Download as pdf or txt
Download as pdf or txt
You are on page 1of 35

A Computational Method for Solitary Internal Waves

in a Continuously Stratified Fluid *

By Bruce Turkington, Alexander Eydeland, and Sheng Wang

An iterative algorithm is presented to compute steady, translational nonlinear


waves in an incompressible fluid with a stable density stratification. The method
yields solitary internal waves as exact solutions of the governing Euler equations
without the restrictions on wave amplitude or density profile involved in the
various approximate models of weakly nonlinear long waves. Computed exam-
ples of large-amplitude solutions in shallow and deep fluid regimes are given. A
natural variational principle derived from the structure of the underlying
dynamical equations forms the basis of the method. The construction of the
numerical algorithm and the analysis of the properties of solutions are both
developed from the same principle.

1. Introduction
Solitary internal waves in density-stratified fluids have been widely observed in
the oceans and the atmosphere, as well as in model experiments on the
laboratory scale. The generating mechanism for these nonlinear long-wave
motions is typically some disturbance of the pycnocline (the region of significant
density variation) which organizes in time into a packet of solitary waves, each
being a localized disturbance of permanent form and constant translational
speed. Field data for solitary waves of depression propagating along the oceanic
thermocline have been documented by a number of researchers [15; 26; 3, 25].
In this environment a wave packet evolves from an initial disturbance caused by

Address for correspondence: Professor Bruce Turkington, Department of Mathematics and Statis-
tics, University of Massachusetts, Amherst, MA 07003.
*The research described in this paper was supported by the National Science Foundation under
grants DMS-9020218 and DMS-8903172.

STUDIES IN APPLIED MATHEMATICS 85:93-127 (1991) 93


Copyright © 1991 by the Massachusetts Institute of Technology
Published by Elsevier Science Publishing Co., Inc. 0022-2526/91/$3.50
94 B. Turkington, A. Eydeland, and S. Wang

a change of tide over a submarine ridge. On the basis of amplitude estimates


inferred from the field measurements, the nonlinearity of the solitary waves
composing such a packet varies from weak to moderate, depending upon their
rank in the packet and the magnitude of the initial disturbance. On the other
hand, strongly nonlinear long waves of elevation have been reported in the
atmospheric boundary layer [11, 12, 14]. In contrast to their oceanic counter-
parts, these solitary waves often appear as isolated disturbances of large
amplitude, which in some cases contain closed eddies. The formation of this
type of surface-based wave is usually associated with the intrusion of an air
current into a nocturnal inversion. In addition, other large-scale, long-lived
disturbances in the atmosphere have been interpreted as steady, translational
solitary waves [23].
Theoretical analyses of solitary-wave phenomena are commonly based on
approximation procedures valid for weakly nonlinear long waves. For waves of
small (finite) amplitude in a general (stable) stratification, the various asymp-
totic regimes are:
(1) shallow fluid (Korteweg-de Vries) [4, 7],
(2) deep fluid (Benjamin-Ono) [5, 13],
(3) intermediate-depth fluid (Joseph, Kubota, et al.) [20, 22],
where the third regime may be viewed as encompassing the first and second
regimes as limiting cases. Alternatively, large-amplitude long waves in a weak
stratification can be treated by similar asymptotic methods [8, 9]. In each of
these first-order theories an approximate solution is obtained which represents
the dominant terms in an asymptotic expansion for the corresponding exact
solution of the governing equations-namely, the two-dimensional Euler equa-
tions for an incompressible, inviscid stratified fluid. The principal small parame-
ter e (say) defining the asymptotic expansions determines the ratio of the
vertical to horizontal length scales involved in the long-wave profile, and is
chosen variously depending upon the undisturbed density stratification (the
pycnocline structure). The balance of nonlinearity with dispersion then dictates
the e-dependence of the nondimensionalized wave amplitude and incremental
wave speed (the difference between the wave speed and the critical phase speed
associated with linear waves in the zero-wave number limit). Since the soliton
solutions of the resulting model equations for the amplitude profile are known
explicitly, these various asymptotic theories provide a complete description of
the wave phenomenon under study whenever the assumption of weak nonlinear-
ity is justified.
As a quantitatively predictive tool, however, such an asymptotic theory may
not be adequate, because the range of validity in e of the first-order approxima-
tions may be too small to include physically realistic waves. Moreover, this range
may depend upon the undisturbed density stratification in a way that is difficult
to specify, as might be expected from the different types of asymptotic analyses
used in the various regimes mentioned above. Indeed, model laboratory tests
[21, 29] have found both good agreement and great disparity between theory
and experiment, depending upon the asymptotic regime in question; these
Solitary Internal Waves 95

investigations have also stressed the lack of a systematic estimate for the range
of validity among the various regimes. (For instance, deep-fluid approximations
appear to require much smaller amplitudes than shallow-fluid approximations.)
Data for waves in the oceans have often been found to agree reasonably well
with first-order asymptotic theory, although some consistent discrepancies have
been noted. On the other hand, the strongly nonlinear waves frequently ob-
served in the atmospheric boundary layer or easily produced in the laboratory
tank almost certainly lie outside of the range of the asymptotic methods. Of
course, it is possible to extend the range of validity of the approximations by
invoking second-order asymptotic expansions such as in [16]. Besides improving
the approximate solutions in general, these refinements are necessary in certain
special circumstances when the first-order theory collapses and hence requires
rescaling [16]. Unfortunately, the derivation of the second-order terms involves
some very complicated analytical calculations and some further numerical
computations when realistic density stratifications are considered. Moreover,
there remains the problem of estimating the (extended) range over which such a
higher-order approximation is valid.
In the present paper we address the problem of computing exact solutions of
the equations governing solitary internal waves without restrictions on the
degree of nonlinearity or the form of the given (continuous) stratification. In
part, our motivation is supplied by the limitations of the various asymptotic
theories for long waves indicated above. In addition, however, our purpose is to
examine the mathematical structure inherent in the exact governing equations
themselves-namely, the (noncanonicai) Hamiltonian structure displayed by
Benjamin [6] and others [1]. As shown in [6], this structure implies a natural
variational principle for steady translational waves, which may be formulated in
several essentially equivalent ways. We adopt a particularly useful formulation
that is expressed in terms of the physically simplest specification of the solitary-
wave disturbance, the unknown function being the vertical displacement of
isopycnal surfaces. In this form the variational principle characterizes a solitary
wave as a minimizer of the (disturbance) kinetic energy subject to a constraint
on the (disturbance) potential energy associated with buoyancy. Besides possess-
ing an appealing physical interpretation, this variational approach unifies the
analytical and computational aspects of the construction of exact solutions.
Consequently, although our main purpose is to describe a numerical algorithm,
we include an explanation of the variational theory and its ramifications in
order to establish a natural and coherent setting for the discussion of the
computational method.
As is well known, finite-amplitude internal waves are solutions of the Dubreil
Jacotin-Long equation [33], a nonlinear elliptic eigenvalue problem in (r/I, A)
where r/I is the streamfunction in a frame translating with speed c, and
A = gh / c 2 • This problem may be brought to a standard semilinear form by
introducing a so-called pseudostreamfunction, and then it is amenable to a
variational analysis. Bona, Bose, and Turner [10] have detailed an existence
theory for periodic (cnoidaO and solitary waves using such an approach. Our
variational principle and its associated existence theory (which is only sketched
96 B. Turkington. A. Eydeland. and S. Wang

by us) parallel theirs, and our computational method can be applied directly to
the problem they analyse. We prefer however to develop the governing semilin-
ear elliptic eigenvalue problem from the primitive variational principle men-
tioned above in order to clarify its physical meaning and its connection with the
general principles of Benjamin [6]. Of course, different (nonvariational) meth-
ods of proving existence are also available, such as the small-amplitude and
large-amplitude results, respectively, of Ter-Krikorov [30] and Amick [2], among
many others.
In a numerical study similar to ours, Tung, Chan, and Kubota [31] have
analysed large-amplitude solitary internal waves in the Boussinesq approxima-
tion using continuation methods and monotone iteration schemes. Although the
iterative algorithm we propose shares some features with their schemes, the two
methods differ fundamentally in the manner in which solutions branches are
parametrized. In our method the wave amplitude measured by the potential
energy constraint is the prescribed parameter, while in their method the wave
speed is prescribed directly through the parameter A. Since our algorithm
determines the multiplier A along with the solution of the constrained minimiza-
tion problem, it requires no a priori information on the (supercritical) wave
speed. This connection with the variational structure simplifies the implementa-
tion of our iterative algorithm and ensures its global convergence.
The outline of the paper is as follows. In Section 2 we derive the primitive
form of the variational principle for solitary waves in a continuously stratified
fluid at rest at infinity. We then reduce the governing constrained minimization
problem to a standard semilinear form in Section 3, and also specify the
technical conditions imposed on the undisturbed density profile. We include in
Section 4 a fairly complete description of the analytical aspects of the varia-
tional theory, discussing the existence and properties of solutions. In Sections 5
and 6 we arrive at our main results concerning the abstract construction and
concrete implementation of the iterative algorithms; we prove the global conver-
gence of the iterative sequence (from an arbitrary admissible initialization) in a
general setting. In Section 7 we document some numerical experiments in-
tended to illustrate the computational method in circumstances that resemble
these encountered in the ocean and atmosphere. Throughout the paper, we
indicate how our general approach can be specialized to handle the simpler
problems occurring under the Boussinesq approximation (appropriate to weak
stratifications), but we avoid this simplification in our main exposition. Finally,
an appendix is attached in which we explain the connection between our
variational principle and Benjamin's energy-impulse principle.

