principal eigenvalues and the term M K𝑑S ∮ 𝑎𝑟 𝑒𝑎 + 𝜕M K𝑔 𝑑𝑠
2
expressed before have a structure such as 1+𝑥 2 𝑑𝑥 − some-
0.8 thing that is well-routine for information-theoretic schemes
such as the dirty-paper-coding-principle. According to what
Q-Function Error

0.6
we have gone over in Definition 2, this case guarantees the
convexity over the secrecy rate.

C OROLLARY 3 − E XAMPLE 2. In case of X M =
0.4
0 e.g. for Torus or Kelin-Bottle, regarding the inequality
S𝑎𝑟 𝑒𝑎 ≥ 𝜋2 SYS 2 M , it is proven that one should send

0.2 little amount of information in the sense that a less amount

0
0 10
(c) Error for Q-function Q𝑡 (𝑠, 𝑎)
Fig. 3: Reward, policy and Q-function error versus the it-
eration regime, while |S| = 𝒰0 + 𝒱0 = 5 + 7 = 12,
|A| = 𝑚𝑎𝑥{𝒰0 , 𝒱0 } = 7, V𝑒𝑛𝑠𝑒𝑚𝑏𝑒𝑙 (𝜁) = 𝜁 2 .
the eigenvlaues may be totally different in every case33 .
Thus, there may exist a Markov-Decision-Process − some-
thing that technically enforces us to propose the following
Reinforcement-learning based algorithm to find the perfect
policy according to the eigenvalues’ distributions.
33 See e.g. [49], [50], [51], [52].
of information leaked by Eve can be guaranteed, aimed at
reducing the amount of non-contractibility and tits relative
20
Iterations
30 40 50
radius. This case and interpretation can be proven as follows
according to the Excision-Theorem37 . This theorem says that
if M0 ⊂ M1 ⊂ M, we say
M0 can be excised if the inclusion
map M \ M
0 , M1 \ M0 has an isomorphism relationship38
with M, M1 . This kind of interpretation can also be proven
by the concept of symplectic capacity as described in the
following remark.
R EMARK 5 − S YMPLECTIC CAPACITY39 . The principle
of Symplectic capacity falls in finding a contractible periodic
34 In order to understand a Markov decision process reinforcement learning
based algorithm, see e.g. [53].
35 See e.g. [49], [50], [51], [52].
36 See e.g. [51], [54], [55].
37 See e.g. [56] to understand what it says.
38 Duality.
39 See e.g. [47], [48].
IEEE, VOL. XX, NO. XX, X 2021 10

( ·) ( ·)

orbit whose period bounds the Hofer-Zehnder capacity on Provisionally speaking, re-call P 𝜁 𝑚𝑎𝑥 , 𝜁 𝑚𝑖𝑛 from Ap-
the energy level which is related to the cylindrical capacity pendix G. Now,
if we are supposed to extend the Markov
as follows. It says for M0 ⊂ M1 ⊂ M, the following process F 𝜁 𝑠𝑒𝑡 , (·) 𝑠𝑒𝑡 ∈ {𝑚𝑖𝑛, 𝑚𝑎𝑥} to a semi-Markov one,
capacity inequality holds while C𝑎 𝑝𝑎𝑐𝑖𝑡 𝑦 denotes the capacity: while the
timing jumps are randomly distributed as well, that
C𝑎 𝑝𝑎𝑐𝑖𝑡 𝑦 M1

 Vℴ𝓁 M
 𝑛1 is, F 𝜁 𝑠𝑒𝑡 is only valid as follows

≤
.
C𝑎 𝑝𝑎𝑐𝑖𝑡 𝑦 M0 Vℴ𝓁 M0

C OROLLARY 4 − E XAMPLE 3. F 𝜁 𝑠𝑒𝑡 (𝑡) , ∀𝑡 𝑠 ≤ 𝑡 < ∀𝑡 𝑠+1 ,

In case of complex or/and





(𝑠+1) (𝑠)
imaginary values40∫ such as X M ∫ = 𝑖, the principal eigenval-
𝑑𝑒 𝑓
(41)


