Rewrite basic concepts of degree generating function of network

and Projects
Ha Pham
January 24, 2024

1 Recall random graph - Network

A random graph is a model network in which some specific set of parameters take fixed values, but the
network is random in other respects.
Example 1.1 (G(m, n - fixe number of nodes and number of edges). A random graph with fixed number of
egdes m and number of vertex n. We choose m pairs of vertices uniformly at random from all possible pairs
and connect them with an edge. Typically one stipulates that the network should be a simple graph, i.e., that
it should have no multiedges or self-edges

Figure 1: Network with self-edge, multiedges and loops

This model is often referred to by its mathematical name G(n, m). The average degree of a node in such
network is
2m
hki =
n
However other properties are not so easy to calculate.
Most mathematical work has actually been conducted on a slightly different model that is considerably
easier to handle such as Rényi-Erdös (ER) network or Poisson random graph or G(n, p).
Example 1.2 (Rényi-Erdös network G(n, p) - fix number of node and probability of connecting two random
chosen node). In G(n, p), we fix the number of vertices, not the number but the probability p of edges between
vertices.
The technical definition of the random graph is not in terms of a single network, but in terms of an
ensemble, a probability distribution over all possible networks. To be specific, G(n, p) is the ensemble of
networks with n vertices in which each simple graph G appears with probability
n
P (G) = pm (1 − p)( 2 )−m
1
 where m is the number of edges in the graph.
Probability that we obtain a graph with exactly m edges in G(n, p) is
 n

n
P (m) = 2 pm (1 − p)( 2 )−m
m

which is a binomial distribution. So the average value of number of edges m is

(n2 ) n
  
n n
m 2 pm (1 − p)( 2 )−m =
X
hmi = p
m=0
m 2

The average degree of a node in a graph of m edges and n nodes is 2mn . So the mean degree of a vertex in
the random graph is
(n2 )
X 2m
hki = c = P (m) = (n − 1)p
m=0
n
A given node is connected randomly and indepenently to each of other n − 1 nodes with probability p. So
probability that it has exactly k neighbors is

ck
 
n−1 k
pk = p (1 − p)n−1−k → e−c
k k!
as n → ∞.
As n → ∞, G(n, p) has tree form with no loop, no multiedge and no self-edge with probability 1. The
network has at most one giant component.
Example 1.3 (Degree distribution). Define pk as fraction of node with degre k in the network. Then pk
can be seen as the probability that a random choosen node in the network has degree k.
Remark that degree distribution does not tell us the complete structure of a network

Figure 2: Two different networks with same degree distribution p1 = 0.4, p2 = 0.6

Example 1.4 (Configuration model - fix sequence degree of nodes in network). The configuration model is
a network with a given degree sequence k1 , k2 , . . . rather than degree distribution. That is exactly degree of
each node in the network is fixed rather than the probability distribution from which the degrees are choosen.

Figure 3: Sequence degree 0, 1, 1, 2, 2, 2, 2, 3, 3, 4 but degree distribution p0 = 0.1, p1 = 0.2, p2 = 0.4, p3 =

0.2, p4 = 0.1

2
 Construction random network from sequence degree: Suppose that Pki is the degree of the node
i for i = 1, . . . , n. Given each node i a total of ki ”stubs”. There are totally ki = 2m stubs. We choose
two stubs uniformly at random and create an edge by connecting them. Then we choose another pairs from
the remaining 2m − 2 stubs and so on.

Figure 4: Each vertex is given a number of “stubs” of edges equal to its desired degree. Then pairs of stubs
are chosen at random and connected together to form edges (dotted line).

Given sequence degree, the total number of edges is 21

P
ki = m is fixed. Thus this model is in some ways
analogous to G(n, m), which also fixes the number of edges.

2 Rewrite degree generating function of (single) network and ap-

plications to analyze network resilience and robustness from
[newman2018networks ]
2.1 Giant component for general undirected network
2.1.1 Generating function of degree and excess degree distribution
Let pk be probability that random node has degree k. The degree generating function is defined by
X
g0 (z) = pk z k (1)
k

Probability that we reach a vertex of degree k upon following an edge is proprotion to kpk . The number of
other neighbor of the node at the end of such edge is called excess degree.
Excess degree of a node is in fact equal to its degree minus 1. Hence probability that a random node has
excess degree k is

(k + 1)pk+1
qk = (2)
hki
P
where hki is the average degree of a node to guarantee the normalization condition qk = 1.
Let g1 (z) be generating function of excess degree distribution, that is
X
g1 (z) = qk z k (3)
k

then using the relationship (2), we have g1 (1) = 1 and

g00 (z)
g1 (z) = (4)
hki

3
2.2 Generating function for number of second neighbor
(2)
Denote pk be probability that a node has exactly k second neighbors and P (2) (k|m) be the conditional
probability of having exactly k second neighbor given that it has first m neighbors (or degree of m). We
have
(2)
X
pk = pm P (2) (k|m) (5)
m≥0

The number of second neighbor is equal to the total excess degree of the first neighbor. Suppose that excess
degree of all nodes are independent Qm then given that the node has m first neighbors, the probability that their
excess degree are j1 , . . . , jm is i=1 qji . Thus

X m
Y
P (2) (k|m) = qji (6)
j1 +···+jm =k i=1

Hence
m
(2)
X X Y
pk = pm qji (7)
m≥0 j1 +···+jm =k i=1

The generating function for the number of second neighbors is

m
(2)
X XX X Y
g (2) (z) = pk z k = pm qji z k (8)
k≥0 k≥0 m≥0j1 +···+jm =k i=1
X X X Ym
= pm qji z ji (9)
m≥0 k≥0 j1 +···+jm =k i=1
X
= pm (g1 (z))m (10)
m≥0

= g0 (g1 (z)) (11)

We derive the powers property of generating function: given a node of degree m, the number of second
neighbour is the total exceed degree of m node. Hence the generating function of the conditional number of
second neighbor given that a node has degree m is (g1 (z))m
X
P (2) (k|m)z k = (g1 (z))m (12)
k

For number of the third neighbor, we use similar argument. Denote P (3) (k|m) be the probability that
there are exactly k third neighbor given that there are m second neighbors. Then the generating function
of the number of the third neighbors is given by
XX
g (3 (z) = p(2)
m P
(3)
(k|m)z k (13)
k≥0 m≥0
X X
= p(2)
m P (3) (k|m)z k (14)
m≥0 k≥0
X
= p(2)
m (g1 (z))
m
(15)
m≥0

= g (2) (g1 (z)) (16)

= g0 (g1 (g1 (z))) (17)

4
We can achieve the generating function of the nth neighbors recursively as following
XX
g (n (z) = p(n−1)
m P (n) (k|m)z k
k≥0 m≥0
X X
= p(n−1)
m P (n) (k|m)z k
m≥0 k≥0
X
= p(n−1)
m (g1 (z))m
m≥0

= g (n−1) (g1 (z))

= g0 (g1 (. . . (g1 (z) . . . )) with n − 1 copies of g1

where P (n) (k|m) is the probability that the number of nth neighbor is k given the number of (n − 1)th is m.
Such generating function allows us to compute the mean degree of neighbors at any distance n
 n−1
c2
cn = (g (n) )0 (1) = (g (n−1) )0 (g1 (1))g10 (1) = (g (n−1) )0 (1)g10 (1) = cn−1 c1 = · · · = c1 (18)
c1

2.2.1 Generating function for small components

Denote πs as probability that a random node belong to a small component of size s. The generating function
for size of small component is
X
h0 (z) = πs z s (19)
s≥1

Let ρs be the probability that a node at the end of an edge belongs to a small component of size s after the
edge is removed. Generating function for this distribution is defined as
X
h1 (z) = ρs z s (20)
s≥0

Denote P (s − 1|k) be the conditional probability that a node belong to a component of size s given that
its degree is k. In other word, after removing this node, the total size of its neighbor components are s − 1.
By total law, we have
X
πs = pk P (s − 1|k) (21)
k

Hence
∞ X
X ∞
h0 (z) = pk P (s − 1|k)z s
s=1 k=0
∞ X
X ∞
=z pk P (s|k)z s
s=0 k=0

P (s|k) is the probability that the k neighbors belong to small components whose size sums to s. Denote si be
the size of the neighbor i which happens with probability ρsi then s1 + · · · + sk = s. Using the independent
of size of small components, we have