2. Variational principle for solitary waves


The physical mechanism of primary importance for internal wave motions in the
atmosphere or oceans is buoyancy, which arises from a coupling between the
density stratification and the gravitational body force. At length and velocity
scales characteristics of real waves, the compressibility and viscosity of the fluid
can be neglected, and similarly the diffusivity of the nonuniform density can be
considered a secondary effect. These traditional assumptions then lead to the
Solitary Internal Waves 97

Boussinesq model, for which the equations governing two-dimensional flows are

Ux + Vy = 0, (2.la)

Pt + uPx + VPy = 0, (2.lb)

p(U t + uU x + vU y ) - Px' (2.Ic)

p( V t + uU x + vv y ) - Py - pg; (2.Id)

here x and yare the horizontal and vertical coordinates, respectively; the
primitive unknowns are density p, pressure p, and velocity (u, v); and g is the
acceleration due to gravity. Benjamin [6] has given an especially neat formula-
tion of the system (2.1) as a pair of (nonlocaD evolution equations for the
density p and the density-weighted vorticity (J' := (pv)x - (pu)y' These equations
are obtained as follows. A streamfunction '" is introduced to enforce (2.la) with
U = "'Y' v = - "'x'
Then the kinematic relation

(J' = Lp"" Lp := - a~ (p a~) - aay (p aay ), (2.2)

holds in the fluid domain, and when supplemented with appropriate boundary
conditions determines '" in terms of p and (J'. (The formally self-adjoint
operator L p is strongly and uniformly elliptic whenever p + ~ p ~ p _ for some
constants p _ < + 00, p + > 0; this condition is clearly compatible with (2.Ib) and
is assumed throughout the sequel.) Upon taking the curl of the vector equation
(2.lc, d), the (p, (J') equations result after some manipulation:

Pt + a(p,,,,) = 0, (2.3a)

(J't + a( (J', " ' ) + a(p, gy - ilV",n = 0, (2.3b)

where the notation a(<p, "') = <Px"'y - <Py"'x is used for the Poisson bracket in the
(x, y) plane.
The present study concerns nonlinear steady translational waves, namely
solutions of (2.3) having the form p = p(x - ct, y), (J' = (J'(x - ct, y) for some
(positive) wave speed c. The fluid is supposed to be bounded by fixed lower and
°
upper surfaces at y = and y = h, respectively; the fluid domain wiII be
denoted by D = {(x, y) E R2: - 00 < x < + 00, 0< y < h}. The governing equa-
tions for such waves reduce to a system for p = p( x, y), (J' = (J'( x, y) in D
expressible as

a(p,,,, - cy) = 0, (2.4a)

a( (J', '" - cy) + a(p, gy - ilV",n = 0, (2.4b)


98 B. Turkington, A. Eydeland, and S. Wang

in which c is also considered to be an unknown. For the sake of simplicity, the


wave is assumed to be a disturbance propagating in a fluid at rest; then the
boundary conditions for (2.2) are simply !/J = 0 on aD. The more general case of
waves in a stratified shear flow is studied in [32].
Solitary waves play a distinguished role both in the field observations of
nonlinear internal waves and in the modern theory of such phenomena. More-
over, there is a particularly simple representation available for these waves
which permits us to recast the (p, u) system as a quasilinear elliptic eigenvalue
problem for a single unknown TJ. The latter problem can be brought to a
classical form, and thus can be effectively treated by analytical and numerical
methods of solutions. This reformulation, which is the basis of our further
analysis, is made as follows. The undisturbed density stratification of the fluid
(at .rest) is specified by a function p(y) satisfying

PE C 2 [O,h], p(y) > 0 and pl(y) < 0 (O~y~h). (2.5)

[As remarked later, this condition can eventually be relaxed to pi ~ 0, p(O) >
p(h).] By virtue of this condition, the undisturbed equilibrium (p, u) = (p(y), 0)
is a stable hydrostatic solution of (2.3). A solitary-wave disturbance is a solution
of (2.4) satisfying the asymptotic conditions

p(x,y) ~ p(y), u(x,y) ~ 0, !/J(x,y) ~ 0 as Ixl ~ 00. (2.6)

When characterizing such a solitary wave it is natural to impose the further


restriction that every isopycnal surface [p(x, y) = const] must connect to x = - 00
and x = + 00, or equivalently, that there are no closed isopycnal surfaces
(entrained eddies). Heuristically, this restriction can be expected to hold when-
ever the wave amplitude is not too large. Under these hypotheses, the density p
can be eliminated in favor of the vertical displacement TJ defined according to
TJ(x, y):= y - p-l(p(X, y)), giving the basic representation

p(x,y) = p(Y-TJ(x,y)). (2.7)

The geometrical meaning of TJ is made plain by Figure 1, where TJ(x, y) is


shown to be the y-displacement of the isopycnal surface passing through the

y c h

-----J x ,y)

~ )
'iI
y - "

__ _ - ____ . y 0
x=0

Figure 1. Geometrical interpretation of the vertical displacement 1/(x, y) of an isopycnal surface


passing through a point (x, y).
Solitary Internal Waves 99

point (x, y) from its undisturbed level at x = ±oo. (We caution that in [4, 5] the
notation TJ has a slightly different meaning.) A wave of elevation (depression) is
represented by TJ> 0 (TJ < 0) in D. Also, TJ = 0 on aD, and TJ ~ 0 as Ixl ~ 00 for
a solitary wave.
The (p, (T) system (2.4) may now be expressed in terms of TJ. [These
manipulations make use of obvious properties of the bracket a( . , . ), an antisym-
metric, bilinear derivation for which a(f( 4»,1/1) = f'( 4> )a( 4>,1/1 ) = a( 4>,/'( 4»1/1 ).]
Equation (2.4a) yields

0= a(p(Y-TJ),I/I-cy) = p'(Y-TJ)a(Y-TJ,I/I-cy),

from which follows

a(Y-TJ,I/I-CTJ) = O. (2.8)

It suffices then to set

1/1 = CTJ; (2.9)

and indeed it is necessary that (2.9) hold whenever TJ; +(1- TJ y)2 *- 0 in D,
since then (2.8) demands that 1/1 - CTJ be constant along each isopycnal Y - TJ =
ji E [0, h] while lim x ~ +00 1/1 - CTJ = O. Furthermore, by (2.7), the requirement
that TJ; + (1- TJ y? *- 0 -is seen to be equivalent to the hypothesis that every
isopycnal connects to x = - 00 and x = + 00. Equation (2.4b) is now reduced to

o= ca(O",TJ - y) + p'(y - TJ)(Y - TJ,gy -tc2IVTJI2)

= a( Y - TJ, cO" + p'( Y - TJ) [gy - tc 2 IVTJI2]),

from which follows

o= a(y - TJ,C(T + p'(y - TJ)[gTJ -tc 2IVTJI2]). (2.10)

It suffices then to set

c - 211 VTJ 12-p'(y -


0"
TJ) = -
g -
-;}TJP'( Y - TJ); (2.11)

and, as in the preceding case of (2.9), it is necessary that (2.11) hold whenever
TJ; + (1- TJyf *- O. Recalling (2.2), the desired nonlinear eigenvalue problem
now follows upon combining (2.9) and (2.11); namely,

MTJ =
ATJ _'( Y-TJ )
-TP in D,
(2.12)
TJ = 0 on aD, TJ ~ 0 as Ix I ~ 00,
100 B. Turkington, A. Eydeland, and S. Wang

where M denotes the quasilinear elliptic operator

M'T1:= -(p(Y-'T1)'T1x)x -(p(Y-'T1)'T1 y)y -H'T1;+'T1;)p'(Y-'T1), (2.13)

and the eigenvalue parameter is defined by

A := gh/c 2 • (2.14)

The governing equation (2.12) for the vertical displacement 'T1 admits a
natural variational formulation, which we use as the basis of our subsequent
analysis. Let objective and constraint functionals, respectively, be defined as

E('T1) = j !I V'T11 2p(Y-'T1)dxdy, (2.15)


D

F('T1) = jf(Y,'T1)dxdy (2.16)


D

where

f(y,'T1):= Ii 117) [p(Y-'T1)-p(Y-n]d~.


o

The variational principle for a solution ('T1, A) of (2.12) may be expressed as

E('T1) ~ min subject to F( 'T1) = A, (2.17)

where A is a given positive constant and the admissible (competing) functions


satisfy the boundary conditions in (2.12). It is straightforward to verify that if 'T1
is a minimizer and A is its associated Lagrange multiplier, then (2.12) holds;
indeed,

£'('T1) = M'T1, F'('T1) = - ~p'(y-'T1)' (2.18)

where a prime denotes the functional derivative in 'T1. Thus, at least in a formal
sense, the variational principle (2.17) furnishes a one-parameter family of
V
solitary-wave solutions 'T1 = 'T1(x - ct, y) with c = gh / A [for a fixed stable
density stratification p( Y)] parametrized by the constraint value A which
determines the amplitude of the wave. A technically precise statement of the
variational principle is given in Section 3.
The physical interpretation of the variational principle (2.17) is quite appeal-
ing. In view of (2.9) it is evident that c 2E( 'T1) equals the kinetic energy of the
wave disturbance with respect to a frame at rest at infinity. From the expres-
sions in (2.16) it is clear that ghF( 'T1) equals the potential energy of the wave
Solitary Internal Waves 101

disturbance; indeed, noticing that f( y, 11) is a line integral over the segment
[y -11, y] (for 11> 0, say), it can be seen that ghF(l1) is precisely the work done
against the restoring force of buoyancy in producing the displacement 11. The
dimensionless group ,,\ = gh / c 2 can be interpreted as the inverse square of an
internal Froude number.
In the appendix, we discuss the relationship of our variational principle to
that introduced in a slightly different form by Benjamin [6]. He shows how his
energy-impulse principle is derived naturally from the (noncanonicaI) Hamilto-
nian structure of the (p, (T) system (2.3) along with its symmetries and associated
conserved quantities. He therefore advocates his form of the variational charac-
terization of solitary waves as the basis for a unified theory of existence and
nonlinear stability. However, it appears to be necessary to reduce his principle
to the more standard and tractable form given in (2.17) before an effective
computational method of solution can be designed.
Yet another variational characterization of steady waves is available in the
work of Bona, Bose, and Turner [10]. Their approach appeals to Yih's [33]
version of the semilinear elliptic equation for the pseudostreamfunction, which
is equivalent to the Dubreil lacotin-Long equation. Certainly, our numerical
method of Section 3 could be applied directly to their form of the variational
problem. The nonlinear (Liouville-type) transformation to standard form that
we utilize in Section 3 is essentially equivalent to the Yih's transformation.
Nevertheless, we prefer to adopt the variational principle (2.17) as the basis for
our investigations for the following reasons: (1) it maintains a connection with
the dynamically fundamental energy-impulse principle; (2) it permits a direct
physical interpretation of the objective and constraint functionals, being ex-
pressed in terms of the vertical displacement 11; (3) it coincides with the
standard form when simplified according to the Boussinesq approximation,
making the transformation in Section 3 unnecessary.

