See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342211425

Analytic solution of the SEIR epidemic model via asymptotic approximant

Preprint · June 2020

CITATIONS READS
0 19

4 authors, including:

Steven J Weinstein Kelly Rogers

Rochester Institute of Technology Rochester Institute of Technology
81 PUBLICATIONS   1,108 CITATIONS    1 PUBLICATION   0 CITATIONS   

SEE PROFILE SEE PROFILE

Nathaniel S. Barlow
Rochester Institute of Technology
40 PUBLICATIONS   153 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Asymptotic Approximants View project

All content following this page was uploaded by Nathaniel S. Barlow on 16 June 2020.

The user has requested enhancement of the downloaded file.

Analytic solution of the SEIR epidemic model via asymptotic approximant
Steven J. Weinstein,1, 2 Morgan S. Holland,1 Kelly E. Rogers,1 and Nathaniel S. Barlow1, a)
1)
School of Mathematical Sciences, Rochester Institute of Technology, Rochester, NY 14623,
USA
2)
Department of Chemical Engineering, Rochester Institute of Technology, Rochester, NY 14623,
USA
(Dated: 16 June 2020)
An analytic solution is obtained to the SEIR Epidemic Model. The solution is created by constructing a
single second-order nonlinear differential equation in ln S and analytically continuing its divergent power
series solution such that it matches the correct long-time exponential damping of the epidemic model. This
is achieved through an asymptotic approximant (Barlow et. al, 2017, Q. Jl Mech. Appl. Math, 70 (1), 21-48)
in the form of a modified symmetric Padé approximant that incorporates this damping. The utility of the
analytical form is demonstrated through its application to the COVID-19 pandemic.

Asymptotic approximants have have been successful at the model to enable future analysis.
creating analytical solutions to many problems in math- The SEIR epidemic model considers the time-evolution
ematical physics1–8 . Like the well-known Padé approxi- of a susceptible population, S(t), interacting with an ex-
mant9,10 , they are constructed to match a primary series posed population, E(t), and infected population, I(t),
expansion in a given region up to any specified order. where t is time. This model is expressed as11
Unlike Padés, however, the form of an asymptotic ap-
proximant is not limited to a ratio of polynomials, and dS
= −βSI (1a)
its structure is chosen to enforce the asymptotic equiva- dt
lence in a region away from the primary series expansion.
By increasing the number of terms in an asymptotic ap- dE
= βSI − αE (1b)
proximant, it converges to the exact solution in these two dt
regions - as well as at all points in between. Convergence
is certainly a necessary condition for a valid approximant; dI
although there is yet no proof, we find that a convergent = αE − γI, (1c)
dt
approximant matches, with high accuracy, the numeri-
cal solutions of the governing equation systems examined with a removed population (recovered + deaths), R(t),
thus far1–8 . evolved by
The COVID-19 outbreak motivates the need to ex-
tend the application of asymptotic approximants to epi- dR
= γI (1d)
demiology models. The method has seen recent success dt
in providing a closed-form solution to the Susceptible- and constraints
Infected-Recovered (SIR) model8 . Here, we extend
the method to the commonly used Susceptible-Exposed- S = S0 , E = E0 , I = I0 , R = R0 at t = 0, (1e)
Infected-Recovered (SEIR) model. This model is formu-
lated as a system of nonlinear ordinary differential equa- where β, α, γ, S0 , E0 , I0 , R0 are non-negative constant
tions, for which no exact analytic solution has yet been parameters11 . Along with initial conditions from (1e),
found. The analytic nature of the asymptotic approx- the solution for S, E, and I may be first obtained
imant derived in what follows is advantageous, in that from (1a) through (1c) and the solution for R subse-
the accuracy and computational expense is not affected quently extracted from (1d).
by the duration of the epidemic prediction; the form is We now manipulate the system (1) into an equivalent
built such that it is accurate in t ∈ [0, ∞) and all points 2nd -order equation in ln S to simplify the analysis that
in between. Depending on the duration, it may be benefi- follows. Equations (1a) and (1b) are added to obtain
cial to replace a numerical solution with the approximant dS dE
within a fitting algorithm that extracts SEIR parameters. + = −αE. (2)
dt dt
En route to the approximant, we also present an alterna-
tive formulation of the SEIR model as a single 2nd -order Solving (1c) for E and substituting into (2) then leads to
nonlinear differential equation in ln S. This form enables
an efficient series solution about t = 0, expansion about d2 I dI dS
+ (γ + α) +α + αγI = 0. (3)
t → ∞, and may itself prove to be an attractive form of dt2 dt dt
Equation (1a) is rewritten as
1 d ln S
a) Electronic mail: [email protected] I=− (4)
β dt
 2

