Métodos computacionales para el estudio de modelos epidemiológicos con incertidumbre - Juan Carlos Cortés

Computational Methods for Random
Epidemiological Models
Métodos Computacionales para el Estudio de
Modelos Epidemiológicos con Incertidumbre
Conferencias de Investigación para Posgrado 2016
Universidad Complutense de Madrid
24 junio de 2016
Prof. Dr. Juan Carlos Cortés
Instituto Universitario de Matemática Multidisciplinar
Universitat Politècnica de València
Métodos Computacionales Estocásticos en Epidemiolog´ıa J.C. Cortés 1

Part I
Ingredients

A naive (but maybe useful) comparison:
Deterministic Random
numbers: a = 3 r.v.’s: A ∼ N(µ = 3;σ2 > 0)
functions: x(t) = 3t s.p.’s: X(t) = At, A ∼ N(µ = 3;σ2 > 0)
There are s.p.’s which are not deﬁned by algebraic formulas as
Wiener process or Brownian motion
{W (t) : t ≥ 0} ≡ {B(t) : t ≥ 0} is called the (standard) Wiener process or Brownian
motion if it satisﬁes the following conditions:
1 It starts at zero w.p. 1: P[{ω ∈ Ω : W (0)(ω) = 0}] = P[W (0) = 0}] = 1.
2 It has stationary increments:
W (t)−W (s)
d
= W (t +h)−W (s +h), ∀h : s,t,s +h,t +h ∈ [0,+∞[.
3 It has independent increments:
W (t2)−W (t1),...,W (tn)−W (tn−1) are independent r.v.’s
∀{ti }n
i=1 : 0 ≤ t1 < t2 < ··· < tn−1 < tn < +∞, n ≥ 1.
4 It is Gaussian with mean zero and variance t: W (t) ∼ N(0;t), ∀t ≥ 0.

Graphical representation of a s.p.

Since a s.p. X(t) = {X(t) : t ∈ T } can be considered as a collection of random vectors
(Xt1 ,...,Xtn ), t1,...,tn ∈ T , n ≥ 1, we can extend the concept of expectation and
covariance for random vectors to s.p.’s and consider these quantities as functions of
t ∈ T :
One-dimensional probabilistic description of a s.p.
Expectation, variance and 1-p.d.f. of a s.p.
Expectation: µX (t) = E[X(t)], t ∈ T .
Variance: σ2
X (t) = V[X(t)] = E[(X(t))2]−(E[X(t)])2
, t ∈ T .
1-p.d.f.: It is the p.d.f. of the r.v. X(t) for every t. It is denoted by f1(x,t).
Two-dimensional probabilistic description of a s.p.
Covariance and 2-p.d.f. of a s.p.
Covariance: CX (t1,t2) = C[Xt1 ,Xt2 ] = E[(X(t1)− µX (t1))(X(t2)− µX (t2))],
t1,t2 ∈ T .
2-p.d.f.: It is the joint p.d.f. of the r.v.’s X(t1) and X(t2) for every t1 and t2. It
is denoted by f2(x1,t1;x2,t2).

Part II
Linear Models

Motivating the linear case: The malthusian population model with migration

p(t )
p(t +Δt )

t
t +Δt

p(t +∆t)−p(t) =
births
bp(t)∆t −
deaths
dp(t)∆t +
immigrants
i∆t −
emigrants
e∆t,
p(t +∆t)−p(t) = kp(t)∆t +m∆t, k = b −d, m = i −e ∈ R,
p(t +∆t)−p(t)
∆t
= kp(t)+m ⇒
˙p(t) = kp(t)+m, t > 0,
p(0) = p0,
Malthusian population model considering migration
˙p(t) = kp(t)+m, t > 0,
p(0) = p0,
, k = b −d, m = i −e.

How can uncertainty be introduced?
There are two main approaches:
Unknown uncertainty: Wiener process or Brownian motion. It requires the
so-called Itô-calculus.
Known uncertainty: It requires the so-called Lp(Ω)-calculus.
Itô-Stochastic Differential Equations (SDE’s)
Assuming, for instance, that the birth-rate coefficient is affected by a Gaussian
perturbation (unknown uncertainty):
˙p(t) = kp(t)+m, t > 0,
p(0) = p0,
, k ⇒ k +λ W (t)
white noise
, k ∈ R, λ > 0,
dp(t)
dt
= (k +λW (t))p(t)+m
dp(t) = (kp(t)+m)dt +λp(t)W (t)dt
dW (t)
dp(t) = (kp(t)+m)dt +λp(t)dW (t)
p(t) = p0 +
t
0
(kp(s)+m)dt +
t
0
λp(s)dW (s)
Itô-type integral
Itô Lemma
−−−−−−−−→ p(t)

Random Differential Equations (RDE’s)
Known uncertainty:
k is positive: k ∼ Exp(λ); k ∼ Be(α;β).
k is negative: k ∼ Un(−2,−0.5); k ∼ N(µ;σ) truncated at (−2,−0.5).
Malthusian population model considering migration
˙p(t) = kp(t)+m, t > 0,
p(0) = p0,
, k = b −d, m = i −e.
In practice the birth, death, immigration, emigration rates and the initial population
are fixed after sampling and measurements, hence it is more realistic to consider that:
k,m,p0 are r.v.’s, defined in a common probability space, (Ω,F,P)
rather than deterministic constants
⇓
This motivates to consider the above model from a stochastic standpoint. As a
consequence, its solution is a stochastic process (s.p.) rather than a classical function.
⇓
The main goals include to compute:
The solution s.p.: p(t) = p(t;ω), ω ∈ Ω.
The mean function: E[p(t)].
The variance function: V[p(t)].

To deal with RDE’s, Lp(Ω)-calculus has demonstrate to be a powerful tool.
p = 2 ⇒ mean square (m.s.) calculus
L2(Ω) = {X : Ω → R, 2-r.v.}
X 2 = E X2 1/2
< +∞
⇒ (L2(Ω), · 2) Banach space
(Ω,F,P) probability space
X : Ω → R is a (continuous absolutely) real random variable (r.v.)
F is a distribution function (d.f.); f is a probability density function (p.d.f.) of X
X 2-r.v. ⇔ E X2
=
Ω
x2
dF(ω) =
R
x2
f (x)dx < +∞
X 2-r.v. ⇒ V[X] = E X2 −(E[X])2
< +∞
Examples

mean square (m.s.) convergence of {Xn : n ≥ 0} ∈ L2(Ω)
Xn
m.s.
−−−→
n→∞
X ⇔ ( Xn −X 2)2
= E (Xn −X)2
−−−→
n→∞
0
Some reasons to select mean square convergence
Zn
m.s.
−−−→
n→∞
Z ⇒
E[Zn] −−−→
n→∞
E[Z],
V[Zn] −−−→
n→∞
V[Z].
⇓
XN (t) =
N
∑
n=0
Xntn
⇓
t ∈ T ﬁxed, ZN = XN (t) ⇒
E[XN (t)] −−−→
N→∞
E[X(t)]
V[XN (t)] −−−→
N→∞
V[X(t)]

However, it would be more desirable to determine the ﬁrst probability density
function (1-p.d.f.), f1(p,t), associated to the solution s.p. p(t) since from it one can
compute, as merely particular cases, the mean and variance functions:
µp(t) = E[p(t)] =
∞
−∞
p f1(p,t)dp,
σ2
p (t) = V[p(t)] =
∞
−∞
p2
f1(p,t)dp −(µp(t))2
.
But in addition, from it one can also compute higher statistical moments:
E[(p(t))k
] =
∞
−∞
pk
f1(p,t)dp, k = 0,1,2,...,
and signiﬁcant information such as the probability of the solution lies within a set of
interest
P[a ≤ p(t) ≤ b] =
b
a
f1(p,t)dp.
This improves the computation of rough bounds like
P[|p(t)− µp(t)| ≥ λ] ≤
(σp(t))2
λ2
,
usually used in practice.

The general random linear diﬀerential equation
Motivated by the previous presentation, in the following we focus on determining the
1-p.d.f., fZ (z,t), of the solution s.p. Z(t) to the general linear random initial value
problem (i.v.p.):
˙Z(t) = AZ(t)+B, t > t0,
Z(t0) = Z0,
where the data Z0, B and A are assumed to be absolutely continuous random
variables (r.v.’s) deﬁned on a common probability space (Ω,F,P), whose domains are
assumed to be:
DZ0
= { z0 = Z0(ω),ω ∈ Ω : z0,1 ≤ z0 ≤ z0,2},
DB = { b = B(ω),ω ∈ Ω : b1 ≤ b ≤ b2},
DA = { a = A(ω),ω ∈ Ω : a1 ≤ a ≤ a2}.
As we shall see later, the unifying element to conduct our study is the Random
Variable Transformation (R.V.T.) method.

