Standardized Maximin E-optimal Designs for the Michaelis Menten Model

Full text

Page 1

Statistica Sinica 13(2003), 1147-1163
STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS FOR
THE MICHAELIS-MENTEN MODEL
Holger Dette1, Viatcheslav B. Melas2and Andrey Pepelyshev2
1Ruhr-Universit¨ at Bochum and2St. Petersburg State University
Abstract: In the Michaelis-Menten model we determine efficient designs by maxi-
mizing a minimum of standardized E-efficiencies. It is shown in many cases that
optimal designs are supported at only two points and that the support points and
corresponding weights can be characterized explicitly. Moreover, a numerical study
indicates that two point designs are usually very efficient, even if they are not
optimal. Some practical recommendations for the design of experiments in the
Michaelis-Menten model are given.
Key words and phrases: Chebyshev system, E-optimal designs, Michaelis-Menten
model, minimax-optimality, standardized optimal designs.
1. Introduction
The Michaelis-Menten model
E[Y | t] =
at
t + b,
t ∈ [0,t0],
(1.1)
is widely used to describe physical and biological phenomena (see Cressie and
Keightley (1979), Johansen (1984), Beverton and Holt (1957), Cornish-Browden
(1979), Hay, Meznarich, DiGiacomo, Hirst and Zerbe (1988) among many others)
and the problem of designing experiments for this model has found considerable
attention in the literature (see Duggleby (1979), Dunn (1988), Rasch (1990),
Boer, Rasch and Hendrix (2000) or Dette and Wong (1999)). An approximate
design ξ is a probability measure with finite support on the interval [0,t0] (see
Kiefer (1974), Silvey (1980) or Pukelsheim (1993)). Here the support points
t1,...,tkrepresent the locations where observations are taken, and the masses
w1,...,wkgive the proportion of the total observations to be taken at the par-
ticular points. If n independent observations with constant variance σ2> 0
have been obtained from the design ξ (possibly with an appropriate rounding of
the quantities nwj, j = 1,...,k, see e.g., Pukelsheim and Rieder (1992)), then
under the assumption of normality, the covariance matrix of the maximum like-
lihood estimate of the parameters (a,b)Tis approximately equal to the matrix

Page 2

1148HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
σ2n−1M−1(ξ,a,b), where
M(ξ,a,b) =
?t0
0
t2
(b + t)2


1
−
a
b+t
a2
(b+t)2
−a
b+t

dξ(t)(1.2)
denotes the information matrix of the design ξ. Note that this matrix depends on
the unknown parameters and, following Chernoff (1953), we call a design locally
φ-optimal if it maximizes a concave function of the information matrix (1.2).
Locally D-optimal designs (φ(M) = log|M|, where |·| denotes the determinant)
have been determined by Rasch (1990) and have equal weights at the points
[b/(2b + 1)]t0and t0. Locally E-optimal designs (φ(M) = λmin(M), where λmin
denotes the minimal eigenvalue) have been found in Dette and Wong (1999) and
have weights 1 − w and w at the points t0and (√2 − 1)t0b · {(2 −√2)t0+ b}−1,
respectively, where the weight w is given by
√2a2/b{b√2 − (4 − 3√2)t0}
2(t0+ b)2+ a2/b2{b√2 − (4 − 3√2)t0}2.
These designs have been criticized for several reasons. On the one hand the
designs depend sensitively on the unknown parameters and are in this sense
not robust. On the other hand, even if prior knowledge about the parameters
is available, it was pointed out in Dette (1997a,b) that, for specific parameter
constellations, E-optimal designs can perform particularly bad. For example if
a is small compared to b, the weight in (1.3) is close to 0 and consequently the
information matrix of the corresponding design is nearly singular. As a numerical
example consider the situation discussed in Cressie and Keightley (1981) where
t0= 2000, the parameter estimates are ˆ a = 43.95 andˆb = 236.53, and a locally E-
optimal design for the parameter (a,b) = (44,237) would take 99.91% at the point
t0. In this case the E-optimal design is inefficient for estimating both parameters
(note that this problem does not appear for the D-optimality criterion).
Some effort has been undertaken to construct robust designs with respect
to the D-criterion (see Song and Wong (1998) or Dette and Biedermann (2002))
but to our knowledge, robust designs based on the E-optimality criterion are not
available in the literature. It is the purpose of the present paper to construct
some. For this, we use a maximin approach introduced by M¨ uller (1995) (see
also Dette (1997a)). A design ξ∗is called standardized maximin φ-optimal if it
maximizes the function
w =
(1.3)
ψ(ξ) =min
a∈A,b∈B
φ(M(ξ,a,b))
max
η
φ(M(η,a,b)),
(1.4)

