www.nature.

com/scientificreports

OPEN Fractional SIR epidemiological

models
Amirhossein Taghvaei1, Tryphon T. Georgiou1*, Larry Norton2 & Allen Tannenbaum3

The purpose of this work is to make a case for epidemiological models with fractional exponent
in the contribution of sub-populations to the incidence rate. More specifically, we question the
standard assumption in the literature on epidemiological models, where the incidence rate dictating
propagation of infections is taken to be proportional to the product between the infected and
susceptible sub-populations; a model that relies on strong mixing between the two groups and
widespread contact between members of the groups. We contend, that contact between infected and
susceptible individuals, especially during the early phases of an epidemic, takes place over a (possibly
diffused) boundary between the respective sub-populations. As a result, the rate of transmission
depends on the product of fractional powers instead. The intuition relies on the fact that infection
grows in geographically concentrated cells, in contrast to the standard product model that relies
on complete mixing of the susceptible to infected sub-populations. We validate the hypothesis of
fractional exponents (1) by numerical simulation for disease propagation in graphs imposing a local
structure to allowed disease transmissions and (2) by fitting the model to the JHU CSSE COVID-19
Data for the period Jan-22-20 to April-30-20, for the countries of Italy, Germany, France, and Spain.

The classical SIR (Susceptible, Infectious, Recovered) model of infectious disease dynamics, and all subsequent
multi-compartmental derivative models, are based on a model for the incidence rate that is taken almost uni-
versally in the form
r(t) = βI(t)S(t), (1)
where I(t), S(t) represent the size of infected and susceptible sub-populations. The proportionality factor β is typi-
cally determined on a case-by-case basis. Thus, if R(t) represents the size of the recovered population, assuming
that all individuals undergo full recovery and thereby the total population S(t) + I(t) + R(t) remains constant,
the most basic model for transmissions is in the form of the following system of equations, known as SIR model,
d
S(t) = −r(t) + ηR(t),
dt
d
I(t) = r(t) − αI(t),
dt
d
R(t) = αI(t) − ηR(t),
dt
where α is the recovery rate (with time constant of recovery τ := 1/α) and η a parameter regulating the rate at
which immunity is lost over time. ­See1,18 for all the details about the SIR model together with an extensive list
of references. Multi-compartmental models that include infected but asymptomatic individuals, deceased, or
exposed, have also been considered. However, throughout, the basic feedback that drives the infection, r(t), is
invariably as in (1).
In departure from this well-studied SIR paradigm, we propose a fractional SIR (fSIR) model with rate
r(t) = βI(t)γ S(t)κ , (2)
where one or, possibly, both sub-populations are scaled by exponents that are typically less than 1. The justifica-
tion for such a model stems from the fact that, at least during the initial phase of an epidemic, infection propa-
gates outwards from infected cells to the general population. In such a scenario, where for instance S(t) ≫ I(t)
(much greater), the boundary of infected cells which would roughly account for most new infections, scales as a

1
Mechanical and Aerospace Engineering, University of Calfornia, Irvine, CA 92697, USA. 2Department of Medicine,
Memorial Sloan Kettering Cancer Center, New York, NY 10021, USA. 3Depts of Computer Science and Applied
Mathematics and Statistics, Stony Brook University, Stony Brook, NY 11794, USA. *email: [email protected]

