When Will A Large Complex System Be Stable

J. theor.
BioL (1985) 113, 153-156
When Will a Large Complex System Be Stable?
JOEL E. COHENt§ AND CHARLES M. NEWMAN~

t Rockefeller University, 1230 York Avenue, New York, N Y I0021-
6399, U.S.A. and :~ Department of Mathematics, University of
Arizona, Tucson, AZ85721, U.S.A.
(Received 26 March t984, and in revised form 26 September 1984)
May (1972, 1973) and Hastings (1982a, b, 1983a, b) announced criteria for
the probable stability or instability, as n]'o~, of systems of n linear ordinary
differential equations or difference equations with random coefficients fixed
in time. However, simple, explicit counter-examples show that, without
some additional conditions, the claims of May and Hastings can be false.
May ( 1972, 1973) announced criteria for the probable stability or instability
of a system of n linear ordinary differential equations with random
coefficients fixed in time, as n increases to infinity. He gave these criteria
the ecological interpretation that, as the number n of species in an ecological
community increases, increasingly severe constraints must be imposed on
the distribution of the interspecific interaction coefficients to assure the
stability of the community. The constraints could be satisfied, as n increases,
by reducing either the connectance (the fraction of interaction coefficients
that are not zero) or the variance of the non-zero interaction coefficients.
This interpretation assumed that the interaction coefficients between species
can usefully be described as random variables fixed in time. May's claims
attracted attention because they appear to contradict an ecological folk taw
that increased complexity (measured by the number n o f interacting species
or the connectance among them) promotes increased stability. Pimm (1984)
has comprehensively reviewed the vexed question of the complexity and
stability o f ecological systems.
Recently Hastings (1982a, b, 1983a, b) announced variants and proofs of
May's claims.
The purpose of this note is to point out that some simple,, explicit
counter-examples (Cohen & Newman 1984a) show that May's and
Hastings's claims are false in the generality with which they are stated. The
counter-examples are quite robust and include broad sets of models that
§ To whom correspondence should be addressed.
153
0022-5193/85/050153 + 04 $03.00/0 (~) 1985 Academic Press Inc. (London) Ltd
154 3. E. C O H E N AND C. M. NEWMAN
fall within the hypotheses of May and Hastings. In the models we have
constructed, May's and Hastings's announced criteria for stability and other
assumptions are satisfied but the probability that the system of differential
or difference equations is stable does not approach 1 as n increases. In
another example, based on Hastings's use of, difference equations, the
announced criteria for instability and other assumptions are satisfied, but
the probability that the system is unstable approaches zero as n increases.
By way of illustration, we shall focus here on one specific claim made
by Hastings (1982a) and the simplest o f our counter-examples.
Hastings (1982a) considered the time-homogeneous system of difference
equations
x,+ I = Bx, ( 1)
where x, is a real n-vector and the random n x n matrix B has all elements
independently and identically distributed and fixed in time. He defined the
system (1) to be stable if the spectral radius r(B), the largest of the absolute
values of the eigenvalues of B, is less than 1.
Hastings supposed that each element Bo of B is 0 with probability 1 - C,
0_<C_<I, and with probability C is drawn, once and for all, from a
distribution with mean and all odd moments 0 and variance a 2. The diagonal
elements of B have the same distribution as the off-diagonal elements of
B. In the course of his argument, Hastings assumed that, as n increases, a
is fixed and C = k / n , where k is a fixed positive constant. He suggested
that data justify the assumption that C = k / n in the context where n is the
number of species in an ecological community and B is a matrix that
describes the species' interactions, which are assumed to be fixed in time.
Under these assumptions, he claimed that if n C a 2 = n Var B~j < 1 then r(B) <
1 with probability approaching 1 as nl'oo, while if n C a 2 > 1 then r ( B ) > 1
with probability approaching 1 as nl'oo.
For a counter-example to Hastings's claims concerning the stability of
the system (1), choose any finite positive constant k. Suppose X o, are
independent normally distributed real random variables with mean 0 and
positive variance a 2. Then the probability p that any one of the X0's exceeds
1 in absolute value is positive. Let (7, = k / n and let each element B~j of the
matrix B be independently and identically distributed, equal to 0 with
probability l - C , and equal to Xo with probability (7,. We have proved
that the probability of instability does not approach 0 as n increases, even
though we can choose ka 2= nC,,a2< 1, contrary to Hastings (1982a).
The assumption in this counter-example that each X~j is normally dis-
tributed may be replaced by the much more general assumption that each
Xij has a positive probability of exceeding 1 in absolute value.
LARGE COMPLEX SYSTEM S.T A B I L I T Y 155
H. M. Hastings has suggested (personal communication, January 1984)