For the sake of clarity in the presentation we will distinguish the following cases:
TYPE I.V.P. CASE
H
˙Z(t) = AZ(t)
Z(t0) = Z0
(I)
I.1 Z0 is a random variable
I.2 A is a random variable
I.3 (Z0,A) is a random vector
NH
˙Z(t) = B
Z(t0) = Z0
(II)
II.1 Z0 is a random variable
II.2 B is a random variable
II.3 (Z0,B) is a random vector
˙Z(t) = AZ(t)+B
Z(t0) = Z0
(III)
III.1 Z0 is a random variable
III.2 B is a random variable
III.3 A is a random variable
III.4 (Z0,B) is a random vector
III.5 (Z0,A) is a random vector
III.6 (B,A) is a random vector
III.7 (Z0,B,A) is a random vector
Z(t) = eA(t−t0)
Z0 +
B
A
eA(t−t0)
−1 , t ≥ t0.
Remarks:
I.V.P. (I): P[{ω ∈ Ω : B(ω) = 0}] = 1; I.V.P. (II): P[{ω ∈ Ω : A(ω) = 0}] = 1.
Hereinafter, deterministic parameters will be written by lower case letters and r.v.’s by capital letters.
Notation for the p.d.f.’s: fZ0
(z0); fZ0,A(z0,a); fZ0,B,A(z0,b,a), etc.
Standard and non-standard p.d.f.’s including copulas can be considered.

Preliminaries on Random Variable Transformation (R.V.T.) method
R.V.T. method: simple scalar version
H : Let X be a continuous r.v. with p.d.f. fX (x) with support S (X) and Y = r(X)
being r a bijective mapping.
T : The p.d.f. of Y , gY (y), is given by:
gY (y) = fX (x = s(y))
ds(y)
dy
, y ∈ S (r(X)).
R.V.T. technique: general scalar version
H : Let X be a r.v. with p.d.f. fX (x) and codomain or support DX = {x : fX (x) > 0}.
Let Y = r(X) be a new r.v. generated by the map r : R −→ R which is assumed to be
continuously diﬀerentiable on DX and such that r (x) = 0 except at a ﬁnite number of
points. Let us suppose that for each y ∈ R, there exist m(y) ≥ 1 points:
x1(y),x2(y),...,xm(y)(y) ∈ DX such that
r(xk (y)) = y, r (xk (y)) = 0, k = 1,2,...,m(y).
T : Then
fY (y) =



m(y)
∑
i=1
fX (xk (y)) r (xk (y))
−1
if m(y) > 0,
0 if m(y) = 0.

Next, we shall present some particular cases of R.V.T. method that will be useful later.
Case I.1: Z(t) = Z0ea(t−t0)
, t ≥ t0.
R.V.T. technique: linear transformation
H : Let X be a continuous r.v. with domain: DX = {x : x1 ≤ x ≤ x2} and p.d.f. fX (x).
T : Then, the p.d.f. fY (y) of the linear transformation Y = αX +β, α = 0 is given
by:
fY (y) =
1
|α|
fX
y −β
α
, where
y1 = αx1 +β ≤ y ≤ αx2 +β = y2 if α > 0,
y1 = αx2 +β ≤ y ≤ αx1 +β = y2 if α < 0.
If α = 0, then Y = β with probability 1 (w.p. 1) and
fY (y) = δ(y −β), −∞ < y < ∞,
where δ(·) denotes the Dirac delta distribution.

Case I.2: Z(t) = z0eA(t−t0)
, t ≥ t0.
R.V.T. technique: exponential transformation
H : Let X be a continuous r.v. with domain: DX = {x : x1 ≤ x ≤ x2} and p.d.f. fX (x).
T : Then the p.d.f. fY (y) of the exponential transformation Y = αeβX +γ, with
αβ = 0 is given by:
fY (y) =
1
|β(y −γ)|
fX
1
β
ln
y −γ
α
,
where
y1 = αeβx1 +γ ≤ y ≤ αeβx2 +γ = y2 if αβ > 0,
y1 = αeβx2 +γ ≤ y ≤ αeβx1 +γ = y2 if αβ < 0.
If α = 0 or β = 0, then Y = α +γ w.p. 1 and
fY (y) = δ(y −(α +γ)), −∞ < y < ∞.

Case I.3: Z(t) = Z0eA(t−t0)
, t ≥ t0.
R.V.T. technique: multi-dimensional version
H : Let X = (X1,...,Xn) be a random vector of dimension n with joint p.d.f. fX(x).
Let r : Rn −→ Rn be a one-to-one deterministic map which is assumed to be continuous
with respect to each one of its arguments, and with continuous partial derivatives.
T : Then, the joint p.d.f. fY(y) of the random vector Y = r(X) is given by
fY(y) = fX (s(y))|Jn|,
where s(y) is the inverse transformation of r(x): x = r−1(y) = s(y) and Jn is the
jacobian of the transformation, i.e.,
Jn = det
∂x
∂y
= det




∂x1
∂y1
··· ∂xn
∂y1
.
..
...
.
..
∂x1
∂yn
··· ∂xn
∂yn



,
which is assumed to be diﬀerent from zero.

Case II.3: Z(t) = Z0 +B(t −t0), t ≥ t0.
R.V.T. technique: sum of two r.v.’s
H : Let (X1,X2) be a continuous random vector with joint p.d.f. fX1,X2
(x1,x2) and
respective domains: DX1
= {x1 : x1,1 ≤ x1 ≤ x1,2} and DX2
= {x2 : x2,1 ≤ x2 ≤ x2,2}.
T : Then the p.d.f. fY1
(y1) of their sum Y1 = X1 +X2 is given by:
fY1
(y1) =
x1,2
x1,1
fX1,X2
(x1,y1 −x1)dx1, y1,1 = x1,1 +x2,1 ≤ y1 ≤ x1,2 +x2,2 = y1,2,
or, equivalently by
fY1
(y1) =
x2,2
x2,1
fX1,X2
(y1 −x2,x2)dx2, y1,1 = x1,1 +x2,1 ≤ y1 ≤ x1,2 +x2,2 = y1,2.
If X1 and X2 are independent r.v.’s, since fX1,X2
(x1,x2) = fX1
(x1)fX2
(x2), being fXi
(xi )
the p.d.f. of Xi , i = 1,2, the p.d.f. of the sum of two independent r.v.’s is just the
convolution of their respective p.d.f.’s:
fY1
(y1) =
x1,2
x1,1
fX1
(x1)fX2
(y1 −x1)dx1, or fY1
(y1) =
x2,2
x2,1
fX1
(y1 −x2)fX2
(x2)dx2.

Case I.3: Z(t) = Z0eA(t−t0)
, t ≥ t0.
R.V.T. technique: product of two r.v.’s
H : Let (X1,X2) be a continuous random vector with joint p.d.f. fX1,X2
(x1,x2) with
respective domains: DX1
= {x1 = 0 : x1,1 ≤ x1 ≤ x1,2} and DX2
= {x2 : x2,1 ≤ x2 ≤ x2,2}.
T : Then the p.d.f. fY1
(y1) of their product Y1 = X1X2 is given by:
fY1
(y1) =
x1,2
x1,1
fX1,X2
x1,
y1
x1
1
|x1|
dx1.
Equivalently, if DX1
= {x1 : x1,1 ≤ x1 ≤ x1,2} and DX2
= {x2 = 0 : x2,1 ≤ x2 ≤ x2,2} then
fY1
(y1) =
x2,2
x2,1
fX1,X2
y1
x2
,x2
1
|x2|
dx2. ( )
If X1 and X2 are independent r.v.’s with p.d.f.’s fX1
(x1) and fX2
(x2), respectively, then
previous formulas write:
fY1
(y1) =
x1,2
x1,1
fX1
(x1)fX2
y1
x1
1
|x1|
dx1, or fY1
(y1) =
x2,2
x2,1
fX1
y1
x2
fX2
(x2)
1
|x2|
dx2,
respectively.

Computing the 1-p.d.f. of the solution s.p. of the general linear random differential
equation: Some study-cases
Case I.1: Z0 is a r.v.
In this case the solution s.p. has the following expression:
Z(t) = Z0ea(t−t0)
, t ≥ t0.
Next, we first fix t : t ≥ t0 and denote Z = Z(t). Then, we apply R.V.T. method
(linear transformation: Y = αX +β, α = 0) to:
α = ea(t−t0)
> 0, β = 0, X = Z0, Y = Z,
and, taking into account that fY (y) = 1
|α| fX
y−β
α and the domain of r.v. Z0, one
gets:
f1(z,t) = e−a(t−t0)
fZ0
z e−a(t−t0)
, z1 ≤ z ≤ z2, t ≥ t0,
where
z1 = z0,1ea(t−t0)
, z2 = z0,2ea(t−t0)
.