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 1

Vol.:(0123456789)
 www.nature.com/scientificreports/

fractional power γ < 1 of the area of the cells, hence I(t)γ . In actuality, due to the diffusive nature of infection-
propagation amongst the general population, the exponent is expected to be larger than 1/2, as it would be in
the continuous limit when the boundary is a smooth curve. Moreover, at least in the early phases of an epidemic,
the exponent of S(t), which is significantly larger than I(t) may turn out to be negligible.
The idea of using fractional exponents in growth models has been motivated from Norton–Simons–Massagué
(NSM) model, a growth model of the form
dV (t)
= aV  (t) − bV (t), (3)
dt
with origins in the 1­ 950s3. This type of model was designed to describe the growth of biological organisms
employing certain energy principles. In the model, the parameters a and b quantify anabolism (growth) and
catabolism (death), respectively. Equation (3) may be interpreted as asserting that the net growth rate of an organ-
ism results from the balance of synthetic and degradative mechanisms. While the rate of the former process fol-
lows a law of allometry (i.e., the rate is proportional to the volume V(t) via a power function), the rate of the latter
process scales linearly with V(t). It is important to note that the two special cases of (3), (1) power law b = 0, and
(2) second-type growth  = 2/3 have already been successfully applied to describe tumor ­growth10,27. The general
case, 0 <  < 1, was introduced i­ n23 to explain the self-seeding hypothesis. Moreover, an important geometri-
cal interpretation was provided i­ n2,22. In these works, the authors relate the exponent  = d/3 to the fractional
Hausdorff dimension of the proliferative tissue, where d denotes the fractal dimension of the tissue. Moreover,
the model (3) has been derived mechanistically by linking tumor growth to metabolic rate and ­vascularization12.
Thus, in a similar spirit to the Norton–Simons–Massagué (NSM) model, herein, we recognize the geometric
constraints imposed on disease propagation by the locality of transmissions around infectious cells. To this
end, we seek to explain the origin of fractional power in (2) by (1) numerical experiments, and (2) fitting such
models to data sets.
Specifically, with respect to (1), we postulate a discrete model where infection propagates over nodes of a
network. The network, representing individuals, is not in general planar, yet it is immersed in R2. While (physi-
cally) neighboring nodes may be densely connected, precluding the graph from being planar, the likelihood of
being connected depends on their physical distance. We observed that such models lead to fractional exponents
in the incidence rate, in agreement with (2).
While the basic mechanism of propagation of pathogens, through contacts (edges) of a social network fol-
lows the extensive foundational work on epidemics over ­networks5,13,19,20, Chapter 16, our proposed network
structure in which edges between nearby nodes are prevalent, renders a number of analytical techniques (such
as the renormalization methods used i­ n19) inapplicable—with graphs failing to display a tree-like structure as in
random networks. Thus, the verification that the postulated contact structure leads to fractional exponents is at
­ f14, may be adapted to reflect the
present empirical. It is anticipated that statistical analysis, e.g., along the lines o
particular structure in the type of networks we consider and lead to more rigorous analytical results.
Regarding (2), we have numerically studied data that is available i­ n7 on the recent COVID-19 epidemic. We
focused on the dynamics of transmission in four countries (Italy, Germany, France, and Spain) during the early
phases of the epidemic. For the specific datasets we considered, as we explained in the results, the exponent γ in
(2) ranges from about 0.6 to 0.8. Coincidentally, the values appear similar to empirically determined exponents
of the NSM model.
we should note that after the initial submission of this work, it came to our attention that fractional powers
in the incidence rate of SIR models have been considered i­ n17,24,29 and going further back t­ o26. The authors o ­ f17
postulated that a threshold level of viral concentration in the population may be needed before an epidemic
takes off, and that such a mechanism could account for nonlinear incidence rates, especially with exponents
γ > 1 that lead to rich dynamical behavior studied i­ n17. The authors ­in29 suggest an exponent slightly less than
1 (specifically 0.97) in their simulations, to model “the biases introduced by time discretization and the fact the
force of infection is disproportionately small at higher densities”. ­In24, the authors consider heterogeneity in the
transmission coefficient β and, for a particular initial distribution of β , obtain an effective fractional model with
exponents larger than 1. In contrast, in our work, the proposed mechanism suggests γ < 1. Moreover, we have
attempted to link the fractional exponents in the dynamics of the epidemic to the structure of the network and,
specifically, to the interface between infected and susceptible subpopulations.

Models of discrete transmissions

To provide insight and justification for our hypothesis on the validity of Eq. (2), we develop a discrete model
for direct transmissions between individuals consisting of nodes (individuals) on a graph that captures contacts
between them. In the present work, the graph is fixed, while in future work we plan to explore the possibility of
time-varying links between nodes as well as modeling control actions, such as social-distancing, so as to study
the effects of such mediation protocols.

Model: probabilistic SIR on a graph. Consider a simple undirected graph of size n with adjacency matrix
A, where Aij = 1 if node i and j are connected, and Aij = 0 otherwise . The graph is used to model the spread
of infection over a network of nodes representing individual people. Every node can be in one the three states
{S, I, R}. We use xi (t) ∈ {S, I, R} to represent the state of node i at time t ∈ {0, 1, 2, 3, . . .}. We consider the unit
of time to be one day. Here, {x1 (t), . . . , xn (t)} evolves, as a Markov chain on 3n states, according to the following
transition probabilities dictating transition at the node level for xi (t) to xi (t + 1),

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 2

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 1.  Structure of the two graph models considered in this paper. Nodes are connected to their 4 nearest
neighbors. The initially infected people are marked in orange.

Parameters Definition
α Recovery rate
β Transmission rate
η Rate of losing immunity
κ Fractional exponent of S
γ Fractional exponent of I
k Number of neighbours of each network node
d Maximal Manhattan distance between neighbours
m Number of additional random edges between randomly selected nodes

Table 1.  Notation for modeling parameters.

(I)
S → I w.p. 1 − (1 − β)Ni (t)

I → R w.p. α (4)
R → S w.p. η.
Here, β is the transmission rate, α is the recovery rate, and η is the susceptibility rate (quantifying loss of immunity
(I)
over time). The notation Ni (t) stands for the number of the neighbors of node i that are infected at time t, i.e.
n
Ni(I) (t) =

Aij 1{xj (t)=I} ,
j=1

where 1{·} = 1 when {·} holds and is 0 otherwise.

With regard to the structure of the graph, specifying contacts between individuals, we describe results con-
sidering the following options:

Two‑dimensional grid‑graph. We work out two rudimentary models where individuals (nodes) are placed on
a 2-dimensional grid (vertex set)
 
V := vij = (i, j) | i, j ∈ {0, 1, . . . , N − 1} .

We carry out experiments for two cases, where the edge set is defined by

(5)
 
E := (vij , vℓk ) | |i − ℓ| + |j − k| ≤ d, ∀i, j, l, k ∈ {0, . . . , N − 1} ,

with d ∈ {1, 2}. Thus, when d = 1, each node is connected to k = 4 nearest neighbors, while in the case where
d = 2 each node is connected to k = 8 nearest neighbors (von Neumann neighborhood with Manhattan distances
1 and 2, respectively). The graph structure for the case d = 1 is depicted in Fig. 1a. In either case, the infectious
model is simulated and the results discussed in the section on experiments.

