Fixed Point Arithmetic

From: David Cressey <cressey73_at_verizon.net>
Date: Thu, 29 Mar 2007 15:32:11 GMT
Message-ID: <%ZQOh.9320$fA2.1606_at_trndny02>



The recent threads on possreps and floating point made me think about fixed point arithmetic, again. I'm going to refer you to the Wikipedia entry on fixed point arithmetic, as a starting point for our discussion.

There's a comparison in there between floating point, binary fixed point, and so called BCD fixed point, or decimal fixed point. At the risk of incorrect usage, I'm going to start out by discussing decimal fixed point arithmetic with a scale factor of 2. The scale factor of two means that there are two digits after the decimal point. The total precision of the fixed point numbers depends on the way they are stored.

An example of a literal in this format is 12345.67

This format is useful for storing and calculating currency amounts, where the smallest unit is one hundredth of the standard unit. If the standard unit is the US dollar, then the penny is the smallest unit. The UK pound is decimalized in this fashion, if I recall correctly.

In the abstract, these FP numbers exhibit closure over addition and subtraction. They also exhibit closure over multiplication by an integer, because multiplication by an integer can be reduced to repetitive addtion or subtraction. In finite computing, additions may produce overflow where floating point numbers would have produced roundoff error instead.

These FP numbers do not have closure over multiplication, because the product may have up to four significant digits after the decimal point, and therefore have to be rounded off for storage purposes.

Anyway, I'm going to say that FP numbers with a scale factor of 2 save an enormous amount of head-scratching when working with bean counting applications. You won't spend time looking for the missing penny. OTOH, if you use these numbers for complicated arithmetic (example: amortization schedule) you have to think about roundoff errors.

When it comes to the stock market, it depends on the country. AFAIK, the NYSE still records stock "points" in units of dollars, with binary fractional points like 1/4, 1/32, and the like. The use of decimal FPs with this kind of data will introduce round off problems that would have been avoided with binary floating point numbers.

It's my understanding that the Australian stock market recently went through a transition from binary fractions like the NYSE to decimal fractions, like the currency. I would have expected some problems to surface in the IT systems that massage that data for precisely this reason.

Does anyone here have any experience with the decimalization of the Australian stock market? Did it cause problems? Did it obviate some prior roundoff problems?

Are fixed point numbers a useful from for DBMSes to support? Received on Thu Mar 29 2007 - 17:32:11 CEST

Original text of this message