Inertial Manifolds

R. T e m a m

Inertial manifolds are new objects that have been re- Dissipative Evolution Systems
cently introduced in relation to the study of the long-
term behavior of solutions of dissipative evolution We are interested in evolutionary dissipative systems.
equations. When they exist, inertial manifolds are fi- Such s y s t e m s arise in p h y s i c s , c h e m i s t r y , me-
nite dimensional invariant smooth manifolds that con- chanics . . . . Usually the state of the system is de-
tain the global attractor and attract all the orbits at an scribed by an element ~ of a Hilbert space H of finite or
exponential rate. Most of the dynamics for the system infinite dimension (the phase space) and its evolution
under consideration takes place on this manifold, pro- on an interval of time I C ~ is defined by a function
ducing a considerable simplification in the study of the
dynamics. Also, the system obtained by restriction to u : I - - * H.
the inertial manifold is a finite dimensional system,
even if the initial system was infinite dimensional. An example of this is a chemical system for which H is
This system, called the inertial system, reproduces Rn and u(t) is an n-tuple (cl(t) . . . . . cn(t)), where cl(t),
most of the dynamical properties of the initial system. . . . . cn(t) are the concentrations at time t of the reac-
In addition to their mathematical significance, iner- tants. In fluid mechanics or meteorology, H is an infi-
tial manifolds m a y have some computational and nite dimensional function space and u(t) ~ H is a func-
physical relevance. As is shown below, inertial mani- tion
folds produce an interaction law between the small J
and large wavelength components of a flow; alterna- u(t) : f~--~ ~,
tively, using a t e r m i n o l o g y n o w c o m m o n in the
physics literature, we can say that the small wave- where f~ C ~3 is the domain filled by the fluid; here
lengths are enslaved by the large ones. This will be u(t) might be the velocity field (at time t) throughout
discussed in more detail below; we shall also relate in- fL or the field of velocities and temperatures.
ertial manifolds to slow manifolds, which are well
known in meteorology for short-term weather forecas-
ting.
From the computational point of view, inertial man-
ifolds and approximate inertial manifolds produce fi-
nite dimensional objects that approximate the attractor
(perhaps a fractal set). They reproduce the coarse
structure of the attractor that we want to approximate,
while they neglect the fine details. Simple approxi-
mate inertial manifolds have been constructed and, al-
though this is still very recent, promising numerical
algorithms have been derived using these manifolds.
This too will be briefly discussed below.

68 THE MATHEMATICAL INTELLIGENCER VOL. 12, NO. 4 9 1990 Springer-Verlag New York
 In many cases of interest the evolution of u is deter-
mined by the solution of an initial value problem for a
differential equation in H, say

du(t)
dt - F(u(t)), t > 0, (1)

u(0) = u0, (2)

where F is a function from H (or part of H) into itself.

When the dynamics of the system is trivial, the system
evolves towards rest or towards a steady state. In (1), Figure 1. Absorbing Set.
(2) this is reflected by the fact that the solution u(t) of
(1), (2) converges as t ~ oo to an equilibrium point u,,
i.e. a solution of sending u 0 to u(t).
Concerning the operators S(t) it may also be appro-
F(u,) = 0. (3) priate, in some situations, to consider only discrete
values of t; for instance, when we consider a Poincar6
For turbulent systems, the dynamics is not trivial. map.
The solutions of (1), (2) remain always time depen- From the physical point of view a dissipative system
dent; their evolution looks essentially unpredictable at is a system dissipating energy; hence some real func-
our present level of understanding of nonlinear dy- tion of u(t) will decrease as t increases. From the math-
namics. Some of the relevant questions are then ematical point of view a well-accepted definition of
dissipativity for a semigroup S(') or equation (1) is the
9 What can we predict concerning the evolution of existence of a bounded absorbing set. This is a bounded
u(t)? set ~0 C H such that:
9 H o w can one effectively compute the solutions of
(1), (2) on a large interval of time? For every bounded set ~ C H,
9 Assuming that we know the solutions of (1), (2), S ( t ) ~ C ~o for all sufficiently large t.
what can we learn from this (abundant) information
that is physically relevant? H o w can we economi- Thus, by contrast with Hamiltonian (nondissipative)
cally extract from a large system a specific item of systems, the orbits do not wander in the whole space
important information? H and they do not fill any open nonempty set in H.
We shall try to show how inertial manifolds could Instead they concentrate in the region ~0. The time at
perhaps provide a new insight to these problems in which a given orbit enters ~0 depends on its initial
nonlinear dynamics. point u 0, but all orbits eventually enter ~0; and the en-
tering time is uniformly bounded for u 0 in a bounded
set ~. (See Figure 1.)
Absorbing Sets