and substituted into (3) to arrive at the 3rd -order equa- obtain the expansion of the nonlinear term ef ≡ PS inn(7).
tion To do so, we solve for the coefficients of S = ãn t by
applying Cauchy’s product rule to the chain-rule result
d3 ln S d2 ln S dS d ln S
+ (γ + α) − αβ + αγ = 0. (5) f 0 S = S 0 and evaluate like-terms to obtain the recursive
dt3 dt2 dt dt expression given by (9d). Although the series solution
Equation (5) may be integrated to obtain given by (9) is an analytic solution to (7), it is only valid
within its radius of convergence and is incapable of cap-
d2 ln S d ln S turing the long-time behavior of the system. This moti-
2
+ (γ + α) − αβS + αγ ln S = C, (6) vates the use of an approximant to analytically continue
dt dt
the series beyond this radius.
where the integration constant The long-time asymptotic behavior of the system (7)
is required to develop our asymptotic approximant, and
C = αγ ln(S0 ) − αβ (E0 + I0 + S0 ) (7a)
so we proceed as follows. It has been proven in prior
is obtained by evaluating the right-hand side of (6) at literature12 that S approaches a limiting value, S∞ , as
t = 0 using (1e) and (1a)-(1c). The form of equation (6) t → ∞, and this corresponds to I → 0 in the same limit.
suggests that the variable substitution f = ln S be made, Thus, f approaches a limiting value, f∞ ≡ ln S∞ , as
and the result is t → ∞. The value of f∞ satisfies the following equation12

d2 f df γ
+ (γ + α) − αβef + αγf = C (7b) ef∞ − (f∞ − ln S0 ) − E0 − I0 − S0 = 0 (10a)
dt2 dt β

where, from (1e) and (4), in the interval

df f∞ ∈ (−∞, ln γ/β). (10b)

f = ln S0 , = −βI0 at t = 0. (7c)
dt
We expand f as t → ∞ as follows:
Once f is solved for in the above, the solution to the
SEIR model is given by f ∼ f∞ + g(t) where g → 0 as t → ∞. (11)
S = ef (8a) Equation (11) is substituted into (7), eg is replaced with
its power series expansion, and terms of O(g 2 ) are ne-
1 df glected to achieve the following linearized equation
I=− (8b)
β dt d2 g dg
+ (γ + α) + αγ − αβef∞ g = 0.


2
(12)
dt dt
γ
R = R0 − (f − ln S0 ) (8c) The general solution to (12) is
β
g = 1 eλ1 t + 2 eλ2 t (13a)
E = E0 + I0 + S0 + R0 − I − S − R. (8d)
 
1
q
The series solution of (7) is given by λ1,2 = −α − γ ± (γ − α)2 + 4αβef∞ (13b)
2
∞
f=
X
an tn , a0 = ln S0 , a1 = −βI0 (9a) where 1 and 2 are unknown constants and λ2 < λ1 < 0
since ef∞ < γ/β from (10b). Thus the long-time asymp-
n=0
totic behavior of f is given by

a2 = [C − (α + γ) a1 + αβS0 − αγa0 ] /2 (9b) f ∼ f∞ + 1 eλ1 t , t → ∞. (14)

Higher order corrections to the expansion (14) may be

αβãn − (γ + α)(n + 1)an+1 − αγan obtained by the method of dominant balance10 as a se-
an+2 = , n>0
(n + 2)(n + 1) ries of more rapidly damped exponentials. However, the
(9c) pattern by which the corrections are asymptotically or-
n dered is not as straightforward as that of the SIR model,
1 X provided in Barlow and Weinstein 8 . In that work, an
ãn>0 = (n − j + 1)an−j+1 ãj , ã0 = S0 . (9d)
n + 1 j=0 asymptotic approximant is constructed as a series of ex-
ponentials that exactly mimics the long-time expansion.
The result (9) is obtained by the standard procedure of In the SEIR model, complications in the higher-order
inserting (9a) into (7) and finding a recursion for the co- asymptotic behavior arise from the competition between
efficients by equating like-terms. It is thus necessary to the two exponentials in (13a). Here, we enforce the
 3

