1 MAY 2002 BOKHOVE 1619

Eulerian Variational Principles for Stratified Hydrostatic Equations

O. BOKHOVE
Faculty of Mathematical Sciences, University of Twente, Enschede, Netherlands

(Manuscript received 19 December 2000, in final form 9 August 2001)

ABSTRACT
The aim of this paper is to advocate the use of variational and Hamiltonian formulations of the hydrostatic
equations of motion in finding new conservative numerical techniques for forecast models. For that reason, the
fundamental conservative structure of the hydrostatic equations of motion is presented. Variational principles
and Hamiltonian formulations of various hydrostatic equations, stratified continuously or in layers, are derived
systematically from an Eulerian perspective in the horizontal and a Lagrangian or material perspective in the
vertical direction. Variational formulations are derived for the hydrostatic incompressible Boussinesq system
and the hydrostatic equations in multiple isentropic and isopycnic layers. The various Hamiltonian formulations
presented share similar Poisson brackets; that is, the (contribution to the) Poisson bracket in each layer or in
the continuously stratified case is the one for the shallow-water equations with the depth replaced by the
appropriate pseudodensity, while the potential (and internal) energy in the Hamiltonian differs in each case. For
hydrostatic equations in material coordinates, either stratified continuously or in layers, the coupling between
layers happens solely through the Hamiltonian, an observation that may aid in searching for conservative
numerical discretizations.

1. Introduction Euler equations and the hydrostatic Euler equations in

Hydrostatic ‘‘primitive’’ equations of motion are
widely used in atmospheric and oceanic fluid dynamics,
in particular in numerical weather prediction and at-
mosphere–ocean climate modeling. In the atmosphere
these equations are derived from the compressible Euler
equations after using the hydrostatic approximation,
while in the ocean the incompressibility condition and
the Boussinesq approximation in addition are often in-
voked.
The aim of the present paper is to enable the use of
variational and Hamiltonian (i.e., geometric) formula-
tions in finding new conservative numerical techniques.
For that reason, I formulate previously unknown geo-
metric formulations of several stratified hydrostatic
models used in geophysical fluid dynamics.
Variational principles and Hamiltonian formulations
of various hydrostatic equations, stratified continuously
or in layers, are derived from an Eulerian perspective
in the horizontal and a Lagrangian or material perspec-
tive in the vertical direction. Material coordinates in the
vertical are at least conceptually advantageous in that
they eliminate the reference to the vertical velocity in
hydrostatic systems. As a starting point, Eulerian var-
iational formulations of the atmospheric nonhydrostatic
Corresponding author address: Dr. O. Bokhove, Faculty of Math-
ematical Sciences, University of Twente, P.O. Box 217, Enschede
7500AE, Netherlands.
E-mail: o.bokhove@math.utwente.nl
isentropic coordinates are outlined, while novel for-
mulations are derived for the hydrostatic incompressible
Boussinesq system and the hydrostatic equations in mul-
tiple isentropic and isopycnic layers. These Eulerian
variational formulations consider the equations of mo-
tion from an Eulerian viewpoint for the horizontal co-
ordinates while a Lagrangian viewpoint is adopted in
the vertical by using material coordinates. Such hydro-
static Eulerian variational formulations in an Eulerian
framework (in the horizontal) can be contrasted with
the more well-known Lagrangian variational formula-
tions in a Lagrangian framework. In the former the hor-
izontal fluid parcels or labels are functions of the fixed
horizontal coordinates and time, while in the latter the
positions of the fluid parcels are functions of the fluid
labels and time. Boussinesq models in isopycnic coor-
dinates are often used in ocean circulation models, such
as the Miami Isopycnic Coordinate Ocean Model (MI-
COM) model (e.g., Bleck and Smith 1990). In contrast,
material coordinates have been less popular in atmo-
spheric numerical models, since the use of terrain-fol-
lowing or sigma coordinates in numerical models ap-
pears to be more efficient near the earth’s boundary.
(Boundary conditions are implemented more easily in
sigma coordinates than in material coordinates.) Nev-
ertheless, the conceptual advantages of isentropic co-
ordinates have been recognized in the atmospheric com-
munity (e.g., Hoskins et al. 1985).
The structure of the paper is as follows. Eulerian var-

q 2002 American Meteorological Society

1620 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 59

iational principles and Hamiltonian formulations of the

continuously stratified hydrostatic systems are outlined
and derived in section 2 and for the layer models in
section 3. Finally, a discussion of the applicability of
the Eulerian variational principles is found in section 4.
The review papers by Shepherd (1990) and Morrison
(1998) provide good introductions to the realm of var-
iational and Hamiltonian fluid dynamics. Conservation
laws are related to variational and Hamiltonian for-
mulations via Noether’s theorem, but I will not explicitly
explore their properties presently and refer readers to
the before-mentioned review papers.