Page 3

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1149
where φ is a concave function and the minimum is taken over certain subsets A
and B of the parameter space specified by the experimenter. In many applications
such a specification is available, see for example Cressie and Keightley (1981,
p.237), where a specific range for the dissociation constant b for the receptor-
estradiol interaction is given. Note that the criterion ψ in (1.4) is a minimum of
φ-efficiencies taken over a certain range of the parameters and is in this sense very
intuitive. We will not use the function φ(M) = λmin(M) corresponding to the E-
optimality criterion directly (because of its deficiencies mentioned in the previous
paragraph), but rather a modified version introduced in Dette (1997b). This uses
a scaling of the elements in the information matrix, similar to the transition from
the covariance to the correlation matrix. The new criterion is carefully defined
in a general context in Section 2, which also gives some preliminary results and
an equivalence theorem for the standardized maximin optimality criterion. In
Section 3 we determine optimal designs for the Michaelis-Menten model and
prove that, for a sufficiently small set B for the parameter b, the standardized
maximin E-optimal design is always supported at two points. To our knowledge
a property of this type has been found in the context of Bayesian or maximin
optimality criteria (see Chaloner and Larntz (1989) or Haines (1995)) but has
never been proved rigorously in a non-trivial situation such as the Michaelis-
Menten model. Numerical results are presented in Section 4, which also gives
some practical recommendations, while Section 5 contains an appendix with a
technical result used for the proofs in Sections 2 and 3.
2. The Standardized Maximin E-optimality Criterion
Consider the regression model
E[Y | t] =
k
?
i=1
aiϕ(t,bi);
t ∈ I,
(2.1)
where ϕ(t,b) is a given function, I is an interval and a = (a1,...,ak)T, b =
(b1,...,bk)Tdenote the vectors of the unknown parameters.
Michaelis-Menten model discussed in the introduction is obtained by the choice
k = 1, I = [0,t0] and ϕ(t,b1) = t/(t + b1). The Fisher information matrix in the
model (2.1) is given by
?
f(t,b)=(f1(t,b),...,f2k(t,b))T=(ϕ(t,b1),ϕ?(t,b1),...,ϕ(t,bk),ϕ?(t,bk))T∈ R2k,
Note that the
M(ξ,a,b) = Da
If(t,b)fT(t,b)dξ(t)Da,
(2.2)
(2.3)

Page 4

1150HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
where f is the vector of regression functions. Derivatives of ϕ are taken with
respect to its second argument, and the matrix Da∈ R2k×2kis defined by




Dette (1997b) points out several drawbacks of designs maximizing λmin(M(ξ,a,
b)). In particular, in many regression models, these designs become inefficient
for estimating the complete vector of parameters, while a standardized version
of the E-optimality criterion avoids these problems. We use this criterion for
the construction of robust designs. To be precise let ej= (0,...,0,1,0,... ,0)T
denote the jth unit vector and ξ∗
the expression eT
inverse of the matrix M. We assume that the jth parameter in (2.1) is estimable
by the design ξ, that is ej ∈ range(M(ξ,a,b)) for all a,b. A standardized E-
optimality criterion is defined as follows (see Dette (1997a,b)). Let
?
denote a diagonal matrix with jth entry proportional to the inverse square root
of the best “variance” obtainable by the choice of an experimental design for
estimating the jth coefficient in (2.1), and define the matrix
Da=







1
0 a 0
00
...
00
00
00
...
...
...
...
...
...
0 0
0 0
0 0
...
1 0
0 a
1
...
0
0
...
...











.
(2.4)
jdenote the locally ej-optimal design minimizing
jM−(ξ,a,b)ej, j = 1,...,2k, where M−denotes a generalized
Ka,b= diag(eT
1M−(ξ∗
1,a,b)e1)−1/2,...,(eT
2kM−(ξ∗
2k,a,b)e2k)−1/2?
(2.5)
C(ξ,b) = (KT
a,bM−(ξ,a,b)Ka,b)−1.
(2.6)
Following Dette (1997a), a design maximizing the function
λmin(C(ξ,b)) = λmin((KT
a,bM−(ξ,a,b)Ka,b)−1) = λmin((KT
1,bM−(ξ,1,b)K1,b)−1)
(2.7)
is termed locally standardized E-optimal (in the last equality we use the notation
1 = (1,...,1)T∈ Rk). Recalling the definition of the matrices in (2.2), (2.3),
(2.4) and Theorem 3.2 in Dette (1997b), it follows that the matrix in (2.6) does
not depend on the linear parameters a1,...,akof (2.1), which justifies the last
equality and our notation C(ξ,b).
Definition 2.1. A design ξ∗is called standardized maximin E-optimal if it
maximizes the function
min
b∈B
η
λmin(C(ξ,b))
max
λmin(C(η,b)),
(2.8)

Page 5

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1151
where the matrix C(ξ,b) is defined in (2.6) and B ⊂ Rkis a given set.
Throughout this paper we assume that the functions f1,...,f2kdefined in
(2.3) generate a Chebyshev system on the interval I (for any b). A set of functions
f1,...,fm: I → R is called a weak Chebyshev system (on the interval I) if there
exists an ε ∈ {−1,1} such that
????????
for all x1,...,xm ∈ I with x1 < x2 < ... < xm. If the inequality in (2.9) is
strict, then {f1,...,fm} is called Chebyshev system. It is well known (see Karlin
and Studden (1966, Theorem II 10.2) that if {f1,...,fm} is a weak Chebyshev
system, then there exists a unique function
ε ·
f1(x1) ... f1(xm)
...
fm(x1) ... fm(xm)
...
...
????????
≥ 0(2.9)
m
?
i=1
c∗
ifi(t) = c∗Tf(t),
(2.10)
with the following properties
(i) |c∗Tf(t)| ≤ 1
(ii) there exist m points s1< ... < smsuch that c∗Tf(si) = (−1)i−1
∀ t ∈ I
(2.11)
i = 1,...,m.
The function c∗Tf(t) is called a Chebyshev polynomial and gives the best approx-
imation of the function f0(t) ≡ 0 by normalized linear combinations of the system
f1,...,fmwith respect to the sup-norm. The points s1,...,smare called Cheby-
shev points and need not be unique. They are unique if 1 ∈ span{f1,...,fm}
and I is a bounded and closed interval, where in this case s1= minx∈Ix,
maxx∈Ix. It is well known (see Studden (1968), Pukelsheim and Studden (1993)
or Imhof and Studden (2001) among others) that in many cases the E- and c-
optimal designs are supported at the Chebyshev points. Our first lemma shows
that this is also the case for the locally ej-optimal designs and the locally stan-
dardized E-optimal design which maximizes the function defined by (2.7).
sm=
Lemma 2.2.
(2.3) generate a Chebyshev system on the interval I with Chebyshev polynomial
fT(t,b)c∗and Chebyshev points t∗
of the functions f1(·,b),...,f2k(·,b) is a weak Chebyshev system on the interval
I, then the following assertions are true:
Assume that (for fixed b) the functions f1(·,b),...,f2k(·,b) in
1< t∗
2< ... < t∗
2k. If any subsystem of 2k − 1

Page 6

1152HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
(a) For every j = 1,...,2k the design ξ∗
j=
?
t∗
1... t∗
w1,j... w2k,j
2k
?
is ej-optimal for
the model (2.1), where the weights are defined by wi,j= [(−1)ieT
/[?2k
ficient of the Chebyshev polynomial defined by (2.10).
(b) The design ξ∗
model (2.1) and λmin(C(ξ∗
iF−1ej]
?=1(−1)?eT
(fi(t∗
?F−1ej], i = 1,...,2k and the matrix F is given by F =
i,j=1. Moreover, eT
j,b))2k
jM−1(ξ∗
j,1,b)ej= (c∗
j)2, where c∗
jis the jth coef-
E= [1/(2k)]?2k
?=1ξ∗
?is locally standardized E-optimal in the
E,b)) = 1/(2k).
Proof. Note that the optimal designs do not depend on the parameters a1,...,ak
and we put aj= 1, j = 1,...,k without loss of generality. For a proof of (a),
consider a fixed j ∈ {1,...,2k} and let h−2
F−1ej)2. Then it is easy to see that JFwj= hjej, where wj= (w1,j,...,w2k,j)T
and the matrix J is defined by J = diag(1,(−1),1,... ,(−1)) ∈ R2k×2k. The
functions f1,...,fj−1,fj+1,...,f2k(for any j = 1,...,2k) generate a Chebyshev
system on the interval I and consequently the quantities (−1)ieT
same sign, which implies that the weights wi,jare all nonnegative. It now follows
from Elfving’s theorem (see Elfving (1952)) that the design ξ∗
from the results of Studden (1968), that hj= c∗
consequence of recent results of Imhof and Studden (2001) (see their Theorem
2.1(a) and the proof of Theorem 2.1(b)) and its proof is therefore omitted.
j
= eT
jM−1(ξ∗
j,1,b)ej= (?2k
?=1(−1)?eT
?
iF−1ejhave the
jis ej-optimal and,
j. Finally, the statement (b) is a
Note that Lemma 2.2 yields for all b that maxηλmin(C(η,b))=λmin(C(ξ∗
= 1/(2k), and the standardized maximin optimality criterion in (2.8) reduces to
the maximization of the function
E,b))
min
b∈Bλmin(C(ξ,b)).
(2.12)
The remaining part of this section is devoted to an equivalence theorem for the
optimality criterion (2.12). To this end let Π denote the set of all probability
measures defined on a σ-field on B, which contains all one-point sets. For a pair
(ξ,π) (where ξ is a design on I and π ∈ Π is a probability measure on B), for
which the minimum eigenvalue of the matrix C(ξ,b) has multiplicity one, we
define the vector pb= pb(ξ) by the relation
pT
bC(ξ,b)pb= λmin(C(ξ,b))(2.13)
and introduce the function
Ψ(t,π) =
?
B|pT
bf(t,b)|2π(db).
(2.14)

