You should specify that you only deal with direct mode AD. Then in the abstract 
the derivative program computes the function f and its derivatives f'. The 
percentage you quote are not general enough to be used as arguments in the 
abstract. 

Author resPonse: Only qualified percentages remain in the abstract. 
Direct mode stipulated.

As you do not define precisely "activity analysis" you should cite a paper that 
does it. For example in pages 1393...1395 
"An automatic differentiation platform: Odyssee", C.Faure, Future Generation 
Computer Systems 21 (2005) 1391--1400. 

Author response: added.

In section 1 you use the term "input/output of conern" instead of the standard 
term "(in-)dependent". 

Author response: changes made.

p 2. section 2. :The last sentence is not correct. If the number of independent is 
small the gain could be large. 

Author response: changes made.

p 3. section 2. :The term "inaccuracies" is not good. It is accurate but cannot be 
decide using static analysis. The paragraph from  "Activity analysis" to "not be 
the case" should be rewritten. 

Author response: paragraph has been rewritten.

p 3. section 3. :Why did you choose ADIC and OpenAD. The fact that your analysis 
is done with fortran and C makes it weaker as both languages cannot be compared. 

Author response: ADIC and OpenAD were chosen because they were readily available
to the authors.  

p4. section 3. : In paragraph "At runtime ... "to inactive". You do not give the 
activity flag in the second case. 

Author response: activity flag is global to the sentence and pertains to both 
cases.  The value of the activity flag is given for each case.

In tables 1.2. it would be clearer if you put the number of input/output w.r.t the 
total number of variables instead on the names of variables. 

Author response: The names of the variables are given to document which AD code 
was generated for our experiments.  Should anyone want to attempt to duplicate our 
results, they would want to know the independent and dependent variables we 
stipulated.  However, the total number of variables (in the original code as well 
as in the AD code) would be an interesting statistic.  The size category gives the 
dimension sizes of independent array variables.

p5. section 4.1: You should not talk about your scripting tool as it is used to 
generate a library if I am correct. The fact that is is generated or hand written 
does not change anything. 
Give a clear name to the strategies. The only meaningfull information is 
static/dynamic/no activity analysis. You give at the same time informations about 
implementation: macro/proc/hand. It makes the tables very unclear. 

Author response: Good point.  Removed all references to various implementation modes.  Showed only autoMacro as Dynamic.

P 6. section 4.2: You analysis the results about percentage of activity in the 
variable. But this information appears no where in the tables. Can you add it. 

Author response: yes. It is now included in the C benchmarks graph.

p 7. section 5.1.: Why did you choose a must-may activity instead of a must-must 
analsis?  An overhead is still present ...

Author response: Our proposed hybrid approach looks to decrease the
number of activity flag checks by statically determining the memory
refences of the mayActive variables where we know they will be MustVary.
It is at those memory references that we can exclude the activity check.
The knowledge of MustUseful as opposed to MayUseful does not add any
usable information for our hybrid approach.  I believe the overhead that
you claim is still present is that of derivitive code that may no longer
be needed at the point in the code where a mayActive variable can be
classified as notUseful.  If a variable can be classified as not useful
then it is not considered active and no dynamic check is done.

p 7. fig 5.: Why did you choose these examples ? 

All along the paper, you do not seem to face the problem of activity analysis 
w.r.t to arrays or structures. If you consider all the cells of an an array to 
have the same activity status you should say so. 
This by itself leads to an overhead in static mode, but the dynamic activity 
analysis does not solve the problem. Or you should duplicate the arrays. 

Author response: Currently in OpenAnalysis, the activity of an array cell agrees 
with the activity of the array.  It is currently on the OpenAnalysis "to do" list 
to expand their treatment of arrays to subsets of array cells.  Dynamically, of 
course, we can track the activity of individual array cells.