Advances in mosquito dynamics modeling∗

Karunia Putra Wijaya1 Thomas Götz1 Edy Soewono2

arXiv:1503.02573v1 [math.OC] 9 Mar 2015

1
Mathematical Institute
University of Koblenz, 56070 Koblenz, Germany
2
Department of Mathematics
Bandung Institute of Technology, 40132 Bandung, Indonesia
March 10, 2015

Abstract
It is preliminarily known that Aedes mosquitoes are very close to
humans and their dwellings, also give rises to a broad spectrum of
diseases: dengue, yellow fever, chikungunya. In this paper, we explore
a multi-age-class model for mosquito population secondarily classi-
fied into indoor-outdoor dynamics. We accentuate a novel design for
the model in which periodicity of the affecting time-varying environ-
mental condition is taken into account. Application of the optimal
control with collocated measure as apposed to the widely-used pro-
totypic smooth time-continuous measure is also considered. Using
two approaches: least-square and maximum likelihood, we estimate
several involving undetermined parameters. We analyze the model
enforceability to biological point of view such as existence, unique-
ness, positivity and boundedness of solution trajectory, also existence
and stability of (non)trivial periodic solution(s) by means of the basic
mosquito offspring number. Some numerical tests are brought along
at the rest of the paper as a compact realistic visualization of the
model.

Keywords: mosquito dynamics modeling; multi-age-class model; indoor-

outdoor dynamics; non-autonomous system; least-square; maximum
∗
This research has been financially supported by Indonesia Endowment Fund for Ed-
ucation (LPDP). Part of this research by third author is supported by the International
Collaboration Grant of Indonesian DGHE 2014

1
 likelihood; optimal control; the basic mosquito offspring number.

Mathematical Subject Classifications: 49J15; 49M05; 92D30;

37B55.

1 Introduction
We consider a mathematical model of mosquito population dynamics within
the framework similar to [28, 29], where the continuous population evolution
is formulated as initial value problem (IVP) of non-autonomous system
ẋ = f(t, x, u; η), t ∈ [0, T ], x(0) = x0  0. (1)
In natural setting, the system equation contains a hyperparameter η whose
appearance describes a collection of measurable intrinsic factors: e.g. natural
births, natural deaths, age-based transitions. Studying the qualitative be-
havior of the solution, one demands the fluctuation phenomena within those
factors to be distinguished as those of which essentially depend on time. As
a consequence of uncertainty in environmental condition, some elements of
η differ in time with possible seasonal trends: monotonic increasing, mono-
tonic decreasing, oscillating, or even fluctuating with Brownian-type move-
ment. Many references have even hypothesized that such intrinsic factors
may behave with periodic streamline in many cases due to environmental
changes: cf. [5, 11, 10, 12, 22, 23]. Ironically, in a national integrated
mosquito management programme, for example, Bonds [3] summarized that
fluctuating meteorology (raindrop, wind speed, air temperature, air humid-
ity, terrestrial radiation, etc) had been out of concern during deployment of
control devices in the field. Therefore, this costed substantial inefficiencies in
mass disposal. This dependency of parameters in time brings the model into
non-autonomous groundwork. As an extrinsic factor, the control measure u
is incorporated into the system for which it plays as a system regulator to-
wards achievement of the general objectives: i.e. minimizing both population
and cost for the control. In other words, the following objective functional
Z TX
def 1
X
J(u) = ωx,i x2i + ωu,j u2j dt (2)
2T 0 i∈I j∈I
x u

attains its minimum for given weighting constants {ωx,i }i∈Ix and {ωu,j }j∈Iu ,
Ix = {1, · · · , 5} and Iu = {1, 2}.
In line with matching the underlying dynamical process of the solution
with that from empirical measurement, we demonstrate estimation of some

2
undetermined parameters in the model by firstly settling them as random
variables. Then by exploiting information from the system, one can charac-
terize the solution as a handling function expressed in terms of such random
variables. We utilize the property of Maximum Likelihood (ML) and Least
Square (LS) which is basically minimization of the difference between the
handling function value-points and the measured data over all possible values
of undetermined parameters in a bounded set. The next problem arises when
analytical solution of the system, or eventually the handling function, can not
be determined explicitly due to complexity of the equation. For this reason,
a schematic Local Linearization (LL) method provides a trade-off between
numerical accuracy and computational outlay. Recently, several regimes in
coping with ML for parameter estimation within epidemic dynamical system
have been explored e.g. in [31, 9, 6].
There have been numerous mathematical papers discussing application
of optimal control scenario to the mosquito reduction issue (see e.g. [28,
29, 8, 24, 7, 18, 30] and some references therein). The authors used proto-
typic autonomous model utmost, encouraging us to propose a novel approach
adopting non-autonomous dynamical system theory. In this paper we restrict
our main scope to the application of temephos and ULV aerosol. Enhance-
ment of indoor-outdoor dynamics and utilization of polynomial collocation
design to the control measure are parts of our interests. The choice of the
design aims at achieving minimization of the costly objective meanwhile pro-
nouncing more efficient and accurate control deployment.

2 Model and analysis

In favor of IVP (1), x denotes the time-variant state which folds five con-
secutive elements, each of which represents the number of: indoor eggs x1 ,
outdoor eggs x2 , indoor larvae x3 , outdoor larvae x4 and adults x5 . A control
measure u is injected into the system as an active feedback regulator whose
elements represent the impact rate for investment of: temephos u1 and ULV
aerosol u2 . Note that we omit writing argument t when it is obvious. To go
into details, the governing system (1) is unfolded as

ẋ1 = pα(t)x5 − β1 x1 − qu1 x1 − µ1 x1 , (3a)

ẋ2 = (1 − p)α(t)x5 − β2 x2 − µ2 x2 , (3b)
ẋ3 = β1 x1 − γ1 x23 − β3 x3 − u1 x3 − ru2 x3 − µ3 x3 , (3c)
ẋ4 = β2 x2 − γ2 x24 − β4 x4 − su2 x4 − µ4 x4 , (3d)
ẋ5 = β3 x3 + β4 x4 − u2 x5 − µ5 x5 . (3e)

3
All the involved parameters are positive and are briefly explained as fol-
def
lows: p, q, r, s are the plausible probabilistic constants; α(t) = ǫ + ǫ0 cos(σt)
(where ǫ > ǫ0 > 0) is the birth rate of potential eggs depending qualitatively
on meteorology distribution; β{1,2,3,4} are the age-transitional rates for the
corresponding classes; γ{1,2} (where γ1 > γ2 ) are the driving forces to the
arising competition amongst larvae and; µ{1,2,3,4,5} are the death rates for the
corresponding classes.
def
It is assumed that the control measure u be in U = Ĉ 0 ([0, T ]; B) a set of
piecewise-continuous functions where B is a bounded block in R2+ . Since f is
uniformly locally Liptschitz continuous on time-state domains and piecewise
continuous on [0, T ], then x should lie in Ĉ 1 ([0, T ]; R5 ) a set of piecewise-
def
differentiable functions. It is defined a form ℓ(f ) = kf k(t) by taking a
norm k·k of f for each t. In this non-autonomous model, we initially denote
x(t) = ν(t, 0, x0 , u) the solution of (1) as a process. In a specified case
when ǫ0 = 0, typical analyses: existence and uniqueness, positive invariance,
existence and stability of equilibria as regards the basic mosquito offspring
number can be referred from our preceding work [29].
Consider M as a non-autonomous set where M ⊆ [0, T ] × R5 . Denote by
def
Mt = {x ∈ R5 : (t, x) ∈ M} t-fiber of M for each t ∈ [0, T ].
Definition 1 (Positively invariant) An autonomous set M ⊆ [0, T ] ×
R5 is called positively invariant under the process ν if for any x ∈ M0 ,
ν(t, 0, x, u) ∈ Mt for all t ≥ 0.
Theorem 1 It holds that M = [0, T ] × R5+ is positively invariant under the
process φ.