Two‑dimensional random graph. We postulate a distribution of nodes on R2 according to a Gaussian mixture

model (Fig. 1b). Each node is connected to its 4 nearest neighbor. Analogous conclusions are drawn and dis-
cussed in the section on experiments as well.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 3

Vol.:(0123456789)
 www.nature.com/scientificreports/

Table 1 includes summary of notation for modeling parameters.

We experiment with a randomized initialization, where a number of small initial cells of infected individu-
als are sprinkled randomly with probability p0 inside the general population. A localized initialization, where a
specified initial collection of neighboring nodes are infected and the contagion begins from these nodes, gives
similar results.
We are interested in studying the dynamics of the subpopulations of susceptible, infected, and recovered
individuals, denoted by S(t), I(t), and R(t), respectively. These quantities are calculated from the state of the
individuals xi (t):
n

ξ(t) = 1{xi (t)=ξ } ,
i=1

for each ξ ∈ {S, I, R}.

We model the change in the number of infected individuals �I(t) := I(t) − I(t − 1) as a function of aggregate
quantities S(t), I(t), and R(t). The change is mainly due to two factors, (1) an increase due to transmission of
infection by contact between infected and susceptible individuals, denoted by �Itrans (t), and (2) a decrease due
to recovery of individuals, denoted by �Irecovery (t). Thus,
�I(t) = �Itrans (t) − �Irecovery (t). (6)
In terms of the state of the Markov chain, �Itrans (t) and Irecovery are given by
n

�Itrans (t) := 1{xi (t)=I,xi (t−1)=S} ,
i=1
n
�Irecovery (t) := 1{xi (t)=R,xi (t−1)=I} ,
i=1

It is expected that �Irecovery (t) ≈ αI(t) on average. However, the model for Itrans is not simple. We hypothesize
the fractional model
�Itrans ≈ cI(t)γ S(t)κ .
where c is a constant, and γ and κ are exponents. We test this hypothesis in the next section by carrying out Monte
Carlo simulations on a network model and fitting the data to the proposed fractional model.

Experiments
Two‑dimensional grid‑graph. Simulation results of infection spread on the two-dimensional grid-graph
with size 100 × 100, for d ∈ {1, 2} (and hence, each node connected to the k ∈ {4, 8} nearest nodes), were car-
ried out for the model (4) and are presented in Figs. 2, 3 and Figs. 4, 5, respectively. The documented experi-
ments were carried out for η = 0.01, initial infection probability p0 = 10−3, combination of parameters for
α = {0.05, 0.1} and β ∈ {0.2, 0.3}, and a time period of 50 days. In order to take into account the randomness
of the stochastic model, 100 Monte Carlo simulations were carried out. Each simulation involves independent
random initialization of the infected individuals, while the structure of the graph remains fixed.
The top panel in each figure depicts the number of susceptible, infected, and recovered individuals as a func-
tion of time, for a single realization of the stochastic model. The bottom panel (of the three panels) in each figure
depicts in 3D the number of newly infected population due to transmission, �Itrans (t), vs. the number of infected
people I(t), and the number of susceptible people S(t), with solid blue lines marking 100 independent simulations.
The center panel represents the two-dimensional projection of the data points on the (log(�Itrans (t)), log(I))
plane.
In order to capture the relationship between �Itrans (t) and I(t) and S(t), a parametric curve of the form
�Itrans (t) = cI(t)γ S(t)κ is fitted to the data-points obtained from each simulation. Thus, the parameters c, γ ,
and κ are obtained by a least-squares fit of the linear relation between respective logarithms (with zero entries
for �Itrans (t), I(t), S(t) being ignored),
log(�Itrans (t)) = log(c) + γ log(I(t)) + κ log(S(t)).
The linear regression implicitly assumes an additive Gaussian perturbation, as in (8), on the log-transformed
data. Such an additive perturbation of the logarithm is natural in population models as it ensures, here, that
�Itrans (t) remains positive and that �Itrans (t) is zero when either S(t) or I(t) is zero; for further discussion, ­see30.
Note that c, γ , κ are random and differ in each realization of the stochastic model. The mean and standard
deviation of γ , κ over 100 simulations are reported in Tables 2a,b for the cases d = 1 and d = 2, respectively.
Also, the curve corresponding to the mean values of the parameters is depicted in the second and third panels for
comparison with the simulation data points (in orange color). We observe that in all cases, the standard devia-
tion of the fitted exponent γ is small, around 0.04. However, the standard deviation over κ is large, in some cases
around 0.4. This is due to some extent to the fact that the simulations were carried out over short time window,
so that the change in S is not sufficiently significant to lead to a reliable exponent κ (i.e., the fractional change in
S is relatively small, hence there is an inherent insensitivity to the actual value of the exponent).

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 4

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 2.  Spread of infection on a two-dimensional grid, with connection to k = 4 nearest neighbors (i.e.,
d = 1 in Eq. (5)) and transmission rate β = 0.2 (the recovery rate α is fixed to either 0.05 or 0.1). The top panels
depict the number of susceptible, infected, and recovered individuals as a function of time over a range of 50
days, for a single realization of the stochastic model described in “Model: probabilistic SIR on a graph”. The
center and bottom panel depict (�Itrans (t), I(t)) and (�Itrans (t), I(t), S(t)) respectively, for 100 Monte-Carlo
simulations (with solid blue lines). The fractional model fit is illustrated with orange dashed curve. The fitting
procedure maximizes the log-likelihood (9) over the free parameters. The value of the fitted exponents are
shown in the legend.

