J. Math. Biol.

46, 48–70 (2003)

Digital Object Identifier (DOI):
Mathematical Biology
10.1007/s00285-002-0170-x

Patrick De Leenheer · Hal Smith

Feedback control for chemostat models

Received: 1 November 2001 / Revised version: 19 August 2002 /
Published online: 19 November 2002 –

c Springer-Verlag 2002

Abstract. It is shown that a chemostat with two organisms can be made coexistent by
means of feedback control of the dilution rate. Remaining freedom in the feedback law can
be used to guarantee robustness or improve particular performance indices. Unfortunately
a topological property prevents coexistence by feedback control for chemostats with more
than two organisms. We apply our results to control bioreactors aimed at producing com-
mercial products through genetically altered organisms. In all our results the coexistence
takes its simplest form: a global asymptotically stable equilibrium point in the interior of the
non-negative orthant.

1. Introduction
The chemostat is a benchmark model in microbial ecology, used for a wide variety
of systems ranging from lakes, waste-water treatment, to reactors for commercial
production of substances by genetically altered organisms. It has the advantage of
being easily implementable in a laboratory and hence the model has been subject to
extensive tests and experiments. It may be used to model the competition of several
organisms for a single nutrient source. A classical result is the ‘competitive exclu-
sion principle’ [22], stating that in the long run only one organism survives while
the others die out. There is a large literature devoted to modifying the chemostat
model to ensure coexistence of the organisms. These are based on suitable manipu-
lation of the chemostat’s operating parameters, the dilution rate [24, 2] or the input
nutrient concentration [24, 9, 5], or by dropping the well-mixed hypothesis [23,
16]. If instead of being constant, an operating parameter is a periodically varying
function of time, coexistence may occur (see [22] for a review).
In this paper we take a different approach based on ideas from control theory.
Two control principles studied in control theory are open-loop control and feedback
control. For results based on the former we refer to [18]. We consider the dilution
rate as a feedback control variable, keeping the input nutrient concentration at a
fixed value. It is shown that if the dilution rate depends affinely on the concen-
trations of two competing organisms, coexistence may be achieved as a globally
asymptotically stable equilibrium point in the interior of the non-negative orthant.
The dilution rate is proportional to the speed of the pump supplying the chemostat

P. De Leenheer, H. Smith: Arizona State University, Department of Mathematics, Tempe,

AZ 85287, USA. e-mail: leenheer@math.la.asu.edu; halsmith@asu.edu
Mathematics Subject Classification (2000): 34H05, 34D23, 93C10
Keywords or phrases: Chemostat – Feedback control – Coexistence
Feedback control for chemostat models 49

with the medium which contains the input nutrient. If accurate measurements of the
concentrations of the competing organisms are available, this implies that a simple
affine control algorithm would allow automated coexistence in the chemostat. In
addition we show that our feedback approach is robust, a feature which open-loop
control methods lack. Actually this difference holds for far more general problems
(e.g. stabilization problems) than the one considered here and in part explains the
success of feedback control when compared to open loop control paradigms. Also
our method is constructive as an explicit feedback law will be provided whereas
only existence of an open-loop control is shown in [18]. Unfortunately we show that
coexistence is not achievable by means of feedback control of the dilution rate for
a chemostat with more than two competing organisms due to a topological obstruc-
tion, while open-loop controls which do achieve this goal are shown to exist in [18].
We apply our results to control bioreactors aimed at producing a commercial
product using genetically altered organisms. Organisms are typically altered by
insertion of a small circle of DNA (a plasmid) coding for the production of the
product and perhaps other useful attributes. Unfortunately, these plasmids may be
unevenly divided between daughter cells during cell division leading occasionally
to the reappearance of the wild-type organism containing no plasmid. Therefore,
one is forced to deal with (i.e., feed) two organisms-the genetically altered one pro-
ducing useful product and the wild-type which is useless for the intended purpose.
Several control strategies have been employed to keep the wild-type organism,
typically a strong competitor against the altered strain, to acceptably low levels,
each with its own set of disadvantages. For example, the plasmid may also code
for resistance to an antibiotic which is then added to the chemostat to select for
the resistant, product-producing, strain [14, 17, 12]. In a similar strategy, the pla-
smid may also code for resistance to a growth inhibitor, supplied in the feed, for
which the wild-type is not resistant [10, 13, 12]. Or, the plasmid may code both
for production of and resistance to a toxin for which the wild-type is susceptible
[15,12]. While these strategies work to control the wild-type organism, the antibiot-
ic or toxin may contaminate the product requiring an expensive separation process.
We propose an alternative control strategy which requires no chemical agents and
expensive post-production separation process. It does, however, require fast and
accurate measurements of cell densities in the effluent in order to implement. It
will be shown that feedback control can be used to control wild-type population
levels and that it achieves superior robustness properties.
In section 2 we introduce the chemostat model. Controlling the chemostat by
means of the dilution rate is discussed in section 3. Section 4 contains the main
result on coexistence of two organisms by means of feedback control of the dilution
rate. In section 5 it is shown that feedback control of the dilution rate is futile for
more than two organisms due to a topological obstruction and we apply our results
for control of bioreactors in section 6. Section 7 contains the proofs.

2. Model of the chemostat

The model of a chemostat is given by the following set of differential equations:

50 P. De Leenheer, H. Smith

2
xi
Ṡ = D(S 0 − S) − fi (S)
γi
i=1
(1)
ẋi = xi (fi (S) − D), i = 1, 2
where
1. S(t) is the concentration of nutrient and xi (t) is the concentration of organism
i in the chemostat at time t.
2. D is the dilution rate of the chemostat (or equivalently, 1/D is the residence
time of a molecule inside the chemostat) and S 0 is the concentration of the
input nutrient. The constants γi are yield constants.
3. The functions fi are called uptake functions and satisfy the following:
• (regularity) fi : R+ → R+ is continuously differentiable and fi (0) = 0.
• (monotonicity) fi is monotonically increasing, i.e. fi > 0 for all S ∈ R+ .
A minimal requirement for any model of a chemostat is that the state components
S, x1 and x2 only take non-negative values for all times t ≥ 0. It can be shown that
system (1) satisfies this requirement. It is convenient to pass to non-dimensional
variables S̄ := S/S 0 , x̄ := x/(γi S 0 ):


2
S̄˙ = D(1 − S̄) − x̄i fi (S 0 S̄)
i=1
x̄˙ i = x̄i (fi (S 0 S̄) − D), i = 1, 2
or, after dropping the bars and writing fi (S) instead of fi (S 0 S) (note that both func-
tions are qualitatively the same and therefore the ‘new’ fi (S) also satisfy regularity
and monotonicity properties):