Proof Let n be (5×5)-matrix representing a collection of all normal vectors

(by rows) to the boundary of R5+ , ∂R5+ . It follows that n = −id where id is
the identity matrix. Notice that at ith boundary, ∂i R5+ ,
[nf (t, x, u)]i |x∈∂i R 5 ,u∈U ≤ 0 uniformly ∀t ≥ 0, i = 1, · · · , 5.
+

This generates evidence showing evolution of the solution-points in ∂R5+ in

counter-direction or at least perpendicular to the corresponding normal vec-
tor. Thus such trajectory of points, emanated from all t ≥ 0, can not leave
R5+ . 
We next consider a decomposition over f as f(t, x, u) = A(t)x+c1 x3 +c2 x24
2

to exemplify further analysis. In this case, A(t) is the Jacobian of the system
evaluated at 0. Let V : [0, +∞) × Ĉ 1 ([0, T ]; R5+ ) × Ĉ 1 ([0, T ]; R5+ ) → R be a
function defined as
def
V (t, x, y) = ℓ(x − y)2 .

4
This function delineates instantly three following conditions: (i) V > 0 if
x 6= y and ≡ 0 if x ≡ y, (ii) V is uniformly locally Lipschitz continuous on
dom(V ) and (iii) for any {xn , yn }n∈N ∈ Ĉ 1 ([0, T ]; R5+ ), limn→∞ V (t, xn , yn ) ≡
0 implies limn→∞ ℓ(xn − yn ) ≡ 0.

Lemma 1 The following assertion applies

lim h−1 (V (t + h, x + hf(t, x, u), y + hf(t, y, u)) − V (t, x, y)) ≤ g(t, V (t, x, y))
h→0−

where g(t, w) is a continuous function exceeding 2ℓ(A(t))w for all t, w ∈ R+ .

In contrast, ℓ(A(t)) denotes a corresponding form applied to a matrix A(t)
induced by ℓ(·).

Proof The inequality follows directly from straight-forward computation

(ref. Cauchy-Schwarz inequality). 

Lemma 2 The following scalar non-autonomous equation

ẇ = g(t, w)

holds these two conditions: (i) g(t, 0) = 0 for all t ∈ R+ and (ii) for each τ ∈
(0, +∞), w ≡ 0 is the only solution on [0, τ ] satisfying w(0) = 0 beforehand.

Remark In our model, A is a matrix-valued function over t whose elements

contain the continuous function α and the piece-wise continuous function u.
Along with the definition, it is clear that both α and u are bounded by some
continuous function, i.e. there exists a continuous function f such that e.g.
ℓ(u) ≤ f for all t. Thus without lost of generality, 1-form ℓ1 (A) induced by
the 1-norm k·k1 constitutes an easiness in proving that any induced form ℓ(A)
is bounded. This can immediately be seen by envisaging the matrix norm
equivalence, in sense that there exists a bounded positive function C(t) such
that ℓ(A(t)) ≤ C(t)ℓ1 (A(t)) for all t and any induced form ℓ(·). Now the
existence of such function g in the form g(t, w) = δ(t)w can be drawn upon
the fact that ℓ(A) is bounded.
We further state that Lemma 1 and Lemma 2 are inherently mild sub-
stances to prove the uniqueness of solution in (1). The notion of uniqueness
in non-autonomous dynamical system dates back to the seminal works by
Murakami [15], Ricciardi-Tubaro [20] and Kato [14], where aforesaid condi-
tions stipulated in Lemma 1 and Lemma 2 set the conforming requirements.
Adopting materials in [13, 14, 27], we summarize the existence and unique-
ness of the solution through the following corollary.

5
Corollary 1 (Uniqueness) IVP (1) has a unique solution x defined on R+
which maps to R5+ for every x0 ∈ R5+ .

It is further imposed the following assumptions on (1):

(H1) the hyperparameter η is chosen to lie in the following set

η ∈ R18

+ : x3 (t; η) ≥ x5 (t; η), x4 (t; η) ≥ x5 (t; η) ∀t ∈ R + and kηk < ∞

(H2) there exist positive continuous functional ϑ and sufficiently large con-
stant L such that η lies in
 
18 c(t) + ϑ(t) 2 3
η ∈ R+ : ℓ(x(t; η)) ≤ L + x5 (t; η) ∀t ∈ R+
2(γ1 + γ2 )
def
where c(t) = max{pα(t) − β1 − 2µ1 , (1 − p)α(t) − β2 − 2µ2 , β1 − β3 −
2µ3 , β2 − β4 − 2µ4 , α(t) + β3 + β4 − 2µ5 }.

Theorem 2 (Lyapunov method) Consider IVP (1). If there exist a Lya-

punov function Ψ : R+ × Ĉ 1 ([0, T ]; R5+ ) → R+ , positive integers ς{1,2,3} , posi-
tive and nonnegative real numbers L, ζ and positive continuous t-functionals
ϑ{1,2,3} where ϑ1 is nondecreasing satisfying the following conditions

(A1) ϑ1 ℓ(x)ς1 ≤ Ψ(t, x) ≤ ϑ2 ℓ(x)ς2

(A2) Ψ̇(t, x) ≤ −ϑ3 ℓ(x)ς3 + L

(A3) Ψ(t, x) − Ψ(t, x)ς3 /ς2 ≤ ζ

for all t ∈ R+ , then the solution x is bounded on R+ .

d
Proof We aim at computing the following total derivative dt Ψ(t, x) exp(at)
where a is a positive constant needed to be determined later. It is clear
(A2)
d
that dt
Ψ(t, x) exp(at) =(Ψ̇(t, x) + aΨ(t, x)) exp(at) ≤ (−ϑ3 ℓ(x)ς3 + L +
 
(A1) (A3)
ϑ3 ς3 /ς2
aΨ(t, x)) exp(at) ≤ − ς3 /ς2 Ψ(t, x) + L + aΨ(t, x) exp(at) ≤ (L +
ϑ2
def ϑ3
aζ) exp(at) by taking a = inf t∈R + ς /ς . This results in
ϑ23 2
 
(L + aζ) (L + aζ)
Ψ(t, x) = Ψ(0, x0 ) + exp(at) − exp(−at)
a a
(A1) (L + aζ) (L + aζ)
≤ ϑ2 (0)kx0 kς2 exp(−at) + ≤ ϑ2 (0)kx0 kς2 + .
a a
6
Using the left-side inequality in (A1), we get
 1/ς1
−1/ς1 ς2 (L + aζ)
ℓ(x) ≤ ϑ1 (0) ϑ2 (0)kx0 k +
a

since ϑ1 is nondecreasing. 

Corollary 2 If (H1) and (H2) are satisfied, then all solutions of (1) are
bounded on R+ .

Proof This is a direct consequence of applying Theorem 2 to (1) with the

following input: ς{1,2,3} = 2, ϑ{1,2} = 1, ϑ3 = ϑ, ζ = 0 and Ψ(t, x) = ℓ(x)2 . 

1
Remark Now we have an appropriate bound for the solution namely (kx0 k2 + L) 2 .

Definition 2 (Fundamental matrix) Consider a non-autonomous linear

system
ż = W (t)z, t ∈ R+ , z(0) = z0 (4)
with Z(t, 0) as fundamental matrix for W (t), i.e. z(t, 0, z0 ) = Z(t, 0)z0 ,
satisfying Z(0, 0) = id. It can further be verified that: (i) Z(t, s)Z(s, 0) =
Z(t, 0), (ii) Z(t, 0)−1 = Z(0, t) and (iii) Z(t, t) = id.

Theorem 3 Consider (4) where W (t) is 2π/σ-periodic. There exist a dif-

ferentiable 2π/σ-periodic matrix Γ1 (t) and a constant matrix Γ2 such that the
according fundamental matrix

Z(t, 0) = Γ1 (t) exp(Γ2 t).

Remark Several highlighting insights regarding Theorem 3:

• Γ1 (t), Γ2 need not to be unique and real even though W (t) is real

• the theorem holds for W (t) complex

• since Z(0, 0) = id, or eventually Γ1 (0) = Γ1 (2π/σ) = id, then Z(2π/

σ, 0) = Γ1 (2π/σ) exp(Γ2 2π/σ) = exp(Γ2 2π/σ).

Definition 3 (Floquet exponent and Floquet multiplier) Let ρ(A) be

the spectrum of A. In Theorem 3, each element of ρ(Γ2 ) is called as Floquet
exponent meanwhile each element of ρ(Z(2π/σ, 0)) is called as Floquet mul-
tiplier.

7
Take a look back to (1). Irrespective to the appearance of the control,
f (t, x, 0) can be identified and decomposed as f (t, x, ǫ0 ) = f 0 (x) + ǫ0 f 1 (t, x)
R1
where f 0 is autonomous and f 1 (t, x) = 0 ∂ǫ∂0 f(t, x, ξǫ0 ) dξ. It is preliminarily
known the following: (i) f(t + 2π/σ, x, ǫ0 ) = f (t, x, ǫ0 ) for all t ∈ R+ and (ii)
if Q is equilibrium point of ẋ = f 0 (x) then f 0 (Q) = 0.

Theorem 4 (Existence of periodic solution) Let f(t, x, ǫ0 ) = f0 (x) +

ǫ0 f1 (t, x) and Q be an equilibrium point of ẋ = f0 (x). If 2π/σρ(∇x f0 (Q)) ∋
2π/σλ ∈ / 2πiZ, then there exist a neighborhood U(Q) and ǫ1 such that for ev-
ery |ǫ0 | < ǫ1 , there exists a 2π/σ-periodic solution of (1) with unique initial
value x0 = x0 (ǫ0 ) ∈ U(Q).

Proof Let x(t, x0 , ǫ0 ) be such solution. Letting x0 ∈ U1 (Q) ⊂ R5+ then x

exists and is unique for all t ∈ R+ by Corollary 1 without lost of generality
def
for |ǫ0 | < ǫ1,1 . Let ζ(t) = ∇x0 x(t, Q, 0). Since ∇x0 x(t, x0 , ǫ0 ) deductively
satisfies the variational equation
d
∇x x(t, x0 , ǫ0 ) = ∇x0 f(t, x(t, x0 , ǫ0 ), ǫ0 )
dt 0
= ∇x f(t, x(t, x0 , ǫ0 ), ǫ0 )∇x0 x(t, x0 , ǫ0 ),

with ∇x0 x(0, x0 , ǫ0 ) = id, then ζ immediately satisfies

d
ζ = ∇x f 0 (Q)ζ with ζ(0) = id.
dt
def
Let S(x0 , ǫ0 ) = x(2π/σ, x0 , ǫ0 ) −x0 . Since f ∈ C 1 (R+ ×R5+ ×(−ǫ1,1 , ǫ1,1 ); R5 ),
then S ∈ C 1 (R5+ ×(−ǫ1,1 , ǫ1,1 ); R5 ). It is clear that S(Q, 0) = 0 and ∇x0 S(Q, 0) =
exp(∇x f 0 (Q)2π/σ) − id. Let v be eigenvector of ∇x f 0 (Q) associated with
eigenvalue λ. We know that (exp(∇x f 0 (Q)2π/σ)−id)v = (exp(2π/σλ)−1)v
making a clearance that exp(2π/σλ) − Q 1 is eigenvalue of ∇x0 S(Q, 0). If 2π/
σλ ∈ / 2πiZ then det(∇x0 S(Q, 0)) = 5i=1 exp(2π/σλi ) − 1 6= 0, making
∇x0 S(Q, 0) invertible. By implicit function theorem, there exist a domain
U2 (Q) × (−ǫ1,2 , ǫ1,2 ) and a smooth x0 (ǫ0 ) for (ǫ0 , x0 (ǫ0 )) defined on this do-
main such that S(x0 (ǫ0 ), ǫ0 ) = 0 or eventually x(2π/σ, x0 (ǫ0 ), ǫ0 ) = x0 (ǫ0 ).
Since f is 2π/σ-periodic over t, then x(t+ 2π/σ, x0 (ǫ0 ), ǫ0 ) = x(t, x0 (ǫ0 ), ǫ0 ) if
def
and only if x(2π/σ, x0 (ǫ0 ), ǫ0 ) = x0 (ǫ0 ). Now letting U(Q) = U1 (Q) ∩ U2 (Q)
and ǫ1 such that (−ǫ1 , ǫ1 ) ⊂ (−ǫ1,1 , ǫ1,1 ) ∩ (−ǫ1,2 , ǫ1,2 ) follows the desired
domain of existence. 
Let b1 = ǫp, b2 = ǫ(1 − p), b3 = β1 , b4 = β2 , b5 = β3 , b6 = β4 and
d1 = β1 + µ1 , d2 = β2 + µ2 , d3 = β3 + µ3 , d4 = β4 + µ4 , d5 = µ5 . With the

8
same technical arrangement using the next generation method [25] as in [29],
we define r
def 3 b1 b3 b5 b2 b4 b6
R(d3 , d4) = + (5)
d5 d1 d3 d5 d2 d4
as the so-called basic mosquito offspring number. In the domain of interest
R5+ , it has been proved in [29] that two equilibria of ẋ = f 0 (x) exist: zero
equilibrium and a positive equilibrium Q.
Lemma 3 There exist two 2π/σ-periodic solutions of (1) in R+ : trivial so-
lution x ≡ 0 if R(d3 , d4) 6= 1 and nontrivial solution x = ν associated with
nontrivial autonomous equilibrium Q if R(d3 + 2γ1 x⋆3 , d4 + 2γ2 x⋆4 ) 6= 1 where
(·, ·, x⋆3 , x⋆4 , ·) = Q.

Definition 4 (Stable periodic solution) A periodic solution ν is said to

be stable if for every ǫ > 0 there exists δ > 0 such that kx0 −ν(τ )k < δ implies
ℓ(x(t, τ, x0 ) − ν(t)) < ǫ for all t ≥ τ . Additionally, if limt→∞ ℓ(x(t, τ, x0 ) −
ν(t)) = 0 then ν is said to be asymptotically stable.
def
Theorem 5 (Stability of periodic solution) Let y = x − ν and irre-
def def
spective to the control, h(t, y) = f(t, y + ν) − f(t, ν) − W (t)y where W (t) =
∇x f(t, ν). If the following conditions hold:
(B1) ℓ(h) ≤ Kℓ(y)2 for some constant K

(B2) all Floquet exponents µ of ν correspond to W (t) lie off the open left-
half plane in C (or the Floquet multipliers lie off the open unit disk in
C)
Rt
(B3) in the decomposition exp( 0 W (s) ds) = Γ1 (t) exp(Γ2 t), Γ1 is bounded
w.r.t. ℓ(·)
then ν is asymptotically stable.

Proof It follows that y satisfies

ẏ = W (t)y + h(t, y),

where h(t, 0) = 0 and ∇y h(t, 0) = 0.

 Rt  (B2)
Observe that ℓ(Z(t, 0)) = ℓ exp( 0 W (s) ds) ≤ ℓ(Γ1 (t))ℓ (exp(Γ2 t)) ≤
(B3)
ℓ(Γ1 (t))C0 exp(−λt) ≤ C exp(−λt) for some positive constants C and λ.
def def
Fix δ where CKδ − λ < 0. Define 0 < a = λ − CKδ and b = Cδ . If

9
ky0 k ≤ b then by continuity of the vector field there exists τ such that y ex-
ists on [0, τ ] satisfying ℓ(y(t)) ≤ δ for all t ∈ [0, τ ]. Generating the solution,
 Rt  (B1)
we obtain ℓ(y) = ℓ Z(t, 0)y0 + Z(t, 0) 0 Z(s, 0)−1h(s, y) ds ≤ [ky0 kC +
Rt
CKδ 0 exp(λs)ℓ(y) ds] exp(−λt). Employing Gronwall’s Lemma, we get
exp(λt)ℓ(y) ≤ ky0 kC exp(CKδt) or ℓ(y) ≤ ky0 kC exp(−at) ≤ δ exp(−at)
such that limt→∞ ℓ(y) = 0. Using some extension theorem, it can be proved
that there exists ǫ such that y is defined on [0, τ + ǫ) by continuity of ℓ(y).
def
For all points, without lost of generality, {tn }n∈N \∞ = {τ +(1−1/n)ǫ}n∈N \∞ ,
note that k(y(ti )k ≤ δ exp(−ati ) < δ for ky0 k < b. This contradicts maxi-
mality of τ . 

Lemma 4 The trivial periodic solution ν ≡ 0 of (1) is asymptotically stable

whenever R(d3 , d4) < 1.

Proof Using direct computation, it can be shown that

K = 0 (6)
 pǫ0 
1 0 0 0 σ
sin(σt)
 0 1 0 0 (1−p)ǫ0
 σ

 sin(σt)
Γ1 (t) = 
 0 0 1 0 0  bounded w.r.t. ℓ(·)
 (7)
 0 0 0 1 0 
0 0 0 0 1
 
−d1 0 0 0 b1
 0 −d2 0 0 b2 
 
Γ2 =  b3
 0 −d3 0 0   (8)
 0 b4 0 −d4 0 
0 0 b5 b6 −d5

resulting Γ1 (2π/σ) = id. Noticing [29] Theorem 3.2, it has been proved
whenever R(d3 , d4 ) < 1 then ρ(Γ2 ) lies off the open left-half plane in C. 
def 1 Rt
Theorem 6 Let ν be nontrivial periodic solution, ν̄{3,4} (t) = t 0
ν{3,4} (s) ds
min def
where ν{3,4} (0) = ν{30,40} and ν{3,4} = mint∈[0,2π/σ] ν̄{3,4} (t). If

R(d3 + 2γ1ν3min , d4 + 2γ2ν4min ) < 1

then ν is asymptotically stable.

10
Proof Let y = x − ν. It is clear that y follows ẏ = g(t, y) where
g(t, y) = A(t)y + c1 (y23 + 2ν3 y3 ) + c2 (y24 + 2ν4 y4 ) by recalling our decomposi-
tion upon f in (1). Now we can state that y delineates non-autonomous
system with perturbance ν{2,3} where 0 is trivial periodic solution. We
can obtain easily the correspondence matrix for linearized system W (t) =
∇x g(t, 0). To save
R t space, we briefly state that the fundamental matrix Z(t) =
Z(t, 0) = exp( 0 W (s) ds) can not be presented easily as ∆1 (t) exp(∆2 t).
def
To continue proceeding, the idea is by choosing ∆1 (t) = Z(t) exp(−∆2 t)
and ∆2 for which exp(∆2 2π/σ) = Z(0)−1 Z(2π/σ). Then this choice sat-
isfies Z(t)Z(0)−1 Z(2π/σ) = ∆1 (t) exp(∆2 t) · exp(∆2 2π/σ) = ∆1 (t + 2π/
σ) exp(∆2 (t + 2π/σ)) = Z(t + 2π/σ) which is nothing but the so-called Flo-
quet theorem. Consider Γ2 = Γ2 (d3 , d4 ) as in (6), it is clear that Z(2π/
σ) = exp (2π/σΓ2 (d3 + 2γ1 ν̄3 (2π/σ), d4 + 2γ2 ν̄4 (2π/σ))). We immediately
get ∆2 = Γ2 (d3 + 2γ1 ν̄3 (2π/σ), d4 + 2γ2 ν̄4 (2π/σ)). If R(d3 + 2γ1 ν3min , d4 +
2γ2 ν4min ) < 1 then R(d3 + 2γ1ν̄3 (2π/σ), d4 + 2γ2 ν̄4 (2π/σ)) < 1, then with sim-
ilar consequence as in Lemma 4, ρ(∆2 ) lies off the open left-half plane in C.
Simultaneously, since R(d3 + 2γ1 ν̄3 (t), d4 + 2γ2 ν̄4 (t)) ≤ R(d3 + 2γ1ν3min , d4 +
2γ2 ν4min ) < 1 then Z is bounded and therefore ∆1 is bounded. The choice of
min
ν{3,4} in [0, 2π/σ] by the fact that ν̄{3,4} have the greatest deviations on their
amplitude at this range. 

3 Parameter estimation
Let η be the hyperparameter of the model (1) and P ⊂ R18 be its feasible
region. Let I := {i ∈ {1, · · · , 18} : ηi unfixed} and θ denotes a vector which
collects all associated parameters whose indices in I. Let Θ ⊂ P respectively
be feasible region for θ. The next key enabling technical simplification is that
one can further rearrange the elements of η as η = (ηf⊤ , θ⊤ )⊤ . In order to
find an estimate of θ, it is essential to identify whether the system in nature
is under control intervention or not. For the sake of simplicity, let us assume
that there are no control treatments during matching process. Fixing ηf and
setting u ≡ 0, we recast IVP (1) as

ẋ(t) = f̄(t, x(t); θ), t ∈ [0, T ], x(0) = x0 , θ ∈ Θ. (9)

def
Let J = {0, · · · , N} and
def
G N = {tj : tj = j∆t, tN = tf , j ∈ J } (10)

be our set of discrete time-points. Taking a good solver for ODE, we assume
that Eq. (9) results in the discrete process φ : G N ×0×R5+ ×Θ → R5+ mapping

11
(tj , 0, x0 , θ) to xj where the sequence {xj }j∈J conforms the regressing path.
def
Let H : R5+ → Rm be a function such that Φ = H ◦φ : G N ×0×R5+ ×Θ → Rm .
Assume (0, x0 ) is fixed, leading to exposition Φ : G N × Θ → Rm .
Let K ⊆ {1, · · · , 5}. Suppose that it is given a data set {T̂, X̂} of time-
state points which folds the sample {T̂il , X̂ il }i∈K
l=1,··· ,ki and let

def
A = {i ∈ K : 0 < T̂il < T, l = 1, · · · , ki }. (11)

The first highlighting processes are given briefly as follows. In practice, since
most T ˆ i is beyond G N , notorious interpolation and extrapolation processes
l
are needed for all i ∈ A and only interpolation process for all i ∈ K\A. The
processes seek all corresponding state-points at all tj based on information
from the known points given in the data set. Since an extrapolation process
suffers from greater uncertainty, thus the higher ki for all i ∈ A will help to
produce more meaningful results. Once we have the data set {T, ˆ X}
ˆ ex- and
interpolated with respect to aforesaid procedure, one draws the refined data
set, {T, X}, where it holds {Tij : j ∈ J } = G N for all i ∈ K and therefore
|K|
X : G N → R+ .
In contrast with easiness in real implementation and the lack of details
in the data, we will always need such function H-like, H, which maps X
|K|
from R+ into Rm as a collection course. We assume that the data for all
state-classes in the model are not necessarily known. Now the corresponding
number m should be taken to satisfy 1 ≤ m ≤ min{|K|, 5}. Working with
def
the same treatment as in the regressing path, we let X = H ◦ X : G N → Rm ,
making X and Φ comparable.

3.1 Least-square approach

Let us define the error of measurement
def
ǫj (θ) = Xj − Φj (θ), j ∈ J. (12)

Let ǫ(θ) = [ǫ1 (θ) · · · ǫ|J | (θ)]. Given an estimate for Θ, now our problem reads
as
def
find θ ∈ Θ such that J(θ) = kǫ(θ)k2F → min . (13)
In this formulation, k·kF denotes the Frobenius norm.

