Since January 2020 Elsevier has created a COVID-19 resource centre with

free information in English and Mandarin on the novel coronavirus COVID-

19. The COVID-19 resource centre is hosted on Elsevier Connect, the
company's public news and information website.

Elsevier hereby grants permission to make all its COVID-19-related

research that is available on the COVID-19 resource centre - including this
research content - immediately available in PubMed Central and other
publicly funded repositories, such as the WHO COVID database with rights
for unrestricted research re-use and analyses in any form or by any means
with acknowledgement of the original source. These permissions are
granted for free by Elsevier for as long as the COVID-19 resource centre
remains active.
 CHAPTER 10

Analytical solutions and parameter

estimation of the SIR epidemic model
Dimiter Prodanova,b
a Sustainability, Environment, Health and Safety, Imec, Leuven, Belgium b ITSDP, IICT, Bulgarian

Academy of Sciences, Sofia, Bulgaria

10.1 Introduction
The origin of the mathematical modeling of epidemics can be traced back to XVIIth cen-
tury with the work of Daniel Bernoulli, who was trained as a physician, and whose results
justified the practice of inoculation against smallpox. This represented a major milestone
of public health. It is interesting to note that some subsequent major steps in the epidemi-
ology of compartmental models were achieved not by pure mathematicians, but by public
health physicians or biologists, such as Sir R.A. Ross, W.H. Hamer, A.G. McKendrick, and
W.O. Kermack [1]. As a notable example of such progress, the SIR (Susceptible-Infected-
Removed) epidemiological model was published in 1927 by Kermack and McKendrick in
1927 to study the plague and cholera epidemics in London and Bombay [2]. Even to date,
the SIR model remains a cornerstone of mathematical epidemiology [3,4]. It should be noted
that the full model derived by Kermack and McKendrick was formulated in terms of convo-
lution integrals, while the more popular form used in the present literature is actually only its
simplified case. The full SIR model included dependence on the age of infection, that is, the
time since an individual becomes infected. In this sense, the full model can be used to provide
a unified approach to compartmental epidemic models. It is also worth mentioning that the
SIR model in discrete time is equivalent to a renewal equation with a geometric generation-
interval distribution with probability parameter γ t [5]. This comes from the fact that the
limiting distribution of the geometric probability distribution is the exponential one.
Outside epidemiology, the SIR model is also extensively used in modeling of online social
networks, viral marketing, diffusion of ideas, spread of computer viruses, financial network
contagion, etc. (see recent survey by Rodrigues and the references therein [6]).
The SIR model can be extended in two directions – either by adding a final state, e.g. “de-
ceased” individuals – D; or by adding one or more intermediate nonobservable populations –

Mathematical Analysis of Infectious Diseases

https://doi.org/10.1016/B978-0-32-390504-6.00015-2 163
Copyright © 2022 Elsevier Inc. All rights reserved.
164 Chapter 10

Figure 10.1: The SIR model input-output diagram.

e.g. “exposed” E individuals. Distinct possibilities include the SEIR and SEIS models, with
an exposed period between getting infected and becoming infective, and SIRS models, with
temporary immunity conveyed upon recovery from the initial infection.
The analytical solution of the SIR model was formulated for the S-variable in a parametric
form in [7]. Recently, Barlow and Weinstein have introduced numerical solutions based on
asymptotic rational approximants [8]. Prodanov established the solution in terms of numeri-
cal inversion of the parametric solution by Newton iteration for the I-variable [9]. The present
chapter focuses on some recent theoretical developments around the SIR model from the per-
spective of real analysis and theory of special functions.

10.2 The SIR model

The full SIR model can be simplified under the assumptions that the rate of infection is con-
stant and there is no demographic variation. The simplified SIR model is a deterministic
model formulated in terms of ordinary differential equations (ODEs) in terms of 3 popula-
tions of individuals. The S population consists of all individuals susceptible to the infection of
concern. The I population comprises the infected individuals. These persons have the disease
and can transmit it to the susceptible individuals. The model can be schematically represented
in the input-output diagram presented in Fig. 10.1.
The standard assumptions behind the SIR model are:
1. A typical member of the population S gets infected by a typical member of the I popula-
tion with certain probability per unit time.
2. Infected individuals recover at rate γ I per unit time.
3. The R population cannot become re-infected and the recovered individuals cannot trans-
mit the disease to others.
4. There is no entry into or departure from the overall population, except possibly through
death from the disease.
It is also assumed that new cases are generated through homogeneous mixing, yielding the
mass-action incidence term βI S, or the standard-incidence term βI S/N [3, Ch. 11]. Note
 Analytical solutions and parameter estimation of the SIR epidemic model 165