2
Ṡ = D(1 − S) − xi fi (S)
i=1
(2)
ẋi = xi (fi (S) − D), i = 1, 2
The equilibrium points of system (2) are:
E0 := (1, 0, 0), E1 := (λ1 , 1 − λ1 , 0) and E2 := (λ2 , 0, 1 − λ2 ) (3)
where the λi – which are assumed to be different – are implicitly defined as follows:
fi (λi ) = D, i = 1, 2 (4)
The principal result concerning the chemostat is the so-called ‘competitive exclu-
sion principle’ and can be stated as follows, see [22]:
Theorem 1. If 0 < λ1 < 1 and if λ1 < λ2 then E1 is a globally asymptotically
stable equilibrium point of system (2) with respect to all initial conditions in the
set {(S, x1 , x2 ) ∈ R+
3 | x > 0, i = 1, 2}.
i

Informally, the exclusion principle states that when two organisms compete for
a single nutrient, one of the organisms is doomed in the long run and loses the com-
petition. This principle also holds in case there are multiple organisms competing
for the nutrient, see also [22].
Feedback control for chemostat models 51

3. Controlling the chemostat by means of the dilution rate

In view of the exclusion principle, one might wonder whether it is possible to

change the long term behavior of the chemostat and make the organisms coexist.
The obvious parameters that are manipulable by the operator running the chemostat
are the dilution rate D and the input nutrient concentration S 0 . In this paper we
will assume that S 0 is fixed and D is manipulable. The intuition behind the idea of
manipulating the dilution rate to achieve coexistence is outlined next.
We will assume throughout the rest of this paper that the graphs of the uptake
functions fi , i = 1, 2 are as depicted in Figure 1. Therefore we introduce the
following standing hypothesis:
H The graphs of the functions f1 and f2 intersect once at S̃:

f1 (S̃) = f2 (S̃) = D̃ (5)

where S̃ ∈ (0, 1). For all S ∈ (0, S̃) the inequality f1 (S) > f2 (S) holds, while for
all S > S̃, f1 (S) < f2 (S) holds. Moreover f2 (S̃) > f1 (S̃).
It follows from theorem 1 that for low values of the dilution rate (D < D̃)
such as Da , organism 1 wins the competition and for higher values (D > D̃, but
D < Dmax ) such as Db , organism 2 wins the competition.
If the goal is coexistence it is therefore tempting to vary the dilution rate between
low and high values. The variation is achievable in different ways:
1. One way is to make the dilution rate a time-dependent function. A particular
case occurs for periodically varying dilution rates: D ≡ D(t) = D(t + T ) for
all t ∈ R and for some T > 0. It has been shown in [22] that in this case it is
indeed possible to make both organisms coexist. Coexistence is expressed by
the existence of a globally attracting periodic solution of system (2) which lies
in int(R+3 ).

2. In this paper we propose a different approach. Instead of letting the dilution

rate depend on time, we let it depend on the state of the chemostat:

D ≡ D(S, x1 , x2 ) (6)

Dmax f2 (S)
Db f1 (S)
D̃
Da

S
0 λa S̃ λb 1

Fig. 1. Different dilution rates result in different winners.

52 P. De Leenheer, H. Smith

where D(S, x1 , x2 ) is some function defined on R+ 3 . Of course this function

should only take non-negative values for obvious physical reasons (dilution
rates cannot be negative), implying that D(S, x1 , x2 ) : R+3 →R .
+
This approach amounts to interpreting the dilution rate as a state feedback con-
trol variable: Its value at a certain moment, depends on the current state of the
chemostat.

Remark 1. We comment briefly on implications of using state feedback.

A mathematical simplification might be anticipated when using state feedback: Pe-
riodically varying dilution rates result in a periodically varying chemostat, while
state feedback leads to a time-independent chemostat.
On the other hand there is a technological problem related to using feedback
control. The current state has to be known to determine the current dilution rate.
But it might be difficult to measure all state variables in an operating chemostat.
For instance, it may not be reasonable to assume that a measurement of the nutrient
concentration is available. We show below that D may be chosen to depend only
on x1 and x2 and not on S. In general the search for or design of feedback laws
depending on all state variables should not be considered as useless, even if it is
not possible to measure all state variables. Suppose that a feedback law has been
found which requires the full state of the system, but that it is impossible to obtain
measurements of say the concentration of the nutrient. The solution to this problem
proposed in control theory is then to find a so-called observer. This is a dynami-
cal system which produces an estimate of the nutrient concentration based on the
available information of the state of the system. Provided an observer can be found,
the next step is then to use this estimate in the original feedback law instead of the
actual value. Of course the challenge is then to prove that the desired goal -which
may be to make the system stable or coexistent for instance- is still achieved.

4. Permanence by means of feedback control

Although we have been using the term coexistence rather informally in the previ-
ous sections and equated it with the survival of all organisms in the long run, we
will instead use the well-known concept of permanence, see e.g. [8], in the rest of
this paper. But first we define positive systems. Consider a system ẋ = f (x) where
x ∈ Rn and f is sufficiently smooth such that existence and uniqueness of solutions
is guaranteed. A (forward) solution starting in x0 at t = 0 is denoted by x(t, x0 )
and is defined on its maximal (forward) interval of existence I(x0 ) (I + (x0 )). The
system is called positive if R+n is a forward invariant set for it, i.e. if

∀x0 ∈ R+
n
: x(t, x0 ) ∈ R+
n
, ∀t ∈ I + (x0 ).

Definition 1. A positive system ẋ = f (x) is permanent if there exists a compact

set K, K ⊂ int(R+n ), which attracts all solutions starting in int(Rn ), i.e.
+

∀x0 ∈ int(R+
n
), ∃T (x0 ) : x(t, x0 ) ∈ K, ∀t ≥ T (x0 ). (7)
Feedback control for chemostat models 53

Permanence might come in different forms. The simplest manifestation occurs if

a positive system possesses an equilibrium point in int(R+ n ) which is globally as-
n ).
ymptotically stable (GAS) with respect to initial conditions in int(R+
Motivated by the discussion in the previous section, we formulate the following
problem:

Permanence Problem: Find – if possible – a feedback D : R+

3 → R such that
+
the following system:

2
Ṡ = D(x)(1 − S) − xi fi (S)
i=1
(8)
ẋi = xi (fi (S) − D(x)), i = 1, 2

where x := (S, x1 , x2 )T , is permanent.

Now we are in a position to state the main result of this section.

Theorem 2. Pick any  in the interval [0, D̃). Then the permanence problem is
solvable by the following affine feedback:

D(x) = k1 x1 + k2 x2 +  (9)

if the gains ki ≥ 0, i = 1, 2 satisfy the following inequalities:

k2 < k̃ < k1 (10)

where k̃ := D̃−
and where k2 > 0 if  = 0.
1−S̃
In particular there exists an equilibrium point x e ∈ int (R+ 3 ) which is GAS for

system (8) with feedback (9) with respect to initial conditions in int(R+3 ).

Conversely, if system (8) with a nonnegative affine feedback D(x) = k1 x1 +

k2 x2 +  possesses a unique equilibrium point in int (R+3 ) which is locally asymp-

totically stable, then the feedback parameters satisfy  < D̃ and (10).

