How to specify constants accurately to DP precision?

I like code that is highly portable even from one programming language to another. For widely used languages, there are FOSS parsers that manage the translation of the code’s syntax. However, when it comes to specifying the value of a constant to DP precision, a variety of problems may arise. For example, the software architects might have decided not to allow you to specify as many digits as are necessary for a unique DP number or you might not know the decimal separator used at execution time.

Imho, the most reliable way to exchange DP numbers across all kinds of borders is to store the DP number as a sequence of 16 hex symbols together with the endianness used for encoding. Obviously, this requires writing as well as testing routines for encoding and decoding, resp., in all programming languages of interest. Moreover, usually the code of high-level programming languages is mainly designed for readability by humans. Most programmers will realise that the sequence "3.14159265358979324" specifies Pi, but I doubt that many of them will notice "LEDP: 40 09 21 FB 54 44 2D 18" (where LE stands in for little endian) has exactly the same meaning.

When looking for an alternative that is highly portable and readable as well, I came across the method of rational approximation, which is valuable particularly in the regime 1E-3 to 1E+3. Usually, a product or ratio of two n-digit numbers has got 2·n significant digits. 17 digits are necessary to specify a unique DP number. Thus, it might be expected that it is possible to uniquely approximate a DP number by a ratio of two integer numbers with at maximum 9 digits, at least, if its magnitude is not too different from one. The following figure demonstrates how the ratios of two integers approximate Pi the closer the more significant digits are used.

rational approximation to Pi

The ratio of the two integers having the smallest amount of significant figures yet exactly reproducing the DP representation of Pi is found to be 245850922 / 78256779 . The figure also gives the notation of the ratios as continued fractions in standard (only positive coefficients) as well as extended (negative coefficients allowed, too) notation. The deviations that can be observed between the blue and red points, resp., in the regime beyond the DP accuracy limit just show the effect of this limit on calculations. Only the values obtained by using Maxima’s data type "bigfloat" are correct.

Concluding, pseudo-code statements like

Double Precision Constant Pi = CDBLE("245850922")/CDBLE("78256779");

are my favourites in the future. It will be interesting for me to observe how long faulty statements like this one

var pi = 3.14159265358979;

survive in documents found on the web. That statement does not yield Pi but a value corresponding to "LEDP: 40 09 21 FB 54 44 2D 11" that is seven ULPs smaller than Pi!


Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: