Dear Dr. Trenberth,

Thank you for your interest in our work.  You asked some interesting
questions concerning Figure 14 in our Global SST analysis paper.  Here
I tried to provide detailed answers.

My answers are illustrated by the supplement to Figure 14 (Fig14sup) 
which can be accessed through
   http://rainbow.ldgo.columbia.edu/~alexeyk/GlobSSTpaper/
If desirable, I can send hardcopy of these figures by airmail.

>In Fig 14 the plots look rather weird.  This is because of the straight
>lines bounding the negative anomalies in a and b.  Neither of the original
>fields look anything like that, e.g. as in Smith et al.  Also, the magnitudes
>are greatly reduced from the original fields.  Is this a consequence of going
>to a 5 degree grid?

Yes, the difference in appearance between plots in Smith et al. (1996)
and in our manuscript are mostly due to regridding to 5x5 grid.  On
the first page of Fig14sup the evolution of the May 1988 NCEP OI field
as it is being regridded to sparser grids is shown: while the first
panel presented on the original 1x1 grid is identical to what you see
in the last panel of Figure 5 (p. 1410) of Smith et al. (1996) (really
minor differences are due to the use of different plotting software
packages), zero line wiggles less and the magnitude of the equatorial
negative anomaly is reduced when data are averaged to 2x2 grid (second
panel).  These changes become much more pronounced when the same data
are presented on 5x5 grid, as on the third panel, which differs from
the Figure 14a of the manuscript (reproduced as the first panel on the
page 2 of Fig14sup) only by the contour interval.

The similar evolution for Smith et al. May 1988 field is shown on the
page 3 of Fig14sup.  However, while the third panel here is certainly
identical to Figure 14b from the manuscript, the first panel differs
from the last panel of Figure 7 (p. 1411) of Smith et al. a bit
stronger than differences in plotting software would warrant. The
explanation is that anomalies in Smith et al. are shown in respect to
NCEP OI normals of 1950-1979.  Smith et al. put to the public domain total
fields, however.  Since I had to compute anomalies from total fields
anyway, I chose to compute them in respect to 1951-1980 normals, to be
a bit more consistent with our analysis, which is based on GOSTA
anomalies from the normals of the latter period.  But, once again, the
zero line straightens up and equatorial anomalies reduce greatly when
the grid is changed to that of 5 degrees.

To summarize, I agree neither that "in Fig 14 the plots look
rather weird," nor that "the original fields look anything like that,
e.g. as in Smith et al.," since to my eye there is a great deal of
resemblance in the plots of our Figure 14 and Smith et al. plots,
reduced size of anomalies being a quantitative rather than a
qualitative difference and a consequence of averaging on a sparser
grid.  As far as the shape of a zero contour is concerned, it is
well-known not to be a robust feature of a contour plot.

>>>How did you do that (interpolation or averaging)?<<<

All fields are regridded on sparser grids by averaging, of course.


>>> If the plot is on a 5 dg grid, why are the gradients so strong in
the figure panels?<<<

I illustrate this on the example of NCEP OI field (Figure 14a or the
1st panel on the 2nd page of Fig14sup), where the gradients are
largest.  Since the anomaly in question is very narrow, its strongest
values are all inside two 5x5 box lines surrounding equator, and
basically all change happens between these lines and those immediately
following polewards (Fig14sup, page 2, panel 2 (colors)).  This sharp
change creates this large gradient (high concentration of contour
lines) on the panel 1 of the same page.  

Actually, page 1 of Fig14sup illustrates how in the process of
regridding the gradient of SST seems less and less uniform, being
shifted from the equator to a zero line.  Of course, this is a
distortion of physical reality, but, after all, nobody would argue
that 5x5 representation of the data is superior to that of 1x1 -- only
in the analysis of historical SST we have to settle with the former,
as we do not have enough data to afford the latter.