3.2 Maximum likelihood approach

Assume that {ǫj (θ)}j∈J are considerably independent and identically dis-
tributed (iid) since it most probably appears commonly that the data is

12
randomly distributed relative to the regressing path. Then we can assume
iid
{ǫj (θ)}j∈J ∼ N (0, Σ) with the corresponding probability density function
(pdf)  
1 1 2
ϕ(ǫj ; θ) = m 1 exp − ∆j (14)
(2π) 2 det(Σ) 2 2
p
where ∆j = ǫj (θ)⊤ Σ−1 ǫj (θ) is Mahalanobis distance from ǫj (θ) to 0 and
det(Σ) is determinant of Σ. The joint pdf (jpdf) for all random variables
{ǫj (θ)}j∈J is given by
   
Y 1 1 2 1 1 2
ϕ(ǫ; θ) = m 1 exp − ∆j = m|J | |J |
exp − k∆k2
j∈J (2π) det(Σ)
2 2 2 (2π) 2 det(Σ) 2 2
(15)
⊤
where ∆ = (∆1 , · · · , ∆|J | ) . Independent from Σ, we get the fact

kǫ(θ)kF → 0 if and only if k∆k2 → 0 if and only if ϕ → max . (16)

So the most sensible way of finding a good θ is by maximizing ϕ, or by

maximizing log ϕ, since log is monotonically increasing.
Theorem 7 (Maximum likelihood for multivariate normal distribution)
iid
Let {ǫj (θ)}j∈J ∼ N (0, Σ) and
def
X def 1
S(θ) = ǫj (θ)ǫj (θ)⊤ , Σ̂(θ) = S(θ)
j∈J
|J |

then Σ̂(θ) = arg maxΣ log ϕ(ǫ; θ) where ϕ(ǫ; θ) is given as in (15).

Remark The (m × m)-random variable S(θ) is called as Wishart ma-

trix which follows Wishart distribution Wm (|J |, Σ) with parameters m (di-
mension of the matrix), |J | (degree of freedom) and Σ (positively defined
covariance matrix). We can always perform the standardization S(θ) ∼
1 1
Wm (|J |, Σ) ⇔ Σ− 2 S(θ)Σ− 2 ∼ Wm (|J |, Im ) where Wm (|J |, Im ) denotes the
standard Wishart distribution. In special case when m = 1, W1 (|J |, 1) =
χ2|J | .
Hence the final constrained optimization problem reads as: find θ ∈ Θ
such that
 
def 1 1 ˆ 2
J(θ) = log ϕ(ǫ; θ) = log m|J | |J |
exp − k∆k2 → max . (17)
(2π) 2 det(Σ̂(θ)) 2 2
q
ˆ ˆ ˆ ⊤ ˆ
In this case, ∆ = (∆1 , · · · , ∆|J | ) and ∆j = ǫj (θ)⊤ Σ̂(θ)−1 ǫj (θ).

13
Remark In line with computation of optimal solution using a derivative-
use method, one has to find the so-called Fisher’s score function F(θ) which
is nothing but the Jacobian ∇θ log ϕ(ǫ; θ) and (for Newton/quasi-Newton)
the information matrix I|J | (θ) which is negative of Hessian ∇2θ log ϕ(ǫ; θ).
These computations require very lengthy expression and therefore one has
to achieve very expensive evaluations. Nevertheless, heuristics should offer
trade-off in direct solving but limit their speed in convergence. Another im-
portant aspect in this problem is that, by giving Θ from the scratch, the
value of the parameter θ on each iterate seems converging to the boundary
of Θ. Initiatively, in this paper, we impose fixed value for all parameters right
up in front, i.e. η = (ηf⊤ , θf⊤ )⊤ , and then perturb the resulting solution with
Gaussian noise along with the covariance matrix Σf . Matching the original
with this perturbed model, one can perceive the process as θf -recovery. Con-
sidering the estimate for Θ, there would be 2 possible methods which can be
used: Wald confidence and the profile likelihood confidence methods. For |J |
very large, the variance var(θf ) ∼ I−1 |J | (θf ) where [var(θf )]ik = [cov(θf,i , θf,k )],
i, k = 1, · · · , |I|. Let z1−τ be (1 − τ )-quantile of a standard normal dis-
tribution and diagZ be a vector composed by selecting out main diagonal
elements of Z. Choosing appropriate τ , we gain Wald confidence interval
running from Wald test: (H0 : θ = θf vs H1 : θ 6= θf ) as
h q q i
Θ = θf − z1−τ diag I−1 (θ
|J | f ), θf + z1−τ diag I−1
(θ
|J | f ) . (18)

One thing we need to make sure that at this large |J |, numerical evaluation
of the inverse information matrix should not be really expensive – one can
approach it with numerical approximation on derivatives. Another method
which is more accurate than Wald confidence method for |J | small is the
profile likelihood confidence method. The profile likelihood confidence in-
terval (also called the likelihood ratio confidence interval) derives from the
asymptotic Chi-square distribution of the likelihood ratio statistics. Let l(Θ)
and u(Θ) denote the lower and upper bound of Θ respectively. It is known
 
ϕ(ǫ; θf )
2 log < χ2|J |−1;1−τ (19)
ϕ(ǫ; θ)
which essentially determines
  
1 2
l(Θ) = arg ϕ(ǫ; θ) : ϕ(ǫ; θ) = ϕ(ǫ; θf ) exp χ|J |−1;1−τ . (20)
2

For the upper bound u(Θ), we take some arbitrary value in (θf , ∞). To
counteract the solution reaching the boundary of the given set, we estimate

14
a small positive number ε and refine the objective in both (13) and (17) using
interior point function as: find θ ∈ Θ such that
def
J(θ) = kǫ(θ)k2F − ε [log(θ − l(Θ)) + log(u(Θ) − θ)] → min (LS)

and analogously: find θ ∈ Θ such that

def
J(θ) = log ϕ(ǫ; θ) + ε [log(θ − l(Θ)) + log(u(Θ) − θ)] → max . (MLE)

3.3 Evaluation of Φ
In order to evaluate Φ, we adopt the property of Local Linearization (LL)
method as it persuades balance between computational outlay and conver-
gence issues. Related to as in [2, 17], the authors suggested to find the
solution of

ẋ(t; θ) = f̄ (tj , x(tj ; θ)) + ∇x f̄(tj , x(tj ; θ))(x(t; θ) − x(tj ; θ)) (21)

on each subinterval [tj , tj+1) where x(t0 ) = x0 and tj , tj+1 ∈ G N . The solution
of (21) is given as the following recursion

xj+1 (θ) = xj (θ) + exp ∇x f̄ j (θ)∆t − id ∇x f̄ j (θ)−1 f̄ j (θ)

 

(22)

providing that ∇x f̄ j (θ) is invertible. In this formulation, xj (θ), f̄ j (θ) and

∇x f̄ j (θ) are abbreviations for x(tj ; θ), f̄(tj , x(tj ; θ)) and ∇x f̄ (tj , x(tj ; θ)) re-
spectively.
Lemma 5 Let us devote to two solutions on the subinterval [tj , tj+1 ). Let
φ be a process representing the analytic solution of (9) and φ∆t be a pro-
cess generated as the solution of (9) using LL method. Assume that f̄ is
uniformly Lipschitz continuous over all prescribed domains of its arguments.
Independent from θ, there exists a positive constant C such that

kφ − φ∆t kL1 ([tj ,tj+1 )) ≤ C∆t2 uniformly ∀tj , tj+1 ∈ G N .

Remark This Lemma exhibits the evidence that the smaller ∆t taken in
numerical computation, the more solution from LL method tends to analytic
solution. One can take a look for the analogous proof of this Lemma in
e.g. [19]. Another important problem needed to be tackled is how we can
efficiently compute matrix exponential in (22). Interested reader can take a
look into Padé approximate for matrix exponential, see e.g. [1, 26]. However,
in this paper we omit the details of this approximate.

