Dear E895 collaborators,
This is the wrapup report of the CSC for the K0s flow manuscript of
SUNYSB. It is lengthy, but it is only part of the material exchanged.
The
objective is to represent fairly the content and nature of these exchanges.
You
will see that there is a happy outcome. We believe that the present
version of
the manuscript is a good one and properly represents SUNYSB's excellent
work.
Our congratulations to them!
Jim Draper (for the CSC)
--------------------------------------------------
On CSC page http://vdgus1.mps.ohio-state.edu/~lisa/CSC/
is the following:
1. Draft dated 26 April, 1999 - this one had sidewards flow and elliptic
flow.
2. Draft dtaed 9 October 1999 - this one dropped the elliptic flow,
but put in
an impact parameter dependence.
This was followed by http://vdgus1.mps.ohio-state.edu/~lisa/CSC/Stage2/
titled
"Issues on Stage 2 draft of K0 Paper of 9 October" noting some of Roy's
points
and the CSC position on them.
Recall that a function of the CSC was to get the answers to questions
raised by
collaborators for questions deemed unanswered and significant. A concatenation
of most of the emails from 7-Nov-1999 to 4-Feb-00 is 70 typed pages
- too much
for here. But to be fair to all requires considerable detail in the
emails
from both sides. The material I include here comes only from >6-Nov
emails
between the CSC and SUNYSB, and it is all of them. It does exclude
however
many emails between members of the CSC during this period. Some of
this omitted
material expresses considerable anguish, on which I'll not elaborate.
The
included emails are not edited except to truncate material repeated
verbatim
from earlier emails and to delete some blank spaces. The former is
clearly
indicated by my repeated comments about truncation.
Note that there was introduction of material in the 9-Oct manuscript
about the
impact parameter dependence of sidewards flow, with rather unexpected
properties. CSC's position was that this is new physics and should
be in a
separate paper later. Otherwise the manuscript should be sent back
to the
full collaboration, having been considerably modified in physics content
since
review by all collaborators.
A major factor in the time line of communications was the observation
that the
7-Nov manuscript had more than twice as many K0s's as the previous
manuscript,
and a 16-Nov phone conversation that left the impression that most
of this
improvement was from retraining the neural net that detected the K0s's.
##############################
From: SMTP%"RLACEY@mail.chem.sunysb.edu" 7-NOV-1999 18:07:10.05
To: DRAPER
CC:
Subj: Response to CSC's Report
Hi to all:
We have decided to give the process one more
shot.. a modified manuscript can be found at:
http://nucwww.chem.sunysb.edu/papers/public/v0_e1b.ps
Our response to "outstanding issues" are outlined below.
>>>> So, that is the first decision that you must make-- to keep the
>>>b-dependence in the paper and send the paper back to the collaboration,
>>>or to drop it and keep the paper in CSC.
The Fig. showing the impact-param dependence of flow has been removed
along with all quantitative material related to impact-parameter dependence.
>> So far as CSC issues, it was generally agreed to drop most of the
>>ones relating only to v2. I have not done a full update of the web
page,
>>(http://vdgus1.mps.ohio-state.edu/~lisa/CSC/)
but have numbered the issues
>>(as reccommended by Jim), for easier reference.
>> From the first list of issues on that page:
>>> Issue 9 (phi_lab dependence of K0 efficiency) remains an open question,
>>as it affects v1 as well. The other issues on that list do not seem
relevant
>>to this paper as it now stands.
We are missing something here !! ... in our previous submission to
the CSC, we
sent the URL for a webpage with supporting materials.. One of these
pages,
shows/compares and discusses the sidewards flow in each of four quadrant's
for
the TPC for several impact-parameters. Within Stats. results from different
quadrant's were shown to be the same ... This being the case,
I am not sure we understand the "open" nature of Issue 9.
>>> From the second list of issues:
>>>> Issue 1 asks for the sideflow from 4 and 8 GeV data, just for
"internal
>>>consumption," not for this paper. This is just to further convince
that
>>>values are stable.
We are currently re-analyzing the data at 4 and 8 AGeV with improved
methods designed to improve stats and the like. We will make these
new results
available to CSC and the collaboration as soon as they are ready.
On the other hand, we belive that it is fair to say that the current
status of the 6 GeV analysis is strong enough to stand on its own:
(a) We have made considerable improvements in the statistical significance
of
the results by processing additional/"new" data.
(b) We have carried out independent analysis' for several different
portions of
the data (ie different ranges of runs) and have found the results to
be the
same (within stats).
(c) We have completely retrained the neural net and re-analyzed the
old data as
well as new data and found that they agree.
(d) We have performed the analysis for multiple impact parameter ranges
and
found a robust anti-flow signal in each case.
(e) We have determined the flow in each of four quadrant's in the TPC
for both
old and new data as well as for several impact-parameter ranges. They
all lead
to a consistent set of results..
ETC ... ETC...ETC...
We could go on but we are sure that the committee has gotten the picture
by
now.
>>>> Issue 2 (changing flow values with changing manuscript version)
had been
>>made a greendot, based on an examination of various manuscripts by
Mike.
Six months ago, we pointed out the fact that in an early version of
the
manuscript, there was a typo in one of the Figure captions. The record
clearly
demonstrates that issue 2 has been resolved.
>>The worry is that it appears that the sideflow has changed
>>*quite a bit* in this new draft, and so this issue comes back to
life. What
>>is happenning here?
>> In the April draft, it said that the antiflow signal
>>was -0.19 +/- 0.03 for b~4-7 fm. In the new one it is:
>b flow
>----- -----
<4 -0.144 +/- 0.02
5-8 -0.115 +/- 0.01
<8 -0.127 +/- 0.01
It would appear that the sideflow for b~4-7 has changed by almost
a factor of 2! ?
One of the committee members pointed out the if one overlays the
"S-curves", then the data lay roughly on top of one another. So the
large
difference in flow value must come from systematic error in the fitting.
Is this right? If so, then the error bars on the flow value should
increase,
yes?
------
Once again, in our earlier submission to CSC, we pointed to a webpage
on which
we compared the old results to the new results. Differences were pointed
out,
which by the way, is not a factor of 2. In addition, we pointed out
that within
error bars the experimental data had not changed. In the earlier analysis
we
did a cubic and a linear fit to the data. The cubic fit gave a better
Chi-square, apparently because of the point at ~ .6 (Ycm). With the
new and
statistically more significant data, we find no significant difference
between
a linear and a cubic fit... YES .. we should have included a larger
systematic
error (associated with the fitting procedure) in the earlier result.