X k
Y
P (s|k) = ρsi
s1 +···+sk =s i=1

5
So
∞ X
X ∞ X k
Y
h0 (z) = z pk ρsi z s
s=0 k=0 s1 +···+sk =s i=1
∞
X ∞
X X k
Y
=z pk ρsi z si
k=0 s=0 s1 +···+sk =s i=1
X∞
=z pk [h1 (z)]k
k=0
= zg0 (h1 (z))

Intuitively, h0 refers for generating function of the sum of ”neighbor component” size while h1 refers for
generating function for the size of a ”neighbor component” following an edge. Given an edge of degree k,
the size of the component that it belongs will be equal to the sum of the size of k ”neighbor component”
and add by 1 to count itself into the ”fullzise component”.
Moreover, in large networks, the removal of a node has no effect on the degree distribution. So if a
neighbor of a removed node has degree k then its probability of belonging to a component of size s is
P (s − 1|k) just as before. After removing, the degree of such node is its orgin excess degree qk . Hence
X
ρs = qk P (s − 1|k) (22)
k

It leads that
X XX
h1 (z) = ρs z s = qk P (s − 1|k)z s = zg1 (h1 (z))
s s k

In summary, we have

h0 (z) = zg0 (h1 (z)) (23)

h1 (z) = zg1 (h1 (z)) (24)

2.2.2 Giant component

Size of giant component is
X
S =1− πs = 1 − h0 (1) = 1 − g0 (h1 (1)) = 1 − g0 (u) (25)

where u = h1 (1) is the the solution of

u = g1 (u) (26)

The equation u = g1 (u) always has a solution u = 1. If it is the unique solution on [0, 1] then S = 0 and
hence there does not exist the giant component.
The giant component exists if and only if the equation g1 (u) = u has another solution on [0, 1). Because
g1 is an increasing and convex function, it happens if and only if

g10 (1) > 1

That is the condition for the existence of giant component.

We have
X 1 X
g10 (z) = kqk z k = k(k + 1)pk+1 z k
hki
k k

6
Hence
X 1 X
g10 (1) = kqk z k = k(k + 1)pk+1
hki
k k
!
1 X X
= (k + 1)2 pk+1 − (k + 1)pk+1
hki
k k
k 2 − hki
=
hki
Hence the condition for the existence of giant component becomes

k 2 − hki
> 1 ⇐⇒ k 2 > 2 hki (27)
hki

Example 2.1 (Poisson random graph). Degree distribution

ck
pk = e−c
k!
where c = hki is the average degree.
The generating function of degree distribution is
X ck k X (zc)k
g0 (z) = e−c z = e−c = e−c ecz = ec(z−1)
k! k!
k k

The generating funciton of excess degree distribution is

g 0 (0)(z) cg0 (z)
g1 (z) = = = g0 (z)
hki c
In order to find the size of giant component, we first find u such that

u = g1 (u) = ec(u−1)

Giant component exists if and only if g10 (1) > 1 or c > 1. In this case, the size of giant component is
S = 1 − ec(u−1)
Example 2.2 (Power law). Pure power law distribution
(
0 if k = 0
pk = k−α (28)
ζ(α) if k > 0

where ζ(α) = k −α .
P
Condition for existence of giant component
We have
1 X −α 1 X −(α−1) ζ(α − 1)
hki = kk = k =
ζ(α) ζ(α) ζ(α)
and
1 X 2 −α 1 X −(α−2) ζ(α − 2)
k2 = k k = k =
ζ(α) ζ(α) ζ(α)
Hence the giant component exists if and only if

k 2 > 2 hki ⇐⇒ ζ(α − 2) > 2ζ(α − 1)

7
which indicates that α < 3.4788 . . .
In practice for power-law tail, the exponent α is between 2 and 3.
In this case, we have
Liα (z) Liα−1 (z)
g0 (z) = , g1 (z) =
ζ(α) zζ(α − 1)
The equation to find the probability u = h1 (u) is

k −α+1 uk
P
Liα−1 (u)
u= =
ζ(α − 1) ζ(α − 1)

There is no explicit solution.

2.3 Percolation and Network Resilience

2.3.1 Percolation
Process of removing some fraction of nodes/edges of a given network along with the edges connected to those
nodes is called percolation or site percolation. Theory of percolation aims to answer the question what
effect of percolation on the performance of the network.
There are two types of percolation: remove node and remove edge which are called site percolation and
bond percolation.
One typical percolation strategy is uniformly random removal of nodes. The mechanism is as following:
a fraction of nodes are selected randomly and removed.
In percolation theory, the fraction of remaining nodes in the graph is parametized by a probability φ.
Such nodes are called to be occupied and the probability φ is called the occupation probability.
For large φ, the giant component may exists. But as φ decreases, the size of giant component is smaller
and smaller and then the giant componnent breaks a part which is called a percolation transition. When a
network contains a giant componnent, it percolates. The point at which the percolation transition occurs is
called the percolation threshold.
In the studies of percolation, the ”components” remains after nodes have been removed are called ”clus-
ter ”. It is helpful to distinguish between the components of the underlying network and the cluster of the
percolation process. The giant component in the percolation process is called the giant cluster. In the
following sections, we recall analysis robustness of several based-node attack strategies.

2.3.2 Uniform removal in the configuration model

Consider a network with degree distribution pk and a percolation process with uniform removal and occupant
probability φ.
Consider a node in the network after percolation process. Then it does not belong to a giant cluster if
and only if it is not connected to giant cluster via any of its neighbors.
Let u be the average probability that a node does not belong to the giant cluster then given a node of
degree k, the total probability that it does not belong to the giant cluster is uk . So average probability that
a node does not belong to a giant component given it is not remove is
X
pk uk = g0 (u)
k

where X
g0 (z) = pk z k
k

Probability that a node is not remove is φ. Hence the probability that a node belong to the giant cluster is

S = φ(1 − g0 (u))
8
Calculation for probability that a node is not connected to the giant via a particular neighbor is as following:
There are two ways that a neighbor of a node does not belong to the giant cluster
1. the neighbor is removed with probability 1 − φ
2. the neighbor is not removed with probability φ but itself is not in the giant component. If such neighbor
is has other k neighbors with probability qk then this case happens with probability φuk
Hence
X X
u= qk (1 − φ + φuk ) = 1 − φ + φ qk uk = 1 − φ + φg1 (u) (29)
k k

where g1 (z) = qk z k is the generating function of the excess degree distribution.

P
The above solution has a trivial solution u = 1. If it does not has other solution on [0, 1) then the giant
cluster does not exist and inversely, the giant cluster exists if and only if (29) has a solution in [0, 1) or
d
dz (1 − φ + φg1 (z)) z=1 > 1. The percolation threshold occurs when

d
(1 − φ + φg1 (z)) =1 (30)
dz z=1

The critical value for percolation probability is

1 hki
φc = = 2 (31)
g10 (1) hk i − hki
Example 2.3 (Poisson random graph). We have
g10 (z) = cec(z−1)
So g10 (1) = c then the critical value
1
φc = .
c
Example 2.4 (Exponential degree distribution).
pk = (1 − e−λ )e−λk
We have
eλ − 1
g0 (k) =
eλ − z
eλ −1
So g00 (z) = (eλ −z)2
. Hence hki = g00 (1) = 1
eλ −1
and
2
g 0 (z) eλ − 1

g1 (z) = 00 =
g0 (1) eλ − z
Critical value for percolation probability
1
φc = = (32)
g10 (1)
The probability u is the solution of the equation
2
eλ − 1

u=1−φ+φ
eλ − u
which satisfies
u2 + (φ − 2eλ )u + φ − 2φeλ + e2λ = 0
and less than 1, which is equal to
Hence we can find explicitly the formula for the size of the giant componnent.

9
2.3.3 Non-uniform removal of nodes
Size of giant cluster
As before, we denote u as the average probability that a node does not belong to the giant cluster via any
neighbor.
Let φk be the probability that a node of degree k is occupied (not removed). Given a node of degree k,
it does not belong to a giant component via a particular neighbor with probability uk . So probability that
it belongs to the giant component is φk (1 − uk ). The average probability that a random node is in the giant
component is
X X X
S= pk φi (1 − uk ) = p k φk − pk φk uk = f0 (1) − f0 (u) (33)
k k k

where
X
f0 (z) = pk φk z k (34)
k

is the generating function for degree distribution of occupied node with

X
f0 (1) = pk φk = φ (35)
k

is the average probability that a vertex is occupied.