Page 7

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1153
Theorem 2.3. Assume that B ⊂ Rkis a compact set and that the assumptions
of Lemma 2.2 are satisfied for all b ∈ B. If for any design ξ with at least 2k
support points and all b ∈ B the minimum eigenvalue of the matrix
C(ξ,b) = K−1
1,bM(ξ,1,b)K−1
1,b
(2.15)
is simple and positive, then a design ξ∗is standardized maximin E-optimal in
the model (2.1) if and only if there exists a probability measure π∗∈ Π such that
max
t∈IΨ(t,π∗) = min
b∈Bλmin(C(ξ∗,b)),
(2.16)
where the function Ψ is defined by (2.14).
Moreover, if ξ∗and π∗satisfy the relation (2.16), then the maximum on the
left hand side is attained for any t ∈ supp(ξ∗) and the minimum on the right
hand side is attained for any b ∈ supp(π∗).
Proof. The proof is based on certain standard arguments of approximate optimal
design theory and for the sake of brevity we only indicate one direction. Let
ξ∗denote a standardized maximin E-optimal design and define for α ∈ (0,1),
ξα = (1 − α)ξ∗+ αξt, where ξt puts all mass at the point t ∈ I. Introducing
φ(ξ,b) = λmin(CK(ξ,b)), we obtain from the optimality of the design ξ∗
min
b∈Bφ(ξα,b) − min
b∈Bφ(ξ∗,b) ≤ 0.
(2.17)
Next define
B∗=
?
b ∈ B
???min
β∈Bφ(ξ∗,β) = φ(ξ∗,b)
?
(2.18)
as the set of points where the minimum of the function φ(ξ∗,·) is attained (note
that B∗depends on the design ξ∗, which is not reflected by our notation) and
introduce the function Q(α) = minπ∈Π
on π is not reflected by our notation). Defining λ∗= Q(0) = minb∈Bφ(ξ∗,b) we
obtain from (2.17), (1/α){Q(α)−Q(0)} = (1/α){Q(α)−λ∗} ≤ 0, which gives in
the limit
∂
∂αQ(α)
π∈˜Π
Here˜Π =
is the eigenvector of the matrix CK(ξ,b) corresponding to the minimum eigen-
value (which has multiplicity 1, by assumption), and therefore we obtain from
(2.19) that
min
π∈˜Π
?
Bφ(ξα,b)π(db) (again the dependence
???α=0+= min
?
B
∂
∂αφ(ξα,b)
???α=0+π(db) ≤ 0.
. Now φ(ξ,b) = pT
(2.19)
?
π ∈ Π |?
Bφ(ξ∗,b)π(db) = λ∗?
bC(ξ,b)pb, where pb
?
B(pT
bf(t,b))2π(db) ≤ λ∗
(2.20)

Page 8

1154HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
for any t. On the other hand we have
?
for any π ∈˜Π, which implies maxt∈I
Comparing this inequality with (2.20) and observing (2.21) shows the existence
of a probability measure π∗such that (2.16) holds. The remaining assertions of
the theorem are shown similarly and the proofs are left to the reader.
B
?
I(pT
bf(t,b))2ξ∗(dt)π(db) =
?
Bφ(ξ∗,b)π(db) = λ∗
(2.21)
?
B(pT
bf(t,b))2π(db) ≥ λ∗for any π ∈˜Π.
Remark 2.4. The theory developed so far is applicable for a broad class of
models including the cases, where the function ϕ(t,b) in (2.1) is given by 1/(t + b),
e−tb, log(t + b). However, for the explicit determination of the standardized
maximin E-optimal designs, several steps have to be performed which differ
for the different models. As a consequence each model has to be investigated
separately. We illustrate this procedure for the Michaelis-Menten model in the
next section.
3. Standardized Maximin E-optimal Designs for the Michaelis-Menten
Model
Note that (1.1) is a special case of (2.1) [k = 1,ϕ(t) = t/(t + b)]. It is easy
to see from (2.9) that the functions
f1(t,b) =
t
t + b,
f2(t,b) =
−t
(t + b)2
(3.1)
generate a Chebyshev system on the interval (0,t0] and any single function has
the same property on the interval (0,t0]. Note that the point t = 0 is obviously
not a support of an optimal design for the Michaelis-Menten model (independent
of the optimality criterion). Therefore it is sufficient to consider this model only
on the interval (0,t0] and all assumptions of Lemma 2.2 and Theorem 2.3 are
satisfied. A straightforward calculation shows that the Chebyshev polynomial is
given by
c∗
1f1(t,b) + c∗
2f2(t,b),
(3.2)
where the functions f1and f2are defined in (3.1), the coefficients and Chebyshev
points are given by
c∗
1= c∗
1(b) =t∗2
1− b2
(t∗
1)2,c∗
2= c∗
2(b) = −c∗
1b(t∗
t∗
1+ b)
1− b
,
(3.3)
t∗
1= t∗
1(b) =
√2t0b
2t0+ 2b +√2b;
t∗
2= t0.
(3.4)