>> Other comments to the new draft:
*>> Central and mid-central impact parameters should be defined when
>> first mentioned.
*** OK ... Done ... ***
>>>* Could you be more specific in the last sentence of reference 25?
This issue has already been addressed in several emails some 6 months
ago. I am not sure why it is back. Here we go again...... the net effect
of the
pion contamination is to reduce the proton flow. Weaker flow leads
to larger
reaction plane dispersion. Therefore, the influence of pion contamination
on
reaction plane determination "is accounted for in the dispersion correction
for
the reaction plane."
We should add here that several months ago, we discussed the fact that
this notion was tested and validated via a barrage of simulations.
>> Some of us feel that the lack of "turnover" in <px> at high rapidity
>> deserves some brief comment. It is different than what we are used
>> to from proton and fragment flow. This might be usefully examined
in
>> a v1 analysis.
I am somewhat amazed by this statement ... Flow values (reflected and
un-reflected) are plotted for ~ -.6 < Ycm <+.6. It is not clear
to me why there
should or should not be a turnover for this rapidity range. In fact,
a
comparison with Heng's plots for proton flow at 6 AGeV (for the same
rapidity
range) clearly shows that there is no turnover.
Finally, we are not sure how a v1 analysis might "usefully examine"
this perceived turnover.
>>> Wording and elaboration suggestions:
>>>(A) The first sentence of paragraph 2: Change "particles" to "kaons".
>>>Grammatically the present "particles" refers to the nucleons.
OK ... done
>>>(B) In the first complete paragraph on page 2 of the mscpt, ~fifth
>>>sentence, it would be useful to know what cuts were used before
>>>"These hypothesized Ks0's were then passed...".
As indicated in Fig. 1, invariant masses between 0.4 and 0.6 were considered.
>>(C) Could you clarify (at least for us) please:
>>About 0.6 of the way down the left hand column of page 2 there
>>is discussion of the inset in Fig. 1 about ctau. This x axis must
>>be distance from the decay vertex. And in converting the slope to
>>ctau, where does the value of gamma come from? Constant or what?
>>A priori or a posteriori?
In the manuscript, we explicitly point out that the decay-length distribution
is given in the rest frame of the K0 (ie the canonical lifetime plot).
Unless
we have missed something (and we may have), we think that a Lorentz
factor is
not inappropriate in the formula given.
>>(D) Line 22 of right hand column of page 2: The word "established"
>>would more appropriately be "assumed". Or maybe this sentence
>>constitutes the definition of "efficiency".
It is clearly stated in the manuscript that the determined distribution
for the
reaction plane should be flat... but it is NOT. Indeed we have ESTABLISHED
that
deviations from this uniformity or flatness is a direct result of deficiencies
in the acceptance of the TPC.
>>(F) Fig. 2: The hollow and solid stars are nearly indistinguishable
>>for the present size of Fig 2, which is about the needed size for
>>publication.
We have corrected this problem..
#########################
From: SMTP%"lisa@pacific.mps.ohio-state.edu" 11-NOV-1999 12:11:17.76
To: DRAPER
CC:
Subj: Re: Response to CSC's Report
Hello Roy and CSC,
As I am working through your response, I have some questions. It could
be that I am confused (imagine that!), so I invite the other CSC members
to
answer as well, if they have the information.
The two things I ask for below are:
1) the URL of the phi-dependence of the K0 efficiency.
2) Clarification on the (changing?) value of F.
I am finding lots of these emails to be kind of confusing. So please,
let's try to answer each other as directly as possible. E.g., for number
(1)
above, please just send me the URL for the K0 efficiency, not something
else.
On Sun, 7 Nov 1999 RLACEY@mail.chem.sunysb.edu wrote:
>> From the first list of issues on that page:
> >>> Issue 9 (phi_lab dependence of K0 efficiency) remains an open
question,
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
I have looked through all (I *think* all) of the email from you guys,
and have found elliptic flow in four quadrants.
But the question asks for the phi-dependence of the K0 efficiency.
Is
there a URL for that?
> >>>> Issue 2 (changing flow values with changing manuscript version)
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
Oh man... the baud rate is very low here....
I am talking here about the NEW draft. I am NOT talking about 6 months
ago. I do not think you sent me a webpage comparing the new and old
results,
where "new" means the draft of 7 November 1999. *** PLEASE CORRECT
ME, SOMEONE,
IF I AM WRONG ***
Factor of 2? 0.19/0.115 = 1.65. Or should I be comparing to b<8?
In
that case it is 0.19/0.127 = 1.50. So it is not factor of 2 (I did
say
"almost"), but it is 50% or 65% less now.
You say that within errorbars the data has not changed, but the flow
values sure have. (0.19 +/- 0.03) .ne. (0.115 +/- 0.01) and (0.19 +/-
0.03)
.ne. (0.127 +/- 0.01).
It seems to me that you should report larger errors in the CURRENT
draft. At least big enough to cover the old value of 0.19. So the errorbar
on
flow should be something at least ~0.07, yes? *** PLEASE CORRECT ME,
SOMEONE,
IF I AM WRONG ***
I notice that this issue is somewhat sidestepped in the draft of 7
November by simply not having a flow value at all. The CSC has not
discussed
it, but I guess this is one way out.
##############################
Note by JED: A lot of emails between CSC members followed, but they
are
omitted here, as noted above. On 19-Nov, Declan announced to the E895-l
that
I was to begin acting as chair of the CSC (membership unchanged). I
accepted
this with the understanding that I would keep written records of all
exchanges
by interacting only via emails. Experience showed me that the ratio
of light
to heat is much improved in this way. The other two members of CSC
concurred.
The wording of all emails below with "CSC" in the Subject line were
concurred
with by all CSC members.
##############################
From: UCDPHY::DRAPER 22-NOV-1999 21:27:45.59
To: SMTP%"RLACEY@mail.chem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@sseos.lbl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: CSC's response to 11/7/99 manuscript.
Hello Roy,
Here is what could be our last specifications from the CSC regarding
your K0s manuscript. In summary, we feel that SB has an excellent piece
of work
and that it is very near to publication. Nice going, guys! We feel
that all of
the matters discussed below can be readily remedied with relatively
little
consumption of time or effort. We all look forward to publication soon.