3. Semilinear form of eigenvalue problem


The variational principle formulated in Section 2 can be transformed to a useful
standard form by an explicit change of unknown. This transformation, which
may be considered the first step in the numerical solution of (2.17), is defined as
follows. Let 11 be replaced by the new unknown

cfJ := s( y) - s( Y -11) with s( y) := rYo


p( z) 1/2 dz. (3.1)

The quasilinear elliptic eigenvalue problem (2.12) in 11 = Y - S-I(S(y)- cfJ) is


then transformed into the following semilinear elliptic eigenvalue problem in cfJ:

-llcfJ + s"(y) - S"(S-I(S(y)-cfJ»)


-2"\
= -h-[Y-S-I(S(Y)-cfJ)]S"(S-I(S(Y)-cfJ») in D,

cfJ = 0 on aD, cfJ ~ 0 as Ixl ~ 00. (3.2)


102 B. Turkington, A. Eydeland, and S. Wang

Moreover, the variational principle (2.17) for (2.12) transforms correspondingly,


and the objective and constraint functionals become

E(cfJ) = f. [~IVcfJI2+e(Y,cfJ)]dxdy, (3.3)


D

F(cfJ) = f.f(y,cp)dxdy, (3.4)


D

where

e(y,cfJ):= ["[s"(y)-s"(s-l(y)-x)]dx
o

= s"(y)cp-Hp(y)_p(S-I(S(y)_cp»], (3.5)

f(y,cfJ) := h-2 f"'( y - S-l( s(y) - X)]S"(S-l( s( y) - X» dX; (3.6)


o

the eigenvalue (multiplier) A transforms identically.


The verification of (3.2)-(3.4), although straightforward, is rather lengthy;
therefore, only the key identities needed to make this transformation will be
documented here. The equation (3.2) follows from the calculations

MTJ = p(Y_TJ)1/2{ -llcfJ+s"(y)-s"(Y-TJ)},

TJP'(Y - "1) = 2p(y - TJ)1/2TJS"(Y - "1).

The expression (3.3), (3.5) is checked by using the identity

f.DilVcfJI2 dxdy = f. {iIV7j12p( y - "1) + H p( y) - p( Y - "1)]


D

+ (1- TJ y ) [p( Y - "1) - p( y) 1/2p( Y - 7j )1/21} dxdy,

and then noticing that the third term on the right-hand side of this identity can
be integrated by parts in y for each fixed x:

tCY -7j)y[p(y -7j) - s'(y)s'(y -7j)] dy


o

= t[p(y) + s"(y)s(y -7j)] dy - s'(h)s(h) + s'(O)s(O)


o

= tS"(y)[s(y-TJ)-s(y)]dy.
o
Solitary Internal Waves 103

The expression (3.4), (3.6) is derived from an alternative form of the definition
(2.16), namely,

f(y,T/) =
-11'o7gj5'(y-Odg;
h

the related formula for f( y, cp) is immediate from the change of variable
X = s(y)- s(y - g).
The transforming function s( y) is strictly increasing, is concave, and belongs
to C 3[0, h], recalling the assumption (2.5). In order to make the above expres-
sions for the functional integrands e(y,cp) and f(y,cp) valid for arbitrary real
values of cp it is convenient to extend the definition of s(y) to y E R. This

°
extension is accomplished most easily by considering the given density stratifica-
tion to be the restriction on .:s; y .:s; h of a smooth function j5( y). The extension
15 E C 2(R) is assumed to be monotone and to tend (rapidly enough) to constant
values at infinity; specifically,

p(y) > 0, j5'(y).:s; ° (y E R),

lim j5( y) = p ±' lim {lj5'(y)I+lyllj5"(y)l} = 0, (3.7)


Y ~ ±oo y ---+ too

where °
satisfies
< p + < p _ < + 00. It is obvious that then the extension s E C 3(R)

s'(y) > 0, s"(y).:s; ° (y E R),

lim s'(y) = pl/2, lim {ls"(y)I+lylls"'(y)i} = 0.


y-+±oo - y~±oo

With these extensions in force, the formulas (3.5) and (3.6) define the functional
integrands e and f for all cp E R. The technical properties satisfied by e and f
may be summarized as follows:

eq,,fq, E C 1([0,h]XR),

e(y,O) = eq,(Y,O) = 0, f( y,O) = fq,( y,O) = °


c/>
lim
-> ±oo
Ieq,q,(y,cp) I = 0,
c/>
lim
-> ±oo
Ifq,q,(y,cp) I = 0. (3.8)

It is also useful to record the expressions

2
p"(y) - j5'(y)
eq,q,(y,O) = 2j5(y) (2j5(y) ] ,
(3.9)
15'( y)
fq,q,(y,O) = - hj5(y)'
104 B. Turkington, A. Eydeland, and S. Wang

Let finite bounds a and b be defined by

a = max [ e.pq,(Y,</» L, b = max[fq,q,(y,</»L, (3.10)

where the maximum is taken over [0, h] X R. Throughout the sequel, we shall
assume that p(y) is given such that a < (7T / h)2. This hypothesis ensures that
the functional E(</» is strictly convex; in fact, the second variation of E(</» then
admits a positive lower bound:

8 2E(</» = j {IV8</>1 2 +e.pq,(y,</»(8</»2}thdy


D

~ f+oo th jh {[( 8</»y]2 _ a( 8</> )2} dy


-00 0

~ [(~r-a]fD(8</»2thdY. (3.11)

On the other hand, the functional F(</» is not necessarily convex, although it is
convex when restricted to </> of small norm, by virtue of (3.9).
The variational principle for the semilinear elliptic eigenvalue problem in
(</>,A)E HJ(D)XR is again

E( </» ~ min subject to F( </» = A, (3.12)

where the objective and constraint functionals are understood to be extended as


above; the Euler-Lagrange equation (3.2) can be written as

- b.</> + eq,( y,</» = Afq,( y,</» in D. (3.13)

The multiplier A is necessarily positive, and hence defines a meaningful wave


speed c according to A = gh / c 2 , because of the identity

j {IV</>12 + </>eq,(y,</»}thdy = Aj</>fq,(y,</»thdy,


D D

in which the left-hand side is positive by the hypothesis that a < (7T / h)2, and
the integrand </>f / y, </» is nonnegative by the properties of f. It is noteworthy
that the above variational formulation in </>, like the equivalent formulation in
11, applies to both waves of elevation (</> > 0) and waves of depression (</> < 0),
and that the construction of the objective and constraint functionals is made
without reference to this distinction. Moreover, the constraint value A is
positive in either case, since f( y, </» is nonnegative for </> of either sign. In
physical terms, this fact is transparent from the interpretation of ghF(cf» as the
(disturbance) potential energy associated with the restoring force of buoyancy;
Solitary Internal Waves 105

indeed, for small disturbances it is approximated by the positive (semi)definite


quadratic form

ghF( 4» = - g
2
f [~(p( y)y) ]4>2 dxdy + 0(114)11 2)
D
as 114>11 ~ o.

The assumptions made in Section 2 to derive the equation (2.12) can now be
weakened in two different ways. First, the hypothesis that p'(y) < 0 for 0 ~ y ~ h
can be replaced by p'(y) ~ 0 [as in (3.7)], permitting layers of homogeneous
fluid in the given stratification. Second, the restriction (on wave amplitude) that
each isopycnal surface connects to x = + 00 and x = - 00, or equivalently that
there are no closed eddies in the flow, can be relaxed. The equations (2.9) and
(2.11), now determined with respect to the extension of p(y) satisfying (3.7), are
seen to be sufficient for the exact equations (2.4) governing a steady transla-
tional wave, although they are not always necessary.
When p( y) is constant in a layer (or some other region), then the ver-
tical displacement function 'T1 is not well defined by the relation p(x, y) =
p(y - 'T1(x, y»; nevertheless, the equation (2.9) may be enforced in such a layer,
thus extending the definition of 71. [It is interesting to note that the extension of
71 is harmonic in such a layer of constant density, and hence it is determined
within the layer by its values on the boundary of the layer, where the vertical
displacement is known. This observation allows us to derive a variational
principle for a two-layer fluid model, say, which can be expressed in terms of
the vertical displacement 'T1 = 71(X) of the interface.]
When the wave amplitude is large enough that there exist points where
'T1(x, y) > y ['T1(x, y) < Y - h] in the case of a wave of elevation [depression], the
relation p(x, y) = p(y - 71(X, y» involves values of y - 71 that lie outside of the
physical range [0, h]. However, the governing equations formally allow solutions
for which the density p is extended in an essentially arbitrary (but suitably
smooth) manner. The extension used here is made a priori by choosing a p( y),
Y E R, according to (3.7). Additional physical considerations are required, of
course, to justify the particular choice of such an extension. Nevertheless, our
approach does allow us to produce valid solutions having closed eddies (regions
which are bounded by a streamline separatrix and within which the streamlines
are closed curves), and permits us to prescribe the density dependence within
the eddy (which usually is taken to be very nearly constant). On the other hand,
we may verify a posteriori that the solution has no such closed eddy, thus
fulfilling our main aim to produce solitary waves which disturb the given density
stratification in a uniquely determined fashion. In that case the extensions
merely eliminate the need to impose any auxiliary conditions (0 ~ y - 71 ~ h) on
the magnitude of the disturbance 'T1 in the course of the analysis or computa-
tion.
The restriction

a < (7T'/h)2 [recall (3.10)] (3.14)