that the dimensionality of β in the two cases is different. In the former case it is [T ]−1 [N]−1 ,
while in the latter it is [T ]−1 .
Model parameters can be interpreted as follows. A disease carrier infects on average β indi-
viduals per day, for an average time of 1/γ days. The β parameter is called disease transmis-
sion rate, while γ – recovery rate. The average number of infections arising from an infected
individual is then given by
β
R0 = ,
γ
called the basic reproduction number. Other useful quantities are the running reproduction
number
S(t)β
R  (t) =
Nγ
and the herd immunity threshold p = 1 − 1/R0 . Typical initial conditions, modeling an out-
break, are S(0) = S0 , I (0) = I0 , R(0) = 0 [2]. The simplified SIR model comprises a set of
three ODEs:
S(t)
Ṡ(t) = −β I (t) (10.1)
N
S(t)
I˙(t) = β I (t) − γ I (t) (10.2)
N
Ṙ(t) = γ I (t) (10.3)
The model assumes a constant overall population N = S + I + R.
The temporal course of the I-variable has a characteristic exponential rise, turnover, and
final slower exponential decline towards 0. The S-variable also has three phases – a slow de-
crease followed by rapid quasi-linear decline followed by slow exponential decline towards
S∞ , while the R-variable has a sigmoid shape. A phase diagram of the model is plotted in
Fig. 10.2.
The model can be re-parametrized using normalized variables as
ṡ = −si (10.4)
γ 1
i̇ = si − gi, g= = (10.5)
β R0
ṙ = gi, (10.6)
subject to normalization s + i + r = 1 and time rescaling τ = βt.
It is, in general, difficult to estimate the contact rate β, which depends on the particular dis-
ease and on social and behavioral factors [3]. On the other hand, different nonmedical inter-
ventions, such as wearing of personal protective equipment, contact restriction or quarantine,
166 Chapter 10

Figure 10.2: Phase plot of the (I, S) plane.

A – Case g=2, the trajectories starting at (4.6316, 1.9) and (6.16(6), 1.2) are indicated with dots.
This corresponds to the phase plot given in [10, Fig. 1A] for α = 0.1, β = 0.2. B – Case g=12, the
trajectories, starting at (13.81(81), 1.1) and (15.3(3), 0.6) are indicated with dots. This
corresponds to the phase plot given in [10, Fig. 1B] for α = 0.1, β = 1.2.

could offer some degree of control over β. The above formulation offers the advantage that
the dynamics imposed by the contact rate β(t), viewed as a function of time, can be absorbed
into the model by time rescaling
 t
τ (t) = β(u)du
−∞

which for a causal step function obviously recovers τ = βt. Note that τ (t) becomes a dimen-
sionless variable.

10.3 Second-order systems equivalent to SIR

The SIR model can be formulated also as several equivalent second-order nonlinear systems
for the different variables. The readers are directed to the original results in [7,10].

10.3.1 A second order differential equation for the i-variable

Recently, Kudryashov et al. gave an equivalent second order differential equation for the i-
variable [10].

Proposition 10.1. The SIR system is reducible to the nonlinear differential equation for the
i-variable:
i̇ 2
ï = −gi 2 − i i̇ + (10.7)
i
 Analytical solutions and parameter estimation of the SIR epidemic model 167

or the system
I2
I˙t = −gI 2 − I It + t (10.8)
I
I˙ = It (10.9)

Proof. From the conservation law ṡ = −i̇ − ṙ = −ig − i̇. Then from Eq. (10.4) s = g + i̇/i.
Differentiating Eq. (10.5) and substituting Eq. (10.4)
   
i̇ i̇ i̇ 2
ï = −g i̇ + i ṡ + i̇s = −g i̇ − i s + s i̇ = −g i̇ + i̇ g +
2
−i g+
2
= −gi 2 + − i i̇
i i i

The advantage of this formulation is that the phase space manifold is parametrized only by
a single parameter. Its phase portrait for values of g=2 and g=12 is shown in Fig. 10.3
 for
reference. The system admits elementary solutions i = i0 e−gt , r = r0 + i0 1 − e−gt which
correspond to s = 0. This can be verified by direct substitution into Eq. (10.7).

10.3.2 A second order differential equation for the s-variable

An equivalent second-order system for the s-variable was given in [7].
Proposition 10.2. The SIR system is reducible to the nonlinear differential equation for the
s-variable
ṡ 2
s̈ = + (s − g)ṡ
s
or the system
S2
S˙t = t + (S − g)St (10.10)
S
Ṡ = St (10.11)

Proof. Differentiating Eq. (10.4) gives

di s̈ + i ṡ ṡ 2 s̈
=− = 2−
dt s s s
and substituting i from Eq. (10.4). Substitution in Eq. (10.5) gives
ṡ 2 s̈ g ṡ
− = − ṡ
s2 s s
from where the result follows.
168 Chapter 10

Figure 10.3: Phase plane portrait of the differential systems equations (10.8)–(10.9).
A, B – (St , S) plot; C, D – (It , I ) A – Case g=2, the trajectories starting at (4.6316, −8.8) and
(6.1667, −7.4) are indicated with dots. C – Case g=2, the trajectories starting at (1.9, 5.0) and
(1.2, 5.0) are indicated with dots. B – Case g=12, the trajectories, starting at (13.8182, −15.2)
and (15.3333, −9.2) are indicated with dots. D – Case g=12, the trajectories, starting at (1.1, 2)
and (0.6, 2.0) are indicated with dots. This corresponds to the phase plot given in [10, Fig. 1B] for
α = 0.1, β = 1.2.

The following proposition is also formulated in [7].

Proposition 10.3. The SIR system is reducible to the nonlinear differential equation for the
r-variable
 r 
−
r̈ = g ṙ e g − 1

or the system
− Rg
Ṙt = gRt e − gRt (10.12)
Ṙ = Rt (10.13)

Proof. Differentiating Eq. (10.6) yields r̈ = g(si − gi) = gsi − g ṙ = −g ṡ − g ṙ. On the other
hand, by Eq. (10.16) we have r = −g log s + c. We use the fixed-point condition s0 = s(0) = g
 Analytical solutions and parameter estimation of the SIR epidemic model 169

so that s = ge−r/g+r0 /g . Since the system is autonomous we can translate the origin to −∞,
where r0 = 0 from where the result follows.