----------------------
CLARITY AND SIMILAR (involves only simple modification of wording):
(1) Some Ref numbers do not appear in sequence in the text.
(2) Last sentence of left hand column of page 1 says a measurement
has been
attributed, but it is actually the zero flow that has been attributed.
So
strike "A measurement which indicated" from the beginning of this sentence.
Similarly strike the comma after the closing paranthesis.
(3) Page 2, left column, line 10: Change "These hypothesized Ks0's
..."
to "All of these hypothesized Ks0's with invariant mass between 0.4
and
0.6 GeV/c**2 ...". This concerns our earlier requested definition of
the
cuts before the neural net.
(4) Page 2, left column, line 15: The word "respectively" has no meaning.
(5) Second paragraph of page 2, first sentence: The word "resulting"
says
literally but erroneoulsy that the embedded spectrum constitutes Fig.
1. So add
"from real data" after "for Ks0's".
(6) Second paragraph, left column, line 11; The term "true Ks0's" has
been
defined as embedded Ks0's, so change "true" there to "real".
(7) The next line, concerning Fig. 2, now shows much more clearly (by
adding
the subscript i under gamma) what was done, so this item is essentially
OK.
However, add the word "laboratory" before "distance" because the beginning
of
the sentence says "in the rest frame".
(8) Page 2, left column, line 20: This concerns the apparent deficit
below
ct~4 cm. Here it is not clear whether the efficiency has been evaluated
and used, but nevertheless fails below 4 cm, or not. The same kind
of
question arises in the next sentence where the region 7-11 cm is described.
This question of efficiency correction should be clarified. And it
should be stated in what bins of y and pt, or whatever, the efficiency
was evaluated during embedding.
(9) Last sentence in same paragraph: One guesses that this means that
~0.03
Ks-'s per event is the yield after correction for efficiency. But it
is an
important number, and it should be made quite unambiguous.
(10) Page 2, right column, line 11: The weight should include the negative
sign for negative rapidity, even though only a small range of negative
rapidity is in Fig. 3.
(11) Last sentence in Ref. 25 (yes again, as the problem was not sufficiently
spelled out earlier -- please tell us if you think the following reasoning
is
incorrect): The correction for the dispersion of the reaction plane
depends on
the number of tracks N. The correction for N tracks with a particular
flow
superposed on a random phi distribution is different from that for
the same N
but with two sets of tracks (pions and protons) with different flows,
including
sign, superposed on a random phi distribution. We believe that the
correction
in question is small, but the logic should be clear. Therefore the
wording of
this last sentence should be something like "The small influence on
Fig. 3
caused by this pion contamination is corrected to appropriate accuracy
by the
dispersion correction for the reaction plane".
(12) In the earlier version there was a brief mention of the estimated
contribution of the antikaons. We believe that this should be used
in the
present mansuscript. It is not addressed at all, and is important logically.
------------------------------
MORE THAN MODIFICATION OF WORDING (but solutions are readily available):
(13) There has been the continuing situation of CSC's asking for the
phi
dependence of the K0 efficiency. It is only peripherally answered by
SB's
already having broken up the <px> in four halves of the TPC. But
this is not
what CSC requested. Furthermore the overlapping of quarters of the
TPC makes
it more difficult to evaluate. It would answer this CSC question to:
(1) Show
us (but not in the mscpt) the efficiency in eight bins of 45 degrees
(lab
frame). (2) Show us the <px> for these same phi bins. (3) Do this
for both
the present analysis and the previous analysis, as the efficiency has
evidently
changed considerably. We believe that this information is readily available
and
easily supplied. We will not be troubled by the statistical uncertainty
that
goes with each such phi bin.
(14) The question of an F number and its uncertainty: The points on
the <px>
plots have been stable within statistical uncertainties as seen in
SB's web
page overlaying the two plots of <px> vs y0. This is plot D of SB's
web
10/3/99. We conclude that SB should quote an F number using only a
linear fit
for all points. This should be done for the new analysis, quoting its
uncertainty from the statistical error bars - and saying that the error
is
statistical. Then repeat F and uncertainty for the older analysis.
Note regarding systematic errors: The significance of a difference
between these two F's depends on how independent are the analyses,
and the
degree to which they are different. So trying to assign a meaningful
systematic uncertainty is not appropriate here. The main older difference
in
the F results are the question of fitting with a linear or a cubic
equation.
Fitting three (or four?) points with a cubic has its problems anyhow.
In this
vein, clarify in the mscpt whether the point (zero,zero) is forced.
Differences
between F for linear and for cubic fits are a different question and
not
relevant here.
(15) Page 2, right column, line 9: We believe that the sentence refering
to
b<~4 and b~ 5 to 8 should be struck, as no such data are to be shown
to clarify
it. If SB insists on retaining this, we note that any response by a
referee
wanting to see further results apportioned in impact parameter would
have to be
declined, else the manuscript would be returned to the full collaboration.
This is in keeping with our earlier discussions.
(16) We understand that SB retrained the neural net. A quick check
of Fig. 1
for 4/26/99 and 11/7/99 manuscripts indicates more than two times as
much data
now. In his phone conversation with you, Mike understood that a large
increase
in efficiency was achieved by retraining the neural net. This is an
important
result, and we want more quantitative information about it. For example,
such
a large improvement is puzzling, as examination of Marvin's work with
the
neural net indicates that such a large increase in yield, if produced
by
widening the criteria, would be accompanied by a much larger increase
in
relative background than is seen in the present Fig. 1. Another part
of this
question is how this improvement in efficiency is distributed in pt,
eta (or
rapidity) and phi.
(17) Panels (a) and (b) of Fig. 2 in the new MS look the same as the
insets
(a) and (b) of the old Fig. 1. With the new and much improved stats,
that Fig.
ought to be updated to reflect the actual data used in the rest of
the
analysis.
-----------------------------
Best regards,
Jim for the CSC
####################################
From: UCDPHY::DRAPER 23-NOV-1999 11:06:20.26
To: SMTP%"RLACEY@mail.chem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@sseos.lbl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: RE: Call
Hi Roy,
This is in response to your email, below. There are a couple of
reasons why I plan to communicate on CSCish matters by email only:
(1) A major reason is that my brain processes and retains information
much
better by sight than by sound.
