Thread: Round integer division
Is it documented anywhere that floating-point numbers round "scientifically", that is 0.5 rounds to the nearest even number? Compare: SELECT round(2.5::real), round(2.5::numeric), round(3.5::real), round(3.5::numeric); generates 2 | 3 | 4 | 4 I stumbled across this when I was trying to use round(a::real/b::real) to generate a rounded result to dividing integers, and noticed sometimes 0.5 was truncated and sometimes it was rounded up. Couldn't find anything about this in the archives or the data type documentation. Is there something obvious that I'm I missing? Thanks, -- Lee Hachadoorian PhD Student, Geography Program in Earth & Environmental Sciences CUNY Graduate Center
On Fri, Jun 25, 2010 at 6:53 PM, Lee Hachadoorian <lee.hachadoorian@gmail.com> wrote: > Is it documented anywhere that floating-point numbers round > "scientifically", that is 0.5 rounds to the nearest even number? Compare: > > SELECT round(2.5::real), round(2.5::numeric), round(3.5::real), > round(3.5::numeric); > > generates > > 2 | 3 | 4 | 4 > > I stumbled across this when I was trying to use round(a::real/b::real) > to generate a rounded result to dividing integers, and noticed sometimes > 0.5 was truncated and sometimes it was rounded up. Couldn't find > anything about this in the archives or the data type documentation. Is > there something obvious that I'm I missing? That all floating point representations are approximate?
On 06/25/2010 07:00 PM, Scott Marlowe wrote: > That all floating point representations are approximate? > But if it's error due to approximation, shouldn't the result be random? I tried this for a handful of larger numbers, and it appears to consistently round to the even number. Wouldn't that have to be intentionally programmed that way? --Lee -- Lee Hachadoorian PhD Student, Geography Program in Earth & Environmental Sciences CUNY Graduate Center
On Friday 25 June 2010 3:53:01 pm Lee Hachadoorian wrote: > Is it documented anywhere that floating-point numbers round > "scientifically", that is 0.5 rounds to the nearest even number? Compare: > > SELECT round(2.5::real), round(2.5::numeric), round(3.5::real), > round(3.5::numeric); > > generates > > 2 | 3 | 4 | 4 > > I stumbled across this when I was trying to use round(a::real/b::real) > to generate a rounded result to dividing integers, and noticed sometimes > 0.5 was truncated and sometimes it was rounded up. Couldn't find > anything about this in the archives or the data type documentation. Is > there something obvious that I'm I missing? > > Thanks, > > -- > > Lee Hachadoorian > PhD Student, Geography > Program in Earth & Environmental Sciences > CUNY Graduate Center See here for explanation: http://archives.postgresql.org/pgsql-general/2010-03/msg00969.php -- Adrian Klaver adrian.klaver@gmail.com
Thanks. That was a fairly recent post, too, but I couldn't come up with the right keyword search to find it. On 06/25/2010 07:25 PM, Adrian Klaver wrote: > On Friday 25 June 2010 3:53:01 pm Lee Hachadoorian wrote: > >> Is it documented anywhere that floating-point numbers round >> "scientifically", that is 0.5 rounds to the nearest even number? >> > See here for explanation: > http://archives.postgresql.org/pgsql-general/2010-03/msg00969.php > > -- Lee Hachadoorian PhD Student, Geography Program in Earth & Environmental Sciences CUNY Graduate Center
On 2010-06-25, Lee Hachadoorian <lee.hachadoorian@gmail.com> wrote: > Is it documented anywhere that floating-point numbers round > "scientifically", that is 0.5 rounds to the nearest even number? That's swiss rounding. And no, as I understand it documented that most arithmetic) is platform specific. Postgres is written in C and the relevant portions of the C standards douments (and discussions thereof) give a good picture of the functioning of postgres arithmetic.
On 2010-06-25, Lee Hachadoorian <lee.hachadoorian@gmail.com> wrote: > > > On 06/25/2010 07:00 PM, Scott Marlowe wrote: >> That all floating point representations are approximate? >> > But if it's error due to approximation, shouldn't the result be random? > I tried this for a handful of larger numbers, and it appears to > consistently round to the even number. Wouldn't that have to be > intentionally programmed that way? Yes it is, an enginneer at intel made that decision in the 70s when the 8087 coprocessor was developed. If you're not running on x86-like hardware the round operation may do something different.
On 06/26/10 17:09, wrote: > On 2010-06-25, Lee Hachadoorian<lee.hachadoorian@gmail.com> wrote: >> Is it documented anywhere that floating-point numbers round >> "scientifically", that is 0.5 rounds to the nearest even number? > > That's swiss rounding. And no, as I understand it documented that > most arithmetic) is platform specific. > > Postgres is written in C and the relevant portions of the > C standards douments (and discussions thereof) give a good > picture of the functioning of postgres arithmetic. IT IS REALLY SAD ! the round(Numeric) does not behave as Math science prescribes :( Shame on PostgresQL !