2. Continuous stratification
a. Stratified equations
The Eulerian variational principle for the Euler equa-
tions follows directly from a transformation of the La-
grangian variational principle. The variables in the La-
grangian framework are the positions x 5 (x, y, z) 5
x(a 3 , t) as function of the fluid parcel labels a 3 5 (a,
b, c) and time t, while in the Eulerian framework their
role is reversed: the fluid labels are variables a 3 5 a 3 (x, FIG. 1. Sketch of the N material layers above the earth’s surface
z, t) as function of space and time. The following Eu- or ocean bottom, which resides at z 5 h b and where the pressure is
p b . ‘‘Material’’ constant n i in layer i is either entropy s i in isentropic
lerian principle for several stratified equations of motion layer or density r i in isopycnic layer models. The pressure at the top
is the difference between the kinetic energy and the of layer i is p i and this layer has thickness h i . The passive atmospheric
potential and internal energy: pressure above the ocean or in the outer atmosphere is p1 5 p a .

05d E 1
t1

t0
dt L EHB a3 ,
da3
dt 2 The following equations of motion emerge from the

5d E E E
t1

t0
dt
DH
dx
hb
hT

dt
variational principle (2.1):
• The Euler equations emerge when we take d H 5 e B
5 1 and r q 5 r and specify an equation of state p 5

3 rq
[1 1
2 2
1
u i 1 R i u i 1 d H r q w 2 2 rgz
2
p(r, s) with pressure p (e.g., Holm 1996). To illustrate
these and upcoming variational calculations, I have
included the straightforward derivation of the Euler

2 e B rU(s, r) ,
] (2.1)
equations from a one-dimensional vertical version of
principle (2.1) in the appendix.
• When we take d H 5 0 and r q 5 r, and specify p 5

) )
](a3 ) p(r, s), the hydrostatic primitive equations arise,
r 5 r R (a) det , and (2.2) which are (in a modified form) still used in numerical
](x, z)
weather and climate prediction. When we transform
]a3 ]a (2.1) from Cartesian to isentropic coordinates in the
1 (v · =)a3 1 w 3 5 0, (2.3) vertical, we arrive at the variational principle inves-
]t ]z tigated by Bokhove (2000).
where the Coriolis parameter f 5 f (x, y) 5 ẑ · = 3 R • Finally, when we take d H 5 e B 5 0 and r q 5 r 0 with
with R 3 5 0, u i 5 dij u j is the horizontal velocity (the r 0 as a constant reference density, the hydrostatic
rule of summation over repeated indices is used and low- incompressible Boussinesq equations appear, which
ercase indices run from 1 to 2); = 5 (]/]x, ]/]y)T ; v and are often used in large-scale oceanography. In the
w are the horizontal and vertical velocity, respectively; next section, the variational principle (2.1) will be
rR is a reference Lagrangian density; g is the constant of transformed from Cartesian to isopycnic coordinates
gravity; DH is the horizontal extent of the domain; U(s, in the vertical and it will be shown that the hydro-
r) the internal energy as function of entropy s and density static Boussinesq equations in isopycnic coordinates
r; and the fluid lies between the topography or earth’s naturally follow from such a transformed variational
surface at z 5 hb and the free surface or outer atmosphere principle.
at z 5 hT .
1 MAY 2002 BOKHOVE 1621

b. Hydrostatic incompressible Boussinesq equations and in which variations are taken with respect to fluid
To derive the variational principle for hydrostatic in- labels a 5 (a, b)T (x, r, t). Fluid parcels a(x, r, t) are
compressible Boussinesq flows, we take d H 5 e B 5 0 thus advected by the horizontal fluid velocity v [ (u,
and r q 5 r 0 . Moreover, since the fluid is incompressible y )T on isopycnic surfaces labeled by coordinate r:
and Boussinesq, we have a pseudodensity,
]a ]a k ]a k
]z v 5 2G21 ⇔ 1 u i i 5 0, (2.6)
s 5 r0 , (2.4) ]t ]t ]x
]r
as the Jacobian between Cartesian and isopycnic co- where the tensor G ki 5 ]a k /]x i , and the pseudodensity
ordinates; that is, r 0 dxdydz 5 sdxdydr. With this def- s is defined by s 5 s 0 (a, r)J(a, b) with the horizontal
inition of pseudodensity, a direct transformation of Jacobian J(a, b) [ ] x a] y b 2 ] x b] y a. We refer to Holm
principle (2.1) to isopycnic (or density) coordinates (1996) and Bokhove (2000) for relations useful in the
gives evaluation of Hamilton’s principle (2.5). In extension to

E 1 2
t1
the atmospheric hydrostatic equations of motion in is-
da entropic coordinates, a free surface has been included
05d dt L B a,
t0
dt to accommodate oceanic applications. Furthermore, the

E E E
system is constrained to be incompressible and the in-
r T (x,t)
ternal energy term is therefore absent (e B 5 0), while
t1

[d dt dx dr
t0 DH r B (x,t)
the potential energy appears in a different form as is
usual. To evaluate the variation of the potential energy,
the Montgomery potential M 5 (p 1 rgz)/r 0 is defined.
3 5[ s 1
2
(x, r, t)u i (x, r, t) 1 R i (x) u i (x, r, t)
] Hence with hydrostasy p z 5 2rg, we find z 5 (r 0 /
g)]M/] r (see Holm 1996). Several boundary terms ap-

6
1 ][z(x, r, t) 2 ] pear after varying (2.5), and cancel one another or are
2 rg , (2.5) zero by using suitable boundary conditions, such as tak-
2 ]r ing p a 5 0 at the free surface. Finally, variation of (2.5)
in which the kinetic minus the potential energy appears, with respect to da k and (da k ) B,T yields

05 E E E t1

t0
dt
DH
dx
rB
rT

dr s (G21 ) kn 1 ]t 1 u ]x 1 u ]x
]u n j
]u n
j
j
]R n
j
2 uj
]R j
]x n
]M
1 n da k
]x 2
1E E 5[ t0
t1

dt
DH
dx s (u n 1 R n )(G21 ) kn 1 ]t 1 u ]x 2]) (da ) 2 [s (u 1 R )(G ) 1 ]t 1 u ]x 2]) (da ) 6 ,
]r T j
]r T
j
rT
k
T n n
21 n
k
]r B j
]r B
j
rB
k
B

(2.7)

where subscripts ( ) B,T denote evaluation at r 5 r B,T , momentum equations are second-order partial differ-
respectively. The equations of motion arising from these ential equations in time for the fluid labels, and are
variations are therefore horizontal advection of the is- Euler–Lagrange equations. The continuity equation ex-
opycnals at the topography and the free surface, that is, pressed in terms of the pseudodensity follows directly
z 5 h B (x) and z 5 h T (x, t), respectively: from the definitions (2.6) and s 5 s 0 (a, r)J(a, b) as
follows:
]r B,T ]r
(da k ) B,T : 1 u k B,T 50 (2.8)
]t ]x k
1 2
]s ]s ]a k ] 2 a m ]a n
5 k0 J(a, b) 1 s0 e i je mn
and the horizontal momentum equations, ]t ]a ]t ]x i]t ]x j

1 2
]u m ]u ]R m ]R ]M ]s ]u j
da k : 1 u k mk 1 u k 2 mk 1 m 5 2u j 2 s j, (2.10)
]t ]x ]x k ]x ]x ]x j
]x
5 0, (2.9)
where e ij and e mn are permutation symbols.
where the relation between s and M is s 5 (r /g)] M/ 2
0
2 The generalized interior momentum corresponding to
]r 2 . As in the isentropic case (see Bokhove 2000), the (2.5) is (e.g., Morrison 1998)
1622 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 59

1 2@ 1 2
da ]a k flows, by dropping the vertical part in the kinetic en-
p k (x, t) 5 dL B a, d ergy, we find,
dt ]t

5 s (G ) 1d (G )
21 m
k
]a
]t
2R 2mn
21 n
j
j

m
05d E E E
t0
t1

dt
DH
dx
hT

hb
dz

5 2s (G21 ) km (u m 1 R m ). (2.11)
An Eulerian action principle follows after a Legendre
transform and may be rewritten in terms of p and a,
or in terms of v and a. In terms of the latter, one finds
3r
[1 1
2
u 1 R i u i 2 gz 2 U(s, r) .
2 i ] (3.1)

We assume the atmosphere to consist of an ideal gas

05d E 5E E [
t1

t0
dt
DH
dx
rT

rB
dr 2s (u m 1 R m )(G21 ) km
]a
]t ]
k with p 5 rRT and R the gas constant, relating pressure
to density and temperature. In that case, U(s, r) 5 c y T
with c y the specific heat at constant volume. From the
first law of thermodynamics, T ds 5 c p dT 2 (1/r)dp,
2 H [u i , a i ] , 6 (2.12) and the ideal gas law, we can find,

1 2 1 2
R/c p
with the Hamiltonian as a Legendre transform p p
T 5 T00 e (s2s00)/cp [u , (3.2)
H [a, v] 5 H [s, v] p00 p00

5
DH
E E dx
rB
rT

dr
2 12 s |v|
1 ](z 2 )
1 rg
1
]r
. (2.13)
2
2 with k 5 R/c p ; specific heat at constant pressure c p ;
potential temperature u; and reference pressure, entropy,
Variations in (2.12) are taken with respect to v and a. and temperature p 00 , s 00 , T 00 ; respectively. For an ideal
Similar variations as in Hamilton’s principle (2.5) yield gas and with hydrostasy ]p/]z 5 2rg, the variational
the horizontal momentum equations (2.9) on isopycnic principle (3.1) can be rewritten as
surfaces for variations, da, the label advection equations
(2.6) for variations dv, and (2.8) for (da k ) B,T . The gen-
eralized momentum corresponding to ]a kB,T /]t is zero,
which signals that the Lagrangian (2.5) is singular (see
05d E E E
t1

t0
dt
DH
dx
hT

hb
dz

e.g., Sudarshan and Mukunda 1974), but only at the

boundary.
A Hamiltonian formulation of the hydrostatic in-
compressible Boussinesq equations in isopycnic co-
3 r
[1 1
2 2
u i 1 R i u i 2 c p rT 1
]( pz)
]z
.
] (3.3)

ordinates, in terms of label variables a and velocity
v, may be derived directly from action principle (2.12)
(Sudarshan and Mukunda 1974)—for simplicity, tak-
ing r B,T constant at the boundaries. The Hamiltonian
formulation described in Holm and Long (1989), in
We will next consider an atmosphere that consists of
several constant entropy layers in each of which the
fluid velocity is assumed to be independent of depth
initially. Hence it will remain so at subsequent times.
With hydrostasy, (3.3) transforms to

E E E
which the advection of boundary isopycnals is not
t1 hT
included, will subsequently appear when we apply the
reduction theory developed by Marsden and Wein- 05d dt dx dz
t0 DH hB
stein (1983) to this Eulerian Hamiltonian fluid-parcel
formulation.

3. Layer models
3
[1
]p 1
]z 2
ui 1 Ri u i 2 g
]( pz)
]z 22 c pu
]p p k
]z p00
k
. (3.4)
]
a. Constant entropy Figure 1 depicts an N-layer atmosphere above topography
When we apply the hydrostatic approximation to z 5 hb and with pressure pa in the passive layer above
the variational principle (2.1) for nonhydrostatic the active atmospheric layers. Integration of (3.4) in z gives

05d E E 5
t0
t1

dt
DH
dx ( p N 2 p b ) 12 u
1
Ni
2
1 R i u Ni 1 gp b h b 2 c p k
p00
uN
( pk11 2 pkb11 )
(k 1 1) N

O [( p 12 u 2
ua
]
N21
1
1 a 2 pa11 ) ai 1 R i uai 2 c p ( pk11 2 pka11
11 )
a52 pk00 (k 1 1) a
1 MAY 2002 BOKHOVE 1623

1 ( p a 2 p2 ) 12 u
1
1i
2
1 R i u1i 2 gp a (h1 1 h2 1 · · · 1 h N 1 h b ) 2 c p
u1
6
( pk11 2 pk2 11 ) .
pk00 (k 1 1) a
(3.5)

Let us also define pseudodensities s N 5 (p b 2 p N )/
g, s a 5 (p a 11 2 p a )/g, and s 1 5 (p 2 2 p a )/g in layers
a 5 1, 2, . . . , N, counting from the first active layer
in the outer atmosphere down to the Nth layer at the
earth’s surface or sea bottom at z 5 h b (x) and with
corresponding velocities v a . Each layer a has thick-
ness h a . The first layer thus reaches from z 5 h b (x)
to h b (x) 1 h N (x, t) with corresponding bottom and
interface pressure p b and p N , respectively. The second
layer starts at h b 1 h N with p 5 p N and finishes at h b
1 h N21 1 h N with p 5 p N21 , and so forth. The last
layer begins at h b 1 h 2 1 · · · 1 h N with p 5 p 2 and
ends at h b 1 h1 1 · · · 1 h N with fixed pressure p 5
p a . An element of mass rdxdz can be integrated along
columns in the vertical across each layer to find a
pseudodensity for each layer. For layer a we thus
obtain [(p a11 2 p a )/g]dx.
With these pseudodensities and their interrelations the
variational principle (3.5) for N layers of constant entropy
or constant potential temperature can be rewritten as,

05d E 1 t1

t0
dt L N aa ,
daa
dt 2
5d E E 5
t1

t0
dt
DH
dx s N 12 u
1
Ni
2
1 R i u Ni 2 gs N h b 2 c p k
uN
00
( p (k11) 2 1p N(k11) ) 1 p a h N
gp (k 1 1) b

O [s 112 u 2
ua
]
N21

3 a ai 1 R i uai 2 c p ( p (k11) 2 pa(k11) ) 2 gsa h b 1 p a ha

a52 gpk00 (k 1 1) a11

1 s1 12 u
1
1i
2
1 R i u1i 2 c p
u1
( p (k11) 2 p a(k11) ) 2 gs 1 h b 1 p a h1 .
gp (k 1 1) 2
k
00
6 (3.6)