>>>I note that for the May 1988 case, the analysed values are extremely
deficient, even in the OSn case.<<<

Well, clearly the OSn field (Fig 14d of the manuscript or 1st panel
on page 5 of Fig14sup) is much better than that of OS (Fig 14c of
the manuscript or the last panel on page 4 of Fig14sup).  The panel
1 of the page 4 of Fig14sup (GOSTA observations) reminds how
difficult is the task which analyses (both ours and Smith et al.'s)
solve here, and how natural for the available observational data
the (wrong) standard OS solution (middle panel, page 4, Fig14sup)
is.  Now, at the stage of EOF calculation (for the period of
1951--1991) we replace GOSTA data by NCEP OI data (regridded on 5x5
grid) for the subperiod of 1982--1991 and obtain the results of OSn
(Figure 14d or the 1st panel of Figure 5 of Fig14sup).  It has now
structural resemblance to NCEP OI solution but the equatorial
negative anomaly is weaker than that of both NCEP OI and Smith et
al. solution.  The explanation is that the cold tongue dynamics is
much better resolved in the NCEP OI data than in GOSTA for the same
period.

When similarly to Smith et al. we estimate EOFs solely from NCEP
OI, we effectively further increase the weight of cold tongue
dynamics among all patterns and obtain OSn+ solution which is
presented on the second pattern of the page 5 of Fig14sup.  This
solution is in no way inferior to that of Smith et al (3rd panel on
page 3): the equatorial -0.8 contour of Smith et al is broken into
a few little contours -- but this is really a minor difference.
Both analyses are inferior to NCEP OI in terms of magnitude of the
equatorial anomaly, but obviously they achieve the maximum of what
the method can give when the data is so poorly sampled in such an
important place.

>>>At the buoy at 110W on the equator, the
anomaly in May 1988 was -4.1 dg C, see Trenberth and Branstator 1992 J.
Climate Feb 1992 p 162 and fig 3. This will change a bit depending on the
base period, but it indicates a problem.<<<

Well, the value of -4.1 C is not immediately interpretable here 
since the  buoy data are essentially pointwise measurements.  
The structure in question is extremely narrow as it 
becomes obvious  when the equatorial measurement is
supplemented by off-equatorial ones.  In May 1988 average total
temperatures on 110W were:
2N: 25.5
Eq: 21.3
2S: 23.9
5S: 26.6,
while of course climatology in this area does not have such a  sharp minimum
on equator. Consistently,  minimum values for equatorial anomalies on 
the page 1 of Fig14sup are much smaller than -4.1 C: 
-3.6 for 1x1 box, 
-3.3 for 2x2 box, 
-2.1 for 5x5 box. 
Of course we do not reproduce -2.1, we reproduce only -0.8 in the OSn+
analysis (with the RMS error conservative estimate of about 0.5 C, 
so that the difference is between 2 and 3 sigmas), but as I wrote above, 
this is about the best one can do with  poor data *and a few existing 
physical patterns which are not easy to distinguish*.

>>> In the dataset you produced is the result OS or OSn?  I guess we can check
by plotting the field but we have not done so yet.<<<

It is OS.  We consider OSn to be a "quick fix" and are not ready to recommend 
it for the usage.  I think it is obvious why we would not recommend to use OSn+
(patterns come solely from the period 1982--1995, which is *not* the most 
characteristic one in terms of historical SSTs), but temporal non-uniformity 
of the dataset used for the pattern estimation in OSn makes us uncomfortable:
it distorts energy distribution over patterns, and additional effort 
will require to recover consistency.  Also we view improvement on estimating 
the covariance structure of SST field as an important and separate project:
OSn is better than OS in some sense, but we'd like to be able to evaluate 
overall the quality of our covariance estimate, to find out if there are other 
glitches, if OSn is generally better than OS, what are other options
(in addition to OS ---> OSn replacement), etc. 

The reason why we consider our standard OS at it is to be an acceptable choice 
is obvious from Figure 13 of the manuscript: even in this "dangerous"
point of 125W and equator, the standard OS provides really great improvement
on raw GOSTA for the entire period of NCEP OI availability (and May 1988
is *not* an exception!)  Discussion of OSn is intended to display one of
existing problems and to demonstrate that there are ways to fix it.  We do
not have a self-consistent way to overcome it as of now, unfortunately.

If you have further questions, please do not hesitate to ask.

Regards,
Alexey