As an indicator of the statistical significance of the fractional SIR model, as compared to the traditional model
with exponents equal to one, we evaluated the standard model selection Akaike Information Criterion (AIC).
The AIC score is defined as
AIC = 2(#parameters − max-log-likelihood), (7)
where the maximum likelihood is obtained by maximizing the probability of observing the data set for the
parameters of the model. The assumed fSIR model is
log(�Itrans (t)) = log(c) + γ log(I(t)) + κ log(S(t)) + ǫ(t), (8)
with ǫ(t) independent Gaussian random noise, for each t, having mean zero and unknown variance denoted by
σ 2. Because of the Gaussian assumption for the noise, the log-likelihood takes the following form
T
T  1  2
log-likelihood = − log(2πσ 2 ) − log(�Itrans (t)) − log(c) − γ log(I(t)) − κ log(S(t)) , (9)
2 2σ 2
t=1

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 5

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 3.  Spread of infection on a two-dimensional grid, with connection to k = 4 nearest neighbors (i.e.,
d = 1 in Eq. (5)) and transmission rate β = 0.3 (the recovery rate α is fixed to either 0.05 or 0.1). The top panels
depict the number of susceptible, infected, and recovered individuals as a function of time over a range of 50
days, for a single realization of the stochastic model described in “Model: probabilistic SIR on a graph”. The
center and bottom panel depict (�Itrans (t), I(t)) and (�Itrans (t), I(t), S(t)) respectively, for 100 Monte–Carlo
simulations (with solid blue lines). The fractional model fit is illustrated with orange dashed curve. The fitting
procedure maximizes the log-likelihood (9) over the free parameters. The value of the fitted exponents are
shown in the legend.

where T is the number of data-points. In our results, we report the log-likelihood normalized by the number of
data points, i.e. T1 log-likelihood.
For the purpose of comparison, we consider four different models: (1) our proposed fSIR model where γ
and κ are free, (2) γ is free and κ = 1, (3) γ = 1 and κ is free, and (4) γ = 1 and κ = 1. The number of unknown
parameters for the first model is 4, namely, c, γ , κ, σ . For the second and the third model, the number of unknown
parameters is 3, and for the last one it is 2. We computed the AIC scores and the normalized maximum log-
likelihood for all four models averaged over 100 Monte-Carlo simulations for the two-dimensional grid network
structure. The results for the setting (Manhattan distance) d = 1 and d = 2 are reported in Tables 3a,b, respec-
tively. The reported values suggest a preferance for the fSIR model over models with exponents equal to one; the
goodness of the fit more than compensates for the increased “complexity in the model” due to the additional
parameters. For illustration purposes, we compare the four models fit to the data in Fig. 6.
In order to study the effects of model parameters α and β on the fitted exponents γ and κ, we carried out addi-
tional experiments and reported the results in Fig. 7. Varying β does not affect the fitted exponent significantly,
however a change in α does. This is consistent with the exponent being impacted by the interface (boundary)
between the infected population and susceptible population; a change in β affects the speed that the interface
propagates. A change in α impacts the interface by changing the rate at which the population inside clusters
become susceptible again.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 6

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 4.  Spread of infection on a two-dimensional grid, with connection to k = 8 nearest neighbors (i.e.,
d = 2 in Eq. (5)) and transmission rate β = 0.2 (the recovery rate α is fixed to either 0.05 or 0.1). The top panels
depict the number of susceptible, infected, and recovered individuals as a function of time over a range of 50
days, for a single realization of the stochastic model described in “Model: probabilistic SIR on a graph”. The
center and bottom panel depict (�Itrans (t), I(t)) and (�Itrans (t), I(t), S(t)) respectively, for 100 Monte-Carlo
simulations (with solid blue lines). The fractional model fit is illustrated with orange dashed curve. The fitting
procedure maximizes the log-likelihood (9) over the free parameters. The value of the fitted exponents are
shown in the legend.

Two‑dimensional random graph. Simulation result on a two-dimensional random graph model were
carried out and documented in Fig. 8. The spatial distribution of 104 node-coordinates ((x, y)-coordinates, cf.
Fig. 1b) has been selected from the mixture of three Gaussians distributions on the node-coordinate plane,
           
1 0 2.0 0.0 1 1 1.44 0.0 1 5 0.64 0.0
N , + N , + N , .
3 0 0.0 0.5 3 5 0.0 0.5 3 1 0.0 1.44

Here, N(v, R) denotes a Gaussian distribution with mean v and covariance R. Initially, for the experiment docu-
mented in Fig. 8, the nodes are connected to their 4 nearest neighbors. The infection spread model is once again
the one specified in (4). The parameters selected for the results in Fig. 8 are β = 0.3 and η = 0.01, while we
compare the effect of α ∈ {0.05, 0.1}. The numerical result contains 100 independent simulations where each
simulation involves independent random construction of the graph according to the Gaussian mixture model,
and also random initialization of infected nodes. As before, the first layer of panels depicts the number of sus-
ceptible, infected, and recovered individuals as a function of time, for a single realization. The two-dimensional
projection of the (�I(t), I(t), S(t)) curve on the (log(�I(t)), log(I(t))) plane is depicted in the second layer of
panels and compared to the distribution of (�I(t), I(t)) point set, while the third layer of panels compares the
fit of curve to the (�I(t), I(t), S(t)) data set in a three-dimensional plot once again in logarithmic scales. The
mean and the standard deviation of the fitted exponents are reported in Table 2c. We carry out the model selec-
tion procedure, described for the two-dimensional grid, for the random graph and report the results in Table 3c.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 7

