The Galerkin vertical representation of temperature
in terms of basis functions 
is:
where 
are coefficients of the expansion in p chosen to minimize
square error under an inner product 
that is just
a weighted pressure average. We split off a reference profile
that does
not vary in space because this can be specified or treated in a separate
calculation--only gradients of T affect the circulation. Note that
is not assumed to be a solution of the system, and that the
spatial average of the model output differs from .
The first trick that aids both numerics and interpretation
is that basis functions for the wind fields are constructed from
approximate analytic solutions. Consider the leading temperature
basis function, 
. Temporarily neglecting
and advection by baroclinic wind,
the solution to the momentum equation (3) for circulations driven by
this baroclinic pressure gradient, satisfying simple lower and upper
boundary conditions, has the form
where
Using such analytically constructed basis functions to derive
nonlinear Galerkin equations
can be highly accurate at low truncation, and provides a physical
association of 
and T components.
The approach (6)-(7) could be used generally but larger gains occur
when the temperature variations that are typically encountered are
restricted by convection. In our approach, we tailor the temperature
basis functions to typical quasi-equilibrium profiles.
One of the main effects of convection on the large scales is to constrain
the vertical temperature profile through the depth of the convecting
region--highly unstable profiles with large convective available potential
energy tend to be modified by convection (e.g., Xu and Emanuel 1989;
Randall and Wang 1992).
For many convection schemes one can explicitly or implicitly define, for
a given set of large scale variables, a quasi-equilibrium temperature profile
that convection brings the large scale temperature towards
(e.g., Manabe et al 1965; Arakawa and Schubert 1974; Betts 1986).
Thus the variations
in quasi-equilibrium temperature profile can be efficiently expressed in the
form (6) but with only N coefficients 
, where
is small. The coefficients 
depend on the large scale
T and q variables. If 
is typical of temperature variations
near deep convective regions, associated velocity variations will be largely
carried in 
. Similarly for shallow convection, and so on.