(2) The brief turnaround time provides for more considered responses
by both
me and the other party. This actually speeds the communication process
where I
am involved by minimizing the tangents and false starts.
You might be concerned that I will not be attentive to emails and may
delay too much in responding. I am not leaving town in the next few
weeks, and
I'll be using the computer most of the time from morning to late evening.
So
there should be no problem.
When I was asked to take over as CSC chair, I made it plain to the
CSC that I would operate in this way. I expect that it will work well.
I believe that progress can be quick and productive.
I'm sending a copy of this to the other CSC folks for initialization.
I do plan to keep the other CSC folks informed of most matters. But
if you
want some one-on-one emails, that is OK with me.
-Jim
#################################
From: SMTP%"RLACEY@mail.chem.sunysb.edu" 30-NOV-1999 20:39:40.88
To: DRAPER
CC:
Subj: Re CSC Report
To: jedraper@UCDAVIS.EDU, lisa@mps.ohio-state.edu, panitkin@sseos.lbl.gov,
keane@ksuvxd.kent.edu
CC: RLACEY@mail.chem.sunysb.edu
Hi Jim:
Enclosed is a response to the questions/issues raised by
CSC. A revised manuscript can be found at:
http://nucwww.chem.sunysb.edu/papers/public/v0_e1b.ps
Thanks
RL
> CLARITY AND SIMILAR (involves only simple modification of wording):
> (1) Some Ref numbers do not appear in sequence in the text.
*** Ref 12 was previously out of sequence ... the sequence is now correct
> (2) Last sentence of left hand column of page 1 says a measurement
has been
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
*** The sentence has been modified to add clarity ****
> (3) Page 2, left column, line 10: Change "These hypothesized Ks0's
..."
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
******* sentence modified to reflect Inv_mass selection *****
> (4) Page 2, left column, line 15: The word "respectively" has no
meaning.
****** Sentence has been modified to make more clear the
meaning of respectively *****
> (5) Second paragraph of page 2, first sentence: The word "resulting"
says
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
****** Modification made to avoid potential ambiguity ********
> (6) Second paragraph, left column, line 11; The term "true Ks0's"
has been
> defined as embedded Ks0's, so change "true" there to "real".
********** OK ... done ..
> (7) The next line, concerning Fig. 2, now shows much more clearly
(by adding
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
******************** OK ... done ...
> (8) Page 2, left column, line 20: This concerns the apparent deficit
below
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** To avoid ambiguity, we now state that the decay length distribution
is
not efficiency corrected. In addition, we point to the fact that a
fit was
performed on the distribution only for a region in which the detection
efficiency was determined to be flat. *** We believe that this approach
is
quite sufficient for the peripheral point being made .. namely, the
neural net
does seperate "true" K0_s's from the combinatoric background. Efficiency
as a
function of ctau was obtained by using K0's from RQMD.
> (9) Last sentence in same paragraph: One guesses that this means
that ~0.03
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** We have modified the sentence to make more transparent, it's
intent.. ie)
it is a raw yield and therefore does not include any efficiency correction.
> (10) Page 2, right column, line 11: The weight should include the
negative
> sign for negative rapidity, even though only a small range of negative
> rapidity is in Fig. 3.
**** The current weight DoES include the negative sign for negative
rapidity
particles ***
> (11) Last sentence in Ref. 25 (yes again, as the problem was not
sufficiently
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** We have modified this sentence to make it clear that the dispersion
correction accounts for the "small" influence the pion contamination
may have
on the extracted flow ******
> (12) In the earlier version there was a brief mention of the estimated
> contribution of the antikaons. We believe that this should be used
in the
**** Recently, we were asked to remove it ... we have now put it back
in its
original form.
> (13) There has been the continuing situation of CSC's asking for
the phi
> dependence of the K0 efficiency. It is only peripherally answered
by SB's
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
****** The phi dependence of the K0 efficiency is not germaine to the
substance of the manuscript. Nonetheless the requested plots can be
found at: http://nucwww.chem.sunysb.edu/e895/protect/v0flow/csc_req.html
It
should be noted that the results presented do not take account of
reaction-plane dispersion, corrections for combinatoric background
and the like
...
We have stated at several meetings and to the GO3 many months ago that
a
relatively small correction (for detection efficiency) is effected
by way of
the pt distributions. Namely, for each rapidity selection we generate
the pt
dependence of the flow. Then, we genenerate the pt distributions (for
the same
rapidity selections) and correct them for possible losses. This correction
is
ubtained by comparing the input (flat) and output pt distributions
from
detailed simulations discussed earlier. Subsequent to this correction,
the flow
for each rapidity is determined by folding the pt distribution with
the pt
dependence of the flow for that rapidity. Please note the similarity
to the
method used and discussed (extensively) for the proton Elliptic flow
manuscript.
Given the currently available space in the manuscript and the improved
statistical significance of the data, we have now included a brief
description
of this procedure.
> (14) The question of an F number and its uncertainty: The points
on the <px>
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
****** This was done in our first report to the CSC ... several weeks
ago.
We repeat here the URL we posted in this earlier mail.
http://nucwww.chem.sunysb.edu/e895/protect/v0flow/v0flow_b_dep.html
> Note regarding systematic errors: The significance of a difference
between
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
*** we have removed explicit values of impact parameter. However, we
do intend
to submit the impact-parameter results to the collaboration in the
very
near future. We do not anticipate a problem here ***************
> (16) We understand that SB retrained the neural net. A quick check
of Fig. 1
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** The apparent improvement in the RAW efficiency has resulted from
improved
training of the neural net + the processing of additional data. We
are not sure
why such an improvement is puzzling ??? A "before and after comparison"
of the
requested distributions can be viewed at:
http://nucwww.chem.sunysb.edu/e895/protect/v0flow/csc_req.html
> (17) Panels (a) and (b) of Fig. 2 in the new MS look the same as
the insets
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** OK ... done *****
###################################
From: UCDPHY::DRAPER 8-DEC-1999 12:42:48.87
To: SMTP%"RLACEY@mail.chem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@sseos.lbl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: CSC's response to 11/30/99 manuscript and email.
Hi Roy,
Here we (CSC) address your 11/30 email responding to our 11/22 report
to you. The numbering is the same as in your response and our report.
The
manuscript being addressed is dated 11/30/99. First, is an overview
of items
(1)-(17). Then two other easy-to-handle points.