Remark 2. Notice that with feedback (9), no knowledge on the nutrient concentra-
tion S(t) is needed to determine the current dilution rate D(x(t)).
On the other hand the concentration x1 (t) and, in general also x2 (t), have to
be known to determine the dilution rate. Only if k2 = 0 – which is an acceptable
choice if  > 0 since it is in accordance with (10) – is it not necessary to determine
x2 .
Moreover since (10) implies that k1 = k2 , any device used to monitor the con-
centrations of organisms (with the intention of using them to determine the current
value of the dilution rate) must be capable of distinguishing between them. For
methods used in practice to measure concentrations we refer to [6, 11, 1].

Remark 3. Recall that for obvious physical reasons, the dilution rate D(x) should
take non-negative values for all x ∈ R+3 . Since  ≥ 0 and k ≥ 0, i = 1, 2, this
i
condition is indeed satisfied. Moreover, recalling that the dilution rate is propor-
tional to the speed of the pump which fills the chemostat, this feedback exhibits a
54 P. De Leenheer, H. Smith

 + k1
l
l
l
 + k̃ l
l+ 
y = k̃(1 − S)
l y = k1 (1 − S) + 
 + k2 aa l
a − S) +  l
y = k2 (1a
f2 (S)
aa l f1 (S)
D̃ aa l
aal
al aa

laa
l a
0 S̃ 1 S

Fig. 2. Bound for k1 and k2 .

desirable property: If the concentrations of the organisms inside the chemostat are
low (x1 and x2 are small), then the dilution rate and thus the speed of the pump is
low. On the contrary, if these values are high, then the pump speed is also high.

Remark 4. It is instructive to consider the equations governing the species con-

centrations of the system resulting from the insertion of the feedback law (9) into
equations (8):

ẋi = xi (fi (S) − k1 x1 − k2 x2 − ), i = 1, 2

Together with an equation for S with constant dilution rate and with terms repre-
senting nutrient recycling of dead biomass, this model is similar to that studied in
e.g. [4, 19] where terms like −k1 x1 − k2 x2 have been interpreted as representing
intra- and inter-specific “interference” competition. There, it is appropriate to dou-
ble index the k as kij . In this sense it is possible to rephrase the resulting action of
our feedback law into a form which is more familiar to mathematical biologists.
Similarly it may inform control theorists of the physical meaning of their feedback
control laws for these models.

Remark 5. A simple graphical interpretation for the inequalities (10) follows from
Figure 2. Denoting the vertical axis in this figure as y, the intersection of the line
y = k1 (1 − S) + , respectively y = k2 (1 − S) + , and the y-axis should be above,
respectively below, the intersection of the line y = k̃(1 − S) +  and the y-axis.

Remark 6. We emphasize that the constraints (10) on the feedback gains entail
some freedom for the feedback law (9). This freedom might be used for different
purposes:
Feedback control for chemostat models 55

k1 + 

k̃ u + 
\
\
\
\
\
\
\
\
\
\ f2u
\ f2
D̃ u
\
k̃ l +  D̃ hhhh \
f1
D̃ l h hhhhh \ f1l
f1u hhhh
k2 + 
 \ hhhh
\ hhhh
hhhh
2fl \ h
0 S̃ l S̃ S̃ u 1 S

Fig. 3. Robustness.

1. Robustness
In many applications the uptake functions are not exactly known, and only
bounds of these functions are available. Denote the upper bounds by fiu (S) and
the lower bounds by fil (S), see Figure 3:

∀S ∈ R+ : fil (S) ≤ fi (S) ≤ fiu (S), i = 1, 2 (11)

It is natural to assume that these bounds also satisfy regularity and monotonicity
conditions.
An important consequence of not knowing the uptake functions exactly is that
the value S̃, needed in the construction of the feedback which achieves per-
manence, is not known either. Instead, only a compact interval I := [S̃ l , S̃ u ]
(which is assumed to be a subset of (0, 1)) is known to contain S̃, see also Figure
3. The limits of the interval I are implicitly given by the following equalities:

f1l (S̃ l ) = f2u (S̃ l ) (:= D̃ l )

(12)
f1u (S̃ u ) = f2l (S̃ u ) (:= D̃ u )

and we assume (this assumption might be relaxed, but we prefer not do to so

because it would not change the argument) that:

∀S ∈ (0, S̃ l ) : f2u (S) < f1l (S)

∀S > S̃ u : f1u (S) < f2l (S)
56 P. De Leenheer, H. Smith

This lack of exact information on the uptake functions might endanger the ca-
pability of the feedback law (9) to render the system (8) permanent. However,
it turns out that if (10) are replaced by:
k1 > k̃ u and k2 < k̃ l (13)

where k̃ u := D̃ −u and k̃ l := D̃ −l , then the feedback law (9) with ki ≥ 0,
u l

1−S̃ 1−S̃
i = 1, 2, still solves the permanence problem. The price paid for this robust-
ness property is a smaller range of feedback gains which is immediate from a
comparison of (10) and (13).

2. Performance indices
The remaining freedom might also be used to modify particular performance
indices, associated to the behavior of the controlled chemostat.
An obvious performance index is the ratio of the values of both organisms at
the interior equilibrium point x e . The proof of theorem 2 will reveal that this
ratio is given by:

x1e k̃ − k2
= (14)
x2e k1 − k̃
showing that any value can be assigned to it without violating the gain con-
straints (10).

Remark 7. Next we give an intuitive explanation of why theorem 2 holds. First

recall from the discussion in the previous section that if D is set to a low value,
then organism 1 wins the competition because the single-organism equilibrium
point corresponding to organism 1 is globally asymptotically stable. Similarly, if
D equals a high value, then the single-organism equilibrium point corresponding
to organism 2 is globally asymptotically stable and organism 2 wins.
Although the single-organism equilibrium points depend on the choice of D,
they should be unstable to achieve permanence. To destabilize the single-organism
equilibrium point corresponding to organism 1, D should take large values in a
neighborhood of this equilibrium point (for if it was small, one would expect it to
be stable). Similarly, to destabilize the equilibrium point corresponding to organism
2, D should take small values in a neighborhood of it. Both conditions are satisfied
for the linear feedback (9) in case (10) holds: Close to the single-organism equilib-
rium point corresponding to organism 1, the dilution rate is close to k1 x1 + , which
is large if k1 is large. A similar argument shows that at the other single-organism
equilibrium point, the dilution rate is small if k2 is small.

Remark 8. The simplest and easiest feedback to implement would be to let k1 =

k2 = k so that D(x) = k(x1 + x2 ) +  requiring no distinction between organ-
isms by the observer. In fact, just this system, with  = 0 and piecewise linear
uptake functions, was numerically simulated in [3], where it is pointed out that a
flow reactor with such negative feedback on the dilution rate is commonly used to
regulate population size and is called a turbidostat. The turbidostat may be useful
Feedback control for chemostat models 57

for pure culture population regulation but it cannot maintain permanence for two
distinct competing species. Indeed, coexistence equilibria can only exist if k takes
a particular value, namely, k = D̃− , determined by the point of intersection of the
1−S̃
growth curves and by the supplied nutrient concentration (here scaled to unity), and
in that case there is a line of equilibria given by x1 + x2 = 1 − S̃. Random effects
can then be expected to eventually eliminate one of the competitors. However, as
pointed out by [3], extinction may take a very long time.