Page 9

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1155
Moreover, from the Cauchy-Binet formula we obtain that, for any design ξ with
at least n ≥ 2 support points, the information matrix C(ξ,b) defined in (2.15)
is positive definite. Because it is a 2 × 2 matrix and the element in the position
(1,2) does not vanish it must have a minimum eigenvalue of multiplicity one.
The following result is now an immediate consequence of Lemma 2.2.
Theorem 3.1.
(a) For j = 1,2 the locally ej-optimal design for the Michaelis-Menten model is
given by
?t∗
w∗
ξ∗
j=
1
j1 − w∗
t0
j
?
,j = 1,2,
(3.5)
where the point t∗
[(2√2 + 3)b]/[(3√2 + 4)b +√2t0] or ω2= 2−1/2.
(b) A standardized locally E-optimal design for the Michaelis-Menten model is
given by
?
w∗1 − w∗
where the point t∗
[2(3 + 2√2)b + t0]/[2√2((3 + 2√2)b + t0)]. Moreover, λmin(C(M(ξ∗
1/2.
1is defined in (3.4) and the weight at this point is ω1 =
ξ∗
E=
t∗
1
t0
?
,
(3.6)
1is defined in (3.4) and the weight at this point is w∗=
E,b)) =
We now concentrate on the standardized maximin E-optimality criterion,
where the parameter space B is a “small” interval. In this case it can be proved
that the standardized maximin E-optimal design ξ∗is always supported at two
points. Note that this fact is intuitively clear because in this case the standardized
maximin E-optimal design ξ∗should be close to a locally standardized E-optimal
design. We mention that this fact has also been observed numerically for Bayesian
D-optimal designs (see Chaloner and Larntz (1989) or Haines (1995)), but to our
knowledge a rigorous proof of this property in a non-trivial context is not known.
Lemma 3.2. Consider the standardized maximin E-optimality criterion defined
by (2.8) with B = [b0− ∆,b0+ ∆] ⊂ R+,∆ > 0 in the Michaelis-Menten model
(1.1). If ∆ is sufficiently small,
(a) any standardized maximin E-optimal design for the Michaelis-Menten model
on the interval [0,t0] is supported at two points including the point t0;
(b) any measure π∗defined by the identity (2.16) satisfies supp(π∗) = ∂B =
{b0− ∆,b0+ ∆}.
Proof.
For a proof of (a) we assume the contrary.
exists a sequence (∆n)n∈Nof positive constants converging to zero such that for,
any n ∈ N, there exists a standardized maximin E-optimal design ξ∗
This means that there
nfor the

Page 10

1156HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
Michaelis-Menten model (with Bn = [b0− ∆n,b0+ ∆n]) which has either at
least three support points or two support points in the interval [0,t0). Because
0 cannot be a support point, ξ∗
t∗
and Theorem 2.3, we obtain for each n ∈ N a probability measure π∗
supp(π∗
nhas at least two support points, say t∗
1,nand
2,n, in the interval (0,t0). Recalling the definition of the function Ψ in (2.14)
nsuch that
n) ⊂ Bn= [b0− ∆n,b0+ ∆n],
Ψ(ti,n,π∗
n) = λ∗
n
i = 1,2,
Ψ?(ti,n,π∗
n) = 0;
i = 1,2,
(3.7)
where the derivative of Ψ(·,π∗
λ∗
sequence (also denoted by ∆n) such that limn→∞ti,n=¯ti
=¯λ. From (3.7) and (2.14) we obtain (note that π∗
measure δb0)
n) is taken with respect to the first argument and
n,b)). A standard argument shows that there exists a sub-
n= minb∈Bnλmin(C(ξ∗
i = 1,2;limn→∞λ∗
n
nconverges weakly to the Dirac
Ψ(¯ti,δb0) = (pT
b0f(¯ti,b0))2=¯λ;
b0f(t,b0))2???t=¯ti= 0;
i = 1,2
(3.8)
Ψ?(¯ti,δb0) =∂
∂t(pT
i = 1,2.
Now the identity f(0,b0) = 0 shows that¯ti ?= 0, i = 1,2. Moreover, if¯t1 =
¯t2, we would obtain the relations Ψ?(¯t1,δb0) = Ψ??(¯t1,δb0) = 0, Ψ(¯t1,δb0) =¯λ,
Ψ(t,δb0) ≤¯λ
polynomial. A similar argument shows that¯t2 < t0. Consequently, we obtain
0 <¯t1<¯t2< t0and (3.8) yields a contradiction to the Chebyshev property of
the system {f1(t,b),f2(t,b)} on the interval (0,t0). The conclusion follows.
For a proof of (b) we use Proposition 5.1 in the Appendix. To be pre-
cise, let s = 2, x = (x1,x2)T= (t,w)T, y = b and define for a design ξt,w =
?
w
1−w
that G is twice continuously differentiable and assumption (a) of Proposition 5.1
is obviously satisfied. For a proof of assumption (b) we note that
∀ t ∈ I. This is impossible because pT
b0f(t,b0) is the Chebyshev
tt0
?
, the function G(t,w,b) = λmin(C(ξ,b)). Then it is easy to see
G(t,w,b) = λmin(C(ξt,w,b)) = λmin(K−1
bM(ξt,w,1,b)K−1
b)
= w(¯ pbf1(t,b))2+ (1 − w)( ¯ pbf2(t0,b))2,
where ¯ pb=K−1
K−1
b
corresponding to its minimum eigenvalue, the matrix M(ξt,w,1,b) is defined
in (2.2), and the matrix Kb= K1,bis given by
?
0
c∗
bpb, pbis an eigenvector of the matrix C(ξt,w,b)=K−1
bM(ξt,w,1,b)
K−1
b
=
c∗
1(b)0
2(b)
?
=
?
c∗
0 c∗
10
2
?
(3.9)