Item (1): Ref numbers are still out of order. Examples are 17, 18,
22, 25.
Items (2) through (9) are now OK.
Item (10) involving negative sign of weight for negative y: Line 14,
right
column, page 2, taken literally implies that the sign is independent
of y.
Items (11) and (12) are now OK.
Item (13): The first piece of this item in the CSC report is "Show
us (but not
in the mscpt) the efficiency in eight bins ...". Although your response
says,
"the requested plots can be found at .../v0flow/csc_req.html", nevertheless
your response does not show the efficiency. However it does take care
of the
other two pieces of this item (13). So please also plot the efficiency,
as
requested.
Items (14) and (15) have been addressed satisfactorily.
++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Item (16), concerning more than a factor of two increase in your recent
yield,
has not been adequately addressed. In part, we stated this item as:
"... a
large increase in efficiency was achieved by retraining the neural
net. This is
an important result, and we want more quantitative information about
it. For
example, such a large improvement is puzzling, as examination of Marvin's
work
with the neural net indicates that such a large increase in yield,
if produced
by widening the criteria, would be accompanied by a much larger increase
in
relative background than is seen in the present Fig. 1. Another part
of this
question is how this improvement in efficiency is distributed in pt,
eta (or
rapidity) and phi."
Your complete reply was: "***** The apparent improvement in the RAW
efficiency has resulted from improved training of the neural net +
the
processing of additional data. We are not sure why such an improvement
is
puzzling ??? A "before and after comparison" of the requested distributions
can be viewed at:
http://nucwww.chem.sunysb.edu/e895/protect/v0flow/csc_req.html"
We CSC felt that it was clear in our request why we were puzzled, but
here is a more quantitative reiteration of our need, which we want
to be
answered. It is stated as below in order to separate the part in your
reply
about additional data from the part about retraining the neural net.
*****
If the portion of the increase in yield that resulted from retraining
the neural net was more than 15%, please comply with the following
portion of
our original request: - viz to explain: "... puzzling, as examination
of
Marvin's work with the neural net indicates that such a large increase
in
yield, if produced by widening the criteria, would be accompanied by
a much
larger increase in relative background than is seen in the present
Fig. 1".
To summarize, if this NN increase in efficiency was more than 15%,
please tell us in significant detail: (a) the size of the increase
in
efficiency, and (b) what produced this increase.
*****
The second part of our need was "Another part of this question is how
this
improvement in efficiency is distributed in pt, eta (or rapidity) and
phi". It
might be argued that your four web pages of plots, viz that part of
the above
web address which you labeled "Item 16: NN K0 kinematic distributions",
answers
this second part. Although this arguement would be only partially valid,
we
will postpone, or possibly drop, further pursuance of this second part
until we
understand the answer to the question reiterated in the paragraph just
above.
+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
Item (17) concerned panels (a) and (b) of Fig. 2 needing to be updated
with
all of the new data. You did this for panel (a), but it is not apparent
what
was done to panel (b). In panel (b) the upper curve looks identical
to the
older plot in its channel-to-channel shape and magnitude. The lower
curve
looks like a new shape on the old scale. And this new shape seems not
to
have the reduction in statistical fluctuations that would be expected
from
more than double the yield.
Next are two other items in the new portions of the 11/30/99 manuscript:
(a) The 24th line from the bottom of the left column of page 2 should
say
"Fig. 2a", not "Fig. 2". This would also match your later allusion
to Fig. 2b.
(b) Line 18 of the left column of page 3 concerns the slope of <px>
vs
y_reduced. Replace "this data" by "these data". The word "value" should
be
replaced by "slope", and "MeV" should be replaced by "MeV/c".
Regards,
Jim (for CSC)
############################
From: SMTP%"RLACEY@sbchem.sunysb.edu" 9-DEC-1999 08:48:42.79
To: DRAPER
CC:
Subj: Re: CSC's response to 11/30/99 manuscript and email.
CC: RLACEY@mail.chem.sunysb.edu, lisa@pacific.mps.ohio-state.edu,
panitkin@sseos.lbl.gov, keane@kent.edu
HI:
Enclosed is our response to the recent CSC report ... We have made
modifications to the manuscript where approrpriate. The resulting version
can
be found at:
http://nucwww.chem.sunysb.edu/papers/public/v0_e1b.ps
> Item (1): Ref numbers are still out of order. Examples are 17, 18,
22, 25.
***** The Refs. have been modified to reflect the correct order
> Item (10) involving negative sign of weight for negative y: Line
14, right
> column, page 2, taken literally implies that the sign is independent
of y.
****** I do not agree with this ... in the manuscript, it is clearly
stated
that " The weight w(y) is assigned a value <p_x>/<p_t>". Since
we all know that
<p_x> is negative for negative rapidities and positive for positive
rapidities,
I do not see any ambiguties in the current description of the weight.
> Item (13): The first piece of this item in the CSC report is "Show
us (but not
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
*** It appears that we continue to oscillate on this issue ...
We DO NOT determine the flow as a function of PHI and then correct
it for
efficiency. We have clearly stated how we address the issue of efficiency.
Namely, for each rapidity slice, we determine the p_t dependence of
the flow as
well as the p_t distribution. We then correct the p_t distribution
for
efficiency by comparing input and output p_t distribution from simulations.
The efficiency corrected p_t distributions are then folded with the
p_t
dependence of the flow to obtain a mean value at each rapidity.....
we have
explained this at several meetings and most recently included verbage
in the
manuscript which not only describes the procedure but also gives a
quantitative
assesment of its outcome.
While the procedure employed (by us) DOES NOT require the efficiency
as a function of PHI. We have determined and provided the
flow for the 8 Phi slices requested by CSC. They clearly show
an essentially flat dependence on I have repeatedly asked for an explanation
for the reason and or reasons why we need to address the
efficiency as a function of PHI ... given that there is no "clear and
present
danger" identified in the p_t method we have used. Suffice to say,
I have not
recieved a response yet. The slow simulations required for efficiency
correction, require inordinate amounts of time. This is very well known.
The
PHI efficiency corrections which are being requested would require
large
amounts of simulation time.. therefore, we believe that there should
be
reasonable justification within the context of the manuscript for doing
them.
Naturally we believe that the information provided thus far is sufficient
for
clarifying the role of efficiency.
It is my hope that the current oscillations on this issue can dampen
sufficiently to allow convergence.
