Thread: [GENERAL] Does NUMERIC lose precision?

Hi everyone. I’m trying to perform some exact precision arithmetic with PostgreSQL’s NUMERIC type. However I can’t seem
toget the unparameterized NUMERIC type to perform exact precision arithmetic: 

# SELECT 2::NUMERIC ^ 64;
               ?column?
---------------------------------------
 18446744073709551616.0000000000000000
(1 row)

While the above operation works fine once I divide 1 by that number the result is an inexact decimal number:

# SELECT 1 / (2::NUMERIC ^ 64);
                ?column?
----------------------------------------
 0.000000000000000000054210108624275222
(1 row)

It doesn't seem to be an issue with the output either as taking the reciprocal yields a different number than I started
with:

# SELECT 1 / (1 / (2::NUMERIC ^ 64));
                         ?column?
-----------------------------------------------------------
 18446744073709551514.042092759729171265910020841463748922
(1 row)

The only way to get an exact result is by specifying an explicit precision and scale:

# SELECT 1 / (2::NUMERIC(96, 64) ^ 64);
                              ?column?
--------------------------------------------------------------------
 0.0000000000000000000542101086242752217003726400434970855712890625
(1 row)

# SELECT 1 / (1 / (2::NUMERIC(96, 64) ^ 64));
                                       ?column?
---------------------------------------------------------------------------------------
 18446744073709551616.0000000000000000000000000000000000000000000000000000000000000000
(1 row)

However this does not seem intuitive from the documentation which states that:

Specifying:

NUMERIC

without any precision or scale creates a column in which numeric values of any precision and scale can be stored, up to
theimplementation limit on precision. A column of this kind will not coerce input values to any particular scale... 

Kaiting Chen <ktchen14@gmail.com> writes:
> Hi everyone. I’m trying to perform some exact precision arithmetic with PostgreSQL’s NUMERIC type. However I can’t
seemto get the unparameterized NUMERIC type to perform exact precision arithmetic: 

Division is inherently inexact, except in special cases.  If you think
that 1/(2^64) should be carried out to enough digits to be exact,
what would you have us do with 1/3?

The actual behavior is that it will choose a result scale (number of
digits after the decimal point) that is dependent on the scales of the
input arguments, but not on their precise values.  I don't recall the
details beyond that.

            regards, tom lane