106 B. Turkington, A. Eydeland, and S. Wang

ensures that the functional E(t/J) is strictly convex. Its physical significance can
be appreciated by writing it in terms of p(y):

max{~p+lh2p"(y)I+lhp'(y)12} < 7T"2p~,


yER

where p + = min p(y). Roughly speaking, this means that the (total) density
change p _ - p + is restricted in terms of the pycnocline thickness. In virtually all
cases of interest involving internal waves in the oceans or atmosphere, the
dimensionless derivatives of density, hp'(Y) and h 2p"(y), are very small com-
pared to the nearly constant density, p(y) """ P +. Consequently, the above
restriction is a very mild one when considering actual geophysical phenomena.
(For this reason, no attempt will be made here to find even weaker hypotheses,
although this is possible in certain cases. The restriction does preclude, for
instance, the study of rapidly stratified fluids with large density changes such as
might approximate the situation of a fluid composed of two layers of different
density.) The classical Boussinesq approximation, which simply neglects any
variation in density apart from that directly involved in the buoyancy effect, is
precisely equivalent to dropping the term e( y, t/J) from our variational problem.
The traditional justification for setting e = 0 may be stated in our notation as
the fact that h 2a «7T"2 for most stratifications of real interest. This widely used
approximation has the advantage that it simplifies the analysis considerably;
for instance, it makes the transformation (3.1) unnecessary, since it replaces
p( y - T/) by p + (say) in the objective functional E( T/), thus putting the func-
tional into standard form. (The transformation degenerates to t/J = plj 2 T/.)
However, some researchers [4, 24] have concluded that the Boussinesq approxi-
mation can be delicate and can lead to significant inaccuracy when finite-ampli-
tude solutions, especially solitary waves, are considered. Therefore, we shall
proceed with our analysis and computation without this simplification, recogniz-
ing that the Boussinesq case may be included in our exact formulation merely by
setting e = O.

4. Properties of solutions
As a preamble to the computational results of the subsequent sections, this
section summarizes some of the properties of solutions that can be demon-
strated analytically from the variational principle given in Section 3. Particular
attention is paid to necessary and/or sufficient conditions for the existence of
solitary waves of elevation or depression depending upon the given undisturbed
density profile p( y ).
Any solution t/J of the variational problem (3.12) may be assumed to possess
the symmetry properties

t/J(x,y) = t/J( - x,y) (4.1a)


Solitary Internal Waves 107

either (+) 4> z 0, 4>x s 0 or (-) 4> s O,4>x z 0 for all x z o.


(4.1b)

This is a consequence of x-translation invariance of the domain and the


functionals E and F. The proof of (4.1) relies on the fact that an admissible
function 4> can be replaced by its x-symmetrization 4>*, namely, that function
which satisfies (4.1) and meas{I4>1 > a} = meas{I4>*1 > a} for all a z O. A stan-
dard argument [10] then yields E(4)*) s E(4)) and F(4)*) = F(4)), as required.
(A more technical argument is needed to infer that 4> = 4>* whenever 4> is a
solution.)
Henceforth, when we refer to a solution of (3.12) or (3.13) we shall assume
that it is an x-symmetrized wave of elevation or depression in the sense of (4.1).
This requires us to exclude certain higher-mode solutions such as are con-
structed analytically in [30] and exhibited experimentally in [13]. It is not known
how to obtain such solutions variationally, except when they have a reflective
symmetry in y.
Various necessary conditions for the existence of (nontrivial) solitary waves,
which may also be viewed as properties of the solution pairs (4), A), can be
inferred from the variational characterization (3.12). The most informative of
these properties expresses a connection between the rate of exponential decay
in x of 4>(x, y) and the value of A (near a critical value Ao). This property can
be obtained by considering "stretching variations" of a solution 4>(x, y), namely,

4>T(X,y):= 4>((l+'T)x,y) for small iT I. ( 4.2)

The test function associated with these variations is

x4>Ax,y) = :'T4>T(x,Y)IT~O·
Then either (3.12) or (3.13) yields the identity

0= :'T{E(4)T)-AF(4>T)}IT~O

= 1. {V4>·V(x4>x)+e",(Y,4»x4>x- Af",(Y,4»x4>x}dxdy
D

= ID {4>; + x[ t1V4>1 2 + e(Y,4» - Af(Y,4»] Jdxdy


= 1. {t4>; - t4>; - e( y,4»
D
+ Af( y,4»} dxdy,

where the integration by parts in the last equality is justified using the fact that
108 B. Turkington, A. Eydeland, and S. Wang

since E( cfJ), F( cfJ) < + 00 there must be a sequence x n ~ + 00 such that


xnfx=x [tl VcfJl 2 + e - Andy ~ 0. From this calculation we obtain the basic iden-
tity n

E(cfJ) - AF(cfJ) = f cfJ;d.xdy,


D
( 4.3)

which we may now apply to derive an estimate on A and a corresponding decay


rate for cfJ.
An immediate consequence of (4.3) is the general estimate

A< E(~) _ E(~) ( 4.4)


- F(~) - ---:::t'

valid for any admissible function ~ in (3.12). A routine analysis of infinitesimal


sinusoidal waves governed by the linearization of (3.13) about cfJ = suggests the
choice
°
~k(X,y) = {'Yk(COSkx)<I>k(Y)' Ikxl :::; 7T'/2,
( 4.5)
0, Ikxl ~ 7T'/2 '

with wavenumber k, where <l>k is the eigenfunction corresponding to the first


eigenvalue Ak of the Sturm-Liouville problem

d <1>k [2
k +e",,,,(y,O) ]<l>k
2
--2-+ = Ad.p.p(y,O)<I>k (O<y<h),
dy

<l>k(O) = ° = <l>k(h). (4.6)

The function ~k is truncated in x so that ~k E HJ(D), and the amplitude factor


'Yk adjusted so that F(~k) = A. In the long-wave limit as k ~ 0, the first
eigenpair (<I>k,A k ) tends to (<I>o,Ao) defined by setting k=O in (4.6). Since
'Yk = O(k- 1/ 2), the limiting form of (4.4) is calculable as the familiar Rayleigh
quotient:

- h[
A :::; liminf E ( ~k) = fo (d<l>oldy)
2
+ e",,,,(y,O)<I>&] dy
k->O F(cfJ k ) ftf",,,,(y,O)<I>&dy = Ao· (4.7)

Written in terms of the phase speed Ck of linear waves, for which A k = gh 1 C1,
this result expresses the important fact that the solitary-wave speed c is
necessarily supercritical:

c ~ Co = supCk . ( 4.8)
k>O
Solitary Internal Waves 109

A sharp form of the estimate (4.7) can be obtained for weakly nonlinear long
waves by choosing suitable approximate solutions as admissible functions 4">. We
shall illustrate this calculation in the Korteweg-de Vries regime only. In this
case, the approximate solution (cPo, AE) of (3.13) is found in the form [7]

cPE(x,y) = 8 2P(8X)<PO(Y) + 0(8 4),


( 4.9)
AE = Ao - 8
2
+ 0(8 4),

where e is a small positive parameter measuring the ratio between the half-
amplitude wavelength I and the (total) fluid depth h; precisely, this asymptotic
regime has characterized the scaling h / 1= O(e) as e - O. The order-e 2 coeffi-
cient in the expansion of AO in (4.9) has been normalized to -l.
Upon substitution of this expansion into (3.13), the resulting order-8 2 equa-
tion is solved by the linear eigenvalue problem (4.6) with k = 0, while the
order-e 4 equation is solvable whenever the amplitude profile function P(X)
(X = ex) satisfies the steady Korteweg-de Vries equation

d 2P
- ] dX
2 + P - IP2 = 0 (-00< X < +(0) ( 4.10)

and P(X) - 0 as IX 1- 00; the coefficients resulting from the calculation of the
solvability condition are

lo <Po dy, 1=31jh [f",,,,,,,(y,0)-A oe,,,,,,,,,(y,0)]<P6 dy , (4.11)


h 2
] =
o

and the eigenfunction in (4.6) is normalized by

jhf",,,,(y,0)<P6dy = 1.
o

The well-known amplitude profile for the solitary wave is then

3 2~ ( 4.12)
P( X) = 2J sech 2/ 1/ 2

provided that 1* 0, and thus the first-order asymptotic solution is completely


determined.
The constraint value in (3.12) is related to e by the expansion

A = F(cP E
) = 3]1/2r 283 +O(e 5), ( 4.13)

which results from a straightforward calculation. In fact, the following expan-


110 B. Turkington, A. Eydeland, and S. Wang

sions hold for small e:

2E( q,t:) = Aoe 3 1 +00

-00
P 2dX + e 5 1 <1>5 dy 1
h

0-00
+00
pI dX

+ e 51'3 Jc(h e1>1>1>(y,O)<1>ody


3 1+ p 00
3
dX + 0(e
5
),
o -00

2F(q,t:) = e31+oop2dX
-00

51 (hi
+e'3Jc ( ) 3
1>1>1>y,O<1>ody 1 +00 3 5
PdX+o(e).
o -00

These expressions, combined with the identity

11
-00
+00
pI dX - 1 +00

-00
p2 dX + J'3
1
1+00

-00
p 3dX = 0,

yield the crucial asymptotic formula

E(q,t:) _ A _ e 2 + 0(e2). ( 4.14)


F(q,t:) - 0

Now, returning to (4.4) with the choice J; = q,t:, we find that for all sufficiently
small e

A ~ Ao - fJe 2 , ( 4.15)

where (} (say) is fixed with 0< fJ < 1. This sharp form of the universal estimate
(4.7) holds under the restriction that the constraint value A = A(e) must be
small. Similar results, however, can be obtained without this restriction by
modifying the above method; for instance, a general estimate of this kind is
established in [10] under some stronger hypotheses on e and f.
The exponential decay of a solution q,(x, y) in x follows directly from an
estimate of the form (4.15) for A. This property is conveniently derived by the
method used in [2], in which a differential inequality is found for the function

Q(s):= f+oofohq,2dxdY (o<s<+oo). ( 4.16)


Solitary Internal Waves 111

By virtue of the governing equation (3.13), an integration by parts yields

Q"(S) = -2f\p(s,Y)cf>As,y)dy
o

= 2f + f 00 h{ IVcf>1 2
+e",(y,cf»cf>-Af",(y,cf»cf> } dxdy
s 0

= 2f
+00
f
h{ IVcf> I2+e",,,,(Y,cf»cf>2- Af,,,,,,(Y,cf»cf>2}dxdy
s 0

+ R(s), ( 4.17)

where the last equality defines a remainder integral R(s) whose integrand is
O(Icf>1 3 ). The Rayleigh principle associated with (4.6) (k = 0) implies that for
each x

Jc(h{ cf>y2+ e",,,,(y,O)cf> 2} dy ~ Ao fh f",,,,(y,O)cf> 2dy,


o 0

which upon substitution into (4.17) gives

Q"(s) ~ 2 (I - A f
A) +00 h
f {cf>;+e",,,,(y,0)cf>2}dxdy + R(s)

r-a]
o s 0

~ 2(1- ;J [( ~ Q( s) + R( s).
Thus, recalling the assumption (3.14), we infer the differential inequality

-Q"(S)+(YO/h)2(A o -A)Q(s):::;; R(s) (0 < s < +(0) (4.18)

for some posItIve constant Yo. Since Q(s),Q'(s)~O and R(s)/Q(s)~O as


s ~ + 00, as is easily checked using (4.1), a routine analysis leads to the decay
rate

Q(s):::;; A1exp ( _ YI(Ao~A)I/2S) ( 4.19)

with positive constants AI < +00, Yl < Yo.


The pointwise bound

cf>(x,y) :::;; Bexp ( _ Y2(Ao~A)I/2IXI), ( 4.20)


112 B. Turkington, A. Eydeland, and S. Wang

with some positive constants Band "Y2' and an analogous bound for IV<t>(x, y)1
can be deduced from (4.19). These results follow easily from standard estimates
in the theory of elliptic partial differential equations [17] applied to (3.13).
Indeed, elliptic regularity theory guarantees that the solution <t> of (3.13)
belongs to C 2 ,u(i5) (0 < ex < 1), and moreover furnishes estimates for
sup{I<t>(x,y)l+hIVcf>(x,y)}l:s+h<x<+oo, O<y<h} in terms of Q(s). The
technical details involved in such an analysis will not be included here, as they
are discussed fully in [10]. [The constants A I and B depend upon the amplitude
of the solution, whereas the "Yi are absolute constants. For example, in the
asymptotic limit for which the constraint value A = A(e) = 0(6 3 ), it can be seen
easily that Al = 0(6 3 ), and with a careful estimation it can be shown that
B = 0(6 2 ).]
Sufficient conditions for the existence of (nontrivial) solitary-wave solutions
can be fashioned on the basis of the foregoing analysis. Indeed, such a
variational existence theory is the main goal of the work of Bona, Bose, and
Turner [10]. Here, we shall be content to sketch the outline of the existence
problem for (3.12). The principal difficulty arising in the problem, of course, is
the lack of compactness properties in an appropriate function space due to the
unboundedness of the domain D. When invoking the direct variational method
for (3.12), this difficulty is overcome by first constructing 2/-periodic (in x)
waves for 1< + 00, and then letting I tend to infinity. Solitary waves are
therefore identified as the limits of long periodic (cnoidaO waves provided that
suitable a priori estimates independent of the half wavelength I can be estab-
lished.
The crucial a priori estimate on periodic waves needed to ensure that a
nontrivial solitary wave is obtained as I ~ + 00 is (4.15)-namely, strict supercrit-
icality of the wave speed. If it is known that Ao - A ~ d for some 6 A > 0
depending on A but independent of I, then the preceding analysis of the
exponential decay of solutions may be invoked for 2/-periodic solutions with
large I. The estimates supplied by elliptic regularity theory combined with (4.19)
then suffice to prove that the limit of the 2/-periodic solutions is a nontrivial
solution of (3.12). (Without such estimates it is possible that the 2/-periodic
solutions tend to zero uniformly in D for a fixed positive constraint value A.)
Setting aside any further technical discussion, we now give some conditions
on e and f [in addition to the structure conditions (3.8)] that enforce the basic
inequality (4.4) for periodic solutions (<t>, A). A straightforward adaptation of the
argument leading to (4.15) then completes the proof of existence. The deriva-
tion of (4.4), however, must be substantially modified in the periodic case
because boundary integrals at x = ± I arise in the formula (4.3), and these extra
terms cannot be treated without some further information. A simple invariant of
(4.3) obtained by taking the usual variations <t>"'(x, y) = (l + T )<t>(x, y) rather
than those in (4.2) is valid in the periodic case:

E(<t» - AF(<t» = if {A[<t>fq,(y,<t»- f(y,<t»]


D

- [<t>eq,(y,<t» - e( y,<t»]} dxdy. ( 4.21)


Solitary Internal Waves 113

With this in hand, the existence of solitary waves can be inferred from sufficient
conditions (imposed on e and f) that guarantee (i) the nonnegativity of the
*"
right-hand side of (4.21) and (ii) J 0 in (4.11); indeed, (i) validates (4.4), and
(ii) then guarantees (4.15). As in the weakly nonlinear theory, the sign of J
distinguishes between waves of elevation (J> 0) and waves of depression
(J < 0). However, the resulting conditions are rather technical and in general
too restrictive for many practical purposes. Therefore, the constructive method
of solution described in the sequel must be invoked to settle the existence
question except when very special density profiles are specified.

5. Iterative algorithm
The main goal of this paper is to introduce a numerical method of solving the
quasilinear elliptic eigenvalue problem (2.12) for solitary waves TJ = TJ( x - ct, y).
In this section we introduce a globally convergent iterative algorithm which
solves the equivalent semilinear problem (3.13) in (¢>, A). Our method takes full
advantage of the associated variational structure (3.12), and in fact it shares
many attributes with a class of algorithms (linearization methods [27] and
sequential quadratic programming [18]) encountered in the constrained-optimi-
zation literature. However, the specific form of our algorithm appears to be
new, especially as it is tailored to the infinite-dimensional problem (3.12)
involving nonlinear partial differential equations.
Throughout this section we shall suppress any discussion of the technical
complications arising from the unboundedness of the domain. As explained in
Section 4, it suffices to compute long periodic waves that accurately approxi-
mate a solitary wave having a known exponential rate of decay in x. Therefore,
we may proceed with the analysis of the algorithm imagining that D is replaced
by D I := {J X I < I < + oo} n D, but we shall continue to write D for the finite
domain for the sake of simplicity in notation.
Let ( " . ) denote the usual dual pairing in the sense of distribution theory, so
that

(u, v) = f uvdxdy
D

whenever (say) U,V E L 2 (D) or u E HJ(D), v E H-1(D). The governing equa-


tion (3.13) interpreted in the weak sense,

f {V¢>'V~+e4>(y,¢»~}dxdy = Aff4>(y,¢»~dxdy for all ~ E HJ( D),


D D

is expressible (in variational form) as

E'(¢» = A£1(¢», (5.1)


114 B. Turkington, A. Eydeland, and S. Wang

where the functional derivatives are

E'(cfJ) = - acfJ + e",(y,cfJ) E H-1(D),


(5.2)
P(cfJ) = f",(y,cfJ) E L 2(D).

The objective functional E is strictly convex by hypothesis, recalling (3.11)


and the hypothesis (3.14). The constraint functional F, on the other hand, is not
necessarily convex, although it is convex in a neighborhood of the origin (which
in fact is quite large in most cases of real interest). A suitable choice of {3 > 0
ensures that

F + {3 E is a convex functional on H J( D) , (5.3)

where (3 depends upon a and b [defined in (3.10)]. The definition of a ~ 0 by

a = max [e",,,,( y,cfJ) L (5.4)

permits the splitting

E = E+ - E_ (5.5)

with

E+(cfJ):= J[iIVcfJI2+%cfJ 2]dxdy,

E_(cfJ):= J[%cfJ2-e(y,cfJ)]dxdy,

in which E + is strictly convex and quadratic, and E _ is convex.