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 5.  Spread of infection on a two-dimensional grid, with connection to k = 8 nearest neighbors (i.e.,
d = 2 in Eq. (5)) and transmission rate β = 0.3 (the recovery rate α is fixed to either 0.05 or 0.1). The top panels
depict the number of susceptible, infected, and recovered individuals as a function of time over a range of 50
days, for a single realization of the stochastic model described in “Model: probabilistic SIR on a graph”. The
center and bottom panel depict (�Itrans (t), I(t)) and (�Itrans (t), I(t), S(t)) respectively, for 100 Monte-Carlo
simulations (with solid blue lines). The fractional model fit is illustrated with orange dashed curve. The fitting
procedure maximizes the log-likelihood (9) over the free parameters. The value of the fitted exponents are
shown in the legend.

Two additional plots that help explain the dependence of γ , κ on the geometry of the network are shown in
Figs. 9 and 10. Specifically, Fig. 9 shows the result of a single Monte Carlo simulation whereas Fig. 10 displays
statistics from 100 simulations. In either figure, the left panel shows the dependence of γ (and in addition, κ in
Fig. 10) as the number k of nearest neighbors to each node varies in {4, 5, 6, 7, 8, 9, 10}. The right panel shows the
dependence of the same quantities as additional edges between randomly selected pairs of nodes are introduced.
In the figure, m ∈ {0, 100, 200, 300, 400, 500, 600, 700} designates the number of additional random edges in this
graph of 104 nodes.
It appears, as hypothesized, that the exponent γ is affected by the local structure of the graph. If vertices are
only connected to nearest neighbors, then γ is small. When random connections are introduced (thereby reduc-
ing the effective diameter of the graph, i.e., “small world” effect), γ increases to 1. Note that only changing the
number of connections to a nearest neighbor does not affect the exponent as shown in Fig. 10a.

COVID‑19 data‑set. We utilized the COVID-19 data repository by the center for systems science and engi-
neering (CSSE) at Johns Hopkins ­University7,9. We study the relationship between the number of newly infected
�Itrans (t) and the total number of infected individuals I(t) for the duration January 22 to September 13, in Italy,
Germany, France, and Spain. The number of newly infected individuals is available as the number of newly
confirmed cases each day. The total number of infected individuals (active cases) is calculated by subtracting the

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 8

Vol:.(1234567890)
www.nature.com/scientificreports/

Params. β = 0.2, α = 0.05 β = 0.2, α = 0.1 β = 0.3, α = 0.05 β = 0.3, α = 0.1

(a) Two-dimensional grid model, with connection to k = 4 nearest neighbors (i.e., d = 1 in
Eq. (5)).
γ 0.67 ± 0.04 0.70 ± 0.05 0.66 ± 0.04 0.70 ± 0.04
κ 0.56 ± 0.43 0.06 ± 0.42 0.66 ± 0.29 0.40 ± 0.24
Cor(γ , κ) 0.73 0.65 0.41 0.61
(b) Two-dimensional grid model, with connection to k = 8 nearest neighbors (i.e., d = 2 in
Eq. (5)).
γ 0.68 ± 0.04 0.76 ± 0.05 0.66 ± 0.06 0.75 ± 0.07
κ 0.62 ± 0.09 0.62 ± 0.08 0.58 ± 0.05 0.62 ± 0.07
Cor(γ , κ) 0.15 0.57 0.64 0.76
Params. β = 0.3, α = 0.05 β = 0.3, α = 0.1
(c) Two-dimensional Gaussian mixture random graph model, with connection to k = 4 nearest
neighbors.
γ 0.62 ± 0.03 0.66 ± 0.04
κ 0.43 ± 0.23 0.22 ± 0.30
Cor(γ , κ) 0.54 0.60

Table 2.  The mean and standard deviation of the fitted exponents γ and κ, and the correlation between γ and
κ, for different network structure. The results are obtained from 100 Monte Carlo Simulation of the stochastic
model described in “Model: probabilistic SIR on a graph”. The infection parameters α and β are assumed to
take different values as indicated in the top row of the tables.