10.4 Indeterminate analytical solution

This section develops the analytical solution in an implicit form. Since there is a first integral
by construction the system can be reduced to any two of the three equations [3, Ch. 2]:
di g
= −1 + (10.14)
ds s
di s
= −1 (10.15)
dr g
dr g
=− (10.16)
ds s
From this formulation
s0 S0 β
Re = N = ≥1
g γ
must hold for the infection to propagate. Re is called the effective reproductive number, while
the basic reproduction number is R0 = Re N [11]. In order to solve the model we will consider
the two equations separately. Direct integration of Eq. (10.14) gives
i = −s + g log s + c (10.17)
where the indeterminate constant c can be determined from the initial conditions.

10.4.1 The s-variable

The s variable can be determined by substitution in Eq. (10.4), resulting in the autonomous
system
ṡ = −s (−s + g log s + c) (10.18)
which can be solved implicitly as

ds
=τ (10.19)
s (s − g log s − c)

The s variable can be represented explicitly in terms of the Lambert W function.

⎛ i−c ⎞
e g ⎠
s = −gW± ⎝− (10.20)
g
170 Chapter 10

where the signs denote the two different real-valued branches of the function. Note, that both
branches are of interest since the argument of the Lambert W function is negative.

10.4.2 The i-variable

Based on Eq. (10.20), Eq. (10.5) can be reduced to the first-order autonomous system
⎛ ⎛ i−c ⎞ ⎞
e ⎠g
i̇ = −gi ⎝W± ⎝− + 1⎠ (10.21)
g

valid for two disjoined domains on the real line. This autonomous ODE can be solved for the
rescaled time variable τ as

di
−   i−c   = gτ (10.22)
e g
i W± − g +1

10.4.3 The r-variable

Finally, the r variable can also be conveniently expressed in terms of i. For this purpose we
solve the differential equation
dr g −1
= =  i−c 
di s −g g
1 + W± − e g

Therefore,
⎛ ⎛ i−c
⎞⎞
e g
r = c1 − g log ⎝−gW± ⎝− ⎠⎠ = c1 − g log s
g

by Proposition 10.5. On the other hand,

⎛ ⎛ ⎞⎞ ⎛ ⎛ ⎞⎞
e
i−c
g  i−c  e
i−c
g
g log ⎝−gW ⎝− ⎠⎠ = g ⎝log e g − W ⎝− ⎠⎠ =
g g
⎛ i−c
⎞
e g
i − c − gW ⎝− ⎠=s +i −c
g
 Analytical solutions and parameter estimation of the SIR epidemic model 171

So that
⎛ i−c
⎞
e g
r = gW ⎝− ⎠ − i + c + c1 (10.23)
g

Since (s0 , 0, 0) is a stable point it follows that

−c 
e g
c + c1 = −gW− − (10.24)
g
Therefore,
⎛ i−c
⎞ −c 
e g e g
r = gW ⎝− ⎠ − gW− − −i (10.25)
g g

Proposition 10.4. i(t) attains a global maximum im = c − g log g − g.

Proof. We use a parametrization for which i(0) = im . Then

  i−im  
−1
i̇(τ ) = −gi W± −e g +1

It follows that i̇(0) = 0, i(0) = im so im is an extremum. In the most elementary way since
W(z) should be real-valued then
i−im
g −1
i − im
−e ≥ −1/e =⇒ ≤0
g
Hence, i ≤ im .

10.5 Inverse parametric solution

The implicit solution Eq. (10.22) can be computed as the definite integral
i
dy
gτ (i) = −   y−c   (10.26)
e g
c+g log g−g y W± − g +1

which requires computation of the Lambert W function on every integration step. However,
this does not seem to be efficient. Alternatively the solution can be computed more efficiently
by change of variables by Proposition 10.6:
i−c

 −gW± − e
g
g dy
gτ (i) = −g (10.27)
g y (g log y − y + c)
172 Chapter 10

On the other hand, from Eq. (10.19) it follows that

 s
dy
τ (s) = − (10.28)
g y (g log y − y + c)
 −c/g   
where the domain of s is [−gW+ −e /g , −gW− −e−c/g /g ].
The τ -variable can also be expressed as function of r. We differentiate Eq. (10.23) by s. From
where
    r−c1
dr 1  e g dr
−gt = dt = dr  =
ds s (s − g log s − c) s=e(c1 −r)/g (r − (c1 + c) ) e
r−c1
g +1
 
where c1 = −gW− −e−c/g /g − c by Eq. (10.24). Therefore,
r y
e g dy  
gτ (r) = − y , q = −gW− −e−c/g /g (10.29)
q−c−g log g
(y − q) e + q
g

for the c-parametrization and the domain of r is

    
0, gW+ −e−c/g /g − gW− −e−c/g /g

Therefore, the following theorem can be formulated.