> Item (16), concerning more than a factor of two increase in your
recent yield,
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
We have benifitted from a factor of 2.6 increase in the raw K0_s yield.
1) A factor of ~ 2 was obtained from additional data processing.
~ 50% more data.
2) A 30 % improvement (new compared to old) was obtained via extensive
additional training of the NN (combination of old (but reprocessed)
and new
data leads to ~ 60% gain ). That is, we increased the training set
for the NN.
It is unclear to us why an improvement in the NNs V0 detection
capeability appears to be "puzzling" to the CSC. The NN's ability to
seperate
"true" V0's from the combinatoric background clearly depends on how
well it is
trained. It is also well known that the training procedure is an iterative
process which depends on the training set. Extension of the training
set very
often results in a "smarter" NN with obvious consequences for the RAW
V0 yield.
> The second part of our need was "Another part of this question is
how this
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
The CSC asked us to address the question of how the improvement in
efficiency is distributed in pt phi etc...
We have provided the CSC with several before (earlier NN training)
and
after (recent training) distributions (mass_inv, phi pt etc) so that
the CSC
can judge for themselves as to wheather or not the improvements in
efficiency
has impacted these distributions. It is very clear that our improved
training
of the NN does not lead to any glaring change in these distributions
ie. this
improved training appears to non-selective (on pt etc..) .... as it
should be.
> Item (17) concerned panels (a) and (b) of Fig. 2 needing to be updated
with
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
Panel (b) of fig 2 shows input (top) and output (bottom) m_inv
distributions for the combinatoric background, with the explicit intent
of
showing that the NN does not "create" V0s. I trust that we have conveyed
its
intent in the manuscript. (1) Barring statistical fluctations, the
shape of the
m_inv distribution of the combinatoric background does not depend on
the number
of events... we believe that the input distribution shown in Fig. 2b
has
sufficient stats for the perpose intended. (2) The lower curve in Fig.
2b is
the resulting m_inv distribution after passage of the input through
the NN. A
well trained NN should reject combinatoric background since this is
one of the
things we train it to do. The output distribution in Fig. 2b clearly
shows
that this is what the NN does. The fact that there appears to be more
statistical fluctuation in the output distribution is GOOD NEWS since
be
believe that we have a better trained NN. That is, we now have a NN
which does
a better job at finding v0s as well as rejecting combinatoric background.
> Next are two other items in the new portions of the 11/30/99 manuscript:
> (a) The 24th line from the bottom of the left column of page 2 should
say...
Ok ... Fig. 2 changed to Fig. 2a
> (b) Line 18 of the left column of page 3 concerns the slope of <px>
vs
> y_reduced. Replace "this data" by "these data". The word "value"
should be
> replaced by "slope", and "MeV" should be replaced by "MeV/c".
Ok ... done ....
Please note that I have followed a suggestion from Declan and have
modified both the placement and verbage of the discussion on the role
of the
anti-kaon.
Thanks
RL
#####################################
From: UCDPHY::DRAPER 14-DEC-1999 16:31:12.08
To: SMTP%"RLACEY@mail.chem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@sseos.lbl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: CSC response to SB's 12/9 mscpt.
Hi Roy,
This is the CSC response to your Ks0 sideward flow manuscript and email
of 12/9/99. The points here are in the same order as in your email.
The item
numbers are those of our 11/22/99 email. Please search on **** for
four short
items still needing attention.
(A) We are happy that you corrected the ordering of the references.
(B) Regarding our former item (10) about the sign of the weight at
negative
rapidity, you reply that the factor <px> obtained from a first pass
in the
iteration is sufficient fulfillment of this requirement. It can be
argued that
the direct route using the sign of rapidity, rather than the indirect
route
using the sign of <px> from a first pass, is conceptually better.
However,
since it is already done the latter way, and since it does not have
a
significant impact on the present results (if not the logic), we drop
this
point.
(C) This is item (13) concerning the phi dependence of efficiency.
We agree
with you that there has been much exchange of words about this, but
it has been
with little illumination of the answer to the question being asked.
In looking
back at the SUNYSB answers to this question, we are hard put to find
any of the
words phi, angl.. or azimu.. - i.e. there is seen no response to the
idea of
the question. It should come as no surprise, judging from our previous
correspondance on this matter, that we do not mean the efficiency of
finding
flow, but rather the efficiency of finding Ks0 particles. **** In the
present
situation, it would be sufficient if you show us a plot of the number
of
"hypothesized Ks0's with 0.4< m_inv< 0.6..." (as defined in your
12/9/99
manuscript, line 18 of left column of page 2) fed to the neural net
as a
function of phi_lab in nonoverlapping 45 degree bins. Surely you must
have this
information available. We are not asking for any new analysis here
- only for
a plot of numbers which we expect that you have at hand.
(D) Item (16) concerns the factor of 2.6 increase in the number of
Ks0's in
your manuscript starting five weeks ago. Below is the last part of
our most
recent statement of this question followed by your entire response,
all
between the &&&&&& marks.
&&&&&&&&&&&&&&&
> To summarize, if this NN increase in efficiency was more than 15%,
please tell
> us in significant detail: (a) the size of the increase in efficiency,
and (b)
> what produced this increase.
We have benifitted from a factor of 2.6 increase in the raw K0_s yield.
1) A factor of ~ 2 was obtained from additional data processing. ~
50% more
data. ((****CSC says the "50%" is puzzling - do you mean twice as many
events
as in the the old data?)).
2) A 30 % improvement (new compared to old) was obtained via extensive
additional training of the NN (combination of old (but reprocessed)
and new
data leads to ~ 60% gain ). That is, we increased the training set
for the NN.
It is unclear to us why an improvement in the NNs V0 detection
capeability appears to be "puzzling" to the CSC. The NN's ability to
seperate
"true" V0's from the combinatoric background clearly depends on how
well it is
trained. It is also well known that the training procedure is an iterative
process which depends on the training set. Extension of the training
set very
often results in a "smarter" NN with obvious consequences for the RAW
V0 yield.
&&&&&&&&&&&&
CSC makes three points here:
(a) This is the first time you have given us any quantitative information
concerning how much of the improvement came from the additional data
and how
much came from matters relating directly to the neural net.