β = 0.2, α = 0.05 β = 0.2, α = 0.1 β = 0.3, α = 0.05 β = 0.3, α = 0.1

Fitting model AIC MLL AIC MLL AIC MLL AIC MLL
(a) Two-dimensional grid model, with connection to k = 4 neighbours (i.e. d = 1 in (5))
γ and κ free −126.64 1.35 −61.38 0.69 −140.59 1.49 −104.23 1.12
γ free κ = 1 −103.28 1.09 −43.37 0.49 −75.48 0.81 −30.35 0.36
γ = 1 and κ free 7.62 −0.02 9.05 −0.03 19.26 −0.13 21.40 −0.15
γ = 1 and κ = 1 35.95 −0.32 24.63 −0.21 32.19 −0.15 23.44 −0.19
(b) Two-dimensional grid model, with connection to k = 8 neighbours (i.e. d = 2 in (5))
γ and κ free −18.84 0.27 −62.87 0.71 −58.93 0.67 −82.37 0.90
γ free κ = 1 51.20 −0.45 14.76 −0.09 85.41 −0.79 63.98 −0.56
γ = 1 and κ free 54.05 −0.48 31.90 −0.26 8.03 −0.02 21.08 −0.15
γ = 1 and κ = 1 60.32 −0.56 32.20 −0.29 83.58 −0.80 64.58 −0.61
β = 0.3, α = 0.05 β = 0.3, α = 0.1
Fitting model AIC MLL AIC MLL
(c) Two-dimensional Gaussian mixture random graph, with connection to k = 4 neighbours
γ and κ free −106.28 1.14 −76.82 0.85
γ free κ = 1 −1.77 0.08 −1.38 0.31
γ = 1 and κ free 36.68 −0.31 36.01 −0.13
γ = 1 and κ = 1 35.80 −0.32 21.25 −0.17
Italy Germany France Spain
Fitting model AIC MLL AIC MLL AIC MLL AIC MLL
(d) Real data from four different countries
γ free 121.33 −0.81 156.18 −1.01 175.54 −1.24 148.34 −1.09
γ =1 216.93 −1.46 203.20 −1.31 219.55 −1.55 204.94 −1.51

Table 3.  Model selection results for simulated and real data. For the simulated data, the fitting models are
based on (8) with parameters c, γ , κ, σ . The models differ in that the exponents γ or κ are fixed or optimally
selected (free). For the real data, the size of susceptible population S(t) is assumed to be constant and the
exponent κ is ignored. The AIC represents the Akaike information criterion defined according to (7) and MLL
denotes the normalized maximum log-likelihood computed by maximizing (9) over free parameters. The
reported results for the simulated data are averaged over 100 Monte-Carlo simulations. For the simulated data,
the infection parameters α and β are assumed to take different values as indicated in the top row of the tables.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 9

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 6.  Comparison of four fSIR models with free and partially specified choice of exponents (with the SIR
model corresponding to γ = 1 and κ = 1) based on simulation data obtained from two-dimensional grid model
with infection parameter α = 0.05 and β = 0.3. The AIC and maximum likelihood scores for these models are
reported in Table 3.

Figure 7.  Dependence of the fitted exponents γ and κ on the model parameters α and β for the two-
dimensional grid model with k = 4 nearest neighbors. (Mean and error bars for standard deviation are based on
100 Monte Carlo simulations).

number of deaths and recovered from the cumulative sum of confirmed cases. The time evolution of the number
of confirmed cases and active cases, for these four different countries, is depicted in Fig. 11. We fitted the model
log(�Itrans (t)) = log(c) + γ log(I(t)) (10)
to the data, for the parameter c and exponent γ , for the first 100 days of the pandemic (starting on January 22).
Due to the fact that S(t) ≫ I(t) during these initial stages of infection spread, S(t) is treated as constant. The
result for the four different countries is depicted in Fig. 12. The model selection result, in comparison to the
model where γ = 1 is fixed, is presented in Table 3d.

Discussion
The recent onset of the COVID-19 pandemic has underscored the urgency of accurate models for the spread
of infectious diseases. These may help guide the allocation of resources, intervention, and mediation strategies,
and may help quantify the impact of lifestyle changes on the progression of the epidemic and the threshold for
herd ­immunity8. In particular, one immediate practical question in the current COVID-19 pandemic is to decide
when an intervention—such as reinstitution of social distancing (after this has been relaxed)—is appropriate.
Thus, while efforts to develop accurate models go back almost a c­ entury16 (see a­ lso4,13), the subject is especially
urgent today.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 10

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 8.  Spread of infection on a mixture of Gaussian random graph model for transmission rate β = 0.3 (the
recovery rate α = 0.05 or α = 0.1).The top panels depict the number of susceptible, infected, and recovered
individuals as a function of time over a range of 50 days, for a single realization of the stochastic model
described in “Model: probabilistic SIR on a graph”. The center and bottom panel depict (�Itrans (t), I(t)) and
(�Itrans (t), I(t), S(t)) respectively, for 100 Monte–Carlo simulations (with solid blue lines). The fractional model
fit is illustrated with orange dashed curve. The fitting procedure maximizes the log-likelihood (9) over the free
parameters. The value of the fitted exponents are shown in the caption.

The main thesis of this work is that models of epidemics, especially during the early phases, incorrectly assume
that the contagion depends on the product of infected and susceptible populations. Contagion takes place at the
boundary of infected cells and as a result it is the topology of the distribution of infected cells that dictate the
spread. An analogous situation takes place in tumor growth, where models suggest (­ see3,21,23) fractional expo-
nent for the contribution of tumor volume, as this more accurately captures the size of the boundary that affects
growth. Thus, based on an analogous rationale, we propose a fractional-power alternative, fSIR, to the standard
SIR model of disease dynamics. The value of exponents depend on a number of factors including the nature of
the boundary between infected cells and the general susceptible population. Specifically, the exponent relates
to the level mixing between infected and healthy individuals at interface between the two sub-populations, and
may be quantified by the diameter of the graph that represents contacts between individuals.
Our thesis is supported by simulation results as well as by fitting this fSIR model to recent COVID-19 datasets.
Specifically, the two-dimensional discrete probabilistic SIR models in Figs. 2 and 3 (with k = 4 nearest-neighbor
connection) and in Figs. 4 and 5 (with k = 8 nearest-neighbor connection), that simulate disease propagation
on a discrete domain of nodes (representing individuals in contact with one another), suggest exponents γ in
the range between 0.66 and 0.76 for the contribution of I(t) on the infection rate. Similar results are observed
in Fig. 8 for a two-dimensional random distribution of nodes (vertex set) with four nearest-neighbor contacts

