At the modest vertical resolutions typical of climate models, simulati
ons produced by models based on semi-Lagrangian approximations tend to
develop a colder tropical tropopause than matching simulations from m
odels with Eulerian approximations, all other components of the model
being the same. The authors examine the source of this relative cold b
ias in the context of the NCAR CCM3 and show that it is primarily due
to insufficient vertical resolution in the standard 18-level model, wh
ich has 3-km spacing near the tropopause. The difference is first diag
nosed with the Held and Suarez idealized forcing to eliminate the comp
lex radiative-convective feedback that affects the tropopause formatio
n in the complete model. In the Held and Suarez case, the tropical sim
ulations converge as the vertical grid layers are halved to produce 36
layers and halved again to produce 72 layers. The semi-Lagrangian app
roximations require extra resolution above the original 18 to capture
the converged tropical tropopause. The Eulerian approximations also ne
ed the increased resolution to capture the single-level tropopause imp
lied by the 36- and 72-level simulations, although with 18 layers it d
oes not produce a colder tropopause, just a thicker multilevel tropopa
use. The authors establish a minimal grid of around 25 levels needed t
o capture the structure of the converged simulation with the Held and
Suarez forcing. The additional resolution is added between 200 and 50
mb, giving a grid spacing of about 1.3 km near the tropopause. With th
is grid the semi-Lagrangian and Eulerian approximations also create th
e same tropical structure in the complete model. With both approximati
ons the convective parameterization is better behaved with the extra u
pper-tropospheric resolution. A benefit to both approximations of the
additional vertical resolution is a reduction of the tropical temperat
ure bias compared to the NCEP reanalysis. The authors also show that t
he Eulerian approximations are prone to stationary grid-scale noise if
the vertical grid is not carefully defined. The semi-Lagrangian shows
no indication of stationary vertical-grid-scale noise. In addition, t
he Eulerian simulation exhibits significantly greater transient vertic
al-grid-scale noise than the semi-Lagrangian.