The pressures at the interface of each layer follow from
the definition of the pseudodensities; for example, p 2 5
gs1 1 p a , . . . , p a 5 g(s1 1 · · · 1 s a21 ) 1 p a , . . . , p b
5 g(s1 1 · · · 1 s N ) 1 p a . Variations in (3.6) are taken
with respect to the fluid labels a a in each layer. For
simplicity we take p a 5 0, although we can generalize
the argument in a consistent manner, resulting in an extra
term p a (H T 2 h1 2 · · · 2 h N 2 h b ) in the integrals of
(3.6), in which H T . h T is the fixed height of a domain
filled partially by the active fluid and partially with a
passive gas of constant pressure p a .
b. Constant density
Consider the Eulerian variational principle for a hy-
drostatic flow in Cartesian coordinates
which is an incomplete variational principle for an
incompressible flow in hydrostatic balance with i 5
1, 2. It follows from (2.1) with d H 5 e B 5 0, r q 5
r. Density can be taken constant in each layer and
the velocity is defined in terms of fluid parcels, which
are advected by the flow. However, the principle
above is incomplete because the constraint imposing
the incompressibility condition is absent. For inviscid
flows in shallow layers of constant density, the ve-
locity in each layer remains independent of the ver-
tical coordinate when it started off so initially. The
integration over the layer depths can thus be taken
without explicitly imposing the incompressibility
condition. We refer again to Fig. 1 with the modifi-
cation that the density, instead of the entropy, is taken

E E E
to be fixed in each layer. To emphasize the analogy
05d
t0
t1

dt
DH
dx
h T (x,t)

h b (x)
dz r
[1 1
2
u 1 R i u i 2 gz ,
2 i ] between the entropy and density layer models we de-
fine s a 5 r a h a . An Eulerian variational principle for
N shallow layer equations arises when we integrate
(3.7) (3.7) across layers:
1624 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 59

05 E E 75
t1

t0
dt
DH
dx sN12 u 1 R 2 u 2 2 g [(s 1 r h ) 2 (r h ) ]@r 6
1
Ni i
i
N
1
N N b
2
N b
2
N

1 O 5s 1 u 1 R 2 u 2 g 1s 1
[ r s r s
2
N21 2
1 1 a a11 a
1 ··· 1 1 ra h b
i N
a ai a a
a51 2 2 r i
r a11 N

1 r h 2 @r 68 ,
1
ra sa11 r s
]
2

2 1 ··· 1 a N a a (3.8)
ra11 rN b

where variations are taken with respect to the fluid labels g(s 1 /r1 1 · · · 1 s N /r N 1 h b ),
a a in the N layers (i 5 1, 2; a 5 1, . . . , N).
 if a 5 1

O s /r 2 ,
1
Ma 5  a21

c. Equations of motion g sa /ra 1 · · · 1 s N /r N 1 h b 1 g a

g51

In both layer models the pseudodensities s a and the  if a 5 2, . . . , N.

velocities are again related to the fluid labels as
(3.13)
sa 5 s0 a (aa )J(aa , ba ), and These potentials emerge from the variational principle
after some manipulation, integration by parts, and after
]aa /]t 1 (va · =)aa 5 0. (3.9) using suitable boundary conditions. Suitable boundary
conditions include quiescence at infinity, periodicity, no
The continuity equations are again like (2.10): normal flow at solid boundaries, vanishing isentropic
layer thickness, or a combination thereof.
]s a /]t 1 = · (s av a ) 5 0. (3.10)

The momentum equations in each layer follow from d. Hamiltonian formulation

(3.6) or (3.8) and are The generalized interior momentum corresponding to
(3.6) or (3.8) is (e.g., Morrison 1998),
da a : ]v a /]t 1 (v a · =)v a 1 f ẑ 3 v a 5 2=M a .

1 2@ 1 2
(3.11) daa ]aak
pak (x, t) 5 dL N aa , d
dt ]t
The Montgomery potentials M a for the isentropic layer
model are defined as

k
5 sa (G21
[
a ) k d mn (Ga ) j
m 21 n
]aaj
]t
2 Rm
]
1 2
pb
M N 5 gh b 1 c pu N 5 2sa (Ga21 ) km (uam 1 R m ), (3.14)
p00
(no summation over a) with (G ) 5 ]a /]x . After using k
a i
k
a
i

Ma 5 gh b 1 c pu N
[1 2 1 2 ]
pb
k

2
pN
k the Legendre transform expressed in terms of v and a,
we find the action principle

E 5E [ O
p00 p00
]aak
]
t1 N

05d sa (uam 1 R m )(G21

O [1 2 1 2 ] k k k dt dx a )k
m

1p 2 , ]t
N21
pg11 pg pa11 a51
1 c pug 2 1 c pua t0 DH
g5a11 p00 p00 00

(3.12) 1 H N [aa , va ] . 6 (3.15)

where a 5 1, . . . , N 2 1, and the ones for the isopycnic The respective Hamiltonians as Legendre transforms H N
layer model are are,

H N [sa , va ] 5 E 5 DH
dx
1
2
s N |v N | 2 1 gs N h b 1 c p k N
u
( p (k11) 2 p N(k11) )
gp00 (k 1 1) b
1 MAY 2002 BOKHOVE 1625