With the above notation in place, the iterative algorithm may be described in
an abstract form. Let cfJ° E HJ(D) be a given initialization satisfying F(cfJ°) = A,
but otherwise arbitrary. The iterative step cfJk ~ cfJk + 1 (k denoting the iteration
number) is defined by solving the quadratic programming subproblem [in cfJ E
HJ(D)]

E+ (cfJ) - E_( cfJk) - \E'- (cfJk),cfJ - cfJk) ~ min

over F( cfJk) + \F'( cfJk) + (3E'( cfJk),cfJ - cfJk) ~ A. (5.6)

Evidently, the iterate cfJ = cfJk + 1 is itself characterized as the solution of a


variational subproblem derived from the variational problem (3.12) by partially
linearizing the objective functional and replacing the nonlinear equality con-
straint with a convexified linear inequality constraint. This abstract characteriza-
Solitary Internal Waves 115

tion defines cfJk+ 1 uniquely, owing to the strict positive-definiteness of E +.


Moreover, the solution cfJk + 1 and its associated multiplier p.,k+ 1 are explicitly
characterized by the corresponding Kuhn-Tucker conditions, the analogue of
the Lagrange-multiplier rule, which are

E~ (cfJk+l) - E'- (cfJk) = p.,k+I[F'( cfJk) + f3E'( cfJk)]' (5.7)

F( cfJk) + (F'( cfJk) + f3 E '( cfJk),cfJ k+1 - cfJk) ~ A, p.,k+1 ~ 0, (5.8)

p.,k+I{F(cfJk)+(F'(cfJk)+f3E'(cfJk),cfJk+l_cfJk)-A} = O. (5.9)

The reader is referred to the standard literature in convex analysis [19, 28] for
an explanation of the stated equivalence.
Assuming for the moment that the above iterative sequence (cfJk, p.,k) is
constructed and converges in an appropriate sense to a limit (cfJ *,p., *), the
manner in which a solution of the problem (3.12), (3.13) is obtained can be seen
as follows. In the limit (k ~oo) the Kuhn-Tucker conditions become

E'( cfJ*) = p.,* [F'( cfJ*) + f3 E '( cfJ*)]' (5.10)

F(cfJ*) ~ A, p.,* ~ 0, (5.11)

p.,*{F(cfJ*)-A} = O. (5.12)

Consequently, cfJ* is a solution of (5.1) with an associated multiplier

p.,*
A* := 1- f3p.,* . (5.13)

Below we show that 0 < p., * < f3 - I must hold for any such solution. Thus, A* is
positive and finite, and the complementary condition (5.12) yields the equality
constraint (3.12). Hence, a critical point cfJ* for the governing variational
problem is found.
The convergence of the iterative sequence rests on the following basic
properties:

E(cfJk+I):::; E(cfJk), (5.14)

F( cfJk) ~ A. (5.15)

An induction argument in k establishes both of these properties jointly. The


assumption F(cfJ°) = A initializes the argument. If F(cfJk) ~ A, then cfJ = cfJk
competes in the minimization (5.6), and so

E(cfJk+I):::; E+(cfJk+l)- E_(cfJk)_(E'-(cfJk),cfJk+l_cfJk):::; E(cfJk), (5.16)


116 B. Turkington, A. Eydeland, and S. Wang

where the convexity of E _ is used in the second inequality. This proves (5.14).
Now, the convexity of F + f3E implies that

F(c{Jk+l) + f3 E (c{Jk+l) ~ F(c{Jk) + f3 E (c{Jk)+(F'(c{Jk)+f3 E '(c{Jk),c{Jk+l_c{Jk)


~ A + f3E( c{Jk). (5.17)

The latter inequality combined with (5.14) produces the inequality F(c{Jk+ 1) ~ A,
which completes the inductive step. Thus, (5.14) and (5.15) are proved for all k.
The following strengthened version of (5.14) is needed to infer the conver-
gence properties of the algorithm:

E+(c{Jk+l_c{Jk) ~ E(c{Jk)- E(c{Jk+l). (5.18)

This claim is verified by retaining the second-order terms involving E + in the


first inequality of (5.16), and then substituting the defining equations (5.7) and
(5.9) into the resulting expression. The required calculation is

E( c{Jk) - E( c{Jk+l) ~ (E'+ (c{Jk+ 1) - E'_ (c{Jk), c{Jk - c{Jk+l) + E + (c{Jk+l - c{Jk)

= J-L k +1(F'(c{Jk)+f3 E '(c{Jk),c{Jk_c{Jk+l) + E+(c{Jk+l_c{Jk)

= J-L k +1{F(c{Jk)-A} + E+(c{Jk+l_c{Jk).

The claimed result then follows by virtue of (5.15).


The strong monotonicity property (5.18) immediately implies that [since
E(c{JkH E* ~ 0]

c{Jk+l - c{Jk ~ 0 strongly in HJ( D). (5.19)

Without a rate for this convergence it cannot be asserted that the (entire)
sequence {c{Jk} converges to a (single) limit c{J*. Nevertheless, the bound E(c{Jk)
~ E(c{J°) ensures at least that every subsequence has a further subsequence
converging weakly in HJ(D). Now, if c{J* denotes any such weak limit point of
{c{Jk}, then (5.19) justifies passing to the limit in (5.7)-(5.9) to obtain (5.10)-(5.12).
A standard argument may then be invoked to show that the corresponding
subsequence of {J-Lk} also converges to J-L* (say). Furthermore, (5.19) enforces
the following alternative valid for any given admissible initialization c{J0: either
{c{Jk} converges to a single limit point or it has an infinite set of limit points,
none of which is isolated. (The straightforward proof is omitted here.) In
general, the lack of a uniqueness theorem for the governing eigenvalue problem
means that the first alternative cannot be guaranteed a priori. This kind of
situation is typical for convergence results that are global in the sense that the
initialization need not be taken close enough to a solution. (By appealing to
elliptic regularity theory the L2-norm convergence stated above can be strength-
ened to C 1-norm convergence.)
An analysis of the rate of local ~onvergence of the iterative sequence to a
(non degenerate) solution will not l_ attempted in the present context. Suffice it
Solitary Internal Waves 117

here to say that the implemented algorithm is observed to converge with a linear
rate to isolated solutions in all of our numerical experiments. The rate of
convergence depends, of course, on the parameters of the problem. For in-
stance, large-amplitude solutions with relatively small half-amplitude wave-
lengths (2 ::;:; 1/ h ::;:; 10) require fewer iterations than smaller-amplitude solutions
with large half-amplitude wavelengths (10::;:; 1/ h ::;:; 50). Similarly, the density
stratification p(y) will affect the convergence rate, slowing it somewhat when
density gradients are increased.

6. Numerical implementation
The iterative algorithm introduced in Section 5 in its abstract form can be easily
translated into an explicit numerical method. Deferring for the moment the
discretization of the domain D (or rather a finite D t ), the explicit construction
of the iterative step </>k ~ (</>k+I,JLk+l) is reducible to the following three parts
in the general case.
(I) Solve the two linear elliptic boundary-value problems (in Uk and w k )

(-.:l + a)u k = a</>k - e",(y,</>k) in D, Uk = 0 on aD, (6.1)

(- A + a)w k = f",(y,</>k) in D, w k = 0 on aD. (6.2)

(II) Evaluate the four integrals

SI = f f",(y,</>k)W k dxdy,
D

S2 = ff",(y,</>k)(</>k_Uk)dxdy,
D

2
S3 = fJIV(</>k_ uk)1 +a(</>k- uk)2]dxdY ,

F( </>k) = f f(y,</>k) dxdy.


D

(III) Define

JLk+1 = max{o, A - F( </>k) + S2 + {3S3 } (6.3)


Sl +2{3S2 + {3 2S3 '

</>k+l = Uk+JLk+l[Wk+{3(</>k_Uk)]. (6.4)

As this construction appears to be somewhat complicated, we shall now


explain how it is equivalent to the Kuhn-Tucker conditions (5.7)-(5.9) for the
118 B. Turkington, A. Eydeland, and S. Wang

defining subproblem (5.6). First, the equation (5.7) is clearly identical to (6.4)
together with (6.1), (6.2); consequently, 4>k+1 is expressed in terms of J,Lk+l.
Then, the complementarity conditions (5.8), (5.9) that determine J,Lk + I can be
written in terms of JLk+ I alone by substituting into them the expression (6.4) for
4>k+I; this results in the alternative: either J,Lk+1 = 0, or J,Lk+1 > 0 and solves

A = F( 4>k) +( f",(Y,4>k) + f3( -/1 + a){ 4>k - Uk),U


k
- 4>k

+ J,Lk+l[ w k + f3( 4>k - Uk)]) .

