You are on page 1of 16

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

net/publication/228924580

A Cell Growth Model Revisited

Article · January 2012

CITATIONS READS

13 119

3 authors:

Gregory Derfel Bruce van Brunt


Ben-Gurion University of the Negev Massey University
44 PUBLICATIONS   606 CITATIONS    89 PUBLICATIONS   942 CITATIONS   

SEE PROFILE SEE PROFILE

Graeme Wake
Massey University
235 PUBLICATIONS   3,498 CITATIONS   

SEE PROFILE

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

Pattern Formation in Reaction-Diffusion Systems with Spatially Varying Parameters View project

Mathematical Modelling View project

All content following this page was uploaded by Graeme Wake on 24 December 2013.

The user has requested enhancement of the downloaded file.


Report Number 09/27

A Cell Growth Model Revisited

by

G. Derfel, B. van Brunt and G.C. Wake

Oxford Centre for Collaborative Applied Mathematics


Mathematical Institute
24 - 29 St Giles’
Oxford
OX1 3LB
England
A Cell Growth Model Revisited
G. Derfel∗, B. van Brunt†, G.C. Wake‡

Abstract
In this paper a stochastic model for the simultaneous growth and di-
vision of a cell-population cohort structured by size is formulated. This
shows that the functional differential equation which describes the steady
form of the steady-size distribution which is approached asymptotically
and satisfies the well-known pantograph equation is more simply derived
via a Poisson process. This firmly establishes the existence of the steady-
size distribution and gives a form for it in terms of a sequence of proba-
bility distribution functions. Also it shows that the pantograph equation
is a key equation for other situations where there is a distinct stochastic
framework.
Key words: steady-size distribution, asymptotic behaviour, Poisson pro-
cess, Pantograph equation.
MOS Classification codes: 35B40, 35B41, 35 L45, 92C17

1 Introduction
In this paper we revisit a cell growth model developed by Hall and Wake [6]. This
model was originally developed to model plant cells [5], however, it has found
applications in tumor growth in humans [2]. A feature of this model is that
a well-known functional differential equation, the pantograph equation, arises
from a separation of variables solution to a Fokker-Planck equation. Specifically,
let n(x, t) denote the number density functions of cells of size x at time t i.e.,for
Rb
0 ≤ a < b the quantity a n(x, t) dx is the number of cells of size between a and
b at time t. The cell growth process is modelled by a modified Fokker-Planck
equation of the form
∂ ∂
n(x, t) = − (gn(x, t)) + α2 Bn(αx, t)
∂t ∂x
− (B + µ) n(x, t), (1.1)
where g is the rate of growth, µ is the rate of death, and B is the rate at which
cells divide into α equally sized daughter cells. Here, α > 1 is a constant. For
∗ Ben-Gurion
University, Israel.
† Massey
University, Institute of Fundamental Sciences, New Zealand.
‡ Massey University, Institute of Information and Mathematical Sciences, New Zealand.

Corresponding author: email g.c.wake@massey.ac.nz.

1
the original model that we study here g, µ and B are positive constants. The
partial differential equation (1.1) is supplemented by the boundary conditions

lim n(x, t) = 0; (1.2)


x→∞

lim n(x, t) = 0; (1.3)
x→∞ ∂x
n(0, t) = 0. (1.4)

The steady size distributions (SSDs) for the number density function correspond
to solutions of the form n(x, t) = N (t)y(x) (i.e., separable solutions). Solutions
of this form yield

N ′ (t) gy ′ (x) By(αx)


= − + α2 − (B + µ)
N (t) y(x) y(x)
= Λ,

where Λ is a constant of separation and ′ denotes differentiation with respect to


the indicated argument. The above relation yields

N (t) = N0 eΛt , (1.5)

where N0 is a constant, and the equation

−gy ′ (x) + α2 By(αx) − (B + µ) y(x) = Λy(x). (1.6)

Under suitable normalization (i.e. choice of N0 ) a solution y ∈ L1 [0, ∞) to


equation (1.6) corresponds to a probability density function in the model. The
boundary conditions (1.2)- (1.4) imply that

lim y(x) = 0, (1.7)


x→∞

lim y (x) = 0, (1.8)
x→∞
y(0) = 0, (1.9)

and requirement that y be a probability density function leads to the conditions


y(x) ≥ 0 for all x ∈ [0, ∞) and
Z ∞
y(x) dx = 1. (1.10)
0

Integrating equation (1.6) from 0 to ∞ gives

Λ = (α − 1)B − µ,

and equation (1.6) reduces to

y ′ (x) + αBy(x) − α2 By(αx) = 0. (1.11)