We assume that the evolution of the system under

consideration is governed by a semi-group of oper- The Global Attractor
ators
S(t):H---~H, t~0, Another aspect of dissipativity is the existence of a
global attractor. This is a compact set ~ C H enjoying
where the following properties:
S(O) = L S(t + "r = S(t) " S('t), t,,r ~ O. (4)
S(t) ,~ = ~l, V t ~ O, (5i)
Knowing the state u(';) of the system at time % its state
at time ~ + t,t > 0 is given by attracts all bounded sets,
i.e., for every bounded set ~ C H,
u(t + "0 = S(t)u('r)(= S('r S(t + "r)u(0)if ,r > 0). d i s t ( S ( t ) ~ , ~ ) := sup inf IIS(t)x - yiJn (5ii)
9 . , ^ xE~yE.~
renas ro u as t ~ + oo.
For example, under suitable hypotheses on F, the ini-
tial value problem (1), (2) is well posed, i.e., for any u 0 In particular, each orbit converges to ~/as t ~ oo 9
E H there exists a unique function u from [0, oo) to H
satisfying (1), (2): in this case S(t) is the mapping dist(S(t)u o, ~ ) --~ 0 as t ~ % V u o. (6)

THE MATHEMATICAL INTELLIGENCER V O L . 12, N O . 4, 1990 69

 Because an abundant literature is already available on t w e e n finite and infinite dimensional dynamical
attractors, we shall just emphasize here that a global systems, already transparent in the results on dimen-
attractor is maximal for the inclusion relation among all sion of attractors, deserves to be further investigated;
attractors and that conditions (5i) and (5ii) make the this was one of the motivations for inertial manifolds.
set ,~ unique, if it exists. If the dynamics is trivial, Another, more general motivation, is the imbedding
reduces to a single equilibrium point (F(u,) = 0). In of the global attractor in a smooth manifold. This is a
the nontrivial case ~ may contain or consist of the set natural problem in geometry and topology which has
of equilibrium points with manifolds connecting them been investigated by several authors under hyperboli-
(the unstable manifolds); orbits of time-periodic solu- city assumptions which may not be easy to verify for
tions; orbits of time-quasiperiodic solutions lying on a specific equations (see in particular M. Shub's book
toms; or even more complicated sets of nonintegral and the references therein).
dimension, i.e., fractals. An inertial manifold for the semigroup {S(t)}t>~o (or
For ordinary differential equations, i.e., when H has for equation (1)) is a smooth manifold ~ of finite di-
finite dimension, the existence of the global attractor mension such that
has been known for many years and some of these
attractors have attracted particular attention (H6non, S(tfiY~ c ~ , V t/> 0, (7i)
Lorenz . . . . ). In the infinite dimensional case the exis-
tence of the attractor has only recently been proved, attracts all the solutions of
(7ii)
for some classes of dissipative evolution equations and (1), (2) at an exponential rate.
for some specific dissipative equations such as the Na-
vier-Stokes equations. By (7ii) we mean that for every u 0 ~ H there exists K1,
When a nonstationary turbulent flow takes place, K2 > 0, such that
the attractor ~ (or part of it) is the natural mathemat-
dist(S(t)uo, ~ ) <<-K1 exp (-K2t ).
ical object for the description of the permanent re-