Example Case I.1: Z0 ∼ N µ;σ2 , µ ∈ R and σ2 > 0
f1(z,t) =
1
√
2πσ2
e
− a(t−t0)+ 1
2σ2 z e−a(t−t0)−µ
2
, −∞ < z < ∞, t ≥ t0.
Example : Z0 ∼ N(0;1), t0 = 0, a = −1.
Z(t) = Z0e−t
, t ≥ t0.

Case I.2: A is a r.v.
Z(t) = z0eA(t−t0)
, t ≥ t0.
Next, we ﬁrst ﬁx t : t > t0 and denote Z = Z(t). Then we apply R.V.T. method
(exponential transformation: Y = αeβX +γ, αβ = 0) to:
α = z0 = 0, β = t −t0 = 0, X = A, γ = 0, Y = Z.
Then, taking into account that fY (y) = 1
|β(y−γ)| fX
1
β ln y−γ
α and
z/z0 = ea(t−t0) > 0 and the domain of r.v. A, one gets:
f1(z,t) =
1
(t −t0)|z|
fA
1
t −t0
ln
z
z0
, z1 ≤ z ≤ z2, t > t0,
where
z1 = z0ea1(t−t0), z2 = z0ea2(t−t0), if z0 > 0,
z1 = z0ea2(t−t0), z2 = z0ea1(t−t0), if z0 < 0.
For t = t0: Z(t) = Z(t0) = z0, which is deterministic. Then its 1-p.d.f. can be written
by the Dirac delta function as follows:
f1(z,t0) = δ(z −z0), −∞ < z < ∞.

Example Case I.2: A ∼ Be(α;β), α,β > 0 and z0 > 0
f1(z,t) = 1
B(α,β)|z|
1
t−t0
α
ln z
z0
α−1
1− 1
t−t0
ln z
z0
β−1
,
z0 ≤ z ≤ z0et−t0 , t > t0.
f1(z,t0) = δ(z −z0), −∞ < z < ∞.
Remark: Since z = z0ea(t−t0) and 0 ≤ a ≤ 1, it is guaranteed that 0 ≤ 1
t−t0
ln z
z0
≤ 1.
Example : A ∼ Be(2;3), t0 = 0, z0 = 1.

Case I.3: (Z0,A) is a random vector
In this case that the solution s.p. has the following expression:
Z(t) = Z1(t)Z2(t), where
Z1(t) = Z0,
Z2(t) = eA(t−t0).
To compute the p.d.f. of Z = Z(t), t : t > t0 ﬁx, ﬁrst we will determine the joint p.d.f.
of Z1 = Z1(t) and Z2 = Z2(t) by R.V.T. method (two–dimensional version) to:
X1 = Z0, X2 = A, r1(z0,a) = z0, r2(z0,a) = ea(t−t0),
Y1 = Z1, Y2 = Z2, s1(z1,z2) = z1, s2(z1,z2) = ln(z2)
t−t0
.
Hence, the Jacobian is given by
J2 =
∂s1(z1,z2)
∂z1
∂s2(z1,z2)
∂z2
=
1
z2(t −t0)
> 0,
therefore
fZ1,Z2
(z1,z2) =
1
z2(t −t0)
fZ0,A z1,
ln(z2)
t −t0
, z0,1 ≤ z1 ≤ z0,2, ea1(t−t0)
≤ z2 ≤ ea2(t−t0)
.

Now, we apply R.V.T. method (product of two r.v.’s: Y1 = X1 X2) to obtain the p.d.f.
of Z = Z1 Z2. As Z2 = eA(t−t0) = 0, we will apply formula ( ):
fY1
(y1) =
x2,2
x2,1
fX1,X2
y1
x2
,x2
1
|x2|
dx2, ( )
to
X1 = Z1 = Z0, X2 = Z2 = eA(t−t0)
> 0, Y1 = Z = Z1 Z2 :
f1(z,t) = fZ (z) =
z2,2
z2,1
fZ1,Z2
z
z2
,z2
1
z2
dz2
=
z2,2
z2,1
fZ0,A
z
z2
,
ln(z2)
t −t0
1
(z2)2(t −t0)
dz2, ˆz1 ≤ z ≤ ˆz2, t > t0,
where
z2,1 = ea1(t−t0)
, z2,2 = ea2(t−t0)
,
ˆz1 = z0,1ea1(t−t0), ˆz2 = z0,2ea2(t−t0), if z0,1 > 0,
ˆz1 = z0,1ea2(t−t0), ˆz2 = z0,2ea2(t−t0), if z0,1 z0,2 ≤ 0,
ˆz1 = z0,1ea2(t−t0), ˆz2 = z0,2ea1(t−t0), if z0,2 < 0.
f1(z0,t0) = fZ0
(z0) =
a2
a1
fZ0,A(z0,a)da, z0,1 ≤ z0 ≤ z0,2.

Example Case I.3: (Z0,A) is a random vector whose components are independent
fZ0,A(z0,a) =
4az0 if 0 < z0, a < 1,
0 otherwise.
we substitute into the obtained formula and after making some simpliﬁcations one
obtains:
f1(z,t) =



4z
(t −t0)2
et−t0
1
ln(z2)
(z2)3
dz2 if 0 ≤ z ≤ 1,
4z
(t −t0)2
et−t0
z
ln(z2)
(z2)3
dz2 if 1 ≤ z ≤ et−t0 ,
t > t0.
Let us take t0 = 0. For t > 0:
f1(z,t) =



z
t2
e−2t
−1+e2t
−2t if 0 ≤ z ≤ 1,
z
t2
−e−2t
(1+2t)+
1+2ln(z)
z2
if 1 ≤ z ≤ et .
t > 0.
For t = 0:
f1(z0,0) = fZ0
(z0) =
1
0
4az0 da = 2z0, z0,1 = 0 < z < 1 = z0,2.

Example : t > 0, t0 = 0.

Case II.3: (Z0,B) is a random vector
Z(t) = Z1(t)+Z2(t), where
Z1(t) = Z0,
Z2(t) = B(t −t0).
To compute the p.d.f. of Z = Z(t), t : t > t0 ﬁx, ﬁrst we will determine the joint p.d.f.
of Z1 = Z1(t) and Z2 = Z2(t) by R.V.T. method (two–dimensional version) to:
X1 = Z0, X2 = B, r1(z0,b) = z0, r2(z0,b) = b(t −t0),
Y1 = Z1, Y2 = Z2, s1(z1,z2) = z1, s2(z1,z2) = z2
t−t0
.
Hence, the Jacobian is given by:
J2 =
∂s1(z1,z2)
∂z1
∂s2(z1,z2)
∂z2
=
1
t −t0
> 0,
therefore
fZ1,Z2
(z1,z2) =
1
t −t0
fZ0,B z1,
z2
t −t0
,
where
z1,1 = z0,1 ≤ z1 ≤ z0,2 = z1,2, z2,1 = b1(t −t0) ≤ z2 ≤ b2(t −t0) = z2,2.

Now, we apply R.V.T. method (sum of two r.v.’s: Y1 = X1 +X2)
fY1
(y1) =
x1,2
x1,1
fX1,X2
(x1,y1 −x1)dx1,
to X1 = Z1, X2 = Z2 and Y1 = Z and we will obtain the p.d.f. of Z = Z1 +Z2:
f1(z,t) = fZ (z) =
z1,2
z1,1
fZ1,Z2
(z1,z −z1)dz1,
=
1
t −t0
z0,2
z0,1
fZ0,B z0,
z −z0
t −t0
dz0, ˆz1 ≤ z ≤ ˆz2, t > t0,
where
ˆz1 = z0,1 +b1(t −t0) ≤ z ≤ z0,2 +b2(t −t0) = ˆz2.
If t = t0, then Z(t) = Z(t0) = Z0 and the 1–p.d.f. is the Z0–marginal p.d.f.:
f1(z0,t0) = fZ0
(z0) =
b2
b1
fZ0,B (z0,b)db, z0,1 ≤ z0 ≤ z0,2.

Example Case II.3: (Z0,B) is a random vector whose components are dependent
fZ0,B (z0,b) =
1
4 + 1
4 (z0)3b − 1
4 z0b3 if −1 < z0 < 1, −1 < b < 1,
0 otherwise.
Example : t > 0, t0 = 0.
We substitute into the obtained formula and after making some simpliﬁcations to
obtain:
f1(z,t) =
1
t
min{1,z+t}
max{z−t,−1}
1
4
+
1
4
(z0)3 z −z0
t
−
1
4
z0
z −z0
t
3
dz0,
−1−t ≤ z ≤ 1+t.

Case III.1: Z0 is a r.v. Here, we illustrate the computation of the mean, variance and
probabilities of interest
Z(t) = ea(t−t0)
Z0 +
b
a
ea(t−t0)
−1 , t ≥ t0.
Next, we ﬁrst ﬁx t : t ≥ t0 and denote Z = Z(t). Then we apply R.V.T. method
(linear transformation: Y = αX +β, α = 0) to:
α = ea(t−t0)
> 0, β =
b
a
ea(t−t0)
−1 , X = Z0, Y = Z.
and, taking into account that fY (y) = 1
|α| fX
y−β
α the domain of r.v. Z0, one gets:
f1(z,t) = e−a(t−t0)
fZ0
e−a(t−t0)
z +
b
a
−
b
a
, z1 ≤ z ≤ z2, t ≥ t0,
where
z1 = z0,1ea(t−t0)
+
b
a
ea(t−t0)
−1 , z2 = z0,2ea(t−t0)
+
b
a
ea(t−t0)
−1 .