2
Equation (1.11) is a special case of the pantograph equation, which has been
studied extensively. A detailed analysis can be found in [9]. The pantograph
equation has found applications ranging from a partition problem in number
theory to the collection of current in an electric train. The reader is directed to
[8] for an overview of the literature and further analysis of the equation.
There are two other problems where the pantograph equation plays a central
rôle that have a distinct statistical flavour, viz. the absorption of light in the
Milky Way [1] and a ruin problem in risk theory [4]. Although these problems
seem distant from the cell growth model, there is nonetheless a concrete link:
all these models are based on the same type of pseudo Poisson process; conse-
quently, they have the same limit distribution. The link is more transparent
using an approach of Derfel [3] that is based on a probabilistic technique. In
the next section we detail this approach and recover some known results about
the cell growth model in a fundamentally different framework. Specifically, we
give a straightforward proof of the existence of an SSD, the derivation of the
pantograph equation for this distribution (invariant measure) and a solution in
the form of a sequence of probability distribution functions.

2 Limit Distribution for Cell Growth


Consider a spatially homogeneous population of cells and suppose that the size
x of a cell grows linearly as a function of time. Suppose further that at random
moments defined by a Poisson process a cell of size x splits into α new cells of
size x/α. Here, α ≥ 1. Specifically, we suppose that the jumps x → x/α occur
in random moments
0 = t0 < t1 < · · · < tn < · · · ,
where the sequence {τn } defined by
τn = tn+1 − tn ,
for n = 1, 2, . . . consists of independently and exponentially distributed random
variables, i.e., for t > 0,
P {τn > t} = e−t .
For simplicity, we assume that the between jumps the cell size x has a unit
rate of growth so that after ∆t time a cell of size x grows to size x + ∆t. This
assumption corresponds to choosing B and g such that αB/g = 1 in the Fokker-
Planck equation. The results detailed below follow mutatis mutandis for a more
general choice of constants.
Lemma 2.1 There exists a limit distribution (invariant measure) for the size
of a cell. This distribution is independent of the initial cell size x0 .
Proof: Consider a cell of initial size x0 that splits (jumps) at random mo-
ments t1 , t2 , . . . , tn , . . .. Let tn− denote the moment immediately before the nth
splitting. The size of a cell at t1− is
x(t1− ) = x0 + τ1 .

3
The cell then splits into α equal parts at t1 so that at t2− the size is
1
x(t2− ) = (x0 + τ1 ) + τ2 .
α
Similarly, at t3−
 
1 1
x(t3− ) = (x0 + τ1 ) + τ2 + τ3
α α
x0 τ2 τ1
= 2
+ τ3 + + 2,
α α α
and in general
x0 τn−1 τ1
x(tn− ) = n−1
+ τn + + · · · + n−1 . (2.1)
α α α
Equation (2.1) shows that there exists a limit distribution for cell size x as
n → ∞ and that this distribution is independent of the initial cell size x0 .
Indeed, the limit distribution function coincides with a probability distribution
function of the random variable
η1 η2 ηn
Z = η0 + + 2 + ··· + n + ··· , (2.2)
α α α
where the ηk are independently, exponentially distributed random variables.


Let 
1 − e−x , x > 0
F (x) = Fτn (x) = 1 − P {τn > x} = (2.3)
0, x ≤ 0,

′ e−x , x > 0
p(x) = pτn (x) = F (x) = (2.4)
0, x < 0.
Denote the probability distribution function (pdf) for (2.2) by z(x) and let
y(x) = z ′ (x). The function y thus corresponds to the probability density func-
tion for (2.2). The next theorem shows that the probability density function
defined by (2.2) satisfies the pantograph equation (1.11) with αB/g = 1 and
y(0) = 0.

Theorem 2.2 The probability distribution function for (2.2) satisfies

z ′ (x) + z(x) = z(αx) (2.5)


z(0) = 0;

the probability density function for (2.2) satisfies

y ′ (x) + y(x) = αy(αx) (2.6)


y(0) = 0.