gime. Ruelle and Takens attempt to explain the tem- The comparison of ~ and ~J~ is transparent; ~ may
poral chaos corresponding to a turbulent flow by con- not be regular and may even be a fractal set, whereas
jecturing that ~ is a complicated, fractal set (strange is required to be a smooth manifold, Lipschitz at
attractor), and that the temporal chaos is due to the least. The equality sign in (50 is replaced by an inclu-
orbits wandering along such a set. For this reason it sion sign in (7i), and while the orbits may converge to
would be, of course, interesting to understand better at an arbitrarily slow rate, they converge to ~ at an
the geometry of such sets, but actually little is known exponential rate. Finally (7ii) obviously implies that
at present. One of the geometrical aspects of attractors C ~l~. If u 0 E ~ , then by (7i) the solution u(t) = S(t)u o
that has been extensively studied, in particular in the of (1), (2) is included in ~ for all t/> 0. The restriction
infinite dimensional situation, is the dimension of the of (1) to ~ is therefore a well defined finite dimen-
a t t r a c t o r - - e i t h e r the Hausdorff dimension or the sional system called the inertial system. Under appro-
fractal dimension. For many dissipative systems it has priate hypotheses the inertial system possesses the
been shown that even if the phase space H has infinite property of asymptotic completeness: by this we mean
dimension, the attractor itself has finite dimension, that to every orbit solution of (1), (2), we can associate
and hence permanent turbulent regimes actually have a companion orbit lying in ~ and producing the same
a finite dimensional structure; this reduction of infinite behavior as t ~ ~:
dimension to finite dimension will be further dis-
cussed below. In some cases, the estimates on the di- Vu 0 E H , 3 v 0 E . ~ a n d ~ > 0 s u c h t h a t
mension of the attractors in terms of the physical data dist(S(t)u o, S(t + T)v0) --~ 0 as t ~ ~.
are physically relevant. For example, in fluid me-
chanics these estimates are in good agreement with Thus, in such cases the inertial manifold produces a
and give a rigorous mathematical proof of some fun- reduction of dimension without any loss of asymptotic
information.
damental aspects of the Kolmogorov and Kraichnan
theories of turbulence; in particular, the maximum Existence theorems of inertial manifolds have been
number of degrees of freedom of a turbulent flow pre- proved for many dissipative differential equations.
dicted by these theories is the same as the dimension Other results include the asymptotic completeness
of the global attractor attached to the flow. and further regularity results, e.g., ~ is a Cl-mani -
fold. Equations for which such results were proved in-
clude reaction-diffusion equations, the Ginsburg-
Inertial Manifolds Landau equation, and pattern-formation equations
such as the Kuramoto-Sivashinsky and the Cahn-Hil-
Infinite dimensional dissipative systems seem to dis- liard equations. However, for many dissipative equa-
play a finite dimensional behavior. The relation be- tions possessing a global finite dimensional attractor,

70 THE MATHEMATICAL INTELUGENCER VOL. 12, NO. 4, 1990

the existence of an inertial manifold is still an open folds S(t)PmH, as t increases; under appropriate hy-
problem. In particular this is the case for the Navier- potheses these manifolds are graphs above PmH con-
Stokes equations, even in space-dimension 2. It is not verging to a limit as t -+ % and the limit is an inertial
certain that an inertial manifold necessarily exists for manifold.
any dissipative evolution equation; nonexistence re- eThe Sacker method reduces the determination of 9 to a
sults have even been proved for reaction-diffusion hyperbolic equation in infinitely many dimensions.
equations and for damped wave equations similar to The e q u a t i o n is easy to derive formally: setting
the Sine-Gordon equation. P = Pm = Pmu , q = qm = Qmu , w e p r o j e c t ( 8 ) onPm H
All the existence results that have been proved give and QmH and obtain the system
an inertial manifold that is a graph. Equation (1) is
rewritten more specifically in the form
dp
du(t) --~ + Ap + PR(p + q) = O,
d-----~+ Au(t) + R(u(t)) = 0. (8)
(11)
dq
+ Aq + QR(p + q) = O.
Here A is an u n b o u n d e d self-adjoint closed positive
operator with domain D(A) C H and R is a nonlinear
operator from D(A) into H. Assuming that A-1 is com- For a trajectory lying on ~ldt, we have
pact, we find by elementary spectral theory the exis-
tence of an orthonormal basis of H consisting of eigen- q(t) = cI)(p(t)), V t i> 0,
vectors wj of A:
and by elimination we readily find
{ Awj = )*jwj, j = 112 . . . . .
0 < ).1 <~ )*2 . . . . Xj--+ ~ a s j - - + oo. - ~ ' ( p ) ' ( P R ( p + ~(p)) + Ap) + Ac~(p) (12)
+ QR(p + C~(p)) = O,
For a given integer m we denote by Pm the orthogonal
projection of H onto the space spanned by w 1. . . . . where ~ ' is the Fr6chet differential of ~.
w m. Let Qm = I + Pm" We look for ~ as the graph of a We can obtain a good idea of the meaning of the
function spectral gap condition (10) by considering the case of
c~: pmH --+ QmH. (9) reaction-diffusion equations with space-periodicity
b o u n d a r y conditions. In this case the condition is
In this case, besides the technical hypotheses, the slightly weaker than (10), namely
major and most restrictive hypothesis is a spectral gap
hm+l - - )*m ~ K4, (13)
condition
and in space-dimension n the )*m are the numbers
)*m+~ -- Xm ~> K3()*m~+l + )*m~), (10) n
(ki)2, kl . . . . . k, E ~. (14)
where K3 is a constant depending on the equation: be- i=1
cause )*m --+ ~ as m --+ % (10) implies that the spec-
By well-known results in number theory any positive
trum of A must contain large gaps.
integer is of the form (14) if n/> 4, and condition (13)
The methods that are available for the determination
is not likely to be satisfied; while if n = 2, there are
of 9 are the following:
arbitrarily large gaps in the set of numbers (14) and
eThe Perron-Lyapunov method, which reduces the deter-
therefore (13) is satisfied for arbitrarily large values of
mination of cI) to the solution of a fixed-point problem
~ - 9 = 9 of a certain mapping ~-. This method re-
sembles that used by Perron and Lyapunov for the QmH
construction of the central manifold, of which inertial
manifolds are a global analog.
eThe Cauchy method is based on the construction of in-
tegral curves. We choose an appropriate large sphere-
like set F in PmH and consider the family of integral
curves
E = tO S(t)F.
t~0

It can be shown that the closure of s is a graph above

PmH and that s is an inertial manifold. Figure 2. Convergence of orbits to the attractor and to the
eThe Hadamard method consists of studying the mani- inertial manifold.