(b) This is the second time that you have written "It is unclear to
us why an
improvement in the NNs V0 detection capeability appears to be "puzzling"
to the
CSC". You might grant that the wording of our original item (16) clearly
explained our puzzlement in the second half of the sentence whose first
half
contained a bit different wording from yours, above. This granting
is
especially reasonable given the apparent possibility that the retraining
of the
NN resulted in a factor ~2 in improvement in yield, based on the information
given to us.
(c) You still have not said whether you changed the size of the parameter
whose
value determines what is assigned as a Ks0.
**** The CSC conclusion from this is that we want a quantitative answer
to (c), just above. The answer to (c) will orient and assist our understanding
of the significance of this 30%.
(E) This concerns item (17). In this we requested that you update Fig.
2
(including 2b) to include all of the data in the rest of the manuscript.
It is
not clear what has been done, as the top curve of 2b looks *identical*,
channel
by channel, to the corresponding curve in the 11/22/99 mscpt (before
this
request). So a significant question is - what have you done? It is
not at all
clear from your description, and we should not guess. Possibly you
have fed
exactly the same combinatoric (as you used for 11/22) events (upper
curve of
2b) to the newly trained neural net, and the result is the lower curve.
Is
this so? Is the relation between the upper curve and the lower curve
of 2b
quantitative now? Then you say: "The fact that there appears to be
more
statistical fluctuation in the output distribution is GOOD NEWS since
be
believe that we have a better trained NN. That is, we now have a NN
which does
a better job at finding v0s as well as rejecting combinatoric background".
This is a puzzle. You said that the retrained neural net improved the
yield by
30%. Are you saying that the factor sqrt(1.3)=1.14 accounts for the
GOOD NEWS?
**** Please explain to us what you did here, including at least the
answer to
the four question marks in this paragraph.
-Jim (for CSC)
###################
From: UCDPHY::DRAPER 17-DEC-1999 22:44:45.71
To: SMTP%"RLACEY@mail.chem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@sseos.lbl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: Declan's recent email
Hi Roy,
As you know, Declan recently sent a message to the CSC interpreting
your phone conversation with him. We appreciate your attempt to clarify
some
of the issues. Nevertheless, please email to the CSC your own complete
reply
to all of the questions in our email to you on 12/14. In this way we
have the
complete story directly from you.
-Jim for CSC
#########################
From: SMTP%"RLACEY@sbchem.sunysb.edu" 21-DEC-1999 19:22:13.05
To: DRAPER
Subj: Re: CSC response to SB's 12/9 mscpt.
CC: RLACEY@mail.chem.sunysb.edu, lisa@pacific.mps.ohio-state.edu,
panitkin@sseos.lbl.gov, keane@kent.edu
Hi Jim et al.
As per your request, a response follows below ...
> (C) This is item (13) concerning the phi dependence of efficiency.
We agree
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
***** We do not have the efficiency for the detection of K0s as a function
of
phi and can not get to them without extensive simulations (as discussed
before). The specific plot requested is not readily available.
Once again, I want to emphasize here that the probability of finding
the reaction plane has been made uniform vs Phi_Lab. Therefore (just
as for a
Spectrometer measurements of flow) we do not need to know the probability
for
K0s vs. Phi_Lab. We do however get the K0 efficiency as a function
of pt (with
acceptance or integration over Phi_Lab). The methods/ideas of the latter
has
been explained in the manuscript as posted, as well as in the published
Pinkenburg PRL.
I have have also included the following paragraph which Declan sent
to
you, which I concur with.
First, Roy says that this info is not readily available for the dataset
in question. Moreover, if I understand correctly, the uniformity of
the phi
dependence of efficiency for finding Ks0 (or possible lack thereof)
is not
relevant to <Px> for Ks0. As long as the RP azimuth is correctly
tagged
event-by-event, all that is required of the Ks0 detection is to correctly
handle the (Pt, y) dependence. Thus, experiments like E877 can determine
proton <Px> in a spectrometer with zero Phi_lab acceptance outside
a single
spike, because an entirely separate subsystem does the RP tagging.
> (D) Item (16) concerns the factor of 2.6 increase in the number of
Ks0's in
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
YES .. Old Data + New Data gives gives twice as many events as in the
old data.
> 2) A 30 % improvement (new compared to old) was obtained via extensive
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
******* Your question misses the point as I will try to explain again..
(1) One has trained the NN anew.
(2) One does NOT use the NN parameter as the final criteria for assigning
a
K0s! Instead, one initially sets the NN parameter (eg. at the same
value as
that obtained from the former training set), and then generates a quality
plot
of the invariant mass (as shown on the web). From this plot, one gets
the peak
area and subtracts the "relatively smooth" background. The NN parameter
choice
that gives the best signal combined with signal-to-background is then
used.
Please note that it is this quality consideration (as exemplified by
the plots
on the web) that indicate improvement in the NN training. The raw magnitude
of
the NN parameter carries no real information after one retrains the
NN.
Orientation of one's understanding of the significance of the 30% can
thus be judged by (a) the number of K0s that appear in the peak after
background subtraction and (2) by the signal-to-background ratio. Plots
with
this information were sent to the CSC (ie webpage URL).
> (E) This concerns item (17). In this we requested that you update
Fig. 2
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
Apparently, we were not clear enough last time ...
Yes, we have fed the same combinatoric background since this background
is indipendent of data set.
Yes, this combinatoric background was "fed" to the NN and the lower
curve is the result.
>Is the relation between the upper curve and the lower curve of 2b
quantitative now?
Yes .. it is.
> Then you say: "The fact that there appears to be more
> statistical fluctuation in the output distribution is GOOD NEWS since
be
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
Here, I try to explain the "Good News" again. The NN training has
increased the fraction of accepted K0s by ~30% (these are from real
events).
The NN training has reduced the number of background K0s that came
from an
input from mixed events (which has nothing to do with 30% or sqrt(1.3).
The
number of combinatoric background which passed the NN is now less so
the lower
curve in 2b is now reduced and hence, its statistical fluctations have
increased.
Hope these new clarifications help.
RL
#############################
From: SMTP%"RLACEY@sbchem.sunysb.edu" 27-DEC-1999 17:53:46.28
To: DRAPER
Subj: Re: Declan's recent email
Hi Jim:
Here's wishing you a happy New Year !!
Several days ago I sent you and the CSC several responces
to qestions raised .... concerning the K0 manuscript.
It would be greatly appreciated if we can get a responce from you
ASAP.