O [12 s |v | 1 gs h 1 c gp (ku 1 1) ( p ]
N21
a 1
1 a a
2
a b p k
(k11)
a11 2 pa(k11) ) 1 s 1|v1| 2 1 gs 1 h b
a52 00 2

1 cp
u1
p (k11) ,
gp (k 1 1) 2
k
00
6 (3.16)

for the entropy model, and

H N [sa , va ] 5 E 7O 5
DH
dx
N

a51
1
2
1 r s
[1 r s
sa |va | 2 1 g sa 1 a a11 1 · · · 1 a N 1 ra h b
2 ra11 rN 2
2

1 2 ]@r 68 ,
ra sa11 r s
2

2 1 · · · 1 a N 1 ra h b a (3.17)
ra11 rN

with h b 5 s N11 /r N11 , for the density model, respectively. bracket is disconnected from the others. The first brack-
Similar variations as in Hamilton’s principles (3.6) or et, in terms of the variables or fields {p a , a a }(x, t), is
(3.8) yield the N horizontal momentum Eqs. (3.11) in an infinite-dimensional canonical bracket, which is
each isentropic or isopycnic layer for da a , and the ad- skew-symmetric [F, G] 5 2[G, F ] and satisfies Jacobi’s
vection Eq. (3.9) for dv a . identity [K, [F, G]] 1 [F, [G, K ]] 1 [G, [K, F ]] 5 0
A Hamiltonian formulation of the hydrostatic equa- (e.g., Morrison 1998). This bracket yields Hamilton’s
tions in the N layers has the form equations and follows directly from the corresponding
action principle (cf. the finite-dimensional case of clas-
dF
5 [F, H N ]. (3.18) sical mechanics). The second and third brackets, which
dt are expressed in terms of the variables {v a , a a } and
In terms of the three sets of variables {a a , p a }, {a a , {v a , s a }, are transformations of this canonical bracket
v a }, and {s a , v a }, we can find the following generalized and are therefore also skew-symmetric and satisfy Ja-
Poisson brackets: cobi’s identity. The third bracket only includes three
variables v a and s a per layer, indicating that there is a
O E dx 5ddaF · dpdG 2 dpdF · ddaG 6
N
symmetry, associated with Noether’s theorem, which
[F, G ] 5 (3.19) allows for this reduction (see e.g., Morrison 1998). The
a51 DH a a a a

5 O E dx 5q ẑ ·
shallow-water equations have a similar bracket (e.g.,
N
dF dG Shepherd 1990) as each layer has here.
3a
a51 DH
dv dv a a

4. Summary and discussion

1 2
dG 1 dF
1 · · = aa Starting with the Eulerian variational principle for the
daa sa dva
Euler equations, I have systematically derived the var-
iational and Hamiltonian formulations of the hydrostatic
1 2 6
dF 1 dG
2 · · = aa (3.20) incompressible Boussinesq equations in isopycnic co-
daa sa dva ordinates, and the hydrostatic multilayer isentropic or

O E dx 5q ẑ · ddvF 3 ddvG 2 dsdF 1= · ddvG 2

isopycnic systems (sections 2 and 3).
N
The various geometric formulations derived in the
5 a present paper all share similar Poisson brackets; that is,
a51 DH a a a a
the (contribution to the) Poisson bracket in each layer
or in the continuously stratified case is the one for the
1 26 ,
dG dF
1 =· (3.21) shallow water equations with the depth h replaced by
dsa d va the appropriate pseudodensity s, while the potential
respectively, where we have defined the potential vor- (and internal) energy in the Hamiltonian differs in each
ticity in each layer as case. For both the layer and continuous models, the
f 1 ẑ · = 3 va coupling between each (infinitesimal) layer happens
qa 5 . (3.22) solely through the Hamiltonian (e.g., Salmon 1982),
sa which is interesting from a numerical viewpoint: any
We notice that the coupling between the layers is fa- numerical progress in preserving part of the Poisson
cilitated only through the Hamiltonian, and that the con- bracket structure for the shallow-water equations, say,
tribution from one layer to the (generalized) Poisson can perhaps extend to these similar hydrostatic systems.
1626 JOURNAL OF THE ATMOSPHERIC SCIENCES VOLUME 59

