numerical methods - error bound in function approximation algorithm -


suppose have set of floating point number "m" bit mantissa , "e" bits exponent. suppose more on want approximate function "f".

from theory know "range reduced function" used , such function derive global function value.

for example let x = (sx,ex,mx) (sign exp , mantissa) then... log2(x) = ex + log2(1.mx) range reduced function "log2(1.mx)".

i have implemented @ present reciprocal, square root, log2 , exp2, i've started work trigonometric functions. wandering if given global error bound (ulp error especially) possible derive error bound range reduced function, there study kind of problem? speaking of log2(x) (as example) lke able say...

"ok want log2(x) k ulp error, achieve given our floating point system need approximate log2(1.mx) p ulp error"

remember said know working floating point number, format generic, classic f32, example e=10, m = 8 end on.

i can't find reference shows such kind of study. reference have (i.e. muller book) doesn't treat topic in way looking kind of paper or similar. know reference?

i'm trying derive such bound myself not easy...

there description of current practice, along proposed improvement , error analysis, @ https://hal.inria.fr/ensl-00086904/document. description of current practice appears consistent overview @ https://docs.oracle.com/cd/e37069_01/html/e39019/z4000ac119729.html, consistent memory of talked problem being mod pi range reduction of trigonometric functions.

i think ieee floating point big step forwards because standardized things @ time when there variety of computer architectures, lowering risks of porting code between them, accuracy requirements implied may have been overkill: many problems constraint on accuracy of output accuracy of input data, not accuracy of calculation of intermediate values.


Comments

Popular posts from this blog

c# - Validate object ID from GET to POST -

node.js - Custom Model Validator SailsJS -

php - Find a regex to take part of Email -