Example Case III.1: Z0 ∼ Exp(λ), λ > 0
f1(z,t) = λe
− a(t−t0)+λ (z+ b
a )e−a(t−t0)− b
a
,
b
a
ea(t−t0)
−1 ≤ z < +∞, t ≥ t0,
Example : λ = 1, t0 = 0 a = −1, b = 1.

Example Case III.1: Computing some statistical properties by the 1–p.d.f.
Moments w.r.t. the origin:
αn(t) = E (Z(t))k
=
∞
b
a ea(t−t0)−1
zk
f1(z,t)dz, k = 0,1,2,...
E[Z(t)] = α1(t) =
−bλ +ea(t−t0)(a+bλ)
λa
,
V[Z(t)] = α2(t)−(α1(t))2
=
e2a(t−t0)
λ2
.
Example : λ = 1, t0 = 0, a = −1, b = 1.
2 4 6 8 10
t
-1.0
-0.5
0.5
1.0
E@ZHtLD
0 1 2 3 4
t
0.2
0.4
0.6
0.8
1.0
Var@ZHtLD

The computation of probabilities can also be carried out directly through the 1-p.d.f.
For instance, it may be of interest to determine the probability that the solution lies
between two ﬁxed values, say, v1 = 2 and v2 = 3:
P[2 ≤ Z ≤ 3] =
3
2
f1(z,t)dz
= −e
λ
a b−(3a+b)ea(−t+t0)
+e
λ
a b−ea(−t+t0) b+aMax 2,
b −1+ea(t−t0)
a
.

Some important remarks regarding the application of R.V.T. technique:
limitations and possibilities
Remark 1: The importance of making an appropriate choice
Let us consider Case III.5. If we write the solution s.p. in the following form:
Z1(t) = Z0eA(t−t0),
Z2(t) = b
A eA(t−t0) −1 ,
then the application of R.V.T. (two–dimensional version) with the following choice:
X1 = Z0, X2 = A, r1(z0,a) = z0ea(t−t0), r2(a) = b
a ea(t−t0) −1 ,
Y1 = Z1, Y2 = Z2, s1(z1,z2) = ? s2(z2) = ?
does not lead to fruitful results since we cannot isolate z0 = s1(z1,z2) and
a = s2(z1,z2) and this would ruin our goal.

Notice that the previous drawback can be overcame as follows:
Z1(t) = Z0 + b
A eA(t−t0),
Z2(t) = − b
A .
and applying R.V.T. (two–dimensional version) with the following choice:
X1 = Z0, X2 = A, r1(z0,a) = z0 + b
a ea(t−t0), r2(a) = −b
a ,
Y1 = Z1, Y2 = Z2, s1(z1,z2) = z1e
b
z2
(t−t0)
+z2, s2(z2) = − b
z2
.
However, sometimes a good choice is not enough to apply R.V.T. method. Take a
meanwhile to deal with the apparent simplest Case III.3 where:
Z = r(A), where r(A) = z0eA(t−t0)
+
b
A
eA(t−t0)
−1 .
Can you isolate the r.v. A?

Lagrange–Bürmann Theorem
H : Suppose z is defined as a function of the variable a by an equation of the form:
z = r(a) where r is analytic about the point a0 where r (a0) = 0.
T : Then, it is possible to invert (or to solve) the equation for a: a = s(z) on a
neighbourhood N (r(a0);δ), δ > 0 of r(a0):
a = s(z) = a0 +
∞
∑
n=1
lim
a→a0
dn−1
dan−1
a−a0
r(a)−r(a0)
n
(z −r(a0))n
n!
, z ∈ N (r(a0);δ), δ > 0.
Step 1: Divide the domain of the map r (or equivalently, the domain of the r.v.
A) into k subintervals: A1,A2,...,Ak where r is monotone.
Step 2: For every subinterval Aj , 1 ≤ j ≤ k , select a0,j ∈ Aj such that
r (a0,j ) = 0. By Lagrange–Bürmann formula, construct the inverse, say sj (z), of
the map r(a) = rj (a) on Aj :
sj (z) = a0,j +
∞
∑
n=1
lim
a→a0,j
dn−1
dan−1
a−a0,j
r(a)−r(a0,j )
n
(z −r(a0,j ))n
n!
, z ∈ N (r(a0,j );δ).
Step 3: Compute the derivative of sj (z):
dsj (z)
dz
=
∞
∑
n=1
lim
a→a0,j
dn−1
dan−1
a−a0,j
r(a)−r(a0,j )
n
(z −r(a0,j ))n−1
(n −1)!
, z ∈ N (r(a0,j );δ).
Step 4: Construct the 1-p.d.f. of Z(t) as follows:
f1(z,t) =
k
∑
j=1
fA(sj (z))
dsj (z)
dz
.

Often, the above inﬁnite series must be truncated at the term Nj to control
computational burden:
sj,Nj
(z) = a0,j +
Nj
∑
n=1
lim
a→a0,j
dn−1
dan−1
a−a0,j
r(a)−r(a0,j )
n
(z −r(a0,j ))n
n!
.
Thus, an approximation of its derivative is:
dsj,Nj
(z)
dz
=
Nj
∑
n=1
lim
a→a0,j
dn−1
dan−1
a−a0,j
r(a)−r(a0,j )
n
(z −r(a0,j ))n−1
(n −1)!
.
Repeating the foregoing process on each interval Aj , 1 ≤ j ≤ k, one gets the
corresponding approximation of f1(z,t) given by:
f1(z,t) =
k
∑
j=1
fA(sj,Nj
(z))
dsj,Nj
(z)
dz
.

Example Case III.3: A ∼ Be(α = 2;β = 3), b = 1, t0 = 0, z0 = 1
Since in this case r(A) is monotone on the whole interval A1 = [0,1], we take k = 1.
In order to carry out computations, A1 has been split into 7 subintervals in accordance
with the process described previously. In each subinterval, an approximation of degree
Nj = 2, 1 ≤ j ≤ 7, has been used.
For the sake of clarity in the representation, due to diﬀerences in the scale the plot has
been split in two pieces: t ∈ [0,1] and t ∈ [1,2].
0.0
0.2
0.4
0.6
0.8
1.0
t
1
2
3
4
5
z
0
4
8
12
16
f1 z,t
0.9
1.2
1.5
1.8
2.1
t
3
6
9
12
z
0.0
0.2
0.4
0.6
0.8
1.0
f1 z,t

Remark 2: Computing the 2–p.d.f. of the solution s.p.
Let us consider Case II.3 and let us ﬁx t1,t2 such as t2 > t1 ≥ t0 and denote
Z1 = Z(t1) and Z2 = Z(t2). All we need to determine the 2–p.d.f. of the solution s.p.
Z(t) is computing the joint p.d.f. of r.v.’s Z1 and Z2. Notice that:
Z(t) = Z0 +B(t −t0).
Then, we apply R.V.T. (two–dimensional version) with the following choice:
X1 = Z0, X2 = B, r1(z0,b) = z0 +b(t1 −t0), r2(z0,b) = z0 +b(t2 −t0),
Y1 = Z1, Y2 = Z2, s1(z1,z2) = z1(t2−t0)−z2(t1−t0)
t2−t1
, s2(z1,z2) = z2−z1
t2−t1
,
Now, taking into account that:
ds1(z1,z2)
dz1
=
t2 −t0
t2 −t1
,
ds1(z1,z2)
dz2
= −
t1 −t0
t2 −t1
,
ds2(z1,z2)
dz1
= −
1
t2 −t1
,
ds2(z1,z2)
dz2
=
1
t2 −t1
,
one obtains the Jacobian:
|J2| =
1
t2 −t1
> 0.

Finally:
f2(z1,t1;z2,t2) = fZ1,Z2
(z1,z2)
= fZ0,B
z1(t2 −t0)−z2(t1 −t0)
t2 −t1
,
z2 −z1
t2 −t1
1
t2 −t1
,
where z1,1 ≤ z1 ≤ z1,2, z2,1 ≤ z2 ≤ z2,2 satisfy
z1,1 = z0,1 +b1(t1 −t0), z1,2 = z0,2 +b2(t1 −t0),
z2,1 = z0,1 +b1(t2 −t0), z2,2 = z0,2 +b2(t2 −t0).
From the 2–p.d.f., we can calculate relevant probabilistic properties such as the
correlation function:
ΓZ (t1,t2) =
∞
−∞
∞
−∞
z1z2f2(z1,t1;z2,t2)dz1 dz2.