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 11

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 9.  Dependence of the fitted exponent γ on graph connectivity for the mixture of Gaussian random
graph model. The infection parameters α = 0.1 and β = 0.3 are fixed. The network parameters, the number of
nearest neighbors k and the number of random additional edges m vary. A single realization is displayed for
each choice of network parameters.

Figure 10.  Dependence of the fitted exponents γ and κ on graph connectivity for the mixture of Gaussian
random graph model. The infection parameters α = 0.1 and β = 0.3 are fixed. The network parameters, the
number of nearest neighbors k and the number of random additional edges m vary. (Mean and error bars for
standard deviation are based on 100 Monte Carlo simulations).

(edge set). Here, the exponent of the I(t) contribution to the infection rate lies in a similar range ({0.62, 0.66} for
the conditions displayed). Two additional plots in Fig. 10 highlight the weak dependence of γ on the number of
short-range contacts (nearest neighbors) and the strong dependence on even a few long-range contacts amongst
the general population.
The fit of the COVID-19 data-set9 gives exponents for the contribution of I(t) on the infection rate in the
range of 0.6–0.8. In this data-set, the value of S(t) (that includes the remaining of a rather large total population)
varies insignificantly over time, and hence may be treated as constant. A limitation of our experiment is that the
value of I(t) is only an estimate since recording of all infected individuals is not guaranteed. Our findings relate to
patterns reported in recent studies on COVID-1911,28. In ­particular28, reports a power growth law of the infected

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 12

Vol:.(1234567890)
www.nature.com/scientificreports/

Figure 11.  Number of confirmed and active cases in four different countries during the period 01/22/2020 to
09/13/2020. The shaded area is the time period where the fractional model is used to fit the data in Fig. 12. The
number of active cases is computed by subtracting the total deaths and recovered from total confirmed cases.

population with time, which is consistent with our proposal since, in general, fractional exponents lead to power
laws (Assuming for simplicity that S(t) is constant, and the recovery rate α = 0, then the fSIR equation simpli-
1
fies to İ = βI(t)γ . When γ ∈ (0, 1), the solution can be readily expressed as I(t) = (c0 + β(1 − γ )t) 1−γ for a
constant c0. This represents a power law in t, unless γ = 1, in which case the solution I(t) = c0 eβt is exponential).
An important direction for future research lies in the better understanding of how fractional powers in
macroscopic dynamics arise from the probabilistic-network epidemic models. The majority of existing literature
on epidemics on networks deals with steady state analysis and is aimed at assessing percolation thresholds and
conditions for the appearance of an endemic state (see the review ­paper25). It will be also of great interest to
explore connections with a body of literature on reaction diffusion d ­ ynamics6,15,31,32 for the purpose of gaining
insight into macroscopic laws for epidemics.
The authors believe that it is imperative that a deeper and more extensive study is carried out, whereupon
the values of I(t), �I(t), R(t) are estimated from more extensive datasets. The effect of mediation efforts, such
as social distancing, should be recorded as well and taken into account by differentiating data for the periods
before and after such mediation protocols take effect. It is the authors’ hope that questions raised in this work,
as to the validity of the basic assumption in SIR models, lead to more reliable and robust ways to estimate the
progression of epidemics as well as the progression of the current COVID-19.

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 13

Vol.:(0123456789)
 www.nature.com/scientificreports/

Figure 12.  Relationship between infection growth due to transmission �Itrans (t) and the number of infected
individuals I(t) (active cases) with COVID-19 in four different countries. The data belongs to the first 100 days
of the infection starting from 01/22/2020.

Code availability
Computer code used in this work is available at https​://githu​b.com/AmirT​ag/fSIR.