Theorem 10.1 (Inverse parametric solution). The inverse parametric solution of the SIR
model is given by the integrals
 s
dy
τ (s) = −
g y (g log y − y + c)
i−c

 −gW± − e
g i
g dy dy
gτ (i) = −g =−    
g y (g log y − y + c) y−c
g
c+g log g−g y W± − e
g +1
r y
e g dy  
gτ (r) = − y , q = −gW− −e−c/g /g
q−c−g log g
(y − q) e + q g

The solutions are plotted in Fig. 10.4. A similar solution for τ (r) was obtained also in [12].
Interestingly, the authors derive an infinite series for τ (r) in terms of upper incomplete
Gamma functions.
 Analytical solutions and parameter estimation of the SIR epidemic model 173

Figure 10.4: Parametric plots of the SIR model.

A – i-variable from equation im -parametrization with parameters g= 1, im = 6.5; B – s and
r-variables with the same parameters. C – i-variable from equation im -parametrization with
parameters g=2.0, im = 6.5; D – s and r-variables with the same parameters. The negative branch is
on the negative axis.

10.5.1 Peak value parametrization

The upper terminal of integration can be determined by the requirement for the real-
valuedness of i. This value of i is denoted as im ; that is
⎛ i −c ⎞
m
e g
W± ⎝− ⎠ = −1
g

Therefore,

im = c + g log g − g (10.30)

The peak-value parametrization is supported by Proposition 10.4.

174 Chapter 10
   z−im  
−1
If we consider formally the phase space z × y = −gz W± −e g + 1 the following
 z−im 
−1
argument allows for the correct branch identification. For i → −∞ W− −e g → −∞
 z−im 
−1
so y < 0; while W+ −e g → 0+ so y > 0. Therefore, if we move the origin as t (0) =
im then conveniently
 i dz
−   z−im   = gτ, τ >0 (10.31)
im z W+ −e g −1 +1
 i dz
−   z−im   = gτ, τ ≤0 (10.32)
im z W− −e g −1 +1

Furthermore, the recovered population under this parametrization is

 i−im   −im 
−1 −1
r = gW± −e g − gW− −e g −i (10.33)

under the same choice of origin.

10.5.2 Initial value parametrization

As customarily accepted, the SIR model can be recast as an initial value problem. In this case,
the indeterminate constant c can be eliminated using the initial condition

i0 = −s0 + g log s0 + c

Therefore,
s
i = i0 + s0 − s + g log s/s0 = 1 − s + g log (10.34)
1 − i0
For this case, the following autonomous differential equation can be formulated:
   
1 − i0 i−1
i̇ = −gi W± − e g +1 (10.35)
g

This can be solved implicitly by separation of variables as

 i dz
−   z−1   = gτ, τ ≤ τm (10.36)
i0 z W− − 1−i
g e
0 g +1
 Analytical solutions and parameter estimation of the SIR epidemic model 175
 i
dz
−   z−1   = gτ, τ > τm (10.37)
i0 z W+ − 1−i0 e g + 1
g

It is noteworthy that the time to the peak of infections τm can be calculated as [9]:
g/s0
log
du
τm =
s0 eu − gu − (s0 + i0 )
0

The result follows by considering the autonomous system (10.14) and fixing the upper termi-
nal of integration s = g. However, by Proposition 10.7 this definite integral can be evaluated
only numerically.

10.6 Analysis of the incidence variable

The incidence i-function of the SIR model appears to be an interesting object of study on its
own. It was demonstrated that the function is nonelementary in the sense of the definition
given below [9].
Definition 10.1. An elementary function is defined as a function built from a finite number
of combinations and compositions of algebraic, exponential, and logarithm functions under
algebraic operations (+, −, ., /)

Allowing for the underlying field to be complex numbers – C, trigonometric functions be-
come elementary as well.
Definition 10.2 (Liouvillian function).
  We say that f (x) is a Liouvillian function if it lies in
some Liouvillian extension of C(x), for some constant field C.

As a first point we establish the nonelementary character of the integral in Eq. (10.22). The
necessary introduction to the theory of differential fields is given in Appendix 10.B. From the
work of Liouville it is known that a function of the form F (x) = f (x)eq(x) , where f , q are
elementary functions, has an elementary antiderivative of the form [13]
 
F (x)dx = f eq dx = heq

for some elementary function h(x) [14]. Therefore, differentiating we obtain

f eq = h eq + hq  eq

so that if eq = 0 h + hq  = f holds. The claim can be strengthened to demand that h be alge-

braic for algebraic f and q (see Theorem 10.1).
176 Chapter 10

Theorem 10.2. The integrals


dξ
I± (ξ ) =   ξ −c  
g
ξ W± − e g +1

are not Liouvillian.

Proof. We use im parametrization. Let c = im + g − g log g. The proof proceeds by change of

variables – first ξ = g log y − yg + g + im ; followed by z = log ((g log y + im + g)/g).
 
dξ y−1
I=    =     dy
ξ −im
g −1
y (g log(y) − gy + im + g) W −ye−y + 1
ξ W± −e +1
  z+ igm +1
dy 1 e
=− = dz
y(g log(y) − gy + im + g) g ee − e
z z+ igm +1

 
since W± −ye−y = −y. The last integral has the form

Aez
dz
ee − Aez
z

which allows for the application of the Liouville theorem in the form of Corollary 10.1. We
can identify

A im
g +1
f ez dx = hez , f (z) = ez , A = e
e − Aez
so that
A
= h (z) + h(z)
− Aez
z
ee
for some unknown algebraic h(z). Since the left-hand side of the equation is transcendental in
z so is the right-hand side, which is a contradiction. Therefore, the integrand has no elemen-
tary antiderivative.