Page 11

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1157
with c∗
equations
1= c∗
1(b) and c∗
2= c∗
2(b) defined by (3.3) (see Lemma 2.2a). Now the
∂
∂tG(t,w,b) =
∂
∂wG(t,w,b) = 0(3.10)
yield the system
w(( ¯ pT
bf(t,b))2)?= 0
( ¯ pT
(3.11)
bf(t,b))2= ( ¯ pT
bf(t0,b))2.
But this means that ¯ pbf(t,b) must be equal (up to a sign) to the Chebyshev
polynomial c∗Tf(t,b) = c∗
¯ pb= c∗and the solution of the first equation in (3.10) with respect to t is uniquely
determined by the interior Chebyshev point t = t∗
hand, we have K−1
this also determines the weight w uniquely. This proves (b) in Proposition 5.1. In
fact it follows by these arguments that the unique solution t∗=t∗(b), w∗=w∗(b)
of the system (3.10) is precisely the interior support point and its corresponding
weight of the locally standardized E-optimal design given in part (b) of Theorem
3.1.
Finally, for a proof of (c), at Proposition 5.1, introduce x∗(b)=(t∗(b),w∗(b))T
and calculate in a straightforward manner
1f1(t,b)+c∗
2f2(t,b) defined by (3.3). Therefore we obtain
1defined in (3.4). On the other
bM(ξt,w,1,b)¯ pb= λmin(C(ξt,w,b))Kb¯ pband, inserting ¯ pb= c∗,
V = (x∗?(b))TJ(b)x∗(b)
?∂
= 2w∗(b)
∂bt∗(b)
?2?
¯ pT
bf(t,b)∂2
∂2t¯ pT
bf(t,b)
????t=t∗(b).
Observing the Chebyshev property of the functions f1,f2, and that the Cheby-
shev polynomial ¯ pT
and the third assumption of Proposition 5.1. Now this proposition shows that
for sufficiently small ∆ and any two point design ξt,wthe function
λmin(C(ξt,w,b))
max
η
bf(t,b) = c∗Tf(t,b) is maximal at t = t∗(b), we obtain V < 0,
Q(ξt,w,b) =
λmin(C(η,b))= 2λmin(C(ξt,w,b))
is concave as a function of b ∈ Bn= [b0−∆,b0+∆] (here the last identity follows
from Lemma 2.2). Consequently, observing part (a) we obtain for sufficiently
small ∆ > 0 that supp(π∗) = {b0− ∆,b0+ ∆}, which proves the second part of
the assertion.
We conclude this section with a more explicit characterization of the stan-
dardized maximin E-optimal designs for the Michaelis-Menten model in the sit-
uation of the previous lemma.
Theorem 3.3. Consider the Michaelis-Menten model (1.1) and the standardized
maximin E-optimality criterion (2.8) with B = [b1,b2] where 0 < b1< b2.