Received: 14 May 2020; Accepted: 11 November 2020

References
1. Bauer, F., Castillo-Chavez, C. & Feng, Z. Mathematical Models in Epidemiology Vol. 69 (Springer, Berlin, 2019).
2. Benzekry, S. et al. Classical mathematical models for description and prediction of experimental tumor growth. PLoS Comput.
Biol. 10(8), e1003800 (2014).
3. Von Bertalanffy, L. Quantitative laws in metabolism and growth. Q. Rev. Biol. 32(3), 217–231 (1957).
4. Bjørnstad, O. N., Shea, K., Krzywinski, M. & Altman, N. Modeling infectious epidemics. Nat. Methods 20, 20 (2020).
5. Capasso, V. Mathematical Structures of Epidemic Systems Vol. 97 (Springer, Berlin, 2008).
6. Colizza, V., Pastor-Satorras, R. & Vespignani, A. Reaction-diffusion processes and metapopulation models in heterogeneous
networks. Nat. Phys. 3(4), 276–282 (2007).
7. Dong, E., Du, H. & Gardner, L. An interactive web-based dashboard to track covid-19 in real time. Lancet. Infect. Dis 20(5), 533–534
(2020).
8. Eker, S. Validity and usefulness of covid-19 models. Human. Soc. Sci. Commun. 7(1), 1–5 (2020).
9. Johns Hopkins University Center for Systems Science and Engineering. 2019 Novel Coronavirus COVID-19 (2019-nCoV) Data
Repository. https​://githu​b.com/CSSEG​ISand​Data/COVID​-19. Accessed 13 Sept 2020 (2020).
10. Gerlee, P. The model of muddle: In search of tumor growth laws. Cancer Res. 73, 2407–2411 (2013).
11. Gomes, M. et al. Individual variation in susceptibility or exposure to sars-cov-2 lowers the herd immunity threshold. medRxiv 20,
20 (2020).
12. Herman, A. B., Savage, V. M. & West, G. B. A quantitative theory of solid tumor growth, metabolic rate and vascularization. PLoS
One 6(9), e22973 (2011).
13. Hethcote, H. W. The mathematics of infectious diseases. SIAM Rev. 42(4), 599–653 (2000).

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 14

Vol:.(1234567890)
www.nature.com/scientificreports/

14. Keeling, M. J., Rand, D. A. & Morris, A. J. Correlation models for childhood epidemics. Proc. R. Soc. Lond. B Biol. Sci. 264(1385),
1149–1156 (1997).
15. Kelton, K. F., Greer, A. L. & Thompson, C. V. Transient nucleation in condensed systems. J. Chem. Phys. 79(12), 6261–6276 (1983).
16. Kermack, W. O. & McKendrick, A. G. A contribution to the mathematical theory of epidemics. Proc. R. Soc. Lond. Ser. A Contain.
Papers Math. Phys. Char. 115(772), 700–721 (1927).
17. Liu, W., Levin, S. A. & Iwasa, Y. Influence of nonlinear incidence rates upon the behavior of sirs epidemiological models. J. Math.
Biol. 23(2), 187–204 (1986).
18. Martcheva, M. An Introduction to Mathematical Epidemiology Vol. 61 (Springer, Berlin, 2015).
19. Newman, M. Spread of epidemic disease on networks. Phys. Rev. E 66(1), 016128 (2002).
20. Newman, M. Networks (Oxford University Press, Oxford, 2018).
21. Norton, L. A gompertzian model of human breast cancer growth. Cancer Res. 48, 7067–7071 (1988).
22. Norton, L. Conceptual and practical implications of breast tissue geometry: Toward a more effective, less toxic therapy. Oncologist
10, 370–381 (2005).
23. Norton, L. & Massagué, J. Is cancer a disease of self-seeding?. Nat. Med. 12(8), 875–878 (2006).
24. Novozhilov, A. S. On the spread of epidemics in a closed heterogeneous population. Math. Biosci. 215(2), 177–185 (2008).
25. Pastor-Satorras, R., Castellano, C., Van Mieghem, P. & Vespignani, A. Epidemic processes in complex networks. Rev. Mod. Phys.
87(3), 925 (2015).
26. Severo, N. C. Generalizations of some stochastic epidemic models. Math. Biosci. 4(3–4), 395–402 (1969).
27. Vaidya, V. G. & Alexandro, F. J. Evaluation of some mathematical models for tumor growth. Int. J. Biomed. Comput. 13, 19–35
(1982).
28. Wodarz, D. & Komarova, N. L. Patterns of the covid19 epidemic spread around the world: Exponential vs power laws. medRxiv
20, 20 (2020).
29. Xia, Y., Bjørnstad, O. N. & Grenfell, B. T. Measles metapopulation dynamics: A gravity model for epidemiological coupling and
dynamics. Am. Nat. 164(2), 267–281 (2004).
30. Xiao, X., White, E. P., Hooten, M. B. & Durham, S. L. On the use of log-transformation vs. nonlinear regression for analyzing
biological power laws. Ecology 92(10), 1887–1894 (2011).
31. Yuste, S. B., Acedo, L. & Lindenberg, K. Reaction front in an a + b → c reaction–subdiffusion process. Phys. Rev. E 69(3), 036126
(2004).
32. Yuste, S. B. & Lindenberg, K. Subdiffusion-limited a + a reactions. Phys. Rev. Lett. 87(11), 118301 (2001).

Acknowledgements
This research was supported by AFOSR Grants FA9550-20-1-0029, FA9550-17-1-0435, NSF Grants 1807664,
1839441, National Institute of Aging Grant R01-AG048769, MSK Cancer Center Support Grant/Core Grant (P30
CA008748), and a Grant from Breast Cancer Research Foundation BCRF-17-193.

Author contributions
T.G. and A.T. initiated the study, A.T. carried out the computational experiments and analysis, all authors A.T.,
T.G., L.N., and A.T. contributed to the writing.

Competing interests
The authors declare no competing interests.

Additional information
Correspondence and requests for materials should be addressed to T.T.G.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the
Creative Commons licence, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from
the copyright holder. To view a copy of this licence, visit http://creat​iveco​mmons​.org/licen​ses/by/4.0/.

© The Author(s) 2020

Scientific Reports | (2020) 10:20882 | https://doi.org/10.1038/s41598-020-77849-7 15

Vol.:(0123456789)