Download Cours 2.

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts
no text concepts found
Transcript
Spread of epidemics on random graphs
Viet Chi TRAN - Université Lille 1 - France
December 11, 2015
1
Course 2 : Convergence of a random individual-based model to Volz’
equations
2
SIR on a configuration model graph
FConfiguration model (CM) : Bollobas (80), Molloy Reed (95),
Durett (07), van der Hofstad (in prep.)
F Description of an SIR epidemics spreading on a configuration
model graph :
I
Infinite system of denumberable equations, Ball and Neal
(2008)).
I
5 ODEs, Volz (2008), Miller (2011).
I
Recently: Barbour Reinert (2014), Janson Luczak Winridge
(2014)
F Individuals are separated into 3 classes :
I
Susceptibles St
Infectious It
I
Removed Rt
I
3
Stochastic model for a finite graph with N vertices
F Only the edges between the I and R individuals are observed.
The degree of each individual is known.
F To each I individual is associated an exponential random clock with
rate α to determine its removal.
F To each open edge (directed to S), we associate a random
exponential clock with rate β.
F When it rings, the edge of an S is chosen at random. We
determine whether its remaining edges are linked with S, I or R-type
individuals.
(a)
(b)
(c)
I
S
I
I
I
I
R
I
R
R
I
I
I
S
I
4
Edge-based quantities
F The idea of Volz is to use network-centric quantities (such as the
number of edges from I to S) rather than node-centric quantities.
F St , It , Rt , St , It , Rt , di , di (St )...
µ finite measure
on N and f bounded or > 0 function:
P
hµ, f i = k ∈N f (k )µ(k ).
5
Edge-based quantities
F The idea of Volz is to use network-centric quantities (such as the
number of edges from I to S) rather than node-centric quantities.
F St , It , Rt , St , It , Rt , di , di (St )...
µ finite measure
on N and f bounded or > 0 function:
P
hµ, f i = k ∈N f (k )µ(k ).
F We introduce the following measures:
X
X
µSt (dk ) =
δdu (dk )
µSI
δdu (St ) (dk )
t (dk ) =
u∈St
µSR
(dk )
t
=
X
u∈It
δdu (St ) (dk )
u∈Rt
5
Edge-based quantities
F The idea of Volz is to use network-centric quantities (such as the
number of edges from I to S) rather than node-centric quantities.
F St , It , Rt , St , It , Rt , di , di (St )...
µ finite measure
on N and f bounded or > 0 function:
P
hµ, f i = k ∈N f (k )µ(k ).
F We introduce the following measures:
X
X
µSt (dk ) =
δdu (dk )
µSI
δdu (St ) (dk )
t (dk ) =
u∈St
µSR
(dk )
t
=
X
u∈It
δdu (St ) (dk )
u∈Rt
This sums up the evolution of the epidemic (but does not allow the
reconstruction of the complicated graph on which the illness
propagates).
X
It = Card(It ) = hµSI
NtSI = hµSI
du (St )
t , 1i,
t , ki =
u∈It
5
Dynamics
F Global force of infection: βNtSI
.
−
F Choice of a given susceptible of degree k : k /NtS− .
So that the rate of infection of a given susceptible of degree k is:
βkptI− .
F The probability that its k − 1 remaining edges are linked to I or R
is:
` m j
p(j, `, m|k − 1, t) =
NtSI −1
−
NtSR
−
k −1
NtS −1
NtSS
−
1j+`+m=k −1 1j<NtSI 1`≤NtSR
−
−
−
6
Dynamics
F Global force of infection: βNtSI
.
−
F Choice of a given susceptible of degree k : k /NtS− .
So that the rate of infection of a given susceptible of degree k is:
βkptI− .
F The probability that its k − 1 remaining edges are linked to I or R
is:
` m j
p(j, `, m|k − 1, t) =
NtSR
NtSI −1
−
−
k −1
NtS −1
NtSS
−
1j+`+m=k −1 1j<NtSI 1`≤NtSR
−
−
−
SR
F To modify the degree distributions µSI
t− (idem for µt− ):
We draw a sequence e = (eu )u∈It− of integers.
• eu is the number of edges to the infectious individual u at t− .
• not all sequences are admissible.
The probability of drawing the sequence e is
Q
du
ρU (e | j, µSI
T− ) =
u∈It−
NtSI −
j+1
eu
1{P eu =j+1, e is admissible} .
6
Renormalization
F We are interested in increasing the number of vertices N without
rescaling the degree distribution.
µN,S , µN,SI , µN,SR .
F We now consider µ(N),S , µ(N),SI and µ(N),SR where for ex:
(N),S
µt
(dk ) =
(N),SI
(idem for µ0
1 N,S
µ (dk )
N t
with
(N),S
lim µ0
N→+∞
(N),SR
with N̄0SI > ε and µ0
= µ̄S0 in MF (N)
with N̄0SR > ε)
7
Renormalization
F We are interested in increasing the number of vertices N without
rescaling the degree distribution.
µN,S , µN,SI , µN,SR .
F We now consider µ(N),S , µ(N),SI and µ(N),SR where for ex:
(N),S
µt
(dk ) =
(N),SI
(idem for µ0
1 N,S
µ (dk )
N t
with
(N),S
lim µ0
N→+∞
(N),SR
with N̄0SI > ε and µ0
= µ̄S0 in MF (N)
with N̄0SR > ε)
F 3 SDE:
(N),SI
(N),SI
(N),SI,f
(N),SI,f
hµt
, f i = hµ0
, f i + At
+ Mt
,
where M (N),SI,f is a square integrable martingale started from 0 and
with previsible quadratic variation in 1/N.
