How can one get in touch with the designers of Excel?

[Omnibus]

Hello Everyone,

This is my first time posting a message in this forum and I hope this here is the right section to post it: Does anyone know who the concrete programmers who designed Excel (say, Excel 2010) within Microsoft are and how one can get in touch with them? Thanks in advance.

 

[Omnibus]

OK, I looked at column C using the formula with and without the offset of -3.356. The absolute values of the numbers in the two instances are practically of the same order of magnitude, save some discrepancies with the beginning and the ending several numbers (and, probably one may argue, there are some more numbers down the line which are an order of magnitude less in absolute value when there is no offset). But even with these discrepancies why should there be such a difference when doing the different calculations (which should lead to the same outcome) on one and the same set of numbers when there is offset and no difference at all when there is no offset (again, I call offset the number -3.356).

 

[Omnibus]

Because one of the limitations in Excel is that it can only store numbers 15-digits long.

As for getting in touch with the MS programmers, I doubt MS would let you do that. The nearest you'll get is MS Tech Support.

You might find the observations in following links eye-opening.
www.forecastingprinciples.com/paperpdf/McCullough.pdf
http://www.practicalstats.com/xlsstats/excelstats.html

Like I said, I understand that but here we're talking about numbers which are practically of the same order of magnitude and the absolute values of their integer part is within the non-zero integer range. So why should the 15 digit accuracy of the same set of numbers play a role in one instance of calculations and not play a role when calculations are done differently? Furthermore, why should there be no difference how you calculate it when the offset of -3.356 is gone?

EDIT: Thanks for the links.

 

[shg]

Floating point numbers get added MOL the same way you do, by aligning the decimal points of each addend . So

      ---------A--------- ----B-----
  1   3.35690037245743000           
  2   0.00478395110765485           
  3   3.36168432356508000 A3: =A1+A2

The last three digits in the first addend aren't there (they are beyond the 15 digits available). The second addend has 15 significant digits, but three are shifted beyond the available precison when the decimals are aligned. So you lose three significant digits in second addend.

 

[Omnibus]

Floating point numbers get added MOL the same way you do, by aligning the decimal points of each addend . So

      ---------A--------- ----B-----
  1   3.35690037245743000           
  2   0.00478395110765485           
  3   3.36168432356508000 A3: =A1+A2

The last three digits in the first addend aren't there (they are beyond the 15 digits available). The second addend has 15 significant digits, but three are shifted beyond the available precison when the decimals are aligned. So you lose three significant digits in second addend.

Yes, but look at the numbers in the first, say, three places after the decimal point which we are actually interested in. They are not affected even if we round it the way the machine would by cutting it off at the, say, seventh or eighth place.

 

[shg]

That's true. But if you keep doing it, the error accumulates.

 

[Omnibus]

That's true. But if you keep doing it, the error accumulates.

So, then why would it accumulate in only one of the instances and not accumulate in the other three?

 

[Omnibus]

Why isn't there a discrepancy in the four different methods of calculation when there is no offset? The numbers are practically of the same order of magnitude? Why don't we see an accumulation of error there?

 

[shg]

So, then why would it accumulate in only one of the instances and not accumulate in the other three?

Sorry, I lacked the ambition to try to understand what your formulas were doing in detail. Your model may have an error, or the result may be due entirely to numerical precision, or ...

Structuring models with consideration for numerical precision is a common problem. You can find many examples where model formulation is driven specifically by precision considerations.

Microsoft has changed many function implementations over the years for improved accuracy (BETADIST are NORMSDIST are two that come to mind, and both require numerical integration). Some of those changes may just involve using more iterations, but I'd bet that others were implemented by changing the model.

 

[Omnibus]

@shg4421,

Thanks for the discussion. I'll have to put some more thought into it. Also, I gather folks here don't think it would do any good to try to get in touch with the designers of Excel (even if I manage to reach them). Thanks everybody.