𝓅𝑖 𝑗 = P𝓇 𝜁 𝑠𝑒𝑡 = 𝑗, 𝑡 𝑠+1 − 𝑡 𝑠 ≤ 𝑡|𝜁 𝑠𝑒𝑡 =𝑖 =
ues and the term M K𝑑S ∮𝑎𝑟1𝑒𝑎 + 𝜕M K𝑔 𝑑𝑠 expressed before

 (𝑠+1) (𝑠)


 P𝓇 𝜁 𝑠𝑒𝑡 = 𝑗 |𝜁 𝑠𝑒𝑡 = 𝑖 ,
have a structure such as 𝑥 𝑑𝑥 − something that may result
in creation of bifurcations in eigenvalues. while (·) 𝑠 stands for the 𝑠−th state.
R EMARK 6 − C ONFORMAL EQUIVALENCE FOR CURVA -
TURES 41 . Two metrics ℊ0 and ℊ𝜙 are conformally equivalent R EMARK 8 − A N OVERVIEW OVER THE POSSIBIL -
if ℊ𝜙 = 𝑒 2𝜙 ℊ0 holds while 𝑒 2𝜙 is called the conformal factor. ITY THEORY [63]-[67]. The following main rules hold in
Now, the following
is satisfied for the relative curvatures [60]: the possibility-theory: (i) the normality axiom indicates that
K 𝜙 = 𝑒 2𝜙 K0 − Δ𝜙 while Δ𝜙 is the Laplacian on the relative 𝜐(𝑠) = 1, ∀𝑠 ∈ S holds; (ii) the non-negativity axiom indicates
surface. that 𝜐(∅) = 0; (iii) degree of possibility is derived by Υ(S) =
R EMARK 7 − DAVIS -K AHAN -T HEOREM42 . Assume Sup 𝜐(𝑠); (iv) degree of possibility is derived by Υ (𝑛𝑒𝑐) (S) =
𝑠 ∈S
M0 ⊂ M and M1 ⊂ M while M0 and M1 are not necessarily 1 − Inf 𝜐(𝑠); (v) the maxitivity axiom45 says that Υ(S1 ∪ S2 ) =
equal nor subsets of each other. There exists the following in 𝑠∉S

relation to the eigen-vectors 𝓋 of M0 and M1 𝑚𝑎𝑥 Υ(S1 ), Υ(S2 ) , ∀S1 , S2 ⊆ S; (vi) the
minitivity axiom
says that Υ(S1 ∩S2 ) = 𝑚𝑖𝑛 Υ(S1 ), Υ(S2 ) , ∀S1 , S2 ⊆ S. Fur-