7
(N),IS,f
F At
(N),IS,f
At
:
Z
=−
t
(N),SI
αhµs
, f ids
0
+
Z tX
(N),I (N),S
µs
(k )
0 k ∈N
×
X
e∈U
X
βkps
ρU (e|j +
psN (j, `, m|k − 1, t)
j+`+1≤k
1, µn,SI
)
s
f (m) +
X
f (du − eu ) − f (du ) ds,
u∈IsN
8
(N),IS,f
F At
:
t
Z
(N),IS,f
=−
At
(N),SI
αhµs
, f ids
0
+
Z tX
(N),I (N),S
µs
(k )
0 k ∈N
X
×
X
βkps
psN (j, `, m|k − 1, t)
j+`+1≤k
ρU (e|j +
1, µn,SI
)
s
e∈U
f (m) +
X
f (du − eu ) − f (du ) ds,
u∈IsN
Th: Under appropriate moment conditions,
(N),S
(N),SI
(N),SR
(µt
, µt
, µt
SR
(µ̄St , µ̄SI
,
µ̄
)
t∈R+
t
t
hµ̄SI
t ,fi
=
hµ̄SI
0 ,fi
)t∈R+ converge to a deterministic limit
Z
t
−
αhµ̄SI
s , f ids
0
Z
+
t
X
0 k ∈N∗
X
βk p̄sI
j+`+m=k −1
× f (m) + (j + 1)
X
k 0 ∈N∗
j, `, m
(p̄sI )j (p̄sR )` (p̄sS )m
k −1
0 k 0 µ̄SI
s (k )
f (k 0 − 1) − f (k 0 )
µ̄Ss (k )ds
SI
hµ̄s , k i
8
Deterministic limit
F Limit equations:
µ̄St (k ) = µ̄S0 (k )θtk ,
θt = e−β
Rt
0
p̄sI ds
hµ̄SI
t , f i = ...
Z t
hµ̄SR
,
f
i
=
αhµ̄SI
t
s , f ids
0
+
Z tX
0 k ∈N
βk p̄sI (k − 1)p̄sR
X
k 0 ∈N
0
k 0 µSR
s (k ) S
µ̄s (k )ds
f (k 0 − 1) − f (k 0 )
SR
N̄s
F This allows us to recover Volz’equations:
• Choosing f ≡ 1 gives S̄t , Īt ,
• Choosing f (k ) = k gives N̄ S , N̄ SI , N̄ SR ,
from which we can deduce p̄I = N̄ SI /N̄ S ...
9
Volz’equations P
Prop: let g(z) =
k ∈N
µ̄S0 (k )z k be the generating function of µ̄S0 .
Z t
θt = exp − β
psI ds
0
S̄t =g(θt ),
Īt = Ī0 +
Z t
β p̄sI θs g 0 (θs ) − αĪs ds
0
N̄ SI
p̄tI = t S = p̄0I +
N̄t
p̄tS =
N̄tSS
N̄tS
Z t
g 00 (θs )
I
I
I
−
β
p̄
(1
−
p̄
)
−
α
p̄
ds.
s
s
s
g 0 (θs )
0
Z t
g 00 (θs ) = p̄0S +
β p̄sI p̄sS 1 − θs 0
ds.
g (θs )
0
β p̄sI p̄sS θs
2
Recall the limit for mixing models:
d S̄t
= − β S̄t Īt ,
dt
d Īt
= β S̄t Īt − αĪt .
dt
10
Volz’equations P
Prop: let g(z) =
k ∈N
µ̄S0 (k )z k be the generating function of µ̄S0 .
Z t
θt = exp − β
psI ds
0
S̄t =g(θt ),
Īt = Ī0 +
Z t
β × N̄sSI − αĪs ds
0
N̄ SI
p̄tI = t S = p̄0I +
N̄t
p̄tS =
N̄tSS
N̄tS
Z t
g 00 (θs )
I
I
I
−
β
p̄
(1
−
p̄
)
−
α
p̄
ds.
s
s
s
g 0 (θs )
0
Z t
g 00 (θs ) = p̄0S +
β p̄sI p̄sS 1 − θs 0
ds.
g (θs )
0
β p̄sI p̄sS θs
2
Recall the limit for mixing models:
d S̄t
= − β S̄t Īt ,
dt
d Īt
= β S̄t Īt − αĪt .
dt
Here:
d S̄t
=g 0 (θt )θ̇t = −βg 0 (θt )θt p̄tI = −β N̄tS p̄tI = −β N̄tSI .
dt
10
Sketch of the proof
(N),SI
(N),S
, 1 + k 5 i < +∞,
Assumption: supN∈N∗ hµ0
, 1 + k 5 i + hµ0
F Tightness: topology on MF (N). Roelly’s criterion.
Aldous-Rebolledo criterion.
|>ε ≤ε
− A(N),SI,f
P |Aτ(N),SI,f
σN
N
P |hM (N),SI,f iτN − hM (N),SI,f iσN | > ε ≤ ε.
F Convergence of the generators.
• The identification of the limit is OK on [0, T ] IF T < τεN where
(N),SI
τεN = inf{t ≥ 0, Nt
< ε}.
F Uniqueness:
• Gronwall’s lemma gives that solutions of the limiting equation
have same mass and same moments of order 1 and 2.
• Uniqueness of the generating function of µ̄IS which solves a
transport equation.
11
600
400
0
200
Sizes of classes
800
1000
Degree distribution of the “initial condition”
0
1
2
3
4
5
6
Time
Prop: For ε > 0, when N → +∞, the degree distribution when after
[εN] infections converges to:
1 X
pk (1 − z ε )k δk
1−ε
k ≥0
where z ε is the solution of 1 − ε = f (1 − z), f being the generating
function of the original degree distribution.
12