0 1

-0.2 0.9

0.8
-0.4

0.7
-0.6

0.6
-0.8
0.5

-1
0.4

-1.2
0.3

-1.4
0.2

-1.6 0.1

-1.8 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

FIG. 1. Analytical and numerical solutions to the SEIR model (1), where the susceptible (S), exposed (E), infected (I), and
recovered (R) populations are represented as a fraction of the total population and t is in units of days. (a) Solution shown
in terms of f ≡ ln S. As the number of terms N is increased, the series solution, denoted fS,N (given by (9), dashed curves),
diverges and the approximant, denoted fA,N (given by (15), solid curves), converges to the exact (numerical) solution (•’s).
Corresponding relative errors are provided in figure 5a. (b) The converged asymptotic approximant for f is used to obtain S,
E, I, and R from (8) shown by solid curves and compared with the numerical solution (closed symbols). The model parameters
values and initial conditions α = 0.466089, β = 0.2, γ = 0.1, S0 = 0.88, E0 = 0.07, I0 = 0.05, and R0 = 0 are taken from
estimates of Ebola virus propagation examined in Rachah and Torres 13 .

leading-order t → ∞ behavior given by (14) and make the An and Bn coefficients of (15) (given α, γ, β, S0 , E0 ,
a more traditional choice for matching with the t = 0 ex- I0 ) is available from the authors16 .
pansion (9). We create an approximant with an embed-
The SEIR approximant (15) is thus an analytic expres-
ded rational function with equal-order numerator and de-
sion that, by construction, matches the correct t → ∞
nominator (i.e., a symmetric Padé approximant10 ), such
behavior given by (14) and whose expansion about t = 0
that it approaches the unknown constant 1 in (14) as
is exact to N th -order. A comparison between the approx-
t → ∞, while converging to the intermediate behavior at
imant solution (15) and the numerical solution to (1) is
shorter times. The assumed SEIR approximant is given
provided in figures 1-4 with the relative error for all four
by
cases provided in figure 5. The indicated error in figure 5
N/2
X is calculated by comparing S(t) to its accurate numerical
An tn solution (assumed to be exact); equivalent curves are ob-
n=0 tained when the other dependent variables of the model
fA,N = f∞ + eλ1 t , N even (15)
N/2
X are used.
1+ Bn tn Figure 1a provides a typical comparison of the N -term
n=1
series solution (9) denoted by fS,N (and dashed lines), the
where the An ’s and Bn ’s are obtained such that the Tay- N -term approximant (15) denoted by fA,N (solid lines),
lor expansion of (15) about t = 0 is exactly (9). Note and the numerical solution (•’s). Note that the series
that, although a rational function is being used in (15), solution has a finite radius of convergence as evidenced
it is not a Padé approximant itself. Padés are only ca- by the poor agreement and divergence from the numer-
pable of capturing tn behavior in the long-time limit, ical solution at larger times, even as additional terms
where n is an integer. The pre-factor eλ1 t is required are included. By contrast, the approximant converges
to make (15) an asymptotic approximant for the SEIR as additional terms are included. For N = 18, the ap-
model. However, we may still make use of fast Padé co- proximant is visibly indistinguishable from the numerical
efficient solvers14,15 by recasting (15) as a Padé for the solution on the scale of figure 1a. Figure 5a provides the
series that results from the Cauchy product between the relative error of the approximant for the data shown in
expansions of e−λ1 t and f −f∞ , the former which is com- figure 1. Increasing the number of terms beyond N = 18
monly known and the latter which is given by (9) (except does improve accuracy up to a point, but a minimum er-
with a0 = ln S0 − f∞ ). A MATLAB code for computing ror barrier is eventually reached of O(10−6 ) at N = 26;
 4