To calculate u, we use the same argument as in random removal percolation. A node A is not connected
to the giant cluster via any of its neighbor if either the neighbor is not occupied or this neighbor is occuppied
but is not connected to giant component via any of its neighbor. Let k be the excess degree of a neighbor,
hence the degree of this neighbor is k + 1. Then it is not occupied with probability 1 − φk+1 and the
probability that it is occupied but not connected to the giant componnent via a particular neighbor (except
the node A) is φk+1 uk . Thus
X
u= qk (1 − φk+1 + φk+1 uk )
k≥0
X X
=1− qk − φk+1 + qk φk+1 uk
k≥0 k≥0

= 1 − f1 (1) + f1 (u)

where
X
f1 (z) = qk φk+1 z k (36)
k≥0

(k+1)pk+1
Using the fact that qk = hki , we have

1 X f 0 (z)
f1 (z) = kpk φk z k−1 = 00 (37)
hki g0 (1)
k≥1

Condition for the existence of the giant component f10 (1) > 1
Find the size of the giant componnent First solve the equation

u = 1 − f1 (1) + f1 (u)

and then compute

S = f0 (1) − f0 (u)
10
 In the case of random attack with occupant probability φ, we have
f0 (z) = φg0 (z)
and hence
g00 (z)
f1 (z) = φ = φg1 (z)
g00 (1)
For the degree attack (remove all vertex with degree equal to or greater than a given value), the equation
to find u typically has no explicit solution.
Example 2.5 (Degree attack for exponential degree distribution). Degree distribution is given by
pk = (1 − e−λ )e−λk , k≥0
We remove all vertices that have degree k0 or greater. That is
(
1 if k < k0
φk = (38)
0 if k ≥ k0
Then we have
0 −1
kX
1 − e−λk0 z k0
eλ − 1
f0 (z) = (1 − e−λ ) e−λk z k = (1 − e−λ ) −λ
= 1 − e−λk0 z k0 λ
1−e z e −z
k=0

We know that hki = g00 (1) = 1

eλ −1
. Hence
2
f 0 (z)  eλ − 1

f1 (z) = 00 = 1 − e−λk0 z k0 − k0 e−λk0 z k0 −1 (e−λ − z)

g0 (1) 1 − eλ z
Need to solve u = 1 − f1 (1) + f1 (u) numerically to compute S
Example 2.6 (Degree attack - Poisson random graph).
ck
pk = e−c
, k ≥ 0, c = hki
k!
We remove all vertices that have degree k0 or greater. Hence
0 −1
kX
ck k
f0 (z) = e−c z
k!
k=0

and
0 −2
kX
ck k
f1 (z) = e−c z
k!
k=0

Example 2.7 (Scale free network).

k −α
pk = , k≥1
ζ(α)
with hki = g00 (1) = ζ(α−1)
ζ(α)
We remove all vertices that have degree k0 or greater. Hence
0 −1
kX
k −α k
f0 (z) = z
ζ(α)
k=1

and
k0 −1 −α 0 −1
kX
ζ(α) X k k −α+1 k−1
f1 (z) = kz k−1 = z
ζ(α − 1) ζ(α) ζ(α − 1)
k=1 k=1

11
2.3.4 Generating function for size of small percolation cluster
We repeat the argument to find generating function of size of small component in a random network to study
size of small percolation cluster.
Let h1 (z) be generating function for the probability that the end of a randomly chosen edge leads to a
percolation cluster of a given number of occupied nodes (generating function for size of a ”neighbor cluster”).
The cluster may contain 0 edge if the mode at the end of the edge is removed with probability 1 − f1 (1) or
may leads to an occupied node with excess degree of k. Hence the size of the origin cluster will be equal to
the sum of the size of k subclusters coming from the node at the end of the edge and 1. Hence
h1 (z) = 1 − f1 (1) + zf1 (h1 (z)) (39)
Let h0 (z) be the generating function for the probability distribution for the size of cluster to which a
random choson node belong to. Then
h0 (z) = 1 − f0 (1) + zf0 (h1 (z)) (40)
Size of giant cluster is
S = 1 − h0 (1) = f0 (1) + f0 (h1 (1)) = f0 (1) + f0 (u) (41)
where u = h1 (1) is the solution to the equation
u = 1 − f1 (1) + f1 (u) (42)
as in the approach in the previous section.

3 Project 1: from Idea of Prof Davide Cassi

3.1 Random but non-uniform removal according to power to node degree
3.1.1 Without cutoff degree - Strategy 1
The probability that a node of degree k ≥ 0 is not removed is
1
φk = (43)
ak +1
for a ≥ 0. So a node with higher degree is more likely to be deleted if a > 1. Inversely if a < 1 then the
higher the degree is, the less likely the node is removed. A node of degree 0 is not removed certainly. The
larger the value of a is, the more likely a node is removed.
Strategy description: We need firstly determine degree of all the node and visit all the node of the
network. For a node of degree k > 0, it is removed with probability 1 − φ)k and kept with probability φk .
For simulation, we need
1. Visit all the node
2. For each node
(a) Determine its degree k
(b) If k > 0, simulate a Bernoulli random number with parameter φk
• if the number is 0, the node is removed
• if the number is 1, the node is kept
For which value of parameter a, this strategy is more efficient than uniformly removal attack?
Question: In this strategy, we need to visit all the node in the graph, it takes a lot of time for large
graphs. Is it possible to replacing the task visiting all by random visiting with random walk on random
graph?
12
3.1.2 With cutoff degree - Strategy 2
We can also combine the above strategy with highest degrees attack strategy in the following way:
(
1
i if i ≤ k0
φi = 1+a (44)
0 if i > k0

Strategy description: remove all the node of degree higher than k0 . For the node with degree k ∈ [0, k0 ],
it is removed with probability 1 − φk . For a = 0, we cover the highest degree attack strategy.
So we modify the above simulation algorithm as following

1. Visit all the node

2. For each node
(a) Determine its degree k
(b) Compare its degree with cutoff degree k0
i. If k > k0 , remove this node
ii. If 0 < k ≤ k0 : Simulate a Bernoulli random number with parameter φk
• if the number is 0, the node is removed
• if the number is 1, the node is kept
The fraction of node to be removed is
X
q =1− φi pk (45)

where pk is the fraction of nodes of degree k in the network.

The objective is to find the percolation transition, LLC and optimal value of a.
Intuitively, this strategy will be more efficient than both highest degrees attack and randomly attack
based on node degree without cutoff degree in strategy 1. In fact, highest degrees attack strategy can be
consider as a particular case of strategy 2 with a = 0.
Question: is it applicable to a simular kind of strategy but for betweenness instead of degree?

3.1.3 Analytic Solution

One can apply directly the result for non-uniform removal in the textbook of Newman[newman2018networks
]. Let pk be the average probability that a node has degree k and φk be the average probability that a node
of degree k is occuppied (not removed). Denote qk = (k+1)p hki
k+1
be the probability that a node has excess
degree k where hki is the average degree of a node.
Define
X
g0 (x) = pk xk (generating function for degree distribution) (46)
k
X
f0 (x) = φk pk xk (generating function for degree of occupied node) (47)
k
1 X f 0 (x)
f1 (x) = φk pk xk = 0 (generating function for excess degree of occupied node) (48)
hki hki

Let h0 be the generating function for size of small percolation cluster then

h0 (x) = 1 − f0 (1) + zf0 (h1 (x)) (49)

13
where h1 (x) is the generating function for size of ”neighbor cluster” and satisfies

h1 (x) = 1 − f1 (1) + zf1 (h1 (x)) (50)

The condition for existence of giant component is f10 (1) > 1. In this case, the size of giant component
(LCC) is S = 1 − u where u is the solution to the equation

u = 1 − f1 (1) + f1 (u) (51)