15
4 Optimal control problem
4.1 Polynomial collocation
Let us assume that the control measures are applied in every n days. The
spacing time between times of application is assigned as h+ne where h, e ∈ R2
def def
and e is a vector containing unities. Let τ 1,k = Ik (h + ne) and τ 2,k =
Ik (h + ne) + ne be two discrete time-points where Ik is a (2 × 2)-diagonal
matrix containing counter. Let ∗ and ∗/ denote respectively the MATLAB
pointwise multiplication and division between two vectors. If diag(Ik ) counts
all elements of the set {0e, 1e, 2e, · · · , (T − n) ∗ /(h + ne)} in a consecutive
manner, then both τ 1,k and τ 2,k count some distinct numbers in R2 . For the
sake of simplicity, assume that (T − n) is divisible by h + ne in correspond
to the operator ∗/. Therefore, we have a finite collection of intervals
def
T = {[τ11,k , τ12,k ) × [τ21,k , τ22,k )}k∈{0,··· ,k(T −n)∗/(h+ne)k∞ } ⊂ R2+ . (23)

Note that for all k ∈ {min ((T − n) ∗ /(h + ne)) , · · · , k(T −n)∗/(h+ne)k∞ },
it appears either [τ11,k , τ12,k ) = ∅ or [τ21,k , τ22,k ) = ∅ simultaneously since one
may need h containing distinct elements. A collection of such counters should
have zero cardinality. Let δτk1 ;τ 2 (t) be some vector-valued function following

pk (t), (t, t) ∈ [τ11,k , τ12,k ) × [τ21,k, τ22,k )

(
def
δτk1 ;τ 2 (t) ≡ 
0, (t, t) ∈ ([0, T ] × [0, T ])\ [τ11,k , τ12,k ) × [τ21,k , τ22,k )
(24)
where pk (t) is a vector-valued polynomial of  degree being arbitrary. Let
def k(T −n)∗/(h+ne)k
δτ 1 ;τ 2 (t) = δτ01 ;τ 2 (t), · · · , δτ 1 ;τ 2 ∞
(t) and

def

p(t) = p0 (t), · · · , pk(T −n)∗/(h+ne)k∞ (t)

be (2 × [k(T − n) ∗ /(h + ne)k∞ + 1])-vector-valued functions. Thus we design

our control measure as
def
c (t) = v ∗ δτ 1 ;τ 2 (t). (25)
In this case, v denotes a vector containing control measure values at T whose
dimension is similar to that of p(t). Given a continuous-time control u(t).
Define a p-collocation Λ(·; p) : U → U such that for c = Λ(u; p), there is a
weighting vector v satisfying c (t) = v ∗ δτ 1 ;τ 2 (t) and ku − ck → min for all
(t, t) ∈ T .

16
4.2 Existence of optimal control
Designate the transformation over time-state variables on the following per-
formance
xi 7→ yi for all i = 1, · · · , 5 and t 7→ y6 .
As a consequence, there exists a function Y such that the non-autonomous
equation (1) is similar to the following autonomous equation

ẏ = Y (y, u), yi (0) = x0 for all i = 1, · · · , 5 and y6 (0) = 0, t ∈ [0, T ]

(26)
where it holds Y6 (y, u) = 1. Define D = {y : ẏ = Y (y, u), t ∈ [0, T ], y(0) =
y0 , u ∈ U} as a set of admissible states. Recall our objective functional
RT
J(u) = 0 g(y, u) dt as in (2) along with this transformation and compose
the optimal control problem as

find (y, u) ∈ D × U such that J(u) → min . (OC)

The following lemma derives one appropriate material to prove existence of

optimal control in (OC).
Lemma 6 The following set
def
S(t, y) = {g(y, u) + γ, Y1 (y, u), · · · , Y6(y, u) : γ ≥ 0, u ∈ B}

is convex for all (t, y) ∈ [0, T ] × R6+ .

Proof Fix (t, y) ∈ [0, T ]×R6+ as an arbitrary choice and write g(y, u)+γ =
κ(u1 , u2 , γ). Keeping in mind that Y = Y (y, u1 , u2 ) and κ = κ(u1 , u2 , γ) are
def
continuous over u and γ. Let B = [0, a1 ]×[0, a2 ] such that (u1 , u2 ) ∈ B. Now
consider that S(t, y) is a set of points ξ ∈ R7 where it structure can be stud-
ied as follows. For fixed u2 = 0 and γ = 0, it is clear that ξ1 = κ(u1 , 0, 0),
ξ{2,4} = Y{2,4} (y, u1 , 0) and ξ{3,5,6,7} are constant. This means that such
points generate a parametric curve in R7 whose projection on each ξ1 ξ2 -
and ξ1 ξ4 -plane are convex quadratic, meanwhile on each ξ1 ξ3 -, ξ1 ξ5 -plane
and so on are straight segments since [0, a1 ] is bounded. If γ goes from 0
to ∞ then this convex curve moves along ξ1 -axis from initial position to
infinity. At this stage, the generated 2D-hyperplane, say P2D , is clearly con-
vex. Moreover, for constant ξ{3,5,6,7} we can identify P2D in ξ1 ξ2 ξ4 -Cartessian
space. If u2 goes from 0 to a2 , then P2D simultaneously moves along new
def
axes called ξ5 - and ξ6 -axis. It is clear that the set P3D = {(ξ1 , ξ2, ξ4 , ξ5 ) ∈
R4 : (ξ1 , ξ2 , ξ4 ) ∈ P2D , ξ5 ∈ [Y4 (y, u1 , a2 ), Y4 (y, u1 , 0)]} is convex, and there-
def
fore is the set P4D = {(ξ1 , ξ2 , ξ4 , ξ5, ξ6 ) ∈ R5 : (ξ1 , ξ2, ξ4 , ξ5 ) ∈ P3D , ξ6 ∈

17
[Y5 (y, u1 , a2 ), Y5 (y, u1 , 0)]}. Then for fixed (t, y), the set S(t, y) = {ξ ∈ R7 :
(ξ1 , ξ2 , ξ4 , ξ5 , ξ6) ∈ P4D , ξ3, ξ7 constant} is also convex. 
Now we are ready to prove the existence of optimal control for our model.
Lemma 7 There exists the only optimal pair (ȳ, ū) for the optimal control
problem (OC).

Proof We refer to Filippov-Cesari’s Theorem [4] to prove the existence of

optimal pair. It states whenever the following conditions hold
1. There exists an admissible pair
2. The set S(t, y) defined in Lemma 6 is convex for every (t, y) ∈ [0, T ]×R6

3. U is compact
4. There exists a number δ > 0 such that every solution ℓ(y) < δ for all
t ∈ [0, T ] and all admissible pairs (y, u),
then there exists such optimal pair. Ads 1 and 3 are trivial, meanwhile ad 2
is proved in Lemma 6. Clearly a well-defined vector field Y in (26) conduces
continuity of y on [0, T ]. By Bounded Value Theorem, one can easily show
that y is bounded on the compact t-domain [0, T ]. This completes the proof.

Definition 5 (Saturation) Let Ĉ([0, T ]; R2+ ) denote a set of piecewise-continuous
functions mapping [0, T ] into R2+ . It is defined the saturation
def
Υ : Ĉ([0, T ]; R2+ ) → U = Ĉ([0, T ]; B)

where the block B = [0, a1 ] × [0, a2 ] as

Υ(u) = max (0, min (a, u)) . (27)

Lemma 8 Let u⋆ ∈ Ĉ([0, T ]; R2+ ) be the optimal control of broadening prob-

lem (OC) by expanding the space for admissible controls. Let

Υ : Ĉ([0, T ]; R2+ ) → U

be a saturation over the control resulting a projected optimal control ū ∈ U,

def
or ū = Υ(u⋆ ). Then (ȳ, ū) ∈ D × U is the solution of the original problem
(OC).

Working with the similar technical arrangement as in [29] Sec. 4, we generate

an appropriate necessary condition for the optimal control as follows.

18
Theorem 8 (Necessary condition) Consider the broadening optimal con-
trol problem (OC) by expanding the space for admissible controls. Let u⋆ ∈
Ĉ([0, T ]; R2+ ) be the minimizer for J and y⋆ ∈ Ĉ 1 ([0, T ]; R6+ ) be the resulting
state. There exists a dual variable z⋆ ∈ Ĉ 1 ([0, T ]; R6) such that the tuple
(y⋆ , z⋆ , u⋆ ) satisfies the following system

⋆ ∂H 
ẏ = with y⋆ (0) = y0  0, (28)
∂z (y⋆ ,z⋆ ,u⋆ )

⋆ ∂H 
ż = − , (29)
∂y (y⋆ ,z⋆ ,u⋆ )
 
⋆ ∂H 
u = arg zero , z⋆ (T ) = 0 (30)
∂u
(y ,z )
 ⋆ ⋆

def
for all t ∈ [0, T ]. The functional H(y, z, u) = g(y, u) + hz, π(y, u)i is Hamil-
tonian functional, meanwhile all equations in (28) are respectively the state,
adjoint, gradient equations and transversality condition.