THE MATHEMATICAL INTELLIGENCER VOL. 12, NO. 4, 1990 7 1

 m. For reaction-diffusion equations the principle of on the manifold. Because the inertial manifold ~2~ is
spatial averaging permits a weakening of (13) and approached at an exponential rate, we see that after a
leads to existence results in space-dimension >2. certain transient period, the orbits lie on ~ during the
In order to resolve the unsolved problems of exis- permanent regime: during the permanent regime small
tence of inertial manifolds one should look perhaps for and large wavelengths phenomena are connected by the
more general forms of inertial manifolds: less regular quasistatic law (15). Alternatively, at each time the
manifolds (e.g., H,Older continuous) and manifolds small wavelengths are solely and fully determined by
that are not graphs (Figure 2). the large ones. We can say also in the terminology of
Haken that the small eddies are enslaved by the large
Turbulence ones.
The existence of an inertial manifold implies the ex-
At this point, it is not clear to what extent inertial istence of an interaction law for the system under
manifolds can help understand turbulence, but we s t u d y . For fluid m e c h a n i c s , t h e r m o h y d r a u l i c s ,
would like to describe here the connection existing be- damped waves . . . . it is not clear whether such laws
tween these theories. should exist. We shall see however that approximate
For the sake of simplicity we consider an evolution forms of such laws do exist.
equation (8) where A is the abstract linear operator
corresponding to an elliptic differential operator asso- Meteorology
ciated with the space-periodicity boundary conditions.
In this case the eigenvectors wi are simply proper com- Before and independently of the concept of inertial
binations of sines and cosines from the Fourier series: manifold, a related concept was developed in metero-
logy, under the name of slow manifold. This concept
sin k. x, cos k" x
has been developed on heuristical grounds for compu-
k = (k1. . . . . kn) ~ ~n, x = (x 1. . . . . x~) ~ ~ ;
tational purposes without any attempt at mathemat-
ical justification.
we are assuming that the period-cube is (0, 2~r)n.
The idea is that the meteorological variables u corre-
We write the series expansion of u
sponding to wind velocity, temperature, and possibly
oo some concentrations, are governed by an equation of
u(t) = ~ uj(t)wj form (1). The permanent regime has been reached, so
j=l
that the orbit u(t) lies on the attractor ~ and in prin-
or ciple a n y m e a s u r e m e n t u 0 should correspond to a
point on ~. Of course this information is of no use
u(t,x) = ~ {ak(t ) sin k . x + bk(t ) cos k. x,}
kEN n
since little is known of this attractor.
In meteorology there is a "natural" decomposition
and consider a d e c o m p o s i t i o n u = Pm + qm of U of H into PmHQm H, where QmH corresponds to small
slightly different from that above; namely for m E ~, wavelengths, called the gravity waves, while PmH cor-
responds to the large wavelengths called the Rossby
Pm = Pro(t,x) = ~ {ak(t) sin k " x + bk(t ) cos k" x} waves. The conjecture leading to the slow manifold is
Ikil~m
for some i
that we have i n d e e d a relation of the form (10),
q,, = q)(Pm), the graph of ~ - w h o s e construction is not
explained here-being the slow manifold 8. Further-
qm = qm(t,x) = ~ {ak(t) sin k" x + bk(t ) cos k- x}. more, the slow manifold 8 contains (0,0) and is tan-
Ikil>m gent to the PmH space at (0,0). The gravity and Rossby
f o r all i
waves are on a totally different scale, a fact that should
be compared to the spectral gap condition (10).
Hence Pm is periodic of period 2~r in each direction and The slow manifold is used in meteorology for short-
qm is the superposition of periodic functions of period term forecasting. The initial data u0 (either measured
2~r/m in each direction. Although the cut-off number or computed) should be on 8 (and even on ~!). This
m is arbitrary here, we can say that Pm corresponds to is not the case, due to errors in measurement and/or
large wavelengths and qm to small wavelengths. computation. For short-term weather forecasting it is
It appears that this phenomenon of interaction of important to have a good initialization (u0 ~ ~ ), so the
small and large wavelengths is an important aspect of measured u 0 is approximated via appropriate algo-
turbulence in general (fluids or waves) and of statis- rithms by a close point 1/0 (a kind of projection) on 8.
tical mechanics. N o w for a system governed by an As mentioned before, no theorem has been proved
equation of form (8), if an inertial manifold of type (9) and there is not even a precise definition of 8 in the
exists, then meteorology literature. There are also some doubts
qm(t) = r (15) about the existence of an exact slow manifold (same

72 THE MATHEMATICAL INTELLIGENCER VOL. 12, NO. 4, 1990

 u 0 a n d on the equation, the n o r m of qm(t) in H (and in
other spaces) becomes and remains small:

[qm(t)[n ~ 8, (16)

where 8 is an appropriate power of )~l/Km+~. A first

consequence of (16) is that the attractor ~ lies in the
strip of width 28, [qm[n ~ 8. Then with an asymptotic
expansion it can be s h o w n that the second equation
(11) can be approximated to first order by

Figure 3. The slow manifold in meteorology. Aqm + QmR(pm)= O.