Example 3.1 (ER network or Poisson network - attack without using cutoff degree - Strategy 1). Consider
an ER graph with parameter p: two random nodes are connected with probability p. An ER(p) network is one
of some ways to build a graph G(m, n) whose number of nodes and edges are fixed (m nodes and n edges).
However it differs from the Quang’s setting for bimodular: fixed the number of wings (open edges) of each
nodes first and then match randomly a pair of wings to create an edge.
Let c = hki be the mean of average degree of a node then

ck
pk = e−c
k!
For a large network with N nodes then c = N p.
We have
X ck
g0 (x) = e−c xk = e−c ecx = ec(x−1)
k!
k≥0
X ck 1 X (cx)k
f0 (x) = e−c x k
= e −c
k! ak + 1 k!(ak + 1)
k≥1 k≥0

f00 (x) X (cx) k

f1 (x) = = e−c
c k!(ak+1 + 1)
k≥0

The giant component exists if f10 (1) > 1 or

X ck
ce−c >1 (52)
k!(ak+2 + 1)
k

Critical value for power a is the solution to the equation

X ck
ce−c =1 (53)
k!(ak+2 + 1)
k

which is useful to find the phase transition

X X (cx)k
q = 1 − f0 (1) = 1 − φk pk = 1 − e−c
k!(ak + 1)
k k≥0

Need to solve the equation

1 − f1 (1) + f1 (u) = u
numerically to find the size of giant component is S = 1 − u.
Example 3.2 (ER network or Poisson network - attack with cutoff degree k0 - Strategy 2). Consider an
ER graph with parameter p: two random nodes are connected with probability p. The construction of ER(p)
network is quite different from a graph G(m, n) whose number of nodes and edges are fixed (m nodes and n
edges).
14
 Let c = hki be the average degree of a node then

ck
pk = e−c
k!
We have
X ck
g0 (x) = e−c xk = e−c ecx = ec(x−1)
k!
k≥0
k0 k
−c
X ck 1 k −c
X0
(cx)k
f0 (x) = e x = e
k! ak + 1 k!(ak + 1)
k=0 k=0
0 −1
kX
f00 (x) (cx) k
f1 (x) = = e−c
c k!(ak+1 + 1)
k=0

The giant component exists if f10 (1) > 1 or

0 −2
kX
ck
ce−c >1 (54)
k!(ak+1 + 1)
k=0

The critical value of the power a when k0 is fixed or critical value of k0 when a is fixed is the solution to the
equation
0 −2
kX
ck
ce−c =1 (55)
k!(ak+1 + 1)
k=0

If k0 is large enough then this critical value can be used to approximate the critical value of a in the case of
attacking without degree threshold.
Need to solve the equation
1 − f1 (1) + f1 (u) = u
numerically to find the size of giant component is S = 1 − u.

Remark 3.1. Since there is no closed form for generating function in strategy 1, we consider strategey 1 as
the limit of strategy 2 when the cutoff degree k0 tends to infinity.
Remark 3.2. Concerning on computing, one may face to error for large k0 due to out of memory problem
to calculate k!. In this case, we may need to use normal approximation to poisson distribution for high degree

pk ≈ P (k − 0.5 ≤ X ≤ k + 0.5)

where X ∼ N (c, c).

Example 3.3 (Scale-free network- Strategy 2). Use the same argument as in the case of ER network. The
degree distribution is given by
(
0 if k = 0
pk = k−α (56)
ζ(α) if k>0

15
 k −α . We have
P
where ζ(α) =
X X k −α
g0 (x) = pk xk = xk (57)
ζ(α)
k k≥1
k0
X X k −α
f0 (x) = φk pk xk = xk (58)
(ak + 1)ζ(α)
k k=1
k
f 0 (x) X0
k 1−α
f1 (x) = 0 = xk−1 (59)
hki (ak + 1)ζ(α − 1)
k=1

Need to solve the equation

1 − f1 (1) + f1 (u) = u
numerically to find the size of giant component is S = 1 − u.

3.2 Network approximation

In the case of facing difficulty to handle with a real structure network (such as in the case of bi-modular), one
try to approximate network structure by an easier structure. The robustness under a given attack strategy
of the origin and approximating network is different. How can we capture the difference of robustness (e.g,
LCC) and make conclusion whether the robustness on approximating network is good enough (close enough)
to the robustness on the origin network.
Such kind of approximation can be applied for real networks. One can firstly estimate the node degree
distribuiton from real network and then calculate the robustness associated to a given attack strategy which
can be used to estimate the true robustness of the network.

Example 3.4. Consider a network of 103 nodes. Among these node, there are 100 nodes of degree 0, 700
nodes of degree 3, 150 nodes of degree 7, 30 node of degree 10 and 20 nodes of degree greater than or equal
to 100. Degree distribution of such network can be estimated by

p0 = 0.1, p3 = 0.7, p7 = 0.15, p10 = 0.03, p100 = 0.02 (60)

Then we can estimate the LCC of this network with respect to degreee attack strategies using degree distri-
bution.
Of course, degree distribution can not capture all the structure of the network. Hence the LCC estimated
by this way can be different from the true LCC that we obtain from attacking the real network but it can be
considered as an approach to forecast LCC.

3.3 Partial degree observation

Consider a network whose node degree is not observed perfectly due to noise such as it is not fully obverved
the number of neighbors of a given individual is suffered from Covid. For a node of true degree k, the
observed degree is effected by a noise k as
dˆk = k − k
where k is a random variable taking values in {0, 1, . . . , k}.
In such kind of hidden model, we face to a problem of mixture distribution. Given a node of observed
degree k, use Bayes’ theorem to infer about its true degree. To continue ...

16
3.3.1 Uniformly random attack
3.3.2 Degree attack

3.4 Question
Does it make sense to modify random walk attack strategy (proposed in Turchetto et all 2013) with the
same idea as Cassi: instead of choosing uniformly a neighbor of vertex u, we choose the next node with
probability depending on its excess degree. For example, consider a vertex in Figure 5 which has three first
neighbors. From the left to the right, their excess degrees are 1, 3 and 2 respectively.

Figure 5: First and Second neighbor illustration

Let
1
qk =
ak + 1
for some a > 0.
Denote these neighbors from the left to the right be a, b, c Then the conditional probabilities that the
next node to be chosen among these three neighbors are
q1 q3 q2
q̃a = , q̃b = , q̃c =
q1 + q3 + q2 q1 + q3 + q2 q1 + q3 + q2
Such setting for random walk on network makes any sense? Is it applicable for node attack?

4 Rewrite multi-degree generating function for analyzing inter-

acting network robustness
4.1 Multidegree generating distribution of interacting network and applications
In this section, we represent the result about generating function on interacting network in [leicht2009percolation
] and resilience evaluation with generating function on interacting (complex) network in [shai2014resilience
]. They are extention arguments of [newman2018networks ] on using degree generating function from
isolated network to complex network. Such results will be applied to analyze the resilience on attacking
inter-model in the next section(s).

4.1.1 Generating function of multidegree and excess degree distribution [leicht2009percolation

]
Consider a complex network with l modules. Take a random node in network i, the probability that it has
exactly k1 neigbors in network 1, k2 neighbors in network 2, . . . , kl neighbor in network l is denoted by
pµk1 ,k2 ,...,kl . The generating function for multidegree distribution of a node in network µ (called a µ-node) is

17
defined by
∞
X
Gµ (x1 , x2 , . . . , xl ) = pµk1 ,k2 ,...,kl xk11 xk22 . . . xkl l (61)
k1 ,...,kl =0

For convenience, we denote x = (x1 , . . . , xl ) and 1 = (1, . . . , 1).

We have
∞ ∞
∂Gµ X X
(x) = kν pik1 ,...,kν ,...,kl xk11 . . . xkνν −1 . . . xkl l (62)
∂xν
kν =1 kp =0,p6=ν

and
∞ ∞
∂Gµ X X µ
(1) = kν pµk1 ,...,kν ,...,kl = hkiν (63)
∂xj
kν =1 kp =0,p6=j

be the average number of neighbor in module ν of a node in module µ.

Following an edge from a node in module ν to a node in network µ (called an ν − µ edge), the remaining
multidegree of such µ-node (denoting for a node in network µ) is called the excess degree.