5. Multiple organisms

We have shown that a chemostat with two competing organisms can be made per-
manent by means of feedback control of the dilution rate. Unfortunately this is no
longer true for chemostats with more than two competitors. The reason for this
failure is one of topological nature: A permanent chemostat should possess at least
one equilibrium point in the interior of the positive orthant. But no matter what feed-
back law is chosen for the dilution rate, chemostats with more than two competing
species (generically) cannot satisfy this property as we show next.
The (non-dimensional) model for a chemostat with multiple, say n, competitors
is readily obtained from the two competitors model and takes the following form:

n
Ṡ = D(x)(1 − S) − xi fi (S)
i=1
(15)
ẋi = xi (fi (S) − D(x)), i = 1, ..., n

where x := (S, x1 , ..., xn )T is the state vector consisting of the concentrations of

the nutrient and the n organisms. As before the n uptake functions fi are assumed to
satisfy regularity and monotonicity conditions and the problem is to find a suitable
feedback function for the dilution rate D : R+ n+1
→ R+ ensuring permanence for
the resulting chemostat.
Notice that if n ≥ 3 the graphs of the uptake functions in general do not have a
nontrivial common intersection point, unless in the (biologically unlikely and un-
realistic) case that the following set of nonlinear equations has a solution S̃ ∈ R+n

with S̃ = 0:

f1 (S) = f2 (S) = ... = fn (S) (16)

On the other hand, the existence of a solution S̃ ∈ R+ n with S̃ = 0 for the equations

(16) is a necessary condition for the existence of an equilibrium point of system

n+1
(15) in int(R+ ), irrespective of the choice for the feedback function D(x). And
it follows from a generalization of theorem 13.3.1 in [8] (in [8] the result is only
proved for so-called Kolmogorov systems, but a careful reading of the proof re-
veals that is still valid for arbitrary positive systems) that in turn this is a necessary
condition for permanence of system (15).
Summarizing, if n ≥ 3, it is in general not possible to find a feedback function
for the dilution rate which ensures permanence of the resulting system (15).
58 P. De Leenheer, H. Smith

6. Application: Producing genetically altered organisms

A potential application area of feedback controlled chemostats is the commercial

production of substances using genetically altered organisms in continuous culture.
The alteration is obtained by inserting DNA in a cell in the form of a plasmid. These
plasmid-bearing organisms are then grown in a bioreactor, modeled by a chemostat.
Unfortunately the plasmid might be lost in the reproduction process, resulting in
the appearance of a second competitor, the plasmid-free organism. It was shown
in [12] that if the plasmid-free organism is a superior competitor (being a superior
competitor can be made mathematically precise), it takes over the bioreactor and
leads to the extinction of the plasmid-bearing organism. This should of course be
prevented by all means. It is sometimes argued that the plasmid-free organism is a
superior competitor because it is not burdened by the extra load of the plasmid. But
if this is true, then there is no hope for efficient long-term production of the pla-
smid-bearing organisms in view of the mentioned extinction result. Moreover, there
have been reports that under certain conditions the plasmid-bearing competitor is
the superior competitor, see [25] and cited references therein. Coexistence of both
organisms seems to be the best possible outcome. It has indeed been shown in [12]
that this can be achieved. In this section we show that permanence is also achievable
by feedback control and leads to superior robustness properties compared to the
case where the dilution rate is constant.
First we introduce a modified chemostat model to describe the competition
between plasmid-bearing and plasmid-free organisms:

2
xi
Ṡ = D(S 0 − S) − fi (S)
γi
i=1
ẋ1 = x1 ((1 − q)f1 (S) − D) (17)
ẋ2 = x2 (f2 (S) − D) + qx1 f1 (S)

where S(t) is the concentration of the nutrient and x1 (t), respectively x2 (t), is the
concentration of the plasmid-bearing organism, respectively plasmid-free organ-
ism. The parameters appearing in (17) have the same meaning as in (1) and the
uptake functions fi , i = 1, 2, satisfy regularity and monotonicity conditions as
before. The parameter q in (17) represents the probability that the plasmid is lost
in the reproduction process and, being a probability, satisfies:

0<q<1

Passing to non-dimensional variables (recall that input nutrient concentration S 0 is

assumed to be a constant) yields the following model:

2
Ṡ = D(1 − S) − xi fi (S)
i=1
ẋ1 = x1 ((1 − q)f1 (S) − D) (18)
ẋ2 = x2 (f2 (S) − D) + qx1 f1 (S)
Feedback control for chemostat models 59

Dmax (1 − q)f1 (S)

D̃ f2 (S)

S
0 λ2 S̃ 1

Fig. 4. Plasmid-bearing ((1 − q)f1 ), plasmid-free (f2 ) uptake functions.

The following standing hypothesis concerning the uptake functions is similar to H

and illustrated in Figure 4:

H’ The graphs of the functions (1 − q)f1 and f2 intersect once at S̃:

(1 − q)f1 (S̃) = f2 (S̃) = D̃ (19)

where S̃ ∈ (0, 1). For all S ∈ (0, S̃) the inequality (1 − q)f1 (S) < f2 (S) holds,
while for all S > S̃, (1 − q)f1 (S) > f2 (S) holds.
Notice that H’ implies that for high nutrient levels (S̃ < S), the plasmid-bearing
organism is a superior competitor, while for low levels (S̃ > S), the plasmid-free
organism is.

6.1. Constant dilution rate

If the dilution rate D is constant, then the possible equilibrium points of system
(18) are given by:

E0 := (1, 0, 0), E2 := (λ2 , 0, 1 − λ2 ) and E ∗ := (λ∗ , x1∗ , x2∗ ) (20)

The washout equilibrium point E0 always exists, while E2 (an equilibrium point
involving only plasmid-free organisms and no plasmid-bearing organisms) exists
only if the following nonlinear equation:

f2 (S) = D (21)

possesses a solution S ≡ λ2 > 0. Obviously, this depends on the function f2 and

the value of D, see Figure 4. Similarly, the equilibrium point E ∗ only exists if
an appropriate nonlinear set of equations is solvable and possesses a solution in
int(R+3 ).

More importantly, notice that no matter what value D takes, there is no equi-
librium point corresponding to plasmid-bearing organisms only. This is different
from the classical chemostat model we discussed before. It also implies that it
60 P. De Leenheer, H. Smith

is not possible to operate the reactor corresponding to the chemostat model (18)
at a desired state where only plasmid-bearing and no plasmid-free organisms are
present. The best one can hope for then is to operate it at the equilibrium point E ∗ ,
if it exists of course. It turns out that this is indeed possible if the dilution rate is
carefully chosen as we show next by recalling one of the principal results in [12].
In the statement of this result we will need the constant Dmax which is defined as
the value of (1 − q)f1 (1).
Theorem 3. If D > Dmax , then E0 is a GAS equilibrium point of system (18) with
respect to all initial conditions in the set P := {(S, x1 , x2 ) ∈ R+
3 | x > 0, i =
i
1, 2}.
If D ∈ (D̃, Dmax ), then E ∗ is a GAS equilibrium point of system (18) with
respect to all initial conditions in the set P .
If D ∈ (0, D̃), then E2 is a GAS equilibrium point of system (18) with respect
to all initial conditions in the set P .
Theorem 3 states that if the dilution rate D is constant, then washout occurs if D is
too high and loss of plasmid-bearing organism if D is too low. Only for moderate
values of D, permanence is possible and the operator running the bioreactor should
be very careful choosing D.