The result of the last paragraph leaves the question about the form of i(τ ) somehow wanting.
Here we demonstrate a functional equation for the i variable exhibiting its non-Liouvillian
character.
 Analytical solutions and parameter estimation of the SIR epidemic model 177

Ritt established that if an elementary function has an elementary inverse it is a sequential

composition of algebraic and transcendental functions [15]. Prelle and Singer proved in
Corollary 3 that if the autonomous system y  = f (y) has an elementary first integral then

dy
g(y) =
f (y)
is also elementary [16]. This presents a direct way of proving that the incidence function
is nonelementary by virtue of Theorem 10.2 but leaves open the question about its non-
Liouvillian character. In order to clarify the form of i(τ ) we make use of the integral I (ξ ).
On the first place, the following identity holds for its kernel:
 ξ −c 
g
W± − e g
1 1
  ξ −c  = −   ξ −c  
e g ξ e g
ξ W± − g +1 ξ W± − g +1

Therefore,
 ξ −c 
g
 W± − g e
dξ
I± (ξ ) = log ξ −   ξ −c  =
e g
ξ W± − g +1
   g log z−z  

 g
dz z − 1 W± − e g
g 


= log ξ +   g log z−z   =

+ 1 
g
(g log z − z + c) W± − e g ξ −c 
g
z=−W± − e g
⎛ ⎛ ⎞⎞
  ξ −c
dz  e g
log ξ +   = log ξ + L ⎝−W± ⎝− ⎠⎠
g log z − z + c 
ξ −c
z=−W± − e g
g g

where we have used the defining identity for the Lambert W function and defined the auxil-
iary function up to a constant as the indefinite integral

dz
L(z) = (10.38)
g log z − z + c
which is a Liouvillian extension nonelementary integral by Proposition 10.8. Exponentiating
and using the notation L(z) it can be seen that
ξ −c 
g
L −W± − e g
exp (I± (ξ )) = ξ e = e−gτ
178 Chapter 10

However, ξ = i(τ ); therefore

i(τ )−c 
g
−gτ −L −W± − e g
i(τ ) = e (10.39)

or
  
i(τ )−c
−gτ −L − ± g −log g
i(τ ) = e , ± (z) = (z ± j π) (10.40)

from which we can infer the non-Liouvillian character of i(τ ) from the form of the equation.
This can be done as follows. For any given τ the RHS is a composition of a Liouvillian func-
tion with a non-Liouvillian one, i.e. the Wright function being not a Liouvillian one [17]. It
can be seen that the only scenario, where i(τ ) is elementary is when the second term, con-
taining the L function vanishes. This is exactly the scenario, when s0 = 0 as indicated by
Eq. (10.7), for in this case i̇ = −gi and i is exponential in τ . To remove the ambiguity in the
definition of L(z) we observe that z = s/g. Therefore,
 z
du
L(z) := − log im
1 g log u/g − u + g + im

so that L(1) = − log im ensuring the identity at τ = 0. Furthermore, it can be claimed that

Theorem 10.3. The incidence function i(t), defined by the differential equation (10.21), is not
Liouvillian.

10.7 Asymptotic analysis of the SIR model

In their seminal publication, Kermack and McKendrick develop asymptotics for r under the
assumption that r/g is small. Recently, Barlow and Weinstein [8] derived Padé approximation
scheme for the s-variable. The approximant for the s-variable by construction, matches the
correct t → ∞ behavior and whose expansion about t = 0 is exact to nth -order.
In the previous publication [9], a double exponential form of the asymptotic for the i-variable
was given, however without performing comprehensive study.
With the hindsight of the functional equation (10.39) a different scheme is more appropriate.
On the first place, the functional equation immediately reveals the long-time asymptotic be-
havior of i.
  
i(τ )−c
−gτ −L − + g −log g
lim i(τ ) = lim e =0
t→∞ τ →∞
 Analytical solutions and parameter estimation of the SIR epidemic model 179

since i(τ ) is monotonously decreasing for τm > 0. Therefore, it follows that

  
−c
−gτ −L − g −log g
i(τ ) ≈ e +
= ke−gτ

for large τ .
Since the solution is Lipshitz everywhere in R one can make use of the Banach Fixed Point
Theorem. Notably, one can use the nonlinear approximation scheme of Daftardar-Gejji-Jafari
for solving the equivalent integral equations [18]. If we treat the equations of the SIR model
as independent we can formally solve Eq. (10.4) as

s(τ ) = s0 e− i(t)dτ

On the other hand, Eq. (10.5) can be transformed formally as

  d (egτ t)
i̇ + gi = si ⇐⇒ e−gτ d egτ i = si dτ ⇐⇒ = s dτ
egτ i
Therefore, formally,

i(τ ) = k2 e−gτ + s(τ )dτ

Starting from the 0th order approximation is i (0) ≈ i0 , it follows that

s (0) ≈ s0 e−i0 τ

This approximation is valid around the fixed point s ≈ g, which we can take as an initial con-
dition s0 = g since then di/ds = 0. However, this corresponds to the peak-value parametriza-
tion so i0 = im , i  (0) = 0. From this, the 1st order approximation for the i-variable becomes
 
g
1−e−im t −gt
i (1) = im e im (10.41)