4
Proof: The proof follows a method developed by Derfel [3], which is based
on the self-similarity of Z. In particular, equation (2.2) can be recast
1 η2 η3  1
Z = η0 + η1 + + 2 + · · · = η0 + Z1 , (2.7)
α α α α
where Z1 has the same distribution as Z.
Given random independent variables w1 and w2 with pdfs G1 (x) and G2 (x)
respectively, the pdf of their sum w1 + w2 is given by the Stieltjes convolution
Z ∞
(G1 #G2 )(x) = G1 (x − t) dG2 (t).
−∞

In addition, for any β > 0 the pdf of βG1 is G1 (x/β). The pdf for Z1 /α is
therefore z(αx) and the pdf for η0 is F (x). Equation (2.7) thus implies

z(x) = z(αx)#F (x). (2.8)

The Stieltjes convolution (2.8) can be expressed as a Laplace convolution. Since


z(x) = 0 for x ≤ 0 and dF (x) = p(x) dx,
Z x
z(αx)#F (x) = z(α(x − t))e−t dt;
0

consequently,
z(x) = z(αx) ∗ p(x), (2.9)
where ∗ denotes the Laplace convolution. Now,
Z x
′ −x
z (x) = z(0)e + α z ′ (α(x − t))e−t dt,
0

and noting that z(0) = 0 integration by parts yields


x
1 x
 −t 
e
Z

−t
z (x) = α − z(α(x − t)) −
z(α(x − t))e dt
α 0 α 0
= z(αx) − z(x).

We thus see that z satisfies (2.5). Equation (2.6) follows immediately from (2.5)
by differentiation noting that z ′ (0) = 0.


3 A Solution Method
Kato and McLeod [9] showed that problems such as (2.5)and (2.6) do not have
unique solutions. Indeed, there are an infinite number of solutions to these prob-
lems. The requirement that solutions to (2.5) are also probability distribution

5
functions, however, resolves this uniqueness problem. In essence, there is only
one solution to (2.5) such that

lim z(x) = 1. (3.1)


x→∞

A similar comment applies to problem (2.6) if condition (1.10) is imposed. These


uniqueness results can be found in [4] and [6].
Problems such as (2.5) and (2.6) can be solved using Dirichlet series or, what
leads to the same thing, Laplace transforms (cf. [6], [8] and [9]). Solutions to
these problems can thus be expressed in the form

X
an e−α x ,
n
z(x) =
n=1
X∞
−αn an e−α x .
n
y(x) =
n=1

The probabilistic interpretation detailed in Section 2, however, brings to the


fore a different solution method. This method entails a sequence generated by
convolutions. We focus exclusively on the probability distribution function.
Let {zn } be the sequence defined by

z0 (x) = 1 − e−x
zn+1 (x) = zn (αx)#F (x), (3.2)

where n ≥ 0 and x ≥ 0. We show that this sequence converges to a probability


distribution function z that is a solution to problem (2.5). One advantage of
this method is that the approximations to the solution preserve the statistical
structure of the problem. Each term of the sequence is a pdf, and it is clear from
the definition of the sequence that zn corresponds to the pdf for the random
variable at the nth splitting.

Lemma 3.1 The sequence {zn } converges uniformly on intervals of the form
I = [0, a], where a > 0.

Proof: We note first that

zn+1 (x) = zn (αx)#F (x)


Z x
= zn (α(x − ξ))e−ξ dξ. (3.3)
0

Since z0 is continuous on [0, ∞) it is clear that zn is also continuous on this


interval for all n ≥ 1. For any continuous function f : [0, ∞) → R and b > 0 let

kf kb = sup |f (ξ)|.
ξ∈[0,b]

6
It is sufficient to show that the series

X
(zn+1 (x) − zn (x)) (3.4)
k=0

is uniformly convergent on I. For x ∈ I,


Z x
|zn+1 (x) − zn (x)| ≤ |zn (α(x − ξ)) − zn−1 (α(x − ξ))| e−ξ dξ
0
≤ kzn − zn−1 kαx Λ,

where
Λ = 1 − e−a .
The above calculation can be repeated to show that

|zn+1 (x) − zn (x)| ≤ kz1 − z0 kαn a Λn . (3.5)

We have
eαx  (α−1)x 
z1 (x) = z0 (x) − e −1 , (3.6)
α−1
so that for all x ∈ [0, ∞)
2
|z1 (x) − z0 (x)| ≤ . (3.7)
α−1
Inequalities (3.5) and (3.7) thus give
2
kzn+1 − zn kαn a ≤ Λn . (3.8)
α−1
Since 0 < Λ < 1, the Weierstrass M test can be used to show that the series
converges uniformly on I.


Lemma 3.2 Each term of the sequence {zn } is a pdf that is differentiable on
[0, ∞). The limit of the sequence is also a pdf.