Page 12

1158HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
(a) For sufficiently small b2− b1a standardized maximin E-optimal design is of
the form ξ∗=
¯ w
1− ¯ w
Theorem 2.3 is of the form π∗=
¯ α
of the optimization problem maxt∈Tmaxt,∈[0,1]minα∈[0,1]g(t,w,α), where the
function g is defined by
?
with ub= (1/2)tr(K−1
and the matrices K−1
b
and M(ξ,1,b) are defined in (3.9) and (1.2), respec-
tively.
(b) The design ξ∗defined in part (a) is standardized maximin E-optimal if and
only if
?t1
t0
?
and the corresponding probability measure π∗in
?b1
b2
1−¯ α
?
. Here (t1, ¯ α, ¯ w) is a solution
g(t,w,α) = α
?
ub1−
u2
b1− vb1
?
b) and vb= (detKb)−2|M(ξt,w,1,b)|,
+ (1 − α)
?
ub2−
?
u2
b2− vb2
?
,
bM(ξt,w,1,b)K−1
max
t∈[0,t0]¯ α(pT
b1f(t,b1))2+ (1 − ¯ α)(pT
b2f(t,b2))2=min
b∈[b1,b2](ub−
?
u2
b− vb).
Proof. Observing the notation C(ξ∗,bi) = K−1
to see that λmin(C(ξ∗,bi)) = ubi−
application of Lemma 3.2 and Theorem 2.3.
biM(ξ∗,bi)K−1
bi− vbi. The proof now follows by a direct
bi
(i = 1,2) it is easy
?
u2
4. A Numerical Example
In this section we take [0,t0] = [0,10] and calculate standardized maximin
E-optimal designs for various sets of the form B = [1,b2]. From Theorem 3.3, the
optimal designs are supported at two points if b2is close to b1= 1. We calculated
the optimal two point designs with the aid of the first part of Theorem 3.3 and
checked its optimality in the class of all designs by an application of part (b) of
that theorem. If b2= 1,2,3,...,7 the standardized maximin E-optimal design is
in fact supported at only two points and the corresponding designs are depicted in
Table 4.1. The table also contains the standardized maximin E-optimal designs
in the class of all standardized locally optimal designs, that is,
ξ∗
loc= argmax
?
min
b∈B
λmin(C(ξ,b))
maxηλmin(C(η,b))
???ξ ∈ Ξ
?
,
(4.1)
where Ξ denotes the class of all designs defined in part (b) of Theorem 3.1 by (3.6).
Note that the determination of these designs is substantially simpler compared
to the calculation of the standardized maximin E-optimal design within the class

Page 13

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1159
of all designs. It is of interest to investigate the corresponding efficiencies of these
designs:
λmin(C(ξ,b))
maxηλmin(C(η,b)).
eff(ξ) = min
b∈B
(4.2)
Table 4.1. The standardized maximin E-optimal designs for the Michaelis-
Menten model on the interval [0,10] with B = [1,b2]. The table contains the
support points x1,x2and weights w1,w2of the design ξ∗and ξ∗
(2.8) and (4.1), while the minimum efficiency is defined by (4.2). In all cases
the standardized maximin E-optimal design is supported at two points.
locdefined by
b2
1234567
x1
x2
w1
w2
10101010 1010 10
ξ∗
loc
0.6040
0.5163
0.4837
0.8134
0.4856
0.5144
0.9602
0.4670
0.5330
1.0708
0.4543
0.5457
1.1579
0.4450
0.5550
1.2286
0.4378
0.5622
1.2874
0.4321
0.5679
x1
x2
w1
w2
1010101010 10 10
ξ∗
0.6040
0.5163
0.4837
0.8169
0.4880
0.5120
0.9693
0.4727
0.5273
1.0847
0.4625
0.5375
1.1757
0.4550
0.5450
1.2498
0.4494
0.5506
1.3111
0.4449
0.5551
ξ∗
loc
ξ∗
eff
eff
1
1
0.9544
0.9544
0.8946
0.8947
0.8448
0.8451
0.8050
0.8053
0.7728
0.7733
0.7465
0.7471
Obviously this quantity is decreasing with increasing values of b2, but even when
b2 = 7 the standardized maximin E-optimal design has an efficiency of 75 %.
Moreover, the standardized maximin E-optimal design ξ∗and the design ξ∗
obtained by maximizing (2.8) in the subclass Ξ of all designs defined in Theorem
3.1(b) have nearly the same performance from a practical point of view.
If b2≥ 8 it follows from Theorem 3.3 that the best two point design is not
optimal within the class of all designs and the standardized maximin E-optimal
design has at least three support points. Some representive results are depicted
in Table 4.2, which also contains the two point designs obtained from the first
part of Theorem 3.3 and the designs ξ∗
show that for b2≥ 8 the standardized maximin E-optimal design is always sup-
ported at three points. We observe that the two point designs are quite efficient
compared to standardized maximin E-optimal designs supported at three points.
For example in the case b2= 20, the two point designs yield approximately 60%
at (4.2) , while the standardized maximin E-optimal design ξ∗gives a minimal
efficiency of approximately 67 %. Only for very large values is the difference
loc
locdefined by (4.1). Our numerical results