A straightforward manipulation with this linear equation in J,Lk+ 1 leads to the


stated formula (6.3). [In the event that solving the latter equation produces
JLk+ 1 < 0, the algorithm sets J,Lk+ I = 0 and relaxes the constraint (5.6) to an
inequality; in practice, this event rarely occurs.]
The above algorithm may be used to seek either waves of elevation or waves
of depression in the sense of (4.1)( + ) or (4.1)( - ), respectively. Indeed, in either
case, if an initialization 4>0 satisfying (4.1) is chosen, then every iterate 4>k will
also satisfy (4.1). For instance, in the elevation case the maximum principle
applied to (6.1) and (6.2) implies that Uk, w k ~ 0, and further (6.4) implies that
4>k+ 1 ~ 0, given that 4>k ~ O. (All of the integrals Si are positive in either case.)
This property of the algorithm is especially helpful when computing branches of
solutions by varying the given parameters defining p( y). Then, choosing the
current solution as initialization for the computation of the solution correspond-
ing to an incremental change in the given parameters, the type (elevation or
depression) of the solution may be maintained along the branch.
The discretization of the algorithm may be accomplished by any standard
approach. In our implementations an FFf method is utilized to solve the
Poisson equations in (I), taking advantage of the fact that they are constant-
coefficient equations in a rectangular domain. As mentioned before, the infinite
strip domain D is truncated by assuming 21-periodicity in x or, alternatively, by
imposing homogeneous Dirichlet conditions at x = ± I. An adaptive choice of
(large enough) I ensures that adequate decay as x ~ ± I is achieved. The
integrals in (II) are computed with a sufficiently accurate quadrature rule
(Simpson's rule). In the non-Boussinesq case, numerical tabulations of sand
S-I [defined in (3.1)] are also required in (I) and (II), since e and fare
determined from p( y ) through these computed functions; this is easily achieved
by a single quadrature to evaluate s on a fine grid on 0:0;; y :0;; 1, and then an
interpolation to find its inverse s - I.
Finally, it is worthwhile to emphasize that a greatly simplified version of the
above algorithm suffices in many cases of real interest-namely, in situations
for which the density stratification is weak and the wave amplitude is small or
moderate. Then, assuming the Boussinesq approximation to be valid, the trans-
formation T] ~ 4> is just 4> = pb/ 2 T] [Po = p(Yo), say]; consequently, the governing
equation takes the form (5.1) with e = 0, meaning that a = 0 and a = O.
Provided also that wave amplitudes are not too large, it may be assumed that
fey, T]) is convex in T], and consequently that (5.3) holds with f3 = 0 within some
Solitary Internal Waves 119

range 17]1 :s; eh. The algorithm thus simplifies to the following:
(I) Solve - LlWk = i,/y, 4>k) in D, w k = 0 on aD.
(II) Evaluate SI' S21 F(4)k) with Uk = o.
(III) Define

Ak+ 1 = max{o, A - F( 4>k) + S2}


SI '
4>k+1 = Ak+1W k .

In this special case the numerical implementation of the algorithm is obviously


easier and cheaper than in the general case. Moreover, the convergence theory
admits a simpler presentation when a = 0 and b = O. In view of these attributes,
this special form of the algorithm is useful in practice as well as in concept,
being the essence of the general form.

7. Computed exampJes
The principal virtue of the iterative algorithm introduced in Sections 5 and 6 is
its computational robustness under a wide range of conditions. The purpose of
this section is to illustrate the application of the algorithm in a few cases
representative of the conditions encountered in practice. Needless to say, the
computed examples included here do not cover the range of applicability of the
algorithm and are not intended to model precisely physical conditions occurring
in the oceans or atmosphere. Nevertheless, our method is available to solve
these real problems and is relatively easy to implement when the given data
p(y) and A are known only from field measurements. The method can also be
extended to allow for an ambient shear flow, but the discussion of such an
extension is given elsewhere [32].
The computational results given below pertain to a class of prototype density
profiles defined by

p(y) = poexp[~tanh
2
2(YO-Y)].
Th' (7.1)

here Yo is a reference level and Po = p(Yo) is a corresponding reference density;


8 and T are dimensionless positive constants and may be interpreted as relative
measures of the density change and thickness, respectively, of the pycnocline.
The Brunt-Vaisiilii or buoyancy frequency defined by N 2 = - gp' /p is deter-
mined from (7.1) to be

N(y) =
g8 )1/2 sech 2( Yo - y
)
(7.2)
( -Th Th'
This class of stratifications contains as limits the familiar special cases:
(1) homogeneous fluid (T -HlO),
(2) two-layer fluid (T ~ 0),
(3) exponential density profile (8, T ~ 00 with finite y = 8/ T).
120 B. Turkington, A. Eydeland, and S. Wang

We naturally express our computed examples in the dimensionless variables


x' = x / h, y' = y / h, TJ' = TJ / h, X = A. Obviously, we can set h = 1 and drop the
primes, and we can normalize Po = 1. In each of the cases considered below a
spatial grid corresponding to 64 Fourier modes in both x and y is chosen, as
this resolution is enough to ensure sufficiently accurate results. In general the
best choice of grid size varies with the problem under consideration as well as
with the accuracy demanded; the number of modes in x depends (weakly) upon
the length 21 of the domain D, = {-I < x < + I, 0 < y < 1}, while the number of
modes in y depends (strongly) on the density profile through T.
In the following examples the stopping criterion

lIq,k+l - q,k11L2 ,:5; 0.00511q,k+lIIL2

is adopted to terminate the iterative algorithm. The constraint value F(q,k),


which satisfies F(q,k);;::: A at each iteration, is observed to converge rapidly to
its specified value A in all cases. On the other hand, the multiplier sequence Ak
*'
(derived from JLk if f3 0) converges more slowly than the corresponding
solution sequence q,k (or TJ k ), as is expected on the basis of elliptic regularity
estimates. The number of iterations needed to fulfill the above-stated stopping
criterion in q,k varies considerably among the examples described below,
besides depending upon the initialization q,0. In the first two cases less than 10
iterations suffice, whereas in the third case (which is quite singular) more than
30 iterations are needed. In each of these cases an initialization based on the
shallow-fluid (KdV) approximation is given, even though the solutions typically
lie far beyond range of validity of that asymptotic regime. The algorithm is
computationally effective, therefore, even with a rough initial guess. Neverthe-
less, the most efficient way to compute branches of solutions is to increment the
prescribed parameters A, 8, T, Yo using the current solution as an initialization.
The number of iterations needed at each incremental step can then be reduced.
The constants a and f3 must be fixed appropriately in the algorithm to
ensure its global convergence. For a given density profile these constants can be
determined a priori independently of A. However, from a practical point of
view it is often preferable to choose a and (especially) f3 adaptively, increasing
them whenever necessary. In our implementation a is fixed a priori while f3 is
taken to be zero in all cases except those involving very strongly nonlinear
waves. The algorithm with f3 = 0 is considerably simpler, and fortunately suf-
fices for most realistic cases.
We now separate for discussion three cases from the class of prototype
densities profiles (7.1), in each case displaying three members of a branch of
computed solitary waves.
First, we consider a branch of elevation waves in a relatively deep fluid with a
thick pycnocline. The fixed parameters are 1 = 4,8 = 0.01, T = 0.5, Yo = 0, while
A increases along the branch: (1) A = 1 X 10- 5, (2) A = 2.5 X 10- 5, (3) A =
5 X 10- 5 • The isopycnal curves (streamlines) for the solutions are displayed in
Figure 2; the corresponding multipliers are (1) A = 2.01 X 10 3 , (2) A= 1.74 X 10 3,
(3) A = 1.52 X 10 3 • The waves vary along this branch from weakly nonlinear (1) to
Solitary Internal Waves 121

(1 )
-------
f--------------

(2)

----
(3)

/~ "

Figure 2. Elevation waves in - 4 < x < 4, a < y < 1 for density profile parameters 8 = 0.01,
T= 0.05, Yo = O. The amplitude constraint varies: (1) A = 1 X 10- 5, (2) A = 2.5 X 10- 5, (3) 5 X 10- 5.
122 B. Turkington, A. Eydeland, and S. Wang

--------
(1 )

(2)
I ~~~-
----------~--

(3)
c-----~.

Figure 3. Depression waves in - 4 < x < 4, 0 < Y < 1 for density profile parameters 8 = 0.01,
T = 0.1 and amplitude constraint A = 1 X 10- 4 • The pycnocline level Yo varies: (1) Yo = 0.75,
(2) Yo = 0.80, (3) Yo = 0.85.
Solitary Internal Waves 123

,'~/
(1)
~' /'/
I--~~/~
I----~~/ ____- - - - j

(2)

~
~~' .. '.
, ---// /
~

(3)

Figure 4. Depression waves in - 10 < x < 10, 0 < y < 1 for density profile parameters 0 = 0.01,
T = 0.2 and amplitude constraint A = 1 X 10- 3• The pycnocline level Yo varies: 0) Yo = 0.68, (2)
Yo = 0.65, (3) Yo = 0.62.
124 B. Turkington, A. Eydeland, and S. Wang

strongly nonlinear (3). The large-amplitude solution (3) does not contain a
closed eddy, but if A is increased slightly more, then a small closed eddy does
emerge. The branch exhibits the expected behavior as the wave amplitude
increases-the wave shape shortens and steepens.
Second, we examine a branch of depression waves in a relatively deep fluid
with a thin pycnocline. The fixed parameters are 1=4, B = 0.01, T = 0.1,
A = 1 X 10- 4 , while Yo increases along the branch as (1) Yo = 0.75, (2) Yo = 0.80,
(3) Yo = 0.85. The streamline portraits are given in Figure 3, and the multipliers
are (1) A = 495, (2) A = 530, (3) A = 577. These moderately nonlinear waves
display a rather sensitive dependence of their shape on changes in the pycno-
cline level. On the basis of examples such as this, it seems fair to say that the
most difficult characteristic of a solitary wave to predict is the half-amplitude
wavelength. Consequently, formal asymptotic theories commonly fail to estimate
this characteristic adequately.
Third, we display a somewhat singular branch of depression waves in a
shallow fluid with a relatively thin pycnocline. The fixed parameters are I = 10,
B = 0.01, T = 0.2, A = 1 X 10- 3 , while Yo decreases along the branch as
(1) Yo = 0.68, (2) Yo = 0.65, (3) Yo = 0.62. The less familiar streamline portraits
given in Figure 4 show a marked broadening of the wave shape as the
pycnocline level decreases. The multipliers corresponding to these strongly
nonlinear solutions are computed to be almost identical (up to numerical error):
(1) A = 503, (2) A = 503, (3) A = 503. This curious phenomenon gives special
support to our claim that computational robustness is enhanced by taking the
constraint value A as a prescribed parameter rather than A, which is fixed in
other methods [31]. On the other hand, our iterative algorithm has a slow
(linear) rate of convergence in this particular case, unlike the typical case of a
strongly nonlinear solution, where it converges at a reasonably fast rate. It is
also worth noting that this broadening effect occurs at a fixed value of A and
does not involve waves containing closed eddies or exceedingly large ampli-
tudes.

