Modelling data

uncertainty in growth
forecasts
Karmeshu*t and F. Lara-Rosano

Institute de Ingenieria, Universidad National Authoma de Mkxico 20, DF, Mexico

(Received May 1985; revised April 1986)

A probabilistic methodology within a dynamic framework is pro-

posed for the study of moments of errors in growth forecasts result-
ing from data uncertainty. This methodology is applied to well-
known evolutionary models of growth, namely exponential and
logistic. Explicit expressions for moments of the stochastic variable
are derived. The paper explores methods based on two-point distri-
bution approach, second-moment analysis, and probability distribu-
tion of parameters. Of these, the two-point distribution is found to
be computationally advantageous. An interesting feature emerging
from the analysis is that the mean and relative fluctuations in the
projected variable of interest are numerically not much different
from the respective ones when the uncertainties in the growth para-
meters are characterized by Gaussian, uniform or two-point distribu-
tion. This, however, holds for forecasting periods which are short
in comparison with the time-scale of the process under study.

Keywords: data uncertainty, two-point distribution, second-

moment analysis, probability distribution.

Growth forecasts are often used to provide the deci- where t = 0 is the initial time and P,, corresponds to
sion-maker with a feeling of dynamic change in social the value of the variable at time t = 0. For a constant
systems and organizations caused by an increase in cer- growth parameter r(t) = r, equation (2a) becomes:
tain state variables of the system, such as population,
earnings or production. The simplest model for growth P(t) = POerr (2b)
forecasts is the exponential growth model, described
The value of p increases exponentially with time
by:
which is why this model is called an exponential growth
model. In reality, the exponential trend cannot be main-
s= r(t)P
tained indefinitely due to the limited supply of
resources. One model which captures this feature of
scarce resources is the well-known logistic growth
where P = P(t) is the variable to be projected and r(t) model1 defined by:
is the growth rate. In general, the growth rate parameter
is not constant but time-dependent and its precise form dP
depends on the problem under consideration. The solu- -=
tion of equation (1) is: dt

where r is the intrinsic per capita growth rate and K

is the carrying capacity of the system. Equation (3)
reduces to that of an exponential growth model for small
values of P when resource constraints are not binding.
* Also at Fundacion Javier Barros Sierra, AC, Centro de Investiga- For large values of P approaching the carrying capacity,
ci6n Prospectiva, Carr. Al Ajusco 203, MCxico, DF, Mexico
t Permanent address: School of Computer and Systems Sciences, Jaw-
however, the growth rate slows down resulting in the
aharlal Nehru University, New Mehrauli Road, New Delhi-110067, growth curve tapering off. Integrating equation (3), P(t)
India becomes:

0307-904X/87/01002-09/$03.00
62 Appl. Math. Modelling, 1987, Vol. 11, February 0 1987 Butterworth & Co. (Publishers) Ltd
 Modelling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano

Taking the expectation of both sides and neglecting

P(t) =
PO the contributions of higher order terms (as one has no
(4)
PO/K+ [l - PO/K] exp(-rt) knowledge of higher moments), approximate expres-
Exponential and logistic growth models are often used sions for the first two moments of y are:
to provide the future projection of the variable P(t).
In order to make these projections, one needs to know
the explicit time variations of the model parameters and
j = qy] -f(n) + $f”(f)
L
the initial values of the variable. Sometimes, variability
in the projections is enormous and may be due to a
u; = [(y - E(jJ))2] = u2[f’(n)]2 (6)
great many reasons.2-9 One major source of variability These results can be easily generalized to situations with
may arise from data uncertainty, e.g. the growth rate many random variables. It may be pointed out that SMA
and initial values may not be known precisely. In prac- will not work when the function f(x) does not have a
tice, these estimates are always uncertain and are ran- first order derivative at x = Z, as it does when it is discon-
dom relative to repeated sample surveys.3 A rational tinuous.
framework for describing uncertainties associated with Rosenblueth”J2 proposed the representation of a
the parameters and initial values treats them as random random variable x in terms of two concentrations PI
variables with a specified first few moments or probabi- and P2 placed at x = x1 and x2. Using this representation,
lity distribution, and can be employed to study the tem- he obtained expressions for the expectation, standard
poral evolution of the projection uncertainties through deviation and coefficient of skewness of y, i.e. y, aY
random differential equations.2 and vYin terms of the first three moments X, a2 and
The purpose of this paper is to investigate the effects vd of x. This representation, besides being computa-
of data uncertainty on the dynamical evolution of the tionally efficient has one significant advantage over that
system. Consequently, interest is centred around the of SMA, namely that it is feasible even when the func-
time development of moments of the state variable P(t). tion f(x) is discontinuous. Furthermore, this represen-
This analysis is expected to provide, in quantitative tation provides an approximate and simple approach
terms, the evolution of uncertainties in projections of to Bayesian statistics. l2 Rosenblueth suggests that the
the variables due to various factors associated with data probability density function P(x) be written as:
uncertainty. The authors propose a method based on
p(x) = P16(x = Xl) + P&x - x2) (7)
two-point distribution in a dynamic set-up, while LindlO
suggests one in a discrete time set-up. where 6( .) is a Dirac delta function. In terms of:
‘$ = V/2 + [l + (V/2)211/2 52 = E, - v
Statistical description of data uncertainty
the parameters Pi and the points Xiare given by:
Data concerning variability in the parameter values and
initial conditions is usually too limited to specify the Pl = 52/G + t-2) p2 = 1 - p, (8)
details of the shape of their probability distributions. &= (Xi-XI/(T i= 1,2 (9)
However, in a number of problems, specification of the
first two moments suffices to estimate the moments of According to Rosenblueth,12 moments of the first three
the errors propagated through the model. The approach orders of y can be obtained from:
widely used by engineers for such types of problems IYI - Y2I
Y = PLY,+ p2Y2 ay = [W211
is the first order second-moment analysis (SMA).5
Another approach is based on the principle of maximum vyuy = (P2 - PIMY - Y2) (10)
entropy, which provides a theoretical framework to con- where:
struct probability distributions in a rational way when
only partial information about the data is available. Yi =fCxi)
These expressions simplify for v = 0, such that:
Second-moment analysis (SMA)
In SMA, all random variables are defined in terms of 51= 52= 1 P, = P2 = 1
their first and second moments and the shapes of the 2
probability distributions of these random variables are
ignored. To illustrate this, consider a random variable ,’ == (Y* + Y2)P fly = lY1- Y21/2 (11)
x, with specified first two moments, expectation and From equations (11) and (7):
variance; E(x) = X and E(x - i)2 = a: respectively. Let xi=j-a,X2=X+(T
y =f(x) be some arbitrary continuous function of X. The
calculation of the first two moments of the random vari- p(x) = ; [x- (X - CT)]+ f [x - (i + CT)]
able y are of interest. To this end Taylor’s expansion (12)
for functionf(x) is written, based on x = R, as: This is the symmetrical two-point distribution due to
Rosenblueth, who also generalized it to the case of n
y =f@) + (x - X)f’(Z) + ; (X - X)2f”(X) + . . . random variables in terms of 2” concentrations which
can be reduced to 2n. l2
where: For the sake of comparison between SMA and the
two-point distribution (TPD), equation (12) gives:
df
[a2fI*=i
f’(R) =
[Iax
*=x f”(n) =
dX2 ,’ = $f(x- 0) + f(X + a)]