6.2. Feedback control

If instead of taking constant values for the dilution rate, we allow it to depend on
the state, system (18) is modified to:

2
Ṡ = D(x)(1 − S) − xi fi (S)
i=1
ẋ1 = x1 ((1 − q)f1 (S) − D(x)) (22)
ẋ2 = x2 (f2 (S) − D(x)) + qx1 f1 (S)
where D(x) is a feedback function, defined for all x = (S, x1 , x2 ) ∈ R+ 3 and tak-

ing values in R+ . The obvious problem is to find – if possible – a suitable feedback

function D(x) such that the resulting chemostat model (22) is permanent.
We are ready to state the main result of this section.
Theorem 4. Pick any  in the interval (0, D̃). Then system (22) with the following
affine feedback:
D(x) = k(x1 + x2 ) +  (23)
is permanent if k satisfies the following inequality:
k > k̃ (24)
D̃−
where k̃ := .
1−S̃
In particular there exists an equilibrium point x e ∈ int(R+
3 ) which is GAS for

system (22) with feedback (23) with respect to initial conditions in int(R+3 ).
Feedback control for chemostat models 61

+k
l
l
l
 + k̃ l
l y = k(1 − S) + 
l
l
y = k̃(1 − S) +  l (1 − q)f1 (S)
l f2 (S)
D̃ l
l
 l
l
l
0 S̃ 1 S

Fig. 5. Bound for k.

Remark 9. As before for the classical chemostat, the feedback (23) is independent
of the nutrient concentration S(t). In control theory the feedback law (23) with
constraint (24) is sometimes called a ‘high gain feedback law’.
A comparison of feedback (9) with feedback (23) reveals that there is only one
gain k, instead of two gains k1 and k2 . There are three reasons for restricting the
number of gain constants here:
1. First, it is possible to achieve permanence of the plasmid model (18) by means
of a single gain, while this was not possible for the classical chemostat model
(8), see the inequalities (10) which imply that k1 = k2 .
2. A single gain constant k offers the advantage of cheaper implementation for
the feedback (23). Indeed, in this case D(x) depends only on the sum of the
concentrations of x1 and x2 and not on their respective values; the observer need
not distinguish one strain from another – a potentially difficult and expensive
process. This contrasts to the case of the classical chemostat model, see remark
2, where a more expensive implementation is necessary.
3. A single gain results in a simplification of the analysis of the resulting model.

Remark 10. As before a simple graphical interpretation can be given for the in-
equality (24), see Figure 5: The intersection of the line y = k(1 − S) + , and the
y-axis should be above the intersection of the line y = k̃(1 − S) +  and the y-axis.

Remark 11. Next we outline why feedback control might be advantageous as a con-
trol strategy. The reason is the superior robustness property achieved with feedback
control, compared to the situation where the dilution rate is constant.
Suppose that the functions (1 − q)f1 and f2 are not known exactly, but only
bounds are available which we denote as (1−q)f1l , (1−q)f1u , f2l and f2u , see Figure
6. These functions are assumed to satisfy regularity and monotonicity properties
62 P. De Leenheer, H. Smith

k+

k̃ u + 
\
\
\
\
\
\
\
\
\
\ (1 − q)f1u
l
Dmax \ (1 − q)f1
D̃ u \ f2
D̃ \ f2l
D̃ l
f2u \
 \
\
(1 − q)f1l \
0 S̃ l S̃ S̃ u 1 S

Fig. 6. Robustness.

and, being bounds for (1 − q)f1 and f2 , satisfy the following:

∀S ∈ R+ : (1 − q)f1l (S) ≤ (1 − q)f1 (S) ≤ (1 − q)f1u (S)
(25)
∀S ∈ R+ : f2l (S) ≤ f2 (S) ≤ f2u (S)

Furthermore we assume that instead of S̃, only a compact interval I := [S̃ u , S̃ l ]

(which is assumed to be a subset of (0, 1)) is known to contains S̃, where the bounds
of I are implicitly given by:
f2l (S̃ l ) = (1 − q)f1u (S̃ l ) (:= D̃ l )
(26)
f2u (S̃ u ) = (1 − q)f1l (S̃ u ) (:= D̃ u )
and where it is assumed that:
∀S ∈ (0, S̃ l ) : (1 − q)f1u (S) < f2l (S)
∀S > S̃ u : f2u (S) < (1 − q)f1l (S)
If the dilution rate is constant, it can be shown that permanence is achievable if
D ∈ (D̃ u , Dmax
l ), where D l
max := (1 − q)f1 (1). Dilution rates outside this in-
l

terval do not guarantee permanence. If D̃ l and Dmaxl are very close to each other,
the operator who is controlling the chemostat should be very careful in setting the
dilution rate.
On the contrary if the dilution rate is implemented as a feedback law as in (23),
it can be shown that permanence is achieved if  ∈ (0, D̃ l ) and k > k̃ u := D̃ −u .
u

1−S̃
In other words, if the gain is chosen high enough, permanence is achieved.
Feedback control for chemostat models 63

Remark 12. We have assumed throughout this section that hypothesis H’ was sat-
isfied, implying that the plasmid-free (plasmid-bearing) organism is the superior
competitor for low (high) nutrient concentrations. An equally interesting case oc-
curs when the opposite holds, in which case the assumption should be changed to
the following:
H” The graphs of the functions (1 − q)f1 and f2 intersect once at S̃:

(1 − q)f1 (S̃) = f2 (S̃) = D̃ (27)

where S̃ ∈ (0, 1). For all S ∈ (0, S̃) the inequality (1 − q)f1 (S) > f2 (S) holds,
while for all S > S̃, (1 − q)f1 (S) < f2 (S) holds.
A similar permanence result can be obtained by means of suitably chosen affine
feedback laws. Since the proof of this result is very similar to that of theorem 4, we
have decided to omit it here.
Finally, in theory there are two remaining cases. One is where the plasmid-bear-
ing competitor is always the superior competitor. This is an (biologically) unrealistic
assumption. The other case is the opposite situation, where the plasmid-free com-
petitor is always the superior competitor. But we already mentioned in the beginning
of this section that in this case there is no hope of achieving permanence. Indeed, a
necessary condition for permanence is that at least for some nutrient concentrations,
the plasmid-bearing competitor is superior.

7. Proofs

7.1. Classical chemostat model

We start by proving that all solutions of a chemostat controlled by an affine feed-

back law for the dilution rate are bounded and that they converge to a particular
subset in R+
3.