Appendix. Energy-impulse variational principle


The (p, 0') system (2.3) in a domain D = {- 00 < x < + 00, 0 < y < h} (say) has
been expressed as a (noncanonicaI) Hamiltonian system by Benjamin [6]. On
this basis, he has documented the symmetries and their concomitant conserved
quantities associated with (2.3), which include the functionals

H(p,O') = f [!O'G O'+gy(p-4»]dxdy


D
p (energy), ( 1)

I(CT) = f/CTdxdy ( x-impulse) , (2)

A(p) = f [a(p) - a(p)] dxdy,


D
B(p,CT) = f CTb(p) dxdy
D
(Casimirs),

(3)
Solitary Internal Waves 125

where a and b are arbitrary (suitably smooth) functions, and Gp: 0' ~ I/J denotes
the Green operator for the linear elliptic boundary-value problem LpI/J = 0' in
D, I/J = 0 on aD. That these functionals are constant in time under evolution
governed by (2.3) may also be verified directly. An elegant variational principle
for steady translational waves p = p(x - ct, y), 0' = O'(x - ct, y) can be fash-
ioned from the above functionals. Let a and b be chosen as

a ( r) = g J;p - 1( r ,) dr' , b(r) = p-l(r), ( 4)

and let modified energy and x-impulse functionals Hand f, respectively, be


defined by

H(p,O') = H(p,O') - A(p), i(p,O') = 1(0') - B(p,O'). (5)

Equivalently, these functionals can be written as

H(p,O') = fD {tIVI/J12p + g f;[Y - p-l(r')] dr'} dxdy,

f(p,O') = !O'[y-p-l(p)]iUdy.
D

Then, a solution pair (p, 0') of (2.4) is a critical point of

fie p, 0') ~ crit subject to f(p,O') = P (6)

with corresponding constraint value P, and the wave speed c is its associated
multiplier. To verify this claim, we calculate the (partial) functional derivatives

Hp = -iIVI/J12 + g[y_p-l(p)]' Ha = I/J, (7a)

0'
fp = fa = y = p-l(p). (7b)
p'(p-l(p») ,

Consequently, the Euler-Lagrange equation H' - cI' = 0 is precisely the system

I/J = C[y_p-l(p)], (8a)

_ i1VI/J1 2 + g[y _ p-l(p)] cO'


(8b)
p'{p-l(p») .

It is straightforward to verify that (8) implies (2.4). In fact, recalling the


126 B. Turkington, A. Eydeland, and S. Wang

definition of 7], we see immediately that (8a) is identical to (2.9) and that (8b)
then becomes equivalent to (2.11).
The variational principle (2.17) that is the basis for the body of the present
paper may be viewed as a reduced form of energy-impulse principle (6). To this
end we note that the unknowns p and 7] are interchangeable according to the
transformation p = jj( y - 7]), as explained in Section 2. Then the kinematical
condition (8a) (namely, 1/1 = C7]) can be used to determine a in terms of 7] and
c; consequently, the dynamical condition (8b) can be expressed in terms of 7]
and c, resulting in our basic equation (2.12). Of course, such a reduction is
possible only for steady configurations (7], c) satisfying the kinematical condition
(8a), not for arbitrary points (p, a) in the phase space for the Hamiltonian
system (2.3). Nevertheless, this procedure does provide a natural connection
between two different variational characterizations of steady solitary waves. In
fact, the two (Lagrangian) functionals, Ii - ci and E - AF, are identified up to
a constant factor under this reduction; indeed, the substitution p = jj(y - 7]),
1/1 = C7] yields

- -
H(p,a)cI(p,a) = I. {I
D Ic
2 2 17) }
IV7]1 jj(Y-7])-g 0 gjj'(y-g)dg dxdy

- C I. C7]L (Y-7)7] dxdy


D
p

-C 2 1. tIV7]1 jj(Y-7])dxdy + ghl.f(Y,7])dxdy


D
2

= -c 2 [E(7])-AF(7])]. (8)

In light of the above remarks we see that the modified form of Benjamin's
variational principle (6) is in a sense reducible to our variational principle (2.17).
The main virtue of (2.17) is that it poses a constrained minimization problem
that is tractable with relatively standard methods of analysis and computation.
Also, once a solution is obtained, the energy-impulse version (6) may then be
invoked to examine the nonlinear stability of the solution by appealing to the
Arnold method [1, 6]. Such a stability analysis is suggested by Benjamin, but will
not be attempted here.

References
1. H. ABARBANEL, D. HOLM, J. MARSDEN, and T. TATIU, Nonlinear stability analysis of stratified
fluid equilibria, Philos. Trans. Roy. Soc. London Ser. A 318:349-409 (1986).
2. C. J. AMICK, Semilinear elliptic eigenvalue problems on an infinite strip with an application to
stratified fluids, Ann. Scuola Norm. Sup. Pisa (4) 11:441-499 (1984).
3. J. APEL, J. HOLBROOK, A. LIU, and J. TSAI, The Sulu sea internal soliton experiment, f. Phys.
Oceangr. 15:1625-1651 (1985).
4. T. B. BENJAMIN, Internal waves of finite amplitude and permanent form, f. Fluid Mech.
25:241-270 (1966).
Solitary Internal Waves 127

5. _ _ , Internal waves of permanent form in fluids of great depth, 1. Fluid Mech. 29:559-592
(1967).
6. _ _ , On the Boussinesq model of two-dimensional wave motions in heterogeneous fluids,
1. Fluid Mech. 165:445-474 (1986).
7. D. J. BENNEY, Long nonlinear waves in fluid flows, 1. Math. Phys. 45:52-63 (1966).
8. D. J. BENNEY and D. R. S. Ko, The propagation of long large amplitude internal waves, Stud.
Appl. Math. 59:187-199 (1978).
9. D. J. BENNEY and R. GRIMSHAW, Large amplitude solitary waves in unbounded stratified fluids,
Stud. Appl. Math. 66:181-187 (1982).
10. J. BONA, D. BOSE, and R. E. L. TURNER, Finite amplitude steady waves in stratified fluids,
1. Math. Pure Appl. 62:389-439 (1983).
11. D. CHRISTIE, M. MUIRHEAD, and A. HALES, On solitary waves in the atmosphere, 1. Atmo-
spheric Sci. 35:805-825 (1978).
12. D. CHRISTIE, K MUIRHEAD, and R. CLARKE, Solitary waves in the lower atmosphere, Nature
293:46-49 (1981).
13. R. E. DAVIS and A. ACRIVOS, Solitary internal waves in deep water, 1. Fluid Mech. 29:593-607
(1967).
14. R. J. DOVIAK and R. GE, An atmospheric solitary gust observed with a Doppler radar, a tall
tower and a surface network, 1. Atmospheric Sci. 41:2559-2573 (1984).
15. A. E. GARGETT, Generation of internal waves in the Strait of Georgia, British Columbia, Deep
Sea Res. 23:17-32 (1976).
16. 1. A. GEAR and R. GRIMSHAW, A second-order theory of solitary waves in shallow fluids, Phys.
Fluids 26:14-29 (1983).
17. D. GILBARG and N. S. TRUDINGER, Elliptic Partial Differential Equations of Second Order,
Springer-Verlag, Berlin, 1977.
18. P. GILL, W. MURRAY, and M. WRIGHT, Practical Optimization, Academic, London, 1981.
19. A. D. IOFFE and V. M. TIHOMIVOV, Theory of Extremal Problems, Elsevier North-Holland, New
York, 1979.
20. R. I. JOSEPH, Solitary waves in a finite depth fluid, J. Phys. A 1O:L225-L227 (1977).
21. C. G. Koop and G. BUTLER, An investigation of internal solitary waves in a two-fluid system,
J. Fluid Mech. 112:225-251 (1981).
22. T. KUBOTA, D. R. S. Ko, and L. DOBBS, Propagation of weakly nonlinear internal waves in a
stratified fluid of finite depth, 1. Hydronaut. 12:157-165 (1978).
23. Y.-L. LIN and R. C. GOFF, A study of mesoscale solitary wave in the atmosphere originating
near a region of deep convection, 1. Atmospheric Sci. 45:194-205 (1988).
24. R. R. LONG, On the Boussinesq approximation and its role in the theory of internal waves,
Tellus 17:46-52 (1965).
25. A. LIU, J. HOLBROOK, and J. APEL, Nonlinear internal wave evolution in the Sulu Sea, 1. Phys.
Oceanogr. 15:1613-1624 (1985).
26. A. R. OSBORNE and T. L. BURCH, Internal solitons in the Andaman Sea, Science 208:451-460
(1980).
27. B. N. PSHENICHNY and Yu. M. DANILIN, Numerical Methods in Extremal Problems, MIR,
Moscow, 1978, Chapter III, Section 6.
28. R. T. ROCKAFELLAR, Convex Analysis, Princeton U.P., Princeton N.J., 1970.
29. H. SEGUR and J. L. HAMMACK, Soliton models of long internal waves, 1. Fluid Mech.
118:285-304 (1982).
30. A. M. TER-KRIKOROV, Theorie exacte des ondes longues stationnaires dans un liquide
heterogene, 1. Mec. 2:351-376 (1963).
31. K-K TUNG, T. F. CHAN, and T. KUBOTA, Large amplitude internal waves of permanent form,
Stud. Appl. Math. 66:1-44 (1982).
32. S. WANG, Doctoral Dissertation, Univ. of Massachusetts, Amherst, in preparation.
33. C. S. YIH, Stratified Fluids, Academic, New York, 1980, Chapter 3.

UNIVERSITY OF MASSACHUSETTS

(Received July 2, 1990)

You might also like