Appl. Math. Modelling, 1987, Vol. 11, February 63

Modeling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano

Uncertainty in the growth rate

=f(%) + ;fyq + $yi) +... (13) To take account of the uncertainty in the growth rate,
suppose that the rate consists of a deterministic part
q? = [f(j + u) - f(i - u)]2/4 r. and a random part r,; then:
= uqf’(x)]’ + cpf’(_qf”‘(.q/3 + ... (14) r=ro+rl (18)
It is interesting to note from equations (5), (6), (13) From equation (2b):
and (14) that for small values of u, so that terms of
P(t) = Poerote’l’ (19)
the order of 03 can be neglected, SMA and TPD give
nearly the same results. E[P(t)] = Poe’o’ E(e’l’) (20)
With a view to studying how TPD fares when ri is distri-
Maximum entropy principle buted as a Gaussian or uniform variate, the first two
moments of P(t) are calculated.
The problem which a decision-maker is usually con-
fronted with is that information about the system is
incomplete and one finds that there is a large number Gaussian distribution
of probability distributions consistent with the informa- Suppose ri is a centred Gaussian random variable, i.e.:
tion available. The question then is: which probability
distribution is the most objective? This dilemma was E(r,) = 0 E(r$ = +
resolved by Jaynes13 in his fundamental paper, in which then, using the well-known result for a centred Gaussian
he demonstrated, using Shannon’s information meas- random variable cp:
ure,14 that maximum entropy probability distributions
(MEPD) are the least biased probability distributions E(e’q) = exp[-E((p2)/2]
with available information. (Information, however, is it is found that:
implicit in the choice of variable). Using MEPD, one
can characterize the probability distribution in terms
of known moments obtained from incomplete data. E[P(t)] = POexp rot + F
If the decision-maker is enabled, from the relevant ( 1
literature, only to infer the maximum and minimum
values of the random variable, then it can be shown dp(t) = f$e*rd e”fr2(@Y - 1)
(22)
that MEPD is the uniform distribution. If, instead, the
At this stage it may be instructive to compare the
decision-maker can infer the first two moments of the
structures of differential equations governing the growth
random variable, then MEPD is a Gaussian distribu-
of Z’(t) in the deterministic case and of E[P(t)] in the
tion.15 MEPD provides a rational theoretical framework
stochastic case. In the former case, with no randomness
allowing the incorporation of additional information as
in the growth parameter:
and when it becomes available. In the context of ecosys-
tem modelling, this framework has been used16 for 1 dP(t)
modelling data uncertainties.
P(t)dr=ro
which is a constant; whereas in the latter case, when
Uncertainty in initial conditions rl is a centred Gaussian variate:
Occasionally, when applying the exponential growth
model, there is uncertainty about the initial condition 1 WWI = r + -&
PO, due to lack of information and/or errors in measure- E[P(t)] dt ’ ’
ment. It is assumed that the first two moments of PO
are known, i.e.: which is no longer a constant. This shows that random-
ness in growth rate parameter plays a crucial role in
E(P,) = PO Var(P,) = IJ$ as much as it converts an autonomous system into a
nonautonomous one with growth rate parameter
Using these results, equation (2b) gives: increasing linearly with time. Equations (21) and (22)
E[P(t)] = FOert (15)
give the coefficient of variation as:

o++(t)= o$e2rt. CV,(t) = (eu:tf2- 1)112 (23)

(16)
When cJ<l:
Hence, the coefficient of variation is:

= ~ UP(t) = -co =
CV,(t)
E[P(t)] PO cvo (17) CVp(t) = UJ 1 + T
i
+ . . . 1’2
1
In this way, the coefficient of variation of the projected It is worth noting that the relative fluctuations increase
variable P(t) continues to have the same value as given with time, but for larger values of c$, the relative fluc-
by the initial conditions, showing that the variability tuation is too high and the projections become unreli-
in P(t) is not enhanced with the passage of time. In able.
the following section, this will be found to be in keen On using a well-known result,l’ the distribution of
contrast with the situations faced when the growth rate P(t) turns out to be lognormal with probability density
is subject to random fluctuations. function (PDF):

64 Appl. Math. Modelling, 1987, Vol. 11, February

 Modelling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano
Table I Exponential growth model: time variation of HP(f)]
using TPD, Gaussian distribution, and SMA for df,= 107, r,
= 0.055, CV_= 0.20)
E(P)(X1061
Gaussian TPD SMA
(24) 0 10 10 10
10 17.4 17.4 17.4
20 30.8 30.8 30.8
30 55 54.9 54.9
Uniform distribution 40 99.4 99.1 99
50 182 180.7 180.1
Sometimes the decision maker can specify a finite inter- 60 337.1 332.4 330.2
val [-b, b] within which the rate rl will lie, such that: JO 632.1 616.3 609.2
80 1200 1150.8 1130
90 2304.5 2162 2103.6
p(rl) = A,-b < rl d b 100 4480.9 4082.7 3927.3

Then:
Second-moment analysis (SMA)
ebt _ e-bt
When the distribution of rl is unknown, SMA is resorted
W(t)1 = Poerd 2bt (25) to. Let the mean and variance of the random variable
rl be 0 and o$. It can easily be seen that:
p$+d(& - e-bt)
u+(t) =
4t2b2 (32)

[bt(eb’ + edbr) - eb[+ epbr] (26) From equations (20) and (32):

Two-point distribution (TPD) E[P(t)] = Poe+ +?qq (33)

Consider the situation when the decision-maker can spe-
cify the first two moments of the random variable rI. The second moment of the variable P(t) is given by:
Suppose rl has zero mean and variance o$. TPD can E[P2(t)] = P$e2Q’(1+ 20$t2) (34)
then be expressed as:
and, hence, the variance:
c+(t) = Pie2Qro$t2 (35)
p(rJ = ; 6(r, - a,) +; 6(r, + UJ (27)
which, for small a&, reduces to equations (22) and (30).
Therefore: It is instructive to make a numerical comparison of
the evolution of moments of P(t) when uncertainty in
growth parameter r is characterized by different proba-
E[erlr]= bility distributions, For the purpose of illustration, a
hypothetical data set is taken. The numerical results
for E[P(t)] and dp(t) are presented in Table I and Figures