A second iteration of the loop results in a nonelementary γ integral. Let


g  
−im τ −gτ
J := e im 1−e dτ

Then, by change of variables

 
  g g −im τ
g g gy  im , im e
J =− y im −1 e im − im dy  =
y=e−im τ
 g
g im
im
180 Chapter 10

where is the upper incomplete Euler’s gamma function. Therefore,

 
g g −im τ
,
im im e
r (1) = gJ = k3  g
g im
im

Matching the long-term limit results in the equation

 
g g e−im t
     im , im
− im −1 − im −1
r (1) = g W+ −e g − W− −e g   (10.42)
g
im

which can be used also for fitting purposes. Then

⎛  ⎞
g g −im τ
⎜ im , im e ⎟
s (2) = k2 e−J = k2 exp ⎝−  g ⎠
g im
im

Matching the initial condition gives

    e a  g
s (2) = g exp (a, a) − a, a e−im t , a= (10.43)
a im
Finally, following the same procedure for the i-variable we obtain
  τ   a   e a
− a,ae−im z ae
i = im exp g e
(2) q
e dz − gτ , q = (a, a) (10.44)
0 a
g
where a = im . The asymptotics of the i-variable are shown in Fig. 10.5.

10.8 Numerical approximation

The i-function can be efficiently approximated by the Newton’s method [9]. The Newton iter-
ation scheme is given as follows for the c-parametrization:
⎛ ⎛ i −c ⎞ ⎞⎛ in −c  ⎞
n  e g
−gW± − g
e g ⎠ ⎜ dy ⎟
in+1 = in + gin ⎝W± ⎝− + 1⎠ ⎝t − ⎠
g g y (g log y − y + c)

(see Theorem 10.1). This form has the advantage of requiring only 1 Lambert W function
evaluation per iteration.
A point of attention here is the choice of the initial value for the iteration scheme. Accord-
ing to the presented analysis, a suitable elementary initializing function is i 1 (t) given by
Eq. (10.41).
 Analytical solutions and parameter estimation of the SIR epidemic model 181

Figure 10.5: Asymptotic solutions i(τ ).

Asymptotic solutions compared to parametric plots of (τ (i), i) parametrized by im = 3.7283 and
g = 2.0. Legends: LB denotes the left branch involving W− (x), RB denotes the right branch
involving W+ (x), Eq. (10.27). A denotes the asymptotic solution. A – the asymptotic is computed
from Eq. (10.41); B – the asymptotic is computed from Eq. (10.44). Plots were produced using the
quad_qags Maxima numerical integration command.

Table 10.1: Incidence parameters.

g R0 T [days] im fit
0.5493 1.8206 6.4849 296.2197 asymptotic
0.4710 2.1233 6.3609 303.9163 sir

10.9 Cast study I: application to influenza A

The influenza data are tabulated in [4]. The task here is to fit the influenza incidence in a
boarding school. The fitting equation is given by

It ∼ N · i (t − T |g, im )

where It is the observed incidence. The fitting functionality was implemented MATLAB® .
Quadratures were estimated by the default MATLAB integration algorithms. The parametric
fitting was conducted using least-squares constrained optimization algorithm. A least square
fit obtained R 2 = 53.1163 for the asymptotic fit and R 2 = 62.9415 for the full model. Fitted
parameters are presented in Table 10.1. The plots are presented in Fig. 10.6. The fitting pro-
duced peak estimate of 296.2197 for the asymptotic equation and 303.9163 for the full model
in excellent agreement with the raw data (see Table 10.1).
182 Chapter 10

Figure 10.6: SIR model fitting of influenza outbreak.

A – estimate of the i variable; B – estimate of the r variable compared with the cumulative number
of cases.

10.10 Cast study II: application to COVID-19

The second case study uses COVID datasets were downloaded from the European Centre for
Disease Prevention and Control (ECDC) website: https://opendata.ecdc.europa.eu/covid19/
casedistribution/csv. The data were imported in the SQLite https://www.sqlite.org database,
filtered by country and transferred to MATLAB for parametric fitting using native routines.
Quadratures were estimated by the default MATLAB integration algorithms. Estimated pa-
rameter values were stored in the same database. The processing is described in [9], For
numerical stability reasons the time variable was rescaled by a factor of 10. The estimate rep-
resented in Fig. 10.8 has been published in [9]. The asymptotic approach is exemplified with
data from Italy (Fig. 10.7).

10.11 Discussion and conclusions

The present chapter demonstrates that the SIR model could be solved completely. Recent
developments have exhibited the solutions of the model in terms of special functions. As
anticipated in previous literature, these functions are nonelementary and in particular the inci-
dence i-function is non-Liouvillian. On the other hand, the asymptotic solutions of the model
could be readily established in terms of iterated exponents or incomplete gamma functions.
In conclusion, the SIR model has proven to be valuable in modeling epidemic outbreaks. In
many situations it has shown excellent agreement with the historical data [3,4] as well as
with the data coming from the present COVID-19 pandemics [9,8,10]. On the other hand, it
 Analytical solutions and parameter estimation of the SIR epidemic model 183

Figure 10.7: Parametric fitting of the first and second waves in Italy.
A, B – case fatality fitting; C, D – incidence fitting; ‘asymp’ refers to parametric fit using the
asymptotic formula Eq. (10.41), ‘sir’ refers to fitting the i-variable computed by numerical inversion
using the parameters estimated by Eq. (10.41). Note the pronounced weekly variation of the
reported numbers.

should be remembered that any model is only a cartoon of reality and it is always important
to challenge model assumptions with the available facts on the ground. A distinction should
be always made between the mathematical consistency or even beauty of a certain model and
its empirical applicability In particular, predictions based on empirically false models may be
harmful, and it is essential to distinguish between assumptions that simplify but do not alter
the predicted effects substantially, and wrong assumptions, which are not supported by the
underlying ecology or social dynamics.
184 Chapter 10

Figure 10.8: Case fatality model for Belgium.