Remark 3: Sometimes the computation of the 1–p.d.f. gives full information of the
solution s.p.
Let us consider Case III.1 for which:
Z(t) = Z0 +
b
a
ea(t−t0)
−
b
a
.
In this case, the solution s.p. at t2 can be represented as follows:
Z(t2) = Z0 + b
a ea(t2−t0) − b
a ,
= ea(t2−t1) Z0 + b
a ea(t1−t0) − b
a
= ea(t2−t1) Z(t1)+ b
a − b
a
= ea(t2−t1)Z(t1)+ b
a ea(t2−t1) −1 .
From this expression we see that the behaviour of the solution Z(t) at the time
instant t2 is deterministically given by a linear transformation of Z(t1). Therefore, the
computation of the 2-p.d.f. is not required.
Let us check it from another point of view!

Let us assume without loss of generality that the expectation of the initial condition is
zero: E[Z0] = 0 and its variance is σ2
Z0
> 0. Then it is easy to check that:
E[Z(ti )] = b
a ea(ti −t0) − b
a , i = 1,2,
σ2
Z(ti ) = σ2
Z0
e2a(ti −t0), i = 1,2,
E[Z(t1)Z(t2)] = σ2
Z0
ea(t2+t1−2t0) + b
a
2
ea(t2+t1−2t0) −ea(t2−t0) −ea(t1−t0) +1 .
Then the correlation coefficient function is given by
ρZ(t1),Z(t2) =
E[Z(t1)Z(t2)]−E[Z(t1)]E[Z(t2)]
σZ(t1)σZ(t2)
= 1.
Z(t2) is completely determined by Z(t1)!
Remark 4: Different but equivalent representations of the 1-p.d.f.
It is important to underline that there usually exist several ways to conduct the study
when applying R.V.T. method, although some of them are easier. Therefore, different
apparently results can appear.

Part III
Nonlinear Models in Epidemiolgy

Motivating the nonlinear case: The SI-type epidemiological model

S
I

β

S(t) number of susceptibles in the time instant t.
I(t) number of infected in the time instant t.
n size of the total population. It is assumed to be constant for all time t.
β > 0 rate of decline in the number of susceptibles.
S (t) = −β
n S(t)[n −S(t)], t > 0,
S(0) = m,
Putting the change of variable: P(t) = S(t)
n ∈ [0,1], the model can be recast as follows
P (t) = −β P(t)[1−P(t)], t > 0,
P(0) = P0 = m/n.

normalized SI-type epidemiological model
P (t) = −β P(t)[1−P(t)], t > 0,
P(0) = P0.
It is more realistic to assume that β and P0 are r.v.’s rather than deterministic
constants. We will assume that they are independent r.v.’s with p.d.f.’s fP0
(p0) and
fβ (β) and domains
DP0
= { p0 = P0(ω),ω ∈ Ω : 0 ≤ p0,1 ≤ p0 ≤ p0,2 ≤ 1},
Dβ = { β = β(ω),ω ∈ Ω : 0 < β1 < β < β2},
respectively.

To compute the 1-p.d.f. of the solution s.p. P(t). To this end, we make several
changes of variables to accommodate the nonlinear SI-model to random linear model
previously studied using the linearization technique:
First change of variable: Q(t) = 1
P(t) . Then, the problem SI-model writes
Q (t) = β Q(t)−β , t > 0,
Q(0) = 1
P0
.
Second change of variable: H(t) = Q(t)−1. This leads
H (t) = β H(t), t > 0,
H(0) = 1
P0
−1.

Using R.V.T. technique one can establish the following result:
Case I.3 of linear random model
H : Let us consider the linear random i.v.p.
˙Z(t) = AZ(t), t > t0,
Z(t0) = Z0,
Z0,A r.v.’s with joint p.d.f. fZ0,A(z0,a) (1)
and domains
DZ0
= {z0 = Z0(ω),ω ∈ Ω : z0,1 ≤ z0 ≤ z0,2}, DA = {a = A(ω),ω ∈ Ω : a1 ≤ a ≤ a2}.
T : Then, the 1-p.d.f. of the solution s.p. Z(t) of (1) is given by
f1(z,t) =
1
t −t0
ea2(t−t0)
ea1(t−t0)
fZ0,A
z
ξ
,
ln(ξ)
t −t0
1
ξ2
dξ, z1 ≤ z ≤ z2, ∀t > t0,
where
z1 = z0,1ea1(t−t0), z2 = z0,2ea2(t−t0), if z0,1 > 0,
z1 = z0,1ea2(t−t0), z2 = z0,2ea2(t−t0), if z0,1 z0,2 ≤ 0,
z1 = z0,1ea2(t−t0), z2 = z0,2ea1(t−t0), if z0,2 < 0.
If t = t0,
f1(z0,t0) =
a2
a1
fZ0,A(z0,a)da, z0,1 ≤ z0 ≤ z0,2.

We identify the inputs of both problems:
H (t) = β H(t), t > 0,
H(0) = 1
P0
−1.
≡
˙Z(t) = AZ(t), t > t0,
Z(t0) = Z0,
Z0 =
1
P0
−1, A = β, Z(t) = H(t), t0 = 0,
and, ﬁxed t > 0, the p.d.f. of r.v. H = H(t) yields
fH (h) =
1
t
ea2t
ea1t
fZ0,A
h
ξ
,
ln(ξ)
t
1
ξ2
dξ =
1
t
ea2t
ea1t
fZ0
h
ξ
fA
ln(ξ)
t
1
ξ2
dξ ,
where independence between r.v.’s Z0 and A has been used.
Now, we need to write fH (h) in terms of the p.d.f.’s of the inputs P0 and β. With this
aim we establish the following specialization of R.V.T. method:

R.V.T. technique: inverse-vertical translation transformation
H : Let c ∈ R and X be an absolutely continuous real r.v. deﬁned on a probability
space (Ω,F,P), with p.d.f. fX (x). Assume that X is a non-zero r.v. and let us denote
by DX the domain of r.v. X, where
DX = I−
x ∪I+
x ,
I−
x = {x = X(ω) ∈ R : −∞ < x < 0, ω ∈ Ω} ,
I+
x = {x = X(ω) ∈ R : 0 < x < +∞, ω ∈ Ω} .
T : Then, the p.d.f. fY (y) of the inverse-vertical translation transformation
Y = 1
X +c is given by
fY (y) =
1
(y −c)2
fX
1
y −c
, y ∈ DY = I−
y ∪I+
y ,
I−
y = {y ∈ R : y < c} ,
I+
y = {y ∈ R : y > c} .
Applying this result to:
X = P0 ,Y = Z0, c = −1, Z0 =
1
P0
−1
one gets
fH (h) =
1
t
ea2t
ea1t
fZ0
h
ξ
fA
ln(ξ)
t
1
ξ2
dξ =
1
t
eβ2t
eβ1t
fP0
ξ
h +ξ
fβ
ln(ξ)
t
1
(h +ξ)2
dξ .

Remember that: P(t) = 1
H(t)+1 , so fixed t, we finally need to recover the p.d.f. fP (p)
of r.v. P = P(t) from the p.d.f. fH (h). With this end, we establish the following result:
R.V.T. technique: inverse-horizontal translation transformation
H : Let d ∈ R and X be an absolutely continuous real r.v. defined on a probability
space (Ω,F,P), with p.d.f. fX (x). Assume that X −d is a non-zero r.v. and let us
denote by DX the domain of r.v. X, where
DX = I−
x ∪I+
x ,
I−
x = {x = X(ω) ∈ R : −∞ < x < d , ω ∈ Ω} ,
I+
x = {x = X(ω) ∈ R : d < x < +∞, ω ∈ Ω} .
T : Then, the p.d.f. fY (y) of the inverse-horizontal translation transformation
Y = 1
X−d is given by
fY (y) =
1
y2
fX
1
y
+d , y ∈ DY = I−
y ∪I+
y ,
I−
y = {y ∈ R : y < 0} ,
I+
y = {y ∈ R : y > 0} .
Applying this result to:
X = H ,Y = P, d = −1,
one gets
fP (p) =
1
p2
fH
1
p
−1 =
1
t
eβ2t
eβ1t
fP0
p ξ
1−p +p ξ
fβ
ln(ξ)
t
1
(1−p +p ξ)2
dξ .

Summarizing,
1-p.d.f. of the solution s.p. of the normalized SI-type epidemiological model
H : Let us consider the random i.v.p.:
P (t) = −β P(t)[1−P(t)], t > 0,
P(0) = P0,
where β and P0 are independent r.v.’s with p.d.f.’s fP0
(p0) and fβ (β) and domains
DP0
= {p0 = P0(ω),ω ∈ Ω : 0 ≤ p0,1 ≤ p0 ≤ p0,2 ≤ 1},
Dβ = {β = β(ω),ω ∈ Ω : 0 ≤ β1 < β < β2},
respectively
T : Then, the 1-p.d.f. of the solution s.p. P(t) is given by:
f1(p,t) =



1
t
eβ2t
eβ1t
fP0
p ξ
1−p +p ξ
fβ
ln(ξ)
t
1
(1−p +p ξ)2
dξ if t > 0,
fP0
(p0) if t = 0.