2
𝑠𝑖𝑛] 𝓋𝑖 (M0 ), 𝓋𝑖 (M1 ) ≤ ||M0 − M1 ||, thermore, the following conditional properties are information-
𝛾𝑥 theoretically satisfied [67]
(40)
𝑑𝑒 𝑓
𝛾 𝑥 > 0 = Min |𝜁𝑖 (M1 ) − 𝜁 𝑗 (M0 )|, ( 𝑗) ( 𝑗)
𝑗≠𝑖 𝜐(𝑠1(𝑖) , 𝑠2 ) = 𝜐(𝑠2 |𝑠1(𝑖) )𝜐(𝑠1(𝑖) ),
while 𝛾 𝑥 > 0 is defined as the least separation distance of the 1 ( 𝑗)
𝜐(𝑠1(𝑖) ) = Max 𝜐(𝑠1(𝑖) , 𝑠2 ),
largest eigen-value(s) from the rest of the spectrum. 𝜛1 (𝑠1(𝑖) ) 𝑠2( 𝑗)
The proof is now completed.  ( 𝑗) 1 ( 𝑗)
𝜐(𝑠2 ) = ( 𝑗)
Max 𝜐(𝑠1(𝑖) , 𝑠2 ),
(𝑖)
𝜛2 (𝑠2 ) 𝑠1
A PPENDIX H ( 𝑗)
𝜛1 (𝑠1(𝑖) ) = Max 𝜐(𝑠2 |𝑠1(𝑖) ) ≤ 1,
P ROOF OF P ROPOSITION 5 ( 𝑗)
𝑠2
( 𝑗) ( 𝑗) (42)
If the attack is a denial-of-service one and if it is perodic, as 𝜛2 (𝑠2 ) = Max 𝜐(𝑠1(𝑖) |𝑠2 ) ≤ 1,
(𝑖)
fully discussed e.g. in [61], there consequently exist two totally 𝑠1
i.i.d and separate scenraios in termf of two separately unstable 𝜐(𝑠1(𝑖) )𝜐(𝑠2 |𝑠1(𝑖) )
( 𝑗)
( 𝑗)
and stable sub-systems between which there is a switching 𝜐(𝑠1(𝑖) |𝑠2 ) = ( 𝑗)
case. Now regarding the facts that: 𝜐(𝑠2 )
( 𝑗) ( 𝑗)
• (i) the switching case theoretically entails a semi-Markov 𝜛2 (𝑠2 )𝜐(𝑠1(𝑖) )𝜐(𝑠2 |𝑠1(𝑖) )
model43 ; and = ( 𝑗) ,
Max 𝜐(𝑠1(𝑘) )𝜐(𝑠2 |𝑠1(𝑘) )

• (ii) the frequency of the occurrence in relation to our 𝑠1
(𝑘)

Markov model may be unavailable, that is, our knowl-

edge of the information is somehow incomplete, the as well as [67]
probability-theory is totally inappropriate44 here, we con- 1 ( 𝑗) ( 𝑗)
𝜐(𝑠3(𝑘) |𝑠1(𝑖) ) = Max 𝜐(𝑠3(𝑘) |𝑠2 )𝜐(𝑠2 |𝑠1(𝑖) ),
sequently need to extend the probability transitions to the 𝜛3 (𝑠1(𝑖) ) 𝑠2( 𝑗)
possibilities ones − where the transitions stand for the (43)
( 𝑗)
strength and casuality − and from a possibility-theoretic 𝜛3 (𝑠1(𝑖) ) = Max 𝜐(𝑠2 |𝑠1(𝑖) ) ≤ 1,
( 𝑗)
𝑠2
point of view;
one can extend the Markov model discussed in the previous
part as below.

40 See e.g. [57], [58], [59] to understand what they technically are. See also
Caldero-Chapoton function.
41 See e.g. [60].
42 See e.g. Theorem 4.5.5. in [42].
43 See e.g. [62] to understand the randomness of the time transitions and 45 For example, if a person is a 50-year-old one, if with the confidence of
the necessity of semi-Markov modeling. 1 we say he/she is an ”aged” person, 0.5 a ”middle-aged” person, and 0 a
44 See e.g. [63], [64], [65], [66] to understand the differences between ”young” one, with the confidence of 1 we can undoubtedly declare that he/she
possibility-theory and probability-theory. is ”adult”.
IEEE, VOL. XX, NO. XX, X 2021 11