Figure 6: An illustration of the remaining edges incident to a node in a network µ reached by following a
random edge between networks ν and µ ([leicht2009percolation ])

For a µ - node following an ν−µ edge with excess degree (k1 , . . . , kν , . . . , kl ), its multidegree is (k1 , . . . , kν +
1, . . . , kl ). Let qkνµ
1 ,...,kn u,...,kl
be probability that a µ-node following a ν−µ edge has excess degree (k1 , . . . , kν , . . . , kl )
then

qkµν1 ,...,kn u,...,kl ∝ (kν + 1)pµ (k1 , . . . , kν + 1, . . . , kl ) (64)

µ
Normalization such probability by hkiν , we have

1
qkµν1 ,...,kn u,...,kl = µ
µ (kν + 1)p (k1 , . . . , kν + 1, . . . , kl ) (65)
hkiν

Generating function of {qkµν1 ,...,kn u,...,kl } is

∞ ∂Gµ
∂xν (x)
X
Gµν (x) = qkµν1 ,...,kn u,...,kl xk11 . . . xkl l = ∂Gµ
(66)
ki =0,i=1,...l ∂xν (1)

18
Example 4.1 ([leicht2009percolation ]). [leicht2009percolation ] considers a concrete example for a
network of two modules α and β. Interconnection in each modules is an ER networks with average node
degree kα and kβ respectively. Each node from module α is connected randomly with any node in module
β with the same probabilty and has in average kαβ neighbors in module β. Similar, a node in module β
has average kβα neighbors in module α. Suppose the indpendence between connections. The multiple degree
distribution is given by
k2
kk1 kβ
α kα k1 α kαβ k2 α
p (k1 , k2 ) = e e
k1 ! k2 !
k1
k kα k2
k β α k
β

pβ (k1 , k2 ) = ekβ 1 ekβ 2

k1 ! k2 !
So the multidegree generating functions are
β
kα (x1 −1) kα (x2 −1)
Gα (x) = Gαβ (x) = Gαα (x) = |e {z } |e {z }
GF for Poisson intraconnection GF for Poisson interconnection
α
Gβ (x) = Gβα (x) = Gββ (x) = ekβ (x1 −1) ekβ (x2 −1)

In fact,
∞
X
k1 k2
Gα (x1 , x2 ) = pα
k1 ,k2 x1 x2
k1 ,k2 =0
k2
∞ k k1 kβ
X k1 α kαβ k2 α
= e kα
e xk1 xk2
k1 ! k2 ! 1 2
k1 ,k2 =0
k β k2
∞ k 1 ∞ k
X k1 α k1 X kαβ k2 α
= e kα
x1 e xk 2
k1 ! k2 ! 2
k1 =0 k2 =0
β
kα (x1 −1) kα (x2 −1)
=e e

In order to compute Gαα , we first take derivative of Gα with respect to x1

∂Gα
(x1 , x2 ) = kα Gα (x1 , x2 )
∂x1
So
α ∂Gα
hkiα = (1, 1) = kα Gα (1, 1) = kα
∂x1
It leads to
∂Gα
∂x1 (x1 , x2 )
Gα,α (x1 , x2 ) = ∂Gα
= Gα (x1 , x2 )
∂x1 (1, 1)

Similary, we obtain the formula for Gαβ , Gβ , Gββ , Gβα .

[leicht2009percolation ] also considers the interacting network whose degree distribution of each module
is power-law with an exponential cutoff but keeps the same Poisson interconnection setting. Multidegree degree
distribution of a β-node is
kβ

kα α
−kβ
kβτ e− κ
pα
k1 ,k2 = kβα e
Liτ (e− 1/κ)
k −τ xk is the τ th polylogarithm.
P
where Liτ (x) = k>0

19
 Multidegree generating function are
β
Gα (x1 , x2 ) = ekα (x1 −1)+kα (x2 −1)
α Liτ (x2 e−1/κ )
Gβ (x1 , x2 ) = ekβ (x1 −1)
Liτ (e−1/κ )

So

Gαα = Gαβ = Gα
α Liτ −1 (x2 e−1/κ )
Gββ (x) = ekβ (x1 −1)
x2 Liτ −1 (e−1/κ )
Gβα (x) = Gβ (x)

Example 4.2 ([shai2015critical ]). A network is divided into m equal-size modules. Both intra and inter
connectivity follows Poisson distribution. The connection inside any module is the same as an ER network
with average intra degree kintra . Each node has average km−1
inter
neighors in each other module. Such model is
a particular case of [leicht2009percolation ] with kα = kintra and kαβ = km−1 inter
for all module α 6= β. The
multidegree generating function is given by
Y kinter
(xj −1)
Gi (x1 , . . . , xm ) = ekintra (xi −1) e m−1 (67)
j6=i

Thus for all i, j we have (

∂Gi kintra Gi (x) if j = i
(x) = kinter
∂xj m−1 Gi (x) 6 i
if j =
Hence
Gij = Gi
If we take account into the fraction βi of node in each module then
Y j βj
Gi (x1 , . . . , xm ) = ekintra (xi −1) eki (xj −1) where kij = kinter (68)
1 − βj
j6=i

Example 4.3 ([dong2018resilience ]). Consider a network of m equal size modules. In each module, only
a fraction r of nodes has interlink with nodes from other modules. So one can divide nodes in a module
in two classes: inward and outward. Take a random node, it belongs to inward class with probability 1 − r
and outward class with probability r. For a node in the inward class, its has intra degree distribution p(k)
and has inter degree 0 with probability 1. For an outward node in module i, the number of intra links
has distribution p(k) with generating function Gii (xi ) and it can connect with a random node from module j
according to a degree distribution with generating function Gji (xj ). Assuming the independence of connection,
the multidegree generating function of an i-node is
Y j
Gi (x) = (1 − r)Gii (xi ) + rGii (xi ) Gi (xj ) (69)
j6=i

where Gji is the generating function for the j-degree of an i-node.

The excess multidegree generating function is given by
∂Gi
∂xj (x)
Gij (x) = ∂Gi
(70)
∂xj (1)

20
In the case that each module is an ER network with average intra-degree k then
Gii (xi ) = ek(xi −1)
If the interconnection degree distribution is also Poisson with average interlink K then for i 6= j, we have
Gji (xj ) = eK(xj −1)
So Y
Gi (x) = (1 − r)ek(xi −1) + rek(xi −1) eK(xj −1)
j6=i

[dong2018resilience ] also provides an explicit formula for scale free module and Poisson interlink.

4.1.2 Generating function for number of second neighbors [leicht2009percolation ]

Repeat the same argument for number of second neighbors in [newman2018networks ], the generating
function for the number of second number of neighbors of a µ-node via a ν-node is
Gµ (1, . . . , Gµν (1, . . . , Gνµ (x)xi =1,i6=µ , 1 . . . , 1) (71)

4.1.3 Generating function for size of small components [leicht2009percolation ]

Following an ν − µ edge and consider the distribution in sizes of the component found by the following
additional outgoing edges.

Figure 7: Topology for size of components reachable by following a random ν − µ edge

([leicht2009percolation ])

Size of component following a ν −µ edge is equal to the total size of component following the edge starting
fromt this µ-node and add by 1 to take account into this µ-node.
Let Hµν (x) be generating function for the size of components following an ν − µ-edge. Given that the
node following an ν − µ edge has excess degree (k1 , . . . , kl ), denote P (s1 , . . . , sl |k1 , . . . , kl ) be the number of
nodes in the module 1,2, . . . , l. belonging to the component following this edge Then
XX
Hµν (x) = qµν (k1 , . . . , kl )P (s1 , . . . , sµ − 1, sl |k1 , . . . , kl ) xs11 . . . xsl l
sj ki
| {z }
ρµν (s1 ,...,sl )
X X
= qµν (k1 , . . . , kl ) P (s1 , . . . , sµ − 1, sl |k1 , . . . , kl )xs11 . . . xsl l
ki sj
X X
= xµ qµν (k1 , . . . , kl ) P (s1 , . . . , sµ , sl |k1 , . . . , kl )xs11 . . . xsµµ . . . xsl l
ki sj

As in [newman2018networks ], using
X X Y
P (s1 , . . . , sl |k1 , . . . , kl ) = ρλµ (ui11 , . . . , uill ) (72)
i1 =1,k1 ,...,il =1,...kl uλ λ
1 +...uk =sλ
λ
λ

21
and decomposing sλ = u1λ + · · · + ukλλ , we obtain
i1 il
u u
X X Y
Hµν (x) = xµ qµν (k1 , . . . , kl ) ρλµ (ui11 , . . . , uil l )x1 1 . . . xl l
ki i1 =1,k1 ,...,il =1,...kl λ

Sum of size of all components following the edges leading to component λ from this µ-node given this node
has is kλ neighbors in module λ has generating function [Hλµ (x)]kλ . We obtain
X
Hµν (x) = xµ qµν (k1 , . . . , kl )(H1µ (x)k1 . . . (Hlµ (x)kl ) (73)
k1 ,...,kl

= xµ Gµν (H1µ (x), Hlµ (x)) (74)

Generating function for components size (s1 , . . . , sl ) containing a µ node is

Hµ (x) = xµ Gµ (H1µ (x)k1 . . . (Hlµ (x)kl ) (75)