From this 1-p.d.f. important information related to SI-epidemiological model can be
computed straightforwardly:
Mean and variance:
µP (t) = E[P(t)] =
∞
−∞
pf1(p,t)dp, (σP (t))2
= V[P(t)] =
∞
−∞
p2
f1(p,t)dp−(µP (t))2
,
Bounds for probabilities upon intervals of interest and more:
P[|P(t)− µP (t)| ≥ λ] ≤
(σP (t))2
λ2
,
P[a ≤ P(t) ≤ b] =
b
a
f1(p,t)dp,
Conﬁdence intervals: Fixed α ∈ (0,1), for each time instant t one can
determine x1(t) and x2(t), such that
1−α = P({ω ∈ Ω : P(t;ω) ∈ [x1(t),x2(t)]}) =
x2(t)
x1(t)
f1(p,t)dp ,
and
x1(t)
0
f1(p,t)dp =
α
2
=
1
x2(t)
f1(p,t)dp .

Further relevant information that can be determined from the 1-p.d.f. includes:
Distribution of time until a given proportion of susceptibles remains in the
population
This distribution answers the following question:
What is the expected time before ρ = 80% of the population remains susceptible?
This distribution is computed from the solution of the SI-model:
P(T) =
P0
eβ T (1−P0)+P0
⇒ {ρ = P(T)} ⇒ T =
1
β
ln
P0(1−ρ)
ρ(1−P0)
.
Now, we apply two-dimensional R.V.T. technique to
X1 = β, Y1 = T, Y1 = r1(X1,X2) =
ln
X2(1−ρ)
ρ(1−X2)
X1
, X1 = s1(Y1,Y2) =
ln
Y2(1−ρ)
ρ(1−Y2)
Y1
,
X2 = P0, Y2 = P0, Y2 = r2(X1,X2) = X2, X2 = s2(Y1,Y2) = Y2,
and taking into account that ∂s2(y1,y2)
∂y1
= 0, the jacobian is
J = −
1
(y1)2
ln
y2(1−ρ)
ρ(1−y2)
= 0,
hence the joint p.d.f. of (Y1,Y2) = (T,P0) is given by
fT,P0
(t,p0) =
1
t2
ln
p0(1−ρ)
ρ(1−p0)
fβ,P0
1
t
ln
p0(1−ρ)
ρ(1−p0)
,p0
=
1
t2
ln
p0(1−ρ)
ρ(1−p0)
fβ
1
t
ln
p0(1−ρ)
ρ(1−p0)
fP0
(p0),

Therefore, the P0-marginal distribution of fT,P0
(t,p0) yields the p.d.f. of T
fT (t;ρ) =
1
t2
min(p0,2,c2)
max(p0,1,c1)
ln
p0(1−ρ)
ρ(1−p0)
fβ
1
t
ln
p0(1−ρ)
ρ(1−p0)
fP0
(p0) dp0, p0 ∈ DP0
,
where
DP0
= {p0 = P0(ω),ω ∈ Ω : 0 ≤ p0,1 ≤ p0 ≤ p0,2 ≤ 1}
and
c1 =
ρ eβ1t
ρeβ1t +(1−ρ)
, c2 =
ρ eβ2t
ρeβ2t +(1−ρ)
.
For t and ρ previously ﬁxed, these values have been determined by imposing that
β1 <
1
t
ln
p0(1−ρ)
ρ(1−p0)
< β2,
being
Dβ = {β = β(ω),ω ∈ Ω : 0 ≤ β1 ≤ β ≤ β2 ≤ 1}.

Modelling the diﬀusion of a new technology
year 1995 1996 1997 1998 1999 2000 2001 2002 2003
penetration
rate (xi )
2.3 7.5 10.2 16.2 37.3 59.9 72.6 81.9 89.3
year 2004 2005 2006 2007 2008 2009 2010 2011 −−
penetration
rate (xi )
91.2 99.2 104.4 108.9 109.6 111.4 111.7 113.9 −−
Remarks:
xi represents the rate of mobile phone lines per 100 inhabitants taking as
reference the Spanish census corresponding to year 2011 updated by INE
(National Statistics Institute of Spain).
xi , may be greater than 100% since any individual can possess more than one
mobile phone line. In order to be able to apply the SI-model, two
transformations on the data listed in previous table will be done.
Transformation:
Pi = 1−xi /115, i = 0,1,...,16 ⇒ 0 ≤ Pi ≤ 1.
1 Standardize the values xi by assuming a saturation value of 115.
2 Since the unknown P(t) of SI-model represents the percentage of susceptibles
instead of infected (i.e., the percentage of people who have already adopted the
mobile phone technology).

year 1995 1996 1997 1998 1999 2000 2001 2002 2003
Pi 0.9800 0.9348 0.9113 0.8591 0.6757 0.4791 0.3687 0.2878 0.2235
year 2004 2005 2006 2007 2008 2009 2010 2011 −−
Pi 0.2070 0.1374 0.0922 0.0530 0.0470 0.0313 0.028695 0.0096 −−
Assumptions:
Pi ∈ (0,1) ⇒ P0 ∼ Be(a;b), β > 0 ⇒ β ∼ Ga(λ;τ).
Fitting the model parameters: Determining a,b,λ,τ:
1 Split the sample data: We take data from t0 = 1995 to t12 = 2007.
2 Minimizing the mean square error:
min
a,b,λ,τ>0
E(a,b,λ,τ) =
12
∑
i=0
(Pi −E[P(t;a,b,λ,τ)])2
=
12
∑
i=0
Pi −
1
0
p f1(p,t)dp ,
2
where
f1(p,t) =
1
t
∞
1
fP0
p ξ
1−p +p ξ
fβ
ln(ξ)
t
1
(1−p +p ξ)2
dξ ,
fP0
p ξ
1−p +p ξ
=
Γ(a+b)
Γ(a)Γ(b)
p ξ
1−p +p ξ
a−1
1−p
1−p +p ξ
b−1
,
fβ
ln(ξ)
t
= λτ ξ− λ
t
1
Γ(τ)
ln(ξ)
t
τ−1
.

Using the Nelder-Mead algorithm we obtain:
a∗
= 114.95, b∗
= 1.83, λ∗
= 27.36, τ∗
= 0.032.
Out[52]=
1995
1997
1999
2001
2003
2005
2007
2009
2011
t
0.0
0.5
1.0
p
0
2
4
f1 p,t
2000 2005 2010 2015 2020 2025
t
0.2
0.4
0.6
0.8
1.0
ΜP t
2000 2005 2010 2015 2020 2025
t
0.05
0.10
0.15
0.20
ΣP t

Validation and prediction using conﬁdence intervals
2000 2005 2010
t
0.2
0.4
0.6
0.8
1.0
P t
real data
real data
expectation
95 Confidence Interval
P.d.f. of the time T until a proportion ρ = 90% of susceptibles remain in the
population
E[T] =
∞
0
tfT (t;0.90)dt = 2.65
0.
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
Ρ
5
10
15
t
0.0
0.1
0.2
0.3
0.4
fT t

Motivating the nonlinear case: The SIS-type epidemiological model
S(t) number of susceptibles in the time instant t.
I(t) number of infected in the time instant t.
n size of the total population. It is assumed to be constant for all time t.
β > 0 rate of decline in the number of susceptibles.
γ > 0 rate of infected that recover from the disease.



S (t) = −βS(t)I(t)+γI(t),
I (t) = βS(t)I(t)−γI(t),
t > 0, S(0) = S0, I(0) = I0,

Taking into account that the solution (S(t),I(t)) of the SIS-model can be written as
follows
S(t) =
γ(1−S0)+(S0β −γ)e(γ−β)t
β(1−S0)+(S0β −γ)e(γ−β)t
,
I(t) =
(β −γ)(1−S0)e(β−γ)t
β(1−S0)e(β−γ)t +S0β −γ
,
t ≥ 0. (2)
Applying the RVT technique, one can establish the following results
1-p.d.f. of the solution s.p. of the SIS-type epidemiological model
f1(s,t) =
Dγ Dβ
fS0,γ,β
ξ +e(ξ−η)t (−1+s)ξ −sη
ξ +e(ξ−η)t (−1+s)η −sη
,ξ,η
e(ξ−η)t (ξ −η)2 dη dξ
(ξ +e(ξ−η)t (−1+s)η −sη)2
,
f1(i,t) =
Dγ Dβ
fS0,γ,β
ξ −η −e(ξ−η)t iξ +iη
ξ −η −e(ξ−η)t iη +iη
,ξ,η
e(ξ−η)t (ξ −η)2dη dξ
(ξ −η −e(ξ−η)t iη +iη)2
,
where
Dβ , Dγ ,
are the domains of r.v.’s β and γ, respectively.