Thanks
RL
#########################
From: UCDPHY::DRAPER 28-DEC-1999 14:39:56.18
To: SMTP%"RLACEY@sbchem.sunysb.edu"
Subj: Just returned from travel.
Hi Roy,
I just returned yesterday from travel. Happy New Year to you too!
Because of the season, it will take a few days to get a response to
you from
the CSC. Hang in there!
Jim
###########################
From: UCDPHY::DRAPER 31-DEC-1999 11:38:50.79
To: SMTP%"RLACEY@sbchem.sunysb.edu"
CC: SMTP%"lisa@pacific.mps.ohio-state.edu",SMTP%"panitkin@cher.star.bnl.gov",
SMTP%"keane@kent.edu",DRAPER
Subj: Hopefully final response from CSC.
Hi Roy,
You're going to believe this! Since your phone call, and 11 hours
after my last checking of email, I received the communique needed for
this
response. It was delayed by Y2k-type computer problems.
This is intended as a friendly letter from collaborators, but for the
record we also must note yet again the vagueness of some of your replies
to our
requests for information. This is the sort of thing that has so delayed
resolution of the prescribed business of this CSC.
Our last communique said:
**** In the present situation, it would be sufficient if you show us
a plot of
the number of "hypothesized Ks0's with 0.4< m_inv< 0.6..." (as
defined in your
12/9/99 manuscript, line 18 of left column of page 2) fed to the neural
net as
a function of phi_lab in nonoverlapping 45 degree bins. Surely you
must have
this information available. We are not asking for any new analysis
here -
only for a plot of numbers which we expect that you have at hand.
-----------------------
You replied:
***** We do not have the efficiency for the detection of K0s as a function
of
phi and can not get to them without extensive simulations (as discussed
before). The specific plot requested is not readily available.
----------------------
CSC now says:
Again we get a specific answer about the efficiency, for which we are
NOT asking, followed by "not readily available" concerning what we
DID ask for.
The word "readily" is vague and uninformative. But we have been outlasted.
---------------------------
***********************************
Note by JED (5-Jan-00): The CSC recognizes that the above efficiency
for
detecting K0s's from their daughters will have a strong dependence
on lab phi.
But a significant question is how different is it from this dependence
for
single particles with which we are familiar. Hence the initial request
for
this information from a collaborator was deemed significant. You have
seen
that the request was modified, but not changed in principal, as information
came to us.
***********************************
Regarding the factor 2.6 increase in K0s yield, we most recently said:
...
(c) You still have not said whether you changed the size of the parameter
whose
value determines what is assigned as a Ks0.
**** The CSC conclusion from this is that we want a quantitative answer
to (c), just above.
-------------------------------
The only answer to this questions is hinted at, but not defined, by
the
phrase between &&& marks below in your reply - namely you
said:
******* Your question misses the point as I will try to explain again..
(1) One has trained the NN anew.
(2) One does NOT use the NN parameter as the final criteria for
assigning a K0s!
Instead, one initially sets the NN parameter (&&& eg. at
the same value as that
obtained from the former training set &&&), and then generates
a quality plot
of the invariant mass (as shown on the web). Etc, etc.
CSC notes that "eg." usually means "for example" and not "this is
what we did". So again the CSC has been outlasted.
********************************
Note by JED (6-Jan-2000): This last statement is ameliorated by the
last email
(from Roy, 5-Jan-00 20:18) in this report.
********************************
In conclusion, the CSC considers that the most important parts of the
most important questions about your most recent manuscript, dated 12/9/99,
are
now sufficiently resolved. Overall there have been substantial modifications/
improvements of the manuscript compared to the version with which the
CSC
commenced.
And we recommend that the manuscript should be submitted to PRL.
We regret the length of time elapsed from the formation of this CSC
until the present recommendation for publication. We believe that the
time
spent by CSC could not have been significantly shortened under the
afore
mentioned circumstances. But at least it's a happy ending.
Congratulations to all at Stoney Brook on an excellent piece of work!
-Jim for the CSC
##############################
From: SMTP%"RLACEY@sbchem.sunysb.edu" 5-JAN-2000 20:18:54.27
To: DRAPER
CC:
Subj: Re: An eensy loose end
CC: Roy Lacey <RLACEY@mail.chem.sunysb.edu>,
Jim Draper <DRAPER@ucdphy.ucdavis.edu>,
Mike Lisa <lisa@mps.ohio-state.edu>,
Sergei Panitkin <panitkin@sseos.lbl.gov>
Hi Declan:
Thank you for your mail requesting further clarification... I do
hope the CSC will be able to issue a report ASAP.
The answer to your question is NO, the value of the neural net (NN)
parameter did not change ...
The point I have tried to make on several occasions now is that a
comparison of the actual values of the neural net parameter is
irrelevant once the NN is retrained. Let me try again:
1) One trains the NN as best as one can via an iterative process ..
2) When one thinks that the NN has been sufficiently trained and
presumably "V0 smart", one begins a process which tries to maximise
the
number of V0s while minimizing the background to an acceptable level...
ie) one performs a search for the best peak to background with the
best
raw V0 yield. This proces obviously includes varying the NN parameter.
If one retrains the NN their is no apriori reason why the NN parameter
should or should not be the same. It is the search for the the best
peak
to background with the largest number of raw V0s which ultimately
determines the value of the NN parameter. Thus, a comparison of the
NN
parameter associated with two different training sets is meaningless.
All the Best.
RL
Declan Keane wrote:
>
> Hi Roy:
>
> The following is a rather straightforward issue, so I wonder if we
can
> easily move it to the stage of being "fully resolved" in the final
CSC
> report:
>
> DRAPER@ucdphy.ucdavis.edu wrote:
> <snip>
> > --------------------------
> > Regarding the factor 2.6 increase in K0s yield, we most recently
said:
> > ...
> > (c) You still have not said whether you changed the size of the
parameter
> > whose value determines what is assigned as a Ks0.
TRUNCATE HERE, AS THIS IS QUOTATION FROM AN EARLIER EMAIL (JED).
> I think you have told me explicitly that the gate on the output neuron
is
> always the same. I'm not saying it has to be that way, and I understand
> that the inv. mass selection is separate from the NN, but if the
answer is
> an unconditional YES, all of the above can be laid to rest immediately.
> Furthermore, if the answer is NO, that explicit NO answer would also
help
> simplify Jim's wrap-up.
>
> Declan
#####################################
FINI