More precisely, after a certain time the orbits enter the
region
[Aqm(t) + QmR(pm(t))[n ~ 82. (17)

Hence the manifold ~ 1 of equation

AQmq~ + QmR(Pmq~) = 0 (18)

appears as an approximate inertial manifold in the

sense that it attracts all orbits in a thin neighborhood
of it, at an exponential rate (Figure 4). Of course the
attractor lies in this neighborhood of ~rd~l; this is one
more item of information on the localization of ~. Also
Figure 4. Localization of the attractor: ~ lies in the darker
shaded region. these localization results of ~ permit the construction
of new, more efficient algorithms for the numerical
approximation of the equation. For example, the usual
problem as for inertial manifolds?). However, it is gen-
Galerkin a p p r o x i m a t i o n w i t h m m o d e s consists of
erally accepted that a sort of approximate slow mani-
projecting equation (8) on PmH; one can also define a
fold does exist a n d the concept, broadly used in short-
term forecasting, is called the nonlinear initialization
nonlinear Galerkin approximation with m modes, which
consists of some sort of projection of the equation (8)
(the linear initialization being a mere projection on
onto ~ 1 . Other approximations based on m-dimen-
PmH) (Figure 3).
sional analytic manifolds are also being investigated.
A p p r o x i m a t e Inertial M a n i f o l d s In finite dimensions the analogs of (17) a n d (18) are
respectively semi-algebraic sets a n d algebraic sets of
In the absence of a definitive result (positive or nega- large dimension; in the infinite dimensional situation
tive) on the existence of inertial manifolds, an inter- they are semi-analytic and analytic sets.
esting alternative is that of approximate inertial manifold.
This topic is the object of current research a n d we shall Further reading
give only a brief description of the concept.
A first idea is to perturb equation (8) and consider a Nonlinear Dynamics and Turbulence (general presentation)
modified equation containing A s, s > 1: J. Gleick, Chaos, New York: Viking Press (1987).
du~ D. Ruelle, The Mathematical Intelligencer 2(3) (1980), 126-137.
- - + ~ ASu~ + Au~ + R(u) = 0,
dt Attractors
where ~ > 0 is small. The eigenvalues )~m are replaced A. V. Babin and M. I. Vishik, J. Soviet Math, 28 (1982),
by ~ K~,. For a suitable s (s ~ so > 1) and for m suffi- 619-627.
A. V. Babin and M. I. Vishik, J. Math. Pures Appl. 62 (1983),
ciently large (m/-- m0(~, so)), the spectral gap condition 441-491.
is fulfilled and the approximate inertial manifold ~ is P. Constantin and C. Foias, Comm. Pure Appl. Math. 38
obtained. However, as e ~ 0, ~ deteriorates, its di- (1985), 1-27.
m e n s i o n m becoming larger a n d larger; we do not P. Constantin, C. Foias and R. Temam, Memoirs Amer. Math.
k n o w if ~J~, converges to a limit. Soc. 314 (1985), and Physica D 30 (1988), 284-296.
P. Constantin, C. Foias, O. Manley, and R. Temam, J. Fluid
Another more natural concept of approximate iner- Mech. 150 (1985), 427-440.
tial manifold results from the following observation: C. Foias and R. Temam, J. Math. Pures Appl. 58 (1979),
after a certain time d e p e n d i n g only on the initial data 339- 368.