Distribution of time until a given proportion of susceptibles and infected remains in
the population
f1(t,ρS ) =
Dγ Dβ
fS0,γ,β
ξ(1+et(ξ−η)(−1+ρS ))−ηρS
ξ +η(et(ξ−η)(−1+ρS )−ρS )
,ξ,η
×
et(ξ−η)(ξ −η)2(1−ρS )|ξ −ηρS |
(ξ +η(et(ξ−η)(−1+ρS )−ρS ))2
dη dξ .
f1(t,ρI ) =
Dγ Dβ
fS0,γ,β
ξ +η(−1+ρI )−et(ξ−η)
ξ −η(1+(−1+et(ξ−η))ρI )
,ξ,η
×
et(ξ−η)(ξ −η)2(ξ +η(−1+ρI ))ρI
(ξ −η(1+(−1+et(ξ−η))ρI ))2
dη dξ .

In epidemiology, the basic reproduction number, R0, associated to an infection is
useful to elucidate whether will spread out or not. In the case of the SIS model, this
value and its relationship with the propagation of the epidemic in the long run is given
by
R0 =
β
γ
,
if R0 < 1 ≡ β < γ, then the diseases will die out as t → +∞,
if R0 > 1 ≡ β > γ, then the diseases will spread out as t → +∞.
This classiﬁcation is easily derived from expression of I(t), or equivalently of S(t),
since
lim
t→+∞
I(t) = lim
t→+∞
(β −γ)(1−S0)e(β−γ)t
β(1−S0)e(β−γ)t +S0β −γ
= 0 if β < γ ,
lim
t→+∞
S(t) = lim
t→+∞
γ(1−S0)+(S0β −γ)e(γ−β)t
β(1−S0)+(S0β −γ)e(γ−β)t
= 1 if β < γ .

In our context, both β and γ are r.v.’s, so that the requirement for epidemic extinction
in the deterministic framework β < γ means the computation of the following
probability in the stochastic scenario
P[S ], S = {ω ∈ Ω : β(ω) < γ(ω)} = {ω ∈ Ω : R0(ω) < 1}. (3)
This key probability can be computed by taking advantage of RVT. Using the mapping
U = (U1,U2)T
= (γ,β)T
, V =
U2
U1
=
β
γ
= R0 ,
one gets
fR0
(r0) =
D(γ)
fγ,β (ξ,r0ξ)|ξ|dξ ,
where fγ,β (·,·) denotes the (γ,β)–marginal distribution of the joint p.d.f. of the
random inputs (S0,γ,β. This allows us to compute the target probability
P[S ] =
1
0 D(γ)
fγ,β (ξ,r0ξ)|ξ|dξ dr0 .

Modelling the the spread of smoking in Spain
year 1987 1993 1995 1997 2001 2003 2006
(tj ) (j = 0) (j = 6) (j = 8) (j = 10) (j = 14) (j = 16) (j = 19)
Sj 0.4488 0.5144 0.5278 0.5514 0.5783 0.6244 0.6467
J = {0,6,8,10,14,16,19}
Assumptions:
S0 ∼ Be(a;b); β > 0 ⇒ β ∼ Exp]0,1000[(λβ ); γ > 0 ⇒ γ ∼ N[0,1](µγ ;σγ ).
Fitting the model parameters: Determining a,b,λβ ,µγ ,σγ :
min
a,b,λβ ,µγ ,σγ >0
E(a,b,λβ ,µγ ,σγ ) = ∑
j∈J
Sj −E[S(tj ;a,b,λβ ,µγ ,σγ )]
2
,
where,
E[S(tj ;a,b,λβ ,µγ ,σγ )] =
1
0
sf1(s,tj )ds , j ∈ J .
f1(s,t) =
1
0
1000
0
fS0
fγ (ξ)fβ (η)
×
e(ξ−η)t (ξ −η)2 dη dξ
(ξ +e(ξ−η)t (−1+s)η −sη)2
.

Note that the p.d.f.’s for input data are
fS0
=
Γ(a+b)
Γ(a)Γ(b)
a−1
×
e(ξ−η)t (−1+s)(η −ξ)
b−1
,
fβ (η) =
λβ e−λβ η
1000
0
λβ e−λβ η
dη
,
and
fγ (ξ) =



e
−
(ξ−µγ )2
2(σγ )2 1
√
2πσγ
1
2 erfc
µγ −1
√
2σγ
− 1
2 erfc
µγ
√
2σγ
, if 0 < ξ ≤ 1,
0, otherwise,
being erfc(z) = 1− 2√
π
z
0
e−t2
dt the complementary error function.

Using the Nelder-Mead algorithm we obtain:
a∗
= 708.755, b∗
= 893.394, λ∗
β = 1362.230, µ∗
γ = 0.0231162, σ∗
γ = 0.0000526.
1987
1989
1991
1993
1995
1997
1999
2001
2003
2005
t
0.4
0.5
0.6
0.7
s
0
20
40
f1(s,t)
1990 2000 2010 2020
t
0.45
0.50
0.55
0.60
0.65
0.70
0.75
μS(t)
1990 2000 2010 2020
t
0.007
0.008
0.009
0.010
0.011
0.012
σS(t)

Validation and prediction using confidence intervals
1990 1995 2000 2005
t
0.45
0.50
0.55
0.60
0.65
S(t)
real data
μS(t)
μS(t)±2σS(t)
year 1987 1993 1995 1997 2001 2003 2006
(tj ) (j = 0) (j = 6) (j = 8) (j = 10) (j = 14) (j = 16) (j = 19)
Confidence level 0.9550 0.9544 0.9545 0.9546 0.9549 0.9550 0.9552
Table: Probabilities associated to the confidence intervals built according to the SIS
model.

Expected time until a certain proportion ρS of the population remains non-smoker
f1(t,ρS ) =
1
0
+∞
0
fS0
ξ(1+et(ξ−η)(−1+ρS ))−ηρS
ξ +η(et(ξ−η)(−1+ρS )−ρS )
fγ (ξ)fβ (η)
×
et(ξ−η)(ξ −η)2(1−ρS )|ξ −ηρS |
(ξ +η(et(ξ−η)(−1+ρS )−ρS ))2
dη dξ ,
ρS 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85
E[TS ] 0.59 4.78 9.42 14.61 20.51 27.32 35.40 45.30 58.10
Table: Expectation of time TS until a proportion, ρS , of the population
remains non-smoker for diﬀerent values ρS .
E[TS ] =
∞
0
tfTS
(t;0.75)dt = 35.4013.
This means that the middle of the year 2023 approximately
represents the average time until 75% of the Spanish men aged
over 16 years old population will be non-smokers.

0.45
0.5
0.55
0.6
0.65
0.7
0.75
0.8
ρS
1987
1992
1997
2002
2007
2012
2017
2022
2027
t
0.0
0.2
0.4
f1(t,ρS)
Figure: Plot of the 1-p.d.f. of the time TS until a proportion
ρS ∈ {0.45,0.50,0.55,0.60,0.65,0.70,0.75} of the population remains susceptible.
Finally, we compute the probability of the event S previously introduced
P[S ] =
1
0
1
0
fγ (ξ)fβ (r0ξ)|ξ|dξ dr0 = 0.999453,
where fβ (η) and fγ (ξ) are the p.d.f.’s of β and γ, respectively. This is in accordance
with the interpretation of the basic reproductive number R0: the percentage of
Spanish smoker men older than 16 years old will likely disappear as t tends to +∞.

The randomized Bertalanffy model: A fish weight growth over the time.
˙W (t) = −λW (t)+η(W (t))2/3, t ≥ t0 ,
W (t0) = W0 .
W (t): fish weight growth at time instant t.
η: intrinsic growth rate.
λ: linear coefficient.
We shall assume that all these inputs are r.v.’s with joint p.d.f. fW0,η,λ (w0,η,λ)
and
P[{ω ∈ Ω : W0(ω) = 0}] = 1, P[{ω ∈ Ω : λ(ω) = 0}] = 1.
Our goals are
1. To determine the 1-p.d.f. of the solution applying RVT method.
2. To use real data in order to assign a reliable probabilistic distribution to random
inputs using an inverse frequentist technique.
3. To construct both punctual and probabilistic predictions based on confidence
intervals.