and [67] [9] K. Guo, K. An, F. Zhou, T. A. Tsiftsis, G. Zheng, ”On the Secrecy Per-
formance of NOMA-Based Integrated Satellite Multiple-Terrestrial Relay
( 𝑗)
Max 𝜐(𝑠3(𝑘) , 𝑠2 |𝑠1(𝑖) ) = Networks With Hardware Impairments,” IEEE Trans. Vehicular Technol.,
( 𝑗) Vol. 70, no. 4, pp. 3661-3676, 2021.
𝑠2
( 𝑗) ( 𝑗)
[10] X. Lai, L. Fan, X. Lei, Y. Deng, G. K. Karagiannidis, ”Secure Mobile
Max 𝜐(𝑠3(𝑘) |𝑠2 , 𝑠1(𝑖) )𝜐(𝑠2 |𝑠1(𝑖) ) = (44) Edge Computing Networks in the Presence of Multiple Eavesdroppers,”
( 𝑗)
𝑠2 IEEE Trans. Commun., Vol. pp, no. 99, pp. 1-1, 2022.
( 𝑗) ( 𝑗) [11] G. Sharma, N. Pandey, A. Singh, R. K. Mallik, ”Secrecy Optimization
Max 𝜐(𝑠3(𝑘) |𝑠2 )𝜐(𝑠2 |𝑠1(𝑖) ), for Diffusion-Based Molecular Timing Channels, ” IEEE Trans. Molecular,
( 𝑗)
𝑠2 Bio. Multi. Commun., Vol. 7, no. 4, pp. 253-261, 2021.
[12] Y. Lou, R. Sun, J. Cheng, D. Nie, G. Qiao, ”Secrecy Outage Analysis
which is also equal to [67] of Two-Hop Decode-and-Forward Mixed RF/UWOC Systems,” IEEE
Commun. Letters, Vol. pp, no. 9, pp. 1-1, 2022.
( 𝑗)
Max 𝜐(𝑠2 |𝑠3(𝑘) , 𝑠1(𝑖) )𝜐(𝑠3(𝑘) |𝑠1(𝑖) ) = [13] T. Lu, L. Chen, J. Zhang, K. Cao, ”Reconfigurable Intelligent Surface
( 𝑗)
𝑠2 Assisted Secret Key Generation in Quasi-Static Environments,” IEEE
( 𝑗) Commun. Letters, Vol. 26, no. 2, pp. 244-248, 2022.
Max 𝜐(𝑠2 |𝑠1(𝑖) )𝜐(𝑠3(𝑘) |𝑠1(𝑖) ) = (45) [14] M. Zamanipour, "A Novelty in Blahut-Arimoto Type Algorithms: Opti-
( 𝑗)
𝑠2 mal Control over Noisy Communication Channels," IEEE Trans. Vehicular
Technol. Vol. 69, no. 6, pp. 6348-6358, 2020.
𝜛3 (𝑠1(𝑖) )𝜐(𝑠3(𝑘) |𝑠1(𝑖) ). [15] P. Algoet, T. Cover, ”A Sandwich Proof of the Shannon-McMillan-
