Closed GoogleCodeExporter closed 8 years ago
This is actually a feature of the IEEE 754 floating point standard and not a
bug in EJML. Take a look at this webpage to learn more:
http://www.concentric.net/~ttwang/tech/javafloat.htm
I double checked the JavaDoc for isIdentical() and it does correctly describe
its behavior. However this is a confusion issue so I have added a comment to
the JavaDoc about NaN and infinite numbers. The functions hasNaN() and
hasUncountable() in MatrixFeatures might be of interest.
Original comment by peter.ab...@gmail.com
on 23 Dec 2010 at 1:14
It is true that any comparison with NaN yields false according to IEEE 754, so
the condition "if( diff > tol )" is false if "diff" is NaN. But this simply
means that the method is probably not implemented correctly, because:
a) A call to isIdentical(m1, m2) should yield exactly the same result as
isIdentical(m1, m2, 0.0) for all matrices m1 and m2, but this is currently
violated
b) I cannot imagine anyone wanting isIdentical(matrixWithNaSsOnly,
matrixWithoutNaNs, 0.0) to return true
Note that JUnit's assertEquals(double a, double b, double tol) is implemented
as follows:
if (Double.compare(expected, actual) == 0)
return;
if (!(Math.abs(expected - actual) <= delta))
failNotEquals(message, new Double(expected), new Double(actual));
MatrixFeatures.isIdential only implements the analog of the second condition,
which, IMHO, returns an incorrect result if any of the two values in the
comparisons is NaN.
Original comment by kaspar.thommen@gmail.com
on 29 Dec 2010 at 3:48
Actually, the link you posted points out this problem in the section "Danger of
Floating-Point Number Optimization":
! (a < b) does not imply that (a >= b).
! (1.0 < NaN) -> true
(1.0 >= NaN) -> false
Original comment by kaspar.thommen@gmail.com
on 29 Dec 2010 at 3:54
I do think those are good points and the result is counter intuitive. Maybe
have isIdentical() do a more rigorous check similar to what JUnit does and
create a new function isEquals() that will only produce good results if all the
numbers are countable. The reason for two different functions is that I
suspect that isEquals() would be much faster than isIdentical().
Original comment by peter.ab...@gmail.com
on 30 Dec 2010 at 9:13
I think the only thing that needs to be done in order to fix isIdentical(m1,
m2, tol) is to replace line 267
if( diff > tol ) {
by
if (!(diff <= tol)) {
This change will ensure that non-NaN diff values won't change the condition,
but if diff is NaN, the condition will be true instead of false now, so the
method would return false, which makes more sense IMHO.
However, it would also mean that two matrices with all NaNs would NOT be
considered equal, which is in accordance with the '==' operator on doubles, but
different from JUnit's behavior and the behavior of Double.equals(). Honestly,
I'm not sure which one to prefer here...
I'm not sure an additional isEquals() is necessary - when would anyone use
isEquals() instead of isIdentical(), and for what cases would they produce
different results?
Original comment by kaspar.thommen@gmail.com
on 3 Jan 2011 at 11:23
Original comment by peter.ab...@gmail.com
on 5 Jan 2011 at 5:59
MatrixFeatures now contains isEquals and isIdentical. isEquals performs the
standard equality test but is designed to always return false if any of the
elements are NaN or infinite. isIdentical() explicitly handles uncountable
numbers and checks to see if they are the same symbol. See SVN version 249 for
this change.
The justification for isEquals is that often a matrix having an uncountable
number is consider a failure condition. Instead of explicitly checking for NaN
or infinity and equality, you can just check for equality.
isIdentical() handles the situation where one wants to make sure each element
is equal to the other or if an element is NaN/Infinite both elements have the
same symbol.
Sound good?
Original comment by peter.ab...@gmail.com
on 12 Jan 2011 at 1:58
Hi Peter,
I'm almost happy :-) Why are you considering two infinite numbers as not being
equal? This contradicts all Java conventions known to me: The '==' operator,
Double.compare(), Double.equals(), JUnit's assertEquals() all consider two
positive or two negative infinity values to be equal.
Regards,
Kaspar
Original comment by kaspar.thommen@gmail.com
on 14 Jan 2011 at 1:51
If you call isEquals(a,b) and not isEquals(a,b,tol) (or set tol to zero) then
you will get what you want. The reason for this is that infinity minus
infinity is infinity.
Original comment by peter.ab...@gmail.com
on 14 Jan 2011 at 3:50
>> "infinity minus infinity is infinity"
I don't think so, try: System.out.println(Double.POSITIVE_INFINITY -
Double.POSITIVE_INFINITY)
Original comment by kaspar.thommen@gmail.com
on 24 Jan 2011 at 7:09
that's a mistake, infinity minus infinity is NaN. However the conclusion still
doesn't change and isEquals(a,b,tol) is following the expected behavior of
floating point arithmetic. It's testing if the magnitude of two numbers are
within tolerance not the == operator.
Original comment by peter.ab...@gmail.com
on 24 Jan 2011 at 12:57
I'm assuming that things are all well and good now so I'm closing this ticket.
Let me know if this is not the case.
Original comment by peter.ab...@gmail.com
on 4 Feb 2011 at 8:59
Original issue reported on code.google.com by
kaspar.thommen@gmail.com
on 23 Dec 2010 at 7:31