note that, to make this assessment, we take the max- and epidemic data after fitting are attributed to the ap-
imum relative error with respect to time (the maxima plicability of the SEIR model and not the approximant.
in figure 5a). For larger values of N , the maximum er- In figures 2-4, a least squares fit to I and R data is used
ror worsens, and the approximant begins to diverge, i.e. to extract SEIR parameters α, β, γ and initial conditions
there is an optimal value of N at which to truncate the S0 and E0 . To do so, the initial values of I0 and R0 are
approximant. Asymptotic approximants do often exhibit taken directly from the COVID data set17 . Additionally,
an optimum truncation6,7 as is typically observed with the time t = 0 is chosen such that disease has progressed
asymptotic expansions in general10 . We emphasize here to a point where initial trends are observed, so that curve
that a numerical solution is not needed to assess conver- shapes are consistent with those reasonably predicted by
gence of approximants to within their optimal truncation; the SEIR model. Adjustments such as this have been
covergence in the Cauchy sense (i.e., distance between well described in fits done in previous work18,19 . The
approximants decreases with increasing N ) may be used. initial guesses for the iterative least-squares fit are taken
In addition to this issue, deficient approximants are pos- from data fits for earlier times than examined here18,19 .
sible with increasing N due to zeroes that can arise in the Our results demonstrate that an asymptotic approxi-
denominator of (15). Such approximants are ignored in mant can be used to provide accurate analytic solutions
assessing convergence. To avoid this behavior, the lowest to the SEIR model. Future work should examine the
number of terms that yields the desired accuracy should ability of the asymptotic approximant technique to yield
be chosen. The convergence of the approximant with closed-form solutions for even more sophisticated epi-
increasing N (up until its optimal trunctation) is a nec- demic models, as well as their endemic counterparts12 .
essary condition for a valid approximant. In figure 1b, 1 N. S. Barlow, A. J. Schultz, S. J. Weinstein, and D. A. Kofke,
the converged (N = 18) asymptotic approximant for f J. Chem. Phys. 137, 204102 (2012).
is used to obtain analytic solutions for S, E, I, and R 2 N. S. Barlow, A. J. Schultz, S. J. Weinstein, and D. A. Kofke,

from (8), which are compared with the numerical solution AIChE J. 60, 3336 (2014).
3 N. S. Barlow, A. J. Schultz, S. J. Weinstein, and D. A. Kofke,
for these quantities. The approximant for N = 18 pro-
vides excellent agreement between these quantities and J. Chem. Phys. 143, 071103:1 (2015).
4 N. S. Barlow, C. R. Stanton, N. Hill, S. J. Weinstein, and A. G.
numerics within the visible scale of the plot, with errors Cio, Q. J. Mech. Appl. Math. 70, 21 (2017).
quantified by figure 5a. 5 N. S. Barlow, S. J. Weinstein, and J. A. Faber, Class. Quant.

Grav. 34, 1 (2017).

6 R. J. Beachley, M. Mistysyn, J. A. Faber, S. J. Weinstein, and
The figure 1 results described above correspond to a
case examined in Rachah and Torres 13 to model an Ebola N. S. Barlow, Class. Quant. Grav. 35, 1 (2018).
7 E. R. Belden, Z. A. Dickman, S. J. Weinstein, A. D. Archibee,
outbreak. In figures 2, 3, and 4, the approximant is ap- E. Burroughs, and N. S. Barlow, Q. J. Mech. Appl. Math. 73,
plied to COVID-19 data17 for Yunan (China), Japan, and 36 (2020).
Sweden, respectively. Figures 5b-d provide the relative 8 N. S. Barlow and S. J. Weinstein, Physica D 408, 1 (2020).
9 G. A. Baker Jr. and J. L. Gammel, J. Math. Anal. Appl. 2, 21
error for these cases; the largest indicated value of N in
each figure is the optimum truncation as discussed above (1961).
10 C. M. Bender and S. A. Orszag, Advanced Mathematical Methods
for figure 5a. Note that we extensively surveyed the avail- for Scientists and Engineers I: Asymptotic Methods and Pertur-
able COVID-19 data, and the results in figures 2-5 are bation Theory (McGraw-Hill, 1978).
representative of the fits and variability in the number of 11 W. O. Kermack and A. G. McKendrick, Proc. Roy. Soc. London

terms needed for convergence of the approximant up to A 115, 700 (1927).

12 H. W. Hethcote, in Mathematical Understanding of Infectious
its optimal truncation.
Disease Dynamics, edited by S. Ma and Y. Xia (World Scientific
Publishing, 2008).
Note that the reported COVID-19 outbreak data17 is 13 A. Rachah and D. F. M. Torres, Math Method Appl. Sci. 40,
provided in terms of confirmed cases, recovered individ- 6155 (2017).
14 “https://github.com/chebfun/chebfun/blob/master/padeapprox.m,”.
uals, and deaths per day. We use recovered + deaths as
15 P. Gonnet, S. Güttel, and L. N. Trefethen, SIAM Rev. 55, 101
an approximation for the removed population R and use
(2013).
confirmed − recovered − deaths as an approximation to 16 https://www.mathworks.com/matlabcentral/fileexchange/77007-
compare with the quantity I of the SEIR model. It is ac- approximantcoefficientsseir.
knowledged that the actual COVID-19 data is influenced 17 John Hopkins University CSSE, “Novel coronavirus (covid-19)