Moreover, when we discretize only the vertical material
coordinate in the Poisson bracket for the continuously
stratified hydrostatic models in a straightforward man-
ner, we can derive the Poisson bracket for the layer
models.
All the hydrostatic models considered have material
or layer coordinates in the vertical. Careful attention
must therefore be paid to the movement of the isen-
tropes and isopycnals or the (internal) interfaces over
the (sloping) boundaries. The movement of these ma-
terial surfaces or interfaces at boundaries tends to be
difficult to model numerically. For numerical bench-
mark purposes, it is thus important to validate numer-
ical simulations against exact solutions in which the
‘‘water line’’ movement is an essential feature. Both
the one-layer shallow-water and the one-layer isentro-
pic models admit such exact solutions. It may be in-
teresting to investigate whether all these special so-
lutions can be inferred from their variational or Ham-
iltonian formulations, and henceforth whether it would
lead us to novel solutions.
This paper provides the conservative structure, which
can form the starting point for the development of con-
servative numerical integration techniques in atmo-
spheric and oceanic forecasting models. The Eulerian
variational principles may form an alternative route for
attempting to find conservative discretizations. Consider
for example the Eulerian variational principle for the
one-layer shallow-water equations, that is, (3.8) with N
5 1 and s 5 h:
therefore, in discretizing the Jacobian of the fluid label
variables and the label advection in Eq (3.9), which
define the height and the velocity, such that there is no
reference to the fluid labels required in the continuity
and momentum equations. It is unclear whether this is
possible or whether one has to settle for an approxi-
mation, for example, one in which the velocity and
height are defined as an average over the particle mo-
tions.
Finally, Eulerian variational principles can form the
starting point for deriving approximate balanced mod-
els, which describe large-scale low-Froude or Rossby
number dynamics (Holm 1996; Bokhove 2001). For
global dynamics in the atmosphere and oceans, the re-
sults of this paper need to be extended to the spherical
case and to quasi-hydrostatic approximations (White
1999; Staniforth 2000). However, there appear to be no
serious obstacles in using the presented techniques in
such a global context (see also, Roulstone and Brice
1995).
Acknowledgments. The research presented in this pa-
per results from a series of seminars for a group of
scientists from the Royal Dutch Meteorological Institute
and the National Research Institute for Mathematics and
Computer Science in the Netherlands. I wish to thank
J. G. Blom, J. E. Frank, D. Lanser, R. Pasmanter, J. G.
Verwer, and W. T. M. Verkleij for all their valuable
comments and discussion. I am grateful for the con-
structive criticism of the three anonymous reviewers.

05d E E 51t1

t0
dt
DH
dx h
1
u 1 Ri u i
2 i 2 APPENDIX

An Eulerian Variational Principle in One

Dimension
1
[
2 g (h 1 h b ) 2 2 h b2
2 ]6 Consider the one-vertical-dimensional form of the
Eulerian variational principle (2.1) for Euler flows:

5 E E 5[1
t1

t0
dt
DH
dx
1
2 i 2
u 1 R i u i 2 g(h 1 h b ) dh
] 05d E [ ]
t1

t0
dt L E a,
da
dt

1 hu idu i . 6 (4.1)
5d E E 5
t1

t0
dt dz r
1 2
2
w 2 gz 2 U(s, r) , 6 (A1)
In weighted residual methods such as the finite-element where the entropy s, the density r, and vertical velocity
method, the general approach is to find a weak for- w are economic shorthands for expressions in terms of
mulation, that arises as an integral of the relevant equa- the vertical fluid parcel variable a(z, t), as follows:
tions of motion multiplied by test functions when there 21

1 2
is no variational principle known. Nobody appears to ]a ]a ]a
s 5 s(a), r5 , w52 . (A2)
have tried to base a discretization on Eulerian variational ]z ]z ]t
principles for (inviscid) fluid equations. However, with Hence, we can substitute (A2) into (A1) and consider
a weighted residual or finite-difference approach, the
spatial part of (4.1) is readily discretized since no spatial
derivatives appear. Due to the presence of fluid labels
in the formulation, one can expect Lagrangian chaotic
05d E E t1

t0
dt dz
[
1 ]a
2 ]z
21

1 21 2
]a
]t
2

2
]a
]z
gz

motion of the fluid labels, in the absence of Eulerian

chaotic motion, which may hamper numerical stability
and accuracy. The crucial and difficult problem lies,
2
]a
]z 1
U s(a),
]a
]z
; 2 ] (A3)
1 MAY 2002 BOKHOVE 1627

variations are thus taken with respect to fluid parcel the use of (A2), we find the more familiar form of the
variable a(z, t) only. Note that label a(z, t) thus plays momentum equation:
the role of the generalized coordinates q i in classical

1 2
mechanics (e.g., Morrison 1998). Alternatively, we can ]w ] 1 2 1 ]p
1 w 1 gz 1 5 0. (A8)
first take the variations of (A1) with respect to w, r, ]t ]z 2 r ]z
and s, and subsequently use the expressions in (A2),
which yields An equation of state and boundary and initial condi-
tions on a(z, t) are required in addition to (A7) and a

05d E t1

t0
dt 1 dw dw 1 dr dr 1 ds ds2
dL E dL E dL E separate continuity equation is then not required. Nev-
ertheless, it arises from parcel advection as follows.
With

