My Math Forum  

Go Back   My Math Forum > College Math Forum > Real Analysis

Real Analysis Real Analysis Math Forum


Reply
 
LinkBack Thread Tools Display Modes
May 23rd, 2010, 04:50 PM   #1
Newbie
 
Joined: May 2010

Posts: 3
Thanks: 0

Precision Arithmetic: A New Floating-Point Arithmetic

ABSTRACT: A new floating-point arithmetic called precision arithmetic is developed to track precision for arithmetic calculations. It uses a novel rounding scheme to avoid excessive rounding error propagation of conventional floating-point arithmetic. Unlike interval arithmetic, its uncertainty tracking is based on statistics and its bounding range is much tighter. Generic standards and systematic methods for validating uncertainty-bearing arithmetics are discussed. The precision arithmetic is found to be better than interval arithmetic in uncertainty-tracking for linear algorithms.

PDF: http://arxiv.org/ftp/cs/papers/0606/0606103.pdf

Please comment. You can also email your comment to me Chengpu@gmail.com directly. I am thinking of start a source forge project on it.
Chengpu is offline  
 
May 23rd, 2010, 07:38 PM   #2
Global Moderator
 
CRGreathouse's Avatar
 
Joined: Nov 2006
From: UTC -5

Posts: 16,046
Thanks: 938

Math Focus: Number theory, computational mathematics, combinatorics, FOM, symbolic logic, TCS, algorithms
Re: Precision Arithmetic: A New Floating-Point Arithmetic

I think the topic is very interesting, something I've worked on before. I'm curious to see how well it performs in practice as well as in theory -- for example, calculating tan(120152108392) or (X x)(Y y)(Z z) for x,y,z uncorrelated and positive.

***

On page 4 you write: "According to central limit theorem [4], most measured values are Gaussian distributed, which has no limited bounding range." But that's not what the central limit theorem says...

On page 4 you also have: "If an bounding range is defined using a statistical rule on bounding leakage, such as the 6?-10^-6 rule for Gaussian distribution [4] (which says the bounding leakage is about 10^-6 for a bounding range of mean 6-fold of standard deviations), there is no guarantee that the calculation result will also obey the 6?-10^-6 rule using interval arithmetic, since interval arithmetic has no statistical ground." Can you give an example of where this fails? I would expect interval arithmetic to give far wider ranges, in general, than 6 standard deviations if that's what the data was given as.

I have no other specific comments through page 14.
CRGreathouse is offline  
May 25th, 2010, 08:25 PM   #3
Newbie
 
Joined: May 2010

Posts: 3
Thanks: 0

Re: Precision Arithmetic: A New Floating-Point Arithmetic

The Gaussian distribution is the de facto underline distribution and central limit theorem is one of the major reason. For example, if you collect data thru A/D, A/D holds signal by integrate signals in capacitor, thus averages signal in hardware. A/D are now the majority way to collect signal. The other way is thru event counter, whose result is usually either Posson distribution or binominal distribution, both of which can be approx well by Gaussian distribution. Of cause, I should have said that the Gaussian distribution is the de facto underline distribution. Thanks.

In my result, Interval arithmetic always broadens the bounding range too fast, e.g, if you start with 6-sigma, after one or two calculations, it become 12-sigma or more. For FFT algorithm, independence arithmetic is the ideal one. You can see this effect clearly.
Chengpu is offline  
June 2nd, 2010, 11:37 AM   #4
Global Moderator
 
CRGreathouse's Avatar
 
Joined: Nov 2006
From: UTC -5

Posts: 16,046
Thanks: 938

Math Focus: Number theory, computational mathematics, combinatorics, FOM, symbolic logic, TCS, algorithms
Re: Precision Arithmetic: A New Floating-Point Arithmetic

Any progress on the SourceForge idea?
CRGreathouse is offline  
Reply

  My Math Forum > College Math Forum > Real Analysis

Tags
arithmetic, floatingpoint, precision



Thread Tools
Display Modes


Similar Threads
Thread Thread Starter Forum Replies Last Post
floating point numbers with modulus(mod) abady-1 Number Theory 5 September 21st, 2013 02:08 PM
Floating Point Notation... Primalscientist Computer Science 9 June 2nd, 2011 06:29 AM
Arbitrary-precision Integer Calculator for bignum arithmetic TuringBombe Number Theory 4 January 25th, 2011 06:55 PM
Precision Arithmetic: A New Floating-Point Arithmetic Chengpu Linear Algebra 0 May 23rd, 2010 04:55 PM
Floating Point Numbers - Single Precision Format (help!) RonnieStokie Algebra 3 December 12th, 2007 09:36 PM





Copyright © 2019 My Math Forum. All rights reserved.