by effects not included in the SEIR model, and this can cases,” https://github.com/CSSEGISandData/COVID-19.
18 M. Peirlinck, K. Linka, F. S. Costabal, and E. Kuhl, Biomech.
affect the ability of the model to closely fit actual COVID
Model. Mechan. doi.org/10.1007/s10237-020-01332-5, 1
data. The data approximations made here are to enable (2020).
comparisons with model predictions. The ability of the 19 K. Linka, M. Peirlinck, F. S. Costabal, and E. Kuhl, Comput.

approximant to match numerical results is unaffected by Method Biomec. doi.org/10.1080/10255842.2020.1759560,

such approximations. Disagreement between the model 1 (2020).
 5

10

8 200

4 150

0 100

-2

-4 50

-6

-8 0
0 10 20 30 40 50 60 70 80 90 100 0 50 100 150

FIG. 2. Analytical and numerical solutions to the SEIR model (1), where S, E, I, R are in units of people and t is in days.
All other notation and labels are the same as in figure 1, except R now also includes deaths. Corresponding relative errors are
provided in figure 5b. SEIR model parameters values and unknown initial conditions are obtained via a least-squares fit to
the Yunan, China COVID-19 outbreak data17 (open symbols). Best fit parameters are α=0.395031, β=0.00333, γ=0.0553093,
S0 =142, and E0 =0. The initial conditions I0 = 44 and R0 =0 are taken directly from the data set17 at a chosen t = 0 (here
January 28, 2020).

10 4
15 2

1.8

1.6

1.4
10

1.2

0.8

5
0.6

0.4

0.2

0 0
0 10 20 30 40 50 60 70 80 90 100 0 10 20 30 40 50 60 70 80 90 100

FIG. 3. Analytical and numerical solutions to the SEIR model (1), where S, E, I, R are in units of people and t is in days.
All other notation and labels are the same as in figure 1, except R now also includes deaths. Corresponding relative errors
are provided in figure 5c. SEIR model parameters values and unknown initial conditions are obtained via a least-squares fit to
the Japan COVID-19 outbreak data17 (open symbols). Best fit parameters are α=0.2332207, β=2.040015×10−5 , γ=0.034334,
S0 =15442, and E0 =0. The initial conditions I0 = 1649 and R0 =529 are taken directly from the data set17 at a chosen t = 0
(here April 1, 2020).
 6

10 4
11 6

10
5

9
4

2
6

1
5

4 0
0 20 40 60 80 100 120 0 20 40 60 80 100 120

FIG. 4. Analytical and numerical solutions to the SEIR model (1), where S, E, I, R are in units of people and t is in days.
All other notation and labels are the same as in figure 1, except R now also includes deaths. Corresponding relative errors are
provided in figure 5d. SEIR model parameters values and unknown initial conditions are obtained via a least-squares fit to
the Sweden COVID-19 outbreak data17 (open symbols). Best fit parameters are α=0.041281, β=1.513332×10−6 , γ=0.004407,
S0 =50306, and E0 =10015. The initial conditions I0 = 1743 and R0 =20 are taken directly from the data set17 at a chosen t = 0
(here March 22, 2020).
 7

10 2
10 0

10 0
10 -2

10 -2
10 -4

10 -4

10 -6
10 -6

10 -8
10 -8

-10
10
10 -10

-12
10
10 -12

10 -14
10 -14

0 10 20 30 40 50 60 70 80 90 100
0 50 100 150

2 2
10 10

10 0 10 0

10 -2 10 -2

-4 -4
10 10

10 -6 10 -6

-8 -8
10 10

-10 -10
10 10

10 -12 10 -12

10 -14 10 -14

0 50 100 150 0 50 100 150

FIG. 5. Relative error of the approximant (15) for increasing N as a function of t (in days). The exact solution is taken to
be the numerical solution of (1), computed using the 4th -order Runge-Kutta scheme with a time-step of 10−4 . The subfigures
(a)-(d) correspond to the cases presented in figure 1-4, respectively. For all figures, N is taken up until optimal truncation is
achieved, indicated by a dashed curve. The cusps in the figures have no physical meaning and simply indicate where the sign
of (SA,N − Sexact ) changes.

View publication stats