A – The parametric fit of the case fatality data; B – Cumulative deaths compared to the estimate
from the r-variable. The origin corresponds to 1st January 2020.

Appendix 10.A The Lambert W function and related integrals

The Lambert W function can be defined implicitly by the equation

W (z)eW (z) = z, z∈C

We observe that by Lemma 10.1 W (z) is transcendental. Furthermore, the Lambert function
obeys the differential equation for x = −1

e−W (x)
W  (x) =
1 + W (x)
The W function is nonelementary and in particular it is non-Liouvillian [17]. Its indefinite
integral is:

x
W (x)dx = xW (x) + −x
W (x)
The Lambert W is a multivalued function. Properties of the W function are given in [19].

Proposition 10.5.
⎛ ⎛ ⎞⎞
 y−c
dy e ⎠⎠ g
 y−c  = g log ⎝−gW ⎝− +C
e g g
1+W − g
 Analytical solutions and parameter estimation of the SIR epidemic model 185

Proof. We differentiate
y−c 
⎛ ⎛ ⎞⎞ y−c
−e
g
y−c g −W g
e g e 1
g ⎝log W ⎝− ⎠⎠ = −  y−c   y−c  =  i−c 
g e g e g 1 + W −ge g
gW − g 1+W − g

Proposition 10.6.
i−c

 i  g
−gW± − e g
dξ dy
  ξ −c   =g
g log g−g+c e g g y (g log y − y + c)
ξ W± − g +1

Proof. We use the change of variables ξ − c = g log y − y and then simplification by the
defining identity of the Lambert W function.

 i  B −1 g
dξ y
  ξ −c   =  g log y−y+c c 
  dy
g log g−g+c A −g
e g g
ξ W − g +1 (g log y − y + c) W − e
g +1
 
B g−y B dy
=   dy = g
A −
y gy log y−y 2 +cy
+ gy log y − y 2 + cy A y (g log y − y + c)
g

where

g log A − A + c = g log g − g + c, g log B − B + c = i

 i−c 
g
Therefore, A = g and B = −gW − e g .

The proof of Theorem 10.2 establishes also the character of two related integrals:

Proposition 10.7. The integral


dy
I=
y + c − ey
is not elementary.
186 Chapter 10

Proof. By change of variables y = ex

 
dy dx
I= =−
y (log y − y + c) e −x −c
x

Proposition 10.8. L(z) is a nonelementary integral.

Proof. This can be established by an argument using change of variables as

  
dz eu du 
L(z) = = 
g log z − z + c −eu + gu + c z=eu
further applying the same reasoning as in the proof of Theorem 10.2.

Appendix 10.B Differential fields

Definition 10.3. Denote by C (x, ci , θi ) the complex-valued ring, generated by the finite set of
rational functions {θi }ni and constants {ci }ni .
Definition 10.4. An element θ is called algebraic if P (x, θ ) = 0 for some polynomial

P (x, t) = t m + am−1 t m−1 + . . . + a0 ,

where ai can be also rational functions of x, or else it is called transcendental.

Lemma 10.1 (Composition lemma). Denote by a and t the algebraic or transcendental ele-
mentary functions, respectively. The following composition rules hold

a ◦ a = a, t ◦ a = t, a◦t =t

Proof. The case a ◦ a when a(x) is a polynomial is trivial. Suppose that a and b are both alge-
braic:
P (x, a) = 0, Q(x, b) = 0
Without loss of generality suppose that ai are polynomial. Formally, b = f¯k (x) for any branch
k with f¯ algebraic since it is a root of a polynomial, where the bar denotes the inverse func-
tion in order to avoid confusion with exponentiation. Therefore,

a ◦ b = bm + am−1 bm−1 + . . . + a0 = f¯km (x) + am−1 f¯km−1 (x) + . . . + a0

is algebraic since it is computed by a finite sequence of algebraic operations.

 Analytical solutions and parameter estimation of the SIR epidemic model 187

Suppose that t = exp(x). Then exp(a) is not algebraic, hence it is transcendental. P (x, ex ) =
exm + am−1 exm−x + . . . + a0 is exponential.

Suppose that t = log(x). Then log(a) is not algebraic, hence it is transcendental.

P (x, log(x)) = logm x + am−1 logm−1 x + . . . + a0 is not algebraic, hence it is transcen-
dental.

In what follows is assumed that the differential field is of characteristic zero and has an alge-
braically closed field of constants. An element y of a differential field is said to be an expo-
nential of an element A if y  = Ay, an exponential of an integral of an element A if y  = Ay;
logarithm of an element A if y  = A /A, and an integral of an element A if y  = A.

The next definition is due to [17].

Definition 10.5. Let (k, ≡ d/dx) be a differential field of characteristic 0. A differential ex-
tension (K, ≡ d/dx) of k is called Liouvillian over k if there are θ1 , . . . , θn ∈ K, such that
K = C(x, θ1 , . . . , θn ) and for all i, at least one of the following

1. θi is algebraic over k(θ1 , . . . , θn−1 )

2. θi ∈ k(θ1 , . . . , θn−1 )
3. θi /θi ∈ k(θ1 , . . . , θn−1 )

holds. The constant subfield C(K) of K is defined to be the set of c in K, such that c = 0.