+ k6(rl
1
+ ur) dr, = cosh(c,.t) (28) (I)-(3).
It may be noted in passing that uncertainty in growth
rate is reflected in all the moments of P(t). The stochastic
Substituting this into equation (20) gives: mean E[P(t)] and the deterministic value P(t) may differ
considerably depending upon the magnitude of a, and
E[P(t)] = POerdcosh(mJ) (29) the forecasting period.
Proceeding on the same lines, the variance becomes:
Correlated errors in initial conditions and growth
o$( t) = Pie2*dsinh2( aJ) (30) rate
Hence, the coefficient of variation is: In the previous sections it is assumed that uncertainties
in initial conditions and growth rate are independent.
CV,(t) = tanh(aJ) (31) In some cases, it may be more realistic to consider the
This equation can be used to estimate the time period case when random errors in these variables have the
in which CV,(t) is within desirable bounds when the same cause and are, therefore, correlated. Expressions
magnitude of dispersion in the growth rate parameter for moments of P(t) are derived in the following subsec-
is given. tions for two cases: (a) Gaussian errors; (b) errors speci-
It is worth noting that the foregoing analysis can easily fied by TPD.
be extended to the case when the first three moments
of the growth rate are specified. In view of the specifica- Gaussian distribution
tion of the third moment, an asymmetrical TPD would
Let the correlation coefficient of random variables r and
be obtained from equations (7)-(9). Using these equa-
PObe such that:
tions, the corresponding first three moments of the
population can be derived. P= E[(~‘o - po)(r - dl/(wJ (36)

Appl. Math. Modelling, 1987, Vol. 11, February 65

Modelling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano

0 20 40 60 60
Time, t (y)
Figure 3 Exponential growth model: coefficients of variation
or projected forecasts using TPD and Gaussian distribution
forr

It is interesting to note that effect of correlated errors

in the expression E[P(t)] is reflected in modifying the
value of the initial mean PO. Depending on the magni-
tudes of the parameters uo, Up and p, this effect may
become significant for increasing time. The expression
for the second moment, however, is not easy to obtain
in this case.
0 20 40 60
Time, t (y)
Figure I Exponential growth model: uncertainty in growth fore- Two-point distribution
casts using TPD (rP, = IO’, r, = 0.055, CV,= 0.2)
Following Rosenblueth, l2 the joint PDF of the random
variables r and POis:

P(T> PO) = T [S(r - (r, - uJ)S(P, - (PO - q))

( 1

+ S(f- - (To + uJ)Wo - (PO + ao))l

+ 7 [S(r- (r, - a,))S(Po- PO+ VI))]
i 1
+ S(r- (ro+ %))Wo- (PO- ~o))l (39)
Using equation (2b), the first moment of P(t) is:
E[P(t)] = [lP,e”@(r, PO) dr dPo (40)
From equation (39) and (40):

E]P(t)l = A$_!! [PO- ao) e(ra-%)’

1-P
+ (PO+ uo) e@-qjf] + -
0 20 40 60 4
Time, r(y)
[(PO- go) e(ro-%,)t+ (PO - co) eCrO-Q]
Figure2 Exponential growth model: uncertainty in growthfore-
casts using Gaussian distribution (rP, = IO’, r, = 0.055, CV,= 0.2)
= [PO+ pa, tanh(u,.t)] eQ’cosh(u,t) (41)
From equation (2b): Proceeding along similar lines, it is found that:

E[Pyt)] = (PO’+ oj)

E[P(t)] = E(Pae”) = $ [E(e”fo+rr)]a=O (37)

Using the result: (I+%) e2’+cosh(2u$) (42)

E(e’+) = exp[-E((p2)/2]
u;(t) = [(PO’+ o$ + 2puoPo) cosh(2u.J)
for a centred Gaussian random variable cpgives:
- (POcash art + puo sinh c$)~] e2ror (43)

E[P(t)] = (P, + crOu&)exp

which, for uncorrelated errors in PO and r, reduces to
(36) (30). It is worth observing that explicit analytical,

66 Appl. Math. Modelling, 1987, Vol. 11, February

 4 Modeling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano

I
I 0 20 40 60 80
0 20 40 60 60
Time, t (y)
Time, r (y) Figure 5 Logistic growth model: time variation of CVp(t) (PO
Figure4 Logistic growth model: growth trajectory of E[P(t)l and =0.1,K=20,r=0.15,cr,=0.09)
deterministic p(t)(f,, = 0.1, K= 20, r= 0.15, a, = 0.09)

E[Z=(t)] = f j: 9, 9,
though approximate, expressions for the moments of r=l j=l k=l

P(t), can be readily obtained by using TPD.