Step 1: To determine the 1-p.d.f. of the solution
Considering the change of variable W (t) = (Z(t))3
, applying the following result
A particular case of RVT
Let Z : Ω → R be an absolutely continuous real r.v. deﬁned on a probability space
(Ω,F,P), with p.d.f. fZ (z). Assume that Z(ω) = 0 for all ω ∈ Ω. Then, the p.d.f.
fW (w) of the transformation W = Z3 is given by
fW (w) =
1
3
fZ
3
√
w |w|−2/3
.
Then, the 1-p.d.f. of the solution W (t) of the Bertalanﬀy model is
f1(w,t)
=
1
3
fZ w1/3
|w|−2/3
=
D(η) D(λ)
fW0,η,λ

 e(1/3)λ(t−t0)λw1/3 +η −e(1/3)λ(t−t0)η
λ
3
ηλ


×
e(1/3)λ(t−t0)λw1/3 +η −e(1/3)λ(t−t0)η
λ
2
e(1/3)λ(t−t0)
|w|−2/3
dλ dη.

Step 2: To assign a probabilistic distribution to random inputs
using real data
ti (years) 1 2 3 4 5 6 7
wi (lbs) 0.2 0.4 0.6 0.9 1 1.3 1.6
ti (years) 8 9 10 11 12 13 14
wi (lbs) 1.8 2.3 2.6 2.9 3.1 3.4 3.7
ti (years) 15 16 17 18 19 20 21
wi (lbs) 4.5 5.2 5.7 6.2 6.5 6.7 6.8
ti (years) 22 23 24 25 26 27 28
wi (lbs) 7.2 8.2 9 9.5 10 10.5 11
ti (years) 29 30 31 32 33
wi (lbs) 11.5 12 12.5 13 14
Table: Fish weights wi for walleye species in lbs every year ti , 1 ≤ i ≤ 33 = N.

To assign a reliable probability distribution to these data, Frequentist Inverse
Technique will be applied.
STEP 2.1: It is assumed that the measured quantity of interest, fish weights (wi ),
are corrupted by measurement errors εi .
wi = W (ti ;q) = W (ti ;w0,η,λ)+εi , 1 ≤ i ≤ 33 = N ,
where errors are assumed i.i.d. and εi ∼ N(0;σ2), being σ > 0 fixed but unknown.
As a consequence of this assignment the probabilistic distribution for the random
vector Q = (W0,η,λ) is assumed to be
Q = (W0,η,λ) ∼ N3(µQ;ΣQ),
where
µQ = ( ˆw0, ˆη,ˆλ) is defined from appropriate estimates of (w0,η,λ).
ΣQ represents the variance-covariance matrix.

STEP 2.2: A least squares ﬁt to the data yields the following
parameter estimates µQ = ( ˆw0, ˆη,ˆλ)
ˆw0 = 0.365934, ˆη = 0.305461, ˆλ = 0.0880184.
The residuals of the ﬁtting are,
εi = W (ti ; ˆw0, ˆη,ˆλ)−wi , 1 ≤ i ≤ 33 = N.
We need to check
i.i.d.
5 10 15 20 25 30
t
-0.4
-0.2
0.2
0.4
0.6
ϵi(t)

normality
Normality Test Statistic p-value
Shapiro-Walk Test 0.958995 0.242077
-0.4 -0.2 0.0 0.2 0.4
-0.4
-0.2
0.0
0.2
0.4
0.6
Normal Theoretical Quantiles
Residuals

STEP 2.3: Determine the sensibility matrix
According to frequentist parameter estimation method ﬁrst we compute the sensitivity
matrix
χ(Q) =











∂W (t1;Q)
∂W0
···
∂W (t33;Q)
∂W0
∂W (t1;Q)
∂η
···
∂W (t33;Q)
∂η
∂W (t1;Q)
∂λ
···
∂W (t33;Q)
∂λ











T
Q=( ˆw0,ˆη,ˆλ)
.
from the solution
W (t) = (W0)1/3
e−(1/3)λ(t−t0)
−
η
λ
e−(1/3)λ(t−t0)
+
η
λ
.

5 10 15 20 25 30
t
1
2
3
4
∂w
∂w0
(ti ; Q)
5 10 15 20 25 30
t
20
40
60
80
100
120
∂w
∂η
(ti ; Q)
5 10 15 20 25 30
t
-200
-150
-100
-50
∂w
∂λ
(ti ; Q)
Figure: Top: Left: ∂W
∂w0
(ti ;Q). Right: ∂W
∂η (ti ;Q). Bottom: ∂W
∂λ (ti ;Q).
ti = i,1 ≤ i ≤ 33 = N.

Then, the following probabilistic distribution has been assigned to model parameters
Q = (W0,η,λ) ∼ N3(µQ;ΣQ),
where
The mean vector has been previously estimated by mean square method
µQ = (0.365934,0.305461,0.0880184).
The variance-covariance matrix is
ΣQ = σ2
(χ(Q))T
χ(Q)
−1
=





0.0029288 −0.000812275 −0.000400288
−0.00081227 0.000268075 0.000136915
−0.000400288 0.000136915 0.0000705259





,
being σ the error standard deviation estimate:
σ =
33
∑
i=1
(εi )2 = 0.214435.

2 5 8 1114172023 26
29
32
t
0
5
10
w
0
5
10
f1(w,t)
1.11.21.31.41.5
1.6
1.7
1.8
1.9
2.
t
0.2
0.3
0.4
0.5
0.6
0.7
w
0
2
4
6
8
f1(w,t)
Figure: Left: 1-p.d.f. of the solution stochastic process to random Bertalanﬀy model
given for all the times of the sample, t ∈ {2,...,33 = N}. Right: Detailed
representation of the 1-p.d.f. for the times t ∈ {1.1,1.2,...,2}.

Step 3: To construct punctual and probabilistic predictions based
on confidence intervals
5 10 15 20 25 30
t
2
4
6
8
10
12
14
W(t)
real data
expectation
99% Confidence interval
Figure: Expectation (solid line) and 99%–confidence intervals (dotted lines). Points
represent fish weigh.

Conclusions and forthcoming work
1 Random Variable Transformation (R.V.T.) method is a powerful tool to compute
the 1-p.d.f. of the solution stochastic process of Random Differential Equations.
2 The application of this technique has been shown for first order lineal and
nonlinear random differential equations, but it can be extended to second-order
differential equations and random difference equations.
3 Extension of the results for systems of both random differential and difference
equations.
4 Application of R.V.T. technique together with numerical methods.

References (Continuous models based on Random Differential Equations)
1 Casabán M.C., Cortés J.C., Romero J.V., Roselló M.D. (2014): Determining the
first probability density function of linear random initial value problems by the
Random Variable Transformation (R.V.T.) technique: A comprehensive study,
Abstr. & Appl. Anal. 2014- ID248512, pp: 1–25.
2 Casabán M.C., Cortés J.C., Navarro-Quilés, A., Romero J.V., Roselló M.D.,
Villanueva, R.J. (2015): Probabilistic solution of the random homogeneous
Riccati differential equation: A comprehensive case-study by using linearization
and the Random Variable Transformation techniques, J. Comput. Appl. Math.
24, pp: 20–35.
Villanueva, R.J. (2016): Computing probabilistic solutions of the Bernoulli
random differential equation, J. Comput. Appl. Math. (accepted).
4 Casabán M.C., Cortés J.C., Romero J.V., Roselló M.D. (2016): Probabilistic
solution of random autonomous first-order linear systems of ordinary differential
equations, Romanian Reports in Physics (accepted).
5 Casabán M.C., Cortés J.C., Romero J.V., Roselló M.D. (2016): Solving random
homogeneous linear second-order differential equations: A full probabilistic
description, Mediterranean J. of Mathematics (accepted).

References (Discrete/Continuous Models based on Random Differential Equations)
1 Casabán M.C., Cortés J.C., Romero J.V., Roselló M.D. (2016): Random
first-order linear discrete models and their probabilistic solution: A
comprehensive study, Abstr. & Appl. Anal. 2016-ID6372108, pp: 1–22.
solution of random homogeneous linear second-order difference equations, Appl.
Math. Lett. 34, pp: 27–33.
solution of random SI-type epidemiological models using the Random Variable
Transformation technique, Comm. Nonl. Sc. Num. Simul. 24(1-3), pp: 86–97.
Villanueva, R.J. (2016): A comprehensive probabilistic solution of random
SIS-type epidemiological models using the Random Variable Transformation
technique, Comm. Nonl. Sc. Num. Simul. 32, pp: 199–210.

Computational Methods for Random
Epidemiological Models
Métodos Computacionales para el Estudio de
Modelos Epidemiológicos con Incertidumbre
Conferencias de Investigación para Posgrado 2016
Universidad Complutense de Madrid
24 junio de 2016
Prof. Dr. Juan Carlos Cortés
Instituto Universitario de Matemática Multidisciplinar
Universitat Politècnica de València

Métodos computacionales para el estudio de modelos epidemiológicos con incertidumbre - Juan Carlos Cortés

More Related Content

What's hot (20)

Viewers also liked (15)

Similar to Métodos computacionales para el estudio de modelos epidemiológicos con incertidumbre - Juan Carlos Cortés (20)

More from Facultad de Informática UCM (20)

Recently uploaded (20)

Métodos computacionales para el estudio de modelos epidemiológicos con incertidumbre - Juan Carlos Cortés