Proposition 1. If  > 0 and ki ≥ 0, i = 1, 2, or if  = 0 and ki > 0, i = 1, 2

then all solutions of system (8) with feedback (9) starting in R+ 3 are bounded. If

x(t) := (S(t), x1 (t), x2 (t)) is a solution starting in R+ , then

3

lim S(t) + x1 (t) + x2 (t) = 1 (28)

t→+∞

Proof. Define the function V : R+

3 → R as follows:
+

1
V (x) := (S + x1 + x2 − 1)2 (29)
2
It follows from the system equations (8) that:

V̇ = −2D(x)V (x) (30)

Case 1.  > 0.
Observe that D(x) > 0 for all x ∈ R+ 3 because k ≥ 0, i = 1, 2 and  > 0. This
i
implies that V̇ ≤ 0 for all x ∈ R+ . Consequently, all solutions of system (8) with
3
3 are bounded since V is radially unbounded in R3 . Furthermore,
feedback (9) in R+ +
64 P. De Leenheer, H. Smith

Lasalle’s invariance principle implies that all solutions converge to the largest in-
variant set E contained in the set M := {x ∈ R+ 3 | V̇ = 0}. As before D(x) > 0 for

x ∈ R+ and therefore (30) implies that M = {x ∈ R+

n 3 | V (x) = 0} or:

lim d(x(t), M) = 0 (31)

t→+∞

where d(a, A) is the (Euclidean) distance between the point a and the set A. The
proof is concluded by observing that (31) and (28) are equivalent.

Case 2.  = 0.
We will prove that since ki > 0, i = 1, 2 all solutions of system (8) with
feedback (9) starting in R+ 3 but not on the S-axis are bounded and if x(t) :=

(S(t), x1 (t), x2 (t)) is such a solution, then (28) holds. Consider (30) but observe
that D(x) is only nonnegative instead of positive in R+ 3 . Lasalle’s invariance princi-

ple learns that solutions converge to the largest invariant set contained in the union
of the S-axis and the set {x ∈ R+ 3 | V (x) = 0}. We show next that no point of the

S-axis can be the ω limit point of a solution which is not initiated on the S-axis
(observe that the S-axis is an invariant set consisting of equilibrium points).
 If S(t0 ) = r at some instant t0 where r ≥ 0 is sufficiently small then Ṡ =
i=1,2 (ki (1 − r) − fi (r))xi which is positive if (x1 , x2 ) = (0, 0). Thus there
exists a real number r0 > 0 such that all solutions not initiated on the S-axis sat-
isfy S(t) > r0 for sufficiently large t. We show next that for sufficiently large t,
D(t) > p0 -where p0 is some positive real number- holds for all solutions not
initiated on the S-axis. Pick any p satisfying fi (r0 ) > p > 0, i = 1, 2. Then if
D(t1 ) = p at some instant t1 , Ḋ ≥ D min(fi (r0 ) − p) > 0 and thus there exists
p0 > 0 such that the above claim holds. This concludes the proof that no point of
the S-axis is an ω limit point and therefore (28) holds.

The proof of the previous result motivates the definition of the following set:

 := {x ∈ R+
3
| S + x1 + x2 = 1} (32)

According to proposition 1 the ω-limit set of every solution of system (8) and feed-
back (9) belongs to the set , which is an invariant set for the system. It may be
conjectured that the asymptotic behavior of solutions on the set  also determines
the asymptotic behavior of solutions of the original system. Although this is not
true in general, we will show later that in this case it is true.
Therefore we study the behavior of the original system restricted to the set .
The dynamics of this system is governed by the differential equation which results
by substituting S by 1 − x1 − x2 in the original system equations:

ẋ1 = x1 (f1 (1 − x1 − x2 ) − D(x))

ẋ2 = x2 (f2 (1 − x1 − x2 ) − D(x)) (33)
x1 (0) ≥ 0, x2 (0) ≥ 0 and x1 (0) + x2 (0) ≤ 1

where D(x) = k1 x1 + k2 x2 +  as before. Notice that if  ∈ [0, D̃) and ki ≥ 0,

i = 1, 2 are chosen to satisfy (10), then system (33) possesses four equilibrium
Feedback control for chemostat models 65

points:
E0r = (0, 0), E1r = (1 − λ1 , 0), E2r = (0, 1 − λ2 ) and E3r = (x1e , x2e ) (34)
where λi and xie , i = 1, 2 are given by:
fi (λi ) = ki (1 − λi ) + , i = 1, 2
1
x1e = (D̃ − (k2 (1 − S̃) + )) (35)
k1 − k 2
1
x2e = ((k1 (1 − S̃) + ) − D̃)
k1 − k 2
The assumptions on  and ki , i = 1, 2 imply that xie > 0 and thus that E3r ∈ int(R+2 ).

Recall also that the values λi , i = 1, 2 can be easily found from the graphs in Figure
2. For example, λ1 is the value of the concentration of the intersection point of the
graph of f1 and the line k1 (1 − S) + .
The asymptotic behavior of system (33) is discussed in the following proposi-
tion.
Proposition 2. If  ∈ [0, D̃) and if the gains ki ≥ 0, i = 1, 2 satisfy the in-
equalities (10), then E3r is a globally asymptotically stable equilibrium point for
system (33) with respect to initial conditions satisfying x1 (0) > 0, x2 (0) > 0 and
x1 (0) + x2 (0) ≤ 1.
Proof. The proof is based on the Poincaré-Bendixson theorem, which sums up the
possible ω-limit sets of two-dimensional systems.
Denote the Jacobian matrices of the equilibrium points Ekr , k = 0, ..., 3 of sys-
tem (33) by Jk and the spectrum of these matrices by σ (Jk ). Simple calculations
and using the assumptions on  and ki , i = 1, 2, leads to the following conclusions:
j
σ (J0 ) = {r01 , r02 } with r0 > 0, j = 1, 2.
σ (J1 ) = {r11 , r12 } with r11 > 0 and r12 < 0.
σ (J2 ) = {r21 , r22 } with r21 > 0 and r22 < 0.
j
σ (J3 ) = {r31 , r32 } with r3 < 0, j = 1, 2.

This means that E0r is a repellor, E3r is locally asymptotically stable and E1r and E2r
are saddle points.
Consider a solution (x1 (t), x2 (t)) with xi (0) > 0, i = 1, 2. A simple applica-
tion of the Butler-McGehee theorem, see e.g. [22], shows that E1r and E2r cannot
belong to the ω-limit set of this solution. The same is true for E0r since it is a repellor.
Notice that system (33) is a competitive system, i.e. the off-diagonal entries of the
Jacobian matrix in all points are negative or zero, implying that the system does
not exhibit periodic solutions, see [7] or [21]. Therefore the only possible ω-limit
set of (x1 (t), x2 (t)) is the equilibrium point E3r or a cycle containing only one
equilibrium point which would necessarily be E3r . But since a cycle cannot contain
a locally asymptotically stable equilibrium point, the ω-limit set of (x(t), x2 (t))
is E3r . This concludes the proof.
66 P. De Leenheer, H. Smith

Finally we are in a position to prove theorem 2.