The adjoint (with transversality condition) and gradient equations can now
be unfolded as

ż1 = −ωx,1 y1 + (β1 + qu1 + µ1 )z1 − β1 z3 , z1 (T ) = 0 (31a)

ż2 = −ωx,2 y2 + (β2 + µ2 )z2 − β2 z4 , z2 (T ) = 0 (31b)
ż3 = −ωx,3 y3 + (2γ1 y3 + β3 + u1 + ru2 + µ3 )z3 − β3 z5 , z3 (T ) = 0 (31c)
ż4 = −ωx,4 y4 + (2γ2 y4 + β4 + su2 + µ4 )z4 − β4 z5 , z4 (T ) = 0 (31d)
ż5 = −ωx,5 y5 + (u2 + µ5 )z5
− [ǫ + ǫ0 cos(σy6 )]pz1 − [ǫ + ǫ0 cos(σy6 )](1 − p)z2 , z5 (T ) = 0 (31e)
ż6 = σǫ0 sin(σy6 )py1 z1 + σǫ0 sin(σy6 )(1 − p)y2 z2 , z6 (T ) = 0 (31f)

and

ωu,1 u1 − qy1 z1 − y3 z3 ≡ 0 (32a)

ωu,2 u2 − ry3 z3 − sy4 z4 − y5 z5 ≡ 0. (32b)

The following algorithm illustrates our scheme to solve (OC).

Remark Note that a termination criterion is included in the Line 11 namely:

when λ < λ0 for sufficiently small λ0 then the program terminates.

19
Algorithm 1 Gradient-based method for solving (OC).
Require: k = 0, an initial guess for the control uk ∈ U, an error tolerance
ǫ > 0 and an initial step-length λ > 0.
1: Compute the p-collocation uk ← Λ(uk ; p).
2: Compute yku ← yk (·; uk ) and zky,u ← zk (·; yk (·; uk )) consecutively from
the state (with forward scheme) and adjoint equation (with backward
scheme).
3: Compute the objective functional J(uk ).
4: Compute u⋆k (yku , zky,u ) from the gradient equation and set u⋆k ←
Λ(u⋆k , p).
5: Update uk+1 (λ) ← uk + λu⋆k and uk+1 ← Υ(uk+1 ). Compute yk+1 u and
k+1
zy,u .
6: Compute J(uk+1 ) and set ∆J ← J(uk+1 ) − J(uk ).
7: if |∆J| < ǫ then
8: ˆ ← (yk+1
Set (ŷ, û, J) u ,u
k+1
, J(uk+1 )) then stop.
9: end if
10: while ∆J ≥ 0 do
11: Update new λ ← arg mins∈[0,λ] φ(s) := J(uk+1 (s)) where φ is a
quadratic function as representative of J with respect to the step-
length s. Note that the solution exists since φ(0), φ′ (0) and φ(λ) can
be computed directly. Compute new uk+1 (λ) ← uk + λu⋆k and set
uk+1 ← Υ(uk+1 ). Then compute new yk+1 u and zk+1
y,u .
k+1 k+1 k
12: Compute J(u ) and set ∆J ← J(u ) − J(u ).
13: if |∆J| < ǫ then
14: ˆ ← (yk+1
Set (ŷ, û, J) u ,u
k+1
, J(uk+1 )) then stop.
15: end if
16: end while
17: Set k ← k + 1.
18: Go to Step 4.
19: return The tuple (ŷ, û, J).ˆ

5 Numerical tests
Table 1 gives an estimate value of all parameters used in the model. In trial
scheme, we aim atPrecovering 2 parameters: θ1 = ǫ0 and θ2 = p. We use
5
H(x) = H(x) = i=1 xi , Σf = 10 uniformly for all classes. We run ge-
netic algorithm as a core program to solve both (LS) and (MLE) with the
following computer specification: operating system OSX 10.9.4, processor
2.6 GHz Intel Core i5, RAM 16 GB, programming language Python 2.7.6
64bits with programming environment Spyder 2.2.4 and accuracy 4 digits.

20
The corresponding result can be looked up in Table 2. The role of two impor-
tant parameters in the model to magnitude of the basic mosquito offspring
number is shown in Fig. 1. Concurrently, from Figs. 2, 3 and 4 we show
the performance and attainment of optimal control to suppress the size of
mosquito population.

ǫ ǫ0 p q r s β1 β2 β3 β4 γ1 γ2 µ1 µ2
3 2 0.4 0.04 0.05 0.05 0.3 0.2 0.08 0.05 0.004 0.0026 0.02 0.01

µ3 µ4 µ5 T h n ωx,{1,2,3,4} ωx,5 ωu,{1,2} a x0 σ

2π
0.02 0.01 0.4 151 [9,19] 1 1 2 4 × 103 [1,1] [21,43,24,37,8] ⌊T /4⌋

Table 1: Estimate for all parameters involved int the model (1).

Problem Scheme 1 Scheme 2 Scheme 3 Scheme 4 Scheme 5

IT 100 1000 2000 5000 10000
(LS) ET 17 s 125 s 323 s 901 s 1987 s
∆t = 0.1 θ1 1.7213 1.7326 1.8001 1.8995 1.9021
θ2 0.2001 0.2567 0.3210 0.3334 0.3789
(MLE) ET 191 s 295 s 692 s 1443 s 3403 s
∆t = 0.1 θ1 1.777 1.8743 1.9123 1.9375 1.9786
θ2 0.2932 0.3031 0.3635 0.3800 0.3994
(LS) ET 185 s 1367 s 3715 s 10101 s 27456 s
∆t = 0.01 θ1 1.8514 1.9522 1.9812 1.9991 1.9999
θ2 0.3465 0.3821 0.4187 0.4097 0.4091
(MLE) ET 2921 s 3229 s 7942 s 19534 s 48047 s
∆t = 0.01 θ1 1.9987 1.9998 1.9999 2.0000 2.0000
θ2 0.3985 0.4023 0.4003 0.4000 0.4000

Table 2: The standard genetic algorithm (GA) performance in solving (LS)

and (MLE). IT = number of iterates, ET = elapsed time, s = second.

6 Concluding remarks
We have exhibited the mosquito population dynamics model using some uni-
fying theories bearing from non-autonomous dynamical system. Imposing
relevant assumptions over all parameters in the model, we prove positivity,

21
 5
2.1
2.1
2.1
1
4.5 temephos
ULV aerosol
2 0.9
2 2
2
4
0.8

1.9 1.9
3.5 1.9 0.7

control (1/day)
1.8 1.8 0.6
3 1.8
ε

0.5
1.7 1.7
2.5 1.7
0.4
1.6 1.6 1.6
2
0.3
1.5 1.5 1.5

1.5 1.4 0.2

1.4 1.4

1.3 1.3 1.3

0.1
1
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
p 0
0 50 100 150
time (day)

Figure 1: The basic mosquito off-

spring number R(d3 , d4) in (p, ǫ)- Figure 2: Optimal control that re-
plane. sults to Fig. 4.

500 350
ind. egg ind. egg
out. egg out. egg
450 ind. larv ind. larv
out. larv out. larv
300
adult adult
400

250
350
all classes (individual)

all classes (individual)

300
200

250

150
200

150
100

100

50
50

0 0
0 50 100 150 0 50 100 150
time (day) time (day)

Figure 3: Uncontrolled dynamics. Figure 4: Controlled dynamics.