The next theorem is due to [14].

Theorem 10.4. If K is an elementary field, then it is closed under differentiation.

An elementary integrability theorem due to Conrad [14].

Theorem 10.5 (Rational Liouville criterion). For f, g ∈ C(x) with f and g nonconstant the
function f (x)eg (x) can be integrated in elementary terms if and only if there exists a rational
function h ∈ C(x) such that h + g  h = f .

The last result can be extended to algebraic functions as follows.

Corollary 10.1 (Algebraic Liouville criterion). For f (x), g(x) algebraic and nonconstant,
the function f (x)eg(x) can be integrated in elementary terms if and only if there exists an al-
gebraic function h(x), for which h + g  h = f .
188 Chapter 10

Proof. Suppose that f and g are arbitrary elementary algebraic functions. Denote the primi-
tive of f by capital F in the juxtaposition f ÷ F . The integral can be integrated by parts
    
I = f (x)e dx = e (x)dF = F (x)e
g(x) g g(x)
− F (x) eg(x) dx

Therefore,

 
f (x) + F (x)g  (x) eg(x) dx = F (x)eg(x)

We observe that g  (x) is elementary by Theorem 10.4. The L.H.S. has the form f eg and since
f (x) + F (x)g  (x) is elementary we can identify

h ≡ F, f1 ≡ f + F g  = h + hg 
 
so that h + hg  eg = (heg ) and the claim follows.

References
[1] F. Brauer, Mathematical epidemiology: past, present, and future, Infectious Disease Modelling 2 (2) (2017
may) 113–127.
[2] W.O. Kermack, A.G. McKendrick, A contribution to the mathematical theory of epidemics, Proceedings
of the Royal Society of London. Series A, Containing Papers of a Mathematical and Physical Character
115 (772) (1927 aug) 700–721.
[3] F. Brauer, P. van den Driessche, J. Wu (Eds.), Mathematical Epidemiology, Springer Berlin Heidelberg, 2008.
[4] M. Martcheva, An Introduction to Mathematical Epidemiology, Springer US, 2015.
[5] D. Champredon, J. Dushoff, D.JD. Earn, Equivalence of the Erlang-distributed SEIR epidemic model and the
renewal equation, SIAM Journal on Applied Mathematics 78 (6) (2018 jan) 3258–3278.
[6] H.S. Rodrigues, Application of SIR epidemiological model: newtrends, International Journal of Applied
Mathematics and Informatics 10 (2016) 92–97.
[7] T. Harko, F.S.N. Lobo, M.K. Mak, Exact analytical solutions of the Susceptible-Infected-Recovered (SIR)
epidemic model and of the SIR model with equal death and birth rates, Applied and Computational Mathe-
matics 236 (2014 jun) 184–194.
[8] N.S. Barlow, S.J. Weinstein, Accurate closed-form solution of the SIR epidemic model, Physica D. Nonlinear
Phenomena 408 (2020 jul) 132540.
[9] D. Prodanov, Analytical Parameter Estimation of the SIR Epidemic Model. Applications to the COVID-19
Pandemic, Entropy, Basel, Switzerland, 2020 Dec, p. 23.
[10] N.A. Kudryashov, M.A. Chmykhov, M. Vigdorowitsch, Analytical features of the SIR model and their appli-
cations to COVID-19, Applied Mathematical Modelling 90 (2021 feb) 466–473.
[11] H. Weiss, The SIR model and the foundations of Public Health, Materials Matemàtics 3 (2013) 1–17.
[12] E. Sadurní, G. Luna-Acosta, Exactly solvable SIR models, their extensions and their application to sensitive
pandemic forecasting, Nonlinear Dynamics 103 (3) (2021 feb) 2955–2971.
[13] M. Rosenlicht, On the explicit solvability of certain transcendental equations, Publications Mathématiques
de l’Institut des Hautes Études Scientifiques 36 (1969) 15–22, Available from http://www.numdam.org/item/
PMIHES_1969__36__15_0.
[14] B. Conard, Impossibility theorems for elementary integration, 2005 Academy Colloquium Series, Clay Math-
ematics Institute, 2005.
 Analytical solutions and parameter estimation of the SIR epidemic model 189

[15] J.F. Ritt, Elementary functions and their inverses, Transactions of the American Mathematical Society 27 (1)
(1925 jan) 68.
[16] M.J. Prelle, M.F. Singer, Elementary first integrals of differential equations, Transactions of the American
Mathematical Society 279 (1) (1983 jan) 215.
[17] M. Bronstein, R.M. Corless, J.H. Davenport, D.J. Jeffrey, Algebraic properties of the Lambert W function
from a result of Rosenlicht and of Liouville, Integral Transforms and Special Functions 19 (10) (2008 oct)
709–712.
[18] V. Daftardar-Gejji, H. Jafari, An iterative method for solving nonlinear functional equations, Journal of Math-
ematical Analysis and Applications 316 (2) (2006 apr) 753–763.
[19] R.M. Corless, G.H. Gonnet, D.E.G. Hare, D.J. Jeffrey, D.E. Knuth, On the Lambert W function, Advances in
Computational Mathematics 5 (1) (1996 dec) 329–359.