Proof of Theorem 2
First we introduce the following variable:

 := S + x1 + x2 − 1 (36)

System (8) with feedback (9) takes the form:

ẋi = xi (fi (1 −  − x1 − x2 ) − D(x)), i = 1, 2

(37)
˙ = −D(x)

where (0) ≥ −1 and xi (0) ≥ 0. From proposition (1) we obtain that limt→+∞
(t) = 0 and that system (37) is uniformly bounded. On the set  := {(x, ) ∈
R+ 2 × [−1, +∞)|  = 0} the dynamics of system (37) is given by system (33).

Notice that system (37) takes the form of system (44) and system (33) the form
of system (45) in the Appendix. This is immediate in the case  > 0. If  = 0 it
follows from the asymptotic inequality D(t) > p0 for large t, which was shown in
the proof of proposition 1. Relying on proposition 2 it can be checked that hypoth-
eses H1–H3 are true for system (33). Hypothesis H4 holds since if system (33)
would possess a cycle in , only E1r and/or E2r could possibly belong to it. Indeed,
it as been shown in the proof of proposition 2 that E0r is a repellor and E3r is locally
asymptotically stable and clearly these equilibrium points cannot belong to a cycle.
But since the stable manifolds of E1r and E2r are portions of the x1 -, respectively
x2 -axis (and both axes are invariant sets for system (33)), they cannot be part of a
cycle of equilibrium points either. Consequently theorem 5 can be applied to sys-
tem (37). In particular it follows that almost all solutions of this system converge
to the equilibrium point (E3r , 0) where E3r is the asymptotically stable equilibri-
um point of system (33). The solutions which do not converge to this equilibrium
point belong to the stable manifolds of the equilibrium points (E1r , 0) and (E2r , 0)
(where E2r and E3r are the saddle points of system (33)), but these stable mani-
folds are subsets of the boundary faces {(x1 , x2 , ) ∈ R3 | x2 = 0}, respectively
{(x1 , x2 , ) ∈ R3 | x1 = 0}, while we are only interested in solutions with initial
condition satisfying xi (0) > 0, i = 1, 2. These facts are easily rephrased for the
original system (8), which concludes the first part of the proof.
The converse statement in the theorem follows from the following argument.
The Jacobian matrix evaluated at the interior equilibrium point E3r :
 
−x̃1 (f1 (S̃) + k1 ) −x̃1 (f1 (S̃) + k2 )
J (E3r ) =
−x̃2 (f2 (S̃) + k1 ) −x̃2 (f2 (S̃) + k2 )

is a Hurwitz matrix (it is always nonsigular by the last part of hypothesis H)

only if k1 > k2 . Since ki ≥ 0 it follows that k1 > 0 and therefore the equali-
ties D(x̃1 , x̃2 ) = k1 x̃1 + k2 x̃2 +  = D̃ imply that  < D̃. Finally, the equilibrium
point E3r belongs to int(R+ 2 ) only if (10) are satisfied which follows immediately

from the expression (35).

Feedback control for chemostat models 67

7.2. Plasmid-bearing and plasmid-free competitors

The proof of theorem 4 is similar to the proof of theorem 2. Therefore we will not
provide all details, but only focus on the differences.
First of all it is easy to prove that all solutions of system (22) with feedback
(23) are bounded if  > 0 and k ≥ 0, and that they converge to the set  (see
(32)). The proof is based on the same arguments as the proof of proposition 1 and
therefore omitted.
The second step is then to study the reduced system on , which is given by
the following differential equation:
ẋ1 = x1 ((1 − q)f1 (1 − x1 − x2 ) − D(x))
ẋ2 = x2 (f2 (1 − x1 − x2 ) − D(x)) + qx1 f1 (1 − x1 − x2 ) (38)
x1 (0) ≥ 0, x2 (0) ≥ 0 and x1 (0) + x2 (0) ≤ 1
where D(x) = k(x1 + x2 ) +  as before. If  ∈ (0, D̃) and the gain k satisfies the
condition (24) of theorem 4, then system (38) possesses three equilibrium points:
E0r = (0, 0), E2r = (0, 1 − λ2 ) and E3r = (x1e , x2e ) (39)
where λ2 is implicitly defined by:
f2 (λ2 ) := k(1 − λ2 ) +  (40)
The value of λ2 can be easily determined graphically, see Figure 7: For given val-
ues of  and k (satisfying the conditions of theorem 4), the intersection of the line
y = k(1 − S) +  and the graph of f2 (S) determine the value of λ2 .
In (39), the values of xie , i = 1, 2, are given by:
(1 − λ∗ )((1 − q)f1 (λ∗ ) − f2 (λ∗ ))
x1e =
f1 (λ∗ ) − f2 (λ∗ )
(41)
(1 − λ∗ )qf1 (λ∗ )
x2e =
f1 (λ∗ ) − f2 (λ∗ )
where λ∗ is implicitly defined by:
(1 − q)f1 (λ∗ ) := k(1 − λ∗ ) +  ≡ D ∗ (42)
As was the case for λ2 , it is easy to determine λ∗ and D ∗ graphically, see Figure
7. Since  and k satisfy the conditions of theorem 4, it can be verified that the
following holds:
S̃ < λ∗ < λ2 < 1 and D ∗ > D̃ (43)
This implies in particular that xie > 0, i = 1, 2 and thus that x e ∈ int(R+ 3 ) as

claimed in the statement of theorem 4.

An important difference between the reduced system (38) and the reduced system
(33) is that system (38) is not a competitive system. Recall that competitive planar
systems cannot have nontrivial periodic solutions and that we used this to determine
the asymptotic behavior of system (33). Fortunately we can also exclude the exis-
tence of nontrivial periodic solutions of system (38) by means of Dulac’s criterion
as we show next.
68 P. De Leenheer, H. Smith

+k
l
l y = k(1 − S) + 
l
 + k̃ l
l+ 
y = k̃(1 − S)
l
l
l (1 − q)f1 (S)
D∗ l f2 (S)
D̃ l
l
 l
l
l
0 S̃ λ∗ λ2 1 S

Fig. 7. Nutrient concentration values at equilibrium points.

Lemma 1. If k ≥ 0 then system (38) does not possess nontrivial periodic solutions.

Proof. Denote the vector field in the right-hand side of system (38) by F (x1 , x2 )
and consider the divergence of the vector field x11x2 F (x1 , x2 ):

1 ∂f1 1 ∂f2 qf1 (1 − x1 − x2 )

− ( (1 − x1 − x2 ) + k) − ( (1 − x1 − x2 ) + k) −
x2 ∂S x1 ∂S x22

which is negative for all xi , i = 1, 2 of interest. Dulac’s criterion for the exclusion
of nontrivial periodic orbits is therefore satisfied which concludes the proof.

The asymptotic behavior of the reduced system (38) is then summarized as follows.
Proposition 3. If  and the gain k satisfy the conditions of theorem 4 then E3r is a
GAS equilibrium point for system (38) with respect to initial conditions satisfying
x1 (0) > 0, x2 (0) > 0 and x1 (0) + x2 (0) ≤ 1.