Proof: The sequence {zn } is defined by a convolution with a pdf F (x). Since
z0 (x) is a pdf, z0 (αx) is also a pdf. Now, z1 (x) = z0 (αx)#F (x). Since z1 is
defined by the convolution of two pdfs, z1 must also be a pdf (cf. [10], pg. 37).
The argument can be repeated to show that zn must be a pdf for all n ≥ 1.
Each zn is continuous on [0, ∞) Equation (3.3) and the Fundamental Theorem
of Calculus therefore imply that the zn are differentiable and
Z x

zn+1 (x) = α zn′ (α(x − ξ))e−ξ , dξ. (3.9)
0

7
Lemma 3.1 shows that there is a z such that zn (x) → z(x) as n → ∞ for
x ∈ [0, ∞). To show that z must be a pdf we study the characteristic functions
associated with the zn . The characteristic function of zn is given by
Z ∞
φn (t) = eitξ zn′ (ξ) dξ.
0

Equation (3.2) implies that

φn+1 (t) = Qn (t)ψ(t),

where Qn is the characteristic function for zn (αx) and


1
ψ=
1 − it
is the characteristic function for F . Now,
Z ∞
Qn (t) = eitξ dzn (αξ)
0
Z ∞
= eitξ/α zn′ (ξ) dξ
0
= φn (t/α);

therefore,
1
φn+1 = φn (t/α),
1 − it
and consequently
n+1
Y −1
it
φn+1 = 1− .
αk
k=0

The product
∞  −1
Y it
1− k
α
k=0

converges uniformly on all compact intervals of R; hence,


∞  −1
Y it
φn (t) → φ(t) = 1− k ,
α
k=0

where φ is continuous on R. We can now appeal to a standard result in prob-


ability theory (cf. [10], pg. 42) that guarantees the existence of a unique pdf z
corresponding to φ; moreover, zn → z as n → ∞.


Theorem 3.3 Let z denote the limit of the sequence defined by (3.2). Then z
is the unique solution to equation (2.5).

8
Proof: Integrating the right hand side of equation (3.9) by parts gives

zn+1 (x) = zn (αx) − zn+1 (x). (3.10)

Now,

|zn+1 (x) − zn′ (x)| = |zn (αx) − zn+1 (x) − (zn−1 (αx) − zn (x))|
≤ |zn (αx) − zn+1 (x)| + |zn−1 (αx) − zn (x)|,

and since {zn } is uniformly convergent on I, the above inequality shows that
{zn′ } is uniformly convergent on I. We thus have that z is differentiable on I
and that zn′ → z ′ as n → ∞. Equation (3.10) therefore implies that z is a
solution to equation (2.5).
The uniqueness of solutions to equation (2.5) satisfying the given boundary
conditions has been established in [4] and [6]. For completeness, however, we
give a proof.
Suppose that there are two distinct solutions z and w to the boundary-value
problem. Let δ = z − w. Then

δ ′ (x) = δ(αx) − δ(x) (3.11)


δ(0) = 0 (3.12)
lim δ(x) = 0. (3.13)
x→∞

By hypothesis, the solutions are distinct and therefore δ(x) 6= 0 for some x > 0.
Without loss of generality we can assume that δ(x) > 0 for some x > 0. Now, δ
is differentiable, a fortiori, continuous for x ≥ 0, and the boundary conditions
(3.12) and (3.13) imply that there exists a global maximum M at some point
0 < xm < ∞. We thus have M = δ(xm ) > 0 and δ ′ (xm ) = 0. Equation
(3.13) implies that δ(αxm ) = δ(xm ); therefore, the global maximum must also
be achieved at αxm . The arguments can be repeated to show that the global
maximum is achieved at αn xm for all n ≥ 0; consequently, δ(αn xm ) = M . Since
M 6= 0, we have the contradiction that limx→∞ δ(x) 6= 0. We thus conclude
that δ(x) = 0 for all x > 0.


Acknowledgements: The publication was completed while an author (GCW)


was a Visiting Fellow at OCCAM (Oxford Centre for Collaborative Applied
Mathematics) under support provided by Award No. KUK-C1-013-04, made
by King Abdullah University of Science and Technology (KAUST).