Breiman Theorem,” The Annals of Probability. Vol. 16, no. 2, pp. 899-909,
In addition, the following equations are also added [68], 1988.
[69] [16] S. Verdu, T. Han, ”The Role of the Asymptotic Equipartition Property
( in Noiseless Source Coding,” IEEE Trans. Info. Theory, Vol. 43, no. 3, pp.
𝜐 (𝑠1 , ··· ,𝑠 𝑁 ) 847-857, 1997.
, 𝜐(𝑠2 , · · · , 𝑠 𝑁 ) ≠ 0,
𝜐(𝑠1 |𝑠2 · · · , 𝑠 𝑁 ) = 𝜐 (𝑠2 , ··· ,𝑠𝑁 ) (46) [17] J. Wang, Z. Xie, G. Yu, ”Decay of scalar curvature on uni-
1, 𝜐(𝑠2 , · · · , 𝑠 𝑁 ) = 0, formly contractible manifolds with finite asymptotic dimension,”
https://arxiv.org/abs/2101.11584, 2021.
as well as [18] B. Tosun, ”Stein domains in C2 with prescribed boundary,” Adv. Geom.,
Ö
Vol. 22, no. 1, pp. 9-22, 2022.
𝜐(𝑠1 , · · · , 𝑠 𝑁 ) = 𝜐 𝑠𝑖 |𝒫𝒶𝓇(𝑠𝑖 ) , (47) [19] J. M. Lee, ”Introduction to smooth manifolds,” G. Texts. Math., 2012.
𝑖 [20] G. Naber, ”Topological methods in Euclidean spaces,” Dover, 2000.
[21] A. Tsiamis, K. Gatsis, G. J. Pappas, ”State-Secrecy Codes for Networked
while the parent elements46 𝒫𝒶𝓇(𝑠𝑖 ) are the ones defined by Linear Systems,” IEEE Trans. Auto. Control, Vol. 65, no. 5, pp. 2001-2015,
the Cartesian poruct of the main set’s domain, aacording to 2020.
[22] B. Anderson, J. Moore, ”Optimal filtering,” Prentice-Hall, 1979.
the following definition for the possibilistic graphs. [23] H. Sun, Z. Wang, ”Minimal Euler Characteristics of 4-manifolds with
D EFINITION 5: P OSSIBILISTIC GRAPH [68], [69]. A 3-manifold groups,” https://arxiv.org/abs/2103.10273, 2021.
possibilistic casual network is defined in terms of the graph [24] R. Caniato, T. Riviere, ”The Unique Tangent Cone Property
𝑑𝑒 𝑓
 