[PO+ (-)‘@[K + (-)‘mk]2
[(PO -(-)GqQ + (K + (-)hK - PO -(-)i~o)q]2 (47)
Logistic growth with uncertainty in data
which, in the limit for large times, tends to [K2 + o$].
The effects of uncertainties in data for the exponential Therefore, in the steady state:
growth model were studied in previous sections. This
limo++(t)= a’+ (48)
section gives the analysis of the projection errors in a km
logistic growth model described by nonlinear differential It may be noted that in the steady state, only fluctuations
equation (3). For illustrative purposes it is assumed that in the carrying capacity contribute to the limiting var-
uncertainties in the initial condition PO and in para- iance of the variable P(t).
meters r and K are mutually independent. The TPD A comparison between the behaviour of the coeffi-
can be written as: cient of variation of the growth forecasts, based on the
logistic and exponential models when the initial condi-
p(r, K, PO)= $[%r- (ro- 4) + S(r-(ro + 4)l tion is uncertain, is of some interest. Logistic model
CV,(t) varies with time as depicted in Figure 5, whereas
x [6(K -(K - a/_)) + 6(K -(K + crK))]
the exponential model CV,(t) remains constant (see
equation (17)). The more general case when the uncer-
x [%Po -(Po - cd) + (PO -(Po + go>1 (44) tainties are correlated can be analysed in a similar man-
ner as in the previous section.
Hence, from equation (4): In conclusion, it may be emphasized that TPD pro-
vides, in an extremely efficient manner, the probabilistic
PoK description of the time development of the moments
E[P(t)l PO+ (K - P,)e-” of the variable P(t) evolving according to logistic
growth. In contrast, when uncertainties in PO, K and
p(r, K, PO) drdKdP, (45) r are assumed to be Gaussian variates, the analytical
expressions E[P( t)] and 4(t) are not amenable to
Substituting from equation (44), this yields:
expressions in closed form.
r2 2 2

E[fYt)l= 67 >: y Updating of projections

r=l j=l k=l
On the basis of an assumed probability distribution of
[PO+ ( -)bo][K + ( -)juK] ri, the moments of P(t) have been obtained so far. Now
suppose that at a subsequent period t = T some new
information about the growth parameter rl is received.
It may then be desirable to use this information to
where:
update the probability distribution of rl, employing
77= expM0 + HkOl (46) Bayesian statistics.12
To illustrate this, if the bounds of rl are known, i.e.:
In the limit of asymptotically large times, it is found
that E[P(t)] tends to the mean carrying capacity K and -bSr,Sb
the variability in POand r, is not reflected in the limiting then a uniform distribution can be used:
value as expected. In the evolutionary stage of logistic
growth, however, these uncertainties in the parameters p(rl) = l/W)
and initial conditions are reflected in E[P(t)], and the where:
growth trajectory of the mean value of Z’(t) deviates
E(r,) = 0 cr, = b/(3)”
from that of the deterministic one, as depicted in Figure
4. Transforming the uniform distribution into a TPD
Proceeding along similar lines: with the same mean and variance:

Appl. Math. Modelling, 1987, Vol. 11, February 67

Modelling data uncertainty in growth forecasts: Karmeshu and F. Lara-Rosano

‘tolerable limits’ if the uncertainty in the growth r&e’s

is not very large and the period of projection is also
of the order of the time-scale, in relation to parameters
with: of the process under study. An important feature of
this analysis is the calculation of analytical expressions
P’(r, = -0;) = P’(r, = a,) = l/2 of the moments of P(t). These explicit expressions can
Defining S as the event providing new information, be used in turn by decision-makers to assess the relative
one may assign the following probabilities: importance of various sources of errors.
It is found that TPD provides in an efficient, although
Pr(Slr, = -gJ = Ly Pr(S(r1 = a,) = p approximate, manner analytical results for a nonlinear
In view of these new probabilities, the distribution of stochastic logistic growth model, whereas analytical
rl is updated by applying Bayes’ theorem as follows: results, with parameters characterized by Gaussian dis-
tribution, are almost impossible to obtain and one may
P”@l= -cJS) = CN/(CY
+ p> have to resort to numerical procedures.
P”(r1 = a&s) = p&X + p)
These probabilities correspond to concentrations of Acknowledgements
asymmetrical TPD (see equations (7)-(9)). In some The authors thank Professor E. Rosenblueth for many
cases it is convenient to use more concentrations to helpful discussions and for reading the manuscript. They
ensure that p(r) converges to a value within the interval also wish to thank the referees for various suggestions.
defined by the concentrations. l2
Having obtained this new estimate of rI at t = T, one
can proceed along similar lines to obtain the moments References
of P(t) using the updated value of the growth rate para- Bright, J. R. ‘Technological forecasting for industry and govern-
meter r, and the revised values of E[P(T)] and v~(T) ment’, Prentice Hall, New York, 1968
Soong, T. T. ‘Random differential equations in science and engi-
as the new initial conditions. neering’, Academic Press, New York, 1973
Sykes, Z. M. ‘Some stochastic versions of the matrix model for
population dynamics’, J. Am. Stat. Assoc., 1969,44, 111-130
Conclusions Pollard, .I. H. ‘On simple approximate calculations appropriate
The modelling of uncertainty in growth forecasts arising to populations with random growth rates’, Theor. Popul. Biol.,
1970,1,208-218
from errors in the data has been discussed. These errors 5 Banjamin, J. R. and Cornell, C. A. ‘Probability, statistics, and
could arise as a result of a large number of sources e.g. decision for civil engineers’, McGraw-Hill, NY, 1970
scarce data, mis-specification of model parameters, and 6 O’Neill, R. V. and Gardner, R. H. ‘Sources of uncertainty in
uncertainty both in the initial conditions and growth ecological models’, in ‘Methodology in systems modelling and
parameters. Various forms of probability distributions simulation’ (eds Zeigler, Elzas, Klir and Oren), North-Holland,
Amsterdam, 1979
characterizing uncertainty in data are proposed. These 7 Karmeshu and Pathria, R. K. ‘Time development of a Markov
probability distributions, namely Gaussian, uniform, process in a finite population-application to diffusion of informa-
and TPD, are used to derive the expressions for the tion’, J. Math. Social., 1980,7,22%240
moments of the variable P(t) evolving in a manner char- 8 Karmeshu and Pathria, R. K. ‘Diffusion of information in a ran-
dom environment’, J. Math. Social., 1980,7,215-227
acterized by exponential or logistic growths described, 9 Stoto, M. A. ‘The accuracy of population projections’, J. Am.
respectively, by either linear or nonlinear differential Stat. Assoc., 1983,78, E-20
equations. 10 Lind, N. C. ‘Modelling of uncertainty in discrete dynamical sys-
These differential equations are studied in the pre- tems’, Appl. Math. Modelling, 1983,7(3), 146-152
sence of random errors in the data within the framework 11 Rosenblueth, E. ‘Point estimates for probability moments’, Proc.
Natl. Acad. Sci., USA, 1975,72,3812-3184
of random differential equations. An interesting aspect 12 Rosenblueth, E. ‘Two-point estimates in probabilities’, Appl.
that emerges from this study is that the randomness Math. Modelling, 1981,5(5), 329-335
in the growth paranfeter modifies the evolution of all 13 Jaynes, T. E. ‘Information theory and statistical mechanics’,
the moments of f’(t). This situation may be compared Phys. Rev., 1957,106,620-630
14 Shannon, C. E. ‘A mathematical theory of communication’, Bell
with the one when the stochasticity in the differential Syst. Tech. J., 1948,27,623-659
equation occurs in an additive manner.2 In this case, 15 Tribus, M. ‘Rational description, decisions and designs’, Perga-
the additive stochasticity does not affect the mean evolu- mon Press, Oxford, 1969
tion of P(t). In view of the randomness appearing in 16 Tiwari, J. L. and Hobbie, J. E. ‘Random differential equations
as models of ecosystems. Initial conditions and parameter specifi-
the parameters, the magnitude of relative fluctuations
cations in terms of maximum entropy distribution’, Math. Biosci.,
in forecasts of P(t) can become extremely large as the 1976,31,37-53
period of forecasting increases. For problems of practi- 17 Hoel, P. G. ‘Introduction to mathematical statistics’, John Wiley,
cal interest, the relative fluctuations can remain within NY, 1954

68 Appl. Math. Modelling, 1987, Vol. 11, February