Such kind of argument comes from power law property of generating function.
Example 4.4 ([leicht2009percolation ]). 1. 2 ER modules and Poisson interlink
β
Hα (x1 , x2 ) = Hαα (x1 , x2 ) = Hαβ (x1 , x2 ) = x1 Gα (Hαα (x), Hβα (x)) = x1 ekα (Hαα (x)−1) ekα (Hβα (x)−1)
α
Hβ (x1 , x2 ) = Hβα (x1 , x2 ) = Hββ (x1 , x2 ) = x2 Gβ (Hαβ (x), Hββ (x)) = x2 ekβ (Hαβ (x)−1) ekβ (Hββ (x)−1)

2. ER module + power law with cutoff degree module + Poisson interlink

β
Hα (x1 , x2 ) = Hαα (x1 , x2 ) = Hαβ (x1 , x2 ) = Hαα (x)Gα (Hαα (x), Hβα (x)) = x1 ekα (Hαα (x)−1) ekα (Hβα (x)−1)
α Liτ (Hββ (x)e−1/κ )
Hβα (x1 , x2 ) = Hβ (x1 , x2 ) = x2 Gβ (Hαβ (x), Hββ (x)) = x2 ekβ (x1 −1)
Liτ (e−1/κ )
α Liτ −1 (Hββ (x)e−1/κ )
Hββ (x1 , x2 ) = x2 Gββ (x1 , x2 ) = ekβ (Hαβ (x)−1
Liτ −1 (e−1/κ )
Example 4.5 ([shai2015critical ]).
Example 4.6 ([dong2018resilience ]).
Example 4.7 ([shai2014resilience ]).

4.1.4 Giant component [leicht2009percolation ]

Probability that a µ-node does not belong to the giant component is Hµ (1). So the fraction of µ-node
belonging to the giant component is

Sµ = 1 − Hµ (1) (76)

Denote uλµ = Hλµ (1) then

Sµ = 1 − Gµ (u1µ , . . . , ulµ ) (77)

where uλµ satisfies the equation

Hλµ (1) = Gλµ (u1µ , . . . , ulµ )
Example 4.8 ([leicht2009percolation ]). 1. 2 ER modules and Poisson interlink
β
Hα (x1 , x2 ) = Hαα (x1 , x2 ) = Hαβ (x1 , x2 ) = x1 Gα (Hαα (x), Hβα (x)) = x1 ekα (Hαα (x)−1) ekα (Hβα (x)−1)
α
Hβ (x1 , x2 ) = Hβα (x1 , x2 ) = Hββ (x1 , x2 ) = x2 Gβ (Hαβ (x), Hββ (x)) = x2 ekβ (Hαβ (x)−1) ekβ (Hββ (x)−1)

22
 2. ER module + power law with cutoff degree module + Poisson interlink
β
Hα (1, 1) = Hαα (1, 1) = Hαβ (1, 1) = Gα (Hαα (1), Hβα (1)) = ekα (Hαα (1)−1) ekα (Hβα (1)−1)
α Liτ (Hββ (1)e−1/κ )
Hβα (1, 1) = Hβ (1, 1) = Gβ (Hαβ (1), Hββ (1)) = ekβ (1−1)
Liτ (e−1/κ )
α Liτ −1 (Hββ (1)e−1/κ )
Hββ (1, 1) = Gββ (1, 1) = ekβ (Hαβ (1)−1 )
Liτ −1 (e−1/κ )
Solve numerically the above system equations to find Hα (1), Hαα (1), Hαβ (1), Hβ (1), Hβα (1), Hββ (1)
to find
Sα = 1 − Gα (Hαα (1), Hβα (1))Sβ = 1 − Gβ (Hαβ (1), Hββ (1))
Example 4.9 ([shai2015critical ]).
Example 4.10 ([dong2018resilience ]).
Example 4.11 ([shai2014resilience ]).
In order to study the phase for the existence of giant component, one uses the condition under which
the average size component diverges where the average number of ν -nodes in the component of a randomly
chosen µ-node is
∂
hsµ iν = Hµ (x)
∂xν x=1

4.2 Percolation and Interacting Network Resilience [shai2014resilience ]

In [shai2014resilience ], they use index i instead of µ and j insteads of ν. It is an extension of [callaway2000network
] for random degree attack and multidegree distribution. But it can be also extented to any degree attack
strategy as in [newman2018networks ].
Denote rki 1 ,k2 ,...,km be probability that a node in module i with multidegree (k1 , k2 , . . . , km ) is occupied.
Define the generating function for the degree and excess degree distributions of occupied node as
∞
X
Fi (x) = rki 1 ,k2 ,...,km pik1 ,k2 ,...,km xk11 xk22 . . . xkmm (78)
k1 ,...,km =0

and
∞
X
Fij (x) = rki 1 ,k2 ,...,km qkij1 ,k2 ,...,km xk11 xk22 . . . xkmm (79)
k1 ,...,km =0

We have
∂Fi
∂xj (x)
Fij (x) = i
(80)
hkij

Define the generating function for the distribution of the number of occupied nodes in the component
reachable by following a randomly chosen ij-edge to an i-node and then following its additional outgoing
links

Jij (x) = 1 − Fij (1) + xi Fij (J1i (x, J2i (x), . . . , Jmi (x)) (81)

The distribution of the number of nodes reachable from a randomly chosen i-node (rather than ij-edge) is
generated by

Ji (x) = 1 − Fi (1) + xi Fi (J1i (x, J2i (x), . . . , Jmi (x)) (82)

23
 Then, the average number of occupant j-nodes in the component of a randomly chosen i-node, is given
by

i ∂Ji
hsij = (1) (83)
∂xj

The threshold for divergence of the above quantity is used as percolation threshold.
Fraction of i-nodes belonging to the giant cluster is

Si = 1 − Ji (1) = Fi (1) − Fi (u1i , . . . , umi ) (84)

where uij = Jij (1) satisfies

uij = 1 − Fij (1) + Fij (u1i , . . . , umi ) (85)

We’ll try to apply this approach to find LCC of a network under degree attack strategies. The degree
attack in [shai2015critical ], [dong2018resilience ] are uniform random attack.

Example 4.12 ([shai2015critical ]). [shai2015critical ] studies uniformly random attack on intercon-
nected node, that is the occupied probability is
(
1 if kj = 0∀j 6= i , that is i has no interlink
rki 1 ,...,km =
q otherwise, that is i has at least one interlink

So  ki
e−kintra kintra −kij
Q P
k−i! j6=i e if j6=i kj = 0
rki 1 ,...,km pik1 ,...,km = j kj
qe−kintra (ki )kintra Q e−kij (ki ) if
P
kj > 0
k−i! j6=i kj ! j6=i

Hence
∞ ki ∞ ki
X X kintra Y j k X X kintra Y j k
Fi (x) = e−kintra xki i e−ki xj j + q e−kintra xki i e−ki xj j
P k − i! P k − i!
j6=i kj =0 ki =0 j6=i j6=i kj >0 ki =0 j6=i
∞ ki
X kintra Y j
= e−kintra xki i e−ki
k − i!
ki =0 j6=i
 
∞ ki ∞ ki
X X k Y j k X X k Y j k
+q e−kintra intra xki i e−ki xj j − e−kintra intra xki i e−ki xj j 
P k − i! P k − i!
j6=i kj ≥0 i
k =0 j6=i k =0
j6=i kj =0 i
j6=i
 
= ekintra (xi −1) e−kinter + q Gi (x) − ekintra (xi −1) e−kinter
= (1 − q)ekintra (xi −1) e−kinter + qGi (x)

where
kij (xj −1)
P
Gi (x) = ekintra (xi −1) e j6=i

The fraction of general node is occupied in each module is

p = Fi (1) = (1 − q)e−kinter + q

Then we obtain the relationship between these occupied probability

p − e−kinter
q=
1 − e−kinter

24
Furthermore
(
∂Fi kintra Fi (x) if j = i
(x) = kinter
∂xj m−1 Gi (x) 6 i
if j =

So we obtain (
Fi (x) if j = i
Fij (x) =
qGi (x) 6 i
if j =
The generating functions for the number of occupied node in a given component following an edge or a node
are satisfied the equations
(
1 − p + xi Fi (J1i (x), . . . , Jmi (x)) if j = i
Jij (x) =
1 − q + qxi Gi (J1i (x), . . . , Jmi (x)) if j 6= i
j
( P
(1 − q)(1 − e−kinter + xi ekintra (Jii (x)−1) e−kinter + qxi ekintra (Jii (x)−1) e j6=i ki (Jji (x)−1) if j = i
= P j
qxi ekintra (Jii (x)−1) e j6=i ki (Jji (x)−1) 6 i
if j =
Ji (x) = 1 − p + xi Fi (J1i (x), . . . , Jmi (x))
kij (Jji (x)−1)
P
= (1 − q)(1 − e−kinter + xi ekintra (Jii (x)−1) e−kinter + qxi ekintra (Jii (x)−1) e j6=i