Page 14

1160HOLGER DETTE, VIATCHESLAV B. MELAS AND ANDREY PEPELYSHEV
between the two and three point designs notable (see the columns with b2= 50
and b2= 100 in Table 4.2).
Table 4.2. The standardized maximin E-optimal designs for the Michaelis-
Menten model on the interval [0,10] with B = [1,b2]. The table contains
the support points xi and weights wi of the design ξ∗and ξ∗
(2.8) and (4.1) and the best two-point design ξ∗
Theorem 3.3, while the minimum efficiency is defined by (4.2). In all cases
the standardized maximin E-optimal design is supported at three points.
locdefined by
twodefined in the first part of
z2
810121416182050 100
x1
x2
w1
w2
1010 10101010 1010 10
ξ∗
loc
1.3371 1.4166 1.4777 1.5262 1.5657 1.5984 1.6261 1.8030
0.4275 0.4203 0.4151 0.4111 0.4079 0.4053 0.4032 0.3902
0.5725 0.5797 0.5849 0.5889 0.5921 0.5947 0.5968 0.6098
1.8753
0.3853
0.6147
x1
x2
w1
w2
1010 1010 1010 101010
ξ∗
two
1.3630 1.4464 1.5103 1.5610 1.6023 1.6367 1.6660 1.8508
0.4413 0.4358 0.4317 0.4286 0.4262 0.4243 0.4228 0.4130
0.5587 0.5642 0.5683 0.5714 0.5738 0.5757 0.5772 0.5870
1.9266
0.4094
0.5906
x1
x2
x3
w1
w2
w3
10 10 1010 1010 10 1010
1.0985 0.9278 0.8757 0.8396 0.8105 0.7965 0.7974 0.8647
2.3340 2.8343 3.1096 3.3276 3.5114 3.6534 3.7205 4.0054 4.1 907
0.4262 0.4002 0.3835 0.3711 0.3603 0.3520 0.3487 0.3362
0.4060 0.3549 0.3456 0.3398 0.3363 0.3350 0.3341 0.3353
0.1678 0.2449 0.2709 0.2890 0.3034 0.3130 0.3172 0.3285
0.9119
ξ∗
0.3320
0.3377
0.3303
ξ∗
loc
ξ∗
two
ξ∗
eff
eff
eff
0.7246 0.6904 0.6649 0.6452 0.6295 0.6167 0.6060 0.5417
0.7253 0.6911 0.6657 0.6461 0.6304 0.6176 0.6070 0.5429
0.7270 0.7034 0.6906 0.6829 0.6779 0.6745 0.6720 0.6562
0.5172
0.5185
0.6499
From these results (and similar results not given) we draw the following con-
clusions. Only in cases with nearly no prior information about the unknown
parameter can the effort of calculating the standardized maximin E-optimal de-
sign within the class of all designs be justified. In all other cases the two point
designs obtained in Theorem 3.3 and the two point designs defined in (4.1) have
a very similar performance as the standardized maximin E-optimal designs. In
many cases the designs from Theorem 3.3 are already standardized maximin E-
optimal as can be checked by the second part of Theorem 3.3. On the other hand
the designs defined by (4.1) are practically indistinguishable from the designs in
Theorem 3.3, and substantially simpler to calculate. Therefore, if it can be as-
sumed that the parameter b varies in an interval [b1,b2] which is not too large,

Page 15

STANDARDIZED MAXIMIN E-OPTIMAL DESIGNS1161
these designs provide a reasonable compromise between efficiently designing an
experiment for the Michaelis-Menten model and dealing with the complexity of
a non-differentiable optimization problem.
Acknowledgements
The authors thank the referees and an associate editor for comments on an
earlier version of this paper, and Isolde Gottschlich who typed most parts of
this paper with considerable technical expertise. The support of the Deutsche
Forschungsgemeinschaft (SFB 475, Komplexit¨ atsreduktion in multivariaten
Datenstrukturen, Teilprojekt A2) is gratefully acknowledged.
Appendix
Proposition A.1. Let G : Ω×I → R, where Ω ⊂ Rsis a compact set and I ⊂ R
an arbitrary interval. Assume that the following conditions are satisfied.
(a) The function G is positive and twice continuously differentiable.
(b) For any y ∈ I,
solution x∗= x∗(y) in Ω.
(c) For all y ∈ I we have (x∗?(y))TJ(y)x∗?(y) < 0, where for fixed y ∈ I
J(y) = (
(x∗(y),y).
For any fixed x ∈ Ω the function Q(y) = Q(x,y) = G(x,y)/(max¯ x∈ΩG(¯ x,y)) is
twice continuously differentiable with respect to y and, if x is sufficiently close to
x∗(y), we have Q??(y) < 0.
Proof. A straightforward calculation shows
∂
∂xG(x,y) =
?
∂
∂x1G(x,y),...,
∂
∂xsG(x,y)
?T
= 0 has a unique
∂2
∂xi∂xjG(x,y)
???x=x∗(y))s
i,j=1is the Jacobian of G evaluated at the point
∂
∂yG(x∗(y),y) =
∂
∂yG(x,y)
???x=x∗(y)+
s
?
i=1
∂
∂xiG(x,y)
???x=x∗(y)·
∂
∂y(x∗
i(y)). (A.1)
Observing condition (b) we obtain
∂2
∂2yG(x∗(y),y) =
∂2
∂2yG(x,y)
s
?
???x=x∗(y)
∂2
∂xi∂xjG(x,y)
−
i,j=1
???x=x∗(y)
∂
∂yx∗
i(y)∂
∂yx∗
j(y).
(A.2)
The second derivative of the function Q (with respect to y) is
= Q1(x,y) + Q2(x,y), where the functions Q1,Q2are defined by
∂2
∂2yQ(y)=∂2
∂2yQ(x,y)
Q1(x,y) =
∂2
∂2yG(x,y)H(y) − G(x,y)H??(y)
H2(y)
,