5 E 5 t1

t0
dt 2
dL E ]a
dw
d
]z [1 2 ]
21
]a
]t
dL
1 Ed
dr
]a
]z 1 2 r5
]a
]z
, and
]a
]t
]a
5 2w 5 2rw,
]z
(A9)

we find the usual continuity equation

1
dL E ]s
ds ]a
ds . 6 (A4) 05
]
]t 1
r2
]a
]z
5
]r
]t
22
] ]a
]z ]t 1 2
The former approach [i.e., using (A3)] emphasizes the
role of the key label variable and the analogy with the ]r ](rw)
5 1 . (A10)
finite-dimensional realm, while the latter approach [i.e., ]t ]z
using (A4)] underscores that the variations of w, r, and
Since entropy s(z, t) 5 s[a(x, t)], we can derive the
s are constrained to independent variations of the label
entropy advection equation from (A9) and, with an
a only. We also need to use the first law of thermo-
equation of state, the Euler equations are complete.
dynamics and its variations:
Taking variations of the different variational princi-
22 ples in this paper is essentially similar to the above
1 2
p ]s ]a ]d a
dU 5 Tds 1 dr 5 T da 1 p , and variation, except that the algebra becomes more in-
r2 ]a ]z ]z volved when more than one spatial dimension is en-
countered. It may be helpful to verify part of the pre-
1
dp 5 c pdT 2 Tds. (A5) sented results in the main text by first restricting atten-
r tion to only one horizontal dimension.
When we vary (A3), we then find

E E [
REFERENCES
t1
]d a ]s
0 5 2d dt dz w 2 rT da
t0
]t ]a Bleck, R., and L. T. Smith, 1990: A wind-driven isopycnic coordinate
model of the North and equatorial Atlantic Ocean. 1: Model
development and supporting experiments. J. Geophys. Res., 95,

1 12 w
1 2
1 gz 1 U 1 2
p ]da
r ]z ] 3273–3285.
Bokhove, O., 2000: On hydrostatic flows in isopycnal coordinates.
J. Fluid Mech., 402, 291–310.

E E
——, 2001: Balanced models in geophysical fluid dynamics: Ham-

5
t0
t1

dt dz
[]w
]t
2T
]s
]z
iltonian formulation and formal stability. The Mathematics of
Atmosphere and Ocean Dynamics, J. Norbury and I. Roulstone,
Eds., Cambridge University Press, 63 pp.
Holm, D. D., 1996: Hamiltonian balance equations. Physica D, 98,

1
] 1 2
]z 2 1
w 1 gz 1 U 1
p
r
da 1 · · · ,
(A6)
2 ] 379–414.
——, and B. Long, 1989: Liapunov instability of ideal stratified equi-
libria in hydrostatic balance. Nonlinearity, 2, 23–25.
Hoskins, B. J., M. E. McIntyre, and A. W. Robertson, 1985: On the
use and significance of isentropic potential vorticity maps. Quart.
where I have left issues regarding boundary contri- J. Roy. Meteor. Soc., 111, 877–946.
Marsden, J. E., and A. Weinstein, 1983: Coadjoint orbits, vortices
butions due to integration by parts to the reader (e.g., and Clebsch variables for incompressible fluids. Physica D, 4,
take da 5 0 at a solid boundary). Hence we find the 394–406.
vertical momentum equation as an Euler–Lagrange Morrison, P. J., 1998: Hamiltonian description of the ideal fluid. Rev.
equation Mod. Phys., 70, 467–521.
Roulstone, I., and S. J. Brice, 1995: A Hamiltonian formulation of
]
]t [12 ]
2
]a
]z
21
]a
]t
1
] 1 2
]z 2 1
w 1 gz 1
1 ]p
r ]z
5 0,

which is second-order in time. If we rewrite (A7) with

2 (A7)
the quasi-hydrostatic equations. Quart. J. Roy. Meteor. Soc., 121,
927–936.
Salmon, R., 1982: The main shape of the thermocline. J. Phys. Ocean-
ogr., 12, 1458–1479.
Shepherd, T. G., 1990: Symmetries, conservation laws, and Hamil-
1628 JOURNAL OF THE ATMOSPHERIC SCIENCES
tonian structure in geophysical fluid dynamics. Advances in Geo-
physics, Vol. 32, Academic Press, 287–338.
Staniforth, A., 2000: Developing efficient unified nonhydrostatic
models. Proceedings of the 50th Anniversary of Numerical
Weather Prediction, Potsdam, Germany, A. Speleat, Ed., Deut-
sche Meteorologische Gesellschaft, e. V., 185–200.
VOLUME 59
Sudarshan, E. C. G., and N. Mukunda, 1974: Classical Dynamics: A
Modern Perspective. John Wiley, 615 pp.
White, A., 1999: Hydrostatic and quasi-hydrostatic versions of the
basic ‘‘New Dynamics’’ equations. Met Office Forecasting Re-
search Tech. Rep. 271, 34 pp.