Let uij = Jij (1) and ui = Ji (1) then we have ui = uii and uij is the solution to the system
j
( P
(1 − q)(1 − e−kinter + ekintra (uii −1) e−kinter + qekintra (uii −1) e j6=i ki (uji −1) if l = i
uil = P j
1 − q + qekintra (uii −1) e j6=i ki (uji −1) if l 6= i
1
In the case that βj = m−1 , we have uii = ui = u for all i and uij = v for all i 6= j such that
(
u = 1 − q − (1 − q)e−kinter + (1 − q)ekintra (u−1) e−kinter + qekintra (u−1) ekinter (v−1)
v = 1 − q + qekintra (u−1) ekinter (v−1)

Hence we obtain the self-consistence equation for u

h i h −kinter
i
(1−ekintra (u−1) )
u = 1 − (1 − q)e−kinter 1 − ekintra (u−1) − q 1 − e(kinter +kintra )(u−1) e−(1−q)e

Let S = 1 − u be the size of giant component then

h −kinter
i
(1−e−kintra S )
S = (1 − q)e−kinter 1 − e−kintra S + q 1 − e−(kinter +kintra )S e−(1−q)e
 

Example 4.13 ([dong2018resilience ]). [dong2018resilience ] focus on uniformly random attack on

any node, that is
rki 1 ,...,km = p
[dong2018resilience ] uses the result in [buldyrev2010catastrophic ] to find the generating function
for multidegree of the network after being attacked - P
recalculate degree for uniformly random attack. For
a network with degree generating function G0 (x) = k pk xk , after randomly removing a fraction 1 − p of
nodes, generating function for the remaining network becomes

Gp0 (x) = G0 (1 − p(1 − x))

Denote 1 − fii and 1 − fij be probability that a random chosen intra and inter-link belong to the giant
component then after a random removal, the probability that a node to connect to the giant component through
25
intra-links and inter-link are p(1 − fii ) and p(1 − fij ). Hence the respective probability that it belong to a
small cluster is 1 − p(1 − fii ) and 1 − p(1 − fij ). Hence fii and fij satisfy the self-consistence equation

fij = Gij [1 − p(1 − f1i ), . . . , 1 − p(1 − fmi )]

Such result matches with the approach in [shai2014resilience ].

In fact, from the uniformly random attack, we have

Fi (x) = pGi (x)

So
Fij (x) = pGij (x)
Denote uij = Jij (1) then it saftisfies the equation

uij = 1 − Fij (1) + Fij (u1i , . . . , umi ) = 1 − p + pGij (u1i , . . . , umi )

For uij = 1 − p(1 − fij ), the above equation becomes

1 − p(1 − fij ) = 1 − p + pGij (1 − p(1 − f1i ), . . . , 1 − p(1 − fmi ))

which is equivalent to
fij = Gij (1 − p(1 − f1i ), . . . , 1 − p(1 − fmi ))

First we can try to find an analytic formula for LCC under different degree attack on the models in these
papers and then for the inter-all model.

5 Projects for interacting network

5.1 Project: Apply intentional attack strategies of [shai2015critical ] on the
model in [dong2018resilience ]
.
In [dong2018resilience ], a modular is divided into 2 groups: non-interconnection and interconnection.
A node in the non-interconnection group has no connection with any node out of its modular but a node in
the non-interconnection group can be connected to a node in another modular by an interlink. The fraction
of nodes in each group are 1 − r and r respectively. For r = 1, we revisit the modular network structure in
[shai2015critical ].
The multidegree distribution of an i-node is
Y m
Y
pik1 ,...,km = (1 − r)pii (ki ) δ0 (kj ) + r pji (kj )
j6=i j=1

where (
1 if x = 0
δ0 (x) =
0 6 0
if x =

26
and pji (k) is probability that an i-node in the interconnection group has k neighbors in module j. Hence
X
Gi (x) = pik1 ,...,km xk11 xkmm
k1 ,...,km
 
X X m
Y
= (1 − r) pii (ki )xki i + r  pji (kj ) xk11 . . . xkmm
ki k1 ,...,km j=1
 
m
k
Y X j
= (1 − r)Gii (xi ) + r  pi (kj )xj j 
j=1 kj
m
Y
= (1 − r)Gii (xi ) + r Gji (xj )
j=1

Probability that randomly an interconnection node is occupied is q. So the generating function for multide-
gree of a random occupied node is
X X
Fi (x) = pik1 ,...,km xk11 . . . xkmm + q pik1 ,...,km xk11 . . . xkmm
P P
j6=i kj =0 j6=i kj >0
 
Y m
Y
= (1 − r) Gii (x) + r (1 − q)Gii (xi ) pji (0) + q Gji (x)
j6=i j=1

and then
∂Fi
∂xj (x)
Fij (x) = j
=
hki i
Using the approach in [shai2014resilience ] to find the condition for the existence of phase transition for
q and then for p. Thus find the size of giant component in the case of existence. The following example
provides a detail computing for the case of network with ER modulars and Poison interlink.
Example 5.1 (Intentional random attack of [shai2015critical ] on [dong2018resilience ] with ER mod-
ular, Poisson interconnection). The degree distributions are

kl
pii (l) = e−k
l!
and
Kl
pji (l) = e−K for j 6= i
l!
The multidegree generating function is
Y
Gi (x) = (1 − r)ek(xi −1) + rek(x1 −1) eK(xj −1)
j6=i

where k is the average intra degree of any node and K is the average inter degree of the interconnection
group in a modular to a given another modular, that is kintra = k and kinter = (m − 1)K.
It leads to
(
∂Gi kGi (x) if i = j
(x) =
Krek(x1 −1) j6=i eK(xj −1) if i 6= j
Q
∂xj

So
i ∂Gi j ∂Gi
hki i = (1) = k and hki i = (1) = rK
∂xi ∂xi
27
and then
(
Gi (x) if i = j
Gij (x) =
rek(x1 −1) j6=i eK(xj −1)
Q
6 j
if i =

Generating function for multidegree of a random occupied i - node is

 
Y j m
Y j
Fi (x) = (1 − r) Gii (xi ) + r (1 − q)Gii (xi ) pi (0) + q Gi (xj )
j6=i j=1
 
Y
= (1 − r) ek(xi −1) + r (1 − q)e−(m−1)K ek(xi −1) + qek(xi −1) eK(xj −1) 
j6=i

and
(
Fi (x) if j = i
Fij (x) =
qek(xi −1) l6=i eK(xl −1)
Q
6 i
if j =

The general occupation probability is

 
p = F (1) = 1 − r + r (1 − q)e−(m−1)K + q

Generating functions for size of small component are satisfied the self-consistence equations

Jij (x) = 1 − Fij (1) + xi Fij (J1i (x), . . . , J1m (x))

Ji (x) = 1 − Fi (1) + xi Fi (J1i (x), . . . , J1m (x))

which satisfies
Jii (x) = Ji (x)
Condition for the existence of giant cluster
The average number of j-node in a component of a random chosen occupied i-node is
m
∂Ji X ∂Fi ∂Jli
hsi ij = (x) = δi (j)Fi (1) + (1) (1)
∂xj x=1 ∂xl ∂xj
l=1
∂Jii X ∂Jli
= δi (j)Fi (1) + kFi (1) (1) + qK (1)
∂xi ∂xj
l6=i

28
 ∂Jij
Denote yij = ∂x1 (1), we have
X ∂Fi ∂Jl1
y11 = F1 (1) + (1)
(1)
∂xl∂x1
l
∂F1 X ∂Fi
= F1 (1) + (1)y11 + (1)yl1
∂x1 ∂xl
l6=1
X
= F1 (1) + kF1 (1)y11 + qrK yl1
l6=1
∂F1j X ∂F1j
y1j = F1j (1) + (1)y11 + (1)yl1 for j 6= 1
∂x1 ∂xl
l6=i
X
= q + qky11 + qK yl1
l6=1
X
yii = kFi (1)yii + qrK yli for i 6= 1
l6=i
X
yij = qkyii + qrK yli for i 6= 1, i 6= j
l6=i

Hence
m
X m
X X m
X X
yii = F1 (1) + k Fi (1)yii + qrK yli = F1 (1) + kF1 (1) yii + qrK yli
i=1 i=1 l6=i i=1 l6=i
X m
X X
yij = (m − 1)q + (m − 1)qk yii + (m − 1)qK yli
i6=j i=1 l6=i
P P
Solving the above system of two variable i yii and l6=i yli , we obtain
m
X F1 (1)[1 − (m − 1)Kq] + (m − 1)Kq 2 r
yii =
i=1
(1 − kF1 (1))(1 − (m − 1)Kq) − (m − 1)Krq 2

which diverges when

(1 − kF1 (1))(1 − (m − 1)Kq) − (m − 1)Krq 2 = 0
Using the same argument in [shai2015critical ], we have the critical occupation probability of interconnected
node, denoted by qc is given by
(
0 if m < m∗
qc = −b+√b2 −4ac (86)
2a if m > m∗

where a =, b =, c = the coefficients of the quadratic in the denominator.

