Subtopic 9 of 9
4.9Precision & rounding
The idea
Floating-point numbers are approximations. Rounding and limited bits mean the usual rules of arithmetic do not always hold.
Think of it like this: A ruler that only has 1 mm marks: you cannot measure 0.3 mm exactly, and adding many small measurements can lose them.
Why it exists
Understanding rounding explains subtle bugs and some GATE trick questions.
How it works
- Many decimals (0.1) have no exact binary form.
- To add, the smaller exponent is aligned, shifting out low bits.
- Adding a tiny number to a huge one can lose the tiny one entirely.
- Hence floating-point addition is not associative.
Worked example: (1e20 + −1e20) + 1 vs 1e20 + (−1e20 + 1)
First gives 1, second gives 0 (the 1 is lost inside −1e20).
Now play with it
- →Read the live results of each experiment; they are computed by your browser.
0.1 + 0.2 (double)0.300000000000000040.1 and 0.2 have no exact binary form, so the sum is slightly off.
(1e20 + −1e20) + 1 (double)1Cancels first, then adds 1: exactly 1.
1e20 + (−1e20 + 1) (double)0The 1 is absorbed by −1e20 (not enough mantissa bits): result 0. Not associative!
16777216 + 1 (single)167772162^24: single precision has 24 significant bits, so +1 is lost.
16777216 + 2 (single)16777218+2 is representable at this magnitude (spacing is 2).
Check yourself
Is floating-point addition associative?
Remember: Floating point is approximate: order of operations matters.
Exam notes (the dense version)
Mantissa has hidden leading 1. Floating-point addition: align exponents (shift smaller), add, normalise, round. Not associative!
Previous year questions: Number Representation & Computer Arithmetic (19)
PYQs are grouped by topic, so this is the full set for Number Representation & Computer Arithmetic; try the ones that match this subtopic. Open all 19 on the topic page.