for Weakly Holomorphic Maps into Projective Algebraic Varieties,”
𝒢𝓅ℴ𝓈𝓈 = 𝑠, 𝒫𝒶𝓇(𝑠), 𝛼𝓅ℴ𝓈𝓈 : 𝜐 𝑠|𝒫𝒶𝓇(𝑠) = 𝛼𝓅ℴ𝓈𝓈 ≠ https://arxiv.org/abs/2108.10371, 2021.
 [25] G. Olikier, P. Absil, ”On the continuity of the tangent cone to the
determinantal variety,” https://arxiv.org/abs/2201.03979, 2022.
1 . [26] F. Fillastre, ”Gauss images of hyperbolic cusps with convex polyhedral
boundary,” Trans. American Math., Vol. 363, no. 10, pp. 5481-5536, 2011.
The proof is now completed.  [27] B. Anderson, M. Ye, ”Exterma without convexity and stability without
Lyapunov,” Com. Info. Sys., Vol. 20, no. 3, 2020.
[28] C. Byrnes, ”On brockett’s necessary condition for stabilizability and the
R EFERENCES topology of Lyapunov functions on R𝑛 ,” Com. Info. Sys., Vol. 8, no. 4,
pp. 333-352, 2008.
[1] Y. Ai, A. Mathur, L. Kong, ”Secure Outage Analysis of FSO Communi- [29] R. A. Vitale, ”The Brunn-Minkowski inequality for random sets,” J.
cations Over Arbitrarily Correlated Málaga Turbulence Channels,” IEEE Multivariate Anal., Vol. 33, no. 2, pp. 286-293, 1990.
Trans. Vehicular Technol., Vol. 70, no. 4, pp. 3961-3965, 2021. [30] S. Boyd, S. P. Boyd, and L. Vandenberghe, ”Convex Optimization.”
[2] Y. Ai, F. A. P. deFigueiredo, L. Kong, ”Secure Vehicular Communica- Cambridge University Press, 2004
tions Through Reconfigurable Intelligent Surfaces,” IEEE Trans. Vehicular [31] W. Liu, ”Decay rates of energy of the 1D damped original nonlinear
Technol., Vol. 70, no. 4, pp. 7272-7276, 2021. wave equation,” Nonlinear Anal. Real W. Apps., Vol. 63, pp. 103-412,
[3] L. Yang, J. Yang, W. Xie, ”Secrecy Performance Analysis of RIS-Aided 2022.
Wireless Communication Systems,” IEEE Trans. Vehicular Technol., Vol. [32] H. J. van Waarde, M. K. Camlibel, ”A Matrix Finsler’s Lemma with
69, no. 10, pp. 12296-12300, 2020. Applications to Data-Driven Control,” https://arxiv.org/abs/2103.13461,
[4] I. Trigui, W. Ajib, W. Zhu, ”Secrecy Outage Probability and Average Rate 2021.
of RIS-Aided Communications Using Quantized Phases,” IEEE Commun. [33] T. S. Blyth, E. F. Robertson, Basic Linear Algebra, 2nd ed., Springer,
Letters, Vol. 25, no. 6, pp. 1820-1824, 2021. 1998.
[5] S. Kavaiya, D. K. Patel, Z. Ding, Y. L. Guan, ”Physical Layer Security [34] L. V. Ahlfors, Complex Analysis, 2nd ed., New York: McGraw-Hill,
in Cognitive Vehicular Networks,” IEEE Trans. Commun., Vol. 69, no. 4, 1970.
pp. 2557-2569, 2021. [35] K. Petersen, M. Pedersen, The Matrix Cookbook,
[6] K. Lee, J. Bang, H. Choi, ”Secrecy Outage Minimization for Wireless- http://matrixcookbook.com, 2012.
Powered Relay Networks With Destination-Assisted Cooperative Jam- [36] Y. Li, G. Geng, Q. Jiang, ”A Parallelized Contour Integral Rayleigh–Ritz
ming,” IEEE IoT. J., Vol. 8, no. 3, pp. 1467-1476, 2021. Method for Computing Critical Eigenvalues of Large-Scale Power Sys-
[7] R. K. Ahiadormey, P. Anokye, H. Jo, C. Song, ”Secrecy Outage Analysis tems,” IEEE Trans. Smart Grid, Vol. 9, no. 4, pp. 3573-3581, 2018.
in NOMA Power Line Communications,” IEEE Commun. Letters, Vol. 25, [37] T. Ikegami, T. Sakurai, “Contour Integral Eigensolver for Non-Hermitian
no. 5, pp. 1448-1452, 2021. Systems: a Rayleigh-Ritz-type Approach,” Taiwanese J. Math., vol. 14, no.
[8] R. Ruby, Q. Pham, K. Wu, A. A. Heidari, H. Chen, ”Enhancing Se- 3A, pp. 825-837, 2010.
crecy Performance of Cooperative NOMA-based IoT Networks via Multi- [38] Y. Li, G. Geng, Q. Jiang, ”A Parallel Contour Integral Method for
Antenna Aided Artificial Noise,” IEEE IoT. J., Vol. pp, no. 99, pp. 1-1, Eigenvalue Analysis of Power Systems,” IEEE Trans. Power Sys., Vol.
2022. 32, no. 1, pp. 624 - 632, 2017.
[39] Alon, Noga, Spencer, Joel H. The Probabilistic Method 2nd ed. John
46 Casual prior samples. Wiley & Sons, Inc., 2000.
IEEE, VOL. XX, NO. XX, X 2021 12