The critical value qc = 0 if qc = 0 make the above quantity diverges, that is

kF1 (1) − 1 = 0

which is equivalent to
∗
k(1 − r + re−(m −1)K
)=0

From this equation the threshold value m∗ depends on k, K, r.

For r = 1, we recover the result in [shai2015critical ].
Find the size of giant cluster
29
 Suppose that q > qc then the giant cluster exists. Denote ui = Ji (1) and uij = Jij (1) which satisfy the
self-consistence equation
ui = 1 − Fi (1) + Fi (u1i , . . . , umi )
uij = 1 − Fij (1) + Fij (u1i , . . . , umi )
From the symmetry and Fi = Fii , we have uii = ui = u1 for all i and uij = u12 for all i 6= j. Let u1 = u
and u12 = v, we have
u = 1 − F1 (1) + F1 (u, v, . . . , v)
h  i h  i
= 1 − 1 − r + r (1 − q)e−(m−1)K + q + (1 − r)ek(u−1) + r (1 − q)e−(m−1)K ek(u−1) + qek(u−1) e(m−1)K(v−1)
v = 1 − F12 (1) + F12 (u, v, . . . , v)
= 1 − q + qek(u−1) e(m−1)K(v−1)
which can be solve numerically.
Let S be size of giant cluster then S = 1 − u.

5.2 Project: Random but nonuniform attack on interacting network on models

proposed in [shai2015critical ] and [dong2018resilience ]
Combining the idea of prof. Cassi and [shai2015critical ], we attack randomly on interconnected node.
The probability that an interconnected node is occupied depends on its multidegree or inter-multidegree.
Qm 1 i
P
 j=1 aij +1 if kj ≤ Kj ∀j, j6=i kj > 0

rki 1 ,...,km = 0 (87)
P
 if j6=i kj > 0, ∃j 6= i such that kj > Kj

1 otherwise
The cutoff degree Kji is given. Such kind of strategy can be modified to degree attack on interconnected
node (idea of Quang Nguyen in the past).
Other degree attack strategies: combining strategy attack on interconnected node and non-interconnected
node.

5.3 Project: attack on high interconnection node (highest degree attack) ??

idea of Quang Nguyen
As in [shai2015critical ], we only attack on the interconnected node but not uniform randomly. The
strategy is to actack on high interconnection node in the sense that total inter degree exceed a given threshold
P j j j
j6=i i > K or an inter degree to at least one module is high enough, that is ki > Ki for some j 6= i and
k
the threshold Kij is given ...
In this section, we study the attack strategies on the model setting in [leicht2009percolation ]

5.4 Attack the interconnected node with high total interconnection degree
5.5 Attack the interconnected node with high interconnection degree to at least
one module
5.6 Project: Inter-all model (Quang Nguyen) - an extension of [dong2018resilience
]
5.6.1 Model description
Consider an interacting network consisting of M module. Each module has a fraction r of nodes with
high inter-connectivity and a fraction 1 − r of nodes with low connectivity. A node of module i in high
30
- interconnectivity group and low -interconnectivity group has a link with a node from module j with
probability Pij1 and Pij2 or average number of neighbors Kij 1
in module j and Kij2
respectively. We also
denote pii (k) as intra - degree distribution of a node in module i. The multi-degree distribution of a node
in module i is hence given by
1 kj 2 kj
Y 1 (Kij ) Y 2 (Kij )
pik1 ,...,km = rpii (ki ) e−Kij + (1 − r)pii (ki ) e−Kij
kj ! kj !
j6=i j6=i

So the multi-degree distribution is given by

P 1 P 2
Kij (xj −1) Kij (xj −1)
Gi (x) = rGii (xi )e j6=i + (1 − r)Gii (xi )e j6=i

If the number of interlink from a random node a module to another modules not Poisson distributed but has
generating function Gil ih
j or Gj depending on the group of high or low connectivity that the node belongs to
then similar to [dong2018resilience ], under the assumption of independent connection, we have
Y Y
Gi (x) = rGii (xi )e Gil i
j (xj ) + (1 − r)Gi (xi ) Gih
j (xj )
j6=i j6=i

In such kind of set up, each pair of nodes is connected randomly with some probability depending on the
groups that they come from. A similar kind of set up for ER network instead of G(m, n). We can fix the
number of edges in each group of a module, not fix the number of interlink and intralink but the probability
of connecting any pair of node in the groups. In the case that we fix both number of node and edge, we face
to the same problem as the network G(m, n) whose node degree distribution is not available and we cannot
find the generating function.

5.6.2 Giant component

5.6.3 Resilience and robustness
In this section, we will apply the results Section 4.2 to obtain analytic formulas for robustness after being
degree attacked.

Random uniformly attack

Highest degree attack
Random attack depending on degree

6 Project: Recalculated degree approach to analyze robustness of

network under degree-based attacked strategies
6.1 Reading recalculate degree for random attack [buldyrev2010catastrophic ]
6.2 Project: recalculated degree for degree attack strategies - need to verify
Consider a random network with degree distribution pk . Select randomly a node A and follow an edge from
this node, then A reach to a node of excess degree k with probability qk . Such neighbor is occupied with
probability φk+1 . Hence probability that a node at the end of an edge is occupied with average probability
X
φ̄ = qk φk+1 (88)
k≥1

31
For a node of degree k, it is occupied with probability φk . Given that, exact j of its neighbors are occupied
with average probability

Ckj φ̄j (1 − φ̄)k−j (89)

Hence the average probability of an occupied node has new degree j after the percolation process is proportion
to
X
p̃j = pk φk Ckj q j (1 − q)k−j (90)
k≥j

Generating function for new degree of an occupied node is

X XX
g̃0 (z) ∝ p̃j z j = pk φk Ckj φ̄j (1 − φ̄)k−j z j
j≥0 j≥0 k≥j
XX
= pk φk Ckj φ̄j (1 − φ̄)k−j z j
k≥0 l≤k
X
= pk φk (B(z))k
k≥0

= f0 (B(z))

where B(z) = φ̄z + 1 − φ̄ and f0 (z) = k≥0 pk φk z k . Using normalization, we have

P

f0 (B(z))
g̃0 (z) =
f0 (1)
Denote q̃k be the probability that a node at the end of a edge from an occupied node is occupied and
has new excess degree k. Given that its excess degree before being attack is l, there are exactly k occupied
neighbor which happens with probability

Clk φ̄k (1 − φ̄)l−k (91)

Hence
X
q̃k = φl+1 ql Clk φ̄k (1 − φ̄)l−k (92)
l≥k

and we obtain the generating function of the excess degree distribution in the new network as following
X
g̃1 (z) = q̃k z k (93)
k≥0
1 XX
= φl+1 ql Clk φ̄k (1 − φ̄)l−k z k (94)
f0 (1)
k≥0 l≥k
1 XX
= φl+1 ql Clk φ̄k (1 − φ̄)l−k z k (95)
f0 (1)
l≥0 k≤l
1 XX
= φl+1 ql (B(z))l (96)
f0 (1)
l≥0 k≤l
1
= f1 (B(z)) (97)
f0 (1)
That is the generating functions for degree and excess degree distribution of the network after the first
attack. Which is useful to evaluate the size of giant cluster and recalculate the degree distibution after the
second attack and so on.
32
Example 6.1 (Poisson random network - random attack).

Example 6.2 (Poisson random network - degree attack).

Example 6.3 (Exponential network - random attack).
Example 6.4 (Exponential network - degree attack).

Example 6.5 (Scale free network - random attack).

Example 6.6 (Scale free network - degree attack).

33