THE MATHEMATICAL INTELLIGENCER VOL. 12, NO. 4, 1990 73

 J. M. Ghidaglia and J. C. Saut, Equations aux d~rivdes partielles
non lindaires dissipatives et syst~mes dynamiques, Paris: Her-
mann (1988).
J. K. Hale, Math. Surveys and Monographs, vol. 25, Provi-
dence, RI: Amer. Math. Soc. (1988).
E. Lieb, Comm. Math. Phys. 92 (1984), 473-480.
R. Marl6, Lecture Notes in Math., vol. 898, New York:
Springer-Verlag (1981).
J. Mallet-Paret, J. Differential Equations 22 (1976), 331.
M. Shub, Global Stability of Dynamical Systems, New York:
Springer-Verlag (1987).
R. Temam, Infinite Dimensional Dynamical Systems in Me-
chanics and Physics, New York: Springer-Verlag (1988).
Inertial Manifolds
P. Constantin, C. Foias, B. Nicolaenko, and R. Temam, Appl.
Math. Sciences, no. 70, New York: Springer-Verlag (1988).
E. Conway, D. Hoff, and J. Smoller, SIAM J. Appl. Math. 35
(1978), 1-16.
C. R. Dering, J. D. Gibbon, D. D. Holm, and B. Nicolaenko,
vol. 99, Contemporary Mathematics, Providence, RI: Amer.
Math. Soc. (1989).
C. Foias, B. Nicolaenko, G. R. Sell, and R. Temam, C.R.
Acad. Sci. Paris, S6rie 1, 301 (1985), 285-288, and J. Math.
Pures Appl. 67 (1988), 197-226.
C. Foias, G. R. Sell, and R. Temam, C.R. Acad. Sci. Paris,
S6rie 1, 301 (1985) 139-141, and J. Differential Equations 73
(1988), 309-353.
J. Mallet-Paret and G. Sell, ]. Amer. Math. Soc. 1 (1983).
B. Nicolaenko, vol. 99, Contemporary Mathematics, Provi-
dence, RI: Amer. Math. Soc. (1989).
J. C. Saut and R. Temam, Proceedings of the Luminy Con-
ference, Math. Modelling Numerical Anal. 23 (1989).
Approximation of Inertial Manifolds
F. Baer and J. J. Tribbia, Mon. Wea. Rev. 105 (1977),
1536-1539.
E. Fabes, M. Luskin and G. R. Sell, J. Differential Equations, to
appear.
C. Foias, M. S. Jolly, I. G. Kevrekides, G. R. Sell, and E. S.
Titi, Physics Letters A 131 (1988), 433-436.
C. Foias, O. Manley, and R. Temam, Math. Modelling Numer-
ical Anal. 22 (1988), 93-118.
C. Foias, G. R. Sell, and E. S. Titi, ]. Dynamics and Differential
Equations (1989, to appear).
C. Foias and R. Temam, C.R. Acad. Sci. Paris, S6rie I, 307
(1988), 5-8, and 67-70.
C. Foias and R. Temam, The algebraic approximation of at-
tractors: the finite dimensional case, Physica D 32 (1988),
163-182.
R. Marl6, Lecture Notes in Math., vol. 597, New York:
Springer-Verlag (1977) 361-378.
M. Marion, Thesis, University Paris-Sud, 1988.
M. Marion and R. Temam, Nonlinear Galerkin Methods,
SIAM J. Numerical Anal. (1988), to appear.
R. Temam, C.R. Acad. Sci. Paris, S4rie II, 306 (1988), 399-402.
R. Ternam, Indiana Univ., Preprint no. 8809, 1988; Jour. Fac.
Sci. of Tokyo, to appear.
R. Temam, Indiana Univ., Preprint no. 8812, 1988; SlAM Jour.
Numer. Anal., to appear.
E. S. Titi, MSI Preprint 88-119, 1988; Jour. Math. Anal. and
Appl., to appear.
J. J. Tribbia, Mon. Wea. Rev., 107 (1979), 704-713 and 112
(1984), 278-284.
Laboratoire d'Analyse Num~rique
C.N.R.S. et Universit~ Paris-Sud
Bdtiment 425, 91405 Orsay, France

74 THE MATHEMATICAL INTELLIGENCER VOL. 12, NO. 4, 1990