22
uniqueness and boundedness of the corresponding solution. With the Floquet
theory, we prove that trivial periodic solution exists if the basic mosquito off-
spring number R(d3 , d4) 6= 1 and is asymptotically stable if R(d3 , d4) < 1.
In this paper, we can not derive a direct relationship between existence
and stability of nontrivial periodic solution ν correspond to nontrivial au-
tonomous equilibrium Q and the basic mosquito offspring number. It states
whenever max R(d3 + 2γ1 x⋆3 , d4 + 2γ2 x⋆4 ), R(d3 + 2γ1ν3min , d4 + 2γ2ν4min ) <


1 (see Lemma 3 and Theorem 6) then ν exists and is asymptotically stable.

From Table 1 and Fig. 1, all set of parameters satisfying R(d3 , d4 ) > 1 makes
nontrivial periodic solution attracting the original solution, therefore it is
graphically asymptotically stable. From Fig. 1, one has to reduce ǫ (meaning
that one has to ensure that meteorology does no longer support mosquito
life) in order to reduce the basic mosquito offspring number significantly. A
condition when R(d3 , d3 ) < 1 guarantees that the mosquito population can
return in insignificant number within finite time and completely die out for
expanded time.
Parameter estimation in our paper bears from necessity of suitable codes
which have extreme reliability in real implementation. Three generic meth-
ods: Local Linearization (LL), Padé approximate and Genetic Algorithm
(GA) come into the play of estimating some parameters in the model. Sum-
marizing the performance of our codes (ref. Table 2), one shows that the
compact program executes in exponential time with the low convergence on
average. Meanwhile, it highlights in the table that MLE scheme converges
more rapid than LS scheme with respect to the number of iterates. Beyond
naive implementation of GA, we examine that this low convergence results
from expensive evaluation of the objective, even rigorous computation of LL
solution on each iterate. Nevertheless, heuristic GA has gained a good deal
in popularity of solving optimization problem with expensive objective. We
strive to work on finding efficient codes of derivative-use method as a striking
resemblance with GA. This in turn enables us to compare both methods in
pursuit of efficient codes within the same framework, even with the higher
size of both the system and undetermined parameters.
We present some brief conclusions from application of optimal control. In
this paper, our work is circumscribed by application of constant collocation.
This can be more or less a key step toward development of efficient deploy-
ment of the control. Further application of polynomial collocation of degree
≥ 1 is needed to propose well-suited program, being readable throughout
academia. By comparing Figs. 3 and 4, we conclude that optimal control
can generally reduce all classes in the mosquito population. From Fig. 2,
it is noticed that application of the ULV aerosol has to be enhanced rather
than that of temephos. It is concluded that the pattern of optimal control

23
fluctuates with same tendency as that of the model dynamics. This means
that the dynamics pattern of the model has a strong influence on the control.

References
[1] Arioli M, Codenotti B, Fassino C. The Padé method for computing the
matrix exponential. Linear Algebra and its Applications 1996; 240:11–
130.

[2] Biscay R, Jimenez JC, Riera JJ, Valdes PA. Local linearization method
for the numerical solution of stochastic differential equations. Annals of
the Institute of Statistical Mathematics 1996; 48(4):631–644.

[3] Bonds JAS. Ultra-low-volume space sprays in mosquito control: a critical

review. Medical and Veterinary Entomology 2012; 26(2):121–130.

[4] Cesari L. Optimization – Theory and applications. Problems with ordinary

differential equations 1983; Springer-Verlag, Berlin-Heidelberg-New York.

[5] Cushing JM. An introduction to structured population dynamics 1998;

SIAM, Philadelphia.

[6] Donneta S, Samson A. Estimation of parameters in incomplete data mod-

els defined by dynamical systems. Journal of Statistical Planning and
Inference 2007; 137(9):2815–2831.

[7] Esteva L, Yang HM. Mathematical model to assess the control of Aedes
aegypti mosquitoes by the sterile insect technique. Mathematical Bio-
sciences 2005; 198:132–147.

[8] Fister KR, McCarthy ML, Oppenheimer SF, Collins C. Optimal control
of insects through sterile insect release and habitat modification. Mathe-
matical Biosciences 2013; 244(2):201–212.

[9] Guedj J, Thiébaut R, Commenges D. Maximum likelihood estimation in

dynamical models of HIV. Biometrics 2007; 63:1198–1206.

[10] Henson SM, Cushing JM. The effect of periodic habitat fluctuations on
a nonlinear insect population model. Journal of Mathematical Biology
1997; 36:201–226.

[11] He D, Earn DJD. Epidemiological effects of seasonal oscillations in birth

rates. Theoretical Population Biology 2007; 72:274–291.

24
[12] Ireland JM, Mestel BD, Norman R.A. The effect of seasonal host birth
rates on disease persistence. Mathematical Biosciences 2007; 206:31–45.

[13] Kato S. Some remarks on nonlinear differential equations in Banach

spaces. Hokkaido Mathematical Journal 1975; 4:205–226.

[14] Kato S. On local and global existence theorems for a nonautonomous dif-
ferential equation in a Banach space. Funkcialaj Ekvacioj 1976; 19:279–
286.

[15] Murakami H. On non-linear ordinary and evolution equations, Funk-

cialaj Ekvacioj 1966; 9:151–162.

[16] Pappalardo M, Passacantando M. Stability for equilibrium problems:

from variational inequalities to dynamical systems. Journal of Optimiza-
tion Theory and Applications 2002; 113:567–582.

[17] Ramos JI. Linearized methods for ordinary differential equations. Ap-
plied Mathematics and Computation 1999; 104:109–129.

[18] Rafikov M, Bevilacqua L, Wysec APP. Optimal control strategy of

malaria vector using genetically modified mosquitoes. Journal of The-
oretical Biology 2009; 258(3):418–425.

[19] Ramos JI, Garcı́a-López CM. Piecewise-linearized methods for initial-

value problems. Applied Mathematics and Computation 1997; 82(2-
3):273–302.

[20] Ricciardi P, Tubaro L. Local existence for differential equations in Ba-

nach space. Bollettino dell’Unione Matematica Italiana 1973; 8:306–316.

[21] Sideris TC. Ordinary differential equations and dynamical systems 2014;
Series: Atlantis Studies in Differential Equations, 2, Atlantis Press.

[22] Sauvage F, Langlais M, Pontier D. Predicting the emergence of human

hantavirus disease using a combination of viral dynamics and rodent de-
mographic patterns. Epidemiology & Infection 2007; 135:46–56.

[23] Sauvage F, Langlais M, Yuccoz NG, Pontier D. Modeling hantavirus in

fluctuating populations of bank voles: the role of indirect transmission
on virus persistence. Journal of Animal Ecology 2003; 72:1–13.

[24] Thomé RCA, Yang HM, Esteva L. Optimal control of Aedes aegypti
mosquitoes by the sterile insect technique and insecticide. Mathematical
Biosciences 2010; 223(1):12–23.

25
[25] van den Driessche P, Watmough J. Reproduction numbers and sub-
threshold endemic equilibria for compartmental models of disease trans-
mission. Mathematical Biosciences 2002; 180:29–48.

[26] van Loan CF. On the limitation of Padé approximation to the matrix
exponential 1977; in Padé and rational approximation, E.B. Saff and R.S.
Varga (eds.), New York: Academic Press.

[27] Webb G. Continuous nonlinear perturbation of linear accretive operators

in Banach spaces. Journal of Functional Analysis 1972; 10:191–203.

[28] Wijaya KP, Götz T, Soewono E, Nuraini N. Temephos spraying and ther-
mal fogging efficacy on Aedes aegypti in homogeneous urban residences.
ScienceAsia 2013; 39S:48–56.

[29] Wijaya KP, Götz T, Soewono E. An optimal control model of mosquito

reduction management in a dengue endemic region. International Journal
of Biomathematics 2014; 7(5):23–45.

[30] Wolf C, Langlais M, Sauvage F, Pontier D. A multi-patch epidemic

model with periodic demography, direct and indirect transmission and
variable maturation rate. Mathematical Population Studies 2006; 13:153–
177.

[31] Yu Y, Liang H. Parameter estimation for HIV ODE models incorporat-

ing longitudinal structure. Statistics and Its Interface 2013; 6:9–18.

26