Proof. Since the proof is very similar to the proof of proposition 2 and we only
provide a sketch of it. It is also based on the Poincaré-Bendixson theorem.
From a calculation of the Jacobian matrices at the equilibrium points of system
(38) follows that E0r is a repellor, E2r is a saddle point and E3r is locally asymptoti-
cally stable. By lemma 1 system (38) does not possess nontrivial periodic solutions
and the Butler-McGehee theorem shows that E2r does not belong to the limit set of
solutions with initial conditions xi (0) > 0, i = 1, 2. Therefore E3r is GAS and this
concludes the proof.

Finally the proof of theorem 4 can be concluded with similar arguments used to
finish the proof of theorem 2 (see the end of the previous subsection). In particular,
theorem 5 is invoked, but since the details are similar as before, we omit the proof.
Feedback control for chemostat models 69

8. Appendix

In this section we state a convergence theorem.

Consider the following system:

ẋ = f (x, y), x ∈ Rn
(44)
ẏ = −γ (x)y, y ∈ R

where f : Rn+1 → Rn and γ : Rn → R+ \ {0} are sufficiently smooth (say at

least of class C 1 ). We assume that D is a forward invariant set for system (44) and
henceforth restrict initial conditions to D. Moreover it is assumed that solutions
of system (44) are uniformly bounded, i.e. there exists a compact subset of D into
which all solutions enter at some time and remain.
Next consider the following system:

ẋ = f (x, 0) (45)

where x ∈  := {x ∈ Rn | (x, 0) ∈ D} ⊂ Rn . For system (45) we introduce the

following set of hypotheses:

H1 There are only a finite number, say p, equilibrium points in  and these are
denoted as x1 , ..., xp .
H2 Denoting the stable manifold of equilibrium point xj as W s (xj ), the following
holds for the dimensions of these manifolds:

∀j = 1, ..., r : dim(W s (xj )) = n

∀j = r + 1, ..., p : dim(W s (xj )) < n

for some r ∈ {1, 2, ..., p}.

p
H3 ∪j =1 W s (xj ) = .
H4 There are no cycles of equilibrium points in .

The following result is then only a slight modification of theorem F.1 in [22].
Theorem 5. If H1–H4 are true, then for some i ∈ {1, ..., p}:

lim (x(t), y(t)) = (xi , 0)

t→+∞

where (x(t), y(t)) is an arbitrary solution of system (44) in D. Moreover,

p
∪j =r+1 W̃ s (xj , 0) has Lebesgue measure zero, where W̃ s (xj , 0) is the stable man-
ifold of equilibrium point (xj , 0) with respect to system (44).

References

1. Bernard, O., Malara, G., Sciandra, A.: The effects of a controlled fluctuating nutrient
environment on continuous cultures of phytoplankton monitored by a computer. J. Exp.
Mar. Biol. Ecol. 197, 263–278 (1996)
2. Butler, G.J., Hsu, S.B., Waltman, P.: A mathematical model of the chemostat with peri-
odic washout rate. SIAM J. Appl. Math. 45, 435–449 (1985)
70 P. De Leenheer, H. Smith

3. Flegr, J.: Two distinct types of natural selection in turbidostat-like and chemostat-like
ecosystems. J. Theor. Biol. 188, 121–126 (1997)
4. Freedman, H.I., Xu, Y.: Models of competition in the chemostat with instantaneous and
delayed nutrient recycling. J. Math. Biol. 31, 513–527 (1993)
5. Hale, J.K., Somolinas, A.S.: Competition for fluctuating nutrient. J. Math. Biol. 18,
255–280 (1983)
6. Hansen, S.R., Hubbel, S.P.: Single-nutrient microbial competition: qualitative agree-
ment between experimental and theoretically forecast outcomes. Science 207, 1491–
1493 (1980)
7. Hirsch, M.W., Smale, S.: Differential equations, dynamical systems and linear algebra,
Academic Press, New York, 1974
8. Hofbauer, J., Sigmund, K.: Evolutionary games and population dynamics, Cambridge
University Press, Cambridge, 1998
9. Hsu, S.B.: A competition model for a seasonally fluctuating nutrient. J. Math. Biol. 9,
115–132 (1980)
10. Hsu, S.B., Waltman, P.: Analysis of a model of two competitors in a chemostat with an
external inhibitor. SIAM J. Appl. Math. 52, 528–540 (1992)
11. Malara, G., Sciandra, A.: A multiparameter phytoplanktonic culture system driven by
microcomputer. J. Appl. Phycol. 3, 235–241 (1991).
12. Hsu, S.B., Waltman, P., Wolkowicz, G.S.K.: Global analysis of a model of plasmid-
bearing, plasmid-free competition in a chemostat. J. Math. Biol. 32, 731–742 (1994)
13. Hsu, S.B., Kuo, T.K.: Global analysis of a model of plasmid-bearing plasmid-free com-
petition in a chemostat with inhibition. J. Math Biol. 34, 41–76 (1995)
14. Hsu, S.B., Waltman, P.: Competition between plasmid-bearing plasmid-free organisms
in selective media. Chem. Eng. Sci. 52, 23–35 (1997)
15. Hsu, S.B., Waltman, P.: Competition in the chemostat when one competitor produces a
toxin, Japan. J. Ind. Appl. Math. 15, 471–490 (1998)
16. Jager, W., So, J.W.-H., Tang, B., Waltman, P.: Competition in the gradostat. J. Math.
Biol. 25, 23–42 (1987)
17. Lenski, R.E., Hattingh, S.: Coexistence of two competitors on one resource and one in-
hibitor: a chemostat model based on bacteria and antibiotics. J. Theor. Biol. 122, 83–93
(1986)
18. Rao, N.S., Roxin, E.O.: Controlled growth of competing species. SIAM J. Appl. Math.
50, 853–864 (1990)
19. Ruan, S., He, X.-Z.: Global stability in chemostat-type competition models with nutrient
recycling. SIAM J. Appl. Math. 58, 170–192 (1998)
20. Smith, H.L.: Competitive coexistence in an oscillating chemostat. SIAM J. Appl. Math.
40, 498–522 (1981)
21. Smith, H.L.: Monotone dynamical systems, AMS, Providence, 1995
22. Smith, H.L., Waltman, P.: The Theory of the Chemostat, Cambridge University Press,
Cambridge, 1995
23. Stephanopoulos, G., Fredrickson, A.G.: Effect of spatial inhomogeneities on the co-
existence of competing microbial populations. Biotech. & Bioeng. XXI, 1491–1498
(1979)
24. Stephanopoulos, G., Fredrickson, A.G., Aris, R.: The growth of competing microbial
populations in CSTR with periodically varying inputs. Amer. Instit. Chem. Eng. J. 25,
863–872 (1979)
25. Stephanopoulos, G., Lapidus, G.R.: Chemostat dynamics of plasmid-bearing, plasmid-
free mixed recombinant cultures. Chem. Eng. Sci. 43, 49–57 (1988)