that one can understand such counter-examples to asymptotic stability in
terms o f the connectedness properties o f the random (directed) graph
obtained from the matrix B by placing a directed edge from i to j if B u # 0.
In the spirit of that suggestion, we note that counter-examples to
asymptotic instability also follow from certain connectedness properties.
When each edge occurs with probability k / n and k < 1, it is known (see
Cohen & Newman 1984b) that there is, as n ~ oo, a strictly positive probabil-
ity that the graph contains no (directed) cycles. It is easy to see that the
corresponding matrix B then has B " = 0 for a sufficiently large power m
so that r(B) = 0. It follows that the probability of stability does not approach
0 as n increases, even though we can choose ka 2= nC, a'-> 1, contrary to
Hastings's (1982a) claim concerning asymptotic instability.
Counter-examples to asymptotic instability which do not require k < 1
are also possible. In particular, we have constructed (Cohen & Newman
1984a) a sequence of random matrices B with independently and identically
distributed elements B u such that l i m , r ~ n V a r B u = o o while l i m , , ~
P(r(B) < 1)= 1, again contrary to Hastings (t982a).
Though they fall within the hypotheses of May and Hastings, our counter-
examples to their instability claims may be ecologically artificial in that we
require either that k be less than 1 or else that X 0 can take on very large
values (with small probability).
Our counter-examples leave open the possibility that May's conjectures
may be true under certain conditions as suggested by numerical studies
(McMurtrie, 1975 and D. E. McClure, personal communication, June 1984).
Recently Geman (1984) has given conditions sufficient to prove the
stability part of May's conjectures. Specifically, let {bo}, i= 1,2,... , j =
1 , 2 , . . . , be independent and identically distributed random variables with
mean zero and finite standard deviation o-. For each n, let B, be the n "<n
matrix with i,j element bu. Geman shows under certain hypotheses on the
moments o f b u that as nToo, the lim sup o f the spectral radius r ( B , / n 1/2)
is less than or equal to o- with probability 1. Robert S. Maier (personal
communication, September 1984) has informed us without providing details
that he has established the instability part of May's conjectures under certain
conditions.
We have also analyzed a model, different from that considered by May
and Hastings, that drops the assumption that the random interaction
coefficients between species are fixed in time (Cohen & Newman 1984a).
For some cases o f this model, which is based on products o f random matrices
(Furstenberg & Kesten, 1960), stability or instability is determined
asymptotically by criteria of exactly the form suggested by May. For other
156 J. E. C O H E N A N D C. M. N E W M A N
cases o f this same model, criteria different in form from those p r o p o s e d by

M a y determine the stability or instability o f the system.
We agree with this observation o f M a y (1971): " I n nature we deal not
with arbitrary c o m p l e x systems, but rather with ones selected by a long a n d
intricate p r o c e s s . . . ; mathematical t h e o r e m s tend to deal with general
c o m p l e x systems, which are quite a n o t h e r matter". H o w e v e r , this is only a
t e n d e n c y o f mathematics, not a requirement. O u r stability results for prod-
ucts o f r a n d o m matrices m a y eventually be e x t e n d e d from the case o f
i n d e p e n d e n t matrices to more realistic stochastic models with n o n l i n e a r
d y n a m i c s and m e m o r y .
Since early investigations ( G a r d n e r & A s h b y 1970) o f the stability o f
r a n d o m systems were motivated by models o f the brain, our analysis o f
models with r a n d o m l y fixed and with r a n d o m l y c h a n g i n g coefficients m a y
be o f interest in n e u r o b i o l o g y as well as in ecology.
R. M. May and H. M. Hastings kindly reviewed earlier drafts of this paper. Our
work was supported in part by U.S. National Science Foundation grants DEB80-
11026 and BSR84-07461 to J.E.C. and MCS80-19384 to C.M.N., and by the hospital-
ity of Mr and Mrs William T. Golden to J.E.C.
REFERENCES
COHEN, J. E. & NEWMAN, C. M~ (1984a). Ann. Prob. 12, 283.
COHEN, J. E. • NEWMAN, C. M. (1984b). A stochastic'theo~ of community food webs: I.
Models and aggregated data. Proc. Roy, Soc. (London) set. B, Section 4, (In press).
FURSTENBERG, H. & KESTEN, H. (1960). Ann. Math. Star 31, 457.
GARDNER, M. R. & ASHBY, W. R. (1970). Nature 228, 784.
GEMAN, S. (1984). Ann. Prob. (In press).
HASTINGS, I-[. M. (t982a). J. Theor. Biol. 97, 155.
HASTINGS, H. M. (1982b). Bull. Am. Math. Soc. 7, 387.
HASTINGS, H. M. ( 1983a). In: Population Biology: Proceedings of the International Conference
Held at the Universi O, of Alberta, 1982, Lecture Notes in Biomathematics. (Freedman, H. I.
& Strobeck, C. eds). Vol, 52, pp, 355-358. New York: Springer-Verlag.
HASTINGS, H. M. (1983b). In: Current Trends in Food Web Theory, Report on a Food Web
Workshop (De Angelis, D. E., Post, W. M. & Sugihara, G. eds). ORNL-5983, Oak Ridge:
Oak Ridge National Laboratory.
MAY, R. M. (1971). Math. Biosci. 12,59.
MAY, R. M. (1972). Nature 238, 413.
MAY, R. M. (1973). Stability and Complexity in Model Ecosystems. Princeton: Princeton
University Press.
MCMURTRIE, R. E. (1975). J. Theor. Biol. 50, 1.
PIMM, S. (1984). Nature 307, 321.

When Will A Large Complex System Be Stable

Uploaded by

Copyright:

Available Formats

When Will A Large Complex System Be Stable

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

When Will A Large Complex System Be Stable

Uploaded by

Copyright:

Available Formats

J. theor.

BioL (1985) 113, 153-156

When Will a Large Complex System Be Stable?

JOEL E. COHENt§ AND CHARLES M. NEWMAN~

H. M. Hastings has suggested (personal communication, January 1984)

cases o f this same model, criteria different in form from those p r o p o s e d by

You might also like