References
[1] Ambartsumyan, V.A., “On the fluctuation of the brightness of the Milky
Way”, Dokl. Akad. Nauk SSSR, vol. 44, pp. 223-226, 1944.
[2] Basse B., Baguley B.,, Marshall E., Joseph W., van Brunt B., Wake G.,
Wall D., “Modelling Cell Death in Human Tumour Cell Lines Exposed to

9
the Anticancer Drug Paclitaxel”, J. Math. Bio. vol. 49 no. 4, pp. 329-357,
2004.
[3] Derfel, G., “Probabilistic method for a class of functional differential equa-
tions”, Ukrain. Mat. Zh. vol. 41, pp. 1322-1327, 1989. English translation
Ukrainian Math. J., vol. 41, pp. 1137-1141, 1990.
[4] Gaver, D.P., “An absorption probablility problem”, J. Math. Anal. Appl.,
vol. 9, N3, pp. 384-393, 1964.
[5] Hall, A.J., Wake, G.C., & Gandar, P.W., “Steady size distributions for
cells in one dimensional plant tissues”, J. Math. Bio., vol. 30, No. 2, pp.
101-123,1991.
[6] Hall, A.J. & Wake, G.C., “A functional differential equation arising in the
modelling of cell-growth”, J. Aust. Math. Soc. Ser. B, vol. 30, pp. 424-435,
1989.

[7] Hirshman, I.I., & Widder, D.V., The Convolution Transform, Princeton
University Press, Princeton, 1955.
[8] Iserles, A., “On the generalized pantograph functional differential equa-
tion”, Euro. Jour. Appl. Math., vol. 4, pp. 1-38, 1993.

[9] Kato T. & McLeod, J.B., “The functional-differential equation y ′ (x) =


ay(λx) + by(x)”, Bull. Amer. Math. Soc., vol. 77, pp. 891-937, 1971.
[10] Pitt, H.R., Integration Measure and Probability, Oliver & Boyd, 1963.

10
RECENT REPORTS
2009
03/09 On the modelling of biological patterns with mechanochemical Moreo
models: insights from analysis and computation Gaffney
Garcia-Aznar
Doblare

04/09 Stability analysis of reaction-diffusion systems with timem- Madzvamuse


dependent coefficients on growing domains Gaffney
Maini

05/09 Onsager reciprocity in premelting solids Peppin


Spannuth
Wettlaufer

06/09 Inherent noise can facilitate coherence in collective swarm motion Yates et al.

07/09 Solving the Coupled System Improves Computational Efficiency Southern


of the Bidomain Equations Plank
Vigmond
Whiteley

08/09 Model reduction using a posteriori analysis Whiteley

09/09 Equilibrium Order Parameters of Liquid Crystals in the Laudau- Majumdar


De Gennes Theory

10/09 Landau-De Gennes theory of nematic liquid crystals: the Oseen- Majumdar
Frank limit and beyond Zarnescu

11/09 A Comparison of Numerical Methods used for Finite Element Pathmanathan


Modelling of Soft Tissue Deformation Gavaghan
Whiteley

12/09 From Individual to Collective Behaviour of Unicellular Organisms: Xue


Recent Results and Open Problems Othmer
Erban

13/09 Stochastic modelling of reaction-diffusion processes: algorithms Erban


for bimolecular reactions Chapman

14/09 Chaste: a test-driven approach to software development for phys- Pitt-Francis et al.
iological modelling

15/09 Block triangular preconditioners for PDE constrained optimiza- Rees


tion Stoll

16/09 From microscopic to macroscopic descriptions of cell migration on Baker


growing domains Yates
Erban
17/09 The Influence of Gene Expression Time Delays on Gierer- Seirin Lee
Meinhardt Pattern Formation Systems Gaffney
Monk

18/09 Analysis of a stochastic chemical system close to a sniper bifurca- Erban et al.
tion of its mean field model

19/09 On the existence and the applications of modified equations for Zygalakis
stochastic differential equations

20/09 Pebble bed: reflector treatment and pressure velocity coupling Charpin et al.

21/09 A finite difference method for free boundary problems Fornberg

22/09 Tangent unit-vector fields: nonabelian homotopy invariants and Majumdar


the Dirichlet energy Robbins
Zyskin

23/09 Morphological instability of a nonequilibrium icecolloid interface Peppin


Majumdar
Wettlaufer

24/09 The effect of polar lipids on tear film dynamics Aydemir


Breward
Witelski

25/09 Preconditioning for active set and projected gradient methods as Stoll
semi-smooth Newton methods for PDE-constrained optimization Wathen
with control constraints

26/09 Functional differential equations arising in cell-growth Wake


Begg

Copies of these, and any other OCCAM reports can be obtained


from:
Oxford Centre for Collaborative Applied Mathematics
Mathematical Institute
24 - 29 St Giles’
Oxford
OX1 3LB
England
www.maths.ox.ac.uk/occam

View publication stats

You might also like