[40] S. Dasgupta, A. Gupta, ”An elementary proof of a theorem of Johnson

and Lindenstrauss,” Random Structs. & Algs. 2002.
[41] S. S. Tohidi, Y. Yildiz, ”Handling actuator magnitude and rate satura-
tion in uncertain over-actuated systems: A modified projection algorithm
approach,” https://arxiv.org/abs/2009.03024, 2020.
[42] R. Vershynin, ”High-dimensional probability: An introduction with
applications in data science,” Cambridge University Press, Vol. 47, 2018.
[43] I Satake, ”The Gauss-Bonnet theorem for V-manifolds,” J. Math. S.
Japan, 1957
[44] C. Allendoerfer, A Weil, ”The gauss-bonnet theorem for riemannian
polyhedra,” Trans. American Math., 1943.
[45] M Gromov, ”Systoles and intersystolic inequalities,” Actes de la table
ronde de geometrie differentielle, 1996.
[46] K. Katz, M. Katz, ”Relative systoles of relative-essential 2−complexes,”
Alg. Geom. Topology, Vol. 11, pp. 101-999, 2011.
[47] M. Bailey, ”Symplectic capacity and convexity”, Toronto Press, 2019.
[48] U. Frauenfelder, ”Finitness of 𝜋1 −sensitive Hofer-Zehnder capacity and
equivariant loop space homology,” J. F. P. Theory Apps., Vol. 19. pp. 3-15,
2017.
[49] H. Inoue, T. Kamada, ”structural instability of friction-induced vibration
by characteristics polynomial plane applied to break squeal,” J. A. Mech.
Design. Sys. Man., Vol. 14, no. 1, 2020.
[50] C Clark, ”The asymptotic distribution of eigenvalues and eigenfunctions
for elliptic boundary value problems,” Siam Review, Vol. 9, no. 4, 1967.
[51] A. Grabsch, ”General truncated linear statistics for the top eigenvalues
of random matrices,” https://arxiv.org/abs/2111.09004, 2021.
[52] E. Gundogdu, V. Constantin, S. Parashar, A. Seifoddini, M. Dang, M.
Salzmann, P. Fua, ”GarNet++: Improving Fast and Accurate Static3D Cloth
Draping by Curvature Loss,” IEEE Trans. Pattern Analysis Machine Intel.
Vol. 44, no. 1, pp. 181-195, 2022.
[53] S. Khodadadian, T. T. Doan, J. Romberg, S. T. Maguluri, ”Finite
Sample Analysis of Two-Time-Scale Natural Actor-Critic Algorithm,”
https://arxiv.org/abs/2101.10506, 2022.
[54] M. Mehta, ”Random matrices,” Elsevier, New york, 2004.
[55] J. Akemann, ”The Oxford handbook of random matrix theory,” Oxford
University Press, 2011.
[56] A. D. Wallace, ”The map excision theorem,” Duke Math. J. Vol. 19, no.
1, pp. 177-182, 1952.
[57] D. Labardini-Fragoso, D. Velasco, ”On a family of Caldero-Chapoton
algebras that have the Laurent phenomenon,” J. Algebra, Vol. 520, pp.
90-135,2019.
[58] A. Blass, ”Seven trees in one,” J. P. A. Algebra, Vol. 103, pp. 1-21,
1995.
[59] G. Cerulli Irelli, D. Labardini-Fragoso, J, Schroer, ”Caldero-Chapoton
algebras,” Trans. Amer. Math. Soc. Vol. 367, pp. 2787-2822, 2015.
[60] V. E. Coll, L. B. Whitt, ”The Flat Plane and a Constructive Proof of
Minding’s Theorem,” https://arxiv.org/abs/1902.06089, 2019.
[61] Y. Zhu, W. Zheng, ”Observer-Based Control for Cyber-Physical Systems
With Periodic DoS Attacks via a Cyclic Switching Strategy,” IEEE Trans.
Auto. Control, Vol. 65, no. 8, pp. 3714-3721, 2020.
[62] S. Wang, J. Park, ”Modeling and analysis of multi-type failures in
wireless body area networks with semi-Markov model,” IEEE Commun.
Letters, Vol. 14, no. 1, pp. 6-8, 2010.
[63] D. Dubois, H. Prade, ”Possibility Theory and Its Applications: Where
Do We Stand?,” Springer Handbook. Comput. Intel., pp. 31-60, 2015.
[64] L.A. Zadeh, ”Fuzzy sets as a basis for a theory of possibility,” Fuzzy
Set. Syst. no. 1, pp. 3-28, 1978.
[65] L.A. Zadeh, ”Fuzzy sets and information granularity,” Adv. Fuzzy Set
Theory. Apps., pp. 3-18, 1979.
[66] L.A. Zadeh, ”Possibility theory and soft data analysis,” Math. Fron.
Social Policy Sc., pp. 69-129, 1982.
[67] W. Mei, ”Formalization of Fuzzy Control in Possibility Theory via Rule
Extraction,” IEEE Access, Vol. 7, pp. 90115-90124, 2019.
[68] S. Benferhat, D. Dubois, L. Garcia, H. Prade, ”Possibilistic logic bases
and possibilistic graphs,” https://arxiv.org/abs/1301.6679, 2013.
[69] S. Benferhat, D. Dubois, S. Kaci, H. Prade, ”Graphical readings of
possibilistic logic bases,” https://arxiv.org/abs/1301.2255, 2013.